| File: | build/source/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp |
| Warning: | line 8987, column 3 Use of memory after it is freed |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===// | ||||||
| 2 | // | ||||||
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | ||||||
| 4 | // See https://llvm.org/LICENSE.txt for license information. | ||||||
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | ||||||
| 6 | // | ||||||
| 7 | //===----------------------------------------------------------------------===// | ||||||
| 8 | // | ||||||
| 9 | // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops | ||||||
| 10 | // and generates target-independent LLVM-IR. | ||||||
| 11 | // The vectorizer uses the TargetTransformInfo analysis to estimate the costs | ||||||
| 12 | // of instructions in order to estimate the profitability of vectorization. | ||||||
| 13 | // | ||||||
| 14 | // The loop vectorizer combines consecutive loop iterations into a single | ||||||
| 15 | // 'wide' iteration. After this transformation the index is incremented | ||||||
| 16 | // by the SIMD vector width, and not by one. | ||||||
| 17 | // | ||||||
| 18 | // This pass has three parts: | ||||||
| 19 | // 1. The main loop pass that drives the different parts. | ||||||
| 20 | // 2. LoopVectorizationLegality - A unit that checks for the legality | ||||||
| 21 | // of the vectorization. | ||||||
| 22 | // 3. InnerLoopVectorizer - A unit that performs the actual | ||||||
| 23 | // widening of instructions. | ||||||
| 24 | // 4. LoopVectorizationCostModel - A unit that checks for the profitability | ||||||
| 25 | // of vectorization. It decides on the optimal vector width, which | ||||||
| 26 | // can be one, if vectorization is not profitable. | ||||||
| 27 | // | ||||||
| 28 | // There is a development effort going on to migrate loop vectorizer to the | ||||||
| 29 | // VPlan infrastructure and to introduce outer loop vectorization support (see | ||||||
| 30 | // docs/Proposal/VectorizationPlan.rst and | ||||||
| 31 | // http://lists.llvm.org/pipermail/llvm-dev/2017-December/119523.html). For this | ||||||
| 32 | // purpose, we temporarily introduced the VPlan-native vectorization path: an | ||||||
| 33 | // alternative vectorization path that is natively implemented on top of the | ||||||
| 34 | // VPlan infrastructure. See EnableVPlanNativePath for enabling. | ||||||
| 35 | // | ||||||
| 36 | //===----------------------------------------------------------------------===// | ||||||
| 37 | // | ||||||
| 38 | // The reduction-variable vectorization is based on the paper: | ||||||
| 39 | // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization. | ||||||
| 40 | // | ||||||
| 41 | // Variable uniformity checks are inspired by: | ||||||
| 42 | // Karrenberg, R. and Hack, S. Whole Function Vectorization. | ||||||
| 43 | // | ||||||
| 44 | // The interleaved access vectorization is based on the paper: | ||||||
| 45 | // Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved | ||||||
| 46 | // Data for SIMD | ||||||
| 47 | // | ||||||
| 48 | // Other ideas/concepts are from: | ||||||
| 49 | // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later. | ||||||
| 50 | // | ||||||
| 51 | // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of | ||||||
| 52 | // Vectorizing Compilers. | ||||||
| 53 | // | ||||||
| 54 | //===----------------------------------------------------------------------===// | ||||||
| 55 | |||||||
| 56 | #include "llvm/Transforms/Vectorize/LoopVectorize.h" | ||||||
| 57 | #include "LoopVectorizationPlanner.h" | ||||||
| 58 | #include "VPRecipeBuilder.h" | ||||||
| 59 | #include "VPlan.h" | ||||||
| 60 | #include "VPlanHCFGBuilder.h" | ||||||
| 61 | #include "VPlanTransforms.h" | ||||||
| 62 | #include "llvm/ADT/APInt.h" | ||||||
| 63 | #include "llvm/ADT/ArrayRef.h" | ||||||
| 64 | #include "llvm/ADT/DenseMap.h" | ||||||
| 65 | #include "llvm/ADT/DenseMapInfo.h" | ||||||
| 66 | #include "llvm/ADT/Hashing.h" | ||||||
| 67 | #include "llvm/ADT/MapVector.h" | ||||||
| 68 | #include "llvm/ADT/STLExtras.h" | ||||||
| 69 | #include "llvm/ADT/SmallPtrSet.h" | ||||||
| 70 | #include "llvm/ADT/SmallSet.h" | ||||||
| 71 | #include "llvm/ADT/SmallVector.h" | ||||||
| 72 | #include "llvm/ADT/Statistic.h" | ||||||
| 73 | #include "llvm/ADT/StringRef.h" | ||||||
| 74 | #include "llvm/ADT/Twine.h" | ||||||
| 75 | #include "llvm/ADT/iterator_range.h" | ||||||
| 76 | #include "llvm/Analysis/AssumptionCache.h" | ||||||
| 77 | #include "llvm/Analysis/BasicAliasAnalysis.h" | ||||||
| 78 | #include "llvm/Analysis/BlockFrequencyInfo.h" | ||||||
| 79 | #include "llvm/Analysis/CFG.h" | ||||||
| 80 | #include "llvm/Analysis/CodeMetrics.h" | ||||||
| 81 | #include "llvm/Analysis/DemandedBits.h" | ||||||
| 82 | #include "llvm/Analysis/GlobalsModRef.h" | ||||||
| 83 | #include "llvm/Analysis/LoopAccessAnalysis.h" | ||||||
| 84 | #include "llvm/Analysis/LoopAnalysisManager.h" | ||||||
| 85 | #include "llvm/Analysis/LoopInfo.h" | ||||||
| 86 | #include "llvm/Analysis/LoopIterator.h" | ||||||
| 87 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | ||||||
| 88 | #include "llvm/Analysis/ProfileSummaryInfo.h" | ||||||
| 89 | #include "llvm/Analysis/ScalarEvolution.h" | ||||||
| 90 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" | ||||||
| 91 | #include "llvm/Analysis/TargetLibraryInfo.h" | ||||||
| 92 | #include "llvm/Analysis/TargetTransformInfo.h" | ||||||
| 93 | #include "llvm/Analysis/ValueTracking.h" | ||||||
| 94 | #include "llvm/Analysis/VectorUtils.h" | ||||||
| 95 | #include "llvm/IR/Attributes.h" | ||||||
| 96 | #include "llvm/IR/BasicBlock.h" | ||||||
| 97 | #include "llvm/IR/CFG.h" | ||||||
| 98 | #include "llvm/IR/Constant.h" | ||||||
| 99 | #include "llvm/IR/Constants.h" | ||||||
| 100 | #include "llvm/IR/DataLayout.h" | ||||||
| 101 | #include "llvm/IR/DebugInfoMetadata.h" | ||||||
| 102 | #include "llvm/IR/DebugLoc.h" | ||||||
| 103 | #include "llvm/IR/DerivedTypes.h" | ||||||
| 104 | #include "llvm/IR/DiagnosticInfo.h" | ||||||
| 105 | #include "llvm/IR/Dominators.h" | ||||||
| 106 | #include "llvm/IR/Function.h" | ||||||
| 107 | #include "llvm/IR/IRBuilder.h" | ||||||
| 108 | #include "llvm/IR/InstrTypes.h" | ||||||
| 109 | #include "llvm/IR/Instruction.h" | ||||||
| 110 | #include "llvm/IR/Instructions.h" | ||||||
| 111 | #include "llvm/IR/IntrinsicInst.h" | ||||||
| 112 | #include "llvm/IR/Intrinsics.h" | ||||||
| 113 | #include "llvm/IR/Metadata.h" | ||||||
| 114 | #include "llvm/IR/Module.h" | ||||||
| 115 | #include "llvm/IR/Operator.h" | ||||||
| 116 | #include "llvm/IR/PatternMatch.h" | ||||||
| 117 | #include "llvm/IR/Type.h" | ||||||
| 118 | #include "llvm/IR/Use.h" | ||||||
| 119 | #include "llvm/IR/User.h" | ||||||
| 120 | #include "llvm/IR/Value.h" | ||||||
| 121 | #include "llvm/IR/ValueHandle.h" | ||||||
| 122 | #include "llvm/IR/Verifier.h" | ||||||
| 123 | #include "llvm/InitializePasses.h" | ||||||
| 124 | #include "llvm/Pass.h" | ||||||
| 125 | #include "llvm/Support/Casting.h" | ||||||
| 126 | #include "llvm/Support/CommandLine.h" | ||||||
| 127 | #include "llvm/Support/Compiler.h" | ||||||
| 128 | #include "llvm/Support/Debug.h" | ||||||
| 129 | #include "llvm/Support/ErrorHandling.h" | ||||||
| 130 | #include "llvm/Support/InstructionCost.h" | ||||||
| 131 | #include "llvm/Support/MathExtras.h" | ||||||
| 132 | #include "llvm/Support/raw_ostream.h" | ||||||
| 133 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | ||||||
| 134 | #include "llvm/Transforms/Utils/InjectTLIMappings.h" | ||||||
| 135 | #include "llvm/Transforms/Utils/LoopSimplify.h" | ||||||
| 136 | #include "llvm/Transforms/Utils/LoopUtils.h" | ||||||
| 137 | #include "llvm/Transforms/Utils/LoopVersioning.h" | ||||||
| 138 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" | ||||||
| 139 | #include "llvm/Transforms/Utils/SizeOpts.h" | ||||||
| 140 | #include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h" | ||||||
| 141 | #include <algorithm> | ||||||
| 142 | #include <cassert> | ||||||
| 143 | #include <cmath> | ||||||
| 144 | #include <cstdint> | ||||||
| 145 | #include <functional> | ||||||
| 146 | #include <iterator> | ||||||
| 147 | #include <limits> | ||||||
| 148 | #include <map> | ||||||
| 149 | #include <memory> | ||||||
| 150 | #include <string> | ||||||
| 151 | #include <tuple> | ||||||
| 152 | #include <utility> | ||||||
| 153 | |||||||
| 154 | using namespace llvm; | ||||||
| 155 | |||||||
| 156 | #define LV_NAME"loop-vectorize" "loop-vectorize" | ||||||
| 157 | #define DEBUG_TYPE"loop-vectorize" LV_NAME"loop-vectorize" | ||||||
| 158 | |||||||
| 159 | #ifndef NDEBUG | ||||||
| 160 | const char VerboseDebug[] = DEBUG_TYPE"loop-vectorize" "-verbose"; | ||||||
| 161 | #endif | ||||||
| 162 | |||||||
| 163 | /// @{ | ||||||
| 164 | /// Metadata attribute names | ||||||
| 165 | const char LLVMLoopVectorizeFollowupAll[] = "llvm.loop.vectorize.followup_all"; | ||||||
| 166 | const char LLVMLoopVectorizeFollowupVectorized[] = | ||||||
| 167 | "llvm.loop.vectorize.followup_vectorized"; | ||||||
| 168 | const char LLVMLoopVectorizeFollowupEpilogue[] = | ||||||
| 169 | "llvm.loop.vectorize.followup_epilogue"; | ||||||
| 170 | /// @} | ||||||
| 171 | |||||||
| 172 | STATISTIC(LoopsVectorized, "Number of loops vectorized")static llvm::Statistic LoopsVectorized = {"loop-vectorize", "LoopsVectorized" , "Number of loops vectorized"}; | ||||||
| 173 | STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization")static llvm::Statistic LoopsAnalyzed = {"loop-vectorize", "LoopsAnalyzed" , "Number of loops analyzed for vectorization"}; | ||||||
| 174 | STATISTIC(LoopsEpilogueVectorized, "Number of epilogues vectorized")static llvm::Statistic LoopsEpilogueVectorized = {"loop-vectorize" , "LoopsEpilogueVectorized", "Number of epilogues vectorized" }; | ||||||
| 175 | |||||||
| 176 | static cl::opt<bool> EnableEpilogueVectorization( | ||||||
| 177 | "enable-epilogue-vectorization", cl::init(true), cl::Hidden, | ||||||
| 178 | cl::desc("Enable vectorization of epilogue loops.")); | ||||||
| 179 | |||||||
| 180 | static cl::opt<unsigned> EpilogueVectorizationForceVF( | ||||||
| 181 | "epilogue-vectorization-force-VF", cl::init(1), cl::Hidden, | ||||||
| 182 | cl::desc("When epilogue vectorization is enabled, and a value greater than " | ||||||
| 183 | "1 is specified, forces the given VF for all applicable epilogue " | ||||||
| 184 | "loops.")); | ||||||
| 185 | |||||||
| 186 | static cl::opt<unsigned> EpilogueVectorizationMinVF( | ||||||
| 187 | "epilogue-vectorization-minimum-VF", cl::init(16), cl::Hidden, | ||||||
| 188 | cl::desc("Only loops with vectorization factor equal to or larger than " | ||||||
| 189 | "the specified value are considered for epilogue vectorization.")); | ||||||
| 190 | |||||||
| 191 | /// Loops with a known constant trip count below this number are vectorized only | ||||||
| 192 | /// if no scalar iteration overheads are incurred. | ||||||
| 193 | static cl::opt<unsigned> TinyTripCountVectorThreshold( | ||||||
| 194 | "vectorizer-min-trip-count", cl::init(16), cl::Hidden, | ||||||
| 195 | cl::desc("Loops with a constant trip count that is smaller than this " | ||||||
| 196 | "value are vectorized only if no scalar iteration overheads " | ||||||
| 197 | "are incurred.")); | ||||||
| 198 | |||||||
| 199 | static cl::opt<unsigned> VectorizeMemoryCheckThreshold( | ||||||
| 200 | "vectorize-memory-check-threshold", cl::init(128), cl::Hidden, | ||||||
| 201 | cl::desc("The maximum allowed number of runtime memory checks")); | ||||||
| 202 | |||||||
| 203 | // Option prefer-predicate-over-epilogue indicates that an epilogue is undesired, | ||||||
| 204 | // that predication is preferred, and this lists all options. I.e., the | ||||||
| 205 | // vectorizer will try to fold the tail-loop (epilogue) into the vector body | ||||||
| 206 | // and predicate the instructions accordingly. If tail-folding fails, there are | ||||||
| 207 | // different fallback strategies depending on these values: | ||||||
| 208 | namespace PreferPredicateTy { | ||||||
| 209 | enum Option { | ||||||
| 210 | ScalarEpilogue = 0, | ||||||
| 211 | PredicateElseScalarEpilogue, | ||||||
| 212 | PredicateOrDontVectorize | ||||||
| 213 | }; | ||||||
| 214 | } // namespace PreferPredicateTy | ||||||
| 215 | |||||||
| 216 | static cl::opt<PreferPredicateTy::Option> PreferPredicateOverEpilogue( | ||||||
| 217 | "prefer-predicate-over-epilogue", | ||||||
| 218 | cl::init(PreferPredicateTy::ScalarEpilogue), | ||||||
| 219 | cl::Hidden, | ||||||
| 220 | cl::desc("Tail-folding and predication preferences over creating a scalar " | ||||||
| 221 | "epilogue loop."), | ||||||
| 222 | cl::values(clEnumValN(PreferPredicateTy::ScalarEpilogue,llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" } | ||||||
| 223 | "scalar-epilogue",llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" } | ||||||
| 224 | "Don't tail-predicate loops, create scalar epilogue")llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" }, | ||||||
| 225 | clEnumValN(PreferPredicateTy::PredicateElseScalarEpilogue,llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." } | ||||||
| 226 | "predicate-else-scalar-epilogue",llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." } | ||||||
| 227 | "prefer tail-folding, create scalar epilogue if tail "llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." } | ||||||
| 228 | "folding fails.")llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." }, | ||||||
| 229 | clEnumValN(PreferPredicateTy::PredicateOrDontVectorize,llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." } | ||||||
| 230 | "predicate-dont-vectorize",llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." } | ||||||
| 231 | "prefers tail-folding, don't attempt vectorization if "llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." } | ||||||
| 232 | "tail-folding fails.")llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." })); | ||||||
| 233 | |||||||
| 234 | static cl::opt<bool> MaximizeBandwidth( | ||||||
| 235 | "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden, | ||||||
| 236 | cl::desc("Maximize bandwidth when selecting vectorization factor which " | ||||||
| 237 | "will be determined by the smallest type in loop.")); | ||||||
| 238 | |||||||
| 239 | static cl::opt<bool> EnableInterleavedMemAccesses( | ||||||
| 240 | "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden, | ||||||
| 241 | cl::desc("Enable vectorization on interleaved memory accesses in a loop")); | ||||||
| 242 | |||||||
| 243 | /// An interleave-group may need masking if it resides in a block that needs | ||||||
| 244 | /// predication, or in order to mask away gaps. | ||||||
| 245 | static cl::opt<bool> EnableMaskedInterleavedMemAccesses( | ||||||
| 246 | "enable-masked-interleaved-mem-accesses", cl::init(false), cl::Hidden, | ||||||
| 247 | cl::desc("Enable vectorization on masked interleaved memory accesses in a loop")); | ||||||
| 248 | |||||||
| 249 | static cl::opt<unsigned> TinyTripCountInterleaveThreshold( | ||||||
| 250 | "tiny-trip-count-interleave-threshold", cl::init(128), cl::Hidden, | ||||||
| 251 | cl::desc("We don't interleave loops with a estimated constant trip count " | ||||||
| 252 | "below this number")); | ||||||
| 253 | |||||||
| 254 | static cl::opt<unsigned> ForceTargetNumScalarRegs( | ||||||
| 255 | "force-target-num-scalar-regs", cl::init(0), cl::Hidden, | ||||||
| 256 | cl::desc("A flag that overrides the target's number of scalar registers.")); | ||||||
| 257 | |||||||
| 258 | static cl::opt<unsigned> ForceTargetNumVectorRegs( | ||||||
| 259 | "force-target-num-vector-regs", cl::init(0), cl::Hidden, | ||||||
| 260 | cl::desc("A flag that overrides the target's number of vector registers.")); | ||||||
| 261 | |||||||
| 262 | static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor( | ||||||
| 263 | "force-target-max-scalar-interleave", cl::init(0), cl::Hidden, | ||||||
| 264 | cl::desc("A flag that overrides the target's max interleave factor for " | ||||||
| 265 | "scalar loops.")); | ||||||
| 266 | |||||||
| 267 | static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor( | ||||||
| 268 | "force-target-max-vector-interleave", cl::init(0), cl::Hidden, | ||||||
| 269 | cl::desc("A flag that overrides the target's max interleave factor for " | ||||||
| 270 | "vectorized loops.")); | ||||||
| 271 | |||||||
| 272 | static cl::opt<unsigned> ForceTargetInstructionCost( | ||||||
| 273 | "force-target-instruction-cost", cl::init(0), cl::Hidden, | ||||||
| 274 | cl::desc("A flag that overrides the target's expected cost for " | ||||||
| 275 | "an instruction to a single constant value. Mostly " | ||||||
| 276 | "useful for getting consistent testing.")); | ||||||
| 277 | |||||||
| 278 | static cl::opt<bool> ForceTargetSupportsScalableVectors( | ||||||
| 279 | "force-target-supports-scalable-vectors", cl::init(false), cl::Hidden, | ||||||
| 280 | cl::desc( | ||||||
| 281 | "Pretend that scalable vectors are supported, even if the target does " | ||||||
| 282 | "not support them. This flag should only be used for testing.")); | ||||||
| 283 | |||||||
| 284 | static cl::opt<unsigned> SmallLoopCost( | ||||||
| 285 | "small-loop-cost", cl::init(20), cl::Hidden, | ||||||
| 286 | cl::desc( | ||||||
| 287 | "The cost of a loop that is considered 'small' by the interleaver.")); | ||||||
| 288 | |||||||
| 289 | static cl::opt<bool> LoopVectorizeWithBlockFrequency( | ||||||
| 290 | "loop-vectorize-with-block-frequency", cl::init(true), cl::Hidden, | ||||||
| 291 | cl::desc("Enable the use of the block frequency analysis to access PGO " | ||||||
| 292 | "heuristics minimizing code growth in cold regions and being more " | ||||||
| 293 | "aggressive in hot regions.")); | ||||||
| 294 | |||||||
| 295 | // Runtime interleave loops for load/store throughput. | ||||||
| 296 | static cl::opt<bool> EnableLoadStoreRuntimeInterleave( | ||||||
| 297 | "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden, | ||||||
| 298 | cl::desc( | ||||||
| 299 | "Enable runtime interleaving until load/store ports are saturated")); | ||||||
| 300 | |||||||
| 301 | /// Interleave small loops with scalar reductions. | ||||||
| 302 | static cl::opt<bool> InterleaveSmallLoopScalarReduction( | ||||||
| 303 | "interleave-small-loop-scalar-reduction", cl::init(false), cl::Hidden, | ||||||
| 304 | cl::desc("Enable interleaving for loops with small iteration counts that " | ||||||
| 305 | "contain scalar reductions to expose ILP.")); | ||||||
| 306 | |||||||
| 307 | /// The number of stores in a loop that are allowed to need predication. | ||||||
| 308 | static cl::opt<unsigned> NumberOfStoresToPredicate( | ||||||
| 309 | "vectorize-num-stores-pred", cl::init(1), cl::Hidden, | ||||||
| 310 | cl::desc("Max number of stores to be predicated behind an if.")); | ||||||
| 311 | |||||||
| 312 | static cl::opt<bool> EnableIndVarRegisterHeur( | ||||||
| 313 | "enable-ind-var-reg-heur", cl::init(true), cl::Hidden, | ||||||
| 314 | cl::desc("Count the induction variable only once when interleaving")); | ||||||
| 315 | |||||||
| 316 | static cl::opt<bool> EnableCondStoresVectorization( | ||||||
| 317 | "enable-cond-stores-vec", cl::init(true), cl::Hidden, | ||||||
| 318 | cl::desc("Enable if predication of stores during vectorization.")); | ||||||
| 319 | |||||||
| 320 | static cl::opt<unsigned> MaxNestedScalarReductionIC( | ||||||
| 321 | "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden, | ||||||
| 322 | cl::desc("The maximum interleave count to use when interleaving a scalar " | ||||||
| 323 | "reduction in a nested loop.")); | ||||||
| 324 | |||||||
| 325 | static cl::opt<bool> | ||||||
| 326 | PreferInLoopReductions("prefer-inloop-reductions", cl::init(false), | ||||||
| 327 | cl::Hidden, | ||||||
| 328 | cl::desc("Prefer in-loop vector reductions, " | ||||||
| 329 | "overriding the targets preference.")); | ||||||
| 330 | |||||||
| 331 | static cl::opt<bool> ForceOrderedReductions( | ||||||
| 332 | "force-ordered-reductions", cl::init(false), cl::Hidden, | ||||||
| 333 | cl::desc("Enable the vectorisation of loops with in-order (strict) " | ||||||
| 334 | "FP reductions")); | ||||||
| 335 | |||||||
| 336 | static cl::opt<bool> PreferPredicatedReductionSelect( | ||||||
| 337 | "prefer-predicated-reduction-select", cl::init(false), cl::Hidden, | ||||||
| 338 | cl::desc( | ||||||
| 339 | "Prefer predicating a reduction operation over an after loop select.")); | ||||||
| 340 | |||||||
| 341 | cl::opt<bool> EnableVPlanNativePath( | ||||||
| 342 | "enable-vplan-native-path", cl::init(false), cl::Hidden, | ||||||
| 343 | cl::desc("Enable VPlan-native vectorization path with " | ||||||
| 344 | "support for outer loop vectorization.")); | ||||||
| 345 | |||||||
| 346 | // This flag enables the stress testing of the VPlan H-CFG construction in the | ||||||
| 347 | // VPlan-native vectorization path. It must be used in conjuction with | ||||||
| 348 | // -enable-vplan-native-path. -vplan-verify-hcfg can also be used to enable the | ||||||
| 349 | // verification of the H-CFGs built. | ||||||
| 350 | static cl::opt<bool> VPlanBuildStressTest( | ||||||
| 351 | "vplan-build-stress-test", cl::init(false), cl::Hidden, | ||||||
| 352 | cl::desc( | ||||||
| 353 | "Build VPlan for every supported loop nest in the function and bail " | ||||||
| 354 | "out right after the build (stress test the VPlan H-CFG construction " | ||||||
| 355 | "in the VPlan-native vectorization path).")); | ||||||
| 356 | |||||||
| 357 | cl::opt<bool> llvm::EnableLoopInterleaving( | ||||||
| 358 | "interleave-loops", cl::init(true), cl::Hidden, | ||||||
| 359 | cl::desc("Enable loop interleaving in Loop vectorization passes")); | ||||||
| 360 | cl::opt<bool> llvm::EnableLoopVectorization( | ||||||
| 361 | "vectorize-loops", cl::init(true), cl::Hidden, | ||||||
| 362 | cl::desc("Run the Loop vectorization passes")); | ||||||
| 363 | |||||||
| 364 | static cl::opt<bool> PrintVPlansInDotFormat( | ||||||
| 365 | "vplan-print-in-dot-format", cl::Hidden, | ||||||
| 366 | cl::desc("Use dot format instead of plain text when dumping VPlans")); | ||||||
| 367 | |||||||
| 368 | static cl::opt<cl::boolOrDefault> ForceSafeDivisor( | ||||||
| 369 | "force-widen-divrem-via-safe-divisor", cl::Hidden, | ||||||
| 370 | cl::desc( | ||||||
| 371 | "Override cost based safe divisor widening for div/rem instructions")); | ||||||
| 372 | |||||||
| 373 | /// A helper function that returns true if the given type is irregular. The | ||||||
| 374 | /// type is irregular if its allocated size doesn't equal the store size of an | ||||||
| 375 | /// element of the corresponding vector type. | ||||||
| 376 | static bool hasIrregularType(Type *Ty, const DataLayout &DL) { | ||||||
| 377 | // Determine if an array of N elements of type Ty is "bitcast compatible" | ||||||
| 378 | // with a <N x Ty> vector. | ||||||
| 379 | // This is only true if there is no padding between the array elements. | ||||||
| 380 | return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty); | ||||||
| 381 | } | ||||||
| 382 | |||||||
| 383 | /// A helper function that returns the reciprocal of the block probability of | ||||||
| 384 | /// predicated blocks. If we return X, we are assuming the predicated block | ||||||
| 385 | /// will execute once for every X iterations of the loop header. | ||||||
| 386 | /// | ||||||
| 387 | /// TODO: We should use actual block probability here, if available. Currently, | ||||||
| 388 | /// we always assume predicated blocks have a 50% chance of executing. | ||||||
| 389 | static unsigned getReciprocalPredBlockProb() { return 2; } | ||||||
| 390 | |||||||
| 391 | /// A helper function that returns an integer or floating-point constant with | ||||||
| 392 | /// value C. | ||||||
| 393 | static Constant *getSignedIntOrFpConstant(Type *Ty, int64_t C) { | ||||||
| 394 | return Ty->isIntegerTy() ? ConstantInt::getSigned(Ty, C) | ||||||
| 395 | : ConstantFP::get(Ty, C); | ||||||
| 396 | } | ||||||
| 397 | |||||||
| 398 | /// Returns "best known" trip count for the specified loop \p L as defined by | ||||||
| 399 | /// the following procedure: | ||||||
| 400 | /// 1) Returns exact trip count if it is known. | ||||||
| 401 | /// 2) Returns expected trip count according to profile data if any. | ||||||
| 402 | /// 3) Returns upper bound estimate if it is known. | ||||||
| 403 | /// 4) Returns std::nullopt if all of the above failed. | ||||||
| 404 | static std::optional<unsigned> getSmallBestKnownTC(ScalarEvolution &SE, | ||||||
| 405 | Loop *L) { | ||||||
| 406 | // Check if exact trip count is known. | ||||||
| 407 | if (unsigned ExpectedTC = SE.getSmallConstantTripCount(L)) | ||||||
| 408 | return ExpectedTC; | ||||||
| 409 | |||||||
| 410 | // Check if there is an expected trip count available from profile data. | ||||||
| 411 | if (LoopVectorizeWithBlockFrequency) | ||||||
| 412 | if (auto EstimatedTC = getLoopEstimatedTripCount(L)) | ||||||
| 413 | return *EstimatedTC; | ||||||
| 414 | |||||||
| 415 | // Check if upper bound estimate is known. | ||||||
| 416 | if (unsigned ExpectedTC = SE.getSmallConstantMaxTripCount(L)) | ||||||
| 417 | return ExpectedTC; | ||||||
| 418 | |||||||
| 419 | return std::nullopt; | ||||||
| 420 | } | ||||||
| 421 | |||||||
| 422 | // Forward declare GeneratedRTChecks. | ||||||
| 423 | class GeneratedRTChecks; | ||||||
| 424 | |||||||
| 425 | namespace llvm { | ||||||
| 426 | |||||||
| 427 | AnalysisKey ShouldRunExtraVectorPasses::Key; | ||||||
| 428 | |||||||
| 429 | /// InnerLoopVectorizer vectorizes loops which contain only one basic | ||||||
| 430 | /// block to a specified vectorization factor (VF). | ||||||
| 431 | /// This class performs the widening of scalars into vectors, or multiple | ||||||
| 432 | /// scalars. This class also implements the following features: | ||||||
| 433 | /// * It inserts an epilogue loop for handling loops that don't have iteration | ||||||
| 434 | /// counts that are known to be a multiple of the vectorization factor. | ||||||
| 435 | /// * It handles the code generation for reduction variables. | ||||||
| 436 | /// * Scalarization (implementation using scalars) of un-vectorizable | ||||||
| 437 | /// instructions. | ||||||
| 438 | /// InnerLoopVectorizer does not perform any vectorization-legality | ||||||
| 439 | /// checks, and relies on the caller to check for the different legality | ||||||
| 440 | /// aspects. The InnerLoopVectorizer relies on the | ||||||
| 441 | /// LoopVectorizationLegality class to provide information about the induction | ||||||
| 442 | /// and reduction variables that were found to a given vectorization factor. | ||||||
| 443 | class InnerLoopVectorizer { | ||||||
| 444 | public: | ||||||
| 445 | InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE, | ||||||
| 446 | LoopInfo *LI, DominatorTree *DT, | ||||||
| 447 | const TargetLibraryInfo *TLI, | ||||||
| 448 | const TargetTransformInfo *TTI, AssumptionCache *AC, | ||||||
| 449 | OptimizationRemarkEmitter *ORE, ElementCount VecWidth, | ||||||
| 450 | ElementCount MinProfitableTripCount, | ||||||
| 451 | unsigned UnrollFactor, LoopVectorizationLegality *LVL, | ||||||
| 452 | LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, | ||||||
| 453 | ProfileSummaryInfo *PSI, GeneratedRTChecks &RTChecks) | ||||||
| 454 | : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI), | ||||||
| 455 | AC(AC), ORE(ORE), VF(VecWidth), UF(UnrollFactor), | ||||||
| 456 | Builder(PSE.getSE()->getContext()), Legal(LVL), Cost(CM), BFI(BFI), | ||||||
| 457 | PSI(PSI), RTChecks(RTChecks) { | ||||||
| 458 | // Query this against the original loop and save it here because the profile | ||||||
| 459 | // of the original loop header may change as the transformation happens. | ||||||
| 460 | OptForSizeBasedOnProfile = llvm::shouldOptimizeForSize( | ||||||
| 461 | OrigLoop->getHeader(), PSI, BFI, PGSOQueryType::IRPass); | ||||||
| 462 | |||||||
| 463 | if (MinProfitableTripCount.isZero()) | ||||||
| 464 | this->MinProfitableTripCount = VecWidth; | ||||||
| 465 | else | ||||||
| 466 | this->MinProfitableTripCount = MinProfitableTripCount; | ||||||
| 467 | } | ||||||
| 468 | |||||||
| 469 | virtual ~InnerLoopVectorizer() = default; | ||||||
| 470 | |||||||
| 471 | /// Create a new empty loop that will contain vectorized instructions later | ||||||
| 472 | /// on, while the old loop will be used as the scalar remainder. Control flow | ||||||
| 473 | /// is generated around the vectorized (and scalar epilogue) loops consisting | ||||||
| 474 | /// of various checks and bypasses. Return the pre-header block of the new | ||||||
| 475 | /// loop and the start value for the canonical induction, if it is != 0. The | ||||||
| 476 | /// latter is the case when vectorizing the epilogue loop. In the case of | ||||||
| 477 | /// epilogue vectorization, this function is overriden to handle the more | ||||||
| 478 | /// complex control flow around the loops. | ||||||
| 479 | virtual std::pair<BasicBlock *, Value *> createVectorizedLoopSkeleton(); | ||||||
| 480 | |||||||
| 481 | /// Fix the vectorized code, taking care of header phi's, live-outs, and more. | ||||||
| 482 | void fixVectorizedLoop(VPTransformState &State, VPlan &Plan); | ||||||
| 483 | |||||||
| 484 | // Return true if any runtime check is added. | ||||||
| 485 | bool areSafetyChecksAdded() { return AddedSafetyChecks; } | ||||||
| 486 | |||||||
| 487 | /// A type for vectorized values in the new loop. Each value from the | ||||||
| 488 | /// original loop, when vectorized, is represented by UF vector values in the | ||||||
| 489 | /// new unrolled loop, where UF is the unroll factor. | ||||||
| 490 | using VectorParts = SmallVector<Value *, 2>; | ||||||
| 491 | |||||||
| 492 | /// A helper function to scalarize a single Instruction in the innermost loop. | ||||||
| 493 | /// Generates a sequence of scalar instances for each lane between \p MinLane | ||||||
| 494 | /// and \p MaxLane, times each part between \p MinPart and \p MaxPart, | ||||||
| 495 | /// inclusive. Uses the VPValue operands from \p RepRecipe instead of \p | ||||||
| 496 | /// Instr's operands. | ||||||
| 497 | void scalarizeInstruction(const Instruction *Instr, | ||||||
| 498 | VPReplicateRecipe *RepRecipe, | ||||||
| 499 | const VPIteration &Instance, bool IfPredicateInstr, | ||||||
| 500 | VPTransformState &State); | ||||||
| 501 | |||||||
| 502 | /// Construct the vector value of a scalarized value \p V one lane at a time. | ||||||
| 503 | void packScalarIntoVectorValue(VPValue *Def, const VPIteration &Instance, | ||||||
| 504 | VPTransformState &State); | ||||||
| 505 | |||||||
| 506 | /// Try to vectorize interleaved access group \p Group with the base address | ||||||
| 507 | /// given in \p Addr, optionally masking the vector operations if \p | ||||||
| 508 | /// BlockInMask is non-null. Use \p State to translate given VPValues to IR | ||||||
| 509 | /// values in the vectorized loop. | ||||||
| 510 | void vectorizeInterleaveGroup(const InterleaveGroup<Instruction> *Group, | ||||||
| 511 | ArrayRef<VPValue *> VPDefs, | ||||||
| 512 | VPTransformState &State, VPValue *Addr, | ||||||
| 513 | ArrayRef<VPValue *> StoredValues, | ||||||
| 514 | VPValue *BlockInMask = nullptr); | ||||||
| 515 | |||||||
| 516 | /// Fix the non-induction PHIs in \p Plan. | ||||||
| 517 | void fixNonInductionPHIs(VPlan &Plan, VPTransformState &State); | ||||||
| 518 | |||||||
| 519 | /// Returns true if the reordering of FP operations is not allowed, but we are | ||||||
| 520 | /// able to vectorize with strict in-order reductions for the given RdxDesc. | ||||||
| 521 | bool useOrderedReductions(const RecurrenceDescriptor &RdxDesc); | ||||||
| 522 | |||||||
| 523 | /// Create a broadcast instruction. This method generates a broadcast | ||||||
| 524 | /// instruction (shuffle) for loop invariant values and for the induction | ||||||
| 525 | /// value. If this is the induction variable then we extend it to N, N+1, ... | ||||||
| 526 | /// this is needed because each iteration in the loop corresponds to a SIMD | ||||||
| 527 | /// element. | ||||||
| 528 | virtual Value *getBroadcastInstrs(Value *V); | ||||||
| 529 | |||||||
| 530 | // Returns the resume value (bc.merge.rdx) for a reduction as | ||||||
| 531 | // generated by fixReduction. | ||||||
| 532 | PHINode *getReductionResumeValue(const RecurrenceDescriptor &RdxDesc); | ||||||
| 533 | |||||||
| 534 | /// Create a new phi node for the induction variable \p OrigPhi to resume | ||||||
| 535 | /// iteration count in the scalar epilogue, from where the vectorized loop | ||||||
| 536 | /// left off. In cases where the loop skeleton is more complicated (eg. | ||||||
| 537 | /// epilogue vectorization) and the resume values can come from an additional | ||||||
| 538 | /// bypass block, the \p AdditionalBypass pair provides information about the | ||||||
| 539 | /// bypass block and the end value on the edge from bypass to this loop. | ||||||
| 540 | PHINode *createInductionResumeValue( | ||||||
| 541 | PHINode *OrigPhi, const InductionDescriptor &ID, | ||||||
| 542 | ArrayRef<BasicBlock *> BypassBlocks, | ||||||
| 543 | std::pair<BasicBlock *, Value *> AdditionalBypass = {nullptr, nullptr}); | ||||||
| 544 | |||||||
| 545 | protected: | ||||||
| 546 | friend class LoopVectorizationPlanner; | ||||||
| 547 | |||||||
| 548 | /// A small list of PHINodes. | ||||||
| 549 | using PhiVector = SmallVector<PHINode *, 4>; | ||||||
| 550 | |||||||
| 551 | /// A type for scalarized values in the new loop. Each value from the | ||||||
| 552 | /// original loop, when scalarized, is represented by UF x VF scalar values | ||||||
| 553 | /// in the new unrolled loop, where UF is the unroll factor and VF is the | ||||||
| 554 | /// vectorization factor. | ||||||
| 555 | using ScalarParts = SmallVector<SmallVector<Value *, 4>, 2>; | ||||||
| 556 | |||||||
| 557 | /// Set up the values of the IVs correctly when exiting the vector loop. | ||||||
| 558 | void fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II, | ||||||
| 559 | Value *VectorTripCount, Value *EndValue, | ||||||
| 560 | BasicBlock *MiddleBlock, BasicBlock *VectorHeader, | ||||||
| 561 | VPlan &Plan); | ||||||
| 562 | |||||||
| 563 | /// Handle all cross-iteration phis in the header. | ||||||
| 564 | void fixCrossIterationPHIs(VPTransformState &State); | ||||||
| 565 | |||||||
| 566 | /// Create the exit value of first order recurrences in the middle block and | ||||||
| 567 | /// update their users. | ||||||
| 568 | void fixFixedOrderRecurrence(VPFirstOrderRecurrencePHIRecipe *PhiR, | ||||||
| 569 | VPTransformState &State); | ||||||
| 570 | |||||||
| 571 | /// Create code for the loop exit value of the reduction. | ||||||
| 572 | void fixReduction(VPReductionPHIRecipe *Phi, VPTransformState &State); | ||||||
| 573 | |||||||
| 574 | /// Clear NSW/NUW flags from reduction instructions if necessary. | ||||||
| 575 | void clearReductionWrapFlags(VPReductionPHIRecipe *PhiR, | ||||||
| 576 | VPTransformState &State); | ||||||
| 577 | |||||||
| 578 | /// Iteratively sink the scalarized operands of a predicated instruction into | ||||||
| 579 | /// the block that was created for it. | ||||||
| 580 | void sinkScalarOperands(Instruction *PredInst); | ||||||
| 581 | |||||||
| 582 | /// Shrinks vector element sizes to the smallest bitwidth they can be legally | ||||||
| 583 | /// represented as. | ||||||
| 584 | void truncateToMinimalBitwidths(VPTransformState &State); | ||||||
| 585 | |||||||
| 586 | /// Returns (and creates if needed) the original loop trip count. | ||||||
| 587 | Value *getOrCreateTripCount(BasicBlock *InsertBlock); | ||||||
| 588 | |||||||
| 589 | /// Returns (and creates if needed) the trip count of the widened loop. | ||||||
| 590 | Value *getOrCreateVectorTripCount(BasicBlock *InsertBlock); | ||||||
| 591 | |||||||
| 592 | /// Returns a bitcasted value to the requested vector type. | ||||||
| 593 | /// Also handles bitcasts of vector<float> <-> vector<pointer> types. | ||||||
| 594 | Value *createBitOrPointerCast(Value *V, VectorType *DstVTy, | ||||||
| 595 | const DataLayout &DL); | ||||||
| 596 | |||||||
| 597 | /// Emit a bypass check to see if the vector trip count is zero, including if | ||||||
| 598 | /// it overflows. | ||||||
| 599 | void emitIterationCountCheck(BasicBlock *Bypass); | ||||||
| 600 | |||||||
| 601 | /// Emit a bypass check to see if all of the SCEV assumptions we've | ||||||
| 602 | /// had to make are correct. Returns the block containing the checks or | ||||||
| 603 | /// nullptr if no checks have been added. | ||||||
| 604 | BasicBlock *emitSCEVChecks(BasicBlock *Bypass); | ||||||
| 605 | |||||||
| 606 | /// Emit bypass checks to check any memory assumptions we may have made. | ||||||
| 607 | /// Returns the block containing the checks or nullptr if no checks have been | ||||||
| 608 | /// added. | ||||||
| 609 | BasicBlock *emitMemRuntimeChecks(BasicBlock *Bypass); | ||||||
| 610 | |||||||
| 611 | /// Emit basic blocks (prefixed with \p Prefix) for the iteration check, | ||||||
| 612 | /// vector loop preheader, middle block and scalar preheader. | ||||||
| 613 | void createVectorLoopSkeleton(StringRef Prefix); | ||||||
| 614 | |||||||
| 615 | /// Create new phi nodes for the induction variables to resume iteration count | ||||||
| 616 | /// in the scalar epilogue, from where the vectorized loop left off. | ||||||
| 617 | /// In cases where the loop skeleton is more complicated (eg. epilogue | ||||||
| 618 | /// vectorization) and the resume values can come from an additional bypass | ||||||
| 619 | /// block, the \p AdditionalBypass pair provides information about the bypass | ||||||
| 620 | /// block and the end value on the edge from bypass to this loop. | ||||||
| 621 | void createInductionResumeValues( | ||||||
| 622 | std::pair<BasicBlock *, Value *> AdditionalBypass = {nullptr, nullptr}); | ||||||
| 623 | |||||||
| 624 | /// Complete the loop skeleton by adding debug MDs, creating appropriate | ||||||
| 625 | /// conditional branches in the middle block, preparing the builder and | ||||||
| 626 | /// running the verifier. Return the preheader of the completed vector loop. | ||||||
| 627 | BasicBlock *completeLoopSkeleton(); | ||||||
| 628 | |||||||
| 629 | /// Collect poison-generating recipes that may generate a poison value that is | ||||||
| 630 | /// used after vectorization, even when their operands are not poison. Those | ||||||
| 631 | /// recipes meet the following conditions: | ||||||
| 632 | /// * Contribute to the address computation of a recipe generating a widen | ||||||
| 633 | /// memory load/store (VPWidenMemoryInstructionRecipe or | ||||||
| 634 | /// VPInterleaveRecipe). | ||||||
| 635 | /// * Such a widen memory load/store has at least one underlying Instruction | ||||||
| 636 | /// that is in a basic block that needs predication and after vectorization | ||||||
| 637 | /// the generated instruction won't be predicated. | ||||||
| 638 | void collectPoisonGeneratingRecipes(VPTransformState &State); | ||||||
| 639 | |||||||
| 640 | /// Allow subclasses to override and print debug traces before/after vplan | ||||||
| 641 | /// execution, when trace information is requested. | ||||||
| 642 | virtual void printDebugTracesAtStart(){}; | ||||||
| 643 | virtual void printDebugTracesAtEnd(){}; | ||||||
| 644 | |||||||
| 645 | /// The original loop. | ||||||
| 646 | Loop *OrigLoop; | ||||||
| 647 | |||||||
| 648 | /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies | ||||||
| 649 | /// dynamic knowledge to simplify SCEV expressions and converts them to a | ||||||
| 650 | /// more usable form. | ||||||
| 651 | PredicatedScalarEvolution &PSE; | ||||||
| 652 | |||||||
| 653 | /// Loop Info. | ||||||
| 654 | LoopInfo *LI; | ||||||
| 655 | |||||||
| 656 | /// Dominator Tree. | ||||||
| 657 | DominatorTree *DT; | ||||||
| 658 | |||||||
| 659 | /// Target Library Info. | ||||||
| 660 | const TargetLibraryInfo *TLI; | ||||||
| 661 | |||||||
| 662 | /// Target Transform Info. | ||||||
| 663 | const TargetTransformInfo *TTI; | ||||||
| 664 | |||||||
| 665 | /// Assumption Cache. | ||||||
| 666 | AssumptionCache *AC; | ||||||
| 667 | |||||||
| 668 | /// Interface to emit optimization remarks. | ||||||
| 669 | OptimizationRemarkEmitter *ORE; | ||||||
| 670 | |||||||
| 671 | /// The vectorization SIMD factor to use. Each vector will have this many | ||||||
| 672 | /// vector elements. | ||||||
| 673 | ElementCount VF; | ||||||
| 674 | |||||||
| 675 | ElementCount MinProfitableTripCount; | ||||||
| 676 | |||||||
| 677 | /// The vectorization unroll factor to use. Each scalar is vectorized to this | ||||||
| 678 | /// many different vector instructions. | ||||||
| 679 | unsigned UF; | ||||||
| 680 | |||||||
| 681 | /// The builder that we use | ||||||
| 682 | IRBuilder<> Builder; | ||||||
| 683 | |||||||
| 684 | // --- Vectorization state --- | ||||||
| 685 | |||||||
| 686 | /// The vector-loop preheader. | ||||||
| 687 | BasicBlock *LoopVectorPreHeader; | ||||||
| 688 | |||||||
| 689 | /// The scalar-loop preheader. | ||||||
| 690 | BasicBlock *LoopScalarPreHeader; | ||||||
| 691 | |||||||
| 692 | /// Middle Block between the vector and the scalar. | ||||||
| 693 | BasicBlock *LoopMiddleBlock; | ||||||
| 694 | |||||||
| 695 | /// The unique ExitBlock of the scalar loop if one exists. Note that | ||||||
| 696 | /// there can be multiple exiting edges reaching this block. | ||||||
| 697 | BasicBlock *LoopExitBlock; | ||||||
| 698 | |||||||
| 699 | /// The scalar loop body. | ||||||
| 700 | BasicBlock *LoopScalarBody; | ||||||
| 701 | |||||||
| 702 | /// A list of all bypass blocks. The first block is the entry of the loop. | ||||||
| 703 | SmallVector<BasicBlock *, 4> LoopBypassBlocks; | ||||||
| 704 | |||||||
| 705 | /// Store instructions that were predicated. | ||||||
| 706 | SmallVector<Instruction *, 4> PredicatedInstructions; | ||||||
| 707 | |||||||
| 708 | /// Trip count of the original loop. | ||||||
| 709 | Value *TripCount = nullptr; | ||||||
| 710 | |||||||
| 711 | /// Trip count of the widened loop (TripCount - TripCount % (VF*UF)) | ||||||
| 712 | Value *VectorTripCount = nullptr; | ||||||
| 713 | |||||||
| 714 | /// The legality analysis. | ||||||
| 715 | LoopVectorizationLegality *Legal; | ||||||
| 716 | |||||||
| 717 | /// The profitablity analysis. | ||||||
| 718 | LoopVectorizationCostModel *Cost; | ||||||
| 719 | |||||||
| 720 | // Record whether runtime checks are added. | ||||||
| 721 | bool AddedSafetyChecks = false; | ||||||
| 722 | |||||||
| 723 | // Holds the end values for each induction variable. We save the end values | ||||||
| 724 | // so we can later fix-up the external users of the induction variables. | ||||||
| 725 | DenseMap<PHINode *, Value *> IVEndValues; | ||||||
| 726 | |||||||
| 727 | /// BFI and PSI are used to check for profile guided size optimizations. | ||||||
| 728 | BlockFrequencyInfo *BFI; | ||||||
| 729 | ProfileSummaryInfo *PSI; | ||||||
| 730 | |||||||
| 731 | // Whether this loop should be optimized for size based on profile guided size | ||||||
| 732 | // optimizatios. | ||||||
| 733 | bool OptForSizeBasedOnProfile; | ||||||
| 734 | |||||||
| 735 | /// Structure to hold information about generated runtime checks, responsible | ||||||
| 736 | /// for cleaning the checks, if vectorization turns out unprofitable. | ||||||
| 737 | GeneratedRTChecks &RTChecks; | ||||||
| 738 | |||||||
| 739 | // Holds the resume values for reductions in the loops, used to set the | ||||||
| 740 | // correct start value of reduction PHIs when vectorizing the epilogue. | ||||||
| 741 | SmallMapVector<const RecurrenceDescriptor *, PHINode *, 4> | ||||||
| 742 | ReductionResumeValues; | ||||||
| 743 | }; | ||||||
| 744 | |||||||
| 745 | class InnerLoopUnroller : public InnerLoopVectorizer { | ||||||
| 746 | public: | ||||||
| 747 | InnerLoopUnroller(Loop *OrigLoop, PredicatedScalarEvolution &PSE, | ||||||
| 748 | LoopInfo *LI, DominatorTree *DT, | ||||||
| 749 | const TargetLibraryInfo *TLI, | ||||||
| 750 | const TargetTransformInfo *TTI, AssumptionCache *AC, | ||||||
| 751 | OptimizationRemarkEmitter *ORE, unsigned UnrollFactor, | ||||||
| 752 | LoopVectorizationLegality *LVL, | ||||||
| 753 | LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, | ||||||
| 754 | ProfileSummaryInfo *PSI, GeneratedRTChecks &Check) | ||||||
| 755 | : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | ||||||
| 756 | ElementCount::getFixed(1), | ||||||
| 757 | ElementCount::getFixed(1), UnrollFactor, LVL, CM, | ||||||
| 758 | BFI, PSI, Check) {} | ||||||
| 759 | |||||||
| 760 | private: | ||||||
| 761 | Value *getBroadcastInstrs(Value *V) override; | ||||||
| 762 | }; | ||||||
| 763 | |||||||
| 764 | /// Encapsulate information regarding vectorization of a loop and its epilogue. | ||||||
| 765 | /// This information is meant to be updated and used across two stages of | ||||||
| 766 | /// epilogue vectorization. | ||||||
| 767 | struct EpilogueLoopVectorizationInfo { | ||||||
| 768 | ElementCount MainLoopVF = ElementCount::getFixed(0); | ||||||
| 769 | unsigned MainLoopUF = 0; | ||||||
| 770 | ElementCount EpilogueVF = ElementCount::getFixed(0); | ||||||
| 771 | unsigned EpilogueUF = 0; | ||||||
| 772 | BasicBlock *MainLoopIterationCountCheck = nullptr; | ||||||
| 773 | BasicBlock *EpilogueIterationCountCheck = nullptr; | ||||||
| 774 | BasicBlock *SCEVSafetyCheck = nullptr; | ||||||
| 775 | BasicBlock *MemSafetyCheck = nullptr; | ||||||
| 776 | Value *TripCount = nullptr; | ||||||
| 777 | Value *VectorTripCount = nullptr; | ||||||
| 778 | |||||||
| 779 | EpilogueLoopVectorizationInfo(ElementCount MVF, unsigned MUF, | ||||||
| 780 | ElementCount EVF, unsigned EUF) | ||||||
| 781 | : MainLoopVF(MVF), MainLoopUF(MUF), EpilogueVF(EVF), EpilogueUF(EUF) { | ||||||
| 782 | assert(EUF == 1 &&(static_cast <bool> (EUF == 1 && "A high UF for the epilogue loop is likely not beneficial." ) ? void (0) : __assert_fail ("EUF == 1 && \"A high UF for the epilogue loop is likely not beneficial.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 783, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 783 | "A high UF for the epilogue loop is likely not beneficial.")(static_cast <bool> (EUF == 1 && "A high UF for the epilogue loop is likely not beneficial." ) ? void (0) : __assert_fail ("EUF == 1 && \"A high UF for the epilogue loop is likely not beneficial.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 783, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 784 | } | ||||||
| 785 | }; | ||||||
| 786 | |||||||
| 787 | /// An extension of the inner loop vectorizer that creates a skeleton for a | ||||||
| 788 | /// vectorized loop that has its epilogue (residual) also vectorized. | ||||||
| 789 | /// The idea is to run the vplan on a given loop twice, firstly to setup the | ||||||
| 790 | /// skeleton and vectorize the main loop, and secondly to complete the skeleton | ||||||
| 791 | /// from the first step and vectorize the epilogue. This is achieved by | ||||||
| 792 | /// deriving two concrete strategy classes from this base class and invoking | ||||||
| 793 | /// them in succession from the loop vectorizer planner. | ||||||
| 794 | class InnerLoopAndEpilogueVectorizer : public InnerLoopVectorizer { | ||||||
| 795 | public: | ||||||
| 796 | InnerLoopAndEpilogueVectorizer( | ||||||
| 797 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, | ||||||
| 798 | DominatorTree *DT, const TargetLibraryInfo *TLI, | ||||||
| 799 | const TargetTransformInfo *TTI, AssumptionCache *AC, | ||||||
| 800 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, | ||||||
| 801 | LoopVectorizationLegality *LVL, llvm::LoopVectorizationCostModel *CM, | ||||||
| 802 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, | ||||||
| 803 | GeneratedRTChecks &Checks) | ||||||
| 804 | : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | ||||||
| 805 | EPI.MainLoopVF, EPI.MainLoopVF, EPI.MainLoopUF, LVL, | ||||||
| 806 | CM, BFI, PSI, Checks), | ||||||
| 807 | EPI(EPI) {} | ||||||
| 808 | |||||||
| 809 | // Override this function to handle the more complex control flow around the | ||||||
| 810 | // three loops. | ||||||
| 811 | std::pair<BasicBlock *, Value *> createVectorizedLoopSkeleton() final { | ||||||
| 812 | return createEpilogueVectorizedLoopSkeleton(); | ||||||
| 813 | } | ||||||
| 814 | |||||||
| 815 | /// The interface for creating a vectorized skeleton using one of two | ||||||
| 816 | /// different strategies, each corresponding to one execution of the vplan | ||||||
| 817 | /// as described above. | ||||||
| 818 | virtual std::pair<BasicBlock *, Value *> | ||||||
| 819 | createEpilogueVectorizedLoopSkeleton() = 0; | ||||||
| 820 | |||||||
| 821 | /// Holds and updates state information required to vectorize the main loop | ||||||
| 822 | /// and its epilogue in two separate passes. This setup helps us avoid | ||||||
| 823 | /// regenerating and recomputing runtime safety checks. It also helps us to | ||||||
| 824 | /// shorten the iteration-count-check path length for the cases where the | ||||||
| 825 | /// iteration count of the loop is so small that the main vector loop is | ||||||
| 826 | /// completely skipped. | ||||||
| 827 | EpilogueLoopVectorizationInfo &EPI; | ||||||
| 828 | }; | ||||||
| 829 | |||||||
| 830 | /// A specialized derived class of inner loop vectorizer that performs | ||||||
| 831 | /// vectorization of *main* loops in the process of vectorizing loops and their | ||||||
| 832 | /// epilogues. | ||||||
| 833 | class EpilogueVectorizerMainLoop : public InnerLoopAndEpilogueVectorizer { | ||||||
| 834 | public: | ||||||
| 835 | EpilogueVectorizerMainLoop( | ||||||
| 836 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, | ||||||
| 837 | DominatorTree *DT, const TargetLibraryInfo *TLI, | ||||||
| 838 | const TargetTransformInfo *TTI, AssumptionCache *AC, | ||||||
| 839 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, | ||||||
| 840 | LoopVectorizationLegality *LVL, llvm::LoopVectorizationCostModel *CM, | ||||||
| 841 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, | ||||||
| 842 | GeneratedRTChecks &Check) | ||||||
| 843 | : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | ||||||
| 844 | EPI, LVL, CM, BFI, PSI, Check) {} | ||||||
| 845 | /// Implements the interface for creating a vectorized skeleton using the | ||||||
| 846 | /// *main loop* strategy (ie the first pass of vplan execution). | ||||||
| 847 | std::pair<BasicBlock *, Value *> createEpilogueVectorizedLoopSkeleton() final; | ||||||
| 848 | |||||||
| 849 | protected: | ||||||
| 850 | /// Emits an iteration count bypass check once for the main loop (when \p | ||||||
| 851 | /// ForEpilogue is false) and once for the epilogue loop (when \p | ||||||
| 852 | /// ForEpilogue is true). | ||||||
| 853 | BasicBlock *emitIterationCountCheck(BasicBlock *Bypass, bool ForEpilogue); | ||||||
| 854 | void printDebugTracesAtStart() override; | ||||||
| 855 | void printDebugTracesAtEnd() override; | ||||||
| 856 | }; | ||||||
| 857 | |||||||
| 858 | // A specialized derived class of inner loop vectorizer that performs | ||||||
| 859 | // vectorization of *epilogue* loops in the process of vectorizing loops and | ||||||
| 860 | // their epilogues. | ||||||
| 861 | class EpilogueVectorizerEpilogueLoop : public InnerLoopAndEpilogueVectorizer { | ||||||
| 862 | public: | ||||||
| 863 | EpilogueVectorizerEpilogueLoop( | ||||||
| 864 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, | ||||||
| 865 | DominatorTree *DT, const TargetLibraryInfo *TLI, | ||||||
| 866 | const TargetTransformInfo *TTI, AssumptionCache *AC, | ||||||
| 867 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, | ||||||
| 868 | LoopVectorizationLegality *LVL, llvm::LoopVectorizationCostModel *CM, | ||||||
| 869 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, | ||||||
| 870 | GeneratedRTChecks &Checks) | ||||||
| 871 | : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | ||||||
| 872 | EPI, LVL, CM, BFI, PSI, Checks) { | ||||||
| 873 | TripCount = EPI.TripCount; | ||||||
| 874 | } | ||||||
| 875 | /// Implements the interface for creating a vectorized skeleton using the | ||||||
| 876 | /// *epilogue loop* strategy (ie the second pass of vplan execution). | ||||||
| 877 | std::pair<BasicBlock *, Value *> createEpilogueVectorizedLoopSkeleton() final; | ||||||
| 878 | |||||||
| 879 | protected: | ||||||
| 880 | /// Emits an iteration count bypass check after the main vector loop has | ||||||
| 881 | /// finished to see if there are any iterations left to execute by either | ||||||
| 882 | /// the vector epilogue or the scalar epilogue. | ||||||
| 883 | BasicBlock *emitMinimumVectorEpilogueIterCountCheck( | ||||||
| 884 | BasicBlock *Bypass, | ||||||
| 885 | BasicBlock *Insert); | ||||||
| 886 | void printDebugTracesAtStart() override; | ||||||
| 887 | void printDebugTracesAtEnd() override; | ||||||
| 888 | }; | ||||||
| 889 | } // end namespace llvm | ||||||
| 890 | |||||||
| 891 | /// Look for a meaningful debug location on the instruction or it's | ||||||
| 892 | /// operands. | ||||||
| 893 | static Instruction *getDebugLocFromInstOrOperands(Instruction *I) { | ||||||
| 894 | if (!I) | ||||||
| 895 | return I; | ||||||
| 896 | |||||||
| 897 | DebugLoc Empty; | ||||||
| 898 | if (I->getDebugLoc() != Empty) | ||||||
| 899 | return I; | ||||||
| 900 | |||||||
| 901 | for (Use &Op : I->operands()) { | ||||||
| 902 | if (Instruction *OpInst = dyn_cast<Instruction>(Op)) | ||||||
| 903 | if (OpInst->getDebugLoc() != Empty) | ||||||
| 904 | return OpInst; | ||||||
| 905 | } | ||||||
| 906 | |||||||
| 907 | return I; | ||||||
| 908 | } | ||||||
| 909 | |||||||
| 910 | /// Write a \p DebugMsg about vectorization to the debug output stream. If \p I | ||||||
| 911 | /// is passed, the message relates to that particular instruction. | ||||||
| 912 | #ifndef NDEBUG | ||||||
| 913 | static void debugVectorizationMessage(const StringRef Prefix, | ||||||
| 914 | const StringRef DebugMsg, | ||||||
| 915 | Instruction *I) { | ||||||
| 916 | dbgs() << "LV: " << Prefix << DebugMsg; | ||||||
| 917 | if (I != nullptr) | ||||||
| 918 | dbgs() << " " << *I; | ||||||
| 919 | else | ||||||
| 920 | dbgs() << '.'; | ||||||
| 921 | dbgs() << '\n'; | ||||||
| 922 | } | ||||||
| 923 | #endif | ||||||
| 924 | |||||||
| 925 | /// Create an analysis remark that explains why vectorization failed | ||||||
| 926 | /// | ||||||
| 927 | /// \p PassName is the name of the pass (e.g. can be AlwaysPrint). \p | ||||||
| 928 | /// RemarkName is the identifier for the remark. If \p I is passed it is an | ||||||
| 929 | /// instruction that prevents vectorization. Otherwise \p TheLoop is used for | ||||||
| 930 | /// the location of the remark. \return the remark object that can be | ||||||
| 931 | /// streamed to. | ||||||
| 932 | static OptimizationRemarkAnalysis createLVAnalysis(const char *PassName, | ||||||
| 933 | StringRef RemarkName, Loop *TheLoop, Instruction *I) { | ||||||
| 934 | Value *CodeRegion = TheLoop->getHeader(); | ||||||
| 935 | DebugLoc DL = TheLoop->getStartLoc(); | ||||||
| 936 | |||||||
| 937 | if (I) { | ||||||
| 938 | CodeRegion = I->getParent(); | ||||||
| 939 | // If there is no debug location attached to the instruction, revert back to | ||||||
| 940 | // using the loop's. | ||||||
| 941 | if (I->getDebugLoc()) | ||||||
| 942 | DL = I->getDebugLoc(); | ||||||
| 943 | } | ||||||
| 944 | |||||||
| 945 | return OptimizationRemarkAnalysis(PassName, RemarkName, DL, CodeRegion); | ||||||
| 946 | } | ||||||
| 947 | |||||||
| 948 | namespace llvm { | ||||||
| 949 | |||||||
| 950 | /// Return a value for Step multiplied by VF. | ||||||
| 951 | Value *createStepForVF(IRBuilderBase &B, Type *Ty, ElementCount VF, | ||||||
| 952 | int64_t Step) { | ||||||
| 953 | assert(Ty->isIntegerTy() && "Expected an integer step")(static_cast <bool> (Ty->isIntegerTy() && "Expected an integer step" ) ? void (0) : __assert_fail ("Ty->isIntegerTy() && \"Expected an integer step\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 953, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 954 | Constant *StepVal = ConstantInt::get(Ty, Step * VF.getKnownMinValue()); | ||||||
| 955 | return VF.isScalable() ? B.CreateVScale(StepVal) : StepVal; | ||||||
| 956 | } | ||||||
| 957 | |||||||
| 958 | /// Return the runtime value for VF. | ||||||
| 959 | Value *getRuntimeVF(IRBuilderBase &B, Type *Ty, ElementCount VF) { | ||||||
| 960 | Constant *EC = ConstantInt::get(Ty, VF.getKnownMinValue()); | ||||||
| 961 | return VF.isScalable() ? B.CreateVScale(EC) : EC; | ||||||
| 962 | } | ||||||
| 963 | |||||||
| 964 | const SCEV *createTripCountSCEV(Type *IdxTy, PredicatedScalarEvolution &PSE) { | ||||||
| 965 | const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount(); | ||||||
| 966 | assert(!isa<SCEVCouldNotCompute>(BackedgeTakenCount) && "Invalid loop count")(static_cast <bool> (!isa<SCEVCouldNotCompute>(BackedgeTakenCount ) && "Invalid loop count") ? void (0) : __assert_fail ("!isa<SCEVCouldNotCompute>(BackedgeTakenCount) && \"Invalid loop count\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 966, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 967 | |||||||
| 968 | ScalarEvolution &SE = *PSE.getSE(); | ||||||
| 969 | |||||||
| 970 | // The exit count might have the type of i64 while the phi is i32. This can | ||||||
| 971 | // happen if we have an induction variable that is sign extended before the | ||||||
| 972 | // compare. The only way that we get a backedge taken count is that the | ||||||
| 973 | // induction variable was signed and as such will not overflow. In such a case | ||||||
| 974 | // truncation is legal. | ||||||
| 975 | if (SE.getTypeSizeInBits(BackedgeTakenCount->getType()) > | ||||||
| 976 | IdxTy->getPrimitiveSizeInBits()) | ||||||
| 977 | BackedgeTakenCount = SE.getTruncateOrNoop(BackedgeTakenCount, IdxTy); | ||||||
| 978 | BackedgeTakenCount = SE.getNoopOrZeroExtend(BackedgeTakenCount, IdxTy); | ||||||
| 979 | |||||||
| 980 | // Get the total trip count from the count by adding 1. | ||||||
| 981 | return SE.getAddExpr(BackedgeTakenCount, | ||||||
| 982 | SE.getOne(BackedgeTakenCount->getType())); | ||||||
| 983 | } | ||||||
| 984 | |||||||
| 985 | static Value *getRuntimeVFAsFloat(IRBuilderBase &B, Type *FTy, | ||||||
| 986 | ElementCount VF) { | ||||||
| 987 | assert(FTy->isFloatingPointTy() && "Expected floating point type!")(static_cast <bool> (FTy->isFloatingPointTy() && "Expected floating point type!") ? void (0) : __assert_fail ( "FTy->isFloatingPointTy() && \"Expected floating point type!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 987, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 988 | Type *IntTy = IntegerType::get(FTy->getContext(), FTy->getScalarSizeInBits()); | ||||||
| 989 | Value *RuntimeVF = getRuntimeVF(B, IntTy, VF); | ||||||
| 990 | return B.CreateUIToFP(RuntimeVF, FTy); | ||||||
| 991 | } | ||||||
| 992 | |||||||
| 993 | void reportVectorizationFailure(const StringRef DebugMsg, | ||||||
| 994 | const StringRef OREMsg, const StringRef ORETag, | ||||||
| 995 | OptimizationRemarkEmitter *ORE, Loop *TheLoop, | ||||||
| 996 | Instruction *I) { | ||||||
| 997 | LLVM_DEBUG(debugVectorizationMessage("Not vectorizing: ", DebugMsg, I))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { debugVectorizationMessage("Not vectorizing: " , DebugMsg, I); } } while (false); | ||||||
| 998 | LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE); | ||||||
| 999 | ORE->emit( | ||||||
| 1000 | createLVAnalysis(Hints.vectorizeAnalysisPassName(), ORETag, TheLoop, I) | ||||||
| 1001 | << "loop not vectorized: " << OREMsg); | ||||||
| 1002 | } | ||||||
| 1003 | |||||||
| 1004 | void reportVectorizationInfo(const StringRef Msg, const StringRef ORETag, | ||||||
| 1005 | OptimizationRemarkEmitter *ORE, Loop *TheLoop, | ||||||
| 1006 | Instruction *I) { | ||||||
| 1007 | LLVM_DEBUG(debugVectorizationMessage("", Msg, I))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { debugVectorizationMessage("", Msg, I); } } while (false); | ||||||
| 1008 | LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE); | ||||||
| 1009 | ORE->emit( | ||||||
| 1010 | createLVAnalysis(Hints.vectorizeAnalysisPassName(), ORETag, TheLoop, I) | ||||||
| 1011 | << Msg); | ||||||
| 1012 | } | ||||||
| 1013 | |||||||
| 1014 | } // end namespace llvm | ||||||
| 1015 | |||||||
| 1016 | #ifndef NDEBUG | ||||||
| 1017 | /// \return string containing a file name and a line # for the given loop. | ||||||
| 1018 | static std::string getDebugLocString(const Loop *L) { | ||||||
| 1019 | std::string Result; | ||||||
| 1020 | if (L) { | ||||||
| 1021 | raw_string_ostream OS(Result); | ||||||
| 1022 | if (const DebugLoc LoopDbgLoc = L->getStartLoc()) | ||||||
| 1023 | LoopDbgLoc.print(OS); | ||||||
| 1024 | else | ||||||
| 1025 | // Just print the module name. | ||||||
| 1026 | OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier(); | ||||||
| 1027 | OS.flush(); | ||||||
| 1028 | } | ||||||
| 1029 | return Result; | ||||||
| 1030 | } | ||||||
| 1031 | #endif | ||||||
| 1032 | |||||||
| 1033 | void InnerLoopVectorizer::collectPoisonGeneratingRecipes( | ||||||
| 1034 | VPTransformState &State) { | ||||||
| 1035 | |||||||
| 1036 | // Collect recipes in the backward slice of `Root` that may generate a poison | ||||||
| 1037 | // value that is used after vectorization. | ||||||
| 1038 | SmallPtrSet<VPRecipeBase *, 16> Visited; | ||||||
| 1039 | auto collectPoisonGeneratingInstrsInBackwardSlice([&](VPRecipeBase *Root) { | ||||||
| 1040 | SmallVector<VPRecipeBase *, 16> Worklist; | ||||||
| 1041 | Worklist.push_back(Root); | ||||||
| 1042 | |||||||
| 1043 | // Traverse the backward slice of Root through its use-def chain. | ||||||
| 1044 | while (!Worklist.empty()) { | ||||||
| 1045 | VPRecipeBase *CurRec = Worklist.back(); | ||||||
| 1046 | Worklist.pop_back(); | ||||||
| 1047 | |||||||
| 1048 | if (!Visited.insert(CurRec).second) | ||||||
| 1049 | continue; | ||||||
| 1050 | |||||||
| 1051 | // Prune search if we find another recipe generating a widen memory | ||||||
| 1052 | // instruction. Widen memory instructions involved in address computation | ||||||
| 1053 | // will lead to gather/scatter instructions, which don't need to be | ||||||
| 1054 | // handled. | ||||||
| 1055 | if (isa<VPWidenMemoryInstructionRecipe>(CurRec) || | ||||||
| 1056 | isa<VPInterleaveRecipe>(CurRec) || | ||||||
| 1057 | isa<VPScalarIVStepsRecipe>(CurRec) || | ||||||
| 1058 | isa<VPCanonicalIVPHIRecipe>(CurRec) || | ||||||
| 1059 | isa<VPActiveLaneMaskPHIRecipe>(CurRec)) | ||||||
| 1060 | continue; | ||||||
| 1061 | |||||||
| 1062 | // This recipe contributes to the address computation of a widen | ||||||
| 1063 | // load/store. Collect recipe if its underlying instruction has | ||||||
| 1064 | // poison-generating flags. | ||||||
| 1065 | Instruction *Instr = CurRec->getUnderlyingInstr(); | ||||||
| 1066 | if (Instr && Instr->hasPoisonGeneratingFlags()) | ||||||
| 1067 | State.MayGeneratePoisonRecipes.insert(CurRec); | ||||||
| 1068 | |||||||
| 1069 | // Add new definitions to the worklist. | ||||||
| 1070 | for (VPValue *operand : CurRec->operands()) | ||||||
| 1071 | if (VPRecipeBase *OpDef = operand->getDefiningRecipe()) | ||||||
| 1072 | Worklist.push_back(OpDef); | ||||||
| 1073 | } | ||||||
| 1074 | }); | ||||||
| 1075 | |||||||
| 1076 | // Traverse all the recipes in the VPlan and collect the poison-generating | ||||||
| 1077 | // recipes in the backward slice starting at the address of a VPWidenRecipe or | ||||||
| 1078 | // VPInterleaveRecipe. | ||||||
| 1079 | auto Iter = depth_first( | ||||||
| 1080 | VPBlockRecursiveTraversalWrapper<VPBlockBase *>(State.Plan->getEntry())); | ||||||
| 1081 | for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Iter)) { | ||||||
| 1082 | for (VPRecipeBase &Recipe : *VPBB) { | ||||||
| 1083 | if (auto *WidenRec = dyn_cast<VPWidenMemoryInstructionRecipe>(&Recipe)) { | ||||||
| 1084 | Instruction &UnderlyingInstr = WidenRec->getIngredient(); | ||||||
| 1085 | VPRecipeBase *AddrDef = WidenRec->getAddr()->getDefiningRecipe(); | ||||||
| 1086 | if (AddrDef && WidenRec->isConsecutive() && | ||||||
| 1087 | Legal->blockNeedsPredication(UnderlyingInstr.getParent())) | ||||||
| 1088 | collectPoisonGeneratingInstrsInBackwardSlice(AddrDef); | ||||||
| 1089 | } else if (auto *InterleaveRec = dyn_cast<VPInterleaveRecipe>(&Recipe)) { | ||||||
| 1090 | VPRecipeBase *AddrDef = InterleaveRec->getAddr()->getDefiningRecipe(); | ||||||
| 1091 | if (AddrDef) { | ||||||
| 1092 | // Check if any member of the interleave group needs predication. | ||||||
| 1093 | const InterleaveGroup<Instruction> *InterGroup = | ||||||
| 1094 | InterleaveRec->getInterleaveGroup(); | ||||||
| 1095 | bool NeedPredication = false; | ||||||
| 1096 | for (int I = 0, NumMembers = InterGroup->getNumMembers(); | ||||||
| 1097 | I < NumMembers; ++I) { | ||||||
| 1098 | Instruction *Member = InterGroup->getMember(I); | ||||||
| 1099 | if (Member) | ||||||
| 1100 | NeedPredication |= | ||||||
| 1101 | Legal->blockNeedsPredication(Member->getParent()); | ||||||
| 1102 | } | ||||||
| 1103 | |||||||
| 1104 | if (NeedPredication) | ||||||
| 1105 | collectPoisonGeneratingInstrsInBackwardSlice(AddrDef); | ||||||
| 1106 | } | ||||||
| 1107 | } | ||||||
| 1108 | } | ||||||
| 1109 | } | ||||||
| 1110 | } | ||||||
| 1111 | |||||||
| 1112 | PHINode *InnerLoopVectorizer::getReductionResumeValue( | ||||||
| 1113 | const RecurrenceDescriptor &RdxDesc) { | ||||||
| 1114 | auto It = ReductionResumeValues.find(&RdxDesc); | ||||||
| 1115 | assert(It != ReductionResumeValues.end() &&(static_cast <bool> (It != ReductionResumeValues.end() && "Expected to find a resume value for the reduction.") ? void (0) : __assert_fail ("It != ReductionResumeValues.end() && \"Expected to find a resume value for the reduction.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1116, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 1116 | "Expected to find a resume value for the reduction.")(static_cast <bool> (It != ReductionResumeValues.end() && "Expected to find a resume value for the reduction.") ? void (0) : __assert_fail ("It != ReductionResumeValues.end() && \"Expected to find a resume value for the reduction.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1116, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 1117 | return It->second; | ||||||
| 1118 | } | ||||||
| 1119 | |||||||
| 1120 | namespace llvm { | ||||||
| 1121 | |||||||
| 1122 | // Loop vectorization cost-model hints how the scalar epilogue loop should be | ||||||
| 1123 | // lowered. | ||||||
| 1124 | enum ScalarEpilogueLowering { | ||||||
| 1125 | |||||||
| 1126 | // The default: allowing scalar epilogues. | ||||||
| 1127 | CM_ScalarEpilogueAllowed, | ||||||
| 1128 | |||||||
| 1129 | // Vectorization with OptForSize: don't allow epilogues. | ||||||
| 1130 | CM_ScalarEpilogueNotAllowedOptSize, | ||||||
| 1131 | |||||||
| 1132 | // A special case of vectorisation with OptForSize: loops with a very small | ||||||
| 1133 | // trip count are considered for vectorization under OptForSize, thereby | ||||||
| 1134 | // making sure the cost of their loop body is dominant, free of runtime | ||||||
| 1135 | // guards and scalar iteration overheads. | ||||||
| 1136 | CM_ScalarEpilogueNotAllowedLowTripLoop, | ||||||
| 1137 | |||||||
| 1138 | // Loop hint predicate indicating an epilogue is undesired. | ||||||
| 1139 | CM_ScalarEpilogueNotNeededUsePredicate, | ||||||
| 1140 | |||||||
| 1141 | // Directive indicating we must either tail fold or not vectorize | ||||||
| 1142 | CM_ScalarEpilogueNotAllowedUsePredicate | ||||||
| 1143 | }; | ||||||
| 1144 | |||||||
| 1145 | /// ElementCountComparator creates a total ordering for ElementCount | ||||||
| 1146 | /// for the purposes of using it in a set structure. | ||||||
| 1147 | struct ElementCountComparator { | ||||||
| 1148 | bool operator()(const ElementCount &LHS, const ElementCount &RHS) const { | ||||||
| 1149 | return std::make_tuple(LHS.isScalable(), LHS.getKnownMinValue()) < | ||||||
| 1150 | std::make_tuple(RHS.isScalable(), RHS.getKnownMinValue()); | ||||||
| 1151 | } | ||||||
| 1152 | }; | ||||||
| 1153 | using ElementCountSet = SmallSet<ElementCount, 16, ElementCountComparator>; | ||||||
| 1154 | |||||||
| 1155 | /// LoopVectorizationCostModel - estimates the expected speedups due to | ||||||
| 1156 | /// vectorization. | ||||||
| 1157 | /// In many cases vectorization is not profitable. This can happen because of | ||||||
| 1158 | /// a number of reasons. In this class we mainly attempt to predict the | ||||||
| 1159 | /// expected speedup/slowdowns due to the supported instruction set. We use the | ||||||
| 1160 | /// TargetTransformInfo to query the different backends for the cost of | ||||||
| 1161 | /// different operations. | ||||||
| 1162 | class LoopVectorizationCostModel { | ||||||
| 1163 | public: | ||||||
| 1164 | LoopVectorizationCostModel(ScalarEpilogueLowering SEL, Loop *L, | ||||||
| 1165 | PredicatedScalarEvolution &PSE, LoopInfo *LI, | ||||||
| 1166 | LoopVectorizationLegality *Legal, | ||||||
| 1167 | const TargetTransformInfo &TTI, | ||||||
| 1168 | const TargetLibraryInfo *TLI, DemandedBits *DB, | ||||||
| 1169 | AssumptionCache *AC, | ||||||
| 1170 | OptimizationRemarkEmitter *ORE, const Function *F, | ||||||
| 1171 | const LoopVectorizeHints *Hints, | ||||||
| 1172 | InterleavedAccessInfo &IAI) | ||||||
| 1173 | : ScalarEpilogueStatus(SEL), TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), | ||||||
| 1174 | TTI(TTI), TLI(TLI), DB(DB), AC(AC), ORE(ORE), TheFunction(F), | ||||||
| 1175 | Hints(Hints), InterleaveInfo(IAI) {} | ||||||
| 1176 | |||||||
| 1177 | /// \return An upper bound for the vectorization factors (both fixed and | ||||||
| 1178 | /// scalable). If the factors are 0, vectorization and interleaving should be | ||||||
| 1179 | /// avoided up front. | ||||||
| 1180 | FixedScalableVFPair computeMaxVF(ElementCount UserVF, unsigned UserIC); | ||||||
| 1181 | |||||||
| 1182 | /// \return True if runtime checks are required for vectorization, and false | ||||||
| 1183 | /// otherwise. | ||||||
| 1184 | bool runtimeChecksRequired(); | ||||||
| 1185 | |||||||
| 1186 | /// \return The most profitable vectorization factor and the cost of that VF. | ||||||
| 1187 | /// This method checks every VF in \p CandidateVFs. If UserVF is not ZERO | ||||||
| 1188 | /// then this vectorization factor will be selected if vectorization is | ||||||
| 1189 | /// possible. | ||||||
| 1190 | VectorizationFactor | ||||||
| 1191 | selectVectorizationFactor(const ElementCountSet &CandidateVFs); | ||||||
| 1192 | |||||||
| 1193 | VectorizationFactor | ||||||
| 1194 | selectEpilogueVectorizationFactor(const ElementCount MaxVF, | ||||||
| 1195 | const LoopVectorizationPlanner &LVP); | ||||||
| 1196 | |||||||
| 1197 | /// Setup cost-based decisions for user vectorization factor. | ||||||
| 1198 | /// \return true if the UserVF is a feasible VF to be chosen. | ||||||
| 1199 | bool selectUserVectorizationFactor(ElementCount UserVF) { | ||||||
| 1200 | collectUniformsAndScalars(UserVF); | ||||||
| 1201 | collectInstsToScalarize(UserVF); | ||||||
| 1202 | return expectedCost(UserVF).first.isValid(); | ||||||
| 1203 | } | ||||||
| 1204 | |||||||
| 1205 | /// \return The size (in bits) of the smallest and widest types in the code | ||||||
| 1206 | /// that needs to be vectorized. We ignore values that remain scalar such as | ||||||
| 1207 | /// 64 bit loop indices. | ||||||
| 1208 | std::pair<unsigned, unsigned> getSmallestAndWidestTypes(); | ||||||
| 1209 | |||||||
| 1210 | /// \return The desired interleave count. | ||||||
| 1211 | /// If interleave count has been specified by metadata it will be returned. | ||||||
| 1212 | /// Otherwise, the interleave count is computed and returned. VF and LoopCost | ||||||
| 1213 | /// are the selected vectorization factor and the cost of the selected VF. | ||||||
| 1214 | unsigned selectInterleaveCount(ElementCount VF, InstructionCost LoopCost); | ||||||
| 1215 | |||||||
| 1216 | /// Memory access instruction may be vectorized in more than one way. | ||||||
| 1217 | /// Form of instruction after vectorization depends on cost. | ||||||
| 1218 | /// This function takes cost-based decisions for Load/Store instructions | ||||||
| 1219 | /// and collects them in a map. This decisions map is used for building | ||||||
| 1220 | /// the lists of loop-uniform and loop-scalar instructions. | ||||||
| 1221 | /// The calculated cost is saved with widening decision in order to | ||||||
| 1222 | /// avoid redundant calculations. | ||||||
| 1223 | void setCostBasedWideningDecision(ElementCount VF); | ||||||
| 1224 | |||||||
| 1225 | /// A struct that represents some properties of the register usage | ||||||
| 1226 | /// of a loop. | ||||||
| 1227 | struct RegisterUsage { | ||||||
| 1228 | /// Holds the number of loop invariant values that are used in the loop. | ||||||
| 1229 | /// The key is ClassID of target-provided register class. | ||||||
| 1230 | SmallMapVector<unsigned, unsigned, 4> LoopInvariantRegs; | ||||||
| 1231 | /// Holds the maximum number of concurrent live intervals in the loop. | ||||||
| 1232 | /// The key is ClassID of target-provided register class. | ||||||
| 1233 | SmallMapVector<unsigned, unsigned, 4> MaxLocalUsers; | ||||||
| 1234 | }; | ||||||
| 1235 | |||||||
| 1236 | /// \return Returns information about the register usages of the loop for the | ||||||
| 1237 | /// given vectorization factors. | ||||||
| 1238 | SmallVector<RegisterUsage, 8> | ||||||
| 1239 | calculateRegisterUsage(ArrayRef<ElementCount> VFs); | ||||||
| 1240 | |||||||
| 1241 | /// Collect values we want to ignore in the cost model. | ||||||
| 1242 | void collectValuesToIgnore(); | ||||||
| 1243 | |||||||
| 1244 | /// Collect all element types in the loop for which widening is needed. | ||||||
| 1245 | void collectElementTypesForWidening(); | ||||||
| 1246 | |||||||
| 1247 | /// Split reductions into those that happen in the loop, and those that happen | ||||||
| 1248 | /// outside. In loop reductions are collected into InLoopReductionChains. | ||||||
| 1249 | void collectInLoopReductions(); | ||||||
| 1250 | |||||||
| 1251 | /// Returns true if we should use strict in-order reductions for the given | ||||||
| 1252 | /// RdxDesc. This is true if the -enable-strict-reductions flag is passed, | ||||||
| 1253 | /// the IsOrdered flag of RdxDesc is set and we do not allow reordering | ||||||
| 1254 | /// of FP operations. | ||||||
| 1255 | bool useOrderedReductions(const RecurrenceDescriptor &RdxDesc) const { | ||||||
| 1256 | return !Hints->allowReordering() && RdxDesc.isOrdered(); | ||||||
| 1257 | } | ||||||
| 1258 | |||||||
| 1259 | /// \returns The smallest bitwidth each instruction can be represented with. | ||||||
| 1260 | /// The vector equivalents of these instructions should be truncated to this | ||||||
| 1261 | /// type. | ||||||
| 1262 | const MapVector<Instruction *, uint64_t> &getMinimalBitwidths() const { | ||||||
| 1263 | return MinBWs; | ||||||
| 1264 | } | ||||||
| 1265 | |||||||
| 1266 | /// \returns True if it is more profitable to scalarize instruction \p I for | ||||||
| 1267 | /// vectorization factor \p VF. | ||||||
| 1268 | bool isProfitableToScalarize(Instruction *I, ElementCount VF) const { | ||||||
| 1269 | assert(VF.isVector() &&(static_cast <bool> (VF.isVector() && "Profitable to scalarize relevant only for VF > 1." ) ? void (0) : __assert_fail ("VF.isVector() && \"Profitable to scalarize relevant only for VF > 1.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1270, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 1270 | "Profitable to scalarize relevant only for VF > 1.")(static_cast <bool> (VF.isVector() && "Profitable to scalarize relevant only for VF > 1." ) ? void (0) : __assert_fail ("VF.isVector() && \"Profitable to scalarize relevant only for VF > 1.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1270, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 1271 | |||||||
| 1272 | // Cost model is not run in the VPlan-native path - return conservative | ||||||
| 1273 | // result until this changes. | ||||||
| 1274 | if (EnableVPlanNativePath) | ||||||
| 1275 | return false; | ||||||
| 1276 | |||||||
| 1277 | auto Scalars = InstsToScalarize.find(VF); | ||||||
| 1278 | assert(Scalars != InstsToScalarize.end() &&(static_cast <bool> (Scalars != InstsToScalarize.end() && "VF not yet analyzed for scalarization profitability") ? void (0) : __assert_fail ("Scalars != InstsToScalarize.end() && \"VF not yet analyzed for scalarization profitability\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1279, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 1279 | "VF not yet analyzed for scalarization profitability")(static_cast <bool> (Scalars != InstsToScalarize.end() && "VF not yet analyzed for scalarization profitability") ? void (0) : __assert_fail ("Scalars != InstsToScalarize.end() && \"VF not yet analyzed for scalarization profitability\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1279, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 1280 | return Scalars->second.find(I) != Scalars->second.end(); | ||||||
| 1281 | } | ||||||
| 1282 | |||||||
| 1283 | /// Returns true if \p I is known to be uniform after vectorization. | ||||||
| 1284 | bool isUniformAfterVectorization(Instruction *I, ElementCount VF) const { | ||||||
| 1285 | if (VF.isScalar()) | ||||||
| 1286 | return true; | ||||||
| 1287 | |||||||
| 1288 | // Cost model is not run in the VPlan-native path - return conservative | ||||||
| 1289 | // result until this changes. | ||||||
| 1290 | if (EnableVPlanNativePath) | ||||||
| 1291 | return false; | ||||||
| 1292 | |||||||
| 1293 | auto UniformsPerVF = Uniforms.find(VF); | ||||||
| 1294 | assert(UniformsPerVF != Uniforms.end() &&(static_cast <bool> (UniformsPerVF != Uniforms.end() && "VF not yet analyzed for uniformity") ? void (0) : __assert_fail ("UniformsPerVF != Uniforms.end() && \"VF not yet analyzed for uniformity\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1295, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 1295 | "VF not yet analyzed for uniformity")(static_cast <bool> (UniformsPerVF != Uniforms.end() && "VF not yet analyzed for uniformity") ? void (0) : __assert_fail ("UniformsPerVF != Uniforms.end() && \"VF not yet analyzed for uniformity\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1295, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 1296 | return UniformsPerVF->second.count(I); | ||||||
| 1297 | } | ||||||
| 1298 | |||||||
| 1299 | /// Returns true if \p I is known to be scalar after vectorization. | ||||||
| 1300 | bool isScalarAfterVectorization(Instruction *I, ElementCount VF) const { | ||||||
| 1301 | if (VF.isScalar()) | ||||||
| 1302 | return true; | ||||||
| 1303 | |||||||
| 1304 | // Cost model is not run in the VPlan-native path - return conservative | ||||||
| 1305 | // result until this changes. | ||||||
| 1306 | if (EnableVPlanNativePath) | ||||||
| 1307 | return false; | ||||||
| 1308 | |||||||
| 1309 | auto ScalarsPerVF = Scalars.find(VF); | ||||||
| 1310 | assert(ScalarsPerVF != Scalars.end() &&(static_cast <bool> (ScalarsPerVF != Scalars.end() && "Scalar values are not calculated for VF") ? void (0) : __assert_fail ("ScalarsPerVF != Scalars.end() && \"Scalar values are not calculated for VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1311, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 1311 | "Scalar values are not calculated for VF")(static_cast <bool> (ScalarsPerVF != Scalars.end() && "Scalar values are not calculated for VF") ? void (0) : __assert_fail ("ScalarsPerVF != Scalars.end() && \"Scalar values are not calculated for VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1311, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 1312 | return ScalarsPerVF->second.count(I); | ||||||
| 1313 | } | ||||||
| 1314 | |||||||
| 1315 | /// \returns True if instruction \p I can be truncated to a smaller bitwidth | ||||||
| 1316 | /// for vectorization factor \p VF. | ||||||
| 1317 | bool canTruncateToMinimalBitwidth(Instruction *I, ElementCount VF) const { | ||||||
| 1318 | return VF.isVector() && MinBWs.find(I) != MinBWs.end() && | ||||||
| 1319 | !isProfitableToScalarize(I, VF) && | ||||||
| 1320 | !isScalarAfterVectorization(I, VF); | ||||||
| 1321 | } | ||||||
| 1322 | |||||||
| 1323 | /// Decision that was taken during cost calculation for memory instruction. | ||||||
| 1324 | enum InstWidening { | ||||||
| 1325 | CM_Unknown, | ||||||
| 1326 | CM_Widen, // For consecutive accesses with stride +1. | ||||||
| 1327 | CM_Widen_Reverse, // For consecutive accesses with stride -1. | ||||||
| 1328 | CM_Interleave, | ||||||
| 1329 | CM_GatherScatter, | ||||||
| 1330 | CM_Scalarize | ||||||
| 1331 | }; | ||||||
| 1332 | |||||||
| 1333 | /// Save vectorization decision \p W and \p Cost taken by the cost model for | ||||||
| 1334 | /// instruction \p I and vector width \p VF. | ||||||
| 1335 | void setWideningDecision(Instruction *I, ElementCount VF, InstWidening W, | ||||||
| 1336 | InstructionCost Cost) { | ||||||
| 1337 | assert(VF.isVector() && "Expected VF >=2")(static_cast <bool> (VF.isVector() && "Expected VF >=2" ) ? void (0) : __assert_fail ("VF.isVector() && \"Expected VF >=2\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1337, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 1338 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost); | ||||||
| 1339 | } | ||||||
| 1340 | |||||||
| 1341 | /// Save vectorization decision \p W and \p Cost taken by the cost model for | ||||||
| 1342 | /// interleaving group \p Grp and vector width \p VF. | ||||||
| 1343 | void setWideningDecision(const InterleaveGroup<Instruction> *Grp, | ||||||
| 1344 | ElementCount VF, InstWidening W, | ||||||
| 1345 | InstructionCost Cost) { | ||||||
| 1346 | assert(VF.isVector() && "Expected VF >=2")(static_cast <bool> (VF.isVector() && "Expected VF >=2" ) ? void (0) : __assert_fail ("VF.isVector() && \"Expected VF >=2\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1346, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 1347 | /// Broadcast this decicion to all instructions inside the group. | ||||||
| 1348 | /// But the cost will be assigned to one instruction only. | ||||||
| 1349 | for (unsigned i = 0; i < Grp->getFactor(); ++i) { | ||||||
| 1350 | if (auto *I = Grp->getMember(i)) { | ||||||
| 1351 | if (Grp->getInsertPos() == I) | ||||||
| 1352 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost); | ||||||
| 1353 | else | ||||||
| 1354 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, 0); | ||||||
| 1355 | } | ||||||
| 1356 | } | ||||||
| 1357 | } | ||||||
| 1358 | |||||||
| 1359 | /// Return the cost model decision for the given instruction \p I and vector | ||||||
| 1360 | /// width \p VF. Return CM_Unknown if this instruction did not pass | ||||||
| 1361 | /// through the cost modeling. | ||||||
| 1362 | InstWidening getWideningDecision(Instruction *I, ElementCount VF) const { | ||||||
| 1363 | assert(VF.isVector() && "Expected VF to be a vector VF")(static_cast <bool> (VF.isVector() && "Expected VF to be a vector VF" ) ? void (0) : __assert_fail ("VF.isVector() && \"Expected VF to be a vector VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1363, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 1364 | // Cost model is not run in the VPlan-native path - return conservative | ||||||
| 1365 | // result until this changes. | ||||||
| 1366 | if (EnableVPlanNativePath) | ||||||
| 1367 | return CM_GatherScatter; | ||||||
| 1368 | |||||||
| 1369 | std::pair<Instruction *, ElementCount> InstOnVF = std::make_pair(I, VF); | ||||||
| 1370 | auto Itr = WideningDecisions.find(InstOnVF); | ||||||
| 1371 | if (Itr == WideningDecisions.end()) | ||||||
| 1372 | return CM_Unknown; | ||||||
| 1373 | return Itr->second.first; | ||||||
| 1374 | } | ||||||
| 1375 | |||||||
| 1376 | /// Return the vectorization cost for the given instruction \p I and vector | ||||||
| 1377 | /// width \p VF. | ||||||
| 1378 | InstructionCost getWideningCost(Instruction *I, ElementCount VF) { | ||||||
| 1379 | assert(VF.isVector() && "Expected VF >=2")(static_cast <bool> (VF.isVector() && "Expected VF >=2" ) ? void (0) : __assert_fail ("VF.isVector() && \"Expected VF >=2\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1379, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 1380 | std::pair<Instruction *, ElementCount> InstOnVF = std::make_pair(I, VF); | ||||||
| 1381 | assert(WideningDecisions.find(InstOnVF) != WideningDecisions.end() &&(static_cast <bool> (WideningDecisions.find(InstOnVF) != WideningDecisions.end() && "The cost is not calculated" ) ? void (0) : __assert_fail ("WideningDecisions.find(InstOnVF) != WideningDecisions.end() && \"The cost is not calculated\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1382, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 1382 | "The cost is not calculated")(static_cast <bool> (WideningDecisions.find(InstOnVF) != WideningDecisions.end() && "The cost is not calculated" ) ? void (0) : __assert_fail ("WideningDecisions.find(InstOnVF) != WideningDecisions.end() && \"The cost is not calculated\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1382, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 1383 | return WideningDecisions[InstOnVF].second; | ||||||
| 1384 | } | ||||||
| 1385 | |||||||
| 1386 | /// Return True if instruction \p I is an optimizable truncate whose operand | ||||||
| 1387 | /// is an induction variable. Such a truncate will be removed by adding a new | ||||||
| 1388 | /// induction variable with the destination type. | ||||||
| 1389 | bool isOptimizableIVTruncate(Instruction *I, ElementCount VF) { | ||||||
| 1390 | // If the instruction is not a truncate, return false. | ||||||
| 1391 | auto *Trunc = dyn_cast<TruncInst>(I); | ||||||
| 1392 | if (!Trunc) | ||||||
| 1393 | return false; | ||||||
| 1394 | |||||||
| 1395 | // Get the source and destination types of the truncate. | ||||||
| 1396 | Type *SrcTy = ToVectorTy(cast<CastInst>(I)->getSrcTy(), VF); | ||||||
| 1397 | Type *DestTy = ToVectorTy(cast<CastInst>(I)->getDestTy(), VF); | ||||||
| 1398 | |||||||
| 1399 | // If the truncate is free for the given types, return false. Replacing a | ||||||
| 1400 | // free truncate with an induction variable would add an induction variable | ||||||
| 1401 | // update instruction to each iteration of the loop. We exclude from this | ||||||
| 1402 | // check the primary induction variable since it will need an update | ||||||
| 1403 | // instruction regardless. | ||||||
| 1404 | Value *Op = Trunc->getOperand(0); | ||||||
| 1405 | if (Op != Legal->getPrimaryInduction() && TTI.isTruncateFree(SrcTy, DestTy)) | ||||||
| 1406 | return false; | ||||||
| 1407 | |||||||
| 1408 | // If the truncated value is not an induction variable, return false. | ||||||
| 1409 | return Legal->isInductionPhi(Op); | ||||||
| 1410 | } | ||||||
| 1411 | |||||||
| 1412 | /// Collects the instructions to scalarize for each predicated instruction in | ||||||
| 1413 | /// the loop. | ||||||
| 1414 | void collectInstsToScalarize(ElementCount VF); | ||||||
| 1415 | |||||||
| 1416 | /// Collect Uniform and Scalar values for the given \p VF. | ||||||
| 1417 | /// The sets depend on CM decision for Load/Store instructions | ||||||
| 1418 | /// that may be vectorized as interleave, gather-scatter or scalarized. | ||||||
| 1419 | void collectUniformsAndScalars(ElementCount VF) { | ||||||
| 1420 | // Do the analysis once. | ||||||
| 1421 | if (VF.isScalar() || Uniforms.find(VF) != Uniforms.end()) | ||||||
| 1422 | return; | ||||||
| 1423 | setCostBasedWideningDecision(VF); | ||||||
| 1424 | collectLoopUniforms(VF); | ||||||
| 1425 | collectLoopScalars(VF); | ||||||
| 1426 | } | ||||||
| 1427 | |||||||
| 1428 | /// Returns true if the target machine supports masked store operation | ||||||
| 1429 | /// for the given \p DataType and kind of access to \p Ptr. | ||||||
| 1430 | bool isLegalMaskedStore(Type *DataType, Value *Ptr, Align Alignment) const { | ||||||
| 1431 | return Legal->isConsecutivePtr(DataType, Ptr) && | ||||||
| 1432 | TTI.isLegalMaskedStore(DataType, Alignment); | ||||||
| 1433 | } | ||||||
| 1434 | |||||||
| 1435 | /// Returns true if the target machine supports masked load operation | ||||||
| 1436 | /// for the given \p DataType and kind of access to \p Ptr. | ||||||
| 1437 | bool isLegalMaskedLoad(Type *DataType, Value *Ptr, Align Alignment) const { | ||||||
| 1438 | return Legal->isConsecutivePtr(DataType, Ptr) && | ||||||
| 1439 | TTI.isLegalMaskedLoad(DataType, Alignment); | ||||||
| 1440 | } | ||||||
| 1441 | |||||||
| 1442 | /// Returns true if the target machine can represent \p V as a masked gather | ||||||
| 1443 | /// or scatter operation. | ||||||
| 1444 | bool isLegalGatherOrScatter(Value *V, | ||||||
| 1445 | ElementCount VF = ElementCount::getFixed(1)) { | ||||||
| 1446 | bool LI = isa<LoadInst>(V); | ||||||
| 1447 | bool SI = isa<StoreInst>(V); | ||||||
| 1448 | if (!LI && !SI) | ||||||
| 1449 | return false; | ||||||
| 1450 | auto *Ty = getLoadStoreType(V); | ||||||
| 1451 | Align Align = getLoadStoreAlignment(V); | ||||||
| 1452 | if (VF.isVector()) | ||||||
| 1453 | Ty = VectorType::get(Ty, VF); | ||||||
| 1454 | return (LI && TTI.isLegalMaskedGather(Ty, Align)) || | ||||||
| 1455 | (SI && TTI.isLegalMaskedScatter(Ty, Align)); | ||||||
| 1456 | } | ||||||
| 1457 | |||||||
| 1458 | /// Returns true if the target machine supports all of the reduction | ||||||
| 1459 | /// variables found for the given VF. | ||||||
| 1460 | bool canVectorizeReductions(ElementCount VF) const { | ||||||
| 1461 | return (all_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool { | ||||||
| 1462 | const RecurrenceDescriptor &RdxDesc = Reduction.second; | ||||||
| 1463 | return TTI.isLegalToVectorizeReduction(RdxDesc, VF); | ||||||
| 1464 | })); | ||||||
| 1465 | } | ||||||
| 1466 | |||||||
| 1467 | /// Given costs for both strategies, return true if the scalar predication | ||||||
| 1468 | /// lowering should be used for div/rem. This incorporates an override | ||||||
| 1469 | /// option so it is not simply a cost comparison. | ||||||
| 1470 | bool isDivRemScalarWithPredication(InstructionCost ScalarCost, | ||||||
| 1471 | InstructionCost SafeDivisorCost) const { | ||||||
| 1472 | switch (ForceSafeDivisor) { | ||||||
| 1473 | case cl::BOU_UNSET: | ||||||
| 1474 | return ScalarCost < SafeDivisorCost; | ||||||
| 1475 | case cl::BOU_TRUE: | ||||||
| 1476 | return false; | ||||||
| 1477 | case cl::BOU_FALSE: | ||||||
| 1478 | return true; | ||||||
| 1479 | }; | ||||||
| 1480 | llvm_unreachable("impossible case value")::llvm::llvm_unreachable_internal("impossible case value", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 1480); | ||||||
| 1481 | } | ||||||
| 1482 | |||||||
| 1483 | /// Returns true if \p I is an instruction which requires predication and | ||||||
| 1484 | /// for which our chosen predication strategy is scalarization (i.e. we | ||||||
| 1485 | /// don't have an alternate strategy such as masking available). | ||||||
| 1486 | /// \p VF is the vectorization factor that will be used to vectorize \p I. | ||||||
| 1487 | bool isScalarWithPredication(Instruction *I, ElementCount VF) const; | ||||||
| 1488 | |||||||
| 1489 | /// Returns true if \p I is an instruction that needs to be predicated | ||||||
| 1490 | /// at runtime. The result is independent of the predication mechanism. | ||||||
| 1491 | /// Superset of instructions that return true for isScalarWithPredication. | ||||||
| 1492 | bool isPredicatedInst(Instruction *I) const; | ||||||
| 1493 | |||||||
| 1494 | /// Return the costs for our two available strategies for lowering a | ||||||
| 1495 | /// div/rem operation which requires speculating at least one lane. | ||||||
| 1496 | /// First result is for scalarization (will be invalid for scalable | ||||||
| 1497 | /// vectors); second is for the safe-divisor strategy. | ||||||
| 1498 | std::pair<InstructionCost, InstructionCost> | ||||||
| 1499 | getDivRemSpeculationCost(Instruction *I, | ||||||
| 1500 | ElementCount VF) const; | ||||||
| 1501 | |||||||
| 1502 | /// Returns true if \p I is a memory instruction with consecutive memory | ||||||
| 1503 | /// access that can be widened. | ||||||
| 1504 | bool memoryInstructionCanBeWidened(Instruction *I, ElementCount VF); | ||||||
| 1505 | |||||||
| 1506 | /// Returns true if \p I is a memory instruction in an interleaved-group | ||||||
| 1507 | /// of memory accesses that can be vectorized with wide vector loads/stores | ||||||
| 1508 | /// and shuffles. | ||||||
| 1509 | bool interleavedAccessCanBeWidened(Instruction *I, ElementCount VF); | ||||||
| 1510 | |||||||
| 1511 | /// Check if \p Instr belongs to any interleaved access group. | ||||||
| 1512 | bool isAccessInterleaved(Instruction *Instr) { | ||||||
| 1513 | return InterleaveInfo.isInterleaved(Instr); | ||||||
| 1514 | } | ||||||
| 1515 | |||||||
| 1516 | /// Get the interleaved access group that \p Instr belongs to. | ||||||
| 1517 | const InterleaveGroup<Instruction> * | ||||||
| 1518 | getInterleavedAccessGroup(Instruction *Instr) { | ||||||
| 1519 | return InterleaveInfo.getInterleaveGroup(Instr); | ||||||
| 1520 | } | ||||||
| 1521 | |||||||
| 1522 | /// Returns true if we're required to use a scalar epilogue for at least | ||||||
| 1523 | /// the final iteration of the original loop. | ||||||
| 1524 | bool requiresScalarEpilogue(ElementCount VF) const { | ||||||
| 1525 | if (!isScalarEpilogueAllowed()) | ||||||
| 1526 | return false; | ||||||
| 1527 | // If we might exit from anywhere but the latch, must run the exiting | ||||||
| 1528 | // iteration in scalar form. | ||||||
| 1529 | if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) | ||||||
| 1530 | return true; | ||||||
| 1531 | return VF.isVector() && InterleaveInfo.requiresScalarEpilogue(); | ||||||
| 1532 | } | ||||||
| 1533 | |||||||
| 1534 | /// Returns true if a scalar epilogue is not allowed due to optsize or a | ||||||
| 1535 | /// loop hint annotation. | ||||||
| 1536 | bool isScalarEpilogueAllowed() const { | ||||||
| 1537 | return ScalarEpilogueStatus == CM_ScalarEpilogueAllowed; | ||||||
| 1538 | } | ||||||
| 1539 | |||||||
| 1540 | /// Returns true if all loop blocks should be masked to fold tail loop. | ||||||
| 1541 | bool foldTailByMasking() const { return FoldTailByMasking; } | ||||||
| 1542 | |||||||
| 1543 | /// Returns true if were tail-folding and want to use the active lane mask | ||||||
| 1544 | /// for vector loop control flow. | ||||||
| 1545 | bool useActiveLaneMaskForControlFlow() const { | ||||||
| 1546 | return FoldTailByMasking && | ||||||
| 1547 | TTI.emitGetActiveLaneMask() == PredicationStyle::DataAndControlFlow; | ||||||
| 1548 | } | ||||||
| 1549 | |||||||
| 1550 | /// Returns true if the instructions in this block requires predication | ||||||
| 1551 | /// for any reason, e.g. because tail folding now requires a predicate | ||||||
| 1552 | /// or because the block in the original loop was predicated. | ||||||
| 1553 | bool blockNeedsPredicationForAnyReason(BasicBlock *BB) const { | ||||||
| 1554 | return foldTailByMasking() || Legal->blockNeedsPredication(BB); | ||||||
| 1555 | } | ||||||
| 1556 | |||||||
| 1557 | /// A SmallMapVector to store the InLoop reduction op chains, mapping phi | ||||||
| 1558 | /// nodes to the chain of instructions representing the reductions. Uses a | ||||||
| 1559 | /// MapVector to ensure deterministic iteration order. | ||||||
| 1560 | using ReductionChainMap = | ||||||
| 1561 | SmallMapVector<PHINode *, SmallVector<Instruction *, 4>, 4>; | ||||||
| 1562 | |||||||
| 1563 | /// Return the chain of instructions representing an inloop reduction. | ||||||
| 1564 | const ReductionChainMap &getInLoopReductionChains() const { | ||||||
| 1565 | return InLoopReductionChains; | ||||||
| 1566 | } | ||||||
| 1567 | |||||||
| 1568 | /// Returns true if the Phi is part of an inloop reduction. | ||||||
| 1569 | bool isInLoopReduction(PHINode *Phi) const { | ||||||
| 1570 | return InLoopReductionChains.count(Phi); | ||||||
| 1571 | } | ||||||
| 1572 | |||||||
| 1573 | /// Estimate cost of an intrinsic call instruction CI if it were vectorized | ||||||
| 1574 | /// with factor VF. Return the cost of the instruction, including | ||||||
| 1575 | /// scalarization overhead if it's needed. | ||||||
| 1576 | InstructionCost getVectorIntrinsicCost(CallInst *CI, ElementCount VF) const; | ||||||
| 1577 | |||||||
| 1578 | /// Estimate cost of a call instruction CI if it were vectorized with factor | ||||||
| 1579 | /// VF. Return the cost of the instruction, including scalarization overhead | ||||||
| 1580 | /// if it's needed. The flag NeedToScalarize shows if the call needs to be | ||||||
| 1581 | /// scalarized - | ||||||
| 1582 | /// i.e. either vector version isn't available, or is too expensive. | ||||||
| 1583 | InstructionCost getVectorCallCost(CallInst *CI, ElementCount VF, | ||||||
| 1584 | bool &NeedToScalarize) const; | ||||||
| 1585 | |||||||
| 1586 | /// Returns true if the per-lane cost of VectorizationFactor A is lower than | ||||||
| 1587 | /// that of B. | ||||||
| 1588 | bool isMoreProfitable(const VectorizationFactor &A, | ||||||
| 1589 | const VectorizationFactor &B) const; | ||||||
| 1590 | |||||||
| 1591 | /// Invalidates decisions already taken by the cost model. | ||||||
| 1592 | void invalidateCostModelingDecisions() { | ||||||
| 1593 | WideningDecisions.clear(); | ||||||
| 1594 | Uniforms.clear(); | ||||||
| 1595 | Scalars.clear(); | ||||||
| 1596 | } | ||||||
| 1597 | |||||||
| 1598 | /// Convenience function that returns the value of vscale_range iff | ||||||
| 1599 | /// vscale_range.min == vscale_range.max or otherwise returns the value | ||||||
| 1600 | /// returned by the corresponding TLI method. | ||||||
| 1601 | std::optional<unsigned> getVScaleForTuning() const; | ||||||
| 1602 | |||||||
| 1603 | private: | ||||||
| 1604 | unsigned NumPredStores = 0; | ||||||
| 1605 | |||||||
| 1606 | /// \return An upper bound for the vectorization factors for both | ||||||
| 1607 | /// fixed and scalable vectorization, where the minimum-known number of | ||||||
| 1608 | /// elements is a power-of-2 larger than zero. If scalable vectorization is | ||||||
| 1609 | /// disabled or unsupported, then the scalable part will be equal to | ||||||
| 1610 | /// ElementCount::getScalable(0). | ||||||
| 1611 | FixedScalableVFPair computeFeasibleMaxVF(unsigned ConstTripCount, | ||||||
| 1612 | ElementCount UserVF, | ||||||
| 1613 | bool FoldTailByMasking); | ||||||
| 1614 | |||||||
| 1615 | /// \return the maximized element count based on the targets vector | ||||||
| 1616 | /// registers and the loop trip-count, but limited to a maximum safe VF. | ||||||
| 1617 | /// This is a helper function of computeFeasibleMaxVF. | ||||||
| 1618 | ElementCount getMaximizedVFForTarget(unsigned ConstTripCount, | ||||||
| 1619 | unsigned SmallestType, | ||||||
| 1620 | unsigned WidestType, | ||||||
| 1621 | ElementCount MaxSafeVF, | ||||||
| 1622 | bool FoldTailByMasking); | ||||||
| 1623 | |||||||
| 1624 | /// \return the maximum legal scalable VF, based on the safe max number | ||||||
| 1625 | /// of elements. | ||||||
| 1626 | ElementCount getMaxLegalScalableVF(unsigned MaxSafeElements); | ||||||
| 1627 | |||||||
| 1628 | /// The vectorization cost is a combination of the cost itself and a boolean | ||||||
| 1629 | /// indicating whether any of the contributing operations will actually | ||||||
| 1630 | /// operate on vector values after type legalization in the backend. If this | ||||||
| 1631 | /// latter value is false, then all operations will be scalarized (i.e. no | ||||||
| 1632 | /// vectorization has actually taken place). | ||||||
| 1633 | using VectorizationCostTy = std::pair<InstructionCost, bool>; | ||||||
| 1634 | |||||||
| 1635 | /// Returns the expected execution cost. The unit of the cost does | ||||||
| 1636 | /// not matter because we use the 'cost' units to compare different | ||||||
| 1637 | /// vector widths. The cost that is returned is *not* normalized by | ||||||
| 1638 | /// the factor width. If \p Invalid is not nullptr, this function | ||||||
| 1639 | /// will add a pair(Instruction*, ElementCount) to \p Invalid for | ||||||
| 1640 | /// each instruction that has an Invalid cost for the given VF. | ||||||
| 1641 | using InstructionVFPair = std::pair<Instruction *, ElementCount>; | ||||||
| 1642 | VectorizationCostTy | ||||||
| 1643 | expectedCost(ElementCount VF, | ||||||
| 1644 | SmallVectorImpl<InstructionVFPair> *Invalid = nullptr); | ||||||
| 1645 | |||||||
| 1646 | /// Returns the execution time cost of an instruction for a given vector | ||||||
| 1647 | /// width. Vector width of one means scalar. | ||||||
| 1648 | VectorizationCostTy getInstructionCost(Instruction *I, ElementCount VF); | ||||||
| 1649 | |||||||
| 1650 | /// The cost-computation logic from getInstructionCost which provides | ||||||
| 1651 | /// the vector type as an output parameter. | ||||||
| 1652 | InstructionCost getInstructionCost(Instruction *I, ElementCount VF, | ||||||
| 1653 | Type *&VectorTy); | ||||||
| 1654 | |||||||
| 1655 | /// Return the cost of instructions in an inloop reduction pattern, if I is | ||||||
| 1656 | /// part of that pattern. | ||||||
| 1657 | std::optional<InstructionCost> | ||||||
| 1658 | getReductionPatternCost(Instruction *I, ElementCount VF, Type *VectorTy, | ||||||
| 1659 | TTI::TargetCostKind CostKind); | ||||||
| 1660 | |||||||
| 1661 | /// Calculate vectorization cost of memory instruction \p I. | ||||||
| 1662 | InstructionCost getMemoryInstructionCost(Instruction *I, ElementCount VF); | ||||||
| 1663 | |||||||
| 1664 | /// The cost computation for scalarized memory instruction. | ||||||
| 1665 | InstructionCost getMemInstScalarizationCost(Instruction *I, ElementCount VF); | ||||||
| 1666 | |||||||
| 1667 | /// The cost computation for interleaving group of memory instructions. | ||||||
| 1668 | InstructionCost getInterleaveGroupCost(Instruction *I, ElementCount VF); | ||||||
| 1669 | |||||||
| 1670 | /// The cost computation for Gather/Scatter instruction. | ||||||
| 1671 | InstructionCost getGatherScatterCost(Instruction *I, ElementCount VF); | ||||||
| 1672 | |||||||
| 1673 | /// The cost computation for widening instruction \p I with consecutive | ||||||
| 1674 | /// memory access. | ||||||
| 1675 | InstructionCost getConsecutiveMemOpCost(Instruction *I, ElementCount VF); | ||||||
| 1676 | |||||||
| 1677 | /// The cost calculation for Load/Store instruction \p I with uniform pointer - | ||||||
| 1678 | /// Load: scalar load + broadcast. | ||||||
| 1679 | /// Store: scalar store + (loop invariant value stored? 0 : extract of last | ||||||
| 1680 | /// element) | ||||||
| 1681 | InstructionCost getUniformMemOpCost(Instruction *I, ElementCount VF); | ||||||
| 1682 | |||||||
| 1683 | /// Estimate the overhead of scalarizing an instruction. This is a | ||||||
| 1684 | /// convenience wrapper for the type-based getScalarizationOverhead API. | ||||||
| 1685 | InstructionCost getScalarizationOverhead(Instruction *I, | ||||||
| 1686 | ElementCount VF) const; | ||||||
| 1687 | |||||||
| 1688 | /// Returns true if an artificially high cost for emulated masked memrefs | ||||||
| 1689 | /// should be used. | ||||||
| 1690 | bool useEmulatedMaskMemRefHack(Instruction *I, ElementCount VF); | ||||||
| 1691 | |||||||
| 1692 | /// Map of scalar integer values to the smallest bitwidth they can be legally | ||||||
| 1693 | /// represented as. The vector equivalents of these values should be truncated | ||||||
| 1694 | /// to this type. | ||||||
| 1695 | MapVector<Instruction *, uint64_t> MinBWs; | ||||||
| 1696 | |||||||
| 1697 | /// A type representing the costs for instructions if they were to be | ||||||
| 1698 | /// scalarized rather than vectorized. The entries are Instruction-Cost | ||||||
| 1699 | /// pairs. | ||||||
| 1700 | using ScalarCostsTy = DenseMap<Instruction *, InstructionCost>; | ||||||
| 1701 | |||||||
| 1702 | /// A set containing all BasicBlocks that are known to present after | ||||||
| 1703 | /// vectorization as a predicated block. | ||||||
| 1704 | DenseMap<ElementCount, SmallPtrSet<BasicBlock *, 4>> | ||||||
| 1705 | PredicatedBBsAfterVectorization; | ||||||
| 1706 | |||||||
| 1707 | /// Records whether it is allowed to have the original scalar loop execute at | ||||||
| 1708 | /// least once. This may be needed as a fallback loop in case runtime | ||||||
| 1709 | /// aliasing/dependence checks fail, or to handle the tail/remainder | ||||||
| 1710 | /// iterations when the trip count is unknown or doesn't divide by the VF, | ||||||
| 1711 | /// or as a peel-loop to handle gaps in interleave-groups. | ||||||
| 1712 | /// Under optsize and when the trip count is very small we don't allow any | ||||||
| 1713 | /// iterations to execute in the scalar loop. | ||||||
| 1714 | ScalarEpilogueLowering ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; | ||||||
| 1715 | |||||||
| 1716 | /// All blocks of loop are to be masked to fold tail of scalar iterations. | ||||||
| 1717 | bool FoldTailByMasking = false; | ||||||
| 1718 | |||||||
| 1719 | /// A map holding scalar costs for different vectorization factors. The | ||||||
| 1720 | /// presence of a cost for an instruction in the mapping indicates that the | ||||||
| 1721 | /// instruction will be scalarized when vectorizing with the associated | ||||||
| 1722 | /// vectorization factor. The entries are VF-ScalarCostTy pairs. | ||||||
| 1723 | DenseMap<ElementCount, ScalarCostsTy> InstsToScalarize; | ||||||
| 1724 | |||||||
| 1725 | /// Holds the instructions known to be uniform after vectorization. | ||||||
| 1726 | /// The data is collected per VF. | ||||||
| 1727 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> Uniforms; | ||||||
| 1728 | |||||||
| 1729 | /// Holds the instructions known to be scalar after vectorization. | ||||||
| 1730 | /// The data is collected per VF. | ||||||
| 1731 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> Scalars; | ||||||
| 1732 | |||||||
| 1733 | /// Holds the instructions (address computations) that are forced to be | ||||||
| 1734 | /// scalarized. | ||||||
| 1735 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> ForcedScalars; | ||||||
| 1736 | |||||||
| 1737 | /// PHINodes of the reductions that should be expanded in-loop along with | ||||||
| 1738 | /// their associated chains of reduction operations, in program order from top | ||||||
| 1739 | /// (PHI) to bottom | ||||||
| 1740 | ReductionChainMap InLoopReductionChains; | ||||||
| 1741 | |||||||
| 1742 | /// A Map of inloop reduction operations and their immediate chain operand. | ||||||
| 1743 | /// FIXME: This can be removed once reductions can be costed correctly in | ||||||
| 1744 | /// vplan. This was added to allow quick lookup to the inloop operations, | ||||||
| 1745 | /// without having to loop through InLoopReductionChains. | ||||||
| 1746 | DenseMap<Instruction *, Instruction *> InLoopReductionImmediateChains; | ||||||
| 1747 | |||||||
| 1748 | /// Returns the expected difference in cost from scalarizing the expression | ||||||
| 1749 | /// feeding a predicated instruction \p PredInst. The instructions to | ||||||
| 1750 | /// scalarize and their scalar costs are collected in \p ScalarCosts. A | ||||||
| 1751 | /// non-negative return value implies the expression will be scalarized. | ||||||
| 1752 | /// Currently, only single-use chains are considered for scalarization. | ||||||
| 1753 | InstructionCost computePredInstDiscount(Instruction *PredInst, | ||||||
| 1754 | ScalarCostsTy &ScalarCosts, | ||||||
| 1755 | ElementCount VF); | ||||||
| 1756 | |||||||
| 1757 | /// Collect the instructions that are uniform after vectorization. An | ||||||
| 1758 | /// instruction is uniform if we represent it with a single scalar value in | ||||||
| 1759 | /// the vectorized loop corresponding to each vector iteration. Examples of | ||||||
| 1760 | /// uniform instructions include pointer operands of consecutive or | ||||||
| 1761 | /// interleaved memory accesses. Note that although uniformity implies an | ||||||
| 1762 | /// instruction will be scalar, the reverse is not true. In general, a | ||||||
| 1763 | /// scalarized instruction will be represented by VF scalar values in the | ||||||
| 1764 | /// vectorized loop, each corresponding to an iteration of the original | ||||||
| 1765 | /// scalar loop. | ||||||
| 1766 | void collectLoopUniforms(ElementCount VF); | ||||||
| 1767 | |||||||
| 1768 | /// Collect the instructions that are scalar after vectorization. An | ||||||
| 1769 | /// instruction is scalar if it is known to be uniform or will be scalarized | ||||||
| 1770 | /// during vectorization. collectLoopScalars should only add non-uniform nodes | ||||||
| 1771 | /// to the list if they are used by a load/store instruction that is marked as | ||||||
| 1772 | /// CM_Scalarize. Non-uniform scalarized instructions will be represented by | ||||||
| 1773 | /// VF values in the vectorized loop, each corresponding to an iteration of | ||||||
| 1774 | /// the original scalar loop. | ||||||
| 1775 | void collectLoopScalars(ElementCount VF); | ||||||
| 1776 | |||||||
| 1777 | /// Keeps cost model vectorization decision and cost for instructions. | ||||||
| 1778 | /// Right now it is used for memory instructions only. | ||||||
| 1779 | using DecisionList = DenseMap<std::pair<Instruction *, ElementCount>, | ||||||
| 1780 | std::pair<InstWidening, InstructionCost>>; | ||||||
| 1781 | |||||||
| 1782 | DecisionList WideningDecisions; | ||||||
| 1783 | |||||||
| 1784 | /// Returns true if \p V is expected to be vectorized and it needs to be | ||||||
| 1785 | /// extracted. | ||||||
| 1786 | bool needsExtract(Value *V, ElementCount VF) const { | ||||||
| 1787 | Instruction *I = dyn_cast<Instruction>(V); | ||||||
| 1788 | if (VF.isScalar() || !I || !TheLoop->contains(I) || | ||||||
| 1789 | TheLoop->isLoopInvariant(I)) | ||||||
| 1790 | return false; | ||||||
| 1791 | |||||||
| 1792 | // Assume we can vectorize V (and hence we need extraction) if the | ||||||
| 1793 | // scalars are not computed yet. This can happen, because it is called | ||||||
| 1794 | // via getScalarizationOverhead from setCostBasedWideningDecision, before | ||||||
| 1795 | // the scalars are collected. That should be a safe assumption in most | ||||||
| 1796 | // cases, because we check if the operands have vectorizable types | ||||||
| 1797 | // beforehand in LoopVectorizationLegality. | ||||||
| 1798 | return Scalars.find(VF) == Scalars.end() || | ||||||
| 1799 | !isScalarAfterVectorization(I, VF); | ||||||
| 1800 | }; | ||||||
| 1801 | |||||||
| 1802 | /// Returns a range containing only operands needing to be extracted. | ||||||
| 1803 | SmallVector<Value *, 4> filterExtractingOperands(Instruction::op_range Ops, | ||||||
| 1804 | ElementCount VF) const { | ||||||
| 1805 | return SmallVector<Value *, 4>(make_filter_range( | ||||||
| 1806 | Ops, [this, VF](Value *V) { return this->needsExtract(V, VF); })); | ||||||
| 1807 | } | ||||||
| 1808 | |||||||
| 1809 | /// Determines if we have the infrastructure to vectorize loop \p L and its | ||||||
| 1810 | /// epilogue, assuming the main loop is vectorized by \p VF. | ||||||
| 1811 | bool isCandidateForEpilogueVectorization(const Loop &L, | ||||||
| 1812 | const ElementCount VF) const; | ||||||
| 1813 | |||||||
| 1814 | /// Returns true if epilogue vectorization is considered profitable, and | ||||||
| 1815 | /// false otherwise. | ||||||
| 1816 | /// \p VF is the vectorization factor chosen for the original loop. | ||||||
| 1817 | bool isEpilogueVectorizationProfitable(const ElementCount VF) const; | ||||||
| 1818 | |||||||
| 1819 | public: | ||||||
| 1820 | /// The loop that we evaluate. | ||||||
| 1821 | Loop *TheLoop; | ||||||
| 1822 | |||||||
| 1823 | /// Predicated scalar evolution analysis. | ||||||
| 1824 | PredicatedScalarEvolution &PSE; | ||||||
| 1825 | |||||||
| 1826 | /// Loop Info analysis. | ||||||
| 1827 | LoopInfo *LI; | ||||||
| 1828 | |||||||
| 1829 | /// Vectorization legality. | ||||||
| 1830 | LoopVectorizationLegality *Legal; | ||||||
| 1831 | |||||||
| 1832 | /// Vector target information. | ||||||
| 1833 | const TargetTransformInfo &TTI; | ||||||
| 1834 | |||||||
| 1835 | /// Target Library Info. | ||||||
| 1836 | const TargetLibraryInfo *TLI; | ||||||
| 1837 | |||||||
| 1838 | /// Demanded bits analysis. | ||||||
| 1839 | DemandedBits *DB; | ||||||
| 1840 | |||||||
| 1841 | /// Assumption cache. | ||||||
| 1842 | AssumptionCache *AC; | ||||||
| 1843 | |||||||
| 1844 | /// Interface to emit optimization remarks. | ||||||
| 1845 | OptimizationRemarkEmitter *ORE; | ||||||
| 1846 | |||||||
| 1847 | const Function *TheFunction; | ||||||
| 1848 | |||||||
| 1849 | /// Loop Vectorize Hint. | ||||||
| 1850 | const LoopVectorizeHints *Hints; | ||||||
| 1851 | |||||||
| 1852 | /// The interleave access information contains groups of interleaved accesses | ||||||
| 1853 | /// with the same stride and close to each other. | ||||||
| 1854 | InterleavedAccessInfo &InterleaveInfo; | ||||||
| 1855 | |||||||
| 1856 | /// Values to ignore in the cost model. | ||||||
| 1857 | SmallPtrSet<const Value *, 16> ValuesToIgnore; | ||||||
| 1858 | |||||||
| 1859 | /// Values to ignore in the cost model when VF > 1. | ||||||
| 1860 | SmallPtrSet<const Value *, 16> VecValuesToIgnore; | ||||||
| 1861 | |||||||
| 1862 | /// All element types found in the loop. | ||||||
| 1863 | SmallPtrSet<Type *, 16> ElementTypesInLoop; | ||||||
| 1864 | |||||||
| 1865 | /// Profitable vector factors. | ||||||
| 1866 | SmallVector<VectorizationFactor, 8> ProfitableVFs; | ||||||
| 1867 | }; | ||||||
| 1868 | } // end namespace llvm | ||||||
| 1869 | |||||||
| 1870 | /// Helper struct to manage generating runtime checks for vectorization. | ||||||
| 1871 | /// | ||||||
| 1872 | /// The runtime checks are created up-front in temporary blocks to allow better | ||||||
| 1873 | /// estimating the cost and un-linked from the existing IR. After deciding to | ||||||
| 1874 | /// vectorize, the checks are moved back. If deciding not to vectorize, the | ||||||
| 1875 | /// temporary blocks are completely removed. | ||||||
| 1876 | class GeneratedRTChecks { | ||||||
| 1877 | /// Basic block which contains the generated SCEV checks, if any. | ||||||
| 1878 | BasicBlock *SCEVCheckBlock = nullptr; | ||||||
| 1879 | |||||||
| 1880 | /// The value representing the result of the generated SCEV checks. If it is | ||||||
| 1881 | /// nullptr, either no SCEV checks have been generated or they have been used. | ||||||
| 1882 | Value *SCEVCheckCond = nullptr; | ||||||
| 1883 | |||||||
| 1884 | /// Basic block which contains the generated memory runtime checks, if any. | ||||||
| 1885 | BasicBlock *MemCheckBlock = nullptr; | ||||||
| 1886 | |||||||
| 1887 | /// The value representing the result of the generated memory runtime checks. | ||||||
| 1888 | /// If it is nullptr, either no memory runtime checks have been generated or | ||||||
| 1889 | /// they have been used. | ||||||
| 1890 | Value *MemRuntimeCheckCond = nullptr; | ||||||
| 1891 | |||||||
| 1892 | DominatorTree *DT; | ||||||
| 1893 | LoopInfo *LI; | ||||||
| 1894 | TargetTransformInfo *TTI; | ||||||
| 1895 | |||||||
| 1896 | SCEVExpander SCEVExp; | ||||||
| 1897 | SCEVExpander MemCheckExp; | ||||||
| 1898 | |||||||
| 1899 | bool CostTooHigh = false; | ||||||
| 1900 | |||||||
| 1901 | public: | ||||||
| 1902 | GeneratedRTChecks(ScalarEvolution &SE, DominatorTree *DT, LoopInfo *LI, | ||||||
| 1903 | TargetTransformInfo *TTI, const DataLayout &DL) | ||||||
| 1904 | : DT(DT), LI(LI), TTI(TTI), SCEVExp(SE, DL, "scev.check"), | ||||||
| 1905 | MemCheckExp(SE, DL, "scev.check") {} | ||||||
| 1906 | |||||||
| 1907 | /// Generate runtime checks in SCEVCheckBlock and MemCheckBlock, so we can | ||||||
| 1908 | /// accurately estimate the cost of the runtime checks. The blocks are | ||||||
| 1909 | /// un-linked from the IR and is added back during vector code generation. If | ||||||
| 1910 | /// there is no vector code generation, the check blocks are removed | ||||||
| 1911 | /// completely. | ||||||
| 1912 | void Create(Loop *L, const LoopAccessInfo &LAI, | ||||||
| 1913 | const SCEVPredicate &UnionPred, ElementCount VF, unsigned IC) { | ||||||
| 1914 | |||||||
| 1915 | // Hard cutoff to limit compile-time increase in case a very large number of | ||||||
| 1916 | // runtime checks needs to be generated. | ||||||
| 1917 | // TODO: Skip cutoff if the loop is guaranteed to execute, e.g. due to | ||||||
| 1918 | // profile info. | ||||||
| 1919 | CostTooHigh = | ||||||
| 1920 | LAI.getNumRuntimePointerChecks() > VectorizeMemoryCheckThreshold; | ||||||
| 1921 | if (CostTooHigh) | ||||||
| 1922 | return; | ||||||
| 1923 | |||||||
| 1924 | BasicBlock *LoopHeader = L->getHeader(); | ||||||
| 1925 | BasicBlock *Preheader = L->getLoopPreheader(); | ||||||
| 1926 | |||||||
| 1927 | // Use SplitBlock to create blocks for SCEV & memory runtime checks to | ||||||
| 1928 | // ensure the blocks are properly added to LoopInfo & DominatorTree. Those | ||||||
| 1929 | // may be used by SCEVExpander. The blocks will be un-linked from their | ||||||
| 1930 | // predecessors and removed from LI & DT at the end of the function. | ||||||
| 1931 | if (!UnionPred.isAlwaysTrue()) { | ||||||
| 1932 | SCEVCheckBlock = SplitBlock(Preheader, Preheader->getTerminator(), DT, LI, | ||||||
| 1933 | nullptr, "vector.scevcheck"); | ||||||
| 1934 | |||||||
| 1935 | SCEVCheckCond = SCEVExp.expandCodeForPredicate( | ||||||
| 1936 | &UnionPred, SCEVCheckBlock->getTerminator()); | ||||||
| 1937 | } | ||||||
| 1938 | |||||||
| 1939 | const auto &RtPtrChecking = *LAI.getRuntimePointerChecking(); | ||||||
| 1940 | if (RtPtrChecking.Need) { | ||||||
| 1941 | auto *Pred = SCEVCheckBlock ? SCEVCheckBlock : Preheader; | ||||||
| 1942 | MemCheckBlock = SplitBlock(Pred, Pred->getTerminator(), DT, LI, nullptr, | ||||||
| 1943 | "vector.memcheck"); | ||||||
| 1944 | |||||||
| 1945 | auto DiffChecks = RtPtrChecking.getDiffChecks(); | ||||||
| 1946 | if (DiffChecks) { | ||||||
| 1947 | Value *RuntimeVF = nullptr; | ||||||
| 1948 | MemRuntimeCheckCond = addDiffRuntimeChecks( | ||||||
| 1949 | MemCheckBlock->getTerminator(), *DiffChecks, MemCheckExp, | ||||||
| 1950 | [VF, &RuntimeVF](IRBuilderBase &B, unsigned Bits) { | ||||||
| 1951 | if (!RuntimeVF) | ||||||
| 1952 | RuntimeVF = getRuntimeVF(B, B.getIntNTy(Bits), VF); | ||||||
| 1953 | return RuntimeVF; | ||||||
| 1954 | }, | ||||||
| 1955 | IC); | ||||||
| 1956 | } else { | ||||||
| 1957 | MemRuntimeCheckCond = | ||||||
| 1958 | addRuntimeChecks(MemCheckBlock->getTerminator(), L, | ||||||
| 1959 | RtPtrChecking.getChecks(), MemCheckExp); | ||||||
| 1960 | } | ||||||
| 1961 | assert(MemRuntimeCheckCond &&(static_cast <bool> (MemRuntimeCheckCond && "no RT checks generated although RtPtrChecking " "claimed checks are required") ? void (0) : __assert_fail ("MemRuntimeCheckCond && \"no RT checks generated although RtPtrChecking \" \"claimed checks are required\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1963, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 1962 | "no RT checks generated although RtPtrChecking "(static_cast <bool> (MemRuntimeCheckCond && "no RT checks generated although RtPtrChecking " "claimed checks are required") ? void (0) : __assert_fail ("MemRuntimeCheckCond && \"no RT checks generated although RtPtrChecking \" \"claimed checks are required\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1963, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 1963 | "claimed checks are required")(static_cast <bool> (MemRuntimeCheckCond && "no RT checks generated although RtPtrChecking " "claimed checks are required") ? void (0) : __assert_fail ("MemRuntimeCheckCond && \"no RT checks generated although RtPtrChecking \" \"claimed checks are required\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 1963, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 1964 | } | ||||||
| 1965 | |||||||
| 1966 | if (!MemCheckBlock && !SCEVCheckBlock) | ||||||
| 1967 | return; | ||||||
| 1968 | |||||||
| 1969 | // Unhook the temporary block with the checks, update various places | ||||||
| 1970 | // accordingly. | ||||||
| 1971 | if (SCEVCheckBlock) | ||||||
| 1972 | SCEVCheckBlock->replaceAllUsesWith(Preheader); | ||||||
| 1973 | if (MemCheckBlock) | ||||||
| 1974 | MemCheckBlock->replaceAllUsesWith(Preheader); | ||||||
| 1975 | |||||||
| 1976 | if (SCEVCheckBlock) { | ||||||
| 1977 | SCEVCheckBlock->getTerminator()->moveBefore(Preheader->getTerminator()); | ||||||
| 1978 | new UnreachableInst(Preheader->getContext(), SCEVCheckBlock); | ||||||
| 1979 | Preheader->getTerminator()->eraseFromParent(); | ||||||
| 1980 | } | ||||||
| 1981 | if (MemCheckBlock) { | ||||||
| 1982 | MemCheckBlock->getTerminator()->moveBefore(Preheader->getTerminator()); | ||||||
| 1983 | new UnreachableInst(Preheader->getContext(), MemCheckBlock); | ||||||
| 1984 | Preheader->getTerminator()->eraseFromParent(); | ||||||
| 1985 | } | ||||||
| 1986 | |||||||
| 1987 | DT->changeImmediateDominator(LoopHeader, Preheader); | ||||||
| 1988 | if (MemCheckBlock) { | ||||||
| 1989 | DT->eraseNode(MemCheckBlock); | ||||||
| 1990 | LI->removeBlock(MemCheckBlock); | ||||||
| 1991 | } | ||||||
| 1992 | if (SCEVCheckBlock) { | ||||||
| 1993 | DT->eraseNode(SCEVCheckBlock); | ||||||
| 1994 | LI->removeBlock(SCEVCheckBlock); | ||||||
| 1995 | } | ||||||
| 1996 | } | ||||||
| 1997 | |||||||
| 1998 | InstructionCost getCost() { | ||||||
| 1999 | if (SCEVCheckBlock || MemCheckBlock) | ||||||
| 2000 | LLVM_DEBUG(dbgs() << "Calculating cost of runtime checks:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Calculating cost of runtime checks:\n" ; } } while (false); | ||||||
| 2001 | |||||||
| 2002 | if (CostTooHigh) { | ||||||
| 2003 | InstructionCost Cost; | ||||||
| 2004 | Cost.setInvalid(); | ||||||
| 2005 | LLVM_DEBUG(dbgs() << " number of checks exceeded threshold\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " number of checks exceeded threshold\n" ; } } while (false); | ||||||
| 2006 | return Cost; | ||||||
| 2007 | } | ||||||
| 2008 | |||||||
| 2009 | InstructionCost RTCheckCost = 0; | ||||||
| 2010 | if (SCEVCheckBlock) | ||||||
| 2011 | for (Instruction &I : *SCEVCheckBlock) { | ||||||
| 2012 | if (SCEVCheckBlock->getTerminator() == &I) | ||||||
| 2013 | continue; | ||||||
| 2014 | InstructionCost C = | ||||||
| 2015 | TTI->getInstructionCost(&I, TTI::TCK_RecipThroughput); | ||||||
| 2016 | LLVM_DEBUG(dbgs() << " " << C << " for " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " " << C << " for " << I << "\n"; } } while (false); | ||||||
| 2017 | RTCheckCost += C; | ||||||
| 2018 | } | ||||||
| 2019 | if (MemCheckBlock) | ||||||
| 2020 | for (Instruction &I : *MemCheckBlock) { | ||||||
| 2021 | if (MemCheckBlock->getTerminator() == &I) | ||||||
| 2022 | continue; | ||||||
| 2023 | InstructionCost C = | ||||||
| 2024 | TTI->getInstructionCost(&I, TTI::TCK_RecipThroughput); | ||||||
| 2025 | LLVM_DEBUG(dbgs() << " " << C << " for " << I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " " << C << " for " << I << "\n"; } } while (false); | ||||||
| 2026 | RTCheckCost += C; | ||||||
| 2027 | } | ||||||
| 2028 | |||||||
| 2029 | if (SCEVCheckBlock || MemCheckBlock) | ||||||
| 2030 | LLVM_DEBUG(dbgs() << "Total cost of runtime checks: " << RTCheckCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Total cost of runtime checks: " << RTCheckCost << "\n"; } } while (false) | ||||||
| 2031 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Total cost of runtime checks: " << RTCheckCost << "\n"; } } while (false); | ||||||
| 2032 | |||||||
| 2033 | return RTCheckCost; | ||||||
| 2034 | } | ||||||
| 2035 | |||||||
| 2036 | /// Remove the created SCEV & memory runtime check blocks & instructions, if | ||||||
| 2037 | /// unused. | ||||||
| 2038 | ~GeneratedRTChecks() { | ||||||
| 2039 | SCEVExpanderCleaner SCEVCleaner(SCEVExp); | ||||||
| 2040 | SCEVExpanderCleaner MemCheckCleaner(MemCheckExp); | ||||||
| 2041 | if (!SCEVCheckCond) | ||||||
| 2042 | SCEVCleaner.markResultUsed(); | ||||||
| 2043 | |||||||
| 2044 | if (!MemRuntimeCheckCond) | ||||||
| 2045 | MemCheckCleaner.markResultUsed(); | ||||||
| 2046 | |||||||
| 2047 | if (MemRuntimeCheckCond) { | ||||||
| 2048 | auto &SE = *MemCheckExp.getSE(); | ||||||
| 2049 | // Memory runtime check generation creates compares that use expanded | ||||||
| 2050 | // values. Remove them before running the SCEVExpanderCleaners. | ||||||
| 2051 | for (auto &I : make_early_inc_range(reverse(*MemCheckBlock))) { | ||||||
| 2052 | if (MemCheckExp.isInsertedInstruction(&I)) | ||||||
| 2053 | continue; | ||||||
| 2054 | SE.forgetValue(&I); | ||||||
| 2055 | I.eraseFromParent(); | ||||||
| 2056 | } | ||||||
| 2057 | } | ||||||
| 2058 | MemCheckCleaner.cleanup(); | ||||||
| 2059 | SCEVCleaner.cleanup(); | ||||||
| 2060 | |||||||
| 2061 | if (SCEVCheckCond) | ||||||
| 2062 | SCEVCheckBlock->eraseFromParent(); | ||||||
| 2063 | if (MemRuntimeCheckCond) | ||||||
| 2064 | MemCheckBlock->eraseFromParent(); | ||||||
| 2065 | } | ||||||
| 2066 | |||||||
| 2067 | /// Adds the generated SCEVCheckBlock before \p LoopVectorPreHeader and | ||||||
| 2068 | /// adjusts the branches to branch to the vector preheader or \p Bypass, | ||||||
| 2069 | /// depending on the generated condition. | ||||||
| 2070 | BasicBlock *emitSCEVChecks(BasicBlock *Bypass, | ||||||
| 2071 | BasicBlock *LoopVectorPreHeader, | ||||||
| 2072 | BasicBlock *LoopExitBlock) { | ||||||
| 2073 | if (!SCEVCheckCond) | ||||||
| 2074 | return nullptr; | ||||||
| 2075 | |||||||
| 2076 | Value *Cond = SCEVCheckCond; | ||||||
| 2077 | // Mark the check as used, to prevent it from being removed during cleanup. | ||||||
| 2078 | SCEVCheckCond = nullptr; | ||||||
| 2079 | if (auto *C = dyn_cast<ConstantInt>(Cond)) | ||||||
| 2080 | if (C->isZero()) | ||||||
| 2081 | return nullptr; | ||||||
| 2082 | |||||||
| 2083 | auto *Pred = LoopVectorPreHeader->getSinglePredecessor(); | ||||||
| 2084 | |||||||
| 2085 | BranchInst::Create(LoopVectorPreHeader, SCEVCheckBlock); | ||||||
| 2086 | // Create new preheader for vector loop. | ||||||
| 2087 | if (auto *PL = LI->getLoopFor(LoopVectorPreHeader)) | ||||||
| 2088 | PL->addBasicBlockToLoop(SCEVCheckBlock, *LI); | ||||||
| 2089 | |||||||
| 2090 | SCEVCheckBlock->getTerminator()->eraseFromParent(); | ||||||
| 2091 | SCEVCheckBlock->moveBefore(LoopVectorPreHeader); | ||||||
| 2092 | Pred->getTerminator()->replaceSuccessorWith(LoopVectorPreHeader, | ||||||
| 2093 | SCEVCheckBlock); | ||||||
| 2094 | |||||||
| 2095 | DT->addNewBlock(SCEVCheckBlock, Pred); | ||||||
| 2096 | DT->changeImmediateDominator(LoopVectorPreHeader, SCEVCheckBlock); | ||||||
| 2097 | |||||||
| 2098 | ReplaceInstWithInst(SCEVCheckBlock->getTerminator(), | ||||||
| 2099 | BranchInst::Create(Bypass, LoopVectorPreHeader, Cond)); | ||||||
| 2100 | return SCEVCheckBlock; | ||||||
| 2101 | } | ||||||
| 2102 | |||||||
| 2103 | /// Adds the generated MemCheckBlock before \p LoopVectorPreHeader and adjusts | ||||||
| 2104 | /// the branches to branch to the vector preheader or \p Bypass, depending on | ||||||
| 2105 | /// the generated condition. | ||||||
| 2106 | BasicBlock *emitMemRuntimeChecks(BasicBlock *Bypass, | ||||||
| 2107 | BasicBlock *LoopVectorPreHeader) { | ||||||
| 2108 | // Check if we generated code that checks in runtime if arrays overlap. | ||||||
| 2109 | if (!MemRuntimeCheckCond) | ||||||
| 2110 | return nullptr; | ||||||
| 2111 | |||||||
| 2112 | auto *Pred = LoopVectorPreHeader->getSinglePredecessor(); | ||||||
| 2113 | Pred->getTerminator()->replaceSuccessorWith(LoopVectorPreHeader, | ||||||
| 2114 | MemCheckBlock); | ||||||
| 2115 | |||||||
| 2116 | DT->addNewBlock(MemCheckBlock, Pred); | ||||||
| 2117 | DT->changeImmediateDominator(LoopVectorPreHeader, MemCheckBlock); | ||||||
| 2118 | MemCheckBlock->moveBefore(LoopVectorPreHeader); | ||||||
| 2119 | |||||||
| 2120 | if (auto *PL = LI->getLoopFor(LoopVectorPreHeader)) | ||||||
| 2121 | PL->addBasicBlockToLoop(MemCheckBlock, *LI); | ||||||
| 2122 | |||||||
| 2123 | ReplaceInstWithInst( | ||||||
| 2124 | MemCheckBlock->getTerminator(), | ||||||
| 2125 | BranchInst::Create(Bypass, LoopVectorPreHeader, MemRuntimeCheckCond)); | ||||||
| 2126 | MemCheckBlock->getTerminator()->setDebugLoc( | ||||||
| 2127 | Pred->getTerminator()->getDebugLoc()); | ||||||
| 2128 | |||||||
| 2129 | // Mark the check as used, to prevent it from being removed during cleanup. | ||||||
| 2130 | MemRuntimeCheckCond = nullptr; | ||||||
| 2131 | return MemCheckBlock; | ||||||
| 2132 | } | ||||||
| 2133 | }; | ||||||
| 2134 | |||||||
| 2135 | // Return true if \p OuterLp is an outer loop annotated with hints for explicit | ||||||
| 2136 | // vectorization. The loop needs to be annotated with #pragma omp simd | ||||||
| 2137 | // simdlen(#) or #pragma clang vectorize(enable) vectorize_width(#). If the | ||||||
| 2138 | // vector length information is not provided, vectorization is not considered | ||||||
| 2139 | // explicit. Interleave hints are not allowed either. These limitations will be | ||||||
| 2140 | // relaxed in the future. | ||||||
| 2141 | // Please, note that we are currently forced to abuse the pragma 'clang | ||||||
| 2142 | // vectorize' semantics. This pragma provides *auto-vectorization hints* | ||||||
| 2143 | // (i.e., LV must check that vectorization is legal) whereas pragma 'omp simd' | ||||||
| 2144 | // provides *explicit vectorization hints* (LV can bypass legal checks and | ||||||
| 2145 | // assume that vectorization is legal). However, both hints are implemented | ||||||
| 2146 | // using the same metadata (llvm.loop.vectorize, processed by | ||||||
| 2147 | // LoopVectorizeHints). This will be fixed in the future when the native IR | ||||||
| 2148 | // representation for pragma 'omp simd' is introduced. | ||||||
| 2149 | static bool isExplicitVecOuterLoop(Loop *OuterLp, | ||||||
| 2150 | OptimizationRemarkEmitter *ORE) { | ||||||
| 2151 | assert(!OuterLp->isInnermost() && "This is not an outer loop")(static_cast <bool> (!OuterLp->isInnermost() && "This is not an outer loop") ? void (0) : __assert_fail ("!OuterLp->isInnermost() && \"This is not an outer loop\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2151, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2152 | LoopVectorizeHints Hints(OuterLp, true /*DisableInterleaving*/, *ORE); | ||||||
| 2153 | |||||||
| 2154 | // Only outer loops with an explicit vectorization hint are supported. | ||||||
| 2155 | // Unannotated outer loops are ignored. | ||||||
| 2156 | if (Hints.getForce() == LoopVectorizeHints::FK_Undefined) | ||||||
| 2157 | return false; | ||||||
| 2158 | |||||||
| 2159 | Function *Fn = OuterLp->getHeader()->getParent(); | ||||||
| 2160 | if (!Hints.allowVectorization(Fn, OuterLp, | ||||||
| 2161 | true /*VectorizeOnlyWhenForced*/)) { | ||||||
| 2162 | LLVM_DEBUG(dbgs() << "LV: Loop hints prevent outer loop vectorization.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints prevent outer loop vectorization.\n" ; } } while (false); | ||||||
| 2163 | return false; | ||||||
| 2164 | } | ||||||
| 2165 | |||||||
| 2166 | if (Hints.getInterleave() > 1) { | ||||||
| 2167 | // TODO: Interleave support is future work. | ||||||
| 2168 | LLVM_DEBUG(dbgs() << "LV: Not vectorizing: Interleave is not supported for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not vectorizing: Interleave is not supported for " "outer loops.\n"; } } while (false) | ||||||
| 2169 | "outer loops.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not vectorizing: Interleave is not supported for " "outer loops.\n"; } } while (false); | ||||||
| 2170 | Hints.emitRemarkWithHints(); | ||||||
| 2171 | return false; | ||||||
| 2172 | } | ||||||
| 2173 | |||||||
| 2174 | return true; | ||||||
| 2175 | } | ||||||
| 2176 | |||||||
| 2177 | static void collectSupportedLoops(Loop &L, LoopInfo *LI, | ||||||
| 2178 | OptimizationRemarkEmitter *ORE, | ||||||
| 2179 | SmallVectorImpl<Loop *> &V) { | ||||||
| 2180 | // Collect inner loops and outer loops without irreducible control flow. For | ||||||
| 2181 | // now, only collect outer loops that have explicit vectorization hints. If we | ||||||
| 2182 | // are stress testing the VPlan H-CFG construction, we collect the outermost | ||||||
| 2183 | // loop of every loop nest. | ||||||
| 2184 | if (L.isInnermost() || VPlanBuildStressTest || | ||||||
| 2185 | (EnableVPlanNativePath && isExplicitVecOuterLoop(&L, ORE))) { | ||||||
| 2186 | LoopBlocksRPO RPOT(&L); | ||||||
| 2187 | RPOT.perform(LI); | ||||||
| 2188 | if (!containsIrreducibleCFG<const BasicBlock *>(RPOT, *LI)) { | ||||||
| 2189 | V.push_back(&L); | ||||||
| 2190 | // TODO: Collect inner loops inside marked outer loops in case | ||||||
| 2191 | // vectorization fails for the outer loop. Do not invoke | ||||||
| 2192 | // 'containsIrreducibleCFG' again for inner loops when the outer loop is | ||||||
| 2193 | // already known to be reducible. We can use an inherited attribute for | ||||||
| 2194 | // that. | ||||||
| 2195 | return; | ||||||
| 2196 | } | ||||||
| 2197 | } | ||||||
| 2198 | for (Loop *InnerL : L) | ||||||
| 2199 | collectSupportedLoops(*InnerL, LI, ORE, V); | ||||||
| 2200 | } | ||||||
| 2201 | |||||||
| 2202 | namespace { | ||||||
| 2203 | |||||||
| 2204 | /// The LoopVectorize Pass. | ||||||
| 2205 | struct LoopVectorize : public FunctionPass { | ||||||
| 2206 | /// Pass identification, replacement for typeid | ||||||
| 2207 | static char ID; | ||||||
| 2208 | |||||||
| 2209 | LoopVectorizePass Impl; | ||||||
| 2210 | |||||||
| 2211 | explicit LoopVectorize(bool InterleaveOnlyWhenForced = false, | ||||||
| 2212 | bool VectorizeOnlyWhenForced = false) | ||||||
| 2213 | : FunctionPass(ID), | ||||||
| 2214 | Impl({InterleaveOnlyWhenForced, VectorizeOnlyWhenForced}) { | ||||||
| 2215 | initializeLoopVectorizePass(*PassRegistry::getPassRegistry()); | ||||||
| 2216 | } | ||||||
| 2217 | |||||||
| 2218 | bool runOnFunction(Function &F) override { | ||||||
| 2219 | if (skipFunction(F)) | ||||||
| 2220 | return false; | ||||||
| 2221 | |||||||
| 2222 | auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); | ||||||
| 2223 | auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | ||||||
| 2224 | auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | ||||||
| 2225 | auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | ||||||
| 2226 | auto *BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI(); | ||||||
| 2227 | auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); | ||||||
| 2228 | auto *TLI = TLIP ? &TLIP->getTLI(F) : nullptr; | ||||||
| 2229 | auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); | ||||||
| 2230 | auto &LAIs = getAnalysis<LoopAccessLegacyAnalysis>().getLAIs(); | ||||||
| 2231 | auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits(); | ||||||
| 2232 | auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); | ||||||
| 2233 | auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); | ||||||
| 2234 | |||||||
| 2235 | return Impl | ||||||
| 2236 | .runImpl(F, *SE, *LI, *TTI, *DT, *BFI, TLI, *DB, *AC, LAIs, *ORE, PSI) | ||||||
| 2237 | .MadeAnyChange; | ||||||
| 2238 | } | ||||||
| 2239 | |||||||
| 2240 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||||
| 2241 | AU.addRequired<AssumptionCacheTracker>(); | ||||||
| 2242 | AU.addRequired<BlockFrequencyInfoWrapperPass>(); | ||||||
| 2243 | AU.addRequired<DominatorTreeWrapperPass>(); | ||||||
| 2244 | AU.addRequired<LoopInfoWrapperPass>(); | ||||||
| 2245 | AU.addRequired<ScalarEvolutionWrapperPass>(); | ||||||
| 2246 | AU.addRequired<TargetTransformInfoWrapperPass>(); | ||||||
| 2247 | AU.addRequired<LoopAccessLegacyAnalysis>(); | ||||||
| 2248 | AU.addRequired<DemandedBitsWrapperPass>(); | ||||||
| 2249 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); | ||||||
| 2250 | AU.addRequired<InjectTLIMappingsLegacy>(); | ||||||
| 2251 | |||||||
| 2252 | // We currently do not preserve loopinfo/dominator analyses with outer loop | ||||||
| 2253 | // vectorization. Until this is addressed, mark these analyses as preserved | ||||||
| 2254 | // only for non-VPlan-native path. | ||||||
| 2255 | // TODO: Preserve Loop and Dominator analyses for VPlan-native path. | ||||||
| 2256 | if (!EnableVPlanNativePath) { | ||||||
| 2257 | AU.addPreserved<LoopInfoWrapperPass>(); | ||||||
| 2258 | AU.addPreserved<DominatorTreeWrapperPass>(); | ||||||
| 2259 | } | ||||||
| 2260 | |||||||
| 2261 | AU.addPreserved<BasicAAWrapperPass>(); | ||||||
| 2262 | AU.addPreserved<GlobalsAAWrapperPass>(); | ||||||
| 2263 | AU.addRequired<ProfileSummaryInfoWrapperPass>(); | ||||||
| 2264 | } | ||||||
| 2265 | }; | ||||||
| 2266 | |||||||
| 2267 | } // end anonymous namespace | ||||||
| 2268 | |||||||
| 2269 | //===----------------------------------------------------------------------===// | ||||||
| 2270 | // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and | ||||||
| 2271 | // LoopVectorizationCostModel and LoopVectorizationPlanner. | ||||||
| 2272 | //===----------------------------------------------------------------------===// | ||||||
| 2273 | |||||||
| 2274 | Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { | ||||||
| 2275 | // We need to place the broadcast of invariant variables outside the loop, | ||||||
| 2276 | // but only if it's proven safe to do so. Else, broadcast will be inside | ||||||
| 2277 | // vector loop body. | ||||||
| 2278 | Instruction *Instr = dyn_cast<Instruction>(V); | ||||||
| 2279 | bool SafeToHoist = OrigLoop->isLoopInvariant(V) && | ||||||
| 2280 | (!Instr || | ||||||
| 2281 | DT->dominates(Instr->getParent(), LoopVectorPreHeader)); | ||||||
| 2282 | // Place the code for broadcasting invariant variables in the new preheader. | ||||||
| 2283 | IRBuilder<>::InsertPointGuard Guard(Builder); | ||||||
| 2284 | if (SafeToHoist) | ||||||
| 2285 | Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); | ||||||
| 2286 | |||||||
| 2287 | // Broadcast the scalar into all locations in the vector. | ||||||
| 2288 | Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast"); | ||||||
| 2289 | |||||||
| 2290 | return Shuf; | ||||||
| 2291 | } | ||||||
| 2292 | |||||||
| 2293 | /// This function adds | ||||||
| 2294 | /// (StartIdx * Step, (StartIdx + 1) * Step, (StartIdx + 2) * Step, ...) | ||||||
| 2295 | /// to each vector element of Val. The sequence starts at StartIndex. | ||||||
| 2296 | /// \p Opcode is relevant for FP induction variable. | ||||||
| 2297 | static Value *getStepVector(Value *Val, Value *StartIdx, Value *Step, | ||||||
| 2298 | Instruction::BinaryOps BinOp, ElementCount VF, | ||||||
| 2299 | IRBuilderBase &Builder) { | ||||||
| 2300 | assert(VF.isVector() && "only vector VFs are supported")(static_cast <bool> (VF.isVector() && "only vector VFs are supported" ) ? void (0) : __assert_fail ("VF.isVector() && \"only vector VFs are supported\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2300, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2301 | |||||||
| 2302 | // Create and check the types. | ||||||
| 2303 | auto *ValVTy = cast<VectorType>(Val->getType()); | ||||||
| 2304 | ElementCount VLen = ValVTy->getElementCount(); | ||||||
| 2305 | |||||||
| 2306 | Type *STy = Val->getType()->getScalarType(); | ||||||
| 2307 | assert((STy->isIntegerTy() || STy->isFloatingPointTy()) &&(static_cast <bool> ((STy->isIntegerTy() || STy-> isFloatingPointTy()) && "Induction Step must be an integer or FP" ) ? void (0) : __assert_fail ("(STy->isIntegerTy() || STy->isFloatingPointTy()) && \"Induction Step must be an integer or FP\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2308, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2308 | "Induction Step must be an integer or FP")(static_cast <bool> ((STy->isIntegerTy() || STy-> isFloatingPointTy()) && "Induction Step must be an integer or FP" ) ? void (0) : __assert_fail ("(STy->isIntegerTy() || STy->isFloatingPointTy()) && \"Induction Step must be an integer or FP\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2308, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2309 | assert(Step->getType() == STy && "Step has wrong type")(static_cast <bool> (Step->getType() == STy && "Step has wrong type") ? void (0) : __assert_fail ("Step->getType() == STy && \"Step has wrong type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2309, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2310 | |||||||
| 2311 | SmallVector<Constant *, 8> Indices; | ||||||
| 2312 | |||||||
| 2313 | // Create a vector of consecutive numbers from zero to VF. | ||||||
| 2314 | VectorType *InitVecValVTy = ValVTy; | ||||||
| 2315 | if (STy->isFloatingPointTy()) { | ||||||
| 2316 | Type *InitVecValSTy = | ||||||
| 2317 | IntegerType::get(STy->getContext(), STy->getScalarSizeInBits()); | ||||||
| 2318 | InitVecValVTy = VectorType::get(InitVecValSTy, VLen); | ||||||
| 2319 | } | ||||||
| 2320 | Value *InitVec = Builder.CreateStepVector(InitVecValVTy); | ||||||
| 2321 | |||||||
| 2322 | // Splat the StartIdx | ||||||
| 2323 | Value *StartIdxSplat = Builder.CreateVectorSplat(VLen, StartIdx); | ||||||
| 2324 | |||||||
| 2325 | if (STy->isIntegerTy()) { | ||||||
| 2326 | InitVec = Builder.CreateAdd(InitVec, StartIdxSplat); | ||||||
| 2327 | Step = Builder.CreateVectorSplat(VLen, Step); | ||||||
| 2328 | assert(Step->getType() == Val->getType() && "Invalid step vec")(static_cast <bool> (Step->getType() == Val->getType () && "Invalid step vec") ? void (0) : __assert_fail ( "Step->getType() == Val->getType() && \"Invalid step vec\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2328, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2329 | // FIXME: The newly created binary instructions should contain nsw/nuw | ||||||
| 2330 | // flags, which can be found from the original scalar operations. | ||||||
| 2331 | Step = Builder.CreateMul(InitVec, Step); | ||||||
| 2332 | return Builder.CreateAdd(Val, Step, "induction"); | ||||||
| 2333 | } | ||||||
| 2334 | |||||||
| 2335 | // Floating point induction. | ||||||
| 2336 | assert((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) &&(static_cast <bool> ((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && "Binary Opcode should be specified for FP induction" ) ? void (0) : __assert_fail ("(BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && \"Binary Opcode should be specified for FP induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2337, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2337 | "Binary Opcode should be specified for FP induction")(static_cast <bool> ((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && "Binary Opcode should be specified for FP induction" ) ? void (0) : __assert_fail ("(BinOp == Instruction::FAdd || BinOp == Instruction::FSub) && \"Binary Opcode should be specified for FP induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2337, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2338 | InitVec = Builder.CreateUIToFP(InitVec, ValVTy); | ||||||
| 2339 | InitVec = Builder.CreateFAdd(InitVec, StartIdxSplat); | ||||||
| 2340 | |||||||
| 2341 | Step = Builder.CreateVectorSplat(VLen, Step); | ||||||
| 2342 | Value *MulOp = Builder.CreateFMul(InitVec, Step); | ||||||
| 2343 | return Builder.CreateBinOp(BinOp, Val, MulOp, "induction"); | ||||||
| 2344 | } | ||||||
| 2345 | |||||||
| 2346 | /// Compute scalar induction steps. \p ScalarIV is the scalar induction | ||||||
| 2347 | /// variable on which to base the steps, \p Step is the size of the step. | ||||||
| 2348 | static void buildScalarSteps(Value *ScalarIV, Value *Step, | ||||||
| 2349 | const InductionDescriptor &ID, VPValue *Def, | ||||||
| 2350 | VPTransformState &State) { | ||||||
| 2351 | IRBuilderBase &Builder = State.Builder; | ||||||
| 2352 | |||||||
| 2353 | // Ensure step has the same type as that of scalar IV. | ||||||
| 2354 | Type *ScalarIVTy = ScalarIV->getType()->getScalarType(); | ||||||
| 2355 | if (ScalarIVTy != Step->getType()) { | ||||||
| 2356 | // TODO: Also use VPDerivedIVRecipe when only the step needs truncating, to | ||||||
| 2357 | // avoid separate truncate here. | ||||||
| 2358 | assert(Step->getType()->isIntegerTy() &&(static_cast <bool> (Step->getType()->isIntegerTy () && "Truncation requires an integer step") ? void ( 0) : __assert_fail ("Step->getType()->isIntegerTy() && \"Truncation requires an integer step\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2359, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2359 | "Truncation requires an integer step")(static_cast <bool> (Step->getType()->isIntegerTy () && "Truncation requires an integer step") ? void ( 0) : __assert_fail ("Step->getType()->isIntegerTy() && \"Truncation requires an integer step\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2359, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2360 | Step = State.Builder.CreateTrunc(Step, ScalarIVTy); | ||||||
| 2361 | } | ||||||
| 2362 | |||||||
| 2363 | // We build scalar steps for both integer and floating-point induction | ||||||
| 2364 | // variables. Here, we determine the kind of arithmetic we will perform. | ||||||
| 2365 | Instruction::BinaryOps AddOp; | ||||||
| 2366 | Instruction::BinaryOps MulOp; | ||||||
| 2367 | if (ScalarIVTy->isIntegerTy()) { | ||||||
| 2368 | AddOp = Instruction::Add; | ||||||
| 2369 | MulOp = Instruction::Mul; | ||||||
| 2370 | } else { | ||||||
| 2371 | AddOp = ID.getInductionOpcode(); | ||||||
| 2372 | MulOp = Instruction::FMul; | ||||||
| 2373 | } | ||||||
| 2374 | |||||||
| 2375 | // Determine the number of scalars we need to generate for each unroll | ||||||
| 2376 | // iteration. | ||||||
| 2377 | bool FirstLaneOnly = vputils::onlyFirstLaneUsed(Def); | ||||||
| 2378 | // Compute the scalar steps and save the results in State. | ||||||
| 2379 | Type *IntStepTy = IntegerType::get(ScalarIVTy->getContext(), | ||||||
| 2380 | ScalarIVTy->getScalarSizeInBits()); | ||||||
| 2381 | Type *VecIVTy = nullptr; | ||||||
| 2382 | Value *UnitStepVec = nullptr, *SplatStep = nullptr, *SplatIV = nullptr; | ||||||
| 2383 | if (!FirstLaneOnly && State.VF.isScalable()) { | ||||||
| 2384 | VecIVTy = VectorType::get(ScalarIVTy, State.VF); | ||||||
| 2385 | UnitStepVec = | ||||||
| 2386 | Builder.CreateStepVector(VectorType::get(IntStepTy, State.VF)); | ||||||
| 2387 | SplatStep = Builder.CreateVectorSplat(State.VF, Step); | ||||||
| 2388 | SplatIV = Builder.CreateVectorSplat(State.VF, ScalarIV); | ||||||
| 2389 | } | ||||||
| 2390 | |||||||
| 2391 | unsigned StartPart = 0; | ||||||
| 2392 | unsigned EndPart = State.UF; | ||||||
| 2393 | unsigned StartLane = 0; | ||||||
| 2394 | unsigned EndLane = FirstLaneOnly ? 1 : State.VF.getKnownMinValue(); | ||||||
| 2395 | if (State.Instance) { | ||||||
| 2396 | StartPart = State.Instance->Part; | ||||||
| 2397 | EndPart = StartPart + 1; | ||||||
| 2398 | StartLane = State.Instance->Lane.getKnownLane(); | ||||||
| 2399 | EndLane = StartLane + 1; | ||||||
| 2400 | } | ||||||
| 2401 | for (unsigned Part = StartPart; Part < EndPart; ++Part) { | ||||||
| 2402 | Value *StartIdx0 = createStepForVF(Builder, IntStepTy, State.VF, Part); | ||||||
| 2403 | |||||||
| 2404 | if (!FirstLaneOnly && State.VF.isScalable()) { | ||||||
| 2405 | auto *SplatStartIdx = Builder.CreateVectorSplat(State.VF, StartIdx0); | ||||||
| 2406 | auto *InitVec = Builder.CreateAdd(SplatStartIdx, UnitStepVec); | ||||||
| 2407 | if (ScalarIVTy->isFloatingPointTy()) | ||||||
| 2408 | InitVec = Builder.CreateSIToFP(InitVec, VecIVTy); | ||||||
| 2409 | auto *Mul = Builder.CreateBinOp(MulOp, InitVec, SplatStep); | ||||||
| 2410 | auto *Add = Builder.CreateBinOp(AddOp, SplatIV, Mul); | ||||||
| 2411 | State.set(Def, Add, Part); | ||||||
| 2412 | // It's useful to record the lane values too for the known minimum number | ||||||
| 2413 | // of elements so we do those below. This improves the code quality when | ||||||
| 2414 | // trying to extract the first element, for example. | ||||||
| 2415 | } | ||||||
| 2416 | |||||||
| 2417 | if (ScalarIVTy->isFloatingPointTy()) | ||||||
| 2418 | StartIdx0 = Builder.CreateSIToFP(StartIdx0, ScalarIVTy); | ||||||
| 2419 | |||||||
| 2420 | for (unsigned Lane = StartLane; Lane < EndLane; ++Lane) { | ||||||
| 2421 | Value *StartIdx = Builder.CreateBinOp( | ||||||
| 2422 | AddOp, StartIdx0, getSignedIntOrFpConstant(ScalarIVTy, Lane)); | ||||||
| 2423 | // The step returned by `createStepForVF` is a runtime-evaluated value | ||||||
| 2424 | // when VF is scalable. Otherwise, it should be folded into a Constant. | ||||||
| 2425 | assert((State.VF.isScalable() || isa<Constant>(StartIdx)) &&(static_cast <bool> ((State.VF.isScalable() || isa<Constant >(StartIdx)) && "Expected StartIdx to be folded to a constant when VF is not " "scalable") ? void (0) : __assert_fail ("(State.VF.isScalable() || isa<Constant>(StartIdx)) && \"Expected StartIdx to be folded to a constant when VF is not \" \"scalable\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2427, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2426 | "Expected StartIdx to be folded to a constant when VF is not "(static_cast <bool> ((State.VF.isScalable() || isa<Constant >(StartIdx)) && "Expected StartIdx to be folded to a constant when VF is not " "scalable") ? void (0) : __assert_fail ("(State.VF.isScalable() || isa<Constant>(StartIdx)) && \"Expected StartIdx to be folded to a constant when VF is not \" \"scalable\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2427, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2427 | "scalable")(static_cast <bool> ((State.VF.isScalable() || isa<Constant >(StartIdx)) && "Expected StartIdx to be folded to a constant when VF is not " "scalable") ? void (0) : __assert_fail ("(State.VF.isScalable() || isa<Constant>(StartIdx)) && \"Expected StartIdx to be folded to a constant when VF is not \" \"scalable\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2427, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2428 | auto *Mul = Builder.CreateBinOp(MulOp, StartIdx, Step); | ||||||
| 2429 | auto *Add = Builder.CreateBinOp(AddOp, ScalarIV, Mul); | ||||||
| 2430 | State.set(Def, Add, VPIteration(Part, Lane)); | ||||||
| 2431 | } | ||||||
| 2432 | } | ||||||
| 2433 | } | ||||||
| 2434 | |||||||
| 2435 | // Generate code for the induction step. Note that induction steps are | ||||||
| 2436 | // required to be loop-invariant | ||||||
| 2437 | static Value *CreateStepValue(const SCEV *Step, ScalarEvolution &SE, | ||||||
| 2438 | Instruction *InsertBefore, | ||||||
| 2439 | Loop *OrigLoop = nullptr) { | ||||||
| 2440 | const DataLayout &DL = SE.getDataLayout(); | ||||||
| 2441 | assert((!OrigLoop || SE.isLoopInvariant(Step, OrigLoop)) &&(static_cast <bool> ((!OrigLoop || SE.isLoopInvariant(Step , OrigLoop)) && "Induction step should be loop invariant" ) ? void (0) : __assert_fail ("(!OrigLoop || SE.isLoopInvariant(Step, OrigLoop)) && \"Induction step should be loop invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2442, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2442 | "Induction step should be loop invariant")(static_cast <bool> ((!OrigLoop || SE.isLoopInvariant(Step , OrigLoop)) && "Induction step should be loop invariant" ) ? void (0) : __assert_fail ("(!OrigLoop || SE.isLoopInvariant(Step, OrigLoop)) && \"Induction step should be loop invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2442, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2443 | if (auto *E = dyn_cast<SCEVUnknown>(Step)) | ||||||
| 2444 | return E->getValue(); | ||||||
| 2445 | |||||||
| 2446 | SCEVExpander Exp(SE, DL, "induction"); | ||||||
| 2447 | return Exp.expandCodeFor(Step, Step->getType(), InsertBefore); | ||||||
| 2448 | } | ||||||
| 2449 | |||||||
| 2450 | /// Compute the transformed value of Index at offset StartValue using step | ||||||
| 2451 | /// StepValue. | ||||||
| 2452 | /// For integer induction, returns StartValue + Index * StepValue. | ||||||
| 2453 | /// For pointer induction, returns StartValue[Index * StepValue]. | ||||||
| 2454 | /// FIXME: The newly created binary instructions should contain nsw/nuw | ||||||
| 2455 | /// flags, which can be found from the original scalar operations. | ||||||
| 2456 | static Value *emitTransformedIndex(IRBuilderBase &B, Value *Index, | ||||||
| 2457 | Value *StartValue, Value *Step, | ||||||
| 2458 | const InductionDescriptor &ID) { | ||||||
| 2459 | Type *StepTy = Step->getType(); | ||||||
| 2460 | Value *CastedIndex = StepTy->isIntegerTy() | ||||||
| 2461 | ? B.CreateSExtOrTrunc(Index, StepTy) | ||||||
| 2462 | : B.CreateCast(Instruction::SIToFP, Index, StepTy); | ||||||
| 2463 | if (CastedIndex != Index) { | ||||||
| 2464 | CastedIndex->setName(CastedIndex->getName() + ".cast"); | ||||||
| 2465 | Index = CastedIndex; | ||||||
| 2466 | } | ||||||
| 2467 | |||||||
| 2468 | // Note: the IR at this point is broken. We cannot use SE to create any new | ||||||
| 2469 | // SCEV and then expand it, hoping that SCEV's simplification will give us | ||||||
