| File: | llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp |
| Warning: | line 5758, column 30 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===- SLPVectorizer.cpp - A bottom up SLP 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 pass implements the Bottom Up SLP vectorizer. It detects consecutive | ||||
| 10 | // stores that can be put together into vector-stores. Next, it attempts to | ||||
| 11 | // construct vectorizable tree using the use-def chains. If a profitable tree | ||||
| 12 | // was found, the SLP vectorizer performs vectorization on the tree. | ||||
| 13 | // | ||||
| 14 | // The pass is inspired by the work described in the paper: | ||||
| 15 | // "Loop-Aware SLP in GCC" by Ira Rosen, Dorit Nuzman, Ayal Zaks. | ||||
| 16 | // | ||||
| 17 | //===----------------------------------------------------------------------===// | ||||
| 18 | |||||
| 19 | #include "llvm/Transforms/Vectorize/SLPVectorizer.h" | ||||
| 20 | #include "llvm/ADT/DenseMap.h" | ||||
| 21 | #include "llvm/ADT/DenseSet.h" | ||||
| 22 | #include "llvm/ADT/Optional.h" | ||||
| 23 | #include "llvm/ADT/PostOrderIterator.h" | ||||
| 24 | #include "llvm/ADT/STLExtras.h" | ||||
| 25 | #include "llvm/ADT/SetVector.h" | ||||
| 26 | #include "llvm/ADT/SmallBitVector.h" | ||||
| 27 | #include "llvm/ADT/SmallPtrSet.h" | ||||
| 28 | #include "llvm/ADT/SmallSet.h" | ||||
| 29 | #include "llvm/ADT/SmallString.h" | ||||
| 30 | #include "llvm/ADT/Statistic.h" | ||||
| 31 | #include "llvm/ADT/iterator.h" | ||||
| 32 | #include "llvm/ADT/iterator_range.h" | ||||
| 33 | #include "llvm/Analysis/AliasAnalysis.h" | ||||
| 34 | #include "llvm/Analysis/AssumptionCache.h" | ||||
| 35 | #include "llvm/Analysis/CodeMetrics.h" | ||||
| 36 | #include "llvm/Analysis/DemandedBits.h" | ||||
| 37 | #include "llvm/Analysis/GlobalsModRef.h" | ||||
| 38 | #include "llvm/Analysis/IVDescriptors.h" | ||||
| 39 | #include "llvm/Analysis/LoopAccessAnalysis.h" | ||||
| 40 | #include "llvm/Analysis/LoopInfo.h" | ||||
| 41 | #include "llvm/Analysis/MemoryLocation.h" | ||||
| 42 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" | ||||
| 43 | #include "llvm/Analysis/ScalarEvolution.h" | ||||
| 44 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" | ||||
| 45 | #include "llvm/Analysis/TargetLibraryInfo.h" | ||||
| 46 | #include "llvm/Analysis/TargetTransformInfo.h" | ||||
| 47 | #include "llvm/Analysis/ValueTracking.h" | ||||
| 48 | #include "llvm/Analysis/VectorUtils.h" | ||||
| 49 | #include "llvm/IR/Attributes.h" | ||||
| 50 | #include "llvm/IR/BasicBlock.h" | ||||
| 51 | #include "llvm/IR/Constant.h" | ||||
| 52 | #include "llvm/IR/Constants.h" | ||||
| 53 | #include "llvm/IR/DataLayout.h" | ||||
| 54 | #include "llvm/IR/DebugLoc.h" | ||||
| 55 | #include "llvm/IR/DerivedTypes.h" | ||||
| 56 | #include "llvm/IR/Dominators.h" | ||||
| 57 | #include "llvm/IR/Function.h" | ||||
| 58 | #include "llvm/IR/IRBuilder.h" | ||||
| 59 | #include "llvm/IR/InstrTypes.h" | ||||
| 60 | #include "llvm/IR/Instruction.h" | ||||
| 61 | #include "llvm/IR/Instructions.h" | ||||
| 62 | #include "llvm/IR/IntrinsicInst.h" | ||||
| 63 | #include "llvm/IR/Intrinsics.h" | ||||
| 64 | #include "llvm/IR/Module.h" | ||||
| 65 | #include "llvm/IR/NoFolder.h" | ||||
| 66 | #include "llvm/IR/Operator.h" | ||||
| 67 | #include "llvm/IR/PatternMatch.h" | ||||
| 68 | #include "llvm/IR/Type.h" | ||||
| 69 | #include "llvm/IR/Use.h" | ||||
| 70 | #include "llvm/IR/User.h" | ||||
| 71 | #include "llvm/IR/Value.h" | ||||
| 72 | #include "llvm/IR/ValueHandle.h" | ||||
| 73 | #include "llvm/IR/Verifier.h" | ||||
| 74 | #include "llvm/InitializePasses.h" | ||||
| 75 | #include "llvm/Pass.h" | ||||
| 76 | #include "llvm/Support/Casting.h" | ||||
| 77 | #include "llvm/Support/CommandLine.h" | ||||
| 78 | #include "llvm/Support/Compiler.h" | ||||
| 79 | #include "llvm/Support/DOTGraphTraits.h" | ||||
| 80 | #include "llvm/Support/Debug.h" | ||||
| 81 | #include "llvm/Support/ErrorHandling.h" | ||||
| 82 | #include "llvm/Support/GraphWriter.h" | ||||
| 83 | #include "llvm/Support/InstructionCost.h" | ||||
| 84 | #include "llvm/Support/KnownBits.h" | ||||
| 85 | #include "llvm/Support/MathExtras.h" | ||||
| 86 | #include "llvm/Support/raw_ostream.h" | ||||
| 87 | #include "llvm/Transforms/Utils/InjectTLIMappings.h" | ||||
| 88 | #include "llvm/Transforms/Utils/LoopUtils.h" | ||||
| 89 | #include "llvm/Transforms/Vectorize.h" | ||||
| 90 | #include <algorithm> | ||||
| 91 | #include <cassert> | ||||
| 92 | #include <cstdint> | ||||
| 93 | #include <iterator> | ||||
| 94 | #include <memory> | ||||
| 95 | #include <set> | ||||
| 96 | #include <string> | ||||
| 97 | #include <tuple> | ||||
| 98 | #include <utility> | ||||
| 99 | #include <vector> | ||||
| 100 | |||||
| 101 | using namespace llvm; | ||||
| 102 | using namespace llvm::PatternMatch; | ||||
| 103 | using namespace slpvectorizer; | ||||
| 104 | |||||
| 105 | #define SV_NAME"slp-vectorizer" "slp-vectorizer" | ||||
| 106 | #define DEBUG_TYPE"SLP" "SLP" | ||||
| 107 | |||||
| 108 | STATISTIC(NumVectorInstructions, "Number of vector instructions generated")static llvm::Statistic NumVectorInstructions = {"SLP", "NumVectorInstructions" , "Number of vector instructions generated"}; | ||||
| 109 | |||||
| 110 | cl::opt<bool> RunSLPVectorization("vectorize-slp", cl::init(true), cl::Hidden, | ||||
| 111 | cl::desc("Run the SLP vectorization passes")); | ||||
| 112 | |||||
| 113 | static cl::opt<int> | ||||
| 114 | SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden, | ||||
| 115 | cl::desc("Only vectorize if you gain more than this " | ||||
| 116 | "number ")); | ||||
| 117 | |||||
| 118 | static cl::opt<bool> | ||||
| 119 | ShouldVectorizeHor("slp-vectorize-hor", cl::init(true), cl::Hidden, | ||||
| 120 | cl::desc("Attempt to vectorize horizontal reductions")); | ||||
| 121 | |||||
| 122 | static cl::opt<bool> ShouldStartVectorizeHorAtStore( | ||||
| 123 | "slp-vectorize-hor-store", cl::init(false), cl::Hidden, | ||||
| 124 | cl::desc( | ||||
| 125 | "Attempt to vectorize horizontal reductions feeding into a store")); | ||||
| 126 | |||||
| 127 | static cl::opt<int> | ||||
| 128 | MaxVectorRegSizeOption("slp-max-reg-size", cl::init(128), cl::Hidden, | ||||
| 129 | cl::desc("Attempt to vectorize for this register size in bits")); | ||||
| 130 | |||||
| 131 | static cl::opt<unsigned> | ||||
| 132 | MaxVFOption("slp-max-vf", cl::init(0), cl::Hidden, | ||||
| 133 | cl::desc("Maximum SLP vectorization factor (0=unlimited)")); | ||||
| 134 | |||||
| 135 | static cl::opt<int> | ||||
| 136 | MaxStoreLookup("slp-max-store-lookup", cl::init(32), cl::Hidden, | ||||
| 137 | cl::desc("Maximum depth of the lookup for consecutive stores.")); | ||||
| 138 | |||||
| 139 | /// Limits the size of scheduling regions in a block. | ||||
| 140 | /// It avoid long compile times for _very_ large blocks where vector | ||||
| 141 | /// instructions are spread over a wide range. | ||||
| 142 | /// This limit is way higher than needed by real-world functions. | ||||
| 143 | static cl::opt<int> | ||||
| 144 | ScheduleRegionSizeBudget("slp-schedule-budget", cl::init(100000), cl::Hidden, | ||||
| 145 | cl::desc("Limit the size of the SLP scheduling region per block")); | ||||
| 146 | |||||
| 147 | static cl::opt<int> MinVectorRegSizeOption( | ||||
| 148 | "slp-min-reg-size", cl::init(128), cl::Hidden, | ||||
| 149 | cl::desc("Attempt to vectorize for this register size in bits")); | ||||
| 150 | |||||
| 151 | static cl::opt<unsigned> RecursionMaxDepth( | ||||
| 152 | "slp-recursion-max-depth", cl::init(12), cl::Hidden, | ||||
| 153 | cl::desc("Limit the recursion depth when building a vectorizable tree")); | ||||
| 154 | |||||
| 155 | static cl::opt<unsigned> MinTreeSize( | ||||
| 156 | "slp-min-tree-size", cl::init(3), cl::Hidden, | ||||
| 157 | cl::desc("Only vectorize small trees if they are fully vectorizable")); | ||||
| 158 | |||||
| 159 | // The maximum depth that the look-ahead score heuristic will explore. | ||||
| 160 | // The higher this value, the higher the compilation time overhead. | ||||
| 161 | static cl::opt<int> LookAheadMaxDepth( | ||||
| 162 | "slp-max-look-ahead-depth", cl::init(2), cl::Hidden, | ||||
| 163 | cl::desc("The maximum look-ahead depth for operand reordering scores")); | ||||
| 164 | |||||
| 165 | // The Look-ahead heuristic goes through the users of the bundle to calculate | ||||
| 166 | // the users cost in getExternalUsesCost(). To avoid compilation time increase | ||||
| 167 | // we limit the number of users visited to this value. | ||||
| 168 | static cl::opt<unsigned> LookAheadUsersBudget( | ||||
| 169 | "slp-look-ahead-users-budget", cl::init(2), cl::Hidden, | ||||
| 170 | cl::desc("The maximum number of users to visit while visiting the " | ||||
| 171 | "predecessors. This prevents compilation time increase.")); | ||||
| 172 | |||||
| 173 | static cl::opt<bool> | ||||
| 174 | ViewSLPTree("view-slp-tree", cl::Hidden, | ||||
| 175 | cl::desc("Display the SLP trees with Graphviz")); | ||||
| 176 | |||||
| 177 | // Limit the number of alias checks. The limit is chosen so that | ||||
| 178 | // it has no negative effect on the llvm benchmarks. | ||||
| 179 | static const unsigned AliasedCheckLimit = 10; | ||||
| 180 | |||||
| 181 | // Another limit for the alias checks: The maximum distance between load/store | ||||
| 182 | // instructions where alias checks are done. | ||||
| 183 | // This limit is useful for very large basic blocks. | ||||
| 184 | static const unsigned MaxMemDepDistance = 160; | ||||
| 185 | |||||
| 186 | /// If the ScheduleRegionSizeBudget is exhausted, we allow small scheduling | ||||
| 187 | /// regions to be handled. | ||||
| 188 | static const int MinScheduleRegionSize = 16; | ||||
| 189 | |||||
| 190 | /// Predicate for the element types that the SLP vectorizer supports. | ||||
| 191 | /// | ||||
| 192 | /// The most important thing to filter here are types which are invalid in LLVM | ||||
| 193 | /// vectors. We also filter target specific types which have absolutely no | ||||
| 194 | /// meaningful vectorization path such as x86_fp80 and ppc_f128. This just | ||||
| 195 | /// avoids spending time checking the cost model and realizing that they will | ||||
| 196 | /// be inevitably scalarized. | ||||
| 197 | static bool isValidElementType(Type *Ty) { | ||||
| 198 | return VectorType::isValidElementType(Ty) && !Ty->isX86_FP80Ty() && | ||||
| 199 | !Ty->isPPC_FP128Ty(); | ||||
| 200 | } | ||||
| 201 | |||||
| 202 | /// \returns true if all of the instructions in \p VL are in the same block or | ||||
| 203 | /// false otherwise. | ||||
| 204 | static bool allSameBlock(ArrayRef<Value *> VL) { | ||||
| 205 | Instruction *I0 = dyn_cast<Instruction>(VL[0]); | ||||
| 206 | if (!I0) | ||||
| 207 | return false; | ||||
| 208 | BasicBlock *BB = I0->getParent(); | ||||
| 209 | for (int I = 1, E = VL.size(); I < E; I++) { | ||||
| 210 | auto *II = dyn_cast<Instruction>(VL[I]); | ||||
| 211 | if (!II) | ||||
| 212 | return false; | ||||
| 213 | |||||
| 214 | if (BB != II->getParent()) | ||||
| 215 | return false; | ||||
| 216 | } | ||||
| 217 | return true; | ||||
| 218 | } | ||||
| 219 | |||||
| 220 | /// \returns True if all of the values in \p VL are constants (but not | ||||
| 221 | /// globals/constant expressions). | ||||
| 222 | static bool allConstant(ArrayRef<Value *> VL) { | ||||
| 223 | // Constant expressions and globals can't be vectorized like normal integer/FP | ||||
| 224 | // constants. | ||||
| 225 | for (Value *i : VL) | ||||
| 226 | if (!isa<Constant>(i) || isa<ConstantExpr>(i) || isa<GlobalValue>(i)) | ||||
| 227 | return false; | ||||
| 228 | return true; | ||||
| 229 | } | ||||
| 230 | |||||
| 231 | /// \returns True if all of the values in \p VL are identical. | ||||
| 232 | static bool isSplat(ArrayRef<Value *> VL) { | ||||
| 233 | for (unsigned i = 1, e = VL.size(); i < e; ++i) | ||||
| 234 | if (VL[i] != VL[0]) | ||||
| 235 | return false; | ||||
| 236 | return true; | ||||
| 237 | } | ||||
| 238 | |||||
| 239 | /// \returns True if \p I is commutative, handles CmpInst and BinaryOperator. | ||||
| 240 | static bool isCommutative(Instruction *I) { | ||||
| 241 | if (auto *Cmp = dyn_cast<CmpInst>(I)) | ||||
| 242 | return Cmp->isCommutative(); | ||||
| 243 | if (auto *BO = dyn_cast<BinaryOperator>(I)) | ||||
| 244 | return BO->isCommutative(); | ||||
| 245 | // TODO: This should check for generic Instruction::isCommutative(), but | ||||
| 246 | // we need to confirm that the caller code correctly handles Intrinsics | ||||
| 247 | // for example (does not have 2 operands). | ||||
| 248 | return false; | ||||
| 249 | } | ||||
| 250 | |||||
| 251 | /// Checks if the vector of instructions can be represented as a shuffle, like: | ||||
| 252 | /// %x0 = extractelement <4 x i8> %x, i32 0 | ||||
| 253 | /// %x3 = extractelement <4 x i8> %x, i32 3 | ||||
| 254 | /// %y1 = extractelement <4 x i8> %y, i32 1 | ||||
| 255 | /// %y2 = extractelement <4 x i8> %y, i32 2 | ||||
| 256 | /// %x0x0 = mul i8 %x0, %x0 | ||||
| 257 | /// %x3x3 = mul i8 %x3, %x3 | ||||
| 258 | /// %y1y1 = mul i8 %y1, %y1 | ||||
| 259 | /// %y2y2 = mul i8 %y2, %y2 | ||||
| 260 | /// %ins1 = insertelement <4 x i8> poison, i8 %x0x0, i32 0 | ||||
| 261 | /// %ins2 = insertelement <4 x i8> %ins1, i8 %x3x3, i32 1 | ||||
| 262 | /// %ins3 = insertelement <4 x i8> %ins2, i8 %y1y1, i32 2 | ||||
| 263 | /// %ins4 = insertelement <4 x i8> %ins3, i8 %y2y2, i32 3 | ||||
| 264 | /// ret <4 x i8> %ins4 | ||||
| 265 | /// can be transformed into: | ||||
| 266 | /// %1 = shufflevector <4 x i8> %x, <4 x i8> %y, <4 x i32> <i32 0, i32 3, i32 5, | ||||
| 267 | /// i32 6> | ||||
| 268 | /// %2 = mul <4 x i8> %1, %1 | ||||
| 269 | /// ret <4 x i8> %2 | ||||
| 270 | /// We convert this initially to something like: | ||||
| 271 | /// %x0 = extractelement <4 x i8> %x, i32 0 | ||||
| 272 | /// %x3 = extractelement <4 x i8> %x, i32 3 | ||||
| 273 | /// %y1 = extractelement <4 x i8> %y, i32 1 | ||||
| 274 | /// %y2 = extractelement <4 x i8> %y, i32 2 | ||||
| 275 | /// %1 = insertelement <4 x i8> poison, i8 %x0, i32 0 | ||||
| 276 | /// %2 = insertelement <4 x i8> %1, i8 %x3, i32 1 | ||||
| 277 | /// %3 = insertelement <4 x i8> %2, i8 %y1, i32 2 | ||||
| 278 | /// %4 = insertelement <4 x i8> %3, i8 %y2, i32 3 | ||||
| 279 | /// %5 = mul <4 x i8> %4, %4 | ||||
| 280 | /// %6 = extractelement <4 x i8> %5, i32 0 | ||||
| 281 | /// %ins1 = insertelement <4 x i8> poison, i8 %6, i32 0 | ||||
| 282 | /// %7 = extractelement <4 x i8> %5, i32 1 | ||||
| 283 | /// %ins2 = insertelement <4 x i8> %ins1, i8 %7, i32 1 | ||||
| 284 | /// %8 = extractelement <4 x i8> %5, i32 2 | ||||
| 285 | /// %ins3 = insertelement <4 x i8> %ins2, i8 %8, i32 2 | ||||
| 286 | /// %9 = extractelement <4 x i8> %5, i32 3 | ||||
| 287 | /// %ins4 = insertelement <4 x i8> %ins3, i8 %9, i32 3 | ||||
| 288 | /// ret <4 x i8> %ins4 | ||||
| 289 | /// InstCombiner transforms this into a shuffle and vector mul | ||||
| 290 | /// Mask will return the Shuffle Mask equivalent to the extracted elements. | ||||
| 291 | /// TODO: Can we split off and reuse the shuffle mask detection from | ||||
| 292 | /// TargetTransformInfo::getInstructionThroughput? | ||||
| 293 | static Optional<TargetTransformInfo::ShuffleKind> | ||||
| 294 | isShuffle(ArrayRef<Value *> VL, SmallVectorImpl<int> &Mask) { | ||||
| 295 | auto *EI0 = cast<ExtractElementInst>(VL[0]); | ||||
| 296 | unsigned Size = | ||||
| 297 | cast<FixedVectorType>(EI0->getVectorOperandType())->getNumElements(); | ||||
| 298 | Value *Vec1 = nullptr; | ||||
| 299 | Value *Vec2 = nullptr; | ||||
| 300 | enum ShuffleMode { Unknown, Select, Permute }; | ||||
| 301 | ShuffleMode CommonShuffleMode = Unknown; | ||||
| 302 | for (unsigned I = 0, E = VL.size(); I < E; ++I) { | ||||
| 303 | auto *EI = cast<ExtractElementInst>(VL[I]); | ||||
| 304 | auto *Vec = EI->getVectorOperand(); | ||||
| 305 | // All vector operands must have the same number of vector elements. | ||||
| 306 | if (cast<FixedVectorType>(Vec->getType())->getNumElements() != Size) | ||||
| 307 | return None; | ||||
| 308 | auto *Idx = dyn_cast<ConstantInt>(EI->getIndexOperand()); | ||||
| 309 | if (!Idx) | ||||
| 310 | return None; | ||||
| 311 | // Undefined behavior if Idx is negative or >= Size. | ||||
| 312 | if (Idx->getValue().uge(Size)) { | ||||
| 313 | Mask.push_back(UndefMaskElem); | ||||
| 314 | continue; | ||||
| 315 | } | ||||
| 316 | unsigned IntIdx = Idx->getValue().getZExtValue(); | ||||
| 317 | Mask.push_back(IntIdx); | ||||
| 318 | // We can extractelement from undef or poison vector. | ||||
| 319 | if (isa<UndefValue>(Vec)) | ||||
| 320 | continue; | ||||
| 321 | // For correct shuffling we have to have at most 2 different vector operands | ||||
| 322 | // in all extractelement instructions. | ||||
| 323 | if (!Vec1 || Vec1 == Vec) | ||||
| 324 | Vec1 = Vec; | ||||
| 325 | else if (!Vec2 || Vec2 == Vec) | ||||
| 326 | Vec2 = Vec; | ||||
| 327 | else | ||||
| 328 | return None; | ||||
| 329 | if (CommonShuffleMode == Permute) | ||||
| 330 | continue; | ||||
| 331 | // If the extract index is not the same as the operation number, it is a | ||||
| 332 | // permutation. | ||||
| 333 | if (IntIdx != I) { | ||||
| 334 | CommonShuffleMode = Permute; | ||||
| 335 | continue; | ||||
| 336 | } | ||||
| 337 | CommonShuffleMode = Select; | ||||
| 338 | } | ||||
| 339 | // If we're not crossing lanes in different vectors, consider it as blending. | ||||
| 340 | if (CommonShuffleMode == Select && Vec2) | ||||
| 341 | return TargetTransformInfo::SK_Select; | ||||
| 342 | // If Vec2 was never used, we have a permutation of a single vector, otherwise | ||||
| 343 | // we have permutation of 2 vectors. | ||||
| 344 | return Vec2 ? TargetTransformInfo::SK_PermuteTwoSrc | ||||
| 345 | : TargetTransformInfo::SK_PermuteSingleSrc; | ||||
| 346 | } | ||||
| 347 | |||||
| 348 | namespace { | ||||
| 349 | |||||
| 350 | /// Main data required for vectorization of instructions. | ||||
| 351 | struct InstructionsState { | ||||
| 352 | /// The very first instruction in the list with the main opcode. | ||||
| 353 | Value *OpValue = nullptr; | ||||
| 354 | |||||
| 355 | /// The main/alternate instruction. | ||||
| 356 | Instruction *MainOp = nullptr; | ||||
| 357 | Instruction *AltOp = nullptr; | ||||
| 358 | |||||
| 359 | /// The main/alternate opcodes for the list of instructions. | ||||
| 360 | unsigned getOpcode() const { | ||||
| 361 | return MainOp ? MainOp->getOpcode() : 0; | ||||
| 362 | } | ||||
| 363 | |||||
| 364 | unsigned getAltOpcode() const { | ||||
| 365 | return AltOp ? AltOp->getOpcode() : 0; | ||||
| 366 | } | ||||
| 367 | |||||
| 368 | /// Some of the instructions in the list have alternate opcodes. | ||||
| 369 | bool isAltShuffle() const { return getOpcode() != getAltOpcode(); } | ||||
| 370 | |||||
| 371 | bool isOpcodeOrAlt(Instruction *I) const { | ||||
| 372 | unsigned CheckedOpcode = I->getOpcode(); | ||||
| 373 | return getOpcode() == CheckedOpcode || getAltOpcode() == CheckedOpcode; | ||||
| 374 | } | ||||
| 375 | |||||
| 376 | InstructionsState() = delete; | ||||
| 377 | InstructionsState(Value *OpValue, Instruction *MainOp, Instruction *AltOp) | ||||
| 378 | : OpValue(OpValue), MainOp(MainOp), AltOp(AltOp) {} | ||||
| 379 | }; | ||||
| 380 | |||||
| 381 | } // end anonymous namespace | ||||
| 382 | |||||
| 383 | /// Chooses the correct key for scheduling data. If \p Op has the same (or | ||||
| 384 | /// alternate) opcode as \p OpValue, the key is \p Op. Otherwise the key is \p | ||||
| 385 | /// OpValue. | ||||
| 386 | static Value *isOneOf(const InstructionsState &S, Value *Op) { | ||||
| 387 | auto *I = dyn_cast<Instruction>(Op); | ||||
| 388 | if (I && S.isOpcodeOrAlt(I)) | ||||
| 389 | return Op; | ||||
| 390 | return S.OpValue; | ||||
| 391 | } | ||||
| 392 | |||||
| 393 | /// \returns true if \p Opcode is allowed as part of of the main/alternate | ||||
| 394 | /// instruction for SLP vectorization. | ||||
| 395 | /// | ||||
| 396 | /// Example of unsupported opcode is SDIV that can potentially cause UB if the | ||||
| 397 | /// "shuffled out" lane would result in division by zero. | ||||
| 398 | static bool isValidForAlternation(unsigned Opcode) { | ||||
| 399 | if (Instruction::isIntDivRem(Opcode)) | ||||
| 400 | return false; | ||||
| 401 | |||||
| 402 | return true; | ||||
| 403 | } | ||||
| 404 | |||||
| 405 | /// \returns analysis of the Instructions in \p VL described in | ||||
| 406 | /// InstructionsState, the Opcode that we suppose the whole list | ||||
| 407 | /// could be vectorized even if its structure is diverse. | ||||
| 408 | static InstructionsState getSameOpcode(ArrayRef<Value *> VL, | ||||
| 409 | unsigned BaseIndex = 0) { | ||||
| 410 | // Make sure these are all Instructions. | ||||
| 411 | if (llvm::any_of(VL, [](Value *V) { return !isa<Instruction>(V); })) | ||||
| 412 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); | ||||
| 413 | |||||
| 414 | bool IsCastOp = isa<CastInst>(VL[BaseIndex]); | ||||
| 415 | bool IsBinOp = isa<BinaryOperator>(VL[BaseIndex]); | ||||
| 416 | unsigned Opcode = cast<Instruction>(VL[BaseIndex])->getOpcode(); | ||||
| 417 | unsigned AltOpcode = Opcode; | ||||
| 418 | unsigned AltIndex = BaseIndex; | ||||
| 419 | |||||
| 420 | // Check for one alternate opcode from another BinaryOperator. | ||||
| 421 | // TODO - generalize to support all operators (types, calls etc.). | ||||
| 422 | for (int Cnt = 0, E = VL.size(); Cnt < E; Cnt++) { | ||||
| 423 | unsigned InstOpcode = cast<Instruction>(VL[Cnt])->getOpcode(); | ||||
| 424 | if (IsBinOp && isa<BinaryOperator>(VL[Cnt])) { | ||||
| 425 | if (InstOpcode == Opcode || InstOpcode == AltOpcode) | ||||
| 426 | continue; | ||||
| 427 | if (Opcode == AltOpcode && isValidForAlternation(InstOpcode) && | ||||
| 428 | isValidForAlternation(Opcode)) { | ||||
| 429 | AltOpcode = InstOpcode; | ||||
| 430 | AltIndex = Cnt; | ||||
| 431 | continue; | ||||
| 432 | } | ||||
| 433 | } else if (IsCastOp && isa<CastInst>(VL[Cnt])) { | ||||
| 434 | Type *Ty0 = cast<Instruction>(VL[BaseIndex])->getOperand(0)->getType(); | ||||
| 435 | Type *Ty1 = cast<Instruction>(VL[Cnt])->getOperand(0)->getType(); | ||||
| 436 | if (Ty0 == Ty1) { | ||||
| 437 | if (InstOpcode == Opcode || InstOpcode == AltOpcode) | ||||
| 438 | continue; | ||||
| 439 | if (Opcode == AltOpcode) { | ||||
| 440 | assert(isValidForAlternation(Opcode) &&(static_cast <bool> (isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && "Cast isn't safe for alternation, logic needs to be updated!" ) ? void (0) : __assert_fail ("isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && \"Cast isn't safe for alternation, logic needs to be updated!\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 442, __extension__ __PRETTY_FUNCTION__)) | ||||
| 441 | isValidForAlternation(InstOpcode) &&(static_cast <bool> (isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && "Cast isn't safe for alternation, logic needs to be updated!" ) ? void (0) : __assert_fail ("isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && \"Cast isn't safe for alternation, logic needs to be updated!\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 442, __extension__ __PRETTY_FUNCTION__)) | ||||
| 442 | "Cast isn't safe for alternation, logic needs to be updated!")(static_cast <bool> (isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && "Cast isn't safe for alternation, logic needs to be updated!" ) ? void (0) : __assert_fail ("isValidForAlternation(Opcode) && isValidForAlternation(InstOpcode) && \"Cast isn't safe for alternation, logic needs to be updated!\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 442, __extension__ __PRETTY_FUNCTION__)); | ||||
| 443 | AltOpcode = InstOpcode; | ||||
| 444 | AltIndex = Cnt; | ||||
| 445 | continue; | ||||
| 446 | } | ||||
| 447 | } | ||||
| 448 | } else if (InstOpcode == Opcode || InstOpcode == AltOpcode) | ||||
| 449 | continue; | ||||
| 450 | return InstructionsState(VL[BaseIndex], nullptr, nullptr); | ||||
| 451 | } | ||||
| 452 | |||||
| 453 | return InstructionsState(VL[BaseIndex], cast<Instruction>(VL[BaseIndex]), | ||||
| 454 | cast<Instruction>(VL[AltIndex])); | ||||
| 455 | } | ||||
| 456 | |||||
| 457 | /// \returns true if all of the values in \p VL have the same type or false | ||||
| 458 | /// otherwise. | ||||
| 459 | static bool allSameType(ArrayRef<Value *> VL) { | ||||
| 460 | Type *Ty = VL[0]->getType(); | ||||
| 461 | for (int i = 1, e = VL.size(); i < e; i++) | ||||
| 462 | if (VL[i]->getType() != Ty) | ||||
| 463 | return false; | ||||
| 464 | |||||
| 465 | return true; | ||||
| 466 | } | ||||
| 467 | |||||
| 468 | /// \returns True if Extract{Value,Element} instruction extracts element Idx. | ||||
| 469 | static Optional<unsigned> getExtractIndex(Instruction *E) { | ||||
| 470 | unsigned Opcode = E->getOpcode(); | ||||
| 471 | assert((Opcode == Instruction::ExtractElement ||(static_cast <bool> ((Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && "Expected extractelement or extractvalue instruction." ) ? void (0) : __assert_fail ("(Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && \"Expected extractelement or extractvalue instruction.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 473, __extension__ __PRETTY_FUNCTION__)) | ||||
| 472 | Opcode == Instruction::ExtractValue) &&(static_cast <bool> ((Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && "Expected extractelement or extractvalue instruction." ) ? void (0) : __assert_fail ("(Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && \"Expected extractelement or extractvalue instruction.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 473, __extension__ __PRETTY_FUNCTION__)) | ||||
| 473 | "Expected extractelement or extractvalue instruction.")(static_cast <bool> ((Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && "Expected extractelement or extractvalue instruction." ) ? void (0) : __assert_fail ("(Opcode == Instruction::ExtractElement || Opcode == Instruction::ExtractValue) && \"Expected extractelement or extractvalue instruction.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 473, __extension__ __PRETTY_FUNCTION__)); | ||||
| 474 | if (Opcode == Instruction::ExtractElement) { | ||||
| 475 | auto *CI = dyn_cast<ConstantInt>(E->getOperand(1)); | ||||
| 476 | if (!CI) | ||||
| 477 | return None; | ||||
| 478 | return CI->getZExtValue(); | ||||
| 479 | } | ||||
| 480 | ExtractValueInst *EI = cast<ExtractValueInst>(E); | ||||
| 481 | if (EI->getNumIndices() != 1) | ||||
| 482 | return None; | ||||
| 483 | return *EI->idx_begin(); | ||||
| 484 | } | ||||
| 485 | |||||
| 486 | /// \returns True if in-tree use also needs extract. This refers to | ||||
| 487 | /// possible scalar operand in vectorized instruction. | ||||
| 488 | static bool InTreeUserNeedToExtract(Value *Scalar, Instruction *UserInst, | ||||
| 489 | TargetLibraryInfo *TLI) { | ||||
| 490 | unsigned Opcode = UserInst->getOpcode(); | ||||
| 491 | switch (Opcode) { | ||||
| 492 | case Instruction::Load: { | ||||
| 493 | LoadInst *LI = cast<LoadInst>(UserInst); | ||||
| 494 | return (LI->getPointerOperand() == Scalar); | ||||
| 495 | } | ||||
| 496 | case Instruction::Store: { | ||||
| 497 | StoreInst *SI = cast<StoreInst>(UserInst); | ||||
| 498 | return (SI->getPointerOperand() == Scalar); | ||||
| 499 | } | ||||
| 500 | case Instruction::Call: { | ||||
| 501 | CallInst *CI = cast<CallInst>(UserInst); | ||||
| 502 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||
| 503 | for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) { | ||||
| 504 | if (hasVectorInstrinsicScalarOpd(ID, i)) | ||||
| 505 | return (CI->getArgOperand(i) == Scalar); | ||||
| 506 | } | ||||
| 507 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | ||||
| 508 | } | ||||
| 509 | default: | ||||
| 510 | return false; | ||||
| 511 | } | ||||
| 512 | } | ||||
| 513 | |||||
| 514 | /// \returns the AA location that is being access by the instruction. | ||||
| 515 | static MemoryLocation getLocation(Instruction *I, AAResults *AA) { | ||||
| 516 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) | ||||
| 517 | return MemoryLocation::get(SI); | ||||
| 518 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | ||||
| 519 | return MemoryLocation::get(LI); | ||||
| 520 | return MemoryLocation(); | ||||
| 521 | } | ||||
| 522 | |||||
| 523 | /// \returns True if the instruction is not a volatile or atomic load/store. | ||||
| 524 | static bool isSimple(Instruction *I) { | ||||
| 525 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) | ||||
| 526 | return LI->isSimple(); | ||||
| 527 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) | ||||
| 528 | return SI->isSimple(); | ||||
| 529 | if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) | ||||
| 530 | return !MI->isVolatile(); | ||||
| 531 | return true; | ||||
| 532 | } | ||||
| 533 | |||||
| 534 | namespace llvm { | ||||
| 535 | |||||
| 536 | static void inversePermutation(ArrayRef<unsigned> Indices, | ||||
| 537 | SmallVectorImpl<int> &Mask) { | ||||
| 538 | Mask.clear(); | ||||
| 539 | const unsigned E = Indices.size(); | ||||
| 540 | Mask.resize(E, E + 1); | ||||
| 541 | for (unsigned I = 0; I < E; ++I) | ||||
| 542 | Mask[Indices[I]] = I; | ||||
| 543 | } | ||||
| 544 | |||||
| 545 | namespace slpvectorizer { | ||||
| 546 | |||||
| 547 | /// Bottom Up SLP Vectorizer. | ||||
| 548 | class BoUpSLP { | ||||
| 549 | struct TreeEntry; | ||||
| 550 | struct ScheduleData; | ||||
| 551 | |||||
| 552 | public: | ||||
| 553 | using ValueList = SmallVector<Value *, 8>; | ||||
| 554 | using InstrList = SmallVector<Instruction *, 16>; | ||||
| 555 | using ValueSet = SmallPtrSet<Value *, 16>; | ||||
| 556 | using StoreList = SmallVector<StoreInst *, 8>; | ||||
| 557 | using ExtraValueToDebugLocsMap = | ||||
| 558 | MapVector<Value *, SmallVector<Instruction *, 2>>; | ||||
| 559 | using OrdersType = SmallVector<unsigned, 4>; | ||||
| 560 | |||||
| 561 | BoUpSLP(Function *Func, ScalarEvolution *Se, TargetTransformInfo *Tti, | ||||
| 562 | TargetLibraryInfo *TLi, AAResults *Aa, LoopInfo *Li, | ||||
| 563 | DominatorTree *Dt, AssumptionCache *AC, DemandedBits *DB, | ||||
| 564 | const DataLayout *DL, OptimizationRemarkEmitter *ORE) | ||||
| 565 | : F(Func), SE(Se), TTI(Tti), TLI(TLi), AA(Aa), LI(Li), DT(Dt), AC(AC), | ||||
| 566 | DB(DB), DL(DL), ORE(ORE), Builder(Se->getContext()) { | ||||
| 567 | CodeMetrics::collectEphemeralValues(F, AC, EphValues); | ||||
| 568 | // Use the vector register size specified by the target unless overridden | ||||
| 569 | // by a command-line option. | ||||
| 570 | // TODO: It would be better to limit the vectorization factor based on | ||||
| 571 | // data type rather than just register size. For example, x86 AVX has | ||||
| 572 | // 256-bit registers, but it does not support integer operations | ||||
| 573 | // at that width (that requires AVX2). | ||||
| 574 | if (MaxVectorRegSizeOption.getNumOccurrences()) | ||||
| 575 | MaxVecRegSize = MaxVectorRegSizeOption; | ||||
| 576 | else | ||||
| 577 | MaxVecRegSize = | ||||
| 578 | TTI->getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector) | ||||
| 579 | .getFixedSize(); | ||||
| 580 | |||||
| 581 | if (MinVectorRegSizeOption.getNumOccurrences()) | ||||
| 582 | MinVecRegSize = MinVectorRegSizeOption; | ||||
| 583 | else | ||||
| 584 | MinVecRegSize = TTI->getMinVectorRegisterBitWidth(); | ||||
| 585 | } | ||||
| 586 | |||||
| 587 | /// Vectorize the tree that starts with the elements in \p VL. | ||||
| 588 | /// Returns the vectorized root. | ||||
| 589 | Value *vectorizeTree(); | ||||
| 590 | |||||
| 591 | /// Vectorize the tree but with the list of externally used values \p | ||||
| 592 | /// ExternallyUsedValues. Values in this MapVector can be replaced but the | ||||
| 593 | /// generated extractvalue instructions. | ||||
| 594 | Value *vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues); | ||||
| 595 | |||||
| 596 | /// \returns the cost incurred by unwanted spills and fills, caused by | ||||
| 597 | /// holding live values over call sites. | ||||
| 598 | InstructionCost getSpillCost() const; | ||||
| 599 | |||||
| 600 | /// \returns the vectorization cost of the subtree that starts at \p VL. | ||||
| 601 | /// A negative number means that this is profitable. | ||||
| 602 | InstructionCost getTreeCost(); | ||||
| 603 | |||||
| 604 | /// Construct a vectorizable tree that starts at \p Roots, ignoring users for | ||||
| 605 | /// the purpose of scheduling and extraction in the \p UserIgnoreLst. | ||||
| 606 | void buildTree(ArrayRef<Value *> Roots, | ||||
| 607 | ArrayRef<Value *> UserIgnoreLst = None); | ||||
| 608 | |||||
| 609 | /// Construct a vectorizable tree that starts at \p Roots, ignoring users for | ||||
| 610 | /// the purpose of scheduling and extraction in the \p UserIgnoreLst taking | ||||
| 611 | /// into account (and updating it, if required) list of externally used | ||||
| 612 | /// values stored in \p ExternallyUsedValues. | ||||
| 613 | void buildTree(ArrayRef<Value *> Roots, | ||||
| 614 | ExtraValueToDebugLocsMap &ExternallyUsedValues, | ||||
| 615 | ArrayRef<Value *> UserIgnoreLst = None); | ||||
| 616 | |||||
| 617 | /// Clear the internal data structures that are created by 'buildTree'. | ||||
| 618 | void deleteTree() { | ||||
| 619 | VectorizableTree.clear(); | ||||
| 620 | ScalarToTreeEntry.clear(); | ||||
| 621 | MustGather.clear(); | ||||
| 622 | ExternalUses.clear(); | ||||
| 623 | NumOpsWantToKeepOrder.clear(); | ||||
| 624 | NumOpsWantToKeepOriginalOrder = 0; | ||||
| 625 | for (auto &Iter : BlocksSchedules) { | ||||
| 626 | BlockScheduling *BS = Iter.second.get(); | ||||
| 627 | BS->clear(); | ||||
| 628 | } | ||||
| 629 | MinBWs.clear(); | ||||
| 630 | InstrElementSize.clear(); | ||||
| 631 | } | ||||
| 632 | |||||
| 633 | unsigned getTreeSize() const { return VectorizableTree.size(); } | ||||
| 634 | |||||
| 635 | /// Perform LICM and CSE on the newly generated gather sequences. | ||||
| 636 | void optimizeGatherSequence(); | ||||
| 637 | |||||
| 638 | /// \returns The best order of instructions for vectorization. | ||||
| 639 | Optional<ArrayRef<unsigned>> bestOrder() const { | ||||
| 640 | assert(llvm::all_of((static_cast <bool> (llvm::all_of( NumOpsWantToKeepOrder , [this](const decltype(NumOpsWantToKeepOrder)::value_type & D) { return D.getFirst().size() == VectorizableTree[0]->Scalars .size(); }) && "All orders must have the same size as number of instructions in " "tree node.") ? void (0) : __assert_fail ("llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst().size() == VectorizableTree[0]->Scalars.size(); }) && \"All orders must have the same size as number of instructions in \" \"tree node.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 647, __extension__ __PRETTY_FUNCTION__)) | ||||
| 641 | NumOpsWantToKeepOrder,(static_cast <bool> (llvm::all_of( NumOpsWantToKeepOrder , [this](const decltype(NumOpsWantToKeepOrder)::value_type & D) { return D.getFirst().size() == VectorizableTree[0]->Scalars .size(); }) && "All orders must have the same size as number of instructions in " "tree node.") ? void (0) : __assert_fail ("llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst().size() == VectorizableTree[0]->Scalars.size(); }) && \"All orders must have the same size as number of instructions in \" \"tree node.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 647, __extension__ __PRETTY_FUNCTION__)) | ||||
| 642 | [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) {(static_cast <bool> (llvm::all_of( NumOpsWantToKeepOrder , [this](const decltype(NumOpsWantToKeepOrder)::value_type & D) { return D.getFirst().size() == VectorizableTree[0]->Scalars .size(); }) && "All orders must have the same size as number of instructions in " "tree node.") ? void (0) : __assert_fail ("llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst().size() == VectorizableTree[0]->Scalars.size(); }) && \"All orders must have the same size as number of instructions in \" \"tree node.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 647, __extension__ __PRETTY_FUNCTION__)) | ||||
| 643 | return D.getFirst().size() ==(static_cast <bool> (llvm::all_of( NumOpsWantToKeepOrder , [this](const decltype(NumOpsWantToKeepOrder)::value_type & D) { return D.getFirst().size() == VectorizableTree[0]->Scalars .size(); }) && "All orders must have the same size as number of instructions in " "tree node.") ? void (0) : __assert_fail ("llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst().size() == VectorizableTree[0]->Scalars.size(); }) && \"All orders must have the same size as number of instructions in \" \"tree node.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 647, __extension__ __PRETTY_FUNCTION__)) | ||||
| 644 | VectorizableTree[0]->Scalars.size();(static_cast <bool> (llvm::all_of( NumOpsWantToKeepOrder , [this](const decltype(NumOpsWantToKeepOrder)::value_type & D) { return D.getFirst().size() == VectorizableTree[0]->Scalars .size(); }) && "All orders must have the same size as number of instructions in " "tree node.") ? void (0) : __assert_fail ("llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst().size() == VectorizableTree[0]->Scalars.size(); }) && \"All orders must have the same size as number of instructions in \" \"tree node.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 647, __extension__ __PRETTY_FUNCTION__)) | ||||
| 645 | }) &&(static_cast <bool> (llvm::all_of( NumOpsWantToKeepOrder , [this](const decltype(NumOpsWantToKeepOrder)::value_type & D) { return D.getFirst().size() == VectorizableTree[0]->Scalars .size(); }) && "All orders must have the same size as number of instructions in " "tree node.") ? void (0) : __assert_fail ("llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst().size() == VectorizableTree[0]->Scalars.size(); }) && \"All orders must have the same size as number of instructions in \" \"tree node.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 647, __extension__ __PRETTY_FUNCTION__)) | ||||
| 646 | "All orders must have the same size as number of instructions in "(static_cast <bool> (llvm::all_of( NumOpsWantToKeepOrder , [this](const decltype(NumOpsWantToKeepOrder)::value_type & D) { return D.getFirst().size() == VectorizableTree[0]->Scalars .size(); }) && "All orders must have the same size as number of instructions in " "tree node.") ? void (0) : __assert_fail ("llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst().size() == VectorizableTree[0]->Scalars.size(); }) && \"All orders must have the same size as number of instructions in \" \"tree node.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 647, __extension__ __PRETTY_FUNCTION__)) | ||||
| 647 | "tree node.")(static_cast <bool> (llvm::all_of( NumOpsWantToKeepOrder , [this](const decltype(NumOpsWantToKeepOrder)::value_type & D) { return D.getFirst().size() == VectorizableTree[0]->Scalars .size(); }) && "All orders must have the same size as number of instructions in " "tree node.") ? void (0) : __assert_fail ("llvm::all_of( NumOpsWantToKeepOrder, [this](const decltype(NumOpsWantToKeepOrder)::value_type &D) { return D.getFirst().size() == VectorizableTree[0]->Scalars.size(); }) && \"All orders must have the same size as number of instructions in \" \"tree node.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 647, __extension__ __PRETTY_FUNCTION__)); | ||||
| 648 | auto I = std::max_element( | ||||
| 649 | NumOpsWantToKeepOrder.begin(), NumOpsWantToKeepOrder.end(), | ||||
| 650 | [](const decltype(NumOpsWantToKeepOrder)::value_type &D1, | ||||
| 651 | const decltype(NumOpsWantToKeepOrder)::value_type &D2) { | ||||
| 652 | return D1.second < D2.second; | ||||
| 653 | }); | ||||
| 654 | if (I == NumOpsWantToKeepOrder.end() || | ||||
| 655 | I->getSecond() <= NumOpsWantToKeepOriginalOrder) | ||||
| 656 | return None; | ||||
| 657 | |||||
| 658 | return makeArrayRef(I->getFirst()); | ||||
| 659 | } | ||||
| 660 | |||||
| 661 | /// Builds the correct order for root instructions. | ||||
| 662 | /// If some leaves have the same instructions to be vectorized, we may | ||||
| 663 | /// incorrectly evaluate the best order for the root node (it is built for the | ||||
| 664 | /// vector of instructions without repeated instructions and, thus, has less | ||||
| 665 | /// elements than the root node). This function builds the correct order for | ||||
| 666 | /// the root node. | ||||
| 667 | /// For example, if the root node is \<a+b, a+c, a+d, f+e\>, then the leaves | ||||
| 668 | /// are \<a, a, a, f\> and \<b, c, d, e\>. When we try to vectorize the first | ||||
| 669 | /// leaf, it will be shrink to \<a, b\>. If instructions in this leaf should | ||||
| 670 | /// be reordered, the best order will be \<1, 0\>. We need to extend this | ||||
| 671 | /// order for the root node. For the root node this order should look like | ||||
| 672 | /// \<3, 0, 1, 2\>. This function extends the order for the reused | ||||
| 673 | /// instructions. | ||||
| 674 | void findRootOrder(OrdersType &Order) { | ||||
| 675 | // If the leaf has the same number of instructions to vectorize as the root | ||||
| 676 | // - order must be set already. | ||||
| 677 | unsigned RootSize = VectorizableTree[0]->Scalars.size(); | ||||
| 678 | if (Order.size() == RootSize) | ||||
| 679 | return; | ||||
| 680 | SmallVector<unsigned, 4> RealOrder(Order.size()); | ||||
| 681 | std::swap(Order, RealOrder); | ||||
| 682 | SmallVector<int, 4> Mask; | ||||
| 683 | inversePermutation(RealOrder, Mask); | ||||
| 684 | Order.assign(Mask.begin(), Mask.end()); | ||||
| 685 | // The leaf has less number of instructions - need to find the true order of | ||||
| 686 | // the root. | ||||
| 687 | // Scan the nodes starting from the leaf back to the root. | ||||
| 688 | const TreeEntry *PNode = VectorizableTree.back().get(); | ||||
| 689 | SmallVector<const TreeEntry *, 4> Nodes(1, PNode); | ||||
| 690 | SmallPtrSet<const TreeEntry *, 4> Visited; | ||||
| 691 | while (!Nodes.empty() && Order.size() != RootSize) { | ||||
| 692 | const TreeEntry *PNode = Nodes.pop_back_val(); | ||||
| 693 | if (!Visited.insert(PNode).second) | ||||
| 694 | continue; | ||||
| 695 | const TreeEntry &Node = *PNode; | ||||
| 696 | for (const EdgeInfo &EI : Node.UserTreeIndices) | ||||
| 697 | if (EI.UserTE) | ||||
| 698 | Nodes.push_back(EI.UserTE); | ||||
| 699 | if (Node.ReuseShuffleIndices.empty()) | ||||
| 700 | continue; | ||||
| 701 | // Build the order for the parent node. | ||||
| 702 | OrdersType NewOrder(Node.ReuseShuffleIndices.size(), RootSize); | ||||
| 703 | SmallVector<unsigned, 4> OrderCounter(Order.size(), 0); | ||||
| 704 | // The algorithm of the order extension is: | ||||
| 705 | // 1. Calculate the number of the same instructions for the order. | ||||
| 706 | // 2. Calculate the index of the new order: total number of instructions | ||||
| 707 | // with order less than the order of the current instruction + reuse | ||||
| 708 | // number of the current instruction. | ||||
| 709 | // 3. The new order is just the index of the instruction in the original | ||||
| 710 | // vector of the instructions. | ||||
| 711 | for (unsigned I : Node.ReuseShuffleIndices) | ||||
| 712 | ++OrderCounter[Order[I]]; | ||||
| 713 | SmallVector<unsigned, 4> CurrentCounter(Order.size(), 0); | ||||
| 714 | for (unsigned I = 0, E = Node.ReuseShuffleIndices.size(); I < E; ++I) { | ||||
| 715 | unsigned ReusedIdx = Node.ReuseShuffleIndices[I]; | ||||
| 716 | unsigned OrderIdx = Order[ReusedIdx]; | ||||
| 717 | unsigned NewIdx = 0; | ||||
| 718 | for (unsigned J = 0; J < OrderIdx; ++J) | ||||
| 719 | NewIdx += OrderCounter[J]; | ||||
| 720 | NewIdx += CurrentCounter[OrderIdx]; | ||||
| 721 | ++CurrentCounter[OrderIdx]; | ||||
| 722 | assert(NewOrder[NewIdx] == RootSize &&(static_cast <bool> (NewOrder[NewIdx] == RootSize && "The order index should not be written already.") ? void (0) : __assert_fail ("NewOrder[NewIdx] == RootSize && \"The order index should not be written already.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 723, __extension__ __PRETTY_FUNCTION__)) | ||||
| 723 | "The order index should not be written already.")(static_cast <bool> (NewOrder[NewIdx] == RootSize && "The order index should not be written already.") ? void (0) : __assert_fail ("NewOrder[NewIdx] == RootSize && \"The order index should not be written already.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 723, __extension__ __PRETTY_FUNCTION__)); | ||||
| 724 | NewOrder[NewIdx] = I; | ||||
| 725 | } | ||||
| 726 | std::swap(Order, NewOrder); | ||||
| 727 | } | ||||
| 728 | assert(Order.size() == RootSize &&(static_cast <bool> (Order.size() == RootSize && "Root node is expected or the size of the order must be the same as " "the number of elements in the root node.") ? void (0) : __assert_fail ("Order.size() == RootSize && \"Root node is expected or the size of the order must be the same as \" \"the number of elements in the root node.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 730, __extension__ __PRETTY_FUNCTION__)) | ||||
| 729 | "Root node is expected or the size of the order must be the same as "(static_cast <bool> (Order.size() == RootSize && "Root node is expected or the size of the order must be the same as " "the number of elements in the root node.") ? void (0) : __assert_fail ("Order.size() == RootSize && \"Root node is expected or the size of the order must be the same as \" \"the number of elements in the root node.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 730, __extension__ __PRETTY_FUNCTION__)) | ||||
| 730 | "the number of elements in the root node.")(static_cast <bool> (Order.size() == RootSize && "Root node is expected or the size of the order must be the same as " "the number of elements in the root node.") ? void (0) : __assert_fail ("Order.size() == RootSize && \"Root node is expected or the size of the order must be the same as \" \"the number of elements in the root node.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 730, __extension__ __PRETTY_FUNCTION__)); | ||||
| 731 | assert(llvm::all_of(Order,(static_cast <bool> (llvm::all_of(Order, [RootSize](unsigned Val) { return Val != RootSize; }) && "All indices must be initialized" ) ? void (0) : __assert_fail ("llvm::all_of(Order, [RootSize](unsigned Val) { return Val != RootSize; }) && \"All indices must be initialized\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 733, __extension__ __PRETTY_FUNCTION__)) | ||||
| 732 | [RootSize](unsigned Val) { return Val != RootSize; }) &&(static_cast <bool> (llvm::all_of(Order, [RootSize](unsigned Val) { return Val != RootSize; }) && "All indices must be initialized" ) ? void (0) : __assert_fail ("llvm::all_of(Order, [RootSize](unsigned Val) { return Val != RootSize; }) && \"All indices must be initialized\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 733, __extension__ __PRETTY_FUNCTION__)) | ||||
| 733 | "All indices must be initialized")(static_cast <bool> (llvm::all_of(Order, [RootSize](unsigned Val) { return Val != RootSize; }) && "All indices must be initialized" ) ? void (0) : __assert_fail ("llvm::all_of(Order, [RootSize](unsigned Val) { return Val != RootSize; }) && \"All indices must be initialized\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 733, __extension__ __PRETTY_FUNCTION__)); | ||||
| 734 | } | ||||
| 735 | |||||
| 736 | /// \return The vector element size in bits to use when vectorizing the | ||||
| 737 | /// expression tree ending at \p V. If V is a store, the size is the width of | ||||
| 738 | /// the stored value. Otherwise, the size is the width of the largest loaded | ||||
| 739 | /// value reaching V. This method is used by the vectorizer to calculate | ||||
| 740 | /// vectorization factors. | ||||
| 741 | unsigned getVectorElementSize(Value *V); | ||||
| 742 | |||||
| 743 | /// Compute the minimum type sizes required to represent the entries in a | ||||
| 744 | /// vectorizable tree. | ||||
| 745 | void computeMinimumValueSizes(); | ||||
| 746 | |||||
| 747 | // \returns maximum vector register size as set by TTI or overridden by cl::opt. | ||||
| 748 | unsigned getMaxVecRegSize() const { | ||||
| 749 | return MaxVecRegSize; | ||||
| 750 | } | ||||
| 751 | |||||
| 752 | // \returns minimum vector register size as set by cl::opt. | ||||
| 753 | unsigned getMinVecRegSize() const { | ||||
| 754 | return MinVecRegSize; | ||||
| 755 | } | ||||
| 756 | |||||
| 757 | unsigned getMaximumVF(unsigned ElemWidth, unsigned Opcode) const { | ||||
| 758 | unsigned MaxVF = MaxVFOption.getNumOccurrences() ? | ||||
| 759 | MaxVFOption : TTI->getMaximumVF(ElemWidth, Opcode); | ||||
| 760 | return MaxVF ? MaxVF : UINT_MAX(2147483647 *2U +1U); | ||||
| 761 | } | ||||
| 762 | |||||
| 763 | /// Check if homogeneous aggregate is isomorphic to some VectorType. | ||||
| 764 | /// Accepts homogeneous multidimensional aggregate of scalars/vectors like | ||||
| 765 | /// {[4 x i16], [4 x i16]}, { <2 x float>, <2 x float> }, | ||||
| 766 | /// {{{i16, i16}, {i16, i16}}, {{i16, i16}, {i16, i16}}} and so on. | ||||
| 767 | /// | ||||
| 768 | /// \returns number of elements in vector if isomorphism exists, 0 otherwise. | ||||
| 769 | unsigned canMapToVector(Type *T, const DataLayout &DL) const; | ||||
| 770 | |||||
| 771 | /// \returns True if the VectorizableTree is both tiny and not fully | ||||
| 772 | /// vectorizable. We do not vectorize such trees. | ||||
| 773 | bool isTreeTinyAndNotFullyVectorizable() const; | ||||
| 774 | |||||
| 775 | /// Assume that a legal-sized 'or'-reduction of shifted/zexted loaded values | ||||
| 776 | /// can be load combined in the backend. Load combining may not be allowed in | ||||
| 777 | /// the IR optimizer, so we do not want to alter the pattern. For example, | ||||
| 778 | /// partially transforming a scalar bswap() pattern into vector code is | ||||
| 779 | /// effectively impossible for the backend to undo. | ||||
| 780 | /// TODO: If load combining is allowed in the IR optimizer, this analysis | ||||
| 781 | /// may not be necessary. | ||||
| 782 | bool isLoadCombineReductionCandidate(RecurKind RdxKind) const; | ||||
| 783 | |||||
| 784 | /// Assume that a vector of stores of bitwise-or/shifted/zexted loaded values | ||||
| 785 | /// can be load combined in the backend. Load combining may not be allowed in | ||||
| 786 | /// the IR optimizer, so we do not want to alter the pattern. For example, | ||||
| 787 | /// partially transforming a scalar bswap() pattern into vector code is | ||||
| 788 | /// effectively impossible for the backend to undo. | ||||
| 789 | /// TODO: If load combining is allowed in the IR optimizer, this analysis | ||||
| 790 | /// may not be necessary. | ||||
| 791 | bool isLoadCombineCandidate() const; | ||||
| 792 | |||||
| 793 | OptimizationRemarkEmitter *getORE() { return ORE; } | ||||
| 794 | |||||
| 795 | /// This structure holds any data we need about the edges being traversed | ||||
| 796 | /// during buildTree_rec(). We keep track of: | ||||
| 797 | /// (i) the user TreeEntry index, and | ||||
| 798 | /// (ii) the index of the edge. | ||||
| 799 | struct EdgeInfo { | ||||
| 800 | EdgeInfo() = default; | ||||
| 801 | EdgeInfo(TreeEntry *UserTE, unsigned EdgeIdx) | ||||
| 802 | : UserTE(UserTE), EdgeIdx(EdgeIdx) {} | ||||
| 803 | /// The user TreeEntry. | ||||
| 804 | TreeEntry *UserTE = nullptr; | ||||
| 805 | /// The operand index of the use. | ||||
| 806 | unsigned EdgeIdx = UINT_MAX(2147483647 *2U +1U); | ||||
| 807 | #ifndef NDEBUG | ||||
| 808 | friend inline raw_ostream &operator<<(raw_ostream &OS, | ||||
| 809 | const BoUpSLP::EdgeInfo &EI) { | ||||
| 810 | EI.dump(OS); | ||||
| 811 | return OS; | ||||
| 812 | } | ||||
| 813 | /// Debug print. | ||||
| 814 | void dump(raw_ostream &OS) const { | ||||
| 815 | OS << "{User:" << (UserTE ? std::to_string(UserTE->Idx) : "null") | ||||
| 816 | << " EdgeIdx:" << EdgeIdx << "}"; | ||||
| 817 | } | ||||
| 818 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { dump(dbgs()); } | ||||
| 819 | #endif | ||||
| 820 | }; | ||||
| 821 | |||||
| 822 | /// A helper data structure to hold the operands of a vector of instructions. | ||||
| 823 | /// This supports a fixed vector length for all operand vectors. | ||||
| 824 | class VLOperands { | ||||
| 825 | /// For each operand we need (i) the value, and (ii) the opcode that it | ||||
| 826 | /// would be attached to if the expression was in a left-linearized form. | ||||
| 827 | /// This is required to avoid illegal operand reordering. | ||||
| 828 | /// For example: | ||||
| 829 | /// \verbatim | ||||
| 830 | /// 0 Op1 | ||||
| 831 | /// |/ | ||||
| 832 | /// Op1 Op2 Linearized + Op2 | ||||
| 833 | /// \ / ----------> |/ | ||||
| 834 | /// - - | ||||
| 835 | /// | ||||
| 836 | /// Op1 - Op2 (0 + Op1) - Op2 | ||||
| 837 | /// \endverbatim | ||||
| 838 | /// | ||||
| 839 | /// Value Op1 is attached to a '+' operation, and Op2 to a '-'. | ||||
| 840 | /// | ||||
| 841 | /// Another way to think of this is to track all the operations across the | ||||
| 842 | /// path from the operand all the way to the root of the tree and to | ||||
| 843 | /// calculate the operation that corresponds to this path. For example, the | ||||
| 844 | /// path from Op2 to the root crosses the RHS of the '-', therefore the | ||||
| 845 | /// corresponding operation is a '-' (which matches the one in the | ||||
| 846 | /// linearized tree, as shown above). | ||||
| 847 | /// | ||||
| 848 | /// For lack of a better term, we refer to this operation as Accumulated | ||||
| 849 | /// Path Operation (APO). | ||||
| 850 | struct OperandData { | ||||
| 851 | OperandData() = default; | ||||
| 852 | OperandData(Value *V, bool APO, bool IsUsed) | ||||
| 853 | : V(V), APO(APO), IsUsed(IsUsed) {} | ||||
| 854 | /// The operand value. | ||||
| 855 | Value *V = nullptr; | ||||
| 856 | /// TreeEntries only allow a single opcode, or an alternate sequence of | ||||
| 857 | /// them (e.g, +, -). Therefore, we can safely use a boolean value for the | ||||
| 858 | /// APO. It is set to 'true' if 'V' is attached to an inverse operation | ||||
| 859 | /// in the left-linearized form (e.g., Sub/Div), and 'false' otherwise | ||||
| 860 | /// (e.g., Add/Mul) | ||||
| 861 | bool APO = false; | ||||
| 862 | /// Helper data for the reordering function. | ||||
| 863 | bool IsUsed = false; | ||||
| 864 | }; | ||||
| 865 | |||||
| 866 | /// During operand reordering, we are trying to select the operand at lane | ||||
| 867 | /// that matches best with the operand at the neighboring lane. Our | ||||
| 868 | /// selection is based on the type of value we are looking for. For example, | ||||
| 869 | /// if the neighboring lane has a load, we need to look for a load that is | ||||
| 870 | /// accessing a consecutive address. These strategies are summarized in the | ||||
| 871 | /// 'ReorderingMode' enumerator. | ||||
| 872 | enum class ReorderingMode { | ||||
| 873 | Load, ///< Matching loads to consecutive memory addresses | ||||
| 874 | Opcode, ///< Matching instructions based on opcode (same or alternate) | ||||
| 875 | Constant, ///< Matching constants | ||||
| 876 | Splat, ///< Matching the same instruction multiple times (broadcast) | ||||
| 877 | Failed, ///< We failed to create a vectorizable group | ||||
| 878 | }; | ||||
| 879 | |||||
| 880 | using OperandDataVec = SmallVector<OperandData, 2>; | ||||
| 881 | |||||
| 882 | /// A vector of operand vectors. | ||||
| 883 | SmallVector<OperandDataVec, 4> OpsVec; | ||||
| 884 | |||||
| 885 | const DataLayout &DL; | ||||
| 886 | ScalarEvolution &SE; | ||||
| 887 | const BoUpSLP &R; | ||||
| 888 | |||||
| 889 | /// \returns the operand data at \p OpIdx and \p Lane. | ||||
| 890 | OperandData &getData(unsigned OpIdx, unsigned Lane) { | ||||
| 891 | return OpsVec[OpIdx][Lane]; | ||||
| 892 | } | ||||
| 893 | |||||
| 894 | /// \returns the operand data at \p OpIdx and \p Lane. Const version. | ||||
| 895 | const OperandData &getData(unsigned OpIdx, unsigned Lane) const { | ||||
| 896 | return OpsVec[OpIdx][Lane]; | ||||
| 897 | } | ||||
| 898 | |||||
| 899 | /// Clears the used flag for all entries. | ||||
| 900 | void clearUsed() { | ||||
| 901 | for (unsigned OpIdx = 0, NumOperands = getNumOperands(); | ||||
| 902 | OpIdx != NumOperands; ++OpIdx) | ||||
| 903 | for (unsigned Lane = 0, NumLanes = getNumLanes(); Lane != NumLanes; | ||||
| 904 | ++Lane) | ||||
| 905 | OpsVec[OpIdx][Lane].IsUsed = false; | ||||
| 906 | } | ||||
| 907 | |||||
| 908 | /// Swap the operand at \p OpIdx1 with that one at \p OpIdx2. | ||||
| 909 | void swap(unsigned OpIdx1, unsigned OpIdx2, unsigned Lane) { | ||||
| 910 | std::swap(OpsVec[OpIdx1][Lane], OpsVec[OpIdx2][Lane]); | ||||
| 911 | } | ||||
| 912 | |||||
| 913 | // The hard-coded scores listed here are not very important. When computing | ||||
| 914 | // the scores of matching one sub-tree with another, we are basically | ||||
| 915 | // counting the number of values that are matching. So even if all scores | ||||
| 916 | // are set to 1, we would still get a decent matching result. | ||||
| 917 | // However, sometimes we have to break ties. For example we may have to | ||||
| 918 | // choose between matching loads vs matching opcodes. This is what these | ||||
| 919 | // scores are helping us with: they provide the order of preference. | ||||
| 920 | |||||
| 921 | /// Loads from consecutive memory addresses, e.g. load(A[i]), load(A[i+1]). | ||||
| 922 | static const int ScoreConsecutiveLoads = 3; | ||||
| 923 | /// ExtractElementInst from same vector and consecutive indexes. | ||||
| 924 | static const int ScoreConsecutiveExtracts = 3; | ||||
| 925 | /// Constants. | ||||
| 926 | static const int ScoreConstants = 2; | ||||
| 927 | /// Instructions with the same opcode. | ||||
| 928 | static const int ScoreSameOpcode = 2; | ||||
| 929 | /// Instructions with alt opcodes (e.g, add + sub). | ||||
| 930 | static const int ScoreAltOpcodes = 1; | ||||
| 931 | /// Identical instructions (a.k.a. splat or broadcast). | ||||
| 932 | static const int ScoreSplat = 1; | ||||
| 933 | /// Matching with an undef is preferable to failing. | ||||
| 934 | static const int ScoreUndef = 1; | ||||
| 935 | /// Score for failing to find a decent match. | ||||
| 936 | static const int ScoreFail = 0; | ||||
| 937 | /// User exteranl to the vectorized code. | ||||
| 938 | static const int ExternalUseCost = 1; | ||||
| 939 | /// The user is internal but in a different lane. | ||||
| 940 | static const int UserInDiffLaneCost = ExternalUseCost; | ||||
| 941 | |||||
| 942 | /// \returns the score of placing \p V1 and \p V2 in consecutive lanes. | ||||
| 943 | static int getShallowScore(Value *V1, Value *V2, const DataLayout &DL, | ||||
| 944 | ScalarEvolution &SE) { | ||||
| 945 | auto *LI1 = dyn_cast<LoadInst>(V1); | ||||
| 946 | auto *LI2 = dyn_cast<LoadInst>(V2); | ||||
| 947 | if (LI1 && LI2) { | ||||
| 948 | if (LI1->getParent() != LI2->getParent()) | ||||
| 949 | return VLOperands::ScoreFail; | ||||
| 950 | |||||
| 951 | Optional<int> Dist = | ||||
| 952 | getPointersDiff(LI1->getPointerOperand(), LI2->getPointerOperand(), | ||||
| 953 | DL, SE, /*StrictCheck=*/true); | ||||
| 954 | return (Dist && *Dist == 1) ? VLOperands::ScoreConsecutiveLoads | ||||
| 955 | : VLOperands::ScoreFail; | ||||
| 956 | } | ||||
| 957 | |||||
| 958 | auto *C1 = dyn_cast<Constant>(V1); | ||||
| 959 | auto *C2 = dyn_cast<Constant>(V2); | ||||
| 960 | if (C1 && C2) | ||||
| 961 | return VLOperands::ScoreConstants; | ||||
| 962 | |||||
| 963 | // Extracts from consecutive indexes of the same vector better score as | ||||
| 964 | // the extracts could be optimized away. | ||||
| 965 | Value *EV; | ||||
| 966 | ConstantInt *Ex1Idx, *Ex2Idx; | ||||
| 967 | if (match(V1, m_ExtractElt(m_Value(EV), m_ConstantInt(Ex1Idx))) && | ||||
| 968 | match(V2, m_ExtractElt(m_Deferred(EV), m_ConstantInt(Ex2Idx))) && | ||||
| 969 | Ex1Idx->getZExtValue() + 1 == Ex2Idx->getZExtValue()) | ||||
| 970 | return VLOperands::ScoreConsecutiveExtracts; | ||||
| 971 | |||||
| 972 | auto *I1 = dyn_cast<Instruction>(V1); | ||||
| 973 | auto *I2 = dyn_cast<Instruction>(V2); | ||||
| 974 | if (I1 && I2) { | ||||
| 975 | if (I1 == I2) | ||||
| 976 | return VLOperands::ScoreSplat; | ||||
| 977 | InstructionsState S = getSameOpcode({I1, I2}); | ||||
| 978 | // Note: Only consider instructions with <= 2 operands to avoid | ||||
| 979 | // complexity explosion. | ||||
| 980 | if (S.getOpcode() && S.MainOp->getNumOperands() <= 2) | ||||
| 981 | return S.isAltShuffle() ? VLOperands::ScoreAltOpcodes | ||||
| 982 | : VLOperands::ScoreSameOpcode; | ||||
| 983 | } | ||||
| 984 | |||||
| 985 | if (isa<UndefValue>(V2)) | ||||
| 986 | return VLOperands::ScoreUndef; | ||||
| 987 | |||||
| 988 | return VLOperands::ScoreFail; | ||||
| 989 | } | ||||
| 990 | |||||
| 991 | /// Holds the values and their lane that are taking part in the look-ahead | ||||
| 992 | /// score calculation. This is used in the external uses cost calculation. | ||||
| 993 | SmallDenseMap<Value *, int> InLookAheadValues; | ||||
| 994 | |||||
| 995 | /// \Returns the additinal cost due to uses of \p LHS and \p RHS that are | ||||
| 996 | /// either external to the vectorized code, or require shuffling. | ||||
| 997 | int getExternalUsesCost(const std::pair<Value *, int> &LHS, | ||||
| 998 | const std::pair<Value *, int> &RHS) { | ||||
| 999 | int Cost = 0; | ||||
| 1000 | std::array<std::pair<Value *, int>, 2> Values = {{LHS, RHS}}; | ||||
| 1001 | for (int Idx = 0, IdxE = Values.size(); Idx != IdxE; ++Idx) { | ||||
| 1002 | Value *V = Values[Idx].first; | ||||
| 1003 | if (isa<Constant>(V)) { | ||||
| 1004 | // Since this is a function pass, it doesn't make semantic sense to | ||||
| 1005 | // walk the users of a subclass of Constant. The users could be in | ||||
| 1006 | // another function, or even another module that happens to be in | ||||
| 1007 | // the same LLVMContext. | ||||
| 1008 | continue; | ||||
| 1009 | } | ||||
| 1010 | |||||
| 1011 | // Calculate the absolute lane, using the minimum relative lane of LHS | ||||
| 1012 | // and RHS as base and Idx as the offset. | ||||
| 1013 | int Ln = std::min(LHS.second, RHS.second) + Idx; | ||||
| 1014 | assert(Ln >= 0 && "Bad lane calculation")(static_cast <bool> (Ln >= 0 && "Bad lane calculation" ) ? void (0) : __assert_fail ("Ln >= 0 && \"Bad lane calculation\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1014, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1015 | unsigned UsersBudget = LookAheadUsersBudget; | ||||
| 1016 | for (User *U : V->users()) { | ||||
| 1017 | if (const TreeEntry *UserTE = R.getTreeEntry(U)) { | ||||
| 1018 | // The user is in the VectorizableTree. Check if we need to insert. | ||||
| 1019 | auto It = llvm::find(UserTE->Scalars, U); | ||||
| 1020 | assert(It != UserTE->Scalars.end() && "U is in UserTE")(static_cast <bool> (It != UserTE->Scalars.end() && "U is in UserTE") ? void (0) : __assert_fail ("It != UserTE->Scalars.end() && \"U is in UserTE\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1020, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1021 | int UserLn = std::distance(UserTE->Scalars.begin(), It); | ||||
| 1022 | assert(UserLn >= 0 && "Bad lane")(static_cast <bool> (UserLn >= 0 && "Bad lane" ) ? void (0) : __assert_fail ("UserLn >= 0 && \"Bad lane\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1022, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1023 | if (UserLn != Ln) | ||||
| 1024 | Cost += UserInDiffLaneCost; | ||||
| 1025 | } else { | ||||
| 1026 | // Check if the user is in the look-ahead code. | ||||
| 1027 | auto It2 = InLookAheadValues.find(U); | ||||
| 1028 | if (It2 != InLookAheadValues.end()) { | ||||
| 1029 | // The user is in the look-ahead code. Check the lane. | ||||
| 1030 | if (It2->second != Ln) | ||||
| 1031 | Cost += UserInDiffLaneCost; | ||||
| 1032 | } else { | ||||
| 1033 | // The user is neither in SLP tree nor in the look-ahead code. | ||||
| 1034 | Cost += ExternalUseCost; | ||||
| 1035 | } | ||||
| 1036 | } | ||||
| 1037 | // Limit the number of visited uses to cap compilation time. | ||||
| 1038 | if (--UsersBudget == 0) | ||||
| 1039 | break; | ||||
| 1040 | } | ||||
| 1041 | } | ||||
| 1042 | return Cost; | ||||
| 1043 | } | ||||
| 1044 | |||||
| 1045 | /// Go through the operands of \p LHS and \p RHS recursively until \p | ||||
| 1046 | /// MaxLevel, and return the cummulative score. For example: | ||||
| 1047 | /// \verbatim | ||||
| 1048 | /// A[0] B[0] A[1] B[1] C[0] D[0] B[1] A[1] | ||||
| 1049 | /// \ / \ / \ / \ / | ||||
| 1050 | /// + + + + | ||||
| 1051 | /// G1 G2 G3 G4 | ||||
| 1052 | /// \endverbatim | ||||
| 1053 | /// The getScoreAtLevelRec(G1, G2) function will try to match the nodes at | ||||
| 1054 | /// each level recursively, accumulating the score. It starts from matching | ||||
| 1055 | /// the additions at level 0, then moves on to the loads (level 1). The | ||||
| 1056 | /// score of G1 and G2 is higher than G1 and G3, because {A[0],A[1]} and | ||||
| 1057 | /// {B[0],B[1]} match with VLOperands::ScoreConsecutiveLoads, while | ||||
| 1058 | /// {A[0],C[0]} has a score of VLOperands::ScoreFail. | ||||
| 1059 | /// Please note that the order of the operands does not matter, as we | ||||
| 1060 | /// evaluate the score of all profitable combinations of operands. In | ||||
| 1061 | /// other words the score of G1 and G4 is the same as G1 and G2. This | ||||
| 1062 | /// heuristic is based on ideas described in: | ||||
| 1063 | /// Look-ahead SLP: Auto-vectorization in the presence of commutative | ||||
| 1064 | /// operations, CGO 2018 by Vasileios Porpodas, Rodrigo C. O. Rocha, | ||||
| 1065 | /// LuÃs F. W. Góes | ||||
| 1066 | int getScoreAtLevelRec(const std::pair<Value *, int> &LHS, | ||||
| 1067 | const std::pair<Value *, int> &RHS, int CurrLevel, | ||||
| 1068 | int MaxLevel) { | ||||
| 1069 | |||||
| 1070 | Value *V1 = LHS.first; | ||||
| 1071 | Value *V2 = RHS.first; | ||||
| 1072 | // Get the shallow score of V1 and V2. | ||||
| 1073 | int ShallowScoreAtThisLevel = | ||||
| 1074 | std::max((int)ScoreFail, getShallowScore(V1, V2, DL, SE) - | ||||
| 1075 | getExternalUsesCost(LHS, RHS)); | ||||
| 1076 | int Lane1 = LHS.second; | ||||
| 1077 | int Lane2 = RHS.second; | ||||
| 1078 | |||||
| 1079 | // If reached MaxLevel, | ||||
| 1080 | // or if V1 and V2 are not instructions, | ||||
| 1081 | // or if they are SPLAT, | ||||
| 1082 | // or if they are not consecutive, early return the current cost. | ||||
| 1083 | auto *I1 = dyn_cast<Instruction>(V1); | ||||
| 1084 | auto *I2 = dyn_cast<Instruction>(V2); | ||||
| 1085 | if (CurrLevel == MaxLevel || !(I1 && I2) || I1 == I2 || | ||||
| 1086 | ShallowScoreAtThisLevel == VLOperands::ScoreFail || | ||||
| 1087 | (isa<LoadInst>(I1) && isa<LoadInst>(I2) && ShallowScoreAtThisLevel)) | ||||
