File: | llvm/lib/Transforms/Vectorize/LoopVectorize.cpp |
Warning: | line 9215, column 34 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 "VPlanPredicator.h" | ||||||
62 | #include "VPlanTransforms.h" | ||||||
63 | #include "llvm/ADT/APInt.h" | ||||||
64 | #include "llvm/ADT/ArrayRef.h" | ||||||
65 | #include "llvm/ADT/DenseMap.h" | ||||||
66 | #include "llvm/ADT/DenseMapInfo.h" | ||||||
67 | #include "llvm/ADT/Hashing.h" | ||||||
68 | #include "llvm/ADT/MapVector.h" | ||||||
69 | #include "llvm/ADT/None.h" | ||||||
70 | #include "llvm/ADT/Optional.h" | ||||||
71 | #include "llvm/ADT/STLExtras.h" | ||||||
72 | #include "llvm/ADT/SmallPtrSet.h" | ||||||
73 | #include "llvm/ADT/SmallSet.h" | ||||||
74 | #include "llvm/ADT/SmallVector.h" | ||||||
75 | #include "llvm/ADT/Statistic.h" | ||||||
76 | #include "llvm/ADT/StringRef.h" | ||||||
77 | #include "llvm/ADT/Twine.h" | ||||||
78 | #include "llvm/ADT/iterator_range.h" | ||||||
79 | #include "llvm/Analysis/AssumptionCache.h" | ||||||
80 | #include "llvm/Analysis/BasicAliasAnalysis.h" | ||||||
81 | #include "llvm/Analysis/BlockFrequencyInfo.h" | ||||||
82 | #include "llvm/Analysis/CFG.h" | ||||||
83 | #include "llvm/Analysis/CodeMetrics.h" | ||||||
84 | #include "llvm/Analysis/DemandedBits.h" | ||||||
85 | #include "llvm/Analysis/GlobalsModRef.h" | ||||||
86 | #include "llvm/Analysis/LoopAccessAnalysis.h" | ||||||
87 | #include "llvm/Analysis/LoopAnalysisManager.h" | ||||||
88 | #include "llvm/Analysis/LoopInfo.h" | ||||||
89 | #include "llvm/Analysis/LoopIterator.h" | ||||||
90 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | ||||||
91 | #include "llvm/Analysis/ProfileSummaryInfo.h" | ||||||
92 | #include "llvm/Analysis/ScalarEvolution.h" | ||||||
93 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" | ||||||
94 | #include "llvm/Analysis/TargetLibraryInfo.h" | ||||||
95 | #include "llvm/Analysis/TargetTransformInfo.h" | ||||||
96 | #include "llvm/Analysis/VectorUtils.h" | ||||||
97 | #include "llvm/IR/Attributes.h" | ||||||
98 | #include "llvm/IR/BasicBlock.h" | ||||||
99 | #include "llvm/IR/CFG.h" | ||||||
100 | #include "llvm/IR/Constant.h" | ||||||
101 | #include "llvm/IR/Constants.h" | ||||||
102 | #include "llvm/IR/DataLayout.h" | ||||||
103 | #include "llvm/IR/DebugInfoMetadata.h" | ||||||
104 | #include "llvm/IR/DebugLoc.h" | ||||||
105 | #include "llvm/IR/DerivedTypes.h" | ||||||
106 | #include "llvm/IR/DiagnosticInfo.h" | ||||||
107 | #include "llvm/IR/Dominators.h" | ||||||
108 | #include "llvm/IR/Function.h" | ||||||
109 | #include "llvm/IR/IRBuilder.h" | ||||||
110 | #include "llvm/IR/InstrTypes.h" | ||||||
111 | #include "llvm/IR/Instruction.h" | ||||||
112 | #include "llvm/IR/Instructions.h" | ||||||
113 | #include "llvm/IR/IntrinsicInst.h" | ||||||
114 | #include "llvm/IR/Intrinsics.h" | ||||||
115 | #include "llvm/IR/LLVMContext.h" | ||||||
116 | #include "llvm/IR/Metadata.h" | ||||||
117 | #include "llvm/IR/Module.h" | ||||||
118 | #include "llvm/IR/Operator.h" | ||||||
119 | #include "llvm/IR/PatternMatch.h" | ||||||
120 | #include "llvm/IR/Type.h" | ||||||
121 | #include "llvm/IR/Use.h" | ||||||
122 | #include "llvm/IR/User.h" | ||||||
123 | #include "llvm/IR/Value.h" | ||||||
124 | #include "llvm/IR/ValueHandle.h" | ||||||
125 | #include "llvm/IR/Verifier.h" | ||||||
126 | #include "llvm/InitializePasses.h" | ||||||
127 | #include "llvm/Pass.h" | ||||||
128 | #include "llvm/Support/Casting.h" | ||||||
129 | #include "llvm/Support/CommandLine.h" | ||||||
130 | #include "llvm/Support/Compiler.h" | ||||||
131 | #include "llvm/Support/Debug.h" | ||||||
132 | #include "llvm/Support/ErrorHandling.h" | ||||||
133 | #include "llvm/Support/InstructionCost.h" | ||||||
134 | #include "llvm/Support/MathExtras.h" | ||||||
135 | #include "llvm/Support/raw_ostream.h" | ||||||
136 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | ||||||
137 | #include "llvm/Transforms/Utils/InjectTLIMappings.h" | ||||||
138 | #include "llvm/Transforms/Utils/LoopSimplify.h" | ||||||
139 | #include "llvm/Transforms/Utils/LoopUtils.h" | ||||||
140 | #include "llvm/Transforms/Utils/LoopVersioning.h" | ||||||
141 | #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" | ||||||
142 | #include "llvm/Transforms/Utils/SizeOpts.h" | ||||||
143 | #include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h" | ||||||
144 | #include <algorithm> | ||||||
145 | #include <cassert> | ||||||
146 | #include <cstdint> | ||||||
147 | #include <cstdlib> | ||||||
148 | #include <functional> | ||||||
149 | #include <iterator> | ||||||
150 | #include <limits> | ||||||
151 | #include <memory> | ||||||
152 | #include <string> | ||||||
153 | #include <tuple> | ||||||
154 | #include <utility> | ||||||
155 | |||||||
156 | using namespace llvm; | ||||||
157 | |||||||
158 | #define LV_NAME"loop-vectorize" "loop-vectorize" | ||||||
159 | #define DEBUG_TYPE"loop-vectorize" LV_NAME"loop-vectorize" | ||||||
160 | |||||||
161 | #ifndef NDEBUG | ||||||
162 | const char VerboseDebug[] = DEBUG_TYPE"loop-vectorize" "-verbose"; | ||||||
163 | #endif | ||||||
164 | |||||||
165 | /// @{ | ||||||
166 | /// Metadata attribute names | ||||||
167 | const char LLVMLoopVectorizeFollowupAll[] = "llvm.loop.vectorize.followup_all"; | ||||||
168 | const char LLVMLoopVectorizeFollowupVectorized[] = | ||||||
169 | "llvm.loop.vectorize.followup_vectorized"; | ||||||
170 | const char LLVMLoopVectorizeFollowupEpilogue[] = | ||||||
171 | "llvm.loop.vectorize.followup_epilogue"; | ||||||
172 | /// @} | ||||||
173 | |||||||
174 | STATISTIC(LoopsVectorized, "Number of loops vectorized")static llvm::Statistic LoopsVectorized = {"loop-vectorize", "LoopsVectorized" , "Number of loops vectorized"}; | ||||||
175 | STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization")static llvm::Statistic LoopsAnalyzed = {"loop-vectorize", "LoopsAnalyzed" , "Number of loops analyzed for vectorization"}; | ||||||
176 | STATISTIC(LoopsEpilogueVectorized, "Number of epilogues vectorized")static llvm::Statistic LoopsEpilogueVectorized = {"loop-vectorize" , "LoopsEpilogueVectorized", "Number of epilogues vectorized" }; | ||||||
177 | |||||||
178 | static cl::opt<bool> EnableEpilogueVectorization( | ||||||
179 | "enable-epilogue-vectorization", cl::init(true), cl::Hidden, | ||||||
180 | cl::desc("Enable vectorization of epilogue loops.")); | ||||||
181 | |||||||
182 | static cl::opt<unsigned> EpilogueVectorizationForceVF( | ||||||
183 | "epilogue-vectorization-force-VF", cl::init(1), cl::Hidden, | ||||||
184 | cl::desc("When epilogue vectorization is enabled, and a value greater than " | ||||||
185 | "1 is specified, forces the given VF for all applicable epilogue " | ||||||
186 | "loops.")); | ||||||
187 | |||||||
188 | static cl::opt<unsigned> EpilogueVectorizationMinVF( | ||||||
189 | "epilogue-vectorization-minimum-VF", cl::init(16), cl::Hidden, | ||||||
190 | cl::desc("Only loops with vectorization factor equal to or larger than " | ||||||
191 | "the specified value are considered for epilogue vectorization.")); | ||||||
192 | |||||||
193 | /// Loops with a known constant trip count below this number are vectorized only | ||||||
194 | /// if no scalar iteration overheads are incurred. | ||||||
195 | static cl::opt<unsigned> TinyTripCountVectorThreshold( | ||||||
196 | "vectorizer-min-trip-count", cl::init(16), cl::Hidden, | ||||||
197 | cl::desc("Loops with a constant trip count that is smaller than this " | ||||||
198 | "value are vectorized only if no scalar iteration overheads " | ||||||
199 | "are incurred.")); | ||||||
200 | |||||||
201 | static cl::opt<unsigned> PragmaVectorizeMemoryCheckThreshold( | ||||||
202 | "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden, | ||||||
203 | cl::desc("The maximum allowed number of runtime memory checks with a " | ||||||
204 | "vectorize(enable) pragma.")); | ||||||
205 | |||||||
206 | // Option prefer-predicate-over-epilogue indicates that an epilogue is undesired, | ||||||
207 | // that predication is preferred, and this lists all options. I.e., the | ||||||
208 | // vectorizer will try to fold the tail-loop (epilogue) into the vector body | ||||||
209 | // and predicate the instructions accordingly. If tail-folding fails, there are | ||||||
210 | // different fallback strategies depending on these values: | ||||||
211 | namespace PreferPredicateTy { | ||||||
212 | enum Option { | ||||||
213 | ScalarEpilogue = 0, | ||||||
214 | PredicateElseScalarEpilogue, | ||||||
215 | PredicateOrDontVectorize | ||||||
216 | }; | ||||||
217 | } // namespace PreferPredicateTy | ||||||
218 | |||||||
219 | static cl::opt<PreferPredicateTy::Option> PreferPredicateOverEpilogue( | ||||||
220 | "prefer-predicate-over-epilogue", | ||||||
221 | cl::init(PreferPredicateTy::ScalarEpilogue), | ||||||
222 | cl::Hidden, | ||||||
223 | cl::desc("Tail-folding and predication preferences over creating a scalar " | ||||||
224 | "epilogue loop."), | ||||||
225 | cl::values(clEnumValN(PreferPredicateTy::ScalarEpilogue,llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" } | ||||||
226 | "scalar-epilogue",llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" } | ||||||
227 | "Don't tail-predicate loops, create scalar epilogue")llvm::cl::OptionEnumValue { "scalar-epilogue", int(PreferPredicateTy ::ScalarEpilogue), "Don't tail-predicate loops, create scalar epilogue" }, | ||||||
228 | clEnumValN(PreferPredicateTy::PredicateElseScalarEpilogue,llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." } | ||||||
229 | "predicate-else-scalar-epilogue",llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." } | ||||||
230 | "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." } | ||||||
231 | "folding fails.")llvm::cl::OptionEnumValue { "predicate-else-scalar-epilogue", int(PreferPredicateTy::PredicateElseScalarEpilogue), "prefer tail-folding, create scalar epilogue if tail " "folding fails." }, | ||||||
232 | clEnumValN(PreferPredicateTy::PredicateOrDontVectorize,llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." } | ||||||
233 | "predicate-dont-vectorize",llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." } | ||||||
234 | "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." } | ||||||
235 | "tail-folding fails.")llvm::cl::OptionEnumValue { "predicate-dont-vectorize", int(PreferPredicateTy ::PredicateOrDontVectorize), "prefers tail-folding, don't attempt vectorization if " "tail-folding fails." })); | ||||||
236 | |||||||
237 | static cl::opt<bool> MaximizeBandwidth( | ||||||
238 | "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden, | ||||||
239 | cl::desc("Maximize bandwidth when selecting vectorization factor which " | ||||||
240 | "will be determined by the smallest type in loop.")); | ||||||
241 | |||||||
242 | static cl::opt<bool> EnableInterleavedMemAccesses( | ||||||
243 | "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden, | ||||||
244 | cl::desc("Enable vectorization on interleaved memory accesses in a loop")); | ||||||
245 | |||||||
246 | /// An interleave-group may need masking if it resides in a block that needs | ||||||
247 | /// predication, or in order to mask away gaps. | ||||||
248 | static cl::opt<bool> EnableMaskedInterleavedMemAccesses( | ||||||
249 | "enable-masked-interleaved-mem-accesses", cl::init(false), cl::Hidden, | ||||||
250 | cl::desc("Enable vectorization on masked interleaved memory accesses in a loop")); | ||||||
251 | |||||||
252 | static cl::opt<unsigned> TinyTripCountInterleaveThreshold( | ||||||
253 | "tiny-trip-count-interleave-threshold", cl::init(128), cl::Hidden, | ||||||
254 | cl::desc("We don't interleave loops with a estimated constant trip count " | ||||||
255 | "below this number")); | ||||||
256 | |||||||
257 | static cl::opt<unsigned> ForceTargetNumScalarRegs( | ||||||
258 | "force-target-num-scalar-regs", cl::init(0), cl::Hidden, | ||||||
259 | cl::desc("A flag that overrides the target's number of scalar registers.")); | ||||||
260 | |||||||
261 | static cl::opt<unsigned> ForceTargetNumVectorRegs( | ||||||
262 | "force-target-num-vector-regs", cl::init(0), cl::Hidden, | ||||||
263 | cl::desc("A flag that overrides the target's number of vector registers.")); | ||||||
264 | |||||||
265 | static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor( | ||||||
266 | "force-target-max-scalar-interleave", cl::init(0), cl::Hidden, | ||||||
267 | cl::desc("A flag that overrides the target's max interleave factor for " | ||||||
268 | "scalar loops.")); | ||||||
269 | |||||||
270 | static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor( | ||||||
271 | "force-target-max-vector-interleave", cl::init(0), cl::Hidden, | ||||||
272 | cl::desc("A flag that overrides the target's max interleave factor for " | ||||||
273 | "vectorized loops.")); | ||||||
274 | |||||||
275 | static cl::opt<unsigned> ForceTargetInstructionCost( | ||||||
276 | "force-target-instruction-cost", cl::init(0), cl::Hidden, | ||||||
277 | cl::desc("A flag that overrides the target's expected cost for " | ||||||
278 | "an instruction to a single constant value. Mostly " | ||||||
279 | "useful for getting consistent testing.")); | ||||||
280 | |||||||
281 | static cl::opt<bool> ForceTargetSupportsScalableVectors( | ||||||
282 | "force-target-supports-scalable-vectors", cl::init(false), cl::Hidden, | ||||||
283 | cl::desc( | ||||||
284 | "Pretend that scalable vectors are supported, even if the target does " | ||||||
285 | "not support them. This flag should only be used for testing.")); | ||||||
286 | |||||||
287 | static cl::opt<unsigned> SmallLoopCost( | ||||||
288 | "small-loop-cost", cl::init(20), cl::Hidden, | ||||||
289 | cl::desc( | ||||||
290 | "The cost of a loop that is considered 'small' by the interleaver.")); | ||||||
291 | |||||||
292 | static cl::opt<bool> LoopVectorizeWithBlockFrequency( | ||||||
293 | "loop-vectorize-with-block-frequency", cl::init(true), cl::Hidden, | ||||||
294 | cl::desc("Enable the use of the block frequency analysis to access PGO " | ||||||
295 | "heuristics minimizing code growth in cold regions and being more " | ||||||
296 | "aggressive in hot regions.")); | ||||||
297 | |||||||
298 | // Runtime interleave loops for load/store throughput. | ||||||
299 | static cl::opt<bool> EnableLoadStoreRuntimeInterleave( | ||||||
300 | "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden, | ||||||
301 | cl::desc( | ||||||
302 | "Enable runtime interleaving until load/store ports are saturated")); | ||||||
303 | |||||||
304 | /// Interleave small loops with scalar reductions. | ||||||
305 | static cl::opt<bool> InterleaveSmallLoopScalarReduction( | ||||||
306 | "interleave-small-loop-scalar-reduction", cl::init(false), cl::Hidden, | ||||||
307 | cl::desc("Enable interleaving for loops with small iteration counts that " | ||||||
308 | "contain scalar reductions to expose ILP.")); | ||||||
309 | |||||||
310 | /// The number of stores in a loop that are allowed to need predication. | ||||||
311 | static cl::opt<unsigned> NumberOfStoresToPredicate( | ||||||
312 | "vectorize-num-stores-pred", cl::init(1), cl::Hidden, | ||||||
313 | cl::desc("Max number of stores to be predicated behind an if.")); | ||||||
314 | |||||||
315 | static cl::opt<bool> EnableIndVarRegisterHeur( | ||||||
316 | "enable-ind-var-reg-heur", cl::init(true), cl::Hidden, | ||||||
317 | cl::desc("Count the induction variable only once when interleaving")); | ||||||
318 | |||||||
319 | static cl::opt<bool> EnableCondStoresVectorization( | ||||||
320 | "enable-cond-stores-vec", cl::init(true), cl::Hidden, | ||||||
321 | cl::desc("Enable if predication of stores during vectorization.")); | ||||||
322 | |||||||
323 | static cl::opt<unsigned> MaxNestedScalarReductionIC( | ||||||
324 | "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden, | ||||||
325 | cl::desc("The maximum interleave count to use when interleaving a scalar " | ||||||
326 | "reduction in a nested loop.")); | ||||||
327 | |||||||
328 | static cl::opt<bool> | ||||||
329 | PreferInLoopReductions("prefer-inloop-reductions", cl::init(false), | ||||||
330 | cl::Hidden, | ||||||
331 | cl::desc("Prefer in-loop vector reductions, " | ||||||
332 | "overriding the targets preference.")); | ||||||
333 | |||||||
334 | static cl::opt<bool> ForceOrderedReductions( | ||||||
335 | "force-ordered-reductions", cl::init(false), cl::Hidden, | ||||||
336 | cl::desc("Enable the vectorisation of loops with in-order (strict) " | ||||||
337 | "FP reductions")); | ||||||
338 | |||||||
339 | static cl::opt<bool> PreferPredicatedReductionSelect( | ||||||
340 | "prefer-predicated-reduction-select", cl::init(false), cl::Hidden, | ||||||
341 | cl::desc( | ||||||
342 | "Prefer predicating a reduction operation over an after loop select.")); | ||||||
343 | |||||||
344 | cl::opt<bool> EnableVPlanNativePath( | ||||||
345 | "enable-vplan-native-path", cl::init(false), cl::Hidden, | ||||||
346 | cl::desc("Enable VPlan-native vectorization path with " | ||||||
347 | "support for outer loop vectorization.")); | ||||||
348 | |||||||
349 | // FIXME: Remove this switch once we have divergence analysis. Currently we | ||||||
350 | // assume divergent non-backedge branches when this switch is true. | ||||||
351 | cl::opt<bool> EnableVPlanPredication( | ||||||
352 | "enable-vplan-predication", cl::init(false), cl::Hidden, | ||||||
353 | cl::desc("Enable VPlan-native vectorization path predicator with " | ||||||
354 | "support for outer loop vectorization.")); | ||||||
355 | |||||||
356 | // This flag enables the stress testing of the VPlan H-CFG construction in the | ||||||
357 | // VPlan-native vectorization path. It must be used in conjuction with | ||||||
358 | // -enable-vplan-native-path. -vplan-verify-hcfg can also be used to enable the | ||||||
359 | // verification of the H-CFGs built. | ||||||
360 | static cl::opt<bool> VPlanBuildStressTest( | ||||||
361 | "vplan-build-stress-test", cl::init(false), cl::Hidden, | ||||||
362 | cl::desc( | ||||||
363 | "Build VPlan for every supported loop nest in the function and bail " | ||||||
364 | "out right after the build (stress test the VPlan H-CFG construction " | ||||||
365 | "in the VPlan-native vectorization path).")); | ||||||
366 | |||||||
367 | cl::opt<bool> llvm::EnableLoopInterleaving( | ||||||
368 | "interleave-loops", cl::init(true), cl::Hidden, | ||||||
369 | cl::desc("Enable loop interleaving in Loop vectorization passes")); | ||||||
370 | cl::opt<bool> llvm::EnableLoopVectorization( | ||||||
371 | "vectorize-loops", cl::init(true), cl::Hidden, | ||||||
372 | cl::desc("Run the Loop vectorization passes")); | ||||||
373 | |||||||
374 | cl::opt<bool> PrintVPlansInDotFormat( | ||||||
375 | "vplan-print-in-dot-format", cl::init(false), cl::Hidden, | ||||||
376 | cl::desc("Use dot format instead of plain text when dumping VPlans")); | ||||||
377 | |||||||
378 | /// A helper function that returns true if the given type is irregular. The | ||||||
379 | /// type is irregular if its allocated size doesn't equal the store size of an | ||||||
380 | /// element of the corresponding vector type. | ||||||
381 | static bool hasIrregularType(Type *Ty, const DataLayout &DL) { | ||||||
382 | // Determine if an array of N elements of type Ty is "bitcast compatible" | ||||||
383 | // with a <N x Ty> vector. | ||||||
384 | // This is only true if there is no padding between the array elements. | ||||||
385 | return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty); | ||||||
386 | } | ||||||
387 | |||||||
388 | /// A helper function that returns the reciprocal of the block probability of | ||||||
389 | /// predicated blocks. If we return X, we are assuming the predicated block | ||||||
390 | /// will execute once for every X iterations of the loop header. | ||||||
391 | /// | ||||||
392 | /// TODO: We should use actual block probability here, if available. Currently, | ||||||
393 | /// we always assume predicated blocks have a 50% chance of executing. | ||||||
394 | static unsigned getReciprocalPredBlockProb() { return 2; } | ||||||
395 | |||||||
396 | /// A helper function that returns an integer or floating-point constant with | ||||||
397 | /// value C. | ||||||
398 | static Constant *getSignedIntOrFpConstant(Type *Ty, int64_t C) { | ||||||
399 | return Ty->isIntegerTy() ? ConstantInt::getSigned(Ty, C) | ||||||
400 | : ConstantFP::get(Ty, C); | ||||||
401 | } | ||||||
402 | |||||||
403 | /// Returns "best known" trip count for the specified loop \p L as defined by | ||||||
404 | /// the following procedure: | ||||||
405 | /// 1) Returns exact trip count if it is known. | ||||||
406 | /// 2) Returns expected trip count according to profile data if any. | ||||||
407 | /// 3) Returns upper bound estimate if it is known. | ||||||
408 | /// 4) Returns None if all of the above failed. | ||||||
409 | static Optional<unsigned> getSmallBestKnownTC(ScalarEvolution &SE, Loop *L) { | ||||||
410 | // Check if exact trip count is known. | ||||||
411 | if (unsigned ExpectedTC = SE.getSmallConstantTripCount(L)) | ||||||
412 | return ExpectedTC; | ||||||
413 | |||||||
414 | // Check if there is an expected trip count available from profile data. | ||||||
415 | if (LoopVectorizeWithBlockFrequency) | ||||||
416 | if (auto EstimatedTC = getLoopEstimatedTripCount(L)) | ||||||
417 | return EstimatedTC; | ||||||
418 | |||||||
419 | // Check if upper bound estimate is known. | ||||||
420 | if (unsigned ExpectedTC = SE.getSmallConstantMaxTripCount(L)) | ||||||
421 | return ExpectedTC; | ||||||
422 | |||||||
423 | return None; | ||||||
424 | } | ||||||
425 | |||||||
426 | // Forward declare GeneratedRTChecks. | ||||||
427 | class GeneratedRTChecks; | ||||||
428 | |||||||
429 | namespace llvm { | ||||||
430 | |||||||
431 | AnalysisKey ShouldRunExtraVectorPasses::Key; | ||||||
432 | |||||||
433 | /// InnerLoopVectorizer vectorizes loops which contain only one basic | ||||||
434 | /// block to a specified vectorization factor (VF). | ||||||
435 | /// This class performs the widening of scalars into vectors, or multiple | ||||||
436 | /// scalars. This class also implements the following features: | ||||||
437 | /// * It inserts an epilogue loop for handling loops that don't have iteration | ||||||
438 | /// counts that are known to be a multiple of the vectorization factor. | ||||||
439 | /// * It handles the code generation for reduction variables. | ||||||
440 | /// * Scalarization (implementation using scalars) of un-vectorizable | ||||||
441 | /// instructions. | ||||||
442 | /// InnerLoopVectorizer does not perform any vectorization-legality | ||||||
443 | /// checks, and relies on the caller to check for the different legality | ||||||
444 | /// aspects. The InnerLoopVectorizer relies on the | ||||||
445 | /// LoopVectorizationLegality class to provide information about the induction | ||||||
446 | /// and reduction variables that were found to a given vectorization factor. | ||||||
447 | class InnerLoopVectorizer { | ||||||
448 | public: | ||||||
449 | InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE, | ||||||
450 | LoopInfo *LI, DominatorTree *DT, | ||||||
451 | const TargetLibraryInfo *TLI, | ||||||
452 | const TargetTransformInfo *TTI, AssumptionCache *AC, | ||||||
453 | OptimizationRemarkEmitter *ORE, ElementCount VecWidth, | ||||||
454 | unsigned UnrollFactor, LoopVectorizationLegality *LVL, | ||||||
455 | LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, | ||||||
456 | ProfileSummaryInfo *PSI, GeneratedRTChecks &RTChecks) | ||||||
457 | : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI), | ||||||
458 | AC(AC), ORE(ORE), VF(VecWidth), UF(UnrollFactor), | ||||||
459 | Builder(PSE.getSE()->getContext()), Legal(LVL), Cost(CM), BFI(BFI), | ||||||
460 | PSI(PSI), RTChecks(RTChecks) { | ||||||
461 | // Query this against the original loop and save it here because the profile | ||||||
462 | // of the original loop header may change as the transformation happens. | ||||||
463 | OptForSizeBasedOnProfile = llvm::shouldOptimizeForSize( | ||||||
464 | OrigLoop->getHeader(), PSI, BFI, PGSOQueryType::IRPass); | ||||||
465 | } | ||||||
466 | |||||||
467 | virtual ~InnerLoopVectorizer() = default; | ||||||
468 | |||||||
469 | /// Create a new empty loop that will contain vectorized instructions later | ||||||
470 | /// on, while the old loop will be used as the scalar remainder. Control flow | ||||||
471 | /// is generated around the vectorized (and scalar epilogue) loops consisting | ||||||
472 | /// of various checks and bypasses. Return the pre-header block of the new | ||||||
473 | /// loop and the start value for the canonical induction, if it is != 0. The | ||||||
474 | /// latter is the case when vectorizing the epilogue loop. In the case of | ||||||
475 | /// epilogue vectorization, this function is overriden to handle the more | ||||||
476 | /// complex control flow around the loops. | ||||||
477 | virtual std::pair<BasicBlock *, Value *> createVectorizedLoopSkeleton(); | ||||||
478 | |||||||
479 | /// Widen a single call instruction within the innermost loop. | ||||||
480 | void widenCallInstruction(CallInst &I, VPValue *Def, VPUser &ArgOperands, | ||||||
481 | VPTransformState &State); | ||||||
482 | |||||||
483 | /// Fix the vectorized code, taking care of header phi's, live-outs, and more. | ||||||
484 | void fixVectorizedLoop(VPTransformState &State); | ||||||
485 | |||||||
486 | // Return true if any runtime check is added. | ||||||
487 | bool areSafetyChecksAdded() { return AddedSafetyChecks; } | ||||||
488 | |||||||
489 | /// A type for vectorized values in the new loop. Each value from the | ||||||
490 | /// original loop, when vectorized, is represented by UF vector values in the | ||||||
491 | /// new unrolled loop, where UF is the unroll factor. | ||||||
492 | using VectorParts = SmallVector<Value *, 2>; | ||||||
493 | |||||||
494 | /// Vectorize a single first-order recurrence or pointer induction PHINode in | ||||||
495 | /// a block. This method handles the induction variable canonicalization. It | ||||||
496 | /// supports both VF = 1 for unrolled loops and arbitrary length vectors. | ||||||
497 | void widenPHIInstruction(Instruction *PN, VPWidenPHIRecipe *PhiR, | ||||||
498 | VPTransformState &State); | ||||||
499 | |||||||
500 | /// A helper function to scalarize a single Instruction in the innermost loop. | ||||||
501 | /// Generates a sequence of scalar instances for each lane between \p MinLane | ||||||
502 | /// and \p MaxLane, times each part between \p MinPart and \p MaxPart, | ||||||
503 | /// inclusive. Uses the VPValue operands from \p RepRecipe instead of \p | ||||||
504 | /// Instr's operands. | ||||||
505 | void scalarizeInstruction(Instruction *Instr, VPReplicateRecipe *RepRecipe, | ||||||
506 | const VPIteration &Instance, bool IfPredicateInstr, | ||||||
507 | VPTransformState &State); | ||||||
508 | |||||||
509 | /// Widen an integer or floating-point induction variable \p IV. If \p Trunc | ||||||
510 | /// is provided, the integer induction variable will first be truncated to | ||||||
511 | /// the corresponding type. \p CanonicalIV is the scalar value generated for | ||||||
512 | /// the canonical induction variable. | ||||||
513 | void widenIntOrFpInduction(PHINode *IV, const InductionDescriptor &ID, | ||||||
514 | Value *Start, TruncInst *Trunc, VPValue *Def, | ||||||
515 | VPTransformState &State, Value *CanonicalIV); | ||||||
516 | |||||||
517 | /// Construct the vector value of a scalarized value \p V one lane at a time. | ||||||
518 | void packScalarIntoVectorValue(VPValue *Def, const VPIteration &Instance, | ||||||
519 | VPTransformState &State); | ||||||
520 | |||||||
521 | /// Try to vectorize interleaved access group \p Group with the base address | ||||||
522 | /// given in \p Addr, optionally masking the vector operations if \p | ||||||
523 | /// BlockInMask is non-null. Use \p State to translate given VPValues to IR | ||||||
524 | /// values in the vectorized loop. | ||||||
525 | void vectorizeInterleaveGroup(const InterleaveGroup<Instruction> *Group, | ||||||
526 | ArrayRef<VPValue *> VPDefs, | ||||||
527 | VPTransformState &State, VPValue *Addr, | ||||||
528 | ArrayRef<VPValue *> StoredValues, | ||||||
529 | VPValue *BlockInMask = nullptr); | ||||||
530 | |||||||
531 | /// Set the debug location in the builder \p Ptr using the debug location in | ||||||
532 | /// \p V. If \p Ptr is None then it uses the class member's Builder. | ||||||
533 | void setDebugLocFromInst(const Value *V, | ||||||
534 | Optional<IRBuilder<> *> CustomBuilder = None); | ||||||
535 | |||||||
536 | /// Fix the non-induction PHIs in the OrigPHIsToFix vector. | ||||||
537 | void fixNonInductionPHIs(VPTransformState &State); | ||||||
538 | |||||||
539 | /// Returns true if the reordering of FP operations is not allowed, but we are | ||||||
540 | /// able to vectorize with strict in-order reductions for the given RdxDesc. | ||||||
541 | bool useOrderedReductions(const RecurrenceDescriptor &RdxDesc); | ||||||
542 | |||||||
543 | /// Create a broadcast instruction. This method generates a broadcast | ||||||
544 | /// instruction (shuffle) for loop invariant values and for the induction | ||||||
545 | /// value. If this is the induction variable then we extend it to N, N+1, ... | ||||||
546 | /// this is needed because each iteration in the loop corresponds to a SIMD | ||||||
547 | /// element. | ||||||
548 | virtual Value *getBroadcastInstrs(Value *V); | ||||||
549 | |||||||
550 | /// Add metadata from one instruction to another. | ||||||
551 | /// | ||||||
552 | /// This includes both the original MDs from \p From and additional ones (\see | ||||||
553 | /// addNewMetadata). Use this for *newly created* instructions in the vector | ||||||
554 | /// loop. | ||||||
555 | void addMetadata(Instruction *To, Instruction *From); | ||||||
556 | |||||||
557 | /// Similar to the previous function but it adds the metadata to a | ||||||
558 | /// vector of instructions. | ||||||
559 | void addMetadata(ArrayRef<Value *> To, Instruction *From); | ||||||
560 | |||||||
561 | protected: | ||||||
562 | friend class LoopVectorizationPlanner; | ||||||
563 | |||||||
564 | /// A small list of PHINodes. | ||||||
565 | using PhiVector = SmallVector<PHINode *, 4>; | ||||||
566 | |||||||
567 | /// A type for scalarized values in the new loop. Each value from the | ||||||
568 | /// original loop, when scalarized, is represented by UF x VF scalar values | ||||||
569 | /// in the new unrolled loop, where UF is the unroll factor and VF is the | ||||||
570 | /// vectorization factor. | ||||||
571 | using ScalarParts = SmallVector<SmallVector<Value *, 4>, 2>; | ||||||
572 | |||||||
573 | /// Set up the values of the IVs correctly when exiting the vector loop. | ||||||
574 | void fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II, | ||||||
575 | Value *CountRoundDown, Value *EndValue, | ||||||
576 | BasicBlock *MiddleBlock); | ||||||
577 | |||||||
578 | /// Introduce a conditional branch (on true, condition to be set later) at the | ||||||
579 | /// end of the header=latch connecting it to itself (across the backedge) and | ||||||
580 | /// to the exit block of \p L. | ||||||
581 | void createHeaderBranch(Loop *L); | ||||||
582 | |||||||
583 | /// Handle all cross-iteration phis in the header. | ||||||
584 | void fixCrossIterationPHIs(VPTransformState &State); | ||||||
585 | |||||||
586 | /// Create the exit value of first order recurrences in the middle block and | ||||||
587 | /// update their users. | ||||||
588 | void fixFirstOrderRecurrence(VPFirstOrderRecurrencePHIRecipe *PhiR, | ||||||
589 | VPTransformState &State); | ||||||
590 | |||||||
591 | /// Create code for the loop exit value of the reduction. | ||||||
592 | void fixReduction(VPReductionPHIRecipe *Phi, VPTransformState &State); | ||||||
593 | |||||||
594 | /// Clear NSW/NUW flags from reduction instructions if necessary. | ||||||
595 | void clearReductionWrapFlags(const RecurrenceDescriptor &RdxDesc, | ||||||
596 | VPTransformState &State); | ||||||
597 | |||||||
598 | /// Fixup the LCSSA phi nodes in the unique exit block. This simply | ||||||
599 | /// means we need to add the appropriate incoming value from the middle | ||||||
600 | /// block as exiting edges from the scalar epilogue loop (if present) are | ||||||
601 | /// already in place, and we exit the vector loop exclusively to the middle | ||||||
602 | /// block. | ||||||
603 | void fixLCSSAPHIs(VPTransformState &State); | ||||||
604 | |||||||
605 | /// Iteratively sink the scalarized operands of a predicated instruction into | ||||||
606 | /// the block that was created for it. | ||||||
607 | void sinkScalarOperands(Instruction *PredInst); | ||||||
608 | |||||||
609 | /// Shrinks vector element sizes to the smallest bitwidth they can be legally | ||||||
610 | /// represented as. | ||||||
611 | void truncateToMinimalBitwidths(VPTransformState &State); | ||||||
612 | |||||||
613 | /// Compute scalar induction steps. \p ScalarIV is the scalar induction | ||||||
614 | /// variable on which to base the steps, \p Step is the size of the step, and | ||||||
615 | /// \p EntryVal is the value from the original loop that maps to the steps. | ||||||
616 | /// Note that \p EntryVal doesn't have to be an induction variable - it | ||||||
617 | /// can also be a truncate instruction. | ||||||
618 | void buildScalarSteps(Value *ScalarIV, Value *Step, Instruction *EntryVal, | ||||||
619 | const InductionDescriptor &ID, VPValue *Def, | ||||||
620 | VPTransformState &State); | ||||||
621 | |||||||
622 | /// Create a vector induction phi node based on an existing scalar one. \p | ||||||
623 | /// EntryVal is the value from the original loop that maps to the vector phi | ||||||
624 | /// node, and \p Step is the loop-invariant step. If \p EntryVal is a | ||||||
625 | /// truncate instruction, instead of widening the original IV, we widen a | ||||||
626 | /// version of the IV truncated to \p EntryVal's type. | ||||||
627 | void createVectorIntOrFpInductionPHI(const InductionDescriptor &II, | ||||||
628 | Value *Step, Value *Start, | ||||||
629 | Instruction *EntryVal, VPValue *Def, | ||||||
630 | VPTransformState &State); | ||||||
631 | |||||||
632 | /// Returns true if an instruction \p I should be scalarized instead of | ||||||
633 | /// vectorized for the chosen vectorization factor. | ||||||
634 | bool shouldScalarizeInstruction(Instruction *I) const; | ||||||
635 | |||||||
636 | /// Returns true if we should generate a scalar version of \p IV. | ||||||
637 | bool needsScalarInduction(Instruction *IV) const; | ||||||
638 | |||||||
639 | /// Returns (and creates if needed) the original loop trip count. | ||||||
640 | Value *getOrCreateTripCount(Loop *NewLoop); | ||||||
641 | |||||||
642 | /// Returns (and creates if needed) the trip count of the widened loop. | ||||||
643 | Value *getOrCreateVectorTripCount(Loop *NewLoop); | ||||||
644 | |||||||
645 | /// Returns a bitcasted value to the requested vector type. | ||||||
646 | /// Also handles bitcasts of vector<float> <-> vector<pointer> types. | ||||||
647 | Value *createBitOrPointerCast(Value *V, VectorType *DstVTy, | ||||||
648 | const DataLayout &DL); | ||||||
649 | |||||||
650 | /// Emit a bypass check to see if the vector trip count is zero, including if | ||||||
651 | /// it overflows. | ||||||
652 | void emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass); | ||||||
653 | |||||||
654 | /// Emit a bypass check to see if all of the SCEV assumptions we've | ||||||
655 | /// had to make are correct. Returns the block containing the checks or | ||||||
656 | /// nullptr if no checks have been added. | ||||||
657 | BasicBlock *emitSCEVChecks(Loop *L, BasicBlock *Bypass); | ||||||
658 | |||||||
659 | /// Emit bypass checks to check any memory assumptions we may have made. | ||||||
660 | /// Returns the block containing the checks or nullptr if no checks have been | ||||||
661 | /// added. | ||||||
662 | BasicBlock *emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass); | ||||||
663 | |||||||
664 | /// Compute the transformed value of Index at offset StartValue using step | ||||||
665 | /// StepValue. | ||||||
666 | /// For integer induction, returns StartValue + Index * StepValue. | ||||||
667 | /// For pointer induction, returns StartValue[Index * StepValue]. | ||||||
668 | /// FIXME: The newly created binary instructions should contain nsw/nuw | ||||||
669 | /// flags, which can be found from the original scalar operations. | ||||||
670 | Value *emitTransformedIndex(IRBuilder<> &B, Value *Index, ScalarEvolution *SE, | ||||||
671 | const DataLayout &DL, | ||||||
672 | const InductionDescriptor &ID, | ||||||
673 | BasicBlock *VectorHeader) const; | ||||||
674 | |||||||
675 | /// Emit basic blocks (prefixed with \p Prefix) for the iteration check, | ||||||
676 | /// vector loop preheader, middle block and scalar preheader. Also | ||||||
677 | /// allocate a loop object for the new vector loop and return it. | ||||||
678 | Loop *createVectorLoopSkeleton(StringRef Prefix); | ||||||
679 | |||||||
680 | /// Create new phi nodes for the induction variables to resume iteration count | ||||||
681 | /// in the scalar epilogue, from where the vectorized loop left off. | ||||||
682 | /// In cases where the loop skeleton is more complicated (eg. epilogue | ||||||
683 | /// vectorization) and the resume values can come from an additional bypass | ||||||
684 | /// block, the \p AdditionalBypass pair provides information about the bypass | ||||||
685 | /// block and the end value on the edge from bypass to this loop. | ||||||
686 | void createInductionResumeValues( | ||||||
687 | Loop *L, | ||||||
688 | std::pair<BasicBlock *, Value *> AdditionalBypass = {nullptr, nullptr}); | ||||||
689 | |||||||
690 | /// Complete the loop skeleton by adding debug MDs, creating appropriate | ||||||
691 | /// conditional branches in the middle block, preparing the builder and | ||||||
692 | /// running the verifier. Take in the vector loop \p L as argument, and return | ||||||
693 | /// the preheader of the completed vector loop. | ||||||
694 | BasicBlock *completeLoopSkeleton(Loop *L, MDNode *OrigLoopID); | ||||||
695 | |||||||
696 | /// Add additional metadata to \p To that was not present on \p Orig. | ||||||
697 | /// | ||||||
698 | /// Currently this is used to add the noalias annotations based on the | ||||||
699 | /// inserted memchecks. Use this for instructions that are *cloned* into the | ||||||
700 | /// vector loop. | ||||||
701 | void addNewMetadata(Instruction *To, const Instruction *Orig); | ||||||
702 | |||||||
703 | /// Collect poison-generating recipes that may generate a poison value that is | ||||||
704 | /// used after vectorization, even when their operands are not poison. Those | ||||||
705 | /// recipes meet the following conditions: | ||||||
706 | /// * Contribute to the address computation of a recipe generating a widen | ||||||
707 | /// memory load/store (VPWidenMemoryInstructionRecipe or | ||||||
708 | /// VPInterleaveRecipe). | ||||||
709 | /// * Such a widen memory load/store has at least one underlying Instruction | ||||||
710 | /// that is in a basic block that needs predication and after vectorization | ||||||
711 | /// the generated instruction won't be predicated. | ||||||
712 | void collectPoisonGeneratingRecipes(VPTransformState &State); | ||||||
713 | |||||||
714 | /// Allow subclasses to override and print debug traces before/after vplan | ||||||
715 | /// execution, when trace information is requested. | ||||||
716 | virtual void printDebugTracesAtStart(){}; | ||||||
717 | virtual void printDebugTracesAtEnd(){}; | ||||||
718 | |||||||
719 | /// The original loop. | ||||||
720 | Loop *OrigLoop; | ||||||
721 | |||||||
722 | /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies | ||||||
723 | /// dynamic knowledge to simplify SCEV expressions and converts them to a | ||||||
724 | /// more usable form. | ||||||
725 | PredicatedScalarEvolution &PSE; | ||||||
726 | |||||||
727 | /// Loop Info. | ||||||
728 | LoopInfo *LI; | ||||||
729 | |||||||
730 | /// Dominator Tree. | ||||||
731 | DominatorTree *DT; | ||||||
732 | |||||||
733 | /// Alias Analysis. | ||||||
734 | AAResults *AA; | ||||||
735 | |||||||
736 | /// Target Library Info. | ||||||
737 | const TargetLibraryInfo *TLI; | ||||||
738 | |||||||
739 | /// Target Transform Info. | ||||||
740 | const TargetTransformInfo *TTI; | ||||||
741 | |||||||
742 | /// Assumption Cache. | ||||||
743 | AssumptionCache *AC; | ||||||
744 | |||||||
745 | /// Interface to emit optimization remarks. | ||||||
746 | OptimizationRemarkEmitter *ORE; | ||||||
747 | |||||||
748 | /// LoopVersioning. It's only set up (non-null) if memchecks were | ||||||
749 | /// used. | ||||||
750 | /// | ||||||
751 | /// This is currently only used to add no-alias metadata based on the | ||||||
752 | /// memchecks. The actually versioning is performed manually. | ||||||
753 | std::unique_ptr<LoopVersioning> LVer; | ||||||
754 | |||||||
755 | /// The vectorization SIMD factor to use. Each vector will have this many | ||||||
756 | /// vector elements. | ||||||
757 | ElementCount VF; | ||||||
758 | |||||||
759 | /// The vectorization unroll factor to use. Each scalar is vectorized to this | ||||||
760 | /// many different vector instructions. | ||||||
761 | unsigned UF; | ||||||
762 | |||||||
763 | /// The builder that we use | ||||||
764 | IRBuilder<> Builder; | ||||||
765 | |||||||
766 | // --- Vectorization state --- | ||||||
767 | |||||||
768 | /// The vector-loop preheader. | ||||||
769 | BasicBlock *LoopVectorPreHeader; | ||||||
770 | |||||||
771 | /// The scalar-loop preheader. | ||||||
772 | BasicBlock *LoopScalarPreHeader; | ||||||
773 | |||||||
774 | /// Middle Block between the vector and the scalar. | ||||||
775 | BasicBlock *LoopMiddleBlock; | ||||||
776 | |||||||
777 | /// The unique ExitBlock of the scalar loop if one exists. Note that | ||||||
778 | /// there can be multiple exiting edges reaching this block. | ||||||
779 | BasicBlock *LoopExitBlock; | ||||||
780 | |||||||
781 | /// The vector loop body. | ||||||
782 | BasicBlock *LoopVectorBody; | ||||||
783 | |||||||
784 | /// The scalar loop body. | ||||||
785 | BasicBlock *LoopScalarBody; | ||||||
786 | |||||||
787 | /// A list of all bypass blocks. The first block is the entry of the loop. | ||||||
788 | SmallVector<BasicBlock *, 4> LoopBypassBlocks; | ||||||
789 | |||||||
790 | /// Store instructions that were predicated. | ||||||
791 | SmallVector<Instruction *, 4> PredicatedInstructions; | ||||||
792 | |||||||
793 | /// Trip count of the original loop. | ||||||
794 | Value *TripCount = nullptr; | ||||||
795 | |||||||
796 | /// Trip count of the widened loop (TripCount - TripCount % (VF*UF)) | ||||||
797 | Value *VectorTripCount = nullptr; | ||||||
798 | |||||||
799 | /// The legality analysis. | ||||||
800 | LoopVectorizationLegality *Legal; | ||||||
801 | |||||||
802 | /// The profitablity analysis. | ||||||
803 | LoopVectorizationCostModel *Cost; | ||||||
804 | |||||||
805 | // Record whether runtime checks are added. | ||||||
806 | bool AddedSafetyChecks = false; | ||||||
807 | |||||||
808 | // Holds the end values for each induction variable. We save the end values | ||||||
809 | // so we can later fix-up the external users of the induction variables. | ||||||
810 | DenseMap<PHINode *, Value *> IVEndValues; | ||||||
811 | |||||||
812 | // Vector of original scalar PHIs whose corresponding widened PHIs need to be | ||||||
813 | // fixed up at the end of vector code generation. | ||||||
814 | SmallVector<PHINode *, 8> OrigPHIsToFix; | ||||||
815 | |||||||
816 | /// BFI and PSI are used to check for profile guided size optimizations. | ||||||
817 | BlockFrequencyInfo *BFI; | ||||||
818 | ProfileSummaryInfo *PSI; | ||||||
819 | |||||||
820 | // Whether this loop should be optimized for size based on profile guided size | ||||||
821 | // optimizatios. | ||||||
822 | bool OptForSizeBasedOnProfile; | ||||||
823 | |||||||
824 | /// Structure to hold information about generated runtime checks, responsible | ||||||
825 | /// for cleaning the checks, if vectorization turns out unprofitable. | ||||||
826 | GeneratedRTChecks &RTChecks; | ||||||
827 | }; | ||||||
828 | |||||||
829 | class InnerLoopUnroller : public InnerLoopVectorizer { | ||||||
830 | public: | ||||||
831 | InnerLoopUnroller(Loop *OrigLoop, PredicatedScalarEvolution &PSE, | ||||||
832 | LoopInfo *LI, DominatorTree *DT, | ||||||
833 | const TargetLibraryInfo *TLI, | ||||||
834 | const TargetTransformInfo *TTI, AssumptionCache *AC, | ||||||
835 | OptimizationRemarkEmitter *ORE, unsigned UnrollFactor, | ||||||
836 | LoopVectorizationLegality *LVL, | ||||||
837 | LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, | ||||||
838 | ProfileSummaryInfo *PSI, GeneratedRTChecks &Check) | ||||||
839 | : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | ||||||
840 | ElementCount::getFixed(1), UnrollFactor, LVL, CM, | ||||||
841 | BFI, PSI, Check) {} | ||||||
842 | |||||||
843 | private: | ||||||
844 | Value *getBroadcastInstrs(Value *V) override; | ||||||
845 | }; | ||||||
846 | |||||||
847 | /// Encapsulate information regarding vectorization of a loop and its epilogue. | ||||||
848 | /// This information is meant to be updated and used across two stages of | ||||||
849 | /// epilogue vectorization. | ||||||
850 | struct EpilogueLoopVectorizationInfo { | ||||||
851 | ElementCount MainLoopVF = ElementCount::getFixed(0); | ||||||
852 | unsigned MainLoopUF = 0; | ||||||
853 | ElementCount EpilogueVF = ElementCount::getFixed(0); | ||||||
854 | unsigned EpilogueUF = 0; | ||||||
855 | BasicBlock *MainLoopIterationCountCheck = nullptr; | ||||||
856 | BasicBlock *EpilogueIterationCountCheck = nullptr; | ||||||
857 | BasicBlock *SCEVSafetyCheck = nullptr; | ||||||
858 | BasicBlock *MemSafetyCheck = nullptr; | ||||||
859 | Value *TripCount = nullptr; | ||||||
860 | Value *VectorTripCount = nullptr; | ||||||
861 | |||||||
862 | EpilogueLoopVectorizationInfo(ElementCount MVF, unsigned MUF, | ||||||
863 | ElementCount EVF, unsigned EUF) | ||||||
864 | : MainLoopVF(MVF), MainLoopUF(MUF), EpilogueVF(EVF), EpilogueUF(EUF) { | ||||||
865 | 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", 866, __extension__ __PRETTY_FUNCTION__)) | ||||||
866 | "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", 866, __extension__ __PRETTY_FUNCTION__)); | ||||||
867 | } | ||||||
868 | }; | ||||||
869 | |||||||
870 | /// An extension of the inner loop vectorizer that creates a skeleton for a | ||||||
871 | /// vectorized loop that has its epilogue (residual) also vectorized. | ||||||
872 | /// The idea is to run the vplan on a given loop twice, firstly to setup the | ||||||
873 | /// skeleton and vectorize the main loop, and secondly to complete the skeleton | ||||||
874 | /// from the first step and vectorize the epilogue. This is achieved by | ||||||
875 | /// deriving two concrete strategy classes from this base class and invoking | ||||||
876 | /// them in succession from the loop vectorizer planner. | ||||||
877 | class InnerLoopAndEpilogueVectorizer : public InnerLoopVectorizer { | ||||||
878 | public: | ||||||
879 | InnerLoopAndEpilogueVectorizer( | ||||||
880 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, | ||||||
881 | DominatorTree *DT, const TargetLibraryInfo *TLI, | ||||||
882 | const TargetTransformInfo *TTI, AssumptionCache *AC, | ||||||
883 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, | ||||||
884 | LoopVectorizationLegality *LVL, llvm::LoopVectorizationCostModel *CM, | ||||||
885 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, | ||||||
886 | GeneratedRTChecks &Checks) | ||||||
887 | : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | ||||||
888 | EPI.MainLoopVF, EPI.MainLoopUF, LVL, CM, BFI, PSI, | ||||||
889 | Checks), | ||||||
890 | EPI(EPI) {} | ||||||
891 | |||||||
892 | // Override this function to handle the more complex control flow around the | ||||||
893 | // three loops. | ||||||
894 | std::pair<BasicBlock *, Value *> | ||||||
895 | createVectorizedLoopSkeleton() final override { | ||||||
896 | return createEpilogueVectorizedLoopSkeleton(); | ||||||
897 | } | ||||||
898 | |||||||
899 | /// The interface for creating a vectorized skeleton using one of two | ||||||
900 | /// different strategies, each corresponding to one execution of the vplan | ||||||
901 | /// as described above. | ||||||
902 | virtual std::pair<BasicBlock *, Value *> | ||||||
903 | createEpilogueVectorizedLoopSkeleton() = 0; | ||||||
904 | |||||||
905 | /// Holds and updates state information required to vectorize the main loop | ||||||
906 | /// and its epilogue in two separate passes. This setup helps us avoid | ||||||
907 | /// regenerating and recomputing runtime safety checks. It also helps us to | ||||||
908 | /// shorten the iteration-count-check path length for the cases where the | ||||||
909 | /// iteration count of the loop is so small that the main vector loop is | ||||||
910 | /// completely skipped. | ||||||
911 | EpilogueLoopVectorizationInfo &EPI; | ||||||
912 | }; | ||||||
913 | |||||||
914 | /// A specialized derived class of inner loop vectorizer that performs | ||||||
915 | /// vectorization of *main* loops in the process of vectorizing loops and their | ||||||
916 | /// epilogues. | ||||||
917 | class EpilogueVectorizerMainLoop : public InnerLoopAndEpilogueVectorizer { | ||||||
918 | public: | ||||||
919 | EpilogueVectorizerMainLoop( | ||||||
920 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, | ||||||
921 | DominatorTree *DT, const TargetLibraryInfo *TLI, | ||||||
922 | const TargetTransformInfo *TTI, AssumptionCache *AC, | ||||||
923 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, | ||||||
924 | LoopVectorizationLegality *LVL, llvm::LoopVectorizationCostModel *CM, | ||||||
925 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, | ||||||
926 | GeneratedRTChecks &Check) | ||||||
927 | : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | ||||||
928 | EPI, LVL, CM, BFI, PSI, Check) {} | ||||||
929 | /// Implements the interface for creating a vectorized skeleton using the | ||||||
930 | /// *main loop* strategy (ie the first pass of vplan execution). | ||||||
931 | std::pair<BasicBlock *, Value *> | ||||||
932 | createEpilogueVectorizedLoopSkeleton() final override; | ||||||
933 | |||||||
934 | protected: | ||||||
935 | /// Emits an iteration count bypass check once for the main loop (when \p | ||||||
936 | /// ForEpilogue is false) and once for the epilogue loop (when \p | ||||||
937 | /// ForEpilogue is true). | ||||||
938 | BasicBlock *emitMinimumIterationCountCheck(Loop *L, BasicBlock *Bypass, | ||||||
939 | bool ForEpilogue); | ||||||
940 | void printDebugTracesAtStart() override; | ||||||
941 | void printDebugTracesAtEnd() override; | ||||||
942 | }; | ||||||
943 | |||||||
944 | // A specialized derived class of inner loop vectorizer that performs | ||||||
945 | // vectorization of *epilogue* loops in the process of vectorizing loops and | ||||||
946 | // their epilogues. | ||||||
947 | class EpilogueVectorizerEpilogueLoop : public InnerLoopAndEpilogueVectorizer { | ||||||
948 | public: | ||||||
949 | EpilogueVectorizerEpilogueLoop( | ||||||
950 | Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, | ||||||
951 | DominatorTree *DT, const TargetLibraryInfo *TLI, | ||||||
952 | const TargetTransformInfo *TTI, AssumptionCache *AC, | ||||||
953 | OptimizationRemarkEmitter *ORE, EpilogueLoopVectorizationInfo &EPI, | ||||||
954 | LoopVectorizationLegality *LVL, llvm::LoopVectorizationCostModel *CM, | ||||||
955 | BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, | ||||||
956 | GeneratedRTChecks &Checks) | ||||||
957 | : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, | ||||||
958 | EPI, LVL, CM, BFI, PSI, Checks) {} | ||||||
959 | /// Implements the interface for creating a vectorized skeleton using the | ||||||
960 | /// *epilogue loop* strategy (ie the second pass of vplan execution). | ||||||
961 | std::pair<BasicBlock *, Value *> | ||||||
962 | createEpilogueVectorizedLoopSkeleton() final override; | ||||||
963 | |||||||
964 | protected: | ||||||
965 | /// Emits an iteration count bypass check after the main vector loop has | ||||||
966 | /// finished to see if there are any iterations left to execute by either | ||||||
967 | /// the vector epilogue or the scalar epilogue. | ||||||
968 | BasicBlock *emitMinimumVectorEpilogueIterCountCheck(Loop *L, | ||||||
969 | BasicBlock *Bypass, | ||||||
970 | BasicBlock *Insert); | ||||||
971 | void printDebugTracesAtStart() override; | ||||||
972 | void printDebugTracesAtEnd() override; | ||||||
973 | }; | ||||||
974 | } // end namespace llvm | ||||||
975 | |||||||
976 | /// Look for a meaningful debug location on the instruction or it's | ||||||
977 | /// operands. | ||||||
978 | static Instruction *getDebugLocFromInstOrOperands(Instruction *I) { | ||||||
979 | if (!I) | ||||||
980 | return I; | ||||||
981 | |||||||
982 | DebugLoc Empty; | ||||||
983 | if (I->getDebugLoc() != Empty) | ||||||
984 | return I; | ||||||
985 | |||||||
986 | for (Use &Op : I->operands()) { | ||||||
987 | if (Instruction *OpInst = dyn_cast<Instruction>(Op)) | ||||||
988 | if (OpInst->getDebugLoc() != Empty) | ||||||
989 | return OpInst; | ||||||
990 | } | ||||||
991 | |||||||
992 | return I; | ||||||
993 | } | ||||||
994 | |||||||
995 | void InnerLoopVectorizer::setDebugLocFromInst( | ||||||
996 | const Value *V, Optional<IRBuilder<> *> CustomBuilder) { | ||||||
997 | IRBuilder<> *B = (CustomBuilder == None) ? &Builder : *CustomBuilder; | ||||||
998 | if (const Instruction *Inst = dyn_cast_or_null<Instruction>(V)) { | ||||||
999 | const DILocation *DIL = Inst->getDebugLoc(); | ||||||
1000 | |||||||
1001 | // When a FSDiscriminator is enabled, we don't need to add the multiply | ||||||
1002 | // factors to the discriminators. | ||||||
1003 | if (DIL && Inst->getFunction()->isDebugInfoForProfiling() && | ||||||
1004 | !isa<DbgInfoIntrinsic>(Inst) && !EnableFSDiscriminator) { | ||||||
1005 | // FIXME: For scalable vectors, assume vscale=1. | ||||||
1006 | auto NewDIL = | ||||||
1007 | DIL->cloneByMultiplyingDuplicationFactor(UF * VF.getKnownMinValue()); | ||||||
1008 | if (NewDIL) | ||||||
1009 | B->SetCurrentDebugLocation(NewDIL.getValue()); | ||||||
1010 | else | ||||||
1011 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Failed to create new discriminator: " << DIL->getFilename() << " Line: " << DIL ->getLine(); } } while (false) | ||||||
1012 | << "Failed to create new discriminator: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Failed to create new discriminator: " << DIL->getFilename() << " Line: " << DIL ->getLine(); } } while (false) | ||||||
1013 | << DIL->getFilename() << " Line: " << DIL->getLine())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Failed to create new discriminator: " << DIL->getFilename() << " Line: " << DIL ->getLine(); } } while (false); | ||||||
1014 | } else | ||||||
1015 | B->SetCurrentDebugLocation(DIL); | ||||||
1016 | } else | ||||||
1017 | B->SetCurrentDebugLocation(DebugLoc()); | ||||||
1018 | } | ||||||
1019 | |||||||
1020 | /// Write a \p DebugMsg about vectorization to the debug output stream. If \p I | ||||||
1021 | /// is passed, the message relates to that particular instruction. | ||||||
1022 | #ifndef NDEBUG | ||||||
1023 | static void debugVectorizationMessage(const StringRef Prefix, | ||||||
1024 | const StringRef DebugMsg, | ||||||
1025 | Instruction *I) { | ||||||
1026 | dbgs() << "LV: " << Prefix << DebugMsg; | ||||||
1027 | if (I != nullptr) | ||||||
1028 | dbgs() << " " << *I; | ||||||
1029 | else | ||||||
1030 | dbgs() << '.'; | ||||||
1031 | dbgs() << '\n'; | ||||||
1032 | } | ||||||
1033 | #endif | ||||||
1034 | |||||||
1035 | /// Create an analysis remark that explains why vectorization failed | ||||||
1036 | /// | ||||||
1037 | /// \p PassName is the name of the pass (e.g. can be AlwaysPrint). \p | ||||||
1038 | /// RemarkName is the identifier for the remark. If \p I is passed it is an | ||||||
1039 | /// instruction that prevents vectorization. Otherwise \p TheLoop is used for | ||||||
1040 | /// the location of the remark. \return the remark object that can be | ||||||
1041 | /// streamed to. | ||||||
1042 | static OptimizationRemarkAnalysis createLVAnalysis(const char *PassName, | ||||||
1043 | StringRef RemarkName, Loop *TheLoop, Instruction *I) { | ||||||
1044 | Value *CodeRegion = TheLoop->getHeader(); | ||||||
1045 | DebugLoc DL = TheLoop->getStartLoc(); | ||||||
1046 | |||||||
1047 | if (I) { | ||||||
1048 | CodeRegion = I->getParent(); | ||||||
1049 | // If there is no debug location attached to the instruction, revert back to | ||||||
1050 | // using the loop's. | ||||||
1051 | if (I->getDebugLoc()) | ||||||
1052 | DL = I->getDebugLoc(); | ||||||
1053 | } | ||||||
1054 | |||||||
1055 | return OptimizationRemarkAnalysis(PassName, RemarkName, DL, CodeRegion); | ||||||
1056 | } | ||||||
1057 | |||||||
1058 | namespace llvm { | ||||||
1059 | |||||||
1060 | /// Return a value for Step multiplied by VF. | ||||||
1061 | Value *createStepForVF(IRBuilder<> &B, Type *Ty, ElementCount VF, | ||||||
1062 | int64_t Step) { | ||||||
1063 | 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", 1063, __extension__ __PRETTY_FUNCTION__)); | ||||||
1064 | Constant *StepVal = ConstantInt::get(Ty, Step * VF.getKnownMinValue()); | ||||||
1065 | return VF.isScalable() ? B.CreateVScale(StepVal) : StepVal; | ||||||
1066 | } | ||||||
1067 | |||||||
1068 | /// Return the runtime value for VF. | ||||||
1069 | Value *getRuntimeVF(IRBuilder<> &B, Type *Ty, ElementCount VF) { | ||||||
1070 | Constant *EC = ConstantInt::get(Ty, VF.getKnownMinValue()); | ||||||
1071 | return VF.isScalable() ? B.CreateVScale(EC) : EC; | ||||||
1072 | } | ||||||
1073 | |||||||
1074 | static Value *getRuntimeVFAsFloat(IRBuilder<> &B, Type *FTy, ElementCount VF) { | ||||||
1075 | 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", 1075, __extension__ __PRETTY_FUNCTION__)); | ||||||
1076 | Type *IntTy = IntegerType::get(FTy->getContext(), FTy->getScalarSizeInBits()); | ||||||
1077 | Value *RuntimeVF = getRuntimeVF(B, IntTy, VF); | ||||||
1078 | return B.CreateUIToFP(RuntimeVF, FTy); | ||||||
1079 | } | ||||||
1080 | |||||||
1081 | void reportVectorizationFailure(const StringRef DebugMsg, | ||||||
1082 | const StringRef OREMsg, const StringRef ORETag, | ||||||
1083 | OptimizationRemarkEmitter *ORE, Loop *TheLoop, | ||||||
1084 | Instruction *I) { | ||||||
1085 | LLVM_DEBUG(debugVectorizationMessage("Not vectorizing: ", DebugMsg, I))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { debugVectorizationMessage("Not vectorizing: " , DebugMsg, I); } } while (false); | ||||||
1086 | LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE); | ||||||
1087 | ORE->emit( | ||||||
1088 | createLVAnalysis(Hints.vectorizeAnalysisPassName(), ORETag, TheLoop, I) | ||||||
1089 | << "loop not vectorized: " << OREMsg); | ||||||
1090 | } | ||||||
1091 | |||||||
1092 | void reportVectorizationInfo(const StringRef Msg, const StringRef ORETag, | ||||||
1093 | OptimizationRemarkEmitter *ORE, Loop *TheLoop, | ||||||
1094 | Instruction *I) { | ||||||
1095 | LLVM_DEBUG(debugVectorizationMessage("", Msg, I))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { debugVectorizationMessage("", Msg, I); } } while (false); | ||||||
1096 | LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE); | ||||||
1097 | ORE->emit( | ||||||
1098 | createLVAnalysis(Hints.vectorizeAnalysisPassName(), ORETag, TheLoop, I) | ||||||
1099 | << Msg); | ||||||
1100 | } | ||||||
1101 | |||||||
1102 | } // end namespace llvm | ||||||
1103 | |||||||
1104 | #ifndef NDEBUG | ||||||
1105 | /// \return string containing a file name and a line # for the given loop. | ||||||
1106 | static std::string getDebugLocString(const Loop *L) { | ||||||
1107 | std::string Result; | ||||||
1108 | if (L) { | ||||||
1109 | raw_string_ostream OS(Result); | ||||||
1110 | if (const DebugLoc LoopDbgLoc = L->getStartLoc()) | ||||||
1111 | LoopDbgLoc.print(OS); | ||||||
1112 | else | ||||||
1113 | // Just print the module name. | ||||||
1114 | OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier(); | ||||||
1115 | OS.flush(); | ||||||
1116 | } | ||||||
1117 | return Result; | ||||||
1118 | } | ||||||
1119 | #endif | ||||||
1120 | |||||||
1121 | void InnerLoopVectorizer::addNewMetadata(Instruction *To, | ||||||
1122 | const Instruction *Orig) { | ||||||
1123 | // If the loop was versioned with memchecks, add the corresponding no-alias | ||||||
1124 | // metadata. | ||||||
1125 | if (LVer && (isa<LoadInst>(Orig) || isa<StoreInst>(Orig))) | ||||||
1126 | LVer->annotateInstWithNoAlias(To, Orig); | ||||||
1127 | } | ||||||
1128 | |||||||
1129 | void InnerLoopVectorizer::collectPoisonGeneratingRecipes( | ||||||
1130 | VPTransformState &State) { | ||||||
1131 | |||||||
1132 | // Collect recipes in the backward slice of `Root` that may generate a poison | ||||||
1133 | // value that is used after vectorization. | ||||||
1134 | SmallPtrSet<VPRecipeBase *, 16> Visited; | ||||||
1135 | auto collectPoisonGeneratingInstrsInBackwardSlice([&](VPRecipeBase *Root) { | ||||||
1136 | SmallVector<VPRecipeBase *, 16> Worklist; | ||||||
1137 | Worklist.push_back(Root); | ||||||
1138 | |||||||
1139 | // Traverse the backward slice of Root through its use-def chain. | ||||||
1140 | while (!Worklist.empty()) { | ||||||
1141 | VPRecipeBase *CurRec = Worklist.back(); | ||||||
1142 | Worklist.pop_back(); | ||||||
1143 | |||||||
1144 | if (!Visited.insert(CurRec).second) | ||||||
1145 | continue; | ||||||
1146 | |||||||
1147 | // Prune search if we find another recipe generating a widen memory | ||||||
1148 | // instruction. Widen memory instructions involved in address computation | ||||||
1149 | // will lead to gather/scatter instructions, which don't need to be | ||||||
1150 | // handled. | ||||||
1151 | if (isa<VPWidenMemoryInstructionRecipe>(CurRec) || | ||||||
1152 | isa<VPInterleaveRecipe>(CurRec) || | ||||||
1153 | isa<VPCanonicalIVPHIRecipe>(CurRec)) | ||||||
1154 | continue; | ||||||
1155 | |||||||
1156 | // This recipe contributes to the address computation of a widen | ||||||
1157 | // load/store. Collect recipe if its underlying instruction has | ||||||
1158 | // poison-generating flags. | ||||||
1159 | Instruction *Instr = CurRec->getUnderlyingInstr(); | ||||||
1160 | if (Instr && Instr->hasPoisonGeneratingFlags()) | ||||||
1161 | State.MayGeneratePoisonRecipes.insert(CurRec); | ||||||
1162 | |||||||
1163 | // Add new definitions to the worklist. | ||||||
1164 | for (VPValue *operand : CurRec->operands()) | ||||||
1165 | if (VPDef *OpDef = operand->getDef()) | ||||||
1166 | Worklist.push_back(cast<VPRecipeBase>(OpDef)); | ||||||
1167 | } | ||||||
1168 | }); | ||||||
1169 | |||||||
1170 | // Traverse all the recipes in the VPlan and collect the poison-generating | ||||||
1171 | // recipes in the backward slice starting at the address of a VPWidenRecipe or | ||||||
1172 | // VPInterleaveRecipe. | ||||||
1173 | auto Iter = depth_first( | ||||||
1174 | VPBlockRecursiveTraversalWrapper<VPBlockBase *>(State.Plan->getEntry())); | ||||||
1175 | for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Iter)) { | ||||||
1176 | for (VPRecipeBase &Recipe : *VPBB) { | ||||||
1177 | if (auto *WidenRec = dyn_cast<VPWidenMemoryInstructionRecipe>(&Recipe)) { | ||||||
1178 | Instruction *UnderlyingInstr = WidenRec->getUnderlyingInstr(); | ||||||
1179 | VPDef *AddrDef = WidenRec->getAddr()->getDef(); | ||||||
1180 | if (AddrDef && WidenRec->isConsecutive() && UnderlyingInstr && | ||||||
1181 | Legal->blockNeedsPredication(UnderlyingInstr->getParent())) | ||||||
1182 | collectPoisonGeneratingInstrsInBackwardSlice( | ||||||
1183 | cast<VPRecipeBase>(AddrDef)); | ||||||
1184 | } else if (auto *InterleaveRec = dyn_cast<VPInterleaveRecipe>(&Recipe)) { | ||||||
1185 | VPDef *AddrDef = InterleaveRec->getAddr()->getDef(); | ||||||
1186 | if (AddrDef) { | ||||||
1187 | // Check if any member of the interleave group needs predication. | ||||||
1188 | const InterleaveGroup<Instruction> *InterGroup = | ||||||
1189 | InterleaveRec->getInterleaveGroup(); | ||||||
1190 | bool NeedPredication = false; | ||||||
1191 | for (int I = 0, NumMembers = InterGroup->getNumMembers(); | ||||||
1192 | I < NumMembers; ++I) { | ||||||
1193 | Instruction *Member = InterGroup->getMember(I); | ||||||
1194 | if (Member) | ||||||
1195 | NeedPredication |= | ||||||
1196 | Legal->blockNeedsPredication(Member->getParent()); | ||||||
1197 | } | ||||||
1198 | |||||||
1199 | if (NeedPredication) | ||||||
1200 | collectPoisonGeneratingInstrsInBackwardSlice( | ||||||
1201 | cast<VPRecipeBase>(AddrDef)); | ||||||
1202 | } | ||||||
1203 | } | ||||||
1204 | } | ||||||
1205 | } | ||||||
1206 | } | ||||||
1207 | |||||||
1208 | void InnerLoopVectorizer::addMetadata(Instruction *To, | ||||||
1209 | Instruction *From) { | ||||||
1210 | propagateMetadata(To, From); | ||||||
1211 | addNewMetadata(To, From); | ||||||
1212 | } | ||||||
1213 | |||||||
1214 | void InnerLoopVectorizer::addMetadata(ArrayRef<Value *> To, | ||||||
1215 | Instruction *From) { | ||||||
1216 | for (Value *V : To) { | ||||||
1217 | if (Instruction *I = dyn_cast<Instruction>(V)) | ||||||
1218 | addMetadata(I, From); | ||||||
1219 | } | ||||||
1220 | } | ||||||
1221 | |||||||
1222 | namespace llvm { | ||||||
1223 | |||||||
1224 | // Loop vectorization cost-model hints how the scalar epilogue loop should be | ||||||
1225 | // lowered. | ||||||
1226 | enum ScalarEpilogueLowering { | ||||||
1227 | |||||||
1228 | // The default: allowing scalar epilogues. | ||||||
1229 | CM_ScalarEpilogueAllowed, | ||||||
1230 | |||||||
1231 | // Vectorization with OptForSize: don't allow epilogues. | ||||||
1232 | CM_ScalarEpilogueNotAllowedOptSize, | ||||||
1233 | |||||||
1234 | // A special case of vectorisation with OptForSize: loops with a very small | ||||||
1235 | // trip count are considered for vectorization under OptForSize, thereby | ||||||
1236 | // making sure the cost of their loop body is dominant, free of runtime | ||||||
1237 | // guards and scalar iteration overheads. | ||||||
1238 | CM_ScalarEpilogueNotAllowedLowTripLoop, | ||||||
1239 | |||||||
1240 | // Loop hint predicate indicating an epilogue is undesired. | ||||||
1241 | CM_ScalarEpilogueNotNeededUsePredicate, | ||||||
1242 | |||||||
1243 | // Directive indicating we must either tail fold or not vectorize | ||||||
1244 | CM_ScalarEpilogueNotAllowedUsePredicate | ||||||
1245 | }; | ||||||
1246 | |||||||
1247 | /// ElementCountComparator creates a total ordering for ElementCount | ||||||
1248 | /// for the purposes of using it in a set structure. | ||||||
1249 | struct ElementCountComparator { | ||||||
1250 | bool operator()(const ElementCount &LHS, const ElementCount &RHS) const { | ||||||
1251 | return std::make_tuple(LHS.isScalable(), LHS.getKnownMinValue()) < | ||||||
1252 | std::make_tuple(RHS.isScalable(), RHS.getKnownMinValue()); | ||||||
1253 | } | ||||||
1254 | }; | ||||||
1255 | using ElementCountSet = SmallSet<ElementCount, 16, ElementCountComparator>; | ||||||
1256 | |||||||
1257 | /// LoopVectorizationCostModel - estimates the expected speedups due to | ||||||
1258 | /// vectorization. | ||||||
1259 | /// In many cases vectorization is not profitable. This can happen because of | ||||||
1260 | /// a number of reasons. In this class we mainly attempt to predict the | ||||||
1261 | /// expected speedup/slowdowns due to the supported instruction set. We use the | ||||||
1262 | /// TargetTransformInfo to query the different backends for the cost of | ||||||
1263 | /// different operations. | ||||||
1264 | class LoopVectorizationCostModel { | ||||||
1265 | public: | ||||||
1266 | LoopVectorizationCostModel(ScalarEpilogueLowering SEL, Loop *L, | ||||||
1267 | PredicatedScalarEvolution &PSE, LoopInfo *LI, | ||||||
1268 | LoopVectorizationLegality *Legal, | ||||||
1269 | const TargetTransformInfo &TTI, | ||||||
1270 | const TargetLibraryInfo *TLI, DemandedBits *DB, | ||||||
1271 | AssumptionCache *AC, | ||||||
1272 | OptimizationRemarkEmitter *ORE, const Function *F, | ||||||
1273 | const LoopVectorizeHints *Hints, | ||||||
1274 | InterleavedAccessInfo &IAI) | ||||||
1275 | : ScalarEpilogueStatus(SEL), TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), | ||||||
1276 | TTI(TTI), TLI(TLI), DB(DB), AC(AC), ORE(ORE), TheFunction(F), | ||||||
1277 | Hints(Hints), InterleaveInfo(IAI) {} | ||||||
1278 | |||||||
1279 | /// \return An upper bound for the vectorization factors (both fixed and | ||||||
1280 | /// scalable). If the factors are 0, vectorization and interleaving should be | ||||||
1281 | /// avoided up front. | ||||||
1282 | FixedScalableVFPair computeMaxVF(ElementCount UserVF, unsigned UserIC); | ||||||
1283 | |||||||
1284 | /// \return True if runtime checks are required for vectorization, and false | ||||||
1285 | /// otherwise. | ||||||
1286 | bool runtimeChecksRequired(); | ||||||
1287 | |||||||
1288 | /// \return The most profitable vectorization factor and the cost of that VF. | ||||||
1289 | /// This method checks every VF in \p CandidateVFs. If UserVF is not ZERO | ||||||
1290 | /// then this vectorization factor will be selected if vectorization is | ||||||
1291 | /// possible. | ||||||
1292 | VectorizationFactor | ||||||
1293 | selectVectorizationFactor(const ElementCountSet &CandidateVFs); | ||||||
1294 | |||||||
1295 | VectorizationFactor | ||||||
1296 | selectEpilogueVectorizationFactor(const ElementCount MaxVF, | ||||||
1297 | const LoopVectorizationPlanner &LVP); | ||||||
1298 | |||||||
1299 | /// Setup cost-based decisions for user vectorization factor. | ||||||
1300 | /// \return true if the UserVF is a feasible VF to be chosen. | ||||||
1301 | bool selectUserVectorizationFactor(ElementCount UserVF) { | ||||||
1302 | collectUniformsAndScalars(UserVF); | ||||||
1303 | collectInstsToScalarize(UserVF); | ||||||
1304 | return expectedCost(UserVF).first.isValid(); | ||||||
1305 | } | ||||||
1306 | |||||||
1307 | /// \return The size (in bits) of the smallest and widest types in the code | ||||||
1308 | /// that needs to be vectorized. We ignore values that remain scalar such as | ||||||
1309 | /// 64 bit loop indices. | ||||||
1310 | std::pair<unsigned, unsigned> getSmallestAndWidestTypes(); | ||||||
1311 | |||||||
1312 | /// \return The desired interleave count. | ||||||
1313 | /// If interleave count has been specified by metadata it will be returned. | ||||||
1314 | /// Otherwise, the interleave count is computed and returned. VF and LoopCost | ||||||
1315 | /// are the selected vectorization factor and the cost of the selected VF. | ||||||
1316 | unsigned selectInterleaveCount(ElementCount VF, unsigned LoopCost); | ||||||
1317 | |||||||
1318 | /// Memory access instruction may be vectorized in more than one way. | ||||||
1319 | /// Form of instruction after vectorization depends on cost. | ||||||
1320 | /// This function takes cost-based decisions for Load/Store instructions | ||||||
1321 | /// and collects them in a map. This decisions map is used for building | ||||||
1322 | /// the lists of loop-uniform and loop-scalar instructions. | ||||||
1323 | /// The calculated cost is saved with widening decision in order to | ||||||
1324 | /// avoid redundant calculations. | ||||||
1325 | void setCostBasedWideningDecision(ElementCount VF); | ||||||
1326 | |||||||
1327 | /// A struct that represents some properties of the register usage | ||||||
1328 | /// of a loop. | ||||||
1329 | struct RegisterUsage { | ||||||
1330 | /// Holds the number of loop invariant values that are used in the loop. | ||||||
1331 | /// The key is ClassID of target-provided register class. | ||||||
1332 | SmallMapVector<unsigned, unsigned, 4> LoopInvariantRegs; | ||||||
1333 | /// Holds the maximum number of concurrent live intervals in the loop. | ||||||
1334 | /// The key is ClassID of target-provided register class. | ||||||
1335 | SmallMapVector<unsigned, unsigned, 4> MaxLocalUsers; | ||||||
1336 | }; | ||||||
1337 | |||||||
1338 | /// \return Returns information about the register usages of the loop for the | ||||||
1339 | /// given vectorization factors. | ||||||
1340 | SmallVector<RegisterUsage, 8> | ||||||
1341 | calculateRegisterUsage(ArrayRef<ElementCount> VFs); | ||||||
1342 | |||||||
1343 | /// Collect values we want to ignore in the cost model. | ||||||
1344 | void collectValuesToIgnore(); | ||||||
1345 | |||||||
1346 | /// Collect all element types in the loop for which widening is needed. | ||||||
1347 | void collectElementTypesForWidening(); | ||||||
1348 | |||||||
1349 | /// Split reductions into those that happen in the loop, and those that happen | ||||||
1350 | /// outside. In loop reductions are collected into InLoopReductionChains. | ||||||
1351 | void collectInLoopReductions(); | ||||||
1352 | |||||||
1353 | /// Returns true if we should use strict in-order reductions for the given | ||||||
1354 | /// RdxDesc. This is true if the -enable-strict-reductions flag is passed, | ||||||
1355 | /// the IsOrdered flag of RdxDesc is set and we do not allow reordering | ||||||
1356 | /// of FP operations. | ||||||
1357 | bool useOrderedReductions(const RecurrenceDescriptor &RdxDesc) { | ||||||
1358 | return !Hints->allowReordering() && RdxDesc.isOrdered(); | ||||||
1359 | } | ||||||
1360 | |||||||
1361 | /// \returns The smallest bitwidth each instruction can be represented with. | ||||||
1362 | /// The vector equivalents of these instructions should be truncated to this | ||||||
1363 | /// type. | ||||||
1364 | const MapVector<Instruction *, uint64_t> &getMinimalBitwidths() const { | ||||||
1365 | return MinBWs; | ||||||
1366 | } | ||||||
1367 | |||||||
1368 | /// \returns True if it is more profitable to scalarize instruction \p I for | ||||||
1369 | /// vectorization factor \p VF. | ||||||
1370 | bool isProfitableToScalarize(Instruction *I, ElementCount VF) const { | ||||||
1371 | 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", 1372, __extension__ __PRETTY_FUNCTION__)) | ||||||
1372 | "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", 1372, __extension__ __PRETTY_FUNCTION__)); | ||||||
1373 | |||||||
1374 | // Cost model is not run in the VPlan-native path - return conservative | ||||||
1375 | // result until this changes. | ||||||
1376 | if (EnableVPlanNativePath) | ||||||
1377 | return false; | ||||||
1378 | |||||||
1379 | auto Scalars = InstsToScalarize.find(VF); | ||||||
1380 | 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", 1381, __extension__ __PRETTY_FUNCTION__)) | ||||||
1381 | "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", 1381, __extension__ __PRETTY_FUNCTION__)); | ||||||
1382 | return Scalars->second.find(I) != Scalars->second.end(); | ||||||
1383 | } | ||||||
1384 | |||||||
1385 | /// Returns true if \p I is known to be uniform after vectorization. | ||||||
1386 | bool isUniformAfterVectorization(Instruction *I, ElementCount VF) const { | ||||||
1387 | if (VF.isScalar()) | ||||||
1388 | return true; | ||||||
1389 | |||||||
1390 | // Cost model is not run in the VPlan-native path - return conservative | ||||||
1391 | // result until this changes. | ||||||
1392 | if (EnableVPlanNativePath) | ||||||
1393 | return false; | ||||||
1394 | |||||||
1395 | auto UniformsPerVF = Uniforms.find(VF); | ||||||
1396 | 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", 1397, __extension__ __PRETTY_FUNCTION__)) | ||||||
1397 | "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", 1397, __extension__ __PRETTY_FUNCTION__)); | ||||||
1398 | return UniformsPerVF->second.count(I); | ||||||
1399 | } | ||||||
1400 | |||||||
1401 | /// Returns true if \p I is known to be scalar after vectorization. | ||||||
1402 | bool isScalarAfterVectorization(Instruction *I, ElementCount VF) const { | ||||||
1403 | if (VF.isScalar()) | ||||||
1404 | return true; | ||||||
1405 | |||||||
1406 | // Cost model is not run in the VPlan-native path - return conservative | ||||||
1407 | // result until this changes. | ||||||
1408 | if (EnableVPlanNativePath) | ||||||
1409 | return false; | ||||||
1410 | |||||||
1411 | auto ScalarsPerVF = Scalars.find(VF); | ||||||
1412 | 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", 1413, __extension__ __PRETTY_FUNCTION__)) | ||||||
1413 | "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", 1413, __extension__ __PRETTY_FUNCTION__)); | ||||||
1414 | return ScalarsPerVF->second.count(I); | ||||||
1415 | } | ||||||
1416 | |||||||
1417 | /// \returns True if instruction \p I can be truncated to a smaller bitwidth | ||||||
1418 | /// for vectorization factor \p VF. | ||||||
1419 | bool canTruncateToMinimalBitwidth(Instruction *I, ElementCount VF) const { | ||||||
1420 | return VF.isVector() && MinBWs.find(I) != MinBWs.end() && | ||||||
1421 | !isProfitableToScalarize(I, VF) && | ||||||
1422 | !isScalarAfterVectorization(I, VF); | ||||||
1423 | } | ||||||
1424 | |||||||
1425 | /// Decision that was taken during cost calculation for memory instruction. | ||||||
1426 | enum InstWidening { | ||||||
1427 | CM_Unknown, | ||||||
1428 | CM_Widen, // For consecutive accesses with stride +1. | ||||||
1429 | CM_Widen_Reverse, // For consecutive accesses with stride -1. | ||||||
1430 | CM_Interleave, | ||||||
1431 | CM_GatherScatter, | ||||||
1432 | CM_Scalarize | ||||||
1433 | }; | ||||||
1434 | |||||||
1435 | /// Save vectorization decision \p W and \p Cost taken by the cost model for | ||||||
1436 | /// instruction \p I and vector width \p VF. | ||||||
1437 | void setWideningDecision(Instruction *I, ElementCount VF, InstWidening W, | ||||||
1438 | InstructionCost Cost) { | ||||||
1439 | 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", 1439, __extension__ __PRETTY_FUNCTION__)); | ||||||
1440 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost); | ||||||
1441 | } | ||||||
1442 | |||||||
1443 | /// Save vectorization decision \p W and \p Cost taken by the cost model for | ||||||
1444 | /// interleaving group \p Grp and vector width \p VF. | ||||||
1445 | void setWideningDecision(const InterleaveGroup<Instruction> *Grp, | ||||||
1446 | ElementCount VF, InstWidening W, | ||||||
1447 | InstructionCost Cost) { | ||||||
1448 | 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", 1448, __extension__ __PRETTY_FUNCTION__)); | ||||||
1449 | /// Broadcast this decicion to all instructions inside the group. | ||||||
1450 | /// But the cost will be assigned to one instruction only. | ||||||
1451 | for (unsigned i = 0; i < Grp->getFactor(); ++i) { | ||||||
1452 | if (auto *I = Grp->getMember(i)) { | ||||||
1453 | if (Grp->getInsertPos() == I) | ||||||
1454 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost); | ||||||
1455 | else | ||||||
1456 | WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, 0); | ||||||
1457 | } | ||||||
1458 | } | ||||||
1459 | } | ||||||
1460 | |||||||
1461 | /// Return the cost model decision for the given instruction \p I and vector | ||||||
1462 | /// width \p VF. Return CM_Unknown if this instruction did not pass | ||||||
1463 | /// through the cost modeling. | ||||||
1464 | InstWidening getWideningDecision(Instruction *I, ElementCount VF) const { | ||||||
1465 | 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", 1465, __extension__ __PRETTY_FUNCTION__)); | ||||||
1466 | // Cost model is not run in the VPlan-native path - return conservative | ||||||
1467 | // result until this changes. | ||||||
1468 | if (EnableVPlanNativePath) | ||||||
1469 | return CM_GatherScatter; | ||||||
1470 | |||||||
1471 | std::pair<Instruction *, ElementCount> InstOnVF = std::make_pair(I, VF); | ||||||
1472 | auto Itr = WideningDecisions.find(InstOnVF); | ||||||
1473 | if (Itr == WideningDecisions.end()) | ||||||
1474 | return CM_Unknown; | ||||||
1475 | return Itr->second.first; | ||||||
1476 | } | ||||||
1477 | |||||||
1478 | /// Return the vectorization cost for the given instruction \p I and vector | ||||||
1479 | /// width \p VF. | ||||||
1480 | InstructionCost getWideningCost(Instruction *I, ElementCount VF) { | ||||||
1481 | 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", 1481, __extension__ __PRETTY_FUNCTION__)); | ||||||
1482 | std::pair<Instruction *, ElementCount> InstOnVF = std::make_pair(I, VF); | ||||||
1483 | 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", 1484, __extension__ __PRETTY_FUNCTION__)) | ||||||
1484 | "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", 1484, __extension__ __PRETTY_FUNCTION__)); | ||||||
1485 | return WideningDecisions[InstOnVF].second; | ||||||
1486 | } | ||||||
1487 | |||||||
1488 | /// Return True if instruction \p I is an optimizable truncate whose operand | ||||||
1489 | /// is an induction variable. Such a truncate will be removed by adding a new | ||||||
1490 | /// induction variable with the destination type. | ||||||
1491 | bool isOptimizableIVTruncate(Instruction *I, ElementCount VF) { | ||||||
1492 | // If the instruction is not a truncate, return false. | ||||||
1493 | auto *Trunc = dyn_cast<TruncInst>(I); | ||||||
1494 | if (!Trunc) | ||||||
1495 | return false; | ||||||
1496 | |||||||
1497 | // Get the source and destination types of the truncate. | ||||||
1498 | Type *SrcTy = ToVectorTy(cast<CastInst>(I)->getSrcTy(), VF); | ||||||
1499 | Type *DestTy = ToVectorTy(cast<CastInst>(I)->getDestTy(), VF); | ||||||
1500 | |||||||
1501 | // If the truncate is free for the given types, return false. Replacing a | ||||||
1502 | // free truncate with an induction variable would add an induction variable | ||||||
1503 | // update instruction to each iteration of the loop. We exclude from this | ||||||
1504 | // check the primary induction variable since it will need an update | ||||||
1505 | // instruction regardless. | ||||||
1506 | Value *Op = Trunc->getOperand(0); | ||||||
1507 | if (Op != Legal->getPrimaryInduction() && TTI.isTruncateFree(SrcTy, DestTy)) | ||||||
1508 | return false; | ||||||
1509 | |||||||
1510 | // If the truncated value is not an induction variable, return false. | ||||||
1511 | return Legal->isInductionPhi(Op); | ||||||
1512 | } | ||||||
1513 | |||||||
1514 | /// Collects the instructions to scalarize for each predicated instruction in | ||||||
1515 | /// the loop. | ||||||
1516 | void collectInstsToScalarize(ElementCount VF); | ||||||
1517 | |||||||
1518 | /// Collect Uniform and Scalar values for the given \p VF. | ||||||
1519 | /// The sets depend on CM decision for Load/Store instructions | ||||||
1520 | /// that may be vectorized as interleave, gather-scatter or scalarized. | ||||||
1521 | void collectUniformsAndScalars(ElementCount VF) { | ||||||
1522 | // Do the analysis once. | ||||||
1523 | if (VF.isScalar() || Uniforms.find(VF) != Uniforms.end()) | ||||||
1524 | return; | ||||||
1525 | setCostBasedWideningDecision(VF); | ||||||
1526 | collectLoopUniforms(VF); | ||||||
1527 | collectLoopScalars(VF); | ||||||
1528 | } | ||||||
1529 | |||||||
1530 | /// Returns true if the target machine supports masked store operation | ||||||
1531 | /// for the given \p DataType and kind of access to \p Ptr. | ||||||
1532 | bool isLegalMaskedStore(Type *DataType, Value *Ptr, Align Alignment) const { | ||||||
1533 | return Legal->isConsecutivePtr(DataType, Ptr) && | ||||||
1534 | TTI.isLegalMaskedStore(DataType, Alignment); | ||||||
1535 | } | ||||||
1536 | |||||||
1537 | /// Returns true if the target machine supports masked load operation | ||||||
1538 | /// for the given \p DataType and kind of access to \p Ptr. | ||||||
1539 | bool isLegalMaskedLoad(Type *DataType, Value *Ptr, Align Alignment) const { | ||||||
1540 | return Legal->isConsecutivePtr(DataType, Ptr) && | ||||||
1541 | TTI.isLegalMaskedLoad(DataType, Alignment); | ||||||
1542 | } | ||||||
1543 | |||||||
1544 | /// Returns true if the target machine can represent \p V as a masked gather | ||||||
1545 | /// or scatter operation. | ||||||
1546 | bool isLegalGatherOrScatter(Value *V, | ||||||
1547 | ElementCount VF = ElementCount::getFixed(1)) { | ||||||
1548 | bool LI = isa<LoadInst>(V); | ||||||
1549 | bool SI = isa<StoreInst>(V); | ||||||
1550 | if (!LI && !SI) | ||||||
1551 | return false; | ||||||
1552 | auto *Ty = getLoadStoreType(V); | ||||||
1553 | Align Align = getLoadStoreAlignment(V); | ||||||
1554 | if (VF.isVector()) | ||||||
1555 | Ty = VectorType::get(Ty, VF); | ||||||
1556 | return (LI && TTI.isLegalMaskedGather(Ty, Align)) || | ||||||
1557 | (SI && TTI.isLegalMaskedScatter(Ty, Align)); | ||||||
1558 | } | ||||||
1559 | |||||||
1560 | /// Returns true if the target machine supports all of the reduction | ||||||
1561 | /// variables found for the given VF. | ||||||
1562 | bool canVectorizeReductions(ElementCount VF) const { | ||||||
1563 | return (all_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool { | ||||||
1564 | const RecurrenceDescriptor &RdxDesc = Reduction.second; | ||||||
1565 | return TTI.isLegalToVectorizeReduction(RdxDesc, VF); | ||||||
1566 | })); | ||||||
1567 | } | ||||||
1568 | |||||||
1569 | /// Returns true if \p I is an instruction that will be scalarized with | ||||||
1570 | /// predication when vectorizing \p I with vectorization factor \p VF. Such | ||||||
1571 | /// instructions include conditional stores and instructions that may divide | ||||||
1572 | /// by zero. | ||||||
1573 | bool isScalarWithPredication(Instruction *I, ElementCount VF) const; | ||||||
1574 | |||||||
1575 | // Returns true if \p I is an instruction that will be predicated either | ||||||
1576 | // through scalar predication or masked load/store or masked gather/scatter. | ||||||
1577 | // \p VF is the vectorization factor that will be used to vectorize \p I. | ||||||
1578 | // Superset of instructions that return true for isScalarWithPredication. | ||||||
1579 | bool isPredicatedInst(Instruction *I, ElementCount VF, | ||||||
1580 | bool IsKnownUniform = false) { | ||||||
1581 | // When we know the load is uniform and the original scalar loop was not | ||||||
1582 | // predicated we don't need to mark it as a predicated instruction. Any | ||||||
1583 | // vectorised blocks created when tail-folding are something artificial we | ||||||
1584 | // have introduced and we know there is always at least one active lane. | ||||||
1585 | // That's why we call Legal->blockNeedsPredication here because it doesn't | ||||||
1586 | // query tail-folding. | ||||||
1587 | if (IsKnownUniform && isa<LoadInst>(I) && | ||||||
1588 | !Legal->blockNeedsPredication(I->getParent())) | ||||||
1589 | return false; | ||||||
1590 | if (!blockNeedsPredicationForAnyReason(I->getParent())) | ||||||
1591 | return false; | ||||||
1592 | // Loads and stores that need some form of masked operation are predicated | ||||||
1593 | // instructions. | ||||||
1594 | if (isa<LoadInst>(I) || isa<StoreInst>(I)) | ||||||
1595 | return Legal->isMaskRequired(I); | ||||||
1596 | return isScalarWithPredication(I, VF); | ||||||
1597 | } | ||||||
1598 | |||||||
1599 | /// Returns true if \p I is a memory instruction with consecutive memory | ||||||
1600 | /// access that can be widened. | ||||||
1601 | bool | ||||||
1602 | memoryInstructionCanBeWidened(Instruction *I, | ||||||
1603 | ElementCount VF = ElementCount::getFixed(1)); | ||||||
1604 | |||||||
1605 | /// Returns true if \p I is a memory instruction in an interleaved-group | ||||||
1606 | /// of memory accesses that can be vectorized with wide vector loads/stores | ||||||
1607 | /// and shuffles. | ||||||
1608 | bool | ||||||
1609 | interleavedAccessCanBeWidened(Instruction *I, | ||||||
1610 | ElementCount VF = ElementCount::getFixed(1)); | ||||||
1611 | |||||||
1612 | /// Check if \p Instr belongs to any interleaved access group. | ||||||
1613 | bool isAccessInterleaved(Instruction *Instr) { | ||||||
1614 | return InterleaveInfo.isInterleaved(Instr); | ||||||
1615 | } | ||||||
1616 | |||||||
1617 | /// Get the interleaved access group that \p Instr belongs to. | ||||||
1618 | const InterleaveGroup<Instruction> * | ||||||
1619 | getInterleavedAccessGroup(Instruction *Instr) { | ||||||
1620 | return InterleaveInfo.getInterleaveGroup(Instr); | ||||||
1621 | } | ||||||
1622 | |||||||
1623 | /// Returns true if we're required to use a scalar epilogue for at least | ||||||
1624 | /// the final iteration of the original loop. | ||||||
1625 | bool requiresScalarEpilogue(ElementCount VF) const { | ||||||
1626 | if (!isScalarEpilogueAllowed()) | ||||||
1627 | return false; | ||||||
1628 | // If we might exit from anywhere but the latch, must run the exiting | ||||||
1629 | // iteration in scalar form. | ||||||
1630 | if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) | ||||||
1631 | return true; | ||||||
1632 | return VF.isVector() && InterleaveInfo.requiresScalarEpilogue(); | ||||||
1633 | } | ||||||
1634 | |||||||
1635 | /// Returns true if a scalar epilogue is not allowed due to optsize or a | ||||||
1636 | /// loop hint annotation. | ||||||
1637 | bool isScalarEpilogueAllowed() const { | ||||||
1638 | return ScalarEpilogueStatus == CM_ScalarEpilogueAllowed; | ||||||
1639 | } | ||||||
1640 | |||||||
1641 | /// Returns true if all loop blocks should be masked to fold tail loop. | ||||||
1642 | bool foldTailByMasking() const { return FoldTailByMasking; } | ||||||
1643 | |||||||
1644 | /// Returns true if the instructions in this block requires predication | ||||||
1645 | /// for any reason, e.g. because tail folding now requires a predicate | ||||||
1646 | /// or because the block in the original loop was predicated. | ||||||
1647 | bool blockNeedsPredicationForAnyReason(BasicBlock *BB) const { | ||||||
1648 | return foldTailByMasking() || Legal->blockNeedsPredication(BB); | ||||||
1649 | } | ||||||
1650 | |||||||
1651 | /// A SmallMapVector to store the InLoop reduction op chains, mapping phi | ||||||
1652 | /// nodes to the chain of instructions representing the reductions. Uses a | ||||||
1653 | /// MapVector to ensure deterministic iteration order. | ||||||
1654 | using ReductionChainMap = | ||||||
1655 | SmallMapVector<PHINode *, SmallVector<Instruction *, 4>, 4>; | ||||||
1656 | |||||||
1657 | /// Return the chain of instructions representing an inloop reduction. | ||||||
1658 | const ReductionChainMap &getInLoopReductionChains() const { | ||||||
1659 | return InLoopReductionChains; | ||||||
1660 | } | ||||||
1661 | |||||||
1662 | /// Returns true if the Phi is part of an inloop reduction. | ||||||
1663 | bool isInLoopReduction(PHINode *Phi) const { | ||||||
1664 | return InLoopReductionChains.count(Phi); | ||||||
1665 | } | ||||||
1666 | |||||||
1667 | /// Estimate cost of an intrinsic call instruction CI if it were vectorized | ||||||
1668 | /// with factor VF. Return the cost of the instruction, including | ||||||
1669 | /// scalarization overhead if it's needed. | ||||||
1670 | InstructionCost getVectorIntrinsicCost(CallInst *CI, ElementCount VF) const; | ||||||
1671 | |||||||
1672 | /// Estimate cost of a call instruction CI if it were vectorized with factor | ||||||
1673 | /// VF. Return the cost of the instruction, including scalarization overhead | ||||||
1674 | /// if it's needed. The flag NeedToScalarize shows if the call needs to be | ||||||
1675 | /// scalarized - | ||||||
1676 | /// i.e. either vector version isn't available, or is too expensive. | ||||||
1677 | InstructionCost getVectorCallCost(CallInst *CI, ElementCount VF, | ||||||
1678 | bool &NeedToScalarize) const; | ||||||
1679 | |||||||
1680 | /// Returns true if the per-lane cost of VectorizationFactor A is lower than | ||||||
1681 | /// that of B. | ||||||
1682 | bool isMoreProfitable(const VectorizationFactor &A, | ||||||
1683 | const VectorizationFactor &B) const; | ||||||
1684 | |||||||
1685 | /// Invalidates decisions already taken by the cost model. | ||||||
1686 | void invalidateCostModelingDecisions() { | ||||||
1687 | WideningDecisions.clear(); | ||||||
1688 | Uniforms.clear(); | ||||||
1689 | Scalars.clear(); | ||||||
1690 | } | ||||||
1691 | |||||||
1692 | private: | ||||||
1693 | unsigned NumPredStores = 0; | ||||||
1694 | |||||||
1695 | /// \return An upper bound for the vectorization factors for both | ||||||
1696 | /// fixed and scalable vectorization, where the minimum-known number of | ||||||
1697 | /// elements is a power-of-2 larger than zero. If scalable vectorization is | ||||||
1698 | /// disabled or unsupported, then the scalable part will be equal to | ||||||
1699 | /// ElementCount::getScalable(0). | ||||||
1700 | FixedScalableVFPair computeFeasibleMaxVF(unsigned ConstTripCount, | ||||||
1701 | ElementCount UserVF, | ||||||
1702 | bool FoldTailByMasking); | ||||||
1703 | |||||||
1704 | /// \return the maximized element count based on the targets vector | ||||||
1705 | /// registers and the loop trip-count, but limited to a maximum safe VF. | ||||||
1706 | /// This is a helper function of computeFeasibleMaxVF. | ||||||
1707 | /// FIXME: MaxSafeVF is currently passed by reference to avoid some obscure | ||||||
1708 | /// issue that occurred on one of the buildbots which cannot be reproduced | ||||||
1709 | /// without having access to the properietary compiler (see comments on | ||||||
1710 | /// D98509). The issue is currently under investigation and this workaround | ||||||
1711 | /// will be removed as soon as possible. | ||||||
1712 | ElementCount getMaximizedVFForTarget(unsigned ConstTripCount, | ||||||
1713 | unsigned SmallestType, | ||||||
1714 | unsigned WidestType, | ||||||
1715 | const ElementCount &MaxSafeVF, | ||||||
1716 | bool FoldTailByMasking); | ||||||
1717 | |||||||
1718 | /// \return the maximum legal scalable VF, based on the safe max number | ||||||
1719 | /// of elements. | ||||||
1720 | ElementCount getMaxLegalScalableVF(unsigned MaxSafeElements); | ||||||
1721 | |||||||
1722 | /// The vectorization cost is a combination of the cost itself and a boolean | ||||||
1723 | /// indicating whether any of the contributing operations will actually | ||||||
1724 | /// operate on vector values after type legalization in the backend. If this | ||||||
1725 | /// latter value is false, then all operations will be scalarized (i.e. no | ||||||
1726 | /// vectorization has actually taken place). | ||||||
1727 | using VectorizationCostTy = std::pair<InstructionCost, bool>; | ||||||
1728 | |||||||
1729 | /// Returns the expected execution cost. The unit of the cost does | ||||||
1730 | /// not matter because we use the 'cost' units to compare different | ||||||
1731 | /// vector widths. The cost that is returned is *not* normalized by | ||||||
1732 | /// the factor width. If \p Invalid is not nullptr, this function | ||||||
1733 | /// will add a pair(Instruction*, ElementCount) to \p Invalid for | ||||||
1734 | /// each instruction that has an Invalid cost for the given VF. | ||||||
1735 | using InstructionVFPair = std::pair<Instruction *, ElementCount>; | ||||||
1736 | VectorizationCostTy | ||||||
1737 | expectedCost(ElementCount VF, | ||||||
1738 | SmallVectorImpl<InstructionVFPair> *Invalid = nullptr); | ||||||
1739 | |||||||
1740 | /// Returns the execution time cost of an instruction for a given vector | ||||||
1741 | /// width. Vector width of one means scalar. | ||||||
1742 | VectorizationCostTy getInstructionCost(Instruction *I, ElementCount VF); | ||||||
1743 | |||||||
1744 | /// The cost-computation logic from getInstructionCost which provides | ||||||
1745 | /// the vector type as an output parameter. | ||||||
1746 | InstructionCost getInstructionCost(Instruction *I, ElementCount VF, | ||||||
1747 | Type *&VectorTy); | ||||||
1748 | |||||||
1749 | /// Return the cost of instructions in an inloop reduction pattern, if I is | ||||||
1750 | /// part of that pattern. | ||||||
1751 | Optional<InstructionCost> | ||||||
1752 | getReductionPatternCost(Instruction *I, ElementCount VF, Type *VectorTy, | ||||||
1753 | TTI::TargetCostKind CostKind); | ||||||
1754 | |||||||
1755 | /// Calculate vectorization cost of memory instruction \p I. | ||||||
1756 | InstructionCost getMemoryInstructionCost(Instruction *I, ElementCount VF); | ||||||
1757 | |||||||
1758 | /// The cost computation for scalarized memory instruction. | ||||||
1759 | InstructionCost getMemInstScalarizationCost(Instruction *I, ElementCount VF); | ||||||
1760 | |||||||
1761 | /// The cost computation for interleaving group of memory instructions. | ||||||
1762 | InstructionCost getInterleaveGroupCost(Instruction *I, ElementCount VF); | ||||||
1763 | |||||||
1764 | /// The cost computation for Gather/Scatter instruction. | ||||||
1765 | InstructionCost getGatherScatterCost(Instruction *I, ElementCount VF); | ||||||
1766 | |||||||
1767 | /// The cost computation for widening instruction \p I with consecutive | ||||||
1768 | /// memory access. | ||||||
1769 | InstructionCost getConsecutiveMemOpCost(Instruction *I, ElementCount VF); | ||||||
1770 | |||||||