| 2470 | // a more optimal code. Unfortunately, attempt of doing so on invalid IR may | ||||||
| 2471 | // lead to various SCEV crashes. So all we can do is to use builder and rely | ||||||
| 2472 | // on InstCombine for future simplifications. Here we handle some trivial | ||||||
| 2473 | // cases only. | ||||||
| 2474 | auto CreateAdd = [&B](Value *X, Value *Y) { | ||||||
| 2475 | assert(X->getType() == Y->getType() && "Types don't match!")(static_cast <bool> (X->getType() == Y->getType() && "Types don't match!") ? void (0) : __assert_fail ( "X->getType() == Y->getType() && \"Types don't match!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2475, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2476 | if (auto *CX = dyn_cast<ConstantInt>(X)) | ||||||
| 2477 | if (CX->isZero()) | ||||||
| 2478 | return Y; | ||||||
| 2479 | if (auto *CY = dyn_cast<ConstantInt>(Y)) | ||||||
| 2480 | if (CY->isZero()) | ||||||
| 2481 | return X; | ||||||
| 2482 | return B.CreateAdd(X, Y); | ||||||
| 2483 | }; | ||||||
| 2484 | |||||||
| 2485 | // We allow X to be a vector type, in which case Y will potentially be | ||||||
| 2486 | // splatted into a vector with the same element count. | ||||||
| 2487 | auto CreateMul = [&B](Value *X, Value *Y) { | ||||||
| 2488 | assert(X->getType()->getScalarType() == Y->getType() &&(static_cast <bool> (X->getType()->getScalarType( ) == Y->getType() && "Types don't match!") ? void ( 0) : __assert_fail ("X->getType()->getScalarType() == Y->getType() && \"Types don't match!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2489, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2489 | "Types don't match!")(static_cast <bool> (X->getType()->getScalarType( ) == Y->getType() && "Types don't match!") ? void ( 0) : __assert_fail ("X->getType()->getScalarType() == Y->getType() && \"Types don't match!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2489, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2490 | if (auto *CX = dyn_cast<ConstantInt>(X)) | ||||||
| 2491 | if (CX->isOne()) | ||||||
| 2492 | return Y; | ||||||
| 2493 | if (auto *CY = dyn_cast<ConstantInt>(Y)) | ||||||
| 2494 | if (CY->isOne()) | ||||||
| 2495 | return X; | ||||||
| 2496 | VectorType *XVTy = dyn_cast<VectorType>(X->getType()); | ||||||
| 2497 | if (XVTy && !isa<VectorType>(Y->getType())) | ||||||
| 2498 | Y = B.CreateVectorSplat(XVTy->getElementCount(), Y); | ||||||
| 2499 | return B.CreateMul(X, Y); | ||||||
| 2500 | }; | ||||||
| 2501 | |||||||
| 2502 | switch (ID.getKind()) { | ||||||
| 2503 | case InductionDescriptor::IK_IntInduction: { | ||||||
| 2504 | assert(!isa<VectorType>(Index->getType()) &&(static_cast <bool> (!isa<VectorType>(Index->getType ()) && "Vector indices not supported for integer inductions yet" ) ? void (0) : __assert_fail ("!isa<VectorType>(Index->getType()) && \"Vector indices not supported for integer inductions yet\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2505, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2505 | "Vector indices not supported for integer inductions yet")(static_cast <bool> (!isa<VectorType>(Index->getType ()) && "Vector indices not supported for integer inductions yet" ) ? void (0) : __assert_fail ("!isa<VectorType>(Index->getType()) && \"Vector indices not supported for integer inductions yet\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2505, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2506 | assert(Index->getType() == StartValue->getType() &&(static_cast <bool> (Index->getType() == StartValue-> getType() && "Index type does not match StartValue type" ) ? void (0) : __assert_fail ("Index->getType() == StartValue->getType() && \"Index type does not match StartValue type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2507, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2507 | "Index type does not match StartValue type")(static_cast <bool> (Index->getType() == StartValue-> getType() && "Index type does not match StartValue type" ) ? void (0) : __assert_fail ("Index->getType() == StartValue->getType() && \"Index type does not match StartValue type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2507, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2508 | if (isa<ConstantInt>(Step) && cast<ConstantInt>(Step)->isMinusOne()) | ||||||
| 2509 | return B.CreateSub(StartValue, Index); | ||||||
| 2510 | auto *Offset = CreateMul(Index, Step); | ||||||
| 2511 | return CreateAdd(StartValue, Offset); | ||||||
| 2512 | } | ||||||
| 2513 | case InductionDescriptor::IK_PtrInduction: { | ||||||
| 2514 | assert(isa<Constant>(Step) &&(static_cast <bool> (isa<Constant>(Step) && "Expected constant step for pointer induction") ? void (0) : __assert_fail ("isa<Constant>(Step) && \"Expected constant step for pointer induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2515, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2515 | "Expected constant step for pointer induction")(static_cast <bool> (isa<Constant>(Step) && "Expected constant step for pointer induction") ? void (0) : __assert_fail ("isa<Constant>(Step) && \"Expected constant step for pointer induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2515, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2516 | return B.CreateGEP(ID.getElementType(), StartValue, CreateMul(Index, Step)); | ||||||
| 2517 | } | ||||||
| 2518 | case InductionDescriptor::IK_FpInduction: { | ||||||
| 2519 | assert(!isa<VectorType>(Index->getType()) &&(static_cast <bool> (!isa<VectorType>(Index->getType ()) && "Vector indices not supported for FP inductions yet" ) ? void (0) : __assert_fail ("!isa<VectorType>(Index->getType()) && \"Vector indices not supported for FP inductions yet\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2520, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2520 | "Vector indices not supported for FP inductions yet")(static_cast <bool> (!isa<VectorType>(Index->getType ()) && "Vector indices not supported for FP inductions yet" ) ? void (0) : __assert_fail ("!isa<VectorType>(Index->getType()) && \"Vector indices not supported for FP inductions yet\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2520, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2521 | assert(Step->getType()->isFloatingPointTy() && "Expected FP Step value")(static_cast <bool> (Step->getType()->isFloatingPointTy () && "Expected FP Step value") ? void (0) : __assert_fail ("Step->getType()->isFloatingPointTy() && \"Expected FP Step value\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2521, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2522 | auto InductionBinOp = ID.getInductionBinOp(); | ||||||
| 2523 | assert(InductionBinOp &&(static_cast <bool> (InductionBinOp && (InductionBinOp ->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode () == Instruction::FSub) && "Original bin op should be defined for FP induction" ) ? void (0) : __assert_fail ("InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && \"Original bin op should be defined for FP induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2526, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2524 | (InductionBinOp->getOpcode() == Instruction::FAdd ||(static_cast <bool> (InductionBinOp && (InductionBinOp ->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode () == Instruction::FSub) && "Original bin op should be defined for FP induction" ) ? void (0) : __assert_fail ("InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && \"Original bin op should be defined for FP induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2526, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2525 | InductionBinOp->getOpcode() == Instruction::FSub) &&(static_cast <bool> (InductionBinOp && (InductionBinOp ->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode () == Instruction::FSub) && "Original bin op should be defined for FP induction" ) ? void (0) : __assert_fail ("InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && \"Original bin op should be defined for FP induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2526, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2526 | "Original bin op should be defined for FP induction")(static_cast <bool> (InductionBinOp && (InductionBinOp ->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode () == Instruction::FSub) && "Original bin op should be defined for FP induction" ) ? void (0) : __assert_fail ("InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && \"Original bin op should be defined for FP induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2526, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2527 | |||||||
| 2528 | Value *MulExp = B.CreateFMul(Step, Index); | ||||||
| 2529 | return B.CreateBinOp(InductionBinOp->getOpcode(), StartValue, MulExp, | ||||||
| 2530 | "induction"); | ||||||
| 2531 | } | ||||||
| 2532 | case InductionDescriptor::IK_NoInduction: | ||||||
| 2533 | return nullptr; | ||||||
| 2534 | } | ||||||
| 2535 | llvm_unreachable("invalid enum")::llvm::llvm_unreachable_internal("invalid enum", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2535); | ||||||
| 2536 | } | ||||||
| 2537 | |||||||
| 2538 | void InnerLoopVectorizer::packScalarIntoVectorValue(VPValue *Def, | ||||||
| 2539 | const VPIteration &Instance, | ||||||
| 2540 | VPTransformState &State) { | ||||||
| 2541 | Value *ScalarInst = State.get(Def, Instance); | ||||||
| 2542 | Value *VectorValue = State.get(Def, Instance.Part); | ||||||
| 2543 | VectorValue = Builder.CreateInsertElement( | ||||||
| 2544 | VectorValue, ScalarInst, | ||||||
| 2545 | Instance.Lane.getAsRuntimeExpr(State.Builder, VF)); | ||||||
| 2546 | State.set(Def, VectorValue, Instance.Part); | ||||||
| 2547 | } | ||||||
| 2548 | |||||||
| 2549 | // Return whether we allow using masked interleave-groups (for dealing with | ||||||
| 2550 | // strided loads/stores that reside in predicated blocks, or for dealing | ||||||
| 2551 | // with gaps). | ||||||
| 2552 | static bool useMaskedInterleavedAccesses(const TargetTransformInfo &TTI) { | ||||||
| 2553 | // If an override option has been passed in for interleaved accesses, use it. | ||||||
| 2554 | if (EnableMaskedInterleavedMemAccesses.getNumOccurrences() > 0) | ||||||
| 2555 | return EnableMaskedInterleavedMemAccesses; | ||||||
| 2556 | |||||||
| 2557 | return TTI.enableMaskedInterleavedAccessVectorization(); | ||||||
| 2558 | } | ||||||
| 2559 | |||||||
| 2560 | // Try to vectorize the interleave group that \p Instr belongs to. | ||||||
| 2561 | // | ||||||
| 2562 | // E.g. Translate following interleaved load group (factor = 3): | ||||||
| 2563 | // for (i = 0; i < N; i+=3) { | ||||||
| 2564 | // R = Pic[i]; // Member of index 0 | ||||||
| 2565 | // G = Pic[i+1]; // Member of index 1 | ||||||
| 2566 | // B = Pic[i+2]; // Member of index 2 | ||||||
| 2567 | // ... // do something to R, G, B | ||||||
| 2568 | // } | ||||||
| 2569 | // To: | ||||||
| 2570 | // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B | ||||||
| 2571 | // %R.vec = shuffle %wide.vec, poison, <0, 3, 6, 9> ; R elements | ||||||
| 2572 | // %G.vec = shuffle %wide.vec, poison, <1, 4, 7, 10> ; G elements | ||||||
| 2573 | // %B.vec = shuffle %wide.vec, poison, <2, 5, 8, 11> ; B elements | ||||||
| 2574 | // | ||||||
| 2575 | // Or translate following interleaved store group (factor = 3): | ||||||
| 2576 | // for (i = 0; i < N; i+=3) { | ||||||
| 2577 | // ... do something to R, G, B | ||||||
| 2578 | // Pic[i] = R; // Member of index 0 | ||||||
| 2579 | // Pic[i+1] = G; // Member of index 1 | ||||||
| 2580 | // Pic[i+2] = B; // Member of index 2 | ||||||
| 2581 | // } | ||||||
| 2582 | // To: | ||||||
| 2583 | // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7> | ||||||
| 2584 | // %B_U.vec = shuffle %B.vec, poison, <0, 1, 2, 3, u, u, u, u> | ||||||
| 2585 | // %interleaved.vec = shuffle %R_G.vec, %B_U.vec, | ||||||
| 2586 | // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements | ||||||
| 2587 | // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B | ||||||
| 2588 | void InnerLoopVectorizer::vectorizeInterleaveGroup( | ||||||
| 2589 | const InterleaveGroup<Instruction> *Group, ArrayRef<VPValue *> VPDefs, | ||||||
| 2590 | VPTransformState &State, VPValue *Addr, ArrayRef<VPValue *> StoredValues, | ||||||
| 2591 | VPValue *BlockInMask) { | ||||||
| 2592 | Instruction *Instr = Group->getInsertPos(); | ||||||
| 2593 | const DataLayout &DL = Instr->getModule()->getDataLayout(); | ||||||
| 2594 | |||||||
| 2595 | // Prepare for the vector type of the interleaved load/store. | ||||||
| 2596 | Type *ScalarTy = getLoadStoreType(Instr); | ||||||
| 2597 | unsigned InterleaveFactor = Group->getFactor(); | ||||||
| 2598 | assert(!VF.isScalable() && "scalable vectors not yet supported.")(static_cast <bool> (!VF.isScalable() && "scalable vectors not yet supported." ) ? void (0) : __assert_fail ("!VF.isScalable() && \"scalable vectors not yet supported.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2598, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2599 | auto *VecTy = VectorType::get(ScalarTy, VF * InterleaveFactor); | ||||||
| 2600 | |||||||
| 2601 | // Prepare for the new pointers. | ||||||
| 2602 | SmallVector<Value *, 2> AddrParts; | ||||||
| 2603 | unsigned Index = Group->getIndex(Instr); | ||||||
| 2604 | |||||||
| 2605 | // TODO: extend the masked interleaved-group support to reversed access. | ||||||
| 2606 | assert((!BlockInMask || !Group->isReverse()) &&(static_cast <bool> ((!BlockInMask || !Group->isReverse ()) && "Reversed masked interleave-group not supported." ) ? void (0) : __assert_fail ("(!BlockInMask || !Group->isReverse()) && \"Reversed masked interleave-group not supported.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2607, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2607 | "Reversed masked interleave-group not supported.")(static_cast <bool> ((!BlockInMask || !Group->isReverse ()) && "Reversed masked interleave-group not supported." ) ? void (0) : __assert_fail ("(!BlockInMask || !Group->isReverse()) && \"Reversed masked interleave-group not supported.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2607, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2608 | |||||||
| 2609 | // If the group is reverse, adjust the index to refer to the last vector lane | ||||||
| 2610 | // instead of the first. We adjust the index from the first vector lane, | ||||||
| 2611 | // rather than directly getting the pointer for lane VF - 1, because the | ||||||
| 2612 | // pointer operand of the interleaved access is supposed to be uniform. For | ||||||
| 2613 | // uniform instructions, we're only required to generate a value for the | ||||||
| 2614 | // first vector lane in each unroll iteration. | ||||||
| 2615 | if (Group->isReverse()) | ||||||
| 2616 | Index += (VF.getKnownMinValue() - 1) * Group->getFactor(); | ||||||
| 2617 | |||||||
| 2618 | for (unsigned Part = 0; Part < UF; Part++) { | ||||||
| 2619 | Value *AddrPart = State.get(Addr, VPIteration(Part, 0)); | ||||||
| 2620 | State.setDebugLocFromInst(AddrPart); | ||||||
| 2621 | |||||||
| 2622 | // Notice current instruction could be any index. Need to adjust the address | ||||||
| 2623 | // to the member of index 0. | ||||||
| 2624 | // | ||||||
| 2625 | // E.g. a = A[i+1]; // Member of index 1 (Current instruction) | ||||||
| 2626 | // b = A[i]; // Member of index 0 | ||||||
| 2627 | // Current pointer is pointed to A[i+1], adjust it to A[i]. | ||||||
| 2628 | // | ||||||
| 2629 | // E.g. A[i+1] = a; // Member of index 1 | ||||||
| 2630 | // A[i] = b; // Member of index 0 | ||||||
| 2631 | // A[i+2] = c; // Member of index 2 (Current instruction) | ||||||
| 2632 | // Current pointer is pointed to A[i+2], adjust it to A[i]. | ||||||
| 2633 | |||||||
| 2634 | bool InBounds = false; | ||||||
| 2635 | if (auto *gep = dyn_cast<GetElementPtrInst>(AddrPart->stripPointerCasts())) | ||||||
| 2636 | InBounds = gep->isInBounds(); | ||||||
| 2637 | AddrPart = Builder.CreateGEP(ScalarTy, AddrPart, Builder.getInt32(-Index)); | ||||||
| 2638 | cast<GetElementPtrInst>(AddrPart)->setIsInBounds(InBounds); | ||||||
| 2639 | |||||||
| 2640 | // Cast to the vector pointer type. | ||||||
| 2641 | unsigned AddressSpace = AddrPart->getType()->getPointerAddressSpace(); | ||||||
| 2642 | Type *PtrTy = VecTy->getPointerTo(AddressSpace); | ||||||
| 2643 | AddrParts.push_back(Builder.CreateBitCast(AddrPart, PtrTy)); | ||||||
| 2644 | } | ||||||
| 2645 | |||||||
| 2646 | State.setDebugLocFromInst(Instr); | ||||||
| 2647 | Value *PoisonVec = PoisonValue::get(VecTy); | ||||||
| 2648 | |||||||
| 2649 | Value *MaskForGaps = nullptr; | ||||||
| 2650 | if (Group->requiresScalarEpilogue() && !Cost->isScalarEpilogueAllowed()) { | ||||||
| 2651 | MaskForGaps = createBitMaskForGaps(Builder, VF.getKnownMinValue(), *Group); | ||||||
| 2652 | assert(MaskForGaps && "Mask for Gaps is required but it is null")(static_cast <bool> (MaskForGaps && "Mask for Gaps is required but it is null" ) ? void (0) : __assert_fail ("MaskForGaps && \"Mask for Gaps is required but it is null\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2652, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2653 | } | ||||||
| 2654 | |||||||
| 2655 | // Vectorize the interleaved load group. | ||||||
| 2656 | if (isa<LoadInst>(Instr)) { | ||||||
| 2657 | // For each unroll part, create a wide load for the group. | ||||||
| 2658 | SmallVector<Value *, 2> NewLoads; | ||||||
| 2659 | for (unsigned Part = 0; Part < UF; Part++) { | ||||||
| 2660 | Instruction *NewLoad; | ||||||
| 2661 | if (BlockInMask || MaskForGaps) { | ||||||
| 2662 | assert(useMaskedInterleavedAccesses(*TTI) &&(static_cast <bool> (useMaskedInterleavedAccesses(*TTI) && "masked interleaved groups are not allowed.") ? void (0) : __assert_fail ("useMaskedInterleavedAccesses(*TTI) && \"masked interleaved groups are not allowed.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2663, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2663 | "masked interleaved groups are not allowed.")(static_cast <bool> (useMaskedInterleavedAccesses(*TTI) && "masked interleaved groups are not allowed.") ? void (0) : __assert_fail ("useMaskedInterleavedAccesses(*TTI) && \"masked interleaved groups are not allowed.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2663, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2664 | Value *GroupMask = MaskForGaps; | ||||||
| 2665 | if (BlockInMask) { | ||||||
| 2666 | Value *BlockInMaskPart = State.get(BlockInMask, Part); | ||||||
| 2667 | Value *ShuffledMask = Builder.CreateShuffleVector( | ||||||
| 2668 | BlockInMaskPart, | ||||||
| 2669 | createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()), | ||||||
| 2670 | "interleaved.mask"); | ||||||
| 2671 | GroupMask = MaskForGaps | ||||||
| 2672 | ? Builder.CreateBinOp(Instruction::And, ShuffledMask, | ||||||
| 2673 | MaskForGaps) | ||||||
| 2674 | : ShuffledMask; | ||||||
| 2675 | } | ||||||
| 2676 | NewLoad = | ||||||
| 2677 | Builder.CreateMaskedLoad(VecTy, AddrParts[Part], Group->getAlign(), | ||||||
| 2678 | GroupMask, PoisonVec, "wide.masked.vec"); | ||||||
| 2679 | } | ||||||
| 2680 | else | ||||||
| 2681 | NewLoad = Builder.CreateAlignedLoad(VecTy, AddrParts[Part], | ||||||
| 2682 | Group->getAlign(), "wide.vec"); | ||||||
| 2683 | Group->addMetadata(NewLoad); | ||||||
| 2684 | NewLoads.push_back(NewLoad); | ||||||
| 2685 | } | ||||||
| 2686 | |||||||
| 2687 | // For each member in the group, shuffle out the appropriate data from the | ||||||
| 2688 | // wide loads. | ||||||
| 2689 | unsigned J = 0; | ||||||
| 2690 | for (unsigned I = 0; I < InterleaveFactor; ++I) { | ||||||
| 2691 | Instruction *Member = Group->getMember(I); | ||||||
| 2692 | |||||||
| 2693 | // Skip the gaps in the group. | ||||||
| 2694 | if (!Member) | ||||||
| 2695 | continue; | ||||||
| 2696 | |||||||
| 2697 | auto StrideMask = | ||||||
| 2698 | createStrideMask(I, InterleaveFactor, VF.getKnownMinValue()); | ||||||
| 2699 | for (unsigned Part = 0; Part < UF; Part++) { | ||||||
| 2700 | Value *StridedVec = Builder.CreateShuffleVector( | ||||||
| 2701 | NewLoads[Part], StrideMask, "strided.vec"); | ||||||
| 2702 | |||||||
| 2703 | // If this member has different type, cast the result type. | ||||||
| 2704 | if (Member->getType() != ScalarTy) { | ||||||
| 2705 | assert(!VF.isScalable() && "VF is assumed to be non scalable.")(static_cast <bool> (!VF.isScalable() && "VF is assumed to be non scalable." ) ? void (0) : __assert_fail ("!VF.isScalable() && \"VF is assumed to be non scalable.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2705, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2706 | VectorType *OtherVTy = VectorType::get(Member->getType(), VF); | ||||||
| 2707 | StridedVec = createBitOrPointerCast(StridedVec, OtherVTy, DL); | ||||||
| 2708 | } | ||||||
| 2709 | |||||||
| 2710 | if (Group->isReverse()) | ||||||
| 2711 | StridedVec = Builder.CreateVectorReverse(StridedVec, "reverse"); | ||||||
| 2712 | |||||||
| 2713 | State.set(VPDefs[J], StridedVec, Part); | ||||||
| 2714 | } | ||||||
| 2715 | ++J; | ||||||
| 2716 | } | ||||||
| 2717 | return; | ||||||
| 2718 | } | ||||||
| 2719 | |||||||
| 2720 | // The sub vector type for current instruction. | ||||||
| 2721 | auto *SubVT = VectorType::get(ScalarTy, VF); | ||||||
| 2722 | |||||||
| 2723 | // Vectorize the interleaved store group. | ||||||
| 2724 | MaskForGaps = createBitMaskForGaps(Builder, VF.getKnownMinValue(), *Group); | ||||||
| 2725 | assert((!MaskForGaps || useMaskedInterleavedAccesses(*TTI)) &&(static_cast <bool> ((!MaskForGaps || useMaskedInterleavedAccesses (*TTI)) && "masked interleaved groups are not allowed." ) ? void (0) : __assert_fail ("(!MaskForGaps || useMaskedInterleavedAccesses(*TTI)) && \"masked interleaved groups are not allowed.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2726, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2726 | "masked interleaved groups are not allowed.")(static_cast <bool> ((!MaskForGaps || useMaskedInterleavedAccesses (*TTI)) && "masked interleaved groups are not allowed." ) ? void (0) : __assert_fail ("(!MaskForGaps || useMaskedInterleavedAccesses(*TTI)) && \"masked interleaved groups are not allowed.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2726, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2727 | assert((!MaskForGaps || !VF.isScalable()) &&(static_cast <bool> ((!MaskForGaps || !VF.isScalable()) && "masking gaps for scalable vectors is not yet supported." ) ? void (0) : __assert_fail ("(!MaskForGaps || !VF.isScalable()) && \"masking gaps for scalable vectors is not yet supported.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2728, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2728 | "masking gaps for scalable vectors is not yet supported.")(static_cast <bool> ((!MaskForGaps || !VF.isScalable()) && "masking gaps for scalable vectors is not yet supported." ) ? void (0) : __assert_fail ("(!MaskForGaps || !VF.isScalable()) && \"masking gaps for scalable vectors is not yet supported.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2728, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2729 | for (unsigned Part = 0; Part < UF; Part++) { | ||||||
| 2730 | // Collect the stored vector from each member. | ||||||
| 2731 | SmallVector<Value *, 4> StoredVecs; | ||||||
| 2732 | unsigned StoredIdx = 0; | ||||||
| 2733 | for (unsigned i = 0; i < InterleaveFactor; i++) { | ||||||
| 2734 | assert((Group->getMember(i) || MaskForGaps) &&(static_cast <bool> ((Group->getMember(i) || MaskForGaps ) && "Fail to get a member from an interleaved store group" ) ? void (0) : __assert_fail ("(Group->getMember(i) || MaskForGaps) && \"Fail to get a member from an interleaved store group\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2735, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2735 | "Fail to get a member from an interleaved store group")(static_cast <bool> ((Group->getMember(i) || MaskForGaps ) && "Fail to get a member from an interleaved store group" ) ? void (0) : __assert_fail ("(Group->getMember(i) || MaskForGaps) && \"Fail to get a member from an interleaved store group\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2735, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2736 | Instruction *Member = Group->getMember(i); | ||||||
| 2737 | |||||||
| 2738 | // Skip the gaps in the group. | ||||||
| 2739 | if (!Member) { | ||||||
| 2740 | Value *Undef = PoisonValue::get(SubVT); | ||||||
| 2741 | StoredVecs.push_back(Undef); | ||||||
| 2742 | continue; | ||||||
| 2743 | } | ||||||
| 2744 | |||||||
| 2745 | Value *StoredVec = State.get(StoredValues[StoredIdx], Part); | ||||||
| 2746 | ++StoredIdx; | ||||||
| 2747 | |||||||
| 2748 | if (Group->isReverse()) | ||||||
| 2749 | StoredVec = Builder.CreateVectorReverse(StoredVec, "reverse"); | ||||||
| 2750 | |||||||
| 2751 | // If this member has different type, cast it to a unified type. | ||||||
| 2752 | |||||||
| 2753 | if (StoredVec->getType() != SubVT) | ||||||
| 2754 | StoredVec = createBitOrPointerCast(StoredVec, SubVT, DL); | ||||||
| 2755 | |||||||
| 2756 | StoredVecs.push_back(StoredVec); | ||||||
| 2757 | } | ||||||
| 2758 | |||||||
| 2759 | // Concatenate all vectors into a wide vector. | ||||||
| 2760 | Value *WideVec = concatenateVectors(Builder, StoredVecs); | ||||||
| 2761 | |||||||
| 2762 | // Interleave the elements in the wide vector. | ||||||
| 2763 | Value *IVec = Builder.CreateShuffleVector( | ||||||
| 2764 | WideVec, createInterleaveMask(VF.getKnownMinValue(), InterleaveFactor), | ||||||
| 2765 | "interleaved.vec"); | ||||||
| 2766 | |||||||
| 2767 | Instruction *NewStoreInstr; | ||||||
| 2768 | if (BlockInMask || MaskForGaps) { | ||||||
| 2769 | Value *GroupMask = MaskForGaps; | ||||||
| 2770 | if (BlockInMask) { | ||||||
| 2771 | Value *BlockInMaskPart = State.get(BlockInMask, Part); | ||||||
| 2772 | Value *ShuffledMask = Builder.CreateShuffleVector( | ||||||
| 2773 | BlockInMaskPart, | ||||||
| 2774 | createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()), | ||||||
| 2775 | "interleaved.mask"); | ||||||
| 2776 | GroupMask = MaskForGaps ? Builder.CreateBinOp(Instruction::And, | ||||||
| 2777 | ShuffledMask, MaskForGaps) | ||||||
| 2778 | : ShuffledMask; | ||||||
| 2779 | } | ||||||
| 2780 | NewStoreInstr = Builder.CreateMaskedStore(IVec, AddrParts[Part], | ||||||
| 2781 | Group->getAlign(), GroupMask); | ||||||
| 2782 | } else | ||||||
| 2783 | NewStoreInstr = | ||||||
| 2784 | Builder.CreateAlignedStore(IVec, AddrParts[Part], Group->getAlign()); | ||||||
| 2785 | |||||||
| 2786 | Group->addMetadata(NewStoreInstr); | ||||||
| 2787 | } | ||||||
| 2788 | } | ||||||
| 2789 | |||||||
| 2790 | void InnerLoopVectorizer::scalarizeInstruction(const Instruction *Instr, | ||||||
| 2791 | VPReplicateRecipe *RepRecipe, | ||||||
| 2792 | const VPIteration &Instance, | ||||||
| 2793 | bool IfPredicateInstr, | ||||||
| 2794 | VPTransformState &State) { | ||||||
| 2795 | assert(!Instr->getType()->isAggregateType() && "Can't handle vectors")(static_cast <bool> (!Instr->getType()->isAggregateType () && "Can't handle vectors") ? void (0) : __assert_fail ("!Instr->getType()->isAggregateType() && \"Can't handle vectors\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2795, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2796 | |||||||
| 2797 | // llvm.experimental.noalias.scope.decl intrinsics must only be duplicated for | ||||||
| 2798 | // the first lane and part. | ||||||
| 2799 | if (isa<NoAliasScopeDeclInst>(Instr)) | ||||||
| 2800 | if (!Instance.isFirstIteration()) | ||||||
| 2801 | return; | ||||||
| 2802 | |||||||
| 2803 | // Does this instruction return a value ? | ||||||
| 2804 | bool IsVoidRetTy = Instr->getType()->isVoidTy(); | ||||||
| 2805 | |||||||
| 2806 | Instruction *Cloned = Instr->clone(); | ||||||
| 2807 | if (!IsVoidRetTy) | ||||||
| 2808 | Cloned->setName(Instr->getName() + ".cloned"); | ||||||
| 2809 | |||||||
| 2810 | // If the scalarized instruction contributes to the address computation of a | ||||||
| 2811 | // widen masked load/store which was in a basic block that needed predication | ||||||
| 2812 | // and is not predicated after vectorization, we can't propagate | ||||||
| 2813 | // poison-generating flags (nuw/nsw, exact, inbounds, etc.). The scalarized | ||||||
| 2814 | // instruction could feed a poison value to the base address of the widen | ||||||
| 2815 | // load/store. | ||||||
| 2816 | if (State.MayGeneratePoisonRecipes.contains(RepRecipe)) | ||||||
| 2817 | Cloned->dropPoisonGeneratingFlags(); | ||||||
| 2818 | |||||||
| 2819 | if (Instr->getDebugLoc()) | ||||||
| 2820 | State.setDebugLocFromInst(Instr); | ||||||
| 2821 | |||||||
| 2822 | // Replace the operands of the cloned instructions with their scalar | ||||||
| 2823 | // equivalents in the new loop. | ||||||
| 2824 | for (const auto &I : enumerate(RepRecipe->operands())) { | ||||||
| 2825 | auto InputInstance = Instance; | ||||||
| 2826 | VPValue *Operand = I.value(); | ||||||
| 2827 | if (vputils::isUniformAfterVectorization(Operand)) | ||||||
| 2828 | InputInstance.Lane = VPLane::getFirstLane(); | ||||||
| 2829 | Cloned->setOperand(I.index(), State.get(Operand, InputInstance)); | ||||||
| 2830 | } | ||||||
| 2831 | State.addNewMetadata(Cloned, Instr); | ||||||
| 2832 | |||||||
| 2833 | // Place the cloned scalar in the new loop. | ||||||
| 2834 | State.Builder.Insert(Cloned); | ||||||
| 2835 | |||||||
| 2836 | State.set(RepRecipe, Cloned, Instance); | ||||||
| 2837 | |||||||
| 2838 | // If we just cloned a new assumption, add it the assumption cache. | ||||||
| 2839 | if (auto *II = dyn_cast<AssumeInst>(Cloned)) | ||||||
| 2840 | AC->registerAssumption(II); | ||||||
| 2841 | |||||||
| 2842 | // End if-block. | ||||||
| 2843 | if (IfPredicateInstr) | ||||||
| 2844 | PredicatedInstructions.push_back(Cloned); | ||||||
| 2845 | } | ||||||
| 2846 | |||||||
| 2847 | Value *InnerLoopVectorizer::getOrCreateTripCount(BasicBlock *InsertBlock) { | ||||||
| 2848 | if (TripCount) | ||||||
| 2849 | return TripCount; | ||||||
| 2850 | |||||||
| 2851 | assert(InsertBlock)(static_cast <bool> (InsertBlock) ? void (0) : __assert_fail ("InsertBlock", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 2851, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2852 | IRBuilder<> Builder(InsertBlock->getTerminator()); | ||||||
| 2853 | // Find the loop boundaries. | ||||||
| 2854 | Type *IdxTy = Legal->getWidestInductionType(); | ||||||
| 2855 | assert(IdxTy && "No type for induction")(static_cast <bool> (IdxTy && "No type for induction" ) ? void (0) : __assert_fail ("IdxTy && \"No type for induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2855, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2856 | const SCEV *ExitCount = createTripCountSCEV(IdxTy, PSE); | ||||||
| 2857 | |||||||
| 2858 | const DataLayout &DL = InsertBlock->getModule()->getDataLayout(); | ||||||
| 2859 | |||||||
| 2860 | // Expand the trip count and place the new instructions in the preheader. | ||||||
| 2861 | // Notice that the pre-header does not change, only the loop body. | ||||||
| 2862 | SCEVExpander Exp(*PSE.getSE(), DL, "induction"); | ||||||
| 2863 | |||||||
| 2864 | // Count holds the overall loop count (N). | ||||||
| 2865 | TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(), | ||||||
| 2866 | InsertBlock->getTerminator()); | ||||||
| 2867 | |||||||
| 2868 | if (TripCount->getType()->isPointerTy()) | ||||||
| 2869 | TripCount = | ||||||
| 2870 | CastInst::CreatePointerCast(TripCount, IdxTy, "exitcount.ptrcnt.to.int", | ||||||
| 2871 | InsertBlock->getTerminator()); | ||||||
| 2872 | |||||||
| 2873 | return TripCount; | ||||||
| 2874 | } | ||||||
| 2875 | |||||||
| 2876 | Value * | ||||||
| 2877 | InnerLoopVectorizer::getOrCreateVectorTripCount(BasicBlock *InsertBlock) { | ||||||
| 2878 | if (VectorTripCount) | ||||||
| 2879 | return VectorTripCount; | ||||||
| 2880 | |||||||
| 2881 | Value *TC = getOrCreateTripCount(InsertBlock); | ||||||
| 2882 | IRBuilder<> Builder(InsertBlock->getTerminator()); | ||||||
| 2883 | |||||||
| 2884 | Type *Ty = TC->getType(); | ||||||
| 2885 | // This is where we can make the step a runtime constant. | ||||||
| 2886 | Value *Step = createStepForVF(Builder, Ty, VF, UF); | ||||||
| 2887 | |||||||
| 2888 | // If the tail is to be folded by masking, round the number of iterations N | ||||||
| 2889 | // up to a multiple of Step instead of rounding down. This is done by first | ||||||
| 2890 | // adding Step-1 and then rounding down. Note that it's ok if this addition | ||||||
| 2891 | // overflows: the vector induction variable will eventually wrap to zero given | ||||||
| 2892 | // that it starts at zero and its Step is a power of two; the loop will then | ||||||
| 2893 | // exit, with the last early-exit vector comparison also producing all-true. | ||||||
| 2894 | // For scalable vectors the VF is not guaranteed to be a power of 2, but this | ||||||
| 2895 | // is accounted for in emitIterationCountCheck that adds an overflow check. | ||||||
| 2896 | if (Cost->foldTailByMasking()) { | ||||||
| 2897 | assert(isPowerOf2_32(VF.getKnownMinValue() * UF) &&(static_cast <bool> (isPowerOf2_32(VF.getKnownMinValue( ) * UF) && "VF*UF must be a power of 2 when folding tail by masking" ) ? void (0) : __assert_fail ("isPowerOf2_32(VF.getKnownMinValue() * UF) && \"VF*UF must be a power of 2 when folding tail by masking\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2898, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2898 | "VF*UF must be a power of 2 when folding tail by masking")(static_cast <bool> (isPowerOf2_32(VF.getKnownMinValue( ) * UF) && "VF*UF must be a power of 2 when folding tail by masking" ) ? void (0) : __assert_fail ("isPowerOf2_32(VF.getKnownMinValue() * UF) && \"VF*UF must be a power of 2 when folding tail by masking\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2898, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2899 | Value *NumLanes = getRuntimeVF(Builder, Ty, VF * UF); | ||||||
| 2900 | TC = Builder.CreateAdd( | ||||||
| 2901 | TC, Builder.CreateSub(NumLanes, ConstantInt::get(Ty, 1)), "n.rnd.up"); | ||||||
| 2902 | } | ||||||
| 2903 | |||||||
| 2904 | // Now we need to generate the expression for the part of the loop that the | ||||||
| 2905 | // vectorized body will execute. This is equal to N - (N % Step) if scalar | ||||||
| 2906 | // iterations are not required for correctness, or N - Step, otherwise. Step | ||||||
| 2907 | // is equal to the vectorization factor (number of SIMD elements) times the | ||||||
| 2908 | // unroll factor (number of SIMD instructions). | ||||||
| 2909 | Value *R = Builder.CreateURem(TC, Step, "n.mod.vf"); | ||||||
| 2910 | |||||||
| 2911 | // There are cases where we *must* run at least one iteration in the remainder | ||||||
| 2912 | // loop. See the cost model for when this can happen. If the step evenly | ||||||
| 2913 | // divides the trip count, we set the remainder to be equal to the step. If | ||||||
| 2914 | // the step does not evenly divide the trip count, no adjustment is necessary | ||||||
| 2915 | // since there will already be scalar iterations. Note that the minimum | ||||||
| 2916 | // iterations check ensures that N >= Step. | ||||||
| 2917 | if (Cost->requiresScalarEpilogue(VF)) { | ||||||
| 2918 | auto *IsZero = Builder.CreateICmpEQ(R, ConstantInt::get(R->getType(), 0)); | ||||||
| 2919 | R = Builder.CreateSelect(IsZero, Step, R); | ||||||
| 2920 | } | ||||||
| 2921 | |||||||
| 2922 | VectorTripCount = Builder.CreateSub(TC, R, "n.vec"); | ||||||
| 2923 | |||||||
| 2924 | return VectorTripCount; | ||||||
| 2925 | } | ||||||
| 2926 | |||||||
| 2927 | Value *InnerLoopVectorizer::createBitOrPointerCast(Value *V, VectorType *DstVTy, | ||||||
| 2928 | const DataLayout &DL) { | ||||||
| 2929 | // Verify that V is a vector type with same number of elements as DstVTy. | ||||||
| 2930 | auto *DstFVTy = cast<FixedVectorType>(DstVTy); | ||||||
| 2931 | unsigned VF = DstFVTy->getNumElements(); | ||||||
| 2932 | auto *SrcVecTy = cast<FixedVectorType>(V->getType()); | ||||||
| 2933 | assert((VF == SrcVecTy->getNumElements()) && "Vector dimensions do not match")(static_cast <bool> ((VF == SrcVecTy->getNumElements ()) && "Vector dimensions do not match") ? void (0) : __assert_fail ("(VF == SrcVecTy->getNumElements()) && \"Vector dimensions do not match\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2933, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2934 | Type *SrcElemTy = SrcVecTy->getElementType(); | ||||||
| 2935 | Type *DstElemTy = DstFVTy->getElementType(); | ||||||
| 2936 | assert((DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) &&(static_cast <bool> ((DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) && "Vector elements must have same size" ) ? void (0) : __assert_fail ("(DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) && \"Vector elements must have same size\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2937, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2937 | "Vector elements must have same size")(static_cast <bool> ((DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) && "Vector elements must have same size" ) ? void (0) : __assert_fail ("(DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) && \"Vector elements must have same size\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2937, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2938 | |||||||
| 2939 | // Do a direct cast if element types are castable. | ||||||
| 2940 | if (CastInst::isBitOrNoopPointerCastable(SrcElemTy, DstElemTy, DL)) { | ||||||
| 2941 | return Builder.CreateBitOrPointerCast(V, DstFVTy); | ||||||
| 2942 | } | ||||||
| 2943 | // V cannot be directly casted to desired vector type. | ||||||
| 2944 | // May happen when V is a floating point vector but DstVTy is a vector of | ||||||
| 2945 | // pointers or vice-versa. Handle this using a two-step bitcast using an | ||||||
| 2946 | // intermediate Integer type for the bitcast i.e. Ptr <-> Int <-> Float. | ||||||
| 2947 | assert((DstElemTy->isPointerTy() != SrcElemTy->isPointerTy()) &&(static_cast <bool> ((DstElemTy->isPointerTy() != SrcElemTy ->isPointerTy()) && "Only one type should be a pointer type" ) ? void (0) : __assert_fail ("(DstElemTy->isPointerTy() != SrcElemTy->isPointerTy()) && \"Only one type should be a pointer type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2948, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2948 | "Only one type should be a pointer type")(static_cast <bool> ((DstElemTy->isPointerTy() != SrcElemTy ->isPointerTy()) && "Only one type should be a pointer type" ) ? void (0) : __assert_fail ("(DstElemTy->isPointerTy() != SrcElemTy->isPointerTy()) && \"Only one type should be a pointer type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2948, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2949 | assert((DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) &&(static_cast <bool> ((DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) && "Only one type should be a floating point type" ) ? void (0) : __assert_fail ("(DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) && \"Only one type should be a floating point type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2950, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 2950 | "Only one type should be a floating point type")(static_cast <bool> ((DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) && "Only one type should be a floating point type" ) ? void (0) : __assert_fail ("(DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) && \"Only one type should be a floating point type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2950, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 2951 | Type *IntTy = | ||||||
| 2952 | IntegerType::getIntNTy(V->getContext(), DL.getTypeSizeInBits(SrcElemTy)); | ||||||
| 2953 | auto *VecIntTy = FixedVectorType::get(IntTy, VF); | ||||||
| 2954 | Value *CastVal = Builder.CreateBitOrPointerCast(V, VecIntTy); | ||||||
| 2955 | return Builder.CreateBitOrPointerCast(CastVal, DstFVTy); | ||||||
| 2956 | } | ||||||
| 2957 | |||||||
| 2958 | void InnerLoopVectorizer::emitIterationCountCheck(BasicBlock *Bypass) { | ||||||
| 2959 | Value *Count = getOrCreateTripCount(LoopVectorPreHeader); | ||||||
| 2960 | // Reuse existing vector loop preheader for TC checks. | ||||||
| 2961 | // Note that new preheader block is generated for vector loop. | ||||||
| 2962 | BasicBlock *const TCCheckBlock = LoopVectorPreHeader; | ||||||
| 2963 | IRBuilder<> Builder(TCCheckBlock->getTerminator()); | ||||||
| 2964 | |||||||
| 2965 | // Generate code to check if the loop's trip count is less than VF * UF, or | ||||||
| 2966 | // equal to it in case a scalar epilogue is required; this implies that the | ||||||
| 2967 | // vector trip count is zero. This check also covers the case where adding one | ||||||
| 2968 | // to the backedge-taken count overflowed leading to an incorrect trip count | ||||||
| 2969 | // of zero. In this case we will also jump to the scalar loop. | ||||||
| 2970 | auto P = Cost->requiresScalarEpilogue(VF) ? ICmpInst::ICMP_ULE | ||||||
| 2971 | : ICmpInst::ICMP_ULT; | ||||||
| 2972 | |||||||
| 2973 | // If tail is to be folded, vector loop takes care of all iterations. | ||||||
| 2974 | Type *CountTy = Count->getType(); | ||||||
| 2975 | Value *CheckMinIters = Builder.getFalse(); | ||||||
| 2976 | auto CreateStep = [&]() -> Value * { | ||||||
| 2977 | // Create step with max(MinProTripCount, UF * VF). | ||||||
| 2978 | if (UF * VF.getKnownMinValue() >= MinProfitableTripCount.getKnownMinValue()) | ||||||
| 2979 | return createStepForVF(Builder, CountTy, VF, UF); | ||||||
| 2980 | |||||||
| 2981 | Value *MinProfTC = | ||||||
| 2982 | createStepForVF(Builder, CountTy, MinProfitableTripCount, 1); | ||||||
| 2983 | if (!VF.isScalable()) | ||||||
| 2984 | return MinProfTC; | ||||||
| 2985 | return Builder.CreateBinaryIntrinsic( | ||||||
| 2986 | Intrinsic::umax, MinProfTC, createStepForVF(Builder, CountTy, VF, UF)); | ||||||
| 2987 | }; | ||||||
| 2988 | |||||||
| 2989 | if (!Cost->foldTailByMasking()) | ||||||
| 2990 | CheckMinIters = | ||||||
| 2991 | Builder.CreateICmp(P, Count, CreateStep(), "min.iters.check"); | ||||||
| 2992 | else if (VF.isScalable()) { | ||||||
| 2993 | // vscale is not necessarily a power-of-2, which means we cannot guarantee | ||||||
| 2994 | // an overflow to zero when updating induction variables and so an | ||||||
| 2995 | // additional overflow check is required before entering the vector loop. | ||||||
| 2996 | |||||||
| 2997 | // Get the maximum unsigned value for the type. | ||||||
| 2998 | Value *MaxUIntTripCount = | ||||||
| 2999 | ConstantInt::get(CountTy, cast<IntegerType>(CountTy)->getMask()); | ||||||
| 3000 | Value *LHS = Builder.CreateSub(MaxUIntTripCount, Count); | ||||||
| 3001 | |||||||
| 3002 | // Don't execute the vector loop if (UMax - n) < (VF * UF). | ||||||
| 3003 | CheckMinIters = Builder.CreateICmp(ICmpInst::ICMP_ULT, LHS, CreateStep()); | ||||||
| 3004 | } | ||||||
| 3005 | |||||||
| 3006 | // Create new preheader for vector loop. | ||||||
| 3007 | LoopVectorPreHeader = | ||||||
| 3008 | SplitBlock(TCCheckBlock, TCCheckBlock->getTerminator(), DT, LI, nullptr, | ||||||
| 3009 | "vector.ph"); | ||||||
| 3010 | |||||||
| 3011 | assert(DT->properlyDominates(DT->getNode(TCCheckBlock),(static_cast <bool> (DT->properlyDominates(DT->getNode (TCCheckBlock), DT->getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass") ? void (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3013, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 3012 | DT->getNode(Bypass)->getIDom()) &&(static_cast <bool> (DT->properlyDominates(DT->getNode (TCCheckBlock), DT->getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass") ? void (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3013, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 3013 | "TC check is expected to dominate Bypass")(static_cast <bool> (DT->properlyDominates(DT->getNode (TCCheckBlock), DT->getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass") ? void (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3013, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3014 | |||||||
| 3015 | // Update dominator for Bypass & LoopExit (if needed). | ||||||
| 3016 | DT->changeImmediateDominator(Bypass, TCCheckBlock); | ||||||
| 3017 | if (!Cost->requiresScalarEpilogue(VF)) | ||||||
| 3018 | // If there is an epilogue which must run, there's no edge from the | ||||||
| 3019 | // middle block to exit blocks and thus no need to update the immediate | ||||||
| 3020 | // dominator of the exit blocks. | ||||||
| 3021 | DT->changeImmediateDominator(LoopExitBlock, TCCheckBlock); | ||||||
| 3022 | |||||||
| 3023 | ReplaceInstWithInst( | ||||||
| 3024 | TCCheckBlock->getTerminator(), | ||||||
| 3025 | BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); | ||||||
| 3026 | LoopBypassBlocks.push_back(TCCheckBlock); | ||||||
| 3027 | } | ||||||
| 3028 | |||||||
| 3029 | BasicBlock *InnerLoopVectorizer::emitSCEVChecks(BasicBlock *Bypass) { | ||||||
| 3030 | BasicBlock *const SCEVCheckBlock = | ||||||
| 3031 | RTChecks.emitSCEVChecks(Bypass, LoopVectorPreHeader, LoopExitBlock); | ||||||
| 3032 | if (!SCEVCheckBlock) | ||||||
| 3033 | return nullptr; | ||||||
| 3034 | |||||||
| 3035 | assert(!(SCEVCheckBlock->getParent()->hasOptSize() ||(static_cast <bool> (!(SCEVCheckBlock->getParent()-> hasOptSize() || (OptForSizeBasedOnProfile && Cost-> Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && "Cannot SCEV check stride or overflow when optimizing for size" ) ? void (0) : __assert_fail ("!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && \"Cannot SCEV check stride or overflow when optimizing for size\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3038, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 3036 | (OptForSizeBasedOnProfile &&(static_cast <bool> (!(SCEVCheckBlock->getParent()-> hasOptSize() || (OptForSizeBasedOnProfile && Cost-> Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && "Cannot SCEV check stride or overflow when optimizing for size" ) ? void (0) : __assert_fail ("!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && \"Cannot SCEV check stride or overflow when optimizing for size\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3038, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 3037 | Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) &&(static_cast <bool> (!(SCEVCheckBlock->getParent()-> hasOptSize() || (OptForSizeBasedOnProfile && Cost-> Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && "Cannot SCEV check stride or overflow when optimizing for size" ) ? void (0) : __assert_fail ("!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && \"Cannot SCEV check stride or overflow when optimizing for size\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3038, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 3038 | "Cannot SCEV check stride or overflow when optimizing for size")(static_cast <bool> (!(SCEVCheckBlock->getParent()-> hasOptSize() || (OptForSizeBasedOnProfile && Cost-> Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && "Cannot SCEV check stride or overflow when optimizing for size" ) ? void (0) : __assert_fail ("!(SCEVCheckBlock->getParent()->hasOptSize() || (OptForSizeBasedOnProfile && Cost->Hints->getForce() != LoopVectorizeHints::FK_Enabled)) && \"Cannot SCEV check stride or overflow when optimizing for size\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3038, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3039 | |||||||
| 3040 | |||||||
| 3041 | // Update dominator only if this is first RT check. | ||||||
| 3042 | if (LoopBypassBlocks.empty()) { | ||||||
| 3043 | DT->changeImmediateDominator(Bypass, SCEVCheckBlock); | ||||||
| 3044 | if (!Cost->requiresScalarEpilogue(VF)) | ||||||
| 3045 | // If there is an epilogue which must run, there's no edge from the | ||||||
| 3046 | // middle block to exit blocks and thus no need to update the immediate | ||||||
| 3047 | // dominator of the exit blocks. | ||||||
| 3048 | DT->changeImmediateDominator(LoopExitBlock, SCEVCheckBlock); | ||||||
| 3049 | } | ||||||
| 3050 | |||||||
| 3051 | LoopBypassBlocks.push_back(SCEVCheckBlock); | ||||||
| 3052 | AddedSafetyChecks = true; | ||||||
| 3053 | return SCEVCheckBlock; | ||||||
| 3054 | } | ||||||
| 3055 | |||||||
| 3056 | BasicBlock *InnerLoopVectorizer::emitMemRuntimeChecks(BasicBlock *Bypass) { | ||||||
| 3057 | // VPlan-native path does not do any analysis for runtime checks currently. | ||||||
| 3058 | if (EnableVPlanNativePath) | ||||||
| 3059 | return nullptr; | ||||||
| 3060 | |||||||
| 3061 | BasicBlock *const MemCheckBlock = | ||||||
| 3062 | RTChecks.emitMemRuntimeChecks(Bypass, LoopVectorPreHeader); | ||||||
| 3063 | |||||||
| 3064 | // Check if we generated code that checks in runtime if arrays overlap. We put | ||||||
| 3065 | // the checks into a separate block to make the more common case of few | ||||||
| 3066 | // elements faster. | ||||||
| 3067 | if (!MemCheckBlock) | ||||||
| 3068 | return nullptr; | ||||||
| 3069 | |||||||
| 3070 | if (MemCheckBlock->getParent()->hasOptSize() || OptForSizeBasedOnProfile) { | ||||||
| 3071 | assert(Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled &&(static_cast <bool> (Cost->Hints->getForce() == LoopVectorizeHints ::FK_Enabled && "Cannot emit memory checks when optimizing for size, unless forced " "to vectorize.") ? void (0) : __assert_fail ("Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && \"Cannot emit memory checks when optimizing for size, unless forced \" \"to vectorize.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3073, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 3072 | "Cannot emit memory checks when optimizing for size, unless forced "(static_cast <bool> (Cost->Hints->getForce() == LoopVectorizeHints ::FK_Enabled && "Cannot emit memory checks when optimizing for size, unless forced " "to vectorize.") ? void (0) : __assert_fail ("Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && \"Cannot emit memory checks when optimizing for size, unless forced \" \"to vectorize.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3073, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 3073 | "to vectorize.")(static_cast <bool> (Cost->Hints->getForce() == LoopVectorizeHints ::FK_Enabled && "Cannot emit memory checks when optimizing for size, unless forced " "to vectorize.") ? void (0) : __assert_fail ("Cost->Hints->getForce() == LoopVectorizeHints::FK_Enabled && \"Cannot emit memory checks when optimizing for size, unless forced \" \"to vectorize.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3073, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3074 | ORE->emit([&]() { | ||||||
| 3075 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationCodeSize", | ||||||
| 3076 | OrigLoop->getStartLoc(), | ||||||
| 3077 | OrigLoop->getHeader()) | ||||||
| 3078 | << "Code-size may be reduced by not forcing " | ||||||
| 3079 | "vectorization, or by source-code modifications " | ||||||
| 3080 | "eliminating the need for runtime checks " | ||||||
| 3081 | "(e.g., adding 'restrict')."; | ||||||
| 3082 | }); | ||||||
| 3083 | } | ||||||
| 3084 | |||||||
| 3085 | LoopBypassBlocks.push_back(MemCheckBlock); | ||||||
| 3086 | |||||||
| 3087 | AddedSafetyChecks = true; | ||||||
| 3088 | |||||||
| 3089 | return MemCheckBlock; | ||||||
| 3090 | } | ||||||
| 3091 | |||||||
| 3092 | void InnerLoopVectorizer::createVectorLoopSkeleton(StringRef Prefix) { | ||||||
| 3093 | LoopScalarBody = OrigLoop->getHeader(); | ||||||
| 3094 | LoopVectorPreHeader = OrigLoop->getLoopPreheader(); | ||||||
| 3095 | assert(LoopVectorPreHeader && "Invalid loop structure")(static_cast <bool> (LoopVectorPreHeader && "Invalid loop structure" ) ? void (0) : __assert_fail ("LoopVectorPreHeader && \"Invalid loop structure\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3095, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3096 | LoopExitBlock = OrigLoop->getUniqueExitBlock(); // may be nullptr | ||||||
| 3097 | assert((LoopExitBlock || Cost->requiresScalarEpilogue(VF)) &&(static_cast <bool> ((LoopExitBlock || Cost->requiresScalarEpilogue (VF)) && "multiple exit loop without required epilogue?" ) ? void (0) : __assert_fail ("(LoopExitBlock || Cost->requiresScalarEpilogue(VF)) && \"multiple exit loop without required epilogue?\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3098, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 3098 | "multiple exit loop without required epilogue?")(static_cast <bool> ((LoopExitBlock || Cost->requiresScalarEpilogue (VF)) && "multiple exit loop without required epilogue?" ) ? void (0) : __assert_fail ("(LoopExitBlock || Cost->requiresScalarEpilogue(VF)) && \"multiple exit loop without required epilogue?\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3098, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3099 | |||||||
| 3100 | LoopMiddleBlock = | ||||||
| 3101 | SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, | ||||||
| 3102 | LI, nullptr, Twine(Prefix) + "middle.block"); | ||||||
| 3103 | LoopScalarPreHeader = | ||||||
| 3104 | SplitBlock(LoopMiddleBlock, LoopMiddleBlock->getTerminator(), DT, LI, | ||||||
| 3105 | nullptr, Twine(Prefix) + "scalar.ph"); | ||||||
| 3106 | |||||||
| 3107 | auto *ScalarLatchTerm = OrigLoop->getLoopLatch()->getTerminator(); | ||||||
| 3108 | |||||||
| 3109 | // Set up the middle block terminator. Two cases: | ||||||
| 3110 | // 1) If we know that we must execute the scalar epilogue, emit an | ||||||
| 3111 | // unconditional branch. | ||||||
| 3112 | // 2) Otherwise, we must have a single unique exit block (due to how we | ||||||
| 3113 | // implement the multiple exit case). In this case, set up a conditional | ||||||
| 3114 | // branch from the middle block to the loop scalar preheader, and the | ||||||
| 3115 | // exit block. completeLoopSkeleton will update the condition to use an | ||||||
| 3116 | // iteration check, if required to decide whether to execute the remainder. | ||||||
| 3117 | BranchInst *BrInst = Cost->requiresScalarEpilogue(VF) ? | ||||||
| 3118 | BranchInst::Create(LoopScalarPreHeader) : | ||||||
| 3119 | BranchInst::Create(LoopExitBlock, LoopScalarPreHeader, | ||||||
| 3120 | Builder.getTrue()); | ||||||
| 3121 | BrInst->setDebugLoc(ScalarLatchTerm->getDebugLoc()); | ||||||
| 3122 | ReplaceInstWithInst(LoopMiddleBlock->getTerminator(), BrInst); | ||||||
| 3123 | |||||||
| 3124 | // Update dominator for loop exit. During skeleton creation, only the vector | ||||||
| 3125 | // pre-header and the middle block are created. The vector loop is entirely | ||||||
| 3126 | // created during VPlan exection. | ||||||
| 3127 | if (!Cost->requiresScalarEpilogue(VF)) | ||||||
| 3128 | // If there is an epilogue which must run, there's no edge from the | ||||||
| 3129 | // middle block to exit blocks and thus no need to update the immediate | ||||||
| 3130 | // dominator of the exit blocks. | ||||||
| 3131 | DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock); | ||||||
| 3132 | } | ||||||
| 3133 | |||||||
| 3134 | PHINode *InnerLoopVectorizer::createInductionResumeValue( | ||||||
| 3135 | PHINode *OrigPhi, const InductionDescriptor &II, | ||||||
| 3136 | ArrayRef<BasicBlock *> BypassBlocks, | ||||||
| 3137 | std::pair<BasicBlock *, Value *> AdditionalBypass) { | ||||||
| 3138 | Value *VectorTripCount = getOrCreateVectorTripCount(LoopVectorPreHeader); | ||||||
| 3139 | assert(VectorTripCount && "Expected valid arguments")(static_cast <bool> (VectorTripCount && "Expected valid arguments" ) ? void (0) : __assert_fail ("VectorTripCount && \"Expected valid arguments\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3139, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3140 | |||||||
| 3141 | Instruction *OldInduction = Legal->getPrimaryInduction(); | ||||||
| 3142 | Value *&EndValue = IVEndValues[OrigPhi]; | ||||||
| 3143 | Value *EndValueFromAdditionalBypass = AdditionalBypass.second; | ||||||
| 3144 | if (OrigPhi == OldInduction) { | ||||||
| 3145 | // We know what the end value is. | ||||||
| 3146 | EndValue = VectorTripCount; | ||||||
| 3147 | } else { | ||||||
| 3148 | IRBuilder<> B(LoopVectorPreHeader->getTerminator()); | ||||||
| 3149 | |||||||
| 3150 | // Fast-math-flags propagate from the original induction instruction. | ||||||
| 3151 | if (II.getInductionBinOp() && isa<FPMathOperator>(II.getInductionBinOp())) | ||||||
| 3152 | B.setFastMathFlags(II.getInductionBinOp()->getFastMathFlags()); | ||||||
| 3153 | |||||||
| 3154 | Value *Step = | ||||||
| 3155 | CreateStepValue(II.getStep(), *PSE.getSE(), &*B.GetInsertPoint()); | ||||||
| 3156 | EndValue = | ||||||
| 3157 | emitTransformedIndex(B, VectorTripCount, II.getStartValue(), Step, II); | ||||||
| 3158 | EndValue->setName("ind.end"); | ||||||
| 3159 | |||||||
| 3160 | // Compute the end value for the additional bypass (if applicable). | ||||||
| 3161 | if (AdditionalBypass.first) { | ||||||
| 3162 | B.SetInsertPoint(&(*AdditionalBypass.first->getFirstInsertionPt())); | ||||||
| 3163 | Value *Step = | ||||||
| 3164 | CreateStepValue(II.getStep(), *PSE.getSE(), &*B.GetInsertPoint()); | ||||||
| 3165 | EndValueFromAdditionalBypass = emitTransformedIndex( | ||||||
| 3166 | B, AdditionalBypass.second, II.getStartValue(), Step, II); | ||||||
| 3167 | EndValueFromAdditionalBypass->setName("ind.end"); | ||||||
| 3168 | } | ||||||
| 3169 | } | ||||||
| 3170 | |||||||
| 3171 | // Create phi nodes to merge from the backedge-taken check block. | ||||||
| 3172 | PHINode *BCResumeVal = PHINode::Create(OrigPhi->getType(), 3, "bc.resume.val", | ||||||
| 3173 | LoopScalarPreHeader->getTerminator()); | ||||||
| 3174 | // Copy original phi DL over to the new one. | ||||||
| 3175 | BCResumeVal->setDebugLoc(OrigPhi->getDebugLoc()); | ||||||
| 3176 | |||||||
| 3177 | // The new PHI merges the original incoming value, in case of a bypass, | ||||||
| 3178 | // or the value at the end of the vectorized loop. | ||||||
| 3179 | BCResumeVal->addIncoming(EndValue, LoopMiddleBlock); | ||||||
| 3180 | |||||||
| 3181 | // Fix the scalar body counter (PHI node). | ||||||
| 3182 | // The old induction's phi node in the scalar body needs the truncated | ||||||
| 3183 | // value. | ||||||
| 3184 | for (BasicBlock *BB : BypassBlocks) | ||||||
| 3185 | BCResumeVal->addIncoming(II.getStartValue(), BB); | ||||||
| 3186 | |||||||
| 3187 | if (AdditionalBypass.first) | ||||||
| 3188 | BCResumeVal->setIncomingValueForBlock(AdditionalBypass.first, | ||||||
| 3189 | EndValueFromAdditionalBypass); | ||||||
| 3190 | return BCResumeVal; | ||||||
| 3191 | } | ||||||
| 3192 | |||||||
| 3193 | void InnerLoopVectorizer::createInductionResumeValues( | ||||||
| 3194 | std::pair<BasicBlock *, Value *> AdditionalBypass) { | ||||||
| 3195 | assert(((AdditionalBypass.first && AdditionalBypass.second) ||(static_cast <bool> (((AdditionalBypass.first && AdditionalBypass.second) || (!AdditionalBypass.first && !AdditionalBypass.second)) && "Inconsistent information about additional bypass." ) ? void (0) : __assert_fail ("((AdditionalBypass.first && AdditionalBypass.second) || (!AdditionalBypass.first && !AdditionalBypass.second)) && \"Inconsistent information about additional bypass.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3197, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 3196 | (!AdditionalBypass.first && !AdditionalBypass.second)) &&(static_cast <bool> (((AdditionalBypass.first && AdditionalBypass.second) || (!AdditionalBypass.first && !AdditionalBypass.second)) && "Inconsistent information about additional bypass." ) ? void (0) : __assert_fail ("((AdditionalBypass.first && AdditionalBypass.second) || (!AdditionalBypass.first && !AdditionalBypass.second)) && \"Inconsistent information about additional bypass.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3197, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 3197 | "Inconsistent information about additional bypass.")(static_cast <bool> (((AdditionalBypass.first && AdditionalBypass.second) || (!AdditionalBypass.first && !AdditionalBypass.second)) && "Inconsistent information about additional bypass." ) ? void (0) : __assert_fail ("((AdditionalBypass.first && AdditionalBypass.second) || (!AdditionalBypass.first && !AdditionalBypass.second)) && \"Inconsistent information about additional bypass.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3197, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3198 | // We are going to resume the execution of the scalar loop. | ||||||
| 3199 | // Go over all of the induction variables that we found and fix the | ||||||
| 3200 | // PHIs that are left in the scalar version of the loop. | ||||||
| 3201 | // The starting values of PHI nodes depend on the counter of the last | ||||||
| 3202 | // iteration in the vectorized loop. | ||||||
| 3203 | // If we come from a bypass edge then we need to start from the original | ||||||
| 3204 | // start value. | ||||||
| 3205 | for (const auto &InductionEntry : Legal->getInductionVars()) { | ||||||
| 3206 | PHINode *OrigPhi = InductionEntry.first; | ||||||
| 3207 | const InductionDescriptor &II = InductionEntry.second; | ||||||
| 3208 | PHINode *BCResumeVal = createInductionResumeValue( | ||||||
| 3209 | OrigPhi, II, LoopBypassBlocks, AdditionalBypass); | ||||||
| 3210 | OrigPhi->setIncomingValueForBlock(LoopScalarPreHeader, BCResumeVal); | ||||||
| 3211 | } | ||||||
| 3212 | } | ||||||
| 3213 | |||||||
| 3214 | BasicBlock *InnerLoopVectorizer::completeLoopSkeleton() { | ||||||
| 3215 | // The trip counts should be cached by now. | ||||||
| 3216 | Value *Count = getOrCreateTripCount(LoopVectorPreHeader); | ||||||
| 3217 | Value *VectorTripCount = getOrCreateVectorTripCount(LoopVectorPreHeader); | ||||||
| 3218 | |||||||
| 3219 | auto *ScalarLatchTerm = OrigLoop->getLoopLatch()->getTerminator(); | ||||||
| 3220 | |||||||
| 3221 | // Add a check in the middle block to see if we have completed | ||||||
| 3222 | // all of the iterations in the first vector loop. Three cases: | ||||||
| 3223 | // 1) If we require a scalar epilogue, there is no conditional branch as | ||||||
| 3224 | // we unconditionally branch to the scalar preheader. Do nothing. | ||||||
| 3225 | // 2) If (N - N%VF) == N, then we *don't* need to run the remainder. | ||||||
| 3226 | // Thus if tail is to be folded, we know we don't need to run the | ||||||
| 3227 | // remainder and we can use the previous value for the condition (true). | ||||||
| 3228 | // 3) Otherwise, construct a runtime check. | ||||||
| 3229 | if (!Cost->requiresScalarEpilogue(VF) && !Cost->foldTailByMasking()) { | ||||||
| 3230 | Instruction *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, | ||||||
| 3231 | Count, VectorTripCount, "cmp.n", | ||||||
| 3232 | LoopMiddleBlock->getTerminator()); | ||||||
| 3233 | |||||||
| 3234 | // Here we use the same DebugLoc as the scalar loop latch terminator instead | ||||||
| 3235 | // of the corresponding compare because they may have ended up with | ||||||
| 3236 | // different line numbers and we want to avoid awkward line stepping while | ||||||
| 3237 | // debugging. Eg. if the compare has got a line number inside the loop. | ||||||
| 3238 | CmpN->setDebugLoc(ScalarLatchTerm->getDebugLoc()); | ||||||
| 3239 | cast<BranchInst>(LoopMiddleBlock->getTerminator())->setCondition(CmpN); | ||||||
| 3240 | } | ||||||
| 3241 | |||||||
| 3242 | #ifdef EXPENSIVE_CHECKS | ||||||
| 3243 | assert(DT->verify(DominatorTree::VerificationLevel::Fast))(static_cast <bool> (DT->verify(DominatorTree::VerificationLevel ::Fast)) ? void (0) : __assert_fail ("DT->verify(DominatorTree::VerificationLevel::Fast)" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3243, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3244 | #endif | ||||||
| 3245 | |||||||
| 3246 | return LoopVectorPreHeader; | ||||||
| 3247 | } | ||||||
| 3248 | |||||||
| 3249 | std::pair<BasicBlock *, Value *> | ||||||
| 3250 | InnerLoopVectorizer::createVectorizedLoopSkeleton() { | ||||||
| 3251 | /* | ||||||
| 3252 | In this function we generate a new loop. The new loop will contain | ||||||
| 3253 | the vectorized instructions while the old loop will continue to run the | ||||||
| 3254 | scalar remainder. | ||||||
| 3255 | |||||||
| 3256 | [ ] <-- loop iteration number check. | ||||||
| 3257 | / | | ||||||
| 3258 | / v | ||||||
| 3259 | | [ ] <-- vector loop bypass (may consist of multiple blocks). | ||||||
| 3260 | | / | | ||||||
| 3261 | | / v | ||||||
| 3262 | || [ ] <-- vector pre header. | ||||||
| 3263 | |/ | | ||||||
| 3264 | | v | ||||||
| 3265 | | [ ] \ | ||||||
| 3266 | | [ ]_| <-- vector loop (created during VPlan execution). | ||||||
| 3267 | | | | ||||||
| 3268 | | v | ||||||
| 3269 | \ -[ ] <--- middle-block. | ||||||
| 3270 | \/ | | ||||||
| 3271 | /\ v | ||||||
| 3272 | | ->[ ] <--- new preheader. | ||||||
| 3273 | | | | ||||||
| 3274 | (opt) v <-- edge from middle to exit iff epilogue is not required. | ||||||
| 3275 | | [ ] \ | ||||||
| 3276 | | [ ]_| <-- old scalar loop to handle remainder (scalar epilogue). | ||||||
| 3277 | \ | | ||||||
| 3278 | \ v | ||||||
| 3279 | >[ ] <-- exit block(s). | ||||||
| 3280 | ... | ||||||
| 3281 | */ | ||||||
| 3282 | |||||||
| 3283 | // Create an empty vector loop, and prepare basic blocks for the runtime | ||||||
| 3284 | // checks. | ||||||
| 3285 | createVectorLoopSkeleton(""); | ||||||
| 3286 | |||||||
| 3287 | // Now, compare the new count to zero. If it is zero skip the vector loop and | ||||||
| 3288 | // jump to the scalar loop. This check also covers the case where the | ||||||
| 3289 | // backedge-taken count is uint##_max: adding one to it will overflow leading | ||||||
| 3290 | // to an incorrect trip count of zero. In this (rare) case we will also jump | ||||||
| 3291 | // to the scalar loop. | ||||||
| 3292 | emitIterationCountCheck(LoopScalarPreHeader); | ||||||
| 3293 | |||||||
| 3294 | // Generate the code to check any assumptions that we've made for SCEV | ||||||
| 3295 | // expressions. | ||||||
| 3296 | emitSCEVChecks(LoopScalarPreHeader); | ||||||
| 3297 | |||||||
| 3298 | // Generate the code that checks in runtime if arrays overlap. We put the | ||||||
| 3299 | // checks into a separate block to make the more common case of few elements | ||||||
| 3300 | // faster. | ||||||
| 3301 | emitMemRuntimeChecks(LoopScalarPreHeader); | ||||||
| 3302 | |||||||
| 3303 | // Emit phis for the new starting index of the scalar loop. | ||||||
| 3304 | createInductionResumeValues(); | ||||||
| 3305 | |||||||
| 3306 | return {completeLoopSkeleton(), nullptr}; | ||||||
| 3307 | } | ||||||
| 3308 | |||||||
| 3309 | // Fix up external users of the induction variable. At this point, we are | ||||||
| 3310 | // in LCSSA form, with all external PHIs that use the IV having one input value, | ||||||
| 3311 | // coming from the remainder loop. We need those PHIs to also have a correct | ||||||
| 3312 | // value for the IV when arriving directly from the middle block. | ||||||
| 3313 | void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi, | ||||||
| 3314 | const InductionDescriptor &II, | ||||||
| 3315 | Value *VectorTripCount, Value *EndValue, | ||||||
| 3316 | BasicBlock *MiddleBlock, | ||||||
| 3317 | BasicBlock *VectorHeader, VPlan &Plan) { | ||||||
| 3318 | // There are two kinds of external IV usages - those that use the value | ||||||
| 3319 | // computed in the last iteration (the PHI) and those that use the penultimate | ||||||
| 3320 | // value (the value that feeds into the phi from the loop latch). | ||||||
| 3321 | // We allow both, but they, obviously, have different values. | ||||||
| 3322 | |||||||
| 3323 | assert(OrigLoop->getUniqueExitBlock() && "Expected a single exit block")(static_cast <bool> (OrigLoop->getUniqueExitBlock() && "Expected a single exit block") ? void (0) : __assert_fail ( "OrigLoop->getUniqueExitBlock() && \"Expected a single exit block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3323, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3324 | |||||||
| 3325 | DenseMap<Value *, Value *> MissingVals; | ||||||
| 3326 | |||||||
| 3327 | // An external user of the last iteration's value should see the value that | ||||||
| 3328 | // the remainder loop uses to initialize its own IV. | ||||||
| 3329 | Value *PostInc = OrigPhi->getIncomingValueForBlock(OrigLoop->getLoopLatch()); | ||||||
| 3330 | for (User *U : PostInc->users()) { | ||||||
| 3331 | Instruction *UI = cast<Instruction>(U); | ||||||
| 3332 | if (!OrigLoop->contains(UI)) { | ||||||
| 3333 | assert(isa<PHINode>(UI) && "Expected LCSSA form")(static_cast <bool> (isa<PHINode>(UI) && "Expected LCSSA form" ) ? void (0) : __assert_fail ("isa<PHINode>(UI) && \"Expected LCSSA form\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3333, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3334 | MissingVals[UI] = EndValue; | ||||||
| 3335 | } | ||||||
| 3336 | } | ||||||
| 3337 | |||||||
| 3338 | // An external user of the penultimate value need to see EndValue - Step. | ||||||
| 3339 | // The simplest way to get this is to recompute it from the constituent SCEVs, | ||||||
| 3340 | // that is Start + (Step * (CRD - 1)). | ||||||
| 3341 | for (User *U : OrigPhi->users()) { | ||||||
| 3342 | auto *UI = cast<Instruction>(U); | ||||||
| 3343 | if (!OrigLoop->contains(UI)) { | ||||||
| 3344 | assert(isa<PHINode>(UI) && "Expected LCSSA form")(static_cast <bool> (isa<PHINode>(UI) && "Expected LCSSA form" ) ? void (0) : __assert_fail ("isa<PHINode>(UI) && \"Expected LCSSA form\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3344, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3345 | |||||||
| 3346 | IRBuilder<> B(MiddleBlock->getTerminator()); | ||||||
| 3347 | |||||||
| 3348 | // Fast-math-flags propagate from the original induction instruction. | ||||||
| 3349 | if (II.getInductionBinOp() && isa<FPMathOperator>(II.getInductionBinOp())) | ||||||
| 3350 | B.setFastMathFlags(II.getInductionBinOp()->getFastMathFlags()); | ||||||
| 3351 | |||||||
| 3352 | Value *CountMinusOne = B.CreateSub( | ||||||
| 3353 | VectorTripCount, ConstantInt::get(VectorTripCount->getType(), 1)); | ||||||
| 3354 | CountMinusOne->setName("cmo"); | ||||||
| 3355 | Value *Step = CreateStepValue(II.getStep(), *PSE.getSE(), | ||||||
| 3356 | VectorHeader->getTerminator()); | ||||||
| 3357 | Value *Escape = | ||||||
| 3358 | emitTransformedIndex(B, CountMinusOne, II.getStartValue(), Step, II); | ||||||
| 3359 | Escape->setName("ind.escape"); | ||||||
| 3360 | MissingVals[UI] = Escape; | ||||||
| 3361 | } | ||||||
| 3362 | } | ||||||
| 3363 | |||||||
| 3364 | for (auto &I : MissingVals) { | ||||||
| 3365 | PHINode *PHI = cast<PHINode>(I.first); | ||||||
| 3366 | // One corner case we have to handle is two IVs "chasing" each-other, | ||||||
| 3367 | // that is %IV2 = phi [...], [ %IV1, %latch ] | ||||||
| 3368 | // In this case, if IV1 has an external use, we need to avoid adding both | ||||||
| 3369 | // "last value of IV1" and "penultimate value of IV2". So, verify that we | ||||||
| 3370 | // don't already have an incoming value for the middle block. | ||||||
| 3371 | if (PHI->getBasicBlockIndex(MiddleBlock) == -1) { | ||||||
| 3372 | PHI->addIncoming(I.second, MiddleBlock); | ||||||
| 3373 | Plan.removeLiveOut(PHI); | ||||||
| 3374 | } | ||||||
| 3375 | } | ||||||
| 3376 | } | ||||||
| 3377 | |||||||
| 3378 | namespace { | ||||||
| 3379 | |||||||
| 3380 | struct CSEDenseMapInfo { | ||||||
| 3381 | static bool canHandle(const Instruction *I) { | ||||||
| 3382 | return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || | ||||||
| 3383 | isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I); | ||||||
| 3384 | } | ||||||
| 3385 | |||||||
| 3386 | static inline Instruction *getEmptyKey() { | ||||||
| 3387 | return DenseMapInfo<Instruction *>::getEmptyKey(); | ||||||
| 3388 | } | ||||||
| 3389 | |||||||
| 3390 | static inline Instruction *getTombstoneKey() { | ||||||
| 3391 | return DenseMapInfo<Instruction *>::getTombstoneKey(); | ||||||
| 3392 | } | ||||||
| 3393 | |||||||
| 3394 | static unsigned getHashValue(const Instruction *I) { | ||||||
| 3395 | assert(canHandle(I) && "Unknown instruction!")(static_cast <bool> (canHandle(I) && "Unknown instruction!" ) ? void (0) : __assert_fail ("canHandle(I) && \"Unknown instruction!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3395, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3396 | return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(), | ||||||
| 3397 | I->value_op_end())); | ||||||
| 3398 | } | ||||||
| 3399 | |||||||
| 3400 | static bool isEqual(const Instruction *LHS, const Instruction *RHS) { | ||||||
| 3401 | if (LHS == getEmptyKey() || RHS == getEmptyKey() || | ||||||
| 3402 | LHS == getTombstoneKey() || RHS == getTombstoneKey()) | ||||||
| 3403 | return LHS == RHS; | ||||||
| 3404 | return LHS->isIdenticalTo(RHS); | ||||||
| 3405 | } | ||||||
| 3406 | }; | ||||||
| 3407 | |||||||
| 3408 | } // end anonymous namespace | ||||||
| 3409 | |||||||
| 3410 | ///Perform cse of induction variable instructions. | ||||||
| 3411 | static void cse(BasicBlock *BB) { | ||||||
| 3412 | // Perform simple cse. | ||||||
| 3413 | SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap; | ||||||
| 3414 | for (Instruction &In : llvm::make_early_inc_range(*BB)) { | ||||||
| 3415 | if (!CSEDenseMapInfo::canHandle(&In)) | ||||||
| 3416 | continue; | ||||||
| 3417 | |||||||
| 3418 | // Check if we can replace this instruction with any of the | ||||||
| 3419 | // visited instructions. | ||||||
| 3420 | if (Instruction *V = CSEMap.lookup(&In)) { | ||||||
| 3421 | In.replaceAllUsesWith(V); | ||||||
| 3422 | In.eraseFromParent(); | ||||||
| 3423 | continue; | ||||||
| 3424 | } | ||||||
| 3425 | |||||||
| 3426 | CSEMap[&In] = &In; | ||||||
| 3427 | } | ||||||
| 3428 | } | ||||||
| 3429 | |||||||
| 3430 | InstructionCost | ||||||
| 3431 | LoopVectorizationCostModel::getVectorCallCost(CallInst *CI, ElementCount VF, | ||||||
| 3432 | bool &NeedToScalarize) const { | ||||||
| 3433 | Function *F = CI->getCalledFunction(); | ||||||
| 3434 | Type *ScalarRetTy = CI->getType(); | ||||||
| 3435 | SmallVector<Type *, 4> Tys, ScalarTys; | ||||||
| 3436 | for (auto &ArgOp : CI->args()) | ||||||
| 3437 | ScalarTys.push_back(ArgOp->getType()); | ||||||
| 3438 | |||||||
| 3439 | // Estimate cost of scalarized vector call. The source operands are assumed | ||||||
| 3440 | // to be vectors, so we need to extract individual elements from there, | ||||||
| 3441 | // execute VF scalar calls, and then gather the result into the vector return | ||||||
| 3442 | // value. | ||||||
| 3443 | InstructionCost ScalarCallCost = | ||||||
| 3444 | TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys, TTI::TCK_RecipThroughput); | ||||||
| 3445 | if (VF.isScalar()) | ||||||
| 3446 | return ScalarCallCost; | ||||||
| 3447 | |||||||
| 3448 | // Compute corresponding vector type for return value and arguments. | ||||||
| 3449 | Type *RetTy = ToVectorTy(ScalarRetTy, VF); | ||||||
| 3450 | for (Type *ScalarTy : ScalarTys) | ||||||
| 3451 | Tys.push_back(ToVectorTy(ScalarTy, VF)); | ||||||
| 3452 | |||||||
| 3453 | // Compute costs of unpacking argument values for the scalar calls and | ||||||
| 3454 | // packing the return values to a vector. | ||||||
| 3455 | InstructionCost ScalarizationCost = getScalarizationOverhead(CI, VF); | ||||||
| 3456 | |||||||
| 3457 | InstructionCost Cost = | ||||||
| 3458 | ScalarCallCost * VF.getKnownMinValue() + ScalarizationCost; | ||||||
| 3459 | |||||||
| 3460 | // If we can't emit a vector call for this function, then the currently found | ||||||
| 3461 | // cost is the cost we need to return. | ||||||
| 3462 | NeedToScalarize = true; | ||||||
| 3463 | VFShape Shape = VFShape::get(*CI, VF, false /*HasGlobalPred*/); | ||||||
| 3464 | Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); | ||||||
| 3465 | |||||||
| 3466 | if (!TLI || CI->isNoBuiltin() || !VecFunc) | ||||||
| 3467 | return Cost; | ||||||
| 3468 | |||||||
| 3469 | // If the corresponding vector cost is cheaper, return its cost. | ||||||
| 3470 | InstructionCost VectorCallCost = | ||||||
| 3471 | TTI.getCallInstrCost(nullptr, RetTy, Tys, TTI::TCK_RecipThroughput); | ||||||
| 3472 | if (VectorCallCost < Cost) { | ||||||
| 3473 | NeedToScalarize = false; | ||||||
| 3474 | Cost = VectorCallCost; | ||||||
| 3475 | } | ||||||
| 3476 | return Cost; | ||||||
| 3477 | } | ||||||
| 3478 | |||||||
| 3479 | static Type *MaybeVectorizeType(Type *Elt, ElementCount VF) { | ||||||
| 3480 | if (VF.isScalar() || (!Elt->isIntOrPtrTy() && !Elt->isFloatingPointTy())) | ||||||
| 3481 | return Elt; | ||||||
| 3482 | return VectorType::get(Elt, VF); | ||||||
| 3483 | } | ||||||
| 3484 | |||||||
| 3485 | InstructionCost | ||||||
| 3486 | LoopVectorizationCostModel::getVectorIntrinsicCost(CallInst *CI, | ||||||
| 3487 | ElementCount VF) const { | ||||||
| 3488 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||||
| 3489 | assert(ID && "Expected intrinsic call!")(static_cast <bool> (ID && "Expected intrinsic call!" ) ? void (0) : __assert_fail ("ID && \"Expected intrinsic call!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3489, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3490 | Type *RetTy = MaybeVectorizeType(CI->getType(), VF); | ||||||
| 3491 | FastMathFlags FMF; | ||||||
| 3492 | if (auto *FPMO = dyn_cast<FPMathOperator>(CI)) | ||||||
| 3493 | FMF = FPMO->getFastMathFlags(); | ||||||
| 3494 | |||||||
| 3495 | SmallVector<const Value *> Arguments(CI->args()); | ||||||
| 3496 | FunctionType *FTy = CI->getCalledFunction()->getFunctionType(); | ||||||
| 3497 | SmallVector<Type *> ParamTys; | ||||||
| 3498 | std::transform(FTy->param_begin(), FTy->param_end(), | ||||||
| 3499 | std::back_inserter(ParamTys), | ||||||
| 3500 | [&](Type *Ty) { return MaybeVectorizeType(Ty, VF); }); | ||||||
| 3501 | |||||||
| 3502 | IntrinsicCostAttributes CostAttrs(ID, RetTy, Arguments, ParamTys, FMF, | ||||||
| 3503 | dyn_cast<IntrinsicInst>(CI)); | ||||||
| 3504 | return TTI.getIntrinsicInstrCost(CostAttrs, | ||||||
| 3505 | TargetTransformInfo::TCK_RecipThroughput); | ||||||
| 3506 | } | ||||||
| 3507 | |||||||
| 3508 | static Type *smallestIntegerVectorType(Type *T1, Type *T2) { | ||||||
| 3509 | auto *I1 = cast<IntegerType>(cast<VectorType>(T1)->getElementType()); | ||||||
| 3510 | auto *I2 = cast<IntegerType>(cast<VectorType>(T2)->getElementType()); | ||||||
| 3511 | return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2; | ||||||
| 3512 | } | ||||||
| 3513 | |||||||
| 3514 | static Type *largestIntegerVectorType(Type *T1, Type *T2) { | ||||||
| 3515 | auto *I1 = cast<IntegerType>(cast<VectorType>(T1)->getElementType()); | ||||||
| 3516 | auto *I2 = cast<IntegerType>(cast<VectorType>(T2)->getElementType()); | ||||||
| 3517 | return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2; | ||||||
| 3518 | } | ||||||
| 3519 | |||||||
| 3520 | void InnerLoopVectorizer::truncateToMinimalBitwidths(VPTransformState &State) { | ||||||
| 3521 | // For every instruction `I` in MinBWs, truncate the operands, create a | ||||||
| 3522 | // truncated version of `I` and reextend its result. InstCombine runs | ||||||
| 3523 | // later and will remove any ext/trunc pairs. | ||||||
| 3524 | SmallPtrSet<Value *, 4> Erased; | ||||||
| 3525 | for (const auto &KV : Cost->getMinimalBitwidths()) { | ||||||
| 3526 | // If the value wasn't vectorized, we must maintain the original scalar | ||||||
| 3527 | // type. The absence of the value from State indicates that it | ||||||
| 3528 | // wasn't vectorized. | ||||||
| 3529 | // FIXME: Should not rely on getVPValue at this point. | ||||||
| 3530 | VPValue *Def = State.Plan->getVPValue(KV.first, true); | ||||||
| 3531 | if (!State.hasAnyVectorValue(Def)) | ||||||
| 3532 | continue; | ||||||
| 3533 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
| 3534 | Value *I = State.get(Def, Part); | ||||||
| 3535 | if (Erased.count(I) || I->use_empty() || !isa<Instruction>(I)) | ||||||
| 3536 | continue; | ||||||
| 3537 | Type *OriginalTy = I->getType(); | ||||||
| 3538 | Type *ScalarTruncatedTy = | ||||||
| 3539 | IntegerType::get(OriginalTy->getContext(), KV.second); | ||||||
| 3540 | auto *TruncatedTy = VectorType::get( | ||||||
| 3541 | ScalarTruncatedTy, cast<VectorType>(OriginalTy)->getElementCount()); | ||||||
| 3542 | if (TruncatedTy == OriginalTy) | ||||||
| 3543 | continue; | ||||||
| 3544 | |||||||
| 3545 | IRBuilder<> B(cast<Instruction>(I)); | ||||||
| 3546 | auto ShrinkOperand = [&](Value *V) -> Value * { | ||||||
| 3547 | if (auto *ZI = dyn_cast<ZExtInst>(V)) | ||||||
| 3548 | if (ZI->getSrcTy() == TruncatedTy) | ||||||
| 3549 | return ZI->getOperand(0); | ||||||
| 3550 | return B.CreateZExtOrTrunc(V, TruncatedTy); | ||||||
| 3551 | }; | ||||||
| 3552 | |||||||
| 3553 | // The actual instruction modification depends on the instruction type, | ||||||
| 3554 | // unfortunately. | ||||||
| 3555 | Value *NewI = nullptr; | ||||||
| 3556 | if (auto *BO = dyn_cast<BinaryOperator>(I)) { | ||||||
| 3557 | NewI = B.CreateBinOp(BO->getOpcode(), ShrinkOperand(BO->getOperand(0)), | ||||||
| 3558 | ShrinkOperand(BO->getOperand(1))); | ||||||
| 3559 | |||||||
| 3560 | // Any wrapping introduced by shrinking this operation shouldn't be | ||||||
| 3561 | // considered undefined behavior. So, we can't unconditionally copy | ||||||
| 3562 | // arithmetic wrapping flags to NewI. | ||||||
| 3563 | cast<BinaryOperator>(NewI)->copyIRFlags(I, /*IncludeWrapFlags=*/false); | ||||||
| 3564 | } else if (auto *CI = dyn_cast<ICmpInst>(I)) { | ||||||
| 3565 | NewI = | ||||||
| 3566 | B.CreateICmp(CI->getPredicate(), ShrinkOperand(CI->getOperand(0)), | ||||||
| 3567 | ShrinkOperand(CI->getOperand(1))); | ||||||
| 3568 | } else if (auto *SI = dyn_cast<SelectInst>(I)) { | ||||||
| 3569 | NewI = B.CreateSelect(SI->getCondition(), | ||||||
| 3570 | ShrinkOperand(SI->getTrueValue()), | ||||||
| 3571 | ShrinkOperand(SI->getFalseValue())); | ||||||
| 3572 | } else if (auto *CI = dyn_cast<CastInst>(I)) { | ||||||
| 3573 | switch (CI->getOpcode()) { | ||||||
| 3574 | default: | ||||||
| 3575 | llvm_unreachable("Unhandled cast!")::llvm::llvm_unreachable_internal("Unhandled cast!", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3575); | ||||||
| 3576 | case Instruction::Trunc: | ||||||
| 3577 | NewI = ShrinkOperand(CI->getOperand(0)); | ||||||
| 3578 | break; | ||||||
| 3579 | case Instruction::SExt: | ||||||
| 3580 | NewI = B.CreateSExtOrTrunc( | ||||||
| 3581 | CI->getOperand(0), | ||||||
| 3582 | smallestIntegerVectorType(OriginalTy, TruncatedTy)); | ||||||
| 3583 | break; | ||||||
| 3584 | case Instruction::ZExt: | ||||||
| 3585 | NewI = B.CreateZExtOrTrunc( | ||||||
| 3586 | CI->getOperand(0), | ||||||
| 3587 | smallestIntegerVectorType(OriginalTy, TruncatedTy)); | ||||||
| 3588 | break; | ||||||
| 3589 | } | ||||||
| 3590 | } else if (auto *SI = dyn_cast<ShuffleVectorInst>(I)) { | ||||||
| 3591 | auto Elements0 = | ||||||
| 3592 | cast<VectorType>(SI->getOperand(0)->getType())->getElementCount(); | ||||||
| 3593 | auto *O0 = B.CreateZExtOrTrunc( | ||||||
| 3594 | SI->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements0)); | ||||||
| 3595 | auto Elements1 = | ||||||
| 3596 | cast<VectorType>(SI->getOperand(1)->getType())->getElementCount(); | ||||||
| 3597 | auto *O1 = B.CreateZExtOrTrunc( | ||||||
| 3598 | SI->getOperand(1), VectorType::get(ScalarTruncatedTy, Elements1)); | ||||||
| 3599 | |||||||
| 3600 | NewI = B.CreateShuffleVector(O0, O1, SI->getShuffleMask()); | ||||||
| 3601 | } else if (isa<LoadInst>(I) || isa<PHINode>(I)) { | ||||||
| 3602 | // Don't do anything with the operands, just extend the result. | ||||||
| 3603 | continue; | ||||||
| 3604 | } else if (auto *IE = dyn_cast<InsertElementInst>(I)) { | ||||||
| 3605 | auto Elements = | ||||||
| 3606 | cast<VectorType>(IE->getOperand(0)->getType())->getElementCount(); | ||||||
| 3607 | auto *O0 = B.CreateZExtOrTrunc( | ||||||
| 3608 | IE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements)); | ||||||
| 3609 | auto *O1 = B.CreateZExtOrTrunc(IE->getOperand(1), ScalarTruncatedTy); | ||||||
| 3610 | NewI = B.CreateInsertElement(O0, O1, IE->getOperand(2)); | ||||||
| 3611 | } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) { | ||||||
| 3612 | auto Elements = | ||||||
| 3613 | cast<VectorType>(EE->getOperand(0)->getType())->getElementCount(); | ||||||
| 3614 | auto *O0 = B.CreateZExtOrTrunc( | ||||||
| 3615 | EE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements)); | ||||||
| 3616 | NewI = B.CreateExtractElement(O0, EE->getOperand(2)); | ||||||
| 3617 | } else { | ||||||
| 3618 | // If we don't know what to do, be conservative and don't do anything. | ||||||
| 3619 | continue; | ||||||
| 3620 | } | ||||||
| 3621 | |||||||
| 3622 | // Lastly, extend the result. | ||||||
| 3623 | NewI->takeName(cast<Instruction>(I)); | ||||||
| 3624 | Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy); | ||||||
| 3625 | I->replaceAllUsesWith(Res); | ||||||
| 3626 | cast<Instruction>(I)->eraseFromParent(); | ||||||
| 3627 | Erased.insert(I); | ||||||
| 3628 | State.reset(Def, Res, Part); | ||||||
| 3629 | } | ||||||
| 3630 | } | ||||||
| 3631 | |||||||
| 3632 | // We'll have created a bunch of ZExts that are now parentless. Clean up. | ||||||
| 3633 | for (const auto &KV : Cost->getMinimalBitwidths()) { | ||||||
| 3634 | // If the value wasn't vectorized, we must maintain the original scalar | ||||||
| 3635 | // type. The absence of the value from State indicates that it | ||||||
| 3636 | // wasn't vectorized. | ||||||
| 3637 | // FIXME: Should not rely on getVPValue at this point. | ||||||
| 3638 | VPValue *Def = State.Plan->getVPValue(KV.first, true); | ||||||
| 3639 | if (!State.hasAnyVectorValue(Def)) | ||||||
| 3640 | continue; | ||||||
| 3641 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
| 3642 | Value *I = State.get(Def, Part); | ||||||
| 3643 | ZExtInst *Inst = dyn_cast<ZExtInst>(I); | ||||||
| 3644 | if (Inst && Inst->use_empty()) { | ||||||
| 3645 | Value *NewI = Inst->getOperand(0); | ||||||
| 3646 | Inst->eraseFromParent(); | ||||||
| 3647 | State.reset(Def, NewI, Part); | ||||||
| 3648 | } | ||||||
| 3649 | } | ||||||
| 3650 | } | ||||||
| 3651 | } | ||||||
| 3652 | |||||||
| 3653 | void InnerLoopVectorizer::fixVectorizedLoop(VPTransformState &State, | ||||||
| 3654 | VPlan &Plan) { | ||||||
| 3655 | // Insert truncates and extends for any truncated instructions as hints to | ||||||
| 3656 | // InstCombine. | ||||||
| 3657 | if (VF.isVector()) | ||||||
| 3658 | truncateToMinimalBitwidths(State); | ||||||
| 3659 | |||||||
| 3660 | // Fix widened non-induction PHIs by setting up the PHI operands. | ||||||
| 3661 | if (EnableVPlanNativePath) | ||||||
| 3662 | fixNonInductionPHIs(Plan, State); | ||||||
| 3663 | |||||||
| 3664 | // At this point every instruction in the original loop is widened to a | ||||||
| 3665 | // vector form. Now we need to fix the recurrences in the loop. These PHI | ||||||
| 3666 | // nodes are currently empty because we did not want to introduce cycles. | ||||||
| 3667 | // This is the second stage of vectorizing recurrences. | ||||||
| 3668 | fixCrossIterationPHIs(State); | ||||||
| 3669 | |||||||
| 3670 | // Forget the original basic block. | ||||||
| 3671 | PSE.getSE()->forgetLoop(OrigLoop); | ||||||
| 3672 | |||||||
| 3673 | VPBasicBlock *LatchVPBB = Plan.getVectorLoopRegion()->getExitingBasicBlock(); | ||||||
| 3674 | Loop *VectorLoop = LI->getLoopFor(State.CFG.VPBB2IRBB[LatchVPBB]); | ||||||
| 3675 | if (Cost->requiresScalarEpilogue(VF)) { | ||||||
| 3676 | // No edge from the middle block to the unique exit block has been inserted | ||||||
| 3677 | // and there is nothing to fix from vector loop; phis should have incoming | ||||||
| 3678 | // from scalar loop only. | ||||||
| 3679 | Plan.clearLiveOuts(); | ||||||
| 3680 | } else { | ||||||
| 3681 | // If we inserted an edge from the middle block to the unique exit block, | ||||||
| 3682 | // update uses outside the loop (phis) to account for the newly inserted | ||||||
| 3683 | // edge. | ||||||
| 3684 | |||||||
| 3685 | // Fix-up external users of the induction variables. | ||||||
| 3686 | for (const auto &Entry : Legal->getInductionVars()) | ||||||
| 3687 | fixupIVUsers(Entry.first, Entry.second, | ||||||
| 3688 | getOrCreateVectorTripCount(VectorLoop->getLoopPreheader()), | ||||||
| 3689 | IVEndValues[Entry.first], LoopMiddleBlock, | ||||||
| 3690 | VectorLoop->getHeader(), Plan); | ||||||
| 3691 | } | ||||||
| 3692 | |||||||
| 3693 | // Fix LCSSA phis not already fixed earlier. Extracts may need to be generated | ||||||
| 3694 | // in the exit block, so update the builder. | ||||||
| 3695 | State.Builder.SetInsertPoint(State.CFG.ExitBB->getFirstNonPHI()); | ||||||
| 3696 | for (const auto &KV : Plan.getLiveOuts()) | ||||||
| 3697 | KV.second->fixPhi(Plan, State); | ||||||
| 3698 | |||||||
| 3699 | for (Instruction *PI : PredicatedInstructions) | ||||||
| 3700 | sinkScalarOperands(&*PI); | ||||||
| 3701 | |||||||
| 3702 | // Remove redundant induction instructions. | ||||||
| 3703 | cse(VectorLoop->getHeader()); | ||||||
| 3704 | |||||||
| 3705 | // Set/update profile weights for the vector and remainder loops as original | ||||||
| 3706 | // loop iterations are now distributed among them. Note that original loop | ||||||
| 3707 | // represented by LoopScalarBody becomes remainder loop after vectorization. | ||||||
| 3708 | // | ||||||
| 3709 | // For cases like foldTailByMasking() and requiresScalarEpiloque() we may | ||||||
| 3710 | // end up getting slightly roughened result but that should be OK since | ||||||
| 3711 | // profile is not inherently precise anyway. Note also possible bypass of | ||||||
| 3712 | // vector code caused by legality checks is ignored, assigning all the weight | ||||||
| 3713 | // to the vector loop, optimistically. | ||||||
| 3714 | // | ||||||
| 3715 | // For scalable vectorization we can't know at compile time how many iterations | ||||||
| 3716 | // of the loop are handled in one vector iteration, so instead assume a pessimistic | ||||||
| 3717 | // vscale of '1'. | ||||||
| 3718 | setProfileInfoAfterUnrolling(LI->getLoopFor(LoopScalarBody), VectorLoop, | ||||||
| 3719 | LI->getLoopFor(LoopScalarBody), | ||||||
| 3720 | VF.getKnownMinValue() * UF); | ||||||
| 3721 | } | ||||||
| 3722 | |||||||
| 3723 | void InnerLoopVectorizer::fixCrossIterationPHIs(VPTransformState &State) { | ||||||
| 3724 | // In order to support recurrences we need to be able to vectorize Phi nodes. | ||||||
| 3725 | // Phi nodes have cycles, so we need to vectorize them in two stages. This is | ||||||
| 3726 | // stage #2: We now need to fix the recurrences by adding incoming edges to | ||||||
| 3727 | // the currently empty PHI nodes. At this point every instruction in the | ||||||
| 3728 | // original loop is widened to a vector form so we can use them to construct | ||||||
| 3729 | // the incoming edges. | ||||||
| 3730 | VPBasicBlock *Header = | ||||||
| 3731 | State.Plan->getVectorLoopRegion()->getEntryBasicBlock(); | ||||||
| 3732 | for (VPRecipeBase &R : Header->phis()) { | ||||||
| 3733 | if (auto *ReductionPhi = dyn_cast<VPReductionPHIRecipe>(&R)) | ||||||
| 3734 | fixReduction(ReductionPhi, State); | ||||||
| 3735 | else if (auto *FOR = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&R)) | ||||||
| 3736 | fixFixedOrderRecurrence(FOR, State); | ||||||
| 3737 | } | ||||||
| 3738 | } | ||||||
| 3739 | |||||||
| 3740 | void InnerLoopVectorizer::fixFixedOrderRecurrence( | ||||||
| 3741 | VPFirstOrderRecurrencePHIRecipe *PhiR, VPTransformState &State) { | ||||||
| 3742 | // This is the second phase of vectorizing first-order recurrences. An | ||||||
| 3743 | // overview of the transformation is described below. Suppose we have the | ||||||
| 3744 | // following loop. | ||||||
| 3745 | // | ||||||
| 3746 | // for (int i = 0; i < n; ++i) | ||||||
| 3747 | // b[i] = a[i] - a[i - 1]; | ||||||
| 3748 | // | ||||||
| 3749 | // There is a first-order recurrence on "a". For this loop, the shorthand | ||||||
| 3750 | // scalar IR looks like: | ||||||
| 3751 | // | ||||||
| 3752 | // scalar.ph: | ||||||
| 3753 | // s_init = a[-1] | ||||||
| 3754 | // br scalar.body | ||||||
| 3755 | // | ||||||
| 3756 | // scalar.body: | ||||||
| 3757 | // i = phi [0, scalar.ph], [i+1, scalar.body] | ||||||
| 3758 | // s1 = phi [s_init, scalar.ph], [s2, scalar.body] | ||||||
| 3759 | // s2 = a[i] | ||||||
| 3760 | // b[i] = s2 - s1 | ||||||
| 3761 | // br cond, scalar.body, ... | ||||||
| 3762 | // | ||||||
| 3763 | // In this example, s1 is a recurrence because it's value depends on the | ||||||
| 3764 | // previous iteration. In the first phase of vectorization, we created a | ||||||
| 3765 | // vector phi v1 for s1. We now complete the vectorization and produce the | ||||||
| 3766 | // shorthand vector IR shown below (for VF = 4, UF = 1). | ||||||
| 3767 | // | ||||||
| 3768 | // vector.ph: | ||||||
| 3769 | // v_init = vector(..., ..., ..., a[-1]) | ||||||
| 3770 | // br vector.body | ||||||
| 3771 | // | ||||||
| 3772 | // vector.body | ||||||
| 3773 | // i = phi [0, vector.ph], [i+4, vector.body] | ||||||
| 3774 | // v1 = phi [v_init, vector.ph], [v2, vector.body] | ||||||
| 3775 | // v2 = a[i, i+1, i+2, i+3]; | ||||||
| 3776 | // v3 = vector(v1(3), v2(0, 1, 2)) | ||||||
| 3777 | // b[i, i+1, i+2, i+3] = v2 - v3 | ||||||
| 3778 | // br cond, vector.body, middle.block | ||||||
| 3779 | // | ||||||
| 3780 | // middle.block: | ||||||
| 3781 | // x = v2(3) | ||||||
| 3782 | // br scalar.ph | ||||||
| 3783 | // | ||||||
| 3784 | // scalar.ph: | ||||||
| 3785 | // s_init = phi [x, middle.block], [a[-1], otherwise] | ||||||
| 3786 | // br scalar.body | ||||||
| 3787 | // | ||||||
| 3788 | // After execution completes the vector loop, we extract the next value of | ||||||
| 3789 | // the recurrence (x) to use as the initial value in the scalar loop. | ||||||
| 3790 | |||||||
| 3791 | // Extract the last vector element in the middle block. This will be the | ||||||
| 3792 | // initial value for the recurrence when jumping to the scalar loop. | ||||||
| 3793 | VPValue *PreviousDef = PhiR->getBackedgeValue(); | ||||||
| 3794 | Value *Incoming = State.get(PreviousDef, UF - 1); | ||||||
| 3795 | auto *ExtractForScalar = Incoming; | ||||||
| 3796 | auto *IdxTy = Builder.getInt32Ty(); | ||||||
| 3797 | if (VF.isVector()) { | ||||||
| 3798 | auto *One = ConstantInt::get(IdxTy, 1); | ||||||
| 3799 | Builder.SetInsertPoint(LoopMiddleBlock->getTerminator()); | ||||||
| 3800 | auto *RuntimeVF = getRuntimeVF(Builder, IdxTy, VF); | ||||||
| 3801 | auto *LastIdx = Builder.CreateSub(RuntimeVF, One); | ||||||
| 3802 | ExtractForScalar = Builder.CreateExtractElement(ExtractForScalar, LastIdx, | ||||||
| 3803 | "vector.recur.extract"); | ||||||
| 3804 | } | ||||||
| 3805 | // Extract the second last element in the middle block if the | ||||||
| 3806 | // Phi is used outside the loop. We need to extract the phi itself | ||||||
| 3807 | // and not the last element (the phi update in the current iteration). This | ||||||
| 3808 | // will be the value when jumping to the exit block from the LoopMiddleBlock, | ||||||
| 3809 | // when the scalar loop is not run at all. | ||||||
| 3810 | Value *ExtractForPhiUsedOutsideLoop = nullptr; | ||||||
| 3811 | if (VF.isVector()) { | ||||||
| 3812 | auto *RuntimeVF = getRuntimeVF(Builder, IdxTy, VF); | ||||||
| 3813 | auto *Idx = Builder.CreateSub(RuntimeVF, ConstantInt::get(IdxTy, 2)); | ||||||
| 3814 | ExtractForPhiUsedOutsideLoop = Builder.CreateExtractElement( | ||||||
| 3815 | Incoming, Idx, "vector.recur.extract.for.phi"); | ||||||
| 3816 | } else if (UF > 1) | ||||||
| 3817 | // When loop is unrolled without vectorizing, initialize | ||||||
| 3818 | // ExtractForPhiUsedOutsideLoop with the value just prior to unrolled value | ||||||
| 3819 | // of `Incoming`. This is analogous to the vectorized case above: extracting | ||||||
| 3820 | // the second last element when VF > 1. | ||||||
| 3821 | ExtractForPhiUsedOutsideLoop = State.get(PreviousDef, UF - 2); | ||||||
| 3822 | |||||||
| 3823 | // Fix the initial value of the original recurrence in the scalar loop. | ||||||
| 3824 | Builder.SetInsertPoint(&*LoopScalarPreHeader->begin()); | ||||||
| 3825 | PHINode *Phi = cast<PHINode>(PhiR->getUnderlyingValue()); | ||||||
| 3826 | auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init"); | ||||||
| 3827 | auto *ScalarInit = PhiR->getStartValue()->getLiveInIRValue(); | ||||||
| 3828 | for (auto *BB : predecessors(LoopScalarPreHeader)) { | ||||||
| 3829 | auto *Incoming = BB == LoopMiddleBlock ? ExtractForScalar : ScalarInit; | ||||||
| 3830 | Start->addIncoming(Incoming, BB); | ||||||
| 3831 | } | ||||||
| 3832 | |||||||
| 3833 | Phi->setIncomingValueForBlock(LoopScalarPreHeader, Start); | ||||||
| 3834 | Phi->setName("scalar.recur"); | ||||||
| 3835 | |||||||
| 3836 | // Finally, fix users of the recurrence outside the loop. The users will need | ||||||
| 3837 | // either the last value of the scalar recurrence or the last value of the | ||||||
| 3838 | // vector recurrence we extracted in the middle block. Since the loop is in | ||||||
| 3839 | // LCSSA form, we just need to find all the phi nodes for the original scalar | ||||||
| 3840 | // recurrence in the exit block, and then add an edge for the middle block. | ||||||
| 3841 | // Note that LCSSA does not imply single entry when the original scalar loop | ||||||
| 3842 | // had multiple exiting edges (as we always run the last iteration in the | ||||||
| 3843 | // scalar epilogue); in that case, there is no edge from middle to exit and | ||||||
| 3844 | // and thus no phis which needed updated. | ||||||
| 3845 | if (!Cost->requiresScalarEpilogue(VF)) | ||||||
| 3846 | for (PHINode &LCSSAPhi : LoopExitBlock->phis()) | ||||||
| 3847 | if (llvm::is_contained(LCSSAPhi.incoming_values(), Phi)) { | ||||||
| 3848 | LCSSAPhi.addIncoming(ExtractForPhiUsedOutsideLoop, LoopMiddleBlock); | ||||||
| 3849 | State.Plan->removeLiveOut(&LCSSAPhi); | ||||||
| 3850 | } | ||||||
| 3851 | } | ||||||
| 3852 | |||||||
| 3853 | void InnerLoopVectorizer::fixReduction(VPReductionPHIRecipe *PhiR, | ||||||
| 3854 | VPTransformState &State) { | ||||||
| 3855 | PHINode *OrigPhi = cast<PHINode>(PhiR->getUnderlyingValue()); | ||||||
| 3856 | // Get it's reduction variable descriptor. | ||||||
| 3857 | assert(Legal->isReductionVariable(OrigPhi) &&(static_cast <bool> (Legal->isReductionVariable(OrigPhi ) && "Unable to find the reduction variable") ? void ( 0) : __assert_fail ("Legal->isReductionVariable(OrigPhi) && \"Unable to find the reduction variable\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3858, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 3858 | "Unable to find the reduction variable")(static_cast <bool> (Legal->isReductionVariable(OrigPhi ) && "Unable to find the reduction variable") ? void ( 0) : __assert_fail ("Legal->isReductionVariable(OrigPhi) && \"Unable to find the reduction variable\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3858, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3859 | const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor(); | ||||||
| 3860 | |||||||
| 3861 | RecurKind RK = RdxDesc.getRecurrenceKind(); | ||||||
| 3862 | TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue(); | ||||||
| 3863 | Instruction *LoopExitInst = RdxDesc.getLoopExitInstr(); | ||||||
| 3864 | State.setDebugLocFromInst(ReductionStartValue); | ||||||
| 3865 | |||||||
| 3866 | VPValue *LoopExitInstDef = PhiR->getBackedgeValue(); | ||||||
| 3867 | // This is the vector-clone of the value that leaves the loop. | ||||||
| 3868 | Type *VecTy = State.get(LoopExitInstDef, 0)->getType(); | ||||||
| 3869 | |||||||
| 3870 | // Wrap flags are in general invalid after vectorization, clear them. | ||||||
| 3871 | clearReductionWrapFlags(PhiR, State); | ||||||
| 3872 | |||||||
| 3873 | // Before each round, move the insertion point right between | ||||||
| 3874 | // the PHIs and the values we are going to write. | ||||||
| 3875 | // This allows us to write both PHINodes and the extractelement | ||||||
| 3876 | // instructions. | ||||||
| 3877 | Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt()); | ||||||
| 3878 | |||||||
| 3879 | State.setDebugLocFromInst(LoopExitInst); | ||||||
| 3880 | |||||||
| 3881 | Type *PhiTy = OrigPhi->getType(); | ||||||
| 3882 | |||||||
| 3883 | VPBasicBlock *LatchVPBB = | ||||||
| 3884 | PhiR->getParent()->getEnclosingLoopRegion()->getExitingBasicBlock(); | ||||||
| 3885 | BasicBlock *VectorLoopLatch = State.CFG.VPBB2IRBB[LatchVPBB]; | ||||||
| 3886 | // If tail is folded by masking, the vector value to leave the loop should be | ||||||
| 3887 | // a Select choosing between the vectorized LoopExitInst and vectorized Phi, | ||||||
| 3888 | // instead of the former. For an inloop reduction the reduction will already | ||||||
| 3889 | // be predicated, and does not need to be handled here. | ||||||
| 3890 | if (Cost->foldTailByMasking() && !PhiR->isInLoop()) { | ||||||
| 3891 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
| 3892 | Value *VecLoopExitInst = State.get(LoopExitInstDef, Part); | ||||||
| 3893 | SelectInst *Sel = nullptr; | ||||||
| 3894 | for (User *U : VecLoopExitInst->users()) { | ||||||
| 3895 | if (isa<SelectInst>(U)) { | ||||||
| 3896 | assert(!Sel && "Reduction exit feeding two selects")(static_cast <bool> (!Sel && "Reduction exit feeding two selects" ) ? void (0) : __assert_fail ("!Sel && \"Reduction exit feeding two selects\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3896, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3897 | Sel = cast<SelectInst>(U); | ||||||
| 3898 | } else | ||||||
| 3899 | assert(isa<PHINode>(U) && "Reduction exit must feed Phi's or select")(static_cast <bool> (isa<PHINode>(U) && "Reduction exit must feed Phi's or select" ) ? void (0) : __assert_fail ("isa<PHINode>(U) && \"Reduction exit must feed Phi's or select\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3899, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3900 | } | ||||||
| 3901 | assert(Sel && "Reduction exit feeds no select")(static_cast <bool> (Sel && "Reduction exit feeds no select" ) ? void (0) : __assert_fail ("Sel && \"Reduction exit feeds no select\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3901, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3902 | State.reset(LoopExitInstDef, Sel, Part); | ||||||
| 3903 | |||||||
| 3904 | if (isa<FPMathOperator>(Sel)) | ||||||
| 3905 | Sel->setFastMathFlags(RdxDesc.getFastMathFlags()); | ||||||
| 3906 | |||||||
| 3907 | // If the target can create a predicated operator for the reduction at no | ||||||
| 3908 | // extra cost in the loop (for example a predicated vadd), it can be | ||||||
| 3909 | // cheaper for the select to remain in the loop than be sunk out of it, | ||||||
| 3910 | // and so use the select value for the phi instead of the old | ||||||
| 3911 | // LoopExitValue. | ||||||
| 3912 | if (PreferPredicatedReductionSelect || | ||||||
| 3913 | TTI->preferPredicatedReductionSelect( | ||||||
| 3914 | RdxDesc.getOpcode(), PhiTy, | ||||||
| 3915 | TargetTransformInfo::ReductionFlags())) { | ||||||
| 3916 | auto *VecRdxPhi = | ||||||
| 3917 | cast<PHINode>(State.get(PhiR, Part)); | ||||||
| 3918 | VecRdxPhi->setIncomingValueForBlock(VectorLoopLatch, Sel); | ||||||
| 3919 | } | ||||||
| 3920 | } | ||||||
| 3921 | } | ||||||
| 3922 | |||||||
| 3923 | // If the vector reduction can be performed in a smaller type, we truncate | ||||||
| 3924 | // then extend the loop exit value to enable InstCombine to evaluate the | ||||||
| 3925 | // entire expression in the smaller type. | ||||||
| 3926 | if (VF.isVector() && PhiTy != RdxDesc.getRecurrenceType()) { | ||||||
| 3927 | assert(!PhiR->isInLoop() && "Unexpected truncated inloop reduction!")(static_cast <bool> (!PhiR->isInLoop() && "Unexpected truncated inloop reduction!" ) ? void (0) : __assert_fail ("!PhiR->isInLoop() && \"Unexpected truncated inloop reduction!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3927, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 3928 | Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF); | ||||||
| 3929 | Builder.SetInsertPoint(VectorLoopLatch->getTerminator()); | ||||||
| 3930 | VectorParts RdxParts(UF); | ||||||
| 3931 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
| 3932 | RdxParts[Part] = State.get(LoopExitInstDef, Part); | ||||||
| 3933 | Value *Trunc = Builder.CreateTrunc(RdxParts[Part], RdxVecTy); | ||||||
| 3934 | Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy) | ||||||
| 3935 | : Builder.CreateZExt(Trunc, VecTy); | ||||||
| 3936 | for (User *U : llvm::make_early_inc_range(RdxParts[Part]->users())) | ||||||
| 3937 | if (U != Trunc) { | ||||||
| 3938 | U->replaceUsesOfWith(RdxParts[Part], Extnd); | ||||||
| 3939 | RdxParts[Part] = Extnd; | ||||||
| 3940 | } | ||||||
| 3941 | } | ||||||
| 3942 | Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt()); | ||||||
| 3943 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
| 3944 | RdxParts[Part] = Builder.CreateTrunc(RdxParts[Part], RdxVecTy); | ||||||
| 3945 | State.reset(LoopExitInstDef, RdxParts[Part], Part); | ||||||
| 3946 | } | ||||||
| 3947 | } | ||||||
| 3948 | |||||||
| 3949 | // Reduce all of the unrolled parts into a single vector. | ||||||
| 3950 | Value *ReducedPartRdx = State.get(LoopExitInstDef, 0); | ||||||
| 3951 | unsigned Op = RecurrenceDescriptor::getOpcode(RK); | ||||||
| 3952 | |||||||
| 3953 | // The middle block terminator has already been assigned a DebugLoc here (the | ||||||
| 3954 | // OrigLoop's single latch terminator). We want the whole middle block to | ||||||
| 3955 | // appear to execute on this line because: (a) it is all compiler generated, | ||||||
| 3956 | // (b) these instructions are always executed after evaluating the latch | ||||||
| 3957 | // conditional branch, and (c) other passes may add new predecessors which | ||||||
| 3958 | // terminate on this line. This is the easiest way to ensure we don't | ||||||
| 3959 | // accidentally cause an extra step back into the loop while debugging. | ||||||
| 3960 | State.setDebugLocFromInst(LoopMiddleBlock->getTerminator()); | ||||||
| 3961 | if (PhiR->isOrdered()) | ||||||
| 3962 | ReducedPartRdx = State.get(LoopExitInstDef, UF - 1); | ||||||
| 3963 | else { | ||||||
| 3964 | // Floating-point operations should have some FMF to enable the reduction. | ||||||
| 3965 | IRBuilderBase::FastMathFlagGuard FMFG(Builder); | ||||||
| 3966 | Builder.setFastMathFlags(RdxDesc.getFastMathFlags()); | ||||||
| 3967 | for (unsigned Part = 1; Part < UF; ++Part) { | ||||||
| 3968 | Value *RdxPart = State.get(LoopExitInstDef, Part); | ||||||
| 3969 | if (Op != Instruction::ICmp && Op != Instruction::FCmp) { | ||||||
| 3970 | ReducedPartRdx = Builder.CreateBinOp( | ||||||
| 3971 | (Instruction::BinaryOps)Op, RdxPart, ReducedPartRdx, "bin.rdx"); | ||||||
| 3972 | } else if (RecurrenceDescriptor::isSelectCmpRecurrenceKind(RK)) | ||||||
| 3973 | ReducedPartRdx = createSelectCmpOp(Builder, ReductionStartValue, RK, | ||||||
| 3974 | ReducedPartRdx, RdxPart); | ||||||
| 3975 | else | ||||||
| 3976 | ReducedPartRdx = createMinMaxOp(Builder, RK, ReducedPartRdx, RdxPart); | ||||||
| 3977 | } | ||||||
| 3978 | } | ||||||
| 3979 | |||||||
| 3980 | // Create the reduction after the loop. Note that inloop reductions create the | ||||||
| 3981 | // target reduction in the loop using a Reduction recipe. | ||||||
| 3982 | if (VF.isVector() && !PhiR->isInLoop()) { | ||||||
| 3983 | ReducedPartRdx = | ||||||
| 3984 | createTargetReduction(Builder, TTI, RdxDesc, ReducedPartRdx, OrigPhi); | ||||||
| 3985 | // If the reduction can be performed in a smaller type, we need to extend | ||||||
| 3986 | // the reduction to the wider type before we branch to the original loop. | ||||||
| 3987 | if (PhiTy != RdxDesc.getRecurrenceType()) | ||||||
| 3988 | ReducedPartRdx = RdxDesc.isSigned() | ||||||
| 3989 | ? Builder.CreateSExt(ReducedPartRdx, PhiTy) | ||||||
| 3990 | : Builder.CreateZExt(ReducedPartRdx, PhiTy); | ||||||
| 3991 | } | ||||||
| 3992 | |||||||
| 3993 | PHINode *ResumePhi = | ||||||
| 3994 | dyn_cast<PHINode>(PhiR->getStartValue()->getUnderlyingValue()); | ||||||
| 3995 | |||||||
| 3996 | // Create a phi node that merges control-flow from the backedge-taken check | ||||||
| 3997 | // block and the middle block. | ||||||
| 3998 | PHINode *BCBlockPhi = PHINode::Create(PhiTy, 2, "bc.merge.rdx", | ||||||
| 3999 | LoopScalarPreHeader->getTerminator()); | ||||||
| 4000 | |||||||
| 4001 | // If we are fixing reductions in the epilogue loop then we should already | ||||||
| 4002 | // have created a bc.merge.rdx Phi after the main vector body. Ensure that | ||||||
| 4003 | // we carry over the incoming values correctly. | ||||||
| 4004 | for (auto *Incoming : predecessors(LoopScalarPreHeader)) { | ||||||
| 4005 | if (Incoming == LoopMiddleBlock) | ||||||
| 4006 | BCBlockPhi->addIncoming(ReducedPartRdx, Incoming); | ||||||
| 4007 | else if (ResumePhi && llvm::is_contained(ResumePhi->blocks(), Incoming)) | ||||||
| 4008 | BCBlockPhi->addIncoming(ResumePhi->getIncomingValueForBlock(Incoming), | ||||||
| 4009 | Incoming); | ||||||
| 4010 | else | ||||||
| 4011 | BCBlockPhi->addIncoming(ReductionStartValue, Incoming); | ||||||
| 4012 | } | ||||||
| 4013 | |||||||
| 4014 | // Set the resume value for this reduction | ||||||
| 4015 | ReductionResumeValues.insert({&RdxDesc, BCBlockPhi}); | ||||||
| 4016 | |||||||
| 4017 | // If there were stores of the reduction value to a uniform memory address | ||||||
| 4018 | // inside the loop, create the final store here. | ||||||
| 4019 | if (StoreInst *SI = RdxDesc.IntermediateStore) { | ||||||
| 4020 | StoreInst *NewSI = | ||||||
| 4021 | Builder.CreateStore(ReducedPartRdx, SI->getPointerOperand()); | ||||||
| 4022 | propagateMetadata(NewSI, SI); | ||||||
| 4023 | |||||||
| 4024 | // If the reduction value is used in other places, | ||||||
| 4025 | // then let the code below create PHI's for that. | ||||||
| 4026 | } | ||||||
| 4027 | |||||||
| 4028 | // Now, we need to fix the users of the reduction variable | ||||||
| 4029 | // inside and outside of the scalar remainder loop. | ||||||
| 4030 | |||||||
| 4031 | // We know that the loop is in LCSSA form. We need to update the PHI nodes | ||||||
| 4032 | // in the exit blocks. See comment on analogous loop in | ||||||
| 4033 | // fixFixedOrderRecurrence for a more complete explaination of the logic. | ||||||
| 4034 | if (!Cost->requiresScalarEpilogue(VF)) | ||||||
| 4035 | for (PHINode &LCSSAPhi : LoopExitBlock->phis()) | ||||||
| 4036 | if (llvm::is_contained(LCSSAPhi.incoming_values(), LoopExitInst)) { | ||||||
| 4037 | LCSSAPhi.addIncoming(ReducedPartRdx, LoopMiddleBlock); | ||||||
| 4038 | State.Plan->removeLiveOut(&LCSSAPhi); | ||||||
| 4039 | } | ||||||
| 4040 | |||||||
| 4041 | // Fix the scalar loop reduction variable with the incoming reduction sum | ||||||
| 4042 | // from the vector body and from the backedge value. | ||||||
| 4043 | int IncomingEdgeBlockIdx = | ||||||
| 4044 | OrigPhi->getBasicBlockIndex(OrigLoop->getLoopLatch()); | ||||||
| 4045 | assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index")(static_cast <bool> (IncomingEdgeBlockIdx >= 0 && "Invalid block index") ? void (0) : __assert_fail ("IncomingEdgeBlockIdx >= 0 && \"Invalid block index\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4045, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 4046 | // Pick the other block. | ||||||
| 4047 | int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); | ||||||
| 4048 | OrigPhi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi); | ||||||
| 4049 | OrigPhi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst); | ||||||
| 4050 | } | ||||||
| 4051 | |||||||
| 4052 | void InnerLoopVectorizer::clearReductionWrapFlags(VPReductionPHIRecipe *PhiR, | ||||||
| 4053 | VPTransformState &State) { | ||||||
| 4054 | const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor(); | ||||||
| 4055 | RecurKind RK = RdxDesc.getRecurrenceKind(); | ||||||
| 4056 | if (RK != RecurKind::Add && RK != RecurKind::Mul) | ||||||
| 4057 | return; | ||||||
| 4058 | |||||||
| 4059 | SmallVector<VPValue *, 8> Worklist; | ||||||
| 4060 | SmallPtrSet<VPValue *, 8> Visited; | ||||||
| 4061 | Worklist.push_back(PhiR); | ||||||
| 4062 | Visited.insert(PhiR); | ||||||
| 4063 | |||||||
| 4064 | while (!Worklist.empty()) { | ||||||
| 4065 | VPValue *Cur = Worklist.pop_back_val(); | ||||||
| 4066 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
| 4067 | Value *V = State.get(Cur, Part); | ||||||
| 4068 | if (!isa<OverflowingBinaryOperator>(V)) | ||||||
| 4069 | break; | ||||||
| 4070 | cast<Instruction>(V)->dropPoisonGeneratingFlags(); | ||||||
| 4071 | } | ||||||
| 4072 | |||||||
| 4073 | for (VPUser *U : Cur->users()) { | ||||||
| 4074 | auto *UserRecipe = dyn_cast<VPRecipeBase>(U); | ||||||
| 4075 | if (!UserRecipe) | ||||||
| 4076 | continue; | ||||||
| 4077 | for (VPValue *V : UserRecipe->definedValues()) | ||||||
| 4078 | if (Visited.insert(V).second) | ||||||
| 4079 | Worklist.push_back(V); | ||||||
| 4080 | } | ||||||
| 4081 | } | ||||||
| 4082 | } | ||||||
| 4083 | |||||||
| 4084 | void InnerLoopVectorizer::sinkScalarOperands(Instruction *PredInst) { | ||||||
| 4085 | // The basic block and loop containing the predicated instruction. | ||||||
| 4086 | auto *PredBB = PredInst->getParent(); | ||||||
| 4087 | auto *VectorLoop = LI->getLoopFor(PredBB); | ||||||
| 4088 | |||||||
| 4089 | // Initialize a worklist with the operands of the predicated instruction. | ||||||
| 4090 | SetVector<Value *> Worklist(PredInst->op_begin(), PredInst->op_end()); | ||||||
| 4091 | |||||||
| 4092 | // Holds instructions that we need to analyze again. An instruction may be | ||||||
| 4093 | // reanalyzed if we don't yet know if we can sink it or not. | ||||||
| 4094 | SmallVector<Instruction *, 8> InstsToReanalyze; | ||||||
| 4095 | |||||||
| 4096 | // Returns true if a given use occurs in the predicated block. Phi nodes use | ||||||
| 4097 | // their operands in their corresponding predecessor blocks. | ||||||
| 4098 | auto isBlockOfUsePredicated = [&](Use &U) -> bool { | ||||||
| 4099 | auto *I = cast<Instruction>(U.getUser()); | ||||||
| 4100 | BasicBlock *BB = I->getParent(); | ||||||
| 4101 | if (auto *Phi = dyn_cast<PHINode>(I)) | ||||||
| 4102 | BB = Phi->getIncomingBlock( | ||||||
| 4103 | PHINode::getIncomingValueNumForOperand(U.getOperandNo())); | ||||||
| 4104 | return BB == PredBB; | ||||||
| 4105 | }; | ||||||
| 4106 | |||||||
| 4107 | // Iteratively sink the scalarized operands of the predicated instruction | ||||||
| 4108 | // into the block we created for it. When an instruction is sunk, it's | ||||||
| 4109 | // operands are then added to the worklist. The algorithm ends after one pass | ||||||
| 4110 | // through the worklist doesn't sink a single instruction. | ||||||
| 4111 | bool Changed; | ||||||
| 4112 | do { | ||||||
| 4113 | // Add the instructions that need to be reanalyzed to the worklist, and | ||||||
| 4114 | // reset the changed indicator. | ||||||
| 4115 | Worklist.insert(InstsToReanalyze.begin(), InstsToReanalyze.end()); | ||||||
| 4116 | InstsToReanalyze.clear(); | ||||||
| 4117 | Changed = false; | ||||||
| 4118 | |||||||
| 4119 | while (!Worklist.empty()) { | ||||||
| 4120 | auto *I = dyn_cast<Instruction>(Worklist.pop_back_val()); | ||||||
| 4121 | |||||||
| 4122 | // We can't sink an instruction if it is a phi node, is not in the loop, | ||||||
| 4123 | // or may have side effects. | ||||||
| 4124 | if (!I || isa<PHINode>(I) || !VectorLoop->contains(I) || | ||||||
| 4125 | I->mayHaveSideEffects()) | ||||||
| 4126 | continue; | ||||||
| 4127 | |||||||
| 4128 | // If the instruction is already in PredBB, check if we can sink its | ||||||
| 4129 | // operands. In that case, VPlan's sinkScalarOperands() succeeded in | ||||||
| 4130 | // sinking the scalar instruction I, hence it appears in PredBB; but it | ||||||
| 4131 | // may have failed to sink I's operands (recursively), which we try | ||||||
| 4132 | // (again) here. | ||||||
| 4133 | if (I->getParent() == PredBB) { | ||||||
| 4134 | Worklist.insert(I->op_begin(), I->op_end()); | ||||||
| 4135 | continue; | ||||||
| 4136 | } | ||||||
| 4137 | |||||||
| 4138 | // It's legal to sink the instruction if all its uses occur in the | ||||||
| 4139 | // predicated block. Otherwise, there's nothing to do yet, and we may | ||||||
| 4140 | // need to reanalyze the instruction. | ||||||
| 4141 | if (!llvm::all_of(I->uses(), isBlockOfUsePredicated)) { | ||||||
| 4142 | InstsToReanalyze.push_back(I); | ||||||
| 4143 | continue; | ||||||
| 4144 | } | ||||||
| 4145 | |||||||
| 4146 | // Move the instruction to the beginning of the predicated block, and add | ||||||
| 4147 | // it's operands to the worklist. | ||||||
| 4148 | I->moveBefore(&*PredBB->getFirstInsertionPt()); | ||||||
| 4149 | Worklist.insert(I->op_begin(), I->op_end()); | ||||||
| 4150 | |||||||
| 4151 | // The sinking may have enabled other instructions to be sunk, so we will | ||||||
| 4152 | // need to iterate. | ||||||
| 4153 | Changed = true; | ||||||
| 4154 | } | ||||||
| 4155 | } while (Changed); | ||||||
| 4156 | } | ||||||
| 4157 | |||||||
| 4158 | void InnerLoopVectorizer::fixNonInductionPHIs(VPlan &Plan, | ||||||
| 4159 | VPTransformState &State) { | ||||||
| 4160 | auto Iter = depth_first( | ||||||
| 4161 | VPBlockRecursiveTraversalWrapper<VPBlockBase *>(Plan.getEntry())); | ||||||
| 4162 | for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Iter)) { | ||||||
| 4163 | for (VPRecipeBase &P : VPBB->phis()) { | ||||||
| 4164 | VPWidenPHIRecipe *VPPhi = dyn_cast<VPWidenPHIRecipe>(&P); | ||||||
| 4165 | if (!VPPhi) | ||||||
| 4166 | continue; | ||||||
| 4167 | PHINode *NewPhi = cast<PHINode>(State.get(VPPhi, 0)); | ||||||
| 4168 | // Make sure the builder has a valid insert point. | ||||||
| 4169 | Builder.SetInsertPoint(NewPhi); | ||||||
| 4170 | for (unsigned i = 0; i < VPPhi->getNumOperands(); ++i) { | ||||||
| 4171 | VPValue *Inc = VPPhi->getIncomingValue(i); | ||||||
| 4172 | VPBasicBlock *VPBB = VPPhi->getIncomingBlock(i); | ||||||
| 4173 | NewPhi->addIncoming(State.get(Inc, 0), State.CFG.VPBB2IRBB[VPBB]); | ||||||
| 4174 | } | ||||||
| 4175 | } | ||||||
| 4176 | } | ||||||
| 4177 | } | ||||||
| 4178 | |||||||
| 4179 | bool InnerLoopVectorizer::useOrderedReductions( | ||||||
| 4180 | const RecurrenceDescriptor &RdxDesc) { | ||||||
| 4181 | return Cost->useOrderedReductions(RdxDesc); | ||||||
| 4182 | } | ||||||
| 4183 | |||||||
| 4184 | void LoopVectorizationCostModel::collectLoopScalars(ElementCount VF) { | ||||||
| 4185 | // We should not collect Scalars more than once per VF. Right now, this | ||||||
| 4186 | // function is called from collectUniformsAndScalars(), which already does | ||||||
| 4187 | // this check. Collecting Scalars for VF=1 does not make any sense. | ||||||
| 4188 | assert(VF.isVector() && Scalars.find(VF) == Scalars.end() &&(static_cast <bool> (VF.isVector() && Scalars.find (VF) == Scalars.end() && "This function should not be visited twice for the same VF" ) ? void (0) : __assert_fail ("VF.isVector() && Scalars.find(VF) == Scalars.end() && \"This function should not be visited twice for the same VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4189, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 4189 | "This function should not be visited twice for the same VF")(static_cast <bool> (VF.isVector() && Scalars.find (VF) == Scalars.end() && "This function should not be visited twice for the same VF" ) ? void (0) : __assert_fail ("VF.isVector() && Scalars.find(VF) == Scalars.end() && \"This function should not be visited twice for the same VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4189, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 4190 | |||||||
| 4191 | // This avoids any chances of creating a REPLICATE recipe during planning | ||||||
| 4192 | // since that would result in generation of scalarized code during execution, | ||||||
| 4193 | // which is not supported for scalable vectors. | ||||||
| 4194 | if (VF.isScalable()) { | ||||||
| 4195 | Scalars[VF].insert(Uniforms[VF].begin(), Uniforms[VF].end()); | ||||||
| 4196 | return; | ||||||
| 4197 | } | ||||||
| 4198 | |||||||
| 4199 | SmallSetVector<Instruction *, 8> Worklist; | ||||||
| 4200 | |||||||
| 4201 | // These sets are used to seed the analysis with pointers used by memory | ||||||
| 4202 | // accesses that will remain scalar. | ||||||
| 4203 | SmallSetVector<Instruction *, 8> ScalarPtrs; | ||||||
| 4204 | SmallPtrSet<Instruction *, 8> PossibleNonScalarPtrs; | ||||||
| 4205 | auto *Latch = TheLoop->getLoopLatch(); | ||||||
| 4206 | |||||||
| 4207 | // A helper that returns true if the use of Ptr by MemAccess will be scalar. | ||||||
| 4208 | // The pointer operands of loads and stores will be scalar as long as the | ||||||
| 4209 | // memory access is not a gather or scatter operation. The value operand of a | ||||||
| 4210 | // store will remain scalar if the store is scalarized. | ||||||
| 4211 | auto isScalarUse = [&](Instruction *MemAccess, Value *Ptr) { | ||||||
| 4212 | InstWidening WideningDecision = getWideningDecision(MemAccess, VF); | ||||||
| 4213 | assert(WideningDecision != CM_Unknown &&(static_cast <bool> (WideningDecision != CM_Unknown && "Widening decision should be ready at this moment") ? void ( 0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4214, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 4214 | "Widening decision should be ready at this moment")(static_cast <bool> (WideningDecision != CM_Unknown && "Widening decision should be ready at this moment") ? void ( 0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4214, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 4215 | if (auto *Store = dyn_cast<StoreInst>(MemAccess)) | ||||||
| 4216 | if (Ptr == Store->getValueOperand()) | ||||||
| 4217 | return WideningDecision == CM_Scalarize; | ||||||
| 4218 | assert(Ptr == getLoadStorePointerOperand(MemAccess) &&(static_cast <bool> (Ptr == getLoadStorePointerOperand( MemAccess) && "Ptr is neither a value or pointer operand" ) ? void (0) : __assert_fail ("Ptr == getLoadStorePointerOperand(MemAccess) && \"Ptr is neither a value or pointer operand\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4219, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 4219 | "Ptr is neither a value or pointer operand")(static_cast <bool> (Ptr == getLoadStorePointerOperand( MemAccess) && "Ptr is neither a value or pointer operand" ) ? void (0) : __assert_fail ("Ptr == getLoadStorePointerOperand(MemAccess) && \"Ptr is neither a value or pointer operand\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4219, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 4220 | return WideningDecision != CM_GatherScatter; | ||||||
| 4221 | }; | ||||||
| 4222 | |||||||
| 4223 | // A helper that returns true if the given value is a bitcast or | ||||||
| 4224 | // getelementptr instruction contained in the loop. | ||||||
| 4225 | auto isLoopVaryingBitCastOrGEP = [&](Value *V) { | ||||||
| 4226 | return ((isa<BitCastInst>(V) && V->getType()->isPointerTy()) || | ||||||
| 4227 | isa<GetElementPtrInst>(V)) && | ||||||
| 4228 | !TheLoop->isLoopInvariant(V); | ||||||
| 4229 | }; | ||||||
| 4230 | |||||||
| 4231 | // A helper that evaluates a memory access's use of a pointer. If the use will | ||||||
| 4232 | // be a scalar use and the pointer is only used by memory accesses, we place | ||||||
| 4233 | // the pointer in ScalarPtrs. Otherwise, the pointer is placed in | ||||||
| 4234 | // PossibleNonScalarPtrs. | ||||||
| 4235 | auto evaluatePtrUse = [&](Instruction *MemAccess, Value *Ptr) { | ||||||
| 4236 | // We only care about bitcast and getelementptr instructions contained in | ||||||
| 4237 | // the loop. | ||||||
| 4238 | if (!isLoopVaryingBitCastOrGEP(Ptr)) | ||||||
| 4239 | return; | ||||||
| 4240 | |||||||
| 4241 | // If the pointer has already been identified as scalar (e.g., if it was | ||||||
| 4242 | // also identified as uniform), there's nothing to do. | ||||||
| 4243 | auto *I = cast<Instruction>(Ptr); | ||||||
| 4244 | if (Worklist.count(I)) | ||||||
| 4245 | return; | ||||||
| 4246 | |||||||
| 4247 | // If the use of the pointer will be a scalar use, and all users of the | ||||||
| 4248 | // pointer are memory accesses, place the pointer in ScalarPtrs. Otherwise, | ||||||
| 4249 | // place the pointer in PossibleNonScalarPtrs. | ||||||
| 4250 | if (isScalarUse(MemAccess, Ptr) && llvm::all_of(I->users(), [&](User *U) { | ||||||
| 4251 | return isa<LoadInst>(U) || isa<StoreInst>(U); | ||||||
| 4252 | })) | ||||||
| 4253 | ScalarPtrs.insert(I); | ||||||
| 4254 | else | ||||||
| 4255 | PossibleNonScalarPtrs.insert(I); | ||||||
| 4256 | }; | ||||||
| 4257 | |||||||
| 4258 | // We seed the scalars analysis with three classes of instructions: (1) | ||||||
| 4259 | // instructions marked uniform-after-vectorization and (2) bitcast, | ||||||
| 4260 | // getelementptr and (pointer) phi instructions used by memory accesses | ||||||
| 4261 | // requiring a scalar use. | ||||||
| 4262 | // | ||||||
| 4263 | // (1) Add to the worklist all instructions that have been identified as | ||||||
| 4264 | // uniform-after-vectorization. | ||||||
| 4265 | Worklist.insert(Uniforms[VF].begin(), Uniforms[VF].end()); | ||||||
| 4266 | |||||||
| 4267 | // (2) Add to the worklist all bitcast and getelementptr instructions used by | ||||||
| 4268 | // memory accesses requiring a scalar use. The pointer operands of loads and | ||||||
| 4269 | // stores will be scalar as long as the memory accesses is not a gather or | ||||||
| 4270 | // scatter operation. The value operand of a store will remain scalar if the | ||||||
| 4271 | // store is scalarized. | ||||||
| 4272 | for (auto *BB : TheLoop->blocks()) | ||||||
| 4273 | for (auto &I : *BB) { | ||||||
| 4274 | if (auto *Load = dyn_cast<LoadInst>(&I)) { | ||||||
| 4275 | evaluatePtrUse(Load, Load->getPointerOperand()); | ||||||
| 4276 | } else if (auto *Store = dyn_cast<StoreInst>(&I)) { | ||||||
| 4277 | evaluatePtrUse(Store, Store->getPointerOperand()); | ||||||
| 4278 | evaluatePtrUse(Store, Store->getValueOperand()); | ||||||
| 4279 | } | ||||||
| 4280 | } | ||||||
| 4281 | for (auto *I : ScalarPtrs) | ||||||
| 4282 | if (!PossibleNonScalarPtrs.count(I)) { | ||||||
| 4283 | LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *I << "\n"; } } while (false); | ||||||
| 4284 | Worklist.insert(I); | ||||||
| 4285 | } | ||||||
| 4286 | |||||||
| 4287 | // Insert the forced scalars. | ||||||
| 4288 | // FIXME: Currently VPWidenPHIRecipe() often creates a dead vector | ||||||
| 4289 | // induction variable when the PHI user is scalarized. | ||||||
| 4290 | auto ForcedScalar = ForcedScalars.find(VF); | ||||||
| 4291 | if (ForcedScalar != ForcedScalars.end()) | ||||||
| 4292 | for (auto *I : ForcedScalar->second) { | ||||||
| 4293 | LLVM_DEBUG(dbgs() << "LV: Found (forced) scalar instruction: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found (forced) scalar instruction: " << *I << "\n"; } } while (false); | ||||||
| 4294 | Worklist.insert(I); | ||||||
| 4295 | } | ||||||
| 4296 | |||||||
| 4297 | // Expand the worklist by looking through any bitcasts and getelementptr | ||||||
| 4298 | // instructions we've already identified as scalar. This is similar to the | ||||||
| 4299 | // expansion step in collectLoopUniforms(); however, here we're only | ||||||
| 4300 | // expanding to include additional bitcasts and getelementptr instructions. | ||||||
| 4301 | unsigned Idx = 0; | ||||||
| 4302 | while (Idx != Worklist.size()) { | ||||||
| 4303 | Instruction *Dst = Worklist[Idx++]; | ||||||
| 4304 | if (!isLoopVaryingBitCastOrGEP(Dst->getOperand(0))) | ||||||
| 4305 | continue; | ||||||
| 4306 | auto *Src = cast<Instruction>(Dst->getOperand(0)); | ||||||
| 4307 | if (llvm::all_of(Src->users(), [&](User *U) -> bool { | ||||||
| 4308 | auto *J = cast<Instruction>(U); | ||||||
| 4309 | return !TheLoop->contains(J) || Worklist.count(J) || | ||||||
| 4310 | ((isa<LoadInst>(J) || isa<StoreInst>(J)) && | ||||||
| 4311 | isScalarUse(J, Src)); | ||||||
| 4312 | })) { | ||||||
| 4313 | Worklist.insert(Src); | ||||||
| 4314 | LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *Src << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *Src << "\n"; } } while (false); | ||||||
| 4315 | } | ||||||
| 4316 | } | ||||||
| 4317 | |||||||
| 4318 | // An induction variable will remain scalar if all users of the induction | ||||||
| 4319 | // variable and induction variable update remain scalar. | ||||||
| 4320 | for (const auto &Induction : Legal->getInductionVars()) { | ||||||
| 4321 | auto *Ind = Induction.first; | ||||||
| 4322 | auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); | ||||||
| 4323 | |||||||
| 4324 | // If tail-folding is applied, the primary induction variable will be used | ||||||
| 4325 | // to feed a vector compare. | ||||||
| 4326 | if (Ind == Legal->getPrimaryInduction() && foldTailByMasking()) | ||||||
| 4327 | continue; | ||||||
| 4328 | |||||||
| 4329 | // Returns true if \p Indvar is a pointer induction that is used directly by | ||||||
| 4330 | // load/store instruction \p I. | ||||||
| 4331 | auto IsDirectLoadStoreFromPtrIndvar = [&](Instruction *Indvar, | ||||||
| 4332 | Instruction *I) { | ||||||
| 4333 | return Induction.second.getKind() == | ||||||
| 4334 | InductionDescriptor::IK_PtrInduction && | ||||||
| 4335 | (isa<LoadInst>(I) || isa<StoreInst>(I)) && | ||||||
| 4336 | Indvar == getLoadStorePointerOperand(I) && isScalarUse(I, Indvar); | ||||||
| 4337 | }; | ||||||
| 4338 | |||||||
| 4339 | // Determine if all users of the induction variable are scalar after | ||||||
| 4340 | // vectorization. | ||||||
| 4341 | auto ScalarInd = llvm::all_of(Ind->users(), [&](User *U) -> bool { | ||||||
| 4342 | auto *I = cast<Instruction>(U); | ||||||
| 4343 | return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) || | ||||||
| 4344 | IsDirectLoadStoreFromPtrIndvar(Ind, I); | ||||||
| 4345 | }); | ||||||
| 4346 | if (!ScalarInd) | ||||||
| 4347 | continue; | ||||||
| 4348 | |||||||
| 4349 | // Determine if all users of the induction variable update instruction are | ||||||
| 4350 | // scalar after vectorization. | ||||||
| 4351 | auto ScalarIndUpdate = | ||||||
| 4352 | llvm::all_of(IndUpdate->users(), [&](User *U) -> bool { | ||||||
| 4353 | auto *I = cast<Instruction>(U); | ||||||
| 4354 | return I == Ind || !TheLoop->contains(I) || Worklist.count(I) || | ||||||
| 4355 | IsDirectLoadStoreFromPtrIndvar(IndUpdate, I); | ||||||
| 4356 | }); | ||||||
| 4357 | if (!ScalarIndUpdate) | ||||||
| 4358 | continue; | ||||||
| 4359 | |||||||
| 4360 | // The induction variable and its update instruction will remain scalar. | ||||||
| 4361 | Worklist.insert(Ind); | ||||||
| 4362 | Worklist.insert(IndUpdate); | ||||||
| 4363 | LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *Ind << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *Ind << "\n"; } } while (false); | ||||||
| 4364 | LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *IndUpdatedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n"; } } while (false) | ||||||
| 4365 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n"; } } while (false); | ||||||
| 4366 | } | ||||||
| 4367 | |||||||
| 4368 | Scalars[VF].insert(Worklist.begin(), Worklist.end()); | ||||||
| 4369 | } | ||||||
| 4370 | |||||||
| 4371 | bool LoopVectorizationCostModel::isScalarWithPredication( | ||||||
| 4372 | Instruction *I, ElementCount VF) const { | ||||||
| 4373 | if (!isPredicatedInst(I)) | ||||||
| 4374 | return false; | ||||||
| 4375 | |||||||
| 4376 | // Do we have a non-scalar lowering for this predicated | ||||||
| 4377 | // instruction? No - it is scalar with predication. | ||||||
| 4378 | switch(I->getOpcode()) { | ||||||
| 4379 | default: | ||||||
| 4380 | return true; | ||||||
| 4381 | case Instruction::Load: | ||||||
| 4382 | case Instruction::Store: { | ||||||
| 4383 | auto *Ptr = getLoadStorePointerOperand(I); | ||||||
| 4384 | auto *Ty = getLoadStoreType(I); | ||||||
| 4385 | Type *VTy = Ty; | ||||||
| 4386 | if (VF.isVector()) | ||||||
| 4387 | VTy = VectorType::get(Ty, VF); | ||||||
| 4388 | const Align Alignment = getLoadStoreAlignment(I); | ||||||
| 4389 | return isa<LoadInst>(I) ? !(isLegalMaskedLoad(Ty, Ptr, Alignment) || | ||||||
| 4390 | TTI.isLegalMaskedGather(VTy, Alignment)) | ||||||
| 4391 | : !(isLegalMaskedStore(Ty, Ptr, Alignment) || | ||||||
| 4392 | TTI.isLegalMaskedScatter(VTy, Alignment)); | ||||||
| 4393 | } | ||||||
| 4394 | case Instruction::UDiv: | ||||||
| 4395 | case Instruction::SDiv: | ||||||
| 4396 | case Instruction::SRem: | ||||||
| 4397 | case Instruction::URem: { | ||||||
| 4398 | // We have the option to use the safe-divisor idiom to avoid predication. | ||||||
| 4399 | // The cost based decision here will always select safe-divisor for | ||||||
| 4400 | // scalable vectors as scalarization isn't legal. | ||||||
| 4401 | const auto [ScalarCost, SafeDivisorCost] = getDivRemSpeculationCost(I, VF); | ||||||
| 4402 | return isDivRemScalarWithPredication(ScalarCost, SafeDivisorCost); | ||||||
| 4403 | } | ||||||
| 4404 | } | ||||||
| 4405 | } | ||||||
| 4406 | |||||||
| 4407 | bool LoopVectorizationCostModel::isPredicatedInst(Instruction *I) const { | ||||||
| 4408 | if (!blockNeedsPredicationForAnyReason(I->getParent())) | ||||||
| 4409 | return false; | ||||||
| 4410 | |||||||
| 4411 | // Can we prove this instruction is safe to unconditionally execute? | ||||||
| 4412 | // If not, we must use some form of predication. | ||||||
| 4413 | switch(I->getOpcode()) { | ||||||
| 4414 | default: | ||||||
| 4415 | return false; | ||||||
| 4416 | case Instruction::Load: | ||||||
| 4417 | case Instruction::Store: { | ||||||
| 4418 | if (!Legal->isMaskRequired(I)) | ||||||
| 4419 | return false; | ||||||
| 4420 | // When we know the load's address is loop invariant and the instruction | ||||||
| 4421 | // in the original scalar loop was unconditionally executed then we | ||||||
| 4422 | // don't need to mark it as a predicated instruction. Tail folding may | ||||||
| 4423 | // introduce additional predication, but we're guaranteed to always have | ||||||
| 4424 | // at least one active lane. We call Legal->blockNeedsPredication here | ||||||
| 4425 | // because it doesn't query tail-folding. For stores, we need to prove | ||||||
| 4426 | // both speculation safety (which follows from the same argument as loads), | ||||||
| 4427 | // but also must prove the value being stored is correct. The easiest | ||||||
| 4428 | // form of the later is to require that all values stored are the same. | ||||||
| 4429 | if (Legal->isUniformMemOp(*I) && | ||||||
| 4430 | (isa<LoadInst>(I) || | ||||||
| 4431 | (isa<StoreInst>(I) && | ||||||
| 4432 | TheLoop->isLoopInvariant(cast<StoreInst>(I)->getValueOperand()))) && | ||||||
| 4433 | !Legal->blockNeedsPredication(I->getParent())) | ||||||
| 4434 | return false; | ||||||
| 4435 | return true; | ||||||
| 4436 | } | ||||||
| 4437 | case Instruction::UDiv: | ||||||
| 4438 | case Instruction::SDiv: | ||||||
| 4439 | case Instruction::SRem: | ||||||
| 4440 | case Instruction::URem: | ||||||
| 4441 | // TODO: We can use the loop-preheader as context point here and get | ||||||
| 4442 | // context sensitive reasoning | ||||||
| 4443 | return !isSafeToSpeculativelyExecute(I); | ||||||
| 4444 | } | ||||||
| 4445 | } | ||||||
| 4446 | |||||||
| 4447 | std::pair<InstructionCost, InstructionCost> | ||||||
| 4448 | LoopVectorizationCostModel::getDivRemSpeculationCost(Instruction *I, | ||||||
| 4449 | ElementCount VF) const { | ||||||
| 4450 | assert(I->getOpcode() == Instruction::UDiv ||(static_cast <bool> (I->getOpcode() == Instruction:: UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode () == Instruction::SRem || I->getOpcode() == Instruction:: URem) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode() == Instruction::SRem || I->getOpcode() == Instruction::URem" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4453, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 4451 | I->getOpcode() == Instruction::SDiv ||(static_cast <bool> (I->getOpcode() == Instruction:: UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode () == Instruction::SRem || I->getOpcode() == Instruction:: URem) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode() == Instruction::SRem || I->getOpcode() == Instruction::URem" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4453, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 4452 | I->getOpcode() == Instruction::SRem ||(static_cast <bool> (I->getOpcode() == Instruction:: UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode () == Instruction::SRem || I->getOpcode() == Instruction:: URem) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode() == Instruction::SRem || I->getOpcode() == Instruction::URem" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4453, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 4453 | I->getOpcode() == Instruction::URem)(static_cast <bool> (I->getOpcode() == Instruction:: UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode () == Instruction::SRem || I->getOpcode() == Instruction:: URem) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode() == Instruction::SRem || I->getOpcode() == Instruction::URem" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4453, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 4454 | assert(!isSafeToSpeculativelyExecute(I))(static_cast <bool> (!isSafeToSpeculativelyExecute(I)) ? void (0) : __assert_fail ("!isSafeToSpeculativelyExecute(I)" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4454, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 4455 | |||||||
| 4456 | const TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||
| 4457 | |||||||
| 4458 | // Scalarization isn't legal for scalable vector types | ||||||
| 4459 | InstructionCost ScalarizationCost = InstructionCost::getInvalid(); | ||||||
| 4460 | if (!VF.isScalable()) { | ||||||
| 4461 | // Get the scalarization cost and scale this amount by the probability of | ||||||
| 4462 | // executing the predicated block. If the instruction is not predicated, | ||||||
| 4463 | // we fall through to the next case. | ||||||
| 4464 | ScalarizationCost = 0; | ||||||
| 4465 | |||||||
| 4466 | // These instructions have a non-void type, so account for the phi nodes | ||||||
| 4467 | // that we will create. This cost is likely to be zero. The phi node | ||||||
| 4468 | // cost, if any, should be scaled by the block probability because it | ||||||
| 4469 | // models a copy at the end of each predicated block. | ||||||
| 4470 | ScalarizationCost += VF.getKnownMinValue() * | ||||||
| 4471 | TTI.getCFInstrCost(Instruction::PHI, CostKind); | ||||||
| 4472 | |||||||
| 4473 | // The cost of the non-predicated instruction. | ||||||
| 4474 | ScalarizationCost += VF.getKnownMinValue() * | ||||||
| 4475 | TTI.getArithmeticInstrCost(I->getOpcode(), I->getType(), CostKind); | ||||||
| 4476 | |||||||
| 4477 | // The cost of insertelement and extractelement instructions needed for | ||||||
| 4478 | // scalarization. | ||||||
| 4479 | ScalarizationCost += getScalarizationOverhead(I, VF); | ||||||
| 4480 | |||||||
| 4481 | // Scale the cost by the probability of executing the predicated blocks. | ||||||
| 4482 | // This assumes the predicated block for each vector lane is equally | ||||||
| 4483 | // likely. | ||||||
| 4484 | ScalarizationCost = ScalarizationCost / getReciprocalPredBlockProb(); | ||||||
| 4485 | } | ||||||
| 4486 | InstructionCost SafeDivisorCost = 0; | ||||||
| 4487 | |||||||
| 4488 | auto *VecTy = ToVectorTy(I->getType(), VF); | ||||||
| 4489 | |||||||
| 4490 | // The cost of the select guard to ensure all lanes are well defined | ||||||
| 4491 | // after we speculate above any internal control flow. | ||||||
| 4492 | SafeDivisorCost += TTI.getCmpSelInstrCost( | ||||||
| 4493 | Instruction::Select, VecTy, | ||||||
| 4494 | ToVectorTy(Type::getInt1Ty(I->getContext()), VF), | ||||||
| 4495 | CmpInst::BAD_ICMP_PREDICATE, CostKind); | ||||||
| 4496 | |||||||
| 4497 | // Certain instructions can be cheaper to vectorize if they have a constant | ||||||
| 4498 | // second vector operand. One example of this are shifts on x86. | ||||||
| 4499 | Value *Op2 = I->getOperand(1); | ||||||
| 4500 | auto Op2Info = TTI.getOperandInfo(Op2); | ||||||
| 4501 | if (Op2Info.Kind == TargetTransformInfo::OK_AnyValue && Legal->isUniform(Op2)) | ||||||
| 4502 | Op2Info.Kind = TargetTransformInfo::OK_UniformValue; | ||||||
| 4503 | |||||||
| 4504 | SmallVector<const Value *, 4> Operands(I->operand_values()); | ||||||
| 4505 | SafeDivisorCost += TTI.getArithmeticInstrCost( | ||||||
| 4506 | I->getOpcode(), VecTy, CostKind, | ||||||
| 4507 | {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None}, | ||||||
| 4508 | Op2Info, Operands, I); | ||||||
| 4509 | return {ScalarizationCost, SafeDivisorCost}; | ||||||
| 4510 | } | ||||||
| 4511 | |||||||
| 4512 | bool LoopVectorizationCostModel::interleavedAccessCanBeWidened( | ||||||
| 4513 | Instruction *I, ElementCount VF) { | ||||||
| 4514 | assert(isAccessInterleaved(I) && "Expecting interleaved access.")(static_cast <bool> (isAccessInterleaved(I) && "Expecting interleaved access." ) ? void (0) : __assert_fail ("isAccessInterleaved(I) && \"Expecting interleaved access.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4514, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 4515 | assert(getWideningDecision(I, VF) == CM_Unknown &&(static_cast <bool> (getWideningDecision(I, VF) == CM_Unknown && "Decision should not be set yet.") ? void (0) : __assert_fail ("getWideningDecision(I, VF) == CM_Unknown && \"Decision should not be set yet.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4516, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 4516 | "Decision should not be set yet.")(static_cast <bool> (getWideningDecision(I, VF) == CM_Unknown && "Decision should not be set yet.") ? void (0) : __assert_fail ("getWideningDecision(I, VF) == CM_Unknown && \"Decision should not be set yet.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4516, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 4517 | auto *Group = getInterleavedAccessGroup(I); | ||||||
| 4518 | assert(Group && "Must have a group.")(static_cast <bool> (Group && "Must have a group." ) ? void (0) : __assert_fail ("Group && \"Must have a group.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4518, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 4519 | |||||||
| 4520 | // If the instruction's allocated size doesn't equal it's type size, it | ||||||
| 4521 | // requires padding and will be scalarized. | ||||||
| 4522 | auto &DL = I->getModule()->getDataLayout(); | ||||||
| 4523 | auto *ScalarTy = getLoadStoreType(I); | ||||||
| 4524 | if (hasIrregularType(ScalarTy, DL)) | ||||||
| 4525 | return false; | ||||||
| 4526 | |||||||
| 4527 | // If the group involves a non-integral pointer, we may not be able to | ||||||
| 4528 | // losslessly cast all values to a common type. | ||||||
| 4529 | unsigned InterleaveFactor = Group->getFactor(); | ||||||
| 4530 | bool ScalarNI = DL.isNonIntegralPointerType(ScalarTy); | ||||||
| 4531 | for (unsigned i = 0; i < InterleaveFactor; i++) { | ||||||
| 4532 | Instruction *Member = Group->getMember(i); | ||||||
| 4533 | if (!Member) | ||||||
| 4534 | continue; | ||||||
| 4535 | auto *MemberTy = getLoadStoreType(Member); | ||||||
| 4536 | bool MemberNI = DL.isNonIntegralPointerType(MemberTy); | ||||||
| 4537 | // Don't coerce non-integral pointers to integers or vice versa. | ||||||
| 4538 | if (MemberNI != ScalarNI) { | ||||||
| 4539 | // TODO: Consider adding special nullptr value case here | ||||||
| 4540 | return false; | ||||||
| 4541 | } else if (MemberNI && ScalarNI && | ||||||
| 4542 | ScalarTy->getPointerAddressSpace() != | ||||||
| 4543 | MemberTy->getPointerAddressSpace()) { | ||||||
| 4544 | return false; | ||||||
| 4545 | } | ||||||
| 4546 | } | ||||||
| 4547 | |||||||
| 4548 | // Check if masking is required. | ||||||
| 4549 | // A Group may need masking for one of two reasons: it resides in a block that | ||||||
| 4550 | // needs predication, or it was decided to use masking to deal with gaps | ||||||
| 4551 | // (either a gap at the end of a load-access that may result in a speculative | ||||||
| 4552 | // load, or any gaps in a store-access). | ||||||
| 4553 | bool PredicatedAccessRequiresMasking = | ||||||
| 4554 | blockNeedsPredicationForAnyReason(I->getParent()) && | ||||||
| 4555 | Legal->isMaskRequired(I); | ||||||
| 4556 | bool LoadAccessWithGapsRequiresEpilogMasking = | ||||||
| 4557 | isa<LoadInst>(I) && Group->requiresScalarEpilogue() && | ||||||
| 4558 | !isScalarEpilogueAllowed(); | ||||||
| 4559 | bool StoreAccessWithGapsRequiresMasking = | ||||||
| 4560 | isa<StoreInst>(I) && (Group->getNumMembers() < Group->getFactor()); | ||||||
| 4561 | if (!PredicatedAccessRequiresMasking && | ||||||
| 4562 | !LoadAccessWithGapsRequiresEpilogMasking && | ||||||
| 4563 | !StoreAccessWithGapsRequiresMasking) | ||||||
| 4564 | return true; | ||||||
| 4565 | |||||||
| 4566 | // If masked interleaving is required, we expect that the user/target had | ||||||
| 4567 | // enabled it, because otherwise it either wouldn't have been created or | ||||||
| 4568 | // it should have been invalidated by the CostModel. | ||||||
| 4569 | assert(useMaskedInterleavedAccesses(TTI) &&(static_cast <bool> (useMaskedInterleavedAccesses(TTI) && "Masked interleave-groups for predicated accesses are not enabled." ) ? void (0) : __assert_fail ("useMaskedInterleavedAccesses(TTI) && \"Masked interleave-groups for predicated accesses are not enabled.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4570, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 4570 | "Masked interleave-groups for predicated accesses are not enabled.")(static_cast <bool> (useMaskedInterleavedAccesses(TTI) && "Masked interleave-groups for predicated accesses are not enabled." ) ? void (0) : __assert_fail ("useMaskedInterleavedAccesses(TTI) && \"Masked interleave-groups for predicated accesses are not enabled.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4570, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 4571 | |||||||
| 4572 | if (Group->isReverse()) | ||||||
| 4573 | return false; | ||||||
| 4574 | |||||||
| 4575 | auto *Ty = getLoadStoreType(I); | ||||||
| 4576 | const Align Alignment = getLoadStoreAlignment(I); | ||||||
| 4577 | return isa<LoadInst>(I) ? TTI.isLegalMaskedLoad(Ty, Alignment) | ||||||
| 4578 | : TTI.isLegalMaskedStore(Ty, Alignment); | ||||||
| 4579 | } | ||||||
| 4580 | |||||||
| 4581 | bool LoopVectorizationCostModel::memoryInstructionCanBeWidened( | ||||||
| 4582 | Instruction *I, ElementCount VF) { | ||||||
| 4583 | // Get and ensure we have a valid memory instruction. | ||||||
| 4584 | assert((isa<LoadInst, StoreInst>(I)) && "Invalid memory instruction")(static_cast <bool> ((isa<LoadInst, StoreInst>(I) ) && "Invalid memory instruction") ? void (0) : __assert_fail ("(isa<LoadInst, StoreInst>(I)) && \"Invalid memory instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4584, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 4585 | |||||||
| 4586 | auto *Ptr = getLoadStorePointerOperand(I); | ||||||
| 4587 | auto *ScalarTy = getLoadStoreType(I); | ||||||
| 4588 | |||||||
| 4589 | // In order to be widened, the pointer should be consecutive, first of all. | ||||||
| 4590 | if (!Legal->isConsecutivePtr(ScalarTy, Ptr)) | ||||||
| 4591 | return false; | ||||||
| 4592 | |||||||
| 4593 | // If the instruction is a store located in a predicated block, it will be | ||||||
| 4594 | // scalarized. | ||||||
| 4595 | if (isScalarWithPredication(I, VF)) | ||||||
| 4596 | return false; | ||||||
| 4597 | |||||||
| 4598 | // If the instruction's allocated size doesn't equal it's type size, it | ||||||
| 4599 | // requires padding and will be scalarized. | ||||||
| 4600 | auto &DL = I->getModule()->getDataLayout(); | ||||||
| 4601 | if (hasIrregularType(ScalarTy, DL)) | ||||||
| 4602 | return false; | ||||||
| 4603 | |||||||
| 4604 | return true; | ||||||
| 4605 | } | ||||||
| 4606 | |||||||
| 4607 | void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) { | ||||||
| 4608 | // We should not collect Uniforms more than once per VF. Right now, | ||||||
| 4609 | // this function is called from collectUniformsAndScalars(), which | ||||||
| 4610 | // already does this check. Collecting Uniforms for VF=1 does not make any | ||||||
| 4611 | // sense. | ||||||
| 4612 | |||||||
| 4613 | assert(VF.isVector() && Uniforms.find(VF) == Uniforms.end() &&(static_cast <bool> (VF.isVector() && Uniforms. find(VF) == Uniforms.end() && "This function should not be visited twice for the same VF" ) ? void (0) : __assert_fail ("VF.isVector() && Uniforms.find(VF) == Uniforms.end() && \"This function should not be visited twice for the same VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4614, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 4614 | "This function should not be visited twice for the same VF")(static_cast <bool> (VF.isVector() && Uniforms. find(VF) == Uniforms.end() && "This function should not be visited twice for the same VF" ) ? void (0) : __assert_fail ("VF.isVector() && Uniforms.find(VF) == Uniforms.end() && \"This function should not be visited twice for the same VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4614, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 4615 | |||||||
| 4616 | // Visit the list of Uniforms. If we'll not find any uniform value, we'll | ||||||
| 4617 | // not analyze again. Uniforms.count(VF) will return 1. | ||||||
| 4618 | Uniforms[VF].clear(); | ||||||
| 4619 | |||||||
| 4620 | // We now know that the loop is vectorizable! | ||||||
| 4621 | // Collect instructions inside the loop that will remain uniform after | ||||||
| 4622 | // vectorization. | ||||||
| 4623 | |||||||
| 4624 | // Global values, params and instructions outside of current loop are out of | ||||||
| 4625 | // scope. | ||||||
| 4626 | auto isOutOfScope = [&](Value *V) -> bool { | ||||||
| 4627 | Instruction *I = dyn_cast<Instruction>(V); | ||||||
| 4628 | return (!I || !TheLoop->contains(I)); | ||||||
| 4629 | }; | ||||||
| 4630 | |||||||
| 4631 | // Worklist containing uniform instructions demanding lane 0. | ||||||
| 4632 | SetVector<Instruction *> Worklist; | ||||||
| 4633 | BasicBlock *Latch = TheLoop->getLoopLatch(); | ||||||
| 4634 | |||||||
| 4635 | // Add uniform instructions demanding lane 0 to the worklist. Instructions | ||||||
| 4636 | // that are scalar with predication must not be considered uniform after | ||||||
| 4637 | // vectorization, because that would create an erroneous replicating region | ||||||
| 4638 | // where only a single instance out of VF should be formed. | ||||||
| 4639 | // TODO: optimize such seldom cases if found important, see PR40816. | ||||||
| 4640 | auto addToWorklistIfAllowed = [&](Instruction *I) -> void { | ||||||
| 4641 | if (isOutOfScope(I)) { | ||||||
| 4642 | LLVM_DEBUG(dbgs() << "LV: Found not uniform due to scope: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found not uniform due to scope: " << *I << "\n"; } } while (false) | ||||||
| 4643 | << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found not uniform due to scope: " << *I << "\n"; } } while (false); | ||||||
| 4644 | return; | ||||||
| 4645 | } | ||||||
| 4646 | if (isScalarWithPredication(I, VF)) { | ||||||
| 4647 | LLVM_DEBUG(dbgs() << "LV: Found not uniform being ScalarWithPredication: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found not uniform being ScalarWithPredication: " << *I << "\n"; } } while (false) | ||||||
| 4648 | << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found not uniform being ScalarWithPredication: " << *I << "\n"; } } while (false); | ||||||
| 4649 | return; | ||||||
| 4650 | } | ||||||
| 4651 | LLVM_DEBUG(dbgs() << "LV: Found uniform instruction: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found uniform instruction: " << *I << "\n"; } } while (false); | ||||||
| 4652 | Worklist.insert(I); | ||||||
| 4653 | }; | ||||||
| 4654 | |||||||
| 4655 | // Start with the conditional branch. If the branch condition is an | ||||||
| 4656 | // instruction contained in the loop that is only used by the branch, it is | ||||||
| 4657 | // uniform. | ||||||
| 4658 | auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0)); | ||||||
| 4659 | if (Cmp && TheLoop->contains(Cmp) && Cmp->hasOneUse()) | ||||||
| 4660 | addToWorklistIfAllowed(Cmp); | ||||||
| 4661 | |||||||
| 4662 | // Return true if all lanes perform the same memory operation, and we can | ||||||
| 4663 | // thus chose to execute only one. | ||||||
| 4664 | auto isUniformMemOpUse = [&](Instruction *I) { | ||||||
| 4665 | if (!Legal->isUniformMemOp(*I)) | ||||||
| 4666 | return false; | ||||||