| 1088 | return ShallowScoreAtThisLevel; | ||||
| 1089 | assert(I1 && I2 && "Should have early exited.")(static_cast <bool> (I1 && I2 && "Should have early exited." ) ? void (0) : __assert_fail ("I1 && I2 && \"Should have early exited.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1089, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1090 | |||||
| 1091 | // Keep track of in-tree values for determining the external-use cost. | ||||
| 1092 | InLookAheadValues[V1] = Lane1; | ||||
| 1093 | InLookAheadValues[V2] = Lane2; | ||||
| 1094 | |||||
| 1095 | // Contains the I2 operand indexes that got matched with I1 operands. | ||||
| 1096 | SmallSet<unsigned, 4> Op2Used; | ||||
| 1097 | |||||
| 1098 | // Recursion towards the operands of I1 and I2. We are trying all possbile | ||||
| 1099 | // operand pairs, and keeping track of the best score. | ||||
| 1100 | for (unsigned OpIdx1 = 0, NumOperands1 = I1->getNumOperands(); | ||||
| 1101 | OpIdx1 != NumOperands1; ++OpIdx1) { | ||||
| 1102 | // Try to pair op1I with the best operand of I2. | ||||
| 1103 | int MaxTmpScore = 0; | ||||
| 1104 | unsigned MaxOpIdx2 = 0; | ||||
| 1105 | bool FoundBest = false; | ||||
| 1106 | // If I2 is commutative try all combinations. | ||||
| 1107 | unsigned FromIdx = isCommutative(I2) ? 0 : OpIdx1; | ||||
| 1108 | unsigned ToIdx = isCommutative(I2) | ||||
| 1109 | ? I2->getNumOperands() | ||||
| 1110 | : std::min(I2->getNumOperands(), OpIdx1 + 1); | ||||
| 1111 | assert(FromIdx <= ToIdx && "Bad index")(static_cast <bool> (FromIdx <= ToIdx && "Bad index" ) ? void (0) : __assert_fail ("FromIdx <= ToIdx && \"Bad index\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1111, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1112 | for (unsigned OpIdx2 = FromIdx; OpIdx2 != ToIdx; ++OpIdx2) { | ||||
| 1113 | // Skip operands already paired with OpIdx1. | ||||
| 1114 | if (Op2Used.count(OpIdx2)) | ||||
| 1115 | continue; | ||||
| 1116 | // Recursively calculate the cost at each level | ||||
| 1117 | int TmpScore = getScoreAtLevelRec({I1->getOperand(OpIdx1), Lane1}, | ||||
| 1118 | {I2->getOperand(OpIdx2), Lane2}, | ||||
| 1119 | CurrLevel + 1, MaxLevel); | ||||
| 1120 | // Look for the best score. | ||||
| 1121 | if (TmpScore > VLOperands::ScoreFail && TmpScore > MaxTmpScore) { | ||||
| 1122 | MaxTmpScore = TmpScore; | ||||
| 1123 | MaxOpIdx2 = OpIdx2; | ||||
| 1124 | FoundBest = true; | ||||
| 1125 | } | ||||
| 1126 | } | ||||
| 1127 | if (FoundBest) { | ||||
| 1128 | // Pair {OpIdx1, MaxOpIdx2} was found to be best. Never revisit it. | ||||
| 1129 | Op2Used.insert(MaxOpIdx2); | ||||
| 1130 | ShallowScoreAtThisLevel += MaxTmpScore; | ||||
| 1131 | } | ||||
| 1132 | } | ||||
| 1133 | return ShallowScoreAtThisLevel; | ||||
| 1134 | } | ||||
| 1135 | |||||
| 1136 | /// \Returns the look-ahead score, which tells us how much the sub-trees | ||||
| 1137 | /// rooted at \p LHS and \p RHS match, the more they match the higher the | ||||
| 1138 | /// score. This helps break ties in an informed way when we cannot decide on | ||||
| 1139 | /// the order of the operands by just considering the immediate | ||||
| 1140 | /// predecessors. | ||||
| 1141 | int getLookAheadScore(const std::pair<Value *, int> &LHS, | ||||
| 1142 | const std::pair<Value *, int> &RHS) { | ||||
| 1143 | InLookAheadValues.clear(); | ||||
| 1144 | return getScoreAtLevelRec(LHS, RHS, 1, LookAheadMaxDepth); | ||||
| 1145 | } | ||||
| 1146 | |||||
| 1147 | // Search all operands in Ops[*][Lane] for the one that matches best | ||||
| 1148 | // Ops[OpIdx][LastLane] and return its opreand index. | ||||
| 1149 | // If no good match can be found, return None. | ||||
| 1150 | Optional<unsigned> | ||||
| 1151 | getBestOperand(unsigned OpIdx, int Lane, int LastLane, | ||||
| 1152 | ArrayRef<ReorderingMode> ReorderingModes) { | ||||
| 1153 | unsigned NumOperands = getNumOperands(); | ||||
| 1154 | |||||
| 1155 | // The operand of the previous lane at OpIdx. | ||||
| 1156 | Value *OpLastLane = getData(OpIdx, LastLane).V; | ||||
| 1157 | |||||
| 1158 | // Our strategy mode for OpIdx. | ||||
| 1159 | ReorderingMode RMode = ReorderingModes[OpIdx]; | ||||
| 1160 | |||||
| 1161 | // The linearized opcode of the operand at OpIdx, Lane. | ||||
| 1162 | bool OpIdxAPO = getData(OpIdx, Lane).APO; | ||||
| 1163 | |||||
| 1164 | // The best operand index and its score. | ||||
| 1165 | // Sometimes we have more than one option (e.g., Opcode and Undefs), so we | ||||
| 1166 | // are using the score to differentiate between the two. | ||||
| 1167 | struct BestOpData { | ||||
| 1168 | Optional<unsigned> Idx = None; | ||||
| 1169 | unsigned Score = 0; | ||||
| 1170 | } BestOp; | ||||
| 1171 | |||||
| 1172 | // Iterate through all unused operands and look for the best. | ||||
| 1173 | for (unsigned Idx = 0; Idx != NumOperands; ++Idx) { | ||||
| 1174 | // Get the operand at Idx and Lane. | ||||
| 1175 | OperandData &OpData = getData(Idx, Lane); | ||||
| 1176 | Value *Op = OpData.V; | ||||
| 1177 | bool OpAPO = OpData.APO; | ||||
| 1178 | |||||
| 1179 | // Skip already selected operands. | ||||
| 1180 | if (OpData.IsUsed) | ||||
| 1181 | continue; | ||||
| 1182 | |||||
| 1183 | // Skip if we are trying to move the operand to a position with a | ||||
| 1184 | // different opcode in the linearized tree form. This would break the | ||||
| 1185 | // semantics. | ||||
| 1186 | if (OpAPO != OpIdxAPO) | ||||
| 1187 | continue; | ||||
| 1188 | |||||
| 1189 | // Look for an operand that matches the current mode. | ||||
| 1190 | switch (RMode) { | ||||
| 1191 | case ReorderingMode::Load: | ||||
| 1192 | case ReorderingMode::Constant: | ||||
| 1193 | case ReorderingMode::Opcode: { | ||||
| 1194 | bool LeftToRight = Lane > LastLane; | ||||
| 1195 | Value *OpLeft = (LeftToRight) ? OpLastLane : Op; | ||||
| 1196 | Value *OpRight = (LeftToRight) ? Op : OpLastLane; | ||||
| 1197 | unsigned Score = | ||||
| 1198 | getLookAheadScore({OpLeft, LastLane}, {OpRight, Lane}); | ||||
| 1199 | if (Score > BestOp.Score) { | ||||
| 1200 | BestOp.Idx = Idx; | ||||
| 1201 | BestOp.Score = Score; | ||||
| 1202 | } | ||||
| 1203 | break; | ||||
| 1204 | } | ||||
| 1205 | case ReorderingMode::Splat: | ||||
| 1206 | if (Op == OpLastLane) | ||||
| 1207 | BestOp.Idx = Idx; | ||||
| 1208 | break; | ||||
| 1209 | case ReorderingMode::Failed: | ||||
| 1210 | return None; | ||||
| 1211 | } | ||||
| 1212 | } | ||||
| 1213 | |||||
| 1214 | if (BestOp.Idx) { | ||||
| 1215 | getData(BestOp.Idx.getValue(), Lane).IsUsed = true; | ||||
| 1216 | return BestOp.Idx; | ||||
| 1217 | } | ||||
| 1218 | // If we could not find a good match return None. | ||||
| 1219 | return None; | ||||
| 1220 | } | ||||
| 1221 | |||||
| 1222 | /// Helper for reorderOperandVecs. \Returns the lane that we should start | ||||
| 1223 | /// reordering from. This is the one which has the least number of operands | ||||
| 1224 | /// that can freely move about. | ||||
| 1225 | unsigned getBestLaneToStartReordering() const { | ||||
| 1226 | unsigned BestLane = 0; | ||||
| 1227 | unsigned Min = UINT_MAX(2147483647 *2U +1U); | ||||
| 1228 | for (unsigned Lane = 0, NumLanes = getNumLanes(); Lane != NumLanes; | ||||
| 1229 | ++Lane) { | ||||
| 1230 | unsigned NumFreeOps = getMaxNumOperandsThatCanBeReordered(Lane); | ||||
| 1231 | if (NumFreeOps < Min) { | ||||
| 1232 | Min = NumFreeOps; | ||||
| 1233 | BestLane = Lane; | ||||
| 1234 | } | ||||
| 1235 | } | ||||
| 1236 | return BestLane; | ||||
| 1237 | } | ||||
| 1238 | |||||
| 1239 | /// \Returns the maximum number of operands that are allowed to be reordered | ||||
| 1240 | /// for \p Lane. This is used as a heuristic for selecting the first lane to | ||||
| 1241 | /// start operand reordering. | ||||
| 1242 | unsigned getMaxNumOperandsThatCanBeReordered(unsigned Lane) const { | ||||
| 1243 | unsigned CntTrue = 0; | ||||
| 1244 | unsigned NumOperands = getNumOperands(); | ||||
| 1245 | // Operands with the same APO can be reordered. We therefore need to count | ||||
| 1246 | // how many of them we have for each APO, like this: Cnt[APO] = x. | ||||
| 1247 | // Since we only have two APOs, namely true and false, we can avoid using | ||||
| 1248 | // a map. Instead we can simply count the number of operands that | ||||
| 1249 | // correspond to one of them (in this case the 'true' APO), and calculate | ||||
| 1250 | // the other by subtracting it from the total number of operands. | ||||
| 1251 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) | ||||
| 1252 | if (getData(OpIdx, Lane).APO) | ||||
| 1253 | ++CntTrue; | ||||
| 1254 | unsigned CntFalse = NumOperands - CntTrue; | ||||
| 1255 | return std::max(CntTrue, CntFalse); | ||||
| 1256 | } | ||||
| 1257 | |||||
| 1258 | /// Go through the instructions in VL and append their operands. | ||||
| 1259 | void appendOperandsOfVL(ArrayRef<Value *> VL) { | ||||
| 1260 | assert(!VL.empty() && "Bad VL")(static_cast <bool> (!VL.empty() && "Bad VL") ? void (0) : __assert_fail ("!VL.empty() && \"Bad VL\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1260, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1261 | assert((empty() || VL.size() == getNumLanes()) &&(static_cast <bool> ((empty() || VL.size() == getNumLanes ()) && "Expected same number of lanes") ? void (0) : __assert_fail ("(empty() || VL.size() == getNumLanes()) && \"Expected same number of lanes\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1262, __extension__ __PRETTY_FUNCTION__)) | ||||
| 1262 | "Expected same number of lanes")(static_cast <bool> ((empty() || VL.size() == getNumLanes ()) && "Expected same number of lanes") ? void (0) : __assert_fail ("(empty() || VL.size() == getNumLanes()) && \"Expected same number of lanes\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1262, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1263 | assert(isa<Instruction>(VL[0]) && "Expected instruction")(static_cast <bool> (isa<Instruction>(VL[0]) && "Expected instruction") ? void (0) : __assert_fail ("isa<Instruction>(VL[0]) && \"Expected instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1263, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1264 | unsigned NumOperands = cast<Instruction>(VL[0])->getNumOperands(); | ||||
| 1265 | OpsVec.resize(NumOperands); | ||||
| 1266 | unsigned NumLanes = VL.size(); | ||||
| 1267 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { | ||||
| 1268 | OpsVec[OpIdx].resize(NumLanes); | ||||
| 1269 | for (unsigned Lane = 0; Lane != NumLanes; ++Lane) { | ||||
| 1270 | assert(isa<Instruction>(VL[Lane]) && "Expected instruction")(static_cast <bool> (isa<Instruction>(VL[Lane]) && "Expected instruction") ? void (0) : __assert_fail ("isa<Instruction>(VL[Lane]) && \"Expected instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1270, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1271 | // Our tree has just 3 nodes: the root and two operands. | ||||
| 1272 | // It is therefore trivial to get the APO. We only need to check the | ||||
| 1273 | // opcode of VL[Lane] and whether the operand at OpIdx is the LHS or | ||||
| 1274 | // RHS operand. The LHS operand of both add and sub is never attached | ||||
| 1275 | // to an inversese operation in the linearized form, therefore its APO | ||||
| 1276 | // is false. The RHS is true only if VL[Lane] is an inverse operation. | ||||
| 1277 | |||||
| 1278 | // Since operand reordering is performed on groups of commutative | ||||
| 1279 | // operations or alternating sequences (e.g., +, -), we can safely | ||||
| 1280 | // tell the inverse operations by checking commutativity. | ||||
| 1281 | bool IsInverseOperation = !isCommutative(cast<Instruction>(VL[Lane])); | ||||
| 1282 | bool APO = (OpIdx == 0) ? false : IsInverseOperation; | ||||
| 1283 | OpsVec[OpIdx][Lane] = {cast<Instruction>(VL[Lane])->getOperand(OpIdx), | ||||
| 1284 | APO, false}; | ||||
| 1285 | } | ||||
| 1286 | } | ||||
| 1287 | } | ||||
| 1288 | |||||
| 1289 | /// \returns the number of operands. | ||||
| 1290 | unsigned getNumOperands() const { return OpsVec.size(); } | ||||
| 1291 | |||||
| 1292 | /// \returns the number of lanes. | ||||
| 1293 | unsigned getNumLanes() const { return OpsVec[0].size(); } | ||||
| 1294 | |||||
| 1295 | /// \returns the operand value at \p OpIdx and \p Lane. | ||||
| 1296 | Value *getValue(unsigned OpIdx, unsigned Lane) const { | ||||
| 1297 | return getData(OpIdx, Lane).V; | ||||
| 1298 | } | ||||
| 1299 | |||||
| 1300 | /// \returns true if the data structure is empty. | ||||
| 1301 | bool empty() const { return OpsVec.empty(); } | ||||
| 1302 | |||||
| 1303 | /// Clears the data. | ||||
| 1304 | void clear() { OpsVec.clear(); } | ||||
| 1305 | |||||
| 1306 | /// \Returns true if there are enough operands identical to \p Op to fill | ||||
| 1307 | /// the whole vector. | ||||
| 1308 | /// Note: This modifies the 'IsUsed' flag, so a cleanUsed() must follow. | ||||
| 1309 | bool shouldBroadcast(Value *Op, unsigned OpIdx, unsigned Lane) { | ||||
| 1310 | bool OpAPO = getData(OpIdx, Lane).APO; | ||||
| 1311 | for (unsigned Ln = 0, Lns = getNumLanes(); Ln != Lns; ++Ln) { | ||||
| 1312 | if (Ln == Lane) | ||||
| 1313 | continue; | ||||
| 1314 | // This is set to true if we found a candidate for broadcast at Lane. | ||||
| 1315 | bool FoundCandidate = false; | ||||
| 1316 | for (unsigned OpI = 0, OpE = getNumOperands(); OpI != OpE; ++OpI) { | ||||
| 1317 | OperandData &Data = getData(OpI, Ln); | ||||
| 1318 | if (Data.APO != OpAPO || Data.IsUsed) | ||||
| 1319 | continue; | ||||
| 1320 | if (Data.V == Op) { | ||||
| 1321 | FoundCandidate = true; | ||||
| 1322 | Data.IsUsed = true; | ||||
| 1323 | break; | ||||
| 1324 | } | ||||
| 1325 | } | ||||
| 1326 | if (!FoundCandidate) | ||||
| 1327 | return false; | ||||
| 1328 | } | ||||
| 1329 | return true; | ||||
| 1330 | } | ||||
| 1331 | |||||
| 1332 | public: | ||||
| 1333 | /// Initialize with all the operands of the instruction vector \p RootVL. | ||||
| 1334 | VLOperands(ArrayRef<Value *> RootVL, const DataLayout &DL, | ||||
| 1335 | ScalarEvolution &SE, const BoUpSLP &R) | ||||
| 1336 | : DL(DL), SE(SE), R(R) { | ||||
| 1337 | // Append all the operands of RootVL. | ||||
| 1338 | appendOperandsOfVL(RootVL); | ||||
| 1339 | } | ||||
| 1340 | |||||
| 1341 | /// \Returns a value vector with the operands across all lanes for the | ||||
| 1342 | /// opearnd at \p OpIdx. | ||||
| 1343 | ValueList getVL(unsigned OpIdx) const { | ||||
| 1344 | ValueList OpVL(OpsVec[OpIdx].size()); | ||||
| 1345 | assert(OpsVec[OpIdx].size() == getNumLanes() &&(static_cast <bool> (OpsVec[OpIdx].size() == getNumLanes () && "Expected same num of lanes across all operands" ) ? void (0) : __assert_fail ("OpsVec[OpIdx].size() == getNumLanes() && \"Expected same num of lanes across all operands\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1346, __extension__ __PRETTY_FUNCTION__)) | ||||
| 1346 | "Expected same num of lanes across all operands")(static_cast <bool> (OpsVec[OpIdx].size() == getNumLanes () && "Expected same num of lanes across all operands" ) ? void (0) : __assert_fail ("OpsVec[OpIdx].size() == getNumLanes() && \"Expected same num of lanes across all operands\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1346, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1347 | for (unsigned Lane = 0, Lanes = getNumLanes(); Lane != Lanes; ++Lane) | ||||
| 1348 | OpVL[Lane] = OpsVec[OpIdx][Lane].V; | ||||
| 1349 | return OpVL; | ||||
| 1350 | } | ||||
| 1351 | |||||
| 1352 | // Performs operand reordering for 2 or more operands. | ||||
| 1353 | // The original operands are in OrigOps[OpIdx][Lane]. | ||||
| 1354 | // The reordered operands are returned in 'SortedOps[OpIdx][Lane]'. | ||||
| 1355 | void reorder() { | ||||
| 1356 | unsigned NumOperands = getNumOperands(); | ||||
| 1357 | unsigned NumLanes = getNumLanes(); | ||||
| 1358 | // Each operand has its own mode. We are using this mode to help us select | ||||
| 1359 | // the instructions for each lane, so that they match best with the ones | ||||
| 1360 | // we have selected so far. | ||||
| 1361 | SmallVector<ReorderingMode, 2> ReorderingModes(NumOperands); | ||||
| 1362 | |||||
| 1363 | // This is a greedy single-pass algorithm. We are going over each lane | ||||
| 1364 | // once and deciding on the best order right away with no back-tracking. | ||||
| 1365 | // However, in order to increase its effectiveness, we start with the lane | ||||
| 1366 | // that has operands that can move the least. For example, given the | ||||
| 1367 | // following lanes: | ||||
| 1368 | // Lane 0 : A[0] = B[0] + C[0] // Visited 3rd | ||||
| 1369 | // Lane 1 : A[1] = C[1] - B[1] // Visited 1st | ||||
| 1370 | // Lane 2 : A[2] = B[2] + C[2] // Visited 2nd | ||||
| 1371 | // Lane 3 : A[3] = C[3] - B[3] // Visited 4th | ||||
| 1372 | // we will start at Lane 1, since the operands of the subtraction cannot | ||||
| 1373 | // be reordered. Then we will visit the rest of the lanes in a circular | ||||
| 1374 | // fashion. That is, Lanes 2, then Lane 0, and finally Lane 3. | ||||
| 1375 | |||||
| 1376 | // Find the first lane that we will start our search from. | ||||
| 1377 | unsigned FirstLane = getBestLaneToStartReordering(); | ||||
| 1378 | |||||
| 1379 | // Initialize the modes. | ||||
| 1380 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { | ||||
| 1381 | Value *OpLane0 = getValue(OpIdx, FirstLane); | ||||
| 1382 | // Keep track if we have instructions with all the same opcode on one | ||||
| 1383 | // side. | ||||
| 1384 | if (isa<LoadInst>(OpLane0)) | ||||
| 1385 | ReorderingModes[OpIdx] = ReorderingMode::Load; | ||||
| 1386 | else if (isa<Instruction>(OpLane0)) { | ||||
| 1387 | // Check if OpLane0 should be broadcast. | ||||
| 1388 | if (shouldBroadcast(OpLane0, OpIdx, FirstLane)) | ||||
| 1389 | ReorderingModes[OpIdx] = ReorderingMode::Splat; | ||||
| 1390 | else | ||||
| 1391 | ReorderingModes[OpIdx] = ReorderingMode::Opcode; | ||||
| 1392 | } | ||||
| 1393 | else if (isa<Constant>(OpLane0)) | ||||
| 1394 | ReorderingModes[OpIdx] = ReorderingMode::Constant; | ||||
| 1395 | else if (isa<Argument>(OpLane0)) | ||||
| 1396 | // Our best hope is a Splat. It may save some cost in some cases. | ||||
| 1397 | ReorderingModes[OpIdx] = ReorderingMode::Splat; | ||||
| 1398 | else | ||||
| 1399 | // NOTE: This should be unreachable. | ||||
| 1400 | ReorderingModes[OpIdx] = ReorderingMode::Failed; | ||||
| 1401 | } | ||||
| 1402 | |||||
| 1403 | // If the initial strategy fails for any of the operand indexes, then we | ||||
| 1404 | // perform reordering again in a second pass. This helps avoid assigning | ||||
| 1405 | // high priority to the failed strategy, and should improve reordering for | ||||
| 1406 | // the non-failed operand indexes. | ||||
| 1407 | for (int Pass = 0; Pass != 2; ++Pass) { | ||||
| 1408 | // Skip the second pass if the first pass did not fail. | ||||
| 1409 | bool StrategyFailed = false; | ||||
| 1410 | // Mark all operand data as free to use. | ||||
| 1411 | clearUsed(); | ||||
| 1412 | // We keep the original operand order for the FirstLane, so reorder the | ||||
| 1413 | // rest of the lanes. We are visiting the nodes in a circular fashion, | ||||
| 1414 | // using FirstLane as the center point and increasing the radius | ||||
| 1415 | // distance. | ||||
| 1416 | for (unsigned Distance = 1; Distance != NumLanes; ++Distance) { | ||||
| 1417 | // Visit the lane on the right and then the lane on the left. | ||||
| 1418 | for (int Direction : {+1, -1}) { | ||||
| 1419 | int Lane = FirstLane + Direction * Distance; | ||||
| 1420 | if (Lane < 0 || Lane >= (int)NumLanes) | ||||
| 1421 | continue; | ||||
| 1422 | int LastLane = Lane - Direction; | ||||
| 1423 | assert(LastLane >= 0 && LastLane < (int)NumLanes &&(static_cast <bool> (LastLane >= 0 && LastLane < (int)NumLanes && "Out of bounds") ? void (0) : __assert_fail ("LastLane >= 0 && LastLane < (int)NumLanes && \"Out of bounds\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1424, __extension__ __PRETTY_FUNCTION__)) | ||||
| 1424 | "Out of bounds")(static_cast <bool> (LastLane >= 0 && LastLane < (int)NumLanes && "Out of bounds") ? void (0) : __assert_fail ("LastLane >= 0 && LastLane < (int)NumLanes && \"Out of bounds\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1424, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1425 | // Look for a good match for each operand. | ||||
| 1426 | for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) { | ||||
| 1427 | // Search for the operand that matches SortedOps[OpIdx][Lane-1]. | ||||
| 1428 | Optional<unsigned> BestIdx = | ||||
| 1429 | getBestOperand(OpIdx, Lane, LastLane, ReorderingModes); | ||||
| 1430 | // By not selecting a value, we allow the operands that follow to | ||||
| 1431 | // select a better matching value. We will get a non-null value in | ||||
| 1432 | // the next run of getBestOperand(). | ||||
| 1433 | if (BestIdx) { | ||||
| 1434 | // Swap the current operand with the one returned by | ||||
| 1435 | // getBestOperand(). | ||||
| 1436 | swap(OpIdx, BestIdx.getValue(), Lane); | ||||
| 1437 | } else { | ||||
| 1438 | // We failed to find a best operand, set mode to 'Failed'. | ||||
| 1439 | ReorderingModes[OpIdx] = ReorderingMode::Failed; | ||||
| 1440 | // Enable the second pass. | ||||
| 1441 | StrategyFailed = true; | ||||
| 1442 | } | ||||
| 1443 | } | ||||
| 1444 | } | ||||
| 1445 | } | ||||
| 1446 | // Skip second pass if the strategy did not fail. | ||||
| 1447 | if (!StrategyFailed) | ||||
| 1448 | break; | ||||
| 1449 | } | ||||
| 1450 | } | ||||
| 1451 | |||||
| 1452 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) | ||||
| 1453 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static StringRef getModeStr(ReorderingMode RMode) { | ||||
| 1454 | switch (RMode) { | ||||
| 1455 | case ReorderingMode::Load: | ||||
| 1456 | return "Load"; | ||||
| 1457 | case ReorderingMode::Opcode: | ||||
| 1458 | return "Opcode"; | ||||
| 1459 | case ReorderingMode::Constant: | ||||
| 1460 | return "Constant"; | ||||
| 1461 | case ReorderingMode::Splat: | ||||
| 1462 | return "Splat"; | ||||
| 1463 | case ReorderingMode::Failed: | ||||
| 1464 | return "Failed"; | ||||
| 1465 | } | ||||
| 1466 | llvm_unreachable("Unimplemented Reordering Type")::llvm::llvm_unreachable_internal("Unimplemented Reordering Type" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1466); | ||||
| 1467 | } | ||||
| 1468 | |||||
| 1469 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static raw_ostream &printMode(ReorderingMode RMode, | ||||
| 1470 | raw_ostream &OS) { | ||||
| 1471 | return OS << getModeStr(RMode); | ||||
| 1472 | } | ||||
| 1473 | |||||
| 1474 | /// Debug print. | ||||
| 1475 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) static void dumpMode(ReorderingMode RMode) { | ||||
| 1476 | printMode(RMode, dbgs()); | ||||
| 1477 | } | ||||
| 1478 | |||||
| 1479 | friend raw_ostream &operator<<(raw_ostream &OS, ReorderingMode RMode) { | ||||
| 1480 | return printMode(RMode, OS); | ||||
| 1481 | } | ||||
| 1482 | |||||
| 1483 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) raw_ostream &print(raw_ostream &OS) const { | ||||
| 1484 | const unsigned Indent = 2; | ||||
| 1485 | unsigned Cnt = 0; | ||||
| 1486 | for (const OperandDataVec &OpDataVec : OpsVec) { | ||||
| 1487 | OS << "Operand " << Cnt++ << "\n"; | ||||
| 1488 | for (const OperandData &OpData : OpDataVec) { | ||||
| 1489 | OS.indent(Indent) << "{"; | ||||
| 1490 | if (Value *V = OpData.V) | ||||
| 1491 | OS << *V; | ||||
| 1492 | else | ||||
| 1493 | OS << "null"; | ||||
| 1494 | OS << ", APO:" << OpData.APO << "}\n"; | ||||
| 1495 | } | ||||
| 1496 | OS << "\n"; | ||||
| 1497 | } | ||||
| 1498 | return OS; | ||||
| 1499 | } | ||||
| 1500 | |||||
| 1501 | /// Debug print. | ||||
| 1502 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { print(dbgs()); } | ||||
| 1503 | #endif | ||||
| 1504 | }; | ||||
| 1505 | |||||
| 1506 | /// Checks if the instruction is marked for deletion. | ||||
| 1507 | bool isDeleted(Instruction *I) const { return DeletedInstructions.count(I); } | ||||
| 1508 | |||||
| 1509 | /// Marks values operands for later deletion by replacing them with Undefs. | ||||
| 1510 | void eraseInstructions(ArrayRef<Value *> AV); | ||||
| 1511 | |||||
| 1512 | ~BoUpSLP(); | ||||
| 1513 | |||||
| 1514 | private: | ||||
| 1515 | /// Checks if all users of \p I are the part of the vectorization tree. | ||||
| 1516 | bool areAllUsersVectorized(Instruction *I) const; | ||||
| 1517 | |||||
| 1518 | /// \returns the cost of the vectorizable entry. | ||||
| 1519 | InstructionCost getEntryCost(const TreeEntry *E); | ||||
| 1520 | |||||
| 1521 | /// This is the recursive part of buildTree. | ||||
| 1522 | void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth, | ||||
| 1523 | const EdgeInfo &EI); | ||||
| 1524 | |||||
| 1525 | /// \returns true if the ExtractElement/ExtractValue instructions in \p VL can | ||||
| 1526 | /// be vectorized to use the original vector (or aggregate "bitcast" to a | ||||
| 1527 | /// vector) and sets \p CurrentOrder to the identity permutation; otherwise | ||||
| 1528 | /// returns false, setting \p CurrentOrder to either an empty vector or a | ||||
| 1529 | /// non-identity permutation that allows to reuse extract instructions. | ||||
| 1530 | bool canReuseExtract(ArrayRef<Value *> VL, Value *OpValue, | ||||
| 1531 | SmallVectorImpl<unsigned> &CurrentOrder) const; | ||||
| 1532 | |||||
| 1533 | /// Vectorize a single entry in the tree. | ||||
| 1534 | Value *vectorizeTree(TreeEntry *E); | ||||
| 1535 | |||||
| 1536 | /// Vectorize a single entry in the tree, starting in \p VL. | ||||
| 1537 | Value *vectorizeTree(ArrayRef<Value *> VL); | ||||
| 1538 | |||||
| 1539 | /// \returns the scalarization cost for this type. Scalarization in this | ||||
| 1540 | /// context means the creation of vectors from a group of scalars. | ||||
| 1541 | InstructionCost | ||||
| 1542 | getGatherCost(FixedVectorType *Ty, | ||||
| 1543 | const DenseSet<unsigned> &ShuffledIndices) const; | ||||
| 1544 | |||||
| 1545 | /// Checks if the gathered \p VL can be represented as shuffle(s) of previous | ||||
| 1546 | /// tree entries. | ||||
| 1547 | /// \returns ShuffleKind, if gathered values can be represented as shuffles of | ||||
| 1548 | /// previous tree entries. \p Mask is filled with the shuffle mask. | ||||
| 1549 | Optional<TargetTransformInfo::ShuffleKind> | ||||
| 1550 | isGatherShuffledEntry(const TreeEntry *TE, SmallVectorImpl<int> &Mask, | ||||
| 1551 | SmallVectorImpl<const TreeEntry *> &Entries); | ||||
| 1552 | |||||
| 1553 | /// \returns the scalarization cost for this list of values. Assuming that | ||||
| 1554 | /// this subtree gets vectorized, we may need to extract the values from the | ||||
| 1555 | /// roots. This method calculates the cost of extracting the values. | ||||
| 1556 | InstructionCost getGatherCost(ArrayRef<Value *> VL) const; | ||||
| 1557 | |||||
| 1558 | /// Set the Builder insert point to one after the last instruction in | ||||
| 1559 | /// the bundle | ||||
| 1560 | void setInsertPointAfterBundle(const TreeEntry *E); | ||||
| 1561 | |||||
| 1562 | /// \returns a vector from a collection of scalars in \p VL. | ||||
| 1563 | Value *gather(ArrayRef<Value *> VL); | ||||
| 1564 | |||||
| 1565 | /// \returns whether the VectorizableTree is fully vectorizable and will | ||||
| 1566 | /// be beneficial even the tree height is tiny. | ||||
| 1567 | bool isFullyVectorizableTinyTree() const; | ||||
| 1568 | |||||
| 1569 | /// Reorder commutative or alt operands to get better probability of | ||||
| 1570 | /// generating vectorized code. | ||||
| 1571 | static void reorderInputsAccordingToOpcode(ArrayRef<Value *> VL, | ||||
| 1572 | SmallVectorImpl<Value *> &Left, | ||||
| 1573 | SmallVectorImpl<Value *> &Right, | ||||
| 1574 | const DataLayout &DL, | ||||
| 1575 | ScalarEvolution &SE, | ||||
| 1576 | const BoUpSLP &R); | ||||
| 1577 | struct TreeEntry { | ||||
| 1578 | using VecTreeTy = SmallVector<std::unique_ptr<TreeEntry>, 8>; | ||||
| 1579 | TreeEntry(VecTreeTy &Container) : Container(Container) {} | ||||
| 1580 | |||||
| 1581 | /// \returns true if the scalars in VL are equal to this entry. | ||||
| 1582 | bool isSame(ArrayRef<Value *> VL) const { | ||||
| 1583 | if (VL.size() == Scalars.size()) | ||||
| 1584 | return std::equal(VL.begin(), VL.end(), Scalars.begin()); | ||||
| 1585 | return VL.size() == ReuseShuffleIndices.size() && | ||||
| 1586 | std::equal( | ||||
| 1587 | VL.begin(), VL.end(), ReuseShuffleIndices.begin(), | ||||
| 1588 | [this](Value *V, int Idx) { return V == Scalars[Idx]; }); | ||||
| 1589 | } | ||||
| 1590 | |||||
| 1591 | /// A vector of scalars. | ||||
| 1592 | ValueList Scalars; | ||||
| 1593 | |||||
| 1594 | /// The Scalars are vectorized into this value. It is initialized to Null. | ||||
| 1595 | Value *VectorizedValue = nullptr; | ||||
| 1596 | |||||
| 1597 | /// Do we need to gather this sequence or vectorize it | ||||
| 1598 | /// (either with vector instruction or with scatter/gather | ||||
| 1599 | /// intrinsics for store/load)? | ||||
| 1600 | enum EntryState { Vectorize, ScatterVectorize, NeedToGather }; | ||||
| 1601 | EntryState State; | ||||
| 1602 | |||||
| 1603 | /// Does this sequence require some shuffling? | ||||
| 1604 | SmallVector<int, 4> ReuseShuffleIndices; | ||||
| 1605 | |||||
| 1606 | /// Does this entry require reordering? | ||||
| 1607 | SmallVector<unsigned, 4> ReorderIndices; | ||||
| 1608 | |||||
| 1609 | /// Points back to the VectorizableTree. | ||||
| 1610 | /// | ||||
| 1611 | /// Only used for Graphviz right now. Unfortunately GraphTrait::NodeRef has | ||||
| 1612 | /// to be a pointer and needs to be able to initialize the child iterator. | ||||
| 1613 | /// Thus we need a reference back to the container to translate the indices | ||||
| 1614 | /// to entries. | ||||
| 1615 | VecTreeTy &Container; | ||||
| 1616 | |||||
| 1617 | /// The TreeEntry index containing the user of this entry. We can actually | ||||
| 1618 | /// have multiple users so the data structure is not truly a tree. | ||||
| 1619 | SmallVector<EdgeInfo, 1> UserTreeIndices; | ||||
| 1620 | |||||
| 1621 | /// The index of this treeEntry in VectorizableTree. | ||||
| 1622 | int Idx = -1; | ||||
| 1623 | |||||
| 1624 | private: | ||||
| 1625 | /// The operands of each instruction in each lane Operands[op_index][lane]. | ||||
| 1626 | /// Note: This helps avoid the replication of the code that performs the | ||||
| 1627 | /// reordering of operands during buildTree_rec() and vectorizeTree(). | ||||
| 1628 | SmallVector<ValueList, 2> Operands; | ||||
| 1629 | |||||
| 1630 | /// The main/alternate instruction. | ||||
| 1631 | Instruction *MainOp = nullptr; | ||||
| 1632 | Instruction *AltOp = nullptr; | ||||
| 1633 | |||||
| 1634 | public: | ||||
| 1635 | /// Set this bundle's \p OpIdx'th operand to \p OpVL. | ||||
| 1636 | void setOperand(unsigned OpIdx, ArrayRef<Value *> OpVL) { | ||||
| 1637 | if (Operands.size() < OpIdx + 1) | ||||
| 1638 | Operands.resize(OpIdx + 1); | ||||
| 1639 | assert(Operands[OpIdx].empty() && "Already resized?")(static_cast <bool> (Operands[OpIdx].empty() && "Already resized?") ? void (0) : __assert_fail ("Operands[OpIdx].empty() && \"Already resized?\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1639, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1640 | Operands[OpIdx].resize(Scalars.size()); | ||||
| 1641 | for (unsigned Lane = 0, E = Scalars.size(); Lane != E; ++Lane) | ||||
| 1642 | Operands[OpIdx][Lane] = OpVL[Lane]; | ||||
| 1643 | } | ||||
| 1644 | |||||
| 1645 | /// Set the operands of this bundle in their original order. | ||||
| 1646 | void setOperandsInOrder() { | ||||
| 1647 | assert(Operands.empty() && "Already initialized?")(static_cast <bool> (Operands.empty() && "Already initialized?" ) ? void (0) : __assert_fail ("Operands.empty() && \"Already initialized?\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1647, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1648 | auto *I0 = cast<Instruction>(Scalars[0]); | ||||
| 1649 | Operands.resize(I0->getNumOperands()); | ||||
| 1650 | unsigned NumLanes = Scalars.size(); | ||||
| 1651 | for (unsigned OpIdx = 0, NumOperands = I0->getNumOperands(); | ||||
| 1652 | OpIdx != NumOperands; ++OpIdx) { | ||||
| 1653 | Operands[OpIdx].resize(NumLanes); | ||||
| 1654 | for (unsigned Lane = 0; Lane != NumLanes; ++Lane) { | ||||
| 1655 | auto *I = cast<Instruction>(Scalars[Lane]); | ||||
| 1656 | assert(I->getNumOperands() == NumOperands &&(static_cast <bool> (I->getNumOperands() == NumOperands && "Expected same number of operands") ? void (0) : __assert_fail ("I->getNumOperands() == NumOperands && \"Expected same number of operands\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1657, __extension__ __PRETTY_FUNCTION__)) | ||||
| 1657 | "Expected same number of operands")(static_cast <bool> (I->getNumOperands() == NumOperands && "Expected same number of operands") ? void (0) : __assert_fail ("I->getNumOperands() == NumOperands && \"Expected same number of operands\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1657, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1658 | Operands[OpIdx][Lane] = I->getOperand(OpIdx); | ||||
| 1659 | } | ||||
| 1660 | } | ||||
| 1661 | } | ||||
| 1662 | |||||
| 1663 | /// \returns the \p OpIdx operand of this TreeEntry. | ||||
| 1664 | ValueList &getOperand(unsigned OpIdx) { | ||||
| 1665 | assert(OpIdx < Operands.size() && "Off bounds")(static_cast <bool> (OpIdx < Operands.size() && "Off bounds") ? void (0) : __assert_fail ("OpIdx < Operands.size() && \"Off bounds\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1665, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1666 | return Operands[OpIdx]; | ||||
| 1667 | } | ||||
| 1668 | |||||
| 1669 | /// \returns the number of operands. | ||||
| 1670 | unsigned getNumOperands() const { return Operands.size(); } | ||||
| 1671 | |||||
| 1672 | /// \return the single \p OpIdx operand. | ||||
| 1673 | Value *getSingleOperand(unsigned OpIdx) const { | ||||
| 1674 | assert(OpIdx < Operands.size() && "Off bounds")(static_cast <bool> (OpIdx < Operands.size() && "Off bounds") ? void (0) : __assert_fail ("OpIdx < Operands.size() && \"Off bounds\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1674, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1675 | assert(!Operands[OpIdx].empty() && "No operand available")(static_cast <bool> (!Operands[OpIdx].empty() && "No operand available") ? void (0) : __assert_fail ("!Operands[OpIdx].empty() && \"No operand available\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1675, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1676 | return Operands[OpIdx][0]; | ||||
| 1677 | } | ||||
| 1678 | |||||
| 1679 | /// Some of the instructions in the list have alternate opcodes. | ||||
| 1680 | bool isAltShuffle() const { | ||||
| 1681 | return getOpcode() != getAltOpcode(); | ||||
| 1682 | } | ||||
| 1683 | |||||
| 1684 | bool isOpcodeOrAlt(Instruction *I) const { | ||||
| 1685 | unsigned CheckedOpcode = I->getOpcode(); | ||||
| 1686 | return (getOpcode() == CheckedOpcode || | ||||
| 1687 | getAltOpcode() == CheckedOpcode); | ||||
| 1688 | } | ||||
| 1689 | |||||
| 1690 | /// Chooses the correct key for scheduling data. If \p Op has the same (or | ||||
| 1691 | /// alternate) opcode as \p OpValue, the key is \p Op. Otherwise the key is | ||||
| 1692 | /// \p OpValue. | ||||
| 1693 | Value *isOneOf(Value *Op) const { | ||||
| 1694 | auto *I = dyn_cast<Instruction>(Op); | ||||
| 1695 | if (I && isOpcodeOrAlt(I)) | ||||
| 1696 | return Op; | ||||
| 1697 | return MainOp; | ||||
| 1698 | } | ||||
| 1699 | |||||
| 1700 | void setOperations(const InstructionsState &S) { | ||||
| 1701 | MainOp = S.MainOp; | ||||
| 1702 | AltOp = S.AltOp; | ||||
| 1703 | } | ||||
| 1704 | |||||
| 1705 | Instruction *getMainOp() const { | ||||
| 1706 | return MainOp; | ||||
| 1707 | } | ||||
| 1708 | |||||
| 1709 | Instruction *getAltOp() const { | ||||
| 1710 | return AltOp; | ||||
| 1711 | } | ||||
| 1712 | |||||
| 1713 | /// The main/alternate opcodes for the list of instructions. | ||||
| 1714 | unsigned getOpcode() const { | ||||
| 1715 | return MainOp ? MainOp->getOpcode() : 0; | ||||
| 1716 | } | ||||
| 1717 | |||||
| 1718 | unsigned getAltOpcode() const { | ||||
| 1719 | return AltOp ? AltOp->getOpcode() : 0; | ||||
| 1720 | } | ||||
| 1721 | |||||
| 1722 | /// Update operations state of this entry if reorder occurred. | ||||
| 1723 | bool updateStateIfReorder() { | ||||
| 1724 | if (ReorderIndices.empty()) | ||||
| 1725 | return false; | ||||
| 1726 | InstructionsState S = getSameOpcode(Scalars, ReorderIndices.front()); | ||||
| 1727 | setOperations(S); | ||||
| 1728 | return true; | ||||
| 1729 | } | ||||
| 1730 | |||||
| 1731 | #ifndef NDEBUG | ||||
| 1732 | /// Debug printer. | ||||
| 1733 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dump() const { | ||||
| 1734 | dbgs() << Idx << ".\n"; | ||||
| 1735 | for (unsigned OpI = 0, OpE = Operands.size(); OpI != OpE; ++OpI) { | ||||
| 1736 | dbgs() << "Operand " << OpI << ":\n"; | ||||
| 1737 | for (const Value *V : Operands[OpI]) | ||||
| 1738 | dbgs().indent(2) << *V << "\n"; | ||||
| 1739 | } | ||||
| 1740 | dbgs() << "Scalars: \n"; | ||||
| 1741 | for (Value *V : Scalars) | ||||
| 1742 | dbgs().indent(2) << *V << "\n"; | ||||
| 1743 | dbgs() << "State: "; | ||||
| 1744 | switch (State) { | ||||
| 1745 | case Vectorize: | ||||
| 1746 | dbgs() << "Vectorize\n"; | ||||
| 1747 | break; | ||||
| 1748 | case ScatterVectorize: | ||||
| 1749 | dbgs() << "ScatterVectorize\n"; | ||||
| 1750 | break; | ||||
| 1751 | case NeedToGather: | ||||
| 1752 | dbgs() << "NeedToGather\n"; | ||||
| 1753 | break; | ||||
| 1754 | } | ||||
| 1755 | dbgs() << "MainOp: "; | ||||
| 1756 | if (MainOp) | ||||
| 1757 | dbgs() << *MainOp << "\n"; | ||||
| 1758 | else | ||||
| 1759 | dbgs() << "NULL\n"; | ||||
| 1760 | dbgs() << "AltOp: "; | ||||
| 1761 | if (AltOp) | ||||
| 1762 | dbgs() << *AltOp << "\n"; | ||||
| 1763 | else | ||||
| 1764 | dbgs() << "NULL\n"; | ||||
| 1765 | dbgs() << "VectorizedValue: "; | ||||
| 1766 | if (VectorizedValue) | ||||
| 1767 | dbgs() << *VectorizedValue << "\n"; | ||||
| 1768 | else | ||||
| 1769 | dbgs() << "NULL\n"; | ||||
| 1770 | dbgs() << "ReuseShuffleIndices: "; | ||||
| 1771 | if (ReuseShuffleIndices.empty()) | ||||
| 1772 | dbgs() << "Empty"; | ||||
| 1773 | else | ||||
| 1774 | for (unsigned ReuseIdx : ReuseShuffleIndices) | ||||
| 1775 | dbgs() << ReuseIdx << ", "; | ||||
| 1776 | dbgs() << "\n"; | ||||
| 1777 | dbgs() << "ReorderIndices: "; | ||||
| 1778 | for (unsigned ReorderIdx : ReorderIndices) | ||||
| 1779 | dbgs() << ReorderIdx << ", "; | ||||
| 1780 | dbgs() << "\n"; | ||||
| 1781 | dbgs() << "UserTreeIndices: "; | ||||
| 1782 | for (const auto &EInfo : UserTreeIndices) | ||||
| 1783 | dbgs() << EInfo << ", "; | ||||
| 1784 | dbgs() << "\n"; | ||||
| 1785 | } | ||||
| 1786 | #endif | ||||
| 1787 | }; | ||||
| 1788 | |||||
| 1789 | #ifndef NDEBUG | ||||
| 1790 | void dumpTreeCosts(const TreeEntry *E, InstructionCost ReuseShuffleCost, | ||||
| 1791 | InstructionCost VecCost, | ||||
| 1792 | InstructionCost ScalarCost) const { | ||||
| 1793 | dbgs() << "SLP: Calculated costs for Tree:\n"; E->dump(); | ||||
| 1794 | dbgs() << "SLP: Costs:\n"; | ||||
| 1795 | dbgs() << "SLP: ReuseShuffleCost = " << ReuseShuffleCost << "\n"; | ||||
| 1796 | dbgs() << "SLP: VectorCost = " << VecCost << "\n"; | ||||
| 1797 | dbgs() << "SLP: ScalarCost = " << ScalarCost << "\n"; | ||||
| 1798 | dbgs() << "SLP: ReuseShuffleCost + VecCost - ScalarCost = " << | ||||
| 1799 | ReuseShuffleCost + VecCost - ScalarCost << "\n"; | ||||
| 1800 | } | ||||
| 1801 | #endif | ||||
| 1802 | |||||
| 1803 | /// Create a new VectorizableTree entry. | ||||
| 1804 | TreeEntry *newTreeEntry(ArrayRef<Value *> VL, Optional<ScheduleData *> Bundle, | ||||
| 1805 | const InstructionsState &S, | ||||
| 1806 | const EdgeInfo &UserTreeIdx, | ||||
| 1807 | ArrayRef<unsigned> ReuseShuffleIndices = None, | ||||
| 1808 | ArrayRef<unsigned> ReorderIndices = None) { | ||||
| 1809 | TreeEntry::EntryState EntryState = | ||||
| 1810 | Bundle ? TreeEntry::Vectorize : TreeEntry::NeedToGather; | ||||
| 1811 | return newTreeEntry(VL, EntryState, Bundle, S, UserTreeIdx, | ||||
| 1812 | ReuseShuffleIndices, ReorderIndices); | ||||
| 1813 | } | ||||
| 1814 | |||||
| 1815 | TreeEntry *newTreeEntry(ArrayRef<Value *> VL, | ||||
| 1816 | TreeEntry::EntryState EntryState, | ||||
| 1817 | Optional<ScheduleData *> Bundle, | ||||
| 1818 | const InstructionsState &S, | ||||
| 1819 | const EdgeInfo &UserTreeIdx, | ||||
| 1820 | ArrayRef<unsigned> ReuseShuffleIndices = None, | ||||
| 1821 | ArrayRef<unsigned> ReorderIndices = None) { | ||||
| 1822 | assert(((!Bundle && EntryState == TreeEntry::NeedToGather) ||(static_cast <bool> (((!Bundle && EntryState == TreeEntry::NeedToGather) || (Bundle && EntryState != TreeEntry::NeedToGather)) && "Need to vectorize gather entry?" ) ? void (0) : __assert_fail ("((!Bundle && EntryState == TreeEntry::NeedToGather) || (Bundle && EntryState != TreeEntry::NeedToGather)) && \"Need to vectorize gather entry?\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1824, __extension__ __PRETTY_FUNCTION__)) | ||||
| 1823 | (Bundle && EntryState != TreeEntry::NeedToGather)) &&(static_cast <bool> (((!Bundle && EntryState == TreeEntry::NeedToGather) || (Bundle && EntryState != TreeEntry::NeedToGather)) && "Need to vectorize gather entry?" ) ? void (0) : __assert_fail ("((!Bundle && EntryState == TreeEntry::NeedToGather) || (Bundle && EntryState != TreeEntry::NeedToGather)) && \"Need to vectorize gather entry?\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1824, __extension__ __PRETTY_FUNCTION__)) | ||||
| 1824 | "Need to vectorize gather entry?")(static_cast <bool> (((!Bundle && EntryState == TreeEntry::NeedToGather) || (Bundle && EntryState != TreeEntry::NeedToGather)) && "Need to vectorize gather entry?" ) ? void (0) : __assert_fail ("((!Bundle && EntryState == TreeEntry::NeedToGather) || (Bundle && EntryState != TreeEntry::NeedToGather)) && \"Need to vectorize gather entry?\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1824, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1825 | VectorizableTree.push_back(std::make_unique<TreeEntry>(VectorizableTree)); | ||||
| 1826 | TreeEntry *Last = VectorizableTree.back().get(); | ||||
| 1827 | Last->Idx = VectorizableTree.size() - 1; | ||||
| 1828 | Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end()); | ||||
| 1829 | Last->State = EntryState; | ||||
| 1830 | Last->ReuseShuffleIndices.append(ReuseShuffleIndices.begin(), | ||||
| 1831 | ReuseShuffleIndices.end()); | ||||
| 1832 | Last->ReorderIndices.append(ReorderIndices.begin(), ReorderIndices.end()); | ||||
| 1833 | Last->setOperations(S); | ||||
| 1834 | if (Last->State != TreeEntry::NeedToGather) { | ||||
| 1835 | for (Value *V : VL) { | ||||
| 1836 | assert(!getTreeEntry(V) && "Scalar already in tree!")(static_cast <bool> (!getTreeEntry(V) && "Scalar already in tree!" ) ? void (0) : __assert_fail ("!getTreeEntry(V) && \"Scalar already in tree!\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1836, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1837 | ScalarToTreeEntry[V] = Last; | ||||
| 1838 | } | ||||
| 1839 | // Update the scheduler bundle to point to this TreeEntry. | ||||
| 1840 | unsigned Lane = 0; | ||||
| 1841 | for (ScheduleData *BundleMember = Bundle.getValue(); BundleMember; | ||||
| 1842 | BundleMember = BundleMember->NextInBundle) { | ||||
| 1843 | BundleMember->TE = Last; | ||||
| 1844 | BundleMember->Lane = Lane; | ||||
| 1845 | ++Lane; | ||||
| 1846 | } | ||||
| 1847 | assert((!Bundle.getValue() || Lane == VL.size()) &&(static_cast <bool> ((!Bundle.getValue() || Lane == VL. size()) && "Bundle and VL out of sync") ? void (0) : __assert_fail ("(!Bundle.getValue() || Lane == VL.size()) && \"Bundle and VL out of sync\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1848, __extension__ __PRETTY_FUNCTION__)) | ||||
| 1848 | "Bundle and VL out of sync")(static_cast <bool> ((!Bundle.getValue() || Lane == VL. size()) && "Bundle and VL out of sync") ? void (0) : __assert_fail ("(!Bundle.getValue() || Lane == VL.size()) && \"Bundle and VL out of sync\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 1848, __extension__ __PRETTY_FUNCTION__)); | ||||
| 1849 | } else { | ||||
| 1850 | MustGather.insert(VL.begin(), VL.end()); | ||||
| 1851 | } | ||||
| 1852 | |||||
| 1853 | if (UserTreeIdx.UserTE) | ||||
| 1854 | Last->UserTreeIndices.push_back(UserTreeIdx); | ||||
| 1855 | |||||
| 1856 | return Last; | ||||
| 1857 | } | ||||
| 1858 | |||||
| 1859 | /// -- Vectorization State -- | ||||
| 1860 | /// Holds all of the tree entries. | ||||
| 1861 | TreeEntry::VecTreeTy VectorizableTree; | ||||
| 1862 | |||||
| 1863 | #ifndef NDEBUG | ||||
| 1864 | /// Debug printer. | ||||
| 1865 | LLVM_DUMP_METHOD__attribute__((noinline)) __attribute__((__used__)) void dumpVectorizableTree() const { | ||||
| 1866 | for (unsigned Id = 0, IdE = VectorizableTree.size(); Id != IdE; ++Id) { | ||||
| 1867 | VectorizableTree[Id]->dump(); | ||||
| 1868 | dbgs() << "\n"; | ||||
| 1869 | } | ||||
| 1870 | } | ||||
| 1871 | #endif | ||||
| 1872 | |||||
| 1873 | TreeEntry *getTreeEntry(Value *V) { return ScalarToTreeEntry.lookup(V); } | ||||
| 1874 | |||||
| 1875 | const TreeEntry *getTreeEntry(Value *V) const { | ||||
| 1876 | return ScalarToTreeEntry.lookup(V); | ||||
| 1877 | } | ||||
| 1878 | |||||
| 1879 | /// Maps a specific scalar to its tree entry. | ||||
| 1880 | SmallDenseMap<Value*, TreeEntry *> ScalarToTreeEntry; | ||||
| 1881 | |||||
| 1882 | /// Maps a value to the proposed vectorizable size. | ||||
| 1883 | SmallDenseMap<Value *, unsigned> InstrElementSize; | ||||
| 1884 | |||||
| 1885 | /// A list of scalars that we found that we need to keep as scalars. | ||||
| 1886 | ValueSet MustGather; | ||||
| 1887 | |||||
| 1888 | /// This POD struct describes one external user in the vectorized tree. | ||||
| 1889 | struct ExternalUser { | ||||
| 1890 | ExternalUser(Value *S, llvm::User *U, int L) | ||||
| 1891 | : Scalar(S), User(U), Lane(L) {} | ||||
| 1892 | |||||
| 1893 | // Which scalar in our function. | ||||
| 1894 | Value *Scalar; | ||||
| 1895 | |||||
| 1896 | // Which user that uses the scalar. | ||||
| 1897 | llvm::User *User; | ||||
| 1898 | |||||
| 1899 | // Which lane does the scalar belong to. | ||||
| 1900 | int Lane; | ||||
| 1901 | }; | ||||
| 1902 | using UserList = SmallVector<ExternalUser, 16>; | ||||
| 1903 | |||||
| 1904 | /// Checks if two instructions may access the same memory. | ||||
| 1905 | /// | ||||
| 1906 | /// \p Loc1 is the location of \p Inst1. It is passed explicitly because it | ||||
| 1907 | /// is invariant in the calling loop. | ||||
| 1908 | bool isAliased(const MemoryLocation &Loc1, Instruction *Inst1, | ||||
| 1909 | Instruction *Inst2) { | ||||
| 1910 | // First check if the result is already in the cache. | ||||
| 1911 | AliasCacheKey key = std::make_pair(Inst1, Inst2); | ||||
| 1912 | Optional<bool> &result = AliasCache[key]; | ||||
| 1913 | if (result.hasValue()) { | ||||
| 1914 | return result.getValue(); | ||||
| 1915 | } | ||||
| 1916 | MemoryLocation Loc2 = getLocation(Inst2, AA); | ||||
| 1917 | bool aliased = true; | ||||
| 1918 | if (Loc1.Ptr && Loc2.Ptr && isSimple(Inst1) && isSimple(Inst2)) { | ||||
| 1919 | // Do the alias check. | ||||
| 1920 | aliased = !AA->isNoAlias(Loc1, Loc2); | ||||
| 1921 | } | ||||
| 1922 | // Store the result in the cache. | ||||
| 1923 | result = aliased; | ||||
| 1924 | return aliased; | ||||
| 1925 | } | ||||
| 1926 | |||||
| 1927 | using AliasCacheKey = std::pair<Instruction *, Instruction *>; | ||||
| 1928 | |||||
| 1929 | /// Cache for alias results. | ||||
| 1930 | /// TODO: consider moving this to the AliasAnalysis itself. | ||||
| 1931 | DenseMap<AliasCacheKey, Optional<bool>> AliasCache; | ||||
| 1932 | |||||
| 1933 | /// Removes an instruction from its block and eventually deletes it. | ||||
| 1934 | /// It's like Instruction::eraseFromParent() except that the actual deletion | ||||
| 1935 | /// is delayed until BoUpSLP is destructed. | ||||
| 1936 | /// This is required to ensure that there are no incorrect collisions in the | ||||
| 1937 | /// AliasCache, which can happen if a new instruction is allocated at the | ||||
| 1938 | /// same address as a previously deleted instruction. | ||||
| 1939 | void eraseInstruction(Instruction *I, bool ReplaceOpsWithUndef = false) { | ||||
| 1940 | auto It = DeletedInstructions.try_emplace(I, ReplaceOpsWithUndef).first; | ||||
| 1941 | It->getSecond() = It->getSecond() && ReplaceOpsWithUndef; | ||||
| 1942 | } | ||||
| 1943 | |||||
| 1944 | /// Temporary store for deleted instructions. Instructions will be deleted | ||||
| 1945 | /// eventually when the BoUpSLP is destructed. | ||||
| 1946 | DenseMap<Instruction *, bool> DeletedInstructions; | ||||
| 1947 | |||||
| 1948 | /// A list of values that need to extracted out of the tree. | ||||
| 1949 | /// This list holds pairs of (Internal Scalar : External User). External User | ||||
| 1950 | /// can be nullptr, it means that this Internal Scalar will be used later, | ||||
| 1951 | /// after vectorization. | ||||
| 1952 | UserList ExternalUses; | ||||
| 1953 | |||||
| 1954 | /// Values used only by @llvm.assume calls. | ||||
| 1955 | SmallPtrSet<const Value *, 32> EphValues; | ||||
| 1956 | |||||
| 1957 | /// Holds all of the instructions that we gathered. | ||||
| 1958 | SetVector<Instruction *> GatherSeq; | ||||
| 1959 | |||||
| 1960 | /// A list of blocks that we are going to CSE. | ||||
| 1961 | SetVector<BasicBlock *> CSEBlocks; | ||||
| 1962 | |||||
| 1963 | /// Contains all scheduling relevant data for an instruction. | ||||
| 1964 | /// A ScheduleData either represents a single instruction or a member of an | ||||
| 1965 | /// instruction bundle (= a group of instructions which is combined into a | ||||
| 1966 | /// vector instruction). | ||||
| 1967 | struct ScheduleData { | ||||
| 1968 | // The initial value for the dependency counters. It means that the | ||||
| 1969 | // dependencies are not calculated yet. | ||||
| 1970 | enum { InvalidDeps = -1 }; | ||||
| 1971 | |||||
| 1972 | ScheduleData() = default; | ||||
| 1973 | |||||
| 1974 | void init(int BlockSchedulingRegionID, Value *OpVal) { | ||||
| 1975 | FirstInBundle = this; | ||||
| 1976 | NextInBundle = nullptr; | ||||
| 1977 | NextLoadStore = nullptr; | ||||
| 1978 | IsScheduled = false; | ||||
| 1979 | SchedulingRegionID = BlockSchedulingRegionID; | ||||
| 1980 | UnscheduledDepsInBundle = UnscheduledDeps; | ||||
| 1981 | clearDependencies(); | ||||
| 1982 | OpValue = OpVal; | ||||
| 1983 | TE = nullptr; | ||||
| 1984 | Lane = -1; | ||||
| 1985 | } | ||||
| 1986 | |||||
| 1987 | /// Returns true if the dependency information has been calculated. | ||||
| 1988 | bool hasValidDependencies() const { return Dependencies != InvalidDeps; } | ||||
| 1989 | |||||
| 1990 | /// Returns true for single instructions and for bundle representatives | ||||
| 1991 | /// (= the head of a bundle). | ||||
| 1992 | bool isSchedulingEntity() const { return FirstInBundle == this; } | ||||
| 1993 | |||||
| 1994 | /// Returns true if it represents an instruction bundle and not only a | ||||
| 1995 | /// single instruction. | ||||
| 1996 | bool isPartOfBundle() const { | ||||
| 1997 | return NextInBundle != nullptr || FirstInBundle != this; | ||||
| 1998 | } | ||||
| 1999 | |||||
| 2000 | /// Returns true if it is ready for scheduling, i.e. it has no more | ||||
| 2001 | /// unscheduled depending instructions/bundles. | ||||
| 2002 | bool isReady() const { | ||||
| 2003 | assert(isSchedulingEntity() &&(static_cast <bool> (isSchedulingEntity() && "can't consider non-scheduling entity for ready list" ) ? void (0) : __assert_fail ("isSchedulingEntity() && \"can't consider non-scheduling entity for ready list\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2004, __extension__ __PRETTY_FUNCTION__)) | ||||
| 2004 | "can't consider non-scheduling entity for ready list")(static_cast <bool> (isSchedulingEntity() && "can't consider non-scheduling entity for ready list" ) ? void (0) : __assert_fail ("isSchedulingEntity() && \"can't consider non-scheduling entity for ready list\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2004, __extension__ __PRETTY_FUNCTION__)); | ||||
| 2005 | return UnscheduledDepsInBundle == 0 && !IsScheduled; | ||||
| 2006 | } | ||||
| 2007 | |||||
| 2008 | /// Modifies the number of unscheduled dependencies, also updating it for | ||||
| 2009 | /// the whole bundle. | ||||
| 2010 | int incrementUnscheduledDeps(int Incr) { | ||||
| 2011 | UnscheduledDeps += Incr; | ||||
| 2012 | return FirstInBundle->UnscheduledDepsInBundle += Incr; | ||||
| 2013 | } | ||||
| 2014 | |||||
| 2015 | /// Sets the number of unscheduled dependencies to the number of | ||||
| 2016 | /// dependencies. | ||||
| 2017 | void resetUnscheduledDeps() { | ||||
| 2018 | incrementUnscheduledDeps(Dependencies - UnscheduledDeps); | ||||
| 2019 | } | ||||
| 2020 | |||||
| 2021 | /// Clears all dependency information. | ||||
| 2022 | void clearDependencies() { | ||||
| 2023 | Dependencies = InvalidDeps; | ||||
| 2024 | resetUnscheduledDeps(); | ||||
| 2025 | MemoryDependencies.clear(); | ||||
| 2026 | } | ||||
| 2027 | |||||
| 2028 | void dump(raw_ostream &os) const { | ||||
| 2029 | if (!isSchedulingEntity()) { | ||||
| 2030 | os << "/ " << *Inst; | ||||
| 2031 | } else if (NextInBundle) { | ||||
| 2032 | os << '[' << *Inst; | ||||
| 2033 | ScheduleData *SD = NextInBundle; | ||||
| 2034 | while (SD) { | ||||
| 2035 | os << ';' << *SD->Inst; | ||||
| 2036 | SD = SD->NextInBundle; | ||||
| 2037 | } | ||||
| 2038 | os << ']'; | ||||
| 2039 | } else { | ||||
| 2040 | os << *Inst; | ||||
| 2041 | } | ||||
| 2042 | } | ||||
| 2043 | |||||
| 2044 | Instruction *Inst = nullptr; | ||||
| 2045 | |||||
| 2046 | /// Points to the head in an instruction bundle (and always to this for | ||||
| 2047 | /// single instructions). | ||||
| 2048 | ScheduleData *FirstInBundle = nullptr; | ||||
| 2049 | |||||
| 2050 | /// Single linked list of all instructions in a bundle. Null if it is a | ||||
| 2051 | /// single instruction. | ||||
| 2052 | ScheduleData *NextInBundle = nullptr; | ||||
| 2053 | |||||
| 2054 | /// Single linked list of all memory instructions (e.g. load, store, call) | ||||
| 2055 | /// in the block - until the end of the scheduling region. | ||||
| 2056 | ScheduleData *NextLoadStore = nullptr; | ||||
| 2057 | |||||
| 2058 | /// The dependent memory instructions. | ||||
| 2059 | /// This list is derived on demand in calculateDependencies(). | ||||
| 2060 | SmallVector<ScheduleData *, 4> MemoryDependencies; | ||||
| 2061 | |||||
| 2062 | /// This ScheduleData is in the current scheduling region if this matches | ||||
| 2063 | /// the current SchedulingRegionID of BlockScheduling. | ||||
| 2064 | int SchedulingRegionID = 0; | ||||
| 2065 | |||||
| 2066 | /// Used for getting a "good" final ordering of instructions. | ||||
| 2067 | int SchedulingPriority = 0; | ||||
| 2068 | |||||
| 2069 | /// The number of dependencies. Constitutes of the number of users of the | ||||
| 2070 | /// instruction plus the number of dependent memory instructions (if any). | ||||
| 2071 | /// This value is calculated on demand. | ||||
| 2072 | /// If InvalidDeps, the number of dependencies is not calculated yet. | ||||
| 2073 | int Dependencies = InvalidDeps; | ||||
| 2074 | |||||
| 2075 | /// The number of dependencies minus the number of dependencies of scheduled | ||||
| 2076 | /// instructions. As soon as this is zero, the instruction/bundle gets ready | ||||
| 2077 | /// for scheduling. | ||||
| 2078 | /// Note that this is negative as long as Dependencies is not calculated. | ||||
| 2079 | int UnscheduledDeps = InvalidDeps; | ||||
| 2080 | |||||
| 2081 | /// The sum of UnscheduledDeps in a bundle. Equals to UnscheduledDeps for | ||||
| 2082 | /// single instructions. | ||||
| 2083 | int UnscheduledDepsInBundle = InvalidDeps; | ||||
| 2084 | |||||
| 2085 | /// True if this instruction is scheduled (or considered as scheduled in the | ||||
| 2086 | /// dry-run). | ||||
| 2087 | bool IsScheduled = false; | ||||
| 2088 | |||||
| 2089 | /// Opcode of the current instruction in the schedule data. | ||||
| 2090 | Value *OpValue = nullptr; | ||||
| 2091 | |||||
| 2092 | /// The TreeEntry that this instruction corresponds to. | ||||
| 2093 | TreeEntry *TE = nullptr; | ||||
| 2094 | |||||
| 2095 | /// The lane of this node in the TreeEntry. | ||||
| 2096 | int Lane = -1; | ||||
| 2097 | }; | ||||
| 2098 | |||||
| 2099 | #ifndef NDEBUG | ||||
| 2100 | friend inline raw_ostream &operator<<(raw_ostream &os, | ||||
| 2101 | const BoUpSLP::ScheduleData &SD) { | ||||
| 2102 | SD.dump(os); | ||||
| 2103 | return os; | ||||
| 2104 | } | ||||
| 2105 | #endif | ||||
| 2106 | |||||
| 2107 | friend struct GraphTraits<BoUpSLP *>; | ||||
| 2108 | friend struct DOTGraphTraits<BoUpSLP *>; | ||||
| 2109 | |||||
| 2110 | /// Contains all scheduling data for a basic block. | ||||
| 2111 | struct BlockScheduling { | ||||
| 2112 | BlockScheduling(BasicBlock *BB) | ||||
| 2113 | : BB(BB), ChunkSize(BB->size()), ChunkPos(ChunkSize) {} | ||||
| 2114 | |||||
| 2115 | void clear() { | ||||
| 2116 | ReadyInsts.clear(); | ||||
| 2117 | ScheduleStart = nullptr; | ||||
| 2118 | ScheduleEnd = nullptr; | ||||
| 2119 | FirstLoadStoreInRegion = nullptr; | ||||
| 2120 | LastLoadStoreInRegion = nullptr; | ||||
| 2121 | |||||
| 2122 | // Reduce the maximum schedule region size by the size of the | ||||
| 2123 | // previous scheduling run. | ||||
| 2124 | ScheduleRegionSizeLimit -= ScheduleRegionSize; | ||||
| 2125 | if (ScheduleRegionSizeLimit < MinScheduleRegionSize) | ||||
| 2126 | ScheduleRegionSizeLimit = MinScheduleRegionSize; | ||||
| 2127 | ScheduleRegionSize = 0; | ||||
| 2128 | |||||
| 2129 | // Make a new scheduling region, i.e. all existing ScheduleData is not | ||||
| 2130 | // in the new region yet. | ||||
| 2131 | ++SchedulingRegionID; | ||||
| 2132 | } | ||||
| 2133 | |||||
| 2134 | ScheduleData *getScheduleData(Value *V) { | ||||
| 2135 | ScheduleData *SD = ScheduleDataMap[V]; | ||||
| 2136 | if (SD && SD->SchedulingRegionID == SchedulingRegionID) | ||||
| 2137 | return SD; | ||||
| 2138 | return nullptr; | ||||
| 2139 | } | ||||
| 2140 | |||||
| 2141 | ScheduleData *getScheduleData(Value *V, Value *Key) { | ||||
| 2142 | if (V == Key) | ||||
| 2143 | return getScheduleData(V); | ||||
| 2144 | auto I = ExtraScheduleDataMap.find(V); | ||||
| 2145 | if (I != ExtraScheduleDataMap.end()) { | ||||
| 2146 | ScheduleData *SD = I->second[Key]; | ||||
| 2147 | if (SD && SD->SchedulingRegionID == SchedulingRegionID) | ||||
| 2148 | return SD; | ||||
| 2149 | } | ||||
| 2150 | return nullptr; | ||||
| 2151 | } | ||||
| 2152 | |||||
| 2153 | bool isInSchedulingRegion(ScheduleData *SD) const { | ||||
| 2154 | return SD->SchedulingRegionID == SchedulingRegionID; | ||||
| 2155 | } | ||||
| 2156 | |||||
| 2157 | /// Marks an instruction as scheduled and puts all dependent ready | ||||
| 2158 | /// instructions into the ready-list. | ||||
| 2159 | template <typename ReadyListType> | ||||
| 2160 | void schedule(ScheduleData *SD, ReadyListType &ReadyList) { | ||||
| 2161 | SD->IsScheduled = true; | ||||
| 2162 | LLVM_DEBUG(dbgs() << "SLP: schedule " << *SD << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: schedule " << *SD << "\n"; } } while (false); | ||||
| 2163 | |||||
| 2164 | ScheduleData *BundleMember = SD; | ||||
| 2165 | while (BundleMember) { | ||||
| 2166 | if (BundleMember->Inst != BundleMember->OpValue) { | ||||
| 2167 | BundleMember = BundleMember->NextInBundle; | ||||
| 2168 | continue; | ||||
| 2169 | } | ||||
| 2170 | // Handle the def-use chain dependencies. | ||||
| 2171 | |||||
| 2172 | // Decrement the unscheduled counter and insert to ready list if ready. | ||||
| 2173 | auto &&DecrUnsched = [this, &ReadyList](Instruction *I) { | ||||
| 2174 | doForAllOpcodes(I, [&ReadyList](ScheduleData *OpDef) { | ||||
| 2175 | if (OpDef && OpDef->hasValidDependencies() && | ||||
| 2176 | OpDef->incrementUnscheduledDeps(-1) == 0) { | ||||
| 2177 | // There are no more unscheduled dependencies after | ||||
| 2178 | // decrementing, so we can put the dependent instruction | ||||
| 2179 | // into the ready list. | ||||
| 2180 | ScheduleData *DepBundle = OpDef->FirstInBundle; | ||||
| 2181 | assert(!DepBundle->IsScheduled &&(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2182, __extension__ __PRETTY_FUNCTION__)) | ||||
| 2182 | "already scheduled bundle gets ready")(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2182, __extension__ __PRETTY_FUNCTION__)); | ||||
| 2183 | ReadyList.insert(DepBundle); | ||||
| 2184 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"; } } while (false) | ||||
| 2185 | << "SLP: gets ready (def): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n"; } } while (false); | ||||
| 2186 | } | ||||
| 2187 | }); | ||||
| 2188 | }; | ||||
| 2189 | |||||
| 2190 | // If BundleMember is a vector bundle, its operands may have been | ||||
| 2191 | // reordered duiring buildTree(). We therefore need to get its operands | ||||
| 2192 | // through the TreeEntry. | ||||
| 2193 | if (TreeEntry *TE = BundleMember->TE) { | ||||
| 2194 | int Lane = BundleMember->Lane; | ||||
| 2195 | assert(Lane >= 0 && "Lane not set")(static_cast <bool> (Lane >= 0 && "Lane not set" ) ? void (0) : __assert_fail ("Lane >= 0 && \"Lane not set\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2195, __extension__ __PRETTY_FUNCTION__)); | ||||
| 2196 | |||||
| 2197 | // Since vectorization tree is being built recursively this assertion | ||||
| 2198 | // ensures that the tree entry has all operands set before reaching | ||||
| 2199 | // this code. Couple of exceptions known at the moment are extracts | ||||
| 2200 | // where their second (immediate) operand is not added. Since | ||||
| 2201 | // immediates do not affect scheduler behavior this is considered | ||||
| 2202 | // okay. | ||||
| 2203 | auto *In = TE->getMainOp(); | ||||
| 2204 | assert(In &&(static_cast <bool> (In && (isa<ExtractValueInst >(In) || isa<ExtractElementInst>(In) || In->getNumOperands () == TE->getNumOperands()) && "Missed TreeEntry operands?" ) ? void (0) : __assert_fail ("In && (isa<ExtractValueInst>(In) || isa<ExtractElementInst>(In) || In->getNumOperands() == TE->getNumOperands()) && \"Missed TreeEntry operands?\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2207, __extension__ __PRETTY_FUNCTION__)) | ||||
| 2205 | (isa<ExtractValueInst>(In) || isa<ExtractElementInst>(In) ||(static_cast <bool> (In && (isa<ExtractValueInst >(In) || isa<ExtractElementInst>(In) || In->getNumOperands () == TE->getNumOperands()) && "Missed TreeEntry operands?" ) ? void (0) : __assert_fail ("In && (isa<ExtractValueInst>(In) || isa<ExtractElementInst>(In) || In->getNumOperands() == TE->getNumOperands()) && \"Missed TreeEntry operands?\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2207, __extension__ __PRETTY_FUNCTION__)) | ||||
| 2206 | In->getNumOperands() == TE->getNumOperands()) &&(static_cast <bool> (In && (isa<ExtractValueInst >(In) || isa<ExtractElementInst>(In) || In->getNumOperands () == TE->getNumOperands()) && "Missed TreeEntry operands?" ) ? void (0) : __assert_fail ("In && (isa<ExtractValueInst>(In) || isa<ExtractElementInst>(In) || In->getNumOperands() == TE->getNumOperands()) && \"Missed TreeEntry operands?\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2207, __extension__ __PRETTY_FUNCTION__)) | ||||
| 2207 | "Missed TreeEntry operands?")(static_cast <bool> (In && (isa<ExtractValueInst >(In) || isa<ExtractElementInst>(In) || In->getNumOperands () == TE->getNumOperands()) && "Missed TreeEntry operands?" ) ? void (0) : __assert_fail ("In && (isa<ExtractValueInst>(In) || isa<ExtractElementInst>(In) || In->getNumOperands() == TE->getNumOperands()) && \"Missed TreeEntry operands?\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2207, __extension__ __PRETTY_FUNCTION__)); | ||||
| 2208 | (void)In; // fake use to avoid build failure when assertions disabled | ||||
| 2209 | |||||
| 2210 | for (unsigned OpIdx = 0, NumOperands = TE->getNumOperands(); | ||||
| 2211 | OpIdx != NumOperands; ++OpIdx) | ||||
| 2212 | if (auto *I = dyn_cast<Instruction>(TE->getOperand(OpIdx)[Lane])) | ||||
| 2213 | DecrUnsched(I); | ||||
| 2214 | } else { | ||||
| 2215 | // If BundleMember is a stand-alone instruction, no operand reordering | ||||
| 2216 | // has taken place, so we directly access its operands. | ||||
| 2217 | for (Use &U : BundleMember->Inst->operands()) | ||||
| 2218 | if (auto *I = dyn_cast<Instruction>(U.get())) | ||||
| 2219 | DecrUnsched(I); | ||||
| 2220 | } | ||||
| 2221 | // Handle the memory dependencies. | ||||
| 2222 | for (ScheduleData *MemoryDepSD : BundleMember->MemoryDependencies) { | ||||
| 2223 | if (MemoryDepSD->incrementUnscheduledDeps(-1) == 0) { | ||||
| 2224 | // There are no more unscheduled dependencies after decrementing, | ||||
| 2225 | // so we can put the dependent instruction into the ready list. | ||||
| 2226 | ScheduleData *DepBundle = MemoryDepSD->FirstInBundle; | ||||
| 2227 | assert(!DepBundle->IsScheduled &&(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2228, __extension__ __PRETTY_FUNCTION__)) | ||||
| 2228 | "already scheduled bundle gets ready")(static_cast <bool> (!DepBundle->IsScheduled && "already scheduled bundle gets ready") ? void (0) : __assert_fail ("!DepBundle->IsScheduled && \"already scheduled bundle gets ready\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2228, __extension__ __PRETTY_FUNCTION__)); | ||||
| 2229 | ReadyList.insert(DepBundle); | ||||
| 2230 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"; } } while (false) | ||||
| 2231 | << "SLP: gets ready (mem): " << *DepBundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n"; } } while (false); | ||||
| 2232 | } | ||||
| 2233 | } | ||||
| 2234 | BundleMember = BundleMember->NextInBundle; | ||||
| 2235 | } | ||||
| 2236 | } | ||||
| 2237 | |||||
| 2238 | void doForAllOpcodes(Value *V, | ||||
| 2239 | function_ref<void(ScheduleData *SD)> Action) { | ||||
| 2240 | if (ScheduleData *SD = getScheduleData(V)) | ||||
| 2241 | Action(SD); | ||||
| 2242 | auto I = ExtraScheduleDataMap.find(V); | ||||
| 2243 | if (I != ExtraScheduleDataMap.end()) | ||||
| 2244 | for (auto &P : I->second) | ||||
| 2245 | if (P.second->SchedulingRegionID == SchedulingRegionID) | ||||
| 2246 | Action(P.second); | ||||
| 2247 | } | ||||
| 2248 | |||||
| 2249 | /// Put all instructions into the ReadyList which are ready for scheduling. | ||||
| 2250 | template <typename ReadyListType> | ||||
