| 1 | //===------- VectorCombine.cpp - Optimize partial vector operations -------===// |
| 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 optimizes scalar/vector interactions using target cost models. The |
| 10 | // transforms implemented here may not fit in traditional loop-based or SLP |
| 11 | // vectorization passes. |
| 12 | // |
| 13 | //===----------------------------------------------------------------------===// |
| 14 | |
| 15 | #include "llvm/Transforms/Vectorize/VectorCombine.h" |
| 16 | #include "llvm/ADT/DenseMap.h" |
| 17 | #include "llvm/ADT/STLExtras.h" |
| 18 | #include "llvm/ADT/ScopeExit.h" |
| 19 | #include "llvm/ADT/SmallVector.h" |
| 20 | #include "llvm/ADT/Statistic.h" |
| 21 | #include "llvm/Analysis/AssumptionCache.h" |
| 22 | #include "llvm/Analysis/BasicAliasAnalysis.h" |
| 23 | #include "llvm/Analysis/GlobalsModRef.h" |
| 24 | #include "llvm/Analysis/InstSimplifyFolder.h" |
| 25 | #include "llvm/Analysis/Loads.h" |
| 26 | #include "llvm/Analysis/TargetFolder.h" |
| 27 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 28 | #include "llvm/Analysis/ValueTracking.h" |
| 29 | #include "llvm/Analysis/VectorUtils.h" |
| 30 | #include "llvm/IR/Dominators.h" |
| 31 | #include "llvm/IR/Function.h" |
| 32 | #include "llvm/IR/IRBuilder.h" |
| 33 | #include "llvm/IR/Instructions.h" |
| 34 | #include "llvm/IR/PatternMatch.h" |
| 35 | #include "llvm/Support/CommandLine.h" |
| 36 | #include "llvm/Support/KnownBits.h" |
| 37 | #include "llvm/Support/MathExtras.h" |
| 38 | #include "llvm/Transforms/Utils/Local.h" |
| 39 | #include "llvm/Transforms/Utils/LoopUtils.h" |
| 40 | #include <numeric> |
| 41 | #include <optional> |
| 42 | #include <queue> |
| 43 | #include <set> |
| 44 | |
| 45 | #define DEBUG_TYPE "vector-combine" |
| 46 | #include "llvm/Transforms/Utils/InstructionWorklist.h" |
| 47 | |
| 48 | using namespace llvm; |
| 49 | using namespace llvm::PatternMatch; |
| 50 | |
| 51 | STATISTIC(NumVecLoad, "Number of vector loads formed" ); |
| 52 | STATISTIC(NumVecCmp, "Number of vector compares formed" ); |
| 53 | STATISTIC(NumVecBO, "Number of vector binops formed" ); |
| 54 | STATISTIC(NumVecCmpBO, "Number of vector compare + binop formed" ); |
| 55 | STATISTIC(NumShufOfBitcast, "Number of shuffles moved after bitcast" ); |
| 56 | STATISTIC(NumScalarOps, "Number of scalar unary + binary ops formed" ); |
| 57 | STATISTIC(NumScalarCmp, "Number of scalar compares formed" ); |
| 58 | STATISTIC(NumScalarIntrinsic, "Number of scalar intrinsic calls formed" ); |
| 59 | |
| 60 | static cl::opt<bool> DisableVectorCombine( |
| 61 | "disable-vector-combine" , cl::init(Val: false), cl::Hidden, |
| 62 | cl::desc("Disable all vector combine transforms" )); |
| 63 | |
| 64 | static cl::opt<bool> ( |
| 65 | "disable-binop-extract-shuffle" , cl::init(Val: false), cl::Hidden, |
| 66 | cl::desc("Disable binop extract to shuffle transforms" )); |
| 67 | |
| 68 | static cl::opt<unsigned> MaxInstrsToScan( |
| 69 | "vector-combine-max-scan-instrs" , cl::init(Val: 30), cl::Hidden, |
| 70 | cl::desc("Max number of instructions to scan for vector combining." )); |
| 71 | |
| 72 | static const unsigned InvalidIndex = std::numeric_limits<unsigned>::max(); |
| 73 | |
| 74 | namespace { |
| 75 | class VectorCombine { |
| 76 | public: |
| 77 | VectorCombine(Function &F, const TargetTransformInfo &TTI, |
| 78 | const DominatorTree &DT, AAResults &AA, AssumptionCache &AC, |
| 79 | const DataLayout *DL, TTI::TargetCostKind CostKind, |
| 80 | bool TryEarlyFoldsOnly) |
| 81 | : F(F), Builder(F.getContext(), InstSimplifyFolder(*DL)), TTI(TTI), |
| 82 | DT(DT), AA(AA), DL(DL), CostKind(CostKind), |
| 83 | SQ(*DL, /*TLI=*/nullptr, &DT, &AC), |
| 84 | TryEarlyFoldsOnly(TryEarlyFoldsOnly) {} |
| 85 | |
| 86 | bool run(); |
| 87 | |
| 88 | private: |
| 89 | Function &F; |
| 90 | IRBuilder<InstSimplifyFolder> Builder; |
| 91 | const TargetTransformInfo &TTI; |
| 92 | const DominatorTree &DT; |
| 93 | AAResults &AA; |
| 94 | const DataLayout *DL; |
| 95 | TTI::TargetCostKind CostKind; |
| 96 | const SimplifyQuery SQ; |
| 97 | |
| 98 | /// If true, only perform beneficial early IR transforms. Do not introduce new |
| 99 | /// vector operations. |
| 100 | bool TryEarlyFoldsOnly; |
| 101 | |
| 102 | InstructionWorklist Worklist; |
| 103 | |
| 104 | /// Next instruction to iterate. It will be updated when it is erased by |
| 105 | /// RecursivelyDeleteTriviallyDeadInstructions. |
| 106 | Instruction *NextInst; |
| 107 | |
| 108 | // TODO: Direct calls from the top-level "run" loop use a plain "Instruction" |
| 109 | // parameter. That should be updated to specific sub-classes because the |
| 110 | // run loop was changed to dispatch on opcode. |
| 111 | bool vectorizeLoadInsert(Instruction &I); |
| 112 | bool widenSubvectorLoad(Instruction &I); |
| 113 | ExtractElementInst *getShuffleExtract(ExtractElementInst *Ext0, |
| 114 | ExtractElementInst *Ext1, |
| 115 | unsigned ) const; |
| 116 | bool isExtractExtractCheap(ExtractElementInst *Ext0, ExtractElementInst *Ext1, |
| 117 | const Instruction &I, |
| 118 | ExtractElementInst *&ConvertToShuffle, |
| 119 | unsigned ); |
| 120 | Value *foldExtExtCmp(Value *V0, Value *V1, Value *ExtIndex, Instruction &I); |
| 121 | Value *foldExtExtBinop(Value *V0, Value *V1, Value *ExtIndex, Instruction &I); |
| 122 | bool foldExtractExtract(Instruction &I); |
| 123 | bool foldInsExtFNeg(Instruction &I); |
| 124 | bool foldInsExtBinop(Instruction &I); |
| 125 | bool foldInsExtVectorToShuffle(Instruction &I); |
| 126 | bool foldBitOpOfCastops(Instruction &I); |
| 127 | bool foldBitOpOfCastConstant(Instruction &I); |
| 128 | bool foldBitcastShuffle(Instruction &I); |
| 129 | bool scalarizeOpOrCmp(Instruction &I); |
| 130 | bool scalarizeVPIntrinsic(Instruction &I); |
| 131 | bool foldExtractedCmps(Instruction &I); |
| 132 | bool foldSelectsFromBitcast(Instruction &I); |
| 133 | bool foldBinopOfReductions(Instruction &I); |
| 134 | bool foldSingleElementStore(Instruction &I); |
| 135 | bool scalarizeLoad(Instruction &I); |
| 136 | bool scalarizeLoadExtract(LoadInst *LI, VectorType *VecTy, Value *Ptr); |
| 137 | bool scalarizeLoadBitcast(LoadInst *LI, VectorType *VecTy, Value *Ptr); |
| 138 | bool scalarizeExtExtract(Instruction &I); |
| 139 | bool foldConcatOfBoolMasks(Instruction &I); |
| 140 | bool foldPermuteOfBinops(Instruction &I); |
| 141 | bool foldShuffleOfBinops(Instruction &I); |
| 142 | bool foldShuffleOfSelects(Instruction &I); |
| 143 | bool foldShuffleOfCastops(Instruction &I); |
| 144 | bool foldShuffleOfShuffles(Instruction &I); |
| 145 | bool foldPermuteOfIntrinsic(Instruction &I); |
| 146 | bool foldShufflesOfLengthChangingShuffles(Instruction &I); |
| 147 | bool foldShuffleOfIntrinsics(Instruction &I); |
| 148 | bool foldShuffleToIdentity(Instruction &I); |
| 149 | bool foldShuffleFromReductions(Instruction &I); |
| 150 | bool foldShuffleChainsToReduce(Instruction &I); |
| 151 | bool foldCastFromReductions(Instruction &I); |
| 152 | bool foldSignBitReductionCmp(Instruction &I); |
| 153 | bool foldReductionZeroTest(Instruction &I); |
| 154 | bool foldICmpEqZeroVectorReduce(Instruction &I); |
| 155 | bool foldEquivalentReductionCmp(Instruction &I); |
| 156 | bool foldReduceAddCmpZero(Instruction &I); |
| 157 | bool foldSelectShuffle(Instruction &I, bool FromReduction = false); |
| 158 | bool foldInterleaveIntrinsics(Instruction &I); |
| 159 | bool foldDeinterleaveIntrinsics(Instruction &I); |
| 160 | bool foldBitcastOfVPLoad(Instruction &I); |
| 161 | bool foldBitOrderReverseAndSwap(Instruction &I); |
| 162 | bool shrinkType(Instruction &I); |
| 163 | bool shrinkLoadForShuffles(Instruction &I); |
| 164 | bool shrinkPhiOfShuffles(Instruction &I); |
| 165 | |
| 166 | void replaceValue(Instruction &Old, Value &New, bool Erase = true) { |
| 167 | LLVM_DEBUG(dbgs() << "VC: Replacing: " << Old << '\n'); |
| 168 | LLVM_DEBUG(dbgs() << " With: " << New << '\n'); |
| 169 | Old.replaceAllUsesWith(V: &New); |
| 170 | if (auto *NewI = dyn_cast<Instruction>(Val: &New)) { |
| 171 | New.takeName(V: &Old); |
| 172 | Worklist.pushUsersToWorkList(I&: *NewI); |
| 173 | Worklist.pushValue(V: NewI); |
| 174 | } |
| 175 | if (Erase && isInstructionTriviallyDead(I: &Old)) { |
| 176 | eraseInstruction(I&: Old); |
| 177 | } else { |
| 178 | Worklist.push(I: &Old); |
| 179 | } |
| 180 | } |
| 181 | |
| 182 | void eraseInstruction(Instruction &I) { |
| 183 | LLVM_DEBUG(dbgs() << "VC: Erasing: " << I << '\n'); |
| 184 | SmallVector<Value *> Ops(I.operands()); |
| 185 | Worklist.remove(I: &I); |
| 186 | I.eraseFromParent(); |
| 187 | |
| 188 | // Push remaining users of the operands and then the operand itself - allows |
| 189 | // further folds that were hindered by OneUse limits. |
| 190 | SmallPtrSet<Value *, 4> Visited; |
| 191 | for (Value *Op : Ops) { |
| 192 | if (!Visited.contains(Ptr: Op)) { |
| 193 | if (auto *OpI = dyn_cast<Instruction>(Val: Op)) { |
| 194 | if (RecursivelyDeleteTriviallyDeadInstructions( |
| 195 | V: OpI, TLI: nullptr, MSSAU: nullptr, AboutToDeleteCallback: [&](Value *V) { |
| 196 | if (auto *I = dyn_cast<Instruction>(Val: V)) { |
| 197 | LLVM_DEBUG(dbgs() << "VC: Erased: " << *I << '\n'); |
| 198 | Worklist.remove(I); |
| 199 | if (I == NextInst) |
| 200 | NextInst = NextInst->getNextNode(); |
| 201 | Visited.insert(Ptr: I); |
| 202 | } |
| 203 | })) |
| 204 | continue; |
| 205 | Worklist.pushUsersToWorkList(I&: *OpI); |
| 206 | Worklist.pushValue(V: OpI); |
| 207 | } |
| 208 | } |
| 209 | } |
| 210 | } |
| 211 | }; |
| 212 | } // namespace |
| 213 | |
| 214 | /// Return the source operand of a potentially bitcasted value. If there is no |
| 215 | /// bitcast, return the input value itself. |
| 216 | static Value *peekThroughBitcasts(Value *V) { |
| 217 | while (auto *BitCast = dyn_cast<BitCastInst>(Val: V)) |
| 218 | V = BitCast->getOperand(i_nocapture: 0); |
| 219 | return V; |
| 220 | } |
| 221 | |
| 222 | static bool canWidenLoad(LoadInst *Load, const TargetTransformInfo &TTI) { |
| 223 | // Do not widen load if atomic/volatile or under asan/hwasan/memtag/tsan. |
| 224 | // The widened load may load data from dirty regions or create data races |
| 225 | // non-existent in the source. |
| 226 | if (!Load || !Load->isSimple() || !Load->hasOneUse() || |
| 227 | Load->getFunction()->hasFnAttribute(Kind: Attribute::SanitizeMemTag) || |
| 228 | mustSuppressSpeculation(LI: *Load)) |
| 229 | return false; |
| 230 | |
| 231 | // We are potentially transforming byte-sized (8-bit) memory accesses, so make |
| 232 | // sure we have all of our type-based constraints in place for this target. |
| 233 | Type *ScalarTy = Load->getType()->getScalarType(); |
| 234 | uint64_t ScalarSize = ScalarTy->getPrimitiveSizeInBits(); |
| 235 | unsigned MinVectorSize = TTI.getMinVectorRegisterBitWidth(); |
| 236 | if (!ScalarSize || !MinVectorSize || MinVectorSize % ScalarSize != 0 || |
| 237 | ScalarSize % 8 != 0) |
| 238 | return false; |
| 239 | |
| 240 | return true; |
| 241 | } |
| 242 | |
| 243 | bool VectorCombine::vectorizeLoadInsert(Instruction &I) { |
| 244 | // Match insert into fixed vector of scalar value. |
| 245 | // TODO: Handle non-zero insert index. |
| 246 | Value *Scalar; |
| 247 | if (!match(V: &I, |
| 248 | P: m_InsertElt(Val: m_Poison(), Elt: m_OneUse(SubPattern: m_Value(V&: Scalar)), Idx: m_ZeroInt()))) |
| 249 | return false; |
| 250 | |
| 251 | // Optionally match an extract from another vector. |
| 252 | Value *X; |
| 253 | bool = match(V: Scalar, P: m_ExtractElt(Val: m_Value(V&: X), Idx: m_ZeroInt())); |
| 254 | if (!HasExtract) |
| 255 | X = Scalar; |
| 256 | |
| 257 | auto *Load = dyn_cast<LoadInst>(Val: X); |
| 258 | if (!canWidenLoad(Load, TTI)) |
| 259 | return false; |
| 260 | |
| 261 | Type *ScalarTy = Scalar->getType(); |
| 262 | uint64_t ScalarSize = ScalarTy->getPrimitiveSizeInBits(); |
| 263 | unsigned MinVectorSize = TTI.getMinVectorRegisterBitWidth(); |
| 264 | |
| 265 | // Check safety of replacing the scalar load with a larger vector load. |
| 266 | // We use minimal alignment (maximum flexibility) because we only care about |
| 267 | // the dereferenceable region. When calculating cost and creating a new op, |
| 268 | // we may use a larger value based on alignment attributes. |
| 269 | Value *SrcPtr = Load->getPointerOperand()->stripPointerCasts(); |
| 270 | assert(isa<PointerType>(SrcPtr->getType()) && "Expected a pointer type" ); |
| 271 | |
| 272 | unsigned MinVecNumElts = MinVectorSize / ScalarSize; |
| 273 | auto *MinVecTy = VectorType::get(ElementType: ScalarTy, NumElements: MinVecNumElts, Scalable: false); |
| 274 | unsigned OffsetEltIndex = 0; |
| 275 | Align Alignment = Load->getAlign(); |
| 276 | if (!isSafeToLoadUnconditionally(V: SrcPtr, Ty: MinVecTy, Alignment: Align(1), DL: *DL, ScanFrom: Load, AC: SQ.AC, |
| 277 | DT: SQ.DT)) { |
| 278 | // It is not safe to load directly from the pointer, but we can still peek |
| 279 | // through gep offsets and check if it safe to load from a base address with |
| 280 | // updated alignment. If it is, we can shuffle the element(s) into place |
| 281 | // after loading. |
| 282 | unsigned OffsetBitWidth = DL->getIndexTypeSizeInBits(Ty: SrcPtr->getType()); |
| 283 | APInt Offset(OffsetBitWidth, 0); |
| 284 | SrcPtr = SrcPtr->stripAndAccumulateInBoundsConstantOffsets(DL: *DL, Offset); |
| 285 | |
| 286 | // We want to shuffle the result down from a high element of a vector, so |
| 287 | // the offset must be positive. |
| 288 | if (Offset.isNegative()) |
| 289 | return false; |
| 290 | |
| 291 | // The offset must be a multiple of the scalar element to shuffle cleanly |
| 292 | // in the element's size. |
| 293 | uint64_t ScalarSizeInBytes = ScalarSize / 8; |
| 294 | if (Offset.urem(RHS: ScalarSizeInBytes) != 0) |
| 295 | return false; |
| 296 | |
| 297 | // If we load MinVecNumElts, will our target element still be loaded? |
| 298 | OffsetEltIndex = Offset.udiv(RHS: ScalarSizeInBytes).getZExtValue(); |
| 299 | if (OffsetEltIndex >= MinVecNumElts) |
| 300 | return false; |
| 301 | |
| 302 | if (!isSafeToLoadUnconditionally(V: SrcPtr, Ty: MinVecTy, Alignment: Align(1), DL: *DL, ScanFrom: Load, |
| 303 | AC: SQ.AC, DT: SQ.DT)) |
| 304 | return false; |
| 305 | |
| 306 | // Update alignment with offset value. Note that the offset could be negated |
| 307 | // to more accurately represent "(new) SrcPtr - Offset = (old) SrcPtr", but |
| 308 | // negation does not change the result of the alignment calculation. |
| 309 | Alignment = commonAlignment(A: Alignment, Offset: Offset.getZExtValue()); |
| 310 | } |
| 311 | |
| 312 | // Original pattern: insertelt undef, load [free casts of] PtrOp, 0 |
| 313 | // Use the greater of the alignment on the load or its source pointer. |
| 314 | Alignment = std::max(a: SrcPtr->getPointerAlignment(DL: *DL), b: Alignment); |
| 315 | Type *LoadTy = Load->getType(); |
| 316 | unsigned AS = Load->getPointerAddressSpace(); |
| 317 | InstructionCost OldCost = |
| 318 | TTI.getMemoryOpCost(Opcode: Instruction::Load, Src: LoadTy, Alignment, AddressSpace: AS, CostKind); |
| 319 | APInt DemandedElts = APInt::getOneBitSet(numBits: MinVecNumElts, BitNo: 0); |
| 320 | OldCost += |
| 321 | TTI.getScalarizationOverhead(Ty: MinVecTy, DemandedElts, |
| 322 | /* Insert */ true, Extract: HasExtract, CostKind); |
| 323 | |
| 324 | // New pattern: load VecPtr |
| 325 | InstructionCost NewCost = |
| 326 | TTI.getMemoryOpCost(Opcode: Instruction::Load, Src: MinVecTy, Alignment, AddressSpace: AS, CostKind); |
| 327 | // Optionally, we are shuffling the loaded vector element(s) into place. |
| 328 | // For the mask set everything but element 0 to undef to prevent poison from |
| 329 | // propagating from the extra loaded memory. This will also optionally |
| 330 | // shrink/grow the vector from the loaded size to the output size. |
| 331 | // We assume this operation has no cost in codegen if there was no offset. |
| 332 | // Note that we could use freeze to avoid poison problems, but then we might |
| 333 | // still need a shuffle to change the vector size. |
| 334 | auto *Ty = cast<FixedVectorType>(Val: I.getType()); |
| 335 | unsigned OutputNumElts = Ty->getNumElements(); |
| 336 | SmallVector<int, 16> Mask(OutputNumElts, PoisonMaskElem); |
| 337 | assert(OffsetEltIndex < MinVecNumElts && "Address offset too big" ); |
| 338 | Mask[0] = OffsetEltIndex; |
| 339 | if (OffsetEltIndex) |
| 340 | NewCost += TTI.getShuffleCost(Kind: TTI::SK_PermuteSingleSrc, DstTy: Ty, SrcTy: MinVecTy, Mask, |
| 341 | CostKind); |
| 342 | |
| 343 | // We can aggressively convert to the vector form because the backend can |
| 344 | // invert this transform if it does not result in a performance win. |
| 345 | if (OldCost < NewCost || !NewCost.isValid()) |
| 346 | return false; |
| 347 | |
| 348 | // It is safe and potentially profitable to load a vector directly: |
| 349 | // inselt undef, load Scalar, 0 --> load VecPtr |
| 350 | IRBuilder<> Builder(Load); |
| 351 | Value *CastedPtr = |
| 352 | Builder.CreatePointerBitCastOrAddrSpaceCast(V: SrcPtr, DestTy: Builder.getPtrTy(AddrSpace: AS)); |
| 353 | Value *VecLd = Builder.CreateAlignedLoad(Ty: MinVecTy, Ptr: CastedPtr, Align: Alignment); |
| 354 | VecLd = Builder.CreateShuffleVector(V: VecLd, Mask); |
| 355 | |
| 356 | replaceValue(Old&: I, New&: *VecLd); |
| 357 | ++NumVecLoad; |
| 358 | return true; |
| 359 | } |
| 360 | |
| 361 | /// If we are loading a vector and then inserting it into a larger vector with |
| 362 | /// undefined elements, try to load the larger vector and eliminate the insert. |
| 363 | /// This removes a shuffle in IR and may allow combining of other loaded values. |
| 364 | bool VectorCombine::widenSubvectorLoad(Instruction &I) { |
| 365 | // Match subvector insert of fixed vector. |
| 366 | auto *Shuf = cast<ShuffleVectorInst>(Val: &I); |
| 367 | if (!Shuf->isIdentityWithPadding()) |
| 368 | return false; |
| 369 | |
| 370 | // Allow a non-canonical shuffle mask that is choosing elements from op1. |
| 371 | unsigned NumOpElts = |
| 372 | cast<FixedVectorType>(Val: Shuf->getOperand(i_nocapture: 0)->getType())->getNumElements(); |
| 373 | unsigned OpIndex = any_of(Range: Shuf->getShuffleMask(), P: [&NumOpElts](int M) { |
| 374 | return M >= (int)(NumOpElts); |
| 375 | }); |
| 376 | |
| 377 | auto *Load = dyn_cast<LoadInst>(Val: Shuf->getOperand(i_nocapture: OpIndex)); |
| 378 | if (!canWidenLoad(Load, TTI)) |
| 379 | return false; |
| 380 | |
| 381 | // We use minimal alignment (maximum flexibility) because we only care about |
| 382 | // the dereferenceable region. When calculating cost and creating a new op, |
| 383 | // we may use a larger value based on alignment attributes. |
| 384 | auto *Ty = cast<FixedVectorType>(Val: I.getType()); |
| 385 | Value *SrcPtr = Load->getPointerOperand()->stripPointerCasts(); |
| 386 | assert(isa<PointerType>(SrcPtr->getType()) && "Expected a pointer type" ); |
| 387 | Align Alignment = Load->getAlign(); |
| 388 | if (!isSafeToLoadUnconditionally(V: SrcPtr, Ty, Alignment: Align(1), DL: *DL, ScanFrom: Load, AC: SQ.AC, |
| 389 | DT: SQ.DT)) |
| 390 | return false; |
| 391 | |
| 392 | Alignment = std::max(a: SrcPtr->getPointerAlignment(DL: *DL), b: Alignment); |
| 393 | Type *LoadTy = Load->getType(); |
| 394 | unsigned AS = Load->getPointerAddressSpace(); |
| 395 | |
| 396 | // Original pattern: insert_subvector (load PtrOp) |
| 397 | // This conservatively assumes that the cost of a subvector insert into an |
| 398 | // undef value is 0. We could add that cost if the cost model accurately |
| 399 | // reflects the real cost of that operation. |
| 400 | InstructionCost OldCost = |
| 401 | TTI.getMemoryOpCost(Opcode: Instruction::Load, Src: LoadTy, Alignment, AddressSpace: AS, CostKind); |
| 402 | |
| 403 | // New pattern: load PtrOp |
| 404 | InstructionCost NewCost = |
| 405 | TTI.getMemoryOpCost(Opcode: Instruction::Load, Src: Ty, Alignment, AddressSpace: AS, CostKind); |
| 406 | |
| 407 | // We can aggressively convert to the vector form because the backend can |
| 408 | // invert this transform if it does not result in a performance win. |
| 409 | if (OldCost < NewCost || !NewCost.isValid()) |
| 410 | return false; |
| 411 | |
| 412 | IRBuilder<> Builder(Load); |
| 413 | Value *CastedPtr = |
| 414 | Builder.CreatePointerBitCastOrAddrSpaceCast(V: SrcPtr, DestTy: Builder.getPtrTy(AddrSpace: AS)); |
| 415 | Value *VecLd = Builder.CreateAlignedLoad(Ty, Ptr: CastedPtr, Align: Alignment); |
| 416 | replaceValue(Old&: I, New&: *VecLd); |
| 417 | ++NumVecLoad; |
| 418 | return true; |
| 419 | } |
| 420 | |
| 421 | /// Determine which, if any, of the inputs should be replaced by a shuffle |
| 422 | /// followed by extract from a different index. |
| 423 | ExtractElementInst *VectorCombine::( |
| 424 | ExtractElementInst *Ext0, ExtractElementInst *Ext1, |
| 425 | unsigned = InvalidIndex) const { |
| 426 | auto *Index0C = dyn_cast<ConstantInt>(Val: Ext0->getIndexOperand()); |
| 427 | auto *Index1C = dyn_cast<ConstantInt>(Val: Ext1->getIndexOperand()); |
| 428 | assert(Index0C && Index1C && "Expected constant extract indexes" ); |
| 429 | |
| 430 | unsigned Index0 = Index0C->getZExtValue(); |
| 431 | unsigned Index1 = Index1C->getZExtValue(); |
| 432 | |
| 433 | // If the extract indexes are identical, no shuffle is needed. |
| 434 | if (Index0 == Index1) |
| 435 | return nullptr; |
| 436 | |
| 437 | Type *VecTy = Ext0->getVectorOperand()->getType(); |
| 438 | assert(VecTy == Ext1->getVectorOperand()->getType() && "Need matching types" ); |
| 439 | InstructionCost Cost0 = |
| 440 | TTI.getVectorInstrCost(I: *Ext0, Val: VecTy, CostKind, Index: Index0); |
| 441 | InstructionCost Cost1 = |
| 442 | TTI.getVectorInstrCost(I: *Ext1, Val: VecTy, CostKind, Index: Index1); |
| 443 | |
| 444 | // If both costs are invalid no shuffle is needed |
| 445 | if (!Cost0.isValid() && !Cost1.isValid()) |
| 446 | return nullptr; |
| 447 | |
| 448 | // We are extracting from 2 different indexes, so one operand must be shuffled |
| 449 | // before performing a vector operation and/or extract. The more expensive |
| 450 | // extract will be replaced by a shuffle. |
| 451 | if (Cost0 > Cost1) |
| 452 | return Ext0; |
| 453 | if (Cost1 > Cost0) |
| 454 | return Ext1; |
| 455 | |
| 456 | // If the costs are equal and there is a preferred extract index, shuffle the |
| 457 | // opposite operand. |
| 458 | if (PreferredExtractIndex == Index0) |
| 459 | return Ext1; |
| 460 | if (PreferredExtractIndex == Index1) |
| 461 | return Ext0; |
| 462 | |
| 463 | // Otherwise, replace the extract with the higher index. |
| 464 | return Index0 > Index1 ? Ext0 : Ext1; |
| 465 | } |
| 466 | |
| 467 | /// Compare the relative costs of 2 extracts followed by scalar operation vs. |
| 468 | /// vector operation(s) followed by extract. Return true if the existing |
| 469 | /// instructions are cheaper than a vector alternative. Otherwise, return false |
| 470 | /// and if one of the extracts should be transformed to a shufflevector, set |
| 471 | /// \p ConvertToShuffle to that extract instruction. |
| 472 | bool VectorCombine::(ExtractElementInst *Ext0, |
| 473 | ExtractElementInst *Ext1, |
| 474 | const Instruction &I, |
| 475 | ExtractElementInst *&ConvertToShuffle, |
| 476 | unsigned ) { |
| 477 | auto *Ext0IndexC = dyn_cast<ConstantInt>(Val: Ext0->getIndexOperand()); |
| 478 | auto *Ext1IndexC = dyn_cast<ConstantInt>(Val: Ext1->getIndexOperand()); |
| 479 | assert(Ext0IndexC && Ext1IndexC && "Expected constant extract indexes" ); |
| 480 | |
| 481 | unsigned Opcode = I.getOpcode(); |
| 482 | Value *Ext0Src = Ext0->getVectorOperand(); |
| 483 | Value *Ext1Src = Ext1->getVectorOperand(); |
| 484 | Type *ScalarTy = Ext0->getType(); |
| 485 | auto *VecTy = cast<VectorType>(Val: Ext0Src->getType()); |
| 486 | InstructionCost ScalarOpCost, VectorOpCost; |
| 487 | |
| 488 | // Get cost estimates for scalar and vector versions of the operation. |
| 489 | bool IsBinOp = Instruction::isBinaryOp(Opcode); |
| 490 | if (IsBinOp) { |
| 491 | ScalarOpCost = TTI.getArithmeticInstrCost(Opcode, Ty: ScalarTy, CostKind); |
| 492 | VectorOpCost = TTI.getArithmeticInstrCost(Opcode, Ty: VecTy, CostKind); |
| 493 | } else { |
| 494 | assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) && |
| 495 | "Expected a compare" ); |
| 496 | CmpInst::Predicate Pred = cast<CmpInst>(Val: I).getPredicate(); |
| 497 | ScalarOpCost = TTI.getCmpSelInstrCost( |
| 498 | Opcode, ValTy: ScalarTy, CondTy: CmpInst::makeCmpResultType(opnd_type: ScalarTy), VecPred: Pred, CostKind); |
| 499 | VectorOpCost = TTI.getCmpSelInstrCost( |
| 500 | Opcode, ValTy: VecTy, CondTy: CmpInst::makeCmpResultType(opnd_type: VecTy), VecPred: Pred, CostKind); |
| 501 | } |
| 502 | |
| 503 | // Get cost estimates for the extract elements. These costs will factor into |
| 504 | // both sequences. |
| 505 | unsigned Ext0Index = Ext0IndexC->getZExtValue(); |
| 506 | unsigned Ext1Index = Ext1IndexC->getZExtValue(); |
| 507 | |
| 508 | InstructionCost = |
| 509 | TTI.getVectorInstrCost(I: *Ext0, Val: VecTy, CostKind, Index: Ext0Index); |
| 510 | InstructionCost = |
| 511 | TTI.getVectorInstrCost(I: *Ext1, Val: VecTy, CostKind, Index: Ext1Index); |
| 512 | |
| 513 | // A more expensive extract will always be replaced by a splat shuffle. |
| 514 | // For example, if Ext0 is more expensive: |
| 515 | // opcode (extelt V0, Ext0), (ext V1, Ext1) --> |
| 516 | // extelt (opcode (splat V0, Ext0), V1), Ext1 |
| 517 | // TODO: Evaluate whether that always results in lowest cost. Alternatively, |
| 518 | // check the cost of creating a broadcast shuffle and shuffling both |
| 519 | // operands to element 0. |
| 520 | unsigned BestExtIndex = Extract0Cost > Extract1Cost ? Ext0Index : Ext1Index; |
| 521 | unsigned BestInsIndex = Extract0Cost > Extract1Cost ? Ext1Index : Ext0Index; |
| 522 | InstructionCost = std::min(a: Extract0Cost, b: Extract1Cost); |
| 523 | |
| 524 | // Extra uses of the extracts mean that we include those costs in the |
| 525 | // vector total because those instructions will not be eliminated. |
| 526 | InstructionCost OldCost, NewCost; |
| 527 | if (Ext0Src == Ext1Src && Ext0Index == Ext1Index) { |
| 528 | // Handle a special case. If the 2 extracts are identical, adjust the |
| 529 | // formulas to account for that. The extra use charge allows for either the |
| 530 | // CSE'd pattern or an unoptimized form with identical values: |
| 531 | // opcode (extelt V, C), (extelt V, C) --> extelt (opcode V, V), C |
| 532 | bool HasUseTax = Ext0 == Ext1 ? !Ext0->hasNUses(N: 2) |
| 533 | : !Ext0->hasOneUse() || !Ext1->hasOneUse(); |
| 534 | OldCost = CheapExtractCost + ScalarOpCost; |
| 535 | NewCost = VectorOpCost + CheapExtractCost + HasUseTax * CheapExtractCost; |
| 536 | } else { |
| 537 | // Handle the general case. Each extract is actually a different value: |
| 538 | // opcode (extelt V0, C0), (extelt V1, C1) --> extelt (opcode V0, V1), C |
| 539 | OldCost = Extract0Cost + Extract1Cost + ScalarOpCost; |
| 540 | NewCost = VectorOpCost + CheapExtractCost + |
| 541 | !Ext0->hasOneUse() * Extract0Cost + |
| 542 | !Ext1->hasOneUse() * Extract1Cost; |
| 543 | } |
| 544 | |
| 545 | ConvertToShuffle = getShuffleExtract(Ext0, Ext1, PreferredExtractIndex); |
| 546 | if (ConvertToShuffle) { |
| 547 | if (IsBinOp && DisableBinopExtractShuffle) |
| 548 | return true; |
| 549 | |
| 550 | // If we are extracting from 2 different indexes, then one operand must be |
| 551 | // shuffled before performing the vector operation. The shuffle mask is |
| 552 | // poison except for 1 lane that is being translated to the remaining |
| 553 | // extraction lane. Therefore, it is a splat shuffle. Ex: |
| 554 | // ShufMask = { poison, poison, 0, poison } |
| 555 | // TODO: The cost model has an option for a "broadcast" shuffle |
| 556 | // (splat-from-element-0), but no option for a more general splat. |
| 557 | if (auto *FixedVecTy = dyn_cast<FixedVectorType>(Val: VecTy)) { |
| 558 | SmallVector<int> ShuffleMask(FixedVecTy->getNumElements(), |
| 559 | PoisonMaskElem); |
| 560 | ShuffleMask[BestInsIndex] = BestExtIndex; |
| 561 | NewCost += TTI.getShuffleCost(Kind: TargetTransformInfo::SK_PermuteSingleSrc, |
| 562 | DstTy: VecTy, SrcTy: VecTy, Mask: ShuffleMask, CostKind, Index: 0, |
| 563 | SubTp: nullptr, Args: {ConvertToShuffle}); |
| 564 | } else { |
| 565 | NewCost += TTI.getShuffleCost(Kind: TargetTransformInfo::SK_PermuteSingleSrc, |
| 566 | DstTy: VecTy, SrcTy: VecTy, Mask: {}, CostKind, Index: 0, SubTp: nullptr, |
| 567 | Args: {ConvertToShuffle}); |
| 568 | } |
| 569 | } |
| 570 | |
| 571 | LLVM_DEBUG(dbgs() << "Found a binop of extractions: " << I << "\n OldCost: " |
| 572 | << OldCost << " vs NewCost: " << NewCost << "\n" ); |
| 573 | |
| 574 | // Aggressively form a vector op if the cost is equal because the transform |
| 575 | // may enable further optimization. |
| 576 | // Codegen can reverse this transform (scalarize) if it was not profitable. |
| 577 | return OldCost < NewCost; |
| 578 | } |
| 579 | |
| 580 | /// Create a shuffle that translates (shifts) 1 element from the input vector |
| 581 | /// to a new element location. |
| 582 | static Value *createShiftShuffle(Value *Vec, unsigned OldIndex, |
| 583 | unsigned NewIndex, IRBuilderBase &Builder) { |
| 584 | // The shuffle mask is poison except for 1 lane that is being translated |
| 585 | // to the new element index. Example for OldIndex == 2 and NewIndex == 0: |
| 586 | // ShufMask = { 2, poison, poison, poison } |
| 587 | auto *VecTy = cast<FixedVectorType>(Val: Vec->getType()); |
| 588 | SmallVector<int, 32> ShufMask(VecTy->getNumElements(), PoisonMaskElem); |
| 589 | ShufMask[NewIndex] = OldIndex; |
| 590 | return Builder.CreateShuffleVector(V: Vec, Mask: ShufMask, Name: "shift" ); |
| 591 | } |
| 592 | |
| 593 | /// Given an extract element instruction with constant index operand, shuffle |
| 594 | /// the source vector (shift the scalar element) to a NewIndex for extraction. |
| 595 | /// Return null if the input can be constant folded, so that we are not creating |
| 596 | /// unnecessary instructions. |
| 597 | static Value *(ExtractElementInst *ExtElt, unsigned NewIndex, |
| 598 | IRBuilderBase &Builder) { |
| 599 | // Shufflevectors can only be created for fixed-width vectors. |
| 600 | Value *X = ExtElt->getVectorOperand(); |
| 601 | if (!isa<FixedVectorType>(Val: X->getType())) |
| 602 | return nullptr; |
| 603 | |
| 604 | // If the extract can be constant-folded, this code is unsimplified. Defer |
| 605 | // to other passes to handle that. |
| 606 | Value *C = ExtElt->getIndexOperand(); |
| 607 | assert(isa<ConstantInt>(C) && "Expected a constant index operand" ); |
| 608 | if (isa<Constant>(Val: X)) |
| 609 | return nullptr; |
| 610 | |
| 611 | Value *Shuf = createShiftShuffle(Vec: X, OldIndex: cast<ConstantInt>(Val: C)->getZExtValue(), |
| 612 | NewIndex, Builder); |
| 613 | return Shuf; |
| 614 | } |
| 615 | |
| 616 | /// Try to reduce extract element costs by converting scalar compares to vector |
| 617 | /// compares followed by extract. |
| 618 | /// cmp (ext0 V0, ExtIndex), (ext1 V1, ExtIndex) |
| 619 | Value *VectorCombine::foldExtExtCmp(Value *V0, Value *V1, Value *ExtIndex, |
| 620 | Instruction &I) { |
| 621 | assert(isa<CmpInst>(&I) && "Expected a compare" ); |
| 622 | |
| 623 | // cmp Pred (extelt V0, ExtIndex), (extelt V1, ExtIndex) |
| 624 | // --> extelt (cmp Pred V0, V1), ExtIndex |
| 625 | ++NumVecCmp; |
| 626 | CmpInst::Predicate Pred = cast<CmpInst>(Val: &I)->getPredicate(); |
| 627 | Value *VecCmp = Builder.CreateCmp(Pred, LHS: V0, RHS: V1); |
| 628 | return Builder.CreateExtractElement(Vec: VecCmp, Idx: ExtIndex, Name: "foldExtExtCmp" ); |
| 629 | } |
| 630 | |
| 631 | /// Try to reduce extract element costs by converting scalar binops to vector |
| 632 | /// binops followed by extract. |
| 633 | /// bo (ext0 V0, ExtIndex), (ext1 V1, ExtIndex) |
| 634 | Value *VectorCombine::foldExtExtBinop(Value *V0, Value *V1, Value *ExtIndex, |
| 635 | Instruction &I) { |
| 636 | assert(isa<BinaryOperator>(&I) && "Expected a binary operator" ); |
| 637 | |
| 638 | // bo (extelt V0, ExtIndex), (extelt V1, ExtIndex) |
| 639 | // --> extelt (bo V0, V1), ExtIndex |
| 640 | ++NumVecBO; |
| 641 | Value *VecBO = Builder.CreateBinOp(Opc: cast<BinaryOperator>(Val: &I)->getOpcode(), LHS: V0, |
| 642 | RHS: V1, Name: "foldExtExtBinop" ); |
| 643 | |
| 644 | // All IR flags are safe to back-propagate because any potential poison |
| 645 | // created in unused vector elements is discarded by the extract. |