| 4667 | if (isa<LoadInst>(I)) | ||||||
| 4668 | // Loading the same address always produces the same result - at least | ||||||
| 4669 | // assuming aliasing and ordering which have already been checked. | ||||||
| 4670 | return true; | ||||||
| 4671 | // Storing the same value on every iteration. | ||||||
| 4672 | return TheLoop->isLoopInvariant(cast<StoreInst>(I)->getValueOperand()); | ||||||
| 4673 | }; | ||||||
| 4674 | |||||||
| 4675 | auto isUniformDecision = [&](Instruction *I, ElementCount VF) { | ||||||
| 4676 | InstWidening WideningDecision = getWideningDecision(I, VF); | ||||||
| 4677 | assert(WideningDecision != CM_Unknown &&(static_cast <bool> (WideningDecision != CM_Unknown && "Widening decision should be ready at this moment") ? void ( 0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4678, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 4678 | "Widening decision should be ready at this moment")(static_cast <bool> (WideningDecision != CM_Unknown && "Widening decision should be ready at this moment") ? void ( 0) : __assert_fail ("WideningDecision != CM_Unknown && \"Widening decision should be ready at this moment\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4678, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 4679 | |||||||
| 4680 | if (isUniformMemOpUse(I)) | ||||||
| 4681 | return true; | ||||||
| 4682 | |||||||
| 4683 | return (WideningDecision == CM_Widen || | ||||||
| 4684 | WideningDecision == CM_Widen_Reverse || | ||||||
| 4685 | WideningDecision == CM_Interleave); | ||||||
| 4686 | }; | ||||||
| 4687 | |||||||
| 4688 | |||||||
| 4689 | // Returns true if Ptr is the pointer operand of a memory access instruction | ||||||
| 4690 | // I, and I is known to not require scalarization. | ||||||
| 4691 | auto isVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool { | ||||||
| 4692 | return getLoadStorePointerOperand(I) == Ptr && isUniformDecision(I, VF); | ||||||
| 4693 | }; | ||||||
| 4694 | |||||||
| 4695 | // Holds a list of values which are known to have at least one uniform use. | ||||||
| 4696 | // Note that there may be other uses which aren't uniform. A "uniform use" | ||||||
| 4697 | // here is something which only demands lane 0 of the unrolled iterations; | ||||||
| 4698 | // it does not imply that all lanes produce the same value (e.g. this is not | ||||||
| 4699 | // the usual meaning of uniform) | ||||||
| 4700 | SetVector<Value *> HasUniformUse; | ||||||
| 4701 | |||||||
| 4702 | // Scan the loop for instructions which are either a) known to have only | ||||||
| 4703 | // lane 0 demanded or b) are uses which demand only lane 0 of their operand. | ||||||
| 4704 | for (auto *BB : TheLoop->blocks()) | ||||||
| 4705 | for (auto &I : *BB) { | ||||||
| 4706 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I)) { | ||||||
| 4707 | switch (II->getIntrinsicID()) { | ||||||
| 4708 | case Intrinsic::sideeffect: | ||||||
| 4709 | case Intrinsic::experimental_noalias_scope_decl: | ||||||
| 4710 | case Intrinsic::assume: | ||||||
| 4711 | case Intrinsic::lifetime_start: | ||||||
| 4712 | case Intrinsic::lifetime_end: | ||||||
| 4713 | if (TheLoop->hasLoopInvariantOperands(&I)) | ||||||
| 4714 | addToWorklistIfAllowed(&I); | ||||||
| 4715 | break; | ||||||
| 4716 | default: | ||||||
| 4717 | break; | ||||||
| 4718 | } | ||||||
| 4719 | } | ||||||
| 4720 | |||||||
| 4721 | // ExtractValue instructions must be uniform, because the operands are | ||||||
| 4722 | // known to be loop-invariant. | ||||||
| 4723 | if (auto *EVI = dyn_cast<ExtractValueInst>(&I)) { | ||||||
| 4724 | assert(isOutOfScope(EVI->getAggregateOperand()) &&(static_cast <bool> (isOutOfScope(EVI->getAggregateOperand ()) && "Expected aggregate value to be loop invariant" ) ? void (0) : __assert_fail ("isOutOfScope(EVI->getAggregateOperand()) && \"Expected aggregate value to be loop invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4725, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 4725 | "Expected aggregate value to be loop invariant")(static_cast <bool> (isOutOfScope(EVI->getAggregateOperand ()) && "Expected aggregate value to be loop invariant" ) ? void (0) : __assert_fail ("isOutOfScope(EVI->getAggregateOperand()) && \"Expected aggregate value to be loop invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4725, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 4726 | addToWorklistIfAllowed(EVI); | ||||||
| 4727 | continue; | ||||||
| 4728 | } | ||||||
| 4729 | |||||||
| 4730 | // If there's no pointer operand, there's nothing to do. | ||||||
| 4731 | auto *Ptr = getLoadStorePointerOperand(&I); | ||||||
| 4732 | if (!Ptr) | ||||||
| 4733 | continue; | ||||||
| 4734 | |||||||
| 4735 | if (isUniformMemOpUse(&I)) | ||||||
| 4736 | addToWorklistIfAllowed(&I); | ||||||
| 4737 | |||||||
| 4738 | if (isUniformDecision(&I, VF)) { | ||||||
| 4739 | assert(isVectorizedMemAccessUse(&I, Ptr) && "consistency check")(static_cast <bool> (isVectorizedMemAccessUse(&I, Ptr ) && "consistency check") ? void (0) : __assert_fail ( "isVectorizedMemAccessUse(&I, Ptr) && \"consistency check\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4739, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 4740 | HasUniformUse.insert(Ptr); | ||||||
| 4741 | } | ||||||
| 4742 | } | ||||||
| 4743 | |||||||
| 4744 | // Add to the worklist any operands which have *only* uniform (e.g. lane 0 | ||||||
| 4745 | // demanding) users. Since loops are assumed to be in LCSSA form, this | ||||||
| 4746 | // disallows uses outside the loop as well. | ||||||
| 4747 | for (auto *V : HasUniformUse) { | ||||||
| 4748 | if (isOutOfScope(V)) | ||||||
| 4749 | continue; | ||||||
| 4750 | auto *I = cast<Instruction>(V); | ||||||
| 4751 | auto UsersAreMemAccesses = | ||||||
| 4752 | llvm::all_of(I->users(), [&](User *U) -> bool { | ||||||
| 4753 | return isVectorizedMemAccessUse(cast<Instruction>(U), V); | ||||||
| 4754 | }); | ||||||
| 4755 | if (UsersAreMemAccesses) | ||||||
| 4756 | addToWorklistIfAllowed(I); | ||||||
| 4757 | } | ||||||
| 4758 | |||||||
| 4759 | // Expand Worklist in topological order: whenever a new instruction | ||||||
| 4760 | // is added , its users should be already inside Worklist. It ensures | ||||||
| 4761 | // a uniform instruction will only be used by uniform instructions. | ||||||
| 4762 | unsigned idx = 0; | ||||||
| 4763 | while (idx != Worklist.size()) { | ||||||
| 4764 | Instruction *I = Worklist[idx++]; | ||||||
| 4765 | |||||||
| 4766 | for (auto *OV : I->operand_values()) { | ||||||
| 4767 | // isOutOfScope operands cannot be uniform instructions. | ||||||
| 4768 | if (isOutOfScope(OV)) | ||||||
| 4769 | continue; | ||||||
| 4770 | // First order recurrence Phi's should typically be considered | ||||||
| 4771 | // non-uniform. | ||||||
| 4772 | auto *OP = dyn_cast<PHINode>(OV); | ||||||
| 4773 | if (OP && Legal->isFixedOrderRecurrence(OP)) | ||||||
| 4774 | continue; | ||||||
| 4775 | // If all the users of the operand are uniform, then add the | ||||||
| 4776 | // operand into the uniform worklist. | ||||||
| 4777 | auto *OI = cast<Instruction>(OV); | ||||||
| 4778 | if (llvm::all_of(OI->users(), [&](User *U) -> bool { | ||||||
| 4779 | auto *J = cast<Instruction>(U); | ||||||
| 4780 | return Worklist.count(J) || isVectorizedMemAccessUse(J, OI); | ||||||
| 4781 | })) | ||||||
| 4782 | addToWorklistIfAllowed(OI); | ||||||
| 4783 | } | ||||||
| 4784 | } | ||||||
| 4785 | |||||||
| 4786 | // For an instruction to be added into Worklist above, all its users inside | ||||||
| 4787 | // the loop should also be in Worklist. However, this condition cannot be | ||||||
| 4788 | // true for phi nodes that form a cyclic dependence. We must process phi | ||||||
| 4789 | // nodes separately. An induction variable will remain uniform if all users | ||||||
| 4790 | // of the induction variable and induction variable update remain uniform. | ||||||
| 4791 | // The code below handles both pointer and non-pointer induction variables. | ||||||
| 4792 | for (const auto &Induction : Legal->getInductionVars()) { | ||||||
| 4793 | auto *Ind = Induction.first; | ||||||
| 4794 | auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); | ||||||
| 4795 | |||||||
| 4796 | // Determine if all users of the induction variable are uniform after | ||||||
| 4797 | // vectorization. | ||||||
| 4798 | auto UniformInd = llvm::all_of(Ind->users(), [&](User *U) -> bool { | ||||||
| 4799 | auto *I = cast<Instruction>(U); | ||||||
| 4800 | return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) || | ||||||
| 4801 | isVectorizedMemAccessUse(I, Ind); | ||||||
| 4802 | }); | ||||||
| 4803 | if (!UniformInd) | ||||||
| 4804 | continue; | ||||||
| 4805 | |||||||
| 4806 | // Determine if all users of the induction variable update instruction are | ||||||
| 4807 | // uniform after vectorization. | ||||||
| 4808 | auto UniformIndUpdate = | ||||||
| 4809 | llvm::all_of(IndUpdate->users(), [&](User *U) -> bool { | ||||||
| 4810 | auto *I = cast<Instruction>(U); | ||||||
| 4811 | return I == Ind || !TheLoop->contains(I) || Worklist.count(I) || | ||||||
| 4812 | isVectorizedMemAccessUse(I, IndUpdate); | ||||||
| 4813 | }); | ||||||
| 4814 | if (!UniformIndUpdate) | ||||||
| 4815 | continue; | ||||||
| 4816 | |||||||
| 4817 | // The induction variable and its update instruction will remain uniform. | ||||||
| 4818 | addToWorklistIfAllowed(Ind); | ||||||
| 4819 | addToWorklistIfAllowed(IndUpdate); | ||||||
| 4820 | } | ||||||
| 4821 | |||||||
| 4822 | Uniforms[VF].insert(Worklist.begin(), Worklist.end()); | ||||||
| 4823 | } | ||||||
| 4824 | |||||||
| 4825 | bool LoopVectorizationCostModel::runtimeChecksRequired() { | ||||||
| 4826 | LLVM_DEBUG(dbgs() << "LV: Performing code size checks.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Performing code size checks.\n" ; } } while (false); | ||||||
| 4827 | |||||||
| 4828 | if (Legal->getRuntimePointerChecking()->Need) { | ||||||
| 4829 | reportVectorizationFailure("Runtime ptr check is required with -Os/-Oz", | ||||||
| 4830 | "runtime pointer checks needed. Enable vectorization of this " | ||||||
| 4831 | "loop with '#pragma clang loop vectorize(enable)' when " | ||||||
| 4832 | "compiling with -Os/-Oz", | ||||||
| 4833 | "CantVersionLoopWithOptForSize", ORE, TheLoop); | ||||||
| 4834 | return true; | ||||||
| 4835 | } | ||||||
| 4836 | |||||||
| 4837 | if (!PSE.getPredicate().isAlwaysTrue()) { | ||||||
| 4838 | reportVectorizationFailure("Runtime SCEV check is required with -Os/-Oz", | ||||||
| 4839 | "runtime SCEV checks needed. Enable vectorization of this " | ||||||
| 4840 | "loop with '#pragma clang loop vectorize(enable)' when " | ||||||
| 4841 | "compiling with -Os/-Oz", | ||||||
| 4842 | "CantVersionLoopWithOptForSize", ORE, TheLoop); | ||||||
| 4843 | return true; | ||||||
| 4844 | } | ||||||
| 4845 | |||||||
| 4846 | // FIXME: Avoid specializing for stride==1 instead of bailing out. | ||||||
| 4847 | if (!Legal->getLAI()->getSymbolicStrides().empty()) { | ||||||
| 4848 | reportVectorizationFailure("Runtime stride check for small trip count", | ||||||
| 4849 | "runtime stride == 1 checks needed. Enable vectorization of " | ||||||
| 4850 | "this loop without such check by compiling with -Os/-Oz", | ||||||
| 4851 | "CantVersionLoopWithOptForSize", ORE, TheLoop); | ||||||
| 4852 | return true; | ||||||
| 4853 | } | ||||||
| 4854 | |||||||
| 4855 | return false; | ||||||
| 4856 | } | ||||||
| 4857 | |||||||
| 4858 | ElementCount | ||||||
| 4859 | LoopVectorizationCostModel::getMaxLegalScalableVF(unsigned MaxSafeElements) { | ||||||
| 4860 | if (!TTI.supportsScalableVectors() && !ForceTargetSupportsScalableVectors) | ||||||
| 4861 | return ElementCount::getScalable(0); | ||||||
| 4862 | |||||||
| 4863 | if (Hints->isScalableVectorizationDisabled()) { | ||||||
| 4864 | reportVectorizationInfo("Scalable vectorization is explicitly disabled", | ||||||
| 4865 | "ScalableVectorizationDisabled", ORE, TheLoop); | ||||||
| 4866 | return ElementCount::getScalable(0); | ||||||
| 4867 | } | ||||||
| 4868 | |||||||
| 4869 | LLVM_DEBUG(dbgs() << "LV: Scalable vectorization is available\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Scalable vectorization is available\n" ; } } while (false); | ||||||
| 4870 | |||||||
| 4871 | auto MaxScalableVF = ElementCount::getScalable( | ||||||
| 4872 | std::numeric_limits<ElementCount::ScalarTy>::max()); | ||||||
| 4873 | |||||||
| 4874 | // Test that the loop-vectorizer can legalize all operations for this MaxVF. | ||||||
| 4875 | // FIXME: While for scalable vectors this is currently sufficient, this should | ||||||
| 4876 | // be replaced by a more detailed mechanism that filters out specific VFs, | ||||||
| 4877 | // instead of invalidating vectorization for a whole set of VFs based on the | ||||||
| 4878 | // MaxVF. | ||||||
| 4879 | |||||||
| 4880 | // Disable scalable vectorization if the loop contains unsupported reductions. | ||||||
| 4881 | if (!canVectorizeReductions(MaxScalableVF)) { | ||||||
| 4882 | reportVectorizationInfo( | ||||||
| 4883 | "Scalable vectorization not supported for the reduction " | ||||||
| 4884 | "operations found in this loop.", | ||||||
| 4885 | "ScalableVFUnfeasible", ORE, TheLoop); | ||||||
| 4886 | return ElementCount::getScalable(0); | ||||||
| 4887 | } | ||||||
| 4888 | |||||||
| 4889 | // Disable scalable vectorization if the loop contains any instructions | ||||||
| 4890 | // with element types not supported for scalable vectors. | ||||||
| 4891 | if (any_of(ElementTypesInLoop, [&](Type *Ty) { | ||||||
| 4892 | return !Ty->isVoidTy() && | ||||||
| 4893 | !this->TTI.isElementTypeLegalForScalableVector(Ty); | ||||||
| 4894 | })) { | ||||||
| 4895 | reportVectorizationInfo("Scalable vectorization is not supported " | ||||||
| 4896 | "for all element types found in this loop.", | ||||||
| 4897 | "ScalableVFUnfeasible", ORE, TheLoop); | ||||||
| 4898 | return ElementCount::getScalable(0); | ||||||
| 4899 | } | ||||||
| 4900 | |||||||
| 4901 | if (Legal->isSafeForAnyVectorWidth()) | ||||||
| 4902 | return MaxScalableVF; | ||||||
| 4903 | |||||||
| 4904 | // Limit MaxScalableVF by the maximum safe dependence distance. | ||||||
| 4905 | std::optional<unsigned> MaxVScale = TTI.getMaxVScale(); | ||||||
| 4906 | if (!MaxVScale && TheFunction->hasFnAttribute(Attribute::VScaleRange)) | ||||||
| 4907 | MaxVScale = | ||||||
| 4908 | TheFunction->getFnAttribute(Attribute::VScaleRange).getVScaleRangeMax(); | ||||||
| 4909 | MaxScalableVF = | ||||||
| 4910 | ElementCount::getScalable(MaxVScale ? (MaxSafeElements / *MaxVScale) : 0); | ||||||
| 4911 | if (!MaxScalableVF) | ||||||
| 4912 | reportVectorizationInfo( | ||||||
| 4913 | "Max legal vector width too small, scalable vectorization " | ||||||
| 4914 | "unfeasible.", | ||||||
| 4915 | "ScalableVFUnfeasible", ORE, TheLoop); | ||||||
| 4916 | |||||||
| 4917 | return MaxScalableVF; | ||||||
| 4918 | } | ||||||
| 4919 | |||||||
| 4920 | FixedScalableVFPair LoopVectorizationCostModel::computeFeasibleMaxVF( | ||||||
| 4921 | unsigned ConstTripCount, ElementCount UserVF, bool FoldTailByMasking) { | ||||||
| 4922 | MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI); | ||||||
| 4923 | unsigned SmallestType, WidestType; | ||||||
| 4924 | std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes(); | ||||||
| 4925 | |||||||
| 4926 | // Get the maximum safe dependence distance in bits computed by LAA. | ||||||
| 4927 | // It is computed by MaxVF * sizeOf(type) * 8, where type is taken from | ||||||
| 4928 | // the memory accesses that is most restrictive (involved in the smallest | ||||||
| 4929 | // dependence distance). | ||||||
| 4930 | unsigned MaxSafeElements = | ||||||
| 4931 | PowerOf2Floor(Legal->getMaxSafeVectorWidthInBits() / WidestType); | ||||||
| 4932 | |||||||
| 4933 | auto MaxSafeFixedVF = ElementCount::getFixed(MaxSafeElements); | ||||||
| 4934 | auto MaxSafeScalableVF = getMaxLegalScalableVF(MaxSafeElements); | ||||||
| 4935 | |||||||
| 4936 | LLVM_DEBUG(dbgs() << "LV: The max safe fixed VF is: " << MaxSafeFixedVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The max safe fixed VF is: " << MaxSafeFixedVF << ".\n"; } } while (false) | ||||||
| 4937 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The max safe fixed VF is: " << MaxSafeFixedVF << ".\n"; } } while (false); | ||||||
| 4938 | LLVM_DEBUG(dbgs() << "LV: The max safe scalable VF is: " << MaxSafeScalableVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The max safe scalable VF is: " << MaxSafeScalableVF << ".\n"; } } while (false) | ||||||
| 4939 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The max safe scalable VF is: " << MaxSafeScalableVF << ".\n"; } } while (false); | ||||||
| 4940 | |||||||
| 4941 | // First analyze the UserVF, fall back if the UserVF should be ignored. | ||||||
| 4942 | if (UserVF) { | ||||||
| 4943 | auto MaxSafeUserVF = | ||||||
| 4944 | UserVF.isScalable() ? MaxSafeScalableVF : MaxSafeFixedVF; | ||||||
| 4945 | |||||||
| 4946 | if (ElementCount::isKnownLE(UserVF, MaxSafeUserVF)) { | ||||||
| 4947 | // If `VF=vscale x N` is safe, then so is `VF=N` | ||||||
| 4948 | if (UserVF.isScalable()) | ||||||
| 4949 | return FixedScalableVFPair( | ||||||
| 4950 | ElementCount::getFixed(UserVF.getKnownMinValue()), UserVF); | ||||||
| 4951 | else | ||||||
| 4952 | return UserVF; | ||||||
| 4953 | } | ||||||
| 4954 | |||||||
| 4955 | assert(ElementCount::isKnownGT(UserVF, MaxSafeUserVF))(static_cast <bool> (ElementCount::isKnownGT(UserVF, MaxSafeUserVF )) ? void (0) : __assert_fail ("ElementCount::isKnownGT(UserVF, MaxSafeUserVF)" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4955, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 4956 | |||||||
| 4957 | // Only clamp if the UserVF is not scalable. If the UserVF is scalable, it | ||||||
| 4958 | // is better to ignore the hint and let the compiler choose a suitable VF. | ||||||
| 4959 | if (!UserVF.isScalable()) { | ||||||
| 4960 | LLVM_DEBUG(dbgs() << "LV: User VF=" << UserVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: User VF=" << UserVF << " is unsafe, clamping to max safe VF=" << MaxSafeFixedVF << ".\n"; } } while (false) | ||||||
| 4961 | << " is unsafe, clamping to max safe VF="do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: User VF=" << UserVF << " is unsafe, clamping to max safe VF=" << MaxSafeFixedVF << ".\n"; } } while (false) | ||||||
| 4962 | << MaxSafeFixedVF << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: User VF=" << UserVF << " is unsafe, clamping to max safe VF=" << MaxSafeFixedVF << ".\n"; } } while (false); | ||||||
| 4963 | ORE->emit([&]() { | ||||||
| 4964 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationFactor", | ||||||
| 4965 | TheLoop->getStartLoc(), | ||||||
| 4966 | TheLoop->getHeader()) | ||||||
| 4967 | << "User-specified vectorization factor " | ||||||
| 4968 | << ore::NV("UserVectorizationFactor", UserVF) | ||||||
| 4969 | << " is unsafe, clamping to maximum safe vectorization factor " | ||||||
| 4970 | << ore::NV("VectorizationFactor", MaxSafeFixedVF); | ||||||
| 4971 | }); | ||||||
| 4972 | return MaxSafeFixedVF; | ||||||
| 4973 | } | ||||||
| 4974 | |||||||
| 4975 | if (!TTI.supportsScalableVectors() && !ForceTargetSupportsScalableVectors) { | ||||||
| 4976 | LLVM_DEBUG(dbgs() << "LV: User VF=" << UserVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: User VF=" << UserVF << " is ignored because scalable vectors are not " "available.\n"; } } while (false) | ||||||
| 4977 | << " is ignored because scalable vectors are not "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: User VF=" << UserVF << " is ignored because scalable vectors are not " "available.\n"; } } while (false) | ||||||
| 4978 | "available.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: User VF=" << UserVF << " is ignored because scalable vectors are not " "available.\n"; } } while (false); | ||||||
| 4979 | ORE->emit([&]() { | ||||||
| 4980 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationFactor", | ||||||
| 4981 | TheLoop->getStartLoc(), | ||||||
| 4982 | TheLoop->getHeader()) | ||||||
| 4983 | << "User-specified vectorization factor " | ||||||
| 4984 | << ore::NV("UserVectorizationFactor", UserVF) | ||||||
| 4985 | << " is ignored because the target does not support scalable " | ||||||
| 4986 | "vectors. The compiler will pick a more suitable value."; | ||||||
| 4987 | }); | ||||||
| 4988 | } else { | ||||||
| 4989 | LLVM_DEBUG(dbgs() << "LV: User VF=" << UserVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: User VF=" << UserVF << " is unsafe. Ignoring scalable UserVF.\n"; } } while (false) | ||||||
| 4990 | << " is unsafe. Ignoring scalable UserVF.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: User VF=" << UserVF << " is unsafe. Ignoring scalable UserVF.\n"; } } while (false); | ||||||
| 4991 | ORE->emit([&]() { | ||||||
| 4992 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationFactor", | ||||||
| 4993 | TheLoop->getStartLoc(), | ||||||
| 4994 | TheLoop->getHeader()) | ||||||
| 4995 | << "User-specified vectorization factor " | ||||||
| 4996 | << ore::NV("UserVectorizationFactor", UserVF) | ||||||
| 4997 | << " is unsafe. Ignoring the hint to let the compiler pick a " | ||||||
| 4998 | "more suitable value."; | ||||||
| 4999 | }); | ||||||
| 5000 | } | ||||||
| 5001 | } | ||||||
| 5002 | |||||||
| 5003 | LLVM_DEBUG(dbgs() << "LV: The Smallest and Widest types: " << SmallestTypedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The Smallest and Widest types: " << SmallestType << " / " << WidestType << " bits.\n"; } } while (false) | ||||||
| 5004 | << " / " << WidestType << " bits.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The Smallest and Widest types: " << SmallestType << " / " << WidestType << " bits.\n"; } } while (false); | ||||||
| 5005 | |||||||
| 5006 | FixedScalableVFPair Result(ElementCount::getFixed(1), | ||||||
| 5007 | ElementCount::getScalable(0)); | ||||||
| 5008 | if (auto MaxVF = | ||||||
| 5009 | getMaximizedVFForTarget(ConstTripCount, SmallestType, WidestType, | ||||||
| 5010 | MaxSafeFixedVF, FoldTailByMasking)) | ||||||
| 5011 | Result.FixedVF = MaxVF; | ||||||
| 5012 | |||||||
| 5013 | if (auto MaxVF = | ||||||
| 5014 | getMaximizedVFForTarget(ConstTripCount, SmallestType, WidestType, | ||||||
| 5015 | MaxSafeScalableVF, FoldTailByMasking)) | ||||||
| 5016 | if (MaxVF.isScalable()) { | ||||||
| 5017 | Result.ScalableVF = MaxVF; | ||||||
| 5018 | LLVM_DEBUG(dbgs() << "LV: Found feasible scalable VF = " << MaxVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found feasible scalable VF = " << MaxVF << "\n"; } } while (false) | ||||||
| 5019 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found feasible scalable VF = " << MaxVF << "\n"; } } while (false); | ||||||
| 5020 | } | ||||||
| 5021 | |||||||
| 5022 | return Result; | ||||||
| 5023 | } | ||||||
| 5024 | |||||||
| 5025 | FixedScalableVFPair | ||||||
| 5026 | LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) { | ||||||
| 5027 | if (Legal->getRuntimePointerChecking()->Need && TTI.hasBranchDivergence()) { | ||||||
| 5028 | // TODO: It may by useful to do since it's still likely to be dynamically | ||||||
| 5029 | // uniform if the target can skip. | ||||||
| 5030 | reportVectorizationFailure( | ||||||
| 5031 | "Not inserting runtime ptr check for divergent target", | ||||||
| 5032 | "runtime pointer checks needed. Not enabled for divergent target", | ||||||
| 5033 | "CantVersionLoopWithDivergentTarget", ORE, TheLoop); | ||||||
| 5034 | return FixedScalableVFPair::getNone(); | ||||||
| 5035 | } | ||||||
| 5036 | |||||||
| 5037 | unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop); | ||||||
| 5038 | LLVM_DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found trip count: " << TC << '\n'; } } while (false); | ||||||
| 5039 | if (TC == 1) { | ||||||
| 5040 | reportVectorizationFailure("Single iteration (non) loop", | ||||||
| 5041 | "loop trip count is one, irrelevant for vectorization", | ||||||
| 5042 | "SingleIterationLoop", ORE, TheLoop); | ||||||
| 5043 | return FixedScalableVFPair::getNone(); | ||||||
| 5044 | } | ||||||
| 5045 | |||||||
| 5046 | switch (ScalarEpilogueStatus) { | ||||||
| 5047 | case CM_ScalarEpilogueAllowed: | ||||||
| 5048 | return computeFeasibleMaxVF(TC, UserVF, false); | ||||||
| 5049 | case CM_ScalarEpilogueNotAllowedUsePredicate: | ||||||
| 5050 | [[fallthrough]]; | ||||||
| 5051 | case CM_ScalarEpilogueNotNeededUsePredicate: | ||||||
| 5052 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: vector predicate hint/switch found.\n" << "LV: Not allowing scalar epilogue, creating predicated " << "vector loop.\n"; } } while (false) | ||||||
| 5053 | dbgs() << "LV: vector predicate hint/switch found.\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: vector predicate hint/switch found.\n" << "LV: Not allowing scalar epilogue, creating predicated " << "vector loop.\n"; } } while (false) | ||||||
| 5054 | << "LV: Not allowing scalar epilogue, creating predicated "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: vector predicate hint/switch found.\n" << "LV: Not allowing scalar epilogue, creating predicated " << "vector loop.\n"; } } while (false) | ||||||
| 5055 | << "vector loop.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: vector predicate hint/switch found.\n" << "LV: Not allowing scalar epilogue, creating predicated " << "vector loop.\n"; } } while (false); | ||||||
| 5056 | break; | ||||||
| 5057 | case CM_ScalarEpilogueNotAllowedLowTripLoop: | ||||||
| 5058 | // fallthrough as a special case of OptForSize | ||||||
| 5059 | case CM_ScalarEpilogueNotAllowedOptSize: | ||||||
| 5060 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedOptSize) | ||||||
| 5061 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not allowing scalar epilogue due to -Os/-Oz.\n" ; } } while (false) | ||||||
| 5062 | dbgs() << "LV: Not allowing scalar epilogue due to -Os/-Oz.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not allowing scalar epilogue due to -Os/-Oz.\n" ; } } while (false); | ||||||
| 5063 | else | ||||||
| 5064 | LLVM_DEBUG(dbgs() << "LV: Not allowing scalar epilogue due to low trip "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not allowing scalar epilogue due to low trip " << "count.\n"; } } while (false) | ||||||
| 5065 | << "count.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not allowing scalar epilogue due to low trip " << "count.\n"; } } while (false); | ||||||
| 5066 | |||||||
| 5067 | // Bail if runtime checks are required, which are not good when optimising | ||||||
| 5068 | // for size. | ||||||
| 5069 | if (runtimeChecksRequired()) | ||||||
| 5070 | return FixedScalableVFPair::getNone(); | ||||||
| 5071 | |||||||
| 5072 | break; | ||||||
| 5073 | } | ||||||
| 5074 | |||||||
| 5075 | // The only loops we can vectorize without a scalar epilogue, are loops with | ||||||
| 5076 | // a bottom-test and a single exiting block. We'd have to handle the fact | ||||||
| 5077 | // that not every instruction executes on the last iteration. This will | ||||||
| 5078 | // require a lane mask which varies through the vector loop body. (TODO) | ||||||
| 5079 | if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) { | ||||||
| 5080 | // If there was a tail-folding hint/switch, but we can't fold the tail by | ||||||
| 5081 | // masking, fallback to a vectorization with a scalar epilogue. | ||||||
| 5082 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotNeededUsePredicate) { | ||||||
| 5083 | LLVM_DEBUG(dbgs() << "LV: Cannot fold tail by masking: vectorize with a "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Cannot fold tail by masking: vectorize with a " "scalar epilogue instead.\n"; } } while (false) | ||||||
| 5084 | "scalar epilogue instead.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Cannot fold tail by masking: vectorize with a " "scalar epilogue instead.\n"; } } while (false); | ||||||
| 5085 | ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; | ||||||
| 5086 | return computeFeasibleMaxVF(TC, UserVF, false); | ||||||
| 5087 | } | ||||||
| 5088 | return FixedScalableVFPair::getNone(); | ||||||
| 5089 | } | ||||||
| 5090 | |||||||
| 5091 | // Now try the tail folding | ||||||
| 5092 | |||||||
| 5093 | // Invalidate interleave groups that require an epilogue if we can't mask | ||||||
| 5094 | // the interleave-group. | ||||||
| 5095 | if (!useMaskedInterleavedAccesses(TTI)) { | ||||||
| 5096 | assert(WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() &&(static_cast <bool> (WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() && "No decisions should have been taken at this point" ) ? void (0) : __assert_fail ("WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() && \"No decisions should have been taken at this point\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5097, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 5097 | "No decisions should have been taken at this point")(static_cast <bool> (WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() && "No decisions should have been taken at this point" ) ? void (0) : __assert_fail ("WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() && \"No decisions should have been taken at this point\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5097, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 5098 | // Note: There is no need to invalidate any cost modeling decisions here, as | ||||||
| 5099 | // non where taken so far. | ||||||
| 5100 | InterleaveInfo.invalidateGroupsRequiringScalarEpilogue(); | ||||||
| 5101 | } | ||||||
| 5102 | |||||||
| 5103 | FixedScalableVFPair MaxFactors = computeFeasibleMaxVF(TC, UserVF, true); | ||||||
| 5104 | // Avoid tail folding if the trip count is known to be a multiple of any VF | ||||||
| 5105 | // we chose. | ||||||
| 5106 | // FIXME: The condition below pessimises the case for fixed-width vectors, | ||||||
| 5107 | // when scalable VFs are also candidates for vectorization. | ||||||
| 5108 | if (MaxFactors.FixedVF.isVector() && !MaxFactors.ScalableVF) { | ||||||
| 5109 | ElementCount MaxFixedVF = MaxFactors.FixedVF; | ||||||
| 5110 | assert((UserVF.isNonZero() || isPowerOf2_32(MaxFixedVF.getFixedValue())) &&(static_cast <bool> ((UserVF.isNonZero() || isPowerOf2_32 (MaxFixedVF.getFixedValue())) && "MaxFixedVF must be a power of 2" ) ? void (0) : __assert_fail ("(UserVF.isNonZero() || isPowerOf2_32(MaxFixedVF.getFixedValue())) && \"MaxFixedVF must be a power of 2\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5111, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 5111 | "MaxFixedVF must be a power of 2")(static_cast <bool> ((UserVF.isNonZero() || isPowerOf2_32 (MaxFixedVF.getFixedValue())) && "MaxFixedVF must be a power of 2" ) ? void (0) : __assert_fail ("(UserVF.isNonZero() || isPowerOf2_32(MaxFixedVF.getFixedValue())) && \"MaxFixedVF must be a power of 2\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5111, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 5112 | unsigned MaxVFtimesIC = UserIC ? MaxFixedVF.getFixedValue() * UserIC | ||||||
| 5113 | : MaxFixedVF.getFixedValue(); | ||||||
| 5114 | ScalarEvolution *SE = PSE.getSE(); | ||||||
| 5115 | const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount(); | ||||||
| 5116 | const SCEV *ExitCount = SE->getAddExpr( | ||||||
| 5117 | BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType())); | ||||||
| 5118 | const SCEV *Rem = SE->getURemExpr( | ||||||
| 5119 | SE->applyLoopGuards(ExitCount, TheLoop), | ||||||
| 5120 | SE->getConstant(BackedgeTakenCount->getType(), MaxVFtimesIC)); | ||||||
| 5121 | if (Rem->isZero()) { | ||||||
| 5122 | // Accept MaxFixedVF if we do not have a tail. | ||||||
| 5123 | LLVM_DEBUG(dbgs() << "LV: No tail will remain for any chosen VF.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: No tail will remain for any chosen VF.\n" ; } } while (false); | ||||||
| 5124 | return MaxFactors; | ||||||
| 5125 | } | ||||||
| 5126 | } | ||||||
| 5127 | |||||||
| 5128 | // If we don't know the precise trip count, or if the trip count that we | ||||||
| 5129 | // found modulo the vectorization factor is not zero, try to fold the tail | ||||||
| 5130 | // by masking. | ||||||
| 5131 | // FIXME: look for a smaller MaxVF that does divide TC rather than masking. | ||||||
| 5132 | if (Legal->prepareToFoldTailByMasking()) { | ||||||
| 5133 | FoldTailByMasking = true; | ||||||
| 5134 | return MaxFactors; | ||||||
| 5135 | } | ||||||
| 5136 | |||||||
| 5137 | // If there was a tail-folding hint/switch, but we can't fold the tail by | ||||||
| 5138 | // masking, fallback to a vectorization with a scalar epilogue. | ||||||
| 5139 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotNeededUsePredicate) { | ||||||
| 5140 | LLVM_DEBUG(dbgs() << "LV: Cannot fold tail by masking: vectorize with a "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Cannot fold tail by masking: vectorize with a " "scalar epilogue instead.\n"; } } while (false) | ||||||
| 5141 | "scalar epilogue instead.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Cannot fold tail by masking: vectorize with a " "scalar epilogue instead.\n"; } } while (false); | ||||||
| 5142 | ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; | ||||||
| 5143 | return MaxFactors; | ||||||
| 5144 | } | ||||||
| 5145 | |||||||
| 5146 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedUsePredicate) { | ||||||
| 5147 | LLVM_DEBUG(dbgs() << "LV: Can't fold tail by masking: don't vectorize\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Can't fold tail by masking: don't vectorize\n" ; } } while (false); | ||||||
| 5148 | return FixedScalableVFPair::getNone(); | ||||||
| 5149 | } | ||||||
| 5150 | |||||||
| 5151 | if (TC == 0) { | ||||||
| 5152 | reportVectorizationFailure( | ||||||
| 5153 | "Unable to calculate the loop count due to complex control flow", | ||||||
| 5154 | "unable to calculate the loop count due to complex control flow", | ||||||
| 5155 | "UnknownLoopCountComplexCFG", ORE, TheLoop); | ||||||
| 5156 | return FixedScalableVFPair::getNone(); | ||||||
| 5157 | } | ||||||
| 5158 | |||||||
| 5159 | reportVectorizationFailure( | ||||||
| 5160 | "Cannot optimize for size and vectorize at the same time.", | ||||||
| 5161 | "cannot optimize for size and vectorize at the same time. " | ||||||
| 5162 | "Enable vectorization of this loop with '#pragma clang loop " | ||||||
| 5163 | "vectorize(enable)' when compiling with -Os/-Oz", | ||||||
| 5164 | "NoTailLoopWithOptForSize", ORE, TheLoop); | ||||||
| 5165 | return FixedScalableVFPair::getNone(); | ||||||
| 5166 | } | ||||||
| 5167 | |||||||
| 5168 | ElementCount LoopVectorizationCostModel::getMaximizedVFForTarget( | ||||||
| 5169 | unsigned ConstTripCount, unsigned SmallestType, unsigned WidestType, | ||||||
| 5170 | ElementCount MaxSafeVF, bool FoldTailByMasking) { | ||||||
| 5171 | bool ComputeScalableMaxVF = MaxSafeVF.isScalable(); | ||||||
| 5172 | const TypeSize WidestRegister = TTI.getRegisterBitWidth( | ||||||
| 5173 | ComputeScalableMaxVF ? TargetTransformInfo::RGK_ScalableVector | ||||||
| 5174 | : TargetTransformInfo::RGK_FixedWidthVector); | ||||||
| 5175 | |||||||
| 5176 | // Convenience function to return the minimum of two ElementCounts. | ||||||
| 5177 | auto MinVF = [](const ElementCount &LHS, const ElementCount &RHS) { | ||||||
| 5178 | assert((LHS.isScalable() == RHS.isScalable()) &&(static_cast <bool> ((LHS.isScalable() == RHS.isScalable ()) && "Scalable flags must match") ? void (0) : __assert_fail ("(LHS.isScalable() == RHS.isScalable()) && \"Scalable flags must match\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5179, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 5179 | "Scalable flags must match")(static_cast <bool> ((LHS.isScalable() == RHS.isScalable ()) && "Scalable flags must match") ? void (0) : __assert_fail ("(LHS.isScalable() == RHS.isScalable()) && \"Scalable flags must match\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5179, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 5180 | return ElementCount::isKnownLT(LHS, RHS) ? LHS : RHS; | ||||||
| 5181 | }; | ||||||
| 5182 | |||||||
| 5183 | // Ensure MaxVF is a power of 2; the dependence distance bound may not be. | ||||||
| 5184 | // Note that both WidestRegister and WidestType may not be a powers of 2. | ||||||
| 5185 | auto MaxVectorElementCount = ElementCount::get( | ||||||
| 5186 | PowerOf2Floor(WidestRegister.getKnownMinSize() / WidestType), | ||||||
| 5187 | ComputeScalableMaxVF); | ||||||
| 5188 | MaxVectorElementCount = MinVF(MaxVectorElementCount, MaxSafeVF); | ||||||
| 5189 | LLVM_DEBUG(dbgs() << "LV: The Widest register safe to use is: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The Widest register safe to use is: " << (MaxVectorElementCount * WidestType) << " bits.\n" ; } } while (false) | ||||||
| 5190 | << (MaxVectorElementCount * WidestType) << " bits.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The Widest register safe to use is: " << (MaxVectorElementCount * WidestType) << " bits.\n" ; } } while (false); | ||||||
| 5191 | |||||||
| 5192 | if (!MaxVectorElementCount) { | ||||||
| 5193 | LLVM_DEBUG(dbgs() << "LV: The target has no "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The target has no " << (ComputeScalableMaxVF ? "scalable" : "fixed") << " vector registers.\n"; } } while (false) | ||||||
| 5194 | << (ComputeScalableMaxVF ? "scalable" : "fixed")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The target has no " << (ComputeScalableMaxVF ? "scalable" : "fixed") << " vector registers.\n"; } } while (false) | ||||||
| 5195 | << " vector registers.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The target has no " << (ComputeScalableMaxVF ? "scalable" : "fixed") << " vector registers.\n"; } } while (false); | ||||||
| 5196 | return ElementCount::getFixed(1); | ||||||
| 5197 | } | ||||||
| 5198 | |||||||
| 5199 | unsigned WidestRegisterMinEC = MaxVectorElementCount.getKnownMinValue(); | ||||||
| 5200 | if (MaxVectorElementCount.isScalable() && | ||||||
| 5201 | TheFunction->hasFnAttribute(Attribute::VScaleRange)) { | ||||||
| 5202 | auto Attr = TheFunction->getFnAttribute(Attribute::VScaleRange); | ||||||
| 5203 | auto Min = Attr.getVScaleRangeMin(); | ||||||
| 5204 | WidestRegisterMinEC *= Min; | ||||||
| 5205 | } | ||||||
| 5206 | if (ConstTripCount && ConstTripCount <= WidestRegisterMinEC && | ||||||
| 5207 | (!FoldTailByMasking || isPowerOf2_32(ConstTripCount))) { | ||||||
| 5208 | // If loop trip count (TC) is known at compile time there is no point in | ||||||
| 5209 | // choosing VF greater than TC (as done in the loop below). Select maximum | ||||||
| 5210 | // power of two which doesn't exceed TC. | ||||||
| 5211 | // If MaxVectorElementCount is scalable, we only fall back on a fixed VF | ||||||
| 5212 | // when the TC is less than or equal to the known number of lanes. | ||||||
| 5213 | auto ClampedConstTripCount = PowerOf2Floor(ConstTripCount); | ||||||
| 5214 | LLVM_DEBUG(dbgs() << "LV: Clamping the MaxVF to maximum power of two not "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Clamping the MaxVF to maximum power of two not " "exceeding the constant trip count: " << ClampedConstTripCount << "\n"; } } while (false) | ||||||
| 5215 | "exceeding the constant trip count: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Clamping the MaxVF to maximum power of two not " "exceeding the constant trip count: " << ClampedConstTripCount << "\n"; } } while (false) | ||||||
| 5216 | << ClampedConstTripCount << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Clamping the MaxVF to maximum power of two not " "exceeding the constant trip count: " << ClampedConstTripCount << "\n"; } } while (false); | ||||||
| 5217 | return ElementCount::getFixed(ClampedConstTripCount); | ||||||
| 5218 | } | ||||||
| 5219 | |||||||
| 5220 | TargetTransformInfo::RegisterKind RegKind = | ||||||
| 5221 | ComputeScalableMaxVF ? TargetTransformInfo::RGK_ScalableVector | ||||||
| 5222 | : TargetTransformInfo::RGK_FixedWidthVector; | ||||||
| 5223 | ElementCount MaxVF = MaxVectorElementCount; | ||||||
| 5224 | if (MaximizeBandwidth || (MaximizeBandwidth.getNumOccurrences() == 0 && | ||||||
| 5225 | TTI.shouldMaximizeVectorBandwidth(RegKind))) { | ||||||
| 5226 | auto MaxVectorElementCountMaxBW = ElementCount::get( | ||||||
| 5227 | PowerOf2Floor(WidestRegister.getKnownMinSize() / SmallestType), | ||||||
| 5228 | ComputeScalableMaxVF); | ||||||
| 5229 | MaxVectorElementCountMaxBW = MinVF(MaxVectorElementCountMaxBW, MaxSafeVF); | ||||||
| 5230 | |||||||
| 5231 | // Collect all viable vectorization factors larger than the default MaxVF | ||||||
| 5232 | // (i.e. MaxVectorElementCount). | ||||||
| 5233 | SmallVector<ElementCount, 8> VFs; | ||||||
| 5234 | for (ElementCount VS = MaxVectorElementCount * 2; | ||||||
| 5235 | ElementCount::isKnownLE(VS, MaxVectorElementCountMaxBW); VS *= 2) | ||||||
| 5236 | VFs.push_back(VS); | ||||||
| 5237 | |||||||
| 5238 | // For each VF calculate its register usage. | ||||||
| 5239 | auto RUs = calculateRegisterUsage(VFs); | ||||||
| 5240 | |||||||
| 5241 | // Select the largest VF which doesn't require more registers than existing | ||||||
| 5242 | // ones. | ||||||
| 5243 | for (int i = RUs.size() - 1; i >= 0; --i) { | ||||||
| 5244 | bool Selected = true; | ||||||
| 5245 | for (auto &pair : RUs[i].MaxLocalUsers) { | ||||||
| 5246 | unsigned TargetNumRegisters = TTI.getNumberOfRegisters(pair.first); | ||||||
| 5247 | if (pair.second > TargetNumRegisters) | ||||||
| 5248 | Selected = false; | ||||||
| 5249 | } | ||||||
| 5250 | if (Selected) { | ||||||
| 5251 | MaxVF = VFs[i]; | ||||||
| 5252 | break; | ||||||
| 5253 | } | ||||||
| 5254 | } | ||||||
| 5255 | if (ElementCount MinVF = | ||||||
| 5256 | TTI.getMinimumVF(SmallestType, ComputeScalableMaxVF)) { | ||||||
| 5257 | if (ElementCount::isKnownLT(MaxVF, MinVF)) { | ||||||
| 5258 | LLVM_DEBUG(dbgs() << "LV: Overriding calculated MaxVF(" << MaxVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Overriding calculated MaxVF(" << MaxVF << ") with target's minimum: " << MinVF << '\n'; } } while (false) | ||||||
| 5259 | << ") with target's minimum: " << MinVF << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Overriding calculated MaxVF(" << MaxVF << ") with target's minimum: " << MinVF << '\n'; } } while (false); | ||||||
| 5260 | MaxVF = MinVF; | ||||||
| 5261 | } | ||||||
| 5262 | } | ||||||
| 5263 | |||||||
| 5264 | // Invalidate any widening decisions we might have made, in case the loop | ||||||
| 5265 | // requires prediction (decided later), but we have already made some | ||||||
| 5266 | // load/store widening decisions. | ||||||
| 5267 | invalidateCostModelingDecisions(); | ||||||
| 5268 | } | ||||||
| 5269 | return MaxVF; | ||||||
| 5270 | } | ||||||
| 5271 | |||||||
| 5272 | std::optional<unsigned> LoopVectorizationCostModel::getVScaleForTuning() const { | ||||||
| 5273 | if (TheFunction->hasFnAttribute(Attribute::VScaleRange)) { | ||||||
| 5274 | auto Attr = TheFunction->getFnAttribute(Attribute::VScaleRange); | ||||||
| 5275 | auto Min = Attr.getVScaleRangeMin(); | ||||||
| 5276 | auto Max = Attr.getVScaleRangeMax(); | ||||||
| 5277 | if (Max && Min == Max) | ||||||
| 5278 | return Max; | ||||||
| 5279 | } | ||||||
| 5280 | |||||||
| 5281 | return TTI.getVScaleForTuning(); | ||||||
| 5282 | } | ||||||
| 5283 | |||||||
| 5284 | bool LoopVectorizationCostModel::isMoreProfitable( | ||||||
| 5285 | const VectorizationFactor &A, const VectorizationFactor &B) const { | ||||||
| 5286 | InstructionCost CostA = A.Cost; | ||||||
| 5287 | InstructionCost CostB = B.Cost; | ||||||
| 5288 | |||||||
| 5289 | unsigned MaxTripCount = PSE.getSE()->getSmallConstantMaxTripCount(TheLoop); | ||||||
| 5290 | |||||||
| 5291 | if (!A.Width.isScalable() && !B.Width.isScalable() && FoldTailByMasking && | ||||||
| 5292 | MaxTripCount) { | ||||||
| 5293 | // If we are folding the tail and the trip count is a known (possibly small) | ||||||
| 5294 | // constant, the trip count will be rounded up to an integer number of | ||||||
| 5295 | // iterations. The total cost will be PerIterationCost*ceil(TripCount/VF), | ||||||
| 5296 | // which we compare directly. When not folding the tail, the total cost will | ||||||
| 5297 | // be PerIterationCost*floor(TC/VF) + Scalar remainder cost, and so is | ||||||
| 5298 | // approximated with the per-lane cost below instead of using the tripcount | ||||||
| 5299 | // as here. | ||||||
| 5300 | auto RTCostA = CostA * divideCeil(MaxTripCount, A.Width.getFixedValue()); | ||||||
| 5301 | auto RTCostB = CostB * divideCeil(MaxTripCount, B.Width.getFixedValue()); | ||||||
| 5302 | return RTCostA < RTCostB; | ||||||
| 5303 | } | ||||||
| 5304 | |||||||
| 5305 | // Improve estimate for the vector width if it is scalable. | ||||||
| 5306 | unsigned EstimatedWidthA = A.Width.getKnownMinValue(); | ||||||
| 5307 | unsigned EstimatedWidthB = B.Width.getKnownMinValue(); | ||||||
| 5308 | if (std::optional<unsigned> VScale = getVScaleForTuning()) { | ||||||
| 5309 | if (A.Width.isScalable()) | ||||||
| 5310 | EstimatedWidthA *= *VScale; | ||||||
| 5311 | if (B.Width.isScalable()) | ||||||
| 5312 | EstimatedWidthB *= *VScale; | ||||||
| 5313 | } | ||||||
| 5314 | |||||||
| 5315 | // Assume vscale may be larger than 1 (or the value being tuned for), | ||||||
| 5316 | // so that scalable vectorization is slightly favorable over fixed-width | ||||||
| 5317 | // vectorization. | ||||||
| 5318 | if (A.Width.isScalable() && !B.Width.isScalable()) | ||||||
| 5319 | return (CostA * B.Width.getFixedValue()) <= (CostB * EstimatedWidthA); | ||||||
| 5320 | |||||||
| 5321 | // To avoid the need for FP division: | ||||||
| 5322 | // (CostA / A.Width) < (CostB / B.Width) | ||||||
| 5323 | // <=> (CostA * B.Width) < (CostB * A.Width) | ||||||
| 5324 | return (CostA * EstimatedWidthB) < (CostB * EstimatedWidthA); | ||||||
| 5325 | } | ||||||
| 5326 | |||||||
| 5327 | VectorizationFactor LoopVectorizationCostModel::selectVectorizationFactor( | ||||||
| 5328 | const ElementCountSet &VFCandidates) { | ||||||
| 5329 | InstructionCost ExpectedCost = expectedCost(ElementCount::getFixed(1)).first; | ||||||
| 5330 | LLVM_DEBUG(dbgs() << "LV: Scalar loop costs: " << ExpectedCost << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Scalar loop costs: " << ExpectedCost << ".\n"; } } while (false); | ||||||
| 5331 | assert(ExpectedCost.isValid() && "Unexpected invalid cost for scalar loop")(static_cast <bool> (ExpectedCost.isValid() && "Unexpected invalid cost for scalar loop" ) ? void (0) : __assert_fail ("ExpectedCost.isValid() && \"Unexpected invalid cost for scalar loop\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5331, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 5332 | assert(VFCandidates.count(ElementCount::getFixed(1)) &&(static_cast <bool> (VFCandidates.count(ElementCount::getFixed (1)) && "Expected Scalar VF to be a candidate") ? void (0) : __assert_fail ("VFCandidates.count(ElementCount::getFixed(1)) && \"Expected Scalar VF to be a candidate\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5333, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 5333 | "Expected Scalar VF to be a candidate")(static_cast <bool> (VFCandidates.count(ElementCount::getFixed (1)) && "Expected Scalar VF to be a candidate") ? void (0) : __assert_fail ("VFCandidates.count(ElementCount::getFixed(1)) && \"Expected Scalar VF to be a candidate\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5333, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 5334 | |||||||
| 5335 | const VectorizationFactor ScalarCost(ElementCount::getFixed(1), ExpectedCost, | ||||||
| 5336 | ExpectedCost); | ||||||
| 5337 | VectorizationFactor ChosenFactor = ScalarCost; | ||||||
| 5338 | |||||||
| 5339 | bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled; | ||||||
| 5340 | if (ForceVectorization && VFCandidates.size() > 1) { | ||||||
| 5341 | // Ignore scalar width, because the user explicitly wants vectorization. | ||||||
| 5342 | // Initialize cost to max so that VF = 2 is, at least, chosen during cost | ||||||
| 5343 | // evaluation. | ||||||
| 5344 | ChosenFactor.Cost = InstructionCost::getMax(); | ||||||
| 5345 | } | ||||||
| 5346 | |||||||
| 5347 | SmallVector<InstructionVFPair> InvalidCosts; | ||||||
| 5348 | for (const auto &i : VFCandidates) { | ||||||
| 5349 | // The cost for scalar VF=1 is already calculated, so ignore it. | ||||||
| 5350 | if (i.isScalar()) | ||||||
| 5351 | continue; | ||||||
| 5352 | |||||||
| 5353 | VectorizationCostTy C = expectedCost(i, &InvalidCosts); | ||||||
| 5354 | VectorizationFactor Candidate(i, C.first, ScalarCost.ScalarCost); | ||||||
| 5355 | |||||||
| 5356 | #ifndef NDEBUG | ||||||
| 5357 | unsigned AssumedMinimumVscale = 1; | ||||||
| 5358 | if (std::optional<unsigned> VScale = getVScaleForTuning()) | ||||||
| 5359 | AssumedMinimumVscale = *VScale; | ||||||
| 5360 | unsigned Width = | ||||||
| 5361 | Candidate.Width.isScalable() | ||||||
| 5362 | ? Candidate.Width.getKnownMinValue() * AssumedMinimumVscale | ||||||
| 5363 | : Candidate.Width.getFixedValue(); | ||||||
| 5364 | LLVM_DEBUG(dbgs() << "LV: Vector loop of width " << ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Vector loop of width " << i << " costs: " << (Candidate.Cost / Width ); } } while (false) | ||||||
| 5365 | << " costs: " << (Candidate.Cost / Width))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Vector loop of width " << i << " costs: " << (Candidate.Cost / Width ); } } while (false); | ||||||
| 5366 | if (i.isScalable()) | ||||||
| 5367 | LLVM_DEBUG(dbgs() << " (assuming a minimum vscale of "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " (assuming a minimum vscale of " << AssumedMinimumVscale << ")"; } } while (false ) | ||||||
| 5368 | << AssumedMinimumVscale << ")")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " (assuming a minimum vscale of " << AssumedMinimumVscale << ")"; } } while (false ); | ||||||
| 5369 | LLVM_DEBUG(dbgs() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << ".\n"; } } while (false ); | ||||||
| 5370 | #endif | ||||||
| 5371 | |||||||
| 5372 | if (!C.second && !ForceVectorization) { | ||||||
| 5373 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not considering vector loop of width " << i << " because it will not generate any vector instructions.\n" ; } } while (false) | ||||||
| 5374 | dbgs() << "LV: Not considering vector loop of width " << ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not considering vector loop of width " << i << " because it will not generate any vector instructions.\n" ; } } while (false) | ||||||
| 5375 | << " because it will not generate any vector instructions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not considering vector loop of width " << i << " because it will not generate any vector instructions.\n" ; } } while (false); | ||||||
| 5376 | continue; | ||||||
| 5377 | } | ||||||
| 5378 | |||||||
| 5379 | // If profitable add it to ProfitableVF list. | ||||||
| 5380 | if (isMoreProfitable(Candidate, ScalarCost)) | ||||||
| 5381 | ProfitableVFs.push_back(Candidate); | ||||||
| 5382 | |||||||
| 5383 | if (isMoreProfitable(Candidate, ChosenFactor)) | ||||||
| 5384 | ChosenFactor = Candidate; | ||||||
| 5385 | } | ||||||
| 5386 | |||||||
| 5387 | // Emit a report of VFs with invalid costs in the loop. | ||||||
| 5388 | if (!InvalidCosts.empty()) { | ||||||
| 5389 | // Group the remarks per instruction, keeping the instruction order from | ||||||
| 5390 | // InvalidCosts. | ||||||
| 5391 | std::map<Instruction *, unsigned> Numbering; | ||||||
| 5392 | unsigned I = 0; | ||||||
| 5393 | for (auto &Pair : InvalidCosts) | ||||||
| 5394 | if (!Numbering.count(Pair.first)) | ||||||
| 5395 | Numbering[Pair.first] = I++; | ||||||
| 5396 | |||||||
| 5397 | // Sort the list, first on instruction(number) then on VF. | ||||||
| 5398 | llvm::sort(InvalidCosts, | ||||||
| 5399 | [&Numbering](InstructionVFPair &A, InstructionVFPair &B) { | ||||||
| 5400 | if (Numbering[A.first] != Numbering[B.first]) | ||||||
| 5401 | return Numbering[A.first] < Numbering[B.first]; | ||||||
| 5402 | ElementCountComparator ECC; | ||||||
| 5403 | return ECC(A.second, B.second); | ||||||
| 5404 | }); | ||||||
| 5405 | |||||||
| 5406 | // For a list of ordered instruction-vf pairs: | ||||||
| 5407 | // [(load, vf1), (load, vf2), (store, vf1)] | ||||||
| 5408 | // Group the instructions together to emit separate remarks for: | ||||||
| 5409 | // load (vf1, vf2) | ||||||
| 5410 | // store (vf1) | ||||||
| 5411 | auto Tail = ArrayRef<InstructionVFPair>(InvalidCosts); | ||||||
| 5412 | auto Subset = ArrayRef<InstructionVFPair>(); | ||||||
| 5413 | do { | ||||||
| 5414 | if (Subset.empty()) | ||||||
| 5415 | Subset = Tail.take_front(1); | ||||||
| 5416 | |||||||
| 5417 | Instruction *I = Subset.front().first; | ||||||
| 5418 | |||||||
| 5419 | // If the next instruction is different, or if there are no other pairs, | ||||||
| 5420 | // emit a remark for the collated subset. e.g. | ||||||
| 5421 | // [(load, vf1), (load, vf2))] | ||||||
| 5422 | // to emit: | ||||||
| 5423 | // remark: invalid costs for 'load' at VF=(vf, vf2) | ||||||
| 5424 | if (Subset == Tail || Tail[Subset.size()].first != I) { | ||||||
| 5425 | std::string OutString; | ||||||
| 5426 | raw_string_ostream OS(OutString); | ||||||
| 5427 | assert(!Subset.empty() && "Unexpected empty range")(static_cast <bool> (!Subset.empty() && "Unexpected empty range" ) ? void (0) : __assert_fail ("!Subset.empty() && \"Unexpected empty range\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5427, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 5428 | OS << "Instruction with invalid costs prevented vectorization at VF=("; | ||||||
| 5429 | for (const auto &Pair : Subset) | ||||||
| 5430 | OS << (Pair.second == Subset.front().second ? "" : ", ") | ||||||
| 5431 | << Pair.second; | ||||||
| 5432 | OS << "):"; | ||||||
| 5433 | if (auto *CI = dyn_cast<CallInst>(I)) | ||||||
| 5434 | OS << " call to " << CI->getCalledFunction()->getName(); | ||||||
| 5435 | else | ||||||
| 5436 | OS << " " << I->getOpcodeName(); | ||||||
| 5437 | OS.flush(); | ||||||
| 5438 | reportVectorizationInfo(OutString, "InvalidCost", ORE, TheLoop, I); | ||||||
| 5439 | Tail = Tail.drop_front(Subset.size()); | ||||||
| 5440 | Subset = {}; | ||||||
| 5441 | } else | ||||||
| 5442 | // Grow the subset by one element | ||||||
| 5443 | Subset = Tail.take_front(Subset.size() + 1); | ||||||
| 5444 | } while (!Tail.empty()); | ||||||
| 5445 | } | ||||||
| 5446 | |||||||
| 5447 | if (!EnableCondStoresVectorization && NumPredStores) { | ||||||
| 5448 | reportVectorizationFailure("There are conditional stores.", | ||||||
| 5449 | "store that is conditionally executed prevents vectorization", | ||||||
| 5450 | "ConditionalStore", ORE, TheLoop); | ||||||
| 5451 | ChosenFactor = ScalarCost; | ||||||
| 5452 | } | ||||||
| 5453 | |||||||
| 5454 | LLVM_DEBUG(if (ForceVectorization && !ChosenFactor.Width.isScalar() &&do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && !ChosenFactor .Width.isScalar() && !isMoreProfitable(ChosenFactor, ScalarCost )) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false) | ||||||
| 5455 | !isMoreProfitable(ChosenFactor, ScalarCost)) dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && !ChosenFactor .Width.isScalar() && !isMoreProfitable(ChosenFactor, ScalarCost )) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false) | ||||||
| 5456 | << "LV: Vectorization seems to be not beneficial, "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && !ChosenFactor .Width.isScalar() && !isMoreProfitable(ChosenFactor, ScalarCost )) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false) | ||||||
| 5457 | << "but was forced by a user.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && !ChosenFactor .Width.isScalar() && !isMoreProfitable(ChosenFactor, ScalarCost )) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false); | ||||||
| 5458 | LLVM_DEBUG(dbgs() << "LV: Selecting VF: " << ChosenFactor.Width << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Selecting VF: " << ChosenFactor.Width << ".\n"; } } while (false); | ||||||
| 5459 | return ChosenFactor; | ||||||
| 5460 | } | ||||||
| 5461 | |||||||
| 5462 | bool LoopVectorizationCostModel::isCandidateForEpilogueVectorization( | ||||||
| 5463 | const Loop &L, ElementCount VF) const { | ||||||
| 5464 | // Cross iteration phis such as reductions need special handling and are | ||||||
| 5465 | // currently unsupported. | ||||||
| 5466 | if (any_of(L.getHeader()->phis(), | ||||||
| 5467 | [&](PHINode &Phi) { return Legal->isFixedOrderRecurrence(&Phi); })) | ||||||
| 5468 | return false; | ||||||
| 5469 | |||||||
| 5470 | // Phis with uses outside of the loop require special handling and are | ||||||
| 5471 | // currently unsupported. | ||||||
| 5472 | for (const auto &Entry : Legal->getInductionVars()) { | ||||||
| 5473 | // Look for uses of the value of the induction at the last iteration. | ||||||
| 5474 | Value *PostInc = Entry.first->getIncomingValueForBlock(L.getLoopLatch()); | ||||||
| 5475 | for (User *U : PostInc->users()) | ||||||
| 5476 | if (!L.contains(cast<Instruction>(U))) | ||||||
| 5477 | return false; | ||||||
| 5478 | // Look for uses of penultimate value of the induction. | ||||||
| 5479 | for (User *U : Entry.first->users()) | ||||||
| 5480 | if (!L.contains(cast<Instruction>(U))) | ||||||
| 5481 | return false; | ||||||
| 5482 | } | ||||||
| 5483 | |||||||
| 5484 | // Epilogue vectorization code has not been auditted to ensure it handles | ||||||
| 5485 | // non-latch exits properly. It may be fine, but it needs auditted and | ||||||
| 5486 | // tested. | ||||||
| 5487 | if (L.getExitingBlock() != L.getLoopLatch()) | ||||||
| 5488 | return false; | ||||||
| 5489 | |||||||
| 5490 | return true; | ||||||
| 5491 | } | ||||||
| 5492 | |||||||
| 5493 | bool LoopVectorizationCostModel::isEpilogueVectorizationProfitable( | ||||||
| 5494 | const ElementCount VF) const { | ||||||
| 5495 | // FIXME: We need a much better cost-model to take different parameters such | ||||||
| 5496 | // as register pressure, code size increase and cost of extra branches into | ||||||
| 5497 | // account. For now we apply a very crude heuristic and only consider loops | ||||||
| 5498 | // with vectorization factors larger than a certain value. | ||||||
| 5499 | |||||||
| 5500 | // Allow the target to opt out entirely. | ||||||
| 5501 | if (!TTI.preferEpilogueVectorization()) | ||||||
| 5502 | return false; | ||||||
| 5503 | |||||||
| 5504 | // We also consider epilogue vectorization unprofitable for targets that don't | ||||||
| 5505 | // consider interleaving beneficial (eg. MVE). | ||||||
| 5506 | if (TTI.getMaxInterleaveFactor(VF.getKnownMinValue()) <= 1) | ||||||
| 5507 | return false; | ||||||
| 5508 | // FIXME: We should consider changing the threshold for scalable | ||||||
| 5509 | // vectors to take VScaleForTuning into account. | ||||||
| 5510 | if (VF.getKnownMinValue() >= EpilogueVectorizationMinVF) | ||||||
| 5511 | return true; | ||||||
| 5512 | return false; | ||||||
| 5513 | } | ||||||
| 5514 | |||||||
| 5515 | VectorizationFactor | ||||||
| 5516 | LoopVectorizationCostModel::selectEpilogueVectorizationFactor( | ||||||
| 5517 | const ElementCount MainLoopVF, const LoopVectorizationPlanner &LVP) { | ||||||
| 5518 | VectorizationFactor Result = VectorizationFactor::Disabled(); | ||||||
| 5519 | if (!EnableEpilogueVectorization) { | ||||||
| 5520 | LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization is disabled.\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization is disabled.\n" ;; } } while (false); | ||||||
| 5521 | return Result; | ||||||
| 5522 | } | ||||||
| 5523 | |||||||
| 5524 | if (!isScalarEpilogueAllowed()) { | ||||||
| 5525 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Unable to vectorize epilogue because no epilogue is " "allowed.\n";; } } while (false) | ||||||
| 5526 | dbgs() << "LEV: Unable to vectorize epilogue because no epilogue is "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Unable to vectorize epilogue because no epilogue is " "allowed.\n";; } } while (false) | ||||||
| 5527 | "allowed.\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Unable to vectorize epilogue because no epilogue is " "allowed.\n";; } } while (false); | ||||||
| 5528 | return Result; | ||||||
| 5529 | } | ||||||
| 5530 | |||||||
| 5531 | // Not really a cost consideration, but check for unsupported cases here to | ||||||
| 5532 | // simplify the logic. | ||||||
| 5533 | if (!isCandidateForEpilogueVectorization(*TheLoop, MainLoopVF)) { | ||||||
| 5534 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Unable to vectorize epilogue because the loop is " "not a supported candidate.\n";; } } while (false) | ||||||
| 5535 | dbgs() << "LEV: Unable to vectorize epilogue because the loop is "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Unable to vectorize epilogue because the loop is " "not a supported candidate.\n";; } } while (false) | ||||||
| 5536 | "not a supported candidate.\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Unable to vectorize epilogue because the loop is " "not a supported candidate.\n";; } } while (false); | ||||||
| 5537 | return Result; | ||||||
| 5538 | } | ||||||
| 5539 | |||||||
| 5540 | if (EpilogueVectorizationForceVF > 1) { | ||||||
| 5541 | LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization factor is forced.\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization factor is forced.\n" ;; } } while (false); | ||||||
| 5542 | ElementCount ForcedEC = ElementCount::getFixed(EpilogueVectorizationForceVF); | ||||||
| 5543 | if (LVP.hasPlanWithVF(ForcedEC)) | ||||||
| 5544 | return {ForcedEC, 0, 0}; | ||||||
| 5545 | else { | ||||||
| 5546 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization forced factor is not viable.\n" ;; } } while (false) | ||||||
| 5547 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization forced factor is not viable.\n" ;; } } while (false) | ||||||
| 5548 | << "LEV: Epilogue vectorization forced factor is not viable.\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization forced factor is not viable.\n" ;; } } while (false); | ||||||
| 5549 | return Result; | ||||||
| 5550 | } | ||||||
| 5551 | } | ||||||
| 5552 | |||||||
| 5553 | if (TheLoop->getHeader()->getParent()->hasOptSize() || | ||||||
| 5554 | TheLoop->getHeader()->getParent()->hasMinSize()) { | ||||||
| 5555 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization skipped due to opt for size.\n" ;; } } while (false) | ||||||
| 5556 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization skipped due to opt for size.\n" ;; } } while (false) | ||||||
| 5557 | << "LEV: Epilogue vectorization skipped due to opt for size.\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization skipped due to opt for size.\n" ;; } } while (false); | ||||||
| 5558 | return Result; | ||||||
| 5559 | } | ||||||
| 5560 | |||||||
| 5561 | if (!isEpilogueVectorizationProfitable(MainLoopVF)) { | ||||||
| 5562 | LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization is not profitable for "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization is not profitable for " "this loop\n"; } } while (false) | ||||||
| 5563 | "this loop\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization is not profitable for " "this loop\n"; } } while (false); | ||||||
| 5564 | return Result; | ||||||
| 5565 | } | ||||||
| 5566 | |||||||
| 5567 | // If MainLoopVF = vscale x 2, and vscale is expected to be 4, then we know | ||||||
| 5568 | // the main loop handles 8 lanes per iteration. We could still benefit from | ||||||
| 5569 | // vectorizing the epilogue loop with VF=4. | ||||||
| 5570 | ElementCount EstimatedRuntimeVF = MainLoopVF; | ||||||
| 5571 | if (MainLoopVF.isScalable()) { | ||||||
| 5572 | EstimatedRuntimeVF = ElementCount::getFixed(MainLoopVF.getKnownMinValue()); | ||||||
| 5573 | if (std::optional<unsigned> VScale = getVScaleForTuning()) | ||||||
| 5574 | EstimatedRuntimeVF *= *VScale; | ||||||
| 5575 | } | ||||||
| 5576 | |||||||
| 5577 | for (auto &NextVF : ProfitableVFs) | ||||||
| 5578 | if (((!NextVF.Width.isScalable() && MainLoopVF.isScalable() && | ||||||
| 5579 | ElementCount::isKnownLT(NextVF.Width, EstimatedRuntimeVF)) || | ||||||
| 5580 | ElementCount::isKnownLT(NextVF.Width, MainLoopVF)) && | ||||||
| 5581 | (Result.Width.isScalar() || isMoreProfitable(NextVF, Result)) && | ||||||
| 5582 | LVP.hasPlanWithVF(NextVF.Width)) | ||||||
| 5583 | Result = NextVF; | ||||||
| 5584 | |||||||
| 5585 | if (Result != VectorizationFactor::Disabled()) | ||||||
| 5586 | LLVM_DEBUG(dbgs() << "LEV: Vectorizing epilogue loop with VF = "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Vectorizing epilogue loop with VF = " << Result.Width << "\n";; } } while (false) | ||||||
| 5587 | << Result.Width << "\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Vectorizing epilogue loop with VF = " << Result.Width << "\n";; } } while (false); | ||||||
| 5588 | return Result; | ||||||
| 5589 | } | ||||||
| 5590 | |||||||
| 5591 | std::pair<unsigned, unsigned> | ||||||
| 5592 | LoopVectorizationCostModel::getSmallestAndWidestTypes() { | ||||||
| 5593 | unsigned MinWidth = -1U; | ||||||
| 5594 | unsigned MaxWidth = 8; | ||||||
| 5595 | const DataLayout &DL = TheFunction->getParent()->getDataLayout(); | ||||||
| 5596 | // For in-loop reductions, no element types are added to ElementTypesInLoop | ||||||
| 5597 | // if there are no loads/stores in the loop. In this case, check through the | ||||||
| 5598 | // reduction variables to determine the maximum width. | ||||||
| 5599 | if (ElementTypesInLoop.empty() && !Legal->getReductionVars().empty()) { | ||||||
| 5600 | // Reset MaxWidth so that we can find the smallest type used by recurrences | ||||||
| 5601 | // in the loop. | ||||||
| 5602 | MaxWidth = -1U; | ||||||
| 5603 | for (const auto &PhiDescriptorPair : Legal->getReductionVars()) { | ||||||
| 5604 | const RecurrenceDescriptor &RdxDesc = PhiDescriptorPair.second; | ||||||
| 5605 | // When finding the min width used by the recurrence we need to account | ||||||
| 5606 | // for casts on the input operands of the recurrence. | ||||||
| 5607 | MaxWidth = std::min<unsigned>( | ||||||
| 5608 | MaxWidth, std::min<unsigned>( | ||||||
| 5609 | RdxDesc.getMinWidthCastToRecurrenceTypeInBits(), | ||||||
| 5610 | RdxDesc.getRecurrenceType()->getScalarSizeInBits())); | ||||||
| 5611 | } | ||||||
| 5612 | } else { | ||||||
| 5613 | for (Type *T : ElementTypesInLoop) { | ||||||
| 5614 | MinWidth = std::min<unsigned>( | ||||||
| 5615 | MinWidth, DL.getTypeSizeInBits(T->getScalarType()).getFixedSize()); | ||||||
| 5616 | MaxWidth = std::max<unsigned>( | ||||||
| 5617 | MaxWidth, DL.getTypeSizeInBits(T->getScalarType()).getFixedSize()); | ||||||
| 5618 | } | ||||||
| 5619 | } | ||||||
| 5620 | return {MinWidth, MaxWidth}; | ||||||
| 5621 | } | ||||||
| 5622 | |||||||
| 5623 | void LoopVectorizationCostModel::collectElementTypesForWidening() { | ||||||
| 5624 | ElementTypesInLoop.clear(); | ||||||
| 5625 | // For each block. | ||||||
| 5626 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||||
| 5627 | // For each instruction in the loop. | ||||||
| 5628 | for (Instruction &I : BB->instructionsWithoutDebug()) { | ||||||
| 5629 | Type *T = I.getType(); | ||||||
| 5630 | |||||||
| 5631 | // Skip ignored values. | ||||||
| 5632 | if (ValuesToIgnore.count(&I)) | ||||||
| 5633 | continue; | ||||||
| 5634 | |||||||
| 5635 | // Only examine Loads, Stores and PHINodes. | ||||||
| 5636 | if (!isa<LoadInst>(I) && !isa<StoreInst>(I) && !isa<PHINode>(I)) | ||||||
| 5637 | continue; | ||||||
| 5638 | |||||||
| 5639 | // Examine PHI nodes that are reduction variables. Update the type to | ||||||
| 5640 | // account for the recurrence type. | ||||||
| 5641 | if (auto *PN = dyn_cast<PHINode>(&I)) { | ||||||
| 5642 | if (!Legal->isReductionVariable(PN)) | ||||||
| 5643 | continue; | ||||||
| 5644 | const RecurrenceDescriptor &RdxDesc = | ||||||
| 5645 | Legal->getReductionVars().find(PN)->second; | ||||||
| 5646 | if (PreferInLoopReductions || useOrderedReductions(RdxDesc) || | ||||||
| 5647 | TTI.preferInLoopReduction(RdxDesc.getOpcode(), | ||||||
| 5648 | RdxDesc.getRecurrenceType(), | ||||||
| 5649 | TargetTransformInfo::ReductionFlags())) | ||||||
| 5650 | continue; | ||||||
| 5651 | T = RdxDesc.getRecurrenceType(); | ||||||
| 5652 | } | ||||||
| 5653 | |||||||
| 5654 | // Examine the stored values. | ||||||
| 5655 | if (auto *ST = dyn_cast<StoreInst>(&I)) | ||||||
| 5656 | T = ST->getValueOperand()->getType(); | ||||||
| 5657 | |||||||
| 5658 | assert(T->isSized() &&(static_cast <bool> (T->isSized() && "Expected the load/store/recurrence type to be sized" ) ? void (0) : __assert_fail ("T->isSized() && \"Expected the load/store/recurrence type to be sized\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5659, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 5659 | "Expected the load/store/recurrence type to be sized")(static_cast <bool> (T->isSized() && "Expected the load/store/recurrence type to be sized" ) ? void (0) : __assert_fail ("T->isSized() && \"Expected the load/store/recurrence type to be sized\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5659, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 5660 | |||||||
| 5661 | ElementTypesInLoop.insert(T); | ||||||
| 5662 | } | ||||||
| 5663 | } | ||||||
| 5664 | } | ||||||
| 5665 | |||||||
| 5666 | unsigned | ||||||
| 5667 | LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF, | ||||||
| 5668 | InstructionCost LoopCost) { | ||||||
| 5669 | // -- The interleave heuristics -- | ||||||
| 5670 | // We interleave the loop in order to expose ILP and reduce the loop overhead. | ||||||
| 5671 | // There are many micro-architectural considerations that we can't predict | ||||||
| 5672 | // at this level. For example, frontend pressure (on decode or fetch) due to | ||||||
| 5673 | // code size, or the number and capabilities of the execution ports. | ||||||
| 5674 | // | ||||||
| 5675 | // We use the following heuristics to select the interleave count: | ||||||
| 5676 | // 1. If the code has reductions, then we interleave to break the cross | ||||||
| 5677 | // iteration dependency. | ||||||
| 5678 | // 2. If the loop is really small, then we interleave to reduce the loop | ||||||
| 5679 | // overhead. | ||||||
| 5680 | // 3. We don't interleave if we think that we will spill registers to memory | ||||||
| 5681 | // due to the increased register pressure. | ||||||
| 5682 | |||||||
| 5683 | if (!isScalarEpilogueAllowed()) | ||||||
| 5684 | return 1; | ||||||
| 5685 | |||||||
| 5686 | // We used the distance for the interleave count. | ||||||
| 5687 | if (Legal->getMaxSafeDepDistBytes() != -1U) | ||||||
| 5688 | return 1; | ||||||
| 5689 | |||||||
| 5690 | auto BestKnownTC = getSmallBestKnownTC(*PSE.getSE(), TheLoop); | ||||||
| 5691 | const bool HasReductions = !Legal->getReductionVars().empty(); | ||||||
| 5692 | // Do not interleave loops with a relatively small known or estimated trip | ||||||
| 5693 | // count. But we will interleave when InterleaveSmallLoopScalarReduction is | ||||||
| 5694 | // enabled, and the code has scalar reductions(HasReductions && VF = 1), | ||||||
| 5695 | // because with the above conditions interleaving can expose ILP and break | ||||||
| 5696 | // cross iteration dependences for reductions. | ||||||
| 5697 | if (BestKnownTC && (*BestKnownTC < TinyTripCountInterleaveThreshold) && | ||||||
| 5698 | !(InterleaveSmallLoopScalarReduction && HasReductions && VF.isScalar())) | ||||||
| 5699 | return 1; | ||||||
| 5700 | |||||||
| 5701 | // If we did not calculate the cost for VF (because the user selected the VF) | ||||||
| 5702 | // then we calculate the cost of VF here. | ||||||
| 5703 | if (LoopCost == 0) { | ||||||
| 5704 | LoopCost = expectedCost(VF).first; | ||||||
| 5705 | assert(LoopCost.isValid() && "Expected to have chosen a VF with valid cost")(static_cast <bool> (LoopCost.isValid() && "Expected to have chosen a VF with valid cost" ) ? void (0) : __assert_fail ("LoopCost.isValid() && \"Expected to have chosen a VF with valid cost\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5705, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 5706 | |||||||
| 5707 | // Loop body is free and there is no need for interleaving. | ||||||
| 5708 | if (LoopCost == 0) | ||||||
| 5709 | return 1; | ||||||
| 5710 | } | ||||||
| 5711 | |||||||
| 5712 | RegisterUsage R = calculateRegisterUsage({VF})[0]; | ||||||
| 5713 | // We divide by these constants so assume that we have at least one | ||||||
| 5714 | // instruction that uses at least one register. | ||||||
| 5715 | for (auto& pair : R.MaxLocalUsers) { | ||||||
| 5716 | pair.second = std::max(pair.second, 1U); | ||||||
| 5717 | } | ||||||
| 5718 | |||||||
| 5719 | // We calculate the interleave count using the following formula. | ||||||
| 5720 | // Subtract the number of loop invariants from the number of available | ||||||
| 5721 | // registers. These registers are used by all of the interleaved instances. | ||||||
| 5722 | // Next, divide the remaining registers by the number of registers that is | ||||||
| 5723 | // required by the loop, in order to estimate how many parallel instances | ||||||
| 5724 | // fit without causing spills. All of this is rounded down if necessary to be | ||||||
| 5725 | // a power of two. We want power of two interleave count to simplify any | ||||||
| 5726 | // addressing operations or alignment considerations. | ||||||
| 5727 | // We also want power of two interleave counts to ensure that the induction | ||||||
| 5728 | // variable of the vector loop wraps to zero, when tail is folded by masking; | ||||||
| 5729 | // this currently happens when OptForSize, in which case IC is set to 1 above. | ||||||
| 5730 | unsigned IC = UINT_MAX(2147483647 *2U +1U); | ||||||
| 5731 | |||||||
| 5732 | for (auto& pair : R.MaxLocalUsers) { | ||||||
| 5733 | unsigned TargetNumRegisters = TTI.getNumberOfRegisters(pair.first); | ||||||
| 5734 | LLVM_DEBUG(dbgs() << "LV: The target has " << TargetNumRegistersdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The target has " << TargetNumRegisters << " registers of " << TTI.getRegisterClassName (pair.first) << " register class\n"; } } while (false) | ||||||
| 5735 | << " registers of "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The target has " << TargetNumRegisters << " registers of " << TTI.getRegisterClassName (pair.first) << " register class\n"; } } while (false) | ||||||
| 5736 | << TTI.getRegisterClassName(pair.first) << " register class\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The target has " << TargetNumRegisters << " registers of " << TTI.getRegisterClassName (pair.first) << " register class\n"; } } while (false); | ||||||
| 5737 | if (VF.isScalar()) { | ||||||
| 5738 | if (ForceTargetNumScalarRegs.getNumOccurrences() > 0) | ||||||
| 5739 | TargetNumRegisters = ForceTargetNumScalarRegs; | ||||||
| 5740 | } else { | ||||||
| 5741 | if (ForceTargetNumVectorRegs.getNumOccurrences() > 0) | ||||||
| 5742 | TargetNumRegisters = ForceTargetNumVectorRegs; | ||||||
| 5743 | } | ||||||
| 5744 | unsigned MaxLocalUsers = pair.second; | ||||||
| 5745 | unsigned LoopInvariantRegs = 0; | ||||||
| 5746 | if (R.LoopInvariantRegs.find(pair.first) != R.LoopInvariantRegs.end()) | ||||||
| 5747 | LoopInvariantRegs = R.LoopInvariantRegs[pair.first]; | ||||||
| 5748 | |||||||
| 5749 | unsigned TmpIC = PowerOf2Floor((TargetNumRegisters - LoopInvariantRegs) / MaxLocalUsers); | ||||||
| 5750 | // Don't count the induction variable as interleaved. | ||||||
| 5751 | if (EnableIndVarRegisterHeur) { | ||||||
| 5752 | TmpIC = | ||||||
| 5753 | PowerOf2Floor((TargetNumRegisters - LoopInvariantRegs - 1) / | ||||||
| 5754 | std::max(1U, (MaxLocalUsers - 1))); | ||||||
| 5755 | } | ||||||
| 5756 | |||||||
| 5757 | IC = std::min(IC, TmpIC); | ||||||
| 5758 | } | ||||||
| 5759 | |||||||
| 5760 | // Clamp the interleave ranges to reasonable counts. | ||||||
| 5761 | unsigned MaxInterleaveCount = | ||||||
| 5762 | TTI.getMaxInterleaveFactor(VF.getKnownMinValue()); | ||||||
| 5763 | |||||||
| 5764 | // Check if the user has overridden the max. | ||||||
| 5765 | if (VF.isScalar()) { | ||||||
| 5766 | if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0) | ||||||
| 5767 | MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor; | ||||||
| 5768 | } else { | ||||||
| 5769 | if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0) | ||||||
| 5770 | MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor; | ||||||
| 5771 | } | ||||||
| 5772 | |||||||
| 5773 | // If trip count is known or estimated compile time constant, limit the | ||||||
| 5774 | // interleave count to be less than the trip count divided by VF, provided it | ||||||
| 5775 | // is at least 1. | ||||||
| 5776 | // | ||||||
| 5777 | // For scalable vectors we can't know if interleaving is beneficial. It may | ||||||
| 5778 | // not be beneficial for small loops if none of the lanes in the second vector | ||||||
| 5779 | // iterations is enabled. However, for larger loops, there is likely to be a | ||||||
| 5780 | // similar benefit as for fixed-width vectors. For now, we choose to leave | ||||||
| 5781 | // the InterleaveCount as if vscale is '1', although if some information about | ||||||
| 5782 | // the vector is known (e.g. min vector size), we can make a better decision. | ||||||
| 5783 | if (BestKnownTC) { | ||||||
| 5784 | MaxInterleaveCount = | ||||||
| 5785 | std::min(*BestKnownTC / VF.getKnownMinValue(), MaxInterleaveCount); | ||||||
| 5786 | // Make sure MaxInterleaveCount is greater than 0. | ||||||
| 5787 | MaxInterleaveCount = std::max(1u, MaxInterleaveCount); | ||||||
| 5788 | } | ||||||
| 5789 | |||||||
| 5790 | assert(MaxInterleaveCount > 0 &&(static_cast <bool> (MaxInterleaveCount > 0 && "Maximum interleave count must be greater than 0") ? void (0 ) : __assert_fail ("MaxInterleaveCount > 0 && \"Maximum interleave count must be greater than 0\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5791, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 5791 | "Maximum interleave count must be greater than 0")(static_cast <bool> (MaxInterleaveCount > 0 && "Maximum interleave count must be greater than 0") ? void (0 ) : __assert_fail ("MaxInterleaveCount > 0 && \"Maximum interleave count must be greater than 0\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5791, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 5792 | |||||||
| 5793 | // Clamp the calculated IC to be between the 1 and the max interleave count | ||||||
| 5794 | // that the target and trip count allows. | ||||||
| 5795 | if (IC > MaxInterleaveCount) | ||||||
| 5796 | IC = MaxInterleaveCount; | ||||||
| 5797 | else | ||||||
| 5798 | // Make sure IC is greater than 0. | ||||||
| 5799 | IC = std::max(1u, IC); | ||||||
| 5800 | |||||||
| 5801 | assert(IC > 0 && "Interleave count must be greater than 0.")(static_cast <bool> (IC > 0 && "Interleave count must be greater than 0." ) ? void (0) : __assert_fail ("IC > 0 && \"Interleave count must be greater than 0.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5801, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 5802 | |||||||
| 5803 | // Interleave if we vectorized this loop and there is a reduction that could | ||||||
| 5804 | // benefit from interleaving. | ||||||
| 5805 | if (VF.isVector() && HasReductions) { | ||||||
| 5806 | LLVM_DEBUG(dbgs() << "LV: Interleaving because of reductions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving because of reductions.\n" ; } } while (false); | ||||||
| 5807 | return IC; | ||||||
| 5808 | } | ||||||
| 5809 | |||||||
| 5810 | // For any scalar loop that either requires runtime checks or predication we | ||||||
| 5811 | // are better off leaving this to the unroller. Note that if we've already | ||||||
| 5812 | // vectorized the loop we will have done the runtime check and so interleaving | ||||||
| 5813 | // won't require further checks. | ||||||
| 5814 | bool ScalarInterleavingRequiresPredication = | ||||||
| 5815 | (VF.isScalar() && any_of(TheLoop->blocks(), [this](BasicBlock *BB) { | ||||||
| 5816 | return Legal->blockNeedsPredication(BB); | ||||||
| 5817 | })); | ||||||
| 5818 | bool ScalarInterleavingRequiresRuntimePointerCheck = | ||||||
| 5819 | (VF.isScalar() && Legal->getRuntimePointerChecking()->Need); | ||||||
| 5820 | |||||||
| 5821 | // We want to interleave small loops in order to reduce the loop overhead and | ||||||
| 5822 | // potentially expose ILP opportunities. | ||||||
| 5823 | LLVM_DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop cost is " << LoopCost << '\n' << "LV: IC is " << IC << '\n' << "LV: VF is " << VF << '\n'; } } while (false) | ||||||
| 5824 | << "LV: IC is " << IC << '\n'do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop cost is " << LoopCost << '\n' << "LV: IC is " << IC << '\n' << "LV: VF is " << VF << '\n'; } } while (false) | ||||||
| 5825 | << "LV: VF is " << VF << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop cost is " << LoopCost << '\n' << "LV: IC is " << IC << '\n' << "LV: VF is " << VF << '\n'; } } while (false); | ||||||
| 5826 | const bool AggressivelyInterleaveReductions = | ||||||
| 5827 | TTI.enableAggressiveInterleaving(HasReductions); | ||||||
| 5828 | if (!ScalarInterleavingRequiresRuntimePointerCheck && | ||||||
| 5829 | !ScalarInterleavingRequiresPredication && LoopCost < SmallLoopCost) { | ||||||
| 5830 | // We assume that the cost overhead is 1 and we use the cost model | ||||||
| 5831 | // to estimate the cost of the loop and interleave until the cost of the | ||||||
| 5832 | // loop overhead is about 5% of the cost of the loop. | ||||||
| 5833 | unsigned SmallIC = std::min( | ||||||
| 5834 | IC, (unsigned)PowerOf2Floor(SmallLoopCost / *LoopCost.getValue())); | ||||||
| 5835 | |||||||
| 5836 | // Interleave until store/load ports (estimated by max interleave count) are | ||||||
| 5837 | // saturated. | ||||||
| 5838 | unsigned NumStores = Legal->getNumStores(); | ||||||
| 5839 | unsigned NumLoads = Legal->getNumLoads(); | ||||||
| 5840 | unsigned StoresIC = IC / (NumStores ? NumStores : 1); | ||||||
| 5841 | unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1); | ||||||
| 5842 | |||||||
| 5843 | // There is little point in interleaving for reductions containing selects | ||||||
| 5844 | // and compares when VF=1 since it may just create more overhead than it's | ||||||
| 5845 | // worth for loops with small trip counts. This is because we still have to | ||||||
| 5846 | // do the final reduction after the loop. | ||||||
| 5847 | bool HasSelectCmpReductions = | ||||||
| 5848 | HasReductions && | ||||||
| 5849 | any_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool { | ||||||
| 5850 | const RecurrenceDescriptor &RdxDesc = Reduction.second; | ||||||
| 5851 | return RecurrenceDescriptor::isSelectCmpRecurrenceKind( | ||||||
| 5852 | RdxDesc.getRecurrenceKind()); | ||||||
| 5853 | }); | ||||||
| 5854 | if (HasSelectCmpReductions) { | ||||||
| 5855 | LLVM_DEBUG(dbgs() << "LV: Not interleaving select-cmp reductions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not interleaving select-cmp reductions.\n" ; } } while (false); | ||||||
| 5856 | return 1; | ||||||
| 5857 | } | ||||||
| 5858 | |||||||
| 5859 | // If we have a scalar reduction (vector reductions are already dealt with | ||||||
| 5860 | // by this point), we can increase the critical path length if the loop | ||||||
| 5861 | // we're interleaving is inside another loop. For tree-wise reductions | ||||||
| 5862 | // set the limit to 2, and for ordered reductions it's best to disable | ||||||
| 5863 | // interleaving entirely. | ||||||
| 5864 | if (HasReductions && TheLoop->getLoopDepth() > 1) { | ||||||
| 5865 | bool HasOrderedReductions = | ||||||
| 5866 | any_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool { | ||||||
| 5867 | const RecurrenceDescriptor &RdxDesc = Reduction.second; | ||||||
| 5868 | return RdxDesc.isOrdered(); | ||||||
| 5869 | }); | ||||||
| 5870 | if (HasOrderedReductions) { | ||||||
| 5871 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not interleaving scalar ordered reductions.\n" ; } } while (false) | ||||||
| 5872 | dbgs() << "LV: Not interleaving scalar ordered reductions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not interleaving scalar ordered reductions.\n" ; } } while (false); | ||||||
| 5873 | return 1; | ||||||
| 5874 | } | ||||||
| 5875 | |||||||
| 5876 | unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC); | ||||||
| 5877 | SmallIC = std::min(SmallIC, F); | ||||||
| 5878 | StoresIC = std::min(StoresIC, F); | ||||||
| 5879 | LoadsIC = std::min(LoadsIC, F); | ||||||
| 5880 | } | ||||||
| 5881 | |||||||
| 5882 | if (EnableLoadStoreRuntimeInterleave && | ||||||
| 5883 | std::max(StoresIC, LoadsIC) > SmallIC) { | ||||||
| 5884 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving to saturate store or load ports.\n" ; } } while (false) | ||||||
| 5885 | dbgs() << "LV: Interleaving to saturate store or load ports.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving to saturate store or load ports.\n" ; } } while (false); | ||||||
| 5886 | return std::max(StoresIC, LoadsIC); | ||||||
| 5887 | } | ||||||
| 5888 | |||||||
| 5889 | // If there are scalar reductions and TTI has enabled aggressive | ||||||
| 5890 | // interleaving for reductions, we will interleave to expose ILP. | ||||||
| 5891 | if (InterleaveSmallLoopScalarReduction && VF.isScalar() && | ||||||
| 5892 | AggressivelyInterleaveReductions) { | ||||||
| 5893 | LLVM_DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving to expose ILP.\n" ; } } while (false); | ||||||
| 5894 | // Interleave no less than SmallIC but not as aggressive as the normal IC | ||||||
| 5895 | // to satisfy the rare situation when resources are too limited. | ||||||
| 5896 | return std::max(IC / 2, SmallIC); | ||||||
| 5897 | } else { | ||||||
| 5898 | LLVM_DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving to reduce branch cost.\n" ; } } while (false); | ||||||
| 5899 | return SmallIC; | ||||||
| 5900 | } | ||||||
| 5901 | } | ||||||
| 5902 | |||||||
| 5903 | // Interleave if this is a large loop (small loops are already dealt with by | ||||||
| 5904 | // this point) that could benefit from interleaving. | ||||||
| 5905 | if (AggressivelyInterleaveReductions) { | ||||||
| 5906 | LLVM_DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving to expose ILP.\n" ; } } while (false); | ||||||
| 5907 | return IC; | ||||||
| 5908 | } | ||||||
| 5909 | |||||||
| 5910 | LLVM_DEBUG(dbgs() << "LV: Not Interleaving.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not Interleaving.\n" ; } } while (false); | ||||||
| 5911 | return 1; | ||||||
| 5912 | } | ||||||
| 5913 | |||||||
| 5914 | SmallVector<LoopVectorizationCostModel::RegisterUsage, 8> | ||||||
| 5915 | LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<ElementCount> VFs) { | ||||||
| 5916 | // This function calculates the register usage by measuring the highest number | ||||||
| 5917 | // of values that are alive at a single location. Obviously, this is a very | ||||||
| 5918 | // rough estimation. We scan the loop in a topological order in order and | ||||||
| 5919 | // assign a number to each instruction. We use RPO to ensure that defs are | ||||||
| 5920 | // met before their users. We assume that each instruction that has in-loop | ||||||
| 5921 | // users starts an interval. We record every time that an in-loop value is | ||||||
| 5922 | // used, so we have a list of the first and last occurrences of each | ||||||
| 5923 | // instruction. Next, we transpose this data structure into a multi map that | ||||||
| 5924 | // holds the list of intervals that *end* at a specific location. This multi | ||||||
| 5925 | // map allows us to perform a linear search. We scan the instructions linearly | ||||||
| 5926 | // and record each time that a new interval starts, by placing it in a set. | ||||||
| 5927 | // If we find this value in the multi-map then we remove it from the set. | ||||||
| 5928 | // The max register usage is the maximum size of the set. | ||||||
| 5929 | // We also search for instructions that are defined outside the loop, but are | ||||||
| 5930 | // used inside the loop. We need this number separately from the max-interval | ||||||
| 5931 | // usage number because when we unroll, loop-invariant values do not take | ||||||
| 5932 | // more register. | ||||||
| 5933 | LoopBlocksDFS DFS(TheLoop); | ||||||
| 5934 | DFS.perform(LI); | ||||||
| 5935 | |||||||
| 5936 | RegisterUsage RU; | ||||||
| 5937 | |||||||
| 5938 | // Each 'key' in the map opens a new interval. The values | ||||||
| 5939 | // of the map are the index of the 'last seen' usage of the | ||||||
| 5940 | // instruction that is the key. | ||||||
| 5941 | using IntervalMap = DenseMap<Instruction *, unsigned>; | ||||||
| 5942 | |||||||
| 5943 | // Maps instruction to its index. | ||||||
| 5944 | SmallVector<Instruction *, 64> IdxToInstr; | ||||||
| 5945 | // Marks the end of each interval. | ||||||
| 5946 | IntervalMap EndPoint; | ||||||
| 5947 | // Saves the list of instruction indices that are used in the loop. | ||||||
| 5948 | SmallPtrSet<Instruction *, 8> Ends; | ||||||
| 5949 | // Saves the list of values that are used in the loop but are defined outside | ||||||
| 5950 | // the loop (not including non-instruction values such as arguments and | ||||||
| 5951 | // constants). | ||||||
| 5952 | SmallPtrSet<Value *, 8> LoopInvariants; | ||||||
| 5953 | |||||||
| 5954 | for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) { | ||||||
| 5955 | for (Instruction &I : BB->instructionsWithoutDebug()) { | ||||||
| 5956 | IdxToInstr.push_back(&I); | ||||||
| 5957 | |||||||
| 5958 | // Save the end location of each USE. | ||||||
| 5959 | for (Value *U : I.operands()) { | ||||||
| 5960 | auto *Instr = dyn_cast<Instruction>(U); | ||||||
| 5961 | |||||||
| 5962 | // Ignore non-instruction values such as arguments, constants, etc. | ||||||
| 5963 | // FIXME: Might need some motivation why these values are ignored. If | ||||||
| 5964 | // for example an argument is used inside the loop it will increase the | ||||||
| 5965 | // register pressure (so shouldn't we add it to LoopInvariants). | ||||||
| 5966 | if (!Instr) | ||||||
| 5967 | continue; | ||||||
| 5968 | |||||||
| 5969 | // If this instruction is outside the loop then record it and continue. | ||||||
| 5970 | if (!TheLoop->contains(Instr)) { | ||||||
| 5971 | LoopInvariants.insert(Instr); | ||||||
| 5972 | continue; | ||||||
| 5973 | } | ||||||
| 5974 | |||||||
| 5975 | // Overwrite previous end points. | ||||||
| 5976 | EndPoint[Instr] = IdxToInstr.size(); | ||||||
| 5977 | Ends.insert(Instr); | ||||||
| 5978 | } | ||||||
| 5979 | } | ||||||
| 5980 | } | ||||||
| 5981 | |||||||
| 5982 | // Saves the list of intervals that end with the index in 'key'. | ||||||
| 5983 | using InstrList = SmallVector<Instruction *, 2>; | ||||||
| 5984 | DenseMap<unsigned, InstrList> TransposeEnds; | ||||||
| 5985 | |||||||
| 5986 | // Transpose the EndPoints to a list of values that end at each index. | ||||||
| 5987 | for (auto &Interval : EndPoint) | ||||||
| 5988 | TransposeEnds[Interval.second].push_back(Interval.first); | ||||||
| 5989 | |||||||
| 5990 | SmallPtrSet<Instruction *, 8> OpenIntervals; | ||||||
| 5991 | SmallVector<RegisterUsage, 8> RUs(VFs.size()); | ||||||
| 5992 | SmallVector<SmallMapVector<unsigned, unsigned, 4>, 8> MaxUsages(VFs.size()); | ||||||
| 5993 | |||||||
| 5994 | LLVM_DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV(REG): Calculating max register usage:\n" ; } } while (false); | ||||||
| 5995 | |||||||
| 5996 | const auto &TTICapture = TTI; | ||||||
| 5997 | auto GetRegUsage = [&TTICapture](Type *Ty, ElementCount VF) -> unsigned { | ||||||
| 5998 | if (Ty->isTokenTy() || !VectorType::isValidElementType(Ty)) | ||||||
| 5999 | return 0; | ||||||
| 6000 | return TTICapture.getRegUsageForType(VectorType::get(Ty, VF)); | ||||||
| 6001 | }; | ||||||
| 6002 | |||||||
| 6003 | for (unsigned int i = 0, s = IdxToInstr.size(); i < s; ++i) { | ||||||
| 6004 | Instruction *I = IdxToInstr[i]; | ||||||
| 6005 | |||||||
| 6006 | // Remove all of the instructions that end at this location. | ||||||
| 6007 | InstrList &List = TransposeEnds[i]; | ||||||
| 6008 | for (Instruction *ToRemove : List) | ||||||
| 6009 | OpenIntervals.erase(ToRemove); | ||||||
| 6010 | |||||||
| 6011 | // Ignore instructions that are never used within the loop. | ||||||
| 6012 | if (!Ends.count(I)) | ||||||
| 6013 | continue; | ||||||
| 6014 | |||||||
| 6015 | // Skip ignored values. | ||||||
| 6016 | if (ValuesToIgnore.count(I)) | ||||||
| 6017 | continue; | ||||||
| 6018 | |||||||
| 6019 | // For each VF find the maximum usage of registers. | ||||||
| 6020 | for (unsigned j = 0, e = VFs.size(); j < e; ++j) { | ||||||
| 6021 | // Count the number of registers used, per register class, given all open | ||||||
| 6022 | // intervals. | ||||||
| 6023 | // Note that elements in this SmallMapVector will be default constructed | ||||||
| 6024 | // as 0. So we can use "RegUsage[ClassID] += n" in the code below even if | ||||||
| 6025 | // there is no previous entry for ClassID. | ||||||
| 6026 | SmallMapVector<unsigned, unsigned, 4> RegUsage; | ||||||
| 6027 | |||||||
| 6028 | if (VFs[j].isScalar()) { | ||||||
| 6029 | for (auto *Inst : OpenIntervals) { | ||||||
| 6030 | unsigned ClassID = | ||||||
| 6031 | TTI.getRegisterClassForType(false, Inst->getType()); | ||||||
| 6032 | // FIXME: The target might use more than one register for the type | ||||||
| 6033 | // even in the scalar case. | ||||||
| 6034 | RegUsage[ClassID] += 1; | ||||||
| 6035 | } | ||||||
| 6036 | } else { | ||||||
| 6037 | collectUniformsAndScalars(VFs[j]); | ||||||
| 6038 | for (auto *Inst : OpenIntervals) { | ||||||
| 6039 | // Skip ignored values for VF > 1. | ||||||
| 6040 | if (VecValuesToIgnore.count(Inst)) | ||||||
| 6041 | continue; | ||||||
| 6042 | if (isScalarAfterVectorization(Inst, VFs[j])) { | ||||||
| 6043 | unsigned ClassID = | ||||||
| 6044 | TTI.getRegisterClassForType(false, Inst->getType()); | ||||||
| 6045 | // FIXME: The target might use more than one register for the type | ||||||
| 6046 | // even in the scalar case. | ||||||
| 6047 | RegUsage[ClassID] += 1; | ||||||
| 6048 | } else { | ||||||
| 6049 | unsigned ClassID = | ||||||
| 6050 | TTI.getRegisterClassForType(true, Inst->getType()); | ||||||
| 6051 | RegUsage[ClassID] += GetRegUsage(Inst->getType(), VFs[j]); | ||||||
| 6052 | } | ||||||
| 6053 | } | ||||||
| 6054 | } | ||||||
| 6055 | |||||||
| 6056 | for (auto& pair : RegUsage) { | ||||||
| 6057 | auto &Entry = MaxUsages[j][pair.first]; | ||||||
| 6058 | Entry = std::max(Entry, pair.second); | ||||||
| 6059 | } | ||||||
| 6060 | } | ||||||
| 6061 | |||||||
| 6062 | LLVM_DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV(REG): At #" << i << " Interval # " << OpenIntervals.size() << '\n'; } } while (false) | ||||||
| 6063 | << OpenIntervals.size() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV(REG): At #" << i << " Interval # " << OpenIntervals.size() << '\n'; } } while (false); | ||||||
| 6064 | |||||||
| 6065 | // Add the current instruction to the list of open intervals. | ||||||
| 6066 | OpenIntervals.insert(I); | ||||||
| 6067 | } | ||||||
| 6068 | |||||||
| 6069 | for (unsigned i = 0, e = VFs.size(); i < e; ++i) { | ||||||
| 6070 | // Note that elements in this SmallMapVector will be default constructed | ||||||
| 6071 | // as 0. So we can use "Invariant[ClassID] += n" in the code below even if | ||||||
| 6072 | // there is no previous entry for ClassID. | ||||||
| 6073 | SmallMapVector<unsigned, unsigned, 4> Invariant; | ||||||
| 6074 | |||||||
| 6075 | for (auto *Inst : LoopInvariants) { | ||||||
| 6076 | // FIXME: The target might use more than one register for the type | ||||||
| 6077 | // even in the scalar case. | ||||||
| 6078 | unsigned Usage = | ||||||
| 6079 | VFs[i].isScalar() ? 1 : GetRegUsage(Inst->getType(), VFs[i]); | ||||||
| 6080 | unsigned ClassID = | ||||||
| 6081 | TTI.getRegisterClassForType(VFs[i].isVector(), Inst->getType()); | ||||||
| 6082 | Invariant[ClassID] += Usage; | ||||||
| 6083 | } | ||||||
| 6084 | |||||||
| 6085 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | ||||||
| 6086 | dbgs() << "LV(REG): VF = " << VFs[i] << '\n';do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | ||||||
| 6087 | dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | ||||||
| 6088 | << " item\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | ||||||
| 6089 | for (const auto &pair : MaxUsages[i]) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | ||||||
| 6090 | dbgs() << "LV(REG): RegisterClass: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | ||||||
| 6091 | << TTI.getRegisterClassName(pair.first) << ", " << pair.seconddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | ||||||
| 6092 | << " registers\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | ||||||
| 6093 | }do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | ||||||
| 6094 | dbgs() << "LV(REG): Found invariant usage: " << Invariant.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | ||||||
| 6095 | << " item\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | ||||||
| 6096 | for (const auto &pair : Invariant) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | ||||||
| 6097 | dbgs() << "LV(REG): RegisterClass: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | ||||||
| 6098 | << TTI.getRegisterClassName(pair.first) << ", " << pair.seconddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | ||||||
| 6099 | << " registers\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | ||||||
| 6100 | }do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false) | ||||||
| 6101 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "LV(REG): VF = " << VFs[i] << '\n'; dbgs() << "LV(REG): Found max usage: " << MaxUsages[i].size() << " item\n"; for (const auto &pair : MaxUsages[i]) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } dbgs() << "LV(REG): Found invariant usage: " << Invariant.size() << " item\n"; for (const auto &pair : Invariant) { dbgs() << "LV(REG): RegisterClass: " << TTI.getRegisterClassName(pair.first) << ", " << pair.second << " registers\n"; } }; } } while (false); | ||||||
| 6102 | |||||||
| 6103 | RU.LoopInvariantRegs = Invariant; | ||||||
| 6104 | RU.MaxLocalUsers = MaxUsages[i]; | ||||||
| 6105 | RUs[i] = RU; | ||||||
| 6106 | } | ||||||
| 6107 | |||||||
| 6108 | return RUs; | ||||||
| 6109 | } | ||||||
| 6110 | |||||||
| 6111 | bool LoopVectorizationCostModel::useEmulatedMaskMemRefHack(Instruction *I, | ||||||
| 6112 | ElementCount VF) { | ||||||
| 6113 | // TODO: Cost model for emulated masked load/store is completely | ||||||
| 6114 | // broken. This hack guides the cost model to use an artificially | ||||||
| 6115 | // high enough value to practically disable vectorization with such | ||||||
| 6116 | // operations, except where previously deployed legality hack allowed | ||||||
| 6117 | // using very low cost values. This is to avoid regressions coming simply | ||||||
| 6118 | // from moving "masked load/store" check from legality to cost model. | ||||||
| 6119 | // Masked Load/Gather emulation was previously never allowed. | ||||||
| 6120 | // Limited number of Masked Store/Scatter emulation was allowed. | ||||||
| 6121 | assert((isPredicatedInst(I)) &&(static_cast <bool> ((isPredicatedInst(I)) && "Expecting a scalar emulated instruction" ) ? void (0) : __assert_fail ("(isPredicatedInst(I)) && \"Expecting a scalar emulated instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6122, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 6122 | "Expecting a scalar emulated instruction")(static_cast <bool> ((isPredicatedInst(I)) && "Expecting a scalar emulated instruction" ) ? void (0) : __assert_fail ("(isPredicatedInst(I)) && \"Expecting a scalar emulated instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6122, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 6123 | return isa<LoadInst>(I) || | ||||||
| 6124 | (isa<StoreInst>(I) && | ||||||
| 6125 | NumPredStores > NumberOfStoresToPredicate); | ||||||
| 6126 | } | ||||||
| 6127 | |||||||
| 6128 | void LoopVectorizationCostModel::collectInstsToScalarize(ElementCount VF) { | ||||||
| 6129 | // If we aren't vectorizing the loop, or if we've already collected the | ||||||
| 6130 | // instructions to scalarize, there's nothing to do. Collection may already | ||||||
| 6131 | // have occurred if we have a user-selected VF and are now computing the | ||||||
| 6132 | // expected cost for interleaving. | ||||||
| 6133 | if (VF.isScalar() || VF.isZero() || | ||||||
| 6134 | InstsToScalarize.find(VF) != InstsToScalarize.end()) | ||||||
| 6135 | return; | ||||||
| 6136 | |||||||
| 6137 | // Initialize a mapping for VF in InstsToScalalarize. If we find that it's | ||||||
| 6138 | // not profitable to scalarize any instructions, the presence of VF in the | ||||||
| 6139 | // map will indicate that we've analyzed it already. | ||||||
| 6140 | ScalarCostsTy &ScalarCostsVF = InstsToScalarize[VF]; | ||||||
| 6141 | |||||||
| 6142 | PredicatedBBsAfterVectorization[VF].clear(); | ||||||
| 6143 | |||||||
| 6144 | // Find all the instructions that are scalar with predication in the loop and | ||||||
| 6145 | // determine if it would be better to not if-convert the blocks they are in. | ||||||
| 6146 | // If so, we also record the instructions to scalarize. | ||||||
| 6147 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||||
| 6148 | if (!blockNeedsPredicationForAnyReason(BB)) | ||||||
| 6149 | continue; | ||||||
| 6150 | for (Instruction &I : *BB) | ||||||
| 6151 | if (isScalarWithPredication(&I, VF)) { | ||||||
| 6152 | ScalarCostsTy ScalarCosts; | ||||||
| 6153 | // Do not apply discount if scalable, because that would lead to | ||||||
| 6154 | // invalid scalarization costs. | ||||||
| 6155 | // Do not apply discount logic if hacked cost is needed | ||||||
| 6156 | // for emulated masked memrefs. | ||||||
| 6157 | if (!VF.isScalable() && !useEmulatedMaskMemRefHack(&I, VF) && | ||||||
| 6158 | computePredInstDiscount(&I, ScalarCosts, VF) >= 0) | ||||||
| 6159 | ScalarCostsVF.insert(ScalarCosts.begin(), ScalarCosts.end()); | ||||||
| 6160 | // Remember that BB will remain after vectorization. | ||||||
| 6161 | PredicatedBBsAfterVectorization[VF].insert(BB); | ||||||
| 6162 | } | ||||||
| 6163 | } | ||||||
| 6164 | } | ||||||
| 6165 | |||||||
| 6166 | InstructionCost LoopVectorizationCostModel::computePredInstDiscount( | ||||||
| 6167 | Instruction *PredInst, ScalarCostsTy &ScalarCosts, ElementCount VF) { | ||||||
| 6168 | assert(!isUniformAfterVectorization(PredInst, VF) &&(static_cast <bool> (!isUniformAfterVectorization(PredInst , VF) && "Instruction marked uniform-after-vectorization will be predicated" ) ? void (0) : __assert_fail ("!isUniformAfterVectorization(PredInst, VF) && \"Instruction marked uniform-after-vectorization will be predicated\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6169, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 6169 | "Instruction marked uniform-after-vectorization will be predicated")(static_cast <bool> (!isUniformAfterVectorization(PredInst , VF) && "Instruction marked uniform-after-vectorization will be predicated" ) ? void (0) : __assert_fail ("!isUniformAfterVectorization(PredInst, VF) && \"Instruction marked uniform-after-vectorization will be predicated\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6169, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 6170 | |||||||
| 6171 | // Initialize the discount to zero, meaning that the scalar version and the | ||||||
| 6172 | // vector version cost the same. | ||||||
| 6173 | InstructionCost Discount = 0; | ||||||
| 6174 | |||||||
| 6175 | // Holds instructions to analyze. The instructions we visit are mapped in | ||||||
| 6176 | // ScalarCosts. Those instructions are the ones that would be scalarized if | ||||||
| 6177 | // we find that the scalar version costs less. | ||||||
| 6178 | SmallVector<Instruction *, 8> Worklist; | ||||||
| 6179 | |||||||
| 6180 | // Returns true if the given instruction can be scalarized. | ||||||
| 6181 | auto canBeScalarized = [&](Instruction *I) -> bool { | ||||||
| 6182 | // We only attempt to scalarize instructions forming a single-use chain | ||||||
| 6183 | // from the original predicated block that would otherwise be vectorized. | ||||||
| 6184 | // Although not strictly necessary, we give up on instructions we know will | ||||||
| 6185 | // already be scalar to avoid traversing chains that are unlikely to be | ||||||
| 6186 | // beneficial. | ||||||
| 6187 | if (!I->hasOneUse() || PredInst->getParent() != I->getParent() || | ||||||
| 6188 | isScalarAfterVectorization(I, VF)) | ||||||
| 6189 | return false; | ||||||
| 6190 | |||||||
| 6191 | // If the instruction is scalar with predication, it will be analyzed | ||||||
| 6192 | // separately. We ignore it within the context of PredInst. | ||||||
| 6193 | if (isScalarWithPredication(I, VF)) | ||||||
| 6194 | return false; | ||||||
| 6195 | |||||||
| 6196 | // If any of the instruction's operands are uniform after vectorization, | ||||||
| 6197 | // the instruction cannot be scalarized. This prevents, for example, a | ||||||
| 6198 | // masked load from being scalarized. | ||||||
| 6199 | // | ||||||
| 6200 | // We assume we will only emit a value for lane zero of an instruction | ||||||
| 6201 | // marked uniform after vectorization, rather than VF identical values. | ||||||
| 6202 | // Thus, if we scalarize an instruction that uses a uniform, we would | ||||||
| 6203 | // create uses of values corresponding to the lanes we aren't emitting code | ||||||
| 6204 | // for. This behavior can be changed by allowing getScalarValue to clone | ||||||
| 6205 | // the lane zero values for uniforms rather than asserting. | ||||||
| 6206 | for (Use &U : I->operands()) | ||||||
| 6207 | if (auto *J = dyn_cast<Instruction>(U.get())) | ||||||
| 6208 | if (isUniformAfterVectorization(J, VF)) | ||||||
| 6209 | return false; | ||||||
| 6210 | |||||||
| 6211 | // Otherwise, we can scalarize the instruction. | ||||||
| 6212 | return true; | ||||||
| 6213 | }; | ||||||
| 6214 | |||||||
| 6215 | // Compute the expected cost discount from scalarizing the entire expression | ||||||
| 6216 | // feeding the predicated instruction. We currently only consider expressions | ||||||
| 6217 | // that are single-use instruction chains. | ||||||
| 6218 | Worklist.push_back(PredInst); | ||||||
| 6219 | while (!Worklist.empty()) { | ||||||
| 6220 | Instruction *I = Worklist.pop_back_val(); | ||||||
| 6221 | |||||||
| 6222 | // If we've already analyzed the instruction, there's nothing to do. | ||||||
| 6223 | if (ScalarCosts.find(I) != ScalarCosts.end()) | ||||||
| 6224 | continue; | ||||||
| 6225 | |||||||
| 6226 | // Compute the cost of the vector instruction. Note that this cost already | ||||||
| 6227 | // includes the scalarization overhead of the predicated instruction. | ||||||
| 6228 | InstructionCost VectorCost = getInstructionCost(I, VF).first; | ||||||
| 6229 | |||||||
| 6230 | // Compute the cost of the scalarized instruction. This cost is the cost of | ||||||
| 6231 | // the instruction as if it wasn't if-converted and instead remained in the | ||||||
| 6232 | // predicated block. We will scale this cost by block probability after | ||||||
| 6233 | // computing the scalarization overhead. | ||||||
| 6234 | InstructionCost ScalarCost = | ||||||
| 6235 | VF.getFixedValue() * | ||||||
| 6236 | getInstructionCost(I, ElementCount::getFixed(1)).first; | ||||||
| 6237 | |||||||
| 6238 | // Compute the scalarization overhead of needed insertelement instructions | ||||||
| 6239 | // and phi nodes. | ||||||
| 6240 | if (isScalarWithPredication(I, VF) && !I->getType()->isVoidTy()) { | ||||||
| 6241 | ScalarCost += TTI.getScalarizationOverhead( | ||||||
| 6242 | cast<VectorType>(ToVectorTy(I->getType(), VF)), | ||||||
| 6243 | APInt::getAllOnes(VF.getFixedValue()), true, false); | ||||||
| 6244 | ScalarCost += | ||||||
| 6245 | VF.getFixedValue() * | ||||||
| 6246 | TTI.getCFInstrCost(Instruction::PHI, TTI::TCK_RecipThroughput); | ||||||
| 6247 | } | ||||||
| 6248 | |||||||
| 6249 | // Compute the scalarization overhead of needed extractelement | ||||||
| 6250 | // instructions. For each of the instruction's operands, if the operand can | ||||||
| 6251 | // be scalarized, add it to the worklist; otherwise, account for the | ||||||
| 6252 | // overhead. | ||||||
| 6253 | for (Use &U : I->operands()) | ||||||
| 6254 | if (auto *J = dyn_cast<Instruction>(U.get())) { | ||||||
| 6255 | assert(VectorType::isValidElementType(J->getType()) &&(static_cast <bool> (VectorType::isValidElementType(J-> getType()) && "Instruction has non-scalar type") ? void (0) : __assert_fail ("VectorType::isValidElementType(J->getType()) && \"Instruction has non-scalar type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6256, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 6256 | "Instruction has non-scalar type")(static_cast <bool> (VectorType::isValidElementType(J-> getType()) && "Instruction has non-scalar type") ? void (0) : __assert_fail ("VectorType::isValidElementType(J->getType()) && \"Instruction has non-scalar type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6256, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 6257 | if (canBeScalarized(J)) | ||||||
| 6258 | Worklist.push_back(J); | ||||||
| 6259 | else if (needsExtract(J, VF)) { | ||||||
| 6260 | ScalarCost += TTI.getScalarizationOverhead( | ||||||
| 6261 | cast<VectorType>(ToVectorTy(J->getType(), VF)), | ||||||
| 6262 | APInt::getAllOnes(VF.getFixedValue()), false, true); | ||||||
| 6263 | } | ||||||
| 6264 | } | ||||||
| 6265 | |||||||
| 6266 | // Scale the total scalar cost by block probability. | ||||||
| 6267 | ScalarCost /= getReciprocalPredBlockProb(); | ||||||
| 6268 | |||||||
| 6269 | // Compute the discount. A non-negative discount means the vector version | ||||||
| 6270 | // of the instruction costs more, and scalarizing would be beneficial. | ||||||
| 6271 | Discount += VectorCost - ScalarCost; | ||||||
| 6272 | ScalarCosts[I] = ScalarCost; | ||||||
| 6273 | } | ||||||
| 6274 | |||||||
| 6275 | return Discount; | ||||||
| 6276 | } | ||||||
| 6277 | |||||||
| 6278 | LoopVectorizationCostModel::VectorizationCostTy | ||||||
| 6279 | LoopVectorizationCostModel::expectedCost( | ||||||
| 6280 | ElementCount VF, SmallVectorImpl<InstructionVFPair> *Invalid) { | ||||||
| 6281 | VectorizationCostTy Cost; | ||||||
| 6282 | |||||||
| 6283 | // For each block. | ||||||
| 6284 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||||
| 6285 | VectorizationCostTy BlockCost; | ||||||
| 6286 | |||||||
| 6287 | // For each instruction in the old loop. | ||||||
| 6288 | for (Instruction &I : BB->instructionsWithoutDebug()) { | ||||||
| 6289 | // Skip ignored values. | ||||||
| 6290 | if (ValuesToIgnore.count(&I) || | ||||||
| 6291 | (VF.isVector() && VecValuesToIgnore.count(&I))) | ||||||
| 6292 | continue; | ||||||
| 6293 | |||||||
| 6294 | VectorizationCostTy C = getInstructionCost(&I, VF); | ||||||
| 6295 | |||||||
| 6296 | // Check if we should override the cost. | ||||||
| 6297 | if (C.first.isValid() && | ||||||
| 6298 | ForceTargetInstructionCost.getNumOccurrences() > 0) | ||||||
| 6299 | C.first = InstructionCost(ForceTargetInstructionCost); | ||||||
| 6300 | |||||||
| 6301 | // Keep a list of instructions with invalid costs. | ||||||
| 6302 | if (Invalid && !C.first.isValid()) | ||||||
| 6303 | Invalid->emplace_back(&I, VF); | ||||||
| 6304 | |||||||
| 6305 | BlockCost.first += C.first; | ||||||
| 6306 | BlockCost.second |= C.second; | ||||||
| 6307 | LLVM_DEBUG(dbgs() << "LV: Found an estimated cost of " << C.firstdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found an estimated cost of " << C.first << " for VF " << VF << " For instruction: " << I << '\n'; } } while (false) | ||||||
| 6308 | << " for VF " << VF << " For instruction: " << Ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found an estimated cost of " << C.first << " for VF " << VF << " For instruction: " << I << '\n'; } } while (false) | ||||||
| 6309 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found an estimated cost of " << C.first << " for VF " << VF << " For instruction: " << I << '\n'; } } while (false); | ||||||
| 6310 | } | ||||||
| 6311 | |||||||
| 6312 | // If we are vectorizing a predicated block, it will have been | ||||||
| 6313 | // if-converted. This means that the block's instructions (aside from | ||||||
| 6314 | // stores and instructions that may divide by zero) will now be | ||||||
| 6315 | // unconditionally executed. For the scalar case, we may not always execute | ||||||
| 6316 | // the predicated block, if it is an if-else block. Thus, scale the block's | ||||||
| 6317 | // cost by the probability of executing it. blockNeedsPredication from | ||||||
| 6318 | // Legal is used so as to not include all blocks in tail folded loops. | ||||||
| 6319 | if (VF.isScalar() && Legal->blockNeedsPredication(BB)) | ||||||
| 6320 | BlockCost.first /= getReciprocalPredBlockProb(); | ||||||
| 6321 | |||||||
| 6322 | Cost.first += BlockCost.first; | ||||||
| 6323 | Cost.second |= BlockCost.second; | ||||||
| 6324 | } | ||||||
| 6325 | |||||||
| 6326 | return Cost; | ||||||
| 6327 | } | ||||||
| 6328 | |||||||
| 6329 | /// Gets Address Access SCEV after verifying that the access pattern | ||||||
| 6330 | /// is loop invariant except the induction variable dependence. | ||||||
| 6331 | /// | ||||||
| 6332 | /// This SCEV can be sent to the Target in order to estimate the address | ||||||
| 6333 | /// calculation cost. | ||||||
| 6334 | static const SCEV *getAddressAccessSCEV( | ||||||
| 6335 | Value *Ptr, | ||||||
| 6336 | LoopVectorizationLegality *Legal, | ||||||
| 6337 | PredicatedScalarEvolution &PSE, | ||||||
| 6338 | const Loop *TheLoop) { | ||||||
| 6339 | |||||||
| 6340 | auto *Gep = dyn_cast<GetElementPtrInst>(Ptr); | ||||||
| 6341 | if (!Gep) | ||||||
| 6342 | return nullptr; | ||||||
| 6343 | |||||||
| 6344 | // We are looking for a gep with all loop invariant indices except for one | ||||||
| 6345 | // which should be an induction variable. | ||||||
| 6346 | auto SE = PSE.getSE(); | ||||||
| 6347 | unsigned NumOperands = Gep->getNumOperands(); | ||||||
| 6348 | for (unsigned i = 1; i < NumOperands; ++i) { | ||||||
| 6349 | Value *Opd = Gep->getOperand(i); | ||||||
| 6350 | if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) && | ||||||
| 6351 | !Legal->isInductionVariable(Opd)) | ||||||
| 6352 | return nullptr; | ||||||
| 6353 | } | ||||||
| 6354 | |||||||
| 6355 | // Now we know we have a GEP ptr, %inv, %ind, %inv. return the Ptr SCEV. | ||||||
| 6356 | return PSE.getSCEV(Ptr); | ||||||
| 6357 | } | ||||||
| 6358 | |||||||
| 6359 | static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) { | ||||||
| 6360 | return Legal->hasStride(I->getOperand(0)) || | ||||||
| 6361 | Legal->hasStride(I->getOperand(1)); | ||||||
| 6362 | } | ||||||
| 6363 | |||||||
| 6364 | InstructionCost | ||||||
| 6365 | LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I, | ||||||
| 6366 | ElementCount VF) { | ||||||
| 6367 | assert(VF.isVector() &&(static_cast <bool> (VF.isVector() && "Scalarization cost of instruction implies vectorization." ) ? void (0) : __assert_fail ("VF.isVector() && \"Scalarization cost of instruction implies vectorization.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6368, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 6368 | "Scalarization cost of instruction implies vectorization.")(static_cast <bool> (VF.isVector() && "Scalarization cost of instruction implies vectorization." ) ? void (0) : __assert_fail ("VF.isVector() && \"Scalarization cost of instruction implies vectorization.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6368, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 6369 | if (VF.isScalable()) | ||||||
| 6370 | return InstructionCost::getInvalid(); | ||||||
| 6371 | |||||||
| 6372 | Type *ValTy = getLoadStoreType(I); | ||||||
| 6373 | auto SE = PSE.getSE(); | ||||||
| 6374 | |||||||
| 6375 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||
| 6376 | Value *Ptr = getLoadStorePointerOperand(I); | ||||||
| 6377 | Type *PtrTy = ToVectorTy(Ptr->getType(), VF); | ||||||
| 6378 | // NOTE: PtrTy is a vector to signal `TTI::getAddressComputationCost` | ||||||
| 6379 | // that it is being called from this specific place. | ||||||
| 6380 | |||||||
| 6381 | // Figure out whether the access is strided and get the stride value | ||||||
| 6382 | // if it's known in compile time | ||||||
| 6383 | const SCEV *PtrSCEV = getAddressAccessSCEV(Ptr, Legal, PSE, TheLoop); | ||||||
| 6384 | |||||||
| 6385 | // Get the cost of the scalar memory instruction and address computation. | ||||||
| 6386 | InstructionCost Cost = | ||||||
| 6387 | VF.getKnownMinValue() * TTI.getAddressComputationCost(PtrTy, SE, PtrSCEV); | ||||||
| 6388 | |||||||
| 6389 | // Don't pass *I here, since it is scalar but will actually be part of a | ||||||
| 6390 | // vectorized loop where the user of it is a vectorized instruction. | ||||||
| 6391 | const Align Alignment = getLoadStoreAlignment(I); | ||||||
| 6392 | Cost += VF.getKnownMinValue() * | ||||||
| 6393 | TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), Alignment, | ||||||
| 6394 | AS, TTI::TCK_RecipThroughput); | ||||||
| 6395 | |||||||
| 6396 | // Get the overhead of the extractelement and insertelement instructions | ||||||
| 6397 | // we might create due to scalarization. | ||||||
| 6398 | Cost += getScalarizationOverhead(I, VF); | ||||||
| 6399 | |||||||
| 6400 | // If we have a predicated load/store, it will need extra i1 extracts and | ||||||
| 6401 | // conditional branches, but may not be executed for each vector lane. Scale | ||||||
| 6402 | // the cost by the probability of executing the predicated block. | ||||||
| 6403 | if (isPredicatedInst(I)) { | ||||||
| 6404 | Cost /= getReciprocalPredBlockProb(); | ||||||
| 6405 | |||||||
| 6406 | // Add the cost of an i1 extract and a branch | ||||||
| 6407 | auto *Vec_i1Ty = | ||||||
| 6408 | VectorType::get(IntegerType::getInt1Ty(ValTy->getContext()), VF); | ||||||
| 6409 | Cost += TTI.getScalarizationOverhead( | ||||||
| 6410 | Vec_i1Ty, APInt::getAllOnes(VF.getKnownMinValue()), | ||||||
| 6411 | /*Insert=*/false, /*Extract=*/true); | ||||||
| 6412 | Cost += TTI.getCFInstrCost(Instruction::Br, TTI::TCK_RecipThroughput); | ||||||
| 6413 | |||||||
| 6414 | if (useEmulatedMaskMemRefHack(I, VF)) | ||||||
| 6415 | // Artificially setting to a high enough value to practically disable | ||||||
| 6416 | // vectorization with such operations. | ||||||
| 6417 | Cost = 3000000; | ||||||
| 6418 | } | ||||||
| 6419 | |||||||
| 6420 | return Cost; | ||||||
| 6421 | } | ||||||
| 6422 | |||||||
| 6423 | InstructionCost | ||||||
| 6424 | LoopVectorizationCostModel::getConsecutiveMemOpCost(Instruction *I, | ||||||
| 6425 | ElementCount VF) { | ||||||
| 6426 | Type *ValTy = getLoadStoreType(I); | ||||||
| 6427 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | ||||||
| 6428 | Value *Ptr = getLoadStorePointerOperand(I); | ||||||
| 6429 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||
| 6430 | int ConsecutiveStride = Legal->isConsecutivePtr(ValTy, Ptr); | ||||||
| 6431 | enum TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||
| 6432 | |||||||
| 6433 | assert((ConsecutiveStride == 1 || ConsecutiveStride == -1) &&(static_cast <bool> ((ConsecutiveStride == 1 || ConsecutiveStride == -1) && "Stride should be 1 or -1 for consecutive memory access" ) ? void (0) : __assert_fail ("(ConsecutiveStride == 1 || ConsecutiveStride == -1) && \"Stride should be 1 or -1 for consecutive memory access\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6434, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 6434 | "Stride should be 1 or -1 for consecutive memory access")(static_cast <bool> ((ConsecutiveStride == 1 || ConsecutiveStride == -1) && "Stride should be 1 or -1 for consecutive memory access" ) ? void (0) : __assert_fail ("(ConsecutiveStride == 1 || ConsecutiveStride == -1) && \"Stride should be 1 or -1 for consecutive memory access\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6434, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 6435 | const Align Alignment = getLoadStoreAlignment(I); | ||||||
| 6436 | InstructionCost Cost = 0; | ||||||
| 6437 | if (Legal->isMaskRequired(I)) { | ||||||
| 6438 | Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS, | ||||||
| 6439 | CostKind); | ||||||
| 6440 | } else { | ||||||
| 6441 | TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(I->getOperand(0)); | ||||||
| 6442 | Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS, | ||||||
| 6443 | CostKind, OpInfo, I); | ||||||
| 6444 | } | ||||||
| 6445 | |||||||
| 6446 | bool Reverse = ConsecutiveStride < 0; | ||||||
| 6447 | if (Reverse) | ||||||
| 6448 | Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, | ||||||
| 6449 | std::nullopt, CostKind, 0); | ||||||
| 6450 | return Cost; | ||||||
| 6451 | } | ||||||
| 6452 | |||||||
| 6453 | InstructionCost | ||||||
| 6454 | LoopVectorizationCostModel::getUniformMemOpCost(Instruction *I, | ||||||
| 6455 | ElementCount VF) { | ||||||
| 6456 | assert(Legal->isUniformMemOp(*I))(static_cast <bool> (Legal->isUniformMemOp(*I)) ? void (0) : __assert_fail ("Legal->isUniformMemOp(*I)", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6456, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 6457 | |||||||
| 6458 | Type *ValTy = getLoadStoreType(I); | ||||||
| 6459 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | ||||||
| 6460 | const Align Alignment = getLoadStoreAlignment(I); | ||||||
| 6461 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||
| 6462 | enum TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||
| 6463 | if (isa<LoadInst>(I)) { | ||||||
| 6464 | return TTI.getAddressComputationCost(ValTy) + | ||||||
| 6465 | TTI.getMemoryOpCost(Instruction::Load, ValTy, Alignment, AS, | ||||||
| 6466 | CostKind) + | ||||||
| 6467 | TTI.getShuffleCost(TargetTransformInfo::SK_Broadcast, VectorTy); | ||||||
| 6468 | } | ||||||
| 6469 | StoreInst *SI = cast<StoreInst>(I); | ||||||
| 6470 | |||||||
| 6471 | bool isLoopInvariantStoreValue = Legal->isUniform(SI->getValueOperand()); | ||||||
| 6472 | return TTI.getAddressComputationCost(ValTy) + | ||||||
| 6473 | TTI.getMemoryOpCost(Instruction::Store, ValTy, Alignment, AS, | ||||||
| 6474 | CostKind) + | ||||||
| 6475 | (isLoopInvariantStoreValue | ||||||
| 6476 | ? 0 | ||||||
| 6477 | : TTI.getVectorInstrCost(Instruction::ExtractElement, VectorTy, | ||||||
| 6478 | VF.getKnownMinValue() - 1)); | ||||||
| 6479 | } | ||||||
| 6480 | |||||||
| 6481 | InstructionCost | ||||||
| 6482 | LoopVectorizationCostModel::getGatherScatterCost(Instruction *I, | ||||||
| 6483 | ElementCount VF) { | ||||||
| 6484 | Type *ValTy = getLoadStoreType(I); | ||||||
| 6485 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | ||||||
| 6486 | const Align Alignment = getLoadStoreAlignment(I); | ||||||
| 6487 | const Value *Ptr = getLoadStorePointerOperand(I); | ||||||
| 6488 | |||||||
| 6489 | return TTI.getAddressComputationCost(VectorTy) + | ||||||
| 6490 | TTI.getGatherScatterOpCost( | ||||||
| 6491 | I->getOpcode(), VectorTy, Ptr, Legal->isMaskRequired(I), Alignment, | ||||||
| 6492 | TargetTransformInfo::TCK_RecipThroughput, I); | ||||||
| 6493 | } | ||||||
| 6494 | |||||||
| 6495 | InstructionCost | ||||||
| 6496 | LoopVectorizationCostModel::getInterleaveGroupCost(Instruction *I, | ||||||
| 6497 | ElementCount VF) { | ||||||
| 6498 | // TODO: Once we have support for interleaving with scalable vectors | ||||||
| 6499 | // we can calculate the cost properly here. | ||||||
| 6500 | if (VF.isScalable()) | ||||||
| 6501 | return InstructionCost::getInvalid(); | ||||||
| 6502 | |||||||
| 6503 | Type *ValTy = getLoadStoreType(I); | ||||||
| 6504 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | ||||||
| 6505 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||
| 6506 | enum TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||
| 6507 | |||||||
| 6508 | auto Group = getInterleavedAccessGroup(I); | ||||||
| 6509 | assert(Group && "Fail to get an interleaved access group.")(static_cast <bool> (Group && "Fail to get an interleaved access group." ) ? void (0) : __assert_fail ("Group && \"Fail to get an interleaved access group.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6509, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 6510 | |||||||
| 6511 | unsigned InterleaveFactor = Group->getFactor(); | ||||||
| 6512 | auto *WideVecTy = VectorType::get(ValTy, VF * InterleaveFactor); | ||||||
| 6513 | |||||||
| 6514 | // Holds the indices of existing members in the interleaved group. | ||||||
| 6515 | SmallVector<unsigned, 4> Indices; | ||||||
| 6516 | for (unsigned IF = 0; IF < InterleaveFactor; IF++) | ||||||
| 6517 | if (Group->getMember(IF)) | ||||||
| 6518 | Indices.push_back(IF); | ||||||
| 6519 | |||||||
| 6520 | // Calculate the cost of the whole interleaved group. | ||||||
| 6521 | bool UseMaskForGaps = | ||||||
| 6522 | (Group->requiresScalarEpilogue() && !isScalarEpilogueAllowed()) || | ||||||
| 6523 | (isa<StoreInst>(I) && (Group->getNumMembers() < Group->getFactor())); | ||||||
| 6524 | InstructionCost Cost = TTI.getInterleavedMemoryOpCost( | ||||||
| 6525 | I->getOpcode(), WideVecTy, Group->getFactor(), Indices, Group->getAlign(), | ||||||
| 6526 | AS, CostKind, Legal->isMaskRequired(I), UseMaskForGaps); | ||||||
| 6527 | |||||||
| 6528 | if (Group->isReverse()) { | ||||||
| 6529 | // TODO: Add support for reversed masked interleaved access. | ||||||
| 6530 | assert(!Legal->isMaskRequired(I) &&(static_cast <bool> (!Legal->isMaskRequired(I) && "Reverse masked interleaved access not supported.") ? void ( 0) : __assert_fail ("!Legal->isMaskRequired(I) && \"Reverse masked interleaved access not supported.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6531, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 6531 | "Reverse masked interleaved access not supported.")(static_cast <bool> (!Legal->isMaskRequired(I) && "Reverse masked interleaved access not supported.") ? void ( 0) : __assert_fail ("!Legal->isMaskRequired(I) && \"Reverse masked interleaved access not supported.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6531, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 6532 | Cost += Group->getNumMembers() * | ||||||
| 6533 | TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, | ||||||
| 6534 | std::nullopt, CostKind, 0); | ||||||
| 6535 | } | ||||||
| 6536 | return Cost; | ||||||
| 6537 | } | ||||||
| 6538 | |||||||
| 6539 | std::optional<InstructionCost> | ||||||
| 6540 | LoopVectorizationCostModel::getReductionPatternCost( | ||||||
| 6541 | Instruction *I, ElementCount VF, Type *Ty, TTI::TargetCostKind CostKind) { | ||||||
| 6542 | using namespace llvm::PatternMatch; | ||||||
| 6543 | // Early exit for no inloop reductions | ||||||
| 6544 | if (InLoopReductionChains.empty() || VF.isScalar() || !isa<VectorType>(Ty)) | ||||||
| 6545 | return std::nullopt; | ||||||
| 6546 | auto *VectorTy = cast<VectorType>(Ty); | ||||||
| 6547 | |||||||
| 6548 | // We are looking for a pattern of, and finding the minimal acceptable cost: | ||||||
| 6549 | // reduce(mul(ext(A), ext(B))) or | ||||||
| 6550 | // reduce(mul(A, B)) or | ||||||
| 6551 | // reduce(ext(A)) or | ||||||
| 6552 | // reduce(A). | ||||||
| 6553 | // The basic idea is that we walk down the tree to do that, finding the root | ||||||
| 6554 | // reduction instruction in InLoopReductionImmediateChains. From there we find | ||||||
| 6555 | // the pattern of mul/ext and test the cost of the entire pattern vs the cost | ||||||
| 6556 | // of the components. If the reduction cost is lower then we return it for the | ||||||
| 6557 | // reduction instruction and 0 for the other instructions in the pattern. If | ||||||
| 6558 | // it is not we return an invalid cost specifying the orignal cost method | ||||||
| 6559 | // should be used. | ||||||
| 6560 | Instruction *RetI = I; | ||||||
| 6561 | if (match(RetI, m_ZExtOrSExt(m_Value()))) { | ||||||
| 6562 | if (!RetI->hasOneUser()) | ||||||
| 6563 | return std::nullopt; | ||||||
| 6564 | RetI = RetI->user_back(); | ||||||
| 6565 | } | ||||||
| 6566 | |||||||
| 6567 | if (match(RetI, m_OneUse(m_Mul(m_Value(), m_Value()))) && | ||||||
| 6568 | RetI->user_back()->getOpcode() == Instruction::Add) { | ||||||
| 6569 | RetI = RetI->user_back(); | ||||||
| 6570 | } | ||||||
| 6571 | |||||||
| 6572 | // Test if the found instruction is a reduction, and if not return an invalid | ||||||
| 6573 | // cost specifying the parent to use the original cost modelling. | ||||||
| 6574 | if (!InLoopReductionImmediateChains.count(RetI)) | ||||||
| 6575 | return std::nullopt; | ||||||
| 6576 | |||||||
| 6577 | // Find the reduction this chain is a part of and calculate the basic cost of | ||||||
| 6578 | // the reduction on its own. | ||||||
| 6579 | Instruction *LastChain = InLoopReductionImmediateChains[RetI]; | ||||||
| 6580 | Instruction *ReductionPhi = LastChain; | ||||||
| 6581 | while (!isa<PHINode>(ReductionPhi)) | ||||||
| 6582 | ReductionPhi = InLoopReductionImmediateChains[ReductionPhi]; | ||||||
| 6583 | |||||||
| 6584 | const RecurrenceDescriptor &RdxDesc = | ||||||
| 6585 | Legal->getReductionVars().find(cast<PHINode>(ReductionPhi))->second; | ||||||
| 6586 | |||||||
| 6587 | InstructionCost BaseCost = TTI.getArithmeticReductionCost( | ||||||
| 6588 | RdxDesc.getOpcode(), VectorTy, RdxDesc.getFastMathFlags(), CostKind); | ||||||
| 6589 | |||||||
| 6590 | // For a call to the llvm.fmuladd intrinsic we need to add the cost of a | ||||||
| 6591 | // normal fmul instruction to the cost of the fadd reduction. | ||||||
| 6592 | if (RdxDesc.getRecurrenceKind() == RecurKind::FMulAdd) | ||||||
| 6593 | BaseCost += | ||||||
| 6594 | TTI.getArithmeticInstrCost(Instruction::FMul, VectorTy, CostKind); | ||||||
| 6595 | |||||||
| 6596 | // If we're using ordered reductions then we can just return the base cost | ||||||
| 6597 | // here, since getArithmeticReductionCost calculates the full ordered | ||||||
| 6598 | // reduction cost when FP reassociation is not allowed. | ||||||
| 6599 | if (useOrderedReductions(RdxDesc)) | ||||||
| 6600 | return BaseCost; | ||||||
| 6601 | |||||||
| 6602 | // Get the operand that was not the reduction chain and match it to one of the | ||||||
| 6603 | // patterns, returning the better cost if it is found. | ||||||
| 6604 | Instruction *RedOp = RetI->getOperand(1) == LastChain | ||||||
| 6605 | ? dyn_cast<Instruction>(RetI->getOperand(0)) | ||||||
| 6606 | : dyn_cast<Instruction>(RetI->getOperand(1)); | ||||||
| 6607 | |||||||
| 6608 | VectorTy = VectorType::get(I->getOperand(0)->getType(), VectorTy); | ||||||
| 6609 | |||||||
| 6610 | Instruction *Op0, *Op1; | ||||||
| 6611 | if (RedOp && RdxDesc.getOpcode() == Instruction::Add && | ||||||
| 6612 | match(RedOp, | ||||||
| 6613 | m_ZExtOrSExt(m_Mul(m_Instruction(Op0), m_Instruction(Op1)))) && | ||||||
| 6614 | match(Op0, m_ZExtOrSExt(m_Value())) && | ||||||
| 6615 | Op0->getOpcode() == Op1->getOpcode() && | ||||||
| 6616 | Op0->getOperand(0)->getType() == Op1->getOperand(0)->getType() && | ||||||
| 6617 | !TheLoop->isLoopInvariant(Op0) && !TheLoop->isLoopInvariant(Op1) && | ||||||
| 6618 | (Op0->getOpcode() == RedOp->getOpcode() || Op0 == Op1)) { | ||||||
| 6619 | |||||||
| 6620 | // Matched reduce.add(ext(mul(ext(A), ext(B))) | ||||||
| 6621 | // Note that the extend opcodes need to all match, or if A==B they will have | ||||||
| 6622 | // been converted to zext(mul(sext(A), sext(A))) as it is known positive, | ||||||
| 6623 | // which is equally fine. | ||||||
| 6624 | bool IsUnsigned = isa<ZExtInst>(Op0); | ||||||
| 6625 | auto *ExtType = VectorType::get(Op0->getOperand(0)->getType(), VectorTy); | ||||||
| 6626 | auto *MulType = VectorType::get(Op0->getType(), VectorTy); | ||||||
| 6627 | |||||||
| 6628 | InstructionCost ExtCost = | ||||||
| 6629 | TTI.getCastInstrCost(Op0->getOpcode(), MulType, ExtType, | ||||||
| 6630 | TTI::CastContextHint::None, CostKind, Op0); | ||||||
| 6631 | InstructionCost MulCost = | ||||||
| 6632 | TTI.getArithmeticInstrCost(Instruction::Mul, MulType, CostKind); | ||||||
| 6633 | InstructionCost Ext2Cost = | ||||||
| 6634 | TTI.getCastInstrCost(RedOp->getOpcode(), VectorTy, MulType, | ||||||
| 6635 | TTI::CastContextHint::None, CostKind, RedOp); | ||||||
| 6636 | |||||||
| 6637 | InstructionCost RedCost = TTI.getMulAccReductionCost( | ||||||
| 6638 | IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, CostKind); | ||||||
| 6639 | |||||||
| 6640 | if (RedCost.isValid() && | ||||||
| 6641 | RedCost < ExtCost * 2 + MulCost + Ext2Cost + BaseCost) | ||||||
| 6642 | return I == RetI ? RedCost : 0; | ||||||
| 6643 | } else if (RedOp && match(RedOp, m_ZExtOrSExt(m_Value())) && | ||||||
| 6644 | !TheLoop->isLoopInvariant(RedOp)) { | ||||||
| 6645 | // Matched reduce(ext(A)) | ||||||
| 6646 | bool IsUnsigned = isa<ZExtInst>(RedOp); | ||||||
| 6647 | auto *ExtType = VectorType::get(RedOp->getOperand(0)->getType(), VectorTy); | ||||||
| 6648 | InstructionCost RedCost = TTI.getExtendedReductionCost( | ||||||
| 6649 | RdxDesc.getOpcode(), IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, | ||||||
| 6650 | RdxDesc.getFastMathFlags(), CostKind); | ||||||
| 6651 | |||||||
| 6652 | InstructionCost ExtCost = | ||||||
| 6653 | TTI.getCastInstrCost(RedOp->getOpcode(), VectorTy, ExtType, | ||||||
| 6654 | TTI::CastContextHint::None, CostKind, RedOp); | ||||||
| 6655 | if (RedCost.isValid() && RedCost < BaseCost + ExtCost) | ||||||
| 6656 | return I == RetI ? RedCost : 0; | ||||||
| 6657 | } else if (RedOp && RdxDesc.getOpcode() == Instruction::Add && | ||||||
| 6658 | match(RedOp, m_Mul(m_Instruction(Op0), m_Instruction(Op1)))) { | ||||||
| 6659 | if (match(Op0, m_ZExtOrSExt(m_Value())) && | ||||||
| 6660 | Op0->getOpcode() == Op1->getOpcode() && | ||||||
| 6661 | !TheLoop->isLoopInvariant(Op0) && !TheLoop->isLoopInvariant(Op1)) { | ||||||
| 6662 | bool IsUnsigned = isa<ZExtInst>(Op0); | ||||||
| 6663 | Type *Op0Ty = Op0->getOperand(0)->getType(); | ||||||
| 6664 | Type *Op1Ty = Op1->getOperand(0)->getType(); | ||||||
| 6665 | Type *LargestOpTy = | ||||||
| 6666 | Op0Ty->getIntegerBitWidth() < Op1Ty->getIntegerBitWidth() ? Op1Ty | ||||||
| 6667 | : Op0Ty; | ||||||
| 6668 | auto *ExtType = VectorType::get(LargestOpTy, VectorTy); | ||||||
| 6669 | |||||||
| 6670 | // Matched reduce.add(mul(ext(A), ext(B))), where the two ext may be of | ||||||
| 6671 | // different sizes. We take the largest type as the ext to reduce, and add | ||||||
| 6672 | // the remaining cost as, for example reduce(mul(ext(ext(A)), ext(B))). | ||||||
| 6673 | InstructionCost ExtCost0 = TTI.getCastInstrCost( | ||||||
| 6674 | Op0->getOpcode(), VectorTy, VectorType::get(Op0Ty, VectorTy), | ||||||
| 6675 | TTI::CastContextHint::None, CostKind, Op0); | ||||||
| 6676 | InstructionCost ExtCost1 = TTI.getCastInstrCost( | ||||||
| 6677 | Op1->getOpcode(), VectorTy, VectorType::get(Op1Ty, VectorTy), | ||||||
| 6678 | TTI::CastContextHint::None, CostKind, Op1); | ||||||
| 6679 | InstructionCost MulCost = | ||||||
| 6680 | TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); | ||||||
| 6681 | |||||||
| 6682 | InstructionCost RedCost = TTI.getMulAccReductionCost( | ||||||
| 6683 | IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, CostKind); | ||||||
| 6684 | InstructionCost ExtraExtCost = 0; | ||||||
| 6685 | if (Op0Ty != LargestOpTy || Op1Ty != LargestOpTy) { | ||||||
| 6686 | Instruction *ExtraExtOp = (Op0Ty != LargestOpTy) ? Op0 : Op1; | ||||||
| 6687 | ExtraExtCost = TTI.getCastInstrCost( | ||||||
| 6688 | ExtraExtOp->getOpcode(), ExtType, | ||||||
| 6689 | VectorType::get(ExtraExtOp->getOperand(0)->getType(), VectorTy), | ||||||
| 6690 | TTI::CastContextHint::None, CostKind, ExtraExtOp); | ||||||
| 6691 | } | ||||||
| 6692 | |||||||
| 6693 | if (RedCost.isValid() && | ||||||
| 6694 | (RedCost + ExtraExtCost) < (ExtCost0 + ExtCost1 + MulCost + BaseCost)) | ||||||
| 6695 | return I == RetI ? RedCost : 0; | ||||||
| 6696 | } else if (!match(I, m_ZExtOrSExt(m_Value()))) { | ||||||
| 6697 | // Matched reduce.add(mul()) | ||||||
| 6698 | InstructionCost MulCost = | ||||||
| 6699 | TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); | ||||||
| 6700 | |||||||
| 6701 | InstructionCost RedCost = TTI.getMulAccReductionCost( | ||||||
| 6702 | true, RdxDesc.getRecurrenceType(), VectorTy, CostKind); | ||||||
| 6703 | |||||||
| 6704 | if (RedCost.isValid() && RedCost < MulCost + BaseCost) | ||||||
| 6705 | return I == RetI ? RedCost : 0; | ||||||
| 6706 | } | ||||||
| 6707 | } | ||||||
| 6708 | |||||||
| 6709 | return I == RetI ? std::optional<InstructionCost>(BaseCost) : std::nullopt; | ||||||
| 6710 | } | ||||||
| 6711 | |||||||
| 6712 | InstructionCost | ||||||
| 6713 | LoopVectorizationCostModel::getMemoryInstructionCost(Instruction *I, | ||||||
| 6714 | ElementCount VF) { | ||||||
| 6715 | // Calculate scalar cost only. Vectorization cost should be ready at this | ||||||
| 6716 | // moment. | ||||||
| 6717 | if (VF.isScalar()) { | ||||||
| 6718 | Type *ValTy = getLoadStoreType(I); | ||||||
| 6719 | const Align Alignment = getLoadStoreAlignment(I); | ||||||
| 6720 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||
| 6721 | |||||||
| 6722 | TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(I->getOperand(0)); | ||||||
| 6723 | return TTI.getAddressComputationCost(ValTy) + | ||||||
| 6724 | TTI.getMemoryOpCost(I->getOpcode(), ValTy, Alignment, AS, | ||||||
| 6725 | TTI::TCK_RecipThroughput, OpInfo, I); | ||||||
| 6726 | } | ||||||
| 6727 | return getWideningCost(I, VF); | ||||||
| 6728 | } | ||||||
| 6729 | |||||||
| 6730 | LoopVectorizationCostModel::VectorizationCostTy | ||||||
| 6731 | LoopVectorizationCostModel::getInstructionCost(Instruction *I, | ||||||
| 6732 | ElementCount VF) { | ||||||
| 6733 | // If we know that this instruction will remain uniform, check the cost of | ||||||
| 6734 | // the scalar version. | ||||||
| 6735 | if (isUniformAfterVectorization(I, VF)) | ||||||
| 6736 | VF = ElementCount::getFixed(1); | ||||||
| 6737 | |||||||
| 6738 | if (VF.isVector() && isProfitableToScalarize(I, VF)) | ||||||
| 6739 | return VectorizationCostTy(InstsToScalarize[VF][I], false); | ||||||
| 6740 | |||||||
| 6741 | // Forced scalars do not have any scalarization overhead. | ||||||
| 6742 | auto ForcedScalar = ForcedScalars.find(VF); | ||||||
| 6743 | if (VF.isVector() && ForcedScalar != ForcedScalars.end()) { | ||||||
| 6744 | auto InstSet = ForcedScalar->second; | ||||||
| 6745 | if (InstSet.count(I)) | ||||||
| 6746 | return VectorizationCostTy( | ||||||
| 6747 | (getInstructionCost(I, ElementCount::getFixed(1)).first * | ||||||
| 6748 | VF.getKnownMinValue()), | ||||||
| 6749 | false); | ||||||
| 6750 | } | ||||||
| 6751 | |||||||
| 6752 | Type *VectorTy; | ||||||
| 6753 | InstructionCost C = getInstructionCost(I, VF, VectorTy); | ||||||
| 6754 | |||||||
| 6755 | bool TypeNotScalarized = false; | ||||||
| 6756 | if (VF.isVector() && VectorTy->isVectorTy()) { | ||||||
| 6757 | if (unsigned NumParts = TTI.getNumberOfParts(VectorTy)) { | ||||||
| 6758 | if (VF.isScalable()) | ||||||
| 6759 | // <vscale x 1 x iN> is assumed to be profitable over iN because | ||||||
| 6760 | // scalable registers are a distinct register class from scalar ones. | ||||||
| 6761 | // If we ever find a target which wants to lower scalable vectors | ||||||
| 6762 | // back to scalars, we'll need to update this code to explicitly | ||||||
| 6763 | // ask TTI about the register class uses for each part. | ||||||
| 6764 | TypeNotScalarized = NumParts <= VF.getKnownMinValue(); | ||||||
| 6765 | else | ||||||
| 6766 | TypeNotScalarized = NumParts < VF.getKnownMinValue(); | ||||||
| 6767 | } else | ||||||
| 6768 | C = InstructionCost::getInvalid(); | ||||||
| 6769 | } | ||||||
| 6770 | return VectorizationCostTy(C, TypeNotScalarized); | ||||||
| 6771 | } | ||||||
| 6772 | |||||||
| 6773 | InstructionCost | ||||||
| 6774 | LoopVectorizationCostModel::getScalarizationOverhead(Instruction *I, | ||||||
| 6775 | ElementCount VF) const { | ||||||
| 6776 | |||||||
| 6777 | // There is no mechanism yet to create a scalable scalarization loop, | ||||||
| 6778 | // so this is currently Invalid. | ||||||
| 6779 | if (VF.isScalable()) | ||||||
| 6780 | return InstructionCost::getInvalid(); | ||||||
| 6781 | |||||||
| 6782 | if (VF.isScalar()) | ||||||
| 6783 | return 0; | ||||||
| 6784 | |||||||
| 6785 | InstructionCost Cost = 0; | ||||||
| 6786 | Type *RetTy = ToVectorTy(I->getType(), VF); | ||||||
| 6787 | if (!RetTy->isVoidTy() && | ||||||
| 6788 | (!isa<LoadInst>(I) || !TTI.supportsEfficientVectorElementLoadStore())) | ||||||
| 6789 | Cost += TTI.getScalarizationOverhead( | ||||||
| 6790 | cast<VectorType>(RetTy), APInt::getAllOnes(VF.getKnownMinValue()), true, | ||||||
| 6791 | false); | ||||||
| 6792 | |||||||
| 6793 | // Some targets keep addresses scalar. | ||||||
| 6794 | if (isa<LoadInst>(I) && !TTI.prefersVectorizedAddressing()) | ||||||
| 6795 | return Cost; | ||||||
| 6796 | |||||||
| 6797 | // Some targets support efficient element stores. | ||||||
| 6798 | if (isa<StoreInst>(I) && TTI.supportsEfficientVectorElementLoadStore()) | ||||||
| 6799 | return Cost; | ||||||
| 6800 | |||||||
| 6801 | // Collect operands to consider. | ||||||
| 6802 | CallInst *CI = dyn_cast<CallInst>(I); | ||||||
| 6803 | Instruction::op_range Ops = CI ? CI->args() : I->operands(); | ||||||
| 6804 | |||||||
| 6805 | // Skip operands that do not require extraction/scalarization and do not incur | ||||||
| 6806 | // any overhead. | ||||||
| 6807 | SmallVector<Type *> Tys; | ||||||
| 6808 | for (auto *V : filterExtractingOperands(Ops, VF)) | ||||||
| 6809 | Tys.push_back(MaybeVectorizeType(V->getType(), VF)); | ||||||
| 6810 | return Cost + TTI.getOperandsScalarizationOverhead( | ||||||
| 6811 | filterExtractingOperands(Ops, VF), Tys); | ||||||
| 6812 | } | ||||||
| 6813 | |||||||
| 6814 | void LoopVectorizationCostModel::setCostBasedWideningDecision(ElementCount VF) { | ||||||
| 6815 | if (VF.isScalar()) | ||||||
| 6816 | return; | ||||||
| 6817 | NumPredStores = 0; | ||||||
| 6818 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||||
| 6819 | // For each instruction in the old loop. | ||||||
| 6820 | for (Instruction &I : *BB) { | ||||||
| 6821 | Value *Ptr = getLoadStorePointerOperand(&I); | ||||||
| 6822 | if (!Ptr) | ||||||
| 6823 | continue; | ||||||
| 6824 | |||||||
| 6825 | // TODO: We should generate better code and update the cost model for | ||||||
| 6826 | // predicated uniform stores. Today they are treated as any other | ||||||
| 6827 | // predicated store (see added test cases in | ||||||
| 6828 | // invariant-store-vectorization.ll). | ||||||
| 6829 | if (isa<StoreInst>(&I) && isScalarWithPredication(&I, VF)) | ||||||
| 6830 | NumPredStores++; | ||||||
| 6831 | |||||||
| 6832 | if (Legal->isUniformMemOp(I)) { | ||||||
| 6833 | auto isLegalToScalarize = [&]() { | ||||||
| 6834 | if (!VF.isScalable()) | ||||||
| 6835 | // Scalarization of fixed length vectors "just works". | ||||||
| 6836 | return true; | ||||||
| 6837 | |||||||
| 6838 | // We have dedicated lowering for unpredicated uniform loads and | ||||||
| 6839 | // stores. Note that even with tail folding we know that at least | ||||||
| 6840 | // one lane is active (i.e. generalized predication is not possible | ||||||
| 6841 | // here), and the logic below depends on this fact. | ||||||
| 6842 | if (!foldTailByMasking()) | ||||||
| 6843 | return true; | ||||||
| 6844 | |||||||
| 6845 | // For scalable vectors, a uniform memop load is always | ||||||
| 6846 | // uniform-by-parts and we know how to scalarize that. | ||||||
| 6847 | if (isa<LoadInst>(I)) | ||||||
| 6848 | return true; | ||||||
| 6849 | |||||||
| 6850 | // A uniform store isn't neccessarily uniform-by-part | ||||||
| 6851 | // and we can't assume scalarization. | ||||||
| 6852 | auto &SI = cast<StoreInst>(I); | ||||||
| 6853 | return TheLoop->isLoopInvariant(SI.getValueOperand()); | ||||||
| 6854 | }; | ||||||
| 6855 | |||||||
| 6856 | const InstructionCost GatherScatterCost = | ||||||
| 6857 | isLegalGatherOrScatter(&I, VF) ? | ||||||
| 6858 | getGatherScatterCost(&I, VF) : InstructionCost::getInvalid(); | ||||||
| 6859 | |||||||
| 6860 | // Load: Scalar load + broadcast | ||||||
| 6861 | // Store: Scalar store + isLoopInvariantStoreValue ? 0 : extract | ||||||
| 6862 | // FIXME: This cost is a significant under-estimate for tail folded | ||||||
| 6863 | // memory ops. | ||||||
| 6864 | const InstructionCost ScalarizationCost = isLegalToScalarize() ? | ||||||
| 6865 | getUniformMemOpCost(&I, VF) : InstructionCost::getInvalid(); | ||||||
| 6866 | |||||||
| 6867 | // Choose better solution for the current VF, Note that Invalid | ||||||
| 6868 | // costs compare as maximumal large. If both are invalid, we get | ||||||
| 6869 | // scalable invalid which signals a failure and a vectorization abort. | ||||||
| 6870 | if (GatherScatterCost < ScalarizationCost) | ||||||
| 6871 | setWideningDecision(&I, VF, CM_GatherScatter, GatherScatterCost); | ||||||
| 6872 | else | ||||||
| 6873 | setWideningDecision(&I, VF, CM_Scalarize, ScalarizationCost); | ||||||
| 6874 | continue; | ||||||
| 6875 | } | ||||||
| 6876 | |||||||
| 6877 | // We assume that widening is the best solution when possible. | ||||||
| 6878 | if (memoryInstructionCanBeWidened(&I, VF)) { | ||||||
| 6879 | InstructionCost Cost = getConsecutiveMemOpCost(&I, VF); | ||||||
| 6880 | int ConsecutiveStride = Legal->isConsecutivePtr( | ||||||
| 6881 | getLoadStoreType(&I), getLoadStorePointerOperand(&I)); | ||||||
| 6882 | assert((ConsecutiveStride == 1 || ConsecutiveStride == -1) &&(static_cast <bool> ((ConsecutiveStride == 1 || ConsecutiveStride == -1) && "Expected consecutive stride.") ? void (0) : __assert_fail ("(ConsecutiveStride == 1 || ConsecutiveStride == -1) && \"Expected consecutive stride.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6883, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 6883 | "Expected consecutive stride.")(static_cast <bool> ((ConsecutiveStride == 1 || ConsecutiveStride == -1) && "Expected consecutive stride.") ? void (0) : __assert_fail ("(ConsecutiveStride == 1 || ConsecutiveStride == -1) && \"Expected consecutive stride.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6883, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 6884 | InstWidening Decision = | ||||||
| 6885 | ConsecutiveStride == 1 ? CM_Widen : CM_Widen_Reverse; | ||||||
| 6886 | setWideningDecision(&I, VF, Decision, Cost); | ||||||
| 6887 | continue; | ||||||
| 6888 | } | ||||||
| 6889 | |||||||
| 6890 | // Choose between Interleaving, Gather/Scatter or Scalarization. | ||||||
| 6891 | InstructionCost InterleaveCost = InstructionCost::getInvalid(); | ||||||
| 6892 | unsigned NumAccesses = 1; | ||||||
| 6893 | if (isAccessInterleaved(&I)) { | ||||||
| 6894 | auto Group = getInterleavedAccessGroup(&I); | ||||||
| 6895 | assert(Group && "Fail to get an interleaved access group.")(static_cast <bool> (Group && "Fail to get an interleaved access group." ) ? void (0) : __assert_fail ("Group && \"Fail to get an interleaved access group.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6895, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 6896 | |||||||
| 6897 | // Make one decision for the whole group. | ||||||
| 6898 | if (getWideningDecision(&I, VF) != CM_Unknown) | ||||||
| 6899 | continue; | ||||||
| 6900 | |||||||
| 6901 | NumAccesses = Group->getNumMembers(); | ||||||
| 6902 | if (interleavedAccessCanBeWidened(&I, VF)) | ||||||
| 6903 | InterleaveCost = getInterleaveGroupCost(&I, VF); | ||||||
| 6904 | } | ||||||
| 6905 | |||||||
| 6906 | InstructionCost GatherScatterCost = | ||||||
| 6907 | isLegalGatherOrScatter(&I, VF) | ||||||
| 6908 | ? getGatherScatterCost(&I, VF) * NumAccesses | ||||||
| 6909 | : InstructionCost::getInvalid(); | ||||||
| 6910 | |||||||
| 6911 | InstructionCost ScalarizationCost = | ||||||
| 6912 | getMemInstScalarizationCost(&I, VF) * NumAccesses; | ||||||
| 6913 | |||||||
| 6914 | // Choose better solution for the current VF, | ||||||
| 6915 | // write down this decision and use it during vectorization. | ||||||
| 6916 | InstructionCost Cost; | ||||||
| 6917 | InstWidening Decision; | ||||||
| 6918 | if (InterleaveCost <= GatherScatterCost && | ||||||
| 6919 | InterleaveCost < ScalarizationCost) { | ||||||
| 6920 | Decision = CM_Interleave; | ||||||
| 6921 | Cost = InterleaveCost; | ||||||
| 6922 | } else if (GatherScatterCost < ScalarizationCost) { | ||||||
| 6923 | Decision = CM_GatherScatter; | ||||||
| 6924 | Cost = GatherScatterCost; | ||||||
| 6925 | } else { | ||||||
| 6926 | Decision = CM_Scalarize; | ||||||
| 6927 | Cost = ScalarizationCost; | ||||||
| 6928 | } | ||||||
| 6929 | // If the instructions belongs to an interleave group, the whole group | ||||||
| 6930 | // receives the same decision. The whole group receives the cost, but | ||||||
| 6931 | // the cost will actually be assigned to one instruction. | ||||||
| 6932 | if (auto Group = getInterleavedAccessGroup(&I)) | ||||||
| 6933 | setWideningDecision(Group, VF, Decision, Cost); | ||||||
| 6934 | else | ||||||
| 6935 | setWideningDecision(&I, VF, Decision, Cost); | ||||||
| 6936 | } | ||||||
| 6937 | } | ||||||
| 6938 | |||||||
| 6939 | // Make sure that any load of address and any other address computation | ||||||
| 6940 | // remains scalar unless there is gather/scatter support. This avoids | ||||||
| 6941 | // inevitable extracts into address registers, and also has the benefit of | ||||||
| 6942 | // activating LSR more, since that pass can't optimize vectorized | ||||||
| 6943 | // addresses. | ||||||
| 6944 | if (TTI.prefersVectorizedAddressing()) | ||||||
| 6945 | return; | ||||||
| 6946 | |||||||
| 6947 | // Start with all scalar pointer uses. | ||||||
| 6948 | SmallPtrSet<Instruction *, 8> AddrDefs; | ||||||
| 6949 | for (BasicBlock *BB : TheLoop->blocks()) | ||||||
| 6950 | for (Instruction &I : *BB) { | ||||||
| 6951 | Instruction *PtrDef = | ||||||
| 6952 | dyn_cast_or_null<Instruction>(getLoadStorePointerOperand(&I)); | ||||||
| 6953 | if (PtrDef && TheLoop->contains(PtrDef) && | ||||||
| 6954 | getWideningDecision(&I, VF) != CM_GatherScatter) | ||||||
| 6955 | AddrDefs.insert(PtrDef); | ||||||
| 6956 | } | ||||||
| 6957 | |||||||
| 6958 | // Add all instructions used to generate the addresses. | ||||||
| 6959 | SmallVector<Instruction *, 4> Worklist; | ||||||
| 6960 | append_range(Worklist, AddrDefs); | ||||||
| 6961 | while (!Worklist.empty()) { | ||||||
| 6962 | Instruction *I = Worklist.pop_back_val(); | ||||||
| 6963 | for (auto &Op : I->operands()) | ||||||
| 6964 | if (auto *InstOp = dyn_cast<Instruction>(Op)) | ||||||
| 6965 | if ((InstOp->getParent() == I->getParent()) && !isa<PHINode>(InstOp) && | ||||||
| 6966 | AddrDefs.insert(InstOp).second) | ||||||
| 6967 | Worklist.push_back(InstOp); | ||||||
| 6968 | } | ||||||
| 6969 | |||||||
| 6970 | for (auto *I : AddrDefs) { | ||||||
| 6971 | if (isa<LoadInst>(I)) { | ||||||
| 6972 | // Setting the desired widening decision should ideally be handled in | ||||||
| 6973 | // by cost functions, but since this involves the task of finding out | ||||||
| 6974 | // if the loaded register is involved in an address computation, it is | ||||||
| 6975 | // instead changed here when we know this is the case. | ||||||
| 6976 | InstWidening Decision = getWideningDecision(I, VF); | ||||||
| 6977 | if (Decision == CM_Widen || Decision == CM_Widen_Reverse) | ||||||
| 6978 | // Scalarize a widened load of address. | ||||||
| 6979 | setWideningDecision( | ||||||
| 6980 | I, VF, CM_Scalarize, | ||||||
| 6981 | (VF.getKnownMinValue() * | ||||||
| 6982 | getMemoryInstructionCost(I, ElementCount::getFixed(1)))); | ||||||
| 6983 | else if (auto Group = getInterleavedAccessGroup(I)) { | ||||||
| 6984 | // Scalarize an interleave group of address loads. | ||||||
| 6985 | for (unsigned I = 0; I < Group->getFactor(); ++I) { | ||||||
| 6986 | if (Instruction *Member = Group->getMember(I)) | ||||||
| 6987 | setWideningDecision( | ||||||
| 6988 | Member, VF, CM_Scalarize, | ||||||
| 6989 | (VF.getKnownMinValue() * | ||||||
| 6990 | getMemoryInstructionCost(Member, ElementCount::getFixed(1)))); | ||||||
| 6991 | } | ||||||
| 6992 | } | ||||||
| 6993 | } else | ||||||
| 6994 | // Make sure I gets scalarized and a cost estimate without | ||||||
| 6995 | // scalarization overhead. | ||||||
| 6996 | ForcedScalars[VF].insert(I); | ||||||
| 6997 | } | ||||||
| 6998 | } | ||||||
| 6999 | |||||||
| 7000 | InstructionCost | ||||||
| 7001 | LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, | ||||||