| 2251 | void initialFillReadyList(ReadyListType &ReadyList) { | ||||
| 2252 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | ||||
| 2253 | doForAllOpcodes(I, [&](ScheduleData *SD) { | ||||
| 2254 | if (SD->isSchedulingEntity() && SD->isReady()) { | ||||
| 2255 | ReadyList.insert(SD); | ||||
| 2256 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: initially in ready list: " << *I << "\n"; } } while (false) | ||||
| 2257 | << "SLP: initially in ready list: " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: initially in ready list: " << *I << "\n"; } } while (false); | ||||
| 2258 | } | ||||
| 2259 | }); | ||||
| 2260 | } | ||||
| 2261 | } | ||||
| 2262 | |||||
| 2263 | /// Checks if a bundle of instructions can be scheduled, i.e. has no | ||||
| 2264 | /// cyclic dependencies. This is only a dry-run, no instructions are | ||||
| 2265 | /// actually moved at this stage. | ||||
| 2266 | /// \returns the scheduling bundle. The returned Optional value is non-None | ||||
| 2267 | /// if \p VL is allowed to be scheduled. | ||||
| 2268 | Optional<ScheduleData *> | ||||
| 2269 | tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP, | ||||
| 2270 | const InstructionsState &S); | ||||
| 2271 | |||||
| 2272 | /// Un-bundles a group of instructions. | ||||
| 2273 | void cancelScheduling(ArrayRef<Value *> VL, Value *OpValue); | ||||
| 2274 | |||||
| 2275 | /// Allocates schedule data chunk. | ||||
| 2276 | ScheduleData *allocateScheduleDataChunks(); | ||||
| 2277 | |||||
| 2278 | /// Extends the scheduling region so that V is inside the region. | ||||
| 2279 | /// \returns true if the region size is within the limit. | ||||
| 2280 | bool extendSchedulingRegion(Value *V, const InstructionsState &S); | ||||
| 2281 | |||||
| 2282 | /// Initialize the ScheduleData structures for new instructions in the | ||||
| 2283 | /// scheduling region. | ||||
| 2284 | void initScheduleData(Instruction *FromI, Instruction *ToI, | ||||
| 2285 | ScheduleData *PrevLoadStore, | ||||
| 2286 | ScheduleData *NextLoadStore); | ||||
| 2287 | |||||
| 2288 | /// Updates the dependency information of a bundle and of all instructions/ | ||||
| 2289 | /// bundles which depend on the original bundle. | ||||
| 2290 | void calculateDependencies(ScheduleData *SD, bool InsertInReadyList, | ||||
| 2291 | BoUpSLP *SLP); | ||||
| 2292 | |||||
| 2293 | /// Sets all instruction in the scheduling region to un-scheduled. | ||||
| 2294 | void resetSchedule(); | ||||
| 2295 | |||||
| 2296 | BasicBlock *BB; | ||||
| 2297 | |||||
| 2298 | /// Simple memory allocation for ScheduleData. | ||||
| 2299 | std::vector<std::unique_ptr<ScheduleData[]>> ScheduleDataChunks; | ||||
| 2300 | |||||
| 2301 | /// The size of a ScheduleData array in ScheduleDataChunks. | ||||
| 2302 | int ChunkSize; | ||||
| 2303 | |||||
| 2304 | /// The allocator position in the current chunk, which is the last entry | ||||
| 2305 | /// of ScheduleDataChunks. | ||||
| 2306 | int ChunkPos; | ||||
| 2307 | |||||
| 2308 | /// Attaches ScheduleData to Instruction. | ||||
| 2309 | /// Note that the mapping survives during all vectorization iterations, i.e. | ||||
| 2310 | /// ScheduleData structures are recycled. | ||||
| 2311 | DenseMap<Value *, ScheduleData *> ScheduleDataMap; | ||||
| 2312 | |||||
| 2313 | /// Attaches ScheduleData to Instruction with the leading key. | ||||
| 2314 | DenseMap<Value *, SmallDenseMap<Value *, ScheduleData *>> | ||||
| 2315 | ExtraScheduleDataMap; | ||||
| 2316 | |||||
| 2317 | struct ReadyList : SmallVector<ScheduleData *, 8> { | ||||
| 2318 | void insert(ScheduleData *SD) { push_back(SD); } | ||||
| 2319 | }; | ||||
| 2320 | |||||
| 2321 | /// The ready-list for scheduling (only used for the dry-run). | ||||
| 2322 | ReadyList ReadyInsts; | ||||
| 2323 | |||||
| 2324 | /// The first instruction of the scheduling region. | ||||
| 2325 | Instruction *ScheduleStart = nullptr; | ||||
| 2326 | |||||
| 2327 | /// The first instruction _after_ the scheduling region. | ||||
| 2328 | Instruction *ScheduleEnd = nullptr; | ||||
| 2329 | |||||
| 2330 | /// The first memory accessing instruction in the scheduling region | ||||
| 2331 | /// (can be null). | ||||
| 2332 | ScheduleData *FirstLoadStoreInRegion = nullptr; | ||||
| 2333 | |||||
| 2334 | /// The last memory accessing instruction in the scheduling region | ||||
| 2335 | /// (can be null). | ||||
| 2336 | ScheduleData *LastLoadStoreInRegion = nullptr; | ||||
| 2337 | |||||
| 2338 | /// The current size of the scheduling region. | ||||
| 2339 | int ScheduleRegionSize = 0; | ||||
| 2340 | |||||
| 2341 | /// The maximum size allowed for the scheduling region. | ||||
| 2342 | int ScheduleRegionSizeLimit = ScheduleRegionSizeBudget; | ||||
| 2343 | |||||
| 2344 | /// The ID of the scheduling region. For a new vectorization iteration this | ||||
| 2345 | /// is incremented which "removes" all ScheduleData from the region. | ||||
| 2346 | // Make sure that the initial SchedulingRegionID is greater than the | ||||
| 2347 | // initial SchedulingRegionID in ScheduleData (which is 0). | ||||
| 2348 | int SchedulingRegionID = 1; | ||||
| 2349 | }; | ||||
| 2350 | |||||
| 2351 | /// Attaches the BlockScheduling structures to basic blocks. | ||||
| 2352 | MapVector<BasicBlock *, std::unique_ptr<BlockScheduling>> BlocksSchedules; | ||||
| 2353 | |||||
| 2354 | /// Performs the "real" scheduling. Done before vectorization is actually | ||||
| 2355 | /// performed in a basic block. | ||||
| 2356 | void scheduleBlock(BlockScheduling *BS); | ||||
| 2357 | |||||
| 2358 | /// List of users to ignore during scheduling and that don't need extracting. | ||||
| 2359 | ArrayRef<Value *> UserIgnoreList; | ||||
| 2360 | |||||
| 2361 | /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of | ||||
| 2362 | /// sorted SmallVectors of unsigned. | ||||
| 2363 | struct OrdersTypeDenseMapInfo { | ||||
| 2364 | static OrdersType getEmptyKey() { | ||||
| 2365 | OrdersType V; | ||||
| 2366 | V.push_back(~1U); | ||||
| 2367 | return V; | ||||
| 2368 | } | ||||
| 2369 | |||||
| 2370 | static OrdersType getTombstoneKey() { | ||||
| 2371 | OrdersType V; | ||||
| 2372 | V.push_back(~2U); | ||||
| 2373 | return V; | ||||
| 2374 | } | ||||
| 2375 | |||||
| 2376 | static unsigned getHashValue(const OrdersType &V) { | ||||
| 2377 | return static_cast<unsigned>(hash_combine_range(V.begin(), V.end())); | ||||
| 2378 | } | ||||
| 2379 | |||||
| 2380 | static bool isEqual(const OrdersType &LHS, const OrdersType &RHS) { | ||||
| 2381 | return LHS == RHS; | ||||
| 2382 | } | ||||
| 2383 | }; | ||||
| 2384 | |||||
| 2385 | /// Contains orders of operations along with the number of bundles that have | ||||
| 2386 | /// operations in this order. It stores only those orders that require | ||||
| 2387 | /// reordering, if reordering is not required it is counted using \a | ||||
| 2388 | /// NumOpsWantToKeepOriginalOrder. | ||||
| 2389 | DenseMap<OrdersType, unsigned, OrdersTypeDenseMapInfo> NumOpsWantToKeepOrder; | ||||
| 2390 | /// Number of bundles that do not require reordering. | ||||
| 2391 | unsigned NumOpsWantToKeepOriginalOrder = 0; | ||||
| 2392 | |||||
| 2393 | // Analysis and block reference. | ||||
| 2394 | Function *F; | ||||
| 2395 | ScalarEvolution *SE; | ||||
| 2396 | TargetTransformInfo *TTI; | ||||
| 2397 | TargetLibraryInfo *TLI; | ||||
| 2398 | AAResults *AA; | ||||
| 2399 | LoopInfo *LI; | ||||
| 2400 | DominatorTree *DT; | ||||
| 2401 | AssumptionCache *AC; | ||||
| 2402 | DemandedBits *DB; | ||||
| 2403 | const DataLayout *DL; | ||||
| 2404 | OptimizationRemarkEmitter *ORE; | ||||
| 2405 | |||||
| 2406 | unsigned MaxVecRegSize; // This is set by TTI or overridden by cl::opt. | ||||
| 2407 | unsigned MinVecRegSize; // Set by cl::opt (default: 128). | ||||
| 2408 | |||||
| 2409 | /// Instruction builder to construct the vectorized tree. | ||||
| 2410 | IRBuilder<> Builder; | ||||
| 2411 | |||||
| 2412 | /// A map of scalar integer values to the smallest bit width with which they | ||||
| 2413 | /// can legally be represented. The values map to (width, signed) pairs, | ||||
| 2414 | /// where "width" indicates the minimum bit width and "signed" is True if the | ||||
| 2415 | /// value must be signed-extended, rather than zero-extended, back to its | ||||
| 2416 | /// original width. | ||||
| 2417 | MapVector<Value *, std::pair<uint64_t, bool>> MinBWs; | ||||
| 2418 | }; | ||||
| 2419 | |||||
| 2420 | } // end namespace slpvectorizer | ||||
| 2421 | |||||
| 2422 | template <> struct GraphTraits<BoUpSLP *> { | ||||
| 2423 | using TreeEntry = BoUpSLP::TreeEntry; | ||||
| 2424 | |||||
| 2425 | /// NodeRef has to be a pointer per the GraphWriter. | ||||
| 2426 | using NodeRef = TreeEntry *; | ||||
| 2427 | |||||
| 2428 | using ContainerTy = BoUpSLP::TreeEntry::VecTreeTy; | ||||
| 2429 | |||||
| 2430 | /// Add the VectorizableTree to the index iterator to be able to return | ||||
| 2431 | /// TreeEntry pointers. | ||||
| 2432 | struct ChildIteratorType | ||||
| 2433 | : public iterator_adaptor_base< | ||||
| 2434 | ChildIteratorType, SmallVector<BoUpSLP::EdgeInfo, 1>::iterator> { | ||||
| 2435 | ContainerTy &VectorizableTree; | ||||
| 2436 | |||||
| 2437 | ChildIteratorType(SmallVector<BoUpSLP::EdgeInfo, 1>::iterator W, | ||||
| 2438 | ContainerTy &VT) | ||||
| 2439 | : ChildIteratorType::iterator_adaptor_base(W), VectorizableTree(VT) {} | ||||
| 2440 | |||||
| 2441 | NodeRef operator*() { return I->UserTE; } | ||||
| 2442 | }; | ||||
| 2443 | |||||
| 2444 | static NodeRef getEntryNode(BoUpSLP &R) { | ||||
| 2445 | return R.VectorizableTree[0].get(); | ||||
| 2446 | } | ||||
| 2447 | |||||
| 2448 | static ChildIteratorType child_begin(NodeRef N) { | ||||
| 2449 | return {N->UserTreeIndices.begin(), N->Container}; | ||||
| 2450 | } | ||||
| 2451 | |||||
| 2452 | static ChildIteratorType child_end(NodeRef N) { | ||||
| 2453 | return {N->UserTreeIndices.end(), N->Container}; | ||||
| 2454 | } | ||||
| 2455 | |||||
| 2456 | /// For the node iterator we just need to turn the TreeEntry iterator into a | ||||
| 2457 | /// TreeEntry* iterator so that it dereferences to NodeRef. | ||||
| 2458 | class nodes_iterator { | ||||
| 2459 | using ItTy = ContainerTy::iterator; | ||||
| 2460 | ItTy It; | ||||
| 2461 | |||||
| 2462 | public: | ||||
| 2463 | nodes_iterator(const ItTy &It2) : It(It2) {} | ||||
| 2464 | NodeRef operator*() { return It->get(); } | ||||
| 2465 | nodes_iterator operator++() { | ||||
| 2466 | ++It; | ||||
| 2467 | return *this; | ||||
| 2468 | } | ||||
| 2469 | bool operator!=(const nodes_iterator &N2) const { return N2.It != It; } | ||||
| 2470 | }; | ||||
| 2471 | |||||
| 2472 | static nodes_iterator nodes_begin(BoUpSLP *R) { | ||||
| 2473 | return nodes_iterator(R->VectorizableTree.begin()); | ||||
| 2474 | } | ||||
| 2475 | |||||
| 2476 | static nodes_iterator nodes_end(BoUpSLP *R) { | ||||
| 2477 | return nodes_iterator(R->VectorizableTree.end()); | ||||
| 2478 | } | ||||
| 2479 | |||||
| 2480 | static unsigned size(BoUpSLP *R) { return R->VectorizableTree.size(); } | ||||
| 2481 | }; | ||||
| 2482 | |||||
| 2483 | template <> struct DOTGraphTraits<BoUpSLP *> : public DefaultDOTGraphTraits { | ||||
| 2484 | using TreeEntry = BoUpSLP::TreeEntry; | ||||
| 2485 | |||||
| 2486 | DOTGraphTraits(bool isSimple = false) : DefaultDOTGraphTraits(isSimple) {} | ||||
| 2487 | |||||
| 2488 | std::string getNodeLabel(const TreeEntry *Entry, const BoUpSLP *R) { | ||||
| 2489 | std::string Str; | ||||
| 2490 | raw_string_ostream OS(Str); | ||||
| 2491 | if (isSplat(Entry->Scalars)) { | ||||
| 2492 | OS << "<splat> " << *Entry->Scalars[0]; | ||||
| 2493 | return Str; | ||||
| 2494 | } | ||||
| 2495 | for (auto V : Entry->Scalars) { | ||||
| 2496 | OS << *V; | ||||
| 2497 | if (llvm::any_of(R->ExternalUses, [&](const BoUpSLP::ExternalUser &EU) { | ||||
| 2498 | return EU.Scalar == V; | ||||
| 2499 | })) | ||||
| 2500 | OS << " <extract>"; | ||||
| 2501 | OS << "\n"; | ||||
| 2502 | } | ||||
| 2503 | return Str; | ||||
| 2504 | } | ||||
| 2505 | |||||
| 2506 | static std::string getNodeAttributes(const TreeEntry *Entry, | ||||
| 2507 | const BoUpSLP *) { | ||||
| 2508 | if (Entry->State == TreeEntry::NeedToGather) | ||||
| 2509 | return "color=red"; | ||||
| 2510 | return ""; | ||||
| 2511 | } | ||||
| 2512 | }; | ||||
| 2513 | |||||
| 2514 | } // end namespace llvm | ||||
| 2515 | |||||
| 2516 | BoUpSLP::~BoUpSLP() { | ||||
| 2517 | for (const auto &Pair : DeletedInstructions) { | ||||
| 2518 | // Replace operands of ignored instructions with Undefs in case if they were | ||||
| 2519 | // marked for deletion. | ||||
| 2520 | if (Pair.getSecond()) { | ||||
| 2521 | Value *Undef = UndefValue::get(Pair.getFirst()->getType()); | ||||
| 2522 | Pair.getFirst()->replaceAllUsesWith(Undef); | ||||
| 2523 | } | ||||
| 2524 | Pair.getFirst()->dropAllReferences(); | ||||
| 2525 | } | ||||
| 2526 | for (const auto &Pair : DeletedInstructions) { | ||||
| 2527 | assert(Pair.getFirst()->use_empty() &&(static_cast <bool> (Pair.getFirst()->use_empty() && "trying to erase instruction with users.") ? void (0) : __assert_fail ("Pair.getFirst()->use_empty() && \"trying to erase instruction with users.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2528, __extension__ __PRETTY_FUNCTION__)) | ||||
| 2528 | "trying to erase instruction with users.")(static_cast <bool> (Pair.getFirst()->use_empty() && "trying to erase instruction with users.") ? void (0) : __assert_fail ("Pair.getFirst()->use_empty() && \"trying to erase instruction with users.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2528, __extension__ __PRETTY_FUNCTION__)); | ||||
| 2529 | Pair.getFirst()->eraseFromParent(); | ||||
| 2530 | } | ||||
| 2531 | #ifdef EXPENSIVE_CHECKS | ||||
| 2532 | // If we could guarantee that this call is not extremely slow, we could | ||||
| 2533 | // remove the ifdef limitation (see PR47712). | ||||
| 2534 | assert(!verifyFunction(*F, &dbgs()))(static_cast <bool> (!verifyFunction(*F, &dbgs())) ? void (0) : __assert_fail ("!verifyFunction(*F, &dbgs())" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2534, __extension__ __PRETTY_FUNCTION__)); | ||||
| 2535 | #endif | ||||
| 2536 | } | ||||
| 2537 | |||||
| 2538 | void BoUpSLP::eraseInstructions(ArrayRef<Value *> AV) { | ||||
| 2539 | for (auto *V : AV) { | ||||
| 2540 | if (auto *I = dyn_cast<Instruction>(V)) | ||||
| 2541 | eraseInstruction(I, /*ReplaceOpsWithUndef=*/true); | ||||
| 2542 | }; | ||||
| 2543 | } | ||||
| 2544 | |||||
| 2545 | void BoUpSLP::buildTree(ArrayRef<Value *> Roots, | ||||
| 2546 | ArrayRef<Value *> UserIgnoreLst) { | ||||
| 2547 | ExtraValueToDebugLocsMap ExternallyUsedValues; | ||||
| 2548 | buildTree(Roots, ExternallyUsedValues, UserIgnoreLst); | ||||
| 2549 | } | ||||
| 2550 | |||||
| 2551 | static int findLaneForValue(ArrayRef<Value *> Scalars, | ||||
| 2552 | ArrayRef<int> ReuseShuffleIndices, Value *V) { | ||||
| 2553 | unsigned FoundLane = std::distance(Scalars.begin(), find(Scalars, V)); | ||||
| 2554 | assert(FoundLane < Scalars.size() && "Couldn't find extract lane")(static_cast <bool> (FoundLane < Scalars.size() && "Couldn't find extract lane") ? void (0) : __assert_fail ("FoundLane < Scalars.size() && \"Couldn't find extract lane\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2554, __extension__ __PRETTY_FUNCTION__)); | ||||
| 2555 | if (!ReuseShuffleIndices.empty()) { | ||||
| 2556 | FoundLane = std::distance(ReuseShuffleIndices.begin(), | ||||
| 2557 | find(ReuseShuffleIndices, FoundLane)); | ||||
| 2558 | } | ||||
| 2559 | return FoundLane; | ||||
| 2560 | } | ||||
| 2561 | |||||
| 2562 | void BoUpSLP::buildTree(ArrayRef<Value *> Roots, | ||||
| 2563 | ExtraValueToDebugLocsMap &ExternallyUsedValues, | ||||
| 2564 | ArrayRef<Value *> UserIgnoreLst) { | ||||
| 2565 | deleteTree(); | ||||
| 2566 | UserIgnoreList = UserIgnoreLst; | ||||
| 2567 | if (!allSameType(Roots)) | ||||
| 2568 | return; | ||||
| 2569 | buildTree_rec(Roots, 0, EdgeInfo()); | ||||
| 2570 | |||||
| 2571 | // Collect the values that we need to extract from the tree. | ||||
| 2572 | for (auto &TEPtr : VectorizableTree) { | ||||
| 2573 | TreeEntry *Entry = TEPtr.get(); | ||||
| 2574 | |||||
| 2575 | // No need to handle users of gathered values. | ||||
| 2576 | if (Entry->State == TreeEntry::NeedToGather) | ||||
| 2577 | continue; | ||||
| 2578 | |||||
| 2579 | // For each lane: | ||||
| 2580 | for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { | ||||
| 2581 | Value *Scalar = Entry->Scalars[Lane]; | ||||
| 2582 | int FoundLane = | ||||
| 2583 | findLaneForValue(Entry->Scalars, Entry->ReuseShuffleIndices, Scalar); | ||||
| 2584 | |||||
| 2585 | // Check if the scalar is externally used as an extra arg. | ||||
| 2586 | auto ExtI = ExternallyUsedValues.find(Scalar); | ||||
| 2587 | if (ExtI != ExternallyUsedValues.end()) { | ||||
| 2588 | LLVM_DEBUG(dbgs() << "SLP: Need to extract: Extra arg from lane "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to extract: Extra arg from lane " << Lane << " from " << *Scalar << ".\n" ; } } while (false) | ||||
| 2589 | << Lane << " from " << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to extract: Extra arg from lane " << Lane << " from " << *Scalar << ".\n" ; } } while (false); | ||||
| 2590 | ExternalUses.emplace_back(Scalar, nullptr, FoundLane); | ||||
| 2591 | } | ||||
| 2592 | for (User *U : Scalar->users()) { | ||||
| 2593 | LLVM_DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Checking user:" << *U << ".\n"; } } while (false); | ||||
| 2594 | |||||
| 2595 | Instruction *UserInst = dyn_cast<Instruction>(U); | ||||
| 2596 | if (!UserInst) | ||||
| 2597 | continue; | ||||
| 2598 | |||||
| 2599 | // Skip in-tree scalars that become vectors | ||||
| 2600 | if (TreeEntry *UseEntry = getTreeEntry(U)) { | ||||
| 2601 | Value *UseScalar = UseEntry->Scalars[0]; | ||||
| 2602 | // Some in-tree scalars will remain as scalar in vectorized | ||||
| 2603 | // instructions. If that is the case, the one in Lane 0 will | ||||
| 2604 | // be used. | ||||
| 2605 | if (UseScalar != U || | ||||
| 2606 | UseEntry->State == TreeEntry::ScatterVectorize || | ||||
| 2607 | !InTreeUserNeedToExtract(Scalar, UserInst, TLI)) { | ||||
| 2608 | LLVM_DEBUG(dbgs() << "SLP: \tInternal user will be removed:" << *Udo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tInternal user will be removed:" << *U << ".\n"; } } while (false) | ||||
| 2609 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tInternal user will be removed:" << *U << ".\n"; } } while (false); | ||||
| 2610 | assert(UseEntry->State != TreeEntry::NeedToGather && "Bad state")(static_cast <bool> (UseEntry->State != TreeEntry::NeedToGather && "Bad state") ? void (0) : __assert_fail ("UseEntry->State != TreeEntry::NeedToGather && \"Bad state\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2610, __extension__ __PRETTY_FUNCTION__)); | ||||
| 2611 | continue; | ||||
| 2612 | } | ||||
| 2613 | } | ||||
| 2614 | |||||
| 2615 | // Ignore users in the user ignore list. | ||||
| 2616 | if (is_contained(UserIgnoreList, UserInst)) | ||||
| 2617 | continue; | ||||
| 2618 | |||||
| 2619 | LLVM_DEBUG(dbgs() << "SLP: Need to extract:" << *U << " from lane "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to extract:" << * U << " from lane " << Lane << " from " << *Scalar << ".\n"; } } while (false) | ||||
| 2620 | << Lane << " from " << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to extract:" << * U << " from lane " << Lane << " from " << *Scalar << ".\n"; } } while (false); | ||||
| 2621 | ExternalUses.push_back(ExternalUser(Scalar, U, FoundLane)); | ||||
| 2622 | } | ||||
| 2623 | } | ||||
| 2624 | } | ||||
| 2625 | } | ||||
| 2626 | |||||
| 2627 | void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth, | ||||
| 2628 | const EdgeInfo &UserTreeIdx) { | ||||
| 2629 | assert((allConstant(VL) || allSameType(VL)) && "Invalid types!")(static_cast <bool> ((allConstant(VL) || allSameType(VL )) && "Invalid types!") ? void (0) : __assert_fail ("(allConstant(VL) || allSameType(VL)) && \"Invalid types!\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2629, __extension__ __PRETTY_FUNCTION__)); | ||||
| |||||
| 2630 | |||||
| 2631 | InstructionsState S = getSameOpcode(VL); | ||||
| 2632 | if (Depth == RecursionMaxDepth) { | ||||
| 2633 | LLVM_DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to max recursion depth.\n" ; } } while (false); | ||||
| 2634 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
| 2635 | return; | ||||
| 2636 | } | ||||
| 2637 | |||||
| 2638 | // Don't handle vectors. | ||||
| 2639 | if (S.OpValue->getType()->isVectorTy()) { | ||||
| 2640 | LLVM_DEBUG(dbgs() << "SLP: Gathering due to vector type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to vector type.\n" ; } } while (false); | ||||
| 2641 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
| 2642 | return; | ||||
| 2643 | } | ||||
| 2644 | |||||
| 2645 | if (StoreInst *SI
| ||||
| 2646 | if (SI->getValueOperand()->getType()->isVectorTy()) { | ||||
| 2647 | LLVM_DEBUG(dbgs() << "SLP: Gathering due to store vector type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to store vector type.\n" ; } } while (false); | ||||
| 2648 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
| 2649 | return; | ||||
| 2650 | } | ||||
| 2651 | |||||
| 2652 | // If all of the operands are identical or constant we have a simple solution. | ||||
| 2653 | if (allConstant(VL) || isSplat(VL) || !allSameBlock(VL) || !S.getOpcode()) { | ||||
| 2654 | LLVM_DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O. \n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to C,S,B,O. \n" ; } } while (false); | ||||
| 2655 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
| 2656 | return; | ||||
| 2657 | } | ||||
| 2658 | |||||
| 2659 | // We now know that this is a vector of instructions of the same type from | ||||
| 2660 | // the same block. | ||||
| 2661 | |||||
| 2662 | // Don't vectorize ephemeral values. | ||||
| 2663 | for (Value *V : VL) { | ||||
| 2664 | if (EphValues.count(V)) { | ||||
| 2665 | LLVM_DEBUG(dbgs() << "SLP: The instruction (" << *Vdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is ephemeral.\n"; } } while (false) | ||||
| 2666 | << ") is ephemeral.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is ephemeral.\n"; } } while (false); | ||||
| 2667 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
| 2668 | return; | ||||
| 2669 | } | ||||
| 2670 | } | ||||
| 2671 | |||||
| 2672 | // Check if this is a duplicate of another entry. | ||||
| 2673 | if (TreeEntry *E = getTreeEntry(S.OpValue)) { | ||||
| 2674 | LLVM_DEBUG(dbgs() << "SLP: \tChecking bundle: " << *S.OpValue << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tChecking bundle: " << *S.OpValue << ".\n"; } } while (false); | ||||
| 2675 | if (!E->isSame(VL)) { | ||||
| 2676 | LLVM_DEBUG(dbgs() << "SLP: Gathering due to partial overlap.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to partial overlap.\n" ; } } while (false); | ||||
| 2677 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
| 2678 | return; | ||||
| 2679 | } | ||||
| 2680 | // Record the reuse of the tree node. FIXME, currently this is only used to | ||||
| 2681 | // properly draw the graph rather than for the actual vectorization. | ||||
| 2682 | E->UserTreeIndices.push_back(UserTreeIdx); | ||||
| 2683 | LLVM_DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *S.OpValuedo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Perfect diamond merge at " << *S.OpValue << ".\n"; } } while (false) | ||||
| 2684 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Perfect diamond merge at " << *S.OpValue << ".\n"; } } while (false); | ||||
| 2685 | return; | ||||
| 2686 | } | ||||
| 2687 | |||||
| 2688 | // Check that none of the instructions in the bundle are already in the tree. | ||||
| 2689 | for (Value *V : VL) { | ||||
| 2690 | auto *I = dyn_cast<Instruction>(V); | ||||
| 2691 | if (!I) | ||||
| 2692 | continue; | ||||
| 2693 | if (getTreeEntry(I)) { | ||||
| 2694 | LLVM_DEBUG(dbgs() << "SLP: The instruction (" << *Vdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is already in tree.\n"; } } while (false) | ||||
| 2695 | << ") is already in tree.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: The instruction (" << * V << ") is already in tree.\n"; } } while (false); | ||||
| 2696 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
| 2697 | return; | ||||
| 2698 | } | ||||
| 2699 | } | ||||
| 2700 | |||||
| 2701 | // If any of the scalars is marked as a value that needs to stay scalar, then | ||||
| 2702 | // we need to gather the scalars. | ||||
| 2703 | // The reduction nodes (stored in UserIgnoreList) also should stay scalar. | ||||
| 2704 | for (Value *V : VL) { | ||||
| 2705 | if (MustGather.count(V) || is_contained(UserIgnoreList, V)) { | ||||
| 2706 | LLVM_DEBUG(dbgs() << "SLP: Gathering due to gathered scalar.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering due to gathered scalar.\n" ; } } while (false); | ||||
| 2707 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
| 2708 | return; | ||||
| 2709 | } | ||||
| 2710 | } | ||||
| 2711 | |||||
| 2712 | // Check that all of the users of the scalars that we want to vectorize are | ||||
| 2713 | // schedulable. | ||||
| 2714 | auto *VL0 = cast<Instruction>(S.OpValue); | ||||
| 2715 | BasicBlock *BB = VL0->getParent(); | ||||
| 2716 | |||||
| 2717 | if (!DT->isReachableFromEntry(BB)) { | ||||
| 2718 | // Don't go into unreachable blocks. They may contain instructions with | ||||
| 2719 | // dependency cycles which confuse the final scheduling. | ||||
| 2720 | LLVM_DEBUG(dbgs() << "SLP: bundle in unreachable block.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: bundle in unreachable block.\n" ; } } while (false); | ||||
| 2721 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
| 2722 | return; | ||||
| 2723 | } | ||||
| 2724 | |||||
| 2725 | // Check that every instruction appears once in this bundle. | ||||
| 2726 | SmallVector<unsigned, 4> ReuseShuffleIndicies; | ||||
| 2727 | SmallVector<Value *, 4> UniqueValues; | ||||
| 2728 | DenseMap<Value *, unsigned> UniquePositions; | ||||
| 2729 | for (Value *V : VL) { | ||||
| 2730 | auto Res = UniquePositions.try_emplace(V, UniqueValues.size()); | ||||
| 2731 | ReuseShuffleIndicies.emplace_back(Res.first->second); | ||||
| 2732 | if (Res.second) | ||||
| 2733 | UniqueValues.emplace_back(V); | ||||
| 2734 | } | ||||
| 2735 | size_t NumUniqueScalarValues = UniqueValues.size(); | ||||
| 2736 | if (NumUniqueScalarValues == VL.size()) { | ||||
| 2737 | ReuseShuffleIndicies.clear(); | ||||
| 2738 | } else { | ||||
| 2739 | LLVM_DEBUG(dbgs() << "SLP: Shuffle for reused scalars.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Shuffle for reused scalars.\n" ; } } while (false); | ||||
| 2740 | if (NumUniqueScalarValues <= 1 || | ||||
| 2741 | !llvm::isPowerOf2_32(NumUniqueScalarValues)) { | ||||
| 2742 | LLVM_DEBUG(dbgs() << "SLP: Scalar used twice in bundle.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Scalar used twice in bundle.\n" ; } } while (false); | ||||
| 2743 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx); | ||||
| 2744 | return; | ||||
| 2745 | } | ||||
| 2746 | VL = UniqueValues; | ||||
| 2747 | } | ||||
| 2748 | |||||
| 2749 | auto &BSRef = BlocksSchedules[BB]; | ||||
| 2750 | if (!BSRef) | ||||
| 2751 | BSRef = std::make_unique<BlockScheduling>(BB); | ||||
| 2752 | |||||
| 2753 | BlockScheduling &BS = *BSRef.get(); | ||||
| 2754 | |||||
| 2755 | Optional<ScheduleData *> Bundle = BS.tryScheduleBundle(VL, this, S); | ||||
| 2756 | if (!Bundle) { | ||||
| 2757 | LLVM_DEBUG(dbgs() << "SLP: We are not able to schedule this bundle!\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: We are not able to schedule this bundle!\n" ; } } while (false); | ||||
| 2758 | assert((!BS.getScheduleData(VL0) ||(static_cast <bool> ((!BS.getScheduleData(VL0) || !BS.getScheduleData (VL0)->isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure" ) ? void (0) : __assert_fail ("(!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2760, __extension__ __PRETTY_FUNCTION__)) | ||||
| 2759 | !BS.getScheduleData(VL0)->isPartOfBundle()) &&(static_cast <bool> ((!BS.getScheduleData(VL0) || !BS.getScheduleData (VL0)->isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure" ) ? void (0) : __assert_fail ("(!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2760, __extension__ __PRETTY_FUNCTION__)) | ||||
| 2760 | "tryScheduleBundle should cancelScheduling on failure")(static_cast <bool> ((!BS.getScheduleData(VL0) || !BS.getScheduleData (VL0)->isPartOfBundle()) && "tryScheduleBundle should cancelScheduling on failure" ) ? void (0) : __assert_fail ("(!BS.getScheduleData(VL0) || !BS.getScheduleData(VL0)->isPartOfBundle()) && \"tryScheduleBundle should cancelScheduling on failure\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 2760, __extension__ __PRETTY_FUNCTION__)); | ||||
| 2761 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 2762 | ReuseShuffleIndicies); | ||||
| 2763 | return; | ||||
| 2764 | } | ||||
| 2765 | LLVM_DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: We are able to schedule this bundle.\n" ; } } while (false); | ||||
| 2766 | |||||
| 2767 | unsigned ShuffleOrOp = S.isAltShuffle() ? | ||||
| 2768 | (unsigned) Instruction::ShuffleVector : S.getOpcode(); | ||||
| 2769 | switch (ShuffleOrOp) { | ||||
| 2770 | case Instruction::PHI: { | ||||
| 2771 | auto *PH = cast<PHINode>(VL0); | ||||
| 2772 | |||||
| 2773 | // Check for terminator values (e.g. invoke). | ||||
| 2774 | for (Value *V : VL) | ||||
| 2775 | for (unsigned I = 0, E = PH->getNumIncomingValues(); I < E; ++I) { | ||||
| 2776 | Instruction *Term = dyn_cast<Instruction>( | ||||
| 2777 | cast<PHINode>(V)->getIncomingValueForBlock( | ||||
| 2778 | PH->getIncomingBlock(I))); | ||||
| 2779 | if (Term && Term->isTerminator()) { | ||||
| 2780 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to swizzle PHINodes (terminator use).\n" ; } } while (false) | ||||
| 2781 | << "SLP: Need to swizzle PHINodes (terminator use).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Need to swizzle PHINodes (terminator use).\n" ; } } while (false); | ||||
| 2782 | BS.cancelScheduling(VL, VL0); | ||||
| 2783 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 2784 | ReuseShuffleIndicies); | ||||
| 2785 | return; | ||||
| 2786 | } | ||||
| 2787 | } | ||||
| 2788 | |||||
| 2789 | TreeEntry *TE = | ||||
| 2790 | newTreeEntry(VL, Bundle, S, UserTreeIdx, ReuseShuffleIndicies); | ||||
| 2791 | LLVM_DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of PHINodes.\n" ; } } while (false); | ||||
| 2792 | |||||
| 2793 | // Keeps the reordered operands to avoid code duplication. | ||||
| 2794 | SmallVector<ValueList, 2> OperandsVec; | ||||
| 2795 | for (unsigned I = 0, E = PH->getNumIncomingValues(); I < E; ++I) { | ||||
| 2796 | ValueList Operands; | ||||
| 2797 | // Prepare the operand vector. | ||||
| 2798 | for (Value *V : VL) | ||||
| 2799 | Operands.push_back(cast<PHINode>(V)->getIncomingValueForBlock( | ||||
| 2800 | PH->getIncomingBlock(I))); | ||||
| 2801 | TE->setOperand(I, Operands); | ||||
| 2802 | OperandsVec.push_back(Operands); | ||||
| 2803 | } | ||||
| 2804 | for (unsigned OpIdx = 0, OpE = OperandsVec.size(); OpIdx != OpE; ++OpIdx) | ||||
| 2805 | buildTree_rec(OperandsVec[OpIdx], Depth + 1, {TE, OpIdx}); | ||||
| 2806 | return; | ||||
| 2807 | } | ||||
| 2808 | case Instruction::ExtractValue: | ||||
| 2809 | case Instruction::ExtractElement: { | ||||
| 2810 | OrdersType CurrentOrder; | ||||
| 2811 | bool Reuse = canReuseExtract(VL, VL0, CurrentOrder); | ||||
| 2812 | if (Reuse) { | ||||
| 2813 | LLVM_DEBUG(dbgs() << "SLP: Reusing or shuffling extract sequence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Reusing or shuffling extract sequence.\n" ; } } while (false); | ||||
| 2814 | ++NumOpsWantToKeepOriginalOrder; | ||||
| 2815 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
| 2816 | ReuseShuffleIndicies); | ||||
| 2817 | // This is a special case, as it does not gather, but at the same time | ||||
| 2818 | // we are not extending buildTree_rec() towards the operands. | ||||
| 2819 | ValueList Op0; | ||||
| 2820 | Op0.assign(VL.size(), VL0->getOperand(0)); | ||||
| 2821 | VectorizableTree.back()->setOperand(0, Op0); | ||||
| 2822 | return; | ||||
| 2823 | } | ||||
| 2824 | if (!CurrentOrder.empty()) { | ||||
| 2825 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: Reusing or shuffling of reordered extract sequence " "with order"; for (unsigned Idx : CurrentOrder) dbgs() << " " << Idx; dbgs() << "\n"; }; } } while (false) | ||||
| 2826 | dbgs() << "SLP: Reusing or shuffling of reordered extract sequence "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: Reusing or shuffling of reordered extract sequence " "with order"; for (unsigned Idx : CurrentOrder) dbgs() << " " << Idx; dbgs() << "\n"; }; } } while (false) | ||||
| 2827 | "with order";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: Reusing or shuffling of reordered extract sequence " "with order"; for (unsigned Idx : CurrentOrder) dbgs() << " " << Idx; dbgs() << "\n"; }; } } while (false) | ||||
| 2828 | for (unsigned Idx : CurrentOrder)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: Reusing or shuffling of reordered extract sequence " "with order"; for (unsigned Idx : CurrentOrder) dbgs() << " " << Idx; dbgs() << "\n"; }; } } while (false) | ||||
| 2829 | dbgs() << " " << Idx;do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: Reusing or shuffling of reordered extract sequence " "with order"; for (unsigned Idx : CurrentOrder) dbgs() << " " << Idx; dbgs() << "\n"; }; } } while (false) | ||||
| 2830 | dbgs() << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: Reusing or shuffling of reordered extract sequence " "with order"; for (unsigned Idx : CurrentOrder) dbgs() << " " << Idx; dbgs() << "\n"; }; } } while (false) | ||||
| 2831 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: Reusing or shuffling of reordered extract sequence " "with order"; for (unsigned Idx : CurrentOrder) dbgs() << " " << Idx; dbgs() << "\n"; }; } } while (false); | ||||
| 2832 | // Insert new order with initial value 0, if it does not exist, | ||||
| 2833 | // otherwise return the iterator to the existing one. | ||||
| 2834 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
| 2835 | ReuseShuffleIndicies, CurrentOrder); | ||||
| 2836 | findRootOrder(CurrentOrder); | ||||
| 2837 | ++NumOpsWantToKeepOrder[CurrentOrder]; | ||||
| 2838 | // This is a special case, as it does not gather, but at the same time | ||||
| 2839 | // we are not extending buildTree_rec() towards the operands. | ||||
| 2840 | ValueList Op0; | ||||
| 2841 | Op0.assign(VL.size(), VL0->getOperand(0)); | ||||
| 2842 | VectorizableTree.back()->setOperand(0, Op0); | ||||
| 2843 | return; | ||||
| 2844 | } | ||||
| 2845 | LLVM_DEBUG(dbgs() << "SLP: Gather extract sequence.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gather extract sequence.\n"; } } while (false); | ||||
| 2846 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 2847 | ReuseShuffleIndicies); | ||||
| 2848 | BS.cancelScheduling(VL, VL0); | ||||
| 2849 | return; | ||||
| 2850 | } | ||||
| 2851 | case Instruction::Load: { | ||||
| 2852 | // Check that a vectorized load would load the same memory as a scalar | ||||
| 2853 | // load. For example, we don't want to vectorize loads that are smaller | ||||
| 2854 | // than 8-bit. Even though we have a packed struct {<i2, i2, i2, i2>} LLVM | ||||
| 2855 | // treats loading/storing it as an i8 struct. If we vectorize loads/stores | ||||
| 2856 | // from such a struct, we read/write packed bits disagreeing with the | ||||
| 2857 | // unvectorized version. | ||||
| 2858 | Type *ScalarTy = VL0->getType(); | ||||
| 2859 | |||||
| 2860 | if (DL->getTypeSizeInBits(ScalarTy) != | ||||
| 2861 | DL->getTypeAllocSizeInBits(ScalarTy)) { | ||||
| 2862 | BS.cancelScheduling(VL, VL0); | ||||
| 2863 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 2864 | ReuseShuffleIndicies); | ||||
| 2865 | LLVM_DEBUG(dbgs() << "SLP: Gathering loads of non-packed type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering loads of non-packed type.\n" ; } } while (false); | ||||
| 2866 | return; | ||||
| 2867 | } | ||||
| 2868 | |||||
| 2869 | // Make sure all loads in the bundle are simple - we can't vectorize | ||||
| 2870 | // atomic or volatile loads. | ||||
| 2871 | SmallVector<Value *, 4> PointerOps(VL.size()); | ||||
| 2872 | auto POIter = PointerOps.begin(); | ||||
| 2873 | for (Value *V : VL) { | ||||
| 2874 | auto *L = cast<LoadInst>(V); | ||||
| 2875 | if (!L->isSimple()) { | ||||
| 2876 | BS.cancelScheduling(VL, VL0); | ||||
| 2877 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 2878 | ReuseShuffleIndicies); | ||||
| 2879 | LLVM_DEBUG(dbgs() << "SLP: Gathering non-simple loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering non-simple loads.\n" ; } } while (false); | ||||
| 2880 | return; | ||||
| 2881 | } | ||||
| 2882 | *POIter = L->getPointerOperand(); | ||||
| 2883 | ++POIter; | ||||
| 2884 | } | ||||
| 2885 | |||||
| 2886 | OrdersType CurrentOrder; | ||||
| 2887 | // Check the order of pointer operands. | ||||
| 2888 | if (llvm::sortPtrAccesses(PointerOps, *DL, *SE, CurrentOrder)) { | ||||
| 2889 | Value *Ptr0; | ||||
| 2890 | Value *PtrN; | ||||
| 2891 | if (CurrentOrder.empty()) { | ||||
| 2892 | Ptr0 = PointerOps.front(); | ||||
| 2893 | PtrN = PointerOps.back(); | ||||
| 2894 | } else { | ||||
| 2895 | Ptr0 = PointerOps[CurrentOrder.front()]; | ||||
| 2896 | PtrN = PointerOps[CurrentOrder.back()]; | ||||
| 2897 | } | ||||
| 2898 | Optional<int> Diff = getPointersDiff(Ptr0, PtrN, *DL, *SE); | ||||
| 2899 | // Check that the sorted loads are consecutive. | ||||
| 2900 | if (static_cast<unsigned>(*Diff) == VL.size() - 1) { | ||||
| 2901 | if (CurrentOrder.empty()) { | ||||
| 2902 | // Original loads are consecutive and does not require reordering. | ||||
| 2903 | ++NumOpsWantToKeepOriginalOrder; | ||||
| 2904 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, | ||||
| 2905 | UserTreeIdx, ReuseShuffleIndicies); | ||||
| 2906 | TE->setOperandsInOrder(); | ||||
| 2907 | LLVM_DEBUG(dbgs() << "SLP: added a vector of loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of loads.\n"; } } while (false); | ||||
| 2908 | } else { | ||||
| 2909 | // Need to reorder. | ||||
| 2910 | TreeEntry *TE = | ||||
| 2911 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
| 2912 | ReuseShuffleIndicies, CurrentOrder); | ||||
| 2913 | TE->setOperandsInOrder(); | ||||
| 2914 | LLVM_DEBUG(dbgs() << "SLP: added a vector of jumbled loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of jumbled loads.\n" ; } } while (false); | ||||
| 2915 | findRootOrder(CurrentOrder); | ||||
| 2916 | ++NumOpsWantToKeepOrder[CurrentOrder]; | ||||
| 2917 | } | ||||
| 2918 | return; | ||||
| 2919 | } | ||||
| 2920 | // Vectorizing non-consecutive loads with `llvm.masked.gather`. | ||||
| 2921 | TreeEntry *TE = newTreeEntry(VL, TreeEntry::ScatterVectorize, Bundle, S, | ||||
| 2922 | UserTreeIdx, ReuseShuffleIndicies); | ||||
| 2923 | TE->setOperandsInOrder(); | ||||
| 2924 | buildTree_rec(PointerOps, Depth + 1, {TE, 0}); | ||||
| 2925 | LLVM_DEBUG(dbgs() << "SLP: added a vector of non-consecutive loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of non-consecutive loads.\n" ; } } while (false); | ||||
| 2926 | return; | ||||
| 2927 | } | ||||
| 2928 | |||||
| 2929 | LLVM_DEBUG(dbgs() << "SLP: Gathering non-consecutive loads.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering non-consecutive loads.\n" ; } } while (false); | ||||
| 2930 | BS.cancelScheduling(VL, VL0); | ||||
| 2931 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 2932 | ReuseShuffleIndicies); | ||||
| 2933 | return; | ||||
| 2934 | } | ||||
| 2935 | case Instruction::ZExt: | ||||
| 2936 | case Instruction::SExt: | ||||
| 2937 | case Instruction::FPToUI: | ||||
| 2938 | case Instruction::FPToSI: | ||||
| 2939 | case Instruction::FPExt: | ||||
| 2940 | case Instruction::PtrToInt: | ||||
| 2941 | case Instruction::IntToPtr: | ||||
| 2942 | case Instruction::SIToFP: | ||||
| 2943 | case Instruction::UIToFP: | ||||
| 2944 | case Instruction::Trunc: | ||||
| 2945 | case Instruction::FPTrunc: | ||||
| 2946 | case Instruction::BitCast: { | ||||
| 2947 | Type *SrcTy = VL0->getOperand(0)->getType(); | ||||
| 2948 | for (Value *V : VL) { | ||||
| 2949 | Type *Ty = cast<Instruction>(V)->getOperand(0)->getType(); | ||||
| 2950 | if (Ty != SrcTy || !isValidElementType(Ty)) { | ||||
| 2951 | BS.cancelScheduling(VL, VL0); | ||||
| 2952 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 2953 | ReuseShuffleIndicies); | ||||
| 2954 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering casts with different src types.\n" ; } } while (false) | ||||
| 2955 | << "SLP: Gathering casts with different src types.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering casts with different src types.\n" ; } } while (false); | ||||
| 2956 | return; | ||||
| 2957 | } | ||||
| 2958 | } | ||||
| 2959 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
| 2960 | ReuseShuffleIndicies); | ||||
| 2961 | LLVM_DEBUG(dbgs() << "SLP: added a vector of casts.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of casts.\n"; } } while (false); | ||||
| 2962 | |||||
| 2963 | TE->setOperandsInOrder(); | ||||
| 2964 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | ||||
| 2965 | ValueList Operands; | ||||
| 2966 | // Prepare the operand vector. | ||||
| 2967 | for (Value *V : VL) | ||||
| 2968 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); | ||||
| 2969 | |||||
| 2970 | buildTree_rec(Operands, Depth + 1, {TE, i}); | ||||
| 2971 | } | ||||
| 2972 | return; | ||||
| 2973 | } | ||||
| 2974 | case Instruction::ICmp: | ||||
| 2975 | case Instruction::FCmp: { | ||||
| 2976 | // Check that all of the compares have the same predicate. | ||||
| 2977 | CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate(); | ||||
| 2978 | CmpInst::Predicate SwapP0 = CmpInst::getSwappedPredicate(P0); | ||||
| 2979 | Type *ComparedTy = VL0->getOperand(0)->getType(); | ||||
| 2980 | for (Value *V : VL) { | ||||
| 2981 | CmpInst *Cmp = cast<CmpInst>(V); | ||||
| 2982 | if ((Cmp->getPredicate() != P0 && Cmp->getPredicate() != SwapP0) || | ||||
| 2983 | Cmp->getOperand(0)->getType() != ComparedTy) { | ||||
| 2984 | BS.cancelScheduling(VL, VL0); | ||||
| 2985 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 2986 | ReuseShuffleIndicies); | ||||
| 2987 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering cmp with different predicate.\n" ; } } while (false) | ||||
| 2988 | << "SLP: Gathering cmp with different predicate.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering cmp with different predicate.\n" ; } } while (false); | ||||
| 2989 | return; | ||||
| 2990 | } | ||||
| 2991 | } | ||||
| 2992 | |||||
| 2993 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
| 2994 | ReuseShuffleIndicies); | ||||
| 2995 | LLVM_DEBUG(dbgs() << "SLP: added a vector of compares.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of compares.\n" ; } } while (false); | ||||
| 2996 | |||||
| 2997 | ValueList Left, Right; | ||||
| 2998 | if (cast<CmpInst>(VL0)->isCommutative()) { | ||||
| 2999 | // Commutative predicate - collect + sort operands of the instructions | ||||
| 3000 | // so that each side is more likely to have the same opcode. | ||||
| 3001 | assert(P0 == SwapP0 && "Commutative Predicate mismatch")(static_cast <bool> (P0 == SwapP0 && "Commutative Predicate mismatch" ) ? void (0) : __assert_fail ("P0 == SwapP0 && \"Commutative Predicate mismatch\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3001, __extension__ __PRETTY_FUNCTION__)); | ||||
| 3002 | reorderInputsAccordingToOpcode(VL, Left, Right, *DL, *SE, *this); | ||||
| 3003 | } else { | ||||
| 3004 | // Collect operands - commute if it uses the swapped predicate. | ||||
| 3005 | for (Value *V : VL) { | ||||
| 3006 | auto *Cmp = cast<CmpInst>(V); | ||||
| 3007 | Value *LHS = Cmp->getOperand(0); | ||||
| 3008 | Value *RHS = Cmp->getOperand(1); | ||||
| 3009 | if (Cmp->getPredicate() != P0) | ||||
| 3010 | std::swap(LHS, RHS); | ||||
| 3011 | Left.push_back(LHS); | ||||
| 3012 | Right.push_back(RHS); | ||||
| 3013 | } | ||||
| 3014 | } | ||||
| 3015 | TE->setOperand(0, Left); | ||||
| 3016 | TE->setOperand(1, Right); | ||||
| 3017 | buildTree_rec(Left, Depth + 1, {TE, 0}); | ||||
| 3018 | buildTree_rec(Right, Depth + 1, {TE, 1}); | ||||
| 3019 | return; | ||||
| 3020 | } | ||||
| 3021 | case Instruction::Select: | ||||
| 3022 | case Instruction::FNeg: | ||||
| 3023 | case Instruction::Add: | ||||
| 3024 | case Instruction::FAdd: | ||||
| 3025 | case Instruction::Sub: | ||||
| 3026 | case Instruction::FSub: | ||||
| 3027 | case Instruction::Mul: | ||||
| 3028 | case Instruction::FMul: | ||||
| 3029 | case Instruction::UDiv: | ||||
| 3030 | case Instruction::SDiv: | ||||
| 3031 | case Instruction::FDiv: | ||||
| 3032 | case Instruction::URem: | ||||
| 3033 | case Instruction::SRem: | ||||
| 3034 | case Instruction::FRem: | ||||
| 3035 | case Instruction::Shl: | ||||
| 3036 | case Instruction::LShr: | ||||
| 3037 | case Instruction::AShr: | ||||
| 3038 | case Instruction::And: | ||||
| 3039 | case Instruction::Or: | ||||
| 3040 | case Instruction::Xor: { | ||||
| 3041 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
| 3042 | ReuseShuffleIndicies); | ||||
| 3043 | LLVM_DEBUG(dbgs() << "SLP: added a vector of un/bin op.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of un/bin op.\n" ; } } while (false); | ||||
| 3044 | |||||
| 3045 | // Sort operands of the instructions so that each side is more likely to | ||||
| 3046 | // have the same opcode. | ||||
| 3047 | if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) { | ||||
| 3048 | ValueList Left, Right; | ||||
| 3049 | reorderInputsAccordingToOpcode(VL, Left, Right, *DL, *SE, *this); | ||||
| 3050 | TE->setOperand(0, Left); | ||||
| 3051 | TE->setOperand(1, Right); | ||||
| 3052 | buildTree_rec(Left, Depth + 1, {TE, 0}); | ||||
| 3053 | buildTree_rec(Right, Depth + 1, {TE, 1}); | ||||
| 3054 | return; | ||||
| 3055 | } | ||||
| 3056 | |||||
| 3057 | TE->setOperandsInOrder(); | ||||
| 3058 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | ||||
| 3059 | ValueList Operands; | ||||
| 3060 | // Prepare the operand vector. | ||||
| 3061 | for (Value *V : VL) | ||||
| 3062 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); | ||||
| 3063 | |||||
| 3064 | buildTree_rec(Operands, Depth + 1, {TE, i}); | ||||
| 3065 | } | ||||
| 3066 | return; | ||||
| 3067 | } | ||||
| 3068 | case Instruction::GetElementPtr: { | ||||
| 3069 | // We don't combine GEPs with complicated (nested) indexing. | ||||
| 3070 | for (Value *V : VL) { | ||||
| 3071 | if (cast<Instruction>(V)->getNumOperands() != 2) { | ||||
| 3072 | LLVM_DEBUG(dbgs() << "SLP: not-vectorizable GEP (nested indexes).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (nested indexes).\n" ; } } while (false); | ||||
| 3073 | BS.cancelScheduling(VL, VL0); | ||||
| 3074 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 3075 | ReuseShuffleIndicies); | ||||
| 3076 | return; | ||||
| 3077 | } | ||||
| 3078 | } | ||||
| 3079 | |||||
| 3080 | // We can't combine several GEPs into one vector if they operate on | ||||
| 3081 | // different types. | ||||
| 3082 | Type *Ty0 = VL0->getOperand(0)->getType(); | ||||
| 3083 | for (Value *V : VL) { | ||||
| 3084 | Type *CurTy = cast<Instruction>(V)->getOperand(0)->getType(); | ||||
| 3085 | if (Ty0 != CurTy) { | ||||
| 3086 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (different types).\n" ; } } while (false) | ||||
| 3087 | << "SLP: not-vectorizable GEP (different types).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (different types).\n" ; } } while (false); | ||||
| 3088 | BS.cancelScheduling(VL, VL0); | ||||
| 3089 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 3090 | ReuseShuffleIndicies); | ||||
| 3091 | return; | ||||
| 3092 | } | ||||
| 3093 | } | ||||
| 3094 | |||||
| 3095 | // We don't combine GEPs with non-constant indexes. | ||||
| 3096 | Type *Ty1 = VL0->getOperand(1)->getType(); | ||||
| 3097 | for (Value *V : VL) { | ||||
| 3098 | auto Op = cast<Instruction>(V)->getOperand(1); | ||||
| 3099 | if (!isa<ConstantInt>(Op) || | ||||
| 3100 | (Op->getType() != Ty1 && | ||||
| 3101 | Op->getType()->getScalarSizeInBits() > | ||||
| 3102 | DL->getIndexSizeInBits( | ||||
| 3103 | V->getType()->getPointerAddressSpace()))) { | ||||
| 3104 | LLVM_DEBUG(dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n" ; } } while (false) | ||||
| 3105 | << "SLP: not-vectorizable GEP (non-constant indexes).\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n" ; } } while (false); | ||||
| 3106 | BS.cancelScheduling(VL, VL0); | ||||
| 3107 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 3108 | ReuseShuffleIndicies); | ||||
| 3109 | return; | ||||
| 3110 | } | ||||
| 3111 | } | ||||
| 3112 | |||||
| 3113 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
| 3114 | ReuseShuffleIndicies); | ||||
| 3115 | LLVM_DEBUG(dbgs() << "SLP: added a vector of GEPs.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of GEPs.\n"; } } while (false); | ||||
| 3116 | TE->setOperandsInOrder(); | ||||
| 3117 | for (unsigned i = 0, e = 2; i < e; ++i) { | ||||
| 3118 | ValueList Operands; | ||||
| 3119 | // Prepare the operand vector. | ||||
| 3120 | for (Value *V : VL) | ||||
| 3121 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); | ||||
| 3122 | |||||
| 3123 | buildTree_rec(Operands, Depth + 1, {TE, i}); | ||||
| 3124 | } | ||||
| 3125 | return; | ||||
| 3126 | } | ||||
| 3127 | case Instruction::Store: { | ||||
| 3128 | // Check if the stores are consecutive or if we need to swizzle them. | ||||
| 3129 | llvm::Type *ScalarTy = cast<StoreInst>(VL0)->getValueOperand()->getType(); | ||||
| 3130 | // Avoid types that are padded when being allocated as scalars, while | ||||
| 3131 | // being packed together in a vector (such as i1). | ||||
| 3132 | if (DL->getTypeSizeInBits(ScalarTy) != | ||||
| 3133 | DL->getTypeAllocSizeInBits(ScalarTy)) { | ||||
| 3134 | BS.cancelScheduling(VL, VL0); | ||||
| 3135 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 3136 | ReuseShuffleIndicies); | ||||
| 3137 | LLVM_DEBUG(dbgs() << "SLP: Gathering stores of non-packed type.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering stores of non-packed type.\n" ; } } while (false); | ||||
| 3138 | return; | ||||
| 3139 | } | ||||
| 3140 | // Make sure all stores in the bundle are simple - we can't vectorize | ||||
| 3141 | // atomic or volatile stores. | ||||
| 3142 | SmallVector<Value *, 4> PointerOps(VL.size()); | ||||
| 3143 | ValueList Operands(VL.size()); | ||||
| 3144 | auto POIter = PointerOps.begin(); | ||||
| 3145 | auto OIter = Operands.begin(); | ||||
| 3146 | for (Value *V : VL) { | ||||
| 3147 | auto *SI = cast<StoreInst>(V); | ||||
| 3148 | if (!SI->isSimple()) { | ||||
| 3149 | BS.cancelScheduling(VL, VL0); | ||||
| 3150 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 3151 | ReuseShuffleIndicies); | ||||
| 3152 | LLVM_DEBUG(dbgs() << "SLP: Gathering non-simple stores.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering non-simple stores.\n" ; } } while (false); | ||||
| 3153 | return; | ||||
| 3154 | } | ||||
| 3155 | *POIter = SI->getPointerOperand(); | ||||
| 3156 | *OIter = SI->getValueOperand(); | ||||
| 3157 | ++POIter; | ||||
| 3158 | ++OIter; | ||||
| 3159 | } | ||||
| 3160 | |||||
| 3161 | OrdersType CurrentOrder; | ||||
| 3162 | // Check the order of pointer operands. | ||||
| 3163 | if (llvm::sortPtrAccesses(PointerOps, *DL, *SE, CurrentOrder)) { | ||||
| 3164 | Value *Ptr0; | ||||
| 3165 | Value *PtrN; | ||||
| 3166 | if (CurrentOrder.empty()) { | ||||
| 3167 | Ptr0 = PointerOps.front(); | ||||
| 3168 | PtrN = PointerOps.back(); | ||||
| 3169 | } else { | ||||
| 3170 | Ptr0 = PointerOps[CurrentOrder.front()]; | ||||
| 3171 | PtrN = PointerOps[CurrentOrder.back()]; | ||||
| 3172 | } | ||||
| 3173 | Optional<int> Dist = getPointersDiff(Ptr0, PtrN, *DL, *SE); | ||||
| 3174 | // Check that the sorted pointer operands are consecutive. | ||||
| 3175 | if (static_cast<unsigned>(*Dist) == VL.size() - 1) { | ||||
| 3176 | if (CurrentOrder.empty()) { | ||||
| 3177 | // Original stores are consecutive and does not require reordering. | ||||
| 3178 | ++NumOpsWantToKeepOriginalOrder; | ||||
| 3179 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, | ||||
| 3180 | UserTreeIdx, ReuseShuffleIndicies); | ||||
| 3181 | TE->setOperandsInOrder(); | ||||
| 3182 | buildTree_rec(Operands, Depth + 1, {TE, 0}); | ||||
| 3183 | LLVM_DEBUG(dbgs() << "SLP: added a vector of stores.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of stores.\n" ; } } while (false); | ||||
| 3184 | } else { | ||||
| 3185 | TreeEntry *TE = | ||||
| 3186 | newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
| 3187 | ReuseShuffleIndicies, CurrentOrder); | ||||
| 3188 | TE->setOperandsInOrder(); | ||||
| 3189 | buildTree_rec(Operands, Depth + 1, {TE, 0}); | ||||
| 3190 | LLVM_DEBUG(dbgs() << "SLP: added a vector of jumbled stores.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a vector of jumbled stores.\n" ; } } while (false); | ||||
| 3191 | findRootOrder(CurrentOrder); | ||||
| 3192 | ++NumOpsWantToKeepOrder[CurrentOrder]; | ||||
| 3193 | } | ||||
| 3194 | return; | ||||
| 3195 | } | ||||
| 3196 | } | ||||
| 3197 | |||||
| 3198 | BS.cancelScheduling(VL, VL0); | ||||
| 3199 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 3200 | ReuseShuffleIndicies); | ||||
| 3201 | LLVM_DEBUG(dbgs() << "SLP: Non-consecutive store.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Non-consecutive store.\n"; } } while (false); | ||||
| 3202 | return; | ||||
| 3203 | } | ||||
| 3204 | case Instruction::Call: { | ||||
| 3205 | // Check if the calls are all to the same vectorizable intrinsic or | ||||
| 3206 | // library function. | ||||
| 3207 | CallInst *CI = cast<CallInst>(VL0); | ||||
| 3208 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||
| 3209 | |||||
| 3210 | VFShape Shape = VFShape::get( | ||||
| 3211 | *CI, ElementCount::getFixed(static_cast<unsigned int>(VL.size())), | ||||
| 3212 | false /*HasGlobalPred*/); | ||||
| 3213 | Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); | ||||
| 3214 | |||||
| 3215 | if (!VecFunc && !isTriviallyVectorizable(ID)) { | ||||
| 3216 | BS.cancelScheduling(VL, VL0); | ||||
| 3217 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 3218 | ReuseShuffleIndicies); | ||||
| 3219 | LLVM_DEBUG(dbgs() << "SLP: Non-vectorizable call.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Non-vectorizable call.\n"; } } while (false); | ||||
| 3220 | return; | ||||
| 3221 | } | ||||
| 3222 | Function *F = CI->getCalledFunction(); | ||||
| 3223 | unsigned NumArgs = CI->getNumArgOperands(); | ||||
| 3224 | SmallVector<Value*, 4> ScalarArgs(NumArgs, nullptr); | ||||
| 3225 | for (unsigned j = 0; j != NumArgs; ++j) | ||||
| 3226 | if (hasVectorInstrinsicScalarOpd(ID, j)) | ||||
| 3227 | ScalarArgs[j] = CI->getArgOperand(j); | ||||
| 3228 | for (Value *V : VL) { | ||||
| 3229 | CallInst *CI2 = dyn_cast<CallInst>(V); | ||||
| 3230 | if (!CI2 || CI2->getCalledFunction() != F || | ||||
| 3231 | getVectorIntrinsicIDForCall(CI2, TLI) != ID || | ||||
| 3232 | (VecFunc && | ||||
| 3233 | VecFunc != VFDatabase(*CI2).getVectorizedFunction(Shape)) || | ||||
| 3234 | !CI->hasIdenticalOperandBundleSchema(*CI2)) { | ||||
| 3235 | BS.cancelScheduling(VL, VL0); | ||||
| 3236 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 3237 | ReuseShuffleIndicies); | ||||
| 3238 | LLVM_DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *Vdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched calls:" << * CI << "!=" << *V << "\n"; } } while (false) | ||||
| 3239 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched calls:" << * CI << "!=" << *V << "\n"; } } while (false); | ||||
| 3240 | return; | ||||
| 3241 | } | ||||
| 3242 | // Some intrinsics have scalar arguments and should be same in order for | ||||
| 3243 | // them to be vectorized. | ||||
| 3244 | for (unsigned j = 0; j != NumArgs; ++j) { | ||||
| 3245 | if (hasVectorInstrinsicScalarOpd(ID, j)) { | ||||
| 3246 | Value *A1J = CI2->getArgOperand(j); | ||||
| 3247 | if (ScalarArgs[j] != A1J) { | ||||
| 3248 | BS.cancelScheduling(VL, VL0); | ||||
| 3249 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 3250 | ReuseShuffleIndicies); | ||||
| 3251 | LLVM_DEBUG(dbgs() << "SLP: mismatched arguments in call:" << *CIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument " << ScalarArgs[j] << "!=" << A1J << "\n"; } } while (false) | ||||
| 3252 | << " argument " << ScalarArgs[j] << "!=" << A1Jdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument " << ScalarArgs[j] << "!=" << A1J << "\n"; } } while (false) | ||||
| 3253 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched arguments in call:" << *CI << " argument " << ScalarArgs[j] << "!=" << A1J << "\n"; } } while (false); | ||||
| 3254 | return; | ||||
| 3255 | } | ||||
| 3256 | } | ||||
| 3257 | } | ||||
| 3258 | // Verify that the bundle operands are identical between the two calls. | ||||
| 3259 | if (CI->hasOperandBundles() && | ||||
| 3260 | !std::equal(CI->op_begin() + CI->getBundleOperandsStartIndex(), | ||||
| 3261 | CI->op_begin() + CI->getBundleOperandsEndIndex(), | ||||
| 3262 | CI2->op_begin() + CI2->getBundleOperandsStartIndex())) { | ||||
| 3263 | BS.cancelScheduling(VL, VL0); | ||||
| 3264 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 3265 | ReuseShuffleIndicies); | ||||
| 3266 | LLVM_DEBUG(dbgs() << "SLP: mismatched bundle operands in calls:"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched bundle operands in calls:" << *CI << "!=" << *V << '\n'; } } while (false) | ||||
| 3267 | << *CI << "!=" << *V << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: mismatched bundle operands in calls:" << *CI << "!=" << *V << '\n'; } } while (false); | ||||
| 3268 | return; | ||||
| 3269 | } | ||||
| 3270 | } | ||||
| 3271 | |||||
| 3272 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
| 3273 | ReuseShuffleIndicies); | ||||
| 3274 | TE->setOperandsInOrder(); | ||||
| 3275 | for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) { | ||||
| 3276 | ValueList Operands; | ||||
| 3277 | // Prepare the operand vector. | ||||
| 3278 | for (Value *V : VL) { | ||||
| 3279 | auto *CI2 = cast<CallInst>(V); | ||||
| 3280 | Operands.push_back(CI2->getArgOperand(i)); | ||||
| 3281 | } | ||||
| 3282 | buildTree_rec(Operands, Depth + 1, {TE, i}); | ||||
| 3283 | } | ||||
| 3284 | return; | ||||
| 3285 | } | ||||
| 3286 | case Instruction::ShuffleVector: { | ||||
| 3287 | // If this is not an alternate sequence of opcode like add-sub | ||||
| 3288 | // then do not vectorize this instruction. | ||||
| 3289 | if (!S.isAltShuffle()) { | ||||
| 3290 | BS.cancelScheduling(VL, VL0); | ||||
| 3291 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 3292 | ReuseShuffleIndicies); | ||||
| 3293 | LLVM_DEBUG(dbgs() << "SLP: ShuffleVector are not vectorized.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: ShuffleVector are not vectorized.\n" ; } } while (false); | ||||
| 3294 | return; | ||||
| 3295 | } | ||||
| 3296 | TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx, | ||||
| 3297 | ReuseShuffleIndicies); | ||||
| 3298 | LLVM_DEBUG(dbgs() << "SLP: added a ShuffleVector op.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: added a ShuffleVector op.\n" ; } } while (false); | ||||
| 3299 | |||||
| 3300 | // Reorder operands if reordering would enable vectorization. | ||||
| 3301 | if (isa<BinaryOperator>(VL0)) { | ||||
| 3302 | ValueList Left, Right; | ||||
| 3303 | reorderInputsAccordingToOpcode(VL, Left, Right, *DL, *SE, *this); | ||||
| 3304 | TE->setOperand(0, Left); | ||||
| 3305 | TE->setOperand(1, Right); | ||||
| 3306 | buildTree_rec(Left, Depth + 1, {TE, 0}); | ||||
| 3307 | buildTree_rec(Right, Depth + 1, {TE, 1}); | ||||
| 3308 | return; | ||||
| 3309 | } | ||||
| 3310 | |||||
| 3311 | TE->setOperandsInOrder(); | ||||
| 3312 | for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) { | ||||
| 3313 | ValueList Operands; | ||||
| 3314 | // Prepare the operand vector. | ||||
| 3315 | for (Value *V : VL) | ||||
| 3316 | Operands.push_back(cast<Instruction>(V)->getOperand(i)); | ||||
| 3317 | |||||
| 3318 | buildTree_rec(Operands, Depth + 1, {TE, i}); | ||||
| 3319 | } | ||||
| 3320 | return; | ||||
| 3321 | } | ||||
| 3322 | default: | ||||
| 3323 | BS.cancelScheduling(VL, VL0); | ||||
| 3324 | newTreeEntry(VL, None /*not vectorized*/, S, UserTreeIdx, | ||||
| 3325 | ReuseShuffleIndicies); | ||||
| 3326 | LLVM_DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Gathering unknown instruction.\n" ; } } while (false); | ||||
| 3327 | return; | ||||
| 3328 | } | ||||
| 3329 | } | ||||
| 3330 | |||||
| 3331 | unsigned BoUpSLP::canMapToVector(Type *T, const DataLayout &DL) const { | ||||
| 3332 | unsigned N = 1; | ||||
| 3333 | Type *EltTy = T; | ||||
| 3334 | |||||
| 3335 | while (isa<StructType>(EltTy) || isa<ArrayType>(EltTy) || | ||||
| 3336 | isa<VectorType>(EltTy)) { | ||||
| 3337 | if (auto *ST = dyn_cast<StructType>(EltTy)) { | ||||
| 3338 | // Check that struct is homogeneous. | ||||
| 3339 | for (const auto *Ty : ST->elements()) | ||||
| 3340 | if (Ty != *ST->element_begin()) | ||||
| 3341 | return 0; | ||||
| 3342 | N *= ST->getNumElements(); | ||||
| 3343 | EltTy = *ST->element_begin(); | ||||
| 3344 | } else if (auto *AT = dyn_cast<ArrayType>(EltTy)) { | ||||
| 3345 | N *= AT->getNumElements(); | ||||
| 3346 | EltTy = AT->getElementType(); | ||||
| 3347 | } else { | ||||
| 3348 | auto *VT = cast<FixedVectorType>(EltTy); | ||||
| 3349 | N *= VT->getNumElements(); | ||||
| 3350 | EltTy = VT->getElementType(); | ||||
| 3351 | } | ||||
| 3352 | } | ||||
| 3353 | |||||
| 3354 | if (!isValidElementType(EltTy)) | ||||
| 3355 | return 0; | ||||
| 3356 | uint64_t VTSize = DL.getTypeStoreSizeInBits(FixedVectorType::get(EltTy, N)); | ||||
| 3357 | if (VTSize < MinVecRegSize || VTSize > MaxVecRegSize || VTSize != DL.getTypeStoreSizeInBits(T)) | ||||
| 3358 | return 0; | ||||
| 3359 | return N; | ||||
| 3360 | } | ||||
| 3361 | |||||
| 3362 | bool BoUpSLP::canReuseExtract(ArrayRef<Value *> VL, Value *OpValue, | ||||
| 3363 | SmallVectorImpl<unsigned> &CurrentOrder) const { | ||||
| 3364 | Instruction *E0 = cast<Instruction>(OpValue); | ||||
| 3365 | assert(E0->getOpcode() == Instruction::ExtractElement ||(static_cast <bool> (E0->getOpcode() == Instruction:: ExtractElement || E0->getOpcode() == Instruction::ExtractValue ) ? void (0) : __assert_fail ("E0->getOpcode() == Instruction::ExtractElement || E0->getOpcode() == Instruction::ExtractValue" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3366, __extension__ __PRETTY_FUNCTION__)) | ||||
| 3366 | E0->getOpcode() == Instruction::ExtractValue)(static_cast <bool> (E0->getOpcode() == Instruction:: ExtractElement || E0->getOpcode() == Instruction::ExtractValue ) ? void (0) : __assert_fail ("E0->getOpcode() == Instruction::ExtractElement || E0->getOpcode() == Instruction::ExtractValue" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3366, __extension__ __PRETTY_FUNCTION__)); | ||||
| 3367 | assert(E0->getOpcode() == getSameOpcode(VL).getOpcode() && "Invalid opcode")(static_cast <bool> (E0->getOpcode() == getSameOpcode (VL).getOpcode() && "Invalid opcode") ? void (0) : __assert_fail ("E0->getOpcode() == getSameOpcode(VL).getOpcode() && \"Invalid opcode\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3367, __extension__ __PRETTY_FUNCTION__)); | ||||
| 3368 | // Check if all of the extracts come from the same vector and from the | ||||
| 3369 | // correct offset. | ||||
| 3370 | Value *Vec = E0->getOperand(0); | ||||
| 3371 | |||||
| 3372 | CurrentOrder.clear(); | ||||
| 3373 | |||||
| 3374 | // We have to extract from a vector/aggregate with the same number of elements. | ||||
| 3375 | unsigned NElts; | ||||
| 3376 | if (E0->getOpcode() == Instruction::ExtractValue) { | ||||
| 3377 | const DataLayout &DL = E0->getModule()->getDataLayout(); | ||||
| 3378 | NElts = canMapToVector(Vec->getType(), DL); | ||||
| 3379 | if (!NElts) | ||||
| 3380 | return false; | ||||
| 3381 | // Check if load can be rewritten as load of vector. | ||||
| 3382 | LoadInst *LI = dyn_cast<LoadInst>(Vec); | ||||
| 3383 | if (!LI || !LI->isSimple() || !LI->hasNUses(VL.size())) | ||||
| 3384 | return false; | ||||
| 3385 | } else { | ||||
| 3386 | NElts = cast<FixedVectorType>(Vec->getType())->getNumElements(); | ||||
| 3387 | } | ||||
| 3388 | |||||
| 3389 | if (NElts != VL.size()) | ||||
| 3390 | return false; | ||||
| 3391 | |||||
| 3392 | // Check that all of the indices extract from the correct offset. | ||||
| 3393 | bool ShouldKeepOrder = true; | ||||
| 3394 | unsigned E = VL.size(); | ||||
| 3395 | // Assign to all items the initial value E + 1 so we can check if the extract | ||||
| 3396 | // instruction index was used already. | ||||
| 3397 | // Also, later we can check that all the indices are used and we have a | ||||
| 3398 | // consecutive access in the extract instructions, by checking that no | ||||
| 3399 | // element of CurrentOrder still has value E + 1. | ||||
| 3400 | CurrentOrder.assign(E, E + 1); | ||||
| 3401 | unsigned I = 0; | ||||
| 3402 | for (; I < E; ++I) { | ||||
| 3403 | auto *Inst = cast<Instruction>(VL[I]); | ||||
| 3404 | if (Inst->getOperand(0) != Vec) | ||||
| 3405 | break; | ||||
| 3406 | Optional<unsigned> Idx = getExtractIndex(Inst); | ||||
| 3407 | if (!Idx) | ||||
| 3408 | break; | ||||
| 3409 | const unsigned ExtIdx = *Idx; | ||||
| 3410 | if (ExtIdx != I) { | ||||
| 3411 | if (ExtIdx >= E || CurrentOrder[ExtIdx] != E + 1) | ||||
| 3412 | break; | ||||
| 3413 | ShouldKeepOrder = false; | ||||
| 3414 | CurrentOrder[ExtIdx] = I; | ||||
| 3415 | } else { | ||||
| 3416 | if (CurrentOrder[I] != E + 1) | ||||
| 3417 | break; | ||||
| 3418 | CurrentOrder[I] = I; | ||||
| 3419 | } | ||||
| 3420 | } | ||||
| 3421 | if (I < E) { | ||||
| 3422 | CurrentOrder.clear(); | ||||
| 3423 | return false; | ||||
| 3424 | } | ||||
| 3425 | |||||
| 3426 | return ShouldKeepOrder; | ||||
| 3427 | } | ||||
| 3428 | |||||
| 3429 | bool BoUpSLP::areAllUsersVectorized(Instruction *I) const { | ||||
| 3430 | return I->hasOneUse() || llvm::all_of(I->users(), [this](User *U) { | ||||
| 3431 | return ScalarToTreeEntry.count(U) > 0; | ||||
| 3432 | }); | ||||
| 3433 | } | ||||
| 3434 | |||||
| 3435 | static std::pair<InstructionCost, InstructionCost> | ||||
| 3436 | getVectorCallCosts(CallInst *CI, FixedVectorType *VecTy, | ||||