| 646 | if (auto *VecBOInst = dyn_cast<Instruction>(Val: VecBO)) |
| 647 | VecBOInst->copyIRFlags(V: &I); |
| 648 | |
| 649 | return Builder.CreateExtractElement(Vec: VecBO, Idx: ExtIndex, Name: "foldExtExtBinop" ); |
| 650 | } |
| 651 | |
| 652 | /// Match an instruction with extracted vector operands. |
| 653 | bool VectorCombine::(Instruction &I) { |
| 654 | // It is not safe to transform things like div, urem, etc. because we may |
| 655 | // create undefined behavior when executing those on unknown vector elements. |
| 656 | if (!isSafeToSpeculativelyExecute(I: &I)) |
| 657 | return false; |
| 658 | |
| 659 | Instruction *I0, *I1; |
| 660 | CmpPredicate Pred = CmpInst::BAD_ICMP_PREDICATE; |
| 661 | if (!match(V: &I, P: m_Cmp(Pred, L: m_Instruction(I&: I0), R: m_Instruction(I&: I1))) && |
| 662 | !match(V: &I, P: m_BinOp(L: m_Instruction(I&: I0), R: m_Instruction(I&: I1)))) |
| 663 | return false; |
| 664 | |
| 665 | Value *V0, *V1; |
| 666 | uint64_t C0, C1; |
| 667 | if (!match(V: I0, P: m_ExtractElt(Val: m_Value(V&: V0), Idx: m_ConstantInt(V&: C0))) || |
| 668 | !match(V: I1, P: m_ExtractElt(Val: m_Value(V&: V1), Idx: m_ConstantInt(V&: C1))) || |
| 669 | V0->getType() != V1->getType()) |
| 670 | return false; |
| 671 | |
| 672 | // For fixed-width vectors, reject out-of-bounds extract indexes |
| 673 | if (auto *FixedVecTy = dyn_cast<FixedVectorType>(Val: V0->getType())) { |
| 674 | unsigned NumElts = FixedVecTy->getNumElements(); |
| 675 | if (C0 >= NumElts || C1 >= NumElts) |
| 676 | return false; |
| 677 | } |
| 678 | |
| 679 | // If the scalar value 'I' is going to be re-inserted into a vector, then try |
| 680 | // to create an extract to that same element. The extract/insert can be |
| 681 | // reduced to a "select shuffle". |
| 682 | // TODO: If we add a larger pattern match that starts from an insert, this |
| 683 | // probably becomes unnecessary. |
| 684 | auto *Ext0 = cast<ExtractElementInst>(Val: I0); |
| 685 | auto *Ext1 = cast<ExtractElementInst>(Val: I1); |
| 686 | uint64_t InsertIndex = InvalidIndex; |
| 687 | if (I.hasOneUse()) |
| 688 | match(V: I.user_back(), |
| 689 | P: m_InsertElt(Val: m_Value(), Elt: m_Value(), Idx: m_ConstantInt(V&: InsertIndex))); |
| 690 | |
| 691 | ExtractElementInst *; |
| 692 | if (isExtractExtractCheap(Ext0, Ext1, I, ConvertToShuffle&: ExtractToChange, PreferredExtractIndex: InsertIndex)) |
| 693 | return false; |
| 694 | |
| 695 | Value *ExtOp0 = Ext0->getVectorOperand(); |
| 696 | Value *ExtOp1 = Ext1->getVectorOperand(); |
| 697 | |
| 698 | if (ExtractToChange) { |
| 699 | unsigned = ExtractToChange == Ext0 ? C1 : C0; |
| 700 | Value *NewExtOp = |
| 701 | translateExtract(ExtElt: ExtractToChange, NewIndex: CheapExtractIdx, Builder); |
| 702 | if (!NewExtOp) |
| 703 | return false; |
| 704 | if (ExtractToChange == Ext0) |
| 705 | ExtOp0 = NewExtOp; |
| 706 | else |
| 707 | ExtOp1 = NewExtOp; |
| 708 | } |
| 709 | |
| 710 | Value *ExtIndex = ExtractToChange == Ext0 ? Ext1->getIndexOperand() |
| 711 | : Ext0->getIndexOperand(); |
| 712 | Value *NewExt = Pred != CmpInst::BAD_ICMP_PREDICATE |
| 713 | ? foldExtExtCmp(V0: ExtOp0, V1: ExtOp1, ExtIndex, I) |
| 714 | : foldExtExtBinop(V0: ExtOp0, V1: ExtOp1, ExtIndex, I); |
| 715 | Worklist.push(I: Ext0); |
| 716 | Worklist.push(I: Ext1); |
| 717 | replaceValue(Old&: I, New&: *NewExt); |
| 718 | return true; |
| 719 | } |
| 720 | |
| 721 | /// Try to replace an extract + scalar fneg + insert with a vector fneg + |
| 722 | /// shuffle. |
| 723 | bool VectorCombine::foldInsExtFNeg(Instruction &I) { |
| 724 | // Match an insert (op (extract)) pattern. |
| 725 | Value *DstVec; |
| 726 | uint64_t ExtIdx, InsIdx; |
| 727 | Instruction *FNeg; |
| 728 | if (!match(V: &I, P: m_InsertElt(Val: m_Value(V&: DstVec), Elt: m_OneUse(SubPattern: m_Instruction(I&: FNeg)), |
| 729 | Idx: m_ConstantInt(V&: InsIdx)))) |
| 730 | return false; |
| 731 | |
| 732 | // Note: This handles the canonical fneg instruction and "fsub -0.0, X". |
| 733 | Value *SrcVec; |
| 734 | Instruction *; |
| 735 | if (!match(V: FNeg, P: m_FNeg(X: m_CombineAnd( |
| 736 | Ps: m_Instruction(I&: Extract), |
| 737 | Ps: m_ExtractElt(Val: m_Value(V&: SrcVec), Idx: m_ConstantInt(V&: ExtIdx)))))) |
| 738 | return false; |
| 739 | |
| 740 | auto *DstVecTy = cast<FixedVectorType>(Val: DstVec->getType()); |
| 741 | auto *DstVecScalarTy = DstVecTy->getScalarType(); |
| 742 | auto *SrcVecTy = dyn_cast<FixedVectorType>(Val: SrcVec->getType()); |
| 743 | if (!SrcVecTy || DstVecScalarTy != SrcVecTy->getScalarType()) |
| 744 | return false; |
| 745 | |
| 746 | // Ignore if insert/extract index is out of bounds or destination vector has |
| 747 | // one element |
| 748 | unsigned NumDstElts = DstVecTy->getNumElements(); |
| 749 | unsigned NumSrcElts = SrcVecTy->getNumElements(); |
| 750 | if (ExtIdx > NumSrcElts || InsIdx >= NumDstElts || NumDstElts == 1) |
| 751 | return false; |
| 752 | |
| 753 | // We are inserting the negated element into the same lane that we extracted |
| 754 | // from. This is equivalent to a select-shuffle that chooses all but the |
| 755 | // negated element from the destination vector. |
| 756 | SmallVector<int> Mask(NumDstElts); |
| 757 | std::iota(first: Mask.begin(), last: Mask.end(), value: 0); |
| 758 | Mask[InsIdx] = (ExtIdx % NumDstElts) + NumDstElts; |
| 759 | InstructionCost OldCost = |
| 760 | TTI.getArithmeticInstrCost(Opcode: Instruction::FNeg, Ty: DstVecScalarTy, CostKind) + |
| 761 | TTI.getVectorInstrCost(I, Val: DstVecTy, CostKind, Index: InsIdx); |
| 762 | |
| 763 | // If the extract has one use, it will be eliminated, so count it in the |
| 764 | // original cost. If it has more than one use, ignore the cost because it will |
| 765 | // be the same before/after. |
| 766 | if (Extract->hasOneUse()) |
| 767 | OldCost += TTI.getVectorInstrCost(I: *Extract, Val: SrcVecTy, CostKind, Index: ExtIdx); |
| 768 | |
| 769 | InstructionCost NewCost = |
| 770 | TTI.getArithmeticInstrCost(Opcode: Instruction::FNeg, Ty: SrcVecTy, CostKind) + |
| 771 | TTI.getShuffleCost(Kind: TargetTransformInfo::SK_PermuteTwoSrc, DstTy: DstVecTy, |
| 772 | SrcTy: DstVecTy, Mask, CostKind); |
| 773 | |
| 774 | bool NeedLenChg = SrcVecTy->getNumElements() != NumDstElts; |
| 775 | // If the lengths of the two vectors are not equal, |
| 776 | // we need to add a length-change vector. Add this cost. |
| 777 | SmallVector<int> SrcMask; |
| 778 | if (NeedLenChg) { |
| 779 | SrcMask.assign(NumElts: NumDstElts, Elt: PoisonMaskElem); |
| 780 | SrcMask[ExtIdx % NumDstElts] = ExtIdx; |
| 781 | NewCost += TTI.getShuffleCost(Kind: TargetTransformInfo::SK_PermuteSingleSrc, |
| 782 | DstTy: DstVecTy, SrcTy: SrcVecTy, Mask: SrcMask, CostKind); |
| 783 | } |
| 784 | |
| 785 | LLVM_DEBUG(dbgs() << "Found an insertion of (extract)fneg : " << I |
| 786 | << "\n OldCost: " << OldCost << " vs NewCost: " << NewCost |
| 787 | << "\n" ); |
| 788 | if (NewCost > OldCost) |
| 789 | return false; |
| 790 | |
| 791 | Value *NewShuf, *LenChgShuf = nullptr; |
| 792 | // insertelt DstVec, (fneg (extractelt SrcVec, Index)), Index |
| 793 | Value *VecFNeg = Builder.CreateFNegFMF(V: SrcVec, FMFSource: FNeg); |
| 794 | if (NeedLenChg) { |
| 795 | // shuffle DstVec, (shuffle (fneg SrcVec), poison, SrcMask), Mask |
| 796 | LenChgShuf = Builder.CreateShuffleVector(V: VecFNeg, Mask: SrcMask); |
| 797 | NewShuf = Builder.CreateShuffleVector(V1: DstVec, V2: LenChgShuf, Mask); |
| 798 | Worklist.pushValue(V: LenChgShuf); |
| 799 | } else { |
| 800 | // shuffle DstVec, (fneg SrcVec), Mask |
| 801 | NewShuf = Builder.CreateShuffleVector(V1: DstVec, V2: VecFNeg, Mask); |
| 802 | } |
| 803 | |
| 804 | Worklist.pushValue(V: VecFNeg); |
| 805 | replaceValue(Old&: I, New&: *NewShuf); |
| 806 | return true; |
| 807 | } |
| 808 | |
| 809 | /// Try to fold insert(binop(x,y),binop(a,b),idx) |
| 810 | /// --> binop(insert(x,a,idx),insert(y,b,idx)) |
| 811 | bool VectorCombine::foldInsExtBinop(Instruction &I) { |
| 812 | BinaryOperator *VecBinOp, *SclBinOp; |
| 813 | uint64_t Index; |
| 814 | if (!match(V: &I, |
| 815 | P: m_InsertElt(Val: m_OneUse(SubPattern: m_BinOp(I&: VecBinOp)), |
| 816 | Elt: m_OneUse(SubPattern: m_BinOp(I&: SclBinOp)), Idx: m_ConstantInt(V&: Index)))) |
| 817 | return false; |
| 818 | |
| 819 | // TODO: Add support for addlike etc. |
| 820 | Instruction::BinaryOps BinOpcode = VecBinOp->getOpcode(); |
| 821 | if (BinOpcode != SclBinOp->getOpcode()) |
| 822 | return false; |
| 823 | |
| 824 | auto *ResultTy = dyn_cast<FixedVectorType>(Val: I.getType()); |
| 825 | if (!ResultTy) |
| 826 | return false; |
| 827 | |
| 828 | // TODO: Attempt to detect m_ExtractElt for scalar operands and convert to |
| 829 | // shuffle? |
| 830 | |
| 831 | InstructionCost OldCost = TTI.getInstructionCost(U: &I, CostKind) + |
| 832 | TTI.getInstructionCost(U: VecBinOp, CostKind) + |
| 833 | TTI.getInstructionCost(U: SclBinOp, CostKind); |
| 834 | InstructionCost NewCost = |
| 835 | TTI.getArithmeticInstrCost(Opcode: BinOpcode, Ty: ResultTy, CostKind) + |
| 836 | TTI.getVectorInstrCost(Opcode: Instruction::InsertElement, Val: ResultTy, CostKind, |
| 837 | Index, Op0: VecBinOp->getOperand(i_nocapture: 0), |
| 838 | Op1: SclBinOp->getOperand(i_nocapture: 0)) + |
| 839 | TTI.getVectorInstrCost(Opcode: Instruction::InsertElement, Val: ResultTy, CostKind, |
| 840 | Index, Op0: VecBinOp->getOperand(i_nocapture: 1), |
| 841 | Op1: SclBinOp->getOperand(i_nocapture: 1)); |
| 842 | |
| 843 | LLVM_DEBUG(dbgs() << "Found an insertion of two binops: " << I |
| 844 | << "\n OldCost: " << OldCost << " vs NewCost: " << NewCost |
| 845 | << "\n" ); |
| 846 | if (NewCost > OldCost) |
| 847 | return false; |
| 848 | |
| 849 | Value *NewIns0 = Builder.CreateInsertElement(Vec: VecBinOp->getOperand(i_nocapture: 0), |
| 850 | NewElt: SclBinOp->getOperand(i_nocapture: 0), Idx: Index); |
| 851 | Value *NewIns1 = Builder.CreateInsertElement(Vec: VecBinOp->getOperand(i_nocapture: 1), |
| 852 | NewElt: SclBinOp->getOperand(i_nocapture: 1), Idx: Index); |
| 853 | Value *NewBO = Builder.CreateBinOp(Opc: BinOpcode, LHS: NewIns0, RHS: NewIns1); |
| 854 | |
| 855 | // Intersect flags from the old binops. |
| 856 | if (auto *NewInst = dyn_cast<Instruction>(Val: NewBO)) { |
| 857 | NewInst->copyIRFlags(V: VecBinOp); |
| 858 | NewInst->andIRFlags(V: SclBinOp); |
| 859 | } |
| 860 | |
| 861 | Worklist.pushValue(V: NewIns0); |
| 862 | Worklist.pushValue(V: NewIns1); |
| 863 | replaceValue(Old&: I, New&: *NewBO); |
| 864 | return true; |
| 865 | } |
| 866 | |
| 867 | /// Match: bitop(castop(x), castop(y)) -> castop(bitop(x, y)) |
| 868 | /// Supports: bitcast, trunc, sext, zext |
| 869 | bool VectorCombine::foldBitOpOfCastops(Instruction &I) { |
| 870 | // Check if this is a bitwise logic operation |
| 871 | auto *BinOp = dyn_cast<BinaryOperator>(Val: &I); |
| 872 | if (!BinOp || !BinOp->isBitwiseLogicOp()) |
| 873 | return false; |
| 874 | |
| 875 | // Get the cast instructions |
| 876 | auto *LHSCast = dyn_cast<CastInst>(Val: BinOp->getOperand(i_nocapture: 0)); |
| 877 | auto *RHSCast = dyn_cast<CastInst>(Val: BinOp->getOperand(i_nocapture: 1)); |
| 878 | if (!LHSCast || !RHSCast) { |
| 879 | LLVM_DEBUG(dbgs() << " One or both operands are not cast instructions\n" ); |
| 880 | return false; |
| 881 | } |
| 882 | |
| 883 | // Both casts must be the same type |
| 884 | Instruction::CastOps CastOpcode = LHSCast->getOpcode(); |
| 885 | if (CastOpcode != RHSCast->getOpcode()) |
| 886 | return false; |
| 887 | |
| 888 | // Only handle supported cast operations |
| 889 | switch (CastOpcode) { |
| 890 | case Instruction::BitCast: |
| 891 | case Instruction::Trunc: |
| 892 | case Instruction::SExt: |
| 893 | case Instruction::ZExt: |
| 894 | break; |
| 895 | default: |
| 896 | return false; |
| 897 | } |
| 898 | |
| 899 | Value *LHSSrc = LHSCast->getOperand(i_nocapture: 0); |
| 900 | Value *RHSSrc = RHSCast->getOperand(i_nocapture: 0); |
| 901 | |
| 902 | // Source types must match |
| 903 | if (LHSSrc->getType() != RHSSrc->getType()) |
| 904 | return false; |
| 905 | |
| 906 | auto *SrcTy = LHSSrc->getType(); |
| 907 | auto *DstTy = I.getType(); |
| 908 | // Bitcasts can handle scalar/vector mixes, such as i16 -> <16 x i1>. |
| 909 | // Other casts only handle vector types with integer elements. |
| 910 | if (CastOpcode != Instruction::BitCast && |
| 911 | (!isa<FixedVectorType>(Val: SrcTy) || !isa<FixedVectorType>(Val: DstTy))) |
| 912 | return false; |
| 913 | |
| 914 | // Only integer scalar/vector values are legal for bitwise logic operations. |
| 915 | if (!SrcTy->getScalarType()->isIntegerTy() || |
| 916 | !DstTy->getScalarType()->isIntegerTy()) |
| 917 | return false; |
| 918 | |
| 919 | // Cost Check : |
| 920 | // OldCost = bitlogic + 2*casts |
| 921 | // NewCost = bitlogic + cast |
| 922 | |
| 923 | // Calculate specific costs for each cast with instruction context |
| 924 | InstructionCost LHSCastCost = TTI.getCastInstrCost( |
| 925 | Opcode: CastOpcode, Dst: DstTy, Src: SrcTy, CCH: TTI::CastContextHint::None, CostKind, I: LHSCast); |
| 926 | InstructionCost RHSCastCost = TTI.getCastInstrCost( |
| 927 | Opcode: CastOpcode, Dst: DstTy, Src: SrcTy, CCH: TTI::CastContextHint::None, CostKind, I: RHSCast); |
| 928 | |
| 929 | InstructionCost OldCost = |
| 930 | TTI.getArithmeticInstrCost(Opcode: BinOp->getOpcode(), Ty: DstTy, CostKind) + |
| 931 | LHSCastCost + RHSCastCost; |
| 932 | |
| 933 | // For new cost, we can't provide an instruction (it doesn't exist yet) |
| 934 | InstructionCost GenericCastCost = TTI.getCastInstrCost( |
| 935 | Opcode: CastOpcode, Dst: DstTy, Src: SrcTy, CCH: TTI::CastContextHint::None, CostKind); |
| 936 | |
| 937 | InstructionCost NewCost = |
| 938 | TTI.getArithmeticInstrCost(Opcode: BinOp->getOpcode(), Ty: SrcTy, CostKind) + |
| 939 | GenericCastCost; |
| 940 | |
| 941 | // Account for multi-use casts using specific costs |
| 942 | if (!LHSCast->hasOneUse()) |
| 943 | NewCost += LHSCastCost; |
| 944 | if (!RHSCast->hasOneUse()) |
| 945 | NewCost += RHSCastCost; |
| 946 | |
| 947 | LLVM_DEBUG(dbgs() << "foldBitOpOfCastops: OldCost=" << OldCost |
| 948 | << " NewCost=" << NewCost << "\n" ); |
| 949 | |
| 950 | if (NewCost > OldCost) |
| 951 | return false; |
| 952 | |
| 953 | // Create the operation on the source type |
| 954 | Value *NewOp = Builder.CreateBinOp(Opc: BinOp->getOpcode(), LHS: LHSSrc, RHS: RHSSrc, |
| 955 | Name: BinOp->getName() + ".inner" ); |
| 956 | if (auto *NewBinOp = dyn_cast<BinaryOperator>(Val: NewOp)) |
| 957 | NewBinOp->copyIRFlags(V: BinOp); |
| 958 | |
| 959 | Worklist.pushValue(V: NewOp); |
| 960 | |
| 961 | // Create the cast operation directly to ensure we get a new instruction |
| 962 | Instruction *NewCast = CastInst::Create(CastOpcode, S: NewOp, Ty: I.getType()); |
| 963 | |
| 964 | // Preserve cast instruction flags |
| 965 | NewCast->copyIRFlags(V: LHSCast); |
| 966 | NewCast->andIRFlags(V: RHSCast); |
| 967 | |
| 968 | // Insert the new instruction |
| 969 | Value *Result = Builder.Insert(I: NewCast); |
| 970 | |
| 971 | replaceValue(Old&: I, New&: *Result); |
| 972 | return true; |
| 973 | } |
| 974 | |
| 975 | /// Match: |
| 976 | // bitop(castop(x), C) -> |
| 977 | // bitop(castop(x), castop(InvC)) -> |
| 978 | // castop(bitop(x, InvC)) |
| 979 | // Supports: bitcast |
| 980 | bool VectorCombine::foldBitOpOfCastConstant(Instruction &I) { |
| 981 | Instruction *LHS; |
| 982 | Constant *C; |
| 983 | |
| 984 | // Check if this is a bitwise logic operation |
| 985 | if (!match(V: &I, P: m_c_BitwiseLogic(L: m_Instruction(I&: LHS), R: m_Constant(C)))) |
| 986 | return false; |
| 987 | |
| 988 | // Get the cast instructions |
| 989 | auto *LHSCast = dyn_cast<CastInst>(Val: LHS); |
| 990 | if (!LHSCast) |
| 991 | return false; |
| 992 | |
| 993 | Instruction::CastOps CastOpcode = LHSCast->getOpcode(); |
| 994 | |
| 995 | // Only handle supported cast operations |
| 996 | switch (CastOpcode) { |
| 997 | case Instruction::BitCast: |
| 998 | case Instruction::ZExt: |
| 999 | case Instruction::SExt: |
| 1000 | case Instruction::Trunc: |
| 1001 | break; |
| 1002 | default: |
| 1003 | return false; |
| 1004 | } |
| 1005 | |
| 1006 | Value *LHSSrc = LHSCast->getOperand(i_nocapture: 0); |
| 1007 | |
| 1008 | auto *SrcTy = LHSSrc->getType(); |
| 1009 | auto *DstTy = I.getType(); |
| 1010 | // Bitcasts can handle scalar/vector mixes, such as i16 -> <16 x i1>. |
| 1011 | // Other casts only handle vector types with integer elements. |
| 1012 | if (CastOpcode != Instruction::BitCast && |
| 1013 | (!isa<FixedVectorType>(Val: SrcTy) || !isa<FixedVectorType>(Val: DstTy))) |
| 1014 | return false; |
| 1015 | |
| 1016 | // Only integer scalar/vector values are legal for bitwise logic operations. |
| 1017 | if (!SrcTy->getScalarType()->isIntegerTy() || |
| 1018 | !DstTy->getScalarType()->isIntegerTy()) |
| 1019 | return false; |
| 1020 | |
| 1021 | // Find the constant InvC, such that castop(InvC) equals to C. |
| 1022 | PreservedCastFlags RHSFlags; |
| 1023 | Constant *InvC = getLosslessInvCast(C, InvCastTo: SrcTy, CastOp: CastOpcode, DL: *DL, Flags: &RHSFlags); |
| 1024 | if (!InvC) |
| 1025 | return false; |
| 1026 | |
| 1027 | // Cost Check : |
| 1028 | // OldCost = bitlogic + cast |
| 1029 | // NewCost = bitlogic + cast |
| 1030 | |
| 1031 | // Calculate specific costs for each cast with instruction context |
| 1032 | InstructionCost LHSCastCost = TTI.getCastInstrCost( |
| 1033 | Opcode: CastOpcode, Dst: DstTy, Src: SrcTy, CCH: TTI::CastContextHint::None, CostKind, I: LHSCast); |
| 1034 | |
| 1035 | InstructionCost OldCost = |
| 1036 | TTI.getArithmeticInstrCost(Opcode: I.getOpcode(), Ty: DstTy, CostKind) + LHSCastCost; |
| 1037 | |
| 1038 | // For new cost, we can't provide an instruction (it doesn't exist yet) |
| 1039 | InstructionCost GenericCastCost = TTI.getCastInstrCost( |
| 1040 | Opcode: CastOpcode, Dst: DstTy, Src: SrcTy, CCH: TTI::CastContextHint::None, CostKind); |
| 1041 | |
| 1042 | InstructionCost NewCost = |
| 1043 | TTI.getArithmeticInstrCost(Opcode: I.getOpcode(), Ty: SrcTy, CostKind) + |
| 1044 | GenericCastCost; |
| 1045 | |
| 1046 | // Account for multi-use casts using specific costs |
| 1047 | if (!LHSCast->hasOneUse()) |
| 1048 | NewCost += LHSCastCost; |
| 1049 | |
| 1050 | LLVM_DEBUG(dbgs() << "foldBitOpOfCastConstant: OldCost=" << OldCost |
| 1051 | << " NewCost=" << NewCost << "\n" ); |
| 1052 | |
| 1053 | if (NewCost > OldCost) |
| 1054 | return false; |
| 1055 | |
| 1056 | // Create the operation on the source type |
| 1057 | Value *NewOp = Builder.CreateBinOp(Opc: (Instruction::BinaryOps)I.getOpcode(), |
| 1058 | LHS: LHSSrc, RHS: InvC, Name: I.getName() + ".inner" ); |
| 1059 | if (auto *NewBinOp = dyn_cast<BinaryOperator>(Val: NewOp)) |
| 1060 | NewBinOp->copyIRFlags(V: &I); |
| 1061 | |
| 1062 | Worklist.pushValue(V: NewOp); |
| 1063 | |
| 1064 | // Create the cast operation directly to ensure we get a new instruction |
| 1065 | Instruction *NewCast = CastInst::Create(CastOpcode, S: NewOp, Ty: I.getType()); |
| 1066 | |
| 1067 | // Preserve cast instruction flags |
| 1068 | if (RHSFlags.NNeg) |
| 1069 | NewCast->setNonNeg(); |
| 1070 | if (RHSFlags.NUW) |
| 1071 | NewCast->setHasNoUnsignedWrap(); |
| 1072 | if (RHSFlags.NSW) |
| 1073 | NewCast->setHasNoSignedWrap(); |
| 1074 | |
| 1075 | NewCast->andIRFlags(V: LHSCast); |
| 1076 | |
| 1077 | // Insert the new instruction |
| 1078 | Value *Result = Builder.Insert(I: NewCast); |
| 1079 | |
| 1080 | replaceValue(Old&: I, New&: *Result); |
| 1081 | return true; |
| 1082 | } |
| 1083 | |
| 1084 | /// If this is a bitcast of a shuffle, try to bitcast the source vector to the |
| 1085 | /// destination type followed by shuffle. This can enable further transforms by |
| 1086 | /// moving bitcasts or shuffles together. |
| 1087 | bool VectorCombine::foldBitcastShuffle(Instruction &I) { |
| 1088 | Value *V0, *V1; |
| 1089 | ArrayRef<int> Mask; |
| 1090 | if (!match(V: &I, P: m_BitCast(Op: m_OneUse( |
| 1091 | SubPattern: m_Shuffle(v1: m_Value(V&: V0), v2: m_Value(V&: V1), mask: m_Mask(Mask)))))) |
| 1092 | return false; |
| 1093 | |
| 1094 | // 1) Do not fold bitcast shuffle for scalable type. First, shuffle cost for |
| 1095 | // scalable type is unknown; Second, we cannot reason if the narrowed shuffle |
| 1096 | // mask for scalable type is a splat or not. |
| 1097 | // 2) Disallow non-vector casts. |
| 1098 | // TODO: We could allow any shuffle. |
| 1099 | auto *DestTy = dyn_cast<FixedVectorType>(Val: I.getType()); |
| 1100 | auto *SrcTy = dyn_cast<FixedVectorType>(Val: V0->getType()); |
| 1101 | if (!DestTy || !SrcTy) |
| 1102 | return false; |
| 1103 | |
| 1104 | unsigned DestEltSize = DestTy->getScalarSizeInBits(); |
| 1105 | unsigned SrcEltSize = SrcTy->getScalarSizeInBits(); |
| 1106 | if (SrcTy->getPrimitiveSizeInBits() % DestEltSize != 0) |
| 1107 | return false; |
| 1108 | |
| 1109 | bool IsUnary = isa<UndefValue>(Val: V1); |
| 1110 | |
| 1111 | // For binary shuffles, only fold bitcast(shuffle(X,Y)) |
| 1112 | // if it won't increase the number of bitcasts. |
| 1113 | if (!IsUnary) { |
| 1114 | auto *BCTy0 = dyn_cast<FixedVectorType>(Val: peekThroughBitcasts(V: V0)->getType()); |
| 1115 | auto *BCTy1 = dyn_cast<FixedVectorType>(Val: peekThroughBitcasts(V: V1)->getType()); |
| 1116 | if (!(BCTy0 && BCTy0->getElementType() == DestTy->getElementType()) && |
| 1117 | !(BCTy1 && BCTy1->getElementType() == DestTy->getElementType())) |
| 1118 | return false; |
| 1119 | } |
| 1120 | |
| 1121 | SmallVector<int, 16> NewMask; |
| 1122 | if (DestEltSize <= SrcEltSize) { |
| 1123 | // The bitcast is from wide to narrow/equal elements. The shuffle mask can |
| 1124 | // always be expanded to the equivalent form choosing narrower elements. |
| 1125 | if (SrcEltSize % DestEltSize != 0) |
| 1126 | return false; |
| 1127 | unsigned ScaleFactor = SrcEltSize / DestEltSize; |
| 1128 | narrowShuffleMaskElts(Scale: ScaleFactor, Mask, ScaledMask&: NewMask); |
| 1129 | } else { |
| 1130 | // The bitcast is from narrow elements to wide elements. The shuffle mask |
| 1131 | // must choose consecutive elements to allow casting first. |
| 1132 | if (DestEltSize % SrcEltSize != 0) |
| 1133 | return false; |
| 1134 | unsigned ScaleFactor = DestEltSize / SrcEltSize; |
| 1135 | if (!widenShuffleMaskElts(Scale: ScaleFactor, Mask, ScaledMask&: NewMask)) |
| 1136 | return false; |
| 1137 | } |
| 1138 | |
| 1139 | // Bitcast the shuffle src - keep its original width but using the destination |
| 1140 | // scalar type. |
| 1141 | unsigned NumSrcElts = SrcTy->getPrimitiveSizeInBits() / DestEltSize; |
| 1142 | auto *NewShuffleTy = |
| 1143 | FixedVectorType::get(ElementType: DestTy->getScalarType(), NumElts: NumSrcElts); |
| 1144 | auto *OldShuffleTy = |
| 1145 | FixedVectorType::get(ElementType: SrcTy->getScalarType(), NumElts: Mask.size()); |
| 1146 | unsigned NumOps = IsUnary ? 1 : 2; |
| 1147 | |
| 1148 | // The new shuffle must not cost more than the old shuffle. |
| 1149 | TargetTransformInfo::ShuffleKind SK = |
| 1150 | IsUnary ? TargetTransformInfo::SK_PermuteSingleSrc |
| 1151 | : TargetTransformInfo::SK_PermuteTwoSrc; |
| 1152 | |
| 1153 | InstructionCost NewCost = |
| 1154 | TTI.getShuffleCost(Kind: SK, DstTy: DestTy, SrcTy: NewShuffleTy, Mask: NewMask, CostKind) + |
| 1155 | (NumOps * TTI.getCastInstrCost(Opcode: Instruction::BitCast, Dst: NewShuffleTy, Src: SrcTy, |
| 1156 | CCH: TargetTransformInfo::CastContextHint::None, |
| 1157 | CostKind)); |
| 1158 | InstructionCost OldCost = |
| 1159 | TTI.getShuffleCost(Kind: SK, DstTy: OldShuffleTy, SrcTy, Mask, CostKind) + |
| 1160 | TTI.getCastInstrCost(Opcode: Instruction::BitCast, Dst: DestTy, Src: OldShuffleTy, |
| 1161 | CCH: TargetTransformInfo::CastContextHint::None, |
| 1162 | CostKind); |
| 1163 | |
| 1164 | LLVM_DEBUG(dbgs() << "Found a bitcasted shuffle: " << I << "\n OldCost: " |
| 1165 | << OldCost << " vs NewCost: " << NewCost << "\n" ); |
| 1166 | |
| 1167 | if (NewCost > OldCost || !NewCost.isValid()) |
| 1168 | return false; |
| 1169 | |
| 1170 | // bitcast (shuf V0, V1, MaskC) --> shuf (bitcast V0), (bitcast V1), MaskC' |
| 1171 | ++NumShufOfBitcast; |
| 1172 | Value *CastV0 = Builder.CreateBitCast(V: peekThroughBitcasts(V: V0), DestTy: NewShuffleTy); |
| 1173 | Value *CastV1 = Builder.CreateBitCast(V: peekThroughBitcasts(V: V1), DestTy: NewShuffleTy); |
| 1174 | Value *Shuf = Builder.CreateShuffleVector(V1: CastV0, V2: CastV1, Mask: NewMask); |
| 1175 | replaceValue(Old&: I, New&: *Shuf); |
| 1176 | return true; |
| 1177 | } |
| 1178 | |
| 1179 | /// VP Intrinsics whose vector operands are both splat values may be simplified |
| 1180 | /// into the scalar version of the operation and the result splatted. This |
| 1181 | /// can lead to scalarization down the line. |
| 1182 | bool VectorCombine::scalarizeVPIntrinsic(Instruction &I) { |
| 1183 | if (!isa<VPIntrinsic>(Val: I)) |
| 1184 | return false; |
| 1185 | VPIntrinsic &VPI = cast<VPIntrinsic>(Val&: I); |
| 1186 | Value *Op0 = VPI.getArgOperand(i: 0); |
| 1187 | Value *Op1 = VPI.getArgOperand(i: 1); |
| 1188 | |
| 1189 | if (!isSplatValue(V: Op0) || !isSplatValue(V: Op1)) |
| 1190 | return false; |
| 1191 | |
| 1192 | // Check getSplatValue early in this function, to avoid doing unnecessary |
| 1193 | // work. |
| 1194 | Value *ScalarOp0 = getSplatValue(V: Op0); |
| 1195 | Value *ScalarOp1 = getSplatValue(V: Op1); |
| 1196 | if (!ScalarOp0 || !ScalarOp1) |
| 1197 | return false; |
| 1198 | |
| 1199 | // For the binary VP intrinsics supported here, the result on disabled lanes |
| 1200 | // is a poison value. For now, only do this simplification if all lanes |
| 1201 | // are active. |
| 1202 | // TODO: Relax the condition that all lanes are active by using insertelement |
| 1203 | // on inactive lanes. |
| 1204 | auto IsAllTrueMask = [](Value *MaskVal) { |
| 1205 | if (Value *SplattedVal = getSplatValue(V: MaskVal)) |
| 1206 | if (auto *ConstValue = dyn_cast<Constant>(Val: SplattedVal)) |
| 1207 | return ConstValue->isAllOnesValue(); |
| 1208 | return false; |
| 1209 | }; |
| 1210 | if (!IsAllTrueMask(VPI.getArgOperand(i: 2))) |
| 1211 | return false; |
| 1212 | |
| 1213 | // Check to make sure we support scalarization of the intrinsic |
| 1214 | Intrinsic::ID IntrID = VPI.getIntrinsicID(); |
| 1215 | if (!VPBinOpIntrinsic::isVPBinOp(ID: IntrID)) |
| 1216 | return false; |
| 1217 | |
| 1218 | // Calculate cost of splatting both operands into vectors and the vector |
| 1219 | // intrinsic |
| 1220 | VectorType *VecTy = cast<VectorType>(Val: VPI.getType()); |
| 1221 | SmallVector<int> Mask; |
| 1222 | if (auto *FVTy = dyn_cast<FixedVectorType>(Val: VecTy)) |
| 1223 | Mask.resize(N: FVTy->getNumElements(), NV: 0); |
| 1224 | InstructionCost SplatCost = |
| 1225 | TTI.getVectorInstrCost(Opcode: Instruction::InsertElement, Val: VecTy, CostKind, Index: 0) + |
| 1226 | TTI.getShuffleCost(Kind: TargetTransformInfo::SK_Broadcast, DstTy: VecTy, SrcTy: VecTy, Mask, |
| 1227 | CostKind); |
| 1228 | |
| 1229 | // Calculate the cost of the VP Intrinsic |
| 1230 | SmallVector<Type *, 4> Args; |
| 1231 | for (Value *V : VPI.args()) |
| 1232 | Args.push_back(Elt: V->getType()); |
| 1233 | IntrinsicCostAttributes Attrs(IntrID, VecTy, Args); |
| 1234 | InstructionCost VectorOpCost = TTI.getIntrinsicInstrCost(ICA: Attrs, CostKind); |
| 1235 | InstructionCost OldCost = 2 * SplatCost + VectorOpCost; |
| 1236 | |
| 1237 | // Determine scalar opcode |
| 1238 | std::optional<unsigned> FunctionalOpcode = |
| 1239 | VPI.getFunctionalOpcode(); |
| 1240 | std::optional<Intrinsic::ID> ScalarIntrID = std::nullopt; |
| 1241 | if (!FunctionalOpcode) { |
| 1242 | ScalarIntrID = VPI.getFunctionalIntrinsicID(); |
| 1243 | if (!ScalarIntrID) |
| 1244 | return false; |
| 1245 | } |
| 1246 | |
| 1247 | // Calculate cost of scalarizing |
| 1248 | InstructionCost ScalarOpCost = 0; |
| 1249 | if (ScalarIntrID) { |
| 1250 | IntrinsicCostAttributes Attrs(*ScalarIntrID, VecTy->getScalarType(), Args); |
| 1251 | ScalarOpCost = TTI.getIntrinsicInstrCost(ICA: Attrs, CostKind); |
| 1252 | } else { |
| 1253 | ScalarOpCost = TTI.getArithmeticInstrCost(Opcode: *FunctionalOpcode, |
| 1254 | Ty: VecTy->getScalarType(), CostKind); |
| 1255 | } |
| 1256 | |
| 1257 | // The existing splats may be kept around if other instructions use them. |
| 1258 | InstructionCost CostToKeepSplats = |
| 1259 | (SplatCost * !Op0->hasOneUse()) + (SplatCost * !Op1->hasOneUse()); |
| 1260 | InstructionCost NewCost = ScalarOpCost + SplatCost + CostToKeepSplats; |
| 1261 | |
| 1262 | LLVM_DEBUG(dbgs() << "Found a VP Intrinsic to scalarize: " << VPI |
| 1263 | << "\n" ); |
| 1264 | LLVM_DEBUG(dbgs() << "Cost of Intrinsic: " << OldCost |
| 1265 | << ", Cost of scalarizing:" << NewCost << "\n" ); |
| 1266 | |
| 1267 | // We want to scalarize unless the vector variant actually has lower cost. |
| 1268 | if (OldCost < NewCost || !NewCost.isValid()) |
| 1269 | return false; |
| 1270 | |
| 1271 | // Scalarize the intrinsic |
| 1272 | ElementCount EC = cast<VectorType>(Val: Op0->getType())->getElementCount(); |
| 1273 | Value *EVL = VPI.getArgOperand(i: 3); |
| 1274 | |
| 1275 | // If the VP op might introduce UB or poison, we can scalarize it provided |
| 1276 | // that we know the EVL > 0: If the EVL is zero, then the original VP op |
| 1277 | // becomes a no-op and thus won't be UB, so make sure we don't introduce UB by |
| 1278 | // scalarizing it. |
| 1279 | bool SafeToSpeculate; |
| 1280 | if (ScalarIntrID) |
| 1281 | SafeToSpeculate = Intrinsic::getFnAttributes(C&: I.getContext(), id: *ScalarIntrID) |
| 1282 | .hasAttribute(Kind: Attribute::AttrKind::Speculatable); |
| 1283 | else |
| 1284 | SafeToSpeculate = isSafeToSpeculativelyExecuteWithOpcode( |
| 1285 | Opcode: *FunctionalOpcode, Inst: &VPI, CtxI: nullptr, AC: SQ.AC, DT: SQ.DT); |
| 1286 | if (!SafeToSpeculate && |
| 1287 | !isKnownNonZero(V: EVL, Q: SimplifyQuery(*DL, SQ.DT, SQ.AC, &VPI))) |
| 1288 | return false; |
| 1289 | |
| 1290 | Value *ScalarVal = |
| 1291 | ScalarIntrID |
| 1292 | ? Builder.CreateIntrinsic(RetTy: VecTy->getScalarType(), ID: *ScalarIntrID, |
| 1293 | Args: {ScalarOp0, ScalarOp1}) |
| 1294 | : Builder.CreateBinOp(Opc: (Instruction::BinaryOps)(*FunctionalOpcode), |
| 1295 | LHS: ScalarOp0, RHS: ScalarOp1); |
| 1296 | |
| 1297 | replaceValue(Old&: VPI, New&: *Builder.CreateVectorSplat(EC, V: ScalarVal)); |
| 1298 | return true; |
| 1299 | } |
| 1300 | |
| 1301 | /// Match a vector op/compare/intrinsic with at least one |
| 1302 | /// inserted scalar operand and convert to scalar op/cmp/intrinsic followed |
| 1303 | /// by insertelement. |
| 1304 | bool VectorCombine::scalarizeOpOrCmp(Instruction &I) { |
| 1305 | auto *UO = dyn_cast<UnaryOperator>(Val: &I); |
| 1306 | auto *BO = dyn_cast<BinaryOperator>(Val: &I); |
| 1307 | auto *CI = dyn_cast<CmpInst>(Val: &I); |
| 1308 | auto *II = dyn_cast<IntrinsicInst>(Val: &I); |
| 1309 | if (!UO && !BO && !CI && !II) |
| 1310 | return false; |
| 1311 | |
| 1312 | // TODO: Allow intrinsics with different argument types |
| 1313 | if (II) { |
| 1314 | if (!isTriviallyVectorizable(ID: II->getIntrinsicID())) |
| 1315 | return false; |
| 1316 | for (auto [Idx, Arg] : enumerate(First: II->args())) |
| 1317 | if (Arg->getType() != II->getType() && |
| 1318 | !isVectorIntrinsicWithScalarOpAtArg(ID: II->getIntrinsicID(), ScalarOpdIdx: Idx, TTI: &TTI)) |
| 1319 | return false; |
| 1320 | } |
| 1321 | |
| 1322 | // Do not convert the vector condition of a vector select into a scalar |
| 1323 | // condition. That may cause problems for codegen because of differences in |
| 1324 | // boolean formats and register-file transfers. |
| 1325 | // TODO: Can we account for that in the cost model? |
| 1326 | if (CI) |
| 1327 | for (User *U : I.users()) |
| 1328 | if (match(V: U, P: m_Select(C: m_Specific(V: &I), L: m_Value(), R: m_Value()))) |
| 1329 | return false; |
| 1330 | |
| 1331 | // Match constant vectors or scalars being inserted into constant vectors: |
| 1332 | // vec_op [VecC0 | (inselt VecC0, V0, Index)], ... |
| 1333 | SmallVector<Value *> VecCs, ScalarOps; |
| 1334 | std::optional<uint64_t> Index; |
| 1335 | |
| 1336 | auto Ops = II ? II->args() : I.operands(); |
| 1337 | for (auto [OpNum, Op] : enumerate(First&: Ops)) { |
| 1338 | Constant *VecC; |
| 1339 | Value *V; |
| 1340 | uint64_t InsIdx = 0; |
| 1341 | if (match(V: Op.get(), P: m_InsertElt(Val: m_Constant(C&: VecC), Elt: m_Value(V), |
| 1342 | Idx: m_ConstantInt(V&: InsIdx)))) { |
| 1343 | // Bail if any inserts are out of bounds. |
| 1344 | VectorType *OpTy = cast<VectorType>(Val: Op->getType()); |