| 7002 | Type *&VectorTy) { | ||||||
| 7003 | Type *RetTy = I->getType(); | ||||||
| 7004 | if (canTruncateToMinimalBitwidth(I, VF)) | ||||||
| 7005 | RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]); | ||||||
| 7006 | auto SE = PSE.getSE(); | ||||||
| 7007 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||
| 7008 | |||||||
| 7009 | auto hasSingleCopyAfterVectorization = [this](Instruction *I, | ||||||
| 7010 | ElementCount VF) -> bool { | ||||||
| 7011 | if (VF.isScalar()) | ||||||
| 7012 | return true; | ||||||
| 7013 | |||||||
| 7014 | auto Scalarized = InstsToScalarize.find(VF); | ||||||
| 7015 | assert(Scalarized != InstsToScalarize.end() &&(static_cast <bool> (Scalarized != InstsToScalarize.end () && "VF not yet analyzed for scalarization profitability" ) ? void (0) : __assert_fail ("Scalarized != InstsToScalarize.end() && \"VF not yet analyzed for scalarization profitability\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7016, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7016 | "VF not yet analyzed for scalarization profitability")(static_cast <bool> (Scalarized != InstsToScalarize.end () && "VF not yet analyzed for scalarization profitability" ) ? void (0) : __assert_fail ("Scalarized != InstsToScalarize.end() && \"VF not yet analyzed for scalarization profitability\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7016, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7017 | return !Scalarized->second.count(I) && | ||||||
| 7018 | llvm::all_of(I->users(), [&](User *U) { | ||||||
| 7019 | auto *UI = cast<Instruction>(U); | ||||||
| 7020 | return !Scalarized->second.count(UI); | ||||||
| 7021 | }); | ||||||
| 7022 | }; | ||||||
| 7023 | (void) hasSingleCopyAfterVectorization; | ||||||
| 7024 | |||||||
| 7025 | if (isScalarAfterVectorization(I, VF)) { | ||||||
| 7026 | // With the exception of GEPs and PHIs, after scalarization there should | ||||||
| 7027 | // only be one copy of the instruction generated in the loop. This is | ||||||
| 7028 | // because the VF is either 1, or any instructions that need scalarizing | ||||||
| 7029 | // have already been dealt with by the the time we get here. As a result, | ||||||
| 7030 | // it means we don't have to multiply the instruction cost by VF. | ||||||
| 7031 | assert(I->getOpcode() == Instruction::GetElementPtr ||(static_cast <bool> (I->getOpcode() == Instruction:: GetElementPtr || I->getOpcode() == Instruction::PHI || (I-> getOpcode() == Instruction::BitCast && I->getType( )->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF )) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::GetElementPtr || I->getOpcode() == Instruction::PHI || (I->getOpcode() == Instruction::BitCast && I->getType()->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF)" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7035, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7032 | I->getOpcode() == Instruction::PHI ||(static_cast <bool> (I->getOpcode() == Instruction:: GetElementPtr || I->getOpcode() == Instruction::PHI || (I-> getOpcode() == Instruction::BitCast && I->getType( )->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF )) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::GetElementPtr || I->getOpcode() == Instruction::PHI || (I->getOpcode() == Instruction::BitCast && I->getType()->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF)" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7035, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7033 | (I->getOpcode() == Instruction::BitCast &&(static_cast <bool> (I->getOpcode() == Instruction:: GetElementPtr || I->getOpcode() == Instruction::PHI || (I-> getOpcode() == Instruction::BitCast && I->getType( )->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF )) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::GetElementPtr || I->getOpcode() == Instruction::PHI || (I->getOpcode() == Instruction::BitCast && I->getType()->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF)" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7035, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7034 | I->getType()->isPointerTy()) ||(static_cast <bool> (I->getOpcode() == Instruction:: GetElementPtr || I->getOpcode() == Instruction::PHI || (I-> getOpcode() == Instruction::BitCast && I->getType( )->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF )) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::GetElementPtr || I->getOpcode() == Instruction::PHI || (I->getOpcode() == Instruction::BitCast && I->getType()->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF)" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7035, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7035 | hasSingleCopyAfterVectorization(I, VF))(static_cast <bool> (I->getOpcode() == Instruction:: GetElementPtr || I->getOpcode() == Instruction::PHI || (I-> getOpcode() == Instruction::BitCast && I->getType( )->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF )) ? void (0) : __assert_fail ("I->getOpcode() == Instruction::GetElementPtr || I->getOpcode() == Instruction::PHI || (I->getOpcode() == Instruction::BitCast && I->getType()->isPointerTy()) || hasSingleCopyAfterVectorization(I, VF)" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7035, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7036 | VectorTy = RetTy; | ||||||
| 7037 | } else | ||||||
| 7038 | VectorTy = ToVectorTy(RetTy, VF); | ||||||
| 7039 | |||||||
| 7040 | // TODO: We need to estimate the cost of intrinsic calls. | ||||||
| 7041 | switch (I->getOpcode()) { | ||||||
| 7042 | case Instruction::GetElementPtr: | ||||||
| 7043 | // We mark this instruction as zero-cost because the cost of GEPs in | ||||||
| 7044 | // vectorized code depends on whether the corresponding memory instruction | ||||||
| 7045 | // is scalarized or not. Therefore, we handle GEPs with the memory | ||||||
| 7046 | // instruction cost. | ||||||
| 7047 | return 0; | ||||||
| 7048 | case Instruction::Br: { | ||||||
| 7049 | // In cases of scalarized and predicated instructions, there will be VF | ||||||
| 7050 | // predicated blocks in the vectorized loop. Each branch around these | ||||||
| 7051 | // blocks requires also an extract of its vector compare i1 element. | ||||||
| 7052 | bool ScalarPredicatedBB = false; | ||||||
| 7053 | BranchInst *BI = cast<BranchInst>(I); | ||||||
| 7054 | if (VF.isVector() && BI->isConditional() && | ||||||
| 7055 | (PredicatedBBsAfterVectorization[VF].count(BI->getSuccessor(0)) || | ||||||
| 7056 | PredicatedBBsAfterVectorization[VF].count(BI->getSuccessor(1)))) | ||||||
| 7057 | ScalarPredicatedBB = true; | ||||||
| 7058 | |||||||
| 7059 | if (ScalarPredicatedBB) { | ||||||
| 7060 | // Not possible to scalarize scalable vector with predicated instructions. | ||||||
| 7061 | if (VF.isScalable()) | ||||||
| 7062 | return InstructionCost::getInvalid(); | ||||||
| 7063 | // Return cost for branches around scalarized and predicated blocks. | ||||||
| 7064 | auto *Vec_i1Ty = | ||||||
| 7065 | VectorType::get(IntegerType::getInt1Ty(RetTy->getContext()), VF); | ||||||
| 7066 | return ( | ||||||
| 7067 | TTI.getScalarizationOverhead( | ||||||
| 7068 | Vec_i1Ty, APInt::getAllOnes(VF.getFixedValue()), false, true) + | ||||||
| 7069 | (TTI.getCFInstrCost(Instruction::Br, CostKind) * VF.getFixedValue())); | ||||||
| 7070 | } else if (I->getParent() == TheLoop->getLoopLatch() || VF.isScalar()) | ||||||
| 7071 | // The back-edge branch will remain, as will all scalar branches. | ||||||
| 7072 | return TTI.getCFInstrCost(Instruction::Br, CostKind); | ||||||
| 7073 | else | ||||||
| 7074 | // This branch will be eliminated by if-conversion. | ||||||
| 7075 | return 0; | ||||||
| 7076 | // Note: We currently assume zero cost for an unconditional branch inside | ||||||
| 7077 | // a predicated block since it will become a fall-through, although we | ||||||
| 7078 | // may decide in the future to call TTI for all branches. | ||||||
| 7079 | } | ||||||
| 7080 | case Instruction::PHI: { | ||||||
| 7081 | auto *Phi = cast<PHINode>(I); | ||||||
| 7082 | |||||||
| 7083 | // First-order recurrences are replaced by vector shuffles inside the loop. | ||||||
| 7084 | if (VF.isVector() && Legal->isFixedOrderRecurrence(Phi)) { | ||||||
| 7085 | SmallVector<int> Mask(VF.getKnownMinValue()); | ||||||
| 7086 | std::iota(Mask.begin(), Mask.end(), VF.getKnownMinValue() - 1); | ||||||
| 7087 | return TTI.getShuffleCost(TargetTransformInfo::SK_Splice, | ||||||
| 7088 | cast<VectorType>(VectorTy), Mask, CostKind, | ||||||
| 7089 | VF.getKnownMinValue() - 1); | ||||||
| 7090 | } | ||||||
| 7091 | |||||||
| 7092 | // Phi nodes in non-header blocks (not inductions, reductions, etc.) are | ||||||
| 7093 | // converted into select instructions. We require N - 1 selects per phi | ||||||
| 7094 | // node, where N is the number of incoming values. | ||||||
| 7095 | if (VF.isVector() && Phi->getParent() != TheLoop->getHeader()) | ||||||
| 7096 | return (Phi->getNumIncomingValues() - 1) * | ||||||
| 7097 | TTI.getCmpSelInstrCost( | ||||||
| 7098 | Instruction::Select, ToVectorTy(Phi->getType(), VF), | ||||||
| 7099 | ToVectorTy(Type::getInt1Ty(Phi->getContext()), VF), | ||||||
| 7100 | CmpInst::BAD_ICMP_PREDICATE, CostKind); | ||||||
| 7101 | |||||||
| 7102 | return TTI.getCFInstrCost(Instruction::PHI, CostKind); | ||||||
| 7103 | } | ||||||
| 7104 | case Instruction::UDiv: | ||||||
| 7105 | case Instruction::SDiv: | ||||||
| 7106 | case Instruction::URem: | ||||||
| 7107 | case Instruction::SRem: | ||||||
| 7108 | if (VF.isVector() && isPredicatedInst(I)) { | ||||||
| 7109 | const auto [ScalarCost, SafeDivisorCost] = getDivRemSpeculationCost(I, VF); | ||||||
| 7110 | return isDivRemScalarWithPredication(ScalarCost, SafeDivisorCost) ? | ||||||
| 7111 | ScalarCost : SafeDivisorCost; | ||||||
| 7112 | } | ||||||
| 7113 | // We've proven all lanes safe to speculate, fall through. | ||||||
| 7114 | [[fallthrough]]; | ||||||
| 7115 | case Instruction::Add: | ||||||
| 7116 | case Instruction::FAdd: | ||||||
| 7117 | case Instruction::Sub: | ||||||
| 7118 | case Instruction::FSub: | ||||||
| 7119 | case Instruction::Mul: | ||||||
| 7120 | case Instruction::FMul: | ||||||
| 7121 | case Instruction::FDiv: | ||||||
| 7122 | case Instruction::FRem: | ||||||
| 7123 | case Instruction::Shl: | ||||||
| 7124 | case Instruction::LShr: | ||||||
| 7125 | case Instruction::AShr: | ||||||
| 7126 | case Instruction::And: | ||||||
| 7127 | case Instruction::Or: | ||||||
| 7128 | case Instruction::Xor: { | ||||||
| 7129 | // Since we will replace the stride by 1 the multiplication should go away. | ||||||
| 7130 | if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal)) | ||||||
| 7131 | return 0; | ||||||
| 7132 | |||||||
| 7133 | // Detect reduction patterns | ||||||
| 7134 | if (auto RedCost = getReductionPatternCost(I, VF, VectorTy, CostKind)) | ||||||
| 7135 | return *RedCost; | ||||||
| 7136 | |||||||
| 7137 | // Certain instructions can be cheaper to vectorize if they have a constant | ||||||
| 7138 | // second vector operand. One example of this are shifts on x86. | ||||||
| 7139 | Value *Op2 = I->getOperand(1); | ||||||
| 7140 | auto Op2Info = TTI.getOperandInfo(Op2); | ||||||
| 7141 | if (Op2Info.Kind == TargetTransformInfo::OK_AnyValue && Legal->isUniform(Op2)) | ||||||
| 7142 | Op2Info.Kind = TargetTransformInfo::OK_UniformValue; | ||||||
| 7143 | |||||||
| 7144 | SmallVector<const Value *, 4> Operands(I->operand_values()); | ||||||
| 7145 | return TTI.getArithmeticInstrCost( | ||||||
| 7146 | I->getOpcode(), VectorTy, CostKind, | ||||||
| 7147 | {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None}, | ||||||
| 7148 | Op2Info, Operands, I); | ||||||
| 7149 | } | ||||||
| 7150 | case Instruction::FNeg: { | ||||||
| 7151 | return TTI.getArithmeticInstrCost( | ||||||
| 7152 | I->getOpcode(), VectorTy, CostKind, | ||||||
| 7153 | {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None}, | ||||||
| 7154 | {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None}, | ||||||
| 7155 | I->getOperand(0), I); | ||||||
| 7156 | } | ||||||
| 7157 | case Instruction::Select: { | ||||||
| 7158 | SelectInst *SI = cast<SelectInst>(I); | ||||||
| 7159 | const SCEV *CondSCEV = SE->getSCEV(SI->getCondition()); | ||||||
| 7160 | bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop)); | ||||||
| 7161 | |||||||
| 7162 | const Value *Op0, *Op1; | ||||||
| 7163 | using namespace llvm::PatternMatch; | ||||||
| 7164 | if (!ScalarCond && (match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1))) || | ||||||
| 7165 | match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1))))) { | ||||||
| 7166 | // select x, y, false --> x & y | ||||||
| 7167 | // select x, true, y --> x | y | ||||||
| 7168 | const auto [Op1VK, Op1VP] = TTI::getOperandInfo(Op0); | ||||||
| 7169 | const auto [Op2VK, Op2VP] = TTI::getOperandInfo(Op1); | ||||||
| 7170 | assert(Op0->getType()->getScalarSizeInBits() == 1 &&(static_cast <bool> (Op0->getType()->getScalarSizeInBits () == 1 && Op1->getType()->getScalarSizeInBits( ) == 1) ? void (0) : __assert_fail ("Op0->getType()->getScalarSizeInBits() == 1 && Op1->getType()->getScalarSizeInBits() == 1" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7171, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7171 | Op1->getType()->getScalarSizeInBits() == 1)(static_cast <bool> (Op0->getType()->getScalarSizeInBits () == 1 && Op1->getType()->getScalarSizeInBits( ) == 1) ? void (0) : __assert_fail ("Op0->getType()->getScalarSizeInBits() == 1 && Op1->getType()->getScalarSizeInBits() == 1" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7171, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7172 | |||||||
| 7173 | SmallVector<const Value *, 2> Operands{Op0, Op1}; | ||||||
| 7174 | return TTI.getArithmeticInstrCost( | ||||||
| 7175 | match(I, m_LogicalOr()) ? Instruction::Or : Instruction::And, VectorTy, | ||||||
| 7176 | CostKind, {Op1VK, Op1VP}, {Op2VK, Op2VP}, Operands, I); | ||||||
| 7177 | } | ||||||
| 7178 | |||||||
| 7179 | Type *CondTy = SI->getCondition()->getType(); | ||||||
| 7180 | if (!ScalarCond) | ||||||
| 7181 | CondTy = VectorType::get(CondTy, VF); | ||||||
| 7182 | |||||||
| 7183 | CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE; | ||||||
| 7184 | if (auto *Cmp = dyn_cast<CmpInst>(SI->getCondition())) | ||||||
| 7185 | Pred = Cmp->getPredicate(); | ||||||
| 7186 | return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy, Pred, | ||||||
| 7187 | CostKind, I); | ||||||
| 7188 | } | ||||||
| 7189 | case Instruction::ICmp: | ||||||
| 7190 | case Instruction::FCmp: { | ||||||
| 7191 | Type *ValTy = I->getOperand(0)->getType(); | ||||||
| 7192 | Instruction *Op0AsInstruction = dyn_cast<Instruction>(I->getOperand(0)); | ||||||
| 7193 | if (canTruncateToMinimalBitwidth(Op0AsInstruction, VF)) | ||||||
| 7194 | ValTy = IntegerType::get(ValTy->getContext(), MinBWs[Op0AsInstruction]); | ||||||
| 7195 | VectorTy = ToVectorTy(ValTy, VF); | ||||||
| 7196 | return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, nullptr, | ||||||
| 7197 | cast<CmpInst>(I)->getPredicate(), CostKind, | ||||||
| 7198 | I); | ||||||
| 7199 | } | ||||||
| 7200 | case Instruction::Store: | ||||||
| 7201 | case Instruction::Load: { | ||||||
| 7202 | ElementCount Width = VF; | ||||||
| 7203 | if (Width.isVector()) { | ||||||
| 7204 | InstWidening Decision = getWideningDecision(I, Width); | ||||||
| 7205 | assert(Decision != CM_Unknown &&(static_cast <bool> (Decision != CM_Unknown && "CM decision should be taken at this point" ) ? void (0) : __assert_fail ("Decision != CM_Unknown && \"CM decision should be taken at this point\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7206, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7206 | "CM decision should be taken at this point")(static_cast <bool> (Decision != CM_Unknown && "CM decision should be taken at this point" ) ? void (0) : __assert_fail ("Decision != CM_Unknown && \"CM decision should be taken at this point\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7206, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7207 | if (getWideningCost(I, VF) == InstructionCost::getInvalid()) | ||||||
| 7208 | return InstructionCost::getInvalid(); | ||||||
| 7209 | if (Decision == CM_Scalarize) | ||||||
| 7210 | Width = ElementCount::getFixed(1); | ||||||
| 7211 | } | ||||||
| 7212 | VectorTy = ToVectorTy(getLoadStoreType(I), Width); | ||||||
| 7213 | return getMemoryInstructionCost(I, VF); | ||||||
| 7214 | } | ||||||
| 7215 | case Instruction::BitCast: | ||||||
| 7216 | if (I->getType()->isPointerTy()) | ||||||
| 7217 | return 0; | ||||||
| 7218 | [[fallthrough]]; | ||||||
| 7219 | case Instruction::ZExt: | ||||||
| 7220 | case Instruction::SExt: | ||||||
| 7221 | case Instruction::FPToUI: | ||||||
| 7222 | case Instruction::FPToSI: | ||||||
| 7223 | case Instruction::FPExt: | ||||||
| 7224 | case Instruction::PtrToInt: | ||||||
| 7225 | case Instruction::IntToPtr: | ||||||
| 7226 | case Instruction::SIToFP: | ||||||
| 7227 | case Instruction::UIToFP: | ||||||
| 7228 | case Instruction::Trunc: | ||||||
| 7229 | case Instruction::FPTrunc: { | ||||||
| 7230 | // Computes the CastContextHint from a Load/Store instruction. | ||||||
| 7231 | auto ComputeCCH = [&](Instruction *I) -> TTI::CastContextHint { | ||||||
| 7232 | assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&(static_cast <bool> ((isa<LoadInst>(I) || isa< StoreInst>(I)) && "Expected a load or a store!") ? void (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected a load or a store!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7233, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7233 | "Expected a load or a store!")(static_cast <bool> ((isa<LoadInst>(I) || isa< StoreInst>(I)) && "Expected a load or a store!") ? void (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Expected a load or a store!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7233, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7234 | |||||||
| 7235 | if (VF.isScalar() || !TheLoop->contains(I)) | ||||||
| 7236 | return TTI::CastContextHint::Normal; | ||||||
| 7237 | |||||||
| 7238 | switch (getWideningDecision(I, VF)) { | ||||||
| 7239 | case LoopVectorizationCostModel::CM_GatherScatter: | ||||||
| 7240 | return TTI::CastContextHint::GatherScatter; | ||||||
| 7241 | case LoopVectorizationCostModel::CM_Interleave: | ||||||
| 7242 | return TTI::CastContextHint::Interleave; | ||||||
| 7243 | case LoopVectorizationCostModel::CM_Scalarize: | ||||||
| 7244 | case LoopVectorizationCostModel::CM_Widen: | ||||||
| 7245 | return Legal->isMaskRequired(I) ? TTI::CastContextHint::Masked | ||||||
| 7246 | : TTI::CastContextHint::Normal; | ||||||
| 7247 | case LoopVectorizationCostModel::CM_Widen_Reverse: | ||||||
| 7248 | return TTI::CastContextHint::Reversed; | ||||||
| 7249 | case LoopVectorizationCostModel::CM_Unknown: | ||||||
| 7250 | llvm_unreachable("Instr did not go through cost modelling?")::llvm::llvm_unreachable_internal("Instr did not go through cost modelling?" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7250); | ||||||
| 7251 | } | ||||||
| 7252 | |||||||
| 7253 | llvm_unreachable("Unhandled case!")::llvm::llvm_unreachable_internal("Unhandled case!", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7253); | ||||||
| 7254 | }; | ||||||
| 7255 | |||||||
| 7256 | unsigned Opcode = I->getOpcode(); | ||||||
| 7257 | TTI::CastContextHint CCH = TTI::CastContextHint::None; | ||||||
| 7258 | // For Trunc, the context is the only user, which must be a StoreInst. | ||||||
| 7259 | if (Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) { | ||||||
| 7260 | if (I->hasOneUse()) | ||||||
| 7261 | if (StoreInst *Store = dyn_cast<StoreInst>(*I->user_begin())) | ||||||
| 7262 | CCH = ComputeCCH(Store); | ||||||
| 7263 | } | ||||||
| 7264 | // For Z/Sext, the context is the operand, which must be a LoadInst. | ||||||
| 7265 | else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt || | ||||||
| 7266 | Opcode == Instruction::FPExt) { | ||||||
| 7267 | if (LoadInst *Load = dyn_cast<LoadInst>(I->getOperand(0))) | ||||||
| 7268 | CCH = ComputeCCH(Load); | ||||||
| 7269 | } | ||||||
| 7270 | |||||||
| 7271 | // We optimize the truncation of induction variables having constant | ||||||
| 7272 | // integer steps. The cost of these truncations is the same as the scalar | ||||||
| 7273 | // operation. | ||||||
| 7274 | if (isOptimizableIVTruncate(I, VF)) { | ||||||
| 7275 | auto *Trunc = cast<TruncInst>(I); | ||||||
| 7276 | return TTI.getCastInstrCost(Instruction::Trunc, Trunc->getDestTy(), | ||||||
| 7277 | Trunc->getSrcTy(), CCH, CostKind, Trunc); | ||||||
| 7278 | } | ||||||
| 7279 | |||||||
| 7280 | // Detect reduction patterns | ||||||
| 7281 | if (auto RedCost = getReductionPatternCost(I, VF, VectorTy, CostKind)) | ||||||
| 7282 | return *RedCost; | ||||||
| 7283 | |||||||
| 7284 | Type *SrcScalarTy = I->getOperand(0)->getType(); | ||||||
| 7285 | Type *SrcVecTy = | ||||||
| 7286 | VectorTy->isVectorTy() ? ToVectorTy(SrcScalarTy, VF) : SrcScalarTy; | ||||||
| 7287 | if (canTruncateToMinimalBitwidth(I, VF)) { | ||||||
| 7288 | // This cast is going to be shrunk. This may remove the cast or it might | ||||||
| 7289 | // turn it into slightly different cast. For example, if MinBW == 16, | ||||||
| 7290 | // "zext i8 %1 to i32" becomes "zext i8 %1 to i16". | ||||||
| 7291 | // | ||||||
| 7292 | // Calculate the modified src and dest types. | ||||||
| 7293 | Type *MinVecTy = VectorTy; | ||||||
| 7294 | if (Opcode == Instruction::Trunc) { | ||||||
| 7295 | SrcVecTy = smallestIntegerVectorType(SrcVecTy, MinVecTy); | ||||||
| 7296 | VectorTy = | ||||||
| 7297 | largestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy); | ||||||
| 7298 | } else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) { | ||||||
| 7299 | SrcVecTy = largestIntegerVectorType(SrcVecTy, MinVecTy); | ||||||
| 7300 | VectorTy = | ||||||
| 7301 | smallestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy); | ||||||
| 7302 | } | ||||||
| 7303 | } | ||||||
| 7304 | |||||||
| 7305 | return TTI.getCastInstrCost(Opcode, VectorTy, SrcVecTy, CCH, CostKind, I); | ||||||
| 7306 | } | ||||||
| 7307 | case Instruction::Call: { | ||||||
| 7308 | if (RecurrenceDescriptor::isFMulAddIntrinsic(I)) | ||||||
| 7309 | if (auto RedCost = getReductionPatternCost(I, VF, VectorTy, CostKind)) | ||||||
| 7310 | return *RedCost; | ||||||
| 7311 | bool NeedToScalarize; | ||||||
| 7312 | CallInst *CI = cast<CallInst>(I); | ||||||
| 7313 | InstructionCost CallCost = getVectorCallCost(CI, VF, NeedToScalarize); | ||||||
| 7314 | if (getVectorIntrinsicIDForCall(CI, TLI)) { | ||||||
| 7315 | InstructionCost IntrinsicCost = getVectorIntrinsicCost(CI, VF); | ||||||
| 7316 | return std::min(CallCost, IntrinsicCost); | ||||||
| 7317 | } | ||||||
| 7318 | return CallCost; | ||||||
| 7319 | } | ||||||
| 7320 | case Instruction::ExtractValue: | ||||||
| 7321 | return TTI.getInstructionCost(I, TTI::TCK_RecipThroughput); | ||||||
| 7322 | case Instruction::Alloca: | ||||||
| 7323 | // We cannot easily widen alloca to a scalable alloca, as | ||||||
| 7324 | // the result would need to be a vector of pointers. | ||||||
| 7325 | if (VF.isScalable()) | ||||||
| 7326 | return InstructionCost::getInvalid(); | ||||||
| 7327 | [[fallthrough]]; | ||||||
| 7328 | default: | ||||||
| 7329 | // This opcode is unknown. Assume that it is the same as 'mul'. | ||||||
| 7330 | return TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); | ||||||
| 7331 | } // end of switch. | ||||||
| 7332 | } | ||||||
| 7333 | |||||||
| 7334 | char LoopVectorize::ID = 0; | ||||||
| 7335 | |||||||
| 7336 | static const char lv_name[] = "Loop Vectorization"; | ||||||
| 7337 | |||||||
| 7338 | INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)static void *initializeLoopVectorizePassOnce(PassRegistry & Registry) { | ||||||
| 7339 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | ||||||
| 7340 | INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)initializeBasicAAWrapperPassPass(Registry); | ||||||
| 7341 | INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)initializeGlobalsAAWrapperPassPass(Registry); | ||||||
| 7342 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); | ||||||
| 7343 | INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)initializeBlockFrequencyInfoWrapperPassPass(Registry); | ||||||
| 7344 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); | ||||||
| 7345 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry); | ||||||
| 7346 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry); | ||||||
| 7347 | INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis)initializeLoopAccessLegacyAnalysisPass(Registry); | ||||||
| 7348 | INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)initializeDemandedBitsWrapperPassPass(Registry); | ||||||
| 7349 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)initializeOptimizationRemarkEmitterWrapperPassPass(Registry); | ||||||
| 7350 | INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)initializeProfileSummaryInfoWrapperPassPass(Registry); | ||||||
| 7351 | INITIALIZE_PASS_DEPENDENCY(InjectTLIMappingsLegacy)initializeInjectTLIMappingsLegacyPass(Registry); | ||||||
| 7352 | INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)PassInfo *PI = new PassInfo( lv_name, "loop-vectorize", & LoopVectorize::ID, PassInfo::NormalCtor_t(callDefaultCtor< LoopVectorize>), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeLoopVectorizePassFlag ; void llvm::initializeLoopVectorizePass(PassRegistry &Registry ) { llvm::call_once(InitializeLoopVectorizePassFlag, initializeLoopVectorizePassOnce , std::ref(Registry)); } | ||||||
| 7353 | |||||||
| 7354 | namespace llvm { | ||||||
| 7355 | |||||||
| 7356 | Pass *createLoopVectorizePass() { return new LoopVectorize(); } | ||||||
| 7357 | |||||||
| 7358 | Pass *createLoopVectorizePass(bool InterleaveOnlyWhenForced, | ||||||
| 7359 | bool VectorizeOnlyWhenForced) { | ||||||
| 7360 | return new LoopVectorize(InterleaveOnlyWhenForced, VectorizeOnlyWhenForced); | ||||||
| 7361 | } | ||||||
| 7362 | |||||||
| 7363 | } // end namespace llvm | ||||||
| 7364 | |||||||
| 7365 | void LoopVectorizationCostModel::collectValuesToIgnore() { | ||||||
| 7366 | // Ignore ephemeral values. | ||||||
| 7367 | CodeMetrics::collectEphemeralValues(TheLoop, AC, ValuesToIgnore); | ||||||
| 7368 | |||||||
| 7369 | // Find all stores to invariant variables. Since they are going to sink | ||||||
| 7370 | // outside the loop we do not need calculate cost for them. | ||||||
| 7371 | for (BasicBlock *BB : TheLoop->blocks()) | ||||||
| 7372 | for (Instruction &I : *BB) { | ||||||
| 7373 | StoreInst *SI; | ||||||
| 7374 | if ((SI = dyn_cast<StoreInst>(&I)) && | ||||||
| 7375 | Legal->isInvariantAddressOfReduction(SI->getPointerOperand())) | ||||||
| 7376 | ValuesToIgnore.insert(&I); | ||||||
| 7377 | } | ||||||
| 7378 | |||||||
| 7379 | // Ignore type-promoting instructions we identified during reduction | ||||||
| 7380 | // detection. | ||||||
| 7381 | for (const auto &Reduction : Legal->getReductionVars()) { | ||||||
| 7382 | const RecurrenceDescriptor &RedDes = Reduction.second; | ||||||
| 7383 | const SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts(); | ||||||
| 7384 | VecValuesToIgnore.insert(Casts.begin(), Casts.end()); | ||||||
| 7385 | } | ||||||
| 7386 | // Ignore type-casting instructions we identified during induction | ||||||
| 7387 | // detection. | ||||||
| 7388 | for (const auto &Induction : Legal->getInductionVars()) { | ||||||
| 7389 | const InductionDescriptor &IndDes = Induction.second; | ||||||
| 7390 | const SmallVectorImpl<Instruction *> &Casts = IndDes.getCastInsts(); | ||||||
| 7391 | VecValuesToIgnore.insert(Casts.begin(), Casts.end()); | ||||||
| 7392 | } | ||||||
| 7393 | } | ||||||
| 7394 | |||||||
| 7395 | void LoopVectorizationCostModel::collectInLoopReductions() { | ||||||
| 7396 | for (const auto &Reduction : Legal->getReductionVars()) { | ||||||
| 7397 | PHINode *Phi = Reduction.first; | ||||||
| 7398 | const RecurrenceDescriptor &RdxDesc = Reduction.second; | ||||||
| 7399 | |||||||
| 7400 | // We don't collect reductions that are type promoted (yet). | ||||||
| 7401 | if (RdxDesc.getRecurrenceType() != Phi->getType()) | ||||||
| 7402 | continue; | ||||||
| 7403 | |||||||
| 7404 | // If the target would prefer this reduction to happen "in-loop", then we | ||||||
| 7405 | // want to record it as such. | ||||||
| 7406 | unsigned Opcode = RdxDesc.getOpcode(); | ||||||
| 7407 | if (!PreferInLoopReductions && !useOrderedReductions(RdxDesc) && | ||||||
| 7408 | !TTI.preferInLoopReduction(Opcode, Phi->getType(), | ||||||
| 7409 | TargetTransformInfo::ReductionFlags())) | ||||||
| 7410 | continue; | ||||||
| 7411 | |||||||
| 7412 | // Check that we can correctly put the reductions into the loop, by | ||||||
| 7413 | // finding the chain of operations that leads from the phi to the loop | ||||||
| 7414 | // exit value. | ||||||
| 7415 | SmallVector<Instruction *, 4> ReductionOperations = | ||||||
| 7416 | RdxDesc.getReductionOpChain(Phi, TheLoop); | ||||||
| 7417 | bool InLoop = !ReductionOperations.empty(); | ||||||
| 7418 | if (InLoop) { | ||||||
| 7419 | InLoopReductionChains[Phi] = ReductionOperations; | ||||||
| 7420 | // Add the elements to InLoopReductionImmediateChains for cost modelling. | ||||||
| 7421 | Instruction *LastChain = Phi; | ||||||
| 7422 | for (auto *I : ReductionOperations) { | ||||||
| 7423 | InLoopReductionImmediateChains[I] = LastChain; | ||||||
| 7424 | LastChain = I; | ||||||
| 7425 | } | ||||||
| 7426 | } | ||||||
| 7427 | LLVM_DEBUG(dbgs() << "LV: Using " << (InLoop ? "inloop" : "out of loop")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Using " << ( InLoop ? "inloop" : "out of loop") << " reduction for phi: " << *Phi << "\n"; } } while (false) | ||||||
| 7428 | << " reduction for phi: " << *Phi << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Using " << ( InLoop ? "inloop" : "out of loop") << " reduction for phi: " << *Phi << "\n"; } } while (false); | ||||||
| 7429 | } | ||||||
| 7430 | } | ||||||
| 7431 | |||||||
| 7432 | // TODO: we could return a pair of values that specify the max VF and | ||||||
| 7433 | // min VF, to be used in `buildVPlans(MinVF, MaxVF)` instead of | ||||||
| 7434 | // `buildVPlans(VF, VF)`. We cannot do it because VPLAN at the moment | ||||||
| 7435 | // doesn't have a cost model that can choose which plan to execute if | ||||||
| 7436 | // more than one is generated. | ||||||
| 7437 | static unsigned determineVPlanVF(const unsigned WidestVectorRegBits, | ||||||
| 7438 | LoopVectorizationCostModel &CM) { | ||||||
| 7439 | unsigned WidestType; | ||||||
| 7440 | std::tie(std::ignore, WidestType) = CM.getSmallestAndWidestTypes(); | ||||||
| 7441 | return WidestVectorRegBits / WidestType; | ||||||
| 7442 | } | ||||||
| 7443 | |||||||
| 7444 | VectorizationFactor | ||||||
| 7445 | LoopVectorizationPlanner::planInVPlanNativePath(ElementCount UserVF) { | ||||||
| 7446 | assert(!UserVF.isScalable() && "scalable vectors not yet supported")(static_cast <bool> (!UserVF.isScalable() && "scalable vectors not yet supported" ) ? void (0) : __assert_fail ("!UserVF.isScalable() && \"scalable vectors not yet supported\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7446, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7447 | ElementCount VF = UserVF; | ||||||
| 7448 | // Outer loop handling: They may require CFG and instruction level | ||||||
| 7449 | // transformations before even evaluating whether vectorization is profitable. | ||||||
| 7450 | // Since we cannot modify the incoming IR, we need to build VPlan upfront in | ||||||
| 7451 | // the vectorization pipeline. | ||||||
| 7452 | if (!OrigLoop->isInnermost()) { | ||||||
| 7453 | // If the user doesn't provide a vectorization factor, determine a | ||||||
| 7454 | // reasonable one. | ||||||
| 7455 | if (UserVF.isZero()) { | ||||||
| 7456 | VF = ElementCount::getFixed(determineVPlanVF( | ||||||
| 7457 | TTI->getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector) | ||||||
| 7458 | .getFixedSize(), | ||||||
| 7459 | CM)); | ||||||
| 7460 | LLVM_DEBUG(dbgs() << "LV: VPlan computed VF " << VF << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: VPlan computed VF " << VF << ".\n"; } } while (false); | ||||||
| 7461 | |||||||
| 7462 | // Make sure we have a VF > 1 for stress testing. | ||||||
| 7463 | if (VPlanBuildStressTest && (VF.isScalar() || VF.isZero())) { | ||||||
| 7464 | LLVM_DEBUG(dbgs() << "LV: VPlan stress testing: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: VPlan stress testing: " << "overriding computed VF.\n"; } } while (false) | ||||||
| 7465 | << "overriding computed VF.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: VPlan stress testing: " << "overriding computed VF.\n"; } } while (false); | ||||||
| 7466 | VF = ElementCount::getFixed(4); | ||||||
| 7467 | } | ||||||
| 7468 | } | ||||||
| 7469 | assert(EnableVPlanNativePath && "VPlan-native path is not enabled.")(static_cast <bool> (EnableVPlanNativePath && "VPlan-native path is not enabled." ) ? void (0) : __assert_fail ("EnableVPlanNativePath && \"VPlan-native path is not enabled.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7469, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7470 | assert(isPowerOf2_32(VF.getKnownMinValue()) &&(static_cast <bool> (isPowerOf2_32(VF.getKnownMinValue( )) && "VF needs to be a power of two") ? void (0) : __assert_fail ("isPowerOf2_32(VF.getKnownMinValue()) && \"VF needs to be a power of two\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7471, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7471 | "VF needs to be a power of two")(static_cast <bool> (isPowerOf2_32(VF.getKnownMinValue( )) && "VF needs to be a power of two") ? void (0) : __assert_fail ("isPowerOf2_32(VF.getKnownMinValue()) && \"VF needs to be a power of two\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7471, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7472 | LLVM_DEBUG(dbgs() << "LV: Using " << (!UserVF.isZero() ? "user " : "")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Using " << ( !UserVF.isZero() ? "user " : "") << "VF " << VF << " to build VPlans.\n"; } } while (false) | ||||||
| 7473 | << "VF " << VF << " to build VPlans.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Using " << ( !UserVF.isZero() ? "user " : "") << "VF " << VF << " to build VPlans.\n"; } } while (false); | ||||||
| 7474 | buildVPlans(VF, VF); | ||||||
| 7475 | |||||||
| 7476 | // For VPlan build stress testing, we bail out after VPlan construction. | ||||||
| 7477 | if (VPlanBuildStressTest) | ||||||
| 7478 | return VectorizationFactor::Disabled(); | ||||||
| 7479 | |||||||
| 7480 | return {VF, 0 /*Cost*/, 0 /* ScalarCost */}; | ||||||
| 7481 | } | ||||||
| 7482 | |||||||
| 7483 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not vectorizing. Inner loops aren't supported in the " "VPlan-native path.\n"; } } while (false) | ||||||
| 7484 | dbgs() << "LV: Not vectorizing. Inner loops aren't supported in the "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not vectorizing. Inner loops aren't supported in the " "VPlan-native path.\n"; } } while (false) | ||||||
| 7485 | "VPlan-native path.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not vectorizing. Inner loops aren't supported in the " "VPlan-native path.\n"; } } while (false); | ||||||
| 7486 | return VectorizationFactor::Disabled(); | ||||||
| 7487 | } | ||||||
| 7488 | |||||||
| 7489 | std::optional<VectorizationFactor> | ||||||
| 7490 | LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) { | ||||||
| 7491 | assert(OrigLoop->isInnermost() && "Inner loop expected.")(static_cast <bool> (OrigLoop->isInnermost() && "Inner loop expected.") ? void (0) : __assert_fail ("OrigLoop->isInnermost() && \"Inner loop expected.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7491, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7492 | FixedScalableVFPair MaxFactors = CM.computeMaxVF(UserVF, UserIC); | ||||||
| 7493 | if (!MaxFactors) // Cases that should not to be vectorized nor interleaved. | ||||||
| 7494 | return std::nullopt; | ||||||
| 7495 | |||||||
| 7496 | // Invalidate interleave groups if all blocks of loop will be predicated. | ||||||
| 7497 | if (CM.blockNeedsPredicationForAnyReason(OrigLoop->getHeader()) && | ||||||
| 7498 | !useMaskedInterleavedAccesses(*TTI)) { | ||||||
| 7499 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Invalidate all interleaved groups due to fold-tail by masking " "which requires masked-interleaved support.\n"; } } while (false ) | ||||||
| 7500 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Invalidate all interleaved groups due to fold-tail by masking " "which requires masked-interleaved support.\n"; } } while (false ) | ||||||
| 7501 | << "LV: Invalidate all interleaved groups due to fold-tail by masking "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Invalidate all interleaved groups due to fold-tail by masking " "which requires masked-interleaved support.\n"; } } while (false ) | ||||||
| 7502 | "which requires masked-interleaved support.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Invalidate all interleaved groups due to fold-tail by masking " "which requires masked-interleaved support.\n"; } } while (false ); | ||||||
| 7503 | if (CM.InterleaveInfo.invalidateGroups()) | ||||||
| 7504 | // Invalidating interleave groups also requires invalidating all decisions | ||||||
| 7505 | // based on them, which includes widening decisions and uniform and scalar | ||||||
| 7506 | // values. | ||||||
| 7507 | CM.invalidateCostModelingDecisions(); | ||||||
| 7508 | } | ||||||
| 7509 | |||||||
| 7510 | ElementCount MaxUserVF = | ||||||
| 7511 | UserVF.isScalable() ? MaxFactors.ScalableVF : MaxFactors.FixedVF; | ||||||
| 7512 | bool UserVFIsLegal = ElementCount::isKnownLE(UserVF, MaxUserVF); | ||||||
| 7513 | if (!UserVF.isZero() && UserVFIsLegal) { | ||||||
| 7514 | assert(isPowerOf2_32(UserVF.getKnownMinValue()) &&(static_cast <bool> (isPowerOf2_32(UserVF.getKnownMinValue ()) && "VF needs to be a power of two") ? void (0) : __assert_fail ("isPowerOf2_32(UserVF.getKnownMinValue()) && \"VF needs to be a power of two\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7515, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7515 | "VF needs to be a power of two")(static_cast <bool> (isPowerOf2_32(UserVF.getKnownMinValue ()) && "VF needs to be a power of two") ? void (0) : __assert_fail ("isPowerOf2_32(UserVF.getKnownMinValue()) && \"VF needs to be a power of two\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7515, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7516 | // Collect the instructions (and their associated costs) that will be more | ||||||
| 7517 | // profitable to scalarize. | ||||||
| 7518 | if (CM.selectUserVectorizationFactor(UserVF)) { | ||||||
| 7519 | LLVM_DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Using user VF " << UserVF << ".\n"; } } while (false); | ||||||
| 7520 | CM.collectInLoopReductions(); | ||||||
| 7521 | buildVPlansWithVPRecipes(UserVF, UserVF); | ||||||
| 7522 | LLVM_DEBUG(printPlans(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { printPlans(dbgs()); } } while (false); | ||||||
| 7523 | return {{UserVF, 0, 0}}; | ||||||
| 7524 | } else | ||||||
| 7525 | reportVectorizationInfo("UserVF ignored because of invalid costs.", | ||||||
| 7526 | "InvalidCost", ORE, OrigLoop); | ||||||
| 7527 | } | ||||||
| 7528 | |||||||
| 7529 | // Populate the set of Vectorization Factor Candidates. | ||||||
| 7530 | ElementCountSet VFCandidates; | ||||||
| 7531 | for (auto VF = ElementCount::getFixed(1); | ||||||
| 7532 | ElementCount::isKnownLE(VF, MaxFactors.FixedVF); VF *= 2) | ||||||
| 7533 | VFCandidates.insert(VF); | ||||||
| 7534 | for (auto VF = ElementCount::getScalable(1); | ||||||
| 7535 | ElementCount::isKnownLE(VF, MaxFactors.ScalableVF); VF *= 2) | ||||||
| 7536 | VFCandidates.insert(VF); | ||||||
| 7537 | |||||||
| 7538 | for (const auto &VF : VFCandidates) { | ||||||
| 7539 | // Collect Uniform and Scalar instructions after vectorization with VF. | ||||||
| 7540 | CM.collectUniformsAndScalars(VF); | ||||||
| 7541 | |||||||
| 7542 | // Collect the instructions (and their associated costs) that will be more | ||||||
| 7543 | // profitable to scalarize. | ||||||
| 7544 | if (VF.isVector()) | ||||||
| 7545 | CM.collectInstsToScalarize(VF); | ||||||
| 7546 | } | ||||||
| 7547 | |||||||
| 7548 | CM.collectInLoopReductions(); | ||||||
| 7549 | buildVPlansWithVPRecipes(ElementCount::getFixed(1), MaxFactors.FixedVF); | ||||||
| 7550 | buildVPlansWithVPRecipes(ElementCount::getScalable(1), MaxFactors.ScalableVF); | ||||||
| 7551 | |||||||
| 7552 | LLVM_DEBUG(printPlans(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { printPlans(dbgs()); } } while (false); | ||||||
| 7553 | if (!MaxFactors.hasVector()) | ||||||
| 7554 | return VectorizationFactor::Disabled(); | ||||||
| 7555 | |||||||
| 7556 | // Select the optimal vectorization factor. | ||||||
| 7557 | VectorizationFactor VF = CM.selectVectorizationFactor(VFCandidates); | ||||||
| 7558 | assert((VF.Width.isScalar() || VF.ScalarCost > 0) && "when vectorizing, the scalar cost must be non-zero.")(static_cast <bool> ((VF.Width.isScalar() || VF.ScalarCost > 0) && "when vectorizing, the scalar cost must be non-zero." ) ? void (0) : __assert_fail ("(VF.Width.isScalar() || VF.ScalarCost > 0) && \"when vectorizing, the scalar cost must be non-zero.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7558, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7559 | return VF; | ||||||
| 7560 | } | ||||||
| 7561 | |||||||
| 7562 | VPlan &LoopVectorizationPlanner::getBestPlanFor(ElementCount VF) const { | ||||||
| 7563 | assert(count_if(VPlans,(static_cast <bool> (count_if(VPlans, [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) == 1 && "Best VF has not a single VPlan." ) ? void (0) : __assert_fail ("count_if(VPlans, [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) == 1 && \"Best VF has not a single VPlan.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7566, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7564 | [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) ==(static_cast <bool> (count_if(VPlans, [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) == 1 && "Best VF has not a single VPlan." ) ? void (0) : __assert_fail ("count_if(VPlans, [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) == 1 && \"Best VF has not a single VPlan.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7566, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7565 | 1 &&(static_cast <bool> (count_if(VPlans, [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) == 1 && "Best VF has not a single VPlan." ) ? void (0) : __assert_fail ("count_if(VPlans, [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) == 1 && \"Best VF has not a single VPlan.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7566, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7566 | "Best VF has not a single VPlan.")(static_cast <bool> (count_if(VPlans, [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) == 1 && "Best VF has not a single VPlan." ) ? void (0) : __assert_fail ("count_if(VPlans, [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) == 1 && \"Best VF has not a single VPlan.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7566, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7567 | |||||||
| 7568 | for (const VPlanPtr &Plan : VPlans) { | ||||||
| 7569 | if (Plan->hasVF(VF)) | ||||||
| 7570 | return *Plan.get(); | ||||||
| 7571 | } | ||||||
| 7572 | llvm_unreachable("No plan found!")::llvm::llvm_unreachable_internal("No plan found!", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7572); | ||||||
| 7573 | } | ||||||
| 7574 | |||||||
| 7575 | static void AddRuntimeUnrollDisableMetaData(Loop *L) { | ||||||
| 7576 | SmallVector<Metadata *, 4> MDs; | ||||||
| 7577 | // Reserve first location for self reference to the LoopID metadata node. | ||||||
| 7578 | MDs.push_back(nullptr); | ||||||
| 7579 | bool IsUnrollMetadata = false; | ||||||
| 7580 | MDNode *LoopID = L->getLoopID(); | ||||||
| 7581 | if (LoopID) { | ||||||
| 7582 | // First find existing loop unrolling disable metadata. | ||||||
| 7583 | for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { | ||||||
| 7584 | auto *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); | ||||||
| 7585 | if (MD) { | ||||||
| 7586 | const auto *S = dyn_cast<MDString>(MD->getOperand(0)); | ||||||
| 7587 | IsUnrollMetadata = | ||||||
| 7588 | S && S->getString().startswith("llvm.loop.unroll.disable"); | ||||||
| 7589 | } | ||||||
| 7590 | MDs.push_back(LoopID->getOperand(i)); | ||||||
| 7591 | } | ||||||
| 7592 | } | ||||||
| 7593 | |||||||
| 7594 | if (!IsUnrollMetadata) { | ||||||
| 7595 | // Add runtime unroll disable metadata. | ||||||
| 7596 | LLVMContext &Context = L->getHeader()->getContext(); | ||||||
| 7597 | SmallVector<Metadata *, 1> DisableOperands; | ||||||
| 7598 | DisableOperands.push_back( | ||||||
| 7599 | MDString::get(Context, "llvm.loop.unroll.runtime.disable")); | ||||||
| 7600 | MDNode *DisableNode = MDNode::get(Context, DisableOperands); | ||||||
| 7601 | MDs.push_back(DisableNode); | ||||||
| 7602 | MDNode *NewLoopID = MDNode::get(Context, MDs); | ||||||
| 7603 | // Set operand 0 to refer to the loop id itself. | ||||||
| 7604 | NewLoopID->replaceOperandWith(0, NewLoopID); | ||||||
| 7605 | L->setLoopID(NewLoopID); | ||||||
| 7606 | } | ||||||
| 7607 | } | ||||||
| 7608 | |||||||
| 7609 | void LoopVectorizationPlanner::executePlan(ElementCount BestVF, unsigned BestUF, | ||||||
| 7610 | VPlan &BestVPlan, | ||||||
| 7611 | InnerLoopVectorizer &ILV, | ||||||
| 7612 | DominatorTree *DT, | ||||||
| 7613 | bool IsEpilogueVectorization) { | ||||||
| 7614 | assert(BestVPlan.hasVF(BestVF) &&(static_cast <bool> (BestVPlan.hasVF(BestVF) && "Trying to execute plan with unsupported VF") ? void (0) : __assert_fail ("BestVPlan.hasVF(BestVF) && \"Trying to execute plan with unsupported VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7615, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7615 | "Trying to execute plan with unsupported VF")(static_cast <bool> (BestVPlan.hasVF(BestVF) && "Trying to execute plan with unsupported VF") ? void (0) : __assert_fail ("BestVPlan.hasVF(BestVF) && \"Trying to execute plan with unsupported VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7615, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7616 | assert(BestVPlan.hasUF(BestUF) &&(static_cast <bool> (BestVPlan.hasUF(BestUF) && "Trying to execute plan with unsupported UF") ? void (0) : __assert_fail ("BestVPlan.hasUF(BestUF) && \"Trying to execute plan with unsupported UF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7617, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7617 | "Trying to execute plan with unsupported UF")(static_cast <bool> (BestVPlan.hasUF(BestUF) && "Trying to execute plan with unsupported UF") ? void (0) : __assert_fail ("BestVPlan.hasUF(BestUF) && \"Trying to execute plan with unsupported UF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7617, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7618 | |||||||
| 7619 | LLVM_DEBUG(dbgs() << "Executing best plan with VF=" << BestVF << ", UF=" << BestUFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Executing best plan with VF=" << BestVF << ", UF=" << BestUF << '\n' ; } } while (false) | ||||||
| 7620 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Executing best plan with VF=" << BestVF << ", UF=" << BestUF << '\n' ; } } while (false); | ||||||
| 7621 | |||||||
| 7622 | // Workaround! Compute the trip count of the original loop and cache it | ||||||
| 7623 | // before we start modifying the CFG. This code has a systemic problem | ||||||
| 7624 | // wherein it tries to run analysis over partially constructed IR; this is | ||||||
| 7625 | // wrong, and not simply for SCEV. The trip count of the original loop | ||||||
| 7626 | // simply happens to be prone to hitting this in practice. In theory, we | ||||||
| 7627 | // can hit the same issue for any SCEV, or ValueTracking query done during | ||||||
| 7628 | // mutation. See PR49900. | ||||||
| 7629 | ILV.getOrCreateTripCount(OrigLoop->getLoopPreheader()); | ||||||
| 7630 | |||||||
| 7631 | if (!IsEpilogueVectorization) | ||||||
| 7632 | VPlanTransforms::optimizeForVFAndUF(BestVPlan, BestVF, BestUF, PSE); | ||||||
| 7633 | |||||||
| 7634 | // Perform the actual loop transformation. | ||||||
| 7635 | |||||||
| 7636 | // 1. Set up the skeleton for vectorization, including vector pre-header and | ||||||
| 7637 | // middle block. The vector loop is created during VPlan execution. | ||||||
| 7638 | VPTransformState State{BestVF, BestUF, LI, DT, ILV.Builder, &ILV, &BestVPlan}; | ||||||
| 7639 | Value *CanonicalIVStartValue; | ||||||
| 7640 | std::tie(State.CFG.PrevBB, CanonicalIVStartValue) = | ||||||
| 7641 | ILV.createVectorizedLoopSkeleton(); | ||||||
| 7642 | |||||||
| 7643 | // Only use noalias metadata when using memory checks guaranteeing no overlap | ||||||
| 7644 | // across all iterations. | ||||||
| 7645 | const LoopAccessInfo *LAI = ILV.Legal->getLAI(); | ||||||
| 7646 | if (LAI && !LAI->getRuntimePointerChecking()->getChecks().empty() && | ||||||
| 7647 | !LAI->getRuntimePointerChecking()->getDiffChecks()) { | ||||||
| 7648 | |||||||
| 7649 | // We currently don't use LoopVersioning for the actual loop cloning but we | ||||||
| 7650 | // still use it to add the noalias metadata. | ||||||
| 7651 | // TODO: Find a better way to re-use LoopVersioning functionality to add | ||||||
| 7652 | // metadata. | ||||||
| 7653 | State.LVer = std::make_unique<LoopVersioning>( | ||||||
| 7654 | *LAI, LAI->getRuntimePointerChecking()->getChecks(), OrigLoop, LI, DT, | ||||||
| 7655 | PSE.getSE()); | ||||||
| 7656 | State.LVer->prepareNoAliasMetadata(); | ||||||
| 7657 | } | ||||||
| 7658 | |||||||
| 7659 | ILV.collectPoisonGeneratingRecipes(State); | ||||||
| 7660 | |||||||
| 7661 | ILV.printDebugTracesAtStart(); | ||||||
| 7662 | |||||||
| 7663 | //===------------------------------------------------===// | ||||||
| 7664 | // | ||||||
| 7665 | // Notice: any optimization or new instruction that go | ||||||
| 7666 | // into the code below should also be implemented in | ||||||
| 7667 | // the cost-model. | ||||||
| 7668 | // | ||||||
| 7669 | //===------------------------------------------------===// | ||||||
| 7670 | |||||||
| 7671 | // 2. Copy and widen instructions from the old loop into the new loop. | ||||||
| 7672 | BestVPlan.prepareToExecute(ILV.getOrCreateTripCount(nullptr), | ||||||
| 7673 | ILV.getOrCreateVectorTripCount(nullptr), | ||||||
| 7674 | CanonicalIVStartValue, State, | ||||||
| 7675 | IsEpilogueVectorization); | ||||||
| 7676 | |||||||
| 7677 | BestVPlan.execute(&State); | ||||||
| 7678 | |||||||
| 7679 | // Keep all loop hints from the original loop on the vector loop (we'll | ||||||
| 7680 | // replace the vectorizer-specific hints below). | ||||||
| 7681 | MDNode *OrigLoopID = OrigLoop->getLoopID(); | ||||||
| 7682 | |||||||
| 7683 | std::optional<MDNode *> VectorizedLoopID = | ||||||
| 7684 | makeFollowupLoopID(OrigLoopID, {LLVMLoopVectorizeFollowupAll, | ||||||
| 7685 | LLVMLoopVectorizeFollowupVectorized}); | ||||||
| 7686 | |||||||
| 7687 | VPBasicBlock *HeaderVPBB = | ||||||
| 7688 | BestVPlan.getVectorLoopRegion()->getEntryBasicBlock(); | ||||||
| 7689 | Loop *L = LI->getLoopFor(State.CFG.VPBB2IRBB[HeaderVPBB]); | ||||||
| 7690 | if (VectorizedLoopID) | ||||||
| 7691 | L->setLoopID(*VectorizedLoopID); | ||||||
| 7692 | else { | ||||||
| 7693 | // Keep all loop hints from the original loop on the vector loop (we'll | ||||||
| 7694 | // replace the vectorizer-specific hints below). | ||||||
| 7695 | if (MDNode *LID = OrigLoop->getLoopID()) | ||||||
| 7696 | L->setLoopID(LID); | ||||||
| 7697 | |||||||
| 7698 | LoopVectorizeHints Hints(L, true, *ORE); | ||||||
| 7699 | Hints.setAlreadyVectorized(); | ||||||
| 7700 | } | ||||||
| 7701 | // Disable runtime unrolling when vectorizing the epilogue loop. | ||||||
| 7702 | if (CanonicalIVStartValue) | ||||||
| 7703 | AddRuntimeUnrollDisableMetaData(L); | ||||||
| 7704 | |||||||
| 7705 | // 3. Fix the vectorized code: take care of header phi's, live-outs, | ||||||
| 7706 | // predication, updating analyses. | ||||||
| 7707 | ILV.fixVectorizedLoop(State, BestVPlan); | ||||||
| 7708 | |||||||
| 7709 | ILV.printDebugTracesAtEnd(); | ||||||
| 7710 | } | ||||||
| 7711 | |||||||
| 7712 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||
| 7713 | void LoopVectorizationPlanner::printPlans(raw_ostream &O) { | ||||||
| 7714 | for (const auto &Plan : VPlans) | ||||||
| 7715 | if (PrintVPlansInDotFormat) | ||||||
| 7716 | Plan->printDOT(O); | ||||||
| 7717 | else | ||||||
| 7718 | Plan->print(O); | ||||||
| 7719 | } | ||||||
| 7720 | #endif | ||||||
| 7721 | |||||||
| 7722 | Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) { return V; } | ||||||
| 7723 | |||||||
| 7724 | //===--------------------------------------------------------------------===// | ||||||
| 7725 | // EpilogueVectorizerMainLoop | ||||||
| 7726 | //===--------------------------------------------------------------------===// | ||||||
| 7727 | |||||||
| 7728 | /// This function is partially responsible for generating the control flow | ||||||
| 7729 | /// depicted in https://llvm.org/docs/Vectorizers.html#epilogue-vectorization. | ||||||
| 7730 | std::pair<BasicBlock *, Value *> | ||||||
| 7731 | EpilogueVectorizerMainLoop::createEpilogueVectorizedLoopSkeleton() { | ||||||
| 7732 | createVectorLoopSkeleton(""); | ||||||
| 7733 | |||||||
| 7734 | // Generate the code to check the minimum iteration count of the vector | ||||||
| 7735 | // epilogue (see below). | ||||||
| 7736 | EPI.EpilogueIterationCountCheck = | ||||||
| 7737 | emitIterationCountCheck(LoopScalarPreHeader, true); | ||||||
| 7738 | EPI.EpilogueIterationCountCheck->setName("iter.check"); | ||||||
| 7739 | |||||||
| 7740 | // Generate the code to check any assumptions that we've made for SCEV | ||||||
| 7741 | // expressions. | ||||||
| 7742 | EPI.SCEVSafetyCheck = emitSCEVChecks(LoopScalarPreHeader); | ||||||
| 7743 | |||||||
| 7744 | // Generate the code that checks at runtime if arrays overlap. We put the | ||||||
| 7745 | // checks into a separate block to make the more common case of few elements | ||||||
| 7746 | // faster. | ||||||
| 7747 | EPI.MemSafetyCheck = emitMemRuntimeChecks(LoopScalarPreHeader); | ||||||
| 7748 | |||||||
| 7749 | // Generate the iteration count check for the main loop, *after* the check | ||||||
| 7750 | // for the epilogue loop, so that the path-length is shorter for the case | ||||||
| 7751 | // that goes directly through the vector epilogue. The longer-path length for | ||||||
| 7752 | // the main loop is compensated for, by the gain from vectorizing the larger | ||||||
| 7753 | // trip count. Note: the branch will get updated later on when we vectorize | ||||||
| 7754 | // the epilogue. | ||||||
| 7755 | EPI.MainLoopIterationCountCheck = | ||||||
| 7756 | emitIterationCountCheck(LoopScalarPreHeader, false); | ||||||
| 7757 | |||||||
| 7758 | // Generate the induction variable. | ||||||
| 7759 | EPI.VectorTripCount = getOrCreateVectorTripCount(LoopVectorPreHeader); | ||||||
| 7760 | |||||||
| 7761 | // Skip induction resume value creation here because they will be created in | ||||||
| 7762 | // the second pass for the scalar loop. The induction resume values for the | ||||||
| 7763 | // inductions in the epilogue loop are created before executing the plan for | ||||||
| 7764 | // the epilogue loop. | ||||||
| 7765 | |||||||
| 7766 | return {completeLoopSkeleton(), nullptr}; | ||||||
| 7767 | } | ||||||
| 7768 | |||||||
| 7769 | void EpilogueVectorizerMainLoop::printDebugTracesAtStart() { | ||||||
| 7770 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n" << "Main Loop VF:" << EPI.MainLoopVF << ", Main Loop UF:" << EPI.MainLoopUF << ", Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | ||||||
| 7771 | dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n" << "Main Loop VF:" << EPI.MainLoopVF << ", Main Loop UF:" << EPI.MainLoopUF << ", Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | ||||||
| 7772 | << "Main Loop VF:" << EPI.MainLoopVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n" << "Main Loop VF:" << EPI.MainLoopVF << ", Main Loop UF:" << EPI.MainLoopUF << ", Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | ||||||
| 7773 | << ", Main Loop UF:" << EPI.MainLoopUFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n" << "Main Loop VF:" << EPI.MainLoopVF << ", Main Loop UF:" << EPI.MainLoopUF << ", Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | ||||||
| 7774 | << ", Epilogue Loop VF:" << EPI.EpilogueVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n" << "Main Loop VF:" << EPI.MainLoopVF << ", Main Loop UF:" << EPI.MainLoopUF << ", Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | ||||||
| 7775 | << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n" << "Main Loop VF:" << EPI.MainLoopVF << ", Main Loop UF:" << EPI.MainLoopUF << ", Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | ||||||
| 7776 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n" << "Main Loop VF:" << EPI.MainLoopVF << ", Main Loop UF:" << EPI.MainLoopUF << ", Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false); | ||||||
| 7777 | } | ||||||
| 7778 | |||||||
| 7779 | void EpilogueVectorizerMainLoop::printDebugTracesAtEnd() { | ||||||
| 7780 | DEBUG_WITH_TYPE(VerboseDebug, {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }; } } while (false) | ||||||
| 7781 | dbgs() << "intermediate fn:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }; } } while (false) | ||||||
| 7782 | << *OrigLoop->getHeader()->getParent() << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }; } } while (false) | ||||||
| 7783 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }; } } while (false); | ||||||
| 7784 | } | ||||||
| 7785 | |||||||
| 7786 | BasicBlock * | ||||||
| 7787 | EpilogueVectorizerMainLoop::emitIterationCountCheck(BasicBlock *Bypass, | ||||||
| 7788 | bool ForEpilogue) { | ||||||
| 7789 | assert(Bypass && "Expected valid bypass basic block.")(static_cast <bool> (Bypass && "Expected valid bypass basic block." ) ? void (0) : __assert_fail ("Bypass && \"Expected valid bypass basic block.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7789, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7790 | ElementCount VFactor = ForEpilogue ? EPI.EpilogueVF : VF; | ||||||
| 7791 | unsigned UFactor = ForEpilogue ? EPI.EpilogueUF : UF; | ||||||
| 7792 | Value *Count = getOrCreateTripCount(LoopVectorPreHeader); | ||||||
| 7793 | // Reuse existing vector loop preheader for TC checks. | ||||||
| 7794 | // Note that new preheader block is generated for vector loop. | ||||||
| 7795 | BasicBlock *const TCCheckBlock = LoopVectorPreHeader; | ||||||
| 7796 | IRBuilder<> Builder(TCCheckBlock->getTerminator()); | ||||||
| 7797 | |||||||
| 7798 | // Generate code to check if the loop's trip count is less than VF * UF of the | ||||||
| 7799 | // main vector loop. | ||||||
| 7800 | auto P = Cost->requiresScalarEpilogue(ForEpilogue ? EPI.EpilogueVF : VF) ? | ||||||
| 7801 | ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT; | ||||||
| 7802 | |||||||
| 7803 | Value *CheckMinIters = Builder.CreateICmp( | ||||||
| 7804 | P, Count, createStepForVF(Builder, Count->getType(), VFactor, UFactor), | ||||||
| 7805 | "min.iters.check"); | ||||||
| 7806 | |||||||
| 7807 | if (!ForEpilogue) | ||||||
| 7808 | TCCheckBlock->setName("vector.main.loop.iter.check"); | ||||||
| 7809 | |||||||
| 7810 | // Create new preheader for vector loop. | ||||||
| 7811 | LoopVectorPreHeader = SplitBlock(TCCheckBlock, TCCheckBlock->getTerminator(), | ||||||
| 7812 | DT, LI, nullptr, "vector.ph"); | ||||||
| 7813 | |||||||
| 7814 | if (ForEpilogue) { | ||||||
| 7815 | assert(DT->properlyDominates(DT->getNode(TCCheckBlock),(static_cast <bool> (DT->properlyDominates(DT->getNode (TCCheckBlock), DT->getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass") ? void (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7817, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7816 | DT->getNode(Bypass)->getIDom()) &&(static_cast <bool> (DT->properlyDominates(DT->getNode (TCCheckBlock), DT->getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass") ? void (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7817, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7817 | "TC check is expected to dominate Bypass")(static_cast <bool> (DT->properlyDominates(DT->getNode (TCCheckBlock), DT->getNode(Bypass)->getIDom()) && "TC check is expected to dominate Bypass") ? void (0) : __assert_fail ("DT->properlyDominates(DT->getNode(TCCheckBlock), DT->getNode(Bypass)->getIDom()) && \"TC check is expected to dominate Bypass\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7817, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7818 | |||||||
| 7819 | // Update dominator for Bypass & LoopExit. | ||||||
| 7820 | DT->changeImmediateDominator(Bypass, TCCheckBlock); | ||||||
| 7821 | if (!Cost->requiresScalarEpilogue(EPI.EpilogueVF)) | ||||||
| 7822 | // For loops with multiple exits, there's no edge from the middle block | ||||||
| 7823 | // to exit blocks (as the epilogue must run) and thus no need to update | ||||||
| 7824 | // the immediate dominator of the exit blocks. | ||||||
| 7825 | DT->changeImmediateDominator(LoopExitBlock, TCCheckBlock); | ||||||
| 7826 | |||||||
| 7827 | LoopBypassBlocks.push_back(TCCheckBlock); | ||||||
| 7828 | |||||||
| 7829 | // Save the trip count so we don't have to regenerate it in the | ||||||
| 7830 | // vec.epilog.iter.check. This is safe to do because the trip count | ||||||
| 7831 | // generated here dominates the vector epilog iter check. | ||||||
| 7832 | EPI.TripCount = Count; | ||||||
| 7833 | } | ||||||
| 7834 | |||||||
| 7835 | ReplaceInstWithInst( | ||||||
| 7836 | TCCheckBlock->getTerminator(), | ||||||
| 7837 | BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); | ||||||
| 7838 | |||||||
| 7839 | return TCCheckBlock; | ||||||
| 7840 | } | ||||||
| 7841 | |||||||
| 7842 | //===--------------------------------------------------------------------===// | ||||||
| 7843 | // EpilogueVectorizerEpilogueLoop | ||||||
| 7844 | //===--------------------------------------------------------------------===// | ||||||
| 7845 | |||||||
| 7846 | /// This function is partially responsible for generating the control flow | ||||||
| 7847 | /// depicted in https://llvm.org/docs/Vectorizers.html#epilogue-vectorization. | ||||||
| 7848 | std::pair<BasicBlock *, Value *> | ||||||
| 7849 | EpilogueVectorizerEpilogueLoop::createEpilogueVectorizedLoopSkeleton() { | ||||||
| 7850 | createVectorLoopSkeleton("vec.epilog."); | ||||||
| 7851 | |||||||
| 7852 | // Now, compare the remaining count and if there aren't enough iterations to | ||||||
| 7853 | // execute the vectorized epilogue skip to the scalar part. | ||||||
| 7854 | BasicBlock *VecEpilogueIterationCountCheck = LoopVectorPreHeader; | ||||||
| 7855 | VecEpilogueIterationCountCheck->setName("vec.epilog.iter.check"); | ||||||
| 7856 | LoopVectorPreHeader = | ||||||
| 7857 | SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, | ||||||
| 7858 | LI, nullptr, "vec.epilog.ph"); | ||||||
| 7859 | emitMinimumVectorEpilogueIterCountCheck(LoopScalarPreHeader, | ||||||
| 7860 | VecEpilogueIterationCountCheck); | ||||||
| 7861 | |||||||
| 7862 | // Adjust the control flow taking the state info from the main loop | ||||||
| 7863 | // vectorization into account. | ||||||
| 7864 | assert(EPI.MainLoopIterationCountCheck && EPI.EpilogueIterationCountCheck &&(static_cast <bool> (EPI.MainLoopIterationCountCheck && EPI.EpilogueIterationCountCheck && "expected this to be saved from the previous pass." ) ? void (0) : __assert_fail ("EPI.MainLoopIterationCountCheck && EPI.EpilogueIterationCountCheck && \"expected this to be saved from the previous pass.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7865, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7865 | "expected this to be saved from the previous pass.")(static_cast <bool> (EPI.MainLoopIterationCountCheck && EPI.EpilogueIterationCountCheck && "expected this to be saved from the previous pass." ) ? void (0) : __assert_fail ("EPI.MainLoopIterationCountCheck && EPI.EpilogueIterationCountCheck && \"expected this to be saved from the previous pass.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7865, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7866 | EPI.MainLoopIterationCountCheck->getTerminator()->replaceUsesOfWith( | ||||||
| 7867 | VecEpilogueIterationCountCheck, LoopVectorPreHeader); | ||||||
| 7868 | |||||||
| 7869 | DT->changeImmediateDominator(LoopVectorPreHeader, | ||||||
| 7870 | EPI.MainLoopIterationCountCheck); | ||||||
| 7871 | |||||||
| 7872 | EPI.EpilogueIterationCountCheck->getTerminator()->replaceUsesOfWith( | ||||||
| 7873 | VecEpilogueIterationCountCheck, LoopScalarPreHeader); | ||||||
| 7874 | |||||||
| 7875 | if (EPI.SCEVSafetyCheck) | ||||||
| 7876 | EPI.SCEVSafetyCheck->getTerminator()->replaceUsesOfWith( | ||||||
| 7877 | VecEpilogueIterationCountCheck, LoopScalarPreHeader); | ||||||
| 7878 | if (EPI.MemSafetyCheck) | ||||||
| 7879 | EPI.MemSafetyCheck->getTerminator()->replaceUsesOfWith( | ||||||
| 7880 | VecEpilogueIterationCountCheck, LoopScalarPreHeader); | ||||||
| 7881 | |||||||
| 7882 | DT->changeImmediateDominator( | ||||||
| 7883 | VecEpilogueIterationCountCheck, | ||||||
| 7884 | VecEpilogueIterationCountCheck->getSinglePredecessor()); | ||||||
| 7885 | |||||||
| 7886 | DT->changeImmediateDominator(LoopScalarPreHeader, | ||||||
| 7887 | EPI.EpilogueIterationCountCheck); | ||||||
| 7888 | if (!Cost->requiresScalarEpilogue(EPI.EpilogueVF)) | ||||||
| 7889 | // If there is an epilogue which must run, there's no edge from the | ||||||
| 7890 | // middle block to exit blocks and thus no need to update the immediate | ||||||
| 7891 | // dominator of the exit blocks. | ||||||
| 7892 | DT->changeImmediateDominator(LoopExitBlock, | ||||||
| 7893 | EPI.EpilogueIterationCountCheck); | ||||||
| 7894 | |||||||
| 7895 | // Keep track of bypass blocks, as they feed start values to the induction and | ||||||
| 7896 | // reduction phis in the scalar loop preheader. | ||||||
| 7897 | if (EPI.SCEVSafetyCheck) | ||||||
| 7898 | LoopBypassBlocks.push_back(EPI.SCEVSafetyCheck); | ||||||
| 7899 | if (EPI.MemSafetyCheck) | ||||||
| 7900 | LoopBypassBlocks.push_back(EPI.MemSafetyCheck); | ||||||
| 7901 | LoopBypassBlocks.push_back(EPI.EpilogueIterationCountCheck); | ||||||
| 7902 | |||||||
| 7903 | // The vec.epilog.iter.check block may contain Phi nodes from inductions or | ||||||
| 7904 | // reductions which merge control-flow from the latch block and the middle | ||||||
| 7905 | // block. Update the incoming values here and move the Phi into the preheader. | ||||||
| 7906 | SmallVector<PHINode *, 4> PhisInBlock; | ||||||
| 7907 | for (PHINode &Phi : VecEpilogueIterationCountCheck->phis()) | ||||||
| 7908 | PhisInBlock.push_back(&Phi); | ||||||
| 7909 | |||||||
| 7910 | for (PHINode *Phi : PhisInBlock) { | ||||||
| 7911 | Phi->moveBefore(LoopVectorPreHeader->getFirstNonPHI()); | ||||||
| 7912 | Phi->replaceIncomingBlockWith( | ||||||
| 7913 | VecEpilogueIterationCountCheck->getSinglePredecessor(), | ||||||
| 7914 | VecEpilogueIterationCountCheck); | ||||||
| 7915 | |||||||
| 7916 | // If the phi doesn't have an incoming value from the | ||||||
| 7917 | // EpilogueIterationCountCheck, we are done. Otherwise remove the incoming | ||||||
| 7918 | // value and also those from other check blocks. This is needed for | ||||||
| 7919 | // reduction phis only. | ||||||
| 7920 | if (none_of(Phi->blocks(), [&](BasicBlock *IncB) { | ||||||
| 7921 | return EPI.EpilogueIterationCountCheck == IncB; | ||||||
| 7922 | })) | ||||||
| 7923 | continue; | ||||||
| 7924 | Phi->removeIncomingValue(EPI.EpilogueIterationCountCheck); | ||||||
| 7925 | if (EPI.SCEVSafetyCheck) | ||||||
| 7926 | Phi->removeIncomingValue(EPI.SCEVSafetyCheck); | ||||||
| 7927 | if (EPI.MemSafetyCheck) | ||||||
| 7928 | Phi->removeIncomingValue(EPI.MemSafetyCheck); | ||||||
| 7929 | } | ||||||
| 7930 | |||||||
| 7931 | // Generate a resume induction for the vector epilogue and put it in the | ||||||
| 7932 | // vector epilogue preheader | ||||||
| 7933 | Type *IdxTy = Legal->getWidestInductionType(); | ||||||
| 7934 | PHINode *EPResumeVal = PHINode::Create(IdxTy, 2, "vec.epilog.resume.val", | ||||||
| 7935 | LoopVectorPreHeader->getFirstNonPHI()); | ||||||
| 7936 | EPResumeVal->addIncoming(EPI.VectorTripCount, VecEpilogueIterationCountCheck); | ||||||
| 7937 | EPResumeVal->addIncoming(ConstantInt::get(IdxTy, 0), | ||||||
| 7938 | EPI.MainLoopIterationCountCheck); | ||||||
| 7939 | |||||||
| 7940 | // Generate induction resume values. These variables save the new starting | ||||||
| 7941 | // indexes for the scalar loop. They are used to test if there are any tail | ||||||
| 7942 | // iterations left once the vector loop has completed. | ||||||
| 7943 | // Note that when the vectorized epilogue is skipped due to iteration count | ||||||
| 7944 | // check, then the resume value for the induction variable comes from | ||||||
| 7945 | // the trip count of the main vector loop, hence passing the AdditionalBypass | ||||||
| 7946 | // argument. | ||||||
| 7947 | createInductionResumeValues({VecEpilogueIterationCountCheck, | ||||||
| 7948 | EPI.VectorTripCount} /* AdditionalBypass */); | ||||||
| 7949 | |||||||
| 7950 | return {completeLoopSkeleton(), EPResumeVal}; | ||||||
| 7951 | } | ||||||
| 7952 | |||||||
| 7953 | BasicBlock * | ||||||
| 7954 | EpilogueVectorizerEpilogueLoop::emitMinimumVectorEpilogueIterCountCheck( | ||||||
| 7955 | BasicBlock *Bypass, BasicBlock *Insert) { | ||||||
| 7956 | |||||||
| 7957 | assert(EPI.TripCount &&(static_cast <bool> (EPI.TripCount && "Expected trip count to have been safed in the first pass." ) ? void (0) : __assert_fail ("EPI.TripCount && \"Expected trip count to have been safed in the first pass.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7958, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7958 | "Expected trip count to have been safed in the first pass.")(static_cast <bool> (EPI.TripCount && "Expected trip count to have been safed in the first pass." ) ? void (0) : __assert_fail ("EPI.TripCount && \"Expected trip count to have been safed in the first pass.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7958, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7959 | assert((static_cast <bool> ((!isa<Instruction>(EPI.TripCount ) || DT->dominates(cast<Instruction>(EPI.TripCount)-> getParent(), Insert)) && "saved trip count does not dominate insertion point." ) ? void (0) : __assert_fail ("(!isa<Instruction>(EPI.TripCount) || DT->dominates(cast<Instruction>(EPI.TripCount)->getParent(), Insert)) && \"saved trip count does not dominate insertion point.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7962, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7960 | (!isa<Instruction>(EPI.TripCount) ||(static_cast <bool> ((!isa<Instruction>(EPI.TripCount ) || DT->dominates(cast<Instruction>(EPI.TripCount)-> getParent(), Insert)) && "saved trip count does not dominate insertion point." ) ? void (0) : __assert_fail ("(!isa<Instruction>(EPI.TripCount) || DT->dominates(cast<Instruction>(EPI.TripCount)->getParent(), Insert)) && \"saved trip count does not dominate insertion point.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7962, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7961 | DT->dominates(cast<Instruction>(EPI.TripCount)->getParent(), Insert)) &&(static_cast <bool> ((!isa<Instruction>(EPI.TripCount ) || DT->dominates(cast<Instruction>(EPI.TripCount)-> getParent(), Insert)) && "saved trip count does not dominate insertion point." ) ? void (0) : __assert_fail ("(!isa<Instruction>(EPI.TripCount) || DT->dominates(cast<Instruction>(EPI.TripCount)->getParent(), Insert)) && \"saved trip count does not dominate insertion point.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7962, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 7962 | "saved trip count does not dominate insertion point.")(static_cast <bool> ((!isa<Instruction>(EPI.TripCount ) || DT->dominates(cast<Instruction>(EPI.TripCount)-> getParent(), Insert)) && "saved trip count does not dominate insertion point." ) ? void (0) : __assert_fail ("(!isa<Instruction>(EPI.TripCount) || DT->dominates(cast<Instruction>(EPI.TripCount)->getParent(), Insert)) && \"saved trip count does not dominate insertion point.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7962, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 7963 | Value *TC = EPI.TripCount; | ||||||
| 7964 | IRBuilder<> Builder(Insert->getTerminator()); | ||||||
| 7965 | Value *Count = Builder.CreateSub(TC, EPI.VectorTripCount, "n.vec.remaining"); | ||||||
| 7966 | |||||||
| 7967 | // Generate code to check if the loop's trip count is less than VF * UF of the | ||||||
| 7968 | // vector epilogue loop. | ||||||
| 7969 | auto P = Cost->requiresScalarEpilogue(EPI.EpilogueVF) ? | ||||||
| 7970 | ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT; | ||||||
| 7971 | |||||||
| 7972 | Value *CheckMinIters = | ||||||
| 7973 | Builder.CreateICmp(P, Count, | ||||||
| 7974 | createStepForVF(Builder, Count->getType(), | ||||||
| 7975 | EPI.EpilogueVF, EPI.EpilogueUF), | ||||||
| 7976 | "min.epilog.iters.check"); | ||||||
| 7977 | |||||||
| 7978 | ReplaceInstWithInst( | ||||||
| 7979 | Insert->getTerminator(), | ||||||
| 7980 | BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); | ||||||
| 7981 | |||||||
| 7982 | LoopBypassBlocks.push_back(Insert); | ||||||
| 7983 | return Insert; | ||||||
| 7984 | } | ||||||
| 7985 | |||||||
| 7986 | void EpilogueVectorizerEpilogueLoop::printDebugTracesAtStart() { | ||||||
| 7987 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (second pass)\n" << "Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | ||||||
| 7988 | dbgs() << "Create Skeleton for epilogue vectorized loop (second pass)\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (second pass)\n" << "Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | ||||||
| 7989 | << "Epilogue Loop VF:" << EPI.EpilogueVFdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (second pass)\n" << "Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | ||||||
| 7990 | << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (second pass)\n" << "Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false) | ||||||
| 7991 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { { dbgs() << "Create Skeleton for epilogue vectorized loop (second pass)\n" << "Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }; } } while (false); | ||||||
| 7992 | } | ||||||
| 7993 | |||||||
| 7994 | void EpilogueVectorizerEpilogueLoop::printDebugTracesAtEnd() { | ||||||
| 7995 | DEBUG_WITH_TYPE(VerboseDebug, {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "final fn:\n" << *OrigLoop ->getHeader()->getParent() << "\n"; }; } } while ( false) | ||||||
| 7996 | dbgs() << "final fn:\n" << *OrigLoop->getHeader()->getParent() << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "final fn:\n" << *OrigLoop ->getHeader()->getParent() << "\n"; }; } } while ( false) | ||||||
| 7997 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "final fn:\n" << *OrigLoop ->getHeader()->getParent() << "\n"; }; } } while ( false); | ||||||
| 7998 | } | ||||||
| 7999 | |||||||
| 8000 | bool LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||
| 8001 | const std::function<bool(ElementCount)> &Predicate, VFRange &Range) { | ||||||
| 8002 | assert(!Range.isEmpty() && "Trying to test an empty VF range.")(static_cast <bool> (!Range.isEmpty() && "Trying to test an empty VF range." ) ? void (0) : __assert_fail ("!Range.isEmpty() && \"Trying to test an empty VF range.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8002, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8003 | bool PredicateAtRangeStart = Predicate(Range.Start); | ||||||
| 8004 | |||||||
| 8005 | for (ElementCount TmpVF = Range.Start * 2; | ||||||
| 8006 | ElementCount::isKnownLT(TmpVF, Range.End); TmpVF *= 2) | ||||||
| 8007 | if (Predicate(TmpVF) != PredicateAtRangeStart) { | ||||||
| 8008 | Range.End = TmpVF; | ||||||
| 8009 | break; | ||||||
| 8010 | } | ||||||
| 8011 | |||||||
| 8012 | return PredicateAtRangeStart; | ||||||
| 8013 | } | ||||||
| 8014 | |||||||
| 8015 | /// Build VPlans for the full range of feasible VF's = {\p MinVF, 2 * \p MinVF, | ||||||
| 8016 | /// 4 * \p MinVF, ..., \p MaxVF} by repeatedly building a VPlan for a sub-range | ||||||
| 8017 | /// of VF's starting at a given VF and extending it as much as possible. Each | ||||||
| 8018 | /// vectorization decision can potentially shorten this sub-range during | ||||||
| 8019 | /// buildVPlan(). | ||||||
| 8020 | void LoopVectorizationPlanner::buildVPlans(ElementCount MinVF, | ||||||
| 8021 | ElementCount MaxVF) { | ||||||
| 8022 | auto MaxVFPlusOne = MaxVF.getWithIncrement(1); | ||||||
| 8023 | for (ElementCount VF = MinVF; ElementCount::isKnownLT(VF, MaxVFPlusOne);) { | ||||||
| 8024 | VFRange SubRange = {VF, MaxVFPlusOne}; | ||||||
| 8025 | VPlans.push_back(buildVPlan(SubRange)); | ||||||
| 8026 | VF = SubRange.End; | ||||||
| 8027 | } | ||||||
| 8028 | } | ||||||
| 8029 | |||||||
| 8030 | VPValue *VPRecipeBuilder::createEdgeMask(BasicBlock *Src, BasicBlock *Dst, | ||||||
| 8031 | VPlanPtr &Plan) { | ||||||
| 8032 | assert(is_contained(predecessors(Dst), Src) && "Invalid edge")(static_cast <bool> (is_contained(predecessors(Dst), Src ) && "Invalid edge") ? void (0) : __assert_fail ("is_contained(predecessors(Dst), Src) && \"Invalid edge\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8032, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8033 | |||||||
| 8034 | // Look for cached value. | ||||||
| 8035 | std::pair<BasicBlock *, BasicBlock *> Edge(Src, Dst); | ||||||
| 8036 | EdgeMaskCacheTy::iterator ECEntryIt = EdgeMaskCache.find(Edge); | ||||||
| 8037 | if (ECEntryIt != EdgeMaskCache.end()) | ||||||
| 8038 | return ECEntryIt->second; | ||||||
| 8039 | |||||||
| 8040 | VPValue *SrcMask = createBlockInMask(Src, Plan); | ||||||
| 8041 | |||||||
| 8042 | // The terminator has to be a branch inst! | ||||||
| 8043 | BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator()); | ||||||
| 8044 | assert(BI && "Unexpected terminator found")(static_cast <bool> (BI && "Unexpected terminator found" ) ? void (0) : __assert_fail ("BI && \"Unexpected terminator found\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8044, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8045 | |||||||
| 8046 | if (!BI->isConditional() || BI->getSuccessor(0) == BI->getSuccessor(1)) | ||||||
| 8047 | return EdgeMaskCache[Edge] = SrcMask; | ||||||
| 8048 | |||||||
| 8049 | // If source is an exiting block, we know the exit edge is dynamically dead | ||||||
| 8050 | // in the vector loop, and thus we don't need to restrict the mask. Avoid | ||||||
| 8051 | // adding uses of an otherwise potentially dead instruction. | ||||||
| 8052 | if (OrigLoop->isLoopExiting(Src)) | ||||||
| 8053 | return EdgeMaskCache[Edge] = SrcMask; | ||||||
| 8054 | |||||||
| 8055 | VPValue *EdgeMask = Plan->getOrAddVPValue(BI->getCondition()); | ||||||
| 8056 | assert(EdgeMask && "No Edge Mask found for condition")(static_cast <bool> (EdgeMask && "No Edge Mask found for condition" ) ? void (0) : __assert_fail ("EdgeMask && \"No Edge Mask found for condition\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8056, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8057 | |||||||
| 8058 | if (BI->getSuccessor(0) != Dst) | ||||||
| 8059 | EdgeMask = Builder.createNot(EdgeMask, BI->getDebugLoc()); | ||||||
| 8060 | |||||||
| 8061 | if (SrcMask) { // Otherwise block in-mask is all-one, no need to AND. | ||||||
| 8062 | // The condition is 'SrcMask && EdgeMask', which is equivalent to | ||||||
| 8063 | // 'select i1 SrcMask, i1 EdgeMask, i1 false'. | ||||||
| 8064 | // The select version does not introduce new UB if SrcMask is false and | ||||||
| 8065 | // EdgeMask is poison. Using 'and' here introduces undefined behavior. | ||||||
| 8066 | VPValue *False = Plan->getOrAddVPValue( | ||||||
| 8067 | ConstantInt::getFalse(BI->getCondition()->getType())); | ||||||
| 8068 | EdgeMask = | ||||||
| 8069 | Builder.createSelect(SrcMask, EdgeMask, False, BI->getDebugLoc()); | ||||||
| 8070 | } | ||||||
| 8071 | |||||||
| 8072 | return EdgeMaskCache[Edge] = EdgeMask; | ||||||
| 8073 | } | ||||||
| 8074 | |||||||
| 8075 | VPValue *VPRecipeBuilder::createBlockInMask(BasicBlock *BB, VPlanPtr &Plan) { | ||||||
| 8076 | assert(OrigLoop->contains(BB) && "Block is not a part of a loop")(static_cast <bool> (OrigLoop->contains(BB) && "Block is not a part of a loop") ? void (0) : __assert_fail ( "OrigLoop->contains(BB) && \"Block is not a part of a loop\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8076, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8077 | |||||||
| 8078 | // Look for cached value. | ||||||
| 8079 | BlockMaskCacheTy::iterator BCEntryIt = BlockMaskCache.find(BB); | ||||||
| 8080 | if (BCEntryIt != BlockMaskCache.end()) | ||||||
| 8081 | return BCEntryIt->second; | ||||||
| 8082 | |||||||
| 8083 | // All-one mask is modelled as no-mask following the convention for masked | ||||||
| 8084 | // load/store/gather/scatter. Initialize BlockMask to no-mask. | ||||||
| 8085 | VPValue *BlockMask = nullptr; | ||||||
| 8086 | |||||||
| 8087 | if (OrigLoop->getHeader() == BB) { | ||||||
| 8088 | if (!CM.blockNeedsPredicationForAnyReason(BB)) | ||||||
| 8089 | return BlockMaskCache[BB] = BlockMask; // Loop incoming mask is all-one. | ||||||
| 8090 | |||||||
| 8091 | assert(CM.foldTailByMasking() && "must fold the tail")(static_cast <bool> (CM.foldTailByMasking() && "must fold the tail" ) ? void (0) : __assert_fail ("CM.foldTailByMasking() && \"must fold the tail\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8091, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8092 | |||||||
| 8093 | // If we're using the active lane mask for control flow, then we get the | ||||||
| 8094 | // mask from the active lane mask PHI that is cached in the VPlan. | ||||||
| 8095 | PredicationStyle EmitGetActiveLaneMask = CM.TTI.emitGetActiveLaneMask(); | ||||||
| 8096 | if (EmitGetActiveLaneMask == PredicationStyle::DataAndControlFlow) | ||||||
| 8097 | return BlockMaskCache[BB] = Plan->getActiveLaneMaskPhi(); | ||||||
| 8098 | |||||||
| 8099 | // Introduce the early-exit compare IV <= BTC to form header block mask. | ||||||
| 8100 | // This is used instead of IV < TC because TC may wrap, unlike BTC. Start by | ||||||
| 8101 | // constructing the desired canonical IV in the header block as its first | ||||||
| 8102 | // non-phi instructions. | ||||||
| 8103 | |||||||
| 8104 | VPBasicBlock *HeaderVPBB = | ||||||
| 8105 | Plan->getVectorLoopRegion()->getEntryBasicBlock(); | ||||||
| 8106 | auto NewInsertionPoint = HeaderVPBB->getFirstNonPhi(); | ||||||
| 8107 | auto *IV = new VPWidenCanonicalIVRecipe(Plan->getCanonicalIV()); | ||||||
| 8108 | HeaderVPBB->insert(IV, HeaderVPBB->getFirstNonPhi()); | ||||||
| 8109 | |||||||
| 8110 | VPBuilder::InsertPointGuard Guard(Builder); | ||||||
| 8111 | Builder.setInsertPoint(HeaderVPBB, NewInsertionPoint); | ||||||
| 8112 | if (EmitGetActiveLaneMask != PredicationStyle::None) { | ||||||
| 8113 | VPValue *TC = Plan->getOrCreateTripCount(); | ||||||
| 8114 | BlockMask = Builder.createNaryOp(VPInstruction::ActiveLaneMask, {IV, TC}, | ||||||
| 8115 | nullptr, "active.lane.mask"); | ||||||
| 8116 | } else { | ||||||
| 8117 | VPValue *BTC = Plan->getOrCreateBackedgeTakenCount(); | ||||||
| 8118 | BlockMask = Builder.createNaryOp(VPInstruction::ICmpULE, {IV, BTC}); | ||||||
| 8119 | } | ||||||
| 8120 | return BlockMaskCache[BB] = BlockMask; | ||||||
| 8121 | } | ||||||
| 8122 | |||||||
| 8123 | // This is the block mask. We OR all incoming edges. | ||||||
| 8124 | for (auto *Predecessor : predecessors(BB)) { | ||||||
| 8125 | VPValue *EdgeMask = createEdgeMask(Predecessor, BB, Plan); | ||||||
| 8126 | if (!EdgeMask) // Mask of predecessor is all-one so mask of block is too. | ||||||
| 8127 | return BlockMaskCache[BB] = EdgeMask; | ||||||
| 8128 | |||||||
| 8129 | if (!BlockMask) { // BlockMask has its initialized nullptr value. | ||||||
| 8130 | BlockMask = EdgeMask; | ||||||
| 8131 | continue; | ||||||
| 8132 | } | ||||||
| 8133 | |||||||
| 8134 | BlockMask = Builder.createOr(BlockMask, EdgeMask, {}); | ||||||
| 8135 | } | ||||||
| 8136 | |||||||
| 8137 | return BlockMaskCache[BB] = BlockMask; | ||||||
| 8138 | } | ||||||
| 8139 | |||||||
| 8140 | VPRecipeBase *VPRecipeBuilder::tryToWidenMemory(Instruction *I, | ||||||
| 8141 | ArrayRef<VPValue *> Operands, | ||||||
| 8142 | VFRange &Range, | ||||||
| 8143 | VPlanPtr &Plan) { | ||||||
| 8144 | assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&(static_cast <bool> ((isa<LoadInst>(I) || isa< StoreInst>(I)) && "Must be called with either a load or store" ) ? void (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Must be called with either a load or store\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8145, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8145 | "Must be called with either a load or store")(static_cast <bool> ((isa<LoadInst>(I) || isa< StoreInst>(I)) && "Must be called with either a load or store" ) ? void (0) : __assert_fail ("(isa<LoadInst>(I) || isa<StoreInst>(I)) && \"Must be called with either a load or store\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8145, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8146 | |||||||
| 8147 | auto willWiden = [&](ElementCount VF) -> bool { | ||||||
| 8148 | LoopVectorizationCostModel::InstWidening Decision = | ||||||
| 8149 | CM.getWideningDecision(I, VF); | ||||||
| 8150 | assert(Decision != LoopVectorizationCostModel::CM_Unknown &&(static_cast <bool> (Decision != LoopVectorizationCostModel ::CM_Unknown && "CM decision should be taken at this point." ) ? void (0) : __assert_fail ("Decision != LoopVectorizationCostModel::CM_Unknown && \"CM decision should be taken at this point.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8151, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8151 | "CM decision should be taken at this point.")(static_cast <bool> (Decision != LoopVectorizationCostModel ::CM_Unknown && "CM decision should be taken at this point." ) ? void (0) : __assert_fail ("Decision != LoopVectorizationCostModel::CM_Unknown && \"CM decision should be taken at this point.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8151, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8152 | if (Decision == LoopVectorizationCostModel::CM_Interleave) | ||||||
| 8153 | return true; | ||||||
| 8154 | if (CM.isScalarAfterVectorization(I, VF) || | ||||||
| 8155 | CM.isProfitableToScalarize(I, VF)) | ||||||
| 8156 | return false; | ||||||
| 8157 | return Decision != LoopVectorizationCostModel::CM_Scalarize; | ||||||
| 8158 | }; | ||||||
| 8159 | |||||||
| 8160 | if (!LoopVectorizationPlanner::getDecisionAndClampRange(willWiden, Range)) | ||||||
| 8161 | return nullptr; | ||||||
| 8162 | |||||||
| 8163 | VPValue *Mask = nullptr; | ||||||
| 8164 | if (Legal->isMaskRequired(I)) | ||||||
| 8165 | Mask = createBlockInMask(I->getParent(), Plan); | ||||||
| 8166 | |||||||
| 8167 | // Determine if the pointer operand of the access is either consecutive or | ||||||
| 8168 | // reverse consecutive. | ||||||
| 8169 | LoopVectorizationCostModel::InstWidening Decision = | ||||||
| 8170 | CM.getWideningDecision(I, Range.Start); | ||||||
| 8171 | bool Reverse = Decision == LoopVectorizationCostModel::CM_Widen_Reverse; | ||||||
| 8172 | bool Consecutive = | ||||||
| 8173 | Reverse || Decision == LoopVectorizationCostModel::CM_Widen; | ||||||
| 8174 | |||||||
| 8175 | if (LoadInst *Load = dyn_cast<LoadInst>(I)) | ||||||
| 8176 | return new VPWidenMemoryInstructionRecipe(*Load, Operands[0], Mask, | ||||||
| 8177 | Consecutive, Reverse); | ||||||
| 8178 | |||||||
| 8179 | StoreInst *Store = cast<StoreInst>(I); | ||||||
| 8180 | return new VPWidenMemoryInstructionRecipe(*Store, Operands[1], Operands[0], | ||||||
| 8181 | Mask, Consecutive, Reverse); | ||||||
| 8182 | } | ||||||
| 8183 | |||||||
| 8184 | /// Creates a VPWidenIntOrFpInductionRecpipe for \p Phi. If needed, it will also | ||||||
| 8185 | /// insert a recipe to expand the step for the induction recipe. | ||||||
| 8186 | static VPWidenIntOrFpInductionRecipe *createWidenInductionRecipes( | ||||||
| 8187 | PHINode *Phi, Instruction *PhiOrTrunc, VPValue *Start, | ||||||
| 8188 | const InductionDescriptor &IndDesc, LoopVectorizationCostModel &CM, | ||||||
| 8189 | VPlan &Plan, ScalarEvolution &SE, Loop &OrigLoop, VFRange &Range) { | ||||||
| 8190 | // Returns true if an instruction \p I should be scalarized instead of | ||||||
| 8191 | // vectorized for the chosen vectorization factor. | ||||||
| 8192 | auto ShouldScalarizeInstruction = [&CM](Instruction *I, ElementCount VF) { | ||||||
| 8193 | return CM.isScalarAfterVectorization(I, VF) || | ||||||
| 8194 | CM.isProfitableToScalarize(I, VF); | ||||||
| 8195 | }; | ||||||
| 8196 | |||||||
| 8197 | bool NeedsScalarIVOnly = LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||
| 8198 | [&](ElementCount VF) { | ||||||
| 8199 | return ShouldScalarizeInstruction(PhiOrTrunc, VF); | ||||||
| 8200 | }, | ||||||
| 8201 | Range); | ||||||
| 8202 | assert(IndDesc.getStartValue() ==(static_cast <bool> (IndDesc.getStartValue() == Phi-> getIncomingValueForBlock(OrigLoop.getLoopPreheader())) ? void (0) : __assert_fail ("IndDesc.getStartValue() == Phi->getIncomingValueForBlock(OrigLoop.getLoopPreheader())" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8203, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8203 | Phi->getIncomingValueForBlock(OrigLoop.getLoopPreheader()))(static_cast <bool> (IndDesc.getStartValue() == Phi-> getIncomingValueForBlock(OrigLoop.getLoopPreheader())) ? void (0) : __assert_fail ("IndDesc.getStartValue() == Phi->getIncomingValueForBlock(OrigLoop.getLoopPreheader())" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8203, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8204 | assert(SE.isLoopInvariant(IndDesc.getStep(), &OrigLoop) &&(static_cast <bool> (SE.isLoopInvariant(IndDesc.getStep (), &OrigLoop) && "step must be loop invariant") ? void (0) : __assert_fail ("SE.isLoopInvariant(IndDesc.getStep(), &OrigLoop) && \"step must be loop invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8205, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8205 | "step must be loop invariant")(static_cast <bool> (SE.isLoopInvariant(IndDesc.getStep (), &OrigLoop) && "step must be loop invariant") ? void (0) : __assert_fail ("SE.isLoopInvariant(IndDesc.getStep(), &OrigLoop) && \"step must be loop invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8205, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8206 | |||||||
| 8207 | VPValue *Step = | ||||||
| 8208 | vputils::getOrCreateVPValueForSCEVExpr(Plan, IndDesc.getStep(), SE); | ||||||
| 8209 | if (auto *TruncI = dyn_cast<TruncInst>(PhiOrTrunc)) { | ||||||
| 8210 | return new VPWidenIntOrFpInductionRecipe(Phi, Start, Step, IndDesc, TruncI, | ||||||
| 8211 | !NeedsScalarIVOnly); | ||||||
| 8212 | } | ||||||
| 8213 | assert(isa<PHINode>(PhiOrTrunc) && "must be a phi node here")(static_cast <bool> (isa<PHINode>(PhiOrTrunc) && "must be a phi node here") ? void (0) : __assert_fail ("isa<PHINode>(PhiOrTrunc) && \"must be a phi node here\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8213, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8214 | return new VPWidenIntOrFpInductionRecipe(Phi, Start, Step, IndDesc, | ||||||
| 8215 | !NeedsScalarIVOnly); | ||||||
| 8216 | } | ||||||
| 8217 | |||||||
| 8218 | VPRecipeBase *VPRecipeBuilder::tryToOptimizeInductionPHI( | ||||||
| 8219 | PHINode *Phi, ArrayRef<VPValue *> Operands, VPlan &Plan, VFRange &Range) { | ||||||
| 8220 | |||||||
| 8221 | // Check if this is an integer or fp induction. If so, build the recipe that | ||||||
| 8222 | // produces its scalar and vector values. | ||||||
| 8223 | if (auto *II = Legal->getIntOrFpInductionDescriptor(Phi)) | ||||||
| 8224 | return createWidenInductionRecipes(Phi, Phi, Operands[0], *II, CM, Plan, | ||||||
| 8225 | *PSE.getSE(), *OrigLoop, Range); | ||||||
| 8226 | |||||||
| 8227 | // Check if this is pointer induction. If so, build the recipe for it. | ||||||
| 8228 | if (auto *II = Legal->getPointerInductionDescriptor(Phi)) { | ||||||
| 8229 | VPValue *Step = vputils::getOrCreateVPValueForSCEVExpr(Plan, II->getStep(), | ||||||
| 8230 | *PSE.getSE()); | ||||||
| 8231 | assert(isa<SCEVConstant>(II->getStep()))(static_cast <bool> (isa<SCEVConstant>(II->getStep ())) ? void (0) : __assert_fail ("isa<SCEVConstant>(II->getStep())" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8231, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8232 | return new VPWidenPointerInductionRecipe( | ||||||
| 8233 | Phi, Operands[0], Step, *II, | ||||||
| 8234 | LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||
| 8235 | [&](ElementCount VF) { | ||||||
| 8236 | return CM.isScalarAfterVectorization(Phi, VF); | ||||||
| 8237 | }, | ||||||
| 8238 | Range)); | ||||||
| 8239 | } | ||||||
| 8240 | return nullptr; | ||||||
| 8241 | } | ||||||
| 8242 | |||||||
| 8243 | VPWidenIntOrFpInductionRecipe *VPRecipeBuilder::tryToOptimizeInductionTruncate( | ||||||
| 8244 | TruncInst *I, ArrayRef<VPValue *> Operands, VFRange &Range, VPlan &Plan) { | ||||||
| 8245 | // Optimize the special case where the source is a constant integer | ||||||
| 8246 | // induction variable. Notice that we can only optimize the 'trunc' case | ||||||
| 8247 | // because (a) FP conversions lose precision, (b) sext/zext may wrap, and | ||||||
| 8248 | // (c) other casts depend on pointer size. | ||||||
| 8249 | |||||||
| 8250 | // Determine whether \p K is a truncation based on an induction variable that | ||||||
| 8251 | // can be optimized. | ||||||
| 8252 | auto isOptimizableIVTruncate = | ||||||
| 8253 | [&](Instruction *K) -> std::function<bool(ElementCount)> { | ||||||
| 8254 | return [=](ElementCount VF) -> bool { | ||||||
| 8255 | return CM.isOptimizableIVTruncate(K, VF); | ||||||
| 8256 | }; | ||||||
| 8257 | }; | ||||||
| 8258 | |||||||
| 8259 | if (LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||
| 8260 | isOptimizableIVTruncate(I), Range)) { | ||||||
| 8261 | |||||||
| 8262 | auto *Phi = cast<PHINode>(I->getOperand(0)); | ||||||
| 8263 | const InductionDescriptor &II = *Legal->getIntOrFpInductionDescriptor(Phi); | ||||||
| 8264 | VPValue *Start = Plan.getOrAddVPValue(II.getStartValue()); | ||||||
| 8265 | return createWidenInductionRecipes(Phi, I, Start, II, CM, Plan, | ||||||
| 8266 | *PSE.getSE(), *OrigLoop, Range); | ||||||
| 8267 | } | ||||||
| 8268 | return nullptr; | ||||||
| 8269 | } | ||||||
| 8270 | |||||||
| 8271 | VPRecipeOrVPValueTy VPRecipeBuilder::tryToBlend(PHINode *Phi, | ||||||
| 8272 | ArrayRef<VPValue *> Operands, | ||||||
| 8273 | VPlanPtr &Plan) { | ||||||
| 8274 | // If all incoming values are equal, the incoming VPValue can be used directly | ||||||
| 8275 | // instead of creating a new VPBlendRecipe. | ||||||
| 8276 | if (llvm::all_equal(Operands)) | ||||||
| 8277 | return Operands[0]; | ||||||
| 8278 | |||||||
| 8279 | unsigned NumIncoming = Phi->getNumIncomingValues(); | ||||||
| 8280 | // For in-loop reductions, we do not need to create an additional select. | ||||||
| 8281 | VPValue *InLoopVal = nullptr; | ||||||
| 8282 | for (unsigned In = 0; In < NumIncoming; In++) { | ||||||
| 8283 | PHINode *PhiOp = | ||||||
| 8284 | dyn_cast_or_null<PHINode>(Operands[In]->getUnderlyingValue()); | ||||||
| 8285 | if (PhiOp && CM.isInLoopReduction(PhiOp)) { | ||||||
| 8286 | assert(!InLoopVal && "Found more than one in-loop reduction!")(static_cast <bool> (!InLoopVal && "Found more than one in-loop reduction!" ) ? void (0) : __assert_fail ("!InLoopVal && \"Found more than one in-loop reduction!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8286, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8287 | InLoopVal = Operands[In]; | ||||||
| 8288 | } | ||||||
| 8289 | } | ||||||
| 8290 | |||||||
| 8291 | assert((!InLoopVal || NumIncoming == 2) &&(static_cast <bool> ((!InLoopVal || NumIncoming == 2) && "Found an in-loop reduction for PHI with unexpected number of " "incoming values") ? void (0) : __assert_fail ("(!InLoopVal || NumIncoming == 2) && \"Found an in-loop reduction for PHI with unexpected number of \" \"incoming values\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8293, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8292 | "Found an in-loop reduction for PHI with unexpected number of "(static_cast <bool> ((!InLoopVal || NumIncoming == 2) && "Found an in-loop reduction for PHI with unexpected number of " "incoming values") ? void (0) : __assert_fail ("(!InLoopVal || NumIncoming == 2) && \"Found an in-loop reduction for PHI with unexpected number of \" \"incoming values\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8293, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8293 | "incoming values")(static_cast <bool> ((!InLoopVal || NumIncoming == 2) && "Found an in-loop reduction for PHI with unexpected number of " "incoming values") ? void (0) : __assert_fail ("(!InLoopVal || NumIncoming == 2) && \"Found an in-loop reduction for PHI with unexpected number of \" \"incoming values\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8293, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8294 | if (InLoopVal) | ||||||
| 8295 | return Operands[Operands[0] == InLoopVal ? 1 : 0]; | ||||||
| 8296 | |||||||
| 8297 | // We know that all PHIs in non-header blocks are converted into selects, so | ||||||
| 8298 | // we don't have to worry about the insertion order and we can just use the | ||||||
| 8299 | // builder. At this point we generate the predication tree. There may be | ||||||
| 8300 | // duplications since this is a simple recursive scan, but future | ||||||
| 8301 | // optimizations will clean it up. | ||||||
| 8302 | SmallVector<VPValue *, 2> OperandsWithMask; | ||||||
| 8303 | |||||||
| 8304 | for (unsigned In = 0; In < NumIncoming; In++) { | ||||||
| 8305 | VPValue *EdgeMask = | ||||||
| 8306 | createEdgeMask(Phi->getIncomingBlock(In), Phi->getParent(), Plan); | ||||||
| 8307 | assert((EdgeMask || NumIncoming == 1) &&(static_cast <bool> ((EdgeMask || NumIncoming == 1) && "Multiple predecessors with one having a full mask") ? void ( 0) : __assert_fail ("(EdgeMask || NumIncoming == 1) && \"Multiple predecessors with one having a full mask\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8308, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8308 | "Multiple predecessors with one having a full mask")(static_cast <bool> ((EdgeMask || NumIncoming == 1) && "Multiple predecessors with one having a full mask") ? void ( 0) : __assert_fail ("(EdgeMask || NumIncoming == 1) && \"Multiple predecessors with one having a full mask\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8308, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8309 | OperandsWithMask.push_back(Operands[In]); | ||||||
| 8310 | if (EdgeMask) | ||||||
| 8311 | OperandsWithMask.push_back(EdgeMask); | ||||||
| 8312 | } | ||||||
| 8313 | return toVPRecipeResult(new VPBlendRecipe(Phi, OperandsWithMask)); | ||||||
| 8314 | } | ||||||
| 8315 | |||||||
| 8316 | VPWidenCallRecipe *VPRecipeBuilder::tryToWidenCall(CallInst *CI, | ||||||
| 8317 | ArrayRef<VPValue *> Operands, | ||||||
| 8318 | VFRange &Range) const { | ||||||
| 8319 | |||||||
| 8320 | bool IsPredicated = LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||
| 8321 | [this, CI](ElementCount VF) { | ||||||
| 8322 | return CM.isScalarWithPredication(CI, VF); | ||||||
| 8323 | }, | ||||||
| 8324 | Range); | ||||||
| 8325 | |||||||
| 8326 | if (IsPredicated) | ||||||
| 8327 | return nullptr; | ||||||
| 8328 | |||||||
| 8329 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||||
| 8330 | if (ID && (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end || | ||||||
| 8331 | ID == Intrinsic::lifetime_start || ID == Intrinsic::sideeffect || | ||||||
| 8332 | ID == Intrinsic::pseudoprobe || | ||||||
| 8333 | ID == Intrinsic::experimental_noalias_scope_decl)) | ||||||
| 8334 | return nullptr; | ||||||
| 8335 | |||||||
| 8336 | ArrayRef<VPValue *> Ops = Operands.take_front(CI->arg_size()); | ||||||
| 8337 | |||||||
| 8338 | // Is it beneficial to perform intrinsic call compared to lib call? | ||||||
| 8339 | bool ShouldUseVectorIntrinsic = | ||||||
| 8340 | ID && LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||
| 8341 | [&](ElementCount VF) -> bool { | ||||||
| 8342 | bool NeedToScalarize = false; | ||||||
| 8343 | // Is it beneficial to perform intrinsic call compared to lib | ||||||
| 8344 | // call? | ||||||
| 8345 | InstructionCost CallCost = | ||||||
| 8346 | CM.getVectorCallCost(CI, VF, NeedToScalarize); | ||||||
| 8347 | InstructionCost IntrinsicCost = | ||||||
| 8348 | CM.getVectorIntrinsicCost(CI, VF); | ||||||
| 8349 | return IntrinsicCost <= CallCost; | ||||||
| 8350 | }, | ||||||
| 8351 | Range); | ||||||
| 8352 | if (ShouldUseVectorIntrinsic) | ||||||
| 8353 | return new VPWidenCallRecipe(*CI, make_range(Ops.begin(), Ops.end()), ID); | ||||||
| 8354 | |||||||
| 8355 | // Is better to call a vectorized version of the function than to to scalarize | ||||||
| 8356 | // the call? | ||||||
| 8357 | auto ShouldUseVectorCall = LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||
| 8358 | [&](ElementCount VF) -> bool { | ||||||
| 8359 | // The following case may be scalarized depending on the VF. | ||||||
| 8360 | // The flag shows whether we can use a usual Call for vectorized | ||||||
| 8361 | // version of the instruction. | ||||||
| 8362 | bool NeedToScalarize = false; | ||||||
| 8363 | CM.getVectorCallCost(CI, VF, NeedToScalarize); | ||||||
| 8364 | return !NeedToScalarize; | ||||||
| 8365 | }, | ||||||
| 8366 | Range); | ||||||
| 8367 | if (ShouldUseVectorCall) | ||||||
| 8368 | return new VPWidenCallRecipe(*CI, make_range(Ops.begin(), Ops.end()), | ||||||
| 8369 | Intrinsic::not_intrinsic); | ||||||
| 8370 | |||||||
| 8371 | return nullptr; | ||||||
| 8372 | } | ||||||
| 8373 | |||||||
| 8374 | bool VPRecipeBuilder::shouldWiden(Instruction *I, VFRange &Range) const { | ||||||
| 8375 | assert(!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) &&(static_cast <bool> (!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) && !isa<StoreInst>(I) && "Instruction should have been handled earlier" ) ? void (0) : __assert_fail ("!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) && !isa<StoreInst>(I) && \"Instruction should have been handled earlier\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8376, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8376 | !isa<StoreInst>(I) && "Instruction should have been handled earlier")(static_cast <bool> (!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) && !isa<StoreInst>(I) && "Instruction should have been handled earlier" ) ? void (0) : __assert_fail ("!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) && !isa<StoreInst>(I) && \"Instruction should have been handled earlier\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8376, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8377 | // Instruction should be widened, unless it is scalar after vectorization, | ||||||
| 8378 | // scalarization is profitable or it is predicated. | ||||||
| 8379 | auto WillScalarize = [this, I](ElementCount VF) -> bool { | ||||||
| 8380 | return CM.isScalarAfterVectorization(I, VF) || | ||||||
| 8381 | CM.isProfitableToScalarize(I, VF) || | ||||||
| 8382 | CM.isScalarWithPredication(I, VF); | ||||||
| 8383 | }; | ||||||
| 8384 | return !LoopVectorizationPlanner::getDecisionAndClampRange(WillScalarize, | ||||||
| 8385 | Range); | ||||||
| 8386 | } | ||||||
| 8387 | |||||||
| 8388 | VPRecipeBase *VPRecipeBuilder::tryToWiden(Instruction *I, | ||||||
| 8389 | ArrayRef<VPValue *> Operands, | ||||||
| 8390 | VPBasicBlock *VPBB, VPlanPtr &Plan) { | ||||||
| 8391 | switch (I->getOpcode()) { | ||||||
| 8392 | default: | ||||||
| 8393 | return nullptr; | ||||||
| 8394 | case Instruction::SDiv: | ||||||
| 8395 | case Instruction::UDiv: | ||||||
| 8396 | case Instruction::SRem: | ||||||
| 8397 | case Instruction::URem: { | ||||||
| 8398 | // If not provably safe, use a select to form a safe divisor before widening the | ||||||
| 8399 | // div/rem operation itself. Otherwise fall through to general handling below. | ||||||
| 8400 | if (CM.isPredicatedInst(I)) { | ||||||
| 8401 | SmallVector<VPValue *> Ops(Operands.begin(), Operands.end()); | ||||||
| 8402 | VPValue *Mask = createBlockInMask(I->getParent(), Plan); | ||||||
| 8403 | VPValue *One = | ||||||
| 8404 | Plan->getOrAddExternalDef(ConstantInt::get(I->getType(), 1u, false)); | ||||||
| 8405 | auto *SafeRHS = | ||||||
| 8406 | new VPInstruction(Instruction::Select, {Mask, Ops[1], One}, | ||||||
| 8407 | I->getDebugLoc()); | ||||||
| 8408 | VPBB->appendRecipe(SafeRHS); | ||||||
| 8409 | Ops[1] = SafeRHS; | ||||||
| 8410 | return new VPWidenRecipe(*I, make_range(Ops.begin(), Ops.end())); | ||||||
| 8411 | } | ||||||
| 8412 | LLVM_FALLTHROUGH[[fallthrough]]; | ||||||
| 8413 | } | ||||||
| 8414 | case Instruction::Add: | ||||||
| 8415 | case Instruction::And: | ||||||
| 8416 | case Instruction::AShr: | ||||||
| 8417 | case Instruction::BitCast: | ||||||
| 8418 | case Instruction::FAdd: | ||||||
| 8419 | case Instruction::FCmp: | ||||||
| 8420 | case Instruction::FDiv: | ||||||
| 8421 | case Instruction::FMul: | ||||||
| 8422 | case Instruction::FNeg: | ||||||
| 8423 | case Instruction::FPExt: | ||||||
| 8424 | case Instruction::FPToSI: | ||||||
| 8425 | case Instruction::FPToUI: | ||||||
| 8426 | case Instruction::FPTrunc: | ||||||
| 8427 | case Instruction::FRem: | ||||||
| 8428 | case Instruction::FSub: | ||||||
| 8429 | case Instruction::ICmp: | ||||||
| 8430 | case Instruction::IntToPtr: | ||||||
| 8431 | case Instruction::LShr: | ||||||
| 8432 | case Instruction::Mul: | ||||||
| 8433 | case Instruction::Or: | ||||||
| 8434 | case Instruction::PtrToInt: | ||||||
| 8435 | case Instruction::Select: | ||||||
| 8436 | case Instruction::SExt: | ||||||
| 8437 | case Instruction::Shl: | ||||||
| 8438 | case Instruction::SIToFP: | ||||||
| 8439 | case Instruction::Sub: | ||||||
| 8440 | case Instruction::Trunc: | ||||||
| 8441 | case Instruction::UIToFP: | ||||||
| 8442 | case Instruction::Xor: | ||||||
| 8443 | case Instruction::ZExt: | ||||||
| 8444 | case Instruction::Freeze: | ||||||
| 8445 | return new VPWidenRecipe(*I, make_range(Operands.begin(), Operands.end())); | ||||||
| 8446 | }; | ||||||
| 8447 | } | ||||||
| 8448 | |||||||
| 8449 | void VPRecipeBuilder::fixHeaderPhis() { | ||||||
| 8450 | BasicBlock *OrigLatch = OrigLoop->getLoopLatch(); | ||||||
| 8451 | for (VPHeaderPHIRecipe *R : PhisToFix) { | ||||||
| 8452 | auto *PN = cast<PHINode>(R->getUnderlyingValue()); | ||||||
| 8453 | VPRecipeBase *IncR = | ||||||
| 8454 | getRecipe(cast<Instruction>(PN->getIncomingValueForBlock(OrigLatch))); | ||||||
| 8455 | R->addOperand(IncR->getVPSingleValue()); | ||||||
| 8456 | } | ||||||
| 8457 | } | ||||||
| 8458 | |||||||
| 8459 | VPBasicBlock *VPRecipeBuilder::handleReplication( | ||||||
| 8460 | Instruction *I, VFRange &Range, VPBasicBlock *VPBB, | ||||||
| 8461 | VPlanPtr &Plan) { | ||||||
| 8462 | bool IsUniform = LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||
| 8463 | [&](ElementCount VF) { return CM.isUniformAfterVectorization(I, VF); }, | ||||||
| 8464 | Range); | ||||||
| 8465 | |||||||
| 8466 | bool IsPredicated = CM.isPredicatedInst(I); | ||||||
| 8467 | |||||||
| 8468 | // Even if the instruction is not marked as uniform, there are certain | ||||||
| 8469 | // intrinsic calls that can be effectively treated as such, so we check for | ||||||
| 8470 | // them here. Conservatively, we only do this for scalable vectors, since | ||||||
| 8471 | // for fixed-width VFs we can always fall back on full scalarization. | ||||||
| 8472 | if (!IsUniform && Range.Start.isScalable() && isa<IntrinsicInst>(I)) { | ||||||
| 8473 | switch (cast<IntrinsicInst>(I)->getIntrinsicID()) { | ||||||
| 8474 | case Intrinsic::assume: | ||||||
| 8475 | case Intrinsic::lifetime_start: | ||||||
| 8476 | case Intrinsic::lifetime_end: | ||||||
| 8477 | // For scalable vectors if one of the operands is variant then we still | ||||||
| 8478 | // want to mark as uniform, which will generate one instruction for just | ||||||
| 8479 | // the first lane of the vector. We can't scalarize the call in the same | ||||||
| 8480 | // way as for fixed-width vectors because we don't know how many lanes | ||||||
| 8481 | // there are. | ||||||
| 8482 | // | ||||||
| 8483 | // The reasons for doing it this way for scalable vectors are: | ||||||
| 8484 | // 1. For the assume intrinsic generating the instruction for the first | ||||||
| 8485 | // lane is still be better than not generating any at all. For | ||||||
| 8486 | // example, the input may be a splat across all lanes. | ||||||
| 8487 | // 2. For the lifetime start/end intrinsics the pointer operand only | ||||||
| 8488 | // does anything useful when the input comes from a stack object, | ||||||
| 8489 | // which suggests it should always be uniform. For non-stack objects | ||||||
| 8490 | // the effect is to poison the object, which still allows us to | ||||||
| 8491 | // remove the call. | ||||||
| 8492 | IsUniform = true; | ||||||
| 8493 | break; | ||||||
| 8494 | default: | ||||||
| 8495 | break; | ||||||
| 8496 | } | ||||||
| 8497 | } | ||||||
| 8498 | |||||||
| 8499 | auto *Recipe = new VPReplicateRecipe(I, Plan->mapToVPValues(I->operands()), | ||||||
| 8500 | IsUniform, IsPredicated); | ||||||
| 8501 | |||||||
| 8502 | // Find if I uses a predicated instruction. If so, it will use its scalar | ||||||
| 8503 | // value. Avoid hoisting the insert-element which packs the scalar value into | ||||||
| 8504 | // a vector value, as that happens iff all users use the vector value. | ||||||
| 8505 | for (VPValue *Op : Recipe->operands()) { | ||||||
| 8506 | auto *PredR = | ||||||
| 8507 | dyn_cast_or_null<VPPredInstPHIRecipe>(Op->getDefiningRecipe()); | ||||||
| 8508 | if (!PredR) | ||||||
| 8509 | continue; | ||||||
| 8510 | auto *RepR = cast<VPReplicateRecipe>( | ||||||
| 8511 | PredR->getOperand(0)->getDefiningRecipe()); | ||||||
| 8512 | assert(RepR->isPredicated() &&(static_cast <bool> (RepR->isPredicated() && "expected Replicate recipe to be predicated") ? void (0) : __assert_fail ("RepR->isPredicated() && \"expected Replicate recipe to be predicated\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8513, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8513 | "expected Replicate recipe to be predicated")(static_cast <bool> (RepR->isPredicated() && "expected Replicate recipe to be predicated") ? void (0) : __assert_fail ("RepR->isPredicated() && \"expected Replicate recipe to be predicated\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8513, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8514 | RepR->setAlsoPack(false); | ||||||
| 8515 | } | ||||||
| 8516 | |||||||
| 8517 | // Finalize the recipe for Instr, first if it is not predicated. | ||||||
| 8518 | if (!IsPredicated) { | ||||||
| 8519 | LLVM_DEBUG(dbgs() << "LV: Scalarizing:" << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Scalarizing:" << *I << "\n"; } } while (false); | ||||||
| 8520 | setRecipe(I, Recipe); | ||||||
| 8521 | Plan->addVPValue(I, Recipe); | ||||||
| 8522 | VPBB->appendRecipe(Recipe); | ||||||
| 8523 | return VPBB; | ||||||
| 8524 | } | ||||||
| 8525 | LLVM_DEBUG(dbgs() << "LV: Scalarizing and predicating:" << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Scalarizing and predicating:" << *I << "\n"; } } while (false); | ||||||
| 8526 | |||||||
| 8527 | VPBlockBase *SingleSucc = VPBB->getSingleSuccessor(); | ||||||
| 8528 | assert(SingleSucc && "VPBB must have a single successor when handling "(static_cast <bool> (SingleSucc && "VPBB must have a single successor when handling " "predicated replication.") ? void (0) : __assert_fail ("SingleSucc && \"VPBB must have a single successor when handling \" \"predicated replication.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8529, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8529 | "predicated replication.")(static_cast <bool> (SingleSucc && "VPBB must have a single successor when handling " "predicated replication.") ? void (0) : __assert_fail ("SingleSucc && \"VPBB must have a single successor when handling \" \"predicated replication.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8529, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8530 | VPBlockUtils::disconnectBlocks(VPBB, SingleSucc); | ||||||
| 8531 | // Record predicated instructions for above packing optimizations. | ||||||
| 8532 | VPBlockBase *Region = createReplicateRegion(Recipe, Plan); | ||||||
| 8533 | VPBlockUtils::insertBlockAfter(Region, VPBB); | ||||||
| 8534 | auto *RegSucc = new VPBasicBlock(); | ||||||
| 8535 | VPBlockUtils::insertBlockAfter(RegSucc, Region); | ||||||
| 8536 | VPBlockUtils::connectBlocks(RegSucc, SingleSucc); | ||||||
| 8537 | return RegSucc; | ||||||
| 8538 | } | ||||||
| 8539 | |||||||
| 8540 | VPRegionBlock * | ||||||
| 8541 | VPRecipeBuilder::createReplicateRegion(VPReplicateRecipe *PredRecipe, | ||||||
| 8542 | VPlanPtr &Plan) { | ||||||
| 8543 | Instruction *Instr = PredRecipe->getUnderlyingInstr(); | ||||||
| 8544 | // Instructions marked for predication are replicated and placed under an | ||||||
| 8545 | // if-then construct to prevent side-effects. | ||||||
| 8546 | // Generate recipes to compute the block mask for this region. | ||||||
| 8547 | VPValue *BlockInMask = createBlockInMask(Instr->getParent(), Plan); | ||||||
| 8548 | |||||||
| 8549 | // Build the triangular if-then region. | ||||||
| 8550 | std::string RegionName = (Twine("pred.") + Instr->getOpcodeName()).str(); | ||||||
| 8551 | assert(Instr->getParent() && "Predicated instruction not in any basic block")(static_cast <bool> (Instr->getParent() && "Predicated instruction not in any basic block" ) ? void (0) : __assert_fail ("Instr->getParent() && \"Predicated instruction not in any basic block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8551, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8552 | auto *BOMRecipe = new VPBranchOnMaskRecipe(BlockInMask); | ||||||
| 8553 | auto *Entry = new VPBasicBlock(Twine(RegionName) + ".entry", BOMRecipe); | ||||||
| 8554 | auto *PHIRecipe = Instr->getType()->isVoidTy() | ||||||
| 8555 | ? nullptr | ||||||
| 8556 | : new VPPredInstPHIRecipe(PredRecipe); | ||||||
| 8557 | if (PHIRecipe) { | ||||||
| 8558 | setRecipe(Instr, PHIRecipe); | ||||||
| 8559 | Plan->addVPValue(Instr, PHIRecipe); | ||||||
| 8560 | } else { | ||||||
| 8561 | setRecipe(Instr, PredRecipe); | ||||||
| 8562 | Plan->addVPValue(Instr, PredRecipe); | ||||||
| 8563 | } | ||||||
| 8564 | |||||||
| 8565 | auto *Exiting = new VPBasicBlock(Twine(RegionName) + ".continue", PHIRecipe); | ||||||
| 8566 | auto *Pred = new VPBasicBlock(Twine(RegionName) + ".if", PredRecipe); | ||||||
| 8567 | VPRegionBlock *Region = new VPRegionBlock(Entry, Exiting, RegionName, true); | ||||||
| 8568 | |||||||
| 8569 | // Note: first set Entry as region entry and then connect successors starting | ||||||
| 8570 | // from it in order, to propagate the "parent" of each VPBasicBlock. | ||||||
| 8571 | VPBlockUtils::insertTwoBlocksAfter(Pred, Exiting, Entry); | ||||||
| 8572 | VPBlockUtils::connectBlocks(Pred, Exiting); | ||||||
| 8573 | |||||||
| 8574 | return Region; | ||||||
| 8575 | } | ||||||
| 8576 | |||||||
| 8577 | VPRecipeOrVPValueTy | ||||||
| 8578 | VPRecipeBuilder::tryToCreateWidenRecipe(Instruction *Instr, | ||||||
| 8579 | ArrayRef<VPValue *> Operands, | ||||||
| 8580 | VFRange &Range, VPBasicBlock *VPBB, | ||||||
| 8581 | VPlanPtr &Plan) { | ||||||
| 8582 | // First, check for specific widening recipes that deal with inductions, Phi | ||||||
| 8583 | // nodes, calls and memory operations. | ||||||
| 8584 | VPRecipeBase *Recipe; | ||||||
| 8585 | if (auto Phi = dyn_cast<PHINode>(Instr)) { | ||||||
| 8586 | if (Phi->getParent() != OrigLoop->getHeader()) | ||||||
| 8587 | return tryToBlend(Phi, Operands, Plan); | ||||||
| 8588 | |||||||
| 8589 | // Always record recipes for header phis. Later first-order recurrence phis | ||||||
| 8590 | // can have earlier phis as incoming values. | ||||||
| 8591 | recordRecipeOf(Phi); | ||||||
| 8592 | |||||||
| 8593 | if ((Recipe = tryToOptimizeInductionPHI(Phi, Operands, *Plan, Range))) | ||||||
| 8594 | return toVPRecipeResult(Recipe); | ||||||
| 8595 | |||||||
| 8596 | VPHeaderPHIRecipe *PhiRecipe = nullptr; | ||||||
| 8597 | assert((Legal->isReductionVariable(Phi) ||(static_cast <bool> ((Legal->isReductionVariable(Phi ) || Legal->isFixedOrderRecurrence(Phi)) && "can only widen reductions and fixed-order recurrences here" ) ? void (0) : __assert_fail ("(Legal->isReductionVariable(Phi) || Legal->isFixedOrderRecurrence(Phi)) && \"can only widen reductions and fixed-order recurrences here\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8599, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8598 | Legal->isFixedOrderRecurrence(Phi)) &&(static_cast <bool> ((Legal->isReductionVariable(Phi ) || Legal->isFixedOrderRecurrence(Phi)) && "can only widen reductions and fixed-order recurrences here" ) ? void (0) : __assert_fail ("(Legal->isReductionVariable(Phi) || Legal->isFixedOrderRecurrence(Phi)) && \"can only widen reductions and fixed-order recurrences here\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8599, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8599 | "can only widen reductions and fixed-order recurrences here")(static_cast <bool> ((Legal->isReductionVariable(Phi ) || Legal->isFixedOrderRecurrence(Phi)) && "can only widen reductions and fixed-order recurrences here" ) ? void (0) : __assert_fail ("(Legal->isReductionVariable(Phi) || Legal->isFixedOrderRecurrence(Phi)) && \"can only widen reductions and fixed-order recurrences here\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8599, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8600 | VPValue *StartV = Operands[0]; | ||||||
| 8601 | if (Legal->isReductionVariable(Phi)) { | ||||||
| 8602 | const RecurrenceDescriptor &RdxDesc = | ||||||
| 8603 | Legal->getReductionVars().find(Phi)->second; | ||||||
| 8604 | assert(RdxDesc.getRecurrenceStartValue() ==(static_cast <bool> (RdxDesc.getRecurrenceStartValue() == Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader ())) ? void (0) : __assert_fail ("RdxDesc.getRecurrenceStartValue() == Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader())" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8605, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8605 | Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader()))(static_cast <bool> (RdxDesc.getRecurrenceStartValue() == Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader ())) ? void (0) : __assert_fail ("RdxDesc.getRecurrenceStartValue() == Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader())" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8605, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8606 | PhiRecipe = new VPReductionPHIRecipe(Phi, RdxDesc, *StartV, | ||||||
| 8607 | CM.isInLoopReduction(Phi), | ||||||
| 8608 | CM.useOrderedReductions(RdxDesc)); | ||||||
| 8609 | } else { | ||||||
| 8610 | // TODO: Currently fixed-order recurrences are modeled as chains of | ||||||
| 8611 | // first-order recurrences. If there are no users of the intermediate | ||||||
| 8612 | // recurrences in the chain, the fixed order recurrence should be modeled | ||||||
| 8613 | // directly, enabling more efficient codegen. | ||||||
| 8614 | PhiRecipe = new VPFirstOrderRecurrencePHIRecipe(Phi, *StartV); | ||||||
| 8615 | } | ||||||
| 8616 | |||||||
| 8617 | // Record the incoming value from the backedge, so we can add the incoming | ||||||
| 8618 | // value from the backedge after all recipes have been created. | ||||||
| 8619 | auto *Inc = cast<Instruction>( | ||||||
| 8620 | Phi->getIncomingValueForBlock(OrigLoop->getLoopLatch())); | ||||||
| 8621 | auto RecipeIter = Ingredient2Recipe.find(Inc); | ||||||
| 8622 | if (RecipeIter == Ingredient2Recipe.end()) | ||||||
| 8623 | recordRecipeOf(Inc); | ||||||
| 8624 | |||||||
| 8625 | PhisToFix.push_back(PhiRecipe); | ||||||
| 8626 | return toVPRecipeResult(PhiRecipe); | ||||||
| 8627 | } | ||||||
| 8628 | |||||||
| 8629 | if (isa<TruncInst>(Instr) && | ||||||
| 8630 | (Recipe = tryToOptimizeInductionTruncate(cast<TruncInst>(Instr), Operands, | ||||||
| 8631 | Range, *Plan))) | ||||||
| 8632 | return toVPRecipeResult(Recipe); | ||||||
| 8633 | |||||||
| 8634 | // All widen recipes below deal only with VF > 1. | ||||||
| 8635 | if (LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||
| 8636 | [&](ElementCount VF) { return VF.isScalar(); }, Range)) | ||||||
| 8637 | return nullptr; | ||||||
| 8638 | |||||||
| 8639 | if (auto *CI = dyn_cast<CallInst>(Instr)) | ||||||
| 8640 | return toVPRecipeResult(tryToWidenCall(CI, Operands, Range)); | ||||||
| 8641 | |||||||
| 8642 | if (isa<LoadInst>(Instr) || isa<StoreInst>(Instr)) | ||||||
| 8643 | return toVPRecipeResult(tryToWidenMemory(Instr, Operands, Range, Plan)); | ||||||
| 8644 | |||||||
| 8645 | if (!shouldWiden(Instr, Range)) | ||||||
| 8646 | return nullptr; | ||||||
| 8647 | |||||||
| 8648 | if (auto GEP = dyn_cast<GetElementPtrInst>(Instr)) | ||||||
| 8649 | return toVPRecipeResult(new VPWidenGEPRecipe( | ||||||
| 8650 | GEP, make_range(Operands.begin(), Operands.end()), OrigLoop)); | ||||||
| 8651 | |||||||
| 8652 | if (auto *SI = dyn_cast<SelectInst>(Instr)) { | ||||||
| 8653 | bool InvariantCond = | ||||||
| 8654 | PSE.getSE()->isLoopInvariant(PSE.getSCEV(SI->getOperand(0)), OrigLoop); | ||||||
| 8655 | return toVPRecipeResult(new VPWidenSelectRecipe( | ||||||
| 8656 | *SI, make_range(Operands.begin(), Operands.end()), InvariantCond)); | ||||||
| 8657 | } | ||||||
| 8658 | |||||||
| 8659 | return toVPRecipeResult(tryToWiden(Instr, Operands, VPBB, Plan)); | ||||||
| 8660 | } | ||||||
| 8661 | |||||||
| 8662 | void LoopVectorizationPlanner::buildVPlansWithVPRecipes(ElementCount MinVF, | ||||||
| 8663 | ElementCount MaxVF) { | ||||||
| 8664 | assert(OrigLoop->isInnermost() && "Inner loop expected.")(static_cast <bool> (OrigLoop->isInnermost() && "Inner loop expected.") ? void (0) : __assert_fail ("OrigLoop->isInnermost() && \"Inner loop expected.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8664, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8665 | |||||||
| 8666 | // Add assume instructions we need to drop to DeadInstructions, to prevent | ||||||
| 8667 | // them from being added to the VPlan. | ||||||
| 8668 | // TODO: We only need to drop assumes in blocks that get flattend. If the | ||||||
| 8669 | // control flow is preserved, we should keep them. | ||||||
| 8670 | SmallPtrSet<Instruction *, 4> DeadInstructions; | ||||||
| 8671 | auto &ConditionalAssumes = Legal->getConditionalAssumes(); | ||||||
| 8672 | DeadInstructions.insert(ConditionalAssumes.begin(), ConditionalAssumes.end()); | ||||||
| 8673 | |||||||
| 8674 | MapVector<Instruction *, Instruction *> &SinkAfter = Legal->getSinkAfter(); | ||||||
| 8675 | // Dead instructions do not need sinking. Remove them from SinkAfter. | ||||||
| 8676 | for (Instruction *I : DeadInstructions) | ||||||
| 8677 | SinkAfter.erase(I); | ||||||
| 8678 | |||||||
| 8679 | // Cannot sink instructions after dead instructions (there won't be any | ||||||
| 8680 | // recipes for them). Instead, find the first non-dead previous instruction. | ||||||
| 8681 | for (auto &P : Legal->getSinkAfter()) { | ||||||
| 8682 | Instruction *SinkTarget = P.second; | ||||||
| 8683 | Instruction *FirstInst = &*SinkTarget->getParent()->begin(); | ||||||
| 8684 | (void)FirstInst; | ||||||
| 8685 | while (DeadInstructions.contains(SinkTarget)) { | ||||||
| 8686 | assert((static_cast <bool> (SinkTarget != FirstInst && "Must find a live instruction (at least the one feeding the " "fixed-order recurrence PHI) before reaching beginning of the block" ) ? void (0) : __assert_fail ("SinkTarget != FirstInst && \"Must find a live instruction (at least the one feeding the \" \"fixed-order recurrence PHI) before reaching beginning of the block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8689, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8687 | SinkTarget != FirstInst &&(static_cast <bool> (SinkTarget != FirstInst && "Must find a live instruction (at least the one feeding the " "fixed-order recurrence PHI) before reaching beginning of the block" ) ? void (0) : __assert_fail ("SinkTarget != FirstInst && \"Must find a live instruction (at least the one feeding the \" \"fixed-order recurrence PHI) before reaching beginning of the block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8689, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8688 | "Must find a live instruction (at least the one feeding the "(static_cast <bool> (SinkTarget != FirstInst && "Must find a live instruction (at least the one feeding the " "fixed-order recurrence PHI) before reaching beginning of the block" ) ? void (0) : __assert_fail ("SinkTarget != FirstInst && \"Must find a live instruction (at least the one feeding the \" \"fixed-order recurrence PHI) before reaching beginning of the block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8689, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8689 | "fixed-order recurrence PHI) before reaching beginning of the block")(static_cast <bool> (SinkTarget != FirstInst && "Must find a live instruction (at least the one feeding the " "fixed-order recurrence PHI) before reaching beginning of the block" ) ? void (0) : __assert_fail ("SinkTarget != FirstInst && \"Must find a live instruction (at least the one feeding the \" \"fixed-order recurrence PHI) before reaching beginning of the block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8689, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8690 | SinkTarget = SinkTarget->getPrevNode(); | ||||||
| 8691 | assert(SinkTarget != P.first &&(static_cast <bool> (SinkTarget != P.first && "sink source equals target, no sinking required" ) ? void (0) : __assert_fail ("SinkTarget != P.first && \"sink source equals target, no sinking required\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8692, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8692 | "sink source equals target, no sinking required")(static_cast <bool> (SinkTarget != P.first && "sink source equals target, no sinking required" ) ? void (0) : __assert_fail ("SinkTarget != P.first && \"sink source equals target, no sinking required\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8692, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8693 | } | ||||||
| 8694 | P.second = SinkTarget; | ||||||
| 8695 | } | ||||||
| 8696 | |||||||
| 8697 | auto MaxVFPlusOne = MaxVF.getWithIncrement(1); | ||||||
| 8698 | for (ElementCount VF = MinVF; ElementCount::isKnownLT(VF, MaxVFPlusOne);) { | ||||||
| 8699 | VFRange SubRange = {VF, MaxVFPlusOne}; | ||||||
| 8700 | VPlans.push_back( | ||||||
| 8701 | buildVPlanWithVPRecipes(SubRange, DeadInstructions, SinkAfter)); | ||||||
| 8702 | VF = SubRange.End; | ||||||
| 8703 | } | ||||||
| 8704 | } | ||||||
| 8705 | |||||||
| 8706 | // Add the necessary canonical IV and branch recipes required to control the | ||||||
| 8707 | // loop. | ||||||
| 8708 | static void addCanonicalIVRecipes(VPlan &Plan, Type *IdxTy, DebugLoc DL, | ||||||
| 8709 | bool HasNUW, | ||||||
| 8710 | bool UseLaneMaskForLoopControlFlow) { | ||||||
| 8711 | Value *StartIdx = ConstantInt::get(IdxTy, 0); | ||||||
| 8712 | auto *StartV = Plan.getOrAddVPValue(StartIdx); | ||||||
| 8713 | |||||||
| 8714 | // Add a VPCanonicalIVPHIRecipe starting at 0 to the header. | ||||||
| 8715 | auto *CanonicalIVPHI = new VPCanonicalIVPHIRecipe(StartV, DL); | ||||||
| 8716 | VPRegionBlock *TopRegion = Plan.getVectorLoopRegion(); | ||||||
| 8717 | VPBasicBlock *Header = TopRegion->getEntryBasicBlock(); | ||||||
| 8718 | Header->insert(CanonicalIVPHI, Header->begin()); | ||||||
| 8719 | |||||||
| 8720 | // Add a CanonicalIVIncrement{NUW} VPInstruction to increment the scalar | ||||||
| 8721 | // IV by VF * UF. | ||||||
| 8722 | auto *CanonicalIVIncrement = | ||||||
| 8723 | new VPInstruction(HasNUW ? VPInstruction::CanonicalIVIncrementNUW | ||||||
| 8724 | : VPInstruction::CanonicalIVIncrement, | ||||||
| 8725 | {CanonicalIVPHI}, DL, "index.next"); | ||||||
| 8726 | CanonicalIVPHI->addOperand(CanonicalIVIncrement); | ||||||
| 8727 | |||||||
| 8728 | VPBasicBlock *EB = TopRegion->getExitingBasicBlock(); | ||||||
| 8729 | EB->appendRecipe(CanonicalIVIncrement); | ||||||
| 8730 | |||||||
| 8731 | if (UseLaneMaskForLoopControlFlow) { | ||||||
| 8732 | // Create the active lane mask instruction in the vplan preheader. | ||||||
| 8733 | VPBasicBlock *Preheader = Plan.getEntry()->getEntryBasicBlock(); | ||||||
| 8734 | |||||||
| 8735 | // We can't use StartV directly in the ActiveLaneMask VPInstruction, since | ||||||
| 8736 | // we have to take unrolling into account. Each part needs to start at | ||||||
| 8737 | // Part * VF | ||||||
| 8738 | auto *CanonicalIVIncrementParts = | ||||||
| 8739 | new VPInstruction(HasNUW ? VPInstruction::CanonicalIVIncrementForPartNUW | ||||||
| 8740 | : VPInstruction::CanonicalIVIncrementForPart, | ||||||
| 8741 | {StartV}, DL, "index.part.next"); | ||||||
| 8742 | Preheader->appendRecipe(CanonicalIVIncrementParts); | ||||||
| 8743 | |||||||
| 8744 | // Create the ActiveLaneMask instruction using the correct start values. | ||||||
| 8745 | VPValue *TC = Plan.getOrCreateTripCount(); | ||||||
| 8746 | auto *EntryALM = new VPInstruction(VPInstruction::ActiveLaneMask, | ||||||
| 8747 | {CanonicalIVIncrementParts, TC}, DL, | ||||||
| 8748 | "active.lane.mask.entry"); | ||||||
| 8749 | Preheader->appendRecipe(EntryALM); | ||||||
| 8750 | |||||||
| 8751 | // Now create the ActiveLaneMaskPhi recipe in the main loop using the | ||||||
| 8752 | // preheader ActiveLaneMask instruction. | ||||||
| 8753 | auto *LaneMaskPhi = new VPActiveLaneMaskPHIRecipe(EntryALM, DebugLoc()); | ||||||
| 8754 | Header->insert(LaneMaskPhi, Header->getFirstNonPhi()); | ||||||
| 8755 | |||||||
| 8756 | // Create the active lane mask for the next iteration of the loop. | ||||||
| 8757 | CanonicalIVIncrementParts = | ||||||
| 8758 | new VPInstruction(HasNUW ? VPInstruction::CanonicalIVIncrementForPartNUW | ||||||
| 8759 | : VPInstruction::CanonicalIVIncrementForPart, | ||||||
| 8760 | {CanonicalIVIncrement}, DL); | ||||||
| 8761 | EB->appendRecipe(CanonicalIVIncrementParts); | ||||||
| 8762 | |||||||
| 8763 | auto *ALM = new VPInstruction(VPInstruction::ActiveLaneMask, | ||||||
| 8764 | {CanonicalIVIncrementParts, TC}, DL, | ||||||
| 8765 | "active.lane.mask.next"); | ||||||
| 8766 | EB->appendRecipe(ALM); | ||||||
| 8767 | LaneMaskPhi->addOperand(ALM); | ||||||
| 8768 | |||||||
| 8769 | // We have to invert the mask here because a true condition means jumping | ||||||
| 8770 | // to the exit block. | ||||||
| 8771 | auto *NotMask = new VPInstruction(VPInstruction::Not, ALM, DL); | ||||||
| 8772 | EB->appendRecipe(NotMask); | ||||||
| 8773 | |||||||
| 8774 | VPInstruction *BranchBack = | ||||||
| 8775 | new VPInstruction(VPInstruction::BranchOnCond, {NotMask}, DL); | ||||||
| 8776 | EB->appendRecipe(BranchBack); | ||||||
| 8777 | } else { | ||||||
| 8778 | // Add the BranchOnCount VPInstruction to the latch. | ||||||
| 8779 | VPInstruction *BranchBack = new VPInstruction( | ||||||
| 8780 | VPInstruction::BranchOnCount, | ||||||
| 8781 | {CanonicalIVIncrement, &Plan.getVectorTripCount()}, DL); | ||||||
| 8782 | EB->appendRecipe(BranchBack); | ||||||
| 8783 | } | ||||||
| 8784 | } | ||||||
| 8785 | |||||||
| 8786 | // Add exit values to \p Plan. VPLiveOuts are added for each LCSSA phi in the | ||||||
| 8787 | // original exit block. | ||||||
| 8788 | static void addUsersInExitBlock(VPBasicBlock *HeaderVPBB, | ||||||
| 8789 | VPBasicBlock *MiddleVPBB, Loop *OrigLoop, | ||||||
| 8790 | VPlan &Plan) { | ||||||
| 8791 | BasicBlock *ExitBB = OrigLoop->getUniqueExitBlock(); | ||||||
| 8792 | BasicBlock *ExitingBB = OrigLoop->getExitingBlock(); | ||||||
| 8793 | // Only handle single-exit loops with unique exit blocks for now. | ||||||
| 8794 | if (!ExitBB || !ExitBB->getSinglePredecessor() || !ExitingBB) | ||||||
| 8795 | return; | ||||||
| 8796 | |||||||
| 8797 | // Introduce VPUsers modeling the exit values. | ||||||
| 8798 | for (PHINode &ExitPhi : ExitBB->phis()) { | ||||||
| 8799 | Value *IncomingValue = | ||||||
| 8800 | ExitPhi.getIncomingValueForBlock(ExitingBB); | ||||||
| 8801 | VPValue *V = Plan.getOrAddVPValue(IncomingValue, true); | ||||||
| 8802 | Plan.addLiveOut(&ExitPhi, V); | ||||||
| 8803 | } | ||||||
| 8804 | } | ||||||
| 8805 | |||||||
| 8806 | VPlanPtr LoopVectorizationPlanner::buildVPlanWithVPRecipes( | ||||||
| 8807 | VFRange &Range, SmallPtrSetImpl<Instruction *> &DeadInstructions, | ||||||
| 8808 | const MapVector<Instruction *, Instruction *> &SinkAfter) { | ||||||
| 8809 | |||||||
| 8810 | SmallPtrSet<const InterleaveGroup<Instruction> *, 1> InterleaveGroups; | ||||||
| 8811 | |||||||
| 8812 | VPRecipeBuilder RecipeBuilder(OrigLoop, TLI, Legal, CM, PSE, Builder); | ||||||
| 8813 | |||||||
| 8814 | // --------------------------------------------------------------------------- | ||||||
| 8815 | // Pre-construction: record ingredients whose recipes we'll need to further | ||||||
| 8816 | // process after constructing the initial VPlan. | ||||||
| 8817 | // --------------------------------------------------------------------------- | ||||||
| 8818 | |||||||
| 8819 | // Mark instructions we'll need to sink later and their targets as | ||||||
| 8820 | // ingredients whose recipe we'll need to record. | ||||||
| 8821 | for (const auto &Entry : SinkAfter) { | ||||||
| 8822 | RecipeBuilder.recordRecipeOf(Entry.first); | ||||||
| 8823 | RecipeBuilder.recordRecipeOf(Entry.second); | ||||||
| 8824 | } | ||||||
| 8825 | for (const auto &Reduction : CM.getInLoopReductionChains()) { | ||||||
| |||||||
| 8826 | PHINode *Phi = Reduction.first; | ||||||
| 8827 | RecurKind Kind = | ||||||
| 8828 | Legal->getReductionVars().find(Phi)->second.getRecurrenceKind(); | ||||||
| 8829 | const SmallVector<Instruction *, 4> &ReductionOperations = Reduction.second; | ||||||
| 8830 | |||||||
| 8831 | RecipeBuilder.recordRecipeOf(Phi); | ||||||
| 8832 | for (const auto &R : ReductionOperations) { | ||||||
| 8833 | RecipeBuilder.recordRecipeOf(R); | ||||||
| 8834 | // For min/max reductions, where we have a pair of icmp/select, we also | ||||||
| 8835 | // need to record the ICmp recipe, so it can be removed later. | ||||||
| 8836 | assert(!RecurrenceDescriptor::isSelectCmpRecurrenceKind(Kind) &&(static_cast <bool> (!RecurrenceDescriptor::isSelectCmpRecurrenceKind (Kind) && "Only min/max recurrences allowed for inloop reductions" ) ? void (0) : __assert_fail ("!RecurrenceDescriptor::isSelectCmpRecurrenceKind(Kind) && \"Only min/max recurrences allowed for inloop reductions\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8837, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8837 | "Only min/max recurrences allowed for inloop reductions")(static_cast <bool> (!RecurrenceDescriptor::isSelectCmpRecurrenceKind (Kind) && "Only min/max recurrences allowed for inloop reductions" ) ? void (0) : __assert_fail ("!RecurrenceDescriptor::isSelectCmpRecurrenceKind(Kind) && \"Only min/max recurrences allowed for inloop reductions\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8837, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8838 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) | ||||||
| 8839 | RecipeBuilder.recordRecipeOf(cast<Instruction>(R->getOperand(0))); | ||||||
| 8840 | } | ||||||
| 8841 | } | ||||||
| 8842 | |||||||
| 8843 | // For each interleave group which is relevant for this (possibly trimmed) | ||||||
| 8844 | // Range, add it to the set of groups to be later applied to the VPlan and add | ||||||
| 8845 | // placeholders for its members' Recipes which we'll be replacing with a | ||||||
| 8846 | // single VPInterleaveRecipe. | ||||||
| 8847 | for (InterleaveGroup<Instruction> *IG : IAI.getInterleaveGroups()) { | ||||||
| 8848 | auto applyIG = [IG, this](ElementCount VF) -> bool { | ||||||
| 8849 | return (VF.isVector() && // Query is illegal for VF == 1 | ||||||
| 8850 | CM.getWideningDecision(IG->getInsertPos(), VF) == | ||||||
| 8851 | LoopVectorizationCostModel::CM_Interleave); | ||||||
| 8852 | }; | ||||||
| 8853 | if (!getDecisionAndClampRange(applyIG, Range)) | ||||||
| 8854 | continue; | ||||||
| 8855 | InterleaveGroups.insert(IG); | ||||||
| 8856 | for (unsigned i = 0; i < IG->getFactor(); i++) | ||||||
| 8857 | if (Instruction *Member = IG->getMember(i)) | ||||||
| 8858 | RecipeBuilder.recordRecipeOf(Member); | ||||||
| 8859 | }; | ||||||
| 8860 | |||||||
| 8861 | // --------------------------------------------------------------------------- | ||||||
| 8862 | // Build initial VPlan: Scan the body of the loop in a topological order to | ||||||
| 8863 | // visit each basic block after having visited its predecessor basic blocks. | ||||||
| 8864 | // --------------------------------------------------------------------------- | ||||||
| 8865 | |||||||
| 8866 | // Create initial VPlan skeleton, starting with a block for the pre-header, | ||||||
| 8867 | // followed by a region for the vector loop, followed by the middle block. The | ||||||
| 8868 | // skeleton vector loop region contains a header and latch block. | ||||||
| 8869 | VPBasicBlock *Preheader = new VPBasicBlock("vector.ph"); | ||||||
| 8870 | auto Plan = std::make_unique<VPlan>(Preheader); | ||||||
| 8871 | |||||||
| 8872 | VPBasicBlock *HeaderVPBB = new VPBasicBlock("vector.body"); | ||||||
| 8873 | VPBasicBlock *LatchVPBB = new VPBasicBlock("vector.latch"); | ||||||
| 8874 | VPBlockUtils::insertBlockAfter(LatchVPBB, HeaderVPBB); | ||||||
| 8875 | auto *TopRegion = new VPRegionBlock(HeaderVPBB, LatchVPBB, "vector loop"); | ||||||
| 8876 | VPBlockUtils::insertBlockAfter(TopRegion, Preheader); | ||||||
| 8877 | VPBasicBlock *MiddleVPBB = new VPBasicBlock("middle.block"); | ||||||
| 8878 | VPBlockUtils::insertBlockAfter(MiddleVPBB, TopRegion); | ||||||
| 8879 | |||||||
| 8880 | Instruction *DLInst = | ||||||
| 8881 | getDebugLocFromInstOrOperands(Legal->getPrimaryInduction()); | ||||||
| 8882 | addCanonicalIVRecipes(*Plan, Legal->getWidestInductionType(), | ||||||
| 8883 | DLInst
| ||||||
| 8884 | !CM.foldTailByMasking(), | ||||||
| 8885 | CM.useActiveLaneMaskForControlFlow()); | ||||||
| 8886 | |||||||
| 8887 | // Scan the body of the loop in a topological order to visit each basic block | ||||||
| 8888 | // after having visited its predecessor basic blocks. | ||||||
| 8889 | LoopBlocksDFS DFS(OrigLoop); | ||||||
| 8890 | DFS.perform(LI); | ||||||
| 8891 | |||||||
| 8892 | VPBasicBlock *VPBB = HeaderVPBB; | ||||||
| 8893 | SmallVector<VPWidenIntOrFpInductionRecipe *> InductionsToMove; | ||||||
| 8894 | for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) { | ||||||
| 8895 | // Relevant instructions from basic block BB will be grouped into VPRecipe | ||||||
| 8896 | // ingredients and fill a new VPBasicBlock. | ||||||
| 8897 | unsigned VPBBsForBB = 0; | ||||||
| 8898 | if (VPBB != HeaderVPBB) | ||||||
| 8899 | VPBB->setName(BB->getName()); | ||||||
| 8900 | Builder.setInsertPoint(VPBB); | ||||||
| 8901 | |||||||
| 8902 | // Introduce each ingredient into VPlan. | ||||||
| 8903 | // TODO: Model and preserve debug intrinsics in VPlan. | ||||||
| 8904 | for (Instruction &I : BB->instructionsWithoutDebug()) { | ||||||
| 8905 | Instruction *Instr = &I; | ||||||
| 8906 | |||||||
| 8907 | // First filter out irrelevant instructions, to ensure no recipes are | ||||||
| 8908 | // built for them. | ||||||
| 8909 | if (isa<BranchInst>(Instr) || DeadInstructions.count(Instr)) | ||||||
| 8910 | continue; | ||||||
| 8911 | |||||||
| 8912 | SmallVector<VPValue *, 4> Operands; | ||||||
| 8913 | auto *Phi = dyn_cast<PHINode>(Instr); | ||||||
| 8914 | if (Phi && Phi->getParent() == OrigLoop->getHeader()) { | ||||||
| 8915 | Operands.push_back(Plan->getOrAddVPValue( | ||||||
| 8916 | Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader()))); | ||||||
| 8917 | } else { | ||||||
| 8918 | auto OpRange = Plan->mapToVPValues(Instr->operands()); | ||||||
| 8919 | Operands = {OpRange.begin(), OpRange.end()}; | ||||||
| 8920 | } | ||||||
| 8921 | |||||||
| 8922 | // Invariant stores inside loop will be deleted and a single store | ||||||
| 8923 | // with the final reduction value will be added to the exit block | ||||||
| 8924 | StoreInst *SI; | ||||||
| 8925 | if ((SI = dyn_cast<StoreInst>(&I)) && | ||||||
| 8926 | Legal->isInvariantAddressOfReduction(SI->getPointerOperand())) | ||||||
| 8927 | continue; | ||||||
| 8928 | |||||||
| 8929 | if (auto RecipeOrValue = RecipeBuilder.tryToCreateWidenRecipe( | ||||||
| 8930 | Instr, Operands, Range, VPBB, Plan)) { | ||||||
| 8931 | // If Instr can be simplified to an existing VPValue, use it. | ||||||
| 8932 | if (RecipeOrValue.is<VPValue *>()) { | ||||||
| 8933 | auto *VPV = RecipeOrValue.get<VPValue *>(); | ||||||
| 8934 | Plan->addVPValue(Instr, VPV); | ||||||
| 8935 | // If the re-used value is a recipe, register the recipe for the | ||||||
| 8936 | // instruction, in case the recipe for Instr needs to be recorded. | ||||||
| 8937 | if (VPRecipeBase *R = VPV->getDefiningRecipe()) | ||||||
| 8938 | RecipeBuilder.setRecipe(Instr, R); | ||||||
| 8939 | continue; | ||||||
| 8940 | } | ||||||
| 8941 | // Otherwise, add the new recipe. | ||||||
| 8942 | VPRecipeBase *Recipe = RecipeOrValue.get<VPRecipeBase *>(); | ||||||
| 8943 | for (auto *Def : Recipe->definedValues()) { | ||||||
| 8944 | auto *UV = Def->getUnderlyingValue(); | ||||||
| 8945 | Plan->addVPValue(UV, Def); | ||||||
| 8946 | } | ||||||
| 8947 | |||||||
| 8948 | if (isa<VPWidenIntOrFpInductionRecipe>(Recipe) && | ||||||
| 8949 | HeaderVPBB->getFirstNonPhi() != VPBB->end()) { | ||||||
| 8950 | // Keep track of VPWidenIntOrFpInductionRecipes not in the phi section | ||||||
| 8951 | // of the header block. That can happen for truncates of induction | ||||||
| 8952 | // variables. Those recipes are moved to the phi section of the header | ||||||
| 8953 | // block after applying SinkAfter, which relies on the original | ||||||
| 8954 | // position of the trunc. | ||||||
| 8955 | assert(isa<TruncInst>(Instr))(static_cast <bool> (isa<TruncInst>(Instr)) ? void (0) : __assert_fail ("isa<TruncInst>(Instr)", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 8955, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8956 | InductionsToMove.push_back( | ||||||
| 8957 | cast<VPWidenIntOrFpInductionRecipe>(Recipe)); | ||||||
| 8958 | } | ||||||
| 8959 | RecipeBuilder.setRecipe(Instr, Recipe); | ||||||
| 8960 | VPBB->appendRecipe(Recipe); | ||||||
| 8961 | continue; | ||||||
| 8962 | } | ||||||
| 8963 | |||||||
| 8964 | // Otherwise, if all widening options failed, Instruction is to be | ||||||
| 8965 | // replicated. This may create a successor for VPBB. | ||||||
| 8966 | VPBasicBlock *NextVPBB = | ||||||
| 8967 | RecipeBuilder.handleReplication(Instr, Range, VPBB, Plan); | ||||||
| 8968 | if (NextVPBB != VPBB) { | ||||||
| 8969 | VPBB = NextVPBB; | ||||||
| 8970 | VPBB->setName(BB->hasName() ? BB->getName() + "." + Twine(VPBBsForBB++) | ||||||
| 8971 | : ""); | ||||||
| 8972 | } | ||||||
| 8973 | } | ||||||
| 8974 | |||||||
| 8975 | VPBlockUtils::insertBlockAfter(new VPBasicBlock(), VPBB); | ||||||
| 8976 | VPBB = cast<VPBasicBlock>(VPBB->getSingleSuccessor()); | ||||||
| 8977 | } | ||||||
| 8978 | |||||||
| 8979 | HeaderVPBB->setName("vector.body"); | ||||||
| 8980 | |||||||
| 8981 | // Fold the last, empty block into its predecessor. | ||||||
| 8982 | VPBB = VPBlockUtils::tryToMergeBlockIntoPredecessor(VPBB); | ||||||
| 8983 | assert(VPBB && "expected to fold last (empty) block")(static_cast <bool> (VPBB && "expected to fold last (empty) block" ) ? void (0) : __assert_fail ("VPBB && \"expected to fold last (empty) block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8983, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8984 | // After here, VPBB should not be used. | ||||||
| 8985 | VPBB = nullptr; | ||||||
| 8986 | |||||||
| 8987 | addUsersInExitBlock(HeaderVPBB, MiddleVPBB, OrigLoop, *Plan); | ||||||
| |||||||
| 8988 | |||||||
| 8989 | assert(isa<VPRegionBlock>(Plan->getVectorLoopRegion()) &&(static_cast <bool> (isa<VPRegionBlock>(Plan-> getVectorLoopRegion()) && !Plan->getVectorLoopRegion ()->getEntryBasicBlock()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock") ? void (0) : __assert_fail ("isa<VPRegionBlock>(Plan->getVectorLoopRegion()) && !Plan->getVectorLoopRegion()->getEntryBasicBlock()->empty() && \"entry block must be set to a VPRegionBlock having a non-empty entry \" \"VPBasicBlock\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8992, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8990 | !Plan->getVectorLoopRegion()->getEntryBasicBlock()->empty() &&(static_cast <bool> (isa<VPRegionBlock>(Plan-> getVectorLoopRegion()) && !Plan->getVectorLoopRegion ()->getEntryBasicBlock()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock") ? void (0) : __assert_fail ("isa<VPRegionBlock>(Plan->getVectorLoopRegion()) && !Plan->getVectorLoopRegion()->getEntryBasicBlock()->empty() && \"entry block must be set to a VPRegionBlock having a non-empty entry \" \"VPBasicBlock\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8992, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8991 | "entry block must be set to a VPRegionBlock having a non-empty entry "(static_cast <bool> (isa<VPRegionBlock>(Plan-> getVectorLoopRegion()) && !Plan->getVectorLoopRegion ()->getEntryBasicBlock()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock") ? void (0) : __assert_fail ("isa<VPRegionBlock>(Plan->getVectorLoopRegion()) && !Plan->getVectorLoopRegion()->getEntryBasicBlock()->empty() && \"entry block must be set to a VPRegionBlock having a non-empty entry \" \"VPBasicBlock\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8992, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 8992 | "VPBasicBlock")(static_cast <bool> (isa<VPRegionBlock>(Plan-> getVectorLoopRegion()) && !Plan->getVectorLoopRegion ()->getEntryBasicBlock()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock") ? void (0) : __assert_fail ("isa<VPRegionBlock>(Plan->getVectorLoopRegion()) && !Plan->getVectorLoopRegion()->getEntryBasicBlock()->empty() && \"entry block must be set to a VPRegionBlock having a non-empty entry \" \"VPBasicBlock\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8992, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 8993 | RecipeBuilder.fixHeaderPhis(); | ||||||
| 8994 | |||||||
| 8995 | // --------------------------------------------------------------------------- | ||||||
| 8996 | // Transform initial VPlan: Apply previously taken decisions, in order, to | ||||||
| 8997 | // bring the VPlan to its final state. | ||||||
| 8998 | // --------------------------------------------------------------------------- | ||||||
| 8999 | |||||||
| 9000 | // Apply Sink-After legal constraints. | ||||||
| 9001 | auto GetReplicateRegion = [](VPRecipeBase *R) -> VPRegionBlock * { | ||||||
| 9002 | auto *Region = dyn_cast_or_null<VPRegionBlock>(R->getParent()->getParent()); | ||||||
| 9003 | if (Region && Region->isReplicator()) { | ||||||
| 9004 | assert(Region->getNumSuccessors() == 1 &&(static_cast <bool> (Region->getNumSuccessors() == 1 && Region->getNumPredecessors() == 1 && "Expected SESE region!" ) ? void (0) : __assert_fail ("Region->getNumSuccessors() == 1 && Region->getNumPredecessors() == 1 && \"Expected SESE region!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9005, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9005 | Region->getNumPredecessors() == 1 && "Expected SESE region!")(static_cast <bool> (Region->getNumSuccessors() == 1 && Region->getNumPredecessors() == 1 && "Expected SESE region!" ) ? void (0) : __assert_fail ("Region->getNumSuccessors() == 1 && Region->getNumPredecessors() == 1 && \"Expected SESE region!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9005, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9006 | assert(R->getParent()->size() == 1 &&(static_cast <bool> (R->getParent()->size() == 1 && "A recipe in an original replicator region must be the only " "recipe in its block") ? void (0) : __assert_fail ("R->getParent()->size() == 1 && \"A recipe in an original replicator region must be the only \" \"recipe in its block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9008, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9007 | "A recipe in an original replicator region must be the only "(static_cast <bool> (R->getParent()->size() == 1 && "A recipe in an original replicator region must be the only " "recipe in its block") ? void (0) : __assert_fail ("R->getParent()->size() == 1 && \"A recipe in an original replicator region must be the only \" \"recipe in its block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9008, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9008 | "recipe in its block")(static_cast <bool> (R->getParent()->size() == 1 && "A recipe in an original replicator region must be the only " "recipe in its block") ? void (0) : __assert_fail ("R->getParent()->size() == 1 && \"A recipe in an original replicator region must be the only \" \"recipe in its block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9008, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9009 | return Region; | ||||||
| 9010 | } | ||||||
| 9011 | return nullptr; | ||||||
| 9012 | }; | ||||||
| 9013 | for (const auto &Entry : SinkAfter) { | ||||||
| 9014 | VPRecipeBase *Sink = RecipeBuilder.getRecipe(Entry.first); | ||||||
| 9015 | VPRecipeBase *Target = RecipeBuilder.getRecipe(Entry.second); | ||||||
| 9016 | |||||||
| 9017 | auto *TargetRegion = GetReplicateRegion(Target); | ||||||
| 9018 | auto *SinkRegion = GetReplicateRegion(Sink); | ||||||
| 9019 | if (!SinkRegion) { | ||||||
| 9020 | // If the sink source is not a replicate region, sink the recipe directly. | ||||||
| 9021 | if (TargetRegion) { | ||||||
| 9022 | // The target is in a replication region, make sure to move Sink to | ||||||
| 9023 | // the block after it, not into the replication region itself. | ||||||
| 9024 | VPBasicBlock *NextBlock = | ||||||
| 9025 | cast<VPBasicBlock>(TargetRegion->getSuccessors().front()); | ||||||
| 9026 | Sink->moveBefore(*NextBlock, NextBlock->getFirstNonPhi()); | ||||||
| 9027 | } else | ||||||
| 9028 | Sink->moveAfter(Target); | ||||||
| 9029 | continue; | ||||||
| 9030 | } | ||||||
| 9031 | |||||||
| 9032 | // The sink source is in a replicate region. Unhook the region from the CFG. | ||||||
| 9033 | auto *SinkPred = SinkRegion->getSinglePredecessor(); | ||||||
| 9034 | auto *SinkSucc = SinkRegion->getSingleSuccessor(); | ||||||
| 9035 | VPBlockUtils::disconnectBlocks(SinkPred, SinkRegion); | ||||||
| 9036 | VPBlockUtils::disconnectBlocks(SinkRegion, SinkSucc); | ||||||
| 9037 | VPBlockUtils::connectBlocks(SinkPred, SinkSucc); | ||||||
| 9038 | |||||||
| 9039 | if (TargetRegion) { | ||||||
| 9040 | // The target recipe is also in a replicate region, move the sink region | ||||||
| 9041 | // after the target region. | ||||||
| 9042 | auto *TargetSucc = TargetRegion->getSingleSuccessor(); | ||||||
| 9043 | VPBlockUtils::disconnectBlocks(TargetRegion, TargetSucc); | ||||||
| 9044 | VPBlockUtils::connectBlocks(TargetRegion, SinkRegion); | ||||||
| 9045 | VPBlockUtils::connectBlocks(SinkRegion, TargetSucc); | ||||||
| 9046 | } else { | ||||||
| 9047 | // The sink source is in a replicate region, we need to move the whole | ||||||
| 9048 | // replicate region, which should only contain a single recipe in the | ||||||
| 9049 | // main block. | ||||||
| 9050 | auto *SplitBlock = | ||||||
| 9051 | Target->getParent()->splitAt(std::next(Target->getIterator())); | ||||||
| 9052 | |||||||
| 9053 | auto *SplitPred = SplitBlock->getSinglePredecessor(); | ||||||
| 9054 | |||||||
| 9055 | VPBlockUtils::disconnectBlocks(SplitPred, SplitBlock); | ||||||
| 9056 | VPBlockUtils::connectBlocks(SplitPred, SinkRegion); | ||||||
| 9057 | VPBlockUtils::connectBlocks(SinkRegion, SplitBlock); | ||||||
| 9058 | } | ||||||
| 9059 | } | ||||||
| 9060 | |||||||
| 9061 | VPlanTransforms::removeRedundantCanonicalIVs(*Plan); | ||||||