1771 | /// The cost calculation for Load/Store instruction \p I with uniform pointer - | ||||||
1772 | /// Load: scalar load + broadcast. | ||||||
1773 | /// Store: scalar store + (loop invariant value stored? 0 : extract of last | ||||||
1774 | /// element) | ||||||
1775 | InstructionCost getUniformMemOpCost(Instruction *I, ElementCount VF); | ||||||
1776 | |||||||
1777 | /// Estimate the overhead of scalarizing an instruction. This is a | ||||||
1778 | /// convenience wrapper for the type-based getScalarizationOverhead API. | ||||||
1779 | InstructionCost getScalarizationOverhead(Instruction *I, | ||||||
1780 | ElementCount VF) const; | ||||||
1781 | |||||||
1782 | /// Returns whether the instruction is a load or store and will be a emitted | ||||||
1783 | /// as a vector operation. | ||||||
1784 | bool isConsecutiveLoadOrStore(Instruction *I); | ||||||
1785 | |||||||
1786 | /// Returns true if an artificially high cost for emulated masked memrefs | ||||||
1787 | /// should be used. | ||||||
1788 | bool useEmulatedMaskMemRefHack(Instruction *I, ElementCount VF); | ||||||
1789 | |||||||
1790 | /// Map of scalar integer values to the smallest bitwidth they can be legally | ||||||
1791 | /// represented as. The vector equivalents of these values should be truncated | ||||||
1792 | /// to this type. | ||||||
1793 | MapVector<Instruction *, uint64_t> MinBWs; | ||||||
1794 | |||||||
1795 | /// A type representing the costs for instructions if they were to be | ||||||
1796 | /// scalarized rather than vectorized. The entries are Instruction-Cost | ||||||
1797 | /// pairs. | ||||||
1798 | using ScalarCostsTy = DenseMap<Instruction *, InstructionCost>; | ||||||
1799 | |||||||
1800 | /// A set containing all BasicBlocks that are known to present after | ||||||
1801 | /// vectorization as a predicated block. | ||||||
1802 | SmallPtrSet<BasicBlock *, 4> PredicatedBBsAfterVectorization; | ||||||
1803 | |||||||
1804 | /// Records whether it is allowed to have the original scalar loop execute at | ||||||
1805 | /// least once. This may be needed as a fallback loop in case runtime | ||||||
1806 | /// aliasing/dependence checks fail, or to handle the tail/remainder | ||||||
1807 | /// iterations when the trip count is unknown or doesn't divide by the VF, | ||||||
1808 | /// or as a peel-loop to handle gaps in interleave-groups. | ||||||
1809 | /// Under optsize and when the trip count is very small we don't allow any | ||||||
1810 | /// iterations to execute in the scalar loop. | ||||||
1811 | ScalarEpilogueLowering ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; | ||||||
1812 | |||||||
1813 | /// All blocks of loop are to be masked to fold tail of scalar iterations. | ||||||
1814 | bool FoldTailByMasking = false; | ||||||
1815 | |||||||
1816 | /// A map holding scalar costs for different vectorization factors. The | ||||||
1817 | /// presence of a cost for an instruction in the mapping indicates that the | ||||||
1818 | /// instruction will be scalarized when vectorizing with the associated | ||||||
1819 | /// vectorization factor. The entries are VF-ScalarCostTy pairs. | ||||||
1820 | DenseMap<ElementCount, ScalarCostsTy> InstsToScalarize; | ||||||
1821 | |||||||
1822 | /// Holds the instructions known to be uniform after vectorization. | ||||||
1823 | /// The data is collected per VF. | ||||||
1824 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> Uniforms; | ||||||
1825 | |||||||
1826 | /// Holds the instructions known to be scalar after vectorization. | ||||||
1827 | /// The data is collected per VF. | ||||||
1828 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> Scalars; | ||||||
1829 | |||||||
1830 | /// Holds the instructions (address computations) that are forced to be | ||||||
1831 | /// scalarized. | ||||||
1832 | DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> ForcedScalars; | ||||||
1833 | |||||||
1834 | /// PHINodes of the reductions that should be expanded in-loop along with | ||||||
1835 | /// their associated chains of reduction operations, in program order from top | ||||||
1836 | /// (PHI) to bottom | ||||||
1837 | ReductionChainMap InLoopReductionChains; | ||||||
1838 | |||||||
1839 | /// A Map of inloop reduction operations and their immediate chain operand. | ||||||
1840 | /// FIXME: This can be removed once reductions can be costed correctly in | ||||||
1841 | /// vplan. This was added to allow quick lookup to the inloop operations, | ||||||
1842 | /// without having to loop through InLoopReductionChains. | ||||||
1843 | DenseMap<Instruction *, Instruction *> InLoopReductionImmediateChains; | ||||||
1844 | |||||||
1845 | /// Returns the expected difference in cost from scalarizing the expression | ||||||
1846 | /// feeding a predicated instruction \p PredInst. The instructions to | ||||||
1847 | /// scalarize and their scalar costs are collected in \p ScalarCosts. A | ||||||
1848 | /// non-negative return value implies the expression will be scalarized. | ||||||
1849 | /// Currently, only single-use chains are considered for scalarization. | ||||||
1850 | int computePredInstDiscount(Instruction *PredInst, ScalarCostsTy &ScalarCosts, | ||||||
1851 | ElementCount VF); | ||||||
1852 | |||||||
1853 | /// Collect the instructions that are uniform after vectorization. An | ||||||
1854 | /// instruction is uniform if we represent it with a single scalar value in | ||||||
1855 | /// the vectorized loop corresponding to each vector iteration. Examples of | ||||||
1856 | /// uniform instructions include pointer operands of consecutive or | ||||||
1857 | /// interleaved memory accesses. Note that although uniformity implies an | ||||||
1858 | /// instruction will be scalar, the reverse is not true. In general, a | ||||||
1859 | /// scalarized instruction will be represented by VF scalar values in the | ||||||
1860 | /// vectorized loop, each corresponding to an iteration of the original | ||||||
1861 | /// scalar loop. | ||||||
1862 | void collectLoopUniforms(ElementCount VF); | ||||||
1863 | |||||||
1864 | /// Collect the instructions that are scalar after vectorization. An | ||||||
1865 | /// instruction is scalar if it is known to be uniform or will be scalarized | ||||||
1866 | /// during vectorization. collectLoopScalars should only add non-uniform nodes | ||||||
1867 | /// to the list if they are used by a load/store instruction that is marked as | ||||||
1868 | /// CM_Scalarize. Non-uniform scalarized instructions will be represented by | ||||||
1869 | /// VF values in the vectorized loop, each corresponding to an iteration of | ||||||
1870 | /// the original scalar loop. | ||||||
1871 | void collectLoopScalars(ElementCount VF); | ||||||
1872 | |||||||
1873 | /// Keeps cost model vectorization decision and cost for instructions. | ||||||
1874 | /// Right now it is used for memory instructions only. | ||||||
1875 | using DecisionList = DenseMap<std::pair<Instruction *, ElementCount>, | ||||||
1876 | std::pair<InstWidening, InstructionCost>>; | ||||||
1877 | |||||||
1878 | DecisionList WideningDecisions; | ||||||
1879 | |||||||
1880 | /// Returns true if \p V is expected to be vectorized and it needs to be | ||||||
1881 | /// extracted. | ||||||
1882 | bool needsExtract(Value *V, ElementCount VF) const { | ||||||
1883 | Instruction *I = dyn_cast<Instruction>(V); | ||||||
1884 | if (VF.isScalar() || !I || !TheLoop->contains(I) || | ||||||
1885 | TheLoop->isLoopInvariant(I)) | ||||||
1886 | return false; | ||||||
1887 | |||||||
1888 | // Assume we can vectorize V (and hence we need extraction) if the | ||||||
1889 | // scalars are not computed yet. This can happen, because it is called | ||||||
1890 | // via getScalarizationOverhead from setCostBasedWideningDecision, before | ||||||
1891 | // the scalars are collected. That should be a safe assumption in most | ||||||
1892 | // cases, because we check if the operands have vectorizable types | ||||||
1893 | // beforehand in LoopVectorizationLegality. | ||||||
1894 | return Scalars.find(VF) == Scalars.end() || | ||||||
1895 | !isScalarAfterVectorization(I, VF); | ||||||
1896 | }; | ||||||
1897 | |||||||
1898 | /// Returns a range containing only operands needing to be extracted. | ||||||
1899 | SmallVector<Value *, 4> filterExtractingOperands(Instruction::op_range Ops, | ||||||
1900 | ElementCount VF) const { | ||||||
1901 | return SmallVector<Value *, 4>(make_filter_range( | ||||||
1902 | Ops, [this, VF](Value *V) { return this->needsExtract(V, VF); })); | ||||||
1903 | } | ||||||
1904 | |||||||
1905 | /// Determines if we have the infrastructure to vectorize loop \p L and its | ||||||
1906 | /// epilogue, assuming the main loop is vectorized by \p VF. | ||||||
1907 | bool isCandidateForEpilogueVectorization(const Loop &L, | ||||||
1908 | const ElementCount VF) const; | ||||||
1909 | |||||||
1910 | /// Returns true if epilogue vectorization is considered profitable, and | ||||||
1911 | /// false otherwise. | ||||||
1912 | /// \p VF is the vectorization factor chosen for the original loop. | ||||||
1913 | bool isEpilogueVectorizationProfitable(const ElementCount VF) const; | ||||||
1914 | |||||||
1915 | public: | ||||||
1916 | /// The loop that we evaluate. | ||||||
1917 | Loop *TheLoop; | ||||||
1918 | |||||||
1919 | /// Predicated scalar evolution analysis. | ||||||
1920 | PredicatedScalarEvolution &PSE; | ||||||
1921 | |||||||
1922 | /// Loop Info analysis. | ||||||
1923 | LoopInfo *LI; | ||||||
1924 | |||||||
1925 | /// Vectorization legality. | ||||||
1926 | LoopVectorizationLegality *Legal; | ||||||
1927 | |||||||
1928 | /// Vector target information. | ||||||
1929 | const TargetTransformInfo &TTI; | ||||||
1930 | |||||||
1931 | /// Target Library Info. | ||||||
1932 | const TargetLibraryInfo *TLI; | ||||||
1933 | |||||||
1934 | /// Demanded bits analysis. | ||||||
1935 | DemandedBits *DB; | ||||||
1936 | |||||||
1937 | /// Assumption cache. | ||||||
1938 | AssumptionCache *AC; | ||||||
1939 | |||||||
1940 | /// Interface to emit optimization remarks. | ||||||
1941 | OptimizationRemarkEmitter *ORE; | ||||||
1942 | |||||||
1943 | const Function *TheFunction; | ||||||
1944 | |||||||
1945 | /// Loop Vectorize Hint. | ||||||
1946 | const LoopVectorizeHints *Hints; | ||||||
1947 | |||||||
1948 | /// The interleave access information contains groups of interleaved accesses | ||||||
1949 | /// with the same stride and close to each other. | ||||||
1950 | InterleavedAccessInfo &InterleaveInfo; | ||||||
1951 | |||||||
1952 | /// Values to ignore in the cost model. | ||||||
1953 | SmallPtrSet<const Value *, 16> ValuesToIgnore; | ||||||
1954 | |||||||
1955 | /// Values to ignore in the cost model when VF > 1. | ||||||
1956 | SmallPtrSet<const Value *, 16> VecValuesToIgnore; | ||||||
1957 | |||||||
1958 | /// All element types found in the loop. | ||||||
1959 | SmallPtrSet<Type *, 16> ElementTypesInLoop; | ||||||
1960 | |||||||
1961 | /// Profitable vector factors. | ||||||
1962 | SmallVector<VectorizationFactor, 8> ProfitableVFs; | ||||||
1963 | }; | ||||||
1964 | } // end namespace llvm | ||||||
1965 | |||||||
1966 | /// Helper struct to manage generating runtime checks for vectorization. | ||||||
1967 | /// | ||||||
1968 | /// The runtime checks are created up-front in temporary blocks to allow better | ||||||
1969 | /// estimating the cost and un-linked from the existing IR. After deciding to | ||||||
1970 | /// vectorize, the checks are moved back. If deciding not to vectorize, the | ||||||
1971 | /// temporary blocks are completely removed. | ||||||
1972 | class GeneratedRTChecks { | ||||||
1973 | /// Basic block which contains the generated SCEV checks, if any. | ||||||
1974 | BasicBlock *SCEVCheckBlock = nullptr; | ||||||
1975 | |||||||
1976 | /// The value representing the result of the generated SCEV checks. If it is | ||||||
1977 | /// nullptr, either no SCEV checks have been generated or they have been used. | ||||||
1978 | Value *SCEVCheckCond = nullptr; | ||||||
1979 | |||||||
1980 | /// Basic block which contains the generated memory runtime checks, if any. | ||||||
1981 | BasicBlock *MemCheckBlock = nullptr; | ||||||
1982 | |||||||
1983 | /// The value representing the result of the generated memory runtime checks. | ||||||
1984 | /// If it is nullptr, either no memory runtime checks have been generated or | ||||||
1985 | /// they have been used. | ||||||
1986 | Value *MemRuntimeCheckCond = nullptr; | ||||||
1987 | |||||||
1988 | DominatorTree *DT; | ||||||
1989 | LoopInfo *LI; | ||||||
1990 | |||||||
1991 | SCEVExpander SCEVExp; | ||||||
1992 | SCEVExpander MemCheckExp; | ||||||
1993 | |||||||
1994 | public: | ||||||
1995 | GeneratedRTChecks(ScalarEvolution &SE, DominatorTree *DT, LoopInfo *LI, | ||||||
1996 | const DataLayout &DL) | ||||||
1997 | : DT(DT), LI(LI), SCEVExp(SE, DL, "scev.check"), | ||||||
1998 | MemCheckExp(SE, DL, "scev.check") {} | ||||||
1999 | |||||||
2000 | /// Generate runtime checks in SCEVCheckBlock and MemCheckBlock, so we can | ||||||
2001 | /// accurately estimate the cost of the runtime checks. The blocks are | ||||||
2002 | /// un-linked from the IR and is added back during vector code generation. If | ||||||
2003 | /// there is no vector code generation, the check blocks are removed | ||||||
2004 | /// completely. | ||||||
2005 | void Create(Loop *L, const LoopAccessInfo &LAI, | ||||||
2006 | const SCEVUnionPredicate &UnionPred) { | ||||||
2007 | |||||||
2008 | BasicBlock *LoopHeader = L->getHeader(); | ||||||
2009 | BasicBlock *Preheader = L->getLoopPreheader(); | ||||||
2010 | |||||||
2011 | // Use SplitBlock to create blocks for SCEV & memory runtime checks to | ||||||
2012 | // ensure the blocks are properly added to LoopInfo & DominatorTree. Those | ||||||
2013 | // may be used by SCEVExpander. The blocks will be un-linked from their | ||||||
2014 | // predecessors and removed from LI & DT at the end of the function. | ||||||
2015 | if (!UnionPred.isAlwaysTrue()) { | ||||||
2016 | SCEVCheckBlock = SplitBlock(Preheader, Preheader->getTerminator(), DT, LI, | ||||||
2017 | nullptr, "vector.scevcheck"); | ||||||
2018 | |||||||
2019 | SCEVCheckCond = SCEVExp.expandCodeForPredicate( | ||||||
2020 | &UnionPred, SCEVCheckBlock->getTerminator()); | ||||||
2021 | } | ||||||
2022 | |||||||
2023 | const auto &RtPtrChecking = *LAI.getRuntimePointerChecking(); | ||||||
2024 | if (RtPtrChecking.Need) { | ||||||
2025 | auto *Pred = SCEVCheckBlock ? SCEVCheckBlock : Preheader; | ||||||
2026 | MemCheckBlock = SplitBlock(Pred, Pred->getTerminator(), DT, LI, nullptr, | ||||||
2027 | "vector.memcheck"); | ||||||
2028 | |||||||
2029 | MemRuntimeCheckCond = | ||||||
2030 | addRuntimeChecks(MemCheckBlock->getTerminator(), L, | ||||||
2031 | RtPtrChecking.getChecks(), MemCheckExp); | ||||||
2032 | 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", 2034, __extension__ __PRETTY_FUNCTION__)) | ||||||
2033 | "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", 2034, __extension__ __PRETTY_FUNCTION__)) | ||||||
2034 | "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", 2034, __extension__ __PRETTY_FUNCTION__)); | ||||||
2035 | } | ||||||
2036 | |||||||
2037 | if (!MemCheckBlock && !SCEVCheckBlock) | ||||||
2038 | return; | ||||||
2039 | |||||||
2040 | // Unhook the temporary block with the checks, update various places | ||||||
2041 | // accordingly. | ||||||
2042 | if (SCEVCheckBlock) | ||||||
2043 | SCEVCheckBlock->replaceAllUsesWith(Preheader); | ||||||
2044 | if (MemCheckBlock) | ||||||
2045 | MemCheckBlock->replaceAllUsesWith(Preheader); | ||||||
2046 | |||||||
2047 | if (SCEVCheckBlock) { | ||||||
2048 | SCEVCheckBlock->getTerminator()->moveBefore(Preheader->getTerminator()); | ||||||
2049 | new UnreachableInst(Preheader->getContext(), SCEVCheckBlock); | ||||||
2050 | Preheader->getTerminator()->eraseFromParent(); | ||||||
2051 | } | ||||||
2052 | if (MemCheckBlock) { | ||||||
2053 | MemCheckBlock->getTerminator()->moveBefore(Preheader->getTerminator()); | ||||||
2054 | new UnreachableInst(Preheader->getContext(), MemCheckBlock); | ||||||
2055 | Preheader->getTerminator()->eraseFromParent(); | ||||||
2056 | } | ||||||
2057 | |||||||
2058 | DT->changeImmediateDominator(LoopHeader, Preheader); | ||||||
2059 | if (MemCheckBlock) { | ||||||
2060 | DT->eraseNode(MemCheckBlock); | ||||||
2061 | LI->removeBlock(MemCheckBlock); | ||||||
2062 | } | ||||||
2063 | if (SCEVCheckBlock) { | ||||||
2064 | DT->eraseNode(SCEVCheckBlock); | ||||||
2065 | LI->removeBlock(SCEVCheckBlock); | ||||||
2066 | } | ||||||
2067 | } | ||||||
2068 | |||||||
2069 | /// Remove the created SCEV & memory runtime check blocks & instructions, if | ||||||
2070 | /// unused. | ||||||
2071 | ~GeneratedRTChecks() { | ||||||
2072 | SCEVExpanderCleaner SCEVCleaner(SCEVExp); | ||||||
2073 | SCEVExpanderCleaner MemCheckCleaner(MemCheckExp); | ||||||
2074 | if (!SCEVCheckCond) | ||||||
2075 | SCEVCleaner.markResultUsed(); | ||||||
2076 | |||||||
2077 | if (!MemRuntimeCheckCond) | ||||||
2078 | MemCheckCleaner.markResultUsed(); | ||||||
2079 | |||||||
2080 | if (MemRuntimeCheckCond) { | ||||||
2081 | auto &SE = *MemCheckExp.getSE(); | ||||||
2082 | // Memory runtime check generation creates compares that use expanded | ||||||
2083 | // values. Remove them before running the SCEVExpanderCleaners. | ||||||
2084 | for (auto &I : make_early_inc_range(reverse(*MemCheckBlock))) { | ||||||
2085 | if (MemCheckExp.isInsertedInstruction(&I)) | ||||||
2086 | continue; | ||||||
2087 | SE.forgetValue(&I); | ||||||
2088 | I.eraseFromParent(); | ||||||
2089 | } | ||||||
2090 | } | ||||||
2091 | MemCheckCleaner.cleanup(); | ||||||
2092 | SCEVCleaner.cleanup(); | ||||||
2093 | |||||||
2094 | if (SCEVCheckCond) | ||||||
2095 | SCEVCheckBlock->eraseFromParent(); | ||||||
2096 | if (MemRuntimeCheckCond) | ||||||
2097 | MemCheckBlock->eraseFromParent(); | ||||||
2098 | } | ||||||
2099 | |||||||
2100 | /// Adds the generated SCEVCheckBlock before \p LoopVectorPreHeader and | ||||||
2101 | /// adjusts the branches to branch to the vector preheader or \p Bypass, | ||||||
2102 | /// depending on the generated condition. | ||||||
2103 | BasicBlock *emitSCEVChecks(Loop *L, BasicBlock *Bypass, | ||||||
2104 | BasicBlock *LoopVectorPreHeader, | ||||||
2105 | BasicBlock *LoopExitBlock) { | ||||||
2106 | if (!SCEVCheckCond) | ||||||
2107 | return nullptr; | ||||||
2108 | if (auto *C = dyn_cast<ConstantInt>(SCEVCheckCond)) | ||||||
2109 | if (C->isZero()) | ||||||
2110 | return nullptr; | ||||||
2111 | |||||||
2112 | auto *Pred = LoopVectorPreHeader->getSinglePredecessor(); | ||||||
2113 | |||||||
2114 | BranchInst::Create(LoopVectorPreHeader, SCEVCheckBlock); | ||||||
2115 | // Create new preheader for vector loop. | ||||||
2116 | if (auto *PL = LI->getLoopFor(LoopVectorPreHeader)) | ||||||
2117 | PL->addBasicBlockToLoop(SCEVCheckBlock, *LI); | ||||||
2118 | |||||||
2119 | SCEVCheckBlock->getTerminator()->eraseFromParent(); | ||||||
2120 | SCEVCheckBlock->moveBefore(LoopVectorPreHeader); | ||||||
2121 | Pred->getTerminator()->replaceSuccessorWith(LoopVectorPreHeader, | ||||||
2122 | SCEVCheckBlock); | ||||||
2123 | |||||||
2124 | DT->addNewBlock(SCEVCheckBlock, Pred); | ||||||
2125 | DT->changeImmediateDominator(LoopVectorPreHeader, SCEVCheckBlock); | ||||||
2126 | |||||||
2127 | ReplaceInstWithInst( | ||||||
2128 | SCEVCheckBlock->getTerminator(), | ||||||
2129 | BranchInst::Create(Bypass, LoopVectorPreHeader, SCEVCheckCond)); | ||||||
2130 | // Mark the check as used, to prevent it from being removed during cleanup. | ||||||
2131 | SCEVCheckCond = nullptr; | ||||||
2132 | return SCEVCheckBlock; | ||||||
2133 | } | ||||||
2134 | |||||||
2135 | /// Adds the generated MemCheckBlock before \p LoopVectorPreHeader and adjusts | ||||||
2136 | /// the branches to branch to the vector preheader or \p Bypass, depending on | ||||||
2137 | /// the generated condition. | ||||||
2138 | BasicBlock *emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass, | ||||||
2139 | BasicBlock *LoopVectorPreHeader) { | ||||||
2140 | // Check if we generated code that checks in runtime if arrays overlap. | ||||||
2141 | if (!MemRuntimeCheckCond) | ||||||
2142 | return nullptr; | ||||||
2143 | |||||||
2144 | auto *Pred = LoopVectorPreHeader->getSinglePredecessor(); | ||||||
2145 | Pred->getTerminator()->replaceSuccessorWith(LoopVectorPreHeader, | ||||||
2146 | MemCheckBlock); | ||||||
2147 | |||||||
2148 | DT->addNewBlock(MemCheckBlock, Pred); | ||||||
2149 | DT->changeImmediateDominator(LoopVectorPreHeader, MemCheckBlock); | ||||||
2150 | MemCheckBlock->moveBefore(LoopVectorPreHeader); | ||||||
2151 | |||||||
2152 | if (auto *PL = LI->getLoopFor(LoopVectorPreHeader)) | ||||||
2153 | PL->addBasicBlockToLoop(MemCheckBlock, *LI); | ||||||
2154 | |||||||
2155 | ReplaceInstWithInst( | ||||||
2156 | MemCheckBlock->getTerminator(), | ||||||
2157 | BranchInst::Create(Bypass, LoopVectorPreHeader, MemRuntimeCheckCond)); | ||||||
2158 | MemCheckBlock->getTerminator()->setDebugLoc( | ||||||
2159 | Pred->getTerminator()->getDebugLoc()); | ||||||
2160 | |||||||
2161 | // Mark the check as used, to prevent it from being removed during cleanup. | ||||||
2162 | MemRuntimeCheckCond = nullptr; | ||||||
2163 | return MemCheckBlock; | ||||||
2164 | } | ||||||
2165 | }; | ||||||
2166 | |||||||
2167 | // Return true if \p OuterLp is an outer loop annotated with hints for explicit | ||||||
2168 | // vectorization. The loop needs to be annotated with #pragma omp simd | ||||||
2169 | // simdlen(#) or #pragma clang vectorize(enable) vectorize_width(#). If the | ||||||
2170 | // vector length information is not provided, vectorization is not considered | ||||||
2171 | // explicit. Interleave hints are not allowed either. These limitations will be | ||||||
2172 | // relaxed in the future. | ||||||
2173 | // Please, note that we are currently forced to abuse the pragma 'clang | ||||||
2174 | // vectorize' semantics. This pragma provides *auto-vectorization hints* | ||||||
2175 | // (i.e., LV must check that vectorization is legal) whereas pragma 'omp simd' | ||||||
2176 | // provides *explicit vectorization hints* (LV can bypass legal checks and | ||||||
2177 | // assume that vectorization is legal). However, both hints are implemented | ||||||
2178 | // using the same metadata (llvm.loop.vectorize, processed by | ||||||
2179 | // LoopVectorizeHints). This will be fixed in the future when the native IR | ||||||
2180 | // representation for pragma 'omp simd' is introduced. | ||||||
2181 | static bool isExplicitVecOuterLoop(Loop *OuterLp, | ||||||
2182 | OptimizationRemarkEmitter *ORE) { | ||||||
2183 | 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", 2183, __extension__ __PRETTY_FUNCTION__)); | ||||||
2184 | LoopVectorizeHints Hints(OuterLp, true /*DisableInterleaving*/, *ORE); | ||||||
2185 | |||||||
2186 | // Only outer loops with an explicit vectorization hint are supported. | ||||||
2187 | // Unannotated outer loops are ignored. | ||||||
2188 | if (Hints.getForce() == LoopVectorizeHints::FK_Undefined) | ||||||
2189 | return false; | ||||||
2190 | |||||||
2191 | Function *Fn = OuterLp->getHeader()->getParent(); | ||||||
2192 | if (!Hints.allowVectorization(Fn, OuterLp, | ||||||
2193 | true /*VectorizeOnlyWhenForced*/)) { | ||||||
2194 | 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); | ||||||
2195 | return false; | ||||||
2196 | } | ||||||
2197 | |||||||
2198 | if (Hints.getInterleave() > 1) { | ||||||
2199 | // TODO: Interleave support is future work. | ||||||
2200 | 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) | ||||||
2201 | "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); | ||||||
2202 | Hints.emitRemarkWithHints(); | ||||||
2203 | return false; | ||||||
2204 | } | ||||||
2205 | |||||||
2206 | return true; | ||||||
2207 | } | ||||||
2208 | |||||||
2209 | static void collectSupportedLoops(Loop &L, LoopInfo *LI, | ||||||
2210 | OptimizationRemarkEmitter *ORE, | ||||||
2211 | SmallVectorImpl<Loop *> &V) { | ||||||
2212 | // Collect inner loops and outer loops without irreducible control flow. For | ||||||
2213 | // now, only collect outer loops that have explicit vectorization hints. If we | ||||||
2214 | // are stress testing the VPlan H-CFG construction, we collect the outermost | ||||||
2215 | // loop of every loop nest. | ||||||
2216 | if (L.isInnermost() || VPlanBuildStressTest || | ||||||
2217 | (EnableVPlanNativePath && isExplicitVecOuterLoop(&L, ORE))) { | ||||||
2218 | LoopBlocksRPO RPOT(&L); | ||||||
2219 | RPOT.perform(LI); | ||||||
2220 | if (!containsIrreducibleCFG<const BasicBlock *>(RPOT, *LI)) { | ||||||
2221 | V.push_back(&L); | ||||||
2222 | // TODO: Collect inner loops inside marked outer loops in case | ||||||
2223 | // vectorization fails for the outer loop. Do not invoke | ||||||
2224 | // 'containsIrreducibleCFG' again for inner loops when the outer loop is | ||||||
2225 | // already known to be reducible. We can use an inherited attribute for | ||||||
2226 | // that. | ||||||
2227 | return; | ||||||
2228 | } | ||||||
2229 | } | ||||||
2230 | for (Loop *InnerL : L) | ||||||
2231 | collectSupportedLoops(*InnerL, LI, ORE, V); | ||||||
2232 | } | ||||||
2233 | |||||||
2234 | namespace { | ||||||
2235 | |||||||
2236 | /// The LoopVectorize Pass. | ||||||
2237 | struct LoopVectorize : public FunctionPass { | ||||||
2238 | /// Pass identification, replacement for typeid | ||||||
2239 | static char ID; | ||||||
2240 | |||||||
2241 | LoopVectorizePass Impl; | ||||||
2242 | |||||||
2243 | explicit LoopVectorize(bool InterleaveOnlyWhenForced = false, | ||||||
2244 | bool VectorizeOnlyWhenForced = false) | ||||||
2245 | : FunctionPass(ID), | ||||||
2246 | Impl({InterleaveOnlyWhenForced, VectorizeOnlyWhenForced}) { | ||||||
2247 | initializeLoopVectorizePass(*PassRegistry::getPassRegistry()); | ||||||
2248 | } | ||||||
2249 | |||||||
2250 | bool runOnFunction(Function &F) override { | ||||||
2251 | if (skipFunction(F)) | ||||||
2252 | return false; | ||||||
2253 | |||||||
2254 | auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); | ||||||
2255 | auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | ||||||
2256 | auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | ||||||
2257 | auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | ||||||
2258 | auto *BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI(); | ||||||
2259 | auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); | ||||||
2260 | auto *TLI = TLIP ? &TLIP->getTLI(F) : nullptr; | ||||||
2261 | auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); | ||||||
2262 | auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); | ||||||
2263 | auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>(); | ||||||
2264 | auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits(); | ||||||
2265 | auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); | ||||||
2266 | auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); | ||||||
2267 | |||||||
2268 | std::function<const LoopAccessInfo &(Loop &)> GetLAA = | ||||||
2269 | [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); }; | ||||||
2270 | |||||||
2271 | return Impl.runImpl(F, *SE, *LI, *TTI, *DT, *BFI, TLI, *DB, *AA, *AC, | ||||||
2272 | GetLAA, *ORE, PSI).MadeAnyChange; | ||||||
2273 | } | ||||||
2274 | |||||||
2275 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||||
2276 | AU.addRequired<AssumptionCacheTracker>(); | ||||||
2277 | AU.addRequired<BlockFrequencyInfoWrapperPass>(); | ||||||
2278 | AU.addRequired<DominatorTreeWrapperPass>(); | ||||||
2279 | AU.addRequired<LoopInfoWrapperPass>(); | ||||||
2280 | AU.addRequired<ScalarEvolutionWrapperPass>(); | ||||||
2281 | AU.addRequired<TargetTransformInfoWrapperPass>(); | ||||||
2282 | AU.addRequired<AAResultsWrapperPass>(); | ||||||
2283 | AU.addRequired<LoopAccessLegacyAnalysis>(); | ||||||
2284 | AU.addRequired<DemandedBitsWrapperPass>(); | ||||||
2285 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); | ||||||
2286 | AU.addRequired<InjectTLIMappingsLegacy>(); | ||||||
2287 | |||||||
2288 | // We currently do not preserve loopinfo/dominator analyses with outer loop | ||||||
2289 | // vectorization. Until this is addressed, mark these analyses as preserved | ||||||
2290 | // only for non-VPlan-native path. | ||||||
2291 | // TODO: Preserve Loop and Dominator analyses for VPlan-native path. | ||||||
2292 | if (!EnableVPlanNativePath) { | ||||||
2293 | AU.addPreserved<LoopInfoWrapperPass>(); | ||||||
2294 | AU.addPreserved<DominatorTreeWrapperPass>(); | ||||||
2295 | } | ||||||
2296 | |||||||
2297 | AU.addPreserved<BasicAAWrapperPass>(); | ||||||
2298 | AU.addPreserved<GlobalsAAWrapperPass>(); | ||||||
2299 | AU.addRequired<ProfileSummaryInfoWrapperPass>(); | ||||||
2300 | } | ||||||
2301 | }; | ||||||
2302 | |||||||
2303 | } // end anonymous namespace | ||||||
2304 | |||||||
2305 | //===----------------------------------------------------------------------===// | ||||||
2306 | // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and | ||||||
2307 | // LoopVectorizationCostModel and LoopVectorizationPlanner. | ||||||
2308 | //===----------------------------------------------------------------------===// | ||||||
2309 | |||||||
2310 | Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) { | ||||||
2311 | // We need to place the broadcast of invariant variables outside the loop, | ||||||
2312 | // but only if it's proven safe to do so. Else, broadcast will be inside | ||||||
2313 | // vector loop body. | ||||||
2314 | Instruction *Instr = dyn_cast<Instruction>(V); | ||||||
2315 | bool SafeToHoist = OrigLoop->isLoopInvariant(V) && | ||||||
2316 | (!Instr || | ||||||
2317 | DT->dominates(Instr->getParent(), LoopVectorPreHeader)); | ||||||
2318 | // Place the code for broadcasting invariant variables in the new preheader. | ||||||
2319 | IRBuilder<>::InsertPointGuard Guard(Builder); | ||||||
2320 | if (SafeToHoist) | ||||||
2321 | Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); | ||||||
2322 | |||||||
2323 | // Broadcast the scalar into all locations in the vector. | ||||||
2324 | Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast"); | ||||||
2325 | |||||||
2326 | return Shuf; | ||||||
2327 | } | ||||||
2328 | |||||||
2329 | /// This function adds | ||||||
2330 | /// (StartIdx * Step, (StartIdx + 1) * Step, (StartIdx + 2) * Step, ...) | ||||||
2331 | /// to each vector element of Val. The sequence starts at StartIndex. | ||||||
2332 | /// \p Opcode is relevant for FP induction variable. | ||||||
2333 | static Value *getStepVector(Value *Val, Value *StartIdx, Value *Step, | ||||||
2334 | Instruction::BinaryOps BinOp, ElementCount VF, | ||||||
2335 | IRBuilder<> &Builder) { | ||||||
2336 | 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", 2336, __extension__ __PRETTY_FUNCTION__)); | ||||||
2337 | |||||||
2338 | // Create and check the types. | ||||||
2339 | auto *ValVTy = cast<VectorType>(Val->getType()); | ||||||
2340 | ElementCount VLen = ValVTy->getElementCount(); | ||||||
2341 | |||||||
2342 | Type *STy = Val->getType()->getScalarType(); | ||||||
2343 | 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", 2344, __extension__ __PRETTY_FUNCTION__)) | ||||||
2344 | "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", 2344, __extension__ __PRETTY_FUNCTION__)); | ||||||
2345 | 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", 2345, __extension__ __PRETTY_FUNCTION__)); | ||||||
2346 | |||||||
2347 | SmallVector<Constant *, 8> Indices; | ||||||
2348 | |||||||
2349 | // Create a vector of consecutive numbers from zero to VF. | ||||||
2350 | VectorType *InitVecValVTy = ValVTy; | ||||||
2351 | Type *InitVecValSTy = STy; | ||||||
2352 | if (STy->isFloatingPointTy()) { | ||||||
2353 | InitVecValSTy = | ||||||
2354 | IntegerType::get(STy->getContext(), STy->getScalarSizeInBits()); | ||||||
2355 | InitVecValVTy = VectorType::get(InitVecValSTy, VLen); | ||||||
2356 | } | ||||||
2357 | Value *InitVec = Builder.CreateStepVector(InitVecValVTy); | ||||||
2358 | |||||||
2359 | // Splat the StartIdx | ||||||
2360 | Value *StartIdxSplat = Builder.CreateVectorSplat(VLen, StartIdx); | ||||||
2361 | |||||||
2362 | if (STy->isIntegerTy()) { | ||||||
2363 | InitVec = Builder.CreateAdd(InitVec, StartIdxSplat); | ||||||
2364 | Step = Builder.CreateVectorSplat(VLen, Step); | ||||||
2365 | 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", 2365, __extension__ __PRETTY_FUNCTION__)); | ||||||
2366 | // FIXME: The newly created binary instructions should contain nsw/nuw | ||||||
2367 | // flags, which can be found from the original scalar operations. | ||||||
2368 | Step = Builder.CreateMul(InitVec, Step); | ||||||
2369 | return Builder.CreateAdd(Val, Step, "induction"); | ||||||
2370 | } | ||||||
2371 | |||||||
2372 | // Floating point induction. | ||||||
2373 | 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", 2374, __extension__ __PRETTY_FUNCTION__)) | ||||||
2374 | "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", 2374, __extension__ __PRETTY_FUNCTION__)); | ||||||
2375 | InitVec = Builder.CreateUIToFP(InitVec, ValVTy); | ||||||
2376 | InitVec = Builder.CreateFAdd(InitVec, StartIdxSplat); | ||||||
2377 | |||||||
2378 | Step = Builder.CreateVectorSplat(VLen, Step); | ||||||
2379 | Value *MulOp = Builder.CreateFMul(InitVec, Step); | ||||||
2380 | return Builder.CreateBinOp(BinOp, Val, MulOp, "induction"); | ||||||
2381 | } | ||||||
2382 | |||||||
2383 | void InnerLoopVectorizer::createVectorIntOrFpInductionPHI( | ||||||
2384 | const InductionDescriptor &II, Value *Step, Value *Start, | ||||||
2385 | Instruction *EntryVal, VPValue *Def, VPTransformState &State) { | ||||||
2386 | IRBuilder<> &Builder = State.Builder; | ||||||
2387 | 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", 2388, __extension__ __PRETTY_FUNCTION__)) | ||||||
2388 | "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", 2388, __extension__ __PRETTY_FUNCTION__)); | ||||||
2389 | |||||||
2390 | // Construct the initial value of the vector IV in the vector loop preheader | ||||||
2391 | auto CurrIP = Builder.saveIP(); | ||||||
2392 | Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator()); | ||||||
2393 | if (isa<TruncInst>(EntryVal)) { | ||||||
2394 | 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", 2395, __extension__ __PRETTY_FUNCTION__)) | ||||||
2395 | "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", 2395, __extension__ __PRETTY_FUNCTION__)); | ||||||
2396 | auto *TruncType = cast<IntegerType>(EntryVal->getType()); | ||||||
2397 | Step = Builder.CreateTrunc(Step, TruncType); | ||||||
2398 | Start = Builder.CreateCast(Instruction::Trunc, Start, TruncType); | ||||||
2399 | } | ||||||
2400 | |||||||
2401 | Value *Zero = getSignedIntOrFpConstant(Start->getType(), 0); | ||||||
2402 | Value *SplatStart = Builder.CreateVectorSplat(State.VF, Start); | ||||||
2403 | Value *SteppedStart = getStepVector( | ||||||
2404 | SplatStart, Zero, Step, II.getInductionOpcode(), State.VF, State.Builder); | ||||||
2405 | |||||||
2406 | // We create vector phi nodes for both integer and floating-point induction | ||||||
2407 | // variables. Here, we determine the kind of arithmetic we will perform. | ||||||
2408 | Instruction::BinaryOps AddOp; | ||||||
2409 | Instruction::BinaryOps MulOp; | ||||||
2410 | if (Step->getType()->isIntegerTy()) { | ||||||
2411 | AddOp = Instruction::Add; | ||||||
2412 | MulOp = Instruction::Mul; | ||||||
2413 | } else { | ||||||
2414 | AddOp = II.getInductionOpcode(); | ||||||
2415 | MulOp = Instruction::FMul; | ||||||
2416 | } | ||||||
2417 | |||||||
2418 | // Multiply the vectorization factor by the step using integer or | ||||||
2419 | // floating-point arithmetic as appropriate. | ||||||
2420 | Type *StepType = Step->getType(); | ||||||
2421 | Value *RuntimeVF; | ||||||
2422 | if (Step->getType()->isFloatingPointTy()) | ||||||
2423 | RuntimeVF = getRuntimeVFAsFloat(Builder, StepType, State.VF); | ||||||
2424 | else | ||||||
2425 | RuntimeVF = getRuntimeVF(Builder, StepType, State.VF); | ||||||
2426 | Value *Mul = Builder.CreateBinOp(MulOp, Step, RuntimeVF); | ||||||
2427 | |||||||
2428 | // Create a vector splat to use in the induction update. | ||||||
2429 | // | ||||||
2430 | // FIXME: If the step is non-constant, we create the vector splat with | ||||||
2431 | // IRBuilder. IRBuilder can constant-fold the multiply, but it doesn't | ||||||
2432 | // handle a constant vector splat. | ||||||
2433 | Value *SplatVF = isa<Constant>(Mul) | ||||||
2434 | ? ConstantVector::getSplat(State.VF, cast<Constant>(Mul)) | ||||||
2435 | : Builder.CreateVectorSplat(State.VF, Mul); | ||||||
2436 | Builder.restoreIP(CurrIP); | ||||||
2437 | |||||||
2438 | // We may need to add the step a number of times, depending on the unroll | ||||||
2439 | // factor. The last of those goes into the PHI. | ||||||
2440 | PHINode *VecInd = PHINode::Create(SteppedStart->getType(), 2, "vec.ind", | ||||||
2441 | &*LoopVectorBody->getFirstInsertionPt()); | ||||||
2442 | VecInd->setDebugLoc(EntryVal->getDebugLoc()); | ||||||
2443 | Instruction *LastInduction = VecInd; | ||||||
2444 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
2445 | State.set(Def, LastInduction, Part); | ||||||
2446 | |||||||
2447 | if (isa<TruncInst>(EntryVal)) | ||||||
2448 | addMetadata(LastInduction, EntryVal); | ||||||
2449 | |||||||
2450 | LastInduction = cast<Instruction>( | ||||||
2451 | Builder.CreateBinOp(AddOp, LastInduction, SplatVF, "step.add")); | ||||||
2452 | LastInduction->setDebugLoc(EntryVal->getDebugLoc()); | ||||||
2453 | } | ||||||
2454 | |||||||
2455 | // Move the last step to the end of the latch block. This ensures consistent | ||||||
2456 | // placement of all induction updates. | ||||||
2457 | auto *LoopVectorLatch = LI->getLoopFor(LoopVectorBody)->getLoopLatch(); | ||||||
2458 | auto *Br = cast<BranchInst>(LoopVectorLatch->getTerminator()); | ||||||
2459 | LastInduction->moveBefore(Br); | ||||||
2460 | LastInduction->setName("vec.ind.next"); | ||||||
2461 | |||||||
2462 | VecInd->addIncoming(SteppedStart, LoopVectorPreHeader); | ||||||
2463 | VecInd->addIncoming(LastInduction, LoopVectorLatch); | ||||||
2464 | } | ||||||
2465 | |||||||
2466 | bool InnerLoopVectorizer::shouldScalarizeInstruction(Instruction *I) const { | ||||||
2467 | return Cost->isScalarAfterVectorization(I, VF) || | ||||||
2468 | Cost->isProfitableToScalarize(I, VF); | ||||||
2469 | } | ||||||
2470 | |||||||
2471 | bool InnerLoopVectorizer::needsScalarInduction(Instruction *IV) const { | ||||||
2472 | if (shouldScalarizeInstruction(IV)) | ||||||
2473 | return true; | ||||||
2474 | auto isScalarInst = [&](User *U) -> bool { | ||||||
2475 | auto *I = cast<Instruction>(U); | ||||||
2476 | return (OrigLoop->contains(I) && shouldScalarizeInstruction(I)); | ||||||
2477 | }; | ||||||
2478 | return llvm::any_of(IV->users(), isScalarInst); | ||||||
2479 | } | ||||||
2480 | |||||||
2481 | /// Returns true if \p ID starts at 0 and has a step of 1. | ||||||
2482 | static bool isCanonicalID(const InductionDescriptor &ID) { | ||||||
2483 | if (!ID.getConstIntStepValue() || !ID.getConstIntStepValue()->isOne()) | ||||||
2484 | return false; | ||||||
2485 | auto *StartC = dyn_cast<ConstantInt>(ID.getStartValue()); | ||||||
2486 | return StartC && StartC->isZero(); | ||||||
2487 | } | ||||||
2488 | |||||||
2489 | void InnerLoopVectorizer::widenIntOrFpInduction( | ||||||
2490 | PHINode *IV, const InductionDescriptor &ID, Value *Start, TruncInst *Trunc, | ||||||
2491 | VPValue *Def, VPTransformState &State, Value *CanonicalIV) { | ||||||
2492 | IRBuilder<> &Builder = State.Builder; | ||||||
2493 | 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", 2493, __extension__ __PRETTY_FUNCTION__)); | ||||||
2494 | assert(!State.VF.isZero() && "VF must be non-zero")(static_cast <bool> (!State.VF.isZero() && "VF must be non-zero" ) ? void (0) : __assert_fail ("!State.VF.isZero() && \"VF must be non-zero\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2494, __extension__ __PRETTY_FUNCTION__)); | ||||||
2495 | |||||||
2496 | // The value from the original loop to which we are mapping the new induction | ||||||
2497 | // variable. | ||||||
2498 | Instruction *EntryVal = Trunc ? cast<Instruction>(Trunc) : IV; | ||||||
2499 | |||||||
2500 | auto &DL = EntryVal->getModule()->getDataLayout(); | ||||||
2501 | |||||||
2502 | // Generate code for the induction step. Note that induction steps are | ||||||
2503 | // required to be loop-invariant | ||||||
2504 | auto CreateStepValue = [&](const SCEV *Step) -> Value * { | ||||||
2505 | assert(PSE.getSE()->isLoopInvariant(Step, OrigLoop) &&(static_cast <bool> (PSE.getSE()->isLoopInvariant(Step , OrigLoop) && "Induction step should be loop invariant" ) ? void (0) : __assert_fail ("PSE.getSE()->isLoopInvariant(Step, OrigLoop) && \"Induction step should be loop invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2506, __extension__ __PRETTY_FUNCTION__)) | ||||||
2506 | "Induction step should be loop invariant")(static_cast <bool> (PSE.getSE()->isLoopInvariant(Step , OrigLoop) && "Induction step should be loop invariant" ) ? void (0) : __assert_fail ("PSE.getSE()->isLoopInvariant(Step, OrigLoop) && \"Induction step should be loop invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2506, __extension__ __PRETTY_FUNCTION__)); | ||||||
2507 | if (PSE.getSE()->isSCEVable(IV->getType())) { | ||||||
2508 | SCEVExpander Exp(*PSE.getSE(), DL, "induction"); | ||||||
2509 | return Exp.expandCodeFor(Step, Step->getType(), | ||||||
2510 | State.CFG.VectorPreHeader->getTerminator()); | ||||||
2511 | } | ||||||
2512 | return cast<SCEVUnknown>(Step)->getValue(); | ||||||
2513 | }; | ||||||
2514 | |||||||
2515 | // The scalar value to broadcast. This is derived from the canonical | ||||||
2516 | // induction variable. If a truncation type is given, truncate the canonical | ||||||
2517 | // induction variable and step. Otherwise, derive these values from the | ||||||
2518 | // induction descriptor. | ||||||
2519 | auto CreateScalarIV = [&](Value *&Step) -> Value * { | ||||||
2520 | Value *ScalarIV = CanonicalIV; | ||||||
2521 | Type *NeededType = IV->getType(); | ||||||
2522 | if (!isCanonicalID(ID) || ScalarIV->getType() != NeededType) { | ||||||
2523 | ScalarIV = | ||||||
2524 | NeededType->isIntegerTy() | ||||||
2525 | ? Builder.CreateSExtOrTrunc(ScalarIV, NeededType) | ||||||
2526 | : Builder.CreateCast(Instruction::SIToFP, ScalarIV, NeededType); | ||||||
2527 | ScalarIV = emitTransformedIndex(Builder, ScalarIV, PSE.getSE(), DL, ID, | ||||||
2528 | State.CFG.PrevBB); | ||||||
2529 | ScalarIV->setName("offset.idx"); | ||||||
2530 | } | ||||||
2531 | if (Trunc) { | ||||||
2532 | auto *TruncType = cast<IntegerType>(Trunc->getType()); | ||||||
2533 | 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", 2534, __extension__ __PRETTY_FUNCTION__)) | ||||||
2534 | "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", 2534, __extension__ __PRETTY_FUNCTION__)); | ||||||
2535 | ScalarIV = Builder.CreateTrunc(ScalarIV, TruncType); | ||||||
2536 | Step = Builder.CreateTrunc(Step, TruncType); | ||||||
2537 | } | ||||||
2538 | return ScalarIV; | ||||||
2539 | }; | ||||||
2540 | |||||||
2541 | // Create the vector values from the scalar IV, in the absence of creating a | ||||||
2542 | // vector IV. | ||||||
2543 | auto CreateSplatIV = [&](Value *ScalarIV, Value *Step) { | ||||||
2544 | Value *Broadcasted = getBroadcastInstrs(ScalarIV); | ||||||
2545 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
2546 | Value *StartIdx; | ||||||
2547 | if (Step->getType()->isFloatingPointTy()) | ||||||
2548 | StartIdx = | ||||||
2549 | getRuntimeVFAsFloat(Builder, Step->getType(), State.VF * Part); | ||||||
2550 | else | ||||||
2551 | StartIdx = getRuntimeVF(Builder, Step->getType(), State.VF * Part); | ||||||
2552 | |||||||
2553 | Value *EntryPart = | ||||||
2554 | getStepVector(Broadcasted, StartIdx, Step, ID.getInductionOpcode(), | ||||||
2555 | State.VF, State.Builder); | ||||||
2556 | State.set(Def, EntryPart, Part); | ||||||
2557 | if (Trunc) | ||||||
2558 | addMetadata(EntryPart, Trunc); | ||||||
2559 | } | ||||||
2560 | }; | ||||||
2561 | |||||||
2562 | // Fast-math-flags propagate from the original induction instruction. | ||||||
2563 | IRBuilder<>::FastMathFlagGuard FMFG(Builder); | ||||||
2564 | if (ID.getInductionBinOp() && isa<FPMathOperator>(ID.getInductionBinOp())) | ||||||
2565 | Builder.setFastMathFlags(ID.getInductionBinOp()->getFastMathFlags()); | ||||||
2566 | |||||||
2567 | // Now do the actual transformations, and start with creating the step value. | ||||||
2568 | Value *Step = CreateStepValue(ID.getStep()); | ||||||
2569 | if (State.VF.isScalar()) { | ||||||
2570 | Value *ScalarIV = CreateScalarIV(Step); | ||||||
2571 | Type *ScalarTy = IntegerType::get(ScalarIV->getContext(), | ||||||
2572 | Step->getType()->getScalarSizeInBits()); | ||||||
2573 | |||||||
2574 | Instruction::BinaryOps IncOp = ID.getInductionOpcode(); | ||||||
2575 | if (IncOp == Instruction::BinaryOpsEnd) | ||||||
2576 | IncOp = Instruction::Add; | ||||||
2577 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
2578 | Value *StartIdx = ConstantInt::get(ScalarTy, Part); | ||||||
2579 | Instruction::BinaryOps MulOp = Instruction::Mul; | ||||||
2580 | if (Step->getType()->isFloatingPointTy()) { | ||||||
2581 | StartIdx = Builder.CreateUIToFP(StartIdx, Step->getType()); | ||||||
2582 | MulOp = Instruction::FMul; | ||||||
2583 | } | ||||||
2584 | |||||||
2585 | Value *Mul = Builder.CreateBinOp(MulOp, StartIdx, Step); | ||||||
2586 | Value *EntryPart = Builder.CreateBinOp(IncOp, ScalarIV, Mul, "induction"); | ||||||
2587 | State.set(Def, EntryPart, Part); | ||||||
2588 | if (Trunc) { | ||||||
2589 | assert(!Step->getType()->isFloatingPointTy() &&(static_cast <bool> (!Step->getType()->isFloatingPointTy () && "fp inductions shouldn't be truncated") ? void ( 0) : __assert_fail ("!Step->getType()->isFloatingPointTy() && \"fp inductions shouldn't be truncated\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2590, __extension__ __PRETTY_FUNCTION__)) | ||||||
2590 | "fp inductions shouldn't be truncated")(static_cast <bool> (!Step->getType()->isFloatingPointTy () && "fp inductions shouldn't be truncated") ? void ( 0) : __assert_fail ("!Step->getType()->isFloatingPointTy() && \"fp inductions shouldn't be truncated\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2590, __extension__ __PRETTY_FUNCTION__)); | ||||||
2591 | addMetadata(EntryPart, Trunc); | ||||||
2592 | } | ||||||
2593 | } | ||||||
2594 | return; | ||||||
2595 | } | ||||||
2596 | |||||||
2597 | // Determine if we want a scalar version of the induction variable. This is | ||||||
2598 | // true if the induction variable itself is not widened, or if it has at | ||||||
2599 | // least one user in the loop that is not widened. | ||||||
2600 | auto NeedsScalarIV = needsScalarInduction(EntryVal); | ||||||
2601 | if (!NeedsScalarIV) { | ||||||
2602 | createVectorIntOrFpInductionPHI(ID, Step, Start, EntryVal, Def, State); | ||||||
2603 | return; | ||||||
2604 | } | ||||||
2605 | |||||||
2606 | // Try to create a new independent vector induction variable. If we can't | ||||||
2607 | // create the phi node, we will splat the scalar induction variable in each | ||||||
2608 | // loop iteration. | ||||||
2609 | if (!shouldScalarizeInstruction(EntryVal)) { | ||||||
2610 | createVectorIntOrFpInductionPHI(ID, Step, Start, EntryVal, Def, State); | ||||||
2611 | Value *ScalarIV = CreateScalarIV(Step); | ||||||
2612 | // Create scalar steps that can be used by instructions we will later | ||||||
2613 | // scalarize. Note that the addition of the scalar steps will not increase | ||||||
2614 | // the number of instructions in the loop in the common case prior to | ||||||
2615 | // InstCombine. We will be trading one vector extract for each scalar step. | ||||||
2616 | buildScalarSteps(ScalarIV, Step, EntryVal, ID, Def, State); | ||||||
2617 | return; | ||||||
2618 | } | ||||||
2619 | |||||||
2620 | // All IV users are scalar instructions, so only emit a scalar IV, not a | ||||||
2621 | // vectorised IV. Except when we tail-fold, then the splat IV feeds the | ||||||
2622 | // predicate used by the masked loads/stores. | ||||||
2623 | Value *ScalarIV = CreateScalarIV(Step); | ||||||
2624 | if (!Cost->isScalarEpilogueAllowed()) | ||||||
2625 | CreateSplatIV(ScalarIV, Step); | ||||||
2626 | buildScalarSteps(ScalarIV, Step, EntryVal, ID, Def, State); | ||||||
2627 | } | ||||||
2628 | |||||||
2629 | void InnerLoopVectorizer::buildScalarSteps(Value *ScalarIV, Value *Step, | ||||||
2630 | Instruction *EntryVal, | ||||||
2631 | const InductionDescriptor &ID, | ||||||
2632 | VPValue *Def, | ||||||
2633 | VPTransformState &State) { | ||||||
2634 | IRBuilder<> &Builder = State.Builder; | ||||||
2635 | // We shouldn't have to build scalar steps if we aren't vectorizing. | ||||||
2636 | assert(State.VF.isVector() && "VF should be greater than one")(static_cast <bool> (State.VF.isVector() && "VF should be greater than one" ) ? void (0) : __assert_fail ("State.VF.isVector() && \"VF should be greater than one\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2636, __extension__ __PRETTY_FUNCTION__)); | ||||||
2637 | // Get the value type and ensure it and the step have the same integer type. | ||||||
2638 | Type *ScalarIVTy = ScalarIV->getType()->getScalarType(); | ||||||
2639 | assert(ScalarIVTy == Step->getType() &&(static_cast <bool> (ScalarIVTy == Step->getType() && "Val and Step should have the same type") ? void (0) : __assert_fail ("ScalarIVTy == Step->getType() && \"Val and Step should have the same type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2640, __extension__ __PRETTY_FUNCTION__)) | ||||||
2640 | "Val and Step should have the same type")(static_cast <bool> (ScalarIVTy == Step->getType() && "Val and Step should have the same type") ? void (0) : __assert_fail ("ScalarIVTy == Step->getType() && \"Val and Step should have the same type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 2640, __extension__ __PRETTY_FUNCTION__)); | ||||||
2641 | |||||||
2642 | // We build scalar steps for both integer and floating-point induction | ||||||
2643 | // variables. Here, we determine the kind of arithmetic we will perform. | ||||||
2644 | Instruction::BinaryOps AddOp; | ||||||
2645 | Instruction::BinaryOps MulOp; | ||||||
2646 | if (ScalarIVTy->isIntegerTy()) { | ||||||
2647 | AddOp = Instruction::Add; | ||||||
2648 | MulOp = Instruction::Mul; | ||||||
2649 | } else { | ||||||
2650 | AddOp = ID.getInductionOpcode(); | ||||||
2651 | MulOp = Instruction::FMul; | ||||||
2652 | } | ||||||
2653 | |||||||
2654 | // Determine the number of scalars we need to generate for each unroll | ||||||
2655 | // iteration. If EntryVal is uniform, we only need to generate the first | ||||||
2656 | // lane. Otherwise, we generate all VF values. | ||||||
2657 | bool IsUniform = | ||||||
2658 | Cost->isUniformAfterVectorization(cast<Instruction>(EntryVal), State.VF); | ||||||
2659 | unsigned Lanes = IsUniform ? 1 : State.VF.getKnownMinValue(); | ||||||
2660 | // Compute the scalar steps and save the results in State. | ||||||
2661 | Type *IntStepTy = IntegerType::get(ScalarIVTy->getContext(), | ||||||
2662 | ScalarIVTy->getScalarSizeInBits()); | ||||||
2663 | Type *VecIVTy = nullptr; | ||||||
2664 | Value *UnitStepVec = nullptr, *SplatStep = nullptr, *SplatIV = nullptr; | ||||||
2665 | if (!IsUniform && State.VF.isScalable()) { | ||||||
2666 | VecIVTy = VectorType::get(ScalarIVTy, State.VF); | ||||||
2667 | UnitStepVec = | ||||||
2668 | Builder.CreateStepVector(VectorType::get(IntStepTy, State.VF)); | ||||||
2669 | SplatStep = Builder.CreateVectorSplat(State.VF, Step); | ||||||
2670 | SplatIV = Builder.CreateVectorSplat(State.VF, ScalarIV); | ||||||
2671 | } | ||||||
2672 | |||||||
2673 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
2674 | Value *StartIdx0 = createStepForVF(Builder, IntStepTy, State.VF, Part); | ||||||
2675 | |||||||
2676 | if (!IsUniform && State.VF.isScalable()) { | ||||||
2677 | auto *SplatStartIdx = Builder.CreateVectorSplat(State.VF, StartIdx0); | ||||||
2678 | auto *InitVec = Builder.CreateAdd(SplatStartIdx, UnitStepVec); | ||||||
2679 | if (ScalarIVTy->isFloatingPointTy()) | ||||||
2680 | InitVec = Builder.CreateSIToFP(InitVec, VecIVTy); | ||||||
2681 | auto *Mul = Builder.CreateBinOp(MulOp, InitVec, SplatStep); | ||||||
2682 | auto *Add = Builder.CreateBinOp(AddOp, SplatIV, Mul); | ||||||
2683 | State.set(Def, Add, Part); | ||||||
2684 | // It's useful to record the lane values too for the known minimum number | ||||||
2685 | // of elements so we do those below. This improves the code quality when | ||||||
2686 | // trying to extract the first element, for example. | ||||||
2687 | } | ||||||
2688 | |||||||
2689 | if (ScalarIVTy->isFloatingPointTy()) | ||||||
2690 | StartIdx0 = Builder.CreateSIToFP(StartIdx0, ScalarIVTy); | ||||||
2691 | |||||||
2692 | for (unsigned Lane = 0; Lane < Lanes; ++Lane) { | ||||||
2693 | Value *StartIdx = Builder.CreateBinOp( | ||||||
2694 | AddOp, StartIdx0, getSignedIntOrFpConstant(ScalarIVTy, Lane)); | ||||||
2695 | // The step returned by `createStepForVF` is a runtime-evaluated value | ||||||
2696 | // when VF is scalable. Otherwise, it should be folded into a Constant. | ||||||
2697 | 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", 2699, __extension__ __PRETTY_FUNCTION__)) | ||||||
2698 | "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", 2699, __extension__ __PRETTY_FUNCTION__)) | ||||||
2699 | "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", 2699, __extension__ __PRETTY_FUNCTION__)); | ||||||
2700 | auto *Mul = Builder.CreateBinOp(MulOp, StartIdx, Step); | ||||||
2701 | auto *Add = Builder.CreateBinOp(AddOp, ScalarIV, Mul); | ||||||
2702 | State.set(Def, Add, VPIteration(Part, Lane)); | ||||||
2703 | } | ||||||
2704 | } | ||||||
2705 | } | ||||||
2706 | |||||||
2707 | void InnerLoopVectorizer::packScalarIntoVectorValue(VPValue *Def, | ||||||
2708 | const VPIteration &Instance, | ||||||
2709 | VPTransformState &State) { | ||||||
2710 | Value *ScalarInst = State.get(Def, Instance); | ||||||
2711 | Value *VectorValue = State.get(Def, Instance.Part); | ||||||
2712 | VectorValue = Builder.CreateInsertElement( | ||||||
2713 | VectorValue, ScalarInst, | ||||||
2714 | Instance.Lane.getAsRuntimeExpr(State.Builder, VF)); | ||||||
2715 | State.set(Def, VectorValue, Instance.Part); | ||||||
2716 | } | ||||||
2717 | |||||||
2718 | // Return whether we allow using masked interleave-groups (for dealing with | ||||||
2719 | // strided loads/stores that reside in predicated blocks, or for dealing | ||||||
2720 | // with gaps). | ||||||
2721 | static bool useMaskedInterleavedAccesses(const TargetTransformInfo &TTI) { | ||||||
2722 | // If an override option has been passed in for interleaved accesses, use it. | ||||||
2723 | if (EnableMaskedInterleavedMemAccesses.getNumOccurrences() > 0) | ||||||
2724 | return EnableMaskedInterleavedMemAccesses; | ||||||
2725 | |||||||
2726 | return TTI.enableMaskedInterleavedAccessVectorization(); | ||||||
2727 | } | ||||||
2728 | |||||||
2729 | // Try to vectorize the interleave group that \p Instr belongs to. | ||||||
2730 | // | ||||||
2731 | // E.g. Translate following interleaved load group (factor = 3): | ||||||
2732 | // for (i = 0; i < N; i+=3) { | ||||||
2733 | // R = Pic[i]; // Member of index 0 | ||||||
2734 | // G = Pic[i+1]; // Member of index 1 | ||||||
2735 | // B = Pic[i+2]; // Member of index 2 | ||||||
2736 | // ... // do something to R, G, B | ||||||
2737 | // } | ||||||
2738 | // To: | ||||||
2739 | // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B | ||||||
2740 | // %R.vec = shuffle %wide.vec, poison, <0, 3, 6, 9> ; R elements | ||||||
2741 | // %G.vec = shuffle %wide.vec, poison, <1, 4, 7, 10> ; G elements | ||||||
2742 | // %B.vec = shuffle %wide.vec, poison, <2, 5, 8, 11> ; B elements | ||||||
2743 | // | ||||||
2744 | // Or translate following interleaved store group (factor = 3): | ||||||
2745 | // for (i = 0; i < N; i+=3) { | ||||||
2746 | // ... do something to R, G, B | ||||||
2747 | // Pic[i] = R; // Member of index 0 | ||||||
2748 | // Pic[i+1] = G; // Member of index 1 | ||||||
2749 | // Pic[i+2] = B; // Member of index 2 | ||||||
2750 | // } | ||||||
2751 | // To: | ||||||
2752 | // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7> | ||||||
2753 | // %B_U.vec = shuffle %B.vec, poison, <0, 1, 2, 3, u, u, u, u> | ||||||
2754 | // %interleaved.vec = shuffle %R_G.vec, %B_U.vec, | ||||||
2755 | // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements | ||||||
2756 | // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B | ||||||
2757 | void InnerLoopVectorizer::vectorizeInterleaveGroup( | ||||||
2758 | const InterleaveGroup<Instruction> *Group, ArrayRef<VPValue *> VPDefs, | ||||||
2759 | VPTransformState &State, VPValue *Addr, ArrayRef<VPValue *> StoredValues, | ||||||
2760 | VPValue *BlockInMask) { | ||||||
2761 | Instruction *Instr = Group->getInsertPos(); | ||||||
2762 | const DataLayout &DL = Instr->getModule()->getDataLayout(); | ||||||
2763 | |||||||
2764 | // Prepare for the vector type of the interleaved load/store. | ||||||
2765 | Type *ScalarTy = getLoadStoreType(Instr); | ||||||
2766 | unsigned InterleaveFactor = Group->getFactor(); | ||||||
2767 | 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", 2767, __extension__ __PRETTY_FUNCTION__)); | ||||||
2768 | auto *VecTy = VectorType::get(ScalarTy, VF * InterleaveFactor); | ||||||
2769 | |||||||
2770 | // Prepare for the new pointers. | ||||||
2771 | SmallVector<Value *, 2> AddrParts; | ||||||
2772 | unsigned Index = Group->getIndex(Instr); | ||||||
2773 | |||||||
2774 | // TODO: extend the masked interleaved-group support to reversed access. | ||||||
2775 | 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", 2776, __extension__ __PRETTY_FUNCTION__)) | ||||||
2776 | "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", 2776, __extension__ __PRETTY_FUNCTION__)); | ||||||
2777 | |||||||
2778 | // If the group is reverse, adjust the index to refer to the last vector lane | ||||||
2779 | // instead of the first. We adjust the index from the first vector lane, | ||||||
2780 | // rather than directly getting the pointer for lane VF - 1, because the | ||||||
2781 | // pointer operand of the interleaved access is supposed to be uniform. For | ||||||
2782 | // uniform instructions, we're only required to generate a value for the | ||||||
2783 | // first vector lane in each unroll iteration. | ||||||
2784 | if (Group->isReverse()) | ||||||
2785 | Index += (VF.getKnownMinValue() - 1) * Group->getFactor(); | ||||||
2786 | |||||||
2787 | for (unsigned Part = 0; Part < UF; Part++) { | ||||||
2788 | Value *AddrPart = State.get(Addr, VPIteration(Part, 0)); | ||||||
2789 | setDebugLocFromInst(AddrPart); | ||||||
2790 | |||||||
2791 | // Notice current instruction could be any index. Need to adjust the address | ||||||
2792 | // to the member of index 0. | ||||||
2793 | // | ||||||
2794 | // E.g. a = A[i+1]; // Member of index 1 (Current instruction) | ||||||
2795 | // b = A[i]; // Member of index 0 | ||||||
2796 | // Current pointer is pointed to A[i+1], adjust it to A[i]. | ||||||
2797 | // | ||||||
2798 | // E.g. A[i+1] = a; // Member of index 1 | ||||||
2799 | // A[i] = b; // Member of index 0 | ||||||
2800 | // A[i+2] = c; // Member of index 2 (Current instruction) | ||||||
2801 | // Current pointer is pointed to A[i+2], adjust it to A[i]. | ||||||
2802 | |||||||
2803 | bool InBounds = false; | ||||||
2804 | if (auto *gep = dyn_cast<GetElementPtrInst>(AddrPart->stripPointerCasts())) | ||||||
2805 | InBounds = gep->isInBounds(); | ||||||
2806 | AddrPart = Builder.CreateGEP(ScalarTy, AddrPart, Builder.getInt32(-Index)); | ||||||
2807 | cast<GetElementPtrInst>(AddrPart)->setIsInBounds(InBounds); | ||||||
2808 | |||||||
2809 | // Cast to the vector pointer type. | ||||||
2810 | unsigned AddressSpace = AddrPart->getType()->getPointerAddressSpace(); | ||||||
2811 | Type *PtrTy = VecTy->getPointerTo(AddressSpace); | ||||||
2812 | AddrParts.push_back(Builder.CreateBitCast(AddrPart, PtrTy)); | ||||||
2813 | } | ||||||
2814 | |||||||
2815 | setDebugLocFromInst(Instr); | ||||||
2816 | Value *PoisonVec = PoisonValue::get(VecTy); | ||||||
2817 | |||||||
2818 | Value *MaskForGaps = nullptr; | ||||||
2819 | if (Group->requiresScalarEpilogue() && !Cost->isScalarEpilogueAllowed()) { | ||||||
2820 | MaskForGaps = createBitMaskForGaps(Builder, VF.getKnownMinValue(), *Group); | ||||||
2821 | 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", 2821, __extension__ __PRETTY_FUNCTION__)); | ||||||
2822 | } | ||||||
2823 | |||||||
2824 | // Vectorize the interleaved load group. | ||||||
2825 | if (isa<LoadInst>(Instr)) { | ||||||
2826 | // For each unroll part, create a wide load for the group. | ||||||
2827 | SmallVector<Value *, 2> NewLoads; | ||||||
2828 | for (unsigned Part = 0; Part < UF; Part++) { | ||||||
2829 | Instruction *NewLoad; | ||||||
2830 | if (BlockInMask || MaskForGaps) { | ||||||
2831 | 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", 2832, __extension__ __PRETTY_FUNCTION__)) | ||||||
2832 | "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", 2832, __extension__ __PRETTY_FUNCTION__)); | ||||||
2833 | Value *GroupMask = MaskForGaps; | ||||||
2834 | if (BlockInMask) { | ||||||
2835 | Value *BlockInMaskPart = State.get(BlockInMask, Part); | ||||||
2836 | Value *ShuffledMask = Builder.CreateShuffleVector( | ||||||
2837 | BlockInMaskPart, | ||||||
2838 | createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()), | ||||||
2839 | "interleaved.mask"); | ||||||
2840 | GroupMask = MaskForGaps | ||||||
2841 | ? Builder.CreateBinOp(Instruction::And, ShuffledMask, | ||||||
2842 | MaskForGaps) | ||||||
2843 | : ShuffledMask; | ||||||
2844 | } | ||||||
2845 | NewLoad = | ||||||
2846 | Builder.CreateMaskedLoad(VecTy, AddrParts[Part], Group->getAlign(), | ||||||
2847 | GroupMask, PoisonVec, "wide.masked.vec"); | ||||||
2848 | } | ||||||
2849 | else | ||||||
2850 | NewLoad = Builder.CreateAlignedLoad(VecTy, AddrParts[Part], | ||||||
2851 | Group->getAlign(), "wide.vec"); | ||||||
2852 | Group->addMetadata(NewLoad); | ||||||
2853 | NewLoads.push_back(NewLoad); | ||||||
2854 | } | ||||||
2855 | |||||||
2856 | // For each member in the group, shuffle out the appropriate data from the | ||||||
2857 | // wide loads. | ||||||
2858 | unsigned J = 0; | ||||||
2859 | for (unsigned I = 0; I < InterleaveFactor; ++I) { | ||||||
2860 | Instruction *Member = Group->getMember(I); | ||||||
2861 | |||||||
2862 | // Skip the gaps in the group. | ||||||
2863 | if (!Member) | ||||||
2864 | continue; | ||||||
2865 | |||||||
2866 | auto StrideMask = | ||||||
2867 | createStrideMask(I, InterleaveFactor, VF.getKnownMinValue()); | ||||||
2868 | for (unsigned Part = 0; Part < UF; Part++) { | ||||||
2869 | Value *StridedVec = Builder.CreateShuffleVector( | ||||||
2870 | NewLoads[Part], StrideMask, "strided.vec"); | ||||||
2871 | |||||||
2872 | // If this member has different type, cast the result type. | ||||||
2873 | if (Member->getType() != ScalarTy) { | ||||||
2874 | 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", 2874, __extension__ __PRETTY_FUNCTION__)); | ||||||
2875 | VectorType *OtherVTy = VectorType::get(Member->getType(), VF); | ||||||
2876 | StridedVec = createBitOrPointerCast(StridedVec, OtherVTy, DL); | ||||||
2877 | } | ||||||
2878 | |||||||
2879 | if (Group->isReverse()) | ||||||
2880 | StridedVec = Builder.CreateVectorReverse(StridedVec, "reverse"); | ||||||
2881 | |||||||
2882 | State.set(VPDefs[J], StridedVec, Part); | ||||||
2883 | } | ||||||
2884 | ++J; | ||||||
2885 | } | ||||||
2886 | return; | ||||||
2887 | } | ||||||
2888 | |||||||
2889 | // The sub vector type for current instruction. | ||||||
2890 | auto *SubVT = VectorType::get(ScalarTy, VF); | ||||||
2891 | |||||||
2892 | // Vectorize the interleaved store group. | ||||||
2893 | MaskForGaps = createBitMaskForGaps(Builder, VF.getKnownMinValue(), *Group); | ||||||
2894 | 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", 2895, __extension__ __PRETTY_FUNCTION__)) | ||||||
2895 | "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", 2895, __extension__ __PRETTY_FUNCTION__)); | ||||||
2896 | 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", 2897, __extension__ __PRETTY_FUNCTION__)) | ||||||
2897 | "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", 2897, __extension__ __PRETTY_FUNCTION__)); | ||||||
2898 | for (unsigned Part = 0; Part < UF; Part++) { | ||||||
2899 | // Collect the stored vector from each member. | ||||||
2900 | SmallVector<Value *, 4> StoredVecs; | ||||||
2901 | for (unsigned i = 0; i < InterleaveFactor; i++) { | ||||||
2902 | 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", 2903, __extension__ __PRETTY_FUNCTION__)) | ||||||
2903 | "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", 2903, __extension__ __PRETTY_FUNCTION__)); | ||||||
2904 | Instruction *Member = Group->getMember(i); | ||||||
2905 | |||||||
2906 | // Skip the gaps in the group. | ||||||
2907 | if (!Member) { | ||||||
2908 | Value *Undef = PoisonValue::get(SubVT); | ||||||
2909 | StoredVecs.push_back(Undef); | ||||||
2910 | continue; | ||||||
2911 | } | ||||||
2912 | |||||||
2913 | Value *StoredVec = State.get(StoredValues[i], Part); | ||||||
2914 | |||||||
2915 | if (Group->isReverse()) | ||||||
2916 | StoredVec = Builder.CreateVectorReverse(StoredVec, "reverse"); | ||||||
2917 | |||||||
2918 | // If this member has different type, cast it to a unified type. | ||||||
2919 | |||||||
2920 | if (StoredVec->getType() != SubVT) | ||||||
2921 | StoredVec = createBitOrPointerCast(StoredVec, SubVT, DL); | ||||||
2922 | |||||||
2923 | StoredVecs.push_back(StoredVec); | ||||||
2924 | } | ||||||
2925 | |||||||
2926 | // Concatenate all vectors into a wide vector. | ||||||
2927 | Value *WideVec = concatenateVectors(Builder, StoredVecs); | ||||||
2928 | |||||||
2929 | // Interleave the elements in the wide vector. | ||||||
2930 | Value *IVec = Builder.CreateShuffleVector( | ||||||
2931 | WideVec, createInterleaveMask(VF.getKnownMinValue(), InterleaveFactor), | ||||||
2932 | "interleaved.vec"); | ||||||
2933 | |||||||
2934 | Instruction *NewStoreInstr; | ||||||
2935 | if (BlockInMask || MaskForGaps) { | ||||||
2936 | Value *GroupMask = MaskForGaps; | ||||||
2937 | if (BlockInMask) { | ||||||
2938 | Value *BlockInMaskPart = State.get(BlockInMask, Part); | ||||||
2939 | Value *ShuffledMask = Builder.CreateShuffleVector( | ||||||
2940 | BlockInMaskPart, | ||||||
2941 | createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()), | ||||||
2942 | "interleaved.mask"); | ||||||
2943 | GroupMask = MaskForGaps ? Builder.CreateBinOp(Instruction::And, | ||||||
2944 | ShuffledMask, MaskForGaps) | ||||||
2945 | : ShuffledMask; | ||||||
2946 | } | ||||||
2947 | NewStoreInstr = Builder.CreateMaskedStore(IVec, AddrParts[Part], | ||||||
2948 | Group->getAlign(), GroupMask); | ||||||
2949 | } else | ||||||
2950 | NewStoreInstr = | ||||||
2951 | Builder.CreateAlignedStore(IVec, AddrParts[Part], Group->getAlign()); | ||||||
2952 | |||||||
2953 | Group->addMetadata(NewStoreInstr); | ||||||
2954 | } | ||||||
2955 | } | ||||||
2956 | |||||||
2957 | void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, | ||||||
2958 | VPReplicateRecipe *RepRecipe, | ||||||
2959 | const VPIteration &Instance, | ||||||
2960 | bool IfPredicateInstr, | ||||||
2961 | VPTransformState &State) { | ||||||
2962 | 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", 2962, __extension__ __PRETTY_FUNCTION__)); | ||||||
2963 | |||||||
2964 | // llvm.experimental.noalias.scope.decl intrinsics must only be duplicated for | ||||||
2965 | // the first lane and part. | ||||||
2966 | if (isa<NoAliasScopeDeclInst>(Instr)) | ||||||
2967 | if (!Instance.isFirstIteration()) | ||||||
2968 | return; | ||||||
2969 | |||||||
2970 | setDebugLocFromInst(Instr); | ||||||
2971 | |||||||
2972 | // Does this instruction return a value ? | ||||||
2973 | bool IsVoidRetTy = Instr->getType()->isVoidTy(); | ||||||
2974 | |||||||
2975 | Instruction *Cloned = Instr->clone(); | ||||||
2976 | if (!IsVoidRetTy) | ||||||
2977 | Cloned->setName(Instr->getName() + ".cloned"); | ||||||
2978 | |||||||
2979 | // If the scalarized instruction contributes to the address computation of a | ||||||
2980 | // widen masked load/store which was in a basic block that needed predication | ||||||
2981 | // and is not predicated after vectorization, we can't propagate | ||||||
2982 | // poison-generating flags (nuw/nsw, exact, inbounds, etc.). The scalarized | ||||||
2983 | // instruction could feed a poison value to the base address of the widen | ||||||
2984 | // load/store. | ||||||
2985 | if (State.MayGeneratePoisonRecipes.contains(RepRecipe)) | ||||||
2986 | Cloned->dropPoisonGeneratingFlags(); | ||||||
2987 | |||||||
2988 | State.Builder.SetInsertPoint(Builder.GetInsertBlock(), | ||||||
2989 | Builder.GetInsertPoint()); | ||||||
2990 | // Replace the operands of the cloned instructions with their scalar | ||||||
2991 | // equivalents in the new loop. | ||||||
2992 | for (auto &I : enumerate(RepRecipe->operands())) { | ||||||
2993 | auto InputInstance = Instance; | ||||||
2994 | VPValue *Operand = I.value(); | ||||||
2995 | if (State.Plan->isUniformAfterVectorization(Operand)) | ||||||
2996 | InputInstance.Lane = VPLane::getFirstLane(); | ||||||
2997 | Cloned->setOperand(I.index(), State.get(Operand, InputInstance)); | ||||||
2998 | } | ||||||
2999 | addNewMetadata(Cloned, Instr); | ||||||
3000 | |||||||
3001 | // Place the cloned scalar in the new loop. | ||||||
3002 | Builder.Insert(Cloned); | ||||||
3003 | |||||||
3004 | State.set(RepRecipe, Cloned, Instance); | ||||||
3005 | |||||||
3006 | // If we just cloned a new assumption, add it the assumption cache. | ||||||
3007 | if (auto *II = dyn_cast<AssumeInst>(Cloned)) | ||||||
3008 | AC->registerAssumption(II); | ||||||
3009 | |||||||
3010 | // End if-block. | ||||||
3011 | if (IfPredicateInstr) | ||||||
3012 | PredicatedInstructions.push_back(Cloned); | ||||||
3013 | } | ||||||
3014 | |||||||
3015 | void InnerLoopVectorizer::createHeaderBranch(Loop *L) { | ||||||
3016 | BasicBlock *Header = L->getHeader(); | ||||||
3017 | assert(!L->getLoopLatch() && "loop should not have a latch at this point")(static_cast <bool> (!L->getLoopLatch() && "loop should not have a latch at this point" ) ? void (0) : __assert_fail ("!L->getLoopLatch() && \"loop should not have a latch at this point\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3017, __extension__ __PRETTY_FUNCTION__)); | ||||||
3018 | |||||||
3019 | IRBuilder<> B(Header->getTerminator()); | ||||||
3020 | Instruction *OldInst = | ||||||
3021 | getDebugLocFromInstOrOperands(Legal->getPrimaryInduction()); | ||||||
3022 | setDebugLocFromInst(OldInst, &B); | ||||||
3023 | |||||||
3024 | // Connect the header to the exit and header blocks and replace the old | ||||||
3025 | // terminator. | ||||||
3026 | B.CreateCondBr(B.getTrue(), L->getUniqueExitBlock(), Header); | ||||||
3027 | |||||||
3028 | // Now we have two terminators. Remove the old one from the block. | ||||||
3029 | Header->getTerminator()->eraseFromParent(); | ||||||
3030 | } | ||||||
3031 | |||||||
3032 | Value *InnerLoopVectorizer::getOrCreateTripCount(Loop *L) { | ||||||
3033 | if (TripCount) | ||||||
3034 | return TripCount; | ||||||
3035 | |||||||
3036 | assert(L && "Create Trip Count for null loop.")(static_cast <bool> (L && "Create Trip Count for null loop." ) ? void (0) : __assert_fail ("L && \"Create Trip Count for null loop.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3036, __extension__ __PRETTY_FUNCTION__)); | ||||||
3037 | IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); | ||||||
3038 | // Find the loop boundaries. | ||||||
3039 | ScalarEvolution *SE = PSE.getSE(); | ||||||
3040 | const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount(); | ||||||
3041 | assert(!isa<SCEVCouldNotCompute>(BackedgeTakenCount) &&(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", 3042, __extension__ __PRETTY_FUNCTION__)) | ||||||
3042 | "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", 3042, __extension__ __PRETTY_FUNCTION__)); | ||||||
3043 | |||||||
3044 | Type *IdxTy = Legal->getWidestInductionType(); | ||||||
3045 | 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", 3045, __extension__ __PRETTY_FUNCTION__)); | ||||||
3046 | |||||||
3047 | // The exit count might have the type of i64 while the phi is i32. This can | ||||||
3048 | // happen if we have an induction variable that is sign extended before the | ||||||
3049 | // compare. The only way that we get a backedge taken count is that the | ||||||
3050 | // induction variable was signed and as such will not overflow. In such a case | ||||||
3051 | // truncation is legal. | ||||||
3052 | if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) > | ||||||
3053 | IdxTy->getPrimitiveSizeInBits()) | ||||||
3054 | BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy); | ||||||
3055 | BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy); | ||||||
3056 | |||||||
3057 | // Get the total trip count from the count by adding 1. | ||||||
3058 | const SCEV *ExitCount = SE->getAddExpr( | ||||||
3059 | BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType())); | ||||||
3060 | |||||||
3061 | const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); | ||||||
3062 | |||||||
3063 | // Expand the trip count and place the new instructions in the preheader. | ||||||
3064 | // Notice that the pre-header does not change, only the loop body. | ||||||
3065 | SCEVExpander Exp(*SE, DL, "induction"); | ||||||
3066 | |||||||
3067 | // Count holds the overall loop count (N). | ||||||
3068 | TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(), | ||||||
3069 | L->getLoopPreheader()->getTerminator()); | ||||||
3070 | |||||||
3071 | if (TripCount->getType()->isPointerTy()) | ||||||
3072 | TripCount = | ||||||
3073 | CastInst::CreatePointerCast(TripCount, IdxTy, "exitcount.ptrcnt.to.int", | ||||||
3074 | L->getLoopPreheader()->getTerminator()); | ||||||
3075 | |||||||
3076 | return TripCount; | ||||||
3077 | } | ||||||
3078 | |||||||
3079 | Value *InnerLoopVectorizer::getOrCreateVectorTripCount(Loop *L) { | ||||||
3080 | if (VectorTripCount) | ||||||
3081 | return VectorTripCount; | ||||||
3082 | |||||||
3083 | Value *TC = getOrCreateTripCount(L); | ||||||
3084 | IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); | ||||||
3085 | |||||||
3086 | Type *Ty = TC->getType(); | ||||||
3087 | // This is where we can make the step a runtime constant. | ||||||
3088 | Value *Step = createStepForVF(Builder, Ty, VF, UF); | ||||||
3089 | |||||||
3090 | // If the tail is to be folded by masking, round the number of iterations N | ||||||
3091 | // up to a multiple of Step instead of rounding down. This is done by first | ||||||
3092 | // adding Step-1 and then rounding down. Note that it's ok if this addition | ||||||
3093 | // overflows: the vector induction variable will eventually wrap to zero given | ||||||
3094 | // that it starts at zero and its Step is a power of two; the loop will then | ||||||
3095 | // exit, with the last early-exit vector comparison also producing all-true. | ||||||
3096 | if (Cost->foldTailByMasking()) { | ||||||
3097 | 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", 3098, __extension__ __PRETTY_FUNCTION__)) | ||||||
3098 | "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", 3098, __extension__ __PRETTY_FUNCTION__)); | ||||||
3099 | Value *NumLanes = getRuntimeVF(Builder, Ty, VF * UF); | ||||||
3100 | TC = Builder.CreateAdd( | ||||||
3101 | TC, Builder.CreateSub(NumLanes, ConstantInt::get(Ty, 1)), "n.rnd.up"); | ||||||
3102 | } | ||||||
3103 | |||||||
3104 | // Now we need to generate the expression for the part of the loop that the | ||||||
3105 | // vectorized body will execute. This is equal to N - (N % Step) if scalar | ||||||
3106 | // iterations are not required for correctness, or N - Step, otherwise. Step | ||||||
3107 | // is equal to the vectorization factor (number of SIMD elements) times the | ||||||
3108 | // unroll factor (number of SIMD instructions). | ||||||
3109 | Value *R = Builder.CreateURem(TC, Step, "n.mod.vf"); | ||||||
3110 | |||||||
3111 | // There are cases where we *must* run at least one iteration in the remainder | ||||||
3112 | // loop. See the cost model for when this can happen. If the step evenly | ||||||
3113 | // divides the trip count, we set the remainder to be equal to the step. If | ||||||
3114 | // the step does not evenly divide the trip count, no adjustment is necessary | ||||||
3115 | // since there will already be scalar iterations. Note that the minimum | ||||||
3116 | // iterations check ensures that N >= Step. | ||||||
3117 | if (Cost->requiresScalarEpilogue(VF)) { | ||||||
3118 | auto *IsZero = Builder.CreateICmpEQ(R, ConstantInt::get(R->getType(), 0)); | ||||||
3119 | R = Builder.CreateSelect(IsZero, Step, R); | ||||||
3120 | } | ||||||
3121 | |||||||
3122 | VectorTripCount = Builder.CreateSub(TC, R, "n.vec"); | ||||||
3123 | |||||||
3124 | return VectorTripCount; | ||||||
3125 | } | ||||||
3126 | |||||||
3127 | Value *InnerLoopVectorizer::createBitOrPointerCast(Value *V, VectorType *DstVTy, | ||||||
3128 | const DataLayout &DL) { | ||||||
3129 | // Verify that V is a vector type with same number of elements as DstVTy. | ||||||
3130 | auto *DstFVTy = cast<FixedVectorType>(DstVTy); | ||||||
3131 | unsigned VF = DstFVTy->getNumElements(); | ||||||
3132 | auto *SrcVecTy = cast<FixedVectorType>(V->getType()); | ||||||
3133 | 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", 3133, __extension__ __PRETTY_FUNCTION__)); | ||||||
3134 | Type *SrcElemTy = SrcVecTy->getElementType(); | ||||||
3135 | Type *DstElemTy = DstFVTy->getElementType(); | ||||||
3136 | 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", 3137, __extension__ __PRETTY_FUNCTION__)) | ||||||
3137 | "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", 3137, __extension__ __PRETTY_FUNCTION__)); | ||||||
3138 | |||||||
3139 | // Do a direct cast if element types are castable. | ||||||
3140 | if (CastInst::isBitOrNoopPointerCastable(SrcElemTy, DstElemTy, DL)) { | ||||||
3141 | return Builder.CreateBitOrPointerCast(V, DstFVTy); | ||||||
3142 | } | ||||||
3143 | // V cannot be directly casted to desired vector type. | ||||||
3144 | // May happen when V is a floating point vector but DstVTy is a vector of | ||||||
3145 | // pointers or vice-versa. Handle this using a two-step bitcast using an | ||||||
3146 | // intermediate Integer type for the bitcast i.e. Ptr <-> Int <-> Float. | ||||||
3147 | 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", 3148, __extension__ __PRETTY_FUNCTION__)) | ||||||
3148 | "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", 3148, __extension__ __PRETTY_FUNCTION__)); | ||||||
3149 | 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", 3150, __extension__ __PRETTY_FUNCTION__)) | ||||||
3150 | "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", 3150, __extension__ __PRETTY_FUNCTION__)); | ||||||
3151 | Type *IntTy = | ||||||
3152 | IntegerType::getIntNTy(V->getContext(), DL.getTypeSizeInBits(SrcElemTy)); | ||||||
3153 | auto *VecIntTy = FixedVectorType::get(IntTy, VF); | ||||||
3154 | Value *CastVal = Builder.CreateBitOrPointerCast(V, VecIntTy); | ||||||
3155 | return Builder.CreateBitOrPointerCast(CastVal, DstFVTy); | ||||||
3156 | } | ||||||
3157 | |||||||
3158 | void InnerLoopVectorizer::emitMinimumIterationCountCheck(Loop *L, | ||||||
3159 | BasicBlock *Bypass) { | ||||||
3160 | Value *Count = getOrCreateTripCount(L); | ||||||
3161 | // Reuse existing vector loop preheader for TC checks. | ||||||
3162 | // Note that new preheader block is generated for vector loop. | ||||||
3163 | BasicBlock *const TCCheckBlock = LoopVectorPreHeader; | ||||||
3164 | IRBuilder<> Builder(TCCheckBlock->getTerminator()); | ||||||
3165 | |||||||
3166 | // Generate code to check if the loop's trip count is less than VF * UF, or | ||||||
3167 | // equal to it in case a scalar epilogue is required; this implies that the | ||||||
3168 | // vector trip count is zero. This check also covers the case where adding one | ||||||
3169 | // to the backedge-taken count overflowed leading to an incorrect trip count | ||||||
3170 | // of zero. In this case we will also jump to the scalar loop. | ||||||
3171 | auto P = Cost->requiresScalarEpilogue(VF) ? ICmpInst::ICMP_ULE | ||||||
3172 | : ICmpInst::ICMP_ULT; | ||||||
3173 | |||||||
3174 | // If tail is to be folded, vector loop takes care of all iterations. | ||||||
3175 | Value *CheckMinIters = Builder.getFalse(); | ||||||
3176 | if (!Cost->foldTailByMasking()) { | ||||||
3177 | Value *Step = createStepForVF(Builder, Count->getType(), VF, UF); | ||||||
3178 | CheckMinIters = Builder.CreateICmp(P, Count, Step, "min.iters.check"); | ||||||
3179 | } | ||||||
3180 | // Create new preheader for vector loop. | ||||||
3181 | LoopVectorPreHeader = | ||||||
3182 | SplitBlock(TCCheckBlock, TCCheckBlock->getTerminator(), DT, LI, nullptr, | ||||||
3183 | "vector.ph"); | ||||||
3184 | |||||||
3185 | 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", 3187, __extension__ __PRETTY_FUNCTION__)) | ||||||
3186 | 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", 3187, __extension__ __PRETTY_FUNCTION__)) | ||||||
3187 | "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", 3187, __extension__ __PRETTY_FUNCTION__)); | ||||||
3188 | |||||||
3189 | // Update dominator for Bypass & LoopExit (if needed). | ||||||
3190 | DT->changeImmediateDominator(Bypass, TCCheckBlock); | ||||||
3191 | if (!Cost->requiresScalarEpilogue(VF)) | ||||||
3192 | // If there is an epilogue which must run, there's no edge from the | ||||||
3193 | // middle block to exit blocks and thus no need to update the immediate | ||||||
3194 | // dominator of the exit blocks. | ||||||
3195 | DT->changeImmediateDominator(LoopExitBlock, TCCheckBlock); | ||||||
3196 | |||||||
3197 | ReplaceInstWithInst( | ||||||
3198 | TCCheckBlock->getTerminator(), | ||||||
3199 | BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); | ||||||
3200 | LoopBypassBlocks.push_back(TCCheckBlock); | ||||||
3201 | } | ||||||
3202 | |||||||
3203 | BasicBlock *InnerLoopVectorizer::emitSCEVChecks(Loop *L, BasicBlock *Bypass) { | ||||||
3204 | |||||||
3205 | BasicBlock *const SCEVCheckBlock = | ||||||
3206 | RTChecks.emitSCEVChecks(L, Bypass, LoopVectorPreHeader, LoopExitBlock); | ||||||
3207 | if (!SCEVCheckBlock) | ||||||
3208 | return nullptr; | ||||||
3209 | |||||||
3210 | 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", 3213, __extension__ __PRETTY_FUNCTION__)) | ||||||
3211 | (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", 3213, __extension__ __PRETTY_FUNCTION__)) | ||||||
3212 | 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", 3213, __extension__ __PRETTY_FUNCTION__)) | ||||||
3213 | "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", 3213, __extension__ __PRETTY_FUNCTION__)); | ||||||
3214 | |||||||
3215 | |||||||
3216 | // Update dominator only if this is first RT check. | ||||||
3217 | if (LoopBypassBlocks.empty()) { | ||||||
3218 | DT->changeImmediateDominator(Bypass, SCEVCheckBlock); | ||||||
3219 | if (!Cost->requiresScalarEpilogue(VF)) | ||||||
3220 | // If there is an epilogue which must run, there's no edge from the | ||||||
3221 | // middle block to exit blocks and thus no need to update the immediate | ||||||
3222 | // dominator of the exit blocks. | ||||||
3223 | DT->changeImmediateDominator(LoopExitBlock, SCEVCheckBlock); | ||||||
3224 | } | ||||||
3225 | |||||||
3226 | LoopBypassBlocks.push_back(SCEVCheckBlock); | ||||||
3227 | AddedSafetyChecks = true; | ||||||
3228 | return SCEVCheckBlock; | ||||||
3229 | } | ||||||
3230 | |||||||
3231 | BasicBlock *InnerLoopVectorizer::emitMemRuntimeChecks(Loop *L, | ||||||
3232 | BasicBlock *Bypass) { | ||||||
3233 | // VPlan-native path does not do any analysis for runtime checks currently. | ||||||
3234 | if (EnableVPlanNativePath) | ||||||
3235 | return nullptr; | ||||||
3236 | |||||||
3237 | BasicBlock *const MemCheckBlock = | ||||||
3238 | RTChecks.emitMemRuntimeChecks(L, Bypass, LoopVectorPreHeader); | ||||||
3239 | |||||||
3240 | // Check if we generated code that checks in runtime if arrays overlap. We put | ||||||
3241 | // the checks into a separate block to make the more common case of few | ||||||
3242 | // elements faster. | ||||||
3243 | if (!MemCheckBlock) | ||||||
3244 | return nullptr; | ||||||
3245 | |||||||
3246 | if (MemCheckBlock->getParent()->hasOptSize() || OptForSizeBasedOnProfile) { | ||||||
3247 | 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", 3249, __extension__ __PRETTY_FUNCTION__)) | ||||||
3248 | "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", 3249, __extension__ __PRETTY_FUNCTION__)) | ||||||
3249 | "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", 3249, __extension__ __PRETTY_FUNCTION__)); | ||||||
3250 | ORE->emit([&]() { | ||||||
3251 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationCodeSize", | ||||||
3252 | L->getStartLoc(), L->getHeader()) | ||||||
3253 | << "Code-size may be reduced by not forcing " | ||||||
3254 | "vectorization, or by source-code modifications " | ||||||
3255 | "eliminating the need for runtime checks " | ||||||
3256 | "(e.g., adding 'restrict')."; | ||||||
3257 | }); | ||||||
3258 | } | ||||||
3259 | |||||||
3260 | LoopBypassBlocks.push_back(MemCheckBlock); | ||||||
3261 | |||||||
3262 | AddedSafetyChecks = true; | ||||||
3263 | |||||||
3264 | // We currently don't use LoopVersioning for the actual loop cloning but we | ||||||
3265 | // still use it to add the noalias metadata. | ||||||
3266 | LVer = std::make_unique<LoopVersioning>( | ||||||
3267 | *Legal->getLAI(), | ||||||
3268 | Legal->getLAI()->getRuntimePointerChecking()->getChecks(), OrigLoop, LI, | ||||||
3269 | DT, PSE.getSE()); | ||||||
3270 | LVer->prepareNoAliasMetadata(); | ||||||
3271 | return MemCheckBlock; | ||||||
3272 | } | ||||||
3273 | |||||||
3274 | Value *InnerLoopVectorizer::emitTransformedIndex( | ||||||
3275 | IRBuilder<> &B, Value *Index, ScalarEvolution *SE, const DataLayout &DL, | ||||||
3276 | const InductionDescriptor &ID, BasicBlock *VectorHeader) const { | ||||||
3277 | |||||||
3278 | SCEVExpander Exp(*SE, DL, "induction"); | ||||||
3279 | auto Step = ID.getStep(); | ||||||
3280 | auto StartValue = ID.getStartValue(); | ||||||
3281 | assert(Index->getType()->getScalarType() == Step->getType() &&(static_cast <bool> (Index->getType()->getScalarType () == Step->getType() && "Index scalar type does not match StepValue type" ) ? void (0) : __assert_fail ("Index->getType()->getScalarType() == Step->getType() && \"Index scalar type does not match StepValue type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3282, __extension__ __PRETTY_FUNCTION__)) | ||||||
3282 | "Index scalar type does not match StepValue type")(static_cast <bool> (Index->getType()->getScalarType () == Step->getType() && "Index scalar type does not match StepValue type" ) ? void (0) : __assert_fail ("Index->getType()->getScalarType() == Step->getType() && \"Index scalar type does not match StepValue type\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3282, __extension__ __PRETTY_FUNCTION__)); | ||||||
3283 | |||||||
3284 | // Note: the IR at this point is broken. We cannot use SE to create any new | ||||||
3285 | // SCEV and then expand it, hoping that SCEV's simplification will give us | ||||||
3286 | // a more optimal code. Unfortunately, attempt of doing so on invalid IR may | ||||||
3287 | // lead to various SCEV crashes. So all we can do is to use builder and rely | ||||||
3288 | // on InstCombine for future simplifications. Here we handle some trivial | ||||||
3289 | // cases only. | ||||||
3290 | auto CreateAdd = [&B](Value *X, Value *Y) { | ||||||
3291 | 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", 3291, __extension__ __PRETTY_FUNCTION__)); | ||||||
3292 | if (auto *CX = dyn_cast<ConstantInt>(X)) | ||||||
3293 | if (CX->isZero()) | ||||||
3294 | return Y; | ||||||
3295 | if (auto *CY = dyn_cast<ConstantInt>(Y)) | ||||||
3296 | if (CY->isZero()) | ||||||
3297 | return X; | ||||||
3298 | return B.CreateAdd(X, Y); | ||||||
3299 | }; | ||||||
3300 | |||||||
3301 | // We allow X to be a vector type, in which case Y will potentially be | ||||||
3302 | // splatted into a vector with the same element count. | ||||||
3303 | auto CreateMul = [&B](Value *X, Value *Y) { | ||||||
3304 | 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", 3305, __extension__ __PRETTY_FUNCTION__)) | ||||||
3305 | "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", 3305, __extension__ __PRETTY_FUNCTION__)); | ||||||
3306 | if (auto *CX = dyn_cast<ConstantInt>(X)) | ||||||
3307 | if (CX->isOne()) | ||||||
3308 | return Y; | ||||||
3309 | if (auto *CY = dyn_cast<ConstantInt>(Y)) | ||||||
3310 | if (CY->isOne()) | ||||||
3311 | return X; | ||||||
3312 | VectorType *XVTy = dyn_cast<VectorType>(X->getType()); | ||||||
3313 | if (XVTy && !isa<VectorType>(Y->getType())) | ||||||
3314 | Y = B.CreateVectorSplat(XVTy->getElementCount(), Y); | ||||||
3315 | return B.CreateMul(X, Y); | ||||||
3316 | }; | ||||||
3317 | |||||||
3318 | // Get a suitable insert point for SCEV expansion. For blocks in the vector | ||||||
3319 | // loop, choose the end of the vector loop header (=VectorHeader), because | ||||||
3320 | // the DomTree is not kept up-to-date for additional blocks generated in the | ||||||
3321 | // vector loop. By using the header as insertion point, we guarantee that the | ||||||
3322 | // expanded instructions dominate all their uses. | ||||||
3323 | auto GetInsertPoint = [this, &B, VectorHeader]() { | ||||||
3324 | BasicBlock *InsertBB = B.GetInsertPoint()->getParent(); | ||||||
3325 | if (InsertBB != LoopVectorBody && | ||||||
3326 | LI->getLoopFor(VectorHeader) == LI->getLoopFor(InsertBB)) | ||||||
3327 | return VectorHeader->getTerminator(); | ||||||
3328 | return &*B.GetInsertPoint(); | ||||||
3329 | }; | ||||||
3330 | |||||||
3331 | switch (ID.getKind()) { | ||||||
3332 | case InductionDescriptor::IK_IntInduction: { | ||||||
3333 | 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", 3334, __extension__ __PRETTY_FUNCTION__)) | ||||||
3334 | "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", 3334, __extension__ __PRETTY_FUNCTION__)); | ||||||
3335 | 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", 3336, __extension__ __PRETTY_FUNCTION__)) | ||||||
3336 | "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", 3336, __extension__ __PRETTY_FUNCTION__)); | ||||||
3337 | if (ID.getConstIntStepValue() && ID.getConstIntStepValue()->isMinusOne()) | ||||||
3338 | return B.CreateSub(StartValue, Index); | ||||||
3339 | auto *Offset = CreateMul( | ||||||
3340 | Index, Exp.expandCodeFor(Step, Index->getType(), GetInsertPoint())); | ||||||
3341 | return CreateAdd(StartValue, Offset); | ||||||
3342 | } | ||||||
3343 | case InductionDescriptor::IK_PtrInduction: { | ||||||
3344 | assert(isa<SCEVConstant>(Step) &&(static_cast <bool> (isa<SCEVConstant>(Step) && "Expected constant step for pointer induction") ? void (0) : __assert_fail ("isa<SCEVConstant>(Step) && \"Expected constant step for pointer induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3345, __extension__ __PRETTY_FUNCTION__)) | ||||||
3345 | "Expected constant step for pointer induction")(static_cast <bool> (isa<SCEVConstant>(Step) && "Expected constant step for pointer induction") ? void (0) : __assert_fail ("isa<SCEVConstant>(Step) && \"Expected constant step for pointer induction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3345, __extension__ __PRETTY_FUNCTION__)); | ||||||
3346 | return B.CreateGEP( | ||||||
3347 | ID.getElementType(), StartValue, | ||||||
3348 | CreateMul(Index, | ||||||
3349 | Exp.expandCodeFor(Step, Index->getType()->getScalarType(), | ||||||
3350 | GetInsertPoint()))); | ||||||
3351 | } | ||||||
3352 | case InductionDescriptor::IK_FpInduction: { | ||||||
3353 | 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", 3354, __extension__ __PRETTY_FUNCTION__)) | ||||||
3354 | "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", 3354, __extension__ __PRETTY_FUNCTION__)); | ||||||
3355 | 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", 3355, __extension__ __PRETTY_FUNCTION__)); | ||||||
3356 | auto InductionBinOp = ID.getInductionBinOp(); | ||||||
3357 | 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", 3360, __extension__ __PRETTY_FUNCTION__)) | ||||||
3358 | (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", 3360, __extension__ __PRETTY_FUNCTION__)) | ||||||
3359 | 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", 3360, __extension__ __PRETTY_FUNCTION__)) | ||||||
3360 | "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", 3360, __extension__ __PRETTY_FUNCTION__)); | ||||||
3361 | |||||||
3362 | Value *StepValue = cast<SCEVUnknown>(Step)->getValue(); | ||||||
3363 | Value *MulExp = B.CreateFMul(StepValue, Index); | ||||||
3364 | return B.CreateBinOp(InductionBinOp->getOpcode(), StartValue, MulExp, | ||||||
3365 | "induction"); | ||||||
3366 | } | ||||||
3367 | case InductionDescriptor::IK_NoInduction: | ||||||
3368 | return nullptr; | ||||||
3369 | } | ||||||
3370 | llvm_unreachable("invalid enum")::llvm::llvm_unreachable_internal("invalid enum", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3370); | ||||||
3371 | } | ||||||
3372 | |||||||
3373 | Loop *InnerLoopVectorizer::createVectorLoopSkeleton(StringRef Prefix) { | ||||||
3374 | LoopScalarBody = OrigLoop->getHeader(); | ||||||
3375 | LoopVectorPreHeader = OrigLoop->getLoopPreheader(); | ||||||
3376 | 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", 3376, __extension__ __PRETTY_FUNCTION__)); | ||||||
3377 | LoopExitBlock = OrigLoop->getUniqueExitBlock(); // may be nullptr | ||||||
3378 | 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", 3379, __extension__ __PRETTY_FUNCTION__)) | ||||||
3379 | "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", 3379, __extension__ __PRETTY_FUNCTION__)); | ||||||
3380 | |||||||
3381 | LoopMiddleBlock = | ||||||
3382 | SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, | ||||||
3383 | LI, nullptr, Twine(Prefix) + "middle.block"); | ||||||
3384 | LoopScalarPreHeader = | ||||||
3385 | SplitBlock(LoopMiddleBlock, LoopMiddleBlock->getTerminator(), DT, LI, | ||||||
3386 | nullptr, Twine(Prefix) + "scalar.ph"); | ||||||
3387 | |||||||
3388 | auto *ScalarLatchTerm = OrigLoop->getLoopLatch()->getTerminator(); | ||||||
3389 | |||||||
3390 | // Set up the middle block terminator. Two cases: | ||||||
3391 | // 1) If we know that we must execute the scalar epilogue, emit an | ||||||
3392 | // unconditional branch. | ||||||
3393 | // 2) Otherwise, we must have a single unique exit block (due to how we | ||||||
3394 | // implement the multiple exit case). In this case, set up a conditonal | ||||||
3395 | // branch from the middle block to the loop scalar preheader, and the | ||||||
3396 | // exit block. completeLoopSkeleton will update the condition to use an | ||||||
3397 | // iteration check, if required to decide whether to execute the remainder. | ||||||
3398 | BranchInst *BrInst = Cost->requiresScalarEpilogue(VF) ? | ||||||
3399 | BranchInst::Create(LoopScalarPreHeader) : | ||||||
3400 | BranchInst::Create(LoopExitBlock, LoopScalarPreHeader, | ||||||
3401 | Builder.getTrue()); | ||||||
3402 | BrInst->setDebugLoc(ScalarLatchTerm->getDebugLoc()); | ||||||
3403 | ReplaceInstWithInst(LoopMiddleBlock->getTerminator(), BrInst); | ||||||
3404 | |||||||
3405 | // We intentionally don't let SplitBlock to update LoopInfo since | ||||||
3406 | // LoopVectorBody should belong to another loop than LoopVectorPreHeader. | ||||||
3407 | // LoopVectorBody is explicitly added to the correct place few lines later. | ||||||
3408 | LoopVectorBody = | ||||||
3409 | SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, | ||||||
3410 | nullptr, nullptr, Twine(Prefix) + "vector.body"); | ||||||
3411 | |||||||
3412 | // Update dominator for loop exit. | ||||||
3413 | if (!Cost->requiresScalarEpilogue(VF)) | ||||||
3414 | // If there is an epilogue which must run, there's no edge from the | ||||||
3415 | // middle block to exit blocks and thus no need to update the immediate | ||||||
3416 | // dominator of the exit blocks. | ||||||
3417 | DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock); | ||||||
3418 | |||||||
3419 | // Create and register the new vector loop. | ||||||
3420 | Loop *Lp = LI->AllocateLoop(); | ||||||
3421 | Loop *ParentLoop = OrigLoop->getParentLoop(); | ||||||
3422 | |||||||
3423 | // Insert the new loop into the loop nest and register the new basic blocks | ||||||