| 3437 | TargetTransformInfo *TTI, TargetLibraryInfo *TLI) { | ||||
| 3438 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||
| 3439 | |||||
| 3440 | // Calculate the cost of the scalar and vector calls. | ||||
| 3441 | SmallVector<Type *, 4> VecTys; | ||||
| 3442 | for (Use &Arg : CI->args()) | ||||
| 3443 | VecTys.push_back( | ||||
| 3444 | FixedVectorType::get(Arg->getType(), VecTy->getNumElements())); | ||||
| 3445 | FastMathFlags FMF; | ||||
| 3446 | if (auto *FPCI = dyn_cast<FPMathOperator>(CI)) | ||||
| 3447 | FMF = FPCI->getFastMathFlags(); | ||||
| 3448 | SmallVector<const Value *> Arguments(CI->arg_begin(), CI->arg_end()); | ||||
| 3449 | IntrinsicCostAttributes CostAttrs(ID, VecTy, Arguments, VecTys, FMF, | ||||
| 3450 | dyn_cast<IntrinsicInst>(CI)); | ||||
| 3451 | auto IntrinsicCost = | ||||
| 3452 | TTI->getIntrinsicInstrCost(CostAttrs, TTI::TCK_RecipThroughput); | ||||
| 3453 | |||||
| 3454 | auto Shape = VFShape::get(*CI, ElementCount::getFixed(static_cast<unsigned>( | ||||
| 3455 | VecTy->getNumElements())), | ||||
| 3456 | false /*HasGlobalPred*/); | ||||
| 3457 | Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape); | ||||
| 3458 | auto LibCost = IntrinsicCost; | ||||
| 3459 | if (!CI->isNoBuiltin() && VecFunc) { | ||||
| 3460 | // Calculate the cost of the vector library call. | ||||
| 3461 | // If the corresponding vector call is cheaper, return its cost. | ||||
| 3462 | LibCost = TTI->getCallInstrCost(nullptr, VecTy, VecTys, | ||||
| 3463 | TTI::TCK_RecipThroughput); | ||||
| 3464 | } | ||||
| 3465 | return {IntrinsicCost, LibCost}; | ||||
| 3466 | } | ||||
| 3467 | |||||
| 3468 | /// Compute the cost of creating a vector of type \p VecTy containing the | ||||
| 3469 | /// extracted values from \p VL. | ||||
| 3470 | static InstructionCost | ||||
| 3471 | computeExtractCost(ArrayRef<Value *> VL, FixedVectorType *VecTy, | ||||
| 3472 | TargetTransformInfo::ShuffleKind ShuffleKind, | ||||
| 3473 | ArrayRef<int> Mask, TargetTransformInfo &TTI) { | ||||
| 3474 | unsigned NumOfParts = TTI.getNumberOfParts(VecTy); | ||||
| 3475 | |||||
| 3476 | if (ShuffleKind != TargetTransformInfo::SK_PermuteSingleSrc || !NumOfParts || | ||||
| 3477 | VecTy->getNumElements() < NumOfParts) | ||||
| 3478 | return TTI.getShuffleCost(ShuffleKind, VecTy, Mask); | ||||
| 3479 | |||||
| 3480 | bool AllConsecutive = true; | ||||
| 3481 | unsigned EltsPerVector = VecTy->getNumElements() / NumOfParts; | ||||
| 3482 | unsigned Idx = -1; | ||||
| 3483 | InstructionCost Cost = 0; | ||||
| 3484 | |||||
| 3485 | // Process extracts in blocks of EltsPerVector to check if the source vector | ||||
| 3486 | // operand can be re-used directly. If not, add the cost of creating a shuffle | ||||
| 3487 | // to extract the values into a vector register. | ||||
| 3488 | for (auto *V : VL) { | ||||
| 3489 | ++Idx; | ||||
| 3490 | |||||
| 3491 | // Reached the start of a new vector registers. | ||||
| 3492 | if (Idx % EltsPerVector == 0) { | ||||
| 3493 | AllConsecutive = true; | ||||
| 3494 | continue; | ||||
| 3495 | } | ||||
| 3496 | |||||
| 3497 | // Check all extracts for a vector register on the target directly | ||||
| 3498 | // extract values in order. | ||||
| 3499 | unsigned CurrentIdx = *getExtractIndex(cast<Instruction>(V)); | ||||
| 3500 | unsigned PrevIdx = *getExtractIndex(cast<Instruction>(VL[Idx - 1])); | ||||
| 3501 | AllConsecutive &= PrevIdx + 1 == CurrentIdx && | ||||
| 3502 | CurrentIdx % EltsPerVector == Idx % EltsPerVector; | ||||
| 3503 | |||||
| 3504 | if (AllConsecutive) | ||||
| 3505 | continue; | ||||
| 3506 | |||||
| 3507 | // Skip all indices, except for the last index per vector block. | ||||
| 3508 | if ((Idx + 1) % EltsPerVector != 0 && Idx + 1 != VL.size()) | ||||
| 3509 | continue; | ||||
| 3510 | |||||
| 3511 | // If we have a series of extracts which are not consecutive and hence | ||||
| 3512 | // cannot re-use the source vector register directly, compute the shuffle | ||||
| 3513 | // cost to extract the a vector with EltsPerVector elements. | ||||
| 3514 | Cost += TTI.getShuffleCost( | ||||
| 3515 | TargetTransformInfo::SK_PermuteSingleSrc, | ||||
| 3516 | FixedVectorType::get(VecTy->getElementType(), EltsPerVector)); | ||||
| 3517 | } | ||||
| 3518 | return Cost; | ||||
| 3519 | } | ||||
| 3520 | |||||
| 3521 | InstructionCost BoUpSLP::getEntryCost(const TreeEntry *E) { | ||||
| 3522 | ArrayRef<Value*> VL = E->Scalars; | ||||
| 3523 | |||||
| 3524 | Type *ScalarTy = VL[0]->getType(); | ||||
| 3525 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) | ||||
| 3526 | ScalarTy = SI->getValueOperand()->getType(); | ||||
| 3527 | else if (CmpInst *CI = dyn_cast<CmpInst>(VL[0])) | ||||
| 3528 | ScalarTy = CI->getOperand(0)->getType(); | ||||
| 3529 | auto *VecTy = FixedVectorType::get(ScalarTy, VL.size()); | ||||
| 3530 | TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput; | ||||
| 3531 | |||||
| 3532 | // If we have computed a smaller type for the expression, update VecTy so | ||||
| 3533 | // that the costs will be accurate. | ||||
| 3534 | if (MinBWs.count(VL[0])) | ||||
| 3535 | VecTy = FixedVectorType::get( | ||||
| 3536 | IntegerType::get(F->getContext(), MinBWs[VL[0]].first), VL.size()); | ||||
| 3537 | |||||
| 3538 | unsigned ReuseShuffleNumbers = E->ReuseShuffleIndices.size(); | ||||
| 3539 | bool NeedToShuffleReuses = !E->ReuseShuffleIndices.empty(); | ||||
| 3540 | InstructionCost ReuseShuffleCost = 0; | ||||
| 3541 | if (NeedToShuffleReuses) { | ||||
| 3542 | ReuseShuffleCost = | ||||
| 3543 | TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, VecTy, | ||||
| 3544 | E->ReuseShuffleIndices); | ||||
| 3545 | } | ||||
| 3546 | // FIXME: it tries to fix a problem with MSVC buildbots. | ||||
| 3547 | TargetTransformInfo &TTIRef = *TTI; | ||||
| 3548 | auto &&AdjustExtractsCost = [this, &TTIRef, CostKind, VL, | ||||
| 3549 | VecTy](InstructionCost &Cost, bool IsGather) { | ||||
| 3550 | DenseMap<Value *, int> ExtractVectorsTys; | ||||
| 3551 | for (auto *V : VL) { | ||||
| 3552 | // If all users of instruction are going to be vectorized and this | ||||
| 3553 | // instruction itself is not going to be vectorized, consider this | ||||
| 3554 | // instruction as dead and remove its cost from the final cost of the | ||||
| 3555 | // vectorized tree. | ||||
| 3556 | if (IsGather && (!areAllUsersVectorized(cast<Instruction>(V)) || | ||||
| 3557 | ScalarToTreeEntry.count(V))) | ||||
| 3558 | continue; | ||||
| 3559 | auto *EE = cast<ExtractElementInst>(V); | ||||
| 3560 | unsigned Idx = *getExtractIndex(EE); | ||||
| 3561 | if (TTIRef.getNumberOfParts(VecTy) != | ||||
| 3562 | TTIRef.getNumberOfParts(EE->getVectorOperandType())) { | ||||
| 3563 | auto It = | ||||
| 3564 | ExtractVectorsTys.try_emplace(EE->getVectorOperand(), Idx).first; | ||||
| 3565 | It->getSecond() = std::min<int>(It->second, Idx); | ||||
| 3566 | } | ||||
| 3567 | // Take credit for instruction that will become dead. | ||||
| 3568 | if (EE->hasOneUse()) { | ||||
| 3569 | Instruction *Ext = EE->user_back(); | ||||
| 3570 | if ((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && | ||||
| 3571 | all_of(Ext->users(), | ||||
| 3572 | [](User *U) { return isa<GetElementPtrInst>(U); })) { | ||||
| 3573 | // Use getExtractWithExtendCost() to calculate the cost of | ||||
| 3574 | // extractelement/ext pair. | ||||
| 3575 | Cost -= | ||||
| 3576 | TTIRef.getExtractWithExtendCost(Ext->getOpcode(), Ext->getType(), | ||||
| 3577 | EE->getVectorOperandType(), Idx); | ||||
| 3578 | // Add back the cost of s|zext which is subtracted separately. | ||||
| 3579 | Cost += TTIRef.getCastInstrCost( | ||||
| 3580 | Ext->getOpcode(), Ext->getType(), EE->getType(), | ||||
| 3581 | TTI::getCastContextHint(Ext), CostKind, Ext); | ||||
| 3582 | continue; | ||||
| 3583 | } | ||||
| 3584 | } | ||||
| 3585 | Cost -= TTIRef.getVectorInstrCost(Instruction::ExtractElement, | ||||
| 3586 | EE->getVectorOperandType(), Idx); | ||||
| 3587 | } | ||||
| 3588 | // Add a cost for subvector extracts/inserts if required. | ||||
| 3589 | for (const auto &Data : ExtractVectorsTys) { | ||||
| 3590 | auto *EEVTy = cast<FixedVectorType>(Data.first->getType()); | ||||
| 3591 | unsigned NumElts = VecTy->getNumElements(); | ||||
| 3592 | if (TTIRef.getNumberOfParts(EEVTy) > TTIRef.getNumberOfParts(VecTy)) { | ||||
| 3593 | unsigned Idx = (Data.second / NumElts) * NumElts; | ||||
| 3594 | unsigned EENumElts = EEVTy->getNumElements(); | ||||
| 3595 | if (Idx + NumElts <= EENumElts) { | ||||
| 3596 | Cost += | ||||
| 3597 | TTIRef.getShuffleCost(TargetTransformInfo::SK_ExtractSubvector, | ||||
| 3598 | EEVTy, None, Idx, VecTy); | ||||
| 3599 | } else { | ||||
| 3600 | // Need to round up the subvector type vectorization factor to avoid a | ||||
| 3601 | // crash in cost model functions. Make SubVT so that Idx + VF of SubVT | ||||
| 3602 | // <= EENumElts. | ||||
| 3603 | auto *SubVT = | ||||
| 3604 | FixedVectorType::get(VecTy->getElementType(), EENumElts - Idx); | ||||
| 3605 | Cost += | ||||
| 3606 | TTIRef.getShuffleCost(TargetTransformInfo::SK_ExtractSubvector, | ||||
| 3607 | EEVTy, None, Idx, SubVT); | ||||
| 3608 | } | ||||
| 3609 | } else { | ||||
| 3610 | Cost += TTIRef.getShuffleCost(TargetTransformInfo::SK_InsertSubvector, | ||||
| 3611 | VecTy, None, 0, EEVTy); | ||||
| 3612 | } | ||||
| 3613 | } | ||||
| 3614 | }; | ||||
| 3615 | if (E->State == TreeEntry::NeedToGather) { | ||||
| 3616 | if (allConstant(VL)) | ||||
| 3617 | return 0; | ||||
| 3618 | if (isSplat(VL)) { | ||||
| 3619 | return ReuseShuffleCost + | ||||
| 3620 | TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, None, | ||||
| 3621 | 0); | ||||
| 3622 | } | ||||
| 3623 | if (E->getOpcode() == Instruction::ExtractElement && | ||||
| 3624 | allSameType(VL) && allSameBlock(VL)) { | ||||
| 3625 | SmallVector<int> Mask; | ||||
| 3626 | Optional<TargetTransformInfo::ShuffleKind> ShuffleKind = | ||||
| 3627 | isShuffle(VL, Mask); | ||||
| 3628 | if (ShuffleKind.hasValue()) { | ||||
| 3629 | InstructionCost Cost = | ||||
| 3630 | computeExtractCost(VL, VecTy, *ShuffleKind, Mask, *TTI); | ||||
| 3631 | AdjustExtractsCost(Cost, /*IsGather=*/true); | ||||
| 3632 | return ReuseShuffleCost + Cost; | ||||
| 3633 | } | ||||
| 3634 | } | ||||
| 3635 | InstructionCost GatherCost = 0; | ||||
| 3636 | SmallVector<int> Mask; | ||||
| 3637 | SmallVector<const TreeEntry *> Entries; | ||||
| 3638 | Optional<TargetTransformInfo::ShuffleKind> Shuffle = | ||||
| 3639 | isGatherShuffledEntry(E, Mask, Entries); | ||||
| 3640 | if (Shuffle.hasValue()) { | ||||
| 3641 | if (ShuffleVectorInst::isIdentityMask(Mask)) { | ||||
| 3642 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: perfect diamond match for gather bundle that starts with " << *VL.front() << ".\n"; } } while (false) | ||||
| 3643 | dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: perfect diamond match for gather bundle that starts with " << *VL.front() << ".\n"; } } while (false) | ||||
| 3644 | << "SLP: perfect diamond match for gather bundle that starts with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: perfect diamond match for gather bundle that starts with " << *VL.front() << ".\n"; } } while (false) | ||||
| 3645 | << *VL.front() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: perfect diamond match for gather bundle that starts with " << *VL.front() << ".\n"; } } while (false); | ||||
| 3646 | } else { | ||||
| 3647 | LLVM_DEBUG(dbgs() << "SLP: shuffled " << Entries.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: shuffled " << Entries. size() << " entries for bundle that starts with " << *VL.front() << ".\n"; } } while (false) | ||||
| 3648 | << " entries for bundle that starts with "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: shuffled " << Entries. size() << " entries for bundle that starts with " << *VL.front() << ".\n"; } } while (false) | ||||
| 3649 | << *VL.front() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: shuffled " << Entries. size() << " entries for bundle that starts with " << *VL.front() << ".\n"; } } while (false); | ||||
| 3650 | GatherCost = TTI->getShuffleCost(*Shuffle, VecTy, Mask); | ||||
| 3651 | } | ||||
| 3652 | } else { | ||||
| 3653 | GatherCost = getGatherCost(VL); | ||||
| 3654 | } | ||||
| 3655 | return ReuseShuffleCost + GatherCost; | ||||
| 3656 | } | ||||
| 3657 | assert((E->State == TreeEntry::Vectorize ||(static_cast <bool> ((E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && "Unhandled state" ) ? void (0) : __assert_fail ("(E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && \"Unhandled state\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3659, __extension__ __PRETTY_FUNCTION__)) | ||||
| 3658 | E->State == TreeEntry::ScatterVectorize) &&(static_cast <bool> ((E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && "Unhandled state" ) ? void (0) : __assert_fail ("(E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && \"Unhandled state\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3659, __extension__ __PRETTY_FUNCTION__)) | ||||
| 3659 | "Unhandled state")(static_cast <bool> ((E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && "Unhandled state" ) ? void (0) : __assert_fail ("(E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && \"Unhandled state\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3659, __extension__ __PRETTY_FUNCTION__)); | ||||
| 3660 | assert(E->getOpcode() && allSameType(VL) && allSameBlock(VL) && "Invalid VL")(static_cast <bool> (E->getOpcode() && allSameType (VL) && allSameBlock(VL) && "Invalid VL") ? void (0) : __assert_fail ("E->getOpcode() && allSameType(VL) && allSameBlock(VL) && \"Invalid VL\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3660, __extension__ __PRETTY_FUNCTION__)); | ||||
| 3661 | Instruction *VL0 = E->getMainOp(); | ||||
| 3662 | unsigned ShuffleOrOp = | ||||
| 3663 | E->isAltShuffle() ? (unsigned)Instruction::ShuffleVector : E->getOpcode(); | ||||
| 3664 | switch (ShuffleOrOp) { | ||||
| 3665 | case Instruction::PHI: | ||||
| 3666 | return 0; | ||||
| 3667 | |||||
| 3668 | case Instruction::ExtractValue: | ||||
| 3669 | case Instruction::ExtractElement: { | ||||
| 3670 | // The common cost of removal ExtractElement/ExtractValue instructions + | ||||
| 3671 | // the cost of shuffles, if required to resuffle the original vector. | ||||
| 3672 | InstructionCost CommonCost = 0; | ||||
| 3673 | if (NeedToShuffleReuses) { | ||||
| 3674 | unsigned Idx = 0; | ||||
| 3675 | for (unsigned I : E->ReuseShuffleIndices) { | ||||
| 3676 | if (ShuffleOrOp == Instruction::ExtractElement) { | ||||
| 3677 | auto *EE = cast<ExtractElementInst>(VL[I]); | ||||
| 3678 | ReuseShuffleCost -= TTI->getVectorInstrCost( | ||||
| 3679 | Instruction::ExtractElement, EE->getVectorOperandType(), | ||||
| 3680 | *getExtractIndex(EE)); | ||||
| 3681 | } else { | ||||
| 3682 | ReuseShuffleCost -= TTI->getVectorInstrCost( | ||||
| 3683 | Instruction::ExtractElement, VecTy, Idx); | ||||
| 3684 | ++Idx; | ||||
| 3685 | } | ||||
| 3686 | } | ||||
| 3687 | Idx = ReuseShuffleNumbers; | ||||
| 3688 | for (Value *V : VL) { | ||||
| 3689 | if (ShuffleOrOp == Instruction::ExtractElement) { | ||||
| 3690 | auto *EE = cast<ExtractElementInst>(V); | ||||
| 3691 | ReuseShuffleCost += TTI->getVectorInstrCost( | ||||
| 3692 | Instruction::ExtractElement, EE->getVectorOperandType(), | ||||
| 3693 | *getExtractIndex(EE)); | ||||
| 3694 | } else { | ||||
| 3695 | --Idx; | ||||
| 3696 | ReuseShuffleCost += TTI->getVectorInstrCost( | ||||
| 3697 | Instruction::ExtractElement, VecTy, Idx); | ||||
| 3698 | } | ||||
| 3699 | } | ||||
| 3700 | CommonCost = ReuseShuffleCost; | ||||
| 3701 | } else if (!E->ReorderIndices.empty()) { | ||||
| 3702 | SmallVector<int> NewMask; | ||||
| 3703 | inversePermutation(E->ReorderIndices, NewMask); | ||||
| 3704 | CommonCost = TTI->getShuffleCost( | ||||
| 3705 | TargetTransformInfo::SK_PermuteSingleSrc, VecTy, NewMask); | ||||
| 3706 | } | ||||
| 3707 | if (ShuffleOrOp == Instruction::ExtractValue) { | ||||
| 3708 | for (unsigned I = 0, E = VL.size(); I < E; ++I) { | ||||
| 3709 | auto *EI = cast<Instruction>(VL[I]); | ||||
| 3710 | // Take credit for instruction that will become dead. | ||||
| 3711 | if (EI->hasOneUse()) { | ||||
| 3712 | Instruction *Ext = EI->user_back(); | ||||
| 3713 | if ((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && | ||||
| 3714 | all_of(Ext->users(), | ||||
| 3715 | [](User *U) { return isa<GetElementPtrInst>(U); })) { | ||||
| 3716 | // Use getExtractWithExtendCost() to calculate the cost of | ||||
| 3717 | // extractelement/ext pair. | ||||
| 3718 | CommonCost -= TTI->getExtractWithExtendCost( | ||||
| 3719 | Ext->getOpcode(), Ext->getType(), VecTy, I); | ||||
| 3720 | // Add back the cost of s|zext which is subtracted separately. | ||||
| 3721 | CommonCost += TTI->getCastInstrCost( | ||||
| 3722 | Ext->getOpcode(), Ext->getType(), EI->getType(), | ||||
| 3723 | TTI::getCastContextHint(Ext), CostKind, Ext); | ||||
| 3724 | continue; | ||||
| 3725 | } | ||||
| 3726 | } | ||||
| 3727 | CommonCost -= | ||||
| 3728 | TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, I); | ||||
| 3729 | } | ||||
| 3730 | } else { | ||||
| 3731 | AdjustExtractsCost(CommonCost, /*IsGather=*/false); | ||||
| 3732 | } | ||||
| 3733 | return CommonCost; | ||||
| 3734 | } | ||||
| 3735 | case Instruction::ZExt: | ||||
| 3736 | case Instruction::SExt: | ||||
| 3737 | case Instruction::FPToUI: | ||||
| 3738 | case Instruction::FPToSI: | ||||
| 3739 | case Instruction::FPExt: | ||||
| 3740 | case Instruction::PtrToInt: | ||||
| 3741 | case Instruction::IntToPtr: | ||||
| 3742 | case Instruction::SIToFP: | ||||
| 3743 | case Instruction::UIToFP: | ||||
| 3744 | case Instruction::Trunc: | ||||
| 3745 | case Instruction::FPTrunc: | ||||
| 3746 | case Instruction::BitCast: { | ||||
| 3747 | Type *SrcTy = VL0->getOperand(0)->getType(); | ||||
| 3748 | InstructionCost ScalarEltCost = | ||||
| 3749 | TTI->getCastInstrCost(E->getOpcode(), ScalarTy, SrcTy, | ||||
| 3750 | TTI::getCastContextHint(VL0), CostKind, VL0); | ||||
| 3751 | if (NeedToShuffleReuses) { | ||||
| 3752 | ReuseShuffleCost -= (ReuseShuffleNumbers - VL.size()) * ScalarEltCost; | ||||
| 3753 | } | ||||
| 3754 | |||||
| 3755 | // Calculate the cost of this instruction. | ||||
| 3756 | InstructionCost ScalarCost = VL.size() * ScalarEltCost; | ||||
| 3757 | |||||
| 3758 | auto *SrcVecTy = FixedVectorType::get(SrcTy, VL.size()); | ||||
| 3759 | InstructionCost VecCost = 0; | ||||
| 3760 | // Check if the values are candidates to demote. | ||||
| 3761 | if (!MinBWs.count(VL0) || VecTy != SrcVecTy) { | ||||
| 3762 | VecCost = | ||||
| 3763 | ReuseShuffleCost + | ||||
| 3764 | TTI->getCastInstrCost(E->getOpcode(), VecTy, SrcVecTy, | ||||
| 3765 | TTI::getCastContextHint(VL0), CostKind, VL0); | ||||
| 3766 | } | ||||
| 3767 | LLVM_DEBUG(dumpTreeCosts(E, ReuseShuffleCost, VecCost, ScalarCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, ReuseShuffleCost, VecCost, ScalarCost ); } } while (false); | ||||
| 3768 | return VecCost - ScalarCost; | ||||
| 3769 | } | ||||
| 3770 | case Instruction::FCmp: | ||||
| 3771 | case Instruction::ICmp: | ||||
| 3772 | case Instruction::Select: { | ||||
| 3773 | // Calculate the cost of this instruction. | ||||
| 3774 | InstructionCost ScalarEltCost = | ||||
| 3775 | TTI->getCmpSelInstrCost(E->getOpcode(), ScalarTy, Builder.getInt1Ty(), | ||||
| 3776 | CmpInst::BAD_ICMP_PREDICATE, CostKind, VL0); | ||||
| 3777 | if (NeedToShuffleReuses) { | ||||
| 3778 | ReuseShuffleCost -= (ReuseShuffleNumbers - VL.size()) * ScalarEltCost; | ||||
| 3779 | } | ||||
| 3780 | auto *MaskTy = FixedVectorType::get(Builder.getInt1Ty(), VL.size()); | ||||
| 3781 | InstructionCost ScalarCost = VecTy->getNumElements() * ScalarEltCost; | ||||
| 3782 | |||||
| 3783 | // Check if all entries in VL are either compares or selects with compares | ||||
| 3784 | // as condition that have the same predicates. | ||||
| 3785 | CmpInst::Predicate VecPred = CmpInst::BAD_ICMP_PREDICATE; | ||||
| 3786 | bool First = true; | ||||
| 3787 | for (auto *V : VL) { | ||||
| 3788 | CmpInst::Predicate CurrentPred; | ||||
| 3789 | auto MatchCmp = m_Cmp(CurrentPred, m_Value(), m_Value()); | ||||
| 3790 | if ((!match(V, m_Select(MatchCmp, m_Value(), m_Value())) && | ||||
| 3791 | !match(V, MatchCmp)) || | ||||
| 3792 | (!First && VecPred != CurrentPred)) { | ||||
| 3793 | VecPred = CmpInst::BAD_ICMP_PREDICATE; | ||||
| 3794 | break; | ||||
| 3795 | } | ||||
| 3796 | First = false; | ||||
| 3797 | VecPred = CurrentPred; | ||||
| 3798 | } | ||||
| 3799 | |||||
| 3800 | InstructionCost VecCost = TTI->getCmpSelInstrCost( | ||||
| 3801 | E->getOpcode(), VecTy, MaskTy, VecPred, CostKind, VL0); | ||||
| 3802 | // Check if it is possible and profitable to use min/max for selects in | ||||
| 3803 | // VL. | ||||
| 3804 | // | ||||
| 3805 | auto IntrinsicAndUse = canConvertToMinOrMaxIntrinsic(VL); | ||||
| 3806 | if (IntrinsicAndUse.first != Intrinsic::not_intrinsic) { | ||||
| 3807 | IntrinsicCostAttributes CostAttrs(IntrinsicAndUse.first, VecTy, | ||||
| 3808 | {VecTy, VecTy}); | ||||
| 3809 | InstructionCost IntrinsicCost = | ||||
| 3810 | TTI->getIntrinsicInstrCost(CostAttrs, CostKind); | ||||
| 3811 | // If the selects are the only uses of the compares, they will be dead | ||||
| 3812 | // and we can adjust the cost by removing their cost. | ||||
| 3813 | if (IntrinsicAndUse.second) | ||||
| 3814 | IntrinsicCost -= | ||||
| 3815 | TTI->getCmpSelInstrCost(Instruction::ICmp, VecTy, MaskTy, | ||||
| 3816 | CmpInst::BAD_ICMP_PREDICATE, CostKind); | ||||
| 3817 | VecCost = std::min(VecCost, IntrinsicCost); | ||||
| 3818 | } | ||||
| 3819 | LLVM_DEBUG(dumpTreeCosts(E, ReuseShuffleCost, VecCost, ScalarCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, ReuseShuffleCost, VecCost, ScalarCost ); } } while (false); | ||||
| 3820 | return ReuseShuffleCost + VecCost - ScalarCost; | ||||
| 3821 | } | ||||
| 3822 | case Instruction::FNeg: | ||||
| 3823 | case Instruction::Add: | ||||
| 3824 | case Instruction::FAdd: | ||||
| 3825 | case Instruction::Sub: | ||||
| 3826 | case Instruction::FSub: | ||||
| 3827 | case Instruction::Mul: | ||||
| 3828 | case Instruction::FMul: | ||||
| 3829 | case Instruction::UDiv: | ||||
| 3830 | case Instruction::SDiv: | ||||
| 3831 | case Instruction::FDiv: | ||||
| 3832 | case Instruction::URem: | ||||
| 3833 | case Instruction::SRem: | ||||
| 3834 | case Instruction::FRem: | ||||
| 3835 | case Instruction::Shl: | ||||
| 3836 | case Instruction::LShr: | ||||
| 3837 | case Instruction::AShr: | ||||
| 3838 | case Instruction::And: | ||||
| 3839 | case Instruction::Or: | ||||
| 3840 | case Instruction::Xor: { | ||||
| 3841 | // Certain instructions can be cheaper to vectorize if they have a | ||||
| 3842 | // constant second vector operand. | ||||
| 3843 | TargetTransformInfo::OperandValueKind Op1VK = | ||||
| 3844 | TargetTransformInfo::OK_AnyValue; | ||||
| 3845 | TargetTransformInfo::OperandValueKind Op2VK = | ||||
| 3846 | TargetTransformInfo::OK_UniformConstantValue; | ||||
| 3847 | TargetTransformInfo::OperandValueProperties Op1VP = | ||||
| 3848 | TargetTransformInfo::OP_None; | ||||
| 3849 | TargetTransformInfo::OperandValueProperties Op2VP = | ||||
| 3850 | TargetTransformInfo::OP_PowerOf2; | ||||
| 3851 | |||||
| 3852 | // If all operands are exactly the same ConstantInt then set the | ||||
| 3853 | // operand kind to OK_UniformConstantValue. | ||||
| 3854 | // If instead not all operands are constants, then set the operand kind | ||||
| 3855 | // to OK_AnyValue. If all operands are constants but not the same, | ||||
| 3856 | // then set the operand kind to OK_NonUniformConstantValue. | ||||
| 3857 | ConstantInt *CInt0 = nullptr; | ||||
| 3858 | for (unsigned i = 0, e = VL.size(); i < e; ++i) { | ||||
| 3859 | const Instruction *I = cast<Instruction>(VL[i]); | ||||
| 3860 | unsigned OpIdx = isa<BinaryOperator>(I) ? 1 : 0; | ||||
| 3861 | ConstantInt *CInt = dyn_cast<ConstantInt>(I->getOperand(OpIdx)); | ||||
| 3862 | if (!CInt) { | ||||
| 3863 | Op2VK = TargetTransformInfo::OK_AnyValue; | ||||
| 3864 | Op2VP = TargetTransformInfo::OP_None; | ||||
| 3865 | break; | ||||
| 3866 | } | ||||
| 3867 | if (Op2VP == TargetTransformInfo::OP_PowerOf2 && | ||||
| 3868 | !CInt->getValue().isPowerOf2()) | ||||
| 3869 | Op2VP = TargetTransformInfo::OP_None; | ||||
| 3870 | if (i == 0) { | ||||
| 3871 | CInt0 = CInt; | ||||
| 3872 | continue; | ||||
| 3873 | } | ||||
| 3874 | if (CInt0 != CInt) | ||||
| 3875 | Op2VK = TargetTransformInfo::OK_NonUniformConstantValue; | ||||
| 3876 | } | ||||
| 3877 | |||||
| 3878 | SmallVector<const Value *, 4> Operands(VL0->operand_values()); | ||||
| 3879 | InstructionCost ScalarEltCost = | ||||
| 3880 | TTI->getArithmeticInstrCost(E->getOpcode(), ScalarTy, CostKind, Op1VK, | ||||
| 3881 | Op2VK, Op1VP, Op2VP, Operands, VL0); | ||||
| 3882 | if (NeedToShuffleReuses) { | ||||
| 3883 | ReuseShuffleCost -= (ReuseShuffleNumbers - VL.size()) * ScalarEltCost; | ||||
| 3884 | } | ||||
| 3885 | InstructionCost ScalarCost = VecTy->getNumElements() * ScalarEltCost; | ||||
| 3886 | InstructionCost VecCost = | ||||
| 3887 | TTI->getArithmeticInstrCost(E->getOpcode(), VecTy, CostKind, Op1VK, | ||||
| 3888 | Op2VK, Op1VP, Op2VP, Operands, VL0); | ||||
| 3889 | LLVM_DEBUG(dumpTreeCosts(E, ReuseShuffleCost, VecCost, ScalarCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, ReuseShuffleCost, VecCost, ScalarCost ); } } while (false); | ||||
| 3890 | return ReuseShuffleCost + VecCost - ScalarCost; | ||||
| 3891 | } | ||||
| 3892 | case Instruction::GetElementPtr: { | ||||
| 3893 | TargetTransformInfo::OperandValueKind Op1VK = | ||||
| 3894 | TargetTransformInfo::OK_AnyValue; | ||||
| 3895 | TargetTransformInfo::OperandValueKind Op2VK = | ||||
| 3896 | TargetTransformInfo::OK_UniformConstantValue; | ||||
| 3897 | |||||
| 3898 | InstructionCost ScalarEltCost = TTI->getArithmeticInstrCost( | ||||
| 3899 | Instruction::Add, ScalarTy, CostKind, Op1VK, Op2VK); | ||||
| 3900 | if (NeedToShuffleReuses) { | ||||
| 3901 | ReuseShuffleCost -= (ReuseShuffleNumbers - VL.size()) * ScalarEltCost; | ||||
| 3902 | } | ||||
| 3903 | InstructionCost ScalarCost = VecTy->getNumElements() * ScalarEltCost; | ||||
| 3904 | InstructionCost VecCost = TTI->getArithmeticInstrCost( | ||||
| 3905 | Instruction::Add, VecTy, CostKind, Op1VK, Op2VK); | ||||
| 3906 | LLVM_DEBUG(dumpTreeCosts(E, ReuseShuffleCost, VecCost, ScalarCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, ReuseShuffleCost, VecCost, ScalarCost ); } } while (false); | ||||
| 3907 | return ReuseShuffleCost + VecCost - ScalarCost; | ||||
| 3908 | } | ||||
| 3909 | case Instruction::Load: { | ||||
| 3910 | // Cost of wide load - cost of scalar loads. | ||||
| 3911 | Align alignment = cast<LoadInst>(VL0)->getAlign(); | ||||
| 3912 | InstructionCost ScalarEltCost = TTI->getMemoryOpCost( | ||||
| 3913 | Instruction::Load, ScalarTy, alignment, 0, CostKind, VL0); | ||||
| 3914 | if (NeedToShuffleReuses) { | ||||
| 3915 | ReuseShuffleCost -= (ReuseShuffleNumbers - VL.size()) * ScalarEltCost; | ||||
| 3916 | } | ||||
| 3917 | InstructionCost ScalarLdCost = VecTy->getNumElements() * ScalarEltCost; | ||||
| 3918 | InstructionCost VecLdCost; | ||||
| 3919 | if (E->State == TreeEntry::Vectorize) { | ||||
| 3920 | VecLdCost = TTI->getMemoryOpCost(Instruction::Load, VecTy, alignment, 0, | ||||
| 3921 | CostKind, VL0); | ||||
| 3922 | } else { | ||||
| 3923 | assert(E->State == TreeEntry::ScatterVectorize && "Unknown EntryState")(static_cast <bool> (E->State == TreeEntry::ScatterVectorize && "Unknown EntryState") ? void (0) : __assert_fail ( "E->State == TreeEntry::ScatterVectorize && \"Unknown EntryState\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3923, __extension__ __PRETTY_FUNCTION__)); | ||||
| 3924 | VecLdCost = TTI->getGatherScatterOpCost( | ||||
| 3925 | Instruction::Load, VecTy, cast<LoadInst>(VL0)->getPointerOperand(), | ||||
| 3926 | /*VariableMask=*/false, alignment, CostKind, VL0); | ||||
| 3927 | } | ||||
| 3928 | if (!NeedToShuffleReuses && !E->ReorderIndices.empty()) { | ||||
| 3929 | SmallVector<int> NewMask; | ||||
| 3930 | inversePermutation(E->ReorderIndices, NewMask); | ||||
| 3931 | VecLdCost += TTI->getShuffleCost( | ||||
| 3932 | TargetTransformInfo::SK_PermuteSingleSrc, VecTy, NewMask); | ||||
| 3933 | } | ||||
| 3934 | LLVM_DEBUG(dumpTreeCosts(E, ReuseShuffleCost, VecLdCost, ScalarLdCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, ReuseShuffleCost, VecLdCost, ScalarLdCost ); } } while (false); | ||||
| 3935 | return ReuseShuffleCost + VecLdCost - ScalarLdCost; | ||||
| 3936 | } | ||||
| 3937 | case Instruction::Store: { | ||||
| 3938 | // We know that we can merge the stores. Calculate the cost. | ||||
| 3939 | bool IsReorder = !E->ReorderIndices.empty(); | ||||
| 3940 | auto *SI = | ||||
| 3941 | cast<StoreInst>(IsReorder ? VL[E->ReorderIndices.front()] : VL0); | ||||
| 3942 | Align Alignment = SI->getAlign(); | ||||
| 3943 | InstructionCost ScalarEltCost = TTI->getMemoryOpCost( | ||||
| 3944 | Instruction::Store, ScalarTy, Alignment, 0, CostKind, VL0); | ||||
| 3945 | InstructionCost ScalarStCost = VecTy->getNumElements() * ScalarEltCost; | ||||
| 3946 | InstructionCost VecStCost = TTI->getMemoryOpCost( | ||||
| 3947 | Instruction::Store, VecTy, Alignment, 0, CostKind, VL0); | ||||
| 3948 | if (IsReorder) { | ||||
| 3949 | SmallVector<int> NewMask; | ||||
| 3950 | inversePermutation(E->ReorderIndices, NewMask); | ||||
| 3951 | VecStCost += TTI->getShuffleCost( | ||||
| 3952 | TargetTransformInfo::SK_PermuteSingleSrc, VecTy, NewMask); | ||||
| 3953 | } | ||||
| 3954 | LLVM_DEBUG(dumpTreeCosts(E, ReuseShuffleCost, VecStCost, ScalarStCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, ReuseShuffleCost, VecStCost, ScalarStCost ); } } while (false); | ||||
| 3955 | return VecStCost - ScalarStCost; | ||||
| 3956 | } | ||||
| 3957 | case Instruction::Call: { | ||||
| 3958 | CallInst *CI = cast<CallInst>(VL0); | ||||
| 3959 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||
| 3960 | |||||
| 3961 | // Calculate the cost of the scalar and vector calls. | ||||
| 3962 | IntrinsicCostAttributes CostAttrs(ID, *CI, 1); | ||||
| 3963 | InstructionCost ScalarEltCost = | ||||
| 3964 | TTI->getIntrinsicInstrCost(CostAttrs, CostKind); | ||||
| 3965 | if (NeedToShuffleReuses) { | ||||
| 3966 | ReuseShuffleCost -= (ReuseShuffleNumbers - VL.size()) * ScalarEltCost; | ||||
| 3967 | } | ||||
| 3968 | InstructionCost ScalarCallCost = VecTy->getNumElements() * ScalarEltCost; | ||||
| 3969 | |||||
| 3970 | auto VecCallCosts = getVectorCallCosts(CI, VecTy, TTI, TLI); | ||||
| 3971 | InstructionCost VecCallCost = | ||||
| 3972 | std::min(VecCallCosts.first, VecCallCosts.second); | ||||
| 3973 | |||||
| 3974 | LLVM_DEBUG(dbgs() << "SLP: Call cost " << VecCallCost - ScalarCallCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost " << VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false) | ||||
| 3975 | << " (" << VecCallCost << "-" << ScalarCallCost << ")"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost " << VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false) | ||||
| 3976 | << " for " << *CI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Call cost " << VecCallCost - ScalarCallCost << " (" << VecCallCost << "-" << ScalarCallCost << ")" << " for " << *CI << "\n"; } } while (false); | ||||
| 3977 | |||||
| 3978 | return ReuseShuffleCost + VecCallCost - ScalarCallCost; | ||||
| 3979 | } | ||||
| 3980 | case Instruction::ShuffleVector: { | ||||
| 3981 | assert(E->isAltShuffle() &&(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ()))) && "Invalid Shuffle Vector Operand") ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3986, __extension__ __PRETTY_FUNCTION__)) | ||||
| 3982 | ((Instruction::isBinaryOp(E->getOpcode()) &&(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ()))) && "Invalid Shuffle Vector Operand") ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3986, __extension__ __PRETTY_FUNCTION__)) | ||||
| 3983 | Instruction::isBinaryOp(E->getAltOpcode())) ||(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ()))) && "Invalid Shuffle Vector Operand") ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3986, __extension__ __PRETTY_FUNCTION__)) | ||||
| 3984 | (Instruction::isCast(E->getOpcode()) &&(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ()))) && "Invalid Shuffle Vector Operand") ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3986, __extension__ __PRETTY_FUNCTION__)) | ||||
| 3985 | Instruction::isCast(E->getAltOpcode()))) &&(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ()))) && "Invalid Shuffle Vector Operand") ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3986, __extension__ __PRETTY_FUNCTION__)) | ||||
| 3986 | "Invalid Shuffle Vector Operand")(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ()))) && "Invalid Shuffle Vector Operand") ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 3986, __extension__ __PRETTY_FUNCTION__)); | ||||
| 3987 | InstructionCost ScalarCost = 0; | ||||
| 3988 | if (NeedToShuffleReuses) { | ||||
| 3989 | for (unsigned Idx : E->ReuseShuffleIndices) { | ||||
| 3990 | Instruction *I = cast<Instruction>(VL[Idx]); | ||||
| 3991 | ReuseShuffleCost -= TTI->getInstructionCost(I, CostKind); | ||||
| 3992 | } | ||||
| 3993 | for (Value *V : VL) { | ||||
| 3994 | Instruction *I = cast<Instruction>(V); | ||||
| 3995 | ReuseShuffleCost += TTI->getInstructionCost(I, CostKind); | ||||
| 3996 | } | ||||
| 3997 | } | ||||
| 3998 | for (Value *V : VL) { | ||||
| 3999 | Instruction *I = cast<Instruction>(V); | ||||
| 4000 | assert(E->isOpcodeOrAlt(I) && "Unexpected main/alternate opcode")(static_cast <bool> (E->isOpcodeOrAlt(I) && "Unexpected main/alternate opcode" ) ? void (0) : __assert_fail ("E->isOpcodeOrAlt(I) && \"Unexpected main/alternate opcode\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4000, __extension__ __PRETTY_FUNCTION__)); | ||||
| 4001 | ScalarCost += TTI->getInstructionCost(I, CostKind); | ||||
| 4002 | } | ||||
| 4003 | // VecCost is equal to sum of the cost of creating 2 vectors | ||||
| 4004 | // and the cost of creating shuffle. | ||||
| 4005 | InstructionCost VecCost = 0; | ||||
| 4006 | if (Instruction::isBinaryOp(E->getOpcode())) { | ||||
| 4007 | VecCost = TTI->getArithmeticInstrCost(E->getOpcode(), VecTy, CostKind); | ||||
| 4008 | VecCost += TTI->getArithmeticInstrCost(E->getAltOpcode(), VecTy, | ||||
| 4009 | CostKind); | ||||
| 4010 | } else { | ||||
| 4011 | Type *Src0SclTy = E->getMainOp()->getOperand(0)->getType(); | ||||
| 4012 | Type *Src1SclTy = E->getAltOp()->getOperand(0)->getType(); | ||||
| 4013 | auto *Src0Ty = FixedVectorType::get(Src0SclTy, VL.size()); | ||||
| 4014 | auto *Src1Ty = FixedVectorType::get(Src1SclTy, VL.size()); | ||||
| 4015 | VecCost = TTI->getCastInstrCost(E->getOpcode(), VecTy, Src0Ty, | ||||
| 4016 | TTI::CastContextHint::None, CostKind); | ||||
| 4017 | VecCost += TTI->getCastInstrCost(E->getAltOpcode(), VecTy, Src1Ty, | ||||
| 4018 | TTI::CastContextHint::None, CostKind); | ||||
| 4019 | } | ||||
| 4020 | |||||
| 4021 | SmallVector<int> Mask(E->Scalars.size()); | ||||
| 4022 | for (unsigned I = 0, End = E->Scalars.size(); I < End; ++I) { | ||||
| 4023 | auto *OpInst = cast<Instruction>(E->Scalars[I]); | ||||
| 4024 | assert(E->isOpcodeOrAlt(OpInst) && "Unexpected main/alternate opcode")(static_cast <bool> (E->isOpcodeOrAlt(OpInst) && "Unexpected main/alternate opcode") ? void (0) : __assert_fail ("E->isOpcodeOrAlt(OpInst) && \"Unexpected main/alternate opcode\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4024, __extension__ __PRETTY_FUNCTION__)); | ||||
| 4025 | Mask[I] = I + (OpInst->getOpcode() == E->getAltOpcode() ? End : 0); | ||||
| 4026 | } | ||||
| 4027 | VecCost += | ||||
| 4028 | TTI->getShuffleCost(TargetTransformInfo::SK_Select, VecTy, Mask, 0); | ||||
| 4029 | LLVM_DEBUG(dumpTreeCosts(E, ReuseShuffleCost, VecCost, ScalarCost))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dumpTreeCosts(E, ReuseShuffleCost, VecCost, ScalarCost ); } } while (false); | ||||
| 4030 | return ReuseShuffleCost + VecCost - ScalarCost; | ||||
| 4031 | } | ||||
| 4032 | default: | ||||
| 4033 | llvm_unreachable("Unknown instruction")::llvm::llvm_unreachable_internal("Unknown instruction", "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4033); | ||||
| 4034 | } | ||||
| 4035 | } | ||||
| 4036 | |||||
| 4037 | bool BoUpSLP::isFullyVectorizableTinyTree() const { | ||||
| 4038 | LLVM_DEBUG(dbgs() << "SLP: Check whether the tree with height "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Check whether the tree with height " << VectorizableTree.size() << " is fully vectorizable .\n" ; } } while (false) | ||||
| 4039 | << VectorizableTree.size() << " is fully vectorizable .\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Check whether the tree with height " << VectorizableTree.size() << " is fully vectorizable .\n" ; } } while (false); | ||||
| 4040 | |||||
| 4041 | // We only handle trees of heights 1 and 2. | ||||
| 4042 | if (VectorizableTree.size() == 1 && | ||||
| 4043 | VectorizableTree[0]->State == TreeEntry::Vectorize) | ||||
| 4044 | return true; | ||||
| 4045 | |||||
| 4046 | if (VectorizableTree.size() != 2) | ||||
| 4047 | return false; | ||||
| 4048 | |||||
| 4049 | // Handle splat and all-constants stores. Also try to vectorize tiny trees | ||||
| 4050 | // with the second gather nodes if they have less scalar operands rather than | ||||
| 4051 | // the initial tree element (may be profitable to shuffle the second gather). | ||||
| 4052 | if (VectorizableTree[0]->State == TreeEntry::Vectorize && | ||||
| 4053 | (allConstant(VectorizableTree[1]->Scalars) || | ||||
| 4054 | isSplat(VectorizableTree[1]->Scalars) || | ||||
| 4055 | (VectorizableTree[1]->State == TreeEntry::NeedToGather && | ||||
| 4056 | VectorizableTree[1]->Scalars.size() < | ||||
| 4057 | VectorizableTree[0]->Scalars.size()))) | ||||
| 4058 | return true; | ||||
| 4059 | |||||
| 4060 | // Gathering cost would be too much for tiny trees. | ||||
| 4061 | if (VectorizableTree[0]->State == TreeEntry::NeedToGather || | ||||
| 4062 | VectorizableTree[1]->State == TreeEntry::NeedToGather) | ||||
| 4063 | return false; | ||||
| 4064 | |||||
| 4065 | return true; | ||||
| 4066 | } | ||||
| 4067 | |||||
| 4068 | static bool isLoadCombineCandidateImpl(Value *Root, unsigned NumElts, | ||||
| 4069 | TargetTransformInfo *TTI) { | ||||
| 4070 | // Look past the root to find a source value. Arbitrarily follow the | ||||
| 4071 | // path through operand 0 of any 'or'. Also, peek through optional | ||||
| 4072 | // shift-left-by-multiple-of-8-bits. | ||||
| 4073 | Value *ZextLoad = Root; | ||||
| 4074 | const APInt *ShAmtC; | ||||
| 4075 | while (!isa<ConstantExpr>(ZextLoad) && | ||||
| 4076 | (match(ZextLoad, m_Or(m_Value(), m_Value())) || | ||||
| 4077 | (match(ZextLoad, m_Shl(m_Value(), m_APInt(ShAmtC))) && | ||||
| 4078 | ShAmtC->urem(8) == 0))) | ||||
| 4079 | ZextLoad = cast<BinaryOperator>(ZextLoad)->getOperand(0); | ||||
| 4080 | |||||
| 4081 | // Check if the input is an extended load of the required or/shift expression. | ||||
| 4082 | Value *LoadPtr; | ||||
| 4083 | if (ZextLoad == Root || !match(ZextLoad, m_ZExt(m_Load(m_Value(LoadPtr))))) | ||||
| 4084 | return false; | ||||
| 4085 | |||||
| 4086 | // Require that the total load bit width is a legal integer type. | ||||
| 4087 | // For example, <8 x i8> --> i64 is a legal integer on a 64-bit target. | ||||
| 4088 | // But <16 x i8> --> i128 is not, so the backend probably can't reduce it. | ||||
| 4089 | Type *SrcTy = LoadPtr->getType()->getPointerElementType(); | ||||
| 4090 | unsigned LoadBitWidth = SrcTy->getIntegerBitWidth() * NumElts; | ||||
| 4091 | if (!TTI->isTypeLegal(IntegerType::get(Root->getContext(), LoadBitWidth))) | ||||
| 4092 | return false; | ||||
| 4093 | |||||
| 4094 | // Everything matched - assume that we can fold the whole sequence using | ||||
| 4095 | // load combining. | ||||
| 4096 | LLVM_DEBUG(dbgs() << "SLP: Assume load combining for tree starting at "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Assume load combining for tree starting at " << *(cast<Instruction>(Root)) << "\n"; } } while (false) | ||||
| 4097 | << *(cast<Instruction>(Root)) << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Assume load combining for tree starting at " << *(cast<Instruction>(Root)) << "\n"; } } while (false); | ||||
| 4098 | |||||
| 4099 | return true; | ||||
| 4100 | } | ||||
| 4101 | |||||
| 4102 | bool BoUpSLP::isLoadCombineReductionCandidate(RecurKind RdxKind) const { | ||||
| 4103 | if (RdxKind != RecurKind::Or) | ||||
| 4104 | return false; | ||||
| 4105 | |||||
| 4106 | unsigned NumElts = VectorizableTree[0]->Scalars.size(); | ||||
| 4107 | Value *FirstReduced = VectorizableTree[0]->Scalars[0]; | ||||
| 4108 | return isLoadCombineCandidateImpl(FirstReduced, NumElts, TTI); | ||||
| 4109 | } | ||||
| 4110 | |||||
| 4111 | bool BoUpSLP::isLoadCombineCandidate() const { | ||||
| 4112 | // Peek through a final sequence of stores and check if all operations are | ||||
| 4113 | // likely to be load-combined. | ||||
| 4114 | unsigned NumElts = VectorizableTree[0]->Scalars.size(); | ||||
| 4115 | for (Value *Scalar : VectorizableTree[0]->Scalars) { | ||||
| 4116 | Value *X; | ||||
| 4117 | if (!match(Scalar, m_Store(m_Value(X), m_Value())) || | ||||
| 4118 | !isLoadCombineCandidateImpl(X, NumElts, TTI)) | ||||
| 4119 | return false; | ||||
| 4120 | } | ||||
| 4121 | return true; | ||||
| 4122 | } | ||||
| 4123 | |||||
| 4124 | bool BoUpSLP::isTreeTinyAndNotFullyVectorizable() const { | ||||
| 4125 | // We can vectorize the tree if its size is greater than or equal to the | ||||
| 4126 | // minimum size specified by the MinTreeSize command line option. | ||||
| 4127 | if (VectorizableTree.size() >= MinTreeSize) | ||||
| 4128 | return false; | ||||
| 4129 | |||||
| 4130 | // If we have a tiny tree (a tree whose size is less than MinTreeSize), we | ||||
| 4131 | // can vectorize it if we can prove it fully vectorizable. | ||||
| 4132 | if (isFullyVectorizableTinyTree()) | ||||
| 4133 | return false; | ||||
| 4134 | |||||
| 4135 | assert(VectorizableTree.empty()(static_cast <bool> (VectorizableTree.empty() ? ExternalUses .empty() : true && "We shouldn't have any external users" ) ? void (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4137, __extension__ __PRETTY_FUNCTION__)) | ||||
| 4136 | ? ExternalUses.empty()(static_cast <bool> (VectorizableTree.empty() ? ExternalUses .empty() : true && "We shouldn't have any external users" ) ? void (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4137, __extension__ __PRETTY_FUNCTION__)) | ||||
| 4137 | : true && "We shouldn't have any external users")(static_cast <bool> (VectorizableTree.empty() ? ExternalUses .empty() : true && "We shouldn't have any external users" ) ? void (0) : __assert_fail ("VectorizableTree.empty() ? ExternalUses.empty() : true && \"We shouldn't have any external users\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4137, __extension__ __PRETTY_FUNCTION__)); | ||||
| 4138 | |||||
| 4139 | // Otherwise, we can't vectorize the tree. It is both tiny and not fully | ||||
| 4140 | // vectorizable. | ||||
| 4141 | return true; | ||||
| 4142 | } | ||||
| 4143 | |||||
| 4144 | InstructionCost BoUpSLP::getSpillCost() const { | ||||
| 4145 | // Walk from the bottom of the tree to the top, tracking which values are | ||||
| 4146 | // live. When we see a call instruction that is not part of our tree, | ||||
| 4147 | // query TTI to see if there is a cost to keeping values live over it | ||||
| 4148 | // (for example, if spills and fills are required). | ||||
| 4149 | unsigned BundleWidth = VectorizableTree.front()->Scalars.size(); | ||||
| 4150 | InstructionCost Cost = 0; | ||||
| 4151 | |||||
| 4152 | SmallPtrSet<Instruction*, 4> LiveValues; | ||||
| 4153 | Instruction *PrevInst = nullptr; | ||||
| 4154 | |||||
| 4155 | // The entries in VectorizableTree are not necessarily ordered by their | ||||
| 4156 | // position in basic blocks. Collect them and order them by dominance so later | ||||
| 4157 | // instructions are guaranteed to be visited first. For instructions in | ||||
| 4158 | // different basic blocks, we only scan to the beginning of the block, so | ||||
| 4159 | // their order does not matter, as long as all instructions in a basic block | ||||
| 4160 | // are grouped together. Using dominance ensures a deterministic order. | ||||
| 4161 | SmallVector<Instruction *, 16> OrderedScalars; | ||||
| 4162 | for (const auto &TEPtr : VectorizableTree) { | ||||
| 4163 | Instruction *Inst = dyn_cast<Instruction>(TEPtr->Scalars[0]); | ||||
| 4164 | if (!Inst) | ||||
| 4165 | continue; | ||||
| 4166 | OrderedScalars.push_back(Inst); | ||||
| 4167 | } | ||||
| 4168 | llvm::stable_sort(OrderedScalars, [this](Instruction *A, Instruction *B) { | ||||
| 4169 | return DT->dominates(B, A); | ||||
| 4170 | }); | ||||
| 4171 | |||||
| 4172 | for (Instruction *Inst : OrderedScalars) { | ||||
| 4173 | if (!PrevInst) { | ||||
| 4174 | PrevInst = Inst; | ||||
| 4175 | continue; | ||||
| 4176 | } | ||||
| 4177 | |||||
| 4178 | // Update LiveValues. | ||||
| 4179 | LiveValues.erase(PrevInst); | ||||
| 4180 | for (auto &J : PrevInst->operands()) { | ||||
| 4181 | if (isa<Instruction>(&*J) && getTreeEntry(&*J)) | ||||
| 4182 | LiveValues.insert(cast<Instruction>(&*J)); | ||||
| 4183 | } | ||||
| 4184 | |||||
| 4185 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: #LV: " << LiveValues .size(); for (auto *X : LiveValues) dbgs() << " " << X->getName(); dbgs() << ", Looking at "; Inst->dump (); }; } } while (false) | ||||
| 4186 | dbgs() << "SLP: #LV: " << LiveValues.size();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: #LV: " << LiveValues .size(); for (auto *X : LiveValues) dbgs() << " " << X->getName(); dbgs() << ", Looking at "; Inst->dump (); }; } } while (false) | ||||
| 4187 | for (auto *X : LiveValues)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: #LV: " << LiveValues .size(); for (auto *X : LiveValues) dbgs() << " " << X->getName(); dbgs() << ", Looking at "; Inst->dump (); }; } } while (false) | ||||
| 4188 | dbgs() << " " << X->getName();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: #LV: " << LiveValues .size(); for (auto *X : LiveValues) dbgs() << " " << X->getName(); dbgs() << ", Looking at "; Inst->dump (); }; } } while (false) | ||||
| 4189 | dbgs() << ", Looking at ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: #LV: " << LiveValues .size(); for (auto *X : LiveValues) dbgs() << " " << X->getName(); dbgs() << ", Looking at "; Inst->dump (); }; } } while (false) | ||||
| 4190 | Inst->dump();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: #LV: " << LiveValues .size(); for (auto *X : LiveValues) dbgs() << " " << X->getName(); dbgs() << ", Looking at "; Inst->dump (); }; } } while (false) | ||||
| 4191 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { { dbgs() << "SLP: #LV: " << LiveValues .size(); for (auto *X : LiveValues) dbgs() << " " << X->getName(); dbgs() << ", Looking at "; Inst->dump (); }; } } while (false); | ||||
| 4192 | |||||
| 4193 | // Now find the sequence of instructions between PrevInst and Inst. | ||||
| 4194 | unsigned NumCalls = 0; | ||||
| 4195 | BasicBlock::reverse_iterator InstIt = ++Inst->getIterator().getReverse(), | ||||
| 4196 | PrevInstIt = | ||||
| 4197 | PrevInst->getIterator().getReverse(); | ||||
| 4198 | while (InstIt != PrevInstIt) { | ||||
| 4199 | if (PrevInstIt == PrevInst->getParent()->rend()) { | ||||
| 4200 | PrevInstIt = Inst->getParent()->rbegin(); | ||||
| 4201 | continue; | ||||
| 4202 | } | ||||
| 4203 | |||||
| 4204 | // Debug information does not impact spill cost. | ||||
| 4205 | if ((isa<CallInst>(&*PrevInstIt) && | ||||
| 4206 | !isa<DbgInfoIntrinsic>(&*PrevInstIt)) && | ||||
| 4207 | &*PrevInstIt != PrevInst) | ||||
| 4208 | NumCalls++; | ||||
| 4209 | |||||
| 4210 | ++PrevInstIt; | ||||
| 4211 | } | ||||
| 4212 | |||||
| 4213 | if (NumCalls) { | ||||
| 4214 | SmallVector<Type*, 4> V; | ||||
| 4215 | for (auto *II : LiveValues) | ||||
| 4216 | V.push_back(FixedVectorType::get(II->getType(), BundleWidth)); | ||||
| 4217 | Cost += NumCalls * TTI->getCostOfKeepingLiveOverCall(V); | ||||
| 4218 | } | ||||
| 4219 | |||||
| 4220 | PrevInst = Inst; | ||||
| 4221 | } | ||||
| 4222 | |||||
| 4223 | return Cost; | ||||
| 4224 | } | ||||
| 4225 | |||||
| 4226 | InstructionCost BoUpSLP::getTreeCost() { | ||||
| 4227 | InstructionCost Cost = 0; | ||||
| 4228 | LLVM_DEBUG(dbgs() << "SLP: Calculating cost for tree of size "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Calculating cost for tree of size " << VectorizableTree.size() << ".\n"; } } while ( false) | ||||
| 4229 | << VectorizableTree.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Calculating cost for tree of size " << VectorizableTree.size() << ".\n"; } } while ( false); | ||||
| 4230 | |||||
| 4231 | unsigned BundleWidth = VectorizableTree[0]->Scalars.size(); | ||||
| 4232 | |||||
| 4233 | for (unsigned I = 0, E = VectorizableTree.size(); I < E; ++I) { | ||||
| 4234 | TreeEntry &TE = *VectorizableTree[I].get(); | ||||
| 4235 | |||||
| 4236 | // We create duplicate tree entries for gather sequences that have multiple | ||||
| 4237 | // uses. However, we should not compute the cost of duplicate sequences. | ||||
| 4238 | // For example, if we have a build vector (i.e., insertelement sequence) | ||||
| 4239 | // that is used by more than one vector instruction, we only need to | ||||
| 4240 | // compute the cost of the insertelement instructions once. The redundant | ||||
| 4241 | // instructions will be eliminated by CSE. | ||||
| 4242 | // | ||||
| 4243 | // We should consider not creating duplicate tree entries for gather | ||||
| 4244 | // sequences, and instead add additional edges to the tree representing | ||||
| 4245 | // their uses. Since such an approach results in fewer total entries, | ||||
| 4246 | // existing heuristics based on tree size may yield different results. | ||||
| 4247 | // | ||||
| 4248 | if (TE.State == TreeEntry::NeedToGather && | ||||
| 4249 | std::any_of(std::next(VectorizableTree.begin(), I + 1), | ||||
| 4250 | VectorizableTree.end(), | ||||
| 4251 | [TE](const std::unique_ptr<TreeEntry> &EntryPtr) { | ||||
| 4252 | return EntryPtr->State == TreeEntry::NeedToGather && | ||||
| 4253 | EntryPtr->isSame(TE.Scalars); | ||||
| 4254 | })) | ||||
| 4255 | continue; | ||||
| 4256 | |||||
| 4257 | InstructionCost C = getEntryCost(&TE); | ||||
| 4258 | Cost += C; | ||||
| 4259 | LLVM_DEBUG(dbgs() << "SLP: Adding cost " << Cdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for bundle that starts with " << *TE.Scalars[0] << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | ||||
| 4260 | << " for bundle that starts with " << *TE.Scalars[0]do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for bundle that starts with " << *TE.Scalars[0] << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | ||||
| 4261 | << ".\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for bundle that starts with " << *TE.Scalars[0] << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false) | ||||
| 4262 | << "SLP: Current total cost = " << Cost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << C << " for bundle that starts with " << *TE.Scalars[0] << ".\n" << "SLP: Current total cost = " << Cost << "\n"; } } while (false); | ||||
| 4263 | } | ||||
| 4264 | |||||
| 4265 | SmallPtrSet<Value *, 16> ExtractCostCalculated; | ||||
| 4266 | InstructionCost ExtractCost = 0; | ||||
| 4267 | for (ExternalUser &EU : ExternalUses) { | ||||
| 4268 | // We only add extract cost once for the same scalar. | ||||
| 4269 | if (!ExtractCostCalculated.insert(EU.Scalar).second) | ||||
| 4270 | continue; | ||||
| 4271 | |||||
| 4272 | // Uses by ephemeral values are free (because the ephemeral value will be | ||||
| 4273 | // removed prior to code generation, and so the extraction will be | ||||
| 4274 | // removed as well). | ||||
| 4275 | if (EphValues.count(EU.User)) | ||||
| 4276 | continue; | ||||
| 4277 | |||||
| 4278 | // If we plan to rewrite the tree in a smaller type, we will need to sign | ||||
| 4279 | // extend the extracted value back to the original type. Here, we account | ||||
| 4280 | // for the extract and the added cost of the sign extend if needed. | ||||
| 4281 | auto *VecTy = FixedVectorType::get(EU.Scalar->getType(), BundleWidth); | ||||
| 4282 | auto *ScalarRoot = VectorizableTree[0]->Scalars[0]; | ||||
| 4283 | if (MinBWs.count(ScalarRoot)) { | ||||
| 4284 | auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); | ||||
| 4285 | auto Extend = | ||||
| 4286 | MinBWs[ScalarRoot].second ? Instruction::SExt : Instruction::ZExt; | ||||
| 4287 | VecTy = FixedVectorType::get(MinTy, BundleWidth); | ||||
| 4288 | ExtractCost += TTI->getExtractWithExtendCost(Extend, EU.Scalar->getType(), | ||||
| 4289 | VecTy, EU.Lane); | ||||
| 4290 | } else { | ||||
| 4291 | ExtractCost += | ||||
| 4292 | TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, EU.Lane); | ||||
| 4293 | } | ||||
| 4294 | } | ||||
| 4295 | |||||
| 4296 | InstructionCost SpillCost = getSpillCost(); | ||||
| 4297 | Cost += SpillCost + ExtractCost; | ||||
| 4298 | |||||
| 4299 | #ifndef NDEBUG | ||||
| 4300 | SmallString<256> Str; | ||||
| 4301 | { | ||||
| 4302 | raw_svector_ostream OS(Str); | ||||
| 4303 | OS << "SLP: Spill Cost = " << SpillCost << ".\n" | ||||
| 4304 | << "SLP: Extract Cost = " << ExtractCost << ".\n" | ||||
| 4305 | << "SLP: Total Cost = " << Cost << ".\n"; | ||||
| 4306 | } | ||||
| 4307 | LLVM_DEBUG(dbgs() << Str)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << Str; } } while (false); | ||||
| 4308 | if (ViewSLPTree) | ||||
| 4309 | ViewGraph(this, "SLP" + F->getName(), false, Str); | ||||
| 4310 | #endif | ||||
| 4311 | |||||
| 4312 | return Cost; | ||||
| 4313 | } | ||||
| 4314 | |||||
| 4315 | Optional<TargetTransformInfo::ShuffleKind> | ||||
| 4316 | BoUpSLP::isGatherShuffledEntry(const TreeEntry *TE, SmallVectorImpl<int> &Mask, | ||||
| 4317 | SmallVectorImpl<const TreeEntry *> &Entries) { | ||||
| 4318 | Mask.assign(TE->Scalars.size(), UndefMaskElem); | ||||
| 4319 | Entries.clear(); | ||||
| 4320 | DenseMap<Value *, const TreeEntry *> UsedValuesEntry; | ||||
| 4321 | unsigned VF = 0; | ||||
| 4322 | // FIXME: Shall be replaced by GetVF function once non-power-2 patch is | ||||
| 4323 | // landed. | ||||
| 4324 | auto &&GetVF = [](const TreeEntry *TE) { | ||||
| 4325 | if (!TE->ReuseShuffleIndices.empty()) | ||||
| 4326 | return TE->ReuseShuffleIndices.size(); | ||||
| 4327 | return TE->Scalars.size(); | ||||
| 4328 | }; | ||||
| 4329 | for (int I = 0, E = TE->Scalars.size(); I < E; ++I) { | ||||
| 4330 | Value *V = TE->Scalars[I]; | ||||
| 4331 | if (isa<UndefValue>(V)) | ||||
| 4332 | continue; | ||||
| 4333 | const TreeEntry *VTE = UsedValuesEntry.lookup(V); | ||||
| 4334 | if (!VTE) { | ||||
| 4335 | if (Entries.size() == 2) | ||||
| 4336 | return None; | ||||
| 4337 | VTE = getTreeEntry(V); | ||||
| 4338 | if (!VTE || find_if( | ||||
| 4339 | VectorizableTree, | ||||
| 4340 | [VTE, TE](const std::unique_ptr<TreeEntry> &EntryPtr) { | ||||
| 4341 | return EntryPtr.get() == VTE || EntryPtr.get() == TE; | ||||
| 4342 | })->get() == TE) { | ||||
| 4343 | // Check if it is used in one of the gathered entries. | ||||
| 4344 | const auto *It = | ||||
| 4345 | find_if(VectorizableTree, | ||||
| 4346 | [V, TE](const std::unique_ptr<TreeEntry> &EntryPtr) { | ||||
| 4347 | return EntryPtr.get() == TE || | ||||
| 4348 | (EntryPtr->State == TreeEntry::NeedToGather && | ||||
| 4349 | is_contained(EntryPtr->Scalars, V)); | ||||
| 4350 | }); | ||||
| 4351 | // The vector factor of shuffled entries must be the same. | ||||
| 4352 | if (It->get() == TE) | ||||
| 4353 | return None; | ||||
| 4354 | VTE = It->get(); | ||||
| 4355 | } | ||||
| 4356 | Entries.push_back(VTE); | ||||
| 4357 | if (Entries.size() == 1) { | ||||
| 4358 | VF = GetVF(VTE); | ||||
| 4359 | } else if (VF != GetVF(VTE)) { | ||||
| 4360 | assert(Entries.size() == 2 && "Expected shuffle of 1 or 2 entries.")(static_cast <bool> (Entries.size() == 2 && "Expected shuffle of 1 or 2 entries." ) ? void (0) : __assert_fail ("Entries.size() == 2 && \"Expected shuffle of 1 or 2 entries.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4360, __extension__ __PRETTY_FUNCTION__)); | ||||
| 4361 | assert(VF > 0 && "Expected non-zero vector factor.")(static_cast <bool> (VF > 0 && "Expected non-zero vector factor." ) ? void (0) : __assert_fail ("VF > 0 && \"Expected non-zero vector factor.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4361, __extension__ __PRETTY_FUNCTION__)); | ||||
| 4362 | return None; | ||||
| 4363 | } | ||||
| 4364 | for (Value *SV : VTE->Scalars) | ||||
| 4365 | UsedValuesEntry.try_emplace(SV, VTE); | ||||
| 4366 | } | ||||
| 4367 | int FoundLane = findLaneForValue(VTE->Scalars, VTE->ReuseShuffleIndices, V); | ||||
| 4368 | Mask[I] = (Entries.front() == VTE ? 0 : VF) + FoundLane; | ||||
| 4369 | // Extra check required by isSingleSourceMaskImpl function (called by | ||||
| 4370 | // ShuffleVectorInst::isSingleSourceMask). | ||||
| 4371 | if (Mask[I] >= 2 * E) | ||||
| 4372 | return None; | ||||
| 4373 | } | ||||
| 4374 | switch (Entries.size()) { | ||||
| 4375 | case 1: | ||||
| 4376 | return TargetTransformInfo::SK_PermuteSingleSrc; | ||||
| 4377 | case 2: | ||||
| 4378 | return TargetTransformInfo::SK_PermuteTwoSrc; | ||||
| 4379 | default: | ||||
| 4380 | break; | ||||
| 4381 | } | ||||
| 4382 | return None; | ||||
| 4383 | } | ||||
| 4384 | |||||
| 4385 | InstructionCost | ||||
| 4386 | BoUpSLP::getGatherCost(FixedVectorType *Ty, | ||||
| 4387 | const DenseSet<unsigned> &ShuffledIndices) const { | ||||
| 4388 | unsigned NumElts = Ty->getNumElements(); | ||||
| 4389 | APInt DemandedElts = APInt::getNullValue(NumElts); | ||||
| 4390 | for (unsigned I = 0; I < NumElts; ++I) | ||||
| 4391 | if (!ShuffledIndices.count(I)) | ||||
| 4392 | DemandedElts.setBit(I); | ||||
| 4393 | InstructionCost Cost = | ||||
| 4394 | TTI->getScalarizationOverhead(Ty, DemandedElts, /*Insert*/ true, | ||||
| 4395 | /*Extract*/ false); | ||||
| 4396 | if (!ShuffledIndices.empty()) | ||||
| 4397 | Cost += TTI->getShuffleCost(TargetTransformInfo::SK_PermuteSingleSrc, Ty); | ||||
| 4398 | return Cost; | ||||
| 4399 | } | ||||
| 4400 | |||||
| 4401 | InstructionCost BoUpSLP::getGatherCost(ArrayRef<Value *> VL) const { | ||||
| 4402 | // Find the type of the operands in VL. | ||||
| 4403 | Type *ScalarTy = VL[0]->getType(); | ||||
| 4404 | if (StoreInst *SI = dyn_cast<StoreInst>(VL[0])) | ||||
| 4405 | ScalarTy = SI->getValueOperand()->getType(); | ||||
| 4406 | auto *VecTy = FixedVectorType::get(ScalarTy, VL.size()); | ||||
| 4407 | // Find the cost of inserting/extracting values from the vector. | ||||
| 4408 | // Check if the same elements are inserted several times and count them as | ||||
| 4409 | // shuffle candidates. | ||||
| 4410 | DenseSet<unsigned> ShuffledElements; | ||||
| 4411 | DenseSet<Value *> UniqueElements; | ||||
| 4412 | // Iterate in reverse order to consider insert elements with the high cost. | ||||
| 4413 | for (unsigned I = VL.size(); I > 0; --I) { | ||||
| 4414 | unsigned Idx = I - 1; | ||||
| 4415 | if (!UniqueElements.insert(VL[Idx]).second) | ||||
| 4416 | ShuffledElements.insert(Idx); | ||||
| 4417 | } | ||||
| 4418 | return getGatherCost(VecTy, ShuffledElements); | ||||
| 4419 | } | ||||
| 4420 | |||||
| 4421 | // Perform operand reordering on the instructions in VL and return the reordered | ||||
| 4422 | // operands in Left and Right. | ||||
| 4423 | void BoUpSLP::reorderInputsAccordingToOpcode(ArrayRef<Value *> VL, | ||||
| 4424 | SmallVectorImpl<Value *> &Left, | ||||
| 4425 | SmallVectorImpl<Value *> &Right, | ||||
| 4426 | const DataLayout &DL, | ||||
| 4427 | ScalarEvolution &SE, | ||||
| 4428 | const BoUpSLP &R) { | ||||
| 4429 | if (VL.empty()) | ||||
| 4430 | return; | ||||
| 4431 | VLOperands Ops(VL, DL, SE, R); | ||||
| 4432 | // Reorder the operands in place. | ||||
| 4433 | Ops.reorder(); | ||||
| 4434 | Left = Ops.getVL(0); | ||||
| 4435 | Right = Ops.getVL(1); | ||||
| 4436 | } | ||||
| 4437 | |||||
| 4438 | void BoUpSLP::setInsertPointAfterBundle(const TreeEntry *E) { | ||||
| 4439 | // Get the basic block this bundle is in. All instructions in the bundle | ||||
| 4440 | // should be in this block. | ||||
| 4441 | auto *Front = E->getMainOp(); | ||||
| 4442 | auto *BB = Front->getParent(); | ||||
| 4443 | assert(llvm::all_of(E->Scalars, [=](Value *V) -> bool {(static_cast <bool> (llvm::all_of(E->Scalars, [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(E->Scalars, [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4446, __extension__ __PRETTY_FUNCTION__)) | ||||
| 4444 | auto *I = cast<Instruction>(V);(static_cast <bool> (llvm::all_of(E->Scalars, [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(E->Scalars, [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4446, __extension__ __PRETTY_FUNCTION__)) | ||||
| 4445 | return !E->isOpcodeOrAlt(I) || I->getParent() == BB;(static_cast <bool> (llvm::all_of(E->Scalars, [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(E->Scalars, [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4446, __extension__ __PRETTY_FUNCTION__)) | ||||
| 4446 | }))(static_cast <bool> (llvm::all_of(E->Scalars, [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })) ? void (0) : __assert_fail ("llvm::all_of(E->Scalars, [=](Value *V) -> bool { auto *I = cast<Instruction>(V); return !E->isOpcodeOrAlt(I) || I->getParent() == BB; })" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4446, __extension__ __PRETTY_FUNCTION__)); | ||||
| 4447 | |||||
| 4448 | // The last instruction in the bundle in program order. | ||||
| 4449 | Instruction *LastInst = nullptr; | ||||
| 4450 | |||||
| 4451 | // Find the last instruction. The common case should be that BB has been | ||||
| 4452 | // scheduled, and the last instruction is VL.back(). So we start with | ||||
| 4453 | // VL.back() and iterate over schedule data until we reach the end of the | ||||
| 4454 | // bundle. The end of the bundle is marked by null ScheduleData. | ||||
| 4455 | if (BlocksSchedules.count(BB)) { | ||||
| 4456 | auto *Bundle = | ||||
| 4457 | BlocksSchedules[BB]->getScheduleData(E->isOneOf(E->Scalars.back())); | ||||
| 4458 | if (Bundle && Bundle->isPartOfBundle()) | ||||
| 4459 | for (; Bundle; Bundle = Bundle->NextInBundle) | ||||
| 4460 | if (Bundle->OpValue == Bundle->Inst) | ||||
| 4461 | LastInst = Bundle->Inst; | ||||
| 4462 | } | ||||
| 4463 | |||||
| 4464 | // LastInst can still be null at this point if there's either not an entry | ||||
| 4465 | // for BB in BlocksSchedules or there's no ScheduleData available for | ||||
| 4466 | // VL.back(). This can be the case if buildTree_rec aborts for various | ||||
| 4467 | // reasons (e.g., the maximum recursion depth is reached, the maximum region | ||||
| 4468 | // size is reached, etc.). ScheduleData is initialized in the scheduling | ||||
| 4469 | // "dry-run". | ||||
| 4470 | // | ||||
| 4471 | // If this happens, we can still find the last instruction by brute force. We | ||||
| 4472 | // iterate forwards from Front (inclusive) until we either see all | ||||
| 4473 | // instructions in the bundle or reach the end of the block. If Front is the | ||||
| 4474 | // last instruction in program order, LastInst will be set to Front, and we | ||||
| 4475 | // will visit all the remaining instructions in the block. | ||||
| 4476 | // | ||||
| 4477 | // One of the reasons we exit early from buildTree_rec is to place an upper | ||||
| 4478 | // bound on compile-time. Thus, taking an additional compile-time hit here is | ||||
| 4479 | // not ideal. However, this should be exceedingly rare since it requires that | ||||
| 4480 | // we both exit early from buildTree_rec and that the bundle be out-of-order | ||||
| 4481 | // (causing us to iterate all the way to the end of the block). | ||||
| 4482 | if (!LastInst) { | ||||
| 4483 | SmallPtrSet<Value *, 16> Bundle(E->Scalars.begin(), E->Scalars.end()); | ||||
| 4484 | for (auto &I : make_range(BasicBlock::iterator(Front), BB->end())) { | ||||
| 4485 | if (Bundle.erase(&I) && E->isOpcodeOrAlt(&I)) | ||||
| 4486 | LastInst = &I; | ||||
| 4487 | if (Bundle.empty()) | ||||
| 4488 | break; | ||||
| 4489 | } | ||||
| 4490 | } | ||||
| 4491 | assert(LastInst && "Failed to find last instruction in bundle")(static_cast <bool> (LastInst && "Failed to find last instruction in bundle" ) ? void (0) : __assert_fail ("LastInst && \"Failed to find last instruction in bundle\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4491, __extension__ __PRETTY_FUNCTION__)); | ||||
| 4492 | |||||
| 4493 | // Set the insertion point after the last instruction in the bundle. Set the | ||||
| 4494 | // debug location to Front. | ||||
| 4495 | Builder.SetInsertPoint(BB, ++LastInst->getIterator()); | ||||
| 4496 | Builder.SetCurrentDebugLocation(Front->getDebugLoc()); | ||||
| 4497 | } | ||||
| 4498 | |||||
| 4499 | Value *BoUpSLP::gather(ArrayRef<Value *> VL) { | ||||
| 4500 | Value *Val0 = | ||||
| 4501 | isa<StoreInst>(VL[0]) ? cast<StoreInst>(VL[0])->getValueOperand() : VL[0]; | ||||
| 4502 | FixedVectorType *VecTy = FixedVectorType::get(Val0->getType(), VL.size()); | ||||
| 4503 | Value *Vec = PoisonValue::get(VecTy); | ||||
| 4504 | unsigned InsIndex = 0; | ||||
| 4505 | for (Value *Val : VL) { | ||||
| 4506 | Vec = Builder.CreateInsertElement(Vec, Val, Builder.getInt32(InsIndex++)); | ||||
| 4507 | auto *InsElt = dyn_cast<InsertElementInst>(Vec); | ||||
| 4508 | if (!InsElt) | ||||
| 4509 | continue; | ||||
| 4510 | GatherSeq.insert(InsElt); | ||||
| 4511 | CSEBlocks.insert(InsElt->getParent()); | ||||
| 4512 | // Add to our 'need-to-extract' list. | ||||
| 4513 | if (TreeEntry *Entry = getTreeEntry(Val)) { | ||||
| 4514 | // Find which lane we need to extract. | ||||
| 4515 | int FoundLane = | ||||
| 4516 | findLaneForValue(Entry->Scalars, Entry->ReuseShuffleIndices, Val); | ||||
| 4517 | ExternalUses.push_back(ExternalUser(Val, InsElt, FoundLane)); | ||||
| 4518 | } | ||||
| 4519 | } | ||||
| 4520 | |||||
| 4521 | return Vec; | ||||
| 4522 | } | ||||
| 4523 | |||||
| 4524 | Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) { | ||||
| 4525 | InstructionsState S = getSameOpcode(VL); | ||||
| 4526 | if (S.getOpcode()) { | ||||
| 4527 | if (TreeEntry *E = getTreeEntry(S.OpValue)) { | ||||
| 4528 | if (E->isSame(VL)) { | ||||
| 4529 | Value *V = vectorizeTree(E); | ||||
| 4530 | if (VL.size() == E->Scalars.size() && !E->ReuseShuffleIndices.empty()) { | ||||
| 4531 | // Reshuffle to get only unique values. | ||||