| 9062 | VPlanTransforms::removeRedundantInductionCasts(*Plan); | ||||||
| 9063 | |||||||
| 9064 | // Now that sink-after is done, move induction recipes for optimized truncates | ||||||
| 9065 | // to the phi section of the header block. | ||||||
| 9066 | for (VPWidenIntOrFpInductionRecipe *Ind : InductionsToMove) | ||||||
| 9067 | Ind->moveBefore(*HeaderVPBB, HeaderVPBB->getFirstNonPhi()); | ||||||
| 9068 | |||||||
| 9069 | // Adjust the recipes for any inloop reductions. | ||||||
| 9070 | adjustRecipesForReductions(cast<VPBasicBlock>(TopRegion->getExiting()), Plan, | ||||||
| 9071 | RecipeBuilder, Range.Start); | ||||||
| 9072 | |||||||
| 9073 | // Introduce a recipe to combine the incoming and previous values of a | ||||||
| 9074 | // fixed-order recurrence. | ||||||
| 9075 | for (VPRecipeBase &R : | ||||||
| 9076 | Plan->getVectorLoopRegion()->getEntryBasicBlock()->phis()) { | ||||||
| 9077 | auto *RecurPhi = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&R); | ||||||
| 9078 | if (!RecurPhi) | ||||||
| 9079 | continue; | ||||||
| 9080 | |||||||
| 9081 | VPRecipeBase *PrevRecipe = &RecurPhi->getBackedgeRecipe(); | ||||||
| 9082 | // Fixed-order recurrences do not contain cycles, so this loop is guaranteed | ||||||
| 9083 | // to terminate. | ||||||
| 9084 | while (auto *PrevPhi = | ||||||
| 9085 | dyn_cast<VPFirstOrderRecurrencePHIRecipe>(PrevRecipe)) | ||||||
| 9086 | PrevRecipe = &PrevPhi->getBackedgeRecipe(); | ||||||
| 9087 | VPBasicBlock *InsertBlock = PrevRecipe->getParent(); | ||||||
| 9088 | auto *Region = GetReplicateRegion(PrevRecipe); | ||||||
| 9089 | if (Region) | ||||||
| 9090 | InsertBlock = dyn_cast<VPBasicBlock>(Region->getSingleSuccessor()); | ||||||
| 9091 | if (!InsertBlock) { | ||||||
| 9092 | InsertBlock = new VPBasicBlock(Region->getName() + ".succ"); | ||||||
| 9093 | VPBlockUtils::insertBlockAfter(InsertBlock, Region); | ||||||
| 9094 | } | ||||||
| 9095 | if (Region || PrevRecipe->isPhi()) | ||||||
| 9096 | Builder.setInsertPoint(InsertBlock, InsertBlock->getFirstNonPhi()); | ||||||
| 9097 | else | ||||||
| 9098 | Builder.setInsertPoint(InsertBlock, std::next(PrevRecipe->getIterator())); | ||||||
| 9099 | |||||||
| 9100 | auto *RecurSplice = cast<VPInstruction>( | ||||||
| 9101 | Builder.createNaryOp(VPInstruction::FirstOrderRecurrenceSplice, | ||||||
| 9102 | {RecurPhi, RecurPhi->getBackedgeValue()})); | ||||||
| 9103 | |||||||
| 9104 | RecurPhi->replaceAllUsesWith(RecurSplice); | ||||||
| 9105 | // Set the first operand of RecurSplice to RecurPhi again, after replacing | ||||||
| 9106 | // all users. | ||||||
| 9107 | RecurSplice->setOperand(0, RecurPhi); | ||||||
| 9108 | } | ||||||
| 9109 | |||||||
| 9110 | // Interleave memory: for each Interleave Group we marked earlier as relevant | ||||||
| 9111 | // for this VPlan, replace the Recipes widening its memory instructions with a | ||||||
| 9112 | // single VPInterleaveRecipe at its insertion point. | ||||||
| 9113 | for (const auto *IG : InterleaveGroups) { | ||||||
| 9114 | auto *Recipe = cast<VPWidenMemoryInstructionRecipe>( | ||||||
| 9115 | RecipeBuilder.getRecipe(IG->getInsertPos())); | ||||||
| 9116 | SmallVector<VPValue *, 4> StoredValues; | ||||||
| 9117 | for (unsigned i = 0; i < IG->getFactor(); ++i) | ||||||
| 9118 | if (auto *SI = dyn_cast_or_null<StoreInst>(IG->getMember(i))) { | ||||||
| 9119 | auto *StoreR = | ||||||
| 9120 | cast<VPWidenMemoryInstructionRecipe>(RecipeBuilder.getRecipe(SI)); | ||||||
| 9121 | StoredValues.push_back(StoreR->getStoredValue()); | ||||||
| 9122 | } | ||||||
| 9123 | |||||||
| 9124 | auto *VPIG = new VPInterleaveRecipe(IG, Recipe->getAddr(), StoredValues, | ||||||
| 9125 | Recipe->getMask()); | ||||||
| 9126 | VPIG->insertBefore(Recipe); | ||||||
| 9127 | unsigned J = 0; | ||||||
| 9128 | for (unsigned i = 0; i < IG->getFactor(); ++i) | ||||||
| 9129 | if (Instruction *Member = IG->getMember(i)) { | ||||||
| 9130 | if (!Member->getType()->isVoidTy()) { | ||||||
| 9131 | VPValue *OriginalV = Plan->getVPValue(Member); | ||||||
| 9132 | Plan->removeVPValueFor(Member); | ||||||
| 9133 | Plan->addVPValue(Member, VPIG->getVPValue(J)); | ||||||
| 9134 | OriginalV->replaceAllUsesWith(VPIG->getVPValue(J)); | ||||||
| 9135 | J++; | ||||||
| 9136 | } | ||||||
| 9137 | RecipeBuilder.getRecipe(Member)->eraseFromParent(); | ||||||
| 9138 | } | ||||||
| 9139 | } | ||||||
| 9140 | |||||||
| 9141 | for (ElementCount VF = Range.Start; ElementCount::isKnownLT(VF, Range.End); | ||||||
| 9142 | VF *= 2) | ||||||
| 9143 | Plan->addVF(VF); | ||||||
| 9144 | Plan->setName("Initial VPlan"); | ||||||
| 9145 | |||||||
| 9146 | // From this point onwards, VPlan-to-VPlan transformations may change the plan | ||||||
| 9147 | // in ways that accessing values using original IR values is incorrect. | ||||||
| 9148 | Plan->disableValue2VPValue(); | ||||||
| 9149 | |||||||
| 9150 | VPlanTransforms::optimizeInductions(*Plan, *PSE.getSE()); | ||||||
| 9151 | VPlanTransforms::removeDeadRecipes(*Plan); | ||||||
| 9152 | VPlanTransforms::sinkScalarOperands(*Plan); | ||||||
| 9153 | VPlanTransforms::mergeReplicateRegions(*Plan); | ||||||
| 9154 | VPlanTransforms::removeRedundantExpandSCEVRecipes(*Plan); | ||||||
| 9155 | |||||||
| 9156 | // Fold Exit block into its predecessor if possible. | ||||||
| 9157 | // TODO: Fold block earlier once all VPlan transforms properly maintain a | ||||||
| 9158 | // VPBasicBlock as exit. | ||||||
| 9159 | VPBlockUtils::tryToMergeBlockIntoPredecessor(TopRegion->getExiting()); | ||||||
| 9160 | |||||||
| 9161 | assert(VPlanVerifier::verifyPlanIsValid(*Plan) && "VPlan is invalid")(static_cast <bool> (VPlanVerifier::verifyPlanIsValid(* Plan) && "VPlan is invalid") ? void (0) : __assert_fail ("VPlanVerifier::verifyPlanIsValid(*Plan) && \"VPlan is invalid\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9161, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9162 | return Plan; | ||||||
| 9163 | } | ||||||
| 9164 | |||||||
| 9165 | VPlanPtr LoopVectorizationPlanner::buildVPlan(VFRange &Range) { | ||||||
| 9166 | // Outer loop handling: They may require CFG and instruction level | ||||||
| 9167 | // transformations before even evaluating whether vectorization is profitable. | ||||||
| 9168 | // Since we cannot modify the incoming IR, we need to build VPlan upfront in | ||||||
| 9169 | // the vectorization pipeline. | ||||||
| 9170 | assert(!OrigLoop->isInnermost())(static_cast <bool> (!OrigLoop->isInnermost()) ? void (0) : __assert_fail ("!OrigLoop->isInnermost()", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 9170, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9171 | assert(EnableVPlanNativePath && "VPlan-native path is not enabled.")(static_cast <bool> (EnableVPlanNativePath && "VPlan-native path is not enabled." ) ? void (0) : __assert_fail ("EnableVPlanNativePath && \"VPlan-native path is not enabled.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9171, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9172 | |||||||
| 9173 | // Create new empty VPlan | ||||||
| 9174 | auto Plan = std::make_unique<VPlan>(); | ||||||
| 9175 | |||||||
| 9176 | // Build hierarchical CFG | ||||||
| 9177 | VPlanHCFGBuilder HCFGBuilder(OrigLoop, LI, *Plan); | ||||||
| 9178 | HCFGBuilder.buildHierarchicalCFG(); | ||||||
| 9179 | |||||||
| 9180 | for (ElementCount VF = Range.Start; ElementCount::isKnownLT(VF, Range.End); | ||||||
| 9181 | VF *= 2) | ||||||
| 9182 | Plan->addVF(VF); | ||||||
| 9183 | |||||||
| 9184 | SmallPtrSet<Instruction *, 1> DeadInstructions; | ||||||
| 9185 | VPlanTransforms::VPInstructionsToVPRecipes( | ||||||
| 9186 | OrigLoop, Plan, | ||||||
| 9187 | [this](PHINode *P) { return Legal->getIntOrFpInductionDescriptor(P); }, | ||||||
| 9188 | DeadInstructions, *PSE.getSE(), *TLI); | ||||||
| 9189 | |||||||
| 9190 | // Remove the existing terminator of the exiting block of the top-most region. | ||||||
| 9191 | // A BranchOnCount will be added instead when adding the canonical IV recipes. | ||||||
| 9192 | auto *Term = | ||||||
| 9193 | Plan->getVectorLoopRegion()->getExitingBasicBlock()->getTerminator(); | ||||||
| 9194 | Term->eraseFromParent(); | ||||||
| 9195 | |||||||
| 9196 | addCanonicalIVRecipes(*Plan, Legal->getWidestInductionType(), DebugLoc(), | ||||||
| 9197 | true, CM.useActiveLaneMaskForControlFlow()); | ||||||
| 9198 | return Plan; | ||||||
| 9199 | } | ||||||
| 9200 | |||||||
| 9201 | // Adjust the recipes for reductions. For in-loop reductions the chain of | ||||||
| 9202 | // instructions leading from the loop exit instr to the phi need to be converted | ||||||
| 9203 | // to reductions, with one operand being vector and the other being the scalar | ||||||
| 9204 | // reduction chain. For other reductions, a select is introduced between the phi | ||||||
| 9205 | // and live-out recipes when folding the tail. | ||||||
| 9206 | void LoopVectorizationPlanner::adjustRecipesForReductions( | ||||||
| 9207 | VPBasicBlock *LatchVPBB, VPlanPtr &Plan, VPRecipeBuilder &RecipeBuilder, | ||||||
| 9208 | ElementCount MinVF) { | ||||||
| 9209 | for (const auto &Reduction : CM.getInLoopReductionChains()) { | ||||||
| 9210 | PHINode *Phi = Reduction.first; | ||||||
| 9211 | const RecurrenceDescriptor &RdxDesc = | ||||||
| 9212 | Legal->getReductionVars().find(Phi)->second; | ||||||
| 9213 | const SmallVector<Instruction *, 4> &ReductionOperations = Reduction.second; | ||||||
| 9214 | |||||||
| 9215 | if (MinVF.isScalar() && !CM.useOrderedReductions(RdxDesc)) | ||||||
| 9216 | continue; | ||||||
| 9217 | |||||||
| 9218 | // ReductionOperations are orders top-down from the phi's use to the | ||||||
| 9219 | // LoopExitValue. We keep a track of the previous item (the Chain) to tell | ||||||
| 9220 | // which of the two operands will remain scalar and which will be reduced. | ||||||
| 9221 | // For minmax the chain will be the select instructions. | ||||||
| 9222 | Instruction *Chain = Phi; | ||||||
| 9223 | for (Instruction *R : ReductionOperations) { | ||||||
| 9224 | VPRecipeBase *WidenRecipe = RecipeBuilder.getRecipe(R); | ||||||
| 9225 | RecurKind Kind = RdxDesc.getRecurrenceKind(); | ||||||
| 9226 | |||||||
| 9227 | VPValue *ChainOp = Plan->getVPValue(Chain); | ||||||
| 9228 | unsigned FirstOpId; | ||||||
| 9229 | assert(!RecurrenceDescriptor::isSelectCmpRecurrenceKind(Kind) &&(static_cast <bool> (!RecurrenceDescriptor::isSelectCmpRecurrenceKind (Kind) && "Only min/max recurrences allowed for inloop reductions" ) ? void (0) : __assert_fail ("!RecurrenceDescriptor::isSelectCmpRecurrenceKind(Kind) && \"Only min/max recurrences allowed for inloop reductions\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9230, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9230 | "Only min/max recurrences allowed for inloop reductions")(static_cast <bool> (!RecurrenceDescriptor::isSelectCmpRecurrenceKind (Kind) && "Only min/max recurrences allowed for inloop reductions" ) ? void (0) : __assert_fail ("!RecurrenceDescriptor::isSelectCmpRecurrenceKind(Kind) && \"Only min/max recurrences allowed for inloop reductions\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9230, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9231 | // Recognize a call to the llvm.fmuladd intrinsic. | ||||||
| 9232 | bool IsFMulAdd = (Kind == RecurKind::FMulAdd); | ||||||
| 9233 | assert((!IsFMulAdd || RecurrenceDescriptor::isFMulAddIntrinsic(R)) &&(static_cast <bool> ((!IsFMulAdd || RecurrenceDescriptor ::isFMulAddIntrinsic(R)) && "Expected instruction to be a call to the llvm.fmuladd intrinsic" ) ? void (0) : __assert_fail ("(!IsFMulAdd || RecurrenceDescriptor::isFMulAddIntrinsic(R)) && \"Expected instruction to be a call to the llvm.fmuladd intrinsic\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9234, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9234 | "Expected instruction to be a call to the llvm.fmuladd intrinsic")(static_cast <bool> ((!IsFMulAdd || RecurrenceDescriptor ::isFMulAddIntrinsic(R)) && "Expected instruction to be a call to the llvm.fmuladd intrinsic" ) ? void (0) : __assert_fail ("(!IsFMulAdd || RecurrenceDescriptor::isFMulAddIntrinsic(R)) && \"Expected instruction to be a call to the llvm.fmuladd intrinsic\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9234, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9235 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { | ||||||
| 9236 | assert(isa<VPWidenSelectRecipe>(WidenRecipe) &&(static_cast <bool> (isa<VPWidenSelectRecipe>(WidenRecipe ) && "Expected to replace a VPWidenSelectSC") ? void ( 0) : __assert_fail ("isa<VPWidenSelectRecipe>(WidenRecipe) && \"Expected to replace a VPWidenSelectSC\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9237, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9237 | "Expected to replace a VPWidenSelectSC")(static_cast <bool> (isa<VPWidenSelectRecipe>(WidenRecipe ) && "Expected to replace a VPWidenSelectSC") ? void ( 0) : __assert_fail ("isa<VPWidenSelectRecipe>(WidenRecipe) && \"Expected to replace a VPWidenSelectSC\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9237, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9238 | FirstOpId = 1; | ||||||
| 9239 | } else { | ||||||
| 9240 | assert((MinVF.isScalar() || isa<VPWidenRecipe>(WidenRecipe) ||(static_cast <bool> ((MinVF.isScalar() || isa<VPWidenRecipe >(WidenRecipe) || (IsFMulAdd && isa<VPWidenCallRecipe >(WidenRecipe))) && "Expected to replace a VPWidenSC" ) ? void (0) : __assert_fail ("(MinVF.isScalar() || isa<VPWidenRecipe>(WidenRecipe) || (IsFMulAdd && isa<VPWidenCallRecipe>(WidenRecipe))) && \"Expected to replace a VPWidenSC\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9242, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9241 | (IsFMulAdd && isa<VPWidenCallRecipe>(WidenRecipe))) &&(static_cast <bool> ((MinVF.isScalar() || isa<VPWidenRecipe >(WidenRecipe) || (IsFMulAdd && isa<VPWidenCallRecipe >(WidenRecipe))) && "Expected to replace a VPWidenSC" ) ? void (0) : __assert_fail ("(MinVF.isScalar() || isa<VPWidenRecipe>(WidenRecipe) || (IsFMulAdd && isa<VPWidenCallRecipe>(WidenRecipe))) && \"Expected to replace a VPWidenSC\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9242, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9242 | "Expected to replace a VPWidenSC")(static_cast <bool> ((MinVF.isScalar() || isa<VPWidenRecipe >(WidenRecipe) || (IsFMulAdd && isa<VPWidenCallRecipe >(WidenRecipe))) && "Expected to replace a VPWidenSC" ) ? void (0) : __assert_fail ("(MinVF.isScalar() || isa<VPWidenRecipe>(WidenRecipe) || (IsFMulAdd && isa<VPWidenCallRecipe>(WidenRecipe))) && \"Expected to replace a VPWidenSC\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9242, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9243 | FirstOpId = 0; | ||||||
| 9244 | } | ||||||
| 9245 | unsigned VecOpId = | ||||||
| 9246 | R->getOperand(FirstOpId) == Chain ? FirstOpId + 1 : FirstOpId; | ||||||
| 9247 | VPValue *VecOp = Plan->getVPValue(R->getOperand(VecOpId)); | ||||||
| 9248 | |||||||
| 9249 | auto *CondOp = CM.blockNeedsPredicationForAnyReason(R->getParent()) | ||||||
| 9250 | ? RecipeBuilder.createBlockInMask(R->getParent(), Plan) | ||||||
| 9251 | : nullptr; | ||||||
| 9252 | |||||||
| 9253 | if (IsFMulAdd) { | ||||||
| 9254 | // If the instruction is a call to the llvm.fmuladd intrinsic then we | ||||||
| 9255 | // need to create an fmul recipe to use as the vector operand for the | ||||||
| 9256 | // fadd reduction. | ||||||
| 9257 | VPInstruction *FMulRecipe = new VPInstruction( | ||||||
| 9258 | Instruction::FMul, {VecOp, Plan->getVPValue(R->getOperand(1))}); | ||||||
| 9259 | FMulRecipe->setFastMathFlags(R->getFastMathFlags()); | ||||||
| 9260 | WidenRecipe->getParent()->insert(FMulRecipe, | ||||||
| 9261 | WidenRecipe->getIterator()); | ||||||
| 9262 | VecOp = FMulRecipe; | ||||||
| 9263 | } | ||||||
| 9264 | VPReductionRecipe *RedRecipe = | ||||||
| 9265 | new VPReductionRecipe(&RdxDesc, R, ChainOp, VecOp, CondOp, TTI); | ||||||
| 9266 | WidenRecipe->getVPSingleValue()->replaceAllUsesWith(RedRecipe); | ||||||
| 9267 | Plan->removeVPValueFor(R); | ||||||
| 9268 | Plan->addVPValue(R, RedRecipe); | ||||||
| 9269 | // Append the recipe to the end of the VPBasicBlock because we need to | ||||||
| 9270 | // ensure that it comes after all of it's inputs, including CondOp. | ||||||
| 9271 | WidenRecipe->getParent()->appendRecipe(RedRecipe); | ||||||
| 9272 | WidenRecipe->getVPSingleValue()->replaceAllUsesWith(RedRecipe); | ||||||
| 9273 | WidenRecipe->eraseFromParent(); | ||||||
| 9274 | |||||||
| 9275 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { | ||||||
| 9276 | VPRecipeBase *CompareRecipe = | ||||||
| 9277 | RecipeBuilder.getRecipe(cast<Instruction>(R->getOperand(0))); | ||||||
| 9278 | assert(isa<VPWidenRecipe>(CompareRecipe) &&(static_cast <bool> (isa<VPWidenRecipe>(CompareRecipe ) && "Expected to replace a VPWidenSC") ? void (0) : __assert_fail ("isa<VPWidenRecipe>(CompareRecipe) && \"Expected to replace a VPWidenSC\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9279, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9279 | "Expected to replace a VPWidenSC")(static_cast <bool> (isa<VPWidenRecipe>(CompareRecipe ) && "Expected to replace a VPWidenSC") ? void (0) : __assert_fail ("isa<VPWidenRecipe>(CompareRecipe) && \"Expected to replace a VPWidenSC\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9279, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9280 | assert(cast<VPWidenRecipe>(CompareRecipe)->getNumUsers() == 0 &&(static_cast <bool> (cast<VPWidenRecipe>(CompareRecipe )->getNumUsers() == 0 && "Expected no remaining users" ) ? void (0) : __assert_fail ("cast<VPWidenRecipe>(CompareRecipe)->getNumUsers() == 0 && \"Expected no remaining users\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9281, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9281 | "Expected no remaining users")(static_cast <bool> (cast<VPWidenRecipe>(CompareRecipe )->getNumUsers() == 0 && "Expected no remaining users" ) ? void (0) : __assert_fail ("cast<VPWidenRecipe>(CompareRecipe)->getNumUsers() == 0 && \"Expected no remaining users\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9281, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9282 | CompareRecipe->eraseFromParent(); | ||||||
| 9283 | } | ||||||
| 9284 | Chain = R; | ||||||
| 9285 | } | ||||||
| 9286 | } | ||||||
| 9287 | |||||||
| 9288 | // If tail is folded by masking, introduce selects between the phi | ||||||
| 9289 | // and the live-out instruction of each reduction, at the beginning of the | ||||||
| 9290 | // dedicated latch block. | ||||||
| 9291 | if (CM.foldTailByMasking()) { | ||||||
| 9292 | Builder.setInsertPoint(LatchVPBB, LatchVPBB->begin()); | ||||||
| 9293 | for (VPRecipeBase &R : | ||||||
| 9294 | Plan->getVectorLoopRegion()->getEntryBasicBlock()->phis()) { | ||||||
| 9295 | VPReductionPHIRecipe *PhiR = dyn_cast<VPReductionPHIRecipe>(&R); | ||||||
| 9296 | if (!PhiR || PhiR->isInLoop()) | ||||||
| 9297 | continue; | ||||||
| 9298 | VPValue *Cond = | ||||||
| 9299 | RecipeBuilder.createBlockInMask(OrigLoop->getHeader(), Plan); | ||||||
| 9300 | VPValue *Red = PhiR->getBackedgeValue(); | ||||||
| 9301 | assert(Red->getDefiningRecipe()->getParent() != LatchVPBB &&(static_cast <bool> (Red->getDefiningRecipe()->getParent () != LatchVPBB && "reduction recipe must be defined before latch" ) ? void (0) : __assert_fail ("Red->getDefiningRecipe()->getParent() != LatchVPBB && \"reduction recipe must be defined before latch\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9302, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9302 | "reduction recipe must be defined before latch")(static_cast <bool> (Red->getDefiningRecipe()->getParent () != LatchVPBB && "reduction recipe must be defined before latch" ) ? void (0) : __assert_fail ("Red->getDefiningRecipe()->getParent() != LatchVPBB && \"reduction recipe must be defined before latch\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9302, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9303 | Builder.createNaryOp(Instruction::Select, {Cond, Red, PhiR}); | ||||||
| 9304 | } | ||||||
| 9305 | } | ||||||
| 9306 | } | ||||||
| 9307 | |||||||
| 9308 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||
| 9309 | void VPInterleaveRecipe::print(raw_ostream &O, const Twine &Indent, | ||||||
| 9310 | VPSlotTracker &SlotTracker) const { | ||||||
| 9311 | O << Indent << "INTERLEAVE-GROUP with factor " << IG->getFactor() << " at "; | ||||||
| 9312 | IG->getInsertPos()->printAsOperand(O, false); | ||||||
| 9313 | O << ", "; | ||||||
| 9314 | getAddr()->printAsOperand(O, SlotTracker); | ||||||
| 9315 | VPValue *Mask = getMask(); | ||||||
| 9316 | if (Mask) { | ||||||
| 9317 | O << ", "; | ||||||
| 9318 | Mask->printAsOperand(O, SlotTracker); | ||||||
| 9319 | } | ||||||
| 9320 | |||||||
| 9321 | unsigned OpIdx = 0; | ||||||
| 9322 | for (unsigned i = 0; i < IG->getFactor(); ++i) { | ||||||
| 9323 | if (!IG->getMember(i)) | ||||||
| 9324 | continue; | ||||||
| 9325 | if (getNumStoreOperands() > 0) { | ||||||
| 9326 | O << "\n" << Indent << " store "; | ||||||
| 9327 | getOperand(1 + OpIdx)->printAsOperand(O, SlotTracker); | ||||||
| 9328 | O << " to index " << i; | ||||||
| 9329 | } else { | ||||||
| 9330 | O << "\n" << Indent << " "; | ||||||
| 9331 | getVPValue(OpIdx)->printAsOperand(O, SlotTracker); | ||||||
| 9332 | O << " = load from index " << i; | ||||||
| 9333 | } | ||||||
| 9334 | ++OpIdx; | ||||||
| 9335 | } | ||||||
| 9336 | } | ||||||
| 9337 | #endif | ||||||
| 9338 | |||||||
| 9339 | void VPWidenIntOrFpInductionRecipe::execute(VPTransformState &State) { | ||||||
| 9340 | assert(!State.Instance && "Int or FP induction being replicated.")(static_cast <bool> (!State.Instance && "Int or FP induction being replicated." ) ? void (0) : __assert_fail ("!State.Instance && \"Int or FP induction being replicated.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9340, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9341 | |||||||
| 9342 | Value *Start = getStartValue()->getLiveInIRValue(); | ||||||
| 9343 | const InductionDescriptor &ID = getInductionDescriptor(); | ||||||
| 9344 | TruncInst *Trunc = getTruncInst(); | ||||||
| 9345 | IRBuilderBase &Builder = State.Builder; | ||||||
| 9346 | assert(IV->getType() == ID.getStartValue()->getType() && "Types must match")(static_cast <bool> (IV->getType() == ID.getStartValue ()->getType() && "Types must match") ? void (0) : __assert_fail ("IV->getType() == ID.getStartValue()->getType() && \"Types must match\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9346, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9347 | assert(State.VF.isVector() && "must have vector VF")(static_cast <bool> (State.VF.isVector() && "must have vector VF" ) ? void (0) : __assert_fail ("State.VF.isVector() && \"must have vector VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9347, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9348 | |||||||
| 9349 | // The value from the original loop to which we are mapping the new induction | ||||||
| 9350 | // variable. | ||||||
| 9351 | Instruction *EntryVal = Trunc ? cast<Instruction>(Trunc) : IV; | ||||||
| 9352 | |||||||
| 9353 | // Fast-math-flags propagate from the original induction instruction. | ||||||
| 9354 | IRBuilder<>::FastMathFlagGuard FMFG(Builder); | ||||||
| 9355 | if (ID.getInductionBinOp() && isa<FPMathOperator>(ID.getInductionBinOp())) | ||||||
| 9356 | Builder.setFastMathFlags(ID.getInductionBinOp()->getFastMathFlags()); | ||||||
| 9357 | |||||||
| 9358 | // Now do the actual transformations, and start with fetching the step value. | ||||||
| 9359 | Value *Step = State.get(getStepValue(), VPIteration(0, 0)); | ||||||
| 9360 | |||||||
| 9361 | assert((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) &&(static_cast <bool> ((isa<PHINode>(EntryVal) || isa <TruncInst>(EntryVal)) && "Expected either an induction phi-node or a truncate of it!" ) ? void (0) : __assert_fail ("(isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) && \"Expected either an induction phi-node or a truncate of it!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9362, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9362 | "Expected either an induction phi-node or a truncate of it!")(static_cast <bool> ((isa<PHINode>(EntryVal) || isa <TruncInst>(EntryVal)) && "Expected either an induction phi-node or a truncate of it!" ) ? void (0) : __assert_fail ("(isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) && \"Expected either an induction phi-node or a truncate of it!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9362, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9363 | |||||||
| 9364 | // Construct the initial value of the vector IV in the vector loop preheader | ||||||
| 9365 | auto CurrIP = Builder.saveIP(); | ||||||
| 9366 | BasicBlock *VectorPH = State.CFG.getPreheaderBBFor(this); | ||||||
| 9367 | Builder.SetInsertPoint(VectorPH->getTerminator()); | ||||||
| 9368 | if (isa<TruncInst>(EntryVal)) { | ||||||
| 9369 | assert(Start->getType()->isIntegerTy() &&(static_cast <bool> (Start->getType()->isIntegerTy () && "Truncation requires an integer type") ? void ( 0) : __assert_fail ("Start->getType()->isIntegerTy() && \"Truncation requires an integer type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9370, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9370 | "Truncation requires an integer type")(static_cast <bool> (Start->getType()->isIntegerTy () && "Truncation requires an integer type") ? void ( 0) : __assert_fail ("Start->getType()->isIntegerTy() && \"Truncation requires an integer type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9370, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9371 | auto *TruncType = cast<IntegerType>(EntryVal->getType()); | ||||||
| 9372 | Step = Builder.CreateTrunc(Step, TruncType); | ||||||
| 9373 | Start = Builder.CreateCast(Instruction::Trunc, Start, TruncType); | ||||||
| 9374 | } | ||||||
| 9375 | |||||||
| 9376 | Value *Zero = getSignedIntOrFpConstant(Start->getType(), 0); | ||||||
| 9377 | Value *SplatStart = Builder.CreateVectorSplat(State.VF, Start); | ||||||
| 9378 | Value *SteppedStart = getStepVector( | ||||||
| 9379 | SplatStart, Zero, Step, ID.getInductionOpcode(), State.VF, State.Builder); | ||||||
| 9380 | |||||||
| 9381 | // We create vector phi nodes for both integer and floating-point induction | ||||||
| 9382 | // variables. Here, we determine the kind of arithmetic we will perform. | ||||||
| 9383 | Instruction::BinaryOps AddOp; | ||||||
| 9384 | Instruction::BinaryOps MulOp; | ||||||
| 9385 | if (Step->getType()->isIntegerTy()) { | ||||||
| 9386 | AddOp = Instruction::Add; | ||||||
| 9387 | MulOp = Instruction::Mul; | ||||||
| 9388 | } else { | ||||||
| 9389 | AddOp = ID.getInductionOpcode(); | ||||||
| 9390 | MulOp = Instruction::FMul; | ||||||
| 9391 | } | ||||||
| 9392 | |||||||
| 9393 | // Multiply the vectorization factor by the step using integer or | ||||||
| 9394 | // floating-point arithmetic as appropriate. | ||||||
| 9395 | Type *StepType = Step->getType(); | ||||||
| 9396 | Value *RuntimeVF; | ||||||
| 9397 | if (Step->getType()->isFloatingPointTy()) | ||||||
| 9398 | RuntimeVF = getRuntimeVFAsFloat(Builder, StepType, State.VF); | ||||||
| 9399 | else | ||||||
| 9400 | RuntimeVF = getRuntimeVF(Builder, StepType, State.VF); | ||||||
| 9401 | Value *Mul = Builder.CreateBinOp(MulOp, Step, RuntimeVF); | ||||||
| 9402 | |||||||
| 9403 | // Create a vector splat to use in the induction update. | ||||||
| 9404 | // | ||||||
| 9405 | // FIXME: If the step is non-constant, we create the vector splat with | ||||||
| 9406 | // IRBuilder. IRBuilder can constant-fold the multiply, but it doesn't | ||||||
| 9407 | // handle a constant vector splat. | ||||||
| 9408 | Value *SplatVF = isa<Constant>(Mul) | ||||||
| 9409 | ? ConstantVector::getSplat(State.VF, cast<Constant>(Mul)) | ||||||
| 9410 | : Builder.CreateVectorSplat(State.VF, Mul); | ||||||
| 9411 | Builder.restoreIP(CurrIP); | ||||||
| 9412 | |||||||
| 9413 | // We may need to add the step a number of times, depending on the unroll | ||||||
| 9414 | // factor. The last of those goes into the PHI. | ||||||
| 9415 | PHINode *VecInd = PHINode::Create(SteppedStart->getType(), 2, "vec.ind", | ||||||
| 9416 | &*State.CFG.PrevBB->getFirstInsertionPt()); | ||||||
| 9417 | VecInd->setDebugLoc(EntryVal->getDebugLoc()); | ||||||
| 9418 | Instruction *LastInduction = VecInd; | ||||||
| 9419 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
| 9420 | State.set(this, LastInduction, Part); | ||||||
| 9421 | |||||||
| 9422 | if (isa<TruncInst>(EntryVal)) | ||||||
| 9423 | State.addMetadata(LastInduction, EntryVal); | ||||||
| 9424 | |||||||
| 9425 | LastInduction = cast<Instruction>( | ||||||
| 9426 | Builder.CreateBinOp(AddOp, LastInduction, SplatVF, "step.add")); | ||||||
| 9427 | LastInduction->setDebugLoc(EntryVal->getDebugLoc()); | ||||||
| 9428 | } | ||||||
| 9429 | |||||||
| 9430 | LastInduction->setName("vec.ind.next"); | ||||||
| 9431 | VecInd->addIncoming(SteppedStart, VectorPH); | ||||||
| 9432 | // Add induction update using an incorrect block temporarily. The phi node | ||||||
| 9433 | // will be fixed after VPlan execution. Note that at this point the latch | ||||||
| 9434 | // block cannot be used, as it does not exist yet. | ||||||
| 9435 | // TODO: Model increment value in VPlan, by turning the recipe into a | ||||||
| 9436 | // multi-def and a subclass of VPHeaderPHIRecipe. | ||||||
| 9437 | VecInd->addIncoming(LastInduction, VectorPH); | ||||||
| 9438 | } | ||||||
| 9439 | |||||||
| 9440 | void VPWidenPointerInductionRecipe::execute(VPTransformState &State) { | ||||||
| 9441 | assert(IndDesc.getKind() == InductionDescriptor::IK_PtrInduction &&(static_cast <bool> (IndDesc.getKind() == InductionDescriptor ::IK_PtrInduction && "Not a pointer induction according to InductionDescriptor!" ) ? void (0) : __assert_fail ("IndDesc.getKind() == InductionDescriptor::IK_PtrInduction && \"Not a pointer induction according to InductionDescriptor!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9442, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9442 | "Not a pointer induction according to InductionDescriptor!")(static_cast <bool> (IndDesc.getKind() == InductionDescriptor ::IK_PtrInduction && "Not a pointer induction according to InductionDescriptor!" ) ? void (0) : __assert_fail ("IndDesc.getKind() == InductionDescriptor::IK_PtrInduction && \"Not a pointer induction according to InductionDescriptor!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9442, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9443 | assert(cast<PHINode>(getUnderlyingInstr())->getType()->isPointerTy() &&(static_cast <bool> (cast<PHINode>(getUnderlyingInstr ())->getType()->isPointerTy() && "Unexpected type." ) ? void (0) : __assert_fail ("cast<PHINode>(getUnderlyingInstr())->getType()->isPointerTy() && \"Unexpected type.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9444, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9444 | "Unexpected type.")(static_cast <bool> (cast<PHINode>(getUnderlyingInstr ())->getType()->isPointerTy() && "Unexpected type." ) ? void (0) : __assert_fail ("cast<PHINode>(getUnderlyingInstr())->getType()->isPointerTy() && \"Unexpected type.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9444, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9445 | |||||||
| 9446 | auto *IVR = getParent()->getPlan()->getCanonicalIV(); | ||||||
| 9447 | PHINode *CanonicalIV = cast<PHINode>(State.get(IVR, 0)); | ||||||
| 9448 | |||||||
| 9449 | if (onlyScalarsGenerated(State.VF)) { | ||||||
| 9450 | // This is the normalized GEP that starts counting at zero. | ||||||
| 9451 | Value *PtrInd = State.Builder.CreateSExtOrTrunc( | ||||||
| 9452 | CanonicalIV, IndDesc.getStep()->getType()); | ||||||
| 9453 | // Determine the number of scalars we need to generate for each unroll | ||||||
| 9454 | // iteration. If the instruction is uniform, we only need to generate the | ||||||
| 9455 | // first lane. Otherwise, we generate all VF values. | ||||||
| 9456 | bool IsUniform = vputils::onlyFirstLaneUsed(this); | ||||||
| 9457 | assert((IsUniform || !State.VF.isScalable()) &&(static_cast <bool> ((IsUniform || !State.VF.isScalable ()) && "Cannot scalarize a scalable VF") ? void (0) : __assert_fail ("(IsUniform || !State.VF.isScalable()) && \"Cannot scalarize a scalable VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9458, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9458 | "Cannot scalarize a scalable VF")(static_cast <bool> ((IsUniform || !State.VF.isScalable ()) && "Cannot scalarize a scalable VF") ? void (0) : __assert_fail ("(IsUniform || !State.VF.isScalable()) && \"Cannot scalarize a scalable VF\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9458, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9459 | unsigned Lanes = IsUniform ? 1 : State.VF.getFixedValue(); | ||||||
| 9460 | |||||||
| 9461 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
| 9462 | Value *PartStart = | ||||||
| 9463 | createStepForVF(State.Builder, PtrInd->getType(), State.VF, Part); | ||||||
| 9464 | |||||||
| 9465 | for (unsigned Lane = 0; Lane < Lanes; ++Lane) { | ||||||
| 9466 | Value *Idx = State.Builder.CreateAdd( | ||||||
| 9467 | PartStart, ConstantInt::get(PtrInd->getType(), Lane)); | ||||||
| 9468 | Value *GlobalIdx = State.Builder.CreateAdd(PtrInd, Idx); | ||||||
| 9469 | |||||||
| 9470 | Value *Step = State.get(getOperand(1), VPIteration(0, Part)); | ||||||
| 9471 | Value *SclrGep = emitTransformedIndex( | ||||||
| 9472 | State.Builder, GlobalIdx, IndDesc.getStartValue(), Step, IndDesc); | ||||||
| 9473 | SclrGep->setName("next.gep"); | ||||||
| 9474 | State.set(this, SclrGep, VPIteration(Part, Lane)); | ||||||
| 9475 | } | ||||||
| 9476 | } | ||||||
| 9477 | return; | ||||||
| 9478 | } | ||||||
| 9479 | |||||||
| 9480 | assert(isa<SCEVConstant>(IndDesc.getStep()) &&(static_cast <bool> (isa<SCEVConstant>(IndDesc.getStep ()) && "Induction step not a SCEV constant!") ? void ( 0) : __assert_fail ("isa<SCEVConstant>(IndDesc.getStep()) && \"Induction step not a SCEV constant!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9481, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9481 | "Induction step not a SCEV constant!")(static_cast <bool> (isa<SCEVConstant>(IndDesc.getStep ()) && "Induction step not a SCEV constant!") ? void ( 0) : __assert_fail ("isa<SCEVConstant>(IndDesc.getStep()) && \"Induction step not a SCEV constant!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9481, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9482 | Type *PhiType = IndDesc.getStep()->getType(); | ||||||
| 9483 | |||||||
| 9484 | // Build a pointer phi | ||||||
| 9485 | Value *ScalarStartValue = getStartValue()->getLiveInIRValue(); | ||||||
| 9486 | Type *ScStValueType = ScalarStartValue->getType(); | ||||||
| 9487 | PHINode *NewPointerPhi = | ||||||
| 9488 | PHINode::Create(ScStValueType, 2, "pointer.phi", CanonicalIV); | ||||||
| 9489 | |||||||
| 9490 | BasicBlock *VectorPH = State.CFG.getPreheaderBBFor(this); | ||||||
| 9491 | NewPointerPhi->addIncoming(ScalarStartValue, VectorPH); | ||||||
| 9492 | |||||||
| 9493 | // A pointer induction, performed by using a gep | ||||||
| 9494 | Instruction *InductionLoc = &*State.Builder.GetInsertPoint(); | ||||||
| 9495 | |||||||
| 9496 | Value *ScalarStepValue = State.get(getOperand(1), VPIteration(0, 0)); | ||||||
| 9497 | Value *RuntimeVF = getRuntimeVF(State.Builder, PhiType, State.VF); | ||||||
| 9498 | Value *NumUnrolledElems = | ||||||
| 9499 | State.Builder.CreateMul(RuntimeVF, ConstantInt::get(PhiType, State.UF)); | ||||||
| 9500 | Value *InductionGEP = GetElementPtrInst::Create( | ||||||
| 9501 | IndDesc.getElementType(), NewPointerPhi, | ||||||
| 9502 | State.Builder.CreateMul(ScalarStepValue, NumUnrolledElems), "ptr.ind", | ||||||
| 9503 | InductionLoc); | ||||||
| 9504 | // Add induction update using an incorrect block temporarily. The phi node | ||||||
| 9505 | // will be fixed after VPlan execution. Note that at this point the latch | ||||||
| 9506 | // block cannot be used, as it does not exist yet. | ||||||
| 9507 | // TODO: Model increment value in VPlan, by turning the recipe into a | ||||||
| 9508 | // multi-def and a subclass of VPHeaderPHIRecipe. | ||||||
| 9509 | NewPointerPhi->addIncoming(InductionGEP, VectorPH); | ||||||
| 9510 | |||||||
| 9511 | // Create UF many actual address geps that use the pointer | ||||||
| 9512 | // phi as base and a vectorized version of the step value | ||||||
| 9513 | // (<step*0, ..., step*N>) as offset. | ||||||
| 9514 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
| 9515 | Type *VecPhiType = VectorType::get(PhiType, State.VF); | ||||||
| 9516 | Value *StartOffsetScalar = | ||||||
| 9517 | State.Builder.CreateMul(RuntimeVF, ConstantInt::get(PhiType, Part)); | ||||||
| 9518 | Value *StartOffset = | ||||||
| 9519 | State.Builder.CreateVectorSplat(State.VF, StartOffsetScalar); | ||||||
| 9520 | // Create a vector of consecutive numbers from zero to VF. | ||||||
| 9521 | StartOffset = State.Builder.CreateAdd( | ||||||
| 9522 | StartOffset, State.Builder.CreateStepVector(VecPhiType)); | ||||||
| 9523 | |||||||
| 9524 | assert(ScalarStepValue == State.get(getOperand(1), VPIteration(0, Part)) &&(static_cast <bool> (ScalarStepValue == State.get(getOperand (1), VPIteration(0, Part)) && "scalar step must be the same across all parts" ) ? void (0) : __assert_fail ("ScalarStepValue == State.get(getOperand(1), VPIteration(0, Part)) && \"scalar step must be the same across all parts\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9525, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9525 | "scalar step must be the same across all parts")(static_cast <bool> (ScalarStepValue == State.get(getOperand (1), VPIteration(0, Part)) && "scalar step must be the same across all parts" ) ? void (0) : __assert_fail ("ScalarStepValue == State.get(getOperand(1), VPIteration(0, Part)) && \"scalar step must be the same across all parts\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9525, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9526 | Value *GEP = State.Builder.CreateGEP( | ||||||
| 9527 | IndDesc.getElementType(), NewPointerPhi, | ||||||
| 9528 | State.Builder.CreateMul( | ||||||
| 9529 | StartOffset, | ||||||
| 9530 | State.Builder.CreateVectorSplat(State.VF, ScalarStepValue), | ||||||
| 9531 | "vector.gep")); | ||||||
| 9532 | State.set(this, GEP, Part); | ||||||
| 9533 | } | ||||||
| 9534 | } | ||||||
| 9535 | |||||||
| 9536 | void VPDerivedIVRecipe::execute(VPTransformState &State) { | ||||||
| 9537 | assert(!State.Instance && "VPDerivedIVRecipe being replicated.")(static_cast <bool> (!State.Instance && "VPDerivedIVRecipe being replicated." ) ? void (0) : __assert_fail ("!State.Instance && \"VPDerivedIVRecipe being replicated.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9537, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9538 | |||||||
| 9539 | // Fast-math-flags propagate from the original induction instruction. | ||||||
| 9540 | IRBuilder<>::FastMathFlagGuard FMFG(State.Builder); | ||||||
| 9541 | if (IndDesc.getInductionBinOp() && | ||||||
| 9542 | isa<FPMathOperator>(IndDesc.getInductionBinOp())) | ||||||
| 9543 | State.Builder.setFastMathFlags( | ||||||
| 9544 | IndDesc.getInductionBinOp()->getFastMathFlags()); | ||||||
| 9545 | |||||||
| 9546 | Value *Step = State.get(getStepValue(), VPIteration(0, 0)); | ||||||
| 9547 | Value *CanonicalIV = State.get(getCanonicalIV(), VPIteration(0, 0)); | ||||||
| 9548 | Value *DerivedIV = | ||||||
| 9549 | emitTransformedIndex(State.Builder, CanonicalIV, | ||||||
| 9550 | getStartValue()->getLiveInIRValue(), Step, IndDesc); | ||||||
| 9551 | DerivedIV->setName("offset.idx"); | ||||||
| 9552 | if (ResultTy != DerivedIV->getType()) { | ||||||
| 9553 | assert(Step->getType()->isIntegerTy() &&(static_cast <bool> (Step->getType()->isIntegerTy () && "Truncation requires an integer step") ? void ( 0) : __assert_fail ("Step->getType()->isIntegerTy() && \"Truncation requires an integer step\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9554, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9554 | "Truncation requires an integer step")(static_cast <bool> (Step->getType()->isIntegerTy () && "Truncation requires an integer step") ? void ( 0) : __assert_fail ("Step->getType()->isIntegerTy() && \"Truncation requires an integer step\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9554, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9555 | DerivedIV = State.Builder.CreateTrunc(DerivedIV, ResultTy); | ||||||
| 9556 | } | ||||||
| 9557 | assert(DerivedIV != CanonicalIV && "IV didn't need transforming?")(static_cast <bool> (DerivedIV != CanonicalIV && "IV didn't need transforming?") ? void (0) : __assert_fail ( "DerivedIV != CanonicalIV && \"IV didn't need transforming?\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9557, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9558 | |||||||
| 9559 | State.set(this, DerivedIV, VPIteration(0, 0)); | ||||||
| 9560 | } | ||||||
| 9561 | |||||||
| 9562 | void VPScalarIVStepsRecipe::execute(VPTransformState &State) { | ||||||
| 9563 | // Fast-math-flags propagate from the original induction instruction. | ||||||
| 9564 | IRBuilder<>::FastMathFlagGuard FMFG(State.Builder); | ||||||
| 9565 | if (IndDesc.getInductionBinOp() && | ||||||
| 9566 | isa<FPMathOperator>(IndDesc.getInductionBinOp())) | ||||||
| 9567 | State.Builder.setFastMathFlags( | ||||||
| 9568 | IndDesc.getInductionBinOp()->getFastMathFlags()); | ||||||
| 9569 | |||||||
| 9570 | Value *BaseIV = State.get(getOperand(0), VPIteration(0, 0)); | ||||||
| 9571 | Value *Step = State.get(getStepValue(), VPIteration(0, 0)); | ||||||
| 9572 | |||||||
| 9573 | buildScalarSteps(BaseIV, Step, IndDesc, this, State); | ||||||
| 9574 | } | ||||||
| 9575 | |||||||
| 9576 | void VPInterleaveRecipe::execute(VPTransformState &State) { | ||||||
| 9577 | assert(!State.Instance && "Interleave group being replicated.")(static_cast <bool> (!State.Instance && "Interleave group being replicated." ) ? void (0) : __assert_fail ("!State.Instance && \"Interleave group being replicated.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9577, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9578 | State.ILV->vectorizeInterleaveGroup(IG, definedValues(), State, getAddr(), | ||||||
| 9579 | getStoredValues(), getMask()); | ||||||
| 9580 | } | ||||||
| 9581 | |||||||
| 9582 | void VPReductionRecipe::execute(VPTransformState &State) { | ||||||
| 9583 | assert(!State.Instance && "Reduction being replicated.")(static_cast <bool> (!State.Instance && "Reduction being replicated." ) ? void (0) : __assert_fail ("!State.Instance && \"Reduction being replicated.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9583, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9584 | Value *PrevInChain = State.get(getChainOp(), 0); | ||||||
| 9585 | RecurKind Kind = RdxDesc->getRecurrenceKind(); | ||||||
| 9586 | bool IsOrdered = State.ILV->useOrderedReductions(*RdxDesc); | ||||||
| 9587 | // Propagate the fast-math flags carried by the underlying instruction. | ||||||
| 9588 | IRBuilderBase::FastMathFlagGuard FMFGuard(State.Builder); | ||||||
| 9589 | State.Builder.setFastMathFlags(RdxDesc->getFastMathFlags()); | ||||||
| 9590 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
| 9591 | Value *NewVecOp = State.get(getVecOp(), Part); | ||||||
| 9592 | if (VPValue *Cond = getCondOp()) { | ||||||
| 9593 | Value *NewCond = State.get(Cond, Part); | ||||||
| 9594 | VectorType *VecTy = cast<VectorType>(NewVecOp->getType()); | ||||||
| 9595 | Value *Iden = RdxDesc->getRecurrenceIdentity( | ||||||
| 9596 | Kind, VecTy->getElementType(), RdxDesc->getFastMathFlags()); | ||||||
| 9597 | Value *IdenVec = | ||||||
| 9598 | State.Builder.CreateVectorSplat(VecTy->getElementCount(), Iden); | ||||||
| 9599 | Value *Select = State.Builder.CreateSelect(NewCond, NewVecOp, IdenVec); | ||||||
| 9600 | NewVecOp = Select; | ||||||
| 9601 | } | ||||||
| 9602 | Value *NewRed; | ||||||
| 9603 | Value *NextInChain; | ||||||
| 9604 | if (IsOrdered) { | ||||||
| 9605 | if (State.VF.isVector()) | ||||||
| 9606 | NewRed = createOrderedReduction(State.Builder, *RdxDesc, NewVecOp, | ||||||
| 9607 | PrevInChain); | ||||||
| 9608 | else | ||||||
| 9609 | NewRed = State.Builder.CreateBinOp( | ||||||
| 9610 | (Instruction::BinaryOps)RdxDesc->getOpcode(Kind), PrevInChain, | ||||||
| 9611 | NewVecOp); | ||||||
| 9612 | PrevInChain = NewRed; | ||||||
| 9613 | } else { | ||||||
| 9614 | PrevInChain = State.get(getChainOp(), Part); | ||||||
| 9615 | NewRed = createTargetReduction(State.Builder, TTI, *RdxDesc, NewVecOp); | ||||||
| 9616 | } | ||||||
| 9617 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { | ||||||
| 9618 | NextInChain = | ||||||
| 9619 | createMinMaxOp(State.Builder, RdxDesc->getRecurrenceKind(), | ||||||
| 9620 | NewRed, PrevInChain); | ||||||
| 9621 | } else if (IsOrdered) | ||||||
| 9622 | NextInChain = NewRed; | ||||||
| 9623 | else | ||||||
| 9624 | NextInChain = State.Builder.CreateBinOp( | ||||||
| 9625 | (Instruction::BinaryOps)RdxDesc->getOpcode(Kind), NewRed, | ||||||
| 9626 | PrevInChain); | ||||||
| 9627 | State.set(this, NextInChain, Part); | ||||||
| 9628 | } | ||||||
| 9629 | } | ||||||
| 9630 | |||||||
| 9631 | void VPReplicateRecipe::execute(VPTransformState &State) { | ||||||
| 9632 | Instruction *UI = getUnderlyingInstr(); | ||||||
| 9633 | if (State.Instance) { // Generate a single instance. | ||||||
| 9634 | assert(!State.VF.isScalable() && "Can't scalarize a scalable vector")(static_cast <bool> (!State.VF.isScalable() && "Can't scalarize a scalable vector" ) ? void (0) : __assert_fail ("!State.VF.isScalable() && \"Can't scalarize a scalable vector\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9634, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9635 | State.ILV->scalarizeInstruction(UI, this, *State.Instance, | ||||||
| 9636 | IsPredicated, State); | ||||||
| 9637 | // Insert scalar instance packing it into a vector. | ||||||
| 9638 | if (AlsoPack && State.VF.isVector()) { | ||||||
| 9639 | // If we're constructing lane 0, initialize to start from poison. | ||||||
| 9640 | if (State.Instance->Lane.isFirstLane()) { | ||||||
| 9641 | assert(!State.VF.isScalable() && "VF is assumed to be non scalable.")(static_cast <bool> (!State.VF.isScalable() && "VF is assumed to be non scalable." ) ? void (0) : __assert_fail ("!State.VF.isScalable() && \"VF is assumed to be non scalable.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9641, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9642 | Value *Poison = PoisonValue::get( | ||||||
| 9643 | VectorType::get(UI->getType(), State.VF)); | ||||||
| 9644 | State.set(this, Poison, State.Instance->Part); | ||||||
| 9645 | } | ||||||
| 9646 | State.ILV->packScalarIntoVectorValue(this, *State.Instance, State); | ||||||
| 9647 | } | ||||||
| 9648 | return; | ||||||
| 9649 | } | ||||||
| 9650 | |||||||
| 9651 | if (IsUniform) { | ||||||
| 9652 | // If the recipe is uniform across all parts (instead of just per VF), only | ||||||
| 9653 | // generate a single instance. | ||||||
| 9654 | if ((isa<LoadInst>(UI) || isa<StoreInst>(UI)) && | ||||||
| 9655 | all_of(operands(), [](VPValue *Op) { | ||||||
| 9656 | return Op->isDefinedOutsideVectorRegions(); | ||||||
| 9657 | })) { | ||||||
| 9658 | State.ILV->scalarizeInstruction(UI, this, VPIteration(0, 0), IsPredicated, | ||||||
| 9659 | State); | ||||||
| 9660 | if (user_begin() != user_end()) { | ||||||
| 9661 | for (unsigned Part = 1; Part < State.UF; ++Part) | ||||||
| 9662 | State.set(this, State.get(this, VPIteration(0, 0)), | ||||||
| 9663 | VPIteration(Part, 0)); | ||||||
| 9664 | } | ||||||
| 9665 | return; | ||||||
| 9666 | } | ||||||
| 9667 | |||||||
| 9668 | // Uniform within VL means we need to generate lane 0 only for each | ||||||
| 9669 | // unrolled copy. | ||||||
| 9670 | for (unsigned Part = 0; Part < State.UF; ++Part) | ||||||
| 9671 | State.ILV->scalarizeInstruction(UI, this, VPIteration(Part, 0), | ||||||
| 9672 | IsPredicated, State); | ||||||
| 9673 | return; | ||||||
| 9674 | } | ||||||
| 9675 | |||||||
| 9676 | // A store of a loop varying value to a loop invariant address only | ||||||
| 9677 | // needs only the last copy of the store. | ||||||
| 9678 | if (isa<StoreInst>(UI) && !getOperand(1)->hasDefiningRecipe()) { | ||||||
| 9679 | auto Lane = VPLane::getLastLaneForVF(State.VF); | ||||||
| 9680 | State.ILV->scalarizeInstruction(UI, this, VPIteration(State.UF - 1, Lane), IsPredicated, | ||||||
| 9681 | State); | ||||||
| 9682 | return; | ||||||
| 9683 | } | ||||||
| 9684 | |||||||
| 9685 | // Generate scalar instances for all VF lanes of all UF parts. | ||||||
| 9686 | assert(!State.VF.isScalable() && "Can't scalarize a scalable vector")(static_cast <bool> (!State.VF.isScalable() && "Can't scalarize a scalable vector" ) ? void (0) : __assert_fail ("!State.VF.isScalable() && \"Can't scalarize a scalable vector\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9686, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9687 | const unsigned EndLane = State.VF.getKnownMinValue(); | ||||||
| 9688 | for (unsigned Part = 0; Part < State.UF; ++Part) | ||||||
| 9689 | for (unsigned Lane = 0; Lane < EndLane; ++Lane) | ||||||
| 9690 | State.ILV->scalarizeInstruction(UI, this, VPIteration(Part, Lane), | ||||||
| 9691 | IsPredicated, State); | ||||||
| 9692 | } | ||||||
| 9693 | |||||||
| 9694 | void VPWidenMemoryInstructionRecipe::execute(VPTransformState &State) { | ||||||
| 9695 | VPValue *StoredValue = isStore() ? getStoredValue() : nullptr; | ||||||
| 9696 | |||||||
| 9697 | // Attempt to issue a wide load. | ||||||
| 9698 | LoadInst *LI = dyn_cast<LoadInst>(&Ingredient); | ||||||
| 9699 | StoreInst *SI = dyn_cast<StoreInst>(&Ingredient); | ||||||
| 9700 | |||||||
| 9701 | assert((LI || SI) && "Invalid Load/Store instruction")(static_cast <bool> ((LI || SI) && "Invalid Load/Store instruction" ) ? void (0) : __assert_fail ("(LI || SI) && \"Invalid Load/Store instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9701, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9702 | assert((!SI || StoredValue) && "No stored value provided for widened store")(static_cast <bool> ((!SI || StoredValue) && "No stored value provided for widened store" ) ? void (0) : __assert_fail ("(!SI || StoredValue) && \"No stored value provided for widened store\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9702, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9703 | assert((!LI || !StoredValue) && "Stored value provided for widened load")(static_cast <bool> ((!LI || !StoredValue) && "Stored value provided for widened load" ) ? void (0) : __assert_fail ("(!LI || !StoredValue) && \"Stored value provided for widened load\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9703, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9704 | |||||||
| 9705 | Type *ScalarDataTy = getLoadStoreType(&Ingredient); | ||||||
| 9706 | |||||||
| 9707 | auto *DataTy = VectorType::get(ScalarDataTy, State.VF); | ||||||
| 9708 | const Align Alignment = getLoadStoreAlignment(&Ingredient); | ||||||
| 9709 | bool CreateGatherScatter = !Consecutive; | ||||||
| 9710 | |||||||
| 9711 | auto &Builder = State.Builder; | ||||||
| 9712 | InnerLoopVectorizer::VectorParts BlockInMaskParts(State.UF); | ||||||
| 9713 | bool isMaskRequired = getMask(); | ||||||
| 9714 | if (isMaskRequired) | ||||||
| 9715 | for (unsigned Part = 0; Part < State.UF; ++Part) | ||||||
| 9716 | BlockInMaskParts[Part] = State.get(getMask(), Part); | ||||||
| 9717 | |||||||
| 9718 | const auto CreateVecPtr = [&](unsigned Part, Value *Ptr) -> Value * { | ||||||
| 9719 | // Calculate the pointer for the specific unroll-part. | ||||||
| 9720 | GetElementPtrInst *PartPtr = nullptr; | ||||||
| 9721 | |||||||
| 9722 | bool InBounds = false; | ||||||
| 9723 | if (auto *gep = dyn_cast<GetElementPtrInst>(Ptr->stripPointerCasts())) | ||||||
| 9724 | InBounds = gep->isInBounds(); | ||||||
| 9725 | if (Reverse) { | ||||||
| 9726 | // If the address is consecutive but reversed, then the | ||||||
| 9727 | // wide store needs to start at the last vector element. | ||||||
| 9728 | // RunTimeVF = VScale * VF.getKnownMinValue() | ||||||
| 9729 | // For fixed-width VScale is 1, then RunTimeVF = VF.getKnownMinValue() | ||||||
| 9730 | Value *RunTimeVF = getRuntimeVF(Builder, Builder.getInt32Ty(), State.VF); | ||||||
| 9731 | // NumElt = -Part * RunTimeVF | ||||||
| 9732 | Value *NumElt = Builder.CreateMul(Builder.getInt32(-Part), RunTimeVF); | ||||||
| 9733 | // LastLane = 1 - RunTimeVF | ||||||
| 9734 | Value *LastLane = Builder.CreateSub(Builder.getInt32(1), RunTimeVF); | ||||||
| 9735 | PartPtr = | ||||||
| 9736 | cast<GetElementPtrInst>(Builder.CreateGEP(ScalarDataTy, Ptr, NumElt)); | ||||||
| 9737 | PartPtr->setIsInBounds(InBounds); | ||||||
| 9738 | PartPtr = cast<GetElementPtrInst>( | ||||||
| 9739 | Builder.CreateGEP(ScalarDataTy, PartPtr, LastLane)); | ||||||
| 9740 | PartPtr->setIsInBounds(InBounds); | ||||||
| 9741 | if (isMaskRequired) // Reverse of a null all-one mask is a null mask. | ||||||
| 9742 | BlockInMaskParts[Part] = | ||||||
| 9743 | Builder.CreateVectorReverse(BlockInMaskParts[Part], "reverse"); | ||||||
| 9744 | } else { | ||||||
| 9745 | Value *Increment = | ||||||
| 9746 | createStepForVF(Builder, Builder.getInt32Ty(), State.VF, Part); | ||||||
| 9747 | PartPtr = cast<GetElementPtrInst>( | ||||||
| 9748 | Builder.CreateGEP(ScalarDataTy, Ptr, Increment)); | ||||||
| 9749 | PartPtr->setIsInBounds(InBounds); | ||||||
| 9750 | } | ||||||
| 9751 | |||||||
| 9752 | unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace(); | ||||||
| 9753 | return Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace)); | ||||||
| 9754 | }; | ||||||
| 9755 | |||||||
| 9756 | // Handle Stores: | ||||||
| 9757 | if (SI) { | ||||||
| 9758 | State.setDebugLocFromInst(SI); | ||||||
| 9759 | |||||||
| 9760 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
| 9761 | Instruction *NewSI = nullptr; | ||||||
| 9762 | Value *StoredVal = State.get(StoredValue, Part); | ||||||
| 9763 | if (CreateGatherScatter) { | ||||||
| 9764 | Value *MaskPart = isMaskRequired ? BlockInMaskParts[Part] : nullptr; | ||||||
| 9765 | Value *VectorGep = State.get(getAddr(), Part); | ||||||
| 9766 | NewSI = Builder.CreateMaskedScatter(StoredVal, VectorGep, Alignment, | ||||||
| 9767 | MaskPart); | ||||||
| 9768 | } else { | ||||||
| 9769 | if (Reverse) { | ||||||
| 9770 | // If we store to reverse consecutive memory locations, then we need | ||||||
| 9771 | // to reverse the order of elements in the stored value. | ||||||
| 9772 | StoredVal = Builder.CreateVectorReverse(StoredVal, "reverse"); | ||||||
| 9773 | // We don't want to update the value in the map as it might be used in | ||||||
| 9774 | // another expression. So don't call resetVectorValue(StoredVal). | ||||||
| 9775 | } | ||||||
| 9776 | auto *VecPtr = | ||||||
| 9777 | CreateVecPtr(Part, State.get(getAddr(), VPIteration(0, 0))); | ||||||
| 9778 | if (isMaskRequired) | ||||||
| 9779 | NewSI = Builder.CreateMaskedStore(StoredVal, VecPtr, Alignment, | ||||||
| 9780 | BlockInMaskParts[Part]); | ||||||
| 9781 | else | ||||||
| 9782 | NewSI = Builder.CreateAlignedStore(StoredVal, VecPtr, Alignment); | ||||||
| 9783 | } | ||||||
| 9784 | State.addMetadata(NewSI, SI); | ||||||
| 9785 | } | ||||||
| 9786 | return; | ||||||
| 9787 | } | ||||||
| 9788 | |||||||
| 9789 | // Handle loads. | ||||||
| 9790 | assert(LI && "Must have a load instruction")(static_cast <bool> (LI && "Must have a load instruction" ) ? void (0) : __assert_fail ("LI && \"Must have a load instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9790, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9791 | State.setDebugLocFromInst(LI); | ||||||
| 9792 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
| 9793 | Value *NewLI; | ||||||
| 9794 | if (CreateGatherScatter) { | ||||||
| 9795 | Value *MaskPart = isMaskRequired ? BlockInMaskParts[Part] : nullptr; | ||||||
| 9796 | Value *VectorGep = State.get(getAddr(), Part); | ||||||
| 9797 | NewLI = Builder.CreateMaskedGather(DataTy, VectorGep, Alignment, MaskPart, | ||||||
| 9798 | nullptr, "wide.masked.gather"); | ||||||
| 9799 | State.addMetadata(NewLI, LI); | ||||||
| 9800 | } else { | ||||||
| 9801 | auto *VecPtr = | ||||||
| 9802 | CreateVecPtr(Part, State.get(getAddr(), VPIteration(0, 0))); | ||||||
| 9803 | if (isMaskRequired) | ||||||
| 9804 | NewLI = Builder.CreateMaskedLoad( | ||||||
| 9805 | DataTy, VecPtr, Alignment, BlockInMaskParts[Part], | ||||||
| 9806 | PoisonValue::get(DataTy), "wide.masked.load"); | ||||||
| 9807 | else | ||||||
| 9808 | NewLI = | ||||||
| 9809 | Builder.CreateAlignedLoad(DataTy, VecPtr, Alignment, "wide.load"); | ||||||
| 9810 | |||||||
| 9811 | // Add metadata to the load, but setVectorValue to the reverse shuffle. | ||||||
| 9812 | State.addMetadata(NewLI, LI); | ||||||
| 9813 | if (Reverse) | ||||||
| 9814 | NewLI = Builder.CreateVectorReverse(NewLI, "reverse"); | ||||||
| 9815 | } | ||||||
| 9816 | |||||||
| 9817 | State.set(getVPSingleValue(), NewLI, Part); | ||||||
| 9818 | } | ||||||
| 9819 | } | ||||||
| 9820 | |||||||
| 9821 | // Determine how to lower the scalar epilogue, which depends on 1) optimising | ||||||
| 9822 | // for minimum code-size, 2) predicate compiler options, 3) loop hints forcing | ||||||
| 9823 | // predication, and 4) a TTI hook that analyses whether the loop is suitable | ||||||
| 9824 | // for predication. | ||||||
| 9825 | static ScalarEpilogueLowering getScalarEpilogueLowering( | ||||||
| 9826 | Function *F, Loop *L, LoopVectorizeHints &Hints, ProfileSummaryInfo *PSI, | ||||||
| 9827 | BlockFrequencyInfo *BFI, TargetTransformInfo *TTI, TargetLibraryInfo *TLI, | ||||||
| 9828 | AssumptionCache *AC, LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT, | ||||||
| 9829 | LoopVectorizationLegality &LVL, InterleavedAccessInfo *IAI) { | ||||||
| 9830 | // 1) OptSize takes precedence over all other options, i.e. if this is set, | ||||||
| 9831 | // don't look at hints or options, and don't request a scalar epilogue. | ||||||
| 9832 | // (For PGSO, as shouldOptimizeForSize isn't currently accessible from | ||||||
| 9833 | // LoopAccessInfo (due to code dependency and not being able to reliably get | ||||||
| 9834 | // PSI/BFI from a loop analysis under NPM), we cannot suppress the collection | ||||||
| 9835 | // of strides in LoopAccessInfo::analyzeLoop() and vectorize without | ||||||
| 9836 | // versioning when the vectorization is forced, unlike hasOptSize. So revert | ||||||
| 9837 | // back to the old way and vectorize with versioning when forced. See D81345.) | ||||||
| 9838 | if (F->hasOptSize() || (llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI, | ||||||
| 9839 | PGSOQueryType::IRPass) && | ||||||
| 9840 | Hints.getForce() != LoopVectorizeHints::FK_Enabled)) | ||||||
| 9841 | return CM_ScalarEpilogueNotAllowedOptSize; | ||||||
| 9842 | |||||||
| 9843 | // 2) If set, obey the directives | ||||||
| 9844 | if (PreferPredicateOverEpilogue.getNumOccurrences()) { | ||||||
| 9845 | switch (PreferPredicateOverEpilogue) { | ||||||
| 9846 | case PreferPredicateTy::ScalarEpilogue: | ||||||
| 9847 | return CM_ScalarEpilogueAllowed; | ||||||
| 9848 | case PreferPredicateTy::PredicateElseScalarEpilogue: | ||||||
| 9849 | return CM_ScalarEpilogueNotNeededUsePredicate; | ||||||
| 9850 | case PreferPredicateTy::PredicateOrDontVectorize: | ||||||
| 9851 | return CM_ScalarEpilogueNotAllowedUsePredicate; | ||||||
| 9852 | }; | ||||||
| 9853 | } | ||||||
| 9854 | |||||||
| 9855 | // 3) If set, obey the hints | ||||||
| 9856 | switch (Hints.getPredicate()) { | ||||||
| 9857 | case LoopVectorizeHints::FK_Enabled: | ||||||
| 9858 | return CM_ScalarEpilogueNotNeededUsePredicate; | ||||||
| 9859 | case LoopVectorizeHints::FK_Disabled: | ||||||
| 9860 | return CM_ScalarEpilogueAllowed; | ||||||
| 9861 | }; | ||||||
| 9862 | |||||||
| 9863 | // 4) if the TTI hook indicates this is profitable, request predication. | ||||||
| 9864 | if (TTI->preferPredicateOverEpilogue(L, LI, *SE, *AC, TLI, DT, &LVL, IAI)) | ||||||
| 9865 | return CM_ScalarEpilogueNotNeededUsePredicate; | ||||||
| 9866 | |||||||
| 9867 | return CM_ScalarEpilogueAllowed; | ||||||
| 9868 | } | ||||||
| 9869 | |||||||
| 9870 | Value *VPTransformState::get(VPValue *Def, unsigned Part) { | ||||||
| 9871 | // If Values have been set for this Def return the one relevant for \p Part. | ||||||
| 9872 | if (hasVectorValue(Def, Part)) | ||||||
| 9873 | return Data.PerPartOutput[Def][Part]; | ||||||
| 9874 | |||||||
| 9875 | if (!hasScalarValue(Def, {Part, 0})) { | ||||||
| 9876 | Value *IRV = Def->getLiveInIRValue(); | ||||||
| 9877 | Value *B = ILV->getBroadcastInstrs(IRV); | ||||||
| 9878 | set(Def, B, Part); | ||||||
| 9879 | return B; | ||||||
| 9880 | } | ||||||
| 9881 | |||||||
| 9882 | Value *ScalarValue = get(Def, {Part, 0}); | ||||||
| 9883 | // If we aren't vectorizing, we can just copy the scalar map values over | ||||||
| 9884 | // to the vector map. | ||||||
| 9885 | if (VF.isScalar()) { | ||||||
| 9886 | set(Def, ScalarValue, Part); | ||||||
| 9887 | return ScalarValue; | ||||||
| 9888 | } | ||||||
| 9889 | |||||||
| 9890 | bool IsUniform = vputils::isUniformAfterVectorization(Def); | ||||||
| 9891 | |||||||
| 9892 | unsigned LastLane = IsUniform ? 0 : VF.getKnownMinValue() - 1; | ||||||
| 9893 | // Check if there is a scalar value for the selected lane. | ||||||
| 9894 | if (!hasScalarValue(Def, {Part, LastLane})) { | ||||||
| 9895 | // At the moment, VPWidenIntOrFpInductionRecipes and VPScalarIVStepsRecipes can also be uniform. | ||||||
| 9896 | assert((isa<VPWidenIntOrFpInductionRecipe>(Def->getDefiningRecipe()) ||(static_cast <bool> ((isa<VPWidenIntOrFpInductionRecipe >(Def->getDefiningRecipe()) || isa<VPScalarIVStepsRecipe >(Def->getDefiningRecipe())) && "unexpected recipe found to be invariant" ) ? void (0) : __assert_fail ("(isa<VPWidenIntOrFpInductionRecipe>(Def->getDefiningRecipe()) || isa<VPScalarIVStepsRecipe>(Def->getDefiningRecipe())) && \"unexpected recipe found to be invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9898, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9897 | isa<VPScalarIVStepsRecipe>(Def->getDefiningRecipe())) &&(static_cast <bool> ((isa<VPWidenIntOrFpInductionRecipe >(Def->getDefiningRecipe()) || isa<VPScalarIVStepsRecipe >(Def->getDefiningRecipe())) && "unexpected recipe found to be invariant" ) ? void (0) : __assert_fail ("(isa<VPWidenIntOrFpInductionRecipe>(Def->getDefiningRecipe()) || isa<VPScalarIVStepsRecipe>(Def->getDefiningRecipe())) && \"unexpected recipe found to be invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9898, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 9898 | "unexpected recipe found to be invariant")(static_cast <bool> ((isa<VPWidenIntOrFpInductionRecipe >(Def->getDefiningRecipe()) || isa<VPScalarIVStepsRecipe >(Def->getDefiningRecipe())) && "unexpected recipe found to be invariant" ) ? void (0) : __assert_fail ("(isa<VPWidenIntOrFpInductionRecipe>(Def->getDefiningRecipe()) || isa<VPScalarIVStepsRecipe>(Def->getDefiningRecipe())) && \"unexpected recipe found to be invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9898, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9899 | IsUniform = true; | ||||||
| 9900 | LastLane = 0; | ||||||
| 9901 | } | ||||||
| 9902 | |||||||
| 9903 | auto *LastInst = cast<Instruction>(get(Def, {Part, LastLane})); | ||||||
| 9904 | // Set the insert point after the last scalarized instruction or after the | ||||||
| 9905 | // last PHI, if LastInst is a PHI. This ensures the insertelement sequence | ||||||
| 9906 | // will directly follow the scalar definitions. | ||||||
| 9907 | auto OldIP = Builder.saveIP(); | ||||||
| 9908 | auto NewIP = | ||||||
| 9909 | isa<PHINode>(LastInst) | ||||||
| 9910 | ? BasicBlock::iterator(LastInst->getParent()->getFirstNonPHI()) | ||||||
| 9911 | : std::next(BasicBlock::iterator(LastInst)); | ||||||
| 9912 | Builder.SetInsertPoint(&*NewIP); | ||||||
| 9913 | |||||||
| 9914 | // However, if we are vectorizing, we need to construct the vector values. | ||||||
| 9915 | // If the value is known to be uniform after vectorization, we can just | ||||||
| 9916 | // broadcast the scalar value corresponding to lane zero for each unroll | ||||||
| 9917 | // iteration. Otherwise, we construct the vector values using | ||||||
| 9918 | // insertelement instructions. Since the resulting vectors are stored in | ||||||
| 9919 | // State, we will only generate the insertelements once. | ||||||
| 9920 | Value *VectorValue = nullptr; | ||||||
| 9921 | if (IsUniform) { | ||||||
| 9922 | VectorValue = ILV->getBroadcastInstrs(ScalarValue); | ||||||
| 9923 | set(Def, VectorValue, Part); | ||||||
| 9924 | } else { | ||||||
| 9925 | // Initialize packing with insertelements to start from undef. | ||||||
| 9926 | assert(!VF.isScalable() && "VF is assumed to be non scalable.")(static_cast <bool> (!VF.isScalable() && "VF is assumed to be non scalable." ) ? void (0) : __assert_fail ("!VF.isScalable() && \"VF is assumed to be non scalable.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9926, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9927 | Value *Undef = PoisonValue::get(VectorType::get(LastInst->getType(), VF)); | ||||||
| 9928 | set(Def, Undef, Part); | ||||||
| 9929 | for (unsigned Lane = 0; Lane < VF.getKnownMinValue(); ++Lane) | ||||||
| 9930 | ILV->packScalarIntoVectorValue(Def, {Part, Lane}, *this); | ||||||
| 9931 | VectorValue = get(Def, Part); | ||||||
| 9932 | } | ||||||
| 9933 | Builder.restoreIP(OldIP); | ||||||
| 9934 | return VectorValue; | ||||||
| 9935 | } | ||||||
| 9936 | |||||||
| 9937 | // Process the loop in the VPlan-native vectorization path. This path builds | ||||||
| 9938 | // VPlan upfront in the vectorization pipeline, which allows to apply | ||||||
| 9939 | // VPlan-to-VPlan transformations from the very beginning without modifying the | ||||||
| 9940 | // input LLVM IR. | ||||||
| 9941 | static bool processLoopInVPlanNativePath( | ||||||
| 9942 | Loop *L, PredicatedScalarEvolution &PSE, LoopInfo *LI, DominatorTree *DT, | ||||||
| 9943 | LoopVectorizationLegality *LVL, TargetTransformInfo *TTI, | ||||||
| 9944 | TargetLibraryInfo *TLI, DemandedBits *DB, AssumptionCache *AC, | ||||||
| 9945 | OptimizationRemarkEmitter *ORE, BlockFrequencyInfo *BFI, | ||||||
| 9946 | ProfileSummaryInfo *PSI, LoopVectorizeHints &Hints, | ||||||
| 9947 | LoopVectorizationRequirements &Requirements) { | ||||||
| 9948 | |||||||
| 9949 | if (isa<SCEVCouldNotCompute>(PSE.getBackedgeTakenCount())) { | ||||||
| 9950 | LLVM_DEBUG(dbgs() << "LV: cannot compute the outer-loop trip count\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: cannot compute the outer-loop trip count\n" ; } } while (false); | ||||||
| 9951 | return false; | ||||||
| 9952 | } | ||||||
| 9953 | assert(EnableVPlanNativePath && "VPlan-native path is disabled.")(static_cast <bool> (EnableVPlanNativePath && "VPlan-native path is disabled." ) ? void (0) : __assert_fail ("EnableVPlanNativePath && \"VPlan-native path is disabled.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9953, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9954 | Function *F = L->getHeader()->getParent(); | ||||||
| 9955 | InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL->getLAI()); | ||||||
| 9956 | |||||||
| 9957 | ScalarEpilogueLowering SEL = getScalarEpilogueLowering( | ||||||
| 9958 | F, L, Hints, PSI, BFI, TTI, TLI, AC, LI, PSE.getSE(), DT, *LVL, &IAI); | ||||||
| 9959 | |||||||
| 9960 | LoopVectorizationCostModel CM(SEL, L, PSE, LI, LVL, *TTI, TLI, DB, AC, ORE, F, | ||||||
| 9961 | &Hints, IAI); | ||||||
| 9962 | // Use the planner for outer loop vectorization. | ||||||
| 9963 | // TODO: CM is not used at this point inside the planner. Turn CM into an | ||||||
| 9964 | // optional argument if we don't need it in the future. | ||||||
| 9965 | LoopVectorizationPlanner LVP(L, LI, TLI, TTI, LVL, CM, IAI, PSE, Hints, ORE); | ||||||
| 9966 | |||||||
| 9967 | // Get user vectorization factor. | ||||||
| 9968 | ElementCount UserVF = Hints.getWidth(); | ||||||
| 9969 | |||||||
| 9970 | CM.collectElementTypesForWidening(); | ||||||
| 9971 | |||||||
| 9972 | // Plan how to best vectorize, return the best VF and its cost. | ||||||
| 9973 | const VectorizationFactor VF = LVP.planInVPlanNativePath(UserVF); | ||||||
| 9974 | |||||||
| 9975 | // If we are stress testing VPlan builds, do not attempt to generate vector | ||||||
| 9976 | // code. Masked vector code generation support will follow soon. | ||||||
| 9977 | // Also, do not attempt to vectorize if no vector code will be produced. | ||||||
| 9978 | if (VPlanBuildStressTest || VectorizationFactor::Disabled() == VF) | ||||||
| 9979 | return false; | ||||||
| 9980 | |||||||
| 9981 | VPlan &BestPlan = LVP.getBestPlanFor(VF.Width); | ||||||
| 9982 | |||||||
| 9983 | { | ||||||
| 9984 | GeneratedRTChecks Checks(*PSE.getSE(), DT, LI, TTI, | ||||||
| 9985 | F->getParent()->getDataLayout()); | ||||||
| 9986 | InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, | ||||||
| 9987 | VF.Width, 1, LVL, &CM, BFI, PSI, Checks); | ||||||
| 9988 | LLVM_DEBUG(dbgs() << "Vectorizing outer loop in \""do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Vectorizing outer loop in \"" << L->getHeader()->getParent()->getName() << "\"\n"; } } while (false) | ||||||
| 9989 | << L->getHeader()->getParent()->getName() << "\"\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Vectorizing outer loop in \"" << L->getHeader()->getParent()->getName() << "\"\n"; } } while (false); | ||||||
| 9990 | LVP.executePlan(VF.Width, 1, BestPlan, LB, DT, false); | ||||||
| 9991 | } | ||||||
| 9992 | |||||||
| 9993 | // Mark the loop as already vectorized to avoid vectorizing again. | ||||||
| 9994 | Hints.setAlreadyVectorized(); | ||||||
| 9995 | assert(!verifyFunction(*L->getHeader()->getParent(), &dbgs()))(static_cast <bool> (!verifyFunction(*L->getHeader() ->getParent(), &dbgs())) ? void (0) : __assert_fail ("!verifyFunction(*L->getHeader()->getParent(), &dbgs())" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9995, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 9996 | return true; | ||||||
| 9997 | } | ||||||
| 9998 | |||||||
| 9999 | // Emit a remark if there are stores to floats that required a floating point | ||||||
| 10000 | // extension. If the vectorized loop was generated with floating point there | ||||||
| 10001 | // will be a performance penalty from the conversion overhead and the change in | ||||||
| 10002 | // the vector width. | ||||||
| 10003 | static void checkMixedPrecision(Loop *L, OptimizationRemarkEmitter *ORE) { | ||||||
| 10004 | SmallVector<Instruction *, 4> Worklist; | ||||||
| 10005 | for (BasicBlock *BB : L->getBlocks()) { | ||||||
| 10006 | for (Instruction &Inst : *BB) { | ||||||
| 10007 | if (auto *S = dyn_cast<StoreInst>(&Inst)) { | ||||||
| 10008 | if (S->getValueOperand()->getType()->isFloatTy()) | ||||||
| 10009 | Worklist.push_back(S); | ||||||
| 10010 | } | ||||||
| 10011 | } | ||||||
| 10012 | } | ||||||
| 10013 | |||||||
| 10014 | // Traverse the floating point stores upwards searching, for floating point | ||||||
| 10015 | // conversions. | ||||||
| 10016 | SmallPtrSet<const Instruction *, 4> Visited; | ||||||
| 10017 | SmallPtrSet<const Instruction *, 4> EmittedRemark; | ||||||
| 10018 | while (!Worklist.empty()) { | ||||||
| 10019 | auto *I = Worklist.pop_back_val(); | ||||||
| 10020 | if (!L->contains(I)) | ||||||
| 10021 | continue; | ||||||
| 10022 | if (!Visited.insert(I).second) | ||||||
| 10023 | continue; | ||||||
| 10024 | |||||||
| 10025 | // Emit a remark if the floating point store required a floating | ||||||
| 10026 | // point conversion. | ||||||
| 10027 | // TODO: More work could be done to identify the root cause such as a | ||||||
| 10028 | // constant or a function return type and point the user to it. | ||||||
| 10029 | if (isa<FPExtInst>(I) && EmittedRemark.insert(I).second) | ||||||
| 10030 | ORE->emit([&]() { | ||||||
| 10031 | return OptimizationRemarkAnalysis(LV_NAME"loop-vectorize", "VectorMixedPrecision", | ||||||
| 10032 | I->getDebugLoc(), L->getHeader()) | ||||||
| 10033 | << "floating point conversion changes vector width. " | ||||||
| 10034 | << "Mixed floating point precision requires an up/down " | ||||||
| 10035 | << "cast that will negatively impact performance."; | ||||||
| 10036 | }); | ||||||
| 10037 | |||||||
| 10038 | for (Use &Op : I->operands()) | ||||||
| 10039 | if (auto *OpI = dyn_cast<Instruction>(Op)) | ||||||
| 10040 | Worklist.push_back(OpI); | ||||||
| 10041 | } | ||||||
| 10042 | } | ||||||
| 10043 | |||||||
| 10044 | static bool areRuntimeChecksProfitable(GeneratedRTChecks &Checks, | ||||||
| 10045 | VectorizationFactor &VF, | ||||||
| 10046 | std::optional<unsigned> VScale, Loop *L, | ||||||
| 10047 | ScalarEvolution &SE) { | ||||||
| 10048 | InstructionCost CheckCost = Checks.getCost(); | ||||||
| 10049 | if (!CheckCost.isValid()) | ||||||
| 10050 | return false; | ||||||
| 10051 | |||||||
| 10052 | // When interleaving only scalar and vector cost will be equal, which in turn | ||||||
| 10053 | // would lead to a divide by 0. Fall back to hard threshold. | ||||||
| 10054 | if (VF.Width.isScalar()) { | ||||||
| 10055 | if (CheckCost > VectorizeMemoryCheckThreshold) { | ||||||
| 10056 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving only is not profitable due to runtime checks\n" ; } } while (false) | ||||||
| 10057 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving only is not profitable due to runtime checks\n" ; } } while (false) | ||||||
| 10058 | << "LV: Interleaving only is not profitable due to runtime checks\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving only is not profitable due to runtime checks\n" ; } } while (false); | ||||||
| 10059 | return false; | ||||||
| 10060 | } | ||||||
| 10061 | return true; | ||||||
| 10062 | } | ||||||
| 10063 | |||||||
| 10064 | // The scalar cost should only be 0 when vectorizing with a user specified VF/IC. In those cases, runtime checks should always be generated. | ||||||
| 10065 | double ScalarC = *VF.ScalarCost.getValue(); | ||||||
| 10066 | if (ScalarC == 0) | ||||||
| 10067 | return true; | ||||||
| 10068 | |||||||
| 10069 | // First, compute the minimum iteration count required so that the vector | ||||||
| 10070 | // loop outperforms the scalar loop. | ||||||
| 10071 | // The total cost of the scalar loop is | ||||||
| 10072 | // ScalarC * TC | ||||||
| 10073 | // where | ||||||
| 10074 | // * TC is the actual trip count of the loop. | ||||||
| 10075 | // * ScalarC is the cost of a single scalar iteration. | ||||||
| 10076 | // | ||||||
| 10077 | // The total cost of the vector loop is | ||||||
| 10078 | // RtC + VecC * (TC / VF) + EpiC | ||||||
| 10079 | // where | ||||||
| 10080 | // * RtC is the cost of the generated runtime checks | ||||||
| 10081 | // * VecC is the cost of a single vector iteration. | ||||||
| 10082 | // * TC is the actual trip count of the loop | ||||||
| 10083 | // * VF is the vectorization factor | ||||||
| 10084 | // * EpiCost is the cost of the generated epilogue, including the cost | ||||||
| 10085 | // of the remaining scalar operations. | ||||||
| 10086 | // | ||||||
| 10087 | // Vectorization is profitable once the total vector cost is less than the | ||||||
| 10088 | // total scalar cost: | ||||||
| 10089 | // RtC + VecC * (TC / VF) + EpiC < ScalarC * TC | ||||||
| 10090 | // | ||||||
| 10091 | // Now we can compute the minimum required trip count TC as | ||||||
| 10092 | // (RtC + EpiC) / (ScalarC - (VecC / VF)) < TC | ||||||
| 10093 | // | ||||||
| 10094 | // For now we assume the epilogue cost EpiC = 0 for simplicity. Note that | ||||||
| 10095 | // the computations are performed on doubles, not integers and the result | ||||||
| 10096 | // is rounded up, hence we get an upper estimate of the TC. | ||||||
| 10097 | unsigned IntVF = VF.Width.getKnownMinValue(); | ||||||
| 10098 | if (VF.Width.isScalable()) { | ||||||
| 10099 | unsigned AssumedMinimumVscale = 1; | ||||||
| 10100 | if (VScale) | ||||||
| 10101 | AssumedMinimumVscale = *VScale; | ||||||
| 10102 | IntVF *= AssumedMinimumVscale; | ||||||
| 10103 | } | ||||||
| 10104 | double VecCOverVF = double(*VF.Cost.getValue()) / IntVF; | ||||||
| 10105 | double RtC = *CheckCost.getValue(); | ||||||
| 10106 | double MinTC1 = RtC / (ScalarC - VecCOverVF); | ||||||
| 10107 | |||||||
| 10108 | // Second, compute a minimum iteration count so that the cost of the | ||||||
| 10109 | // runtime checks is only a fraction of the total scalar loop cost. This | ||||||
| 10110 | // adds a loop-dependent bound on the overhead incurred if the runtime | ||||||
| 10111 | // checks fail. In case the runtime checks fail, the cost is RtC + ScalarC | ||||||
| 10112 | // * TC. To bound the runtime check to be a fraction 1/X of the scalar | ||||||
| 10113 | // cost, compute | ||||||
| 10114 | // RtC < ScalarC * TC * (1 / X) ==> RtC * X / ScalarC < TC | ||||||
| 10115 | double MinTC2 = RtC * 10 / ScalarC; | ||||||
| 10116 | |||||||
| 10117 | // Now pick the larger minimum. If it is not a multiple of VF, choose the | ||||||
| 10118 | // next closest multiple of VF. This should partly compensate for ignoring | ||||||
| 10119 | // the epilogue cost. | ||||||
| 10120 | uint64_t MinTC = std::ceil(std::max(MinTC1, MinTC2)); | ||||||
| 10121 | VF.MinProfitableTripCount = ElementCount::getFixed(alignTo(MinTC, IntVF)); | ||||||
| 10122 | |||||||
| 10123 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Minimum required TC for runtime checks to be profitable:" << VF.MinProfitableTripCount << "\n"; } } while ( false) | ||||||
| 10124 | dbgs() << "LV: Minimum required TC for runtime checks to be profitable:"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Minimum required TC for runtime checks to be profitable:" << VF.MinProfitableTripCount << "\n"; } } while ( false) | ||||||
| 10125 | << VF.MinProfitableTripCount << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Minimum required TC for runtime checks to be profitable:" << VF.MinProfitableTripCount << "\n"; } } while ( false); | ||||||
| 10126 | |||||||
| 10127 | // Skip vectorization if the expected trip count is less than the minimum | ||||||
| 10128 | // required trip count. | ||||||
| 10129 | if (auto ExpectedTC = getSmallBestKnownTC(SE, L)) { | ||||||
| 10130 | if (ElementCount::isKnownLT(ElementCount::getFixed(*ExpectedTC), | ||||||
| 10131 | VF.MinProfitableTripCount)) { | ||||||
| 10132 | LLVM_DEBUG(dbgs() << "LV: Vectorization is not beneficial: expected "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Vectorization is not beneficial: expected " "trip count < minimum profitable VF (" << *ExpectedTC << " < " << VF.MinProfitableTripCount << ")\n"; } } while (false) | ||||||
| 10133 | "trip count < minimum profitable VF ("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Vectorization is not beneficial: expected " "trip count < minimum profitable VF (" << *ExpectedTC << " < " << VF.MinProfitableTripCount << ")\n"; } } while (false) | ||||||
| 10134 | << *ExpectedTC << " < " << VF.MinProfitableTripCountdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Vectorization is not beneficial: expected " "trip count < minimum profitable VF (" << *ExpectedTC << " < " << VF.MinProfitableTripCount << ")\n"; } } while (false) | ||||||
| 10135 | << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Vectorization is not beneficial: expected " "trip count < minimum profitable VF (" << *ExpectedTC << " < " << VF.MinProfitableTripCount << ")\n"; } } while (false); | ||||||
| 10136 | |||||||
| 10137 | return false; | ||||||
| 10138 | } | ||||||
| 10139 | } | ||||||
| 10140 | return true; | ||||||
| 10141 | } | ||||||
| 10142 | |||||||
| 10143 | LoopVectorizePass::LoopVectorizePass(LoopVectorizeOptions Opts) | ||||||
| 10144 | : InterleaveOnlyWhenForced(Opts.InterleaveOnlyWhenForced || | ||||||
| 10145 | !EnableLoopInterleaving), | ||||||
| 10146 | VectorizeOnlyWhenForced(Opts.VectorizeOnlyWhenForced || | ||||||
| 10147 | !EnableLoopVectorization) {} | ||||||
| 10148 | |||||||
| 10149 | bool LoopVectorizePass::processLoop(Loop *L) { | ||||||
| 10150 | assert((EnableVPlanNativePath || L->isInnermost()) &&(static_cast <bool> ((EnableVPlanNativePath || L->isInnermost ()) && "VPlan-native path is not enabled. Only process inner loops." ) ? void (0) : __assert_fail ("(EnableVPlanNativePath || L->isInnermost()) && \"VPlan-native path is not enabled. Only process inner loops.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10151, __extension__ __PRETTY_FUNCTION__)) | ||||||
| 10151 | "VPlan-native path is not enabled. Only process inner loops.")(static_cast <bool> ((EnableVPlanNativePath || L->isInnermost ()) && "VPlan-native path is not enabled. Only process inner loops." ) ? void (0) : __assert_fail ("(EnableVPlanNativePath || L->isInnermost()) && \"VPlan-native path is not enabled. Only process inner loops.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10151, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 10152 | |||||||
| 10153 | #ifndef NDEBUG | ||||||
| 10154 | const std::string DebugLocStr = getDebugLocString(L); | ||||||
| 10155 | #endif /* NDEBUG */ | ||||||
| 10156 | |||||||
| 10157 | LLVM_DEBUG(dbgs() << "\nLV: Checking a loop in '"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "\nLV: Checking a loop in '" << L->getHeader()->getParent()->getName() << "' from " << DebugLocStr << "\n"; } } while (false ) | ||||||
| 10158 | << L->getHeader()->getParent()->getName() << "' from "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "\nLV: Checking a loop in '" << L->getHeader()->getParent()->getName() << "' from " << DebugLocStr << "\n"; } } while (false ) | ||||||
| 10159 | << DebugLocStr << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "\nLV: Checking a loop in '" << L->getHeader()->getParent()->getName() << "' from " << DebugLocStr << "\n"; } } while (false ); | ||||||
| 10160 | |||||||
| 10161 | LoopVectorizeHints Hints(L, InterleaveOnlyWhenForced, *ORE, TTI); | ||||||
| 10162 | |||||||
| 10163 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | ||||||
| 10164 | dbgs() << "LV: Loop hints:"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | ||||||
| 10165 | << " force="do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | ||||||
| 10166 | << (Hints.getForce() == LoopVectorizeHints::FK_Disableddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | ||||||
| 10167 | ? "disabled"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | ||||||
| 10168 | : (Hints.getForce() == LoopVectorizeHints::FK_Enableddo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | ||||||
| 10169 | ? "enabled"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | ||||||
| 10170 | : "?"))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | ||||||
| 10171 | << " width=" << Hints.getWidth()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false) | ||||||
| 10172 | << " interleave=" << Hints.getInterleave() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints:: FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints ::FK_Enabled ? "enabled" : "?")) << " width=" << Hints .getWidth() << " interleave=" << Hints.getInterleave () << "\n"; } } while (false); | ||||||
| 10173 | |||||||
| 10174 | // Function containing loop | ||||||
| 10175 | Function *F = L->getHeader()->getParent(); | ||||||
| 10176 | |||||||
| 10177 | // Looking at the diagnostic output is the only way to determine if a loop | ||||||
| 10178 | // was vectorized (other than looking at the IR or machine code), so it | ||||||
| 10179 | // is important to generate an optimization remark for each loop. Most of | ||||||
| 10180 | // these messages are generated as OptimizationRemarkAnalysis. Remarks | ||||||
| 10181 | // generated as OptimizationRemark and OptimizationRemarkMissed are | ||||||
| 10182 | // less verbose reporting vectorized loops and unvectorized loops that may | ||||||
| 10183 | // benefit from vectorization, respectively. | ||||||
| 10184 | |||||||
| 10185 | if (!Hints.allowVectorization(F, L, VectorizeOnlyWhenForced)) { | ||||||
| 10186 | LLVM_DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Loop hints prevent vectorization.\n" ; } } while (false); | ||||||
| 10187 | return false; | ||||||
| 10188 | } | ||||||
| 10189 | |||||||
| 10190 | PredicatedScalarEvolution PSE(*SE, *L); | ||||||
| 10191 | |||||||
| 10192 | // Check if it is legal to vectorize the loop. | ||||||
| 10193 | LoopVectorizationRequirements Requirements; | ||||||
| 10194 | LoopVectorizationLegality LVL(L, PSE, DT, TTI, TLI, F, *LAIs, LI, ORE, | ||||||
| 10195 | &Requirements, &Hints, DB, AC, BFI, PSI); | ||||||
| 10196 | if (!LVL.canVectorize(EnableVPlanNativePath)) { | ||||||
| 10197 | LLVM_DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not vectorizing: Cannot prove legality.\n" ; } } while (false); | ||||||
| 10198 | Hints.emitRemarkWithHints(); | ||||||
| 10199 | return false; | ||||||
| 10200 | } | ||||||
| 10201 | |||||||
| 10202 | // Entrance to the VPlan-native vectorization path. Outer loops are processed | ||||||
| 10203 | // here. They may require CFG and instruction level transformations before | ||||||
| 10204 | // even evaluating whether vectorization is profitable. Since we cannot modify | ||||||
| 10205 | // the incoming IR, we need to build VPlan upfront in the vectorization | ||||||
| 10206 | // pipeline. | ||||||
| 10207 | if (!L->isInnermost()) | ||||||
| 10208 | return processLoopInVPlanNativePath(L, PSE, LI, DT, &LVL, TTI, TLI, DB, AC, | ||||||
| 10209 | ORE, BFI, PSI, Hints, Requirements); | ||||||
| 10210 | |||||||
| 10211 | assert(L->isInnermost() && "Inner loop expected.")(static_cast <bool> (L->isInnermost() && "Inner loop expected." ) ? void (0) : __assert_fail ("L->isInnermost() && \"Inner loop expected.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10211, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 10212 | |||||||
| 10213 | InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL.getLAI()); | ||||||
| 10214 | bool UseInterleaved = TTI->enableInterleavedAccessVectorization(); | ||||||
| 10215 | |||||||
| 10216 | // If an override option has been passed in for interleaved accesses, use it. | ||||||
| 10217 | if (EnableInterleavedMemAccesses.getNumOccurrences() > 0) | ||||||
| 10218 | UseInterleaved = EnableInterleavedMemAccesses; | ||||||
| 10219 | |||||||
| 10220 | // Analyze interleaved memory accesses. | ||||||
| 10221 | if (UseInterleaved) | ||||||
| 10222 | IAI.analyzeInterleaving(useMaskedInterleavedAccesses(*TTI)); | ||||||
| 10223 | |||||||
| 10224 | // Check the function attributes and profiles to find out if this function | ||||||
| 10225 | // should be optimized for size. | ||||||
| 10226 | ScalarEpilogueLowering SEL = getScalarEpilogueLowering( | ||||||
| 10227 | F, L, Hints, PSI, BFI, TTI, TLI, AC, LI, PSE.getSE(), DT, LVL, &IAI); | ||||||
| 10228 | |||||||
| 10229 | // Check the loop for a trip count threshold: vectorize loops with a tiny trip | ||||||
| 10230 | // count by optimizing for size, to minimize overheads. | ||||||
| 10231 | auto ExpectedTC = getSmallBestKnownTC(*SE, L); | ||||||
| 10232 | if (ExpectedTC && *ExpectedTC < TinyTripCountVectorThreshold) { | ||||||
| 10233 | LLVM_DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a loop with a very small trip count. " << "This loop is worth vectorizing only if no scalar " << "iteration overheads are incurred."; } } while (false ) | ||||||
| 10234 | << "This loop is worth vectorizing only if no scalar "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a loop with a very small trip count. " << "This loop is worth vectorizing only if no scalar " << "iteration overheads are incurred."; } } while (false ) | ||||||
| 10235 | << "iteration overheads are incurred.")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a loop with a very small trip count. " << "This loop is worth vectorizing only if no scalar " << "iteration overheads are incurred."; } } while (false ); | ||||||
| 10236 | if (Hints.getForce() == LoopVectorizeHints::FK_Enabled) | ||||||
| 10237 | LLVM_DEBUG(dbgs() << " But vectorizing was explicitly forced.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " But vectorizing was explicitly forced.\n" ; } } while (false); | ||||||
| 10238 | else { | ||||||
| 10239 | if (*ExpectedTC > TTI->getMinTripCountTailFoldingThreshold()) { | ||||||
| 10240 | LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "\n"; } } while (false); | ||||||
| 10241 | SEL = CM_ScalarEpilogueNotAllowedLowTripLoop; | ||||||
| 10242 | } else { | ||||||
| 10243 | LLVM_DEBUG(dbgs() << " But the target considers the trip count too "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " But the target considers the trip count too " "small to consider vectorizing.\n"; } } while (false) | ||||||
| 10244 | "small to consider vectorizing.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " But the target considers the trip count too " "small to consider vectorizing.\n"; } } while (false); | ||||||
| 10245 | reportVectorizationFailure( | ||||||
| 10246 | "The trip count is below the minial threshold value.", | ||||||
| 10247 | "loop trip count is too low, avoiding vectorization", | ||||||
| 10248 | "LowTripCount", ORE, L); | ||||||
| 10249 | Hints.emitRemarkWithHints(); | ||||||
| 10250 | return false; | ||||||
| 10251 | } | ||||||
| 10252 | } | ||||||
| 10253 | } | ||||||
| 10254 | |||||||
| 10255 | // Check the function attributes to see if implicit floats or vectors are | ||||||
| 10256 | // allowed. | ||||||
| 10257 | if (F->hasFnAttribute(Attribute::NoImplicitFloat)) { | ||||||
| 10258 | reportVectorizationFailure( | ||||||
| 10259 | "Can't vectorize when the NoImplicitFloat attribute is used", | ||||||
| 10260 | "loop not vectorized due to NoImplicitFloat attribute", | ||||||
| 10261 | "NoImplicitFloat", ORE, L); | ||||||
| 10262 | Hints.emitRemarkWithHints(); | ||||||
| 10263 | return false; | ||||||
| 10264 | } | ||||||
| 10265 | |||||||
| 10266 | // Check if the target supports potentially unsafe FP vectorization. | ||||||
| 10267 | // FIXME: Add a check for the type of safety issue (denormal, signaling) | ||||||
| 10268 | // for the target we're vectorizing for, to make sure none of the | ||||||
| 10269 | // additional fp-math flags can help. | ||||||
| 10270 | if (Hints.isPotentiallyUnsafe() && | ||||||
| 10271 | TTI->isFPVectorizationPotentiallyUnsafe()) { | ||||||
| 10272 | reportVectorizationFailure( | ||||||
| 10273 | "Potentially unsafe FP op prevents vectorization", | ||||||
| 10274 | "loop not vectorized due to unsafe FP support.", | ||||||
| 10275 | "UnsafeFP", ORE, L); | ||||||
| 10276 | Hints.emitRemarkWithHints(); | ||||||
| 10277 | return false; | ||||||
| 10278 | } | ||||||
| 10279 | |||||||
| 10280 | bool AllowOrderedReductions; | ||||||
| 10281 | // If the flag is set, use that instead and override the TTI behaviour. | ||||||
| 10282 | if (ForceOrderedReductions.getNumOccurrences() > 0) | ||||||
| 10283 | AllowOrderedReductions = ForceOrderedReductions; | ||||||
| 10284 | else | ||||||
| 10285 | AllowOrderedReductions = TTI->enableOrderedReductions(); | ||||||
| 10286 | if (!LVL.canVectorizeFPMath(AllowOrderedReductions)) { | ||||||
| 10287 | ORE->emit([&]() { | ||||||
| 10288 | auto *ExactFPMathInst = Requirements.getExactFPInst(); | ||||||
| 10289 | return OptimizationRemarkAnalysisFPCommute(DEBUG_TYPE"loop-vectorize", "CantReorderFPOps", | ||||||
| 10290 | ExactFPMathInst->getDebugLoc(), | ||||||
| 10291 | ExactFPMathInst->getParent()) | ||||||
| 10292 | << "loop not vectorized: cannot prove it is safe to reorder " | ||||||
| 10293 | "floating-point operations"; | ||||||
| 10294 | }); | ||||||
| 10295 | LLVM_DEBUG(dbgs() << "LV: loop not vectorized: cannot prove it is safe to "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: loop not vectorized: cannot prove it is safe to " "reorder floating-point operations\n"; } } while (false) | ||||||
| 10296 | "reorder floating-point operations\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: loop not vectorized: cannot prove it is safe to " "reorder floating-point operations\n"; } } while (false); | ||||||
| 10297 | Hints.emitRemarkWithHints(); | ||||||
| 10298 | return false; | ||||||
| 10299 | } | ||||||
| 10300 | |||||||
| 10301 | // Use the cost model. | ||||||
| 10302 | LoopVectorizationCostModel CM(SEL, L, PSE, LI, &LVL, *TTI, TLI, DB, AC, ORE, | ||||||
| 10303 | F, &Hints, IAI); | ||||||
| 10304 | CM.collectValuesToIgnore(); | ||||||
| 10305 | CM.collectElementTypesForWidening(); | ||||||
| 10306 | |||||||
| 10307 | // Use the planner for vectorization. | ||||||
| 10308 | LoopVectorizationPlanner LVP(L, LI, TLI, TTI, &LVL, CM, IAI, PSE, Hints, ORE); | ||||||
| 10309 | |||||||
| 10310 | // Get user vectorization factor and interleave count. | ||||||
| 10311 | ElementCount UserVF = Hints.getWidth(); | ||||||
| 10312 | unsigned UserIC = Hints.getInterleave(); | ||||||
| 10313 | |||||||
| 10314 | // Plan how to best vectorize, return the best VF and its cost. | ||||||
| 10315 | std::optional<VectorizationFactor> MaybeVF = LVP.plan(UserVF, UserIC); | ||||||
| 10316 | |||||||
| 10317 | VectorizationFactor VF = VectorizationFactor::Disabled(); | ||||||
| 10318 | unsigned IC = 1; | ||||||
| 10319 | |||||||
| 10320 | GeneratedRTChecks Checks(*PSE.getSE(), DT, LI, TTI, | ||||||
| 10321 | F->getParent()->getDataLayout()); | ||||||
| 10322 | if (MaybeVF) { | ||||||
| 10323 | VF = *MaybeVF; | ||||||
| 10324 | // Select the interleave count. | ||||||
| 10325 | IC = CM.selectInterleaveCount(VF.Width, VF.Cost); | ||||||
| 10326 | |||||||
| 10327 | unsigned SelectedIC = std::max(IC, UserIC); | ||||||
| 10328 | // Optimistically generate runtime checks if they are needed. Drop them if | ||||||
| 10329 | // they turn out to not be profitable. | ||||||
| 10330 | if (VF.Width.isVector() || SelectedIC > 1) | ||||||
| 10331 | Checks.Create(L, *LVL.getLAI(), PSE.getPredicate(), VF.Width, SelectedIC); | ||||||
| 10332 | |||||||
| 10333 | // Check if it is profitable to vectorize with runtime checks. | ||||||
| 10334 | bool ForceVectorization = | ||||||
| 10335 | Hints.getForce() == LoopVectorizeHints::FK_Enabled; | ||||||
| 10336 | if (!ForceVectorization && | ||||||
| 10337 | !areRuntimeChecksProfitable(Checks, VF, CM.getVScaleForTuning(), L, | ||||||
| 10338 | *PSE.getSE())) { | ||||||
| 10339 | ORE->emit([&]() { | ||||||
| 10340 | return OptimizationRemarkAnalysisAliasing( | ||||||
| 10341 | DEBUG_TYPE"loop-vectorize", "CantReorderMemOps", L->getStartLoc(), | ||||||
| 10342 | L->getHeader()) | ||||||
| 10343 | << "loop not vectorized: cannot prove it is safe to reorder " | ||||||
| 10344 | "memory operations"; | ||||||
| 10345 | }); | ||||||
| 10346 | LLVM_DEBUG(dbgs() << "LV: Too many memory checks needed.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Too many memory checks needed.\n" ; } } while (false); | ||||||
| 10347 | Hints.emitRemarkWithHints(); | ||||||
| 10348 | return false; | ||||||
| 10349 | } | ||||||
| 10350 | } | ||||||
| 10351 | |||||||
| 10352 | // Identify the diagnostic messages that should be produced. | ||||||
| 10353 | std::pair<StringRef, std::string> VecDiagMsg, IntDiagMsg; | ||||||
| 10354 | bool VectorizeLoop = true, InterleaveLoop = true; | ||||||
| 10355 | if (VF.Width.isScalar()) { | ||||||
| 10356 | LLVM_DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Vectorization is possible but not beneficial.\n" ; } } while (false); | ||||||
| 10357 | VecDiagMsg = std::make_pair( | ||||||
| 10358 | "VectorizationNotBeneficial", | ||||||
| 10359 | "the cost-model indicates that vectorization is not beneficial"); | ||||||
| 10360 | VectorizeLoop = false; | ||||||
| 10361 | } | ||||||
| 10362 | |||||||
| 10363 | if (!MaybeVF && UserIC > 1) { | ||||||
| 10364 | // Tell the user interleaving was avoided up-front, despite being explicitly | ||||||
| 10365 | // requested. | ||||||
| 10366 | LLVM_DEBUG(dbgs() << "LV: Ignoring UserIC, because vectorization and "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Ignoring UserIC, because vectorization and " "interleaving should be avoided up front\n"; } } while (false ) | ||||||
| 10367 | "interleaving should be avoided up front\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Ignoring UserIC, because vectorization and " "interleaving should be avoided up front\n"; } } while (false ); | ||||||
| 10368 | IntDiagMsg = std::make_pair( | ||||||
| 10369 | "InterleavingAvoided", | ||||||
| 10370 | "Ignoring UserIC, because interleaving was avoided up front"); | ||||||
| 10371 | InterleaveLoop = false; | ||||||
| 10372 | } else if (IC == 1 && UserIC <= 1) { | ||||||
| 10373 | // Tell the user interleaving is not beneficial. | ||||||
| 10374 | LLVM_DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving is not beneficial.\n" ; } } while (false); | ||||||
| 10375 | IntDiagMsg = std::make_pair( | ||||||
| 10376 | "InterleavingNotBeneficial", | ||||||
| 10377 | "the cost-model indicates that interleaving is not beneficial"); | ||||||
| 10378 | InterleaveLoop = false; | ||||||
| 10379 | if (UserIC == 1) { | ||||||
| 10380 | IntDiagMsg.first = "InterleavingNotBeneficialAndDisabled"; | ||||||
| 10381 | IntDiagMsg.second += | ||||||
| 10382 | " and is explicitly disabled or interleave count is set to 1"; | ||||||
| 10383 | } | ||||||
| 10384 | } else if (IC > 1 && UserIC == 1) { | ||||||
| 10385 | // Tell the user interleaving is beneficial, but it explicitly disabled. | ||||||
| 10386 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving is beneficial but is explicitly disabled." ; } } while (false) | ||||||
| 10387 | dbgs() << "LV: Interleaving is beneficial but is explicitly disabled.")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving is beneficial but is explicitly disabled." ; } } while (false); | ||||||
| 10388 | IntDiagMsg = std::make_pair( | ||||||
| 10389 | "InterleavingBeneficialButDisabled", | ||||||
| 10390 | "the cost-model indicates that interleaving is beneficial " | ||||||
| 10391 | "but is explicitly disabled or interleave count is set to 1"); | ||||||
| 10392 | InterleaveLoop = false; | ||||||
| 10393 | } | ||||||
| 10394 | |||||||
| 10395 | // Override IC if user provided an interleave count. | ||||||
| 10396 | IC = UserIC > 0 ? UserIC : IC; | ||||||
| 10397 | |||||||
| 10398 | // Emit diagnostic messages, if any. | ||||||
| 10399 | const char *VAPassName = Hints.vectorizeAnalysisPassName(); | ||||||
| 10400 | if (!VectorizeLoop && !InterleaveLoop) { | ||||||
| 10401 | // Do not vectorize or interleaving the loop. | ||||||
| 10402 | ORE->emit([&]() { | ||||||
| 10403 | return OptimizationRemarkMissed(VAPassName, VecDiagMsg.first, | ||||||
| 10404 | L->getStartLoc(), L->getHeader()) | ||||||
| 10405 | << VecDiagMsg.second; | ||||||
| 10406 | }); | ||||||
| 10407 | ORE->emit([&]() { | ||||||
| 10408 | return OptimizationRemarkMissed(LV_NAME"loop-vectorize", IntDiagMsg.first, | ||||||
| 10409 | L->getStartLoc(), L->getHeader()) | ||||||
| 10410 | << IntDiagMsg.second; | ||||||
| 10411 | }); | ||||||
| 10412 | return false; | ||||||
| 10413 | } else if (!VectorizeLoop && InterleaveLoop) { | ||||||
| 10414 | LLVM_DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleave Count is " << IC << '\n'; } } while (false); | ||||||
| 10415 | ORE->emit([&]() { | ||||||
| 10416 | return OptimizationRemarkAnalysis(VAPassName, VecDiagMsg.first, | ||||||
| 10417 | L->getStartLoc(), L->getHeader()) | ||||||
| 10418 | << VecDiagMsg.second; | ||||||
| 10419 | }); | ||||||
| 10420 | } else if (VectorizeLoop && !InterleaveLoop) { | ||||||
| 10421 | LLVM_DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Widthdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " << DebugLocStr << '\n'; } } while (false) | ||||||
| 10422 | << ") in " << DebugLocStr << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " << DebugLocStr << '\n'; } } while (false); | ||||||
| 10423 | ORE->emit([&]() { | ||||||
| 10424 | return OptimizationRemarkAnalysis(LV_NAME"loop-vectorize", IntDiagMsg.first, | ||||||
| 10425 | L->getStartLoc(), L->getHeader()) | ||||||
| 10426 | << IntDiagMsg.second; | ||||||
| 10427 | }); | ||||||
| 10428 | } else if (VectorizeLoop && InterleaveLoop) { | ||||||
| 10429 | LLVM_DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Widthdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " << DebugLocStr << '\n'; } } while (false) | ||||||
| 10430 | << ") in " << DebugLocStr << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " << DebugLocStr << '\n'; } } while (false); | ||||||
| 10431 | LLVM_DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleave Count is " << IC << '\n'; } } while (false); | ||||||
| 10432 | } | ||||||
| 10433 | |||||||
| 10434 | bool DisableRuntimeUnroll = false; | ||||||
| 10435 | MDNode *OrigLoopID = L->getLoopID(); | ||||||
| 10436 | { | ||||||
| 10437 | using namespace ore; | ||||||
| 10438 | if (!VectorizeLoop) { | ||||||
| 10439 | assert(IC > 1 && "interleave count should not be 1 or 0")(static_cast <bool> (IC > 1 && "interleave count should not be 1 or 0" ) ? void (0) : __assert_fail ("IC > 1 && \"interleave count should not be 1 or 0\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10439, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 10440 | // If we decided that it is not legal to vectorize the loop, then | ||||||
| 10441 | // interleave it. | ||||||
| 10442 | InnerLoopUnroller Unroller(L, PSE, LI, DT, TLI, TTI, AC, ORE, IC, &LVL, | ||||||
| 10443 | &CM, BFI, PSI, Checks); | ||||||
| 10444 | |||||||
| 10445 | VPlan &BestPlan = LVP.getBestPlanFor(VF.Width); | ||||||
| 10446 | LVP.executePlan(VF.Width, IC, BestPlan, Unroller, DT, false); | ||||||
| 10447 | |||||||
| 10448 | ORE->emit([&]() { | ||||||
| 10449 | return OptimizationRemark(LV_NAME"loop-vectorize", "Interleaved", L->getStartLoc(), | ||||||
| 10450 | L->getHeader()) | ||||||
| 10451 | << "interleaved loop (interleaved count: " | ||||||
| 10452 | << NV("InterleaveCount", IC) << ")"; | ||||||
| 10453 | }); | ||||||
| 10454 | } else { | ||||||
| 10455 | // If we decided that it is *legal* to vectorize the loop, then do it. | ||||||
| 10456 | |||||||
| 10457 | // Consider vectorizing the epilogue too if it's profitable. | ||||||
| 10458 | VectorizationFactor EpilogueVF = | ||||||
| 10459 | CM.selectEpilogueVectorizationFactor(VF.Width, LVP); | ||||||
| 10460 | if (EpilogueVF.Width.isVector()) { | ||||||
| 10461 | |||||||
| 10462 | // The first pass vectorizes the main loop and creates a scalar epilogue | ||||||
| 10463 | // to be vectorized by executing the plan (potentially with a different | ||||||
| 10464 | // factor) again shortly afterwards. | ||||||
| 10465 | EpilogueLoopVectorizationInfo EPI(VF.Width, IC, EpilogueVF.Width, 1); | ||||||
| 10466 | EpilogueVectorizerMainLoop MainILV(L, PSE, LI, DT, TLI, TTI, AC, ORE, | ||||||
| 10467 | EPI, &LVL, &CM, BFI, PSI, Checks); | ||||||
| 10468 | |||||||
| 10469 | VPlan &BestMainPlan = LVP.getBestPlanFor(EPI.MainLoopVF); | ||||||
| 10470 | LVP.executePlan(EPI.MainLoopVF, EPI.MainLoopUF, BestMainPlan, MainILV, | ||||||
| 10471 | DT, true); | ||||||
| 10472 | ++LoopsVectorized; | ||||||
| 10473 | |||||||
| 10474 | // Second pass vectorizes the epilogue and adjusts the control flow | ||||||
| 10475 | // edges from the first pass. | ||||||
| 10476 | EPI.MainLoopVF = EPI.EpilogueVF; | ||||||
| 10477 | EPI.MainLoopUF = EPI.EpilogueUF; | ||||||
| 10478 | EpilogueVectorizerEpilogueLoop EpilogILV(L, PSE, LI, DT, TLI, TTI, AC, | ||||||
| 10479 | ORE, EPI, &LVL, &CM, BFI, PSI, | ||||||
| 10480 | Checks); | ||||||
| 10481 | |||||||
| 10482 | VPlan &BestEpiPlan = LVP.getBestPlanFor(EPI.EpilogueVF); | ||||||
| 10483 | VPRegionBlock *VectorLoop = BestEpiPlan.getVectorLoopRegion(); | ||||||
| 10484 | VPBasicBlock *Header = VectorLoop->getEntryBasicBlock(); | ||||||
| 10485 | Header->setName("vec.epilog.vector.body"); | ||||||
| 10486 | |||||||
| 10487 | // Ensure that the start values for any VPWidenIntOrFpInductionRecipe, | ||||||
| 10488 | // VPWidenPointerInductionRecipe and VPReductionPHIRecipes are updated | ||||||
| 10489 | // before vectorizing the epilogue loop. | ||||||
| 10490 | for (VPRecipeBase &R : Header->phis()) { | ||||||
| 10491 | if (isa<VPCanonicalIVPHIRecipe>(&R)) | ||||||
| 10492 | continue; | ||||||
| 10493 | |||||||
| 10494 | Value *ResumeV = nullptr; | ||||||
| 10495 | // TODO: Move setting of resume values to prepareToExecute. | ||||||
| 10496 | if (auto *ReductionPhi = dyn_cast<VPReductionPHIRecipe>(&R)) { | ||||||
| 10497 | ResumeV = MainILV.getReductionResumeValue( | ||||||
| 10498 | ReductionPhi->getRecurrenceDescriptor()); | ||||||
| 10499 | } else { | ||||||
| 10500 | // Create induction resume values for both widened pointer and | ||||||
| 10501 | // integer/fp inductions and update the start value of the induction | ||||||
| 10502 | // recipes to use the resume value. | ||||||
| 10503 | PHINode *IndPhi = nullptr; | ||||||
| 10504 | const InductionDescriptor *ID; | ||||||
| 10505 | if (auto *Ind = dyn_cast<VPWidenPointerInductionRecipe>(&R)) { | ||||||
| 10506 | IndPhi = cast<PHINode>(Ind->getUnderlyingValue()); | ||||||
| 10507 | ID = &Ind->getInductionDescriptor(); | ||||||
| 10508 | } else { | ||||||
| 10509 | auto *WidenInd = cast<VPWidenIntOrFpInductionRecipe>(&R); | ||||||
| 10510 | IndPhi = WidenInd->getPHINode(); | ||||||
| 10511 | ID = &WidenInd->getInductionDescriptor(); | ||||||
| 10512 | } | ||||||
| 10513 | |||||||
| 10514 | ResumeV = MainILV.createInductionResumeValue( | ||||||
| 10515 | IndPhi, *ID, {EPI.MainLoopIterationCountCheck}); | ||||||
| 10516 | } | ||||||
| 10517 | assert(ResumeV && "Must have a resume value")(static_cast <bool> (ResumeV && "Must have a resume value" ) ? void (0) : __assert_fail ("ResumeV && \"Must have a resume value\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10517, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 10518 | VPValue *StartVal = BestEpiPlan.getOrAddExternalDef(ResumeV); | ||||||
| 10519 | cast<VPHeaderPHIRecipe>(&R)->setStartValue(StartVal); | ||||||
| 10520 | } | ||||||
| 10521 | |||||||
| 10522 | LVP.executePlan(EPI.EpilogueVF, EPI.EpilogueUF, BestEpiPlan, EpilogILV, | ||||||
| 10523 | DT, true); | ||||||
| 10524 | ++LoopsEpilogueVectorized; | ||||||
| 10525 | |||||||
| 10526 | if (!MainILV.areSafetyChecksAdded()) | ||||||
| 10527 | DisableRuntimeUnroll = true; | ||||||
| 10528 | } else { | ||||||
| 10529 | InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, | ||||||
| 10530 | VF.MinProfitableTripCount, IC, &LVL, &CM, BFI, | ||||||
| 10531 | PSI, Checks); | ||||||
| 10532 | |||||||
| 10533 | VPlan &BestPlan = LVP.getBestPlanFor(VF.Width); | ||||||
| 10534 | LVP.executePlan(VF.Width, IC, BestPlan, LB, DT, false); | ||||||
| 10535 | ++LoopsVectorized; | ||||||
| 10536 | |||||||
| 10537 | // Add metadata to disable runtime unrolling a scalar loop when there | ||||||
| 10538 | // are no runtime checks about strides and memory. A scalar loop that is | ||||||
| 10539 | // rarely used is not worth unrolling. | ||||||
| 10540 | if (!LB.areSafetyChecksAdded()) | ||||||
| 10541 | DisableRuntimeUnroll = true; | ||||||
| 10542 | } | ||||||
| 10543 | // Report the vectorization decision. | ||||||
| 10544 | ORE->emit([&]() { | ||||||
| 10545 | return OptimizationRemark(LV_NAME"loop-vectorize", "Vectorized", L->getStartLoc(), | ||||||
| 10546 | L->getHeader()) | ||||||
| 10547 | << "vectorized loop (vectorization width: " | ||||||
| 10548 | << NV("VectorizationFactor", VF.Width) | ||||||
| 10549 | << ", interleaved count: " << NV("InterleaveCount", IC) << ")"; | ||||||
| 10550 | }); | ||||||
| 10551 | } | ||||||
| 10552 | |||||||
| 10553 | if (ORE->allowExtraAnalysis(LV_NAME"loop-vectorize")) | ||||||
| 10554 | checkMixedPrecision(L, ORE); | ||||||
| 10555 | } | ||||||
| 10556 | |||||||
| 10557 | std::optional<MDNode *> RemainderLoopID = | ||||||
| 10558 | makeFollowupLoopID(OrigLoopID, {LLVMLoopVectorizeFollowupAll, | ||||||
| 10559 | LLVMLoopVectorizeFollowupEpilogue}); | ||||||
| 10560 | if (RemainderLoopID) { | ||||||
| 10561 | L->setLoopID(*RemainderLoopID); | ||||||
| 10562 | } else { | ||||||
| 10563 | if (DisableRuntimeUnroll) | ||||||
| 10564 | AddRuntimeUnrollDisableMetaData(L); | ||||||
| 10565 | |||||||
| 10566 | // Mark the loop as already vectorized to avoid vectorizing again. | ||||||
| 10567 | Hints.setAlreadyVectorized(); | ||||||
| 10568 | } | ||||||
| 10569 | |||||||
| 10570 | assert(!verifyFunction(*L->getHeader()->getParent(), &dbgs()))(static_cast <bool> (!verifyFunction(*L->getHeader() ->getParent(), &dbgs())) ? void (0) : __assert_fail ("!verifyFunction(*L->getHeader()->getParent(), &dbgs())" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10570, __extension__ __PRETTY_FUNCTION__)); | ||||||
| 10571 | return true; | ||||||
| 10572 | } | ||||||
| 10573 | |||||||
| 10574 | LoopVectorizeResult LoopVectorizePass::runImpl( | ||||||
| 10575 | Function &F, ScalarEvolution &SE_, LoopInfo &LI_, TargetTransformInfo &TTI_, | ||||||
| 10576 | DominatorTree &DT_, BlockFrequencyInfo &BFI_, TargetLibraryInfo *TLI_, | ||||||
| 10577 | DemandedBits &DB_, AssumptionCache &AC_, LoopAccessInfoManager &LAIs_, | ||||||
| 10578 | OptimizationRemarkEmitter &ORE_, ProfileSummaryInfo *PSI_) { | ||||||
| 10579 | SE = &SE_; | ||||||
| 10580 | LI = &LI_; | ||||||
| 10581 | TTI = &TTI_; | ||||||
| 10582 | DT = &DT_; | ||||||
| 10583 | BFI = &BFI_; | ||||||
| 10584 | TLI = TLI_; | ||||||
| 10585 | AC = &AC_; | ||||||
| 10586 | LAIs = &LAIs_; | ||||||
| 10587 | DB = &DB_; | ||||||
| 10588 | ORE = &ORE_; | ||||||
| 10589 | PSI = PSI_; | ||||||
| 10590 | |||||||
| 10591 | // Don't attempt if | ||||||
| 10592 | // 1. the target claims to have no vector registers, and | ||||||
| 10593 | // 2. interleaving won't help ILP. | ||||||
| 10594 | // | ||||||
| 10595 | // The second condition is necessary because, even if the target has no | ||||||
| 10596 | // vector registers, loop vectorization may still enable scalar | ||||||
| 10597 | // interleaving. | ||||||
| 10598 | if (!TTI->getNumberOfRegisters(TTI->getRegisterClassForType(true)) && | ||||||
| 10599 | TTI->getMaxInterleaveFactor(1) < 2) | ||||||
| 10600 | return LoopVectorizeResult(false, false); | ||||||
| 10601 | |||||||
| 10602 | bool Changed = false, CFGChanged = false; | ||||||
| 10603 | |||||||
| 10604 | // The vectorizer requires loops to be in simplified form. | ||||||
| 10605 | // Since simplification may add new inner loops, it has to run before the | ||||||
| 10606 | // legality and profitability checks. This means running the loop vectorizer | ||||||
| 10607 | // will simplify all loops, regardless of whether anything end up being | ||||||
| 10608 | // vectorized. | ||||||
| 10609 | for (const auto &L : *LI) | ||||||
| 10610 | Changed |= CFGChanged |= | ||||||
| 10611 | simplifyLoop(L, DT, LI, SE, AC, nullptr, false /* PreserveLCSSA */); | ||||||
| 10612 | |||||||
| 10613 | // Build up a worklist of inner-loops to vectorize. This is necessary as | ||||||
| 10614 | // the act of vectorizing or partially unrolling a loop creates new loops | ||||||
| 10615 | // and can invalidate iterators across the loops. | ||||||
| 10616 | SmallVector<Loop *, 8> Worklist; | ||||||
| 10617 | |||||||
| 10618 | for (Loop *L : *LI) | ||||||
| 10619 | collectSupportedLoops(*L, LI, ORE, Worklist); | ||||||
| 10620 | |||||||
| 10621 | LoopsAnalyzed += Worklist.size(); | ||||||
| 10622 | |||||||
| 10623 | // Now walk the identified inner loops. | ||||||
| 10624 | while (!Worklist.empty()) { | ||||||
| 10625 | Loop *L = Worklist.pop_back_val(); | ||||||
| 10626 | |||||||
| 10627 | // For the inner loops we actually process, form LCSSA to simplify the | ||||||
| 10628 | // transform. | ||||||
| 10629 | Changed |= formLCSSARecursively(*L, *DT, LI, SE); | ||||||
| 10630 | |||||||
| 10631 | Changed |= CFGChanged |= processLoop(L); | ||||||
| 10632 | } | ||||||
| 10633 | |||||||
| 10634 | // Process each loop nest in the function. | ||||||
| 10635 | return LoopVectorizeResult(Changed, CFGChanged); | ||||||
| 10636 | } | ||||||
| 10637 | |||||||
| 10638 | PreservedAnalyses LoopVectorizePass::run(Function &F, | ||||||
| 10639 | FunctionAnalysisManager &AM) { | ||||||
| 10640 | auto &LI = AM.getResult<LoopAnalysis>(F); | ||||||
| 10641 | // There are no loops in the function. Return before computing other expensive | ||||||
| 10642 | // analyses. | ||||||
| 10643 | if (LI.empty()) | ||||||
| 10644 | return PreservedAnalyses::all(); | ||||||
| 10645 | auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F); | ||||||
| 10646 | auto &TTI = AM.getResult<TargetIRAnalysis>(F); | ||||||
| 10647 | auto &DT = AM.getResult<DominatorTreeAnalysis>(F); | ||||||
| 10648 | auto &BFI = AM.getResult<BlockFrequencyAnalysis>(F); | ||||||
| 10649 | auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); | ||||||
| 10650 | auto &AC = AM.getResult<AssumptionAnalysis>(F); | ||||||
| 10651 | auto &DB = AM.getResult<DemandedBitsAnalysis>(F); | ||||||
| 10652 | auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F); | ||||||
| 10653 | |||||||
| 10654 | LoopAccessInfoManager &LAIs = AM.getResult<LoopAccessAnalysis>(F); | ||||||
| 10655 | auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F); | ||||||
| 10656 | ProfileSummaryInfo *PSI = | ||||||
| 10657 | MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent()); | ||||||
| 10658 | LoopVectorizeResult Result = | ||||||
| 10659 | runImpl(F, SE, LI, TTI, DT, BFI, &TLI, DB, AC, LAIs, ORE, PSI); | ||||||
| 10660 | if (!Result.MadeAnyChange) | ||||||
| 10661 | return PreservedAnalyses::all(); | ||||||
| 10662 | PreservedAnalyses PA; | ||||||
| 10663 | |||||||
| 10664 | // We currently do not preserve loopinfo/dominator analyses with outer loop | ||||||
| 10665 | // vectorization. Until this is addressed, mark these analyses as preserved | ||||||
| 10666 | // only for non-VPlan-native path. | ||||||
| 10667 | // TODO: Preserve Loop and Dominator analyses for VPlan-native path. | ||||||
| 10668 | if (!EnableVPlanNativePath) { | ||||||
| 10669 | PA.preserve<LoopAnalysis>(); | ||||||
| 10670 | PA.preserve<DominatorTreeAnalysis>(); | ||||||
| 10671 | } | ||||||
| 10672 | |||||||
| 10673 | if (Result.MadeCFGChange) { | ||||||
| 10674 | // Making CFG changes likely means a loop got vectorized. Indicate that | ||||||
| 10675 | // extra simplification passes should be run. | ||||||
| 10676 | // TODO: MadeCFGChanges is not a prefect proxy. Extra passes should only | ||||||
| 10677 | // be run if runtime checks have been added. | ||||||
| 10678 | AM.getResult<ShouldRunExtraVectorPasses>(F); | ||||||
| 10679 | PA.preserve<ShouldRunExtraVectorPasses>(); | ||||||
| 10680 | } else { | ||||||
| 10681 | PA.preserveSet<CFGAnalyses>(); | ||||||
| 10682 | } | ||||||
| 10683 | return PA; | ||||||
| 10684 | } | ||||||
| 10685 | |||||||
| 10686 | void LoopVectorizePass::printPipeline( | ||||||
| 10687 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { | ||||||
| 10688 | static_cast<PassInfoMixin<LoopVectorizePass> *>(this)->printPipeline( | ||||||
| 10689 | OS, MapClassName2PassName); | ||||||
| 10690 | |||||||
| 10691 | OS << "<"; | ||||||
| 10692 | OS << (InterleaveOnlyWhenForced ? "" : "no-") << "interleave-forced-only;"; | ||||||
| 10693 | OS << (VectorizeOnlyWhenForced ? "" : "no-") << "vectorize-forced-only;"; | ||||||
| 10694 | OS << ">"; | ||||||
| 10695 | } |
| 1 | // Components for manipulating sequences of characters -*- C++ -*- |
| 2 | |
| 3 | // Copyright (C) 1997-2020 Free Software Foundation, Inc. |
| 4 | // |
| 5 | // This file is part of the GNU ISO C++ Library. This library is free |
| 6 | // software; you can redistribute it and/or modify it under the |
| 7 | // terms of the GNU General Public License as published by the |
| 8 | // Free Software Foundation; either version 3, or (at your option) |
| 9 | // any later version. |
| 10 | |
| 11 | // This library is distributed in the hope that it will be useful, |
| 12 | // but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 13 | // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 14 | // GNU General Public License for more details. |
| 15 | |
| 16 | // Under Section 7 of GPL version 3, you are granted additional |
| 17 | // permissions described in the GCC Runtime Library Exception, version |
| 18 | // 3.1, as published by the Free Software Foundation. |
| 19 | |
| 20 | // You should have received a copy of the GNU General Public License and |
| 21 | // a copy of the GCC Runtime Library Exception along with this program; |
| 22 | // see the files COPYING3 and COPYING.RUNTIME respectively. If not, see |
| 23 | // <http://www.gnu.org/licenses/>. |
| 24 | |
| 25 | /** @file bits/basic_string.h |
| 26 | * This is an internal header file, included by other library headers. |
| 27 | * Do not attempt to use it directly. @headername{string} |
| 28 | */ |
| 29 | |
| 30 | // |
| 31 | // ISO C++ 14882: 21 Strings library |
| 32 | // |
| 33 | |
| 34 | #ifndef _BASIC_STRING_H1 |
| 35 | #define _BASIC_STRING_H1 1 |
| 36 | |
| 37 | #pragma GCC system_header |
| 38 | |
| 39 | #include <ext/atomicity.h> |
| 40 | #include <ext/alloc_traits.h> |
| 41 | #include <debug/debug.h> |
| 42 | |
| 43 | #if __cplusplus201703L >= 201103L |
| 44 | #include <initializer_list> |
| 45 | #endif |
| 46 | |
| 47 | #if __cplusplus201703L >= 201703L |
| 48 | # include <string_view> |
| 49 | #endif |
| 50 | |
| 51 | |
| 52 | namespace std _GLIBCXX_VISIBILITY(default)__attribute__ ((__visibility__ ("default"))) |
| 53 | { |
| 54 | _GLIBCXX_BEGIN_NAMESPACE_VERSION |
| 55 | |
| 56 | #if _GLIBCXX_USE_CXX11_ABI1 |
| 57 | _GLIBCXX_BEGIN_NAMESPACE_CXX11namespace __cxx11 { |
| 58 | /** |
| 59 | * @class basic_string basic_string.h <string> |
| 60 | * @brief Managing sequences of characters and character-like objects. |
| 61 | * |
| 62 | * @ingroup strings |
| 63 | * @ingroup sequences |
| 64 | * |
| 65 | * @tparam _CharT Type of character |
| 66 | * @tparam _Traits Traits for character type, defaults to |
| 67 | * char_traits<_CharT>. |
| 68 | * @tparam _Alloc Allocator type, defaults to allocator<_CharT>. |
| 69 | * |
| 70 | * Meets the requirements of a <a href="tables.html#65">container</a>, a |
| 71 | * <a href="tables.html#66">reversible container</a>, and a |
| 72 | * <a href="tables.html#67">sequence</a>. Of the |
| 73 | * <a href="tables.html#68">optional sequence requirements</a>, only |
| 74 | * @c push_back, @c at, and @c %array access are supported. |
| 75 | */ |
| 76 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 77 | class basic_string |
| 78 | { |
| 79 | typedef typename __gnu_cxx::__alloc_traits<_Alloc>::template |
| 80 | rebind<_CharT>::other _Char_alloc_type; |
| 81 | typedef __gnu_cxx::__alloc_traits<_Char_alloc_type> _Alloc_traits; |
| 82 | |
| 83 | // Types: |
| 84 | public: |
| 85 | typedef _Traits traits_type; |
| 86 | typedef typename _Traits::char_type value_type; |
| 87 | typedef _Char_alloc_type allocator_type; |
| 88 | typedef typename _Alloc_traits::size_type size_type; |
| 89 | typedef typename _Alloc_traits::difference_type difference_type; |
| 90 | typedef typename _Alloc_traits::reference reference; |
| 91 | typedef typename _Alloc_traits::const_reference const_reference; |
| 92 | typedef typename _Alloc_traits::pointer pointer; |
| 93 | typedef typename _Alloc_traits::const_pointer const_pointer; |
| 94 | typedef __gnu_cxx::__normal_iterator<pointer, basic_string> iterator; |
| 95 | typedef __gnu_cxx::__normal_iterator<const_pointer, basic_string> |
| 96 | const_iterator; |
| 97 | typedef std::reverse_iterator<const_iterator> const_reverse_iterator; |
| 98 | typedef std::reverse_iterator<iterator> reverse_iterator; |
| 99 | |
| 100 | /// Value returned by various member functions when they fail. |
| 101 | static const size_type npos = static_cast<size_type>(-1); |
| 102 | |
| 103 | protected: |
| 104 | // type used for positions in insert, erase etc. |
| 105 | #if __cplusplus201703L < 201103L |
| 106 | typedef iterator __const_iterator; |
| 107 | #else |
| 108 | typedef const_iterator __const_iterator; |
| 109 | #endif |
| 110 | |
| 111 | private: |
| 112 | #if __cplusplus201703L >= 201703L |
| 113 | // A helper type for avoiding boiler-plate. |
| 114 | typedef basic_string_view<_CharT, _Traits> __sv_type; |
| 115 | |
| 116 | template<typename _Tp, typename _Res> |
| 117 | using _If_sv = enable_if_t< |
| 118 | __and_<is_convertible<const _Tp&, __sv_type>, |
| 119 | __not_<is_convertible<const _Tp*, const basic_string*>>, |
| 120 | __not_<is_convertible<const _Tp&, const _CharT*>>>::value, |
| 121 | _Res>; |
| 122 | |
| 123 | // Allows an implicit conversion to __sv_type. |
| 124 | static __sv_type |
| 125 | _S_to_string_view(__sv_type __svt) noexcept |
| 126 | { return __svt; } |
| 127 | |
| 128 | // Wraps a string_view by explicit conversion and thus |
| 129 | // allows to add an internal constructor that does not |
| 130 | // participate in overload resolution when a string_view |
| 131 | // is provided. |
| 132 | struct __sv_wrapper |
| 133 | { |
| 134 | explicit __sv_wrapper(__sv_type __sv) noexcept : _M_sv(__sv) { } |
| 135 | __sv_type _M_sv; |
| 136 | }; |
| 137 | |
| 138 | /** |
| 139 | * @brief Only internally used: Construct string from a string view |
| 140 | * wrapper. |
| 141 | * @param __svw string view wrapper. |
| 142 | * @param __a Allocator to use. |
| 143 | */ |
| 144 | explicit |
| 145 | basic_string(__sv_wrapper __svw, const _Alloc& __a) |
| 146 | : basic_string(__svw._M_sv.data(), __svw._M_sv.size(), __a) { } |
| 147 | #endif |
| 148 | |
| 149 | // Use empty-base optimization: http://www.cantrip.org/emptyopt.html |
| 150 | struct _Alloc_hider : allocator_type // TODO check __is_final |
| 151 | { |
| 152 | #if __cplusplus201703L < 201103L |
| 153 | _Alloc_hider(pointer __dat, const _Alloc& __a = _Alloc()) |
| 154 | : allocator_type(__a), _M_p(__dat) { } |
| 155 | #else |
| 156 | _Alloc_hider(pointer __dat, const _Alloc& __a) |
| 157 | : allocator_type(__a), _M_p(__dat) { } |
| 158 | |
| 159 | _Alloc_hider(pointer __dat, _Alloc&& __a = _Alloc()) |
| 160 | : allocator_type(std::move(__a)), _M_p(__dat) { } |
| 161 | #endif |
| 162 | |
| 163 | pointer _M_p; // The actual data. |
| 164 | }; |
| 165 | |
| 166 | _Alloc_hider _M_dataplus; |
| 167 | size_type _M_string_length; |
| 168 | |
| 169 | enum { _S_local_capacity = 15 / sizeof(_CharT) }; |
| 170 | |
| 171 | union |
| 172 | { |
| 173 | _CharT _M_local_buf[_S_local_capacity + 1]; |
| 174 | size_type _M_allocated_capacity; |
| 175 | }; |
| 176 | |
| 177 | void |
| 178 | _M_data(pointer __p) |
| 179 | { _M_dataplus._M_p = __p; } |
| 180 | |
| 181 | void |
| 182 | _M_length(size_type __length) |
| 183 | { _M_string_length = __length; } |
| 184 | |
| 185 | pointer |
| 186 | _M_data() const |
| 187 | { return _M_dataplus._M_p; } |
| 188 | |
| 189 | pointer |
| 190 | _M_local_data() |
| 191 | { |
| 192 | #if __cplusplus201703L >= 201103L |
| 193 | return std::pointer_traits<pointer>::pointer_to(*_M_local_buf); |
| 194 | #else |
| 195 | return pointer(_M_local_buf); |
| 196 | #endif |
| 197 | } |
| 198 | |
| 199 | const_pointer |
| 200 | _M_local_data() const |
| 201 | { |
| 202 | #if __cplusplus201703L >= 201103L |
| 203 | return std::pointer_traits<const_pointer>::pointer_to(*_M_local_buf); |
| 204 | #else |
| 205 | return const_pointer(_M_local_buf); |
| 206 | #endif |
| 207 | } |
| 208 | |
| 209 | void |
| 210 | _M_capacity(size_type __capacity) |
| 211 | { _M_allocated_capacity = __capacity; } |
| 212 | |
| 213 | void |
| 214 | _M_set_length(size_type __n) |
| 215 | { |
| 216 | _M_length(__n); |
| 217 | traits_type::assign(_M_data()[__n], _CharT()); |
| 218 | } |
| 219 | |
| 220 | bool |
| 221 | _M_is_local() const |
| 222 | { return _M_data() == _M_local_data(); } |
| 223 | |
| 224 | // Create & Destroy |
| 225 | pointer |
| 226 | _M_create(size_type&, size_type); |
| 227 | |
| 228 | void |
| 229 | _M_dispose() |
| 230 | { |
| 231 | if (!_M_is_local()) |
| 232 | _M_destroy(_M_allocated_capacity); |
| 233 | } |
| 234 | |
| 235 | void |
| 236 | _M_destroy(size_type __size) throw() |
| 237 | { _Alloc_traits::deallocate(_M_get_allocator(), _M_data(), __size + 1); } |
| 238 | |
| 239 | // _M_construct_aux is used to implement the 21.3.1 para 15 which |
| 240 | // requires special behaviour if _InIterator is an integral type |
| 241 | template<typename _InIterator> |
| 242 | void |
| 243 | _M_construct_aux(_InIterator __beg, _InIterator __end, |
| 244 | std::__false_type) |
| 245 | { |
| 246 | typedef typename iterator_traits<_InIterator>::iterator_category _Tag; |
| 247 | _M_construct(__beg, __end, _Tag()); |
| 248 | } |
| 249 | |
| 250 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
| 251 | // 438. Ambiguity in the "do the right thing" clause |
| 252 | template<typename _Integer> |
| 253 | void |
| 254 | _M_construct_aux(_Integer __beg, _Integer __end, std::__true_type) |
| 255 | { _M_construct_aux_2(static_cast<size_type>(__beg), __end); } |
| 256 | |
| 257 | void |
| 258 | _M_construct_aux_2(size_type __req, _CharT __c) |
| 259 | { _M_construct(__req, __c); } |
| 260 | |
| 261 | template<typename _InIterator> |
| 262 | void |
| 263 | _M_construct(_InIterator __beg, _InIterator __end) |
| 264 | { |
| 265 | typedef typename std::__is_integer<_InIterator>::__type _Integral; |
| 266 | _M_construct_aux(__beg, __end, _Integral()); |
| 267 | } |
| 268 | |
| 269 | // For Input Iterators, used in istreambuf_iterators, etc. |
| 270 | template<typename _InIterator> |
| 271 | void |
| 272 | _M_construct(_InIterator __beg, _InIterator __end, |
| 273 | std::input_iterator_tag); |
| 274 | |
| 275 | // For forward_iterators up to random_access_iterators, used for |
| 276 | // string::iterator, _CharT*, etc. |
| 277 | template<typename _FwdIterator> |
| 278 | void |
| 279 | _M_construct(_FwdIterator __beg, _FwdIterator __end, |
| 280 | std::forward_iterator_tag); |
| 281 | |
| 282 | void |
| 283 | _M_construct(size_type __req, _CharT __c); |
| 284 | |
| 285 | allocator_type& |
| 286 | _M_get_allocator() |
| 287 | { return _M_dataplus; } |
| 288 | |
| 289 | const allocator_type& |
| 290 | _M_get_allocator() const |
| 291 | { return _M_dataplus; } |
| 292 | |
| 293 | private: |
| 294 | |
| 295 | #ifdef _GLIBCXX_DISAMBIGUATE_REPLACE_INST |
| 296 | // The explicit instantiations in misc-inst.cc require this due to |
| 297 | // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64063 |
| 298 | template<typename _Tp, bool _Requires = |
| 299 | !__are_same<_Tp, _CharT*>::__value |
| 300 | && !__are_same<_Tp, const _CharT*>::__value |
| 301 | && !__are_same<_Tp, iterator>::__value |
| 302 | && !__are_same<_Tp, const_iterator>::__value> |
| 303 | struct __enable_if_not_native_iterator |
| 304 | { typedef basic_string& __type; }; |
| 305 | template<typename _Tp> |
| 306 | struct __enable_if_not_native_iterator<_Tp, false> { }; |
| 307 | #endif |
| 308 | |
| 309 | size_type |
| 310 | _M_check(size_type __pos, const char* __s) const |
| 311 | { |
| 312 | if (__pos > this->size()) |
| 313 | __throw_out_of_range_fmt(__N("%s: __pos (which is %zu) > "("%s: __pos (which is %zu) > " "this->size() (which is %zu)" ) |
| 314 | "this->size() (which is %zu)")("%s: __pos (which is %zu) > " "this->size() (which is %zu)" ), |
| 315 | __s, __pos, this->size()); |
| 316 | return __pos; |
| 317 | } |
| 318 | |
| 319 | void |
| 320 | _M_check_length(size_type __n1, size_type __n2, const char* __s) const |
| 321 | { |
| 322 | if (this->max_size() - (this->size() - __n1) < __n2) |
| 323 | __throw_length_error(__N(__s)(__s)); |
| 324 | } |
| 325 | |
| 326 | |
| 327 | // NB: _M_limit doesn't check for a bad __pos value. |
| 328 | size_type |
| 329 | _M_limit(size_type __pos, size_type __off) const _GLIBCXX_NOEXCEPTnoexcept |
| 330 | { |
| 331 | const bool __testoff = __off < this->size() - __pos; |
| 332 | return __testoff ? __off : this->size() - __pos; |
| 333 | } |
| 334 | |
| 335 | // True if _Rep and source do not overlap. |
| 336 | bool |
| 337 | _M_disjunct(const _CharT* __s) const _GLIBCXX_NOEXCEPTnoexcept |
| 338 | { |
| 339 | return (less<const _CharT*>()(__s, _M_data()) |
| 340 | || less<const _CharT*>()(_M_data() + this->size(), __s)); |
| 341 | } |
| 342 | |
| 343 | // When __n = 1 way faster than the general multichar |
| 344 | // traits_type::copy/move/assign. |
| 345 | static void |
| 346 | _S_copy(_CharT* __d, const _CharT* __s, size_type __n) |
| 347 | { |
| 348 | if (__n == 1) |
| 349 | traits_type::assign(*__d, *__s); |
| 350 | else |
| 351 | traits_type::copy(__d, __s, __n); |
| 352 | } |
| 353 | |
| 354 | static void |
| 355 | _S_move(_CharT* __d, const _CharT* __s, size_type __n) |
| 356 | { |
| 357 | if (__n == 1) |
| 358 | traits_type::assign(*__d, *__s); |
| 359 | else |
| 360 | traits_type::move(__d, __s, __n); |
| 361 | } |
| 362 | |
| 363 | static void |
| 364 | _S_assign(_CharT* __d, size_type __n, _CharT __c) |
| 365 | { |
| 366 | if (__n == 1) |
| 367 | traits_type::assign(*__d, __c); |
| 368 | else |
| 369 | traits_type::assign(__d, __n, __c); |
| 370 | } |
| 371 | |
| 372 | // _S_copy_chars is a separate template to permit specialization |
| 373 | // to optimize for the common case of pointers as iterators. |
| 374 | template<class _Iterator> |
| 375 | static void |
| 376 | _S_copy_chars(_CharT* __p, _Iterator __k1, _Iterator __k2) |
| 377 | { |
| 378 | for (; __k1 != __k2; ++__k1, (void)++__p) |
| 379 | traits_type::assign(*__p, *__k1); // These types are off. |
| 380 | } |
| 381 | |
| 382 | static void |
| 383 | _S_copy_chars(_CharT* __p, iterator __k1, iterator __k2) _GLIBCXX_NOEXCEPTnoexcept |
| 384 | { _S_copy_chars(__p, __k1.base(), __k2.base()); } |
| 385 | |
| 386 | static void |
| 387 | _S_copy_chars(_CharT* __p, const_iterator __k1, const_iterator __k2) |
| 388 | _GLIBCXX_NOEXCEPTnoexcept |
| 389 | { _S_copy_chars(__p, __k1.base(), __k2.base()); } |
| 390 | |
| 391 | static void |
| 392 | _S_copy_chars(_CharT* __p, _CharT* __k1, _CharT* __k2) _GLIBCXX_NOEXCEPTnoexcept |
| 393 | { _S_copy(__p, __k1, __k2 - __k1); } |
| 394 | |
| 395 | static void |
| 396 | _S_copy_chars(_CharT* __p, const _CharT* __k1, const _CharT* __k2) |
| 397 | _GLIBCXX_NOEXCEPTnoexcept |
| 398 | { _S_copy(__p, __k1, __k2 - __k1); } |
| 399 | |
| 400 | static int |
| 401 | _S_compare(size_type __n1, size_type __n2) _GLIBCXX_NOEXCEPTnoexcept |
| 402 | { |
| 403 | const difference_type __d = difference_type(__n1 - __n2); |
| 404 | |
| 405 | if (__d > __gnu_cxx::__numeric_traits<int>::__max) |
| 406 | return __gnu_cxx::__numeric_traits<int>::__max; |
| 407 | else if (__d < __gnu_cxx::__numeric_traits<int>::__min) |
| 408 | return __gnu_cxx::__numeric_traits<int>::__min; |
| 409 | else |
| 410 | return int(__d); |
| 411 | } |
| 412 | |
| 413 | void |
| 414 | _M_assign(const basic_string&); |
| 415 | |
| 416 | void |
| 417 | _M_mutate(size_type __pos, size_type __len1, const _CharT* __s, |
| 418 | size_type __len2); |
| 419 | |
| 420 | void |
| 421 | _M_erase(size_type __pos, size_type __n); |
| 422 | |
| 423 | public: |
| 424 | // Construct/copy/destroy: |
| 425 | // NB: We overload ctors in some cases instead of using default |
| 426 | // arguments, per 17.4.4.4 para. 2 item 2. |
| 427 | |
| 428 | /** |
| 429 | * @brief Default constructor creates an empty string. |
| 430 | */ |
| 431 | basic_string() |
| 432 | _GLIBCXX_NOEXCEPT_IF(is_nothrow_default_constructible<_Alloc>::value)noexcept(is_nothrow_default_constructible<_Alloc>::value ) |
| 433 | : _M_dataplus(_M_local_data()) |
| 434 | { _M_set_length(0); } |
| 435 | |
| 436 | /** |
| 437 | * @brief Construct an empty string using allocator @a a. |
| 438 | */ |
| 439 | explicit |
| 440 | basic_string(const _Alloc& __a) _GLIBCXX_NOEXCEPTnoexcept |
| 441 | : _M_dataplus(_M_local_data(), __a) |
| 442 | { _M_set_length(0); } |
| 443 | |
| 444 | /** |
| 445 | * @brief Construct string with copy of value of @a __str. |
| 446 | * @param __str Source string. |
| 447 | */ |
| 448 | basic_string(const basic_string& __str) |
| 449 | : _M_dataplus(_M_local_data(), |
| 450 | _Alloc_traits::_S_select_on_copy(__str._M_get_allocator())) |
| 451 | { _M_construct(__str._M_data(), __str._M_data() + __str.length()); } |
| 452 | |
| 453 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
| 454 | // 2583. no way to supply an allocator for basic_string(str, pos) |
| 455 | /** |
| 456 | * @brief Construct string as copy of a substring. |
| 457 | * @param __str Source string. |
| 458 | * @param __pos Index of first character to copy from. |
| 459 | * @param __a Allocator to use. |
| 460 | */ |
| 461 | basic_string(const basic_string& __str, size_type __pos, |
| 462 | const _Alloc& __a = _Alloc()) |
| 463 | : _M_dataplus(_M_local_data(), __a) |
| 464 | { |
| 465 | const _CharT* __start = __str._M_data() |
| 466 | + __str._M_check(__pos, "basic_string::basic_string"); |
| 467 | _M_construct(__start, __start + __str._M_limit(__pos, npos)); |
| 468 | } |
| 469 | |
| 470 | /** |
| 471 | * @brief Construct string as copy of a substring. |
| 472 | * @param __str Source string. |
| 473 | * @param __pos Index of first character to copy from. |
| 474 | * @param __n Number of characters to copy. |
| 475 | */ |
| 476 | basic_string(const basic_string& __str, size_type __pos, |
| 477 | size_type __n) |
| 478 | : _M_dataplus(_M_local_data()) |
| 479 | { |
| 480 | const _CharT* __start = __str._M_data() |
| 481 | + __str._M_check(__pos, "basic_string::basic_string"); |
| 482 | _M_construct(__start, __start + __str._M_limit(__pos, __n)); |
| 483 | } |
| 484 | |
| 485 | /** |
| 486 | * @brief Construct string as copy of a substring. |
| 487 | * @param __str Source string. |
| 488 | * @param __pos Index of first character to copy from. |
| 489 | * @param __n Number of characters to copy. |
| 490 | * @param __a Allocator to use. |
| 491 | */ |
| 492 | basic_string(const basic_string& __str, size_type __pos, |
| 493 | size_type __n, const _Alloc& __a) |
| 494 | : _M_dataplus(_M_local_data(), __a) |
| 495 | { |
| 496 | const _CharT* __start |
| 497 | = __str._M_data() + __str._M_check(__pos, "string::string"); |
| 498 | _M_construct(__start, __start + __str._M_limit(__pos, __n)); |
| 499 | } |
| 500 | |
| 501 | /** |
| 502 | * @brief Construct string initialized by a character %array. |
| 503 | * @param __s Source character %array. |
| 504 | * @param __n Number of characters to copy. |
| 505 | * @param __a Allocator to use (default is default allocator). |
| 506 | * |
| 507 | * NB: @a __s must have at least @a __n characters, '\\0' |
| 508 | * has no special meaning. |
| 509 | */ |
| 510 | basic_string(const _CharT* __s, size_type __n, |
| 511 | const _Alloc& __a = _Alloc()) |
| 512 | : _M_dataplus(_M_local_data(), __a) |
| 513 | { _M_construct(__s, __s + __n); } |
| 514 | |
| 515 | /** |
| 516 | * @brief Construct string as copy of a C string. |
| 517 | * @param __s Source C string. |
| 518 | * @param __a Allocator to use (default is default allocator). |
| 519 | */ |
| 520 | #if __cpp_deduction_guides201703L && ! defined _GLIBCXX_DEFINING_STRING_INSTANTIATIONS |
| 521 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
| 522 | // 3076. basic_string CTAD ambiguity |
| 523 | template<typename = _RequireAllocator<_Alloc>> |
| 524 | #endif |
| 525 | basic_string(const _CharT* __s, const _Alloc& __a = _Alloc()) |
| 526 | : _M_dataplus(_M_local_data(), __a) |
| 527 | { _M_construct(__s, __s ? __s + traits_type::length(__s) : __s+npos); } |
| 528 | |
| 529 | /** |
| 530 | * @brief Construct string as multiple characters. |
| 531 | * @param __n Number of characters. |
| 532 | * @param __c Character to use. |
| 533 | * @param __a Allocator to use (default is default allocator). |
| 534 | */ |
| 535 | #if __cpp_deduction_guides201703L && ! defined _GLIBCXX_DEFINING_STRING_INSTANTIATIONS |
| 536 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
| 537 | // 3076. basic_string CTAD ambiguity |
| 538 | template<typename = _RequireAllocator<_Alloc>> |
| 539 | #endif |
| 540 | basic_string(size_type __n, _CharT __c, const _Alloc& __a = _Alloc()) |
| 541 | : _M_dataplus(_M_local_data(), __a) |
| 542 | { _M_construct(__n, __c); } |
| 543 | |
| 544 | #if __cplusplus201703L >= 201103L |
| 545 | /** |
| 546 | * @brief Move construct string. |
| 547 | * @param __str Source string. |
| 548 | * |
| 549 | * The newly-created string contains the exact contents of @a __str. |
| 550 | * @a __str is a valid, but unspecified string. |
| 551 | **/ |
| 552 | basic_string(basic_string&& __str) noexcept |
| 553 | : _M_dataplus(_M_local_data(), std::move(__str._M_get_allocator())) |
| 554 | { |
| 555 | if (__str._M_is_local()) |
| 556 | { |
| 557 | traits_type::copy(_M_local_buf, __str._M_local_buf, |
| 558 | _S_local_capacity + 1); |
| 559 | } |
| 560 | else |
| 561 | { |
| 562 | _M_data(__str._M_data()); |
| 563 | _M_capacity(__str._M_allocated_capacity); |
| 564 | } |
| 565 | |
| 566 | // Must use _M_length() here not _M_set_length() because |
| 567 | // basic_stringbuf relies on writing into unallocated capacity so |
| 568 | // we mess up the contents if we put a '\0' in the string. |
| 569 | _M_length(__str.length()); |
| 570 | __str._M_data(__str._M_local_data()); |
| 571 | __str._M_set_length(0); |
| 572 | } |
| 573 | |
| 574 | /** |
| 575 | * @brief Construct string from an initializer %list. |
| 576 | * @param __l std::initializer_list of characters. |
| 577 | * @param __a Allocator to use (default is default allocator). |
| 578 | */ |
| 579 | basic_string(initializer_list<_CharT> __l, const _Alloc& __a = _Alloc()) |
| 580 | : _M_dataplus(_M_local_data(), __a) |
| 581 | { _M_construct(__l.begin(), __l.end()); } |
| 582 | |
| 583 | basic_string(const basic_string& __str, const _Alloc& __a) |
| 584 | : _M_dataplus(_M_local_data(), __a) |
| 585 | { _M_construct(__str.begin(), __str.end()); } |
| 586 | |
| 587 | basic_string(basic_string&& __str, const _Alloc& __a) |
| 588 | noexcept(_Alloc_traits::_S_always_equal()) |
| 589 | : _M_dataplus(_M_local_data(), __a) |
| 590 | { |
| 591 | if (__str._M_is_local()) |
| 592 | { |
| 593 | traits_type::copy(_M_local_buf, __str._M_local_buf, |
| 594 | _S_local_capacity + 1); |
| 595 | _M_length(__str.length()); |
| 596 | __str._M_set_length(0); |
| 597 | } |
| 598 | else if (_Alloc_traits::_S_always_equal() |
| 599 | || __str.get_allocator() == __a) |
| 600 | { |
| 601 | _M_data(__str._M_data()); |
| 602 | _M_length(__str.length()); |
| 603 | _M_capacity(__str._M_allocated_capacity); |
| 604 | __str._M_data(__str._M_local_buf); |
| 605 | __str._M_set_length(0); |
| 606 | } |
| 607 | else |
| 608 | _M_construct(__str.begin(), __str.end()); |
| 609 | } |
| 610 | |
| 611 | #endif // C++11 |
| 612 | |
| 613 | /** |
| 614 | * @brief Construct string as copy of a range. |
| 615 | * @param __beg Start of range. |
| 616 | * @param __end End of range. |
| 617 | * @param __a Allocator to use (default is default allocator). |
| 618 | */ |
| 619 | #if __cplusplus201703L >= 201103L |
| 620 | template<typename _InputIterator, |
| 621 | typename = std::_RequireInputIter<_InputIterator>> |
| 622 | #else |
| 623 | template<typename _InputIterator> |
| 624 | #endif |
| 625 | basic_string(_InputIterator __beg, _InputIterator __end, |
| 626 | const _Alloc& __a = _Alloc()) |
| 627 | : _M_dataplus(_M_local_data(), __a) |
| 628 | { _M_construct(__beg, __end); } |
| 629 | |
| 630 | #if __cplusplus201703L >= 201703L |
| 631 | /** |
| 632 | * @brief Construct string from a substring of a string_view. |
| 633 | * @param __t Source object convertible to string view. |
| 634 | * @param __pos The index of the first character to copy from __t. |
| 635 | * @param __n The number of characters to copy from __t. |
| 636 | * @param __a Allocator to use. |
| 637 | */ |
| 638 | template<typename _Tp, typename = _If_sv<_Tp, void>> |
| 639 | basic_string(const _Tp& __t, size_type __pos, size_type __n, |
| 640 | const _Alloc& __a = _Alloc()) |
| 641 | : basic_string(_S_to_string_view(__t).substr(__pos, __n), __a) { } |
| 642 | |
| 643 | /** |
| 644 | * @brief Construct string from a string_view. |
| 645 | * @param __t Source object convertible to string view. |
| 646 | * @param __a Allocator to use (default is default allocator). |
| 647 | */ |
| 648 | template<typename _Tp, typename = _If_sv<_Tp, void>> |
| 649 | explicit |
| 650 | basic_string(const _Tp& __t, const _Alloc& __a = _Alloc()) |
| 651 | : basic_string(__sv_wrapper(_S_to_string_view(__t)), __a) { } |
| 652 | #endif // C++17 |
| 653 | |
| 654 | /** |
| 655 | * @brief Destroy the string instance. |
| 656 | */ |
| 657 | ~basic_string() |
| 658 | { _M_dispose(); } |
| 659 | |
| 660 | /** |
| 661 | * @brief Assign the value of @a str to this string. |
| 662 | * @param __str Source string. |
| 663 | */ |
| 664 | basic_string& |
| 665 | operator=(const basic_string& __str) |
| 666 | { |
| 667 | return this->assign(__str); |
| 668 | } |
| 669 | |
| 670 | /** |
| 671 | * @brief Copy contents of @a s into this string. |
| 672 | * @param __s Source null-terminated string. |
| 673 | */ |
| 674 | basic_string& |
| 675 | operator=(const _CharT* __s) |
| 676 | { return this->assign(__s); } |
| 677 | |
| 678 | /** |
| 679 | * @brief Set value to string of length 1. |
| 680 | * @param __c Source character. |
| 681 | * |
| 682 | * Assigning to a character makes this string length 1 and |
| 683 | * (*this)[0] == @a c. |
| 684 | */ |
| 685 | basic_string& |
| 686 | operator=(_CharT __c) |
| 687 | { |
| 688 | this->assign(1, __c); |
| 689 | return *this; |
| 690 | } |
| 691 | |
| 692 | #if __cplusplus201703L >= 201103L |
| 693 | /** |
| 694 | * @brief Move assign the value of @a str to this string. |
| 695 | * @param __str Source string. |
| 696 | * |
| 697 | * The contents of @a str are moved into this string (without copying). |
| 698 | * @a str is a valid, but unspecified string. |
| 699 | **/ |
| 700 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
| 701 | // 2063. Contradictory requirements for string move assignment |
| 702 | basic_string& |
| 703 | operator=(basic_string&& __str) |
| 704 | noexcept(_Alloc_traits::_S_nothrow_move()) |
| 705 | { |
| 706 | if (!_M_is_local() && _Alloc_traits::_S_propagate_on_move_assign() |
| 707 | && !_Alloc_traits::_S_always_equal() |
| 708 | && _M_get_allocator() != __str._M_get_allocator()) |
| 709 | { |
| 710 | // Destroy existing storage before replacing allocator. |
| 711 | _M_destroy(_M_allocated_capacity); |
| 712 | _M_data(_M_local_data()); |
| 713 | _M_set_length(0); |
| 714 | } |
| 715 | // Replace allocator if POCMA is true. |
| 716 | std::__alloc_on_move(_M_get_allocator(), __str._M_get_allocator()); |
| 717 | |
| 718 | if (__str._M_is_local()) |
| 719 | { |
| 720 | // We've always got room for a short string, just copy it. |
| 721 | if (__str.size()) |
| 722 | this->_S_copy(_M_data(), __str._M_data(), __str.size()); |
| 723 | _M_set_length(__str.size()); |
| 724 | } |
| 725 | else if (_Alloc_traits::_S_propagate_on_move_assign() |
| 726 | || _Alloc_traits::_S_always_equal() |
| 727 | || _M_get_allocator() == __str._M_get_allocator()) |
| 728 | { |
| 729 | // Just move the allocated pointer, our allocator can free it. |
| 730 | pointer __data = nullptr; |
| 731 | size_type __capacity; |
| 732 | if (!_M_is_local()) |
| 733 | { |
| 734 | if (_Alloc_traits::_S_always_equal()) |
| 735 | { |
| 736 | // __str can reuse our existing storage. |
| 737 | __data = _M_data(); |
| 738 | __capacity = _M_allocated_capacity; |
| 739 | } |
| 740 | else // __str can't use it, so free it. |
| 741 | _M_destroy(_M_allocated_capacity); |
| 742 | } |
| 743 | |
| 744 | _M_data(__str._M_data()); |
| 745 | _M_length(__str.length()); |
| 746 | _M_capacity(__str._M_allocated_capacity); |
| 747 | if (__data) |
| 748 | { |
| 749 | __str._M_data(__data); |
| 750 | __str._M_capacity(__capacity); |
| 751 | } |
| 752 | else |
| 753 | __str._M_data(__str._M_local_buf); |
| 754 | } |
| 755 | else // Need to do a deep copy |
| 756 | assign(__str); |
| 757 | __str.clear(); |
| 758 | return *this; |
| 759 | } |
| 760 | |
| 761 | /** |
| 762 | * @brief Set value to string constructed from initializer %list. |
| 763 | * @param __l std::initializer_list. |
| 764 | */ |
| 765 | basic_string& |
| 766 | operator=(initializer_list<_CharT> __l) |
| 767 | { |
| 768 | this->assign(__l.begin(), __l.size()); |
| 769 | return *this; |
| 770 | } |
| 771 | #endif // C++11 |
| 772 | |
| 773 | #if __cplusplus201703L >= 201703L |
| 774 | /** |
| 775 | * @brief Set value to string constructed from a string_view. |
| 776 | * @param __svt An object convertible to string_view. |
| 777 | */ |
| 778 | template<typename _Tp> |
| 779 | _If_sv<_Tp, basic_string&> |
| 780 | operator=(const _Tp& __svt) |
| 781 | { return this->assign(__svt); } |
| 782 | |
| 783 | /** |
| 784 | * @brief Convert to a string_view. |
| 785 | * @return A string_view. |
| 786 | */ |
| 787 | operator __sv_type() const noexcept |
| 788 | { return __sv_type(data(), size()); } |
| 789 | #endif // C++17 |
| 790 | |
| 791 | // Iterators: |
| 792 | /** |
| 793 | * Returns a read/write iterator that points to the first character in |
| 794 | * the %string. |
| 795 | */ |
| 796 | iterator |
| 797 | begin() _GLIBCXX_NOEXCEPTnoexcept |
| 798 | { return iterator(_M_data()); } |
| 799 | |
| 800 | /** |
| 801 | * Returns a read-only (constant) iterator that points to the first |
| 802 | * character in the %string. |
| 803 | */ |
| 804 | const_iterator |
| 805 | begin() const _GLIBCXX_NOEXCEPTnoexcept |
| 806 | { return const_iterator(_M_data()); } |
| 807 | |
| 808 | /** |
| 809 | * Returns a read/write iterator that points one past the last |
| 810 | * character in the %string. |
| 811 | */ |
| 812 | iterator |
| 813 | end() _GLIBCXX_NOEXCEPTnoexcept |
| 814 | { return iterator(_M_data() + this->size()); } |
| 815 | |
| 816 | /** |
| 817 | * Returns a read-only (constant) iterator that points one past the |
| 818 | * last character in the %string. |
| 819 | */ |
| 820 | const_iterator |
| 821 | end() const _GLIBCXX_NOEXCEPTnoexcept |
| 822 | { return const_iterator(_M_data() + this->size()); } |
| 823 | |
| 824 | /** |
| 825 | * Returns a read/write reverse iterator that points to the last |
| 826 | * character in the %string. Iteration is done in reverse element |
| 827 | * order. |
| 828 | */ |
| 829 | reverse_iterator |
| 830 | rbegin() _GLIBCXX_NOEXCEPTnoexcept |
| 831 | { return reverse_iterator(this->end()); } |
| 832 | |
| 833 | /** |
| 834 | * Returns a read-only (constant) reverse iterator that points |
| 835 | * to the last character in the %string. Iteration is done in |
| 836 | * reverse element order. |
| 837 | */ |
| 838 | const_reverse_iterator |
| 839 | rbegin() const _GLIBCXX_NOEXCEPTnoexcept |
| 840 | { return const_reverse_iterator(this->end()); } |
| 841 | |
| 842 | /** |
| 843 | * Returns a read/write reverse iterator that points to one before the |
| 844 | * first character in the %string. Iteration is done in reverse |
| 845 | * element order. |
| 846 | */ |
| 847 | reverse_iterator |
| 848 | rend() _GLIBCXX_NOEXCEPTnoexcept |
| 849 | { return reverse_iterator(this->begin()); } |
| 850 | |
| 851 | /** |
| 852 | * Returns a read-only (constant) reverse iterator that points |
| 853 | * to one before the first character in the %string. Iteration |
| 854 | * is done in reverse element order. |
| 855 | */ |
| 856 | const_reverse_iterator |
| 857 | rend() const _GLIBCXX_NOEXCEPTnoexcept |
| 858 | { return const_reverse_iterator(this->begin()); } |
| 859 | |
| 860 | #if __cplusplus201703L >= 201103L |
| 861 | /** |
| 862 | * Returns a read-only (constant) iterator that points to the first |
| 863 | * character in the %string. |
| 864 | */ |
| 865 | const_iterator |
| 866 | cbegin() const noexcept |
| 867 | { return const_iterator(this->_M_data()); } |
| 868 | |
| 869 | /** |
| 870 | * Returns a read-only (constant) iterator that points one past the |
| 871 | * last character in the %string. |
| 872 | */ |
| 873 | const_iterator |
| 874 | cend() const noexcept |
| 875 | { return const_iterator(this->_M_data() + this->size()); } |
| 876 | |
| 877 | /** |
| 878 | * Returns a read-only (constant) reverse iterator that points |
| 879 | * to the last character in the %string. Iteration is done in |
| 880 | * reverse element order. |
| 881 | */ |
| 882 | const_reverse_iterator |
| 883 | crbegin() const noexcept |
| 884 | { return const_reverse_iterator(this->end()); } |
| 885 | |
| 886 | /** |
| 887 | * Returns a read-only (constant) reverse iterator that points |
| 888 | * to one before the first character in the %string. Iteration |
| 889 | * is done in reverse element order. |
| 890 | */ |
| 891 | const_reverse_iterator |
| 892 | crend() const noexcept |
| 893 | { return const_reverse_iterator(this->begin()); } |
| 894 | #endif |
| 895 | |
| 896 | public: |
| 897 | // Capacity: |
| 898 | /// Returns the number of characters in the string, not including any |
| 899 | /// null-termination. |
| 900 | size_type |
| 901 | size() const _GLIBCXX_NOEXCEPTnoexcept |
| 902 | { return _M_string_length; } |
| 903 | |
| 904 | /// Returns the number of characters in the string, not including any |
| 905 | /// null-termination. |
| 906 | size_type |
| 907 | length() const _GLIBCXX_NOEXCEPTnoexcept |
| 908 | { return _M_string_length; } |
| 909 | |
| 910 | /// Returns the size() of the largest possible %string. |
| 911 | size_type |
| 912 | max_size() const _GLIBCXX_NOEXCEPTnoexcept |
| 913 | { return (_Alloc_traits::max_size(_M_get_allocator()) - 1) / 2; } |
| 914 | |
| 915 | /** |
| 916 | * @brief Resizes the %string to the specified number of characters. |
| 917 | * @param __n Number of characters the %string should contain. |
| 918 | * @param __c Character to fill any new elements. |
| 919 | * |
| 920 | * This function will %resize the %string to the specified |
| 921 | * number of characters. If the number is smaller than the |
| 922 | * %string's current size the %string is truncated, otherwise |
| 923 | * the %string is extended and new elements are %set to @a __c. |
| 924 | */ |
| 925 | void |
| 926 | resize(size_type __n, _CharT __c); |
| 927 | |
| 928 | /** |
| 929 | * @brief Resizes the %string to the specified number of characters. |
| 930 | * @param __n Number of characters the %string should contain. |
| 931 | * |
| 932 | * This function will resize the %string to the specified length. If |
| 933 | * the new size is smaller than the %string's current size the %string |
| 934 | * is truncated, otherwise the %string is extended and new characters |
| 935 | * are default-constructed. For basic types such as char, this means |
| 936 | * setting them to 0. |
| 937 | */ |
| 938 | void |
| 939 | resize(size_type __n) |
| 940 | { this->resize(__n, _CharT()); } |
| 941 | |
| 942 | #if __cplusplus201703L >= 201103L |
| 943 | /// A non-binding request to reduce capacity() to size(). |
| 944 | void |
| 945 | shrink_to_fit() noexcept |
| 946 | { |
| 947 | #if __cpp_exceptions |
| 948 | if (capacity() > size()) |
| 949 | { |
| 950 | try |
| 951 | { reserve(0); } |
| 952 | catch(...) |
| 953 | { } |
| 954 | } |
| 955 | #endif |
| 956 | } |
| 957 | #endif |
| 958 | |
| 959 | /** |
| 960 | * Returns the total number of characters that the %string can hold |
| 961 | * before needing to allocate more memory. |
| 962 | */ |
| 963 | size_type |
| 964 | capacity() const _GLIBCXX_NOEXCEPTnoexcept |
| 965 | { |
| 966 | return _M_is_local() ? size_type(_S_local_capacity) |
| 967 | : _M_allocated_capacity; |
| 968 | } |
| 969 | |
| 970 | /** |
| 971 | * @brief Attempt to preallocate enough memory for specified number of |
| 972 | * characters. |
| 973 | * @param __res_arg Number of characters required. |
| 974 | * @throw std::length_error If @a __res_arg exceeds @c max_size(). |
| 975 | * |
| 976 | * This function attempts to reserve enough memory for the |
| 977 | * %string to hold the specified number of characters. If the |
| 978 | * number requested is more than max_size(), length_error is |
| 979 | * thrown. |
| 980 | * |
| 981 | * The advantage of this function is that if optimal code is a |
| 982 | * necessity and the user can determine the string length that will be |
| 983 | * required, the user can reserve the memory in %advance, and thus |
| 984 | * prevent a possible reallocation of memory and copying of %string |
| 985 | * data. |
| 986 | */ |
| 987 | void |
| 988 | reserve(size_type __res_arg = 0); |
| 989 | |
| 990 | /** |
| 991 | * Erases the string, making it empty. |
| 992 | */ |
| 993 | void |
| 994 | clear() _GLIBCXX_NOEXCEPTnoexcept |
| 995 | { _M_set_length(0); } |
| 996 | |
| 997 | /** |
| 998 | * Returns true if the %string is empty. Equivalent to |
| 999 | * <code>*this == ""</code>. |
| 1000 | */ |
| 1001 | _GLIBCXX_NODISCARD[[__nodiscard__]] bool |
| 1002 | empty() const _GLIBCXX_NOEXCEPTnoexcept |
| 1003 | { return this->size() == 0; } |
| 1004 | |
| 1005 | // Element access: |
| 1006 | /** |
| 1007 | * @brief Subscript access to the data contained in the %string. |
| 1008 | * @param __pos The index of the character to access. |
| 1009 | * @return Read-only (constant) reference to the character. |