| 1345 | if (OpTy->getElementCount().getKnownMinValue() <= InsIdx) |
| 1346 | return false; |
| 1347 | // All inserts must have the same index. |
| 1348 | // TODO: Deal with mismatched index constants and variable indexes? |
| 1349 | if (!Index) |
| 1350 | Index = InsIdx; |
| 1351 | else if (InsIdx != *Index) |
| 1352 | return false; |
| 1353 | VecCs.push_back(Elt: VecC); |
| 1354 | ScalarOps.push_back(Elt: V); |
| 1355 | } else if (II && isVectorIntrinsicWithScalarOpAtArg(ID: II->getIntrinsicID(), |
| 1356 | ScalarOpdIdx: OpNum, TTI: &TTI)) { |
| 1357 | VecCs.push_back(Elt: Op.get()); |
| 1358 | ScalarOps.push_back(Elt: Op.get()); |
| 1359 | } else if (match(V: Op.get(), P: m_Constant(C&: VecC))) { |
| 1360 | VecCs.push_back(Elt: VecC); |
| 1361 | ScalarOps.push_back(Elt: nullptr); |
| 1362 | } else { |
| 1363 | return false; |
| 1364 | } |
| 1365 | } |
| 1366 | |
| 1367 | // Bail if all operands are constant. |
| 1368 | if (!Index.has_value()) |
| 1369 | return false; |
| 1370 | |
| 1371 | VectorType *VecTy = cast<VectorType>(Val: I.getType()); |
| 1372 | Type *ScalarTy = VecTy->getScalarType(); |
| 1373 | assert(VecTy->isVectorTy() && |
| 1374 | (ScalarTy->isIntegerTy() || ScalarTy->isFloatingPointTy() || |
| 1375 | ScalarTy->isPointerTy()) && |
| 1376 | "Unexpected types for insert element into binop or cmp" ); |
| 1377 | |
| 1378 | unsigned Opcode = I.getOpcode(); |
| 1379 | InstructionCost ScalarOpCost, VectorOpCost; |
| 1380 | if (CI) { |
| 1381 | CmpInst::Predicate Pred = CI->getPredicate(); |
| 1382 | ScalarOpCost = TTI.getCmpSelInstrCost( |
| 1383 | Opcode, ValTy: ScalarTy, CondTy: CmpInst::makeCmpResultType(opnd_type: ScalarTy), VecPred: Pred, CostKind); |
| 1384 | VectorOpCost = TTI.getCmpSelInstrCost( |
| 1385 | Opcode, ValTy: VecTy, CondTy: CmpInst::makeCmpResultType(opnd_type: VecTy), VecPred: Pred, CostKind); |
| 1386 | } else if (UO || BO) { |
| 1387 | ScalarOpCost = TTI.getArithmeticInstrCost(Opcode, Ty: ScalarTy, CostKind); |
| 1388 | VectorOpCost = TTI.getArithmeticInstrCost(Opcode, Ty: VecTy, CostKind); |
| 1389 | } else { |
| 1390 | IntrinsicCostAttributes ScalarICA( |
| 1391 | II->getIntrinsicID(), ScalarTy, |
| 1392 | SmallVector<Type *>(II->arg_size(), ScalarTy)); |
| 1393 | ScalarOpCost = TTI.getIntrinsicInstrCost(ICA: ScalarICA, CostKind); |
| 1394 | IntrinsicCostAttributes VectorICA( |
| 1395 | II->getIntrinsicID(), VecTy, |
| 1396 | SmallVector<Type *>(II->arg_size(), VecTy)); |
| 1397 | VectorOpCost = TTI.getIntrinsicInstrCost(ICA: VectorICA, CostKind); |
| 1398 | } |
| 1399 | |
| 1400 | // Fold the vector constants in the original vectors into a new base vector to |
| 1401 | // get more accurate cost modelling. |
| 1402 | Value *NewVecC = nullptr; |
| 1403 | if (CI) |
| 1404 | NewVecC = simplifyCmpInst(Predicate: CI->getPredicate(), LHS: VecCs[0], RHS: VecCs[1], Q: SQ); |
| 1405 | else if (UO) |
| 1406 | NewVecC = |
| 1407 | simplifyUnOp(Opcode: UO->getOpcode(), Op: VecCs[0], FMF: UO->getFastMathFlags(), Q: SQ); |
| 1408 | else if (BO) |
| 1409 | NewVecC = simplifyBinOp(Opcode: BO->getOpcode(), LHS: VecCs[0], RHS: VecCs[1], Q: SQ); |
| 1410 | else if (II) |
| 1411 | NewVecC = simplifyCall(Call: II, Callee: II->getCalledOperand(), Args: VecCs, Q: SQ); |
| 1412 | |
| 1413 | if (!NewVecC) |
| 1414 | return false; |
| 1415 | |
| 1416 | // Get cost estimate for the insert element. This cost will factor into |
| 1417 | // both sequences. |
| 1418 | InstructionCost OldCost = VectorOpCost; |
| 1419 | InstructionCost NewCost = |
| 1420 | ScalarOpCost + TTI.getVectorInstrCost(Opcode: Instruction::InsertElement, Val: VecTy, |
| 1421 | CostKind, Index: *Index, Op0: NewVecC); |
| 1422 | |
| 1423 | for (auto [Idx, Op, VecC, Scalar] : enumerate(First&: Ops, Rest&: VecCs, Rest&: ScalarOps)) { |
| 1424 | if (!Scalar || (II && isVectorIntrinsicWithScalarOpAtArg( |
| 1425 | ID: II->getIntrinsicID(), ScalarOpdIdx: Idx, TTI: &TTI))) |
| 1426 | continue; |
| 1427 | InstructionCost InsertCost = TTI.getVectorInstrCost( |
| 1428 | Opcode: Instruction::InsertElement, Val: VecTy, CostKind, Index: *Index, Op0: VecC, Op1: Scalar); |
| 1429 | OldCost += InsertCost; |
| 1430 | NewCost += !Op->hasOneUse() * InsertCost; |
| 1431 | } |
| 1432 | |
| 1433 | // We want to scalarize unless the vector variant actually has lower cost. |
| 1434 | if (OldCost < NewCost || !NewCost.isValid()) |
| 1435 | return false; |
| 1436 | |
| 1437 | // vec_op (inselt VecC0, V0, Index), (inselt VecC1, V1, Index) --> |
| 1438 | // inselt NewVecC, (scalar_op V0, V1), Index |
| 1439 | if (CI) |
| 1440 | ++NumScalarCmp; |
| 1441 | else if (UO || BO) |
| 1442 | ++NumScalarOps; |
| 1443 | else |
| 1444 | ++NumScalarIntrinsic; |
| 1445 | |
| 1446 | // For constant cases, extract the scalar element, this should constant fold. |
| 1447 | for (auto [OpIdx, Scalar, VecC] : enumerate(First&: ScalarOps, Rest&: VecCs)) |
| 1448 | if (!Scalar) |
| 1449 | ScalarOps[OpIdx] = ConstantExpr::getExtractElement( |
| 1450 | Vec: cast<Constant>(Val: VecC), Idx: Builder.getInt64(C: *Index)); |
| 1451 | |
| 1452 | Value *Scalar; |
| 1453 | if (CI) |
| 1454 | Scalar = Builder.CreateCmp(Pred: CI->getPredicate(), LHS: ScalarOps[0], RHS: ScalarOps[1]); |
| 1455 | else if (UO || BO) |
| 1456 | Scalar = Builder.CreateNAryOp(Opc: Opcode, Ops: ScalarOps); |
| 1457 | else |
| 1458 | Scalar = Builder.CreateIntrinsic(RetTy: ScalarTy, ID: II->getIntrinsicID(), Args: ScalarOps); |
| 1459 | |
| 1460 | Scalar->setName(I.getName() + ".scalar" ); |
| 1461 | |
| 1462 | // All IR flags are safe to back-propagate. There is no potential for extra |
| 1463 | // poison to be created by the scalar instruction. |
| 1464 | if (auto *ScalarInst = dyn_cast<Instruction>(Val: Scalar)) |
| 1465 | ScalarInst->copyIRFlags(V: &I); |
| 1466 | |
| 1467 | Value *Insert = Builder.CreateInsertElement(Vec: NewVecC, NewElt: Scalar, Idx: *Index); |
| 1468 | replaceValue(Old&: I, New&: *Insert); |
| 1469 | return true; |
| 1470 | } |
| 1471 | |
| 1472 | /// Try to combine a scalar binop + 2 scalar compares of extracted elements of |
| 1473 | /// a vector into vector operations followed by extract. Note: The SLP pass |
| 1474 | /// may miss this pattern because of implementation problems. |
| 1475 | bool VectorCombine::(Instruction &I) { |
| 1476 | auto *BI = dyn_cast<BinaryOperator>(Val: &I); |
| 1477 | |
| 1478 | // We are looking for a scalar binop of booleans. |
| 1479 | // binop i1 (cmp Pred I0, C0), (cmp Pred I1, C1) |
| 1480 | if (!BI || !I.getType()->isIntegerTy(BitWidth: 1)) |
| 1481 | return false; |
| 1482 | |
| 1483 | // The compare predicates should match, and each compare should have a |
| 1484 | // constant operand. |
| 1485 | Value *B0 = I.getOperand(i: 0), *B1 = I.getOperand(i: 1); |
| 1486 | Instruction *I0, *I1; |
| 1487 | Constant *C0, *C1; |
| 1488 | CmpPredicate P0, P1; |
| 1489 | if (!match(V: B0, P: m_Cmp(Pred&: P0, L: m_Instruction(I&: I0), R: m_Constant(C&: C0))) || |
| 1490 | !match(V: B1, P: m_Cmp(Pred&: P1, L: m_Instruction(I&: I1), R: m_Constant(C&: C1)))) |
| 1491 | return false; |
| 1492 | |
| 1493 | auto MatchingPred = CmpPredicate::getMatching(A: P0, B: P1); |
| 1494 | if (!MatchingPred) |
| 1495 | return false; |
| 1496 | |
| 1497 | // The compare operands must be extracts of the same vector with constant |
| 1498 | // extract indexes. |
| 1499 | Value *X; |
| 1500 | uint64_t Index0, Index1; |
| 1501 | if (!match(V: I0, P: m_ExtractElt(Val: m_Value(V&: X), Idx: m_ConstantInt(V&: Index0))) || |
| 1502 | !match(V: I1, P: m_ExtractElt(Val: m_Specific(V: X), Idx: m_ConstantInt(V&: Index1)))) |
| 1503 | return false; |
| 1504 | |
| 1505 | auto *Ext0 = cast<ExtractElementInst>(Val: I0); |
| 1506 | auto *Ext1 = cast<ExtractElementInst>(Val: I1); |
| 1507 | ExtractElementInst *ConvertToShuf = getShuffleExtract(Ext0, Ext1, PreferredExtractIndex: CostKind); |
| 1508 | if (!ConvertToShuf) |
| 1509 | return false; |
| 1510 | assert((ConvertToShuf == Ext0 || ConvertToShuf == Ext1) && |
| 1511 | "Unknown ExtractElementInst" ); |
| 1512 | |
| 1513 | // The original scalar pattern is: |
| 1514 | // binop i1 (cmp Pred (ext X, Index0), C0), (cmp Pred (ext X, Index1), C1) |
| 1515 | CmpInst::Predicate Pred = *MatchingPred; |
| 1516 | unsigned CmpOpcode = |
| 1517 | CmpInst::isFPPredicate(P: Pred) ? Instruction::FCmp : Instruction::ICmp; |
| 1518 | auto *VecTy = dyn_cast<FixedVectorType>(Val: X->getType()); |
| 1519 | if (!VecTy) |
| 1520 | return false; |
| 1521 | |
| 1522 | if (Index0 >= VecTy->getNumElements() || Index1 >= VecTy->getNumElements()) |
| 1523 | return false; |
| 1524 | |
| 1525 | InstructionCost Ext0Cost = |
| 1526 | TTI.getVectorInstrCost(I: *Ext0, Val: VecTy, CostKind, Index: Index0); |
| 1527 | InstructionCost Ext1Cost = |
| 1528 | TTI.getVectorInstrCost(I: *Ext1, Val: VecTy, CostKind, Index: Index1); |
| 1529 | InstructionCost CmpCost = TTI.getCmpSelInstrCost( |
| 1530 | Opcode: CmpOpcode, ValTy: I0->getType(), CondTy: CmpInst::makeCmpResultType(opnd_type: I0->getType()), VecPred: Pred, |
| 1531 | CostKind); |
| 1532 | |
| 1533 | InstructionCost OldCost = |
| 1534 | Ext0Cost + Ext1Cost + CmpCost * 2 + |
| 1535 | TTI.getArithmeticInstrCost(Opcode: I.getOpcode(), Ty: I.getType(), CostKind); |
| 1536 | |
| 1537 | // The proposed vector pattern is: |
| 1538 | // vcmp = cmp Pred X, VecC |
| 1539 | // ext (binop vNi1 vcmp, (shuffle vcmp, Index1)), Index0 |
| 1540 | int CheapIndex = ConvertToShuf == Ext0 ? Index1 : Index0; |
| 1541 | int ExpensiveIndex = ConvertToShuf == Ext0 ? Index0 : Index1; |
| 1542 | auto *CmpTy = cast<FixedVectorType>(Val: CmpInst::makeCmpResultType(opnd_type: VecTy)); |
| 1543 | InstructionCost NewCost = TTI.getCmpSelInstrCost( |
| 1544 | Opcode: CmpOpcode, ValTy: VecTy, CondTy: CmpInst::makeCmpResultType(opnd_type: VecTy), VecPred: Pred, CostKind); |
| 1545 | SmallVector<int, 32> ShufMask(VecTy->getNumElements(), PoisonMaskElem); |
| 1546 | ShufMask[CheapIndex] = ExpensiveIndex; |
| 1547 | NewCost += TTI.getShuffleCost(Kind: TargetTransformInfo::SK_PermuteSingleSrc, DstTy: CmpTy, |
| 1548 | SrcTy: CmpTy, Mask: ShufMask, CostKind); |
| 1549 | NewCost += TTI.getArithmeticInstrCost(Opcode: I.getOpcode(), Ty: CmpTy, CostKind); |
| 1550 | NewCost += TTI.getVectorInstrCost(I: *Ext0, Val: CmpTy, CostKind, Index: CheapIndex); |
| 1551 | NewCost += Ext0->hasOneUse() ? 0 : Ext0Cost; |
| 1552 | NewCost += Ext1->hasOneUse() ? 0 : Ext1Cost; |
| 1553 | |
| 1554 | // Aggressively form vector ops if the cost is equal because the transform |
| 1555 | // may enable further optimization. |
| 1556 | // Codegen can reverse this transform (scalarize) if it was not profitable. |
| 1557 | if (OldCost < NewCost || !NewCost.isValid()) |
| 1558 | return false; |
| 1559 | |
| 1560 | // Create a vector constant from the 2 scalar constants. |
| 1561 | SmallVector<Constant *, 32> CmpC(VecTy->getNumElements(), |
| 1562 | PoisonValue::get(T: VecTy->getElementType())); |
| 1563 | CmpC[Index0] = C0; |
| 1564 | CmpC[Index1] = C1; |
| 1565 | Value *VCmp = Builder.CreateCmp(Pred, LHS: X, RHS: ConstantVector::get(V: CmpC)); |
| 1566 | Value *Shuf = createShiftShuffle(Vec: VCmp, OldIndex: ExpensiveIndex, NewIndex: CheapIndex, Builder); |
| 1567 | Value *LHS = ConvertToShuf == Ext0 ? Shuf : VCmp; |
| 1568 | Value *RHS = ConvertToShuf == Ext0 ? VCmp : Shuf; |
| 1569 | Value *VecLogic = Builder.CreateBinOp(Opc: BI->getOpcode(), LHS, RHS); |
| 1570 | Value *NewExt = Builder.CreateExtractElement(Vec: VecLogic, Idx: CheapIndex); |
| 1571 | replaceValue(Old&: I, New&: *NewExt); |
| 1572 | ++NumVecCmpBO; |
| 1573 | return true; |
| 1574 | } |
| 1575 | |
| 1576 | /// Try to fold scalar selects that select between extracted elements and zero |
| 1577 | /// into extracting from a vector select. This is rooted at the bitcast. |
| 1578 | /// |
| 1579 | /// This pattern arises when a vector is bitcast to a smaller element type, |
| 1580 | /// elements are extracted, and then conditionally selected with zero: |
| 1581 | /// |
| 1582 | /// %bc = bitcast <4 x i32> %src to <16 x i8> |
| 1583 | /// %e0 = extractelement <16 x i8> %bc, i32 0 |
| 1584 | /// %s0 = select i1 %cond, i8 %e0, i8 0 |
| 1585 | /// %e1 = extractelement <16 x i8> %bc, i32 1 |
| 1586 | /// %s1 = select i1 %cond, i8 %e1, i8 0 |
| 1587 | /// ... |
| 1588 | /// |
| 1589 | /// Transforms to: |
| 1590 | /// %sel = select i1 %cond, <4 x i32> %src, <4 x i32> zeroinitializer |
| 1591 | /// %bc = bitcast <4 x i32> %sel to <16 x i8> |
| 1592 | /// %e0 = extractelement <16 x i8> %bc, i32 0 |
| 1593 | /// %e1 = extractelement <16 x i8> %bc, i32 1 |
| 1594 | /// ... |
| 1595 | /// |
| 1596 | /// This is profitable because vector select on wider types produces fewer |
| 1597 | /// select/cndmask instructions than scalar selects on each element. |
| 1598 | bool VectorCombine::foldSelectsFromBitcast(Instruction &I) { |
| 1599 | auto *BC = dyn_cast<BitCastInst>(Val: &I); |
| 1600 | if (!BC) |
| 1601 | return false; |
| 1602 | |
| 1603 | FixedVectorType *SrcVecTy = dyn_cast<FixedVectorType>(Val: BC->getSrcTy()); |
| 1604 | FixedVectorType *DstVecTy = dyn_cast<FixedVectorType>(Val: BC->getDestTy()); |
| 1605 | if (!SrcVecTy || !DstVecTy) |
| 1606 | return false; |
| 1607 | |
| 1608 | // Source must be 32-bit or 64-bit elements, destination must be smaller |
| 1609 | // integer elements. Zero in all these types is all-bits-zero. |
| 1610 | Type *SrcEltTy = SrcVecTy->getElementType(); |
| 1611 | Type *DstEltTy = DstVecTy->getElementType(); |
| 1612 | unsigned SrcEltBits = SrcEltTy->getPrimitiveSizeInBits(); |
| 1613 | unsigned DstEltBits = DstEltTy->getPrimitiveSizeInBits(); |
| 1614 | |
| 1615 | if (SrcEltBits != 32 && SrcEltBits != 64) |
| 1616 | return false; |
| 1617 | |
| 1618 | if (!DstEltTy->isIntegerTy() || DstEltBits >= SrcEltBits) |
| 1619 | return false; |
| 1620 | |
| 1621 | // Check profitability using TTI before collecting users. |
| 1622 | Type *CondTy = CmpInst::makeCmpResultType(opnd_type: DstEltTy); |
| 1623 | Type *VecCondTy = CmpInst::makeCmpResultType(opnd_type: SrcVecTy); |
| 1624 | |
| 1625 | InstructionCost ScalarSelCost = |
| 1626 | TTI.getCmpSelInstrCost(Opcode: Instruction::Select, ValTy: DstEltTy, CondTy, |
| 1627 | VecPred: CmpInst::BAD_ICMP_PREDICATE, CostKind); |
| 1628 | InstructionCost VecSelCost = |
| 1629 | TTI.getCmpSelInstrCost(Opcode: Instruction::Select, ValTy: SrcVecTy, CondTy: VecCondTy, |
| 1630 | VecPred: CmpInst::BAD_ICMP_PREDICATE, CostKind); |
| 1631 | |
| 1632 | // We need at least this many selects for vectorization to be profitable. |
| 1633 | // VecSelCost < ScalarSelCost * NumSelects => NumSelects > VecSelCost / |
| 1634 | // ScalarSelCost |
| 1635 | if (!ScalarSelCost.isValid() || ScalarSelCost == 0) |
| 1636 | return false; |
| 1637 | |
| 1638 | unsigned MinSelects = (VecSelCost.getValue() / ScalarSelCost.getValue()) + 1; |
| 1639 | |
| 1640 | // Quick check: if bitcast doesn't have enough users, bail early. |
| 1641 | if (!BC->hasNUsesOrMore(N: MinSelects)) |
| 1642 | return false; |
| 1643 | |
| 1644 | // Collect all select users that match the pattern, grouped by condition. |
| 1645 | // Pattern: select i1 %cond, (extractelement %bc, idx), 0 |
| 1646 | DenseMap<Value *, SmallVector<SelectInst *, 8>> CondToSelects; |
| 1647 | |
| 1648 | for (User *U : BC->users()) { |
| 1649 | auto *Ext = dyn_cast<ExtractElementInst>(Val: U); |
| 1650 | if (!Ext) |
| 1651 | continue; |
| 1652 | |
| 1653 | for (User *ExtUser : Ext->users()) { |
| 1654 | Value *Cond; |
| 1655 | // Match: select i1 %cond, %ext, 0 |
| 1656 | if (match(V: ExtUser, P: m_Select(C: m_Value(V&: Cond), L: m_Specific(V: Ext), R: m_Zero())) && |
| 1657 | Cond->getType()->isIntegerTy(BitWidth: 1)) |
| 1658 | CondToSelects[Cond].push_back(Elt: cast<SelectInst>(Val: ExtUser)); |
| 1659 | } |
| 1660 | } |
| 1661 | |
| 1662 | if (CondToSelects.empty()) |
| 1663 | return false; |
| 1664 | |
| 1665 | bool MadeChange = false; |
| 1666 | Value *SrcVec = BC->getOperand(i_nocapture: 0); |
| 1667 | |
| 1668 | // Process each group of selects with the same condition. |
| 1669 | for (auto [Cond, Selects] : CondToSelects) { |
| 1670 | // Only profitable if vector select cost < total scalar select cost. |
| 1671 | if (Selects.size() < MinSelects) { |
| 1672 | LLVM_DEBUG(dbgs() << "VectorCombine: foldSelectsFromBitcast not " |
| 1673 | << "profitable (VecCost=" << VecSelCost |
| 1674 | << ", ScalarCost=" << ScalarSelCost |
| 1675 | << ", NumSelects=" << Selects.size() << ")\n" ); |
| 1676 | continue; |
| 1677 | } |
| 1678 | |
| 1679 | // Create the vector select and bitcast once for this condition. |
| 1680 | auto InsertPt = std::next(x: BC->getIterator()); |
| 1681 | |
| 1682 | if (auto *CondInst = dyn_cast<Instruction>(Val: Cond)) |
| 1683 | if (DT.dominates(Def: BC, User: CondInst)) |
| 1684 | InsertPt = std::next(x: CondInst->getIterator()); |
| 1685 | |
| 1686 | Builder.SetInsertPoint(InsertPt); |
| 1687 | Value *VecSel = |
| 1688 | Builder.CreateSelect(C: Cond, True: SrcVec, False: Constant::getNullValue(Ty: SrcVecTy)); |
| 1689 | Value *NewBC = Builder.CreateBitCast(V: VecSel, DestTy: DstVecTy); |
| 1690 | |
| 1691 | // Replace each scalar select with an extract from the new bitcast. |
| 1692 | for (SelectInst *Sel : Selects) { |
| 1693 | auto *Ext = cast<ExtractElementInst>(Val: Sel->getTrueValue()); |
| 1694 | Value *Idx = Ext->getIndexOperand(); |
| 1695 | |
| 1696 | Builder.SetInsertPoint(Sel); |
| 1697 | Value *NewExt = Builder.CreateExtractElement(Vec: NewBC, Idx); |
| 1698 | replaceValue(Old&: *Sel, New&: *NewExt); |
| 1699 | MadeChange = true; |
| 1700 | } |
| 1701 | |
| 1702 | LLVM_DEBUG(dbgs() << "VectorCombine: folded " << Selects.size() |
| 1703 | << " selects into vector select\n" ); |
| 1704 | } |
| 1705 | |
| 1706 | return MadeChange; |
| 1707 | } |
| 1708 | |
| 1709 | static void analyzeCostOfVecReduction(const IntrinsicInst &II, |
| 1710 | TTI::TargetCostKind CostKind, |
| 1711 | const TargetTransformInfo &TTI, |
| 1712 | InstructionCost &CostBeforeReduction, |
| 1713 | InstructionCost &CostAfterReduction) { |
| 1714 | Instruction *Op0, *Op1; |
| 1715 | auto *RedOp = dyn_cast<Instruction>(Val: II.getOperand(i_nocapture: 0)); |
| 1716 | auto *VecRedTy = cast<VectorType>(Val: II.getOperand(i_nocapture: 0)->getType()); |
| 1717 | unsigned ReductionOpc = |
| 1718 | getArithmeticReductionInstruction(RdxID: II.getIntrinsicID()); |
| 1719 | if (RedOp && match(V: RedOp, P: m_ZExtOrSExt(Op: m_Value()))) { |
| 1720 | bool IsUnsigned = isa<ZExtInst>(Val: RedOp); |
| 1721 | auto *ExtType = cast<VectorType>(Val: RedOp->getOperand(i: 0)->getType()); |
| 1722 | |
| 1723 | CostBeforeReduction = |
| 1724 | TTI.getCastInstrCost(Opcode: RedOp->getOpcode(), Dst: VecRedTy, Src: ExtType, |
| 1725 | CCH: TTI::CastContextHint::None, CostKind, I: RedOp); |
| 1726 | CostAfterReduction = |
| 1727 | TTI.getExtendedReductionCost(Opcode: ReductionOpc, IsUnsigned, ResTy: II.getType(), |
| 1728 | Ty: ExtType, FMF: FastMathFlags(), CostKind); |
| 1729 | return; |
| 1730 | } |
| 1731 | if (RedOp && II.getIntrinsicID() == Intrinsic::vector_reduce_add && |
| 1732 | match(V: RedOp, |
| 1733 | P: m_ZExtOrSExt(Op: m_Mul(L: m_Instruction(I&: Op0), R: m_Instruction(I&: Op1)))) && |
| 1734 | match(V: Op0, P: m_ZExtOrSExt(Op: m_Value())) && |
| 1735 | Op0->getOpcode() == Op1->getOpcode() && |
| 1736 | Op0->getOperand(i: 0)->getType() == Op1->getOperand(i: 0)->getType() && |
| 1737 | (Op0->getOpcode() == RedOp->getOpcode() || Op0 == Op1)) { |
| 1738 | // Matched reduce.add(ext(mul(ext(A), ext(B))) |
| 1739 | bool IsUnsigned = isa<ZExtInst>(Val: Op0); |
| 1740 | auto *ExtType = cast<VectorType>(Val: Op0->getOperand(i: 0)->getType()); |
| 1741 | VectorType *MulType = VectorType::get(ElementType: Op0->getType(), Other: VecRedTy); |
| 1742 | |
| 1743 | InstructionCost ExtCost = |
| 1744 | TTI.getCastInstrCost(Opcode: Op0->getOpcode(), Dst: MulType, Src: ExtType, |
| 1745 | CCH: TTI::CastContextHint::None, CostKind, I: Op0); |
| 1746 | InstructionCost MulCost = |
| 1747 | TTI.getArithmeticInstrCost(Opcode: Instruction::Mul, Ty: MulType, CostKind); |
| 1748 | InstructionCost Ext2Cost = |
| 1749 | TTI.getCastInstrCost(Opcode: RedOp->getOpcode(), Dst: VecRedTy, Src: MulType, |
| 1750 | CCH: TTI::CastContextHint::None, CostKind, I: RedOp); |
| 1751 | |
| 1752 | CostBeforeReduction = ExtCost * 2 + MulCost + Ext2Cost; |
| 1753 | CostAfterReduction = TTI.getMulAccReductionCost( |
| 1754 | IsUnsigned, RedOpcode: ReductionOpc, ResTy: II.getType(), Ty: ExtType, CostKind); |
| 1755 | return; |
| 1756 | } |
| 1757 | CostAfterReduction = TTI.getArithmeticReductionCost(Opcode: ReductionOpc, Ty: VecRedTy, |
| 1758 | FMF: std::nullopt, CostKind); |
| 1759 | } |
| 1760 | |
| 1761 | bool VectorCombine::foldBinopOfReductions(Instruction &I) { |
| 1762 | Instruction::BinaryOps BinOpOpc = cast<BinaryOperator>(Val: &I)->getOpcode(); |
| 1763 | Intrinsic::ID ReductionIID = getReductionForBinop(Opc: BinOpOpc); |
| 1764 | if (BinOpOpc == Instruction::Sub) |
| 1765 | ReductionIID = Intrinsic::vector_reduce_add; |
| 1766 | if (ReductionIID == Intrinsic::not_intrinsic) |
| 1767 | return false; |
| 1768 | // FP reductions have a start-value operand that this fold doesn't handle. |
| 1769 | if (ReductionIID == Intrinsic::vector_reduce_fadd || |
| 1770 | ReductionIID == Intrinsic::vector_reduce_fmul) |
| 1771 | return false; |
| 1772 | |
| 1773 | auto checkIntrinsicAndGetItsArgument = [](Value *V, |
| 1774 | Intrinsic::ID IID) -> Value * { |
| 1775 | auto *II = dyn_cast<IntrinsicInst>(Val: V); |
| 1776 | if (!II) |
| 1777 | return nullptr; |
| 1778 | if (II->getIntrinsicID() == IID && II->hasOneUse()) |
| 1779 | return II->getArgOperand(i: 0); |
| 1780 | return nullptr; |
| 1781 | }; |
| 1782 | |
| 1783 | Value *V0 = checkIntrinsicAndGetItsArgument(I.getOperand(i: 0), ReductionIID); |
| 1784 | if (!V0) |
| 1785 | return false; |
| 1786 | Value *V1 = checkIntrinsicAndGetItsArgument(I.getOperand(i: 1), ReductionIID); |
| 1787 | if (!V1) |
| 1788 | return false; |
| 1789 | |
| 1790 | auto *VTy = cast<VectorType>(Val: V0->getType()); |
| 1791 | if (V1->getType() != VTy) |
| 1792 | return false; |
| 1793 | const auto &II0 = *cast<IntrinsicInst>(Val: I.getOperand(i: 0)); |
| 1794 | const auto &II1 = *cast<IntrinsicInst>(Val: I.getOperand(i: 1)); |
| 1795 | unsigned ReductionOpc = |
| 1796 | getArithmeticReductionInstruction(RdxID: II0.getIntrinsicID()); |
| 1797 | |
| 1798 | InstructionCost OldCost = 0; |
| 1799 | InstructionCost NewCost = 0; |
| 1800 | InstructionCost CostOfRedOperand0 = 0; |
| 1801 | InstructionCost CostOfRed0 = 0; |
| 1802 | InstructionCost CostOfRedOperand1 = 0; |
| 1803 | InstructionCost CostOfRed1 = 0; |
| 1804 | analyzeCostOfVecReduction(II: II0, CostKind, TTI, CostBeforeReduction&: CostOfRedOperand0, CostAfterReduction&: CostOfRed0); |
| 1805 | analyzeCostOfVecReduction(II: II1, CostKind, TTI, CostBeforeReduction&: CostOfRedOperand1, CostAfterReduction&: CostOfRed1); |
| 1806 | OldCost = CostOfRed0 + CostOfRed1 + TTI.getInstructionCost(U: &I, CostKind); |
| 1807 | NewCost = |
| 1808 | CostOfRedOperand0 + CostOfRedOperand1 + |
| 1809 | TTI.getArithmeticInstrCost(Opcode: BinOpOpc, Ty: VTy, CostKind) + |
| 1810 | TTI.getArithmeticReductionCost(Opcode: ReductionOpc, Ty: VTy, FMF: std::nullopt, CostKind); |
| 1811 | if (NewCost >= OldCost || !NewCost.isValid()) |
| 1812 | return false; |
| 1813 | |
| 1814 | LLVM_DEBUG(dbgs() << "Found two mergeable reductions: " << I |
| 1815 | << "\n OldCost: " << OldCost << " vs NewCost: " << NewCost |
| 1816 | << "\n" ); |
| 1817 | Value *VectorBO; |
| 1818 | if (BinOpOpc == Instruction::Or) |
| 1819 | VectorBO = Builder.CreateOr(LHS: V0, RHS: V1, Name: "" , |
| 1820 | IsDisjoint: cast<PossiblyDisjointInst>(Val&: I).isDisjoint()); |
| 1821 | else |
| 1822 | VectorBO = Builder.CreateBinOp(Opc: BinOpOpc, LHS: V0, RHS: V1); |
| 1823 | |
| 1824 | Value *Rdx = Builder.CreateIntrinsic(ID: ReductionIID, OverloadTypes: {VTy}, Args: {VectorBO}); |
| 1825 | replaceValue(Old&: I, New&: *Rdx); |
| 1826 | return true; |
| 1827 | } |
| 1828 | |
| 1829 | // Check if memory loc modified between two instrs in the same BB |
| 1830 | static bool isMemModifiedBetween(BasicBlock::iterator Begin, |
| 1831 | BasicBlock::iterator End, |
| 1832 | const MemoryLocation &Loc, AAResults &AA) { |
| 1833 | unsigned NumScanned = 0; |
| 1834 | return std::any_of(first: Begin, last: End, pred: [&](const Instruction &Instr) { |
| 1835 | return isModSet(MRI: AA.getModRefInfo(I: &Instr, OptLoc: Loc)) || |
| 1836 | ++NumScanned > MaxInstrsToScan; |
| 1837 | }); |
| 1838 | } |
| 1839 | |
| 1840 | namespace { |
| 1841 | /// Helper class to indicate whether a vector index can be safely scalarized and |
| 1842 | /// if a freeze needs to be inserted. |
| 1843 | class ScalarizationResult { |
| 1844 | enum class StatusTy { Unsafe, Safe, SafeWithFreeze }; |
| 1845 | |
| 1846 | StatusTy Status; |
| 1847 | Value *ToFreeze; |
| 1848 | |
| 1849 | ScalarizationResult(StatusTy Status, Value *ToFreeze = nullptr) |
| 1850 | : Status(Status), ToFreeze(ToFreeze) {} |
| 1851 | |
| 1852 | public: |
| 1853 | ScalarizationResult(const ScalarizationResult &Other) = default; |
| 1854 | ~ScalarizationResult() { |
| 1855 | assert(!ToFreeze && "freeze() not called with ToFreeze being set" ); |
| 1856 | } |
| 1857 | |
| 1858 | static ScalarizationResult unsafe() { return {StatusTy::Unsafe}; } |
| 1859 | static ScalarizationResult safe() { return {StatusTy::Safe}; } |
| 1860 | static ScalarizationResult safeWithFreeze(Value *ToFreeze) { |
| 1861 | return {StatusTy::SafeWithFreeze, ToFreeze}; |
| 1862 | } |
| 1863 | |
| 1864 | /// Returns true if the index can be scalarize without requiring a freeze. |
| 1865 | bool isSafe() const { return Status == StatusTy::Safe; } |
| 1866 | /// Returns true if the index cannot be scalarized. |
| 1867 | bool isUnsafe() const { return Status == StatusTy::Unsafe; } |
| 1868 | /// Returns true if the index can be scalarize, but requires inserting a |
| 1869 | /// freeze. |
| 1870 | bool isSafeWithFreeze() const { return Status == StatusTy::SafeWithFreeze; } |
| 1871 | |
| 1872 | /// Reset the state of Unsafe and clear ToFreze if set. |
| 1873 | void discard() { |
| 1874 | ToFreeze = nullptr; |
| 1875 | Status = StatusTy::Unsafe; |
| 1876 | } |
| 1877 | |
| 1878 | /// Freeze the ToFreeze and update the use in \p User to use it. |
| 1879 | void freeze(IRBuilderBase &Builder, Instruction &UserI) { |
| 1880 | assert(isSafeWithFreeze() && |
| 1881 | "should only be used when freezing is required" ); |
| 1882 | assert(is_contained(ToFreeze->users(), &UserI) && |
| 1883 | "UserI must be a user of ToFreeze" ); |
| 1884 | IRBuilder<>::InsertPointGuard Guard(Builder); |
| 1885 | Builder.SetInsertPoint(cast<Instruction>(Val: &UserI)); |
| 1886 | Value *Frozen = |
| 1887 | Builder.CreateFreeze(V: ToFreeze, Name: ToFreeze->getName() + ".frozen" ); |
| 1888 | for (Use &U : make_early_inc_range(Range: (UserI.operands()))) |
| 1889 | if (U.get() == ToFreeze) |
| 1890 | U.set(Frozen); |
| 1891 | |
| 1892 | ToFreeze = nullptr; |
| 1893 | } |
| 1894 | }; |
| 1895 | } // namespace |
| 1896 | |
| 1897 | /// Check if it is legal to scalarize a memory access to \p VecTy at index \p |
| 1898 | /// Idx. \p Idx must access a valid vector element. |
| 1899 | static ScalarizationResult canScalarizeAccess(VectorType *VecTy, Value *Idx, |
| 1900 | const SimplifyQuery &SQ) { |
| 1901 | // We do checks for both fixed vector types and scalable vector types. |
| 1902 | // This is the number of elements of fixed vector types, |
| 1903 | // or the minimum number of elements of scalable vector types. |
| 1904 | uint64_t NumElements = VecTy->getElementCount().getKnownMinValue(); |
| 1905 | unsigned IntWidth = Idx->getType()->getScalarSizeInBits(); |
| 1906 | |
| 1907 | if (auto *C = dyn_cast<ConstantInt>(Val: Idx)) { |
| 1908 | if (C->getValue().ult(RHS: NumElements)) |
| 1909 | return ScalarizationResult::safe(); |
| 1910 | return ScalarizationResult::unsafe(); |
| 1911 | } |
| 1912 | |
| 1913 | // Always unsafe if the index type can't handle all inbound values. |
| 1914 | if (!llvm::isUIntN(N: IntWidth, x: NumElements)) |
| 1915 | return ScalarizationResult::unsafe(); |
| 1916 | |
| 1917 | APInt Zero(IntWidth, 0); |
| 1918 | APInt MaxElts(IntWidth, NumElements); |
| 1919 | ConstantRange ValidIndices(Zero, MaxElts); |
| 1920 | ConstantRange IdxRange(IntWidth, true); |
| 1921 | |
| 1922 | if (isGuaranteedNotToBePoison(V: Idx, AC: SQ.AC, CtxI: SQ.CxtI, DT: SQ.DT)) { |
| 1923 | if (ValidIndices.contains( |
| 1924 | CR: computeConstantRange(V: Idx, /*ForSigned=*/false, SQ))) |
| 1925 | return ScalarizationResult::safe(); |
| 1926 | return ScalarizationResult::unsafe(); |
| 1927 | } |
| 1928 | |
| 1929 | // If the index may be poison, check if we can insert a freeze before the |
| 1930 | // range of the index is restricted. |
| 1931 | Value *IdxBase; |
| 1932 | ConstantInt *CI; |
| 1933 | if (match(V: Idx, P: m_And(L: m_Value(V&: IdxBase), R: m_ConstantInt(CI)))) { |
| 1934 | IdxRange = IdxRange.binaryAnd(Other: CI->getValue()); |
| 1935 | } else if (match(V: Idx, P: m_URem(L: m_Value(V&: IdxBase), R: m_ConstantInt(CI)))) { |
| 1936 | IdxRange = IdxRange.urem(Other: CI->getValue()); |
| 1937 | } |
| 1938 | |
| 1939 | if (ValidIndices.contains(CR: IdxRange)) |
| 1940 | return ScalarizationResult::safeWithFreeze(ToFreeze: IdxBase); |
| 1941 | return ScalarizationResult::unsafe(); |
| 1942 | } |
| 1943 | |
| 1944 | /// The memory operation on a vector of \p ScalarType had alignment of |
| 1945 | /// \p VectorAlignment. Compute the maximal, but conservatively correct, |
| 1946 | /// alignment that will be valid for the memory operation on a single scalar |
| 1947 | /// element of the same type with index \p Idx. |
| 1948 | static Align computeAlignmentAfterScalarization(Align VectorAlignment, |
| 1949 | Type *ScalarType, Value *Idx, |
| 1950 | const DataLayout &DL) { |
| 1951 | if (auto *C = dyn_cast<ConstantInt>(Val: Idx)) |
| 1952 | return commonAlignment(A: VectorAlignment, |
| 1953 | Offset: C->getZExtValue() * DL.getTypeStoreSize(Ty: ScalarType)); |
| 1954 | return commonAlignment(A: VectorAlignment, Offset: DL.getTypeStoreSize(Ty: ScalarType)); |
| 1955 | } |
| 1956 | |
| 1957 | // Combine patterns like: |
| 1958 | // %0 = load <4 x i32>, <4 x i32>* %a |
| 1959 | // %1 = insertelement <4 x i32> %0, i32 %b, i32 1 |
| 1960 | // store <4 x i32> %1, <4 x i32>* %a |
| 1961 | // to: |
| 1962 | // %0 = bitcast <4 x i32>* %a to i32* |
| 1963 | // %1 = getelementptr inbounds i32, i32* %0, i64 0, i64 1 |
| 1964 | // store i32 %b, i32* %1 |
| 1965 | bool VectorCombine::foldSingleElementStore(Instruction &I) { |
| 1966 | if (!TTI.allowVectorElementIndexingUsingGEP()) |
| 1967 | return false; |
| 1968 | auto *SI = cast<StoreInst>(Val: &I); |
| 1969 | if (!SI->isSimple() || !isa<VectorType>(Val: SI->getValueOperand()->getType())) |
| 1970 | return false; |
| 1971 | |
| 1972 | // TODO: Combine more complicated patterns (multiple insert) by referencing |
| 1973 | // TargetTransformInfo. |
| 1974 | Instruction *Source; |
| 1975 | Value *NewElement; |
| 1976 | Value *Idx; |
| 1977 | if (!match(V: SI->getValueOperand(), |
| 1978 | P: m_InsertElt(Val: m_Instruction(I&: Source), Elt: m_Value(V&: NewElement), |
| 1979 | Idx: m_Value(V&: Idx)))) |
| 1980 | return false; |
| 1981 | |
| 1982 | if (auto *Load = dyn_cast<LoadInst>(Val: Source)) { |
| 1983 | auto VecTy = cast<VectorType>(Val: SI->getValueOperand()->getType()); |
| 1984 | Value *SrcAddr = Load->getPointerOperand()->stripPointerCasts(); |
| 1985 | // Don't optimize for atomic/volatile load or store. Ensure memory is not |
| 1986 | // modified between, vector type matches store size, and index is inbounds. |
| 1987 | if (!Load->isSimple() || Load->getParent() != SI->getParent() || |
| 1988 | !DL->typeSizeEqualsStoreSize(Ty: Load->getType()->getScalarType()) || |
| 1989 | SrcAddr != SI->getPointerOperand()->stripPointerCasts()) |
| 1990 | return false; |
| 1991 | |
| 1992 | if (isMemModifiedBetween(Begin: Load->getIterator(), End: SI->getIterator(), |
| 1993 | Loc: MemoryLocation::get(SI), AA)) |
| 1994 | return false; |
| 1995 | auto ScalarizableIdx = |
| 1996 | canScalarizeAccess(VecTy, Idx, SQ: SQ.getWithInstruction(I: Load)); |
| 1997 | if (ScalarizableIdx.isUnsafe()) |
| 1998 | return false; |
| 1999 | |
| 2000 | // Ensure we add the load back to the worklist BEFORE its users so they can |
| 2001 | // erased in the correct order. |
| 2002 | Worklist.push(I: Load); |
| 2003 | |
| 2004 | if (ScalarizableIdx.isSafeWithFreeze()) |