| 4532 | // If some of the scalars are duplicated in the vectorization tree | ||||
| 4533 | // entry, we do not vectorize them but instead generate a mask for the | ||||
| 4534 | // reuses. But if there are several users of the same entry, they may | ||||
| 4535 | // have different vectorization factors. This is especially important | ||||
| 4536 | // for PHI nodes. In this case, we need to adapt the resulting | ||||
| 4537 | // instruction for the user vectorization factor and have to reshuffle | ||||
| 4538 | // it again to take only unique elements of the vector. Without this | ||||
| 4539 | // code the function incorrectly returns reduced vector instruction | ||||
| 4540 | // with the same elements, not with the unique ones. | ||||
| 4541 | // block: | ||||
| 4542 | // %phi = phi <2 x > { .., %entry} {%shuffle, %block} | ||||
| 4543 | // %2 = shuffle <2 x > %phi, %poison, <4 x > <0, 0, 1, 1> | ||||
| 4544 | // ... (use %2) | ||||
| 4545 | // %shuffle = shuffle <2 x> %2, poison, <2 x> {0, 2} | ||||
| 4546 | // br %block | ||||
| 4547 | SmallVector<int, 4> UniqueIdxs; | ||||
| 4548 | SmallSet<int, 4> UsedIdxs; | ||||
| 4549 | int Pos = 0; | ||||
| 4550 | for (int Idx : E->ReuseShuffleIndices) { | ||||
| 4551 | if (UsedIdxs.insert(Idx).second) | ||||
| 4552 | UniqueIdxs.emplace_back(Pos); | ||||
| 4553 | ++Pos; | ||||
| 4554 | } | ||||
| 4555 | V = Builder.CreateShuffleVector(V, UniqueIdxs, "shrink.shuffle"); | ||||
| 4556 | } | ||||
| 4557 | return V; | ||||
| 4558 | } | ||||
| 4559 | } | ||||
| 4560 | } | ||||
| 4561 | |||||
| 4562 | // Check that every instruction appears once in this bundle. | ||||
| 4563 | SmallVector<int, 4> ReuseShuffleIndicies; | ||||
| 4564 | SmallVector<Value *, 4> UniqueValues; | ||||
| 4565 | if (VL.size() > 2) { | ||||
| 4566 | DenseMap<Value *, unsigned> UniquePositions; | ||||
| 4567 | for (Value *V : VL) { | ||||
| 4568 | auto Res = UniquePositions.try_emplace(V, UniqueValues.size()); | ||||
| 4569 | ReuseShuffleIndicies.emplace_back(Res.first->second); | ||||
| 4570 | if (Res.second || isa<Constant>(V)) | ||||
| 4571 | UniqueValues.emplace_back(V); | ||||
| 4572 | } | ||||
| 4573 | // Do not shuffle single element or if number of unique values is not power | ||||
| 4574 | // of 2. | ||||
| 4575 | if (UniqueValues.size() == VL.size() || UniqueValues.size() <= 1 || | ||||
| 4576 | !llvm::isPowerOf2_32(UniqueValues.size())) | ||||
| 4577 | ReuseShuffleIndicies.clear(); | ||||
| 4578 | else | ||||
| 4579 | VL = UniqueValues; | ||||
| 4580 | } | ||||
| 4581 | |||||
| 4582 | Value *Vec = gather(VL); | ||||
| 4583 | if (!ReuseShuffleIndicies.empty()) { | ||||
| 4584 | Vec = Builder.CreateShuffleVector(Vec, ReuseShuffleIndicies, "shuffle"); | ||||
| 4585 | if (auto *I = dyn_cast<Instruction>(Vec)) { | ||||
| 4586 | GatherSeq.insert(I); | ||||
| 4587 | CSEBlocks.insert(I->getParent()); | ||||
| 4588 | } | ||||
| 4589 | } | ||||
| 4590 | return Vec; | ||||
| 4591 | } | ||||
| 4592 | |||||
| 4593 | namespace { | ||||
| 4594 | /// Merges shuffle masks and emits final shuffle instruction, if required. | ||||
| 4595 | class ShuffleInstructionBuilder { | ||||
| 4596 | IRBuilderBase &Builder; | ||||
| 4597 | bool IsFinalized = false; | ||||
| 4598 | SmallVector<int, 4> Mask; | ||||
| 4599 | |||||
| 4600 | public: | ||||
| 4601 | ShuffleInstructionBuilder(IRBuilderBase &Builder) : Builder(Builder) {} | ||||
| 4602 | |||||
| 4603 | /// Adds a mask, inverting it before applying. | ||||
| 4604 | void addInversedMask(ArrayRef<unsigned> SubMask) { | ||||
| 4605 | if (SubMask.empty()) | ||||
| 4606 | return; | ||||
| 4607 | SmallVector<int, 4> NewMask; | ||||
| 4608 | inversePermutation(SubMask, NewMask); | ||||
| 4609 | addMask(NewMask); | ||||
| 4610 | } | ||||
| 4611 | |||||
| 4612 | /// Functions adds masks, merging them into single one. | ||||
| 4613 | void addMask(ArrayRef<unsigned> SubMask) { | ||||
| 4614 | SmallVector<int, 4> NewMask(SubMask.begin(), SubMask.end()); | ||||
| 4615 | addMask(NewMask); | ||||
| 4616 | } | ||||
| 4617 | |||||
| 4618 | void addMask(ArrayRef<int> SubMask) { | ||||
| 4619 | if (SubMask.empty()) | ||||
| 4620 | return; | ||||
| 4621 | if (Mask.empty()) { | ||||
| 4622 | Mask.append(SubMask.begin(), SubMask.end()); | ||||
| 4623 | return; | ||||
| 4624 | } | ||||
| 4625 | SmallVector<int, 4> NewMask(SubMask.size(), SubMask.size()); | ||||
| 4626 | int TermValue = std::min(Mask.size(), SubMask.size()); | ||||
| 4627 | for (int I = 0, E = SubMask.size(); I < E; ++I) { | ||||
| 4628 | if (SubMask[I] >= TermValue || Mask[SubMask[I]] >= TermValue) { | ||||
| 4629 | NewMask[I] = E; | ||||
| 4630 | continue; | ||||
| 4631 | } | ||||
| 4632 | NewMask[I] = Mask[SubMask[I]]; | ||||
| 4633 | } | ||||
| 4634 | Mask.swap(NewMask); | ||||
| 4635 | } | ||||
| 4636 | |||||
| 4637 | Value *finalize(Value *V) { | ||||
| 4638 | IsFinalized = true; | ||||
| 4639 | if (Mask.empty()) | ||||
| 4640 | return V; | ||||
| 4641 | return Builder.CreateShuffleVector(V, Mask, "shuffle"); | ||||
| 4642 | } | ||||
| 4643 | |||||
| 4644 | ~ShuffleInstructionBuilder() { | ||||
| 4645 | assert((IsFinalized || Mask.empty()) &&(static_cast <bool> ((IsFinalized || Mask.empty()) && "Shuffle construction must be finalized.") ? void (0) : __assert_fail ("(IsFinalized || Mask.empty()) && \"Shuffle construction must be finalized.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4646, __extension__ __PRETTY_FUNCTION__)) | ||||
| 4646 | "Shuffle construction must be finalized.")(static_cast <bool> ((IsFinalized || Mask.empty()) && "Shuffle construction must be finalized.") ? void (0) : __assert_fail ("(IsFinalized || Mask.empty()) && \"Shuffle construction must be finalized.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4646, __extension__ __PRETTY_FUNCTION__)); | ||||
| 4647 | } | ||||
| 4648 | }; | ||||
| 4649 | } // namespace | ||||
| 4650 | |||||
| 4651 | Value *BoUpSLP::vectorizeTree(TreeEntry *E) { | ||||
| 4652 | IRBuilder<>::InsertPointGuard Guard(Builder); | ||||
| 4653 | |||||
| 4654 | if (E->VectorizedValue) { | ||||
| 4655 | LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n"; } } while (false); | ||||
| 4656 | return E->VectorizedValue; | ||||
| 4657 | } | ||||
| 4658 | |||||
| 4659 | ShuffleInstructionBuilder ShuffleBuilder(Builder); | ||||
| 4660 | bool NeedToShuffleReuses = !E->ReuseShuffleIndices.empty(); | ||||
| 4661 | if (E->State == TreeEntry::NeedToGather) { | ||||
| 4662 | setInsertPointAfterBundle(E); | ||||
| 4663 | Value *Vec; | ||||
| 4664 | SmallVector<int> Mask; | ||||
| 4665 | SmallVector<const TreeEntry *> Entries; | ||||
| 4666 | Optional<TargetTransformInfo::ShuffleKind> Shuffle = | ||||
| 4667 | isGatherShuffledEntry(E, Mask, Entries); | ||||
| 4668 | if (Shuffle.hasValue()) { | ||||
| 4669 | assert((Entries.size() == 1 || Entries.size() == 2) &&(static_cast <bool> ((Entries.size() == 1 || Entries.size () == 2) && "Expected shuffle of 1 or 2 entries.") ? void (0) : __assert_fail ("(Entries.size() == 1 || Entries.size() == 2) && \"Expected shuffle of 1 or 2 entries.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4670, __extension__ __PRETTY_FUNCTION__)) | ||||
| 4670 | "Expected shuffle of 1 or 2 entries.")(static_cast <bool> ((Entries.size() == 1 || Entries.size () == 2) && "Expected shuffle of 1 or 2 entries.") ? void (0) : __assert_fail ("(Entries.size() == 1 || Entries.size() == 2) && \"Expected shuffle of 1 or 2 entries.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4670, __extension__ __PRETTY_FUNCTION__)); | ||||
| 4671 | Vec = Builder.CreateShuffleVector(Entries.front()->VectorizedValue, | ||||
| 4672 | Entries.back()->VectorizedValue, Mask); | ||||
| 4673 | } else { | ||||
| 4674 | Vec = gather(E->Scalars); | ||||
| 4675 | } | ||||
| 4676 | if (NeedToShuffleReuses) { | ||||
| 4677 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | ||||
| 4678 | Vec = ShuffleBuilder.finalize(Vec); | ||||
| 4679 | if (auto *I = dyn_cast<Instruction>(Vec)) { | ||||
| 4680 | GatherSeq.insert(I); | ||||
| 4681 | CSEBlocks.insert(I->getParent()); | ||||
| 4682 | } | ||||
| 4683 | } | ||||
| 4684 | E->VectorizedValue = Vec; | ||||
| 4685 | return Vec; | ||||
| 4686 | } | ||||
| 4687 | |||||
| 4688 | assert((E->State == TreeEntry::Vectorize ||(static_cast <bool> ((E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && "Unhandled state" ) ? void (0) : __assert_fail ("(E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && \"Unhandled state\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4690, __extension__ __PRETTY_FUNCTION__)) | ||||
| 4689 | E->State == TreeEntry::ScatterVectorize) &&(static_cast <bool> ((E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && "Unhandled state" ) ? void (0) : __assert_fail ("(E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && \"Unhandled state\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4690, __extension__ __PRETTY_FUNCTION__)) | ||||
| 4690 | "Unhandled state")(static_cast <bool> ((E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && "Unhandled state" ) ? void (0) : __assert_fail ("(E->State == TreeEntry::Vectorize || E->State == TreeEntry::ScatterVectorize) && \"Unhandled state\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4690, __extension__ __PRETTY_FUNCTION__)); | ||||
| 4691 | unsigned ShuffleOrOp = | ||||
| 4692 | E->isAltShuffle() ? (unsigned)Instruction::ShuffleVector : E->getOpcode(); | ||||
| 4693 | Instruction *VL0 = E->getMainOp(); | ||||
| 4694 | Type *ScalarTy = VL0->getType(); | ||||
| 4695 | if (auto *Store = dyn_cast<StoreInst>(VL0)) | ||||
| 4696 | ScalarTy = Store->getValueOperand()->getType(); | ||||
| 4697 | auto *VecTy = FixedVectorType::get(ScalarTy, E->Scalars.size()); | ||||
| 4698 | switch (ShuffleOrOp) { | ||||
| 4699 | case Instruction::PHI: { | ||||
| 4700 | auto *PH = cast<PHINode>(VL0); | ||||
| 4701 | Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI()); | ||||
| 4702 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); | ||||
| 4703 | PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues()); | ||||
| 4704 | Value *V = NewPhi; | ||||
| 4705 | if (NeedToShuffleReuses) | ||||
| 4706 | V = Builder.CreateShuffleVector(V, E->ReuseShuffleIndices, "shuffle"); | ||||
| 4707 | |||||
| 4708 | E->VectorizedValue = V; | ||||
| 4709 | |||||
| 4710 | // PHINodes may have multiple entries from the same block. We want to | ||||
| 4711 | // visit every block once. | ||||
| 4712 | SmallPtrSet<BasicBlock*, 4> VisitedBBs; | ||||
| 4713 | |||||
| 4714 | for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) { | ||||
| 4715 | ValueList Operands; | ||||
| 4716 | BasicBlock *IBB = PH->getIncomingBlock(i); | ||||
| 4717 | |||||
| 4718 | if (!VisitedBBs.insert(IBB).second) { | ||||
| 4719 | NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB); | ||||
| 4720 | continue; | ||||
| 4721 | } | ||||
| 4722 | |||||
| 4723 | Builder.SetInsertPoint(IBB->getTerminator()); | ||||
| 4724 | Builder.SetCurrentDebugLocation(PH->getDebugLoc()); | ||||
| 4725 | Value *Vec = vectorizeTree(E->getOperand(i)); | ||||
| 4726 | NewPhi->addIncoming(Vec, IBB); | ||||
| 4727 | } | ||||
| 4728 | |||||
| 4729 | assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&(static_cast <bool> (NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && "Invalid number of incoming values" ) ? void (0) : __assert_fail ("NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && \"Invalid number of incoming values\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4730, __extension__ __PRETTY_FUNCTION__)) | ||||
| 4730 | "Invalid number of incoming values")(static_cast <bool> (NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && "Invalid number of incoming values" ) ? void (0) : __assert_fail ("NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() && \"Invalid number of incoming values\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4730, __extension__ __PRETTY_FUNCTION__)); | ||||
| 4731 | return V; | ||||
| 4732 | } | ||||
| 4733 | |||||
| 4734 | case Instruction::ExtractElement: { | ||||
| 4735 | Value *V = E->getSingleOperand(0); | ||||
| 4736 | Builder.SetInsertPoint(VL0); | ||||
| 4737 | ShuffleBuilder.addInversedMask(E->ReorderIndices); | ||||
| 4738 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | ||||
| 4739 | V = ShuffleBuilder.finalize(V); | ||||
| 4740 | E->VectorizedValue = V; | ||||
| 4741 | return V; | ||||
| 4742 | } | ||||
| 4743 | case Instruction::ExtractValue: { | ||||
| 4744 | auto *LI = cast<LoadInst>(E->getSingleOperand(0)); | ||||
| 4745 | Builder.SetInsertPoint(LI); | ||||
| 4746 | auto *PtrTy = PointerType::get(VecTy, LI->getPointerAddressSpace()); | ||||
| 4747 | Value *Ptr = Builder.CreateBitCast(LI->getOperand(0), PtrTy); | ||||
| 4748 | LoadInst *V = Builder.CreateAlignedLoad(VecTy, Ptr, LI->getAlign()); | ||||
| 4749 | Value *NewV = propagateMetadata(V, E->Scalars); | ||||
| 4750 | ShuffleBuilder.addInversedMask(E->ReorderIndices); | ||||
| 4751 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | ||||
| 4752 | NewV = ShuffleBuilder.finalize(NewV); | ||||
| 4753 | E->VectorizedValue = NewV; | ||||
| 4754 | return NewV; | ||||
| 4755 | } | ||||
| 4756 | case Instruction::ZExt: | ||||
| 4757 | case Instruction::SExt: | ||||
| 4758 | case Instruction::FPToUI: | ||||
| 4759 | case Instruction::FPToSI: | ||||
| 4760 | case Instruction::FPExt: | ||||
| 4761 | case Instruction::PtrToInt: | ||||
| 4762 | case Instruction::IntToPtr: | ||||
| 4763 | case Instruction::SIToFP: | ||||
| 4764 | case Instruction::UIToFP: | ||||
| 4765 | case Instruction::Trunc: | ||||
| 4766 | case Instruction::FPTrunc: | ||||
| 4767 | case Instruction::BitCast: { | ||||
| 4768 | setInsertPointAfterBundle(E); | ||||
| 4769 | |||||
| 4770 | Value *InVec = vectorizeTree(E->getOperand(0)); | ||||
| 4771 | |||||
| 4772 | if (E->VectorizedValue) { | ||||
| 4773 | LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | ||||
| 4774 | return E->VectorizedValue; | ||||
| 4775 | } | ||||
| 4776 | |||||
| 4777 | auto *CI = cast<CastInst>(VL0); | ||||
| 4778 | Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy); | ||||
| 4779 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | ||||
| 4780 | V = ShuffleBuilder.finalize(V); | ||||
| 4781 | |||||
| 4782 | E->VectorizedValue = V; | ||||
| 4783 | ++NumVectorInstructions; | ||||
| 4784 | return V; | ||||
| 4785 | } | ||||
| 4786 | case Instruction::FCmp: | ||||
| 4787 | case Instruction::ICmp: { | ||||
| 4788 | setInsertPointAfterBundle(E); | ||||
| 4789 | |||||
| 4790 | Value *L = vectorizeTree(E->getOperand(0)); | ||||
| 4791 | Value *R = vectorizeTree(E->getOperand(1)); | ||||
| 4792 | |||||
| 4793 | if (E->VectorizedValue) { | ||||
| 4794 | LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | ||||
| 4795 | return E->VectorizedValue; | ||||
| 4796 | } | ||||
| 4797 | |||||
| 4798 | CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate(); | ||||
| 4799 | Value *V = Builder.CreateCmp(P0, L, R); | ||||
| 4800 | propagateIRFlags(V, E->Scalars, VL0); | ||||
| 4801 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | ||||
| 4802 | V = ShuffleBuilder.finalize(V); | ||||
| 4803 | |||||
| 4804 | E->VectorizedValue = V; | ||||
| 4805 | ++NumVectorInstructions; | ||||
| 4806 | return V; | ||||
| 4807 | } | ||||
| 4808 | case Instruction::Select: { | ||||
| 4809 | setInsertPointAfterBundle(E); | ||||
| 4810 | |||||
| 4811 | Value *Cond = vectorizeTree(E->getOperand(0)); | ||||
| 4812 | Value *True = vectorizeTree(E->getOperand(1)); | ||||
| 4813 | Value *False = vectorizeTree(E->getOperand(2)); | ||||
| 4814 | |||||
| 4815 | if (E->VectorizedValue) { | ||||
| 4816 | LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | ||||
| 4817 | return E->VectorizedValue; | ||||
| 4818 | } | ||||
| 4819 | |||||
| 4820 | Value *V = Builder.CreateSelect(Cond, True, False); | ||||
| 4821 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | ||||
| 4822 | V = ShuffleBuilder.finalize(V); | ||||
| 4823 | |||||
| 4824 | E->VectorizedValue = V; | ||||
| 4825 | ++NumVectorInstructions; | ||||
| 4826 | return V; | ||||
| 4827 | } | ||||
| 4828 | case Instruction::FNeg: { | ||||
| 4829 | setInsertPointAfterBundle(E); | ||||
| 4830 | |||||
| 4831 | Value *Op = vectorizeTree(E->getOperand(0)); | ||||
| 4832 | |||||
| 4833 | if (E->VectorizedValue) { | ||||
| 4834 | LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | ||||
| 4835 | return E->VectorizedValue; | ||||
| 4836 | } | ||||
| 4837 | |||||
| 4838 | Value *V = Builder.CreateUnOp( | ||||
| 4839 | static_cast<Instruction::UnaryOps>(E->getOpcode()), Op); | ||||
| 4840 | propagateIRFlags(V, E->Scalars, VL0); | ||||
| 4841 | if (auto *I = dyn_cast<Instruction>(V)) | ||||
| 4842 | V = propagateMetadata(I, E->Scalars); | ||||
| 4843 | |||||
| 4844 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | ||||
| 4845 | V = ShuffleBuilder.finalize(V); | ||||
| 4846 | |||||
| 4847 | E->VectorizedValue = V; | ||||
| 4848 | ++NumVectorInstructions; | ||||
| 4849 | |||||
| 4850 | return V; | ||||
| 4851 | } | ||||
| 4852 | case Instruction::Add: | ||||
| 4853 | case Instruction::FAdd: | ||||
| 4854 | case Instruction::Sub: | ||||
| 4855 | case Instruction::FSub: | ||||
| 4856 | case Instruction::Mul: | ||||
| 4857 | case Instruction::FMul: | ||||
| 4858 | case Instruction::UDiv: | ||||
| 4859 | case Instruction::SDiv: | ||||
| 4860 | case Instruction::FDiv: | ||||
| 4861 | case Instruction::URem: | ||||
| 4862 | case Instruction::SRem: | ||||
| 4863 | case Instruction::FRem: | ||||
| 4864 | case Instruction::Shl: | ||||
| 4865 | case Instruction::LShr: | ||||
| 4866 | case Instruction::AShr: | ||||
| 4867 | case Instruction::And: | ||||
| 4868 | case Instruction::Or: | ||||
| 4869 | case Instruction::Xor: { | ||||
| 4870 | setInsertPointAfterBundle(E); | ||||
| 4871 | |||||
| 4872 | Value *LHS = vectorizeTree(E->getOperand(0)); | ||||
| 4873 | Value *RHS = vectorizeTree(E->getOperand(1)); | ||||
| 4874 | |||||
| 4875 | if (E->VectorizedValue) { | ||||
| 4876 | LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | ||||
| 4877 | return E->VectorizedValue; | ||||
| 4878 | } | ||||
| 4879 | |||||
| 4880 | Value *V = Builder.CreateBinOp( | ||||
| 4881 | static_cast<Instruction::BinaryOps>(E->getOpcode()), LHS, | ||||
| 4882 | RHS); | ||||
| 4883 | propagateIRFlags(V, E->Scalars, VL0); | ||||
| 4884 | if (auto *I = dyn_cast<Instruction>(V)) | ||||
| 4885 | V = propagateMetadata(I, E->Scalars); | ||||
| 4886 | |||||
| 4887 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | ||||
| 4888 | V = ShuffleBuilder.finalize(V); | ||||
| 4889 | |||||
| 4890 | E->VectorizedValue = V; | ||||
| 4891 | ++NumVectorInstructions; | ||||
| 4892 | |||||
| 4893 | return V; | ||||
| 4894 | } | ||||
| 4895 | case Instruction::Load: { | ||||
| 4896 | // Loads are inserted at the head of the tree because we don't want to | ||||
| 4897 | // sink them all the way down past store instructions. | ||||
| 4898 | bool IsReorder = E->updateStateIfReorder(); | ||||
| 4899 | if (IsReorder) | ||||
| 4900 | VL0 = E->getMainOp(); | ||||
| 4901 | setInsertPointAfterBundle(E); | ||||
| 4902 | |||||
| 4903 | LoadInst *LI = cast<LoadInst>(VL0); | ||||
| 4904 | Instruction *NewLI; | ||||
| 4905 | unsigned AS = LI->getPointerAddressSpace(); | ||||
| 4906 | Value *PO = LI->getPointerOperand(); | ||||
| 4907 | if (E->State == TreeEntry::Vectorize) { | ||||
| 4908 | |||||
| 4909 | Value *VecPtr = Builder.CreateBitCast(PO, VecTy->getPointerTo(AS)); | ||||
| 4910 | |||||
| 4911 | // The pointer operand uses an in-tree scalar so we add the new BitCast | ||||
| 4912 | // to ExternalUses list to make sure that an extract will be generated | ||||
| 4913 | // in the future. | ||||
| 4914 | if (getTreeEntry(PO)) | ||||
| 4915 | ExternalUses.emplace_back(PO, cast<User>(VecPtr), 0); | ||||
| 4916 | |||||
| 4917 | NewLI = Builder.CreateAlignedLoad(VecTy, VecPtr, LI->getAlign()); | ||||
| 4918 | } else { | ||||
| 4919 | assert(E->State == TreeEntry::ScatterVectorize && "Unhandled state")(static_cast <bool> (E->State == TreeEntry::ScatterVectorize && "Unhandled state") ? void (0) : __assert_fail ("E->State == TreeEntry::ScatterVectorize && \"Unhandled state\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 4919, __extension__ __PRETTY_FUNCTION__)); | ||||
| 4920 | Value *VecPtr = vectorizeTree(E->getOperand(0)); | ||||
| 4921 | // Use the minimum alignment of the gathered loads. | ||||
| 4922 | Align CommonAlignment = LI->getAlign(); | ||||
| 4923 | for (Value *V : E->Scalars) | ||||
| 4924 | CommonAlignment = | ||||
| 4925 | commonAlignment(CommonAlignment, cast<LoadInst>(V)->getAlign()); | ||||
| 4926 | NewLI = Builder.CreateMaskedGather(VecPtr, CommonAlignment); | ||||
| 4927 | } | ||||
| 4928 | Value *V = propagateMetadata(NewLI, E->Scalars); | ||||
| 4929 | |||||
| 4930 | ShuffleBuilder.addInversedMask(E->ReorderIndices); | ||||
| 4931 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | ||||
| 4932 | V = ShuffleBuilder.finalize(V); | ||||
| 4933 | E->VectorizedValue = V; | ||||
| 4934 | ++NumVectorInstructions; | ||||
| 4935 | return V; | ||||
| 4936 | } | ||||
| 4937 | case Instruction::Store: { | ||||
| 4938 | bool IsReorder = !E->ReorderIndices.empty(); | ||||
| 4939 | auto *SI = cast<StoreInst>( | ||||
| 4940 | IsReorder ? E->Scalars[E->ReorderIndices.front()] : VL0); | ||||
| 4941 | unsigned AS = SI->getPointerAddressSpace(); | ||||
| 4942 | |||||
| 4943 | setInsertPointAfterBundle(E); | ||||
| 4944 | |||||
| 4945 | Value *VecValue = vectorizeTree(E->getOperand(0)); | ||||
| 4946 | ShuffleBuilder.addMask(E->ReorderIndices); | ||||
| 4947 | VecValue = ShuffleBuilder.finalize(VecValue); | ||||
| 4948 | |||||
| 4949 | Value *ScalarPtr = SI->getPointerOperand(); | ||||
| 4950 | Value *VecPtr = Builder.CreateBitCast( | ||||
| 4951 | ScalarPtr, VecValue->getType()->getPointerTo(AS)); | ||||
| 4952 | StoreInst *ST = Builder.CreateAlignedStore(VecValue, VecPtr, | ||||
| 4953 | SI->getAlign()); | ||||
| 4954 | |||||
| 4955 | // The pointer operand uses an in-tree scalar, so add the new BitCast to | ||||
| 4956 | // ExternalUses to make sure that an extract will be generated in the | ||||
| 4957 | // future. | ||||
| 4958 | if (getTreeEntry(ScalarPtr)) | ||||
| 4959 | ExternalUses.push_back(ExternalUser(ScalarPtr, cast<User>(VecPtr), 0)); | ||||
| 4960 | |||||
| 4961 | Value *V = propagateMetadata(ST, E->Scalars); | ||||
| 4962 | |||||
| 4963 | E->VectorizedValue = V; | ||||
| 4964 | ++NumVectorInstructions; | ||||
| 4965 | return V; | ||||
| 4966 | } | ||||
| 4967 | case Instruction::GetElementPtr: { | ||||
| 4968 | setInsertPointAfterBundle(E); | ||||
| 4969 | |||||
| 4970 | Value *Op0 = vectorizeTree(E->getOperand(0)); | ||||
| 4971 | |||||
| 4972 | std::vector<Value *> OpVecs; | ||||
| 4973 | for (int j = 1, e = cast<GetElementPtrInst>(VL0)->getNumOperands(); j < e; | ||||
| 4974 | ++j) { | ||||
| 4975 | ValueList &VL = E->getOperand(j); | ||||
| 4976 | // Need to cast all elements to the same type before vectorization to | ||||
| 4977 | // avoid crash. | ||||
| 4978 | Type *VL0Ty = VL0->getOperand(j)->getType(); | ||||
| 4979 | Type *Ty = llvm::all_of( | ||||
| 4980 | VL, [VL0Ty](Value *V) { return VL0Ty == V->getType(); }) | ||||
| 4981 | ? VL0Ty | ||||
| 4982 | : DL->getIndexType(cast<GetElementPtrInst>(VL0) | ||||
| 4983 | ->getPointerOperandType() | ||||
| 4984 | ->getScalarType()); | ||||
| 4985 | for (Value *&V : VL) { | ||||
| 4986 | auto *CI = cast<ConstantInt>(V); | ||||
| 4987 | V = ConstantExpr::getIntegerCast(CI, Ty, | ||||
| 4988 | CI->getValue().isSignBitSet()); | ||||
| 4989 | } | ||||
| 4990 | Value *OpVec = vectorizeTree(VL); | ||||
| 4991 | OpVecs.push_back(OpVec); | ||||
| 4992 | } | ||||
| 4993 | |||||
| 4994 | Value *V = Builder.CreateGEP( | ||||
| 4995 | cast<GetElementPtrInst>(VL0)->getSourceElementType(), Op0, OpVecs); | ||||
| 4996 | if (Instruction *I = dyn_cast<Instruction>(V)) | ||||
| 4997 | V = propagateMetadata(I, E->Scalars); | ||||
| 4998 | |||||
| 4999 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | ||||
| 5000 | V = ShuffleBuilder.finalize(V); | ||||
| 5001 | |||||
| 5002 | E->VectorizedValue = V; | ||||
| 5003 | ++NumVectorInstructions; | ||||
| 5004 | |||||
| 5005 | return V; | ||||
| 5006 | } | ||||
| 5007 | case Instruction::Call: { | ||||
| 5008 | CallInst *CI = cast<CallInst>(VL0); | ||||
| 5009 | setInsertPointAfterBundle(E); | ||||
| 5010 | |||||
| 5011 | Intrinsic::ID IID = Intrinsic::not_intrinsic; | ||||
| 5012 | if (Function *FI = CI->getCalledFunction()) | ||||
| 5013 | IID = FI->getIntrinsicID(); | ||||
| 5014 | |||||
| 5015 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); | ||||
| 5016 | |||||
| 5017 | auto VecCallCosts = getVectorCallCosts(CI, VecTy, TTI, TLI); | ||||
| 5018 | bool UseIntrinsic = ID != Intrinsic::not_intrinsic && | ||||
| 5019 | VecCallCosts.first <= VecCallCosts.second; | ||||
| 5020 | |||||
| 5021 | Value *ScalarArg = nullptr; | ||||
| 5022 | std::vector<Value *> OpVecs; | ||||
| 5023 | for (int j = 0, e = CI->getNumArgOperands(); j < e; ++j) { | ||||
| 5024 | ValueList OpVL; | ||||
| 5025 | // Some intrinsics have scalar arguments. This argument should not be | ||||
| 5026 | // vectorized. | ||||
| 5027 | if (UseIntrinsic && hasVectorInstrinsicScalarOpd(IID, j)) { | ||||
| 5028 | CallInst *CEI = cast<CallInst>(VL0); | ||||
| 5029 | ScalarArg = CEI->getArgOperand(j); | ||||
| 5030 | OpVecs.push_back(CEI->getArgOperand(j)); | ||||
| 5031 | continue; | ||||
| 5032 | } | ||||
| 5033 | |||||
| 5034 | Value *OpVec = vectorizeTree(E->getOperand(j)); | ||||
| 5035 | LLVM_DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n"; } } while (false); | ||||
| 5036 | OpVecs.push_back(OpVec); | ||||
| 5037 | } | ||||
| 5038 | |||||
| 5039 | Function *CF; | ||||
| 5040 | if (!UseIntrinsic) { | ||||
| 5041 | VFShape Shape = | ||||
| 5042 | VFShape::get(*CI, ElementCount::getFixed(static_cast<unsigned>( | ||||
| 5043 | VecTy->getNumElements())), | ||||
| 5044 | false /*HasGlobalPred*/); | ||||
| 5045 | CF = VFDatabase(*CI).getVectorizedFunction(Shape); | ||||
| 5046 | } else { | ||||
| 5047 | Type *Tys[] = {FixedVectorType::get(CI->getType(), E->Scalars.size())}; | ||||
| 5048 | CF = Intrinsic::getDeclaration(F->getParent(), ID, Tys); | ||||
| 5049 | } | ||||
| 5050 | |||||
| 5051 | SmallVector<OperandBundleDef, 1> OpBundles; | ||||
| 5052 | CI->getOperandBundlesAsDefs(OpBundles); | ||||
| 5053 | Value *V = Builder.CreateCall(CF, OpVecs, OpBundles); | ||||
| 5054 | |||||
| 5055 | // The scalar argument uses an in-tree scalar so we add the new vectorized | ||||
| 5056 | // call to ExternalUses list to make sure that an extract will be | ||||
| 5057 | // generated in the future. | ||||
| 5058 | if (ScalarArg && getTreeEntry(ScalarArg)) | ||||
| 5059 | ExternalUses.push_back(ExternalUser(ScalarArg, cast<User>(V), 0)); | ||||
| 5060 | |||||
| 5061 | propagateIRFlags(V, E->Scalars, VL0); | ||||
| 5062 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | ||||
| 5063 | V = ShuffleBuilder.finalize(V); | ||||
| 5064 | |||||
| 5065 | E->VectorizedValue = V; | ||||
| 5066 | ++NumVectorInstructions; | ||||
| 5067 | return V; | ||||
| 5068 | } | ||||
| 5069 | case Instruction::ShuffleVector: { | ||||
| 5070 | assert(E->isAltShuffle() &&(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ()))) && "Invalid Shuffle Vector Operand") ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5075, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5071 | ((Instruction::isBinaryOp(E->getOpcode()) &&(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ()))) && "Invalid Shuffle Vector Operand") ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5075, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5072 | Instruction::isBinaryOp(E->getAltOpcode())) ||(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ()))) && "Invalid Shuffle Vector Operand") ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5075, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5073 | (Instruction::isCast(E->getOpcode()) &&(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ()))) && "Invalid Shuffle Vector Operand") ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5075, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5074 | Instruction::isCast(E->getAltOpcode()))) &&(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ()))) && "Invalid Shuffle Vector Operand") ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5075, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5075 | "Invalid Shuffle Vector Operand")(static_cast <bool> (E->isAltShuffle() && (( Instruction::isBinaryOp(E->getOpcode()) && Instruction ::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E ->getOpcode()) && Instruction::isCast(E->getAltOpcode ()))) && "Invalid Shuffle Vector Operand") ? void (0) : __assert_fail ("E->isAltShuffle() && ((Instruction::isBinaryOp(E->getOpcode()) && Instruction::isBinaryOp(E->getAltOpcode())) || (Instruction::isCast(E->getOpcode()) && Instruction::isCast(E->getAltOpcode()))) && \"Invalid Shuffle Vector Operand\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5075, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5076 | |||||
| 5077 | Value *LHS = nullptr, *RHS = nullptr; | ||||
| 5078 | if (Instruction::isBinaryOp(E->getOpcode())) { | ||||
| 5079 | setInsertPointAfterBundle(E); | ||||
| 5080 | LHS = vectorizeTree(E->getOperand(0)); | ||||
| 5081 | RHS = vectorizeTree(E->getOperand(1)); | ||||
| 5082 | } else { | ||||
| 5083 | setInsertPointAfterBundle(E); | ||||
| 5084 | LHS = vectorizeTree(E->getOperand(0)); | ||||
| 5085 | } | ||||
| 5086 | |||||
| 5087 | if (E->VectorizedValue) { | ||||
| 5088 | LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n"; } } while (false); | ||||
| 5089 | return E->VectorizedValue; | ||||
| 5090 | } | ||||
| 5091 | |||||
| 5092 | Value *V0, *V1; | ||||
| 5093 | if (Instruction::isBinaryOp(E->getOpcode())) { | ||||
| 5094 | V0 = Builder.CreateBinOp( | ||||
| 5095 | static_cast<Instruction::BinaryOps>(E->getOpcode()), LHS, RHS); | ||||
| 5096 | V1 = Builder.CreateBinOp( | ||||
| 5097 | static_cast<Instruction::BinaryOps>(E->getAltOpcode()), LHS, RHS); | ||||
| 5098 | } else { | ||||
| 5099 | V0 = Builder.CreateCast( | ||||
| 5100 | static_cast<Instruction::CastOps>(E->getOpcode()), LHS, VecTy); | ||||
| 5101 | V1 = Builder.CreateCast( | ||||
| 5102 | static_cast<Instruction::CastOps>(E->getAltOpcode()), LHS, VecTy); | ||||
| 5103 | } | ||||
| 5104 | |||||
| 5105 | // Create shuffle to take alternate operations from the vector. | ||||
| 5106 | // Also, gather up main and alt scalar ops to propagate IR flags to | ||||
| 5107 | // each vector operation. | ||||
| 5108 | ValueList OpScalars, AltScalars; | ||||
| 5109 | unsigned e = E->Scalars.size(); | ||||
| 5110 | SmallVector<int, 8> Mask(e); | ||||
| 5111 | for (unsigned i = 0; i < e; ++i) { | ||||
| 5112 | auto *OpInst = cast<Instruction>(E->Scalars[i]); | ||||
| 5113 | assert(E->isOpcodeOrAlt(OpInst) && "Unexpected main/alternate opcode")(static_cast <bool> (E->isOpcodeOrAlt(OpInst) && "Unexpected main/alternate opcode") ? void (0) : __assert_fail ("E->isOpcodeOrAlt(OpInst) && \"Unexpected main/alternate opcode\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5113, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5114 | if (OpInst->getOpcode() == E->getAltOpcode()) { | ||||
| 5115 | Mask[i] = e + i; | ||||
| 5116 | AltScalars.push_back(E->Scalars[i]); | ||||
| 5117 | } else { | ||||
| 5118 | Mask[i] = i; | ||||
| 5119 | OpScalars.push_back(E->Scalars[i]); | ||||
| 5120 | } | ||||
| 5121 | } | ||||
| 5122 | |||||
| 5123 | propagateIRFlags(V0, OpScalars); | ||||
| 5124 | propagateIRFlags(V1, AltScalars); | ||||
| 5125 | |||||
| 5126 | Value *V = Builder.CreateShuffleVector(V0, V1, Mask); | ||||
| 5127 | if (Instruction *I = dyn_cast<Instruction>(V)) | ||||
| 5128 | V = propagateMetadata(I, E->Scalars); | ||||
| 5129 | ShuffleBuilder.addMask(E->ReuseShuffleIndices); | ||||
| 5130 | V = ShuffleBuilder.finalize(V); | ||||
| 5131 | |||||
| 5132 | E->VectorizedValue = V; | ||||
| 5133 | ++NumVectorInstructions; | ||||
| 5134 | |||||
| 5135 | return V; | ||||
| 5136 | } | ||||
| 5137 | default: | ||||
| 5138 | llvm_unreachable("unknown inst")::llvm::llvm_unreachable_internal("unknown inst", "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5138); | ||||
| 5139 | } | ||||
| 5140 | return nullptr; | ||||
| 5141 | } | ||||
| 5142 | |||||
| 5143 | Value *BoUpSLP::vectorizeTree() { | ||||
| 5144 | ExtraValueToDebugLocsMap ExternallyUsedValues; | ||||
| 5145 | return vectorizeTree(ExternallyUsedValues); | ||||
| 5146 | } | ||||
| 5147 | |||||
| 5148 | Value * | ||||
| 5149 | BoUpSLP::vectorizeTree(ExtraValueToDebugLocsMap &ExternallyUsedValues) { | ||||
| 5150 | // All blocks must be scheduled before any instructions are inserted. | ||||
| 5151 | for (auto &BSIter : BlocksSchedules) { | ||||
| 5152 | scheduleBlock(BSIter.second.get()); | ||||
| 5153 | } | ||||
| 5154 | |||||
| 5155 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | ||||
| 5156 | auto *VectorRoot = vectorizeTree(VectorizableTree[0].get()); | ||||
| 5157 | |||||
| 5158 | // If the vectorized tree can be rewritten in a smaller type, we truncate the | ||||
| 5159 | // vectorized root. InstCombine will then rewrite the entire expression. We | ||||
| 5160 | // sign extend the extracted values below. | ||||
| 5161 | auto *ScalarRoot = VectorizableTree[0]->Scalars[0]; | ||||
| 5162 | if (MinBWs.count(ScalarRoot)) { | ||||
| 5163 | if (auto *I = dyn_cast<Instruction>(VectorRoot)) { | ||||
| 5164 | // If current instr is a phi and not the last phi, insert it after the | ||||
| 5165 | // last phi node. | ||||
| 5166 | if (isa<PHINode>(I)) | ||||
| 5167 | Builder.SetInsertPoint(&*I->getParent()->getFirstInsertionPt()); | ||||
| 5168 | else | ||||
| 5169 | Builder.SetInsertPoint(&*++BasicBlock::iterator(I)); | ||||
| 5170 | } | ||||
| 5171 | auto BundleWidth = VectorizableTree[0]->Scalars.size(); | ||||
| 5172 | auto *MinTy = IntegerType::get(F->getContext(), MinBWs[ScalarRoot].first); | ||||
| 5173 | auto *VecTy = FixedVectorType::get(MinTy, BundleWidth); | ||||
| 5174 | auto *Trunc = Builder.CreateTrunc(VectorRoot, VecTy); | ||||
| 5175 | VectorizableTree[0]->VectorizedValue = Trunc; | ||||
| 5176 | } | ||||
| 5177 | |||||
| 5178 | LLVM_DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Extracting " << ExternalUses .size() << " values .\n"; } } while (false) | ||||
| 5179 | << " values .\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Extracting " << ExternalUses .size() << " values .\n"; } } while (false); | ||||
| 5180 | |||||
| 5181 | // If necessary, sign-extend or zero-extend ScalarRoot to the larger type | ||||
| 5182 | // specified by ScalarType. | ||||
| 5183 | auto extend = [&](Value *ScalarRoot, Value *Ex, Type *ScalarType) { | ||||
| 5184 | if (!MinBWs.count(ScalarRoot)) | ||||
| 5185 | return Ex; | ||||
| 5186 | if (MinBWs[ScalarRoot].second) | ||||
| 5187 | return Builder.CreateSExt(Ex, ScalarType); | ||||
| 5188 | return Builder.CreateZExt(Ex, ScalarType); | ||||
| 5189 | }; | ||||
| 5190 | |||||
| 5191 | // Extract all of the elements with the external uses. | ||||
| 5192 | for (const auto &ExternalUse : ExternalUses) { | ||||
| 5193 | Value *Scalar = ExternalUse.Scalar; | ||||
| 5194 | llvm::User *User = ExternalUse.User; | ||||
| 5195 | |||||
| 5196 | // Skip users that we already RAUW. This happens when one instruction | ||||
| 5197 | // has multiple uses of the same value. | ||||
| 5198 | if (User && !is_contained(Scalar->users(), User)) | ||||
| 5199 | continue; | ||||
| 5200 | TreeEntry *E = getTreeEntry(Scalar); | ||||
| 5201 | assert(E && "Invalid scalar")(static_cast <bool> (E && "Invalid scalar") ? void (0) : __assert_fail ("E && \"Invalid scalar\"", "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5201, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5202 | assert(E->State != TreeEntry::NeedToGather &&(static_cast <bool> (E->State != TreeEntry::NeedToGather && "Extracting from a gather list") ? void (0) : __assert_fail ("E->State != TreeEntry::NeedToGather && \"Extracting from a gather list\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5203, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5203 | "Extracting from a gather list")(static_cast <bool> (E->State != TreeEntry::NeedToGather && "Extracting from a gather list") ? void (0) : __assert_fail ("E->State != TreeEntry::NeedToGather && \"Extracting from a gather list\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5203, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5204 | |||||
| 5205 | Value *Vec = E->VectorizedValue; | ||||
| 5206 | assert(Vec && "Can't find vectorizable value")(static_cast <bool> (Vec && "Can't find vectorizable value" ) ? void (0) : __assert_fail ("Vec && \"Can't find vectorizable value\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5206, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5207 | |||||
| 5208 | Value *Lane = Builder.getInt32(ExternalUse.Lane); | ||||
| 5209 | // If User == nullptr, the Scalar is used as extra arg. Generate | ||||
| 5210 | // ExtractElement instruction and update the record for this scalar in | ||||
| 5211 | // ExternallyUsedValues. | ||||
| 5212 | if (!User) { | ||||
| 5213 | assert(ExternallyUsedValues.count(Scalar) &&(static_cast <bool> (ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? void (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5215, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5214 | "Scalar with nullptr as an external user must be registered in "(static_cast <bool> (ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? void (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5215, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5215 | "ExternallyUsedValues map")(static_cast <bool> (ExternallyUsedValues.count(Scalar) && "Scalar with nullptr as an external user must be registered in " "ExternallyUsedValues map") ? void (0) : __assert_fail ("ExternallyUsedValues.count(Scalar) && \"Scalar with nullptr as an external user must be registered in \" \"ExternallyUsedValues map\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5215, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5216 | if (auto *VecI = dyn_cast<Instruction>(Vec)) { | ||||
| 5217 | Builder.SetInsertPoint(VecI->getParent(), | ||||
| 5218 | std::next(VecI->getIterator())); | ||||
| 5219 | } else { | ||||
| 5220 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | ||||
| 5221 | } | ||||
| 5222 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | ||||
| 5223 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | ||||
| 5224 | CSEBlocks.insert(cast<Instruction>(Scalar)->getParent()); | ||||
| 5225 | auto &Locs = ExternallyUsedValues[Scalar]; | ||||
| 5226 | ExternallyUsedValues.insert({Ex, Locs}); | ||||
| 5227 | ExternallyUsedValues.erase(Scalar); | ||||
| 5228 | // Required to update internally referenced instructions. | ||||
| 5229 | Scalar->replaceAllUsesWith(Ex); | ||||
| 5230 | continue; | ||||
| 5231 | } | ||||
| 5232 | |||||
| 5233 | // Generate extracts for out-of-tree users. | ||||
| 5234 | // Find the insertion point for the extractelement lane. | ||||
| 5235 | if (auto *VecI = dyn_cast<Instruction>(Vec)) { | ||||
| 5236 | if (PHINode *PH = dyn_cast<PHINode>(User)) { | ||||
| 5237 | for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) { | ||||
| 5238 | if (PH->getIncomingValue(i) == Scalar) { | ||||
| 5239 | Instruction *IncomingTerminator = | ||||
| 5240 | PH->getIncomingBlock(i)->getTerminator(); | ||||
| 5241 | if (isa<CatchSwitchInst>(IncomingTerminator)) { | ||||
| 5242 | Builder.SetInsertPoint(VecI->getParent(), | ||||
| 5243 | std::next(VecI->getIterator())); | ||||
| 5244 | } else { | ||||
| 5245 | Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator()); | ||||
| 5246 | } | ||||
| 5247 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | ||||
| 5248 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | ||||
| 5249 | CSEBlocks.insert(PH->getIncomingBlock(i)); | ||||
| 5250 | PH->setOperand(i, Ex); | ||||
| 5251 | } | ||||
| 5252 | } | ||||
| 5253 | } else { | ||||
| 5254 | Builder.SetInsertPoint(cast<Instruction>(User)); | ||||
| 5255 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | ||||
| 5256 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | ||||
| 5257 | CSEBlocks.insert(cast<Instruction>(User)->getParent()); | ||||
| 5258 | User->replaceUsesOfWith(Scalar, Ex); | ||||
| 5259 | } | ||||
| 5260 | } else { | ||||
| 5261 | Builder.SetInsertPoint(&F->getEntryBlock().front()); | ||||
| 5262 | Value *Ex = Builder.CreateExtractElement(Vec, Lane); | ||||
| 5263 | Ex = extend(ScalarRoot, Ex, Scalar->getType()); | ||||
| 5264 | CSEBlocks.insert(&F->getEntryBlock()); | ||||
| 5265 | User->replaceUsesOfWith(Scalar, Ex); | ||||
| 5266 | } | ||||
| 5267 | |||||
| 5268 | LLVM_DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Replaced:" << *User << ".\n"; } } while (false); | ||||
| 5269 | } | ||||
| 5270 | |||||
| 5271 | // For each vectorized value: | ||||
| 5272 | for (auto &TEPtr : VectorizableTree) { | ||||
| 5273 | TreeEntry *Entry = TEPtr.get(); | ||||
| 5274 | |||||
| 5275 | // No need to handle users of gathered values. | ||||
| 5276 | if (Entry->State == TreeEntry::NeedToGather) | ||||
| 5277 | continue; | ||||
| 5278 | |||||
| 5279 | assert(Entry->VectorizedValue && "Can't find vectorizable value")(static_cast <bool> (Entry->VectorizedValue && "Can't find vectorizable value") ? void (0) : __assert_fail ( "Entry->VectorizedValue && \"Can't find vectorizable value\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5279, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5280 | |||||
| 5281 | // For each lane: | ||||
| 5282 | for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) { | ||||
| 5283 | Value *Scalar = Entry->Scalars[Lane]; | ||||
| 5284 | |||||
| 5285 | #ifndef NDEBUG | ||||
| 5286 | Type *Ty = Scalar->getType(); | ||||
| 5287 | if (!Ty->isVoidTy()) { | ||||
| 5288 | for (User *U : Scalar->users()) { | ||||
| 5289 | LLVM_DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tvalidating user:" << *U << ".\n"; } } while (false); | ||||
| 5290 | |||||
| 5291 | // It is legal to delete users in the ignorelist. | ||||
| 5292 | assert((getTreeEntry(U) || is_contained(UserIgnoreList, U)) &&(static_cast <bool> ((getTreeEntry(U) || is_contained(UserIgnoreList , U)) && "Deleting out-of-tree value") ? void (0) : __assert_fail ("(getTreeEntry(U) || is_contained(UserIgnoreList, U)) && \"Deleting out-of-tree value\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5293, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5293 | "Deleting out-of-tree value")(static_cast <bool> ((getTreeEntry(U) || is_contained(UserIgnoreList , U)) && "Deleting out-of-tree value") ? void (0) : __assert_fail ("(getTreeEntry(U) || is_contained(UserIgnoreList, U)) && \"Deleting out-of-tree value\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5293, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5294 | } | ||||
| 5295 | } | ||||
| 5296 | #endif | ||||
| 5297 | LLVM_DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: \tErasing scalar:" << * Scalar << ".\n"; } } while (false); | ||||
| 5298 | eraseInstruction(cast<Instruction>(Scalar)); | ||||
| 5299 | } | ||||
| 5300 | } | ||||
| 5301 | |||||
| 5302 | Builder.ClearInsertionPoint(); | ||||
| 5303 | InstrElementSize.clear(); | ||||
| 5304 | |||||
| 5305 | return VectorizableTree[0]->VectorizedValue; | ||||
| 5306 | } | ||||
| 5307 | |||||
| 5308 | void BoUpSLP::optimizeGatherSequence() { | ||||
| 5309 | LLVM_DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Optimizing " << GatherSeq .size() << " gather sequences instructions.\n"; } } while (false) | ||||
| 5310 | << " gather sequences instructions.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Optimizing " << GatherSeq .size() << " gather sequences instructions.\n"; } } while (false); | ||||
| 5311 | // LICM InsertElementInst sequences. | ||||
| 5312 | for (Instruction *I : GatherSeq) { | ||||
| 5313 | if (isDeleted(I)) | ||||
| 5314 | continue; | ||||
| 5315 | |||||
| 5316 | // Check if this block is inside a loop. | ||||
| 5317 | Loop *L = LI->getLoopFor(I->getParent()); | ||||
| 5318 | if (!L) | ||||
| 5319 | continue; | ||||
| 5320 | |||||
| 5321 | // Check if it has a preheader. | ||||
| 5322 | BasicBlock *PreHeader = L->getLoopPreheader(); | ||||
| 5323 | if (!PreHeader) | ||||
| 5324 | continue; | ||||
| 5325 | |||||
| 5326 | // If the vector or the element that we insert into it are | ||||
| 5327 | // instructions that are defined in this basic block then we can't | ||||
| 5328 | // hoist this instruction. | ||||
| 5329 | auto *Op0 = dyn_cast<Instruction>(I->getOperand(0)); | ||||
| 5330 | auto *Op1 = dyn_cast<Instruction>(I->getOperand(1)); | ||||
| 5331 | if (Op0 && L->contains(Op0)) | ||||
| 5332 | continue; | ||||
| 5333 | if (Op1 && L->contains(Op1)) | ||||
| 5334 | continue; | ||||
| 5335 | |||||
| 5336 | // We can hoist this instruction. Move it to the pre-header. | ||||
| 5337 | I->moveBefore(PreHeader->getTerminator()); | ||||
| 5338 | } | ||||
| 5339 | |||||
| 5340 | // Make a list of all reachable blocks in our CSE queue. | ||||
| 5341 | SmallVector<const DomTreeNode *, 8> CSEWorkList; | ||||
| 5342 | CSEWorkList.reserve(CSEBlocks.size()); | ||||
| 5343 | for (BasicBlock *BB : CSEBlocks) | ||||
| 5344 | if (DomTreeNode *N = DT->getNode(BB)) { | ||||
| 5345 | assert(DT->isReachableFromEntry(N))(static_cast <bool> (DT->isReachableFromEntry(N)) ? void (0) : __assert_fail ("DT->isReachableFromEntry(N)", "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5345, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5346 | CSEWorkList.push_back(N); | ||||
| 5347 | } | ||||
| 5348 | |||||
| 5349 | // Sort blocks by domination. This ensures we visit a block after all blocks | ||||
| 5350 | // dominating it are visited. | ||||
| 5351 | llvm::stable_sort(CSEWorkList, | ||||
| 5352 | [this](const DomTreeNode *A, const DomTreeNode *B) { | ||||
| 5353 | return DT->properlyDominates(A, B); | ||||
| 5354 | }); | ||||
| 5355 | |||||
| 5356 | // Perform O(N^2) search over the gather sequences and merge identical | ||||
| 5357 | // instructions. TODO: We can further optimize this scan if we split the | ||||
| 5358 | // instructions into different buckets based on the insert lane. | ||||
| 5359 | SmallVector<Instruction *, 16> Visited; | ||||
| 5360 | for (auto I = CSEWorkList.begin(), E = CSEWorkList.end(); I != E; ++I) { | ||||
| 5361 | assert(*I &&(static_cast <bool> (*I && (I == CSEWorkList.begin () || !DT->dominates(*I, *std::prev(I))) && "Worklist not sorted properly!" ) ? void (0) : __assert_fail ("*I && (I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) && \"Worklist not sorted properly!\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5363, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5362 | (I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) &&(static_cast <bool> (*I && (I == CSEWorkList.begin () || !DT->dominates(*I, *std::prev(I))) && "Worklist not sorted properly!" ) ? void (0) : __assert_fail ("*I && (I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) && \"Worklist not sorted properly!\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5363, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5363 | "Worklist not sorted properly!")(static_cast <bool> (*I && (I == CSEWorkList.begin () || !DT->dominates(*I, *std::prev(I))) && "Worklist not sorted properly!" ) ? void (0) : __assert_fail ("*I && (I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) && \"Worklist not sorted properly!\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5363, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5364 | BasicBlock *BB = (*I)->getBlock(); | ||||
| 5365 | // For all instructions in blocks containing gather sequences: | ||||
| 5366 | for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) { | ||||
| 5367 | Instruction *In = &*it++; | ||||
| 5368 | if (isDeleted(In)) | ||||
| 5369 | continue; | ||||
| 5370 | if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In)) | ||||
| 5371 | continue; | ||||
| 5372 | |||||
| 5373 | // Check if we can replace this instruction with any of the | ||||
| 5374 | // visited instructions. | ||||
| 5375 | for (Instruction *v : Visited) { | ||||
| 5376 | if (In->isIdenticalTo(v) && | ||||
| 5377 | DT->dominates(v->getParent(), In->getParent())) { | ||||
| 5378 | In->replaceAllUsesWith(v); | ||||
| 5379 | eraseInstruction(In); | ||||
| 5380 | In = nullptr; | ||||
| 5381 | break; | ||||
| 5382 | } | ||||
| 5383 | } | ||||
| 5384 | if (In) { | ||||
| 5385 | assert(!is_contained(Visited, In))(static_cast <bool> (!is_contained(Visited, In)) ? void (0) : __assert_fail ("!is_contained(Visited, In)", "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5385, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5386 | Visited.push_back(In); | ||||
| 5387 | } | ||||
| 5388 | } | ||||
| 5389 | } | ||||
| 5390 | CSEBlocks.clear(); | ||||
| 5391 | GatherSeq.clear(); | ||||
| 5392 | } | ||||
| 5393 | |||||
| 5394 | // Groups the instructions to a bundle (which is then a single scheduling entity) | ||||
| 5395 | // and schedules instructions until the bundle gets ready. | ||||
| 5396 | Optional<BoUpSLP::ScheduleData *> | ||||
| 5397 | BoUpSLP::BlockScheduling::tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP, | ||||
| 5398 | const InstructionsState &S) { | ||||
| 5399 | if (isa<PHINode>(S.OpValue)) | ||||
| 5400 | return nullptr; | ||||
| 5401 | |||||
| 5402 | // Initialize the instruction bundle. | ||||
| 5403 | Instruction *OldScheduleEnd = ScheduleEnd; | ||||
| 5404 | ScheduleData *PrevInBundle = nullptr; | ||||
| 5405 | ScheduleData *Bundle = nullptr; | ||||
| 5406 | bool ReSchedule = false; | ||||
| 5407 | LLVM_DEBUG(dbgs() << "SLP: bundle: " << *S.OpValue << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: bundle: " << *S.OpValue << "\n"; } } while (false); | ||||
| 5408 | |||||
| 5409 | auto &&TryScheduleBundle = [this, OldScheduleEnd, SLP](bool ReSchedule, | ||||
| 5410 | ScheduleData *Bundle) { | ||||
| 5411 | // The scheduling region got new instructions at the lower end (or it is a | ||||
| 5412 | // new region for the first bundle). This makes it necessary to | ||||
| 5413 | // recalculate all dependencies. | ||||
| 5414 | // It is seldom that this needs to be done a second time after adding the | ||||
| 5415 | // initial bundle to the region. | ||||
| 5416 | if (ScheduleEnd != OldScheduleEnd) { | ||||
| 5417 | for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) | ||||
| 5418 | doForAllOpcodes(I, [](ScheduleData *SD) { SD->clearDependencies(); }); | ||||
| 5419 | ReSchedule = true; | ||||
| 5420 | } | ||||
| 5421 | if (ReSchedule
| ||||
| 5422 | resetSchedule(); | ||||
| 5423 | initialFillReadyList(ReadyInsts); | ||||
| 5424 | } | ||||
| 5425 | if (Bundle
| ||||
| 5426 | LLVM_DEBUG(dbgs() << "SLP: try schedule bundle " << *Bundledo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: try schedule bundle " << *Bundle << " in block " << BB->getName() << "\n"; } } while (false) | ||||
| 5427 | << " in block " << BB->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: try schedule bundle " << *Bundle << " in block " << BB->getName() << "\n"; } } while (false); | ||||
| 5428 | calculateDependencies(Bundle, /*InsertInReadyList=*/true, SLP); | ||||
| 5429 | } | ||||
| 5430 | |||||
| 5431 | // Now try to schedule the new bundle or (if no bundle) just calculate | ||||
| 5432 | // dependencies. As soon as the bundle is "ready" it means that there are no | ||||
| 5433 | // cyclic dependencies and we can schedule it. Note that's important that we | ||||
| 5434 | // don't "schedule" the bundle yet (see cancelScheduling). | ||||
| 5435 | while (((!Bundle && ReSchedule) || (Bundle && !Bundle->isReady())) && | ||||
| 5436 | !ReadyInsts.empty()) { | ||||
| 5437 | ScheduleData *Picked = ReadyInsts.pop_back_val(); | ||||
| 5438 | if (Picked->isSchedulingEntity() && Picked->isReady()) | ||||
| 5439 | schedule(Picked, ReadyInsts); | ||||
| 5440 | } | ||||
| 5441 | }; | ||||
| 5442 | |||||
| 5443 | // Make sure that the scheduling region contains all | ||||
| 5444 | // instructions of the bundle. | ||||
| 5445 | for (Value *V : VL) { | ||||
| 5446 | if (!extendSchedulingRegion(V, S)) { | ||||
| 5447 | // If the scheduling region got new instructions at the lower end (or it | ||||
| 5448 | // is a new region for the first bundle). This makes it necessary to | ||||
| 5449 | // recalculate all dependencies. | ||||
| 5450 | // Otherwise the compiler may crash trying to incorrectly calculate | ||||
| 5451 | // dependencies and emit instruction in the wrong order at the actual | ||||
| 5452 | // scheduling. | ||||
| 5453 | TryScheduleBundle(/*ReSchedule=*/false, nullptr); | ||||
| 5454 | return None; | ||||
| 5455 | } | ||||
| 5456 | } | ||||
| 5457 | |||||
| 5458 | for (Value *V : VL) { | ||||
| 5459 | ScheduleData *BundleMember = getScheduleData(V); | ||||
| 5460 | assert(BundleMember &&(static_cast <bool> (BundleMember && "no ScheduleData for bundle member (maybe not in same basic block)" ) ? void (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member (maybe not in same basic block)\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5461, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5461 | "no ScheduleData for bundle member (maybe not in same basic block)")(static_cast <bool> (BundleMember && "no ScheduleData for bundle member (maybe not in same basic block)" ) ? void (0) : __assert_fail ("BundleMember && \"no ScheduleData for bundle member (maybe not in same basic block)\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5461, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5462 | if (BundleMember->IsScheduled) { | ||||
| 5463 | // A bundle member was scheduled as single instruction before and now | ||||
| 5464 | // needs to be scheduled as part of the bundle. We just get rid of the | ||||
| 5465 | // existing schedule. | ||||
| 5466 | LLVM_DEBUG(dbgs() << "SLP: reset schedule because " << *BundleMemberdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: reset schedule because " << *BundleMember << " was already scheduled\n"; } } while (false) | ||||
| 5467 | << " was already scheduled\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: reset schedule because " << *BundleMember << " was already scheduled\n"; } } while (false); | ||||
| 5468 | ReSchedule = true; | ||||
| 5469 | } | ||||
| 5470 | assert(BundleMember->isSchedulingEntity() &&(static_cast <bool> (BundleMember->isSchedulingEntity () && "bundle member already part of other bundle") ? void (0) : __assert_fail ("BundleMember->isSchedulingEntity() && \"bundle member already part of other bundle\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5471, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5471 | "bundle member already part of other bundle")(static_cast <bool> (BundleMember->isSchedulingEntity () && "bundle member already part of other bundle") ? void (0) : __assert_fail ("BundleMember->isSchedulingEntity() && \"bundle member already part of other bundle\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5471, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5472 | if (PrevInBundle
| ||||
| 5473 | PrevInBundle->NextInBundle = BundleMember; | ||||
| 5474 | } else { | ||||
| 5475 | Bundle = BundleMember; | ||||
| 5476 | } | ||||
| 5477 | BundleMember->UnscheduledDepsInBundle = 0; | ||||
| 5478 | Bundle->UnscheduledDepsInBundle += BundleMember->UnscheduledDeps; | ||||
| 5479 | |||||
| 5480 | // Group the instructions to a bundle. | ||||
| 5481 | BundleMember->FirstInBundle = Bundle; | ||||
| 5482 | PrevInBundle = BundleMember; | ||||
| 5483 | } | ||||
| 5484 | assert(Bundle && "Failed to find schedule bundle")(static_cast <bool> (Bundle && "Failed to find schedule bundle" ) ? void (0) : __assert_fail ("Bundle && \"Failed to find schedule bundle\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5484, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5485 | TryScheduleBundle(ReSchedule, Bundle); | ||||
| 5486 | if (!Bundle->isReady()) { | ||||
| 5487 | cancelScheduling(VL, S.OpValue); | ||||
| 5488 | return None; | ||||
| 5489 | } | ||||
| 5490 | return Bundle; | ||||
| 5491 | } | ||||
| 5492 | |||||
| 5493 | void BoUpSLP::BlockScheduling::cancelScheduling(ArrayRef<Value *> VL, | ||||
| 5494 | Value *OpValue) { | ||||
| 5495 | if (isa<PHINode>(OpValue)) | ||||
| 5496 | return; | ||||
| 5497 | |||||
| 5498 | ScheduleData *Bundle = getScheduleData(OpValue); | ||||
| 5499 | LLVM_DEBUG(dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n"; } } while (false); | ||||
| 5500 | assert(!Bundle->IsScheduled &&(static_cast <bool> (!Bundle->IsScheduled && "Can't cancel bundle which is already scheduled") ? void (0) : __assert_fail ("!Bundle->IsScheduled && \"Can't cancel bundle which is already scheduled\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5501, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5501 | "Can't cancel bundle which is already scheduled")(static_cast <bool> (!Bundle->IsScheduled && "Can't cancel bundle which is already scheduled") ? void (0) : __assert_fail ("!Bundle->IsScheduled && \"Can't cancel bundle which is already scheduled\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5501, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5502 | assert(Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() &&(static_cast <bool> (Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && "tried to unbundle something which is not a bundle" ) ? void (0) : __assert_fail ("Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && \"tried to unbundle something which is not a bundle\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5503, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5503 | "tried to unbundle something which is not a bundle")(static_cast <bool> (Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && "tried to unbundle something which is not a bundle" ) ? void (0) : __assert_fail ("Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() && \"tried to unbundle something which is not a bundle\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5503, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5504 | |||||
| 5505 | // Un-bundle: make single instructions out of the bundle. | ||||
| 5506 | ScheduleData *BundleMember = Bundle; | ||||
| 5507 | while (BundleMember) { | ||||
| 5508 | assert(BundleMember->FirstInBundle == Bundle && "corrupt bundle links")(static_cast <bool> (BundleMember->FirstInBundle == Bundle && "corrupt bundle links") ? void (0) : __assert_fail ("BundleMember->FirstInBundle == Bundle && \"corrupt bundle links\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5508, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5509 | BundleMember->FirstInBundle = BundleMember; | ||||
| 5510 | ScheduleData *Next = BundleMember->NextInBundle; | ||||
| 5511 | BundleMember->NextInBundle = nullptr; | ||||
| 5512 | BundleMember->UnscheduledDepsInBundle = BundleMember->UnscheduledDeps; | ||||
| 5513 | if (BundleMember->UnscheduledDepsInBundle == 0) { | ||||
| 5514 | ReadyInsts.insert(BundleMember); | ||||
| 5515 | } | ||||
| 5516 | BundleMember = Next; | ||||
| 5517 | } | ||||
| 5518 | } | ||||
| 5519 | |||||
| 5520 | BoUpSLP::ScheduleData *BoUpSLP::BlockScheduling::allocateScheduleDataChunks() { | ||||
| 5521 | // Allocate a new ScheduleData for the instruction. | ||||
| 5522 | if (ChunkPos >= ChunkSize) { | ||||
| 5523 | ScheduleDataChunks.push_back(std::make_unique<ScheduleData[]>(ChunkSize)); | ||||
| 5524 | ChunkPos = 0; | ||||
| 5525 | } | ||||
| 5526 | return &(ScheduleDataChunks.back()[ChunkPos++]); | ||||
| 5527 | } | ||||
| 5528 | |||||
| 5529 | bool BoUpSLP::BlockScheduling::extendSchedulingRegion(Value *V, | ||||
| 5530 | const InstructionsState &S) { | ||||
| 5531 | if (getScheduleData(V, isOneOf(S, V))) | ||||
| 5532 | return true; | ||||
| 5533 | Instruction *I = dyn_cast<Instruction>(V); | ||||
| 5534 | assert(I && "bundle member must be an instruction")(static_cast <bool> (I && "bundle member must be an instruction" ) ? void (0) : __assert_fail ("I && \"bundle member must be an instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5534, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5535 | assert(!isa<PHINode>(I) && "phi nodes don't need to be scheduled")(static_cast <bool> (!isa<PHINode>(I) && "phi nodes don't need to be scheduled" ) ? void (0) : __assert_fail ("!isa<PHINode>(I) && \"phi nodes don't need to be scheduled\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5535, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5536 | auto &&CheckSheduleForI = [this, &S](Instruction *I) -> bool { | ||||
| 5537 | ScheduleData *ISD = getScheduleData(I); | ||||
| 5538 | if (!ISD) | ||||
| 5539 | return false; | ||||
| 5540 | assert(isInSchedulingRegion(ISD) &&(static_cast <bool> (isInSchedulingRegion(ISD) && "ScheduleData not in scheduling region") ? void (0) : __assert_fail ("isInSchedulingRegion(ISD) && \"ScheduleData not in scheduling region\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5541, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5541 | "ScheduleData not in scheduling region")(static_cast <bool> (isInSchedulingRegion(ISD) && "ScheduleData not in scheduling region") ? void (0) : __assert_fail ("isInSchedulingRegion(ISD) && \"ScheduleData not in scheduling region\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5541, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5542 | ScheduleData *SD = allocateScheduleDataChunks(); | ||||
| 5543 | SD->Inst = I; | ||||
| 5544 | SD->init(SchedulingRegionID, S.OpValue); | ||||
| 5545 | ExtraScheduleDataMap[I][S.OpValue] = SD; | ||||
| 5546 | return true; | ||||
| 5547 | }; | ||||
| 5548 | if (CheckSheduleForI(I)) | ||||
| 5549 | return true; | ||||
| 5550 | if (!ScheduleStart) { | ||||
| 5551 | // It's the first instruction in the new region. | ||||
| 5552 | initScheduleData(I, I->getNextNode(), nullptr, nullptr); | ||||
| 5553 | ScheduleStart = I; | ||||
| 5554 | ScheduleEnd = I->getNextNode(); | ||||
| 5555 | if (isOneOf(S, I) != I) | ||||
| 5556 | CheckSheduleForI(I); | ||||
| 5557 | assert(ScheduleEnd && "tried to vectorize a terminator?")(static_cast <bool> (ScheduleEnd && "tried to vectorize a terminator?" ) ? void (0) : __assert_fail ("ScheduleEnd && \"tried to vectorize a terminator?\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5557, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5558 | LLVM_DEBUG(dbgs() << "SLP: initialize schedule region to " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: initialize schedule region to " << *I << "\n"; } } while (false); | ||||
| 5559 | return true; | ||||
| 5560 | } | ||||
| 5561 | // Search up and down at the same time, because we don't know if the new | ||||
| 5562 | // instruction is above or below the existing scheduling region. | ||||
| 5563 | BasicBlock::reverse_iterator UpIter = | ||||
| 5564 | ++ScheduleStart->getIterator().getReverse(); | ||||
| 5565 | BasicBlock::reverse_iterator UpperEnd = BB->rend(); | ||||
| 5566 | BasicBlock::iterator DownIter = ScheduleEnd->getIterator(); | ||||
| 5567 | BasicBlock::iterator LowerEnd = BB->end(); | ||||
| 5568 | while (UpIter != UpperEnd && DownIter != LowerEnd && &*UpIter != I && | ||||
| 5569 | &*DownIter != I) { | ||||
| 5570 | if (++ScheduleRegionSize > ScheduleRegionSizeLimit) { | ||||
| 5571 | LLVM_DEBUG(dbgs() << "SLP: exceeded schedule region size limit\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: exceeded schedule region size limit\n" ; } } while (false); | ||||
| 5572 | return false; | ||||
| 5573 | } | ||||
| 5574 | |||||
| 5575 | ++UpIter; | ||||
| 5576 | ++DownIter; | ||||
| 5577 | } | ||||
| 5578 | if (DownIter == LowerEnd || (UpIter != UpperEnd && &*UpIter == I)) { | ||||
| 5579 | assert(I->getParent() == ScheduleStart->getParent() &&(static_cast <bool> (I->getParent() == ScheduleStart ->getParent() && "Instruction is in wrong basic block." ) ? void (0) : __assert_fail ("I->getParent() == ScheduleStart->getParent() && \"Instruction is in wrong basic block.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5580, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5580 | "Instruction is in wrong basic block.")(static_cast <bool> (I->getParent() == ScheduleStart ->getParent() && "Instruction is in wrong basic block." ) ? void (0) : __assert_fail ("I->getParent() == ScheduleStart->getParent() && \"Instruction is in wrong basic block.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5580, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5581 | initScheduleData(I, ScheduleStart, nullptr, FirstLoadStoreInRegion); | ||||
| 5582 | ScheduleStart = I; | ||||
| 5583 | if (isOneOf(S, I) != I) | ||||
| 5584 | CheckSheduleForI(I); | ||||
| 5585 | LLVM_DEBUG(dbgs() << "SLP: extend schedule region start to " << *Ido { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: extend schedule region start to " << *I << "\n"; } } while (false) | ||||
| 5586 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: extend schedule region start to " << *I << "\n"; } } while (false); | ||||
| 5587 | return true; | ||||
| 5588 | } | ||||
| 5589 | assert((UpIter == UpperEnd || (DownIter != LowerEnd && &*DownIter == I)) &&(static_cast <bool> ((UpIter == UpperEnd || (DownIter != LowerEnd && &*DownIter == I)) && "Expected to reach top of the basic block or instruction down the " "lower end.") ? void (0) : __assert_fail ("(UpIter == UpperEnd || (DownIter != LowerEnd && &*DownIter == I)) && \"Expected to reach top of the basic block or instruction down the \" \"lower end.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5591, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5590 | "Expected to reach top of the basic block or instruction down the "(static_cast <bool> ((UpIter == UpperEnd || (DownIter != LowerEnd && &*DownIter == I)) && "Expected to reach top of the basic block or instruction down the " "lower end.") ? void (0) : __assert_fail ("(UpIter == UpperEnd || (DownIter != LowerEnd && &*DownIter == I)) && \"Expected to reach top of the basic block or instruction down the \" \"lower end.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5591, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5591 | "lower end.")(static_cast <bool> ((UpIter == UpperEnd || (DownIter != LowerEnd && &*DownIter == I)) && "Expected to reach top of the basic block or instruction down the " "lower end.") ? void (0) : __assert_fail ("(UpIter == UpperEnd || (DownIter != LowerEnd && &*DownIter == I)) && \"Expected to reach top of the basic block or instruction down the \" \"lower end.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5591, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5592 | assert(I->getParent() == ScheduleEnd->getParent() &&(static_cast <bool> (I->getParent() == ScheduleEnd-> getParent() && "Instruction is in wrong basic block." ) ? void (0) : __assert_fail ("I->getParent() == ScheduleEnd->getParent() && \"Instruction is in wrong basic block.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5593, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5593 | "Instruction is in wrong basic block.")(static_cast <bool> (I->getParent() == ScheduleEnd-> getParent() && "Instruction is in wrong basic block." ) ? void (0) : __assert_fail ("I->getParent() == ScheduleEnd->getParent() && \"Instruction is in wrong basic block.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5593, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5594 | initScheduleData(ScheduleEnd, I->getNextNode(), LastLoadStoreInRegion, | ||||