| 1010 | * |
| 1011 | * This operator allows for easy, array-style, data access. |
| 1012 | * Note that data access with this operator is unchecked and |
| 1013 | * out_of_range lookups are not defined. (For checked lookups |
| 1014 | * see at().) |
| 1015 | */ |
| 1016 | const_reference |
| 1017 | operator[] (size_type __pos) const _GLIBCXX_NOEXCEPTnoexcept |
| 1018 | { |
| 1019 | __glibcxx_assert(__pos <= size()); |
| 1020 | return _M_data()[__pos]; |
| 1021 | } |
| 1022 | |
| 1023 | /** |
| 1024 | * @brief Subscript access to the data contained in the %string. |
| 1025 | * @param __pos The index of the character to access. |
| 1026 | * @return Read/write reference to the character. |
| 1027 | * |
| 1028 | * This operator allows for easy, array-style, data access. |
| 1029 | * Note that data access with this operator is unchecked and |
| 1030 | * out_of_range lookups are not defined. (For checked lookups |
| 1031 | * see at().) |
| 1032 | */ |
| 1033 | reference |
| 1034 | operator[](size_type __pos) |
| 1035 | { |
| 1036 | // Allow pos == size() both in C++98 mode, as v3 extension, |
| 1037 | // and in C++11 mode. |
| 1038 | __glibcxx_assert(__pos <= size()); |
| 1039 | // In pedantic mode be strict in C++98 mode. |
| 1040 | _GLIBCXX_DEBUG_PEDASSERT(__cplusplus >= 201103L || __pos < size()); |
| 1041 | return _M_data()[__pos]; |
| 1042 | } |
| 1043 | |
| 1044 | /** |
| 1045 | * @brief Provides access to the data contained in the %string. |
| 1046 | * @param __n The index of the character to access. |
| 1047 | * @return Read-only (const) reference to the character. |
| 1048 | * @throw std::out_of_range If @a n is an invalid index. |
| 1049 | * |
| 1050 | * This function provides for safer data access. The parameter is |
| 1051 | * first checked that it is in the range of the string. The function |
| 1052 | * throws out_of_range if the check fails. |
| 1053 | */ |
| 1054 | const_reference |
| 1055 | at(size_type __n) const |
| 1056 | { |
| 1057 | if (__n >= this->size()) |
| 1058 | __throw_out_of_range_fmt(__N("basic_string::at: __n "("basic_string::at: __n " "(which is %zu) >= this->size() " "(which is %zu)") |
| 1059 | "(which is %zu) >= this->size() "("basic_string::at: __n " "(which is %zu) >= this->size() " "(which is %zu)") |
| 1060 | "(which is %zu)")("basic_string::at: __n " "(which is %zu) >= this->size() " "(which is %zu)"), |
| 1061 | __n, this->size()); |
| 1062 | return _M_data()[__n]; |
| 1063 | } |
| 1064 | |
| 1065 | /** |
| 1066 | * @brief Provides access to the data contained in the %string. |
| 1067 | * @param __n The index of the character to access. |
| 1068 | * @return Read/write reference to the character. |
| 1069 | * @throw std::out_of_range If @a n is an invalid index. |
| 1070 | * |
| 1071 | * This function provides for safer data access. The parameter is |
| 1072 | * first checked that it is in the range of the string. The function |
| 1073 | * throws out_of_range if the check fails. |
| 1074 | */ |
| 1075 | reference |
| 1076 | at(size_type __n) |
| 1077 | { |
| 1078 | if (__n >= size()) |
| 1079 | __throw_out_of_range_fmt(__N("basic_string::at: __n "("basic_string::at: __n " "(which is %zu) >= this->size() " "(which is %zu)") |
| 1080 | "(which is %zu) >= this->size() "("basic_string::at: __n " "(which is %zu) >= this->size() " "(which is %zu)") |
| 1081 | "(which is %zu)")("basic_string::at: __n " "(which is %zu) >= this->size() " "(which is %zu)"), |
| 1082 | __n, this->size()); |
| 1083 | return _M_data()[__n]; |
| 1084 | } |
| 1085 | |
| 1086 | #if __cplusplus201703L >= 201103L |
| 1087 | /** |
| 1088 | * Returns a read/write reference to the data at the first |
| 1089 | * element of the %string. |
| 1090 | */ |
| 1091 | reference |
| 1092 | front() noexcept |
| 1093 | { |
| 1094 | __glibcxx_assert(!empty()); |
| 1095 | return operator[](0); |
| 1096 | } |
| 1097 | |
| 1098 | /** |
| 1099 | * Returns a read-only (constant) reference to the data at the first |
| 1100 | * element of the %string. |
| 1101 | */ |
| 1102 | const_reference |
| 1103 | front() const noexcept |
| 1104 | { |
| 1105 | __glibcxx_assert(!empty()); |
| 1106 | return operator[](0); |
| 1107 | } |
| 1108 | |
| 1109 | /** |
| 1110 | * Returns a read/write reference to the data at the last |
| 1111 | * element of the %string. |
| 1112 | */ |
| 1113 | reference |
| 1114 | back() noexcept |
| 1115 | { |
| 1116 | __glibcxx_assert(!empty()); |
| 1117 | return operator[](this->size() - 1); |
| 1118 | } |
| 1119 | |
| 1120 | /** |
| 1121 | * Returns a read-only (constant) reference to the data at the |
| 1122 | * last element of the %string. |
| 1123 | */ |
| 1124 | const_reference |
| 1125 | back() const noexcept |
| 1126 | { |
| 1127 | __glibcxx_assert(!empty()); |
| 1128 | return operator[](this->size() - 1); |
| 1129 | } |
| 1130 | #endif |
| 1131 | |
| 1132 | // Modifiers: |
| 1133 | /** |
| 1134 | * @brief Append a string to this string. |
| 1135 | * @param __str The string to append. |
| 1136 | * @return Reference to this string. |
| 1137 | */ |
| 1138 | basic_string& |
| 1139 | operator+=(const basic_string& __str) |
| 1140 | { return this->append(__str); } |
| 1141 | |
| 1142 | /** |
| 1143 | * @brief Append a C string. |
| 1144 | * @param __s The C string to append. |
| 1145 | * @return Reference to this string. |
| 1146 | */ |
| 1147 | basic_string& |
| 1148 | operator+=(const _CharT* __s) |
| 1149 | { return this->append(__s); } |
| 1150 | |
| 1151 | /** |
| 1152 | * @brief Append a character. |
| 1153 | * @param __c The character to append. |
| 1154 | * @return Reference to this string. |
| 1155 | */ |
| 1156 | basic_string& |
| 1157 | operator+=(_CharT __c) |
| 1158 | { |
| 1159 | this->push_back(__c); |
| 1160 | return *this; |
| 1161 | } |
| 1162 | |
| 1163 | #if __cplusplus201703L >= 201103L |
| 1164 | /** |
| 1165 | * @brief Append an initializer_list of characters. |
| 1166 | * @param __l The initializer_list of characters to be appended. |
| 1167 | * @return Reference to this string. |
| 1168 | */ |
| 1169 | basic_string& |
| 1170 | operator+=(initializer_list<_CharT> __l) |
| 1171 | { return this->append(__l.begin(), __l.size()); } |
| 1172 | #endif // C++11 |
| 1173 | |
| 1174 | #if __cplusplus201703L >= 201703L |
| 1175 | /** |
| 1176 | * @brief Append a string_view. |
| 1177 | * @param __svt An object convertible to string_view to be appended. |
| 1178 | * @return Reference to this string. |
| 1179 | */ |
| 1180 | template<typename _Tp> |
| 1181 | _If_sv<_Tp, basic_string&> |
| 1182 | operator+=(const _Tp& __svt) |
| 1183 | { return this->append(__svt); } |
| 1184 | #endif // C++17 |
| 1185 | |
| 1186 | /** |
| 1187 | * @brief Append a string to this string. |
| 1188 | * @param __str The string to append. |
| 1189 | * @return Reference to this string. |
| 1190 | */ |
| 1191 | basic_string& |
| 1192 | append(const basic_string& __str) |
| 1193 | { return _M_append(__str._M_data(), __str.size()); } |
| 1194 | |
| 1195 | /** |
| 1196 | * @brief Append a substring. |
| 1197 | * @param __str The string to append. |
| 1198 | * @param __pos Index of the first character of str to append. |
| 1199 | * @param __n The number of characters to append. |
| 1200 | * @return Reference to this string. |
| 1201 | * @throw std::out_of_range if @a __pos is not a valid index. |
| 1202 | * |
| 1203 | * This function appends @a __n characters from @a __str |
| 1204 | * starting at @a __pos to this string. If @a __n is is larger |
| 1205 | * than the number of available characters in @a __str, the |
| 1206 | * remainder of @a __str is appended. |
| 1207 | */ |
| 1208 | basic_string& |
| 1209 | append(const basic_string& __str, size_type __pos, size_type __n = npos) |
| 1210 | { return _M_append(__str._M_data() |
| 1211 | + __str._M_check(__pos, "basic_string::append"), |
| 1212 | __str._M_limit(__pos, __n)); } |
| 1213 | |
| 1214 | /** |
| 1215 | * @brief Append a C substring. |
| 1216 | * @param __s The C string to append. |
| 1217 | * @param __n The number of characters to append. |
| 1218 | * @return Reference to this string. |
| 1219 | */ |
| 1220 | basic_string& |
| 1221 | append(const _CharT* __s, size_type __n) |
| 1222 | { |
| 1223 | __glibcxx_requires_string_len(__s, __n); |
| 1224 | _M_check_length(size_type(0), __n, "basic_string::append"); |
| 1225 | return _M_append(__s, __n); |
| 1226 | } |
| 1227 | |
| 1228 | /** |
| 1229 | * @brief Append a C string. |
| 1230 | * @param __s The C string to append. |
| 1231 | * @return Reference to this string. |
| 1232 | */ |
| 1233 | basic_string& |
| 1234 | append(const _CharT* __s) |
| 1235 | { |
| 1236 | __glibcxx_requires_string(__s); |
| 1237 | const size_type __n = traits_type::length(__s); |
| 1238 | _M_check_length(size_type(0), __n, "basic_string::append"); |
| 1239 | return _M_append(__s, __n); |
| 1240 | } |
| 1241 | |
| 1242 | /** |
| 1243 | * @brief Append multiple characters. |
| 1244 | * @param __n The number of characters to append. |
| 1245 | * @param __c The character to use. |
| 1246 | * @return Reference to this string. |
| 1247 | * |
| 1248 | * Appends __n copies of __c to this string. |
| 1249 | */ |
| 1250 | basic_string& |
| 1251 | append(size_type __n, _CharT __c) |
| 1252 | { return _M_replace_aux(this->size(), size_type(0), __n, __c); } |
| 1253 | |
| 1254 | #if __cplusplus201703L >= 201103L |
| 1255 | /** |
| 1256 | * @brief Append an initializer_list of characters. |
| 1257 | * @param __l The initializer_list of characters to append. |
| 1258 | * @return Reference to this string. |
| 1259 | */ |
| 1260 | basic_string& |
| 1261 | append(initializer_list<_CharT> __l) |
| 1262 | { return this->append(__l.begin(), __l.size()); } |
| 1263 | #endif // C++11 |
| 1264 | |
| 1265 | /** |
| 1266 | * @brief Append a range of characters. |
| 1267 | * @param __first Iterator referencing the first character to append. |
| 1268 | * @param __last Iterator marking the end of the range. |
| 1269 | * @return Reference to this string. |
| 1270 | * |
| 1271 | * Appends characters in the range [__first,__last) to this string. |
| 1272 | */ |
| 1273 | #if __cplusplus201703L >= 201103L |
| 1274 | template<class _InputIterator, |
| 1275 | typename = std::_RequireInputIter<_InputIterator>> |
| 1276 | #else |
| 1277 | template<class _InputIterator> |
| 1278 | #endif |
| 1279 | basic_string& |
| 1280 | append(_InputIterator __first, _InputIterator __last) |
| 1281 | { return this->replace(end(), end(), __first, __last); } |
| 1282 | |
| 1283 | #if __cplusplus201703L >= 201703L |
| 1284 | /** |
| 1285 | * @brief Append a string_view. |
| 1286 | * @param __svt An object convertible to string_view to be appended. |
| 1287 | * @return Reference to this string. |
| 1288 | */ |
| 1289 | template<typename _Tp> |
| 1290 | _If_sv<_Tp, basic_string&> |
| 1291 | append(const _Tp& __svt) |
| 1292 | { |
| 1293 | __sv_type __sv = __svt; |
| 1294 | return this->append(__sv.data(), __sv.size()); |
| 1295 | } |
| 1296 | |
| 1297 | /** |
| 1298 | * @brief Append a range of characters from a string_view. |
| 1299 | * @param __svt An object convertible to string_view to be appended from. |
| 1300 | * @param __pos The position in the string_view to append from. |
| 1301 | * @param __n The number of characters to append from the string_view. |
| 1302 | * @return Reference to this string. |
| 1303 | */ |
| 1304 | template<typename _Tp> |
| 1305 | _If_sv<_Tp, basic_string&> |
| 1306 | append(const _Tp& __svt, size_type __pos, size_type __n = npos) |
| 1307 | { |
| 1308 | __sv_type __sv = __svt; |
| 1309 | return _M_append(__sv.data() |
| 1310 | + std::__sv_check(__sv.size(), __pos, "basic_string::append"), |
| 1311 | std::__sv_limit(__sv.size(), __pos, __n)); |
| 1312 | } |
| 1313 | #endif // C++17 |
| 1314 | |
| 1315 | /** |
| 1316 | * @brief Append a single character. |
| 1317 | * @param __c Character to append. |
| 1318 | */ |
| 1319 | void |
| 1320 | push_back(_CharT __c) |
| 1321 | { |
| 1322 | const size_type __size = this->size(); |
| 1323 | if (__size + 1 > this->capacity()) |
| 1324 | this->_M_mutate(__size, size_type(0), 0, size_type(1)); |
| 1325 | traits_type::assign(this->_M_data()[__size], __c); |
| 1326 | this->_M_set_length(__size + 1); |
| 1327 | } |
| 1328 | |
| 1329 | /** |
| 1330 | * @brief Set value to contents of another string. |
| 1331 | * @param __str Source string to use. |
| 1332 | * @return Reference to this string. |
| 1333 | */ |
| 1334 | basic_string& |
| 1335 | assign(const basic_string& __str) |
| 1336 | { |
| 1337 | #if __cplusplus201703L >= 201103L |
| 1338 | if (_Alloc_traits::_S_propagate_on_copy_assign()) |
| 1339 | { |
| 1340 | if (!_Alloc_traits::_S_always_equal() && !_M_is_local() |
| 1341 | && _M_get_allocator() != __str._M_get_allocator()) |
| 1342 | { |
| 1343 | // Propagating allocator cannot free existing storage so must |
| 1344 | // deallocate it before replacing current allocator. |
| 1345 | if (__str.size() <= _S_local_capacity) |
| 1346 | { |
| 1347 | _M_destroy(_M_allocated_capacity); |
| 1348 | _M_data(_M_local_data()); |
| 1349 | _M_set_length(0); |
| 1350 | } |
| 1351 | else |
| 1352 | { |
| 1353 | const auto __len = __str.size(); |
| 1354 | auto __alloc = __str._M_get_allocator(); |
| 1355 | // If this allocation throws there are no effects: |
| 1356 | auto __ptr = _Alloc_traits::allocate(__alloc, __len + 1); |
| 1357 | _M_destroy(_M_allocated_capacity); |
| 1358 | _M_data(__ptr); |
| 1359 | _M_capacity(__len); |
| 1360 | _M_set_length(__len); |
| 1361 | } |
| 1362 | } |
| 1363 | std::__alloc_on_copy(_M_get_allocator(), __str._M_get_allocator()); |
| 1364 | } |
| 1365 | #endif |
| 1366 | this->_M_assign(__str); |
| 1367 | return *this; |
| 1368 | } |
| 1369 | |
| 1370 | #if __cplusplus201703L >= 201103L |
| 1371 | /** |
| 1372 | * @brief Set value to contents of another string. |
| 1373 | * @param __str Source string to use. |
| 1374 | * @return Reference to this string. |
| 1375 | * |
| 1376 | * This function sets this string to the exact contents of @a __str. |
| 1377 | * @a __str is a valid, but unspecified string. |
| 1378 | */ |
| 1379 | basic_string& |
| 1380 | assign(basic_string&& __str) |
| 1381 | noexcept(_Alloc_traits::_S_nothrow_move()) |
| 1382 | { |
| 1383 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
| 1384 | // 2063. Contradictory requirements for string move assignment |
| 1385 | return *this = std::move(__str); |
| 1386 | } |
| 1387 | #endif // C++11 |
| 1388 | |
| 1389 | /** |
| 1390 | * @brief Set value to a substring of a string. |
| 1391 | * @param __str The string to use. |
| 1392 | * @param __pos Index of the first character of str. |
| 1393 | * @param __n Number of characters to use. |
| 1394 | * @return Reference to this string. |
| 1395 | * @throw std::out_of_range if @a pos is not a valid index. |
| 1396 | * |
| 1397 | * This function sets this string to the substring of @a __str |
| 1398 | * consisting of @a __n characters at @a __pos. If @a __n is |
| 1399 | * is larger than the number of available characters in @a |
| 1400 | * __str, the remainder of @a __str is used. |
| 1401 | */ |
| 1402 | basic_string& |
| 1403 | assign(const basic_string& __str, size_type __pos, size_type __n = npos) |
| 1404 | { return _M_replace(size_type(0), this->size(), __str._M_data() |
| 1405 | + __str._M_check(__pos, "basic_string::assign"), |
| 1406 | __str._M_limit(__pos, __n)); } |
| 1407 | |
| 1408 | /** |
| 1409 | * @brief Set value to a C substring. |
| 1410 | * @param __s The C string to use. |
| 1411 | * @param __n Number of characters to use. |
| 1412 | * @return Reference to this string. |
| 1413 | * |
| 1414 | * This function sets the value of this string to the first @a __n |
| 1415 | * characters of @a __s. If @a __n is is larger than the number of |
| 1416 | * available characters in @a __s, the remainder of @a __s is used. |
| 1417 | */ |
| 1418 | basic_string& |
| 1419 | assign(const _CharT* __s, size_type __n) |
| 1420 | { |
| 1421 | __glibcxx_requires_string_len(__s, __n); |
| 1422 | return _M_replace(size_type(0), this->size(), __s, __n); |
| 1423 | } |
| 1424 | |
| 1425 | /** |
| 1426 | * @brief Set value to contents of a C string. |
| 1427 | * @param __s The C string to use. |
| 1428 | * @return Reference to this string. |
| 1429 | * |
| 1430 | * This function sets the value of this string to the value of @a __s. |
| 1431 | * The data is copied, so there is no dependence on @a __s once the |
| 1432 | * function returns. |
| 1433 | */ |
| 1434 | basic_string& |
| 1435 | assign(const _CharT* __s) |
| 1436 | { |
| 1437 | __glibcxx_requires_string(__s); |
| 1438 | return _M_replace(size_type(0), this->size(), __s, |
| 1439 | traits_type::length(__s)); |
| 1440 | } |
| 1441 | |
| 1442 | /** |
| 1443 | * @brief Set value to multiple characters. |
| 1444 | * @param __n Length of the resulting string. |
| 1445 | * @param __c The character to use. |
| 1446 | * @return Reference to this string. |
| 1447 | * |
| 1448 | * This function sets the value of this string to @a __n copies of |
| 1449 | * character @a __c. |
| 1450 | */ |
| 1451 | basic_string& |
| 1452 | assign(size_type __n, _CharT __c) |
| 1453 | { return _M_replace_aux(size_type(0), this->size(), __n, __c); } |
| 1454 | |
| 1455 | /** |
| 1456 | * @brief Set value to a range of characters. |
| 1457 | * @param __first Iterator referencing the first character to append. |
| 1458 | * @param __last Iterator marking the end of the range. |
| 1459 | * @return Reference to this string. |
| 1460 | * |
| 1461 | * Sets value of string to characters in the range [__first,__last). |
| 1462 | */ |
| 1463 | #if __cplusplus201703L >= 201103L |
| 1464 | template<class _InputIterator, |
| 1465 | typename = std::_RequireInputIter<_InputIterator>> |
| 1466 | #else |
| 1467 | template<class _InputIterator> |
| 1468 | #endif |
| 1469 | basic_string& |
| 1470 | assign(_InputIterator __first, _InputIterator __last) |
| 1471 | { return this->replace(begin(), end(), __first, __last); } |
| 1472 | |
| 1473 | #if __cplusplus201703L >= 201103L |
| 1474 | /** |
| 1475 | * @brief Set value to an initializer_list of characters. |
| 1476 | * @param __l The initializer_list of characters to assign. |
| 1477 | * @return Reference to this string. |
| 1478 | */ |
| 1479 | basic_string& |
| 1480 | assign(initializer_list<_CharT> __l) |
| 1481 | { return this->assign(__l.begin(), __l.size()); } |
| 1482 | #endif // C++11 |
| 1483 | |
| 1484 | #if __cplusplus201703L >= 201703L |
| 1485 | /** |
| 1486 | * @brief Set value from a string_view. |
| 1487 | * @param __svt The source object convertible to string_view. |
| 1488 | * @return Reference to this string. |
| 1489 | */ |
| 1490 | template<typename _Tp> |
| 1491 | _If_sv<_Tp, basic_string&> |
| 1492 | assign(const _Tp& __svt) |
| 1493 | { |
| 1494 | __sv_type __sv = __svt; |
| 1495 | return this->assign(__sv.data(), __sv.size()); |
| 1496 | } |
| 1497 | |
| 1498 | /** |
| 1499 | * @brief Set value from a range of characters in a string_view. |
| 1500 | * @param __svt The source object convertible to string_view. |
| 1501 | * @param __pos The position in the string_view to assign from. |
| 1502 | * @param __n The number of characters to assign. |
| 1503 | * @return Reference to this string. |
| 1504 | */ |
| 1505 | template<typename _Tp> |
| 1506 | _If_sv<_Tp, basic_string&> |
| 1507 | assign(const _Tp& __svt, size_type __pos, size_type __n = npos) |
| 1508 | { |
| 1509 | __sv_type __sv = __svt; |
| 1510 | return _M_replace(size_type(0), this->size(), |
| 1511 | __sv.data() |
| 1512 | + std::__sv_check(__sv.size(), __pos, "basic_string::assign"), |
| 1513 | std::__sv_limit(__sv.size(), __pos, __n)); |
| 1514 | } |
| 1515 | #endif // C++17 |
| 1516 | |
| 1517 | #if __cplusplus201703L >= 201103L |
| 1518 | /** |
| 1519 | * @brief Insert multiple characters. |
| 1520 | * @param __p Const_iterator referencing location in string to |
| 1521 | * insert at. |
| 1522 | * @param __n Number of characters to insert |
| 1523 | * @param __c The character to insert. |
| 1524 | * @return Iterator referencing the first inserted char. |
| 1525 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1526 | * |
| 1527 | * Inserts @a __n copies of character @a __c starting at the |
| 1528 | * position referenced by iterator @a __p. If adding |
| 1529 | * characters causes the length to exceed max_size(), |
| 1530 | * length_error is thrown. The value of the string doesn't |
| 1531 | * change if an error is thrown. |
| 1532 | */ |
| 1533 | iterator |
| 1534 | insert(const_iterator __p, size_type __n, _CharT __c) |
| 1535 | { |
| 1536 | _GLIBCXX_DEBUG_PEDASSERT(__p >= begin() && __p <= end()); |
| 1537 | const size_type __pos = __p - begin(); |
| 1538 | this->replace(__p, __p, __n, __c); |
| 1539 | return iterator(this->_M_data() + __pos); |
| 1540 | } |
| 1541 | #else |
| 1542 | /** |
| 1543 | * @brief Insert multiple characters. |
| 1544 | * @param __p Iterator referencing location in string to insert at. |
| 1545 | * @param __n Number of characters to insert |
| 1546 | * @param __c The character to insert. |
| 1547 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1548 | * |
| 1549 | * Inserts @a __n copies of character @a __c starting at the |
| 1550 | * position referenced by iterator @a __p. If adding |
| 1551 | * characters causes the length to exceed max_size(), |
| 1552 | * length_error is thrown. The value of the string doesn't |
| 1553 | * change if an error is thrown. |
| 1554 | */ |
| 1555 | void |
| 1556 | insert(iterator __p, size_type __n, _CharT __c) |
| 1557 | { this->replace(__p, __p, __n, __c); } |
| 1558 | #endif |
| 1559 | |
| 1560 | #if __cplusplus201703L >= 201103L |
| 1561 | /** |
| 1562 | * @brief Insert a range of characters. |
| 1563 | * @param __p Const_iterator referencing location in string to |
| 1564 | * insert at. |
| 1565 | * @param __beg Start of range. |
| 1566 | * @param __end End of range. |
| 1567 | * @return Iterator referencing the first inserted char. |
| 1568 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1569 | * |
| 1570 | * Inserts characters in range [beg,end). If adding characters |
| 1571 | * causes the length to exceed max_size(), length_error is |
| 1572 | * thrown. The value of the string doesn't change if an error |
| 1573 | * is thrown. |
| 1574 | */ |
| 1575 | template<class _InputIterator, |
| 1576 | typename = std::_RequireInputIter<_InputIterator>> |
| 1577 | iterator |
| 1578 | insert(const_iterator __p, _InputIterator __beg, _InputIterator __end) |
| 1579 | { |
| 1580 | _GLIBCXX_DEBUG_PEDASSERT(__p >= begin() && __p <= end()); |
| 1581 | const size_type __pos = __p - begin(); |
| 1582 | this->replace(__p, __p, __beg, __end); |
| 1583 | return iterator(this->_M_data() + __pos); |
| 1584 | } |
| 1585 | #else |
| 1586 | /** |
| 1587 | * @brief Insert a range of characters. |
| 1588 | * @param __p Iterator referencing location in string to insert at. |
| 1589 | * @param __beg Start of range. |
| 1590 | * @param __end End of range. |
| 1591 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1592 | * |
| 1593 | * Inserts characters in range [__beg,__end). If adding |
| 1594 | * characters causes the length to exceed max_size(), |
| 1595 | * length_error is thrown. The value of the string doesn't |
| 1596 | * change if an error is thrown. |
| 1597 | */ |
| 1598 | template<class _InputIterator> |
| 1599 | void |
| 1600 | insert(iterator __p, _InputIterator __beg, _InputIterator __end) |
| 1601 | { this->replace(__p, __p, __beg, __end); } |
| 1602 | #endif |
| 1603 | |
| 1604 | #if __cplusplus201703L >= 201103L |
| 1605 | /** |
| 1606 | * @brief Insert an initializer_list of characters. |
| 1607 | * @param __p Iterator referencing location in string to insert at. |
| 1608 | * @param __l The initializer_list of characters to insert. |
| 1609 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1610 | */ |
| 1611 | iterator |
| 1612 | insert(const_iterator __p, initializer_list<_CharT> __l) |
| 1613 | { return this->insert(__p, __l.begin(), __l.end()); } |
| 1614 | |
| 1615 | #ifdef _GLIBCXX_DEFINING_STRING_INSTANTIATIONS |
| 1616 | // See PR libstdc++/83328 |
| 1617 | void |
| 1618 | insert(iterator __p, initializer_list<_CharT> __l) |
| 1619 | { |
| 1620 | _GLIBCXX_DEBUG_PEDASSERT(__p >= begin() && __p <= end()); |
| 1621 | this->insert(__p - begin(), __l.begin(), __l.size()); |
| 1622 | } |
| 1623 | #endif |
| 1624 | #endif // C++11 |
| 1625 | |
| 1626 | /** |
| 1627 | * @brief Insert value of a string. |
| 1628 | * @param __pos1 Position in string to insert at. |
| 1629 | * @param __str The string to insert. |
| 1630 | * @return Reference to this string. |
| 1631 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1632 | * |
| 1633 | * Inserts value of @a __str starting at @a __pos1. If adding |
| 1634 | * characters causes the length to exceed max_size(), |
| 1635 | * length_error is thrown. The value of the string doesn't |
| 1636 | * change if an error is thrown. |
| 1637 | */ |
| 1638 | basic_string& |
| 1639 | insert(size_type __pos1, const basic_string& __str) |
| 1640 | { return this->replace(__pos1, size_type(0), |
| 1641 | __str._M_data(), __str.size()); } |
| 1642 | |
| 1643 | /** |
| 1644 | * @brief Insert a substring. |
| 1645 | * @param __pos1 Position in string to insert at. |
| 1646 | * @param __str The string to insert. |
| 1647 | * @param __pos2 Start of characters in str to insert. |
| 1648 | * @param __n Number of characters to insert. |
| 1649 | * @return Reference to this string. |
| 1650 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1651 | * @throw std::out_of_range If @a pos1 > size() or |
| 1652 | * @a __pos2 > @a str.size(). |
| 1653 | * |
| 1654 | * Starting at @a pos1, insert @a __n character of @a __str |
| 1655 | * beginning with @a __pos2. If adding characters causes the |
| 1656 | * length to exceed max_size(), length_error is thrown. If @a |
| 1657 | * __pos1 is beyond the end of this string or @a __pos2 is |
| 1658 | * beyond the end of @a __str, out_of_range is thrown. The |
| 1659 | * value of the string doesn't change if an error is thrown. |
| 1660 | */ |
| 1661 | basic_string& |
| 1662 | insert(size_type __pos1, const basic_string& __str, |
| 1663 | size_type __pos2, size_type __n = npos) |
| 1664 | { return this->replace(__pos1, size_type(0), __str._M_data() |
| 1665 | + __str._M_check(__pos2, "basic_string::insert"), |
| 1666 | __str._M_limit(__pos2, __n)); } |
| 1667 | |
| 1668 | /** |
| 1669 | * @brief Insert a C substring. |
| 1670 | * @param __pos Position in string to insert at. |
| 1671 | * @param __s The C string to insert. |
| 1672 | * @param __n The number of characters to insert. |
| 1673 | * @return Reference to this string. |
| 1674 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1675 | * @throw std::out_of_range If @a __pos is beyond the end of this |
| 1676 | * string. |
| 1677 | * |
| 1678 | * Inserts the first @a __n characters of @a __s starting at @a |
| 1679 | * __pos. If adding characters causes the length to exceed |
| 1680 | * max_size(), length_error is thrown. If @a __pos is beyond |
| 1681 | * end(), out_of_range is thrown. The value of the string |
| 1682 | * doesn't change if an error is thrown. |
| 1683 | */ |
| 1684 | basic_string& |
| 1685 | insert(size_type __pos, const _CharT* __s, size_type __n) |
| 1686 | { return this->replace(__pos, size_type(0), __s, __n); } |
| 1687 | |
| 1688 | /** |
| 1689 | * @brief Insert a C string. |
| 1690 | * @param __pos Position in string to insert at. |
| 1691 | * @param __s The C string to insert. |
| 1692 | * @return Reference to this string. |
| 1693 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1694 | * @throw std::out_of_range If @a pos is beyond the end of this |
| 1695 | * string. |
| 1696 | * |
| 1697 | * Inserts the first @a n characters of @a __s starting at @a __pos. If |
| 1698 | * adding characters causes the length to exceed max_size(), |
| 1699 | * length_error is thrown. If @a __pos is beyond end(), out_of_range is |
| 1700 | * thrown. The value of the string doesn't change if an error is |
| 1701 | * thrown. |
| 1702 | */ |
| 1703 | basic_string& |
| 1704 | insert(size_type __pos, const _CharT* __s) |
| 1705 | { |
| 1706 | __glibcxx_requires_string(__s); |
| 1707 | return this->replace(__pos, size_type(0), __s, |
| 1708 | traits_type::length(__s)); |
| 1709 | } |
| 1710 | |
| 1711 | /** |
| 1712 | * @brief Insert multiple characters. |
| 1713 | * @param __pos Index in string to insert at. |
| 1714 | * @param __n Number of characters to insert |
| 1715 | * @param __c The character to insert. |
| 1716 | * @return Reference to this string. |
| 1717 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1718 | * @throw std::out_of_range If @a __pos is beyond the end of this |
| 1719 | * string. |
| 1720 | * |
| 1721 | * Inserts @a __n copies of character @a __c starting at index |
| 1722 | * @a __pos. If adding characters causes the length to exceed |
| 1723 | * max_size(), length_error is thrown. If @a __pos > length(), |
| 1724 | * out_of_range is thrown. The value of the string doesn't |
| 1725 | * change if an error is thrown. |
| 1726 | */ |
| 1727 | basic_string& |
| 1728 | insert(size_type __pos, size_type __n, _CharT __c) |
| 1729 | { return _M_replace_aux(_M_check(__pos, "basic_string::insert"), |
| 1730 | size_type(0), __n, __c); } |
| 1731 | |
| 1732 | /** |
| 1733 | * @brief Insert one character. |
| 1734 | * @param __p Iterator referencing position in string to insert at. |
| 1735 | * @param __c The character to insert. |
| 1736 | * @return Iterator referencing newly inserted char. |
| 1737 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1738 | * |
| 1739 | * Inserts character @a __c at position referenced by @a __p. |
| 1740 | * If adding character causes the length to exceed max_size(), |
| 1741 | * length_error is thrown. If @a __p is beyond end of string, |
| 1742 | * out_of_range is thrown. The value of the string doesn't |
| 1743 | * change if an error is thrown. |
| 1744 | */ |
| 1745 | iterator |
| 1746 | insert(__const_iterator __p, _CharT __c) |
| 1747 | { |
| 1748 | _GLIBCXX_DEBUG_PEDASSERT(__p >= begin() && __p <= end()); |
| 1749 | const size_type __pos = __p - begin(); |
| 1750 | _M_replace_aux(__pos, size_type(0), size_type(1), __c); |
| 1751 | return iterator(_M_data() + __pos); |
| 1752 | } |
| 1753 | |
| 1754 | #if __cplusplus201703L >= 201703L |
| 1755 | /** |
| 1756 | * @brief Insert a string_view. |
| 1757 | * @param __pos Position in string to insert at. |
| 1758 | * @param __svt The object convertible to string_view to insert. |
| 1759 | * @return Reference to this string. |
| 1760 | */ |
| 1761 | template<typename _Tp> |
| 1762 | _If_sv<_Tp, basic_string&> |
| 1763 | insert(size_type __pos, const _Tp& __svt) |
| 1764 | { |
| 1765 | __sv_type __sv = __svt; |
| 1766 | return this->insert(__pos, __sv.data(), __sv.size()); |
| 1767 | } |
| 1768 | |
| 1769 | /** |
| 1770 | * @brief Insert a string_view. |
| 1771 | * @param __pos1 Position in string to insert at. |
| 1772 | * @param __svt The object convertible to string_view to insert from. |
| 1773 | * @param __pos2 Start of characters in str to insert. |
| 1774 | * @param __n The number of characters to insert. |
| 1775 | * @return Reference to this string. |
| 1776 | */ |
| 1777 | template<typename _Tp> |
| 1778 | _If_sv<_Tp, basic_string&> |
| 1779 | insert(size_type __pos1, const _Tp& __svt, |
| 1780 | size_type __pos2, size_type __n = npos) |
| 1781 | { |
| 1782 | __sv_type __sv = __svt; |
| 1783 | return this->replace(__pos1, size_type(0), |
| 1784 | __sv.data() |
| 1785 | + std::__sv_check(__sv.size(), __pos2, "basic_string::insert"), |
| 1786 | std::__sv_limit(__sv.size(), __pos2, __n)); |
| 1787 | } |
| 1788 | #endif // C++17 |
| 1789 | |
| 1790 | /** |
| 1791 | * @brief Remove characters. |
| 1792 | * @param __pos Index of first character to remove (default 0). |
| 1793 | * @param __n Number of characters to remove (default remainder). |
| 1794 | * @return Reference to this string. |
| 1795 | * @throw std::out_of_range If @a pos is beyond the end of this |
| 1796 | * string. |
| 1797 | * |
| 1798 | * Removes @a __n characters from this string starting at @a |
| 1799 | * __pos. The length of the string is reduced by @a __n. If |
| 1800 | * there are < @a __n characters to remove, the remainder of |
| 1801 | * the string is truncated. If @a __p is beyond end of string, |
| 1802 | * out_of_range is thrown. The value of the string doesn't |
| 1803 | * change if an error is thrown. |
| 1804 | */ |
| 1805 | basic_string& |
| 1806 | erase(size_type __pos = 0, size_type __n = npos) |
| 1807 | { |
| 1808 | _M_check(__pos, "basic_string::erase"); |
| 1809 | if (__n == npos) |
| 1810 | this->_M_set_length(__pos); |
| 1811 | else if (__n != 0) |
| 1812 | this->_M_erase(__pos, _M_limit(__pos, __n)); |
| 1813 | return *this; |
| 1814 | } |
| 1815 | |
| 1816 | /** |
| 1817 | * @brief Remove one character. |
| 1818 | * @param __position Iterator referencing the character to remove. |
| 1819 | * @return iterator referencing same location after removal. |
| 1820 | * |
| 1821 | * Removes the character at @a __position from this string. The value |
| 1822 | * of the string doesn't change if an error is thrown. |
| 1823 | */ |
| 1824 | iterator |
| 1825 | erase(__const_iterator __position) |
| 1826 | { |
| 1827 | _GLIBCXX_DEBUG_PEDASSERT(__position >= begin() |
| 1828 | && __position < end()); |
| 1829 | const size_type __pos = __position - begin(); |
| 1830 | this->_M_erase(__pos, size_type(1)); |
| 1831 | return iterator(_M_data() + __pos); |
| 1832 | } |
| 1833 | |
| 1834 | /** |
| 1835 | * @brief Remove a range of characters. |
| 1836 | * @param __first Iterator referencing the first character to remove. |
| 1837 | * @param __last Iterator referencing the end of the range. |
| 1838 | * @return Iterator referencing location of first after removal. |
| 1839 | * |
| 1840 | * Removes the characters in the range [first,last) from this string. |
| 1841 | * The value of the string doesn't change if an error is thrown. |
| 1842 | */ |
| 1843 | iterator |
| 1844 | erase(__const_iterator __first, __const_iterator __last) |
| 1845 | { |
| 1846 | _GLIBCXX_DEBUG_PEDASSERT(__first >= begin() && __first <= __last |
| 1847 | && __last <= end()); |
| 1848 | const size_type __pos = __first - begin(); |
| 1849 | if (__last == end()) |
| 1850 | this->_M_set_length(__pos); |
| 1851 | else |
| 1852 | this->_M_erase(__pos, __last - __first); |
| 1853 | return iterator(this->_M_data() + __pos); |
| 1854 | } |
| 1855 | |
| 1856 | #if __cplusplus201703L >= 201103L |
| 1857 | /** |
| 1858 | * @brief Remove the last character. |
| 1859 | * |
| 1860 | * The string must be non-empty. |
| 1861 | */ |
| 1862 | void |
| 1863 | pop_back() noexcept |
| 1864 | { |
| 1865 | __glibcxx_assert(!empty()); |
| 1866 | _M_erase(size() - 1, 1); |
| 1867 | } |
| 1868 | #endif // C++11 |
| 1869 | |
| 1870 | /** |
| 1871 | * @brief Replace characters with value from another string. |
| 1872 | * @param __pos Index of first character to replace. |
| 1873 | * @param __n Number of characters to be replaced. |
| 1874 | * @param __str String to insert. |
| 1875 | * @return Reference to this string. |
| 1876 | * @throw std::out_of_range If @a pos is beyond the end of this |
| 1877 | * string. |
| 1878 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1879 | * |
| 1880 | * Removes the characters in the range [__pos,__pos+__n) from |
| 1881 | * this string. In place, the value of @a __str is inserted. |
| 1882 | * If @a __pos is beyond end of string, out_of_range is thrown. |
| 1883 | * If the length of the result exceeds max_size(), length_error |
| 1884 | * is thrown. The value of the string doesn't change if an |
| 1885 | * error is thrown. |
| 1886 | */ |
| 1887 | basic_string& |
| 1888 | replace(size_type __pos, size_type __n, const basic_string& __str) |
| 1889 | { return this->replace(__pos, __n, __str._M_data(), __str.size()); } |
| 1890 | |
| 1891 | /** |
| 1892 | * @brief Replace characters with value from another string. |
| 1893 | * @param __pos1 Index of first character to replace. |
| 1894 | * @param __n1 Number of characters to be replaced. |
| 1895 | * @param __str String to insert. |
| 1896 | * @param __pos2 Index of first character of str to use. |
| 1897 | * @param __n2 Number of characters from str to use. |
| 1898 | * @return Reference to this string. |
| 1899 | * @throw std::out_of_range If @a __pos1 > size() or @a __pos2 > |
| 1900 | * __str.size(). |
| 1901 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1902 | * |
| 1903 | * Removes the characters in the range [__pos1,__pos1 + n) from this |
| 1904 | * string. In place, the value of @a __str is inserted. If @a __pos is |
| 1905 | * beyond end of string, out_of_range is thrown. If the length of the |
| 1906 | * result exceeds max_size(), length_error is thrown. The value of the |
| 1907 | * string doesn't change if an error is thrown. |
| 1908 | */ |
| 1909 | basic_string& |
| 1910 | replace(size_type __pos1, size_type __n1, const basic_string& __str, |
| 1911 | size_type __pos2, size_type __n2 = npos) |
| 1912 | { return this->replace(__pos1, __n1, __str._M_data() |
| 1913 | + __str._M_check(__pos2, "basic_string::replace"), |
| 1914 | __str._M_limit(__pos2, __n2)); } |
| 1915 | |
| 1916 | /** |
| 1917 | * @brief Replace characters with value of a C substring. |
| 1918 | * @param __pos Index of first character to replace. |
| 1919 | * @param __n1 Number of characters to be replaced. |
| 1920 | * @param __s C string to insert. |
| 1921 | * @param __n2 Number of characters from @a s to use. |
| 1922 | * @return Reference to this string. |
| 1923 | * @throw std::out_of_range If @a pos1 > size(). |
| 1924 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1925 | * |
| 1926 | * Removes the characters in the range [__pos,__pos + __n1) |
| 1927 | * from this string. In place, the first @a __n2 characters of |
| 1928 | * @a __s are inserted, or all of @a __s if @a __n2 is too large. If |
| 1929 | * @a __pos is beyond end of string, out_of_range is thrown. If |
| 1930 | * the length of result exceeds max_size(), length_error is |
| 1931 | * thrown. The value of the string doesn't change if an error |
| 1932 | * is thrown. |
| 1933 | */ |
| 1934 | basic_string& |
| 1935 | replace(size_type __pos, size_type __n1, const _CharT* __s, |
| 1936 | size_type __n2) |
| 1937 | { |
| 1938 | __glibcxx_requires_string_len(__s, __n2); |
| 1939 | return _M_replace(_M_check(__pos, "basic_string::replace"), |
| 1940 | _M_limit(__pos, __n1), __s, __n2); |
| 1941 | } |
| 1942 | |
| 1943 | /** |
| 1944 | * @brief Replace characters with value of a C string. |
| 1945 | * @param __pos Index of first character to replace. |
| 1946 | * @param __n1 Number of characters to be replaced. |
| 1947 | * @param __s C string to insert. |
| 1948 | * @return Reference to this string. |
| 1949 | * @throw std::out_of_range If @a pos > size(). |
| 1950 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1951 | * |
| 1952 | * Removes the characters in the range [__pos,__pos + __n1) |
| 1953 | * from this string. In place, the characters of @a __s are |
| 1954 | * inserted. If @a __pos is beyond end of string, out_of_range |
| 1955 | * is thrown. If the length of result exceeds max_size(), |
| 1956 | * length_error is thrown. The value of the string doesn't |
| 1957 | * change if an error is thrown. |
| 1958 | */ |
| 1959 | basic_string& |
| 1960 | replace(size_type __pos, size_type __n1, const _CharT* __s) |
| 1961 | { |
| 1962 | __glibcxx_requires_string(__s); |
| 1963 | return this->replace(__pos, __n1, __s, traits_type::length(__s)); |
| 1964 | } |
| 1965 | |
| 1966 | /** |
| 1967 | * @brief Replace characters with multiple characters. |
| 1968 | * @param __pos Index of first character to replace. |
| 1969 | * @param __n1 Number of characters to be replaced. |
| 1970 | * @param __n2 Number of characters to insert. |
| 1971 | * @param __c Character to insert. |
| 1972 | * @return Reference to this string. |
| 1973 | * @throw std::out_of_range If @a __pos > size(). |
| 1974 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1975 | * |
| 1976 | * Removes the characters in the range [pos,pos + n1) from this |
| 1977 | * string. In place, @a __n2 copies of @a __c are inserted. |
| 1978 | * If @a __pos is beyond end of string, out_of_range is thrown. |
| 1979 | * If the length of result exceeds max_size(), length_error is |
| 1980 | * thrown. The value of the string doesn't change if an error |
| 1981 | * is thrown. |
| 1982 | */ |
| 1983 | basic_string& |
| 1984 | replace(size_type __pos, size_type __n1, size_type __n2, _CharT __c) |
| 1985 | { return _M_replace_aux(_M_check(__pos, "basic_string::replace"), |
| 1986 | _M_limit(__pos, __n1), __n2, __c); } |
| 1987 | |
| 1988 | /** |
| 1989 | * @brief Replace range of characters with string. |
| 1990 | * @param __i1 Iterator referencing start of range to replace. |
| 1991 | * @param __i2 Iterator referencing end of range to replace. |
| 1992 | * @param __str String value to insert. |
| 1993 | * @return Reference to this string. |
| 1994 | * @throw std::length_error If new length exceeds @c max_size(). |
| 1995 | * |
| 1996 | * Removes the characters in the range [__i1,__i2). In place, |
| 1997 | * the value of @a __str is inserted. If the length of result |
| 1998 | * exceeds max_size(), length_error is thrown. The value of |
| 1999 | * the string doesn't change if an error is thrown. |
| 2000 | */ |
| 2001 | basic_string& |
| 2002 | replace(__const_iterator __i1, __const_iterator __i2, |
| 2003 | const basic_string& __str) |
| 2004 | { return this->replace(__i1, __i2, __str._M_data(), __str.size()); } |
| 2005 | |
| 2006 | /** |
| 2007 | * @brief Replace range of characters with C substring. |
| 2008 | * @param __i1 Iterator referencing start of range to replace. |
| 2009 | * @param __i2 Iterator referencing end of range to replace. |
| 2010 | * @param __s C string value to insert. |
| 2011 | * @param __n Number of characters from s to insert. |
| 2012 | * @return Reference to this string. |
| 2013 | * @throw std::length_error If new length exceeds @c max_size(). |
| 2014 | * |
| 2015 | * Removes the characters in the range [__i1,__i2). In place, |
| 2016 | * the first @a __n characters of @a __s are inserted. If the |
| 2017 | * length of result exceeds max_size(), length_error is thrown. |
| 2018 | * The value of the string doesn't change if an error is |
| 2019 | * thrown. |
| 2020 | */ |
| 2021 | basic_string& |
| 2022 | replace(__const_iterator __i1, __const_iterator __i2, |
| 2023 | const _CharT* __s, size_type __n) |
| 2024 | { |
| 2025 | _GLIBCXX_DEBUG_PEDASSERT(begin() <= __i1 && __i1 <= __i2 |
| 2026 | && __i2 <= end()); |
| 2027 | return this->replace(__i1 - begin(), __i2 - __i1, __s, __n); |
| 2028 | } |
| 2029 | |
| 2030 | /** |
| 2031 | * @brief Replace range of characters with C string. |
| 2032 | * @param __i1 Iterator referencing start of range to replace. |
| 2033 | * @param __i2 Iterator referencing end of range to replace. |
| 2034 | * @param __s C string value to insert. |
| 2035 | * @return Reference to this string. |
| 2036 | * @throw std::length_error If new length exceeds @c max_size(). |
| 2037 | * |
| 2038 | * Removes the characters in the range [__i1,__i2). In place, |
| 2039 | * the characters of @a __s are inserted. If the length of |
| 2040 | * result exceeds max_size(), length_error is thrown. The |
| 2041 | * value of the string doesn't change if an error is thrown. |
| 2042 | */ |
| 2043 | basic_string& |
| 2044 | replace(__const_iterator __i1, __const_iterator __i2, const _CharT* __s) |
| 2045 | { |
| 2046 | __glibcxx_requires_string(__s); |
| 2047 | return this->replace(__i1, __i2, __s, traits_type::length(__s)); |
| 2048 | } |
| 2049 | |
| 2050 | /** |
| 2051 | * @brief Replace range of characters with multiple characters |
| 2052 | * @param __i1 Iterator referencing start of range to replace. |
| 2053 | * @param __i2 Iterator referencing end of range to replace. |
| 2054 | * @param __n Number of characters to insert. |
| 2055 | * @param __c Character to insert. |
| 2056 | * @return Reference to this string. |
| 2057 | * @throw std::length_error If new length exceeds @c max_size(). |
| 2058 | * |
| 2059 | * Removes the characters in the range [__i1,__i2). In place, |
| 2060 | * @a __n copies of @a __c are inserted. If the length of |
| 2061 | * result exceeds max_size(), length_error is thrown. The |
| 2062 | * value of the string doesn't change if an error is thrown. |
| 2063 | */ |
| 2064 | basic_string& |
| 2065 | replace(__const_iterator __i1, __const_iterator __i2, size_type __n, |
| 2066 | _CharT __c) |
| 2067 | { |
| 2068 | _GLIBCXX_DEBUG_PEDASSERT(begin() <= __i1 && __i1 <= __i2 |
| 2069 | && __i2 <= end()); |
| 2070 | return _M_replace_aux(__i1 - begin(), __i2 - __i1, __n, __c); |
| 2071 | } |
| 2072 | |
| 2073 | /** |
| 2074 | * @brief Replace range of characters with range. |
| 2075 | * @param __i1 Iterator referencing start of range to replace. |
| 2076 | * @param __i2 Iterator referencing end of range to replace. |
| 2077 | * @param __k1 Iterator referencing start of range to insert. |
| 2078 | * @param __k2 Iterator referencing end of range to insert. |
| 2079 | * @return Reference to this string. |
| 2080 | * @throw std::length_error If new length exceeds @c max_size(). |
| 2081 | * |
| 2082 | * Removes the characters in the range [__i1,__i2). In place, |
| 2083 | * characters in the range [__k1,__k2) are inserted. If the |
| 2084 | * length of result exceeds max_size(), length_error is thrown. |
| 2085 | * The value of the string doesn't change if an error is |
| 2086 | * thrown. |
| 2087 | */ |
| 2088 | #if __cplusplus201703L >= 201103L |
| 2089 | template<class _InputIterator, |
| 2090 | typename = std::_RequireInputIter<_InputIterator>> |
| 2091 | basic_string& |
| 2092 | replace(const_iterator __i1, const_iterator __i2, |
| 2093 | _InputIterator __k1, _InputIterator __k2) |
| 2094 | { |
| 2095 | _GLIBCXX_DEBUG_PEDASSERT(begin() <= __i1 && __i1 <= __i2 |
| 2096 | && __i2 <= end()); |
| 2097 | __glibcxx_requires_valid_range(__k1, __k2); |
| 2098 | return this->_M_replace_dispatch(__i1, __i2, __k1, __k2, |
| 2099 | std::__false_type()); |
| 2100 | } |
| 2101 | #else |
| 2102 | template<class _InputIterator> |
| 2103 | #ifdef _GLIBCXX_DISAMBIGUATE_REPLACE_INST |
| 2104 | typename __enable_if_not_native_iterator<_InputIterator>::__type |
| 2105 | #else |
| 2106 | basic_string& |
| 2107 | #endif |
| 2108 | replace(iterator __i1, iterator __i2, |
| 2109 | _InputIterator __k1, _InputIterator __k2) |
| 2110 | { |
| 2111 | _GLIBCXX_DEBUG_PEDASSERT(begin() <= __i1 && __i1 <= __i2 |
| 2112 | && __i2 <= end()); |
| 2113 | __glibcxx_requires_valid_range(__k1, __k2); |
| 2114 | typedef typename std::__is_integer<_InputIterator>::__type _Integral; |
| 2115 | return _M_replace_dispatch(__i1, __i2, __k1, __k2, _Integral()); |
| 2116 | } |
| 2117 | #endif |
| 2118 | |
| 2119 | // Specializations for the common case of pointer and iterator: |
| 2120 | // useful to avoid the overhead of temporary buffering in _M_replace. |
| 2121 | basic_string& |
| 2122 | replace(__const_iterator __i1, __const_iterator __i2, |
| 2123 | _CharT* __k1, _CharT* __k2) |
| 2124 | { |
| 2125 | _GLIBCXX_DEBUG_PEDASSERT(begin() <= __i1 && __i1 <= __i2 |
| 2126 | && __i2 <= end()); |
| 2127 | __glibcxx_requires_valid_range(__k1, __k2); |
| 2128 | return this->replace(__i1 - begin(), __i2 - __i1, |
| 2129 | __k1, __k2 - __k1); |
| 2130 | } |
| 2131 | |
| 2132 | basic_string& |
| 2133 | replace(__const_iterator __i1, __const_iterator __i2, |
| 2134 | const _CharT* __k1, const _CharT* __k2) |
| 2135 | { |
| 2136 | _GLIBCXX_DEBUG_PEDASSERT(begin() <= __i1 && __i1 <= __i2 |
| 2137 | && __i2 <= end()); |
| 2138 | __glibcxx_requires_valid_range(__k1, __k2); |
| 2139 | return this->replace(__i1 - begin(), __i2 - __i1, |
| 2140 | __k1, __k2 - __k1); |
| 2141 | } |
| 2142 | |
| 2143 | basic_string& |
| 2144 | replace(__const_iterator __i1, __const_iterator __i2, |
| 2145 | iterator __k1, iterator __k2) |
| 2146 | { |
| 2147 | _GLIBCXX_DEBUG_PEDASSERT(begin() <= __i1 && __i1 <= __i2 |
| 2148 | && __i2 <= end()); |
| 2149 | __glibcxx_requires_valid_range(__k1, __k2); |
| 2150 | return this->replace(__i1 - begin(), __i2 - __i1, |
| 2151 | __k1.base(), __k2 - __k1); |
| 2152 | } |
| 2153 | |
| 2154 | basic_string& |
| 2155 | replace(__const_iterator __i1, __const_iterator __i2, |
| 2156 | const_iterator __k1, const_iterator __k2) |
| 2157 | { |
| 2158 | _GLIBCXX_DEBUG_PEDASSERT(begin() <= __i1 && __i1 <= __i2 |
| 2159 | && __i2 <= end()); |
| 2160 | __glibcxx_requires_valid_range(__k1, __k2); |
| 2161 | return this->replace(__i1 - begin(), __i2 - __i1, |
| 2162 | __k1.base(), __k2 - __k1); |
| 2163 | } |
| 2164 | |
| 2165 | #if __cplusplus201703L >= 201103L |
| 2166 | /** |
| 2167 | * @brief Replace range of characters with initializer_list. |
| 2168 | * @param __i1 Iterator referencing start of range to replace. |
| 2169 | * @param __i2 Iterator referencing end of range to replace. |
| 2170 | * @param __l The initializer_list of characters to insert. |
| 2171 | * @return Reference to this string. |
| 2172 | * @throw std::length_error If new length exceeds @c max_size(). |
| 2173 | * |
| 2174 | * Removes the characters in the range [__i1,__i2). In place, |
| 2175 | * characters in the range [__k1,__k2) are inserted. If the |
| 2176 | * length of result exceeds max_size(), length_error is thrown. |
| 2177 | * The value of the string doesn't change if an error is |
| 2178 | * thrown. |
| 2179 | */ |
| 2180 | basic_string& replace(const_iterator __i1, const_iterator __i2, |
| 2181 | initializer_list<_CharT> __l) |
| 2182 | { return this->replace(__i1, __i2, __l.begin(), __l.size()); } |
| 2183 | #endif // C++11 |
| 2184 | |
| 2185 | #if __cplusplus201703L >= 201703L |
| 2186 | /** |
| 2187 | * @brief Replace range of characters with string_view. |
| 2188 | * @param __pos The position to replace at. |
| 2189 | * @param __n The number of characters to replace. |
| 2190 | * @param __svt The object convertible to string_view to insert. |
| 2191 | * @return Reference to this string. |
| 2192 | */ |
| 2193 | template<typename _Tp> |
| 2194 | _If_sv<_Tp, basic_string&> |
| 2195 | replace(size_type __pos, size_type __n, const _Tp& __svt) |
| 2196 | { |
| 2197 | __sv_type __sv = __svt; |
| 2198 | return this->replace(__pos, __n, __sv.data(), __sv.size()); |
| 2199 | } |
| 2200 | |
| 2201 | /** |
| 2202 | * @brief Replace range of characters with string_view. |
| 2203 | * @param __pos1 The position to replace at. |
| 2204 | * @param __n1 The number of characters to replace. |
| 2205 | * @param __svt The object convertible to string_view to insert from. |
| 2206 | * @param __pos2 The position in the string_view to insert from. |
| 2207 | * @param __n2 The number of characters to insert. |
| 2208 | * @return Reference to this string. |
| 2209 | */ |
| 2210 | template<typename _Tp> |
| 2211 | _If_sv<_Tp, basic_string&> |
| 2212 | replace(size_type __pos1, size_type __n1, const _Tp& __svt, |
| 2213 | size_type __pos2, size_type __n2 = npos) |
| 2214 | { |
| 2215 | __sv_type __sv = __svt; |
| 2216 | return this->replace(__pos1, __n1, |
| 2217 | __sv.data() |
| 2218 | + std::__sv_check(__sv.size(), __pos2, "basic_string::replace"), |
| 2219 | std::__sv_limit(__sv.size(), __pos2, __n2)); |
| 2220 | } |
| 2221 | |
| 2222 | /** |
| 2223 | * @brief Replace range of characters with string_view. |
| 2224 | * @param __i1 An iterator referencing the start position |
| 2225 | to replace at. |
| 2226 | * @param __i2 An iterator referencing the end position |
| 2227 | for the replace. |
| 2228 | * @param __svt The object convertible to string_view to insert from. |
| 2229 | * @return Reference to this string. |
| 2230 | */ |
| 2231 | template<typename _Tp> |
| 2232 | _If_sv<_Tp, basic_string&> |
| 2233 | replace(const_iterator __i1, const_iterator __i2, const _Tp& __svt) |
| 2234 | { |
| 2235 | __sv_type __sv = __svt; |
| 2236 | return this->replace(__i1 - begin(), __i2 - __i1, __sv); |
| 2237 | } |
| 2238 | #endif // C++17 |
| 2239 | |
| 2240 | private: |
| 2241 | template<class _Integer> |
| 2242 | basic_string& |
| 2243 | _M_replace_dispatch(const_iterator __i1, const_iterator __i2, |
| 2244 | _Integer __n, _Integer __val, __true_type) |
| 2245 | { return _M_replace_aux(__i1 - begin(), __i2 - __i1, __n, __val); } |
| 2246 | |
| 2247 | template<class _InputIterator> |
| 2248 | basic_string& |
| 2249 | _M_replace_dispatch(const_iterator __i1, const_iterator __i2, |
| 2250 | _InputIterator __k1, _InputIterator __k2, |
| 2251 | __false_type); |
| 2252 | |
| 2253 | basic_string& |
| 2254 | _M_replace_aux(size_type __pos1, size_type __n1, size_type __n2, |
| 2255 | _CharT __c); |
| 2256 | |
| 2257 | basic_string& |
| 2258 | _M_replace(size_type __pos, size_type __len1, const _CharT* __s, |
| 2259 | const size_type __len2); |
| 2260 | |
| 2261 | basic_string& |
| 2262 | _M_append(const _CharT* __s, size_type __n); |
| 2263 | |
| 2264 | public: |
| 2265 | |
| 2266 | /** |
| 2267 | * @brief Copy substring into C string. |
| 2268 | * @param __s C string to copy value into. |
| 2269 | * @param __n Number of characters to copy. |
| 2270 | * @param __pos Index of first character to copy. |
| 2271 | * @return Number of characters actually copied |
| 2272 | * @throw std::out_of_range If __pos > size(). |
| 2273 | * |
| 2274 | * Copies up to @a __n characters starting at @a __pos into the |
| 2275 | * C string @a __s. If @a __pos is %greater than size(), |
| 2276 | * out_of_range is thrown. |
| 2277 | */ |
| 2278 | size_type |
| 2279 | copy(_CharT* __s, size_type __n, size_type __pos = 0) const; |
| 2280 | |
| 2281 | /** |
| 2282 | * @brief Swap contents with another string. |
| 2283 | * @param __s String to swap with. |
| 2284 | * |
| 2285 | * Exchanges the contents of this string with that of @a __s in constant |
| 2286 | * time. |
| 2287 | */ |
| 2288 | void |
| 2289 | swap(basic_string& __s) _GLIBCXX_NOEXCEPTnoexcept; |
| 2290 | |
| 2291 | // String operations: |
| 2292 | /** |
| 2293 | * @brief Return const pointer to null-terminated contents. |
| 2294 | * |
| 2295 | * This is a handle to internal data. Do not modify or dire things may |
| 2296 | * happen. |
| 2297 | */ |
| 2298 | const _CharT* |
| 2299 | c_str() const _GLIBCXX_NOEXCEPTnoexcept |
| 2300 | { return _M_data(); } |
| 2301 | |
| 2302 | /** |
| 2303 | * @brief Return const pointer to contents. |
| 2304 | * |
| 2305 | * This is a pointer to internal data. It is undefined to modify |
| 2306 | * the contents through the returned pointer. To get a pointer that |
| 2307 | * allows modifying the contents use @c &str[0] instead, |
| 2308 | * (or in C++17 the non-const @c str.data() overload). |
| 2309 | */ |
| 2310 | const _CharT* |
| 2311 | data() const _GLIBCXX_NOEXCEPTnoexcept |
| 2312 | { return _M_data(); } |
| 2313 | |
| 2314 | #if __cplusplus201703L >= 201703L |
| 2315 | /** |
| 2316 | * @brief Return non-const pointer to contents. |
| 2317 | * |
| 2318 | * This is a pointer to the character sequence held by the string. |
| 2319 | * Modifying the characters in the sequence is allowed. |
| 2320 | */ |
| 2321 | _CharT* |
| 2322 | data() noexcept |
| 2323 | { return _M_data(); } |
| 2324 | #endif |
| 2325 | |
| 2326 | /** |
| 2327 | * @brief Return copy of allocator used to construct this string. |
| 2328 | */ |
| 2329 | allocator_type |
| 2330 | get_allocator() const _GLIBCXX_NOEXCEPTnoexcept |
| 2331 | { return _M_get_allocator(); } |
| 2332 | |
| 2333 | /** |
| 2334 | * @brief Find position of a C substring. |
| 2335 | * @param __s C string to locate. |
| 2336 | * @param __pos Index of character to search from. |
| 2337 | * @param __n Number of characters from @a s to search for. |
| 2338 | * @return Index of start of first occurrence. |
| 2339 | * |
| 2340 | * Starting from @a __pos, searches forward for the first @a |
| 2341 | * __n characters in @a __s within this string. If found, |
| 2342 | * returns the index where it begins. If not found, returns |
| 2343 | * npos. |
| 2344 | */ |
| 2345 | size_type |
| 2346 | find(const _CharT* __s, size_type __pos, size_type __n) const |
| 2347 | _GLIBCXX_NOEXCEPTnoexcept; |
| 2348 | |
| 2349 | /** |
| 2350 | * @brief Find position of a string. |
| 2351 | * @param __str String to locate. |
| 2352 | * @param __pos Index of character to search from (default 0). |
| 2353 | * @return Index of start of first occurrence. |
| 2354 | * |
| 2355 | * Starting from @a __pos, searches forward for value of @a __str within |
| 2356 | * this string. If found, returns the index where it begins. If not |
| 2357 | * found, returns npos. |
| 2358 | */ |
| 2359 | size_type |
| 2360 | find(const basic_string& __str, size_type __pos = 0) const |
| 2361 | _GLIBCXX_NOEXCEPTnoexcept |
| 2362 | { return this->find(__str.data(), __pos, __str.size()); } |
| 2363 | |
| 2364 | #if __cplusplus201703L >= 201703L |
| 2365 | /** |
| 2366 | * @brief Find position of a string_view. |
| 2367 | * @param __svt The object convertible to string_view to locate. |
| 2368 | * @param __pos Index of character to search from (default 0). |
| 2369 | * @return Index of start of first occurrence. |
| 2370 | */ |
| 2371 | template<typename _Tp> |
| 2372 | _If_sv<_Tp, size_type> |
| 2373 | find(const _Tp& __svt, size_type __pos = 0) const |
| 2374 | noexcept(is_same<_Tp, __sv_type>::value) |
| 2375 | { |
| 2376 | __sv_type __sv = __svt; |
| 2377 | return this->find(__sv.data(), __pos, __sv.size()); |
| 2378 | } |
| 2379 | #endif // C++17 |
| 2380 | |
| 2381 | /** |
| 2382 | * @brief Find position of a C string. |
| 2383 | * @param __s C string to locate. |
| 2384 | * @param __pos Index of character to search from (default 0). |
| 2385 | * @return Index of start of first occurrence. |
| 2386 | * |
| 2387 | * Starting from @a __pos, searches forward for the value of @a |
| 2388 | * __s within this string. If found, returns the index where |
| 2389 | * it begins. If not found, returns npos. |
| 2390 | */ |
| 2391 | size_type |
| 2392 | find(const _CharT* __s, size_type __pos = 0) const _GLIBCXX_NOEXCEPTnoexcept |
| 2393 | { |
| 2394 | __glibcxx_requires_string(__s); |
| 2395 | return this->find(__s, __pos, traits_type::length(__s)); |
| 2396 | } |
| 2397 | |
| 2398 | /** |
| 2399 | * @brief Find position of a character. |
| 2400 | * @param __c Character to locate. |
| 2401 | * @param __pos Index of character to search from (default 0). |
| 2402 | * @return Index of first occurrence. |
| 2403 | * |
| 2404 | * Starting from @a __pos, searches forward for @a __c within |
| 2405 | * this string. If found, returns the index where it was |
| 2406 | * found. If not found, returns npos. |
| 2407 | */ |
| 2408 | size_type |
| 2409 | find(_CharT __c, size_type __pos = 0) const _GLIBCXX_NOEXCEPTnoexcept; |
| 2410 | |
| 2411 | /** |
| 2412 | * @brief Find last position of a string. |
| 2413 | * @param __str String to locate. |
| 2414 | * @param __pos Index of character to search back from (default end). |
| 2415 | * @return Index of start of last occurrence. |
| 2416 | * |
| 2417 | * Starting from @a __pos, searches backward for value of @a |
| 2418 | * __str within this string. If found, returns the index where |
| 2419 | * it begins. If not found, returns npos. |
| 2420 | */ |
| 2421 | size_type |
| 2422 | rfind(const basic_string& __str, size_type __pos = npos) const |
| 2423 | _GLIBCXX_NOEXCEPTnoexcept |
| 2424 | { return this->rfind(__str.data(), __pos, __str.size()); } |
| 2425 | |
| 2426 | #if __cplusplus201703L >= 201703L |
| 2427 | /** |
| 2428 | * @brief Find last position of a string_view. |
| 2429 | * @param __svt The object convertible to string_view to locate. |
| 2430 | * @param __pos Index of character to search back from (default end). |
| 2431 | * @return Index of start of last occurrence. |
| 2432 | */ |
| 2433 | template<typename _Tp> |
| 2434 | _If_sv<_Tp, size_type> |
| 2435 | rfind(const _Tp& __svt, size_type __pos = npos) const |
| 2436 | noexcept(is_same<_Tp, __sv_type>::value) |
| 2437 | { |
| 2438 | __sv_type __sv = __svt; |
| 2439 | return this->rfind(__sv.data(), __pos, __sv.size()); |
| 2440 | } |
| 2441 | #endif // C++17 |
| 2442 | |
| 2443 | /** |
| 2444 | * @brief Find last position of a C substring. |
| 2445 | * @param __s C string to locate. |
| 2446 | * @param __pos Index of character to search back from. |
| 2447 | * @param __n Number of characters from s to search for. |
| 2448 | * @return Index of start of last occurrence. |
| 2449 | * |
| 2450 | * Starting from @a __pos, searches backward for the first @a |
| 2451 | * __n characters in @a __s within this string. If found, |
| 2452 | * returns the index where it begins. If not found, returns |
| 2453 | * npos. |
| 2454 | */ |
| 2455 | size_type |
| 2456 | rfind(const _CharT* __s, size_type __pos, size_type __n) const |
| 2457 | _GLIBCXX_NOEXCEPTnoexcept; |
| 2458 | |
| 2459 | /** |
| 2460 | * @brief Find last position of a C string. |
| 2461 | * @param __s C string to locate. |
| 2462 | * @param __pos Index of character to start search at (default end). |
| 2463 | * @return Index of start of last occurrence. |
| 2464 | * |
| 2465 | * Starting from @a __pos, searches backward for the value of |
| 2466 | * @a __s within this string. If found, returns the index |
| 2467 | * where it begins. If not found, returns npos. |
| 2468 | */ |
| 2469 | size_type |
| 2470 | rfind(const _CharT* __s, size_type __pos = npos) const |
| 2471 | { |
| 2472 | __glibcxx_requires_string(__s); |
| 2473 | return this->rfind(__s, __pos, traits_type::length(__s)); |
| 2474 | } |
| 2475 | |
| 2476 | /** |
| 2477 | * @brief Find last position of a character. |
| 2478 | * @param __c Character to locate. |
| 2479 | * @param __pos Index of character to search back from (default end). |
| 2480 | * @return Index of last occurrence. |
| 2481 | * |
| 2482 | * Starting from @a __pos, searches backward for @a __c within |
| 2483 | * this string. If found, returns the index where it was |
| 2484 | * found. If not found, returns npos. |
| 2485 | */ |
| 2486 | size_type |
| 2487 | rfind(_CharT __c, size_type __pos = npos) const _GLIBCXX_NOEXCEPTnoexcept; |
| 2488 | |
| 2489 | /** |
| 2490 | * @brief Find position of a character of string. |
| 2491 | * @param __str String containing characters to locate. |
| 2492 | * @param __pos Index of character to search from (default 0). |
| 2493 | * @return Index of first occurrence. |
| 2494 | * |
| 2495 | * Starting from @a __pos, searches forward for one of the |
| 2496 | * characters of @a __str within this string. If found, |
| 2497 | * returns the index where it was found. If not found, returns |
| 2498 | * npos. |
| 2499 | */ |
| 2500 | size_type |
| 2501 | find_first_of(const basic_string& __str, size_type __pos = 0) const |
| 2502 | _GLIBCXX_NOEXCEPTnoexcept |
| 2503 | { return this->find_first_of(__str.data(), __pos, __str.size()); } |
| 2504 | |
| 2505 | #if __cplusplus201703L >= 201703L |
| 2506 | /** |
| 2507 | * @brief Find position of a character of a string_view. |
| 2508 | * @param __svt An object convertible to string_view containing |
| 2509 | * characters to locate. |
| 2510 | * @param __pos Index of character to search from (default 0). |
| 2511 | * @return Index of first occurrence. |
| 2512 | */ |
| 2513 | template<typename _Tp> |
| 2514 | _If_sv<_Tp, size_type> |
| 2515 | find_first_of(const _Tp& __svt, size_type __pos = 0) const |
| 2516 | noexcept(is_same<_Tp, __sv_type>::value) |
| 2517 | { |
| 2518 | __sv_type __sv = __svt; |
| 2519 | return this->find_first_of(__sv.data(), __pos, __sv.size()); |
| 2520 | } |
| 2521 | #endif // C++17 |
| 2522 | |
| 2523 | /** |
| 2524 | * @brief Find position of a character of C substring. |
| 2525 | * @param __s String containing characters to locate. |
| 2526 | * @param __pos Index of character to search from. |
| 2527 | * @param __n Number of characters from s to search for. |
| 2528 | * @return Index of first occurrence. |
| 2529 | * |
| 2530 | * Starting from @a __pos, searches forward for one of the |
| 2531 | * first @a __n characters of @a __s within this string. If |
| 2532 | * found, returns the index where it was found. If not found, |
| 2533 | * returns npos. |
| 2534 | */ |
| 2535 | size_type |
| 2536 | find_first_of(const _CharT* __s, size_type __pos, size_type __n) const |
| 2537 | _GLIBCXX_NOEXCEPTnoexcept; |
| 2538 | |
| 2539 | /** |
| 2540 | * @brief Find position of a character of C string. |
| 2541 | * @param __s String containing characters to locate. |
| 2542 | * @param __pos Index of character to search from (default 0). |
| 2543 | * @return Index of first occurrence. |
| 2544 | * |
| 2545 | * Starting from @a __pos, searches forward for one of the |
| 2546 | * characters of @a __s within this string. If found, returns |
| 2547 | * the index where it was found. If not found, returns npos. |
| 2548 | */ |
| 2549 | size_type |
| 2550 | find_first_of(const _CharT* __s, size_type __pos = 0) const |
| 2551 | _GLIBCXX_NOEXCEPTnoexcept |
| 2552 | { |
| 2553 | __glibcxx_requires_string(__s); |
| 2554 | return this->find_first_of(__s, __pos, traits_type::length(__s)); |
| 2555 | } |
| 2556 | |
| 2557 | /** |
| 2558 | * @brief Find position of a character. |
| 2559 | * @param __c Character to locate. |
| 2560 | * @param __pos Index of character to search from (default 0). |
| 2561 | * @return Index of first occurrence. |
| 2562 | * |
| 2563 | * Starting from @a __pos, searches forward for the character |
| 2564 | * @a __c within this string. If found, returns the index |
| 2565 | * where it was found. If not found, returns npos. |
| 2566 | * |
| 2567 | * Note: equivalent to find(__c, __pos). |
| 2568 | */ |
| 2569 | size_type |
| 2570 | find_first_of(_CharT __c, size_type __pos = 0) const _GLIBCXX_NOEXCEPTnoexcept |
| 2571 | { return this->find(__c, __pos); } |
| 2572 | |
| 2573 | /** |
| 2574 | * @brief Find last position of a character of string. |
| 2575 | * @param __str String containing characters to locate. |
| 2576 | * @param __pos Index of character to search back from (default end). |
| 2577 | * @return Index of last occurrence. |
| 2578 | * |
| 2579 | * Starting from @a __pos, searches backward for one of the |
| 2580 | * characters of @a __str within this string. If found, |
| 2581 | * returns the index where it was found. If not found, returns |
| 2582 | * npos. |
| 2583 | */ |
| 2584 | size_type |
| 2585 | find_last_of(const basic_string& __str, size_type __pos = npos) const |
| 2586 | _GLIBCXX_NOEXCEPTnoexcept |
| 2587 | { return this->find_last_of(__str.data(), __pos, __str.size()); } |
| 2588 | |
| 2589 | #if __cplusplus201703L >= 201703L |
| 2590 | /** |
| 2591 | * @brief Find last position of a character of string. |
| 2592 | * @param __svt An object convertible to string_view containing |
| 2593 | * characters to locate. |
| 2594 | * @param __pos Index of character to search back from (default end). |
| 2595 | * @return Index of last occurrence. |
| 2596 | */ |
| 2597 | template<typename _Tp> |
| 2598 | _If_sv<_Tp, size_type> |
| 2599 | find_last_of(const _Tp& __svt, size_type __pos = npos) const |
| 2600 | noexcept(is_same<_Tp, __sv_type>::value) |
| 2601 | { |
| 2602 | __sv_type __sv = __svt; |
| 2603 | return this->find_last_of(__sv.data(), __pos, __sv.size()); |
| 2604 | } |
| 2605 | #endif // C++17 |
| 2606 | |
| 2607 | /** |
| 2608 | * @brief Find last position of a character of C substring. |
| 2609 | * @param __s C string containing characters to locate. |
| 2610 | * @param __pos Index of character to search back from. |
| 2611 | * @param __n Number of characters from s to search for. |
| 2612 | * @return Index of last occurrence. |
| 2613 | * |
| 2614 | * Starting from @a __pos, searches backward for one of the |
| 2615 | * first @a __n characters of @a __s within this string. If |
| 2616 | * found, returns the index where it was found. If not found, |
| 2617 | * returns npos. |
| 2618 | */ |
| 2619 | size_type |
| 2620 | find_last_of(const _CharT* __s, size_type __pos, size_type __n) const |
| 2621 | _GLIBCXX_NOEXCEPTnoexcept; |
| 2622 | |
| 2623 | /** |
| 2624 | * @brief Find last position of a character of C string. |
| 2625 | * @param __s C string containing characters to locate. |
| 2626 | * @param __pos Index of character to search back from (default end). |
| 2627 | * @return Index of last occurrence. |
| 2628 | * |
| 2629 | * Starting from @a __pos, searches backward for one of the |
| 2630 | * characters of @a __s within this string. If found, returns |
| 2631 | * the index where it was found. If not found, returns npos. |
| 2632 | */ |
| 2633 | size_type |
| 2634 | find_last_of(const _CharT* __s, size_type __pos = npos) const |
| 2635 | _GLIBCXX_NOEXCEPTnoexcept |
| 2636 | { |
| 2637 | __glibcxx_requires_string(__s); |
| 2638 | return this->find_last_of(__s, __pos, traits_type::length(__s)); |
| 2639 | } |
| 2640 | |
| 2641 | /** |
| 2642 | * @brief Find last position of a character. |
| 2643 | * @param __c Character to locate. |
| 2644 | * @param __pos Index of character to search back from (default end). |
| 2645 | * @return Index of last occurrence. |
| 2646 | * |
| 2647 | * Starting from @a __pos, searches backward for @a __c within |
| 2648 | * this string. If found, returns the index where it was |
| 2649 | * found. If not found, returns npos. |
| 2650 | * |
| 2651 | * Note: equivalent to rfind(__c, __pos). |
| 2652 | */ |
| 2653 | size_type |
| 2654 | find_last_of(_CharT __c, size_type __pos = npos) const _GLIBCXX_NOEXCEPTnoexcept |
| 2655 | { return this->rfind(__c, __pos); } |
| 2656 | |
| 2657 | /** |
| 2658 | * @brief Find position of a character not in string. |
| 2659 | * @param __str String containing characters to avoid. |
| 2660 | * @param __pos Index of character to search from (default 0). |
| 2661 | * @return Index of first occurrence. |
| 2662 | * |
| 2663 | * Starting from @a __pos, searches forward for a character not contained |
| 2664 | * in @a __str within this string. If found, returns the index where it |
| 2665 | * was found. If not found, returns npos. |
| 2666 | */ |
| 2667 | size_type |
| 2668 | find_first_not_of(const basic_string& __str, size_type __pos = 0) const |
| 2669 | _GLIBCXX_NOEXCEPTnoexcept |
| 2670 | { return this->find_first_not_of(__str.data(), __pos, __str.size()); } |
| 2671 | |
| 2672 | #if __cplusplus201703L >= 201703L |
| 2673 | /** |
| 2674 | * @brief Find position of a character not in a string_view. |
| 2675 | * @param __svt A object convertible to string_view containing |
| 2676 | * characters to avoid. |
| 2677 | * @param __pos Index of character to search from (default 0). |
| 2678 | * @return Index of first occurrence. |
| 2679 | */ |
| 2680 | template<typename _Tp> |
| 2681 | _If_sv<_Tp, size_type> |
| 2682 | find_first_not_of(const _Tp& __svt, size_type __pos = 0) const |
| 2683 | noexcept(is_same<_Tp, __sv_type>::value) |
| 2684 | { |
| 2685 | __sv_type __sv = __svt; |
| 2686 | return this->find_first_not_of(__sv.data(), __pos, __sv.size()); |
| 2687 | } |
| 2688 | #endif // C++17 |
| 2689 | |
| 2690 | /** |
| 2691 | * @brief Find position of a character not in C substring. |
| 2692 | * @param __s C string containing characters to avoid. |
| 2693 | * @param __pos Index of character to search from. |
| 2694 | * @param __n Number of characters from __s to consider. |
| 2695 | * @return Index of first occurrence. |
| 2696 | * |
| 2697 | * Starting from @a __pos, searches forward for a character not |
| 2698 | * contained in the first @a __n characters of @a __s within |
| 2699 | * this string. If found, returns the index where it was |
| 2700 | * found. If not found, returns npos. |
| 2701 | */ |
| 2702 | size_type |
| 2703 | find_first_not_of(const _CharT* __s, size_type __pos, |
| 2704 | size_type __n) const _GLIBCXX_NOEXCEPTnoexcept; |
| 2705 | |
| 2706 | /** |
| 2707 | * @brief Find position of a character not in C string. |
| 2708 | * @param __s C string containing characters to avoid. |
| 2709 | * @param __pos Index of character to search from (default 0). |
| 2710 | * @return Index of first occurrence. |
| 2711 | * |
| 2712 | * Starting from @a __pos, searches forward for a character not |
| 2713 | * contained in @a __s within this string. If found, returns |
| 2714 | * the index where it was found. If not found, returns npos. |
| 2715 | */ |
| 2716 | size_type |
| 2717 | find_first_not_of(const _CharT* __s, size_type __pos = 0) const |
| 2718 | _GLIBCXX_NOEXCEPTnoexcept |
| 2719 | { |
| 2720 | __glibcxx_requires_string(__s); |
| 2721 | return this->find_first_not_of(__s, __pos, traits_type::length(__s)); |
| 2722 | } |
| 2723 | |
| 2724 | /** |
| 2725 | * @brief Find position of a different character. |
| 2726 | * @param __c Character to avoid. |
| 2727 | * @param __pos Index of character to search from (default 0). |
| 2728 | * @return Index of first occurrence. |
| 2729 | * |
| 2730 | * Starting from @a __pos, searches forward for a character |
| 2731 | * other than @a __c within this string. If found, returns the |
| 2732 | * index where it was found. If not found, returns npos. |
| 2733 | */ |
| 2734 | size_type |
| 2735 | find_first_not_of(_CharT __c, size_type __pos = 0) const |
| 2736 | _GLIBCXX_NOEXCEPTnoexcept; |
| 2737 | |
| 2738 | /** |
| 2739 | * @brief Find last position of a character not in string. |
| 2740 | * @param __str String containing characters to avoid. |
| 2741 | * @param __pos Index of character to search back from (default end). |
| 2742 | * @return Index of last occurrence. |
| 2743 | * |
| 2744 | * Starting from @a __pos, searches backward for a character |
| 2745 | * not contained in @a __str within this string. If found, |
| 2746 | * returns the index where it was found. If not found, returns |
| 2747 | * npos. |
| 2748 | */ |
| 2749 | size_type |
| 2750 | find_last_not_of(const basic_string& __str, size_type __pos = npos) const |
| 2751 | _GLIBCXX_NOEXCEPTnoexcept |
| 2752 | { return this->find_last_not_of(__str.data(), __pos, __str.size()); } |
| 2753 | |
| 2754 | #if __cplusplus201703L >= 201703L |
| 2755 | /** |
| 2756 | * @brief Find last position of a character not in a string_view. |
| 2757 | * @param __svt An object convertible to string_view containing |
| 2758 | * characters to avoid. |
| 2759 | * @param __pos Index of character to search back from (default end). |
| 2760 | * @return Index of last occurrence. |
| 2761 | */ |
| 2762 | template<typename _Tp> |
| 2763 | _If_sv<_Tp, size_type> |
| 2764 | find_last_not_of(const _Tp& __svt, size_type __pos = npos) const |
| 2765 | noexcept(is_same<_Tp, __sv_type>::value) |
| 2766 | { |
| 2767 | __sv_type __sv = __svt; |
| 2768 | return this->find_last_not_of(__sv.data(), __pos, __sv.size()); |
| 2769 | } |
| 2770 | #endif // C++17 |
| 2771 | |
| 2772 | /** |
| 2773 | * @brief Find last position of a character not in C substring. |
| 2774 | * @param __s C string containing characters to avoid. |
| 2775 | * @param __pos Index of character to search back from. |
| 2776 | * @param __n Number of characters from s to consider. |
| 2777 | * @return Index of last occurrence. |
| 2778 | * |
| 2779 | * Starting from @a __pos, searches backward for a character not |
| 2780 | * contained in the first @a __n characters of @a __s within this string. |
| 2781 | * If found, returns the index where it was found. If not found, |
| 2782 | * returns npos. |
| 2783 | */ |
| 2784 | size_type |
| 2785 | find_last_not_of(const _CharT* __s, size_type __pos, |
| 2786 | size_type __n) const _GLIBCXX_NOEXCEPTnoexcept; |
| 2787 | /** |
| 2788 | * @brief Find last position of a character not in C string. |
| 2789 | * @param __s C string containing characters to avoid. |
| 2790 | * @param __pos Index of character to search back from (default end). |
| 2791 | * @return Index of last occurrence. |
| 2792 | * |
| 2793 | * Starting from @a __pos, searches backward for a character |
| 2794 | * not contained in @a __s within this string. If found, |
| 2795 | * returns the index where it was found. If not found, returns |
| 2796 | * npos. |
| 2797 | */ |
| 2798 | size_type |
| 2799 | find_last_not_of(const _CharT* __s, size_type __pos = npos) const |
| 2800 | _GLIBCXX_NOEXCEPTnoexcept |
| 2801 | { |
| 2802 | __glibcxx_requires_string(__s); |
| 2803 | return this->find_last_not_of(__s, __pos, traits_type::length(__s)); |
| 2804 | } |
| 2805 | |
| 2806 | /** |
| 2807 | * @brief Find last position of a different character. |
| 2808 | * @param __c Character to avoid. |
| 2809 | * @param __pos Index of character to search back from (default end). |
| 2810 | * @return Index of last occurrence. |
| 2811 | * |
| 2812 | * Starting from @a __pos, searches backward for a character other than |
| 2813 | * @a __c within this string. If found, returns the index where it was |
| 2814 | * found. If not found, returns npos. |
| 2815 | */ |
| 2816 | size_type |
| 2817 | find_last_not_of(_CharT __c, size_type __pos = npos) const |
| 2818 | _GLIBCXX_NOEXCEPTnoexcept; |
| 2819 | |
| 2820 | /** |
| 2821 | * @brief Get a substring. |
| 2822 | * @param __pos Index of first character (default 0). |
| 2823 | * @param __n Number of characters in substring (default remainder). |
| 2824 | * @return The new string. |
| 2825 | * @throw std::out_of_range If __pos > size(). |
| 2826 | * |
| 2827 | * Construct and return a new string using the @a __n |
| 2828 | * characters starting at @a __pos. If the string is too |
| 2829 | * short, use the remainder of the characters. If @a __pos is |
| 2830 | * beyond the end of the string, out_of_range is thrown. |
| 2831 | */ |
| 2832 | basic_string |
| 2833 | substr(size_type __pos = 0, size_type __n = npos) const |
| 2834 | { return basic_string(*this, |
| 2835 | _M_check(__pos, "basic_string::substr"), __n); } |
| 2836 | |
| 2837 | /** |
| 2838 | * @brief Compare to a string. |
| 2839 | * @param __str String to compare against. |
| 2840 | * @return Integer < 0, 0, or > 0. |
| 2841 | * |
| 2842 | * Returns an integer < 0 if this string is ordered before @a |
| 2843 | * __str, 0 if their values are equivalent, or > 0 if this |
| 2844 | * string is ordered after @a __str. Determines the effective |
| 2845 | * length rlen of the strings to compare as the smallest of |
| 2846 | * size() and str.size(). The function then compares the two |
| 2847 | * strings by calling traits::compare(data(), str.data(),rlen). |
| 2848 | * If the result of the comparison is nonzero returns it, |
| 2849 | * otherwise the shorter one is ordered first. |
| 2850 | */ |
| 2851 | int |
| 2852 | compare(const basic_string& __str) const |
| 2853 | { |
| 2854 | const size_type __size = this->size(); |
| 2855 | const size_type __osize = __str.size(); |
| 2856 | const size_type __len = std::min(__size, __osize); |
| 2857 | |
| 2858 | int __r = traits_type::compare(_M_data(), __str.data(), __len); |
| 2859 | if (!__r) |
| 2860 | __r = _S_compare(__size, __osize); |
| 2861 | return __r; |
| 2862 | } |
| 2863 | |
| 2864 | #if __cplusplus201703L >= 201703L |
| 2865 | /** |
| 2866 | * @brief Compare to a string_view. |
| 2867 | * @param __svt An object convertible to string_view to compare against. |
| 2868 | * @return Integer < 0, 0, or > 0. |
| 2869 | */ |
| 2870 | template<typename _Tp> |
| 2871 | _If_sv<_Tp, int> |
| 2872 | compare(const _Tp& __svt) const |
| 2873 | noexcept(is_same<_Tp, __sv_type>::value) |
| 2874 | { |
| 2875 | __sv_type __sv = __svt; |
| 2876 | const size_type __size = this->size(); |
| 2877 | const size_type __osize = __sv.size(); |
| 2878 | const size_type __len = std::min(__size, __osize); |
| 2879 | |
| 2880 | int __r = traits_type::compare(_M_data(), __sv.data(), __len); |
| 2881 | if (!__r) |
| 2882 | __r = _S_compare(__size, __osize); |
| 2883 | return __r; |
| 2884 | } |
| 2885 | |
| 2886 | /** |
| 2887 | * @brief Compare to a string_view. |
| 2888 | * @param __pos A position in the string to start comparing from. |
| 2889 | * @param __n The number of characters to compare. |
| 2890 | * @param __svt An object convertible to string_view to compare |
| 2891 | * against. |
| 2892 | * @return Integer < 0, 0, or > 0. |
| 2893 | */ |
| 2894 | template<typename _Tp> |
| 2895 | _If_sv<_Tp, int> |
| 2896 | compare(size_type __pos, size_type __n, const _Tp& __svt) const |
| 2897 | noexcept(is_same<_Tp, __sv_type>::value) |
| 2898 | { |
| 2899 | __sv_type __sv = __svt; |
| 2900 | return __sv_type(*this).substr(__pos, __n).compare(__sv); |
| 2901 | } |
| 2902 | |
| 2903 | /** |
| 2904 | * @brief Compare to a string_view. |
| 2905 | * @param __pos1 A position in the string to start comparing from. |
| 2906 | * @param __n1 The number of characters to compare. |
| 2907 | * @param __svt An object convertible to string_view to compare |
| 2908 | * against. |
| 2909 | * @param __pos2 A position in the string_view to start comparing from. |
| 2910 | * @param __n2 The number of characters to compare. |
| 2911 | * @return Integer < 0, 0, or > 0. |
| 2912 | */ |
| 2913 | template<typename _Tp> |
| 2914 | _If_sv<_Tp, int> |
| 2915 | compare(size_type __pos1, size_type __n1, const _Tp& __svt, |
| 2916 | size_type __pos2, size_type __n2 = npos) const |
| 2917 | noexcept(is_same<_Tp, __sv_type>::value) |
| 2918 | { |
| 2919 | __sv_type __sv = __svt; |
| 2920 | return __sv_type(*this) |
| 2921 | .substr(__pos1, __n1).compare(__sv.substr(__pos2, __n2)); |
| 2922 | } |
| 2923 | #endif // C++17 |
| 2924 | |
| 2925 | /** |
| 2926 | * @brief Compare substring to a string. |
| 2927 | * @param __pos Index of first character of substring. |
| 2928 | * @param __n Number of characters in substring. |
| 2929 | * @param __str String to compare against. |
| 2930 | * @return Integer < 0, 0, or > 0. |
| 2931 | * |
| 2932 | * Form the substring of this string from the @a __n characters |
| 2933 | * starting at @a __pos. Returns an integer < 0 if the |
| 2934 | * substring is ordered before @a __str, 0 if their values are |
| 2935 | * equivalent, or > 0 if the substring is ordered after @a |
| 2936 | * __str. Determines the effective length rlen of the strings |
| 2937 | * to compare as the smallest of the length of the substring |
| 2938 | * and @a __str.size(). The function then compares the two |
| 2939 | * strings by calling |
| 2940 | * traits::compare(substring.data(),str.data(),rlen). If the |
| 2941 | * result of the comparison is nonzero returns it, otherwise |
| 2942 | * the shorter one is ordered first. |
| 2943 | */ |
| 2944 | int |
| 2945 | compare(size_type __pos, size_type __n, const basic_string& __str) const; |
| 2946 | |
| 2947 | /** |
| 2948 | * @brief Compare substring to a substring. |
| 2949 | * @param __pos1 Index of first character of substring. |
| 2950 | * @param __n1 Number of characters in substring. |
| 2951 | * @param __str String to compare against. |
| 2952 | * @param __pos2 Index of first character of substring of str. |
| 2953 | * @param __n2 Number of characters in substring of str. |
| 2954 | * @return Integer < 0, 0, or > 0. |
| 2955 | * |
| 2956 | * Form the substring of this string from the @a __n1 |
| 2957 | * characters starting at @a __pos1. Form the substring of @a |
| 2958 | * __str from the @a __n2 characters starting at @a __pos2. |
| 2959 | * Returns an integer < 0 if this substring is ordered before |
| 2960 | * the substring of @a __str, 0 if their values are equivalent, |
| 2961 | * or > 0 if this substring is ordered after the substring of |
| 2962 | * @a __str. Determines the effective length rlen of the |
| 2963 | * strings to compare as the smallest of the lengths of the |
| 2964 | * substrings. The function then compares the two strings by |
| 2965 | * calling |
| 2966 | * traits::compare(substring.data(),str.substr(pos2,n2).data(),rlen). |
| 2967 | * If the result of the comparison is nonzero returns it, |
| 2968 | * otherwise the shorter one is ordered first. |
| 2969 | */ |
| 2970 | int |
| 2971 | compare(size_type __pos1, size_type __n1, const basic_string& __str, |
| 2972 | size_type __pos2, size_type __n2 = npos) const; |
| 2973 | |
| 2974 | /** |
| 2975 | * @brief Compare to a C string. |
| 2976 | * @param __s C string to compare against. |
| 2977 | * @return Integer < 0, 0, or > 0. |
| 2978 | * |
| 2979 | * Returns an integer < 0 if this string is ordered before @a __s, 0 if |
| 2980 | * their values are equivalent, or > 0 if this string is ordered after |
| 2981 | * @a __s. Determines the effective length rlen of the strings to |
| 2982 | * compare as the smallest of size() and the length of a string |
| 2983 | * constructed from @a __s. The function then compares the two strings |
| 2984 | * by calling traits::compare(data(),s,rlen). If the result of the |
| 2985 | * comparison is nonzero returns it, otherwise the shorter one is |
| 2986 | * ordered first. |
| 2987 | */ |
| 2988 | int |
| 2989 | compare(const _CharT* __s) const _GLIBCXX_NOEXCEPTnoexcept; |
| 2990 | |
| 2991 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
| 2992 | // 5 String::compare specification questionable |
| 2993 | /** |
| 2994 | * @brief Compare substring to a C string. |
| 2995 | * @param __pos Index of first character of substring. |
| 2996 | * @param __n1 Number of characters in substring. |
| 2997 | * @param __s C string to compare against. |
| 2998 | * @return Integer < 0, 0, or > 0. |
| 2999 | * |
| 3000 | * Form the substring of this string from the @a __n1 |
| 3001 | * characters starting at @a pos. Returns an integer < 0 if |
| 3002 | * the substring is ordered before @a __s, 0 if their values |
| 3003 | * are equivalent, or > 0 if the substring is ordered after @a |
| 3004 | * __s. Determines the effective length rlen of the strings to |
| 3005 | * compare as the smallest of the length of the substring and |
| 3006 | * the length of a string constructed from @a __s. The |
| 3007 | * function then compares the two string by calling |
| 3008 | * traits::compare(substring.data(),__s,rlen). If the result of |
| 3009 | * the comparison is nonzero returns it, otherwise the shorter |
| 3010 | * one is ordered first. |
| 3011 | */ |
| 3012 | int |
| 3013 | compare(size_type __pos, size_type __n1, const _CharT* __s) const; |
| 3014 | |
| 3015 | /** |
| 3016 | * @brief Compare substring against a character %array. |
| 3017 | * @param __pos Index of first character of substring. |
| 3018 | * @param __n1 Number of characters in substring. |
| 3019 | * @param __s character %array to compare against. |
| 3020 | * @param __n2 Number of characters of s. |
| 3021 | * @return Integer < 0, 0, or > 0. |
| 3022 | * |
| 3023 | * Form the substring of this string from the @a __n1 |
| 3024 | * characters starting at @a __pos. Form a string from the |
| 3025 | * first @a __n2 characters of @a __s. Returns an integer < 0 |
| 3026 | * if this substring is ordered before the string from @a __s, |
| 3027 | * 0 if their values are equivalent, or > 0 if this substring |
| 3028 | * is ordered after the string from @a __s. Determines the |
| 3029 | * effective length rlen of the strings to compare as the |
| 3030 | * smallest of the length of the substring and @a __n2. The |
| 3031 | * function then compares the two strings by calling |
| 3032 | * traits::compare(substring.data(),s,rlen). If the result of |
| 3033 | * the comparison is nonzero returns it, otherwise the shorter |
| 3034 | * one is ordered first. |
| 3035 | * |
| 3036 | * NB: s must have at least n2 characters, '\\0' has |
| 3037 | * no special meaning. |
| 3038 | */ |
| 3039 | int |
| 3040 | compare(size_type __pos, size_type __n1, const _CharT* __s, |
| 3041 | size_type __n2) const; |
| 3042 | |
| 3043 | #if __cplusplus201703L > 201703L |
| 3044 | bool |
| 3045 | starts_with(basic_string_view<_CharT, _Traits> __x) const noexcept |
| 3046 | { return __sv_type(this->data(), this->size()).starts_with(__x); } |
| 3047 | |
| 3048 | bool |
| 3049 | starts_with(_CharT __x) const noexcept |
| 3050 | { return __sv_type(this->data(), this->size()).starts_with(__x); } |
| 3051 | |
| 3052 | bool |
| 3053 | starts_with(const _CharT* __x) const noexcept |
| 3054 | { return __sv_type(this->data(), this->size()).starts_with(__x); } |
| 3055 | |
| 3056 | bool |
| 3057 | ends_with(basic_string_view<_CharT, _Traits> __x) const noexcept |
| 3058 | { return __sv_type(this->data(), this->size()).ends_with(__x); } |
| 3059 | |
| 3060 | bool |
| 3061 | ends_with(_CharT __x) const noexcept |
| 3062 | { return __sv_type(this->data(), this->size()).ends_with(__x); } |
| 3063 | |
| 3064 | bool |
| 3065 | ends_with(const _CharT* __x) const noexcept |
| 3066 | { return __sv_type(this->data(), this->size()).ends_with(__x); } |
| 3067 | #endif // C++20 |
| 3068 | |
| 3069 | // Allow basic_stringbuf::__xfer_bufptrs to call _M_length: |
| 3070 | template<typename, typename, typename> friend class basic_stringbuf; |
| 3071 | }; |
| 3072 | _GLIBCXX_END_NAMESPACE_CXX11} |
| 3073 | #else // !_GLIBCXX_USE_CXX11_ABI |
| 3074 | // Reference-counted COW string implentation |
| 3075 | |
| 3076 | /** |
| 3077 | * @class basic_string basic_string.h <string> |
| 3078 | * @brief Managing sequences of characters and character-like objects. |
| 3079 | * |
| 3080 | * @ingroup strings |
| 3081 | * @ingroup sequences |
| 3082 | * |
| 3083 | * @tparam _CharT Type of character |
| 3084 | * @tparam _Traits Traits for character type, defaults to |
| 3085 | * char_traits<_CharT>. |
| 3086 | * @tparam _Alloc Allocator type, defaults to allocator<_CharT>. |
| 3087 | * |
| 3088 | * Meets the requirements of a <a href="tables.html#65">container</a>, a |
| 3089 | * <a href="tables.html#66">reversible container</a>, and a |
| 3090 | * <a href="tables.html#67">sequence</a>. Of the |
| 3091 | * <a href="tables.html#68">optional sequence requirements</a>, only |
| 3092 | * @c push_back, @c at, and @c %array access are supported. |
| 3093 | * |
| 3094 | * @doctodo |
| 3095 | * |
| 3096 | * |
| 3097 | * Documentation? What's that? |
| 3098 | * Nathan Myers <ncm@cantrip.org>. |
| 3099 | * |
| 3100 | * A string looks like this: |
| 3101 | * |
| 3102 | * @code |
| 3103 | * [_Rep] |
| 3104 | * _M_length |
| 3105 | * [basic_string<char_type>] _M_capacity |
| 3106 | * _M_dataplus _M_refcount |
| 3107 | * _M_p ----------------> unnamed array of char_type |
| 3108 | * @endcode |
| 3109 | * |
| 3110 | * Where the _M_p points to the first character in the string, and |
| 3111 | * you cast it to a pointer-to-_Rep and subtract 1 to get a |
| 3112 | * pointer to the header. |
| 3113 | * |
| 3114 | * This approach has the enormous advantage that a string object |
| 3115 | * requires only one allocation. All the ugliness is confined |
| 3116 | * within a single %pair of inline functions, which each compile to |
| 3117 | * a single @a add instruction: _Rep::_M_data(), and |
| 3118 | * string::_M_rep(); and the allocation function which gets a |
| 3119 | * block of raw bytes and with room enough and constructs a _Rep |
| 3120 | * object at the front. |
| 3121 | * |
| 3122 | * The reason you want _M_data pointing to the character %array and |
| 3123 | * not the _Rep is so that the debugger can see the string |
| 3124 | * contents. (Probably we should add a non-inline member to get |
| 3125 | * the _Rep for the debugger to use, so users can check the actual |
| 3126 | * string length.) |
| 3127 | * |
| 3128 | * Note that the _Rep object is a POD so that you can have a |
| 3129 | * static <em>empty string</em> _Rep object already @a constructed before |
| 3130 | * static constructors have run. The reference-count encoding is |
| 3131 | * chosen so that a 0 indicates one reference, so you never try to |
| 3132 | * destroy the empty-string _Rep object. |
| 3133 | * |
| 3134 | * All but the last paragraph is considered pretty conventional |
| 3135 | * for a C++ string implementation. |
| 3136 | */ |
| 3137 | // 21.3 Template class basic_string |
| 3138 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 3139 | class basic_string |
| 3140 | { |
| 3141 | typedef typename __gnu_cxx::__alloc_traits<_Alloc>::template |
| 3142 | rebind<_CharT>::other _CharT_alloc_type; |
| 3143 | typedef __gnu_cxx::__alloc_traits<_CharT_alloc_type> _CharT_alloc_traits; |
| 3144 | |
| 3145 | // Types: |
| 3146 | public: |
| 3147 | typedef _Traits traits_type; |
| 3148 | typedef typename _Traits::char_type value_type; |
| 3149 | typedef _Alloc allocator_type; |
| 3150 | typedef typename _CharT_alloc_type::size_type size_type; |
| 3151 | typedef typename _CharT_alloc_type::difference_type difference_type; |
| 3152 | #if __cplusplus201703L < 201103L |
| 3153 | typedef typename _CharT_alloc_type::reference reference; |
| 3154 | typedef typename _CharT_alloc_type::const_reference const_reference; |
| 3155 | #else |
| 3156 | typedef value_type& reference; |
| 3157 | typedef const value_type& const_reference; |
| 3158 | #endif |
| 3159 | typedef typename _CharT_alloc_traits::pointer pointer; |
| 3160 | typedef typename _CharT_alloc_traits::const_pointer const_pointer; |
| 3161 | typedef __gnu_cxx::__normal_iterator<pointer, basic_string> iterator; |
| 3162 | typedef __gnu_cxx::__normal_iterator<const_pointer, basic_string> |
| 3163 | const_iterator; |
| 3164 | typedef std::reverse_iterator<const_iterator> const_reverse_iterator; |
| 3165 | typedef std::reverse_iterator<iterator> reverse_iterator; |
| 3166 | |
| 3167 | protected: |
| 3168 | // type used for positions in insert, erase etc. |
| 3169 | typedef iterator __const_iterator; |
| 3170 | |
| 3171 | private: |
| 3172 | // _Rep: string representation |
| 3173 | // Invariants: |
| 3174 | // 1. String really contains _M_length + 1 characters: due to 21.3.4 |
| 3175 | // must be kept null-terminated. |
| 3176 | // 2. _M_capacity >= _M_length |
| 3177 | // Allocated memory is always (_M_capacity + 1) * sizeof(_CharT). |
| 3178 | // 3. _M_refcount has three states: |
| 3179 | // -1: leaked, one reference, no ref-copies allowed, non-const. |
| 3180 | // 0: one reference, non-const. |
| 3181 | // n>0: n + 1 references, operations require a lock, const. |
| 3182 | // 4. All fields==0 is an empty string, given the extra storage |
| 3183 | // beyond-the-end for a null terminator; thus, the shared |
| 3184 | // empty string representation needs no constructor. |
| 3185 | |
| 3186 | struct _Rep_base |
| 3187 | { |
| 3188 | size_type _M_length; |
| 3189 | size_type _M_capacity; |
| 3190 | _Atomic_word _M_refcount; |
| 3191 | }; |
| 3192 | |
| 3193 | struct _Rep : _Rep_base |
| 3194 | { |
| 3195 | // Types: |
| 3196 | typedef typename __gnu_cxx::__alloc_traits<_Alloc>::template |
| 3197 | rebind<char>::other _Raw_bytes_alloc; |
| 3198 | |
| 3199 | // (Public) Data members: |
| 3200 | |
| 3201 | // The maximum number of individual char_type elements of an |
| 3202 | // individual string is determined by _S_max_size. This is the |
| 3203 | // value that will be returned by max_size(). (Whereas npos |
| 3204 | // is the maximum number of bytes the allocator can allocate.) |
| 3205 | // If one was to divvy up the theoretical largest size string, |
| 3206 | // with a terminating character and m _CharT elements, it'd |
| 3207 | // look like this: |
| 3208 | // npos = sizeof(_Rep) + (m * sizeof(_CharT)) + sizeof(_CharT) |
| 3209 | // Solving for m: |
| 3210 | // m = ((npos - sizeof(_Rep))/sizeof(CharT)) - 1 |
| 3211 | // In addition, this implementation quarters this amount. |
| 3212 | static const size_type _S_max_size; |
| 3213 | static const _CharT _S_terminal; |
| 3214 | |
| 3215 | // The following storage is init'd to 0 by the linker, resulting |
| 3216 | // (carefully) in an empty string with one reference. |
| 3217 | static size_type _S_empty_rep_storage[]; |
| 3218 | |
| 3219 | static _Rep& |
| 3220 | _S_empty_rep() _GLIBCXX_NOEXCEPTnoexcept |
| 3221 | { |
| 3222 | // NB: Mild hack to avoid strict-aliasing warnings. Note that |
| 3223 | // _S_empty_rep_storage is never modified and the punning should |
| 3224 | // be reasonably safe in this case. |
| 3225 | void* __p = reinterpret_cast<void*>(&_S_empty_rep_storage); |
| 3226 | return *reinterpret_cast<_Rep*>(__p); |
| 3227 | } |
| 3228 | |
| 3229 | bool |
| 3230 | _M_is_leaked() const _GLIBCXX_NOEXCEPTnoexcept |
| 3231 | { |
| 3232 | #if defined(__GTHREADS1) |
| 3233 | // _M_refcount is mutated concurrently by _M_refcopy/_M_dispose, |
| 3234 | // so we need to use an atomic load. However, _M_is_leaked |
| 3235 | // predicate does not change concurrently (i.e. the string is either |
| 3236 | // leaked or not), so a relaxed load is enough. |
| 3237 | return __atomic_load_n(&this->_M_refcount, __ATOMIC_RELAXED0) < 0; |
| 3238 | #else |
| 3239 | return this->_M_refcount < 0; |
| 3240 | #endif |
| 3241 | } |
| 3242 | |
| 3243 | bool |
| 3244 | _M_is_shared() const _GLIBCXX_NOEXCEPTnoexcept |
| 3245 | { |
| 3246 | #if defined(__GTHREADS1) |
| 3247 | // _M_refcount is mutated concurrently by _M_refcopy/_M_dispose, |
| 3248 | // so we need to use an atomic load. Another thread can drop last |
| 3249 | // but one reference concurrently with this check, so we need this |
| 3250 | // load to be acquire to synchronize with release fetch_and_add in |
| 3251 | // _M_dispose. |
| 3252 | return __atomic_load_n(&this->_M_refcount, __ATOMIC_ACQUIRE2) > 0; |
| 3253 | #else |
| 3254 | return this->_M_refcount > 0; |
| 3255 | #endif |
| 3256 | } |
| 3257 | |
| 3258 | void |
| 3259 | _M_set_leaked() _GLIBCXX_NOEXCEPTnoexcept |
| 3260 | { this->_M_refcount = -1; } |
| 3261 | |
| 3262 | void |
| 3263 | _M_set_sharable() _GLIBCXX_NOEXCEPTnoexcept |
| 3264 | { this->_M_refcount = 0; } |
| 3265 | |
| 3266 | void |
| 3267 | _M_set_length_and_sharable(size_type __n) _GLIBCXX_NOEXCEPTnoexcept |
| 3268 | { |
| 3269 | #if _GLIBCXX_FULLY_DYNAMIC_STRING0 == 0 |
| 3270 | if (__builtin_expect(this != &_S_empty_rep(), false)) |
| 3271 | #endif |
| 3272 | { |
| 3273 | this->_M_set_sharable(); // One reference. |
| 3274 | this->_M_length = __n; |
| 3275 | traits_type::assign(this->_M_refdata()[__n], _S_terminal); |
| 3276 | // grrr. (per 21.3.4) |
| 3277 | // You cannot leave those LWG people alone for a second. |
| 3278 | } |
| 3279 | } |
| 3280 | |
| 3281 | _CharT* |
| 3282 | _M_refdata() throw() |
| 3283 | { return reinterpret_cast<_CharT*>(this + 1); } |
| 3284 | |
| 3285 | _CharT* |
| 3286 | _M_grab(const _Alloc& __alloc1, const _Alloc& __alloc2) |
| 3287 | { |
| 3288 | return (!_M_is_leaked() && __alloc1 == __alloc2) |
| 3289 | ? _M_refcopy() : _M_clone(__alloc1); |
| 3290 | } |
| 3291 | |
| 3292 | // Create & Destroy |
| 3293 | static _Rep* |
| 3294 | _S_create(size_type, size_type, const _Alloc&); |
| 3295 | |
| 3296 | void |
| 3297 | _M_dispose(const _Alloc& __a) _GLIBCXX_NOEXCEPTnoexcept |
| 3298 | { |
| 3299 | #if _GLIBCXX_FULLY_DYNAMIC_STRING0 == 0 |
| 3300 | if (__builtin_expect(this != &_S_empty_rep(), false)) |
| 3301 | #endif |
| 3302 | { |
| 3303 | // Be race-detector-friendly. For more info see bits/c++config. |
| 3304 | _GLIBCXX_SYNCHRONIZATION_HAPPENS_BEFORE(&this->_M_refcount); |
| 3305 | // Decrement of _M_refcount is acq_rel, because: |
| 3306 | // - all but last decrements need to release to synchronize with |
| 3307 | // the last decrement that will delete the object. |
| 3308 | // - the last decrement needs to acquire to synchronize with |
| 3309 | // all the previous decrements. |
| 3310 | // - last but one decrement needs to release to synchronize with |
| 3311 | // the acquire load in _M_is_shared that will conclude that |
| 3312 | // the object is not shared anymore. |
| 3313 | if (__gnu_cxx::__exchange_and_add_dispatch(&this->_M_refcount, |
| 3314 | -1) <= 0) |
| 3315 | { |
| 3316 | _GLIBCXX_SYNCHRONIZATION_HAPPENS_AFTER(&this->_M_refcount); |
| 3317 | _M_destroy(__a); |
| 3318 | } |
| 3319 | } |
| 3320 | } // XXX MT |
| 3321 | |
| 3322 | void |
| 3323 | _M_destroy(const _Alloc&) throw(); |
| 3324 | |
| 3325 | _CharT* |
| 3326 | _M_refcopy() throw() |
| 3327 | { |
| 3328 | #if _GLIBCXX_FULLY_DYNAMIC_STRING0 == 0 |
| 3329 | if (__builtin_expect(this != &_S_empty_rep(), false)) |
| 3330 | #endif |
| 3331 | __gnu_cxx::__atomic_add_dispatch(&this->_M_refcount, 1); |
| 3332 | return _M_refdata(); |
| 3333 | } // XXX MT |
| 3334 | |
| 3335 | _CharT* |
| 3336 | _M_clone(const _Alloc&, size_type __res = 0); |
| 3337 | }; |
| 3338 | |
| 3339 | // Use empty-base optimization: http://www.cantrip.org/emptyopt.html |
| 3340 | struct _Alloc_hider : _Alloc |
| 3341 | { |
| 3342 | _Alloc_hider(_CharT* __dat, const _Alloc& __a) _GLIBCXX_NOEXCEPTnoexcept |
| 3343 | : _Alloc(__a), _M_p(__dat) { } |
| 3344 | |
| 3345 | _CharT* _M_p; // The actual data. |
| 3346 | }; |
| 3347 | |
| 3348 | public: |
| 3349 | // Data Members (public): |
| 3350 | // NB: This is an unsigned type, and thus represents the maximum |
| 3351 | // size that the allocator can hold. |
| 3352 | /// Value returned by various member functions when they fail. |
| 3353 | static const size_type npos = static_cast<size_type>(-1); |
| 3354 | |
| 3355 | private: |
| 3356 | // Data Members (private): |
| 3357 | mutable _Alloc_hider _M_dataplus; |
| 3358 | |
| 3359 | _CharT* |
| 3360 | _M_data() const _GLIBCXX_NOEXCEPTnoexcept |
| 3361 | { return _M_dataplus._M_p; } |
| 3362 | |
| 3363 | _CharT* |
| 3364 | _M_data(_CharT* __p) _GLIBCXX_NOEXCEPTnoexcept |
| 3365 | { return (_M_dataplus._M_p = __p); } |
| 3366 | |
| 3367 | _Rep* |
| 3368 | _M_rep() const _GLIBCXX_NOEXCEPTnoexcept |
| 3369 | { return &((reinterpret_cast<_Rep*> (_M_data()))[-1]); } |
| 3370 | |
| 3371 | // For the internal use we have functions similar to `begin'/`end' |
| 3372 | // but they do not call _M_leak. |
| 3373 | iterator |
| 3374 | _M_ibegin() const _GLIBCXX_NOEXCEPTnoexcept |
| 3375 | { return iterator(_M_data()); } |
| 3376 | |
| 3377 | iterator |
| 3378 | _M_iend() const _GLIBCXX_NOEXCEPTnoexcept |
| 3379 | { return iterator(_M_data() + this->size()); } |
| 3380 | |
| 3381 | void |
| 3382 | _M_leak() // for use in begin() & non-const op[] |
| 3383 | { |
| 3384 | if (!_M_rep()->_M_is_leaked()) |
| 3385 | _M_leak_hard(); |
| 3386 | } |
| 3387 | |
| 3388 | size_type |
| 3389 | _M_check(size_type __pos, const char* __s) const |
| 3390 | { |
| 3391 | if (__pos > this->size()) |
| 3392 | __throw_out_of_range_fmt(__N("%s: __pos (which is %zu) > "("%s: __pos (which is %zu) > " "this->size() (which is %zu)" ) |
| 3393 | "this->size() (which is %zu)")("%s: __pos (which is %zu) > " "this->size() (which is %zu)" ), |
| 3394 | __s, __pos, this->size()); |
| 3395 | return __pos; |
| 3396 | } |
| 3397 | |
| 3398 | void |
| 3399 | _M_check_length(size_type __n1, size_type __n2, const char* __s) const |
| 3400 | { |
| 3401 | if (this->max_size() - (this->size() - __n1) < __n2) |
| 3402 | __throw_length_error(__N(__s)(__s)); |
| 3403 | } |
| 3404 | |
| 3405 | // NB: _M_limit doesn't check for a bad __pos value. |
| 3406 | size_type |
| 3407 | _M_limit(size_type __pos, size_type __off) const _GLIBCXX_NOEXCEPTnoexcept |
| 3408 | { |
| 3409 | const bool __testoff = __off < this->size() - __pos; |
| 3410 | return __testoff ? __off : this->size() - __pos; |
| 3411 | } |
| 3412 | |
| 3413 | // True if _Rep and source do not overlap. |
| 3414 | bool |
| 3415 | _M_disjunct(const _CharT* __s) const _GLIBCXX_NOEXCEPTnoexcept |
| 3416 | { |
| 3417 | return (less<const _CharT*>()(__s, _M_data()) |
| 3418 | || less<const _CharT*>()(_M_data() + this->size(), __s)); |
| 3419 | } |
| 3420 | |
| 3421 | // When __n = 1 way faster than the general multichar |
| 3422 | // traits_type::copy/move/assign. |
| 3423 | static void |
| 3424 | _M_copy(_CharT* __d, const _CharT* __s, size_type __n) _GLIBCXX_NOEXCEPTnoexcept |
| 3425 | { |
| 3426 | if (__n == 1) |
| 3427 | traits_type::assign(*__d, *__s); |
| 3428 | else |
| 3429 | traits_type::copy(__d, __s, __n); |
| 3430 | } |
| 3431 | |
| 3432 | static void |
| 3433 | _M_move(_CharT* __d, const _CharT* __s, size_type __n) _GLIBCXX_NOEXCEPTnoexcept |
| 3434 | { |
| 3435 | if (__n == 1) |
| 3436 | traits_type::assign(*__d, *__s); |
| 3437 | else |
| 3438 | traits_type::move(__d, __s, __n); |
| 3439 | } |
| 3440 | |
| 3441 | static void |
| 3442 | _M_assign(_CharT* __d, size_type __n, _CharT __c) _GLIBCXX_NOEXCEPTnoexcept |
| 3443 | { |
| 3444 | if (__n == 1) |
| 3445 | traits_type::assign(*__d, __c); |
| 3446 | else |
| 3447 | traits_type::assign(__d, __n, __c); |
| 3448 | } |
| 3449 | |
| 3450 | // _S_copy_chars is a separate template to permit specialization |
| 3451 | // to optimize for the common case of pointers as iterators. |
| 3452 | template<class _Iterator> |
| 3453 | static void |
| 3454 | _S_copy_chars(_CharT* __p, _Iterator __k1, _Iterator __k2) |
| 3455 | { |
| 3456 | for (; __k1 != __k2; ++__k1, (void)++__p) |
| 3457 | traits_type::assign(*__p, *__k1); // These types are off. |
| 3458 | } |
| 3459 | |
| 3460 | static void |
| 3461 | _S_copy_chars(_CharT* __p, iterator __k1, iterator __k2) _GLIBCXX_NOEXCEPTnoexcept |
| 3462 | { _S_copy_chars(__p, __k1.base(), __k2.base()); } |
| 3463 | |
| 3464 | static void |
| 3465 | _S_copy_chars(_CharT* __p, const_iterator __k1, const_iterator __k2) |
| 3466 | _GLIBCXX_NOEXCEPTnoexcept |
| 3467 | { _S_copy_chars(__p, __k1.base(), __k2.base()); } |
| 3468 | |
| 3469 | static void |
| 3470 | _S_copy_chars(_CharT* __p, _CharT* __k1, _CharT* __k2) _GLIBCXX_NOEXCEPTnoexcept |
| 3471 | { _M_copy(__p, __k1, __k2 - __k1); } |
| 3472 | |
| 3473 | static void |
| 3474 | _S_copy_chars(_CharT* __p, const _CharT* __k1, const _CharT* __k2) |
| 3475 | _GLIBCXX_NOEXCEPTnoexcept |
| 3476 | { _M_copy(__p, __k1, __k2 - __k1); } |
| 3477 | |
| 3478 | static int |
| 3479 | _S_compare(size_type __n1, size_type __n2) _GLIBCXX_NOEXCEPTnoexcept |
| 3480 | { |
| 3481 | const difference_type __d = difference_type(__n1 - __n2); |
| 3482 | |
| 3483 | if (__d > __gnu_cxx::__numeric_traits<int>::__max) |
| 3484 | return __gnu_cxx::__numeric_traits<int>::__max; |
| 3485 | else if (__d < __gnu_cxx::__numeric_traits<int>::__min) |
| 3486 | return __gnu_cxx::__numeric_traits<int>::__min; |
| 3487 | else |
| 3488 | return int(__d); |
| 3489 | } |
| 3490 | |
| 3491 | void |
| 3492 | _M_mutate(size_type __pos, size_type __len1, size_type __len2); |
| 3493 | |
| 3494 | void |
| 3495 | _M_leak_hard(); |
| 3496 | |
| 3497 | static _Rep& |
| 3498 | _S_empty_rep() _GLIBCXX_NOEXCEPTnoexcept |
| 3499 | { return _Rep::_S_empty_rep(); } |
| 3500 | |
| 3501 | #if __cplusplus201703L >= 201703L |
| 3502 | // A helper type for avoiding boiler-plate. |
| 3503 | typedef basic_string_view<_CharT, _Traits> __sv_type; |
| 3504 | |
| 3505 | template<typename _Tp, typename _Res> |
| 3506 | using _If_sv = enable_if_t< |
| 3507 | __and_<is_convertible<const _Tp&, __sv_type>, |
| 3508 | __not_<is_convertible<const _Tp*, const basic_string*>>, |
| 3509 | __not_<is_convertible<const _Tp&, const _CharT*>>>::value, |
| 3510 | _Res>; |
| 3511 | |
| 3512 | // Allows an implicit conversion to __sv_type. |
| 3513 | static __sv_type |
| 3514 | _S_to_string_view(__sv_type __svt) noexcept |
| 3515 | { return __svt; } |
| 3516 | |
| 3517 | // Wraps a string_view by explicit conversion and thus |
| 3518 | // allows to add an internal constructor that does not |
| 3519 | // participate in overload resolution when a string_view |
| 3520 | // is provided. |
| 3521 | struct __sv_wrapper |
| 3522 | { |
| 3523 | explicit __sv_wrapper(__sv_type __sv) noexcept : _M_sv(__sv) { } |
| 3524 | __sv_type _M_sv; |
| 3525 | }; |
| 3526 | |
| 3527 | /** |
| 3528 | * @brief Only internally used: Construct string from a string view |
| 3529 | * wrapper. |
| 3530 | * @param __svw string view wrapper. |
| 3531 | * @param __a Allocator to use. |
| 3532 | */ |
| 3533 | explicit |
| 3534 | basic_string(__sv_wrapper __svw, const _Alloc& __a) |
| 3535 | : basic_string(__svw._M_sv.data(), __svw._M_sv.size(), __a) { } |
| 3536 | #endif |
| 3537 | |
| 3538 | public: |
| 3539 | // Construct/copy/destroy: |
| 3540 | // NB: We overload ctors in some cases instead of using default |
| 3541 | // arguments, per 17.4.4.4 para. 2 item 2. |
| 3542 | |
| 3543 | /** |
| 3544 | * @brief Default constructor creates an empty string. |
| 3545 | */ |
| 3546 | basic_string() |
| 3547 | #if _GLIBCXX_FULLY_DYNAMIC_STRING0 == 0 |
| 3548 | _GLIBCXX_NOEXCEPTnoexcept |
| 3549 | : _M_dataplus(_S_empty_rep()._M_refdata(), _Alloc()) |
| 3550 | #else |
| 3551 | : _M_dataplus(_S_construct(size_type(), _CharT(), _Alloc()), _Alloc()) |
| 3552 | #endif |
| 3553 | { } |
| 3554 | |
| 3555 | /** |
| 3556 | * @brief Construct an empty string using allocator @a a. |
| 3557 | */ |
| 3558 | explicit |
| 3559 | basic_string(const _Alloc& __a); |
| 3560 | |
| 3561 | // NB: per LWG issue 42, semantics different from IS: |
| 3562 | /** |
| 3563 | * @brief Construct string with copy of value of @a str. |
| 3564 | * @param __str Source string. |
| 3565 | */ |
| 3566 | basic_string(const basic_string& __str); |
| 3567 | |
| 3568 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
| 3569 | // 2583. no way to supply an allocator for basic_string(str, pos) |
| 3570 | /** |
| 3571 | * @brief Construct string as copy of a substring. |
| 3572 | * @param __str Source string. |
| 3573 | * @param __pos Index of first character to copy from. |
| 3574 | * @param __a Allocator to use. |
| 3575 | */ |
| 3576 | basic_string(const basic_string& __str, size_type __pos, |
| 3577 | const _Alloc& __a = _Alloc()); |
| 3578 | |
| 3579 | /** |
| 3580 | * @brief Construct string as copy of a substring. |
| 3581 | * @param __str Source string. |
| 3582 | * @param __pos Index of first character to copy from. |
| 3583 | * @param __n Number of characters to copy. |
| 3584 | */ |
| 3585 | basic_string(const basic_string& __str, size_type __pos, |
| 3586 | size_type __n); |
| 3587 | /** |
| 3588 | * @brief Construct string as copy of a substring. |
| 3589 | * @param __str Source string. |
| 3590 | * @param __pos Index of first character to copy from. |
| 3591 | * @param __n Number of characters to copy. |
| 3592 | * @param __a Allocator to use. |
| 3593 | */ |
| 3594 | basic_string(const basic_string& __str, size_type __pos, |
| 3595 | size_type __n, const _Alloc& __a); |
| 3596 | |
| 3597 | /** |
| 3598 | * @brief Construct string initialized by a character %array. |
| 3599 | * @param __s Source character %array. |
| 3600 | * @param __n Number of characters to copy. |
| 3601 | * @param __a Allocator to use (default is default allocator). |
| 3602 | * |
| 3603 | * NB: @a __s must have at least @a __n characters, '\\0' |
| 3604 | * has no special meaning. |
| 3605 | */ |
| 3606 | basic_string(const _CharT* __s, size_type __n, |
| 3607 | const _Alloc& __a = _Alloc()); |
| 3608 | |
| 3609 | /** |
| 3610 | * @brief Construct string as copy of a C string. |
| 3611 | * @param __s Source C string. |
| 3612 | * @param __a Allocator to use (default is default allocator). |
| 3613 | */ |
| 3614 | #if __cpp_deduction_guides201703L && ! defined _GLIBCXX_DEFINING_STRING_INSTANTIATIONS |
| 3615 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
| 3616 | // 3076. basic_string CTAD ambiguity |
| 3617 | template<typename = _RequireAllocator<_Alloc>> |
| 3618 | #endif |
| 3619 | basic_string(const _CharT* __s, const _Alloc& __a = _Alloc()) |
| 3620 | : _M_dataplus(_S_construct(__s, __s ? __s + traits_type::length(__s) : |
| 3621 | __s + npos, __a), __a) |
| 3622 | { } |
| 3623 | |
| 3624 | /** |
| 3625 | * @brief Construct string as multiple characters. |
| 3626 | * @param __n Number of characters. |
| 3627 | * @param __c Character to use. |
| 3628 | * @param __a Allocator to use (default is default allocator). |
| 3629 | */ |
| 3630 | basic_string(size_type __n, _CharT __c, const _Alloc& __a = _Alloc()); |
| 3631 | |
| 3632 | #if __cplusplus201703L >= 201103L |
| 3633 | /** |
| 3634 | * @brief Move construct string. |
| 3635 | * @param __str Source string. |
| 3636 | * |
| 3637 | * The newly-created string contains the exact contents of @a __str. |
| 3638 | * @a __str is a valid, but unspecified string. |
| 3639 | **/ |
| 3640 | basic_string(basic_string&& __str) |
| 3641 | #if _GLIBCXX_FULLY_DYNAMIC_STRING0 == 0 |
| 3642 | noexcept // FIXME C++11: should always be noexcept. |
| 3643 | #endif |
| 3644 | : _M_dataplus(std::move(__str._M_dataplus)) |
| 3645 | { |
| 3646 | #if _GLIBCXX_FULLY_DYNAMIC_STRING0 == 0 |
| 3647 | __str._M_data(_S_empty_rep()._M_refdata()); |
| 3648 | #else |
| 3649 | __str._M_data(_S_construct(size_type(), _CharT(), get_allocator())); |
| 3650 | #endif |
| 3651 | } |
| 3652 | |
| 3653 | /** |
| 3654 | * @brief Construct string from an initializer %list. |
| 3655 | * @param __l std::initializer_list of characters. |
| 3656 | * @param __a Allocator to use (default is default allocator). |
| 3657 | */ |
| 3658 | basic_string(initializer_list<_CharT> __l, const _Alloc& __a = _Alloc()); |
| 3659 | |
| 3660 | basic_string(const basic_string& __str, const _Alloc& __a) |
| 3661 | : _M_dataplus(__str._M_rep()->_M_grab(__a, __str.get_allocator()), __a) |
| 3662 | { } |
| 3663 | |
| 3664 | basic_string(basic_string&& __str, const _Alloc& __a) |
| 3665 | : _M_dataplus(__str._M_data(), __a) |
| 3666 | { |
| 3667 | if (__a == __str.get_allocator()) |
| 3668 | { |
| 3669 | #if _GLIBCXX_FULLY_DYNAMIC_STRING0 == 0 |
| 3670 | __str._M_data(_S_empty_rep()._M_refdata()); |
| 3671 | #else |
| 3672 | __str._M_data(_S_construct(size_type(), _CharT(), __a)); |
| 3673 | #endif |
| 3674 | } |
| 3675 | else |
| 3676 | _M_dataplus._M_p = _S_construct(__str.begin(), __str.end(), __a); |
| 3677 | } |
| 3678 | #endif // C++11 |
| 3679 | |
| 3680 | /** |
| 3681 | * @brief Construct string as copy of a range. |
| 3682 | * @param __beg Start of range. |
| 3683 | * @param __end End of range. |
| 3684 | * @param __a Allocator to use (default is default allocator). |
| 3685 | */ |
| 3686 | template<class _InputIterator> |
| 3687 | basic_string(_InputIterator __beg, _InputIterator __end, |
| 3688 | const _Alloc& __a = _Alloc()); |
| 3689 | |
| 3690 | #if __cplusplus201703L >= 201703L |
| 3691 | /** |
| 3692 | * @brief Construct string from a substring of a string_view. |
| 3693 | * @param __t Source object convertible to string view. |
| 3694 | * @param __pos The index of the first character to copy from __t. |
| 3695 | * @param __n The number of characters to copy from __t. |
| 3696 | * @param __a Allocator to use. |
| 3697 | */ |
| 3698 | template<typename _Tp, typename = _If_sv<_Tp, void>> |
| 3699 | basic_string(const _Tp& __t, size_type __pos, size_type __n, |
| 3700 | const _Alloc& __a = _Alloc()) |
| 3701 | : basic_string(_S_to_string_view(__t).substr(__pos, __n), __a) { } |
| 3702 | |
| 3703 | /** |
| 3704 | * @brief Construct string from a string_view. |
| 3705 | * @param __t Source object convertible to string view. |
| 3706 | * @param __a Allocator to use (default is default allocator). |
| 3707 | */ |
| 3708 | template<typename _Tp, typename = _If_sv<_Tp, void>> |
| 3709 | explicit |
| 3710 | basic_string(const _Tp& __t, const _Alloc& __a = _Alloc()) |
| 3711 | : basic_string(__sv_wrapper(_S_to_string_view(__t)), __a) { } |
| 3712 | #endif // C++17 |
| 3713 | |
| 3714 | /** |
| 3715 | * @brief Destroy the string instance. |
| 3716 | */ |
| 3717 | ~basic_string() _GLIBCXX_NOEXCEPTnoexcept |
| 3718 | { _M_rep()->_M_dispose(this->get_allocator()); } |
| 3719 | |
| 3720 | /** |
| 3721 | * @brief Assign the value of @a str to this string. |
| 3722 | * @param __str Source string. |
| 3723 | */ |
| 3724 | basic_string& |
| 3725 | operator=(const basic_string& __str) |
| 3726 | { return this->assign(__str); } |
| 3727 | |
| 3728 | /** |
| 3729 | * @brief Copy contents of @a s into this string. |
| 3730 | * @param __s Source null-terminated string. |
| 3731 | */ |
| 3732 | basic_string& |
| 3733 | operator=(const _CharT* __s) |
| 3734 | { return this->assign(__s); } |
| 3735 | |
| 3736 | /** |
| 3737 | * @brief Set value to string of length 1. |
| 3738 | * @param __c Source character. |
| 3739 | * |
| 3740 | * Assigning to a character makes this string length 1 and |
| 3741 | * (*this)[0] == @a c. |
| 3742 | */ |
| 3743 | basic_string& |
| 3744 | operator=(_CharT __c) |
| 3745 | { |
| 3746 | this->assign(1, __c); |
| 3747 | return *this; |
| 3748 | } |
| 3749 | |
| 3750 | #if __cplusplus201703L >= 201103L |
| 3751 | /** |
| 3752 | * @brief Move assign the value of @a str to this string. |
| 3753 | * @param __str Source string. |
| 3754 | * |
| 3755 | * The contents of @a str are moved into this string (without copying). |
| 3756 | * @a str is a valid, but unspecified string. |
| 3757 | **/ |
| 3758 | basic_string& |
| 3759 | operator=(basic_string&& __str) |
| 3760 | _GLIBCXX_NOEXCEPT_IF(allocator_traits<_Alloc>::is_always_equal::value)noexcept(allocator_traits<_Alloc>::is_always_equal::value ) |
| 3761 | { |
| 3762 | // NB: DR 1204. |
| 3763 | this->swap(__str); |
| 3764 | return *this; |
| 3765 | } |
| 3766 | |
| 3767 | /** |
| 3768 | * @brief Set value to string constructed from initializer %list. |
| 3769 | * @param __l std::initializer_list. |
| 3770 | */ |
| 3771 | basic_string& |
| 3772 | operator=(initializer_list<_CharT> __l) |
| 3773 | { |
| 3774 | this->assign(__l.begin(), __l.size()); |
| 3775 | return *this; |
| 3776 | } |
| 3777 | #endif // C++11 |
| 3778 | |
| 3779 | #if __cplusplus201703L >= 201703L |
| 3780 | /** |
| 3781 | * @brief Set value to string constructed from a string_view. |
| 3782 | * @param __svt An object convertible to string_view. |
| 3783 | */ |
| 3784 | template<typename _Tp> |
| 3785 | _If_sv<_Tp, basic_string&> |
| 3786 | operator=(const _Tp& __svt) |
| 3787 | { return this->assign(__svt); } |
| 3788 | |
| 3789 | /** |
| 3790 | * @brief Convert to a string_view. |
| 3791 | * @return A string_view. |
| 3792 | */ |
| 3793 | operator __sv_type() const noexcept |
| 3794 | { return __sv_type(data(), size()); } |
| 3795 | #endif // C++17 |
| 3796 | |
| 3797 | // Iterators: |
| 3798 | /** |
| 3799 | * Returns a read/write iterator that points to the first character in |
| 3800 | * the %string. Unshares the string. |
| 3801 | */ |
| 3802 | iterator |
| 3803 | begin() // FIXME C++11: should be noexcept. |
| 3804 | { |
| 3805 | _M_leak(); |
| 3806 | return iterator(_M_data()); |
| 3807 | } |
| 3808 | |
| 3809 | /** |
| 3810 | * Returns a read-only (constant) iterator that points to the first |
| 3811 | * character in the %string. |
| 3812 | */ |
| 3813 | const_iterator |
| 3814 | begin() const _GLIBCXX_NOEXCEPTnoexcept |
| 3815 | { return const_iterator(_M_data()); } |
| 3816 | |
| 3817 | /** |
| 3818 | * Returns a read/write iterator that points one past the last |
| 3819 | * character in the %string. Unshares the string. |
| 3820 | */ |
| 3821 | iterator |
| 3822 | end() // FIXME C++11: should be noexcept. |
| 3823 | { |
| 3824 | _M_leak(); |
| 3825 | return iterator(_M_data() + this->size()); |
| 3826 | } |
| 3827 | |
| 3828 | /** |
| 3829 | * Returns a read-only (constant) iterator that points one past the |
| 3830 | * last character in the %string. |
| 3831 | */ |
| 3832 | const_iterator |
| 3833 | end() const _GLIBCXX_NOEXCEPTnoexcept |
| 3834 | { return const_iterator(_M_data() + this->size()); } |
| 3835 | |
| 3836 | /** |
| 3837 | * Returns a read/write reverse iterator that points to the last |
| 3838 | * character in the %string. Iteration is done in reverse element |
| 3839 | * order. Unshares the string. |
| 3840 | */ |
| 3841 | reverse_iterator |
| 3842 | rbegin() // FIXME C++11: should be noexcept. |
| 3843 | { return reverse_iterator(this->end()); } |
| 3844 | |
| 3845 | /** |
| 3846 | * Returns a read-only (constant) reverse iterator that points |
| 3847 | * to the last character in the %string. Iteration is done in |
| 3848 | * reverse element order. |
| 3849 | */ |
| 3850 | const_reverse_iterator |
| 3851 | rbegin() const _GLIBCXX_NOEXCEPTnoexcept |
| 3852 | { return const_reverse_iterator(this->end()); } |
| 3853 | |
| 3854 | /** |
| 3855 | * Returns a read/write reverse iterator that points to one before the |
| 3856 | * first character in the %string. Iteration is done in reverse |
| 3857 | * element order. Unshares the string. |
| 3858 | */ |
| 3859 | reverse_iterator |
| 3860 | rend() // FIXME C++11: should be noexcept. |
| 3861 | { return reverse_iterator(this->begin()); } |
| 3862 | |
| 3863 | /** |
| 3864 | * Returns a read-only (constant) reverse iterator that points |
| 3865 | * to one before the first character in the %string. Iteration |
| 3866 | * is done in reverse element order. |
| 3867 | */ |
| 3868 | const_reverse_iterator |
| 3869 | rend() const _GLIBCXX_NOEXCEPTnoexcept |
| 3870 | { return const_reverse_iterator(this->begin()); } |
| 3871 | |
| 3872 | #if __cplusplus201703L >= 201103L |
| 3873 | /** |
| 3874 | * Returns a read-only (constant) iterator that points to the first |
| 3875 | * character in the %string. |
| 3876 | */ |
| 3877 | const_iterator |
| 3878 | cbegin() const noexcept |
| 3879 | { return const_iterator(this->_M_data()); } |
| 3880 | |
| 3881 | /** |
| 3882 | * Returns a read-only (constant) iterator that points one past the |
| 3883 | * last character in the %string. |
| 3884 | */ |
| 3885 | const_iterator |
| 3886 | cend() const noexcept |
| 3887 | { return const_iterator(this->_M_data() + this->size()); } |
| 3888 | |
| 3889 | /** |
| 3890 | * Returns a read-only (constant) reverse iterator that points |
| 3891 | * to the last character in the %string. Iteration is done in |
| 3892 | * reverse element order. |
| 3893 | */ |
| 3894 | const_reverse_iterator |
| 3895 | crbegin() const noexcept |
| 3896 | { return const_reverse_iterator(this->end()); } |
| 3897 | |
| 3898 | /** |
| 3899 | * Returns a read-only (constant) reverse iterator that points |
| 3900 | * to one before the first character in the %string. Iteration |
| 3901 | * is done in reverse element order. |
| 3902 | */ |
| 3903 | const_reverse_iterator |
| 3904 | crend() const noexcept |
| 3905 | { return const_reverse_iterator(this->begin()); } |
| 3906 | #endif |
| 3907 | |
| 3908 | public: |
| 3909 | // Capacity: |
| 3910 | /// Returns the number of characters in the string, not including any |
| 3911 | /// null-termination. |
| 3912 | size_type |
| 3913 | size() const _GLIBCXX_NOEXCEPTnoexcept |
| 3914 | { return _M_rep()->_M_length; } |
| 3915 | |
| 3916 | /// Returns the number of characters in the string, not including any |
| 3917 | /// null-termination. |
| 3918 | size_type |
| 3919 | length() const _GLIBCXX_NOEXCEPTnoexcept |
| 3920 | { return _M_rep()->_M_length; } |
| 3921 | |
| 3922 | /// Returns the size() of the largest possible %string. |
| 3923 | size_type |
| 3924 | max_size() const _GLIBCXX_NOEXCEPTnoexcept |
| 3925 | { return _Rep::_S_max_size; } |
| 3926 | |
| 3927 | /** |
| 3928 | * @brief Resizes the %string to the specified number of characters. |
| 3929 | * @param __n Number of characters the %string should contain. |
| 3930 | * @param __c Character to fill any new elements. |
| 3931 | * |
| 3932 | * This function will %resize the %string to the specified |
| 3933 | * number of characters. If the number is smaller than the |
| 3934 | * %string's current size the %string is truncated, otherwise |
| 3935 | * the %string is extended and new elements are %set to @a __c. |
| 3936 | */ |
| 3937 | void |
| 3938 | resize(size_type __n, _CharT __c); |
| 3939 | |
| 3940 | /** |
| 3941 | * @brief Resizes the %string to the specified number of characters. |
| 3942 | * @param __n Number of characters the %string should contain. |
| 3943 | * |
| 3944 | * This function will resize the %string to the specified length. If |
| 3945 | * the new size is smaller than the %string's current size the %string |
| 3946 | * is truncated, otherwise the %string is extended and new characters |
| 3947 | * are default-constructed. For basic types such as char, this means |
| 3948 | * setting them to 0. |
| 3949 | */ |
| 3950 | void |
| 3951 | resize(size_type __n) |
| 3952 | { this->resize(__n, _CharT()); } |
| 3953 | |
| 3954 | #if __cplusplus201703L >= 201103L |
| 3955 | /// A non-binding request to reduce capacity() to size(). |
| 3956 | void |
| 3957 | shrink_to_fit() _GLIBCXX_NOEXCEPTnoexcept |
| 3958 | { |
| 3959 | #if __cpp_exceptions |
| 3960 | if (capacity() > size()) |
| 3961 | { |
| 3962 | try |
| 3963 | { reserve(0); } |
| 3964 | catch(...) |
| 3965 | { } |
| 3966 | } |
| 3967 | #endif |
| 3968 | } |
| 3969 | #endif |
| 3970 | |
| 3971 | /** |
| 3972 | * Returns the total number of characters that the %string can hold |
| 3973 | * before needing to allocate more memory. |
| 3974 | */ |
| 3975 | size_type |
| 3976 | capacity() const _GLIBCXX_NOEXCEPTnoexcept |
| 3977 | { return _M_rep()->_M_capacity; } |
| 3978 | |
| 3979 | /** |
| 3980 | * @brief Attempt to preallocate enough memory for specified number of |
| 3981 | * characters. |
| 3982 | * @param __res_arg Number of characters required. |
| 3983 | * @throw std::length_error If @a __res_arg exceeds @c max_size(). |
| 3984 | * |
| 3985 | * This function attempts to reserve enough memory for the |
| 3986 | * %string to hold the specified number of characters. If the |
| 3987 | * number requested is more than max_size(), length_error is |
| 3988 | * thrown. |
| 3989 | * |
| 3990 | * The advantage of this function is that if optimal code is a |
| 3991 | * necessity and the user can determine the string length that will be |
| 3992 | * required, the user can reserve the memory in %advance, and thus |
| 3993 | * prevent a possible reallocation of memory and copying of %string |
| 3994 | * data. |
| 3995 | */ |
| 3996 | void |
| 3997 | reserve(size_type __res_arg = 0); |
| 3998 | |
| 3999 | /** |
| 4000 | * Erases the string, making it empty. |
| 4001 | */ |
| 4002 | #if _GLIBCXX_FULLY_DYNAMIC_STRING0 == 0 |
| 4003 | void |
| 4004 | clear() _GLIBCXX_NOEXCEPTnoexcept |
| 4005 | { |
| 4006 | if (_M_rep()->_M_is_shared()) |
| 4007 | { |
| 4008 | _M_rep()->_M_dispose(this->get_allocator()); |
| 4009 | _M_data(_S_empty_rep()._M_refdata()); |
| 4010 | } |
| 4011 | else |
| 4012 | _M_rep()->_M_set_length_and_sharable(0); |
| 4013 | } |
| 4014 | #else |
| 4015 | // PR 56166: this should not throw. |
| 4016 | void |
| 4017 | clear() |
| 4018 | { _M_mutate(0, this->size(), 0); } |
| 4019 | #endif |
| 4020 | |
| 4021 | /** |
| 4022 | * Returns true if the %string is empty. Equivalent to |
| 4023 | * <code>*this == ""</code>. |
| 4024 | */ |
| 4025 | _GLIBCXX_NODISCARD[[__nodiscard__]] bool |
| 4026 | empty() const _GLIBCXX_NOEXCEPTnoexcept |
| 4027 | { return this->size() == 0; } |
| 4028 | |
| 4029 | // Element access: |
| 4030 | /** |
| 4031 | * @brief Subscript access to the data contained in the %string. |
| 4032 | * @param __pos The index of the character to access. |
| 4033 | * @return Read-only (constant) reference to the character. |
| 4034 | * |
| 4035 | * This operator allows for easy, array-style, data access. |
| 4036 | * Note that data access with this operator is unchecked and |
| 4037 | * out_of_range lookups are not defined. (For checked lookups |
| 4038 | * see at().) |
| 4039 | */ |
| 4040 | const_reference |
| 4041 | operator[] (size_type __pos) const _GLIBCXX_NOEXCEPTnoexcept |
| 4042 | { |
| 4043 | __glibcxx_assert(__pos <= size()); |
| 4044 | return _M_data()[__pos]; |
| 4045 | } |
| 4046 | |
| 4047 | /** |
| 4048 | * @brief Subscript access to the data contained in the %string. |
| 4049 | * @param __pos The index of the character to access. |
| 4050 | * @return Read/write reference to the character. |
| 4051 | * |
| 4052 | * This operator allows for easy, array-style, data access. |
| 4053 | * Note that data access with this operator is unchecked and |
| 4054 | * out_of_range lookups are not defined. (For checked lookups |
| 4055 | * see at().) Unshares the string. |
| 4056 | */ |
| 4057 | reference |
| 4058 | operator[](size_type __pos) |
| 4059 | { |
| 4060 | // Allow pos == size() both in C++98 mode, as v3 extension, |
| 4061 | // and in C++11 mode. |
| 4062 | __glibcxx_assert(__pos <= size()); |
| 4063 | // In pedantic mode be strict in C++98 mode. |
| 4064 | _GLIBCXX_DEBUG_PEDASSERT(__cplusplus >= 201103L || __pos < size()); |
| 4065 | _M_leak(); |
| 4066 | return _M_data()[__pos]; |
| 4067 | } |
| 4068 | |
| 4069 | /** |
| 4070 | * @brief Provides access to the data contained in the %string. |
| 4071 | * @param __n The index of the character to access. |
| 4072 | * @return Read-only (const) reference to the character. |
| 4073 | * @throw std::out_of_range If @a n is an invalid index. |
| 4074 | * |
| 4075 | * This function provides for safer data access. The parameter is |
| 4076 | * first checked that it is in the range of the string. The function |
| 4077 | * throws out_of_range if the check fails. |
| 4078 | */ |
| 4079 | const_reference |
| 4080 | at(size_type __n) const |
| 4081 | { |
| 4082 | if (__n >= this->size()) |
| 4083 | __throw_out_of_range_fmt(__N("basic_string::at: __n "("basic_string::at: __n " "(which is %zu) >= this->size() " "(which is %zu)") |
| 4084 | "(which is %zu) >= this->size() "("basic_string::at: __n " "(which is %zu) >= this->size() " "(which is %zu)") |
| 4085 | "(which is %zu)")("basic_string::at: __n " "(which is %zu) >= this->size() " "(which is %zu)"), |
| 4086 | __n, this->size()); |
| 4087 | return _M_data()[__n]; |
| 4088 | } |
| 4089 | |
| 4090 | /** |
| 4091 | * @brief Provides access to the data contained in the %string. |
| 4092 | * @param __n The index of the character to access. |
| 4093 | * @return Read/write reference to the character. |
| 4094 | * @throw std::out_of_range If @a n is an invalid index. |
| 4095 | * |
| 4096 | * This function provides for safer data access. The parameter is |
| 4097 | * first checked that it is in the range of the string. The function |
| 4098 | * throws out_of_range if the check fails. Success results in |
| 4099 | * unsharing the string. |
| 4100 | */ |
| 4101 | reference |
| 4102 | at(size_type __n) |
| 4103 | { |
| 4104 | if (__n >= size()) |
| 4105 | __throw_out_of_range_fmt(__N("basic_string::at: __n "("basic_string::at: __n " "(which is %zu) >= this->size() " "(which is %zu)") |
| 4106 | "(which is %zu) >= this->size() "("basic_string::at: __n " "(which is %zu) >= this->size() " "(which is %zu)") |
| 4107 | "(which is %zu)")("basic_string::at: __n " "(which is %zu) >= this->size() " "(which is %zu)"), |
| 4108 | __n, this->size()); |
| 4109 | _M_leak(); |
| 4110 | return _M_data()[__n]; |
| 4111 | } |
| 4112 | |
| 4113 | #if __cplusplus201703L >= 201103L |
| 4114 | /** |
| 4115 | * Returns a read/write reference to the data at the first |
| 4116 | * element of the %string. |
| 4117 | */ |
| 4118 | reference |
| 4119 | front() |
| 4120 | { |
| 4121 | __glibcxx_assert(!empty()); |
| 4122 | return operator[](0); |
| 4123 | } |
| 4124 | |
| 4125 | /** |
| 4126 | * Returns a read-only (constant) reference to the data at the first |
| 4127 | * element of the %string. |
| 4128 | */ |
| 4129 | const_reference |
| 4130 | front() const noexcept |
| 4131 | { |
| 4132 | __glibcxx_assert(!empty()); |
| 4133 | return operator[](0); |
| 4134 | } |
| 4135 | |
| 4136 | /** |
| 4137 | * Returns a read/write reference to the data at the last |
| 4138 | * element of the %string. |
| 4139 | */ |
| 4140 | reference |
| 4141 | back() |
| 4142 | { |
| 4143 | __glibcxx_assert(!empty()); |
| 4144 | return operator[](this->size() - 1); |
| 4145 | } |
| 4146 | |
| 4147 | /** |
| 4148 | * Returns a read-only (constant) reference to the data at the |
| 4149 | * last element of the %string. |
| 4150 | */ |
| 4151 | const_reference |
| 4152 | back() const noexcept |
| 4153 | { |
| 4154 | __glibcxx_assert(!empty()); |
| 4155 | return operator[](this->size() - 1); |
| 4156 | } |
| 4157 | #endif |
| 4158 | |
| 4159 | // Modifiers: |
| 4160 | /** |
| 4161 | * @brief Append a string to this string. |
| 4162 | * @param __str The string to append. |
| 4163 | * @return Reference to this string. |
| 4164 | */ |
| 4165 | basic_string& |
| 4166 | operator+=(const basic_string& __str) |
| 4167 | { return this->append(__str); } |
| 4168 | |
| 4169 | /** |
| 4170 | * @brief Append a C string. |
| 4171 | * @param __s The C string to append. |
| 4172 | * @return Reference to this string. |
| 4173 | */ |
| 4174 | basic_string& |
| 4175 | operator+=(const _CharT* __s) |
| 4176 | { return this->append(__s); } |
| 4177 | |
| 4178 | /** |
| 4179 | * @brief Append a character. |
| 4180 | * @param __c The character to append. |
| 4181 | * @return Reference to this string. |
| 4182 | */ |
| 4183 | basic_string& |
| 4184 | operator+=(_CharT __c) |
| 4185 | { |
| 4186 | this->push_back(__c); |
| 4187 | return *this; |
| 4188 | } |
| 4189 | |
| 4190 | #if __cplusplus201703L >= 201103L |
| 4191 | /** |
| 4192 | * @brief Append an initializer_list of characters. |
| 4193 | * @param __l The initializer_list of characters to be appended. |
| 4194 | * @return Reference to this string. |
| 4195 | */ |
| 4196 | basic_string& |
| 4197 | operator+=(initializer_list<_CharT> __l) |
| 4198 | { return this->append(__l.begin(), __l.size()); } |
| 4199 | #endif // C++11 |
| 4200 | |
| 4201 | #if __cplusplus201703L >= 201703L |
| 4202 | /** |
| 4203 | * @brief Append a string_view. |
| 4204 | * @param __svt The object convertible to string_view to be appended. |
| 4205 | * @return Reference to this string. |
| 4206 | */ |
| 4207 | template<typename _Tp> |
| 4208 | _If_sv<_Tp, basic_string&> |
| 4209 | operator+=(const _Tp& __svt) |
| 4210 | { return this->append(__svt); } |
| 4211 | #endif // C++17 |
| 4212 | |
| 4213 | /** |
| 4214 | * @brief Append a string to this string. |
| 4215 | * @param __str The string to append. |
| 4216 | * @return Reference to this string. |
| 4217 | */ |
| 4218 | basic_string& |
| 4219 | append(const basic_string& __str); |
| 4220 | |
| 4221 | /** |
| 4222 | * @brief Append a substring. |
| 4223 | * @param __str The string to append. |
| 4224 | * @param __pos Index of the first character of str to append. |
| 4225 | * @param __n The number of characters to append. |
| 4226 | * @return Reference to this string. |
| 4227 | * @throw std::out_of_range if @a __pos is not a valid index. |
| 4228 | * |
| 4229 | * This function appends @a __n characters from @a __str |
| 4230 | * starting at @a __pos to this string. If @a __n is is larger |
| 4231 | * than the number of available characters in @a __str, the |
| 4232 | * remainder of @a __str is appended. |
| 4233 | */ |
| 4234 | basic_string& |
| 4235 | append(const basic_string& __str, size_type __pos, size_type __n = npos); |
| 4236 | |
| 4237 | /** |
| 4238 | * @brief Append a C substring. |
| 4239 | * @param __s The C string to append. |
| 4240 | * @param __n The number of characters to append. |
| 4241 | * @return Reference to this string. |
| 4242 | */ |
| 4243 | basic_string& |
| 4244 | append(const _CharT* __s, size_type __n); |
| 4245 | |
| 4246 | /** |
| 4247 | * @brief Append a C string. |
| 4248 | * @param __s The C string to append. |
| 4249 | * @return Reference to this string. |
| 4250 | */ |
| 4251 | basic_string& |
| 4252 | append(const _CharT* __s) |
| 4253 | { |
| 4254 | __glibcxx_requires_string(__s); |
| 4255 | return this->append(__s, traits_type::length(__s)); |
| 4256 | } |
| 4257 | |
| 4258 | /** |
| 4259 | * @brief Append multiple characters. |
| 4260 | * @param __n The number of characters to append. |
| 4261 | * @param __c The character to use. |
| 4262 | * @return Reference to this string. |
| 4263 | * |
| 4264 | * Appends __n copies of __c to this string. |
| 4265 | */ |
| 4266 | basic_string& |
| 4267 | append(size_type __n, _CharT __c); |
| 4268 | |
| 4269 | #if __cplusplus201703L >= 201103L |
| 4270 | /** |
| 4271 | * @brief Append an initializer_list of characters. |
| 4272 | * @param __l The initializer_list of characters to append. |
| 4273 | * @return Reference to this string. |
| 4274 | */ |
| 4275 | basic_string& |
| 4276 | append(initializer_list<_CharT> __l) |
| 4277 | { return this->append(__l.begin(), __l.size()); } |
| 4278 | #endif // C++11 |
| 4279 | |
| 4280 | /** |
| 4281 | * @brief Append a range of characters. |
| 4282 | * @param __first Iterator referencing the first character to append. |
| 4283 | * @param __last Iterator marking the end of the range. |
| 4284 | * @return Reference to this string. |
| 4285 | * |
| 4286 | * Appends characters in the range [__first,__last) to this string. |
| 4287 | */ |
| 4288 | template<class _InputIterator> |
| 4289 | basic_string& |
| 4290 | append(_InputIterator __first, _InputIterator __last) |
| 4291 | { return this->replace(_M_iend(), _M_iend(), __first, __last); } |
| 4292 | |
| 4293 | #if __cplusplus201703L >= 201703L |
| 4294 | /** |
| 4295 | * @brief Append a string_view. |
| 4296 | * @param __svt The object convertible to string_view to be appended. |
| 4297 | * @return Reference to this string. |
| 4298 | */ |
| 4299 | template<typename _Tp> |
| 4300 | _If_sv<_Tp, basic_string&> |
| 4301 | append(const _Tp& __svt) |
| 4302 | { |
| 4303 | __sv_type __sv = __svt; |
| 4304 | return this->append(__sv.data(), __sv.size()); |
| 4305 | } |
| 4306 | |
| 4307 | /** |
| 4308 | * @brief Append a range of characters from a string_view. |
| 4309 | * @param __svt The object convertible to string_view to be appended |
| 4310 | * from. |
| 4311 | * @param __pos The position in the string_view to append from. |
| 4312 | * @param __n The number of characters to append from the string_view. |
| 4313 | * @return Reference to this string. |
| 4314 | */ |
| 4315 | template<typename _Tp> |
| 4316 | _If_sv<_Tp, basic_string&> |
| 4317 | append(const _Tp& __svt, size_type __pos, size_type __n = npos) |
| 4318 | { |
| 4319 | __sv_type __sv = __svt; |
| 4320 | return append(__sv.data() |
| 4321 | + std::__sv_check(__sv.size(), __pos, "basic_string::append"), |
| 4322 | std::__sv_limit(__sv.size(), __pos, __n)); |
| 4323 | } |
| 4324 | #endif // C++17 |
| 4325 | |
| 4326 | /** |
| 4327 | * @brief Append a single character. |
| 4328 | * @param __c Character to append. |
| 4329 | */ |
| 4330 | void |
| 4331 | push_back(_CharT __c) |
| 4332 | { |
| 4333 | const size_type __len = 1 + this->size(); |
| 4334 | if (__len > this->capacity() || _M_rep()->_M_is_shared()) |
| 4335 | this->reserve(__len); |
| 4336 | traits_type::assign(_M_data()[this->size()], __c); |
| 4337 | _M_rep()->_M_set_length_and_sharable(__len); |
| 4338 | } |
| 4339 | |
| 4340 | /** |
| 4341 | * @brief Set value to contents of another string. |
| 4342 | * @param __str Source string to use. |
| 4343 | * @return Reference to this string. |
| 4344 | */ |
| 4345 | basic_string& |
| 4346 | assign(const basic_string& __str); |
| 4347 | |
| 4348 | #if __cplusplus201703L >= 201103L |
| 4349 | /** |
| 4350 | * @brief Set value to contents of another string. |
| 4351 | * @param __str Source string to use. |
| 4352 | * @return Reference to this string. |
| 4353 | * |
| 4354 | * This function sets this string to the exact contents of @a __str. |
| 4355 | * @a __str is a valid, but unspecified string. |
| 4356 | */ |
| 4357 | basic_string& |
| 4358 | assign(basic_string&& __str) |
| 4359 | noexcept(allocator_traits<_Alloc>::is_always_equal::value) |
| 4360 | { |
| 4361 | this->swap(__str); |
| 4362 | return *this; |
| 4363 | } |
| 4364 | #endif // C++11 |
| 4365 | |
| 4366 | /** |
| 4367 | * @brief Set value to a substring of a string. |
| 4368 | * @param __str The string to use. |
| 4369 | * @param __pos Index of the first character of str. |
| 4370 | * @param __n Number of characters to use. |
| 4371 | * @return Reference to this string. |
| 4372 | * @throw std::out_of_range if @a pos is not a valid index. |
| 4373 | * |
| 4374 | * This function sets this string to the substring of @a __str |
| 4375 | * consisting of @a __n characters at @a __pos. If @a __n is |
| 4376 | * is larger than the number of available characters in @a |
| 4377 | * __str, the remainder of @a __str is used. |
| 4378 | */ |
| 4379 | basic_string& |
| 4380 | assign(const basic_string& __str, size_type __pos, size_type __n = npos) |
| 4381 | { return this->assign(__str._M_data() |
| 4382 | + __str._M_check(__pos, "basic_string::assign"), |
| 4383 | __str._M_limit(__pos, __n)); } |
| 4384 | |
| 4385 | /** |
| 4386 | * @brief Set value to a C substring. |
| 4387 | * @param __s The C string to use. |
| 4388 | * @param __n Number of characters to use. |
| 4389 | * @return Reference to this string. |
| 4390 | * |
| 4391 | * This function sets the value of this string to the first @a __n |
| 4392 | * characters of @a __s. If @a __n is is larger than the number of |
| 4393 | * available characters in @a __s, the remainder of @a __s is used. |
| 4394 | */ |
| 4395 | basic_string& |
| 4396 | assign(const _CharT* __s, size_type __n); |
| 4397 | |
| 4398 | /** |
| 4399 | * @brief Set value to contents of a C string. |
| 4400 | * @param __s The C string to use. |
| 4401 | * @return Reference to this string. |
| 4402 | * |
| 4403 | * This function sets the value of this string to the value of @a __s. |
| 4404 | * The data is copied, so there is no dependence on @a __s once the |
| 4405 | * function returns. |
| 4406 | */ |
| 4407 | basic_string& |
| 4408 | assign(const _CharT* __s) |
| 4409 | { |
| 4410 | __glibcxx_requires_string(__s); |
| 4411 | return this->assign(__s, traits_type::length(__s)); |
| 4412 | } |
| 4413 | |
| 4414 | /** |
| 4415 | * @brief Set value to multiple characters. |
| 4416 | * @param __n Length of the resulting string. |
| 4417 | * @param __c The character to use. |
| 4418 | * @return Reference to this string. |
| 4419 | * |
| 4420 | * This function sets the value of this string to @a __n copies of |
| 4421 | * character @a __c. |
| 4422 | */ |
| 4423 | basic_string& |
| 4424 | assign(size_type __n, _CharT __c) |
| 4425 | { return _M_replace_aux(size_type(0), this->size(), __n, __c); } |
| 4426 | |
| 4427 | /** |
| 4428 | * @brief Set value to a range of characters. |
| 4429 | * @param __first Iterator referencing the first character to append. |
| 4430 | * @param __last Iterator marking the end of the range. |
| 4431 | * @return Reference to this string. |
| 4432 | * |
| 4433 | * Sets value of string to characters in the range [__first,__last). |
| 4434 | */ |
| 4435 | template<class _InputIterator> |
| 4436 | basic_string& |
| 4437 | assign(_InputIterator __first, _InputIterator __last) |
| 4438 | { return this->replace(_M_ibegin(), _M_iend(), __first, __last); } |
| 4439 | |
| 4440 | #if __cplusplus201703L >= 201103L |
| 4441 | /** |
| 4442 | * @brief Set value to an initializer_list of characters. |
| 4443 | * @param __l The initializer_list of characters to assign. |
| 4444 | * @return Reference to this string. |
| 4445 | */ |
| 4446 | basic_string& |
| 4447 | assign(initializer_list<_CharT> __l) |
| 4448 | { return this->assign(__l.begin(), __l.size()); } |
| 4449 | #endif // C++11 |
| 4450 | |
| 4451 | #if __cplusplus201703L >= 201703L |
| 4452 | /** |
| 4453 | * @brief Set value from a string_view. |
| 4454 | * @param __svt The source object convertible to string_view. |
| 4455 | * @return Reference to this string. |
| 4456 | */ |
| 4457 | template<typename _Tp> |
| 4458 | _If_sv<_Tp, basic_string&> |
| 4459 | assign(const _Tp& __svt) |
| 4460 | { |
| 4461 | __sv_type __sv = __svt; |
| 4462 | return this->assign(__sv.data(), __sv.size()); |
| 4463 | } |
| 4464 | |
| 4465 | /** |
| 4466 | * @brief Set value from a range of characters in a string_view. |
| 4467 | * @param __svt The source object convertible to string_view. |
| 4468 | * @param __pos The position in the string_view to assign from. |
| 4469 | * @param __n The number of characters to assign. |
| 4470 | * @return Reference to this string. |
| 4471 | */ |
| 4472 | template<typename _Tp> |
| 4473 | _If_sv<_Tp, basic_string&> |
| 4474 | assign(const _Tp& __svt, size_type __pos, size_type __n = npos) |
| 4475 | { |
| 4476 | __sv_type __sv = __svt; |
| 4477 | return assign(__sv.data() |
| 4478 | + std::__sv_check(__sv.size(), __pos, "basic_string::assign"), |
| 4479 | std::__sv_limit(__sv.size(), __pos, __n)); |
| 4480 | } |
| 4481 | #endif // C++17 |
| 4482 | |
| 4483 | /** |
| 4484 | * @brief Insert multiple characters. |
| 4485 | * @param __p Iterator referencing location in string to insert at. |
| 4486 | * @param __n Number of characters to insert |
| 4487 | * @param __c The character to insert. |
| 4488 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4489 | * |
| 4490 | * Inserts @a __n copies of character @a __c starting at the |
| 4491 | * position referenced by iterator @a __p. If adding |
| 4492 | * characters causes the length to exceed max_size(), |
| 4493 | * length_error is thrown. The value of the string doesn't |
| 4494 | * change if an error is thrown. |
| 4495 | */ |
| 4496 | void |
| 4497 | insert(iterator __p, size_type __n, _CharT __c) |
| 4498 | { this->replace(__p, __p, __n, __c); } |
| 4499 | |
| 4500 | /** |
| 4501 | * @brief Insert a range of characters. |
| 4502 | * @param __p Iterator referencing location in string to insert at. |
| 4503 | * @param __beg Start of range. |
| 4504 | * @param __end End of range. |
| 4505 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4506 | * |
| 4507 | * Inserts characters in range [__beg,__end). If adding |
| 4508 | * characters causes the length to exceed max_size(), |
| 4509 | * length_error is thrown. The value of the string doesn't |
| 4510 | * change if an error is thrown. |
| 4511 | */ |
| 4512 | template<class _InputIterator> |
| 4513 | void |
| 4514 | insert(iterator __p, _InputIterator __beg, _InputIterator __end) |
| 4515 | { this->replace(__p, __p, __beg, __end); } |
| 4516 | |
| 4517 | #if __cplusplus201703L >= 201103L |
| 4518 | /** |
| 4519 | * @brief Insert an initializer_list of characters. |
| 4520 | * @param __p Iterator referencing location in string to insert at. |
| 4521 | * @param __l The initializer_list of characters to insert. |
| 4522 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4523 | */ |
| 4524 | void |
| 4525 | insert(iterator __p, initializer_list<_CharT> __l) |
| 4526 | { |
| 4527 | _GLIBCXX_DEBUG_PEDASSERT(__p >= _M_ibegin() && __p <= _M_iend()); |
| 4528 | this->insert(__p - _M_ibegin(), __l.begin(), __l.size()); |
| 4529 | } |
| 4530 | #endif // C++11 |
| 4531 | |
| 4532 | /** |
| 4533 | * @brief Insert value of a string. |
| 4534 | * @param __pos1 Position in string to insert at. |
| 4535 | * @param __str The string to insert. |
| 4536 | * @return Reference to this string. |
| 4537 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4538 | * |
| 4539 | * Inserts value of @a __str starting at @a __pos1. If adding |
| 4540 | * characters causes the length to exceed max_size(), |
| 4541 | * length_error is thrown. The value of the string doesn't |
| 4542 | * change if an error is thrown. |
| 4543 | */ |
| 4544 | basic_string& |
| 4545 | insert(size_type __pos1, const basic_string& __str) |
| 4546 | { return this->insert(__pos1, __str, size_type(0), __str.size()); } |
| 4547 | |
| 4548 | /** |
| 4549 | * @brief Insert a substring. |
| 4550 | * @param __pos1 Position in string to insert at. |
| 4551 | * @param __str The string to insert. |
| 4552 | * @param __pos2 Start of characters in str to insert. |
| 4553 | * @param __n Number of characters to insert. |
| 4554 | * @return Reference to this string. |
| 4555 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4556 | * @throw std::out_of_range If @a pos1 > size() or |
| 4557 | * @a __pos2 > @a str.size(). |
| 4558 | * |
| 4559 | * Starting at @a pos1, insert @a __n character of @a __str |
| 4560 | * beginning with @a __pos2. If adding characters causes the |
| 4561 | * length to exceed max_size(), length_error is thrown. If @a |
| 4562 | * __pos1 is beyond the end of this string or @a __pos2 is |
| 4563 | * beyond the end of @a __str, out_of_range is thrown. The |
| 4564 | * value of the string doesn't change if an error is thrown. |
| 4565 | */ |
| 4566 | basic_string& |
| 4567 | insert(size_type __pos1, const basic_string& __str, |
| 4568 | size_type __pos2, size_type __n = npos) |
| 4569 | { return this->insert(__pos1, __str._M_data() |
| 4570 | + __str._M_check(__pos2, "basic_string::insert"), |
| 4571 | __str._M_limit(__pos2, __n)); } |
| 4572 | |
| 4573 | /** |
| 4574 | * @brief Insert a C substring. |
| 4575 | * @param __pos Position in string to insert at. |
| 4576 | * @param __s The C string to insert. |
| 4577 | * @param __n The number of characters to insert. |
| 4578 | * @return Reference to this string. |
| 4579 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4580 | * @throw std::out_of_range If @a __pos is beyond the end of this |
| 4581 | * string. |
| 4582 | * |
| 4583 | * Inserts the first @a __n characters of @a __s starting at @a |
| 4584 | * __pos. If adding characters causes the length to exceed |
| 4585 | * max_size(), length_error is thrown. If @a __pos is beyond |
| 4586 | * end(), out_of_range is thrown. The value of the string |
| 4587 | * doesn't change if an error is thrown. |
| 4588 | */ |
| 4589 | basic_string& |
| 4590 | insert(size_type __pos, const _CharT* __s, size_type __n); |
| 4591 | |
| 4592 | /** |
| 4593 | * @brief Insert a C string. |
| 4594 | * @param __pos Position in string to insert at. |
| 4595 | * @param __s The C string to insert. |
| 4596 | * @return Reference to this string. |
| 4597 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4598 | * @throw std::out_of_range If @a pos is beyond the end of this |
| 4599 | * string. |
| 4600 | * |
| 4601 | * Inserts the first @a n characters of @a __s starting at @a __pos. If |
| 4602 | * adding characters causes the length to exceed max_size(), |
| 4603 | * length_error is thrown. If @a __pos is beyond end(), out_of_range is |
| 4604 | * thrown. The value of the string doesn't change if an error is |
| 4605 | * thrown. |
| 4606 | */ |
| 4607 | basic_string& |
| 4608 | insert(size_type __pos, const _CharT* __s) |
| 4609 | { |
| 4610 | __glibcxx_requires_string(__s); |
| 4611 | return this->insert(__pos, __s, traits_type::length(__s)); |
| 4612 | } |
| 4613 | |
| 4614 | /** |
| 4615 | * @brief Insert multiple characters. |
| 4616 | * @param __pos Index in string to insert at. |
| 4617 | * @param __n Number of characters to insert |
| 4618 | * @param __c The character to insert. |
| 4619 | * @return Reference to this string. |
| 4620 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4621 | * @throw std::out_of_range If @a __pos is beyond the end of this |
| 4622 | * string. |
| 4623 | * |
| 4624 | * Inserts @a __n copies of character @a __c starting at index |
| 4625 | * @a __pos. If adding characters causes the length to exceed |
| 4626 | * max_size(), length_error is thrown. If @a __pos > length(), |
| 4627 | * out_of_range is thrown. The value of the string doesn't |
| 4628 | * change if an error is thrown. |
| 4629 | */ |
| 4630 | basic_string& |
| 4631 | insert(size_type __pos, size_type __n, _CharT __c) |
| 4632 | { return _M_replace_aux(_M_check(__pos, "basic_string::insert"), |
| 4633 | size_type(0), __n, __c); } |
| 4634 | |
| 4635 | /** |
| 4636 | * @brief Insert one character. |
| 4637 | * @param __p Iterator referencing position in string to insert at. |
| 4638 | * @param __c The character to insert. |
| 4639 | * @return Iterator referencing newly inserted char. |
| 4640 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4641 | * |
| 4642 | * Inserts character @a __c at position referenced by @a __p. |
| 4643 | * If adding character causes the length to exceed max_size(), |
| 4644 | * length_error is thrown. If @a __p is beyond end of string, |
| 4645 | * out_of_range is thrown. The value of the string doesn't |
| 4646 | * change if an error is thrown. |
| 4647 | */ |
| 4648 | iterator |
| 4649 | insert(iterator __p, _CharT __c) |
| 4650 | { |
| 4651 | _GLIBCXX_DEBUG_PEDASSERT(__p >= _M_ibegin() && __p <= _M_iend()); |
| 4652 | const size_type __pos = __p - _M_ibegin(); |
| 4653 | _M_replace_aux(__pos, size_type(0), size_type(1), __c); |
| 4654 | _M_rep()->_M_set_leaked(); |
| 4655 | return iterator(_M_data() + __pos); |
| 4656 | } |
| 4657 | |
| 4658 | #if __cplusplus201703L >= 201703L |
| 4659 | /** |
| 4660 | * @brief Insert a string_view. |
| 4661 | * @param __pos Position in string to insert at. |
| 4662 | * @param __svt The object convertible to string_view to insert. |
| 4663 | * @return Reference to this string. |
| 4664 | */ |
| 4665 | template<typename _Tp> |
| 4666 | _If_sv<_Tp, basic_string&> |
| 4667 | insert(size_type __pos, const _Tp& __svt) |
| 4668 | { |
| 4669 | __sv_type __sv = __svt; |
| 4670 | return this->insert(__pos, __sv.data(), __sv.size()); |
| 4671 | } |
| 4672 | |
| 4673 | /** |
| 4674 | * @brief Insert a string_view. |
| 4675 | * @param __pos Position in string to insert at. |
| 4676 | * @param __svt The object convertible to string_view to insert from. |
| 4677 | * @param __pos Position in string_view to insert |
| 4678 | * from. |
| 4679 | * @param __n The number of characters to insert. |
| 4680 | * @return Reference to this string. |
| 4681 | */ |
| 4682 | template<typename _Tp> |
| 4683 | _If_sv<_Tp, basic_string&> |
| 4684 | insert(size_type __pos1, const _Tp& __svt, |
| 4685 | size_type __pos2, size_type __n = npos) |
| 4686 | { |
| 4687 | __sv_type __sv = __svt; |
| 4688 | return this->replace(__pos1, size_type(0), __sv.data() |