| 2005 | ScalarizableIdx.freeze(Builder, UserI&: *cast<Instruction>(Val: Idx)); |
| 2006 | Value *GEP = Builder.CreateInBoundsGEP( |
| 2007 | Ty: SI->getValueOperand()->getType(), Ptr: SI->getPointerOperand(), |
| 2008 | IdxList: {ConstantInt::get(Ty: Idx->getType(), V: 0), Idx}); |
| 2009 | StoreInst *NSI = Builder.CreateStore(Val: NewElement, Ptr: GEP); |
| 2010 | NSI->copyMetadata(SrcInst: *SI); |
| 2011 | Align ScalarOpAlignment = computeAlignmentAfterScalarization( |
| 2012 | VectorAlignment: std::max(a: SI->getAlign(), b: Load->getAlign()), ScalarType: NewElement->getType(), Idx, |
| 2013 | DL: *DL); |
| 2014 | NSI->setAlignment(ScalarOpAlignment); |
| 2015 | replaceValue(Old&: I, New&: *NSI); |
| 2016 | eraseInstruction(I); |
| 2017 | return true; |
| 2018 | } |
| 2019 | |
| 2020 | return false; |
| 2021 | } |
| 2022 | |
| 2023 | /// Try to scalarize vector loads feeding extractelement or bitcast |
| 2024 | /// instructions. |
| 2025 | bool VectorCombine::scalarizeLoad(Instruction &I) { |
| 2026 | Value *Ptr; |
| 2027 | if (!match(V: &I, P: m_Load(Op: m_Value(V&: Ptr)))) |
| 2028 | return false; |
| 2029 | |
| 2030 | auto *LI = cast<LoadInst>(Val: &I); |
| 2031 | auto *VecTy = cast<VectorType>(Val: LI->getType()); |
| 2032 | |
| 2033 | // The isSimple() check could be isUnordered(), but for now we cowardly |
| 2034 | // refuse to handle even unordered atomics. |
| 2035 | if (!LI->isSimple() || !DL->typeSizeEqualsStoreSize(Ty: VecTy->getScalarType())) |
| 2036 | return false; |
| 2037 | |
| 2038 | bool = true; |
| 2039 | bool AllBitcasts = true; |
| 2040 | Instruction *LastCheckedInst = LI; |
| 2041 | unsigned NumInstChecked = 0; |
| 2042 | |
| 2043 | // Check what type of users we have (must either all be extracts or |
| 2044 | // bitcasts) and ensure no memory modifications between the load and |
| 2045 | // its users. |
| 2046 | for (User *U : LI->users()) { |
| 2047 | auto *UI = dyn_cast<Instruction>(Val: U); |
| 2048 | if (!UI || UI->getParent() != LI->getParent()) |
| 2049 | return false; |
| 2050 | |
| 2051 | // If any user is waiting to be erased, then bail out as this will |
| 2052 | // distort the cost calculation and possibly lead to infinite loops. |
| 2053 | if (UI->use_empty()) |
| 2054 | return false; |
| 2055 | |
| 2056 | if (!isa<ExtractElementInst>(Val: UI)) |
| 2057 | AllExtracts = false; |
| 2058 | if (!isa<BitCastInst>(Val: UI)) |
| 2059 | AllBitcasts = false; |
| 2060 | |
| 2061 | // Check if any instruction between the load and the user may modify memory. |
| 2062 | if (LastCheckedInst->comesBefore(Other: UI)) { |
| 2063 | for (Instruction &I : |
| 2064 | make_range(x: std::next(x: LI->getIterator()), y: UI->getIterator())) { |
| 2065 | // Bail out if we reached the check limit or the instruction may write |
| 2066 | // to memory. |
| 2067 | if (NumInstChecked == MaxInstrsToScan || I.mayWriteToMemory()) |
| 2068 | return false; |
| 2069 | NumInstChecked++; |
| 2070 | } |
| 2071 | LastCheckedInst = UI; |
| 2072 | } |
| 2073 | } |
| 2074 | |
| 2075 | if (AllExtracts) |
| 2076 | return scalarizeLoadExtract(LI, VecTy, Ptr); |
| 2077 | if (AllBitcasts) |
| 2078 | return scalarizeLoadBitcast(LI, VecTy, Ptr); |
| 2079 | return false; |
| 2080 | } |
| 2081 | |
| 2082 | /// Try to scalarize vector loads feeding extractelement instructions. |
| 2083 | bool VectorCombine::(LoadInst *LI, VectorType *VecTy, |
| 2084 | Value *Ptr) { |
| 2085 | if (!TTI.allowVectorElementIndexingUsingGEP()) |
| 2086 | return false; |
| 2087 | |
| 2088 | DenseMap<ExtractElementInst *, ScalarizationResult> NeedFreeze; |
| 2089 | llvm::scope_exit FailureGuard([&]() { |
| 2090 | // If the transform is aborted, discard the ScalarizationResults. |
| 2091 | for (auto &Pair : NeedFreeze) |
| 2092 | Pair.second.discard(); |
| 2093 | }); |
| 2094 | |
| 2095 | InstructionCost OriginalCost = |
| 2096 | TTI.getMemoryOpCost(Opcode: Instruction::Load, Src: VecTy, Alignment: LI->getAlign(), |
| 2097 | AddressSpace: LI->getPointerAddressSpace(), CostKind); |
| 2098 | InstructionCost ScalarizedCost = 0; |
| 2099 | |
| 2100 | for (User *U : LI->users()) { |
| 2101 | auto *UI = cast<ExtractElementInst>(Val: U); |
| 2102 | |
| 2103 | auto ScalarIdx = canScalarizeAccess(VecTy, Idx: UI->getIndexOperand(), |
| 2104 | SQ: SQ.getWithInstruction(I: LI)); |
| 2105 | if (ScalarIdx.isUnsafe()) |
| 2106 | return false; |
| 2107 | if (ScalarIdx.isSafeWithFreeze()) { |
| 2108 | NeedFreeze.try_emplace(Key: UI, Args&: ScalarIdx); |
| 2109 | ScalarIdx.discard(); |
| 2110 | } |
| 2111 | |
| 2112 | auto *Index = dyn_cast<ConstantInt>(Val: UI->getIndexOperand()); |
| 2113 | OriginalCost += |
| 2114 | TTI.getVectorInstrCost(Opcode: Instruction::ExtractElement, Val: VecTy, CostKind, |
| 2115 | Index: Index ? Index->getZExtValue() : -1); |
| 2116 | ScalarizedCost += |
| 2117 | TTI.getMemoryOpCost(Opcode: Instruction::Load, Src: VecTy->getElementType(), |
| 2118 | Alignment: Align(1), AddressSpace: LI->getPointerAddressSpace(), CostKind); |
| 2119 | ScalarizedCost += TTI.getAddressComputationCost(PtrTy: LI->getPointerOperandType(), |
| 2120 | SE: nullptr, Ptr: nullptr, CostKind); |
| 2121 | } |
| 2122 | |
| 2123 | LLVM_DEBUG(dbgs() << "Found all extractions of a vector load: " << *LI |
| 2124 | << "\n LoadExtractCost: " << OriginalCost |
| 2125 | << " vs ScalarizedCost: " << ScalarizedCost << "\n" ); |
| 2126 | |
| 2127 | if (ScalarizedCost >= OriginalCost) |
| 2128 | return false; |
| 2129 | |
| 2130 | // Ensure we add the load back to the worklist BEFORE its users so they can |
| 2131 | // erased in the correct order. |
| 2132 | Worklist.push(I: LI); |
| 2133 | |
| 2134 | Type *ElemType = VecTy->getElementType(); |
| 2135 | |
| 2136 | // Replace extracts with narrow scalar loads. |
| 2137 | for (User *U : LI->users()) { |
| 2138 | auto *EI = cast<ExtractElementInst>(Val: U); |
| 2139 | Value *Idx = EI->getIndexOperand(); |
| 2140 | |
| 2141 | // Insert 'freeze' for poison indexes. |
| 2142 | auto It = NeedFreeze.find(Val: EI); |
| 2143 | if (It != NeedFreeze.end()) |
| 2144 | It->second.freeze(Builder, UserI&: *cast<Instruction>(Val: Idx)); |
| 2145 | |
| 2146 | Builder.SetInsertPoint(EI); |
| 2147 | Value *GEP = |
| 2148 | Builder.CreateInBoundsGEP(Ty: VecTy, Ptr, IdxList: {Builder.getInt32(C: 0), Idx}); |
| 2149 | auto *NewLoad = cast<LoadInst>( |
| 2150 | Val: Builder.CreateLoad(Ty: ElemType, Ptr: GEP, Name: EI->getName() + ".scalar" )); |
| 2151 | |
| 2152 | Align ScalarOpAlignment = |
| 2153 | computeAlignmentAfterScalarization(VectorAlignment: LI->getAlign(), ScalarType: ElemType, Idx, DL: *DL); |
| 2154 | NewLoad->setAlignment(ScalarOpAlignment); |
| 2155 | |
| 2156 | if (auto *ConstIdx = dyn_cast<ConstantInt>(Val: Idx)) { |
| 2157 | size_t Offset = ConstIdx->getZExtValue() * DL->getTypeStoreSize(Ty: ElemType); |
| 2158 | AAMDNodes OldAAMD = LI->getAAMetadata(); |
| 2159 | NewLoad->setAAMetadata(OldAAMD.adjustForAccess(Offset, AccessTy: ElemType, DL: *DL)); |
| 2160 | } |
| 2161 | |
| 2162 | replaceValue(Old&: *EI, New&: *NewLoad, Erase: false); |
| 2163 | } |
| 2164 | |
| 2165 | FailureGuard.release(); |
| 2166 | return true; |
| 2167 | } |
| 2168 | |
| 2169 | /// Try to scalarize vector loads feeding bitcast instructions. |
| 2170 | bool VectorCombine::scalarizeLoadBitcast(LoadInst *LI, VectorType *VecTy, |
| 2171 | Value *Ptr) { |
| 2172 | InstructionCost OriginalCost = |
| 2173 | TTI.getMemoryOpCost(Opcode: Instruction::Load, Src: VecTy, Alignment: LI->getAlign(), |
| 2174 | AddressSpace: LI->getPointerAddressSpace(), CostKind); |
| 2175 | |
| 2176 | Type *TargetScalarType = nullptr; |
| 2177 | unsigned VecBitWidth = DL->getTypeSizeInBits(Ty: VecTy); |
| 2178 | |
| 2179 | for (User *U : LI->users()) { |
| 2180 | auto *BC = cast<BitCastInst>(Val: U); |
| 2181 | |
| 2182 | Type *DestTy = BC->getDestTy(); |
| 2183 | if (!DestTy->isIntegerTy() && !DestTy->isFloatingPointTy()) |
| 2184 | return false; |
| 2185 | |
| 2186 | unsigned DestBitWidth = DL->getTypeSizeInBits(Ty: DestTy); |
| 2187 | if (DestBitWidth != VecBitWidth) |
| 2188 | return false; |
| 2189 | |
| 2190 | // All bitcasts must target the same scalar type. |
| 2191 | if (!TargetScalarType) |
| 2192 | TargetScalarType = DestTy; |
| 2193 | else if (TargetScalarType != DestTy) |
| 2194 | return false; |
| 2195 | |
| 2196 | OriginalCost += |
| 2197 | TTI.getCastInstrCost(Opcode: Instruction::BitCast, Dst: TargetScalarType, Src: VecTy, |
| 2198 | CCH: TTI.getCastContextHint(I: BC), CostKind, I: BC); |
| 2199 | } |
| 2200 | |
| 2201 | if (!TargetScalarType) |
| 2202 | return false; |
| 2203 | |
| 2204 | assert(!LI->user_empty() && "Unexpected load without bitcast users" ); |
| 2205 | InstructionCost ScalarizedCost = |
| 2206 | TTI.getMemoryOpCost(Opcode: Instruction::Load, Src: TargetScalarType, Alignment: LI->getAlign(), |
| 2207 | AddressSpace: LI->getPointerAddressSpace(), CostKind); |
| 2208 | |
| 2209 | LLVM_DEBUG(dbgs() << "Found vector load feeding only bitcasts: " << *LI |
| 2210 | << "\n OriginalCost: " << OriginalCost |
| 2211 | << " vs ScalarizedCost: " << ScalarizedCost << "\n" ); |
| 2212 | |
| 2213 | if (ScalarizedCost >= OriginalCost) |
| 2214 | return false; |
| 2215 | |
| 2216 | // Ensure we add the load back to the worklist BEFORE its users so they can |
| 2217 | // erased in the correct order. |
| 2218 | Worklist.push(I: LI); |
| 2219 | |
| 2220 | Builder.SetInsertPoint(LI); |
| 2221 | auto *ScalarLoad = |
| 2222 | Builder.CreateLoad(Ty: TargetScalarType, Ptr, Name: LI->getName() + ".scalar" ); |
| 2223 | ScalarLoad->setAlignment(LI->getAlign()); |
| 2224 | ScalarLoad->copyMetadata(SrcInst: *LI); |
| 2225 | |
| 2226 | // Replace all bitcast users with the scalar load. |
| 2227 | for (User *U : LI->users()) { |
| 2228 | auto *BC = cast<BitCastInst>(Val: U); |
| 2229 | replaceValue(Old&: *BC, New&: *ScalarLoad, Erase: false); |
| 2230 | } |
| 2231 | |
| 2232 | return true; |
| 2233 | } |
| 2234 | |
| 2235 | bool VectorCombine::(Instruction &I) { |
| 2236 | if (!TTI.allowVectorElementIndexingUsingGEP()) |
| 2237 | return false; |
| 2238 | auto *Ext = dyn_cast<ZExtInst>(Val: &I); |
| 2239 | if (!Ext) |
| 2240 | return false; |
| 2241 | |
| 2242 | // Try to convert a vector zext feeding only extracts to a set of scalar |
| 2243 | // (Src << ExtIdx *Size) & (Size -1) |
| 2244 | // if profitable . |
| 2245 | auto *SrcTy = dyn_cast<FixedVectorType>(Val: Ext->getOperand(i_nocapture: 0)->getType()); |
| 2246 | if (!SrcTy) |
| 2247 | return false; |
| 2248 | auto *DstTy = cast<FixedVectorType>(Val: Ext->getType()); |
| 2249 | |
| 2250 | Type *ScalarDstTy = DstTy->getElementType(); |
| 2251 | if (DL->getTypeSizeInBits(Ty: SrcTy) != DL->getTypeSizeInBits(Ty: ScalarDstTy)) |
| 2252 | return false; |
| 2253 | |
| 2254 | InstructionCost VectorCost = |
| 2255 | TTI.getCastInstrCost(Opcode: Instruction::ZExt, Dst: DstTy, Src: SrcTy, |
| 2256 | CCH: TTI::CastContextHint::None, CostKind, I: Ext); |
| 2257 | unsigned ExtCnt = 0; |
| 2258 | bool ExtLane0 = false; |
| 2259 | for (User *U : Ext->users()) { |
| 2260 | uint64_t Idx; |
| 2261 | if (!match(V: U, P: m_ExtractElt(Val: m_Value(), Idx: m_ConstantInt(V&: Idx)))) |
| 2262 | return false; |
| 2263 | if (cast<Instruction>(Val: U)->use_empty()) |
| 2264 | continue; |
| 2265 | ExtCnt += 1; |
| 2266 | ExtLane0 |= !Idx; |
| 2267 | VectorCost += TTI.getVectorInstrCost(Opcode: Instruction::ExtractElement, Val: DstTy, |
| 2268 | CostKind, Index: Idx, Op0: U); |
| 2269 | } |
| 2270 | |
| 2271 | InstructionCost ScalarCost = |
| 2272 | ExtCnt * TTI.getArithmeticInstrCost( |
| 2273 | Opcode: Instruction::And, Ty: ScalarDstTy, CostKind, |
| 2274 | Opd1Info: {.Kind: TTI::OK_AnyValue, .Properties: TTI::OP_None}, |
| 2275 | Opd2Info: {.Kind: TTI::OK_NonUniformConstantValue, .Properties: TTI::OP_None}) + |
| 2276 | (ExtCnt - ExtLane0) * |
| 2277 | TTI.getArithmeticInstrCost( |
| 2278 | Opcode: Instruction::LShr, Ty: ScalarDstTy, CostKind, |
| 2279 | Opd1Info: {.Kind: TTI::OK_AnyValue, .Properties: TTI::OP_None}, |
| 2280 | Opd2Info: {.Kind: TTI::OK_NonUniformConstantValue, .Properties: TTI::OP_None}); |
| 2281 | if (ScalarCost > VectorCost) |
| 2282 | return false; |
| 2283 | |
| 2284 | Value *ScalarV = Ext->getOperand(i_nocapture: 0); |
| 2285 | if (!isGuaranteedNotToBePoison(V: ScalarV, AC: SQ.AC, CtxI: dyn_cast<Instruction>(Val: ScalarV), |
| 2286 | DT: SQ.DT)) { |
| 2287 | // Check wether all lanes are extracted, all extracts trigger UB |
| 2288 | // on poison, and the last extract (and hence all previous ones) |
| 2289 | // are guaranteed to execute if Ext executes. If so, we do not |
| 2290 | // need to insert a freeze. |
| 2291 | SmallDenseSet<ConstantInt *, 8> ; |
| 2292 | bool = true; |
| 2293 | ExtractElementInst * = nullptr; |
| 2294 | BasicBlock *ExtBB = Ext->getParent(); |
| 2295 | for (User *U : Ext->users()) { |
| 2296 | auto * = cast<ExtractElementInst>(Val: U); |
| 2297 | if (Extract->getParent() != ExtBB || !programUndefinedIfPoison(Inst: Extract)) { |
| 2298 | AllExtractsTriggerUB = false; |
| 2299 | break; |
| 2300 | } |
| 2301 | ExtractedLanes.insert(V: cast<ConstantInt>(Val: Extract->getIndexOperand())); |
| 2302 | if (!LastExtract || LastExtract->comesBefore(Other: Extract)) |
| 2303 | LastExtract = Extract; |
| 2304 | } |
| 2305 | if (ExtractedLanes.size() != DstTy->getNumElements() || |
| 2306 | !AllExtractsTriggerUB || |
| 2307 | !isGuaranteedToTransferExecutionToSuccessor(Begin: Ext->getIterator(), |
| 2308 | End: LastExtract->getIterator())) |
| 2309 | ScalarV = Builder.CreateFreeze(V: ScalarV); |
| 2310 | } |
| 2311 | ScalarV = Builder.CreateBitCast( |
| 2312 | V: ScalarV, |
| 2313 | DestTy: IntegerType::get(C&: SrcTy->getContext(), NumBits: DL->getTypeSizeInBits(Ty: SrcTy))); |
| 2314 | uint64_t SrcEltSizeInBits = DL->getTypeSizeInBits(Ty: SrcTy->getElementType()); |
| 2315 | uint64_t TotalBits = DL->getTypeSizeInBits(Ty: SrcTy); |
| 2316 | APInt EltBitMask = APInt::getLowBitsSet(numBits: TotalBits, loBitsSet: SrcEltSizeInBits); |
| 2317 | Type *PackedTy = IntegerType::get(C&: SrcTy->getContext(), NumBits: TotalBits); |
| 2318 | Value *Mask = ConstantInt::get(Ty: PackedTy, V: EltBitMask); |
| 2319 | for (User *U : Ext->users()) { |
| 2320 | auto * = cast<ExtractElementInst>(Val: U); |
| 2321 | uint64_t Idx = |
| 2322 | cast<ConstantInt>(Val: Extract->getIndexOperand())->getZExtValue(); |
| 2323 | uint64_t ShiftAmt = |
| 2324 | DL->isBigEndian() |
| 2325 | ? (TotalBits - SrcEltSizeInBits - Idx * SrcEltSizeInBits) |
| 2326 | : (Idx * SrcEltSizeInBits); |
| 2327 | Value *LShr = Builder.CreateLShr(LHS: ScalarV, RHS: ShiftAmt); |
| 2328 | Value *And = Builder.CreateAnd(LHS: LShr, RHS: Mask); |
| 2329 | U->replaceAllUsesWith(V: And); |
| 2330 | } |
| 2331 | return true; |
| 2332 | } |
| 2333 | |
| 2334 | /// Try to fold "(or (zext (bitcast X)), (shl (zext (bitcast Y)), C))" |
| 2335 | /// to "(bitcast (concat X, Y))" |
| 2336 | /// where X/Y are bitcasted from i1 mask vectors. |
| 2337 | bool VectorCombine::foldConcatOfBoolMasks(Instruction &I) { |
| 2338 | Type *Ty = I.getType(); |
| 2339 | if (!Ty->isIntegerTy()) |
| 2340 | return false; |
| 2341 | |
| 2342 | // TODO: Add big endian test coverage |
| 2343 | if (DL->isBigEndian()) |
| 2344 | return false; |
| 2345 | |
| 2346 | // Restrict to disjoint cases so the mask vectors aren't overlapping. |
| 2347 | Instruction *X, *Y; |
| 2348 | if (!match(V: &I, P: m_DisjointOr(L: m_Instruction(I&: X), R: m_Instruction(I&: Y)))) |
| 2349 | return false; |
| 2350 | |
| 2351 | // Allow both sources to contain shl, to handle more generic pattern: |
| 2352 | // "(or (shl (zext (bitcast X)), C1), (shl (zext (bitcast Y)), C2))" |
| 2353 | Value *SrcX; |
| 2354 | uint64_t ShAmtX = 0; |
| 2355 | if (!match(V: X, P: m_OneUse(SubPattern: m_ZExt(Op: m_OneUse(SubPattern: m_BitCast(Op: m_Value(V&: SrcX)))))) && |
| 2356 | !match(V: X, P: m_OneUse( |
| 2357 | SubPattern: m_Shl(L: m_OneUse(SubPattern: m_ZExt(Op: m_OneUse(SubPattern: m_BitCast(Op: m_Value(V&: SrcX))))), |
| 2358 | R: m_ConstantInt(V&: ShAmtX))))) |
| 2359 | return false; |
| 2360 | |
| 2361 | Value *SrcY; |
| 2362 | uint64_t ShAmtY = 0; |
| 2363 | if (!match(V: Y, P: m_OneUse(SubPattern: m_ZExt(Op: m_OneUse(SubPattern: m_BitCast(Op: m_Value(V&: SrcY)))))) && |
| 2364 | !match(V: Y, P: m_OneUse( |
| 2365 | SubPattern: m_Shl(L: m_OneUse(SubPattern: m_ZExt(Op: m_OneUse(SubPattern: m_BitCast(Op: m_Value(V&: SrcY))))), |
| 2366 | R: m_ConstantInt(V&: ShAmtY))))) |
| 2367 | return false; |
| 2368 | |
| 2369 | // Canonicalize larger shift to the RHS. |
| 2370 | if (ShAmtX > ShAmtY) { |
| 2371 | std::swap(a&: X, b&: Y); |
| 2372 | std::swap(a&: SrcX, b&: SrcY); |
| 2373 | std::swap(a&: ShAmtX, b&: ShAmtY); |
| 2374 | } |
| 2375 | |
| 2376 | // Ensure both sources are matching vXi1 bool mask types, and that the shift |
| 2377 | // difference is the mask width so they can be easily concatenated together. |
| 2378 | uint64_t ShAmtDiff = ShAmtY - ShAmtX; |
| 2379 | unsigned NumSHL = (ShAmtX > 0) + (ShAmtY > 0); |
| 2380 | unsigned BitWidth = Ty->getPrimitiveSizeInBits(); |
| 2381 | auto *MaskTy = dyn_cast<FixedVectorType>(Val: SrcX->getType()); |
| 2382 | if (!MaskTy || SrcX->getType() != SrcY->getType() || |
| 2383 | !MaskTy->getElementType()->isIntegerTy(BitWidth: 1) || |
| 2384 | MaskTy->getNumElements() != ShAmtDiff || |
| 2385 | MaskTy->getNumElements() > (BitWidth / 2)) |
| 2386 | return false; |
| 2387 | |
| 2388 | auto *ConcatTy = FixedVectorType::getDoubleElementsVectorType(VTy: MaskTy); |
| 2389 | auto *ConcatIntTy = |
| 2390 | Type::getIntNTy(C&: Ty->getContext(), N: ConcatTy->getNumElements()); |
| 2391 | auto *MaskIntTy = Type::getIntNTy(C&: Ty->getContext(), N: ShAmtDiff); |
| 2392 | |
| 2393 | SmallVector<int, 32> ConcatMask(ConcatTy->getNumElements()); |
| 2394 | std::iota(first: ConcatMask.begin(), last: ConcatMask.end(), value: 0); |
| 2395 | |
| 2396 | // TODO: Is it worth supporting multi use cases? |
| 2397 | InstructionCost OldCost = 0; |
| 2398 | OldCost += TTI.getArithmeticInstrCost(Opcode: Instruction::Or, Ty, CostKind); |
| 2399 | OldCost += |
| 2400 | NumSHL * TTI.getArithmeticInstrCost(Opcode: Instruction::Shl, Ty, CostKind); |
| 2401 | OldCost += 2 * TTI.getCastInstrCost(Opcode: Instruction::ZExt, Dst: Ty, Src: MaskIntTy, |
| 2402 | CCH: TTI::CastContextHint::None, CostKind); |
| 2403 | OldCost += 2 * TTI.getCastInstrCost(Opcode: Instruction::BitCast, Dst: MaskIntTy, Src: MaskTy, |
| 2404 | CCH: TTI::CastContextHint::None, CostKind); |
| 2405 | |
| 2406 | InstructionCost NewCost = 0; |
| 2407 | NewCost += TTI.getShuffleCost(Kind: TargetTransformInfo::SK_PermuteTwoSrc, DstTy: ConcatTy, |
| 2408 | SrcTy: MaskTy, Mask: ConcatMask, CostKind); |
| 2409 | NewCost += TTI.getCastInstrCost(Opcode: Instruction::BitCast, Dst: ConcatIntTy, Src: ConcatTy, |
| 2410 | CCH: TTI::CastContextHint::None, CostKind); |
| 2411 | if (Ty != ConcatIntTy) |
| 2412 | NewCost += TTI.getCastInstrCost(Opcode: Instruction::ZExt, Dst: Ty, Src: ConcatIntTy, |
| 2413 | CCH: TTI::CastContextHint::None, CostKind); |
| 2414 | if (ShAmtX > 0) |
| 2415 | NewCost += TTI.getArithmeticInstrCost(Opcode: Instruction::Shl, Ty, CostKind); |
| 2416 | |
| 2417 | LLVM_DEBUG(dbgs() << "Found a concatenation of bitcasted bool masks: " << I |
| 2418 | << "\n OldCost: " << OldCost << " vs NewCost: " << NewCost |
| 2419 | << "\n" ); |
| 2420 | |
| 2421 | if (NewCost > OldCost) |
| 2422 | return false; |
| 2423 | |
| 2424 | // Build bool mask concatenation, bitcast back to scalar integer, and perform |
| 2425 | // any residual zero-extension or shifting. |
| 2426 | Value *Concat = Builder.CreateShuffleVector(V1: SrcX, V2: SrcY, Mask: ConcatMask); |
| 2427 | Worklist.pushValue(V: Concat); |
| 2428 | |
| 2429 | Value *Result = Builder.CreateBitCast(V: Concat, DestTy: ConcatIntTy); |
| 2430 | |
| 2431 | if (Ty != ConcatIntTy) { |
| 2432 | Worklist.pushValue(V: Result); |
| 2433 | Result = Builder.CreateZExt(V: Result, DestTy: Ty); |
| 2434 | } |
| 2435 | |
| 2436 | if (ShAmtX > 0) { |
| 2437 | Worklist.pushValue(V: Result); |
| 2438 | Result = Builder.CreateShl(LHS: Result, RHS: ShAmtX); |
| 2439 | } |
| 2440 | |
| 2441 | replaceValue(Old&: I, New&: *Result); |
| 2442 | return true; |
| 2443 | } |
| 2444 | |
| 2445 | /// Try to convert "shuffle (binop (shuffle, shuffle)), undef" |
| 2446 | /// --> "binop (shuffle), (shuffle)". |
| 2447 | bool VectorCombine::foldPermuteOfBinops(Instruction &I) { |
| 2448 | BinaryOperator *BinOp; |
| 2449 | ArrayRef<int> OuterMask; |
| 2450 | if (!match(V: &I, P: m_Shuffle(v1: m_BinOp(I&: BinOp), v2: m_Undef(), mask: m_Mask(OuterMask)))) |
| 2451 | return false; |
| 2452 | |
| 2453 | // Don't introduce poison into div/rem. |
| 2454 | if (BinOp->isIntDivRem() && llvm::is_contained(Range&: OuterMask, Element: PoisonMaskElem)) |
| 2455 | return false; |
| 2456 | |
| 2457 | Value *Op00, *Op01, *Op10, *Op11; |
| 2458 | ArrayRef<int> Mask0, Mask1; |
| 2459 | bool Match0 = match(V: BinOp->getOperand(i_nocapture: 0), |
| 2460 | P: m_Shuffle(v1: m_Value(V&: Op00), v2: m_Value(V&: Op01), mask: m_Mask(Mask0))); |
| 2461 | bool Match1 = match(V: BinOp->getOperand(i_nocapture: 1), |
| 2462 | P: m_Shuffle(v1: m_Value(V&: Op10), v2: m_Value(V&: Op11), mask: m_Mask(Mask1))); |
| 2463 | if (!Match0 && !Match1) |
| 2464 | return false; |
| 2465 | |
| 2466 | Op00 = Match0 ? Op00 : BinOp->getOperand(i_nocapture: 0); |
| 2467 | Op01 = Match0 ? Op01 : BinOp->getOperand(i_nocapture: 0); |
| 2468 | Op10 = Match1 ? Op10 : BinOp->getOperand(i_nocapture: 1); |
| 2469 | Op11 = Match1 ? Op11 : BinOp->getOperand(i_nocapture: 1); |
| 2470 | |
| 2471 | Instruction::BinaryOps Opcode = BinOp->getOpcode(); |
| 2472 | auto *ShuffleDstTy = dyn_cast<FixedVectorType>(Val: I.getType()); |
| 2473 | auto *BinOpTy = dyn_cast<FixedVectorType>(Val: BinOp->getType()); |
| 2474 | auto *Op0Ty = dyn_cast<FixedVectorType>(Val: Op00->getType()); |
| 2475 | auto *Op1Ty = dyn_cast<FixedVectorType>(Val: Op10->getType()); |
| 2476 | if (!ShuffleDstTy || !BinOpTy || !Op0Ty || !Op1Ty) |
| 2477 | return false; |
| 2478 | |
| 2479 | unsigned NumSrcElts = BinOpTy->getNumElements(); |
| 2480 | |
| 2481 | // Don't accept shuffles that reference the second operand in |
| 2482 | // div/rem or if its an undef arg. |
| 2483 | if ((BinOp->isIntDivRem() || !isa<PoisonValue>(Val: I.getOperand(i: 1))) && |
| 2484 | any_of(Range&: OuterMask, P: [NumSrcElts](int M) { return M >= (int)NumSrcElts; })) |
| 2485 | return false; |
| 2486 | |
| 2487 | // Merge outer / inner (or identity if no match) shuffles. |
| 2488 | SmallVector<int> NewMask0, NewMask1; |
| 2489 | for (int M : OuterMask) { |
| 2490 | if (M < 0 || M >= (int)NumSrcElts) { |
| 2491 | NewMask0.push_back(Elt: PoisonMaskElem); |
| 2492 | NewMask1.push_back(Elt: PoisonMaskElem); |
| 2493 | } else { |
| 2494 | NewMask0.push_back(Elt: Match0 ? Mask0[M] : M); |
| 2495 | NewMask1.push_back(Elt: Match1 ? Mask1[M] : M); |
| 2496 | } |
| 2497 | } |
| 2498 | |
| 2499 | unsigned NumOpElts = Op0Ty->getNumElements(); |
| 2500 | bool IsIdentity0 = ShuffleDstTy == Op0Ty && |
| 2501 | all_of(Range&: NewMask0, P: [NumOpElts](int M) { return M < (int)NumOpElts; }) && |
| 2502 | ShuffleVectorInst::isIdentityMask(Mask: NewMask0, NumSrcElts: NumOpElts); |
| 2503 | bool IsIdentity1 = ShuffleDstTy == Op1Ty && |
| 2504 | all_of(Range&: NewMask1, P: [NumOpElts](int M) { return M < (int)NumOpElts; }) && |
| 2505 | ShuffleVectorInst::isIdentityMask(Mask: NewMask1, NumSrcElts: NumOpElts); |
| 2506 | |
| 2507 | InstructionCost NewCost = 0; |
| 2508 | // Try to merge shuffles across the binop if the new shuffles are not costly. |
| 2509 | InstructionCost BinOpCost = |
| 2510 | TTI.getArithmeticInstrCost(Opcode, Ty: BinOpTy, CostKind); |
| 2511 | InstructionCost OldCost = |
| 2512 | BinOpCost + TTI.getShuffleCost(Kind: TargetTransformInfo::SK_PermuteSingleSrc, |
| 2513 | DstTy: ShuffleDstTy, SrcTy: BinOpTy, Mask: OuterMask, CostKind, |
| 2514 | Index: 0, SubTp: nullptr, Args: {BinOp}, CxtI: &I); |
| 2515 | if (!BinOp->hasOneUse()) |
| 2516 | NewCost += BinOpCost; |
| 2517 | |
| 2518 | if (Match0) { |
| 2519 | InstructionCost Shuf0Cost = TTI.getShuffleCost( |
| 2520 | Kind: TargetTransformInfo::SK_PermuteTwoSrc, DstTy: BinOpTy, SrcTy: Op0Ty, Mask: Mask0, CostKind, |
| 2521 | Index: 0, SubTp: nullptr, Args: {Op00, Op01}, CxtI: cast<Instruction>(Val: BinOp->getOperand(i_nocapture: 0))); |
| 2522 | OldCost += Shuf0Cost; |
| 2523 | if (!BinOp->hasOneUse() || !BinOp->getOperand(i_nocapture: 0)->hasOneUse()) |
| 2524 | NewCost += Shuf0Cost; |
| 2525 | } |
| 2526 | if (Match1) { |
| 2527 | InstructionCost Shuf1Cost = TTI.getShuffleCost( |
| 2528 | Kind: TargetTransformInfo::SK_PermuteTwoSrc, DstTy: BinOpTy, SrcTy: Op1Ty, Mask: Mask1, CostKind, |
| 2529 | Index: 0, SubTp: nullptr, Args: {Op10, Op11}, CxtI: cast<Instruction>(Val: BinOp->getOperand(i_nocapture: 1))); |
| 2530 | OldCost += Shuf1Cost; |
| 2531 | if (!BinOp->hasOneUse() || !BinOp->getOperand(i_nocapture: 1)->hasOneUse()) |
| 2532 | NewCost += Shuf1Cost; |
| 2533 | } |
| 2534 | |
| 2535 | NewCost += TTI.getArithmeticInstrCost(Opcode, Ty: ShuffleDstTy, CostKind); |
| 2536 | |
| 2537 | if (!IsIdentity0) |
| 2538 | NewCost += |
| 2539 | TTI.getShuffleCost(Kind: TargetTransformInfo::SK_PermuteTwoSrc, DstTy: ShuffleDstTy, |
| 2540 | SrcTy: Op0Ty, Mask: NewMask0, CostKind, Index: 0, SubTp: nullptr, Args: {Op00, Op01}); |
| 2541 | if (!IsIdentity1) |
| 2542 | NewCost += |
| 2543 | TTI.getShuffleCost(Kind: TargetTransformInfo::SK_PermuteTwoSrc, DstTy: ShuffleDstTy, |
| 2544 | SrcTy: Op1Ty, Mask: NewMask1, CostKind, Index: 0, SubTp: nullptr, Args: {Op10, Op11}); |
| 2545 | |
| 2546 | LLVM_DEBUG(dbgs() << "Found a shuffle feeding a shuffled binop: " << I |
| 2547 | << "\n OldCost: " << OldCost << " vs NewCost: " << NewCost |
| 2548 | << "\n" ); |
| 2549 | |
| 2550 | // If costs are equal, still fold as we reduce instruction count. |
| 2551 | if (NewCost > OldCost) |
| 2552 | return false; |
| 2553 | |
| 2554 | Value *LHS = |
| 2555 | IsIdentity0 ? Op00 : Builder.CreateShuffleVector(V1: Op00, V2: Op01, Mask: NewMask0); |
| 2556 | Value *RHS = |
| 2557 | IsIdentity1 ? Op10 : Builder.CreateShuffleVector(V1: Op10, V2: Op11, Mask: NewMask1); |
| 2558 | Value *NewBO = Builder.CreateBinOp(Opc: Opcode, LHS, RHS); |
| 2559 | |
| 2560 | // Intersect flags from the old binops. |
| 2561 | if (auto *NewInst = dyn_cast<Instruction>(Val: NewBO)) |
| 2562 | NewInst->copyIRFlags(V: BinOp); |
| 2563 | |
| 2564 | Worklist.pushValue(V: LHS); |
| 2565 | Worklist.pushValue(V: RHS); |
| 2566 | replaceValue(Old&: I, New&: *NewBO); |
| 2567 | return true; |
| 2568 | } |
| 2569 | |
| 2570 | /// Try to convert "shuffle (binop), (binop)" into "binop (shuffle), (shuffle)". |
| 2571 | /// Try to convert "shuffle (cmpop), (cmpop)" into "cmpop (shuffle), (shuffle)". |
| 2572 | bool VectorCombine::foldShuffleOfBinops(Instruction &I) { |
| 2573 | ArrayRef<int> OldMask; |
| 2574 | Instruction *LHS, *RHS; |
| 2575 | if (!match(V: &I, P: m_Shuffle(v1: m_Instruction(I&: LHS), v2: m_Instruction(I&: RHS), |
| 2576 | mask: m_Mask(OldMask)))) |
| 2577 | return false; |
| 2578 | |
| 2579 | // TODO: Add support for addlike etc. |
| 2580 | if (LHS->getOpcode() != RHS->getOpcode()) |
| 2581 | return false; |
| 2582 | |
| 2583 | Value *X, *Y, *Z, *W; |
| 2584 | bool IsCommutative = false; |
| 2585 | CmpPredicate PredLHS = CmpInst::BAD_ICMP_PREDICATE; |
| 2586 | CmpPredicate PredRHS = CmpInst::BAD_ICMP_PREDICATE; |
| 2587 | if (match(V: LHS, P: m_BinOp(L: m_Value(V&: X), R: m_Value(V&: Y))) && |
| 2588 | match(V: RHS, P: m_BinOp(L: m_Value(V&: Z), R: m_Value(V&: W)))) { |
| 2589 | auto *BO = cast<BinaryOperator>(Val: LHS); |
| 2590 | // Don't introduce poison into div/rem. |
| 2591 | if (llvm::is_contained(Range&: OldMask, Element: PoisonMaskElem) && BO->isIntDivRem()) |
| 2592 | return false; |
| 2593 | IsCommutative = BinaryOperator::isCommutative(Opcode: BO->getOpcode()); |
| 2594 | } else if (match(V: LHS, P: m_Cmp(Pred&: PredLHS, L: m_Value(V&: X), R: m_Value(V&: Y))) && |
| 2595 | match(V: RHS, P: m_Cmp(Pred&: PredRHS, L: m_Value(V&: Z), R: m_Value(V&: W))) && |
| 2596 | (CmpInst::Predicate)PredLHS == (CmpInst::Predicate)PredRHS) { |
| 2597 | IsCommutative = cast<CmpInst>(Val: LHS)->isCommutative(); |
| 2598 | } else |
| 2599 | return false; |
| 2600 | |
| 2601 | auto *ShuffleDstTy = dyn_cast<FixedVectorType>(Val: I.getType()); |
| 2602 | auto *BinResTy = dyn_cast<FixedVectorType>(Val: LHS->getType()); |
| 2603 | auto *BinOpTy = dyn_cast<FixedVectorType>(Val: X->getType()); |
| 2604 | if (!ShuffleDstTy || !BinResTy || !BinOpTy || X->getType() != Z->getType()) |
| 2605 | return false; |
| 2606 | |
| 2607 | bool SameBinOp = LHS == RHS; |
| 2608 | unsigned NumSrcElts = BinOpTy->getNumElements(); |
| 2609 | |
| 2610 | // If we have something like "add X, Y" and "add Z, X", swap ops to match. |
| 2611 | if (IsCommutative && X != Z && Y != W && (X == W || Y == Z)) |
| 2612 | std::swap(a&: X, b&: Y); |
| 2613 | |
| 2614 | auto ConvertToUnary = [NumSrcElts](int &M) { |
| 2615 | if (M >= (int)NumSrcElts) |
| 2616 | M -= NumSrcElts; |
| 2617 | }; |
| 2618 | |
| 2619 | SmallVector<int> NewMask0(OldMask); |
| 2620 | TargetTransformInfo::ShuffleKind SK0 = TargetTransformInfo::SK_PermuteTwoSrc; |
| 2621 | TTI::OperandValueInfo Op0Info = TTI.commonOperandInfo(X, Y: Z); |
| 2622 | if (X == Z) { |
| 2623 | llvm::for_each(Range&: NewMask0, F: ConvertToUnary); |
| 2624 | SK0 = TargetTransformInfo::SK_PermuteSingleSrc; |
| 2625 | Z = PoisonValue::get(T: BinOpTy); |
| 2626 | } |
| 2627 | |
| 2628 | SmallVector<int> NewMask1(OldMask); |
| 2629 | TargetTransformInfo::ShuffleKind SK1 = TargetTransformInfo::SK_PermuteTwoSrc; |
| 2630 | TTI::OperandValueInfo Op1Info = TTI.commonOperandInfo(X: Y, Y: W); |
| 2631 | if (Y == W) { |
| 2632 | llvm::for_each(Range&: NewMask1, F: ConvertToUnary); |
| 2633 | SK1 = TargetTransformInfo::SK_PermuteSingleSrc; |
| 2634 | W = PoisonValue::get(T: BinOpTy); |
| 2635 | } |
| 2636 | |
| 2637 | // Try to replace a binop with a shuffle if the shuffle is not costly. |
| 2638 | // When SameBinOp, only count the binop cost once. |
| 2639 | InstructionCost LHSCost = TTI.getInstructionCost(U: LHS, CostKind); |
| 2640 | InstructionCost RHSCost = TTI.getInstructionCost(U: RHS, CostKind); |
| 2641 | |
| 2642 | InstructionCost OldCost = LHSCost; |
| 2643 | if (!SameBinOp) { |
| 2644 | OldCost += RHSCost; |
| 2645 | } |
| 2646 | OldCost += TTI.getShuffleCost(Kind: TargetTransformInfo::SK_PermuteTwoSrc, |
| 2647 | DstTy: ShuffleDstTy, SrcTy: BinResTy, Mask: OldMask, CostKind, Index: 0, |
| 2648 | SubTp: nullptr, Args: {LHS, RHS}, CxtI: &I); |
| 2649 | |
| 2650 | // Handle shuffle(binop(shuffle(x),y),binop(z,shuffle(w))) style patterns |
| 2651 | // where one use shuffles have gotten split across the binop/cmp. These |
| 2652 | // often allow a major reduction in total cost that wouldn't happen as |
| 2653 | // individual folds. |
| 2654 | auto MergeInner = [&](Value *&Op, int Offset, MutableArrayRef<int> Mask, |
| 2655 | TTI::TargetCostKind CostKind) -> bool { |
| 2656 | Value *InnerOp; |
| 2657 | ArrayRef<int> InnerMask; |
| 2658 | if (match(V: Op, P: m_OneUse(SubPattern: m_Shuffle(v1: m_Value(V&: InnerOp), v2: m_Undef(), |
| 2659 | mask: m_Mask(InnerMask)))) && |
| 2660 | InnerOp->getType() == Op->getType() && |
| 2661 | all_of(Range&: InnerMask, |
| 2662 | P: [NumSrcElts](int M) { return M < (int)NumSrcElts; })) { |
| 2663 | for (int &M : Mask) |
| 2664 | if (Offset <= M && M < (int)(Offset + NumSrcElts)) { |
| 2665 | M = InnerMask[M - Offset]; |
| 2666 | M = 0 <= M ? M + Offset : M; |
| 2667 | } |
| 2668 | OldCost += TTI.getInstructionCost(U: cast<Instruction>(Val: Op), CostKind); |
| 2669 | Op = InnerOp; |
| 2670 | return true; |
| 2671 | } |
| 2672 | return false; |
| 2673 | }; |
| 2674 | bool ReducedInstCount = false; |
| 2675 | ReducedInstCount |= MergeInner(X, 0, NewMask0, CostKind); |
| 2676 | ReducedInstCount |= MergeInner(Y, 0, NewMask1, CostKind); |
| 2677 | ReducedInstCount |= MergeInner(Z, NumSrcElts, NewMask0, CostKind); |
| 2678 | ReducedInstCount |= MergeInner(W, NumSrcElts, NewMask1, CostKind); |
| 2679 | bool SingleSrcBinOp = (X == Y) && (Z == W) && (NewMask0 == NewMask1); |
| 2680 | // SingleSrcBinOp only reduces instruction count if we also eliminate the |
| 2681 | // original binop(s). If binops have multiple uses, they won't be eliminated. |
| 2682 | ReducedInstCount |= SingleSrcBinOp && LHS->hasOneUser() && RHS->hasOneUser(); |
| 2683 | |
| 2684 | // For concat shuffles of i1 vectors where both binops are one-use, the |