| 5595 | nullptr); | ||||
| 5596 | ScheduleEnd = I->getNextNode(); | ||||
| 5597 | if (isOneOf(S, I) != I) | ||||
| 5598 | CheckSheduleForI(I); | ||||
| 5599 | assert(ScheduleEnd && "tried to vectorize a terminator?")(static_cast <bool> (ScheduleEnd && "tried to vectorize a terminator?" ) ? void (0) : __assert_fail ("ScheduleEnd && \"tried to vectorize a terminator?\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5599, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5600 | LLVM_DEBUG(dbgs() << "SLP: extend schedule region end to " << *I << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: extend schedule region end to " << *I << "\n"; } } while (false); | ||||
| 5601 | return true; | ||||
| 5602 | } | ||||
| 5603 | |||||
| 5604 | void BoUpSLP::BlockScheduling::initScheduleData(Instruction *FromI, | ||||
| 5605 | Instruction *ToI, | ||||
| 5606 | ScheduleData *PrevLoadStore, | ||||
| 5607 | ScheduleData *NextLoadStore) { | ||||
| 5608 | ScheduleData *CurrentLoadStore = PrevLoadStore; | ||||
| 5609 | for (Instruction *I = FromI; I != ToI; I = I->getNextNode()) { | ||||
| 5610 | ScheduleData *SD = ScheduleDataMap[I]; | ||||
| 5611 | if (!SD) { | ||||
| 5612 | SD = allocateScheduleDataChunks(); | ||||
| 5613 | ScheduleDataMap[I] = SD; | ||||
| 5614 | SD->Inst = I; | ||||
| 5615 | } | ||||
| 5616 | assert(!isInSchedulingRegion(SD) &&(static_cast <bool> (!isInSchedulingRegion(SD) && "new ScheduleData already in scheduling region") ? void (0) : __assert_fail ("!isInSchedulingRegion(SD) && \"new ScheduleData already in scheduling region\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5617, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5617 | "new ScheduleData already in scheduling region")(static_cast <bool> (!isInSchedulingRegion(SD) && "new ScheduleData already in scheduling region") ? void (0) : __assert_fail ("!isInSchedulingRegion(SD) && \"new ScheduleData already in scheduling region\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5617, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5618 | SD->init(SchedulingRegionID, I); | ||||
| 5619 | |||||
| 5620 | if (I->mayReadOrWriteMemory() && | ||||
| 5621 | (!isa<IntrinsicInst>(I) || | ||||
| 5622 | (cast<IntrinsicInst>(I)->getIntrinsicID() != Intrinsic::sideeffect && | ||||
| 5623 | cast<IntrinsicInst>(I)->getIntrinsicID() != | ||||
| 5624 | Intrinsic::pseudoprobe))) { | ||||
| 5625 | // Update the linked list of memory accessing instructions. | ||||
| 5626 | if (CurrentLoadStore) { | ||||
| 5627 | CurrentLoadStore->NextLoadStore = SD; | ||||
| 5628 | } else { | ||||
| 5629 | FirstLoadStoreInRegion = SD; | ||||
| 5630 | } | ||||
| 5631 | CurrentLoadStore = SD; | ||||
| 5632 | } | ||||
| 5633 | } | ||||
| 5634 | if (NextLoadStore) { | ||||
| 5635 | if (CurrentLoadStore) | ||||
| 5636 | CurrentLoadStore->NextLoadStore = NextLoadStore; | ||||
| 5637 | } else { | ||||
| 5638 | LastLoadStoreInRegion = CurrentLoadStore; | ||||
| 5639 | } | ||||
| 5640 | } | ||||
| 5641 | |||||
| 5642 | void BoUpSLP::BlockScheduling::calculateDependencies(ScheduleData *SD, | ||||
| 5643 | bool InsertInReadyList, | ||||
| 5644 | BoUpSLP *SLP) { | ||||
| 5645 | assert(SD->isSchedulingEntity())(static_cast <bool> (SD->isSchedulingEntity()) ? void (0) : __assert_fail ("SD->isSchedulingEntity()", "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5645, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5646 | |||||
| 5647 | SmallVector<ScheduleData *, 10> WorkList; | ||||
| 5648 | WorkList.push_back(SD); | ||||
| 5649 | |||||
| 5650 | while (!WorkList.empty()) { | ||||
| 5651 | ScheduleData *SD = WorkList.pop_back_val(); | ||||
| 5652 | |||||
| 5653 | ScheduleData *BundleMember = SD; | ||||
| 5654 | while (BundleMember) { | ||||
| 5655 | assert(isInSchedulingRegion(BundleMember))(static_cast <bool> (isInSchedulingRegion(BundleMember) ) ? void (0) : __assert_fail ("isInSchedulingRegion(BundleMember)" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5655, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5656 | if (!BundleMember->hasValidDependencies()) { | ||||
| 5657 | |||||
| 5658 | LLVM_DEBUG(dbgs() << "SLP: update deps of " << *BundleMemberdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: update deps of " << *BundleMember << "\n"; } } while (false) | ||||
| 5659 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: update deps of " << *BundleMember << "\n"; } } while (false); | ||||
| 5660 | BundleMember->Dependencies = 0; | ||||
| 5661 | BundleMember->resetUnscheduledDeps(); | ||||
| 5662 | |||||
| 5663 | // Handle def-use chain dependencies. | ||||
| 5664 | if (BundleMember->OpValue != BundleMember->Inst) { | ||||
| 5665 | ScheduleData *UseSD = getScheduleData(BundleMember->Inst); | ||||
| 5666 | if (UseSD && isInSchedulingRegion(UseSD->FirstInBundle)) { | ||||
| 5667 | BundleMember->Dependencies++; | ||||
| 5668 | ScheduleData *DestBundle = UseSD->FirstInBundle; | ||||
| 5669 | if (!DestBundle->IsScheduled) | ||||
| 5670 | BundleMember->incrementUnscheduledDeps(1); | ||||
| 5671 | if (!DestBundle->hasValidDependencies()) | ||||
| 5672 | WorkList.push_back(DestBundle); | ||||
| 5673 | } | ||||
| 5674 | } else { | ||||
| 5675 | for (User *U : BundleMember->Inst->users()) { | ||||
| 5676 | if (isa<Instruction>(U)) { | ||||
| 5677 | ScheduleData *UseSD = getScheduleData(U); | ||||
| 5678 | if (UseSD && isInSchedulingRegion(UseSD->FirstInBundle)) { | ||||
| 5679 | BundleMember->Dependencies++; | ||||
| 5680 | ScheduleData *DestBundle = UseSD->FirstInBundle; | ||||
| 5681 | if (!DestBundle->IsScheduled) | ||||
| 5682 | BundleMember->incrementUnscheduledDeps(1); | ||||
| 5683 | if (!DestBundle->hasValidDependencies()) | ||||
| 5684 | WorkList.push_back(DestBundle); | ||||
| 5685 | } | ||||
| 5686 | } else { | ||||
| 5687 | // I'm not sure if this can ever happen. But we need to be safe. | ||||
| 5688 | // This lets the instruction/bundle never be scheduled and | ||||
| 5689 | // eventually disable vectorization. | ||||
| 5690 | BundleMember->Dependencies++; | ||||
| 5691 | BundleMember->incrementUnscheduledDeps(1); | ||||
| 5692 | } | ||||
| 5693 | } | ||||
| 5694 | } | ||||
| 5695 | |||||
| 5696 | // Handle the memory dependencies. | ||||
| 5697 | ScheduleData *DepDest = BundleMember->NextLoadStore; | ||||
| 5698 | if (DepDest) { | ||||
| 5699 | Instruction *SrcInst = BundleMember->Inst; | ||||
| 5700 | MemoryLocation SrcLoc = getLocation(SrcInst, SLP->AA); | ||||
| 5701 | bool SrcMayWrite = BundleMember->Inst->mayWriteToMemory(); | ||||
| 5702 | unsigned numAliased = 0; | ||||
| 5703 | unsigned DistToSrc = 1; | ||||
| 5704 | |||||
| 5705 | while (DepDest) { | ||||
| 5706 | assert(isInSchedulingRegion(DepDest))(static_cast <bool> (isInSchedulingRegion(DepDest)) ? void (0) : __assert_fail ("isInSchedulingRegion(DepDest)", "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5706, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5707 | |||||
| 5708 | // We have two limits to reduce the complexity: | ||||
| 5709 | // 1) AliasedCheckLimit: It's a small limit to reduce calls to | ||||
| 5710 | // SLP->isAliased (which is the expensive part in this loop). | ||||
| 5711 | // 2) MaxMemDepDistance: It's for very large blocks and it aborts | ||||
| 5712 | // the whole loop (even if the loop is fast, it's quadratic). | ||||
| 5713 | // It's important for the loop break condition (see below) to | ||||
| 5714 | // check this limit even between two read-only instructions. | ||||
| 5715 | if (DistToSrc >= MaxMemDepDistance || | ||||
| 5716 | ((SrcMayWrite || DepDest->Inst->mayWriteToMemory()) && | ||||
| 5717 | (numAliased >= AliasedCheckLimit || | ||||
| 5718 | SLP->isAliased(SrcLoc, SrcInst, DepDest->Inst)))) { | ||||
| 5719 | |||||
| 5720 | // We increment the counter only if the locations are aliased | ||||
| 5721 | // (instead of counting all alias checks). This gives a better | ||||
| 5722 | // balance between reduced runtime and accurate dependencies. | ||||
| 5723 | numAliased++; | ||||
| 5724 | |||||
| 5725 | DepDest->MemoryDependencies.push_back(BundleMember); | ||||
| 5726 | BundleMember->Dependencies++; | ||||
| 5727 | ScheduleData *DestBundle = DepDest->FirstInBundle; | ||||
| 5728 | if (!DestBundle->IsScheduled) { | ||||
| 5729 | BundleMember->incrementUnscheduledDeps(1); | ||||
| 5730 | } | ||||
| 5731 | if (!DestBundle->hasValidDependencies()) { | ||||
| 5732 | WorkList.push_back(DestBundle); | ||||
| 5733 | } | ||||
| 5734 | } | ||||
| 5735 | DepDest = DepDest->NextLoadStore; | ||||
| 5736 | |||||
| 5737 | // Example, explaining the loop break condition: Let's assume our | ||||
| 5738 | // starting instruction is i0 and MaxMemDepDistance = 3. | ||||
| 5739 | // | ||||
| 5740 | // +--------v--v--v | ||||
| 5741 | // i0,i1,i2,i3,i4,i5,i6,i7,i8 | ||||
| 5742 | // +--------^--^--^ | ||||
| 5743 | // | ||||
| 5744 | // MaxMemDepDistance let us stop alias-checking at i3 and we add | ||||
| 5745 | // dependencies from i0 to i3,i4,.. (even if they are not aliased). | ||||
| 5746 | // Previously we already added dependencies from i3 to i6,i7,i8 | ||||
| 5747 | // (because of MaxMemDepDistance). As we added a dependency from | ||||
| 5748 | // i0 to i3, we have transitive dependencies from i0 to i6,i7,i8 | ||||
| 5749 | // and we can abort this loop at i6. | ||||
| 5750 | if (DistToSrc >= 2 * MaxMemDepDistance) | ||||
| 5751 | break; | ||||
| 5752 | DistToSrc++; | ||||
| 5753 | } | ||||
| 5754 | } | ||||
| 5755 | } | ||||
| 5756 | BundleMember = BundleMember->NextInBundle; | ||||
| 5757 | } | ||||
| 5758 | if (InsertInReadyList
| ||||
| |||||
| 5759 | ReadyInsts.push_back(SD); | ||||
| 5760 | LLVM_DEBUG(dbgs() << "SLP: gets ready on update: " << *SD->Instdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready on update: " << *SD->Inst << "\n"; } } while (false) | ||||
| 5761 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: gets ready on update: " << *SD->Inst << "\n"; } } while (false); | ||||
| 5762 | } | ||||
| 5763 | } | ||||
| 5764 | } | ||||
| 5765 | |||||
| 5766 | void BoUpSLP::BlockScheduling::resetSchedule() { | ||||
| 5767 | assert(ScheduleStart &&(static_cast <bool> (ScheduleStart && "tried to reset schedule on block which has not been scheduled" ) ? void (0) : __assert_fail ("ScheduleStart && \"tried to reset schedule on block which has not been scheduled\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5768, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5768 | "tried to reset schedule on block which has not been scheduled")(static_cast <bool> (ScheduleStart && "tried to reset schedule on block which has not been scheduled" ) ? void (0) : __assert_fail ("ScheduleStart && \"tried to reset schedule on block which has not been scheduled\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5768, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5769 | for (Instruction *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) { | ||||
| 5770 | doForAllOpcodes(I, [&](ScheduleData *SD) { | ||||
| 5771 | assert(isInSchedulingRegion(SD) &&(static_cast <bool> (isInSchedulingRegion(SD) && "ScheduleData not in scheduling region") ? void (0) : __assert_fail ("isInSchedulingRegion(SD) && \"ScheduleData not in scheduling region\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5772, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5772 | "ScheduleData not in scheduling region")(static_cast <bool> (isInSchedulingRegion(SD) && "ScheduleData not in scheduling region") ? void (0) : __assert_fail ("isInSchedulingRegion(SD) && \"ScheduleData not in scheduling region\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5772, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5773 | SD->IsScheduled = false; | ||||
| 5774 | SD->resetUnscheduledDeps(); | ||||
| 5775 | }); | ||||
| 5776 | } | ||||
| 5777 | ReadyInsts.clear(); | ||||
| 5778 | } | ||||
| 5779 | |||||
| 5780 | void BoUpSLP::scheduleBlock(BlockScheduling *BS) { | ||||
| 5781 | if (!BS->ScheduleStart) | ||||
| 5782 | return; | ||||
| 5783 | |||||
| 5784 | LLVM_DEBUG(dbgs() << "SLP: schedule block " << BS->BB->getName() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: schedule block " << BS ->BB->getName() << "\n"; } } while (false); | ||||
| 5785 | |||||
| 5786 | BS->resetSchedule(); | ||||
| 5787 | |||||
| 5788 | // For the real scheduling we use a more sophisticated ready-list: it is | ||||
| 5789 | // sorted by the original instruction location. This lets the final schedule | ||||
| 5790 | // be as close as possible to the original instruction order. | ||||
| 5791 | struct ScheduleDataCompare { | ||||
| 5792 | bool operator()(ScheduleData *SD1, ScheduleData *SD2) const { | ||||
| 5793 | return SD2->SchedulingPriority < SD1->SchedulingPriority; | ||||
| 5794 | } | ||||
| 5795 | }; | ||||
| 5796 | std::set<ScheduleData *, ScheduleDataCompare> ReadyInsts; | ||||
| 5797 | |||||
| 5798 | // Ensure that all dependency data is updated and fill the ready-list with | ||||
| 5799 | // initial instructions. | ||||
| 5800 | int Idx = 0; | ||||
| 5801 | int NumToSchedule = 0; | ||||
| 5802 | for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd; | ||||
| 5803 | I = I->getNextNode()) { | ||||
| 5804 | BS->doForAllOpcodes(I, [this, &Idx, &NumToSchedule, BS](ScheduleData *SD) { | ||||
| 5805 | assert(SD->isPartOfBundle() ==(static_cast <bool> (SD->isPartOfBundle() == (getTreeEntry (SD->Inst) != nullptr) && "scheduler and vectorizer bundle mismatch" ) ? void (0) : __assert_fail ("SD->isPartOfBundle() == (getTreeEntry(SD->Inst) != nullptr) && \"scheduler and vectorizer bundle mismatch\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5807, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5806 | (getTreeEntry(SD->Inst) != nullptr) &&(static_cast <bool> (SD->isPartOfBundle() == (getTreeEntry (SD->Inst) != nullptr) && "scheduler and vectorizer bundle mismatch" ) ? void (0) : __assert_fail ("SD->isPartOfBundle() == (getTreeEntry(SD->Inst) != nullptr) && \"scheduler and vectorizer bundle mismatch\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5807, __extension__ __PRETTY_FUNCTION__)) | ||||
| 5807 | "scheduler and vectorizer bundle mismatch")(static_cast <bool> (SD->isPartOfBundle() == (getTreeEntry (SD->Inst) != nullptr) && "scheduler and vectorizer bundle mismatch" ) ? void (0) : __assert_fail ("SD->isPartOfBundle() == (getTreeEntry(SD->Inst) != nullptr) && \"scheduler and vectorizer bundle mismatch\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5807, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5808 | SD->FirstInBundle->SchedulingPriority = Idx++; | ||||
| 5809 | if (SD->isSchedulingEntity()) { | ||||
| 5810 | BS->calculateDependencies(SD, false, this); | ||||
| 5811 | NumToSchedule++; | ||||
| 5812 | } | ||||
| 5813 | }); | ||||
| 5814 | } | ||||
| 5815 | BS->initialFillReadyList(ReadyInsts); | ||||
| 5816 | |||||
| 5817 | Instruction *LastScheduledInst = BS->ScheduleEnd; | ||||
| 5818 | |||||
| 5819 | // Do the "real" scheduling. | ||||
| 5820 | while (!ReadyInsts.empty()) { | ||||
| 5821 | ScheduleData *picked = *ReadyInsts.begin(); | ||||
| 5822 | ReadyInsts.erase(ReadyInsts.begin()); | ||||
| 5823 | |||||
| 5824 | // Move the scheduled instruction(s) to their dedicated places, if not | ||||
| 5825 | // there yet. | ||||
| 5826 | ScheduleData *BundleMember = picked; | ||||
| 5827 | while (BundleMember) { | ||||
| 5828 | Instruction *pickedInst = BundleMember->Inst; | ||||
| 5829 | if (LastScheduledInst->getNextNode() != pickedInst) { | ||||
| 5830 | BS->BB->getInstList().remove(pickedInst); | ||||
| 5831 | BS->BB->getInstList().insert(LastScheduledInst->getIterator(), | ||||
| 5832 | pickedInst); | ||||
| 5833 | } | ||||
| 5834 | LastScheduledInst = pickedInst; | ||||
| 5835 | BundleMember = BundleMember->NextInBundle; | ||||
| 5836 | } | ||||
| 5837 | |||||
| 5838 | BS->schedule(picked, ReadyInsts); | ||||
| 5839 | NumToSchedule--; | ||||
| 5840 | } | ||||
| 5841 | assert(NumToSchedule == 0 && "could not schedule all instructions")(static_cast <bool> (NumToSchedule == 0 && "could not schedule all instructions" ) ? void (0) : __assert_fail ("NumToSchedule == 0 && \"could not schedule all instructions\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 5841, __extension__ __PRETTY_FUNCTION__)); | ||||
| 5842 | |||||
| 5843 | // Avoid duplicate scheduling of the block. | ||||
| 5844 | BS->ScheduleStart = nullptr; | ||||
| 5845 | } | ||||
| 5846 | |||||
| 5847 | unsigned BoUpSLP::getVectorElementSize(Value *V) { | ||||
| 5848 | // If V is a store, just return the width of the stored value (or value | ||||
| 5849 | // truncated just before storing) without traversing the expression tree. | ||||
| 5850 | // This is the common case. | ||||
| 5851 | if (auto *Store = dyn_cast<StoreInst>(V)) { | ||||
| 5852 | if (auto *Trunc = dyn_cast<TruncInst>(Store->getValueOperand())) | ||||
| 5853 | return DL->getTypeSizeInBits(Trunc->getSrcTy()); | ||||
| 5854 | return DL->getTypeSizeInBits(Store->getValueOperand()->getType()); | ||||
| 5855 | } | ||||
| 5856 | |||||
| 5857 | auto E = InstrElementSize.find(V); | ||||
| 5858 | if (E != InstrElementSize.end()) | ||||
| 5859 | return E->second; | ||||
| 5860 | |||||
| 5861 | // If V is not a store, we can traverse the expression tree to find loads | ||||
| 5862 | // that feed it. The type of the loaded value may indicate a more suitable | ||||
| 5863 | // width than V's type. We want to base the vector element size on the width | ||||
| 5864 | // of memory operations where possible. | ||||
| 5865 | SmallVector<std::pair<Instruction *, BasicBlock *>, 16> Worklist; | ||||
| 5866 | SmallPtrSet<Instruction *, 16> Visited; | ||||
| 5867 | if (auto *I = dyn_cast<Instruction>(V)) { | ||||
| 5868 | Worklist.emplace_back(I, I->getParent()); | ||||
| 5869 | Visited.insert(I); | ||||
| 5870 | } | ||||
| 5871 | |||||
| 5872 | // Traverse the expression tree in bottom-up order looking for loads. If we | ||||
| 5873 | // encounter an instruction we don't yet handle, we give up. | ||||
| 5874 | auto Width = 0u; | ||||
| 5875 | while (!Worklist.empty()) { | ||||
| 5876 | Instruction *I; | ||||
| 5877 | BasicBlock *Parent; | ||||
| 5878 | std::tie(I, Parent) = Worklist.pop_back_val(); | ||||
| 5879 | |||||
| 5880 | // We should only be looking at scalar instructions here. If the current | ||||
| 5881 | // instruction has a vector type, skip. | ||||
| 5882 | auto *Ty = I->getType(); | ||||
| 5883 | if (isa<VectorType>(Ty)) | ||||
| 5884 | continue; | ||||
| 5885 | |||||
| 5886 | // If the current instruction is a load, update MaxWidth to reflect the | ||||
| 5887 | // width of the loaded value. | ||||
| 5888 | if (isa<LoadInst>(I) || isa<ExtractElementInst>(I) || | ||||
| 5889 | isa<ExtractValueInst>(I)) | ||||
| 5890 | Width = std::max<unsigned>(Width, DL->getTypeSizeInBits(Ty)); | ||||
| 5891 | |||||
| 5892 | // Otherwise, we need to visit the operands of the instruction. We only | ||||
| 5893 | // handle the interesting cases from buildTree here. If an operand is an | ||||
| 5894 | // instruction we haven't yet visited and from the same basic block as the | ||||
| 5895 | // user or the use is a PHI node, we add it to the worklist. | ||||
| 5896 | else if (isa<PHINode>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I) || | ||||
| 5897 | isa<CmpInst>(I) || isa<SelectInst>(I) || isa<BinaryOperator>(I) || | ||||
| 5898 | isa<UnaryOperator>(I)) { | ||||
| 5899 | for (Use &U : I->operands()) | ||||
| 5900 | if (auto *J = dyn_cast<Instruction>(U.get())) | ||||
| 5901 | if (Visited.insert(J).second && | ||||
| 5902 | (isa<PHINode>(I) || J->getParent() == Parent)) | ||||
| 5903 | Worklist.emplace_back(J, J->getParent()); | ||||
| 5904 | } else { | ||||
| 5905 | break; | ||||
| 5906 | } | ||||
| 5907 | } | ||||
| 5908 | |||||
| 5909 | // If we didn't encounter a memory access in the expression tree, or if we | ||||
| 5910 | // gave up for some reason, just return the width of V. Otherwise, return the | ||||
| 5911 | // maximum width we found. | ||||
| 5912 | if (!Width) { | ||||
| 5913 | if (auto *CI = dyn_cast<CmpInst>(V)) | ||||
| 5914 | V = CI->getOperand(0); | ||||
| 5915 | Width = DL->getTypeSizeInBits(V->getType()); | ||||
| 5916 | } | ||||
| 5917 | |||||
| 5918 | for (Instruction *I : Visited) | ||||
| 5919 | InstrElementSize[I] = Width; | ||||
| 5920 | |||||
| 5921 | return Width; | ||||
| 5922 | } | ||||
| 5923 | |||||
| 5924 | // Determine if a value V in a vectorizable expression Expr can be demoted to a | ||||
| 5925 | // smaller type with a truncation. We collect the values that will be demoted | ||||
| 5926 | // in ToDemote and additional roots that require investigating in Roots. | ||||
| 5927 | static bool collectValuesToDemote(Value *V, SmallPtrSetImpl<Value *> &Expr, | ||||
| 5928 | SmallVectorImpl<Value *> &ToDemote, | ||||
| 5929 | SmallVectorImpl<Value *> &Roots) { | ||||
| 5930 | // We can always demote constants. | ||||
| 5931 | if (isa<Constant>(V)) { | ||||
| 5932 | ToDemote.push_back(V); | ||||
| 5933 | return true; | ||||
| 5934 | } | ||||
| 5935 | |||||
| 5936 | // If the value is not an instruction in the expression with only one use, it | ||||
| 5937 | // cannot be demoted. | ||||
| 5938 | auto *I = dyn_cast<Instruction>(V); | ||||
| 5939 | if (!I || !I->hasOneUse() || !Expr.count(I)) | ||||
| 5940 | return false; | ||||
| 5941 | |||||
| 5942 | switch (I->getOpcode()) { | ||||
| 5943 | |||||
| 5944 | // We can always demote truncations and extensions. Since truncations can | ||||
| 5945 | // seed additional demotion, we save the truncated value. | ||||
| 5946 | case Instruction::Trunc: | ||||
| 5947 | Roots.push_back(I->getOperand(0)); | ||||
| 5948 | break; | ||||
| 5949 | case Instruction::ZExt: | ||||
| 5950 | case Instruction::SExt: | ||||
| 5951 | break; | ||||
| 5952 | |||||
| 5953 | // We can demote certain binary operations if we can demote both of their | ||||
| 5954 | // operands. | ||||
| 5955 | case Instruction::Add: | ||||
| 5956 | case Instruction::Sub: | ||||
| 5957 | case Instruction::Mul: | ||||
| 5958 | case Instruction::And: | ||||
| 5959 | case Instruction::Or: | ||||
| 5960 | case Instruction::Xor: | ||||
| 5961 | if (!collectValuesToDemote(I->getOperand(0), Expr, ToDemote, Roots) || | ||||
| 5962 | !collectValuesToDemote(I->getOperand(1), Expr, ToDemote, Roots)) | ||||
| 5963 | return false; | ||||
| 5964 | break; | ||||
| 5965 | |||||
| 5966 | // We can demote selects if we can demote their true and false values. | ||||
| 5967 | case Instruction::Select: { | ||||
| 5968 | SelectInst *SI = cast<SelectInst>(I); | ||||
| 5969 | if (!collectValuesToDemote(SI->getTrueValue(), Expr, ToDemote, Roots) || | ||||
| 5970 | !collectValuesToDemote(SI->getFalseValue(), Expr, ToDemote, Roots)) | ||||
| 5971 | return false; | ||||
| 5972 | break; | ||||
| 5973 | } | ||||
| 5974 | |||||
| 5975 | // We can demote phis if we can demote all their incoming operands. Note that | ||||
| 5976 | // we don't need to worry about cycles since we ensure single use above. | ||||
| 5977 | case Instruction::PHI: { | ||||
| 5978 | PHINode *PN = cast<PHINode>(I); | ||||
| 5979 | for (Value *IncValue : PN->incoming_values()) | ||||
| 5980 | if (!collectValuesToDemote(IncValue, Expr, ToDemote, Roots)) | ||||
| 5981 | return false; | ||||
| 5982 | break; | ||||
| 5983 | } | ||||
| 5984 | |||||
| 5985 | // Otherwise, conservatively give up. | ||||
| 5986 | default: | ||||
| 5987 | return false; | ||||
| 5988 | } | ||||
| 5989 | |||||
| 5990 | // Record the value that we can demote. | ||||
| 5991 | ToDemote.push_back(V); | ||||
| 5992 | return true; | ||||
| 5993 | } | ||||
| 5994 | |||||
| 5995 | void BoUpSLP::computeMinimumValueSizes() { | ||||
| 5996 | // If there are no external uses, the expression tree must be rooted by a | ||||
| 5997 | // store. We can't demote in-memory values, so there is nothing to do here. | ||||
| 5998 | if (ExternalUses.empty()) | ||||
| 5999 | return; | ||||
| 6000 | |||||
| 6001 | // We only attempt to truncate integer expressions. | ||||
| 6002 | auto &TreeRoot = VectorizableTree[0]->Scalars; | ||||
| 6003 | auto *TreeRootIT = dyn_cast<IntegerType>(TreeRoot[0]->getType()); | ||||
| 6004 | if (!TreeRootIT) | ||||
| 6005 | return; | ||||
| 6006 | |||||
| 6007 | // If the expression is not rooted by a store, these roots should have | ||||
| 6008 | // external uses. We will rely on InstCombine to rewrite the expression in | ||||
| 6009 | // the narrower type. However, InstCombine only rewrites single-use values. | ||||
| 6010 | // This means that if a tree entry other than a root is used externally, it | ||||
| 6011 | // must have multiple uses and InstCombine will not rewrite it. The code | ||||
| 6012 | // below ensures that only the roots are used externally. | ||||
| 6013 | SmallPtrSet<Value *, 32> Expr(TreeRoot.begin(), TreeRoot.end()); | ||||
| 6014 | for (auto &EU : ExternalUses) | ||||
| 6015 | if (!Expr.erase(EU.Scalar)) | ||||
| 6016 | return; | ||||
| 6017 | if (!Expr.empty()) | ||||
| 6018 | return; | ||||
| 6019 | |||||
| 6020 | // Collect the scalar values of the vectorizable expression. We will use this | ||||
| 6021 | // context to determine which values can be demoted. If we see a truncation, | ||||
| 6022 | // we mark it as seeding another demotion. | ||||
| 6023 | for (auto &EntryPtr : VectorizableTree) | ||||
| 6024 | Expr.insert(EntryPtr->Scalars.begin(), EntryPtr->Scalars.end()); | ||||
| 6025 | |||||
| 6026 | // Ensure the roots of the vectorizable tree don't form a cycle. They must | ||||
| 6027 | // have a single external user that is not in the vectorizable tree. | ||||
| 6028 | for (auto *Root : TreeRoot) | ||||
| 6029 | if (!Root->hasOneUse() || Expr.count(*Root->user_begin())) | ||||
| 6030 | return; | ||||
| 6031 | |||||
| 6032 | // Conservatively determine if we can actually truncate the roots of the | ||||
| 6033 | // expression. Collect the values that can be demoted in ToDemote and | ||||
| 6034 | // additional roots that require investigating in Roots. | ||||
| 6035 | SmallVector<Value *, 32> ToDemote; | ||||
| 6036 | SmallVector<Value *, 4> Roots; | ||||
| 6037 | for (auto *Root : TreeRoot) | ||||
| 6038 | if (!collectValuesToDemote(Root, Expr, ToDemote, Roots)) | ||||
| 6039 | return; | ||||
| 6040 | |||||
| 6041 | // The maximum bit width required to represent all the values that can be | ||||
| 6042 | // demoted without loss of precision. It would be safe to truncate the roots | ||||
| 6043 | // of the expression to this width. | ||||
| 6044 | auto MaxBitWidth = 8u; | ||||
| 6045 | |||||
| 6046 | // We first check if all the bits of the roots are demanded. If they're not, | ||||
| 6047 | // we can truncate the roots to this narrower type. | ||||
| 6048 | for (auto *Root : TreeRoot) { | ||||
| 6049 | auto Mask = DB->getDemandedBits(cast<Instruction>(Root)); | ||||
| 6050 | MaxBitWidth = std::max<unsigned>( | ||||
| 6051 | Mask.getBitWidth() - Mask.countLeadingZeros(), MaxBitWidth); | ||||
| 6052 | } | ||||
| 6053 | |||||
| 6054 | // True if the roots can be zero-extended back to their original type, rather | ||||
| 6055 | // than sign-extended. We know that if the leading bits are not demanded, we | ||||
| 6056 | // can safely zero-extend. So we initialize IsKnownPositive to True. | ||||
| 6057 | bool IsKnownPositive = true; | ||||
| 6058 | |||||
| 6059 | // If all the bits of the roots are demanded, we can try a little harder to | ||||
| 6060 | // compute a narrower type. This can happen, for example, if the roots are | ||||
| 6061 | // getelementptr indices. InstCombine promotes these indices to the pointer | ||||
| 6062 | // width. Thus, all their bits are technically demanded even though the | ||||
| 6063 | // address computation might be vectorized in a smaller type. | ||||
| 6064 | // | ||||
| 6065 | // We start by looking at each entry that can be demoted. We compute the | ||||
| 6066 | // maximum bit width required to store the scalar by using ValueTracking to | ||||
| 6067 | // compute the number of high-order bits we can truncate. | ||||
| 6068 | if (MaxBitWidth == DL->getTypeSizeInBits(TreeRoot[0]->getType()) && | ||||
| 6069 | llvm::all_of(TreeRoot, [](Value *R) { | ||||
| 6070 | assert(R->hasOneUse() && "Root should have only one use!")(static_cast <bool> (R->hasOneUse() && "Root should have only one use!" ) ? void (0) : __assert_fail ("R->hasOneUse() && \"Root should have only one use!\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6070, __extension__ __PRETTY_FUNCTION__)); | ||||
| 6071 | return isa<GetElementPtrInst>(R->user_back()); | ||||
| 6072 | })) { | ||||
| 6073 | MaxBitWidth = 8u; | ||||
| 6074 | |||||
| 6075 | // Determine if the sign bit of all the roots is known to be zero. If not, | ||||
| 6076 | // IsKnownPositive is set to False. | ||||
| 6077 | IsKnownPositive = llvm::all_of(TreeRoot, [&](Value *R) { | ||||
| 6078 | KnownBits Known = computeKnownBits(R, *DL); | ||||
| 6079 | return Known.isNonNegative(); | ||||
| 6080 | }); | ||||
| 6081 | |||||
| 6082 | // Determine the maximum number of bits required to store the scalar | ||||
| 6083 | // values. | ||||
| 6084 | for (auto *Scalar : ToDemote) { | ||||
| 6085 | auto NumSignBits = ComputeNumSignBits(Scalar, *DL, 0, AC, nullptr, DT); | ||||
| 6086 | auto NumTypeBits = DL->getTypeSizeInBits(Scalar->getType()); | ||||
| 6087 | MaxBitWidth = std::max<unsigned>(NumTypeBits - NumSignBits, MaxBitWidth); | ||||
| 6088 | } | ||||
| 6089 | |||||
| 6090 | // If we can't prove that the sign bit is zero, we must add one to the | ||||
| 6091 | // maximum bit width to account for the unknown sign bit. This preserves | ||||
| 6092 | // the existing sign bit so we can safely sign-extend the root back to the | ||||
| 6093 | // original type. Otherwise, if we know the sign bit is zero, we will | ||||
| 6094 | // zero-extend the root instead. | ||||
| 6095 | // | ||||
| 6096 | // FIXME: This is somewhat suboptimal, as there will be cases where adding | ||||
| 6097 | // one to the maximum bit width will yield a larger-than-necessary | ||||
| 6098 | // type. In general, we need to add an extra bit only if we can't | ||||
| 6099 | // prove that the upper bit of the original type is equal to the | ||||
| 6100 | // upper bit of the proposed smaller type. If these two bits are the | ||||
| 6101 | // same (either zero or one) we know that sign-extending from the | ||||
| 6102 | // smaller type will result in the same value. Here, since we can't | ||||
| 6103 | // yet prove this, we are just making the proposed smaller type | ||||
| 6104 | // larger to ensure correctness. | ||||
| 6105 | if (!IsKnownPositive) | ||||
| 6106 | ++MaxBitWidth; | ||||
| 6107 | } | ||||
| 6108 | |||||
| 6109 | // Round MaxBitWidth up to the next power-of-two. | ||||
| 6110 | if (!isPowerOf2_64(MaxBitWidth)) | ||||
| 6111 | MaxBitWidth = NextPowerOf2(MaxBitWidth); | ||||
| 6112 | |||||
| 6113 | // If the maximum bit width we compute is less than the with of the roots' | ||||
| 6114 | // type, we can proceed with the narrowing. Otherwise, do nothing. | ||||
| 6115 | if (MaxBitWidth >= TreeRootIT->getBitWidth()) | ||||
| 6116 | return; | ||||
| 6117 | |||||
| 6118 | // If we can truncate the root, we must collect additional values that might | ||||
| 6119 | // be demoted as a result. That is, those seeded by truncations we will | ||||
| 6120 | // modify. | ||||
| 6121 | while (!Roots.empty()) | ||||
| 6122 | collectValuesToDemote(Roots.pop_back_val(), Expr, ToDemote, Roots); | ||||
| 6123 | |||||
| 6124 | // Finally, map the values we can demote to the maximum bit with we computed. | ||||
| 6125 | for (auto *Scalar : ToDemote) | ||||
| 6126 | MinBWs[Scalar] = std::make_pair(MaxBitWidth, !IsKnownPositive); | ||||
| 6127 | } | ||||
| 6128 | |||||
| 6129 | namespace { | ||||
| 6130 | |||||
| 6131 | /// The SLPVectorizer Pass. | ||||
| 6132 | struct SLPVectorizer : public FunctionPass { | ||||
| 6133 | SLPVectorizerPass Impl; | ||||
| 6134 | |||||
| 6135 | /// Pass identification, replacement for typeid | ||||
| 6136 | static char ID; | ||||
| 6137 | |||||
| 6138 | explicit SLPVectorizer() : FunctionPass(ID) { | ||||
| 6139 | initializeSLPVectorizerPass(*PassRegistry::getPassRegistry()); | ||||
| 6140 | } | ||||
| 6141 | |||||
| 6142 | bool doInitialization(Module &M) override { | ||||
| 6143 | return false; | ||||
| 6144 | } | ||||
| 6145 | |||||
| 6146 | bool runOnFunction(Function &F) override { | ||||
| 6147 | if (skipFunction(F)) | ||||
| 6148 | return false; | ||||
| 6149 | |||||
| 6150 | auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); | ||||
| 6151 | auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | ||||
| 6152 | auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); | ||||
| 6153 | auto *TLI = TLIP ? &TLIP->getTLI(F) : nullptr; | ||||
| 6154 | auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); | ||||
| 6155 | auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | ||||
| 6156 | auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | ||||
| 6157 | auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); | ||||
| 6158 | auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits(); | ||||
| 6159 | auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE(); | ||||
| 6160 | |||||
| 6161 | return Impl.runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); | ||||
| 6162 | } | ||||
| 6163 | |||||
| 6164 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||
| 6165 | FunctionPass::getAnalysisUsage(AU); | ||||
| 6166 | AU.addRequired<AssumptionCacheTracker>(); | ||||
| 6167 | AU.addRequired<ScalarEvolutionWrapperPass>(); | ||||
| 6168 | AU.addRequired<AAResultsWrapperPass>(); | ||||
| 6169 | AU.addRequired<TargetTransformInfoWrapperPass>(); | ||||
| 6170 | AU.addRequired<LoopInfoWrapperPass>(); | ||||
| 6171 | AU.addRequired<DominatorTreeWrapperPass>(); | ||||
| 6172 | AU.addRequired<DemandedBitsWrapperPass>(); | ||||
| 6173 | AU.addRequired<OptimizationRemarkEmitterWrapperPass>(); | ||||
| 6174 | AU.addRequired<InjectTLIMappingsLegacy>(); | ||||
| 6175 | AU.addPreserved<LoopInfoWrapperPass>(); | ||||
| 6176 | AU.addPreserved<DominatorTreeWrapperPass>(); | ||||
| 6177 | AU.addPreserved<AAResultsWrapperPass>(); | ||||
| 6178 | AU.addPreserved<GlobalsAAWrapperPass>(); | ||||
| 6179 | AU.setPreservesCFG(); | ||||
| 6180 | } | ||||
| 6181 | }; | ||||
| 6182 | |||||
| 6183 | } // end anonymous namespace | ||||
| 6184 | |||||
| 6185 | PreservedAnalyses SLPVectorizerPass::run(Function &F, FunctionAnalysisManager &AM) { | ||||
| 6186 | auto *SE = &AM.getResult<ScalarEvolutionAnalysis>(F); | ||||
| 6187 | auto *TTI = &AM.getResult<TargetIRAnalysis>(F); | ||||
| 6188 | auto *TLI = AM.getCachedResult<TargetLibraryAnalysis>(F); | ||||
| 6189 | auto *AA = &AM.getResult<AAManager>(F); | ||||
| 6190 | auto *LI = &AM.getResult<LoopAnalysis>(F); | ||||
| 6191 | auto *DT = &AM.getResult<DominatorTreeAnalysis>(F); | ||||
| 6192 | auto *AC = &AM.getResult<AssumptionAnalysis>(F); | ||||
| 6193 | auto *DB = &AM.getResult<DemandedBitsAnalysis>(F); | ||||
| 6194 | auto *ORE = &AM.getResult<OptimizationRemarkEmitterAnalysis>(F); | ||||
| 6195 | |||||
| 6196 | bool Changed = runImpl(F, SE, TTI, TLI, AA, LI, DT, AC, DB, ORE); | ||||
| 6197 | if (!Changed) | ||||
| 6198 | return PreservedAnalyses::all(); | ||||
| 6199 | |||||
| 6200 | PreservedAnalyses PA; | ||||
| 6201 | PA.preserveSet<CFGAnalyses>(); | ||||
| 6202 | PA.preserve<AAManager>(); | ||||
| 6203 | PA.preserve<GlobalsAA>(); | ||||
| 6204 | return PA; | ||||
| 6205 | } | ||||
| 6206 | |||||
| 6207 | bool SLPVectorizerPass::runImpl(Function &F, ScalarEvolution *SE_, | ||||
| 6208 | TargetTransformInfo *TTI_, | ||||
| 6209 | TargetLibraryInfo *TLI_, AAResults *AA_, | ||||
| 6210 | LoopInfo *LI_, DominatorTree *DT_, | ||||
| 6211 | AssumptionCache *AC_, DemandedBits *DB_, | ||||
| 6212 | OptimizationRemarkEmitter *ORE_) { | ||||
| 6213 | if (!RunSLPVectorization) | ||||
| 6214 | return false; | ||||
| 6215 | SE = SE_; | ||||
| 6216 | TTI = TTI_; | ||||
| 6217 | TLI = TLI_; | ||||
| 6218 | AA = AA_; | ||||
| 6219 | LI = LI_; | ||||
| 6220 | DT = DT_; | ||||
| 6221 | AC = AC_; | ||||
| 6222 | DB = DB_; | ||||
| 6223 | DL = &F.getParent()->getDataLayout(); | ||||
| 6224 | |||||
| 6225 | Stores.clear(); | ||||
| 6226 | GEPs.clear(); | ||||
| 6227 | bool Changed = false; | ||||
| 6228 | |||||
| 6229 | // If the target claims to have no vector registers don't attempt | ||||
| 6230 | // vectorization. | ||||
| 6231 | if (!TTI->getNumberOfRegisters(TTI->getRegisterClassForType(true))) | ||||
| 6232 | return false; | ||||
| 6233 | |||||
| 6234 | // Don't vectorize when the attribute NoImplicitFloat is used. | ||||
| 6235 | if (F.hasFnAttribute(Attribute::NoImplicitFloat)) | ||||
| 6236 | return false; | ||||
| 6237 | |||||
| 6238 | LLVM_DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n"; } } while (false); | ||||
| 6239 | |||||
| 6240 | // Use the bottom up slp vectorizer to construct chains that start with | ||||
| 6241 | // store instructions. | ||||
| 6242 | BoUpSLP R(&F, SE, TTI, TLI, AA, LI, DT, AC, DB, DL, ORE_); | ||||
| 6243 | |||||
| 6244 | // A general note: the vectorizer must use BoUpSLP::eraseInstruction() to | ||||
| 6245 | // delete instructions. | ||||
| 6246 | |||||
| 6247 | // Scan the blocks in the function in post order. | ||||
| 6248 | for (auto BB : post_order(&F.getEntryBlock())) { | ||||
| 6249 | collectSeedInstructions(BB); | ||||
| 6250 | |||||
| 6251 | // Vectorize trees that end at stores. | ||||
| 6252 | if (!Stores.empty()) { | ||||
| 6253 | LLVM_DEBUG(dbgs() << "SLP: Found stores for " << Stores.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found stores for " << Stores .size() << " underlying objects.\n"; } } while (false) | ||||
| 6254 | << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found stores for " << Stores .size() << " underlying objects.\n"; } } while (false); | ||||
| 6255 | Changed |= vectorizeStoreChains(R); | ||||
| 6256 | } | ||||
| 6257 | |||||
| 6258 | // Vectorize trees that end at reductions. | ||||
| 6259 | Changed |= vectorizeChainsInBlock(BB, R); | ||||
| 6260 | |||||
| 6261 | // Vectorize the index computations of getelementptr instructions. This | ||||
| 6262 | // is primarily intended to catch gather-like idioms ending at | ||||
| 6263 | // non-consecutive loads. | ||||
| 6264 | if (!GEPs.empty()) { | ||||
| 6265 | LLVM_DEBUG(dbgs() << "SLP: Found GEPs for " << GEPs.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found GEPs for " << GEPs .size() << " underlying objects.\n"; } } while (false) | ||||
| 6266 | << " underlying objects.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found GEPs for " << GEPs .size() << " underlying objects.\n"; } } while (false); | ||||
| 6267 | Changed |= vectorizeGEPIndices(BB, R); | ||||
| 6268 | } | ||||
| 6269 | } | ||||
| 6270 | |||||
| 6271 | if (Changed) { | ||||
| 6272 | R.optimizeGatherSequence(); | ||||
| 6273 | LLVM_DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: vectorized \"" << F.getName () << "\"\n"; } } while (false); | ||||
| 6274 | } | ||||
| 6275 | return Changed; | ||||
| 6276 | } | ||||
| 6277 | |||||
| 6278 | bool SLPVectorizerPass::vectorizeStoreChain(ArrayRef<Value *> Chain, BoUpSLP &R, | ||||
| 6279 | unsigned Idx) { | ||||
| 6280 | LLVM_DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << Chain.size()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << Chain.size() << "\n"; } } while (false) | ||||
| 6281 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << Chain.size() << "\n"; } } while (false); | ||||
| 6282 | const unsigned Sz = R.getVectorElementSize(Chain[0]); | ||||
| 6283 | const unsigned MinVF = R.getMinVecRegSize() / Sz; | ||||
| 6284 | unsigned VF = Chain.size(); | ||||
| 6285 | |||||
| 6286 | if (!isPowerOf2_32(Sz) || !isPowerOf2_32(VF) || VF < 2 || VF < MinVF) | ||||
| 6287 | return false; | ||||
| 6288 | |||||
| 6289 | LLVM_DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << Idxdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << VF << " stores at offset " << Idx << "\n"; } } while ( false) | ||||
| 6290 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << VF << " stores at offset " << Idx << "\n"; } } while ( false); | ||||
| 6291 | |||||
| 6292 | R.buildTree(Chain); | ||||
| 6293 | Optional<ArrayRef<unsigned>> Order = R.bestOrder(); | ||||
| 6294 | // TODO: Handle orders of size less than number of elements in the vector. | ||||
| 6295 | if (Order && Order->size() == Chain.size()) { | ||||
| 6296 | // TODO: reorder tree nodes without tree rebuilding. | ||||
| 6297 | SmallVector<Value *, 4> ReorderedOps(Chain.rbegin(), Chain.rend()); | ||||
| 6298 | llvm::transform(*Order, ReorderedOps.begin(), | ||||
| 6299 | [Chain](const unsigned Idx) { return Chain[Idx]; }); | ||||
| 6300 | R.buildTree(ReorderedOps); | ||||
| 6301 | } | ||||
| 6302 | if (R.isTreeTinyAndNotFullyVectorizable()) | ||||
| 6303 | return false; | ||||
| 6304 | if (R.isLoadCombineCandidate()) | ||||
| 6305 | return false; | ||||
| 6306 | |||||
| 6307 | R.computeMinimumValueSizes(); | ||||
| 6308 | |||||
| 6309 | InstructionCost Cost = R.getTreeCost(); | ||||
| 6310 | |||||
| 6311 | LLVM_DEBUG(dbgs() << "SLP: Found cost = " << Cost << " for VF =" << VF << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Found cost = " << Cost << " for VF =" << VF << "\n"; } } while (false ); | ||||
| 6312 | if (Cost < -SLPCostThreshold) { | ||||
| 6313 | LLVM_DEBUG(dbgs() << "SLP: Decided to vectorize cost = " << Cost << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Decided to vectorize cost = " << Cost << "\n"; } } while (false); | ||||
| 6314 | |||||
| 6315 | using namespace ore; | ||||
| 6316 | |||||
| 6317 | R.getORE()->emit(OptimizationRemark(SV_NAME"slp-vectorizer", "StoresVectorized", | ||||
| 6318 | cast<StoreInst>(Chain[0])) | ||||
| 6319 | << "Stores SLP vectorized with cost " << NV("Cost", Cost) | ||||
| 6320 | << " and with tree size " | ||||
| 6321 | << NV("TreeSize", R.getTreeSize())); | ||||
| 6322 | |||||
| 6323 | R.vectorizeTree(); | ||||
| 6324 | return true; | ||||
| 6325 | } | ||||
| 6326 | |||||
| 6327 | return false; | ||||
| 6328 | } | ||||
| 6329 | |||||
| 6330 | bool SLPVectorizerPass::vectorizeStores(ArrayRef<StoreInst *> Stores, | ||||
| 6331 | BoUpSLP &R) { | ||||
| 6332 | // We may run into multiple chains that merge into a single chain. We mark the | ||||
| 6333 | // stores that we vectorized so that we don't visit the same store twice. | ||||
| 6334 | BoUpSLP::ValueSet VectorizedStores; | ||||
| 6335 | bool Changed = false; | ||||
| 6336 | |||||
| 6337 | int E = Stores.size(); | ||||
| 6338 | SmallBitVector Tails(E, false); | ||||
| 6339 | int MaxIter = MaxStoreLookup.getValue(); | ||||
| 6340 | SmallVector<std::pair<int, int>, 16> ConsecutiveChain( | ||||
| 6341 | E, std::make_pair(E, INT_MAX2147483647)); | ||||
| 6342 | SmallVector<SmallBitVector, 4> CheckedPairs(E, SmallBitVector(E, false)); | ||||
| 6343 | int IterCnt; | ||||
| 6344 | auto &&FindConsecutiveAccess = [this, &Stores, &Tails, &IterCnt, MaxIter, | ||||
| 6345 | &CheckedPairs, | ||||
| 6346 | &ConsecutiveChain](int K, int Idx) { | ||||
| 6347 | if (IterCnt >= MaxIter) | ||||
| 6348 | return true; | ||||
| 6349 | if (CheckedPairs[Idx].test(K)) | ||||
| 6350 | return ConsecutiveChain[K].second == 1 && | ||||
| 6351 | ConsecutiveChain[K].first == Idx; | ||||
| 6352 | ++IterCnt; | ||||
| 6353 | CheckedPairs[Idx].set(K); | ||||
| 6354 | CheckedPairs[K].set(Idx); | ||||
| 6355 | Optional<int> Diff = getPointersDiff(Stores[K]->getPointerOperand(), | ||||
| 6356 | Stores[Idx]->getPointerOperand(), *DL, | ||||
| 6357 | *SE, /*StrictCheck=*/true); | ||||
| 6358 | if (!Diff || *Diff == 0) | ||||
| 6359 | return false; | ||||
| 6360 | int Val = *Diff; | ||||
| 6361 | if (Val < 0) { | ||||
| 6362 | if (ConsecutiveChain[Idx].second > -Val) { | ||||
| 6363 | Tails.set(K); | ||||
| 6364 | ConsecutiveChain[Idx] = std::make_pair(K, -Val); | ||||
| 6365 | } | ||||
| 6366 | return false; | ||||
| 6367 | } | ||||
| 6368 | if (ConsecutiveChain[K].second <= Val) | ||||
| 6369 | return false; | ||||
| 6370 | |||||
| 6371 | Tails.set(Idx); | ||||
| 6372 | ConsecutiveChain[K] = std::make_pair(Idx, Val); | ||||
| 6373 | return Val == 1; | ||||
| 6374 | }; | ||||
| 6375 | // Do a quadratic search on all of the given stores in reverse order and find | ||||
| 6376 | // all of the pairs of stores that follow each other. | ||||
| 6377 | for (int Idx = E - 1; Idx >= 0; --Idx) { | ||||
| 6378 | // If a store has multiple consecutive store candidates, search according | ||||
| 6379 | // to the sequence: Idx-1, Idx+1, Idx-2, Idx+2, ... | ||||
| 6380 | // This is because usually pairing with immediate succeeding or preceding | ||||
| 6381 | // candidate create the best chance to find slp vectorization opportunity. | ||||
| 6382 | const int MaxLookDepth = std::max(E - Idx, Idx + 1); | ||||
| 6383 | IterCnt = 0; | ||||
| 6384 | for (int Offset = 1, F = MaxLookDepth; Offset < F; ++Offset) | ||||
| 6385 | if ((Idx >= Offset && FindConsecutiveAccess(Idx - Offset, Idx)) || | ||||
| 6386 | (Idx + Offset < E && FindConsecutiveAccess(Idx + Offset, Idx))) | ||||
| 6387 | break; | ||||
| 6388 | } | ||||
| 6389 | |||||
| 6390 | // Tracks if we tried to vectorize stores starting from the given tail | ||||
| 6391 | // already. | ||||
| 6392 | SmallBitVector TriedTails(E, false); | ||||
| 6393 | // For stores that start but don't end a link in the chain: | ||||
| 6394 | for (int Cnt = E; Cnt > 0; --Cnt) { | ||||
| 6395 | int I = Cnt - 1; | ||||
| 6396 | if (ConsecutiveChain[I].first == E || Tails.test(I)) | ||||
| 6397 | continue; | ||||
| 6398 | // We found a store instr that starts a chain. Now follow the chain and try | ||||
| 6399 | // to vectorize it. | ||||
| 6400 | BoUpSLP::ValueList Operands; | ||||
| 6401 | // Collect the chain into a list. | ||||
| 6402 | while (I != E && !VectorizedStores.count(Stores[I])) { | ||||
| 6403 | Operands.push_back(Stores[I]); | ||||
| 6404 | Tails.set(I); | ||||
| 6405 | if (ConsecutiveChain[I].second != 1) { | ||||
| 6406 | // Mark the new end in the chain and go back, if required. It might be | ||||
| 6407 | // required if the original stores come in reversed order, for example. | ||||
| 6408 | if (ConsecutiveChain[I].first != E && | ||||
| 6409 | Tails.test(ConsecutiveChain[I].first) && !TriedTails.test(I) && | ||||
| 6410 | !VectorizedStores.count(Stores[ConsecutiveChain[I].first])) { | ||||
| 6411 | TriedTails.set(I); | ||||
| 6412 | Tails.reset(ConsecutiveChain[I].first); | ||||
| 6413 | if (Cnt < ConsecutiveChain[I].first + 2) | ||||
| 6414 | Cnt = ConsecutiveChain[I].first + 2; | ||||
| 6415 | } | ||||
| 6416 | break; | ||||
| 6417 | } | ||||
| 6418 | // Move to the next value in the chain. | ||||
| 6419 | I = ConsecutiveChain[I].first; | ||||
| 6420 | } | ||||
| 6421 | assert(!Operands.empty() && "Expected non-empty list of stores.")(static_cast <bool> (!Operands.empty() && "Expected non-empty list of stores." ) ? void (0) : __assert_fail ("!Operands.empty() && \"Expected non-empty list of stores.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6421, __extension__ __PRETTY_FUNCTION__)); | ||||
| 6422 | |||||
| 6423 | unsigned MaxVecRegSize = R.getMaxVecRegSize(); | ||||
| 6424 | unsigned EltSize = R.getVectorElementSize(Operands[0]); | ||||
| 6425 | unsigned MaxElts = llvm::PowerOf2Floor(MaxVecRegSize / EltSize); | ||||
| 6426 | |||||
| 6427 | unsigned MinVF = std::max(2U, R.getMinVecRegSize() / EltSize); | ||||
| 6428 | unsigned MaxVF = std::min(R.getMaximumVF(EltSize, Instruction::Store), | ||||
| 6429 | MaxElts); | ||||
| 6430 | |||||
| 6431 | // FIXME: Is division-by-2 the correct step? Should we assert that the | ||||
| 6432 | // register size is a power-of-2? | ||||
| 6433 | unsigned StartIdx = 0; | ||||
| 6434 | for (unsigned Size = MaxVF; Size >= MinVF; Size /= 2) { | ||||
| 6435 | for (unsigned Cnt = StartIdx, E = Operands.size(); Cnt + Size <= E;) { | ||||
| 6436 | ArrayRef<Value *> Slice = makeArrayRef(Operands).slice(Cnt, Size); | ||||
| 6437 | if (!VectorizedStores.count(Slice.front()) && | ||||
| 6438 | !VectorizedStores.count(Slice.back()) && | ||||
| 6439 | vectorizeStoreChain(Slice, R, Cnt)) { | ||||
| 6440 | // Mark the vectorized stores so that we don't vectorize them again. | ||||
| 6441 | VectorizedStores.insert(Slice.begin(), Slice.end()); | ||||
| 6442 | Changed = true; | ||||
| 6443 | // If we vectorized initial block, no need to try to vectorize it | ||||
| 6444 | // again. | ||||
| 6445 | if (Cnt == StartIdx) | ||||
| 6446 | StartIdx += Size; | ||||
| 6447 | Cnt += Size; | ||||
| 6448 | continue; | ||||
| 6449 | } | ||||
| 6450 | ++Cnt; | ||||
| 6451 | } | ||||
| 6452 | // Check if the whole array was vectorized already - exit. | ||||
| 6453 | if (StartIdx >= Operands.size()) | ||||
| 6454 | break; | ||||
| 6455 | } | ||||
| 6456 | } | ||||
| 6457 | |||||
| 6458 | return Changed; | ||||
| 6459 | } | ||||
| 6460 | |||||
| 6461 | void SLPVectorizerPass::collectSeedInstructions(BasicBlock *BB) { | ||||
| 6462 | // Initialize the collections. We will make a single pass over the block. | ||||
| 6463 | Stores.clear(); | ||||
| 6464 | GEPs.clear(); | ||||
| 6465 | |||||
| 6466 | // Visit the store and getelementptr instructions in BB and organize them in | ||||
| 6467 | // Stores and GEPs according to the underlying objects of their pointer | ||||
| 6468 | // operands. | ||||
| 6469 | for (Instruction &I : *BB) { | ||||
| 6470 | // Ignore store instructions that are volatile or have a pointer operand | ||||
| 6471 | // that doesn't point to a scalar type. | ||||
| 6472 | if (auto *SI = dyn_cast<StoreInst>(&I)) { | ||||
| 6473 | if (!SI->isSimple()) | ||||
| 6474 | continue; | ||||
| 6475 | if (!isValidElementType(SI->getValueOperand()->getType())) | ||||
| 6476 | continue; | ||||
| 6477 | Stores[getUnderlyingObject(SI->getPointerOperand())].push_back(SI); | ||||
| 6478 | } | ||||
| 6479 | |||||
| 6480 | // Ignore getelementptr instructions that have more than one index, a | ||||
| 6481 | // constant index, or a pointer operand that doesn't point to a scalar | ||||
| 6482 | // type. | ||||
| 6483 | else if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) { | ||||
| 6484 | auto Idx = GEP->idx_begin()->get(); | ||||
| 6485 | if (GEP->getNumIndices() > 1 || isa<Constant>(Idx)) | ||||
| 6486 | continue; | ||||
| 6487 | if (!isValidElementType(Idx->getType())) | ||||
| 6488 | continue; | ||||
| 6489 | if (GEP->getType()->isVectorTy()) | ||||
| 6490 | continue; | ||||
| 6491 | GEPs[GEP->getPointerOperand()].push_back(GEP); | ||||
| 6492 | } | ||||
| 6493 | } | ||||
| 6494 | } | ||||
| 6495 | |||||
| 6496 | bool SLPVectorizerPass::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) { | ||||
| 6497 | if (!A || !B) | ||||
| 6498 | return false; | ||||
| 6499 | Value *VL[] = {A, B}; | ||||
| 6500 | return tryToVectorizeList(VL, R, /*AllowReorder=*/true); | ||||
| 6501 | } | ||||
| 6502 | |||||
| 6503 | bool SLPVectorizerPass::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R, | ||||
| 6504 | bool AllowReorder, | ||||
| 6505 | ArrayRef<Value *> InsertUses) { | ||||
| 6506 | if (VL.size() < 2) | ||||
| 6507 | return false; | ||||
| 6508 | |||||
| 6509 | LLVM_DEBUG(dbgs() << "SLP: Trying to vectorize a list of length = "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize a list of length = " << VL.size() << ".\n"; } } while (false) | ||||
| 6510 | << VL.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize a list of length = " << VL.size() << ".\n"; } } while (false); | ||||
| 6511 | |||||
| 6512 | // Check that all of the parts are instructions of the same type, | ||||
| 6513 | // we permit an alternate opcode via InstructionsState. | ||||
| 6514 | InstructionsState S = getSameOpcode(VL); | ||||
| 6515 | if (!S.getOpcode()) | ||||
| 6516 | return false; | ||||
| 6517 | |||||
| 6518 | Instruction *I0 = cast<Instruction>(S.OpValue); | ||||
| 6519 | // Make sure invalid types (including vector type) are rejected before | ||||
| 6520 | // determining vectorization factor for scalar instructions. | ||||
| 6521 | for (Value *V : VL) { | ||||
| 6522 | Type *Ty = V->getType(); | ||||
| 6523 | if (!isValidElementType(Ty)) { | ||||
| 6524 | // NOTE: the following will give user internal llvm type name, which may | ||||
| 6525 | // not be useful. | ||||
| 6526 | R.getORE()->emit([&]() { | ||||
| 6527 | std::string type_str; | ||||
| 6528 | llvm::raw_string_ostream rso(type_str); | ||||
| 6529 | Ty->print(rso); | ||||
| 6530 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "UnsupportedType", I0) | ||||
| 6531 | << "Cannot SLP vectorize list: type " | ||||
| 6532 | << rso.str() + " is unsupported by vectorizer"; | ||||
| 6533 | }); | ||||
| 6534 | return false; | ||||
| 6535 | } | ||||
| 6536 | } | ||||
| 6537 | |||||
| 6538 | unsigned Sz = R.getVectorElementSize(I0); | ||||
| 6539 | unsigned MinVF = std::max(2U, R.getMinVecRegSize() / Sz); | ||||
| 6540 | unsigned MaxVF = std::max<unsigned>(PowerOf2Floor(VL.size()), MinVF); | ||||
| 6541 | MaxVF = std::min(R.getMaximumVF(Sz, S.getOpcode()), MaxVF); | ||||
| 6542 | if (MaxVF < 2) { | ||||
| 6543 | R.getORE()->emit([&]() { | ||||
| 6544 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "SmallVF", I0) | ||||
| 6545 | << "Cannot SLP vectorize list: vectorization factor " | ||||
| 6546 | << "less than 2 is not supported"; | ||||
| 6547 | }); | ||||
| 6548 | return false; | ||||
| 6549 | } | ||||
| 6550 | |||||
| 6551 | bool Changed = false; | ||||
| 6552 | bool CandidateFound = false; | ||||
| 6553 | InstructionCost MinCost = SLPCostThreshold.getValue(); | ||||
| 6554 | |||||
| 6555 | bool CompensateUseCost = | ||||
| 6556 | !InsertUses.empty() && llvm::all_of(InsertUses, [](const Value *V) { | ||||
| 6557 | return V && isa<InsertElementInst>(V); | ||||
| 6558 | }); | ||||
| 6559 | assert((!CompensateUseCost || InsertUses.size() == VL.size()) &&(static_cast <bool> ((!CompensateUseCost || InsertUses. size() == VL.size()) && "Each scalar expected to have an associated InsertElement user." ) ? void (0) : __assert_fail ("(!CompensateUseCost || InsertUses.size() == VL.size()) && \"Each scalar expected to have an associated InsertElement user.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6560, __extension__ __PRETTY_FUNCTION__)) | ||||
| 6560 | "Each scalar expected to have an associated InsertElement user.")(static_cast <bool> ((!CompensateUseCost || InsertUses. size() == VL.size()) && "Each scalar expected to have an associated InsertElement user." ) ? void (0) : __assert_fail ("(!CompensateUseCost || InsertUses.size() == VL.size()) && \"Each scalar expected to have an associated InsertElement user.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6560, __extension__ __PRETTY_FUNCTION__)); | ||||
| 6561 | |||||
| 6562 | unsigned NextInst = 0, MaxInst = VL.size(); | ||||
| 6563 | for (unsigned VF = MaxVF; NextInst + 1 < MaxInst && VF >= MinVF; VF /= 2) { | ||||
| 6564 | // No actual vectorization should happen, if number of parts is the same as | ||||
| 6565 | // provided vectorization factor (i.e. the scalar type is used for vector | ||||
| 6566 | // code during codegen). | ||||
| 6567 | auto *VecTy = FixedVectorType::get(VL[0]->getType(), VF); | ||||
| 6568 | if (TTI->getNumberOfParts(VecTy) == VF) | ||||
| 6569 | continue; | ||||
| 6570 | for (unsigned I = NextInst; I < MaxInst; ++I) { | ||||
| 6571 | unsigned OpsWidth = 0; | ||||
| 6572 | |||||
| 6573 | if (I + VF > MaxInst) | ||||
| 6574 | OpsWidth = MaxInst - I; | ||||
| 6575 | else | ||||
| 6576 | OpsWidth = VF; | ||||
| 6577 | |||||
| 6578 | if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2) | ||||
| 6579 | break; | ||||
| 6580 | |||||
| 6581 | ArrayRef<Value *> Ops = VL.slice(I, OpsWidth); | ||||
| 6582 | // Check that a previous iteration of this loop did not delete the Value. | ||||
| 6583 | if (llvm::any_of(Ops, [&R](Value *V) { | ||||
| 6584 | auto *I = dyn_cast<Instruction>(V); | ||||
| 6585 | return I && R.isDeleted(I); | ||||
| 6586 | })) | ||||
| 6587 | continue; | ||||
| 6588 | |||||
| 6589 | LLVM_DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"; } } while (false) | ||||
| 6590 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing " << OpsWidth << " operations " << "\n"; } } while (false); | ||||
| 6591 | |||||
| 6592 | R.buildTree(Ops); | ||||
| 6593 | Optional<ArrayRef<unsigned>> Order = R.bestOrder(); | ||||
| 6594 | // TODO: check if we can allow reordering for more cases. | ||||
| 6595 | if (AllowReorder && Order) { | ||||
| 6596 | // TODO: reorder tree nodes without tree rebuilding. | ||||
| 6597 | // Conceptually, there is nothing actually preventing us from trying to | ||||
| 6598 | // reorder a larger list. In fact, we do exactly this when vectorizing | ||||
| 6599 | // reductions. However, at this point, we only expect to get here when | ||||
| 6600 | // there are exactly two operations. | ||||
| 6601 | assert(Ops.size() == 2)(static_cast <bool> (Ops.size() == 2) ? void (0) : __assert_fail ("Ops.size() == 2", "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6601, __extension__ __PRETTY_FUNCTION__)); | ||||
| 6602 | Value *ReorderedOps[] = {Ops[1], Ops[0]}; | ||||
| 6603 | R.buildTree(ReorderedOps, None); | ||||
| 6604 | } | ||||
| 6605 | if (R.isTreeTinyAndNotFullyVectorizable()) | ||||
| 6606 | continue; | ||||
| 6607 | |||||
| 6608 | R.computeMinimumValueSizes(); | ||||
| 6609 | InstructionCost Cost = R.getTreeCost(); | ||||
| 6610 | CandidateFound = true; | ||||
| 6611 | if (CompensateUseCost) { | ||||
| 6612 | // TODO: Use TTI's getScalarizationOverhead for sequence of inserts | ||||
| 6613 | // rather than sum of single inserts as the latter may overestimate | ||||
| 6614 | // cost. This work should imply improving cost estimation for extracts | ||||
| 6615 | // that added in for external (for vectorization tree) users,i.e. that | ||||
| 6616 | // part should also switch to same interface. | ||||
| 6617 | // For example, the following case is projected code after SLP: | ||||
| 6618 | // %4 = extractelement <4 x i64> %3, i32 0 | ||||
| 6619 | // %v0 = insertelement <4 x i64> poison, i64 %4, i32 0 | ||||
| 6620 | // %5 = extractelement <4 x i64> %3, i32 1 | ||||
| 6621 | // %v1 = insertelement <4 x i64> %v0, i64 %5, i32 1 | ||||
| 6622 | // %6 = extractelement <4 x i64> %3, i32 2 | ||||
| 6623 | // %v2 = insertelement <4 x i64> %v1, i64 %6, i32 2 | ||||
| 6624 | // %7 = extractelement <4 x i64> %3, i32 3 | ||||
| 6625 | // %v3 = insertelement <4 x i64> %v2, i64 %7, i32 3 | ||||
| 6626 | // | ||||
| 6627 | // Extracts here added by SLP in order to feed users (the inserts) of | ||||
| 6628 | // original scalars and contribute to "ExtractCost" at cost evaluation. | ||||
| 6629 | // The inserts in turn form sequence to build an aggregate that | ||||
| 6630 | // detected by findBuildAggregate routine. | ||||
| 6631 | // SLP makes an assumption that such sequence will be optimized away | ||||
| 6632 | // later (instcombine) so it tries to compensate ExctractCost with | ||||
| 6633 | // cost of insert sequence. | ||||
| 6634 | // Current per element cost calculation approach is not quite accurate | ||||
| 6635 | // and tends to create bias toward favoring vectorization. | ||||
| 6636 | // Switching to the TTI interface might help a bit. | ||||
| 6637 | // Alternative solution could be pattern-match to detect a no-op or | ||||
| 6638 | // shuffle. | ||||
| 6639 | InstructionCost UserCost = 0; | ||||
| 6640 | for (unsigned Lane = 0; Lane < OpsWidth; Lane++) { | ||||
| 6641 | auto *IE = cast<InsertElementInst>(InsertUses[I + Lane]); | ||||
| 6642 | if (auto *CI = dyn_cast<ConstantInt>(IE->getOperand(2))) | ||||
| 6643 | UserCost += TTI->getVectorInstrCost( | ||||
| 6644 | Instruction::InsertElement, IE->getType(), CI->getZExtValue()); | ||||
| 6645 | } | ||||
| 6646 | LLVM_DEBUG(dbgs() << "SLP: Compensate cost of users by: " << UserCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Compensate cost of users by: " << UserCost << ".\n"; } } while (false) | ||||
| 6647 | << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Compensate cost of users by: " << UserCost << ".\n"; } } while (false); | ||||
| 6648 | Cost -= UserCost; | ||||
| 6649 | } | ||||
| 6650 | |||||
| 6651 | MinCost = std::min(MinCost, Cost); | ||||
| 6652 | |||||
| 6653 | if (Cost < -SLPCostThreshold) { | ||||
| 6654 | LLVM_DEBUG(dbgs() << "SLP: Vectorizing list at cost:" << Cost << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorizing list at cost:" << Cost << ".\n"; } } while (false); | ||||
| 6655 | R.getORE()->emit(OptimizationRemark(SV_NAME"slp-vectorizer", "VectorizedList", | ||||
| 6656 | cast<Instruction>(Ops[0])) | ||||
| 6657 | << "SLP vectorized with cost " << ore::NV("Cost", Cost) | ||||
| 6658 | << " and with tree size " | ||||
| 6659 | << ore::NV("TreeSize", R.getTreeSize())); | ||||
| 6660 | |||||
| 6661 | R.vectorizeTree(); | ||||
| 6662 | // Move to the next bundle. | ||||
| 6663 | I += VF - 1; | ||||
| 6664 | NextInst = I + 1; | ||||
| 6665 | Changed = true; | ||||
| 6666 | } | ||||
| 6667 | } | ||||
| 6668 | } | ||||
| 6669 | |||||
| 6670 | if (!Changed && CandidateFound) { | ||||
| 6671 | R.getORE()->emit([&]() { | ||||
| 6672 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "NotBeneficial", I0) | ||||
| 6673 | << "List vectorization was possible but not beneficial with cost " | ||||
| 6674 | << ore::NV("Cost", MinCost) << " >= " | ||||
| 6675 | << ore::NV("Treshold", -SLPCostThreshold); | ||||
| 6676 | }); | ||||
| 6677 | } else if (!Changed) { | ||||
| 6678 | R.getORE()->emit([&]() { | ||||
| 6679 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "NotPossible", I0) | ||||
| 6680 | << "Cannot SLP vectorize list: vectorization was impossible" | ||||
| 6681 | << " with available vectorization factors"; | ||||
| 6682 | }); | ||||
| 6683 | } | ||||
| 6684 | return Changed; | ||||
| 6685 | } | ||||
| 6686 | |||||
| 6687 | bool SLPVectorizerPass::tryToVectorize(Instruction *I, BoUpSLP &R) { | ||||
| 6688 | if (!I) | ||||
| 6689 | return false; | ||||
| 6690 | |||||
| 6691 | if (!isa<BinaryOperator>(I) && !isa<CmpInst>(I)) | ||||
| 6692 | return false; | ||||
| 6693 | |||||
| 6694 | Value *P = I->getParent(); | ||||
| 6695 | |||||
| 6696 | // Vectorize in current basic block only. | ||||
| 6697 | auto *Op0 = dyn_cast<Instruction>(I->getOperand(0)); | ||||
| 6698 | auto *Op1 = dyn_cast<Instruction>(I->getOperand(1)); | ||||
| 6699 | if (!Op0 || !Op1 || Op0->getParent() != P || Op1->getParent() != P) | ||||
| 6700 | return false; | ||||
| 6701 | |||||
| 6702 | // Try to vectorize V. | ||||
| 6703 | if (tryToVectorizePair(Op0, Op1, R)) | ||||
| 6704 | return true; | ||||
| 6705 | |||||
| 6706 | auto *A = dyn_cast<BinaryOperator>(Op0); | ||||
| 6707 | auto *B = dyn_cast<BinaryOperator>(Op1); | ||||
| 6708 | // Try to skip B. | ||||
| 6709 | if (B && B->hasOneUse()) { | ||||
| 6710 | auto *B0 = dyn_cast<BinaryOperator>(B->getOperand(0)); | ||||
| 6711 | auto *B1 = dyn_cast<BinaryOperator>(B->getOperand(1)); | ||||
| 6712 | if (B0 && B0->getParent() == P && tryToVectorizePair(A, B0, R)) | ||||
| 6713 | return true; | ||||
| 6714 | if (B1 && B1->getParent() == P && tryToVectorizePair(A, B1, R)) | ||||
| 6715 | return true; | ||||
| 6716 | } | ||||
| 6717 | |||||
| 6718 | // Try to skip A. | ||||
| 6719 | if (A && A->hasOneUse()) { | ||||
| 6720 | auto *A0 = dyn_cast<BinaryOperator>(A->getOperand(0)); | ||||
| 6721 | auto *A1 = dyn_cast<BinaryOperator>(A->getOperand(1)); | ||||
| 6722 | if (A0 && A0->getParent() == P && tryToVectorizePair(A0, B, R)) | ||||
| 6723 | return true; | ||||
| 6724 | if (A1 && A1->getParent() == P && tryToVectorizePair(A1, B, R)) | ||||
| 6725 | return true; | ||||
| 6726 | } | ||||
| 6727 | return false; | ||||
| 6728 | } | ||||
| 6729 | |||||
| 6730 | namespace { | ||||
| 6731 | |||||
| 6732 | /// Model horizontal reductions. | ||||
| 6733 | /// | ||||
| 6734 | /// A horizontal reduction is a tree of reduction instructions that has values | ||||
| 6735 | /// that can be put into a vector as its leaves. For example: | ||||
| 6736 | /// | ||||
| 6737 | /// mul mul mul mul | ||||
| 6738 | /// \ / \ / | ||||
| 6739 | /// + + | ||||
| 6740 | /// \ / | ||||
| 6741 | /// + | ||||
| 6742 | /// This tree has "mul" as its leaf values and "+" as its reduction | ||||
| 6743 | /// instructions. A reduction can feed into a store or a binary operation | ||||
| 6744 | /// feeding a phi. | ||||
| 6745 | /// ... | ||||
| 6746 | /// \ / | ||||
| 6747 | /// + | ||||
| 6748 | /// | | ||||
| 6749 | /// phi += | ||||
| 6750 | /// | ||||
| 6751 | /// Or: | ||||
| 6752 | /// ... | ||||
| 6753 | /// \ / | ||||
| 6754 | /// + | ||||
| 6755 | /// | | ||||
| 6756 | /// *p = | ||||
| 6757 | /// | ||||
| 6758 | class HorizontalReduction { | ||||
| 6759 | using ReductionOpsType = SmallVector<Value *, 16>; | ||||
| 6760 | using ReductionOpsListType = SmallVector<ReductionOpsType, 2>; | ||||
| 6761 | ReductionOpsListType ReductionOps; | ||||
| 6762 | SmallVector<Value *, 32> ReducedVals; | ||||
| 6763 | // Use map vector to make stable output. | ||||
| 6764 | MapVector<Instruction *, Value *> ExtraArgs; | ||||
| 6765 | WeakTrackingVH ReductionRoot; | ||||
| 6766 | /// The type of reduction operation. | ||||
| 6767 | RecurKind RdxKind; | ||||
| 6768 | |||||
| 6769 | /// Checks if instruction is associative and can be vectorized. | ||||
| 6770 | static bool isVectorizable(RecurKind Kind, Instruction *I) { | ||||
| 6771 | if (Kind == RecurKind::None) | ||||
| 6772 | return false; | ||||
| 6773 | if (RecurrenceDescriptor::isIntMinMaxRecurrenceKind(Kind)) | ||||
| 6774 | return true; | ||||
| 6775 | |||||
| 6776 | if (Kind == RecurKind::FMax || Kind == RecurKind::FMin) { | ||||
| 6777 | // FP min/max are associative except for NaN and -0.0. We do not | ||||
| 6778 | // have to rule out -0.0 here because the intrinsic semantics do not | ||||
| 6779 | // specify a fixed result for it. | ||||
| 6780 | return I->getFastMathFlags().noNaNs(); | ||||
| 6781 | } | ||||
| 6782 | |||||
| 6783 | return I->isAssociative(); | ||||
| 6784 | } | ||||
| 6785 | |||||