| 4689 | + std::__sv_check(__sv.size(), __pos2, "basic_string::insert"), |
| 4690 | std::__sv_limit(__sv.size(), __pos2, __n)); |
| 4691 | } |
| 4692 | #endif // C++17 |
| 4693 | |
| 4694 | /** |
| 4695 | * @brief Remove characters. |
| 4696 | * @param __pos Index of first character to remove (default 0). |
| 4697 | * @param __n Number of characters to remove (default remainder). |
| 4698 | * @return Reference to this string. |
| 4699 | * @throw std::out_of_range If @a pos is beyond the end of this |
| 4700 | * string. |
| 4701 | * |
| 4702 | * Removes @a __n characters from this string starting at @a |
| 4703 | * __pos. The length of the string is reduced by @a __n. If |
| 4704 | * there are < @a __n characters to remove, the remainder of |
| 4705 | * the string is truncated. If @a __p is beyond end of string, |
| 4706 | * out_of_range is thrown. The value of the string doesn't |
| 4707 | * change if an error is thrown. |
| 4708 | */ |
| 4709 | basic_string& |
| 4710 | erase(size_type __pos = 0, size_type __n = npos) |
| 4711 | { |
| 4712 | _M_mutate(_M_check(__pos, "basic_string::erase"), |
| 4713 | _M_limit(__pos, __n), size_type(0)); |
| 4714 | return *this; |
| 4715 | } |
| 4716 | |
| 4717 | /** |
| 4718 | * @brief Remove one character. |
| 4719 | * @param __position Iterator referencing the character to remove. |
| 4720 | * @return iterator referencing same location after removal. |
| 4721 | * |
| 4722 | * Removes the character at @a __position from this string. The value |
| 4723 | * of the string doesn't change if an error is thrown. |
| 4724 | */ |
| 4725 | iterator |
| 4726 | erase(iterator __position) |
| 4727 | { |
| 4728 | _GLIBCXX_DEBUG_PEDASSERT(__position >= _M_ibegin() |
| 4729 | && __position < _M_iend()); |
| 4730 | const size_type __pos = __position - _M_ibegin(); |
| 4731 | _M_mutate(__pos, size_type(1), size_type(0)); |
| 4732 | _M_rep()->_M_set_leaked(); |
| 4733 | return iterator(_M_data() + __pos); |
| 4734 | } |
| 4735 | |
| 4736 | /** |
| 4737 | * @brief Remove a range of characters. |
| 4738 | * @param __first Iterator referencing the first character to remove. |
| 4739 | * @param __last Iterator referencing the end of the range. |
| 4740 | * @return Iterator referencing location of first after removal. |
| 4741 | * |
| 4742 | * Removes the characters in the range [first,last) from this string. |
| 4743 | * The value of the string doesn't change if an error is thrown. |
| 4744 | */ |
| 4745 | iterator |
| 4746 | erase(iterator __first, iterator __last); |
| 4747 | |
| 4748 | #if __cplusplus201703L >= 201103L |
| 4749 | /** |
| 4750 | * @brief Remove the last character. |
| 4751 | * |
| 4752 | * The string must be non-empty. |
| 4753 | */ |
| 4754 | void |
| 4755 | pop_back() // FIXME C++11: should be noexcept. |
| 4756 | { |
| 4757 | __glibcxx_assert(!empty()); |
| 4758 | erase(size() - 1, 1); |
| 4759 | } |
| 4760 | #endif // C++11 |
| 4761 | |
| 4762 | /** |
| 4763 | * @brief Replace characters with value from another string. |
| 4764 | * @param __pos Index of first character to replace. |
| 4765 | * @param __n Number of characters to be replaced. |
| 4766 | * @param __str String to insert. |
| 4767 | * @return Reference to this string. |
| 4768 | * @throw std::out_of_range If @a pos is beyond the end of this |
| 4769 | * string. |
| 4770 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4771 | * |
| 4772 | * Removes the characters in the range [__pos,__pos+__n) from |
| 4773 | * this string. In place, the value of @a __str is inserted. |
| 4774 | * If @a __pos is beyond end of string, out_of_range is thrown. |
| 4775 | * If the length of the result exceeds max_size(), length_error |
| 4776 | * is thrown. The value of the string doesn't change if an |
| 4777 | * error is thrown. |
| 4778 | */ |
| 4779 | basic_string& |
| 4780 | replace(size_type __pos, size_type __n, const basic_string& __str) |
| 4781 | { return this->replace(__pos, __n, __str._M_data(), __str.size()); } |
| 4782 | |
| 4783 | /** |
| 4784 | * @brief Replace characters with value from another string. |
| 4785 | * @param __pos1 Index of first character to replace. |
| 4786 | * @param __n1 Number of characters to be replaced. |
| 4787 | * @param __str String to insert. |
| 4788 | * @param __pos2 Index of first character of str to use. |
| 4789 | * @param __n2 Number of characters from str to use. |
| 4790 | * @return Reference to this string. |
| 4791 | * @throw std::out_of_range If @a __pos1 > size() or @a __pos2 > |
| 4792 | * __str.size(). |
| 4793 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4794 | * |
| 4795 | * Removes the characters in the range [__pos1,__pos1 + n) from this |
| 4796 | * string. In place, the value of @a __str is inserted. If @a __pos is |
| 4797 | * beyond end of string, out_of_range is thrown. If the length of the |
| 4798 | * result exceeds max_size(), length_error is thrown. The value of the |
| 4799 | * string doesn't change if an error is thrown. |
| 4800 | */ |
| 4801 | basic_string& |
| 4802 | replace(size_type __pos1, size_type __n1, const basic_string& __str, |
| 4803 | size_type __pos2, size_type __n2 = npos) |
| 4804 | { return this->replace(__pos1, __n1, __str._M_data() |
| 4805 | + __str._M_check(__pos2, "basic_string::replace"), |
| 4806 | __str._M_limit(__pos2, __n2)); } |
| 4807 | |
| 4808 | /** |
| 4809 | * @brief Replace characters with value of a C substring. |
| 4810 | * @param __pos Index of first character to replace. |
| 4811 | * @param __n1 Number of characters to be replaced. |
| 4812 | * @param __s C string to insert. |
| 4813 | * @param __n2 Number of characters from @a s to use. |
| 4814 | * @return Reference to this string. |
| 4815 | * @throw std::out_of_range If @a pos1 > size(). |
| 4816 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4817 | * |
| 4818 | * Removes the characters in the range [__pos,__pos + __n1) |
| 4819 | * from this string. In place, the first @a __n2 characters of |
| 4820 | * @a __s are inserted, or all of @a __s if @a __n2 is too large. If |
| 4821 | * @a __pos is beyond end of string, out_of_range is thrown. If |
| 4822 | * the length of result exceeds max_size(), length_error is |
| 4823 | * thrown. The value of the string doesn't change if an error |
| 4824 | * is thrown. |
| 4825 | */ |
| 4826 | basic_string& |
| 4827 | replace(size_type __pos, size_type __n1, const _CharT* __s, |
| 4828 | size_type __n2); |
| 4829 | |
| 4830 | /** |
| 4831 | * @brief Replace characters with value of a C string. |
| 4832 | * @param __pos Index of first character to replace. |
| 4833 | * @param __n1 Number of characters to be replaced. |
| 4834 | * @param __s C string to insert. |
| 4835 | * @return Reference to this string. |
| 4836 | * @throw std::out_of_range If @a pos > size(). |
| 4837 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4838 | * |
| 4839 | * Removes the characters in the range [__pos,__pos + __n1) |
| 4840 | * from this string. In place, the characters of @a __s are |
| 4841 | * inserted. If @a __pos is beyond end of string, out_of_range |
| 4842 | * is thrown. If the length of result exceeds max_size(), |
| 4843 | * length_error is thrown. The value of the string doesn't |
| 4844 | * change if an error is thrown. |
| 4845 | */ |
| 4846 | basic_string& |
| 4847 | replace(size_type __pos, size_type __n1, const _CharT* __s) |
| 4848 | { |
| 4849 | __glibcxx_requires_string(__s); |
| 4850 | return this->replace(__pos, __n1, __s, traits_type::length(__s)); |
| 4851 | } |
| 4852 | |
| 4853 | /** |
| 4854 | * @brief Replace characters with multiple characters. |
| 4855 | * @param __pos Index of first character to replace. |
| 4856 | * @param __n1 Number of characters to be replaced. |
| 4857 | * @param __n2 Number of characters to insert. |
| 4858 | * @param __c Character to insert. |
| 4859 | * @return Reference to this string. |
| 4860 | * @throw std::out_of_range If @a __pos > size(). |
| 4861 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4862 | * |
| 4863 | * Removes the characters in the range [pos,pos + n1) from this |
| 4864 | * string. In place, @a __n2 copies of @a __c are inserted. |
| 4865 | * If @a __pos is beyond end of string, out_of_range is thrown. |
| 4866 | * If the length of result exceeds max_size(), length_error is |
| 4867 | * thrown. The value of the string doesn't change if an error |
| 4868 | * is thrown. |
| 4869 | */ |
| 4870 | basic_string& |
| 4871 | replace(size_type __pos, size_type __n1, size_type __n2, _CharT __c) |
| 4872 | { return _M_replace_aux(_M_check(__pos, "basic_string::replace"), |
| 4873 | _M_limit(__pos, __n1), __n2, __c); } |
| 4874 | |
| 4875 | /** |
| 4876 | * @brief Replace range of characters with string. |
| 4877 | * @param __i1 Iterator referencing start of range to replace. |
| 4878 | * @param __i2 Iterator referencing end of range to replace. |
| 4879 | * @param __str String value to insert. |
| 4880 | * @return Reference to this string. |
| 4881 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4882 | * |
| 4883 | * Removes the characters in the range [__i1,__i2). In place, |
| 4884 | * the value of @a __str is inserted. If the length of result |
| 4885 | * exceeds max_size(), length_error is thrown. The value of |
| 4886 | * the string doesn't change if an error is thrown. |
| 4887 | */ |
| 4888 | basic_string& |
| 4889 | replace(iterator __i1, iterator __i2, const basic_string& __str) |
| 4890 | { return this->replace(__i1, __i2, __str._M_data(), __str.size()); } |
| 4891 | |
| 4892 | /** |
| 4893 | * @brief Replace range of characters with C substring. |
| 4894 | * @param __i1 Iterator referencing start of range to replace. |
| 4895 | * @param __i2 Iterator referencing end of range to replace. |
| 4896 | * @param __s C string value to insert. |
| 4897 | * @param __n Number of characters from s to insert. |
| 4898 | * @return Reference to this string. |
| 4899 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4900 | * |
| 4901 | * Removes the characters in the range [__i1,__i2). In place, |
| 4902 | * the first @a __n characters of @a __s are inserted. If the |
| 4903 | * length of result exceeds max_size(), length_error is thrown. |
| 4904 | * The value of the string doesn't change if an error is |
| 4905 | * thrown. |
| 4906 | */ |
| 4907 | basic_string& |
| 4908 | replace(iterator __i1, iterator __i2, const _CharT* __s, size_type __n) |
| 4909 | { |
| 4910 | _GLIBCXX_DEBUG_PEDASSERT(_M_ibegin() <= __i1 && __i1 <= __i2 |
| 4911 | && __i2 <= _M_iend()); |
| 4912 | return this->replace(__i1 - _M_ibegin(), __i2 - __i1, __s, __n); |
| 4913 | } |
| 4914 | |
| 4915 | /** |
| 4916 | * @brief Replace range of characters with C string. |
| 4917 | * @param __i1 Iterator referencing start of range to replace. |
| 4918 | * @param __i2 Iterator referencing end of range to replace. |
| 4919 | * @param __s C string value to insert. |
| 4920 | * @return Reference to this string. |
| 4921 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4922 | * |
| 4923 | * Removes the characters in the range [__i1,__i2). In place, |
| 4924 | * the characters of @a __s are inserted. If the length of |
| 4925 | * result exceeds max_size(), length_error is thrown. The |
| 4926 | * value of the string doesn't change if an error is thrown. |
| 4927 | */ |
| 4928 | basic_string& |
| 4929 | replace(iterator __i1, iterator __i2, const _CharT* __s) |
| 4930 | { |
| 4931 | __glibcxx_requires_string(__s); |
| 4932 | return this->replace(__i1, __i2, __s, traits_type::length(__s)); |
| 4933 | } |
| 4934 | |
| 4935 | /** |
| 4936 | * @brief Replace range of characters with multiple characters |
| 4937 | * @param __i1 Iterator referencing start of range to replace. |
| 4938 | * @param __i2 Iterator referencing end of range to replace. |
| 4939 | * @param __n Number of characters to insert. |
| 4940 | * @param __c Character to insert. |
| 4941 | * @return Reference to this string. |
| 4942 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4943 | * |
| 4944 | * Removes the characters in the range [__i1,__i2). In place, |
| 4945 | * @a __n copies of @a __c are inserted. If the length of |
| 4946 | * result exceeds max_size(), length_error is thrown. The |
| 4947 | * value of the string doesn't change if an error is thrown. |
| 4948 | */ |
| 4949 | basic_string& |
| 4950 | replace(iterator __i1, iterator __i2, size_type __n, _CharT __c) |
| 4951 | { |
| 4952 | _GLIBCXX_DEBUG_PEDASSERT(_M_ibegin() <= __i1 && __i1 <= __i2 |
| 4953 | && __i2 <= _M_iend()); |
| 4954 | return _M_replace_aux(__i1 - _M_ibegin(), __i2 - __i1, __n, __c); |
| 4955 | } |
| 4956 | |
| 4957 | /** |
| 4958 | * @brief Replace range of characters with range. |
| 4959 | * @param __i1 Iterator referencing start of range to replace. |
| 4960 | * @param __i2 Iterator referencing end of range to replace. |
| 4961 | * @param __k1 Iterator referencing start of range to insert. |
| 4962 | * @param __k2 Iterator referencing end of range to insert. |
| 4963 | * @return Reference to this string. |
| 4964 | * @throw std::length_error If new length exceeds @c max_size(). |
| 4965 | * |
| 4966 | * Removes the characters in the range [__i1,__i2). In place, |
| 4967 | * characters in the range [__k1,__k2) are inserted. If the |
| 4968 | * length of result exceeds max_size(), length_error is thrown. |
| 4969 | * The value of the string doesn't change if an error is |
| 4970 | * thrown. |
| 4971 | */ |
| 4972 | template<class _InputIterator> |
| 4973 | basic_string& |
| 4974 | replace(iterator __i1, iterator __i2, |
| 4975 | _InputIterator __k1, _InputIterator __k2) |
| 4976 | { |
| 4977 | _GLIBCXX_DEBUG_PEDASSERT(_M_ibegin() <= __i1 && __i1 <= __i2 |
| 4978 | && __i2 <= _M_iend()); |
| 4979 | __glibcxx_requires_valid_range(__k1, __k2); |
| 4980 | typedef typename std::__is_integer<_InputIterator>::__type _Integral; |
| 4981 | return _M_replace_dispatch(__i1, __i2, __k1, __k2, _Integral()); |
| 4982 | } |
| 4983 | |
| 4984 | // Specializations for the common case of pointer and iterator: |
| 4985 | // useful to avoid the overhead of temporary buffering in _M_replace. |
| 4986 | basic_string& |
| 4987 | replace(iterator __i1, iterator __i2, _CharT* __k1, _CharT* __k2) |
| 4988 | { |
| 4989 | _GLIBCXX_DEBUG_PEDASSERT(_M_ibegin() <= __i1 && __i1 <= __i2 |
| 4990 | && __i2 <= _M_iend()); |
| 4991 | __glibcxx_requires_valid_range(__k1, __k2); |
| 4992 | return this->replace(__i1 - _M_ibegin(), __i2 - __i1, |
| 4993 | __k1, __k2 - __k1); |
| 4994 | } |
| 4995 | |
| 4996 | basic_string& |
| 4997 | replace(iterator __i1, iterator __i2, |
| 4998 | const _CharT* __k1, const _CharT* __k2) |
| 4999 | { |
| 5000 | _GLIBCXX_DEBUG_PEDASSERT(_M_ibegin() <= __i1 && __i1 <= __i2 |
| 5001 | && __i2 <= _M_iend()); |
| 5002 | __glibcxx_requires_valid_range(__k1, __k2); |
| 5003 | return this->replace(__i1 - _M_ibegin(), __i2 - __i1, |
| 5004 | __k1, __k2 - __k1); |
| 5005 | } |
| 5006 | |
| 5007 | basic_string& |
| 5008 | replace(iterator __i1, iterator __i2, iterator __k1, iterator __k2) |
| 5009 | { |
| 5010 | _GLIBCXX_DEBUG_PEDASSERT(_M_ibegin() <= __i1 && __i1 <= __i2 |
| 5011 | && __i2 <= _M_iend()); |
| 5012 | __glibcxx_requires_valid_range(__k1, __k2); |
| 5013 | return this->replace(__i1 - _M_ibegin(), __i2 - __i1, |
| 5014 | __k1.base(), __k2 - __k1); |
| 5015 | } |
| 5016 | |
| 5017 | basic_string& |
| 5018 | replace(iterator __i1, iterator __i2, |
| 5019 | const_iterator __k1, const_iterator __k2) |
| 5020 | { |
| 5021 | _GLIBCXX_DEBUG_PEDASSERT(_M_ibegin() <= __i1 && __i1 <= __i2 |
| 5022 | && __i2 <= _M_iend()); |
| 5023 | __glibcxx_requires_valid_range(__k1, __k2); |
| 5024 | return this->replace(__i1 - _M_ibegin(), __i2 - __i1, |
| 5025 | __k1.base(), __k2 - __k1); |
| 5026 | } |
| 5027 | |
| 5028 | #if __cplusplus201703L >= 201103L |
| 5029 | /** |
| 5030 | * @brief Replace range of characters with initializer_list. |
| 5031 | * @param __i1 Iterator referencing start of range to replace. |
| 5032 | * @param __i2 Iterator referencing end of range to replace. |
| 5033 | * @param __l The initializer_list of characters to insert. |
| 5034 | * @return Reference to this string. |
| 5035 | * @throw std::length_error If new length exceeds @c max_size(). |
| 5036 | * |
| 5037 | * Removes the characters in the range [__i1,__i2). In place, |
| 5038 | * characters in the range [__k1,__k2) are inserted. If the |
| 5039 | * length of result exceeds max_size(), length_error is thrown. |
| 5040 | * The value of the string doesn't change if an error is |
| 5041 | * thrown. |
| 5042 | */ |
| 5043 | basic_string& replace(iterator __i1, iterator __i2, |
| 5044 | initializer_list<_CharT> __l) |
| 5045 | { return this->replace(__i1, __i2, __l.begin(), __l.end()); } |
| 5046 | #endif // C++11 |
| 5047 | |
| 5048 | #if __cplusplus201703L >= 201703L |
| 5049 | /** |
| 5050 | * @brief Replace range of characters with string_view. |
| 5051 | * @param __pos The position to replace at. |
| 5052 | * @param __n The number of characters to replace. |
| 5053 | * @param __svt The object convertible to string_view to insert. |
| 5054 | * @return Reference to this string. |
| 5055 | */ |
| 5056 | template<typename _Tp> |
| 5057 | _If_sv<_Tp, basic_string&> |
| 5058 | replace(size_type __pos, size_type __n, const _Tp& __svt) |
| 5059 | { |
| 5060 | __sv_type __sv = __svt; |
| 5061 | return this->replace(__pos, __n, __sv.data(), __sv.size()); |
| 5062 | } |
| 5063 | |
| 5064 | /** |
| 5065 | * @brief Replace range of characters with string_view. |
| 5066 | * @param __pos1 The position to replace at. |
| 5067 | * @param __n1 The number of characters to replace. |
| 5068 | * @param __svt The object convertible to string_view to insert from. |
| 5069 | * @param __pos2 The position in the string_view to insert from. |
| 5070 | * @param __n2 The number of characters to insert. |
| 5071 | * @return Reference to this string. |
| 5072 | */ |
| 5073 | template<typename _Tp> |
| 5074 | _If_sv<_Tp, basic_string&> |
| 5075 | replace(size_type __pos1, size_type __n1, const _Tp& __svt, |
| 5076 | size_type __pos2, size_type __n2 = npos) |
| 5077 | { |
| 5078 | __sv_type __sv = __svt; |
| 5079 | return this->replace(__pos1, __n1, |
| 5080 | __sv.data() |
| 5081 | + std::__sv_check(__sv.size(), __pos2, "basic_string::replace"), |
| 5082 | std::__sv_limit(__sv.size(), __pos2, __n2)); |
| 5083 | } |
| 5084 | |
| 5085 | /** |
| 5086 | * @brief Replace range of characters with string_view. |
| 5087 | * @param __i1 An iterator referencing the start position |
| 5088 | to replace at. |
| 5089 | * @param __i2 An iterator referencing the end position |
| 5090 | for the replace. |
| 5091 | * @param __svt The object convertible to string_view to insert from. |
| 5092 | * @return Reference to this string. |
| 5093 | */ |
| 5094 | template<typename _Tp> |
| 5095 | _If_sv<_Tp, basic_string&> |
| 5096 | replace(const_iterator __i1, const_iterator __i2, const _Tp& __svt) |
| 5097 | { |
| 5098 | __sv_type __sv = __svt; |
| 5099 | return this->replace(__i1 - begin(), __i2 - __i1, __sv); |
| 5100 | } |
| 5101 | #endif // C++17 |
| 5102 | |
| 5103 | private: |
| 5104 | template<class _Integer> |
| 5105 | basic_string& |
| 5106 | _M_replace_dispatch(iterator __i1, iterator __i2, _Integer __n, |
| 5107 | _Integer __val, __true_type) |
| 5108 | { return _M_replace_aux(__i1 - _M_ibegin(), __i2 - __i1, __n, __val); } |
| 5109 | |
| 5110 | template<class _InputIterator> |
| 5111 | basic_string& |
| 5112 | _M_replace_dispatch(iterator __i1, iterator __i2, _InputIterator __k1, |
| 5113 | _InputIterator __k2, __false_type); |
| 5114 | |
| 5115 | basic_string& |
| 5116 | _M_replace_aux(size_type __pos1, size_type __n1, size_type __n2, |
| 5117 | _CharT __c); |
| 5118 | |
| 5119 | basic_string& |
| 5120 | _M_replace_safe(size_type __pos1, size_type __n1, const _CharT* __s, |
| 5121 | size_type __n2); |
| 5122 | |
| 5123 | // _S_construct_aux is used to implement the 21.3.1 para 15 which |
| 5124 | // requires special behaviour if _InIter is an integral type |
| 5125 | template<class _InIterator> |
| 5126 | static _CharT* |
| 5127 | _S_construct_aux(_InIterator __beg, _InIterator __end, |
| 5128 | const _Alloc& __a, __false_type) |
| 5129 | { |
| 5130 | typedef typename iterator_traits<_InIterator>::iterator_category _Tag; |
| 5131 | return _S_construct(__beg, __end, __a, _Tag()); |
| 5132 | } |
| 5133 | |
| 5134 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
| 5135 | // 438. Ambiguity in the "do the right thing" clause |
| 5136 | template<class _Integer> |
| 5137 | static _CharT* |
| 5138 | _S_construct_aux(_Integer __beg, _Integer __end, |
| 5139 | const _Alloc& __a, __true_type) |
| 5140 | { return _S_construct_aux_2(static_cast<size_type>(__beg), |
| 5141 | __end, __a); } |
| 5142 | |
| 5143 | static _CharT* |
| 5144 | _S_construct_aux_2(size_type __req, _CharT __c, const _Alloc& __a) |
| 5145 | { return _S_construct(__req, __c, __a); } |
| 5146 | |
| 5147 | template<class _InIterator> |
| 5148 | static _CharT* |
| 5149 | _S_construct(_InIterator __beg, _InIterator __end, const _Alloc& __a) |
| 5150 | { |
| 5151 | typedef typename std::__is_integer<_InIterator>::__type _Integral; |
| 5152 | return _S_construct_aux(__beg, __end, __a, _Integral()); |
| 5153 | } |
| 5154 | |
| 5155 | // For Input Iterators, used in istreambuf_iterators, etc. |
| 5156 | template<class _InIterator> |
| 5157 | static _CharT* |
| 5158 | _S_construct(_InIterator __beg, _InIterator __end, const _Alloc& __a, |
| 5159 | input_iterator_tag); |
| 5160 | |
| 5161 | // For forward_iterators up to random_access_iterators, used for |
| 5162 | // string::iterator, _CharT*, etc. |
| 5163 | template<class _FwdIterator> |
| 5164 | static _CharT* |
| 5165 | _S_construct(_FwdIterator __beg, _FwdIterator __end, const _Alloc& __a, |
| 5166 | forward_iterator_tag); |
| 5167 | |
| 5168 | static _CharT* |
| 5169 | _S_construct(size_type __req, _CharT __c, const _Alloc& __a); |
| 5170 | |
| 5171 | public: |
| 5172 | |
| 5173 | /** |
| 5174 | * @brief Copy substring into C string. |
| 5175 | * @param __s C string to copy value into. |
| 5176 | * @param __n Number of characters to copy. |
| 5177 | * @param __pos Index of first character to copy. |
| 5178 | * @return Number of characters actually copied |
| 5179 | * @throw std::out_of_range If __pos > size(). |
| 5180 | * |
| 5181 | * Copies up to @a __n characters starting at @a __pos into the |
| 5182 | * C string @a __s. If @a __pos is %greater than size(), |
| 5183 | * out_of_range is thrown. |
| 5184 | */ |
| 5185 | size_type |
| 5186 | copy(_CharT* __s, size_type __n, size_type __pos = 0) const; |
| 5187 | |
| 5188 | /** |
| 5189 | * @brief Swap contents with another string. |
| 5190 | * @param __s String to swap with. |
| 5191 | * |
| 5192 | * Exchanges the contents of this string with that of @a __s in constant |
| 5193 | * time. |
| 5194 | */ |
| 5195 | void |
| 5196 | swap(basic_string& __s) |
| 5197 | _GLIBCXX_NOEXCEPT_IF(allocator_traits<_Alloc>::is_always_equal::value)noexcept(allocator_traits<_Alloc>::is_always_equal::value ); |
| 5198 | |
| 5199 | // String operations: |
| 5200 | /** |
| 5201 | * @brief Return const pointer to null-terminated contents. |
| 5202 | * |
| 5203 | * This is a handle to internal data. Do not modify or dire things may |
| 5204 | * happen. |
| 5205 | */ |
| 5206 | const _CharT* |
| 5207 | c_str() const _GLIBCXX_NOEXCEPTnoexcept |
| 5208 | { return _M_data(); } |
| 5209 | |
| 5210 | /** |
| 5211 | * @brief Return const pointer to contents. |
| 5212 | * |
| 5213 | * This is a pointer to internal data. It is undefined to modify |
| 5214 | * the contents through the returned pointer. To get a pointer that |
| 5215 | * allows modifying the contents use @c &str[0] instead, |
| 5216 | * (or in C++17 the non-const @c str.data() overload). |
| 5217 | */ |
| 5218 | const _CharT* |
| 5219 | data() const _GLIBCXX_NOEXCEPTnoexcept |
| 5220 | { return _M_data(); } |
| 5221 | |
| 5222 | #if __cplusplus201703L >= 201703L |
| 5223 | /** |
| 5224 | * @brief Return non-const pointer to contents. |
| 5225 | * |
| 5226 | * This is a pointer to the character sequence held by the string. |
| 5227 | * Modifying the characters in the sequence is allowed. |
| 5228 | */ |
| 5229 | _CharT* |
| 5230 | data() noexcept |
| 5231 | { |
| 5232 | _M_leak(); |
| 5233 | return _M_data(); |
| 5234 | } |
| 5235 | #endif |
| 5236 | |
| 5237 | /** |
| 5238 | * @brief Return copy of allocator used to construct this string. |
| 5239 | */ |
| 5240 | allocator_type |
| 5241 | get_allocator() const _GLIBCXX_NOEXCEPTnoexcept |
| 5242 | { return _M_dataplus; } |
| 5243 | |
| 5244 | /** |
| 5245 | * @brief Find position of a C substring. |
| 5246 | * @param __s C string to locate. |
| 5247 | * @param __pos Index of character to search from. |
| 5248 | * @param __n Number of characters from @a s to search for. |
| 5249 | * @return Index of start of first occurrence. |
| 5250 | * |
| 5251 | * Starting from @a __pos, searches forward for the first @a |
| 5252 | * __n characters in @a __s within this string. If found, |
| 5253 | * returns the index where it begins. If not found, returns |
| 5254 | * npos. |
| 5255 | */ |
| 5256 | size_type |
| 5257 | find(const _CharT* __s, size_type __pos, size_type __n) const |
| 5258 | _GLIBCXX_NOEXCEPTnoexcept; |
| 5259 | |
| 5260 | /** |
| 5261 | * @brief Find position of a string. |
| 5262 | * @param __str String to locate. |
| 5263 | * @param __pos Index of character to search from (default 0). |
| 5264 | * @return Index of start of first occurrence. |
| 5265 | * |
| 5266 | * Starting from @a __pos, searches forward for value of @a __str within |
| 5267 | * this string. If found, returns the index where it begins. If not |
| 5268 | * found, returns npos. |
| 5269 | */ |
| 5270 | size_type |
| 5271 | find(const basic_string& __str, size_type __pos = 0) const |
| 5272 | _GLIBCXX_NOEXCEPTnoexcept |
| 5273 | { return this->find(__str.data(), __pos, __str.size()); } |
| 5274 | |
| 5275 | /** |
| 5276 | * @brief Find position of a C string. |
| 5277 | * @param __s C string to locate. |
| 5278 | * @param __pos Index of character to search from (default 0). |
| 5279 | * @return Index of start of first occurrence. |
| 5280 | * |
| 5281 | * Starting from @a __pos, searches forward for the value of @a |
| 5282 | * __s within this string. If found, returns the index where |
| 5283 | * it begins. If not found, returns npos. |
| 5284 | */ |
| 5285 | size_type |
| 5286 | find(const _CharT* __s, size_type __pos = 0) const _GLIBCXX_NOEXCEPTnoexcept |
| 5287 | { |
| 5288 | __glibcxx_requires_string(__s); |
| 5289 | return this->find(__s, __pos, traits_type::length(__s)); |
| 5290 | } |
| 5291 | |
| 5292 | /** |
| 5293 | * @brief Find position of a character. |
| 5294 | * @param __c Character to locate. |
| 5295 | * @param __pos Index of character to search from (default 0). |
| 5296 | * @return Index of first occurrence. |
| 5297 | * |
| 5298 | * Starting from @a __pos, searches forward for @a __c within |
| 5299 | * this string. If found, returns the index where it was |
| 5300 | * found. If not found, returns npos. |
| 5301 | */ |
| 5302 | size_type |
| 5303 | find(_CharT __c, size_type __pos = 0) const _GLIBCXX_NOEXCEPTnoexcept; |
| 5304 | |
| 5305 | #if __cplusplus201703L >= 201703L |
| 5306 | /** |
| 5307 | * @brief Find position of a string_view. |
| 5308 | * @param __svt The object convertible to string_view to locate. |
| 5309 | * @param __pos Index of character to search from (default 0). |
| 5310 | * @return Index of start of first occurrence. |
| 5311 | */ |
| 5312 | template<typename _Tp> |
| 5313 | _If_sv<_Tp, size_type> |
| 5314 | find(const _Tp& __svt, size_type __pos = 0) const |
| 5315 | noexcept(is_same<_Tp, __sv_type>::value) |
| 5316 | { |
| 5317 | __sv_type __sv = __svt; |
| 5318 | return this->find(__sv.data(), __pos, __sv.size()); |
| 5319 | } |
| 5320 | #endif // C++17 |
| 5321 | |
| 5322 | /** |
| 5323 | * @brief Find last position of a string. |
| 5324 | * @param __str String to locate. |
| 5325 | * @param __pos Index of character to search back from (default end). |
| 5326 | * @return Index of start of last occurrence. |
| 5327 | * |
| 5328 | * Starting from @a __pos, searches backward for value of @a |
| 5329 | * __str within this string. If found, returns the index where |
| 5330 | * it begins. If not found, returns npos. |
| 5331 | */ |
| 5332 | size_type |
| 5333 | rfind(const basic_string& __str, size_type __pos = npos) const |
| 5334 | _GLIBCXX_NOEXCEPTnoexcept |
| 5335 | { return this->rfind(__str.data(), __pos, __str.size()); } |
| 5336 | |
| 5337 | /** |
| 5338 | * @brief Find last position of a C substring. |
| 5339 | * @param __s C string to locate. |
| 5340 | * @param __pos Index of character to search back from. |
| 5341 | * @param __n Number of characters from s to search for. |
| 5342 | * @return Index of start of last occurrence. |
| 5343 | * |
| 5344 | * Starting from @a __pos, searches backward for the first @a |
| 5345 | * __n characters in @a __s within this string. If found, |
| 5346 | * returns the index where it begins. If not found, returns |
| 5347 | * npos. |
| 5348 | */ |
| 5349 | size_type |
| 5350 | rfind(const _CharT* __s, size_type __pos, size_type __n) const |
| 5351 | _GLIBCXX_NOEXCEPTnoexcept; |
| 5352 | |
| 5353 | /** |
| 5354 | * @brief Find last position of a C string. |
| 5355 | * @param __s C string to locate. |
| 5356 | * @param __pos Index of character to start search at (default end). |
| 5357 | * @return Index of start of last occurrence. |
| 5358 | * |
| 5359 | * Starting from @a __pos, searches backward for the value of |
| 5360 | * @a __s within this string. If found, returns the index |
| 5361 | * where it begins. If not found, returns npos. |
| 5362 | */ |
| 5363 | size_type |
| 5364 | rfind(const _CharT* __s, size_type __pos = npos) const _GLIBCXX_NOEXCEPTnoexcept |
| 5365 | { |
| 5366 | __glibcxx_requires_string(__s); |
| 5367 | return this->rfind(__s, __pos, traits_type::length(__s)); |
| 5368 | } |
| 5369 | |
| 5370 | /** |
| 5371 | * @brief Find last position of a character. |
| 5372 | * @param __c Character to locate. |
| 5373 | * @param __pos Index of character to search back from (default end). |
| 5374 | * @return Index of last occurrence. |
| 5375 | * |
| 5376 | * Starting from @a __pos, searches backward for @a __c within |
| 5377 | * this string. If found, returns the index where it was |
| 5378 | * found. If not found, returns npos. |
| 5379 | */ |
| 5380 | size_type |
| 5381 | rfind(_CharT __c, size_type __pos = npos) const _GLIBCXX_NOEXCEPTnoexcept; |
| 5382 | |
| 5383 | #if __cplusplus201703L >= 201703L |
| 5384 | /** |
| 5385 | * @brief Find last position of a string_view. |
| 5386 | * @param __svt The object convertible to string_view to locate. |
| 5387 | * @param __pos Index of character to search back from (default end). |
| 5388 | * @return Index of start of last occurrence. |
| 5389 | */ |
| 5390 | template<typename _Tp> |
| 5391 | _If_sv<_Tp, size_type> |
| 5392 | rfind(const _Tp& __svt, size_type __pos = npos) const |
| 5393 | noexcept(is_same<_Tp, __sv_type>::value) |
| 5394 | { |
| 5395 | __sv_type __sv = __svt; |
| 5396 | return this->rfind(__sv.data(), __pos, __sv.size()); |
| 5397 | } |
| 5398 | #endif // C++17 |
| 5399 | |
| 5400 | /** |
| 5401 | * @brief Find position of a character of string. |
| 5402 | * @param __str String containing characters to locate. |
| 5403 | * @param __pos Index of character to search from (default 0). |
| 5404 | * @return Index of first occurrence. |
| 5405 | * |
| 5406 | * Starting from @a __pos, searches forward for one of the |
| 5407 | * characters of @a __str within this string. If found, |
| 5408 | * returns the index where it was found. If not found, returns |
| 5409 | * npos. |
| 5410 | */ |
| 5411 | size_type |
| 5412 | find_first_of(const basic_string& __str, size_type __pos = 0) const |
| 5413 | _GLIBCXX_NOEXCEPTnoexcept |
| 5414 | { return this->find_first_of(__str.data(), __pos, __str.size()); } |
| 5415 | |
| 5416 | /** |
| 5417 | * @brief Find position of a character of C substring. |
| 5418 | * @param __s String containing characters to locate. |
| 5419 | * @param __pos Index of character to search from. |
| 5420 | * @param __n Number of characters from s to search for. |
| 5421 | * @return Index of first occurrence. |
| 5422 | * |
| 5423 | * Starting from @a __pos, searches forward for one of the |
| 5424 | * first @a __n characters of @a __s within this string. If |
| 5425 | * found, returns the index where it was found. If not found, |
| 5426 | * returns npos. |
| 5427 | */ |
| 5428 | size_type |
| 5429 | find_first_of(const _CharT* __s, size_type __pos, size_type __n) const |
| 5430 | _GLIBCXX_NOEXCEPTnoexcept; |
| 5431 | |
| 5432 | /** |
| 5433 | * @brief Find position of a character of C string. |
| 5434 | * @param __s String containing characters to locate. |
| 5435 | * @param __pos Index of character to search from (default 0). |
| 5436 | * @return Index of first occurrence. |
| 5437 | * |
| 5438 | * Starting from @a __pos, searches forward for one of the |
| 5439 | * characters of @a __s within this string. If found, returns |
| 5440 | * the index where it was found. If not found, returns npos. |
| 5441 | */ |
| 5442 | size_type |
| 5443 | find_first_of(const _CharT* __s, size_type __pos = 0) const |
| 5444 | _GLIBCXX_NOEXCEPTnoexcept |
| 5445 | { |
| 5446 | __glibcxx_requires_string(__s); |
| 5447 | return this->find_first_of(__s, __pos, traits_type::length(__s)); |
| 5448 | } |
| 5449 | |
| 5450 | /** |
| 5451 | * @brief Find position of a character. |
| 5452 | * @param __c Character to locate. |
| 5453 | * @param __pos Index of character to search from (default 0). |
| 5454 | * @return Index of first occurrence. |
| 5455 | * |
| 5456 | * Starting from @a __pos, searches forward for the character |
| 5457 | * @a __c within this string. If found, returns the index |
| 5458 | * where it was found. If not found, returns npos. |
| 5459 | * |
| 5460 | * Note: equivalent to find(__c, __pos). |
| 5461 | */ |
| 5462 | size_type |
| 5463 | find_first_of(_CharT __c, size_type __pos = 0) const _GLIBCXX_NOEXCEPTnoexcept |
| 5464 | { return this->find(__c, __pos); } |
| 5465 | |
| 5466 | #if __cplusplus201703L >= 201703L |
| 5467 | /** |
| 5468 | * @brief Find position of a character of a string_view. |
| 5469 | * @param __svt An object convertible to string_view containing |
| 5470 | * characters to locate. |
| 5471 | * @param __pos Index of character to search from (default 0). |
| 5472 | * @return Index of first occurrence. |
| 5473 | */ |
| 5474 | template<typename _Tp> |
| 5475 | _If_sv<_Tp, size_type> |
| 5476 | find_first_of(const _Tp& __svt, size_type __pos = 0) const |
| 5477 | noexcept(is_same<_Tp, __sv_type>::value) |
| 5478 | { |
| 5479 | __sv_type __sv = __svt; |
| 5480 | return this->find_first_of(__sv.data(), __pos, __sv.size()); |
| 5481 | } |
| 5482 | #endif // C++17 |
| 5483 | |
| 5484 | /** |
| 5485 | * @brief Find last position of a character of string. |
| 5486 | * @param __str String containing characters to locate. |
| 5487 | * @param __pos Index of character to search back from (default end). |
| 5488 | * @return Index of last occurrence. |
| 5489 | * |
| 5490 | * Starting from @a __pos, searches backward for one of the |
| 5491 | * characters of @a __str within this string. If found, |
| 5492 | * returns the index where it was found. If not found, returns |
| 5493 | * npos. |
| 5494 | */ |
| 5495 | size_type |
| 5496 | find_last_of(const basic_string& __str, size_type __pos = npos) const |
| 5497 | _GLIBCXX_NOEXCEPTnoexcept |
| 5498 | { return this->find_last_of(__str.data(), __pos, __str.size()); } |
| 5499 | |
| 5500 | /** |
| 5501 | * @brief Find last position of a character of C substring. |
| 5502 | * @param __s C string containing characters to locate. |
| 5503 | * @param __pos Index of character to search back from. |
| 5504 | * @param __n Number of characters from s to search for. |
| 5505 | * @return Index of last occurrence. |
| 5506 | * |
| 5507 | * Starting from @a __pos, searches backward for one of the |
| 5508 | * first @a __n characters of @a __s within this string. If |
| 5509 | * found, returns the index where it was found. If not found, |
| 5510 | * returns npos. |
| 5511 | */ |
| 5512 | size_type |
| 5513 | find_last_of(const _CharT* __s, size_type __pos, size_type __n) const |
| 5514 | _GLIBCXX_NOEXCEPTnoexcept; |
| 5515 | |
| 5516 | /** |
| 5517 | * @brief Find last position of a character of C string. |
| 5518 | * @param __s C string containing characters to locate. |
| 5519 | * @param __pos Index of character to search back from (default end). |
| 5520 | * @return Index of last occurrence. |
| 5521 | * |
| 5522 | * Starting from @a __pos, searches backward for one of the |
| 5523 | * characters of @a __s within this string. If found, returns |
| 5524 | * the index where it was found. If not found, returns npos. |
| 5525 | */ |
| 5526 | size_type |
| 5527 | find_last_of(const _CharT* __s, size_type __pos = npos) const |
| 5528 | _GLIBCXX_NOEXCEPTnoexcept |
| 5529 | { |
| 5530 | __glibcxx_requires_string(__s); |
| 5531 | return this->find_last_of(__s, __pos, traits_type::length(__s)); |
| 5532 | } |
| 5533 | |
| 5534 | /** |
| 5535 | * @brief Find last position of a character. |
| 5536 | * @param __c Character to locate. |
| 5537 | * @param __pos Index of character to search back from (default end). |
| 5538 | * @return Index of last occurrence. |
| 5539 | * |
| 5540 | * Starting from @a __pos, searches backward for @a __c within |
| 5541 | * this string. If found, returns the index where it was |
| 5542 | * found. If not found, returns npos. |
| 5543 | * |
| 5544 | * Note: equivalent to rfind(__c, __pos). |
| 5545 | */ |
| 5546 | size_type |
| 5547 | find_last_of(_CharT __c, size_type __pos = npos) const _GLIBCXX_NOEXCEPTnoexcept |
| 5548 | { return this->rfind(__c, __pos); } |
| 5549 | |
| 5550 | #if __cplusplus201703L >= 201703L |
| 5551 | /** |
| 5552 | * @brief Find last position of a character of string. |
| 5553 | * @param __svt An object convertible to string_view containing |
| 5554 | * characters to locate. |
| 5555 | * @param __pos Index of character to search back from (default end). |
| 5556 | * @return Index of last occurrence. |
| 5557 | */ |
| 5558 | template<typename _Tp> |
| 5559 | _If_sv<_Tp, size_type> |
| 5560 | find_last_of(const _Tp& __svt, size_type __pos = npos) const |
| 5561 | noexcept(is_same<_Tp, __sv_type>::value) |
| 5562 | { |
| 5563 | __sv_type __sv = __svt; |
| 5564 | return this->find_last_of(__sv.data(), __pos, __sv.size()); |
| 5565 | } |
| 5566 | #endif // C++17 |
| 5567 | |
| 5568 | /** |
| 5569 | * @brief Find position of a character not in string. |
| 5570 | * @param __str String containing characters to avoid. |
| 5571 | * @param __pos Index of character to search from (default 0). |
| 5572 | * @return Index of first occurrence. |
| 5573 | * |
| 5574 | * Starting from @a __pos, searches forward for a character not contained |
| 5575 | * in @a __str within this string. If found, returns the index where it |
| 5576 | * was found. If not found, returns npos. |
| 5577 | */ |
| 5578 | size_type |
| 5579 | find_first_not_of(const basic_string& __str, size_type __pos = 0) const |
| 5580 | _GLIBCXX_NOEXCEPTnoexcept |
| 5581 | { return this->find_first_not_of(__str.data(), __pos, __str.size()); } |
| 5582 | |
| 5583 | /** |
| 5584 | * @brief Find position of a character not in C substring. |
| 5585 | * @param __s C string containing characters to avoid. |
| 5586 | * @param __pos Index of character to search from. |
| 5587 | * @param __n Number of characters from __s to consider. |
| 5588 | * @return Index of first occurrence. |
| 5589 | * |
| 5590 | * Starting from @a __pos, searches forward for a character not |
| 5591 | * contained in the first @a __n characters of @a __s within |
| 5592 | * this string. If found, returns the index where it was |
| 5593 | * found. If not found, returns npos. |
| 5594 | */ |
| 5595 | size_type |
| 5596 | find_first_not_of(const _CharT* __s, size_type __pos, |
| 5597 | size_type __n) const _GLIBCXX_NOEXCEPTnoexcept; |
| 5598 | |
| 5599 | /** |
| 5600 | * @brief Find position of a character not in C string. |
| 5601 | * @param __s C string containing characters to avoid. |
| 5602 | * @param __pos Index of character to search from (default 0). |
| 5603 | * @return Index of first occurrence. |
| 5604 | * |
| 5605 | * Starting from @a __pos, searches forward for a character not |
| 5606 | * contained in @a __s within this string. If found, returns |
| 5607 | * the index where it was found. If not found, returns npos. |
| 5608 | */ |
| 5609 | size_type |
| 5610 | find_first_not_of(const _CharT* __s, size_type __pos = 0) const |
| 5611 | _GLIBCXX_NOEXCEPTnoexcept |
| 5612 | { |
| 5613 | __glibcxx_requires_string(__s); |
| 5614 | return this->find_first_not_of(__s, __pos, traits_type::length(__s)); |
| 5615 | } |
| 5616 | |
| 5617 | /** |
| 5618 | * @brief Find position of a different character. |
| 5619 | * @param __c Character to avoid. |
| 5620 | * @param __pos Index of character to search from (default 0). |
| 5621 | * @return Index of first occurrence. |
| 5622 | * |
| 5623 | * Starting from @a __pos, searches forward for a character |
| 5624 | * other than @a __c within this string. If found, returns the |
| 5625 | * index where it was found. If not found, returns npos. |
| 5626 | */ |
| 5627 | size_type |
| 5628 | find_first_not_of(_CharT __c, size_type __pos = 0) const |
| 5629 | _GLIBCXX_NOEXCEPTnoexcept; |
| 5630 | |
| 5631 | #if __cplusplus201703L >= 201703L |
| 5632 | /** |
| 5633 | * @brief Find position of a character not in a string_view. |
| 5634 | * @param __svt An object convertible to string_view containing |
| 5635 | * characters to avoid. |
| 5636 | * @param __pos Index of character to search from (default 0). |
| 5637 | * @return Index of first occurrence. |
| 5638 | */ |
| 5639 | template<typename _Tp> |
| 5640 | _If_sv<_Tp, size_type> |
| 5641 | find_first_not_of(const _Tp& __svt, size_type __pos = 0) const |
| 5642 | noexcept(is_same<_Tp, __sv_type>::value) |
| 5643 | { |
| 5644 | __sv_type __sv = __svt; |
| 5645 | return this->find_first_not_of(__sv.data(), __pos, __sv.size()); |
| 5646 | } |
| 5647 | #endif // C++17 |
| 5648 | |
| 5649 | /** |
| 5650 | * @brief Find last position of a character not in string. |
| 5651 | * @param __str String containing characters to avoid. |
| 5652 | * @param __pos Index of character to search back from (default end). |
| 5653 | * @return Index of last occurrence. |
| 5654 | * |
| 5655 | * Starting from @a __pos, searches backward for a character |
| 5656 | * not contained in @a __str within this string. If found, |
| 5657 | * returns the index where it was found. If not found, returns |
| 5658 | * npos. |
| 5659 | */ |
| 5660 | size_type |
| 5661 | find_last_not_of(const basic_string& __str, size_type __pos = npos) const |
| 5662 | _GLIBCXX_NOEXCEPTnoexcept |
| 5663 | { return this->find_last_not_of(__str.data(), __pos, __str.size()); } |
| 5664 | |
| 5665 | /** |
| 5666 | * @brief Find last position of a character not in C substring. |
| 5667 | * @param __s C string containing characters to avoid. |
| 5668 | * @param __pos Index of character to search back from. |
| 5669 | * @param __n Number of characters from s to consider. |
| 5670 | * @return Index of last occurrence. |
| 5671 | * |
| 5672 | * Starting from @a __pos, searches backward for a character not |
| 5673 | * contained in the first @a __n characters of @a __s within this string. |
| 5674 | * If found, returns the index where it was found. If not found, |
| 5675 | * returns npos. |
| 5676 | */ |
| 5677 | size_type |
| 5678 | find_last_not_of(const _CharT* __s, size_type __pos, |
| 5679 | size_type __n) const _GLIBCXX_NOEXCEPTnoexcept; |
| 5680 | /** |
| 5681 | * @brief Find last position of a character not in C string. |
| 5682 | * @param __s C string containing characters to avoid. |
| 5683 | * @param __pos Index of character to search back from (default end). |
| 5684 | * @return Index of last occurrence. |
| 5685 | * |
| 5686 | * Starting from @a __pos, searches backward for a character |
| 5687 | * not contained in @a __s within this string. If found, |
| 5688 | * returns the index where it was found. If not found, returns |
| 5689 | * npos. |
| 5690 | */ |
| 5691 | size_type |
| 5692 | find_last_not_of(const _CharT* __s, size_type __pos = npos) const |
| 5693 | _GLIBCXX_NOEXCEPTnoexcept |
| 5694 | { |
| 5695 | __glibcxx_requires_string(__s); |
| 5696 | return this->find_last_not_of(__s, __pos, traits_type::length(__s)); |
| 5697 | } |
| 5698 | |
| 5699 | /** |
| 5700 | * @brief Find last position of a different character. |
| 5701 | * @param __c Character to avoid. |
| 5702 | * @param __pos Index of character to search back from (default end). |
| 5703 | * @return Index of last occurrence. |
| 5704 | * |
| 5705 | * Starting from @a __pos, searches backward for a character other than |
| 5706 | * @a __c within this string. If found, returns the index where it was |
| 5707 | * found. If not found, returns npos. |
| 5708 | */ |
| 5709 | size_type |
| 5710 | find_last_not_of(_CharT __c, size_type __pos = npos) const |
| 5711 | _GLIBCXX_NOEXCEPTnoexcept; |
| 5712 | |
| 5713 | #if __cplusplus201703L >= 201703L |
| 5714 | /** |
| 5715 | * @brief Find last position of a character not in a string_view. |
| 5716 | * @param __svt An object convertible to string_view containing |
| 5717 | * characters to avoid. |
| 5718 | * @param __pos Index of character to search back from (default end). |
| 5719 | * @return Index of last occurrence. |
| 5720 | */ |
| 5721 | template<typename _Tp> |
| 5722 | _If_sv<_Tp, size_type> |
| 5723 | find_last_not_of(const _Tp& __svt, size_type __pos = npos) const |
| 5724 | noexcept(is_same<_Tp, __sv_type>::value) |
| 5725 | { |
| 5726 | __sv_type __sv = __svt; |
| 5727 | return this->find_last_not_of(__sv.data(), __pos, __sv.size()); |
| 5728 | } |
| 5729 | #endif // C++17 |
| 5730 | |
| 5731 | /** |
| 5732 | * @brief Get a substring. |
| 5733 | * @param __pos Index of first character (default 0). |
| 5734 | * @param __n Number of characters in substring (default remainder). |
| 5735 | * @return The new string. |
| 5736 | * @throw std::out_of_range If __pos > size(). |
| 5737 | * |
| 5738 | * Construct and return a new string using the @a __n |
| 5739 | * characters starting at @a __pos. If the string is too |
| 5740 | * short, use the remainder of the characters. If @a __pos is |
| 5741 | * beyond the end of the string, out_of_range is thrown. |
| 5742 | */ |
| 5743 | basic_string |
| 5744 | substr(size_type __pos = 0, size_type __n = npos) const |
| 5745 | { return basic_string(*this, |
| 5746 | _M_check(__pos, "basic_string::substr"), __n); } |
| 5747 | |
| 5748 | /** |
| 5749 | * @brief Compare to a string. |
| 5750 | * @param __str String to compare against. |
| 5751 | * @return Integer < 0, 0, or > 0. |
| 5752 | * |
| 5753 | * Returns an integer < 0 if this string is ordered before @a |
| 5754 | * __str, 0 if their values are equivalent, or > 0 if this |
| 5755 | * string is ordered after @a __str. Determines the effective |
| 5756 | * length rlen of the strings to compare as the smallest of |
| 5757 | * size() and str.size(). The function then compares the two |
| 5758 | * strings by calling traits::compare(data(), str.data(),rlen). |
| 5759 | * If the result of the comparison is nonzero returns it, |
| 5760 | * otherwise the shorter one is ordered first. |
| 5761 | */ |
| 5762 | int |
| 5763 | compare(const basic_string& __str) const |
| 5764 | { |
| 5765 | const size_type __size = this->size(); |
| 5766 | const size_type __osize = __str.size(); |
| 5767 | const size_type __len = std::min(__size, __osize); |
| 5768 | |
| 5769 | int __r = traits_type::compare(_M_data(), __str.data(), __len); |
| 5770 | if (!__r) |
| 5771 | __r = _S_compare(__size, __osize); |
| 5772 | return __r; |
| 5773 | } |
| 5774 | |
| 5775 | #if __cplusplus201703L >= 201703L |
| 5776 | /** |
| 5777 | * @brief Compare to a string_view. |
| 5778 | * @param __svt An object convertible to string_view to compare against. |
| 5779 | * @return Integer < 0, 0, or > 0. |
| 5780 | */ |
| 5781 | template<typename _Tp> |
| 5782 | _If_sv<_Tp, int> |
| 5783 | compare(const _Tp& __svt) const |
| 5784 | noexcept(is_same<_Tp, __sv_type>::value) |
| 5785 | { |
| 5786 | __sv_type __sv = __svt; |
| 5787 | const size_type __size = this->size(); |
| 5788 | const size_type __osize = __sv.size(); |
| 5789 | const size_type __len = std::min(__size, __osize); |
| 5790 | |
| 5791 | int __r = traits_type::compare(_M_data(), __sv.data(), __len); |
| 5792 | if (!__r) |
| 5793 | __r = _S_compare(__size, __osize); |
| 5794 | return __r; |
| 5795 | } |
| 5796 | |
| 5797 | /** |
| 5798 | * @brief Compare to a string_view. |
| 5799 | * @param __pos A position in the string to start comparing from. |
| 5800 | * @param __n The number of characters to compare. |
| 5801 | * @param __svt An object convertible to string_view to compare |
| 5802 | * against. |
| 5803 | * @return Integer < 0, 0, or > 0. |
| 5804 | */ |
| 5805 | template<typename _Tp> |
| 5806 | _If_sv<_Tp, int> |
| 5807 | compare(size_type __pos, size_type __n, const _Tp& __svt) const |
| 5808 | noexcept(is_same<_Tp, __sv_type>::value) |
| 5809 | { |
| 5810 | __sv_type __sv = __svt; |
| 5811 | return __sv_type(*this).substr(__pos, __n).compare(__sv); |
| 5812 | } |
| 5813 | |
| 5814 | /** |
| 5815 | * @brief Compare to a string_view. |
| 5816 | * @param __pos1 A position in the string to start comparing from. |
| 5817 | * @param __n1 The number of characters to compare. |
| 5818 | * @param __svt An object convertible to string_view to compare |
| 5819 | * against. |
| 5820 | * @param __pos2 A position in the string_view to start comparing from. |
| 5821 | * @param __n2 The number of characters to compare. |
| 5822 | * @return Integer < 0, 0, or > 0. |
| 5823 | */ |
| 5824 | template<typename _Tp> |
| 5825 | _If_sv<_Tp, int> |
| 5826 | compare(size_type __pos1, size_type __n1, const _Tp& __svt, |
| 5827 | size_type __pos2, size_type __n2 = npos) const |
| 5828 | noexcept(is_same<_Tp, __sv_type>::value) |
| 5829 | { |
| 5830 | __sv_type __sv = __svt; |
| 5831 | return __sv_type(*this) |
| 5832 | .substr(__pos1, __n1).compare(__sv.substr(__pos2, __n2)); |
| 5833 | } |
| 5834 | #endif // C++17 |
| 5835 | |
| 5836 | /** |
| 5837 | * @brief Compare substring to a string. |
| 5838 | * @param __pos Index of first character of substring. |
| 5839 | * @param __n Number of characters in substring. |
| 5840 | * @param __str String to compare against. |
| 5841 | * @return Integer < 0, 0, or > 0. |
| 5842 | * |
| 5843 | * Form the substring of this string from the @a __n characters |
| 5844 | * starting at @a __pos. Returns an integer < 0 if the |
| 5845 | * substring is ordered before @a __str, 0 if their values are |
| 5846 | * equivalent, or > 0 if the substring is ordered after @a |
| 5847 | * __str. Determines the effective length rlen of the strings |
| 5848 | * to compare as the smallest of the length of the substring |
| 5849 | * and @a __str.size(). The function then compares the two |
| 5850 | * strings by calling |
| 5851 | * traits::compare(substring.data(),str.data(),rlen). If the |
| 5852 | * result of the comparison is nonzero returns it, otherwise |
| 5853 | * the shorter one is ordered first. |
| 5854 | */ |
| 5855 | int |
| 5856 | compare(size_type __pos, size_type __n, const basic_string& __str) const; |
| 5857 | |
| 5858 | /** |
| 5859 | * @brief Compare substring to a substring. |
| 5860 | * @param __pos1 Index of first character of substring. |
| 5861 | * @param __n1 Number of characters in substring. |
| 5862 | * @param __str String to compare against. |
| 5863 | * @param __pos2 Index of first character of substring of str. |
| 5864 | * @param __n2 Number of characters in substring of str. |
| 5865 | * @return Integer < 0, 0, or > 0. |
| 5866 | * |
| 5867 | * Form the substring of this string from the @a __n1 |
| 5868 | * characters starting at @a __pos1. Form the substring of @a |
| 5869 | * __str from the @a __n2 characters starting at @a __pos2. |
| 5870 | * Returns an integer < 0 if this substring is ordered before |
| 5871 | * the substring of @a __str, 0 if their values are equivalent, |
| 5872 | * or > 0 if this substring is ordered after the substring of |
| 5873 | * @a __str. Determines the effective length rlen of the |
| 5874 | * strings to compare as the smallest of the lengths of the |
| 5875 | * substrings. The function then compares the two strings by |
| 5876 | * calling |
| 5877 | * traits::compare(substring.data(),str.substr(pos2,n2).data(),rlen). |
| 5878 | * If the result of the comparison is nonzero returns it, |
| 5879 | * otherwise the shorter one is ordered first. |
| 5880 | */ |
| 5881 | int |
| 5882 | compare(size_type __pos1, size_type __n1, const basic_string& __str, |
| 5883 | size_type __pos2, size_type __n2 = npos) const; |
| 5884 | |
| 5885 | /** |
| 5886 | * @brief Compare to a C string. |
| 5887 | * @param __s C string to compare against. |
| 5888 | * @return Integer < 0, 0, or > 0. |
| 5889 | * |
| 5890 | * Returns an integer < 0 if this string is ordered before @a __s, 0 if |
| 5891 | * their values are equivalent, or > 0 if this string is ordered after |
| 5892 | * @a __s. Determines the effective length rlen of the strings to |
| 5893 | * compare as the smallest of size() and the length of a string |
| 5894 | * constructed from @a __s. The function then compares the two strings |
| 5895 | * by calling traits::compare(data(),s,rlen). If the result of the |
| 5896 | * comparison is nonzero returns it, otherwise the shorter one is |
| 5897 | * ordered first. |
| 5898 | */ |
| 5899 | int |
| 5900 | compare(const _CharT* __s) const _GLIBCXX_NOEXCEPTnoexcept; |
| 5901 | |
| 5902 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
| 5903 | // 5 String::compare specification questionable |
| 5904 | /** |
| 5905 | * @brief Compare substring to a C string. |
| 5906 | * @param __pos Index of first character of substring. |
| 5907 | * @param __n1 Number of characters in substring. |
| 5908 | * @param __s C string to compare against. |
| 5909 | * @return Integer < 0, 0, or > 0. |
| 5910 | * |
| 5911 | * Form the substring of this string from the @a __n1 |
| 5912 | * characters starting at @a pos. Returns an integer < 0 if |
| 5913 | * the substring is ordered before @a __s, 0 if their values |
| 5914 | * are equivalent, or > 0 if the substring is ordered after @a |
| 5915 | * __s. Determines the effective length rlen of the strings to |
| 5916 | * compare as the smallest of the length of the substring and |
| 5917 | * the length of a string constructed from @a __s. The |
| 5918 | * function then compares the two string by calling |
| 5919 | * traits::compare(substring.data(),__s,rlen). If the result of |
| 5920 | * the comparison is nonzero returns it, otherwise the shorter |
| 5921 | * one is ordered first. |
| 5922 | */ |
| 5923 | int |
| 5924 | compare(size_type __pos, size_type __n1, const _CharT* __s) const; |
| 5925 | |
| 5926 | /** |
| 5927 | * @brief Compare substring against a character %array. |
| 5928 | * @param __pos Index of first character of substring. |
| 5929 | * @param __n1 Number of characters in substring. |
| 5930 | * @param __s character %array to compare against. |
| 5931 | * @param __n2 Number of characters of s. |
| 5932 | * @return Integer < 0, 0, or > 0. |
| 5933 | * |
| 5934 | * Form the substring of this string from the @a __n1 |
| 5935 | * characters starting at @a __pos. Form a string from the |
| 5936 | * first @a __n2 characters of @a __s. Returns an integer < 0 |
| 5937 | * if this substring is ordered before the string from @a __s, |
| 5938 | * 0 if their values are equivalent, or > 0 if this substring |
| 5939 | * is ordered after the string from @a __s. Determines the |
| 5940 | * effective length rlen of the strings to compare as the |
| 5941 | * smallest of the length of the substring and @a __n2. The |
| 5942 | * function then compares the two strings by calling |
| 5943 | * traits::compare(substring.data(),s,rlen). If the result of |
| 5944 | * the comparison is nonzero returns it, otherwise the shorter |
| 5945 | * one is ordered first. |
| 5946 | * |
| 5947 | * NB: s must have at least n2 characters, '\\0' has |
| 5948 | * no special meaning. |
| 5949 | */ |
| 5950 | int |
| 5951 | compare(size_type __pos, size_type __n1, const _CharT* __s, |
| 5952 | size_type __n2) const; |
| 5953 | |
| 5954 | #if __cplusplus201703L > 201703L |
| 5955 | bool |
| 5956 | starts_with(basic_string_view<_CharT, _Traits> __x) const noexcept |
| 5957 | { return __sv_type(this->data(), this->size()).starts_with(__x); } |
| 5958 | |
| 5959 | bool |
| 5960 | starts_with(_CharT __x) const noexcept |
| 5961 | { return __sv_type(this->data(), this->size()).starts_with(__x); } |
| 5962 | |
| 5963 | bool |
| 5964 | starts_with(const _CharT* __x) const noexcept |
| 5965 | { return __sv_type(this->data(), this->size()).starts_with(__x); } |
| 5966 | |
| 5967 | bool |
| 5968 | ends_with(basic_string_view<_CharT, _Traits> __x) const noexcept |
| 5969 | { return __sv_type(this->data(), this->size()).ends_with(__x); } |
| 5970 | |
| 5971 | bool |
| 5972 | ends_with(_CharT __x) const noexcept |
| 5973 | { return __sv_type(this->data(), this->size()).ends_with(__x); } |
| 5974 | |
| 5975 | bool |
| 5976 | ends_with(const _CharT* __x) const noexcept |
| 5977 | { return __sv_type(this->data(), this->size()).ends_with(__x); } |
| 5978 | #endif // C++20 |
| 5979 | |
| 5980 | # ifdef _GLIBCXX_TM_TS_INTERNAL |
| 5981 | friend void |
| 5982 | ::_txnal_cow_string_C1_for_exceptions(void* that, const char* s, |
| 5983 | void* exc); |
| 5984 | friend const char* |
| 5985 | ::_txnal_cow_string_c_str(const void *that); |
| 5986 | friend void |
| 5987 | ::_txnal_cow_string_D1(void *that); |
| 5988 | friend void |
| 5989 | ::_txnal_cow_string_D1_commit(void *that); |
| 5990 | # endif |
| 5991 | }; |
| 5992 | #endif // !_GLIBCXX_USE_CXX11_ABI |
| 5993 | |
| 5994 | #if __cpp_deduction_guides201703L >= 201606 |
| 5995 | _GLIBCXX_BEGIN_NAMESPACE_CXX11namespace __cxx11 { |
| 5996 | template<typename _InputIterator, typename _CharT |
| 5997 | = typename iterator_traits<_InputIterator>::value_type, |
| 5998 | typename _Allocator = allocator<_CharT>, |
| 5999 | typename = _RequireInputIter<_InputIterator>, |
| 6000 | typename = _RequireAllocator<_Allocator>> |
| 6001 | basic_string(_InputIterator, _InputIterator, _Allocator = _Allocator()) |
| 6002 | -> basic_string<_CharT, char_traits<_CharT>, _Allocator>; |
| 6003 | |
| 6004 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
| 6005 | // 3075. basic_string needs deduction guides from basic_string_view |
| 6006 | template<typename _CharT, typename _Traits, |
| 6007 | typename _Allocator = allocator<_CharT>, |
| 6008 | typename = _RequireAllocator<_Allocator>> |
| 6009 | basic_string(basic_string_view<_CharT, _Traits>, const _Allocator& = _Allocator()) |
| 6010 | -> basic_string<_CharT, _Traits, _Allocator>; |
| 6011 | |
| 6012 | template<typename _CharT, typename _Traits, |
| 6013 | typename _Allocator = allocator<_CharT>, |
| 6014 | typename = _RequireAllocator<_Allocator>> |
| 6015 | basic_string(basic_string_view<_CharT, _Traits>, |
| 6016 | typename basic_string<_CharT, _Traits, _Allocator>::size_type, |
| 6017 | typename basic_string<_CharT, _Traits, _Allocator>::size_type, |
| 6018 | const _Allocator& = _Allocator()) |
| 6019 | -> basic_string<_CharT, _Traits, _Allocator>; |
| 6020 | _GLIBCXX_END_NAMESPACE_CXX11} |
| 6021 | #endif |
| 6022 | |
| 6023 | // operator+ |
| 6024 | /** |
| 6025 | * @brief Concatenate two strings. |
| 6026 | * @param __lhs First string. |
| 6027 | * @param __rhs Last string. |
| 6028 | * @return New string with value of @a __lhs followed by @a __rhs. |
| 6029 | */ |
| 6030 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6031 | basic_string<_CharT, _Traits, _Alloc> |
| 6032 | operator+(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6033 | const basic_string<_CharT, _Traits, _Alloc>& __rhs) |
| 6034 | { |
| 6035 | basic_string<_CharT, _Traits, _Alloc> __str(__lhs); |
| 6036 | __str.append(__rhs); |
| 6037 | return __str; |
| 6038 | } |
| 6039 | |
| 6040 | /** |
| 6041 | * @brief Concatenate C string and string. |
| 6042 | * @param __lhs First string. |
| 6043 | * @param __rhs Last string. |
| 6044 | * @return New string with value of @a __lhs followed by @a __rhs. |
| 6045 | */ |
| 6046 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6047 | basic_string<_CharT,_Traits,_Alloc> |
| 6048 | operator+(const _CharT* __lhs, |
| 6049 | const basic_string<_CharT,_Traits,_Alloc>& __rhs); |
| 6050 | |
| 6051 | /** |
| 6052 | * @brief Concatenate character and string. |
| 6053 | * @param __lhs First string. |
| 6054 | * @param __rhs Last string. |
| 6055 | * @return New string with @a __lhs followed by @a __rhs. |
| 6056 | */ |
| 6057 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6058 | basic_string<_CharT,_Traits,_Alloc> |
| 6059 | operator+(_CharT __lhs, const basic_string<_CharT,_Traits,_Alloc>& __rhs); |
| 6060 | |
| 6061 | /** |
| 6062 | * @brief Concatenate string and C string. |
| 6063 | * @param __lhs First string. |
| 6064 | * @param __rhs Last string. |
| 6065 | * @return New string with @a __lhs followed by @a __rhs. |
| 6066 | */ |
| 6067 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6068 | inline basic_string<_CharT, _Traits, _Alloc> |
| 6069 | operator+(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6070 | const _CharT* __rhs) |
| 6071 | { |
| 6072 | basic_string<_CharT, _Traits, _Alloc> __str(__lhs); |
| 6073 | __str.append(__rhs); |
| 6074 | return __str; |
| 6075 | } |
| 6076 | |
| 6077 | /** |
| 6078 | * @brief Concatenate string and character. |
| 6079 | * @param __lhs First string. |
| 6080 | * @param __rhs Last string. |
| 6081 | * @return New string with @a __lhs followed by @a __rhs. |
| 6082 | */ |
| 6083 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6084 | inline basic_string<_CharT, _Traits, _Alloc> |
| 6085 | operator+(const basic_string<_CharT, _Traits, _Alloc>& __lhs, _CharT __rhs) |
| 6086 | { |
| 6087 | typedef basic_string<_CharT, _Traits, _Alloc> __string_type; |
| 6088 | typedef typename __string_type::size_type __size_type; |
| 6089 | __string_type __str(__lhs); |
| 6090 | __str.append(__size_type(1), __rhs); |
| 6091 | return __str; |
| 6092 | } |
| 6093 | |
| 6094 | #if __cplusplus201703L >= 201103L |
| 6095 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6096 | inline basic_string<_CharT, _Traits, _Alloc> |
| 6097 | operator+(basic_string<_CharT, _Traits, _Alloc>&& __lhs, |
| 6098 | const basic_string<_CharT, _Traits, _Alloc>& __rhs) |
| 6099 | { return std::move(__lhs.append(__rhs)); } |
| 6100 | |
| 6101 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6102 | inline basic_string<_CharT, _Traits, _Alloc> |
| 6103 | operator+(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6104 | basic_string<_CharT, _Traits, _Alloc>&& __rhs) |
| 6105 | { return std::move(__rhs.insert(0, __lhs)); } |
| 6106 | |
| 6107 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6108 | inline basic_string<_CharT, _Traits, _Alloc> |
| 6109 | operator+(basic_string<_CharT, _Traits, _Alloc>&& __lhs, |
| 6110 | basic_string<_CharT, _Traits, _Alloc>&& __rhs) |
| 6111 | { |
| 6112 | #if _GLIBCXX_USE_CXX11_ABI1 |
| 6113 | using _Alloc_traits = allocator_traits<_Alloc>; |
| 6114 | bool __use_rhs = false; |
| 6115 | if _GLIBCXX17_CONSTEXPRconstexpr (typename _Alloc_traits::is_always_equal{}) |
| 6116 | __use_rhs = true; |
| 6117 | else if (__lhs.get_allocator() == __rhs.get_allocator()) |
| 6118 | __use_rhs = true; |
| 6119 | if (__use_rhs) |
| 6120 | #endif |
| 6121 | { |
| 6122 | const auto __size = __lhs.size() + __rhs.size(); |
| 6123 | if (__size > __lhs.capacity() && __size <= __rhs.capacity()) |
| 6124 | return std::move(__rhs.insert(0, __lhs)); |
| 6125 | } |
| 6126 | return std::move(__lhs.append(__rhs)); |
| 6127 | } |
| 6128 | |
| 6129 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6130 | inline basic_string<_CharT, _Traits, _Alloc> |
| 6131 | operator+(const _CharT* __lhs, |
| 6132 | basic_string<_CharT, _Traits, _Alloc>&& __rhs) |
| 6133 | { return std::move(__rhs.insert(0, __lhs)); } |
| 6134 | |
| 6135 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6136 | inline basic_string<_CharT, _Traits, _Alloc> |
| 6137 | operator+(_CharT __lhs, |
| 6138 | basic_string<_CharT, _Traits, _Alloc>&& __rhs) |
| 6139 | { return std::move(__rhs.insert(0, 1, __lhs)); } |
| 6140 | |
| 6141 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6142 | inline basic_string<_CharT, _Traits, _Alloc> |
| 6143 | operator+(basic_string<_CharT, _Traits, _Alloc>&& __lhs, |
| 6144 | const _CharT* __rhs) |
| 6145 | { return std::move(__lhs.append(__rhs)); } |
| 6146 | |
| 6147 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6148 | inline basic_string<_CharT, _Traits, _Alloc> |
| 6149 | operator+(basic_string<_CharT, _Traits, _Alloc>&& __lhs, |
| 6150 | _CharT __rhs) |
| 6151 | { return std::move(__lhs.append(1, __rhs)); } |
| 6152 | #endif |
| 6153 | |
| 6154 | // operator == |
| 6155 | /** |
| 6156 | * @brief Test equivalence of two strings. |
| 6157 | * @param __lhs First string. |
| 6158 | * @param __rhs Second string. |
| 6159 | * @return True if @a __lhs.compare(@a __rhs) == 0. False otherwise. |
| 6160 | */ |
| 6161 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6162 | inline bool |
| 6163 | operator==(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6164 | const basic_string<_CharT, _Traits, _Alloc>& __rhs) |
| 6165 | _GLIBCXX_NOEXCEPTnoexcept |
| 6166 | { return __lhs.compare(__rhs) == 0; } |
| 6167 | |
| 6168 | template<typename _CharT> |
| 6169 | inline |
| 6170 | typename __gnu_cxx::__enable_if<__is_char<_CharT>::__value, bool>::__type |
| 6171 | operator==(const basic_string<_CharT>& __lhs, |
| 6172 | const basic_string<_CharT>& __rhs) _GLIBCXX_NOEXCEPTnoexcept |
| 6173 | { return (__lhs.size() == __rhs.size() |
| 6174 | && !std::char_traits<_CharT>::compare(__lhs.data(), __rhs.data(), |
| 6175 | __lhs.size())); } |
| 6176 | |
| 6177 | /** |
| 6178 | * @brief Test equivalence of string and C string. |
| 6179 | * @param __lhs String. |
| 6180 | * @param __rhs C string. |
| 6181 | * @return True if @a __lhs.compare(@a __rhs) == 0. False otherwise. |
| 6182 | */ |
| 6183 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6184 | inline bool |
| 6185 | operator==(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6186 | const _CharT* __rhs) |
| 6187 | { return __lhs.compare(__rhs) == 0; } |
| 6188 | |
| 6189 | #if __cpp_lib_three_way_comparison |
| 6190 | /** |
| 6191 | * @brief Three-way comparison of a string and a C string. |
| 6192 | * @param __lhs A string. |
| 6193 | * @param __rhs A null-terminated string. |
| 6194 | * @return A value indicating whether `__lhs` is less than, equal to, |
| 6195 | * greater than, or incomparable with `__rhs`. |
| 6196 | */ |
| 6197 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6198 | inline auto |
| 6199 | operator<=>(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6200 | const basic_string<_CharT, _Traits, _Alloc>& __rhs) noexcept |
| 6201 | -> decltype(__detail::__char_traits_cmp_cat<_Traits>(0)) |
| 6202 | { return __detail::__char_traits_cmp_cat<_Traits>(__lhs.compare(__rhs)); } |
| 6203 | |
| 6204 | /** |
| 6205 | * @brief Three-way comparison of a string and a C string. |
| 6206 | * @param __lhs A string. |
| 6207 | * @param __rhs A null-terminated string. |
| 6208 | * @return A value indicating whether `__lhs` is less than, equal to, |
| 6209 | * greater than, or incomparable with `__rhs`. |
| 6210 | */ |
| 6211 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6212 | inline auto |
| 6213 | operator<=>(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6214 | const _CharT* __rhs) noexcept |
| 6215 | -> decltype(__detail::__char_traits_cmp_cat<_Traits>(0)) |
| 6216 | { return __detail::__char_traits_cmp_cat<_Traits>(__lhs.compare(__rhs)); } |
| 6217 | #else |
| 6218 | /** |
| 6219 | * @brief Test equivalence of C string and string. |
| 6220 | * @param __lhs C string. |
| 6221 | * @param __rhs String. |
| 6222 | * @return True if @a __rhs.compare(@a __lhs) == 0. False otherwise. |
| 6223 | */ |
| 6224 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6225 | inline bool |
| 6226 | operator==(const _CharT* __lhs, |
| 6227 | const basic_string<_CharT, _Traits, _Alloc>& __rhs) |
| 6228 | { return __rhs.compare(__lhs) == 0; } |
| 6229 | |
| 6230 | // operator != |
| 6231 | /** |
| 6232 | * @brief Test difference of two strings. |
| 6233 | * @param __lhs First string. |
| 6234 | * @param __rhs Second string. |
| 6235 | * @return True if @a __lhs.compare(@a __rhs) != 0. False otherwise. |
| 6236 | */ |
| 6237 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6238 | inline bool |
| 6239 | operator!=(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6240 | const basic_string<_CharT, _Traits, _Alloc>& __rhs) |
| 6241 | _GLIBCXX_NOEXCEPTnoexcept |
| 6242 | { return !(__lhs == __rhs); } |
| 6243 | |
| 6244 | /** |
| 6245 | * @brief Test difference of C string and string. |
| 6246 | * @param __lhs C string. |
| 6247 | * @param __rhs String. |
| 6248 | * @return True if @a __rhs.compare(@a __lhs) != 0. False otherwise. |
| 6249 | */ |
| 6250 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6251 | inline bool |
| 6252 | operator!=(const _CharT* __lhs, |
| 6253 | const basic_string<_CharT, _Traits, _Alloc>& __rhs) |
| 6254 | { return !(__lhs == __rhs); } |
| 6255 | |
| 6256 | /** |
| 6257 | * @brief Test difference of string and C string. |
| 6258 | * @param __lhs String. |
| 6259 | * @param __rhs C string. |
| 6260 | * @return True if @a __lhs.compare(@a __rhs) != 0. False otherwise. |
| 6261 | */ |
| 6262 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6263 | inline bool |
| 6264 | operator!=(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6265 | const _CharT* __rhs) |
| 6266 | { return !(__lhs == __rhs); } |
| 6267 | |
| 6268 | // operator < |
| 6269 | /** |
| 6270 | * @brief Test if string precedes string. |
| 6271 | * @param __lhs First string. |
| 6272 | * @param __rhs Second string. |
| 6273 | * @return True if @a __lhs precedes @a __rhs. False otherwise. |
| 6274 | */ |
| 6275 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6276 | inline bool |
| 6277 | operator<(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6278 | const basic_string<_CharT, _Traits, _Alloc>& __rhs) |
| 6279 | _GLIBCXX_NOEXCEPTnoexcept |
| 6280 | { return __lhs.compare(__rhs) < 0; } |
| 6281 | |
| 6282 | /** |
| 6283 | * @brief Test if string precedes C string. |
| 6284 | * @param __lhs String. |
| 6285 | * @param __rhs C string. |
| 6286 | * @return True if @a __lhs precedes @a __rhs. False otherwise. |
| 6287 | */ |
| 6288 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6289 | inline bool |
| 6290 | operator<(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6291 | const _CharT* __rhs) |
| 6292 | { return __lhs.compare(__rhs) < 0; } |
| 6293 | |
| 6294 | /** |
| 6295 | * @brief Test if C string precedes string. |
| 6296 | * @param __lhs C string. |
| 6297 | * @param __rhs String. |
| 6298 | * @return True if @a __lhs precedes @a __rhs. False otherwise. |
| 6299 | */ |
| 6300 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6301 | inline bool |
| 6302 | operator<(const _CharT* __lhs, |
| 6303 | const basic_string<_CharT, _Traits, _Alloc>& __rhs) |
| 6304 | { return __rhs.compare(__lhs) > 0; } |
| 6305 | |
| 6306 | // operator > |
| 6307 | /** |
| 6308 | * @brief Test if string follows string. |
| 6309 | * @param __lhs First string. |
| 6310 | * @param __rhs Second string. |
| 6311 | * @return True if @a __lhs follows @a __rhs. False otherwise. |
| 6312 | */ |
| 6313 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6314 | inline bool |
| 6315 | operator>(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6316 | const basic_string<_CharT, _Traits, _Alloc>& __rhs) |
| 6317 | _GLIBCXX_NOEXCEPTnoexcept |
| 6318 | { return __lhs.compare(__rhs) > 0; } |
| 6319 | |
| 6320 | /** |
| 6321 | * @brief Test if string follows C string. |
| 6322 | * @param __lhs String. |
| 6323 | * @param __rhs C string. |
| 6324 | * @return True if @a __lhs follows @a __rhs. False otherwise. |
| 6325 | */ |
| 6326 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6327 | inline bool |
| 6328 | operator>(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6329 | const _CharT* __rhs) |
| 6330 | { return __lhs.compare(__rhs) > 0; } |
| 6331 | |
| 6332 | /** |
| 6333 | * @brief Test if C string follows string. |
| 6334 | * @param __lhs C string. |
| 6335 | * @param __rhs String. |
| 6336 | * @return True if @a __lhs follows @a __rhs. False otherwise. |
| 6337 | */ |
| 6338 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6339 | inline bool |
| 6340 | operator>(const _CharT* __lhs, |
| 6341 | const basic_string<_CharT, _Traits, _Alloc>& __rhs) |
| 6342 | { return __rhs.compare(__lhs) < 0; } |
| 6343 | |
| 6344 | // operator <= |
| 6345 | /** |
| 6346 | * @brief Test if string doesn't follow string. |
| 6347 | * @param __lhs First string. |
| 6348 | * @param __rhs Second string. |
| 6349 | * @return True if @a __lhs doesn't follow @a __rhs. False otherwise. |
| 6350 | */ |
| 6351 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6352 | inline bool |
| 6353 | operator<=(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6354 | const basic_string<_CharT, _Traits, _Alloc>& __rhs) |
| 6355 | _GLIBCXX_NOEXCEPTnoexcept |
| 6356 | { return __lhs.compare(__rhs) <= 0; } |
| 6357 | |
| 6358 | /** |
| 6359 | * @brief Test if string doesn't follow C string. |
| 6360 | * @param __lhs String. |
| 6361 | * @param __rhs C string. |
| 6362 | * @return True if @a __lhs doesn't follow @a __rhs. False otherwise. |
| 6363 | */ |
| 6364 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6365 | inline bool |
| 6366 | operator<=(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6367 | const _CharT* __rhs) |
| 6368 | { return __lhs.compare(__rhs) <= 0; } |
| 6369 | |
| 6370 | /** |
| 6371 | * @brief Test if C string doesn't follow string. |
| 6372 | * @param __lhs C string. |
| 6373 | * @param __rhs String. |
| 6374 | * @return True if @a __lhs doesn't follow @a __rhs. False otherwise. |
| 6375 | */ |
| 6376 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6377 | inline bool |
| 6378 | operator<=(const _CharT* __lhs, |
| 6379 | const basic_string<_CharT, _Traits, _Alloc>& __rhs) |
| 6380 | { return __rhs.compare(__lhs) >= 0; } |
| 6381 | |
| 6382 | // operator >= |
| 6383 | /** |
| 6384 | * @brief Test if string doesn't precede string. |
| 6385 | * @param __lhs First string. |
| 6386 | * @param __rhs Second string. |
| 6387 | * @return True if @a __lhs doesn't precede @a __rhs. False otherwise. |
| 6388 | */ |
| 6389 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6390 | inline bool |
| 6391 | operator>=(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6392 | const basic_string<_CharT, _Traits, _Alloc>& __rhs) |
| 6393 | _GLIBCXX_NOEXCEPTnoexcept |
| 6394 | { return __lhs.compare(__rhs) >= 0; } |
| 6395 | |
| 6396 | /** |
| 6397 | * @brief Test if string doesn't precede C string. |
| 6398 | * @param __lhs String. |
| 6399 | * @param __rhs C string. |
| 6400 | * @return True if @a __lhs doesn't precede @a __rhs. False otherwise. |
| 6401 | */ |
| 6402 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6403 | inline bool |
| 6404 | operator>=(const basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6405 | const _CharT* __rhs) |
| 6406 | { return __lhs.compare(__rhs) >= 0; } |
| 6407 | |
| 6408 | /** |
| 6409 | * @brief Test if C string doesn't precede string. |
| 6410 | * @param __lhs C string. |
| 6411 | * @param __rhs String. |
| 6412 | * @return True if @a __lhs doesn't precede @a __rhs. False otherwise. |
| 6413 | */ |
| 6414 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6415 | inline bool |
| 6416 | operator>=(const _CharT* __lhs, |
| 6417 | const basic_string<_CharT, _Traits, _Alloc>& __rhs) |
| 6418 | { return __rhs.compare(__lhs) <= 0; } |
| 6419 | #endif // three-way comparison |
| 6420 | |
| 6421 | /** |
| 6422 | * @brief Swap contents of two strings. |
| 6423 | * @param __lhs First string. |
| 6424 | * @param __rhs Second string. |
| 6425 | * |
| 6426 | * Exchanges the contents of @a __lhs and @a __rhs in constant time. |
| 6427 | */ |
| 6428 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6429 | inline void |
| 6430 | swap(basic_string<_CharT, _Traits, _Alloc>& __lhs, |
| 6431 | basic_string<_CharT, _Traits, _Alloc>& __rhs) |
| 6432 | _GLIBCXX_NOEXCEPT_IF(noexcept(__lhs.swap(__rhs)))noexcept(noexcept(__lhs.swap(__rhs))) |
| 6433 | { __lhs.swap(__rhs); } |
| 6434 | |
| 6435 | |
| 6436 | /** |
| 6437 | * @brief Read stream into a string. |
| 6438 | * @param __is Input stream. |
| 6439 | * @param __str Buffer to store into. |
| 6440 | * @return Reference to the input stream. |
| 6441 | * |
| 6442 | * Stores characters from @a __is into @a __str until whitespace is |
| 6443 | * found, the end of the stream is encountered, or str.max_size() |
| 6444 | * is reached. If is.width() is non-zero, that is the limit on the |
| 6445 | * number of characters stored into @a __str. Any previous |
| 6446 | * contents of @a __str are erased. |
| 6447 | */ |
| 6448 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6449 | basic_istream<_CharT, _Traits>& |
| 6450 | operator>>(basic_istream<_CharT, _Traits>& __is, |
| 6451 | basic_string<_CharT, _Traits, _Alloc>& __str); |
| 6452 | |
| 6453 | template<> |
| 6454 | basic_istream<char>& |
| 6455 | operator>>(basic_istream<char>& __is, basic_string<char>& __str); |
| 6456 | |
| 6457 | /** |
| 6458 | * @brief Write string to a stream. |
| 6459 | * @param __os Output stream. |
| 6460 | * @param __str String to write out. |
| 6461 | * @return Reference to the output stream. |
| 6462 | * |
| 6463 | * Output characters of @a __str into os following the same rules as for |
| 6464 | * writing a C string. |
| 6465 | */ |
| 6466 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6467 | inline basic_ostream<_CharT, _Traits>& |
| 6468 | operator<<(basic_ostream<_CharT, _Traits>& __os, |
| 6469 | const basic_string<_CharT, _Traits, _Alloc>& __str) |
| 6470 | { |
| 6471 | // _GLIBCXX_RESOLVE_LIB_DEFECTS |
| 6472 | // 586. string inserter not a formatted function |
| 6473 | return __ostream_insert(__os, __str.data(), __str.size()); |
| 6474 | } |
| 6475 | |
| 6476 | /** |
| 6477 | * @brief Read a line from stream into a string. |
| 6478 | * @param __is Input stream. |
| 6479 | * @param __str Buffer to store into. |
| 6480 | * @param __delim Character marking end of line. |
| 6481 | * @return Reference to the input stream. |
| 6482 | * |
| 6483 | * Stores characters from @a __is into @a __str until @a __delim is |
| 6484 | * found, the end of the stream is encountered, or str.max_size() |
| 6485 | * is reached. Any previous contents of @a __str are erased. If |
| 6486 | * @a __delim is encountered, it is extracted but not stored into |
| 6487 | * @a __str. |
| 6488 | */ |
| 6489 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6490 | basic_istream<_CharT, _Traits>& |
| 6491 | getline(basic_istream<_CharT, _Traits>& __is, |
| 6492 | basic_string<_CharT, _Traits, _Alloc>& __str, _CharT __delim); |
| 6493 | |
| 6494 | /** |
| 6495 | * @brief Read a line from stream into a string. |
| 6496 | * @param __is Input stream. |
| 6497 | * @param __str Buffer to store into. |
| 6498 | * @return Reference to the input stream. |
| 6499 | * |
| 6500 | * Stores characters from is into @a __str until '\n' is |
| 6501 | * found, the end of the stream is encountered, or str.max_size() |
| 6502 | * is reached. Any previous contents of @a __str are erased. If |
| 6503 | * end of line is encountered, it is extracted but not stored into |
| 6504 | * @a __str. |
| 6505 | */ |
| 6506 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6507 | inline basic_istream<_CharT, _Traits>& |
| 6508 | getline(basic_istream<_CharT, _Traits>& __is, |
| 6509 | basic_string<_CharT, _Traits, _Alloc>& __str) |
| 6510 | { return std::getline(__is, __str, __is.widen('\n')); } |
| 6511 | |
| 6512 | #if __cplusplus201703L >= 201103L |
| 6513 | /// Read a line from an rvalue stream into a string. |
| 6514 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6515 | inline basic_istream<_CharT, _Traits>& |
| 6516 | getline(basic_istream<_CharT, _Traits>&& __is, |
| 6517 | basic_string<_CharT, _Traits, _Alloc>& __str, _CharT __delim) |
| 6518 | { return std::getline(__is, __str, __delim); } |
| 6519 | |
| 6520 | /// Read a line from an rvalue stream into a string. |
| 6521 | template<typename _CharT, typename _Traits, typename _Alloc> |
| 6522 | inline basic_istream<_CharT, _Traits>& |
| 6523 | getline(basic_istream<_CharT, _Traits>&& __is, |
| 6524 | basic_string<_CharT, _Traits, _Alloc>& __str) |
| 6525 | { return std::getline(__is, __str); } |
| 6526 | #endif |
| 6527 | |
| 6528 | template<> |
| 6529 | basic_istream<char>& |
| 6530 | getline(basic_istream<char>& __in, basic_string<char>& __str, |
| 6531 | char __delim); |
| 6532 | |
| 6533 | #ifdef _GLIBCXX_USE_WCHAR_T1 |
| 6534 | template<> |
| 6535 | basic_istream<wchar_t>& |
| 6536 | getline(basic_istream<wchar_t>& __in, basic_string<wchar_t>& __str, |
| 6537 | wchar_t __delim); |
| 6538 | #endif |
| 6539 | |
| 6540 | _GLIBCXX_END_NAMESPACE_VERSION |
| 6541 | } // namespace |
| 6542 | |
| 6543 | #if __cplusplus201703L >= 201103L |
| 6544 | |
| 6545 | #include <ext/string_conversions.h> |
| 6546 | #include <bits/charconv.h> |
| 6547 | |
| 6548 | namespace std _GLIBCXX_VISIBILITY(default)__attribute__ ((__visibility__ ("default"))) |
| 6549 | { |
| 6550 | _GLIBCXX_BEGIN_NAMESPACE_VERSION |
| 6551 | _GLIBCXX_BEGIN_NAMESPACE_CXX11namespace __cxx11 { |
| 6552 | |
| 6553 | #if _GLIBCXX_USE_C99_STDLIB1 |
| 6554 | // 21.4 Numeric Conversions [string.conversions]. |
| 6555 | inline int |
| 6556 | stoi(const string& __str, size_t* __idx = 0, int __base = 10) |
| 6557 | { return __gnu_cxx::__stoa<long, int>(&std::strtol, "stoi", __str.c_str(), |
| 6558 | __idx, __base); } |
| 6559 | |
| 6560 | inline long |
| 6561 | stol(const string& __str, size_t* __idx = 0, int __base = 10) |
| 6562 | { return __gnu_cxx::__stoa(&std::strtol, "stol", __str.c_str(), |
| 6563 | __idx, __base); } |
| 6564 | |
| 6565 | inline unsigned long |
| 6566 | stoul(const string& __str, size_t* __idx = 0, int __base = 10) |
| 6567 | { return __gnu_cxx::__stoa(&std::strtoul, "stoul", __str.c_str(), |
| 6568 | __idx, __base); } |
| 6569 | |
| 6570 | inline long long |
| 6571 | stoll(const string& __str, size_t* __idx = 0, int __base = 10) |
| 6572 | { return __gnu_cxx::__stoa(&std::strtoll, "stoll", __str.c_str(), |
| 6573 | __idx, __base); } |
| 6574 | |
| 6575 | inline unsigned long long |
| 6576 | stoull(const string& __str, size_t* __idx = 0, int __base = 10) |
| 6577 | { return __gnu_cxx::__stoa(&std::strtoull, "stoull", __str.c_str(), |
| 6578 | __idx, __base); } |
| 6579 | |
| 6580 | // NB: strtof vs strtod. |
| 6581 | inline float |
| 6582 | stof(const string& __str, size_t* __idx = 0) |
| 6583 | { return __gnu_cxx::__stoa(&std::strtof, "stof", __str.c_str(), __idx); } |
| 6584 | |
| 6585 | inline double |
| 6586 | stod(const string& __str, size_t* __idx = 0) |
| 6587 | { return __gnu_cxx::__stoa(&std::strtod, "stod", __str.c_str(), __idx); } |
| 6588 | |
| 6589 | inline long double |
| 6590 | stold(const string& __str, size_t* __idx = 0) |
| 6591 | { return __gnu_cxx::__stoa(&std::strtold, "stold", __str.c_str(), __idx); } |
| 6592 | #endif // _GLIBCXX_USE_C99_STDLIB |
| 6593 | |
| 6594 | // DR 1261. Insufficent overloads for to_string / to_wstring |
| 6595 | |
| 6596 | inline string |
| 6597 | to_string(int __val) |
| 6598 | { |
| 6599 | const bool __neg = __val < 0; |
| 6600 | const unsigned __uval = __neg ? (unsigned)~__val + 1u : __val; |
| 6601 | const auto __len = __detail::__to_chars_len(__uval); |
| 6602 | string __str(__neg + __len, '-'); |
| 6603 | __detail::__to_chars_10_impl(&__str[__neg], __len, __uval); |
| 6604 | return __str; |
| 6605 | } |
| 6606 | |
| 6607 | inline string |
| 6608 | to_string(unsigned __val) |
| 6609 | { |
| 6610 | string __str(__detail::__to_chars_len(__val), '\0'); |
| 6611 | __detail::__to_chars_10_impl(&__str[0], __str.size(), __val); |
| 6612 | return __str; |
| 6613 | } |
| 6614 | |
| 6615 | inline string |
| 6616 | to_string(long __val) |
| 6617 | { |
| 6618 | const bool __neg = __val < 0; |
| 6619 | const unsigned long __uval = __neg ? (unsigned long)~__val + 1ul : __val; |
| 6620 | const auto __len = __detail::__to_chars_len(__uval); |
| 6621 | string __str(__neg + __len, '-'); |
| 6622 | __detail::__to_chars_10_impl(&__str[__neg], __len, __uval); |
| 6623 | return __str; |
| 6624 | } |
| 6625 | |
| 6626 | inline string |
| 6627 | to_string(unsigned long __val) |
| 6628 | { |
| 6629 | string __str(__detail::__to_chars_len(__val), '\0'); |
| 6630 | __detail::__to_chars_10_impl(&__str[0], __str.size(), __val); |
| 6631 | return __str; |
| 6632 | } |
| 6633 | |
| 6634 | inline string |
| 6635 | to_string(long long __val) |
| 6636 | { |
| 6637 | const bool __neg = __val < 0; |
| 6638 | const unsigned long long __uval |
| 6639 | = __neg ? (unsigned long long)~__val + 1ull : __val; |
| 6640 | const auto __len = __detail::__to_chars_len(__uval); |
| 6641 | string __str(__neg + __len, '-'); |
| 6642 | __detail::__to_chars_10_impl(&__str[__neg], __len, __uval); |
| 6643 | return __str; |
| 6644 | } |
| 6645 | |
| 6646 | inline string |
| 6647 | to_string(unsigned long long __val) |
| 6648 | { |
| 6649 | string __str(__detail::__to_chars_len(__val), '\0'); |
| 6650 | __detail::__to_chars_10_impl(&__str[0], __str.size(), __val); |
| 6651 | return __str; |
| 6652 | } |
| 6653 | |
| 6654 | #if _GLIBCXX_USE_C99_STDIO1 |
| 6655 | // NB: (v)snprintf vs sprintf. |
| 6656 | |
| 6657 | inline string |
| 6658 | to_string(float __val) |
| 6659 | { |
| 6660 | const int __n = |
| 6661 | __gnu_cxx::__numeric_traits<float>::__max_exponent10 + 20; |
| 6662 | return __gnu_cxx::__to_xstring<string>(&std::vsnprintf, __n, |
| 6663 | "%f", __val); |
| 6664 | } |
| 6665 | |
| 6666 | inline string |
| 6667 | to_string(double __val) |
| 6668 | { |
| 6669 | const int __n = |
| 6670 | __gnu_cxx::__numeric_traits<double>::__max_exponent10 + 20; |
| 6671 | return __gnu_cxx::__to_xstring<string>(&std::vsnprintf, __n, |
| 6672 | "%f", __val); |
| 6673 | } |
| 6674 | |
| 6675 | inline string |
| 6676 | to_string(long double __val) |
| 6677 | { |
| 6678 | const int __n = |
| 6679 | __gnu_cxx::__numeric_traits<long double>::__max_exponent10 + 20; |
| 6680 | return __gnu_cxx::__to_xstring<string>(&std::vsnprintf, __n, |
| 6681 | "%Lf", __val); |
| 6682 | } |
| 6683 | #endif // _GLIBCXX_USE_C99_STDIO |
| 6684 | |
| 6685 | #if defined(_GLIBCXX_USE_WCHAR_T1) && _GLIBCXX_USE_C99_WCHAR1 |
| 6686 | inline int |
| 6687 | stoi(const wstring& __str, size_t* __idx = 0, int __base = 10) |
| 6688 | { return __gnu_cxx::__stoa<long, int>(&std::wcstol, "stoi", __str.c_str(), |
| 6689 | __idx, __base); } |
| 6690 | |
| 6691 | inline long |
| 6692 | stol(const wstring& __str, size_t* __idx = 0, int __base = 10) |
| 6693 | { return __gnu_cxx::__stoa(&std::wcstol, "stol", __str.c_str(), |
| 6694 | __idx, __base); } |
| 6695 | |
| 6696 | inline unsigned long |
| 6697 | stoul(const wstring& __str, size_t* __idx = 0, int __base = 10) |
| 6698 | { return __gnu_cxx::__stoa(&std::wcstoul, "stoul", __str.c_str(), |
| 6699 | __idx, __base); } |
| 6700 | |
| 6701 | inline long long |
| 6702 | stoll(const wstring& __str, size_t* __idx = 0, int __base = 10) |
| 6703 | { return __gnu_cxx::__stoa(&std::wcstoll, "stoll", __str.c_str(), |
| 6704 | __idx, __base); } |
| 6705 | |
| 6706 | inline unsigned long long |
| 6707 | stoull(const wstring& __str, size_t* __idx = 0, int __base = 10) |
| 6708 | { return __gnu_cxx::__stoa(&std::wcstoull, "stoull", __str.c_str(), |
| 6709 | __idx, __base); } |
| 6710 | |
| 6711 | // NB: wcstof vs wcstod. |
| 6712 | inline float |
| 6713 | stof(const wstring& __str, size_t* __idx = 0) |
| 6714 | { return __gnu_cxx::__stoa(&std::wcstof, "stof", __str.c_str(), __idx); } |
| 6715 | |
| 6716 | inline double |
| 6717 | stod(const wstring& __str, size_t* __idx = 0) |
| 6718 | { return __gnu_cxx::__stoa(&std::wcstod, "stod", __str.c_str(), __idx); } |
| 6719 | |
| 6720 | inline long double |
| 6721 | stold(const wstring& __str, size_t* __idx = 0) |
| 6722 | { return __gnu_cxx::__stoa(&std::wcstold, "stold", __str.c_str(), __idx); } |
| 6723 | |
| 6724 | #ifndef _GLIBCXX_HAVE_BROKEN_VSWPRINTF |
| 6725 | // DR 1261. |
| 6726 | inline wstring |
| 6727 | to_wstring(int __val) |
| 6728 | { return __gnu_cxx::__to_xstring<wstring>(&std::vswprintf, 4 * sizeof(int), |
| 6729 | L"%d", __val); } |
| 6730 | |
| 6731 | inline wstring |
| 6732 | to_wstring(unsigned __val) |
| 6733 | { return __gnu_cxx::__to_xstring<wstring>(&std::vswprintf, |
| 6734 | 4 * sizeof(unsigned), |
| 6735 | L"%u", __val); } |
| 6736 | |
| 6737 | inline wstring |
| 6738 | to_wstring(long __val) |
| 6739 | { return __gnu_cxx::__to_xstring<wstring>(&std::vswprintf, 4 * sizeof(long), |
| 6740 | L"%ld", __val); } |
| 6741 | |
| 6742 | inline wstring |
| 6743 | to_wstring(unsigned long __val) |
| 6744 | { return __gnu_cxx::__to_xstring<wstring>(&std::vswprintf, |
| 6745 | 4 * sizeof(unsigned long), |
| 6746 | L"%lu", __val); } |
| 6747 | |
| 6748 | inline wstring |
| 6749 | to_wstring(long long __val) |
| 6750 | { return __gnu_cxx::__to_xstring<wstring>(&std::vswprintf, |
| 6751 | 4 * sizeof(long long), |
| 6752 | L"%lld", __val); } |
| 6753 | |
| 6754 | inline wstring |
| 6755 | to_wstring(unsigned long long __val) |
| 6756 | { return __gnu_cxx::__to_xstring<wstring>(&std::vswprintf, |
| 6757 | 4 * sizeof(unsigned long long), |
| 6758 | L"%llu", __val); } |
| 6759 | |
| 6760 | inline wstring |
| 6761 | to_wstring(float __val) |
| 6762 | { |
| 6763 | const int __n = |
| 6764 | __gnu_cxx::__numeric_traits<float>::__max_exponent10 + 20; |
| 6765 | return __gnu_cxx::__to_xstring<wstring>(&std::vswprintf, __n, |
| 6766 | L"%f", __val); |
| 6767 | } |
| 6768 | |
| 6769 | inline wstring |
| 6770 | to_wstring(double __val) |
| 6771 | { |
| 6772 | const int __n = |
| 6773 | __gnu_cxx::__numeric_traits<double>::__max_exponent10 + 20; |
| 6774 | return __gnu_cxx::__to_xstring<wstring>(&std::vswprintf, __n, |
| 6775 | L"%f", __val); |
| 6776 | } |
| 6777 | |
| 6778 | inline wstring |
| 6779 | to_wstring(long double __val) |
| 6780 | { |
| 6781 | const int __n = |
| 6782 | __gnu_cxx::__numeric_traits<long double>::__max_exponent10 + 20; |
| 6783 | return __gnu_cxx::__to_xstring<wstring>(&std::vswprintf, __n, |
| 6784 | L"%Lf", __val); |
| 6785 | } |
| 6786 | #endif // _GLIBCXX_HAVE_BROKEN_VSWPRINTF |
| 6787 | #endif // _GLIBCXX_USE_WCHAR_T && _GLIBCXX_USE_C99_WCHAR |
| 6788 | |
| 6789 | _GLIBCXX_END_NAMESPACE_CXX11} |
| 6790 | _GLIBCXX_END_NAMESPACE_VERSION |
| 6791 | } // namespace |
| 6792 | |
| 6793 | #endif /* C++11 */ |
| 6794 | |
| 6795 | #if __cplusplus201703L >= 201103L |
| 6796 | |
| 6797 | #include <bits/functional_hash.h> |
| 6798 | |
| 6799 | namespace std _GLIBCXX_VISIBILITY(default)__attribute__ ((__visibility__ ("default"))) |
| 6800 | { |
| 6801 | _GLIBCXX_BEGIN_NAMESPACE_VERSION |
| 6802 | |
| 6803 | // DR 1182. |
| 6804 | |
| 6805 | #ifndef _GLIBCXX_COMPATIBILITY_CXX0X |
| 6806 | /// std::hash specialization for string. |
| 6807 | template<> |
| 6808 | struct hash<string> |
| 6809 | : public __hash_base<size_t, string> |
| 6810 | { |
| 6811 | size_t |
| 6812 | operator()(const string& __s) const noexcept |
| 6813 | { return std::_Hash_impl::hash(__s.data(), __s.length()); } |
| 6814 | }; |
| 6815 | |
| 6816 | template<> |
| 6817 | struct __is_fast_hash<hash<string>> : std::false_type |
| 6818 | { }; |
| 6819 | |
| 6820 | #ifdef _GLIBCXX_USE_WCHAR_T1 |
| 6821 | /// std::hash specialization for wstring. |
| 6822 | template<> |
| 6823 | struct hash<wstring> |
| 6824 | : public __hash_base<size_t, wstring> |
| 6825 | { |
| 6826 | size_t |
| 6827 | operator()(const wstring& __s) const noexcept |
| 6828 | { return std::_Hash_impl::hash(__s.data(), |
| 6829 | __s.length() * sizeof(wchar_t)); } |
| 6830 | }; |
| 6831 | |
| 6832 | template<> |
| 6833 | struct __is_fast_hash<hash<wstring>> : std::false_type |
| 6834 | { }; |
| 6835 | #endif |
| 6836 | #endif /* _GLIBCXX_COMPATIBILITY_CXX0X */ |
| 6837 | |
| 6838 | #ifdef _GLIBCXX_USE_CHAR8_T |
| 6839 | /// std::hash specialization for u8string. |
| 6840 | template<> |
| 6841 | struct hash<u8string> |
| 6842 | : public __hash_base<size_t, u8string> |
| 6843 | { |
| 6844 | size_t |
| 6845 | operator()(const u8string& __s) const noexcept |
| 6846 | { return std::_Hash_impl::hash(__s.data(), |
| 6847 | __s.length() * sizeof(char8_t)); } |
| 6848 | }; |
| 6849 | |
| 6850 | template<> |
| 6851 | struct __is_fast_hash<hash<u8string>> : std::false_type |
| 6852 | { }; |
| 6853 | #endif |
| 6854 | |
| 6855 | /// std::hash specialization for u16string. |
| 6856 | template<> |
| 6857 | struct hash<u16string> |
| 6858 | : public __hash_base<size_t, u16string> |
| 6859 | { |
| 6860 | size_t |
| 6861 | operator()(const u16string& __s) const noexcept |
| 6862 | { return std::_Hash_impl::hash(__s.data(), |
| 6863 | __s.length() * sizeof(char16_t)); } |
| 6864 | }; |
| 6865 | |
| 6866 | template<> |
| 6867 | struct __is_fast_hash<hash<u16string>> : std::false_type |
| 6868 | { }; |
| 6869 | |
| 6870 | /// std::hash specialization for u32string. |
| 6871 | template<> |
| 6872 | struct hash<u32string> |
| 6873 | : public __hash_base<size_t, u32string> |
| 6874 | { |
| 6875 | size_t |
| 6876 | operator()(const u32string& __s) const noexcept |
| 6877 | { return std::_Hash_impl::hash(__s.data(), |
| 6878 | __s.length() * sizeof(char32_t)); } |
| 6879 | }; |
| 6880 | |
| 6881 | template<> |
| 6882 | struct __is_fast_hash<hash<u32string>> : std::false_type |
| 6883 | { }; |
| 6884 | |
| 6885 | #if __cplusplus201703L >= 201402L |
| 6886 | |
| 6887 | #define __cpp_lib_string_udls201304 201304 |
| 6888 | |
| 6889 | inline namespace literals |
| 6890 | { |
| 6891 | inline namespace string_literals |
| 6892 | { |
| 6893 | #pragma GCC diagnostic push |
| 6894 | #pragma GCC diagnostic ignored "-Wliteral-suffix" |
| 6895 | _GLIBCXX_DEFAULT_ABI_TAG__attribute ((__abi_tag__ ("cxx11"))) |
| 6896 | inline basic_string<char> |
| 6897 | operator""s(const char* __str, size_t __len) |
| 6898 | { return basic_string<char>{__str, __len}; } |
| 6899 | |
| 6900 | #ifdef _GLIBCXX_USE_WCHAR_T1 |
| 6901 | _GLIBCXX_DEFAULT_ABI_TAG__attribute ((__abi_tag__ ("cxx11"))) |
| 6902 | inline basic_string<wchar_t> |
| 6903 | operator""s(const wchar_t* __str, size_t __len) |
| 6904 | { return basic_string<wchar_t>{__str, __len}; } |
| 6905 | #endif |
| 6906 | |
| 6907 | #ifdef _GLIBCXX_USE_CHAR8_T |
| 6908 | _GLIBCXX_DEFAULT_ABI_TAG__attribute ((__abi_tag__ ("cxx11"))) |
| 6909 | inline basic_string<char8_t> |
| 6910 | operator""s(const char8_t* __str, size_t __len) |
| 6911 | { return basic_string<char8_t>{__str, __len}; } |
| 6912 | #endif |
| 6913 | |
| 6914 | _GLIBCXX_DEFAULT_ABI_TAG__attribute ((__abi_tag__ ("cxx11"))) |
| 6915 | inline basic_string<char16_t> |
| 6916 | operator""s(const char16_t* __str, size_t __len) |
| 6917 | { return basic_string<char16_t>{__str, __len}; } |
| 6918 | |
| 6919 | _GLIBCXX_DEFAULT_ABI_TAG__attribute ((__abi_tag__ ("cxx11"))) |
| 6920 | inline basic_string<char32_t> |
| 6921 | operator""s(const char32_t* __str, size_t __len) |
| 6922 | { return basic_string<char32_t>{__str, __len}; } |
| 6923 | |
| 6924 | #pragma GCC diagnostic pop |
| 6925 | } // inline namespace string_literals |
| 6926 | } // inline namespace literals |
| 6927 | |
| 6928 | #if __cplusplus201703L >= 201703L |
| 6929 | namespace __detail::__variant |
| 6930 | { |
| 6931 | template<typename> struct _Never_valueless_alt; // see <variant> |
| 6932 | |
| 6933 | // Provide the strong exception-safety guarantee when emplacing a |
| 6934 | // basic_string into a variant, but only if moving the string cannot throw. |
| 6935 | template<typename _Tp, typename _Traits, typename _Alloc> |
| 6936 | struct _Never_valueless_alt<std::basic_string<_Tp, _Traits, _Alloc>> |
| 6937 | : __and_< |
| 6938 | is_nothrow_move_constructible<std::basic_string<_Tp, _Traits, _Alloc>>, |
| 6939 | is_nothrow_move_assignable<std::basic_string<_Tp, _Traits, _Alloc>> |
| 6940 | >::type |
| 6941 | { }; |
| 6942 | } // namespace __detail::__variant |
| 6943 | #endif // C++17 |
| 6944 | #endif // C++14 |
| 6945 | |
| 6946 | _GLIBCXX_END_NAMESPACE_VERSION |
| 6947 | } // namespace std |
| 6948 | |
| 6949 | #endif // C++11 |
| 6950 | |
| 6951 | #endif /* _BASIC_STRING_H */ |
| 1 | //===- VPlan.h - Represent A Vectorizer Plan --------------------*- C++ -*-===// | ||||||||||||
| 2 | // | ||||||||||||
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | ||||||||||||
| 4 | // See https://llvm.org/LICENSE.txt for license information. | ||||||||||||
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | ||||||||||||
| 6 | // | ||||||||||||
| 7 | //===----------------------------------------------------------------------===// | ||||||||||||
| 8 | // | ||||||||||||
| 9 | /// \file | ||||||||||||
| 10 | /// This file contains the declarations of the Vectorization Plan base classes: | ||||||||||||
| 11 | /// 1. VPBasicBlock and VPRegionBlock that inherit from a common pure virtual | ||||||||||||
| 12 | /// VPBlockBase, together implementing a Hierarchical CFG; | ||||||||||||
| 13 | /// 2. Specializations of GraphTraits that allow VPBlockBase graphs to be | ||||||||||||
| 14 | /// treated as proper graphs for generic algorithms; | ||||||||||||
| 15 | /// 3. Pure virtual VPRecipeBase serving as the base class for recipes contained | ||||||||||||
| 16 | /// within VPBasicBlocks; | ||||||||||||
| 17 | /// 4. VPInstruction, a concrete Recipe and VPUser modeling a single planned | ||||||||||||
| 18 | /// instruction; | ||||||||||||
| 19 | /// 5. The VPlan class holding a candidate for vectorization; | ||||||||||||
| 20 | /// 6. The VPlanPrinter class providing a way to print a plan in dot format; | ||||||||||||
| 21 | /// These are documented in docs/VectorizationPlan.rst. | ||||||||||||
| 22 | // | ||||||||||||
| 23 | //===----------------------------------------------------------------------===// | ||||||||||||
| 24 | |||||||||||||
| 25 | #ifndef LLVM_TRANSFORMS_VECTORIZE_VPLAN_H | ||||||||||||
| 26 | #define LLVM_TRANSFORMS_VECTORIZE_VPLAN_H | ||||||||||||
| 27 | |||||||||||||
| 28 | #include "VPlanValue.h" | ||||||||||||
| 29 | #include "llvm/ADT/DenseMap.h" | ||||||||||||
| 30 | #include "llvm/ADT/DepthFirstIterator.h" | ||||||||||||
| 31 | #include "llvm/ADT/GraphTraits.h" | ||||||||||||
| 32 | #include "llvm/ADT/MapVector.h" | ||||||||||||
| 33 | #include "llvm/ADT/SmallBitVector.h" | ||||||||||||
| 34 | #include "llvm/ADT/SmallPtrSet.h" | ||||||||||||
| 35 | #include "llvm/ADT/SmallVector.h" | ||||||||||||
| 36 | #include "llvm/ADT/Twine.h" | ||||||||||||
| 37 | #include "llvm/ADT/ilist.h" | ||||||||||||
| 38 | #include "llvm/ADT/ilist_node.h" | ||||||||||||
| 39 | #include "llvm/Analysis/LoopInfo.h" | ||||||||||||
| 40 | #include "llvm/Analysis/VectorUtils.h" | ||||||||||||
| 41 | #include "llvm/IR/DebugLoc.h" | ||||||||||||
| 42 | #include "llvm/IR/FMF.h" | ||||||||||||
| 43 | #include "llvm/Transforms/Utils/LoopVersioning.h" | ||||||||||||
| 44 | #include <algorithm> | ||||||||||||
| 45 | #include <cassert> | ||||||||||||
| 46 | #include <cstddef> | ||||||||||||
| 47 | #include <string> | ||||||||||||
| 48 | |||||||||||||
| 49 | namespace llvm { | ||||||||||||
| 50 | |||||||||||||
| 51 | class BasicBlock; | ||||||||||||
| 52 | class DominatorTree; | ||||||||||||
| 53 | class InductionDescriptor; | ||||||||||||
| 54 | class InnerLoopVectorizer; | ||||||||||||
| 55 | class IRBuilderBase; | ||||||||||||
| 56 | class LoopInfo; | ||||||||||||
| 57 | class PredicateScalarEvolution; | ||||||||||||
| 58 | class raw_ostream; | ||||||||||||
| 59 | class RecurrenceDescriptor; | ||||||||||||
| 60 | class SCEV; | ||||||||||||
| 61 | class Type; | ||||||||||||
| 62 | class VPBasicBlock; | ||||||||||||
| 63 | class VPRegionBlock; | ||||||||||||
| 64 | class VPlan; | ||||||||||||
| 65 | class VPReplicateRecipe; | ||||||||||||
| 66 | class VPlanSlp; | ||||||||||||
| 67 | class Value; | ||||||||||||
| 68 | |||||||||||||
| 69 | namespace Intrinsic { | ||||||||||||
| 70 | typedef unsigned ID; | ||||||||||||
| 71 | } | ||||||||||||
| 72 | |||||||||||||
| 73 | /// Returns a calculation for the total number of elements for a given \p VF. | ||||||||||||
| 74 | /// For fixed width vectors this value is a constant, whereas for scalable | ||||||||||||
| 75 | /// vectors it is an expression determined at runtime. | ||||||||||||
| 76 | Value *getRuntimeVF(IRBuilderBase &B, Type *Ty, ElementCount VF); | ||||||||||||
| 77 | |||||||||||||
| 78 | /// Return a value for Step multiplied by VF. | ||||||||||||
| 79 | Value *createStepForVF(IRBuilderBase &B, Type *Ty, ElementCount VF, | ||||||||||||
| 80 | int64_t Step); | ||||||||||||
| 81 | |||||||||||||
| 82 | const SCEV *createTripCountSCEV(Type *IdxTy, PredicatedScalarEvolution &PSE); | ||||||||||||
| 83 | |||||||||||||
| 84 | /// A range of powers-of-2 vectorization factors with fixed start and | ||||||||||||
| 85 | /// adjustable end. The range includes start and excludes end, e.g.,: | ||||||||||||
| 86 | /// [1, 9) = {1, 2, 4, 8} | ||||||||||||
| 87 | struct VFRange { | ||||||||||||
| 88 | // A power of 2. | ||||||||||||
| 89 | const ElementCount Start; | ||||||||||||
| 90 | |||||||||||||
| 91 | // Need not be a power of 2. If End <= Start range is empty. | ||||||||||||
| 92 | ElementCount End; | ||||||||||||
| 93 | |||||||||||||
| 94 | bool isEmpty() const { | ||||||||||||
| 95 | return End.getKnownMinValue() <= Start.getKnownMinValue(); | ||||||||||||
| 96 | } | ||||||||||||
| 97 | |||||||||||||
| 98 | VFRange(const ElementCount &Start, const ElementCount &End) | ||||||||||||
| 99 | : Start(Start), End(End) { | ||||||||||||
| 100 | assert(Start.isScalable() == End.isScalable() &&(static_cast <bool> (Start.isScalable() == End.isScalable () && "Both Start and End should have the same scalable flag" ) ? void (0) : __assert_fail ("Start.isScalable() == End.isScalable() && \"Both Start and End should have the same scalable flag\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 101, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 101 | "Both Start and End should have the same scalable flag")(static_cast <bool> (Start.isScalable() == End.isScalable () && "Both Start and End should have the same scalable flag" ) ? void (0) : __assert_fail ("Start.isScalable() == End.isScalable() && \"Both Start and End should have the same scalable flag\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 101, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 102 | assert(isPowerOf2_32(Start.getKnownMinValue()) &&(static_cast <bool> (isPowerOf2_32(Start.getKnownMinValue ()) && "Expected Start to be a power of 2") ? void (0 ) : __assert_fail ("isPowerOf2_32(Start.getKnownMinValue()) && \"Expected Start to be a power of 2\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 103, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 103 | "Expected Start to be a power of 2")(static_cast <bool> (isPowerOf2_32(Start.getKnownMinValue ()) && "Expected Start to be a power of 2") ? void (0 ) : __assert_fail ("isPowerOf2_32(Start.getKnownMinValue()) && \"Expected Start to be a power of 2\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 103, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 104 | } | ||||||||||||
| 105 | }; | ||||||||||||
| 106 | |||||||||||||
| 107 | using VPlanPtr = std::unique_ptr<VPlan>; | ||||||||||||
| 108 | |||||||||||||
| 109 | /// In what follows, the term "input IR" refers to code that is fed into the | ||||||||||||
| 110 | /// vectorizer whereas the term "output IR" refers to code that is generated by | ||||||||||||
| 111 | /// the vectorizer. | ||||||||||||
| 112 | |||||||||||||
| 113 | /// VPLane provides a way to access lanes in both fixed width and scalable | ||||||||||||
| 114 | /// vectors, where for the latter the lane index sometimes needs calculating | ||||||||||||
| 115 | /// as a runtime expression. | ||||||||||||
| 116 | class VPLane { | ||||||||||||
| 117 | public: | ||||||||||||
| 118 | /// Kind describes how to interpret Lane. | ||||||||||||
| 119 | enum class Kind : uint8_t { | ||||||||||||
| 120 | /// For First, Lane is the index into the first N elements of a | ||||||||||||
| 121 | /// fixed-vector <N x <ElTy>> or a scalable vector <vscale x N x <ElTy>>. | ||||||||||||
| 122 | First, | ||||||||||||
| 123 | /// For ScalableLast, Lane is the offset from the start of the last | ||||||||||||
| 124 | /// N-element subvector in a scalable vector <vscale x N x <ElTy>>. For | ||||||||||||
| 125 | /// example, a Lane of 0 corresponds to lane `(vscale - 1) * N`, a Lane of | ||||||||||||
| 126 | /// 1 corresponds to `((vscale - 1) * N) + 1`, etc. | ||||||||||||
| 127 | ScalableLast | ||||||||||||
| 128 | }; | ||||||||||||
| 129 | |||||||||||||
| 130 | private: | ||||||||||||
| 131 | /// in [0..VF) | ||||||||||||
| 132 | unsigned Lane; | ||||||||||||
| 133 | |||||||||||||
| 134 | /// Indicates how the Lane should be interpreted, as described above. | ||||||||||||
| 135 | Kind LaneKind; | ||||||||||||
| 136 | |||||||||||||
| 137 | public: | ||||||||||||
| 138 | VPLane(unsigned Lane, Kind LaneKind) : Lane(Lane), LaneKind(LaneKind) {} | ||||||||||||
| 139 | |||||||||||||
| 140 | static VPLane getFirstLane() { return VPLane(0, VPLane::Kind::First); } | ||||||||||||
| 141 | |||||||||||||
| 142 | static VPLane getLastLaneForVF(const ElementCount &VF) { | ||||||||||||
| 143 | unsigned LaneOffset = VF.getKnownMinValue() - 1; | ||||||||||||
| 144 | Kind LaneKind; | ||||||||||||
| 145 | if (VF.isScalable()) | ||||||||||||
| 146 | // In this case 'LaneOffset' refers to the offset from the start of the | ||||||||||||
| 147 | // last subvector with VF.getKnownMinValue() elements. | ||||||||||||
| 148 | LaneKind = VPLane::Kind::ScalableLast; | ||||||||||||
| 149 | else | ||||||||||||
| 150 | LaneKind = VPLane::Kind::First; | ||||||||||||
| 151 | return VPLane(LaneOffset, LaneKind); | ||||||||||||
| 152 | } | ||||||||||||
| 153 | |||||||||||||
| 154 | /// Returns a compile-time known value for the lane index and asserts if the | ||||||||||||
| 155 | /// lane can only be calculated at runtime. | ||||||||||||
| 156 | unsigned getKnownLane() const { | ||||||||||||
| 157 | assert(LaneKind == Kind::First)(static_cast <bool> (LaneKind == Kind::First) ? void (0 ) : __assert_fail ("LaneKind == Kind::First", "llvm/lib/Transforms/Vectorize/VPlan.h" , 157, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 158 | return Lane; | ||||||||||||
| 159 | } | ||||||||||||
| 160 | |||||||||||||
| 161 | /// Returns an expression describing the lane index that can be used at | ||||||||||||
| 162 | /// runtime. | ||||||||||||
| 163 | Value *getAsRuntimeExpr(IRBuilderBase &Builder, const ElementCount &VF) const; | ||||||||||||
| 164 | |||||||||||||
| 165 | /// Returns the Kind of lane offset. | ||||||||||||
| 166 | Kind getKind() const { return LaneKind; } | ||||||||||||
| 167 | |||||||||||||
| 168 | /// Returns true if this is the first lane of the whole vector. | ||||||||||||
| 169 | bool isFirstLane() const { return Lane == 0 && LaneKind == Kind::First; } | ||||||||||||
| 170 | |||||||||||||
| 171 | /// Maps the lane to a cache index based on \p VF. | ||||||||||||
| 172 | unsigned mapToCacheIndex(const ElementCount &VF) const { | ||||||||||||
| 173 | switch (LaneKind) { | ||||||||||||
| 174 | case VPLane::Kind::ScalableLast: | ||||||||||||
| 175 | assert(VF.isScalable() && Lane < VF.getKnownMinValue())(static_cast <bool> (VF.isScalable() && Lane < VF.getKnownMinValue()) ? void (0) : __assert_fail ("VF.isScalable() && Lane < VF.getKnownMinValue()" , "llvm/lib/Transforms/Vectorize/VPlan.h", 175, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 176 | return VF.getKnownMinValue() + Lane; | ||||||||||||
| 177 | default: | ||||||||||||
| 178 | assert(Lane < VF.getKnownMinValue())(static_cast <bool> (Lane < VF.getKnownMinValue()) ? void (0) : __assert_fail ("Lane < VF.getKnownMinValue()", "llvm/lib/Transforms/Vectorize/VPlan.h", 178, __extension__ __PRETTY_FUNCTION__ )); | ||||||||||||
| 179 | return Lane; | ||||||||||||
| 180 | } | ||||||||||||
| 181 | } | ||||||||||||
| 182 | |||||||||||||
| 183 | /// Returns the maxmimum number of lanes that we are able to consider | ||||||||||||
| 184 | /// caching for \p VF. | ||||||||||||
| 185 | static unsigned getNumCachedLanes(const ElementCount &VF) { | ||||||||||||
| 186 | return VF.getKnownMinValue() * (VF.isScalable() ? 2 : 1); | ||||||||||||
| 187 | } | ||||||||||||
| 188 | }; | ||||||||||||
| 189 | |||||||||||||
| 190 | /// VPIteration represents a single point in the iteration space of the output | ||||||||||||
| 191 | /// (vectorized and/or unrolled) IR loop. | ||||||||||||
| 192 | struct VPIteration { | ||||||||||||
| 193 | /// in [0..UF) | ||||||||||||
| 194 | unsigned Part; | ||||||||||||
| 195 | |||||||||||||
| 196 | VPLane Lane; | ||||||||||||
| 197 | |||||||||||||
| 198 | VPIteration(unsigned Part, unsigned Lane, | ||||||||||||
| 199 | VPLane::Kind Kind = VPLane::Kind::First) | ||||||||||||
| 200 | : Part(Part), Lane(Lane, Kind) {} | ||||||||||||
| 201 | |||||||||||||
| 202 | VPIteration(unsigned Part, const VPLane &Lane) : Part(Part), Lane(Lane) {} | ||||||||||||
| 203 | |||||||||||||
| 204 | bool isFirstIteration() const { return Part == 0 && Lane.isFirstLane(); } | ||||||||||||
| 205 | }; | ||||||||||||
| 206 | |||||||||||||
| 207 | /// VPTransformState holds information passed down when "executing" a VPlan, | ||||||||||||
| 208 | /// needed for generating the output IR. | ||||||||||||
| 209 | struct VPTransformState { | ||||||||||||
| 210 | VPTransformState(ElementCount VF, unsigned UF, LoopInfo *LI, | ||||||||||||
| 211 | DominatorTree *DT, IRBuilderBase &Builder, | ||||||||||||
| 212 | InnerLoopVectorizer *ILV, VPlan *Plan) | ||||||||||||
| 213 | : VF(VF), UF(UF), LI(LI), DT(DT), Builder(Builder), ILV(ILV), Plan(Plan), | ||||||||||||
| 214 | LVer(nullptr) {} | ||||||||||||
| 215 | |||||||||||||
| 216 | /// The chosen Vectorization and Unroll Factors of the loop being vectorized. | ||||||||||||
| 217 | ElementCount VF; | ||||||||||||
| 218 | unsigned UF; | ||||||||||||
| 219 | |||||||||||||
| 220 | /// Hold the indices to generate specific scalar instructions. Null indicates | ||||||||||||
| 221 | /// that all instances are to be generated, using either scalar or vector | ||||||||||||
| 222 | /// instructions. | ||||||||||||
| 223 | std::optional<VPIteration> Instance; | ||||||||||||
| 224 | |||||||||||||
| 225 | struct DataState { | ||||||||||||
| 226 | /// A type for vectorized values in the new loop. Each value from the | ||||||||||||
| 227 | /// original loop, when vectorized, is represented by UF vector values in | ||||||||||||
| 228 | /// the new unrolled loop, where UF is the unroll factor. | ||||||||||||
| 229 | typedef SmallVector<Value *, 2> PerPartValuesTy; | ||||||||||||
| 230 | |||||||||||||
| 231 | DenseMap<VPValue *, PerPartValuesTy> PerPartOutput; | ||||||||||||
| 232 | |||||||||||||
| 233 | using ScalarsPerPartValuesTy = SmallVector<SmallVector<Value *, 4>, 2>; | ||||||||||||
| 234 | DenseMap<VPValue *, ScalarsPerPartValuesTy> PerPartScalars; | ||||||||||||
| 235 | } Data; | ||||||||||||
| 236 | |||||||||||||
| 237 | /// Get the generated Value for a given VPValue and a given Part. Note that | ||||||||||||
| 238 | /// as some Defs are still created by ILV and managed in its ValueMap, this | ||||||||||||
| 239 | /// method will delegate the call to ILV in such cases in order to provide | ||||||||||||
| 240 | /// callers a consistent API. | ||||||||||||
| 241 | /// \see set. | ||||||||||||
| 242 | Value *get(VPValue *Def, unsigned Part); | ||||||||||||
| 243 | |||||||||||||
| 244 | /// Get the generated Value for a given VPValue and given Part and Lane. | ||||||||||||
| 245 | Value *get(VPValue *Def, const VPIteration &Instance); | ||||||||||||
| 246 | |||||||||||||
| 247 | bool hasVectorValue(VPValue *Def, unsigned Part) { | ||||||||||||
| 248 | auto I = Data.PerPartOutput.find(Def); | ||||||||||||
| 249 | return I != Data.PerPartOutput.end() && Part < I->second.size() && | ||||||||||||
| 250 | I->second[Part]; | ||||||||||||
| 251 | } | ||||||||||||
| 252 | |||||||||||||
| 253 | bool hasAnyVectorValue(VPValue *Def) const { | ||||||||||||
| 254 | return Data.PerPartOutput.find(Def) != Data.PerPartOutput.end(); | ||||||||||||
| 255 | } | ||||||||||||
| 256 | |||||||||||||
| 257 | bool hasScalarValue(VPValue *Def, VPIteration Instance) { | ||||||||||||
| 258 | auto I = Data.PerPartScalars.find(Def); | ||||||||||||
| 259 | if (I == Data.PerPartScalars.end()) | ||||||||||||
| 260 | return false; | ||||||||||||
| 261 | unsigned CacheIdx = Instance.Lane.mapToCacheIndex(VF); | ||||||||||||
| 262 | return Instance.Part < I->second.size() && | ||||||||||||
| 263 | CacheIdx < I->second[Instance.Part].size() && | ||||||||||||
| 264 | I->second[Instance.Part][CacheIdx]; | ||||||||||||
| 265 | } | ||||||||||||
| 266 | |||||||||||||
| 267 | /// Set the generated Value for a given VPValue and a given Part. | ||||||||||||
| 268 | void set(VPValue *Def, Value *V, unsigned Part) { | ||||||||||||
| 269 | if (!Data.PerPartOutput.count(Def)) { | ||||||||||||
| 270 | DataState::PerPartValuesTy Entry(UF); | ||||||||||||
| 271 | Data.PerPartOutput[Def] = Entry; | ||||||||||||
| 272 | } | ||||||||||||
| 273 | Data.PerPartOutput[Def][Part] = V; | ||||||||||||
| 274 | } | ||||||||||||
| 275 | /// Reset an existing vector value for \p Def and a given \p Part. | ||||||||||||
| 276 | void reset(VPValue *Def, Value *V, unsigned Part) { | ||||||||||||
| 277 | auto Iter = Data.PerPartOutput.find(Def); | ||||||||||||
| 278 | assert(Iter != Data.PerPartOutput.end() &&(static_cast <bool> (Iter != Data.PerPartOutput.end() && "need to overwrite existing value") ? void (0) : __assert_fail ("Iter != Data.PerPartOutput.end() && \"need to overwrite existing value\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 279, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 279 | "need to overwrite existing value")(static_cast <bool> (Iter != Data.PerPartOutput.end() && "need to overwrite existing value") ? void (0) : __assert_fail ("Iter != Data.PerPartOutput.end() && \"need to overwrite existing value\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 279, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 280 | Iter->second[Part] = V; | ||||||||||||
| 281 | } | ||||||||||||
| 282 | |||||||||||||
| 283 | /// Set the generated scalar \p V for \p Def and the given \p Instance. | ||||||||||||
| 284 | void set(VPValue *Def, Value *V, const VPIteration &Instance) { | ||||||||||||
| 285 | auto Iter = Data.PerPartScalars.insert({Def, {}}); | ||||||||||||
| 286 | auto &PerPartVec = Iter.first->second; | ||||||||||||
| 287 | while (PerPartVec.size() <= Instance.Part) | ||||||||||||
| 288 | PerPartVec.emplace_back(); | ||||||||||||
| 289 | auto &Scalars = PerPartVec[Instance.Part]; | ||||||||||||
| 290 | unsigned CacheIdx = Instance.Lane.mapToCacheIndex(VF); | ||||||||||||
| 291 | while (Scalars.size() <= CacheIdx) | ||||||||||||
| 292 | Scalars.push_back(nullptr); | ||||||||||||
| 293 | assert(!Scalars[CacheIdx] && "should overwrite existing value")(static_cast <bool> (!Scalars[CacheIdx] && "should overwrite existing value" ) ? void (0) : __assert_fail ("!Scalars[CacheIdx] && \"should overwrite existing value\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 293, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 294 | Scalars[CacheIdx] = V; | ||||||||||||
| 295 | } | ||||||||||||
| 296 | |||||||||||||
| 297 | /// Reset an existing scalar value for \p Def and a given \p Instance. | ||||||||||||
| 298 | void reset(VPValue *Def, Value *V, const VPIteration &Instance) { | ||||||||||||
| 299 | auto Iter = Data.PerPartScalars.find(Def); | ||||||||||||
| 300 | assert(Iter != Data.PerPartScalars.end() &&(static_cast <bool> (Iter != Data.PerPartScalars.end() && "need to overwrite existing value") ? void (0) : __assert_fail ("Iter != Data.PerPartScalars.end() && \"need to overwrite existing value\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 301, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 301 | "need to overwrite existing value")(static_cast <bool> (Iter != Data.PerPartScalars.end() && "need to overwrite existing value") ? void (0) : __assert_fail ("Iter != Data.PerPartScalars.end() && \"need to overwrite existing value\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 301, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 302 | assert(Instance.Part < Iter->second.size() &&(static_cast <bool> (Instance.Part < Iter->second .size() && "need to overwrite existing value") ? void (0) : __assert_fail ("Instance.Part < Iter->second.size() && \"need to overwrite existing value\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 303, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 303 | "need to overwrite existing value")(static_cast <bool> (Instance.Part < Iter->second .size() && "need to overwrite existing value") ? void (0) : __assert_fail ("Instance.Part < Iter->second.size() && \"need to overwrite existing value\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 303, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 304 | unsigned CacheIdx = Instance.Lane.mapToCacheIndex(VF); | ||||||||||||
| 305 | assert(CacheIdx < Iter->second[Instance.Part].size() &&(static_cast <bool> (CacheIdx < Iter->second[Instance .Part].size() && "need to overwrite existing value") ? void (0) : __assert_fail ("CacheIdx < Iter->second[Instance.Part].size() && \"need to overwrite existing value\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 306, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 306 | "need to overwrite existing value")(static_cast <bool> (CacheIdx < Iter->second[Instance .Part].size() && "need to overwrite existing value") ? void (0) : __assert_fail ("CacheIdx < Iter->second[Instance.Part].size() && \"need to overwrite existing value\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 306, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 307 | Iter->second[Instance.Part][CacheIdx] = V; | ||||||||||||
| 308 | } | ||||||||||||
| 309 | |||||||||||||
| 310 | /// Add additional metadata to \p To that was not present on \p Orig. | ||||||||||||
| 311 | /// | ||||||||||||
| 312 | /// Currently this is used to add the noalias annotations based on the | ||||||||||||
| 313 | /// inserted memchecks. Use this for instructions that are *cloned* into the | ||||||||||||
| 314 | /// vector loop. | ||||||||||||
| 315 | void addNewMetadata(Instruction *To, const Instruction *Orig); | ||||||||||||
| 316 | |||||||||||||
| 317 | /// Add metadata from one instruction to another. | ||||||||||||
| 318 | /// | ||||||||||||
| 319 | /// This includes both the original MDs from \p From and additional ones (\see | ||||||||||||
| 320 | /// addNewMetadata). Use this for *newly created* instructions in the vector | ||||||||||||
| 321 | /// loop. | ||||||||||||
| 322 | void addMetadata(Instruction *To, Instruction *From); | ||||||||||||
| 323 | |||||||||||||
| 324 | /// Similar to the previous function but it adds the metadata to a | ||||||||||||
| 325 | /// vector of instructions. | ||||||||||||
| 326 | void addMetadata(ArrayRef<Value *> To, Instruction *From); | ||||||||||||
| 327 | |||||||||||||
| 328 | /// Set the debug location in the builder using the debug location in \p V. | ||||||||||||
| 329 | void setDebugLocFromInst(const Value *V); | ||||||||||||
| 330 | |||||||||||||
| 331 | /// Hold state information used when constructing the CFG of the output IR, | ||||||||||||
| 332 | /// traversing the VPBasicBlocks and generating corresponding IR BasicBlocks. | ||||||||||||
| 333 | struct CFGState { | ||||||||||||
| 334 | /// The previous VPBasicBlock visited. Initially set to null. | ||||||||||||
| 335 | VPBasicBlock *PrevVPBB = nullptr; | ||||||||||||
| 336 | |||||||||||||
| 337 | /// The previous IR BasicBlock created or used. Initially set to the new | ||||||||||||
| 338 | /// header BasicBlock. | ||||||||||||
| 339 | BasicBlock *PrevBB = nullptr; | ||||||||||||
| 340 | |||||||||||||
| 341 | /// The last IR BasicBlock in the output IR. Set to the exit block of the | ||||||||||||
| 342 | /// vector loop. | ||||||||||||
| 343 | BasicBlock *ExitBB = nullptr; | ||||||||||||
| 344 | |||||||||||||
| 345 | /// A mapping of each VPBasicBlock to the corresponding BasicBlock. In case | ||||||||||||
| 346 | /// of replication, maps the BasicBlock of the last replica created. | ||||||||||||
| 347 | SmallDenseMap<VPBasicBlock *, BasicBlock *> VPBB2IRBB; | ||||||||||||
| 348 | |||||||||||||
| 349 | CFGState() = default; | ||||||||||||
| 350 | |||||||||||||
| 351 | /// Returns the BasicBlock* mapped to the pre-header of the loop region | ||||||||||||
| 352 | /// containing \p R. | ||||||||||||
| 353 | BasicBlock *getPreheaderBBFor(VPRecipeBase *R); | ||||||||||||
| 354 | } CFG; | ||||||||||||
| 355 | |||||||||||||
| 356 | /// Hold a pointer to LoopInfo to register new basic blocks in the loop. | ||||||||||||
| 357 | LoopInfo *LI; | ||||||||||||
| 358 | |||||||||||||
| 359 | /// Hold a pointer to Dominator Tree to register new basic blocks in the loop. | ||||||||||||
| 360 | DominatorTree *DT; | ||||||||||||
| 361 | |||||||||||||
| 362 | /// Hold a reference to the IRBuilder used to generate output IR code. | ||||||||||||
| 363 | IRBuilderBase &Builder; | ||||||||||||
| 364 | |||||||||||||
| 365 | VPValue2ValueTy VPValue2Value; | ||||||||||||
| 366 | |||||||||||||
| 367 | /// Hold the canonical scalar IV of the vector loop (start=0, step=VF*UF). | ||||||||||||
| 368 | Value *CanonicalIV = nullptr; | ||||||||||||
| 369 | |||||||||||||
| 370 | /// Hold a pointer to InnerLoopVectorizer to reuse its IR generation methods. | ||||||||||||
| 371 | InnerLoopVectorizer *ILV; | ||||||||||||
| 372 | |||||||||||||
| 373 | /// Pointer to the VPlan code is generated for. | ||||||||||||
| 374 | VPlan *Plan; | ||||||||||||
| 375 | |||||||||||||
| 376 | /// Holds recipes that may generate a poison value that is used after | ||||||||||||
| 377 | /// vectorization, even when their operands are not poison. | ||||||||||||
| 378 | SmallPtrSet<VPRecipeBase *, 16> MayGeneratePoisonRecipes; | ||||||||||||
| 379 | |||||||||||||
| 380 | /// The loop object for the current parent region, or nullptr. | ||||||||||||
| 381 | Loop *CurrentVectorLoop = nullptr; | ||||||||||||
| 382 | |||||||||||||
| 383 | /// LoopVersioning. It's only set up (non-null) if memchecks were | ||||||||||||
| 384 | /// used. | ||||||||||||
| 385 | /// | ||||||||||||
| 386 | /// This is currently only used to add no-alias metadata based on the | ||||||||||||
| 387 | /// memchecks. The actually versioning is performed manually. | ||||||||||||
| 388 | std::unique_ptr<LoopVersioning> LVer; | ||||||||||||
| 389 | }; | ||||||||||||
| 390 | |||||||||||||
| 391 | /// VPBlockBase is the building block of the Hierarchical Control-Flow Graph. | ||||||||||||
| 392 | /// A VPBlockBase can be either a VPBasicBlock or a VPRegionBlock. | ||||||||||||
| 393 | class VPBlockBase { | ||||||||||||
| 394 | friend class VPBlockUtils; | ||||||||||||
| 395 | |||||||||||||
| 396 | const unsigned char SubclassID; ///< Subclass identifier (for isa/dyn_cast). | ||||||||||||
| 397 | |||||||||||||
| 398 | /// An optional name for the block. | ||||||||||||
| 399 | std::string Name; | ||||||||||||
| 400 | |||||||||||||
| 401 | /// The immediate VPRegionBlock which this VPBlockBase belongs to, or null if | ||||||||||||
| 402 | /// it is a topmost VPBlockBase. | ||||||||||||
| 403 | VPRegionBlock *Parent = nullptr; | ||||||||||||
| 404 | |||||||||||||
| 405 | /// List of predecessor blocks. | ||||||||||||
| 406 | SmallVector<VPBlockBase *, 1> Predecessors; | ||||||||||||
| 407 | |||||||||||||
| 408 | /// List of successor blocks. | ||||||||||||
| 409 | SmallVector<VPBlockBase *, 1> Successors; | ||||||||||||
| 410 | |||||||||||||
| 411 | /// VPlan containing the block. Can only be set on the entry block of the | ||||||||||||
| 412 | /// plan. | ||||||||||||
| 413 | VPlan *Plan = nullptr; | ||||||||||||
| 414 | |||||||||||||
| 415 | /// Add \p Successor as the last successor to this block. | ||||||||||||
| 416 | void appendSuccessor(VPBlockBase *Successor) { | ||||||||||||
| 417 | assert(Successor && "Cannot add nullptr successor!")(static_cast <bool> (Successor && "Cannot add nullptr successor!" ) ? void (0) : __assert_fail ("Successor && \"Cannot add nullptr successor!\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 417, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 418 | Successors.push_back(Successor); | ||||||||||||
| 419 | } | ||||||||||||
| 420 | |||||||||||||
| 421 | /// Add \p Predecessor as the last predecessor to this block. | ||||||||||||
| 422 | void appendPredecessor(VPBlockBase *Predecessor) { | ||||||||||||
| 423 | assert(Predecessor && "Cannot add nullptr predecessor!")(static_cast <bool> (Predecessor && "Cannot add nullptr predecessor!" ) ? void (0) : __assert_fail ("Predecessor && \"Cannot add nullptr predecessor!\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 423, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 424 | Predecessors.push_back(Predecessor); | ||||||||||||
| 425 | } | ||||||||||||
| 426 | |||||||||||||
| 427 | /// Remove \p Predecessor from the predecessors of this block. | ||||||||||||
| 428 | void removePredecessor(VPBlockBase *Predecessor) { | ||||||||||||
| 429 | auto Pos = find(Predecessors, Predecessor); | ||||||||||||
| 430 | assert(Pos && "Predecessor does not exist")(static_cast <bool> (Pos && "Predecessor does not exist" ) ? void (0) : __assert_fail ("Pos && \"Predecessor does not exist\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 430, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 431 | Predecessors.erase(Pos); | ||||||||||||
| 432 | } | ||||||||||||
| 433 | |||||||||||||
| 434 | /// Remove \p Successor from the successors of this block. | ||||||||||||
| 435 | void removeSuccessor(VPBlockBase *Successor) { | ||||||||||||
| 436 | auto Pos = find(Successors, Successor); | ||||||||||||
| 437 | assert(Pos && "Successor does not exist")(static_cast <bool> (Pos && "Successor does not exist" ) ? void (0) : __assert_fail ("Pos && \"Successor does not exist\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 437, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 438 | Successors.erase(Pos); | ||||||||||||
| 439 | } | ||||||||||||
| 440 | |||||||||||||
| 441 | protected: | ||||||||||||
| 442 | VPBlockBase(const unsigned char SC, const std::string &N) | ||||||||||||
| 443 | : SubclassID(SC), Name(N) {} | ||||||||||||
| 444 | |||||||||||||
| 445 | public: | ||||||||||||
| 446 | /// An enumeration for keeping track of the concrete subclass of VPBlockBase | ||||||||||||
| 447 | /// that are actually instantiated. Values of this enumeration are kept in the | ||||||||||||
| 448 | /// SubclassID field of the VPBlockBase objects. They are used for concrete | ||||||||||||
| 449 | /// type identification. | ||||||||||||
| 450 | using VPBlockTy = enum { VPBasicBlockSC, VPRegionBlockSC }; | ||||||||||||
| 451 | |||||||||||||
| 452 | using VPBlocksTy = SmallVectorImpl<VPBlockBase *>; | ||||||||||||
| 453 | |||||||||||||
| 454 | virtual ~VPBlockBase() = default; | ||||||||||||
| 455 | |||||||||||||
| 456 | const std::string &getName() const { return Name; } | ||||||||||||
| 457 | |||||||||||||
| 458 | void setName(const Twine &newName) { Name = newName.str(); } | ||||||||||||
| 459 | |||||||||||||
| 460 | /// \return an ID for the concrete type of this object. | ||||||||||||
| 461 | /// This is used to implement the classof checks. This should not be used | ||||||||||||
| 462 | /// for any other purpose, as the values may change as LLVM evolves. | ||||||||||||
| 463 | unsigned getVPBlockID() const { return SubclassID; } | ||||||||||||
| 464 | |||||||||||||
| 465 | VPRegionBlock *getParent() { return Parent; } | ||||||||||||
| 466 | const VPRegionBlock *getParent() const { return Parent; } | ||||||||||||
| 467 | |||||||||||||
| 468 | /// \return A pointer to the plan containing the current block. | ||||||||||||
| 469 | VPlan *getPlan(); | ||||||||||||
| 470 | const VPlan *getPlan() const; | ||||||||||||
| 471 | |||||||||||||
| 472 | /// Sets the pointer of the plan containing the block. The block must be the | ||||||||||||
| 473 | /// entry block into the VPlan. | ||||||||||||
| 474 | void setPlan(VPlan *ParentPlan); | ||||||||||||
| 475 | |||||||||||||
| 476 | void setParent(VPRegionBlock *P) { Parent = P; } | ||||||||||||
| 477 | |||||||||||||
| 478 | /// \return the VPBasicBlock that is the entry of this VPBlockBase, | ||||||||||||
| 479 | /// recursively, if the latter is a VPRegionBlock. Otherwise, if this | ||||||||||||
| 480 | /// VPBlockBase is a VPBasicBlock, it is returned. | ||||||||||||
| 481 | const VPBasicBlock *getEntryBasicBlock() const; | ||||||||||||
| 482 | VPBasicBlock *getEntryBasicBlock(); | ||||||||||||
| 483 | |||||||||||||
| 484 | /// \return the VPBasicBlock that is the exiting this VPBlockBase, | ||||||||||||
| 485 | /// recursively, if the latter is a VPRegionBlock. Otherwise, if this | ||||||||||||
| 486 | /// VPBlockBase is a VPBasicBlock, it is returned. | ||||||||||||
| 487 | const VPBasicBlock *getExitingBasicBlock() const; | ||||||||||||
| 488 | VPBasicBlock *getExitingBasicBlock(); | ||||||||||||
| 489 | |||||||||||||
| 490 | const VPBlocksTy &getSuccessors() const { return Successors; } | ||||||||||||
| 491 | VPBlocksTy &getSuccessors() { return Successors; } | ||||||||||||
| 492 | |||||||||||||
| 493 | iterator_range<VPBlockBase **> successors() { return Successors; } | ||||||||||||
| 494 | |||||||||||||
| 495 | const VPBlocksTy &getPredecessors() const { return Predecessors; } | ||||||||||||
| 496 | VPBlocksTy &getPredecessors() { return Predecessors; } | ||||||||||||
| 497 | |||||||||||||
| 498 | /// \return the successor of this VPBlockBase if it has a single successor. | ||||||||||||
| 499 | /// Otherwise return a null pointer. | ||||||||||||
| 500 | VPBlockBase *getSingleSuccessor() const { | ||||||||||||
| 501 | return (Successors.size() == 1 ? *Successors.begin() : nullptr); | ||||||||||||
| 502 | } | ||||||||||||
| 503 | |||||||||||||
| 504 | /// \return the predecessor of this VPBlockBase if it has a single | ||||||||||||
| 505 | /// predecessor. Otherwise return a null pointer. | ||||||||||||
| 506 | VPBlockBase *getSinglePredecessor() const { | ||||||||||||
| 507 | return (Predecessors.size() == 1 ? *Predecessors.begin() : nullptr); | ||||||||||||
| 508 | } | ||||||||||||
| 509 | |||||||||||||
| 510 | size_t getNumSuccessors() const { return Successors.size(); } | ||||||||||||
| 511 | size_t getNumPredecessors() const { return Predecessors.size(); } | ||||||||||||
| 512 | |||||||||||||
| 513 | /// An Enclosing Block of a block B is any block containing B, including B | ||||||||||||
| 514 | /// itself. \return the closest enclosing block starting from "this", which | ||||||||||||
| 515 | /// has successors. \return the root enclosing block if all enclosing blocks | ||||||||||||
| 516 | /// have no successors. | ||||||||||||
| 517 | VPBlockBase *getEnclosingBlockWithSuccessors(); | ||||||||||||
| 518 | |||||||||||||
| 519 | /// \return the closest enclosing block starting from "this", which has | ||||||||||||
| 520 | /// predecessors. \return the root enclosing block if all enclosing blocks | ||||||||||||
| 521 | /// have no predecessors. | ||||||||||||
| 522 | VPBlockBase *getEnclosingBlockWithPredecessors(); | ||||||||||||
| 523 | |||||||||||||
| 524 | /// \return the successors either attached directly to this VPBlockBase or, if | ||||||||||||
| 525 | /// this VPBlockBase is the exit block of a VPRegionBlock and has no | ||||||||||||
| 526 | /// successors of its own, search recursively for the first enclosing | ||||||||||||
| 527 | /// VPRegionBlock that has successors and return them. If no such | ||||||||||||
| 528 | /// VPRegionBlock exists, return the (empty) successors of the topmost | ||||||||||||
| 529 | /// VPBlockBase reached. | ||||||||||||
| 530 | const VPBlocksTy &getHierarchicalSuccessors() { | ||||||||||||
| 531 | return getEnclosingBlockWithSuccessors()->getSuccessors(); | ||||||||||||
| 532 | } | ||||||||||||
| 533 | |||||||||||||
| 534 | /// \return the hierarchical successor of this VPBlockBase if it has a single | ||||||||||||
| 535 | /// hierarchical successor. Otherwise return a null pointer. | ||||||||||||
| 536 | VPBlockBase *getSingleHierarchicalSuccessor() { | ||||||||||||
| 537 | return getEnclosingBlockWithSuccessors()->getSingleSuccessor(); | ||||||||||||
| 538 | } | ||||||||||||
| 539 | |||||||||||||
| 540 | /// \return the predecessors either attached directly to this VPBlockBase or, | ||||||||||||
| 541 | /// if this VPBlockBase is the entry block of a VPRegionBlock and has no | ||||||||||||
| 542 | /// predecessors of its own, search recursively for the first enclosing | ||||||||||||
| 543 | /// VPRegionBlock that has predecessors and return them. If no such | ||||||||||||
| 544 | /// VPRegionBlock exists, return the (empty) predecessors of the topmost | ||||||||||||
| 545 | /// VPBlockBase reached. | ||||||||||||
| 546 | const VPBlocksTy &getHierarchicalPredecessors() { | ||||||||||||
| 547 | return getEnclosingBlockWithPredecessors()->getPredecessors(); | ||||||||||||
| 548 | } | ||||||||||||
| 549 | |||||||||||||
| 550 | /// \return the hierarchical predecessor of this VPBlockBase if it has a | ||||||||||||
| 551 | /// single hierarchical predecessor. Otherwise return a null pointer. | ||||||||||||
| 552 | VPBlockBase *getSingleHierarchicalPredecessor() { | ||||||||||||
| 553 | return getEnclosingBlockWithPredecessors()->getSinglePredecessor(); | ||||||||||||
| 554 | } | ||||||||||||
| 555 | |||||||||||||
| 556 | /// Set a given VPBlockBase \p Successor as the single successor of this | ||||||||||||
| 557 | /// VPBlockBase. This VPBlockBase is not added as predecessor of \p Successor. | ||||||||||||
| 558 | /// This VPBlockBase must have no successors. | ||||||||||||
| 559 | void setOneSuccessor(VPBlockBase *Successor) { | ||||||||||||
| 560 | assert(Successors.empty() && "Setting one successor when others exist.")(static_cast <bool> (Successors.empty() && "Setting one successor when others exist." ) ? void (0) : __assert_fail ("Successors.empty() && \"Setting one successor when others exist.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 560, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 561 | appendSuccessor(Successor); | ||||||||||||
| 562 | } | ||||||||||||
| 563 | |||||||||||||
| 564 | /// Set two given VPBlockBases \p IfTrue and \p IfFalse to be the two | ||||||||||||
| 565 | /// successors of this VPBlockBase. This VPBlockBase is not added as | ||||||||||||
| 566 | /// predecessor of \p IfTrue or \p IfFalse. This VPBlockBase must have no | ||||||||||||
| 567 | /// successors. | ||||||||||||
| 568 | void setTwoSuccessors(VPBlockBase *IfTrue, VPBlockBase *IfFalse) { | ||||||||||||
| 569 | assert(Successors.empty() && "Setting two successors when others exist.")(static_cast <bool> (Successors.empty() && "Setting two successors when others exist." ) ? void (0) : __assert_fail ("Successors.empty() && \"Setting two successors when others exist.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 569, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 570 | appendSuccessor(IfTrue); | ||||||||||||
| 571 | appendSuccessor(IfFalse); | ||||||||||||
| 572 | } | ||||||||||||
| 573 | |||||||||||||
| 574 | /// Set each VPBasicBlock in \p NewPreds as predecessor of this VPBlockBase. | ||||||||||||
| 575 | /// This VPBlockBase must have no predecessors. This VPBlockBase is not added | ||||||||||||
| 576 | /// as successor of any VPBasicBlock in \p NewPreds. | ||||||||||||
| 577 | void setPredecessors(ArrayRef<VPBlockBase *> NewPreds) { | ||||||||||||
| 578 | assert(Predecessors.empty() && "Block predecessors already set.")(static_cast <bool> (Predecessors.empty() && "Block predecessors already set." ) ? void (0) : __assert_fail ("Predecessors.empty() && \"Block predecessors already set.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 578, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 579 | for (auto *Pred : NewPreds) | ||||||||||||
| 580 | appendPredecessor(Pred); | ||||||||||||
| 581 | } | ||||||||||||
| 582 | |||||||||||||
| 583 | /// Remove all the predecessor of this block. | ||||||||||||
| 584 | void clearPredecessors() { Predecessors.clear(); } | ||||||||||||
| 585 | |||||||||||||
| 586 | /// Remove all the successors of this block. | ||||||||||||
| 587 | void clearSuccessors() { Successors.clear(); } | ||||||||||||
| 588 | |||||||||||||
| 589 | /// The method which generates the output IR that correspond to this | ||||||||||||
| 590 | /// VPBlockBase, thereby "executing" the VPlan. | ||||||||||||
| 591 | virtual void execute(VPTransformState *State) = 0; | ||||||||||||
| 592 | |||||||||||||
| 593 | /// Delete all blocks reachable from a given VPBlockBase, inclusive. | ||||||||||||
| 594 | static void deleteCFG(VPBlockBase *Entry); | ||||||||||||
| 595 | |||||||||||||
| 596 | /// Return true if it is legal to hoist instructions into this block. | ||||||||||||
| 597 | bool isLegalToHoistInto() { | ||||||||||||
| 598 | // There are currently no constraints that prevent an instruction to be | ||||||||||||
| 599 | // hoisted into a VPBlockBase. | ||||||||||||
| 600 | return true; | ||||||||||||
| 601 | } | ||||||||||||
| 602 | |||||||||||||
| 603 | /// Replace all operands of VPUsers in the block with \p NewValue and also | ||||||||||||
| 604 | /// replaces all uses of VPValues defined in the block with NewValue. | ||||||||||||
| 605 | virtual void dropAllReferences(VPValue *NewValue) = 0; | ||||||||||||
| 606 | |||||||||||||
| 607 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 608 | void printAsOperand(raw_ostream &OS, bool PrintType) const { | ||||||||||||
| 609 | OS << getName(); | ||||||||||||
| 610 | } | ||||||||||||
| 611 | |||||||||||||
| 612 | /// Print plain-text dump of this VPBlockBase to \p O, prefixing all lines | ||||||||||||
| 613 | /// with \p Indent. \p SlotTracker is used to print unnamed VPValue's using | ||||||||||||
| 614 | /// consequtive numbers. | ||||||||||||
| 615 | /// | ||||||||||||
| 616 | /// Note that the numbering is applied to the whole VPlan, so printing | ||||||||||||
| 617 | /// individual blocks is consistent with the whole VPlan printing. | ||||||||||||
| 618 | virtual void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 619 | VPSlotTracker &SlotTracker) const = 0; | ||||||||||||
| 620 | |||||||||||||
| 621 | /// Print plain-text dump of this VPlan to \p O. | ||||||||||||
| 622 | void print(raw_ostream &O) const { | ||||||||||||
| 623 | VPSlotTracker SlotTracker(getPlan()); | ||||||||||||
| 624 | print(O, "", SlotTracker); | ||||||||||||
| 625 | } | ||||||||||||
| 626 | |||||||||||||
| 627 | /// Print the successors of this block to \p O, prefixing all lines with \p | ||||||||||||
| 628 | /// Indent. | ||||||||||||
| 629 | void printSuccessors(raw_ostream &O, const Twine &Indent) const; | ||||||||||||
| 630 | |||||||||||||
| 631 | /// Dump this VPBlockBase to dbgs(). | ||||||||||||
| 632 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { print(dbgs()); } | ||||||||||||
| 633 | #endif | ||||||||||||
| 634 | }; | ||||||||||||
| 635 | |||||||||||||
| 636 | /// A value that is used outside the VPlan. The operand of the user needs to be | ||||||||||||
| 637 | /// added to the associated LCSSA phi node. | ||||||||||||
| 638 | class VPLiveOut : public VPUser { | ||||||||||||
| 639 | PHINode *Phi; | ||||||||||||
| 640 | |||||||||||||
| 641 | public: | ||||||||||||
| 642 | VPLiveOut(PHINode *Phi, VPValue *Op) | ||||||||||||
| 643 | : VPUser({Op}, VPUser::VPUserID::LiveOut), Phi(Phi) {} | ||||||||||||
| 644 | |||||||||||||
| 645 | /// Fixup the wrapped LCSSA phi node in the unique exit block. This simply | ||||||||||||
| 646 | /// means we need to add the appropriate incoming value from the middle | ||||||||||||
| 647 | /// block as exiting edges from the scalar epilogue loop (if present) are | ||||||||||||
| 648 | /// already in place, and we exit the vector loop exclusively to the middle | ||||||||||||
| 649 | /// block. | ||||||||||||
| 650 | void fixPhi(VPlan &Plan, VPTransformState &State); | ||||||||||||
| 651 | |||||||||||||
| 652 | /// Returns true if the VPLiveOut uses scalars of operand \p Op. | ||||||||||||
| 653 | bool usesScalars(const VPValue *Op) const override { | ||||||||||||
| 654 | assert(is_contained(operands(), Op) &&(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 655, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 655 | "Op must be an operand of the recipe")(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 655, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 656 | return true; | ||||||||||||
| 657 | } | ||||||||||||
| 658 | |||||||||||||
| 659 | PHINode *getPhi() const { return Phi; } | ||||||||||||
| 660 | }; | ||||||||||||
| 661 | |||||||||||||
| 662 | /// VPRecipeBase is a base class modeling a sequence of one or more output IR | ||||||||||||
| 663 | /// instructions. VPRecipeBase owns the the VPValues it defines through VPDef | ||||||||||||
| 664 | /// and is responsible for deleting its defined values. Single-value | ||||||||||||
| 665 | /// VPRecipeBases that also inherit from VPValue must make sure to inherit from | ||||||||||||
| 666 | /// VPRecipeBase before VPValue. | ||||||||||||
| 667 | class VPRecipeBase : public ilist_node_with_parent<VPRecipeBase, VPBasicBlock>, | ||||||||||||
| 668 | public VPDef, | ||||||||||||
| 669 | public VPUser { | ||||||||||||
| 670 | friend VPBasicBlock; | ||||||||||||
| 671 | friend class VPBlockUtils; | ||||||||||||
| 672 | |||||||||||||
| 673 | /// Each VPRecipe belongs to a single VPBasicBlock. | ||||||||||||
| 674 | VPBasicBlock *Parent = nullptr; | ||||||||||||
| 675 | |||||||||||||
| 676 | public: | ||||||||||||
| 677 | VPRecipeBase(const unsigned char SC, ArrayRef<VPValue *> Operands) | ||||||||||||
| 678 | : VPDef(SC), VPUser(Operands, VPUser::VPUserID::Recipe) {} | ||||||||||||
| 679 | |||||||||||||
| 680 | template <typename IterT> | ||||||||||||
| 681 | VPRecipeBase(const unsigned char SC, iterator_range<IterT> Operands) | ||||||||||||
| 682 | : VPDef(SC), VPUser(Operands, VPUser::VPUserID::Recipe) {} | ||||||||||||
| 683 | virtual ~VPRecipeBase() = default; | ||||||||||||
| 684 | |||||||||||||
| 685 | /// \return the VPBasicBlock which this VPRecipe belongs to. | ||||||||||||
| 686 | VPBasicBlock *getParent() { return Parent; } | ||||||||||||
| 687 | const VPBasicBlock *getParent() const { return Parent; } | ||||||||||||
| 688 | |||||||||||||
| 689 | /// The method which generates the output IR instructions that correspond to | ||||||||||||
| 690 | /// this VPRecipe, thereby "executing" the VPlan. | ||||||||||||
| 691 | virtual void execute(VPTransformState &State) = 0; | ||||||||||||
| 692 | |||||||||||||
| 693 | /// Insert an unlinked recipe into a basic block immediately before | ||||||||||||
| 694 | /// the specified recipe. | ||||||||||||
| 695 | void insertBefore(VPRecipeBase *InsertPos); | ||||||||||||
| 696 | /// Insert an unlinked recipe into \p BB immediately before the insertion | ||||||||||||
| 697 | /// point \p IP; | ||||||||||||
| 698 | void insertBefore(VPBasicBlock &BB, iplist<VPRecipeBase>::iterator IP); | ||||||||||||
| 699 | |||||||||||||
| 700 | /// Insert an unlinked Recipe into a basic block immediately after | ||||||||||||
| 701 | /// the specified Recipe. | ||||||||||||
| 702 | void insertAfter(VPRecipeBase *InsertPos); | ||||||||||||
| 703 | |||||||||||||
| 704 | /// Unlink this recipe from its current VPBasicBlock and insert it into | ||||||||||||
| 705 | /// the VPBasicBlock that MovePos lives in, right after MovePos. | ||||||||||||
| 706 | void moveAfter(VPRecipeBase *MovePos); | ||||||||||||
| 707 | |||||||||||||
| 708 | /// Unlink this recipe and insert into BB before I. | ||||||||||||
| 709 | /// | ||||||||||||
| 710 | /// \pre I is a valid iterator into BB. | ||||||||||||
| 711 | void moveBefore(VPBasicBlock &BB, iplist<VPRecipeBase>::iterator I); | ||||||||||||
| 712 | |||||||||||||
| 713 | /// This method unlinks 'this' from the containing basic block, but does not | ||||||||||||
| 714 | /// delete it. | ||||||||||||
| 715 | void removeFromParent(); | ||||||||||||
| 716 | |||||||||||||
| 717 | /// This method unlinks 'this' from the containing basic block and deletes it. | ||||||||||||
| 718 | /// | ||||||||||||
| 719 | /// \returns an iterator pointing to the element after the erased one | ||||||||||||
| 720 | iplist<VPRecipeBase>::iterator eraseFromParent(); | ||||||||||||
| 721 | |||||||||||||
| 722 | /// Returns the underlying instruction, if the recipe is a VPValue or nullptr | ||||||||||||
| 723 | /// otherwise. | ||||||||||||
| 724 | Instruction *getUnderlyingInstr() { | ||||||||||||
| 725 | return cast<Instruction>(getVPSingleValue()->getUnderlyingValue()); | ||||||||||||
| 726 | } | ||||||||||||
| 727 | const Instruction *getUnderlyingInstr() const { | ||||||||||||
| 728 | return cast<Instruction>(getVPSingleValue()->getUnderlyingValue()); | ||||||||||||
| 729 | } | ||||||||||||
| 730 | |||||||||||||
| 731 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
| 732 | static inline bool classof(const VPDef *D) { | ||||||||||||
| 733 | // All VPDefs are also VPRecipeBases. | ||||||||||||
| 734 | return true; | ||||||||||||
| 735 | } | ||||||||||||
| 736 | |||||||||||||
| 737 | static inline bool classof(const VPUser *U) { | ||||||||||||
| 738 | return U->getVPUserID() == VPUser::VPUserID::Recipe; | ||||||||||||
| 739 | } | ||||||||||||
| 740 | |||||||||||||
| 741 | /// Returns true if the recipe may have side-effects. | ||||||||||||
| 742 | bool mayHaveSideEffects() const; | ||||||||||||
| 743 | |||||||||||||
| 744 | /// Returns true for PHI-like recipes. | ||||||||||||
| 745 | bool isPhi() const { | ||||||||||||
| 746 | return getVPDefID() >= VPFirstPHISC && getVPDefID() <= VPLastPHISC; | ||||||||||||
| 747 | } | ||||||||||||
| 748 | |||||||||||||
| 749 | /// Returns true if the recipe may read from memory. | ||||||||||||
| 750 | bool mayReadFromMemory() const; | ||||||||||||
| 751 | |||||||||||||
| 752 | /// Returns true if the recipe may write to memory. | ||||||||||||
| 753 | bool mayWriteToMemory() const; | ||||||||||||
| 754 | |||||||||||||
| 755 | /// Returns true if the recipe may read from or write to memory. | ||||||||||||
| 756 | bool mayReadOrWriteMemory() const { | ||||||||||||
| 757 | return mayReadFromMemory() || mayWriteToMemory(); | ||||||||||||
| 758 | } | ||||||||||||
| 759 | }; | ||||||||||||
| 760 | |||||||||||||
| 761 | inline bool VPUser::classof(const VPDef *Def) { | ||||||||||||
| 762 | return Def->getVPDefID() == VPRecipeBase::VPInstructionSC || | ||||||||||||
| 763 | Def->getVPDefID() == VPRecipeBase::VPWidenSC || | ||||||||||||
| 764 | Def->getVPDefID() == VPRecipeBase::VPWidenCallSC || | ||||||||||||
| 765 | Def->getVPDefID() == VPRecipeBase::VPWidenSelectSC || | ||||||||||||
| 766 | Def->getVPDefID() == VPRecipeBase::VPWidenGEPSC || | ||||||||||||
| 767 | Def->getVPDefID() == VPRecipeBase::VPBlendSC || | ||||||||||||
| 768 | Def->getVPDefID() == VPRecipeBase::VPInterleaveSC || | ||||||||||||
| 769 | Def->getVPDefID() == VPRecipeBase::VPReplicateSC || | ||||||||||||
| 770 | Def->getVPDefID() == VPRecipeBase::VPReductionSC || | ||||||||||||
| 771 | Def->getVPDefID() == VPRecipeBase::VPBranchOnMaskSC || | ||||||||||||
| 772 | Def->getVPDefID() == VPRecipeBase::VPWidenMemoryInstructionSC; | ||||||||||||
| 773 | } | ||||||||||||
| 774 | |||||||||||||
| 775 | // Helper macro to define common classof implementations for recipes. | ||||||||||||
| 776 | #define VP_CLASSOF_IMPL(VPDefID, VPValueID)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPDefID; } static inline bool classof(const VPValue *V) { return V->getVPValueID() == VPValueID; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast<VPRecipeBase >(U); return R && R->getVPDefID() == VPDefID; } static inline bool classof(const VPRecipeBase *R) { return R ->getVPDefID() == VPDefID; } \ | ||||||||||||
| 777 | static inline bool classof(const VPDef *D) { \ | ||||||||||||
| 778 | return D->getVPDefID() == VPDefID; \ | ||||||||||||
| 779 | } \ | ||||||||||||
| 780 | static inline bool classof(const VPValue *V) { \ | ||||||||||||
| 781 | return V->getVPValueID() == VPValueID; \ | ||||||||||||
| 782 | } \ | ||||||||||||
| 783 | static inline bool classof(const VPUser *U) { \ | ||||||||||||
| 784 | auto *R = dyn_cast<VPRecipeBase>(U); \ | ||||||||||||
| 785 | return R && R->getVPDefID() == VPDefID; \ | ||||||||||||
| 786 | } \ | ||||||||||||
| 787 | static inline bool classof(const VPRecipeBase *R) { \ | ||||||||||||
| 788 | return R->getVPDefID() == VPDefID; \ | ||||||||||||
| 789 | } | ||||||||||||
| 790 | |||||||||||||
| 791 | /// This is a concrete Recipe that models a single VPlan-level instruction. | ||||||||||||
| 792 | /// While as any Recipe it may generate a sequence of IR instructions when | ||||||||||||
| 793 | /// executed, these instructions would always form a single-def expression as | ||||||||||||
| 794 | /// the VPInstruction is also a single def-use vertex. | ||||||||||||
| 795 | class VPInstruction : public VPRecipeBase, public VPValue { | ||||||||||||
| 796 | friend class VPlanSlp; | ||||||||||||
| 797 | |||||||||||||
| 798 | public: | ||||||||||||
| 799 | /// VPlan opcodes, extending LLVM IR with idiomatics instructions. | ||||||||||||
| 800 | enum { | ||||||||||||
| 801 | FirstOrderRecurrenceSplice = | ||||||||||||
| 802 | Instruction::OtherOpsEnd + 1, // Combines the incoming and previous | ||||||||||||
| 803 | // values of a first-order recurrence. | ||||||||||||
| 804 | Not, | ||||||||||||
| 805 | ICmpULE, | ||||||||||||
| 806 | SLPLoad, | ||||||||||||
| 807 | SLPStore, | ||||||||||||
| 808 | ActiveLaneMask, | ||||||||||||
| 809 | CanonicalIVIncrement, | ||||||||||||
| 810 | CanonicalIVIncrementNUW, | ||||||||||||
| 811 | // The next two are similar to the above, but instead increment the | ||||||||||||
| 812 | // canonical IV separately for each unrolled part. | ||||||||||||
| 813 | CanonicalIVIncrementForPart, | ||||||||||||
| 814 | CanonicalIVIncrementForPartNUW, | ||||||||||||
| 815 | BranchOnCount, | ||||||||||||
| 816 | BranchOnCond | ||||||||||||
| 817 | }; | ||||||||||||
| 818 | |||||||||||||
| 819 | private: | ||||||||||||
| 820 | typedef unsigned char OpcodeTy; | ||||||||||||
| 821 | OpcodeTy Opcode; | ||||||||||||
| 822 | FastMathFlags FMF; | ||||||||||||
| 823 | DebugLoc DL; | ||||||||||||
| 824 | |||||||||||||
| 825 | /// An optional name that can be used for the generated IR instruction. | ||||||||||||
| 826 | const std::string Name; | ||||||||||||
| 827 | |||||||||||||
| 828 | /// Utility method serving execute(): generates a single instance of the | ||||||||||||
| 829 | /// modeled instruction. | ||||||||||||
| 830 | void generateInstruction(VPTransformState &State, unsigned Part); | ||||||||||||
| 831 | |||||||||||||
| 832 | protected: | ||||||||||||
| 833 | void setUnderlyingInstr(Instruction *I) { setUnderlyingValue(I); } | ||||||||||||
| 834 | |||||||||||||
| 835 | public: | ||||||||||||
| 836 | VPInstruction(unsigned Opcode, ArrayRef<VPValue *> Operands, DebugLoc DL, | ||||||||||||
| 837 | const Twine &Name = "") | ||||||||||||
| 838 | : VPRecipeBase(VPRecipeBase::VPInstructionSC, Operands), | ||||||||||||
| 839 | VPValue(VPValue::VPVInstructionSC, nullptr, this), Opcode(Opcode), | ||||||||||||
| 840 | DL(DL), Name(Name.str()) {} | ||||||||||||
| 841 | |||||||||||||
| 842 | VPInstruction(unsigned Opcode, std::initializer_list<VPValue *> Operands, | ||||||||||||
| 843 | DebugLoc DL = {}, const Twine &Name = "") | ||||||||||||
| 844 | : VPInstruction(Opcode, ArrayRef<VPValue *>(Operands), DL, Name) {} | ||||||||||||
| 845 | |||||||||||||
| 846 | VP_CLASSOF_IMPL(VPRecipeBase::VPInstructionSC, VPValue::VPVInstructionSC)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPInstructionSC; } static inline bool classof (const VPValue *V) { return V->getVPValueID() == VPValue:: VPVInstructionSC; } static inline bool classof(const VPUser * U) { auto *R = dyn_cast<VPRecipeBase>(U); return R && R->getVPDefID() == VPRecipeBase::VPInstructionSC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID () == VPRecipeBase::VPInstructionSC; } | ||||||||||||
| 847 | |||||||||||||