| 2685 | // transform keeps the same instruction count but canonicalises to a single |
| 2686 | // wider binop, enabling downstream folds (e.g. NOT(XOR(concat(a,b), |
| 2687 | // concat(c,d))) -> XNOR(concat(a,b),concat(c,d)) on AVX-512 mask regs). |
| 2688 | // Restrict to BinaryOperator (not CmpInst) since narrow comparisons may |
| 2689 | // be cheaper than wide ones on some targets (e.g. AVX-512 vpcmpeq). |
| 2690 | ReducedInstCount |= cast<ShuffleVectorInst>(Val: &I)->isConcat() && |
| 2691 | I.getType()->getScalarType()->isIntegerTy(BitWidth: 1) && |
| 2692 | isa<BinaryOperator>(Val: LHS) && LHS->hasOneUser() && |
| 2693 | RHS->hasOneUser(); |
| 2694 | |
| 2695 | auto *ShuffleCmpTy = |
| 2696 | FixedVectorType::get(ElementType: BinOpTy->getElementType(), FVTy: ShuffleDstTy); |
| 2697 | InstructionCost NewCost = TTI.getShuffleCost( |
| 2698 | Kind: SK0, DstTy: ShuffleCmpTy, SrcTy: BinOpTy, Mask: NewMask0, CostKind, Index: 0, SubTp: nullptr, Args: {X, Z}); |
| 2699 | if (!SingleSrcBinOp) |
| 2700 | NewCost += TTI.getShuffleCost(Kind: SK1, DstTy: ShuffleCmpTy, SrcTy: BinOpTy, Mask: NewMask1, |
| 2701 | CostKind, Index: 0, SubTp: nullptr, Args: {Y, W}); |
| 2702 | |
| 2703 | if (PredLHS == CmpInst::BAD_ICMP_PREDICATE) { |
| 2704 | NewCost += TTI.getArithmeticInstrCost(Opcode: LHS->getOpcode(), Ty: ShuffleDstTy, |
| 2705 | CostKind, Opd1Info: Op0Info, Opd2Info: Op1Info); |
| 2706 | } else { |
| 2707 | NewCost += |
| 2708 | TTI.getCmpSelInstrCost(Opcode: LHS->getOpcode(), ValTy: ShuffleCmpTy, CondTy: ShuffleDstTy, |
| 2709 | VecPred: PredLHS, CostKind, Op1Info: Op0Info, Op2Info: Op1Info); |
| 2710 | } |
| 2711 | // If LHS/RHS have other uses, we need to account for the cost of keeping |
| 2712 | // the original instructions. When SameBinOp, only add the cost once. |
| 2713 | if (!LHS->hasOneUser()) |
| 2714 | NewCost += LHSCost; |
| 2715 | if (!SameBinOp && !RHS->hasOneUser()) |
| 2716 | NewCost += RHSCost; |
| 2717 | |
| 2718 | LLVM_DEBUG(dbgs() << "Found a shuffle feeding two binops: " << I |
| 2719 | << "\n OldCost: " << OldCost << " vs NewCost: " << NewCost |
| 2720 | << "\n" ); |
| 2721 | |
| 2722 | // If either shuffle will constant fold away, then fold for the same cost as |
| 2723 | // we will reduce the instruction count. |
| 2724 | ReducedInstCount |= (isa<Constant>(Val: X) && isa<Constant>(Val: Z)) || |
| 2725 | (isa<Constant>(Val: Y) && isa<Constant>(Val: W)); |
| 2726 | if (ReducedInstCount ? (NewCost > OldCost) : (NewCost >= OldCost)) |
| 2727 | return false; |
| 2728 | |
| 2729 | Value *Shuf0 = Builder.CreateShuffleVector(V1: X, V2: Z, Mask: NewMask0); |
| 2730 | Value *Shuf1 = |
| 2731 | SingleSrcBinOp ? Shuf0 : Builder.CreateShuffleVector(V1: Y, V2: W, Mask: NewMask1); |
| 2732 | Value *NewBO = PredLHS == CmpInst::BAD_ICMP_PREDICATE |
| 2733 | ? Builder.CreateBinOp( |
| 2734 | Opc: cast<BinaryOperator>(Val: LHS)->getOpcode(), LHS: Shuf0, RHS: Shuf1) |
| 2735 | : Builder.CreateCmp(Pred: PredLHS, LHS: Shuf0, RHS: Shuf1); |
| 2736 | |
| 2737 | // Intersect flags from the old binops. |
| 2738 | if (auto *NewInst = dyn_cast<Instruction>(Val: NewBO)) { |
| 2739 | NewInst->copyIRFlags(V: LHS); |
| 2740 | NewInst->andIRFlags(V: RHS); |
| 2741 | } |
| 2742 | |
| 2743 | Worklist.pushValue(V: Shuf0); |
| 2744 | Worklist.pushValue(V: Shuf1); |
| 2745 | replaceValue(Old&: I, New&: *NewBO); |
| 2746 | return true; |
| 2747 | } |
| 2748 | |
| 2749 | /// Try to convert, |
| 2750 | /// (shuffle(select(c1,t1,f1)), (select(c2,t2,f2)), m) into |
| 2751 | /// (select (shuffle c1,c2,m), (shuffle t1,t2,m), (shuffle f1,f2,m)) |
| 2752 | bool VectorCombine::foldShuffleOfSelects(Instruction &I) { |
| 2753 | ArrayRef<int> Mask; |
| 2754 | Value *C1, *T1, *F1, *C2, *T2, *F2; |
| 2755 | if (!match(V: &I, P: m_Shuffle(v1: m_Select(C: m_Value(V&: C1), L: m_Value(V&: T1), R: m_Value(V&: F1)), |
| 2756 | v2: m_Select(C: m_Value(V&: C2), L: m_Value(V&: T2), R: m_Value(V&: F2)), |
| 2757 | mask: m_Mask(Mask)))) |
| 2758 | return false; |
| 2759 | |
| 2760 | auto *Sel1 = cast<Instruction>(Val: I.getOperand(i: 0)); |
| 2761 | auto *Sel2 = cast<Instruction>(Val: I.getOperand(i: 1)); |
| 2762 | |
| 2763 | auto *C1VecTy = dyn_cast<FixedVectorType>(Val: C1->getType()); |
| 2764 | auto *C2VecTy = dyn_cast<FixedVectorType>(Val: C2->getType()); |
| 2765 | if (!C1VecTy || !C2VecTy || C1VecTy != C2VecTy) |
| 2766 | return false; |
| 2767 | |
| 2768 | auto *SI0FOp = dyn_cast<FPMathOperator>(Val: I.getOperand(i: 0)); |
| 2769 | auto *SI1FOp = dyn_cast<FPMathOperator>(Val: I.getOperand(i: 1)); |
| 2770 | // SelectInsts must have the same FMF. |
| 2771 | if (((SI0FOp == nullptr) != (SI1FOp == nullptr)) || |
| 2772 | ((SI0FOp != nullptr) && |
| 2773 | (SI0FOp->getFastMathFlags() != SI1FOp->getFastMathFlags()))) |
| 2774 | return false; |
| 2775 | |
| 2776 | auto *SrcVecTy = cast<FixedVectorType>(Val: T1->getType()); |
| 2777 | auto *DstVecTy = cast<FixedVectorType>(Val: I.getType()); |
| 2778 | auto SK = TargetTransformInfo::SK_PermuteTwoSrc; |
| 2779 | auto SelOp = Instruction::Select; |
| 2780 | |
| 2781 | InstructionCost CostSel1 = TTI.getCmpSelInstrCost( |
| 2782 | Opcode: SelOp, ValTy: SrcVecTy, CondTy: C1VecTy, VecPred: CmpInst::BAD_ICMP_PREDICATE, CostKind); |
| 2783 | InstructionCost CostSel2 = TTI.getCmpSelInstrCost( |
| 2784 | Opcode: SelOp, ValTy: SrcVecTy, CondTy: C2VecTy, VecPred: CmpInst::BAD_ICMP_PREDICATE, CostKind); |
| 2785 | |
| 2786 | InstructionCost OldCost = |
| 2787 | CostSel1 + CostSel2 + |
| 2788 | TTI.getShuffleCost(Kind: SK, DstTy: DstVecTy, SrcTy: SrcVecTy, Mask, CostKind, Index: 0, SubTp: nullptr, |
| 2789 | Args: {I.getOperand(i: 0), I.getOperand(i: 1)}, CxtI: &I); |
| 2790 | |
| 2791 | InstructionCost NewCost = TTI.getShuffleCost( |
| 2792 | Kind: SK, DstTy: FixedVectorType::get(ElementType: C1VecTy->getScalarType(), NumElts: Mask.size()), SrcTy: C1VecTy, |
| 2793 | Mask, CostKind, Index: 0, SubTp: nullptr, Args: {C1, C2}); |
| 2794 | NewCost += TTI.getShuffleCost(Kind: SK, DstTy: DstVecTy, SrcTy: SrcVecTy, Mask, CostKind, Index: 0, |
| 2795 | SubTp: nullptr, Args: {T1, T2}); |
| 2796 | NewCost += TTI.getShuffleCost(Kind: SK, DstTy: DstVecTy, SrcTy: SrcVecTy, Mask, CostKind, Index: 0, |
| 2797 | SubTp: nullptr, Args: {F1, F2}); |
| 2798 | auto *C1C2ShuffledVecTy = FixedVectorType::get( |
| 2799 | ElementType: Type::getInt1Ty(C&: I.getContext()), NumElts: DstVecTy->getNumElements()); |
| 2800 | NewCost += TTI.getCmpSelInstrCost(Opcode: SelOp, ValTy: DstVecTy, CondTy: C1C2ShuffledVecTy, |
| 2801 | VecPred: CmpInst::BAD_ICMP_PREDICATE, CostKind); |
| 2802 | |
| 2803 | if (!Sel1->hasOneUse()) |
| 2804 | NewCost += CostSel1; |
| 2805 | if (!Sel2->hasOneUse()) |
| 2806 | NewCost += CostSel2; |
| 2807 | |
| 2808 | LLVM_DEBUG(dbgs() << "Found a shuffle feeding two selects: " << I |
| 2809 | << "\n OldCost: " << OldCost << " vs NewCost: " << NewCost |
| 2810 | << "\n" ); |
| 2811 | if (NewCost > OldCost) |
| 2812 | return false; |
| 2813 | |
| 2814 | Value *ShuffleCmp = Builder.CreateShuffleVector(V1: C1, V2: C2, Mask); |
| 2815 | Value *ShuffleTrue = Builder.CreateShuffleVector(V1: T1, V2: T2, Mask); |
| 2816 | Value *ShuffleFalse = Builder.CreateShuffleVector(V1: F1, V2: F2, Mask); |
| 2817 | Value *NewSel; |
| 2818 | // We presuppose that the SelectInsts have the same FMF. |
| 2819 | if (SI0FOp) |
| 2820 | NewSel = Builder.CreateSelectFMF(C: ShuffleCmp, True: ShuffleTrue, False: ShuffleFalse, |
| 2821 | FMFSource: SI0FOp->getFastMathFlags()); |
| 2822 | else |
| 2823 | NewSel = Builder.CreateSelect(C: ShuffleCmp, True: ShuffleTrue, False: ShuffleFalse); |
| 2824 | |
| 2825 | Worklist.pushValue(V: ShuffleCmp); |
| 2826 | Worklist.pushValue(V: ShuffleTrue); |
| 2827 | Worklist.pushValue(V: ShuffleFalse); |
| 2828 | replaceValue(Old&: I, New&: *NewSel); |
| 2829 | return true; |
| 2830 | } |
| 2831 | |
| 2832 | /// Try to convert "shuffle (castop), (castop)" with a shared castop operand |
| 2833 | /// into "castop (shuffle)". |
| 2834 | bool VectorCombine::foldShuffleOfCastops(Instruction &I) { |
| 2835 | Value *V0, *V1; |
| 2836 | ArrayRef<int> OldMask; |
| 2837 | if (!match(V: &I, P: m_Shuffle(v1: m_Value(V&: V0), v2: m_Value(V&: V1), mask: m_Mask(OldMask)))) |
| 2838 | return false; |
| 2839 | |
| 2840 | // Check whether this is a binary shuffle. |
| 2841 | bool IsBinaryShuffle = !isa<UndefValue>(Val: V1); |
| 2842 | |
| 2843 | auto *C0 = dyn_cast<CastInst>(Val: V0); |
| 2844 | auto *C1 = dyn_cast<CastInst>(Val: V1); |
| 2845 | if (!C0 || (IsBinaryShuffle && !C1)) |
| 2846 | return false; |
| 2847 | |
| 2848 | Instruction::CastOps Opcode = C0->getOpcode(); |
| 2849 | |
| 2850 | // If this is allowed, foldShuffleOfCastops can get stuck in a loop |
| 2851 | // with foldBitcastOfShuffle. Reject in favor of foldBitcastOfShuffle. |
| 2852 | if (!IsBinaryShuffle && Opcode == Instruction::BitCast) |
| 2853 | return false; |
| 2854 | |
| 2855 | if (IsBinaryShuffle) { |
| 2856 | if (C0->getSrcTy() != C1->getSrcTy()) |
| 2857 | return false; |
| 2858 | // Handle shuffle(zext_nneg(x), sext(y)) -> sext(shuffle(x,y)) folds. |
| 2859 | if (Opcode != C1->getOpcode()) { |
| 2860 | if (match(V: C0, P: m_SExtLike(Op: m_Value())) && match(V: C1, P: m_SExtLike(Op: m_Value()))) |
| 2861 | Opcode = Instruction::SExt; |
| 2862 | else |
| 2863 | return false; |
| 2864 | } |
| 2865 | } |
| 2866 | |
| 2867 | auto *ShuffleDstTy = dyn_cast<FixedVectorType>(Val: I.getType()); |
| 2868 | auto *CastDstTy = dyn_cast<FixedVectorType>(Val: C0->getDestTy()); |
| 2869 | auto *CastSrcTy = dyn_cast<FixedVectorType>(Val: C0->getSrcTy()); |
| 2870 | if (!ShuffleDstTy || !CastDstTy || !CastSrcTy) |
| 2871 | return false; |
| 2872 | |
| 2873 | unsigned NumSrcElts = CastSrcTy->getNumElements(); |
| 2874 | unsigned NumDstElts = CastDstTy->getNumElements(); |
| 2875 | assert((NumDstElts == NumSrcElts || Opcode == Instruction::BitCast) && |
| 2876 | "Only bitcasts expected to alter src/dst element counts" ); |
| 2877 | |
| 2878 | // Check for bitcasting of unscalable vector types. |
| 2879 | // e.g. <32 x i40> -> <40 x i32> |
| 2880 | if (NumDstElts != NumSrcElts && (NumSrcElts % NumDstElts) != 0 && |
| 2881 | (NumDstElts % NumSrcElts) != 0) |
| 2882 | return false; |
| 2883 | |
| 2884 | SmallVector<int, 16> NewMask; |
| 2885 | if (NumSrcElts >= NumDstElts) { |
| 2886 | // The bitcast is from wide to narrow/equal elements. The shuffle mask can |
| 2887 | // always be expanded to the equivalent form choosing narrower elements. |
| 2888 | assert(NumSrcElts % NumDstElts == 0 && "Unexpected shuffle mask" ); |
| 2889 | unsigned ScaleFactor = NumSrcElts / NumDstElts; |
| 2890 | narrowShuffleMaskElts(Scale: ScaleFactor, Mask: OldMask, ScaledMask&: NewMask); |
| 2891 | } else { |
| 2892 | // The bitcast is from narrow elements to wide elements. The shuffle mask |
| 2893 | // must choose consecutive elements to allow casting first. |
| 2894 | assert(NumDstElts % NumSrcElts == 0 && "Unexpected shuffle mask" ); |
| 2895 | unsigned ScaleFactor = NumDstElts / NumSrcElts; |
| 2896 | if (!widenShuffleMaskElts(Scale: ScaleFactor, Mask: OldMask, ScaledMask&: NewMask)) |
| 2897 | return false; |
| 2898 | } |
| 2899 | |
| 2900 | auto *NewShuffleDstTy = |
| 2901 | FixedVectorType::get(ElementType: CastSrcTy->getScalarType(), NumElts: NewMask.size()); |
| 2902 | |
| 2903 | // Try to replace a castop with a shuffle if the shuffle is not costly. |
| 2904 | InstructionCost CostC0 = |
| 2905 | TTI.getCastInstrCost(Opcode: C0->getOpcode(), Dst: CastDstTy, Src: CastSrcTy, |
| 2906 | CCH: TTI::CastContextHint::None, CostKind, I: C0); |
| 2907 | |
| 2908 | TargetTransformInfo::ShuffleKind ShuffleKind; |
| 2909 | if (IsBinaryShuffle) |
| 2910 | ShuffleKind = TargetTransformInfo::SK_PermuteTwoSrc; |
| 2911 | else |
| 2912 | ShuffleKind = TargetTransformInfo::SK_PermuteSingleSrc; |
| 2913 | |
| 2914 | InstructionCost OldCost = CostC0; |
| 2915 | OldCost += TTI.getShuffleCost(Kind: ShuffleKind, DstTy: ShuffleDstTy, SrcTy: CastDstTy, Mask: OldMask, |
| 2916 | CostKind, Index: 0, SubTp: nullptr, Args: {}, CxtI: &I); |
| 2917 | |
| 2918 | InstructionCost NewCost = TTI.getShuffleCost(Kind: ShuffleKind, DstTy: NewShuffleDstTy, |
| 2919 | SrcTy: CastSrcTy, Mask: NewMask, CostKind); |
| 2920 | NewCost += TTI.getCastInstrCost(Opcode, Dst: ShuffleDstTy, Src: NewShuffleDstTy, |
| 2921 | CCH: TTI::CastContextHint::None, CostKind); |
| 2922 | if (!C0->hasOneUse()) |
| 2923 | NewCost += CostC0; |
| 2924 | if (IsBinaryShuffle) { |
| 2925 | InstructionCost CostC1 = |
| 2926 | TTI.getCastInstrCost(Opcode: C1->getOpcode(), Dst: CastDstTy, Src: CastSrcTy, |
| 2927 | CCH: TTI::CastContextHint::None, CostKind, I: C1); |
| 2928 | OldCost += CostC1; |
| 2929 | if (!C1->hasOneUse()) |
| 2930 | NewCost += CostC1; |
| 2931 | } |
| 2932 | |
| 2933 | LLVM_DEBUG(dbgs() << "Found a shuffle feeding two casts: " << I |
| 2934 | << "\n OldCost: " << OldCost << " vs NewCost: " << NewCost |
| 2935 | << "\n" ); |
| 2936 | if (NewCost > OldCost) |
| 2937 | return false; |
| 2938 | |
| 2939 | Value *Shuf; |
| 2940 | if (IsBinaryShuffle) |
| 2941 | Shuf = Builder.CreateShuffleVector(V1: C0->getOperand(i_nocapture: 0), V2: C1->getOperand(i_nocapture: 0), |
| 2942 | Mask: NewMask); |
| 2943 | else |
| 2944 | Shuf = Builder.CreateShuffleVector(V: C0->getOperand(i_nocapture: 0), Mask: NewMask); |
| 2945 | |
| 2946 | Value *Cast = Builder.CreateCast(Op: Opcode, V: Shuf, DestTy: ShuffleDstTy); |
| 2947 | |
| 2948 | // Intersect flags from the old casts. |
| 2949 | if (auto *NewInst = dyn_cast<Instruction>(Val: Cast)) { |
| 2950 | NewInst->copyIRFlags(V: C0); |
| 2951 | if (IsBinaryShuffle) |
| 2952 | NewInst->andIRFlags(V: C1); |
| 2953 | } |
| 2954 | |
| 2955 | Worklist.pushValue(V: Shuf); |
| 2956 | replaceValue(Old&: I, New&: *Cast); |
| 2957 | return true; |
| 2958 | } |
| 2959 | |
| 2960 | /// Try to convert any of: |
| 2961 | /// "shuffle (shuffle x, y), (shuffle y, x)" |
| 2962 | /// "shuffle (shuffle x, undef), (shuffle y, undef)" |
| 2963 | /// "shuffle (shuffle x, undef), y" |
| 2964 | /// "shuffle x, (shuffle y, undef)" |
| 2965 | /// into "shuffle x, y". |
| 2966 | bool VectorCombine::foldShuffleOfShuffles(Instruction &I) { |
| 2967 | ArrayRef<int> OuterMask; |
| 2968 | Value *OuterV0, *OuterV1; |
| 2969 | if (!match(V: &I, |
| 2970 | P: m_Shuffle(v1: m_Value(V&: OuterV0), v2: m_Value(V&: OuterV1), mask: m_Mask(OuterMask)))) |
| 2971 | return false; |
| 2972 | |
| 2973 | ArrayRef<int> InnerMask0, InnerMask1; |
| 2974 | Value *X0, *X1, *Y0, *Y1; |
| 2975 | bool Match0 = |
| 2976 | match(V: OuterV0, P: m_Shuffle(v1: m_Value(V&: X0), v2: m_Value(V&: Y0), mask: m_Mask(InnerMask0))); |
| 2977 | bool Match1 = |
| 2978 | match(V: OuterV1, P: m_Shuffle(v1: m_Value(V&: X1), v2: m_Value(V&: Y1), mask: m_Mask(InnerMask1))); |
| 2979 | if (!Match0 && !Match1) |
| 2980 | return false; |
| 2981 | |
| 2982 | // If the outer shuffle is a permute, then create a fake inner all-poison |
| 2983 | // shuffle. This is easier than accounting for length-changing shuffles below. |
| 2984 | SmallVector<int, 16> PoisonMask1; |
| 2985 | if (!Match1 && isa<PoisonValue>(Val: OuterV1)) { |
| 2986 | X1 = X0; |
| 2987 | Y1 = Y0; |
| 2988 | PoisonMask1.append(NumInputs: InnerMask0.size(), Elt: PoisonMaskElem); |
| 2989 | InnerMask1 = PoisonMask1; |
| 2990 | Match1 = true; // fake match |
| 2991 | } |
| 2992 | |
| 2993 | X0 = Match0 ? X0 : OuterV0; |
| 2994 | Y0 = Match0 ? Y0 : OuterV0; |
| 2995 | X1 = Match1 ? X1 : OuterV1; |
| 2996 | Y1 = Match1 ? Y1 : OuterV1; |
| 2997 | auto *ShuffleDstTy = dyn_cast<FixedVectorType>(Val: I.getType()); |
| 2998 | auto *ShuffleSrcTy = dyn_cast<FixedVectorType>(Val: X0->getType()); |
| 2999 | auto *ShuffleImmTy = dyn_cast<FixedVectorType>(Val: OuterV0->getType()); |
| 3000 | if (!ShuffleDstTy || !ShuffleSrcTy || !ShuffleImmTy || |
| 3001 | X0->getType() != X1->getType()) |
| 3002 | return false; |
| 3003 | |
| 3004 | unsigned NumSrcElts = ShuffleSrcTy->getNumElements(); |
| 3005 | unsigned NumImmElts = ShuffleImmTy->getNumElements(); |
| 3006 | |
| 3007 | // Attempt to merge shuffles, matching upto 2 source operands. |
| 3008 | // Replace index to a poison arg with PoisonMaskElem. |
| 3009 | // Bail if either inner masks reference an undef arg. |
| 3010 | SmallVector<int, 16> NewMask(OuterMask); |
| 3011 | Value *NewX = nullptr, *NewY = nullptr; |
| 3012 | for (int &M : NewMask) { |
| 3013 | Value *Src = nullptr; |
| 3014 | if (0 <= M && M < (int)NumImmElts) { |
| 3015 | Src = OuterV0; |
| 3016 | if (Match0) { |
| 3017 | M = InnerMask0[M]; |
| 3018 | Src = M >= (int)NumSrcElts ? Y0 : X0; |
| 3019 | M = M >= (int)NumSrcElts ? (M - NumSrcElts) : M; |
| 3020 | } |
| 3021 | } else if (M >= (int)NumImmElts) { |
| 3022 | Src = OuterV1; |
| 3023 | M -= NumImmElts; |
| 3024 | if (Match1) { |
| 3025 | M = InnerMask1[M]; |
| 3026 | Src = M >= (int)NumSrcElts ? Y1 : X1; |
| 3027 | M = M >= (int)NumSrcElts ? (M - NumSrcElts) : M; |
| 3028 | } |
| 3029 | } |
| 3030 | if (Src && M != PoisonMaskElem) { |
| 3031 | assert(0 <= M && M < (int)NumSrcElts && "Unexpected shuffle mask index" ); |
| 3032 | if (isa<UndefValue>(Val: Src)) { |
| 3033 | // We've referenced an undef element - if its poison, update the shuffle |
| 3034 | // mask, else bail. |
| 3035 | if (!isa<PoisonValue>(Val: Src)) |
| 3036 | return false; |
| 3037 | M = PoisonMaskElem; |
| 3038 | continue; |
| 3039 | } |
| 3040 | if (!NewX || NewX == Src) { |
| 3041 | NewX = Src; |
| 3042 | continue; |
| 3043 | } |
| 3044 | if (!NewY || NewY == Src) { |
| 3045 | M += NumSrcElts; |
| 3046 | NewY = Src; |
| 3047 | continue; |
| 3048 | } |
| 3049 | return false; |
| 3050 | } |
| 3051 | } |
| 3052 | |
| 3053 | if (!NewX) { |
| 3054 | replaceValue(Old&: I, New&: *PoisonValue::get(T: ShuffleDstTy)); |
| 3055 | return true; |
| 3056 | } |
| 3057 | |
| 3058 | if (!NewY) |
| 3059 | NewY = PoisonValue::get(T: ShuffleSrcTy); |
| 3060 | |
| 3061 | // Have we folded to an Identity shuffle? |
| 3062 | if (ShuffleVectorInst::isIdentityMask(Mask: NewMask, NumSrcElts)) { |
| 3063 | replaceValue(Old&: I, New&: *NewX); |
| 3064 | return true; |
| 3065 | } |
| 3066 | |
| 3067 | // Try to merge the shuffles if the new shuffle is not costly. |
| 3068 | InstructionCost InnerCost0 = 0; |
| 3069 | if (Match0) |
| 3070 | InnerCost0 = TTI.getInstructionCost(U: cast<User>(Val: OuterV0), CostKind); |
| 3071 | |
| 3072 | InstructionCost InnerCost1 = 0; |
| 3073 | if (Match1) |
| 3074 | InnerCost1 = TTI.getInstructionCost(U: cast<User>(Val: OuterV1), CostKind); |
| 3075 | |
| 3076 | InstructionCost OuterCost = TTI.getInstructionCost(U: &I, CostKind); |
| 3077 | |
| 3078 | InstructionCost OldCost = InnerCost0 + InnerCost1 + OuterCost; |
| 3079 | |
| 3080 | bool IsUnary = all_of(Range&: NewMask, P: [&](int M) { return M < (int)NumSrcElts; }); |
| 3081 | TargetTransformInfo::ShuffleKind SK = |
| 3082 | IsUnary ? TargetTransformInfo::SK_PermuteSingleSrc |
| 3083 | : TargetTransformInfo::SK_PermuteTwoSrc; |
| 3084 | InstructionCost NewCost = |
| 3085 | TTI.getShuffleCost(Kind: SK, DstTy: ShuffleDstTy, SrcTy: ShuffleSrcTy, Mask: NewMask, CostKind, Index: 0, |
| 3086 | SubTp: nullptr, Args: {NewX, NewY}); |
| 3087 | if (!OuterV0->hasOneUse()) |
| 3088 | NewCost += InnerCost0; |
| 3089 | if (!OuterV1->hasOneUse()) |
| 3090 | NewCost += InnerCost1; |
| 3091 | |
| 3092 | LLVM_DEBUG(dbgs() << "Found a shuffle feeding two shuffles: " << I |
| 3093 | << "\n OldCost: " << OldCost << " vs NewCost: " << NewCost |
| 3094 | << "\n" ); |
| 3095 | if (NewCost > OldCost) |
| 3096 | return false; |
| 3097 | |
| 3098 | Value *Shuf = Builder.CreateShuffleVector(V1: NewX, V2: NewY, Mask: NewMask); |
| 3099 | replaceValue(Old&: I, New&: *Shuf); |
| 3100 | return true; |
| 3101 | } |
| 3102 | |
| 3103 | /// Try to convert a chain of length-preserving shuffles that are fed by |
| 3104 | /// length-changing shuffles from the same source, e.g. a chain of length 3: |
| 3105 | /// |
| 3106 | /// "shuffle (shuffle (shuffle x, (shuffle y, undef)), |
| 3107 | /// (shuffle y, undef)), |
| 3108 | // (shuffle y, undef)" |
| 3109 | /// |
| 3110 | /// into a single shuffle fed by a length-changing shuffle: |
| 3111 | /// |
| 3112 | /// "shuffle x, (shuffle y, undef)" |
| 3113 | /// |
| 3114 | /// Such chains arise e.g. from folding extract/insert sequences. |
| 3115 | bool VectorCombine::foldShufflesOfLengthChangingShuffles(Instruction &I) { |
| 3116 | FixedVectorType *TrunkType = dyn_cast<FixedVectorType>(Val: I.getType()); |
| 3117 | if (!TrunkType) |
| 3118 | return false; |
| 3119 | |
| 3120 | unsigned ChainLength = 0; |
| 3121 | SmallVector<int> Mask; |
| 3122 | SmallVector<int> YMask; |
| 3123 | InstructionCost OldCost = 0; |
| 3124 | InstructionCost NewCost = 0; |
| 3125 | Value *Trunk = &I; |
| 3126 | unsigned NumTrunkElts = TrunkType->getNumElements(); |
| 3127 | Value *Y = nullptr; |
| 3128 | |
| 3129 | for (;;) { |
| 3130 | // Match the current trunk against (commutations of) the pattern |
| 3131 | // "shuffle trunk', (shuffle y, undef)" |
| 3132 | ArrayRef<int> OuterMask; |
| 3133 | Value *OuterV0, *OuterV1; |
| 3134 | if (ChainLength != 0 && !Trunk->hasOneUse()) |
| 3135 | break; |
| 3136 | if (!match(V: Trunk, P: m_Shuffle(v1: m_Value(V&: OuterV0), v2: m_Value(V&: OuterV1), |
| 3137 | mask: m_Mask(OuterMask)))) |
| 3138 | break; |
| 3139 | if (OuterV0->getType() != TrunkType) { |
| 3140 | // This shuffle is not length-preserving, so it cannot be part of the |
| 3141 | // chain. |
| 3142 | break; |
| 3143 | } |
| 3144 | |
| 3145 | ArrayRef<int> InnerMask0, InnerMask1; |
| 3146 | Value *A0, *A1, *B0, *B1; |
| 3147 | bool Match0 = |
| 3148 | match(V: OuterV0, P: m_Shuffle(v1: m_Value(V&: A0), v2: m_Value(V&: B0), mask: m_Mask(InnerMask0))); |
| 3149 | bool Match1 = |
| 3150 | match(V: OuterV1, P: m_Shuffle(v1: m_Value(V&: A1), v2: m_Value(V&: B1), mask: m_Mask(InnerMask1))); |
| 3151 | bool Match0Leaf = Match0 && A0->getType() != I.getType(); |
| 3152 | bool Match1Leaf = Match1 && A1->getType() != I.getType(); |
| 3153 | if (Match0Leaf == Match1Leaf) { |
| 3154 | // Only handle the case of exactly one leaf in each step. The "two leaves" |
| 3155 | // case is handled by foldShuffleOfShuffles. |
| 3156 | break; |
| 3157 | } |
| 3158 | |
| 3159 | SmallVector<int> CommutedOuterMask; |
| 3160 | if (Match0Leaf) { |
| 3161 | std::swap(a&: OuterV0, b&: OuterV1); |
| 3162 | std::swap(a&: InnerMask0, b&: InnerMask1); |
| 3163 | std::swap(a&: A0, b&: A1); |
| 3164 | std::swap(a&: B0, b&: B1); |
| 3165 | llvm::append_range(C&: CommutedOuterMask, R&: OuterMask); |
| 3166 | for (int &M : CommutedOuterMask) { |
| 3167 | if (M == PoisonMaskElem) |
| 3168 | continue; |
| 3169 | if (M < (int)NumTrunkElts) |
| 3170 | M += NumTrunkElts; |
| 3171 | else |
| 3172 | M -= NumTrunkElts; |
| 3173 | } |
| 3174 | OuterMask = CommutedOuterMask; |
| 3175 | } |
| 3176 | if (!OuterV1->hasOneUse()) |
| 3177 | break; |
| 3178 | |
| 3179 | if (!isa<UndefValue>(Val: A1)) { |
| 3180 | if (!Y) |
| 3181 | Y = A1; |
| 3182 | else if (Y != A1) |
| 3183 | break; |
| 3184 | } |
| 3185 | if (!isa<UndefValue>(Val: B1)) { |
| 3186 | if (!Y) |
| 3187 | Y = B1; |
| 3188 | else if (Y != B1) |
| 3189 | break; |
| 3190 | } |
| 3191 | |
| 3192 | auto *YType = cast<FixedVectorType>(Val: A1->getType()); |
| 3193 | int NumLeafElts = YType->getNumElements(); |
| 3194 | SmallVector<int> LocalYMask(InnerMask1); |
| 3195 | for (int &M : LocalYMask) { |
| 3196 | if (M >= NumLeafElts) |
| 3197 | M -= NumLeafElts; |
| 3198 | } |
| 3199 | |
| 3200 | InstructionCost LocalOldCost = |
| 3201 | TTI.getInstructionCost(U: cast<User>(Val: Trunk), CostKind) + |
| 3202 | TTI.getInstructionCost(U: cast<User>(Val: OuterV1), CostKind); |
| 3203 | |
| 3204 | // Handle the initial (start of chain) case. |
| 3205 | if (!ChainLength) { |
| 3206 | Mask.assign(AR: OuterMask); |
| 3207 | YMask.assign(RHS: LocalYMask); |
| 3208 | OldCost = NewCost = LocalOldCost; |
| 3209 | Trunk = OuterV0; |
| 3210 | ChainLength++; |
| 3211 | continue; |
| 3212 | } |
| 3213 | |
| 3214 | // For the non-root case, first attempt to combine masks. |
| 3215 | SmallVector<int> NewYMask(YMask); |
| 3216 | bool Valid = true; |
| 3217 | for (auto [CombinedM, LeafM] : llvm::zip(t&: NewYMask, u&: LocalYMask)) { |
| 3218 | if (LeafM == -1 || CombinedM == LeafM) |
| 3219 | continue; |
| 3220 | if (CombinedM == -1) { |
| 3221 | CombinedM = LeafM; |
| 3222 | } else { |
| 3223 | Valid = false; |
| 3224 | break; |
| 3225 | } |
| 3226 | } |
| 3227 | if (!Valid) |
| 3228 | break; |
| 3229 | |
| 3230 | SmallVector<int> NewMask; |
| 3231 | NewMask.reserve(N: NumTrunkElts); |
| 3232 | for (int M : Mask) { |
| 3233 | if (M < 0 || M >= static_cast<int>(NumTrunkElts)) |
| 3234 | NewMask.push_back(Elt: M); |
| 3235 | else |
| 3236 | NewMask.push_back(Elt: OuterMask[M]); |
| 3237 | } |
| 3238 | |
| 3239 | // Break the chain if adding this new step complicates the shuffles such |
| 3240 | // that it would increase the new cost by more than the old cost of this |
| 3241 | // step. |
| 3242 | InstructionCost LocalNewCost = |
| 3243 | TTI.getShuffleCost(Kind: TargetTransformInfo::SK_PermuteSingleSrc, DstTy: TrunkType, |
| 3244 | SrcTy: YType, Mask: NewYMask, CostKind) + |
| 3245 | TTI.getShuffleCost(Kind: TargetTransformInfo::SK_PermuteTwoSrc, DstTy: TrunkType, |
| 3246 | SrcTy: TrunkType, Mask: NewMask, CostKind); |
| 3247 | |
| 3248 | if (LocalNewCost >= NewCost && LocalOldCost < LocalNewCost - NewCost) |
| 3249 | break; |
| 3250 | |
| 3251 | LLVM_DEBUG({ |
| 3252 | if (ChainLength == 1) { |
| 3253 | dbgs() << "Found chain of shuffles fed by length-changing shuffles: " |
| 3254 | << I << '\n'; |
| 3255 | } |
| 3256 | dbgs() << " next chain link: " << *Trunk << '\n' |
| 3257 | << " old cost: " << (OldCost + LocalOldCost) |
| 3258 | << " new cost: " << LocalNewCost << '\n'; |
| 3259 | }); |
| 3260 | |
| 3261 | Mask = NewMask; |
| 3262 | YMask = NewYMask; |
| 3263 | OldCost += LocalOldCost; |
| 3264 | NewCost = LocalNewCost; |
| 3265 | Trunk = OuterV0; |
| 3266 | ChainLength++; |
| 3267 | } |
| 3268 | if (ChainLength <= 1) |
| 3269 | return false; |
| 3270 | |
| 3271 | // Bail out if all leaves were poison. |
| 3272 | if (!Y) |
| 3273 | return false; |
| 3274 | |
| 3275 | if (llvm::all_of(Range&: Mask, P: [&](int M) { |
| 3276 | return M < 0 || M >= static_cast<int>(NumTrunkElts); |
| 3277 | })) { |
| 3278 | // Produce a canonical simplified form if all elements are sourced from Y. |
| 3279 | for (int &M : Mask) { |
| 3280 | if (M >= static_cast<int>(NumTrunkElts)) |
| 3281 | M = YMask[M - NumTrunkElts]; |
| 3282 | } |
| 3283 | Value *Root = |
| 3284 | Builder.CreateShuffleVector(V1: Y, V2: PoisonValue::get(T: Y->getType()), Mask); |
| 3285 | replaceValue(Old&: I, New&: *Root); |
| 3286 | return true; |
| 3287 | } |
| 3288 | |
| 3289 | Value *Leaf = |
| 3290 | Builder.CreateShuffleVector(V1: Y, V2: PoisonValue::get(T: Y->getType()), Mask: YMask); |
| 3291 | Value *Root = Builder.CreateShuffleVector(V1: Trunk, V2: Leaf, Mask); |
| 3292 | replaceValue(Old&: I, New&: *Root); |
| 3293 | return true; |
| 3294 | } |
| 3295 | |
| 3296 | /// Try to convert |
| 3297 | /// "shuffle (intrinsic), (intrinsic)" into "intrinsic (shuffle), (shuffle)". |
| 3298 | bool VectorCombine::foldShuffleOfIntrinsics(Instruction &I) { |
| 3299 | Value *V0, *V1; |
| 3300 | ArrayRef<int> OldMask; |
| 3301 | if (!match(V: &I, P: m_Shuffle(v1: m_Value(V&: V0), v2: m_Value(V&: V1), mask: m_Mask(OldMask)))) |
| 3302 | return false; |
| 3303 | |
| 3304 | auto *II0 = dyn_cast<IntrinsicInst>(Val: V0); |
| 3305 | auto *II1 = dyn_cast<IntrinsicInst>(Val: V1); |
| 3306 | if (!II0 || !II1) |
| 3307 | return false; |
| 3308 | |
| 3309 | Intrinsic::ID IID = II0->getIntrinsicID(); |
| 3310 | if (IID != II1->getIntrinsicID()) |
| 3311 | return false; |
| 3312 | InstructionCost CostII0 = |
| 3313 | TTI.getIntrinsicInstrCost(ICA: IntrinsicCostAttributes(IID, *II0), CostKind); |
| 3314 | InstructionCost CostII1 = |
| 3315 | TTI.getIntrinsicInstrCost(ICA: IntrinsicCostAttributes(IID, *II1), CostKind); |
| 3316 | |
| 3317 | auto *ShuffleDstTy = dyn_cast<FixedVectorType>(Val: I.getType()); |
| 3318 | auto *II0Ty = dyn_cast<FixedVectorType>(Val: II0->getType()); |
| 3319 | if (!ShuffleDstTy || !II0Ty) |
| 3320 | return false; |
| 3321 | |
| 3322 | if (!isTriviallyVectorizable(ID: IID)) |
| 3323 | return false; |
| 3324 | |
| 3325 | for (unsigned I = 0, E = II0->arg_size(); I != E; ++I) { |
| 3326 | Value *Arg0 = II0->getArgOperand(i: I); |
| 3327 | Value *Arg1 = II1->getArgOperand(i: I); |
| 3328 | if (isVectorIntrinsicWithScalarOpAtArg(ID: IID, ScalarOpdIdx: I, TTI: &TTI)) { |
| 3329 | // Scalar operands must be identical. |
| 3330 | if (Arg0 != Arg1) |
| 3331 | return false; |
| 3332 | } else if (Arg0->getType() != Arg1->getType()) { |
| 3333 | // The corresponding vector operands are shuffled together, so they must |
| 3334 | // share the same type. For intrinsics overloaded on their operand type |
| 3335 | // (e.g. llvm.fptosi.sat), two calls can produce the same result type |
| 3336 | // from different operand types; shuffling those would be invalid. |
| 3337 | return false; |
| 3338 | } |
| 3339 | } |
| 3340 | |
| 3341 | InstructionCost OldCost = |
| 3342 | CostII0 + CostII1 + |
| 3343 | TTI.getShuffleCost(Kind: TargetTransformInfo::SK_PermuteTwoSrc, DstTy: ShuffleDstTy, |
| 3344 | SrcTy: II0Ty, Mask: OldMask, CostKind, Index: 0, SubTp: nullptr, Args: {II0, II1}, CxtI: &I); |
| 3345 | |
| 3346 | SmallVector<Type *> NewArgsTy; |
| 3347 | InstructionCost NewCost = 0; |
| 3348 | SmallDenseSet<std::pair<Value *, Value *>> SeenOperandPairs; |
| 3349 | for (unsigned I = 0, E = II0->arg_size(); I != E; ++I) { |
| 3350 | if (isVectorIntrinsicWithScalarOpAtArg(ID: IID, ScalarOpdIdx: I, TTI: &TTI)) { |
| 3351 | NewArgsTy.push_back(Elt: II0->getArgOperand(i: I)->getType()); |
| 3352 | } else { |
| 3353 | auto *VecTy = cast<FixedVectorType>(Val: II0->getArgOperand(i: I)->getType()); |
| 3354 | auto *ArgTy = FixedVectorType::get(ElementType: VecTy->getElementType(), |
| 3355 | NumElts: ShuffleDstTy->getNumElements()); |
| 3356 | NewArgsTy.push_back(Elt: ArgTy); |
| 3357 | std::pair<Value *, Value *> OperandPair = |
| 3358 | std::make_pair(x: II0->getArgOperand(i: I), y: II1->getArgOperand(i: I)); |
| 3359 | if (!SeenOperandPairs.insert(V: OperandPair).second) { |
| 3360 | // We've already computed the cost for this operand pair. |
| 3361 | continue; |
| 3362 | } |
| 3363 | NewCost += TTI.getShuffleCost( |
| 3364 | Kind: TargetTransformInfo::SK_PermuteTwoSrc, DstTy: ArgTy, SrcTy: VecTy, Mask: OldMask, |
| 3365 | CostKind, Index: 0, SubTp: nullptr, Args: {II0->getArgOperand(i: I), II1->getArgOperand(i: I)}); |
| 3366 | } |
| 3367 | } |
| 3368 | IntrinsicCostAttributes NewAttr(IID, ShuffleDstTy, NewArgsTy); |
| 3369 | |
| 3370 | NewCost += TTI.getIntrinsicInstrCost(ICA: NewAttr, CostKind); |
| 3371 | if (!II0->hasOneUse()) |
| 3372 | NewCost += CostII0; |
| 3373 | if (II1 != II0 && !II1->hasOneUse()) |
| 3374 | NewCost += CostII1; |
| 3375 | |
| 3376 | LLVM_DEBUG(dbgs() << "Found a shuffle feeding two intrinsics: " << I |
| 3377 | << "\n OldCost: " << OldCost << " vs NewCost: " << NewCost |
| 3378 | << "\n" ); |
| 3379 | |
| 3380 | if (NewCost > OldCost) |
| 3381 | return false; |
| 3382 | |
| 3383 | SmallVector<Value *> NewArgs; |
| 3384 | SmallDenseMap<std::pair<Value *, Value *>, Value *> ShuffleCache; |
| 3385 | for (unsigned I = 0, E = II0->arg_size(); I != E; ++I) |
| 3386 | if (isVectorIntrinsicWithScalarOpAtArg(ID: IID, ScalarOpdIdx: I, TTI: &TTI)) { |
| 3387 | NewArgs.push_back(Elt: II0->getArgOperand(i: I)); |
| 3388 | } else { |
| 3389 | std::pair<Value *, Value *> OperandPair = |
| 3390 | std::make_pair(x: II0->getArgOperand(i: I), y: II1->getArgOperand(i: I)); |
| 3391 | auto It = ShuffleCache.find(Val: OperandPair); |
| 3392 | if (It != ShuffleCache.end()) { |
| 3393 | // Reuse previously created shuffle for this operand pair. |
| 3394 | NewArgs.push_back(Elt: It->second); |
| 3395 | continue; |
| 3396 | } |
| 3397 | Value *Shuf = Builder.CreateShuffleVector(V1: II0->getArgOperand(i: I), |
| 3398 | V2: II1->getArgOperand(i: I), Mask: OldMask); |
| 3399 | ShuffleCache[OperandPair] = Shuf; |
| 3400 | NewArgs.push_back(Elt: Shuf); |
| 3401 | Worklist.pushValue(V: Shuf); |
| 3402 | } |
| 3403 | Value *NewIntrinsic = Builder.CreateIntrinsic(RetTy: ShuffleDstTy, ID: IID, Args: NewArgs); |
| 3404 | |
| 3405 | // Intersect flags from the old intrinsics. |
| 3406 | if (auto *NewInst = dyn_cast<Instruction>(Val: NewIntrinsic)) { |