| 6786 | /// Checks if the ParentStackElem.first should be marked as a reduction | ||||
| 6787 | /// operation with an extra argument or as extra argument itself. | ||||
| 6788 | void markExtraArg(std::pair<Instruction *, unsigned> &ParentStackElem, | ||||
| 6789 | Value *ExtraArg) { | ||||
| 6790 | if (ExtraArgs.count(ParentStackElem.first)) { | ||||
| 6791 | ExtraArgs[ParentStackElem.first] = nullptr; | ||||
| 6792 | // We ran into something like: | ||||
| 6793 | // ParentStackElem.first = ExtraArgs[ParentStackElem.first] + ExtraArg. | ||||
| 6794 | // The whole ParentStackElem.first should be considered as an extra value | ||||
| 6795 | // in this case. | ||||
| 6796 | // Do not perform analysis of remaining operands of ParentStackElem.first | ||||
| 6797 | // instruction, this whole instruction is an extra argument. | ||||
| 6798 | ParentStackElem.second = getNumberOfOperands(ParentStackElem.first); | ||||
| 6799 | } else { | ||||
| 6800 | // We ran into something like: | ||||
| 6801 | // ParentStackElem.first += ... + ExtraArg + ... | ||||
| 6802 | ExtraArgs[ParentStackElem.first] = ExtraArg; | ||||
| 6803 | } | ||||
| 6804 | } | ||||
| 6805 | |||||
| 6806 | /// Creates reduction operation with the current opcode. | ||||
| 6807 | static Value *createOp(IRBuilder<> &Builder, RecurKind Kind, Value *LHS, | ||||
| 6808 | Value *RHS, const Twine &Name, bool UseSelect) { | ||||
| 6809 | unsigned RdxOpcode = RecurrenceDescriptor::getOpcode(Kind); | ||||
| 6810 | switch (Kind) { | ||||
| 6811 | case RecurKind::Add: | ||||
| 6812 | case RecurKind::Mul: | ||||
| 6813 | case RecurKind::Or: | ||||
| 6814 | case RecurKind::And: | ||||
| 6815 | case RecurKind::Xor: | ||||
| 6816 | case RecurKind::FAdd: | ||||
| 6817 | case RecurKind::FMul: | ||||
| 6818 | return Builder.CreateBinOp((Instruction::BinaryOps)RdxOpcode, LHS, RHS, | ||||
| 6819 | Name); | ||||
| 6820 | case RecurKind::FMax: | ||||
| 6821 | return Builder.CreateBinaryIntrinsic(Intrinsic::maxnum, LHS, RHS); | ||||
| 6822 | case RecurKind::FMin: | ||||
| 6823 | return Builder.CreateBinaryIntrinsic(Intrinsic::minnum, LHS, RHS); | ||||
| 6824 | case RecurKind::SMax: | ||||
| 6825 | if (UseSelect) { | ||||
| 6826 | Value *Cmp = Builder.CreateICmpSGT(LHS, RHS, Name); | ||||
| 6827 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | ||||
| 6828 | } | ||||
| 6829 | return Builder.CreateBinaryIntrinsic(Intrinsic::smax, LHS, RHS); | ||||
| 6830 | case RecurKind::SMin: | ||||
| 6831 | if (UseSelect) { | ||||
| 6832 | Value *Cmp = Builder.CreateICmpSLT(LHS, RHS, Name); | ||||
| 6833 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | ||||
| 6834 | } | ||||
| 6835 | return Builder.CreateBinaryIntrinsic(Intrinsic::smin, LHS, RHS); | ||||
| 6836 | case RecurKind::UMax: | ||||
| 6837 | if (UseSelect) { | ||||
| 6838 | Value *Cmp = Builder.CreateICmpUGT(LHS, RHS, Name); | ||||
| 6839 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | ||||
| 6840 | } | ||||
| 6841 | return Builder.CreateBinaryIntrinsic(Intrinsic::umax, LHS, RHS); | ||||
| 6842 | case RecurKind::UMin: | ||||
| 6843 | if (UseSelect) { | ||||
| 6844 | Value *Cmp = Builder.CreateICmpULT(LHS, RHS, Name); | ||||
| 6845 | return Builder.CreateSelect(Cmp, LHS, RHS, Name); | ||||
| 6846 | } | ||||
| 6847 | return Builder.CreateBinaryIntrinsic(Intrinsic::umin, LHS, RHS); | ||||
| 6848 | default: | ||||
| 6849 | llvm_unreachable("Unknown reduction operation.")::llvm::llvm_unreachable_internal("Unknown reduction operation." , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6849); | ||||
| 6850 | } | ||||
| 6851 | } | ||||
| 6852 | |||||
| 6853 | /// Creates reduction operation with the current opcode with the IR flags | ||||
| 6854 | /// from \p ReductionOps. | ||||
| 6855 | static Value *createOp(IRBuilder<> &Builder, RecurKind RdxKind, Value *LHS, | ||||
| 6856 | Value *RHS, const Twine &Name, | ||||
| 6857 | const ReductionOpsListType &ReductionOps) { | ||||
| 6858 | bool UseSelect = ReductionOps.size() == 2; | ||||
| 6859 | assert((!UseSelect || isa<SelectInst>(ReductionOps[1][0])) &&(static_cast <bool> ((!UseSelect || isa<SelectInst> (ReductionOps[1][0])) && "Expected cmp + select pairs for reduction" ) ? void (0) : __assert_fail ("(!UseSelect || isa<SelectInst>(ReductionOps[1][0])) && \"Expected cmp + select pairs for reduction\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6860, __extension__ __PRETTY_FUNCTION__)) | ||||
| 6860 | "Expected cmp + select pairs for reduction")(static_cast <bool> ((!UseSelect || isa<SelectInst> (ReductionOps[1][0])) && "Expected cmp + select pairs for reduction" ) ? void (0) : __assert_fail ("(!UseSelect || isa<SelectInst>(ReductionOps[1][0])) && \"Expected cmp + select pairs for reduction\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6860, __extension__ __PRETTY_FUNCTION__)); | ||||
| 6861 | Value *Op = createOp(Builder, RdxKind, LHS, RHS, Name, UseSelect); | ||||
| 6862 | if (RecurrenceDescriptor::isIntMinMaxRecurrenceKind(RdxKind)) { | ||||
| 6863 | if (auto *Sel = dyn_cast<SelectInst>(Op)) { | ||||
| 6864 | propagateIRFlags(Sel->getCondition(), ReductionOps[0]); | ||||
| 6865 | propagateIRFlags(Op, ReductionOps[1]); | ||||
| 6866 | return Op; | ||||
| 6867 | } | ||||
| 6868 | } | ||||
| 6869 | propagateIRFlags(Op, ReductionOps[0]); | ||||
| 6870 | return Op; | ||||
| 6871 | } | ||||
| 6872 | |||||
| 6873 | /// Creates reduction operation with the current opcode with the IR flags | ||||
| 6874 | /// from \p I. | ||||
| 6875 | static Value *createOp(IRBuilder<> &Builder, RecurKind RdxKind, Value *LHS, | ||||
| 6876 | Value *RHS, const Twine &Name, Instruction *I) { | ||||
| 6877 | auto *SelI = dyn_cast<SelectInst>(I); | ||||
| 6878 | Value *Op = createOp(Builder, RdxKind, LHS, RHS, Name, SelI != nullptr); | ||||
| 6879 | if (SelI && RecurrenceDescriptor::isIntMinMaxRecurrenceKind(RdxKind)) { | ||||
| 6880 | if (auto *Sel = dyn_cast<SelectInst>(Op)) | ||||
| 6881 | propagateIRFlags(Sel->getCondition(), SelI->getCondition()); | ||||
| 6882 | } | ||||
| 6883 | propagateIRFlags(Op, I); | ||||
| 6884 | return Op; | ||||
| 6885 | } | ||||
| 6886 | |||||
| 6887 | static RecurKind getRdxKind(Instruction *I) { | ||||
| 6888 | assert(I && "Expected instruction for reduction matching")(static_cast <bool> (I && "Expected instruction for reduction matching" ) ? void (0) : __assert_fail ("I && \"Expected instruction for reduction matching\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 6888, __extension__ __PRETTY_FUNCTION__)); | ||||
| 6889 | TargetTransformInfo::ReductionFlags RdxFlags; | ||||
| 6890 | if (match(I, m_Add(m_Value(), m_Value()))) | ||||
| 6891 | return RecurKind::Add; | ||||
| 6892 | if (match(I, m_Mul(m_Value(), m_Value()))) | ||||
| 6893 | return RecurKind::Mul; | ||||
| 6894 | if (match(I, m_And(m_Value(), m_Value()))) | ||||
| 6895 | return RecurKind::And; | ||||
| 6896 | if (match(I, m_Or(m_Value(), m_Value()))) | ||||
| 6897 | return RecurKind::Or; | ||||
| 6898 | if (match(I, m_Xor(m_Value(), m_Value()))) | ||||
| 6899 | return RecurKind::Xor; | ||||
| 6900 | if (match(I, m_FAdd(m_Value(), m_Value()))) | ||||
| 6901 | return RecurKind::FAdd; | ||||
| 6902 | if (match(I, m_FMul(m_Value(), m_Value()))) | ||||
| 6903 | return RecurKind::FMul; | ||||
| 6904 | |||||
| 6905 | if (match(I, m_Intrinsic<Intrinsic::maxnum>(m_Value(), m_Value()))) | ||||
| 6906 | return RecurKind::FMax; | ||||
| 6907 | if (match(I, m_Intrinsic<Intrinsic::minnum>(m_Value(), m_Value()))) | ||||
| 6908 | return RecurKind::FMin; | ||||
| 6909 | |||||
| 6910 | // This matches either cmp+select or intrinsics. SLP is expected to handle | ||||
| 6911 | // either form. | ||||
| 6912 | // TODO: If we are canonicalizing to intrinsics, we can remove several | ||||
| 6913 | // special-case paths that deal with selects. | ||||
| 6914 | if (match(I, m_SMax(m_Value(), m_Value()))) | ||||
| 6915 | return RecurKind::SMax; | ||||
| 6916 | if (match(I, m_SMin(m_Value(), m_Value()))) | ||||
| 6917 | return RecurKind::SMin; | ||||
| 6918 | if (match(I, m_UMax(m_Value(), m_Value()))) | ||||
| 6919 | return RecurKind::UMax; | ||||
| 6920 | if (match(I, m_UMin(m_Value(), m_Value()))) | ||||
| 6921 | return RecurKind::UMin; | ||||
| 6922 | |||||
| 6923 | if (auto *Select = dyn_cast<SelectInst>(I)) { | ||||
| 6924 | // Try harder: look for min/max pattern based on instructions producing | ||||
| 6925 | // same values such as: select ((cmp Inst1, Inst2), Inst1, Inst2). | ||||
| 6926 | // During the intermediate stages of SLP, it's very common to have | ||||
| 6927 | // pattern like this (since optimizeGatherSequence is run only once | ||||
| 6928 | // at the end): | ||||
| 6929 | // %1 = extractelement <2 x i32> %a, i32 0 | ||||
| 6930 | // %2 = extractelement <2 x i32> %a, i32 1 | ||||
| 6931 | // %cond = icmp sgt i32 %1, %2 | ||||
| 6932 | // %3 = extractelement <2 x i32> %a, i32 0 | ||||
| 6933 | // %4 = extractelement <2 x i32> %a, i32 1 | ||||
| 6934 | // %select = select i1 %cond, i32 %3, i32 %4 | ||||
| 6935 | CmpInst::Predicate Pred; | ||||
| 6936 | Instruction *L1; | ||||
| 6937 | Instruction *L2; | ||||
| 6938 | |||||
| 6939 | Value *LHS = Select->getTrueValue(); | ||||
| 6940 | Value *RHS = Select->getFalseValue(); | ||||
| 6941 | Value *Cond = Select->getCondition(); | ||||
| 6942 | |||||
| 6943 | // TODO: Support inverse predicates. | ||||
| 6944 | if (match(Cond, m_Cmp(Pred, m_Specific(LHS), m_Instruction(L2)))) { | ||||
| 6945 | if (!isa<ExtractElementInst>(RHS) || | ||||
| 6946 | !L2->isIdenticalTo(cast<Instruction>(RHS))) | ||||
| 6947 | return RecurKind::None; | ||||
| 6948 | } else if (match(Cond, m_Cmp(Pred, m_Instruction(L1), m_Specific(RHS)))) { | ||||
| 6949 | if (!isa<ExtractElementInst>(LHS) || | ||||
| 6950 | !L1->isIdenticalTo(cast<Instruction>(LHS))) | ||||
| 6951 | return RecurKind::None; | ||||
| 6952 | } else { | ||||
| 6953 | if (!isa<ExtractElementInst>(LHS) || !isa<ExtractElementInst>(RHS)) | ||||
| 6954 | return RecurKind::None; | ||||
| 6955 | if (!match(Cond, m_Cmp(Pred, m_Instruction(L1), m_Instruction(L2))) || | ||||
| 6956 | !L1->isIdenticalTo(cast<Instruction>(LHS)) || | ||||
| 6957 | !L2->isIdenticalTo(cast<Instruction>(RHS))) | ||||
| 6958 | return RecurKind::None; | ||||
| 6959 | } | ||||
| 6960 | |||||
| 6961 | TargetTransformInfo::ReductionFlags RdxFlags; | ||||
| 6962 | switch (Pred) { | ||||
| 6963 | default: | ||||
| 6964 | return RecurKind::None; | ||||
| 6965 | case CmpInst::ICMP_SGT: | ||||
| 6966 | case CmpInst::ICMP_SGE: | ||||
| 6967 | return RecurKind::SMax; | ||||
| 6968 | case CmpInst::ICMP_SLT: | ||||
| 6969 | case CmpInst::ICMP_SLE: | ||||
| 6970 | return RecurKind::SMin; | ||||
| 6971 | case CmpInst::ICMP_UGT: | ||||
| 6972 | case CmpInst::ICMP_UGE: | ||||
| 6973 | return RecurKind::UMax; | ||||
| 6974 | case CmpInst::ICMP_ULT: | ||||
| 6975 | case CmpInst::ICMP_ULE: | ||||
| 6976 | return RecurKind::UMin; | ||||
| 6977 | } | ||||
| 6978 | } | ||||
| 6979 | return RecurKind::None; | ||||
| 6980 | } | ||||
| 6981 | |||||
| 6982 | /// Get the index of the first operand. | ||||
| 6983 | static unsigned getFirstOperandIndex(Instruction *I) { | ||||
| 6984 | return isa<SelectInst>(I) ? 1 : 0; | ||||
| 6985 | } | ||||
| 6986 | |||||
| 6987 | /// Total number of operands in the reduction operation. | ||||
| 6988 | static unsigned getNumberOfOperands(Instruction *I) { | ||||
| 6989 | return isa<SelectInst>(I) ? 3 : 2; | ||||
| 6990 | } | ||||
| 6991 | |||||
| 6992 | /// Checks if the instruction is in basic block \p BB. | ||||
| 6993 | /// For a min/max reduction check that both compare and select are in \p BB. | ||||
| 6994 | static bool hasSameParent(Instruction *I, BasicBlock *BB, bool IsRedOp) { | ||||
| 6995 | auto *Sel = dyn_cast<SelectInst>(I); | ||||
| 6996 | if (IsRedOp && Sel) { | ||||
| 6997 | auto *Cmp = cast<Instruction>(Sel->getCondition()); | ||||
| 6998 | return Sel->getParent() == BB && Cmp->getParent() == BB; | ||||
| 6999 | } | ||||
| 7000 | return I->getParent() == BB; | ||||
| 7001 | } | ||||
| 7002 | |||||
| 7003 | /// Expected number of uses for reduction operations/reduced values. | ||||
| 7004 | static bool hasRequiredNumberOfUses(bool MatchCmpSel, Instruction *I) { | ||||
| 7005 | // SelectInst must be used twice while the condition op must have single | ||||
| 7006 | // use only. | ||||
| 7007 | if (MatchCmpSel) { | ||||
| 7008 | if (auto *Sel = dyn_cast<SelectInst>(I)) | ||||
| 7009 | return Sel->hasNUses(2) && Sel->getCondition()->hasOneUse(); | ||||
| 7010 | return I->hasNUses(2); | ||||
| 7011 | } | ||||
| 7012 | |||||
| 7013 | // Arithmetic reduction operation must be used once only. | ||||
| 7014 | return I->hasOneUse(); | ||||
| 7015 | } | ||||
| 7016 | |||||
| 7017 | /// Initializes the list of reduction operations. | ||||
| 7018 | void initReductionOps(Instruction *I) { | ||||
| 7019 | if (isa<SelectInst>(I)) | ||||
| 7020 | ReductionOps.assign(2, ReductionOpsType()); | ||||
| 7021 | else | ||||
| 7022 | ReductionOps.assign(1, ReductionOpsType()); | ||||
| 7023 | } | ||||
| 7024 | |||||
| 7025 | /// Add all reduction operations for the reduction instruction \p I. | ||||
| 7026 | void addReductionOps(Instruction *I) { | ||||
| 7027 | if (auto *Sel = dyn_cast<SelectInst>(I)) { | ||||
| 7028 | ReductionOps[0].emplace_back(Sel->getCondition()); | ||||
| 7029 | ReductionOps[1].emplace_back(Sel); | ||||
| 7030 | } else { | ||||
| 7031 | ReductionOps[0].emplace_back(I); | ||||
| 7032 | } | ||||
| 7033 | } | ||||
| 7034 | |||||
| 7035 | static Value *getLHS(RecurKind Kind, Instruction *I) { | ||||
| 7036 | if (Kind == RecurKind::None) | ||||
| 7037 | return nullptr; | ||||
| 7038 | return I->getOperand(getFirstOperandIndex(I)); | ||||
| 7039 | } | ||||
| 7040 | static Value *getRHS(RecurKind Kind, Instruction *I) { | ||||
| 7041 | if (Kind == RecurKind::None) | ||||
| 7042 | return nullptr; | ||||
| 7043 | return I->getOperand(getFirstOperandIndex(I) + 1); | ||||
| 7044 | } | ||||
| 7045 | |||||
| 7046 | public: | ||||
| 7047 | HorizontalReduction() = default; | ||||
| 7048 | |||||
| 7049 | /// Try to find a reduction tree. | ||||
| 7050 | bool matchAssociativeReduction(PHINode *Phi, Instruction *B) { | ||||
| 7051 | assert((!Phi || is_contained(Phi->operands(), B)) &&(static_cast <bool> ((!Phi || is_contained(Phi->operands (), B)) && "Phi needs to use the binary operator") ? void (0) : __assert_fail ("(!Phi || is_contained(Phi->operands(), B)) && \"Phi needs to use the binary operator\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7052, __extension__ __PRETTY_FUNCTION__)) | ||||
| 7052 | "Phi needs to use the binary operator")(static_cast <bool> ((!Phi || is_contained(Phi->operands (), B)) && "Phi needs to use the binary operator") ? void (0) : __assert_fail ("(!Phi || is_contained(Phi->operands(), B)) && \"Phi needs to use the binary operator\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7052, __extension__ __PRETTY_FUNCTION__)); | ||||
| 7053 | |||||
| 7054 | RdxKind = getRdxKind(B); | ||||
| 7055 | |||||
| 7056 | // We could have a initial reductions that is not an add. | ||||
| 7057 | // r *= v1 + v2 + v3 + v4 | ||||
| 7058 | // In such a case start looking for a tree rooted in the first '+'. | ||||
| 7059 | if (Phi) { | ||||
| 7060 | if (getLHS(RdxKind, B) == Phi) { | ||||
| 7061 | Phi = nullptr; | ||||
| 7062 | B = dyn_cast<Instruction>(getRHS(RdxKind, B)); | ||||
| 7063 | if (!B) | ||||
| 7064 | return false; | ||||
| 7065 | RdxKind = getRdxKind(B); | ||||
| 7066 | } else if (getRHS(RdxKind, B) == Phi) { | ||||
| 7067 | Phi = nullptr; | ||||
| 7068 | B = dyn_cast<Instruction>(getLHS(RdxKind, B)); | ||||
| 7069 | if (!B) | ||||
| 7070 | return false; | ||||
| 7071 | RdxKind = getRdxKind(B); | ||||
| 7072 | } | ||||
| 7073 | } | ||||
| 7074 | |||||
| 7075 | if (!isVectorizable(RdxKind, B)) | ||||
| 7076 | return false; | ||||
| 7077 | |||||
| 7078 | // Analyze "regular" integer/FP types for reductions - no target-specific | ||||
| 7079 | // types or pointers. | ||||
| 7080 | Type *Ty = B->getType(); | ||||
| 7081 | if (!isValidElementType(Ty) || Ty->isPointerTy()) | ||||
| 7082 | return false; | ||||
| 7083 | |||||
| 7084 | // Though the ultimate reduction may have multiple uses, its condition must | ||||
| 7085 | // have only single use. | ||||
| 7086 | if (auto *SI = dyn_cast<SelectInst>(B)) | ||||
| 7087 | if (!SI->getCondition()->hasOneUse()) | ||||
| 7088 | return false; | ||||
| 7089 | |||||
| 7090 | ReductionRoot = B; | ||||
| 7091 | |||||
| 7092 | // The opcode for leaf values that we perform a reduction on. | ||||
| 7093 | // For example: load(x) + load(y) + load(z) + fptoui(w) | ||||
| 7094 | // The leaf opcode for 'w' does not match, so we don't include it as a | ||||
| 7095 | // potential candidate for the reduction. | ||||
| 7096 | unsigned LeafOpcode = 0; | ||||
| 7097 | |||||
| 7098 | // Post order traverse the reduction tree starting at B. We only handle true | ||||
| 7099 | // trees containing only binary operators. | ||||
| 7100 | SmallVector<std::pair<Instruction *, unsigned>, 32> Stack; | ||||
| 7101 | Stack.push_back(std::make_pair(B, getFirstOperandIndex(B))); | ||||
| 7102 | initReductionOps(B); | ||||
| 7103 | while (!Stack.empty()) { | ||||
| 7104 | Instruction *TreeN = Stack.back().first; | ||||
| 7105 | unsigned EdgeToVisit = Stack.back().second++; | ||||
| 7106 | const RecurKind TreeRdxKind = getRdxKind(TreeN); | ||||
| 7107 | bool IsReducedValue = TreeRdxKind != RdxKind; | ||||
| 7108 | |||||
| 7109 | // Postorder visit. | ||||
| 7110 | if (IsReducedValue || EdgeToVisit == getNumberOfOperands(TreeN)) { | ||||
| 7111 | if (IsReducedValue) | ||||
| 7112 | ReducedVals.push_back(TreeN); | ||||
| 7113 | else { | ||||
| 7114 | auto ExtraArgsIter = ExtraArgs.find(TreeN); | ||||
| 7115 | if (ExtraArgsIter != ExtraArgs.end() && !ExtraArgsIter->second) { | ||||
| 7116 | // Check if TreeN is an extra argument of its parent operation. | ||||
| 7117 | if (Stack.size() <= 1) { | ||||
| 7118 | // TreeN can't be an extra argument as it is a root reduction | ||||
| 7119 | // operation. | ||||
| 7120 | return false; | ||||
| 7121 | } | ||||
| 7122 | // Yes, TreeN is an extra argument, do not add it to a list of | ||||
| 7123 | // reduction operations. | ||||
| 7124 | // Stack[Stack.size() - 2] always points to the parent operation. | ||||
| 7125 | markExtraArg(Stack[Stack.size() - 2], TreeN); | ||||
| 7126 | ExtraArgs.erase(TreeN); | ||||
| 7127 | } else | ||||
| 7128 | addReductionOps(TreeN); | ||||
| 7129 | } | ||||
| 7130 | // Retract. | ||||
| 7131 | Stack.pop_back(); | ||||
| 7132 | continue; | ||||
| 7133 | } | ||||
| 7134 | |||||
| 7135 | // Visit left or right. | ||||
| 7136 | Value *EdgeVal = TreeN->getOperand(EdgeToVisit); | ||||
| 7137 | auto *EdgeInst = dyn_cast<Instruction>(EdgeVal); | ||||
| 7138 | if (!EdgeInst) { | ||||
| 7139 | // Edge value is not a reduction instruction or a leaf instruction. | ||||
| 7140 | // (It may be a constant, function argument, or something else.) | ||||
| 7141 | markExtraArg(Stack.back(), EdgeVal); | ||||
| 7142 | continue; | ||||
| 7143 | } | ||||
| 7144 | RecurKind EdgeRdxKind = getRdxKind(EdgeInst); | ||||
| 7145 | // Continue analysis if the next operand is a reduction operation or | ||||
| 7146 | // (possibly) a leaf value. If the leaf value opcode is not set, | ||||
| 7147 | // the first met operation != reduction operation is considered as the | ||||
| 7148 | // leaf opcode. | ||||
| 7149 | // Only handle trees in the current basic block. | ||||
| 7150 | // Each tree node needs to have minimal number of users except for the | ||||
| 7151 | // ultimate reduction. | ||||
| 7152 | const bool IsRdxInst = EdgeRdxKind == RdxKind; | ||||
| 7153 | if (EdgeInst != Phi && EdgeInst != B && | ||||
| 7154 | hasSameParent(EdgeInst, B->getParent(), IsRdxInst) && | ||||
| 7155 | hasRequiredNumberOfUses(isa<SelectInst>(B), EdgeInst) && | ||||
| 7156 | (!LeafOpcode || LeafOpcode == EdgeInst->getOpcode() || IsRdxInst)) { | ||||
| 7157 | if (IsRdxInst) { | ||||
| 7158 | // We need to be able to reassociate the reduction operations. | ||||
| 7159 | if (!isVectorizable(EdgeRdxKind, EdgeInst)) { | ||||
| 7160 | // I is an extra argument for TreeN (its parent operation). | ||||
| 7161 | markExtraArg(Stack.back(), EdgeInst); | ||||
| 7162 | continue; | ||||
| 7163 | } | ||||
| 7164 | } else if (!LeafOpcode) { | ||||
| 7165 | LeafOpcode = EdgeInst->getOpcode(); | ||||
| 7166 | } | ||||
| 7167 | Stack.push_back( | ||||
| 7168 | std::make_pair(EdgeInst, getFirstOperandIndex(EdgeInst))); | ||||
| 7169 | continue; | ||||
| 7170 | } | ||||
| 7171 | // I is an extra argument for TreeN (its parent operation). | ||||
| 7172 | markExtraArg(Stack.back(), EdgeInst); | ||||
| 7173 | } | ||||
| 7174 | return true; | ||||
| 7175 | } | ||||
| 7176 | |||||
| 7177 | /// Attempt to vectorize the tree found by matchAssociativeReduction. | ||||
| 7178 | bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) { | ||||
| 7179 | // If there are a sufficient number of reduction values, reduce | ||||
| 7180 | // to a nearby power-of-2. We can safely generate oversized | ||||
| 7181 | // vectors and rely on the backend to split them to legal sizes. | ||||
| 7182 | unsigned NumReducedVals = ReducedVals.size(); | ||||
| 7183 | if (NumReducedVals < 4) | ||||
| 7184 | return false; | ||||
| 7185 | |||||
| 7186 | // Intersect the fast-math-flags from all reduction operations. | ||||
| 7187 | FastMathFlags RdxFMF; | ||||
| 7188 | RdxFMF.set(); | ||||
| 7189 | for (ReductionOpsType &RdxOp : ReductionOps) { | ||||
| 7190 | for (Value *RdxVal : RdxOp) { | ||||
| 7191 | if (auto *FPMO = dyn_cast<FPMathOperator>(RdxVal)) | ||||
| 7192 | RdxFMF &= FPMO->getFastMathFlags(); | ||||
| 7193 | } | ||||
| 7194 | } | ||||
| 7195 | |||||
| 7196 | IRBuilder<> Builder(cast<Instruction>(ReductionRoot)); | ||||
| 7197 | Builder.setFastMathFlags(RdxFMF); | ||||
| 7198 | |||||
| 7199 | BoUpSLP::ExtraValueToDebugLocsMap ExternallyUsedValues; | ||||
| 7200 | // The same extra argument may be used several times, so log each attempt | ||||
| 7201 | // to use it. | ||||
| 7202 | for (const std::pair<Instruction *, Value *> &Pair : ExtraArgs) { | ||||
| 7203 | assert(Pair.first && "DebugLoc must be set.")(static_cast <bool> (Pair.first && "DebugLoc must be set." ) ? void (0) : __assert_fail ("Pair.first && \"DebugLoc must be set.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7203, __extension__ __PRETTY_FUNCTION__)); | ||||
| 7204 | ExternallyUsedValues[Pair.second].push_back(Pair.first); | ||||
| 7205 | } | ||||
| 7206 | |||||
| 7207 | // The compare instruction of a min/max is the insertion point for new | ||||
| 7208 | // instructions and may be replaced with a new compare instruction. | ||||
| 7209 | auto getCmpForMinMaxReduction = [](Instruction *RdxRootInst) { | ||||
| 7210 | assert(isa<SelectInst>(RdxRootInst) &&(static_cast <bool> (isa<SelectInst>(RdxRootInst) && "Expected min/max reduction to have select root instruction" ) ? void (0) : __assert_fail ("isa<SelectInst>(RdxRootInst) && \"Expected min/max reduction to have select root instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7211, __extension__ __PRETTY_FUNCTION__)) | ||||
| 7211 | "Expected min/max reduction to have select root instruction")(static_cast <bool> (isa<SelectInst>(RdxRootInst) && "Expected min/max reduction to have select root instruction" ) ? void (0) : __assert_fail ("isa<SelectInst>(RdxRootInst) && \"Expected min/max reduction to have select root instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7211, __extension__ __PRETTY_FUNCTION__)); | ||||
| 7212 | Value *ScalarCond = cast<SelectInst>(RdxRootInst)->getCondition(); | ||||
| 7213 | assert(isa<Instruction>(ScalarCond) &&(static_cast <bool> (isa<Instruction>(ScalarCond) && "Expected min/max reduction to have compare condition" ) ? void (0) : __assert_fail ("isa<Instruction>(ScalarCond) && \"Expected min/max reduction to have compare condition\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7214, __extension__ __PRETTY_FUNCTION__)) | ||||
| 7214 | "Expected min/max reduction to have compare condition")(static_cast <bool> (isa<Instruction>(ScalarCond) && "Expected min/max reduction to have compare condition" ) ? void (0) : __assert_fail ("isa<Instruction>(ScalarCond) && \"Expected min/max reduction to have compare condition\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7214, __extension__ __PRETTY_FUNCTION__)); | ||||
| 7215 | return cast<Instruction>(ScalarCond); | ||||
| 7216 | }; | ||||
| 7217 | |||||
| 7218 | // The reduction root is used as the insertion point for new instructions, | ||||
| 7219 | // so set it as externally used to prevent it from being deleted. | ||||
| 7220 | ExternallyUsedValues[ReductionRoot]; | ||||
| 7221 | SmallVector<Value *, 16> IgnoreList; | ||||
| 7222 | for (ReductionOpsType &RdxOp : ReductionOps) | ||||
| 7223 | IgnoreList.append(RdxOp.begin(), RdxOp.end()); | ||||
| 7224 | |||||
| 7225 | unsigned ReduxWidth = PowerOf2Floor(NumReducedVals); | ||||
| 7226 | if (NumReducedVals > ReduxWidth) { | ||||
| 7227 | // In the loop below, we are building a tree based on a window of | ||||
| 7228 | // 'ReduxWidth' values. | ||||
| 7229 | // If the operands of those values have common traits (compare predicate, | ||||
| 7230 | // constant operand, etc), then we want to group those together to | ||||
| 7231 | // minimize the cost of the reduction. | ||||
| 7232 | |||||
| 7233 | // TODO: This should be extended to count common operands for | ||||
| 7234 | // compares and binops. | ||||
| 7235 | |||||
| 7236 | // Step 1: Count the number of times each compare predicate occurs. | ||||
| 7237 | SmallDenseMap<unsigned, unsigned> PredCountMap; | ||||
| 7238 | for (Value *RdxVal : ReducedVals) { | ||||
| 7239 | CmpInst::Predicate Pred; | ||||
| 7240 | if (match(RdxVal, m_Cmp(Pred, m_Value(), m_Value()))) | ||||
| 7241 | ++PredCountMap[Pred]; | ||||
| 7242 | } | ||||
| 7243 | // Step 2: Sort the values so the most common predicates come first. | ||||
| 7244 | stable_sort(ReducedVals, [&PredCountMap](Value *A, Value *B) { | ||||
| 7245 | CmpInst::Predicate PredA, PredB; | ||||
| 7246 | if (match(A, m_Cmp(PredA, m_Value(), m_Value())) && | ||||
| 7247 | match(B, m_Cmp(PredB, m_Value(), m_Value()))) { | ||||
| 7248 | return PredCountMap[PredA] > PredCountMap[PredB]; | ||||
| 7249 | } | ||||
| 7250 | return false; | ||||
| 7251 | }); | ||||
| 7252 | } | ||||
| 7253 | |||||
| 7254 | Value *VectorizedTree = nullptr; | ||||
| 7255 | unsigned i = 0; | ||||
| 7256 | while (i < NumReducedVals - ReduxWidth + 1 && ReduxWidth > 2) { | ||||
| 7257 | ArrayRef<Value *> VL(&ReducedVals[i], ReduxWidth); | ||||
| 7258 | V.buildTree(VL, ExternallyUsedValues, IgnoreList); | ||||
| 7259 | Optional<ArrayRef<unsigned>> Order = V.bestOrder(); | ||||
| 7260 | if (Order) { | ||||
| 7261 | assert(Order->size() == VL.size() &&(static_cast <bool> (Order->size() == VL.size() && "Order size must be the same as number of vectorized " "instructions." ) ? void (0) : __assert_fail ("Order->size() == VL.size() && \"Order size must be the same as number of vectorized \" \"instructions.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7263, __extension__ __PRETTY_FUNCTION__)) | ||||
| 7262 | "Order size must be the same as number of vectorized "(static_cast <bool> (Order->size() == VL.size() && "Order size must be the same as number of vectorized " "instructions." ) ? void (0) : __assert_fail ("Order->size() == VL.size() && \"Order size must be the same as number of vectorized \" \"instructions.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7263, __extension__ __PRETTY_FUNCTION__)) | ||||
| 7263 | "instructions.")(static_cast <bool> (Order->size() == VL.size() && "Order size must be the same as number of vectorized " "instructions." ) ? void (0) : __assert_fail ("Order->size() == VL.size() && \"Order size must be the same as number of vectorized \" \"instructions.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7263, __extension__ __PRETTY_FUNCTION__)); | ||||
| 7264 | // TODO: reorder tree nodes without tree rebuilding. | ||||
| 7265 | SmallVector<Value *, 4> ReorderedOps(VL.size()); | ||||
| 7266 | llvm::transform(*Order, ReorderedOps.begin(), | ||||
| 7267 | [VL](const unsigned Idx) { return VL[Idx]; }); | ||||
| 7268 | V.buildTree(ReorderedOps, ExternallyUsedValues, IgnoreList); | ||||
| 7269 | } | ||||
| 7270 | if (V.isTreeTinyAndNotFullyVectorizable()) | ||||
| 7271 | break; | ||||
| 7272 | if (V.isLoadCombineReductionCandidate(RdxKind)) | ||||
| 7273 | break; | ||||
| 7274 | |||||
| 7275 | V.computeMinimumValueSizes(); | ||||
| 7276 | |||||
| 7277 | // Estimate cost. | ||||
| 7278 | InstructionCost TreeCost = V.getTreeCost(); | ||||
| 7279 | InstructionCost ReductionCost = | ||||
| 7280 | getReductionCost(TTI, ReducedVals[i], ReduxWidth); | ||||
| 7281 | InstructionCost Cost = TreeCost + ReductionCost; | ||||
| 7282 | if (!Cost.isValid()) { | ||||
| 7283 | LLVM_DEBUG(dbgs() << "Encountered invalid baseline cost.\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "Encountered invalid baseline cost.\n" ; } } while (false); | ||||
| 7284 | return false; | ||||
| 7285 | } | ||||
| 7286 | if (Cost >= -SLPCostThreshold) { | ||||
| 7287 | V.getORE()->emit([&]() { | ||||
| 7288 | return OptimizationRemarkMissed(SV_NAME"slp-vectorizer", "HorSLPNotBeneficial", | ||||
| 7289 | cast<Instruction>(VL[0])) | ||||
| 7290 | << "Vectorizing horizontal reduction is possible" | ||||
| 7291 | << "but not beneficial with cost " << ore::NV("Cost", Cost) | ||||
| 7292 | << " and threshold " | ||||
| 7293 | << ore::NV("Threshold", -SLPCostThreshold); | ||||
| 7294 | }); | ||||
| 7295 | break; | ||||
| 7296 | } | ||||
| 7297 | |||||
| 7298 | LLVM_DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost << ". (HorRdx)\n"; } } while (false) | ||||
| 7299 | << Cost << ". (HorRdx)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost << ". (HorRdx)\n"; } } while (false); | ||||
| 7300 | V.getORE()->emit([&]() { | ||||
| 7301 | return OptimizationRemark(SV_NAME"slp-vectorizer", "VectorizedHorizontalReduction", | ||||
| 7302 | cast<Instruction>(VL[0])) | ||||
| 7303 | << "Vectorized horizontal reduction with cost " | ||||
| 7304 | << ore::NV("Cost", Cost) << " and with tree size " | ||||
| 7305 | << ore::NV("TreeSize", V.getTreeSize()); | ||||
| 7306 | }); | ||||
| 7307 | |||||
| 7308 | // Vectorize a tree. | ||||
| 7309 | DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc(); | ||||
| 7310 | Value *VectorizedRoot = V.vectorizeTree(ExternallyUsedValues); | ||||
| 7311 | |||||
| 7312 | // Emit a reduction. If the root is a select (min/max idiom), the insert | ||||
| 7313 | // point is the compare condition of that select. | ||||
| 7314 | Instruction *RdxRootInst = cast<Instruction>(ReductionRoot); | ||||
| 7315 | if (isa<SelectInst>(RdxRootInst)) | ||||
| 7316 | Builder.SetInsertPoint(getCmpForMinMaxReduction(RdxRootInst)); | ||||
| 7317 | else | ||||
| 7318 | Builder.SetInsertPoint(RdxRootInst); | ||||
| 7319 | |||||
| 7320 | Value *ReducedSubTree = | ||||
| 7321 | emitReduction(VectorizedRoot, Builder, ReduxWidth, TTI); | ||||
| 7322 | |||||
| 7323 | if (!VectorizedTree) { | ||||
| 7324 | // Initialize the final value in the reduction. | ||||
| 7325 | VectorizedTree = ReducedSubTree; | ||||
| 7326 | } else { | ||||
| 7327 | // Update the final value in the reduction. | ||||
| 7328 | Builder.SetCurrentDebugLocation(Loc); | ||||
| 7329 | VectorizedTree = createOp(Builder, RdxKind, VectorizedTree, | ||||
| 7330 | ReducedSubTree, "op.rdx", ReductionOps); | ||||
| 7331 | } | ||||
| 7332 | i += ReduxWidth; | ||||
| 7333 | ReduxWidth = PowerOf2Floor(NumReducedVals - i); | ||||
| 7334 | } | ||||
| 7335 | |||||
| 7336 | if (VectorizedTree) { | ||||
| 7337 | // Finish the reduction. | ||||
| 7338 | for (; i < NumReducedVals; ++i) { | ||||
| 7339 | auto *I = cast<Instruction>(ReducedVals[i]); | ||||
| 7340 | Builder.SetCurrentDebugLocation(I->getDebugLoc()); | ||||
| 7341 | VectorizedTree = | ||||
| 7342 | createOp(Builder, RdxKind, VectorizedTree, I, "", ReductionOps); | ||||
| 7343 | } | ||||
| 7344 | for (auto &Pair : ExternallyUsedValues) { | ||||
| 7345 | // Add each externally used value to the final reduction. | ||||
| 7346 | for (auto *I : Pair.second) { | ||||
| 7347 | Builder.SetCurrentDebugLocation(I->getDebugLoc()); | ||||
| 7348 | VectorizedTree = createOp(Builder, RdxKind, VectorizedTree, | ||||
| 7349 | Pair.first, "op.extra", I); | ||||
| 7350 | } | ||||
| 7351 | } | ||||
| 7352 | |||||
| 7353 | ReductionRoot->replaceAllUsesWith(VectorizedTree); | ||||
| 7354 | |||||
| 7355 | // Mark all scalar reduction ops for deletion, they are replaced by the | ||||
| 7356 | // vector reductions. | ||||
| 7357 | V.eraseInstructions(IgnoreList); | ||||
| 7358 | } | ||||
| 7359 | return VectorizedTree != nullptr; | ||||
| 7360 | } | ||||
| 7361 | |||||
| 7362 | unsigned numReductionValues() const { return ReducedVals.size(); } | ||||
| 7363 | |||||
| 7364 | private: | ||||
| 7365 | /// Calculate the cost of a reduction. | ||||
| 7366 | InstructionCost getReductionCost(TargetTransformInfo *TTI, | ||||
| 7367 | Value *FirstReducedVal, | ||||
| 7368 | unsigned ReduxWidth) { | ||||
| 7369 | Type *ScalarTy = FirstReducedVal->getType(); | ||||
| 7370 | FixedVectorType *VectorTy = FixedVectorType::get(ScalarTy, ReduxWidth); | ||||
| 7371 | InstructionCost VectorCost, ScalarCost; | ||||
| 7372 | switch (RdxKind) { | ||||
| 7373 | case RecurKind::Add: | ||||
| 7374 | case RecurKind::Mul: | ||||
| 7375 | case RecurKind::Or: | ||||
| 7376 | case RecurKind::And: | ||||
| 7377 | case RecurKind::Xor: | ||||
| 7378 | case RecurKind::FAdd: | ||||
| 7379 | case RecurKind::FMul: { | ||||
| 7380 | unsigned RdxOpcode = RecurrenceDescriptor::getOpcode(RdxKind); | ||||
| 7381 | VectorCost = TTI->getArithmeticReductionCost(RdxOpcode, VectorTy, | ||||
| 7382 | /*IsPairwiseForm=*/false); | ||||
| 7383 | ScalarCost = TTI->getArithmeticInstrCost(RdxOpcode, ScalarTy); | ||||
| 7384 | break; | ||||
| 7385 | } | ||||
| 7386 | case RecurKind::FMax: | ||||
| 7387 | case RecurKind::FMin: { | ||||
| 7388 | auto *VecCondTy = cast<VectorType>(CmpInst::makeCmpResultType(VectorTy)); | ||||
| 7389 | VectorCost = | ||||
| 7390 | TTI->getMinMaxReductionCost(VectorTy, VecCondTy, | ||||
| 7391 | /*pairwise=*/false, /*unsigned=*/false); | ||||
| 7392 | ScalarCost = | ||||
| 7393 | TTI->getCmpSelInstrCost(Instruction::FCmp, ScalarTy) + | ||||
| 7394 | TTI->getCmpSelInstrCost(Instruction::Select, ScalarTy, | ||||
| 7395 | CmpInst::makeCmpResultType(ScalarTy)); | ||||
| 7396 | break; | ||||
| 7397 | } | ||||
| 7398 | case RecurKind::SMax: | ||||
| 7399 | case RecurKind::SMin: | ||||
| 7400 | case RecurKind::UMax: | ||||
| 7401 | case RecurKind::UMin: { | ||||
| 7402 | auto *VecCondTy = cast<VectorType>(CmpInst::makeCmpResultType(VectorTy)); | ||||
| 7403 | bool IsUnsigned = | ||||
| 7404 | RdxKind == RecurKind::UMax || RdxKind == RecurKind::UMin; | ||||
| 7405 | VectorCost = | ||||
| 7406 | TTI->getMinMaxReductionCost(VectorTy, VecCondTy, | ||||
| 7407 | /*IsPairwiseForm=*/false, IsUnsigned); | ||||
| 7408 | ScalarCost = | ||||
| 7409 | TTI->getCmpSelInstrCost(Instruction::ICmp, ScalarTy) + | ||||
| 7410 | TTI->getCmpSelInstrCost(Instruction::Select, ScalarTy, | ||||
| 7411 | CmpInst::makeCmpResultType(ScalarTy)); | ||||
| 7412 | break; | ||||
| 7413 | } | ||||
| 7414 | default: | ||||
| 7415 | llvm_unreachable("Expected arithmetic or min/max reduction operation")::llvm::llvm_unreachable_internal("Expected arithmetic or min/max reduction operation" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7415); | ||||
| 7416 | } | ||||
| 7417 | |||||
| 7418 | // Scalar cost is repeated for N-1 elements. | ||||
| 7419 | ScalarCost *= (ReduxWidth - 1); | ||||
| 7420 | LLVM_DEBUG(dbgs() << "SLP: Adding cost " << VectorCost - ScalarCostdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << VectorCost - ScalarCost << " for reduction that starts with " << *FirstReducedVal << " (It is a splitting reduction)\n" ; } } while (false) | ||||
| 7421 | << " for reduction that starts with " << *FirstReducedValdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << VectorCost - ScalarCost << " for reduction that starts with " << *FirstReducedVal << " (It is a splitting reduction)\n" ; } } while (false) | ||||
| 7422 | << " (It is a splitting reduction)\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Adding cost " << VectorCost - ScalarCost << " for reduction that starts with " << *FirstReducedVal << " (It is a splitting reduction)\n" ; } } while (false); | ||||
| 7423 | return VectorCost - ScalarCost; | ||||
| 7424 | } | ||||
| 7425 | |||||
| 7426 | /// Emit a horizontal reduction of the vectorized value. | ||||
| 7427 | Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder, | ||||
| 7428 | unsigned ReduxWidth, const TargetTransformInfo *TTI) { | ||||
| 7429 | assert(VectorizedValue && "Need to have a vectorized tree node")(static_cast <bool> (VectorizedValue && "Need to have a vectorized tree node" ) ? void (0) : __assert_fail ("VectorizedValue && \"Need to have a vectorized tree node\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7429, __extension__ __PRETTY_FUNCTION__)); | ||||
| 7430 | assert(isPowerOf2_32(ReduxWidth) &&(static_cast <bool> (isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now") ? void (0) : __assert_fail ("isPowerOf2_32(ReduxWidth) && \"We only handle power-of-two reductions for now\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7431, __extension__ __PRETTY_FUNCTION__)) | ||||
| 7431 | "We only handle power-of-two reductions for now")(static_cast <bool> (isPowerOf2_32(ReduxWidth) && "We only handle power-of-two reductions for now") ? void (0) : __assert_fail ("isPowerOf2_32(ReduxWidth) && \"We only handle power-of-two reductions for now\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7431, __extension__ __PRETTY_FUNCTION__)); | ||||
| 7432 | |||||
| 7433 | return createSimpleTargetReduction(Builder, TTI, VectorizedValue, RdxKind, | ||||
| 7434 | ReductionOps.back()); | ||||
| 7435 | } | ||||
| 7436 | }; | ||||
| 7437 | |||||
| 7438 | } // end anonymous namespace | ||||
| 7439 | |||||
| 7440 | static Optional<unsigned> getAggregateSize(Instruction *InsertInst) { | ||||
| 7441 | if (auto *IE = dyn_cast<InsertElementInst>(InsertInst)) | ||||
| 7442 | return cast<FixedVectorType>(IE->getType())->getNumElements(); | ||||
| 7443 | |||||
| 7444 | unsigned AggregateSize = 1; | ||||
| 7445 | auto *IV = cast<InsertValueInst>(InsertInst); | ||||
| 7446 | Type *CurrentType = IV->getType(); | ||||
| 7447 | do { | ||||
| 7448 | if (auto *ST = dyn_cast<StructType>(CurrentType)) { | ||||
| 7449 | for (auto *Elt : ST->elements()) | ||||
| 7450 | if (Elt != ST->getElementType(0)) // check homogeneity | ||||
| 7451 | return None; | ||||
| 7452 | AggregateSize *= ST->getNumElements(); | ||||
| 7453 | CurrentType = ST->getElementType(0); | ||||
| 7454 | } else if (auto *AT = dyn_cast<ArrayType>(CurrentType)) { | ||||
| 7455 | AggregateSize *= AT->getNumElements(); | ||||
| 7456 | CurrentType = AT->getElementType(); | ||||
| 7457 | } else if (auto *VT = dyn_cast<FixedVectorType>(CurrentType)) { | ||||
| 7458 | AggregateSize *= VT->getNumElements(); | ||||
| 7459 | return AggregateSize; | ||||
| 7460 | } else if (CurrentType->isSingleValueType()) { | ||||
| 7461 | return AggregateSize; | ||||
| 7462 | } else { | ||||
| 7463 | return None; | ||||
| 7464 | } | ||||
| 7465 | } while (true); | ||||
| 7466 | } | ||||
| 7467 | |||||
| 7468 | static Optional<unsigned> getOperandIndex(Instruction *InsertInst, | ||||
| 7469 | unsigned OperandOffset) { | ||||
| 7470 | unsigned OperandIndex = OperandOffset; | ||||
| 7471 | if (auto *IE = dyn_cast<InsertElementInst>(InsertInst)) { | ||||
| 7472 | if (auto *CI = dyn_cast<ConstantInt>(IE->getOperand(2))) { | ||||
| 7473 | auto *VT = cast<FixedVectorType>(IE->getType()); | ||||
| 7474 | OperandIndex *= VT->getNumElements(); | ||||
| 7475 | OperandIndex += CI->getZExtValue(); | ||||
| 7476 | return OperandIndex; | ||||
| 7477 | } | ||||
| 7478 | return None; | ||||
| 7479 | } | ||||
| 7480 | |||||
| 7481 | auto *IV = cast<InsertValueInst>(InsertInst); | ||||
| 7482 | Type *CurrentType = IV->getType(); | ||||
| 7483 | for (unsigned int Index : IV->indices()) { | ||||
| 7484 | if (auto *ST = dyn_cast<StructType>(CurrentType)) { | ||||
| 7485 | OperandIndex *= ST->getNumElements(); | ||||
| 7486 | CurrentType = ST->getElementType(Index); | ||||
| 7487 | } else if (auto *AT = dyn_cast<ArrayType>(CurrentType)) { | ||||
| 7488 | OperandIndex *= AT->getNumElements(); | ||||
| 7489 | CurrentType = AT->getElementType(); | ||||
| 7490 | } else { | ||||
| 7491 | return None; | ||||
| 7492 | } | ||||
| 7493 | OperandIndex += Index; | ||||
| 7494 | } | ||||
| 7495 | return OperandIndex; | ||||
| 7496 | } | ||||
| 7497 | |||||
| 7498 | static bool findBuildAggregate_rec(Instruction *LastInsertInst, | ||||
| 7499 | TargetTransformInfo *TTI, | ||||
| 7500 | SmallVectorImpl<Value *> &BuildVectorOpds, | ||||
| 7501 | SmallVectorImpl<Value *> &InsertElts, | ||||
| 7502 | unsigned OperandOffset) { | ||||
| 7503 | do { | ||||
| 7504 | Value *InsertedOperand = LastInsertInst->getOperand(1); | ||||
| 7505 | Optional<unsigned> OperandIndex = | ||||
| 7506 | getOperandIndex(LastInsertInst, OperandOffset); | ||||
| 7507 | if (!OperandIndex) | ||||
| 7508 | return false; | ||||
| 7509 | if (isa<InsertElementInst>(InsertedOperand) || | ||||
| 7510 | isa<InsertValueInst>(InsertedOperand)) { | ||||
| 7511 | if (!findBuildAggregate_rec(cast<Instruction>(InsertedOperand), TTI, | ||||
| 7512 | BuildVectorOpds, InsertElts, *OperandIndex)) | ||||
| 7513 | return false; | ||||
| 7514 | } else { | ||||
| 7515 | BuildVectorOpds[*OperandIndex] = InsertedOperand; | ||||
| 7516 | InsertElts[*OperandIndex] = LastInsertInst; | ||||
| 7517 | } | ||||
| 7518 | if (isa<UndefValue>(LastInsertInst->getOperand(0))) | ||||
| 7519 | return true; | ||||
| 7520 | LastInsertInst = dyn_cast<Instruction>(LastInsertInst->getOperand(0)); | ||||
| 7521 | } while (LastInsertInst != nullptr && | ||||
| 7522 | (isa<InsertValueInst>(LastInsertInst) || | ||||
| 7523 | isa<InsertElementInst>(LastInsertInst)) && | ||||
| 7524 | LastInsertInst->hasOneUse()); | ||||
| 7525 | return false; | ||||
| 7526 | } | ||||
| 7527 | |||||
| 7528 | /// Recognize construction of vectors like | ||||
| 7529 | /// %ra = insertelement <4 x float> poison, float %s0, i32 0 | ||||
| 7530 | /// %rb = insertelement <4 x float> %ra, float %s1, i32 1 | ||||
| 7531 | /// %rc = insertelement <4 x float> %rb, float %s2, i32 2 | ||||
| 7532 | /// %rd = insertelement <4 x float> %rc, float %s3, i32 3 | ||||
| 7533 | /// starting from the last insertelement or insertvalue instruction. | ||||
| 7534 | /// | ||||
| 7535 | /// Also recognize homogeneous aggregates like {<2 x float>, <2 x float>}, | ||||
| 7536 | /// {{float, float}, {float, float}}, [2 x {float, float}] and so on. | ||||
| 7537 | /// See llvm/test/Transforms/SLPVectorizer/X86/pr42022.ll for examples. | ||||
| 7538 | /// | ||||
| 7539 | /// Assume LastInsertInst is of InsertElementInst or InsertValueInst type. | ||||
| 7540 | /// | ||||
| 7541 | /// \return true if it matches. | ||||
| 7542 | static bool findBuildAggregate(Instruction *LastInsertInst, | ||||
| 7543 | TargetTransformInfo *TTI, | ||||
| 7544 | SmallVectorImpl<Value *> &BuildVectorOpds, | ||||
| 7545 | SmallVectorImpl<Value *> &InsertElts) { | ||||
| 7546 | |||||
| 7547 | assert((isa<InsertElementInst>(LastInsertInst) ||(static_cast <bool> ((isa<InsertElementInst>(LastInsertInst ) || isa<InsertValueInst>(LastInsertInst)) && "Expected insertelement or insertvalue instruction!" ) ? void (0) : __assert_fail ("(isa<InsertElementInst>(LastInsertInst) || isa<InsertValueInst>(LastInsertInst)) && \"Expected insertelement or insertvalue instruction!\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7549, __extension__ __PRETTY_FUNCTION__)) | ||||
| 7548 | isa<InsertValueInst>(LastInsertInst)) &&(static_cast <bool> ((isa<InsertElementInst>(LastInsertInst ) || isa<InsertValueInst>(LastInsertInst)) && "Expected insertelement or insertvalue instruction!" ) ? void (0) : __assert_fail ("(isa<InsertElementInst>(LastInsertInst) || isa<InsertValueInst>(LastInsertInst)) && \"Expected insertelement or insertvalue instruction!\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7549, __extension__ __PRETTY_FUNCTION__)) | ||||
| 7549 | "Expected insertelement or insertvalue instruction!")(static_cast <bool> ((isa<InsertElementInst>(LastInsertInst ) || isa<InsertValueInst>(LastInsertInst)) && "Expected insertelement or insertvalue instruction!" ) ? void (0) : __assert_fail ("(isa<InsertElementInst>(LastInsertInst) || isa<InsertValueInst>(LastInsertInst)) && \"Expected insertelement or insertvalue instruction!\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7549, __extension__ __PRETTY_FUNCTION__)); | ||||
| 7550 | |||||
| 7551 | assert((BuildVectorOpds.empty() && InsertElts.empty()) &&(static_cast <bool> ((BuildVectorOpds.empty() && InsertElts.empty()) && "Expected empty result vectors!" ) ? void (0) : __assert_fail ("(BuildVectorOpds.empty() && InsertElts.empty()) && \"Expected empty result vectors!\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7552, __extension__ __PRETTY_FUNCTION__)) | ||||
| 7552 | "Expected empty result vectors!")(static_cast <bool> ((BuildVectorOpds.empty() && InsertElts.empty()) && "Expected empty result vectors!" ) ? void (0) : __assert_fail ("(BuildVectorOpds.empty() && InsertElts.empty()) && \"Expected empty result vectors!\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 7552, __extension__ __PRETTY_FUNCTION__)); | ||||
| 7553 | |||||
| 7554 | Optional<unsigned> AggregateSize = getAggregateSize(LastInsertInst); | ||||
| 7555 | if (!AggregateSize) | ||||
| 7556 | return false; | ||||
| 7557 | BuildVectorOpds.resize(*AggregateSize); | ||||
| 7558 | InsertElts.resize(*AggregateSize); | ||||
| 7559 | |||||
| 7560 | if (findBuildAggregate_rec(LastInsertInst, TTI, BuildVectorOpds, InsertElts, | ||||
| 7561 | 0)) { | ||||
| 7562 | llvm::erase_value(BuildVectorOpds, nullptr); | ||||
| 7563 | llvm::erase_value(InsertElts, nullptr); | ||||
| 7564 | if (BuildVectorOpds.size() >= 2) | ||||
| 7565 | return true; | ||||
| 7566 | } | ||||
| 7567 | |||||
| 7568 | return false; | ||||
| 7569 | } | ||||
| 7570 | |||||
| 7571 | static bool PhiTypeSorterFunc(Value *V, Value *V2) { | ||||
| 7572 | return V->getType() < V2->getType(); | ||||
| 7573 | } | ||||
| 7574 | |||||
| 7575 | /// Try and get a reduction value from a phi node. | ||||
| 7576 | /// | ||||
| 7577 | /// Given a phi node \p P in a block \p ParentBB, consider possible reductions | ||||
| 7578 | /// if they come from either \p ParentBB or a containing loop latch. | ||||
| 7579 | /// | ||||
| 7580 | /// \returns A candidate reduction value if possible, or \code nullptr \endcode | ||||
| 7581 | /// if not possible. | ||||
| 7582 | static Value *getReductionValue(const DominatorTree *DT, PHINode *P, | ||||
| 7583 | BasicBlock *ParentBB, LoopInfo *LI) { | ||||
| 7584 | // There are situations where the reduction value is not dominated by the | ||||
| 7585 | // reduction phi. Vectorizing such cases has been reported to cause | ||||
| 7586 | // miscompiles. See PR25787. | ||||
| 7587 | auto DominatedReduxValue = [&](Value *R) { | ||||
| 7588 | return isa<Instruction>(R) && | ||||
| 7589 | DT->dominates(P->getParent(), cast<Instruction>(R)->getParent()); | ||||
| 7590 | }; | ||||
| 7591 | |||||
| 7592 | Value *Rdx = nullptr; | ||||
| 7593 | |||||
| 7594 | // Return the incoming value if it comes from the same BB as the phi node. | ||||
| 7595 | if (P->getIncomingBlock(0) == ParentBB) { | ||||
| 7596 | Rdx = P->getIncomingValue(0); | ||||
| 7597 | } else if (P->getIncomingBlock(1) == ParentBB) { | ||||
| 7598 | Rdx = P->getIncomingValue(1); | ||||
| 7599 | } | ||||
| 7600 | |||||
| 7601 | if (Rdx && DominatedReduxValue(Rdx)) | ||||
| 7602 | return Rdx; | ||||
| 7603 | |||||
| 7604 | // Otherwise, check whether we have a loop latch to look at. | ||||
| 7605 | Loop *BBL = LI->getLoopFor(ParentBB); | ||||
| 7606 | if (!BBL) | ||||
| 7607 | return nullptr; | ||||
| 7608 | BasicBlock *BBLatch = BBL->getLoopLatch(); | ||||
| 7609 | if (!BBLatch) | ||||
| 7610 | return nullptr; | ||||
| 7611 | |||||
| 7612 | // There is a loop latch, return the incoming value if it comes from | ||||
| 7613 | // that. This reduction pattern occasionally turns up. | ||||
| 7614 | if (P->getIncomingBlock(0) == BBLatch) { | ||||
| 7615 | Rdx = P->getIncomingValue(0); | ||||
| 7616 | } else if (P->getIncomingBlock(1) == BBLatch) { | ||||
| 7617 | Rdx = P->getIncomingValue(1); | ||||
| 7618 | } | ||||
| 7619 | |||||
| 7620 | if (Rdx && DominatedReduxValue(Rdx)) | ||||
| 7621 | return Rdx; | ||||
| 7622 | |||||
| 7623 | return nullptr; | ||||
| 7624 | } | ||||
| 7625 | |||||
| 7626 | static bool matchRdxBop(Instruction *I, Value *&V0, Value *&V1) { | ||||
| 7627 | if (match(I, m_BinOp(m_Value(V0), m_Value(V1)))) | ||||
| 7628 | return true; | ||||
| 7629 | if (match(I, m_Intrinsic<Intrinsic::maxnum>(m_Value(V0), m_Value(V1)))) | ||||
| 7630 | return true; | ||||
| 7631 | if (match(I, m_Intrinsic<Intrinsic::minnum>(m_Value(V0), m_Value(V1)))) | ||||
| 7632 | return true; | ||||
| 7633 | if (match(I, m_Intrinsic<Intrinsic::smax>(m_Value(V0), m_Value(V1)))) | ||||
| 7634 | return true; | ||||
| 7635 | if (match(I, m_Intrinsic<Intrinsic::smin>(m_Value(V0), m_Value(V1)))) | ||||
| 7636 | return true; | ||||
| 7637 | if (match(I, m_Intrinsic<Intrinsic::umax>(m_Value(V0), m_Value(V1)))) | ||||
| 7638 | return true; | ||||
| 7639 | if (match(I, m_Intrinsic<Intrinsic::umin>(m_Value(V0), m_Value(V1)))) | ||||
| 7640 | return true; | ||||
| 7641 | return false; | ||||
| 7642 | } | ||||
| 7643 | |||||
| 7644 | /// Attempt to reduce a horizontal reduction. | ||||
| 7645 | /// If it is legal to match a horizontal reduction feeding the phi node \a P | ||||
| 7646 | /// with reduction operators \a Root (or one of its operands) in a basic block | ||||
| 7647 | /// \a BB, then check if it can be done. If horizontal reduction is not found | ||||
| 7648 | /// and root instruction is a binary operation, vectorization of the operands is | ||||
| 7649 | /// attempted. | ||||
| 7650 | /// \returns true if a horizontal reduction was matched and reduced or operands | ||||
| 7651 | /// of one of the binary instruction were vectorized. | ||||
| 7652 | /// \returns false if a horizontal reduction was not matched (or not possible) | ||||
| 7653 | /// or no vectorization of any binary operation feeding \a Root instruction was | ||||
| 7654 | /// performed. | ||||
| 7655 | static bool tryToVectorizeHorReductionOrInstOperands( | ||||
| 7656 | PHINode *P, Instruction *Root, BasicBlock *BB, BoUpSLP &R, | ||||
| 7657 | TargetTransformInfo *TTI, | ||||
| 7658 | const function_ref<bool(Instruction *, BoUpSLP &)> Vectorize) { | ||||
| 7659 | if (!ShouldVectorizeHor) | ||||
| 7660 | return false; | ||||
| 7661 | |||||
| 7662 | if (!Root) | ||||
| 7663 | return false; | ||||
| 7664 | |||||
| 7665 | if (Root->getParent() != BB || isa<PHINode>(Root)) | ||||
| 7666 | return false; | ||||
| 7667 | // Start analysis starting from Root instruction. If horizontal reduction is | ||||
| 7668 | // found, try to vectorize it. If it is not a horizontal reduction or | ||||
| 7669 | // vectorization is not possible or not effective, and currently analyzed | ||||
| 7670 | // instruction is a binary operation, try to vectorize the operands, using | ||||
| 7671 | // pre-order DFS traversal order. If the operands were not vectorized, repeat | ||||
| 7672 | // the same procedure considering each operand as a possible root of the | ||||
| 7673 | // horizontal reduction. | ||||
| 7674 | // Interrupt the process if the Root instruction itself was vectorized or all | ||||
| 7675 | // sub-trees not higher that RecursionMaxDepth were analyzed/vectorized. | ||||
| 7676 | // Skip the analysis of CmpInsts.Compiler implements postanalysis of the | ||||
| 7677 | // CmpInsts so we can skip extra attempts in | ||||
| 7678 | // tryToVectorizeHorReductionOrInstOperands and save compile time. | ||||
| 7679 | SmallVector<std::pair<Instruction *, unsigned>, 8> Stack(1, {Root, 0}); | ||||
| 7680 | SmallPtrSet<Value *, 8> VisitedInstrs; | ||||
| 7681 | bool Res = false; | ||||
| 7682 | while (!Stack.empty()) { | ||||
| 7683 | Instruction *Inst; | ||||
| 7684 | unsigned Level; | ||||
| 7685 | std::tie(Inst, Level) = Stack.pop_back_val(); | ||||
| 7686 | Value *B0, *B1; | ||||
| 7687 | bool IsBinop = matchRdxBop(Inst, B0, B1); | ||||
| 7688 | bool IsSelect = match(Inst, m_Select(m_Value(), m_Value(), m_Value())); | ||||
| 7689 | if (IsBinop || IsSelect) { | ||||
| 7690 | HorizontalReduction HorRdx; | ||||
| 7691 | if (HorRdx.matchAssociativeReduction(P, Inst)) { | ||||
| 7692 | if (HorRdx.tryToReduce(R, TTI)) { | ||||
| 7693 | Res = true; | ||||
| 7694 | // Set P to nullptr to avoid re-analysis of phi node in | ||||
| 7695 | // matchAssociativeReduction function unless this is the root node. | ||||
| 7696 | P = nullptr; | ||||
| 7697 | continue; | ||||
| 7698 | } | ||||
| 7699 | } | ||||
| 7700 | if (P && IsBinop) { | ||||
| 7701 | Inst = dyn_cast<Instruction>(B0); | ||||
| 7702 | if (Inst == P) | ||||
| 7703 | Inst = dyn_cast<Instruction>(B1); | ||||
| 7704 | if (!Inst) { | ||||
| 7705 | // Set P to nullptr to avoid re-analysis of phi node in | ||||
| 7706 | // matchAssociativeReduction function unless this is the root node. | ||||
| 7707 | P = nullptr; | ||||
| 7708 | continue; | ||||
| 7709 | } | ||||
| 7710 | } | ||||
| 7711 | } | ||||
| 7712 | // Set P to nullptr to avoid re-analysis of phi node in | ||||
| 7713 | // matchAssociativeReduction function unless this is the root node. | ||||
| 7714 | P = nullptr; | ||||
| 7715 | // Do not try to vectorize CmpInst operands, this is done separately. | ||||
| 7716 | if (!isa<CmpInst>(Inst) && Vectorize(Inst, R)) { | ||||
| 7717 | Res = true; | ||||
| 7718 | continue; | ||||
| 7719 | } | ||||
| 7720 | |||||
| 7721 | // Try to vectorize operands. | ||||
| 7722 | // Continue analysis for the instruction from the same basic block only to | ||||
| 7723 | // save compile time. | ||||
| 7724 | if (++Level < RecursionMaxDepth) | ||||
| 7725 | for (auto *Op : Inst->operand_values()) | ||||
| 7726 | if (VisitedInstrs.insert(Op).second) | ||||
| 7727 | if (auto *I = dyn_cast<Instruction>(Op)) | ||||
| 7728 | // Do not try to vectorize CmpInst operands, this is done | ||||
| 7729 | // separately. | ||||
| 7730 | if (!isa<PHINode>(I) && !isa<CmpInst>(I) && !R.isDeleted(I) && | ||||
| 7731 | I->getParent() == BB) | ||||
| 7732 | Stack.emplace_back(I, Level); | ||||
| 7733 | } | ||||
| 7734 | return Res; | ||||
| 7735 | } | ||||
| 7736 | |||||
| 7737 | bool SLPVectorizerPass::vectorizeRootInstruction(PHINode *P, Value *V, | ||||
| 7738 | BasicBlock *BB, BoUpSLP &R, | ||||
| 7739 | TargetTransformInfo *TTI) { | ||||
| 7740 | auto *I = dyn_cast_or_null<Instruction>(V); | ||||
| 7741 | if (!I) | ||||
| 7742 | return false; | ||||
| 7743 | |||||
| 7744 | if (!isa<BinaryOperator>(I)) | ||||
| 7745 | P = nullptr; | ||||
| 7746 | // Try to match and vectorize a horizontal reduction. | ||||
| 7747 | auto &&ExtraVectorization = [this](Instruction *I, BoUpSLP &R) -> bool { | ||||
| 7748 | return tryToVectorize(I, R); | ||||
| 7749 | }; | ||||
| 7750 | return tryToVectorizeHorReductionOrInstOperands(P, I, BB, R, TTI, | ||||
| 7751 | ExtraVectorization); | ||||
| 7752 | } | ||||
| 7753 | |||||
| 7754 | bool SLPVectorizerPass::vectorizeInsertValueInst(InsertValueInst *IVI, | ||||
| 7755 | BasicBlock *BB, BoUpSLP &R) { | ||||
| 7756 | const DataLayout &DL = BB->getModule()->getDataLayout(); | ||||
| 7757 | if (!R.canMapToVector(IVI->getType(), DL)) | ||||
| 7758 | return false; | ||||
| 7759 | |||||
| 7760 | SmallVector<Value *, 16> BuildVectorOpds; | ||||
| 7761 | SmallVector<Value *, 16> BuildVectorInsts; | ||||
| 7762 | if (!findBuildAggregate(IVI, TTI, BuildVectorOpds, BuildVectorInsts)) | ||||
| 7763 | return false; | ||||
| 7764 | |||||
| 7765 | LLVM_DEBUG(dbgs() << "SLP: array mappable to vector: " << *IVI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: array mappable to vector: " << *IVI << "\n"; } } while (false); | ||||
| 7766 | // Aggregate value is unlikely to be processed in vector register, we need to | ||||
| 7767 | // extract scalars into scalar registers, so NeedExtraction is set true. | ||||
| 7768 | return tryToVectorizeList(BuildVectorOpds, R, /*AllowReorder=*/false, | ||||
| 7769 | BuildVectorInsts); | ||||
| 7770 | } | ||||
| 7771 | |||||
| 7772 | bool SLPVectorizerPass::vectorizeInsertElementInst(InsertElementInst *IEI, | ||||
| 7773 | BasicBlock *BB, BoUpSLP &R) { | ||||
| 7774 | SmallVector<Value *, 16> BuildVectorInsts; | ||||
| 7775 | SmallVector<Value *, 16> BuildVectorOpds; | ||||
| 7776 | SmallVector<int> Mask; | ||||
| 7777 | if (!findBuildAggregate(IEI, TTI, BuildVectorOpds, BuildVectorInsts) || | ||||
| 7778 | (llvm::all_of(BuildVectorOpds, | ||||
| 7779 | [](Value *V) { return isa<ExtractElementInst>(V); }) && | ||||
| 7780 | isShuffle(BuildVectorOpds, Mask))) | ||||
| 7781 | return false; | ||||
| 7782 | |||||
| 7783 | // Vectorize starting with the build vector operands ignoring the BuildVector | ||||
| 7784 | // instructions for the purpose of scheduling and user extraction. | ||||
| 7785 | return tryToVectorizeList(BuildVectorOpds, R, /*AllowReorder=*/false, | ||||
| 7786 | BuildVectorInsts); | ||||
| 7787 | } | ||||
| 7788 | |||||
| 7789 | bool SLPVectorizerPass::vectorizeSimpleInstructions( | ||||
| 7790 | SmallVectorImpl<Instruction *> &Instructions, BasicBlock *BB, BoUpSLP &R, | ||||
| 7791 | bool AtTerminator) { | ||||
| 7792 | bool OpsChanged = false; | ||||
| 7793 | SmallVector<Instruction *, 4> PostponedCmps; | ||||
| 7794 | for (auto *I : reverse(Instructions)) { | ||||
| 7795 | if (R.isDeleted(I)) | ||||
| 7796 | continue; | ||||
| 7797 | if (auto *LastInsertValue = dyn_cast<InsertValueInst>(I)) | ||||
| 7798 | OpsChanged |= vectorizeInsertValueInst(LastInsertValue, BB, R); | ||||
| 7799 | else if (auto *LastInsertElem = dyn_cast<InsertElementInst>(I)) | ||||
| 7800 | OpsChanged |= vectorizeInsertElementInst(LastInsertElem, BB, R); | ||||
| 7801 | else if (isa<CmpInst>(I)) | ||||
| 7802 | PostponedCmps.push_back(I); | ||||
| 7803 | } | ||||
| 7804 | if (AtTerminator) { | ||||
| 7805 | // Try to find reductions first. | ||||
| 7806 | for (Instruction *I : PostponedCmps) { | ||||
| 7807 | if (R.isDeleted(I)) | ||||
| 7808 | continue; | ||||
| 7809 | for (Value *Op : I->operands()) | ||||
| 7810 | OpsChanged |= vectorizeRootInstruction(nullptr, Op, BB, R, TTI); | ||||
| 7811 | } | ||||
| 7812 | // Try to vectorize operands as vector bundles. | ||||
| 7813 | for (Instruction *I : PostponedCmps) { | ||||
| 7814 | if (R.isDeleted(I)) | ||||
| 7815 | continue; | ||||
| 7816 | OpsChanged |= tryToVectorize(I, R); | ||||
| 7817 | } | ||||
| 7818 | Instructions.clear(); | ||||
| 7819 | } else { | ||||
| 7820 | // Insert in reverse order since the PostponedCmps vector was filled in | ||||
| 7821 | // reverse order. | ||||
| 7822 | Instructions.assign(PostponedCmps.rbegin(), PostponedCmps.rend()); | ||||
| 7823 | } | ||||
| 7824 | return OpsChanged; | ||||
| 7825 | } | ||||
| 7826 | |||||
| 7827 | bool SLPVectorizerPass::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) { | ||||
| 7828 | bool Changed = false; | ||||
| 7829 | SmallVector<Value *, 4> Incoming; | ||||
| 7830 | SmallPtrSet<Value *, 16> VisitedInstrs; | ||||
| 7831 | |||||
| 7832 | bool HaveVectorizedPhiNodes = true; | ||||
| 7833 | while (HaveVectorizedPhiNodes) { | ||||
| 7834 | HaveVectorizedPhiNodes = false; | ||||
| 7835 | |||||
| 7836 | // Collect the incoming values from the PHIs. | ||||
| 7837 | Incoming.clear(); | ||||
| 7838 | for (Instruction &I : *BB) { | ||||
| 7839 | PHINode *P = dyn_cast<PHINode>(&I); | ||||
| 7840 | if (!P) | ||||
| 7841 | break; | ||||
| 7842 | |||||
| 7843 | if (!VisitedInstrs.count(P) && !R.isDeleted(P)) | ||||
| 7844 | Incoming.push_back(P); | ||||
| 7845 | } | ||||
| 7846 | |||||
| 7847 | // Sort by type. | ||||
| 7848 | llvm::stable_sort(Incoming, PhiTypeSorterFunc); | ||||
| 7849 | |||||
| 7850 | // Try to vectorize elements base on their type. | ||||
| 7851 | for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(), | ||||
| 7852 | E = Incoming.end(); | ||||
| 7853 | IncIt != E;) { | ||||
| 7854 | |||||
| 7855 | // Look for the next elements with the same type. | ||||
| 7856 | SmallVector<Value *, 4>::iterator SameTypeIt = IncIt; | ||||
| 7857 | while (SameTypeIt != E && | ||||
| 7858 | (*SameTypeIt)->getType() == (*IncIt)->getType()) { | ||||
| 7859 | VisitedInstrs.insert(*SameTypeIt); | ||||
| 7860 | ++SameTypeIt; | ||||
| 7861 | } | ||||
| 7862 | |||||
| 7863 | // Try to vectorize them. | ||||
| 7864 | unsigned NumElts = (SameTypeIt - IncIt); | ||||
| 7865 | LLVM_DEBUG(dbgs() << "SLP: Trying to vectorize starting at PHIs ("do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n"; } } while (false) | ||||
| 7866 | << NumElts << ")\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n"; } } while (false); | ||||
| 7867 | // The order in which the phi nodes appear in the program does not matter. | ||||
| 7868 | // So allow tryToVectorizeList to reorder them if it is beneficial. This | ||||
| 7869 | // is done when there are exactly two elements since tryToVectorizeList | ||||
| 7870 | // asserts that there are only two values when AllowReorder is true. | ||||
| 7871 | bool AllowReorder = NumElts == 2; | ||||
| 7872 | if (NumElts > 1 && | ||||
| 7873 | tryToVectorizeList(makeArrayRef(IncIt, NumElts), R, AllowReorder)) { | ||||
| 7874 | // Success start over because instructions might have been changed. | ||||
| 7875 | HaveVectorizedPhiNodes = true; | ||||
| 7876 | Changed = true; | ||||
| 7877 | break; | ||||
| 7878 | } | ||||
| 7879 | |||||
| 7880 | // Start over at the next instruction of a different type (or the end). | ||||
| 7881 | IncIt = SameTypeIt; | ||||
| 7882 | } | ||||
| 7883 | } | ||||
| 7884 | |||||
| 7885 | VisitedInstrs.clear(); | ||||
| 7886 | |||||
| 7887 | SmallVector<Instruction *, 8> PostProcessInstructions; | ||||
| 7888 | SmallDenseSet<Instruction *, 4> KeyNodes; | ||||
| 7889 | for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { | ||||
| 7890 | // Skip instructions with scalable type. The num of elements is unknown at | ||||
| 7891 | // compile-time for scalable type. | ||||
| 7892 | if (isa<ScalableVectorType>(it->getType())) | ||||
| 7893 | continue; | ||||
| 7894 | |||||
| 7895 | // Skip instructions marked for the deletion. | ||||
| 7896 | if (R.isDeleted(&*it)) | ||||
| 7897 | continue; | ||||
| 7898 | // We may go through BB multiple times so skip the one we have checked. | ||||
| 7899 | if (!VisitedInstrs.insert(&*it).second) { | ||||
| 7900 | if (it->use_empty() && KeyNodes.contains(&*it) && | ||||
| 7901 | vectorizeSimpleInstructions(PostProcessInstructions, BB, R, | ||||
| 7902 | it->isTerminator())) { | ||||
| 7903 | // We would like to start over since some instructions are deleted | ||||
| 7904 | // and the iterator may become invalid value. | ||||
| 7905 | Changed = true; | ||||
| 7906 | it = BB->begin(); | ||||
| 7907 | e = BB->end(); | ||||
| 7908 | } | ||||
| 7909 | continue; | ||||
| 7910 | } | ||||
| 7911 | |||||
| 7912 | if (isa<DbgInfoIntrinsic>(it)) | ||||
| 7913 | continue; | ||||
| 7914 | |||||
| 7915 | // Try to vectorize reductions that use PHINodes. | ||||
| 7916 | if (PHINode *P = dyn_cast<PHINode>(it)) { | ||||
| 7917 | // Check that the PHI is a reduction PHI. | ||||
| 7918 | if (P->getNumIncomingValues() == 2) { | ||||
| 7919 | // Try to match and vectorize a horizontal reduction. | ||||
| 7920 | if (vectorizeRootInstruction(P, getReductionValue(DT, P, BB, LI), BB, R, | ||||
| 7921 | TTI)) { | ||||
| 7922 | Changed = true; | ||||
| 7923 | it = BB->begin(); | ||||
| 7924 | e = BB->end(); | ||||
| 7925 | continue; | ||||
| 7926 | } | ||||
| 7927 | } | ||||
| 7928 | // Try to vectorize the incoming values of the PHI, to catch reductions | ||||
| 7929 | // that feed into PHIs. | ||||
| 7930 | for (unsigned I = 0, E = P->getNumIncomingValues(); I != E; I++) { | ||||
| 7931 | // Skip if the incoming block is the current BB for now. Also, bypass | ||||
| 7932 | // unreachable IR for efficiency and to avoid crashing. | ||||
| 7933 | // TODO: Collect the skipped incoming values and try to vectorize them | ||||
| 7934 | // after processing BB. | ||||
| 7935 | if (BB == P->getIncomingBlock(I) || | ||||
| 7936 | !DT->isReachableFromEntry(P->getIncomingBlock(I))) | ||||
| 7937 | continue; | ||||
| 7938 | |||||
| 7939 | Changed |= vectorizeRootInstruction(nullptr, P->getIncomingValue(I), | ||||
| 7940 | P->getIncomingBlock(I), R, TTI); | ||||
| 7941 | } | ||||
| 7942 | continue; | ||||
| 7943 | } | ||||
| 7944 | |||||
| 7945 | // Ran into an instruction without users, like terminator, or function call | ||||
| 7946 | // with ignored return value, store. Ignore unused instructions (basing on | ||||
| 7947 | // instruction type, except for CallInst and InvokeInst). | ||||
| 7948 | if (it->use_empty() && (it->getType()->isVoidTy() || isa<CallInst>(it) || | ||||
| 7949 | isa<InvokeInst>(it))) { | ||||
| 7950 | KeyNodes.insert(&*it); | ||||
| 7951 | bool OpsChanged = false; | ||||
| 7952 | if (ShouldStartVectorizeHorAtStore || !isa<StoreInst>(it)) { | ||||
| 7953 | for (auto *V : it->operand_values()) { | ||||
| 7954 | // Try to match and vectorize a horizontal reduction. | ||||
| 7955 | OpsChanged |= vectorizeRootInstruction(nullptr, V, BB, R, TTI); | ||||
| 7956 | } | ||||
| 7957 | } | ||||
| 7958 | // Start vectorization of post-process list of instructions from the | ||||
| 7959 | // top-tree instructions to try to vectorize as many instructions as | ||||
| 7960 | // possible. | ||||
| 7961 | OpsChanged |= vectorizeSimpleInstructions(PostProcessInstructions, BB, R, | ||||
| 7962 | it->isTerminator()); | ||||
| 7963 | if (OpsChanged) { | ||||
| 7964 | // We would like to start over since some instructions are deleted | ||||
| 7965 | // and the iterator may become invalid value. | ||||
| 7966 | Changed = true; | ||||
| 7967 | it = BB->begin(); | ||||
| 7968 | e = BB->end(); | ||||
| 7969 | continue; | ||||
| 7970 | } | ||||
| 7971 | } | ||||
| 7972 | |||||
| 7973 | if (isa<InsertElementInst>(it) || isa<CmpInst>(it) || | ||||
| 7974 | isa<InsertValueInst>(it)) | ||||
| 7975 | PostProcessInstructions.push_back(&*it); | ||||
| 7976 | } | ||||
| 7977 | |||||
| 7978 | return Changed; | ||||
| 7979 | } | ||||
| 7980 | |||||
| 7981 | bool SLPVectorizerPass::vectorizeGEPIndices(BasicBlock *BB, BoUpSLP &R) { | ||||
| 7982 | auto Changed = false; | ||||
| 7983 | for (auto &Entry : GEPs) { | ||||