3424 | // before calling any utilities such as SCEV that require valid LoopInfo. | ||||||
3425 | if (ParentLoop) { | ||||||
3426 | ParentLoop->addChildLoop(Lp); | ||||||
3427 | } else { | ||||||
3428 | LI->addTopLevelLoop(Lp); | ||||||
3429 | } | ||||||
3430 | Lp->addBasicBlockToLoop(LoopVectorBody, *LI); | ||||||
3431 | return Lp; | ||||||
3432 | } | ||||||
3433 | |||||||
3434 | void InnerLoopVectorizer::createInductionResumeValues( | ||||||
3435 | Loop *L, std::pair<BasicBlock *, Value *> AdditionalBypass) { | ||||||
3436 | 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", 3438, __extension__ __PRETTY_FUNCTION__)) | ||||||
3437 | (!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", 3438, __extension__ __PRETTY_FUNCTION__)) | ||||||
3438 | "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", 3438, __extension__ __PRETTY_FUNCTION__)); | ||||||
3439 | |||||||
3440 | Value *VectorTripCount = getOrCreateVectorTripCount(L); | ||||||
3441 | assert(VectorTripCount && L && "Expected valid arguments")(static_cast <bool> (VectorTripCount && L && "Expected valid arguments") ? void (0) : __assert_fail ("VectorTripCount && L && \"Expected valid arguments\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3441, __extension__ __PRETTY_FUNCTION__)); | ||||||
3442 | // We are going to resume the execution of the scalar loop. | ||||||
3443 | // Go over all of the induction variables that we found and fix the | ||||||
3444 | // PHIs that are left in the scalar version of the loop. | ||||||
3445 | // The starting values of PHI nodes depend on the counter of the last | ||||||
3446 | // iteration in the vectorized loop. | ||||||
3447 | // If we come from a bypass edge then we need to start from the original | ||||||
3448 | // start value. | ||||||
3449 | Instruction *OldInduction = Legal->getPrimaryInduction(); | ||||||
3450 | for (auto &InductionEntry : Legal->getInductionVars()) { | ||||||
3451 | PHINode *OrigPhi = InductionEntry.first; | ||||||
3452 | InductionDescriptor II = InductionEntry.second; | ||||||
3453 | |||||||
3454 | // Create phi nodes to merge from the backedge-taken check block. | ||||||
3455 | PHINode *BCResumeVal = | ||||||
3456 | PHINode::Create(OrigPhi->getType(), 3, "bc.resume.val", | ||||||
3457 | LoopScalarPreHeader->getTerminator()); | ||||||
3458 | // Copy original phi DL over to the new one. | ||||||
3459 | BCResumeVal->setDebugLoc(OrigPhi->getDebugLoc()); | ||||||
3460 | Value *&EndValue = IVEndValues[OrigPhi]; | ||||||
3461 | Value *EndValueFromAdditionalBypass = AdditionalBypass.second; | ||||||
3462 | if (OrigPhi == OldInduction) { | ||||||
3463 | // We know what the end value is. | ||||||
3464 | EndValue = VectorTripCount; | ||||||
3465 | } else { | ||||||
3466 | IRBuilder<> B(L->getLoopPreheader()->getTerminator()); | ||||||
3467 | |||||||
3468 | // Fast-math-flags propagate from the original induction instruction. | ||||||
3469 | if (II.getInductionBinOp() && isa<FPMathOperator>(II.getInductionBinOp())) | ||||||
3470 | B.setFastMathFlags(II.getInductionBinOp()->getFastMathFlags()); | ||||||
3471 | |||||||
3472 | Type *StepType = II.getStep()->getType(); | ||||||
3473 | Instruction::CastOps CastOp = | ||||||
3474 | CastInst::getCastOpcode(VectorTripCount, true, StepType, true); | ||||||
3475 | Value *CRD = B.CreateCast(CastOp, VectorTripCount, StepType, "cast.crd"); | ||||||
3476 | const DataLayout &DL = LoopScalarBody->getModule()->getDataLayout(); | ||||||
3477 | EndValue = | ||||||
3478 | emitTransformedIndex(B, CRD, PSE.getSE(), DL, II, LoopVectorBody); | ||||||
3479 | EndValue->setName("ind.end"); | ||||||
3480 | |||||||
3481 | // Compute the end value for the additional bypass (if applicable). | ||||||
3482 | if (AdditionalBypass.first) { | ||||||
3483 | B.SetInsertPoint(&(*AdditionalBypass.first->getFirstInsertionPt())); | ||||||
3484 | CastOp = CastInst::getCastOpcode(AdditionalBypass.second, true, | ||||||
3485 | StepType, true); | ||||||
3486 | CRD = | ||||||
3487 | B.CreateCast(CastOp, AdditionalBypass.second, StepType, "cast.crd"); | ||||||
3488 | EndValueFromAdditionalBypass = | ||||||
3489 | emitTransformedIndex(B, CRD, PSE.getSE(), DL, II, LoopVectorBody); | ||||||
3490 | EndValueFromAdditionalBypass->setName("ind.end"); | ||||||
3491 | } | ||||||
3492 | } | ||||||
3493 | // The new PHI merges the original incoming value, in case of a bypass, | ||||||
3494 | // or the value at the end of the vectorized loop. | ||||||
3495 | BCResumeVal->addIncoming(EndValue, LoopMiddleBlock); | ||||||
3496 | |||||||
3497 | // Fix the scalar body counter (PHI node). | ||||||
3498 | // The old induction's phi node in the scalar body needs the truncated | ||||||
3499 | // value. | ||||||
3500 | for (BasicBlock *BB : LoopBypassBlocks) | ||||||
3501 | BCResumeVal->addIncoming(II.getStartValue(), BB); | ||||||
3502 | |||||||
3503 | if (AdditionalBypass.first) | ||||||
3504 | BCResumeVal->setIncomingValueForBlock(AdditionalBypass.first, | ||||||
3505 | EndValueFromAdditionalBypass); | ||||||
3506 | |||||||
3507 | OrigPhi->setIncomingValueForBlock(LoopScalarPreHeader, BCResumeVal); | ||||||
3508 | } | ||||||
3509 | } | ||||||
3510 | |||||||
3511 | BasicBlock *InnerLoopVectorizer::completeLoopSkeleton(Loop *L, | ||||||
3512 | MDNode *OrigLoopID) { | ||||||
3513 | assert(L && "Expected valid loop.")(static_cast <bool> (L && "Expected valid loop." ) ? void (0) : __assert_fail ("L && \"Expected valid loop.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3513, __extension__ __PRETTY_FUNCTION__)); | ||||||
3514 | |||||||
3515 | // The trip counts should be cached by now. | ||||||
3516 | Value *Count = getOrCreateTripCount(L); | ||||||
3517 | Value *VectorTripCount = getOrCreateVectorTripCount(L); | ||||||
3518 | |||||||
3519 | auto *ScalarLatchTerm = OrigLoop->getLoopLatch()->getTerminator(); | ||||||
3520 | |||||||
3521 | // Add a check in the middle block to see if we have completed | ||||||
3522 | // all of the iterations in the first vector loop. Three cases: | ||||||
3523 | // 1) If we require a scalar epilogue, there is no conditional branch as | ||||||
3524 | // we unconditionally branch to the scalar preheader. Do nothing. | ||||||
3525 | // 2) If (N - N%VF) == N, then we *don't* need to run the remainder. | ||||||
3526 | // Thus if tail is to be folded, we know we don't need to run the | ||||||
3527 | // remainder and we can use the previous value for the condition (true). | ||||||
3528 | // 3) Otherwise, construct a runtime check. | ||||||
3529 | if (!Cost->requiresScalarEpilogue(VF) && !Cost->foldTailByMasking()) { | ||||||
3530 | Instruction *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, | ||||||
3531 | Count, VectorTripCount, "cmp.n", | ||||||
3532 | LoopMiddleBlock->getTerminator()); | ||||||
3533 | |||||||
3534 | // Here we use the same DebugLoc as the scalar loop latch terminator instead | ||||||
3535 | // of the corresponding compare because they may have ended up with | ||||||
3536 | // different line numbers and we want to avoid awkward line stepping while | ||||||
3537 | // debugging. Eg. if the compare has got a line number inside the loop. | ||||||
3538 | CmpN->setDebugLoc(ScalarLatchTerm->getDebugLoc()); | ||||||
3539 | cast<BranchInst>(LoopMiddleBlock->getTerminator())->setCondition(CmpN); | ||||||
3540 | } | ||||||
3541 | |||||||
3542 | // Get ready to start creating new instructions into the vectorized body. | ||||||
3543 | assert(LoopVectorPreHeader == L->getLoopPreheader() &&(static_cast <bool> (LoopVectorPreHeader == L->getLoopPreheader () && "Inconsistent vector loop preheader") ? void (0 ) : __assert_fail ("LoopVectorPreHeader == L->getLoopPreheader() && \"Inconsistent vector loop preheader\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3544, __extension__ __PRETTY_FUNCTION__)) | ||||||
3544 | "Inconsistent vector loop preheader")(static_cast <bool> (LoopVectorPreHeader == L->getLoopPreheader () && "Inconsistent vector loop preheader") ? void (0 ) : __assert_fail ("LoopVectorPreHeader == L->getLoopPreheader() && \"Inconsistent vector loop preheader\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3544, __extension__ __PRETTY_FUNCTION__)); | ||||||
3545 | Builder.SetInsertPoint(&*LoopVectorBody->getFirstInsertionPt()); | ||||||
3546 | |||||||
3547 | #ifdef EXPENSIVE_CHECKS | ||||||
3548 | 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", 3548, __extension__ __PRETTY_FUNCTION__)); | ||||||
3549 | LI->verify(*DT); | ||||||
3550 | #endif | ||||||
3551 | |||||||
3552 | return LoopVectorPreHeader; | ||||||
3553 | } | ||||||
3554 | |||||||
3555 | std::pair<BasicBlock *, Value *> | ||||||
3556 | InnerLoopVectorizer::createVectorizedLoopSkeleton() { | ||||||
3557 | /* | ||||||
3558 | In this function we generate a new loop. The new loop will contain | ||||||
3559 | the vectorized instructions while the old loop will continue to run the | ||||||
3560 | scalar remainder. | ||||||
3561 | |||||||
3562 | [ ] <-- loop iteration number check. | ||||||
3563 | / | | ||||||
3564 | / v | ||||||
3565 | | [ ] <-- vector loop bypass (may consist of multiple blocks). | ||||||
3566 | | / | | ||||||
3567 | | / v | ||||||
3568 | || [ ] <-- vector pre header. | ||||||
3569 | |/ | | ||||||
3570 | | v | ||||||
3571 | | [ ] \ | ||||||
3572 | | [ ]_| <-- vector loop. | ||||||
3573 | | | | ||||||
3574 | | v | ||||||
3575 | \ -[ ] <--- middle-block. | ||||||
3576 | \/ | | ||||||
3577 | /\ v | ||||||
3578 | | ->[ ] <--- new preheader. | ||||||
3579 | | | | ||||||
3580 | (opt) v <-- edge from middle to exit iff epilogue is not required. | ||||||
3581 | | [ ] \ | ||||||
3582 | | [ ]_| <-- old scalar loop to handle remainder (scalar epilogue). | ||||||
3583 | \ | | ||||||
3584 | \ v | ||||||
3585 | >[ ] <-- exit block(s). | ||||||
3586 | ... | ||||||
3587 | */ | ||||||
3588 | |||||||
3589 | // Get the metadata of the original loop before it gets modified. | ||||||
3590 | MDNode *OrigLoopID = OrigLoop->getLoopID(); | ||||||
3591 | |||||||
3592 | // Workaround! Compute the trip count of the original loop and cache it | ||||||
3593 | // before we start modifying the CFG. This code has a systemic problem | ||||||
3594 | // wherein it tries to run analysis over partially constructed IR; this is | ||||||
3595 | // wrong, and not simply for SCEV. The trip count of the original loop | ||||||
3596 | // simply happens to be prone to hitting this in practice. In theory, we | ||||||
3597 | // can hit the same issue for any SCEV, or ValueTracking query done during | ||||||
3598 | // mutation. See PR49900. | ||||||
3599 | getOrCreateTripCount(OrigLoop); | ||||||
3600 | |||||||
3601 | // Create an empty vector loop, and prepare basic blocks for the runtime | ||||||
3602 | // checks. | ||||||
3603 | Loop *Lp = createVectorLoopSkeleton(""); | ||||||
3604 | |||||||
3605 | // Now, compare the new count to zero. If it is zero skip the vector loop and | ||||||
3606 | // jump to the scalar loop. This check also covers the case where the | ||||||
3607 | // backedge-taken count is uint##_max: adding one to it will overflow leading | ||||||
3608 | // to an incorrect trip count of zero. In this (rare) case we will also jump | ||||||
3609 | // to the scalar loop. | ||||||
3610 | emitMinimumIterationCountCheck(Lp, LoopScalarPreHeader); | ||||||
3611 | |||||||
3612 | // Generate the code to check any assumptions that we've made for SCEV | ||||||
3613 | // expressions. | ||||||
3614 | emitSCEVChecks(Lp, LoopScalarPreHeader); | ||||||
3615 | |||||||
3616 | // Generate the code that checks in runtime if arrays overlap. We put the | ||||||
3617 | // checks into a separate block to make the more common case of few elements | ||||||
3618 | // faster. | ||||||
3619 | emitMemRuntimeChecks(Lp, LoopScalarPreHeader); | ||||||
3620 | |||||||
3621 | createHeaderBranch(Lp); | ||||||
3622 | |||||||
3623 | // Emit phis for the new starting index of the scalar loop. | ||||||
3624 | createInductionResumeValues(Lp); | ||||||
3625 | |||||||
3626 | return {completeLoopSkeleton(Lp, OrigLoopID), nullptr}; | ||||||
3627 | } | ||||||
3628 | |||||||
3629 | // Fix up external users of the induction variable. At this point, we are | ||||||
3630 | // in LCSSA form, with all external PHIs that use the IV having one input value, | ||||||
3631 | // coming from the remainder loop. We need those PHIs to also have a correct | ||||||
3632 | // value for the IV when arriving directly from the middle block. | ||||||
3633 | void InnerLoopVectorizer::fixupIVUsers(PHINode *OrigPhi, | ||||||
3634 | const InductionDescriptor &II, | ||||||
3635 | Value *CountRoundDown, Value *EndValue, | ||||||
3636 | BasicBlock *MiddleBlock) { | ||||||
3637 | // There are two kinds of external IV usages - those that use the value | ||||||
3638 | // computed in the last iteration (the PHI) and those that use the penultimate | ||||||
3639 | // value (the value that feeds into the phi from the loop latch). | ||||||
3640 | // We allow both, but they, obviously, have different values. | ||||||
3641 | |||||||
3642 | 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", 3642, __extension__ __PRETTY_FUNCTION__)); | ||||||
3643 | |||||||
3644 | DenseMap<Value *, Value *> MissingVals; | ||||||
3645 | |||||||
3646 | // An external user of the last iteration's value should see the value that | ||||||
3647 | // the remainder loop uses to initialize its own IV. | ||||||
3648 | Value *PostInc = OrigPhi->getIncomingValueForBlock(OrigLoop->getLoopLatch()); | ||||||
3649 | for (User *U : PostInc->users()) { | ||||||
3650 | Instruction *UI = cast<Instruction>(U); | ||||||
3651 | if (!OrigLoop->contains(UI)) { | ||||||
3652 | 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", 3652, __extension__ __PRETTY_FUNCTION__)); | ||||||
3653 | MissingVals[UI] = EndValue; | ||||||
3654 | } | ||||||
3655 | } | ||||||
3656 | |||||||
3657 | // An external user of the penultimate value need to see EndValue - Step. | ||||||
3658 | // The simplest way to get this is to recompute it from the constituent SCEVs, | ||||||
3659 | // that is Start + (Step * (CRD - 1)). | ||||||
3660 | for (User *U : OrigPhi->users()) { | ||||||
3661 | auto *UI = cast<Instruction>(U); | ||||||
3662 | if (!OrigLoop->contains(UI)) { | ||||||
3663 | const DataLayout &DL = | ||||||
3664 | OrigLoop->getHeader()->getModule()->getDataLayout(); | ||||||
3665 | 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", 3665, __extension__ __PRETTY_FUNCTION__)); | ||||||
3666 | |||||||
3667 | IRBuilder<> B(MiddleBlock->getTerminator()); | ||||||
3668 | |||||||
3669 | // Fast-math-flags propagate from the original induction instruction. | ||||||
3670 | if (II.getInductionBinOp() && isa<FPMathOperator>(II.getInductionBinOp())) | ||||||
3671 | B.setFastMathFlags(II.getInductionBinOp()->getFastMathFlags()); | ||||||
3672 | |||||||
3673 | Value *CountMinusOne = B.CreateSub( | ||||||
3674 | CountRoundDown, ConstantInt::get(CountRoundDown->getType(), 1)); | ||||||
3675 | Value *CMO = | ||||||
3676 | !II.getStep()->getType()->isIntegerTy() | ||||||
3677 | ? B.CreateCast(Instruction::SIToFP, CountMinusOne, | ||||||
3678 | II.getStep()->getType()) | ||||||
3679 | : B.CreateSExtOrTrunc(CountMinusOne, II.getStep()->getType()); | ||||||
3680 | CMO->setName("cast.cmo"); | ||||||
3681 | Value *Escape = | ||||||
3682 | emitTransformedIndex(B, CMO, PSE.getSE(), DL, II, LoopVectorBody); | ||||||
3683 | Escape->setName("ind.escape"); | ||||||
3684 | MissingVals[UI] = Escape; | ||||||
3685 | } | ||||||
3686 | } | ||||||
3687 | |||||||
3688 | for (auto &I : MissingVals) { | ||||||
3689 | PHINode *PHI = cast<PHINode>(I.first); | ||||||
3690 | // One corner case we have to handle is two IVs "chasing" each-other, | ||||||
3691 | // that is %IV2 = phi [...], [ %IV1, %latch ] | ||||||
3692 | // In this case, if IV1 has an external use, we need to avoid adding both | ||||||
3693 | // "last value of IV1" and "penultimate value of IV2". So, verify that we | ||||||
3694 | // don't already have an incoming value for the middle block. | ||||||
3695 | if (PHI->getBasicBlockIndex(MiddleBlock) == -1) | ||||||
3696 | PHI->addIncoming(I.second, MiddleBlock); | ||||||
3697 | } | ||||||
3698 | } | ||||||
3699 | |||||||
3700 | namespace { | ||||||
3701 | |||||||
3702 | struct CSEDenseMapInfo { | ||||||
3703 | static bool canHandle(const Instruction *I) { | ||||||
3704 | return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) || | ||||||
3705 | isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I); | ||||||
3706 | } | ||||||
3707 | |||||||
3708 | static inline Instruction *getEmptyKey() { | ||||||
3709 | return DenseMapInfo<Instruction *>::getEmptyKey(); | ||||||
3710 | } | ||||||
3711 | |||||||
3712 | static inline Instruction *getTombstoneKey() { | ||||||
3713 | return DenseMapInfo<Instruction *>::getTombstoneKey(); | ||||||
3714 | } | ||||||
3715 | |||||||
3716 | static unsigned getHashValue(const Instruction *I) { | ||||||
3717 | 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", 3717, __extension__ __PRETTY_FUNCTION__)); | ||||||
3718 | return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(), | ||||||
3719 | I->value_op_end())); | ||||||
3720 | } | ||||||
3721 | |||||||
3722 | static bool isEqual(const Instruction *LHS, const Instruction *RHS) { | ||||||
3723 | if (LHS == getEmptyKey() || RHS == getEmptyKey() || | ||||||
3724 | LHS == getTombstoneKey() || RHS == getTombstoneKey()) | ||||||
3725 | return LHS == RHS; | ||||||
3726 | return LHS->isIdenticalTo(RHS); | ||||||
3727 | } | ||||||
3728 | }; | ||||||
3729 | |||||||
3730 | } // end anonymous namespace | ||||||
3731 | |||||||
3732 | ///Perform cse of induction variable instructions. | ||||||
3733 | static void cse(BasicBlock *BB) { | ||||||
3734 | // Perform simple cse. | ||||||
3735 | SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap; | ||||||
3736 | for (Instruction &In : llvm::make_early_inc_range(*BB)) { | ||||||
3737 | if (!CSEDenseMapInfo::canHandle(&In)) | ||||||
3738 | continue; | ||||||
3739 | |||||||
3740 | // Check if we can replace this instruction with any of the | ||||||
3741 | // visited instructions. | ||||||
3742 | if (Instruction *V = CSEMap.lookup(&In)) { | ||||||
3743 | In.replaceAllUsesWith(V); | ||||||
3744 | In.eraseFromParent(); | ||||||
3745 | continue; | ||||||
3746 | } | ||||||
3747 | |||||||
3748 | CSEMap[&In] = &In; | ||||||
3749 | } | ||||||
3750 | } | ||||||
3751 | |||||||
3752 | InstructionCost | ||||||
3753 | LoopVectorizationCostModel::getVectorCallCost(CallInst *CI, ElementCount VF, | ||||||
3754 | bool &NeedToScalarize) const { | ||||||
3755 | Function *F = CI->getCalledFunction(); | ||||||
3756 | Type *ScalarRetTy = CI->getType(); | ||||||
3757 | SmallVector<Type *, 4> Tys, ScalarTys; | ||||||
3758 | for (auto &ArgOp : CI->args()) | ||||||
3759 | ScalarTys.push_back(ArgOp->getType()); | ||||||
3760 | |||||||
3761 | // Estimate cost of scalarized vector call. The source operands are assumed | ||||||
3762 | // to be vectors, so we need to extract individual elements from there, | ||||||
3763 | // execute VF scalar calls, and then gather the result into the vector return | ||||||
3764 | // value. | ||||||
3765 | InstructionCost ScalarCallCost = | ||||||
3766 | TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys, TTI::TCK_RecipThroughput); | ||||||
3767 | if (VF.isScalar()) | ||||||
3768 | return ScalarCallCost; | ||||||
3769 | |||||||
3770 | // Compute corresponding vector type for return value and arguments. | ||||||
3771 | Type *RetTy = ToVectorTy(ScalarRetTy, VF); | ||||||
3772 | for (Type *ScalarTy : ScalarTys) | ||||||
3773 | Tys.push_back(ToVectorTy(ScalarTy, VF)); | ||||||
3774 | |||||||
3775 | // Compute costs of unpacking argument values for the scalar calls and | ||||||
3776 | // packing the return values to a vector. | ||||||
3777 | InstructionCost ScalarizationCost = getScalarizationOverhead(CI, VF); | ||||||
3778 | |||||||
3779 | InstructionCost Cost = | ||||||
3780 | ScalarCallCost * VF.getKnownMinValue() + ScalarizationCost; | ||||||
3781 | |||||||
3782 | // If we can't emit a vector call for this function, then the currently found | ||||||
3783 | // cost is the cost we need to return. | ||||||
3784 | NeedToScalarize = true; | ||||||
3785 | VFShape Shape = VFShape::get(*CI, VF, false /*HasGlobalPred*/); | ||||||
3786 | Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); | ||||||
3787 | |||||||
3788 | if (!TLI || CI->isNoBuiltin() || !VecFunc) | ||||||
3789 | return Cost; | ||||||
3790 | |||||||
3791 | // If the corresponding vector cost is cheaper, return its cost. | ||||||
3792 | InstructionCost VectorCallCost = | ||||||
3793 | TTI.getCallInstrCost(nullptr, RetTy, Tys, TTI::TCK_RecipThroughput); | ||||||
3794 | if (VectorCallCost < Cost) { | ||||||
3795 | NeedToScalarize = false; | ||||||
3796 | Cost = VectorCallCost; | ||||||
3797 | } | ||||||
3798 | return Cost; | ||||||
3799 | } | ||||||
3800 | |||||||
3801 | static Type *MaybeVectorizeType(Type *Elt, ElementCount VF) { | ||||||
3802 | if (VF.isScalar() || (!Elt->isIntOrPtrTy() && !Elt->isFloatingPointTy())) | ||||||
3803 | return Elt; | ||||||
3804 | return VectorType::get(Elt, VF); | ||||||
3805 | } | ||||||
3806 | |||||||
3807 | InstructionCost | ||||||
3808 | LoopVectorizationCostModel::getVectorIntrinsicCost(CallInst *CI, | ||||||
3809 | ElementCount VF) const { | ||||||
3810 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||||
3811 | 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", 3811, __extension__ __PRETTY_FUNCTION__)); | ||||||
3812 | Type *RetTy = MaybeVectorizeType(CI->getType(), VF); | ||||||
3813 | FastMathFlags FMF; | ||||||
3814 | if (auto *FPMO = dyn_cast<FPMathOperator>(CI)) | ||||||
3815 | FMF = FPMO->getFastMathFlags(); | ||||||
3816 | |||||||
3817 | SmallVector<const Value *> Arguments(CI->args()); | ||||||
3818 | FunctionType *FTy = CI->getCalledFunction()->getFunctionType(); | ||||||
3819 | SmallVector<Type *> ParamTys; | ||||||
3820 | std::transform(FTy->param_begin(), FTy->param_end(), | ||||||
3821 | std::back_inserter(ParamTys), | ||||||
3822 | [&](Type *Ty) { return MaybeVectorizeType(Ty, VF); }); | ||||||
3823 | |||||||
3824 | IntrinsicCostAttributes CostAttrs(ID, RetTy, Arguments, ParamTys, FMF, | ||||||
3825 | dyn_cast<IntrinsicInst>(CI)); | ||||||
3826 | return TTI.getIntrinsicInstrCost(CostAttrs, | ||||||
3827 | TargetTransformInfo::TCK_RecipThroughput); | ||||||
3828 | } | ||||||
3829 | |||||||
3830 | static Type *smallestIntegerVectorType(Type *T1, Type *T2) { | ||||||
3831 | auto *I1 = cast<IntegerType>(cast<VectorType>(T1)->getElementType()); | ||||||
3832 | auto *I2 = cast<IntegerType>(cast<VectorType>(T2)->getElementType()); | ||||||
3833 | return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2; | ||||||
3834 | } | ||||||
3835 | |||||||
3836 | static Type *largestIntegerVectorType(Type *T1, Type *T2) { | ||||||
3837 | auto *I1 = cast<IntegerType>(cast<VectorType>(T1)->getElementType()); | ||||||
3838 | auto *I2 = cast<IntegerType>(cast<VectorType>(T2)->getElementType()); | ||||||
3839 | return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2; | ||||||
3840 | } | ||||||
3841 | |||||||
3842 | void InnerLoopVectorizer::truncateToMinimalBitwidths(VPTransformState &State) { | ||||||
3843 | // For every instruction `I` in MinBWs, truncate the operands, create a | ||||||
3844 | // truncated version of `I` and reextend its result. InstCombine runs | ||||||
3845 | // later and will remove any ext/trunc pairs. | ||||||
3846 | SmallPtrSet<Value *, 4> Erased; | ||||||
3847 | for (const auto &KV : Cost->getMinimalBitwidths()) { | ||||||
3848 | // If the value wasn't vectorized, we must maintain the original scalar | ||||||
3849 | // type. The absence of the value from State indicates that it | ||||||
3850 | // wasn't vectorized. | ||||||
3851 | // FIXME: Should not rely on getVPValue at this point. | ||||||
3852 | VPValue *Def = State.Plan->getVPValue(KV.first, true); | ||||||
3853 | if (!State.hasAnyVectorValue(Def)) | ||||||
3854 | continue; | ||||||
3855 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
3856 | Value *I = State.get(Def, Part); | ||||||
3857 | if (Erased.count(I) || I->use_empty() || !isa<Instruction>(I)) | ||||||
3858 | continue; | ||||||
3859 | Type *OriginalTy = I->getType(); | ||||||
3860 | Type *ScalarTruncatedTy = | ||||||
3861 | IntegerType::get(OriginalTy->getContext(), KV.second); | ||||||
3862 | auto *TruncatedTy = VectorType::get( | ||||||
3863 | ScalarTruncatedTy, cast<VectorType>(OriginalTy)->getElementCount()); | ||||||
3864 | if (TruncatedTy == OriginalTy) | ||||||
3865 | continue; | ||||||
3866 | |||||||
3867 | IRBuilder<> B(cast<Instruction>(I)); | ||||||
3868 | auto ShrinkOperand = [&](Value *V) -> Value * { | ||||||
3869 | if (auto *ZI = dyn_cast<ZExtInst>(V)) | ||||||
3870 | if (ZI->getSrcTy() == TruncatedTy) | ||||||
3871 | return ZI->getOperand(0); | ||||||
3872 | return B.CreateZExtOrTrunc(V, TruncatedTy); | ||||||
3873 | }; | ||||||
3874 | |||||||
3875 | // The actual instruction modification depends on the instruction type, | ||||||
3876 | // unfortunately. | ||||||
3877 | Value *NewI = nullptr; | ||||||
3878 | if (auto *BO = dyn_cast<BinaryOperator>(I)) { | ||||||
3879 | NewI = B.CreateBinOp(BO->getOpcode(), ShrinkOperand(BO->getOperand(0)), | ||||||
3880 | ShrinkOperand(BO->getOperand(1))); | ||||||
3881 | |||||||
3882 | // Any wrapping introduced by shrinking this operation shouldn't be | ||||||
3883 | // considered undefined behavior. So, we can't unconditionally copy | ||||||
3884 | // arithmetic wrapping flags to NewI. | ||||||
3885 | cast<BinaryOperator>(NewI)->copyIRFlags(I, /*IncludeWrapFlags=*/false); | ||||||
3886 | } else if (auto *CI = dyn_cast<ICmpInst>(I)) { | ||||||
3887 | NewI = | ||||||
3888 | B.CreateICmp(CI->getPredicate(), ShrinkOperand(CI->getOperand(0)), | ||||||
3889 | ShrinkOperand(CI->getOperand(1))); | ||||||
3890 | } else if (auto *SI = dyn_cast<SelectInst>(I)) { | ||||||
3891 | NewI = B.CreateSelect(SI->getCondition(), | ||||||
3892 | ShrinkOperand(SI->getTrueValue()), | ||||||
3893 | ShrinkOperand(SI->getFalseValue())); | ||||||
3894 | } else if (auto *CI = dyn_cast<CastInst>(I)) { | ||||||
3895 | switch (CI->getOpcode()) { | ||||||
3896 | default: | ||||||
3897 | llvm_unreachable("Unhandled cast!")::llvm::llvm_unreachable_internal("Unhandled cast!", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 3897); | ||||||
3898 | case Instruction::Trunc: | ||||||
3899 | NewI = ShrinkOperand(CI->getOperand(0)); | ||||||
3900 | break; | ||||||
3901 | case Instruction::SExt: | ||||||
3902 | NewI = B.CreateSExtOrTrunc( | ||||||
3903 | CI->getOperand(0), | ||||||
3904 | smallestIntegerVectorType(OriginalTy, TruncatedTy)); | ||||||
3905 | break; | ||||||
3906 | case Instruction::ZExt: | ||||||
3907 | NewI = B.CreateZExtOrTrunc( | ||||||
3908 | CI->getOperand(0), | ||||||
3909 | smallestIntegerVectorType(OriginalTy, TruncatedTy)); | ||||||
3910 | break; | ||||||
3911 | } | ||||||
3912 | } else if (auto *SI = dyn_cast<ShuffleVectorInst>(I)) { | ||||||
3913 | auto Elements0 = | ||||||
3914 | cast<VectorType>(SI->getOperand(0)->getType())->getElementCount(); | ||||||
3915 | auto *O0 = B.CreateZExtOrTrunc( | ||||||
3916 | SI->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements0)); | ||||||
3917 | auto Elements1 = | ||||||
3918 | cast<VectorType>(SI->getOperand(1)->getType())->getElementCount(); | ||||||
3919 | auto *O1 = B.CreateZExtOrTrunc( | ||||||
3920 | SI->getOperand(1), VectorType::get(ScalarTruncatedTy, Elements1)); | ||||||
3921 | |||||||
3922 | NewI = B.CreateShuffleVector(O0, O1, SI->getShuffleMask()); | ||||||
3923 | } else if (isa<LoadInst>(I) || isa<PHINode>(I)) { | ||||||
3924 | // Don't do anything with the operands, just extend the result. | ||||||
3925 | continue; | ||||||
3926 | } else if (auto *IE = dyn_cast<InsertElementInst>(I)) { | ||||||
3927 | auto Elements = | ||||||
3928 | cast<VectorType>(IE->getOperand(0)->getType())->getElementCount(); | ||||||
3929 | auto *O0 = B.CreateZExtOrTrunc( | ||||||
3930 | IE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements)); | ||||||
3931 | auto *O1 = B.CreateZExtOrTrunc(IE->getOperand(1), ScalarTruncatedTy); | ||||||
3932 | NewI = B.CreateInsertElement(O0, O1, IE->getOperand(2)); | ||||||
3933 | } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) { | ||||||
3934 | auto Elements = | ||||||
3935 | cast<VectorType>(EE->getOperand(0)->getType())->getElementCount(); | ||||||
3936 | auto *O0 = B.CreateZExtOrTrunc( | ||||||
3937 | EE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements)); | ||||||
3938 | NewI = B.CreateExtractElement(O0, EE->getOperand(2)); | ||||||
3939 | } else { | ||||||
3940 | // If we don't know what to do, be conservative and don't do anything. | ||||||
3941 | continue; | ||||||
3942 | } | ||||||
3943 | |||||||
3944 | // Lastly, extend the result. | ||||||
3945 | NewI->takeName(cast<Instruction>(I)); | ||||||
3946 | Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy); | ||||||
3947 | I->replaceAllUsesWith(Res); | ||||||
3948 | cast<Instruction>(I)->eraseFromParent(); | ||||||
3949 | Erased.insert(I); | ||||||
3950 | State.reset(Def, Res, Part); | ||||||
3951 | } | ||||||
3952 | } | ||||||
3953 | |||||||
3954 | // We'll have created a bunch of ZExts that are now parentless. Clean up. | ||||||
3955 | for (const auto &KV : Cost->getMinimalBitwidths()) { | ||||||
3956 | // If the value wasn't vectorized, we must maintain the original scalar | ||||||
3957 | // type. The absence of the value from State indicates that it | ||||||
3958 | // wasn't vectorized. | ||||||
3959 | // FIXME: Should not rely on getVPValue at this point. | ||||||
3960 | VPValue *Def = State.Plan->getVPValue(KV.first, true); | ||||||
3961 | if (!State.hasAnyVectorValue(Def)) | ||||||
3962 | continue; | ||||||
3963 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
3964 | Value *I = State.get(Def, Part); | ||||||
3965 | ZExtInst *Inst = dyn_cast<ZExtInst>(I); | ||||||
3966 | if (Inst && Inst->use_empty()) { | ||||||
3967 | Value *NewI = Inst->getOperand(0); | ||||||
3968 | Inst->eraseFromParent(); | ||||||
3969 | State.reset(Def, NewI, Part); | ||||||
3970 | } | ||||||
3971 | } | ||||||
3972 | } | ||||||
3973 | } | ||||||
3974 | |||||||
3975 | void InnerLoopVectorizer::fixVectorizedLoop(VPTransformState &State) { | ||||||
3976 | // Insert truncates and extends for any truncated instructions as hints to | ||||||
3977 | // InstCombine. | ||||||
3978 | if (VF.isVector()) | ||||||
3979 | truncateToMinimalBitwidths(State); | ||||||
3980 | |||||||
3981 | // Fix widened non-induction PHIs by setting up the PHI operands. | ||||||
3982 | if (OrigPHIsToFix.size()) { | ||||||
3983 | assert(EnableVPlanNativePath &&(static_cast <bool> (EnableVPlanNativePath && "Unexpected non-induction PHIs for fixup in non VPlan-native path" ) ? void (0) : __assert_fail ("EnableVPlanNativePath && \"Unexpected non-induction PHIs for fixup in non VPlan-native path\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3984, __extension__ __PRETTY_FUNCTION__)) | ||||||
3984 | "Unexpected non-induction PHIs for fixup in non VPlan-native path")(static_cast <bool> (EnableVPlanNativePath && "Unexpected non-induction PHIs for fixup in non VPlan-native path" ) ? void (0) : __assert_fail ("EnableVPlanNativePath && \"Unexpected non-induction PHIs for fixup in non VPlan-native path\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 3984, __extension__ __PRETTY_FUNCTION__)); | ||||||
3985 | fixNonInductionPHIs(State); | ||||||
3986 | } | ||||||
3987 | |||||||
3988 | // At this point every instruction in the original loop is widened to a | ||||||
3989 | // vector form. Now we need to fix the recurrences in the loop. These PHI | ||||||
3990 | // nodes are currently empty because we did not want to introduce cycles. | ||||||
3991 | // This is the second stage of vectorizing recurrences. | ||||||
3992 | fixCrossIterationPHIs(State); | ||||||
3993 | |||||||
3994 | // Forget the original basic block. | ||||||
3995 | PSE.getSE()->forgetLoop(OrigLoop); | ||||||
3996 | |||||||
3997 | // If we inserted an edge from the middle block to the unique exit block, | ||||||
3998 | // update uses outside the loop (phis) to account for the newly inserted | ||||||
3999 | // edge. | ||||||
4000 | if (!Cost->requiresScalarEpilogue(VF)) { | ||||||
4001 | // Fix-up external users of the induction variables. | ||||||
4002 | for (auto &Entry : Legal->getInductionVars()) | ||||||
4003 | fixupIVUsers(Entry.first, Entry.second, | ||||||
4004 | getOrCreateVectorTripCount(LI->getLoopFor(LoopVectorBody)), | ||||||
4005 | IVEndValues[Entry.first], LoopMiddleBlock); | ||||||
4006 | |||||||
4007 | fixLCSSAPHIs(State); | ||||||
4008 | } | ||||||
4009 | |||||||
4010 | for (Instruction *PI : PredicatedInstructions) | ||||||
4011 | sinkScalarOperands(&*PI); | ||||||
4012 | |||||||
4013 | // Remove redundant induction instructions. | ||||||
4014 | cse(LoopVectorBody); | ||||||
4015 | |||||||
4016 | // Set/update profile weights for the vector and remainder loops as original | ||||||
4017 | // loop iterations are now distributed among them. Note that original loop | ||||||
4018 | // represented by LoopScalarBody becomes remainder loop after vectorization. | ||||||
4019 | // | ||||||
4020 | // For cases like foldTailByMasking() and requiresScalarEpiloque() we may | ||||||
4021 | // end up getting slightly roughened result but that should be OK since | ||||||
4022 | // profile is not inherently precise anyway. Note also possible bypass of | ||||||
4023 | // vector code caused by legality checks is ignored, assigning all the weight | ||||||
4024 | // to the vector loop, optimistically. | ||||||
4025 | // | ||||||
4026 | // For scalable vectorization we can't know at compile time how many iterations | ||||||
4027 | // of the loop are handled in one vector iteration, so instead assume a pessimistic | ||||||
4028 | // vscale of '1'. | ||||||
4029 | setProfileInfoAfterUnrolling( | ||||||
4030 | LI->getLoopFor(LoopScalarBody), LI->getLoopFor(LoopVectorBody), | ||||||
4031 | LI->getLoopFor(LoopScalarBody), VF.getKnownMinValue() * UF); | ||||||
4032 | } | ||||||
4033 | |||||||
4034 | void InnerLoopVectorizer::fixCrossIterationPHIs(VPTransformState &State) { | ||||||
4035 | // In order to support recurrences we need to be able to vectorize Phi nodes. | ||||||
4036 | // Phi nodes have cycles, so we need to vectorize them in two stages. This is | ||||||
4037 | // stage #2: We now need to fix the recurrences by adding incoming edges to | ||||||
4038 | // the currently empty PHI nodes. At this point every instruction in the | ||||||
4039 | // original loop is widened to a vector form so we can use them to construct | ||||||
4040 | // the incoming edges. | ||||||
4041 | VPBasicBlock *Header = State.Plan->getEntry()->getEntryBasicBlock(); | ||||||
4042 | for (VPRecipeBase &R : Header->phis()) { | ||||||
4043 | if (auto *ReductionPhi = dyn_cast<VPReductionPHIRecipe>(&R)) | ||||||
4044 | fixReduction(ReductionPhi, State); | ||||||
4045 | else if (auto *FOR = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&R)) | ||||||
4046 | fixFirstOrderRecurrence(FOR, State); | ||||||
4047 | } | ||||||
4048 | } | ||||||
4049 | |||||||
4050 | void InnerLoopVectorizer::fixFirstOrderRecurrence( | ||||||
4051 | VPFirstOrderRecurrencePHIRecipe *PhiR, VPTransformState &State) { | ||||||
4052 | // This is the second phase of vectorizing first-order recurrences. An | ||||||
4053 | // overview of the transformation is described below. Suppose we have the | ||||||
4054 | // following loop. | ||||||
4055 | // | ||||||
4056 | // for (int i = 0; i < n; ++i) | ||||||
4057 | // b[i] = a[i] - a[i - 1]; | ||||||
4058 | // | ||||||
4059 | // There is a first-order recurrence on "a". For this loop, the shorthand | ||||||
4060 | // scalar IR looks like: | ||||||
4061 | // | ||||||
4062 | // scalar.ph: | ||||||
4063 | // s_init = a[-1] | ||||||
4064 | // br scalar.body | ||||||
4065 | // | ||||||
4066 | // scalar.body: | ||||||
4067 | // i = phi [0, scalar.ph], [i+1, scalar.body] | ||||||
4068 | // s1 = phi [s_init, scalar.ph], [s2, scalar.body] | ||||||
4069 | // s2 = a[i] | ||||||
4070 | // b[i] = s2 - s1 | ||||||
4071 | // br cond, scalar.body, ... | ||||||
4072 | // | ||||||
4073 | // In this example, s1 is a recurrence because it's value depends on the | ||||||
4074 | // previous iteration. In the first phase of vectorization, we created a | ||||||
4075 | // vector phi v1 for s1. We now complete the vectorization and produce the | ||||||
4076 | // shorthand vector IR shown below (for VF = 4, UF = 1). | ||||||
4077 | // | ||||||
4078 | // vector.ph: | ||||||
4079 | // v_init = vector(..., ..., ..., a[-1]) | ||||||
4080 | // br vector.body | ||||||
4081 | // | ||||||
4082 | // vector.body | ||||||
4083 | // i = phi [0, vector.ph], [i+4, vector.body] | ||||||
4084 | // v1 = phi [v_init, vector.ph], [v2, vector.body] | ||||||
4085 | // v2 = a[i, i+1, i+2, i+3]; | ||||||
4086 | // v3 = vector(v1(3), v2(0, 1, 2)) | ||||||
4087 | // b[i, i+1, i+2, i+3] = v2 - v3 | ||||||
4088 | // br cond, vector.body, middle.block | ||||||
4089 | // | ||||||
4090 | // middle.block: | ||||||
4091 | // x = v2(3) | ||||||
4092 | // br scalar.ph | ||||||
4093 | // | ||||||
4094 | // scalar.ph: | ||||||
4095 | // s_init = phi [x, middle.block], [a[-1], otherwise] | ||||||
4096 | // br scalar.body | ||||||
4097 | // | ||||||
4098 | // After execution completes the vector loop, we extract the next value of | ||||||
4099 | // the recurrence (x) to use as the initial value in the scalar loop. | ||||||
4100 | |||||||
4101 | // Extract the last vector element in the middle block. This will be the | ||||||
4102 | // initial value for the recurrence when jumping to the scalar loop. | ||||||
4103 | VPValue *PreviousDef = PhiR->getBackedgeValue(); | ||||||
4104 | Value *Incoming = State.get(PreviousDef, UF - 1); | ||||||
4105 | auto *ExtractForScalar = Incoming; | ||||||
4106 | auto *IdxTy = Builder.getInt32Ty(); | ||||||
4107 | if (VF.isVector()) { | ||||||
4108 | auto *One = ConstantInt::get(IdxTy, 1); | ||||||
4109 | Builder.SetInsertPoint(LoopMiddleBlock->getTerminator()); | ||||||
4110 | auto *RuntimeVF = getRuntimeVF(Builder, IdxTy, VF); | ||||||
4111 | auto *LastIdx = Builder.CreateSub(RuntimeVF, One); | ||||||
4112 | ExtractForScalar = Builder.CreateExtractElement(ExtractForScalar, LastIdx, | ||||||
4113 | "vector.recur.extract"); | ||||||
4114 | } | ||||||
4115 | // Extract the second last element in the middle block if the | ||||||
4116 | // Phi is used outside the loop. We need to extract the phi itself | ||||||
4117 | // and not the last element (the phi update in the current iteration). This | ||||||
4118 | // will be the value when jumping to the exit block from the LoopMiddleBlock, | ||||||
4119 | // when the scalar loop is not run at all. | ||||||
4120 | Value *ExtractForPhiUsedOutsideLoop = nullptr; | ||||||
4121 | if (VF.isVector()) { | ||||||
4122 | auto *RuntimeVF = getRuntimeVF(Builder, IdxTy, VF); | ||||||
4123 | auto *Idx = Builder.CreateSub(RuntimeVF, ConstantInt::get(IdxTy, 2)); | ||||||
4124 | ExtractForPhiUsedOutsideLoop = Builder.CreateExtractElement( | ||||||
4125 | Incoming, Idx, "vector.recur.extract.for.phi"); | ||||||
4126 | } else if (UF > 1) | ||||||
4127 | // When loop is unrolled without vectorizing, initialize | ||||||
4128 | // ExtractForPhiUsedOutsideLoop with the value just prior to unrolled value | ||||||
4129 | // of `Incoming`. This is analogous to the vectorized case above: extracting | ||||||
4130 | // the second last element when VF > 1. | ||||||
4131 | ExtractForPhiUsedOutsideLoop = State.get(PreviousDef, UF - 2); | ||||||
4132 | |||||||
4133 | // Fix the initial value of the original recurrence in the scalar loop. | ||||||
4134 | Builder.SetInsertPoint(&*LoopScalarPreHeader->begin()); | ||||||
4135 | PHINode *Phi = cast<PHINode>(PhiR->getUnderlyingValue()); | ||||||
4136 | auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init"); | ||||||
4137 | auto *ScalarInit = PhiR->getStartValue()->getLiveInIRValue(); | ||||||
4138 | for (auto *BB : predecessors(LoopScalarPreHeader)) { | ||||||
4139 | auto *Incoming = BB == LoopMiddleBlock ? ExtractForScalar : ScalarInit; | ||||||
4140 | Start->addIncoming(Incoming, BB); | ||||||
4141 | } | ||||||
4142 | |||||||
4143 | Phi->setIncomingValueForBlock(LoopScalarPreHeader, Start); | ||||||
4144 | Phi->setName("scalar.recur"); | ||||||
4145 | |||||||
4146 | // Finally, fix users of the recurrence outside the loop. The users will need | ||||||
4147 | // either the last value of the scalar recurrence or the last value of the | ||||||
4148 | // vector recurrence we extracted in the middle block. Since the loop is in | ||||||
4149 | // LCSSA form, we just need to find all the phi nodes for the original scalar | ||||||
4150 | // recurrence in the exit block, and then add an edge for the middle block. | ||||||
4151 | // Note that LCSSA does not imply single entry when the original scalar loop | ||||||
4152 | // had multiple exiting edges (as we always run the last iteration in the | ||||||
4153 | // scalar epilogue); in that case, there is no edge from middle to exit and | ||||||
4154 | // and thus no phis which needed updated. | ||||||
4155 | if (!Cost->requiresScalarEpilogue(VF)) | ||||||
4156 | for (PHINode &LCSSAPhi : LoopExitBlock->phis()) | ||||||
4157 | if (llvm::is_contained(LCSSAPhi.incoming_values(), Phi)) | ||||||
4158 | LCSSAPhi.addIncoming(ExtractForPhiUsedOutsideLoop, LoopMiddleBlock); | ||||||
4159 | } | ||||||
4160 | |||||||
4161 | void InnerLoopVectorizer::fixReduction(VPReductionPHIRecipe *PhiR, | ||||||
4162 | VPTransformState &State) { | ||||||
4163 | PHINode *OrigPhi = cast<PHINode>(PhiR->getUnderlyingValue()); | ||||||
4164 | // Get it's reduction variable descriptor. | ||||||
4165 | 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", 4166, __extension__ __PRETTY_FUNCTION__)) | ||||||
4166 | "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", 4166, __extension__ __PRETTY_FUNCTION__)); | ||||||
4167 | const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor(); | ||||||
4168 | |||||||
4169 | RecurKind RK = RdxDesc.getRecurrenceKind(); | ||||||
4170 | TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue(); | ||||||
4171 | Instruction *LoopExitInst = RdxDesc.getLoopExitInstr(); | ||||||
4172 | setDebugLocFromInst(ReductionStartValue); | ||||||
4173 | |||||||
4174 | VPValue *LoopExitInstDef = PhiR->getBackedgeValue(); | ||||||
4175 | // This is the vector-clone of the value that leaves the loop. | ||||||
4176 | Type *VecTy = State.get(LoopExitInstDef, 0)->getType(); | ||||||
4177 | |||||||
4178 | // Wrap flags are in general invalid after vectorization, clear them. | ||||||
4179 | clearReductionWrapFlags(RdxDesc, State); | ||||||
4180 | |||||||
4181 | // Before each round, move the insertion point right between | ||||||
4182 | // the PHIs and the values we are going to write. | ||||||
4183 | // This allows us to write both PHINodes and the extractelement | ||||||
4184 | // instructions. | ||||||
4185 | Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt()); | ||||||
4186 | |||||||
4187 | setDebugLocFromInst(LoopExitInst); | ||||||
4188 | |||||||
4189 | Type *PhiTy = OrigPhi->getType(); | ||||||
4190 | // If tail is folded by masking, the vector value to leave the loop should be | ||||||
4191 | // a Select choosing between the vectorized LoopExitInst and vectorized Phi, | ||||||
4192 | // instead of the former. For an inloop reduction the reduction will already | ||||||
4193 | // be predicated, and does not need to be handled here. | ||||||
4194 | if (Cost->foldTailByMasking() && !PhiR->isInLoop()) { | ||||||
4195 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
4196 | Value *VecLoopExitInst = State.get(LoopExitInstDef, Part); | ||||||
4197 | Value *Sel = nullptr; | ||||||
4198 | for (User *U : VecLoopExitInst->users()) { | ||||||
4199 | if (isa<SelectInst>(U)) { | ||||||
4200 | 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", 4200, __extension__ __PRETTY_FUNCTION__)); | ||||||
4201 | Sel = U; | ||||||
4202 | } else | ||||||
4203 | 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", 4203, __extension__ __PRETTY_FUNCTION__)); | ||||||
4204 | } | ||||||
4205 | 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", 4205, __extension__ __PRETTY_FUNCTION__)); | ||||||
4206 | State.reset(LoopExitInstDef, Sel, Part); | ||||||
4207 | |||||||
4208 | // If the target can create a predicated operator for the reduction at no | ||||||
4209 | // extra cost in the loop (for example a predicated vadd), it can be | ||||||
4210 | // cheaper for the select to remain in the loop than be sunk out of it, | ||||||
4211 | // and so use the select value for the phi instead of the old | ||||||
4212 | // LoopExitValue. | ||||||
4213 | if (PreferPredicatedReductionSelect || | ||||||
4214 | TTI->preferPredicatedReductionSelect( | ||||||
4215 | RdxDesc.getOpcode(), PhiTy, | ||||||
4216 | TargetTransformInfo::ReductionFlags())) { | ||||||
4217 | auto *VecRdxPhi = | ||||||
4218 | cast<PHINode>(State.get(PhiR, Part)); | ||||||
4219 | VecRdxPhi->setIncomingValueForBlock( | ||||||
4220 | LI->getLoopFor(LoopVectorBody)->getLoopLatch(), Sel); | ||||||
4221 | } | ||||||
4222 | } | ||||||
4223 | } | ||||||
4224 | |||||||
4225 | // If the vector reduction can be performed in a smaller type, we truncate | ||||||
4226 | // then extend the loop exit value to enable InstCombine to evaluate the | ||||||
4227 | // entire expression in the smaller type. | ||||||
4228 | if (VF.isVector() && PhiTy != RdxDesc.getRecurrenceType()) { | ||||||
4229 | 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", 4229, __extension__ __PRETTY_FUNCTION__)); | ||||||
4230 | Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF); | ||||||
4231 | Builder.SetInsertPoint( | ||||||
4232 | LI->getLoopFor(LoopVectorBody)->getLoopLatch()->getTerminator()); | ||||||
4233 | VectorParts RdxParts(UF); | ||||||
4234 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
4235 | RdxParts[Part] = State.get(LoopExitInstDef, Part); | ||||||
4236 | Value *Trunc = Builder.CreateTrunc(RdxParts[Part], RdxVecTy); | ||||||
4237 | Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy) | ||||||
4238 | : Builder.CreateZExt(Trunc, VecTy); | ||||||
4239 | for (User *U : llvm::make_early_inc_range(RdxParts[Part]->users())) | ||||||
4240 | if (U != Trunc) { | ||||||
4241 | U->replaceUsesOfWith(RdxParts[Part], Extnd); | ||||||
4242 | RdxParts[Part] = Extnd; | ||||||
4243 | } | ||||||
4244 | } | ||||||
4245 | Builder.SetInsertPoint(&*LoopMiddleBlock->getFirstInsertionPt()); | ||||||
4246 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
4247 | RdxParts[Part] = Builder.CreateTrunc(RdxParts[Part], RdxVecTy); | ||||||
4248 | State.reset(LoopExitInstDef, RdxParts[Part], Part); | ||||||
4249 | } | ||||||
4250 | } | ||||||
4251 | |||||||
4252 | // Reduce all of the unrolled parts into a single vector. | ||||||
4253 | Value *ReducedPartRdx = State.get(LoopExitInstDef, 0); | ||||||
4254 | unsigned Op = RecurrenceDescriptor::getOpcode(RK); | ||||||
4255 | |||||||
4256 | // The middle block terminator has already been assigned a DebugLoc here (the | ||||||
4257 | // OrigLoop's single latch terminator). We want the whole middle block to | ||||||
4258 | // appear to execute on this line because: (a) it is all compiler generated, | ||||||
4259 | // (b) these instructions are always executed after evaluating the latch | ||||||
4260 | // conditional branch, and (c) other passes may add new predecessors which | ||||||
4261 | // terminate on this line. This is the easiest way to ensure we don't | ||||||
4262 | // accidentally cause an extra step back into the loop while debugging. | ||||||
4263 | setDebugLocFromInst(LoopMiddleBlock->getTerminator()); | ||||||
4264 | if (PhiR->isOrdered()) | ||||||
4265 | ReducedPartRdx = State.get(LoopExitInstDef, UF - 1); | ||||||
4266 | else { | ||||||
4267 | // Floating-point operations should have some FMF to enable the reduction. | ||||||
4268 | IRBuilderBase::FastMathFlagGuard FMFG(Builder); | ||||||
4269 | Builder.setFastMathFlags(RdxDesc.getFastMathFlags()); | ||||||
4270 | for (unsigned Part = 1; Part < UF; ++Part) { | ||||||
4271 | Value *RdxPart = State.get(LoopExitInstDef, Part); | ||||||
4272 | if (Op != Instruction::ICmp && Op != Instruction::FCmp) { | ||||||
4273 | ReducedPartRdx = Builder.CreateBinOp( | ||||||
4274 | (Instruction::BinaryOps)Op, RdxPart, ReducedPartRdx, "bin.rdx"); | ||||||
4275 | } else if (RecurrenceDescriptor::isSelectCmpRecurrenceKind(RK)) | ||||||
4276 | ReducedPartRdx = createSelectCmpOp(Builder, ReductionStartValue, RK, | ||||||
4277 | ReducedPartRdx, RdxPart); | ||||||
4278 | else | ||||||
4279 | ReducedPartRdx = createMinMaxOp(Builder, RK, ReducedPartRdx, RdxPart); | ||||||
4280 | } | ||||||
4281 | } | ||||||
4282 | |||||||
4283 | // Create the reduction after the loop. Note that inloop reductions create the | ||||||
4284 | // target reduction in the loop using a Reduction recipe. | ||||||
4285 | if (VF.isVector() && !PhiR->isInLoop()) { | ||||||
4286 | ReducedPartRdx = | ||||||
4287 | createTargetReduction(Builder, TTI, RdxDesc, ReducedPartRdx, OrigPhi); | ||||||
4288 | // If the reduction can be performed in a smaller type, we need to extend | ||||||
4289 | // the reduction to the wider type before we branch to the original loop. | ||||||
4290 | if (PhiTy != RdxDesc.getRecurrenceType()) | ||||||
4291 | ReducedPartRdx = RdxDesc.isSigned() | ||||||
4292 | ? Builder.CreateSExt(ReducedPartRdx, PhiTy) | ||||||
4293 | : Builder.CreateZExt(ReducedPartRdx, PhiTy); | ||||||
4294 | } | ||||||
4295 | |||||||
4296 | // Create a phi node that merges control-flow from the backedge-taken check | ||||||
4297 | // block and the middle block. | ||||||
4298 | PHINode *BCBlockPhi = PHINode::Create(PhiTy, 2, "bc.merge.rdx", | ||||||
4299 | LoopScalarPreHeader->getTerminator()); | ||||||
4300 | for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) | ||||||
4301 | BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]); | ||||||
4302 | BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock); | ||||||
4303 | |||||||
4304 | // Now, we need to fix the users of the reduction variable | ||||||
4305 | // inside and outside of the scalar remainder loop. | ||||||
4306 | |||||||
4307 | // We know that the loop is in LCSSA form. We need to update the PHI nodes | ||||||
4308 | // in the exit blocks. See comment on analogous loop in | ||||||
4309 | // fixFirstOrderRecurrence for a more complete explaination of the logic. | ||||||
4310 | if (!Cost->requiresScalarEpilogue(VF)) | ||||||
4311 | for (PHINode &LCSSAPhi : LoopExitBlock->phis()) | ||||||
4312 | if (llvm::is_contained(LCSSAPhi.incoming_values(), LoopExitInst)) | ||||||
4313 | LCSSAPhi.addIncoming(ReducedPartRdx, LoopMiddleBlock); | ||||||
4314 | |||||||
4315 | // Fix the scalar loop reduction variable with the incoming reduction sum | ||||||
4316 | // from the vector body and from the backedge value. | ||||||
4317 | int IncomingEdgeBlockIdx = | ||||||
4318 | OrigPhi->getBasicBlockIndex(OrigLoop->getLoopLatch()); | ||||||
4319 | 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", 4319, __extension__ __PRETTY_FUNCTION__)); | ||||||
4320 | // Pick the other block. | ||||||
4321 | int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1); | ||||||
4322 | OrigPhi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi); | ||||||
4323 | OrigPhi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst); | ||||||
4324 | } | ||||||
4325 | |||||||
4326 | void InnerLoopVectorizer::clearReductionWrapFlags(const RecurrenceDescriptor &RdxDesc, | ||||||
4327 | VPTransformState &State) { | ||||||
4328 | RecurKind RK = RdxDesc.getRecurrenceKind(); | ||||||
4329 | if (RK != RecurKind::Add && RK != RecurKind::Mul) | ||||||
4330 | return; | ||||||
4331 | |||||||
4332 | Instruction *LoopExitInstr = RdxDesc.getLoopExitInstr(); | ||||||
4333 | assert(LoopExitInstr && "null loop exit instruction")(static_cast <bool> (LoopExitInstr && "null loop exit instruction" ) ? void (0) : __assert_fail ("LoopExitInstr && \"null loop exit instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4333, __extension__ __PRETTY_FUNCTION__)); | ||||||
4334 | SmallVector<Instruction *, 8> Worklist; | ||||||
4335 | SmallPtrSet<Instruction *, 8> Visited; | ||||||
4336 | Worklist.push_back(LoopExitInstr); | ||||||
4337 | Visited.insert(LoopExitInstr); | ||||||
4338 | |||||||
4339 | while (!Worklist.empty()) { | ||||||
4340 | Instruction *Cur = Worklist.pop_back_val(); | ||||||
4341 | if (isa<OverflowingBinaryOperator>(Cur)) | ||||||
4342 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
4343 | // FIXME: Should not rely on getVPValue at this point. | ||||||
4344 | Value *V = State.get(State.Plan->getVPValue(Cur, true), Part); | ||||||
4345 | cast<Instruction>(V)->dropPoisonGeneratingFlags(); | ||||||
4346 | } | ||||||
4347 | |||||||
4348 | for (User *U : Cur->users()) { | ||||||
4349 | Instruction *UI = cast<Instruction>(U); | ||||||
4350 | if ((Cur != LoopExitInstr || OrigLoop->contains(UI->getParent())) && | ||||||
4351 | Visited.insert(UI).second) | ||||||
4352 | Worklist.push_back(UI); | ||||||
4353 | } | ||||||
4354 | } | ||||||
4355 | } | ||||||
4356 | |||||||
4357 | void InnerLoopVectorizer::fixLCSSAPHIs(VPTransformState &State) { | ||||||
4358 | for (PHINode &LCSSAPhi : LoopExitBlock->phis()) { | ||||||
4359 | if (LCSSAPhi.getBasicBlockIndex(LoopMiddleBlock) != -1) | ||||||
4360 | // Some phis were already hand updated by the reduction and recurrence | ||||||
4361 | // code above, leave them alone. | ||||||
4362 | continue; | ||||||
4363 | |||||||
4364 | auto *IncomingValue = LCSSAPhi.getIncomingValue(0); | ||||||
4365 | // Non-instruction incoming values will have only one value. | ||||||
4366 | |||||||
4367 | VPLane Lane = VPLane::getFirstLane(); | ||||||
4368 | if (isa<Instruction>(IncomingValue) && | ||||||
4369 | !Cost->isUniformAfterVectorization(cast<Instruction>(IncomingValue), | ||||||
4370 | VF)) | ||||||
4371 | Lane = VPLane::getLastLaneForVF(VF); | ||||||
4372 | |||||||
4373 | // Can be a loop invariant incoming value or the last scalar value to be | ||||||
4374 | // extracted from the vectorized loop. | ||||||
4375 | // FIXME: Should not rely on getVPValue at this point. | ||||||
4376 | Builder.SetInsertPoint(LoopMiddleBlock->getTerminator()); | ||||||
4377 | Value *lastIncomingValue = | ||||||
4378 | OrigLoop->isLoopInvariant(IncomingValue) | ||||||
4379 | ? IncomingValue | ||||||
4380 | : State.get(State.Plan->getVPValue(IncomingValue, true), | ||||||
4381 | VPIteration(UF - 1, Lane)); | ||||||
4382 | LCSSAPhi.addIncoming(lastIncomingValue, LoopMiddleBlock); | ||||||
4383 | } | ||||||
4384 | } | ||||||
4385 | |||||||
4386 | void InnerLoopVectorizer::sinkScalarOperands(Instruction *PredInst) { | ||||||
4387 | // The basic block and loop containing the predicated instruction. | ||||||
4388 | auto *PredBB = PredInst->getParent(); | ||||||
4389 | auto *VectorLoop = LI->getLoopFor(PredBB); | ||||||
4390 | |||||||
4391 | // Initialize a worklist with the operands of the predicated instruction. | ||||||
4392 | SetVector<Value *> Worklist(PredInst->op_begin(), PredInst->op_end()); | ||||||
4393 | |||||||
4394 | // Holds instructions that we need to analyze again. An instruction may be | ||||||
4395 | // reanalyzed if we don't yet know if we can sink it or not. | ||||||
4396 | SmallVector<Instruction *, 8> InstsToReanalyze; | ||||||
4397 | |||||||
4398 | // Returns true if a given use occurs in the predicated block. Phi nodes use | ||||||
4399 | // their operands in their corresponding predecessor blocks. | ||||||
4400 | auto isBlockOfUsePredicated = [&](Use &U) -> bool { | ||||||
4401 | auto *I = cast<Instruction>(U.getUser()); | ||||||
4402 | BasicBlock *BB = I->getParent(); | ||||||
4403 | if (auto *Phi = dyn_cast<PHINode>(I)) | ||||||
4404 | BB = Phi->getIncomingBlock( | ||||||
4405 | PHINode::getIncomingValueNumForOperand(U.getOperandNo())); | ||||||
4406 | return BB == PredBB; | ||||||
4407 | }; | ||||||
4408 | |||||||
4409 | // Iteratively sink the scalarized operands of the predicated instruction | ||||||
4410 | // into the block we created for it. When an instruction is sunk, it's | ||||||
4411 | // operands are then added to the worklist. The algorithm ends after one pass | ||||||
4412 | // through the worklist doesn't sink a single instruction. | ||||||
4413 | bool Changed; | ||||||
4414 | do { | ||||||
4415 | // Add the instructions that need to be reanalyzed to the worklist, and | ||||||
4416 | // reset the changed indicator. | ||||||
4417 | Worklist.insert(InstsToReanalyze.begin(), InstsToReanalyze.end()); | ||||||
4418 | InstsToReanalyze.clear(); | ||||||
4419 | Changed = false; | ||||||
4420 | |||||||
4421 | while (!Worklist.empty()) { | ||||||
4422 | auto *I = dyn_cast<Instruction>(Worklist.pop_back_val()); | ||||||
4423 | |||||||
4424 | // We can't sink an instruction if it is a phi node, is not in the loop, | ||||||
4425 | // or may have side effects. | ||||||
4426 | if (!I || isa<PHINode>(I) || !VectorLoop->contains(I) || | ||||||
4427 | I->mayHaveSideEffects()) | ||||||
4428 | continue; | ||||||
4429 | |||||||
4430 | // If the instruction is already in PredBB, check if we can sink its | ||||||
4431 | // operands. In that case, VPlan's sinkScalarOperands() succeeded in | ||||||
4432 | // sinking the scalar instruction I, hence it appears in PredBB; but it | ||||||
4433 | // may have failed to sink I's operands (recursively), which we try | ||||||
4434 | // (again) here. | ||||||
4435 | if (I->getParent() == PredBB) { | ||||||
4436 | Worklist.insert(I->op_begin(), I->op_end()); | ||||||
4437 | continue; | ||||||
4438 | } | ||||||
4439 | |||||||
4440 | // It's legal to sink the instruction if all its uses occur in the | ||||||
4441 | // predicated block. Otherwise, there's nothing to do yet, and we may | ||||||
4442 | // need to reanalyze the instruction. | ||||||
4443 | if (!llvm::all_of(I->uses(), isBlockOfUsePredicated)) { | ||||||
4444 | InstsToReanalyze.push_back(I); | ||||||
4445 | continue; | ||||||
4446 | } | ||||||
4447 | |||||||
4448 | // Move the instruction to the beginning of the predicated block, and add | ||||||
4449 | // it's operands to the worklist. | ||||||
4450 | I->moveBefore(&*PredBB->getFirstInsertionPt()); | ||||||
4451 | Worklist.insert(I->op_begin(), I->op_end()); | ||||||
4452 | |||||||
4453 | // The sinking may have enabled other instructions to be sunk, so we will | ||||||
4454 | // need to iterate. | ||||||
4455 | Changed = true; | ||||||
4456 | } | ||||||
4457 | } while (Changed); | ||||||
4458 | } | ||||||
4459 | |||||||
4460 | void InnerLoopVectorizer::fixNonInductionPHIs(VPTransformState &State) { | ||||||
4461 | for (PHINode *OrigPhi : OrigPHIsToFix) { | ||||||
4462 | VPWidenPHIRecipe *VPPhi = | ||||||
4463 | cast<VPWidenPHIRecipe>(State.Plan->getVPValue(OrigPhi)); | ||||||
4464 | PHINode *NewPhi = cast<PHINode>(State.get(VPPhi, 0)); | ||||||
4465 | // Make sure the builder has a valid insert point. | ||||||
4466 | Builder.SetInsertPoint(NewPhi); | ||||||
4467 | for (unsigned i = 0; i < VPPhi->getNumOperands(); ++i) { | ||||||
4468 | VPValue *Inc = VPPhi->getIncomingValue(i); | ||||||
4469 | VPBasicBlock *VPBB = VPPhi->getIncomingBlock(i); | ||||||
4470 | NewPhi->addIncoming(State.get(Inc, 0), State.CFG.VPBB2IRBB[VPBB]); | ||||||
4471 | } | ||||||
4472 | } | ||||||
4473 | } | ||||||
4474 | |||||||
4475 | bool InnerLoopVectorizer::useOrderedReductions( | ||||||
4476 | const RecurrenceDescriptor &RdxDesc) { | ||||||
4477 | return Cost->useOrderedReductions(RdxDesc); | ||||||
4478 | } | ||||||
4479 | |||||||
4480 | void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN, | ||||||
4481 | VPWidenPHIRecipe *PhiR, | ||||||
4482 | VPTransformState &State) { | ||||||
4483 | PHINode *P = cast<PHINode>(PN); | ||||||
4484 | if (EnableVPlanNativePath) { | ||||||
4485 | // Currently we enter here in the VPlan-native path for non-induction | ||||||
4486 | // PHIs where all control flow is uniform. We simply widen these PHIs. | ||||||
4487 | // Create a vector phi with no operands - the vector phi operands will be | ||||||
4488 | // set at the end of vector code generation. | ||||||
4489 | Type *VecTy = (State.VF.isScalar()) | ||||||
4490 | ? PN->getType() | ||||||
4491 | : VectorType::get(PN->getType(), State.VF); | ||||||
4492 | Value *VecPhi = Builder.CreatePHI(VecTy, PN->getNumOperands(), "vec.phi"); | ||||||
4493 | State.set(PhiR, VecPhi, 0); | ||||||
4494 | OrigPHIsToFix.push_back(P); | ||||||
4495 | |||||||
4496 | return; | ||||||
4497 | } | ||||||
4498 | |||||||
4499 | assert(PN->getParent() == OrigLoop->getHeader() &&(static_cast <bool> (PN->getParent() == OrigLoop-> getHeader() && "Non-header phis should have been handled elsewhere" ) ? void (0) : __assert_fail ("PN->getParent() == OrigLoop->getHeader() && \"Non-header phis should have been handled elsewhere\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4500, __extension__ __PRETTY_FUNCTION__)) | ||||||
4500 | "Non-header phis should have been handled elsewhere")(static_cast <bool> (PN->getParent() == OrigLoop-> getHeader() && "Non-header phis should have been handled elsewhere" ) ? void (0) : __assert_fail ("PN->getParent() == OrigLoop->getHeader() && \"Non-header phis should have been handled elsewhere\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4500, __extension__ __PRETTY_FUNCTION__)); | ||||||
4501 | |||||||
4502 | // In order to support recurrences we need to be able to vectorize Phi nodes. | ||||||
4503 | // Phi nodes have cycles, so we need to vectorize them in two stages. This is | ||||||
4504 | // stage #1: We create a new vector PHI node with no incoming edges. We'll use | ||||||
4505 | // this value when we vectorize all of the instructions that use the PHI. | ||||||
4506 | |||||||
4507 | assert(!Legal->isReductionVariable(P) &&(static_cast <bool> (!Legal->isReductionVariable(P) && "reductions should be handled elsewhere") ? void (0) : __assert_fail ("!Legal->isReductionVariable(P) && \"reductions should be handled elsewhere\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4508, __extension__ __PRETTY_FUNCTION__)) | ||||||
4508 | "reductions should be handled elsewhere")(static_cast <bool> (!Legal->isReductionVariable(P) && "reductions should be handled elsewhere") ? void (0) : __assert_fail ("!Legal->isReductionVariable(P) && \"reductions should be handled elsewhere\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4508, __extension__ __PRETTY_FUNCTION__)); | ||||||
4509 | |||||||
4510 | setDebugLocFromInst(P); | ||||||
4511 | |||||||
4512 | // This PHINode must be an induction variable. | ||||||
4513 | // Make sure that we know about it. | ||||||
4514 | assert(Legal->getInductionVars().count(P) && "Not an induction variable")(static_cast <bool> (Legal->getInductionVars().count (P) && "Not an induction variable") ? void (0) : __assert_fail ("Legal->getInductionVars().count(P) && \"Not an induction variable\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4514, __extension__ __PRETTY_FUNCTION__)); | ||||||
4515 | |||||||
4516 | InductionDescriptor II = Legal->getInductionVars().lookup(P); | ||||||
4517 | const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout(); | ||||||
4518 | |||||||
4519 | auto *IVR = PhiR->getParent()->getPlan()->getCanonicalIV(); | ||||||
4520 | PHINode *CanonicalIV = cast<PHINode>(State.get(IVR, 0)); | ||||||
4521 | |||||||
4522 | // FIXME: The newly created binary instructions should contain nsw/nuw flags, | ||||||
4523 | // which can be found from the original scalar operations. | ||||||
4524 | switch (II.getKind()) { | ||||||
4525 | case InductionDescriptor::IK_NoInduction: | ||||||
4526 | llvm_unreachable("Unknown induction")::llvm::llvm_unreachable_internal("Unknown induction", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 4526); | ||||||
4527 | case InductionDescriptor::IK_IntInduction: | ||||||
4528 | case InductionDescriptor::IK_FpInduction: | ||||||
4529 | llvm_unreachable("Integer/fp induction is handled elsewhere.")::llvm::llvm_unreachable_internal("Integer/fp induction is handled elsewhere." , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4529); | ||||||
4530 | case InductionDescriptor::IK_PtrInduction: { | ||||||
4531 | // Handle the pointer induction variable case. | ||||||
4532 | assert(P->getType()->isPointerTy() && "Unexpected type.")(static_cast <bool> (P->getType()->isPointerTy() && "Unexpected type.") ? void (0) : __assert_fail ("P->getType()->isPointerTy() && \"Unexpected type.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4532, __extension__ __PRETTY_FUNCTION__)); | ||||||
4533 | |||||||
4534 | if (Cost->isScalarAfterVectorization(P, State.VF)) { | ||||||
4535 | // This is the normalized GEP that starts counting at zero. | ||||||
4536 | Value *PtrInd = | ||||||
4537 | Builder.CreateSExtOrTrunc(CanonicalIV, II.getStep()->getType()); | ||||||
4538 | // Determine the number of scalars we need to generate for each unroll | ||||||
4539 | // iteration. If the instruction is uniform, we only need to generate the | ||||||
4540 | // first lane. Otherwise, we generate all VF values. | ||||||
4541 | bool IsUniform = Cost->isUniformAfterVectorization(P, State.VF); | ||||||
4542 | 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", 4543, __extension__ __PRETTY_FUNCTION__)) | ||||||
4543 | "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", 4543, __extension__ __PRETTY_FUNCTION__)); | ||||||
4544 | unsigned Lanes = IsUniform ? 1 : State.VF.getFixedValue(); | ||||||
4545 | |||||||
4546 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
4547 | Value *PartStart = | ||||||
4548 | createStepForVF(Builder, PtrInd->getType(), VF, Part); | ||||||
4549 | |||||||
4550 | for (unsigned Lane = 0; Lane < Lanes; ++Lane) { | ||||||
4551 | Value *Idx = Builder.CreateAdd( | ||||||
4552 | PartStart, ConstantInt::get(PtrInd->getType(), Lane)); | ||||||
4553 | Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx); | ||||||
4554 | Value *SclrGep = emitTransformedIndex(Builder, GlobalIdx, PSE.getSE(), | ||||||
4555 | DL, II, State.CFG.PrevBB); | ||||||
4556 | SclrGep->setName("next.gep"); | ||||||
4557 | State.set(PhiR, SclrGep, VPIteration(Part, Lane)); | ||||||
4558 | } | ||||||
4559 | } | ||||||
4560 | return; | ||||||
4561 | } | ||||||
4562 | assert(isa<SCEVConstant>(II.getStep()) &&(static_cast <bool> (isa<SCEVConstant>(II.getStep ()) && "Induction step not a SCEV constant!") ? void ( 0) : __assert_fail ("isa<SCEVConstant>(II.getStep()) && \"Induction step not a SCEV constant!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4563, __extension__ __PRETTY_FUNCTION__)) | ||||||
4563 | "Induction step not a SCEV constant!")(static_cast <bool> (isa<SCEVConstant>(II.getStep ()) && "Induction step not a SCEV constant!") ? void ( 0) : __assert_fail ("isa<SCEVConstant>(II.getStep()) && \"Induction step not a SCEV constant!\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4563, __extension__ __PRETTY_FUNCTION__)); | ||||||
4564 | Type *PhiType = II.getStep()->getType(); | ||||||
4565 | |||||||
4566 | // Build a pointer phi | ||||||
4567 | Value *ScalarStartValue = PhiR->getStartValue()->getLiveInIRValue(); | ||||||
4568 | Type *ScStValueType = ScalarStartValue->getType(); | ||||||
4569 | PHINode *NewPointerPhi = | ||||||
4570 | PHINode::Create(ScStValueType, 2, "pointer.phi", CanonicalIV); | ||||||
4571 | NewPointerPhi->addIncoming(ScalarStartValue, LoopVectorPreHeader); | ||||||
4572 | |||||||
4573 | // A pointer induction, performed by using a gep | ||||||
4574 | BasicBlock *LoopLatch = LI->getLoopFor(LoopVectorBody)->getLoopLatch(); | ||||||
4575 | Instruction *InductionLoc = LoopLatch->getTerminator(); | ||||||
4576 | const SCEV *ScalarStep = II.getStep(); | ||||||
4577 | SCEVExpander Exp(*PSE.getSE(), DL, "induction"); | ||||||
4578 | Value *ScalarStepValue = | ||||||
4579 | Exp.expandCodeFor(ScalarStep, PhiType, InductionLoc); | ||||||
4580 | Value *RuntimeVF = getRuntimeVF(Builder, PhiType, VF); | ||||||
4581 | Value *NumUnrolledElems = | ||||||
4582 | Builder.CreateMul(RuntimeVF, ConstantInt::get(PhiType, State.UF)); | ||||||
4583 | Value *InductionGEP = GetElementPtrInst::Create( | ||||||
4584 | II.getElementType(), NewPointerPhi, | ||||||
4585 | Builder.CreateMul(ScalarStepValue, NumUnrolledElems), "ptr.ind", | ||||||
4586 | InductionLoc); | ||||||
4587 | NewPointerPhi->addIncoming(InductionGEP, LoopLatch); | ||||||
4588 | |||||||
4589 | // Create UF many actual address geps that use the pointer | ||||||
4590 | // phi as base and a vectorized version of the step value | ||||||
4591 | // (<step*0, ..., step*N>) as offset. | ||||||
4592 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
4593 | Type *VecPhiType = VectorType::get(PhiType, State.VF); | ||||||
4594 | Value *StartOffsetScalar = | ||||||
4595 | Builder.CreateMul(RuntimeVF, ConstantInt::get(PhiType, Part)); | ||||||
4596 | Value *StartOffset = | ||||||
4597 | Builder.CreateVectorSplat(State.VF, StartOffsetScalar); | ||||||
4598 | // Create a vector of consecutive numbers from zero to VF. | ||||||
4599 | StartOffset = | ||||||
4600 | Builder.CreateAdd(StartOffset, Builder.CreateStepVector(VecPhiType)); | ||||||
4601 | |||||||
4602 | Value *GEP = Builder.CreateGEP( | ||||||
4603 | II.getElementType(), NewPointerPhi, | ||||||
4604 | Builder.CreateMul( | ||||||
4605 | StartOffset, Builder.CreateVectorSplat(State.VF, ScalarStepValue), | ||||||
4606 | "vector.gep")); | ||||||
4607 | State.set(PhiR, GEP, Part); | ||||||
4608 | } | ||||||
4609 | } | ||||||
4610 | } | ||||||
4611 | } | ||||||
4612 | |||||||
4613 | /// A helper function for checking whether an integer division-related | ||||||
4614 | /// instruction may divide by zero (in which case it must be predicated if | ||||||
4615 | /// executed conditionally in the scalar code). | ||||||
4616 | /// TODO: It may be worthwhile to generalize and check isKnownNonZero(). | ||||||
4617 | /// Non-zero divisors that are non compile-time constants will not be | ||||||
4618 | /// converted into multiplication, so we will still end up scalarizing | ||||||
4619 | /// the division, but can do so w/o predication. | ||||||
4620 | static bool mayDivideByZero(Instruction &I) { | ||||||
4621 | assert((I.getOpcode() == Instruction::UDiv ||(static_cast <bool> ((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction ::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction" ) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4625, __extension__ __PRETTY_FUNCTION__)) | ||||||
4622 | I.getOpcode() == Instruction::SDiv ||(static_cast <bool> ((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction ::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction" ) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4625, __extension__ __PRETTY_FUNCTION__)) | ||||||
4623 | I.getOpcode() == Instruction::URem ||(static_cast <bool> ((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction ::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction" ) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4625, __extension__ __PRETTY_FUNCTION__)) | ||||||
4624 | I.getOpcode() == Instruction::SRem) &&(static_cast <bool> ((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction ::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction" ) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4625, __extension__ __PRETTY_FUNCTION__)) | ||||||
4625 | "Unexpected instruction")(static_cast <bool> ((I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction ::URem || I.getOpcode() == Instruction::SRem) && "Unexpected instruction" ) ? void (0) : __assert_fail ("(I.getOpcode() == Instruction::UDiv || I.getOpcode() == Instruction::SDiv || I.getOpcode() == Instruction::URem || I.getOpcode() == Instruction::SRem) && \"Unexpected instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4625, __extension__ __PRETTY_FUNCTION__)); | ||||||
4626 | Value *Divisor = I.getOperand(1); | ||||||
4627 | auto *CInt = dyn_cast<ConstantInt>(Divisor); | ||||||
4628 | return !CInt || CInt->isZero(); | ||||||
4629 | } | ||||||
4630 | |||||||
4631 | void InnerLoopVectorizer::widenCallInstruction(CallInst &I, VPValue *Def, | ||||||
4632 | VPUser &ArgOperands, | ||||||
4633 | VPTransformState &State) { | ||||||
4634 | assert(!isa<DbgInfoIntrinsic>(I) &&(static_cast <bool> (!isa<DbgInfoIntrinsic>(I) && "DbgInfoIntrinsic should have been dropped during VPlan construction" ) ? void (0) : __assert_fail ("!isa<DbgInfoIntrinsic>(I) && \"DbgInfoIntrinsic should have been dropped during VPlan construction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4635, __extension__ __PRETTY_FUNCTION__)) | ||||||
4635 | "DbgInfoIntrinsic should have been dropped during VPlan construction")(static_cast <bool> (!isa<DbgInfoIntrinsic>(I) && "DbgInfoIntrinsic should have been dropped during VPlan construction" ) ? void (0) : __assert_fail ("!isa<DbgInfoIntrinsic>(I) && \"DbgInfoIntrinsic should have been dropped during VPlan construction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4635, __extension__ __PRETTY_FUNCTION__)); | ||||||
4636 | setDebugLocFromInst(&I); | ||||||
4637 | |||||||
4638 | Module *M = I.getParent()->getParent()->getParent(); | ||||||
4639 | auto *CI = cast<CallInst>(&I); | ||||||
4640 | |||||||
4641 | SmallVector<Type *, 4> Tys; | ||||||
4642 | for (Value *ArgOperand : CI->args()) | ||||||
4643 | Tys.push_back(ToVectorTy(ArgOperand->getType(), VF.getKnownMinValue())); | ||||||
4644 | |||||||
4645 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||||
4646 | |||||||
4647 | // The flag shows whether we use Intrinsic or a usual Call for vectorized | ||||||
4648 | // version of the instruction. | ||||||
4649 | // Is it beneficial to perform intrinsic call compared to lib call? | ||||||
4650 | bool NeedToScalarize = false; | ||||||
4651 | InstructionCost CallCost = Cost->getVectorCallCost(CI, VF, NeedToScalarize); | ||||||
4652 | InstructionCost IntrinsicCost = ID ? Cost->getVectorIntrinsicCost(CI, VF) : 0; | ||||||
4653 | bool UseVectorIntrinsic = ID && IntrinsicCost <= CallCost; | ||||||
4654 | assert((UseVectorIntrinsic || !NeedToScalarize) &&(static_cast <bool> ((UseVectorIntrinsic || !NeedToScalarize ) && "Instruction should be scalarized elsewhere.") ? void (0) : __assert_fail ("(UseVectorIntrinsic || !NeedToScalarize) && \"Instruction should be scalarized elsewhere.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4655, __extension__ __PRETTY_FUNCTION__)) | ||||||
4655 | "Instruction should be scalarized elsewhere.")(static_cast <bool> ((UseVectorIntrinsic || !NeedToScalarize ) && "Instruction should be scalarized elsewhere.") ? void (0) : __assert_fail ("(UseVectorIntrinsic || !NeedToScalarize) && \"Instruction should be scalarized elsewhere.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4655, __extension__ __PRETTY_FUNCTION__)); | ||||||
4656 | assert((IntrinsicCost.isValid() || CallCost.isValid()) &&(static_cast <bool> ((IntrinsicCost.isValid() || CallCost .isValid()) && "Either the intrinsic cost or vector call cost must be valid" ) ? void (0) : __assert_fail ("(IntrinsicCost.isValid() || CallCost.isValid()) && \"Either the intrinsic cost or vector call cost must be valid\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4657, __extension__ __PRETTY_FUNCTION__)) | ||||||
4657 | "Either the intrinsic cost or vector call cost must be valid")(static_cast <bool> ((IntrinsicCost.isValid() || CallCost .isValid()) && "Either the intrinsic cost or vector call cost must be valid" ) ? void (0) : __assert_fail ("(IntrinsicCost.isValid() || CallCost.isValid()) && \"Either the intrinsic cost or vector call cost must be valid\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4657, __extension__ __PRETTY_FUNCTION__)); | ||||||
4658 | |||||||
4659 | for (unsigned Part = 0; Part < UF; ++Part) { | ||||||
4660 | SmallVector<Type *, 2> TysForDecl = {CI->getType()}; | ||||||
4661 | SmallVector<Value *, 4> Args; | ||||||
4662 | for (auto &I : enumerate(ArgOperands.operands())) { | ||||||
4663 | // Some intrinsics have a scalar argument - don't replace it with a | ||||||
4664 | // vector. | ||||||
4665 | Value *Arg; | ||||||
4666 | if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, I.index())) | ||||||
4667 | Arg = State.get(I.value(), Part); | ||||||
4668 | else { | ||||||
4669 | Arg = State.get(I.value(), VPIteration(0, 0)); | ||||||
4670 | if (hasVectorInstrinsicOverloadedScalarOpd(ID, I.index())) | ||||||
4671 | TysForDecl.push_back(Arg->getType()); | ||||||
4672 | } | ||||||
4673 | Args.push_back(Arg); | ||||||
4674 | } | ||||||
4675 | |||||||
4676 | Function *VectorF; | ||||||
4677 | if (UseVectorIntrinsic) { | ||||||
4678 | // Use vector version of the intrinsic. | ||||||
4679 | if (VF.isVector()) | ||||||
4680 | TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF); | ||||||
4681 | VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl); | ||||||
4682 | assert(VectorF && "Can't retrieve vector intrinsic.")(static_cast <bool> (VectorF && "Can't retrieve vector intrinsic." ) ? void (0) : __assert_fail ("VectorF && \"Can't retrieve vector intrinsic.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4682, __extension__ __PRETTY_FUNCTION__)); | ||||||
4683 | } else { | ||||||
4684 | // Use vector version of the function call. | ||||||
4685 | const VFShape Shape = VFShape::get(*CI, VF, false /*HasGlobalPred*/); | ||||||
4686 | #ifndef NDEBUG | ||||||
4687 | assert(VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr &&(static_cast <bool> (VFDatabase(*CI).getVectorizedFunction (Shape) != nullptr && "Can't create vector function." ) ? void (0) : __assert_fail ("VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr && \"Can't create vector function.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4688, __extension__ __PRETTY_FUNCTION__)) | ||||||
4688 | "Can't create vector function.")(static_cast <bool> (VFDatabase(*CI).getVectorizedFunction (Shape) != nullptr && "Can't create vector function." ) ? void (0) : __assert_fail ("VFDatabase(*CI).getVectorizedFunction(Shape) != nullptr && \"Can't create vector function.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 4688, __extension__ __PRETTY_FUNCTION__)); | ||||||
4689 | #endif | ||||||
4690 | VectorF = VFDatabase(*CI).getVectorizedFunction(Shape); | ||||||
4691 | } | ||||||
4692 | SmallVector<OperandBundleDef, 1> OpBundles; | ||||||
4693 | CI->getOperandBundlesAsDefs(OpBundles); | ||||||
4694 | CallInst *V = Builder.CreateCall(VectorF, Args, OpBundles); | ||||||
4695 | |||||||
4696 | if (isa<FPMathOperator>(V)) | ||||||
4697 | V->copyFastMathFlags(CI); | ||||||
4698 | |||||||
4699 | State.set(Def, V, Part); | ||||||
4700 | addMetadata(V, &I); | ||||||
4701 | } | ||||||
4702 | } | ||||||
4703 | |||||||
4704 | void LoopVectorizationCostModel::collectLoopScalars(ElementCount VF) { | ||||||
4705 | // We should not collect Scalars more than once per VF. Right now, this | ||||||
4706 | // function is called from collectUniformsAndScalars(), which already does | ||||||
4707 | // this check. Collecting Scalars for VF=1 does not make any sense. | ||||||
4708 | 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", 4709, __extension__ __PRETTY_FUNCTION__)) | ||||||
4709 | "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", 4709, __extension__ __PRETTY_FUNCTION__)); | ||||||
4710 | |||||||
4711 | SmallSetVector<Instruction *, 8> Worklist; | ||||||
4712 | |||||||
4713 | // These sets are used to seed the analysis with pointers used by memory | ||||||
4714 | // accesses that will remain scalar. | ||||||
4715 | SmallSetVector<Instruction *, 8> ScalarPtrs; | ||||||
4716 | SmallPtrSet<Instruction *, 8> PossibleNonScalarPtrs; | ||||||
4717 | auto *Latch = TheLoop->getLoopLatch(); | ||||||
4718 | |||||||
4719 | // A helper that returns true if the use of Ptr by MemAccess will be scalar. | ||||||
4720 | // The pointer operands of loads and stores will be scalar as long as the | ||||||
4721 | // memory access is not a gather or scatter operation. The value operand of a | ||||||
4722 | // store will remain scalar if the store is scalarized. | ||||||
4723 | auto isScalarUse = [&](Instruction *MemAccess, Value *Ptr) { | ||||||
4724 | InstWidening WideningDecision = getWideningDecision(MemAccess, VF); | ||||||
4725 | 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", 4726, __extension__ __PRETTY_FUNCTION__)) | ||||||
4726 | "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", 4726, __extension__ __PRETTY_FUNCTION__)); | ||||||
4727 | if (auto *Store = dyn_cast<StoreInst>(MemAccess)) | ||||||
4728 | if (Ptr == Store->getValueOperand()) | ||||||
4729 | return WideningDecision == CM_Scalarize; | ||||||
4730 | 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", 4731, __extension__ __PRETTY_FUNCTION__)) | ||||||
4731 | "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", 4731, __extension__ __PRETTY_FUNCTION__)); | ||||||
4732 | return WideningDecision != CM_GatherScatter; | ||||||
4733 | }; | ||||||
4734 | |||||||
4735 | // A helper that returns true if the given value is a bitcast or | ||||||
4736 | // getelementptr instruction contained in the loop. | ||||||
4737 | auto isLoopVaryingBitCastOrGEP = [&](Value *V) { | ||||||
4738 | return ((isa<BitCastInst>(V) && V->getType()->isPointerTy()) || | ||||||
4739 | isa<GetElementPtrInst>(V)) && | ||||||
4740 | !TheLoop->isLoopInvariant(V); | ||||||
4741 | }; | ||||||
4742 | |||||||
4743 | // A helper that evaluates a memory access's use of a pointer. If the use will | ||||||
4744 | // be a scalar use and the pointer is only used by memory accesses, we place | ||||||
4745 | // the pointer in ScalarPtrs. Otherwise, the pointer is placed in | ||||||
4746 | // PossibleNonScalarPtrs. | ||||||
4747 | auto evaluatePtrUse = [&](Instruction *MemAccess, Value *Ptr) { | ||||||
4748 | // We only care about bitcast and getelementptr instructions contained in | ||||||
4749 | // the loop. | ||||||
4750 | if (!isLoopVaryingBitCastOrGEP(Ptr)) | ||||||
4751 | return; | ||||||
4752 | |||||||
4753 | // If the pointer has already been identified as scalar (e.g., if it was | ||||||
4754 | // also identified as uniform), there's nothing to do. | ||||||
4755 | auto *I = cast<Instruction>(Ptr); | ||||||
4756 | if (Worklist.count(I)) | ||||||
4757 | return; | ||||||
4758 | |||||||
4759 | // If the use of the pointer will be a scalar use, and all users of the | ||||||
4760 | // pointer are memory accesses, place the pointer in ScalarPtrs. Otherwise, | ||||||
4761 | // place the pointer in PossibleNonScalarPtrs. | ||||||
4762 | if (isScalarUse(MemAccess, Ptr) && llvm::all_of(I->users(), [&](User *U) { | ||||||
4763 | return isa<LoadInst>(U) || isa<StoreInst>(U); | ||||||
4764 | })) | ||||||
4765 | ScalarPtrs.insert(I); | ||||||
4766 | else | ||||||
4767 | PossibleNonScalarPtrs.insert(I); | ||||||
4768 | }; | ||||||
4769 | |||||||
4770 | // We seed the scalars analysis with three classes of instructions: (1) | ||||||
4771 | // instructions marked uniform-after-vectorization and (2) bitcast, | ||||||
4772 | // getelementptr and (pointer) phi instructions used by memory accesses | ||||||
4773 | // requiring a scalar use. | ||||||
4774 | // | ||||||
4775 | // (1) Add to the worklist all instructions that have been identified as | ||||||
4776 | // uniform-after-vectorization. | ||||||
4777 | Worklist.insert(Uniforms[VF].begin(), Uniforms[VF].end()); | ||||||
4778 | |||||||
4779 | // (2) Add to the worklist all bitcast and getelementptr instructions used by | ||||||
4780 | // memory accesses requiring a scalar use. The pointer operands of loads and | ||||||
4781 | // stores will be scalar as long as the memory accesses is not a gather or | ||||||
4782 | // scatter operation. The value operand of a store will remain scalar if the | ||||||
4783 | // store is scalarized. | ||||||
4784 | for (auto *BB : TheLoop->blocks()) | ||||||
4785 | for (auto &I : *BB) { | ||||||
4786 | if (auto *Load = dyn_cast<LoadInst>(&I)) { | ||||||
4787 | evaluatePtrUse(Load, Load->getPointerOperand()); | ||||||
4788 | } else if (auto *Store = dyn_cast<StoreInst>(&I)) { | ||||||
4789 | evaluatePtrUse(Store, Store->getPointerOperand()); | ||||||
4790 | evaluatePtrUse(Store, Store->getValueOperand()); | ||||||
4791 | } | ||||||
4792 | } | ||||||
4793 | for (auto *I : ScalarPtrs) | ||||||
4794 | if (!PossibleNonScalarPtrs.count(I)) { | ||||||
4795 | 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); | ||||||
4796 | Worklist.insert(I); | ||||||
4797 | } | ||||||
4798 | |||||||
4799 | // Insert the forced scalars. | ||||||
4800 | // FIXME: Currently widenPHIInstruction() often creates a dead vector | ||||||
4801 | // induction variable when the PHI user is scalarized. | ||||||
4802 | auto ForcedScalar = ForcedScalars.find(VF); | ||||||
4803 | if (ForcedScalar != ForcedScalars.end()) | ||||||
4804 | for (auto *I : ForcedScalar->second) | ||||||
4805 | Worklist.insert(I); | ||||||
4806 | |||||||
4807 | // Expand the worklist by looking through any bitcasts and getelementptr | ||||||
4808 | // instructions we've already identified as scalar. This is similar to the | ||||||
4809 | // expansion step in collectLoopUniforms(); however, here we're only | ||||||
4810 | // expanding to include additional bitcasts and getelementptr instructions. | ||||||
4811 | unsigned Idx = 0; | ||||||
4812 | while (Idx != Worklist.size()) { | ||||||
4813 | Instruction *Dst = Worklist[Idx++]; | ||||||
4814 | if (!isLoopVaryingBitCastOrGEP(Dst->getOperand(0))) | ||||||
4815 | continue; | ||||||
4816 | auto *Src = cast<Instruction>(Dst->getOperand(0)); | ||||||
4817 | if (llvm::all_of(Src->users(), [&](User *U) -> bool { | ||||||
4818 | auto *J = cast<Instruction>(U); | ||||||
4819 | return !TheLoop->contains(J) || Worklist.count(J) || | ||||||
4820 | ((isa<LoadInst>(J) || isa<StoreInst>(J)) && | ||||||
4821 | isScalarUse(J, Src)); | ||||||
4822 | })) { | ||||||
4823 | Worklist.insert(Src); | ||||||
4824 | 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); | ||||||
4825 | } | ||||||
4826 | } | ||||||
4827 | |||||||
4828 | // An induction variable will remain scalar if all users of the induction | ||||||
4829 | // variable and induction variable update remain scalar. | ||||||
4830 | for (auto &Induction : Legal->getInductionVars()) { | ||||||
4831 | auto *Ind = Induction.first; | ||||||
4832 | auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); | ||||||
4833 | |||||||
4834 | // If tail-folding is applied, the primary induction variable will be used | ||||||
4835 | // to feed a vector compare. | ||||||
4836 | if (Ind == Legal->getPrimaryInduction() && foldTailByMasking()) | ||||||
4837 | continue; | ||||||
4838 | |||||||
4839 | // Returns true if \p Indvar is a pointer induction that is used directly by | ||||||
4840 | // load/store instruction \p I. | ||||||
4841 | auto IsDirectLoadStoreFromPtrIndvar = [&](Instruction *Indvar, | ||||||
4842 | Instruction *I) { | ||||||
4843 | return Induction.second.getKind() == | ||||||
4844 | InductionDescriptor::IK_PtrInduction && | ||||||
4845 | (isa<LoadInst>(I) || isa<StoreInst>(I)) && | ||||||
4846 | Indvar == getLoadStorePointerOperand(I) && isScalarUse(I, Indvar); | ||||||
4847 | }; | ||||||
4848 | |||||||
4849 | // Determine if all users of the induction variable are scalar after | ||||||
4850 | // vectorization. | ||||||
4851 | auto ScalarInd = llvm::all_of(Ind->users(), [&](User *U) -> bool { | ||||||
4852 | auto *I = cast<Instruction>(U); | ||||||
4853 | return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) || | ||||||
4854 | IsDirectLoadStoreFromPtrIndvar(Ind, I); | ||||||
4855 | }); | ||||||
4856 | if (!ScalarInd) | ||||||
4857 | continue; | ||||||
4858 | |||||||
4859 | // Determine if all users of the induction variable update instruction are | ||||||
4860 | // scalar after vectorization. | ||||||
4861 | auto ScalarIndUpdate = | ||||||
4862 | llvm::all_of(IndUpdate->users(), [&](User *U) -> bool { | ||||||
4863 | auto *I = cast<Instruction>(U); | ||||||
4864 | return I == Ind || !TheLoop->contains(I) || Worklist.count(I) || | ||||||
4865 | IsDirectLoadStoreFromPtrIndvar(IndUpdate, I); | ||||||
4866 | }); | ||||||
4867 | if (!ScalarIndUpdate) | ||||||
4868 | continue; | ||||||
4869 | |||||||
4870 | // The induction variable and its update instruction will remain scalar. | ||||||
4871 | Worklist.insert(Ind); | ||||||
4872 | Worklist.insert(IndUpdate); | ||||||
4873 | 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); | ||||||
4874 | LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *IndUpdatedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n"; } } while (false) | ||||||
4875 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n"; } } while (false); | ||||||
4876 | } | ||||||
4877 | |||||||
4878 | Scalars[VF].insert(Worklist.begin(), Worklist.end()); | ||||||
4879 | } | ||||||
4880 | |||||||
4881 | bool LoopVectorizationCostModel::isScalarWithPredication( | ||||||
4882 | Instruction *I, ElementCount VF) const { | ||||||
4883 | if (!blockNeedsPredicationForAnyReason(I->getParent())) | ||||||
4884 | return false; | ||||||
4885 | switch(I->getOpcode()) { | ||||||
4886 | default: | ||||||
4887 | break; | ||||||
4888 | case Instruction::Load: | ||||||
4889 | case Instruction::Store: { | ||||||
4890 | if (!Legal->isMaskRequired(I)) | ||||||
4891 | return false; | ||||||
4892 | auto *Ptr = getLoadStorePointerOperand(I); | ||||||
4893 | auto *Ty = getLoadStoreType(I); | ||||||
4894 | Type *VTy = Ty; | ||||||
4895 | if (VF.isVector()) | ||||||
4896 | VTy = VectorType::get(Ty, VF); | ||||||
4897 | const Align Alignment = getLoadStoreAlignment(I); | ||||||
4898 | return isa<LoadInst>(I) ? !(isLegalMaskedLoad(Ty, Ptr, Alignment) || | ||||||
4899 | TTI.isLegalMaskedGather(VTy, Alignment)) | ||||||
4900 | : !(isLegalMaskedStore(Ty, Ptr, Alignment) || | ||||||
4901 | TTI.isLegalMaskedScatter(VTy, Alignment)); | ||||||
4902 | } | ||||||
4903 | case Instruction::UDiv: | ||||||
4904 | case Instruction::SDiv: | ||||||
4905 | case Instruction::SRem: | ||||||
4906 | case Instruction::URem: | ||||||
4907 | return mayDivideByZero(*I); | ||||||
4908 | } | ||||||
4909 | return false; | ||||||
4910 | } | ||||||
4911 | |||||||
4912 | bool LoopVectorizationCostModel::interleavedAccessCanBeWidened( | ||||||
4913 | Instruction *I, ElementCount VF) { | ||||||
4914 | 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", 4914, __extension__ __PRETTY_FUNCTION__)); | ||||||
4915 | 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", 4916, __extension__ __PRETTY_FUNCTION__)) | ||||||
4916 | "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", 4916, __extension__ __PRETTY_FUNCTION__)); | ||||||
4917 | auto *Group = getInterleavedAccessGroup(I); | ||||||
4918 | 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", 4918, __extension__ __PRETTY_FUNCTION__)); | ||||||
4919 | |||||||
4920 | // If the instruction's allocated size doesn't equal it's type size, it | ||||||
4921 | // requires padding and will be scalarized. | ||||||
4922 | auto &DL = I->getModule()->getDataLayout(); | ||||||
4923 | auto *ScalarTy = getLoadStoreType(I); | ||||||
4924 | if (hasIrregularType(ScalarTy, DL)) | ||||||
4925 | return false; | ||||||
4926 | |||||||
4927 | // Check if masking is required. | ||||||
4928 | // A Group may need masking for one of two reasons: it resides in a block that | ||||||
4929 | // needs predication, or it was decided to use masking to deal with gaps | ||||||
4930 | // (either a gap at the end of a load-access that may result in a speculative | ||||||
4931 | // load, or any gaps in a store-access). | ||||||
4932 | bool PredicatedAccessRequiresMasking = | ||||||
4933 | blockNeedsPredicationForAnyReason(I->getParent()) && | ||||||
4934 | Legal->isMaskRequired(I); | ||||||
4935 | bool LoadAccessWithGapsRequiresEpilogMasking = | ||||||
4936 | isa<LoadInst>(I) && Group->requiresScalarEpilogue() && | ||||||
4937 | !isScalarEpilogueAllowed(); | ||||||
4938 | bool StoreAccessWithGapsRequiresMasking = | ||||||
4939 | isa<StoreInst>(I) && (Group->getNumMembers() < Group->getFactor()); | ||||||
4940 | if (!PredicatedAccessRequiresMasking && | ||||||
4941 | !LoadAccessWithGapsRequiresEpilogMasking && | ||||||
4942 | !StoreAccessWithGapsRequiresMasking) | ||||||
4943 | return true; | ||||||
4944 | |||||||
4945 | // If masked interleaving is required, we expect that the user/target had | ||||||
4946 | // enabled it, because otherwise it either wouldn't have been created or | ||||||
4947 | // it should have been invalidated by the CostModel. | ||||||
4948 | 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", 4949, __extension__ __PRETTY_FUNCTION__)) | ||||||
4949 | "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", 4949, __extension__ __PRETTY_FUNCTION__)); | ||||||
4950 | |||||||
4951 | if (Group->isReverse()) | ||||||
4952 | return false; | ||||||
4953 | |||||||
4954 | auto *Ty = getLoadStoreType(I); | ||||||
4955 | const Align Alignment = getLoadStoreAlignment(I); | ||||||
4956 | return isa<LoadInst>(I) ? TTI.isLegalMaskedLoad(Ty, Alignment) | ||||||
4957 | : TTI.isLegalMaskedStore(Ty, Alignment); | ||||||
4958 | } | ||||||
4959 | |||||||
4960 | bool LoopVectorizationCostModel::memoryInstructionCanBeWidened( | ||||||
4961 | Instruction *I, ElementCount VF) { | ||||||
4962 | // Get and ensure we have a valid memory instruction. | ||||||
4963 | 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", 4963, __extension__ __PRETTY_FUNCTION__)); | ||||||
4964 | |||||||
4965 | auto *Ptr = getLoadStorePointerOperand(I); | ||||||
4966 | auto *ScalarTy = getLoadStoreType(I); | ||||||
4967 | |||||||
4968 | // In order to be widened, the pointer should be consecutive, first of all. | ||||||
4969 | if (!Legal->isConsecutivePtr(ScalarTy, Ptr)) | ||||||
4970 | return false; | ||||||
4971 | |||||||
4972 | // If the instruction is a store located in a predicated block, it will be | ||||||
4973 | // scalarized. | ||||||
4974 | if (isScalarWithPredication(I, VF)) | ||||||
4975 | return false; | ||||||
4976 | |||||||
4977 | // If the instruction's allocated size doesn't equal it's type size, it | ||||||
4978 | // requires padding and will be scalarized. | ||||||
4979 | auto &DL = I->getModule()->getDataLayout(); | ||||||
4980 | if (hasIrregularType(ScalarTy, DL)) | ||||||
4981 | return false; | ||||||
4982 | |||||||
4983 | return true; | ||||||
4984 | } | ||||||
4985 | |||||||
4986 | void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) { | ||||||
4987 | // We should not collect Uniforms more than once per VF. Right now, | ||||||
4988 | // this function is called from collectUniformsAndScalars(), which | ||||||
4989 | // already does this check. Collecting Uniforms for VF=1 does not make any | ||||||
4990 | // sense. | ||||||
4991 | |||||||
4992 | 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", 4993, __extension__ __PRETTY_FUNCTION__)) | ||||||
4993 | "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", 4993, __extension__ __PRETTY_FUNCTION__)); | ||||||
4994 | |||||||
4995 | // Visit the list of Uniforms. If we'll not find any uniform value, we'll | ||||||
4996 | // not analyze again. Uniforms.count(VF) will return 1. | ||||||
4997 | Uniforms[VF].clear(); | ||||||
4998 | |||||||
4999 | // We now know that the loop is vectorizable! | ||||||
5000 | // Collect instructions inside the loop that will remain uniform after | ||||||
5001 | // vectorization. | ||||||
5002 | |||||||
5003 | // Global values, params and instructions outside of current loop are out of | ||||||
5004 | // scope. | ||||||
5005 | auto isOutOfScope = [&](Value *V) -> bool { | ||||||
5006 | Instruction *I = dyn_cast<Instruction>(V); | ||||||
5007 | return (!I || !TheLoop->contains(I)); | ||||||
5008 | }; | ||||||
5009 | |||||||
5010 | // Worklist containing uniform instructions demanding lane 0. | ||||||
5011 | SetVector<Instruction *> Worklist; | ||||||
5012 | BasicBlock *Latch = TheLoop->getLoopLatch(); | ||||||
5013 | |||||||
5014 | // Add uniform instructions demanding lane 0 to the worklist. Instructions | ||||||
5015 | // that are scalar with predication must not be considered uniform after | ||||||
5016 | // vectorization, because that would create an erroneous replicating region | ||||||
5017 | // where only a single instance out of VF should be formed. | ||||||
5018 | // TODO: optimize such seldom cases if found important, see PR40816. | ||||||
5019 | auto addToWorklistIfAllowed = [&](Instruction *I) -> void { | ||||||
5020 | if (isOutOfScope(I)) { | ||||||
5021 | 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) | ||||||
5022 | << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found not uniform due to scope: " << *I << "\n"; } } while (false); | ||||||
5023 | return; | ||||||
5024 | } | ||||||
5025 | if (isScalarWithPredication(I, VF)) { | ||||||
5026 | 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) | ||||||