| 3407 | NewInst->copyIRFlags(V: II0); |
| 3408 | NewInst->andIRFlags(V: II1); |
| 3409 | } |
| 3410 | |
| 3411 | replaceValue(Old&: I, New&: *NewIntrinsic); |
| 3412 | return true; |
| 3413 | } |
| 3414 | |
| 3415 | /// Try to convert |
| 3416 | /// "shuffle (intrinsic), (poison/undef)" into "intrinsic (shuffle)". |
| 3417 | bool VectorCombine::foldPermuteOfIntrinsic(Instruction &I) { |
| 3418 | Value *V0; |
| 3419 | ArrayRef<int> Mask; |
| 3420 | if (!match(V: &I, P: m_Shuffle(v1: m_Value(V&: V0), v2: m_Undef(), mask: m_Mask(Mask)))) |
| 3421 | return false; |
| 3422 | |
| 3423 | auto *II0 = dyn_cast<IntrinsicInst>(Val: V0); |
| 3424 | if (!II0) |
| 3425 | return false; |
| 3426 | |
| 3427 | auto *ShuffleDstTy = dyn_cast<FixedVectorType>(Val: I.getType()); |
| 3428 | auto *IntrinsicSrcTy = dyn_cast<FixedVectorType>(Val: II0->getType()); |
| 3429 | if (!ShuffleDstTy || !IntrinsicSrcTy) |
| 3430 | return false; |
| 3431 | |
| 3432 | // Validate it's a pure permute, mask should only reference the first vector |
| 3433 | unsigned NumSrcElts = IntrinsicSrcTy->getNumElements(); |
| 3434 | if (any_of(Range&: Mask, P: [NumSrcElts](int M) { return M >= (int)NumSrcElts; })) |
| 3435 | return false; |
| 3436 | |
| 3437 | Intrinsic::ID IID = II0->getIntrinsicID(); |
| 3438 | if (!isTriviallyVectorizable(ID: IID)) |
| 3439 | return false; |
| 3440 | |
| 3441 | // Cost analysis |
| 3442 | InstructionCost IntrinsicCost = |
| 3443 | TTI.getIntrinsicInstrCost(ICA: IntrinsicCostAttributes(IID, *II0), CostKind); |
| 3444 | InstructionCost OldCost = |
| 3445 | IntrinsicCost + |
| 3446 | TTI.getShuffleCost(Kind: TargetTransformInfo::SK_PermuteSingleSrc, DstTy: ShuffleDstTy, |
| 3447 | SrcTy: IntrinsicSrcTy, Mask, CostKind, Index: 0, SubTp: nullptr, Args: {V0}, CxtI: &I); |
| 3448 | |
| 3449 | SmallVector<Type *> NewArgsTy; |
| 3450 | InstructionCost NewCost = 0; |
| 3451 | for (unsigned I = 0, E = II0->arg_size(); I != E; ++I) { |
| 3452 | if (isVectorIntrinsicWithScalarOpAtArg(ID: IID, ScalarOpdIdx: I, TTI: &TTI)) { |
| 3453 | NewArgsTy.push_back(Elt: II0->getArgOperand(i: I)->getType()); |
| 3454 | } else { |
| 3455 | auto *VecTy = cast<FixedVectorType>(Val: II0->getArgOperand(i: I)->getType()); |
| 3456 | auto *ArgTy = FixedVectorType::get(ElementType: VecTy->getElementType(), |
| 3457 | NumElts: ShuffleDstTy->getNumElements()); |
| 3458 | NewArgsTy.push_back(Elt: ArgTy); |
| 3459 | NewCost += TTI.getShuffleCost(Kind: TargetTransformInfo::SK_PermuteSingleSrc, |
| 3460 | DstTy: ArgTy, SrcTy: VecTy, Mask, CostKind, Index: 0, SubTp: nullptr, |
| 3461 | Args: {II0->getArgOperand(i: I)}); |
| 3462 | } |
| 3463 | } |
| 3464 | IntrinsicCostAttributes NewAttr(IID, ShuffleDstTy, NewArgsTy); |
| 3465 | NewCost += TTI.getIntrinsicInstrCost(ICA: NewAttr, CostKind); |
| 3466 | |
| 3467 | // If the intrinsic has multiple uses, we need to account for the cost of |
| 3468 | // keeping the original intrinsic around. |
| 3469 | if (!II0->hasOneUse()) |
| 3470 | NewCost += IntrinsicCost; |
| 3471 | |
| 3472 | LLVM_DEBUG(dbgs() << "Found a permute of intrinsic: " << I << "\n OldCost: " |
| 3473 | << OldCost << " vs NewCost: " << NewCost << "\n" ); |
| 3474 | |
| 3475 | if (NewCost > OldCost) |
| 3476 | return false; |
| 3477 | |
| 3478 | // Transform |
| 3479 | SmallVector<Value *> NewArgs; |
| 3480 | for (unsigned I = 0, E = II0->arg_size(); I != E; ++I) { |
| 3481 | if (isVectorIntrinsicWithScalarOpAtArg(ID: IID, ScalarOpdIdx: I, TTI: &TTI)) { |
| 3482 | NewArgs.push_back(Elt: II0->getArgOperand(i: I)); |
| 3483 | } else { |
| 3484 | Value *Shuf = Builder.CreateShuffleVector(V: II0->getArgOperand(i: I), Mask); |
| 3485 | NewArgs.push_back(Elt: Shuf); |
| 3486 | Worklist.pushValue(V: Shuf); |
| 3487 | } |
| 3488 | } |
| 3489 | |
| 3490 | Value *NewIntrinsic = Builder.CreateIntrinsic(RetTy: ShuffleDstTy, ID: IID, Args: NewArgs); |
| 3491 | |
| 3492 | if (auto *NewInst = dyn_cast<Instruction>(Val: NewIntrinsic)) |
| 3493 | NewInst->copyIRFlags(V: II0); |
| 3494 | |
| 3495 | replaceValue(Old&: I, New&: *NewIntrinsic); |
| 3496 | return true; |
| 3497 | } |
| 3498 | |
| 3499 | using InstLane = std::pair<Value *, int>; |
| 3500 | |
| 3501 | static InstLane lookThroughShuffles(Value *V, int Lane) { |
| 3502 | while (auto *SV = dyn_cast<ShuffleVectorInst>(Val: V)) { |
| 3503 | unsigned NumElts = |
| 3504 | cast<FixedVectorType>(Val: SV->getOperand(i_nocapture: 0)->getType())->getNumElements(); |
| 3505 | int M = SV->getMaskValue(Elt: Lane); |
| 3506 | if (M < 0) |
| 3507 | return {nullptr, PoisonMaskElem}; |
| 3508 | if (static_cast<unsigned>(M) < NumElts) { |
| 3509 | V = SV->getOperand(i_nocapture: 0); |
| 3510 | Lane = M; |
| 3511 | } else { |
| 3512 | V = SV->getOperand(i_nocapture: 1); |
| 3513 | Lane = M - NumElts; |
| 3514 | } |
| 3515 | } |
| 3516 | return InstLane{V, Lane}; |
| 3517 | } |
| 3518 | |
| 3519 | static SmallVector<InstLane> |
| 3520 | generateInstLaneVectorFromOperand(ArrayRef<InstLane> Item, int Op) { |
| 3521 | SmallVector<InstLane> NItem; |
| 3522 | for (InstLane IL : Item) { |
| 3523 | auto [U, Lane] = IL; |
| 3524 | InstLane OpLane = |
| 3525 | U ? lookThroughShuffles(V: cast<Instruction>(Val: U)->getOperand(i: Op), Lane) |
| 3526 | : InstLane{nullptr, PoisonMaskElem}; |
| 3527 | NItem.emplace_back(Args&: OpLane); |
| 3528 | } |
| 3529 | return NItem; |
| 3530 | } |
| 3531 | |
| 3532 | /// Detect concat of multiple values into a vector |
| 3533 | static bool isFreeConcat(ArrayRef<InstLane> Item, TTI::TargetCostKind CostKind, |
| 3534 | const TargetTransformInfo &TTI) { |
| 3535 | auto *Ty = cast<FixedVectorType>(Val: Item.front().first->getType()); |
| 3536 | unsigned NumElts = Ty->getNumElements(); |
| 3537 | if (Item.size() == NumElts || NumElts == 1 || Item.size() % NumElts != 0) |
| 3538 | return false; |
| 3539 | |
| 3540 | // Check that the concat is free, usually meaning that the type will be split |
| 3541 | // during legalization. |
| 3542 | SmallVector<int, 16> ConcatMask(NumElts * 2); |
| 3543 | std::iota(first: ConcatMask.begin(), last: ConcatMask.end(), value: 0); |
| 3544 | if (TTI.getShuffleCost(Kind: TTI::SK_PermuteTwoSrc, |
| 3545 | DstTy: FixedVectorType::get(ElementType: Ty->getScalarType(), NumElts: NumElts * 2), |
| 3546 | SrcTy: Ty, Mask: ConcatMask, CostKind) != 0) |
| 3547 | return false; |
| 3548 | |
| 3549 | unsigned NumSlices = Item.size() / NumElts; |
| 3550 | // Currently we generate a tree of shuffles for the concats, which limits us |
| 3551 | // to a power2. |
| 3552 | if (!isPowerOf2_32(Value: NumSlices)) |
| 3553 | return false; |
| 3554 | for (unsigned Slice = 0; Slice < NumSlices; ++Slice) { |
| 3555 | Value *SliceV = Item[Slice * NumElts].first; |
| 3556 | if (!SliceV || SliceV->getType() != Ty) |
| 3557 | return false; |
| 3558 | for (unsigned Elt = 0; Elt < NumElts; ++Elt) { |
| 3559 | auto [V, Lane] = Item[Slice * NumElts + Elt]; |
| 3560 | if (Lane != static_cast<int>(Elt) || SliceV != V) |
| 3561 | return false; |
| 3562 | } |
| 3563 | } |
| 3564 | return true; |
| 3565 | } |
| 3566 | |
| 3567 | static Value * |
| 3568 | generateNewInstTree(ArrayRef<InstLane> Item, Use *From, |
| 3569 | const DenseSet<std::pair<Value *, Use *>> &IdentityLeafs, |
| 3570 | const DenseSet<std::pair<Value *, Use *>> &SplatLeafs, |
| 3571 | const DenseSet<std::pair<Value *, Use *>> &ConcatLeafs, |
| 3572 | IRBuilderBase &Builder, const TargetTransformInfo *TTI) { |
| 3573 | auto [FrontV, FrontLane] = Item.front(); |
| 3574 | |
| 3575 | if (IdentityLeafs.contains(V: std::make_pair(x&: FrontV, y&: From))) { |
| 3576 | return FrontV; |
| 3577 | } |
| 3578 | if (SplatLeafs.contains(V: std::make_pair(x&: FrontV, y&: From))) { |
| 3579 | SmallVector<int, 16> Mask(Item.size(), FrontLane); |
| 3580 | return Builder.CreateShuffleVector(V: FrontV, Mask); |
| 3581 | } |
| 3582 | if (ConcatLeafs.contains(V: std::make_pair(x&: FrontV, y&: From))) { |
| 3583 | unsigned NumElts = |
| 3584 | cast<FixedVectorType>(Val: FrontV->getType())->getNumElements(); |
| 3585 | SmallVector<Value *> Values(Item.size() / NumElts, nullptr); |
| 3586 | for (unsigned S = 0; S < Values.size(); ++S) |
| 3587 | Values[S] = Item[S * NumElts].first; |
| 3588 | |
| 3589 | while (Values.size() > 1) { |
| 3590 | NumElts *= 2; |
| 3591 | SmallVector<int, 16> Mask(NumElts, 0); |
| 3592 | std::iota(first: Mask.begin(), last: Mask.end(), value: 0); |
| 3593 | SmallVector<Value *> NewValues(Values.size() / 2, nullptr); |
| 3594 | for (unsigned S = 0; S < NewValues.size(); ++S) |
| 3595 | NewValues[S] = |
| 3596 | Builder.CreateShuffleVector(V1: Values[S * 2], V2: Values[S * 2 + 1], Mask); |
| 3597 | Values = NewValues; |
| 3598 | } |
| 3599 | return Values[0]; |
| 3600 | } |
| 3601 | |
| 3602 | auto *I = cast<Instruction>(Val: FrontV); |
| 3603 | |
| 3604 | // Handle vector bitcasts that change element count. We cannot use |
| 3605 | // generateInstLaneVectorFromOperand for these because the lane indices |
| 3606 | // don't map 1:1 through the bitcast. |
| 3607 | if (auto *BitCast = dyn_cast<BitCastInst>(Val: I)) { |
| 3608 | auto *BCDstTy = dyn_cast<FixedVectorType>(Val: BitCast->getDestTy()); |
| 3609 | auto *BCSrcTy = dyn_cast<FixedVectorType>(Val: BitCast->getSrcTy()); |
| 3610 | if (BCDstTy && BCSrcTy && |
| 3611 | BCDstTy->getElementCount() != BCSrcTy->getElementCount()) { |
| 3612 | unsigned DstElts = BCDstTy->getNumElements(); |
| 3613 | unsigned SrcElts = BCSrcTy->getNumElements(); |
| 3614 | SmallVector<InstLane> NewItem; |
| 3615 | if (DstElts > SrcElts) { |
| 3616 | // Widening: compress operand Item. |
| 3617 | unsigned R = DstElts / SrcElts; |
| 3618 | if (Item.size() % R != 0) |
| 3619 | return nullptr; |
| 3620 | for (unsigned Idx = 0, E = Item.size(); Idx < E; Idx += R) { |
| 3621 | auto [V, Lane] = Item[Idx]; |
| 3622 | if (!V) { |
| 3623 | NewItem.push_back(Elt: {nullptr, PoisonMaskElem}); |
| 3624 | continue; |
| 3625 | } |
| 3626 | NewItem.push_back( |
| 3627 | Elt: lookThroughShuffles(V: cast<Operator>(Val: V)->getOperand(i: 0), Lane: Lane / R)); |
| 3628 | } |
| 3629 | } else { |
| 3630 | // Narrowing: expand operand Item. |
| 3631 | unsigned R = SrcElts / DstElts; |
| 3632 | for (auto [V, Lane] : Item) { |
| 3633 | if (!V) { |
| 3634 | NewItem.append(NumInputs: R, Elt: {nullptr, PoisonMaskElem}); |
| 3635 | continue; |
| 3636 | } |
| 3637 | Value *Op = cast<Operator>(Val: V)->getOperand(i: 0); |
| 3638 | for (unsigned J = 0; J < R; ++J) |
| 3639 | NewItem.push_back(Elt: lookThroughShuffles(V: Op, Lane: Lane * R + J)); |
| 3640 | } |
| 3641 | } |
| 3642 | Value *Op = generateNewInstTree(Item: NewItem, From: &BitCast->getOperandUse(i: 0), |
| 3643 | IdentityLeafs, SplatLeafs, ConcatLeafs, |
| 3644 | Builder, TTI); |
| 3645 | return Builder.CreateBitCast( |
| 3646 | V: Op, DestTy: FixedVectorType::get(ElementType: BCDstTy->getScalarType(), NumElts: Item.size())); |
| 3647 | } |
| 3648 | } |
| 3649 | auto *II = dyn_cast<IntrinsicInst>(Val: I); |
| 3650 | unsigned NumOps = I->getNumOperands() - (II ? 1 : 0); |
| 3651 | SmallVector<Value *> Ops(NumOps); |
| 3652 | for (unsigned Idx = 0; Idx < NumOps; Idx++) { |
| 3653 | if (II && |
| 3654 | isVectorIntrinsicWithScalarOpAtArg(ID: II->getIntrinsicID(), ScalarOpdIdx: Idx, TTI)) { |
| 3655 | Ops[Idx] = II->getOperand(i_nocapture: Idx); |
| 3656 | continue; |
| 3657 | } |
| 3658 | Ops[Idx] = generateNewInstTree(Item: generateInstLaneVectorFromOperand(Item, Op: Idx), |
| 3659 | From: &I->getOperandUse(i: Idx), IdentityLeafs, |
| 3660 | SplatLeafs, ConcatLeafs, Builder, TTI); |
| 3661 | } |
| 3662 | |
| 3663 | SmallVector<Value *, 8> ValueList; |
| 3664 | for (const auto &Lane : Item) |
| 3665 | if (Lane.first) |
| 3666 | ValueList.push_back(Elt: Lane.first); |
| 3667 | |
| 3668 | Type *DstTy = |
| 3669 | FixedVectorType::get(ElementType: I->getType()->getScalarType(), NumElts: Item.size()); |
| 3670 | if (auto *BI = dyn_cast<BinaryOperator>(Val: I)) { |
| 3671 | auto *Value = Builder.CreateBinOp(Opc: (Instruction::BinaryOps)BI->getOpcode(), |
| 3672 | LHS: Ops[0], RHS: Ops[1]); |
| 3673 | propagateIRFlags(I: Value, VL: ValueList); |
| 3674 | return Value; |
| 3675 | } |
| 3676 | if (auto *CI = dyn_cast<CmpInst>(Val: I)) { |
| 3677 | auto *Value = Builder.CreateCmp(Pred: CI->getPredicate(), LHS: Ops[0], RHS: Ops[1]); |
| 3678 | propagateIRFlags(I: Value, VL: ValueList); |
| 3679 | return Value; |
| 3680 | } |
| 3681 | if (auto *SI = dyn_cast<SelectInst>(Val: I)) { |
| 3682 | auto *Value = Builder.CreateSelect(C: Ops[0], True: Ops[1], False: Ops[2], Name: "" , MDFrom: SI); |
| 3683 | propagateIRFlags(I: Value, VL: ValueList); |
| 3684 | return Value; |
| 3685 | } |
| 3686 | if (auto *CI = dyn_cast<CastInst>(Val: I)) { |
| 3687 | auto *Value = Builder.CreateCast(Op: CI->getOpcode(), V: Ops[0], DestTy: DstTy); |
| 3688 | propagateIRFlags(I: Value, VL: ValueList); |
| 3689 | return Value; |
| 3690 | } |
| 3691 | if (II) { |
| 3692 | auto *Value = Builder.CreateIntrinsic(RetTy: DstTy, ID: II->getIntrinsicID(), Args: Ops); |
| 3693 | propagateIRFlags(I: Value, VL: ValueList); |
| 3694 | return Value; |
| 3695 | } |
| 3696 | assert(isa<UnaryInstruction>(I) && "Unexpected instruction type in Generate" ); |
| 3697 | auto *Value = |
| 3698 | Builder.CreateUnOp(Opc: (Instruction::UnaryOps)I->getOpcode(), V: Ops[0]); |
| 3699 | propagateIRFlags(I: Value, VL: ValueList); |
| 3700 | return Value; |
| 3701 | } |
| 3702 | |
| 3703 | // Starting from a shuffle, look up through operands tracking the shuffled index |
| 3704 | // of each lane. If we can simplify away the shuffles to identities then |
| 3705 | // do so. |
| 3706 | bool VectorCombine::foldShuffleToIdentity(Instruction &I) { |
| 3707 | auto *Ty = dyn_cast<FixedVectorType>(Val: I.getType()); |
| 3708 | if (!Ty || I.use_empty()) |
| 3709 | return false; |
| 3710 | |
| 3711 | SmallVector<InstLane> Start(Ty->getNumElements()); |
| 3712 | for (unsigned M = 0, E = Ty->getNumElements(); M < E; ++M) |
| 3713 | Start[M] = lookThroughShuffles(V: &I, Lane: M); |
| 3714 | |
| 3715 | SmallVector<std::pair<SmallVector<InstLane>, Use *>> Worklist; |
| 3716 | Worklist.push_back(Elt: std::make_pair(x&: Start, y: &*I.use_begin())); |
| 3717 | DenseSet<std::pair<Value *, Use *>> IdentityLeafs, SplatLeafs, ConcatLeafs; |
| 3718 | unsigned NumVisited = 0; |
| 3719 | bool TraversedElCountChangingBitcast = false; |
| 3720 | |
| 3721 | while (!Worklist.empty()) { |
| 3722 | if (++NumVisited > MaxInstrsToScan) |
| 3723 | return false; |
| 3724 | |
| 3725 | auto ItemFrom = Worklist.pop_back_val(); |
| 3726 | auto Item = ItemFrom.first; |
| 3727 | auto From = ItemFrom.second; |
| 3728 | auto [FrontV, FrontLane] = Item.front(); |
| 3729 | |
| 3730 | // If we found an undef first lane then bail out to keep things simple. |
| 3731 | if (!FrontV) |
| 3732 | return false; |
| 3733 | |
| 3734 | // Helper to peek through bitcasts to the same value. |
| 3735 | auto IsEquiv = [&](Value *X, Value *Y) { |
| 3736 | return X->getType() == Y->getType() && |
| 3737 | peekThroughBitcasts(V: X) == peekThroughBitcasts(V: Y); |
| 3738 | }; |
| 3739 | |
| 3740 | // Look for an identity value. |
| 3741 | if (FrontLane == 0 && |
| 3742 | cast<FixedVectorType>(Val: FrontV->getType())->getNumElements() == |
| 3743 | Item.size() && |
| 3744 | all_of(Range: drop_begin(RangeOrContainer: enumerate(First&: Item)), P: [IsEquiv, Item](const auto &E) { |
| 3745 | Value *FrontV = Item.front().first; |
| 3746 | return !E.value().first || (IsEquiv(E.value().first, FrontV) && |
| 3747 | E.value().second == (int)E.index()); |
| 3748 | })) { |
| 3749 | IdentityLeafs.insert(V: std::make_pair(x&: FrontV, y&: From)); |
| 3750 | continue; |
| 3751 | } |
| 3752 | // Look for constants, for the moment only supporting constant splats. |
| 3753 | if (auto *C = dyn_cast<Constant>(Val: FrontV); |
| 3754 | C && C->getSplatValue() && |
| 3755 | all_of(Range: drop_begin(RangeOrContainer&: Item), P: [Item](InstLane &IL) { |
| 3756 | Value *FrontV = Item.front().first; |
| 3757 | Value *V = IL.first; |
| 3758 | return !V || (isa<Constant>(Val: V) && |
| 3759 | cast<Constant>(Val: V)->getSplatValue() == |
| 3760 | cast<Constant>(Val: FrontV)->getSplatValue()); |
| 3761 | })) { |
| 3762 | SplatLeafs.insert(V: std::make_pair(x&: FrontV, y&: From)); |
| 3763 | continue; |
| 3764 | } |
| 3765 | // Look for a splat value. |
| 3766 | if (all_of(Range: drop_begin(RangeOrContainer&: Item), P: [Item](InstLane &IL) { |
| 3767 | auto [FrontV, FrontLane] = Item.front(); |
| 3768 | auto [V, Lane] = IL; |
| 3769 | return !V || (V == FrontV && Lane == FrontLane); |
| 3770 | })) { |
| 3771 | SplatLeafs.insert(V: std::make_pair(x&: FrontV, y&: From)); |
| 3772 | continue; |
| 3773 | } |
| 3774 | |
| 3775 | // We need each element to be the same type of value, and check that each |
| 3776 | // element has a single use. |
| 3777 | auto CheckLaneIsEquivalentToFirst = [Item](InstLane IL) { |
| 3778 | Value *FrontV = Item.front().first; |
| 3779 | if (!IL.first) |
| 3780 | return true; |
| 3781 | Value *V = IL.first; |
| 3782 | if (auto *I = dyn_cast<Instruction>(Val: V); I && !I->hasOneUser()) |
| 3783 | return false; |
| 3784 | if (V->getValueID() != FrontV->getValueID()) |
| 3785 | return false; |
| 3786 | if (auto *CI = dyn_cast<CmpInst>(Val: V)) |
| 3787 | if (CI->getPredicate() != cast<CmpInst>(Val: FrontV)->getPredicate()) |
| 3788 | return false; |
| 3789 | if (auto *CI = dyn_cast<CastInst>(Val: V)) |
| 3790 | if (CI->getSrcTy()->getScalarType() != |
| 3791 | cast<CastInst>(Val: FrontV)->getSrcTy()->getScalarType()) |
| 3792 | return false; |
| 3793 | if (auto *SI = dyn_cast<SelectInst>(Val: V)) |
| 3794 | if (!isa<VectorType>(Val: SI->getOperand(i_nocapture: 0)->getType()) || |
| 3795 | SI->getOperand(i_nocapture: 0)->getType() != |
| 3796 | cast<SelectInst>(Val: FrontV)->getOperand(i_nocapture: 0)->getType()) |
| 3797 | return false; |
| 3798 | if (isa<CallInst>(Val: V) && !isa<IntrinsicInst>(Val: V)) |
| 3799 | return false; |
| 3800 | auto *II = dyn_cast<IntrinsicInst>(Val: V); |
| 3801 | return !II || (isa<IntrinsicInst>(Val: FrontV) && |
| 3802 | II->getIntrinsicID() == |
| 3803 | cast<IntrinsicInst>(Val: FrontV)->getIntrinsicID() && |
| 3804 | !II->hasOperandBundles()); |
| 3805 | }; |
| 3806 | if (all_of(Range: drop_begin(RangeOrContainer&: Item), P: CheckLaneIsEquivalentToFirst)) { |
| 3807 | // Check the operator is one that we support. |
| 3808 | if (isa<BinaryOperator, CmpInst>(Val: FrontV)) { |
| 3809 | // We exclude div/rem in case they hit UB from poison lanes. |
| 3810 | if (auto *BO = dyn_cast<BinaryOperator>(Val: FrontV); |
| 3811 | BO && BO->isIntDivRem()) |
| 3812 | return false; |
| 3813 | Worklist.emplace_back(Args: generateInstLaneVectorFromOperand(Item, Op: 0), |
| 3814 | Args: &cast<Instruction>(Val: FrontV)->getOperandUse(i: 0)); |
| 3815 | Worklist.emplace_back(Args: generateInstLaneVectorFromOperand(Item, Op: 1), |
| 3816 | Args: &cast<Instruction>(Val: FrontV)->getOperandUse(i: 1)); |
| 3817 | continue; |
| 3818 | } else if (isa<UnaryOperator, TruncInst, ZExtInst, SExtInst, FPToSIInst, |
| 3819 | FPToUIInst, SIToFPInst, UIToFPInst>(Val: FrontV)) { |
| 3820 | Worklist.emplace_back(Args: generateInstLaneVectorFromOperand(Item, Op: 0), |
| 3821 | Args: &cast<Instruction>(Val: FrontV)->getOperandUse(i: 0)); |
| 3822 | continue; |
| 3823 | } else if (auto *BitCast = dyn_cast<BitCastInst>(Val: FrontV)) { |
| 3824 | auto *BCDstTy = dyn_cast<FixedVectorType>(Val: BitCast->getDestTy()); |
| 3825 | auto *BCSrcTy = dyn_cast<FixedVectorType>(Val: BitCast->getSrcTy()); |
| 3826 | if (BCDstTy && BCSrcTy) { |
| 3827 | ElementCount DstEC = BCDstTy->getElementCount(); |
| 3828 | ElementCount SrcEC = BCSrcTy->getElementCount(); |
| 3829 | if (DstEC == SrcEC) { |
| 3830 | // Same element count - simple pass-through. |
| 3831 | Worklist.emplace_back(Args: generateInstLaneVectorFromOperand(Item, Op: 0), |
| 3832 | Args: &BitCast->getOperandUse(i: 0)); |
| 3833 | continue; |
| 3834 | } |
| 3835 | unsigned DstElts = DstEC.getFixedValue(); |
| 3836 | unsigned SrcElts = SrcEC.getFixedValue(); |
| 3837 | if (DstElts > SrcElts && DstElts % SrcElts == 0) { |
| 3838 | // Widening bitcast (e.g. <2 x i32> -> <4 x i16>). Compress |
| 3839 | // consecutive groups of R destination lanes into one source |
| 3840 | // lane. |
| 3841 | unsigned R = DstElts / SrcElts; |
| 3842 | SmallVector<InstLane> NItem; |
| 3843 | bool Valid = Item.size() % R == 0; |
| 3844 | for (unsigned Idx = 0, E = Item.size(); Valid && Idx < E; |
| 3845 | Idx += R) { |
| 3846 | auto [V0, L0] = Item[Idx]; |
| 3847 | if (!V0) { |
| 3848 | if (any_of(Range: ArrayRef(Item).slice(N: Idx + 1, M: R - 1), |
| 3849 | P: [](InstLane IL) { return IL.first != nullptr; })) { |
| 3850 | Valid = false; |
| 3851 | break; |
| 3852 | } |
| 3853 | NItem.push_back(Elt: {nullptr, PoisonMaskElem}); |
| 3854 | continue; |
| 3855 | } |
| 3856 | if (L0 % R != 0) { |
| 3857 | Valid = false; |
| 3858 | break; |
| 3859 | } |
| 3860 | for (unsigned J = 1; J < R; ++J) { |
| 3861 | auto [VJ, LJ] = Item[Idx + J]; |
| 3862 | if (!VJ || VJ != V0 || LJ != L0 + (int)J) { |
| 3863 | Valid = false; |
| 3864 | break; |
| 3865 | } |
| 3866 | } |
| 3867 | if (!Valid) |
| 3868 | break; |
| 3869 | NItem.push_back(Elt: lookThroughShuffles( |
| 3870 | V: cast<Operator>(Val: V0)->getOperand(i: 0), Lane: L0 / R)); |
| 3871 | } |
| 3872 | if (Valid) { |
| 3873 | TraversedElCountChangingBitcast = true; |
| 3874 | Worklist.emplace_back(Args&: NItem, Args: &BitCast->getOperandUse(i: 0)); |
| 3875 | continue; |
| 3876 | } |
| 3877 | } else if (SrcElts > DstElts && SrcElts % DstElts == 0) { |
| 3878 | // Narrowing bitcast (e.g. <4 x i16> -> <2 x i32>). Expand |
| 3879 | // each destination lane into R source lanes. |
| 3880 | unsigned R = SrcElts / DstElts; |
| 3881 | SmallVector<InstLane> NItem; |
| 3882 | for (auto [V, Lane] : Item) { |
| 3883 | if (!V) { |
| 3884 | NItem.append(NumInputs: R, Elt: {nullptr, PoisonMaskElem}); |
| 3885 | continue; |
| 3886 | } |
| 3887 | Value *Op = cast<Operator>(Val: V)->getOperand(i: 0); |
| 3888 | for (unsigned J = 0; J < R; ++J) |
| 3889 | NItem.push_back(Elt: lookThroughShuffles(V: Op, Lane: Lane * R + J)); |
| 3890 | } |
| 3891 | TraversedElCountChangingBitcast = true; |
| 3892 | Worklist.emplace_back(Args&: NItem, Args: &BitCast->getOperandUse(i: 0)); |
| 3893 | continue; |
| 3894 | } |
| 3895 | } |
| 3896 | } else if (auto *Sel = dyn_cast<SelectInst>(Val: FrontV)) { |
| 3897 | Worklist.emplace_back(Args: generateInstLaneVectorFromOperand(Item, Op: 0), |
| 3898 | Args: &Sel->getOperandUse(i: 0)); |
| 3899 | Worklist.emplace_back(Args: generateInstLaneVectorFromOperand(Item, Op: 1), |
| 3900 | Args: &Sel->getOperandUse(i: 1)); |
| 3901 | Worklist.emplace_back(Args: generateInstLaneVectorFromOperand(Item, Op: 2), |
| 3902 | Args: &Sel->getOperandUse(i: 2)); |
| 3903 | continue; |
| 3904 | } else if (auto *II = dyn_cast<IntrinsicInst>(Val: FrontV); |
| 3905 | II && isTriviallyVectorizable(ID: II->getIntrinsicID()) && |
| 3906 | !II->hasOperandBundles()) { |
| 3907 | for (unsigned Op = 0, E = II->getNumOperands() - 1; Op < E; Op++) { |
| 3908 | if (isVectorIntrinsicWithScalarOpAtArg(ID: II->getIntrinsicID(), ScalarOpdIdx: Op, |
| 3909 | TTI: &TTI)) { |
| 3910 | if (!all_of(Range: drop_begin(RangeOrContainer&: Item), P: [Item, Op](InstLane &IL) { |
| 3911 | Value *FrontV = Item.front().first; |
| 3912 | Value *V = IL.first; |
| 3913 | return !V || (cast<Instruction>(Val: V)->getOperand(i: Op) == |
| 3914 | cast<Instruction>(Val: FrontV)->getOperand(i: Op)); |
| 3915 | })) |
| 3916 | return false; |
| 3917 | continue; |
| 3918 | } |
| 3919 | Worklist.emplace_back(Args: generateInstLaneVectorFromOperand(Item, Op), |
| 3920 | Args: &cast<Instruction>(Val: FrontV)->getOperandUse(i: Op)); |
| 3921 | } |
| 3922 | continue; |
| 3923 | } |
| 3924 | } |
| 3925 | |
| 3926 | if (isFreeConcat(Item, CostKind, TTI)) { |
| 3927 | ConcatLeafs.insert(V: std::make_pair(x&: FrontV, y&: From)); |
| 3928 | continue; |
| 3929 | } |
| 3930 | |
| 3931 | return false; |
| 3932 | } |
| 3933 | |
| 3934 | if (NumVisited <= 1) |
| 3935 | return false; |
| 3936 | |
| 3937 | // If the only non-leaf node traversed was a single bitcast that changes |
| 3938 | // element count, the fold would just commute the bitcast and shuffle. |
| 3939 | // foldBitcastShuffle does the reverse transform, causing an infinite loop. |
| 3940 | if (NumVisited == 2 && TraversedElCountChangingBitcast) |
| 3941 | return false; |
| 3942 | |
| 3943 | LLVM_DEBUG(dbgs() << "Found a superfluous identity shuffle: " << I << "\n" ); |
| 3944 | |
| 3945 | // If we got this far, we know the shuffles are superfluous and can be |
| 3946 | // removed. Scan through again and generate the new tree of instructions. |
| 3947 | Builder.SetInsertPoint(&I); |
| 3948 | Value *V = generateNewInstTree(Item: Start, From: &*I.use_begin(), IdentityLeafs, |
| 3949 | SplatLeafs, ConcatLeafs, Builder, TTI: &TTI); |
| 3950 | replaceValue(Old&: I, New&: *V); |
| 3951 | return true; |
| 3952 | } |
| 3953 | |
| 3954 | /// Given a commutative reduction, the order of the input lanes does not alter |
| 3955 | /// the results. We can use this to remove certain shuffles feeding the |
| 3956 | /// reduction, removing the need to shuffle at all. |
| 3957 | bool VectorCombine::foldShuffleFromReductions(Instruction &I) { |
| 3958 | auto *II = dyn_cast<IntrinsicInst>(Val: &I); |
| 3959 | if (!II) |
| 3960 | return false; |
| 3961 | switch (II->getIntrinsicID()) { |
| 3962 | case Intrinsic::vector_reduce_add: |
| 3963 | case Intrinsic::vector_reduce_mul: |
| 3964 | case Intrinsic::vector_reduce_and: |
| 3965 | case Intrinsic::vector_reduce_or: |
| 3966 | case Intrinsic::vector_reduce_xor: |
| 3967 | case Intrinsic::vector_reduce_smin: |
| 3968 | case Intrinsic::vector_reduce_smax: |
| 3969 | case Intrinsic::vector_reduce_umin: |
| 3970 | case Intrinsic::vector_reduce_umax: |
| 3971 | break; |
| 3972 | default: |
| 3973 | return false; |
| 3974 | } |
| 3975 | |
| 3976 | // Find all the inputs when looking through operations that do not alter the |
| 3977 | // lane order (binops, for example). Currently we look for a single shuffle, |
| 3978 | // and can ignore splat values. |
| 3979 | std::queue<Value *> Worklist; |
| 3980 | SmallPtrSet<Value *, 4> Visited; |
| 3981 | ShuffleVectorInst *Shuffle = nullptr; |
| 3982 | if (auto *Op = dyn_cast<Instruction>(Val: I.getOperand(i: 0))) |
| 3983 | Worklist.push(x: Op); |
| 3984 | |
| 3985 | while (!Worklist.empty()) { |
| 3986 | Value *CV = Worklist.front(); |
| 3987 | Worklist.pop(); |
| 3988 | if (Visited.contains(Ptr: CV)) |
| 3989 | continue; |
| 3990 | |
| 3991 | // Splats don't change the order, so can be safely ignored. |
| 3992 | if (isSplatValue(V: CV)) |
| 3993 | continue; |
| 3994 | |
| 3995 | Visited.insert(Ptr: CV); |
| 3996 | |
| 3997 | if (auto *CI = dyn_cast<Instruction>(Val: CV)) { |
| 3998 | if (CI->isBinaryOp()) { |
| 3999 | for (auto *Op : CI->operand_values()) |
| 4000 | Worklist.push(x: Op); |
| 4001 | continue; |
| 4002 | } else if (auto *SV = dyn_cast<ShuffleVectorInst>(Val: CI)) { |
| 4003 | if (Shuffle && Shuffle != SV) |
| 4004 | return false; |
| 4005 | Shuffle = SV; |
| 4006 | continue; |
| 4007 | } |
| 4008 | } |
| 4009 | |
| 4010 | // Anything else is currently an unknown node. |
| 4011 | return false; |
| 4012 | } |
| 4013 | |
| 4014 | if (!Shuffle) |
| 4015 | return false; |
| 4016 | |
| 4017 | // Check all uses of the binary ops and shuffles are also included in the |
| 4018 | // lane-invariant operations (Visited should be the list of lanewise |
| 4019 | // instructions, including the shuffle that we found). |
| 4020 | for (auto *V : Visited) |
| 4021 | for (auto *U : V->users()) |
| 4022 | if (!Visited.contains(Ptr: U) && U != &I) |
| 4023 | return false; |
| 4024 | |
| 4025 | FixedVectorType *VecType = |
| 4026 | dyn_cast<FixedVectorType>(Val: II->getOperand(i_nocapture: 0)->getType()); |
| 4027 | if (!VecType) |
| 4028 | return false; |
| 4029 | FixedVectorType *ShuffleInputType = |
| 4030 | dyn_cast<FixedVectorType>(Val: Shuffle->getOperand(i_nocapture: 0)->getType()); |
| 4031 | if (!ShuffleInputType) |
| 4032 | return false; |
| 4033 | unsigned NumInputElts = ShuffleInputType->getNumElements(); |
| 4034 | |
| 4035 | // Find the mask from sorting the lanes into order. This is most likely to |
| 4036 | // become a identity or concat mask. Undef elements are pushed to the end. |
| 4037 | SmallVector<int> ConcatMask; |
| 4038 | Shuffle->getShuffleMask(Result&: ConcatMask); |
| 4039 | sort(C&: ConcatMask, Comp: [](int X, int Y) { return (unsigned)X < (unsigned)Y; }); |
| 4040 | bool UsesSecondVec = |
| 4041 | any_of(Range&: ConcatMask, P: [&](int M) { return M >= (int)NumInputElts; }); |
| 4042 | |
| 4043 | InstructionCost OldCost = TTI.getShuffleCost( |
| 4044 | Kind: UsesSecondVec ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc, DstTy: VecType, |
| 4045 | SrcTy: ShuffleInputType, Mask: Shuffle->getShuffleMask(), CostKind); |
| 4046 | InstructionCost NewCost = TTI.getShuffleCost( |
| 4047 | Kind: UsesSecondVec ? TTI::SK_PermuteTwoSrc : TTI::SK_PermuteSingleSrc, DstTy: VecType, |
| 4048 | SrcTy: ShuffleInputType, Mask: ConcatMask, CostKind); |
| 4049 | |
| 4050 | LLVM_DEBUG(dbgs() << "Found a reduction feeding from a shuffle: " << *Shuffle |
| 4051 | << "\n" ); |
| 4052 | LLVM_DEBUG(dbgs() << " OldCost: " << OldCost << " vs NewCost: " << NewCost |
| 4053 | << "\n" ); |
| 4054 | bool MadeChanges = false; |
| 4055 | if (NewCost < OldCost) { |
| 4056 | Builder.SetInsertPoint(Shuffle); |
| 4057 | Value *NewShuffle = Builder.CreateShuffleVector( |
| 4058 | V1: Shuffle->getOperand(i_nocapture: 0), V2: Shuffle->getOperand(i_nocapture: 1), Mask: ConcatMask); |
| 4059 | LLVM_DEBUG(dbgs() << "Created new shuffle: " << *NewShuffle << "\n" ); |
| 4060 | replaceValue(Old&: *Shuffle, New&: *NewShuffle); |
| 4061 | return true; |
| 4062 | } |
| 4063 | |
| 4064 | // See if we can re-use foldSelectShuffle, getting it to reduce the size of |
| 4065 | // the shuffle into a nicer order, as it can ignore the order of the shuffles. |
| 4066 | MadeChanges |= foldSelectShuffle(I&: *Shuffle, FromReduction: true); |
| 4067 | return MadeChanges; |
| 4068 | } |
| 4069 | |
| 4070 | /// Try to fold a chain of shuffles and ops feeding extractelement(..., 0) |
| 4071 | /// into llvm.vector.reduce.*, by tracking which lanes contribute to the |
| 4072 | /// extracted lane and reducing the widest vector whose lanes each contribute |
| 4073 | /// once. |
| 4074 | /// |
| 4075 | /// For example: |
| 4076 | /// |
| 4077 | /// %lo = shufflevector <4 x i32> %a, poison, <2 x i32> <i32 0, i32 1> |
| 4078 | /// %hi = shufflevector <4 x i32> %a, poison, <2 x i32> <i32 2, i32 3> |
| 4079 | /// %s = add <2 x i32> %lo, %hi |
| 4080 | /// %sh = shufflevector <2 x i32> %s, poison, <2 x i32> <i32 1, i32 poison> |
| 4081 | /// %r = add <2 x i32> %s, %sh |
| 4082 | /// %e = extractelement <2 x i32> %r, i64 0 |
| 4083 | /// |
| 4084 | /// transforms to: |
| 4085 | /// |
| 4086 | /// %e = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> %a) |
| 4087 | bool VectorCombine::foldShuffleChainsToReduce(Instruction &I) { |
| 4088 | Value *VecOpEE; |
| 4089 | if (!match(V: &I, P: m_ExtractElt(Val: m_Value(V&: VecOpEE), Idx: m_Zero()))) |
| 4090 | return false; |
| 4091 | |
| 4092 | auto *FVT = dyn_cast<FixedVectorType>(Val: VecOpEE->getType()); |
| 4093 | if (!FVT) |
| 4094 | return false; |
| 4095 | |
| 4096 | if (FVT->getNumElements() < 2) |
| 4097 | return false; |
| 4098 | |
| 4099 | std::optional<Instruction::BinaryOps> CommonBinOp; |
| 4100 | std::optional<Intrinsic::ID> CommonCallOp; |
| 4101 | |
| 4102 | if (auto *BO = dyn_cast<BinaryOperator>(Val: VecOpEE)) { |
| 4103 | if (!getReductionForBinop(Opc: BO->getOpcode())) |
| 4104 | return false; |
| 4105 | CommonBinOp = BO->getOpcode(); |
| 4106 | } else if (auto *MMI = dyn_cast<MinMaxIntrinsic>(Val: VecOpEE)) { |
| 4107 | CommonCallOp = MMI->getIntrinsicID(); |
| 4108 | } else { |
| 4109 | return false; |
| 4110 | } |
| 4111 | |
| 4112 | // For floating-point reductions, track FMF intersection across all binops. |
| 4113 | FastMathFlags CommonFMF; |
| 4114 | bool IsFloatReduction = false; |
| 4115 | |
| 4116 | // A chain node is one we walk through, either a matching-opcode binop/min-max |
| 4117 | // or a single-source shuffle. Anything else is a leaf source. |
| 4118 | auto IsChainNode = [&](Value *V) { |
| 4119 | if (auto *BO = dyn_cast<BinaryOperator>(Val: V)) |
| 4120 | return CommonBinOp && BO->getOpcode() == *CommonBinOp; |