| 7984 | // If the getelementptr list has fewer than two elements, there's nothing | ||||
| 7985 | // to do. | ||||
| 7986 | if (Entry.second.size() < 2) | ||||
| 7987 | continue; | ||||
| 7988 | |||||
| 7989 | LLVM_DEBUG(dbgs() << "SLP: Analyzing a getelementptr list of length "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a getelementptr list of length " << Entry.second.size() << ".\n"; } } while (false ) | ||||
| 7990 | << Entry.second.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a getelementptr list of length " << Entry.second.size() << ".\n"; } } while (false ); | ||||
| 7991 | |||||
| 7992 | // Process the GEP list in chunks suitable for the target's supported | ||||
| 7993 | // vector size. If a vector register can't hold 1 element, we are done. We | ||||
| 7994 | // are trying to vectorize the index computations, so the maximum number of | ||||
| 7995 | // elements is based on the size of the index expression, rather than the | ||||
| 7996 | // size of the GEP itself (the target's pointer size). | ||||
| 7997 | unsigned MaxVecRegSize = R.getMaxVecRegSize(); | ||||
| 7998 | unsigned EltSize = R.getVectorElementSize(*Entry.second[0]->idx_begin()); | ||||
| 7999 | if (MaxVecRegSize < EltSize) | ||||
| 8000 | continue; | ||||
| 8001 | |||||
| 8002 | unsigned MaxElts = MaxVecRegSize / EltSize; | ||||
| 8003 | for (unsigned BI = 0, BE = Entry.second.size(); BI < BE; BI += MaxElts) { | ||||
| 8004 | auto Len = std::min<unsigned>(BE - BI, MaxElts); | ||||
| 8005 | ArrayRef<GetElementPtrInst *> GEPList(&Entry.second[BI], Len); | ||||
| 8006 | |||||
| 8007 | // Initialize a set a candidate getelementptrs. Note that we use a | ||||
| 8008 | // SetVector here to preserve program order. If the index computations | ||||
| 8009 | // are vectorizable and begin with loads, we want to minimize the chance | ||||
| 8010 | // of having to reorder them later. | ||||
| 8011 | SetVector<Value *> Candidates(GEPList.begin(), GEPList.end()); | ||||
| 8012 | |||||
| 8013 | // Some of the candidates may have already been vectorized after we | ||||
| 8014 | // initially collected them. If so, they are marked as deleted, so remove | ||||
| 8015 | // them from the set of candidates. | ||||
| 8016 | Candidates.remove_if( | ||||
| 8017 | [&R](Value *I) { return R.isDeleted(cast<Instruction>(I)); }); | ||||
| 8018 | |||||
| 8019 | // Remove from the set of candidates all pairs of getelementptrs with | ||||
| 8020 | // constant differences. Such getelementptrs are likely not good | ||||
| 8021 | // candidates for vectorization in a bottom-up phase since one can be | ||||
| 8022 | // computed from the other. We also ensure all candidate getelementptr | ||||
| 8023 | // indices are unique. | ||||
| 8024 | for (int I = 0, E = GEPList.size(); I < E && Candidates.size() > 1; ++I) { | ||||
| 8025 | auto *GEPI = GEPList[I]; | ||||
| 8026 | if (!Candidates.count(GEPI)) | ||||
| 8027 | continue; | ||||
| 8028 | auto *SCEVI = SE->getSCEV(GEPList[I]); | ||||
| 8029 | for (int J = I + 1; J < E && Candidates.size() > 1; ++J) { | ||||
| 8030 | auto *GEPJ = GEPList[J]; | ||||
| 8031 | auto *SCEVJ = SE->getSCEV(GEPList[J]); | ||||
| 8032 | if (isa<SCEVConstant>(SE->getMinusSCEV(SCEVI, SCEVJ))) { | ||||
| 8033 | Candidates.remove(GEPI); | ||||
| 8034 | Candidates.remove(GEPJ); | ||||
| 8035 | } else if (GEPI->idx_begin()->get() == GEPJ->idx_begin()->get()) { | ||||
| 8036 | Candidates.remove(GEPJ); | ||||
| 8037 | } | ||||
| 8038 | } | ||||
| 8039 | } | ||||
| 8040 | |||||
| 8041 | // We break out of the above computation as soon as we know there are | ||||
| 8042 | // fewer than two candidates remaining. | ||||
| 8043 | if (Candidates.size() < 2) | ||||
| 8044 | continue; | ||||
| 8045 | |||||
| 8046 | // Add the single, non-constant index of each candidate to the bundle. We | ||||
| 8047 | // ensured the indices met these constraints when we originally collected | ||||
| 8048 | // the getelementptrs. | ||||
| 8049 | SmallVector<Value *, 16> Bundle(Candidates.size()); | ||||
| 8050 | auto BundleIndex = 0u; | ||||
| 8051 | for (auto *V : Candidates) { | ||||
| 8052 | auto *GEP = cast<GetElementPtrInst>(V); | ||||
| 8053 | auto *GEPIdx = GEP->idx_begin()->get(); | ||||
| 8054 | assert(GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx))(static_cast <bool> (GEP->getNumIndices() == 1 || !isa <Constant>(GEPIdx)) ? void (0) : __assert_fail ("GEP->getNumIndices() == 1 || !isa<Constant>(GEPIdx)" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp" , 8054, __extension__ __PRETTY_FUNCTION__)); | ||||
| 8055 | Bundle[BundleIndex++] = GEPIdx; | ||||
| 8056 | } | ||||
| 8057 | |||||
| 8058 | // Try and vectorize the indices. We are currently only interested in | ||||
| 8059 | // gather-like cases of the form: | ||||
| 8060 | // | ||||
| 8061 | // ... = g[a[0] - b[0]] + g[a[1] - b[1]] + ... | ||||
| 8062 | // | ||||
| 8063 | // where the loads of "a", the loads of "b", and the subtractions can be | ||||
| 8064 | // performed in parallel. It's likely that detecting this pattern in a | ||||
| 8065 | // bottom-up phase will be simpler and less costly than building a | ||||
| 8066 | // full-blown top-down phase beginning at the consecutive loads. | ||||
| 8067 | Changed |= tryToVectorizeList(Bundle, R); | ||||
| 8068 | } | ||||
| 8069 | } | ||||
| 8070 | return Changed; | ||||
| 8071 | } | ||||
| 8072 | |||||
| 8073 | bool SLPVectorizerPass::vectorizeStoreChains(BoUpSLP &R) { | ||||
| 8074 | bool Changed = false; | ||||
| 8075 | // Attempt to sort and vectorize each of the store-groups. | ||||
| 8076 | for (StoreListMap::iterator it = Stores.begin(), e = Stores.end(); it != e; | ||||
| 8077 | ++it) { | ||||
| 8078 | if (it->second.size() < 2) | ||||
| 8079 | continue; | ||||
| 8080 | |||||
| 8081 | LLVM_DEBUG(dbgs() << "SLP: Analyzing a store chain of length "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << it->second.size() << ".\n"; } } while (false ) | ||||
| 8082 | << it->second.size() << ".\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("SLP")) { dbgs() << "SLP: Analyzing a store chain of length " << it->second.size() << ".\n"; } } while (false ); | ||||
| 8083 | |||||
| 8084 | Changed |= vectorizeStores(it->second, R); | ||||
| 8085 | } | ||||
| 8086 | return Changed; | ||||
| 8087 | } | ||||
| 8088 | |||||
| 8089 | char SLPVectorizer::ID = 0; | ||||
| 8090 | |||||
| 8091 | static const char lv_name[] = "SLP Vectorizer"; | ||||
| 8092 | |||||
| 8093 | INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)static void *initializeSLPVectorizerPassOnce(PassRegistry & Registry) { | ||||
| 8094 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); | ||||
| 8095 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | ||||
| 8096 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry); | ||||
| 8097 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)initializeScalarEvolutionWrapperPassPass(Registry); | ||||
| 8098 | INITIALIZE_PASS_DEPENDENCY(LoopSimplify)initializeLoopSimplifyPass(Registry); | ||||
| 8099 | INITIALIZE_PASS_DEPENDENCY(DemandedBitsWrapperPass)initializeDemandedBitsWrapperPassPass(Registry); | ||||
| 8100 | INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)initializeOptimizationRemarkEmitterWrapperPassPass(Registry); | ||||
| 8101 | INITIALIZE_PASS_DEPENDENCY(InjectTLIMappingsLegacy)initializeInjectTLIMappingsLegacyPass(Registry); | ||||
| 8102 | INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)PassInfo *PI = new PassInfo( lv_name, "slp-vectorizer", & SLPVectorizer::ID, PassInfo::NormalCtor_t(callDefaultCtor< SLPVectorizer>), false, false); Registry.registerPass(*PI, true); return PI; } static llvm::once_flag InitializeSLPVectorizerPassFlag ; void llvm::initializeSLPVectorizerPass(PassRegistry &Registry ) { llvm::call_once(InitializeSLPVectorizerPassFlag, initializeSLPVectorizerPassOnce , std::ref(Registry)); } | ||||
| 8103 | |||||
| 8104 | Pass *llvm::createSLPVectorizerPass() { return new SLPVectorizer(); } |
| 1 | //===- llvm/ADT/SmallVector.h - 'Normally small' vectors --------*- 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 | // This file defines the SmallVector class. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #ifndef LLVM_ADT_SMALLVECTOR_H |
| 14 | #define LLVM_ADT_SMALLVECTOR_H |
| 15 | |
| 16 | #include "llvm/ADT/iterator_range.h" |
| 17 | #include "llvm/Support/Compiler.h" |
| 18 | #include "llvm/Support/ErrorHandling.h" |
| 19 | #include "llvm/Support/MathExtras.h" |
| 20 | #include "llvm/Support/MemAlloc.h" |
| 21 | #include "llvm/Support/type_traits.h" |
| 22 | #include <algorithm> |
| 23 | #include <cassert> |
| 24 | #include <cstddef> |
| 25 | #include <cstdlib> |
| 26 | #include <cstring> |
| 27 | #include <initializer_list> |
| 28 | #include <iterator> |
| 29 | #include <limits> |
| 30 | #include <memory> |
| 31 | #include <new> |
| 32 | #include <type_traits> |
| 33 | #include <utility> |
| 34 | |
| 35 | namespace llvm { |
| 36 | |
| 37 | /// This is all the stuff common to all SmallVectors. |
| 38 | /// |
| 39 | /// The template parameter specifies the type which should be used to hold the |
| 40 | /// Size and Capacity of the SmallVector, so it can be adjusted. |
| 41 | /// Using 32 bit size is desirable to shrink the size of the SmallVector. |
| 42 | /// Using 64 bit size is desirable for cases like SmallVector<char>, where a |
| 43 | /// 32 bit size would limit the vector to ~4GB. SmallVectors are used for |
| 44 | /// buffering bitcode output - which can exceed 4GB. |
| 45 | template <class Size_T> class SmallVectorBase { |
| 46 | protected: |
| 47 | void *BeginX; |
| 48 | Size_T Size = 0, Capacity; |
| 49 | |
| 50 | /// The maximum value of the Size_T used. |
| 51 | static constexpr size_t SizeTypeMax() { |
| 52 | return std::numeric_limits<Size_T>::max(); |
| 53 | } |
| 54 | |
| 55 | SmallVectorBase() = delete; |
| 56 | SmallVectorBase(void *FirstEl, size_t TotalCapacity) |
| 57 | : BeginX(FirstEl), Capacity(TotalCapacity) {} |
| 58 | |
| 59 | /// This is a helper for \a grow() that's out of line to reduce code |
| 60 | /// duplication. This function will report a fatal error if it can't grow at |
| 61 | /// least to \p MinSize. |
| 62 | void *mallocForGrow(size_t MinSize, size_t TSize, size_t &NewCapacity); |
| 63 | |
| 64 | /// This is an implementation of the grow() method which only works |
| 65 | /// on POD-like data types and is out of line to reduce code duplication. |
| 66 | /// This function will report a fatal error if it cannot increase capacity. |
| 67 | void grow_pod(void *FirstEl, size_t MinSize, size_t TSize); |
| 68 | |
| 69 | public: |
| 70 | size_t size() const { return Size; } |
| 71 | size_t capacity() const { return Capacity; } |
| 72 | |
| 73 | LLVM_NODISCARD[[clang::warn_unused_result]] bool empty() const { return !Size; } |
| 74 | |
| 75 | /// Set the array size to \p N, which the current array must have enough |
| 76 | /// capacity for. |
| 77 | /// |
| 78 | /// This does not construct or destroy any elements in the vector. |
| 79 | /// |
| 80 | /// Clients can use this in conjunction with capacity() to write past the end |
| 81 | /// of the buffer when they know that more elements are available, and only |
| 82 | /// update the size later. This avoids the cost of value initializing elements |
| 83 | /// which will only be overwritten. |
| 84 | void set_size(size_t N) { |
| 85 | assert(N <= capacity())(static_cast <bool> (N <= capacity()) ? void (0) : __assert_fail ("N <= capacity()", "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/include/llvm/ADT/SmallVector.h" , 85, __extension__ __PRETTY_FUNCTION__)); |
| 86 | Size = N; |
| 87 | } |
| 88 | }; |
| 89 | |
| 90 | template <class T> |
| 91 | using SmallVectorSizeType = |
| 92 | typename std::conditional<sizeof(T) < 4 && sizeof(void *) >= 8, uint64_t, |
| 93 | uint32_t>::type; |
| 94 | |
| 95 | /// Figure out the offset of the first element. |
| 96 | template <class T, typename = void> struct SmallVectorAlignmentAndSize { |
| 97 | alignas(SmallVectorBase<SmallVectorSizeType<T>>) char Base[sizeof( |
| 98 | SmallVectorBase<SmallVectorSizeType<T>>)]; |
| 99 | alignas(T) char FirstEl[sizeof(T)]; |
| 100 | }; |
| 101 | |
| 102 | /// This is the part of SmallVectorTemplateBase which does not depend on whether |
| 103 | /// the type T is a POD. The extra dummy template argument is used by ArrayRef |
| 104 | /// to avoid unnecessarily requiring T to be complete. |
| 105 | template <typename T, typename = void> |
| 106 | class SmallVectorTemplateCommon |
| 107 | : public SmallVectorBase<SmallVectorSizeType<T>> { |
| 108 | using Base = SmallVectorBase<SmallVectorSizeType<T>>; |
| 109 | |
| 110 | /// Find the address of the first element. For this pointer math to be valid |
| 111 | /// with small-size of 0 for T with lots of alignment, it's important that |
| 112 | /// SmallVectorStorage is properly-aligned even for small-size of 0. |
| 113 | void *getFirstEl() const { |
| 114 | return const_cast<void *>(reinterpret_cast<const void *>( |
| 115 | reinterpret_cast<const char *>(this) + |
| 116 | offsetof(SmallVectorAlignmentAndSize<T>, FirstEl)__builtin_offsetof(SmallVectorAlignmentAndSize<T>, FirstEl ))); |
| 117 | } |
| 118 | // Space after 'FirstEl' is clobbered, do not add any instance vars after it. |
| 119 | |
| 120 | protected: |
| 121 | SmallVectorTemplateCommon(size_t Size) : Base(getFirstEl(), Size) {} |
| 122 | |
| 123 | void grow_pod(size_t MinSize, size_t TSize) { |
| 124 | Base::grow_pod(getFirstEl(), MinSize, TSize); |
| 125 | } |
| 126 | |
| 127 | /// Return true if this is a smallvector which has not had dynamic |
| 128 | /// memory allocated for it. |
| 129 | bool isSmall() const { return this->BeginX == getFirstEl(); } |
| 130 | |
| 131 | /// Put this vector in a state of being small. |
| 132 | void resetToSmall() { |
| 133 | this->BeginX = getFirstEl(); |
| 134 | this->Size = this->Capacity = 0; // FIXME: Setting Capacity to 0 is suspect. |
| 135 | } |
| 136 | |
| 137 | /// Return true if V is an internal reference to the given range. |
| 138 | bool isReferenceToRange(const void *V, const void *First, const void *Last) const { |
| 139 | // Use std::less to avoid UB. |
| 140 | std::less<> LessThan; |
| 141 | return !LessThan(V, First) && LessThan(V, Last); |
| 142 | } |
| 143 | |
| 144 | /// Return true if V is an internal reference to this vector. |
| 145 | bool isReferenceToStorage(const void *V) const { |
| 146 | return isReferenceToRange(V, this->begin(), this->end()); |
| 147 | } |
| 148 | |
| 149 | /// Return true if First and Last form a valid (possibly empty) range in this |
| 150 | /// vector's storage. |
| 151 | bool isRangeInStorage(const void *First, const void *Last) const { |
| 152 | // Use std::less to avoid UB. |
| 153 | std::less<> LessThan; |
| 154 | return !LessThan(First, this->begin()) && !LessThan(Last, First) && |
| 155 | !LessThan(this->end(), Last); |
| 156 | } |
| 157 | |
| 158 | /// Return true unless Elt will be invalidated by resizing the vector to |
| 159 | /// NewSize. |
| 160 | bool isSafeToReferenceAfterResize(const void *Elt, size_t NewSize) { |
| 161 | // Past the end. |
| 162 | if (LLVM_LIKELY(!isReferenceToStorage(Elt))__builtin_expect((bool)(!isReferenceToStorage(Elt)), true)) |
| 163 | return true; |
| 164 | |
| 165 | // Return false if Elt will be destroyed by shrinking. |
| 166 | if (NewSize <= this->size()) |
| 167 | return Elt < this->begin() + NewSize; |
| 168 | |
| 169 | // Return false if we need to grow. |
| 170 | return NewSize <= this->capacity(); |
| 171 | } |
| 172 | |
| 173 | /// Check whether Elt will be invalidated by resizing the vector to NewSize. |
| 174 | void assertSafeToReferenceAfterResize(const void *Elt, size_t NewSize) { |
| 175 | assert(isSafeToReferenceAfterResize(Elt, NewSize) &&(static_cast <bool> (isSafeToReferenceAfterResize(Elt, NewSize ) && "Attempting to reference an element of the vector in an operation " "that invalidates it") ? void (0) : __assert_fail ("isSafeToReferenceAfterResize(Elt, NewSize) && \"Attempting to reference an element of the vector in an operation \" \"that invalidates it\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/include/llvm/ADT/SmallVector.h" , 177, __extension__ __PRETTY_FUNCTION__)) |
| 176 | "Attempting to reference an element of the vector in an operation "(static_cast <bool> (isSafeToReferenceAfterResize(Elt, NewSize ) && "Attempting to reference an element of the vector in an operation " "that invalidates it") ? void (0) : __assert_fail ("isSafeToReferenceAfterResize(Elt, NewSize) && \"Attempting to reference an element of the vector in an operation \" \"that invalidates it\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/include/llvm/ADT/SmallVector.h" , 177, __extension__ __PRETTY_FUNCTION__)) |
| 177 | "that invalidates it")(static_cast <bool> (isSafeToReferenceAfterResize(Elt, NewSize ) && "Attempting to reference an element of the vector in an operation " "that invalidates it") ? void (0) : __assert_fail ("isSafeToReferenceAfterResize(Elt, NewSize) && \"Attempting to reference an element of the vector in an operation \" \"that invalidates it\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/include/llvm/ADT/SmallVector.h" , 177, __extension__ __PRETTY_FUNCTION__)); |
| 178 | } |
| 179 | |
| 180 | /// Check whether Elt will be invalidated by increasing the size of the |
| 181 | /// vector by N. |
| 182 | void assertSafeToAdd(const void *Elt, size_t N = 1) { |
| 183 | this->assertSafeToReferenceAfterResize(Elt, this->size() + N); |
| 184 | } |
| 185 | |
| 186 | /// Check whether any part of the range will be invalidated by clearing. |
| 187 | void assertSafeToReferenceAfterClear(const T *From, const T *To) { |
| 188 | if (From == To) |
| 189 | return; |
| 190 | this->assertSafeToReferenceAfterResize(From, 0); |
| 191 | this->assertSafeToReferenceAfterResize(To - 1, 0); |
| 192 | } |
| 193 | template < |
| 194 | class ItTy, |
| 195 | std::enable_if_t<!std::is_same<std::remove_const_t<ItTy>, T *>::value, |
| 196 | bool> = false> |
| 197 | void assertSafeToReferenceAfterClear(ItTy, ItTy) {} |
| 198 | |
| 199 | /// Check whether any part of the range will be invalidated by growing. |
| 200 | void assertSafeToAddRange(const T *From, const T *To) { |
| 201 | if (From == To) |
| 202 | return; |
| 203 | this->assertSafeToAdd(From, To - From); |
| 204 | this->assertSafeToAdd(To - 1, To - From); |
| 205 | } |
| 206 | template < |
| 207 | class ItTy, |
| 208 | std::enable_if_t<!std::is_same<std::remove_const_t<ItTy>, T *>::value, |
| 209 | bool> = false> |
| 210 | void assertSafeToAddRange(ItTy, ItTy) {} |
| 211 | |
| 212 | /// Reserve enough space to add one element, and return the updated element |
| 213 | /// pointer in case it was a reference to the storage. |
| 214 | template <class U> |
| 215 | static const T *reserveForParamAndGetAddressImpl(U *This, const T &Elt, |
| 216 | size_t N) { |
| 217 | size_t NewSize = This->size() + N; |
| 218 | if (LLVM_LIKELY(NewSize <= This->capacity())__builtin_expect((bool)(NewSize <= This->capacity()), true )) |
| 219 | return &Elt; |
| 220 | |
| 221 | bool ReferencesStorage = false; |
| 222 | int64_t Index = -1; |
| 223 | if (!U::TakesParamByValue) { |
| 224 | if (LLVM_UNLIKELY(This->isReferenceToStorage(&Elt))__builtin_expect((bool)(This->isReferenceToStorage(&Elt )), false)) { |
| 225 | ReferencesStorage = true; |
| 226 | Index = &Elt - This->begin(); |
| 227 | } |
| 228 | } |
| 229 | This->grow(NewSize); |
| 230 | return ReferencesStorage ? This->begin() + Index : &Elt; |
| 231 | } |
| 232 | |
| 233 | public: |
| 234 | using size_type = size_t; |
| 235 | using difference_type = ptrdiff_t; |
| 236 | using value_type = T; |
| 237 | using iterator = T *; |
| 238 | using const_iterator = const T *; |
| 239 | |
| 240 | using const_reverse_iterator = std::reverse_iterator<const_iterator>; |
| 241 | using reverse_iterator = std::reverse_iterator<iterator>; |
| 242 | |
| 243 | using reference = T &; |
| 244 | using const_reference = const T &; |
| 245 | using pointer = T *; |
| 246 | using const_pointer = const T *; |
| 247 | |
| 248 | using Base::capacity; |
| 249 | using Base::empty; |
| 250 | using Base::size; |
| 251 | |
| 252 | // forward iterator creation methods. |
| 253 | iterator begin() { return (iterator)this->BeginX; } |
| 254 | const_iterator begin() const { return (const_iterator)this->BeginX; } |
| 255 | iterator end() { return begin() + size(); } |
| 256 | const_iterator end() const { return begin() + size(); } |
| 257 | |
| 258 | // reverse iterator creation methods. |
| 259 | reverse_iterator rbegin() { return reverse_iterator(end()); } |
| 260 | const_reverse_iterator rbegin() const{ return const_reverse_iterator(end()); } |
| 261 | reverse_iterator rend() { return reverse_iterator(begin()); } |
| 262 | const_reverse_iterator rend() const { return const_reverse_iterator(begin());} |
| 263 | |
| 264 | size_type size_in_bytes() const { return size() * sizeof(T); } |
| 265 | size_type max_size() const { |
| 266 | return std::min(this->SizeTypeMax(), size_type(-1) / sizeof(T)); |
| 267 | } |
| 268 | |
| 269 | size_t capacity_in_bytes() const { return capacity() * sizeof(T); } |
| 270 | |
| 271 | /// Return a pointer to the vector's buffer, even if empty(). |
| 272 | pointer data() { return pointer(begin()); } |
| 273 | /// Return a pointer to the vector's buffer, even if empty(). |
| 274 | const_pointer data() const { return const_pointer(begin()); } |
| 275 | |
| 276 | reference operator[](size_type idx) { |
| 277 | assert(idx < size())(static_cast <bool> (idx < size()) ? void (0) : __assert_fail ("idx < size()", "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/include/llvm/ADT/SmallVector.h" , 277, __extension__ __PRETTY_FUNCTION__)); |
| 278 | return begin()[idx]; |
| 279 | } |
| 280 | const_reference operator[](size_type idx) const { |
| 281 | assert(idx < size())(static_cast <bool> (idx < size()) ? void (0) : __assert_fail ("idx < size()", "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/include/llvm/ADT/SmallVector.h" , 281, __extension__ __PRETTY_FUNCTION__)); |
| 282 | return begin()[idx]; |
| 283 | } |
| 284 | |
| 285 | reference front() { |
| 286 | assert(!empty())(static_cast <bool> (!empty()) ? void (0) : __assert_fail ("!empty()", "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/include/llvm/ADT/SmallVector.h" , 286, __extension__ __PRETTY_FUNCTION__)); |
| 287 | return begin()[0]; |
| 288 | } |
| 289 | const_reference front() const { |
| 290 | assert(!empty())(static_cast <bool> (!empty()) ? void (0) : __assert_fail ("!empty()", "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/include/llvm/ADT/SmallVector.h" , 290, __extension__ __PRETTY_FUNCTION__)); |
| 291 | return begin()[0]; |
| 292 | } |
| 293 | |
| 294 | reference back() { |
| 295 | assert(!empty())(static_cast <bool> (!empty()) ? void (0) : __assert_fail ("!empty()", "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/include/llvm/ADT/SmallVector.h" , 295, __extension__ __PRETTY_FUNCTION__)); |
| 296 | return end()[-1]; |
| 297 | } |
| 298 | const_reference back() const { |
| 299 | assert(!empty())(static_cast <bool> (!empty()) ? void (0) : __assert_fail ("!empty()", "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/include/llvm/ADT/SmallVector.h" , 299, __extension__ __PRETTY_FUNCTION__)); |
| 300 | return end()[-1]; |
| 301 | } |
| 302 | }; |
| 303 | |
| 304 | /// SmallVectorTemplateBase<TriviallyCopyable = false> - This is where we put |
| 305 | /// method implementations that are designed to work with non-trivial T's. |
| 306 | /// |
| 307 | /// We approximate is_trivially_copyable with trivial move/copy construction and |
| 308 | /// trivial destruction. While the standard doesn't specify that you're allowed |
| 309 | /// copy these types with memcpy, there is no way for the type to observe this. |
| 310 | /// This catches the important case of std::pair<POD, POD>, which is not |
| 311 | /// trivially assignable. |
| 312 | template <typename T, bool = (is_trivially_copy_constructible<T>::value) && |
| 313 | (is_trivially_move_constructible<T>::value) && |
| 314 | std::is_trivially_destructible<T>::value> |
| 315 | class SmallVectorTemplateBase : public SmallVectorTemplateCommon<T> { |
| 316 | friend class SmallVectorTemplateCommon<T>; |
| 317 | |
| 318 | protected: |
| 319 | static constexpr bool TakesParamByValue = false; |
| 320 | using ValueParamT = const T &; |
| 321 | |
| 322 | SmallVectorTemplateBase(size_t Size) : SmallVectorTemplateCommon<T>(Size) {} |
| 323 | |
| 324 | static void destroy_range(T *S, T *E) { |
| 325 | while (S != E) { |
| 326 | --E; |
| 327 | E->~T(); |
| 328 | } |
| 329 | } |
| 330 | |
| 331 | /// Move the range [I, E) into the uninitialized memory starting with "Dest", |
| 332 | /// constructing elements as needed. |
| 333 | template<typename It1, typename It2> |
| 334 | static void uninitialized_move(It1 I, It1 E, It2 Dest) { |
| 335 | std::uninitialized_copy(std::make_move_iterator(I), |
| 336 | std::make_move_iterator(E), Dest); |
| 337 | } |
| 338 | |
| 339 | /// Copy the range [I, E) onto the uninitialized memory starting with "Dest", |
| 340 | /// constructing elements as needed. |
| 341 | template<typename It1, typename It2> |
| 342 | static void uninitialized_copy(It1 I, It1 E, It2 Dest) { |
| 343 | std::uninitialized_copy(I, E, Dest); |
| 344 | } |
| 345 | |
| 346 | /// Grow the allocated memory (without initializing new elements), doubling |
| 347 | /// the size of the allocated memory. Guarantees space for at least one more |
| 348 | /// element, or MinSize more elements if specified. |
| 349 | void grow(size_t MinSize = 0); |
| 350 | |
| 351 | /// Create a new allocation big enough for \p MinSize and pass back its size |
| 352 | /// in \p NewCapacity. This is the first section of \a grow(). |
| 353 | T *mallocForGrow(size_t MinSize, size_t &NewCapacity) { |
| 354 | return static_cast<T *>( |
| 355 | SmallVectorBase<SmallVectorSizeType<T>>::mallocForGrow( |
| 356 | MinSize, sizeof(T), NewCapacity)); |
| 357 | } |
| 358 | |
| 359 | /// Move existing elements over to the new allocation \p NewElts, the middle |
| 360 | /// section of \a grow(). |
| 361 | void moveElementsForGrow(T *NewElts); |
| 362 | |
| 363 | /// Transfer ownership of the allocation, finishing up \a grow(). |
| 364 | void takeAllocationForGrow(T *NewElts, size_t NewCapacity); |
| 365 | |
| 366 | /// Reserve enough space to add one element, and return the updated element |
| 367 | /// pointer in case it was a reference to the storage. |
| 368 | const T *reserveForParamAndGetAddress(const T &Elt, size_t N = 1) { |
| 369 | return this->reserveForParamAndGetAddressImpl(this, Elt, N); |
| 370 | } |
| 371 | |
| 372 | /// Reserve enough space to add one element, and return the updated element |
| 373 | /// pointer in case it was a reference to the storage. |
| 374 | T *reserveForParamAndGetAddress(T &Elt, size_t N = 1) { |
| 375 | return const_cast<T *>( |
| 376 | this->reserveForParamAndGetAddressImpl(this, Elt, N)); |
| 377 | } |
| 378 | |
| 379 | static T &&forward_value_param(T &&V) { return std::move(V); } |
| 380 | static const T &forward_value_param(const T &V) { return V; } |
| 381 | |
| 382 | void growAndAssign(size_t NumElts, const T &Elt) { |
| 383 | // Grow manually in case Elt is an internal reference. |
| 384 | size_t NewCapacity; |
| 385 | T *NewElts = mallocForGrow(NumElts, NewCapacity); |
| 386 | std::uninitialized_fill_n(NewElts, NumElts, Elt); |
| 387 | this->destroy_range(this->begin(), this->end()); |
| 388 | takeAllocationForGrow(NewElts, NewCapacity); |
| 389 | this->set_size(NumElts); |
| 390 | } |
| 391 | |
| 392 | template <typename... ArgTypes> T &growAndEmplaceBack(ArgTypes &&... Args) { |
| 393 | // Grow manually in case one of Args is an internal reference. |
| 394 | size_t NewCapacity; |
| 395 | T *NewElts = mallocForGrow(0, NewCapacity); |
| 396 | ::new ((void *)(NewElts + this->size())) T(std::forward<ArgTypes>(Args)...); |
| 397 | moveElementsForGrow(NewElts); |
| 398 | takeAllocationForGrow(NewElts, NewCapacity); |
| 399 | this->set_size(this->size() + 1); |
| 400 | return this->back(); |
| 401 | } |
| 402 | |
| 403 | public: |
| 404 | void push_back(const T &Elt) { |
| 405 | const T *EltPtr = reserveForParamAndGetAddress(Elt); |
| 406 | ::new ((void *)this->end()) T(*EltPtr); |
| 407 | this->set_size(this->size() + 1); |
| 408 | } |
| 409 | |
| 410 | void push_back(T &&Elt) { |
| 411 | T *EltPtr = reserveForParamAndGetAddress(Elt); |
| 412 | ::new ((void *)this->end()) T(::std::move(*EltPtr)); |
| 413 | this->set_size(this->size() + 1); |
| 414 | } |
| 415 | |
| 416 | void pop_back() { |
| 417 | this->set_size(this->size() - 1); |
| 418 | this->end()->~T(); |
| 419 | } |
| 420 | }; |
| 421 | |
| 422 | // Define this out-of-line to dissuade the C++ compiler from inlining it. |
| 423 | template <typename T, bool TriviallyCopyable> |
| 424 | void SmallVectorTemplateBase<T, TriviallyCopyable>::grow(size_t MinSize) { |
| 425 | size_t NewCapacity; |
| 426 | T *NewElts = mallocForGrow(MinSize, NewCapacity); |
| 427 | moveElementsForGrow(NewElts); |
| 428 | takeAllocationForGrow(NewElts, NewCapacity); |
| 429 | } |
| 430 | |
| 431 | // Define this out-of-line to dissuade the C++ compiler from inlining it. |
| 432 | template <typename T, bool TriviallyCopyable> |
| 433 | void SmallVectorTemplateBase<T, TriviallyCopyable>::moveElementsForGrow( |
| 434 | T *NewElts) { |
| 435 | // Move the elements over. |
| 436 | this->uninitialized_move(this->begin(), this->end(), NewElts); |
| 437 | |
| 438 | // Destroy the original elements. |
| 439 | destroy_range(this->begin(), this->end()); |
| 440 | } |
| 441 | |
| 442 | // Define this out-of-line to dissuade the C++ compiler from inlining it. |
| 443 | template <typename T, bool TriviallyCopyable> |
| 444 | void SmallVectorTemplateBase<T, TriviallyCopyable>::takeAllocationForGrow( |
| 445 | T *NewElts, size_t NewCapacity) { |
| 446 | // If this wasn't grown from the inline copy, deallocate the old space. |
| 447 | if (!this->isSmall()) |
| 448 | free(this->begin()); |
| 449 | |
| 450 | this->BeginX = NewElts; |
| 451 | this->Capacity = NewCapacity; |
| 452 | } |
| 453 | |
| 454 | /// SmallVectorTemplateBase<TriviallyCopyable = true> - This is where we put |
| 455 | /// method implementations that are designed to work with trivially copyable |
| 456 | /// T's. This allows using memcpy in place of copy/move construction and |
| 457 | /// skipping destruction. |
| 458 | template <typename T> |
| 459 | class SmallVectorTemplateBase<T, true> : public SmallVectorTemplateCommon<T> { |
| 460 | friend class SmallVectorTemplateCommon<T>; |
| 461 | |
| 462 | protected: |
| 463 | /// True if it's cheap enough to take parameters by value. Doing so avoids |
| 464 | /// overhead related to mitigations for reference invalidation. |
| 465 | static constexpr bool TakesParamByValue = sizeof(T) <= 2 * sizeof(void *); |
| 466 | |
| 467 | /// Either const T& or T, depending on whether it's cheap enough to take |
| 468 | /// parameters by value. |
| 469 | using ValueParamT = |
| 470 | typename std::conditional<TakesParamByValue, T, const T &>::type; |
| 471 | |
| 472 | SmallVectorTemplateBase(size_t Size) : SmallVectorTemplateCommon<T>(Size) {} |
| 473 | |
| 474 | // No need to do a destroy loop for POD's. |
| 475 | static void destroy_range(T *, T *) {} |
| 476 | |
| 477 | /// Move the range [I, E) onto the uninitialized memory |
| 478 | /// starting with "Dest", constructing elements into it as needed. |
| 479 | template<typename It1, typename It2> |
| 480 | static void uninitialized_move(It1 I, It1 E, It2 Dest) { |
| 481 | // Just do a copy. |
| 482 | uninitialized_copy(I, E, Dest); |
| 483 | } |
| 484 | |
| 485 | /// Copy the range [I, E) onto the uninitialized memory |
| 486 | /// starting with "Dest", constructing elements into it as needed. |
| 487 | template<typename It1, typename It2> |
| 488 | static void uninitialized_copy(It1 I, It1 E, It2 Dest) { |
| 489 | // Arbitrary iterator types; just use the basic implementation. |
| 490 | std::uninitialized_copy(I, E, Dest); |
| 491 | } |
| 492 | |
| 493 | /// Copy the range [I, E) onto the uninitialized memory |
| 494 | /// starting with "Dest", constructing elements into it as needed. |
| 495 | template <typename T1, typename T2> |
| 496 | static void uninitialized_copy( |
| 497 | T1 *I, T1 *E, T2 *Dest, |
| 498 | std::enable_if_t<std::is_same<typename std::remove_const<T1>::type, |
| 499 | T2>::value> * = nullptr) { |
| 500 | // Use memcpy for PODs iterated by pointers (which includes SmallVector |
| 501 | // iterators): std::uninitialized_copy optimizes to memmove, but we can |
| 502 | // use memcpy here. Note that I and E are iterators and thus might be |
| 503 | // invalid for memcpy if they are equal. |
| 504 | if (I != E) |
| 505 | memcpy(reinterpret_cast<void *>(Dest), I, (E - I) * sizeof(T)); |
| 506 | } |
| 507 | |
| 508 | /// Double the size of the allocated memory, guaranteeing space for at |
| 509 | /// least one more element or MinSize if specified. |
| 510 | void grow(size_t MinSize = 0) { this->grow_pod(MinSize, sizeof(T)); } |
| 511 | |
| 512 | /// Reserve enough space to add one element, and return the updated element |
| 513 | /// pointer in case it was a reference to the storage. |
| 514 | const T *reserveForParamAndGetAddress(const T &Elt, size_t N = 1) { |
| 515 | return this->reserveForParamAndGetAddressImpl(this, Elt, N); |
| 516 | } |
| 517 | |
| 518 | /// Reserve enough space to add one element, and return the updated element |
| 519 | /// pointer in case it was a reference to the storage. |
| 520 | T *reserveForParamAndGetAddress(T &Elt, size_t N = 1) { |
| 521 | return const_cast<T *>( |
| 522 | this->reserveForParamAndGetAddressImpl(this, Elt, N)); |
| 523 | } |
| 524 | |
| 525 | /// Copy \p V or return a reference, depending on \a ValueParamT. |
| 526 | static ValueParamT forward_value_param(ValueParamT V) { return V; } |
| 527 | |
| 528 | void growAndAssign(size_t NumElts, T Elt) { |
| 529 | // Elt has been copied in case it's an internal reference, side-stepping |
| 530 | // reference invalidation problems without losing the realloc optimization. |
| 531 | this->set_size(0); |
| 532 | this->grow(NumElts); |
| 533 | std::uninitialized_fill_n(this->begin(), NumElts, Elt); |
| 534 | this->set_size(NumElts); |
| 535 | } |
| 536 | |
| 537 | template <typename... ArgTypes> T &growAndEmplaceBack(ArgTypes &&... Args) { |
| 538 | // Use push_back with a copy in case Args has an internal reference, |
| 539 | // side-stepping reference invalidation problems without losing the realloc |
| 540 | // optimization. |
| 541 | push_back(T(std::forward<ArgTypes>(Args)...)); |
| 542 | return this->back(); |
| 543 | } |
| 544 | |
| 545 | public: |
| 546 | void push_back(ValueParamT Elt) { |
| 547 | const T *EltPtr = reserveForParamAndGetAddress(Elt); |
| 548 | memcpy(reinterpret_cast<void *>(this->end()), EltPtr, sizeof(T)); |
| 549 | this->set_size(this->size() + 1); |
| 550 | } |
| 551 | |
| 552 | void pop_back() { this->set_size(this->size() - 1); } |
| 553 | }; |
| 554 | |
| 555 | /// This class consists of common code factored out of the SmallVector class to |
| 556 | /// reduce code duplication based on the SmallVector 'N' template parameter. |
| 557 | template <typename T> |
| 558 | class SmallVectorImpl : public SmallVectorTemplateBase<T> { |
| 559 | using SuperClass = SmallVectorTemplateBase<T>; |
| 560 | |
| 561 | public: |
| 562 | using iterator = typename SuperClass::iterator; |
| 563 | using const_iterator = typename SuperClass::const_iterator; |
| 564 | using reference = typename SuperClass::reference; |
| 565 | using size_type = typename SuperClass::size_type; |
| 566 | |
| 567 | protected: |
| 568 | using SmallVectorTemplateBase<T>::TakesParamByValue; |
| 569 | using ValueParamT = typename SuperClass::ValueParamT; |
| 570 | |
| 571 | // Default ctor - Initialize to empty. |
| 572 | explicit SmallVectorImpl(unsigned N) |
| 573 | : SmallVectorTemplateBase<T>(N) {} |
| 574 | |
| 575 | public: |
| 576 | SmallVectorImpl(const SmallVectorImpl &) = delete; |
| 577 | |
| 578 | ~SmallVectorImpl() { |
| 579 | // Subclass has already destructed this vector's elements. |
| 580 | // If this wasn't grown from the inline copy, deallocate the old space. |
| 581 | if (!this->isSmall()) |
| 582 | free(this->begin()); |
| 583 | } |
| 584 | |
| 585 | void clear() { |
| 586 | this->destroy_range(this->begin(), this->end()); |
| 587 | this->Size = 0; |
| 588 | } |
| 589 | |
| 590 | private: |
| 591 | template <bool ForOverwrite> void resizeImpl(size_type N) { |
| 592 | if (N < this->size()) { |
| 593 | this->pop_back_n(this->size() - N); |
| 594 | } else if (N > this->size()) { |
| 595 | this->reserve(N); |
| 596 | for (auto I = this->end(), E = this->begin() + N; I != E; ++I) |
| 597 | if (ForOverwrite) |
| 598 | new (&*I) T; |
| 599 | else |
| 600 | new (&*I) T(); |
| 601 | this->set_size(N); |
| 602 | } |
| 603 | } |
| 604 | |
| 605 | public: |
| 606 | void resize(size_type N) { resizeImpl<false>(N); } |
| 607 | |
| 608 | /// Like resize, but \ref T is POD, the new values won't be initialized. |
| 609 | void resize_for_overwrite(size_type N) { resizeImpl<true>(N); } |
| 610 | |
| 611 | void resize(size_type N, ValueParamT NV) { |
| 612 | if (N == this->size()) |
| 613 | return; |
| 614 | |
| 615 | if (N < this->size()) { |
| 616 | this->pop_back_n(this->size() - N); |
| 617 | return; |
| 618 | } |
| 619 | |
| 620 | // N > this->size(). Defer to append. |
| 621 | this->append(N - this->size(), NV); |
| 622 | } |
| 623 | |
| 624 | void reserve(size_type N) { |
| 625 | if (this->capacity() < N) |
| 626 | this->grow(N); |
| 627 | } |
| 628 | |
| 629 | void pop_back_n(size_type NumItems) { |
| 630 | assert(this->size() >= NumItems)(static_cast <bool> (this->size() >= NumItems) ? void (0) : __assert_fail ("this->size() >= NumItems", "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/include/llvm/ADT/SmallVector.h" , 630, __extension__ __PRETTY_FUNCTION__)); |
| 631 | this->destroy_range(this->end() - NumItems, this->end()); |
| 632 | this->set_size(this->size() - NumItems); |
| 633 | } |
| 634 | |
| 635 | LLVM_NODISCARD[[clang::warn_unused_result]] T pop_back_val() { |
| 636 | T Result = ::std::move(this->back()); |
| 637 | this->pop_back(); |
| 638 | return Result; |
| 639 | } |
| 640 | |
| 641 | void swap(SmallVectorImpl &RHS); |
| 642 | |
| 643 | /// Add the specified range to the end of the SmallVector. |
| 644 | template <typename in_iter, |
| 645 | typename = std::enable_if_t<std::is_convertible< |
| 646 | typename std::iterator_traits<in_iter>::iterator_category, |
| 647 | std::input_iterator_tag>::value>> |
| 648 | void append(in_iter in_start, in_iter in_end) { |
| 649 | this->assertSafeToAddRange(in_start, in_end); |
| 650 | size_type NumInputs = std::distance(in_start, in_end); |
| 651 | this->reserve(this->size() + NumInputs); |
| 652 | this->uninitialized_copy(in_start, in_end, this->end()); |
| 653 | this->set_size(this->size() + NumInputs); |
| 654 | } |
| 655 | |
| 656 | /// Append \p NumInputs copies of \p Elt to the end. |
| 657 | void append(size_type NumInputs, ValueParamT Elt) { |
| 658 | const T *EltPtr = this->reserveForParamAndGetAddress(Elt, NumInputs); |
| 659 | std::uninitialized_fill_n(this->end(), NumInputs, *EltPtr); |
| 660 | this->set_size(this->size() + NumInputs); |
| 661 | } |
| 662 | |
| 663 | void append(std::initializer_list<T> IL) { |
| 664 | append(IL.begin(), IL.end()); |
| 665 | } |
| 666 | |
| 667 | void append(const SmallVectorImpl &RHS) { append(RHS.begin(), RHS.end()); } |
| 668 | |
| 669 | void assign(size_type NumElts, ValueParamT Elt) { |
| 670 | // Note that Elt could be an internal reference. |
| 671 | if (NumElts > this->capacity()) { |
| 672 | this->growAndAssign(NumElts, Elt); |
| 673 | return; |
| 674 | } |
| 675 | |
| 676 | // Assign over existing elements. |
| 677 | std::fill_n(this->begin(), std::min(NumElts, this->size()), Elt); |
| 678 | if (NumElts > this->size()) |
| 679 | std::uninitialized_fill_n(this->end(), NumElts - this->size(), Elt); |
| 680 | else if (NumElts < this->size()) |
| 681 | this->destroy_range(this->begin() + NumElts, this->end()); |
| 682 | this->set_size(NumElts); |
| 683 | } |
| 684 | |
| 685 | // FIXME: Consider assigning over existing elements, rather than clearing & |
| 686 | // re-initializing them - for all assign(...) variants. |
| 687 | |
| 688 | template <typename in_iter, |
| 689 | typename = std::enable_if_t<std::is_convertible< |
| 690 | typename std::iterator_traits<in_iter>::iterator_category, |
| 691 | std::input_iterator_tag>::value>> |
| 692 | void assign(in_iter in_start, in_iter in_end) { |
| 693 | this->assertSafeToReferenceAfterClear(in_start, in_end); |
| 694 | clear(); |
| 695 | append(in_start, in_end); |
| 696 | } |
| 697 | |
| 698 | void assign(std::initializer_list<T> IL) { |
| 699 | clear(); |
| 700 | append(IL); |
| 701 | } |
| 702 | |
| 703 | void assign(const SmallVectorImpl &RHS) { assign(RHS.begin(), RHS.end()); } |
| 704 | |
| 705 | iterator erase(const_iterator CI) { |
| 706 | // Just cast away constness because this is a non-const member function. |
| 707 | iterator I = const_cast<iterator>(CI); |
| 708 | |
| 709 | assert(this->isReferenceToStorage(CI) && "Iterator to erase is out of bounds.")(static_cast <bool> (this->isReferenceToStorage(CI) && "Iterator to erase is out of bounds.") ? void (0) : __assert_fail ("this->isReferenceToStorage(CI) && \"Iterator to erase is out of bounds.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/include/llvm/ADT/SmallVector.h" , 709, __extension__ __PRETTY_FUNCTION__)); |
| 710 | |
| 711 | iterator N = I; |
| 712 | // Shift all elts down one. |
| 713 | std::move(I+1, this->end(), I); |
| 714 | // Drop the last elt. |
| 715 | this->pop_back(); |
| 716 | return(N); |
| 717 | } |
| 718 | |
| 719 | iterator erase(const_iterator CS, const_iterator CE) { |
| 720 | // Just cast away constness because this is a non-const member function. |
| 721 | iterator S = const_cast<iterator>(CS); |
| 722 | iterator E = const_cast<iterator>(CE); |
| 723 | |
| 724 | assert(this->isRangeInStorage(S, E) && "Range to erase is out of bounds.")(static_cast <bool> (this->isRangeInStorage(S, E) && "Range to erase is out of bounds.") ? void (0) : __assert_fail ("this->isRangeInStorage(S, E) && \"Range to erase is out of bounds.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/include/llvm/ADT/SmallVector.h" , 724, __extension__ __PRETTY_FUNCTION__)); |
| 725 | |
| 726 | iterator N = S; |
| 727 | // Shift all elts down. |
| 728 | iterator I = std::move(E, this->end(), S); |
| 729 | // Drop the last elts. |
| 730 | this->destroy_range(I, this->end()); |
| 731 | this->set_size(I - this->begin()); |
| 732 | return(N); |
| 733 | } |
| 734 | |
| 735 | private: |
| 736 | template <class ArgType> iterator insert_one_impl(iterator I, ArgType &&Elt) { |
| 737 | // Callers ensure that ArgType is derived from T. |
| 738 | static_assert( |
| 739 | std::is_same<std::remove_const_t<std::remove_reference_t<ArgType>>, |
| 740 | T>::value, |
| 741 | "ArgType must be derived from T!"); |
| 742 | |
| 743 | if (I == this->end()) { // Important special case for empty vector. |
| 744 | this->push_back(::std::forward<ArgType>(Elt)); |
| 745 | return this->end()-1; |
| 746 | } |
| 747 | |
| 748 | assert(this->isReferenceToStorage(I) && "Insertion iterator is out of bounds.")(static_cast <bool> (this->isReferenceToStorage(I) && "Insertion iterator is out of bounds.") ? void (0) : __assert_fail ("this->isReferenceToStorage(I) && \"Insertion iterator is out of bounds.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/include/llvm/ADT/SmallVector.h" , 748, __extension__ __PRETTY_FUNCTION__)); |
| 749 | |
| 750 | // Grow if necessary. |
| 751 | size_t Index = I - this->begin(); |
| 752 | std::remove_reference_t<ArgType> *EltPtr = |
| 753 | this->reserveForParamAndGetAddress(Elt); |
| 754 | I = this->begin() + Index; |
| 755 | |
| 756 | ::new ((void*) this->end()) T(::std::move(this->back())); |
| 757 | // Push everything else over. |
| 758 | std::move_backward(I, this->end()-1, this->end()); |
| 759 | this->set_size(this->size() + 1); |
| 760 | |
| 761 | // If we just moved the element we're inserting, be sure to update |
| 762 | // the reference (never happens if TakesParamByValue). |
| 763 | static_assert(!TakesParamByValue || std::is_same<ArgType, T>::value, |
| 764 | "ArgType must be 'T' when taking by value!"); |
| 765 | if (!TakesParamByValue && this->isReferenceToRange(EltPtr, I, this->end())) |
| 766 | ++EltPtr; |
| 767 | |
| 768 | *I = ::std::forward<ArgType>(*EltPtr); |
| 769 | return I; |
| 770 | } |
| 771 | |
| 772 | public: |
| 773 | iterator insert(iterator I, T &&Elt) { |
| 774 | return insert_one_impl(I, this->forward_value_param(std::move(Elt))); |
| 775 | } |
| 776 | |
| 777 | iterator insert(iterator I, const T &Elt) { |
| 778 | return insert_one_impl(I, this->forward_value_param(Elt)); |
| 779 | } |
| 780 | |
| 781 | iterator insert(iterator I, size_type NumToInsert, ValueParamT Elt) { |
| 782 | // Convert iterator to elt# to avoid invalidating iterator when we reserve() |
| 783 | size_t InsertElt = I - this->begin(); |
| 784 | |
| 785 | if (I == this->end()) { // Important special case for empty vector. |
| 786 | append(NumToInsert, Elt); |
| 787 | return this->begin()+InsertElt; |
| 788 | } |
| 789 | |
| 790 | assert(this->isReferenceToStorage(I) && "Insertion iterator is out of bounds.")(static_cast <bool> (this->isReferenceToStorage(I) && "Insertion iterator is out of bounds.") ? void (0) : __assert_fail ("this->isReferenceToStorage(I) && \"Insertion iterator is out of bounds.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/include/llvm/ADT/SmallVector.h" , 790, __extension__ __PRETTY_FUNCTION__)); |
| 791 | |
| 792 | // Ensure there is enough space, and get the (maybe updated) address of |
| 793 | // Elt. |
| 794 | const T *EltPtr = this->reserveForParamAndGetAddress(Elt, NumToInsert); |
| 795 | |
| 796 | // Uninvalidate the iterator. |
| 797 | I = this->begin()+InsertElt; |
| 798 | |
| 799 | // If there are more elements between the insertion point and the end of the |
| 800 | // range than there are being inserted, we can use a simple approach to |
| 801 | // insertion. Since we already reserved space, we know that this won't |
| 802 | // reallocate the vector. |
| 803 | if (size_t(this->end()-I) >= NumToInsert) { |
| 804 | T *OldEnd = this->end(); |
| 805 | append(std::move_iterator<iterator>(this->end() - NumToInsert), |
| 806 | std::move_iterator<iterator>(this->end())); |
| 807 | |
| 808 | // Copy the existing elements that get replaced. |
| 809 | std::move_backward(I, OldEnd-NumToInsert, OldEnd); |
| 810 | |
| 811 | // If we just moved the element we're inserting, be sure to update |
| 812 | // the reference (never happens if TakesParamByValue). |
| 813 | if (!TakesParamByValue && I <= EltPtr && EltPtr < this->end()) |
| 814 | EltPtr += NumToInsert; |
| 815 | |
| 816 | std::fill_n(I, NumToInsert, *EltPtr); |
| 817 | return I; |
| 818 | } |
| 819 | |
| 820 | // Otherwise, we're inserting more elements than exist already, and we're |
| 821 | // not inserting at the end. |
| 822 | |
| 823 | // Move over the elements that we're about to overwrite. |
| 824 | T *OldEnd = this->end(); |
| 825 | this->set_size(this->size() + NumToInsert); |
| 826 | size_t NumOverwritten = OldEnd-I; |
| 827 | this->uninitialized_move(I, OldEnd, this->end()-NumOverwritten); |
| 828 | |
| 829 | // If we just moved the element we're inserting, be sure to update |
| 830 | // the reference (never happens if TakesParamByValue). |
| 831 | if (!TakesParamByValue && I <= EltPtr && EltPtr < this->end()) |
| 832 | EltPtr += NumToInsert; |
| 833 | |
| 834 | // Replace the overwritten part. |
| 835 | std::fill_n(I, NumOverwritten, *EltPtr); |
| 836 | |
| 837 | // Insert the non-overwritten middle part. |
| 838 | std::uninitialized_fill_n(OldEnd, NumToInsert - NumOverwritten, *EltPtr); |
| 839 | return I; |
| 840 | } |
| 841 | |
| 842 | template <typename ItTy, |
| 843 | typename = std::enable_if_t<std::is_convertible< |
| 844 | typename std::iterator_traits<ItTy>::iterator_category, |
| 845 | std::input_iterator_tag>::value>> |
| 846 | iterator insert(iterator I, ItTy From, ItTy To) { |
| 847 | // Convert iterator to elt# to avoid invalidating iterator when we reserve() |
| 848 | size_t InsertElt = I - this->begin(); |
| 849 | |
| 850 | if (I == this->end()) { // Important special case for empty vector. |
| 851 | append(From, To); |
| 852 | return this->begin()+InsertElt; |
| 853 | } |
| 854 | |
| 855 | assert(this->isReferenceToStorage(I) && "Insertion iterator is out of bounds.")(static_cast <bool> (this->isReferenceToStorage(I) && "Insertion iterator is out of bounds.") ? void (0) : __assert_fail ("this->isReferenceToStorage(I) && \"Insertion iterator is out of bounds.\"" , "/build/llvm-toolchain-snapshot-13~++20210506100649+6304c0836a4d/llvm/include/llvm/ADT/SmallVector.h" , 855, __extension__ __PRETTY_FUNCTION__)); |
| 856 | |
| 857 | // Check that the reserve that follows doesn't invalidate the iterators. |
| 858 | this->assertSafeToAddRange(From, To); |
| 859 | |
| 860 | size_t NumToInsert = std::distance(From, To); |
| 861 | |
| 862 | // Ensure there is enough space. |
| 863 | reserve(this->size() + NumToInsert); |
| 864 | |
| 865 | // Uninvalidate the iterator. |
| 866 | I = this->begin()+InsertElt; |
| 867 | |
| 868 | // If there are more elements between the insertion point and the end of the |
| 869 | // range than there are being inserted, we can use a simple approach to |
| 870 | // insertion. Since we already reserved space, we know that this won't |
| 871 | // reallocate the vector. |
| 872 | if (size_t(this->end()-I) >= NumToInsert) { |
| 873 | T *OldEnd = this->end(); |
| 874 | append(std::move_iterator<iterator>(this->end() - NumToInsert), |
| 875 | std::move_iterator<iterator>(this->end())); |
| 876 | |
| 877 | // Copy the existing elements that get replaced. |
| 878 | std::move_backward(I, OldEnd-NumToInsert, OldEnd); |
| 879 | |
| 880 | std::copy(From, To, I); |
| 881 | return I; |
| 882 | } |
| 883 | |
| 884 | // Otherwise, we're inserting more elements than exist already, and we're |
| 885 | // not inserting at the end. |
| 886 | |
| 887 | // Move over the elements that we're about to overwrite. |
| 888 | T *OldEnd = this->end(); |
| 889 | this->set_size(this->size() + NumToInsert); |
| 890 | size_t NumOverwritten = OldEnd-I; |
| 891 | this->uninitialized_move(I, OldEnd, this->end()-NumOverwritten); |
| 892 | |
| 893 | // Replace the overwritten part. |
| 894 | for (T *J = I; NumOverwritten > 0; --NumOverwritten) { |
| 895 | *J = *From; |
| 896 | ++J; ++From; |
| 897 | } |
| 898 | |
| 899 | // Insert the non-overwritten middle part. |
| 900 | this->uninitialized_copy(From, To, OldEnd); |
| 901 | return I; |
| 902 | } |
| 903 | |
| 904 | void insert(iterator I, std::initializer_list<T> IL) { |
| 905 | insert(I, IL.begin(), IL.end()); |
| 906 | } |
| 907 | |
| 908 | template <typename... ArgTypes> reference emplace_back(ArgTypes &&... Args) { |
| 909 | if (LLVM_UNLIKELY(this->size() >= this->capacity())__builtin_expect((bool)(this->size() >= this->capacity ()), false)) |
| 910 | return this->growAndEmplaceBack(std::forward<ArgTypes>(Args)...); |
| 911 | |
| 912 | ::new ((void *)this->end()) T(std::forward<ArgTypes>(Args)...); |
| 913 | this->set_size(this->size() + 1); |
| 914 | return this->back(); |
| 915 | } |
| 916 | |
| 917 | SmallVectorImpl &operator=(const SmallVectorImpl &RHS); |
| 918 | |
| 919 | SmallVectorImpl &operator=(SmallVectorImpl &&RHS); |
| 920 | |
| 921 | bool operator==(const SmallVectorImpl &RHS) const { |
| 922 | if (this->size() != RHS.size()) return false; |
| 923 | return std::equal(this->begin(), this->end(), RHS.begin()); |
| 924 | } |
| 925 | bool operator!=(const SmallVectorImpl &RHS) const { |
| 926 | return !(*this == RHS); |
| 927 | } |
| 928 | |
| 929 | bool operator<(const SmallVectorImpl &RHS) const { |
| 930 | return std::lexicographical_compare(this->begin(), this->end(), |
| 931 | RHS.begin(), RHS.end()); |
| 932 | } |
| 933 | }; |
| 934 | |
| 935 | template <typename T> |
| 936 | void SmallVectorImpl<T>::swap(SmallVectorImpl<T> &RHS) { |
| 937 | if (this == &RHS) return; |
| 938 | |
| 939 | // We can only avoid copying elements if neither vector is small. |
| 940 | if (!this->isSmall() && !RHS.isSmall()) { |
| 941 | std::swap(this->BeginX, RHS.BeginX); |
| 942 | std::swap(this->Size, RHS.Size); |
| 943 | std::swap(this->Capacity, RHS.Capacity); |
| 944 | return; |
| 945 | } |
| 946 | this->reserve(RHS.size()); |
| 947 | RHS.reserve(this->size()); |
| 948 | |
| 949 | // Swap the shared elements. |
| 950 | size_t NumShared = this->size(); |
| 951 | if (NumShared > RHS.size()) NumShared = RHS.size(); |
| 952 | for (size_type i = 0; i != NumShared; ++i) |
| 953 | std::swap((*this)[i], RHS[i]); |
| 954 | |
| 955 | // Copy over the extra elts. |
| 956 | if (this->size() > RHS.size()) { |
| 957 | size_t EltDiff = this->size() - RHS.size(); |
| 958 | this->uninitialized_copy(this->begin()+NumShared, this->end(), RHS.end()); |
| 959 | RHS.set_size(RHS.size() + EltDiff); |
| 960 | this->destroy_range(this->begin()+NumShared, this->end()); |
| 961 | this->set_size(NumShared); |
| 962 | } else if (RHS.size() > this->size()) { |
| 963 | size_t EltDiff = RHS.size() - this->size(); |
| 964 | this->uninitialized_copy(RHS.begin()+NumShared, RHS.end(), this->end()); |
| 965 | this->set_size(this->size() + EltDiff); |
| 966 | this->destroy_range(RHS.begin()+NumShared, RHS.end()); |
| 967 | RHS.set_size(NumShared); |
| 968 | } |
| 969 | } |
| 970 | |
| 971 | template <typename T> |
| 972 | SmallVectorImpl<T> &SmallVectorImpl<T>:: |
| 973 | operator=(const SmallVectorImpl<T> &RHS) { |
| 974 | // Avoid self-assignment. |
| 975 | if (this == &RHS) return *this; |
| 976 | |
| 977 | // If we already have sufficient space, assign the common elements, then |
| 978 | // destroy any excess. |
| 979 | size_t RHSSize = RHS.size(); |
| 980 | size_t CurSize = this->size(); |
| 981 | if (CurSize >= RHSSize) { |
| 982 | // Assign common elements. |
| 983 | iterator NewEnd; |
| 984 | if (RHSSize) |
| 985 | NewEnd = std::copy(RHS.begin(), RHS.begin()+RHSSize, this->begin()); |
| 986 | else |
| 987 | NewEnd = this->begin(); |
| 988 | |
| 989 | // Destroy excess elements. |
| 990 | this->destroy_range(NewEnd, this->end()); |
| 991 | |
| 992 | // Trim. |
| 993 | this->set_size(RHSSize); |
| 994 | return *this; |
| 995 | } |
| 996 | |
| 997 | // If we have to grow to have enough elements, destroy the current elements. |
| 998 | // This allows us to avoid copying them during the grow. |
| 999 | // FIXME: don't do this if they're efficiently moveable. |
| 1000 | if (this->capacity() < RHSSize) { |
| 1001 | // Destroy current elements. |
| 1002 | this->clear(); |
| 1003 | CurSize = 0; |
| 1004 | this->grow(RHSSize); |
| 1005 | } else if (CurSize) { |
| 1006 | // Otherwise, use assignment for the already-constructed elements. |
| 1007 | std::copy(RHS.begin(), RHS.begin()+CurSize, this->begin()); |
| 1008 | } |
| 1009 | |
| 1010 | // Copy construct the new elements in place. |
| 1011 | this->uninitialized_copy(RHS.begin()+CurSize, RHS.end(), |
| 1012 | this->begin()+CurSize); |
| 1013 | |
| 1014 | // Set end. |
| 1015 | this->set_size(RHSSize); |
| 1016 | return *this; |
| 1017 | } |
| 1018 | |
| 1019 | template <typename T> |
| 1020 | SmallVectorImpl<T> &SmallVectorImpl<T>::operator=(SmallVectorImpl<T> &&RHS) { |
| 1021 | // Avoid self-assignment. |
| 1022 | if (this == &RHS) return *this; |
| 1023 | |
| 1024 | // If the RHS isn't small, clear this vector and then steal its buffer. |
| 1025 | if (!RHS.isSmall()) { |
| 1026 | this->destroy_range(this->begin(), this->end()); |
| 1027 | if (!this->isSmall()) free(this->begin()); |
| 1028 | this->BeginX = RHS.BeginX; |
| 1029 | this->Size = RHS.Size; |
| 1030 | this->Capacity = RHS.Capacity; |
| 1031 | RHS.resetToSmall(); |
| 1032 | return *this; |
| 1033 | } |
| 1034 | |
| 1035 | // If we already have sufficient space, assign the common elements, then |
| 1036 | // destroy any excess. |
| 1037 | size_t RHSSize = RHS.size(); |
| 1038 | size_t CurSize = this->size(); |
| 1039 | if (CurSize >= RHSSize) { |
| 1040 | // Assign common elements. |
| 1041 | iterator NewEnd = this->begin(); |
| 1042 | if (RHSSize) |
| 1043 | NewEnd = std::move(RHS.begin(), RHS.end(), NewEnd); |
| 1044 | |
| 1045 | // Destroy excess elements and trim the bounds. |
| 1046 | this->destroy_range(NewEnd, this->end()); |
| 1047 | this->set_size(RHSSize); |
| 1048 | |
| 1049 | // Clear the RHS. |
| 1050 | RHS.clear(); |
| 1051 | |
| 1052 | return *this; |
| 1053 | } |
| 1054 | |
| 1055 | // If we have to grow to have enough elements, destroy the current elements. |
| 1056 | // This allows us to avoid copying them during the grow. |
| 1057 | // FIXME: this may not actually make any sense if we can efficiently move |
| 1058 | // elements. |
| 1059 | if (this->capacity() < RHSSize) { |
| 1060 | // Destroy current elements. |
| 1061 | this->clear(); |
| 1062 | CurSize = 0; |
| 1063 | this->grow(RHSSize); |
| 1064 | } else if (CurSize) { |
| 1065 | // Otherwise, use assignment for the already-constructed elements. |
| 1066 | std::move(RHS.begin(), RHS.begin()+CurSize, this->begin()); |
| 1067 | } |
| 1068 | |
| 1069 | // Move-construct the new elements in place. |
| 1070 | this->uninitialized_move(RHS.begin()+CurSize, RHS.end(), |
| 1071 | this->begin()+CurSize); |
| 1072 | |
| 1073 | // Set end. |
| 1074 | this->set_size(RHSSize); |
| 1075 | |
| 1076 | RHS.clear(); |
| 1077 | return *this; |
| 1078 | } |
| 1079 | |
| 1080 | /// Storage for the SmallVector elements. This is specialized for the N=0 case |
| 1081 | /// to avoid allocating unnecessary storage. |
| 1082 | template <typename T, unsigned N> |
| 1083 | struct SmallVectorStorage { |
| 1084 | alignas(T) char InlineElts[N * sizeof(T)]; |
| 1085 | }; |
| 1086 | |
| 1087 | /// We need the storage to be properly aligned even for small-size of 0 so that |
| 1088 | /// the pointer math in \a SmallVectorTemplateCommon::getFirstEl() is |
| 1089 | /// well-defined. |
| 1090 | template <typename T> struct alignas(T) SmallVectorStorage<T, 0> {}; |
| 1091 | |
| 1092 | /// Forward declaration of SmallVector so that |
| 1093 | /// calculateSmallVectorDefaultInlinedElements can reference |
| 1094 | /// `sizeof(SmallVector<T, 0>)`. |
| 1095 | template <typename T, unsigned N> class LLVM_GSL_OWNER[[gsl::Owner]] SmallVector; |
| 1096 | |
| 1097 | /// Helper class for calculating the default number of inline elements for |
| 1098 | /// `SmallVector<T>`. |
| 1099 | /// |
| 1100 | /// This should be migrated to a constexpr function when our minimum |
| 1101 | /// compiler support is enough for multi-statement constexpr functions. |
| 1102 | template <typename T> struct CalculateSmallVectorDefaultInlinedElements { |
| 1103 | // Parameter controlling the default number of inlined elements |
| 1104 | // for `SmallVector<T>`. |
| 1105 | // |
| 1106 | // The default number of inlined elements ensures that |
| 1107 | // 1. There is at least one inlined element. |
| 1108 | // 2. `sizeof(SmallVector<T>) <= kPreferredSmallVectorSizeof` unless |
| 1109 | // it contradicts 1. |
| 1110 | static constexpr size_t kPreferredSmallVectorSizeof = 64; |
| 1111 | |
| 1112 | // static_assert that sizeof(T) is not "too big". |
| 1113 | // |
| 1114 | // Because our policy guarantees at least one inlined element, it is possible |
| 1115 | // for an arbitrarily large inlined element to allocate an arbitrarily large |
| 1116 | // amount of inline storage. We generally consider it an antipattern for a |
| 1117 | // SmallVector to allocate an excessive amount of inline storage, so we want |
| 1118 | // to call attention to these cases and make sure that users are making an |
| 1119 | // intentional decision if they request a lot of inline storage. |
| 1120 | // |
| 1121 | // We want this assertion to trigger in pathological cases, but otherwise |
| 1122 | // not be too easy to hit. To accomplish that, the cutoff is actually somewhat |
| 1123 | // larger than kPreferredSmallVectorSizeof (otherwise, |
| 1124 | // `SmallVector<SmallVector<T>>` would be one easy way to trip it, and that |
| 1125 | // pattern seems useful in practice). |
| 1126 | // |
| 1127 | // One wrinkle is that this assertion is in theory non-portable, since |
| 1128 | // sizeof(T) is in general platform-dependent. However, we don't expect this |
| 1129 | // to be much of an issue, because most LLVM development happens on 64-bit |
| 1130 | // hosts, and therefore sizeof(T) is expected to *decrease* when compiled for |
| 1131 | // 32-bit hosts, dodging the issue. The reverse situation, where development |
| 1132 | // happens on a 32-bit host and then fails due to sizeof(T) *increasing* on a |
| 1133 | // 64-bit host, is expected to be very rare. |
| 1134 | static_assert( |
| 1135 | sizeof(T) <= 256, |
| 1136 | "You are trying to use a default number of inlined elements for " |
| 1137 | "`SmallVector<T>` but `sizeof(T)` is really big! Please use an " |
| 1138 | "explicit number of inlined elements with `SmallVector<T, N>` to make " |
| 1139 | "sure you really want that much inline storage."); |
| 1140 | |
| 1141 | // Discount the size of the header itself when calculating the maximum inline |
| 1142 | // bytes. |
| 1143 | static constexpr size_t PreferredInlineBytes = |
| 1144 | kPreferredSmallVectorSizeof - sizeof(SmallVector<T, 0>); |
| 1145 | static constexpr size_t NumElementsThatFit = PreferredInlineBytes / sizeof(T); |
| 1146 | static constexpr size_t value = |
| 1147 | NumElementsThatFit == 0 ? 1 : NumElementsThatFit; |
| 1148 | }; |
| 1149 | |
| 1150 | /// This is a 'vector' (really, a variable-sized array), optimized |
| 1151 | /// for the case when the array is small. It contains some number of elements |
| 1152 | /// in-place, which allows it to avoid heap allocation when the actual number of |
| 1153 | /// elements is below that threshold. This allows normal "small" cases to be |
| 1154 | /// fast without losing generality for large inputs. |
| 1155 | /// |
| 1156 | /// \note |
| 1157 | /// In the absence of a well-motivated choice for the number of inlined |
| 1158 | /// elements \p N, it is recommended to use \c SmallVector<T> (that is, |
| 1159 | /// omitting the \p N). This will choose a default number of inlined elements |
| 1160 | /// reasonable for allocation on the stack (for example, trying to keep \c |
| 1161 | /// sizeof(SmallVector<T>) around 64 bytes). |
| 1162 | /// |
| 1163 | /// \warning This does not attempt to be exception safe. |
| 1164 | /// |
| 1165 | /// \see https://llvm.org/docs/ProgrammersManual.html#llvm-adt-smallvector-h |
| 1166 | template <typename T, |
| 1167 | unsigned N = CalculateSmallVectorDefaultInlinedElements<T>::value> |
| 1168 | class LLVM_GSL_OWNER[[gsl::Owner]] SmallVector : public SmallVectorImpl<T>, |
| 1169 | SmallVectorStorage<T, N> { |
| 1170 | public: |
| 1171 | SmallVector() : SmallVectorImpl<T>(N) {} |
| 1172 | |
| 1173 | ~SmallVector() { |
| 1174 | // Destroy the constructed elements in the vector. |
| 1175 | this->destroy_range(this->begin(), this->end()); |
| 1176 | } |
| 1177 | |
| 1178 | explicit SmallVector(size_t Size, const T &Value = T()) |
| 1179 | : SmallVectorImpl<T>(N) { |
| 1180 | this->assign(Size, Value); |
| 1181 | } |
| 1182 | |
| 1183 | template <typename ItTy, |
| 1184 | typename = std::enable_if_t<std::is_convertible< |
| 1185 | typename std::iterator_traits<ItTy>::iterator_category, |
| 1186 | std::input_iterator_tag>::value>> |
| 1187 | SmallVector(ItTy S, ItTy E) : SmallVectorImpl<T>(N) { |
| 1188 | this->append(S, E); |
| 1189 | } |
| 1190 | |
| 1191 | template <typename RangeTy> |
| 1192 | explicit SmallVector(const iterator_range<RangeTy> &R) |
| 1193 | : SmallVectorImpl<T>(N) { |
| 1194 | this->append(R.begin(), R.end()); |
| 1195 | } |
| 1196 | |
| 1197 | SmallVector(std::initializer_list<T> IL) : SmallVectorImpl<T>(N) { |
| 1198 | this->assign(IL); |
| 1199 | } |
| 1200 | |
| 1201 | SmallVector(const SmallVector &RHS) : SmallVectorImpl<T>(N) { |
| 1202 | if (!RHS.empty()) |
| 1203 | SmallVectorImpl<T>::operator=(RHS); |
| 1204 | } |
| 1205 | |
| 1206 | SmallVector &operator=(const SmallVector &RHS) { |
| 1207 | SmallVectorImpl<T>::operator=(RHS); |
| 1208 | return *this; |
| 1209 | } |
| 1210 | |
| 1211 | SmallVector(SmallVector &&RHS) : SmallVectorImpl<T>(N) { |
| 1212 | if (!RHS.empty()) |
| 1213 | SmallVectorImpl<T>::operator=(::std::move(RHS)); |
| 1214 | } |
| 1215 | |
| 1216 | SmallVector(SmallVectorImpl<T> &&RHS) : SmallVectorImpl<T>(N) { |
| 1217 | if (!RHS.empty()) |
| 1218 | SmallVectorImpl<T>::operator=(::std::move(RHS)); |
| 1219 | } |
| 1220 | |
| 1221 | SmallVector &operator=(SmallVector &&RHS) { |
| 1222 | SmallVectorImpl<T>::operator=(::std::move(RHS)); |
| 1223 | return *this; |
| 1224 | } |
| 1225 | |
| 1226 | SmallVector &operator=(SmallVectorImpl<T> &&RHS) { |
| 1227 | SmallVectorImpl<T>::operator=(::std::move(RHS)); |
| 1228 | return *this; |
| 1229 | } |
| 1230 | |
| 1231 | SmallVector &operator=(std::initializer_list<T> IL) { |
| 1232 | this->assign(IL); |
| 1233 | return *this; |
| 1234 | } |
| 1235 | }; |
| 1236 | |
| 1237 | template <typename T, unsigned N> |
| 1238 | inline size_t capacity_in_bytes(const SmallVector<T, N> &X) { |
| 1239 | return X.capacity_in_bytes(); |
| 1240 | } |
| 1241 | |
| 1242 | /// Given a range of type R, iterate the entire range and return a |
| 1243 | /// SmallVector with elements of the vector. This is useful, for example, |
| 1244 | /// when you want to iterate a range and then sort the results. |
| 1245 | template <unsigned Size, typename R> |
| 1246 | SmallVector<typename std::remove_const<typename std::remove_reference< |
| 1247 | decltype(*std::begin(std::declval<R &>()))>::type>::type, |
| 1248 | Size> |
| 1249 | to_vector(R &&Range) { |
| 1250 | return {std::begin(Range), std::end(Range)}; |
| 1251 | } |
| 1252 | |
| 1253 | } // end namespace llvm |
| 1254 | |
| 1255 | namespace std { |
| 1256 | |
| 1257 | /// Implement std::swap in terms of SmallVector swap. |
| 1258 | template<typename T> |
| 1259 | inline void |
| 1260 | swap(llvm::SmallVectorImpl<T> &LHS, llvm::SmallVectorImpl<T> &RHS) { |
| 1261 | LHS.swap(RHS); |
| 1262 | } |
| 1263 | |
| 1264 | /// Implement std::swap in terms of SmallVector swap. |
| 1265 | template<typename T, unsigned N> |
| 1266 | inline void |
| 1267 | swap(llvm::SmallVector<T, N> &LHS, llvm::SmallVector<T, N> &RHS) { |
| 1268 | LHS.swap(RHS); |
| 1269 | } |
| 1270 | |
| 1271 | } // end namespace std |
| 1272 | |
| 1273 | #endif // LLVM_ADT_SMALLVECTOR_H |