5027 | << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found not uniform being ScalarWithPredication: " << *I << "\n"; } } while (false); | ||||||
5028 | return; | ||||||
5029 | } | ||||||
5030 | 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); | ||||||
5031 | Worklist.insert(I); | ||||||
5032 | }; | ||||||
5033 | |||||||
5034 | // Start with the conditional branch. If the branch condition is an | ||||||
5035 | // instruction contained in the loop that is only used by the branch, it is | ||||||
5036 | // uniform. | ||||||
5037 | auto *Cmp = dyn_cast<Instruction>(Latch->getTerminator()->getOperand(0)); | ||||||
5038 | if (Cmp && TheLoop->contains(Cmp) && Cmp->hasOneUse()) | ||||||
5039 | addToWorklistIfAllowed(Cmp); | ||||||
5040 | |||||||
5041 | auto isUniformDecision = [&](Instruction *I, ElementCount VF) { | ||||||
5042 | InstWidening WideningDecision = getWideningDecision(I, VF); | ||||||
5043 | 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", 5044, __extension__ __PRETTY_FUNCTION__)) | ||||||
5044 | "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", 5044, __extension__ __PRETTY_FUNCTION__)); | ||||||
5045 | |||||||
5046 | // A uniform memory op is itself uniform. We exclude uniform stores | ||||||
5047 | // here as they demand the last lane, not the first one. | ||||||
5048 | if (isa<LoadInst>(I) && Legal->isUniformMemOp(*I)) { | ||||||
5049 | assert(WideningDecision == CM_Scalarize)(static_cast <bool> (WideningDecision == CM_Scalarize) ? void (0) : __assert_fail ("WideningDecision == CM_Scalarize" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 5049, __extension__ __PRETTY_FUNCTION__)); | ||||||
5050 | return true; | ||||||
5051 | } | ||||||
5052 | |||||||
5053 | return (WideningDecision == CM_Widen || | ||||||
5054 | WideningDecision == CM_Widen_Reverse || | ||||||
5055 | WideningDecision == CM_Interleave); | ||||||
5056 | }; | ||||||
5057 | |||||||
5058 | |||||||
5059 | // Returns true if Ptr is the pointer operand of a memory access instruction | ||||||
5060 | // I, and I is known to not require scalarization. | ||||||
5061 | auto isVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool { | ||||||
5062 | return getLoadStorePointerOperand(I) == Ptr && isUniformDecision(I, VF); | ||||||
5063 | }; | ||||||
5064 | |||||||
5065 | // Holds a list of values which are known to have at least one uniform use. | ||||||
5066 | // Note that there may be other uses which aren't uniform. A "uniform use" | ||||||
5067 | // here is something which only demands lane 0 of the unrolled iterations; | ||||||
5068 | // it does not imply that all lanes produce the same value (e.g. this is not | ||||||
5069 | // the usual meaning of uniform) | ||||||
5070 | SetVector<Value *> HasUniformUse; | ||||||
5071 | |||||||
5072 | // Scan the loop for instructions which are either a) known to have only | ||||||
5073 | // lane 0 demanded or b) are uses which demand only lane 0 of their operand. | ||||||
5074 | for (auto *BB : TheLoop->blocks()) | ||||||
5075 | for (auto &I : *BB) { | ||||||
5076 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I)) { | ||||||
5077 | switch (II->getIntrinsicID()) { | ||||||
5078 | case Intrinsic::sideeffect: | ||||||
5079 | case Intrinsic::experimental_noalias_scope_decl: | ||||||
5080 | case Intrinsic::assume: | ||||||
5081 | case Intrinsic::lifetime_start: | ||||||
5082 | case Intrinsic::lifetime_end: | ||||||
5083 | if (TheLoop->hasLoopInvariantOperands(&I)) | ||||||
5084 | addToWorklistIfAllowed(&I); | ||||||
5085 | break; | ||||||
5086 | default: | ||||||
5087 | break; | ||||||
5088 | } | ||||||
5089 | } | ||||||
5090 | |||||||
5091 | // ExtractValue instructions must be uniform, because the operands are | ||||||
5092 | // known to be loop-invariant. | ||||||
5093 | if (auto *EVI = dyn_cast<ExtractValueInst>(&I)) { | ||||||
5094 | 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", 5095, __extension__ __PRETTY_FUNCTION__)) | ||||||
5095 | "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", 5095, __extension__ __PRETTY_FUNCTION__)); | ||||||
5096 | addToWorklistIfAllowed(EVI); | ||||||
5097 | continue; | ||||||
5098 | } | ||||||
5099 | |||||||
5100 | // If there's no pointer operand, there's nothing to do. | ||||||
5101 | auto *Ptr = getLoadStorePointerOperand(&I); | ||||||
5102 | if (!Ptr) | ||||||
5103 | continue; | ||||||
5104 | |||||||
5105 | // A uniform memory op is itself uniform. We exclude uniform stores | ||||||
5106 | // here as they demand the last lane, not the first one. | ||||||
5107 | if (isa<LoadInst>(I) && Legal->isUniformMemOp(I)) | ||||||
5108 | addToWorklistIfAllowed(&I); | ||||||
5109 | |||||||
5110 | if (isUniformDecision(&I, VF)) { | ||||||
5111 | 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", 5111, __extension__ __PRETTY_FUNCTION__)); | ||||||
5112 | HasUniformUse.insert(Ptr); | ||||||
5113 | } | ||||||
5114 | } | ||||||
5115 | |||||||
5116 | // Add to the worklist any operands which have *only* uniform (e.g. lane 0 | ||||||
5117 | // demanding) users. Since loops are assumed to be in LCSSA form, this | ||||||
5118 | // disallows uses outside the loop as well. | ||||||
5119 | for (auto *V : HasUniformUse) { | ||||||
5120 | if (isOutOfScope(V)) | ||||||
5121 | continue; | ||||||
5122 | auto *I = cast<Instruction>(V); | ||||||
5123 | auto UsersAreMemAccesses = | ||||||
5124 | llvm::all_of(I->users(), [&](User *U) -> bool { | ||||||
5125 | return isVectorizedMemAccessUse(cast<Instruction>(U), V); | ||||||
5126 | }); | ||||||
5127 | if (UsersAreMemAccesses) | ||||||
5128 | addToWorklistIfAllowed(I); | ||||||
5129 | } | ||||||
5130 | |||||||
5131 | // Expand Worklist in topological order: whenever a new instruction | ||||||
5132 | // is added , its users should be already inside Worklist. It ensures | ||||||
5133 | // a uniform instruction will only be used by uniform instructions. | ||||||
5134 | unsigned idx = 0; | ||||||
5135 | while (idx != Worklist.size()) { | ||||||
5136 | Instruction *I = Worklist[idx++]; | ||||||
5137 | |||||||
5138 | for (auto OV : I->operand_values()) { | ||||||
5139 | // isOutOfScope operands cannot be uniform instructions. | ||||||
5140 | if (isOutOfScope(OV)) | ||||||
5141 | continue; | ||||||
5142 | // First order recurrence Phi's should typically be considered | ||||||
5143 | // non-uniform. | ||||||
5144 | auto *OP = dyn_cast<PHINode>(OV); | ||||||
5145 | if (OP && Legal->isFirstOrderRecurrence(OP)) | ||||||
5146 | continue; | ||||||
5147 | // If all the users of the operand are uniform, then add the | ||||||
5148 | // operand into the uniform worklist. | ||||||
5149 | auto *OI = cast<Instruction>(OV); | ||||||
5150 | if (llvm::all_of(OI->users(), [&](User *U) -> bool { | ||||||
5151 | auto *J = cast<Instruction>(U); | ||||||
5152 | return Worklist.count(J) || isVectorizedMemAccessUse(J, OI); | ||||||
5153 | })) | ||||||
5154 | addToWorklistIfAllowed(OI); | ||||||
5155 | } | ||||||
5156 | } | ||||||
5157 | |||||||
5158 | // For an instruction to be added into Worklist above, all its users inside | ||||||
5159 | // the loop should also be in Worklist. However, this condition cannot be | ||||||
5160 | // true for phi nodes that form a cyclic dependence. We must process phi | ||||||
5161 | // nodes separately. An induction variable will remain uniform if all users | ||||||
5162 | // of the induction variable and induction variable update remain uniform. | ||||||
5163 | // The code below handles both pointer and non-pointer induction variables. | ||||||
5164 | for (auto &Induction : Legal->getInductionVars()) { | ||||||
5165 | auto *Ind = Induction.first; | ||||||
5166 | auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); | ||||||
5167 | |||||||
5168 | // Determine if all users of the induction variable are uniform after | ||||||
5169 | // vectorization. | ||||||
5170 | auto UniformInd = llvm::all_of(Ind->users(), [&](User *U) -> bool { | ||||||
5171 | auto *I = cast<Instruction>(U); | ||||||
5172 | return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) || | ||||||
5173 | isVectorizedMemAccessUse(I, Ind); | ||||||
5174 | }); | ||||||
5175 | if (!UniformInd) | ||||||
5176 | continue; | ||||||
5177 | |||||||
5178 | // Determine if all users of the induction variable update instruction are | ||||||
5179 | // uniform after vectorization. | ||||||
5180 | auto UniformIndUpdate = | ||||||
5181 | llvm::all_of(IndUpdate->users(), [&](User *U) -> bool { | ||||||
5182 | auto *I = cast<Instruction>(U); | ||||||
5183 | return I == Ind || !TheLoop->contains(I) || Worklist.count(I) || | ||||||
5184 | isVectorizedMemAccessUse(I, IndUpdate); | ||||||
5185 | }); | ||||||
5186 | if (!UniformIndUpdate) | ||||||
5187 | continue; | ||||||
5188 | |||||||
5189 | // The induction variable and its update instruction will remain uniform. | ||||||
5190 | addToWorklistIfAllowed(Ind); | ||||||
5191 | addToWorklistIfAllowed(IndUpdate); | ||||||
5192 | } | ||||||
5193 | |||||||
5194 | Uniforms[VF].insert(Worklist.begin(), Worklist.end()); | ||||||
5195 | } | ||||||
5196 | |||||||
5197 | bool LoopVectorizationCostModel::runtimeChecksRequired() { | ||||||
5198 | 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); | ||||||
5199 | |||||||
5200 | if (Legal->getRuntimePointerChecking()->Need) { | ||||||
5201 | reportVectorizationFailure("Runtime ptr check is required with -Os/-Oz", | ||||||
5202 | "runtime pointer checks needed. Enable vectorization of this " | ||||||
5203 | "loop with '#pragma clang loop vectorize(enable)' when " | ||||||
5204 | "compiling with -Os/-Oz", | ||||||
5205 | "CantVersionLoopWithOptForSize", ORE, TheLoop); | ||||||
5206 | return true; | ||||||
5207 | } | ||||||
5208 | |||||||
5209 | if (!PSE.getUnionPredicate().getPredicates().empty()) { | ||||||
5210 | reportVectorizationFailure("Runtime SCEV check is required with -Os/-Oz", | ||||||
5211 | "runtime SCEV checks needed. Enable vectorization of this " | ||||||
5212 | "loop with '#pragma clang loop vectorize(enable)' when " | ||||||
5213 | "compiling with -Os/-Oz", | ||||||
5214 | "CantVersionLoopWithOptForSize", ORE, TheLoop); | ||||||
5215 | return true; | ||||||
5216 | } | ||||||
5217 | |||||||
5218 | // FIXME: Avoid specializing for stride==1 instead of bailing out. | ||||||
5219 | if (!Legal->getLAI()->getSymbolicStrides().empty()) { | ||||||
5220 | reportVectorizationFailure("Runtime stride check for small trip count", | ||||||
5221 | "runtime stride == 1 checks needed. Enable vectorization of " | ||||||
5222 | "this loop without such check by compiling with -Os/-Oz", | ||||||
5223 | "CantVersionLoopWithOptForSize", ORE, TheLoop); | ||||||
5224 | return true; | ||||||
5225 | } | ||||||
5226 | |||||||
5227 | return false; | ||||||
5228 | } | ||||||
5229 | |||||||
5230 | ElementCount | ||||||
5231 | LoopVectorizationCostModel::getMaxLegalScalableVF(unsigned MaxSafeElements) { | ||||||
5232 | if (!TTI.supportsScalableVectors() && !ForceTargetSupportsScalableVectors) | ||||||
5233 | return ElementCount::getScalable(0); | ||||||
5234 | |||||||
5235 | if (Hints->isScalableVectorizationDisabled()) { | ||||||
5236 | reportVectorizationInfo("Scalable vectorization is explicitly disabled", | ||||||
5237 | "ScalableVectorizationDisabled", ORE, TheLoop); | ||||||
5238 | return ElementCount::getScalable(0); | ||||||
5239 | } | ||||||
5240 | |||||||
5241 | 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); | ||||||
5242 | |||||||
5243 | auto MaxScalableVF = ElementCount::getScalable( | ||||||
5244 | std::numeric_limits<ElementCount::ScalarTy>::max()); | ||||||
5245 | |||||||
5246 | // Test that the loop-vectorizer can legalize all operations for this MaxVF. | ||||||
5247 | // FIXME: While for scalable vectors this is currently sufficient, this should | ||||||
5248 | // be replaced by a more detailed mechanism that filters out specific VFs, | ||||||
5249 | // instead of invalidating vectorization for a whole set of VFs based on the | ||||||
5250 | // MaxVF. | ||||||
5251 | |||||||
5252 | // Disable scalable vectorization if the loop contains unsupported reductions. | ||||||
5253 | if (!canVectorizeReductions(MaxScalableVF)) { | ||||||
5254 | reportVectorizationInfo( | ||||||
5255 | "Scalable vectorization not supported for the reduction " | ||||||
5256 | "operations found in this loop.", | ||||||
5257 | "ScalableVFUnfeasible", ORE, TheLoop); | ||||||
5258 | return ElementCount::getScalable(0); | ||||||
5259 | } | ||||||
5260 | |||||||
5261 | // Disable scalable vectorization if the loop contains any instructions | ||||||
5262 | // with element types not supported for scalable vectors. | ||||||
5263 | if (any_of(ElementTypesInLoop, [&](Type *Ty) { | ||||||
5264 | return !Ty->isVoidTy() && | ||||||
5265 | !this->TTI.isElementTypeLegalForScalableVector(Ty); | ||||||
5266 | })) { | ||||||
5267 | reportVectorizationInfo("Scalable vectorization is not supported " | ||||||
5268 | "for all element types found in this loop.", | ||||||
5269 | "ScalableVFUnfeasible", ORE, TheLoop); | ||||||
5270 | return ElementCount::getScalable(0); | ||||||
5271 | } | ||||||
5272 | |||||||
5273 | if (Legal->isSafeForAnyVectorWidth()) | ||||||
5274 | return MaxScalableVF; | ||||||
5275 | |||||||
5276 | // Limit MaxScalableVF by the maximum safe dependence distance. | ||||||
5277 | Optional<unsigned> MaxVScale = TTI.getMaxVScale(); | ||||||
5278 | if (!MaxVScale && TheFunction->hasFnAttribute(Attribute::VScaleRange)) | ||||||
5279 | MaxVScale = | ||||||
5280 | TheFunction->getFnAttribute(Attribute::VScaleRange).getVScaleRangeMax(); | ||||||
5281 | MaxScalableVF = ElementCount::getScalable( | ||||||
5282 | MaxVScale ? (MaxSafeElements / MaxVScale.getValue()) : 0); | ||||||
5283 | if (!MaxScalableVF) | ||||||
5284 | reportVectorizationInfo( | ||||||
5285 | "Max legal vector width too small, scalable vectorization " | ||||||
5286 | "unfeasible.", | ||||||
5287 | "ScalableVFUnfeasible", ORE, TheLoop); | ||||||
5288 | |||||||
5289 | return MaxScalableVF; | ||||||
5290 | } | ||||||
5291 | |||||||
5292 | FixedScalableVFPair LoopVectorizationCostModel::computeFeasibleMaxVF( | ||||||
5293 | unsigned ConstTripCount, ElementCount UserVF, bool FoldTailByMasking) { | ||||||
5294 | MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI); | ||||||
5295 | unsigned SmallestType, WidestType; | ||||||
5296 | std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes(); | ||||||
5297 | |||||||
5298 | // Get the maximum safe dependence distance in bits computed by LAA. | ||||||
5299 | // It is computed by MaxVF * sizeOf(type) * 8, where type is taken from | ||||||
5300 | // the memory accesses that is most restrictive (involved in the smallest | ||||||
5301 | // dependence distance). | ||||||
5302 | unsigned MaxSafeElements = | ||||||
5303 | PowerOf2Floor(Legal->getMaxSafeVectorWidthInBits() / WidestType); | ||||||
5304 | |||||||
5305 | auto MaxSafeFixedVF = ElementCount::getFixed(MaxSafeElements); | ||||||
5306 | auto MaxSafeScalableVF = getMaxLegalScalableVF(MaxSafeElements); | ||||||
5307 | |||||||
5308 | 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) | ||||||
5309 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The max safe fixed VF is: " << MaxSafeFixedVF << ".\n"; } } while (false); | ||||||
5310 | 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) | ||||||
5311 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The max safe scalable VF is: " << MaxSafeScalableVF << ".\n"; } } while (false); | ||||||
5312 | |||||||
5313 | // First analyze the UserVF, fall back if the UserVF should be ignored. | ||||||
5314 | if (UserVF) { | ||||||
5315 | auto MaxSafeUserVF = | ||||||
5316 | UserVF.isScalable() ? MaxSafeScalableVF : MaxSafeFixedVF; | ||||||
5317 | |||||||
5318 | if (ElementCount::isKnownLE(UserVF, MaxSafeUserVF)) { | ||||||
5319 | // If `VF=vscale x N` is safe, then so is `VF=N` | ||||||
5320 | if (UserVF.isScalable()) | ||||||
5321 | return FixedScalableVFPair( | ||||||
5322 | ElementCount::getFixed(UserVF.getKnownMinValue()), UserVF); | ||||||
5323 | else | ||||||
5324 | return UserVF; | ||||||
5325 | } | ||||||
5326 | |||||||
5327 | 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", 5327, __extension__ __PRETTY_FUNCTION__)); | ||||||
5328 | |||||||
5329 | // Only clamp if the UserVF is not scalable. If the UserVF is scalable, it | ||||||
5330 | // is better to ignore the hint and let the compiler choose a suitable VF. | ||||||
5331 | if (!UserVF.isScalable()) { | ||||||
5332 | 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) | ||||||
5333 | << " 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) | ||||||
5334 | << 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); | ||||||
5335 | ORE->emit([&]() { | ||||||
5336 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationFactor", | ||||||
5337 | TheLoop->getStartLoc(), | ||||||
5338 | TheLoop->getHeader()) | ||||||
5339 | << "User-specified vectorization factor " | ||||||
5340 | << ore::NV("UserVectorizationFactor", UserVF) | ||||||
5341 | << " is unsafe, clamping to maximum safe vectorization factor " | ||||||
5342 | << ore::NV("VectorizationFactor", MaxSafeFixedVF); | ||||||
5343 | }); | ||||||
5344 | return MaxSafeFixedVF; | ||||||
5345 | } | ||||||
5346 | |||||||
5347 | if (!TTI.supportsScalableVectors() && !ForceTargetSupportsScalableVectors) { | ||||||
5348 | 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) | ||||||
5349 | << " 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) | ||||||
5350 | "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); | ||||||
5351 | ORE->emit([&]() { | ||||||
5352 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationFactor", | ||||||
5353 | TheLoop->getStartLoc(), | ||||||
5354 | TheLoop->getHeader()) | ||||||
5355 | << "User-specified vectorization factor " | ||||||
5356 | << ore::NV("UserVectorizationFactor", UserVF) | ||||||
5357 | << " is ignored because the target does not support scalable " | ||||||
5358 | "vectors. The compiler will pick a more suitable value."; | ||||||
5359 | }); | ||||||
5360 | } else { | ||||||
5361 | 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) | ||||||
5362 | << " 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); | ||||||
5363 | ORE->emit([&]() { | ||||||
5364 | return OptimizationRemarkAnalysis(DEBUG_TYPE"loop-vectorize", "VectorizationFactor", | ||||||
5365 | TheLoop->getStartLoc(), | ||||||
5366 | TheLoop->getHeader()) | ||||||
5367 | << "User-specified vectorization factor " | ||||||
5368 | << ore::NV("UserVectorizationFactor", UserVF) | ||||||
5369 | << " is unsafe. Ignoring the hint to let the compiler pick a " | ||||||
5370 | "more suitable value."; | ||||||
5371 | }); | ||||||
5372 | } | ||||||
5373 | } | ||||||
5374 | |||||||
5375 | 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) | ||||||
5376 | << " / " << WidestType << " bits.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: The Smallest and Widest types: " << SmallestType << " / " << WidestType << " bits.\n"; } } while (false); | ||||||
5377 | |||||||
5378 | FixedScalableVFPair Result(ElementCount::getFixed(1), | ||||||
5379 | ElementCount::getScalable(0)); | ||||||
5380 | if (auto MaxVF = | ||||||
5381 | getMaximizedVFForTarget(ConstTripCount, SmallestType, WidestType, | ||||||
5382 | MaxSafeFixedVF, FoldTailByMasking)) | ||||||
5383 | Result.FixedVF = MaxVF; | ||||||
5384 | |||||||
5385 | if (auto MaxVF = | ||||||
5386 | getMaximizedVFForTarget(ConstTripCount, SmallestType, WidestType, | ||||||
5387 | MaxSafeScalableVF, FoldTailByMasking)) | ||||||
5388 | if (MaxVF.isScalable()) { | ||||||
5389 | Result.ScalableVF = MaxVF; | ||||||
5390 | 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) | ||||||
5391 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found feasible scalable VF = " << MaxVF << "\n"; } } while (false); | ||||||
5392 | } | ||||||
5393 | |||||||
5394 | return Result; | ||||||
5395 | } | ||||||
5396 | |||||||
5397 | FixedScalableVFPair | ||||||
5398 | LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) { | ||||||
5399 | if (Legal->getRuntimePointerChecking()->Need && TTI.hasBranchDivergence()) { | ||||||
5400 | // TODO: It may by useful to do since it's still likely to be dynamically | ||||||
5401 | // uniform if the target can skip. | ||||||
5402 | reportVectorizationFailure( | ||||||
5403 | "Not inserting runtime ptr check for divergent target", | ||||||
5404 | "runtime pointer checks needed. Not enabled for divergent target", | ||||||
5405 | "CantVersionLoopWithDivergentTarget", ORE, TheLoop); | ||||||
5406 | return FixedScalableVFPair::getNone(); | ||||||
5407 | } | ||||||
5408 | |||||||
5409 | unsigned TC = PSE.getSE()->getSmallConstantTripCount(TheLoop); | ||||||
5410 | 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); | ||||||
5411 | if (TC == 1) { | ||||||
5412 | reportVectorizationFailure("Single iteration (non) loop", | ||||||
5413 | "loop trip count is one, irrelevant for vectorization", | ||||||
5414 | "SingleIterationLoop", ORE, TheLoop); | ||||||
5415 | return FixedScalableVFPair::getNone(); | ||||||
5416 | } | ||||||
5417 | |||||||
5418 | switch (ScalarEpilogueStatus) { | ||||||
5419 | case CM_ScalarEpilogueAllowed: | ||||||
5420 | return computeFeasibleMaxVF(TC, UserVF, false); | ||||||
5421 | case CM_ScalarEpilogueNotAllowedUsePredicate: | ||||||
5422 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | ||||||
5423 | case CM_ScalarEpilogueNotNeededUsePredicate: | ||||||
5424 | 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) | ||||||
5425 | 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) | ||||||
5426 | << "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) | ||||||
5427 | << "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); | ||||||
5428 | break; | ||||||
5429 | case CM_ScalarEpilogueNotAllowedLowTripLoop: | ||||||
5430 | // fallthrough as a special case of OptForSize | ||||||
5431 | case CM_ScalarEpilogueNotAllowedOptSize: | ||||||
5432 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedOptSize) | ||||||
5433 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not allowing scalar epilogue due to -Os/-Oz.\n" ; } } while (false) | ||||||
5434 | 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); | ||||||
5435 | else | ||||||
5436 | 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) | ||||||
5437 | << "count.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not allowing scalar epilogue due to low trip " << "count.\n"; } } while (false); | ||||||
5438 | |||||||
5439 | // Bail if runtime checks are required, which are not good when optimising | ||||||
5440 | // for size. | ||||||
5441 | if (runtimeChecksRequired()) | ||||||
5442 | return FixedScalableVFPair::getNone(); | ||||||
5443 | |||||||
5444 | break; | ||||||
5445 | } | ||||||
5446 | |||||||
5447 | // The only loops we can vectorize without a scalar epilogue, are loops with | ||||||
5448 | // a bottom-test and a single exiting block. We'd have to handle the fact | ||||||
5449 | // that not every instruction executes on the last iteration. This will | ||||||
5450 | // require a lane mask which varies through the vector loop body. (TODO) | ||||||
5451 | if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) { | ||||||
5452 | // If there was a tail-folding hint/switch, but we can't fold the tail by | ||||||
5453 | // masking, fallback to a vectorization with a scalar epilogue. | ||||||
5454 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotNeededUsePredicate) { | ||||||
5455 | 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) | ||||||
5456 | "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); | ||||||
5457 | ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; | ||||||
5458 | return computeFeasibleMaxVF(TC, UserVF, false); | ||||||
5459 | } | ||||||
5460 | return FixedScalableVFPair::getNone(); | ||||||
5461 | } | ||||||
5462 | |||||||
5463 | // Now try the tail folding | ||||||
5464 | |||||||
5465 | // Invalidate interleave groups that require an epilogue if we can't mask | ||||||
5466 | // the interleave-group. | ||||||
5467 | if (!useMaskedInterleavedAccesses(TTI)) { | ||||||
5468 | 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", 5469, __extension__ __PRETTY_FUNCTION__)) | ||||||
5469 | "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", 5469, __extension__ __PRETTY_FUNCTION__)); | ||||||
5470 | // Note: There is no need to invalidate any cost modeling decisions here, as | ||||||
5471 | // non where taken so far. | ||||||
5472 | InterleaveInfo.invalidateGroupsRequiringScalarEpilogue(); | ||||||
5473 | } | ||||||
5474 | |||||||
5475 | FixedScalableVFPair MaxFactors = computeFeasibleMaxVF(TC, UserVF, true); | ||||||
5476 | // Avoid tail folding if the trip count is known to be a multiple of any VF | ||||||
5477 | // we chose. | ||||||
5478 | // FIXME: The condition below pessimises the case for fixed-width vectors, | ||||||
5479 | // when scalable VFs are also candidates for vectorization. | ||||||
5480 | if (MaxFactors.FixedVF.isVector() && !MaxFactors.ScalableVF) { | ||||||
5481 | ElementCount MaxFixedVF = MaxFactors.FixedVF; | ||||||
5482 | 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", 5483, __extension__ __PRETTY_FUNCTION__)) | ||||||
5483 | "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", 5483, __extension__ __PRETTY_FUNCTION__)); | ||||||
5484 | unsigned MaxVFtimesIC = UserIC ? MaxFixedVF.getFixedValue() * UserIC | ||||||
5485 | : MaxFixedVF.getFixedValue(); | ||||||
5486 | ScalarEvolution *SE = PSE.getSE(); | ||||||
5487 | const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount(); | ||||||
5488 | const SCEV *ExitCount = SE->getAddExpr( | ||||||
5489 | BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType())); | ||||||
5490 | const SCEV *Rem = SE->getURemExpr( | ||||||
5491 | SE->applyLoopGuards(ExitCount, TheLoop), | ||||||
5492 | SE->getConstant(BackedgeTakenCount->getType(), MaxVFtimesIC)); | ||||||
5493 | if (Rem->isZero()) { | ||||||
5494 | // Accept MaxFixedVF if we do not have a tail. | ||||||
5495 | 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); | ||||||
5496 | return MaxFactors; | ||||||
5497 | } | ||||||
5498 | } | ||||||
5499 | |||||||
5500 | // For scalable vectors don't use tail folding for low trip counts or | ||||||
5501 | // optimizing for code size. We only permit this if the user has explicitly | ||||||
5502 | // requested it. | ||||||
5503 | if (ScalarEpilogueStatus != CM_ScalarEpilogueNotNeededUsePredicate && | ||||||
5504 | ScalarEpilogueStatus != CM_ScalarEpilogueNotAllowedUsePredicate && | ||||||
5505 | MaxFactors.ScalableVF.isVector()) | ||||||
5506 | MaxFactors.ScalableVF = ElementCount::getScalable(0); | ||||||
5507 | |||||||
5508 | // If we don't know the precise trip count, or if the trip count that we | ||||||
5509 | // found modulo the vectorization factor is not zero, try to fold the tail | ||||||
5510 | // by masking. | ||||||
5511 | // FIXME: look for a smaller MaxVF that does divide TC rather than masking. | ||||||
5512 | if (Legal->prepareToFoldTailByMasking()) { | ||||||
5513 | FoldTailByMasking = true; | ||||||
5514 | return MaxFactors; | ||||||
5515 | } | ||||||
5516 | |||||||
5517 | // If there was a tail-folding hint/switch, but we can't fold the tail by | ||||||
5518 | // masking, fallback to a vectorization with a scalar epilogue. | ||||||
5519 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotNeededUsePredicate) { | ||||||
5520 | 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) | ||||||
5521 | "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); | ||||||
5522 | ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; | ||||||
5523 | return MaxFactors; | ||||||
5524 | } | ||||||
5525 | |||||||
5526 | if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedUsePredicate) { | ||||||
5527 | 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); | ||||||
5528 | return FixedScalableVFPair::getNone(); | ||||||
5529 | } | ||||||
5530 | |||||||
5531 | if (TC == 0) { | ||||||
5532 | reportVectorizationFailure( | ||||||
5533 | "Unable to calculate the loop count due to complex control flow", | ||||||
5534 | "unable to calculate the loop count due to complex control flow", | ||||||
5535 | "UnknownLoopCountComplexCFG", ORE, TheLoop); | ||||||
5536 | return FixedScalableVFPair::getNone(); | ||||||
5537 | } | ||||||
5538 | |||||||
5539 | reportVectorizationFailure( | ||||||
5540 | "Cannot optimize for size and vectorize at the same time.", | ||||||
5541 | "cannot optimize for size and vectorize at the same time. " | ||||||
5542 | "Enable vectorization of this loop with '#pragma clang loop " | ||||||
5543 | "vectorize(enable)' when compiling with -Os/-Oz", | ||||||
5544 | "NoTailLoopWithOptForSize", ORE, TheLoop); | ||||||
5545 | return FixedScalableVFPair::getNone(); | ||||||
5546 | } | ||||||
5547 | |||||||
5548 | ElementCount LoopVectorizationCostModel::getMaximizedVFForTarget( | ||||||
5549 | unsigned ConstTripCount, unsigned SmallestType, unsigned WidestType, | ||||||
5550 | const ElementCount &MaxSafeVF, bool FoldTailByMasking) { | ||||||
5551 | bool ComputeScalableMaxVF = MaxSafeVF.isScalable(); | ||||||
5552 | TypeSize WidestRegister = TTI.getRegisterBitWidth( | ||||||
5553 | ComputeScalableMaxVF ? TargetTransformInfo::RGK_ScalableVector | ||||||
5554 | : TargetTransformInfo::RGK_FixedWidthVector); | ||||||
5555 | |||||||
5556 | // Convenience function to return the minimum of two ElementCounts. | ||||||
5557 | auto MinVF = [](const ElementCount &LHS, const ElementCount &RHS) { | ||||||
5558 | 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", 5559, __extension__ __PRETTY_FUNCTION__)) | ||||||
5559 | "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", 5559, __extension__ __PRETTY_FUNCTION__)); | ||||||
5560 | return ElementCount::isKnownLT(LHS, RHS) ? LHS : RHS; | ||||||
5561 | }; | ||||||
5562 | |||||||
5563 | // Ensure MaxVF is a power of 2; the dependence distance bound may not be. | ||||||
5564 | // Note that both WidestRegister and WidestType may not be a powers of 2. | ||||||
5565 | auto MaxVectorElementCount = ElementCount::get( | ||||||
5566 | PowerOf2Floor(WidestRegister.getKnownMinSize() / WidestType), | ||||||
5567 | ComputeScalableMaxVF); | ||||||
5568 | MaxVectorElementCount = MinVF(MaxVectorElementCount, MaxSafeVF); | ||||||
5569 | 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) | ||||||
5570 | << (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); | ||||||
5571 | |||||||
5572 | if (!MaxVectorElementCount) { | ||||||
5573 | 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) | ||||||
5574 | << (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) | ||||||
5575 | << " 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); | ||||||
5576 | return ElementCount::getFixed(1); | ||||||
5577 | } | ||||||
5578 | |||||||
5579 | const auto TripCountEC = ElementCount::getFixed(ConstTripCount); | ||||||
5580 | if (ConstTripCount && | ||||||
5581 | ElementCount::isKnownLE(TripCountEC, MaxVectorElementCount) && | ||||||
5582 | (!FoldTailByMasking || isPowerOf2_32(ConstTripCount))) { | ||||||
5583 | // If loop trip count (TC) is known at compile time there is no point in | ||||||
5584 | // choosing VF greater than TC (as done in the loop below). Select maximum | ||||||
5585 | // power of two which doesn't exceed TC. | ||||||
5586 | // If MaxVectorElementCount is scalable, we only fall back on a fixed VF | ||||||
5587 | // when the TC is less than or equal to the known number of lanes. | ||||||
5588 | auto ClampedConstTripCount = PowerOf2Floor(ConstTripCount); | ||||||
5589 | 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) | ||||||
5590 | "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) | ||||||
5591 | << 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); | ||||||
5592 | return ElementCount::getFixed(ClampedConstTripCount); | ||||||
5593 | } | ||||||
5594 | |||||||
5595 | ElementCount MaxVF = MaxVectorElementCount; | ||||||
5596 | if (TTI.shouldMaximizeVectorBandwidth() || | ||||||
5597 | (MaximizeBandwidth && isScalarEpilogueAllowed())) { | ||||||
5598 | auto MaxVectorElementCountMaxBW = ElementCount::get( | ||||||
5599 | PowerOf2Floor(WidestRegister.getKnownMinSize() / SmallestType), | ||||||
5600 | ComputeScalableMaxVF); | ||||||
5601 | MaxVectorElementCountMaxBW = MinVF(MaxVectorElementCountMaxBW, MaxSafeVF); | ||||||
5602 | |||||||
5603 | // Collect all viable vectorization factors larger than the default MaxVF | ||||||
5604 | // (i.e. MaxVectorElementCount). | ||||||
5605 | SmallVector<ElementCount, 8> VFs; | ||||||
5606 | for (ElementCount VS = MaxVectorElementCount * 2; | ||||||
5607 | ElementCount::isKnownLE(VS, MaxVectorElementCountMaxBW); VS *= 2) | ||||||
5608 | VFs.push_back(VS); | ||||||
5609 | |||||||
5610 | // For each VF calculate its register usage. | ||||||
5611 | auto RUs = calculateRegisterUsage(VFs); | ||||||
5612 | |||||||
5613 | // Select the largest VF which doesn't require more registers than existing | ||||||
5614 | // ones. | ||||||
5615 | for (int i = RUs.size() - 1; i >= 0; --i) { | ||||||
5616 | bool Selected = true; | ||||||
5617 | for (auto &pair : RUs[i].MaxLocalUsers) { | ||||||
5618 | unsigned TargetNumRegisters = TTI.getNumberOfRegisters(pair.first); | ||||||
5619 | if (pair.second > TargetNumRegisters) | ||||||
5620 | Selected = false; | ||||||
5621 | } | ||||||
5622 | if (Selected) { | ||||||
5623 | MaxVF = VFs[i]; | ||||||
5624 | break; | ||||||
5625 | } | ||||||
5626 | } | ||||||
5627 | if (ElementCount MinVF = | ||||||
5628 | TTI.getMinimumVF(SmallestType, ComputeScalableMaxVF)) { | ||||||
5629 | if (ElementCount::isKnownLT(MaxVF, MinVF)) { | ||||||
5630 | 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) | ||||||
5631 | << ") 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); | ||||||
5632 | MaxVF = MinVF; | ||||||
5633 | } | ||||||
5634 | } | ||||||
5635 | } | ||||||
5636 | return MaxVF; | ||||||
5637 | } | ||||||
5638 | |||||||
5639 | bool LoopVectorizationCostModel::isMoreProfitable( | ||||||
5640 | const VectorizationFactor &A, const VectorizationFactor &B) const { | ||||||
5641 | InstructionCost CostA = A.Cost; | ||||||
5642 | InstructionCost CostB = B.Cost; | ||||||
5643 | |||||||
5644 | unsigned MaxTripCount = PSE.getSE()->getSmallConstantMaxTripCount(TheLoop); | ||||||
5645 | |||||||
5646 | if (!A.Width.isScalable() && !B.Width.isScalable() && FoldTailByMasking && | ||||||
5647 | MaxTripCount) { | ||||||
5648 | // If we are folding the tail and the trip count is a known (possibly small) | ||||||
5649 | // constant, the trip count will be rounded up to an integer number of | ||||||
5650 | // iterations. The total cost will be PerIterationCost*ceil(TripCount/VF), | ||||||
5651 | // which we compare directly. When not folding the tail, the total cost will | ||||||
5652 | // be PerIterationCost*floor(TC/VF) + Scalar remainder cost, and so is | ||||||
5653 | // approximated with the per-lane cost below instead of using the tripcount | ||||||
5654 | // as here. | ||||||
5655 | auto RTCostA = CostA * divideCeil(MaxTripCount, A.Width.getFixedValue()); | ||||||
5656 | auto RTCostB = CostB * divideCeil(MaxTripCount, B.Width.getFixedValue()); | ||||||
5657 | return RTCostA < RTCostB; | ||||||
5658 | } | ||||||
5659 | |||||||
5660 | // Improve estimate for the vector width if it is scalable. | ||||||
5661 | unsigned EstimatedWidthA = A.Width.getKnownMinValue(); | ||||||
5662 | unsigned EstimatedWidthB = B.Width.getKnownMinValue(); | ||||||
5663 | if (Optional<unsigned> VScale = TTI.getVScaleForTuning()) { | ||||||
5664 | if (A.Width.isScalable()) | ||||||
5665 | EstimatedWidthA *= VScale.getValue(); | ||||||
5666 | if (B.Width.isScalable()) | ||||||
5667 | EstimatedWidthB *= VScale.getValue(); | ||||||
5668 | } | ||||||
5669 | |||||||
5670 | // Assume vscale may be larger than 1 (or the value being tuned for), | ||||||
5671 | // so that scalable vectorization is slightly favorable over fixed-width | ||||||
5672 | // vectorization. | ||||||
5673 | if (A.Width.isScalable() && !B.Width.isScalable()) | ||||||
5674 | return (CostA * B.Width.getFixedValue()) <= (CostB * EstimatedWidthA); | ||||||
5675 | |||||||
5676 | // To avoid the need for FP division: | ||||||
5677 | // (CostA / A.Width) < (CostB / B.Width) | ||||||
5678 | // <=> (CostA * B.Width) < (CostB * A.Width) | ||||||
5679 | return (CostA * EstimatedWidthB) < (CostB * EstimatedWidthA); | ||||||
5680 | } | ||||||
5681 | |||||||
5682 | VectorizationFactor LoopVectorizationCostModel::selectVectorizationFactor( | ||||||
5683 | const ElementCountSet &VFCandidates) { | ||||||
5684 | InstructionCost ExpectedCost = expectedCost(ElementCount::getFixed(1)).first; | ||||||
5685 | 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); | ||||||
5686 | 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", 5686, __extension__ __PRETTY_FUNCTION__)); | ||||||
5687 | 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", 5688, __extension__ __PRETTY_FUNCTION__)) | ||||||
5688 | "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", 5688, __extension__ __PRETTY_FUNCTION__)); | ||||||
5689 | |||||||
5690 | const VectorizationFactor ScalarCost(ElementCount::getFixed(1), ExpectedCost); | ||||||
5691 | VectorizationFactor ChosenFactor = ScalarCost; | ||||||
5692 | |||||||
5693 | bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled; | ||||||
5694 | if (ForceVectorization && VFCandidates.size() > 1) { | ||||||
5695 | // Ignore scalar width, because the user explicitly wants vectorization. | ||||||
5696 | // Initialize cost to max so that VF = 2 is, at least, chosen during cost | ||||||
5697 | // evaluation. | ||||||
5698 | ChosenFactor.Cost = InstructionCost::getMax(); | ||||||
5699 | } | ||||||
5700 | |||||||
5701 | SmallVector<InstructionVFPair> InvalidCosts; | ||||||
5702 | for (const auto &i : VFCandidates) { | ||||||
5703 | // The cost for scalar VF=1 is already calculated, so ignore it. | ||||||
5704 | if (i.isScalar()) | ||||||
5705 | continue; | ||||||
5706 | |||||||
5707 | VectorizationCostTy C = expectedCost(i, &InvalidCosts); | ||||||
5708 | VectorizationFactor Candidate(i, C.first); | ||||||
5709 | |||||||
5710 | #ifndef NDEBUG | ||||||
5711 | unsigned AssumedMinimumVscale = 1; | ||||||
5712 | if (Optional<unsigned> VScale = TTI.getVScaleForTuning()) | ||||||
5713 | AssumedMinimumVscale = VScale.getValue(); | ||||||
5714 | unsigned Width = | ||||||
5715 | Candidate.Width.isScalable() | ||||||
5716 | ? Candidate.Width.getKnownMinValue() * AssumedMinimumVscale | ||||||
5717 | : Candidate.Width.getFixedValue(); | ||||||
5718 | 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) | ||||||
5719 | << " 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); | ||||||
5720 | if (i.isScalable()) | ||||||
5721 | 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 ) | ||||||
5722 | << AssumedMinimumVscale << ")")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << " (assuming a minimum vscale of " << AssumedMinimumVscale << ")"; } } while (false ); | ||||||
5723 | LLVM_DEBUG(dbgs() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << ".\n"; } } while (false ); | ||||||
5724 | #endif | ||||||
5725 | |||||||
5726 | if (!C.second && !ForceVectorization) { | ||||||
5727 | 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) | ||||||
5728 | 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) | ||||||
5729 | << " 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); | ||||||
5730 | continue; | ||||||
5731 | } | ||||||
5732 | |||||||
5733 | // If profitable add it to ProfitableVF list. | ||||||
5734 | if (isMoreProfitable(Candidate, ScalarCost)) | ||||||
5735 | ProfitableVFs.push_back(Candidate); | ||||||
5736 | |||||||
5737 | if (isMoreProfitable(Candidate, ChosenFactor)) | ||||||
5738 | ChosenFactor = Candidate; | ||||||
5739 | } | ||||||
5740 | |||||||
5741 | // Emit a report of VFs with invalid costs in the loop. | ||||||
5742 | if (!InvalidCosts.empty()) { | ||||||
5743 | // Group the remarks per instruction, keeping the instruction order from | ||||||
5744 | // InvalidCosts. | ||||||
5745 | std::map<Instruction *, unsigned> Numbering; | ||||||
5746 | unsigned I = 0; | ||||||
5747 | for (auto &Pair : InvalidCosts) | ||||||
5748 | if (!Numbering.count(Pair.first)) | ||||||
5749 | Numbering[Pair.first] = I++; | ||||||
5750 | |||||||
5751 | // Sort the list, first on instruction(number) then on VF. | ||||||
5752 | llvm::sort(InvalidCosts, | ||||||
5753 | [&Numbering](InstructionVFPair &A, InstructionVFPair &B) { | ||||||
5754 | if (Numbering[A.first] != Numbering[B.first]) | ||||||
5755 | return Numbering[A.first] < Numbering[B.first]; | ||||||
5756 | ElementCountComparator ECC; | ||||||
5757 | return ECC(A.second, B.second); | ||||||
5758 | }); | ||||||
5759 | |||||||
5760 | // For a list of ordered instruction-vf pairs: | ||||||
5761 | // [(load, vf1), (load, vf2), (store, vf1)] | ||||||
5762 | // Group the instructions together to emit separate remarks for: | ||||||
5763 | // load (vf1, vf2) | ||||||
5764 | // store (vf1) | ||||||
5765 | auto Tail = ArrayRef<InstructionVFPair>(InvalidCosts); | ||||||
5766 | auto Subset = ArrayRef<InstructionVFPair>(); | ||||||
5767 | do { | ||||||
5768 | if (Subset.empty()) | ||||||
5769 | Subset = Tail.take_front(1); | ||||||
5770 | |||||||
5771 | Instruction *I = Subset.front().first; | ||||||
5772 | |||||||
5773 | // If the next instruction is different, or if there are no other pairs, | ||||||
5774 | // emit a remark for the collated subset. e.g. | ||||||
5775 | // [(load, vf1), (load, vf2))] | ||||||
5776 | // to emit: | ||||||
5777 | // remark: invalid costs for 'load' at VF=(vf, vf2) | ||||||
5778 | if (Subset == Tail || Tail[Subset.size()].first != I) { | ||||||
5779 | std::string OutString; | ||||||
5780 | raw_string_ostream OS(OutString); | ||||||
5781 | 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", 5781, __extension__ __PRETTY_FUNCTION__)); | ||||||
5782 | OS << "Instruction with invalid costs prevented vectorization at VF=("; | ||||||
5783 | for (auto &Pair : Subset) | ||||||
5784 | OS << (Pair.second == Subset.front().second ? "" : ", ") | ||||||
5785 | << Pair.second; | ||||||
5786 | OS << "):"; | ||||||
5787 | if (auto *CI = dyn_cast<CallInst>(I)) | ||||||
5788 | OS << " call to " << CI->getCalledFunction()->getName(); | ||||||
5789 | else | ||||||
5790 | OS << " " << I->getOpcodeName(); | ||||||
5791 | OS.flush(); | ||||||
5792 | reportVectorizationInfo(OutString, "InvalidCost", ORE, TheLoop, I); | ||||||
5793 | Tail = Tail.drop_front(Subset.size()); | ||||||
5794 | Subset = {}; | ||||||
5795 | } else | ||||||
5796 | // Grow the subset by one element | ||||||
5797 | Subset = Tail.take_front(Subset.size() + 1); | ||||||
5798 | } while (!Tail.empty()); | ||||||
5799 | } | ||||||
5800 | |||||||
5801 | if (!EnableCondStoresVectorization && NumPredStores) { | ||||||
5802 | reportVectorizationFailure("There are conditional stores.", | ||||||
5803 | "store that is conditionally executed prevents vectorization", | ||||||
5804 | "ConditionalStore", ORE, TheLoop); | ||||||
5805 | ChosenFactor = ScalarCost; | ||||||
5806 | } | ||||||
5807 | |||||||
5808 | LLVM_DEBUG(if (ForceVectorization && !ChosenFactor.Width.isScalar() &&do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && !ChosenFactor .Width.isScalar() && ChosenFactor.Cost >= ScalarCost .Cost) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false) | ||||||
5809 | ChosenFactor.Cost >= ScalarCost.Cost) dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && !ChosenFactor .Width.isScalar() && ChosenFactor.Cost >= ScalarCost .Cost) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false) | ||||||
5810 | << "LV: Vectorization seems to be not beneficial, "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && !ChosenFactor .Width.isScalar() && ChosenFactor.Cost >= ScalarCost .Cost) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false) | ||||||
5811 | << "but was forced by a user.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { if (ForceVectorization && !ChosenFactor .Width.isScalar() && ChosenFactor.Cost >= ScalarCost .Cost) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"; } } while (false); | ||||||
5812 | 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); | ||||||
5813 | return ChosenFactor; | ||||||
5814 | } | ||||||
5815 | |||||||
5816 | bool LoopVectorizationCostModel::isCandidateForEpilogueVectorization( | ||||||
5817 | const Loop &L, ElementCount VF) const { | ||||||
5818 | // Cross iteration phis such as reductions need special handling and are | ||||||
5819 | // currently unsupported. | ||||||
5820 | if (any_of(L.getHeader()->phis(), [&](PHINode &Phi) { | ||||||
5821 | return Legal->isFirstOrderRecurrence(&Phi) || | ||||||
5822 | Legal->isReductionVariable(&Phi); | ||||||
5823 | })) | ||||||
5824 | return false; | ||||||
5825 | |||||||
5826 | // Phis with uses outside of the loop require special handling and are | ||||||
5827 | // currently unsupported. | ||||||
5828 | for (auto &Entry : Legal->getInductionVars()) { | ||||||
5829 | // Look for uses of the value of the induction at the last iteration. | ||||||
5830 | Value *PostInc = Entry.first->getIncomingValueForBlock(L.getLoopLatch()); | ||||||
5831 | for (User *U : PostInc->users()) | ||||||
5832 | if (!L.contains(cast<Instruction>(U))) | ||||||
5833 | return false; | ||||||
5834 | // Look for uses of penultimate value of the induction. | ||||||
5835 | for (User *U : Entry.first->users()) | ||||||
5836 | if (!L.contains(cast<Instruction>(U))) | ||||||
5837 | return false; | ||||||
5838 | } | ||||||
5839 | |||||||
5840 | // Induction variables that are widened require special handling that is | ||||||
5841 | // currently not supported. | ||||||
5842 | if (any_of(Legal->getInductionVars(), [&](auto &Entry) { | ||||||
5843 | return !(this->isScalarAfterVectorization(Entry.first, VF) || | ||||||
5844 | this->isProfitableToScalarize(Entry.first, VF)); | ||||||
5845 | })) | ||||||
5846 | return false; | ||||||
5847 | |||||||
5848 | // Epilogue vectorization code has not been auditted to ensure it handles | ||||||
5849 | // non-latch exits properly. It may be fine, but it needs auditted and | ||||||
5850 | // tested. | ||||||
5851 | if (L.getExitingBlock() != L.getLoopLatch()) | ||||||
5852 | return false; | ||||||
5853 | |||||||
5854 | return true; | ||||||
5855 | } | ||||||
5856 | |||||||
5857 | bool LoopVectorizationCostModel::isEpilogueVectorizationProfitable( | ||||||
5858 | const ElementCount VF) const { | ||||||
5859 | // FIXME: We need a much better cost-model to take different parameters such | ||||||
5860 | // as register pressure, code size increase and cost of extra branches into | ||||||
5861 | // account. For now we apply a very crude heuristic and only consider loops | ||||||
5862 | // with vectorization factors larger than a certain value. | ||||||
5863 | // We also consider epilogue vectorization unprofitable for targets that don't | ||||||
5864 | // consider interleaving beneficial (eg. MVE). | ||||||
5865 | if (TTI.getMaxInterleaveFactor(VF.getKnownMinValue()) <= 1) | ||||||
5866 | return false; | ||||||
5867 | if (VF.getFixedValue() >= EpilogueVectorizationMinVF) | ||||||
5868 | return true; | ||||||
5869 | return false; | ||||||
5870 | } | ||||||
5871 | |||||||
5872 | VectorizationFactor | ||||||
5873 | LoopVectorizationCostModel::selectEpilogueVectorizationFactor( | ||||||
5874 | const ElementCount MainLoopVF, const LoopVectorizationPlanner &LVP) { | ||||||
5875 | VectorizationFactor Result = VectorizationFactor::Disabled(); | ||||||
5876 | if (!EnableEpilogueVectorization) { | ||||||
5877 | 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); | ||||||
5878 | return Result; | ||||||
5879 | } | ||||||
5880 | |||||||
5881 | if (!isScalarEpilogueAllowed()) { | ||||||
5882 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Unable to vectorize epilogue because no epilogue is " "allowed.\n";; } } while (false) | ||||||
5883 | 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) | ||||||
5884 | "allowed.\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Unable to vectorize epilogue because no epilogue is " "allowed.\n";; } } while (false); | ||||||
5885 | return Result; | ||||||
5886 | } | ||||||
5887 | |||||||
5888 | // Not really a cost consideration, but check for unsupported cases here to | ||||||
5889 | // simplify the logic. | ||||||
5890 | if (!isCandidateForEpilogueVectorization(*TheLoop, MainLoopVF)) { | ||||||
5891 | 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) | ||||||
5892 | 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) | ||||||
5893 | "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); | ||||||
5894 | return Result; | ||||||
5895 | } | ||||||
5896 | |||||||
5897 | if (EpilogueVectorizationForceVF > 1) { | ||||||
5898 | 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); | ||||||
5899 | ElementCount ForcedEC = ElementCount::getFixed(EpilogueVectorizationForceVF); | ||||||
5900 | if (LVP.hasPlanWithVF(ForcedEC)) | ||||||
5901 | return {ForcedEC, 0}; | ||||||
5902 | else { | ||||||
5903 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization forced factor is not viable.\n" ;; } } while (false) | ||||||
5904 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization forced factor is not viable.\n" ;; } } while (false) | ||||||
5905 | << "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); | ||||||
5906 | return Result; | ||||||
5907 | } | ||||||
5908 | } | ||||||
5909 | |||||||
5910 | if (TheLoop->getHeader()->getParent()->hasOptSize() || | ||||||
5911 | TheLoop->getHeader()->getParent()->hasMinSize()) { | ||||||
5912 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization skipped due to opt for size.\n" ;; } } while (false) | ||||||
5913 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization skipped due to opt for size.\n" ;; } } while (false) | ||||||
5914 | << "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); | ||||||
5915 | return Result; | ||||||
5916 | } | ||||||
5917 | |||||||
5918 | auto FixedMainLoopVF = ElementCount::getFixed(MainLoopVF.getKnownMinValue()); | ||||||
5919 | if (MainLoopVF.isScalable()) | ||||||
5920 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization using scalable vectors not " "yet supported. Converting to fixed-width (VF=" << FixedMainLoopVF << ") instead\n"; } } while (false) | ||||||
5921 | dbgs() << "LEV: Epilogue vectorization using scalable vectors not "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization using scalable vectors not " "yet supported. Converting to fixed-width (VF=" << FixedMainLoopVF << ") instead\n"; } } while (false) | ||||||
5922 | "yet supported. Converting to fixed-width (VF="do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization using scalable vectors not " "yet supported. Converting to fixed-width (VF=" << FixedMainLoopVF << ") instead\n"; } } while (false) | ||||||
5923 | << FixedMainLoopVF << ") instead\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization using scalable vectors not " "yet supported. Converting to fixed-width (VF=" << FixedMainLoopVF << ") instead\n"; } } while (false); | ||||||
5924 | |||||||
5925 | if (!isEpilogueVectorizationProfitable(FixedMainLoopVF)) { | ||||||
5926 | 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) | ||||||
5927 | "this loop\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Epilogue vectorization is not profitable for " "this loop\n"; } } while (false); | ||||||
5928 | return Result; | ||||||
5929 | } | ||||||
5930 | |||||||
5931 | for (auto &NextVF : ProfitableVFs) | ||||||
5932 | if (ElementCount::isKnownLT(NextVF.Width, FixedMainLoopVF) && | ||||||
5933 | (Result.Width.getFixedValue() == 1 || | ||||||
5934 | isMoreProfitable(NextVF, Result)) && | ||||||
5935 | LVP.hasPlanWithVF(NextVF.Width)) | ||||||
5936 | Result = NextVF; | ||||||
5937 | |||||||
5938 | if (Result != VectorizationFactor::Disabled()) | ||||||
5939 | 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.getFixedValue() << "\n";; } } while (false) | ||||||
5940 | << Result.Width.getFixedValue() << "\n";)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LEV: Vectorizing epilogue loop with VF = " << Result.Width.getFixedValue() << "\n";; } } while (false); | ||||||
5941 | return Result; | ||||||
5942 | } | ||||||
5943 | |||||||
5944 | std::pair<unsigned, unsigned> | ||||||
5945 | LoopVectorizationCostModel::getSmallestAndWidestTypes() { | ||||||
5946 | unsigned MinWidth = -1U; | ||||||
5947 | unsigned MaxWidth = 8; | ||||||
5948 | const DataLayout &DL = TheFunction->getParent()->getDataLayout(); | ||||||
5949 | // For in-loop reductions, no element types are added to ElementTypesInLoop | ||||||
5950 | // if there are no loads/stores in the loop. In this case, check through the | ||||||
5951 | // reduction variables to determine the maximum width. | ||||||
5952 | if (ElementTypesInLoop.empty() && !Legal->getReductionVars().empty()) { | ||||||
5953 | // Reset MaxWidth so that we can find the smallest type used by recurrences | ||||||
5954 | // in the loop. | ||||||
5955 | MaxWidth = -1U; | ||||||
5956 | for (auto &PhiDescriptorPair : Legal->getReductionVars()) { | ||||||
5957 | const RecurrenceDescriptor &RdxDesc = PhiDescriptorPair.second; | ||||||
5958 | // When finding the min width used by the recurrence we need to account | ||||||
5959 | // for casts on the input operands of the recurrence. | ||||||
5960 | MaxWidth = std::min<unsigned>( | ||||||
5961 | MaxWidth, std::min<unsigned>( | ||||||
5962 | RdxDesc.getMinWidthCastToRecurrenceTypeInBits(), | ||||||
5963 | RdxDesc.getRecurrenceType()->getScalarSizeInBits())); | ||||||
5964 | } | ||||||
5965 | } else { | ||||||
5966 | for (Type *T : ElementTypesInLoop) { | ||||||
5967 | MinWidth = std::min<unsigned>( | ||||||
5968 | MinWidth, DL.getTypeSizeInBits(T->getScalarType()).getFixedSize()); | ||||||
5969 | MaxWidth = std::max<unsigned>( | ||||||
5970 | MaxWidth, DL.getTypeSizeInBits(T->getScalarType()).getFixedSize()); | ||||||
5971 | } | ||||||
5972 | } | ||||||
5973 | return {MinWidth, MaxWidth}; | ||||||
5974 | } | ||||||
5975 | |||||||
5976 | void LoopVectorizationCostModel::collectElementTypesForWidening() { | ||||||
5977 | ElementTypesInLoop.clear(); | ||||||
5978 | // For each block. | ||||||
5979 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||||
5980 | // For each instruction in the loop. | ||||||
5981 | for (Instruction &I : BB->instructionsWithoutDebug()) { | ||||||
5982 | Type *T = I.getType(); | ||||||
5983 | |||||||
5984 | // Skip ignored values. | ||||||
5985 | if (ValuesToIgnore.count(&I)) | ||||||
5986 | continue; | ||||||
5987 | |||||||
5988 | // Only examine Loads, Stores and PHINodes. | ||||||
5989 | if (!isa<LoadInst>(I) && !isa<StoreInst>(I) && !isa<PHINode>(I)) | ||||||
5990 | continue; | ||||||
5991 | |||||||
5992 | // Examine PHI nodes that are reduction variables. Update the type to | ||||||
5993 | // account for the recurrence type. | ||||||
5994 | if (auto *PN = dyn_cast<PHINode>(&I)) { | ||||||
5995 | if (!Legal->isReductionVariable(PN)) | ||||||
5996 | continue; | ||||||
5997 | const RecurrenceDescriptor &RdxDesc = | ||||||
5998 | Legal->getReductionVars().find(PN)->second; | ||||||
5999 | if (PreferInLoopReductions || useOrderedReductions(RdxDesc) || | ||||||
6000 | TTI.preferInLoopReduction(RdxDesc.getOpcode(), | ||||||
6001 | RdxDesc.getRecurrenceType(), | ||||||
6002 | TargetTransformInfo::ReductionFlags())) | ||||||
6003 | continue; | ||||||
6004 | T = RdxDesc.getRecurrenceType(); | ||||||
6005 | } | ||||||
6006 | |||||||
6007 | // Examine the stored values. | ||||||
6008 | if (auto *ST = dyn_cast<StoreInst>(&I)) | ||||||
6009 | T = ST->getValueOperand()->getType(); | ||||||
6010 | |||||||
6011 | 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", 6012, __extension__ __PRETTY_FUNCTION__)) | ||||||
6012 | "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", 6012, __extension__ __PRETTY_FUNCTION__)); | ||||||
6013 | |||||||
6014 | ElementTypesInLoop.insert(T); | ||||||
6015 | } | ||||||
6016 | } | ||||||
6017 | } | ||||||
6018 | |||||||
6019 | unsigned LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF, | ||||||
6020 | unsigned LoopCost) { | ||||||
6021 | // -- The interleave heuristics -- | ||||||
6022 | // We interleave the loop in order to expose ILP and reduce the loop overhead. | ||||||
6023 | // There are many micro-architectural considerations that we can't predict | ||||||
6024 | // at this level. For example, frontend pressure (on decode or fetch) due to | ||||||
6025 | // code size, or the number and capabilities of the execution ports. | ||||||
6026 | // | ||||||
6027 | // We use the following heuristics to select the interleave count: | ||||||
6028 | // 1. If the code has reductions, then we interleave to break the cross | ||||||
6029 | // iteration dependency. | ||||||
6030 | // 2. If the loop is really small, then we interleave to reduce the loop | ||||||
6031 | // overhead. | ||||||
6032 | // 3. We don't interleave if we think that we will spill registers to memory | ||||||
6033 | // due to the increased register pressure. | ||||||
6034 | |||||||
6035 | if (!isScalarEpilogueAllowed()) | ||||||
6036 | return 1; | ||||||
6037 | |||||||
6038 | // We used the distance for the interleave count. | ||||||
6039 | if (Legal->getMaxSafeDepDistBytes() != -1U) | ||||||
6040 | return 1; | ||||||
6041 | |||||||
6042 | auto BestKnownTC = getSmallBestKnownTC(*PSE.getSE(), TheLoop); | ||||||
6043 | const bool HasReductions = !Legal->getReductionVars().empty(); | ||||||
6044 | // Do not interleave loops with a relatively small known or estimated trip | ||||||
6045 | // count. But we will interleave when InterleaveSmallLoopScalarReduction is | ||||||
6046 | // enabled, and the code has scalar reductions(HasReductions && VF = 1), | ||||||
6047 | // because with the above conditions interleaving can expose ILP and break | ||||||
6048 | // cross iteration dependences for reductions. | ||||||
6049 | if (BestKnownTC && (*BestKnownTC < TinyTripCountInterleaveThreshold) && | ||||||
6050 | !(InterleaveSmallLoopScalarReduction && HasReductions && VF.isScalar())) | ||||||
6051 | return 1; | ||||||
6052 | |||||||
6053 | RegisterUsage R = calculateRegisterUsage({VF})[0]; | ||||||
6054 | // We divide by these constants so assume that we have at least one | ||||||
6055 | // instruction that uses at least one register. | ||||||
6056 | for (auto& pair : R.MaxLocalUsers) { | ||||||
6057 | pair.second = std::max(pair.second, 1U); | ||||||
6058 | } | ||||||
6059 | |||||||
6060 | // We calculate the interleave count using the following formula. | ||||||
6061 | // Subtract the number of loop invariants from the number of available | ||||||
6062 | // registers. These registers are used by all of the interleaved instances. | ||||||
6063 | // Next, divide the remaining registers by the number of registers that is | ||||||
6064 | // required by the loop, in order to estimate how many parallel instances | ||||||
6065 | // fit without causing spills. All of this is rounded down if necessary to be | ||||||
6066 | // a power of two. We want power of two interleave count to simplify any | ||||||
6067 | // addressing operations or alignment considerations. | ||||||
6068 | // We also want power of two interleave counts to ensure that the induction | ||||||
6069 | // variable of the vector loop wraps to zero, when tail is folded by masking; | ||||||
6070 | // this currently happens when OptForSize, in which case IC is set to 1 above. | ||||||
6071 | unsigned IC = UINT_MAX(2147483647 *2U +1U); | ||||||
6072 | |||||||
6073 | for (auto& pair : R.MaxLocalUsers) { | ||||||
6074 | unsigned TargetNumRegisters = TTI.getNumberOfRegisters(pair.first); | ||||||
6075 | 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) | ||||||
6076 | << " 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) | ||||||
6077 | << 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); | ||||||
6078 | if (VF.isScalar()) { | ||||||
6079 | if (ForceTargetNumScalarRegs.getNumOccurrences() > 0) | ||||||
6080 | TargetNumRegisters = ForceTargetNumScalarRegs; | ||||||
6081 | } else { | ||||||
6082 | if (ForceTargetNumVectorRegs.getNumOccurrences() > 0) | ||||||
6083 | TargetNumRegisters = ForceTargetNumVectorRegs; | ||||||
6084 | } | ||||||
6085 | unsigned MaxLocalUsers = pair.second; | ||||||
6086 | unsigned LoopInvariantRegs = 0; | ||||||
6087 | if (R.LoopInvariantRegs.find(pair.first) != R.LoopInvariantRegs.end()) | ||||||
6088 | LoopInvariantRegs = R.LoopInvariantRegs[pair.first]; | ||||||
6089 | |||||||
6090 | unsigned TmpIC = PowerOf2Floor((TargetNumRegisters - LoopInvariantRegs) / MaxLocalUsers); | ||||||
6091 | // Don't count the induction variable as interleaved. | ||||||
6092 | if (EnableIndVarRegisterHeur) { | ||||||
6093 | TmpIC = | ||||||
6094 | PowerOf2Floor((TargetNumRegisters - LoopInvariantRegs - 1) / | ||||||
6095 | std::max(1U, (MaxLocalUsers - 1))); | ||||||
6096 | } | ||||||
6097 | |||||||
6098 | IC = std::min(IC, TmpIC); | ||||||
6099 | } | ||||||
6100 | |||||||
6101 | // Clamp the interleave ranges to reasonable counts. | ||||||
6102 | unsigned MaxInterleaveCount = | ||||||
6103 | TTI.getMaxInterleaveFactor(VF.getKnownMinValue()); | ||||||
6104 | |||||||
6105 | // Check if the user has overridden the max. | ||||||
6106 | if (VF.isScalar()) { | ||||||
6107 | if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0) | ||||||
6108 | MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor; | ||||||
6109 | } else { | ||||||
6110 | if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0) | ||||||
6111 | MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor; | ||||||
6112 | } | ||||||
6113 | |||||||
6114 | // If trip count is known or estimated compile time constant, limit the | ||||||
6115 | // interleave count to be less than the trip count divided by VF, provided it | ||||||
6116 | // is at least 1. | ||||||
6117 | // | ||||||
6118 | // For scalable vectors we can't know if interleaving is beneficial. It may | ||||||
6119 | // not be beneficial for small loops if none of the lanes in the second vector | ||||||
6120 | // iterations is enabled. However, for larger loops, there is likely to be a | ||||||
6121 | // similar benefit as for fixed-width vectors. For now, we choose to leave | ||||||
6122 | // the InterleaveCount as if vscale is '1', although if some information about | ||||||
6123 | // the vector is known (e.g. min vector size), we can make a better decision. | ||||||
6124 | if (BestKnownTC) { | ||||||
6125 | MaxInterleaveCount = | ||||||
6126 | std::min(*BestKnownTC / VF.getKnownMinValue(), MaxInterleaveCount); | ||||||
6127 | // Make sure MaxInterleaveCount is greater than 0. | ||||||
6128 | MaxInterleaveCount = std::max(1u, MaxInterleaveCount); | ||||||
6129 | } | ||||||
6130 | |||||||
6131 | 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", 6132, __extension__ __PRETTY_FUNCTION__)) | ||||||
6132 | "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", 6132, __extension__ __PRETTY_FUNCTION__)); | ||||||
6133 | |||||||
6134 | // Clamp the calculated IC to be between the 1 and the max interleave count | ||||||
6135 | // that the target and trip count allows. | ||||||
6136 | if (IC > MaxInterleaveCount) | ||||||
6137 | IC = MaxInterleaveCount; | ||||||
6138 | else | ||||||
6139 | // Make sure IC is greater than 0. | ||||||
6140 | IC = std::max(1u, IC); | ||||||
6141 | |||||||
6142 | 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", 6142, __extension__ __PRETTY_FUNCTION__)); | ||||||
6143 | |||||||
6144 | // If we did not calculate the cost for VF (because the user selected the VF) | ||||||
6145 | // then we calculate the cost of VF here. | ||||||
6146 | if (LoopCost == 0) { | ||||||
6147 | InstructionCost C = expectedCost(VF).first; | ||||||
6148 | assert(C.isValid() && "Expected to have chosen a VF with valid cost")(static_cast <bool> (C.isValid() && "Expected to have chosen a VF with valid cost" ) ? void (0) : __assert_fail ("C.isValid() && \"Expected to have chosen a VF with valid cost\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6148, __extension__ __PRETTY_FUNCTION__)); | ||||||
6149 | LoopCost = *C.getValue(); | ||||||
6150 | } | ||||||
6151 | |||||||
6152 | assert(LoopCost && "Non-zero loop cost expected")(static_cast <bool> (LoopCost && "Non-zero loop cost expected" ) ? void (0) : __assert_fail ("LoopCost && \"Non-zero loop cost expected\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6152, __extension__ __PRETTY_FUNCTION__)); | ||||||
6153 | |||||||
6154 | // Interleave if we vectorized this loop and there is a reduction that could | ||||||
6155 | // benefit from interleaving. | ||||||
6156 | if (VF.isVector() && HasReductions) { | ||||||
6157 | 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); | ||||||
6158 | return IC; | ||||||
6159 | } | ||||||
6160 | |||||||
6161 | // Note that if we've already vectorized the loop we will have done the | ||||||
6162 | // runtime check and so interleaving won't require further checks. | ||||||
6163 | bool InterleavingRequiresRuntimePointerCheck = | ||||||
6164 | (VF.isScalar() && Legal->getRuntimePointerChecking()->Need); | ||||||
6165 | |||||||
6166 | // We want to interleave small loops in order to reduce the loop overhead and | ||||||
6167 | // potentially expose ILP opportunities. | ||||||
6168 | 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) | ||||||
6169 | << "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) | ||||||
6170 | << "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); | ||||||
6171 | const bool AggressivelyInterleaveReductions = | ||||||
6172 | TTI.enableAggressiveInterleaving(HasReductions); | ||||||
6173 | if (!InterleavingRequiresRuntimePointerCheck && LoopCost < SmallLoopCost) { | ||||||
6174 | // We assume that the cost overhead is 1 and we use the cost model | ||||||
6175 | // to estimate the cost of the loop and interleave until the cost of the | ||||||
6176 | // loop overhead is about 5% of the cost of the loop. | ||||||
6177 | unsigned SmallIC = | ||||||
6178 | std::min(IC, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost)); | ||||||
6179 | |||||||
6180 | // Interleave until store/load ports (estimated by max interleave count) are | ||||||
6181 | // saturated. | ||||||
6182 | unsigned NumStores = Legal->getNumStores(); | ||||||
6183 | unsigned NumLoads = Legal->getNumLoads(); | ||||||
6184 | unsigned StoresIC = IC / (NumStores ? NumStores : 1); | ||||||
6185 | unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1); | ||||||
6186 | |||||||
6187 | // There is little point in interleaving for reductions containing selects | ||||||
6188 | // and compares when VF=1 since it may just create more overhead than it's | ||||||
6189 | // worth for loops with small trip counts. This is because we still have to | ||||||
6190 | // do the final reduction after the loop. | ||||||
6191 | bool HasSelectCmpReductions = | ||||||
6192 | HasReductions && | ||||||
6193 | any_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool { | ||||||
6194 | const RecurrenceDescriptor &RdxDesc = Reduction.second; | ||||||
6195 | return RecurrenceDescriptor::isSelectCmpRecurrenceKind( | ||||||
6196 | RdxDesc.getRecurrenceKind()); | ||||||
6197 | }); | ||||||
6198 | if (HasSelectCmpReductions) { | ||||||
6199 | 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); | ||||||
6200 | return 1; | ||||||
6201 | } | ||||||
6202 | |||||||
6203 | // If we have a scalar reduction (vector reductions are already dealt with | ||||||
6204 | // by this point), we can increase the critical path length if the loop | ||||||
6205 | // we're interleaving is inside another loop. For tree-wise reductions | ||||||
6206 | // set the limit to 2, and for ordered reductions it's best to disable | ||||||
6207 | // interleaving entirely. | ||||||
6208 | if (HasReductions && TheLoop->getLoopDepth() > 1) { | ||||||
6209 | bool HasOrderedReductions = | ||||||
6210 | any_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool { | ||||||
6211 | const RecurrenceDescriptor &RdxDesc = Reduction.second; | ||||||
6212 | return RdxDesc.isOrdered(); | ||||||
6213 | }); | ||||||
6214 | if (HasOrderedReductions) { | ||||||
6215 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not interleaving scalar ordered reductions.\n" ; } } while (false) | ||||||
6216 | 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); | ||||||
6217 | return 1; | ||||||
6218 | } | ||||||
6219 | |||||||
6220 | unsigned F = static_cast<unsigned>(MaxNestedScalarReductionIC); | ||||||
6221 | SmallIC = std::min(SmallIC, F); | ||||||
6222 | StoresIC = std::min(StoresIC, F); | ||||||
6223 | LoadsIC = std::min(LoadsIC, F); | ||||||
6224 | } | ||||||
6225 | |||||||
6226 | if (EnableLoadStoreRuntimeInterleave && | ||||||
6227 | std::max(StoresIC, LoadsIC) > SmallIC) { | ||||||
6228 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving to saturate store or load ports.\n" ; } } while (false) | ||||||
6229 | 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); | ||||||
6230 | return std::max(StoresIC, LoadsIC); | ||||||
6231 | } | ||||||
6232 | |||||||
6233 | // If there are scalar reductions and TTI has enabled aggressive | ||||||
6234 | // interleaving for reductions, we will interleave to expose ILP. | ||||||
6235 | if (InterleaveSmallLoopScalarReduction && VF.isScalar() && | ||||||
6236 | AggressivelyInterleaveReductions) { | ||||||
6237 | 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); | ||||||
6238 | // Interleave no less than SmallIC but not as aggressive as the normal IC | ||||||
6239 | // to satisfy the rare situation when resources are too limited. | ||||||
6240 | return std::max(IC / 2, SmallIC); | ||||||
6241 | } else { | ||||||
6242 | 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); | ||||||
6243 | return SmallIC; | ||||||
6244 | } | ||||||
6245 | } | ||||||
6246 | |||||||
6247 | // Interleave if this is a large loop (small loops are already dealt with by | ||||||
6248 | // this point) that could benefit from interleaving. | ||||||
6249 | if (AggressivelyInterleaveReductions) { | ||||||
6250 | 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); | ||||||
6251 | return IC; | ||||||
6252 | } | ||||||
6253 | |||||||
6254 | LLVM_DEBUG(dbgs() << "LV: Not Interleaving.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Not Interleaving.\n" ; } } while (false); | ||||||
6255 | return 1; | ||||||
6256 | } | ||||||
6257 | |||||||
6258 | SmallVector<LoopVectorizationCostModel::RegisterUsage, 8> | ||||||
6259 | LoopVectorizationCostModel::calculateRegisterUsage(ArrayRef<ElementCount> VFs) { | ||||||
6260 | // This function calculates the register usage by measuring the highest number | ||||||
6261 | // of values that are alive at a single location. Obviously, this is a very | ||||||
6262 | // rough estimation. We scan the loop in a topological order in order and | ||||||
6263 | // assign a number to each instruction. We use RPO to ensure that defs are | ||||||
6264 | // met before their users. We assume that each instruction that has in-loop | ||||||
6265 | // users starts an interval. We record every time that an in-loop value is | ||||||
6266 | // used, so we have a list of the first and last occurrences of each | ||||||
6267 | // instruction. Next, we transpose this data structure into a multi map that | ||||||
6268 | // holds the list of intervals that *end* at a specific location. This multi | ||||||
6269 | // map allows us to perform a linear search. We scan the instructions linearly | ||||||
6270 | // and record each time that a new interval starts, by placing it in a set. | ||||||
6271 | // If we find this value in the multi-map then we remove it from the set. | ||||||
6272 | // The max register usage is the maximum size of the set. | ||||||
6273 | // We also search for instructions that are defined outside the loop, but are | ||||||
6274 | // used inside the loop. We need this number separately from the max-interval | ||||||
6275 | // usage number because when we unroll, loop-invariant values do not take | ||||||
6276 | // more register. | ||||||
6277 | LoopBlocksDFS DFS(TheLoop); | ||||||
6278 | DFS.perform(LI); | ||||||
6279 | |||||||
6280 | RegisterUsage RU; | ||||||
6281 | |||||||
6282 | // Each 'key' in the map opens a new interval. The values | ||||||
6283 | // of the map are the index of the 'last seen' usage of the | ||||||
6284 | // instruction that is the key. | ||||||
6285 | using IntervalMap = DenseMap<Instruction *, unsigned>; | ||||||
6286 | |||||||
6287 | // Maps instruction to its index. | ||||||
6288 | SmallVector<Instruction *, 64> IdxToInstr; | ||||||
6289 | // Marks the end of each interval. | ||||||
6290 | IntervalMap EndPoint; | ||||||
6291 | // Saves the list of instruction indices that are used in the loop. | ||||||
6292 | SmallPtrSet<Instruction *, 8> Ends; | ||||||
6293 | // Saves the list of values that are used in the loop but are | ||||||
6294 | // defined outside the loop, such as arguments and constants. | ||||||
6295 | SmallPtrSet<Value *, 8> LoopInvariants; | ||||||
6296 | |||||||
6297 | for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) { | ||||||
6298 | for (Instruction &I : BB->instructionsWithoutDebug()) { | ||||||
6299 | IdxToInstr.push_back(&I); | ||||||
6300 | |||||||
6301 | // Save the end location of each USE. | ||||||
6302 | for (Value *U : I.operands()) { | ||||||
6303 | auto *Instr = dyn_cast<Instruction>(U); | ||||||
6304 | |||||||
6305 | // Ignore non-instruction values such as arguments, constants, etc. | ||||||
6306 | if (!Instr) | ||||||
6307 | continue; | ||||||
6308 | |||||||
6309 | // If this instruction is outside the loop then record it and continue. | ||||||
6310 | if (!TheLoop->contains(Instr)) { | ||||||
6311 | LoopInvariants.insert(Instr); | ||||||
6312 | continue; | ||||||
6313 | } | ||||||
6314 | |||||||
6315 | // Overwrite previous end points. | ||||||
6316 | EndPoint[Instr] = IdxToInstr.size(); | ||||||
6317 | Ends.insert(Instr); | ||||||
6318 | } | ||||||
6319 | } | ||||||
6320 | } | ||||||
6321 | |||||||
6322 | // Saves the list of intervals that end with the index in 'key'. | ||||||
6323 | using InstrList = SmallVector<Instruction *, 2>; | ||||||
6324 | DenseMap<unsigned, InstrList> TransposeEnds; | ||||||
6325 | |||||||
6326 | // Transpose the EndPoints to a list of values that end at each index. | ||||||
6327 | for (auto &Interval : EndPoint) | ||||||
6328 | TransposeEnds[Interval.second].push_back(Interval.first); | ||||||
6329 | |||||||
6330 | SmallPtrSet<Instruction *, 8> OpenIntervals; | ||||||
6331 | SmallVector<RegisterUsage, 8> RUs(VFs.size()); | ||||||
6332 | SmallVector<SmallMapVector<unsigned, unsigned, 4>, 8> MaxUsages(VFs.size()); | ||||||
6333 | |||||||
6334 | 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); | ||||||
6335 | |||||||
6336 | // A lambda that gets the register usage for the given type and VF. | ||||||
6337 | const auto &TTICapture = TTI; | ||||||
6338 | auto GetRegUsage = [&TTICapture](Type *Ty, ElementCount VF) -> unsigned { | ||||||
6339 | if (Ty->isTokenTy() || !VectorType::isValidElementType(Ty)) | ||||||
6340 | return 0; | ||||||
6341 | InstructionCost::CostType RegUsage = | ||||||
6342 | *TTICapture.getRegUsageForType(VectorType::get(Ty, VF)).getValue(); | ||||||
6343 | assert(RegUsage >= 0 && RegUsage <= std::numeric_limits<unsigned>::max() &&(static_cast <bool> (RegUsage >= 0 && RegUsage <= std::numeric_limits<unsigned>::max() && "Nonsensical values for register usage." ) ? void (0) : __assert_fail ("RegUsage >= 0 && RegUsage <= std::numeric_limits<unsigned>::max() && \"Nonsensical values for register usage.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6344, __extension__ __PRETTY_FUNCTION__)) | ||||||
6344 | "Nonsensical values for register usage.")(static_cast <bool> (RegUsage >= 0 && RegUsage <= std::numeric_limits<unsigned>::max() && "Nonsensical values for register usage." ) ? void (0) : __assert_fail ("RegUsage >= 0 && RegUsage <= std::numeric_limits<unsigned>::max() && \"Nonsensical values for register usage.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6344, __extension__ __PRETTY_FUNCTION__)); | ||||||
6345 | return RegUsage; | ||||||
6346 | }; | ||||||
6347 | |||||||
6348 | for (unsigned int i = 0, s = IdxToInstr.size(); i < s; ++i) { | ||||||
6349 | Instruction *I = IdxToInstr[i]; | ||||||
6350 | |||||||
6351 | // Remove all of the instructions that end at this location. | ||||||
6352 | InstrList &List = TransposeEnds[i]; | ||||||
6353 | for (Instruction *ToRemove : List) | ||||||
6354 | OpenIntervals.erase(ToRemove); | ||||||
6355 | |||||||
6356 | // Ignore instructions that are never used within the loop. | ||||||
6357 | if (!Ends.count(I)) | ||||||
6358 | continue; | ||||||
6359 | |||||||
6360 | // Skip ignored values. | ||||||
6361 | if (ValuesToIgnore.count(I)) | ||||||
6362 | continue; | ||||||
6363 | |||||||
6364 | // For each VF find the maximum usage of registers. | ||||||
6365 | for (unsigned j = 0, e = VFs.size(); j < e; ++j) { | ||||||
6366 | // Count the number of live intervals. | ||||||
6367 | SmallMapVector<unsigned, unsigned, 4> RegUsage; | ||||||
6368 | |||||||
6369 | if (VFs[j].isScalar()) { | ||||||
6370 | for (auto Inst : OpenIntervals) { | ||||||
6371 | unsigned ClassID = TTI.getRegisterClassForType(false, Inst->getType()); | ||||||
6372 | if (RegUsage.find(ClassID) == RegUsage.end()) | ||||||
6373 | RegUsage[ClassID] = 1; | ||||||
6374 | else | ||||||
6375 | RegUsage[ClassID] += 1; | ||||||
6376 | } | ||||||
6377 | } else { | ||||||
6378 | collectUniformsAndScalars(VFs[j]); | ||||||
6379 | for (auto Inst : OpenIntervals) { | ||||||
6380 | // Skip ignored values for VF > 1. | ||||||
6381 | if (VecValuesToIgnore.count(Inst)) | ||||||
6382 | continue; | ||||||
6383 | if (isScalarAfterVectorization(Inst, VFs[j])) { | ||||||
6384 | unsigned ClassID = TTI.getRegisterClassForType(false, Inst->getType()); | ||||||
6385 | if (RegUsage.find(ClassID) == RegUsage.end()) | ||||||
6386 | RegUsage[ClassID] = 1; | ||||||
6387 | else | ||||||
6388 | RegUsage[ClassID] += 1; | ||||||
6389 | } else { | ||||||
6390 | unsigned ClassID = TTI.getRegisterClassForType(true, Inst->getType()); | ||||||
6391 | if (RegUsage.find(ClassID) == RegUsage.end()) | ||||||
6392 | RegUsage[ClassID] = GetRegUsage(Inst->getType(), VFs[j]); | ||||||
6393 | else | ||||||
6394 | RegUsage[ClassID] += GetRegUsage(Inst->getType(), VFs[j]); | ||||||
6395 | } | ||||||
6396 | } | ||||||
6397 | } | ||||||
6398 | |||||||
6399 | for (auto& pair : RegUsage) { | ||||||
6400 | if (MaxUsages[j].find(pair.first) != MaxUsages[j].end()) | ||||||
6401 | MaxUsages[j][pair.first] = std::max(MaxUsages[j][pair.first], pair.second); | ||||||
6402 | else | ||||||
6403 | MaxUsages[j][pair.first] = pair.second; | ||||||
6404 | } | ||||||
6405 | } | ||||||
6406 | |||||||
6407 | 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) | ||||||
6408 | << OpenIntervals.size() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV(REG): At #" << i << " Interval # " << OpenIntervals.size() << '\n'; } } while (false); | ||||||
6409 | |||||||
6410 | // Add the current instruction to the list of open intervals. | ||||||
6411 | OpenIntervals.insert(I); | ||||||
6412 | } | ||||||
6413 | |||||||
6414 | for (unsigned i = 0, e = VFs.size(); i < e; ++i) { | ||||||
6415 | SmallMapVector<unsigned, unsigned, 4> Invariant; | ||||||
6416 | |||||||
6417 | for (auto Inst : LoopInvariants) { | ||||||
6418 | unsigned Usage = | ||||||
6419 | VFs[i].isScalar() ? 1 : GetRegUsage(Inst->getType(), VFs[i]); | ||||||
6420 | unsigned ClassID = | ||||||
6421 | TTI.getRegisterClassForType(VFs[i].isVector(), Inst->getType()); | ||||||
6422 | if (Invariant.find(ClassID) == Invariant.end()) | ||||||
6423 | Invariant[ClassID] = Usage; | ||||||
6424 | else | ||||||
6425 | Invariant[ClassID] += Usage; | ||||||
6426 | } | ||||||
6427 | |||||||
6428 | 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) | ||||||
6429 | 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) | ||||||
6430 | 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) | ||||||
6431 | << " 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) | ||||||
6432 | 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) | ||||||
6433 | 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) | ||||||
6434 | << 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) | ||||||
6435 | << " 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) | ||||||
6436 | }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) | ||||||
6437 | 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) | ||||||
6438 | << " 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) | ||||||
6439 | 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) | ||||||
6440 | 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) | ||||||
6441 | << 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) | ||||||
6442 | << " 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) | ||||||
6443 | }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) | ||||||
6444 | })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); | ||||||
6445 | |||||||
6446 | RU.LoopInvariantRegs = Invariant; | ||||||
6447 | RU.MaxLocalUsers = MaxUsages[i]; | ||||||
6448 | RUs[i] = RU; | ||||||
6449 | } | ||||||
6450 | |||||||
6451 | return RUs; | ||||||
6452 | } | ||||||
6453 | |||||||
6454 | bool LoopVectorizationCostModel::useEmulatedMaskMemRefHack(Instruction *I, | ||||||
6455 | ElementCount VF) { | ||||||
6456 | // TODO: Cost model for emulated masked load/store is completely | ||||||
6457 | // broken. This hack guides the cost model to use an artificially | ||||||
6458 | // high enough value to practically disable vectorization with such | ||||||
6459 | // operations, except where previously deployed legality hack allowed | ||||||
6460 | // using very low cost values. This is to avoid regressions coming simply | ||||||
6461 | // from moving "masked load/store" check from legality to cost model. | ||||||
6462 | // Masked Load/Gather emulation was previously never allowed. | ||||||
6463 | // Limited number of Masked Store/Scatter emulation was allowed. | ||||||
6464 | assert(isPredicatedInst(I, VF) && "Expecting a scalar emulated instruction")(static_cast <bool> (isPredicatedInst(I, VF) && "Expecting a scalar emulated instruction") ? void (0) : __assert_fail ("isPredicatedInst(I, VF) && \"Expecting a scalar emulated instruction\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 6464, __extension__ __PRETTY_FUNCTION__)); | ||||||
6465 | return isa<LoadInst>(I) || | ||||||
6466 | (isa<StoreInst>(I) && | ||||||
6467 | NumPredStores > NumberOfStoresToPredicate); | ||||||
6468 | } | ||||||
6469 | |||||||
6470 | void LoopVectorizationCostModel::collectInstsToScalarize(ElementCount VF) { | ||||||
6471 | // If we aren't vectorizing the loop, or if we've already collected the | ||||||
6472 | // instructions to scalarize, there's nothing to do. Collection may already | ||||||
6473 | // have occurred if we have a user-selected VF and are now computing the | ||||||
6474 | // expected cost for interleaving. | ||||||
6475 | if (VF.isScalar() || VF.isZero() || | ||||||
6476 | InstsToScalarize.find(VF) != InstsToScalarize.end()) | ||||||
6477 | return; | ||||||
6478 | |||||||
6479 | // Initialize a mapping for VF in InstsToScalalarize. If we find that it's | ||||||
6480 | // not profitable to scalarize any instructions, the presence of VF in the | ||||||
6481 | // map will indicate that we've analyzed it already. | ||||||
6482 | ScalarCostsTy &ScalarCostsVF = InstsToScalarize[VF]; | ||||||
6483 | |||||||
6484 | // Find all the instructions that are scalar with predication in the loop and | ||||||
6485 | // determine if it would be better to not if-convert the blocks they are in. | ||||||
6486 | // If so, we also record the instructions to scalarize. | ||||||
6487 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||||
6488 | if (!blockNeedsPredicationForAnyReason(BB)) | ||||||
6489 | continue; | ||||||
6490 | for (Instruction &I : *BB) | ||||||
6491 | if (isScalarWithPredication(&I, VF)) { | ||||||
6492 | ScalarCostsTy ScalarCosts; | ||||||
6493 | // Do not apply discount if scalable, because that would lead to | ||||||
6494 | // invalid scalarization costs. | ||||||
6495 | // Do not apply discount logic if hacked cost is needed | ||||||
6496 | // for emulated masked memrefs. | ||||||
6497 | if (!VF.isScalable() && !useEmulatedMaskMemRefHack(&I, VF) && | ||||||
6498 | computePredInstDiscount(&I, ScalarCosts, VF) >= 0) | ||||||
6499 | ScalarCostsVF.insert(ScalarCosts.begin(), ScalarCosts.end()); | ||||||
6500 | // Remember that BB will remain after vectorization. | ||||||
6501 | PredicatedBBsAfterVectorization.insert(BB); | ||||||
6502 | } | ||||||
6503 | } | ||||||
6504 | } | ||||||
6505 | |||||||
6506 | int LoopVectorizationCostModel::computePredInstDiscount( | ||||||
6507 | Instruction *PredInst, ScalarCostsTy &ScalarCosts, ElementCount VF) { | ||||||
6508 | 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", 6509, __extension__ __PRETTY_FUNCTION__)) | ||||||
6509 | "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", 6509, __extension__ __PRETTY_FUNCTION__)); | ||||||
6510 | |||||||
6511 | // Initialize the discount to zero, meaning that the scalar version and the | ||||||
6512 | // vector version cost the same. | ||||||
6513 | InstructionCost Discount = 0; | ||||||
6514 | |||||||
6515 | // Holds instructions to analyze. The instructions we visit are mapped in | ||||||
6516 | // ScalarCosts. Those instructions are the ones that would be scalarized if | ||||||
6517 | // we find that the scalar version costs less. | ||||||
6518 | SmallVector<Instruction *, 8> Worklist; | ||||||
6519 | |||||||
6520 | // Returns true if the given instruction can be scalarized. | ||||||
6521 | auto canBeScalarized = [&](Instruction *I) -> bool { | ||||||
6522 | // We only attempt to scalarize instructions forming a single-use chain | ||||||
6523 | // from the original predicated block that would otherwise be vectorized. | ||||||
6524 | // Although not strictly necessary, we give up on instructions we know will | ||||||
6525 | // already be scalar to avoid traversing chains that are unlikely to be | ||||||
6526 | // beneficial. | ||||||
6527 | if (!I->hasOneUse() || PredInst->getParent() != I->getParent() || | ||||||
6528 | isScalarAfterVectorization(I, VF)) | ||||||
6529 | return false; | ||||||
6530 | |||||||
6531 | // If the instruction is scalar with predication, it will be analyzed | ||||||
6532 | // separately. We ignore it within the context of PredInst. | ||||||
6533 | if (isScalarWithPredication(I, VF)) | ||||||
6534 | return false; | ||||||
6535 | |||||||
6536 | // If any of the instruction's operands are uniform after vectorization, | ||||||
6537 | // the instruction cannot be scalarized. This prevents, for example, a | ||||||
6538 | // masked load from being scalarized. | ||||||
6539 | // | ||||||
6540 | // We assume we will only emit a value for lane zero of an instruction | ||||||
6541 | // marked uniform after vectorization, rather than VF identical values. | ||||||
6542 | // Thus, if we scalarize an instruction that uses a uniform, we would | ||||||
6543 | // create uses of values corresponding to the lanes we aren't emitting code | ||||||
6544 | // for. This behavior can be changed by allowing getScalarValue to clone | ||||||
6545 | // the lane zero values for uniforms rather than asserting. | ||||||
6546 | for (Use &U : I->operands()) | ||||||
6547 | if (auto *J = dyn_cast<Instruction>(U.get())) | ||||||
6548 | if (isUniformAfterVectorization(J, VF)) | ||||||
6549 | return false; | ||||||
6550 | |||||||
6551 | // Otherwise, we can scalarize the instruction. | ||||||
6552 | return true; | ||||||
6553 | }; | ||||||
6554 | |||||||
6555 | // Compute the expected cost discount from scalarizing the entire expression | ||||||
6556 | // feeding the predicated instruction. We currently only consider expressions | ||||||
6557 | // that are single-use instruction chains. | ||||||
6558 | Worklist.push_back(PredInst); | ||||||
6559 | while (!Worklist.empty()) { | ||||||
6560 | Instruction *I = Worklist.pop_back_val(); | ||||||
6561 | |||||||
6562 | // If we've already analyzed the instruction, there's nothing to do. | ||||||
6563 | if (ScalarCosts.find(I) != ScalarCosts.end()) | ||||||
6564 | continue; | ||||||
6565 | |||||||
6566 | // Compute the cost of the vector instruction. Note that this cost already | ||||||
6567 | // includes the scalarization overhead of the predicated instruction. | ||||||
6568 | InstructionCost VectorCost = getInstructionCost(I, VF).first; | ||||||
6569 | |||||||
6570 | // Compute the cost of the scalarized instruction. This cost is the cost of | ||||||
6571 | // the instruction as if it wasn't if-converted and instead remained in the | ||||||
6572 | // predicated block. We will scale this cost by block probability after | ||||||
6573 | // computing the scalarization overhead. | ||||||
6574 | InstructionCost ScalarCost = | ||||||
6575 | VF.getFixedValue() * | ||||||
6576 | getInstructionCost(I, ElementCount::getFixed(1)).first; | ||||||
6577 | |||||||
6578 | // Compute the scalarization overhead of needed insertelement instructions | ||||||
6579 | // and phi nodes. | ||||||
6580 | if (isScalarWithPredication(I, VF) && !I->getType()->isVoidTy()) { | ||||||
6581 | ScalarCost += TTI.getScalarizationOverhead( | ||||||
6582 | cast<VectorType>(ToVectorTy(I->getType(), VF)), | ||||||
6583 | APInt::getAllOnes(VF.getFixedValue()), true, false); | ||||||
6584 | ScalarCost += | ||||||
6585 | VF.getFixedValue() * | ||||||
6586 | TTI.getCFInstrCost(Instruction::PHI, TTI::TCK_RecipThroughput); | ||||||
6587 | } | ||||||
6588 | |||||||
6589 | // Compute the scalarization overhead of needed extractelement | ||||||
6590 | // instructions. For each of the instruction's operands, if the operand can | ||||||
6591 | // be scalarized, add it to the worklist; otherwise, account for the | ||||||
6592 | // overhead. | ||||||
6593 | for (Use &U : I->operands()) | ||||||
6594 | if (auto *J = dyn_cast<Instruction>(U.get())) { | ||||||
6595 | 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", 6596, __extension__ __PRETTY_FUNCTION__)) | ||||||
6596 | "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", 6596, __extension__ __PRETTY_FUNCTION__)); | ||||||
6597 | if (canBeScalarized(J)) | ||||||
6598 | Worklist.push_back(J); | ||||||
6599 | else if (needsExtract(J, VF)) { | ||||||
6600 | ScalarCost += TTI.getScalarizationOverhead( | ||||||
6601 | cast<VectorType>(ToVectorTy(J->getType(), VF)), | ||||||
6602 | APInt::getAllOnes(VF.getFixedValue()), false, true); | ||||||
6603 | } | ||||||
6604 | } | ||||||
6605 | |||||||
6606 | // Scale the total scalar cost by block probability. | ||||||
6607 | ScalarCost /= getReciprocalPredBlockProb(); | ||||||
6608 | |||||||
6609 | // Compute the discount. A non-negative discount means the vector version | ||||||
6610 | // of the instruction costs more, and scalarizing would be beneficial. | ||||||
6611 | Discount += VectorCost - ScalarCost; | ||||||
6612 | ScalarCosts[I] = ScalarCost; | ||||||
6613 | } | ||||||
6614 | |||||||
6615 | return *Discount.getValue(); | ||||||
6616 | } | ||||||
6617 | |||||||
6618 | LoopVectorizationCostModel::VectorizationCostTy | ||||||
6619 | LoopVectorizationCostModel::expectedCost( | ||||||
6620 | ElementCount VF, SmallVectorImpl<InstructionVFPair> *Invalid) { | ||||||
6621 | VectorizationCostTy Cost; | ||||||
6622 | |||||||
6623 | // For each block. | ||||||
6624 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||||
6625 | VectorizationCostTy BlockCost; | ||||||
6626 | |||||||
6627 | // For each instruction in the old loop. | ||||||
6628 | for (Instruction &I : BB->instructionsWithoutDebug()) { | ||||||
6629 | // Skip ignored values. | ||||||
6630 | if (ValuesToIgnore.count(&I) || | ||||||
6631 | (VF.isVector() && VecValuesToIgnore.count(&I))) | ||||||
6632 | continue; | ||||||
6633 | |||||||
6634 | VectorizationCostTy C = getInstructionCost(&I, VF); | ||||||
6635 | |||||||
6636 | // Check if we should override the cost. | ||||||
6637 | if (C.first.isValid() && | ||||||
6638 | ForceTargetInstructionCost.getNumOccurrences() > 0) | ||||||
6639 | C.first = InstructionCost(ForceTargetInstructionCost); | ||||||
6640 | |||||||
6641 | // Keep a list of instructions with invalid costs. | ||||||
6642 | if (Invalid && !C.first.isValid()) | ||||||
6643 | Invalid->emplace_back(&I, VF); | ||||||
6644 | |||||||
6645 | BlockCost.first += C.first; | ||||||
6646 | BlockCost.second |= C.second; | ||||||
6647 | 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) | ||||||
6648 | << " 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) | ||||||
6649 | << '\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); | ||||||
6650 | } | ||||||
6651 | |||||||
6652 | // If we are vectorizing a predicated block, it will have been | ||||||
6653 | // if-converted. This means that the block's instructions (aside from | ||||||
6654 | // stores and instructions that may divide by zero) will now be | ||||||
6655 | // unconditionally executed. For the scalar case, we may not always execute | ||||||
6656 | // the predicated block, if it is an if-else block. Thus, scale the block's | ||||||
6657 | // cost by the probability of executing it. blockNeedsPredication from | ||||||
6658 | // Legal is used so as to not include all blocks in tail folded loops. | ||||||
6659 | if (VF.isScalar() && Legal->blockNeedsPredication(BB)) | ||||||
6660 | BlockCost.first /= getReciprocalPredBlockProb(); | ||||||
6661 | |||||||
6662 | Cost.first += BlockCost.first; | ||||||
6663 | Cost.second |= BlockCost.second; | ||||||
6664 | } | ||||||
6665 | |||||||
6666 | return Cost; | ||||||
6667 | } | ||||||
6668 | |||||||
6669 | /// Gets Address Access SCEV after verifying that the access pattern | ||||||
6670 | /// is loop invariant except the induction variable dependence. | ||||||
6671 | /// | ||||||
6672 | /// This SCEV can be sent to the Target in order to estimate the address | ||||||
6673 | /// calculation cost. | ||||||
6674 | static const SCEV *getAddressAccessSCEV( | ||||||
6675 | Value *Ptr, | ||||||
6676 | LoopVectorizationLegality *Legal, | ||||||
6677 | PredicatedScalarEvolution &PSE, | ||||||
6678 | const Loop *TheLoop) { | ||||||
6679 | |||||||
6680 | auto *Gep = dyn_cast<GetElementPtrInst>(Ptr); | ||||||
6681 | if (!Gep) | ||||||
6682 | return nullptr; | ||||||
6683 | |||||||
6684 | // We are looking for a gep with all loop invariant indices except for one | ||||||
6685 | // which should be an induction variable. | ||||||
6686 | auto SE = PSE.getSE(); | ||||||
6687 | unsigned NumOperands = Gep->getNumOperands(); | ||||||
6688 | for (unsigned i = 1; i < NumOperands; ++i) { | ||||||
6689 | Value *Opd = Gep->getOperand(i); | ||||||
6690 | if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) && | ||||||
6691 | !Legal->isInductionVariable(Opd)) | ||||||
6692 | return nullptr; | ||||||
6693 | } | ||||||
6694 | |||||||
6695 | // Now we know we have a GEP ptr, %inv, %ind, %inv. return the Ptr SCEV. | ||||||
6696 | return PSE.getSCEV(Ptr); | ||||||
6697 | } | ||||||
6698 | |||||||
6699 | static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) { | ||||||
6700 | return Legal->hasStride(I->getOperand(0)) || | ||||||
6701 | Legal->hasStride(I->getOperand(1)); | ||||||
6702 | } | ||||||
6703 | |||||||
6704 | InstructionCost | ||||||
6705 | LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I, | ||||||
6706 | ElementCount VF) { | ||||||
6707 | 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", 6708, __extension__ __PRETTY_FUNCTION__)) | ||||||
6708 | "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", 6708, __extension__ __PRETTY_FUNCTION__)); | ||||||
6709 | if (VF.isScalable()) | ||||||
6710 | return InstructionCost::getInvalid(); | ||||||
6711 | |||||||
6712 | Type *ValTy = getLoadStoreType(I); | ||||||
6713 | auto SE = PSE.getSE(); | ||||||
6714 | |||||||
6715 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||
6716 | Value *Ptr = getLoadStorePointerOperand(I); | ||||||
6717 | Type *PtrTy = ToVectorTy(Ptr->getType(), VF); | ||||||
6718 | // NOTE: PtrTy is a vector to signal `TTI::getAddressComputationCost` | ||||||
6719 | // that it is being called from this specific place. | ||||||
6720 | |||||||
6721 | // Figure out whether the access is strided and get the stride value | ||||||
6722 | // if it's known in compile time | ||||||
6723 | const SCEV *PtrSCEV = getAddressAccessSCEV(Ptr, Legal, PSE, TheLoop); | ||||||
6724 | |||||||
6725 | // Get the cost of the scalar memory instruction and address computation. | ||||||
6726 | InstructionCost Cost = | ||||||
6727 | VF.getKnownMinValue() * TTI.getAddressComputationCost(PtrTy, SE, PtrSCEV); | ||||||
6728 | |||||||
6729 | // Don't pass *I here, since it is scalar but will actually be part of a | ||||||