| 848 | VPInstruction *clone() const { | ||||||||||||
| 849 | SmallVector<VPValue *, 2> Operands(operands()); | ||||||||||||
| 850 | return new VPInstruction(Opcode, Operands, DL, Name); | ||||||||||||
| 851 | } | ||||||||||||
| 852 | |||||||||||||
| 853 | unsigned getOpcode() const { return Opcode; } | ||||||||||||
| 854 | |||||||||||||
| 855 | /// Generate the instruction. | ||||||||||||
| 856 | /// TODO: We currently execute only per-part unless a specific instance is | ||||||||||||
| 857 | /// provided. | ||||||||||||
| 858 | void execute(VPTransformState &State) override; | ||||||||||||
| 859 | |||||||||||||
| 860 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 861 | /// Print the VPInstruction to \p O. | ||||||||||||
| 862 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 863 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 864 | |||||||||||||
| 865 | /// Print the VPInstruction to dbgs() (for debugging). | ||||||||||||
| 866 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const; | ||||||||||||
| 867 | #endif | ||||||||||||
| 868 | |||||||||||||
| 869 | /// Return true if this instruction may modify memory. | ||||||||||||
| 870 | bool mayWriteToMemory() const { | ||||||||||||
| 871 | // TODO: we can use attributes of the called function to rule out memory | ||||||||||||
| 872 | // modifications. | ||||||||||||
| 873 | return Opcode == Instruction::Store || Opcode == Instruction::Call || | ||||||||||||
| 874 | Opcode == Instruction::Invoke || Opcode == SLPStore; | ||||||||||||
| 875 | } | ||||||||||||
| 876 | |||||||||||||
| 877 | bool hasResult() const { | ||||||||||||
| 878 | // CallInst may or may not have a result, depending on the called function. | ||||||||||||
| 879 | // Conservatively return calls have results for now. | ||||||||||||
| 880 | switch (getOpcode()) { | ||||||||||||
| 881 | case Instruction::Ret: | ||||||||||||
| 882 | case Instruction::Br: | ||||||||||||
| 883 | case Instruction::Store: | ||||||||||||
| 884 | case Instruction::Switch: | ||||||||||||
| 885 | case Instruction::IndirectBr: | ||||||||||||
| 886 | case Instruction::Resume: | ||||||||||||
| 887 | case Instruction::CatchRet: | ||||||||||||
| 888 | case Instruction::Unreachable: | ||||||||||||
| 889 | case Instruction::Fence: | ||||||||||||
| 890 | case Instruction::AtomicRMW: | ||||||||||||
| 891 | case VPInstruction::BranchOnCond: | ||||||||||||
| 892 | case VPInstruction::BranchOnCount: | ||||||||||||
| 893 | return false; | ||||||||||||
| 894 | default: | ||||||||||||
| 895 | return true; | ||||||||||||
| 896 | } | ||||||||||||
| 897 | } | ||||||||||||
| 898 | |||||||||||||
| 899 | /// Set the fast-math flags. | ||||||||||||
| 900 | void setFastMathFlags(FastMathFlags FMFNew); | ||||||||||||
| 901 | |||||||||||||
| 902 | /// Returns true if the recipe only uses the first lane of operand \p Op. | ||||||||||||
| 903 | bool onlyFirstLaneUsed(const VPValue *Op) const override { | ||||||||||||
| 904 | assert(is_contained(operands(), Op) &&(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 905, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 905 | "Op must be an operand of the recipe")(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 905, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 906 | if (getOperand(0) != Op) | ||||||||||||
| 907 | return false; | ||||||||||||
| 908 | switch (getOpcode()) { | ||||||||||||
| 909 | default: | ||||||||||||
| 910 | return false; | ||||||||||||
| 911 | case VPInstruction::ActiveLaneMask: | ||||||||||||
| 912 | case VPInstruction::CanonicalIVIncrement: | ||||||||||||
| 913 | case VPInstruction::CanonicalIVIncrementNUW: | ||||||||||||
| 914 | case VPInstruction::CanonicalIVIncrementForPart: | ||||||||||||
| 915 | case VPInstruction::CanonicalIVIncrementForPartNUW: | ||||||||||||
| 916 | case VPInstruction::BranchOnCount: | ||||||||||||
| 917 | return true; | ||||||||||||
| 918 | }; | ||||||||||||
| 919 | llvm_unreachable("switch should return")::llvm::llvm_unreachable_internal("switch should return", "llvm/lib/Transforms/Vectorize/VPlan.h" , 919); | ||||||||||||
| 920 | } | ||||||||||||
| 921 | }; | ||||||||||||
| 922 | |||||||||||||
| 923 | /// VPWidenRecipe is a recipe for producing a copy of vector type its | ||||||||||||
| 924 | /// ingredient. This recipe covers most of the traditional vectorization cases | ||||||||||||
| 925 | /// where each ingredient transforms into a vectorized version of itself. | ||||||||||||
| 926 | class VPWidenRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
| 927 | public: | ||||||||||||
| 928 | template <typename IterT> | ||||||||||||
| 929 | VPWidenRecipe(Instruction &I, iterator_range<IterT> Operands) | ||||||||||||
| 930 | : VPRecipeBase(VPRecipeBase::VPWidenSC, Operands), | ||||||||||||
| 931 | VPValue(VPValue::VPVWidenSC, &I, this) {} | ||||||||||||
| 932 | |||||||||||||
| 933 | ~VPWidenRecipe() override = default; | ||||||||||||
| 934 | |||||||||||||
| 935 | VP_CLASSOF_IMPL(VPRecipeBase::VPWidenSC, VPValue::VPVWidenSC)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPWidenSC; } static inline bool classof(const VPValue *V) { return V->getVPValueID() == VPValue::VPVWidenSC ; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast <VPRecipeBase>(U); return R && R->getVPDefID () == VPRecipeBase::VPWidenSC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID() == VPRecipeBase ::VPWidenSC; } | ||||||||||||
| 936 | |||||||||||||
| 937 | /// Produce widened copies of all Ingredients. | ||||||||||||
| 938 | void execute(VPTransformState &State) override; | ||||||||||||
| 939 | |||||||||||||
| 940 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 941 | /// Print the recipe. | ||||||||||||
| 942 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 943 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 944 | #endif | ||||||||||||
| 945 | }; | ||||||||||||
| 946 | |||||||||||||
| 947 | /// A recipe for widening Call instructions. | ||||||||||||
| 948 | class VPWidenCallRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
| 949 | /// ID of the vector intrinsic to call when widening the call. If set the | ||||||||||||
| 950 | /// Intrinsic::not_intrinsic, a library call will be used instead. | ||||||||||||
| 951 | Intrinsic::ID VectorIntrinsicID; | ||||||||||||
| 952 | |||||||||||||
| 953 | public: | ||||||||||||
| 954 | template <typename IterT> | ||||||||||||
| 955 | VPWidenCallRecipe(CallInst &I, iterator_range<IterT> CallArguments, | ||||||||||||
| 956 | Intrinsic::ID VectorIntrinsicID) | ||||||||||||
| 957 | : VPRecipeBase(VPRecipeBase::VPWidenCallSC, CallArguments), | ||||||||||||
| 958 | VPValue(VPValue::VPVWidenCallSC, &I, this), | ||||||||||||
| 959 | VectorIntrinsicID(VectorIntrinsicID) {} | ||||||||||||
| 960 | |||||||||||||
| 961 | ~VPWidenCallRecipe() override = default; | ||||||||||||
| 962 | |||||||||||||
| 963 | VP_CLASSOF_IMPL(VPRecipeBase::VPWidenCallSC, VPValue::VPVWidenCallSC)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPWidenCallSC; } static inline bool classof (const VPValue *V) { return V->getVPValueID() == VPValue:: VPVWidenCallSC; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast<VPRecipeBase>(U); return R && R->getVPDefID() == VPRecipeBase::VPWidenCallSC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID () == VPRecipeBase::VPWidenCallSC; } | ||||||||||||
| 964 | |||||||||||||
| 965 | /// Produce a widened version of the call instruction. | ||||||||||||
| 966 | void execute(VPTransformState &State) override; | ||||||||||||
| 967 | |||||||||||||
| 968 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 969 | /// Print the recipe. | ||||||||||||
| 970 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 971 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 972 | #endif | ||||||||||||
| 973 | }; | ||||||||||||
| 974 | |||||||||||||
| 975 | /// A recipe for widening select instructions. | ||||||||||||
| 976 | class VPWidenSelectRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
| 977 | |||||||||||||
| 978 | /// Is the condition of the select loop invariant? | ||||||||||||
| 979 | bool InvariantCond; | ||||||||||||
| 980 | |||||||||||||
| 981 | public: | ||||||||||||
| 982 | template <typename IterT> | ||||||||||||
| 983 | VPWidenSelectRecipe(SelectInst &I, iterator_range<IterT> Operands, | ||||||||||||
| 984 | bool InvariantCond) | ||||||||||||
| 985 | : VPRecipeBase(VPRecipeBase::VPWidenSelectSC, Operands), | ||||||||||||
| 986 | VPValue(VPValue::VPVWidenSelectSC, &I, this), | ||||||||||||
| 987 | InvariantCond(InvariantCond) {} | ||||||||||||
| 988 | |||||||||||||
| 989 | ~VPWidenSelectRecipe() override = default; | ||||||||||||
| 990 | |||||||||||||
| 991 | VP_CLASSOF_IMPL(VPRecipeBase::VPWidenSelectSC, VPValue::VPVWidenSelectSC)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPWidenSelectSC; } static inline bool classof (const VPValue *V) { return V->getVPValueID() == VPValue:: VPVWidenSelectSC; } static inline bool classof(const VPUser * U) { auto *R = dyn_cast<VPRecipeBase>(U); return R && R->getVPDefID() == VPRecipeBase::VPWidenSelectSC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID () == VPRecipeBase::VPWidenSelectSC; } | ||||||||||||
| 992 | |||||||||||||
| 993 | /// Produce a widened version of the select instruction. | ||||||||||||
| 994 | void execute(VPTransformState &State) override; | ||||||||||||
| 995 | |||||||||||||
| 996 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 997 | /// Print the recipe. | ||||||||||||
| 998 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 999 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1000 | #endif | ||||||||||||
| 1001 | }; | ||||||||||||
| 1002 | |||||||||||||
| 1003 | /// A recipe for handling GEP instructions. | ||||||||||||
| 1004 | class VPWidenGEPRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
| 1005 | bool IsPtrLoopInvariant; | ||||||||||||
| 1006 | SmallBitVector IsIndexLoopInvariant; | ||||||||||||
| 1007 | |||||||||||||
| 1008 | public: | ||||||||||||
| 1009 | template <typename IterT> | ||||||||||||
| 1010 | VPWidenGEPRecipe(GetElementPtrInst *GEP, iterator_range<IterT> Operands) | ||||||||||||
| 1011 | : VPRecipeBase(VPRecipeBase::VPWidenGEPSC, Operands), | ||||||||||||
| 1012 | VPValue(VPWidenGEPSC, GEP, this), | ||||||||||||
| 1013 | IsIndexLoopInvariant(GEP->getNumIndices(), false) {} | ||||||||||||
| 1014 | |||||||||||||
| 1015 | template <typename IterT> | ||||||||||||
| 1016 | VPWidenGEPRecipe(GetElementPtrInst *GEP, iterator_range<IterT> Operands, | ||||||||||||
| 1017 | Loop *OrigLoop) | ||||||||||||
| 1018 | : VPRecipeBase(VPRecipeBase::VPWidenGEPSC, Operands), | ||||||||||||
| 1019 | VPValue(VPValue::VPVWidenGEPSC, GEP, this), | ||||||||||||
| 1020 | IsIndexLoopInvariant(GEP->getNumIndices(), false) { | ||||||||||||
| 1021 | IsPtrLoopInvariant = OrigLoop->isLoopInvariant(GEP->getPointerOperand()); | ||||||||||||
| 1022 | for (auto Index : enumerate(GEP->indices())) | ||||||||||||
| 1023 | IsIndexLoopInvariant[Index.index()] = | ||||||||||||
| 1024 | OrigLoop->isLoopInvariant(Index.value().get()); | ||||||||||||
| 1025 | } | ||||||||||||
| 1026 | ~VPWidenGEPRecipe() override = default; | ||||||||||||
| 1027 | |||||||||||||
| 1028 | VP_CLASSOF_IMPL(VPRecipeBase::VPWidenGEPSC, VPValue::VPVWidenGEPSC)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPWidenGEPSC; } static inline bool classof (const VPValue *V) { return V->getVPValueID() == VPValue:: VPVWidenGEPSC; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast<VPRecipeBase>(U); return R && R->getVPDefID() == VPRecipeBase::VPWidenGEPSC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID () == VPRecipeBase::VPWidenGEPSC; } | ||||||||||||
| 1029 | |||||||||||||
| 1030 | /// Generate the gep nodes. | ||||||||||||
| 1031 | void execute(VPTransformState &State) override; | ||||||||||||
| 1032 | |||||||||||||
| 1033 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1034 | /// Print the recipe. | ||||||||||||
| 1035 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1036 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1037 | #endif | ||||||||||||
| 1038 | }; | ||||||||||||
| 1039 | |||||||||||||
| 1040 | /// A recipe for handling phi nodes of integer and floating-point inductions, | ||||||||||||
| 1041 | /// producing their vector values. | ||||||||||||
| 1042 | class VPWidenIntOrFpInductionRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
| 1043 | PHINode *IV; | ||||||||||||
| 1044 | const InductionDescriptor &IndDesc; | ||||||||||||
| 1045 | bool NeedsVectorIV; | ||||||||||||
| 1046 | |||||||||||||
| 1047 | public: | ||||||||||||
| 1048 | VPWidenIntOrFpInductionRecipe(PHINode *IV, VPValue *Start, VPValue *Step, | ||||||||||||
| 1049 | const InductionDescriptor &IndDesc, | ||||||||||||
| 1050 | bool NeedsVectorIV) | ||||||||||||
| 1051 | : VPRecipeBase(VPWidenIntOrFpInductionSC, {Start, Step}), | ||||||||||||
| 1052 | VPValue(IV, this), IV(IV), IndDesc(IndDesc), | ||||||||||||
| 1053 | NeedsVectorIV(NeedsVectorIV) {} | ||||||||||||
| 1054 | |||||||||||||
| 1055 | VPWidenIntOrFpInductionRecipe(PHINode *IV, VPValue *Start, VPValue *Step, | ||||||||||||
| 1056 | const InductionDescriptor &IndDesc, | ||||||||||||
| 1057 | TruncInst *Trunc, bool NeedsVectorIV) | ||||||||||||
| 1058 | : VPRecipeBase(VPWidenIntOrFpInductionSC, {Start, Step}), | ||||||||||||
| 1059 | VPValue(Trunc, this), IV(IV), IndDesc(IndDesc), | ||||||||||||
| 1060 | NeedsVectorIV(NeedsVectorIV) {} | ||||||||||||
| 1061 | |||||||||||||
| 1062 | ~VPWidenIntOrFpInductionRecipe() override = default; | ||||||||||||
| 1063 | |||||||||||||
| 1064 | VP_CLASSOF_IMPL(VPRecipeBase::VPWidenIntOrFpInductionSC,static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPWidenIntOrFpInductionSC; } static inline bool classof(const VPValue *V) { return V->getVPValueID() == VPValue::VPVWidenIntOrFpInductionSC; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast<VPRecipeBase >(U); return R && R->getVPDefID() == VPRecipeBase ::VPWidenIntOrFpInductionSC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID() == VPRecipeBase ::VPWidenIntOrFpInductionSC; } | ||||||||||||
| 1065 | VPValue::VPVWidenIntOrFpInductionSC)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPWidenIntOrFpInductionSC; } static inline bool classof(const VPValue *V) { return V->getVPValueID() == VPValue::VPVWidenIntOrFpInductionSC; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast<VPRecipeBase >(U); return R && R->getVPDefID() == VPRecipeBase ::VPWidenIntOrFpInductionSC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID() == VPRecipeBase ::VPWidenIntOrFpInductionSC; } | ||||||||||||
| 1066 | |||||||||||||
| 1067 | /// Generate the vectorized and scalarized versions of the phi node as | ||||||||||||
| 1068 | /// needed by their users. | ||||||||||||
| 1069 | void execute(VPTransformState &State) override; | ||||||||||||
| 1070 | |||||||||||||
| 1071 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1072 | /// Print the recipe. | ||||||||||||
| 1073 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1074 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1075 | #endif | ||||||||||||
| 1076 | |||||||||||||
| 1077 | /// Returns the start value of the induction. | ||||||||||||
| 1078 | VPValue *getStartValue() { return getOperand(0); } | ||||||||||||
| 1079 | const VPValue *getStartValue() const { return getOperand(0); } | ||||||||||||
| 1080 | |||||||||||||
| 1081 | /// Returns the step value of the induction. | ||||||||||||
| 1082 | VPValue *getStepValue() { return getOperand(1); } | ||||||||||||
| 1083 | const VPValue *getStepValue() const { return getOperand(1); } | ||||||||||||
| 1084 | |||||||||||||
| 1085 | /// Returns the first defined value as TruncInst, if it is one or nullptr | ||||||||||||
| 1086 | /// otherwise. | ||||||||||||
| 1087 | TruncInst *getTruncInst() { | ||||||||||||
| 1088 | return dyn_cast_or_null<TruncInst>(getVPValue(0)->getUnderlyingValue()); | ||||||||||||
| 1089 | } | ||||||||||||
| 1090 | const TruncInst *getTruncInst() const { | ||||||||||||
| 1091 | return dyn_cast_or_null<TruncInst>(getVPValue(0)->getUnderlyingValue()); | ||||||||||||
| 1092 | } | ||||||||||||
| 1093 | |||||||||||||
| 1094 | PHINode *getPHINode() { return IV; } | ||||||||||||
| 1095 | |||||||||||||
| 1096 | /// Returns the induction descriptor for the recipe. | ||||||||||||
| 1097 | const InductionDescriptor &getInductionDescriptor() const { return IndDesc; } | ||||||||||||
| 1098 | |||||||||||||
| 1099 | /// Returns true if the induction is canonical, i.e. starting at 0 and | ||||||||||||
| 1100 | /// incremented by UF * VF (= the original IV is incremented by 1). | ||||||||||||
| 1101 | bool isCanonical() const; | ||||||||||||
| 1102 | |||||||||||||
| 1103 | /// Returns the scalar type of the induction. | ||||||||||||
| 1104 | const Type *getScalarType() const { | ||||||||||||
| 1105 | const TruncInst *TruncI = getTruncInst(); | ||||||||||||
| 1106 | return TruncI ? TruncI->getType() : IV->getType(); | ||||||||||||
| 1107 | } | ||||||||||||
| 1108 | |||||||||||||
| 1109 | /// Returns true if a vector phi needs to be created for the induction. | ||||||||||||
| 1110 | bool needsVectorIV() const { return NeedsVectorIV; } | ||||||||||||
| 1111 | }; | ||||||||||||
| 1112 | |||||||||||||
| 1113 | /// A pure virtual base class for all recipes modeling header phis, including | ||||||||||||
| 1114 | /// phis for first order recurrences, pointer inductions and reductions. The | ||||||||||||
| 1115 | /// start value is the first operand of the recipe and the incoming value from | ||||||||||||
| 1116 | /// the backedge is the second operand. | ||||||||||||
| 1117 | /// | ||||||||||||
| 1118 | /// Inductions are modeled using the following sub-classes: | ||||||||||||
| 1119 | /// * VPCanonicalIVPHIRecipe: Canonical scalar induction of the vector loop, | ||||||||||||
| 1120 | /// starting at a specified value (zero for the main vector loop, the resume | ||||||||||||
| 1121 | /// value for the epilogue vector loop) and stepping by 1. The induction | ||||||||||||
| 1122 | /// controls exiting of the vector loop by comparing against the vector trip | ||||||||||||
| 1123 | /// count. Produces a single scalar PHI for the induction value per | ||||||||||||
| 1124 | /// iteration. | ||||||||||||
| 1125 | /// * VPWidenIntOrFpInductionRecipe: Generates vector values for integer and | ||||||||||||
| 1126 | /// floating point inductions with arbitrary start and step values. Produces | ||||||||||||
| 1127 | /// a vector PHI per-part. | ||||||||||||
| 1128 | /// * VPDerivedIVRecipe: Converts the canonical IV value to the corresponding | ||||||||||||
| 1129 | /// value of an IV with different start and step values. Produces a single | ||||||||||||
| 1130 | /// scalar value per iteration | ||||||||||||
| 1131 | /// * VPScalarIVStepsRecipe: Generates scalar values per-lane based on a | ||||||||||||
| 1132 | /// canonical or derived induction. | ||||||||||||
| 1133 | /// * VPWidenPointerInductionRecipe: Generate vector and scalar values for a | ||||||||||||
| 1134 | /// pointer induction. Produces either a vector PHI per-part or scalar values | ||||||||||||
| 1135 | /// per-lane based on the canonical induction. | ||||||||||||
| 1136 | class VPHeaderPHIRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
| 1137 | protected: | ||||||||||||
| 1138 | VPHeaderPHIRecipe(unsigned char VPVID, unsigned char VPDefID, PHINode *Phi, | ||||||||||||
| 1139 | VPValue *Start = nullptr) | ||||||||||||
| 1140 | : VPRecipeBase(VPDefID, {}), VPValue(VPVID, Phi, this) { | ||||||||||||
| 1141 | if (Start) | ||||||||||||
| 1142 | addOperand(Start); | ||||||||||||
| 1143 | } | ||||||||||||
| 1144 | |||||||||||||
| 1145 | public: | ||||||||||||
| 1146 | ~VPHeaderPHIRecipe() override = default; | ||||||||||||
| 1147 | |||||||||||||
| 1148 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
| 1149 | static inline bool classof(const VPRecipeBase *B) { | ||||||||||||
| 1150 | return B->getVPDefID() >= VPRecipeBase::VPFirstHeaderPHISC && | ||||||||||||
| 1151 | B->getVPDefID() <= VPRecipeBase::VPLastPHISC; | ||||||||||||
| 1152 | } | ||||||||||||
| 1153 | static inline bool classof(const VPValue *V) { | ||||||||||||
| 1154 | return V->getVPValueID() >= VPValue::VPVFirstHeaderPHISC && | ||||||||||||
| 1155 | V->getVPValueID() <= VPValue::VPVLastPHISC; | ||||||||||||
| 1156 | } | ||||||||||||
| 1157 | |||||||||||||
| 1158 | /// Generate the phi nodes. | ||||||||||||
| 1159 | void execute(VPTransformState &State) override = 0; | ||||||||||||
| 1160 | |||||||||||||
| 1161 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1162 | /// Print the recipe. | ||||||||||||
| 1163 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1164 | VPSlotTracker &SlotTracker) const override = 0; | ||||||||||||
| 1165 | #endif | ||||||||||||
| 1166 | |||||||||||||
| 1167 | /// Returns the start value of the phi, if one is set. | ||||||||||||
| 1168 | VPValue *getStartValue() { | ||||||||||||
| 1169 | return getNumOperands() == 0 ? nullptr : getOperand(0); | ||||||||||||
| 1170 | } | ||||||||||||
| 1171 | VPValue *getStartValue() const { | ||||||||||||
| 1172 | return getNumOperands() == 0 ? nullptr : getOperand(0); | ||||||||||||
| 1173 | } | ||||||||||||
| 1174 | |||||||||||||
| 1175 | /// Update the start value of the recipe. | ||||||||||||
| 1176 | void setStartValue(VPValue *V) { setOperand(0, V); } | ||||||||||||
| 1177 | |||||||||||||
| 1178 | /// Returns the incoming value from the loop backedge. | ||||||||||||
| 1179 | VPValue *getBackedgeValue() { | ||||||||||||
| 1180 | return getOperand(1); | ||||||||||||
| 1181 | } | ||||||||||||
| 1182 | |||||||||||||
| 1183 | /// Returns the backedge value as a recipe. The backedge value is guaranteed | ||||||||||||
| 1184 | /// to be a recipe. | ||||||||||||
| 1185 | VPRecipeBase &getBackedgeRecipe() { | ||||||||||||
| 1186 | return *getBackedgeValue()->getDefiningRecipe(); | ||||||||||||
| 1187 | } | ||||||||||||
| 1188 | }; | ||||||||||||
| 1189 | |||||||||||||
| 1190 | class VPWidenPointerInductionRecipe : public VPHeaderPHIRecipe { | ||||||||||||
| 1191 | const InductionDescriptor &IndDesc; | ||||||||||||
| 1192 | |||||||||||||
| 1193 | bool IsScalarAfterVectorization; | ||||||||||||
| 1194 | |||||||||||||
| 1195 | public: | ||||||||||||
| 1196 | /// Create a new VPWidenPointerInductionRecipe for \p Phi with start value \p | ||||||||||||
| 1197 | /// Start. | ||||||||||||
| 1198 | VPWidenPointerInductionRecipe(PHINode *Phi, VPValue *Start, VPValue *Step, | ||||||||||||
| 1199 | const InductionDescriptor &IndDesc, | ||||||||||||
| 1200 | bool IsScalarAfterVectorization) | ||||||||||||
| 1201 | : VPHeaderPHIRecipe(VPVWidenPointerInductionSC, VPWidenPointerInductionSC, | ||||||||||||
| 1202 | Phi), | ||||||||||||
| 1203 | IndDesc(IndDesc), | ||||||||||||
| 1204 | IsScalarAfterVectorization(IsScalarAfterVectorization) { | ||||||||||||
| 1205 | addOperand(Start); | ||||||||||||
| 1206 | addOperand(Step); | ||||||||||||
| 1207 | } | ||||||||||||
| 1208 | |||||||||||||
| 1209 | ~VPWidenPointerInductionRecipe() override = default; | ||||||||||||
| 1210 | |||||||||||||
| 1211 | VP_CLASSOF_IMPL(VPRecipeBase::VPWidenPointerInductionSC,static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPWidenPointerInductionSC; } static inline bool classof(const VPValue *V) { return V->getVPValueID() == VPValue::VPVWidenPointerInductionSC; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast<VPRecipeBase >(U); return R && R->getVPDefID() == VPRecipeBase ::VPWidenPointerInductionSC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID() == VPRecipeBase ::VPWidenPointerInductionSC; } | ||||||||||||
| 1212 | VPValue::VPVWidenPointerInductionSC)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPWidenPointerInductionSC; } static inline bool classof(const VPValue *V) { return V->getVPValueID() == VPValue::VPVWidenPointerInductionSC; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast<VPRecipeBase >(U); return R && R->getVPDefID() == VPRecipeBase ::VPWidenPointerInductionSC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID() == VPRecipeBase ::VPWidenPointerInductionSC; } | ||||||||||||
| 1213 | |||||||||||||
| 1214 | /// Generate vector values for the pointer induction. | ||||||||||||
| 1215 | void execute(VPTransformState &State) override; | ||||||||||||
| 1216 | |||||||||||||
| 1217 | /// Returns true if only scalar values will be generated. | ||||||||||||
| 1218 | bool onlyScalarsGenerated(ElementCount VF); | ||||||||||||
| 1219 | |||||||||||||
| 1220 | /// Returns the induction descriptor for the recipe. | ||||||||||||
| 1221 | const InductionDescriptor &getInductionDescriptor() const { return IndDesc; } | ||||||||||||
| 1222 | |||||||||||||
| 1223 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1224 | /// Print the recipe. | ||||||||||||
| 1225 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1226 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1227 | #endif | ||||||||||||
| 1228 | }; | ||||||||||||
| 1229 | |||||||||||||
| 1230 | /// A recipe for handling header phis that are widened in the vector loop. | ||||||||||||
| 1231 | /// In the VPlan native path, all incoming VPValues & VPBasicBlock pairs are | ||||||||||||
| 1232 | /// managed in the recipe directly. | ||||||||||||
| 1233 | class VPWidenPHIRecipe : public VPHeaderPHIRecipe { | ||||||||||||
| 1234 | /// List of incoming blocks. Only used in the VPlan native path. | ||||||||||||
| 1235 | SmallVector<VPBasicBlock *, 2> IncomingBlocks; | ||||||||||||
| 1236 | |||||||||||||
| 1237 | public: | ||||||||||||
| 1238 | /// Create a new VPWidenPHIRecipe for \p Phi with start value \p Start. | ||||||||||||
| 1239 | VPWidenPHIRecipe(PHINode *Phi, VPValue *Start = nullptr) | ||||||||||||
| 1240 | : VPHeaderPHIRecipe(VPVWidenPHISC, VPWidenPHISC, Phi) { | ||||||||||||
| 1241 | if (Start) | ||||||||||||
| 1242 | addOperand(Start); | ||||||||||||
| 1243 | } | ||||||||||||
| 1244 | |||||||||||||
| 1245 | ~VPWidenPHIRecipe() override = default; | ||||||||||||
| 1246 | |||||||||||||
| 1247 | VP_CLASSOF_IMPL(VPRecipeBase::VPWidenPHISC, VPValue::VPVWidenPHISC)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPWidenPHISC; } static inline bool classof (const VPValue *V) { return V->getVPValueID() == VPValue:: VPVWidenPHISC; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast<VPRecipeBase>(U); return R && R->getVPDefID() == VPRecipeBase::VPWidenPHISC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID () == VPRecipeBase::VPWidenPHISC; } | ||||||||||||
| 1248 | |||||||||||||
| 1249 | /// Generate the phi/select nodes. | ||||||||||||
| 1250 | void execute(VPTransformState &State) override; | ||||||||||||
| 1251 | |||||||||||||
| 1252 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1253 | /// Print the recipe. | ||||||||||||
| 1254 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1255 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1256 | #endif | ||||||||||||
| 1257 | |||||||||||||
| 1258 | /// Adds a pair (\p IncomingV, \p IncomingBlock) to the phi. | ||||||||||||
| 1259 | void addIncoming(VPValue *IncomingV, VPBasicBlock *IncomingBlock) { | ||||||||||||
| 1260 | addOperand(IncomingV); | ||||||||||||
| 1261 | IncomingBlocks.push_back(IncomingBlock); | ||||||||||||
| 1262 | } | ||||||||||||
| 1263 | |||||||||||||
| 1264 | /// Returns the \p I th incoming VPBasicBlock. | ||||||||||||
| 1265 | VPBasicBlock *getIncomingBlock(unsigned I) { return IncomingBlocks[I]; } | ||||||||||||
| 1266 | |||||||||||||
| 1267 | /// Returns the \p I th incoming VPValue. | ||||||||||||
| 1268 | VPValue *getIncomingValue(unsigned I) { return getOperand(I); } | ||||||||||||
| 1269 | }; | ||||||||||||
| 1270 | |||||||||||||
| 1271 | /// A recipe for handling first-order recurrence phis. The start value is the | ||||||||||||
| 1272 | /// first operand of the recipe and the incoming value from the backedge is the | ||||||||||||
| 1273 | /// second operand. | ||||||||||||
| 1274 | struct VPFirstOrderRecurrencePHIRecipe : public VPHeaderPHIRecipe { | ||||||||||||
| 1275 | VPFirstOrderRecurrencePHIRecipe(PHINode *Phi, VPValue &Start) | ||||||||||||
| 1276 | : VPHeaderPHIRecipe(VPVFirstOrderRecurrencePHISC, | ||||||||||||
| 1277 | VPFirstOrderRecurrencePHISC, Phi, &Start) {} | ||||||||||||
| 1278 | |||||||||||||
| 1279 | VP_CLASSOF_IMPL(VPRecipeBase::VPFirstOrderRecurrencePHISC,static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPFirstOrderRecurrencePHISC; } static inline bool classof(const VPValue *V) { return V->getVPValueID() == VPValue::VPVFirstOrderRecurrencePHISC; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast<VPRecipeBase >(U); return R && R->getVPDefID() == VPRecipeBase ::VPFirstOrderRecurrencePHISC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID() == VPRecipeBase ::VPFirstOrderRecurrencePHISC; } | ||||||||||||
| 1280 | VPValue::VPVFirstOrderRecurrencePHISC)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPFirstOrderRecurrencePHISC; } static inline bool classof(const VPValue *V) { return V->getVPValueID() == VPValue::VPVFirstOrderRecurrencePHISC; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast<VPRecipeBase >(U); return R && R->getVPDefID() == VPRecipeBase ::VPFirstOrderRecurrencePHISC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID() == VPRecipeBase ::VPFirstOrderRecurrencePHISC; } | ||||||||||||
| 1281 | |||||||||||||
| 1282 | static inline bool classof(const VPHeaderPHIRecipe *R) { | ||||||||||||
| 1283 | return R->getVPDefID() == VPRecipeBase::VPFirstOrderRecurrencePHISC; | ||||||||||||
| 1284 | } | ||||||||||||
| 1285 | |||||||||||||
| 1286 | void execute(VPTransformState &State) override; | ||||||||||||
| 1287 | |||||||||||||
| 1288 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1289 | /// Print the recipe. | ||||||||||||
| 1290 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1291 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1292 | #endif | ||||||||||||
| 1293 | }; | ||||||||||||
| 1294 | |||||||||||||
| 1295 | /// A recipe for handling reduction phis. The start value is the first operand | ||||||||||||
| 1296 | /// of the recipe and the incoming value from the backedge is the second | ||||||||||||
| 1297 | /// operand. | ||||||||||||
| 1298 | class VPReductionPHIRecipe : public VPHeaderPHIRecipe { | ||||||||||||
| 1299 | /// Descriptor for the reduction. | ||||||||||||
| 1300 | const RecurrenceDescriptor &RdxDesc; | ||||||||||||
| 1301 | |||||||||||||
| 1302 | /// The phi is part of an in-loop reduction. | ||||||||||||
| 1303 | bool IsInLoop; | ||||||||||||
| 1304 | |||||||||||||
| 1305 | /// The phi is part of an ordered reduction. Requires IsInLoop to be true. | ||||||||||||
| 1306 | bool IsOrdered; | ||||||||||||
| 1307 | |||||||||||||
| 1308 | public: | ||||||||||||
| 1309 | /// Create a new VPReductionPHIRecipe for the reduction \p Phi described by \p | ||||||||||||
| 1310 | /// RdxDesc. | ||||||||||||
| 1311 | VPReductionPHIRecipe(PHINode *Phi, const RecurrenceDescriptor &RdxDesc, | ||||||||||||
| 1312 | VPValue &Start, bool IsInLoop = false, | ||||||||||||
| 1313 | bool IsOrdered = false) | ||||||||||||
| 1314 | : VPHeaderPHIRecipe(VPVReductionPHISC, VPReductionPHISC, Phi, &Start), | ||||||||||||
| 1315 | RdxDesc(RdxDesc), IsInLoop(IsInLoop), IsOrdered(IsOrdered) { | ||||||||||||
| 1316 | assert((!IsOrdered || IsInLoop) && "IsOrdered requires IsInLoop")(static_cast <bool> ((!IsOrdered || IsInLoop) && "IsOrdered requires IsInLoop") ? void (0) : __assert_fail ("(!IsOrdered || IsInLoop) && \"IsOrdered requires IsInLoop\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1316, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 1317 | } | ||||||||||||
| 1318 | |||||||||||||
| 1319 | ~VPReductionPHIRecipe() override = default; | ||||||||||||
| 1320 | |||||||||||||
| 1321 | VP_CLASSOF_IMPL(VPRecipeBase::VPReductionPHISC, VPValue::VPVReductionPHISC)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPReductionPHISC; } static inline bool classof (const VPValue *V) { return V->getVPValueID() == VPValue:: VPVReductionPHISC; } static inline bool classof(const VPUser * U) { auto *R = dyn_cast<VPRecipeBase>(U); return R && R->getVPDefID() == VPRecipeBase::VPReductionPHISC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID () == VPRecipeBase::VPReductionPHISC; } | ||||||||||||
| 1322 | |||||||||||||
| 1323 | static inline bool classof(const VPHeaderPHIRecipe *R) { | ||||||||||||
| 1324 | return R->getVPDefID() == VPRecipeBase::VPReductionPHISC; | ||||||||||||
| 1325 | } | ||||||||||||
| 1326 | |||||||||||||
| 1327 | /// Generate the phi/select nodes. | ||||||||||||
| 1328 | void execute(VPTransformState &State) override; | ||||||||||||
| 1329 | |||||||||||||
| 1330 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1331 | /// Print the recipe. | ||||||||||||
| 1332 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1333 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1334 | #endif | ||||||||||||
| 1335 | |||||||||||||
| 1336 | const RecurrenceDescriptor &getRecurrenceDescriptor() const { | ||||||||||||
| 1337 | return RdxDesc; | ||||||||||||
| 1338 | } | ||||||||||||
| 1339 | |||||||||||||
| 1340 | /// Returns true, if the phi is part of an ordered reduction. | ||||||||||||
| 1341 | bool isOrdered() const { return IsOrdered; } | ||||||||||||
| 1342 | |||||||||||||
| 1343 | /// Returns true, if the phi is part of an in-loop reduction. | ||||||||||||
| 1344 | bool isInLoop() const { return IsInLoop; } | ||||||||||||
| 1345 | }; | ||||||||||||
| 1346 | |||||||||||||
| 1347 | /// A recipe for vectorizing a phi-node as a sequence of mask-based select | ||||||||||||
| 1348 | /// instructions. | ||||||||||||
| 1349 | class VPBlendRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
| 1350 | PHINode *Phi; | ||||||||||||
| 1351 | |||||||||||||
| 1352 | public: | ||||||||||||
| 1353 | /// The blend operation is a User of the incoming values and of their | ||||||||||||
| 1354 | /// respective masks, ordered [I0, M0, I1, M1, ...]. Note that a single value | ||||||||||||
| 1355 | /// might be incoming with a full mask for which there is no VPValue. | ||||||||||||
| 1356 | VPBlendRecipe(PHINode *Phi, ArrayRef<VPValue *> Operands) | ||||||||||||
| 1357 | : VPRecipeBase(VPBlendSC, Operands), | ||||||||||||
| 1358 | VPValue(VPValue::VPVBlendSC, Phi, this), Phi(Phi) { | ||||||||||||
| 1359 | assert(Operands.size() > 0 &&(static_cast <bool> (Operands.size() > 0 && ( (Operands.size() == 1) || (Operands.size() % 2 == 0)) && "Expected either a single incoming value or a positive even number " "of operands") ? void (0) : __assert_fail ("Operands.size() > 0 && ((Operands.size() == 1) || (Operands.size() % 2 == 0)) && \"Expected either a single incoming value or a positive even number \" \"of operands\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1362, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 1360 | ((Operands.size() == 1) || (Operands.size() % 2 == 0)) &&(static_cast <bool> (Operands.size() > 0 && ( (Operands.size() == 1) || (Operands.size() % 2 == 0)) && "Expected either a single incoming value or a positive even number " "of operands") ? void (0) : __assert_fail ("Operands.size() > 0 && ((Operands.size() == 1) || (Operands.size() % 2 == 0)) && \"Expected either a single incoming value or a positive even number \" \"of operands\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1362, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 1361 | "Expected either a single incoming value or a positive even number "(static_cast <bool> (Operands.size() > 0 && ( (Operands.size() == 1) || (Operands.size() % 2 == 0)) && "Expected either a single incoming value or a positive even number " "of operands") ? void (0) : __assert_fail ("Operands.size() > 0 && ((Operands.size() == 1) || (Operands.size() % 2 == 0)) && \"Expected either a single incoming value or a positive even number \" \"of operands\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1362, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 1362 | "of operands")(static_cast <bool> (Operands.size() > 0 && ( (Operands.size() == 1) || (Operands.size() % 2 == 0)) && "Expected either a single incoming value or a positive even number " "of operands") ? void (0) : __assert_fail ("Operands.size() > 0 && ((Operands.size() == 1) || (Operands.size() % 2 == 0)) && \"Expected either a single incoming value or a positive even number \" \"of operands\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1362, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 1363 | } | ||||||||||||
| 1364 | |||||||||||||
| 1365 | VP_CLASSOF_IMPL(VPRecipeBase::VPBlendSC, VPValue::VPVBlendSC)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPBlendSC; } static inline bool classof(const VPValue *V) { return V->getVPValueID() == VPValue::VPVBlendSC ; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast <VPRecipeBase>(U); return R && R->getVPDefID () == VPRecipeBase::VPBlendSC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID() == VPRecipeBase ::VPBlendSC; } | ||||||||||||
| 1366 | |||||||||||||
| 1367 | /// Return the number of incoming values, taking into account that a single | ||||||||||||
| 1368 | /// incoming value has no mask. | ||||||||||||
| 1369 | unsigned getNumIncomingValues() const { return (getNumOperands() + 1) / 2; } | ||||||||||||
| 1370 | |||||||||||||
| 1371 | /// Return incoming value number \p Idx. | ||||||||||||
| 1372 | VPValue *getIncomingValue(unsigned Idx) const { return getOperand(Idx * 2); } | ||||||||||||
| 1373 | |||||||||||||
| 1374 | /// Return mask number \p Idx. | ||||||||||||
| 1375 | VPValue *getMask(unsigned Idx) const { return getOperand(Idx * 2 + 1); } | ||||||||||||
| 1376 | |||||||||||||
| 1377 | /// Generate the phi/select nodes. | ||||||||||||
| 1378 | void execute(VPTransformState &State) override; | ||||||||||||
| 1379 | |||||||||||||
| 1380 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1381 | /// Print the recipe. | ||||||||||||
| 1382 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1383 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1384 | #endif | ||||||||||||
| 1385 | |||||||||||||
| 1386 | /// Returns true if the recipe only uses the first lane of operand \p Op. | ||||||||||||
| 1387 | bool onlyFirstLaneUsed(const VPValue *Op) const override { | ||||||||||||
| 1388 | assert(is_contained(operands(), Op) &&(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1389, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 1389 | "Op must be an operand of the recipe")(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1389, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 1390 | // Recursing through Blend recipes only, must terminate at header phi's the | ||||||||||||
| 1391 | // latest. | ||||||||||||
| 1392 | return all_of(users(), | ||||||||||||
| 1393 | [this](VPUser *U) { return U->onlyFirstLaneUsed(this); }); | ||||||||||||
| 1394 | } | ||||||||||||
| 1395 | }; | ||||||||||||
| 1396 | |||||||||||||
| 1397 | /// VPInterleaveRecipe is a recipe for transforming an interleave group of load | ||||||||||||
| 1398 | /// or stores into one wide load/store and shuffles. The first operand of a | ||||||||||||
| 1399 | /// VPInterleave recipe is the address, followed by the stored values, followed | ||||||||||||
| 1400 | /// by an optional mask. | ||||||||||||
| 1401 | class VPInterleaveRecipe : public VPRecipeBase { | ||||||||||||
| 1402 | const InterleaveGroup<Instruction> *IG; | ||||||||||||
| 1403 | |||||||||||||
| 1404 | bool HasMask = false; | ||||||||||||
| 1405 | |||||||||||||
| 1406 | public: | ||||||||||||
| 1407 | VPInterleaveRecipe(const InterleaveGroup<Instruction> *IG, VPValue *Addr, | ||||||||||||
| 1408 | ArrayRef<VPValue *> StoredValues, VPValue *Mask) | ||||||||||||
| 1409 | : VPRecipeBase(VPInterleaveSC, {Addr}), IG(IG) { | ||||||||||||
| 1410 | for (unsigned i = 0; i < IG->getFactor(); ++i) | ||||||||||||
| 1411 | if (Instruction *I = IG->getMember(i)) { | ||||||||||||
| 1412 | if (I->getType()->isVoidTy()) | ||||||||||||
| 1413 | continue; | ||||||||||||
| 1414 | new VPValue(I, this); | ||||||||||||
| 1415 | } | ||||||||||||
| 1416 | |||||||||||||
| 1417 | for (auto *SV : StoredValues) | ||||||||||||
| 1418 | addOperand(SV); | ||||||||||||
| 1419 | if (Mask) { | ||||||||||||
| 1420 | HasMask = true; | ||||||||||||
| 1421 | addOperand(Mask); | ||||||||||||
| 1422 | } | ||||||||||||
| 1423 | } | ||||||||||||
| 1424 | ~VPInterleaveRecipe() override = default; | ||||||||||||
| 1425 | |||||||||||||
| 1426 | static inline bool classof(const VPDef *D) { | ||||||||||||
| 1427 | return D->getVPDefID() == VPRecipeBase::VPInterleaveSC; | ||||||||||||
| 1428 | } | ||||||||||||
| 1429 | |||||||||||||
| 1430 | /// Return the address accessed by this recipe. | ||||||||||||
| 1431 | VPValue *getAddr() const { | ||||||||||||
| 1432 | return getOperand(0); // Address is the 1st, mandatory operand. | ||||||||||||
| 1433 | } | ||||||||||||
| 1434 | |||||||||||||
| 1435 | /// Return the mask used by this recipe. Note that a full mask is represented | ||||||||||||
| 1436 | /// by a nullptr. | ||||||||||||
| 1437 | VPValue *getMask() const { | ||||||||||||
| 1438 | // Mask is optional and therefore the last, currently 2nd operand. | ||||||||||||
| 1439 | return HasMask ? getOperand(getNumOperands() - 1) : nullptr; | ||||||||||||
| 1440 | } | ||||||||||||
| 1441 | |||||||||||||
| 1442 | /// Return the VPValues stored by this interleave group. If it is a load | ||||||||||||
| 1443 | /// interleave group, return an empty ArrayRef. | ||||||||||||
| 1444 | ArrayRef<VPValue *> getStoredValues() const { | ||||||||||||
| 1445 | // The first operand is the address, followed by the stored values, followed | ||||||||||||
| 1446 | // by an optional mask. | ||||||||||||
| 1447 | return ArrayRef<VPValue *>(op_begin(), getNumOperands()) | ||||||||||||
| 1448 | .slice(1, getNumStoreOperands()); | ||||||||||||
| 1449 | } | ||||||||||||
| 1450 | |||||||||||||
| 1451 | /// Generate the wide load or store, and shuffles. | ||||||||||||
| 1452 | void execute(VPTransformState &State) override; | ||||||||||||
| 1453 | |||||||||||||
| 1454 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1455 | /// Print the recipe. | ||||||||||||
| 1456 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1457 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1458 | #endif | ||||||||||||
| 1459 | |||||||||||||
| 1460 | const InterleaveGroup<Instruction> *getInterleaveGroup() { return IG; } | ||||||||||||
| 1461 | |||||||||||||
| 1462 | /// Returns the number of stored operands of this interleave group. Returns 0 | ||||||||||||
| 1463 | /// for load interleave groups. | ||||||||||||
| 1464 | unsigned getNumStoreOperands() const { | ||||||||||||
| 1465 | return getNumOperands() - (HasMask ? 2 : 1); | ||||||||||||
| 1466 | } | ||||||||||||
| 1467 | |||||||||||||
| 1468 | /// The recipe only uses the first lane of the address. | ||||||||||||
| 1469 | bool onlyFirstLaneUsed(const VPValue *Op) const override { | ||||||||||||
| 1470 | assert(is_contained(operands(), Op) &&(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1471, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 1471 | "Op must be an operand of the recipe")(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1471, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 1472 | return Op == getAddr() && !llvm::is_contained(getStoredValues(), Op); | ||||||||||||
| 1473 | } | ||||||||||||
| 1474 | }; | ||||||||||||
| 1475 | |||||||||||||
| 1476 | /// A recipe to represent inloop reduction operations, performing a reduction on | ||||||||||||
| 1477 | /// a vector operand into a scalar value, and adding the result to a chain. | ||||||||||||
| 1478 | /// The Operands are {ChainOp, VecOp, [Condition]}. | ||||||||||||
| 1479 | class VPReductionRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
| 1480 | /// The recurrence decriptor for the reduction in question. | ||||||||||||
| 1481 | const RecurrenceDescriptor *RdxDesc; | ||||||||||||
| 1482 | /// Pointer to the TTI, needed to create the target reduction | ||||||||||||
| 1483 | const TargetTransformInfo *TTI; | ||||||||||||
| 1484 | |||||||||||||
| 1485 | public: | ||||||||||||
| 1486 | VPReductionRecipe(const RecurrenceDescriptor *R, Instruction *I, | ||||||||||||
| 1487 | VPValue *ChainOp, VPValue *VecOp, VPValue *CondOp, | ||||||||||||
| 1488 | const TargetTransformInfo *TTI) | ||||||||||||
| 1489 | : VPRecipeBase(VPRecipeBase::VPReductionSC, {ChainOp, VecOp}), | ||||||||||||
| 1490 | VPValue(VPValue::VPVReductionSC, I, this), RdxDesc(R), TTI(TTI) { | ||||||||||||
| 1491 | if (CondOp) | ||||||||||||
| 1492 | addOperand(CondOp); | ||||||||||||
| 1493 | } | ||||||||||||
| 1494 | |||||||||||||
| 1495 | ~VPReductionRecipe() override = default; | ||||||||||||
| 1496 | |||||||||||||
| 1497 | VP_CLASSOF_IMPL(VPRecipeBase::VPReductionSC, VPValue::VPVReductionSC)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPReductionSC; } static inline bool classof (const VPValue *V) { return V->getVPValueID() == VPValue:: VPVReductionSC; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast<VPRecipeBase>(U); return R && R->getVPDefID() == VPRecipeBase::VPReductionSC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID () == VPRecipeBase::VPReductionSC; } | ||||||||||||
| 1498 | |||||||||||||
| 1499 | /// Generate the reduction in the loop | ||||||||||||
| 1500 | void execute(VPTransformState &State) override; | ||||||||||||
| 1501 | |||||||||||||
| 1502 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1503 | /// Print the recipe. | ||||||||||||
| 1504 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1505 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1506 | #endif | ||||||||||||
| 1507 | |||||||||||||
| 1508 | /// The VPValue of the scalar Chain being accumulated. | ||||||||||||
| 1509 | VPValue *getChainOp() const { return getOperand(0); } | ||||||||||||
| 1510 | /// The VPValue of the vector value to be reduced. | ||||||||||||
| 1511 | VPValue *getVecOp() const { return getOperand(1); } | ||||||||||||
| 1512 | /// The VPValue of the condition for the block. | ||||||||||||
| 1513 | VPValue *getCondOp() const { | ||||||||||||
| 1514 | return getNumOperands() > 2 ? getOperand(2) : nullptr; | ||||||||||||
| 1515 | } | ||||||||||||
| 1516 | }; | ||||||||||||
| 1517 | |||||||||||||
| 1518 | /// VPReplicateRecipe replicates a given instruction producing multiple scalar | ||||||||||||
| 1519 | /// copies of the original scalar type, one per lane, instead of producing a | ||||||||||||
| 1520 | /// single copy of widened type for all lanes. If the instruction is known to be | ||||||||||||
| 1521 | /// uniform only one copy, per lane zero, will be generated. | ||||||||||||
| 1522 | class VPReplicateRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
| 1523 | /// Indicator if only a single replica per lane is needed. | ||||||||||||
| 1524 | bool IsUniform; | ||||||||||||
| 1525 | |||||||||||||
| 1526 | /// Indicator if the replicas are also predicated. | ||||||||||||
| 1527 | bool IsPredicated; | ||||||||||||
| 1528 | |||||||||||||
| 1529 | /// Indicator if the scalar values should also be packed into a vector. | ||||||||||||
| 1530 | bool AlsoPack; | ||||||||||||
| 1531 | |||||||||||||
| 1532 | public: | ||||||||||||
| 1533 | template <typename IterT> | ||||||||||||
| 1534 | VPReplicateRecipe(Instruction *I, iterator_range<IterT> Operands, | ||||||||||||
| 1535 | bool IsUniform, bool IsPredicated = false) | ||||||||||||
| 1536 | : VPRecipeBase(VPReplicateSC, Operands), VPValue(VPVReplicateSC, I, this), | ||||||||||||
| 1537 | IsUniform(IsUniform), IsPredicated(IsPredicated) { | ||||||||||||
| 1538 | // Retain the previous behavior of predicateInstructions(), where an | ||||||||||||
| 1539 | // insert-element of a predicated instruction got hoisted into the | ||||||||||||
| 1540 | // predicated basic block iff it was its only user. This is achieved by | ||||||||||||
| 1541 | // having predicated instructions also pack their values into a vector by | ||||||||||||
| 1542 | // default unless they have a replicated user which uses their scalar value. | ||||||||||||
| 1543 | AlsoPack = IsPredicated && !I->use_empty(); | ||||||||||||
| 1544 | } | ||||||||||||
| 1545 | |||||||||||||
| 1546 | ~VPReplicateRecipe() override = default; | ||||||||||||
| 1547 | |||||||||||||
| 1548 | VP_CLASSOF_IMPL(VPRecipeBase::VPReplicateSC, VPValue::VPVReplicateSC)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPReplicateSC; } static inline bool classof (const VPValue *V) { return V->getVPValueID() == VPValue:: VPVReplicateSC; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast<VPRecipeBase>(U); return R && R->getVPDefID() == VPRecipeBase::VPReplicateSC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID () == VPRecipeBase::VPReplicateSC; } | ||||||||||||
| 1549 | |||||||||||||
| 1550 | /// Generate replicas of the desired Ingredient. Replicas will be generated | ||||||||||||
| 1551 | /// for all parts and lanes unless a specific part and lane are specified in | ||||||||||||
| 1552 | /// the \p State. | ||||||||||||
| 1553 | void execute(VPTransformState &State) override; | ||||||||||||
| 1554 | |||||||||||||
| 1555 | void setAlsoPack(bool Pack) { AlsoPack = Pack; } | ||||||||||||
| 1556 | |||||||||||||
| 1557 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1558 | /// Print the recipe. | ||||||||||||
| 1559 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1560 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1561 | #endif | ||||||||||||
| 1562 | |||||||||||||
| 1563 | bool isUniform() const { return IsUniform; } | ||||||||||||
| 1564 | |||||||||||||
| 1565 | bool isPacked() const { return AlsoPack; } | ||||||||||||
| 1566 | |||||||||||||
| 1567 | bool isPredicated() const { return IsPredicated; } | ||||||||||||
| 1568 | |||||||||||||
| 1569 | /// Returns true if the recipe only uses the first lane of operand \p Op. | ||||||||||||
| 1570 | bool onlyFirstLaneUsed(const VPValue *Op) const override { | ||||||||||||
| 1571 | assert(is_contained(operands(), Op) &&(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1572, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 1572 | "Op must be an operand of the recipe")(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1572, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 1573 | return isUniform(); | ||||||||||||
| 1574 | } | ||||||||||||
| 1575 | |||||||||||||
| 1576 | /// Returns true if the recipe uses scalars of operand \p Op. | ||||||||||||
| 1577 | bool usesScalars(const VPValue *Op) const override { | ||||||||||||
| 1578 | assert(is_contained(operands(), Op) &&(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1579, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 1579 | "Op must be an operand of the recipe")(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1579, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 1580 | return true; | ||||||||||||
| 1581 | } | ||||||||||||
| 1582 | }; | ||||||||||||
| 1583 | |||||||||||||
| 1584 | /// A recipe for generating conditional branches on the bits of a mask. | ||||||||||||
| 1585 | class VPBranchOnMaskRecipe : public VPRecipeBase { | ||||||||||||
| 1586 | public: | ||||||||||||
| 1587 | VPBranchOnMaskRecipe(VPValue *BlockInMask) | ||||||||||||
| 1588 | : VPRecipeBase(VPBranchOnMaskSC, {}) { | ||||||||||||
| 1589 | if (BlockInMask) // nullptr means all-one mask. | ||||||||||||
| 1590 | addOperand(BlockInMask); | ||||||||||||
| 1591 | } | ||||||||||||
| 1592 | |||||||||||||
| 1593 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
| 1594 | static inline bool classof(const VPDef *D) { | ||||||||||||
| 1595 | return D->getVPDefID() == VPRecipeBase::VPBranchOnMaskSC; | ||||||||||||
| 1596 | } | ||||||||||||
| 1597 | |||||||||||||
| 1598 | /// Generate the extraction of the appropriate bit from the block mask and the | ||||||||||||
| 1599 | /// conditional branch. | ||||||||||||
| 1600 | void execute(VPTransformState &State) override; | ||||||||||||
| 1601 | |||||||||||||
| 1602 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1603 | /// Print the recipe. | ||||||||||||
| 1604 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1605 | VPSlotTracker &SlotTracker) const override { | ||||||||||||
| 1606 | O << Indent << "BRANCH-ON-MASK "; | ||||||||||||
| 1607 | if (VPValue *Mask = getMask()) | ||||||||||||
| 1608 | Mask->printAsOperand(O, SlotTracker); | ||||||||||||
| 1609 | else | ||||||||||||
| 1610 | O << " All-One"; | ||||||||||||
| 1611 | } | ||||||||||||
| 1612 | #endif | ||||||||||||
| 1613 | |||||||||||||
| 1614 | /// Return the mask used by this recipe. Note that a full mask is represented | ||||||||||||
| 1615 | /// by a nullptr. | ||||||||||||
| 1616 | VPValue *getMask() const { | ||||||||||||
| 1617 | assert(getNumOperands() <= 1 && "should have either 0 or 1 operands")(static_cast <bool> (getNumOperands() <= 1 && "should have either 0 or 1 operands") ? void (0) : __assert_fail ("getNumOperands() <= 1 && \"should have either 0 or 1 operands\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1617, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 1618 | // Mask is optional. | ||||||||||||
| 1619 | return getNumOperands() == 1 ? getOperand(0) : nullptr; | ||||||||||||
| 1620 | } | ||||||||||||
| 1621 | |||||||||||||
| 1622 | /// Returns true if the recipe uses scalars of operand \p Op. | ||||||||||||
| 1623 | bool usesScalars(const VPValue *Op) const override { | ||||||||||||
| 1624 | assert(is_contained(operands(), Op) &&(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1625, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 1625 | "Op must be an operand of the recipe")(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1625, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 1626 | return true; | ||||||||||||
| 1627 | } | ||||||||||||
| 1628 | }; | ||||||||||||
| 1629 | |||||||||||||
| 1630 | /// VPPredInstPHIRecipe is a recipe for generating the phi nodes needed when | ||||||||||||
| 1631 | /// control converges back from a Branch-on-Mask. The phi nodes are needed in | ||||||||||||
| 1632 | /// order to merge values that are set under such a branch and feed their uses. | ||||||||||||
| 1633 | /// The phi nodes can be scalar or vector depending on the users of the value. | ||||||||||||
| 1634 | /// This recipe works in concert with VPBranchOnMaskRecipe. | ||||||||||||
| 1635 | class VPPredInstPHIRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
| 1636 | public: | ||||||||||||
| 1637 | /// Construct a VPPredInstPHIRecipe given \p PredInst whose value needs a phi | ||||||||||||
| 1638 | /// nodes after merging back from a Branch-on-Mask. | ||||||||||||
| 1639 | VPPredInstPHIRecipe(VPValue *PredV) | ||||||||||||
| 1640 | : VPRecipeBase(VPPredInstPHISC, PredV), | ||||||||||||
| 1641 | VPValue(VPValue::VPVPredInstPHI, nullptr, this) {} | ||||||||||||
| 1642 | ~VPPredInstPHIRecipe() override = default; | ||||||||||||
| 1643 | |||||||||||||
| 1644 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
| 1645 | static inline bool classof(const VPDef *D) { | ||||||||||||
| 1646 | return D->getVPDefID() == VPRecipeBase::VPPredInstPHISC; | ||||||||||||
| 1647 | } | ||||||||||||
| 1648 | |||||||||||||
| 1649 | /// Generates phi nodes for live-outs as needed to retain SSA form. | ||||||||||||
| 1650 | void execute(VPTransformState &State) override; | ||||||||||||
| 1651 | |||||||||||||
| 1652 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1653 | /// Print the recipe. | ||||||||||||
| 1654 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1655 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1656 | #endif | ||||||||||||
| 1657 | |||||||||||||
| 1658 | /// Returns true if the recipe uses scalars of operand \p Op. | ||||||||||||
| 1659 | bool usesScalars(const VPValue *Op) const override { | ||||||||||||
| 1660 | assert(is_contained(operands(), Op) &&(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1661, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 1661 | "Op must be an operand of the recipe")(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1661, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 1662 | return true; | ||||||||||||
| 1663 | } | ||||||||||||
| 1664 | }; | ||||||||||||
| 1665 | |||||||||||||
| 1666 | /// A Recipe for widening load/store operations. | ||||||||||||
| 1667 | /// The recipe uses the following VPValues: | ||||||||||||
| 1668 | /// - For load: Address, optional mask | ||||||||||||
| 1669 | /// - For store: Address, stored value, optional mask | ||||||||||||
| 1670 | /// TODO: We currently execute only per-part unless a specific instance is | ||||||||||||
| 1671 | /// provided. | ||||||||||||
| 1672 | class VPWidenMemoryInstructionRecipe : public VPRecipeBase { | ||||||||||||
| 1673 | Instruction &Ingredient; | ||||||||||||
| 1674 | |||||||||||||
| 1675 | // Whether the loaded-from / stored-to addresses are consecutive. | ||||||||||||
| 1676 | bool Consecutive; | ||||||||||||
| 1677 | |||||||||||||
| 1678 | // Whether the consecutive loaded/stored addresses are in reverse order. | ||||||||||||
| 1679 | bool Reverse; | ||||||||||||
| 1680 | |||||||||||||
| 1681 | void setMask(VPValue *Mask) { | ||||||||||||
| 1682 | if (!Mask) | ||||||||||||
| 1683 | return; | ||||||||||||
| 1684 | addOperand(Mask); | ||||||||||||
| 1685 | } | ||||||||||||
| 1686 | |||||||||||||
| 1687 | bool isMasked() const { | ||||||||||||
| 1688 | return isStore() ? getNumOperands() == 3 : getNumOperands() == 2; | ||||||||||||
| 1689 | } | ||||||||||||
| 1690 | |||||||||||||
| 1691 | public: | ||||||||||||
| 1692 | VPWidenMemoryInstructionRecipe(LoadInst &Load, VPValue *Addr, VPValue *Mask, | ||||||||||||
| 1693 | bool Consecutive, bool Reverse) | ||||||||||||
| 1694 | : VPRecipeBase(VPWidenMemoryInstructionSC, {Addr}), Ingredient(Load), | ||||||||||||
| 1695 | Consecutive(Consecutive), Reverse(Reverse) { | ||||||||||||
| 1696 | assert((Consecutive || !Reverse) && "Reverse implies consecutive")(static_cast <bool> ((Consecutive || !Reverse) && "Reverse implies consecutive") ? void (0) : __assert_fail ("(Consecutive || !Reverse) && \"Reverse implies consecutive\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1696, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 1697 | new VPValue(VPValue::VPVMemoryInstructionSC, &Load, this); | ||||||||||||
| 1698 | setMask(Mask); | ||||||||||||
| 1699 | } | ||||||||||||
| 1700 | |||||||||||||
| 1701 | VPWidenMemoryInstructionRecipe(StoreInst &Store, VPValue *Addr, | ||||||||||||
| 1702 | VPValue *StoredValue, VPValue *Mask, | ||||||||||||
| 1703 | bool Consecutive, bool Reverse) | ||||||||||||
| 1704 | : VPRecipeBase(VPWidenMemoryInstructionSC, {Addr, StoredValue}), | ||||||||||||
| 1705 | Ingredient(Store), Consecutive(Consecutive), Reverse(Reverse) { | ||||||||||||
| 1706 | assert((Consecutive || !Reverse) && "Reverse implies consecutive")(static_cast <bool> ((Consecutive || !Reverse) && "Reverse implies consecutive") ? void (0) : __assert_fail ("(Consecutive || !Reverse) && \"Reverse implies consecutive\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1706, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 1707 | setMask(Mask); | ||||||||||||
| 1708 | } | ||||||||||||
| 1709 | |||||||||||||
| 1710 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
| 1711 | static inline bool classof(const VPDef *D) { | ||||||||||||
| 1712 | return D->getVPDefID() == VPRecipeBase::VPWidenMemoryInstructionSC; | ||||||||||||
| 1713 | } | ||||||||||||
| 1714 | |||||||||||||
| 1715 | /// Return the address accessed by this recipe. | ||||||||||||
| 1716 | VPValue *getAddr() const { | ||||||||||||
| 1717 | return getOperand(0); // Address is the 1st, mandatory operand. | ||||||||||||
| 1718 | } | ||||||||||||
| 1719 | |||||||||||||
| 1720 | /// Return the mask used by this recipe. Note that a full mask is represented | ||||||||||||
| 1721 | /// by a nullptr. | ||||||||||||
| 1722 | VPValue *getMask() const { | ||||||||||||
| 1723 | // Mask is optional and therefore the last operand. | ||||||||||||
| 1724 | return isMasked() ? getOperand(getNumOperands() - 1) : nullptr; | ||||||||||||
| 1725 | } | ||||||||||||
| 1726 | |||||||||||||
| 1727 | /// Returns true if this recipe is a store. | ||||||||||||
| 1728 | bool isStore() const { return isa<StoreInst>(Ingredient); } | ||||||||||||
| 1729 | |||||||||||||
| 1730 | /// Return the address accessed by this recipe. | ||||||||||||
| 1731 | VPValue *getStoredValue() const { | ||||||||||||
| 1732 | assert(isStore() && "Stored value only available for store instructions")(static_cast <bool> (isStore() && "Stored value only available for store instructions" ) ? void (0) : __assert_fail ("isStore() && \"Stored value only available for store instructions\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1732, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 1733 | return getOperand(1); // Stored value is the 2nd, mandatory operand. | ||||||||||||
| 1734 | } | ||||||||||||
| 1735 | |||||||||||||
| 1736 | // Return whether the loaded-from / stored-to addresses are consecutive. | ||||||||||||
| 1737 | bool isConsecutive() const { return Consecutive; } | ||||||||||||
| 1738 | |||||||||||||
| 1739 | // Return whether the consecutive loaded/stored addresses are in reverse | ||||||||||||
| 1740 | // order. | ||||||||||||
| 1741 | bool isReverse() const { return Reverse; } | ||||||||||||
| 1742 | |||||||||||||
| 1743 | /// Generate the wide load/store. | ||||||||||||
| 1744 | void execute(VPTransformState &State) override; | ||||||||||||
| 1745 | |||||||||||||
| 1746 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1747 | /// Print the recipe. | ||||||||||||
| 1748 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1749 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1750 | #endif | ||||||||||||
| 1751 | |||||||||||||
| 1752 | /// Returns true if the recipe only uses the first lane of operand \p Op. | ||||||||||||
| 1753 | bool onlyFirstLaneUsed(const VPValue *Op) const override { | ||||||||||||
| 1754 | assert(is_contained(operands(), Op) &&(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1755, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 1755 | "Op must be an operand of the recipe")(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1755, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 1756 | |||||||||||||
| 1757 | // Widened, consecutive memory operations only demand the first lane of | ||||||||||||
| 1758 | // their address, unless the same operand is also stored. That latter can | ||||||||||||
| 1759 | // happen with opaque pointers. | ||||||||||||
| 1760 | return Op == getAddr() && isConsecutive() && | ||||||||||||
| 1761 | (!isStore() || Op != getStoredValue()); | ||||||||||||
| 1762 | } | ||||||||||||
| 1763 | |||||||||||||
| 1764 | Instruction &getIngredient() const { return Ingredient; } | ||||||||||||
| 1765 | }; | ||||||||||||
| 1766 | |||||||||||||
| 1767 | /// Recipe to expand a SCEV expression. | ||||||||||||
| 1768 | class VPExpandSCEVRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
| 1769 | const SCEV *Expr; | ||||||||||||
| 1770 | ScalarEvolution &SE; | ||||||||||||
| 1771 | |||||||||||||
| 1772 | public: | ||||||||||||
| 1773 | VPExpandSCEVRecipe(const SCEV *Expr, ScalarEvolution &SE) | ||||||||||||
| 1774 | : VPRecipeBase(VPExpandSCEVSC, {}), VPValue(nullptr, this), Expr(Expr), | ||||||||||||
| 1775 | SE(SE) {} | ||||||||||||
| 1776 | |||||||||||||
| 1777 | ~VPExpandSCEVRecipe() override = default; | ||||||||||||
| 1778 | |||||||||||||
| 1779 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
| 1780 | static inline bool classof(const VPDef *D) { | ||||||||||||
| 1781 | return D->getVPDefID() == VPExpandSCEVSC; | ||||||||||||
| 1782 | } | ||||||||||||
| 1783 | |||||||||||||
| 1784 | /// Generate a canonical vector induction variable of the vector loop, with | ||||||||||||
| 1785 | void execute(VPTransformState &State) override; | ||||||||||||
| 1786 | |||||||||||||
| 1787 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1788 | /// Print the recipe. | ||||||||||||
| 1789 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1790 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1791 | #endif | ||||||||||||
| 1792 | |||||||||||||
| 1793 | const SCEV *getSCEV() const { return Expr; } | ||||||||||||
| 1794 | }; | ||||||||||||
| 1795 | |||||||||||||
| 1796 | /// Canonical scalar induction phi of the vector loop. Starting at the specified | ||||||||||||
| 1797 | /// start value (either 0 or the resume value when vectorizing the epilogue | ||||||||||||
| 1798 | /// loop). VPWidenCanonicalIVRecipe represents the vector version of the | ||||||||||||
| 1799 | /// canonical induction variable. | ||||||||||||
| 1800 | class VPCanonicalIVPHIRecipe : public VPHeaderPHIRecipe { | ||||||||||||
| 1801 | DebugLoc DL; | ||||||||||||
| 1802 | |||||||||||||
| 1803 | public: | ||||||||||||
| 1804 | VPCanonicalIVPHIRecipe(VPValue *StartV, DebugLoc DL) | ||||||||||||
| 1805 | : VPHeaderPHIRecipe(VPValue::VPVCanonicalIVPHISC, VPCanonicalIVPHISC, | ||||||||||||
| 1806 | nullptr, StartV), | ||||||||||||
| 1807 | DL(DL) {} | ||||||||||||
| 1808 | |||||||||||||
| 1809 | ~VPCanonicalIVPHIRecipe() override = default; | ||||||||||||
| 1810 | |||||||||||||
| 1811 | VP_CLASSOF_IMPL(VPRecipeBase::VPCanonicalIVPHISC,static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPCanonicalIVPHISC; } static inline bool classof (const VPValue *V) { return V->getVPValueID() == VPValue:: VPVCanonicalIVPHISC; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast<VPRecipeBase>(U); return R && R->getVPDefID() == VPRecipeBase::VPCanonicalIVPHISC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID () == VPRecipeBase::VPCanonicalIVPHISC; } | ||||||||||||
| 1812 | VPValue::VPVCanonicalIVPHISC)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPCanonicalIVPHISC; } static inline bool classof (const VPValue *V) { return V->getVPValueID() == VPValue:: VPVCanonicalIVPHISC; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast<VPRecipeBase>(U); return R && R->getVPDefID() == VPRecipeBase::VPCanonicalIVPHISC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID () == VPRecipeBase::VPCanonicalIVPHISC; } | ||||||||||||
| 1813 | |||||||||||||
| 1814 | static inline bool classof(const VPHeaderPHIRecipe *D) { | ||||||||||||
| 1815 | return D->getVPDefID() == VPCanonicalIVPHISC; | ||||||||||||
| 1816 | } | ||||||||||||
| 1817 | |||||||||||||
| 1818 | /// Generate the canonical scalar induction phi of the vector loop. | ||||||||||||
| 1819 | void execute(VPTransformState &State) override; | ||||||||||||
| 1820 | |||||||||||||
| 1821 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1822 | /// Print the recipe. | ||||||||||||
| 1823 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1824 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1825 | #endif | ||||||||||||
| 1826 | |||||||||||||
| 1827 | /// Returns the scalar type of the induction. | ||||||||||||
| 1828 | const Type *getScalarType() const { | ||||||||||||
| 1829 | return getOperand(0)->getLiveInIRValue()->getType(); | ||||||||||||
| 1830 | } | ||||||||||||
| 1831 | |||||||||||||
| 1832 | /// Returns true if the recipe only uses the first lane of operand \p Op. | ||||||||||||
| 1833 | bool onlyFirstLaneUsed(const VPValue *Op) const override { | ||||||||||||
| 1834 | assert(is_contained(operands(), Op) &&(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1835, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 1835 | "Op must be an operand of the recipe")(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1835, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 1836 | return true; | ||||||||||||
| 1837 | } | ||||||||||||
| 1838 | |||||||||||||
| 1839 | /// Check if the induction described by \p ID is canonical, i.e. has the same | ||||||||||||
| 1840 | /// start, step (of 1), and type as the canonical IV. | ||||||||||||
| 1841 | bool isCanonical(const InductionDescriptor &ID, Type *Ty) const; | ||||||||||||
| 1842 | }; | ||||||||||||
| 1843 | |||||||||||||
| 1844 | /// A recipe for generating the active lane mask for the vector loop that is | ||||||||||||
| 1845 | /// used to predicate the vector operations. | ||||||||||||
| 1846 | /// TODO: It would be good to use the existing VPWidenPHIRecipe instead and | ||||||||||||
| 1847 | /// remove VPActiveLaneMaskPHIRecipe. | ||||||||||||
| 1848 | class VPActiveLaneMaskPHIRecipe : public VPHeaderPHIRecipe { | ||||||||||||
| 1849 | DebugLoc DL; | ||||||||||||
| 1850 | |||||||||||||
| 1851 | public: | ||||||||||||
| 1852 | VPActiveLaneMaskPHIRecipe(VPValue *StartMask, DebugLoc DL) | ||||||||||||
| 1853 | : VPHeaderPHIRecipe(VPValue::VPVActiveLaneMaskPHISC, | ||||||||||||
| 1854 | VPActiveLaneMaskPHISC, nullptr, StartMask), | ||||||||||||
| 1855 | DL(DL) {} | ||||||||||||
| 1856 | |||||||||||||
| 1857 | ~VPActiveLaneMaskPHIRecipe() override = default; | ||||||||||||
| 1858 | |||||||||||||
| 1859 | VP_CLASSOF_IMPL(VPRecipeBase::VPActiveLaneMaskPHISC,static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPActiveLaneMaskPHISC; } static inline bool classof(const VPValue *V) { return V->getVPValueID() == VPValue ::VPVActiveLaneMaskPHISC; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast<VPRecipeBase>(U); return R && R->getVPDefID() == VPRecipeBase::VPActiveLaneMaskPHISC; } static inline bool classof(const VPRecipeBase *R) { return R ->getVPDefID() == VPRecipeBase::VPActiveLaneMaskPHISC; } | ||||||||||||
| 1860 | VPValue::VPVActiveLaneMaskPHISC)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPActiveLaneMaskPHISC; } static inline bool classof(const VPValue *V) { return V->getVPValueID() == VPValue ::VPVActiveLaneMaskPHISC; } static inline bool classof(const VPUser *U) { auto *R = dyn_cast<VPRecipeBase>(U); return R && R->getVPDefID() == VPRecipeBase::VPActiveLaneMaskPHISC; } static inline bool classof(const VPRecipeBase *R) { return R ->getVPDefID() == VPRecipeBase::VPActiveLaneMaskPHISC; } | ||||||||||||
| 1861 | |||||||||||||
| 1862 | static inline bool classof(const VPHeaderPHIRecipe *D) { | ||||||||||||
| 1863 | return D->getVPDefID() == VPActiveLaneMaskPHISC; | ||||||||||||
| 1864 | } | ||||||||||||
| 1865 | |||||||||||||
| 1866 | /// Generate the active lane mask phi of the vector loop. | ||||||||||||
| 1867 | void execute(VPTransformState &State) override; | ||||||||||||
| 1868 | |||||||||||||
| 1869 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1870 | /// Print the recipe. | ||||||||||||
| 1871 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1872 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1873 | #endif | ||||||||||||
| 1874 | }; | ||||||||||||
| 1875 | |||||||||||||
| 1876 | /// A Recipe for widening the canonical induction variable of the vector loop. | ||||||||||||
| 1877 | class VPWidenCanonicalIVRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
| 1878 | public: | ||||||||||||
| 1879 | VPWidenCanonicalIVRecipe(VPCanonicalIVPHIRecipe *CanonicalIV) | ||||||||||||
| 1880 | : VPRecipeBase(VPWidenCanonicalIVSC, {CanonicalIV}), | ||||||||||||
| 1881 | VPValue(VPValue::VPVWidenCanonicalIVSC, nullptr, this) {} | ||||||||||||
| 1882 | |||||||||||||
| 1883 | ~VPWidenCanonicalIVRecipe() override = default; | ||||||||||||
| 1884 | |||||||||||||
| 1885 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
| 1886 | static inline bool classof(const VPDef *D) { | ||||||||||||
| 1887 | return D->getVPDefID() == VPRecipeBase::VPWidenCanonicalIVSC; | ||||||||||||
| 1888 | } | ||||||||||||
| 1889 | |||||||||||||
| 1890 | /// Extra classof implementations to allow directly casting from VPUser -> | ||||||||||||
| 1891 | /// VPWidenCanonicalIVRecipe. | ||||||||||||
| 1892 | static inline bool classof(const VPUser *U) { | ||||||||||||
| 1893 | auto *R = dyn_cast<VPRecipeBase>(U); | ||||||||||||
| 1894 | return R && R->getVPDefID() == VPRecipeBase::VPWidenCanonicalIVSC; | ||||||||||||
| 1895 | } | ||||||||||||
| 1896 | static inline bool classof(const VPRecipeBase *R) { | ||||||||||||
| 1897 | return R->getVPDefID() == VPRecipeBase::VPWidenCanonicalIVSC; | ||||||||||||
| 1898 | } | ||||||||||||
| 1899 | |||||||||||||
| 1900 | /// Generate a canonical vector induction variable of the vector loop, with | ||||||||||||
| 1901 | /// start = {<Part*VF, Part*VF+1, ..., Part*VF+VF-1> for 0 <= Part < UF}, and | ||||||||||||
| 1902 | /// step = <VF*UF, VF*UF, ..., VF*UF>. | ||||||||||||
| 1903 | void execute(VPTransformState &State) override; | ||||||||||||
| 1904 | |||||||||||||
| 1905 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1906 | /// Print the recipe. | ||||||||||||
| 1907 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1908 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1909 | #endif | ||||||||||||
| 1910 | |||||||||||||
| 1911 | /// Returns the scalar type of the induction. | ||||||||||||
| 1912 | const Type *getScalarType() const { | ||||||||||||
| 1913 | return cast<VPCanonicalIVPHIRecipe>(getOperand(0)->getDefiningRecipe()) | ||||||||||||
| 1914 | ->getScalarType(); | ||||||||||||
| 1915 | } | ||||||||||||
| 1916 | }; | ||||||||||||
| 1917 | |||||||||||||
| 1918 | /// A recipe for converting the canonical IV value to the corresponding value of | ||||||||||||
| 1919 | /// an IV with different start and step values, using Start + CanonicalIV * | ||||||||||||
| 1920 | /// Step. | ||||||||||||
| 1921 | class VPDerivedIVRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
| 1922 | /// The type of the result value. It may be smaller than the type of the | ||||||||||||
| 1923 | /// induction and in this case it will get truncated to ResultTy. | ||||||||||||
| 1924 | Type *ResultTy; | ||||||||||||
| 1925 | |||||||||||||
| 1926 | /// Induction descriptor for the induction the canonical IV is transformed to. | ||||||||||||
| 1927 | const InductionDescriptor &IndDesc; | ||||||||||||
| 1928 | |||||||||||||
| 1929 | public: | ||||||||||||
| 1930 | VPDerivedIVRecipe(const InductionDescriptor &IndDesc, VPValue *Start, | ||||||||||||
| 1931 | VPCanonicalIVPHIRecipe *CanonicalIV, VPValue *Step, | ||||||||||||
| 1932 | Type *ResultTy) | ||||||||||||
| 1933 | : VPRecipeBase(VPDerivedIVSC, {Start, CanonicalIV, Step}), | ||||||||||||
| 1934 | VPValue(VPVDerivedIVSC, nullptr, this), ResultTy(ResultTy), | ||||||||||||
| 1935 | IndDesc(IndDesc) {} | ||||||||||||
| 1936 | |||||||||||||
| 1937 | ~VPDerivedIVRecipe() override = default; | ||||||||||||
| 1938 | |||||||||||||
| 1939 | VP_CLASSOF_IMPL(VPRecipeBase::VPDerivedIVSC, VPValue::VPVInstructionSC)static inline bool classof(const VPDef *D) { return D->getVPDefID () == VPRecipeBase::VPDerivedIVSC; } static inline bool classof (const VPValue *V) { return V->getVPValueID() == VPValue:: VPVInstructionSC; } static inline bool classof(const VPUser * U) { auto *R = dyn_cast<VPRecipeBase>(U); return R && R->getVPDefID() == VPRecipeBase::VPDerivedIVSC; } static inline bool classof(const VPRecipeBase *R) { return R->getVPDefID () == VPRecipeBase::VPDerivedIVSC; } | ||||||||||||
| 1940 | |||||||||||||
| 1941 | /// Generate the transformed value of the induction at offset StartValue (1. | ||||||||||||
| 1942 | /// operand) + IV (2. operand) * StepValue (3, operand). | ||||||||||||
| 1943 | void execute(VPTransformState &State) override; | ||||||||||||
| 1944 | |||||||||||||
| 1945 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1946 | /// Print the recipe. | ||||||||||||
| 1947 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1948 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1949 | #endif | ||||||||||||
| 1950 | |||||||||||||
| 1951 | VPValue *getStartValue() const { return getOperand(0); } | ||||||||||||
| 1952 | VPValue *getCanonicalIV() const { return getOperand(1); } | ||||||||||||
| 1953 | VPValue *getStepValue() const { return getOperand(2); } | ||||||||||||
| 1954 | |||||||||||||
| 1955 | /// Returns true if the recipe only uses the first lane of operand \p Op. | ||||||||||||
| 1956 | bool onlyFirstLaneUsed(const VPValue *Op) const override { | ||||||||||||
| 1957 | assert(is_contained(operands(), Op) &&(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1958, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 1958 | "Op must be an operand of the recipe")(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1958, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 1959 | return true; | ||||||||||||
| 1960 | } | ||||||||||||
| 1961 | }; | ||||||||||||
| 1962 | |||||||||||||
| 1963 | /// A recipe for handling phi nodes of integer and floating-point inductions, | ||||||||||||
| 1964 | /// producing their scalar values. | ||||||||||||
| 1965 | class VPScalarIVStepsRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
| 1966 | const InductionDescriptor &IndDesc; | ||||||||||||
| 1967 | |||||||||||||
| 1968 | public: | ||||||||||||
| 1969 | VPScalarIVStepsRecipe(const InductionDescriptor &IndDesc, VPValue *IV, | ||||||||||||
| 1970 | VPValue *Step) | ||||||||||||
| 1971 | : VPRecipeBase(VPScalarIVStepsSC, {IV, Step}), VPValue(nullptr, this), | ||||||||||||
| 1972 | IndDesc(IndDesc) {} | ||||||||||||
| 1973 | |||||||||||||
| 1974 | ~VPScalarIVStepsRecipe() override = default; | ||||||||||||
| 1975 | |||||||||||||
| 1976 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
| 1977 | static inline bool classof(const VPDef *D) { | ||||||||||||
| 1978 | return D->getVPDefID() == VPRecipeBase::VPScalarIVStepsSC; | ||||||||||||
| 1979 | } | ||||||||||||
| 1980 | /// Extra classof implementations to allow directly casting from VPUser -> | ||||||||||||
| 1981 | /// VPScalarIVStepsRecipe. | ||||||||||||
| 1982 | static inline bool classof(const VPUser *U) { | ||||||||||||
| 1983 | auto *R = dyn_cast<VPRecipeBase>(U); | ||||||||||||
| 1984 | return R && R->getVPDefID() == VPRecipeBase::VPScalarIVStepsSC; | ||||||||||||
| 1985 | } | ||||||||||||
| 1986 | static inline bool classof(const VPRecipeBase *R) { | ||||||||||||
| 1987 | return R->getVPDefID() == VPRecipeBase::VPScalarIVStepsSC; | ||||||||||||
| 1988 | } | ||||||||||||
| 1989 | |||||||||||||
| 1990 | /// Generate the scalarized versions of the phi node as needed by their users. | ||||||||||||
| 1991 | void execute(VPTransformState &State) override; | ||||||||||||
| 1992 | |||||||||||||
| 1993 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 1994 | /// Print the recipe. | ||||||||||||
| 1995 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 1996 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 1997 | #endif | ||||||||||||
| 1998 | |||||||||||||
| 1999 | VPValue *getStepValue() const { return getOperand(1); } | ||||||||||||
| 2000 | |||||||||||||
| 2001 | /// Returns true if the recipe only uses the first lane of operand \p Op. | ||||||||||||
| 2002 | bool onlyFirstLaneUsed(const VPValue *Op) const override { | ||||||||||||
| 2003 | assert(is_contained(operands(), Op) &&(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2004, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 2004 | "Op must be an operand of the recipe")(static_cast <bool> (is_contained(operands(), Op) && "Op must be an operand of the recipe") ? void (0) : __assert_fail ("is_contained(operands(), Op) && \"Op must be an operand of the recipe\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2004, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2005 | return true; | ||||||||||||
| 2006 | } | ||||||||||||
| 2007 | }; | ||||||||||||
| 2008 | |||||||||||||
| 2009 | /// VPBasicBlock serves as the leaf of the Hierarchical Control-Flow Graph. It | ||||||||||||
| 2010 | /// holds a sequence of zero or more VPRecipe's each representing a sequence of | ||||||||||||
| 2011 | /// output IR instructions. All PHI-like recipes must come before any non-PHI recipes. | ||||||||||||
| 2012 | class VPBasicBlock : public VPBlockBase { | ||||||||||||
| 2013 | public: | ||||||||||||
| 2014 | using RecipeListTy = iplist<VPRecipeBase>; | ||||||||||||
| 2015 | |||||||||||||
| 2016 | private: | ||||||||||||
| 2017 | /// The VPRecipes held in the order of output instructions to generate. | ||||||||||||
| 2018 | RecipeListTy Recipes; | ||||||||||||
| 2019 | |||||||||||||
| 2020 | public: | ||||||||||||
| 2021 | VPBasicBlock(const Twine &Name = "", VPRecipeBase *Recipe = nullptr) | ||||||||||||
| 2022 | : VPBlockBase(VPBasicBlockSC, Name.str()) { | ||||||||||||
| 2023 | if (Recipe) | ||||||||||||
| 2024 | appendRecipe(Recipe); | ||||||||||||
| 2025 | } | ||||||||||||
| 2026 | |||||||||||||
| 2027 | ~VPBasicBlock() override { | ||||||||||||
| 2028 | while (!Recipes.empty()) | ||||||||||||
| 2029 | Recipes.pop_back(); | ||||||||||||
| 2030 | } | ||||||||||||
| 2031 | |||||||||||||
| 2032 | /// Instruction iterators... | ||||||||||||
| 2033 | using iterator = RecipeListTy::iterator; | ||||||||||||
| 2034 | using const_iterator = RecipeListTy::const_iterator; | ||||||||||||
| 2035 | using reverse_iterator = RecipeListTy::reverse_iterator; | ||||||||||||
| 2036 | using const_reverse_iterator = RecipeListTy::const_reverse_iterator; | ||||||||||||
| 2037 | |||||||||||||
| 2038 | //===--------------------------------------------------------------------===// | ||||||||||||
| 2039 | /// Recipe iterator methods | ||||||||||||
| 2040 | /// | ||||||||||||
| 2041 | inline iterator begin() { return Recipes.begin(); } | ||||||||||||
| 2042 | inline const_iterator begin() const { return Recipes.begin(); } | ||||||||||||
| 2043 | inline iterator end() { return Recipes.end(); } | ||||||||||||
| 2044 | inline const_iterator end() const { return Recipes.end(); } | ||||||||||||
| 2045 | |||||||||||||
| 2046 | inline reverse_iterator rbegin() { return Recipes.rbegin(); } | ||||||||||||
| 2047 | inline const_reverse_iterator rbegin() const { return Recipes.rbegin(); } | ||||||||||||
| 2048 | inline reverse_iterator rend() { return Recipes.rend(); } | ||||||||||||
| 2049 | inline const_reverse_iterator rend() const { return Recipes.rend(); } | ||||||||||||
| 2050 | |||||||||||||
| 2051 | inline size_t size() const { return Recipes.size(); } | ||||||||||||
| 2052 | inline bool empty() const { return Recipes.empty(); } | ||||||||||||
| 2053 | inline const VPRecipeBase &front() const { return Recipes.front(); } | ||||||||||||
| 2054 | inline VPRecipeBase &front() { return Recipes.front(); } | ||||||||||||
| 2055 | inline const VPRecipeBase &back() const { return Recipes.back(); } | ||||||||||||
| 2056 | inline VPRecipeBase &back() { return Recipes.back(); } | ||||||||||||
| 2057 | |||||||||||||
| 2058 | /// Returns a reference to the list of recipes. | ||||||||||||
| 2059 | RecipeListTy &getRecipeList() { return Recipes; } | ||||||||||||
| 2060 | |||||||||||||
| 2061 | /// Returns a pointer to a member of the recipe list. | ||||||||||||
| 2062 | static RecipeListTy VPBasicBlock::*getSublistAccess(VPRecipeBase *) { | ||||||||||||
| 2063 | return &VPBasicBlock::Recipes; | ||||||||||||
| 2064 | } | ||||||||||||
| 2065 | |||||||||||||
| 2066 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
| 2067 | static inline bool classof(const VPBlockBase *V) { | ||||||||||||
| 2068 | return V->getVPBlockID() == VPBlockBase::VPBasicBlockSC; | ||||||||||||
| 2069 | } | ||||||||||||
| 2070 | |||||||||||||
| 2071 | void insert(VPRecipeBase *Recipe, iterator InsertPt) { | ||||||||||||
| 2072 | assert(Recipe && "No recipe to append.")(static_cast <bool> (Recipe && "No recipe to append." ) ? void (0) : __assert_fail ("Recipe && \"No recipe to append.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2072, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2073 | assert(!Recipe->Parent && "Recipe already in VPlan")(static_cast <bool> (!Recipe->Parent && "Recipe already in VPlan" ) ? void (0) : __assert_fail ("!Recipe->Parent && \"Recipe already in VPlan\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2073, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2074 | Recipe->Parent = this; | ||||||||||||
| 2075 | Recipes.insert(InsertPt, Recipe); | ||||||||||||
| 2076 | } | ||||||||||||
| 2077 | |||||||||||||
| 2078 | /// Augment the existing recipes of a VPBasicBlock with an additional | ||||||||||||
| 2079 | /// \p Recipe as the last recipe. | ||||||||||||
| 2080 | void appendRecipe(VPRecipeBase *Recipe) { insert(Recipe, end()); } | ||||||||||||
| 2081 | |||||||||||||
| 2082 | /// The method which generates the output IR instructions that correspond to | ||||||||||||
| 2083 | /// this VPBasicBlock, thereby "executing" the VPlan. | ||||||||||||
| 2084 | void execute(VPTransformState *State) override; | ||||||||||||
| 2085 | |||||||||||||
| 2086 | /// Return the position of the first non-phi node recipe in the block. | ||||||||||||
| 2087 | iterator getFirstNonPhi(); | ||||||||||||
| 2088 | |||||||||||||
| 2089 | /// Returns an iterator range over the PHI-like recipes in the block. | ||||||||||||
| 2090 | iterator_range<iterator> phis() { | ||||||||||||
| 2091 | return make_range(begin(), getFirstNonPhi()); | ||||||||||||
| 2092 | } | ||||||||||||
| 2093 | |||||||||||||
| 2094 | void dropAllReferences(VPValue *NewValue) override; | ||||||||||||
| 2095 | |||||||||||||
| 2096 | /// Split current block at \p SplitAt by inserting a new block between the | ||||||||||||
| 2097 | /// current block and its successors and moving all recipes starting at | ||||||||||||
| 2098 | /// SplitAt to the new block. Returns the new block. | ||||||||||||
| 2099 | VPBasicBlock *splitAt(iterator SplitAt); | ||||||||||||
| 2100 | |||||||||||||
| 2101 | VPRegionBlock *getEnclosingLoopRegion(); | ||||||||||||
| 2102 | |||||||||||||
| 2103 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 2104 | /// Print this VPBsicBlock to \p O, prefixing all lines with \p Indent. \p | ||||||||||||
| 2105 | /// SlotTracker is used to print unnamed VPValue's using consequtive numbers. | ||||||||||||
| 2106 | /// | ||||||||||||
| 2107 | /// Note that the numbering is applied to the whole VPlan, so printing | ||||||||||||
| 2108 | /// individual blocks is consistent with the whole VPlan printing. | ||||||||||||
| 2109 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 2110 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 2111 | using VPBlockBase::print; // Get the print(raw_stream &O) version. | ||||||||||||
| 2112 | #endif | ||||||||||||
| 2113 | |||||||||||||
| 2114 | /// If the block has multiple successors, return the branch recipe terminating | ||||||||||||
| 2115 | /// the block. If there are no or only a single successor, return nullptr; | ||||||||||||
| 2116 | VPRecipeBase *getTerminator(); | ||||||||||||
| 2117 | const VPRecipeBase *getTerminator() const; | ||||||||||||
| 2118 | |||||||||||||
| 2119 | /// Returns true if the block is exiting it's parent region. | ||||||||||||
| 2120 | bool isExiting() const; | ||||||||||||
| 2121 | |||||||||||||
| 2122 | private: | ||||||||||||
| 2123 | /// Create an IR BasicBlock to hold the output instructions generated by this | ||||||||||||
| 2124 | /// VPBasicBlock, and return it. Update the CFGState accordingly. | ||||||||||||
| 2125 | BasicBlock *createEmptyBasicBlock(VPTransformState::CFGState &CFG); | ||||||||||||
| 2126 | }; | ||||||||||||
| 2127 | |||||||||||||
| 2128 | /// VPRegionBlock represents a collection of VPBasicBlocks and VPRegionBlocks | ||||||||||||
| 2129 | /// which form a Single-Entry-Single-Exiting subgraph of the output IR CFG. | ||||||||||||
| 2130 | /// A VPRegionBlock may indicate that its contents are to be replicated several | ||||||||||||
| 2131 | /// times. This is designed to support predicated scalarization, in which a | ||||||||||||
| 2132 | /// scalar if-then code structure needs to be generated VF * UF times. Having | ||||||||||||
| 2133 | /// this replication indicator helps to keep a single model for multiple | ||||||||||||
| 2134 | /// candidate VF's. The actual replication takes place only once the desired VF | ||||||||||||
| 2135 | /// and UF have been determined. | ||||||||||||
| 2136 | class VPRegionBlock : public VPBlockBase { | ||||||||||||
| 2137 | /// Hold the Single Entry of the SESE region modelled by the VPRegionBlock. | ||||||||||||
| 2138 | VPBlockBase *Entry; | ||||||||||||
| 2139 | |||||||||||||
| 2140 | /// Hold the Single Exiting block of the SESE region modelled by the | ||||||||||||
| 2141 | /// VPRegionBlock. | ||||||||||||
| 2142 | VPBlockBase *Exiting; | ||||||||||||
| 2143 | |||||||||||||
| 2144 | /// An indicator whether this region is to generate multiple replicated | ||||||||||||
| 2145 | /// instances of output IR corresponding to its VPBlockBases. | ||||||||||||
| 2146 | bool IsReplicator; | ||||||||||||
| 2147 | |||||||||||||
| 2148 | public: | ||||||||||||
| 2149 | VPRegionBlock(VPBlockBase *Entry, VPBlockBase *Exiting, | ||||||||||||
| 2150 | const std::string &Name = "", bool IsReplicator = false) | ||||||||||||
| 2151 | : VPBlockBase(VPRegionBlockSC, Name), Entry(Entry), Exiting(Exiting), | ||||||||||||
| 2152 | IsReplicator(IsReplicator) { | ||||||||||||
| 2153 | assert(Entry->getPredecessors().empty() && "Entry block has predecessors.")(static_cast <bool> (Entry->getPredecessors().empty( ) && "Entry block has predecessors.") ? void (0) : __assert_fail ("Entry->getPredecessors().empty() && \"Entry block has predecessors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2153, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2154 | assert(Exiting->getSuccessors().empty() && "Exit block has successors.")(static_cast <bool> (Exiting->getSuccessors().empty( ) && "Exit block has successors.") ? void (0) : __assert_fail ("Exiting->getSuccessors().empty() && \"Exit block has successors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2154, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2155 | Entry->setParent(this); | ||||||||||||
| 2156 | Exiting->setParent(this); | ||||||||||||
| 2157 | } | ||||||||||||
| 2158 | VPRegionBlock(const std::string &Name = "", bool IsReplicator = false) | ||||||||||||
| 2159 | : VPBlockBase(VPRegionBlockSC, Name), Entry(nullptr), Exiting(nullptr), | ||||||||||||
| 2160 | IsReplicator(IsReplicator) {} | ||||||||||||
| 2161 | |||||||||||||
| 2162 | ~VPRegionBlock() override { | ||||||||||||
| 2163 | if (Entry) { | ||||||||||||
| 2164 | VPValue DummyValue; | ||||||||||||
| 2165 | Entry->dropAllReferences(&DummyValue); | ||||||||||||
| 2166 | deleteCFG(Entry); | ||||||||||||
| 2167 | } | ||||||||||||
| 2168 | } | ||||||||||||
| 2169 | |||||||||||||
| 2170 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
| 2171 | static inline bool classof(const VPBlockBase *V) { | ||||||||||||
| 2172 | return V->getVPBlockID() == VPBlockBase::VPRegionBlockSC; | ||||||||||||
| 2173 | } | ||||||||||||
| 2174 | |||||||||||||
| 2175 | const VPBlockBase *getEntry() const { return Entry; } | ||||||||||||
| 2176 | VPBlockBase *getEntry() { return Entry; } | ||||||||||||
| 2177 | |||||||||||||
| 2178 | /// Set \p EntryBlock as the entry VPBlockBase of this VPRegionBlock. \p | ||||||||||||
| 2179 | /// EntryBlock must have no predecessors. | ||||||||||||
| 2180 | void setEntry(VPBlockBase *EntryBlock) { | ||||||||||||
| 2181 | assert(EntryBlock->getPredecessors().empty() &&(static_cast <bool> (EntryBlock->getPredecessors().empty () && "Entry block cannot have predecessors.") ? void (0) : __assert_fail ("EntryBlock->getPredecessors().empty() && \"Entry block cannot have predecessors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2182, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 2182 | "Entry block cannot have predecessors.")(static_cast <bool> (EntryBlock->getPredecessors().empty () && "Entry block cannot have predecessors.") ? void (0) : __assert_fail ("EntryBlock->getPredecessors().empty() && \"Entry block cannot have predecessors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2182, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2183 | Entry = EntryBlock; | ||||||||||||
| 2184 | EntryBlock->setParent(this); | ||||||||||||
| 2185 | } | ||||||||||||
| 2186 | |||||||||||||
| 2187 | // FIXME: DominatorTreeBase is doing 'A->getParent()->front()'. 'front' is a | ||||||||||||
| 2188 | // specific interface of llvm::Function, instead of using | ||||||||||||
| 2189 | // GraphTraints::getEntryNode. We should add a new template parameter to | ||||||||||||
| 2190 | // DominatorTreeBase representing the Graph type. | ||||||||||||
| 2191 | VPBlockBase &front() const { return *Entry; } | ||||||||||||
| 2192 | |||||||||||||
| 2193 | const VPBlockBase *getExiting() const { return Exiting; } | ||||||||||||
| 2194 | VPBlockBase *getExiting() { return Exiting; } | ||||||||||||
| 2195 | |||||||||||||
| 2196 | /// Set \p ExitingBlock as the exiting VPBlockBase of this VPRegionBlock. \p | ||||||||||||
| 2197 | /// ExitingBlock must have no successors. | ||||||||||||
| 2198 | void setExiting(VPBlockBase *ExitingBlock) { | ||||||||||||
| 2199 | assert(ExitingBlock->getSuccessors().empty() &&(static_cast <bool> (ExitingBlock->getSuccessors().empty () && "Exit block cannot have successors.") ? void (0 ) : __assert_fail ("ExitingBlock->getSuccessors().empty() && \"Exit block cannot have successors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2200, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 2200 | "Exit block cannot have successors.")(static_cast <bool> (ExitingBlock->getSuccessors().empty () && "Exit block cannot have successors.") ? void (0 ) : __assert_fail ("ExitingBlock->getSuccessors().empty() && \"Exit block cannot have successors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2200, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2201 | Exiting = ExitingBlock; | ||||||||||||
| 2202 | ExitingBlock->setParent(this); | ||||||||||||
| 2203 | } | ||||||||||||
| 2204 | |||||||||||||
| 2205 | /// Returns the pre-header VPBasicBlock of the loop region. | ||||||||||||
| 2206 | VPBasicBlock *getPreheaderVPBB() { | ||||||||||||
| 2207 | assert(!isReplicator() && "should only get pre-header of loop regions")(static_cast <bool> (!isReplicator() && "should only get pre-header of loop regions" ) ? void (0) : __assert_fail ("!isReplicator() && \"should only get pre-header of loop regions\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2207, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2208 | return getSinglePredecessor()->getExitingBasicBlock(); | ||||||||||||
| 2209 | } | ||||||||||||
| 2210 | |||||||||||||
| 2211 | /// An indicator whether this region is to generate multiple replicated | ||||||||||||
| 2212 | /// instances of output IR corresponding to its VPBlockBases. | ||||||||||||
| 2213 | bool isReplicator() const { return IsReplicator; } | ||||||||||||
| 2214 | |||||||||||||
| 2215 | /// The method which generates the output IR instructions that correspond to | ||||||||||||
| 2216 | /// this VPRegionBlock, thereby "executing" the VPlan. | ||||||||||||
| 2217 | void execute(VPTransformState *State) override; | ||||||||||||
| 2218 | |||||||||||||
| 2219 | void dropAllReferences(VPValue *NewValue) override; | ||||||||||||
| 2220 | |||||||||||||
| 2221 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 2222 | /// Print this VPRegionBlock to \p O (recursively), prefixing all lines with | ||||||||||||
| 2223 | /// \p Indent. \p SlotTracker is used to print unnamed VPValue's using | ||||||||||||
| 2224 | /// consequtive numbers. | ||||||||||||
| 2225 | /// | ||||||||||||
| 2226 | /// Note that the numbering is applied to the whole VPlan, so printing | ||||||||||||
| 2227 | /// individual regions is consistent with the whole VPlan printing. | ||||||||||||
| 2228 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
| 2229 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
| 2230 | using VPBlockBase::print; // Get the print(raw_stream &O) version. | ||||||||||||
| 2231 | #endif | ||||||||||||
| 2232 | }; | ||||||||||||
| 2233 | |||||||||||||
| 2234 | //===----------------------------------------------------------------------===// | ||||||||||||
| 2235 | // GraphTraits specializations for VPlan Hierarchical Control-Flow Graphs // | ||||||||||||
| 2236 | //===----------------------------------------------------------------------===// | ||||||||||||
| 2237 | |||||||||||||
| 2238 | // The following set of template specializations implement GraphTraits to treat | ||||||||||||
| 2239 | // any VPBlockBase as a node in a graph of VPBlockBases. It's important to note | ||||||||||||
| 2240 | // that VPBlockBase traits don't recurse into VPRegioBlocks, i.e., if the | ||||||||||||
| 2241 | // VPBlockBase is a VPRegionBlock, this specialization provides access to its | ||||||||||||
| 2242 | // successors/predecessors but not to the blocks inside the region. | ||||||||||||
| 2243 | |||||||||||||
| 2244 | template <> struct GraphTraits<VPBlockBase *> { | ||||||||||||
| 2245 | using NodeRef = VPBlockBase *; | ||||||||||||
| 2246 | using ChildIteratorType = SmallVectorImpl<VPBlockBase *>::iterator; | ||||||||||||
| 2247 | |||||||||||||
| 2248 | static NodeRef getEntryNode(NodeRef N) { return N; } | ||||||||||||
| 2249 | |||||||||||||
| 2250 | static inline ChildIteratorType child_begin(NodeRef N) { | ||||||||||||
| 2251 | return N->getSuccessors().begin(); | ||||||||||||
| 2252 | } | ||||||||||||
| 2253 | |||||||||||||
| 2254 | static inline ChildIteratorType child_end(NodeRef N) { | ||||||||||||
| 2255 | return N->getSuccessors().end(); | ||||||||||||
| 2256 | } | ||||||||||||
| 2257 | }; | ||||||||||||
| 2258 | |||||||||||||
| 2259 | template <> struct GraphTraits<const VPBlockBase *> { | ||||||||||||
| 2260 | using NodeRef = const VPBlockBase *; | ||||||||||||
| 2261 | using ChildIteratorType = SmallVectorImpl<VPBlockBase *>::const_iterator; | ||||||||||||
| 2262 | |||||||||||||
| 2263 | static NodeRef getEntryNode(NodeRef N) { return N; } | ||||||||||||
| 2264 | |||||||||||||
| 2265 | static inline ChildIteratorType child_begin(NodeRef N) { | ||||||||||||
| 2266 | return N->getSuccessors().begin(); | ||||||||||||
| 2267 | } | ||||||||||||
| 2268 | |||||||||||||
| 2269 | static inline ChildIteratorType child_end(NodeRef N) { | ||||||||||||
| 2270 | return N->getSuccessors().end(); | ||||||||||||
| 2271 | } | ||||||||||||
| 2272 | }; | ||||||||||||
| 2273 | |||||||||||||
| 2274 | // Inverse order specialization for VPBasicBlocks. Predecessors are used instead | ||||||||||||
| 2275 | // of successors for the inverse traversal. | ||||||||||||
| 2276 | template <> struct GraphTraits<Inverse<VPBlockBase *>> { | ||||||||||||
| 2277 | using NodeRef = VPBlockBase *; | ||||||||||||
| 2278 | using ChildIteratorType = SmallVectorImpl<VPBlockBase *>::iterator; | ||||||||||||
| 2279 | |||||||||||||
| 2280 | static NodeRef getEntryNode(Inverse<NodeRef> B) { return B.Graph; } | ||||||||||||
| 2281 | |||||||||||||
| 2282 | static inline ChildIteratorType child_begin(NodeRef N) { | ||||||||||||
| 2283 | return N->getPredecessors().begin(); | ||||||||||||
| 2284 | } | ||||||||||||
| 2285 | |||||||||||||
| 2286 | static inline ChildIteratorType child_end(NodeRef N) { | ||||||||||||
| 2287 | return N->getPredecessors().end(); | ||||||||||||
| 2288 | } | ||||||||||||
| 2289 | }; | ||||||||||||
| 2290 | |||||||||||||
| 2291 | // The following set of template specializations implement GraphTraits to | ||||||||||||
| 2292 | // treat VPRegionBlock as a graph and recurse inside its nodes. It's important | ||||||||||||
| 2293 | // to note that the blocks inside the VPRegionBlock are treated as VPBlockBases | ||||||||||||
| 2294 | // (i.e., no dyn_cast is performed, VPBlockBases specialization is used), so | ||||||||||||
| 2295 | // there won't be automatic recursion into other VPBlockBases that turn to be | ||||||||||||
| 2296 | // VPRegionBlocks. | ||||||||||||
| 2297 | |||||||||||||
| 2298 | template <> | ||||||||||||
| 2299 | struct GraphTraits<VPRegionBlock *> : public GraphTraits<VPBlockBase *> { | ||||||||||||
| 2300 | using GraphRef = VPRegionBlock *; | ||||||||||||
| 2301 | using nodes_iterator = df_iterator<NodeRef>; | ||||||||||||
| 2302 | |||||||||||||
| 2303 | static NodeRef getEntryNode(GraphRef N) { return N->getEntry(); } | ||||||||||||
| 2304 | |||||||||||||
| 2305 | static nodes_iterator nodes_begin(GraphRef N) { | ||||||||||||
| 2306 | return nodes_iterator::begin(N->getEntry()); | ||||||||||||
| 2307 | } | ||||||||||||
| 2308 | |||||||||||||
| 2309 | static nodes_iterator nodes_end(GraphRef N) { | ||||||||||||
| 2310 | // df_iterator::end() returns an empty iterator so the node used doesn't | ||||||||||||
| 2311 | // matter. | ||||||||||||
| 2312 | return nodes_iterator::end(N); | ||||||||||||
| 2313 | } | ||||||||||||
| 2314 | }; | ||||||||||||
| 2315 | |||||||||||||
| 2316 | template <> | ||||||||||||
| 2317 | struct GraphTraits<const VPRegionBlock *> | ||||||||||||
| 2318 | : public GraphTraits<const VPBlockBase *> { | ||||||||||||
| 2319 | using GraphRef = const VPRegionBlock *; | ||||||||||||
| 2320 | using nodes_iterator = df_iterator<NodeRef>; | ||||||||||||
| 2321 | |||||||||||||
| 2322 | static NodeRef getEntryNode(GraphRef N) { return N->getEntry(); } | ||||||||||||
| 2323 | |||||||||||||
| 2324 | static nodes_iterator nodes_begin(GraphRef N) { | ||||||||||||
| 2325 | return nodes_iterator::begin(N->getEntry()); | ||||||||||||
| 2326 | } | ||||||||||||
| 2327 | |||||||||||||
| 2328 | static nodes_iterator nodes_end(GraphRef N) { | ||||||||||||
| 2329 | // df_iterator::end() returns an empty iterator so the node used doesn't | ||||||||||||
| 2330 | // matter. | ||||||||||||
| 2331 | return nodes_iterator::end(N); | ||||||||||||
| 2332 | } | ||||||||||||
| 2333 | }; | ||||||||||||
| 2334 | |||||||||||||
| 2335 | template <> | ||||||||||||
| 2336 | struct GraphTraits<Inverse<VPRegionBlock *>> | ||||||||||||
| 2337 | : public GraphTraits<Inverse<VPBlockBase *>> { | ||||||||||||
| 2338 | using GraphRef = VPRegionBlock *; | ||||||||||||
| 2339 | using nodes_iterator = df_iterator<NodeRef>; | ||||||||||||
| 2340 | |||||||||||||
| 2341 | static NodeRef getEntryNode(Inverse<GraphRef> N) { | ||||||||||||
| 2342 | return N.Graph->getExiting(); | ||||||||||||
| 2343 | } | ||||||||||||
| 2344 | |||||||||||||
| 2345 | static nodes_iterator nodes_begin(GraphRef N) { | ||||||||||||
| 2346 | return nodes_iterator::begin(N->getExiting()); | ||||||||||||
| 2347 | } | ||||||||||||
| 2348 | |||||||||||||
| 2349 | static nodes_iterator nodes_end(GraphRef N) { | ||||||||||||
| 2350 | // df_iterator::end() returns an empty iterator so the node used doesn't | ||||||||||||
| 2351 | // matter. | ||||||||||||
| 2352 | return nodes_iterator::end(N); | ||||||||||||
| 2353 | } | ||||||||||||
| 2354 | }; | ||||||||||||
| 2355 | |||||||||||||
| 2356 | /// Iterator to traverse all successors of a VPBlockBase node. This includes the | ||||||||||||
| 2357 | /// entry node of VPRegionBlocks. Exit blocks of a region implicitly have their | ||||||||||||
| 2358 | /// parent region's successors. This ensures all blocks in a region are visited | ||||||||||||
| 2359 | /// before any blocks in a successor region when doing a reverse post-order | ||||||||||||
| 2360 | // traversal of the graph. | ||||||||||||
| 2361 | template <typename BlockPtrTy> | ||||||||||||
| 2362 | class VPAllSuccessorsIterator | ||||||||||||
| 2363 | : public iterator_facade_base<VPAllSuccessorsIterator<BlockPtrTy>, | ||||||||||||
| 2364 | std::forward_iterator_tag, VPBlockBase> { | ||||||||||||
| 2365 | BlockPtrTy Block; | ||||||||||||
| 2366 | /// Index of the current successor. For VPBasicBlock nodes, this simply is the | ||||||||||||
| 2367 | /// index for the successor array. For VPRegionBlock, SuccessorIdx == 0 is | ||||||||||||
| 2368 | /// used for the region's entry block, and SuccessorIdx - 1 are the indices | ||||||||||||
| 2369 | /// for the successor array. | ||||||||||||
| 2370 | size_t SuccessorIdx; | ||||||||||||
| 2371 | |||||||||||||
| 2372 | static BlockPtrTy getBlockWithSuccs(BlockPtrTy Current) { | ||||||||||||
| 2373 | while (Current && Current->getNumSuccessors() == 0) | ||||||||||||
| 2374 | Current = Current->getParent(); | ||||||||||||
| 2375 | return Current; | ||||||||||||
| 2376 | } | ||||||||||||
| 2377 | |||||||||||||
| 2378 | /// Templated helper to dereference successor \p SuccIdx of \p Block. Used by | ||||||||||||
| 2379 | /// both the const and non-const operator* implementations. | ||||||||||||
| 2380 | template <typename T1> static T1 deref(T1 Block, unsigned SuccIdx) { | ||||||||||||
| 2381 | if (auto *R = dyn_cast<VPRegionBlock>(Block)) { | ||||||||||||
| 2382 | if (SuccIdx == 0) | ||||||||||||
| 2383 | return R->getEntry(); | ||||||||||||
| 2384 | SuccIdx--; | ||||||||||||
| 2385 | } | ||||||||||||
| 2386 | |||||||||||||
| 2387 | // For exit blocks, use the next parent region with successors. | ||||||||||||
| 2388 | return getBlockWithSuccs(Block)->getSuccessors()[SuccIdx]; | ||||||||||||
| 2389 | } | ||||||||||||
| 2390 | |||||||||||||
| 2391 | public: | ||||||||||||
| 2392 | VPAllSuccessorsIterator(BlockPtrTy Block, size_t Idx = 0) | ||||||||||||
| 2393 | : Block(Block), SuccessorIdx(Idx) {} | ||||||||||||
| 2394 | VPAllSuccessorsIterator(const VPAllSuccessorsIterator &Other) | ||||||||||||
| 2395 | : Block(Other.Block), SuccessorIdx(Other.SuccessorIdx) {} | ||||||||||||
| 2396 | |||||||||||||
| 2397 | VPAllSuccessorsIterator &operator=(const VPAllSuccessorsIterator &R) { | ||||||||||||
| 2398 | Block = R.Block; | ||||||||||||
| 2399 | SuccessorIdx = R.SuccessorIdx; | ||||||||||||
| 2400 | return *this; | ||||||||||||
| 2401 | } | ||||||||||||
| 2402 | |||||||||||||
| 2403 | static VPAllSuccessorsIterator end(BlockPtrTy Block) { | ||||||||||||
| 2404 | BlockPtrTy ParentWithSuccs = getBlockWithSuccs(Block); | ||||||||||||
| 2405 | unsigned NumSuccessors = ParentWithSuccs | ||||||||||||
| 2406 | ? ParentWithSuccs->getNumSuccessors() | ||||||||||||
| 2407 | : Block->getNumSuccessors(); | ||||||||||||
| 2408 | |||||||||||||
| 2409 | if (auto *R = dyn_cast<VPRegionBlock>(Block)) | ||||||||||||
| 2410 | return {R, NumSuccessors + 1}; | ||||||||||||
| 2411 | return {Block, NumSuccessors}; | ||||||||||||
| 2412 | } | ||||||||||||
| 2413 | |||||||||||||
| 2414 | bool operator==(const VPAllSuccessorsIterator &R) const { | ||||||||||||
| 2415 | return Block == R.Block && SuccessorIdx == R.SuccessorIdx; | ||||||||||||
| 2416 | } | ||||||||||||
| 2417 | |||||||||||||
| 2418 | const VPBlockBase *operator*() const { return deref(Block, SuccessorIdx); } | ||||||||||||
| 2419 | |||||||||||||
| 2420 | BlockPtrTy operator*() { return deref(Block, SuccessorIdx); } | ||||||||||||
| 2421 | |||||||||||||
| 2422 | VPAllSuccessorsIterator &operator++() { | ||||||||||||
| 2423 | SuccessorIdx++; | ||||||||||||
| 2424 | return *this; | ||||||||||||
| 2425 | } | ||||||||||||
| 2426 | |||||||||||||
| 2427 | VPAllSuccessorsIterator operator++(int X) { | ||||||||||||
| 2428 | VPAllSuccessorsIterator Orig = *this; | ||||||||||||
| 2429 | SuccessorIdx++; | ||||||||||||
| 2430 | return Orig; | ||||||||||||
| 2431 | } | ||||||||||||
| 2432 | }; | ||||||||||||
| 2433 | |||||||||||||
| 2434 | /// Helper for GraphTraits specialization that traverses through VPRegionBlocks. | ||||||||||||
| 2435 | template <typename BlockTy> class VPBlockRecursiveTraversalWrapper { | ||||||||||||
| 2436 | BlockTy Entry; | ||||||||||||
| 2437 | |||||||||||||
| 2438 | public: | ||||||||||||
| 2439 | VPBlockRecursiveTraversalWrapper(BlockTy Entry) : Entry(Entry) {} | ||||||||||||
| 2440 | BlockTy getEntry() { return Entry; } | ||||||||||||
| 2441 | }; | ||||||||||||
| 2442 | |||||||||||||
| 2443 | /// GraphTraits specialization to recursively traverse VPBlockBase nodes, | ||||||||||||
| 2444 | /// including traversing through VPRegionBlocks. Exit blocks of a region | ||||||||||||
| 2445 | /// implicitly have their parent region's successors. This ensures all blocks in | ||||||||||||
| 2446 | /// a region are visited before any blocks in a successor region when doing a | ||||||||||||
| 2447 | /// reverse post-order traversal of the graph. | ||||||||||||
| 2448 | template <> | ||||||||||||
| 2449 | struct GraphTraits<VPBlockRecursiveTraversalWrapper<VPBlockBase *>> { | ||||||||||||
| 2450 | using NodeRef = VPBlockBase *; | ||||||||||||
| 2451 | using ChildIteratorType = VPAllSuccessorsIterator<VPBlockBase *>; | ||||||||||||
| 2452 | |||||||||||||
| 2453 | static NodeRef | ||||||||||||
| 2454 | getEntryNode(VPBlockRecursiveTraversalWrapper<VPBlockBase *> N) { | ||||||||||||
| 2455 | return N.getEntry(); | ||||||||||||
| 2456 | } | ||||||||||||
| 2457 | |||||||||||||
| 2458 | static inline ChildIteratorType child_begin(NodeRef N) { | ||||||||||||
| 2459 | return ChildIteratorType(N); | ||||||||||||
| 2460 | } | ||||||||||||
| 2461 | |||||||||||||
| 2462 | static inline ChildIteratorType child_end(NodeRef N) { | ||||||||||||
| 2463 | return ChildIteratorType::end(N); | ||||||||||||
| 2464 | } | ||||||||||||
| 2465 | }; | ||||||||||||
| 2466 | |||||||||||||
| 2467 | template <> | ||||||||||||
| 2468 | struct GraphTraits<VPBlockRecursiveTraversalWrapper<const VPBlockBase *>> { | ||||||||||||
| 2469 | using NodeRef = const VPBlockBase *; | ||||||||||||
| 2470 | using ChildIteratorType = VPAllSuccessorsIterator<const VPBlockBase *>; | ||||||||||||
| 2471 | |||||||||||||
| 2472 | static NodeRef | ||||||||||||
| 2473 | getEntryNode(VPBlockRecursiveTraversalWrapper<const VPBlockBase *> N) { | ||||||||||||
| 2474 | return N.getEntry(); | ||||||||||||
| 2475 | } | ||||||||||||
| 2476 | |||||||||||||
| 2477 | static inline ChildIteratorType child_begin(NodeRef N) { | ||||||||||||
| 2478 | return ChildIteratorType(N); | ||||||||||||
| 2479 | } | ||||||||||||
| 2480 | |||||||||||||
| 2481 | static inline ChildIteratorType child_end(NodeRef N) { | ||||||||||||
| 2482 | return ChildIteratorType::end(N); | ||||||||||||
| 2483 | } | ||||||||||||
| 2484 | }; | ||||||||||||
| 2485 | |||||||||||||
| 2486 | /// VPlan models a candidate for vectorization, encoding various decisions take | ||||||||||||
| 2487 | /// to produce efficient output IR, including which branches, basic-blocks and | ||||||||||||
| 2488 | /// output IR instructions to generate, and their cost. VPlan holds a | ||||||||||||
| 2489 | /// Hierarchical-CFG of VPBasicBlocks and VPRegionBlocks rooted at an Entry | ||||||||||||
| 2490 | /// VPBlock. | ||||||||||||
| 2491 | class VPlan { | ||||||||||||
| 2492 | friend class VPlanPrinter; | ||||||||||||
| 2493 | friend class VPSlotTracker; | ||||||||||||
| 2494 | |||||||||||||
| 2495 | /// Hold the single entry to the Hierarchical CFG of the VPlan. | ||||||||||||
| 2496 | VPBlockBase *Entry; | ||||||||||||
| 2497 | |||||||||||||
| 2498 | /// Holds the VFs applicable to this VPlan. | ||||||||||||
| 2499 | SmallSetVector<ElementCount, 2> VFs; | ||||||||||||
| 2500 | |||||||||||||
| 2501 | /// Holds the UFs applicable to this VPlan. If empty, the VPlan is valid for | ||||||||||||
| 2502 | /// any UF. | ||||||||||||
| 2503 | SmallSetVector<unsigned, 2> UFs; | ||||||||||||
| 2504 | |||||||||||||
| 2505 | /// Holds the name of the VPlan, for printing. | ||||||||||||
| 2506 | std::string Name; | ||||||||||||
| 2507 | |||||||||||||
| 2508 | /// Holds all the external definitions created for this VPlan. External | ||||||||||||
| 2509 | /// definitions must be immutable and hold a pointer to their underlying IR. | ||||||||||||
| 2510 | DenseMap<Value *, VPValue *> VPExternalDefs; | ||||||||||||
| 2511 | |||||||||||||
| 2512 | /// Represents the trip count of the original loop, for folding | ||||||||||||
| 2513 | /// the tail. | ||||||||||||
| 2514 | VPValue *TripCount = nullptr; | ||||||||||||
| 2515 | |||||||||||||
| 2516 | /// Represents the backedge taken count of the original loop, for folding | ||||||||||||
| 2517 | /// the tail. It equals TripCount - 1. | ||||||||||||
| 2518 | VPValue *BackedgeTakenCount = nullptr; | ||||||||||||
| 2519 | |||||||||||||
| 2520 | /// Represents the vector trip count. | ||||||||||||
| 2521 | VPValue VectorTripCount; | ||||||||||||
| 2522 | |||||||||||||
| 2523 | /// Holds a mapping between Values and their corresponding VPValue inside | ||||||||||||
| 2524 | /// VPlan. | ||||||||||||
| 2525 | Value2VPValueTy Value2VPValue; | ||||||||||||
| 2526 | |||||||||||||
| 2527 | /// Contains all VPValues that been allocated by addVPValue directly and need | ||||||||||||
| 2528 | /// to be free when the plan's destructor is called. | ||||||||||||
| 2529 | SmallVector<VPValue *, 16> VPValuesToFree; | ||||||||||||
| 2530 | |||||||||||||
| 2531 | /// Indicates whether it is safe use the Value2VPValue mapping or if the | ||||||||||||
| 2532 | /// mapping cannot be used any longer, because it is stale. | ||||||||||||
| 2533 | bool Value2VPValueEnabled = true; | ||||||||||||
| 2534 | |||||||||||||
| 2535 | /// Values used outside the plan. | ||||||||||||
| 2536 | MapVector<PHINode *, VPLiveOut *> LiveOuts; | ||||||||||||
| 2537 | |||||||||||||
| 2538 | public: | ||||||||||||
| 2539 | VPlan(VPBlockBase *Entry = nullptr) : Entry(Entry) { | ||||||||||||
| 2540 | if (Entry) | ||||||||||||
| 2541 | Entry->setPlan(this); | ||||||||||||
| 2542 | } | ||||||||||||
| 2543 | |||||||||||||
| 2544 | ~VPlan() { | ||||||||||||
| 2545 | clearLiveOuts(); | ||||||||||||
| 2546 | |||||||||||||
| 2547 | if (Entry) { | ||||||||||||
| 2548 | VPValue DummyValue; | ||||||||||||
| 2549 | for (VPBlockBase *Block : depth_first(Entry)) | ||||||||||||
| 2550 | Block->dropAllReferences(&DummyValue); | ||||||||||||
| 2551 | |||||||||||||
| 2552 | VPBlockBase::deleteCFG(Entry); | ||||||||||||
| 2553 | } | ||||||||||||
| 2554 | for (VPValue *VPV : VPValuesToFree) | ||||||||||||
| 2555 | delete VPV; | ||||||||||||
| 2556 | if (TripCount) | ||||||||||||
| 2557 | delete TripCount; | ||||||||||||
| 2558 | if (BackedgeTakenCount) | ||||||||||||
| 2559 | delete BackedgeTakenCount; | ||||||||||||
| 2560 | for (auto &P : VPExternalDefs) | ||||||||||||
| 2561 | delete P.second; | ||||||||||||
| 2562 | } | ||||||||||||
| 2563 | |||||||||||||
| 2564 | /// Prepare the plan for execution, setting up the required live-in values. | ||||||||||||
| 2565 | void prepareToExecute(Value *TripCount, Value *VectorTripCount, | ||||||||||||
| 2566 | Value *CanonicalIVStartValue, VPTransformState &State, | ||||||||||||
| 2567 | bool IsEpilogueVectorization); | ||||||||||||
| 2568 | |||||||||||||
| 2569 | /// Generate the IR code for this VPlan. | ||||||||||||
| 2570 | void execute(VPTransformState *State); | ||||||||||||
| 2571 | |||||||||||||
| 2572 | VPBlockBase *getEntry() { return Entry; } | ||||||||||||
| 2573 | const VPBlockBase *getEntry() const { return Entry; } | ||||||||||||
| 2574 | |||||||||||||
| 2575 | VPBlockBase *setEntry(VPBlockBase *Block) { | ||||||||||||
| 2576 | Entry = Block; | ||||||||||||
| 2577 | Block->setPlan(this); | ||||||||||||
| 2578 | return Entry; | ||||||||||||
| 2579 | } | ||||||||||||
| 2580 | |||||||||||||
| 2581 | /// The trip count of the original loop. | ||||||||||||
| 2582 | VPValue *getOrCreateTripCount() { | ||||||||||||
| 2583 | if (!TripCount) | ||||||||||||
| 2584 | TripCount = new VPValue(); | ||||||||||||
| 2585 | return TripCount; | ||||||||||||
| 2586 | } | ||||||||||||
| 2587 | |||||||||||||
| 2588 | /// The backedge taken count of the original loop. | ||||||||||||
| 2589 | VPValue *getOrCreateBackedgeTakenCount() { | ||||||||||||
| 2590 | if (!BackedgeTakenCount) | ||||||||||||
| 2591 | BackedgeTakenCount = new VPValue(); | ||||||||||||
| 2592 | return BackedgeTakenCount; | ||||||||||||
| 2593 | } | ||||||||||||
| 2594 | |||||||||||||
| 2595 | /// The vector trip count. | ||||||||||||
| 2596 | VPValue &getVectorTripCount() { return VectorTripCount; } | ||||||||||||
| 2597 | |||||||||||||
| 2598 | /// Mark the plan to indicate that using Value2VPValue is not safe any | ||||||||||||
| 2599 | /// longer, because it may be stale. | ||||||||||||
| 2600 | void disableValue2VPValue() { Value2VPValueEnabled = false; } | ||||||||||||
| 2601 | |||||||||||||
| 2602 | void addVF(ElementCount VF) { VFs.insert(VF); } | ||||||||||||
| 2603 | |||||||||||||
| 2604 | void setVF(ElementCount VF) { | ||||||||||||
| 2605 | assert(hasVF(VF) && "Cannot set VF not already in plan")(static_cast <bool> (hasVF(VF) && "Cannot set VF not already in plan" ) ? void (0) : __assert_fail ("hasVF(VF) && \"Cannot set VF not already in plan\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2605, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2606 | VFs.clear(); | ||||||||||||
| 2607 | VFs.insert(VF); | ||||||||||||
| 2608 | } | ||||||||||||
| 2609 | |||||||||||||
| 2610 | bool hasVF(ElementCount VF) { return VFs.count(VF); } | ||||||||||||
| 2611 | |||||||||||||
| 2612 | bool hasScalarVFOnly() const { return VFs.size() == 1 && VFs[0].isScalar(); } | ||||||||||||
| 2613 | |||||||||||||
| 2614 | bool hasUF(unsigned UF) const { return UFs.empty() || UFs.contains(UF); } | ||||||||||||
| 2615 | |||||||||||||
| 2616 | void setUF(unsigned UF) { | ||||||||||||
| 2617 | assert(hasUF(UF) && "Cannot set the UF not already in plan")(static_cast <bool> (hasUF(UF) && "Cannot set the UF not already in plan" ) ? void (0) : __assert_fail ("hasUF(UF) && \"Cannot set the UF not already in plan\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2617, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2618 | UFs.clear(); | ||||||||||||
| 2619 | UFs.insert(UF); | ||||||||||||
| 2620 | } | ||||||||||||
| 2621 | |||||||||||||
| 2622 | /// Return a string with the name of the plan and the applicable VFs and UFs. | ||||||||||||
| 2623 | std::string getName() const; | ||||||||||||
| 2624 | |||||||||||||
| 2625 | void setName(const Twine &newName) { Name = newName.str(); } | ||||||||||||
| 2626 | |||||||||||||
| 2627 | /// Get the existing or add a new external definition for \p V. | ||||||||||||
| 2628 | VPValue *getOrAddExternalDef(Value *V) { | ||||||||||||
| 2629 | auto I = VPExternalDefs.insert({V, nullptr}); | ||||||||||||
| 2630 | if (I.second) | ||||||||||||
| 2631 | I.first->second = new VPValue(V); | ||||||||||||
| 2632 | return I.first->second; | ||||||||||||
| 2633 | } | ||||||||||||
| 2634 | |||||||||||||
| 2635 | void addVPValue(Value *V) { | ||||||||||||
| 2636 | assert(Value2VPValueEnabled &&(static_cast <bool> (Value2VPValueEnabled && "IR value to VPValue mapping may be out of date!" ) ? void (0) : __assert_fail ("Value2VPValueEnabled && \"IR value to VPValue mapping may be out of date!\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2637, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 2637 | "IR value to VPValue mapping may be out of date!")(static_cast <bool> (Value2VPValueEnabled && "IR value to VPValue mapping may be out of date!" ) ? void (0) : __assert_fail ("Value2VPValueEnabled && \"IR value to VPValue mapping may be out of date!\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2637, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2638 | assert(V && "Trying to add a null Value to VPlan")(static_cast <bool> (V && "Trying to add a null Value to VPlan" ) ? void (0) : __assert_fail ("V && \"Trying to add a null Value to VPlan\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2638, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2639 | assert(!Value2VPValue.count(V) && "Value already exists in VPlan")(static_cast <bool> (!Value2VPValue.count(V) && "Value already exists in VPlan") ? void (0) : __assert_fail ( "!Value2VPValue.count(V) && \"Value already exists in VPlan\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2639, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2640 | VPValue *VPV = new VPValue(V); | ||||||||||||
| 2641 | Value2VPValue[V] = VPV; | ||||||||||||
| 2642 | VPValuesToFree.push_back(VPV); | ||||||||||||
| 2643 | } | ||||||||||||
| 2644 | |||||||||||||
| 2645 | void addVPValue(Value *V, VPValue *VPV) { | ||||||||||||
| 2646 | assert(Value2VPValueEnabled && "Value2VPValue mapping may be out of date!")(static_cast <bool> (Value2VPValueEnabled && "Value2VPValue mapping may be out of date!" ) ? void (0) : __assert_fail ("Value2VPValueEnabled && \"Value2VPValue mapping may be out of date!\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2646, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2647 | assert(V && "Trying to add a null Value to VPlan")(static_cast <bool> (V && "Trying to add a null Value to VPlan" ) ? void (0) : __assert_fail ("V && \"Trying to add a null Value to VPlan\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2647, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2648 | assert(!Value2VPValue.count(V) && "Value already exists in VPlan")(static_cast <bool> (!Value2VPValue.count(V) && "Value already exists in VPlan") ? void (0) : __assert_fail ( "!Value2VPValue.count(V) && \"Value already exists in VPlan\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2648, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2649 | Value2VPValue[V] = VPV; | ||||||||||||
| 2650 | } | ||||||||||||
| 2651 | |||||||||||||
| 2652 | /// Returns the VPValue for \p V. \p OverrideAllowed can be used to disable | ||||||||||||
| 2653 | /// checking whether it is safe to query VPValues using IR Values. | ||||||||||||
| 2654 | VPValue *getVPValue(Value *V, bool OverrideAllowed = false) { | ||||||||||||
| 2655 | assert((OverrideAllowed || isa<Constant>(V) || Value2VPValueEnabled) &&(static_cast <bool> ((OverrideAllowed || isa<Constant >(V) || Value2VPValueEnabled) && "Value2VPValue mapping may be out of date!" ) ? void (0) : __assert_fail ("(OverrideAllowed || isa<Constant>(V) || Value2VPValueEnabled) && \"Value2VPValue mapping may be out of date!\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2656, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 2656 | "Value2VPValue mapping may be out of date!")(static_cast <bool> ((OverrideAllowed || isa<Constant >(V) || Value2VPValueEnabled) && "Value2VPValue mapping may be out of date!" ) ? void (0) : __assert_fail ("(OverrideAllowed || isa<Constant>(V) || Value2VPValueEnabled) && \"Value2VPValue mapping may be out of date!\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2656, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2657 | assert(V && "Trying to get the VPValue of a null Value")(static_cast <bool> (V && "Trying to get the VPValue of a null Value" ) ? void (0) : __assert_fail ("V && \"Trying to get the VPValue of a null Value\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2657, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2658 | assert(Value2VPValue.count(V) && "Value does not exist in VPlan")(static_cast <bool> (Value2VPValue.count(V) && "Value does not exist in VPlan" ) ? void (0) : __assert_fail ("Value2VPValue.count(V) && \"Value does not exist in VPlan\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2658, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2659 | return Value2VPValue[V]; | ||||||||||||
| 2660 | } | ||||||||||||
| 2661 | |||||||||||||
| 2662 | /// Gets the VPValue or adds a new one (if none exists yet) for \p V. \p | ||||||||||||
| 2663 | /// OverrideAllowed can be used to disable checking whether it is safe to | ||||||||||||
| 2664 | /// query VPValues using IR Values. | ||||||||||||
| 2665 | VPValue *getOrAddVPValue(Value *V, bool OverrideAllowed = false) { | ||||||||||||
| 2666 | assert((OverrideAllowed || isa<Constant>(V) || Value2VPValueEnabled) &&(static_cast <bool> ((OverrideAllowed || isa<Constant >(V) || Value2VPValueEnabled) && "Value2VPValue mapping may be out of date!" ) ? void (0) : __assert_fail ("(OverrideAllowed || isa<Constant>(V) || Value2VPValueEnabled) && \"Value2VPValue mapping may be out of date!\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2667, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 2667 | "Value2VPValue mapping may be out of date!")(static_cast <bool> ((OverrideAllowed || isa<Constant >(V) || Value2VPValueEnabled) && "Value2VPValue mapping may be out of date!" ) ? void (0) : __assert_fail ("(OverrideAllowed || isa<Constant>(V) || Value2VPValueEnabled) && \"Value2VPValue mapping may be out of date!\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2667, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2668 | assert(V && "Trying to get or add the VPValue of a null Value")(static_cast <bool> (V && "Trying to get or add the VPValue of a null Value" ) ? void (0) : __assert_fail ("V && \"Trying to get or add the VPValue of a null Value\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2668, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2669 | if (!Value2VPValue.count(V)) | ||||||||||||
| 2670 | addVPValue(V); | ||||||||||||
| 2671 | return getVPValue(V); | ||||||||||||
| 2672 | } | ||||||||||||
| 2673 | |||||||||||||
| 2674 | void removeVPValueFor(Value *V) { | ||||||||||||
| 2675 | assert(Value2VPValueEnabled &&(static_cast <bool> (Value2VPValueEnabled && "IR value to VPValue mapping may be out of date!" ) ? void (0) : __assert_fail ("Value2VPValueEnabled && \"IR value to VPValue mapping may be out of date!\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2676, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 2676 | "IR value to VPValue mapping may be out of date!")(static_cast <bool> (Value2VPValueEnabled && "IR value to VPValue mapping may be out of date!" ) ? void (0) : __assert_fail ("Value2VPValueEnabled && \"IR value to VPValue mapping may be out of date!\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2676, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2677 | Value2VPValue.erase(V); | ||||||||||||
| 2678 | } | ||||||||||||
| 2679 | |||||||||||||
| 2680 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 2681 | /// Print this VPlan to \p O. | ||||||||||||
| 2682 | void print(raw_ostream &O) const; | ||||||||||||
| 2683 | |||||||||||||
| 2684 | /// Print this VPlan in DOT format to \p O. | ||||||||||||
| 2685 | void printDOT(raw_ostream &O) const; | ||||||||||||
| 2686 | |||||||||||||
| 2687 | /// Dump the plan to stderr (for debugging). | ||||||||||||
| 2688 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const; | ||||||||||||
| 2689 | #endif | ||||||||||||
| 2690 | |||||||||||||
| 2691 | /// Returns a range mapping the values the range \p Operands to their | ||||||||||||
| 2692 | /// corresponding VPValues. | ||||||||||||
| 2693 | iterator_range<mapped_iterator<Use *, std::function<VPValue *(Value *)>>> | ||||||||||||
| 2694 | mapToVPValues(User::op_range Operands) { | ||||||||||||
| 2695 | std::function<VPValue *(Value *)> Fn = [this](Value *Op) { | ||||||||||||
| 2696 | return getOrAddVPValue(Op); | ||||||||||||
| 2697 | }; | ||||||||||||
| 2698 | return map_range(Operands, Fn); | ||||||||||||
| 2699 | } | ||||||||||||
| 2700 | |||||||||||||
| 2701 | /// Returns the VPRegionBlock of the vector loop. | ||||||||||||
| 2702 | VPRegionBlock *getVectorLoopRegion() { | ||||||||||||
| 2703 | return cast<VPRegionBlock>(getEntry()->getSingleSuccessor()); | ||||||||||||
| 2704 | } | ||||||||||||
| 2705 | const VPRegionBlock *getVectorLoopRegion() const { | ||||||||||||
| 2706 | return cast<VPRegionBlock>(getEntry()->getSingleSuccessor()); | ||||||||||||
| 2707 | } | ||||||||||||
| 2708 | |||||||||||||
| 2709 | /// Returns the canonical induction recipe of the vector loop. | ||||||||||||
| 2710 | VPCanonicalIVPHIRecipe *getCanonicalIV() { | ||||||||||||
| 2711 | VPBasicBlock *EntryVPBB = getVectorLoopRegion()->getEntryBasicBlock(); | ||||||||||||
| 2712 | if (EntryVPBB->empty()) { | ||||||||||||
| 2713 | // VPlan native path. | ||||||||||||
| 2714 | EntryVPBB = cast<VPBasicBlock>(EntryVPBB->getSingleSuccessor()); | ||||||||||||
| 2715 | } | ||||||||||||
| 2716 | return cast<VPCanonicalIVPHIRecipe>(&*EntryVPBB->begin()); | ||||||||||||
| 2717 | } | ||||||||||||
| 2718 | |||||||||||||
| 2719 | /// Find and return the VPActiveLaneMaskPHIRecipe from the header - there | ||||||||||||
| 2720 | /// be only one at most. If there isn't one, then return nullptr. | ||||||||||||
| 2721 | VPActiveLaneMaskPHIRecipe *getActiveLaneMaskPhi(); | ||||||||||||
| 2722 | |||||||||||||
| 2723 | void addLiveOut(PHINode *PN, VPValue *V); | ||||||||||||
| 2724 | |||||||||||||
| 2725 | void clearLiveOuts() { | ||||||||||||
| 2726 | for (auto &KV : LiveOuts) | ||||||||||||
| 2727 | delete KV.second; | ||||||||||||
| 2728 | LiveOuts.clear(); | ||||||||||||
| 2729 | } | ||||||||||||
| 2730 | |||||||||||||
| 2731 | void removeLiveOut(PHINode *PN) { | ||||||||||||
| 2732 | delete LiveOuts[PN]; | ||||||||||||
| 2733 | LiveOuts.erase(PN); | ||||||||||||
| 2734 | } | ||||||||||||
| 2735 | |||||||||||||
| 2736 | const MapVector<PHINode *, VPLiveOut *> &getLiveOuts() const { | ||||||||||||
| 2737 | return LiveOuts; | ||||||||||||
| 2738 | } | ||||||||||||
| 2739 | |||||||||||||
| 2740 | private: | ||||||||||||
| 2741 | /// Add to the given dominator tree the header block and every new basic block | ||||||||||||
| 2742 | /// that was created between it and the latch block, inclusive. | ||||||||||||
| 2743 | static void updateDominatorTree(DominatorTree *DT, BasicBlock *LoopLatchBB, | ||||||||||||
| 2744 | BasicBlock *LoopPreHeaderBB, | ||||||||||||
| 2745 | BasicBlock *LoopExitBB); | ||||||||||||
| 2746 | }; | ||||||||||||
| 2747 | |||||||||||||
| 2748 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 2749 | /// VPlanPrinter prints a given VPlan to a given output stream. The printing is | ||||||||||||
| 2750 | /// indented and follows the dot format. | ||||||||||||
| 2751 | class VPlanPrinter { | ||||||||||||
| 2752 | raw_ostream &OS; | ||||||||||||
| 2753 | const VPlan &Plan; | ||||||||||||
| 2754 | unsigned Depth = 0; | ||||||||||||
| 2755 | unsigned TabWidth = 2; | ||||||||||||
| 2756 | std::string Indent; | ||||||||||||
| 2757 | unsigned BID = 0; | ||||||||||||
| 2758 | SmallDenseMap<const VPBlockBase *, unsigned> BlockID; | ||||||||||||
| 2759 | |||||||||||||
| 2760 | VPSlotTracker SlotTracker; | ||||||||||||
| 2761 | |||||||||||||
| 2762 | /// Handle indentation. | ||||||||||||
| 2763 | void bumpIndent(int b) { Indent = std::string((Depth += b) * TabWidth, ' '); } | ||||||||||||
| 2764 | |||||||||||||
| 2765 | /// Print a given \p Block of the Plan. | ||||||||||||
| 2766 | void dumpBlock(const VPBlockBase *Block); | ||||||||||||
| 2767 | |||||||||||||
| 2768 | /// Print the information related to the CFG edges going out of a given | ||||||||||||
| 2769 | /// \p Block, followed by printing the successor blocks themselves. | ||||||||||||
| 2770 | void dumpEdges(const VPBlockBase *Block); | ||||||||||||
| 2771 | |||||||||||||
| 2772 | /// Print a given \p BasicBlock, including its VPRecipes, followed by printing | ||||||||||||
| 2773 | /// its successor blocks. | ||||||||||||
| 2774 | void dumpBasicBlock(const VPBasicBlock *BasicBlock); | ||||||||||||
| 2775 | |||||||||||||
| 2776 | /// Print a given \p Region of the Plan. | ||||||||||||
| 2777 | void dumpRegion(const VPRegionBlock *Region); | ||||||||||||
| 2778 | |||||||||||||
| 2779 | unsigned getOrCreateBID(const VPBlockBase *Block) { | ||||||||||||
| 2780 | return BlockID.count(Block) ? BlockID[Block] : BlockID[Block] = BID++; | ||||||||||||
| 2781 | } | ||||||||||||
| 2782 | |||||||||||||
| 2783 | Twine getOrCreateName(const VPBlockBase *Block); | ||||||||||||
| 2784 | |||||||||||||
| 2785 | Twine getUID(const VPBlockBase *Block); | ||||||||||||
| 2786 | |||||||||||||
| 2787 | /// Print the information related to a CFG edge between two VPBlockBases. | ||||||||||||
| 2788 | void drawEdge(const VPBlockBase *From, const VPBlockBase *To, bool Hidden, | ||||||||||||
| 2789 | const Twine &Label); | ||||||||||||
| 2790 | |||||||||||||
| 2791 | public: | ||||||||||||
| 2792 | VPlanPrinter(raw_ostream &O, const VPlan &P) | ||||||||||||
| 2793 | : OS(O), Plan(P), SlotTracker(&P) {} | ||||||||||||
| 2794 | |||||||||||||
| 2795 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump(); | ||||||||||||
| 2796 | }; | ||||||||||||
| 2797 | |||||||||||||
| 2798 | struct VPlanIngredient { | ||||||||||||
| 2799 | const Value *V; | ||||||||||||
| 2800 | |||||||||||||
| 2801 | VPlanIngredient(const Value *V) : V(V) {} | ||||||||||||
| 2802 | |||||||||||||
| 2803 | void print(raw_ostream &O) const; | ||||||||||||
| 2804 | }; | ||||||||||||
| 2805 | |||||||||||||
| 2806 | inline raw_ostream &operator<<(raw_ostream &OS, const VPlanIngredient &I) { | ||||||||||||
| 2807 | I.print(OS); | ||||||||||||
| 2808 | return OS; | ||||||||||||
| 2809 | } | ||||||||||||
| 2810 | |||||||||||||
| 2811 | inline raw_ostream &operator<<(raw_ostream &OS, const VPlan &Plan) { | ||||||||||||
| 2812 | Plan.print(OS); | ||||||||||||
| 2813 | return OS; | ||||||||||||
| 2814 | } | ||||||||||||
| 2815 | #endif | ||||||||||||
| 2816 | |||||||||||||
| 2817 | //===----------------------------------------------------------------------===// | ||||||||||||
| 2818 | // VPlan Utilities | ||||||||||||
| 2819 | //===----------------------------------------------------------------------===// | ||||||||||||
| 2820 | |||||||||||||
| 2821 | /// Class that provides utilities for VPBlockBases in VPlan. | ||||||||||||
| 2822 | class VPBlockUtils { | ||||||||||||
| 2823 | public: | ||||||||||||
| 2824 | VPBlockUtils() = delete; | ||||||||||||
| 2825 | |||||||||||||
| 2826 | /// Insert disconnected VPBlockBase \p NewBlock after \p BlockPtr. Add \p | ||||||||||||
| 2827 | /// NewBlock as successor of \p BlockPtr and \p BlockPtr as predecessor of \p | ||||||||||||
| 2828 | /// NewBlock, and propagate \p BlockPtr parent to \p NewBlock. \p BlockPtr's | ||||||||||||
| 2829 | /// successors are moved from \p BlockPtr to \p NewBlock. \p NewBlock must | ||||||||||||
| 2830 | /// have neither successors nor predecessors. | ||||||||||||
| 2831 | static void insertBlockAfter(VPBlockBase *NewBlock, VPBlockBase *BlockPtr) { | ||||||||||||
| 2832 | assert(NewBlock->getSuccessors().empty() &&(static_cast <bool> (NewBlock->getSuccessors().empty () && NewBlock->getPredecessors().empty() && "Can't insert new block with predecessors or successors.") ? void (0) : __assert_fail ("NewBlock->getSuccessors().empty() && NewBlock->getPredecessors().empty() && \"Can't insert new block with predecessors or successors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2834, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 2833 | NewBlock->getPredecessors().empty() &&(static_cast <bool> (NewBlock->getSuccessors().empty () && NewBlock->getPredecessors().empty() && "Can't insert new block with predecessors or successors.") ? void (0) : __assert_fail ("NewBlock->getSuccessors().empty() && NewBlock->getPredecessors().empty() && \"Can't insert new block with predecessors or successors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2834, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 2834 | "Can't insert new block with predecessors or successors.")(static_cast <bool> (NewBlock->getSuccessors().empty () && NewBlock->getPredecessors().empty() && "Can't insert new block with predecessors or successors.") ? void (0) : __assert_fail ("NewBlock->getSuccessors().empty() && NewBlock->getPredecessors().empty() && \"Can't insert new block with predecessors or successors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2834, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2835 | NewBlock->setParent(BlockPtr->getParent()); | ||||||||||||
| 2836 | SmallVector<VPBlockBase *> Succs(BlockPtr->successors()); | ||||||||||||
| 2837 | for (VPBlockBase *Succ : Succs) { | ||||||||||||
| 2838 | disconnectBlocks(BlockPtr, Succ); | ||||||||||||
| 2839 | connectBlocks(NewBlock, Succ); | ||||||||||||
| 2840 | } | ||||||||||||
| 2841 | connectBlocks(BlockPtr, NewBlock); | ||||||||||||
| 2842 | } | ||||||||||||
| 2843 | |||||||||||||
| 2844 | /// Insert disconnected VPBlockBases \p IfTrue and \p IfFalse after \p | ||||||||||||
| 2845 | /// BlockPtr. Add \p IfTrue and \p IfFalse as succesors of \p BlockPtr and \p | ||||||||||||
| 2846 | /// BlockPtr as predecessor of \p IfTrue and \p IfFalse. Propagate \p BlockPtr | ||||||||||||
| 2847 | /// parent to \p IfTrue and \p IfFalse. \p BlockPtr must have no successors | ||||||||||||
| 2848 | /// and \p IfTrue and \p IfFalse must have neither successors nor | ||||||||||||
| 2849 | /// predecessors. | ||||||||||||
| 2850 | static void insertTwoBlocksAfter(VPBlockBase *IfTrue, VPBlockBase *IfFalse, | ||||||||||||
| 2851 | VPBlockBase *BlockPtr) { | ||||||||||||
| 2852 | assert(IfTrue->getSuccessors().empty() &&(static_cast <bool> (IfTrue->getSuccessors().empty() && "Can't insert IfTrue with successors.") ? void (0 ) : __assert_fail ("IfTrue->getSuccessors().empty() && \"Can't insert IfTrue with successors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2853, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 2853 | "Can't insert IfTrue with successors.")(static_cast <bool> (IfTrue->getSuccessors().empty() && "Can't insert IfTrue with successors.") ? void (0 ) : __assert_fail ("IfTrue->getSuccessors().empty() && \"Can't insert IfTrue with successors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2853, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2854 | assert(IfFalse->getSuccessors().empty() &&(static_cast <bool> (IfFalse->getSuccessors().empty( ) && "Can't insert IfFalse with successors.") ? void ( 0) : __assert_fail ("IfFalse->getSuccessors().empty() && \"Can't insert IfFalse with successors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2855, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 2855 | "Can't insert IfFalse with successors.")(static_cast <bool> (IfFalse->getSuccessors().empty( ) && "Can't insert IfFalse with successors.") ? void ( 0) : __assert_fail ("IfFalse->getSuccessors().empty() && \"Can't insert IfFalse with successors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2855, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2856 | BlockPtr->setTwoSuccessors(IfTrue, IfFalse); | ||||||||||||
| 2857 | IfTrue->setPredecessors({BlockPtr}); | ||||||||||||
| 2858 | IfFalse->setPredecessors({BlockPtr}); | ||||||||||||
| 2859 | IfTrue->setParent(BlockPtr->getParent()); | ||||||||||||
| 2860 | IfFalse->setParent(BlockPtr->getParent()); | ||||||||||||
| 2861 | } | ||||||||||||
| 2862 | |||||||||||||
| 2863 | /// Connect VPBlockBases \p From and \p To bi-directionally. Append \p To to | ||||||||||||
| 2864 | /// the successors of \p From and \p From to the predecessors of \p To. Both | ||||||||||||
| 2865 | /// VPBlockBases must have the same parent, which can be null. Both | ||||||||||||
| 2866 | /// VPBlockBases can be already connected to other VPBlockBases. | ||||||||||||
| 2867 | static void connectBlocks(VPBlockBase *From, VPBlockBase *To) { | ||||||||||||
| 2868 | assert((From->getParent() == To->getParent()) &&(static_cast <bool> ((From->getParent() == To->getParent ()) && "Can't connect two block with different parents" ) ? void (0) : __assert_fail ("(From->getParent() == To->getParent()) && \"Can't connect two block with different parents\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2869, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 2869 | "Can't connect two block with different parents")(static_cast <bool> ((From->getParent() == To->getParent ()) && "Can't connect two block with different parents" ) ? void (0) : __assert_fail ("(From->getParent() == To->getParent()) && \"Can't connect two block with different parents\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2869, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2870 | assert(From->getNumSuccessors() < 2 &&(static_cast <bool> (From->getNumSuccessors() < 2 && "Blocks can't have more than two successors.") ? void (0) : __assert_fail ("From->getNumSuccessors() < 2 && \"Blocks can't have more than two successors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2871, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
| 2871 | "Blocks can't have more than two successors.")(static_cast <bool> (From->getNumSuccessors() < 2 && "Blocks can't have more than two successors.") ? void (0) : __assert_fail ("From->getNumSuccessors() < 2 && \"Blocks can't have more than two successors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2871, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2872 | From->appendSuccessor(To); | ||||||||||||
| 2873 | To->appendPredecessor(From); | ||||||||||||
| 2874 | } | ||||||||||||
| 2875 | |||||||||||||
| 2876 | /// Disconnect VPBlockBases \p From and \p To bi-directionally. Remove \p To | ||||||||||||
| 2877 | /// from the successors of \p From and \p From from the predecessors of \p To. | ||||||||||||
| 2878 | static void disconnectBlocks(VPBlockBase *From, VPBlockBase *To) { | ||||||||||||
| 2879 | assert(To && "Successor to disconnect is null.")(static_cast <bool> (To && "Successor to disconnect is null." ) ? void (0) : __assert_fail ("To && \"Successor to disconnect is null.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2879, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 2880 | From->removeSuccessor(To); | ||||||||||||
| 2881 | To->removePredecessor(From); | ||||||||||||
| 2882 | } | ||||||||||||
| 2883 | |||||||||||||
| 2884 | /// Try to merge \p Block into its single predecessor, if \p Block is a | ||||||||||||
| 2885 | /// VPBasicBlock and its predecessor has a single successor. Returns a pointer | ||||||||||||
| 2886 | /// to the predecessor \p Block was merged into or nullptr otherwise. | ||||||||||||
| 2887 | static VPBasicBlock *tryToMergeBlockIntoPredecessor(VPBlockBase *Block) { | ||||||||||||
| 2888 | auto *VPBB = dyn_cast<VPBasicBlock>(Block); | ||||||||||||
| 2889 | auto *PredVPBB = | ||||||||||||
| 2890 | dyn_cast_or_null<VPBasicBlock>(Block->getSinglePredecessor()); | ||||||||||||
| 2891 | if (!VPBB
| ||||||||||||
| 2892 | return nullptr; | ||||||||||||
| 2893 | |||||||||||||
| 2894 | for (VPRecipeBase &R : make_early_inc_range(*VPBB)) | ||||||||||||
| 2895 | R.moveBefore(*PredVPBB, PredVPBB->end()); | ||||||||||||
| 2896 | VPBlockUtils::disconnectBlocks(PredVPBB, VPBB); | ||||||||||||
| 2897 | auto *ParentRegion = cast<VPRegionBlock>(Block->getParent()); | ||||||||||||
| 2898 | if (ParentRegion->getExiting() == Block) | ||||||||||||
| 2899 | ParentRegion->setExiting(PredVPBB); | ||||||||||||
| 2900 | SmallVector<VPBlockBase *> Successors(Block->successors()); | ||||||||||||
| 2901 | for (auto *Succ : Successors) { | ||||||||||||
| 2902 | VPBlockUtils::disconnectBlocks(Block, Succ); | ||||||||||||
| 2903 | VPBlockUtils::connectBlocks(PredVPBB, Succ); | ||||||||||||
| 2904 | } | ||||||||||||
| 2905 | delete Block; | ||||||||||||
| 2906 | return PredVPBB; | ||||||||||||
| 2907 | } | ||||||||||||
| 2908 | |||||||||||||
| 2909 | /// Return an iterator range over \p Range which only includes \p BlockTy | ||||||||||||
| 2910 | /// blocks. The accesses are casted to \p BlockTy. | ||||||||||||
| 2911 | template <typename BlockTy, typename T> | ||||||||||||
| 2912 | static auto blocksOnly(const T &Range) { | ||||||||||||
| 2913 | // Create BaseTy with correct const-ness based on BlockTy. | ||||||||||||
| 2914 | using BaseTy = std::conditional_t<std::is_const<BlockTy>::value, | ||||||||||||
| 2915 | const VPBlockBase, VPBlockBase>; | ||||||||||||
| 2916 | |||||||||||||
| 2917 | // We need to first create an iterator range over (const) BlocktTy & instead | ||||||||||||
| 2918 | // of (const) BlockTy * for filter_range to work properly. | ||||||||||||
| 2919 | auto Mapped = | ||||||||||||
| 2920 | map_range(Range, [](BaseTy *Block) -> BaseTy & { return *Block; }); | ||||||||||||
| 2921 | auto Filter = make_filter_range( | ||||||||||||
| 2922 | Mapped, [](BaseTy &Block) { return isa<BlockTy>(&Block); }); | ||||||||||||
| 2923 | return map_range(Filter, [](BaseTy &Block) -> BlockTy * { | ||||||||||||
| 2924 | return cast<BlockTy>(&Block); | ||||||||||||
| 2925 | }); | ||||||||||||
| 2926 | } | ||||||||||||
| 2927 | }; | ||||||||||||
| 2928 | |||||||||||||
| 2929 | class VPInterleavedAccessInfo { | ||||||||||||
| 2930 | DenseMap<VPInstruction *, InterleaveGroup<VPInstruction> *> | ||||||||||||
| 2931 | InterleaveGroupMap; | ||||||||||||
| 2932 | |||||||||||||
| 2933 | /// Type for mapping of instruction based interleave groups to VPInstruction | ||||||||||||
| 2934 | /// interleave groups | ||||||||||||
| 2935 | using Old2NewTy = DenseMap<InterleaveGroup<Instruction> *, | ||||||||||||
| 2936 | InterleaveGroup<VPInstruction> *>; | ||||||||||||
| 2937 | |||||||||||||
| 2938 | /// Recursively \p Region and populate VPlan based interleave groups based on | ||||||||||||
| 2939 | /// \p IAI. | ||||||||||||
| 2940 | void visitRegion(VPRegionBlock *Region, Old2NewTy &Old2New, | ||||||||||||
| 2941 | InterleavedAccessInfo &IAI); | ||||||||||||
| 2942 | /// Recursively traverse \p Block and populate VPlan based interleave groups | ||||||||||||
| 2943 | /// based on \p IAI. | ||||||||||||
| 2944 | void visitBlock(VPBlockBase *Block, Old2NewTy &Old2New, | ||||||||||||
| 2945 | InterleavedAccessInfo &IAI); | ||||||||||||
| 2946 | |||||||||||||
| 2947 | public: | ||||||||||||
| 2948 | VPInterleavedAccessInfo(VPlan &Plan, InterleavedAccessInfo &IAI); | ||||||||||||
| 2949 | |||||||||||||
| 2950 | ~VPInterleavedAccessInfo() { | ||||||||||||
| 2951 | SmallPtrSet<InterleaveGroup<VPInstruction> *, 4> DelSet; | ||||||||||||
| 2952 | // Avoid releasing a pointer twice. | ||||||||||||
| 2953 | for (auto &I : InterleaveGroupMap) | ||||||||||||
| 2954 | DelSet.insert(I.second); | ||||||||||||
| 2955 | for (auto *Ptr : DelSet) | ||||||||||||
| 2956 | delete Ptr; | ||||||||||||
| 2957 | } | ||||||||||||
| 2958 | |||||||||||||
| 2959 | /// Get the interleave group that \p Instr belongs to. | ||||||||||||
| 2960 | /// | ||||||||||||
| 2961 | /// \returns nullptr if doesn't have such group. | ||||||||||||
| 2962 | InterleaveGroup<VPInstruction> * | ||||||||||||
| 2963 | getInterleaveGroup(VPInstruction *Instr) const { | ||||||||||||
| 2964 | return InterleaveGroupMap.lookup(Instr); | ||||||||||||
| 2965 | } | ||||||||||||
| 2966 | }; | ||||||||||||
| 2967 | |||||||||||||
| 2968 | /// Class that maps (parts of) an existing VPlan to trees of combined | ||||||||||||
| 2969 | /// VPInstructions. | ||||||||||||
| 2970 | class VPlanSlp { | ||||||||||||
| 2971 | enum class OpMode { Failed, Load, Opcode }; | ||||||||||||
| 2972 | |||||||||||||
| 2973 | /// A DenseMapInfo implementation for using SmallVector<VPValue *, 4> as | ||||||||||||
| 2974 | /// DenseMap keys. | ||||||||||||
| 2975 | struct BundleDenseMapInfo { | ||||||||||||
| 2976 | static SmallVector<VPValue *, 4> getEmptyKey() { | ||||||||||||
| 2977 | return {reinterpret_cast<VPValue *>(-1)}; | ||||||||||||
| 2978 | } | ||||||||||||
| 2979 | |||||||||||||
| 2980 | static SmallVector<VPValue *, 4> getTombstoneKey() { | ||||||||||||
| 2981 | return {reinterpret_cast<VPValue *>(-2)}; | ||||||||||||
| 2982 | } | ||||||||||||
| 2983 | |||||||||||||
| 2984 | static unsigned getHashValue(const SmallVector<VPValue *, 4> &V) { | ||||||||||||
| 2985 | return static_cast<unsigned>(hash_combine_range(V.begin(), V.end())); | ||||||||||||
| 2986 | } | ||||||||||||
| 2987 | |||||||||||||
| 2988 | static bool isEqual(const SmallVector<VPValue *, 4> &LHS, | ||||||||||||
| 2989 | const SmallVector<VPValue *, 4> &RHS) { | ||||||||||||
| 2990 | return LHS == RHS; | ||||||||||||
| 2991 | } | ||||||||||||
| 2992 | }; | ||||||||||||
| 2993 | |||||||||||||
| 2994 | /// Mapping of values in the original VPlan to a combined VPInstruction. | ||||||||||||
| 2995 | DenseMap<SmallVector<VPValue *, 4>, VPInstruction *, BundleDenseMapInfo> | ||||||||||||
| 2996 | BundleToCombined; | ||||||||||||
| 2997 | |||||||||||||
| 2998 | VPInterleavedAccessInfo &IAI; | ||||||||||||
| 2999 | |||||||||||||
| 3000 | /// Basic block to operate on. For now, only instructions in a single BB are | ||||||||||||
| 3001 | /// considered. | ||||||||||||
| 3002 | const VPBasicBlock &BB; | ||||||||||||
| 3003 | |||||||||||||
| 3004 | /// Indicates whether we managed to combine all visited instructions or not. | ||||||||||||
| 3005 | bool CompletelySLP = true; | ||||||||||||
| 3006 | |||||||||||||
| 3007 | /// Width of the widest combined bundle in bits. | ||||||||||||
| 3008 | unsigned WidestBundleBits = 0; | ||||||||||||
| 3009 | |||||||||||||
| 3010 | using MultiNodeOpTy = | ||||||||||||
| 3011 | typename std::pair<VPInstruction *, SmallVector<VPValue *, 4>>; | ||||||||||||
| 3012 | |||||||||||||
| 3013 | // Input operand bundles for the current multi node. Each multi node operand | ||||||||||||
| 3014 | // bundle contains values not matching the multi node's opcode. They will | ||||||||||||
| 3015 | // be reordered in reorderMultiNodeOps, once we completed building a | ||||||||||||
| 3016 | // multi node. | ||||||||||||
| 3017 | SmallVector<MultiNodeOpTy, 4> MultiNodeOps; | ||||||||||||
| 3018 | |||||||||||||
| 3019 | /// Indicates whether we are building a multi node currently. | ||||||||||||
| 3020 | bool MultiNodeActive = false; | ||||||||||||
| 3021 | |||||||||||||
| 3022 | /// Check if we can vectorize Operands together. | ||||||||||||
| 3023 | bool areVectorizable(ArrayRef<VPValue *> Operands) const; | ||||||||||||
| 3024 | |||||||||||||
| 3025 | /// Add combined instruction \p New for the bundle \p Operands. | ||||||||||||
| 3026 | void addCombined(ArrayRef<VPValue *> Operands, VPInstruction *New); | ||||||||||||
| 3027 | |||||||||||||
| 3028 | /// Indicate we hit a bundle we failed to combine. Returns nullptr for now. | ||||||||||||
| 3029 | VPInstruction *markFailed(); | ||||||||||||
| 3030 | |||||||||||||
| 3031 | /// Reorder operands in the multi node to maximize sequential memory access | ||||||||||||
| 3032 | /// and commutative operations. | ||||||||||||
| 3033 | SmallVector<MultiNodeOpTy, 4> reorderMultiNodeOps(); | ||||||||||||
| 3034 | |||||||||||||
| 3035 | /// Choose the best candidate to use for the lane after \p Last. The set of | ||||||||||||
| 3036 | /// candidates to choose from are values with an opcode matching \p Last's | ||||||||||||
| 3037 | /// or loads consecutive to \p Last. | ||||||||||||
| 3038 | std::pair<OpMode, VPValue *> getBest(OpMode Mode, VPValue *Last, | ||||||||||||
| 3039 | SmallPtrSetImpl<VPValue *> &Candidates, | ||||||||||||
| 3040 | VPInterleavedAccessInfo &IAI); | ||||||||||||
| 3041 | |||||||||||||
| 3042 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
| 3043 | /// Print bundle \p Values to dbgs(). | ||||||||||||
| 3044 | void dumpBundle(ArrayRef<VPValue *> Values); | ||||||||||||
| 3045 | #endif | ||||||||||||
| 3046 | |||||||||||||
| 3047 | public: | ||||||||||||
| 3048 | VPlanSlp(VPInterleavedAccessInfo &IAI, VPBasicBlock &BB) : IAI(IAI), BB(BB) {} | ||||||||||||
| 3049 | |||||||||||||
| 3050 | ~VPlanSlp() = default; | ||||||||||||
| 3051 | |||||||||||||
| 3052 | /// Tries to build an SLP tree rooted at \p Operands and returns a | ||||||||||||
| 3053 | /// VPInstruction combining \p Operands, if they can be combined. | ||||||||||||
| 3054 | VPInstruction *buildGraph(ArrayRef<VPValue *> Operands); | ||||||||||||
| 3055 | |||||||||||||
| 3056 | /// Return the width of the widest combined bundle in bits. | ||||||||||||
| 3057 | unsigned getWidestBundleBits() const { return WidestBundleBits; } | ||||||||||||
| 3058 | |||||||||||||
| 3059 | /// Return true if all visited instruction can be combined. | ||||||||||||
| 3060 | bool isCompletelySLP() const { return CompletelySLP; } | ||||||||||||
| 3061 | }; | ||||||||||||
| 3062 | |||||||||||||
| 3063 | namespace vputils { | ||||||||||||
| 3064 | |||||||||||||
| 3065 | /// Returns true if only the first lane of \p Def is used. | ||||||||||||
| 3066 | bool onlyFirstLaneUsed(VPValue *Def); | ||||||||||||
| 3067 | |||||||||||||
| 3068 | /// Get or create a VPValue that corresponds to the expansion of \p Expr. If \p | ||||||||||||
| 3069 | /// Expr is a SCEVConstant or SCEVUnknown, return a VPValue wrapping the live-in | ||||||||||||
| 3070 | /// value. Otherwise return a VPExpandSCEVRecipe to expand \p Expr. If \p Plan's | ||||||||||||
| 3071 | /// pre-header already contains a recipe expanding \p Expr, return it. If not, | ||||||||||||
| 3072 | /// create a new one. | ||||||||||||
| 3073 | VPValue *getOrCreateVPValueForSCEVExpr(VPlan &Plan, const SCEV *Expr, | ||||||||||||
| 3074 | ScalarEvolution &SE); | ||||||||||||
| 3075 | |||||||||||||
| 3076 | /// Returns true if \p VPV is uniform after vectorization. | ||||||||||||
| 3077 | inline bool isUniformAfterVectorization(VPValue *VPV) { | ||||||||||||
| 3078 | // A value defined outside the vector region must be uniform after | ||||||||||||
| 3079 | // vectorization inside a vector region. | ||||||||||||
| 3080 | if (VPV->isDefinedOutsideVectorRegions()) | ||||||||||||
| 3081 | return true; | ||||||||||||
| 3082 | VPRecipeBase *Def = VPV->getDefiningRecipe(); | ||||||||||||
| 3083 | assert(Def && "Must have definition for value defined inside vector region")(static_cast <bool> (Def && "Must have definition for value defined inside vector region" ) ? void (0) : __assert_fail ("Def && \"Must have definition for value defined inside vector region\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 3083, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
| 3084 | if (auto Rep = dyn_cast<VPReplicateRecipe>(Def)) | ||||||||||||
| 3085 | return Rep->isUniform(); | ||||||||||||
| 3086 | return false; | ||||||||||||
| 3087 | } | ||||||||||||
| 3088 | } // end namespace vputils | ||||||||||||
| 3089 | |||||||||||||
| 3090 | } // end namespace llvm | ||||||||||||
| 3091 | |||||||||||||
| 3092 | #endif // LLVM_TRANSFORMS_VECTORIZE_VPLAN_H |