| 4121 | if (auto *MMI = dyn_cast<MinMaxIntrinsic>(Val: V)) |
| 4122 | return CommonCallOp && MMI->getIntrinsicID() == *CommonCallOp; |
| 4123 | if (auto *SVI = dyn_cast<ShuffleVectorInst>(Val: V)) |
| 4124 | return isa<PoisonValue>(Val: SVI->getOperand(i_nocapture: 1)); |
| 4125 | return false; |
| 4126 | }; |
| 4127 | |
| 4128 | // Collect the chain, building Nodes in postorder. Bail if the chain is empty |
| 4129 | // or exceeds MaxChainNodes. |
| 4130 | constexpr unsigned MaxChainNodes = 32; |
| 4131 | SmallSetVector<Value *, 16> Nodes; |
| 4132 | SmallSetVector<Value *, 4> Sources; |
| 4133 | unsigned NumVisited = 0; |
| 4134 | auto AddSource = [&](Value *V) { |
| 4135 | if (!isa<FixedVectorType>(Val: V->getType())) |
| 4136 | return false; |
| 4137 | Sources.insert(X: V); |
| 4138 | return true; |
| 4139 | }; |
| 4140 | auto Walk = [&](Value *V, auto &&Walk) -> bool { |
| 4141 | if (Nodes.contains(key: V) || Sources.contains(key: V)) |
| 4142 | return true; |
| 4143 | if (++NumVisited > MaxChainNodes) |
| 4144 | return false; |
| 4145 | if (!IsChainNode(V)) |
| 4146 | return AddSource(V); |
| 4147 | // Chain shuffles always have poison as op1, so only op0 matters. |
| 4148 | auto *U = cast<Instruction>(Val: V); |
| 4149 | unsigned NumOps = isa<ShuffleVectorInst>(Val: U) ? 1 : 2; |
| 4150 | for (unsigned I = 0; I != NumOps; ++I) |
| 4151 | if (!Walk(U->getOperand(i: I), Walk)) |
| 4152 | return false; |
| 4153 | if (isa<ShuffleVectorInst>(Val: U) || Nodes.contains(key: U->getOperand(i: 0)) || |
| 4154 | Nodes.contains(key: U->getOperand(i: 1))) { |
| 4155 | Nodes.insert(X: V); |
| 4156 | return true; |
| 4157 | } |
| 4158 | // Both operands are leaves so treat this binop as a source rather than |
| 4159 | // walking into it. |
| 4160 | return AddSource(V); |
| 4161 | }; |
| 4162 | if (!Walk(VecOpEE, Walk) || Nodes.empty()) |
| 4163 | return false; |
| 4164 | |
| 4165 | bool IsIdempotent = |
| 4166 | CommonCallOp || (CommonBinOp && Instruction::isIdempotent(Opcode: *CommonBinOp)); |
| 4167 | |
| 4168 | // For FP reductions, require reassoc on every binop and collect FMF. |
| 4169 | for (Value *V : Nodes) { |
| 4170 | auto *BinOp = dyn_cast<BinaryOperator>(Val: V); |
| 4171 | if (!BinOp || !BinOp->getType()->isFPOrFPVectorTy()) |
| 4172 | continue; |
| 4173 | if (!BinOp->hasAllowReassoc()) |
| 4174 | return false; |
| 4175 | if (!IsFloatReduction) { |
| 4176 | CommonFMF = BinOp->getFastMathFlags(); |
| 4177 | IsFloatReduction = true; |
| 4178 | } else { |
| 4179 | CommonFMF &= BinOp->getFastMathFlags(); |
| 4180 | } |
| 4181 | } |
| 4182 | |
| 4183 | // Top-down demanded elements. For each chain value, track which lanes feed |
| 4184 | // the extracted lane 0 and which feed it more than once. Reverse postorder |
| 4185 | // visits every use before its value. A binop forwards its demand to both |
| 4186 | // operands and a shuffle follows its mask back to the source lane. |
| 4187 | struct Demand { |
| 4188 | APInt Lanes; |
| 4189 | APInt Duplicates; |
| 4190 | }; |
| 4191 | DenseMap<Value *, Demand> Demands; |
| 4192 | auto DemandOf = [&](Value *V) -> Demand & { |
| 4193 | unsigned N = cast<FixedVectorType>(Val: V->getType())->getNumElements(); |
| 4194 | Demand &D = Demands[V]; |
| 4195 | if (D.Lanes.getBitWidth() != N) |
| 4196 | D.Lanes = D.Duplicates = APInt::getZero(numBits: N); |
| 4197 | return D; |
| 4198 | }; |
| 4199 | DemandOf(VecOpEE).Lanes.setBit(0); |
| 4200 | for (Value *V : reverse(C&: Nodes)) { |
| 4201 | Demand DV = Demands.lookup(Val: V); |
| 4202 | if (DV.Lanes.isZero()) |
| 4203 | continue; |
| 4204 | if (auto *SVI = dyn_cast<ShuffleVectorInst>(Val: V)) { |
| 4205 | ArrayRef<int> Mask = SVI->getShuffleMask(); |
| 4206 | Demand &DS = DemandOf(SVI->getOperand(i_nocapture: 0)); |
| 4207 | for (unsigned I = 0, E = Mask.size(); I != E; ++I) { |
| 4208 | // Skip lanes that are undemanded or map to poison. |
| 4209 | if (!DV.Lanes[I] || Mask[I] < 0 || |
| 4210 | (unsigned)Mask[I] >= DS.Lanes.getBitWidth()) |
| 4211 | continue; |
| 4212 | if (DS.Lanes[Mask[I]] || DV.Duplicates[I]) |
| 4213 | DS.Duplicates.setBit(Mask[I]); |
| 4214 | DS.Lanes.setBit(Mask[I]); |
| 4215 | } |
| 4216 | } else { |
| 4217 | auto *U = cast<User>(Val: V); |
| 4218 | for (Value *Op : {U->getOperand(i: 0), U->getOperand(i: 1)}) { |
| 4219 | Demand &DOp = DemandOf(Op); |
| 4220 | // Lanes demanded through more than one path accumulate in Duplicates. |
| 4221 | DOp.Duplicates |= DV.Duplicates | (DOp.Lanes & DV.Lanes); |
| 4222 | DOp.Lanes |= DV.Lanes; |
| 4223 | } |
| 4224 | } |
| 4225 | } |
| 4226 | |
| 4227 | // Reducing V replaces the entire chain, so every contribution to the result |
| 4228 | // must flow through V. Reject if anything above V reads outside the chain. |
| 4229 | auto CoversChain = [&](Value *V) { |
| 4230 | SmallVector<Value *, 8> Worklist(1, VecOpEE); |
| 4231 | SmallPtrSet<Value *, 8> Seen; |
| 4232 | Seen.insert(Ptr: VecOpEE); |
| 4233 | while (!Worklist.empty()) { |
| 4234 | auto *U = cast<Instruction>(Val: Worklist.pop_back_val()); |
| 4235 | unsigned NumOps = isa<ShuffleVectorInst>(Val: U) ? 1 : 2; |
| 4236 | for (unsigned I = 0; I != NumOps; ++I) { |
| 4237 | Value *Op = U->getOperand(i: I); |
| 4238 | if (Op == V || !Seen.insert(Ptr: Op).second) |
| 4239 | continue; |
| 4240 | if (!Nodes.contains(key: Op)) |
| 4241 | return false; |
| 4242 | Worklist.push_back(Elt: Op); |
| 4243 | } |
| 4244 | } |
| 4245 | return true; |
| 4246 | }; |
| 4247 | |
| 4248 | // Reduce a single cleanly demanded source if there is one, otherwise the |
| 4249 | // deepest intermediate that covers the chain. |
| 4250 | struct ReductionCut { |
| 4251 | Value *Src; |
| 4252 | APInt Elts; |
| 4253 | }; |
| 4254 | std::optional<ReductionCut> Cut; |
| 4255 | for (Value *S : Sources) { |
| 4256 | auto It = Demands.find(Val: S); |
| 4257 | if (It == Demands.end() || It->second.Lanes.isZero()) |
| 4258 | continue; |
| 4259 | if (Cut || (!IsIdempotent && !It->second.Duplicates.isZero())) { |
| 4260 | Cut.reset(); |
| 4261 | break; |
| 4262 | } |
| 4263 | Cut = ReductionCut{.Src: S, .Elts: It->second.Lanes}; |
| 4264 | } |
| 4265 | if (!Cut) { |
| 4266 | for (Value *V : Nodes) { |
| 4267 | if (!isa<BinaryOperator>(Val: V) && !isa<MinMaxIntrinsic>(Val: V)) |
| 4268 | continue; |
| 4269 | auto It = Demands.find(Val: V); |
| 4270 | if (It == Demands.end() || !It->second.Lanes.isAllOnes()) |
| 4271 | continue; |
| 4272 | if (!IsIdempotent && !It->second.Duplicates.isZero()) |
| 4273 | continue; |
| 4274 | if (!CoversChain(V)) |
| 4275 | continue; |
| 4276 | Cut = ReductionCut{.Src: V, .Elts: It->second.Lanes}; |
| 4277 | break; |
| 4278 | } |
| 4279 | } |
| 4280 | // Reducing one lane is just an extract and can refold forever. |
| 4281 | if (!Cut || Cut->Elts.popcount() < 2) |
| 4282 | return false; |
| 4283 | |
| 4284 | Intrinsic::ID ReducedOp = |
| 4285 | (CommonCallOp ? getMinMaxReductionIntrinsicID(IID: *CommonCallOp) |
| 4286 | : getReductionForBinop(Opc: *CommonBinOp)); |
| 4287 | if (!ReducedOp) |
| 4288 | return false; |
| 4289 | |
| 4290 | InstructionCost OrigCost = 0; |
| 4291 | for (Value *V : Nodes) |
| 4292 | OrigCost += TTI.getInstructionCost(U: cast<Instruction>(Val: V), CostKind); |
| 4293 | |
| 4294 | auto *SrcVT = cast<FixedVectorType>(Val: Cut->Src->getType()); |
| 4295 | bool IsPartialReduction = !Cut->Elts.isAllOnes(); |
| 4296 | FixedVectorType *ReduceVecTy = |
| 4297 | IsPartialReduction |
| 4298 | ? FixedVectorType::get(ElementType: FVT->getElementType(), NumElts: Cut->Elts.popcount()) |
| 4299 | : SrcVT; |
| 4300 | |
| 4301 | SmallVector<int> ; |
| 4302 | InstructionCost NewCost = 0; |
| 4303 | if (IsPartialReduction) { |
| 4304 | for (unsigned I = 0, E = Cut->Elts.getBitWidth(); I != E; ++I) |
| 4305 | if (Cut->Elts[I]) |
| 4306 | ExtractMask.push_back(Elt: I); |
| 4307 | unsigned SubIdx = 0, SubLen; |
| 4308 | auto SK = Cut->Elts.isShiftedMask(MaskIdx&: SubIdx, MaskLen&: SubLen) |
| 4309 | ? TargetTransformInfo::SK_ExtractSubvector |
| 4310 | : TargetTransformInfo::SK_PermuteSingleSrc; |
| 4311 | NewCost += TTI.getShuffleCost(Kind: SK, DstTy: ReduceVecTy, SrcTy: SrcVT, Mask: ExtractMask, CostKind, |
| 4312 | Index: SubIdx, SubTp: ReduceVecTy); |
| 4313 | } |
| 4314 | |
| 4315 | IntrinsicCostAttributes ICA( |
| 4316 | ReducedOp, ReduceVecTy->getElementType(), |
| 4317 | IsFloatReduction |
| 4318 | ? SmallVector<Type *, 2>{ReduceVecTy->getElementType(), ReduceVecTy} |
| 4319 | : SmallVector<Type *, 2>{ReduceVecTy}, |
| 4320 | IsFloatReduction ? CommonFMF : FastMathFlags()); |
| 4321 | NewCost += TTI.getIntrinsicInstrCost(ICA, CostKind); |
| 4322 | |
| 4323 | LLVM_DEBUG(dbgs() << "Found reduction shuffle chain: " << I << "\n OldCost : " |
| 4324 | << OrigCost << " vs NewCost: " << NewCost << "\n" ); |
| 4325 | |
| 4326 | if (!OrigCost.isValid() || !NewCost.isValid()) |
| 4327 | return false; |
| 4328 | |
| 4329 | if (VecOpEE->hasOneUse() ? (NewCost > OrigCost) : (NewCost >= OrigCost)) |
| 4330 | return false; |
| 4331 | |
| 4332 | Value *ReduceInput = Cut->Src; |
| 4333 | if (IsPartialReduction) |
| 4334 | ReduceInput = Builder.CreateShuffleVector(V: Cut->Src, Mask: ExtractMask); |
| 4335 | |
| 4336 | Value *ReducedResult; |
| 4337 | if (IsFloatReduction) { |
| 4338 | Value *Identity = ConstantExpr::getBinOpIdentity( |
| 4339 | Opcode: *CommonBinOp, Ty: ReduceVecTy->getElementType(), /*AllowRHSConstant=*/false, |
| 4340 | NSZ: CommonFMF.noSignedZeros()); |
| 4341 | ReducedResult = Builder.CreateIntrinsic(ID: ReducedOp, OverloadTypes: {ReduceVecTy}, |
| 4342 | Args: {Identity, ReduceInput}, FMFSource: CommonFMF); |
| 4343 | } else { |
| 4344 | ReducedResult = |
| 4345 | Builder.CreateIntrinsic(ID: ReducedOp, OverloadTypes: {ReduceVecTy}, Args: {ReduceInput}); |
| 4346 | } |
| 4347 | replaceValue(Old&: I, New&: *ReducedResult); |
| 4348 | |
| 4349 | return true; |
| 4350 | } |
| 4351 | |
| 4352 | /// Determine if its more efficient to fold: |
| 4353 | /// reduce(trunc(x)) -> trunc(reduce(x)). |
| 4354 | /// reduce(sext(x)) -> sext(reduce(x)). |
| 4355 | /// reduce(zext(x)) -> zext(reduce(x)). |
| 4356 | bool VectorCombine::foldCastFromReductions(Instruction &I) { |
| 4357 | auto *II = dyn_cast<IntrinsicInst>(Val: &I); |
| 4358 | if (!II) |
| 4359 | return false; |
| 4360 | |
| 4361 | bool TruncOnly = false; |
| 4362 | Intrinsic::ID IID = II->getIntrinsicID(); |
| 4363 | switch (IID) { |
| 4364 | case Intrinsic::vector_reduce_add: |
| 4365 | case Intrinsic::vector_reduce_mul: |
| 4366 | TruncOnly = true; |
| 4367 | break; |
| 4368 | case Intrinsic::vector_reduce_and: |
| 4369 | case Intrinsic::vector_reduce_or: |
| 4370 | case Intrinsic::vector_reduce_xor: |
| 4371 | break; |
| 4372 | default: |
| 4373 | return false; |
| 4374 | } |
| 4375 | |
| 4376 | unsigned ReductionOpc = getArithmeticReductionInstruction(RdxID: IID); |
| 4377 | Value *ReductionSrc = I.getOperand(i: 0); |
| 4378 | |
| 4379 | Value *Src; |
| 4380 | if (!match(V: ReductionSrc, P: m_OneUse(SubPattern: m_Trunc(Op: m_Value(V&: Src)))) && |
| 4381 | (TruncOnly || !match(V: ReductionSrc, P: m_OneUse(SubPattern: m_ZExtOrSExt(Op: m_Value(V&: Src)))))) |
| 4382 | return false; |
| 4383 | |
| 4384 | auto CastOpc = |
| 4385 | (Instruction::CastOps)cast<Instruction>(Val: ReductionSrc)->getOpcode(); |
| 4386 | |
| 4387 | auto *SrcTy = cast<VectorType>(Val: Src->getType()); |
| 4388 | auto *ReductionSrcTy = cast<VectorType>(Val: ReductionSrc->getType()); |
| 4389 | Type *ResultTy = I.getType(); |
| 4390 | |
| 4391 | InstructionCost OldCost = TTI.getArithmeticReductionCost( |
| 4392 | Opcode: ReductionOpc, Ty: ReductionSrcTy, FMF: std::nullopt, CostKind); |
| 4393 | OldCost += TTI.getCastInstrCost(Opcode: CastOpc, Dst: ReductionSrcTy, Src: SrcTy, |
| 4394 | CCH: TTI::CastContextHint::None, CostKind, |
| 4395 | I: cast<CastInst>(Val: ReductionSrc)); |
| 4396 | InstructionCost NewCost = |
| 4397 | TTI.getArithmeticReductionCost(Opcode: ReductionOpc, Ty: SrcTy, FMF: std::nullopt, |
| 4398 | CostKind) + |
| 4399 | TTI.getCastInstrCost(Opcode: CastOpc, Dst: ResultTy, Src: ReductionSrcTy->getScalarType(), |
| 4400 | CCH: TTI::CastContextHint::None, CostKind); |
| 4401 | |
| 4402 | if (OldCost <= NewCost || !NewCost.isValid()) |
| 4403 | return false; |
| 4404 | |
| 4405 | Value *NewReduction = Builder.CreateIntrinsic(RetTy: SrcTy->getScalarType(), |
| 4406 | ID: II->getIntrinsicID(), Args: {Src}); |
| 4407 | Value *NewCast = Builder.CreateCast(Op: CastOpc, V: NewReduction, DestTy: ResultTy); |
| 4408 | replaceValue(Old&: I, New&: *NewCast); |
| 4409 | return true; |
| 4410 | } |
| 4411 | |
| 4412 | /// Fold: |
| 4413 | /// icmp pred (reduce.{add,or,and,umax,umin}(signbit_extract(x))), C |
| 4414 | /// into: |
| 4415 | /// icmp sgt/slt (reduce.{or,umax,and,umin}(x)), -1/0 |
| 4416 | /// |
| 4417 | /// Sign-bit reductions produce values with known semantics: |
| 4418 | /// - reduce.{or,umax}: 0 if no element is negative, 1 if any is |
| 4419 | /// - reduce.{and,umin}: 1 if all elements are negative, 0 if any isn't |
| 4420 | /// - reduce.add: count of negative elements (0 to NumElts) |
| 4421 | /// |
| 4422 | /// Both lshr and ashr are supported: |
| 4423 | /// - lshr produces 0 or 1, so reduce.add range is [0, N] |
| 4424 | /// - ashr produces 0 or -1, so reduce.add range is [-N, 0] |
| 4425 | /// |
| 4426 | /// The fold generalizes to multiple source vectors combined with the same |
| 4427 | /// operation as the reduction. For example: |
| 4428 | /// reduce.or(or(shr A, shr B)) conceptually extends the vector |
| 4429 | /// For reduce.add, this changes the count to M*N where M is the number of |
| 4430 | /// source vectors. |
| 4431 | /// |
| 4432 | /// We transform to a direct sign check on the original vector using |
| 4433 | /// reduce.{or,umax} or reduce.{and,umin}. |
| 4434 | /// |
| 4435 | /// In spirit, it's similar to foldSignBitCheck in InstCombine. |
| 4436 | bool VectorCombine::foldSignBitReductionCmp(Instruction &I) { |
| 4437 | CmpPredicate Pred; |
| 4438 | IntrinsicInst *ReduceOp; |
| 4439 | const APInt *CmpVal; |
| 4440 | if (!match(V: &I, |
| 4441 | P: m_ICmp(Pred, L: m_OneUse(SubPattern: m_AnyIntrinsic(I&: ReduceOp)), R: m_APInt(Res&: CmpVal)))) |
| 4442 | return false; |
| 4443 | |
| 4444 | Intrinsic::ID OrigIID = ReduceOp->getIntrinsicID(); |
| 4445 | switch (OrigIID) { |
| 4446 | case Intrinsic::vector_reduce_or: |
| 4447 | case Intrinsic::vector_reduce_umax: |
| 4448 | case Intrinsic::vector_reduce_and: |
| 4449 | case Intrinsic::vector_reduce_umin: |
| 4450 | case Intrinsic::vector_reduce_add: |
| 4451 | break; |
| 4452 | default: |
| 4453 | return false; |
| 4454 | } |
| 4455 | |
| 4456 | Value *ReductionSrc = ReduceOp->getArgOperand(i: 0); |
| 4457 | auto *VecTy = dyn_cast<FixedVectorType>(Val: ReductionSrc->getType()); |
| 4458 | if (!VecTy) |
| 4459 | return false; |
| 4460 | |
| 4461 | unsigned BitWidth = VecTy->getScalarSizeInBits(); |
| 4462 | if (BitWidth == 1) |
| 4463 | return false; |
| 4464 | |
| 4465 | unsigned NumElts = VecTy->getNumElements(); |
| 4466 | |
| 4467 | // Determine the expected tree opcode for multi-vector patterns. |
| 4468 | // The tree opcode must match the reduction's underlying operation. |
| 4469 | // |
| 4470 | // TODO: for pairs of equivalent operators, we should match both, |
| 4471 | // not only the most common. |
| 4472 | Instruction::BinaryOps TreeOpcode; |
| 4473 | switch (OrigIID) { |
| 4474 | case Intrinsic::vector_reduce_or: |
| 4475 | case Intrinsic::vector_reduce_umax: |
| 4476 | TreeOpcode = Instruction::Or; |
| 4477 | break; |
| 4478 | case Intrinsic::vector_reduce_and: |
| 4479 | case Intrinsic::vector_reduce_umin: |
| 4480 | TreeOpcode = Instruction::And; |
| 4481 | break; |
| 4482 | case Intrinsic::vector_reduce_add: |
| 4483 | TreeOpcode = Instruction::Add; |
| 4484 | break; |
| 4485 | default: |
| 4486 | llvm_unreachable("Unexpected intrinsic" ); |
| 4487 | } |
| 4488 | |
| 4489 | // Collect sign-bit extraction leaves from an associative tree of TreeOpcode. |
| 4490 | // The tree conceptually extends the vector being reduced. |
| 4491 | SmallVector<Value *, 8> Worklist; |
| 4492 | SmallVector<Value *, 8> Sources; // Original vectors (X in shr X, BW-1) |
| 4493 | Worklist.push_back(Elt: ReductionSrc); |
| 4494 | std::optional<bool> IsAShr; |
| 4495 | constexpr unsigned MaxSources = 8; |
| 4496 | |
| 4497 | // Calculate old cost: all shifts + tree ops + reduction |
| 4498 | InstructionCost OldCost = TTI.getInstructionCost(U: ReduceOp, CostKind); |
| 4499 | |
| 4500 | while (!Worklist.empty() && Worklist.size() <= MaxSources && |
| 4501 | Sources.size() <= MaxSources) { |
| 4502 | Value *V = Worklist.pop_back_val(); |
| 4503 | |
| 4504 | // Try to match sign-bit extraction: shr X, (bitwidth-1) |
| 4505 | Value *X; |
| 4506 | if (match(V, P: m_OneUse(SubPattern: m_Shr(L: m_Value(V&: X), R: m_SpecificInt(V: BitWidth - 1))))) { |
| 4507 | auto *Shr = cast<Instruction>(Val: V); |
| 4508 | |
| 4509 | // All shifts must be the same type (all lshr or all ashr) |
| 4510 | bool ThisIsAShr = Shr->getOpcode() == Instruction::AShr; |
| 4511 | if (!IsAShr) |
| 4512 | IsAShr = ThisIsAShr; |
| 4513 | else if (*IsAShr != ThisIsAShr) |
| 4514 | return false; |
| 4515 | |
| 4516 | Sources.push_back(Elt: X); |
| 4517 | |
| 4518 | // As part of the fold, we remove all of the shifts, so we need to keep |
| 4519 | // track of their costs. |
| 4520 | OldCost += TTI.getInstructionCost(U: Shr, CostKind); |
| 4521 | |
| 4522 | continue; |
| 4523 | } |
| 4524 | |
| 4525 | // Try to extend through a tree node of the expected opcode |
| 4526 | Value *A, *B; |
| 4527 | if (!match(V, P: m_OneUse(SubPattern: m_BinOp(Opcode: TreeOpcode, L: m_Value(V&: A), R: m_Value(V&: B))))) |
| 4528 | return false; |
| 4529 | |
| 4530 | // We are potentially replacing these operations as well, so we add them |
| 4531 | // to the costs. |
| 4532 | OldCost += TTI.getInstructionCost(U: cast<Instruction>(Val: V), CostKind); |
| 4533 | |
| 4534 | Worklist.push_back(Elt: A); |
| 4535 | Worklist.push_back(Elt: B); |
| 4536 | } |
| 4537 | |
| 4538 | // Must have at least one source and not exceed limit |
| 4539 | if (Sources.empty() || Sources.size() > MaxSources || |
| 4540 | Worklist.size() > MaxSources || !IsAShr) |
| 4541 | return false; |
| 4542 | |
| 4543 | unsigned NumSources = Sources.size(); |
| 4544 | |
| 4545 | // For reduce.add, the total count must fit as a signed integer. |
| 4546 | // Range is [0, M*N] for lshr or [-M*N, 0] for ashr. |
| 4547 | if (OrigIID == Intrinsic::vector_reduce_add && |
| 4548 | !isIntN(N: BitWidth, x: NumSources * NumElts)) |
| 4549 | return false; |
| 4550 | |
| 4551 | // Compute the boundary value when all elements are negative: |
| 4552 | // - Per-element contribution: 1 for lshr, -1 for ashr |
| 4553 | // - For add: M*N (total elements across all sources); for others: just 1 |
| 4554 | unsigned Count = |
| 4555 | (OrigIID == Intrinsic::vector_reduce_add) ? NumSources * NumElts : 1; |
| 4556 | APInt NegativeVal(CmpVal->getBitWidth(), Count); |
| 4557 | if (*IsAShr) |
| 4558 | NegativeVal.negate(); |
| 4559 | |
| 4560 | // Range is [min(0, AllNegVal), max(0, AllNegVal)] |
| 4561 | APInt Zero = APInt::getZero(numBits: CmpVal->getBitWidth()); |
| 4562 | APInt RangeLow = APIntOps::smin(A: Zero, B: NegativeVal); |
| 4563 | APInt RangeHigh = APIntOps::smax(A: Zero, B: NegativeVal); |
| 4564 | |
| 4565 | // Determine comparison semantics: |
| 4566 | // - IsEq: true for equality test, false for inequality |
| 4567 | // - TestsNegative: true if testing against AllNegVal, false for zero |
| 4568 | // |
| 4569 | // In addition to EQ/NE against 0 or AllNegVal, we support inequalities |
| 4570 | // that fold to boundary tests given the narrow value range: |
| 4571 | // < RangeHigh -> != RangeHigh |
| 4572 | // > RangeHigh-1 -> == RangeHigh |
| 4573 | // > RangeLow -> != RangeLow |
| 4574 | // < RangeLow+1 -> == RangeLow |
| 4575 | // |
| 4576 | // For inequalities, we work with signed predicates only. Unsigned predicates |
| 4577 | // are canonicalized to signed when the range is non-negative (where they are |
| 4578 | // equivalent). When the range includes negative values, unsigned predicates |
| 4579 | // would have different semantics due to wrap-around, so we reject them. |
| 4580 | if (!ICmpInst::isEquality(P: Pred) && !ICmpInst::isSigned(Pred)) { |
| 4581 | if (RangeLow.isNegative()) |
| 4582 | return false; |
| 4583 | Pred = ICmpInst::getSignedPredicate(Pred); |
| 4584 | } |
| 4585 | |
| 4586 | bool IsEq; |
| 4587 | bool TestsNegative; |
| 4588 | if (ICmpInst::isEquality(P: Pred)) { |
| 4589 | if (CmpVal->isZero()) { |
| 4590 | TestsNegative = false; |
| 4591 | } else if (*CmpVal == NegativeVal) { |
| 4592 | TestsNegative = true; |
| 4593 | } else { |
| 4594 | return false; |
| 4595 | } |
| 4596 | IsEq = Pred == ICmpInst::ICMP_EQ; |
| 4597 | } else if (Pred == ICmpInst::ICMP_SLT && *CmpVal == RangeHigh) { |
| 4598 | IsEq = false; |
| 4599 | TestsNegative = (RangeHigh == NegativeVal); |
| 4600 | } else if (Pred == ICmpInst::ICMP_SGT && *CmpVal == RangeHigh - 1) { |
| 4601 | IsEq = true; |
| 4602 | TestsNegative = (RangeHigh == NegativeVal); |
| 4603 | } else if (Pred == ICmpInst::ICMP_SGT && *CmpVal == RangeLow) { |
| 4604 | IsEq = false; |
| 4605 | TestsNegative = (RangeLow == NegativeVal); |
| 4606 | } else if (Pred == ICmpInst::ICMP_SLT && *CmpVal == RangeLow + 1) { |
| 4607 | IsEq = true; |
| 4608 | TestsNegative = (RangeLow == NegativeVal); |
| 4609 | } else { |
| 4610 | return false; |
| 4611 | } |
| 4612 | |
| 4613 | // For this fold we support four types of checks: |
| 4614 | // |
| 4615 | // 1. All lanes are negative - AllNeg |
| 4616 | // 2. All lanes are non-negative - AllNonNeg |
| 4617 | // 3. At least one negative lane - AnyNeg |
| 4618 | // 4. At least one non-negative lane - AnyNonNeg |
| 4619 | // |
| 4620 | // For each case, we can generate the following code: |
| 4621 | // |
| 4622 | // 1. AllNeg - reduce.and/umin(X) < 0 |
| 4623 | // 2. AllNonNeg - reduce.or/umax(X) > -1 |
| 4624 | // 3. AnyNeg - reduce.or/umax(X) < 0 |
| 4625 | // 4. AnyNonNeg - reduce.and/umin(X) > -1 |
| 4626 | // |
| 4627 | // The table below shows the aggregation of all supported cases |
| 4628 | // using these four cases. |
| 4629 | // |
| 4630 | // Reduction | == 0 | != 0 | == MAX | != MAX |
| 4631 | // ------------+-----------+-----------+-----------+----------- |
| 4632 | // or/umax | AllNonNeg | AnyNeg | AnyNeg | AllNonNeg |
| 4633 | // and/umin | AnyNonNeg | AllNeg | AllNeg | AnyNonNeg |
| 4634 | // add | AllNonNeg | AnyNeg | AllNeg | AnyNonNeg |
| 4635 | // |
| 4636 | // NOTE: MAX = 1 for or/and/umax/umin, and the vector size N for add |
| 4637 | // |
| 4638 | // For easier codegen and check inversion, we use the following encoding: |
| 4639 | // |
| 4640 | // 1. Bit-3 === requires or/umax (1) or and/umin (0) check |
| 4641 | // 2. Bit-2 === checks < 0 (1) or > -1 (0) |
| 4642 | // 3. Bit-1 === universal (1) or existential (0) check |
| 4643 | // |
| 4644 | // AnyNeg = 0b110: uses or/umax, checks negative, any-check |
| 4645 | // AllNonNeg = 0b101: uses or/umax, checks non-neg, all-check |
| 4646 | // AnyNonNeg = 0b000: uses and/umin, checks non-neg, any-check |
| 4647 | // AllNeg = 0b011: uses and/umin, checks negative, all-check |
| 4648 | // |
| 4649 | // XOR with 0b011 inverts the check (swaps all/any and neg/non-neg). |
| 4650 | // |
| 4651 | enum CheckKind : unsigned { |
| 4652 | AnyNonNeg = 0b000, |
| 4653 | AllNeg = 0b011, |
| 4654 | AllNonNeg = 0b101, |
| 4655 | AnyNeg = 0b110, |
| 4656 | }; |
| 4657 | // Return true if we fold this check into or/umax and false for and/umin |
| 4658 | auto RequiresOr = [](CheckKind C) -> bool { return C & 0b100; }; |
| 4659 | // Return true if we should check if result is negative and false otherwise |
| 4660 | auto IsNegativeCheck = [](CheckKind C) -> bool { return C & 0b010; }; |
| 4661 | // Logically invert the check |
| 4662 | auto Invert = [](CheckKind C) { return CheckKind(C ^ 0b011); }; |
| 4663 | |
| 4664 | CheckKind Base; |
| 4665 | switch (OrigIID) { |
| 4666 | case Intrinsic::vector_reduce_or: |
| 4667 | case Intrinsic::vector_reduce_umax: |
| 4668 | Base = TestsNegative ? AnyNeg : AllNonNeg; |
| 4669 | break; |
| 4670 | case Intrinsic::vector_reduce_and: |
| 4671 | case Intrinsic::vector_reduce_umin: |
| 4672 | Base = TestsNegative ? AllNeg : AnyNonNeg; |
| 4673 | break; |
| 4674 | case Intrinsic::vector_reduce_add: |
| 4675 | Base = TestsNegative ? AllNeg : AllNonNeg; |
| 4676 | break; |
| 4677 | default: |
| 4678 | llvm_unreachable("Unexpected intrinsic" ); |
| 4679 | } |
| 4680 | |
| 4681 | CheckKind Check = IsEq ? Base : Invert(Base); |
| 4682 | |
| 4683 | auto PickCheaper = [&](Intrinsic::ID Arith, Intrinsic::ID MinMax) { |
| 4684 | InstructionCost ArithCost = |
| 4685 | TTI.getArithmeticReductionCost(Opcode: getArithmeticReductionInstruction(RdxID: Arith), |
| 4686 | Ty: VecTy, FMF: std::nullopt, CostKind); |
| 4687 | InstructionCost MinMaxCost = |
| 4688 | TTI.getMinMaxReductionCost(IID: getMinMaxReductionIntrinsicOp(RdxID: MinMax), Ty: VecTy, |
| 4689 | FMF: FastMathFlags(), CostKind); |
| 4690 | return ArithCost <= MinMaxCost ? std::make_pair(x&: Arith, y&: ArithCost) |
| 4691 | : std::make_pair(x&: MinMax, y&: MinMaxCost); |
| 4692 | }; |
| 4693 | |
| 4694 | // Choose output reduction based on encoding's MSB |
| 4695 | auto [NewIID, NewCost] = RequiresOr(Check) |
| 4696 | ? PickCheaper(Intrinsic::vector_reduce_or, |
| 4697 | Intrinsic::vector_reduce_umax) |
| 4698 | : PickCheaper(Intrinsic::vector_reduce_and, |
| 4699 | Intrinsic::vector_reduce_umin); |
| 4700 | |
| 4701 | // Add cost of combining multiple sources with or/and |
| 4702 | if (NumSources > 1) { |
| 4703 | unsigned CombineOpc = |
| 4704 | RequiresOr(Check) ? Instruction::Or : Instruction::And; |
| 4705 | NewCost += TTI.getArithmeticInstrCost(Opcode: CombineOpc, Ty: VecTy, CostKind) * |
| 4706 | (NumSources - 1); |
| 4707 | } |
| 4708 | |
| 4709 | LLVM_DEBUG(dbgs() << "Found sign-bit reduction cmp: " << I << "\n OldCost: " |
| 4710 | << OldCost << " vs NewCost: " << NewCost << "\n" ); |
| 4711 | |
| 4712 | if (NewCost > OldCost) |
| 4713 | return false; |
| 4714 | |
| 4715 | // Generate the combined input and reduction |
| 4716 | Builder.SetInsertPoint(&I); |
| 4717 | Type *ScalarTy = VecTy->getScalarType(); |
| 4718 | |
| 4719 | Value *Input; |
| 4720 | if (NumSources == 1) { |
| 4721 | Input = Sources[0]; |
| 4722 | } else { |
| 4723 | // Combine sources with or/and based on check type |
| 4724 | Input = RequiresOr(Check) ? Builder.CreateOr(Ops: Sources) |
| 4725 | : Builder.CreateAnd(Ops: Sources); |
| 4726 | } |
| 4727 | |
| 4728 | Value *NewReduce = Builder.CreateIntrinsic(RetTy: ScalarTy, ID: NewIID, Args: {Input}); |
| 4729 | Value *NewCmp = IsNegativeCheck(Check) ? Builder.CreateIsNeg(Arg: NewReduce) |
| 4730 | : Builder.CreateIsNotNeg(Arg: NewReduce); |
| 4731 | replaceValue(Old&: I, New&: *NewCmp); |
| 4732 | return true; |
| 4733 | } |
| 4734 | |
| 4735 | /// Fold a zero test of reduce.or or reduce.umax into a boolean reduction. |
| 4736 | /// |
| 4737 | /// Vectorization may produce IR that compares the result of a scalar reduction |
| 4738 | /// with zero. Depending on the target, lowering a reduction and a scalar |
| 4739 | /// comparison separately can cost more than reducing lane-wise comparison |
| 4740 | /// results. This fold creates the latter form only when it is not costlier. |
| 4741 | /// |
| 4742 | /// Before: |
| 4743 | /// %r = call iT @llvm.vector.reduce.or.vNiT(<N x iT> %x) |
| 4744 | /// %cmp = icmp ne iT %r, 0 |
| 4745 | /// |
| 4746 | /// After: |
| 4747 | /// %lane.cmp = icmp ne <N x iT> %x, zeroinitializer |
| 4748 | /// %cmp = call i1 @llvm.vector.reduce.or.vNi1(<N x i1> %lane.cmp) |
| 4749 | /// |
| 4750 | /// `reduce.or` and `reduce.umax` are non-zero when at least one lane is |
| 4751 | /// non-zero. Therefore, `icmp ne` uses the existential `reduce.or` test. |
| 4752 | /// Conversely, `icmp eq` must check that every lane is zero, so it uses the |
| 4753 | /// universal `reduce.and` test. |
| 4754 | /// |
| 4755 | /// Before: |
| 4756 | /// %r = call iT @llvm.vector.reduce.umax.vNiT(<N x iT> %x) |
| 4757 | /// %cmp = icmp eq iT %r, 0 |
| 4758 | /// |
| 4759 | /// After: |
| 4760 | /// %lane.cmp = icmp eq <N x iT> %x, zeroinitializer |
| 4761 | /// %cmp = call i1 @llvm.vector.reduce.and.vNi1(<N x i1> %lane.cmp) |
| 4762 | bool VectorCombine::foldReductionZeroTest(Instruction &I) { |
| 4763 | CmpPredicate Pred; |
| 4764 | Value *Op; |
| 4765 | |
| 4766 | if (!match(V: &I, P: m_c_ICmp(Pred, L: m_Value(V&: Op), R: m_Zero())) || |
| 4767 | !ICmpInst::isEquality(P: Pred)) |
| 4768 | return false; |
| 4769 | |
| 4770 | auto *II = dyn_cast<IntrinsicInst>(Val: Op); |
| 4771 | if (!II || !II->hasOneUse()) |
| 4772 | return false; |
| 4773 | |
| 4774 | auto ReduceID = II->getIntrinsicID(); |
| 4775 | if (ReduceID != Intrinsic::vector_reduce_or && |
| 4776 | ReduceID != Intrinsic::vector_reduce_umax) |
| 4777 | return false; |
| 4778 | |
| 4779 | Value *Vec = II->getArgOperand(i: 0); |
| 4780 | auto *VecTy = dyn_cast<FixedVectorType>(Val: Vec->getType()); |
| 4781 | if (!VecTy || !VecTy->getElementType()->isIntegerTy()) |
| 4782 | return false; |
| 4783 | |
| 4784 | // Map the scalar zero test to an any-lane or all-lane boolean reduction. |
| 4785 | Intrinsic::ID NewIID = (Pred == ICmpInst::ICMP_NE) |
| 4786 | ? Intrinsic::vector_reduce_or |
| 4787 | : Intrinsic::vector_reduce_and; |
| 4788 | |
| 4789 | // This is not an unconditional canonicalization: compare the cost of the |
| 4790 | // original scalar reduction and compare with the vector compare and i1 |
| 4791 | // reduction replacement for both reduce.or and reduce.umax. |
| 4792 | InstructionCost OldCost = TTI.getInstructionCost(U: II, CostKind) + |
| 4793 | TTI.getInstructionCost(U: &I, CostKind); |
| 4794 | |
| 4795 | auto *CmpTy = cast<VectorType>(Val: CmpInst::makeCmpResultType(opnd_type: VecTy)); |
| 4796 | InstructionCost NewCost = |
| 4797 | TTI.getCmpSelInstrCost(Opcode: Instruction::ICmp, ValTy: VecTy, CondTy: CmpTy, VecPred: Pred, CostKind); |
| 4798 | NewCost += TTI.getArithmeticReductionCost( |
| 4799 | Opcode: getArithmeticReductionInstruction(RdxID: NewIID), Ty: CmpTy, FMF: std::nullopt, CostKind); |
| 4800 | |
| 4801 | LLVM_DEBUG(dbgs() << "Found a reduction zero test: " << I << "\n OldCost: " |
| 4802 | << OldCost << " vs NewCost: " << NewCost << "\n" ); |
| 4803 | |
| 4804 | if (!OldCost.isValid() || !NewCost.isValid() || NewCost > OldCost) |
| 4805 | return false; |
| 4806 | |
| 4807 | Builder.SetInsertPoint(&I); |
| 4808 | Value *NewCmp = Builder.CreateICmp(P: Pred, LHS: Vec, RHS: Constant::getNullValue(Ty: VecTy)); |
| 4809 | Value *NewReduce = Builder.CreateIntrinsic(ID: NewIID, OverloadTypes: {CmpTy}, Args: {NewCmp}); |
| 4810 | replaceValue(Old&: I, New&: *NewReduce); |
| 4811 | return true; |
| 4812 | } |
| 4813 | |
| 4814 | /// vector.reduce.OP f(X_i) == 0 -> vector.reduce.OP X_i == 0 |
| 4815 | /// |
| 4816 | /// We can prove it for cases when: |
| 4817 | /// |
| 4818 | /// 1. OP X_i == 0 <=> \forall i \in [1, N] X_i == 0 |
| 4819 | /// 1'. OP X_i == 0 <=> \exists j \in [1, N] X_j == 0 |
| 4820 | /// 2. f(x) == 0 <=> x == 0 |
| 4821 | /// |
| 4822 | /// From 1 and 2 (or 1' and 2), we can infer that |
| 4823 | /// |
| 4824 | /// OP f(X_i) == 0 <=> OP X_i == 0. |
| 4825 | /// |
| 4826 | /// (1) |
| 4827 | /// OP f(X_i) == 0 <=> \forall i \in [1, N] f(X_i) == 0 |
| 4828 | /// (2) |
| 4829 | /// <=> \forall i \in [1, N] X_i == 0 |
| 4830 | /// (1) |
| 4831 | /// <=> OP(X_i) == 0 |
| 4832 | /// |
| 4833 | /// For some of the OP's and f's, we need to have domain constraints on X |
| 4834 | /// to ensure properties 1 (or 1') and 2. |
| 4835 | bool VectorCombine::foldICmpEqZeroVectorReduce(Instruction &I) { |
| 4836 | CmpPredicate Pred; |
| 4837 | Value *Op; |
| 4838 | if (!match(V: &I, P: m_ICmp(Pred, L: m_Value(V&: Op), R: m_Zero())) || |
| 4839 | !ICmpInst::isEquality(P: Pred)) |
| 4840 | return false; |
| 4841 | |
| 4842 | auto *II = dyn_cast<IntrinsicInst>(Val: Op); |
| 4843 | if (!II) |
| 4844 | return false; |
| 4845 | |
| 4846 | switch (II->getIntrinsicID()) { |
| 4847 | case Intrinsic::vector_reduce_add: |
| 4848 | case Intrinsic::vector_reduce_or: |
| 4849 | case Intrinsic::vector_reduce_umin: |
| 4850 | case Intrinsic::vector_reduce_umax: |
| 4851 | case Intrinsic::vector_reduce_smin: |
| 4852 | case Intrinsic::vector_reduce_smax: |
| 4853 | break; |
| 4854 | default: |
| 4855 | return false; |
| 4856 | } |
| 4857 | |