6730 | // vectorized loop where the user of it is a vectorized instruction. | ||||||
6731 | const Align Alignment = getLoadStoreAlignment(I); | ||||||
6732 | Cost += VF.getKnownMinValue() * | ||||||
6733 | TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), Alignment, | ||||||
6734 | AS, TTI::TCK_RecipThroughput); | ||||||
6735 | |||||||
6736 | // Get the overhead of the extractelement and insertelement instructions | ||||||
6737 | // we might create due to scalarization. | ||||||
6738 | Cost += getScalarizationOverhead(I, VF); | ||||||
6739 | |||||||
6740 | // If we have a predicated load/store, it will need extra i1 extracts and | ||||||
6741 | // conditional branches, but may not be executed for each vector lane. Scale | ||||||
6742 | // the cost by the probability of executing the predicated block. | ||||||
6743 | if (isPredicatedInst(I, VF)) { | ||||||
6744 | Cost /= getReciprocalPredBlockProb(); | ||||||
6745 | |||||||
6746 | // Add the cost of an i1 extract and a branch | ||||||
6747 | auto *Vec_i1Ty = | ||||||
6748 | VectorType::get(IntegerType::getInt1Ty(ValTy->getContext()), VF); | ||||||
6749 | Cost += TTI.getScalarizationOverhead( | ||||||
6750 | Vec_i1Ty, APInt::getAllOnes(VF.getKnownMinValue()), | ||||||
6751 | /*Insert=*/false, /*Extract=*/true); | ||||||
6752 | Cost += TTI.getCFInstrCost(Instruction::Br, TTI::TCK_RecipThroughput); | ||||||
6753 | |||||||
6754 | if (useEmulatedMaskMemRefHack(I, VF)) | ||||||
6755 | // Artificially setting to a high enough value to practically disable | ||||||
6756 | // vectorization with such operations. | ||||||
6757 | Cost = 3000000; | ||||||
6758 | } | ||||||
6759 | |||||||
6760 | return Cost; | ||||||
6761 | } | ||||||
6762 | |||||||
6763 | InstructionCost | ||||||
6764 | LoopVectorizationCostModel::getConsecutiveMemOpCost(Instruction *I, | ||||||
6765 | ElementCount VF) { | ||||||
6766 | Type *ValTy = getLoadStoreType(I); | ||||||
6767 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | ||||||
6768 | Value *Ptr = getLoadStorePointerOperand(I); | ||||||
6769 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||
6770 | int ConsecutiveStride = Legal->isConsecutivePtr(ValTy, Ptr); | ||||||
6771 | enum TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||
6772 | |||||||
6773 | 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", 6774, __extension__ __PRETTY_FUNCTION__)) | ||||||
6774 | "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", 6774, __extension__ __PRETTY_FUNCTION__)); | ||||||
6775 | const Align Alignment = getLoadStoreAlignment(I); | ||||||
6776 | InstructionCost Cost = 0; | ||||||
6777 | if (Legal->isMaskRequired(I)) | ||||||
6778 | Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS, | ||||||
6779 | CostKind); | ||||||
6780 | else | ||||||
6781 | Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS, | ||||||
6782 | CostKind, I); | ||||||
6783 | |||||||
6784 | bool Reverse = ConsecutiveStride < 0; | ||||||
6785 | if (Reverse) | ||||||
6786 | Cost += | ||||||
6787 | TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, None, 0); | ||||||
6788 | return Cost; | ||||||
6789 | } | ||||||
6790 | |||||||
6791 | InstructionCost | ||||||
6792 | LoopVectorizationCostModel::getUniformMemOpCost(Instruction *I, | ||||||
6793 | ElementCount VF) { | ||||||
6794 | assert(Legal->isUniformMemOp(*I))(static_cast <bool> (Legal->isUniformMemOp(*I)) ? void (0) : __assert_fail ("Legal->isUniformMemOp(*I)", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 6794, __extension__ __PRETTY_FUNCTION__)); | ||||||
6795 | |||||||
6796 | Type *ValTy = getLoadStoreType(I); | ||||||
6797 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | ||||||
6798 | const Align Alignment = getLoadStoreAlignment(I); | ||||||
6799 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||
6800 | enum TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||
6801 | if (isa<LoadInst>(I)) { | ||||||
6802 | return TTI.getAddressComputationCost(ValTy) + | ||||||
6803 | TTI.getMemoryOpCost(Instruction::Load, ValTy, Alignment, AS, | ||||||
6804 | CostKind) + | ||||||
6805 | TTI.getShuffleCost(TargetTransformInfo::SK_Broadcast, VectorTy); | ||||||
6806 | } | ||||||
6807 | StoreInst *SI = cast<StoreInst>(I); | ||||||
6808 | |||||||
6809 | bool isLoopInvariantStoreValue = Legal->isUniform(SI->getValueOperand()); | ||||||
6810 | return TTI.getAddressComputationCost(ValTy) + | ||||||
6811 | TTI.getMemoryOpCost(Instruction::Store, ValTy, Alignment, AS, | ||||||
6812 | CostKind) + | ||||||
6813 | (isLoopInvariantStoreValue | ||||||
6814 | ? 0 | ||||||
6815 | : TTI.getVectorInstrCost(Instruction::ExtractElement, VectorTy, | ||||||
6816 | VF.getKnownMinValue() - 1)); | ||||||
6817 | } | ||||||
6818 | |||||||
6819 | InstructionCost | ||||||
6820 | LoopVectorizationCostModel::getGatherScatterCost(Instruction *I, | ||||||
6821 | ElementCount VF) { | ||||||
6822 | Type *ValTy = getLoadStoreType(I); | ||||||
6823 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | ||||||
6824 | const Align Alignment = getLoadStoreAlignment(I); | ||||||
6825 | const Value *Ptr = getLoadStorePointerOperand(I); | ||||||
6826 | |||||||
6827 | return TTI.getAddressComputationCost(VectorTy) + | ||||||
6828 | TTI.getGatherScatterOpCost( | ||||||
6829 | I->getOpcode(), VectorTy, Ptr, Legal->isMaskRequired(I), Alignment, | ||||||
6830 | TargetTransformInfo::TCK_RecipThroughput, I); | ||||||
6831 | } | ||||||
6832 | |||||||
6833 | InstructionCost | ||||||
6834 | LoopVectorizationCostModel::getInterleaveGroupCost(Instruction *I, | ||||||
6835 | ElementCount VF) { | ||||||
6836 | // TODO: Once we have support for interleaving with scalable vectors | ||||||
6837 | // we can calculate the cost properly here. | ||||||
6838 | if (VF.isScalable()) | ||||||
6839 | return InstructionCost::getInvalid(); | ||||||
6840 | |||||||
6841 | Type *ValTy = getLoadStoreType(I); | ||||||
6842 | auto *VectorTy = cast<VectorType>(ToVectorTy(ValTy, VF)); | ||||||
6843 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||
6844 | |||||||
6845 | auto Group = getInterleavedAccessGroup(I); | ||||||
6846 | 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", 6846, __extension__ __PRETTY_FUNCTION__)); | ||||||
6847 | |||||||
6848 | unsigned InterleaveFactor = Group->getFactor(); | ||||||
6849 | auto *WideVecTy = VectorType::get(ValTy, VF * InterleaveFactor); | ||||||
6850 | |||||||
6851 | // Holds the indices of existing members in the interleaved group. | ||||||
6852 | SmallVector<unsigned, 4> Indices; | ||||||
6853 | for (unsigned IF = 0; IF < InterleaveFactor; IF++) | ||||||
6854 | if (Group->getMember(IF)) | ||||||
6855 | Indices.push_back(IF); | ||||||
6856 | |||||||
6857 | // Calculate the cost of the whole interleaved group. | ||||||
6858 | bool UseMaskForGaps = | ||||||
6859 | (Group->requiresScalarEpilogue() && !isScalarEpilogueAllowed()) || | ||||||
6860 | (isa<StoreInst>(I) && (Group->getNumMembers() < Group->getFactor())); | ||||||
6861 | InstructionCost Cost = TTI.getInterleavedMemoryOpCost( | ||||||
6862 | I->getOpcode(), WideVecTy, Group->getFactor(), Indices, Group->getAlign(), | ||||||
6863 | AS, TTI::TCK_RecipThroughput, Legal->isMaskRequired(I), UseMaskForGaps); | ||||||
6864 | |||||||
6865 | if (Group->isReverse()) { | ||||||
6866 | // TODO: Add support for reversed masked interleaved access. | ||||||
6867 | 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", 6868, __extension__ __PRETTY_FUNCTION__)) | ||||||
6868 | "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", 6868, __extension__ __PRETTY_FUNCTION__)); | ||||||
6869 | Cost += | ||||||
6870 | Group->getNumMembers() * | ||||||
6871 | TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, None, 0); | ||||||
6872 | } | ||||||
6873 | return Cost; | ||||||
6874 | } | ||||||
6875 | |||||||
6876 | Optional<InstructionCost> LoopVectorizationCostModel::getReductionPatternCost( | ||||||
6877 | Instruction *I, ElementCount VF, Type *Ty, TTI::TargetCostKind CostKind) { | ||||||
6878 | using namespace llvm::PatternMatch; | ||||||
6879 | // Early exit for no inloop reductions | ||||||
6880 | if (InLoopReductionChains.empty() || VF.isScalar() || !isa<VectorType>(Ty)) | ||||||
6881 | return None; | ||||||
6882 | auto *VectorTy = cast<VectorType>(Ty); | ||||||
6883 | |||||||
6884 | // We are looking for a pattern of, and finding the minimal acceptable cost: | ||||||
6885 | // reduce(mul(ext(A), ext(B))) or | ||||||
6886 | // reduce(mul(A, B)) or | ||||||
6887 | // reduce(ext(A)) or | ||||||
6888 | // reduce(A). | ||||||
6889 | // The basic idea is that we walk down the tree to do that, finding the root | ||||||
6890 | // reduction instruction in InLoopReductionImmediateChains. From there we find | ||||||
6891 | // the pattern of mul/ext and test the cost of the entire pattern vs the cost | ||||||
6892 | // of the components. If the reduction cost is lower then we return it for the | ||||||
6893 | // reduction instruction and 0 for the other instructions in the pattern. If | ||||||
6894 | // it is not we return an invalid cost specifying the orignal cost method | ||||||
6895 | // should be used. | ||||||
6896 | Instruction *RetI = I; | ||||||
6897 | if (match(RetI, m_ZExtOrSExt(m_Value()))) { | ||||||
6898 | if (!RetI->hasOneUser()) | ||||||
6899 | return None; | ||||||
6900 | RetI = RetI->user_back(); | ||||||
6901 | } | ||||||
6902 | if (match(RetI, m_Mul(m_Value(), m_Value())) && | ||||||
6903 | RetI->user_back()->getOpcode() == Instruction::Add) { | ||||||
6904 | if (!RetI->hasOneUser()) | ||||||
6905 | return None; | ||||||
6906 | RetI = RetI->user_back(); | ||||||
6907 | } | ||||||
6908 | |||||||
6909 | // Test if the found instruction is a reduction, and if not return an invalid | ||||||
6910 | // cost specifying the parent to use the original cost modelling. | ||||||
6911 | if (!InLoopReductionImmediateChains.count(RetI)) | ||||||
6912 | return None; | ||||||
6913 | |||||||
6914 | // Find the reduction this chain is a part of and calculate the basic cost of | ||||||
6915 | // the reduction on its own. | ||||||
6916 | Instruction *LastChain = InLoopReductionImmediateChains[RetI]; | ||||||
6917 | Instruction *ReductionPhi = LastChain; | ||||||
6918 | while (!isa<PHINode>(ReductionPhi)) | ||||||
6919 | ReductionPhi = InLoopReductionImmediateChains[ReductionPhi]; | ||||||
6920 | |||||||
6921 | const RecurrenceDescriptor &RdxDesc = | ||||||
6922 | Legal->getReductionVars().find(cast<PHINode>(ReductionPhi))->second; | ||||||
6923 | |||||||
6924 | InstructionCost BaseCost = TTI.getArithmeticReductionCost( | ||||||
6925 | RdxDesc.getOpcode(), VectorTy, RdxDesc.getFastMathFlags(), CostKind); | ||||||
6926 | |||||||
6927 | // For a call to the llvm.fmuladd intrinsic we need to add the cost of a | ||||||
6928 | // normal fmul instruction to the cost of the fadd reduction. | ||||||
6929 | if (RdxDesc.getRecurrenceKind() == RecurKind::FMulAdd) | ||||||
6930 | BaseCost += | ||||||
6931 | TTI.getArithmeticInstrCost(Instruction::FMul, VectorTy, CostKind); | ||||||
6932 | |||||||
6933 | // If we're using ordered reductions then we can just return the base cost | ||||||
6934 | // here, since getArithmeticReductionCost calculates the full ordered | ||||||
6935 | // reduction cost when FP reassociation is not allowed. | ||||||
6936 | if (useOrderedReductions(RdxDesc)) | ||||||
6937 | return BaseCost; | ||||||
6938 | |||||||
6939 | // Get the operand that was not the reduction chain and match it to one of the | ||||||
6940 | // patterns, returning the better cost if it is found. | ||||||
6941 | Instruction *RedOp = RetI->getOperand(1) == LastChain | ||||||
6942 | ? dyn_cast<Instruction>(RetI->getOperand(0)) | ||||||
6943 | : dyn_cast<Instruction>(RetI->getOperand(1)); | ||||||
6944 | |||||||
6945 | VectorTy = VectorType::get(I->getOperand(0)->getType(), VectorTy); | ||||||
6946 | |||||||
6947 | Instruction *Op0, *Op1; | ||||||
6948 | if (RedOp && | ||||||
6949 | match(RedOp, | ||||||
6950 | m_ZExtOrSExt(m_Mul(m_Instruction(Op0), m_Instruction(Op1)))) && | ||||||
6951 | match(Op0, m_ZExtOrSExt(m_Value())) && | ||||||
6952 | Op0->getOpcode() == Op1->getOpcode() && | ||||||
6953 | Op0->getOperand(0)->getType() == Op1->getOperand(0)->getType() && | ||||||
6954 | !TheLoop->isLoopInvariant(Op0) && !TheLoop->isLoopInvariant(Op1) && | ||||||
6955 | (Op0->getOpcode() == RedOp->getOpcode() || Op0 == Op1)) { | ||||||
6956 | |||||||
6957 | // Matched reduce(ext(mul(ext(A), ext(B))) | ||||||
6958 | // Note that the extend opcodes need to all match, or if A==B they will have | ||||||
6959 | // been converted to zext(mul(sext(A), sext(A))) as it is known positive, | ||||||
6960 | // which is equally fine. | ||||||
6961 | bool IsUnsigned = isa<ZExtInst>(Op0); | ||||||
6962 | auto *ExtType = VectorType::get(Op0->getOperand(0)->getType(), VectorTy); | ||||||
6963 | auto *MulType = VectorType::get(Op0->getType(), VectorTy); | ||||||
6964 | |||||||
6965 | InstructionCost ExtCost = | ||||||
6966 | TTI.getCastInstrCost(Op0->getOpcode(), MulType, ExtType, | ||||||
6967 | TTI::CastContextHint::None, CostKind, Op0); | ||||||
6968 | InstructionCost MulCost = | ||||||
6969 | TTI.getArithmeticInstrCost(Instruction::Mul, MulType, CostKind); | ||||||
6970 | InstructionCost Ext2Cost = | ||||||
6971 | TTI.getCastInstrCost(RedOp->getOpcode(), VectorTy, MulType, | ||||||
6972 | TTI::CastContextHint::None, CostKind, RedOp); | ||||||
6973 | |||||||
6974 | InstructionCost RedCost = TTI.getExtendedAddReductionCost( | ||||||
6975 | /*IsMLA=*/true, IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, | ||||||
6976 | CostKind); | ||||||
6977 | |||||||
6978 | if (RedCost.isValid() && | ||||||
6979 | RedCost < ExtCost * 2 + MulCost + Ext2Cost + BaseCost) | ||||||
6980 | return I == RetI ? RedCost : 0; | ||||||
6981 | } else if (RedOp && match(RedOp, m_ZExtOrSExt(m_Value())) && | ||||||
6982 | !TheLoop->isLoopInvariant(RedOp)) { | ||||||
6983 | // Matched reduce(ext(A)) | ||||||
6984 | bool IsUnsigned = isa<ZExtInst>(RedOp); | ||||||
6985 | auto *ExtType = VectorType::get(RedOp->getOperand(0)->getType(), VectorTy); | ||||||
6986 | InstructionCost RedCost = TTI.getExtendedAddReductionCost( | ||||||
6987 | /*IsMLA=*/false, IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, | ||||||
6988 | CostKind); | ||||||
6989 | |||||||
6990 | InstructionCost ExtCost = | ||||||
6991 | TTI.getCastInstrCost(RedOp->getOpcode(), VectorTy, ExtType, | ||||||
6992 | TTI::CastContextHint::None, CostKind, RedOp); | ||||||
6993 | if (RedCost.isValid() && RedCost < BaseCost + ExtCost) | ||||||
6994 | return I == RetI ? RedCost : 0; | ||||||
6995 | } else if (RedOp && | ||||||
6996 | match(RedOp, m_Mul(m_Instruction(Op0), m_Instruction(Op1)))) { | ||||||
6997 | if (match(Op0, m_ZExtOrSExt(m_Value())) && | ||||||
6998 | Op0->getOpcode() == Op1->getOpcode() && | ||||||
6999 | !TheLoop->isLoopInvariant(Op0) && !TheLoop->isLoopInvariant(Op1)) { | ||||||
7000 | bool IsUnsigned = isa<ZExtInst>(Op0); | ||||||
7001 | Type *Op0Ty = Op0->getOperand(0)->getType(); | ||||||
7002 | Type *Op1Ty = Op1->getOperand(0)->getType(); | ||||||
7003 | Type *LargestOpTy = | ||||||
7004 | Op0Ty->getIntegerBitWidth() < Op1Ty->getIntegerBitWidth() ? Op1Ty | ||||||
7005 | : Op0Ty; | ||||||
7006 | auto *ExtType = VectorType::get(LargestOpTy, VectorTy); | ||||||
7007 | |||||||
7008 | // Matched reduce(mul(ext(A), ext(B))), where the two ext may be of | ||||||
7009 | // different sizes. We take the largest type as the ext to reduce, and add | ||||||
7010 | // the remaining cost as, for example reduce(mul(ext(ext(A)), ext(B))). | ||||||
7011 | InstructionCost ExtCost0 = TTI.getCastInstrCost( | ||||||
7012 | Op0->getOpcode(), VectorTy, VectorType::get(Op0Ty, VectorTy), | ||||||
7013 | TTI::CastContextHint::None, CostKind, Op0); | ||||||
7014 | InstructionCost ExtCost1 = TTI.getCastInstrCost( | ||||||
7015 | Op1->getOpcode(), VectorTy, VectorType::get(Op1Ty, VectorTy), | ||||||
7016 | TTI::CastContextHint::None, CostKind, Op1); | ||||||
7017 | InstructionCost MulCost = | ||||||
7018 | TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); | ||||||
7019 | |||||||
7020 | InstructionCost RedCost = TTI.getExtendedAddReductionCost( | ||||||
7021 | /*IsMLA=*/true, IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, | ||||||
7022 | CostKind); | ||||||
7023 | InstructionCost ExtraExtCost = 0; | ||||||
7024 | if (Op0Ty != LargestOpTy || Op1Ty != LargestOpTy) { | ||||||
7025 | Instruction *ExtraExtOp = (Op0Ty != LargestOpTy) ? Op0 : Op1; | ||||||
7026 | ExtraExtCost = TTI.getCastInstrCost( | ||||||
7027 | ExtraExtOp->getOpcode(), ExtType, | ||||||
7028 | VectorType::get(ExtraExtOp->getOperand(0)->getType(), VectorTy), | ||||||
7029 | TTI::CastContextHint::None, CostKind, ExtraExtOp); | ||||||
7030 | } | ||||||
7031 | |||||||
7032 | if (RedCost.isValid() && | ||||||
7033 | (RedCost + ExtraExtCost) < (ExtCost0 + ExtCost1 + MulCost + BaseCost)) | ||||||
7034 | return I == RetI ? RedCost : 0; | ||||||
7035 | } else if (!match(I, m_ZExtOrSExt(m_Value()))) { | ||||||
7036 | // Matched reduce(mul()) | ||||||
7037 | InstructionCost MulCost = | ||||||
7038 | TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); | ||||||
7039 | |||||||
7040 | InstructionCost RedCost = TTI.getExtendedAddReductionCost( | ||||||
7041 | /*IsMLA=*/true, true, RdxDesc.getRecurrenceType(), VectorTy, | ||||||
7042 | CostKind); | ||||||
7043 | |||||||
7044 | if (RedCost.isValid() && RedCost < MulCost + BaseCost) | ||||||
7045 | return I == RetI ? RedCost : 0; | ||||||
7046 | } | ||||||
7047 | } | ||||||
7048 | |||||||
7049 | return I == RetI ? Optional<InstructionCost>(BaseCost) : None; | ||||||
7050 | } | ||||||
7051 | |||||||
7052 | InstructionCost | ||||||
7053 | LoopVectorizationCostModel::getMemoryInstructionCost(Instruction *I, | ||||||
7054 | ElementCount VF) { | ||||||
7055 | // Calculate scalar cost only. Vectorization cost should be ready at this | ||||||
7056 | // moment. | ||||||
7057 | if (VF.isScalar()) { | ||||||
7058 | Type *ValTy = getLoadStoreType(I); | ||||||
7059 | const Align Alignment = getLoadStoreAlignment(I); | ||||||
7060 | unsigned AS = getLoadStoreAddressSpace(I); | ||||||
7061 | |||||||
7062 | return TTI.getAddressComputationCost(ValTy) + | ||||||
7063 | TTI.getMemoryOpCost(I->getOpcode(), ValTy, Alignment, AS, | ||||||
7064 | TTI::TCK_RecipThroughput, I); | ||||||
7065 | } | ||||||
7066 | return getWideningCost(I, VF); | ||||||
7067 | } | ||||||
7068 | |||||||
7069 | LoopVectorizationCostModel::VectorizationCostTy | ||||||
7070 | LoopVectorizationCostModel::getInstructionCost(Instruction *I, | ||||||
7071 | ElementCount VF) { | ||||||
7072 | // If we know that this instruction will remain uniform, check the cost of | ||||||
7073 | // the scalar version. | ||||||
7074 | if (isUniformAfterVectorization(I, VF)) | ||||||
7075 | VF = ElementCount::getFixed(1); | ||||||
7076 | |||||||
7077 | if (VF.isVector() && isProfitableToScalarize(I, VF)) | ||||||
7078 | return VectorizationCostTy(InstsToScalarize[VF][I], false); | ||||||
7079 | |||||||
7080 | // Forced scalars do not have any scalarization overhead. | ||||||
7081 | auto ForcedScalar = ForcedScalars.find(VF); | ||||||
7082 | if (VF.isVector() && ForcedScalar != ForcedScalars.end()) { | ||||||
7083 | auto InstSet = ForcedScalar->second; | ||||||
7084 | if (InstSet.count(I)) | ||||||
7085 | return VectorizationCostTy( | ||||||
7086 | (getInstructionCost(I, ElementCount::getFixed(1)).first * | ||||||
7087 | VF.getKnownMinValue()), | ||||||
7088 | false); | ||||||
7089 | } | ||||||
7090 | |||||||
7091 | Type *VectorTy; | ||||||
7092 | InstructionCost C = getInstructionCost(I, VF, VectorTy); | ||||||
7093 | |||||||
7094 | bool TypeNotScalarized = false; | ||||||
7095 | if (VF.isVector() && VectorTy->isVectorTy()) { | ||||||
7096 | unsigned NumParts = TTI.getNumberOfParts(VectorTy); | ||||||
7097 | if (NumParts) | ||||||
7098 | TypeNotScalarized = NumParts < VF.getKnownMinValue(); | ||||||
7099 | else | ||||||
7100 | C = InstructionCost::getInvalid(); | ||||||
7101 | } | ||||||
7102 | return VectorizationCostTy(C, TypeNotScalarized); | ||||||
7103 | } | ||||||
7104 | |||||||
7105 | InstructionCost | ||||||
7106 | LoopVectorizationCostModel::getScalarizationOverhead(Instruction *I, | ||||||
7107 | ElementCount VF) const { | ||||||
7108 | |||||||
7109 | // There is no mechanism yet to create a scalable scalarization loop, | ||||||
7110 | // so this is currently Invalid. | ||||||
7111 | if (VF.isScalable()) | ||||||
7112 | return InstructionCost::getInvalid(); | ||||||
7113 | |||||||
7114 | if (VF.isScalar()) | ||||||
7115 | return 0; | ||||||
7116 | |||||||
7117 | InstructionCost Cost = 0; | ||||||
7118 | Type *RetTy = ToVectorTy(I->getType(), VF); | ||||||
7119 | if (!RetTy->isVoidTy() && | ||||||
7120 | (!isa<LoadInst>(I) || !TTI.supportsEfficientVectorElementLoadStore())) | ||||||
7121 | Cost += TTI.getScalarizationOverhead( | ||||||
7122 | cast<VectorType>(RetTy), APInt::getAllOnes(VF.getKnownMinValue()), true, | ||||||
7123 | false); | ||||||
7124 | |||||||
7125 | // Some targets keep addresses scalar. | ||||||
7126 | if (isa<LoadInst>(I) && !TTI.prefersVectorizedAddressing()) | ||||||
7127 | return Cost; | ||||||
7128 | |||||||
7129 | // Some targets support efficient element stores. | ||||||
7130 | if (isa<StoreInst>(I) && TTI.supportsEfficientVectorElementLoadStore()) | ||||||
7131 | return Cost; | ||||||
7132 | |||||||
7133 | // Collect operands to consider. | ||||||
7134 | CallInst *CI = dyn_cast<CallInst>(I); | ||||||
7135 | Instruction::op_range Ops = CI ? CI->args() : I->operands(); | ||||||
7136 | |||||||
7137 | // Skip operands that do not require extraction/scalarization and do not incur | ||||||
7138 | // any overhead. | ||||||
7139 | SmallVector<Type *> Tys; | ||||||
7140 | for (auto *V : filterExtractingOperands(Ops, VF)) | ||||||
7141 | Tys.push_back(MaybeVectorizeType(V->getType(), VF)); | ||||||
7142 | return Cost + TTI.getOperandsScalarizationOverhead( | ||||||
7143 | filterExtractingOperands(Ops, VF), Tys); | ||||||
7144 | } | ||||||
7145 | |||||||
7146 | void LoopVectorizationCostModel::setCostBasedWideningDecision(ElementCount VF) { | ||||||
7147 | if (VF.isScalar()) | ||||||
7148 | return; | ||||||
7149 | NumPredStores = 0; | ||||||
7150 | for (BasicBlock *BB : TheLoop->blocks()) { | ||||||
7151 | // For each instruction in the old loop. | ||||||
7152 | for (Instruction &I : *BB) { | ||||||
7153 | Value *Ptr = getLoadStorePointerOperand(&I); | ||||||
7154 | if (!Ptr) | ||||||
7155 | continue; | ||||||
7156 | |||||||
7157 | // TODO: We should generate better code and update the cost model for | ||||||
7158 | // predicated uniform stores. Today they are treated as any other | ||||||
7159 | // predicated store (see added test cases in | ||||||
7160 | // invariant-store-vectorization.ll). | ||||||
7161 | if (isa<StoreInst>(&I) && isScalarWithPredication(&I, VF)) | ||||||
7162 | NumPredStores++; | ||||||
7163 | |||||||
7164 | if (Legal->isUniformMemOp(I)) { | ||||||
7165 | // TODO: Avoid replicating loads and stores instead of | ||||||
7166 | // relying on instcombine to remove them. | ||||||
7167 | // Load: Scalar load + broadcast | ||||||
7168 | // Store: Scalar store + isLoopInvariantStoreValue ? 0 : extract | ||||||
7169 | InstructionCost Cost; | ||||||
7170 | if (isa<StoreInst>(&I) && VF.isScalable() && | ||||||
7171 | isLegalGatherOrScatter(&I, VF)) { | ||||||
7172 | Cost = getGatherScatterCost(&I, VF); | ||||||
7173 | setWideningDecision(&I, VF, CM_GatherScatter, Cost); | ||||||
7174 | } else { | ||||||
7175 | assert((isa<LoadInst>(&I) || !VF.isScalable()) &&(static_cast <bool> ((isa<LoadInst>(&I) || !VF .isScalable()) && "Cannot yet scalarize uniform stores" ) ? void (0) : __assert_fail ("(isa<LoadInst>(&I) || !VF.isScalable()) && \"Cannot yet scalarize uniform stores\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7176, __extension__ __PRETTY_FUNCTION__)) | ||||||
7176 | "Cannot yet scalarize uniform stores")(static_cast <bool> ((isa<LoadInst>(&I) || !VF .isScalable()) && "Cannot yet scalarize uniform stores" ) ? void (0) : __assert_fail ("(isa<LoadInst>(&I) || !VF.isScalable()) && \"Cannot yet scalarize uniform stores\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 7176, __extension__ __PRETTY_FUNCTION__)); | ||||||
7177 | Cost = getUniformMemOpCost(&I, VF); | ||||||
7178 | setWideningDecision(&I, VF, CM_Scalarize, Cost); | ||||||
7179 | } | ||||||
7180 | continue; | ||||||
7181 | } | ||||||
7182 | |||||||
7183 | // We assume that widening is the best solution when possible. | ||||||
7184 | if (memoryInstructionCanBeWidened(&I, VF)) { | ||||||
7185 | InstructionCost Cost = getConsecutiveMemOpCost(&I, VF); | ||||||
7186 | int ConsecutiveStride = Legal->isConsecutivePtr( | ||||||
7187 | getLoadStoreType(&I), getLoadStorePointerOperand(&I)); | ||||||
7188 | 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", 7189, __extension__ __PRETTY_FUNCTION__)) | ||||||
7189 | "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", 7189, __extension__ __PRETTY_FUNCTION__)); | ||||||
7190 | InstWidening Decision = | ||||||
7191 | ConsecutiveStride == 1 ? CM_Widen : CM_Widen_Reverse; | ||||||
7192 | setWideningDecision(&I, VF, Decision, Cost); | ||||||
7193 | continue; | ||||||
7194 | } | ||||||
7195 | |||||||
7196 | // Choose between Interleaving, Gather/Scatter or Scalarization. | ||||||
7197 | InstructionCost InterleaveCost = InstructionCost::getInvalid(); | ||||||
7198 | unsigned NumAccesses = 1; | ||||||
7199 | if (isAccessInterleaved(&I)) { | ||||||
7200 | auto Group = getInterleavedAccessGroup(&I); | ||||||
7201 | 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", 7201, __extension__ __PRETTY_FUNCTION__)); | ||||||
7202 | |||||||
7203 | // Make one decision for the whole group. | ||||||
7204 | if (getWideningDecision(&I, VF) != CM_Unknown) | ||||||
7205 | continue; | ||||||
7206 | |||||||
7207 | NumAccesses = Group->getNumMembers(); | ||||||
7208 | if (interleavedAccessCanBeWidened(&I, VF)) | ||||||
7209 | InterleaveCost = getInterleaveGroupCost(&I, VF); | ||||||
7210 | } | ||||||
7211 | |||||||
7212 | InstructionCost GatherScatterCost = | ||||||
7213 | isLegalGatherOrScatter(&I, VF) | ||||||
7214 | ? getGatherScatterCost(&I, VF) * NumAccesses | ||||||
7215 | : InstructionCost::getInvalid(); | ||||||
7216 | |||||||
7217 | InstructionCost ScalarizationCost = | ||||||
7218 | getMemInstScalarizationCost(&I, VF) * NumAccesses; | ||||||
7219 | |||||||
7220 | // Choose better solution for the current VF, | ||||||
7221 | // write down this decision and use it during vectorization. | ||||||
7222 | InstructionCost Cost; | ||||||
7223 | InstWidening Decision; | ||||||
7224 | if (InterleaveCost <= GatherScatterCost && | ||||||
7225 | InterleaveCost < ScalarizationCost) { | ||||||
7226 | Decision = CM_Interleave; | ||||||
7227 | Cost = InterleaveCost; | ||||||
7228 | } else if (GatherScatterCost < ScalarizationCost) { | ||||||
7229 | Decision = CM_GatherScatter; | ||||||
7230 | Cost = GatherScatterCost; | ||||||
7231 | } else { | ||||||
7232 | Decision = CM_Scalarize; | ||||||
7233 | Cost = ScalarizationCost; | ||||||
7234 | } | ||||||
7235 | // If the instructions belongs to an interleave group, the whole group | ||||||
7236 | // receives the same decision. The whole group receives the cost, but | ||||||
7237 | // the cost will actually be assigned to one instruction. | ||||||
7238 | if (auto Group = getInterleavedAccessGroup(&I)) | ||||||
7239 | setWideningDecision(Group, VF, Decision, Cost); | ||||||
7240 | else | ||||||
7241 | setWideningDecision(&I, VF, Decision, Cost); | ||||||
7242 | } | ||||||
7243 | } | ||||||
7244 | |||||||
7245 | // Make sure that any load of address and any other address computation | ||||||
7246 | // remains scalar unless there is gather/scatter support. This avoids | ||||||
7247 | // inevitable extracts into address registers, and also has the benefit of | ||||||
7248 | // activating LSR more, since that pass can't optimize vectorized | ||||||
7249 | // addresses. | ||||||
7250 | if (TTI.prefersVectorizedAddressing()) | ||||||
7251 | return; | ||||||
7252 | |||||||
7253 | // Start with all scalar pointer uses. | ||||||
7254 | SmallPtrSet<Instruction *, 8> AddrDefs; | ||||||
7255 | for (BasicBlock *BB : TheLoop->blocks()) | ||||||
7256 | for (Instruction &I : *BB) { | ||||||
7257 | Instruction *PtrDef = | ||||||
7258 | dyn_cast_or_null<Instruction>(getLoadStorePointerOperand(&I)); | ||||||
7259 | if (PtrDef && TheLoop->contains(PtrDef) && | ||||||
7260 | getWideningDecision(&I, VF) != CM_GatherScatter) | ||||||
7261 | AddrDefs.insert(PtrDef); | ||||||
7262 | } | ||||||
7263 | |||||||
7264 | // Add all instructions used to generate the addresses. | ||||||
7265 | SmallVector<Instruction *, 4> Worklist; | ||||||
7266 | append_range(Worklist, AddrDefs); | ||||||
7267 | while (!Worklist.empty()) { | ||||||
7268 | Instruction *I = Worklist.pop_back_val(); | ||||||
7269 | for (auto &Op : I->operands()) | ||||||
7270 | if (auto *InstOp = dyn_cast<Instruction>(Op)) | ||||||
7271 | if ((InstOp->getParent() == I->getParent()) && !isa<PHINode>(InstOp) && | ||||||
7272 | AddrDefs.insert(InstOp).second) | ||||||
7273 | Worklist.push_back(InstOp); | ||||||
7274 | } | ||||||
7275 | |||||||
7276 | for (auto *I : AddrDefs) { | ||||||
7277 | if (isa<LoadInst>(I)) { | ||||||
7278 | // Setting the desired widening decision should ideally be handled in | ||||||
7279 | // by cost functions, but since this involves the task of finding out | ||||||
7280 | // if the loaded register is involved in an address computation, it is | ||||||
7281 | // instead changed here when we know this is the case. | ||||||
7282 | InstWidening Decision = getWideningDecision(I, VF); | ||||||
7283 | if (Decision == CM_Widen || Decision == CM_Widen_Reverse) | ||||||
7284 | // Scalarize a widened load of address. | ||||||
7285 | setWideningDecision( | ||||||
7286 | I, VF, CM_Scalarize, | ||||||
7287 | (VF.getKnownMinValue() * | ||||||
7288 | getMemoryInstructionCost(I, ElementCount::getFixed(1)))); | ||||||
7289 | else if (auto Group = getInterleavedAccessGroup(I)) { | ||||||
7290 | // Scalarize an interleave group of address loads. | ||||||
7291 | for (unsigned I = 0; I < Group->getFactor(); ++I) { | ||||||
7292 | if (Instruction *Member = Group->getMember(I)) | ||||||
7293 | setWideningDecision( | ||||||
7294 | Member, VF, CM_Scalarize, | ||||||
7295 | (VF.getKnownMinValue() * | ||||||
7296 | getMemoryInstructionCost(Member, ElementCount::getFixed(1)))); | ||||||
7297 | } | ||||||
7298 | } | ||||||
7299 | } else | ||||||
7300 | // Make sure I gets scalarized and a cost estimate without | ||||||
7301 | // scalarization overhead. | ||||||
7302 | ForcedScalars[VF].insert(I); | ||||||
7303 | } | ||||||
7304 | } | ||||||
7305 | |||||||
7306 | InstructionCost | ||||||
7307 | LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF, | ||||||
7308 | Type *&VectorTy) { | ||||||
7309 | Type *RetTy = I->getType(); | ||||||
7310 | if (canTruncateToMinimalBitwidth(I, VF)) | ||||||
7311 | RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]); | ||||||
7312 | auto SE = PSE.getSE(); | ||||||
7313 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||||
7314 | |||||||
7315 | auto hasSingleCopyAfterVectorization = [this](Instruction *I, | ||||||
7316 | ElementCount VF) -> bool { | ||||||
7317 | if (VF.isScalar()) | ||||||
7318 | return true; | ||||||
7319 | |||||||
7320 | auto Scalarized = InstsToScalarize.find(VF); | ||||||
7321 | 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", 7322, __extension__ __PRETTY_FUNCTION__)) | ||||||
7322 | "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", 7322, __extension__ __PRETTY_FUNCTION__)); | ||||||
7323 | return !Scalarized->second.count(I) && | ||||||
7324 | llvm::all_of(I->users(), [&](User *U) { | ||||||
7325 | auto *UI = cast<Instruction>(U); | ||||||
7326 | return !Scalarized->second.count(UI); | ||||||
7327 | }); | ||||||
7328 | }; | ||||||
7329 | (void) hasSingleCopyAfterVectorization; | ||||||
7330 | |||||||
7331 | if (isScalarAfterVectorization(I, VF)) { | ||||||
7332 | // With the exception of GEPs and PHIs, after scalarization there should | ||||||
7333 | // only be one copy of the instruction generated in the loop. This is | ||||||
7334 | // because the VF is either 1, or any instructions that need scalarizing | ||||||
7335 | // have already been dealt with by the the time we get here. As a result, | ||||||
7336 | // it means we don't have to multiply the instruction cost by VF. | ||||||
7337 | 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", 7341, __extension__ __PRETTY_FUNCTION__)) | ||||||
7338 | 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", 7341, __extension__ __PRETTY_FUNCTION__)) | ||||||
7339 | (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", 7341, __extension__ __PRETTY_FUNCTION__)) | ||||||
7340 | 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", 7341, __extension__ __PRETTY_FUNCTION__)) | ||||||
7341 | 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", 7341, __extension__ __PRETTY_FUNCTION__)); | ||||||
7342 | VectorTy = RetTy; | ||||||
7343 | } else | ||||||
7344 | VectorTy = ToVectorTy(RetTy, VF); | ||||||
7345 | |||||||
7346 | // TODO: We need to estimate the cost of intrinsic calls. | ||||||
7347 | switch (I->getOpcode()) { | ||||||
7348 | case Instruction::GetElementPtr: | ||||||
7349 | // We mark this instruction as zero-cost because the cost of GEPs in | ||||||
7350 | // vectorized code depends on whether the corresponding memory instruction | ||||||
7351 | // is scalarized or not. Therefore, we handle GEPs with the memory | ||||||
7352 | // instruction cost. | ||||||
7353 | return 0; | ||||||
7354 | case Instruction::Br: { | ||||||
7355 | // In cases of scalarized and predicated instructions, there will be VF | ||||||
7356 | // predicated blocks in the vectorized loop. Each branch around these | ||||||
7357 | // blocks requires also an extract of its vector compare i1 element. | ||||||
7358 | bool ScalarPredicatedBB = false; | ||||||
7359 | BranchInst *BI = cast<BranchInst>(I); | ||||||
7360 | if (VF.isVector() && BI->isConditional() && | ||||||
7361 | (PredicatedBBsAfterVectorization.count(BI->getSuccessor(0)) || | ||||||
7362 | PredicatedBBsAfterVectorization.count(BI->getSuccessor(1)))) | ||||||
7363 | ScalarPredicatedBB = true; | ||||||
7364 | |||||||
7365 | if (ScalarPredicatedBB) { | ||||||
7366 | // Not possible to scalarize scalable vector with predicated instructions. | ||||||
7367 | if (VF.isScalable()) | ||||||
7368 | return InstructionCost::getInvalid(); | ||||||
7369 | // Return cost for branches around scalarized and predicated blocks. | ||||||
7370 | auto *Vec_i1Ty = | ||||||
7371 | VectorType::get(IntegerType::getInt1Ty(RetTy->getContext()), VF); | ||||||
7372 | return ( | ||||||
7373 | TTI.getScalarizationOverhead( | ||||||
7374 | Vec_i1Ty, APInt::getAllOnes(VF.getFixedValue()), false, true) + | ||||||
7375 | (TTI.getCFInstrCost(Instruction::Br, CostKind) * VF.getFixedValue())); | ||||||
7376 | } else if (I->getParent() == TheLoop->getLoopLatch() || VF.isScalar()) | ||||||
7377 | // The back-edge branch will remain, as will all scalar branches. | ||||||
7378 | return TTI.getCFInstrCost(Instruction::Br, CostKind); | ||||||
7379 | else | ||||||
7380 | // This branch will be eliminated by if-conversion. | ||||||
7381 | return 0; | ||||||
7382 | // Note: We currently assume zero cost for an unconditional branch inside | ||||||
7383 | // a predicated block since it will become a fall-through, although we | ||||||
7384 | // may decide in the future to call TTI for all branches. | ||||||
7385 | } | ||||||
7386 | case Instruction::PHI: { | ||||||
7387 | auto *Phi = cast<PHINode>(I); | ||||||
7388 | |||||||
7389 | // First-order recurrences are replaced by vector shuffles inside the loop. | ||||||
7390 | // NOTE: Don't use ToVectorTy as SK_ExtractSubvector expects a vector type. | ||||||
7391 | if (VF.isVector() && Legal->isFirstOrderRecurrence(Phi)) | ||||||
7392 | return TTI.getShuffleCost( | ||||||
7393 | TargetTransformInfo::SK_ExtractSubvector, cast<VectorType>(VectorTy), | ||||||
7394 | None, VF.getKnownMinValue() - 1, FixedVectorType::get(RetTy, 1)); | ||||||
7395 | |||||||
7396 | // Phi nodes in non-header blocks (not inductions, reductions, etc.) are | ||||||
7397 | // converted into select instructions. We require N - 1 selects per phi | ||||||
7398 | // node, where N is the number of incoming values. | ||||||
7399 | if (VF.isVector() && Phi->getParent() != TheLoop->getHeader()) | ||||||
7400 | return (Phi->getNumIncomingValues() - 1) * | ||||||
7401 | TTI.getCmpSelInstrCost( | ||||||
7402 | Instruction::Select, ToVectorTy(Phi->getType(), VF), | ||||||
7403 | ToVectorTy(Type::getInt1Ty(Phi->getContext()), VF), | ||||||
7404 | CmpInst::BAD_ICMP_PREDICATE, CostKind); | ||||||
7405 | |||||||
7406 | return TTI.getCFInstrCost(Instruction::PHI, CostKind); | ||||||
7407 | } | ||||||
7408 | case Instruction::UDiv: | ||||||
7409 | case Instruction::SDiv: | ||||||
7410 | case Instruction::URem: | ||||||
7411 | case Instruction::SRem: | ||||||
7412 | // If we have a predicated instruction, it may not be executed for each | ||||||
7413 | // vector lane. Get the scalarization cost and scale this amount by the | ||||||
7414 | // probability of executing the predicated block. If the instruction is not | ||||||
7415 | // predicated, we fall through to the next case. | ||||||
7416 | if (VF.isVector() && isScalarWithPredication(I, VF)) { | ||||||
7417 | InstructionCost Cost = 0; | ||||||
7418 | |||||||
7419 | // These instructions have a non-void type, so account for the phi nodes | ||||||
7420 | // that we will create. This cost is likely to be zero. The phi node | ||||||
7421 | // cost, if any, should be scaled by the block probability because it | ||||||
7422 | // models a copy at the end of each predicated block. | ||||||
7423 | Cost += VF.getKnownMinValue() * | ||||||
7424 | TTI.getCFInstrCost(Instruction::PHI, CostKind); | ||||||
7425 | |||||||
7426 | // The cost of the non-predicated instruction. | ||||||
7427 | Cost += VF.getKnownMinValue() * | ||||||
7428 | TTI.getArithmeticInstrCost(I->getOpcode(), RetTy, CostKind); | ||||||
7429 | |||||||
7430 | // The cost of insertelement and extractelement instructions needed for | ||||||
7431 | // scalarization. | ||||||
7432 | Cost += getScalarizationOverhead(I, VF); | ||||||
7433 | |||||||
7434 | // Scale the cost by the probability of executing the predicated blocks. | ||||||
7435 | // This assumes the predicated block for each vector lane is equally | ||||||
7436 | // likely. | ||||||
7437 | return Cost / getReciprocalPredBlockProb(); | ||||||
7438 | } | ||||||
7439 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | ||||||
7440 | case Instruction::Add: | ||||||
7441 | case Instruction::FAdd: | ||||||
7442 | case Instruction::Sub: | ||||||
7443 | case Instruction::FSub: | ||||||
7444 | case Instruction::Mul: | ||||||
7445 | case Instruction::FMul: | ||||||
7446 | case Instruction::FDiv: | ||||||
7447 | case Instruction::FRem: | ||||||
7448 | case Instruction::Shl: | ||||||
7449 | case Instruction::LShr: | ||||||
7450 | case Instruction::AShr: | ||||||
7451 | case Instruction::And: | ||||||
7452 | case Instruction::Or: | ||||||
7453 | case Instruction::Xor: { | ||||||
7454 | // Since we will replace the stride by 1 the multiplication should go away. | ||||||
7455 | if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal)) | ||||||
7456 | return 0; | ||||||
7457 | |||||||
7458 | // Detect reduction patterns | ||||||
7459 | if (auto RedCost = getReductionPatternCost(I, VF, VectorTy, CostKind)) | ||||||
7460 | return *RedCost; | ||||||
7461 | |||||||
7462 | // Certain instructions can be cheaper to vectorize if they have a constant | ||||||
7463 | // second vector operand. One example of this are shifts on x86. | ||||||
7464 | Value *Op2 = I->getOperand(1); | ||||||
7465 | TargetTransformInfo::OperandValueProperties Op2VP; | ||||||
7466 | TargetTransformInfo::OperandValueKind Op2VK = | ||||||
7467 | TTI.getOperandInfo(Op2, Op2VP); | ||||||
7468 | if (Op2VK == TargetTransformInfo::OK_AnyValue && Legal->isUniform(Op2)) | ||||||
7469 | Op2VK = TargetTransformInfo::OK_UniformValue; | ||||||
7470 | |||||||
7471 | SmallVector<const Value *, 4> Operands(I->operand_values()); | ||||||
7472 | return TTI.getArithmeticInstrCost( | ||||||
7473 | I->getOpcode(), VectorTy, CostKind, TargetTransformInfo::OK_AnyValue, | ||||||
7474 | Op2VK, TargetTransformInfo::OP_None, Op2VP, Operands, I); | ||||||
7475 | } | ||||||
7476 | case Instruction::FNeg: { | ||||||
7477 | return TTI.getArithmeticInstrCost( | ||||||
7478 | I->getOpcode(), VectorTy, CostKind, TargetTransformInfo::OK_AnyValue, | ||||||
7479 | TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None, | ||||||
7480 | TargetTransformInfo::OP_None, I->getOperand(0), I); | ||||||
7481 | } | ||||||
7482 | case Instruction::Select: { | ||||||
7483 | SelectInst *SI = cast<SelectInst>(I); | ||||||
7484 | const SCEV *CondSCEV = SE->getSCEV(SI->getCondition()); | ||||||
7485 | bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop)); | ||||||
7486 | |||||||
7487 | const Value *Op0, *Op1; | ||||||
7488 | using namespace llvm::PatternMatch; | ||||||
7489 | if (!ScalarCond && (match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1))) || | ||||||
7490 | match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1))))) { | ||||||
7491 | // select x, y, false --> x & y | ||||||
7492 | // select x, true, y --> x | y | ||||||
7493 | TTI::OperandValueProperties Op1VP = TTI::OP_None; | ||||||
7494 | TTI::OperandValueProperties Op2VP = TTI::OP_None; | ||||||
7495 | TTI::OperandValueKind Op1VK = TTI::getOperandInfo(Op0, Op1VP); | ||||||
7496 | TTI::OperandValueKind Op2VK = TTI::getOperandInfo(Op1, Op2VP); | ||||||
7497 | 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", 7498, __extension__ __PRETTY_FUNCTION__)) | ||||||
7498 | 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", 7498, __extension__ __PRETTY_FUNCTION__)); | ||||||
7499 | |||||||
7500 | SmallVector<const Value *, 2> Operands{Op0, Op1}; | ||||||
7501 | return TTI.getArithmeticInstrCost( | ||||||
7502 | match(I, m_LogicalOr()) ? Instruction::Or : Instruction::And, VectorTy, | ||||||
7503 | CostKind, Op1VK, Op2VK, Op1VP, Op2VP, Operands, I); | ||||||
7504 | } | ||||||
7505 | |||||||
7506 | Type *CondTy = SI->getCondition()->getType(); | ||||||
7507 | if (!ScalarCond) | ||||||
7508 | CondTy = VectorType::get(CondTy, VF); | ||||||
7509 | |||||||
7510 | CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE; | ||||||
7511 | if (auto *Cmp = dyn_cast<CmpInst>(SI->getCondition())) | ||||||
7512 | Pred = Cmp->getPredicate(); | ||||||
7513 | return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy, Pred, | ||||||
7514 | CostKind, I); | ||||||
7515 | } | ||||||
7516 | case Instruction::ICmp: | ||||||
7517 | case Instruction::FCmp: { | ||||||
7518 | Type *ValTy = I->getOperand(0)->getType(); | ||||||
7519 | Instruction *Op0AsInstruction = dyn_cast<Instruction>(I->getOperand(0)); | ||||||
7520 | if (canTruncateToMinimalBitwidth(Op0AsInstruction, VF)) | ||||||
7521 | ValTy = IntegerType::get(ValTy->getContext(), MinBWs[Op0AsInstruction]); | ||||||
7522 | VectorTy = ToVectorTy(ValTy, VF); | ||||||
7523 | return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, nullptr, | ||||||
7524 | cast<CmpInst>(I)->getPredicate(), CostKind, | ||||||
7525 | I); | ||||||
7526 | } | ||||||
7527 | case Instruction::Store: | ||||||
7528 | case Instruction::Load: { | ||||||
7529 | ElementCount Width = VF; | ||||||
7530 | if (Width.isVector()) { | ||||||
7531 | InstWidening Decision = getWideningDecision(I, Width); | ||||||
7532 | 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", 7533, __extension__ __PRETTY_FUNCTION__)) | ||||||
7533 | "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", 7533, __extension__ __PRETTY_FUNCTION__)); | ||||||
7534 | if (Decision == CM_Scalarize) | ||||||
7535 | Width = ElementCount::getFixed(1); | ||||||
7536 | } | ||||||
7537 | VectorTy = ToVectorTy(getLoadStoreType(I), Width); | ||||||
7538 | return getMemoryInstructionCost(I, VF); | ||||||
7539 | } | ||||||
7540 | case Instruction::BitCast: | ||||||
7541 | if (I->getType()->isPointerTy()) | ||||||
7542 | return 0; | ||||||
7543 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | ||||||
7544 | case Instruction::ZExt: | ||||||
7545 | case Instruction::SExt: | ||||||
7546 | case Instruction::FPToUI: | ||||||
7547 | case Instruction::FPToSI: | ||||||
7548 | case Instruction::FPExt: | ||||||
7549 | case Instruction::PtrToInt: | ||||||
7550 | case Instruction::IntToPtr: | ||||||
7551 | case Instruction::SIToFP: | ||||||
7552 | case Instruction::UIToFP: | ||||||
7553 | case Instruction::Trunc: | ||||||
7554 | case Instruction::FPTrunc: { | ||||||
7555 | // Computes the CastContextHint from a Load/Store instruction. | ||||||
7556 | auto ComputeCCH = [&](Instruction *I) -> TTI::CastContextHint { | ||||||
7557 | 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", 7558, __extension__ __PRETTY_FUNCTION__)) | ||||||
7558 | "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", 7558, __extension__ __PRETTY_FUNCTION__)); | ||||||
7559 | |||||||
7560 | if (VF.isScalar() || !TheLoop->contains(I)) | ||||||
7561 | return TTI::CastContextHint::Normal; | ||||||
7562 | |||||||
7563 | switch (getWideningDecision(I, VF)) { | ||||||
7564 | case LoopVectorizationCostModel::CM_GatherScatter: | ||||||
7565 | return TTI::CastContextHint::GatherScatter; | ||||||
7566 | case LoopVectorizationCostModel::CM_Interleave: | ||||||
7567 | return TTI::CastContextHint::Interleave; | ||||||
7568 | case LoopVectorizationCostModel::CM_Scalarize: | ||||||
7569 | case LoopVectorizationCostModel::CM_Widen: | ||||||
7570 | return Legal->isMaskRequired(I) ? TTI::CastContextHint::Masked | ||||||
7571 | : TTI::CastContextHint::Normal; | ||||||
7572 | case LoopVectorizationCostModel::CM_Widen_Reverse: | ||||||
7573 | return TTI::CastContextHint::Reversed; | ||||||
7574 | case LoopVectorizationCostModel::CM_Unknown: | ||||||
7575 | 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", 7575); | ||||||
7576 | } | ||||||
7577 | |||||||
7578 | llvm_unreachable("Unhandled case!")::llvm::llvm_unreachable_internal("Unhandled case!", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7578); | ||||||
7579 | }; | ||||||
7580 | |||||||
7581 | unsigned Opcode = I->getOpcode(); | ||||||
7582 | TTI::CastContextHint CCH = TTI::CastContextHint::None; | ||||||
7583 | // For Trunc, the context is the only user, which must be a StoreInst. | ||||||
7584 | if (Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) { | ||||||
7585 | if (I->hasOneUse()) | ||||||
7586 | if (StoreInst *Store = dyn_cast<StoreInst>(*I->user_begin())) | ||||||
7587 | CCH = ComputeCCH(Store); | ||||||
7588 | } | ||||||
7589 | // For Z/Sext, the context is the operand, which must be a LoadInst. | ||||||
7590 | else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt || | ||||||
7591 | Opcode == Instruction::FPExt) { | ||||||
7592 | if (LoadInst *Load = dyn_cast<LoadInst>(I->getOperand(0))) | ||||||
7593 | CCH = ComputeCCH(Load); | ||||||
7594 | } | ||||||
7595 | |||||||
7596 | // We optimize the truncation of induction variables having constant | ||||||
7597 | // integer steps. The cost of these truncations is the same as the scalar | ||||||
7598 | // operation. | ||||||
7599 | if (isOptimizableIVTruncate(I, VF)) { | ||||||
7600 | auto *Trunc = cast<TruncInst>(I); | ||||||
7601 | return TTI.getCastInstrCost(Instruction::Trunc, Trunc->getDestTy(), | ||||||
7602 | Trunc->getSrcTy(), CCH, CostKind, Trunc); | ||||||
7603 | } | ||||||
7604 | |||||||
7605 | // Detect reduction patterns | ||||||
7606 | if (auto RedCost = getReductionPatternCost(I, VF, VectorTy, CostKind)) | ||||||
7607 | return *RedCost; | ||||||
7608 | |||||||
7609 | Type *SrcScalarTy = I->getOperand(0)->getType(); | ||||||
7610 | Type *SrcVecTy = | ||||||
7611 | VectorTy->isVectorTy() ? ToVectorTy(SrcScalarTy, VF) : SrcScalarTy; | ||||||
7612 | if (canTruncateToMinimalBitwidth(I, VF)) { | ||||||
7613 | // This cast is going to be shrunk. This may remove the cast or it might | ||||||
7614 | // turn it into slightly different cast. For example, if MinBW == 16, | ||||||
7615 | // "zext i8 %1 to i32" becomes "zext i8 %1 to i16". | ||||||
7616 | // | ||||||
7617 | // Calculate the modified src and dest types. | ||||||
7618 | Type *MinVecTy = VectorTy; | ||||||
7619 | if (Opcode == Instruction::Trunc) { | ||||||
7620 | SrcVecTy = smallestIntegerVectorType(SrcVecTy, MinVecTy); | ||||||
7621 | VectorTy = | ||||||
7622 | largestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy); | ||||||
7623 | } else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) { | ||||||
7624 | SrcVecTy = largestIntegerVectorType(SrcVecTy, MinVecTy); | ||||||
7625 | VectorTy = | ||||||
7626 | smallestIntegerVectorType(ToVectorTy(I->getType(), VF), MinVecTy); | ||||||
7627 | } | ||||||
7628 | } | ||||||
7629 | |||||||
7630 | return TTI.getCastInstrCost(Opcode, VectorTy, SrcVecTy, CCH, CostKind, I); | ||||||
7631 | } | ||||||
7632 | case Instruction::Call: { | ||||||
7633 | if (RecurrenceDescriptor::isFMulAddIntrinsic(I)) | ||||||
7634 | if (auto RedCost = getReductionPatternCost(I, VF, VectorTy, CostKind)) | ||||||
7635 | return *RedCost; | ||||||
7636 | bool NeedToScalarize; | ||||||
7637 | CallInst *CI = cast<CallInst>(I); | ||||||
7638 | InstructionCost CallCost = getVectorCallCost(CI, VF, NeedToScalarize); | ||||||
7639 | if (getVectorIntrinsicIDForCall(CI, TLI)) { | ||||||
7640 | InstructionCost IntrinsicCost = getVectorIntrinsicCost(CI, VF); | ||||||
7641 | return std::min(CallCost, IntrinsicCost); | ||||||
7642 | } | ||||||
7643 | return CallCost; | ||||||
7644 | } | ||||||
7645 | case Instruction::ExtractValue: | ||||||
7646 | return TTI.getInstructionCost(I, TTI::TCK_RecipThroughput); | ||||||
7647 | case Instruction::Alloca: | ||||||
7648 | // We cannot easily widen alloca to a scalable alloca, as | ||||||
7649 | // the result would need to be a vector of pointers. | ||||||
7650 | if (VF.isScalable()) | ||||||
7651 | return InstructionCost::getInvalid(); | ||||||
7652 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | ||||||
7653 | default: | ||||||
7654 | // This opcode is unknown. Assume that it is the same as 'mul'. | ||||||
7655 | return TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); | ||||||
7656 | } // end of switch. | ||||||
7657 | } | ||||||
7658 | |||||||
7659 | char LoopVectorize::ID = 0; | ||||||
7660 | |||||||
7661 | static const char lv_name[] = "Loop Vectorization"; | ||||||
7662 | |||||||
7663 | INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)static void *initializeLoopVectorizePassOnce(PassRegistry & Registry) { | ||||||
7664 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | ||||||
7665 | INITIALIZE_PASS_DEPENDENCY(BasicAAWrapperPass)initializeBasicAAWrapperPassPass(Registry); | ||||||
7666 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); | ||||||
7667 | INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)initializeGlobalsAAWrapperPassPass(Registry); | ||||||
7668 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); | ||||||
7669 | INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)initializeBlockFrequencyInfoWrapperPassPass(Registry); | ||||||
7670 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); | ||||||
7671 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry); | ||||||
7672 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry); | ||||||
7673 | INITIALIZE_PASS_DEPENDENCY(LoopAccessLegacyAnalysis)initializeLoopAccessLegacyAnalysisPass(Registry); | ||||||
7674 | INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)initializeDemandedBitsWrapperPassPass(Registry); | ||||||
7675 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)initializeOptimizationRemarkEmitterWrapperPassPass(Registry); | ||||||
7676 | INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)initializeProfileSummaryInfoWrapperPassPass(Registry); | ||||||
7677 | INITIALIZE_PASS_DEPENDENCY(InjectTLIMappingsLegacy)initializeInjectTLIMappingsLegacyPass(Registry); | ||||||
7678 | 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)); } | ||||||
7679 | |||||||
7680 | namespace llvm { | ||||||
7681 | |||||||
7682 | Pass *createLoopVectorizePass() { return new LoopVectorize(); } | ||||||
7683 | |||||||
7684 | Pass *createLoopVectorizePass(bool InterleaveOnlyWhenForced, | ||||||
7685 | bool VectorizeOnlyWhenForced) { | ||||||
7686 | return new LoopVectorize(InterleaveOnlyWhenForced, VectorizeOnlyWhenForced); | ||||||
7687 | } | ||||||
7688 | |||||||
7689 | } // end namespace llvm | ||||||
7690 | |||||||
7691 | bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) { | ||||||
7692 | // Check if the pointer operand of a load or store instruction is | ||||||
7693 | // consecutive. | ||||||
7694 | if (auto *Ptr = getLoadStorePointerOperand(Inst)) | ||||||
7695 | return Legal->isConsecutivePtr(getLoadStoreType(Inst), Ptr); | ||||||
7696 | return false; | ||||||
7697 | } | ||||||
7698 | |||||||
7699 | void LoopVectorizationCostModel::collectValuesToIgnore() { | ||||||
7700 | // Ignore ephemeral values. | ||||||
7701 | CodeMetrics::collectEphemeralValues(TheLoop, AC, ValuesToIgnore); | ||||||
7702 | |||||||
7703 | // Ignore type-promoting instructions we identified during reduction | ||||||
7704 | // detection. | ||||||
7705 | for (auto &Reduction : Legal->getReductionVars()) { | ||||||
7706 | const RecurrenceDescriptor &RedDes = Reduction.second; | ||||||
7707 | const SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts(); | ||||||
7708 | VecValuesToIgnore.insert(Casts.begin(), Casts.end()); | ||||||
7709 | } | ||||||
7710 | // Ignore type-casting instructions we identified during induction | ||||||
7711 | // detection. | ||||||
7712 | for (auto &Induction : Legal->getInductionVars()) { | ||||||
7713 | const InductionDescriptor &IndDes = Induction.second; | ||||||
7714 | const SmallVectorImpl<Instruction *> &Casts = IndDes.getCastInsts(); | ||||||
7715 | VecValuesToIgnore.insert(Casts.begin(), Casts.end()); | ||||||
7716 | } | ||||||
7717 | } | ||||||
7718 | |||||||
7719 | void LoopVectorizationCostModel::collectInLoopReductions() { | ||||||
7720 | for (auto &Reduction : Legal->getReductionVars()) { | ||||||
7721 | PHINode *Phi = Reduction.first; | ||||||
7722 | const RecurrenceDescriptor &RdxDesc = Reduction.second; | ||||||
7723 | |||||||
7724 | // We don't collect reductions that are type promoted (yet). | ||||||
7725 | if (RdxDesc.getRecurrenceType() != Phi->getType()) | ||||||
7726 | continue; | ||||||
7727 | |||||||
7728 | // If the target would prefer this reduction to happen "in-loop", then we | ||||||
7729 | // want to record it as such. | ||||||
7730 | unsigned Opcode = RdxDesc.getOpcode(); | ||||||
7731 | if (!PreferInLoopReductions && !useOrderedReductions(RdxDesc) && | ||||||
7732 | !TTI.preferInLoopReduction(Opcode, Phi->getType(), | ||||||
7733 | TargetTransformInfo::ReductionFlags())) | ||||||
7734 | continue; | ||||||
7735 | |||||||
7736 | // Check that we can correctly put the reductions into the loop, by | ||||||
7737 | // finding the chain of operations that leads from the phi to the loop | ||||||
7738 | // exit value. | ||||||
7739 | SmallVector<Instruction *, 4> ReductionOperations = | ||||||
7740 | RdxDesc.getReductionOpChain(Phi, TheLoop); | ||||||
7741 | bool InLoop = !ReductionOperations.empty(); | ||||||
7742 | if (InLoop) { | ||||||
7743 | InLoopReductionChains[Phi] = ReductionOperations; | ||||||
7744 | // Add the elements to InLoopReductionImmediateChains for cost modelling. | ||||||
7745 | Instruction *LastChain = Phi; | ||||||
7746 | for (auto *I : ReductionOperations) { | ||||||
7747 | InLoopReductionImmediateChains[I] = LastChain; | ||||||
7748 | LastChain = I; | ||||||
7749 | } | ||||||
7750 | } | ||||||
7751 | 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) | ||||||
7752 | << " 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); | ||||||
7753 | } | ||||||
7754 | } | ||||||
7755 | |||||||
7756 | // TODO: we could return a pair of values that specify the max VF and | ||||||
7757 | // min VF, to be used in `buildVPlans(MinVF, MaxVF)` instead of | ||||||
7758 | // `buildVPlans(VF, VF)`. We cannot do it because VPLAN at the moment | ||||||
7759 | // doesn't have a cost model that can choose which plan to execute if | ||||||
7760 | // more than one is generated. | ||||||
7761 | static unsigned determineVPlanVF(const unsigned WidestVectorRegBits, | ||||||
7762 | LoopVectorizationCostModel &CM) { | ||||||
7763 | unsigned WidestType; | ||||||
7764 | std::tie(std::ignore, WidestType) = CM.getSmallestAndWidestTypes(); | ||||||
7765 | return WidestVectorRegBits / WidestType; | ||||||
7766 | } | ||||||
7767 | |||||||
7768 | VectorizationFactor | ||||||
7769 | LoopVectorizationPlanner::planInVPlanNativePath(ElementCount UserVF) { | ||||||
7770 | 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", 7770, __extension__ __PRETTY_FUNCTION__)); | ||||||
7771 | ElementCount VF = UserVF; | ||||||
7772 | // Outer loop handling: They may require CFG and instruction level | ||||||
7773 | // transformations before even evaluating whether vectorization is profitable. | ||||||
7774 | // Since we cannot modify the incoming IR, we need to build VPlan upfront in | ||||||
7775 | // the vectorization pipeline. | ||||||
7776 | if (!OrigLoop->isInnermost()) { | ||||||
7777 | // If the user doesn't provide a vectorization factor, determine a | ||||||
7778 | // reasonable one. | ||||||
7779 | if (UserVF.isZero()) { | ||||||
7780 | VF = ElementCount::getFixed(determineVPlanVF( | ||||||
7781 | TTI->getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector) | ||||||
7782 | .getFixedSize(), | ||||||
7783 | CM)); | ||||||
7784 | 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); | ||||||
7785 | |||||||
7786 | // Make sure we have a VF > 1 for stress testing. | ||||||
7787 | if (VPlanBuildStressTest && (VF.isScalar() || VF.isZero())) { | ||||||
7788 | 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) | ||||||
7789 | << "overriding computed VF.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: VPlan stress testing: " << "overriding computed VF.\n"; } } while (false); | ||||||
7790 | VF = ElementCount::getFixed(4); | ||||||
7791 | } | ||||||
7792 | } | ||||||
7793 | 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", 7793, __extension__ __PRETTY_FUNCTION__)); | ||||||
7794 | 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", 7795, __extension__ __PRETTY_FUNCTION__)) | ||||||
7795 | "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", 7795, __extension__ __PRETTY_FUNCTION__)); | ||||||
7796 | 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) | ||||||
7797 | << "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); | ||||||
7798 | buildVPlans(VF, VF); | ||||||
7799 | |||||||
7800 | // For VPlan build stress testing, we bail out after VPlan construction. | ||||||
7801 | if (VPlanBuildStressTest) | ||||||
7802 | return VectorizationFactor::Disabled(); | ||||||
7803 | |||||||
7804 | return {VF, 0 /*Cost*/}; | ||||||
7805 | } | ||||||
7806 | |||||||
7807 | 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) | ||||||
7808 | 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) | ||||||
7809 | "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); | ||||||
7810 | return VectorizationFactor::Disabled(); | ||||||
7811 | } | ||||||
7812 | |||||||
7813 | Optional<VectorizationFactor> | ||||||
7814 | LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) { | ||||||
7815 | 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", 7815, __extension__ __PRETTY_FUNCTION__)); | ||||||
7816 | FixedScalableVFPair MaxFactors = CM.computeMaxVF(UserVF, UserIC); | ||||||
7817 | if (!MaxFactors) // Cases that should not to be vectorized nor interleaved. | ||||||
7818 | return None; | ||||||
7819 | |||||||
7820 | // Invalidate interleave groups if all blocks of loop will be predicated. | ||||||
7821 | if (CM.blockNeedsPredicationForAnyReason(OrigLoop->getHeader()) && | ||||||
7822 | !useMaskedInterleavedAccesses(*TTI)) { | ||||||
7823 | 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 ) | ||||||
7824 | 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 ) | ||||||
7825 | << "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 ) | ||||||
7826 | "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 ); | ||||||
7827 | if (CM.InterleaveInfo.invalidateGroups()) | ||||||
7828 | // Invalidating interleave groups also requires invalidating all decisions | ||||||
7829 | // based on them, which includes widening decisions and uniform and scalar | ||||||
7830 | // values. | ||||||
7831 | CM.invalidateCostModelingDecisions(); | ||||||
7832 | } | ||||||
7833 | |||||||
7834 | ElementCount MaxUserVF = | ||||||
7835 | UserVF.isScalable() ? MaxFactors.ScalableVF : MaxFactors.FixedVF; | ||||||
7836 | bool UserVFIsLegal = ElementCount::isKnownLE(UserVF, MaxUserVF); | ||||||
7837 | if (!UserVF.isZero() && UserVFIsLegal) { | ||||||
7838 | 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", 7839, __extension__ __PRETTY_FUNCTION__)) | ||||||
7839 | "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", 7839, __extension__ __PRETTY_FUNCTION__)); | ||||||
7840 | // Collect the instructions (and their associated costs) that will be more | ||||||
7841 | // profitable to scalarize. | ||||||
7842 | if (CM.selectUserVectorizationFactor(UserVF)) { | ||||||
7843 | 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); | ||||||
7844 | CM.collectInLoopReductions(); | ||||||
7845 | buildVPlansWithVPRecipes(UserVF, UserVF); | ||||||
7846 | LLVM_DEBUG(printPlans(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { printPlans(dbgs()); } } while (false); | ||||||
7847 | return {{UserVF, 0}}; | ||||||
7848 | } else | ||||||
7849 | reportVectorizationInfo("UserVF ignored because of invalid costs.", | ||||||
7850 | "InvalidCost", ORE, OrigLoop); | ||||||
7851 | } | ||||||
7852 | |||||||
7853 | // Populate the set of Vectorization Factor Candidates. | ||||||
7854 | ElementCountSet VFCandidates; | ||||||
7855 | for (auto VF = ElementCount::getFixed(1); | ||||||
7856 | ElementCount::isKnownLE(VF, MaxFactors.FixedVF); VF *= 2) | ||||||
7857 | VFCandidates.insert(VF); | ||||||
7858 | for (auto VF = ElementCount::getScalable(1); | ||||||
7859 | ElementCount::isKnownLE(VF, MaxFactors.ScalableVF); VF *= 2) | ||||||
7860 | VFCandidates.insert(VF); | ||||||
7861 | |||||||
7862 | for (const auto &VF : VFCandidates) { | ||||||
7863 | // Collect Uniform and Scalar instructions after vectorization with VF. | ||||||
7864 | CM.collectUniformsAndScalars(VF); | ||||||
7865 | |||||||
7866 | // Collect the instructions (and their associated costs) that will be more | ||||||
7867 | // profitable to scalarize. | ||||||
7868 | if (VF.isVector()) | ||||||
7869 | CM.collectInstsToScalarize(VF); | ||||||
7870 | } | ||||||
7871 | |||||||
7872 | CM.collectInLoopReductions(); | ||||||
7873 | buildVPlansWithVPRecipes(ElementCount::getFixed(1), MaxFactors.FixedVF); | ||||||
7874 | buildVPlansWithVPRecipes(ElementCount::getScalable(1), MaxFactors.ScalableVF); | ||||||
7875 | |||||||
7876 | LLVM_DEBUG(printPlans(dbgs()))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { printPlans(dbgs()); } } while (false); | ||||||
7877 | if (!MaxFactors.hasVector()) | ||||||
7878 | return VectorizationFactor::Disabled(); | ||||||
7879 | |||||||
7880 | // Select the optimal vectorization factor. | ||||||
7881 | auto SelectedVF = CM.selectVectorizationFactor(VFCandidates); | ||||||
7882 | |||||||
7883 | // Check if it is profitable to vectorize with runtime checks. | ||||||
7884 | unsigned NumRuntimePointerChecks = Requirements.getNumRuntimePointerChecks(); | ||||||
7885 | if (SelectedVF.Width.getKnownMinValue() > 1 && NumRuntimePointerChecks) { | ||||||
7886 | bool PragmaThresholdReached = | ||||||
7887 | NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold; | ||||||
7888 | bool ThresholdReached = | ||||||
7889 | NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold; | ||||||
7890 | if ((ThresholdReached && !Hints.allowReordering()) || | ||||||
7891 | PragmaThresholdReached) { | ||||||
7892 | ORE->emit([&]() { | ||||||
7893 | return OptimizationRemarkAnalysisAliasing( | ||||||
7894 | DEBUG_TYPE"loop-vectorize", "CantReorderMemOps", OrigLoop->getStartLoc(), | ||||||
7895 | OrigLoop->getHeader()) | ||||||
7896 | << "loop not vectorized: cannot prove it is safe to reorder " | ||||||
7897 | "memory operations"; | ||||||
7898 | }); | ||||||
7899 | 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); | ||||||
7900 | Hints.emitRemarkWithHints(); | ||||||
7901 | return VectorizationFactor::Disabled(); | ||||||
7902 | } | ||||||
7903 | } | ||||||
7904 | return SelectedVF; | ||||||
7905 | } | ||||||
7906 | |||||||
7907 | VPlan &LoopVectorizationPlanner::getBestPlanFor(ElementCount VF) const { | ||||||
7908 | 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", 7911, __extension__ __PRETTY_FUNCTION__)) | ||||||
7909 | [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", 7911, __extension__ __PRETTY_FUNCTION__)) | ||||||
7910 | 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", 7911, __extension__ __PRETTY_FUNCTION__)) | ||||||
7911 | "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", 7911, __extension__ __PRETTY_FUNCTION__)); | ||||||
7912 | |||||||
7913 | for (const VPlanPtr &Plan : VPlans) { | ||||||
7914 | if (Plan->hasVF(VF)) | ||||||
7915 | return *Plan.get(); | ||||||
7916 | } | ||||||
7917 | llvm_unreachable("No plan found!")::llvm::llvm_unreachable_internal("No plan found!", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 7917); | ||||||
7918 | } | ||||||
7919 | |||||||
7920 | static void AddRuntimeUnrollDisableMetaData(Loop *L) { | ||||||
7921 | SmallVector<Metadata *, 4> MDs; | ||||||
7922 | // Reserve first location for self reference to the LoopID metadata node. | ||||||
7923 | MDs.push_back(nullptr); | ||||||
7924 | bool IsUnrollMetadata = false; | ||||||
7925 | MDNode *LoopID = L->getLoopID(); | ||||||
7926 | if (LoopID) { | ||||||
7927 | // First find existing loop unrolling disable metadata. | ||||||
7928 | for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) { | ||||||
7929 | auto *MD = dyn_cast<MDNode>(LoopID->getOperand(i)); | ||||||
7930 | if (MD) { | ||||||
7931 | const auto *S = dyn_cast<MDString>(MD->getOperand(0)); | ||||||
7932 | IsUnrollMetadata = | ||||||
7933 | S && S->getString().startswith("llvm.loop.unroll.disable"); | ||||||
7934 | } | ||||||
7935 | MDs.push_back(LoopID->getOperand(i)); | ||||||
7936 | } | ||||||
7937 | } | ||||||
7938 | |||||||
7939 | if (!IsUnrollMetadata) { | ||||||
7940 | // Add runtime unroll disable metadata. | ||||||
7941 | LLVMContext &Context = L->getHeader()->getContext(); | ||||||
7942 | SmallVector<Metadata *, 1> DisableOperands; | ||||||
7943 | DisableOperands.push_back( | ||||||
7944 | MDString::get(Context, "llvm.loop.unroll.runtime.disable")); | ||||||
7945 | MDNode *DisableNode = MDNode::get(Context, DisableOperands); | ||||||
7946 | MDs.push_back(DisableNode); | ||||||
7947 | MDNode *NewLoopID = MDNode::get(Context, MDs); | ||||||
7948 | // Set operand 0 to refer to the loop id itself. | ||||||
7949 | NewLoopID->replaceOperandWith(0, NewLoopID); | ||||||
7950 | L->setLoopID(NewLoopID); | ||||||
7951 | } | ||||||
7952 | } | ||||||
7953 | |||||||
7954 | void LoopVectorizationPlanner::executePlan(ElementCount BestVF, unsigned BestUF, | ||||||
7955 | VPlan &BestVPlan, | ||||||
7956 | InnerLoopVectorizer &ILV, | ||||||
7957 | DominatorTree *DT) { | ||||||
7958 | 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) | ||||||
7959 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "Executing best plan with VF=" << BestVF << ", UF=" << BestUF << '\n' ; } } while (false); | ||||||
7960 | |||||||
7961 | // Perform the actual loop transformation. | ||||||
7962 | |||||||
7963 | // 1. Create a new empty loop. Unlink the old loop and connect the new one. | ||||||
7964 | VPTransformState State{BestVF, BestUF, LI, DT, ILV.Builder, &ILV, &BestVPlan}; | ||||||
7965 | Value *CanonicalIVStartValue; | ||||||
7966 | std::tie(State.CFG.PrevBB, CanonicalIVStartValue) = | ||||||
7967 | ILV.createVectorizedLoopSkeleton(); | ||||||
7968 | ILV.collectPoisonGeneratingRecipes(State); | ||||||
7969 | |||||||
7970 | ILV.printDebugTracesAtStart(); | ||||||
7971 | |||||||
7972 | //===------------------------------------------------===// | ||||||
7973 | // | ||||||
7974 | // Notice: any optimization or new instruction that go | ||||||
7975 | // into the code below should also be implemented in | ||||||
7976 | // the cost-model. | ||||||
7977 | // | ||||||
7978 | //===------------------------------------------------===// | ||||||
7979 | |||||||
7980 | // 2. Copy and widen instructions from the old loop into the new loop. | ||||||
7981 | BestVPlan.prepareToExecute(ILV.getOrCreateTripCount(nullptr), | ||||||
7982 | ILV.getOrCreateVectorTripCount(nullptr), | ||||||
7983 | CanonicalIVStartValue, State); | ||||||
7984 | BestVPlan.execute(&State); | ||||||
7985 | |||||||
7986 | // Keep all loop hints from the original loop on the vector loop (we'll | ||||||
7987 | // replace the vectorizer-specific hints below). | ||||||
7988 | MDNode *OrigLoopID = OrigLoop->getLoopID(); | ||||||
7989 | |||||||
7990 | Optional<MDNode *> VectorizedLoopID = | ||||||
7991 | makeFollowupLoopID(OrigLoopID, {LLVMLoopVectorizeFollowupAll, | ||||||
7992 | LLVMLoopVectorizeFollowupVectorized}); | ||||||
7993 | |||||||
7994 | Loop *L = LI->getLoopFor(State.CFG.PrevBB); | ||||||
7995 | if (VectorizedLoopID.hasValue()) | ||||||
7996 | L->setLoopID(VectorizedLoopID.getValue()); | ||||||
7997 | else { | ||||||
7998 | // Keep all loop hints from the original loop on the vector loop (we'll | ||||||
7999 | // replace the vectorizer-specific hints below). | ||||||
8000 | if (MDNode *LID = OrigLoop->getLoopID()) | ||||||
8001 | L->setLoopID(LID); | ||||||
8002 | |||||||
8003 | LoopVectorizeHints Hints(L, true, *ORE); | ||||||
8004 | Hints.setAlreadyVectorized(); | ||||||
8005 | } | ||||||
8006 | // Disable runtime unrolling when vectorizing the epilogue loop. | ||||||
8007 | if (CanonicalIVStartValue) | ||||||
8008 | AddRuntimeUnrollDisableMetaData(L); | ||||||
8009 | |||||||
8010 | // 3. Fix the vectorized code: take care of header phi's, live-outs, | ||||||
8011 | // predication, updating analyses. | ||||||
8012 | ILV.fixVectorizedLoop(State); | ||||||
8013 | |||||||
8014 | ILV.printDebugTracesAtEnd(); | ||||||
8015 | } | ||||||
8016 | |||||||
8017 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||
8018 | void LoopVectorizationPlanner::printPlans(raw_ostream &O) { | ||||||
8019 | for (const auto &Plan : VPlans) | ||||||
8020 | if (PrintVPlansInDotFormat) | ||||||
8021 | Plan->printDOT(O); | ||||||
8022 | else | ||||||
8023 | Plan->print(O); | ||||||
8024 | } | ||||||
8025 | #endif | ||||||
8026 | |||||||
8027 | void LoopVectorizationPlanner::collectTriviallyDeadInstructions( | ||||||
8028 | SmallPtrSetImpl<Instruction *> &DeadInstructions) { | ||||||
8029 | |||||||
8030 | // We create new control-flow for the vectorized loop, so the original exit | ||||||
8031 | // conditions will be dead after vectorization if it's only used by the | ||||||
8032 | // terminator | ||||||
8033 | SmallVector<BasicBlock*> ExitingBlocks; | ||||||
8034 | OrigLoop->getExitingBlocks(ExitingBlocks); | ||||||
8035 | for (auto *BB : ExitingBlocks) { | ||||||
8036 | auto *Cmp = dyn_cast<Instruction>(BB->getTerminator()->getOperand(0)); | ||||||
8037 | if (!Cmp || !Cmp->hasOneUse()) | ||||||
8038 | continue; | ||||||
8039 | |||||||
8040 | // TODO: we should introduce a getUniqueExitingBlocks on Loop | ||||||
8041 | if (!DeadInstructions.insert(Cmp).second) | ||||||
8042 | continue; | ||||||
8043 | |||||||
8044 | // The operands of the icmp is often a dead trunc, used by IndUpdate. | ||||||
8045 | // TODO: can recurse through operands in general | ||||||
8046 | for (Value *Op : Cmp->operands()) { | ||||||
8047 | if (isa<TruncInst>(Op) && Op->hasOneUse()) | ||||||
8048 | DeadInstructions.insert(cast<Instruction>(Op)); | ||||||
8049 | } | ||||||
8050 | } | ||||||
8051 | |||||||
8052 | // We create new "steps" for induction variable updates to which the original | ||||||
8053 | // induction variables map. An original update instruction will be dead if | ||||||
8054 | // all its users except the induction variable are dead. | ||||||
8055 | auto *Latch = OrigLoop->getLoopLatch(); | ||||||
8056 | for (auto &Induction : Legal->getInductionVars()) { | ||||||
8057 | PHINode *Ind = Induction.first; | ||||||
8058 | auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch)); | ||||||
8059 | |||||||
8060 | // If the tail is to be folded by masking, the primary induction variable, | ||||||
8061 | // if exists, isn't dead: it will be used for masking. Don't kill it. | ||||||
8062 | if (CM.foldTailByMasking() && IndUpdate == Legal->getPrimaryInduction()) | ||||||
8063 | continue; | ||||||
8064 | |||||||
8065 | if (llvm::all_of(IndUpdate->users(), [&](User *U) -> bool { | ||||||
8066 | return U == Ind || DeadInstructions.count(cast<Instruction>(U)); | ||||||
8067 | })) | ||||||
8068 | DeadInstructions.insert(IndUpdate); | ||||||
8069 | } | ||||||
8070 | } | ||||||
8071 | |||||||
8072 | Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) { return V; } | ||||||
8073 | |||||||
8074 | //===--------------------------------------------------------------------===// | ||||||
8075 | // EpilogueVectorizerMainLoop | ||||||
8076 | //===--------------------------------------------------------------------===// | ||||||
8077 | |||||||
8078 | /// This function is partially responsible for generating the control flow | ||||||
8079 | /// depicted in https://llvm.org/docs/Vectorizers.html#epilogue-vectorization. | ||||||
8080 | std::pair<BasicBlock *, Value *> | ||||||
8081 | EpilogueVectorizerMainLoop::createEpilogueVectorizedLoopSkeleton() { | ||||||
8082 | MDNode *OrigLoopID = OrigLoop->getLoopID(); | ||||||
8083 | Loop *Lp = createVectorLoopSkeleton(""); | ||||||
8084 | |||||||
8085 | // Generate the code to check the minimum iteration count of the vector | ||||||
8086 | // epilogue (see below). | ||||||
8087 | EPI.EpilogueIterationCountCheck = | ||||||
8088 | emitMinimumIterationCountCheck(Lp, LoopScalarPreHeader, true); | ||||||
8089 | EPI.EpilogueIterationCountCheck->setName("iter.check"); | ||||||
8090 | |||||||
8091 | // Generate the code to check any assumptions that we've made for SCEV | ||||||
8092 | // expressions. | ||||||
8093 | EPI.SCEVSafetyCheck = emitSCEVChecks(Lp, LoopScalarPreHeader); | ||||||
8094 | |||||||
8095 | // Generate the code that checks at runtime if arrays overlap. We put the | ||||||
8096 | // checks into a separate block to make the more common case of few elements | ||||||
8097 | // faster. | ||||||
8098 | EPI.MemSafetyCheck = emitMemRuntimeChecks(Lp, LoopScalarPreHeader); | ||||||
8099 | |||||||
8100 | // Generate the iteration count check for the main loop, *after* the check | ||||||
8101 | // for the epilogue loop, so that the path-length is shorter for the case | ||||||
8102 | // that goes directly through the vector epilogue. The longer-path length for | ||||||
8103 | // the main loop is compensated for, by the gain from vectorizing the larger | ||||||
8104 | // trip count. Note: the branch will get updated later on when we vectorize | ||||||
8105 | // the epilogue. | ||||||
8106 | EPI.MainLoopIterationCountCheck = | ||||||
8107 | emitMinimumIterationCountCheck(Lp, LoopScalarPreHeader, false); | ||||||
8108 | |||||||
8109 | // Generate the induction variable. | ||||||
8110 | Value *CountRoundDown = getOrCreateVectorTripCount(Lp); | ||||||
8111 | EPI.VectorTripCount = CountRoundDown; | ||||||
8112 | createHeaderBranch(Lp); | ||||||
8113 | |||||||
8114 | // Skip induction resume value creation here because they will be created in | ||||||
8115 | // the second pass. If we created them here, they wouldn't be used anyway, | ||||||
8116 | // because the vplan in the second pass still contains the inductions from the | ||||||
8117 | // original loop. | ||||||
8118 | |||||||
8119 | return {completeLoopSkeleton(Lp, OrigLoopID), nullptr}; | ||||||
8120 | } | ||||||
8121 | |||||||
8122 | void EpilogueVectorizerMainLoop::printDebugTracesAtStart() { | ||||||
8123 | 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) | ||||||
8124 | 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) | ||||||
8125 | << "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) | ||||||
8126 | << ", 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) | ||||||
8127 | << ", 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) | ||||||
8128 | << ", 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) | ||||||
8129 | })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); | ||||||
8130 | } | ||||||
8131 | |||||||
8132 | void EpilogueVectorizerMainLoop::printDebugTracesAtEnd() { | ||||||
8133 | DEBUG_WITH_TYPE(VerboseDebug, {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }; } } while (false) | ||||||
8134 | dbgs() << "intermediate fn:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }; } } while (false) | ||||||
8135 | << *OrigLoop->getHeader()->getParent() << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }; } } while (false) | ||||||
8136 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }; } } while (false); | ||||||
8137 | } | ||||||
8138 | |||||||
8139 | BasicBlock *EpilogueVectorizerMainLoop::emitMinimumIterationCountCheck( | ||||||
8140 | Loop *L, BasicBlock *Bypass, bool ForEpilogue) { | ||||||
8141 | assert(L && "Expected valid Loop.")(static_cast <bool> (L && "Expected valid Loop." ) ? void (0) : __assert_fail ("L && \"Expected valid Loop.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8141, __extension__ __PRETTY_FUNCTION__)); | ||||||
8142 | 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", 8142, __extension__ __PRETTY_FUNCTION__)); | ||||||
8143 | ElementCount VFactor = ForEpilogue ? EPI.EpilogueVF : VF; | ||||||
8144 | unsigned UFactor = ForEpilogue ? EPI.EpilogueUF : UF; | ||||||
8145 | Value *Count = getOrCreateTripCount(L); | ||||||
8146 | // Reuse existing vector loop preheader for TC checks. | ||||||
8147 | // Note that new preheader block is generated for vector loop. | ||||||
8148 | BasicBlock *const TCCheckBlock = LoopVectorPreHeader; | ||||||
8149 | IRBuilder<> Builder(TCCheckBlock->getTerminator()); | ||||||
8150 | |||||||
8151 | // Generate code to check if the loop's trip count is less than VF * UF of the | ||||||
8152 | // main vector loop. | ||||||
8153 | auto P = Cost->requiresScalarEpilogue(ForEpilogue ? EPI.EpilogueVF : VF) ? | ||||||
8154 | ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT; | ||||||
8155 | |||||||
8156 | Value *CheckMinIters = Builder.CreateICmp( | ||||||
8157 | P, Count, createStepForVF(Builder, Count->getType(), VFactor, UFactor), | ||||||
8158 | "min.iters.check"); | ||||||
8159 | |||||||
8160 | if (!ForEpilogue) | ||||||
8161 | TCCheckBlock->setName("vector.main.loop.iter.check"); | ||||||
8162 | |||||||
8163 | // Create new preheader for vector loop. | ||||||
8164 | LoopVectorPreHeader = SplitBlock(TCCheckBlock, TCCheckBlock->getTerminator(), | ||||||
8165 | DT, LI, nullptr, "vector.ph"); | ||||||
8166 | |||||||
8167 | if (ForEpilogue) { | ||||||
8168 | 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", 8170, __extension__ __PRETTY_FUNCTION__)) | ||||||
8169 | 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", 8170, __extension__ __PRETTY_FUNCTION__)) | ||||||
8170 | "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", 8170, __extension__ __PRETTY_FUNCTION__)); | ||||||
8171 | |||||||
8172 | // Update dominator for Bypass & LoopExit. | ||||||
8173 | DT->changeImmediateDominator(Bypass, TCCheckBlock); | ||||||
8174 | if (!Cost->requiresScalarEpilogue(EPI.EpilogueVF)) | ||||||
8175 | // For loops with multiple exits, there's no edge from the middle block | ||||||
8176 | // to exit blocks (as the epilogue must run) and thus no need to update | ||||||
8177 | // the immediate dominator of the exit blocks. | ||||||
8178 | DT->changeImmediateDominator(LoopExitBlock, TCCheckBlock); | ||||||
8179 | |||||||
8180 | LoopBypassBlocks.push_back(TCCheckBlock); | ||||||
8181 | |||||||
8182 | // Save the trip count so we don't have to regenerate it in the | ||||||
8183 | // vec.epilog.iter.check. This is safe to do because the trip count | ||||||
8184 | // generated here dominates the vector epilog iter check. | ||||||
8185 | EPI.TripCount = Count; | ||||||
8186 | } | ||||||
8187 | |||||||
8188 | ReplaceInstWithInst( | ||||||
8189 | TCCheckBlock->getTerminator(), | ||||||
8190 | BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); | ||||||
8191 | |||||||
8192 | return TCCheckBlock; | ||||||
8193 | } | ||||||
8194 | |||||||
8195 | //===--------------------------------------------------------------------===// | ||||||
8196 | // EpilogueVectorizerEpilogueLoop | ||||||
8197 | //===--------------------------------------------------------------------===// | ||||||
8198 | |||||||
8199 | /// This function is partially responsible for generating the control flow | ||||||
8200 | /// depicted in https://llvm.org/docs/Vectorizers.html#epilogue-vectorization. | ||||||
8201 | std::pair<BasicBlock *, Value *> | ||||||
8202 | EpilogueVectorizerEpilogueLoop::createEpilogueVectorizedLoopSkeleton() { | ||||||
8203 | MDNode *OrigLoopID = OrigLoop->getLoopID(); | ||||||
8204 | Loop *Lp = createVectorLoopSkeleton("vec.epilog."); | ||||||
8205 | |||||||
8206 | // Now, compare the remaining count and if there aren't enough iterations to | ||||||
8207 | // execute the vectorized epilogue skip to the scalar part. | ||||||
8208 | BasicBlock *VecEpilogueIterationCountCheck = LoopVectorPreHeader; | ||||||
8209 | VecEpilogueIterationCountCheck->setName("vec.epilog.iter.check"); | ||||||
8210 | LoopVectorPreHeader = | ||||||
8211 | SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, | ||||||
8212 | LI, nullptr, "vec.epilog.ph"); | ||||||
8213 | emitMinimumVectorEpilogueIterCountCheck(Lp, LoopScalarPreHeader, | ||||||
8214 | VecEpilogueIterationCountCheck); | ||||||
8215 | |||||||
8216 | // Adjust the control flow taking the state info from the main loop | ||||||
8217 | // vectorization into account. | ||||||
8218 | 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", 8219, __extension__ __PRETTY_FUNCTION__)) | ||||||
8219 | "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", 8219, __extension__ __PRETTY_FUNCTION__)); | ||||||
8220 | EPI.MainLoopIterationCountCheck->getTerminator()->replaceUsesOfWith( | ||||||
8221 | VecEpilogueIterationCountCheck, LoopVectorPreHeader); | ||||||
8222 | |||||||
8223 | DT->changeImmediateDominator(LoopVectorPreHeader, | ||||||
8224 | EPI.MainLoopIterationCountCheck); | ||||||
8225 | |||||||
8226 | EPI.EpilogueIterationCountCheck->getTerminator()->replaceUsesOfWith( | ||||||
8227 | VecEpilogueIterationCountCheck, LoopScalarPreHeader); | ||||||
8228 | |||||||
8229 | if (EPI.SCEVSafetyCheck) | ||||||
8230 | EPI.SCEVSafetyCheck->getTerminator()->replaceUsesOfWith( | ||||||
8231 | VecEpilogueIterationCountCheck, LoopScalarPreHeader); | ||||||
8232 | if (EPI.MemSafetyCheck) | ||||||
8233 | EPI.MemSafetyCheck->getTerminator()->replaceUsesOfWith( | ||||||
8234 | VecEpilogueIterationCountCheck, LoopScalarPreHeader); | ||||||
8235 | |||||||
8236 | DT->changeImmediateDominator( | ||||||
8237 | VecEpilogueIterationCountCheck, | ||||||
8238 | VecEpilogueIterationCountCheck->getSinglePredecessor()); | ||||||
8239 | |||||||
8240 | DT->changeImmediateDominator(LoopScalarPreHeader, | ||||||
8241 | EPI.EpilogueIterationCountCheck); | ||||||
8242 | if (!Cost->requiresScalarEpilogue(EPI.EpilogueVF)) | ||||||
8243 | // If there is an epilogue which must run, there's no edge from the | ||||||
8244 | // middle block to exit blocks and thus no need to update the immediate | ||||||
8245 | // dominator of the exit blocks. | ||||||
8246 | DT->changeImmediateDominator(LoopExitBlock, | ||||||
8247 | EPI.EpilogueIterationCountCheck); | ||||||
8248 | |||||||
8249 | // Keep track of bypass blocks, as they feed start values to the induction | ||||||
8250 | // phis in the scalar loop preheader. | ||||||
8251 | if (EPI.SCEVSafetyCheck) | ||||||
8252 | LoopBypassBlocks.push_back(EPI.SCEVSafetyCheck); | ||||||
8253 | if (EPI.MemSafetyCheck) | ||||||
8254 | LoopBypassBlocks.push_back(EPI.MemSafetyCheck); | ||||||
8255 | LoopBypassBlocks.push_back(EPI.EpilogueIterationCountCheck); | ||||||
8256 | |||||||
8257 | // Generate a resume induction for the vector epilogue and put it in the | ||||||
8258 | // vector epilogue preheader | ||||||
8259 | Type *IdxTy = Legal->getWidestInductionType(); | ||||||
8260 | PHINode *EPResumeVal = PHINode::Create(IdxTy, 2, "vec.epilog.resume.val", | ||||||
8261 | LoopVectorPreHeader->getFirstNonPHI()); | ||||||
8262 | EPResumeVal->addIncoming(EPI.VectorTripCount, VecEpilogueIterationCountCheck); | ||||||
8263 | EPResumeVal->addIncoming(ConstantInt::get(IdxTy, 0), | ||||||
8264 | EPI.MainLoopIterationCountCheck); | ||||||
8265 | |||||||
8266 | // Generate the induction variable. | ||||||
8267 | createHeaderBranch(Lp); | ||||||
8268 | |||||||
8269 | // Generate induction resume values. These variables save the new starting | ||||||
8270 | // indexes for the scalar loop. They are used to test if there are any tail | ||||||
8271 | // iterations left once the vector loop has completed. | ||||||
8272 | // Note that when the vectorized epilogue is skipped due to iteration count | ||||||
8273 | // check, then the resume value for the induction variable comes from | ||||||
8274 | // the trip count of the main vector loop, hence passing the AdditionalBypass | ||||||
8275 | // argument. | ||||||
8276 | createInductionResumeValues(Lp, {VecEpilogueIterationCountCheck, | ||||||
8277 | EPI.VectorTripCount} /* AdditionalBypass */); | ||||||
8278 | |||||||
8279 | return {completeLoopSkeleton(Lp, OrigLoopID), EPResumeVal}; | ||||||
8280 | } | ||||||
8281 | |||||||
8282 | BasicBlock * | ||||||
8283 | EpilogueVectorizerEpilogueLoop::emitMinimumVectorEpilogueIterCountCheck( | ||||||
8284 | Loop *L, BasicBlock *Bypass, BasicBlock *Insert) { | ||||||
8285 | |||||||
8286 | 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", 8287, __extension__ __PRETTY_FUNCTION__)) | ||||||
8287 | "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", 8287, __extension__ __PRETTY_FUNCTION__)); | ||||||
8288 | 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", 8291, __extension__ __PRETTY_FUNCTION__)) | ||||||
8289 | (!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", 8291, __extension__ __PRETTY_FUNCTION__)) | ||||||
8290 | 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", 8291, __extension__ __PRETTY_FUNCTION__)) | ||||||
8291 | "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", 8291, __extension__ __PRETTY_FUNCTION__)); | ||||||
8292 | Value *TC = EPI.TripCount; | ||||||
8293 | IRBuilder<> Builder(Insert->getTerminator()); | ||||||
8294 | Value *Count = Builder.CreateSub(TC, EPI.VectorTripCount, "n.vec.remaining"); | ||||||
8295 | |||||||
8296 | // Generate code to check if the loop's trip count is less than VF * UF of the | ||||||
8297 | // vector epilogue loop. | ||||||
8298 | auto P = Cost->requiresScalarEpilogue(EPI.EpilogueVF) ? | ||||||
8299 | ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT; | ||||||
8300 | |||||||
8301 | Value *CheckMinIters = | ||||||
8302 | Builder.CreateICmp(P, Count, | ||||||
8303 | createStepForVF(Builder, Count->getType(), | ||||||
8304 | EPI.EpilogueVF, EPI.EpilogueUF), | ||||||
8305 | "min.epilog.iters.check"); | ||||||
8306 | |||||||
8307 | ReplaceInstWithInst( | ||||||
8308 | Insert->getTerminator(), | ||||||
8309 | BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters)); | ||||||
8310 | |||||||
8311 | LoopBypassBlocks.push_back(Insert); | ||||||
8312 | return Insert; | ||||||
8313 | } | ||||||
8314 | |||||||
8315 | void EpilogueVectorizerEpilogueLoop::printDebugTracesAtStart() { | ||||||
8316 | 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) | ||||||
8317 | 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) | ||||||
8318 | << "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) | ||||||
8319 | << ", 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) | ||||||
8320 | })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); | ||||||
8321 | } | ||||||
8322 | |||||||
8323 | void EpilogueVectorizerEpilogueLoop::printDebugTracesAtEnd() { | ||||||
8324 | DEBUG_WITH_TYPE(VerboseDebug, {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "final fn:\n" << *OrigLoop ->getHeader()->getParent() << "\n"; }; } } while ( false) | ||||||
8325 | 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) | ||||||
8326 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType (VerboseDebug)) { { dbgs() << "final fn:\n" << *OrigLoop ->getHeader()->getParent() << "\n"; }; } } while ( false); | ||||||
8327 | } | ||||||
8328 | |||||||
8329 | bool LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||
8330 | const std::function<bool(ElementCount)> &Predicate, VFRange &Range) { | ||||||
8331 | 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", 8331, __extension__ __PRETTY_FUNCTION__)); | ||||||
8332 | bool PredicateAtRangeStart = Predicate(Range.Start); | ||||||
8333 | |||||||
8334 | for (ElementCount TmpVF = Range.Start * 2; | ||||||
8335 | ElementCount::isKnownLT(TmpVF, Range.End); TmpVF *= 2) | ||||||
8336 | if (Predicate(TmpVF) != PredicateAtRangeStart) { | ||||||
8337 | Range.End = TmpVF; | ||||||
8338 | break; | ||||||
8339 | } | ||||||
8340 | |||||||
8341 | return PredicateAtRangeStart; | ||||||
8342 | } | ||||||
8343 | |||||||
8344 | /// Build VPlans for the full range of feasible VF's = {\p MinVF, 2 * \p MinVF, | ||||||
8345 | /// 4 * \p MinVF, ..., \p MaxVF} by repeatedly building a VPlan for a sub-range | ||||||
8346 | /// of VF's starting at a given VF and extending it as much as possible. Each | ||||||
8347 | /// vectorization decision can potentially shorten this sub-range during | ||||||
8348 | /// buildVPlan(). | ||||||
8349 | void LoopVectorizationPlanner::buildVPlans(ElementCount MinVF, | ||||||
8350 | ElementCount MaxVF) { | ||||||
8351 | auto MaxVFPlusOne = MaxVF.getWithIncrement(1); | ||||||
8352 | for (ElementCount VF = MinVF; ElementCount::isKnownLT(VF, MaxVFPlusOne);) { | ||||||
8353 | VFRange SubRange = {VF, MaxVFPlusOne}; | ||||||
8354 | VPlans.push_back(buildVPlan(SubRange)); | ||||||
8355 | VF = SubRange.End; | ||||||
8356 | } | ||||||
8357 | } | ||||||
8358 | |||||||
8359 | VPValue *VPRecipeBuilder::createEdgeMask(BasicBlock *Src, BasicBlock *Dst, | ||||||
8360 | VPlanPtr &Plan) { | ||||||
8361 | 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", 8361, __extension__ __PRETTY_FUNCTION__)); | ||||||
8362 | |||||||
8363 | // Look for cached value. | ||||||
8364 | std::pair<BasicBlock *, BasicBlock *> Edge(Src, Dst); | ||||||
8365 | EdgeMaskCacheTy::iterator ECEntryIt = EdgeMaskCache.find(Edge); | ||||||
8366 | if (ECEntryIt != EdgeMaskCache.end()) | ||||||
8367 | return ECEntryIt->second; | ||||||
8368 | |||||||
8369 | VPValue *SrcMask = createBlockInMask(Src, Plan); | ||||||
8370 | |||||||
8371 | // The terminator has to be a branch inst! | ||||||
8372 | BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator()); | ||||||
8373 | 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", 8373, __extension__ __PRETTY_FUNCTION__)); | ||||||
8374 | |||||||
8375 | if (!BI->isConditional() || BI->getSuccessor(0) == BI->getSuccessor(1)) | ||||||
8376 | return EdgeMaskCache[Edge] = SrcMask; | ||||||
8377 | |||||||
8378 | // If source is an exiting block, we know the exit edge is dynamically dead | ||||||
8379 | // in the vector loop, and thus we don't need to restrict the mask. Avoid | ||||||
8380 | // adding uses of an otherwise potentially dead instruction. | ||||||
8381 | if (OrigLoop->isLoopExiting(Src)) | ||||||
8382 | return EdgeMaskCache[Edge] = SrcMask; | ||||||
8383 | |||||||
8384 | VPValue *EdgeMask = Plan->getOrAddVPValue(BI->getCondition()); | ||||||
8385 | 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", 8385, __extension__ __PRETTY_FUNCTION__)); | ||||||
8386 | |||||||
8387 | if (BI->getSuccessor(0) != Dst) | ||||||
8388 | EdgeMask = Builder.createNot(EdgeMask, BI->getDebugLoc()); | ||||||
8389 | |||||||
8390 | if (SrcMask) { // Otherwise block in-mask is all-one, no need to AND. | ||||||
8391 | // The condition is 'SrcMask && EdgeMask', which is equivalent to | ||||||
8392 | // 'select i1 SrcMask, i1 EdgeMask, i1 false'. | ||||||
8393 | // The select version does not introduce new UB if SrcMask is false and | ||||||
8394 | // EdgeMask is poison. Using 'and' here introduces undefined behavior. | ||||||
8395 | VPValue *False = Plan->getOrAddVPValue( | ||||||
8396 | ConstantInt::getFalse(BI->getCondition()->getType())); | ||||||
8397 | EdgeMask = | ||||||
8398 | Builder.createSelect(SrcMask, EdgeMask, False, BI->getDebugLoc()); | ||||||
8399 | } | ||||||
8400 | |||||||
8401 | return EdgeMaskCache[Edge] = EdgeMask; | ||||||
8402 | } | ||||||
8403 | |||||||
8404 | VPValue *VPRecipeBuilder::createBlockInMask(BasicBlock *BB, VPlanPtr &Plan) { | ||||||
8405 | 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", 8405, __extension__ __PRETTY_FUNCTION__)); | ||||||
8406 | |||||||
8407 | // Look for cached value. | ||||||
8408 | BlockMaskCacheTy::iterator BCEntryIt = BlockMaskCache.find(BB); | ||||||
8409 | if (BCEntryIt != BlockMaskCache.end()) | ||||||
8410 | return BCEntryIt->second; | ||||||
8411 | |||||||
8412 | // All-one mask is modelled as no-mask following the convention for masked | ||||||
8413 | // load/store/gather/scatter. Initialize BlockMask to no-mask. | ||||||
8414 | VPValue *BlockMask = nullptr; | ||||||
8415 | |||||||
8416 | if (OrigLoop->getHeader() == BB) { | ||||||
8417 | if (!CM.blockNeedsPredicationForAnyReason(BB)) | ||||||
8418 | return BlockMaskCache[BB] = BlockMask; // Loop incoming mask is all-one. | ||||||
8419 | |||||||
8420 | // Introduce the early-exit compare IV <= BTC to form header block mask. | ||||||
8421 | // This is used instead of IV < TC because TC may wrap, unlike BTC. Start by | ||||||
8422 | // constructing the desired canonical IV in the header block as its first | ||||||
8423 | // non-phi instructions. | ||||||
8424 | 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", 8424, __extension__ __PRETTY_FUNCTION__)); | ||||||
8425 | VPBasicBlock *HeaderVPBB = Plan->getEntry()->getEntryBasicBlock(); | ||||||
8426 | auto NewInsertionPoint = HeaderVPBB->getFirstNonPhi(); | ||||||
8427 | |||||||
8428 | VPValue *IV = nullptr; | ||||||
8429 | if (Legal->getPrimaryInduction()) | ||||||
8430 | IV = Plan->getOrAddVPValue(Legal->getPrimaryInduction()); | ||||||
8431 | else { | ||||||
8432 | auto *IVRecipe = new VPWidenCanonicalIVRecipe(Plan->getCanonicalIV()); | ||||||
8433 | HeaderVPBB->insert(IVRecipe, NewInsertionPoint); | ||||||
8434 | IV = IVRecipe; | ||||||
8435 | } | ||||||
8436 | |||||||
8437 | VPBuilder::InsertPointGuard Guard(Builder); | ||||||
8438 | Builder.setInsertPoint(HeaderVPBB, NewInsertionPoint); | ||||||
8439 | if (CM.TTI.emitGetActiveLaneMask()) { | ||||||
8440 | VPValue *TC = Plan->getOrCreateTripCount(); | ||||||
8441 | BlockMask = Builder.createNaryOp(VPInstruction::ActiveLaneMask, {IV, TC}); | ||||||
8442 | } else { | ||||||
8443 | VPValue *BTC = Plan->getOrCreateBackedgeTakenCount(); | ||||||
8444 | BlockMask = Builder.createNaryOp(VPInstruction::ICmpULE, {IV, BTC}); | ||||||
8445 | } | ||||||
8446 | return BlockMaskCache[BB] = BlockMask; | ||||||
8447 | } | ||||||
8448 | |||||||
8449 | // This is the block mask. We OR all incoming edges. | ||||||
8450 | for (auto *Predecessor : predecessors(BB)) { | ||||||
8451 | VPValue *EdgeMask = createEdgeMask(Predecessor, BB, Plan); | ||||||
8452 | if (!EdgeMask) // Mask of predecessor is all-one so mask of block is too. | ||||||
8453 | return BlockMaskCache[BB] = EdgeMask; | ||||||
8454 | |||||||
8455 | if (!BlockMask) { // BlockMask has its initialized nullptr value. | ||||||
8456 | BlockMask = EdgeMask; | ||||||
8457 | continue; | ||||||
8458 | } | ||||||
8459 | |||||||
8460 | BlockMask = Builder.createOr(BlockMask, EdgeMask, {}); | ||||||
8461 | } | ||||||
8462 | |||||||
8463 | return BlockMaskCache[BB] = BlockMask; | ||||||
8464 | } | ||||||
8465 | |||||||
8466 | VPRecipeBase *VPRecipeBuilder::tryToWidenMemory(Instruction *I, | ||||||
8467 | ArrayRef<VPValue *> Operands, | ||||||
8468 | VFRange &Range, | ||||||
8469 | VPlanPtr &Plan) { | ||||||
8470 | 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", 8471, __extension__ __PRETTY_FUNCTION__)) | ||||||
8471 | "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", 8471, __extension__ __PRETTY_FUNCTION__)); | ||||||
8472 | |||||||
8473 | auto willWiden = [&](ElementCount VF) -> bool { | ||||||
8474 | if (VF.isScalar()) | ||||||
8475 | return false; | ||||||
8476 | LoopVectorizationCostModel::InstWidening Decision = | ||||||
8477 | CM.getWideningDecision(I, VF); | ||||||
8478 | 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", 8479, __extension__ __PRETTY_FUNCTION__)) | ||||||
8479 | "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", 8479, __extension__ __PRETTY_FUNCTION__)); | ||||||
8480 | if (Decision == LoopVectorizationCostModel::CM_Interleave) | ||||||
8481 | return true; | ||||||
8482 | if (CM.isScalarAfterVectorization(I, VF) || | ||||||
8483 | CM.isProfitableToScalarize(I, VF)) | ||||||
8484 | return false; | ||||||
8485 | return Decision != LoopVectorizationCostModel::CM_Scalarize; | ||||||
8486 | }; | ||||||
8487 | |||||||
8488 | if (!LoopVectorizationPlanner::getDecisionAndClampRange(willWiden, Range)) | ||||||
8489 | return nullptr; | ||||||
8490 | |||||||
8491 | VPValue *Mask = nullptr; | ||||||
8492 | if (Legal->isMaskRequired(I)) | ||||||
8493 | Mask = createBlockInMask(I->getParent(), Plan); | ||||||
8494 | |||||||
8495 | // Determine if the pointer operand of the access is either consecutive or | ||||||
8496 | // reverse consecutive. | ||||||
8497 | LoopVectorizationCostModel::InstWidening Decision = | ||||||
8498 | CM.getWideningDecision(I, Range.Start); | ||||||
8499 | bool Reverse = Decision == LoopVectorizationCostModel::CM_Widen_Reverse; | ||||||
8500 | bool Consecutive = | ||||||
8501 | Reverse || Decision == LoopVectorizationCostModel::CM_Widen; | ||||||
8502 | |||||||
8503 | if (LoadInst *Load = dyn_cast<LoadInst>(I)) | ||||||
8504 | return new VPWidenMemoryInstructionRecipe(*Load, Operands[0], Mask, | ||||||
8505 | Consecutive, Reverse); | ||||||
8506 | |||||||
8507 | StoreInst *Store = cast<StoreInst>(I); | ||||||
8508 | return new VPWidenMemoryInstructionRecipe(*Store, Operands[1], Operands[0], | ||||||
8509 | Mask, Consecutive, Reverse); | ||||||
8510 | } | ||||||
8511 | |||||||
8512 | VPWidenIntOrFpInductionRecipe * | ||||||
8513 | VPRecipeBuilder::tryToOptimizeInductionPHI(PHINode *Phi, | ||||||
8514 | ArrayRef<VPValue *> Operands) const { | ||||||
8515 | // Check if this is an integer or fp induction. If so, build the recipe that | ||||||
8516 | // produces its scalar and vector values. | ||||||
8517 | if (auto *II = Legal->getIntOrFpInductionDescriptor(Phi)) { | ||||||
8518 | assert(II->getStartValue() ==(static_cast <bool> (II->getStartValue() == Phi-> getIncomingValueForBlock(OrigLoop->getLoopPreheader())) ? void (0) : __assert_fail ("II->getStartValue() == Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader())" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8519, __extension__ __PRETTY_FUNCTION__)) | ||||||
8519 | Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader()))(static_cast <bool> (II->getStartValue() == Phi-> getIncomingValueForBlock(OrigLoop->getLoopPreheader())) ? void (0) : __assert_fail ("II->getStartValue() == Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader())" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8519, __extension__ __PRETTY_FUNCTION__)); | ||||||
8520 | return new VPWidenIntOrFpInductionRecipe(Phi, Operands[0], *II); | ||||||
8521 | } | ||||||
8522 | |||||||
8523 | return nullptr; | ||||||
8524 | } | ||||||
8525 | |||||||
8526 | VPWidenIntOrFpInductionRecipe *VPRecipeBuilder::tryToOptimizeInductionTruncate( | ||||||
8527 | TruncInst *I, ArrayRef<VPValue *> Operands, VFRange &Range, | ||||||
8528 | VPlan &Plan) const { | ||||||
8529 | // Optimize the special case where the source is a constant integer | ||||||
8530 | // induction variable. Notice that we can only optimize the 'trunc' case | ||||||
8531 | // because (a) FP conversions lose precision, (b) sext/zext may wrap, and | ||||||
8532 | // (c) other casts depend on pointer size. | ||||||
8533 | |||||||
8534 | // Determine whether \p K is a truncation based on an induction variable that | ||||||
8535 | // can be optimized. | ||||||
8536 | auto isOptimizableIVTruncate = | ||||||
8537 | [&](Instruction *K) -> std::function<bool(ElementCount)> { | ||||||
8538 | return [=](ElementCount VF) -> bool { | ||||||
8539 | return CM.isOptimizableIVTruncate(K, VF); | ||||||
8540 | }; | ||||||
8541 | }; | ||||||
8542 | |||||||
8543 | if (LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||
8544 | isOptimizableIVTruncate(I), Range)) { | ||||||
8545 | |||||||
8546 | auto *Phi = cast<PHINode>(I->getOperand(0)); | ||||||
8547 | const InductionDescriptor &II = *Legal->getIntOrFpInductionDescriptor(Phi); | ||||||
8548 | VPValue *Start = Plan.getOrAddVPValue(II.getStartValue()); | ||||||
8549 | return new VPWidenIntOrFpInductionRecipe(Phi, Start, II, I); | ||||||
8550 | } | ||||||
8551 | return nullptr; | ||||||
8552 | } | ||||||
8553 | |||||||
8554 | VPRecipeOrVPValueTy VPRecipeBuilder::tryToBlend(PHINode *Phi, | ||||||
8555 | ArrayRef<VPValue *> Operands, | ||||||
8556 | VPlanPtr &Plan) { | ||||||
8557 | // If all incoming values are equal, the incoming VPValue can be used directly | ||||||
8558 | // instead of creating a new VPBlendRecipe. | ||||||
8559 | VPValue *FirstIncoming = Operands[0]; | ||||||
8560 | if (all_of(Operands, [FirstIncoming](const VPValue *Inc) { | ||||||
8561 | return FirstIncoming == Inc; | ||||||
8562 | })) { | ||||||
8563 | return Operands[0]; | ||||||
8564 | } | ||||||
8565 | |||||||
8566 | // We know that all PHIs in non-header blocks are converted into selects, so | ||||||
8567 | // we don't have to worry about the insertion order and we can just use the | ||||||
8568 | // builder. At this point we generate the predication tree. There may be | ||||||
8569 | // duplications since this is a simple recursive scan, but future | ||||||
8570 | // optimizations will clean it up. | ||||||
8571 | SmallVector<VPValue *, 2> OperandsWithMask; | ||||||
8572 | unsigned NumIncoming = Phi->getNumIncomingValues(); | ||||||
8573 | |||||||
8574 | for (unsigned In = 0; In < NumIncoming; In++) { | ||||||
8575 | VPValue *EdgeMask = | ||||||
8576 | createEdgeMask(Phi->getIncomingBlock(In), Phi->getParent(), Plan); | ||||||
8577 | 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", 8578, __extension__ __PRETTY_FUNCTION__)) | ||||||
8578 | "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", 8578, __extension__ __PRETTY_FUNCTION__)); | ||||||
8579 | OperandsWithMask.push_back(Operands[In]); | ||||||
8580 | if (EdgeMask) | ||||||
8581 | OperandsWithMask.push_back(EdgeMask); | ||||||
8582 | } | ||||||
8583 | return toVPRecipeResult(new VPBlendRecipe(Phi, OperandsWithMask)); | ||||||
8584 | } | ||||||
8585 | |||||||
8586 | VPWidenCallRecipe *VPRecipeBuilder::tryToWidenCall(CallInst *CI, | ||||||
8587 | ArrayRef<VPValue *> Operands, | ||||||
8588 | VFRange &Range) const { | ||||||
8589 | |||||||
8590 | bool IsPredicated = LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||
8591 | [this, CI](ElementCount VF) { | ||||||
8592 | return CM.isScalarWithPredication(CI, VF); | ||||||
8593 | }, | ||||||
8594 | Range); | ||||||
8595 | |||||||
8596 | if (IsPredicated) | ||||||
8597 | return nullptr; | ||||||
8598 | |||||||
8599 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||||
8600 | if (ID && (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end || | ||||||
8601 | ID == Intrinsic::lifetime_start || ID == Intrinsic::sideeffect || | ||||||
8602 | ID == Intrinsic::pseudoprobe || | ||||||
8603 | ID == Intrinsic::experimental_noalias_scope_decl)) | ||||||
8604 | return nullptr; | ||||||
8605 | |||||||
8606 | auto willWiden = [&](ElementCount VF) -> bool { | ||||||
8607 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||||
8608 | // The following case may be scalarized depending on the VF. | ||||||
8609 | // The flag shows whether we use Intrinsic or a usual Call for vectorized | ||||||
8610 | // version of the instruction. | ||||||
8611 | // Is it beneficial to perform intrinsic call compared to lib call? | ||||||
8612 | bool NeedToScalarize = false; | ||||||
8613 | InstructionCost CallCost = CM.getVectorCallCost(CI, VF, NeedToScalarize); | ||||||
8614 | InstructionCost IntrinsicCost = ID ? CM.getVectorIntrinsicCost(CI, VF) : 0; | ||||||
8615 | bool UseVectorIntrinsic = ID && IntrinsicCost <= CallCost; | ||||||
8616 | return UseVectorIntrinsic || !NeedToScalarize; | ||||||
8617 | }; | ||||||
8618 | |||||||
8619 | if (!LoopVectorizationPlanner::getDecisionAndClampRange(willWiden, Range)) | ||||||
8620 | return nullptr; | ||||||
8621 | |||||||
8622 | ArrayRef<VPValue *> Ops = Operands.take_front(CI->arg_size()); | ||||||
8623 | return new VPWidenCallRecipe(*CI, make_range(Ops.begin(), Ops.end())); | ||||||
8624 | } | ||||||
8625 | |||||||
8626 | bool VPRecipeBuilder::shouldWiden(Instruction *I, VFRange &Range) const { | ||||||
8627 | 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", 8628, __extension__ __PRETTY_FUNCTION__)) | ||||||
8628 | !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", 8628, __extension__ __PRETTY_FUNCTION__)); | ||||||
8629 | // Instruction should be widened, unless it is scalar after vectorization, | ||||||
8630 | // scalarization is profitable or it is predicated. | ||||||
8631 | auto WillScalarize = [this, I](ElementCount VF) -> bool { | ||||||
8632 | return CM.isScalarAfterVectorization(I, VF) || | ||||||
8633 | CM.isProfitableToScalarize(I, VF) || | ||||||
8634 | CM.isScalarWithPredication(I, VF); | ||||||
8635 | }; | ||||||
8636 | return !LoopVectorizationPlanner::getDecisionAndClampRange(WillScalarize, | ||||||
8637 | Range); | ||||||
8638 | } | ||||||
8639 | |||||||
8640 | VPWidenRecipe *VPRecipeBuilder::tryToWiden(Instruction *I, | ||||||
8641 | ArrayRef<VPValue *> Operands) const { | ||||||
8642 | auto IsVectorizableOpcode = [](unsigned Opcode) { | ||||||
8643 | switch (Opcode) { | ||||||
8644 | case Instruction::Add: | ||||||
8645 | case Instruction::And: | ||||||
8646 | case Instruction::AShr: | ||||||
8647 | case Instruction::BitCast: | ||||||
8648 | case Instruction::FAdd: | ||||||
8649 | case Instruction::FCmp: | ||||||
8650 | case Instruction::FDiv: | ||||||
8651 | case Instruction::FMul: | ||||||
8652 | case Instruction::FNeg: | ||||||
8653 | case Instruction::FPExt: | ||||||
8654 | case Instruction::FPToSI: | ||||||
8655 | case Instruction::FPToUI: | ||||||
8656 | case Instruction::FPTrunc: | ||||||
8657 | case Instruction::FRem: | ||||||
8658 | case Instruction::FSub: | ||||||
8659 | case Instruction::ICmp: | ||||||
8660 | case Instruction::IntToPtr: | ||||||
8661 | case Instruction::LShr: | ||||||
8662 | case Instruction::Mul: | ||||||
8663 | case Instruction::Or: | ||||||
8664 | case Instruction::PtrToInt: | ||||||
8665 | case Instruction::SDiv: | ||||||
8666 | case Instruction::Select: | ||||||
8667 | case Instruction::SExt: | ||||||
8668 | case Instruction::Shl: | ||||||
8669 | case Instruction::SIToFP: | ||||||
8670 | case Instruction::SRem: | ||||||
8671 | case Instruction::Sub: | ||||||
8672 | case Instruction::Trunc: | ||||||
8673 | case Instruction::UDiv: | ||||||
8674 | case Instruction::UIToFP: | ||||||
8675 | case Instruction::URem: | ||||||
8676 | case Instruction::Xor: | ||||||
8677 | case Instruction::ZExt: | ||||||
8678 | return true; | ||||||
8679 | } | ||||||
8680 | return false; | ||||||
8681 | }; | ||||||
8682 | |||||||
8683 | if (!IsVectorizableOpcode(I->getOpcode())) | ||||||
8684 | return nullptr; | ||||||
8685 | |||||||
8686 | // Success: widen this instruction. | ||||||
8687 | return new VPWidenRecipe(*I, make_range(Operands.begin(), Operands.end())); | ||||||
8688 | } | ||||||
8689 | |||||||
8690 | void VPRecipeBuilder::fixHeaderPhis() { | ||||||
8691 | BasicBlock *OrigLatch = OrigLoop->getLoopLatch(); | ||||||
8692 | for (VPHeaderPHIRecipe *R : PhisToFix) { | ||||||
8693 | auto *PN = cast<PHINode>(R->getUnderlyingValue()); | ||||||
8694 | VPRecipeBase *IncR = | ||||||
8695 | getRecipe(cast<Instruction>(PN->getIncomingValueForBlock(OrigLatch))); | ||||||
8696 | R->addOperand(IncR->getVPSingleValue()); | ||||||
8697 | } | ||||||
8698 | } | ||||||
8699 | |||||||
8700 | VPBasicBlock *VPRecipeBuilder::handleReplication( | ||||||
8701 | Instruction *I, VFRange &Range, VPBasicBlock *VPBB, | ||||||
8702 | VPlanPtr &Plan) { | ||||||
8703 | bool IsUniform = LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||
8704 | [&](ElementCount VF) { return CM.isUniformAfterVectorization(I, VF); }, | ||||||
8705 | Range); | ||||||
8706 | |||||||
8707 | bool IsPredicated = LoopVectorizationPlanner::getDecisionAndClampRange( | ||||||
8708 | [&](ElementCount VF) { return CM.isPredicatedInst(I, VF, IsUniform); }, | ||||||
8709 | Range); | ||||||
8710 | |||||||
8711 | // Even if the instruction is not marked as uniform, there are certain | ||||||
8712 | // intrinsic calls that can be effectively treated as such, so we check for | ||||||
8713 | // them here. Conservatively, we only do this for scalable vectors, since | ||||||
8714 | // for fixed-width VFs we can always fall back on full scalarization. | ||||||
8715 | if (!IsUniform && Range.Start.isScalable() && isa<IntrinsicInst>(I)) { | ||||||
8716 | switch (cast<IntrinsicInst>(I)->getIntrinsicID()) { | ||||||
8717 | case Intrinsic::assume: | ||||||
8718 | case Intrinsic::lifetime_start: | ||||||
8719 | case Intrinsic::lifetime_end: | ||||||
8720 | // For scalable vectors if one of the operands is variant then we still | ||||||
8721 | // want to mark as uniform, which will generate one instruction for just | ||||||
8722 | // the first lane of the vector. We can't scalarize the call in the same | ||||||
8723 | // way as for fixed-width vectors because we don't know how many lanes | ||||||
8724 | // there are. | ||||||
8725 | // | ||||||
8726 | // The reasons for doing it this way for scalable vectors are: | ||||||
8727 | // 1. For the assume intrinsic generating the instruction for the first | ||||||
8728 | // lane is still be better than not generating any at all. For | ||||||
8729 | // example, the input may be a splat across all lanes. | ||||||
8730 | // 2. For the lifetime start/end intrinsics the pointer operand only | ||||||
8731 | // does anything useful when the input comes from a stack object, | ||||||
8732 | // which suggests it should always be uniform. For non-stack objects | ||||||
8733 | // the effect is to poison the object, which still allows us to | ||||||
8734 | // remove the call. | ||||||
8735 | IsUniform = true; | ||||||
8736 | break; | ||||||
8737 | default: | ||||||
8738 | break; | ||||||
8739 | } | ||||||
8740 | } | ||||||
8741 | |||||||
8742 | auto *Recipe = new VPReplicateRecipe(I, Plan->mapToVPValues(I->operands()), | ||||||
8743 | IsUniform, IsPredicated); | ||||||
8744 | setRecipe(I, Recipe); | ||||||
8745 | Plan->addVPValue(I, Recipe); | ||||||
8746 | |||||||
8747 | // Find if I uses a predicated instruction. If so, it will use its scalar | ||||||
8748 | // value. Avoid hoisting the insert-element which packs the scalar value into | ||||||
8749 | // a vector value, as that happens iff all users use the vector value. | ||||||
8750 | for (VPValue *Op : Recipe->operands()) { | ||||||
8751 | auto *PredR = dyn_cast_or_null<VPPredInstPHIRecipe>(Op->getDef()); | ||||||
8752 | if (!PredR) | ||||||
8753 | continue; | ||||||
8754 | auto *RepR = | ||||||
8755 | cast_or_null<VPReplicateRecipe>(PredR->getOperand(0)->getDef()); | ||||||
8756 | 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", 8757, __extension__ __PRETTY_FUNCTION__)) | ||||||
8757 | "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", 8757, __extension__ __PRETTY_FUNCTION__)); | ||||||
8758 | RepR->setAlsoPack(false); | ||||||
8759 | } | ||||||
8760 | |||||||
8761 | // Finalize the recipe for Instr, first if it is not predicated. | ||||||
8762 | if (!IsPredicated) { | ||||||
8763 | LLVM_DEBUG(dbgs() << "LV: Scalarizing:" << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Scalarizing:" << *I << "\n"; } } while (false); | ||||||
8764 | VPBB->appendRecipe(Recipe); | ||||||
8765 | return VPBB; | ||||||
8766 | } | ||||||
8767 | 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); | ||||||
8768 | |||||||
8769 | VPBlockBase *SingleSucc = VPBB->getSingleSuccessor(); | ||||||
8770 | 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", 8771, __extension__ __PRETTY_FUNCTION__)) | ||||||
8771 | "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", 8771, __extension__ __PRETTY_FUNCTION__)); | ||||||
8772 | VPBlockUtils::disconnectBlocks(VPBB, SingleSucc); | ||||||
8773 | // Record predicated instructions for above packing optimizations. | ||||||
8774 | VPBlockBase *Region = createReplicateRegion(I, Recipe, Plan); | ||||||
8775 | VPBlockUtils::insertBlockAfter(Region, VPBB); | ||||||
8776 | auto *RegSucc = new VPBasicBlock(); | ||||||
8777 | VPBlockUtils::insertBlockAfter(RegSucc, Region); | ||||||
8778 | VPBlockUtils::connectBlocks(RegSucc, SingleSucc); | ||||||
8779 | return RegSucc; | ||||||
8780 | } | ||||||
8781 | |||||||
8782 | VPRegionBlock *VPRecipeBuilder::createReplicateRegion(Instruction *Instr, | ||||||
8783 | VPRecipeBase *PredRecipe, | ||||||
8784 | VPlanPtr &Plan) { | ||||||
8785 | // Instructions marked for predication are replicated and placed under an | ||||||
8786 | // if-then construct to prevent side-effects. | ||||||
8787 | |||||||
8788 | // Generate recipes to compute the block mask for this region. | ||||||
8789 | VPValue *BlockInMask = createBlockInMask(Instr->getParent(), Plan); | ||||||
8790 | |||||||
8791 | // Build the triangular if-then region. | ||||||
8792 | std::string RegionName = (Twine("pred.") + Instr->getOpcodeName()).str(); | ||||||
8793 | 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", 8793, __extension__ __PRETTY_FUNCTION__)); | ||||||
8794 | auto *BOMRecipe = new VPBranchOnMaskRecipe(BlockInMask); | ||||||
8795 | auto *Entry = new VPBasicBlock(Twine(RegionName) + ".entry", BOMRecipe); | ||||||
8796 | auto *PHIRecipe = Instr->getType()->isVoidTy() | ||||||
8797 | ? nullptr | ||||||
8798 | : new VPPredInstPHIRecipe(Plan->getOrAddVPValue(Instr)); | ||||||
8799 | if (PHIRecipe) { | ||||||
8800 | Plan->removeVPValueFor(Instr); | ||||||
8801 | Plan->addVPValue(Instr, PHIRecipe); | ||||||
8802 | } | ||||||
8803 | auto *Exit = new VPBasicBlock(Twine(RegionName) + ".continue", PHIRecipe); | ||||||
8804 | auto *Pred = new VPBasicBlock(Twine(RegionName) + ".if", PredRecipe); | ||||||
8805 | VPRegionBlock *Region = new VPRegionBlock(Entry, Exit, RegionName, true); | ||||||
8806 | |||||||
8807 | // Note: first set Entry as region entry and then connect successors starting | ||||||
8808 | // from it in order, to propagate the "parent" of each VPBasicBlock. | ||||||
8809 | VPBlockUtils::insertTwoBlocksAfter(Pred, Exit, BlockInMask, Entry); | ||||||
8810 | VPBlockUtils::connectBlocks(Pred, Exit); | ||||||
8811 | |||||||
8812 | return Region; | ||||||
8813 | } | ||||||
8814 | |||||||
8815 | VPRecipeOrVPValueTy | ||||||
8816 | VPRecipeBuilder::tryToCreateWidenRecipe(Instruction *Instr, | ||||||
8817 | ArrayRef<VPValue *> Operands, | ||||||
8818 | VFRange &Range, VPlanPtr &Plan) { | ||||||
8819 | // First, check for specific widening recipes that deal with calls, memory | ||||||
8820 | // operations, inductions and Phi nodes. | ||||||
8821 | if (auto *CI = dyn_cast<CallInst>(Instr)) | ||||||
8822 | return toVPRecipeResult(tryToWidenCall(CI, Operands, Range)); | ||||||
8823 | |||||||
8824 | if (isa<LoadInst>(Instr) || isa<StoreInst>(Instr)) | ||||||
8825 | return toVPRecipeResult(tryToWidenMemory(Instr, Operands, Range, Plan)); | ||||||
8826 | |||||||
8827 | VPRecipeBase *Recipe; | ||||||
8828 | if (auto Phi = dyn_cast<PHINode>(Instr)) { | ||||||
8829 | if (Phi->getParent() != OrigLoop->getHeader()) | ||||||
8830 | return tryToBlend(Phi, Operands, Plan); | ||||||
8831 | if ((Recipe = tryToOptimizeInductionPHI(Phi, Operands))) | ||||||
8832 | return toVPRecipeResult(Recipe); | ||||||
8833 | |||||||
8834 | VPHeaderPHIRecipe *PhiRecipe = nullptr; | ||||||
8835 | if (Legal->isReductionVariable(Phi) || Legal->isFirstOrderRecurrence(Phi)) { | ||||||
8836 | VPValue *StartV = Operands[0]; | ||||||
8837 | if (Legal->isReductionVariable(Phi)) { | ||||||
8838 | const RecurrenceDescriptor &RdxDesc = | ||||||
8839 | Legal->getReductionVars().find(Phi)->second; | ||||||
8840 | 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", 8841, __extension__ __PRETTY_FUNCTION__)) | ||||||
8841 | 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", 8841, __extension__ __PRETTY_FUNCTION__)); | ||||||
8842 | PhiRecipe = new VPReductionPHIRecipe(Phi, RdxDesc, *StartV, | ||||||
8843 | CM.isInLoopReduction(Phi), | ||||||
8844 | CM.useOrderedReductions(RdxDesc)); | ||||||
8845 | } else { | ||||||
8846 | PhiRecipe = new VPFirstOrderRecurrencePHIRecipe(Phi, *StartV); | ||||||
8847 | } | ||||||
8848 | |||||||
8849 | // Record the incoming value from the backedge, so we can add the incoming | ||||||
8850 | // value from the backedge after all recipes have been created. | ||||||
8851 | recordRecipeOf(cast<Instruction>( | ||||||
8852 | Phi->getIncomingValueForBlock(OrigLoop->getLoopLatch()))); | ||||||
8853 | PhisToFix.push_back(PhiRecipe); | ||||||
8854 | } else { | ||||||
8855 | // TODO: record backedge value for remaining pointer induction phis. | ||||||
8856 | assert(Phi->getType()->isPointerTy() &&(static_cast <bool> (Phi->getType()->isPointerTy( ) && "only pointer phis should be handled here") ? void (0) : __assert_fail ("Phi->getType()->isPointerTy() && \"only pointer phis should be handled here\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8857, __extension__ __PRETTY_FUNCTION__)) | ||||||
8857 | "only pointer phis should be handled here")(static_cast <bool> (Phi->getType()->isPointerTy( ) && "only pointer phis should be handled here") ? void (0) : __assert_fail ("Phi->getType()->isPointerTy() && \"only pointer phis should be handled here\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8857, __extension__ __PRETTY_FUNCTION__)); | ||||||
8858 | assert(Legal->getInductionVars().count(Phi) &&(static_cast <bool> (Legal->getInductionVars().count (Phi) && "Not an induction variable") ? void (0) : __assert_fail ("Legal->getInductionVars().count(Phi) && \"Not an induction variable\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8859, __extension__ __PRETTY_FUNCTION__)) | ||||||
8859 | "Not an induction variable")(static_cast <bool> (Legal->getInductionVars().count (Phi) && "Not an induction variable") ? void (0) : __assert_fail ("Legal->getInductionVars().count(Phi) && \"Not an induction variable\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8859, __extension__ __PRETTY_FUNCTION__)); | ||||||
8860 | InductionDescriptor II = Legal->getInductionVars().lookup(Phi); | ||||||
8861 | VPValue *Start = Plan->getOrAddVPValue(II.getStartValue()); | ||||||
8862 | PhiRecipe = new VPWidenPHIRecipe(Phi, Start); | ||||||
8863 | } | ||||||
8864 | |||||||
8865 | return toVPRecipeResult(PhiRecipe); | ||||||
8866 | } | ||||||
8867 | |||||||
8868 | if (isa<TruncInst>(Instr) && | ||||||
8869 | (Recipe = tryToOptimizeInductionTruncate(cast<TruncInst>(Instr), Operands, | ||||||
8870 | Range, *Plan))) | ||||||
8871 | return toVPRecipeResult(Recipe); | ||||||
8872 | |||||||
8873 | if (!shouldWiden(Instr, Range)) | ||||||
8874 | return nullptr; | ||||||
8875 | |||||||
8876 | if (auto GEP = dyn_cast<GetElementPtrInst>(Instr)) | ||||||
8877 | return toVPRecipeResult(new VPWidenGEPRecipe( | ||||||
8878 | GEP, make_range(Operands.begin(), Operands.end()), OrigLoop)); | ||||||
8879 | |||||||
8880 | if (auto *SI = dyn_cast<SelectInst>(Instr)) { | ||||||
8881 | bool InvariantCond = | ||||||
8882 | PSE.getSE()->isLoopInvariant(PSE.getSCEV(SI->getOperand(0)), OrigLoop); | ||||||
8883 | return toVPRecipeResult(new VPWidenSelectRecipe( | ||||||
8884 | *SI, make_range(Operands.begin(), Operands.end()), InvariantCond)); | ||||||
8885 | } | ||||||
8886 | |||||||
8887 | return toVPRecipeResult(tryToWiden(Instr, Operands)); | ||||||
8888 | } | ||||||
8889 | |||||||
8890 | void LoopVectorizationPlanner::buildVPlansWithVPRecipes(ElementCount MinVF, | ||||||
8891 | ElementCount MaxVF) { | ||||||
8892 | 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", 8892, __extension__ __PRETTY_FUNCTION__)); | ||||||
8893 | |||||||
8894 | // Collect instructions from the original loop that will become trivially dead | ||||||
8895 | // in the vectorized loop. We don't need to vectorize these instructions. For | ||||||
8896 | // example, original induction update instructions can become dead because we | ||||||
8897 | // separately emit induction "steps" when generating code for the new loop. | ||||||
8898 | // Similarly, we create a new latch condition when setting up the structure | ||||||
8899 | // of the new loop, so the old one can become dead. | ||||||
8900 | SmallPtrSet<Instruction *, 4> DeadInstructions; | ||||||
8901 | collectTriviallyDeadInstructions(DeadInstructions); | ||||||
8902 | |||||||
8903 | // Add assume instructions we need to drop to DeadInstructions, to prevent | ||||||
8904 | // them from being added to the VPlan. | ||||||
8905 | // TODO: We only need to drop assumes in blocks that get flattend. If the | ||||||
8906 | // control flow is preserved, we should keep them. | ||||||
8907 | auto &ConditionalAssumes = Legal->getConditionalAssumes(); | ||||||
8908 | DeadInstructions.insert(ConditionalAssumes.begin(), ConditionalAssumes.end()); | ||||||
8909 | |||||||
8910 | MapVector<Instruction *, Instruction *> &SinkAfter = Legal->getSinkAfter(); | ||||||
8911 | // Dead instructions do not need sinking. Remove them from SinkAfter. | ||||||
8912 | for (Instruction *I : DeadInstructions) | ||||||
8913 | SinkAfter.erase(I); | ||||||
8914 | |||||||
8915 | // Cannot sink instructions after dead instructions (there won't be any | ||||||
8916 | // recipes for them). Instead, find the first non-dead previous instruction. | ||||||
8917 | for (auto &P : Legal->getSinkAfter()) { | ||||||
8918 | Instruction *SinkTarget = P.second; | ||||||
8919 | Instruction *FirstInst = &*SinkTarget->getParent()->begin(); | ||||||
8920 | (void)FirstInst; | ||||||
8921 | while (DeadInstructions.contains(SinkTarget)) { | ||||||
8922 | assert((static_cast <bool> (SinkTarget != FirstInst && "Must find a live instruction (at least the one feeding the " "first-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 \" \"first-order recurrence PHI) before reaching beginning of the block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8925, __extension__ __PRETTY_FUNCTION__)) | ||||||
8923 | SinkTarget != FirstInst &&(static_cast <bool> (SinkTarget != FirstInst && "Must find a live instruction (at least the one feeding the " "first-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 \" \"first-order recurrence PHI) before reaching beginning of the block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8925, __extension__ __PRETTY_FUNCTION__)) | ||||||
8924 | "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 " "first-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 \" \"first-order recurrence PHI) before reaching beginning of the block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8925, __extension__ __PRETTY_FUNCTION__)) | ||||||
8925 | "first-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 " "first-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 \" \"first-order recurrence PHI) before reaching beginning of the block\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 8925, __extension__ __PRETTY_FUNCTION__)); | ||||||
8926 | SinkTarget = SinkTarget->getPrevNode(); | ||||||
8927 | 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", 8928, __extension__ __PRETTY_FUNCTION__)) | ||||||
8928 | "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", 8928, __extension__ __PRETTY_FUNCTION__)); | ||||||
8929 | } | ||||||
8930 | P.second = SinkTarget; | ||||||
8931 | } | ||||||
8932 | |||||||
8933 | auto MaxVFPlusOne = MaxVF.getWithIncrement(1); | ||||||
8934 | for (ElementCount VF = MinVF; ElementCount::isKnownLT(VF, MaxVFPlusOne);) { | ||||||
8935 | VFRange SubRange = {VF, MaxVFPlusOne}; | ||||||
8936 | VPlans.push_back( | ||||||
8937 | buildVPlanWithVPRecipes(SubRange, DeadInstructions, SinkAfter)); | ||||||
8938 | VF = SubRange.End; | ||||||
8939 | } | ||||||
8940 | } | ||||||
8941 | |||||||
8942 | // Add a VPCanonicalIVPHIRecipe starting at 0 to the header, a | ||||||
8943 | // CanonicalIVIncrement{NUW} VPInstruction to increment it by VF * UF and a | ||||||
8944 | // BranchOnCount VPInstruction to the latch. | ||||||
8945 | static void addCanonicalIVRecipes(VPlan &Plan, Type *IdxTy, DebugLoc DL, | ||||||
8946 | bool HasNUW, bool IsVPlanNative) { | ||||||
8947 | Value *StartIdx = ConstantInt::get(IdxTy, 0); | ||||||
8948 | auto *StartV = Plan.getOrAddVPValue(StartIdx); | ||||||
8949 | |||||||
8950 | auto *CanonicalIVPHI = new VPCanonicalIVPHIRecipe(StartV, DL); | ||||||
8951 | VPRegionBlock *TopRegion = Plan.getVectorLoopRegion(); | ||||||
8952 | VPBasicBlock *Header = TopRegion->getEntryBasicBlock(); | ||||||
8953 | if (IsVPlanNative) | ||||||
8954 | Header = cast<VPBasicBlock>(Header->getSingleSuccessor()); | ||||||
8955 | Header->insert(CanonicalIVPHI, Header->begin()); | ||||||
8956 | |||||||
8957 | auto *CanonicalIVIncrement = | ||||||
8958 | new VPInstruction(HasNUW ? VPInstruction::CanonicalIVIncrementNUW | ||||||
8959 | : VPInstruction::CanonicalIVIncrement, | ||||||
8960 | {CanonicalIVPHI}, DL); | ||||||
8961 | CanonicalIVPHI->addOperand(CanonicalIVIncrement); | ||||||
8962 | |||||||
8963 | VPBasicBlock *EB = TopRegion->getExitBasicBlock(); | ||||||
8964 | if (IsVPlanNative) { | ||||||
8965 | EB = cast<VPBasicBlock>(EB->getSinglePredecessor()); | ||||||
8966 | EB->setCondBit(nullptr); | ||||||
8967 | } | ||||||
8968 | EB->appendRecipe(CanonicalIVIncrement); | ||||||
8969 | |||||||
8970 | auto *BranchOnCount = | ||||||
8971 | new VPInstruction(VPInstruction::BranchOnCount, | ||||||
8972 | {CanonicalIVIncrement, &Plan.getVectorTripCount()}, DL); | ||||||
8973 | EB->appendRecipe(BranchOnCount); | ||||||
8974 | } | ||||||
8975 | |||||||
8976 | VPlanPtr LoopVectorizationPlanner::buildVPlanWithVPRecipes( | ||||||
8977 | VFRange &Range, SmallPtrSetImpl<Instruction *> &DeadInstructions, | ||||||
8978 | const MapVector<Instruction *, Instruction *> &SinkAfter) { | ||||||
8979 | |||||||
8980 | SmallPtrSet<const InterleaveGroup<Instruction> *, 1> InterleaveGroups; | ||||||
8981 | |||||||
8982 | VPRecipeBuilder RecipeBuilder(OrigLoop, TLI, Legal, CM, PSE, Builder); | ||||||
8983 | |||||||
8984 | // --------------------------------------------------------------------------- | ||||||
8985 | // Pre-construction: record ingredients whose recipes we'll need to further | ||||||
8986 | // process after constructing the initial VPlan. | ||||||
8987 | // --------------------------------------------------------------------------- | ||||||
8988 | |||||||
8989 | // Mark instructions we'll need to sink later and their targets as | ||||||
8990 | // ingredients whose recipe we'll need to record. | ||||||
8991 | for (auto &Entry : SinkAfter) { | ||||||
8992 | RecipeBuilder.recordRecipeOf(Entry.first); | ||||||
8993 | RecipeBuilder.recordRecipeOf(Entry.second); | ||||||
8994 | } | ||||||
8995 | for (auto &Reduction : CM.getInLoopReductionChains()) { | ||||||
| |||||||
8996 | PHINode *Phi = Reduction.first; | ||||||
8997 | RecurKind Kind = | ||||||
8998 | Legal->getReductionVars().find(Phi)->second.getRecurrenceKind(); | ||||||
8999 | const SmallVector<Instruction *, 4> &ReductionOperations = Reduction.second; | ||||||
9000 | |||||||
9001 | RecipeBuilder.recordRecipeOf(Phi); | ||||||
9002 | for (auto &R : ReductionOperations) { | ||||||
9003 | RecipeBuilder.recordRecipeOf(R); | ||||||
9004 | // For min/max reducitons, where we have a pair of icmp/select, we also | ||||||
9005 | // need to record the ICmp recipe, so it can be removed later. | ||||||
9006 | 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", 9007, __extension__ __PRETTY_FUNCTION__)) | ||||||
9007 | "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", 9007, __extension__ __PRETTY_FUNCTION__)); | ||||||
9008 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) | ||||||
9009 | RecipeBuilder.recordRecipeOf(cast<Instruction>(R->getOperand(0))); | ||||||
9010 | } | ||||||
9011 | } | ||||||
9012 | |||||||
9013 | // For each interleave group which is relevant for this (possibly trimmed) | ||||||
9014 | // Range, add it to the set of groups to be later applied to the VPlan and add | ||||||
9015 | // placeholders for its members' Recipes which we'll be replacing with a | ||||||
9016 | // single VPInterleaveRecipe. | ||||||
9017 | for (InterleaveGroup<Instruction> *IG : IAI.getInterleaveGroups()) { | ||||||
9018 | auto applyIG = [IG, this](ElementCount VF) -> bool { | ||||||
9019 | return (VF.isVector() && // Query is illegal for VF == 1 | ||||||
9020 | CM.getWideningDecision(IG->getInsertPos(), VF) == | ||||||
9021 | LoopVectorizationCostModel::CM_Interleave); | ||||||
9022 | }; | ||||||
9023 | if (!getDecisionAndClampRange(applyIG, Range)) | ||||||
9024 | continue; | ||||||
9025 | InterleaveGroups.insert(IG); | ||||||
9026 | for (unsigned i = 0; i < IG->getFactor(); i++) | ||||||
9027 | if (Instruction *Member = IG->getMember(i)) | ||||||
9028 | RecipeBuilder.recordRecipeOf(Member); | ||||||
9029 | }; | ||||||
9030 | |||||||
9031 | // --------------------------------------------------------------------------- | ||||||
9032 | // Build initial VPlan: Scan the body of the loop in a topological order to | ||||||
9033 | // visit each basic block after having visited its predecessor basic blocks. | ||||||
9034 | // --------------------------------------------------------------------------- | ||||||
9035 | |||||||
9036 | // Create initial VPlan skeleton, with separate header and latch blocks. | ||||||
9037 | VPBasicBlock *HeaderVPBB = new VPBasicBlock(); | ||||||
9038 | VPBasicBlock *LatchVPBB = new VPBasicBlock("vector.latch"); | ||||||
9039 | VPBlockUtils::insertBlockAfter(LatchVPBB, HeaderVPBB); | ||||||
9040 | auto *TopRegion = new VPRegionBlock(HeaderVPBB, LatchVPBB, "vector loop"); | ||||||
9041 | auto Plan = std::make_unique<VPlan>(TopRegion); | ||||||
9042 | |||||||
9043 | Instruction *DLInst = | ||||||
9044 | getDebugLocFromInstOrOperands(Legal->getPrimaryInduction()); | ||||||
9045 | addCanonicalIVRecipes(*Plan, Legal->getWidestInductionType(), | ||||||
9046 | DLInst
| ||||||
9047 | !CM.foldTailByMasking(), false); | ||||||
9048 | |||||||
9049 | // Scan the body of the loop in a topological order to visit each basic block | ||||||
9050 | // after having visited its predecessor basic blocks. | ||||||
9051 | LoopBlocksDFS DFS(OrigLoop); | ||||||
9052 | DFS.perform(LI); | ||||||
9053 | |||||||
9054 | VPBasicBlock *VPBB = HeaderVPBB; | ||||||
9055 | SmallVector<VPWidenIntOrFpInductionRecipe *> InductionsToMove; | ||||||
9056 | for (BasicBlock *BB : make_range(DFS.beginRPO(), DFS.endRPO())) { | ||||||
9057 | // Relevant instructions from basic block BB will be grouped into VPRecipe | ||||||
9058 | // ingredients and fill a new VPBasicBlock. | ||||||
9059 | unsigned VPBBsForBB = 0; | ||||||
9060 | VPBB->setName(BB->getName()); | ||||||
9061 | Builder.setInsertPoint(VPBB); | ||||||
9062 | |||||||
9063 | // Introduce each ingredient into VPlan. | ||||||
9064 | // TODO: Model and preserve debug instrinsics in VPlan. | ||||||
9065 | for (Instruction &I : BB->instructionsWithoutDebug()) { | ||||||
9066 | Instruction *Instr = &I; | ||||||
9067 | |||||||
9068 | // First filter out irrelevant instructions, to ensure no recipes are | ||||||
9069 | // built for them. | ||||||
9070 | if (isa<BranchInst>(Instr) || DeadInstructions.count(Instr)) | ||||||
9071 | continue; | ||||||
9072 | |||||||
9073 | SmallVector<VPValue *, 4> Operands; | ||||||
9074 | auto *Phi = dyn_cast<PHINode>(Instr); | ||||||
9075 | if (Phi && Phi->getParent() == OrigLoop->getHeader()) { | ||||||
9076 | Operands.push_back(Plan->getOrAddVPValue( | ||||||
9077 | Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader()))); | ||||||
9078 | } else { | ||||||
9079 | auto OpRange = Plan->mapToVPValues(Instr->operands()); | ||||||
9080 | Operands = {OpRange.begin(), OpRange.end()}; | ||||||
9081 | } | ||||||
9082 | if (auto RecipeOrValue = RecipeBuilder.tryToCreateWidenRecipe( | ||||||
9083 | Instr, Operands, Range, Plan)) { | ||||||
9084 | // If Instr can be simplified to an existing VPValue, use it. | ||||||
9085 | if (RecipeOrValue.is<VPValue *>()) { | ||||||
9086 | auto *VPV = RecipeOrValue.get<VPValue *>(); | ||||||
9087 | Plan->addVPValue(Instr, VPV); | ||||||
9088 | // If the re-used value is a recipe, register the recipe for the | ||||||
9089 | // instruction, in case the recipe for Instr needs to be recorded. | ||||||
9090 | if (auto *R = dyn_cast_or_null<VPRecipeBase>(VPV->getDef())) | ||||||
9091 | RecipeBuilder.setRecipe(Instr, R); | ||||||
9092 | continue; | ||||||
9093 | } | ||||||
9094 | // Otherwise, add the new recipe. | ||||||
9095 | VPRecipeBase *Recipe = RecipeOrValue.get<VPRecipeBase *>(); | ||||||
9096 | for (auto *Def : Recipe->definedValues()) { | ||||||
9097 | auto *UV = Def->getUnderlyingValue(); | ||||||
9098 | Plan->addVPValue(UV, Def); | ||||||
9099 | } | ||||||
9100 | |||||||
9101 | if (isa<VPWidenIntOrFpInductionRecipe>(Recipe) && | ||||||
9102 | HeaderVPBB->getFirstNonPhi() != VPBB->end()) { | ||||||
9103 | // Keep track of VPWidenIntOrFpInductionRecipes not in the phi section | ||||||
9104 | // of the header block. That can happen for truncates of induction | ||||||
9105 | // variables. Those recipes are moved to the phi section of the header | ||||||
9106 | // block after applying SinkAfter, which relies on the original | ||||||
9107 | // position of the trunc. | ||||||
9108 | assert(isa<TruncInst>(Instr))(static_cast <bool> (isa<TruncInst>(Instr)) ? void (0) : __assert_fail ("isa<TruncInst>(Instr)", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 9108, __extension__ __PRETTY_FUNCTION__)); | ||||||
9109 | InductionsToMove.push_back( | ||||||
9110 | cast<VPWidenIntOrFpInductionRecipe>(Recipe)); | ||||||
9111 | } | ||||||
9112 | RecipeBuilder.setRecipe(Instr, Recipe); | ||||||
9113 | VPBB->appendRecipe(Recipe); | ||||||
9114 | continue; | ||||||
9115 | } | ||||||
9116 | |||||||
9117 | // Otherwise, if all widening options failed, Instruction is to be | ||||||
9118 | // replicated. This may create a successor for VPBB. | ||||||
9119 | VPBasicBlock *NextVPBB = | ||||||
9120 | RecipeBuilder.handleReplication(Instr, Range, VPBB, Plan); | ||||||
9121 | if (NextVPBB != VPBB) { | ||||||
9122 | VPBB = NextVPBB; | ||||||
9123 | VPBB->setName(BB->hasName() ? BB->getName() + "." + Twine(VPBBsForBB++) | ||||||
9124 | : ""); | ||||||
9125 | } | ||||||
9126 | } | ||||||
9127 | |||||||
9128 | VPBlockUtils::insertBlockAfter(new VPBasicBlock(), VPBB); | ||||||
9129 | VPBB = cast<VPBasicBlock>(VPBB->getSingleSuccessor()); | ||||||
9130 | } | ||||||
9131 | |||||||
9132 | // Fold the last, empty block into its predecessor. | ||||||
9133 | VPBB = VPBlockUtils::tryToMergeBlockIntoPredecessor(VPBB); | ||||||
9134 | 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", 9134, __extension__ __PRETTY_FUNCTION__)); | ||||||
9135 | // After here, VPBB should not be used. | ||||||
9136 | VPBB = nullptr; | ||||||
9137 | |||||||
9138 | assert(isa<VPRegionBlock>(Plan->getEntry()) &&(static_cast <bool> (isa<VPRegionBlock>(Plan-> getEntry()) && !Plan->getEntry()->getEntryBasicBlock ()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock") ? void (0) : __assert_fail ("isa<VPRegionBlock>(Plan->getEntry()) && !Plan->getEntry()->getEntryBasicBlock()->empty() && \"entry block must be set to a VPRegionBlock having a non-empty entry \" \"VPBasicBlock\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9141, __extension__ __PRETTY_FUNCTION__)) | ||||||
9139 | !Plan->getEntry()->getEntryBasicBlock()->empty() &&(static_cast <bool> (isa<VPRegionBlock>(Plan-> getEntry()) && !Plan->getEntry()->getEntryBasicBlock ()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock") ? void (0) : __assert_fail ("isa<VPRegionBlock>(Plan->getEntry()) && !Plan->getEntry()->getEntryBasicBlock()->empty() && \"entry block must be set to a VPRegionBlock having a non-empty entry \" \"VPBasicBlock\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9141, __extension__ __PRETTY_FUNCTION__)) | ||||||
9140 | "entry block must be set to a VPRegionBlock having a non-empty entry "(static_cast <bool> (isa<VPRegionBlock>(Plan-> getEntry()) && !Plan->getEntry()->getEntryBasicBlock ()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock") ? void (0) : __assert_fail ("isa<VPRegionBlock>(Plan->getEntry()) && !Plan->getEntry()->getEntryBasicBlock()->empty() && \"entry block must be set to a VPRegionBlock having a non-empty entry \" \"VPBasicBlock\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9141, __extension__ __PRETTY_FUNCTION__)) | ||||||
9141 | "VPBasicBlock")(static_cast <bool> (isa<VPRegionBlock>(Plan-> getEntry()) && !Plan->getEntry()->getEntryBasicBlock ()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock") ? void (0) : __assert_fail ("isa<VPRegionBlock>(Plan->getEntry()) && !Plan->getEntry()->getEntryBasicBlock()->empty() && \"entry block must be set to a VPRegionBlock having a non-empty entry \" \"VPBasicBlock\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9141, __extension__ __PRETTY_FUNCTION__)); | ||||||
9142 | RecipeBuilder.fixHeaderPhis(); | ||||||
9143 | |||||||
9144 | // --------------------------------------------------------------------------- | ||||||
9145 | // Transform initial VPlan: Apply previously taken decisions, in order, to | ||||||
9146 | // bring the VPlan to its final state. | ||||||
9147 | // --------------------------------------------------------------------------- | ||||||
9148 | |||||||
9149 | // Apply Sink-After legal constraints. | ||||||
9150 | auto GetReplicateRegion = [](VPRecipeBase *R) -> VPRegionBlock * { | ||||||
9151 | auto *Region = dyn_cast_or_null<VPRegionBlock>(R->getParent()->getParent()); | ||||||
9152 | if (Region && Region->isReplicator()) { | ||||||
9153 | 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", 9154, __extension__ __PRETTY_FUNCTION__)) | ||||||
9154 | 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", 9154, __extension__ __PRETTY_FUNCTION__)); | ||||||
9155 | 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", 9157, __extension__ __PRETTY_FUNCTION__)) | ||||||
9156 | "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", 9157, __extension__ __PRETTY_FUNCTION__)) | ||||||
9157 | "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", 9157, __extension__ __PRETTY_FUNCTION__)); | ||||||
9158 | return Region; | ||||||
9159 | } | ||||||
9160 | return nullptr; | ||||||
9161 | }; | ||||||
9162 | for (auto &Entry : SinkAfter) { | ||||||
9163 | VPRecipeBase *Sink = RecipeBuilder.getRecipe(Entry.first); | ||||||
9164 | VPRecipeBase *Target = RecipeBuilder.getRecipe(Entry.second); | ||||||
9165 | |||||||
9166 | auto *TargetRegion = GetReplicateRegion(Target); | ||||||
9167 | auto *SinkRegion = GetReplicateRegion(Sink); | ||||||
9168 | if (!SinkRegion) { | ||||||
9169 | // If the sink source is not a replicate region, sink the recipe directly. | ||||||
9170 | if (TargetRegion) { | ||||||
9171 | // The target is in a replication region, make sure to move Sink to | ||||||
9172 | // the block after it, not into the replication region itself. | ||||||
9173 | VPBasicBlock *NextBlock = | ||||||
9174 | cast<VPBasicBlock>(TargetRegion->getSuccessors().front()); | ||||||
9175 | Sink->moveBefore(*NextBlock, NextBlock->getFirstNonPhi()); | ||||||
9176 | } else | ||||||
9177 | Sink->moveAfter(Target); | ||||||
9178 | continue; | ||||||
9179 | } | ||||||
9180 | |||||||
9181 | // The sink source is in a replicate region. Unhook the region from the CFG. | ||||||
9182 | auto *SinkPred = SinkRegion->getSinglePredecessor(); | ||||||
9183 | auto *SinkSucc = SinkRegion->getSingleSuccessor(); | ||||||
9184 | VPBlockUtils::disconnectBlocks(SinkPred, SinkRegion); | ||||||
9185 | VPBlockUtils::disconnectBlocks(SinkRegion, SinkSucc); | ||||||
9186 | VPBlockUtils::connectBlocks(SinkPred, SinkSucc); | ||||||
9187 | |||||||
9188 | if (TargetRegion) { | ||||||
9189 | // The target recipe is also in a replicate region, move the sink region | ||||||
9190 | // after the target region. | ||||||
9191 | auto *TargetSucc = TargetRegion->getSingleSuccessor(); | ||||||
9192 | VPBlockUtils::disconnectBlocks(TargetRegion, TargetSucc); | ||||||
9193 | VPBlockUtils::connectBlocks(TargetRegion, SinkRegion); | ||||||
9194 | VPBlockUtils::connectBlocks(SinkRegion, TargetSucc); | ||||||
9195 | } else { | ||||||
9196 | // The sink source is in a replicate region, we need to move the whole | ||||||
9197 | // replicate region, which should only contain a single recipe in the | ||||||
9198 | // main block. | ||||||
9199 | auto *SplitBlock = | ||||||
9200 | Target->getParent()->splitAt(std::next(Target->getIterator())); | ||||||
9201 | |||||||
9202 | auto *SplitPred = SplitBlock->getSinglePredecessor(); | ||||||
9203 | |||||||
9204 | VPBlockUtils::disconnectBlocks(SplitPred, SplitBlock); | ||||||
9205 | VPBlockUtils::connectBlocks(SplitPred, SinkRegion); | ||||||
9206 | VPBlockUtils::connectBlocks(SinkRegion, SplitBlock); | ||||||
9207 | } | ||||||
9208 | } | ||||||
9209 | |||||||
9210 | VPlanTransforms::removeRedundantInductionCasts(*Plan); | ||||||
9211 | |||||||
9212 | // Now that sink-after is done, move induction recipes for optimized truncates | ||||||
9213 | // to the phi section of the header block. | ||||||
9214 | for (VPWidenIntOrFpInductionRecipe *Ind : InductionsToMove) | ||||||
9215 | Ind->moveBefore(*HeaderVPBB, HeaderVPBB->getFirstNonPhi()); | ||||||
| |||||||
9216 | |||||||
9217 | // Adjust the recipes for any inloop reductions. | ||||||
9218 | adjustRecipesForReductions(cast<VPBasicBlock>(TopRegion->getExit()), Plan, | ||||||
9219 | RecipeBuilder, Range.Start); | ||||||
9220 | |||||||
9221 | // Introduce a recipe to combine the incoming and previous values of a | ||||||
9222 | // first-order recurrence. | ||||||
9223 | for (VPRecipeBase &R : Plan->getEntry()->getEntryBasicBlock()->phis()) { | ||||||
9224 | auto *RecurPhi = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&R); | ||||||
9225 | if (!RecurPhi) | ||||||
9226 | continue; | ||||||
9227 | |||||||
9228 | VPRecipeBase *PrevRecipe = RecurPhi->getBackedgeRecipe(); | ||||||
9229 | VPBasicBlock *InsertBlock = PrevRecipe->getParent(); | ||||||
9230 | auto *Region = GetReplicateRegion(PrevRecipe); | ||||||
9231 | if (Region) | ||||||
9232 | InsertBlock = cast<VPBasicBlock>(Region->getSingleSuccessor()); | ||||||
9233 | if (Region || PrevRecipe->isPhi()) | ||||||
9234 | Builder.setInsertPoint(InsertBlock, InsertBlock->getFirstNonPhi()); | ||||||
9235 | else | ||||||
9236 | Builder.setInsertPoint(InsertBlock, std::next(PrevRecipe->getIterator())); | ||||||
9237 | |||||||
9238 | auto *RecurSplice = cast<VPInstruction>( | ||||||
9239 | Builder.createNaryOp(VPInstruction::FirstOrderRecurrenceSplice, | ||||||
9240 | {RecurPhi, RecurPhi->getBackedgeValue()})); | ||||||
9241 | |||||||
9242 | RecurPhi->replaceAllUsesWith(RecurSplice); | ||||||
9243 | // Set the first operand of RecurSplice to RecurPhi again, after replacing | ||||||
9244 | // all users. | ||||||
9245 | RecurSplice->setOperand(0, RecurPhi); | ||||||
9246 | } | ||||||
9247 | |||||||
9248 | // Interleave memory: for each Interleave Group we marked earlier as relevant | ||||||
9249 | // for this VPlan, replace the Recipes widening its memory instructions with a | ||||||
9250 | // single VPInterleaveRecipe at its insertion point. | ||||||
9251 | for (auto IG : InterleaveGroups) { | ||||||
9252 | auto *Recipe = cast<VPWidenMemoryInstructionRecipe>( | ||||||
9253 | RecipeBuilder.getRecipe(IG->getInsertPos())); | ||||||
9254 | SmallVector<VPValue *, 4> StoredValues; | ||||||
9255 | for (unsigned i = 0; i < IG->getFactor(); ++i) | ||||||
9256 | if (auto *SI = dyn_cast_or_null<StoreInst>(IG->getMember(i))) { | ||||||
9257 | auto *StoreR = | ||||||
9258 | cast<VPWidenMemoryInstructionRecipe>(RecipeBuilder.getRecipe(SI)); | ||||||
9259 | StoredValues.push_back(StoreR->getStoredValue()); | ||||||
9260 | } | ||||||
9261 | |||||||
9262 | auto *VPIG = new VPInterleaveRecipe(IG, Recipe->getAddr(), StoredValues, | ||||||
9263 | Recipe->getMask()); | ||||||
9264 | VPIG->insertBefore(Recipe); | ||||||
9265 | unsigned J = 0; | ||||||
9266 | for (unsigned i = 0; i < IG->getFactor(); ++i) | ||||||
9267 | if (Instruction *Member = IG->getMember(i)) { | ||||||
9268 | if (!Member->getType()->isVoidTy()) { | ||||||
9269 | VPValue *OriginalV = Plan->getVPValue(Member); | ||||||
9270 | Plan->removeVPValueFor(Member); | ||||||
9271 | Plan->addVPValue(Member, VPIG->getVPValue(J)); | ||||||
9272 | OriginalV->replaceAllUsesWith(VPIG->getVPValue(J)); | ||||||
9273 | J++; | ||||||
9274 | } | ||||||
9275 | RecipeBuilder.getRecipe(Member)->eraseFromParent(); | ||||||
9276 | } | ||||||
9277 | } | ||||||
9278 | |||||||
9279 | // From this point onwards, VPlan-to-VPlan transformations may change the plan | ||||||
9280 | // in ways that accessing values using original IR values is incorrect. | ||||||
9281 | Plan->disableValue2VPValue(); | ||||||
9282 | |||||||
9283 | VPlanTransforms::sinkScalarOperands(*Plan); | ||||||
9284 | VPlanTransforms::mergeReplicateRegions(*Plan); | ||||||
9285 | |||||||
9286 | std::string PlanName; | ||||||
9287 | raw_string_ostream RSO(PlanName); | ||||||
9288 | ElementCount VF = Range.Start; | ||||||
9289 | Plan->addVF(VF); | ||||||
9290 | RSO << "Initial VPlan for VF={" << VF; | ||||||
9291 | for (VF *= 2; ElementCount::isKnownLT(VF, Range.End); VF *= 2) { | ||||||
9292 | Plan->addVF(VF); | ||||||
9293 | RSO << "," << VF; | ||||||
9294 | } | ||||||
9295 | RSO << "},UF>=1"; | ||||||
9296 | RSO.flush(); | ||||||
9297 | Plan->setName(PlanName); | ||||||
9298 | |||||||
9299 | // Fold Exit block into its predecessor if possible. | ||||||
9300 | // TODO: Fold block earlier once all VPlan transforms properly maintain a | ||||||
9301 | // VPBasicBlock as exit. | ||||||
9302 | VPBlockUtils::tryToMergeBlockIntoPredecessor(TopRegion->getExit()); | ||||||
9303 | |||||||
9304 | 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", 9304, __extension__ __PRETTY_FUNCTION__)); | ||||||
9305 | return Plan; | ||||||
9306 | } | ||||||
9307 | |||||||
9308 | VPlanPtr LoopVectorizationPlanner::buildVPlan(VFRange &Range) { | ||||||
9309 | // Outer loop handling: They may require CFG and instruction level | ||||||
9310 | // transformations before even evaluating whether vectorization is profitable. | ||||||
9311 | // Since we cannot modify the incoming IR, we need to build VPlan upfront in | ||||||
9312 | // the vectorization pipeline. | ||||||
9313 | assert(!OrigLoop->isInnermost())(static_cast <bool> (!OrigLoop->isInnermost()) ? void (0) : __assert_fail ("!OrigLoop->isInnermost()", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 9313, __extension__ __PRETTY_FUNCTION__)); | ||||||
9314 | 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", 9314, __extension__ __PRETTY_FUNCTION__)); | ||||||
9315 | |||||||
9316 | // Create new empty VPlan | ||||||
9317 | auto Plan = std::make_unique<VPlan>(); | ||||||
9318 | |||||||
9319 | // Build hierarchical CFG | ||||||
9320 | VPlanHCFGBuilder HCFGBuilder(OrigLoop, LI, *Plan); | ||||||
9321 | HCFGBuilder.buildHierarchicalCFG(); | ||||||
9322 | |||||||
9323 | for (ElementCount VF = Range.Start; ElementCount::isKnownLT(VF, Range.End); | ||||||
9324 | VF *= 2) | ||||||
9325 | Plan->addVF(VF); | ||||||
9326 | |||||||
9327 | if (EnableVPlanPredication) { | ||||||
9328 | VPlanPredicator VPP(*Plan); | ||||||
9329 | VPP.predicate(); | ||||||
9330 | |||||||
9331 | // Avoid running transformation to recipes until masked code generation in | ||||||
9332 | // VPlan-native path is in place. | ||||||
9333 | return Plan; | ||||||
9334 | } | ||||||
9335 | |||||||
9336 | SmallPtrSet<Instruction *, 1> DeadInstructions; | ||||||
9337 | VPlanTransforms::VPInstructionsToVPRecipes( | ||||||
9338 | OrigLoop, Plan, | ||||||
9339 | [this](PHINode *P) { return Legal->getIntOrFpInductionDescriptor(P); }, | ||||||
9340 | DeadInstructions, *PSE.getSE()); | ||||||
9341 | |||||||
9342 | addCanonicalIVRecipes(*Plan, Legal->getWidestInductionType(), DebugLoc(), | ||||||
9343 | true, true); | ||||||
9344 | return Plan; | ||||||
9345 | } | ||||||
9346 | |||||||
9347 | // Adjust the recipes for reductions. For in-loop reductions the chain of | ||||||
9348 | // instructions leading from the loop exit instr to the phi need to be converted | ||||||
9349 | // to reductions, with one operand being vector and the other being the scalar | ||||||
9350 | // reduction chain. For other reductions, a select is introduced between the phi | ||||||
9351 | // and live-out recipes when folding the tail. | ||||||
9352 | void LoopVectorizationPlanner::adjustRecipesForReductions( | ||||||
9353 | VPBasicBlock *LatchVPBB, VPlanPtr &Plan, VPRecipeBuilder &RecipeBuilder, | ||||||
9354 | ElementCount MinVF) { | ||||||
9355 | for (auto &Reduction : CM.getInLoopReductionChains()) { | ||||||
9356 | PHINode *Phi = Reduction.first; | ||||||
9357 | const RecurrenceDescriptor &RdxDesc = | ||||||
9358 | Legal->getReductionVars().find(Phi)->second; | ||||||
9359 | const SmallVector<Instruction *, 4> &ReductionOperations = Reduction.second; | ||||||
9360 | |||||||
9361 | if (MinVF.isScalar() && !CM.useOrderedReductions(RdxDesc)) | ||||||
9362 | continue; | ||||||
9363 | |||||||
9364 | // ReductionOperations are orders top-down from the phi's use to the | ||||||
9365 | // LoopExitValue. We keep a track of the previous item (the Chain) to tell | ||||||
9366 | // which of the two operands will remain scalar and which will be reduced. | ||||||
9367 | // For minmax the chain will be the select instructions. | ||||||
9368 | Instruction *Chain = Phi; | ||||||
9369 | for (Instruction *R : ReductionOperations) { | ||||||
9370 | VPRecipeBase *WidenRecipe = RecipeBuilder.getRecipe(R); | ||||||
9371 | RecurKind Kind = RdxDesc.getRecurrenceKind(); | ||||||
9372 | |||||||
9373 | VPValue *ChainOp = Plan->getVPValue(Chain); | ||||||
9374 | unsigned FirstOpId; | ||||||
9375 | 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", 9376, __extension__ __PRETTY_FUNCTION__)) | ||||||
9376 | "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", 9376, __extension__ __PRETTY_FUNCTION__)); | ||||||
9377 | // Recognize a call to the llvm.fmuladd intrinsic. | ||||||
9378 | bool IsFMulAdd = (Kind == RecurKind::FMulAdd); | ||||||
9379 | 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", 9380, __extension__ __PRETTY_FUNCTION__)) | ||||||
9380 | "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", 9380, __extension__ __PRETTY_FUNCTION__)); | ||||||
9381 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { | ||||||
9382 | 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", 9383, __extension__ __PRETTY_FUNCTION__)) | ||||||
9383 | "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", 9383, __extension__ __PRETTY_FUNCTION__)); | ||||||
9384 | FirstOpId = 1; | ||||||
9385 | } else { | ||||||
9386 | 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", 9388, __extension__ __PRETTY_FUNCTION__)) | ||||||
9387 | (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", 9388, __extension__ __PRETTY_FUNCTION__)) | ||||||
9388 | "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", 9388, __extension__ __PRETTY_FUNCTION__)); | ||||||
9389 | FirstOpId = 0; | ||||||
9390 | } | ||||||
9391 | unsigned VecOpId = | ||||||
9392 | R->getOperand(FirstOpId) == Chain ? FirstOpId + 1 : FirstOpId; | ||||||
9393 | VPValue *VecOp = Plan->getVPValue(R->getOperand(VecOpId)); | ||||||
9394 | |||||||
9395 | auto *CondOp = CM.foldTailByMasking() | ||||||
9396 | ? RecipeBuilder.createBlockInMask(R->getParent(), Plan) | ||||||
9397 | : nullptr; | ||||||
9398 | |||||||
9399 | if (IsFMulAdd) { | ||||||
9400 | // If the instruction is a call to the llvm.fmuladd intrinsic then we | ||||||
9401 | // need to create an fmul recipe to use as the vector operand for the | ||||||
9402 | // fadd reduction. | ||||||
9403 | VPInstruction *FMulRecipe = new VPInstruction( | ||||||
9404 | Instruction::FMul, {VecOp, Plan->getVPValue(R->getOperand(1))}); | ||||||
9405 | FMulRecipe->setFastMathFlags(R->getFastMathFlags()); | ||||||
9406 | WidenRecipe->getParent()->insert(FMulRecipe, | ||||||
9407 | WidenRecipe->getIterator()); | ||||||
9408 | VecOp = FMulRecipe; | ||||||
9409 | } | ||||||
9410 | VPReductionRecipe *RedRecipe = | ||||||
9411 | new VPReductionRecipe(&RdxDesc, R, ChainOp, VecOp, CondOp, TTI); | ||||||
9412 | WidenRecipe->getVPSingleValue()->replaceAllUsesWith(RedRecipe); | ||||||
9413 | Plan->removeVPValueFor(R); | ||||||
9414 | Plan->addVPValue(R, RedRecipe); | ||||||
9415 | WidenRecipe->getParent()->insert(RedRecipe, WidenRecipe->getIterator()); | ||||||
9416 | WidenRecipe->getVPSingleValue()->replaceAllUsesWith(RedRecipe); | ||||||
9417 | WidenRecipe->eraseFromParent(); | ||||||
9418 | |||||||
9419 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { | ||||||
9420 | VPRecipeBase *CompareRecipe = | ||||||
9421 | RecipeBuilder.getRecipe(cast<Instruction>(R->getOperand(0))); | ||||||
9422 | 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", 9423, __extension__ __PRETTY_FUNCTION__)) | ||||||
9423 | "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", 9423, __extension__ __PRETTY_FUNCTION__)); | ||||||
9424 | 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", 9425, __extension__ __PRETTY_FUNCTION__)) | ||||||
9425 | "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", 9425, __extension__ __PRETTY_FUNCTION__)); | ||||||
9426 | CompareRecipe->eraseFromParent(); | ||||||
9427 | } | ||||||
9428 | Chain = R; | ||||||
9429 | } | ||||||
9430 | } | ||||||
9431 | |||||||
9432 | // If tail is folded by masking, introduce selects between the phi | ||||||
9433 | // and the live-out instruction of each reduction, at the beginning of the | ||||||
9434 | // dedicated latch block. | ||||||
9435 | if (CM.foldTailByMasking()) { | ||||||
9436 | Builder.setInsertPoint(LatchVPBB, LatchVPBB->begin()); | ||||||
9437 | for (VPRecipeBase &R : Plan->getEntry()->getEntryBasicBlock()->phis()) { | ||||||
9438 | VPReductionPHIRecipe *PhiR = dyn_cast<VPReductionPHIRecipe>(&R); | ||||||
9439 | if (!PhiR || PhiR->isInLoop()) | ||||||
9440 | continue; | ||||||
9441 | VPValue *Cond = | ||||||
9442 | RecipeBuilder.createBlockInMask(OrigLoop->getHeader(), Plan); | ||||||
9443 | VPValue *Red = PhiR->getBackedgeValue(); | ||||||
9444 | assert(cast<VPRecipeBase>(Red->getDef())->getParent() != LatchVPBB &&(static_cast <bool> (cast<VPRecipeBase>(Red->getDef ())->getParent() != LatchVPBB && "reduction recipe must be defined before latch" ) ? void (0) : __assert_fail ("cast<VPRecipeBase>(Red->getDef())->getParent() != LatchVPBB && \"reduction recipe must be defined before latch\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9445, __extension__ __PRETTY_FUNCTION__)) | ||||||
9445 | "reduction recipe must be defined before latch")(static_cast <bool> (cast<VPRecipeBase>(Red->getDef ())->getParent() != LatchVPBB && "reduction recipe must be defined before latch" ) ? void (0) : __assert_fail ("cast<VPRecipeBase>(Red->getDef())->getParent() != LatchVPBB && \"reduction recipe must be defined before latch\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9445, __extension__ __PRETTY_FUNCTION__)); | ||||||
9446 | Builder.createNaryOp(Instruction::Select, {Cond, Red, PhiR}); | ||||||
9447 | } | ||||||
9448 | } | ||||||
9449 | } | ||||||
9450 | |||||||
9451 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||
9452 | void VPInterleaveRecipe::print(raw_ostream &O, const Twine &Indent, | ||||||
9453 | VPSlotTracker &SlotTracker) const { | ||||||
9454 | O << Indent << "INTERLEAVE-GROUP with factor " << IG->getFactor() << " at "; | ||||||
9455 | IG->getInsertPos()->printAsOperand(O, false); | ||||||
9456 | O << ", "; | ||||||
9457 | getAddr()->printAsOperand(O, SlotTracker); | ||||||
9458 | VPValue *Mask = getMask(); | ||||||
9459 | if (Mask) { | ||||||
9460 | O << ", "; | ||||||
9461 | Mask->printAsOperand(O, SlotTracker); | ||||||
9462 | } | ||||||
9463 | |||||||
9464 | unsigned OpIdx = 0; | ||||||
9465 | for (unsigned i = 0; i < IG->getFactor(); ++i) { | ||||||
9466 | if (!IG->getMember(i)) | ||||||
9467 | continue; | ||||||
9468 | if (getNumStoreOperands() > 0) { | ||||||
9469 | O << "\n" << Indent << " store "; | ||||||
9470 | getOperand(1 + OpIdx)->printAsOperand(O, SlotTracker); | ||||||
9471 | O << " to index " << i; | ||||||
9472 | } else { | ||||||
9473 | O << "\n" << Indent << " "; | ||||||
9474 | getVPValue(OpIdx)->printAsOperand(O, SlotTracker); | ||||||
9475 | O << " = load from index " << i; | ||||||
9476 | } | ||||||
9477 | ++OpIdx; | ||||||
9478 | } | ||||||
9479 | } | ||||||
9480 | #endif | ||||||
9481 | |||||||
9482 | void VPWidenCallRecipe::execute(VPTransformState &State) { | ||||||
9483 | State.ILV->widenCallInstruction(*cast<CallInst>(getUnderlyingInstr()), this, | ||||||
9484 | *this, State); | ||||||
9485 | } | ||||||
9486 | |||||||
9487 | void VPWidenSelectRecipe::execute(VPTransformState &State) { | ||||||
9488 | auto &I = *cast<SelectInst>(getUnderlyingInstr()); | ||||||
9489 | State.ILV->setDebugLocFromInst(&I); | ||||||
9490 | |||||||
9491 | // The condition can be loop invariant but still defined inside the | ||||||
9492 | // loop. This means that we can't just use the original 'cond' value. | ||||||
9493 | // We have to take the 'vectorized' value and pick the first lane. | ||||||
9494 | // Instcombine will make this a no-op. | ||||||
9495 | auto *InvarCond = | ||||||
9496 | InvariantCond ? State.get(getOperand(0), VPIteration(0, 0)) : nullptr; | ||||||
9497 | |||||||
9498 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
9499 | Value *Cond = InvarCond ? InvarCond : State.get(getOperand(0), Part); | ||||||
9500 | Value *Op0 = State.get(getOperand(1), Part); | ||||||
9501 | Value *Op1 = State.get(getOperand(2), Part); | ||||||
9502 | Value *Sel = State.Builder.CreateSelect(Cond, Op0, Op1); | ||||||
9503 | State.set(this, Sel, Part); | ||||||
9504 | State.ILV->addMetadata(Sel, &I); | ||||||
9505 | } | ||||||
9506 | } | ||||||
9507 | |||||||
9508 | void VPWidenRecipe::execute(VPTransformState &State) { | ||||||
9509 | auto &I = *cast<Instruction>(getUnderlyingValue()); | ||||||
9510 | auto &Builder = State.Builder; | ||||||
9511 | switch (I.getOpcode()) { | ||||||
9512 | case Instruction::Call: | ||||||
9513 | case Instruction::Br: | ||||||
9514 | case Instruction::PHI: | ||||||
9515 | case Instruction::GetElementPtr: | ||||||
9516 | case Instruction::Select: | ||||||
9517 | llvm_unreachable("This instruction is handled by a different recipe.")::llvm::llvm_unreachable_internal("This instruction is handled by a different recipe." , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9517); | ||||||
9518 | case Instruction::UDiv: | ||||||
9519 | case Instruction::SDiv: | ||||||
9520 | case Instruction::SRem: | ||||||
9521 | case Instruction::URem: | ||||||
9522 | case Instruction::Add: | ||||||
9523 | case Instruction::FAdd: | ||||||
9524 | case Instruction::Sub: | ||||||
9525 | case Instruction::FSub: | ||||||
9526 | case Instruction::FNeg: | ||||||
9527 | case Instruction::Mul: | ||||||
9528 | case Instruction::FMul: | ||||||
9529 | case Instruction::FDiv: | ||||||
9530 | case Instruction::FRem: | ||||||
9531 | case Instruction::Shl: | ||||||
9532 | case Instruction::LShr: | ||||||
9533 | case Instruction::AShr: | ||||||
9534 | case Instruction::And: | ||||||
9535 | case Instruction::Or: | ||||||
9536 | case Instruction::Xor: { | ||||||
9537 | // Just widen unops and binops. | ||||||
9538 | State.ILV->setDebugLocFromInst(&I); | ||||||
9539 | |||||||
9540 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
9541 | SmallVector<Value *, 2> Ops; | ||||||
9542 | for (VPValue *VPOp : operands()) | ||||||
9543 | Ops.push_back(State.get(VPOp, Part)); | ||||||
9544 | |||||||
9545 | Value *V = Builder.CreateNAryOp(I.getOpcode(), Ops); | ||||||
9546 | |||||||
9547 | if (auto *VecOp = dyn_cast<Instruction>(V)) { | ||||||
9548 | VecOp->copyIRFlags(&I); | ||||||
9549 | |||||||
9550 | // If the instruction is vectorized and was in a basic block that needed | ||||||
9551 | // predication, we can't propagate poison-generating flags (nuw/nsw, | ||||||
9552 | // exact, etc.). The control flow has been linearized and the | ||||||
9553 | // instruction is no longer guarded by the predicate, which could make | ||||||
9554 | // the flag properties to no longer hold. | ||||||
9555 | if (State.MayGeneratePoisonRecipes.contains(this)) | ||||||
9556 | VecOp->dropPoisonGeneratingFlags(); | ||||||
9557 | } | ||||||
9558 | |||||||
9559 | // Use this vector value for all users of the original instruction. | ||||||
9560 | State.set(this, V, Part); | ||||||
9561 | State.ILV->addMetadata(V, &I); | ||||||
9562 | } | ||||||
9563 | |||||||
9564 | break; | ||||||
9565 | } | ||||||
9566 | case Instruction::ICmp: | ||||||
9567 | case Instruction::FCmp: { | ||||||
9568 | // Widen compares. Generate vector compares. | ||||||
9569 | bool FCmp = (I.getOpcode() == Instruction::FCmp); | ||||||
9570 | auto *Cmp = cast<CmpInst>(&I); | ||||||
9571 | State.ILV->setDebugLocFromInst(Cmp); | ||||||
9572 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
9573 | Value *A = State.get(getOperand(0), Part); | ||||||
9574 | Value *B = State.get(getOperand(1), Part); | ||||||
9575 | Value *C = nullptr; | ||||||
9576 | if (FCmp) { | ||||||
9577 | // Propagate fast math flags. | ||||||
9578 | IRBuilder<>::FastMathFlagGuard FMFG(Builder); | ||||||
9579 | Builder.setFastMathFlags(Cmp->getFastMathFlags()); | ||||||
9580 | C = Builder.CreateFCmp(Cmp->getPredicate(), A, B); | ||||||
9581 | } else { | ||||||
9582 | C = Builder.CreateICmp(Cmp->getPredicate(), A, B); | ||||||
9583 | } | ||||||
9584 | State.set(this, C, Part); | ||||||
9585 | State.ILV->addMetadata(C, &I); | ||||||
9586 | } | ||||||
9587 | |||||||
9588 | break; | ||||||
9589 | } | ||||||
9590 | |||||||
9591 | case Instruction::ZExt: | ||||||
9592 | case Instruction::SExt: | ||||||
9593 | case Instruction::FPToUI: | ||||||
9594 | case Instruction::FPToSI: | ||||||
9595 | case Instruction::FPExt: | ||||||
9596 | case Instruction::PtrToInt: | ||||||
9597 | case Instruction::IntToPtr: | ||||||
9598 | case Instruction::SIToFP: | ||||||
9599 | case Instruction::UIToFP: | ||||||
9600 | case Instruction::Trunc: | ||||||
9601 | case Instruction::FPTrunc: | ||||||
9602 | case Instruction::BitCast: { | ||||||
9603 | auto *CI = cast<CastInst>(&I); | ||||||
9604 | State.ILV->setDebugLocFromInst(CI); | ||||||
9605 | |||||||
9606 | /// Vectorize casts. | ||||||
9607 | Type *DestTy = (State.VF.isScalar()) | ||||||
9608 | ? CI->getType() | ||||||
9609 | : VectorType::get(CI->getType(), State.VF); | ||||||
9610 | |||||||
9611 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
9612 | Value *A = State.get(getOperand(0), Part); | ||||||
9613 | Value *Cast = Builder.CreateCast(CI->getOpcode(), A, DestTy); | ||||||
9614 | State.set(this, Cast, Part); | ||||||
9615 | State.ILV->addMetadata(Cast, &I); | ||||||
9616 | } | ||||||
9617 | break; | ||||||
9618 | } | ||||||
9619 | default: | ||||||
9620 | // This instruction is not vectorized by simple widening. | ||||||
9621 | LLVM_DEBUG(dbgs() << "LV: Found an unhandled instruction: " << I)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found an unhandled instruction: " << I; } } while (false); | ||||||
9622 | llvm_unreachable("Unhandled instruction!")::llvm::llvm_unreachable_internal("Unhandled instruction!", "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp" , 9622); | ||||||
9623 | } // end of switch. | ||||||
9624 | } | ||||||
9625 | |||||||
9626 | void VPWidenGEPRecipe::execute(VPTransformState &State) { | ||||||
9627 | auto *GEP = cast<GetElementPtrInst>(getUnderlyingInstr()); | ||||||
9628 | // Construct a vector GEP by widening the operands of the scalar GEP as | ||||||
9629 | // necessary. We mark the vector GEP 'inbounds' if appropriate. A GEP | ||||||
9630 | // results in a vector of pointers when at least one operand of the GEP | ||||||
9631 | // is vector-typed. Thus, to keep the representation compact, we only use | ||||||
9632 | // vector-typed operands for loop-varying values. | ||||||
9633 | |||||||
9634 | if (State.VF.isVector() && IsPtrLoopInvariant && IsIndexLoopInvariant.all()) { | ||||||
9635 | // If we are vectorizing, but the GEP has only loop-invariant operands, | ||||||
9636 | // the GEP we build (by only using vector-typed operands for | ||||||
9637 | // loop-varying values) would be a scalar pointer. Thus, to ensure we | ||||||
9638 | // produce a vector of pointers, we need to either arbitrarily pick an | ||||||
9639 | // operand to broadcast, or broadcast a clone of the original GEP. | ||||||
9640 | // Here, we broadcast a clone of the original. | ||||||
9641 | // | ||||||
9642 | // TODO: If at some point we decide to scalarize instructions having | ||||||
9643 | // loop-invariant operands, this special case will no longer be | ||||||
9644 | // required. We would add the scalarization decision to | ||||||
9645 | // collectLoopScalars() and teach getVectorValue() to broadcast | ||||||
9646 | // the lane-zero scalar value. | ||||||
9647 | auto *Clone = State.Builder.Insert(GEP->clone()); | ||||||
9648 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
9649 | Value *EntryPart = State.Builder.CreateVectorSplat(State.VF, Clone); | ||||||
9650 | State.set(this, EntryPart, Part); | ||||||
9651 | State.ILV->addMetadata(EntryPart, GEP); | ||||||
9652 | } | ||||||
9653 | } else { | ||||||
9654 | // If the GEP has at least one loop-varying operand, we are sure to | ||||||
9655 | // produce a vector of pointers. But if we are only unrolling, we want | ||||||
9656 | // to produce a scalar GEP for each unroll part. Thus, the GEP we | ||||||
9657 | // produce with the code below will be scalar (if VF == 1) or vector | ||||||
9658 | // (otherwise). Note that for the unroll-only case, we still maintain | ||||||
9659 | // values in the vector mapping with initVector, as we do for other | ||||||
9660 | // instructions. | ||||||
9661 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
9662 | // The pointer operand of the new GEP. If it's loop-invariant, we | ||||||
9663 | // won't broadcast it. | ||||||
9664 | auto *Ptr = IsPtrLoopInvariant | ||||||
9665 | ? State.get(getOperand(0), VPIteration(0, 0)) | ||||||
9666 | : State.get(getOperand(0), Part); | ||||||
9667 | |||||||
9668 | // Collect all the indices for the new GEP. If any index is | ||||||
9669 | // loop-invariant, we won't broadcast it. | ||||||
9670 | SmallVector<Value *, 4> Indices; | ||||||
9671 | for (unsigned I = 1, E = getNumOperands(); I < E; I++) { | ||||||
9672 | VPValue *Operand = getOperand(I); | ||||||
9673 | if (IsIndexLoopInvariant[I - 1]) | ||||||
9674 | Indices.push_back(State.get(Operand, VPIteration(0, 0))); | ||||||
9675 | else | ||||||
9676 | Indices.push_back(State.get(Operand, Part)); | ||||||
9677 | } | ||||||
9678 | |||||||
9679 | // If the GEP instruction is vectorized and was in a basic block that | ||||||
9680 | // needed predication, we can't propagate the poison-generating 'inbounds' | ||||||
9681 | // flag. The control flow has been linearized and the GEP is no longer | ||||||
9682 | // guarded by the predicate, which could make the 'inbounds' properties to | ||||||
9683 | // no longer hold. | ||||||
9684 | bool IsInBounds = | ||||||
9685 | GEP->isInBounds() && State.MayGeneratePoisonRecipes.count(this) == 0; | ||||||
9686 | |||||||
9687 | // Create the new GEP. Note that this GEP may be a scalar if VF == 1, | ||||||
9688 | // but it should be a vector, otherwise. | ||||||
9689 | auto *NewGEP = IsInBounds | ||||||
9690 | ? State.Builder.CreateInBoundsGEP( | ||||||
9691 | GEP->getSourceElementType(), Ptr, Indices) | ||||||
9692 | : State.Builder.CreateGEP(GEP->getSourceElementType(), | ||||||
9693 | Ptr, Indices); | ||||||
9694 | assert((State.VF.isScalar() || NewGEP->getType()->isVectorTy()) &&(static_cast <bool> ((State.VF.isScalar() || NewGEP-> getType()->isVectorTy()) && "NewGEP is not a pointer vector" ) ? void (0) : __assert_fail ("(State.VF.isScalar() || NewGEP->getType()->isVectorTy()) && \"NewGEP is not a pointer vector\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9695, __extension__ __PRETTY_FUNCTION__)) | ||||||
9695 | "NewGEP is not a pointer vector")(static_cast <bool> ((State.VF.isScalar() || NewGEP-> getType()->isVectorTy()) && "NewGEP is not a pointer vector" ) ? void (0) : __assert_fail ("(State.VF.isScalar() || NewGEP->getType()->isVectorTy()) && \"NewGEP is not a pointer vector\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9695, __extension__ __PRETTY_FUNCTION__)); | ||||||
9696 | State.set(this, NewGEP, Part); | ||||||
9697 | State.ILV->addMetadata(NewGEP, GEP); | ||||||
9698 | } | ||||||
9699 | } | ||||||
9700 | } | ||||||
9701 | |||||||
9702 | void VPWidenIntOrFpInductionRecipe::execute(VPTransformState &State) { | ||||||
9703 | 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", 9703, __extension__ __PRETTY_FUNCTION__)); | ||||||
9704 | auto *CanonicalIV = State.get(getParent()->getPlan()->getCanonicalIV(), 0); | ||||||
9705 | State.ILV->widenIntOrFpInduction(IV, getInductionDescriptor(), | ||||||
9706 | getStartValue()->getLiveInIRValue(), | ||||||
9707 | getTruncInst(), this, State, CanonicalIV); | ||||||
9708 | } | ||||||
9709 | |||||||
9710 | void VPWidenPHIRecipe::execute(VPTransformState &State) { | ||||||
9711 | State.ILV->widenPHIInstruction(cast<PHINode>(getUnderlyingValue()), this, | ||||||
9712 | State); | ||||||
9713 | } | ||||||
9714 | |||||||
9715 | void VPBlendRecipe::execute(VPTransformState &State) { | ||||||
9716 | State.ILV->setDebugLocFromInst(Phi, &State.Builder); | ||||||
9717 | // We know that all PHIs in non-header blocks are converted into | ||||||
9718 | // selects, so we don't have to worry about the insertion order and we | ||||||
9719 | // can just use the builder. | ||||||
9720 | // At this point we generate the predication tree. There may be | ||||||
9721 | // duplications since this is a simple recursive scan, but future | ||||||
9722 | // optimizations will clean it up. | ||||||
9723 | |||||||
9724 | unsigned NumIncoming = getNumIncomingValues(); | ||||||
9725 | |||||||
9726 | // Generate a sequence of selects of the form: | ||||||
9727 | // SELECT(Mask3, In3, | ||||||
9728 | // SELECT(Mask2, In2, | ||||||
9729 | // SELECT(Mask1, In1, | ||||||
9730 | // In0))) | ||||||
9731 | // Note that Mask0 is never used: lanes for which no path reaches this phi and | ||||||
9732 | // are essentially undef are taken from In0. | ||||||
9733 | InnerLoopVectorizer::VectorParts Entry(State.UF); | ||||||
9734 | for (unsigned In = 0; In < NumIncoming; ++In) { | ||||||
9735 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
9736 | // We might have single edge PHIs (blocks) - use an identity | ||||||
9737 | // 'select' for the first PHI operand. | ||||||
9738 | Value *In0 = State.get(getIncomingValue(In), Part); | ||||||
9739 | if (In == 0) | ||||||
9740 | Entry[Part] = In0; // Initialize with the first incoming value. | ||||||
9741 | else { | ||||||
9742 | // Select between the current value and the previous incoming edge | ||||||
9743 | // based on the incoming mask. | ||||||
9744 | Value *Cond = State.get(getMask(In), Part); | ||||||
9745 | Entry[Part] = | ||||||
9746 | State.Builder.CreateSelect(Cond, In0, Entry[Part], "predphi"); | ||||||
9747 | } | ||||||
9748 | } | ||||||
9749 | } | ||||||
9750 | for (unsigned Part = 0; Part < State.UF; ++Part) | ||||||
9751 | State.set(this, Entry[Part], Part); | ||||||
9752 | } | ||||||
9753 | |||||||
9754 | void VPInterleaveRecipe::execute(VPTransformState &State) { | ||||||
9755 | 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", 9755, __extension__ __PRETTY_FUNCTION__)); | ||||||
9756 | State.ILV->vectorizeInterleaveGroup(IG, definedValues(), State, getAddr(), | ||||||
9757 | getStoredValues(), getMask()); | ||||||
9758 | } | ||||||
9759 | |||||||
9760 | void VPReductionRecipe::execute(VPTransformState &State) { | ||||||
9761 | 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", 9761, __extension__ __PRETTY_FUNCTION__)); | ||||||
9762 | Value *PrevInChain = State.get(getChainOp(), 0); | ||||||
9763 | RecurKind Kind = RdxDesc->getRecurrenceKind(); | ||||||
9764 | bool IsOrdered = State.ILV->useOrderedReductions(*RdxDesc); | ||||||
9765 | // Propagate the fast-math flags carried by the underlying instruction. | ||||||
9766 | IRBuilderBase::FastMathFlagGuard FMFGuard(State.Builder); | ||||||
9767 | State.Builder.setFastMathFlags(RdxDesc->getFastMathFlags()); | ||||||
9768 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
9769 | Value *NewVecOp = State.get(getVecOp(), Part); | ||||||
9770 | if (VPValue *Cond = getCondOp()) { | ||||||
9771 | Value *NewCond = State.get(Cond, Part); | ||||||
9772 | VectorType *VecTy = cast<VectorType>(NewVecOp->getType()); | ||||||
9773 | Value *Iden = RdxDesc->getRecurrenceIdentity( | ||||||
9774 | Kind, VecTy->getElementType(), RdxDesc->getFastMathFlags()); | ||||||
9775 | Value *IdenVec = | ||||||
9776 | State.Builder.CreateVectorSplat(VecTy->getElementCount(), Iden); | ||||||
9777 | Value *Select = State.Builder.CreateSelect(NewCond, NewVecOp, IdenVec); | ||||||
9778 | NewVecOp = Select; | ||||||
9779 | } | ||||||
9780 | Value *NewRed; | ||||||
9781 | Value *NextInChain; | ||||||
9782 | if (IsOrdered) { | ||||||
9783 | if (State.VF.isVector()) | ||||||
9784 | NewRed = createOrderedReduction(State.Builder, *RdxDesc, NewVecOp, | ||||||
9785 | PrevInChain); | ||||||
9786 | else | ||||||
9787 | NewRed = State.Builder.CreateBinOp( | ||||||
9788 | (Instruction::BinaryOps)RdxDesc->getOpcode(Kind), PrevInChain, | ||||||
9789 | NewVecOp); | ||||||
9790 | PrevInChain = NewRed; | ||||||
9791 | } else { | ||||||
9792 | PrevInChain = State.get(getChainOp(), Part); | ||||||
9793 | NewRed = createTargetReduction(State.Builder, TTI, *RdxDesc, NewVecOp); | ||||||
9794 | } | ||||||
9795 | if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { | ||||||
9796 | NextInChain = | ||||||
9797 | createMinMaxOp(State.Builder, RdxDesc->getRecurrenceKind(), | ||||||
9798 | NewRed, PrevInChain); | ||||||
9799 | } else if (IsOrdered) | ||||||
9800 | NextInChain = NewRed; | ||||||
9801 | else | ||||||
9802 | NextInChain = State.Builder.CreateBinOp( | ||||||
9803 | (Instruction::BinaryOps)RdxDesc->getOpcode(Kind), NewRed, | ||||||
9804 | PrevInChain); | ||||||
9805 | State.set(this, NextInChain, Part); | ||||||
9806 | } | ||||||
9807 | } | ||||||
9808 | |||||||
9809 | void VPReplicateRecipe::execute(VPTransformState &State) { | ||||||
9810 | if (State.Instance) { // Generate a single instance. | ||||||
9811 | 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", 9811, __extension__ __PRETTY_FUNCTION__)); | ||||||
9812 | State.ILV->scalarizeInstruction(getUnderlyingInstr(), this, *State.Instance, | ||||||
9813 | IsPredicated, State); | ||||||
9814 | // Insert scalar instance packing it into a vector. | ||||||
9815 | if (AlsoPack && State.VF.isVector()) { | ||||||
9816 | // If we're constructing lane 0, initialize to start from poison. | ||||||
9817 | if (State.Instance->Lane.isFirstLane()) { | ||||||
9818 | 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", 9818, __extension__ __PRETTY_FUNCTION__)); | ||||||
9819 | Value *Poison = PoisonValue::get( | ||||||
9820 | VectorType::get(getUnderlyingValue()->getType(), State.VF)); | ||||||
9821 | State.set(this, Poison, State.Instance->Part); | ||||||
9822 | } | ||||||
9823 | State.ILV->packScalarIntoVectorValue(this, *State.Instance, State); | ||||||
9824 | } | ||||||
9825 | return; | ||||||
9826 | } | ||||||
9827 | |||||||
9828 | // Generate scalar instances for all VF lanes of all UF parts, unless the | ||||||
9829 | // instruction is uniform inwhich case generate only the first lane for each | ||||||
9830 | // of the UF parts. | ||||||
9831 | unsigned EndLane = IsUniform ? 1 : State.VF.getKnownMinValue(); | ||||||
9832 | assert((!State.VF.isScalable() || IsUniform) &&(static_cast <bool> ((!State.VF.isScalable() || IsUniform ) && "Can't scalarize a scalable vector") ? void (0) : __assert_fail ("(!State.VF.isScalable() || IsUniform) && \"Can't scalarize a scalable vector\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9833, __extension__ __PRETTY_FUNCTION__)) | ||||||
9833 | "Can't scalarize a scalable vector")(static_cast <bool> ((!State.VF.isScalable() || IsUniform ) && "Can't scalarize a scalable vector") ? void (0) : __assert_fail ("(!State.VF.isScalable() || IsUniform) && \"Can't scalarize a scalable vector\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9833, __extension__ __PRETTY_FUNCTION__)); | ||||||
9834 | for (unsigned Part = 0; Part < State.UF; ++Part) | ||||||
9835 | for (unsigned Lane = 0; Lane < EndLane; ++Lane) | ||||||
9836 | State.ILV->scalarizeInstruction(getUnderlyingInstr(), this, | ||||||
9837 | VPIteration(Part, Lane), IsPredicated, | ||||||
9838 | State); | ||||||
9839 | } | ||||||
9840 | |||||||
9841 | void VPBranchOnMaskRecipe::execute(VPTransformState &State) { | ||||||
9842 | assert(State.Instance && "Branch on Mask works only on single instance.")(static_cast <bool> (State.Instance && "Branch on Mask works only on single instance." ) ? void (0) : __assert_fail ("State.Instance && \"Branch on Mask works only on single instance.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9842, __extension__ __PRETTY_FUNCTION__)); | ||||||
9843 | |||||||
9844 | unsigned Part = State.Instance->Part; | ||||||
9845 | unsigned Lane = State.Instance->Lane.getKnownLane(); | ||||||
9846 | |||||||
9847 | Value *ConditionBit = nullptr; | ||||||
9848 | VPValue *BlockInMask = getMask(); | ||||||
9849 | if (BlockInMask) { | ||||||
9850 | ConditionBit = State.get(BlockInMask, Part); | ||||||
9851 | if (ConditionBit->getType()->isVectorTy()) | ||||||
9852 | ConditionBit = State.Builder.CreateExtractElement( | ||||||
9853 | ConditionBit, State.Builder.getInt32(Lane)); | ||||||
9854 | } else // Block in mask is all-one. | ||||||
9855 | ConditionBit = State.Builder.getTrue(); | ||||||
9856 | |||||||
9857 | // Replace the temporary unreachable terminator with a new conditional branch, | ||||||
9858 | // whose two destinations will be set later when they are created. | ||||||
9859 | auto *CurrentTerminator = State.CFG.PrevBB->getTerminator(); | ||||||
9860 | assert(isa<UnreachableInst>(CurrentTerminator) &&(static_cast <bool> (isa<UnreachableInst>(CurrentTerminator ) && "Expected to replace unreachable terminator with conditional branch." ) ? void (0) : __assert_fail ("isa<UnreachableInst>(CurrentTerminator) && \"Expected to replace unreachable terminator with conditional branch.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9861, __extension__ __PRETTY_FUNCTION__)) | ||||||
9861 | "Expected to replace unreachable terminator with conditional branch.")(static_cast <bool> (isa<UnreachableInst>(CurrentTerminator ) && "Expected to replace unreachable terminator with conditional branch." ) ? void (0) : __assert_fail ("isa<UnreachableInst>(CurrentTerminator) && \"Expected to replace unreachable terminator with conditional branch.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9861, __extension__ __PRETTY_FUNCTION__)); | ||||||
9862 | auto *CondBr = BranchInst::Create(State.CFG.PrevBB, nullptr, ConditionBit); | ||||||
9863 | CondBr->setSuccessor(0, nullptr); | ||||||
9864 | ReplaceInstWithInst(CurrentTerminator, CondBr); | ||||||
9865 | } | ||||||
9866 | |||||||
9867 | void VPPredInstPHIRecipe::execute(VPTransformState &State) { | ||||||
9868 | assert(State.Instance && "Predicated instruction PHI works per instance.")(static_cast <bool> (State.Instance && "Predicated instruction PHI works per instance." ) ? void (0) : __assert_fail ("State.Instance && \"Predicated instruction PHI works per instance.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9868, __extension__ __PRETTY_FUNCTION__)); | ||||||
9869 | Instruction *ScalarPredInst = | ||||||
9870 | cast<Instruction>(State.get(getOperand(0), *State.Instance)); | ||||||
9871 | BasicBlock *PredicatedBB = ScalarPredInst->getParent(); | ||||||
9872 | BasicBlock *PredicatingBB = PredicatedBB->getSinglePredecessor(); | ||||||
9873 | assert(PredicatingBB && "Predicated block has no single predecessor.")(static_cast <bool> (PredicatingBB && "Predicated block has no single predecessor." ) ? void (0) : __assert_fail ("PredicatingBB && \"Predicated block has no single predecessor.\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9873, __extension__ __PRETTY_FUNCTION__)); | ||||||
9874 | assert(isa<VPReplicateRecipe>(getOperand(0)) &&(static_cast <bool> (isa<VPReplicateRecipe>(getOperand (0)) && "operand must be VPReplicateRecipe") ? void ( 0) : __assert_fail ("isa<VPReplicateRecipe>(getOperand(0)) && \"operand must be VPReplicateRecipe\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9875, __extension__ __PRETTY_FUNCTION__)) | ||||||
9875 | "operand must be VPReplicateRecipe")(static_cast <bool> (isa<VPReplicateRecipe>(getOperand (0)) && "operand must be VPReplicateRecipe") ? void ( 0) : __assert_fail ("isa<VPReplicateRecipe>(getOperand(0)) && \"operand must be VPReplicateRecipe\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 9875, __extension__ __PRETTY_FUNCTION__)); | ||||||
9876 | |||||||
9877 | // By current pack/unpack logic we need to generate only a single phi node: if | ||||||
9878 | // a vector value for the predicated instruction exists at this point it means | ||||||
9879 | // the instruction has vector users only, and a phi for the vector value is | ||||||
9880 | // needed. In this case the recipe of the predicated instruction is marked to | ||||||
9881 | // also do that packing, thereby "hoisting" the insert-element sequence. | ||||||
9882 | // Otherwise, a phi node for the scalar value is needed. | ||||||
9883 | unsigned Part = State.Instance->Part; | ||||||
9884 | if (State.hasVectorValue(getOperand(0), Part)) { | ||||||
9885 | Value *VectorValue = State.get(getOperand(0), Part); | ||||||
9886 | InsertElementInst *IEI = cast<InsertElementInst>(VectorValue); | ||||||
9887 | PHINode *VPhi = State.Builder.CreatePHI(IEI->getType(), 2); | ||||||
9888 | VPhi->addIncoming(IEI->getOperand(0), PredicatingBB); // Unmodified vector. | ||||||
9889 | VPhi->addIncoming(IEI, PredicatedBB); // New vector with inserted element. | ||||||
9890 | if (State.hasVectorValue(this, Part)) | ||||||
9891 | State.reset(this, VPhi, Part); | ||||||
9892 | else | ||||||
9893 | State.set(this, VPhi, Part); | ||||||
9894 | // NOTE: Currently we need to update the value of the operand, so the next | ||||||
9895 | // predicated iteration inserts its generated value in the correct vector. | ||||||
9896 | State.reset(getOperand(0), VPhi, Part); | ||||||
9897 | } else { | ||||||
9898 | Type *PredInstType = getOperand(0)->getUnderlyingValue()->getType(); | ||||||
9899 | PHINode *Phi = State.Builder.CreatePHI(PredInstType, 2); | ||||||
9900 | Phi->addIncoming(PoisonValue::get(ScalarPredInst->getType()), | ||||||
9901 | PredicatingBB); | ||||||
9902 | Phi->addIncoming(ScalarPredInst, PredicatedBB); | ||||||
9903 | if (State.hasScalarValue(this, *State.Instance)) | ||||||
9904 | State.reset(this, Phi, *State.Instance); | ||||||
9905 | else | ||||||
9906 | State.set(this, Phi, *State.Instance); | ||||||
9907 | // NOTE: Currently we need to update the value of the operand, so the next | ||||||
9908 | // predicated iteration inserts its generated value in the correct vector. | ||||||
9909 | State.reset(getOperand(0), Phi, *State.Instance); | ||||||
9910 | } | ||||||
9911 | } | ||||||
9912 | |||||||
9913 | void VPWidenMemoryInstructionRecipe::execute(VPTransformState &State) { | ||||||
9914 | VPValue *StoredValue = isStore() ? getStoredValue() : nullptr; | ||||||
9915 | |||||||
9916 | // Attempt to issue a wide load. | ||||||
9917 | LoadInst *LI = dyn_cast<LoadInst>(&Ingredient); | ||||||
9918 | StoreInst *SI = dyn_cast<StoreInst>(&Ingredient); | ||||||
9919 | |||||||
9920 | 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", 9920, __extension__ __PRETTY_FUNCTION__)); | ||||||
9921 | 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", 9921, __extension__ __PRETTY_FUNCTION__)); | ||||||
9922 | 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", 9922, __extension__ __PRETTY_FUNCTION__)); | ||||||
9923 | |||||||
9924 | Type *ScalarDataTy = getLoadStoreType(&Ingredient); | ||||||
9925 | |||||||
9926 | auto *DataTy = VectorType::get(ScalarDataTy, State.VF); | ||||||
9927 | const Align Alignment = getLoadStoreAlignment(&Ingredient); | ||||||
9928 | bool CreateGatherScatter = !Consecutive; | ||||||
9929 | |||||||
9930 | auto &Builder = State.Builder; | ||||||
9931 | InnerLoopVectorizer::VectorParts BlockInMaskParts(State.UF); | ||||||
9932 | bool isMaskRequired = getMask(); | ||||||
9933 | if (isMaskRequired) | ||||||
9934 | for (unsigned Part = 0; Part < State.UF; ++Part) | ||||||
9935 | BlockInMaskParts[Part] = State.get(getMask(), Part); | ||||||
9936 | |||||||
9937 | const auto CreateVecPtr = [&](unsigned Part, Value *Ptr) -> Value * { | ||||||
9938 | // Calculate the pointer for the specific unroll-part. | ||||||
9939 | GetElementPtrInst *PartPtr = nullptr; | ||||||
9940 | |||||||
9941 | bool InBounds = false; | ||||||
9942 | if (auto *gep = dyn_cast<GetElementPtrInst>(Ptr->stripPointerCasts())) | ||||||
9943 | InBounds = gep->isInBounds(); | ||||||
9944 | if (Reverse) { | ||||||
9945 | // If the address is consecutive but reversed, then the | ||||||
9946 | // wide store needs to start at the last vector element. | ||||||
9947 | // RunTimeVF = VScale * VF.getKnownMinValue() | ||||||
9948 | // For fixed-width VScale is 1, then RunTimeVF = VF.getKnownMinValue() | ||||||
9949 | Value *RunTimeVF = getRuntimeVF(Builder, Builder.getInt32Ty(), State.VF); | ||||||
9950 | // NumElt = -Part * RunTimeVF | ||||||
9951 | Value *NumElt = Builder.CreateMul(Builder.getInt32(-Part), RunTimeVF); | ||||||
9952 | // LastLane = 1 - RunTimeVF | ||||||
9953 | Value *LastLane = Builder.CreateSub(Builder.getInt32(1), RunTimeVF); | ||||||
9954 | PartPtr = | ||||||
9955 | cast<GetElementPtrInst>(Builder.CreateGEP(ScalarDataTy, Ptr, NumElt)); | ||||||
9956 | PartPtr->setIsInBounds(InBounds); | ||||||
9957 | PartPtr = cast<GetElementPtrInst>( | ||||||
9958 | Builder.CreateGEP(ScalarDataTy, PartPtr, LastLane)); | ||||||
9959 | PartPtr->setIsInBounds(InBounds); | ||||||
9960 | if (isMaskRequired) // Reverse of a null all-one mask is a null mask. | ||||||
9961 | BlockInMaskParts[Part] = | ||||||
9962 | Builder.CreateVectorReverse(BlockInMaskParts[Part], "reverse"); | ||||||
9963 | } else { | ||||||
9964 | Value *Increment = | ||||||
9965 | createStepForVF(Builder, Builder.getInt32Ty(), State.VF, Part); | ||||||
9966 | PartPtr = cast<GetElementPtrInst>( | ||||||
9967 | Builder.CreateGEP(ScalarDataTy, Ptr, Increment)); | ||||||
9968 | PartPtr->setIsInBounds(InBounds); | ||||||
9969 | } | ||||||
9970 | |||||||
9971 | unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace(); | ||||||
9972 | return Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace)); | ||||||
9973 | }; | ||||||
9974 | |||||||
9975 | // Handle Stores: | ||||||
9976 | if (SI) { | ||||||
9977 | State.ILV->setDebugLocFromInst(SI); | ||||||
9978 | |||||||
9979 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
9980 | Instruction *NewSI = nullptr; | ||||||
9981 | Value *StoredVal = State.get(StoredValue, Part); | ||||||
9982 | if (CreateGatherScatter) { | ||||||
9983 | Value *MaskPart = isMaskRequired ? BlockInMaskParts[Part] : nullptr; | ||||||
9984 | Value *VectorGep = State.get(getAddr(), Part); | ||||||
9985 | NewSI = Builder.CreateMaskedScatter(StoredVal, VectorGep, Alignment, | ||||||
9986 | MaskPart); | ||||||
9987 | } else { | ||||||
9988 | if (Reverse) { | ||||||
9989 | // If we store to reverse consecutive memory locations, then we need | ||||||
9990 | // to reverse the order of elements in the stored value. | ||||||
9991 | StoredVal = Builder.CreateVectorReverse(StoredVal, "reverse"); | ||||||
9992 | // We don't want to update the value in the map as it might be used in | ||||||
9993 | // another expression. So don't call resetVectorValue(StoredVal). | ||||||
9994 | } | ||||||
9995 | auto *VecPtr = | ||||||
9996 | CreateVecPtr(Part, State.get(getAddr(), VPIteration(0, 0))); | ||||||
9997 | if (isMaskRequired) | ||||||
9998 | NewSI = Builder.CreateMaskedStore(StoredVal, VecPtr, Alignment, | ||||||
9999 | BlockInMaskParts[Part]); | ||||||
10000 | else | ||||||
10001 | NewSI = Builder.CreateAlignedStore(StoredVal, VecPtr, Alignment); | ||||||
10002 | } | ||||||
10003 | State.ILV->addMetadata(NewSI, SI); | ||||||
10004 | } | ||||||
10005 | return; | ||||||
10006 | } | ||||||
10007 | |||||||
10008 | // Handle loads. | ||||||
10009 | 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", 10009, __extension__ __PRETTY_FUNCTION__)); | ||||||
10010 | State.ILV->setDebugLocFromInst(LI); | ||||||
10011 | for (unsigned Part = 0; Part < State.UF; ++Part) { | ||||||
10012 | Value *NewLI; | ||||||
10013 | if (CreateGatherScatter) { | ||||||
10014 | Value *MaskPart = isMaskRequired ? BlockInMaskParts[Part] : nullptr; | ||||||
10015 | Value *VectorGep = State.get(getAddr(), Part); | ||||||
10016 | NewLI = Builder.CreateMaskedGather(DataTy, VectorGep, Alignment, MaskPart, | ||||||
10017 | nullptr, "wide.masked.gather"); | ||||||
10018 | State.ILV->addMetadata(NewLI, LI); | ||||||
10019 | } else { | ||||||
10020 | auto *VecPtr = | ||||||
10021 | CreateVecPtr(Part, State.get(getAddr(), VPIteration(0, 0))); | ||||||
10022 | if (isMaskRequired) | ||||||
10023 | NewLI = Builder.CreateMaskedLoad( | ||||||
10024 | DataTy, VecPtr, Alignment, BlockInMaskParts[Part], | ||||||
10025 | PoisonValue::get(DataTy), "wide.masked.load"); | ||||||
10026 | else | ||||||
10027 | NewLI = | ||||||
10028 | Builder.CreateAlignedLoad(DataTy, VecPtr, Alignment, "wide.load"); | ||||||
10029 | |||||||
10030 | // Add metadata to the load, but setVectorValue to the reverse shuffle. | ||||||
10031 | State.ILV->addMetadata(NewLI, LI); | ||||||
10032 | if (Reverse) | ||||||
10033 | NewLI = Builder.CreateVectorReverse(NewLI, "reverse"); | ||||||
10034 | } | ||||||
10035 | |||||||
10036 | State.set(this, NewLI, Part); | ||||||
10037 | } | ||||||
10038 | } | ||||||
10039 | |||||||
10040 | // Determine how to lower the scalar epilogue, which depends on 1) optimising | ||||||
10041 | // for minimum code-size, 2) predicate compiler options, 3) loop hints forcing | ||||||
10042 | // predication, and 4) a TTI hook that analyses whether the loop is suitable | ||||||
10043 | // for predication. | ||||||
10044 | static ScalarEpilogueLowering getScalarEpilogueLowering( | ||||||
10045 | Function *F, Loop *L, LoopVectorizeHints &Hints, ProfileSummaryInfo *PSI, | ||||||
10046 | BlockFrequencyInfo *BFI, TargetTransformInfo *TTI, TargetLibraryInfo *TLI, | ||||||
10047 | AssumptionCache *AC, LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT, | ||||||
10048 | LoopVectorizationLegality &LVL) { | ||||||
10049 | // 1) OptSize takes precedence over all other options, i.e. if this is set, | ||||||
10050 | // don't look at hints or options, and don't request a scalar epilogue. | ||||||
10051 | // (For PGSO, as shouldOptimizeForSize isn't currently accessible from | ||||||
10052 | // LoopAccessInfo (due to code dependency and not being able to reliably get | ||||||
10053 | // PSI/BFI from a loop analysis under NPM), we cannot suppress the collection | ||||||
10054 | // of strides in LoopAccessInfo::analyzeLoop() and vectorize without | ||||||
10055 | // versioning when the vectorization is forced, unlike hasOptSize. So revert | ||||||
10056 | // back to the old way and vectorize with versioning when forced. See D81345.) | ||||||
10057 | if (F->hasOptSize() || (llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI, | ||||||
10058 | PGSOQueryType::IRPass) && | ||||||
10059 | Hints.getForce() != LoopVectorizeHints::FK_Enabled)) | ||||||
10060 | return CM_ScalarEpilogueNotAllowedOptSize; | ||||||
10061 | |||||||
10062 | // 2) If set, obey the directives | ||||||
10063 | if (PreferPredicateOverEpilogue.getNumOccurrences()) { | ||||||
10064 | switch (PreferPredicateOverEpilogue) { | ||||||
10065 | case PreferPredicateTy::ScalarEpilogue: | ||||||
10066 | return CM_ScalarEpilogueAllowed; | ||||||
10067 | case PreferPredicateTy::PredicateElseScalarEpilogue: | ||||||
10068 | return CM_ScalarEpilogueNotNeededUsePredicate; | ||||||
10069 | case PreferPredicateTy::PredicateOrDontVectorize: | ||||||
10070 | return CM_ScalarEpilogueNotAllowedUsePredicate; | ||||||
10071 | }; | ||||||
10072 | } | ||||||
10073 | |||||||
10074 | // 3) If set, obey the hints | ||||||
10075 | switch (Hints.getPredicate()) { | ||||||
10076 | case LoopVectorizeHints::FK_Enabled: | ||||||
10077 | return CM_ScalarEpilogueNotNeededUsePredicate; | ||||||
10078 | case LoopVectorizeHints::FK_Disabled: | ||||||
10079 | return CM_ScalarEpilogueAllowed; | ||||||
10080 | }; | ||||||
10081 | |||||||
10082 | // 4) if the TTI hook indicates this is profitable, request predication. | ||||||
10083 | if (TTI->preferPredicateOverEpilogue(L, LI, *SE, *AC, TLI, DT, | ||||||
10084 | LVL.getLAI())) | ||||||
10085 | return CM_ScalarEpilogueNotNeededUsePredicate; | ||||||
10086 | |||||||
10087 | return CM_ScalarEpilogueAllowed; | ||||||
10088 | } | ||||||
10089 | |||||||
10090 | Value *VPTransformState::get(VPValue *Def, unsigned Part) { | ||||||
10091 | // If Values have been set for this Def return the one relevant for \p Part. | ||||||
10092 | if (hasVectorValue(Def, Part)) | ||||||
10093 | return Data.PerPartOutput[Def][Part]; | ||||||
10094 | |||||||
10095 | if (!hasScalarValue(Def, {Part, 0})) { | ||||||
10096 | Value *IRV = Def->getLiveInIRValue(); | ||||||
10097 | Value *B = ILV->getBroadcastInstrs(IRV); | ||||||
10098 | set(Def, B, Part); | ||||||
10099 | return B; | ||||||
10100 | } | ||||||
10101 | |||||||
10102 | Value *ScalarValue = get(Def, {Part, 0}); | ||||||
10103 | // If we aren't vectorizing, we can just copy the scalar map values over | ||||||
10104 | // to the vector map. | ||||||
10105 | if (VF.isScalar()) { | ||||||
10106 | set(Def, ScalarValue, Part); | ||||||
10107 | return ScalarValue; | ||||||
10108 | } | ||||||
10109 | |||||||
10110 | auto *RepR = dyn_cast<VPReplicateRecipe>(Def); | ||||||
10111 | bool IsUniform = RepR && RepR->isUniform(); | ||||||
10112 | |||||||
10113 | unsigned LastLane = IsUniform ? 0 : VF.getKnownMinValue() - 1; | ||||||
10114 | // Check if there is a scalar value for the selected lane. | ||||||
10115 | if (!hasScalarValue(Def, {Part, LastLane})) { | ||||||
10116 | // At the moment, VPWidenIntOrFpInductionRecipes can also be uniform. | ||||||
10117 | assert(isa<VPWidenIntOrFpInductionRecipe>(Def->getDef()) &&(static_cast <bool> (isa<VPWidenIntOrFpInductionRecipe >(Def->getDef()) && "unexpected recipe found to be invariant" ) ? void (0) : __assert_fail ("isa<VPWidenIntOrFpInductionRecipe>(Def->getDef()) && \"unexpected recipe found to be invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10118, __extension__ __PRETTY_FUNCTION__)) | ||||||
10118 | "unexpected recipe found to be invariant")(static_cast <bool> (isa<VPWidenIntOrFpInductionRecipe >(Def->getDef()) && "unexpected recipe found to be invariant" ) ? void (0) : __assert_fail ("isa<VPWidenIntOrFpInductionRecipe>(Def->getDef()) && \"unexpected recipe found to be invariant\"" , "llvm/lib/Transforms/Vectorize/LoopVectorize.cpp", 10118, __extension__ __PRETTY_FUNCTION__)); | ||||||
10119 | IsUniform = true; | ||||||
10120 | LastLane = 0; | ||||||
10121 | } | ||||||
10122 | |||||||
10123 | auto *LastInst = cast<Instruction>(get(Def, {Part, LastLane})); | ||||||
10124 | // Set the insert point after the last scalarized instruction or after the | ||||||
10125 | // last PHI, if LastInst is a PHI. This ensures the insertelement sequence | ||||||
10126 | // will directly follow the scalar definitions. | ||||||
10127 | auto OldIP = Builder.saveIP(); | ||||||
10128 | auto NewIP = | ||||||
10129 | isa<PHINode>(LastInst) | ||||||
10130 | ? BasicBlock::iterator(LastInst->getParent()->getFirstNonPHI()) | ||||||
10131 | : std::next(BasicBlock::iterator(LastInst)); | ||||||
10132 | Builder.SetInsertPoint(&*NewIP); | ||||||
10133 | |||||||
10134 | // However, if we are vectorizing, we need to construct the vector values. | ||||||
10135 | // If the value is known to be uniform after vectorization, we can just | ||||||
10136 | // broadcast the scalar value corresponding to lane zero for each unroll | ||||||
10137 | // iteration. Otherwise, we construct the vector values using | ||||||
10138 | // insertelement instructions. Since the resulting vectors are stored in | ||||||
10139 | // State, we will only generate the insertelements once. | ||||||
10140 | Value *VectorValue = nullptr; | ||||||
10141 | if (IsUniform) { | ||||||
10142 | VectorValue = ILV->getBroadcastInstrs(ScalarValue); | ||||||
10143 | set(Def, VectorValue, Part); | ||||||
10144 | } else { | ||||||
10145 | // Initialize packing with insertelements to start from undef. | ||||||
10146 | 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", 10146, __extension__ __PRETTY_FUNCTION__)); | ||||||
10147 | Value *Undef = PoisonValue::get(VectorType::get(LastInst->getType(), VF)); | ||||||
10148 | set(Def, Undef, Part); | ||||||
10149 | for (unsigned Lane = 0; Lane < VF.getKnownMinValue(); ++Lane) | ||||||
10150 | ILV->packScalarIntoVectorValue(Def, {Part, Lane}, *this); | ||||||
10151 | VectorValue = get(Def, Part); | ||||||
10152 | } | ||||||
10153 | Builder.restoreIP(OldIP); | ||||||
10154 | return VectorValue; | ||||||
10155 | } | ||||||
10156 | |||||||
10157 | // Process the loop in the VPlan-native vectorization path. This path builds | ||||||
10158 | // VPlan upfront in the vectorization pipeline, which allows to apply | ||||||
10159 | // VPlan-to-VPlan transformations from the very beginning without modifying the | ||||||
10160 | // input LLVM IR. | ||||||
10161 | static bool processLoopInVPlanNativePath( | ||||||
10162 | Loop *L, PredicatedScalarEvolution &PSE, LoopInfo *LI, DominatorTree *DT, | ||||||
10163 | LoopVectorizationLegality *LVL, TargetTransformInfo *TTI, | ||||||
10164 | TargetLibraryInfo *TLI, DemandedBits *DB, AssumptionCache *AC, | ||||||
10165 | OptimizationRemarkEmitter *ORE, BlockFrequencyInfo *BFI, | ||||||
10166 | ProfileSummaryInfo *PSI, LoopVectorizeHints &Hints, | ||||||
10167 | LoopVectorizationRequirements &Requirements) { | ||||||
10168 | |||||||
10169 | if (isa<SCEVCouldNotCompute>(PSE.getBackedgeTakenCount())) { | ||||||
10170 | 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); | ||||||
10171 | return false; | ||||||
10172 | } | ||||||
10173 | 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", 10173, __extension__ __PRETTY_FUNCTION__)); | ||||||
10174 | Function *F = L->getHeader()->getParent(); | ||||||
10175 | InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL->getLAI()); | ||||||
10176 | |||||||
10177 | ScalarEpilogueLowering SEL = getScalarEpilogueLowering( | ||||||
10178 | F, L, Hints, PSI, BFI, TTI, TLI, AC, LI, PSE.getSE(), DT, *LVL); | ||||||
10179 | |||||||
10180 | LoopVectorizationCostModel CM(SEL, L, PSE, LI, LVL, *TTI, TLI, DB, AC, ORE, F, | ||||||
10181 | &Hints, IAI); | ||||||
10182 | // Use the planner for outer loop vectorization. | ||||||
10183 | // TODO: CM is not used at this point inside the planner. Turn CM into an | ||||||
10184 | // optional argument if we don't need it in the future. | ||||||
10185 | LoopVectorizationPlanner LVP(L, LI, TLI, TTI, LVL, CM, IAI, PSE, Hints, | ||||||
10186 | Requirements, ORE); | ||||||
10187 | |||||||
10188 | // Get user vectorization factor. | ||||||
10189 | ElementCount UserVF = Hints.getWidth(); | ||||||
10190 | |||||||
10191 | CM.collectElementTypesForWidening(); | ||||||
10192 | |||||||
10193 | // Plan how to best vectorize, return the best VF and its cost. | ||||||
10194 | const VectorizationFactor VF = LVP.planInVPlanNativePath(UserVF); | ||||||
10195 | |||||||
10196 | // If we are stress testing VPlan builds, do not attempt to generate vector | ||||||
10197 | // code. Masked vector code generation support will follow soon. | ||||||
10198 | // Also, do not attempt to vectorize if no vector code will be produced. | ||||||
10199 | if (VPlanBuildStressTest || EnableVPlanPredication || | ||||||
10200 | VectorizationFactor::Disabled() == VF) | ||||||
10201 | return false; | ||||||
10202 | |||||||
10203 | VPlan &BestPlan = LVP.getBestPlanFor(VF.Width); | ||||||
10204 | |||||||
10205 | { | ||||||
10206 | GeneratedRTChecks Checks(*PSE.getSE(), DT, LI, | ||||||
10207 | F->getParent()->getDataLayout()); | ||||||
10208 | InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, 1, LVL, | ||||||
10209 | &CM, BFI, PSI, Checks); | ||||||
10210 | 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) | ||||||
10211 | << 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); | ||||||
10212 | LVP.executePlan(VF.Width, 1, BestPlan, LB, DT); | ||||||
10213 | } | ||||||
10214 | |||||||
10215 | // Mark the loop as already vectorized to avoid vectorizing again. | ||||||
10216 | Hints.setAlreadyVectorized(); | ||||||
10217 | 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", 10217, __extension__ __PRETTY_FUNCTION__)); | ||||||
10218 | return true; | ||||||
10219 | } | ||||||
10220 | |||||||
10221 | // Emit a remark if there are stores to floats that required a floating point | ||||||
10222 | // extension. If the vectorized loop was generated with floating point there | ||||||
10223 | // will be a performance penalty from the conversion overhead and the change in | ||||||
10224 | // the vector width. | ||||||
10225 | static void checkMixedPrecision(Loop *L, OptimizationRemarkEmitter *ORE) { | ||||||
10226 | SmallVector<Instruction *, 4> Worklist; | ||||||
10227 | for (BasicBlock *BB : L->getBlocks()) { | ||||||
10228 | for (Instruction &Inst : *BB) { | ||||||
10229 | if (auto *S = dyn_cast<StoreInst>(&Inst)) { | ||||||
10230 | if (S->getValueOperand()->getType()->isFloatTy()) | ||||||
10231 | Worklist.push_back(S); | ||||||
10232 | } | ||||||
10233 | } | ||||||
10234 | } | ||||||
10235 | |||||||
10236 | // Traverse the floating point stores upwards searching, for floating point | ||||||
10237 | // conversions. | ||||||
10238 | SmallPtrSet<const Instruction *, 4> Visited; | ||||||
10239 | SmallPtrSet<const Instruction *, 4> EmittedRemark; | ||||||
10240 | while (!Worklist.empty()) { | ||||||
10241 | auto *I = Worklist.pop_back_val(); | ||||||
10242 | if (!L->contains(I)) | ||||||
10243 | continue; | ||||||
10244 | if (!Visited.insert(I).second) | ||||||
10245 | continue; | ||||||
10246 | |||||||
10247 | // Emit a remark if the floating point store required a floating | ||||||
10248 | // point conversion. | ||||||
10249 | // TODO: More work could be done to identify the root cause such as a | ||||||
10250 | // constant or a function return type and point the user to it. | ||||||
10251 | if (isa<FPExtInst>(I) && EmittedRemark.insert(I).second) | ||||||
10252 | ORE->emit([&]() { | ||||||
10253 | return OptimizationRemarkAnalysis(LV_NAME"loop-vectorize", "VectorMixedPrecision", | ||||||
10254 | I->getDebugLoc(), L->getHeader()) | ||||||
10255 | << "floating point conversion changes vector width. " | ||||||
10256 | << "Mixed floating point precision requires an up/down " | ||||||
10257 | << "cast that will negatively impact performance."; | ||||||
10258 | }); | ||||||
10259 | |||||||
10260 | for (Use &Op : I->operands()) | ||||||
10261 | if (auto *OpI = dyn_cast<Instruction>(Op)) | ||||||
10262 | Worklist.push_back(OpI); | ||||||
10263 | } | ||||||
10264 | } | ||||||
10265 | |||||||
10266 | LoopVectorizePass::LoopVectorizePass(LoopVectorizeOptions Opts) | ||||||
10267 | : InterleaveOnlyWhenForced(Opts.InterleaveOnlyWhenForced || | ||||||
10268 | !EnableLoopInterleaving), | ||||||
10269 | VectorizeOnlyWhenForced(Opts.VectorizeOnlyWhenForced || | ||||||
10270 | !EnableLoopVectorization) {} | ||||||
10271 | |||||||
10272 | bool LoopVectorizePass::processLoop(Loop *L) { | ||||||
10273 | 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", 10274, __extension__ __PRETTY_FUNCTION__)) | ||||||
10274 | "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", 10274, __extension__ __PRETTY_FUNCTION__)); | ||||||
10275 | |||||||
10276 | #ifndef NDEBUG | ||||||
10277 | const std::string DebugLocStr = getDebugLocString(L); | ||||||
10278 | #endif /* NDEBUG */ | ||||||
10279 | |||||||
10280 | 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 ) | ||||||
10281 | << 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 ) | ||||||
10282 | << 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 ); | ||||||
10283 | |||||||
10284 | LoopVectorizeHints Hints(L, InterleaveOnlyWhenForced, *ORE, TTI); | ||||||
10285 | |||||||
10286 | 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) | ||||||
10287 | 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) | ||||||
10288 | << " 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) | ||||||
10289 | << (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) | ||||||
10290 | ? "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) | ||||||
10291 | : (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) | ||||||
10292 | ? "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) | ||||||
10293 | : "?"))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) | ||||||
10294 | << " 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) | ||||||
10295 | << " 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); | ||||||
10296 | |||||||
10297 | // Function containing loop | ||||||
10298 | Function *F = L->getHeader()->getParent(); | ||||||
10299 | |||||||
10300 | // Looking at the diagnostic output is the only way to determine if a loop | ||||||
10301 | // was vectorized (other than looking at the IR or machine code), so it | ||||||
10302 | // is important to generate an optimization remark for each loop. Most of | ||||||
10303 | // these messages are generated as OptimizationRemarkAnalysis. Remarks | ||||||
10304 | // generated as OptimizationRemark and OptimizationRemarkMissed are | ||||||
10305 | // less verbose reporting vectorized loops and unvectorized loops that may | ||||||
10306 | // benefit from vectorization, respectively. | ||||||
10307 | |||||||
10308 | if (!Hints.allowVectorization(F, L, VectorizeOnlyWhenForced)) { | ||||||
10309 | 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); | ||||||
10310 | return false; | ||||||
10311 | } | ||||||
10312 | |||||||
10313 | PredicatedScalarEvolution PSE(*SE, *L); | ||||||
10314 | |||||||
10315 | // Check if it is legal to vectorize the loop. | ||||||
10316 | LoopVectorizationRequirements Requirements; | ||||||
10317 | LoopVectorizationLegality LVL(L, PSE, DT, TTI, TLI, AA, F, GetLAA, LI, ORE, | ||||||
10318 | &Requirements, &Hints, DB, AC, BFI, PSI); | ||||||
10319 | if (!LVL.canVectorize(EnableVPlanNativePath)) { | ||||||
10320 | 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); | ||||||
10321 | Hints.emitRemarkWithHints(); | ||||||
10322 | return false; | ||||||
10323 | } | ||||||
10324 | |||||||
10325 | // Check the function attributes and profiles to find out if this function | ||||||
10326 | // should be optimized for size. | ||||||
10327 | ScalarEpilogueLowering SEL = getScalarEpilogueLowering( | ||||||
10328 | F, L, Hints, PSI, BFI, TTI, TLI, AC, LI, PSE.getSE(), DT, LVL); | ||||||
10329 | |||||||
10330 | // Entrance to the VPlan-native vectorization path. Outer loops are processed | ||||||
10331 | // here. They may require CFG and instruction level transformations before | ||||||
10332 | // even evaluating whether vectorization is profitable. Since we cannot modify | ||||||
10333 | // the incoming IR, we need to build VPlan upfront in the vectorization | ||||||
10334 | // pipeline. | ||||||
10335 | if (!L->isInnermost()) | ||||||
10336 | return processLoopInVPlanNativePath(L, PSE, LI, DT, &LVL, TTI, TLI, DB, AC, | ||||||
10337 | ORE, BFI, PSI, Hints, Requirements); | ||||||
10338 | |||||||
10339 | 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", 10339, __extension__ __PRETTY_FUNCTION__)); | ||||||
10340 | |||||||
10341 | // Check the loop for a trip count threshold: vectorize loops with a tiny trip | ||||||
10342 | // count by optimizing for size, to minimize overheads. | ||||||
10343 | auto ExpectedTC = getSmallBestKnownTC(*SE, L); | ||||||
10344 | if (ExpectedTC && *ExpectedTC < TinyTripCountVectorThreshold) { | ||||||
10345 | 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 ) | ||||||
10346 | << "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 ) | ||||||
10347 | << "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 ); | ||||||
10348 | if (Hints.getForce() == LoopVectorizeHints::FK_Enabled) | ||||||
10349 | 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); | ||||||
10350 | else { | ||||||
10351 | LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "\n"; } } while (false); | ||||||
10352 | SEL = CM_ScalarEpilogueNotAllowedLowTripLoop; | ||||||
10353 | } | ||||||
10354 | } | ||||||
10355 | |||||||
10356 | // Check the function attributes to see if implicit floats are allowed. | ||||||
10357 | // FIXME: This check doesn't seem possibly correct -- what if the loop is | ||||||
10358 | // an integer loop and the vector instructions selected are purely integer | ||||||
10359 | // vector instructions? | ||||||
10360 | if (F->hasFnAttribute(Attribute::NoImplicitFloat)) { | ||||||
10361 | reportVectorizationFailure( | ||||||
10362 | "Can't vectorize when the NoImplicitFloat attribute is used", | ||||||
10363 | "loop not vectorized due to NoImplicitFloat attribute", | ||||||
10364 | "NoImplicitFloat", ORE, L); | ||||||
10365 | Hints.emitRemarkWithHints(); | ||||||
10366 | return false; | ||||||
10367 | } | ||||||
10368 | |||||||
10369 | // Check if the target supports potentially unsafe FP vectorization. | ||||||
10370 | // FIXME: Add a check for the type of safety issue (denormal, signaling) | ||||||
10371 | // for the target we're vectorizing for, to make sure none of the | ||||||
10372 | // additional fp-math flags can help. | ||||||
10373 | if (Hints.isPotentiallyUnsafe() && | ||||||
10374 | TTI->isFPVectorizationPotentiallyUnsafe()) { | ||||||
10375 | reportVectorizationFailure( | ||||||
10376 | "Potentially unsafe FP op prevents vectorization", | ||||||
10377 | "loop not vectorized due to unsafe FP support.", | ||||||
10378 | "UnsafeFP", ORE, L); | ||||||
10379 | Hints.emitRemarkWithHints(); | ||||||
10380 | return false; | ||||||
10381 | } | ||||||
10382 | |||||||
10383 | bool AllowOrderedReductions; | ||||||
10384 | // If the flag is set, use that instead and override the TTI behaviour. | ||||||
10385 | if (ForceOrderedReductions.getNumOccurrences() > 0) | ||||||
10386 | AllowOrderedReductions = ForceOrderedReductions; | ||||||
10387 | else | ||||||
10388 | AllowOrderedReductions = TTI->enableOrderedReductions(); | ||||||
10389 | if (!LVL.canVectorizeFPMath(AllowOrderedReductions)) { | ||||||
10390 | ORE->emit([&]() { | ||||||
10391 | auto *ExactFPMathInst = Requirements.getExactFPInst(); | ||||||
10392 | return OptimizationRemarkAnalysisFPCommute(DEBUG_TYPE"loop-vectorize", "CantReorderFPOps", | ||||||
10393 | ExactFPMathInst->getDebugLoc(), | ||||||
10394 | ExactFPMathInst->getParent()) | ||||||
10395 | << "loop not vectorized: cannot prove it is safe to reorder " | ||||||
10396 | "floating-point operations"; | ||||||
10397 | }); | ||||||
10398 | 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) | ||||||
10399 | "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); | ||||||
10400 | Hints.emitRemarkWithHints(); | ||||||
10401 | return false; | ||||||
10402 | } | ||||||
10403 | |||||||
10404 | bool UseInterleaved = TTI->enableInterleavedAccessVectorization(); | ||||||
10405 | InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL.getLAI()); | ||||||
10406 | |||||||
10407 | // If an override option has been passed in for interleaved accesses, use it. | ||||||
10408 | if (EnableInterleavedMemAccesses.getNumOccurrences() > 0) | ||||||
10409 | UseInterleaved = EnableInterleavedMemAccesses; | ||||||
10410 | |||||||
10411 | // Analyze interleaved memory accesses. | ||||||
10412 | if (UseInterleaved) { | ||||||
10413 | IAI.analyzeInterleaving(useMaskedInterleavedAccesses(*TTI)); | ||||||
10414 | } | ||||||
10415 | |||||||
10416 | // Use the cost model. | ||||||
10417 | LoopVectorizationCostModel CM(SEL, L, PSE, LI, &LVL, *TTI, TLI, DB, AC, ORE, | ||||||
10418 | F, &Hints, IAI); | ||||||
10419 | CM.collectValuesToIgnore(); | ||||||
10420 | CM.collectElementTypesForWidening(); | ||||||
10421 | |||||||
10422 | // Use the planner for vectorization. | ||||||
10423 | LoopVectorizationPlanner LVP(L, LI, TLI, TTI, &LVL, CM, IAI, PSE, Hints, | ||||||
10424 | Requirements, ORE); | ||||||
10425 | |||||||
10426 | // Get user vectorization factor and interleave count. | ||||||
10427 | ElementCount UserVF = Hints.getWidth(); | ||||||
10428 | unsigned UserIC = Hints.getInterleave(); | ||||||
10429 | |||||||
10430 | // Plan how to best vectorize, return the best VF and its cost. | ||||||
10431 | Optional<VectorizationFactor> MaybeVF = LVP.plan(UserVF, UserIC); | ||||||
10432 | |||||||
10433 | VectorizationFactor VF = VectorizationFactor::Disabled(); | ||||||
10434 | unsigned IC = 1; | ||||||
10435 | |||||||
10436 | if (MaybeVF) { | ||||||
10437 | VF = *MaybeVF; | ||||||
10438 | // Select the interleave count. | ||||||
10439 | IC = CM.selectInterleaveCount(VF.Width, *VF.Cost.getValue()); | ||||||
10440 | } | ||||||
10441 | |||||||
10442 | // Identify the diagnostic messages that should be produced. | ||||||
10443 | std::pair<StringRef, std::string> VecDiagMsg, IntDiagMsg; | ||||||
10444 | bool VectorizeLoop = true, InterleaveLoop = true; | ||||||
10445 | if (VF.Width.isScalar()) { | ||||||
10446 | 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); | ||||||
10447 | VecDiagMsg = std::make_pair( | ||||||
10448 | "VectorizationNotBeneficial", | ||||||
10449 | "the cost-model indicates that vectorization is not beneficial"); | ||||||
10450 | VectorizeLoop = false; | ||||||
10451 | } | ||||||
10452 | |||||||
10453 | if (!MaybeVF && UserIC > 1) { | ||||||
10454 | // Tell the user interleaving was avoided up-front, despite being explicitly | ||||||
10455 | // requested. | ||||||
10456 | 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 ) | ||||||
10457 | "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 ); | ||||||
10458 | IntDiagMsg = std::make_pair( | ||||||
10459 | "InterleavingAvoided", | ||||||
10460 | "Ignoring UserIC, because interleaving was avoided up front"); | ||||||
10461 | InterleaveLoop = false; | ||||||
10462 | } else if (IC == 1 && UserIC <= 1) { | ||||||
10463 | // Tell the user interleaving is not beneficial. | ||||||
10464 | 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); | ||||||
10465 | IntDiagMsg = std::make_pair( | ||||||
10466 | "InterleavingNotBeneficial", | ||||||
10467 | "the cost-model indicates that interleaving is not beneficial"); | ||||||
10468 | InterleaveLoop = false; | ||||||
10469 | if (UserIC == 1) { | ||||||
10470 | IntDiagMsg.first = "InterleavingNotBeneficialAndDisabled"; | ||||||
10471 | IntDiagMsg.second += | ||||||
10472 | " and is explicitly disabled or interleave count is set to 1"; | ||||||
10473 | } | ||||||
10474 | } else if (IC > 1 && UserIC == 1) { | ||||||
10475 | // Tell the user interleaving is beneficial, but it explicitly disabled. | ||||||
10476 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Interleaving is beneficial but is explicitly disabled." ; } } while (false) | ||||||
10477 | 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); | ||||||
10478 | IntDiagMsg = std::make_pair( | ||||||
10479 | "InterleavingBeneficialButDisabled", | ||||||
10480 | "the cost-model indicates that interleaving is beneficial " | ||||||
10481 | "but is explicitly disabled or interleave count is set to 1"); | ||||||
10482 | InterleaveLoop = false; | ||||||
10483 | } | ||||||
10484 | |||||||
10485 | // Override IC if user provided an interleave count. | ||||||
10486 | IC = UserIC > 0 ? UserIC : IC; | ||||||
10487 | |||||||
10488 | // Emit diagnostic messages, if any. | ||||||
10489 | const char *VAPassName = Hints.vectorizeAnalysisPassName(); | ||||||
10490 | if (!VectorizeLoop && !InterleaveLoop) { | ||||||
10491 | // Do not vectorize or interleaving the loop. | ||||||
10492 | ORE->emit([&]() { | ||||||
10493 | return OptimizationRemarkMissed(VAPassName, VecDiagMsg.first, | ||||||
10494 | L->getStartLoc(), L->getHeader()) | ||||||
10495 | << VecDiagMsg.second; | ||||||
10496 | }); | ||||||
10497 | ORE->emit([&]() { | ||||||
10498 | return OptimizationRemarkMissed(LV_NAME"loop-vectorize", IntDiagMsg.first, | ||||||
10499 | L->getStartLoc(), L->getHeader()) | ||||||
10500 | << IntDiagMsg.second; | ||||||
10501 | }); | ||||||
10502 | return false; | ||||||
10503 | } else if (!VectorizeLoop && InterleaveLoop) { | ||||||
10504 | 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); | ||||||
10505 | ORE->emit([&]() { | ||||||
10506 | return OptimizationRemarkAnalysis(VAPassName, VecDiagMsg.first, | ||||||
10507 | L->getStartLoc(), L->getHeader()) | ||||||
10508 | << VecDiagMsg.second; | ||||||
10509 | }); | ||||||
10510 | } else if (VectorizeLoop && !InterleaveLoop) { | ||||||
10511 | 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) | ||||||
10512 | << ") in " << DebugLocStr << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " << DebugLocStr << '\n'; } } while (false); | ||||||
10513 | ORE->emit([&]() { | ||||||
10514 | return OptimizationRemarkAnalysis(LV_NAME"loop-vectorize", IntDiagMsg.first, | ||||||
10515 | L->getStartLoc(), L->getHeader()) | ||||||
10516 | << IntDiagMsg.second; | ||||||
10517 | }); | ||||||
10518 | } else if (VectorizeLoop && InterleaveLoop) { | ||||||
10519 | 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) | ||||||
10520 | << ") in " << DebugLocStr << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("loop-vectorize")) { dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " << DebugLocStr << '\n'; } } while (false); | ||||||
10521 | 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); | ||||||
10522 | } | ||||||
10523 | |||||||
10524 | bool DisableRuntimeUnroll = false; | ||||||
10525 | MDNode *OrigLoopID = L->getLoopID(); | ||||||
10526 | { | ||||||
10527 | // Optimistically generate runtime checks. Drop them if they turn out to not | ||||||
10528 | // be profitable. Limit the scope of Checks, so the cleanup happens | ||||||
10529 | // immediately after vector codegeneration is done. | ||||||
10530 | GeneratedRTChecks Checks(*PSE.getSE(), DT, LI, | ||||||
10531 | F->getParent()->getDataLayout()); | ||||||
10532 | if (!VF.Width.isScalar() || IC > 1) | ||||||
10533 | Checks.Create(L, *LVL.getLAI(), PSE.getUnionPredicate()); | ||||||
10534 | |||||||
10535 | using namespace ore; | ||||||
10536 | if (!VectorizeLoop) { | ||||||
10537 | 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", 10537, __extension__ __PRETTY_FUNCTION__)); | ||||||
10538 | // If we decided that it is not legal to vectorize the loop, then | ||||||
10539 | // interleave it. | ||||||
10540 | InnerLoopUnroller Unroller(L, PSE, LI, DT, TLI, TTI, AC, ORE, IC, &LVL, | ||||||
10541 | &CM, BFI, PSI, Checks); | ||||||
10542 | |||||||
10543 | VPlan &BestPlan = LVP.getBestPlanFor(VF.Width); | ||||||
10544 | LVP.executePlan(VF.Width, IC, BestPlan, Unroller, DT); | ||||||
10545 | |||||||
10546 | ORE->emit([&]() { | ||||||
10547 | return OptimizationRemark(LV_NAME"loop-vectorize", "Interleaved", L->getStartLoc(), | ||||||
10548 | L->getHeader()) | ||||||
10549 | << "interleaved loop (interleaved count: " | ||||||
10550 | << NV("InterleaveCount", IC) << ")"; | ||||||
10551 | }); | ||||||
10552 | } else { | ||||||
10553 | // If we decided that it is *legal* to vectorize the loop, then do it. | ||||||
10554 | |||||||
10555 | // Consider vectorizing the epilogue too if it's profitable. | ||||||
10556 | VectorizationFactor EpilogueVF = | ||||||
10557 | CM.selectEpilogueVectorizationFactor(VF.Width, LVP); | ||||||
10558 | if (EpilogueVF.Width.isVector()) { | ||||||
10559 | |||||||
10560 | // The first pass vectorizes the main loop and creates a scalar epilogue | ||||||
10561 | // to be vectorized by executing the plan (potentially with a different | ||||||
10562 | // factor) again shortly afterwards. | ||||||
10563 | EpilogueLoopVectorizationInfo EPI(VF.Width, IC, EpilogueVF.Width, 1); | ||||||
10564 | EpilogueVectorizerMainLoop MainILV(L, PSE, LI, DT, TLI, TTI, AC, ORE, | ||||||
10565 | EPI, &LVL, &CM, BFI, PSI, Checks); | ||||||
10566 | |||||||
10567 | VPlan &BestMainPlan = LVP.getBestPlanFor(EPI.MainLoopVF); | ||||||
10568 | LVP.executePlan(EPI.MainLoopVF, EPI.MainLoopUF, BestMainPlan, MainILV, | ||||||
10569 | DT); | ||||||
10570 | ++LoopsVectorized; | ||||||
10571 | |||||||
10572 | simplifyLoop(L, DT, LI, SE, AC, nullptr, false /* PreserveLCSSA */); | ||||||
10573 | formLCSSARecursively(*L, *DT, LI, SE); | ||||||
10574 | |||||||
10575 | // Second pass vectorizes the epilogue and adjusts the control flow | ||||||
10576 | // edges from the first pass. | ||||||
10577 | EPI.MainLoopVF = EPI.EpilogueVF; | ||||||
10578 | EPI.MainLoopUF = EPI.EpilogueUF; | ||||||
10579 | EpilogueVectorizerEpilogueLoop EpilogILV(L, PSE, LI, DT, TLI, TTI, AC, | ||||||
10580 | ORE, EPI, &LVL, &CM, BFI, PSI, | ||||||
10581 | Checks); | ||||||
10582 | |||||||
10583 | VPlan &BestEpiPlan = LVP.getBestPlanFor(EPI.EpilogueVF); | ||||||
10584 | LVP.executePlan(EPI.EpilogueVF, EPI.EpilogueUF, BestEpiPlan, EpilogILV, | ||||||
10585 | DT); | ||||||
10586 | ++LoopsEpilogueVectorized; | ||||||
10587 | |||||||
10588 | if (!MainILV.areSafetyChecksAdded()) | ||||||
10589 | DisableRuntimeUnroll = true; | ||||||
10590 | } else { | ||||||
10591 | InnerLoopVectorizer LB(L, PSE, LI, DT, TLI, TTI, AC, ORE, VF.Width, IC, | ||||||
10592 | &LVL, &CM, BFI, PSI, Checks); | ||||||
10593 | |||||||
10594 | VPlan &BestPlan = LVP.getBestPlanFor(VF.Width); | ||||||
10595 | LVP.executePlan(VF.Width, IC, BestPlan, LB, DT); | ||||||
10596 | ++LoopsVectorized; | ||||||
10597 | |||||||
10598 | // Add metadata to disable runtime unrolling a scalar loop when there | ||||||
10599 | // are no runtime checks about strides and memory. A scalar loop that is | ||||||
10600 | // rarely used is not worth unrolling. | ||||||
10601 | if (!LB.areSafetyChecksAdded()) | ||||||
10602 | DisableRuntimeUnroll = true; | ||||||
10603 | } | ||||||
10604 | // Report the vectorization decision. | ||||||
10605 | ORE->emit([&]() { | ||||||
10606 | return OptimizationRemark(LV_NAME"loop-vectorize", "Vectorized", L->getStartLoc(), | ||||||
10607 | L->getHeader()) | ||||||
10608 | << "vectorized loop (vectorization width: " | ||||||
10609 | << NV("VectorizationFactor", VF.Width) | ||||||
10610 | << ", interleaved count: " << NV("InterleaveCount", IC) << ")"; | ||||||
10611 | }); | ||||||
10612 | } | ||||||
10613 | |||||||
10614 | if (ORE->allowExtraAnalysis(LV_NAME"loop-vectorize")) | ||||||
10615 | checkMixedPrecision(L, ORE); | ||||||
10616 | } | ||||||
10617 | |||||||
10618 | Optional<MDNode *> RemainderLoopID = | ||||||
10619 | makeFollowupLoopID(OrigLoopID, {LLVMLoopVectorizeFollowupAll, | ||||||
10620 | LLVMLoopVectorizeFollowupEpilogue}); | ||||||
10621 | if (RemainderLoopID.hasValue()) { | ||||||
10622 | L->setLoopID(RemainderLoopID.getValue()); | ||||||
10623 | } else { | ||||||
10624 | if (DisableRuntimeUnroll) | ||||||
10625 | AddRuntimeUnrollDisableMetaData(L); | ||||||
10626 | |||||||
10627 | // Mark the loop as already vectorized to avoid vectorizing again. | ||||||
10628 | Hints.setAlreadyVectorized(); | ||||||
10629 | } | ||||||
10630 | |||||||
10631 | 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", 10631, __extension__ __PRETTY_FUNCTION__)); | ||||||
10632 | return true; | ||||||
10633 | } | ||||||
10634 | |||||||
10635 | LoopVectorizeResult LoopVectorizePass::runImpl( | ||||||
10636 | Function &F, ScalarEvolution &SE_, LoopInfo &LI_, TargetTransformInfo &TTI_, | ||||||
10637 | DominatorTree &DT_, BlockFrequencyInfo &BFI_, TargetLibraryInfo *TLI_, | ||||||
10638 | DemandedBits &DB_, AAResults &AA_, AssumptionCache &AC_, | ||||||
10639 | std::function<const LoopAccessInfo &(Loop &)> &GetLAA_, | ||||||
10640 | OptimizationRemarkEmitter &ORE_, ProfileSummaryInfo *PSI_) { | ||||||
10641 | SE = &SE_; | ||||||
10642 | LI = &LI_; | ||||||
10643 | TTI = &TTI_; | ||||||
10644 | DT = &DT_; | ||||||
10645 | BFI = &BFI_; | ||||||
10646 | TLI = TLI_; | ||||||
10647 | AA = &AA_; | ||||||
10648 | AC = &AC_; | ||||||
10649 | GetLAA = &GetLAA_; | ||||||
10650 | DB = &DB_; | ||||||
10651 | ORE = &ORE_; | ||||||
10652 | PSI = PSI_; | ||||||
10653 | |||||||
10654 | // Don't attempt if | ||||||
10655 | // 1. the target claims to have no vector registers, and | ||||||
10656 | // 2. interleaving won't help ILP. | ||||||
10657 | // | ||||||
10658 | // The second condition is necessary because, even if the target has no | ||||||
10659 | // vector registers, loop vectorization may still enable scalar | ||||||
10660 | // interleaving. | ||||||
10661 | if (!TTI->getNumberOfRegisters(TTI->getRegisterClassForType(true)) && | ||||||
10662 | TTI->getMaxInterleaveFactor(1) < 2) | ||||||
10663 | return LoopVectorizeResult(false, false); | ||||||
10664 | |||||||
10665 | bool Changed = false, CFGChanged = false; | ||||||
10666 | |||||||
10667 | // The vectorizer requires loops to be in simplified form. | ||||||
10668 | // Since simplification may add new inner loops, it has to run before the | ||||||
10669 | // legality and profitability checks. This means running the loop vectorizer | ||||||
10670 | // will simplify all loops, regardless of whether anything end up being | ||||||
10671 | // vectorized. | ||||||
10672 | for (auto &L : *LI) | ||||||
10673 | Changed |= CFGChanged |= | ||||||
10674 | simplifyLoop(L, DT, LI, SE, AC, nullptr, false /* PreserveLCSSA */); | ||||||
10675 | |||||||
10676 | // Build up a worklist of inner-loops to vectorize. This is necessary as | ||||||
10677 | // the act of vectorizing or partially unrolling a loop creates new loops | ||||||
10678 | // and can invalidate iterators across the loops. | ||||||
10679 | SmallVector<Loop *, 8> Worklist; | ||||||
10680 | |||||||
10681 | for (Loop *L : *LI) | ||||||
10682 | collectSupportedLoops(*L, LI, ORE, Worklist); | ||||||
10683 | |||||||
10684 | LoopsAnalyzed += Worklist.size(); | ||||||
10685 | |||||||
10686 | // Now walk the identified inner loops. | ||||||
10687 | while (!Worklist.empty()) { | ||||||
10688 | Loop *L = Worklist.pop_back_val(); | ||||||
10689 | |||||||
10690 | // For the inner loops we actually process, form LCSSA to simplify the | ||||||
10691 | // transform. | ||||||
10692 | Changed |= formLCSSARecursively(*L, *DT, LI, SE); | ||||||
10693 | |||||||
10694 | Changed |= CFGChanged |= processLoop(L); | ||||||
10695 | } | ||||||
10696 | |||||||
10697 | // Process each loop nest in the function. | ||||||
10698 | return LoopVectorizeResult(Changed, CFGChanged); | ||||||
10699 | } | ||||||
10700 | |||||||
10701 | PreservedAnalyses LoopVectorizePass::run(Function &F, | ||||||
10702 | FunctionAnalysisManager &AM) { | ||||||
10703 | auto &SE = AM.getResult<ScalarEvolutionAnalysis>(F); | ||||||
10704 | auto &LI = AM.getResult<LoopAnalysis>(F); | ||||||
10705 | auto &TTI = AM.getResult<TargetIRAnalysis>(F); | ||||||
10706 | auto &DT = AM.getResult<DominatorTreeAnalysis>(F); | ||||||
10707 | auto &BFI = AM.getResult<BlockFrequencyAnalysis>(F); | ||||||
10708 | auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); | ||||||
10709 | auto &AA = AM.getResult<AAManager>(F); | ||||||
10710 | auto &AC = AM.getResult<AssumptionAnalysis>(F); | ||||||
10711 | auto &DB = AM.getResult<DemandedBitsAnalysis>(F); | ||||||
10712 | auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F); | ||||||
10713 | |||||||
10714 | auto &LAM = AM.getResult<LoopAnalysisManagerFunctionProxy>(F).getManager(); | ||||||
10715 | std::function<const LoopAccessInfo &(Loop &)> GetLAA = | ||||||
10716 | [&](Loop &L) -> const LoopAccessInfo & { | ||||||
10717 | LoopStandardAnalysisResults AR = {AA, AC, DT, LI, SE, | ||||||
10718 | TLI, TTI, nullptr, nullptr, nullptr}; | ||||||
10719 | return LAM.getResult<LoopAccessAnalysis>(L, AR); | ||||||
10720 | }; | ||||||
10721 | auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F); | ||||||
10722 | ProfileSummaryInfo *PSI = | ||||||
10723 | MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent()); | ||||||
10724 | LoopVectorizeResult Result = | ||||||
10725 | runImpl(F, SE, LI, TTI, DT, BFI, &TLI, DB, AA, AC, GetLAA, ORE, PSI); | ||||||
10726 | if (!Result.MadeAnyChange) | ||||||
10727 | return PreservedAnalyses::all(); | ||||||
10728 | PreservedAnalyses PA; | ||||||
10729 | |||||||
10730 | // We currently do not preserve loopinfo/dominator analyses with outer loop | ||||||
10731 | // vectorization. Until this is addressed, mark these analyses as preserved | ||||||
10732 | // only for non-VPlan-native path. | ||||||
10733 | // TODO: Preserve Loop and Dominator analyses for VPlan-native path. | ||||||
10734 | if (!EnableVPlanNativePath) { | ||||||
10735 | PA.preserve<LoopAnalysis>(); | ||||||
10736 | PA.preserve<DominatorTreeAnalysis>(); | ||||||
10737 | } | ||||||
10738 | |||||||
10739 | if (Result.MadeCFGChange) { | ||||||
10740 | // Making CFG changes likely means a loop got vectorized. Indicate that | ||||||
10741 | // extra simplification passes should be run. | ||||||
10742 | // TODO: MadeCFGChanges is not a prefect proxy. Extra passes should only | ||||||
10743 | // be run if runtime checks have been added. | ||||||
10744 | AM.getResult<ShouldRunExtraVectorPasses>(F); | ||||||
10745 | PA.preserve<ShouldRunExtraVectorPasses>(); | ||||||
10746 | } else { | ||||||
10747 | PA.preserveSet<CFGAnalyses>(); | ||||||
10748 | } | ||||||
10749 | return PA; | ||||||
10750 | } | ||||||
10751 | |||||||
10752 | void LoopVectorizePass::printPipeline( | ||||||
10753 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { | ||||||
10754 | static_cast<PassInfoMixin<LoopVectorizePass> *>(this)->printPipeline( | ||||||
10755 | OS, MapClassName2PassName); | ||||||
10756 | |||||||
10757 | OS << "<"; | ||||||
10758 | OS << (InterleaveOnlyWhenForced ? "" : "no-") << "interleave-forced-only;"; | ||||||
10759 | OS << (VectorizeOnlyWhenForced ? "" : "no-") << "vectorize-forced-only;"; | ||||||
10760 | OS << ">"; | ||||||
10761 | } |
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 __cplusplus201402L >= 201103L |
44 | #include <initializer_list> |
45 | #endif |
46 | |
47 | #if __cplusplus201402L >= 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 __cplusplus201402L < 201103L |
106 | typedef iterator __const_iterator; |
107 | #else |
108 | typedef const_iterator __const_iterator; |
109 | #endif |
110 | |
111 | private: |
112 | #if __cplusplus201402L >= 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 __cplusplus201402L < 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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_guides && ! 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_guides && ! 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L > 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 __cplusplus201402L < 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 __cplusplus201402L >= 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_guides && ! 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L > 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_guides >= 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 __cplusplus201402L >= 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_CONSTEXPR (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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 __cplusplus201402L >= 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 "VPlanLoopInfo.h" | ||||||||||||
29 | #include "VPlanValue.h" | ||||||||||||
30 | #include "llvm/ADT/DenseMap.h" | ||||||||||||
31 | #include "llvm/ADT/DepthFirstIterator.h" | ||||||||||||
32 | #include "llvm/ADT/GraphTraits.h" | ||||||||||||
33 | #include "llvm/ADT/Optional.h" | ||||||||||||
34 | #include "llvm/ADT/SmallBitVector.h" | ||||||||||||
35 | #include "llvm/ADT/SmallPtrSet.h" | ||||||||||||
36 | #include "llvm/ADT/SmallSet.h" | ||||||||||||
37 | #include "llvm/ADT/SmallVector.h" | ||||||||||||
38 | #include "llvm/ADT/Twine.h" | ||||||||||||
39 | #include "llvm/ADT/ilist.h" | ||||||||||||
40 | #include "llvm/ADT/ilist_node.h" | ||||||||||||
41 | #include "llvm/Analysis/VectorUtils.h" | ||||||||||||
42 | #include "llvm/IR/DebugLoc.h" | ||||||||||||
43 | #include "llvm/IR/IRBuilder.h" | ||||||||||||
44 | #include "llvm/Support/InstructionCost.h" | ||||||||||||
45 | #include <algorithm> | ||||||||||||
46 | #include <cassert> | ||||||||||||
47 | #include <cstddef> | ||||||||||||
48 | #include <map> | ||||||||||||
49 | #include <string> | ||||||||||||
50 | |||||||||||||
51 | namespace llvm { | ||||||||||||
52 | |||||||||||||
53 | class BasicBlock; | ||||||||||||
54 | class DominatorTree; | ||||||||||||
55 | class InductionDescriptor; | ||||||||||||
56 | class InnerLoopVectorizer; | ||||||||||||
57 | class LoopInfo; | ||||||||||||
58 | class raw_ostream; | ||||||||||||
59 | class RecurrenceDescriptor; | ||||||||||||
60 | class Value; | ||||||||||||
61 | class VPBasicBlock; | ||||||||||||
62 | class VPRegionBlock; | ||||||||||||
63 | class VPlan; | ||||||||||||
64 | class VPReplicateRecipe; | ||||||||||||
65 | class VPlanSlp; | ||||||||||||
66 | |||||||||||||
67 | /// Returns a calculation for the total number of elements for a given \p VF. | ||||||||||||
68 | /// For fixed width vectors this value is a constant, whereas for scalable | ||||||||||||
69 | /// vectors it is an expression determined at runtime. | ||||||||||||
70 | Value *getRuntimeVF(IRBuilder<> &B, Type *Ty, ElementCount VF); | ||||||||||||
71 | |||||||||||||
72 | /// Return a value for Step multiplied by VF. | ||||||||||||
73 | Value *createStepForVF(IRBuilder<> &B, Type *Ty, ElementCount VF, int64_t Step); | ||||||||||||
74 | |||||||||||||
75 | /// A range of powers-of-2 vectorization factors with fixed start and | ||||||||||||
76 | /// adjustable end. The range includes start and excludes end, e.g.,: | ||||||||||||
77 | /// [1, 9) = {1, 2, 4, 8} | ||||||||||||
78 | struct VFRange { | ||||||||||||
79 | // A power of 2. | ||||||||||||
80 | const ElementCount Start; | ||||||||||||
81 | |||||||||||||
82 | // Need not be a power of 2. If End <= Start range is empty. | ||||||||||||
83 | ElementCount End; | ||||||||||||
84 | |||||||||||||
85 | bool isEmpty() const { | ||||||||||||
86 | return End.getKnownMinValue() <= Start.getKnownMinValue(); | ||||||||||||
87 | } | ||||||||||||
88 | |||||||||||||
89 | VFRange(const ElementCount &Start, const ElementCount &End) | ||||||||||||
90 | : Start(Start), End(End) { | ||||||||||||
91 | 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", 92, __extension__ __PRETTY_FUNCTION__ )) | ||||||||||||
92 | "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", 92, __extension__ __PRETTY_FUNCTION__ )); | ||||||||||||
93 | 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", 94, __extension__ __PRETTY_FUNCTION__ )) | ||||||||||||
94 | "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", 94, __extension__ __PRETTY_FUNCTION__ )); | ||||||||||||
95 | } | ||||||||||||
96 | }; | ||||||||||||
97 | |||||||||||||
98 | using VPlanPtr = std::unique_ptr<VPlan>; | ||||||||||||
99 | |||||||||||||
100 | /// In what follows, the term "input IR" refers to code that is fed into the | ||||||||||||
101 | /// vectorizer whereas the term "output IR" refers to code that is generated by | ||||||||||||
102 | /// the vectorizer. | ||||||||||||
103 | |||||||||||||
104 | /// VPLane provides a way to access lanes in both fixed width and scalable | ||||||||||||
105 | /// vectors, where for the latter the lane index sometimes needs calculating | ||||||||||||
106 | /// as a runtime expression. | ||||||||||||
107 | class VPLane { | ||||||||||||
108 | public: | ||||||||||||
109 | /// Kind describes how to interpret Lane. | ||||||||||||
110 | enum class Kind : uint8_t { | ||||||||||||
111 | /// For First, Lane is the index into the first N elements of a | ||||||||||||
112 | /// fixed-vector <N x <ElTy>> or a scalable vector <vscale x N x <ElTy>>. | ||||||||||||
113 | First, | ||||||||||||
114 | /// For ScalableLast, Lane is the offset from the start of the last | ||||||||||||
115 | /// N-element subvector in a scalable vector <vscale x N x <ElTy>>. For | ||||||||||||
116 | /// example, a Lane of 0 corresponds to lane `(vscale - 1) * N`, a Lane of | ||||||||||||
117 | /// 1 corresponds to `((vscale - 1) * N) + 1`, etc. | ||||||||||||
118 | ScalableLast | ||||||||||||
119 | }; | ||||||||||||
120 | |||||||||||||
121 | private: | ||||||||||||
122 | /// in [0..VF) | ||||||||||||
123 | unsigned Lane; | ||||||||||||
124 | |||||||||||||
125 | /// Indicates how the Lane should be interpreted, as described above. | ||||||||||||
126 | Kind LaneKind; | ||||||||||||
127 | |||||||||||||
128 | public: | ||||||||||||
129 | VPLane(unsigned Lane, Kind LaneKind) : Lane(Lane), LaneKind(LaneKind) {} | ||||||||||||
130 | |||||||||||||
131 | static VPLane getFirstLane() { return VPLane(0, VPLane::Kind::First); } | ||||||||||||
132 | |||||||||||||
133 | static VPLane getLastLaneForVF(const ElementCount &VF) { | ||||||||||||
134 | unsigned LaneOffset = VF.getKnownMinValue() - 1; | ||||||||||||
135 | Kind LaneKind; | ||||||||||||
136 | if (VF.isScalable()) | ||||||||||||
137 | // In this case 'LaneOffset' refers to the offset from the start of the | ||||||||||||
138 | // last subvector with VF.getKnownMinValue() elements. | ||||||||||||
139 | LaneKind = VPLane::Kind::ScalableLast; | ||||||||||||
140 | else | ||||||||||||
141 | LaneKind = VPLane::Kind::First; | ||||||||||||
142 | return VPLane(LaneOffset, LaneKind); | ||||||||||||
143 | } | ||||||||||||
144 | |||||||||||||
145 | /// Returns a compile-time known value for the lane index and asserts if the | ||||||||||||
146 | /// lane can only be calculated at runtime. | ||||||||||||
147 | unsigned getKnownLane() const { | ||||||||||||
148 | assert(LaneKind == Kind::First)(static_cast <bool> (LaneKind == Kind::First) ? void (0 ) : __assert_fail ("LaneKind == Kind::First", "llvm/lib/Transforms/Vectorize/VPlan.h" , 148, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
149 | return Lane; | ||||||||||||
150 | } | ||||||||||||
151 | |||||||||||||
152 | /// Returns an expression describing the lane index that can be used at | ||||||||||||
153 | /// runtime. | ||||||||||||
154 | Value *getAsRuntimeExpr(IRBuilder<> &Builder, const ElementCount &VF) const; | ||||||||||||
155 | |||||||||||||
156 | /// Returns the Kind of lane offset. | ||||||||||||
157 | Kind getKind() const { return LaneKind; } | ||||||||||||
158 | |||||||||||||
159 | /// Returns true if this is the first lane of the whole vector. | ||||||||||||
160 | bool isFirstLane() const { return Lane == 0 && LaneKind == Kind::First; } | ||||||||||||
161 | |||||||||||||
162 | /// Maps the lane to a cache index based on \p VF. | ||||||||||||
163 | unsigned mapToCacheIndex(const ElementCount &VF) const { | ||||||||||||
164 | switch (LaneKind) { | ||||||||||||
165 | case VPLane::Kind::ScalableLast: | ||||||||||||
166 | 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", 166, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
167 | return VF.getKnownMinValue() + Lane; | ||||||||||||
168 | default: | ||||||||||||
169 | assert(Lane < VF.getKnownMinValue())(static_cast <bool> (Lane < VF.getKnownMinValue()) ? void (0) : __assert_fail ("Lane < VF.getKnownMinValue()", "llvm/lib/Transforms/Vectorize/VPlan.h", 169, __extension__ __PRETTY_FUNCTION__ )); | ||||||||||||
170 | return Lane; | ||||||||||||
171 | } | ||||||||||||
172 | } | ||||||||||||
173 | |||||||||||||
174 | /// Returns the maxmimum number of lanes that we are able to consider | ||||||||||||
175 | /// caching for \p VF. | ||||||||||||
176 | static unsigned getNumCachedLanes(const ElementCount &VF) { | ||||||||||||
177 | return VF.getKnownMinValue() * (VF.isScalable() ? 2 : 1); | ||||||||||||
178 | } | ||||||||||||
179 | }; | ||||||||||||
180 | |||||||||||||
181 | /// VPIteration represents a single point in the iteration space of the output | ||||||||||||
182 | /// (vectorized and/or unrolled) IR loop. | ||||||||||||
183 | struct VPIteration { | ||||||||||||
184 | /// in [0..UF) | ||||||||||||
185 | unsigned Part; | ||||||||||||
186 | |||||||||||||
187 | VPLane Lane; | ||||||||||||
188 | |||||||||||||
189 | VPIteration(unsigned Part, unsigned Lane, | ||||||||||||
190 | VPLane::Kind Kind = VPLane::Kind::First) | ||||||||||||
191 | : Part(Part), Lane(Lane, Kind) {} | ||||||||||||
192 | |||||||||||||
193 | VPIteration(unsigned Part, const VPLane &Lane) : Part(Part), Lane(Lane) {} | ||||||||||||
194 | |||||||||||||
195 | bool isFirstIteration() const { return Part == 0 && Lane.isFirstLane(); } | ||||||||||||
196 | }; | ||||||||||||
197 | |||||||||||||
198 | /// VPTransformState holds information passed down when "executing" a VPlan, | ||||||||||||
199 | /// needed for generating the output IR. | ||||||||||||
200 | struct VPTransformState { | ||||||||||||
201 | VPTransformState(ElementCount VF, unsigned UF, LoopInfo *LI, | ||||||||||||
202 | DominatorTree *DT, IRBuilder<> &Builder, | ||||||||||||
203 | InnerLoopVectorizer *ILV, VPlan *Plan) | ||||||||||||
204 | : VF(VF), UF(UF), LI(LI), DT(DT), Builder(Builder), ILV(ILV), Plan(Plan) { | ||||||||||||
205 | } | ||||||||||||
206 | |||||||||||||
207 | /// The chosen Vectorization and Unroll Factors of the loop being vectorized. | ||||||||||||
208 | ElementCount VF; | ||||||||||||
209 | unsigned UF; | ||||||||||||
210 | |||||||||||||
211 | /// Hold the indices to generate specific scalar instructions. Null indicates | ||||||||||||
212 | /// that all instances are to be generated, using either scalar or vector | ||||||||||||
213 | /// instructions. | ||||||||||||
214 | Optional<VPIteration> Instance; | ||||||||||||
215 | |||||||||||||
216 | struct DataState { | ||||||||||||
217 | /// A type for vectorized values in the new loop. Each value from the | ||||||||||||
218 | /// original loop, when vectorized, is represented by UF vector values in | ||||||||||||
219 | /// the new unrolled loop, where UF is the unroll factor. | ||||||||||||
220 | typedef SmallVector<Value *, 2> PerPartValuesTy; | ||||||||||||
221 | |||||||||||||
222 | DenseMap<VPValue *, PerPartValuesTy> PerPartOutput; | ||||||||||||
223 | |||||||||||||
224 | using ScalarsPerPartValuesTy = SmallVector<SmallVector<Value *, 4>, 2>; | ||||||||||||
225 | DenseMap<VPValue *, ScalarsPerPartValuesTy> PerPartScalars; | ||||||||||||
226 | } Data; | ||||||||||||
227 | |||||||||||||
228 | /// Get the generated Value for a given VPValue and a given Part. Note that | ||||||||||||
229 | /// as some Defs are still created by ILV and managed in its ValueMap, this | ||||||||||||
230 | /// method will delegate the call to ILV in such cases in order to provide | ||||||||||||
231 | /// callers a consistent API. | ||||||||||||
232 | /// \see set. | ||||||||||||
233 | Value *get(VPValue *Def, unsigned Part); | ||||||||||||
234 | |||||||||||||
235 | /// Get the generated Value for a given VPValue and given Part and Lane. | ||||||||||||
236 | Value *get(VPValue *Def, const VPIteration &Instance); | ||||||||||||
237 | |||||||||||||
238 | bool hasVectorValue(VPValue *Def, unsigned Part) { | ||||||||||||
239 | auto I = Data.PerPartOutput.find(Def); | ||||||||||||
240 | return I != Data.PerPartOutput.end() && Part < I->second.size() && | ||||||||||||
241 | I->second[Part]; | ||||||||||||
242 | } | ||||||||||||
243 | |||||||||||||
244 | bool hasAnyVectorValue(VPValue *Def) const { | ||||||||||||
245 | return Data.PerPartOutput.find(Def) != Data.PerPartOutput.end(); | ||||||||||||
246 | } | ||||||||||||
247 | |||||||||||||
248 | bool hasScalarValue(VPValue *Def, VPIteration Instance) { | ||||||||||||
249 | auto I = Data.PerPartScalars.find(Def); | ||||||||||||
250 | if (I == Data.PerPartScalars.end()) | ||||||||||||
251 | return false; | ||||||||||||
252 | unsigned CacheIdx = Instance.Lane.mapToCacheIndex(VF); | ||||||||||||
253 | return Instance.Part < I->second.size() && | ||||||||||||
254 | CacheIdx < I->second[Instance.Part].size() && | ||||||||||||
255 | I->second[Instance.Part][CacheIdx]; | ||||||||||||
256 | } | ||||||||||||
257 | |||||||||||||
258 | /// Set the generated Value for a given VPValue and a given Part. | ||||||||||||
259 | void set(VPValue *Def, Value *V, unsigned Part) { | ||||||||||||
260 | if (!Data.PerPartOutput.count(Def)) { | ||||||||||||
261 | DataState::PerPartValuesTy Entry(UF); | ||||||||||||
262 | Data.PerPartOutput[Def] = Entry; | ||||||||||||
263 | } | ||||||||||||
264 | Data.PerPartOutput[Def][Part] = V; | ||||||||||||
265 | } | ||||||||||||
266 | /// Reset an existing vector value for \p Def and a given \p Part. | ||||||||||||
267 | void reset(VPValue *Def, Value *V, unsigned Part) { | ||||||||||||
268 | auto Iter = Data.PerPartOutput.find(Def); | ||||||||||||
269 | 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", 270, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
270 | "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", 270, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
271 | Iter->second[Part] = V; | ||||||||||||
272 | } | ||||||||||||
273 | |||||||||||||
274 | /// Set the generated scalar \p V for \p Def and the given \p Instance. | ||||||||||||
275 | void set(VPValue *Def, Value *V, const VPIteration &Instance) { | ||||||||||||
276 | auto Iter = Data.PerPartScalars.insert({Def, {}}); | ||||||||||||
277 | auto &PerPartVec = Iter.first->second; | ||||||||||||
278 | while (PerPartVec.size() <= Instance.Part) | ||||||||||||
279 | PerPartVec.emplace_back(); | ||||||||||||
280 | auto &Scalars = PerPartVec[Instance.Part]; | ||||||||||||
281 | unsigned CacheIdx = Instance.Lane.mapToCacheIndex(VF); | ||||||||||||
282 | while (Scalars.size() <= CacheIdx) | ||||||||||||
283 | Scalars.push_back(nullptr); | ||||||||||||
284 | 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", 284, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
285 | Scalars[CacheIdx] = V; | ||||||||||||
286 | } | ||||||||||||
287 | |||||||||||||
288 | /// Reset an existing scalar value for \p Def and a given \p Instance. | ||||||||||||
289 | void reset(VPValue *Def, Value *V, const VPIteration &Instance) { | ||||||||||||
290 | auto Iter = Data.PerPartScalars.find(Def); | ||||||||||||
291 | 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", 292, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
292 | "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", 292, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
293 | 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", 294, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
294 | "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", 294, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
295 | unsigned CacheIdx = Instance.Lane.mapToCacheIndex(VF); | ||||||||||||
296 | 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", 297, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
297 | "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", 297, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
298 | Iter->second[Instance.Part][CacheIdx] = V; | ||||||||||||
299 | } | ||||||||||||
300 | |||||||||||||
301 | /// Hold state information used when constructing the CFG of the output IR, | ||||||||||||
302 | /// traversing the VPBasicBlocks and generating corresponding IR BasicBlocks. | ||||||||||||
303 | struct CFGState { | ||||||||||||
304 | /// The previous VPBasicBlock visited. Initially set to null. | ||||||||||||
305 | VPBasicBlock *PrevVPBB = nullptr; | ||||||||||||
306 | |||||||||||||
307 | /// The previous IR BasicBlock created or used. Initially set to the new | ||||||||||||
308 | /// header BasicBlock. | ||||||||||||
309 | BasicBlock *PrevBB = nullptr; | ||||||||||||
310 | |||||||||||||
311 | /// The last IR BasicBlock in the output IR. Set to the new latch | ||||||||||||
312 | /// BasicBlock, used for placing the newly created BasicBlocks. | ||||||||||||
313 | BasicBlock *LastBB = nullptr; | ||||||||||||
314 | |||||||||||||
315 | /// The IR BasicBlock that is the preheader of the vector loop in the output | ||||||||||||
316 | /// IR. | ||||||||||||
317 | /// FIXME: The vector preheader should also be modeled in VPlan, so any code | ||||||||||||
318 | /// that needs to be added to the preheader gets directly generated by | ||||||||||||
319 | /// VPlan. There should be no need to manage a pointer to the IR BasicBlock. | ||||||||||||
320 | BasicBlock *VectorPreHeader = nullptr; | ||||||||||||
321 | |||||||||||||
322 | /// A mapping of each VPBasicBlock to the corresponding BasicBlock. In case | ||||||||||||
323 | /// of replication, maps the BasicBlock of the last replica created. | ||||||||||||
324 | SmallDenseMap<VPBasicBlock *, BasicBlock *> VPBB2IRBB; | ||||||||||||
325 | |||||||||||||
326 | /// Vector of VPBasicBlocks whose terminator instruction needs to be fixed | ||||||||||||
327 | /// up at the end of vector code generation. | ||||||||||||
328 | SmallVector<VPBasicBlock *, 8> VPBBsToFix; | ||||||||||||
329 | |||||||||||||
330 | CFGState() = default; | ||||||||||||
331 | } CFG; | ||||||||||||
332 | |||||||||||||
333 | /// Hold a pointer to LoopInfo to register new basic blocks in the loop. | ||||||||||||
334 | LoopInfo *LI; | ||||||||||||
335 | |||||||||||||
336 | /// Hold a pointer to Dominator Tree to register new basic blocks in the loop. | ||||||||||||
337 | DominatorTree *DT; | ||||||||||||
338 | |||||||||||||
339 | /// Hold a reference to the IRBuilder used to generate output IR code. | ||||||||||||
340 | IRBuilder<> &Builder; | ||||||||||||
341 | |||||||||||||
342 | VPValue2ValueTy VPValue2Value; | ||||||||||||
343 | |||||||||||||
344 | /// Hold the canonical scalar IV of the vector loop (start=0, step=VF*UF). | ||||||||||||
345 | Value *CanonicalIV = nullptr; | ||||||||||||
346 | |||||||||||||
347 | /// Hold a pointer to InnerLoopVectorizer to reuse its IR generation methods. | ||||||||||||
348 | InnerLoopVectorizer *ILV; | ||||||||||||
349 | |||||||||||||
350 | /// Pointer to the VPlan code is generated for. | ||||||||||||
351 | VPlan *Plan; | ||||||||||||
352 | |||||||||||||
353 | /// Holds recipes that may generate a poison value that is used after | ||||||||||||
354 | /// vectorization, even when their operands are not poison. | ||||||||||||
355 | SmallPtrSet<VPRecipeBase *, 16> MayGeneratePoisonRecipes; | ||||||||||||
356 | }; | ||||||||||||
357 | |||||||||||||
358 | /// VPUsers instance used by VPBlockBase to manage CondBit and the block | ||||||||||||
359 | /// predicate. Currently VPBlockUsers are used in VPBlockBase for historical | ||||||||||||
360 | /// reasons, but in the future the only VPUsers should either be recipes or | ||||||||||||
361 | /// live-outs.VPBlockBase uses. | ||||||||||||
362 | struct VPBlockUser : public VPUser { | ||||||||||||
363 | VPBlockUser() : VPUser({}, VPUserID::Block) {} | ||||||||||||
364 | |||||||||||||
365 | VPValue *getSingleOperandOrNull() { | ||||||||||||
366 | if (getNumOperands() == 1) | ||||||||||||
367 | return getOperand(0); | ||||||||||||
368 | |||||||||||||
369 | return nullptr; | ||||||||||||
370 | } | ||||||||||||
371 | const VPValue *getSingleOperandOrNull() const { | ||||||||||||
372 | if (getNumOperands() == 1) | ||||||||||||
373 | return getOperand(0); | ||||||||||||
374 | |||||||||||||
375 | return nullptr; | ||||||||||||
376 | } | ||||||||||||
377 | |||||||||||||
378 | void resetSingleOpUser(VPValue *NewVal) { | ||||||||||||
379 | assert(getNumOperands() <= 1 && "Didn't expect more than one operand!")(static_cast <bool> (getNumOperands() <= 1 && "Didn't expect more than one operand!") ? void (0) : __assert_fail ("getNumOperands() <= 1 && \"Didn't expect more than one operand!\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 379, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
380 | if (!NewVal) { | ||||||||||||
381 | if (getNumOperands() == 1) | ||||||||||||
382 | removeLastOperand(); | ||||||||||||
383 | return; | ||||||||||||
384 | } | ||||||||||||
385 | |||||||||||||
386 | if (getNumOperands() == 1) | ||||||||||||
387 | setOperand(0, NewVal); | ||||||||||||
388 | else | ||||||||||||
389 | addOperand(NewVal); | ||||||||||||
390 | } | ||||||||||||
391 | }; | ||||||||||||
392 | |||||||||||||
393 | /// VPBlockBase is the building block of the Hierarchical Control-Flow Graph. | ||||||||||||
394 | /// A VPBlockBase can be either a VPBasicBlock or a VPRegionBlock. | ||||||||||||
395 | class VPBlockBase { | ||||||||||||
396 | friend class VPBlockUtils; | ||||||||||||
397 | |||||||||||||
398 | const unsigned char SubclassID; ///< Subclass identifier (for isa/dyn_cast). | ||||||||||||
399 | |||||||||||||
400 | /// An optional name for the block. | ||||||||||||
401 | std::string Name; | ||||||||||||
402 | |||||||||||||
403 | /// The immediate VPRegionBlock which this VPBlockBase belongs to, or null if | ||||||||||||
404 | /// it is a topmost VPBlockBase. | ||||||||||||
405 | VPRegionBlock *Parent = nullptr; | ||||||||||||
406 | |||||||||||||
407 | /// List of predecessor blocks. | ||||||||||||
408 | SmallVector<VPBlockBase *, 1> Predecessors; | ||||||||||||
409 | |||||||||||||
410 | /// List of successor blocks. | ||||||||||||
411 | SmallVector<VPBlockBase *, 1> Successors; | ||||||||||||
412 | |||||||||||||
413 | /// Successor selector managed by a VPUser. For blocks with zero or one | ||||||||||||
414 | /// successors, there is no operand. Otherwise there is exactly one operand | ||||||||||||
415 | /// which is the branch condition. | ||||||||||||
416 | VPBlockUser CondBitUser; | ||||||||||||
417 | |||||||||||||
418 | /// If the block is predicated, its predicate is stored as an operand of this | ||||||||||||
419 | /// VPUser to maintain the def-use relations. Otherwise there is no operand | ||||||||||||
420 | /// here. | ||||||||||||
421 | VPBlockUser PredicateUser; | ||||||||||||
422 | |||||||||||||
423 | /// VPlan containing the block. Can only be set on the entry block of the | ||||||||||||
424 | /// plan. | ||||||||||||
425 | VPlan *Plan = nullptr; | ||||||||||||
426 | |||||||||||||
427 | /// Add \p Successor as the last successor to this block. | ||||||||||||
428 | void appendSuccessor(VPBlockBase *Successor) { | ||||||||||||
429 | 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", 429, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
430 | Successors.push_back(Successor); | ||||||||||||
431 | } | ||||||||||||
432 | |||||||||||||
433 | /// Add \p Predecessor as the last predecessor to this block. | ||||||||||||
434 | void appendPredecessor(VPBlockBase *Predecessor) { | ||||||||||||
435 | 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", 435, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
436 | Predecessors.push_back(Predecessor); | ||||||||||||
437 | } | ||||||||||||
438 | |||||||||||||
439 | /// Remove \p Predecessor from the predecessors of this block. | ||||||||||||
440 | void removePredecessor(VPBlockBase *Predecessor) { | ||||||||||||
441 | auto Pos = find(Predecessors, Predecessor); | ||||||||||||
442 | 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", 442, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
443 | Predecessors.erase(Pos); | ||||||||||||
444 | } | ||||||||||||
445 | |||||||||||||
446 | /// Remove \p Successor from the successors of this block. | ||||||||||||
447 | void removeSuccessor(VPBlockBase *Successor) { | ||||||||||||
448 | auto Pos = find(Successors, Successor); | ||||||||||||
449 | 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", 449, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
450 | Successors.erase(Pos); | ||||||||||||
451 | } | ||||||||||||
452 | |||||||||||||
453 | protected: | ||||||||||||
454 | VPBlockBase(const unsigned char SC, const std::string &N) | ||||||||||||
455 | : SubclassID(SC), Name(N) {} | ||||||||||||
456 | |||||||||||||
457 | public: | ||||||||||||
458 | /// An enumeration for keeping track of the concrete subclass of VPBlockBase | ||||||||||||
459 | /// that are actually instantiated. Values of this enumeration are kept in the | ||||||||||||
460 | /// SubclassID field of the VPBlockBase objects. They are used for concrete | ||||||||||||
461 | /// type identification. | ||||||||||||
462 | using VPBlockTy = enum { VPBasicBlockSC, VPRegionBlockSC }; | ||||||||||||
463 | |||||||||||||
464 | using VPBlocksTy = SmallVectorImpl<VPBlockBase *>; | ||||||||||||
465 | |||||||||||||
466 | virtual ~VPBlockBase() = default; | ||||||||||||
467 | |||||||||||||
468 | const std::string &getName() const { return Name; } | ||||||||||||
469 | |||||||||||||
470 | void setName(const Twine &newName) { Name = newName.str(); } | ||||||||||||
471 | |||||||||||||
472 | /// \return an ID for the concrete type of this object. | ||||||||||||
473 | /// This is used to implement the classof checks. This should not be used | ||||||||||||
474 | /// for any other purpose, as the values may change as LLVM evolves. | ||||||||||||
475 | unsigned getVPBlockID() const { return SubclassID; } | ||||||||||||
476 | |||||||||||||
477 | VPRegionBlock *getParent() { return Parent; } | ||||||||||||
478 | const VPRegionBlock *getParent() const { return Parent; } | ||||||||||||
479 | |||||||||||||
480 | /// \return A pointer to the plan containing the current block. | ||||||||||||
481 | VPlan *getPlan(); | ||||||||||||
482 | const VPlan *getPlan() const; | ||||||||||||
483 | |||||||||||||
484 | /// Sets the pointer of the plan containing the block. The block must be the | ||||||||||||
485 | /// entry block into the VPlan. | ||||||||||||
486 | void setPlan(VPlan *ParentPlan); | ||||||||||||
487 | |||||||||||||
488 | void setParent(VPRegionBlock *P) { Parent = P; } | ||||||||||||
489 | |||||||||||||
490 | /// \return the VPBasicBlock that is the entry of this VPBlockBase, | ||||||||||||
491 | /// recursively, if the latter is a VPRegionBlock. Otherwise, if this | ||||||||||||
492 | /// VPBlockBase is a VPBasicBlock, it is returned. | ||||||||||||
493 | const VPBasicBlock *getEntryBasicBlock() const; | ||||||||||||
494 | VPBasicBlock *getEntryBasicBlock(); | ||||||||||||
495 | |||||||||||||
496 | /// \return the VPBasicBlock that is the exit of this VPBlockBase, | ||||||||||||
497 | /// recursively, if the latter is a VPRegionBlock. Otherwise, if this | ||||||||||||
498 | /// VPBlockBase is a VPBasicBlock, it is returned. | ||||||||||||
499 | const VPBasicBlock *getExitBasicBlock() const; | ||||||||||||
500 | VPBasicBlock *getExitBasicBlock(); | ||||||||||||
501 | |||||||||||||
502 | const VPBlocksTy &getSuccessors() const { return Successors; } | ||||||||||||
503 | VPBlocksTy &getSuccessors() { return Successors; } | ||||||||||||
504 | |||||||||||||
505 | iterator_range<VPBlockBase **> successors() { return Successors; } | ||||||||||||
506 | |||||||||||||
507 | const VPBlocksTy &getPredecessors() const { return Predecessors; } | ||||||||||||
508 | VPBlocksTy &getPredecessors() { return Predecessors; } | ||||||||||||
509 | |||||||||||||
510 | /// \return the successor of this VPBlockBase if it has a single successor. | ||||||||||||
511 | /// Otherwise return a null pointer. | ||||||||||||
512 | VPBlockBase *getSingleSuccessor() const { | ||||||||||||
513 | return (Successors.size() == 1 ? *Successors.begin() : nullptr); | ||||||||||||
514 | } | ||||||||||||
515 | |||||||||||||
516 | /// \return the predecessor of this VPBlockBase if it has a single | ||||||||||||
517 | /// predecessor. Otherwise return a null pointer. | ||||||||||||
518 | VPBlockBase *getSinglePredecessor() const { | ||||||||||||
519 | return (Predecessors.size() == 1 ? *Predecessors.begin() : nullptr); | ||||||||||||
520 | } | ||||||||||||
521 | |||||||||||||
522 | size_t getNumSuccessors() const { return Successors.size(); } | ||||||||||||
523 | size_t getNumPredecessors() const { return Predecessors.size(); } | ||||||||||||
524 | |||||||||||||
525 | /// An Enclosing Block of a block B is any block containing B, including B | ||||||||||||
526 | /// itself. \return the closest enclosing block starting from "this", which | ||||||||||||
527 | /// has successors. \return the root enclosing block if all enclosing blocks | ||||||||||||
528 | /// have no successors. | ||||||||||||
529 | VPBlockBase *getEnclosingBlockWithSuccessors(); | ||||||||||||
530 | |||||||||||||
531 | /// \return the closest enclosing block starting from "this", which has | ||||||||||||
532 | /// predecessors. \return the root enclosing block if all enclosing blocks | ||||||||||||
533 | /// have no predecessors. | ||||||||||||
534 | VPBlockBase *getEnclosingBlockWithPredecessors(); | ||||||||||||
535 | |||||||||||||
536 | /// \return the successors either attached directly to this VPBlockBase or, if | ||||||||||||
537 | /// this VPBlockBase is the exit block of a VPRegionBlock and has no | ||||||||||||
538 | /// successors of its own, search recursively for the first enclosing | ||||||||||||
539 | /// VPRegionBlock that has successors and return them. If no such | ||||||||||||
540 | /// VPRegionBlock exists, return the (empty) successors of the topmost | ||||||||||||
541 | /// VPBlockBase reached. | ||||||||||||
542 | const VPBlocksTy &getHierarchicalSuccessors() { | ||||||||||||
543 | return getEnclosingBlockWithSuccessors()->getSuccessors(); | ||||||||||||
544 | } | ||||||||||||
545 | |||||||||||||
546 | /// \return the hierarchical successor of this VPBlockBase if it has a single | ||||||||||||
547 | /// hierarchical successor. Otherwise return a null pointer. | ||||||||||||
548 | VPBlockBase *getSingleHierarchicalSuccessor() { | ||||||||||||
549 | return getEnclosingBlockWithSuccessors()->getSingleSuccessor(); | ||||||||||||
550 | } | ||||||||||||
551 | |||||||||||||
552 | /// \return the predecessors either attached directly to this VPBlockBase or, | ||||||||||||
553 | /// if this VPBlockBase is the entry block of a VPRegionBlock and has no | ||||||||||||
554 | /// predecessors of its own, search recursively for the first enclosing | ||||||||||||
555 | /// VPRegionBlock that has predecessors and return them. If no such | ||||||||||||
556 | /// VPRegionBlock exists, return the (empty) predecessors of the topmost | ||||||||||||
557 | /// VPBlockBase reached. | ||||||||||||
558 | const VPBlocksTy &getHierarchicalPredecessors() { | ||||||||||||
559 | return getEnclosingBlockWithPredecessors()->getPredecessors(); | ||||||||||||
560 | } | ||||||||||||
561 | |||||||||||||
562 | /// \return the hierarchical predecessor of this VPBlockBase if it has a | ||||||||||||
563 | /// single hierarchical predecessor. Otherwise return a null pointer. | ||||||||||||
564 | VPBlockBase *getSingleHierarchicalPredecessor() { | ||||||||||||
565 | return getEnclosingBlockWithPredecessors()->getSinglePredecessor(); | ||||||||||||
566 | } | ||||||||||||
567 | |||||||||||||
568 | /// \return the condition bit selecting the successor. | ||||||||||||
569 | VPValue *getCondBit(); | ||||||||||||
570 | /// \return the condition bit selecting the successor. | ||||||||||||
571 | const VPValue *getCondBit() const; | ||||||||||||
572 | /// Set the condition bit selecting the successor. | ||||||||||||
573 | void setCondBit(VPValue *CV); | ||||||||||||
574 | |||||||||||||
575 | /// \return the block's predicate. | ||||||||||||
576 | VPValue *getPredicate(); | ||||||||||||
577 | /// \return the block's predicate. | ||||||||||||
578 | const VPValue *getPredicate() const; | ||||||||||||
579 | /// Set the block's predicate. | ||||||||||||
580 | void setPredicate(VPValue *Pred); | ||||||||||||
581 | |||||||||||||
582 | /// Set a given VPBlockBase \p Successor as the single successor of this | ||||||||||||
583 | /// VPBlockBase. This VPBlockBase is not added as predecessor of \p Successor. | ||||||||||||
584 | /// This VPBlockBase must have no successors. | ||||||||||||
585 | void setOneSuccessor(VPBlockBase *Successor) { | ||||||||||||
586 | 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", 586, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
587 | appendSuccessor(Successor); | ||||||||||||
588 | } | ||||||||||||
589 | |||||||||||||
590 | /// Set two given VPBlockBases \p IfTrue and \p IfFalse to be the two | ||||||||||||
591 | /// successors of this VPBlockBase. \p Condition is set as the successor | ||||||||||||
592 | /// selector. This VPBlockBase is not added as predecessor of \p IfTrue or \p | ||||||||||||
593 | /// IfFalse. This VPBlockBase must have no successors. | ||||||||||||
594 | void setTwoSuccessors(VPBlockBase *IfTrue, VPBlockBase *IfFalse, | ||||||||||||
595 | VPValue *Condition) { | ||||||||||||
596 | 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", 596, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
597 | assert(Condition && "Setting two successors without condition!")(static_cast <bool> (Condition && "Setting two successors without condition!" ) ? void (0) : __assert_fail ("Condition && \"Setting two successors without condition!\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 597, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
598 | setCondBit(Condition); | ||||||||||||
599 | appendSuccessor(IfTrue); | ||||||||||||
600 | appendSuccessor(IfFalse); | ||||||||||||
601 | } | ||||||||||||
602 | |||||||||||||
603 | /// Set each VPBasicBlock in \p NewPreds as predecessor of this VPBlockBase. | ||||||||||||
604 | /// This VPBlockBase must have no predecessors. This VPBlockBase is not added | ||||||||||||
605 | /// as successor of any VPBasicBlock in \p NewPreds. | ||||||||||||
606 | void setPredecessors(ArrayRef<VPBlockBase *> NewPreds) { | ||||||||||||
607 | 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", 607, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
608 | for (auto *Pred : NewPreds) | ||||||||||||
609 | appendPredecessor(Pred); | ||||||||||||
610 | } | ||||||||||||
611 | |||||||||||||
612 | /// Remove all the predecessor of this block. | ||||||||||||
613 | void clearPredecessors() { Predecessors.clear(); } | ||||||||||||
614 | |||||||||||||
615 | /// Remove all the successors of this block and set to null its condition bit | ||||||||||||
616 | void clearSuccessors() { | ||||||||||||
617 | Successors.clear(); | ||||||||||||
618 | setCondBit(nullptr); | ||||||||||||
619 | } | ||||||||||||
620 | |||||||||||||
621 | /// The method which generates the output IR that correspond to this | ||||||||||||
622 | /// VPBlockBase, thereby "executing" the VPlan. | ||||||||||||
623 | virtual void execute(struct VPTransformState *State) = 0; | ||||||||||||
624 | |||||||||||||
625 | /// Delete all blocks reachable from a given VPBlockBase, inclusive. | ||||||||||||
626 | static void deleteCFG(VPBlockBase *Entry); | ||||||||||||
627 | |||||||||||||
628 | /// Return true if it is legal to hoist instructions into this block. | ||||||||||||
629 | bool isLegalToHoistInto() { | ||||||||||||
630 | // There are currently no constraints that prevent an instruction to be | ||||||||||||
631 | // hoisted into a VPBlockBase. | ||||||||||||
632 | return true; | ||||||||||||
633 | } | ||||||||||||
634 | |||||||||||||
635 | /// Replace all operands of VPUsers in the block with \p NewValue and also | ||||||||||||
636 | /// replaces all uses of VPValues defined in the block with NewValue. | ||||||||||||
637 | virtual void dropAllReferences(VPValue *NewValue) = 0; | ||||||||||||
638 | |||||||||||||
639 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
640 | void printAsOperand(raw_ostream &OS, bool PrintType) const { | ||||||||||||
641 | OS << getName(); | ||||||||||||
642 | } | ||||||||||||
643 | |||||||||||||
644 | /// Print plain-text dump of this VPBlockBase to \p O, prefixing all lines | ||||||||||||
645 | /// with \p Indent. \p SlotTracker is used to print unnamed VPValue's using | ||||||||||||
646 | /// consequtive numbers. | ||||||||||||
647 | /// | ||||||||||||
648 | /// Note that the numbering is applied to the whole VPlan, so printing | ||||||||||||
649 | /// individual blocks is consistent with the whole VPlan printing. | ||||||||||||
650 | virtual void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
651 | VPSlotTracker &SlotTracker) const = 0; | ||||||||||||
652 | |||||||||||||
653 | /// Print plain-text dump of this VPlan to \p O. | ||||||||||||
654 | void print(raw_ostream &O) const { | ||||||||||||
655 | VPSlotTracker SlotTracker(getPlan()); | ||||||||||||
656 | print(O, "", SlotTracker); | ||||||||||||
657 | } | ||||||||||||
658 | |||||||||||||
659 | /// Print the successors of this block to \p O, prefixing all lines with \p | ||||||||||||
660 | /// Indent. | ||||||||||||
661 | void printSuccessors(raw_ostream &O, const Twine &Indent) const; | ||||||||||||
662 | |||||||||||||
663 | /// Dump this VPBlockBase to dbgs(). | ||||||||||||
664 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { print(dbgs()); } | ||||||||||||
665 | #endif | ||||||||||||
666 | }; | ||||||||||||
667 | |||||||||||||
668 | /// VPRecipeBase is a base class modeling a sequence of one or more output IR | ||||||||||||
669 | /// instructions. VPRecipeBase owns the the VPValues it defines through VPDef | ||||||||||||
670 | /// and is responsible for deleting its defined values. Single-value | ||||||||||||
671 | /// VPRecipeBases that also inherit from VPValue must make sure to inherit from | ||||||||||||
672 | /// VPRecipeBase before VPValue. | ||||||||||||
673 | class VPRecipeBase : public ilist_node_with_parent<VPRecipeBase, VPBasicBlock>, | ||||||||||||
674 | public VPDef, | ||||||||||||
675 | public VPUser { | ||||||||||||
676 | friend VPBasicBlock; | ||||||||||||
677 | friend class VPBlockUtils; | ||||||||||||
678 | |||||||||||||
679 | /// Each VPRecipe belongs to a single VPBasicBlock. | ||||||||||||
680 | VPBasicBlock *Parent = nullptr; | ||||||||||||
681 | |||||||||||||
682 | public: | ||||||||||||
683 | VPRecipeBase(const unsigned char SC, ArrayRef<VPValue *> Operands) | ||||||||||||
684 | : VPDef(SC), VPUser(Operands, VPUser::VPUserID::Recipe) {} | ||||||||||||
685 | |||||||||||||
686 | template <typename IterT> | ||||||||||||
687 | VPRecipeBase(const unsigned char SC, iterator_range<IterT> Operands) | ||||||||||||
688 | : VPDef(SC), VPUser(Operands, VPUser::VPUserID::Recipe) {} | ||||||||||||
689 | virtual ~VPRecipeBase() = default; | ||||||||||||
690 | |||||||||||||
691 | /// \return the VPBasicBlock which this VPRecipe belongs to. | ||||||||||||
692 | VPBasicBlock *getParent() { return Parent; } | ||||||||||||
693 | const VPBasicBlock *getParent() const { return Parent; } | ||||||||||||
694 | |||||||||||||
695 | /// The method which generates the output IR instructions that correspond to | ||||||||||||
696 | /// this VPRecipe, thereby "executing" the VPlan. | ||||||||||||
697 | virtual void execute(struct VPTransformState &State) = 0; | ||||||||||||
698 | |||||||||||||
699 | /// Insert an unlinked recipe into a basic block immediately before | ||||||||||||
700 | /// the specified recipe. | ||||||||||||
701 | void insertBefore(VPRecipeBase *InsertPos); | ||||||||||||
702 | |||||||||||||
703 | /// Insert an unlinked Recipe into a basic block immediately after | ||||||||||||
704 | /// the specified Recipe. | ||||||||||||
705 | void insertAfter(VPRecipeBase *InsertPos); | ||||||||||||
706 | |||||||||||||
707 | /// Unlink this recipe from its current VPBasicBlock and insert it into | ||||||||||||
708 | /// the VPBasicBlock that MovePos lives in, right after MovePos. | ||||||||||||
709 | void moveAfter(VPRecipeBase *MovePos); | ||||||||||||
710 | |||||||||||||
711 | /// Unlink this recipe and insert into BB before I. | ||||||||||||
712 | /// | ||||||||||||
713 | /// \pre I is a valid iterator into BB. | ||||||||||||
714 | void moveBefore(VPBasicBlock &BB, iplist<VPRecipeBase>::iterator I); | ||||||||||||
715 | |||||||||||||
716 | /// This method unlinks 'this' from the containing basic block, but does not | ||||||||||||
717 | /// delete it. | ||||||||||||
718 | void removeFromParent(); | ||||||||||||
719 | |||||||||||||
720 | /// This method unlinks 'this' from the containing basic block and deletes it. | ||||||||||||
721 | /// | ||||||||||||
722 | /// \returns an iterator pointing to the element after the erased one | ||||||||||||
723 | iplist<VPRecipeBase>::iterator eraseFromParent(); | ||||||||||||
724 | |||||||||||||
725 | /// Returns the underlying instruction, if the recipe is a VPValue or nullptr | ||||||||||||
726 | /// otherwise. | ||||||||||||
727 | Instruction *getUnderlyingInstr() { | ||||||||||||
728 | return cast<Instruction>(getVPSingleValue()->getUnderlyingValue()); | ||||||||||||
729 | } | ||||||||||||
730 | const Instruction *getUnderlyingInstr() const { | ||||||||||||
731 | return cast<Instruction>(getVPSingleValue()->getUnderlyingValue()); | ||||||||||||
732 | } | ||||||||||||
733 | |||||||||||||
734 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
735 | static inline bool classof(const VPDef *D) { | ||||||||||||
736 | // All VPDefs are also VPRecipeBases. | ||||||||||||
737 | return true; | ||||||||||||
738 | } | ||||||||||||
739 | |||||||||||||
740 | static inline bool classof(const VPUser *U) { | ||||||||||||
741 | return U->getVPUserID() == VPUser::VPUserID::Recipe; | ||||||||||||
742 | } | ||||||||||||
743 | |||||||||||||
744 | /// Returns true if the recipe may have side-effects. | ||||||||||||
745 | bool mayHaveSideEffects() const; | ||||||||||||
746 | |||||||||||||
747 | /// Returns true for PHI-like recipes. | ||||||||||||
748 | bool isPhi() const { | ||||||||||||
749 | return getVPDefID() >= VPFirstPHISC && getVPDefID() <= VPLastPHISC; | ||||||||||||
750 | } | ||||||||||||
751 | |||||||||||||
752 | /// Returns true if the recipe may read from memory. | ||||||||||||
753 | bool mayReadFromMemory() const; | ||||||||||||
754 | |||||||||||||
755 | /// Returns true if the recipe may write to memory. | ||||||||||||
756 | bool mayWriteToMemory() const; | ||||||||||||
757 | |||||||||||||
758 | /// Returns true if the recipe may read from or write to memory. | ||||||||||||
759 | bool mayReadOrWriteMemory() const { | ||||||||||||
760 | return mayReadFromMemory() || mayWriteToMemory(); | ||||||||||||
761 | } | ||||||||||||
762 | }; | ||||||||||||
763 | |||||||||||||
764 | inline bool VPUser::classof(const VPDef *Def) { | ||||||||||||
765 | return Def->getVPDefID() == VPRecipeBase::VPInstructionSC || | ||||||||||||
766 | Def->getVPDefID() == VPRecipeBase::VPWidenSC || | ||||||||||||
767 | Def->getVPDefID() == VPRecipeBase::VPWidenCallSC || | ||||||||||||
768 | Def->getVPDefID() == VPRecipeBase::VPWidenSelectSC || | ||||||||||||
769 | Def->getVPDefID() == VPRecipeBase::VPWidenGEPSC || | ||||||||||||
770 | Def->getVPDefID() == VPRecipeBase::VPBlendSC || | ||||||||||||
771 | Def->getVPDefID() == VPRecipeBase::VPInterleaveSC || | ||||||||||||
772 | Def->getVPDefID() == VPRecipeBase::VPReplicateSC || | ||||||||||||
773 | Def->getVPDefID() == VPRecipeBase::VPReductionSC || | ||||||||||||
774 | Def->getVPDefID() == VPRecipeBase::VPBranchOnMaskSC || | ||||||||||||
775 | Def->getVPDefID() == VPRecipeBase::VPWidenMemoryInstructionSC; | ||||||||||||
776 | } | ||||||||||||
777 | |||||||||||||
778 | /// This is a concrete Recipe that models a single VPlan-level instruction. | ||||||||||||
779 | /// While as any Recipe it may generate a sequence of IR instructions when | ||||||||||||
780 | /// executed, these instructions would always form a single-def expression as | ||||||||||||
781 | /// the VPInstruction is also a single def-use vertex. | ||||||||||||
782 | class VPInstruction : public VPRecipeBase, public VPValue { | ||||||||||||
783 | friend class VPlanSlp; | ||||||||||||
784 | |||||||||||||
785 | public: | ||||||||||||
786 | /// VPlan opcodes, extending LLVM IR with idiomatics instructions. | ||||||||||||
787 | enum { | ||||||||||||
788 | FirstOrderRecurrenceSplice = | ||||||||||||
789 | Instruction::OtherOpsEnd + 1, // Combines the incoming and previous | ||||||||||||
790 | // values of a first-order recurrence. | ||||||||||||
791 | Not, | ||||||||||||
792 | ICmpULE, | ||||||||||||
793 | SLPLoad, | ||||||||||||
794 | SLPStore, | ||||||||||||
795 | ActiveLaneMask, | ||||||||||||
796 | CanonicalIVIncrement, | ||||||||||||
797 | CanonicalIVIncrementNUW, | ||||||||||||
798 | BranchOnCount, | ||||||||||||
799 | }; | ||||||||||||
800 | |||||||||||||
801 | private: | ||||||||||||
802 | typedef unsigned char OpcodeTy; | ||||||||||||
803 | OpcodeTy Opcode; | ||||||||||||
804 | FastMathFlags FMF; | ||||||||||||
805 | DebugLoc DL; | ||||||||||||
806 | |||||||||||||
807 | /// Utility method serving execute(): generates a single instance of the | ||||||||||||
808 | /// modeled instruction. | ||||||||||||
809 | void generateInstruction(VPTransformState &State, unsigned Part); | ||||||||||||
810 | |||||||||||||
811 | protected: | ||||||||||||
812 | void setUnderlyingInstr(Instruction *I) { setUnderlyingValue(I); } | ||||||||||||
813 | |||||||||||||
814 | public: | ||||||||||||
815 | VPInstruction(unsigned Opcode, ArrayRef<VPValue *> Operands, DebugLoc DL) | ||||||||||||
816 | : VPRecipeBase(VPRecipeBase::VPInstructionSC, Operands), | ||||||||||||
817 | VPValue(VPValue::VPVInstructionSC, nullptr, this), Opcode(Opcode), | ||||||||||||
818 | DL(DL) {} | ||||||||||||
819 | |||||||||||||
820 | VPInstruction(unsigned Opcode, std::initializer_list<VPValue *> Operands, | ||||||||||||
821 | DebugLoc DL = {}) | ||||||||||||
822 | : VPInstruction(Opcode, ArrayRef<VPValue *>(Operands), DL) {} | ||||||||||||
823 | |||||||||||||
824 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
825 | static inline bool classof(const VPValue *V) { | ||||||||||||
826 | return V->getVPValueID() == VPValue::VPVInstructionSC; | ||||||||||||
827 | } | ||||||||||||
828 | |||||||||||||
829 | VPInstruction *clone() const { | ||||||||||||
830 | SmallVector<VPValue *, 2> Operands(operands()); | ||||||||||||
831 | return new VPInstruction(Opcode, Operands, DL); | ||||||||||||
832 | } | ||||||||||||
833 | |||||||||||||
834 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
835 | static inline bool classof(const VPDef *R) { | ||||||||||||
836 | return R->getVPDefID() == VPRecipeBase::VPInstructionSC; | ||||||||||||
837 | } | ||||||||||||
838 | |||||||||||||
839 | /// Extra classof implementations to allow directly casting from VPUser -> | ||||||||||||
840 | /// VPInstruction. | ||||||||||||
841 | static inline bool classof(const VPUser *U) { | ||||||||||||
842 | auto *R = dyn_cast<VPRecipeBase>(U); | ||||||||||||
843 | return R && R->getVPDefID() == VPRecipeBase::VPInstructionSC; | ||||||||||||
844 | } | ||||||||||||
845 | static inline bool classof(const VPRecipeBase *R) { | ||||||||||||
846 | return R->getVPDefID() == VPRecipeBase::VPInstructionSC; | ||||||||||||
847 | } | ||||||||||||
848 | |||||||||||||
849 | unsigned getOpcode() const { return Opcode; } | ||||||||||||
850 | |||||||||||||
851 | /// Generate the instruction. | ||||||||||||
852 | /// TODO: We currently execute only per-part unless a specific instance is | ||||||||||||
853 | /// provided. | ||||||||||||
854 | void execute(VPTransformState &State) override; | ||||||||||||
855 | |||||||||||||
856 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
857 | /// Print the VPInstruction to \p O. | ||||||||||||
858 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
859 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
860 | |||||||||||||
861 | /// Print the VPInstruction to dbgs() (for debugging). | ||||||||||||
862 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const; | ||||||||||||
863 | #endif | ||||||||||||
864 | |||||||||||||
865 | /// Return true if this instruction may modify memory. | ||||||||||||
866 | bool mayWriteToMemory() const { | ||||||||||||
867 | // TODO: we can use attributes of the called function to rule out memory | ||||||||||||
868 | // modifications. | ||||||||||||
869 | return Opcode == Instruction::Store || Opcode == Instruction::Call || | ||||||||||||
870 | Opcode == Instruction::Invoke || Opcode == SLPStore; | ||||||||||||
871 | } | ||||||||||||
872 | |||||||||||||
873 | bool hasResult() const { | ||||||||||||
874 | // CallInst may or may not have a result, depending on the called function. | ||||||||||||
875 | // Conservatively return calls have results for now. | ||||||||||||
876 | switch (getOpcode()) { | ||||||||||||
877 | case Instruction::Ret: | ||||||||||||
878 | case Instruction::Br: | ||||||||||||
879 | case Instruction::Store: | ||||||||||||
880 | case Instruction::Switch: | ||||||||||||
881 | case Instruction::IndirectBr: | ||||||||||||
882 | case Instruction::Resume: | ||||||||||||
883 | case Instruction::CatchRet: | ||||||||||||
884 | case Instruction::Unreachable: | ||||||||||||
885 | case Instruction::Fence: | ||||||||||||
886 | case Instruction::AtomicRMW: | ||||||||||||
887 | case VPInstruction::BranchOnCount: | ||||||||||||
888 | return false; | ||||||||||||
889 | default: | ||||||||||||
890 | return true; | ||||||||||||
891 | } | ||||||||||||
892 | } | ||||||||||||
893 | |||||||||||||
894 | /// Set the fast-math flags. | ||||||||||||
895 | void setFastMathFlags(FastMathFlags FMFNew); | ||||||||||||
896 | }; | ||||||||||||
897 | |||||||||||||
898 | /// VPWidenRecipe is a recipe for producing a copy of vector type its | ||||||||||||
899 | /// ingredient. This recipe covers most of the traditional vectorization cases | ||||||||||||
900 | /// where each ingredient transforms into a vectorized version of itself. | ||||||||||||
901 | class VPWidenRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
902 | public: | ||||||||||||
903 | template <typename IterT> | ||||||||||||
904 | VPWidenRecipe(Instruction &I, iterator_range<IterT> Operands) | ||||||||||||
905 | : VPRecipeBase(VPRecipeBase::VPWidenSC, Operands), | ||||||||||||
906 | VPValue(VPValue::VPVWidenSC, &I, this) {} | ||||||||||||
907 | |||||||||||||
908 | ~VPWidenRecipe() override = default; | ||||||||||||
909 | |||||||||||||
910 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
911 | static inline bool classof(const VPDef *D) { | ||||||||||||
912 | return D->getVPDefID() == VPRecipeBase::VPWidenSC; | ||||||||||||
913 | } | ||||||||||||
914 | static inline bool classof(const VPValue *V) { | ||||||||||||
915 | return V->getVPValueID() == VPValue::VPVWidenSC; | ||||||||||||
916 | } | ||||||||||||
917 | |||||||||||||
918 | /// Produce widened copies of all Ingredients. | ||||||||||||
919 | void execute(VPTransformState &State) override; | ||||||||||||
920 | |||||||||||||
921 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
922 | /// Print the recipe. | ||||||||||||
923 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
924 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
925 | #endif | ||||||||||||
926 | }; | ||||||||||||
927 | |||||||||||||
928 | /// A recipe for widening Call instructions. | ||||||||||||
929 | class VPWidenCallRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
930 | |||||||||||||
931 | public: | ||||||||||||
932 | template <typename IterT> | ||||||||||||
933 | VPWidenCallRecipe(CallInst &I, iterator_range<IterT> CallArguments) | ||||||||||||
934 | : VPRecipeBase(VPRecipeBase::VPWidenCallSC, CallArguments), | ||||||||||||
935 | VPValue(VPValue::VPVWidenCallSC, &I, this) {} | ||||||||||||
936 | |||||||||||||
937 | ~VPWidenCallRecipe() override = default; | ||||||||||||
938 | |||||||||||||
939 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
940 | static inline bool classof(const VPDef *D) { | ||||||||||||
941 | return D->getVPDefID() == VPRecipeBase::VPWidenCallSC; | ||||||||||||
942 | } | ||||||||||||
943 | |||||||||||||
944 | /// Produce a widened version of the call instruction. | ||||||||||||
945 | void execute(VPTransformState &State) override; | ||||||||||||
946 | |||||||||||||
947 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
948 | /// Print the recipe. | ||||||||||||
949 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
950 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
951 | #endif | ||||||||||||
952 | }; | ||||||||||||
953 | |||||||||||||
954 | /// A recipe for widening select instructions. | ||||||||||||
955 | class VPWidenSelectRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
956 | |||||||||||||
957 | /// Is the condition of the select loop invariant? | ||||||||||||
958 | bool InvariantCond; | ||||||||||||
959 | |||||||||||||
960 | public: | ||||||||||||
961 | template <typename IterT> | ||||||||||||
962 | VPWidenSelectRecipe(SelectInst &I, iterator_range<IterT> Operands, | ||||||||||||
963 | bool InvariantCond) | ||||||||||||
964 | : VPRecipeBase(VPRecipeBase::VPWidenSelectSC, Operands), | ||||||||||||
965 | VPValue(VPValue::VPVWidenSelectSC, &I, this), | ||||||||||||
966 | InvariantCond(InvariantCond) {} | ||||||||||||
967 | |||||||||||||
968 | ~VPWidenSelectRecipe() override = default; | ||||||||||||
969 | |||||||||||||
970 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
971 | static inline bool classof(const VPDef *D) { | ||||||||||||
972 | return D->getVPDefID() == VPRecipeBase::VPWidenSelectSC; | ||||||||||||
973 | } | ||||||||||||
974 | |||||||||||||
975 | /// Produce a widened version of the select instruction. | ||||||||||||
976 | void execute(VPTransformState &State) override; | ||||||||||||
977 | |||||||||||||
978 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
979 | /// Print the recipe. | ||||||||||||
980 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
981 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
982 | #endif | ||||||||||||
983 | }; | ||||||||||||
984 | |||||||||||||
985 | /// A recipe for handling GEP instructions. | ||||||||||||
986 | class VPWidenGEPRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
987 | bool IsPtrLoopInvariant; | ||||||||||||
988 | SmallBitVector IsIndexLoopInvariant; | ||||||||||||
989 | |||||||||||||
990 | public: | ||||||||||||
991 | template <typename IterT> | ||||||||||||
992 | VPWidenGEPRecipe(GetElementPtrInst *GEP, iterator_range<IterT> Operands) | ||||||||||||
993 | : VPRecipeBase(VPRecipeBase::VPWidenGEPSC, Operands), | ||||||||||||
994 | VPValue(VPWidenGEPSC, GEP, this), | ||||||||||||
995 | IsIndexLoopInvariant(GEP->getNumIndices(), false) {} | ||||||||||||
996 | |||||||||||||
997 | template <typename IterT> | ||||||||||||
998 | VPWidenGEPRecipe(GetElementPtrInst *GEP, iterator_range<IterT> Operands, | ||||||||||||
999 | Loop *OrigLoop) | ||||||||||||
1000 | : VPRecipeBase(VPRecipeBase::VPWidenGEPSC, Operands), | ||||||||||||
1001 | VPValue(VPValue::VPVWidenGEPSC, GEP, this), | ||||||||||||
1002 | IsIndexLoopInvariant(GEP->getNumIndices(), false) { | ||||||||||||
1003 | IsPtrLoopInvariant = OrigLoop->isLoopInvariant(GEP->getPointerOperand()); | ||||||||||||
1004 | for (auto Index : enumerate(GEP->indices())) | ||||||||||||
1005 | IsIndexLoopInvariant[Index.index()] = | ||||||||||||
1006 | OrigLoop->isLoopInvariant(Index.value().get()); | ||||||||||||
1007 | } | ||||||||||||
1008 | ~VPWidenGEPRecipe() override = default; | ||||||||||||
1009 | |||||||||||||
1010 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1011 | static inline bool classof(const VPDef *D) { | ||||||||||||
1012 | return D->getVPDefID() == VPRecipeBase::VPWidenGEPSC; | ||||||||||||
1013 | } | ||||||||||||
1014 | |||||||||||||
1015 | /// Generate the gep nodes. | ||||||||||||
1016 | void execute(VPTransformState &State) override; | ||||||||||||
1017 | |||||||||||||
1018 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
1019 | /// Print the recipe. | ||||||||||||
1020 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
1021 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
1022 | #endif | ||||||||||||
1023 | }; | ||||||||||||
1024 | |||||||||||||
1025 | /// A recipe for handling phi nodes of integer and floating-point inductions, | ||||||||||||
1026 | /// producing their vector and scalar values. | ||||||||||||
1027 | class VPWidenIntOrFpInductionRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
1028 | PHINode *IV; | ||||||||||||
1029 | const InductionDescriptor &IndDesc; | ||||||||||||
1030 | |||||||||||||
1031 | public: | ||||||||||||
1032 | VPWidenIntOrFpInductionRecipe(PHINode *IV, VPValue *Start, | ||||||||||||
1033 | const InductionDescriptor &IndDesc) | ||||||||||||
1034 | : VPRecipeBase(VPWidenIntOrFpInductionSC, {Start}), VPValue(IV, this), | ||||||||||||
1035 | IV(IV), IndDesc(IndDesc) {} | ||||||||||||
1036 | |||||||||||||
1037 | VPWidenIntOrFpInductionRecipe(PHINode *IV, VPValue *Start, | ||||||||||||
1038 | const InductionDescriptor &IndDesc, | ||||||||||||
1039 | TruncInst *Trunc) | ||||||||||||
1040 | : VPRecipeBase(VPWidenIntOrFpInductionSC, {Start}), VPValue(Trunc, this), | ||||||||||||
1041 | IV(IV), IndDesc(IndDesc) {} | ||||||||||||
1042 | |||||||||||||
1043 | ~VPWidenIntOrFpInductionRecipe() override = default; | ||||||||||||
1044 | |||||||||||||
1045 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1046 | static inline bool classof(const VPDef *D) { | ||||||||||||
1047 | return D->getVPDefID() == VPRecipeBase::VPWidenIntOrFpInductionSC; | ||||||||||||
1048 | } | ||||||||||||
1049 | |||||||||||||
1050 | /// Generate the vectorized and scalarized versions of the phi node as | ||||||||||||
1051 | /// needed by their users. | ||||||||||||
1052 | void execute(VPTransformState &State) override; | ||||||||||||
1053 | |||||||||||||
1054 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
1055 | /// Print the recipe. | ||||||||||||
1056 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
1057 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
1058 | #endif | ||||||||||||
1059 | |||||||||||||
1060 | /// Returns the start value of the induction. | ||||||||||||
1061 | VPValue *getStartValue() { return getOperand(0); } | ||||||||||||
1062 | |||||||||||||
1063 | /// Returns the first defined value as TruncInst, if it is one or nullptr | ||||||||||||
1064 | /// otherwise. | ||||||||||||
1065 | TruncInst *getTruncInst() { | ||||||||||||
1066 | return dyn_cast_or_null<TruncInst>(getVPValue(0)->getUnderlyingValue()); | ||||||||||||
1067 | } | ||||||||||||
1068 | const TruncInst *getTruncInst() const { | ||||||||||||
1069 | return dyn_cast_or_null<TruncInst>(getVPValue(0)->getUnderlyingValue()); | ||||||||||||
1070 | } | ||||||||||||
1071 | |||||||||||||
1072 | /// Returns the induction descriptor for the recipe. | ||||||||||||
1073 | const InductionDescriptor &getInductionDescriptor() const { return IndDesc; } | ||||||||||||
1074 | }; | ||||||||||||
1075 | |||||||||||||
1076 | /// A pure virtual base class for all recipes modeling header phis, including | ||||||||||||
1077 | /// phis for first order recurrences, pointer inductions and reductions. The | ||||||||||||
1078 | /// start value is the first operand of the recipe and the incoming value from | ||||||||||||
1079 | /// the backedge is the second operand. | ||||||||||||
1080 | class VPHeaderPHIRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
1081 | protected: | ||||||||||||
1082 | VPHeaderPHIRecipe(unsigned char VPVID, unsigned char VPDefID, PHINode *Phi, | ||||||||||||
1083 | VPValue *Start = nullptr) | ||||||||||||
1084 | : VPRecipeBase(VPDefID, {}), VPValue(VPVID, Phi, this) { | ||||||||||||
1085 | if (Start) | ||||||||||||
1086 | addOperand(Start); | ||||||||||||
1087 | } | ||||||||||||
1088 | |||||||||||||
1089 | public: | ||||||||||||
1090 | ~VPHeaderPHIRecipe() override = default; | ||||||||||||
1091 | |||||||||||||
1092 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1093 | static inline bool classof(const VPRecipeBase *B) { | ||||||||||||
1094 | return B->getVPDefID() == VPRecipeBase::VPCanonicalIVPHISC || | ||||||||||||
1095 | B->getVPDefID() == VPRecipeBase::VPFirstOrderRecurrencePHISC || | ||||||||||||
1096 | B->getVPDefID() == VPRecipeBase::VPReductionPHISC || | ||||||||||||
1097 | B->getVPDefID() == VPRecipeBase::VPWidenIntOrFpInductionSC || | ||||||||||||
1098 | B->getVPDefID() == VPRecipeBase::VPWidenPHISC; | ||||||||||||
1099 | } | ||||||||||||
1100 | static inline bool classof(const VPValue *V) { | ||||||||||||
1101 | return V->getVPValueID() == VPValue::VPVCanonicalIVPHISC || | ||||||||||||
1102 | V->getVPValueID() == VPValue::VPVFirstOrderRecurrencePHISC || | ||||||||||||
1103 | V->getVPValueID() == VPValue::VPVReductionPHISC || | ||||||||||||
1104 | V->getVPValueID() == VPValue::VPVWidenIntOrFpInductionSC || | ||||||||||||
1105 | V->getVPValueID() == VPValue::VPVWidenPHISC; | ||||||||||||
1106 | } | ||||||||||||
1107 | |||||||||||||
1108 | /// Generate the phi nodes. | ||||||||||||
1109 | void execute(VPTransformState &State) override = 0; | ||||||||||||
1110 | |||||||||||||
1111 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
1112 | /// Print the recipe. | ||||||||||||
1113 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
1114 | VPSlotTracker &SlotTracker) const override = 0; | ||||||||||||
1115 | #endif | ||||||||||||
1116 | |||||||||||||
1117 | /// Returns the start value of the phi, if one is set. | ||||||||||||
1118 | VPValue *getStartValue() { | ||||||||||||
1119 | return getNumOperands() == 0 ? nullptr : getOperand(0); | ||||||||||||
1120 | } | ||||||||||||
1121 | |||||||||||||
1122 | /// Returns the incoming value from the loop backedge. | ||||||||||||
1123 | VPValue *getBackedgeValue() { | ||||||||||||
1124 | return getOperand(1); | ||||||||||||
1125 | } | ||||||||||||
1126 | |||||||||||||
1127 | /// Returns the backedge value as a recipe. The backedge value is guaranteed | ||||||||||||
1128 | /// to be a recipe. | ||||||||||||
1129 | VPRecipeBase *getBackedgeRecipe() { | ||||||||||||
1130 | return cast<VPRecipeBase>(getBackedgeValue()->getDef()); | ||||||||||||
1131 | } | ||||||||||||
1132 | }; | ||||||||||||
1133 | |||||||||||||
1134 | /// A recipe for handling header phis that are widened in the vector loop. | ||||||||||||
1135 | /// In the VPlan native path, all incoming VPValues & VPBasicBlock pairs are | ||||||||||||
1136 | /// managed in the recipe directly. | ||||||||||||
1137 | class VPWidenPHIRecipe : public VPHeaderPHIRecipe { | ||||||||||||
1138 | /// List of incoming blocks. Only used in the VPlan native path. | ||||||||||||
1139 | SmallVector<VPBasicBlock *, 2> IncomingBlocks; | ||||||||||||
1140 | |||||||||||||
1141 | public: | ||||||||||||
1142 | /// Create a new VPWidenPHIRecipe for \p Phi with start value \p Start. | ||||||||||||
1143 | VPWidenPHIRecipe(PHINode *Phi, VPValue *Start = nullptr) | ||||||||||||
1144 | : VPHeaderPHIRecipe(VPVWidenPHISC, VPWidenPHISC, Phi) { | ||||||||||||
1145 | if (Start) | ||||||||||||
1146 | addOperand(Start); | ||||||||||||
1147 | } | ||||||||||||
1148 | |||||||||||||
1149 | ~VPWidenPHIRecipe() override = default; | ||||||||||||
1150 | |||||||||||||
1151 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1152 | static inline bool classof(const VPRecipeBase *B) { | ||||||||||||
1153 | return B->getVPDefID() == VPRecipeBase::VPWidenPHISC; | ||||||||||||
1154 | } | ||||||||||||
1155 | static inline bool classof(const VPHeaderPHIRecipe *R) { | ||||||||||||
1156 | return R->getVPDefID() == VPRecipeBase::VPWidenPHISC; | ||||||||||||
1157 | } | ||||||||||||
1158 | static inline bool classof(const VPValue *V) { | ||||||||||||
1159 | return V->getVPValueID() == VPValue::VPVWidenPHISC; | ||||||||||||
1160 | } | ||||||||||||
1161 | |||||||||||||
1162 | /// Generate the phi/select nodes. | ||||||||||||
1163 | void execute(VPTransformState &State) override; | ||||||||||||
1164 | |||||||||||||
1165 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
1166 | /// Print the recipe. | ||||||||||||
1167 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
1168 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
1169 | #endif | ||||||||||||
1170 | |||||||||||||
1171 | /// Adds a pair (\p IncomingV, \p IncomingBlock) to the phi. | ||||||||||||
1172 | void addIncoming(VPValue *IncomingV, VPBasicBlock *IncomingBlock) { | ||||||||||||
1173 | addOperand(IncomingV); | ||||||||||||
1174 | IncomingBlocks.push_back(IncomingBlock); | ||||||||||||
1175 | } | ||||||||||||
1176 | |||||||||||||
1177 | /// Returns the \p I th incoming VPBasicBlock. | ||||||||||||
1178 | VPBasicBlock *getIncomingBlock(unsigned I) { return IncomingBlocks[I]; } | ||||||||||||
1179 | |||||||||||||
1180 | /// Returns the \p I th incoming VPValue. | ||||||||||||
1181 | VPValue *getIncomingValue(unsigned I) { return getOperand(I); } | ||||||||||||
1182 | }; | ||||||||||||
1183 | |||||||||||||
1184 | /// A recipe for handling first-order recurrence phis. The start value is the | ||||||||||||
1185 | /// first operand of the recipe and the incoming value from the backedge is the | ||||||||||||
1186 | /// second operand. | ||||||||||||
1187 | struct VPFirstOrderRecurrencePHIRecipe : public VPHeaderPHIRecipe { | ||||||||||||
1188 | VPFirstOrderRecurrencePHIRecipe(PHINode *Phi, VPValue &Start) | ||||||||||||
1189 | : VPHeaderPHIRecipe(VPVFirstOrderRecurrencePHISC, | ||||||||||||
1190 | VPFirstOrderRecurrencePHISC, Phi, &Start) {} | ||||||||||||
1191 | |||||||||||||
1192 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1193 | static inline bool classof(const VPRecipeBase *R) { | ||||||||||||
1194 | return R->getVPDefID() == VPRecipeBase::VPFirstOrderRecurrencePHISC; | ||||||||||||
1195 | } | ||||||||||||
1196 | static inline bool classof(const VPHeaderPHIRecipe *R) { | ||||||||||||
1197 | return R->getVPDefID() == VPRecipeBase::VPFirstOrderRecurrencePHISC; | ||||||||||||
1198 | } | ||||||||||||
1199 | static inline bool classof(const VPValue *V) { | ||||||||||||
1200 | return V->getVPValueID() == VPValue::VPVFirstOrderRecurrencePHISC; | ||||||||||||
1201 | } | ||||||||||||
1202 | |||||||||||||
1203 | void execute(VPTransformState &State) override; | ||||||||||||
1204 | |||||||||||||
1205 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
1206 | /// Print the recipe. | ||||||||||||
1207 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
1208 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
1209 | #endif | ||||||||||||
1210 | }; | ||||||||||||
1211 | |||||||||||||
1212 | /// A recipe for handling reduction phis. The start value is the first operand | ||||||||||||
1213 | /// of the recipe and the incoming value from the backedge is the second | ||||||||||||
1214 | /// operand. | ||||||||||||
1215 | class VPReductionPHIRecipe : public VPHeaderPHIRecipe { | ||||||||||||
1216 | /// Descriptor for the reduction. | ||||||||||||
1217 | const RecurrenceDescriptor &RdxDesc; | ||||||||||||
1218 | |||||||||||||
1219 | /// The phi is part of an in-loop reduction. | ||||||||||||
1220 | bool IsInLoop; | ||||||||||||
1221 | |||||||||||||
1222 | /// The phi is part of an ordered reduction. Requires IsInLoop to be true. | ||||||||||||
1223 | bool IsOrdered; | ||||||||||||
1224 | |||||||||||||
1225 | public: | ||||||||||||
1226 | /// Create a new VPReductionPHIRecipe for the reduction \p Phi described by \p | ||||||||||||
1227 | /// RdxDesc. | ||||||||||||
1228 | VPReductionPHIRecipe(PHINode *Phi, const RecurrenceDescriptor &RdxDesc, | ||||||||||||
1229 | VPValue &Start, bool IsInLoop = false, | ||||||||||||
1230 | bool IsOrdered = false) | ||||||||||||
1231 | : VPHeaderPHIRecipe(VPVReductionPHISC, VPReductionPHISC, Phi, &Start), | ||||||||||||
1232 | RdxDesc(RdxDesc), IsInLoop(IsInLoop), IsOrdered(IsOrdered) { | ||||||||||||
1233 | 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", 1233, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
1234 | } | ||||||||||||
1235 | |||||||||||||
1236 | ~VPReductionPHIRecipe() override = default; | ||||||||||||
1237 | |||||||||||||
1238 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1239 | static inline bool classof(const VPRecipeBase *R) { | ||||||||||||
1240 | return R->getVPDefID() == VPRecipeBase::VPReductionPHISC; | ||||||||||||
1241 | } | ||||||||||||
1242 | static inline bool classof(const VPHeaderPHIRecipe *R) { | ||||||||||||
1243 | return R->getVPDefID() == VPRecipeBase::VPReductionPHISC; | ||||||||||||
1244 | } | ||||||||||||
1245 | static inline bool classof(const VPValue *V) { | ||||||||||||
1246 | return V->getVPValueID() == VPValue::VPVReductionPHISC; | ||||||||||||
1247 | } | ||||||||||||
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 | const RecurrenceDescriptor &getRecurrenceDescriptor() const { | ||||||||||||
1259 | return RdxDesc; | ||||||||||||
1260 | } | ||||||||||||
1261 | |||||||||||||
1262 | /// Returns true, if the phi is part of an ordered reduction. | ||||||||||||
1263 | bool isOrdered() const { return IsOrdered; } | ||||||||||||
1264 | |||||||||||||
1265 | /// Returns true, if the phi is part of an in-loop reduction. | ||||||||||||
1266 | bool isInLoop() const { return IsInLoop; } | ||||||||||||
1267 | }; | ||||||||||||
1268 | |||||||||||||
1269 | /// A recipe for vectorizing a phi-node as a sequence of mask-based select | ||||||||||||
1270 | /// instructions. | ||||||||||||
1271 | class VPBlendRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
1272 | PHINode *Phi; | ||||||||||||
1273 | |||||||||||||
1274 | public: | ||||||||||||
1275 | /// The blend operation is a User of the incoming values and of their | ||||||||||||
1276 | /// respective masks, ordered [I0, M0, I1, M1, ...]. Note that a single value | ||||||||||||
1277 | /// might be incoming with a full mask for which there is no VPValue. | ||||||||||||
1278 | VPBlendRecipe(PHINode *Phi, ArrayRef<VPValue *> Operands) | ||||||||||||
1279 | : VPRecipeBase(VPBlendSC, Operands), | ||||||||||||
1280 | VPValue(VPValue::VPVBlendSC, Phi, this), Phi(Phi) { | ||||||||||||
1281 | 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", 1284, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
1282 | ((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", 1284, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
1283 | "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", 1284, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
1284 | "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", 1284, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
1285 | } | ||||||||||||
1286 | |||||||||||||
1287 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1288 | static inline bool classof(const VPDef *D) { | ||||||||||||
1289 | return D->getVPDefID() == VPRecipeBase::VPBlendSC; | ||||||||||||
1290 | } | ||||||||||||
1291 | |||||||||||||
1292 | /// Return the number of incoming values, taking into account that a single | ||||||||||||
1293 | /// incoming value has no mask. | ||||||||||||
1294 | unsigned getNumIncomingValues() const { return (getNumOperands() + 1) / 2; } | ||||||||||||
1295 | |||||||||||||
1296 | /// Return incoming value number \p Idx. | ||||||||||||
1297 | VPValue *getIncomingValue(unsigned Idx) const { return getOperand(Idx * 2); } | ||||||||||||
1298 | |||||||||||||
1299 | /// Return mask number \p Idx. | ||||||||||||
1300 | VPValue *getMask(unsigned Idx) const { return getOperand(Idx * 2 + 1); } | ||||||||||||
1301 | |||||||||||||
1302 | /// Generate the phi/select nodes. | ||||||||||||
1303 | void execute(VPTransformState &State) override; | ||||||||||||
1304 | |||||||||||||
1305 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
1306 | /// Print the recipe. | ||||||||||||
1307 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
1308 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
1309 | #endif | ||||||||||||
1310 | }; | ||||||||||||
1311 | |||||||||||||
1312 | /// VPInterleaveRecipe is a recipe for transforming an interleave group of load | ||||||||||||
1313 | /// or stores into one wide load/store and shuffles. The first operand of a | ||||||||||||
1314 | /// VPInterleave recipe is the address, followed by the stored values, followed | ||||||||||||
1315 | /// by an optional mask. | ||||||||||||
1316 | class VPInterleaveRecipe : public VPRecipeBase { | ||||||||||||
1317 | const InterleaveGroup<Instruction> *IG; | ||||||||||||
1318 | |||||||||||||
1319 | bool HasMask = false; | ||||||||||||
1320 | |||||||||||||
1321 | public: | ||||||||||||
1322 | VPInterleaveRecipe(const InterleaveGroup<Instruction> *IG, VPValue *Addr, | ||||||||||||
1323 | ArrayRef<VPValue *> StoredValues, VPValue *Mask) | ||||||||||||
1324 | : VPRecipeBase(VPInterleaveSC, {Addr}), IG(IG) { | ||||||||||||
1325 | for (unsigned i = 0; i < IG->getFactor(); ++i) | ||||||||||||
1326 | if (Instruction *I = IG->getMember(i)) { | ||||||||||||
1327 | if (I->getType()->isVoidTy()) | ||||||||||||
1328 | continue; | ||||||||||||
1329 | new VPValue(I, this); | ||||||||||||
1330 | } | ||||||||||||
1331 | |||||||||||||
1332 | for (auto *SV : StoredValues) | ||||||||||||
1333 | addOperand(SV); | ||||||||||||
1334 | if (Mask) { | ||||||||||||
1335 | HasMask = true; | ||||||||||||
1336 | addOperand(Mask); | ||||||||||||
1337 | } | ||||||||||||
1338 | } | ||||||||||||
1339 | ~VPInterleaveRecipe() override = default; | ||||||||||||
1340 | |||||||||||||
1341 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1342 | static inline bool classof(const VPDef *D) { | ||||||||||||
1343 | return D->getVPDefID() == VPRecipeBase::VPInterleaveSC; | ||||||||||||
1344 | } | ||||||||||||
1345 | |||||||||||||
1346 | /// Return the address accessed by this recipe. | ||||||||||||
1347 | VPValue *getAddr() const { | ||||||||||||
1348 | return getOperand(0); // Address is the 1st, mandatory operand. | ||||||||||||
1349 | } | ||||||||||||
1350 | |||||||||||||
1351 | /// Return the mask used by this recipe. Note that a full mask is represented | ||||||||||||
1352 | /// by a nullptr. | ||||||||||||
1353 | VPValue *getMask() const { | ||||||||||||
1354 | // Mask is optional and therefore the last, currently 2nd operand. | ||||||||||||
1355 | return HasMask ? getOperand(getNumOperands() - 1) : nullptr; | ||||||||||||
1356 | } | ||||||||||||
1357 | |||||||||||||
1358 | /// Return the VPValues stored by this interleave group. If it is a load | ||||||||||||
1359 | /// interleave group, return an empty ArrayRef. | ||||||||||||
1360 | ArrayRef<VPValue *> getStoredValues() const { | ||||||||||||
1361 | // The first operand is the address, followed by the stored values, followed | ||||||||||||
1362 | // by an optional mask. | ||||||||||||
1363 | return ArrayRef<VPValue *>(op_begin(), getNumOperands()) | ||||||||||||
1364 | .slice(1, getNumStoreOperands()); | ||||||||||||
1365 | } | ||||||||||||
1366 | |||||||||||||
1367 | /// Generate the wide load or store, and shuffles. | ||||||||||||
1368 | void execute(VPTransformState &State) override; | ||||||||||||
1369 | |||||||||||||
1370 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
1371 | /// Print the recipe. | ||||||||||||
1372 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
1373 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
1374 | #endif | ||||||||||||
1375 | |||||||||||||
1376 | const InterleaveGroup<Instruction> *getInterleaveGroup() { return IG; } | ||||||||||||
1377 | |||||||||||||
1378 | /// Returns the number of stored operands of this interleave group. Returns 0 | ||||||||||||
1379 | /// for load interleave groups. | ||||||||||||
1380 | unsigned getNumStoreOperands() const { | ||||||||||||
1381 | return getNumOperands() - (HasMask ? 2 : 1); | ||||||||||||
1382 | } | ||||||||||||
1383 | }; | ||||||||||||
1384 | |||||||||||||
1385 | /// A recipe to represent inloop reduction operations, performing a reduction on | ||||||||||||
1386 | /// a vector operand into a scalar value, and adding the result to a chain. | ||||||||||||
1387 | /// The Operands are {ChainOp, VecOp, [Condition]}. | ||||||||||||
1388 | class VPReductionRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
1389 | /// The recurrence decriptor for the reduction in question. | ||||||||||||
1390 | const RecurrenceDescriptor *RdxDesc; | ||||||||||||
1391 | /// Pointer to the TTI, needed to create the target reduction | ||||||||||||
1392 | const TargetTransformInfo *TTI; | ||||||||||||
1393 | |||||||||||||
1394 | public: | ||||||||||||
1395 | VPReductionRecipe(const RecurrenceDescriptor *R, Instruction *I, | ||||||||||||
1396 | VPValue *ChainOp, VPValue *VecOp, VPValue *CondOp, | ||||||||||||
1397 | const TargetTransformInfo *TTI) | ||||||||||||
1398 | : VPRecipeBase(VPRecipeBase::VPReductionSC, {ChainOp, VecOp}), | ||||||||||||
1399 | VPValue(VPValue::VPVReductionSC, I, this), RdxDesc(R), TTI(TTI) { | ||||||||||||
1400 | if (CondOp) | ||||||||||||
1401 | addOperand(CondOp); | ||||||||||||
1402 | } | ||||||||||||
1403 | |||||||||||||
1404 | ~VPReductionRecipe() override = default; | ||||||||||||
1405 | |||||||||||||
1406 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1407 | static inline bool classof(const VPValue *V) { | ||||||||||||
1408 | return V->getVPValueID() == VPValue::VPVReductionSC; | ||||||||||||
1409 | } | ||||||||||||
1410 | |||||||||||||
1411 | /// Generate the reduction in the loop | ||||||||||||
1412 | void execute(VPTransformState &State) override; | ||||||||||||
1413 | |||||||||||||
1414 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
1415 | /// Print the recipe. | ||||||||||||
1416 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
1417 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
1418 | #endif | ||||||||||||
1419 | |||||||||||||
1420 | /// The VPValue of the scalar Chain being accumulated. | ||||||||||||
1421 | VPValue *getChainOp() const { return getOperand(0); } | ||||||||||||
1422 | /// The VPValue of the vector value to be reduced. | ||||||||||||
1423 | VPValue *getVecOp() const { return getOperand(1); } | ||||||||||||
1424 | /// The VPValue of the condition for the block. | ||||||||||||
1425 | VPValue *getCondOp() const { | ||||||||||||
1426 | return getNumOperands() > 2 ? getOperand(2) : nullptr; | ||||||||||||
1427 | } | ||||||||||||
1428 | }; | ||||||||||||
1429 | |||||||||||||
1430 | /// VPReplicateRecipe replicates a given instruction producing multiple scalar | ||||||||||||
1431 | /// copies of the original scalar type, one per lane, instead of producing a | ||||||||||||
1432 | /// single copy of widened type for all lanes. If the instruction is known to be | ||||||||||||
1433 | /// uniform only one copy, per lane zero, will be generated. | ||||||||||||
1434 | class VPReplicateRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
1435 | /// Indicator if only a single replica per lane is needed. | ||||||||||||
1436 | bool IsUniform; | ||||||||||||
1437 | |||||||||||||
1438 | /// Indicator if the replicas are also predicated. | ||||||||||||
1439 | bool IsPredicated; | ||||||||||||
1440 | |||||||||||||
1441 | /// Indicator if the scalar values should also be packed into a vector. | ||||||||||||
1442 | bool AlsoPack; | ||||||||||||
1443 | |||||||||||||
1444 | public: | ||||||||||||
1445 | template <typename IterT> | ||||||||||||
1446 | VPReplicateRecipe(Instruction *I, iterator_range<IterT> Operands, | ||||||||||||
1447 | bool IsUniform, bool IsPredicated = false) | ||||||||||||
1448 | : VPRecipeBase(VPReplicateSC, Operands), VPValue(VPVReplicateSC, I, this), | ||||||||||||
1449 | IsUniform(IsUniform), IsPredicated(IsPredicated) { | ||||||||||||
1450 | // Retain the previous behavior of predicateInstructions(), where an | ||||||||||||
1451 | // insert-element of a predicated instruction got hoisted into the | ||||||||||||
1452 | // predicated basic block iff it was its only user. This is achieved by | ||||||||||||
1453 | // having predicated instructions also pack their values into a vector by | ||||||||||||
1454 | // default unless they have a replicated user which uses their scalar value. | ||||||||||||
1455 | AlsoPack = IsPredicated && !I->use_empty(); | ||||||||||||
1456 | } | ||||||||||||
1457 | |||||||||||||
1458 | ~VPReplicateRecipe() override = default; | ||||||||||||
1459 | |||||||||||||
1460 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1461 | static inline bool classof(const VPDef *D) { | ||||||||||||
1462 | return D->getVPDefID() == VPRecipeBase::VPReplicateSC; | ||||||||||||
1463 | } | ||||||||||||
1464 | |||||||||||||
1465 | static inline bool classof(const VPValue *V) { | ||||||||||||
1466 | return V->getVPValueID() == VPValue::VPVReplicateSC; | ||||||||||||
1467 | } | ||||||||||||
1468 | |||||||||||||
1469 | /// Generate replicas of the desired Ingredient. Replicas will be generated | ||||||||||||
1470 | /// for all parts and lanes unless a specific part and lane are specified in | ||||||||||||
1471 | /// the \p State. | ||||||||||||
1472 | void execute(VPTransformState &State) override; | ||||||||||||
1473 | |||||||||||||
1474 | void setAlsoPack(bool Pack) { AlsoPack = Pack; } | ||||||||||||
1475 | |||||||||||||
1476 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
1477 | /// Print the recipe. | ||||||||||||
1478 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
1479 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
1480 | #endif | ||||||||||||
1481 | |||||||||||||
1482 | bool isUniform() const { return IsUniform; } | ||||||||||||
1483 | |||||||||||||
1484 | bool isPacked() const { return AlsoPack; } | ||||||||||||
1485 | |||||||||||||
1486 | bool isPredicated() const { return IsPredicated; } | ||||||||||||
1487 | }; | ||||||||||||
1488 | |||||||||||||
1489 | /// A recipe for generating conditional branches on the bits of a mask. | ||||||||||||
1490 | class VPBranchOnMaskRecipe : public VPRecipeBase { | ||||||||||||
1491 | public: | ||||||||||||
1492 | VPBranchOnMaskRecipe(VPValue *BlockInMask) | ||||||||||||
1493 | : VPRecipeBase(VPBranchOnMaskSC, {}) { | ||||||||||||
1494 | if (BlockInMask) // nullptr means all-one mask. | ||||||||||||
1495 | addOperand(BlockInMask); | ||||||||||||
1496 | } | ||||||||||||
1497 | |||||||||||||
1498 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1499 | static inline bool classof(const VPDef *D) { | ||||||||||||
1500 | return D->getVPDefID() == VPRecipeBase::VPBranchOnMaskSC; | ||||||||||||
1501 | } | ||||||||||||
1502 | |||||||||||||
1503 | /// Generate the extraction of the appropriate bit from the block mask and the | ||||||||||||
1504 | /// conditional branch. | ||||||||||||
1505 | void execute(VPTransformState &State) override; | ||||||||||||
1506 | |||||||||||||
1507 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
1508 | /// Print the recipe. | ||||||||||||
1509 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
1510 | VPSlotTracker &SlotTracker) const override { | ||||||||||||
1511 | O << Indent << "BRANCH-ON-MASK "; | ||||||||||||
1512 | if (VPValue *Mask = getMask()) | ||||||||||||
1513 | Mask->printAsOperand(O, SlotTracker); | ||||||||||||
1514 | else | ||||||||||||
1515 | O << " All-One"; | ||||||||||||
1516 | } | ||||||||||||
1517 | #endif | ||||||||||||
1518 | |||||||||||||
1519 | /// Return the mask used by this recipe. Note that a full mask is represented | ||||||||||||
1520 | /// by a nullptr. | ||||||||||||
1521 | VPValue *getMask() const { | ||||||||||||
1522 | 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", 1522, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
1523 | // Mask is optional. | ||||||||||||
1524 | return getNumOperands() == 1 ? getOperand(0) : nullptr; | ||||||||||||
1525 | } | ||||||||||||
1526 | }; | ||||||||||||
1527 | |||||||||||||
1528 | /// VPPredInstPHIRecipe is a recipe for generating the phi nodes needed when | ||||||||||||
1529 | /// control converges back from a Branch-on-Mask. The phi nodes are needed in | ||||||||||||
1530 | /// order to merge values that are set under such a branch and feed their uses. | ||||||||||||
1531 | /// The phi nodes can be scalar or vector depending on the users of the value. | ||||||||||||
1532 | /// This recipe works in concert with VPBranchOnMaskRecipe. | ||||||||||||
1533 | class VPPredInstPHIRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
1534 | public: | ||||||||||||
1535 | /// Construct a VPPredInstPHIRecipe given \p PredInst whose value needs a phi | ||||||||||||
1536 | /// nodes after merging back from a Branch-on-Mask. | ||||||||||||
1537 | VPPredInstPHIRecipe(VPValue *PredV) | ||||||||||||
1538 | : VPRecipeBase(VPPredInstPHISC, PredV), | ||||||||||||
1539 | VPValue(VPValue::VPVPredInstPHI, nullptr, this) {} | ||||||||||||
1540 | ~VPPredInstPHIRecipe() override = default; | ||||||||||||
1541 | |||||||||||||
1542 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1543 | static inline bool classof(const VPDef *D) { | ||||||||||||
1544 | return D->getVPDefID() == VPRecipeBase::VPPredInstPHISC; | ||||||||||||
1545 | } | ||||||||||||
1546 | |||||||||||||
1547 | /// Generates phi nodes for live-outs as needed to retain SSA form. | ||||||||||||
1548 | void execute(VPTransformState &State) override; | ||||||||||||
1549 | |||||||||||||
1550 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
1551 | /// Print the recipe. | ||||||||||||
1552 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
1553 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
1554 | #endif | ||||||||||||
1555 | }; | ||||||||||||
1556 | |||||||||||||
1557 | /// A Recipe for widening load/store operations. | ||||||||||||
1558 | /// The recipe uses the following VPValues: | ||||||||||||
1559 | /// - For load: Address, optional mask | ||||||||||||
1560 | /// - For store: Address, stored value, optional mask | ||||||||||||
1561 | /// TODO: We currently execute only per-part unless a specific instance is | ||||||||||||
1562 | /// provided. | ||||||||||||
1563 | class VPWidenMemoryInstructionRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
1564 | Instruction &Ingredient; | ||||||||||||
1565 | |||||||||||||
1566 | // Whether the loaded-from / stored-to addresses are consecutive. | ||||||||||||
1567 | bool Consecutive; | ||||||||||||
1568 | |||||||||||||
1569 | // Whether the consecutive loaded/stored addresses are in reverse order. | ||||||||||||
1570 | bool Reverse; | ||||||||||||
1571 | |||||||||||||
1572 | void setMask(VPValue *Mask) { | ||||||||||||
1573 | if (!Mask) | ||||||||||||
1574 | return; | ||||||||||||
1575 | addOperand(Mask); | ||||||||||||
1576 | } | ||||||||||||
1577 | |||||||||||||
1578 | bool isMasked() const { | ||||||||||||
1579 | return isStore() ? getNumOperands() == 3 : getNumOperands() == 2; | ||||||||||||
1580 | } | ||||||||||||
1581 | |||||||||||||
1582 | public: | ||||||||||||
1583 | VPWidenMemoryInstructionRecipe(LoadInst &Load, VPValue *Addr, VPValue *Mask, | ||||||||||||
1584 | bool Consecutive, bool Reverse) | ||||||||||||
1585 | : VPRecipeBase(VPWidenMemoryInstructionSC, {Addr}), | ||||||||||||
1586 | VPValue(VPValue::VPVMemoryInstructionSC, &Load, this), Ingredient(Load), | ||||||||||||
1587 | Consecutive(Consecutive), Reverse(Reverse) { | ||||||||||||
1588 | 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", 1588, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
1589 | setMask(Mask); | ||||||||||||
1590 | } | ||||||||||||
1591 | |||||||||||||
1592 | VPWidenMemoryInstructionRecipe(StoreInst &Store, VPValue *Addr, | ||||||||||||
1593 | VPValue *StoredValue, VPValue *Mask, | ||||||||||||
1594 | bool Consecutive, bool Reverse) | ||||||||||||
1595 | : VPRecipeBase(VPWidenMemoryInstructionSC, {Addr, StoredValue}), | ||||||||||||
1596 | VPValue(VPValue::VPVMemoryInstructionSC, &Store, this), | ||||||||||||
1597 | Ingredient(Store), Consecutive(Consecutive), Reverse(Reverse) { | ||||||||||||
1598 | 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", 1598, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
1599 | setMask(Mask); | ||||||||||||
1600 | } | ||||||||||||
1601 | |||||||||||||
1602 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1603 | static inline bool classof(const VPDef *D) { | ||||||||||||
1604 | return D->getVPDefID() == VPRecipeBase::VPWidenMemoryInstructionSC; | ||||||||||||
1605 | } | ||||||||||||
1606 | |||||||||||||
1607 | /// Return the address accessed by this recipe. | ||||||||||||
1608 | VPValue *getAddr() const { | ||||||||||||
1609 | return getOperand(0); // Address is the 1st, mandatory operand. | ||||||||||||
1610 | } | ||||||||||||
1611 | |||||||||||||
1612 | /// Return the mask used by this recipe. Note that a full mask is represented | ||||||||||||
1613 | /// by a nullptr. | ||||||||||||
1614 | VPValue *getMask() const { | ||||||||||||
1615 | // Mask is optional and therefore the last operand. | ||||||||||||
1616 | return isMasked() ? getOperand(getNumOperands() - 1) : nullptr; | ||||||||||||
1617 | } | ||||||||||||
1618 | |||||||||||||
1619 | /// Returns true if this recipe is a store. | ||||||||||||
1620 | bool isStore() const { return isa<StoreInst>(Ingredient); } | ||||||||||||
1621 | |||||||||||||
1622 | /// Return the address accessed by this recipe. | ||||||||||||
1623 | VPValue *getStoredValue() const { | ||||||||||||
1624 | 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", 1624, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
1625 | return getOperand(1); // Stored value is the 2nd, mandatory operand. | ||||||||||||
1626 | } | ||||||||||||
1627 | |||||||||||||
1628 | // Return whether the loaded-from / stored-to addresses are consecutive. | ||||||||||||
1629 | bool isConsecutive() const { return Consecutive; } | ||||||||||||
1630 | |||||||||||||
1631 | // Return whether the consecutive loaded/stored addresses are in reverse | ||||||||||||
1632 | // order. | ||||||||||||
1633 | bool isReverse() const { return Reverse; } | ||||||||||||
1634 | |||||||||||||
1635 | /// Generate the wide load/store. | ||||||||||||
1636 | void execute(VPTransformState &State) override; | ||||||||||||
1637 | |||||||||||||
1638 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
1639 | /// Print the recipe. | ||||||||||||
1640 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
1641 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
1642 | #endif | ||||||||||||
1643 | }; | ||||||||||||
1644 | |||||||||||||
1645 | /// Canonical scalar induction phi of the vector loop. Starting at the specified | ||||||||||||
1646 | /// start value (either 0 or the resume value when vectorizing the epilogue | ||||||||||||
1647 | /// loop). VPWidenCanonicalIVRecipe represents the vector version of the | ||||||||||||
1648 | /// canonical induction variable. | ||||||||||||
1649 | class VPCanonicalIVPHIRecipe : public VPHeaderPHIRecipe { | ||||||||||||
1650 | DebugLoc DL; | ||||||||||||
1651 | |||||||||||||
1652 | public: | ||||||||||||
1653 | VPCanonicalIVPHIRecipe(VPValue *StartV, DebugLoc DL) | ||||||||||||
1654 | : VPHeaderPHIRecipe(VPValue::VPVCanonicalIVPHISC, VPCanonicalIVPHISC, | ||||||||||||
1655 | nullptr, StartV), | ||||||||||||
1656 | DL(DL) {} | ||||||||||||
1657 | |||||||||||||
1658 | ~VPCanonicalIVPHIRecipe() override = default; | ||||||||||||
1659 | |||||||||||||
1660 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1661 | static inline bool classof(const VPDef *D) { | ||||||||||||
1662 | return D->getVPDefID() == VPCanonicalIVPHISC; | ||||||||||||
1663 | } | ||||||||||||
1664 | |||||||||||||
1665 | /// Generate the canonical scalar induction phi of the vector loop. | ||||||||||||
1666 | void execute(VPTransformState &State) override; | ||||||||||||
1667 | |||||||||||||
1668 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
1669 | /// Print the recipe. | ||||||||||||
1670 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
1671 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
1672 | #endif | ||||||||||||
1673 | }; | ||||||||||||
1674 | |||||||||||||
1675 | /// A Recipe for widening the canonical induction variable of the vector loop. | ||||||||||||
1676 | class VPWidenCanonicalIVRecipe : public VPRecipeBase, public VPValue { | ||||||||||||
1677 | public: | ||||||||||||
1678 | VPWidenCanonicalIVRecipe(VPCanonicalIVPHIRecipe *CanonicalIV) | ||||||||||||
1679 | : VPRecipeBase(VPWidenCanonicalIVSC, {CanonicalIV}), | ||||||||||||
1680 | VPValue(VPValue::VPVWidenCanonicalIVSC, nullptr, this) {} | ||||||||||||
1681 | |||||||||||||
1682 | ~VPWidenCanonicalIVRecipe() override = default; | ||||||||||||
1683 | |||||||||||||
1684 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1685 | static inline bool classof(const VPDef *D) { | ||||||||||||
1686 | return D->getVPDefID() == VPRecipeBase::VPWidenCanonicalIVSC; | ||||||||||||
1687 | } | ||||||||||||
1688 | |||||||||||||
1689 | /// Generate a canonical vector induction variable of the vector loop, with | ||||||||||||
1690 | /// start = {<Part*VF, Part*VF+1, ..., Part*VF+VF-1> for 0 <= Part < UF}, and | ||||||||||||
1691 | /// step = <VF*UF, VF*UF, ..., VF*UF>. | ||||||||||||
1692 | void execute(VPTransformState &State) override; | ||||||||||||
1693 | |||||||||||||
1694 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
1695 | /// Print the recipe. | ||||||||||||
1696 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
1697 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
1698 | #endif | ||||||||||||
1699 | }; | ||||||||||||
1700 | |||||||||||||
1701 | /// VPBasicBlock serves as the leaf of the Hierarchical Control-Flow Graph. It | ||||||||||||
1702 | /// holds a sequence of zero or more VPRecipe's each representing a sequence of | ||||||||||||
1703 | /// output IR instructions. All PHI-like recipes must come before any non-PHI recipes. | ||||||||||||
1704 | class VPBasicBlock : public VPBlockBase { | ||||||||||||
1705 | public: | ||||||||||||
1706 | using RecipeListTy = iplist<VPRecipeBase>; | ||||||||||||
1707 | |||||||||||||
1708 | private: | ||||||||||||
1709 | /// The VPRecipes held in the order of output instructions to generate. | ||||||||||||
1710 | RecipeListTy Recipes; | ||||||||||||
1711 | |||||||||||||
1712 | public: | ||||||||||||
1713 | VPBasicBlock(const Twine &Name = "", VPRecipeBase *Recipe = nullptr) | ||||||||||||
1714 | : VPBlockBase(VPBasicBlockSC, Name.str()) { | ||||||||||||
1715 | if (Recipe) | ||||||||||||
1716 | appendRecipe(Recipe); | ||||||||||||
1717 | } | ||||||||||||
1718 | |||||||||||||
1719 | ~VPBasicBlock() override { | ||||||||||||
1720 | while (!Recipes.empty()) | ||||||||||||
1721 | Recipes.pop_back(); | ||||||||||||
1722 | } | ||||||||||||
1723 | |||||||||||||
1724 | /// Instruction iterators... | ||||||||||||
1725 | using iterator = RecipeListTy::iterator; | ||||||||||||
1726 | using const_iterator = RecipeListTy::const_iterator; | ||||||||||||
1727 | using reverse_iterator = RecipeListTy::reverse_iterator; | ||||||||||||
1728 | using const_reverse_iterator = RecipeListTy::const_reverse_iterator; | ||||||||||||
1729 | |||||||||||||
1730 | //===--------------------------------------------------------------------===// | ||||||||||||
1731 | /// Recipe iterator methods | ||||||||||||
1732 | /// | ||||||||||||
1733 | inline iterator begin() { return Recipes.begin(); } | ||||||||||||
1734 | inline const_iterator begin() const { return Recipes.begin(); } | ||||||||||||
1735 | inline iterator end() { return Recipes.end(); } | ||||||||||||
1736 | inline const_iterator end() const { return Recipes.end(); } | ||||||||||||
1737 | |||||||||||||
1738 | inline reverse_iterator rbegin() { return Recipes.rbegin(); } | ||||||||||||
1739 | inline const_reverse_iterator rbegin() const { return Recipes.rbegin(); } | ||||||||||||
1740 | inline reverse_iterator rend() { return Recipes.rend(); } | ||||||||||||
1741 | inline const_reverse_iterator rend() const { return Recipes.rend(); } | ||||||||||||
1742 | |||||||||||||
1743 | inline size_t size() const { return Recipes.size(); } | ||||||||||||
1744 | inline bool empty() const { return Recipes.empty(); } | ||||||||||||
1745 | inline const VPRecipeBase &front() const { return Recipes.front(); } | ||||||||||||
1746 | inline VPRecipeBase &front() { return Recipes.front(); } | ||||||||||||
1747 | inline const VPRecipeBase &back() const { return Recipes.back(); } | ||||||||||||
1748 | inline VPRecipeBase &back() { return Recipes.back(); } | ||||||||||||
1749 | |||||||||||||
1750 | /// Returns a reference to the list of recipes. | ||||||||||||
1751 | RecipeListTy &getRecipeList() { return Recipes; } | ||||||||||||
1752 | |||||||||||||
1753 | /// Returns a pointer to a member of the recipe list. | ||||||||||||
1754 | static RecipeListTy VPBasicBlock::*getSublistAccess(VPRecipeBase *) { | ||||||||||||
1755 | return &VPBasicBlock::Recipes; | ||||||||||||
1756 | } | ||||||||||||
1757 | |||||||||||||
1758 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1759 | static inline bool classof(const VPBlockBase *V) { | ||||||||||||
1760 | return V->getVPBlockID() == VPBlockBase::VPBasicBlockSC; | ||||||||||||
1761 | } | ||||||||||||
1762 | |||||||||||||
1763 | void insert(VPRecipeBase *Recipe, iterator InsertPt) { | ||||||||||||
1764 | 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", 1764, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
1765 | 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", 1765, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
1766 | Recipe->Parent = this; | ||||||||||||
1767 | Recipes.insert(InsertPt, Recipe); | ||||||||||||
1768 | } | ||||||||||||
1769 | |||||||||||||
1770 | /// Augment the existing recipes of a VPBasicBlock with an additional | ||||||||||||
1771 | /// \p Recipe as the last recipe. | ||||||||||||
1772 | void appendRecipe(VPRecipeBase *Recipe) { insert(Recipe, end()); } | ||||||||||||
1773 | |||||||||||||
1774 | /// The method which generates the output IR instructions that correspond to | ||||||||||||
1775 | /// this VPBasicBlock, thereby "executing" the VPlan. | ||||||||||||
1776 | void execute(struct VPTransformState *State) override; | ||||||||||||
1777 | |||||||||||||
1778 | /// Return the position of the first non-phi node recipe in the block. | ||||||||||||
1779 | iterator getFirstNonPhi(); | ||||||||||||
1780 | |||||||||||||
1781 | /// Returns an iterator range over the PHI-like recipes in the block. | ||||||||||||
1782 | iterator_range<iterator> phis() { | ||||||||||||
1783 | return make_range(begin(), getFirstNonPhi()); | ||||||||||||
1784 | } | ||||||||||||
1785 | |||||||||||||
1786 | void dropAllReferences(VPValue *NewValue) override; | ||||||||||||
1787 | |||||||||||||
1788 | /// Split current block at \p SplitAt by inserting a new block between the | ||||||||||||
1789 | /// current block and its successors and moving all recipes starting at | ||||||||||||
1790 | /// SplitAt to the new block. Returns the new block. | ||||||||||||
1791 | VPBasicBlock *splitAt(iterator SplitAt); | ||||||||||||
1792 | |||||||||||||
1793 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
1794 | /// Print this VPBsicBlock to \p O, prefixing all lines with \p Indent. \p | ||||||||||||
1795 | /// SlotTracker is used to print unnamed VPValue's using consequtive numbers. | ||||||||||||
1796 | /// | ||||||||||||
1797 | /// Note that the numbering is applied to the whole VPlan, so printing | ||||||||||||
1798 | /// individual blocks is consistent with the whole VPlan printing. | ||||||||||||
1799 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
1800 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
1801 | using VPBlockBase::print; // Get the print(raw_stream &O) version. | ||||||||||||
1802 | #endif | ||||||||||||
1803 | |||||||||||||
1804 | private: | ||||||||||||
1805 | /// Create an IR BasicBlock to hold the output instructions generated by this | ||||||||||||
1806 | /// VPBasicBlock, and return it. Update the CFGState accordingly. | ||||||||||||
1807 | BasicBlock *createEmptyBasicBlock(VPTransformState::CFGState &CFG); | ||||||||||||
1808 | }; | ||||||||||||
1809 | |||||||||||||
1810 | /// VPRegionBlock represents a collection of VPBasicBlocks and VPRegionBlocks | ||||||||||||
1811 | /// which form a Single-Entry-Single-Exit subgraph of the output IR CFG. | ||||||||||||
1812 | /// A VPRegionBlock may indicate that its contents are to be replicated several | ||||||||||||
1813 | /// times. This is designed to support predicated scalarization, in which a | ||||||||||||
1814 | /// scalar if-then code structure needs to be generated VF * UF times. Having | ||||||||||||
1815 | /// this replication indicator helps to keep a single model for multiple | ||||||||||||
1816 | /// candidate VF's. The actual replication takes place only once the desired VF | ||||||||||||
1817 | /// and UF have been determined. | ||||||||||||
1818 | class VPRegionBlock : public VPBlockBase { | ||||||||||||
1819 | /// Hold the Single Entry of the SESE region modelled by the VPRegionBlock. | ||||||||||||
1820 | VPBlockBase *Entry; | ||||||||||||
1821 | |||||||||||||
1822 | /// Hold the Single Exit of the SESE region modelled by the VPRegionBlock. | ||||||||||||
1823 | VPBlockBase *Exit; | ||||||||||||
1824 | |||||||||||||
1825 | /// An indicator whether this region is to generate multiple replicated | ||||||||||||
1826 | /// instances of output IR corresponding to its VPBlockBases. | ||||||||||||
1827 | bool IsReplicator; | ||||||||||||
1828 | |||||||||||||
1829 | public: | ||||||||||||
1830 | VPRegionBlock(VPBlockBase *Entry, VPBlockBase *Exit, | ||||||||||||
1831 | const std::string &Name = "", bool IsReplicator = false) | ||||||||||||
1832 | : VPBlockBase(VPRegionBlockSC, Name), Entry(Entry), Exit(Exit), | ||||||||||||
1833 | IsReplicator(IsReplicator) { | ||||||||||||
1834 | 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", 1834, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
1835 | assert(Exit->getSuccessors().empty() && "Exit block has successors.")(static_cast <bool> (Exit->getSuccessors().empty() && "Exit block has successors.") ? void (0) : __assert_fail ("Exit->getSuccessors().empty() && \"Exit block has successors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1835, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
1836 | Entry->setParent(this); | ||||||||||||
1837 | Exit->setParent(this); | ||||||||||||
1838 | } | ||||||||||||
1839 | VPRegionBlock(const std::string &Name = "", bool IsReplicator = false) | ||||||||||||
1840 | : VPBlockBase(VPRegionBlockSC, Name), Entry(nullptr), Exit(nullptr), | ||||||||||||
1841 | IsReplicator(IsReplicator) {} | ||||||||||||
1842 | |||||||||||||
1843 | ~VPRegionBlock() override { | ||||||||||||
1844 | if (Entry) { | ||||||||||||
1845 | VPValue DummyValue; | ||||||||||||
1846 | Entry->dropAllReferences(&DummyValue); | ||||||||||||
1847 | deleteCFG(Entry); | ||||||||||||
1848 | } | ||||||||||||
1849 | } | ||||||||||||
1850 | |||||||||||||
1851 | /// Method to support type inquiry through isa, cast, and dyn_cast. | ||||||||||||
1852 | static inline bool classof(const VPBlockBase *V) { | ||||||||||||
1853 | return V->getVPBlockID() == VPBlockBase::VPRegionBlockSC; | ||||||||||||
1854 | } | ||||||||||||
1855 | |||||||||||||
1856 | const VPBlockBase *getEntry() const { return Entry; } | ||||||||||||
1857 | VPBlockBase *getEntry() { return Entry; } | ||||||||||||
1858 | |||||||||||||
1859 | /// Set \p EntryBlock as the entry VPBlockBase of this VPRegionBlock. \p | ||||||||||||
1860 | /// EntryBlock must have no predecessors. | ||||||||||||
1861 | void setEntry(VPBlockBase *EntryBlock) { | ||||||||||||
1862 | 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", 1863, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
1863 | "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", 1863, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
1864 | Entry = EntryBlock; | ||||||||||||
1865 | EntryBlock->setParent(this); | ||||||||||||
1866 | } | ||||||||||||
1867 | |||||||||||||
1868 | // FIXME: DominatorTreeBase is doing 'A->getParent()->front()'. 'front' is a | ||||||||||||
1869 | // specific interface of llvm::Function, instead of using | ||||||||||||
1870 | // GraphTraints::getEntryNode. We should add a new template parameter to | ||||||||||||
1871 | // DominatorTreeBase representing the Graph type. | ||||||||||||
1872 | VPBlockBase &front() const { return *Entry; } | ||||||||||||
1873 | |||||||||||||
1874 | const VPBlockBase *getExit() const { return Exit; } | ||||||||||||
1875 | VPBlockBase *getExit() { return Exit; } | ||||||||||||
1876 | |||||||||||||
1877 | /// Set \p ExitBlock as the exit VPBlockBase of this VPRegionBlock. \p | ||||||||||||
1878 | /// ExitBlock must have no successors. | ||||||||||||
1879 | void setExit(VPBlockBase *ExitBlock) { | ||||||||||||
1880 | assert(ExitBlock->getSuccessors().empty() &&(static_cast <bool> (ExitBlock->getSuccessors().empty () && "Exit block cannot have successors.") ? void (0 ) : __assert_fail ("ExitBlock->getSuccessors().empty() && \"Exit block cannot have successors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1881, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
1881 | "Exit block cannot have successors.")(static_cast <bool> (ExitBlock->getSuccessors().empty () && "Exit block cannot have successors.") ? void (0 ) : __assert_fail ("ExitBlock->getSuccessors().empty() && \"Exit block cannot have successors.\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 1881, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
1882 | Exit = ExitBlock; | ||||||||||||
1883 | ExitBlock->setParent(this); | ||||||||||||
1884 | } | ||||||||||||
1885 | |||||||||||||
1886 | /// An indicator whether this region is to generate multiple replicated | ||||||||||||
1887 | /// instances of output IR corresponding to its VPBlockBases. | ||||||||||||
1888 | bool isReplicator() const { return IsReplicator; } | ||||||||||||
1889 | |||||||||||||
1890 | /// The method which generates the output IR instructions that correspond to | ||||||||||||
1891 | /// this VPRegionBlock, thereby "executing" the VPlan. | ||||||||||||
1892 | void execute(struct VPTransformState *State) override; | ||||||||||||
1893 | |||||||||||||
1894 | void dropAllReferences(VPValue *NewValue) override; | ||||||||||||
1895 | |||||||||||||
1896 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
1897 | /// Print this VPRegionBlock to \p O (recursively), prefixing all lines with | ||||||||||||
1898 | /// \p Indent. \p SlotTracker is used to print unnamed VPValue's using | ||||||||||||
1899 | /// consequtive numbers. | ||||||||||||
1900 | /// | ||||||||||||
1901 | /// Note that the numbering is applied to the whole VPlan, so printing | ||||||||||||
1902 | /// individual regions is consistent with the whole VPlan printing. | ||||||||||||
1903 | void print(raw_ostream &O, const Twine &Indent, | ||||||||||||
1904 | VPSlotTracker &SlotTracker) const override; | ||||||||||||
1905 | using VPBlockBase::print; // Get the print(raw_stream &O) version. | ||||||||||||
1906 | #endif | ||||||||||||
1907 | }; | ||||||||||||
1908 | |||||||||||||
1909 | //===----------------------------------------------------------------------===// | ||||||||||||
1910 | // GraphTraits specializations for VPlan Hierarchical Control-Flow Graphs // | ||||||||||||
1911 | //===----------------------------------------------------------------------===// | ||||||||||||
1912 | |||||||||||||
1913 | // The following set of template specializations implement GraphTraits to treat | ||||||||||||
1914 | // any VPBlockBase as a node in a graph of VPBlockBases. It's important to note | ||||||||||||
1915 | // that VPBlockBase traits don't recurse into VPRegioBlocks, i.e., if the | ||||||||||||
1916 | // VPBlockBase is a VPRegionBlock, this specialization provides access to its | ||||||||||||
1917 | // successors/predecessors but not to the blocks inside the region. | ||||||||||||
1918 | |||||||||||||
1919 | template <> struct GraphTraits<VPBlockBase *> { | ||||||||||||
1920 | using NodeRef = VPBlockBase *; | ||||||||||||
1921 | using ChildIteratorType = SmallVectorImpl<VPBlockBase *>::iterator; | ||||||||||||
1922 | |||||||||||||
1923 | static NodeRef getEntryNode(NodeRef N) { return N; } | ||||||||||||
1924 | |||||||||||||
1925 | static inline ChildIteratorType child_begin(NodeRef N) { | ||||||||||||
1926 | return N->getSuccessors().begin(); | ||||||||||||
1927 | } | ||||||||||||
1928 | |||||||||||||
1929 | static inline ChildIteratorType child_end(NodeRef N) { | ||||||||||||
1930 | return N->getSuccessors().end(); | ||||||||||||
1931 | } | ||||||||||||
1932 | }; | ||||||||||||
1933 | |||||||||||||
1934 | template <> struct GraphTraits<const VPBlockBase *> { | ||||||||||||
1935 | using NodeRef = const VPBlockBase *; | ||||||||||||
1936 | using ChildIteratorType = SmallVectorImpl<VPBlockBase *>::const_iterator; | ||||||||||||
1937 | |||||||||||||
1938 | static NodeRef getEntryNode(NodeRef N) { return N; } | ||||||||||||
1939 | |||||||||||||
1940 | static inline ChildIteratorType child_begin(NodeRef N) { | ||||||||||||
1941 | return N->getSuccessors().begin(); | ||||||||||||
1942 | } | ||||||||||||
1943 | |||||||||||||
1944 | static inline ChildIteratorType child_end(NodeRef N) { | ||||||||||||
1945 | return N->getSuccessors().end(); | ||||||||||||
1946 | } | ||||||||||||
1947 | }; | ||||||||||||
1948 | |||||||||||||
1949 | // Inverse order specialization for VPBasicBlocks. Predecessors are used instead | ||||||||||||
1950 | // of successors for the inverse traversal. | ||||||||||||
1951 | template <> struct GraphTraits<Inverse<VPBlockBase *>> { | ||||||||||||
1952 | using NodeRef = VPBlockBase *; | ||||||||||||
1953 | using ChildIteratorType = SmallVectorImpl<VPBlockBase *>::iterator; | ||||||||||||
1954 | |||||||||||||
1955 | static NodeRef getEntryNode(Inverse<NodeRef> B) { return B.Graph; } | ||||||||||||
1956 | |||||||||||||
1957 | static inline ChildIteratorType child_begin(NodeRef N) { | ||||||||||||
1958 | return N->getPredecessors().begin(); | ||||||||||||
1959 | } | ||||||||||||
1960 | |||||||||||||
1961 | static inline ChildIteratorType child_end(NodeRef N) { | ||||||||||||
1962 | return N->getPredecessors().end(); | ||||||||||||
1963 | } | ||||||||||||
1964 | }; | ||||||||||||
1965 | |||||||||||||
1966 | // The following set of template specializations implement GraphTraits to | ||||||||||||
1967 | // treat VPRegionBlock as a graph and recurse inside its nodes. It's important | ||||||||||||
1968 | // to note that the blocks inside the VPRegionBlock are treated as VPBlockBases | ||||||||||||
1969 | // (i.e., no dyn_cast is performed, VPBlockBases specialization is used), so | ||||||||||||
1970 | // there won't be automatic recursion into other VPBlockBases that turn to be | ||||||||||||
1971 | // VPRegionBlocks. | ||||||||||||
1972 | |||||||||||||
1973 | template <> | ||||||||||||
1974 | struct GraphTraits<VPRegionBlock *> : public GraphTraits<VPBlockBase *> { | ||||||||||||
1975 | using GraphRef = VPRegionBlock *; | ||||||||||||
1976 | using nodes_iterator = df_iterator<NodeRef>; | ||||||||||||
1977 | |||||||||||||
1978 | static NodeRef getEntryNode(GraphRef N) { return N->getEntry(); } | ||||||||||||
1979 | |||||||||||||
1980 | static nodes_iterator nodes_begin(GraphRef N) { | ||||||||||||
1981 | return nodes_iterator::begin(N->getEntry()); | ||||||||||||
1982 | } | ||||||||||||
1983 | |||||||||||||
1984 | static nodes_iterator nodes_end(GraphRef N) { | ||||||||||||
1985 | // df_iterator::end() returns an empty iterator so the node used doesn't | ||||||||||||
1986 | // matter. | ||||||||||||
1987 | return nodes_iterator::end(N); | ||||||||||||
1988 | } | ||||||||||||
1989 | }; | ||||||||||||
1990 | |||||||||||||
1991 | template <> | ||||||||||||
1992 | struct GraphTraits<const VPRegionBlock *> | ||||||||||||
1993 | : public GraphTraits<const VPBlockBase *> { | ||||||||||||
1994 | using GraphRef = const VPRegionBlock *; | ||||||||||||
1995 | using nodes_iterator = df_iterator<NodeRef>; | ||||||||||||
1996 | |||||||||||||
1997 | static NodeRef getEntryNode(GraphRef N) { return N->getEntry(); } | ||||||||||||
1998 | |||||||||||||
1999 | static nodes_iterator nodes_begin(GraphRef N) { | ||||||||||||
2000 | return nodes_iterator::begin(N->getEntry()); | ||||||||||||
2001 | } | ||||||||||||
2002 | |||||||||||||
2003 | static nodes_iterator nodes_end(GraphRef N) { | ||||||||||||
2004 | // df_iterator::end() returns an empty iterator so the node used doesn't | ||||||||||||
2005 | // matter. | ||||||||||||
2006 | return nodes_iterator::end(N); | ||||||||||||
2007 | } | ||||||||||||
2008 | }; | ||||||||||||
2009 | |||||||||||||
2010 | template <> | ||||||||||||
2011 | struct GraphTraits<Inverse<VPRegionBlock *>> | ||||||||||||
2012 | : public GraphTraits<Inverse<VPBlockBase *>> { | ||||||||||||
2013 | using GraphRef = VPRegionBlock *; | ||||||||||||
2014 | using nodes_iterator = df_iterator<NodeRef>; | ||||||||||||
2015 | |||||||||||||
2016 | static NodeRef getEntryNode(Inverse<GraphRef> N) { | ||||||||||||
2017 | return N.Graph->getExit(); | ||||||||||||
2018 | } | ||||||||||||
2019 | |||||||||||||
2020 | static nodes_iterator nodes_begin(GraphRef N) { | ||||||||||||
2021 | return nodes_iterator::begin(N->getExit()); | ||||||||||||
2022 | } | ||||||||||||
2023 | |||||||||||||
2024 | static nodes_iterator nodes_end(GraphRef N) { | ||||||||||||
2025 | // df_iterator::end() returns an empty iterator so the node used doesn't | ||||||||||||
2026 | // matter. | ||||||||||||
2027 | return nodes_iterator::end(N); | ||||||||||||
2028 | } | ||||||||||||
2029 | }; | ||||||||||||
2030 | |||||||||||||
2031 | /// Iterator to traverse all successors of a VPBlockBase node. This includes the | ||||||||||||
2032 | /// entry node of VPRegionBlocks. Exit blocks of a region implicitly have their | ||||||||||||
2033 | /// parent region's successors. This ensures all blocks in a region are visited | ||||||||||||
2034 | /// before any blocks in a successor region when doing a reverse post-order | ||||||||||||
2035 | // traversal of the graph. | ||||||||||||
2036 | template <typename BlockPtrTy> | ||||||||||||
2037 | class VPAllSuccessorsIterator | ||||||||||||
2038 | : public iterator_facade_base<VPAllSuccessorsIterator<BlockPtrTy>, | ||||||||||||
2039 | std::forward_iterator_tag, VPBlockBase> { | ||||||||||||
2040 | BlockPtrTy Block; | ||||||||||||
2041 | /// Index of the current successor. For VPBasicBlock nodes, this simply is the | ||||||||||||
2042 | /// index for the successor array. For VPRegionBlock, SuccessorIdx == 0 is | ||||||||||||
2043 | /// used for the region's entry block, and SuccessorIdx - 1 are the indices | ||||||||||||
2044 | /// for the successor array. | ||||||||||||
2045 | size_t SuccessorIdx; | ||||||||||||
2046 | |||||||||||||
2047 | static BlockPtrTy getBlockWithSuccs(BlockPtrTy Current) { | ||||||||||||
2048 | while (Current && Current->getNumSuccessors() == 0) | ||||||||||||
2049 | Current = Current->getParent(); | ||||||||||||
2050 | return Current; | ||||||||||||
2051 | } | ||||||||||||
2052 | |||||||||||||
2053 | /// Templated helper to dereference successor \p SuccIdx of \p Block. Used by | ||||||||||||
2054 | /// both the const and non-const operator* implementations. | ||||||||||||
2055 | template <typename T1> static T1 deref(T1 Block, unsigned SuccIdx) { | ||||||||||||
2056 | if (auto *R = dyn_cast<VPRegionBlock>(Block)) { | ||||||||||||
2057 | if (SuccIdx == 0) | ||||||||||||
2058 | return R->getEntry(); | ||||||||||||
2059 | SuccIdx--; | ||||||||||||
2060 | } | ||||||||||||
2061 | |||||||||||||
2062 | // For exit blocks, use the next parent region with successors. | ||||||||||||
2063 | return getBlockWithSuccs(Block)->getSuccessors()[SuccIdx]; | ||||||||||||
2064 | } | ||||||||||||
2065 | |||||||||||||
2066 | public: | ||||||||||||
2067 | VPAllSuccessorsIterator(BlockPtrTy Block, size_t Idx = 0) | ||||||||||||
2068 | : Block(Block), SuccessorIdx(Idx) {} | ||||||||||||
2069 | VPAllSuccessorsIterator(const VPAllSuccessorsIterator &Other) | ||||||||||||
2070 | : Block(Other.Block), SuccessorIdx(Other.SuccessorIdx) {} | ||||||||||||
2071 | |||||||||||||
2072 | VPAllSuccessorsIterator &operator=(const VPAllSuccessorsIterator &R) { | ||||||||||||
2073 | Block = R.Block; | ||||||||||||
2074 | SuccessorIdx = R.SuccessorIdx; | ||||||||||||
2075 | return *this; | ||||||||||||
2076 | } | ||||||||||||
2077 | |||||||||||||
2078 | static VPAllSuccessorsIterator end(BlockPtrTy Block) { | ||||||||||||
2079 | BlockPtrTy ParentWithSuccs = getBlockWithSuccs(Block); | ||||||||||||
2080 | unsigned NumSuccessors = ParentWithSuccs | ||||||||||||
2081 | ? ParentWithSuccs->getNumSuccessors() | ||||||||||||
2082 | : Block->getNumSuccessors(); | ||||||||||||
2083 | |||||||||||||
2084 | if (auto *R = dyn_cast<VPRegionBlock>(Block)) | ||||||||||||
2085 | return {R, NumSuccessors + 1}; | ||||||||||||
2086 | return {Block, NumSuccessors}; | ||||||||||||
2087 | } | ||||||||||||
2088 | |||||||||||||
2089 | bool operator==(const VPAllSuccessorsIterator &R) const { | ||||||||||||
2090 | return Block == R.Block && SuccessorIdx == R.SuccessorIdx; | ||||||||||||
2091 | } | ||||||||||||
2092 | |||||||||||||
2093 | const VPBlockBase *operator*() const { return deref(Block, SuccessorIdx); } | ||||||||||||
2094 | |||||||||||||
2095 | BlockPtrTy operator*() { return deref(Block, SuccessorIdx); } | ||||||||||||
2096 | |||||||||||||
2097 | VPAllSuccessorsIterator &operator++() { | ||||||||||||
2098 | SuccessorIdx++; | ||||||||||||
2099 | return *this; | ||||||||||||
2100 | } | ||||||||||||
2101 | |||||||||||||
2102 | VPAllSuccessorsIterator operator++(int X) { | ||||||||||||
2103 | VPAllSuccessorsIterator Orig = *this; | ||||||||||||
2104 | SuccessorIdx++; | ||||||||||||
2105 | return Orig; | ||||||||||||
2106 | } | ||||||||||||
2107 | }; | ||||||||||||
2108 | |||||||||||||
2109 | /// Helper for GraphTraits specialization that traverses through VPRegionBlocks. | ||||||||||||
2110 | template <typename BlockTy> class VPBlockRecursiveTraversalWrapper { | ||||||||||||
2111 | BlockTy Entry; | ||||||||||||
2112 | |||||||||||||
2113 | public: | ||||||||||||
2114 | VPBlockRecursiveTraversalWrapper(BlockTy Entry) : Entry(Entry) {} | ||||||||||||
2115 | BlockTy getEntry() { return Entry; } | ||||||||||||
2116 | }; | ||||||||||||
2117 | |||||||||||||
2118 | /// GraphTraits specialization to recursively traverse VPBlockBase nodes, | ||||||||||||
2119 | /// including traversing through VPRegionBlocks. Exit blocks of a region | ||||||||||||
2120 | /// implicitly have their parent region's successors. This ensures all blocks in | ||||||||||||
2121 | /// a region are visited before any blocks in a successor region when doing a | ||||||||||||
2122 | /// reverse post-order traversal of the graph. | ||||||||||||
2123 | template <> | ||||||||||||
2124 | struct GraphTraits<VPBlockRecursiveTraversalWrapper<VPBlockBase *>> { | ||||||||||||
2125 | using NodeRef = VPBlockBase *; | ||||||||||||
2126 | using ChildIteratorType = VPAllSuccessorsIterator<VPBlockBase *>; | ||||||||||||
2127 | |||||||||||||
2128 | static NodeRef | ||||||||||||
2129 | getEntryNode(VPBlockRecursiveTraversalWrapper<VPBlockBase *> N) { | ||||||||||||
2130 | return N.getEntry(); | ||||||||||||
2131 | } | ||||||||||||
2132 | |||||||||||||
2133 | static inline ChildIteratorType child_begin(NodeRef N) { | ||||||||||||
2134 | return ChildIteratorType(N); | ||||||||||||
2135 | } | ||||||||||||
2136 | |||||||||||||
2137 | static inline ChildIteratorType child_end(NodeRef N) { | ||||||||||||
2138 | return ChildIteratorType::end(N); | ||||||||||||
2139 | } | ||||||||||||
2140 | }; | ||||||||||||
2141 | |||||||||||||
2142 | template <> | ||||||||||||
2143 | struct GraphTraits<VPBlockRecursiveTraversalWrapper<const VPBlockBase *>> { | ||||||||||||
2144 | using NodeRef = const VPBlockBase *; | ||||||||||||
2145 | using ChildIteratorType = VPAllSuccessorsIterator<const VPBlockBase *>; | ||||||||||||
2146 | |||||||||||||
2147 | static NodeRef | ||||||||||||
2148 | getEntryNode(VPBlockRecursiveTraversalWrapper<const VPBlockBase *> N) { | ||||||||||||
2149 | return N.getEntry(); | ||||||||||||
2150 | } | ||||||||||||
2151 | |||||||||||||
2152 | static inline ChildIteratorType child_begin(NodeRef N) { | ||||||||||||
2153 | return ChildIteratorType(N); | ||||||||||||
2154 | } | ||||||||||||
2155 | |||||||||||||
2156 | static inline ChildIteratorType child_end(NodeRef N) { | ||||||||||||
2157 | return ChildIteratorType::end(N); | ||||||||||||
2158 | } | ||||||||||||
2159 | }; | ||||||||||||
2160 | |||||||||||||
2161 | /// VPlan models a candidate for vectorization, encoding various decisions take | ||||||||||||
2162 | /// to produce efficient output IR, including which branches, basic-blocks and | ||||||||||||
2163 | /// output IR instructions to generate, and their cost. VPlan holds a | ||||||||||||
2164 | /// Hierarchical-CFG of VPBasicBlocks and VPRegionBlocks rooted at an Entry | ||||||||||||
2165 | /// VPBlock. | ||||||||||||
2166 | class VPlan { | ||||||||||||
2167 | friend class VPlanPrinter; | ||||||||||||
2168 | friend class VPSlotTracker; | ||||||||||||
2169 | |||||||||||||
2170 | /// Hold the single entry to the Hierarchical CFG of the VPlan. | ||||||||||||
2171 | VPBlockBase *Entry; | ||||||||||||
2172 | |||||||||||||
2173 | /// Holds the VFs applicable to this VPlan. | ||||||||||||
2174 | SmallSetVector<ElementCount, 2> VFs; | ||||||||||||
2175 | |||||||||||||
2176 | /// Holds the name of the VPlan, for printing. | ||||||||||||
2177 | std::string Name; | ||||||||||||
2178 | |||||||||||||
2179 | /// Holds all the external definitions created for this VPlan. | ||||||||||||
2180 | // TODO: Introduce a specific representation for external definitions in | ||||||||||||
2181 | // VPlan. External definitions must be immutable and hold a pointer to its | ||||||||||||
2182 | // underlying IR that will be used to implement its structural comparison | ||||||||||||
2183 | // (operators '==' and '<'). | ||||||||||||
2184 | SetVector<VPValue *> VPExternalDefs; | ||||||||||||
2185 | |||||||||||||
2186 | /// Represents the trip count of the original loop, for folding | ||||||||||||
2187 | /// the tail. | ||||||||||||
2188 | VPValue *TripCount = nullptr; | ||||||||||||
2189 | |||||||||||||
2190 | /// Represents the backedge taken count of the original loop, for folding | ||||||||||||
2191 | /// the tail. It equals TripCount - 1. | ||||||||||||
2192 | VPValue *BackedgeTakenCount = nullptr; | ||||||||||||
2193 | |||||||||||||
2194 | /// Represents the vector trip count. | ||||||||||||
2195 | VPValue VectorTripCount; | ||||||||||||
2196 | |||||||||||||
2197 | /// Holds a mapping between Values and their corresponding VPValue inside | ||||||||||||
2198 | /// VPlan. | ||||||||||||
2199 | Value2VPValueTy Value2VPValue; | ||||||||||||
2200 | |||||||||||||
2201 | /// Contains all VPValues that been allocated by addVPValue directly and need | ||||||||||||
2202 | /// to be free when the plan's destructor is called. | ||||||||||||
2203 | SmallVector<VPValue *, 16> VPValuesToFree; | ||||||||||||
2204 | |||||||||||||
2205 | /// Holds the VPLoopInfo analysis for this VPlan. | ||||||||||||
2206 | VPLoopInfo VPLInfo; | ||||||||||||
2207 | |||||||||||||
2208 | /// Indicates whether it is safe use the Value2VPValue mapping or if the | ||||||||||||
2209 | /// mapping cannot be used any longer, because it is stale. | ||||||||||||
2210 | bool Value2VPValueEnabled = true; | ||||||||||||
2211 | |||||||||||||
2212 | public: | ||||||||||||
2213 | VPlan(VPBlockBase *Entry = nullptr) : Entry(Entry) { | ||||||||||||
2214 | if (Entry) | ||||||||||||
2215 | Entry->setPlan(this); | ||||||||||||
2216 | } | ||||||||||||
2217 | |||||||||||||
2218 | ~VPlan() { | ||||||||||||
2219 | if (Entry) { | ||||||||||||
2220 | VPValue DummyValue; | ||||||||||||
2221 | for (VPBlockBase *Block : depth_first(Entry)) | ||||||||||||
2222 | Block->dropAllReferences(&DummyValue); | ||||||||||||
2223 | |||||||||||||
2224 | VPBlockBase::deleteCFG(Entry); | ||||||||||||
2225 | } | ||||||||||||
2226 | for (VPValue *VPV : VPValuesToFree) | ||||||||||||
2227 | delete VPV; | ||||||||||||
2228 | if (TripCount) | ||||||||||||
2229 | delete TripCount; | ||||||||||||
2230 | if (BackedgeTakenCount) | ||||||||||||
2231 | delete BackedgeTakenCount; | ||||||||||||
2232 | for (VPValue *Def : VPExternalDefs) | ||||||||||||
2233 | delete Def; | ||||||||||||
2234 | } | ||||||||||||
2235 | |||||||||||||
2236 | /// Prepare the plan for execution, setting up the required live-in values. | ||||||||||||
2237 | void prepareToExecute(Value *TripCount, Value *VectorTripCount, | ||||||||||||
2238 | Value *CanonicalIVStartValue, VPTransformState &State); | ||||||||||||
2239 | |||||||||||||
2240 | /// Generate the IR code for this VPlan. | ||||||||||||
2241 | void execute(struct VPTransformState *State); | ||||||||||||
2242 | |||||||||||||
2243 | VPBlockBase *getEntry() { return Entry; } | ||||||||||||
2244 | const VPBlockBase *getEntry() const { return Entry; } | ||||||||||||
2245 | |||||||||||||
2246 | VPBlockBase *setEntry(VPBlockBase *Block) { | ||||||||||||
2247 | Entry = Block; | ||||||||||||
2248 | Block->setPlan(this); | ||||||||||||
2249 | return Entry; | ||||||||||||
2250 | } | ||||||||||||
2251 | |||||||||||||
2252 | /// The trip count of the original loop. | ||||||||||||
2253 | VPValue *getOrCreateTripCount() { | ||||||||||||
2254 | if (!TripCount) | ||||||||||||
2255 | TripCount = new VPValue(); | ||||||||||||
2256 | return TripCount; | ||||||||||||
2257 | } | ||||||||||||
2258 | |||||||||||||
2259 | /// The backedge taken count of the original loop. | ||||||||||||
2260 | VPValue *getOrCreateBackedgeTakenCount() { | ||||||||||||
2261 | if (!BackedgeTakenCount) | ||||||||||||
2262 | BackedgeTakenCount = new VPValue(); | ||||||||||||
2263 | return BackedgeTakenCount; | ||||||||||||
2264 | } | ||||||||||||
2265 | |||||||||||||
2266 | /// The vector trip count. | ||||||||||||
2267 | VPValue &getVectorTripCount() { return VectorTripCount; } | ||||||||||||
2268 | |||||||||||||
2269 | /// Mark the plan to indicate that using Value2VPValue is not safe any | ||||||||||||
2270 | /// longer, because it may be stale. | ||||||||||||
2271 | void disableValue2VPValue() { Value2VPValueEnabled = false; } | ||||||||||||
2272 | |||||||||||||
2273 | void addVF(ElementCount VF) { VFs.insert(VF); } | ||||||||||||
2274 | |||||||||||||
2275 | bool hasVF(ElementCount VF) { return VFs.count(VF); } | ||||||||||||
2276 | |||||||||||||
2277 | const std::string &getName() const { return Name; } | ||||||||||||
2278 | |||||||||||||
2279 | void setName(const Twine &newName) { Name = newName.str(); } | ||||||||||||
2280 | |||||||||||||
2281 | /// Add \p VPVal to the pool of external definitions if it's not already | ||||||||||||
2282 | /// in the pool. | ||||||||||||
2283 | void addExternalDef(VPValue *VPVal) { VPExternalDefs.insert(VPVal); } | ||||||||||||
2284 | |||||||||||||
2285 | void addVPValue(Value *V) { | ||||||||||||
2286 | 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", 2287, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
2287 | "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", 2287, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2288 | 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", 2288, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2289 | 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", 2289, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2290 | VPValue *VPV = new VPValue(V); | ||||||||||||
2291 | Value2VPValue[V] = VPV; | ||||||||||||
2292 | VPValuesToFree.push_back(VPV); | ||||||||||||
2293 | } | ||||||||||||
2294 | |||||||||||||
2295 | void addVPValue(Value *V, VPValue *VPV) { | ||||||||||||
2296 | 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", 2296, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2297 | 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", 2297, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2298 | 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", 2298, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2299 | Value2VPValue[V] = VPV; | ||||||||||||
2300 | } | ||||||||||||
2301 | |||||||||||||
2302 | /// Returns the VPValue for \p V. \p OverrideAllowed can be used to disable | ||||||||||||
2303 | /// checking whether it is safe to query VPValues using IR Values. | ||||||||||||
2304 | VPValue *getVPValue(Value *V, bool OverrideAllowed = false) { | ||||||||||||
2305 | 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", 2306, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
2306 | "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", 2306, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2307 | 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", 2307, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2308 | 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", 2308, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2309 | return Value2VPValue[V]; | ||||||||||||
2310 | } | ||||||||||||
2311 | |||||||||||||
2312 | /// Gets the VPValue or adds a new one (if none exists yet) for \p V. \p | ||||||||||||
2313 | /// OverrideAllowed can be used to disable checking whether it is safe to | ||||||||||||
2314 | /// query VPValues using IR Values. | ||||||||||||
2315 | VPValue *getOrAddVPValue(Value *V, bool OverrideAllowed = false) { | ||||||||||||
2316 | 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", 2317, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
2317 | "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", 2317, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2318 | 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", 2318, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2319 | if (!Value2VPValue.count(V)) | ||||||||||||
2320 | addVPValue(V); | ||||||||||||
2321 | return getVPValue(V); | ||||||||||||
2322 | } | ||||||||||||
2323 | |||||||||||||
2324 | void removeVPValueFor(Value *V) { | ||||||||||||
2325 | 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", 2326, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
2326 | "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", 2326, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2327 | Value2VPValue.erase(V); | ||||||||||||
2328 | } | ||||||||||||
2329 | |||||||||||||
2330 | /// Return the VPLoopInfo analysis for this VPlan. | ||||||||||||
2331 | VPLoopInfo &getVPLoopInfo() { return VPLInfo; } | ||||||||||||
2332 | const VPLoopInfo &getVPLoopInfo() const { return VPLInfo; } | ||||||||||||
2333 | |||||||||||||
2334 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
2335 | /// Print this VPlan to \p O. | ||||||||||||
2336 | void print(raw_ostream &O) const; | ||||||||||||
2337 | |||||||||||||
2338 | /// Print this VPlan in DOT format to \p O. | ||||||||||||
2339 | void printDOT(raw_ostream &O) const; | ||||||||||||
2340 | |||||||||||||
2341 | /// Dump the plan to stderr (for debugging). | ||||||||||||
2342 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const; | ||||||||||||
2343 | #endif | ||||||||||||
2344 | |||||||||||||
2345 | /// Returns a range mapping the values the range \p Operands to their | ||||||||||||
2346 | /// corresponding VPValues. | ||||||||||||
2347 | iterator_range<mapped_iterator<Use *, std::function<VPValue *(Value *)>>> | ||||||||||||
2348 | mapToVPValues(User::op_range Operands) { | ||||||||||||
2349 | std::function<VPValue *(Value *)> Fn = [this](Value *Op) { | ||||||||||||
2350 | return getOrAddVPValue(Op); | ||||||||||||
2351 | }; | ||||||||||||
2352 | return map_range(Operands, Fn); | ||||||||||||
2353 | } | ||||||||||||
2354 | |||||||||||||
2355 | /// Returns true if \p VPV is uniform after vectorization. | ||||||||||||
2356 | bool isUniformAfterVectorization(VPValue *VPV) const { | ||||||||||||
2357 | auto RepR = dyn_cast_or_null<VPReplicateRecipe>(VPV->getDef()); | ||||||||||||
2358 | return !VPV->getDef() || (RepR && RepR->isUniform()); | ||||||||||||
2359 | } | ||||||||||||
2360 | |||||||||||||
2361 | /// Returns the VPRegionBlock of the vector loop. | ||||||||||||
2362 | VPRegionBlock *getVectorLoopRegion() { | ||||||||||||
2363 | return cast<VPRegionBlock>(getEntry()); | ||||||||||||
2364 | } | ||||||||||||
2365 | |||||||||||||
2366 | /// Returns the canonical induction recipe of the vector loop. | ||||||||||||
2367 | VPCanonicalIVPHIRecipe *getCanonicalIV() { | ||||||||||||
2368 | VPBasicBlock *EntryVPBB = getVectorLoopRegion()->getEntryBasicBlock(); | ||||||||||||
2369 | if (EntryVPBB->empty()) { | ||||||||||||
2370 | // VPlan native path. | ||||||||||||
2371 | EntryVPBB = cast<VPBasicBlock>(EntryVPBB->getSingleSuccessor()); | ||||||||||||
2372 | } | ||||||||||||
2373 | return cast<VPCanonicalIVPHIRecipe>(&*EntryVPBB->begin()); | ||||||||||||
2374 | } | ||||||||||||
2375 | |||||||||||||
2376 | private: | ||||||||||||
2377 | /// Add to the given dominator tree the header block and every new basic block | ||||||||||||
2378 | /// that was created between it and the latch block, inclusive. | ||||||||||||
2379 | static void updateDominatorTree(DominatorTree *DT, BasicBlock *LoopLatchBB, | ||||||||||||
2380 | BasicBlock *LoopPreHeaderBB, | ||||||||||||
2381 | BasicBlock *LoopExitBB); | ||||||||||||
2382 | }; | ||||||||||||
2383 | |||||||||||||
2384 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
2385 | /// VPlanPrinter prints a given VPlan to a given output stream. The printing is | ||||||||||||
2386 | /// indented and follows the dot format. | ||||||||||||
2387 | class VPlanPrinter { | ||||||||||||
2388 | raw_ostream &OS; | ||||||||||||
2389 | const VPlan &Plan; | ||||||||||||
2390 | unsigned Depth = 0; | ||||||||||||
2391 | unsigned TabWidth = 2; | ||||||||||||
2392 | std::string Indent; | ||||||||||||
2393 | unsigned BID = 0; | ||||||||||||
2394 | SmallDenseMap<const VPBlockBase *, unsigned> BlockID; | ||||||||||||
2395 | |||||||||||||
2396 | VPSlotTracker SlotTracker; | ||||||||||||
2397 | |||||||||||||
2398 | /// Handle indentation. | ||||||||||||
2399 | void bumpIndent(int b) { Indent = std::string((Depth += b) * TabWidth, ' '); } | ||||||||||||
2400 | |||||||||||||
2401 | /// Print a given \p Block of the Plan. | ||||||||||||
2402 | void dumpBlock(const VPBlockBase *Block); | ||||||||||||
2403 | |||||||||||||
2404 | /// Print the information related to the CFG edges going out of a given | ||||||||||||
2405 | /// \p Block, followed by printing the successor blocks themselves. | ||||||||||||
2406 | void dumpEdges(const VPBlockBase *Block); | ||||||||||||
2407 | |||||||||||||
2408 | /// Print a given \p BasicBlock, including its VPRecipes, followed by printing | ||||||||||||
2409 | /// its successor blocks. | ||||||||||||
2410 | void dumpBasicBlock(const VPBasicBlock *BasicBlock); | ||||||||||||
2411 | |||||||||||||
2412 | /// Print a given \p Region of the Plan. | ||||||||||||
2413 | void dumpRegion(const VPRegionBlock *Region); | ||||||||||||
2414 | |||||||||||||
2415 | unsigned getOrCreateBID(const VPBlockBase *Block) { | ||||||||||||
2416 | return BlockID.count(Block) ? BlockID[Block] : BlockID[Block] = BID++; | ||||||||||||
2417 | } | ||||||||||||
2418 | |||||||||||||
2419 | Twine getOrCreateName(const VPBlockBase *Block); | ||||||||||||
2420 | |||||||||||||
2421 | Twine getUID(const VPBlockBase *Block); | ||||||||||||
2422 | |||||||||||||
2423 | /// Print the information related to a CFG edge between two VPBlockBases. | ||||||||||||
2424 | void drawEdge(const VPBlockBase *From, const VPBlockBase *To, bool Hidden, | ||||||||||||
2425 | const Twine &Label); | ||||||||||||
2426 | |||||||||||||
2427 | public: | ||||||||||||
2428 | VPlanPrinter(raw_ostream &O, const VPlan &P) | ||||||||||||
2429 | : OS(O), Plan(P), SlotTracker(&P) {} | ||||||||||||
2430 | |||||||||||||
2431 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump(); | ||||||||||||
2432 | }; | ||||||||||||
2433 | |||||||||||||
2434 | struct VPlanIngredient { | ||||||||||||
2435 | const Value *V; | ||||||||||||
2436 | |||||||||||||
2437 | VPlanIngredient(const Value *V) : V(V) {} | ||||||||||||
2438 | |||||||||||||
2439 | void print(raw_ostream &O) const; | ||||||||||||
2440 | }; | ||||||||||||
2441 | |||||||||||||
2442 | inline raw_ostream &operator<<(raw_ostream &OS, const VPlanIngredient &I) { | ||||||||||||
2443 | I.print(OS); | ||||||||||||
2444 | return OS; | ||||||||||||
2445 | } | ||||||||||||
2446 | |||||||||||||
2447 | inline raw_ostream &operator<<(raw_ostream &OS, const VPlan &Plan) { | ||||||||||||
2448 | Plan.print(OS); | ||||||||||||
2449 | return OS; | ||||||||||||
2450 | } | ||||||||||||
2451 | #endif | ||||||||||||
2452 | |||||||||||||
2453 | //===----------------------------------------------------------------------===// | ||||||||||||
2454 | // VPlan Utilities | ||||||||||||
2455 | //===----------------------------------------------------------------------===// | ||||||||||||
2456 | |||||||||||||
2457 | /// Class that provides utilities for VPBlockBases in VPlan. | ||||||||||||
2458 | class VPBlockUtils { | ||||||||||||
2459 | public: | ||||||||||||
2460 | VPBlockUtils() = delete; | ||||||||||||
2461 | |||||||||||||
2462 | /// Insert disconnected VPBlockBase \p NewBlock after \p BlockPtr. Add \p | ||||||||||||
2463 | /// NewBlock as successor of \p BlockPtr and \p BlockPtr as predecessor of \p | ||||||||||||
2464 | /// NewBlock, and propagate \p BlockPtr parent to \p NewBlock. \p BlockPtr's | ||||||||||||
2465 | /// successors are moved from \p BlockPtr to \p NewBlock and \p BlockPtr's | ||||||||||||
2466 | /// conditional bit is propagated to \p NewBlock. \p NewBlock must have | ||||||||||||
2467 | /// neither successors nor predecessors. | ||||||||||||
2468 | static void insertBlockAfter(VPBlockBase *NewBlock, VPBlockBase *BlockPtr) { | ||||||||||||
2469 | 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", 2471, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
2470 | 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", 2471, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
2471 | "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", 2471, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2472 | NewBlock->setParent(BlockPtr->getParent()); | ||||||||||||
2473 | SmallVector<VPBlockBase *> Succs(BlockPtr->successors()); | ||||||||||||
2474 | for (VPBlockBase *Succ : Succs) { | ||||||||||||
2475 | disconnectBlocks(BlockPtr, Succ); | ||||||||||||
2476 | connectBlocks(NewBlock, Succ); | ||||||||||||
2477 | } | ||||||||||||
2478 | NewBlock->setCondBit(BlockPtr->getCondBit()); | ||||||||||||
2479 | BlockPtr->setCondBit(nullptr); | ||||||||||||
2480 | connectBlocks(BlockPtr, NewBlock); | ||||||||||||
2481 | } | ||||||||||||
2482 | |||||||||||||
2483 | /// Insert disconnected VPBlockBases \p IfTrue and \p IfFalse after \p | ||||||||||||
2484 | /// BlockPtr. Add \p IfTrue and \p IfFalse as succesors of \p BlockPtr and \p | ||||||||||||
2485 | /// BlockPtr as predecessor of \p IfTrue and \p IfFalse. Propagate \p BlockPtr | ||||||||||||
2486 | /// parent to \p IfTrue and \p IfFalse. \p Condition is set as the successor | ||||||||||||
2487 | /// selector. \p BlockPtr must have no successors and \p IfTrue and \p IfFalse | ||||||||||||
2488 | /// must have neither successors nor predecessors. | ||||||||||||
2489 | static void insertTwoBlocksAfter(VPBlockBase *IfTrue, VPBlockBase *IfFalse, | ||||||||||||
2490 | VPValue *Condition, VPBlockBase *BlockPtr) { | ||||||||||||
2491 | 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", 2492, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
2492 | "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", 2492, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2493 | 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", 2494, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
2494 | "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", 2494, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2495 | BlockPtr->setTwoSuccessors(IfTrue, IfFalse, Condition); | ||||||||||||
2496 | IfTrue->setPredecessors({BlockPtr}); | ||||||||||||
2497 | IfFalse->setPredecessors({BlockPtr}); | ||||||||||||
2498 | IfTrue->setParent(BlockPtr->getParent()); | ||||||||||||
2499 | IfFalse->setParent(BlockPtr->getParent()); | ||||||||||||
2500 | } | ||||||||||||
2501 | |||||||||||||
2502 | /// Connect VPBlockBases \p From and \p To bi-directionally. Append \p To to | ||||||||||||
2503 | /// the successors of \p From and \p From to the predecessors of \p To. Both | ||||||||||||
2504 | /// VPBlockBases must have the same parent, which can be null. Both | ||||||||||||
2505 | /// VPBlockBases can be already connected to other VPBlockBases. | ||||||||||||
2506 | static void connectBlocks(VPBlockBase *From, VPBlockBase *To) { | ||||||||||||
2507 | 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", 2508, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
2508 | "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", 2508, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2509 | 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", 2510, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
2510 | "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", 2510, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2511 | From->appendSuccessor(To); | ||||||||||||
2512 | To->appendPredecessor(From); | ||||||||||||
2513 | } | ||||||||||||
2514 | |||||||||||||
2515 | /// Disconnect VPBlockBases \p From and \p To bi-directionally. Remove \p To | ||||||||||||
2516 | /// from the successors of \p From and \p From from the predecessors of \p To. | ||||||||||||
2517 | static void disconnectBlocks(VPBlockBase *From, VPBlockBase *To) { | ||||||||||||
2518 | 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", 2518, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2519 | From->removeSuccessor(To); | ||||||||||||
2520 | To->removePredecessor(From); | ||||||||||||
2521 | } | ||||||||||||
2522 | |||||||||||||
2523 | /// Try to merge \p Block into its single predecessor, if \p Block is a | ||||||||||||
2524 | /// VPBasicBlock and its predecessor has a single successor. Returns a pointer | ||||||||||||
2525 | /// to the predecessor \p Block was merged into or nullptr otherwise. | ||||||||||||
2526 | static VPBasicBlock *tryToMergeBlockIntoPredecessor(VPBlockBase *Block) { | ||||||||||||
2527 | auto *VPBB = dyn_cast<VPBasicBlock>(Block); | ||||||||||||
2528 | auto *PredVPBB = | ||||||||||||
2529 | dyn_cast_or_null<VPBasicBlock>(Block->getSinglePredecessor()); | ||||||||||||
2530 | if (!VPBB
| ||||||||||||
2531 | return nullptr; | ||||||||||||
2532 | |||||||||||||
2533 | for (VPRecipeBase &R : make_early_inc_range(*VPBB)) | ||||||||||||
2534 | R.moveBefore(*PredVPBB, PredVPBB->end()); | ||||||||||||
2535 | VPBlockUtils::disconnectBlocks(PredVPBB, VPBB); | ||||||||||||
2536 | auto *ParentRegion = cast<VPRegionBlock>(Block->getParent()); | ||||||||||||
2537 | if (ParentRegion->getExit() == Block) | ||||||||||||
2538 | ParentRegion->setExit(PredVPBB); | ||||||||||||
2539 | SmallVector<VPBlockBase *> Successors(Block->successors()); | ||||||||||||
2540 | for (auto *Succ : Successors) { | ||||||||||||
2541 | VPBlockUtils::disconnectBlocks(Block, Succ); | ||||||||||||
2542 | VPBlockUtils::connectBlocks(PredVPBB, Succ); | ||||||||||||
2543 | } | ||||||||||||
2544 | delete Block; | ||||||||||||
2545 | return PredVPBB; | ||||||||||||
2546 | } | ||||||||||||
2547 | |||||||||||||
2548 | /// Returns true if the edge \p FromBlock -> \p ToBlock is a back-edge. | ||||||||||||
2549 | static bool isBackEdge(const VPBlockBase *FromBlock, | ||||||||||||
2550 | const VPBlockBase *ToBlock, const VPLoopInfo *VPLI) { | ||||||||||||
2551 | assert(FromBlock->getParent() == ToBlock->getParent() &&(static_cast <bool> (FromBlock->getParent() == ToBlock ->getParent() && FromBlock->getParent() && "Must be in same region") ? void (0) : __assert_fail ("FromBlock->getParent() == ToBlock->getParent() && FromBlock->getParent() && \"Must be in same region\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2552, __extension__ __PRETTY_FUNCTION__)) | ||||||||||||
2552 | FromBlock->getParent() && "Must be in same region")(static_cast <bool> (FromBlock->getParent() == ToBlock ->getParent() && FromBlock->getParent() && "Must be in same region") ? void (0) : __assert_fail ("FromBlock->getParent() == ToBlock->getParent() && FromBlock->getParent() && \"Must be in same region\"" , "llvm/lib/Transforms/Vectorize/VPlan.h", 2552, __extension__ __PRETTY_FUNCTION__)); | ||||||||||||
2553 | const VPLoop *FromLoop = VPLI->getLoopFor(FromBlock); | ||||||||||||
2554 | const VPLoop *ToLoop = VPLI->getLoopFor(ToBlock); | ||||||||||||
2555 | if (!FromLoop || !ToLoop || FromLoop != ToLoop) | ||||||||||||
2556 | return false; | ||||||||||||
2557 | |||||||||||||
2558 | // A back-edge is a branch from the loop latch to its header. | ||||||||||||
2559 | return ToLoop->isLoopLatch(FromBlock) && ToBlock == ToLoop->getHeader(); | ||||||||||||
2560 | } | ||||||||||||
2561 | |||||||||||||
2562 | /// Returns true if \p Block is a loop latch | ||||||||||||
2563 | static bool blockIsLoopLatch(const VPBlockBase *Block, | ||||||||||||
2564 | const VPLoopInfo *VPLInfo) { | ||||||||||||
2565 | if (const VPLoop *ParentVPL = VPLInfo->getLoopFor(Block)) | ||||||||||||
2566 | return ParentVPL->isLoopLatch(Block); | ||||||||||||
2567 | |||||||||||||
2568 | return false; | ||||||||||||
2569 | } | ||||||||||||
2570 | |||||||||||||
2571 | /// Count and return the number of succesors of \p PredBlock excluding any | ||||||||||||
2572 | /// backedges. | ||||||||||||
2573 | static unsigned countSuccessorsNoBE(VPBlockBase *PredBlock, | ||||||||||||
2574 | VPLoopInfo *VPLI) { | ||||||||||||
2575 | unsigned Count = 0; | ||||||||||||
2576 | for (VPBlockBase *SuccBlock : PredBlock->getSuccessors()) { | ||||||||||||
2577 | if (!VPBlockUtils::isBackEdge(PredBlock, SuccBlock, VPLI)) | ||||||||||||
2578 | Count++; | ||||||||||||
2579 | } | ||||||||||||
2580 | return Count; | ||||||||||||
2581 | } | ||||||||||||
2582 | |||||||||||||
2583 | /// Return an iterator range over \p Range which only includes \p BlockTy | ||||||||||||
2584 | /// blocks. The accesses are casted to \p BlockTy. | ||||||||||||
2585 | template <typename BlockTy, typename T> | ||||||||||||
2586 | static auto blocksOnly(const T &Range) { | ||||||||||||
2587 | // Create BaseTy with correct const-ness based on BlockTy. | ||||||||||||
2588 | using BaseTy = | ||||||||||||
2589 | typename std::conditional<std::is_const<BlockTy>::value, | ||||||||||||
2590 | const VPBlockBase, VPBlockBase>::type; | ||||||||||||
2591 | |||||||||||||
2592 | // We need to first create an iterator range over (const) BlocktTy & instead | ||||||||||||
2593 | // of (const) BlockTy * for filter_range to work properly. | ||||||||||||
2594 | auto Mapped = | ||||||||||||
2595 | map_range(Range, [](BaseTy *Block) -> BaseTy & { return *Block; }); | ||||||||||||
2596 | auto Filter = make_filter_range( | ||||||||||||
2597 | Mapped, [](BaseTy &Block) { return isa<BlockTy>(&Block); }); | ||||||||||||
2598 | return map_range(Filter, [](BaseTy &Block) -> BlockTy * { | ||||||||||||
2599 | return cast<BlockTy>(&Block); | ||||||||||||
2600 | }); | ||||||||||||
2601 | } | ||||||||||||
2602 | }; | ||||||||||||
2603 | |||||||||||||
2604 | class VPInterleavedAccessInfo { | ||||||||||||
2605 | DenseMap<VPInstruction *, InterleaveGroup<VPInstruction> *> | ||||||||||||
2606 | InterleaveGroupMap; | ||||||||||||
2607 | |||||||||||||
2608 | /// Type for mapping of instruction based interleave groups to VPInstruction | ||||||||||||
2609 | /// interleave groups | ||||||||||||
2610 | using Old2NewTy = DenseMap<InterleaveGroup<Instruction> *, | ||||||||||||
2611 | InterleaveGroup<VPInstruction> *>; | ||||||||||||
2612 | |||||||||||||
2613 | /// Recursively \p Region and populate VPlan based interleave groups based on | ||||||||||||
2614 | /// \p IAI. | ||||||||||||
2615 | void visitRegion(VPRegionBlock *Region, Old2NewTy &Old2New, | ||||||||||||
2616 | InterleavedAccessInfo &IAI); | ||||||||||||
2617 | /// Recursively traverse \p Block and populate VPlan based interleave groups | ||||||||||||
2618 | /// based on \p IAI. | ||||||||||||
2619 | void visitBlock(VPBlockBase *Block, Old2NewTy &Old2New, | ||||||||||||
2620 | InterleavedAccessInfo &IAI); | ||||||||||||
2621 | |||||||||||||
2622 | public: | ||||||||||||
2623 | VPInterleavedAccessInfo(VPlan &Plan, InterleavedAccessInfo &IAI); | ||||||||||||
2624 | |||||||||||||
2625 | ~VPInterleavedAccessInfo() { | ||||||||||||
2626 | SmallPtrSet<InterleaveGroup<VPInstruction> *, 4> DelSet; | ||||||||||||
2627 | // Avoid releasing a pointer twice. | ||||||||||||
2628 | for (auto &I : InterleaveGroupMap) | ||||||||||||
2629 | DelSet.insert(I.second); | ||||||||||||
2630 | for (auto *Ptr : DelSet) | ||||||||||||
2631 | delete Ptr; | ||||||||||||
2632 | } | ||||||||||||
2633 | |||||||||||||
2634 | /// Get the interleave group that \p Instr belongs to. | ||||||||||||
2635 | /// | ||||||||||||
2636 | /// \returns nullptr if doesn't have such group. | ||||||||||||
2637 | InterleaveGroup<VPInstruction> * | ||||||||||||
2638 | getInterleaveGroup(VPInstruction *Instr) const { | ||||||||||||
2639 | return InterleaveGroupMap.lookup(Instr); | ||||||||||||
2640 | } | ||||||||||||
2641 | }; | ||||||||||||
2642 | |||||||||||||
2643 | /// Class that maps (parts of) an existing VPlan to trees of combined | ||||||||||||
2644 | /// VPInstructions. | ||||||||||||
2645 | class VPlanSlp { | ||||||||||||
2646 | enum class OpMode { Failed, Load, Opcode }; | ||||||||||||
2647 | |||||||||||||
2648 | /// A DenseMapInfo implementation for using SmallVector<VPValue *, 4> as | ||||||||||||
2649 | /// DenseMap keys. | ||||||||||||
2650 | struct BundleDenseMapInfo { | ||||||||||||
2651 | static SmallVector<VPValue *, 4> getEmptyKey() { | ||||||||||||
2652 | return {reinterpret_cast<VPValue *>(-1)}; | ||||||||||||
2653 | } | ||||||||||||
2654 | |||||||||||||
2655 | static SmallVector<VPValue *, 4> getTombstoneKey() { | ||||||||||||
2656 | return {reinterpret_cast<VPValue *>(-2)}; | ||||||||||||
2657 | } | ||||||||||||
2658 | |||||||||||||
2659 | static unsigned getHashValue(const SmallVector<VPValue *, 4> &V) { | ||||||||||||
2660 | return static_cast<unsigned>(hash_combine_range(V.begin(), V.end())); | ||||||||||||
2661 | } | ||||||||||||
2662 | |||||||||||||
2663 | static bool isEqual(const SmallVector<VPValue *, 4> &LHS, | ||||||||||||
2664 | const SmallVector<VPValue *, 4> &RHS) { | ||||||||||||
2665 | return LHS == RHS; | ||||||||||||
2666 | } | ||||||||||||
2667 | }; | ||||||||||||
2668 | |||||||||||||
2669 | /// Mapping of values in the original VPlan to a combined VPInstruction. | ||||||||||||
2670 | DenseMap<SmallVector<VPValue *, 4>, VPInstruction *, BundleDenseMapInfo> | ||||||||||||
2671 | BundleToCombined; | ||||||||||||
2672 | |||||||||||||
2673 | VPInterleavedAccessInfo &IAI; | ||||||||||||
2674 | |||||||||||||
2675 | /// Basic block to operate on. For now, only instructions in a single BB are | ||||||||||||
2676 | /// considered. | ||||||||||||
2677 | const VPBasicBlock &BB; | ||||||||||||
2678 | |||||||||||||
2679 | /// Indicates whether we managed to combine all visited instructions or not. | ||||||||||||
2680 | bool CompletelySLP = true; | ||||||||||||
2681 | |||||||||||||
2682 | /// Width of the widest combined bundle in bits. | ||||||||||||
2683 | unsigned WidestBundleBits = 0; | ||||||||||||
2684 | |||||||||||||
2685 | using MultiNodeOpTy = | ||||||||||||
2686 | typename std::pair<VPInstruction *, SmallVector<VPValue *, 4>>; | ||||||||||||
2687 | |||||||||||||
2688 | // Input operand bundles for the current multi node. Each multi node operand | ||||||||||||
2689 | // bundle contains values not matching the multi node's opcode. They will | ||||||||||||
2690 | // be reordered in reorderMultiNodeOps, once we completed building a | ||||||||||||
2691 | // multi node. | ||||||||||||
2692 | SmallVector<MultiNodeOpTy, 4> MultiNodeOps; | ||||||||||||
2693 | |||||||||||||
2694 | /// Indicates whether we are building a multi node currently. | ||||||||||||
2695 | bool MultiNodeActive = false; | ||||||||||||
2696 | |||||||||||||
2697 | /// Check if we can vectorize Operands together. | ||||||||||||
2698 | bool areVectorizable(ArrayRef<VPValue *> Operands) const; | ||||||||||||
2699 | |||||||||||||
2700 | /// Add combined instruction \p New for the bundle \p Operands. | ||||||||||||
2701 | void addCombined(ArrayRef<VPValue *> Operands, VPInstruction *New); | ||||||||||||
2702 | |||||||||||||
2703 | /// Indicate we hit a bundle we failed to combine. Returns nullptr for now. | ||||||||||||
2704 | VPInstruction *markFailed(); | ||||||||||||
2705 | |||||||||||||
2706 | /// Reorder operands in the multi node to maximize sequential memory access | ||||||||||||
2707 | /// and commutative operations. | ||||||||||||
2708 | SmallVector<MultiNodeOpTy, 4> reorderMultiNodeOps(); | ||||||||||||
2709 | |||||||||||||
2710 | /// Choose the best candidate to use for the lane after \p Last. The set of | ||||||||||||
2711 | /// candidates to choose from are values with an opcode matching \p Last's | ||||||||||||
2712 | /// or loads consecutive to \p Last. | ||||||||||||
2713 | std::pair<OpMode, VPValue *> getBest(OpMode Mode, VPValue *Last, | ||||||||||||
2714 | SmallPtrSetImpl<VPValue *> &Candidates, | ||||||||||||
2715 | VPInterleavedAccessInfo &IAI); | ||||||||||||
2716 | |||||||||||||
2717 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||||||||||
2718 | /// Print bundle \p Values to dbgs(). | ||||||||||||
2719 | void dumpBundle(ArrayRef<VPValue *> Values); | ||||||||||||
2720 | #endif | ||||||||||||
2721 | |||||||||||||
2722 | public: | ||||||||||||
2723 | VPlanSlp(VPInterleavedAccessInfo &IAI, VPBasicBlock &BB) : IAI(IAI), BB(BB) {} | ||||||||||||
2724 | |||||||||||||
2725 | ~VPlanSlp() = default; | ||||||||||||
2726 | |||||||||||||
2727 | /// Tries to build an SLP tree rooted at \p Operands and returns a | ||||||||||||
2728 | /// VPInstruction combining \p Operands, if they can be combined. | ||||||||||||
2729 | VPInstruction *buildGraph(ArrayRef<VPValue *> Operands); | ||||||||||||
2730 | |||||||||||||
2731 | /// Return the width of the widest combined bundle in bits. | ||||||||||||
2732 | unsigned getWidestBundleBits() const { return WidestBundleBits; } | ||||||||||||
2733 | |||||||||||||
2734 | /// Return true if all visited instruction can be combined. | ||||||||||||
2735 | bool isCompletelySLP() const { return CompletelySLP; } | ||||||||||||
2736 | }; | ||||||||||||
2737 | } // end namespace llvm | ||||||||||||
2738 | |||||||||||||
2739 | #endif // LLVM_TRANSFORMS_VECTORIZE_VPLAN_H |