| 4858 | Value *InnerOp = II->getArgOperand(i: 0); |
| 4859 | |
| 4860 | // TODO: fixed vector type might be too restrictive |
| 4861 | if (!II->hasOneUse() || !isa<FixedVectorType>(Val: InnerOp->getType())) |
| 4862 | return false; |
| 4863 | |
| 4864 | Value *X = nullptr; |
| 4865 | |
| 4866 | // Check for zero-preserving operations where f(x) = 0 <=> x = 0 |
| 4867 | // |
| 4868 | // 1. f(x) = shl nuw x, y for arbitrary y |
| 4869 | // 2. f(x) = mul nuw x, c for defined c != 0 |
| 4870 | // 3. f(x) = zext x |
| 4871 | // 4. f(x) = sext x |
| 4872 | // 5. f(x) = neg x |
| 4873 | // |
| 4874 | if (!(match(V: InnerOp, P: m_NUWShl(L: m_Value(V&: X), R: m_Value())) || // Case 1 |
| 4875 | match(V: InnerOp, P: m_NUWMul(L: m_Value(V&: X), R: m_NonZeroInt())) || // Case 2 |
| 4876 | match(V: InnerOp, P: m_ZExt(Op: m_Value(V&: X))) || // Case 3 |
| 4877 | match(V: InnerOp, P: m_SExt(Op: m_Value(V&: X))) || // Case 4 |
| 4878 | match(V: InnerOp, P: m_Neg(V: m_Value(V&: X))) // Case 5 |
| 4879 | )) |
| 4880 | return false; |
| 4881 | |
| 4882 | SimplifyQuery S = SQ.getWithInstruction(I: &I); |
| 4883 | auto *XTy = cast<FixedVectorType>(Val: X->getType()); |
| 4884 | |
| 4885 | // Check for domain constraints for all supported reductions. |
| 4886 | // |
| 4887 | // a. OR X_i - has property 1 for every X |
| 4888 | // b. UMAX X_i - has property 1 for every X |
| 4889 | // c. UMIN X_i - has property 1' for every X |
| 4890 | // d. SMAX X_i - has property 1 for X >= 0 |
| 4891 | // e. SMIN X_i - has property 1' for X >= 0 |
| 4892 | // f. ADD X_i - has property 1 for X >= 0 && ADD X_i doesn't sign wrap |
| 4893 | // |
| 4894 | // In order for the proof to work, we need 1 (or 1') to be true for both |
| 4895 | // OP f(X_i) and OP X_i and that's why below we check constraints twice. |
| 4896 | // |
| 4897 | // NOTE: ADD X_i holds property 1 for a mirror case as well, i.e. when |
| 4898 | // X <= 0 && ADD X_i doesn't sign wrap. However, due to the nature |
| 4899 | // of known bits, we can't reasonably hold knowledge of "either 0 |
| 4900 | // or negative". |
| 4901 | switch (II->getIntrinsicID()) { |
| 4902 | case Intrinsic::vector_reduce_add: { |
| 4903 | // We need to check that both X_i and f(X_i) have enough leading |
| 4904 | // zeros to not overflow. |
| 4905 | KnownBits KnownX = computeKnownBits(V: X, Q: S); |
| 4906 | KnownBits KnownFX = computeKnownBits(V: InnerOp, Q: S); |
| 4907 | unsigned NumElems = XTy->getNumElements(); |
| 4908 | // Adding N elements loses at most ceil(log2(N)) leading bits. |
| 4909 | unsigned LostBits = Log2_32_Ceil(Value: NumElems); |
| 4910 | unsigned LeadingZerosX = KnownX.countMinLeadingZeros(); |
| 4911 | unsigned LeadingZerosFX = KnownFX.countMinLeadingZeros(); |
| 4912 | // Need at least one leading zero left after summation to ensure no overflow |
| 4913 | if (LeadingZerosX <= LostBits || LeadingZerosFX <= LostBits) |
| 4914 | return false; |
| 4915 | |
| 4916 | // We are not checking whether X or f(X) are positive explicitly because |
| 4917 | // we implicitly checked for it when we checked if both cases have enough |
| 4918 | // leading zeros to not wrap addition. |
| 4919 | break; |
| 4920 | } |
| 4921 | case Intrinsic::vector_reduce_smin: |
| 4922 | case Intrinsic::vector_reduce_smax: |
| 4923 | // Check whether X >= 0 and f(X) >= 0 |
| 4924 | if (!isKnownNonNegative(V: InnerOp, SQ: S) || !isKnownNonNegative(V: X, SQ: S)) |
| 4925 | return false; |
| 4926 | |
| 4927 | break; |
| 4928 | default: |
| 4929 | break; |
| 4930 | }; |
| 4931 | |
| 4932 | LLVM_DEBUG(dbgs() << "Found a reduction to 0 comparison with removable op: " |
| 4933 | << *II << "\n" ); |
| 4934 | |
| 4935 | // For zext/sext, check if the transform is profitable using cost model. |
| 4936 | // For other operations (shl, mul, neg), we're removing an instruction |
| 4937 | // while keeping the same reduction type, so it's always profitable. |
| 4938 | if (isa<ZExtInst>(Val: InnerOp) || isa<SExtInst>(Val: InnerOp)) { |
| 4939 | auto *FXTy = cast<FixedVectorType>(Val: InnerOp->getType()); |
| 4940 | Intrinsic::ID IID = II->getIntrinsicID(); |
| 4941 | |
| 4942 | InstructionCost ExtCost = TTI.getCastInstrCost( |
| 4943 | Opcode: cast<CastInst>(Val: InnerOp)->getOpcode(), Dst: FXTy, Src: XTy, |
| 4944 | CCH: TTI::CastContextHint::None, CostKind, I: cast<CastInst>(Val: InnerOp)); |
| 4945 | |
| 4946 | InstructionCost OldReduceCost, NewReduceCost; |
| 4947 | switch (IID) { |
| 4948 | case Intrinsic::vector_reduce_add: |
| 4949 | case Intrinsic::vector_reduce_or: |
| 4950 | OldReduceCost = TTI.getArithmeticReductionCost( |
| 4951 | Opcode: getArithmeticReductionInstruction(RdxID: IID), Ty: FXTy, FMF: std::nullopt, CostKind); |
| 4952 | NewReduceCost = TTI.getArithmeticReductionCost( |
| 4953 | Opcode: getArithmeticReductionInstruction(RdxID: IID), Ty: XTy, FMF: std::nullopt, CostKind); |
| 4954 | break; |
| 4955 | case Intrinsic::vector_reduce_umin: |
| 4956 | case Intrinsic::vector_reduce_umax: |
| 4957 | case Intrinsic::vector_reduce_smin: |
| 4958 | case Intrinsic::vector_reduce_smax: |
| 4959 | OldReduceCost = TTI.getMinMaxReductionCost( |
| 4960 | IID: getMinMaxReductionIntrinsicOp(RdxID: IID), Ty: FXTy, FMF: FastMathFlags(), CostKind); |
| 4961 | NewReduceCost = TTI.getMinMaxReductionCost( |
| 4962 | IID: getMinMaxReductionIntrinsicOp(RdxID: IID), Ty: XTy, FMF: FastMathFlags(), CostKind); |
| 4963 | break; |
| 4964 | default: |
| 4965 | llvm_unreachable("Unexpected reduction" ); |
| 4966 | } |
| 4967 | |
| 4968 | InstructionCost OldCost = OldReduceCost + ExtCost; |
| 4969 | InstructionCost NewCost = |
| 4970 | NewReduceCost + (InnerOp->hasOneUse() ? 0 : ExtCost); |
| 4971 | |
| 4972 | LLVM_DEBUG(dbgs() << "Found a removable extension before reduction: " |
| 4973 | << *InnerOp << "\n OldCost: " << OldCost |
| 4974 | << " vs NewCost: " << NewCost << "\n" ); |
| 4975 | |
| 4976 | // We consider transformation to still be potentially beneficial even |
| 4977 | // when the costs are the same because we might remove a use from f(X) |
| 4978 | // and unlock other optimizations. Equal costs would just mean that we |
| 4979 | // didn't make it worse in the worst case. |
| 4980 | if (NewCost > OldCost) |
| 4981 | return false; |
| 4982 | } |
| 4983 | |
| 4984 | // Since we support zext and sext as f, we might change the scalar type |
| 4985 | // of the intrinsic. |
| 4986 | Type *Ty = XTy->getScalarType(); |
| 4987 | Value *NewReduce = Builder.CreateIntrinsic(RetTy: Ty, ID: II->getIntrinsicID(), Args: {X}); |
| 4988 | Value *NewCmp = |
| 4989 | Builder.CreateICmp(P: Pred, LHS: NewReduce, RHS: ConstantInt::getNullValue(Ty)); |
| 4990 | replaceValue(Old&: I, New&: *NewCmp); |
| 4991 | return true; |
| 4992 | } |
| 4993 | |
| 4994 | /// Fold comparisons of reduce.or/reduce.and with reduce.umax/reduce.umin |
| 4995 | /// based on cost, preserving the comparison semantics. |
| 4996 | /// |
| 4997 | /// We use two fundamental properties for each pair: |
| 4998 | /// |
| 4999 | /// 1. or(X) == 0 <=> umax(X) == 0 |
| 5000 | /// 2. or(X) == 1 <=> umax(X) == 1 |
| 5001 | /// 3. sign(or(X)) == sign(umax(X)) |
| 5002 | /// |
| 5003 | /// 1. and(X) == -1 <=> umin(X) == -1 |
| 5004 | /// 2. and(X) == -2 <=> umin(X) == -2 |
| 5005 | /// 3. sign(and(X)) == sign(umin(X)) |
| 5006 | /// |
| 5007 | /// From these we can infer the following transformations: |
| 5008 | /// a. or(X) ==/!= 0 <-> umax(X) ==/!= 0 |
| 5009 | /// b. or(X) s< 0 <-> umax(X) s< 0 |
| 5010 | /// c. or(X) s> -1 <-> umax(X) s> -1 |
| 5011 | /// d. or(X) s< 1 <-> umax(X) s< 1 |
| 5012 | /// e. or(X) ==/!= 1 <-> umax(X) ==/!= 1 |
| 5013 | /// f. or(X) s< 2 <-> umax(X) s< 2 |
| 5014 | /// g. and(X) ==/!= -1 <-> umin(X) ==/!= -1 |
| 5015 | /// h. and(X) s< 0 <-> umin(X) s< 0 |
| 5016 | /// i. and(X) s> -1 <-> umin(X) s> -1 |
| 5017 | /// j. and(X) s> -2 <-> umin(X) s> -2 |
| 5018 | /// k. and(X) ==/!= -2 <-> umin(X) ==/!= -2 |
| 5019 | /// l. and(X) s> -3 <-> umin(X) s> -3 |
| 5020 | /// |
| 5021 | bool VectorCombine::foldEquivalentReductionCmp(Instruction &I) { |
| 5022 | CmpPredicate Pred; |
| 5023 | Value *ReduceOp; |
| 5024 | const APInt *CmpVal; |
| 5025 | if (!match(V: &I, P: m_ICmp(Pred, L: m_Value(V&: ReduceOp), R: m_APInt(Res&: CmpVal)))) |
| 5026 | return false; |
| 5027 | |
| 5028 | auto *II = dyn_cast<IntrinsicInst>(Val: ReduceOp); |
| 5029 | if (!II || !II->hasOneUse()) |
| 5030 | return false; |
| 5031 | |
| 5032 | const auto IsValidOrUmaxCmp = [&]() { |
| 5033 | // or === umax for i1 |
| 5034 | if (CmpVal->getBitWidth() == 1) |
| 5035 | return true; |
| 5036 | |
| 5037 | // Cases a and e |
| 5038 | bool IsEquality = |
| 5039 | (CmpVal->isZero() || CmpVal->isOne()) && ICmpInst::isEquality(P: Pred); |
| 5040 | // Case c |
| 5041 | bool IsPositive = CmpVal->isAllOnes() && Pred == ICmpInst::ICMP_SGT; |
| 5042 | // Cases b, d, and f |
| 5043 | bool IsNegative = (CmpVal->isZero() || CmpVal->isOne() || *CmpVal == 2) && |
| 5044 | Pred == ICmpInst::ICMP_SLT; |
| 5045 | return IsEquality || IsPositive || IsNegative; |
| 5046 | }; |
| 5047 | |
| 5048 | const auto IsValidAndUminCmp = [&]() { |
| 5049 | // and === umin for i1 |
| 5050 | if (CmpVal->getBitWidth() == 1) |
| 5051 | return true; |
| 5052 | |
| 5053 | const auto LeadingOnes = CmpVal->countl_one(); |
| 5054 | |
| 5055 | // Cases g and k |
| 5056 | bool IsEquality = |
| 5057 | (CmpVal->isAllOnes() || LeadingOnes + 1 == CmpVal->getBitWidth()) && |
| 5058 | ICmpInst::isEquality(P: Pred); |
| 5059 | // Case h |
| 5060 | bool IsNegative = CmpVal->isZero() && Pred == ICmpInst::ICMP_SLT; |
| 5061 | // Cases i, j, and l |
| 5062 | bool IsPositive = |
| 5063 | // if the number has at least N - 2 leading ones |
| 5064 | // and the two LSBs are: |
| 5065 | // - 1 x 1 -> -1 |
| 5066 | // - 1 x 0 -> -2 |
| 5067 | // - 0 x 1 -> -3 |
| 5068 | LeadingOnes + 2 >= CmpVal->getBitWidth() && |
| 5069 | ((*CmpVal)[0] || (*CmpVal)[1]) && Pred == ICmpInst::ICMP_SGT; |
| 5070 | return IsEquality || IsNegative || IsPositive; |
| 5071 | }; |
| 5072 | |
| 5073 | Intrinsic::ID OriginalIID = II->getIntrinsicID(); |
| 5074 | Intrinsic::ID AlternativeIID; |
| 5075 | |
| 5076 | // Check if this is a valid comparison pattern and determine the alternate |
| 5077 | // reduction intrinsic. |
| 5078 | switch (OriginalIID) { |
| 5079 | case Intrinsic::vector_reduce_or: |
| 5080 | if (!IsValidOrUmaxCmp()) |
| 5081 | return false; |
| 5082 | AlternativeIID = Intrinsic::vector_reduce_umax; |
| 5083 | break; |
| 5084 | case Intrinsic::vector_reduce_umax: |
| 5085 | if (!IsValidOrUmaxCmp()) |
| 5086 | return false; |
| 5087 | AlternativeIID = Intrinsic::vector_reduce_or; |
| 5088 | break; |
| 5089 | case Intrinsic::vector_reduce_and: |
| 5090 | if (!IsValidAndUminCmp()) |
| 5091 | return false; |
| 5092 | AlternativeIID = Intrinsic::vector_reduce_umin; |
| 5093 | break; |
| 5094 | case Intrinsic::vector_reduce_umin: |
| 5095 | if (!IsValidAndUminCmp()) |
| 5096 | return false; |
| 5097 | AlternativeIID = Intrinsic::vector_reduce_and; |
| 5098 | break; |
| 5099 | default: |
| 5100 | return false; |
| 5101 | } |
| 5102 | |
| 5103 | Value *X = II->getArgOperand(i: 0); |
| 5104 | auto *VecTy = dyn_cast<FixedVectorType>(Val: X->getType()); |
| 5105 | if (!VecTy) |
| 5106 | return false; |
| 5107 | |
| 5108 | const auto GetReductionCost = [&](Intrinsic::ID IID) -> InstructionCost { |
| 5109 | unsigned ReductionOpc = getArithmeticReductionInstruction(RdxID: IID); |
| 5110 | if (ReductionOpc != Instruction::ICmp) |
| 5111 | return TTI.getArithmeticReductionCost(Opcode: ReductionOpc, Ty: VecTy, FMF: std::nullopt, |
| 5112 | CostKind); |
| 5113 | return TTI.getMinMaxReductionCost(IID: getMinMaxReductionIntrinsicOp(RdxID: IID), Ty: VecTy, |
| 5114 | FMF: FastMathFlags(), CostKind); |
| 5115 | }; |
| 5116 | |
| 5117 | InstructionCost OrigCost = GetReductionCost(OriginalIID); |
| 5118 | InstructionCost AltCost = GetReductionCost(AlternativeIID); |
| 5119 | |
| 5120 | LLVM_DEBUG(dbgs() << "Found equivalent reduction cmp: " << I |
| 5121 | << "\n OrigCost: " << OrigCost |
| 5122 | << " vs AltCost: " << AltCost << "\n" ); |
| 5123 | |
| 5124 | if (AltCost >= OrigCost) |
| 5125 | return false; |
| 5126 | |
| 5127 | Builder.SetInsertPoint(&I); |
| 5128 | Type *ScalarTy = VecTy->getScalarType(); |
| 5129 | Value *NewReduce = Builder.CreateIntrinsic(RetTy: ScalarTy, ID: AlternativeIID, Args: {X}); |
| 5130 | Value *NewCmp = |
| 5131 | Builder.CreateICmp(P: Pred, LHS: NewReduce, RHS: ConstantInt::get(Ty: ScalarTy, V: *CmpVal)); |
| 5132 | |
| 5133 | replaceValue(Old&: I, New&: *NewCmp); |
| 5134 | return true; |
| 5135 | } |
| 5136 | |
| 5137 | /// Used by foldReduceAddCmpZero to check if we can prove that a value is |
| 5138 | /// non-positive. |
| 5139 | /// KnownBits cannot see sext <? x i1> as non-positive: each top bit equals a |
| 5140 | /// single unknown input bit, which a per-bit lattice cannot track. The fold's |
| 5141 | /// target shape is popcount-style sums of <N x i1> valid/invalid masks (e.g. |
| 5142 | /// ray-intersection hits) tested for any-hit. |
| 5143 | /// Previous attempts to approximate the known bits of such expressions were |
| 5144 | /// using a fully recursive value tracking approach to infer a constant range |
| 5145 | /// but ultimately turned to be too expensive in compile time. |
| 5146 | static bool isKnownNonPositive(const Value *V, const SimplifyQuery &SQ, |
| 5147 | unsigned Depth = 0) { |
| 5148 | constexpr unsigned MaxLocalDepth = 2; |
| 5149 | if (Depth > MaxLocalDepth) |
| 5150 | return false; |
| 5151 | |
| 5152 | auto NumSignBits = [&](const Value *X) { |
| 5153 | return ComputeNumSignBits(Op: X, DL: SQ.DL, AC: SQ.AC, CxtI: SQ.CxtI, DT: SQ.DT); |
| 5154 | }; |
| 5155 | if (NumSignBits(V) == V->getType()->getScalarSizeInBits()) |
| 5156 | return true; |
| 5157 | |
| 5158 | Value *A, *B; |
| 5159 | if (match(V, P: m_Add(L: m_Value(V&: A), R: m_Value(V&: B)))) |
| 5160 | return NumSignBits(A) >= 2 && NumSignBits(B) >= 2 && |
| 5161 | isKnownNonPositive(V: A, SQ, Depth: Depth + 1) && |
| 5162 | isKnownNonPositive(V: B, SQ, Depth: Depth + 1); |
| 5163 | |
| 5164 | return computeKnownBits(V, Q: SQ).isNonPositive(); |
| 5165 | } |
| 5166 | |
| 5167 | /// Fold (icmp pred (reduce.add X), 0) to (icmp pred' (reduce.or X), 0) when X |
| 5168 | /// has lanes known to all be non-negative or all non-positive, so that |
| 5169 | /// sum == 0 iff every lane is 0. Falls back to reduce.umax if reduce.or is |
| 5170 | /// more expensive on the target. |
| 5171 | bool VectorCombine::foldReduceAddCmpZero(Instruction &I) { |
| 5172 | CmpPredicate Pred; |
| 5173 | Value *Vec; |
| 5174 | if (!match(V: &I, P: m_ICmp(Pred, |
| 5175 | L: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::vector_reduce_add>( |
| 5176 | Ops: m_Value(V&: Vec))), |
| 5177 | R: m_Zero()))) |
| 5178 | return false; |
| 5179 | |
| 5180 | auto *VecTy = dyn_cast<FixedVectorType>(Val: Vec->getType()); |
| 5181 | if (!VecTy || VecTy->getNumElements() < 2) |
| 5182 | return false; |
| 5183 | |
| 5184 | SimplifyQuery Q = SQ.getWithInstruction(I: &I); |
| 5185 | bool IsNonNegative = isKnownNonNegative(V: Vec, SQ: Q); |
| 5186 | bool IsNonPositive = !IsNonNegative && isKnownNonPositive(V: Vec, SQ: Q); |
| 5187 | if (!IsNonNegative && !IsNonPositive) |
| 5188 | return false; |
| 5189 | |
| 5190 | // Summing NumElts lanes can consume up to log2(NumElts) sign bits. Require |
| 5191 | // strictly more headroom than that so the sum cannot wrap to zero. |
| 5192 | unsigned NumElts = VecTy->getNumElements(); |
| 5193 | unsigned NumSignBits = ComputeNumSignBits(Op: Vec, DL: *DL, AC: SQ.AC, CxtI: &I, DT: &DT); |
| 5194 | if (Log2_32(Value: NumElts) >= NumSignBits) |
| 5195 | return false; |
| 5196 | |
| 5197 | ICmpInst::Predicate NewPred; |
| 5198 | switch (Pred) { |
| 5199 | case ICmpInst::ICMP_EQ: |
| 5200 | case ICmpInst::ICMP_ULE: |
| 5201 | case ICmpInst::ICMP_SLE: |
| 5202 | case ICmpInst::ICMP_SGE: |
| 5203 | NewPred = ICmpInst::ICMP_EQ; |
| 5204 | break; |
| 5205 | case ICmpInst::ICMP_NE: |
| 5206 | case ICmpInst::ICMP_UGT: |
| 5207 | case ICmpInst::ICMP_SGT: |
| 5208 | case ICmpInst::ICMP_SLT: |
| 5209 | NewPred = ICmpInst::ICMP_NE; |
| 5210 | break; |
| 5211 | default: |
| 5212 | return false; |
| 5213 | } |
| 5214 | |
| 5215 | // SGT and SLE on a non-positive tree, and SLT and SGE on a non-negative |
| 5216 | // tree, are tautologies (always true or always false). Leave those to |
| 5217 | // InstCombine rather than mapping them here. Remaining signed inequalities |
| 5218 | // also need one extra sign bit so the sum cannot flip sign. |
| 5219 | if (!IsNonNegative && |
| 5220 | (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE)) |
| 5221 | return false; |
| 5222 | if (!IsNonPositive && |
| 5223 | (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE)) |
| 5224 | return false; |
| 5225 | if ((Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SLE || |
| 5226 | Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SGE) && |
| 5227 | Log2_32(Value: NumElts) >= NumSignBits - 1) |
| 5228 | return false; |
| 5229 | |
| 5230 | InstructionCost OrigCost = TTI.getArithmeticReductionCost( |
| 5231 | Opcode: Instruction::Add, Ty: VecTy, FMF: std::nullopt, CostKind); |
| 5232 | InstructionCost OrCost = TTI.getArithmeticReductionCost( |
| 5233 | Opcode: Instruction::Or, Ty: VecTy, FMF: std::nullopt, CostKind); |
| 5234 | InstructionCost UmaxCost = TTI.getMinMaxReductionCost( |
| 5235 | IID: Intrinsic::umax, Ty: VecTy, FMF: FastMathFlags(), CostKind); |
| 5236 | if (!OrCost.isValid() && !UmaxCost.isValid()) |
| 5237 | return false; |
| 5238 | bool UseOr = OrCost.isValid() && (!UmaxCost.isValid() || OrCost <= UmaxCost); |
| 5239 | InstructionCost AltCost = UseOr ? OrCost : UmaxCost; |
| 5240 | if (AltCost > OrigCost) |
| 5241 | return false; |
| 5242 | |
| 5243 | Builder.SetInsertPoint(&I); |
| 5244 | Value *NewReduce = UseOr ? Builder.CreateOrReduce(Src: Vec) |
| 5245 | : Builder.CreateIntrinsic( |
| 5246 | ID: Intrinsic::vector_reduce_umax, OverloadTypes: {VecTy}, Args: {Vec}); |
| 5247 | Worklist.pushValue(V: NewReduce); |
| 5248 | Value *NewCmp = Builder.CreateICmp( |
| 5249 | P: NewPred, LHS: NewReduce, RHS: ConstantInt::getNullValue(Ty: VecTy->getScalarType())); |
| 5250 | replaceValue(Old&: I, New&: *NewCmp); |
| 5251 | return true; |
| 5252 | } |
| 5253 | |
| 5254 | /// Returns true if this ShuffleVectorInst eventually feeds into a |
| 5255 | /// vector reduction intrinsic (e.g., vector_reduce_add) by only following |
| 5256 | /// chains of shuffles and binary operators (in any combination/order). |
| 5257 | /// The search does not go deeper than the given Depth. |
| 5258 | static bool feedsIntoVectorReduction(ShuffleVectorInst *SVI) { |
| 5259 | constexpr unsigned MaxVisited = 32; |
| 5260 | SmallPtrSet<Instruction *, 8> Visited; |
| 5261 | SmallVector<Instruction *, 4> WorkList; |
| 5262 | bool FoundReduction = false; |
| 5263 | |
| 5264 | WorkList.push_back(Elt: SVI); |
| 5265 | while (!WorkList.empty()) { |
| 5266 | Instruction *I = WorkList.pop_back_val(); |
| 5267 | for (User *U : I->users()) { |
| 5268 | auto *UI = cast<Instruction>(Val: U); |
| 5269 | if (!UI || !Visited.insert(Ptr: UI).second) |
| 5270 | continue; |
| 5271 | if (Visited.size() > MaxVisited) |
| 5272 | return false; |
| 5273 | if (auto *II = dyn_cast<IntrinsicInst>(Val: UI)) { |
| 5274 | // More than one reduction reached |
| 5275 | if (FoundReduction) |
| 5276 | return false; |
| 5277 | switch (II->getIntrinsicID()) { |
| 5278 | case Intrinsic::vector_reduce_add: |
| 5279 | case Intrinsic::vector_reduce_mul: |
| 5280 | case Intrinsic::vector_reduce_and: |
| 5281 | case Intrinsic::vector_reduce_or: |
| 5282 | case Intrinsic::vector_reduce_xor: |
| 5283 | case Intrinsic::vector_reduce_smin: |
| 5284 | case Intrinsic::vector_reduce_smax: |
| 5285 | case Intrinsic::vector_reduce_umin: |
| 5286 | case Intrinsic::vector_reduce_umax: |
| 5287 | FoundReduction = true; |
| 5288 | continue; |
| 5289 | default: |
| 5290 | return false; |
| 5291 | } |
| 5292 | } |
| 5293 | |
| 5294 | if (!isa<BinaryOperator>(Val: UI) && !isa<ShuffleVectorInst>(Val: UI)) |
| 5295 | return false; |
| 5296 | |
| 5297 | WorkList.emplace_back(Args&: UI); |
| 5298 | } |
| 5299 | } |
| 5300 | return FoundReduction; |
| 5301 | } |
| 5302 | |
| 5303 | /// This method looks for groups of shuffles acting on binops, of the form: |
| 5304 | /// %x = shuffle ... |
| 5305 | /// %y = shuffle ... |
| 5306 | /// %a = binop %x, %y |
| 5307 | /// %b = binop %x, %y |
| 5308 | /// shuffle %a, %b, selectmask |
| 5309 | /// We may, especially if the shuffle is wider than legal, be able to convert |
| 5310 | /// the shuffle to a form where only parts of a and b need to be computed. On |
| 5311 | /// architectures with no obvious "select" shuffle, this can reduce the total |
| 5312 | /// number of operations if the target reports them as cheaper. |
| 5313 | bool VectorCombine::foldSelectShuffle(Instruction &I, bool FromReduction) { |
| 5314 | auto *SVI = cast<ShuffleVectorInst>(Val: &I); |
| 5315 | auto *VT = cast<FixedVectorType>(Val: I.getType()); |
| 5316 | auto *Op0 = dyn_cast<Instruction>(Val: SVI->getOperand(i_nocapture: 0)); |
| 5317 | auto *Op1 = dyn_cast<Instruction>(Val: SVI->getOperand(i_nocapture: 1)); |
| 5318 | if (!Op0 || !Op1 || Op0 == Op1 || !Op0->isBinaryOp() || !Op1->isBinaryOp() || |
| 5319 | VT != Op0->getType()) |
| 5320 | return false; |
| 5321 | |
| 5322 | auto *SVI0A = dyn_cast<Instruction>(Val: Op0->getOperand(i: 0)); |
| 5323 | auto *SVI0B = dyn_cast<Instruction>(Val: Op0->getOperand(i: 1)); |
| 5324 | auto *SVI1A = dyn_cast<Instruction>(Val: Op1->getOperand(i: 0)); |
| 5325 | auto *SVI1B = dyn_cast<Instruction>(Val: Op1->getOperand(i: 1)); |
| 5326 | SmallPtrSet<Instruction *, 4> InputShuffles({SVI0A, SVI0B, SVI1A, SVI1B}); |
| 5327 | auto checkSVNonOpUses = [&](Instruction *I) { |
| 5328 | if (!I || I->getOperand(i: 0)->getType() != VT) |
| 5329 | return true; |
| 5330 | return any_of(Range: I->users(), P: [&](User *U) { |
| 5331 | return U != Op0 && U != Op1 && |
| 5332 | !(isa<ShuffleVectorInst>(Val: U) && |
| 5333 | (InputShuffles.contains(Ptr: cast<Instruction>(Val: U)) || |
| 5334 | isInstructionTriviallyDead(I: cast<Instruction>(Val: U)))); |
| 5335 | }); |
| 5336 | }; |
| 5337 | if (checkSVNonOpUses(SVI0A) || checkSVNonOpUses(SVI0B) || |
| 5338 | checkSVNonOpUses(SVI1A) || checkSVNonOpUses(SVI1B)) |
| 5339 | return false; |
| 5340 | |
| 5341 | // Collect all the uses that are shuffles that we can transform together. We |
| 5342 | // may not have a single shuffle, but a group that can all be transformed |
| 5343 | // together profitably. |
| 5344 | SmallVector<ShuffleVectorInst *> Shuffles; |
| 5345 | auto collectShuffles = [&](Instruction *I) { |
| 5346 | for (auto *U : I->users()) { |
| 5347 | auto *SV = dyn_cast<ShuffleVectorInst>(Val: U); |
| 5348 | if (!SV || SV->getType() != VT) |
| 5349 | return false; |
| 5350 | if ((SV->getOperand(i_nocapture: 0) != Op0 && SV->getOperand(i_nocapture: 0) != Op1) || |
| 5351 | (SV->getOperand(i_nocapture: 1) != Op0 && SV->getOperand(i_nocapture: 1) != Op1)) |
| 5352 | return false; |
| 5353 | if (!llvm::is_contained(Range&: Shuffles, Element: SV)) |
| 5354 | Shuffles.push_back(Elt: SV); |
| 5355 | } |
| 5356 | return true; |
| 5357 | }; |
| 5358 | if (!collectShuffles(Op0) || !collectShuffles(Op1)) |
| 5359 | return false; |
| 5360 | // From a reduction, we need to be processing a single shuffle, otherwise the |
| 5361 | // other uses will not be lane-invariant. |
| 5362 | if (FromReduction && Shuffles.size() > 1) |
| 5363 | return false; |
| 5364 | |
| 5365 | // Add any shuffle uses for the shuffles we have found, to include them in our |
| 5366 | // cost calculations. |
| 5367 | if (!FromReduction) { |
| 5368 | for (size_t Idx = 0, E = Shuffles.size(); Idx != E; ++Idx) { |
| 5369 | for (auto *U : Shuffles[Idx]->users()) { |
| 5370 | ShuffleVectorInst *SSV = dyn_cast<ShuffleVectorInst>(Val: U); |
| 5371 | if (SSV && isa<UndefValue>(Val: SSV->getOperand(i_nocapture: 1)) && SSV->getType() == VT) |
| 5372 | Shuffles.push_back(Elt: SSV); |
| 5373 | } |
| 5374 | } |
| 5375 | } |
| 5376 | |
| 5377 | // For each of the output shuffles, we try to sort all the first vector |
| 5378 | // elements to the beginning, followed by the second array elements at the |
| 5379 | // end. If the binops are legalized to smaller vectors, this may reduce total |
| 5380 | // number of binops. We compute the ReconstructMask mask needed to convert |
| 5381 | // back to the original lane order. |
| 5382 | SmallVector<std::pair<int, int>> V1, V2; |
| 5383 | SmallVector<SmallVector<int>> OrigReconstructMasks; |
| 5384 | int MaxV1Elt = 0, MaxV2Elt = 0; |
| 5385 | unsigned NumElts = VT->getNumElements(); |
| 5386 | for (ShuffleVectorInst *SVN : Shuffles) { |
| 5387 | SmallVector<int> Mask; |
| 5388 | SVN->getShuffleMask(Result&: Mask); |
| 5389 | |
| 5390 | // Check the operands are the same as the original, or reversed (in which |
| 5391 | // case we need to commute the mask). |
| 5392 | Value *SVOp0 = SVN->getOperand(i_nocapture: 0); |
| 5393 | Value *SVOp1 = SVN->getOperand(i_nocapture: 1); |
| 5394 | if (isa<UndefValue>(Val: SVOp1)) { |
| 5395 | auto *SSV = cast<ShuffleVectorInst>(Val: SVOp0); |
| 5396 | SVOp0 = SSV->getOperand(i_nocapture: 0); |
| 5397 | SVOp1 = SSV->getOperand(i_nocapture: 1); |
| 5398 | for (int &Elem : Mask) { |
| 5399 | if (Elem >= static_cast<int>(SSV->getShuffleMask().size())) |
| 5400 | return false; |
| 5401 | Elem = Elem < 0 ? Elem : SSV->getMaskValue(Elt: Elem); |
| 5402 | } |
| 5403 | } |
| 5404 | if (SVOp0 == Op1 && SVOp1 == Op0) { |
| 5405 | std::swap(a&: SVOp0, b&: SVOp1); |
| 5406 | ShuffleVectorInst::commuteShuffleMask(Mask, InVecNumElts: NumElts); |
| 5407 | } |
| 5408 | if (SVOp0 != Op0 || SVOp1 != Op1) |
| 5409 | return false; |
| 5410 | |
| 5411 | // Calculate the reconstruction mask for this shuffle, as the mask needed to |
| 5412 | // take the packed values from Op0/Op1 and reconstructing to the original |
| 5413 | // order. |
| 5414 | SmallVector<int> ReconstructMask; |
| 5415 | for (unsigned I = 0; I < Mask.size(); I++) { |
| 5416 | if (Mask[I] < 0) { |
| 5417 | ReconstructMask.push_back(Elt: -1); |
| 5418 | } else if (Mask[I] < static_cast<int>(NumElts)) { |
| 5419 | MaxV1Elt = std::max(a: MaxV1Elt, b: Mask[I]); |
| 5420 | auto It = find_if(Range&: V1, P: [&](const std::pair<int, int> &A) { |
| 5421 | return Mask[I] == A.first; |
| 5422 | }); |
| 5423 | if (It != V1.end()) |
| 5424 | ReconstructMask.push_back(Elt: It - V1.begin()); |
| 5425 | else { |
| 5426 | ReconstructMask.push_back(Elt: V1.size()); |
| 5427 | V1.emplace_back(Args&: Mask[I], Args: V1.size()); |
| 5428 | } |
| 5429 | } else { |
| 5430 | MaxV2Elt = std::max<int>(a: MaxV2Elt, b: Mask[I] - NumElts); |
| 5431 | auto It = find_if(Range&: V2, P: [&](const std::pair<int, int> &A) { |
| 5432 | return Mask[I] - static_cast<int>(NumElts) == A.first; |
| 5433 | }); |
| 5434 | if (It != V2.end()) |
| 5435 | ReconstructMask.push_back(Elt: NumElts + It - V2.begin()); |
| 5436 | else { |
| 5437 | ReconstructMask.push_back(Elt: NumElts + V2.size()); |
| 5438 | V2.emplace_back(Args: Mask[I] - NumElts, Args: NumElts + V2.size()); |
| 5439 | } |
| 5440 | } |
| 5441 | } |
| 5442 | |
| 5443 | // For reductions, we know that the lane ordering out doesn't alter the |
| 5444 | // result. In-order can help simplify the shuffle away. |
| 5445 | if (FromReduction) |
| 5446 | sort(C&: ReconstructMask); |
| 5447 | OrigReconstructMasks.push_back(Elt: std::move(ReconstructMask)); |
| 5448 | } |
| 5449 | |
| 5450 | // If the Maximum element used from V1 and V2 are not larger than the new |
| 5451 | // vectors, the vectors are already packes and performing the optimization |
| 5452 | // again will likely not help any further. This also prevents us from getting |
| 5453 | // stuck in a cycle in case the costs do not also rule it out. |
| 5454 | if (V1.empty() || V2.empty() || |
| 5455 | (MaxV1Elt == static_cast<int>(V1.size()) - 1 && |
| 5456 | MaxV2Elt == static_cast<int>(V2.size()) - 1)) |
| 5457 | return false; |
| 5458 | |
| 5459 | // GetBaseMaskValue takes one of the inputs, which may either be a shuffle, a |
| 5460 | // shuffle of another shuffle, or not a shuffle (that is treated like a |
| 5461 | // identity shuffle). |
| 5462 | auto GetBaseMaskValue = [&](Instruction *I, int M) { |
| 5463 | auto *SV = dyn_cast<ShuffleVectorInst>(Val: I); |
| 5464 | if (!SV) |
| 5465 | return M; |
| 5466 | if (isa<UndefValue>(Val: SV->getOperand(i_nocapture: 1))) |
| 5467 | if (auto *SSV = dyn_cast<ShuffleVectorInst>(Val: SV->getOperand(i_nocapture: 0))) |
| 5468 | if (InputShuffles.contains(Ptr: SSV)) |
| 5469 | return SSV->getMaskValue(Elt: SV->getMaskValue(Elt: M)); |
| 5470 | return SV->getMaskValue(Elt: M); |
| 5471 | }; |
| 5472 | |
| 5473 | // Attempt to sort the inputs my ascending mask values to make simpler input |
| 5474 | // shuffles and push complex shuffles down to the uses. We sort on the first |
| 5475 | // of the two input shuffle orders, to try and get at least one input into a |
| 5476 | // nice order. |
| 5477 | auto SortBase = [&](Instruction *A, std::pair<int, int> X, |
| 5478 | std::pair<int, int> Y) { |
| 5479 | int MXA = GetBaseMaskValue(A, X.first); |
| 5480 | int MYA = GetBaseMaskValue(A, Y.first); |
| 5481 | return MXA < MYA; |
| 5482 | }; |
| 5483 | stable_sort(Range&: V1, C: [&](std::pair<int, int> A, std::pair<int, int> B) { |
| 5484 | return SortBase(SVI0A, A, B); |
| 5485 | }); |
| 5486 | stable_sort(Range&: V2, C: [&](std::pair<int, int> A, std::pair<int, int> B) { |
| 5487 | return SortBase(SVI1A, A, B); |
| 5488 | }); |
| 5489 | // Calculate our ReconstructMasks from the OrigReconstructMasks and the |
| 5490 | // modified order of the input shuffles. |
| 5491 | SmallVector<SmallVector<int>> ReconstructMasks; |
| 5492 | for (const auto &Mask : OrigReconstructMasks) { |
| 5493 | SmallVector<int> ReconstructMask; |
| 5494 | for (int M : Mask) { |
| 5495 | auto FindIndex = [](const SmallVector<std::pair<int, int>> &V, int M) { |
| 5496 | auto It = find_if(Range: V, P: [M](auto A) { return A.second == M; }); |
| 5497 | assert(It != V.end() && "Expected all entries in Mask" ); |
| 5498 | return std::distance(first: V.begin(), last: It); |
| 5499 | }; |
| 5500 | if (M < 0) |
| 5501 | ReconstructMask.push_back(Elt: -1); |
| 5502 | else if (M < static_cast<int>(NumElts)) { |
| 5503 | ReconstructMask.push_back(Elt: FindIndex(V1, M)); |
| 5504 | } else { |
| 5505 | ReconstructMask.push_back(Elt: NumElts + FindIndex(V2, M)); |
| 5506 | } |
| 5507 | } |
| 5508 | ReconstructMasks.push_back(Elt: std::move(ReconstructMask)); |
| 5509 | } |
| 5510 | |
| 5511 | // Calculate the masks needed for the new input shuffles, which get padded |
| 5512 | // with undef |
| 5513 | SmallVector<int> V1A, V1B, V2A, V2B; |
| 5514 | for (unsigned I = 0; I < V1.size(); I++) { |
| 5515 | V1A.push_back(Elt: GetBaseMaskValue(SVI0A, V1[I].first)); |
| 5516 | V1B.push_back(Elt: GetBaseMaskValue(SVI0B, V1[I].first)); |
| 5517 | } |
| 5518 | for (unsigned I = 0; I < V2.size(); I++) { |
| 5519 | V2A.push_back(Elt: GetBaseMaskValue(SVI1A, V2[I].first)); |
| 5520 | V2B.push_back(Elt: GetBaseMaskValue(SVI1B, V2[I].first)); |
| 5521 | } |
| 5522 | while (V1A.size() < NumElts) { |
| 5523 | V1A.push_back(Elt: PoisonMaskElem); |
| 5524 | V1B.push_back(Elt: PoisonMaskElem); |
| 5525 | } |
| 5526 | while (V2A.size() < NumElts) { |
| 5527 | V2A.push_back(Elt: PoisonMaskElem); |
| 5528 | V2B.push_back(Elt: PoisonMaskElem); |
| 5529 | } |
| 5530 | |
| 5531 | auto AddShuffleCost = [&](InstructionCost C, Instruction *I) { |
| 5532 | auto *SV = dyn_cast<ShuffleVectorInst>(Val: I); |
| 5533 | if (!SV) |
| 5534 | return C; |
| 5535 | return C + TTI.getShuffleCost(Kind: isa<UndefValue>(Val: SV->getOperand(i_nocapture: 1)) |
| 5536 | ? TTI::SK_PermuteSingleSrc |
| 5537 | : TTI::SK_PermuteTwoSrc, |
| 5538 | DstTy: VT, SrcTy: VT, Mask: SV->getShuffleMask(), CostKind); |
| 5539 | }; |
| 5540 | auto AddShuffleMaskCost = [&](InstructionCost C, ArrayRef<int> Mask) { |
| 5541 | return C + |
| 5542 | TTI.getShuffleCost(Kind: TTI::SK_PermuteTwoSrc, DstTy: VT, SrcTy: VT, Mask, CostKind); |
| 5543 | }; |
| 5544 | |
| 5545 | unsigned ElementSize = VT->getElementType()->getPrimitiveSizeInBits(); |
| 5546 | unsigned MaxVectorSize = |
| 5547 | TTI.getRegisterBitWidth(K: TargetTransformInfo::RGK_FixedWidthVector); |
| 5548 | unsigned MaxElementsInVector = MaxVectorSize / ElementSize; |
| 5549 | if (MaxElementsInVector == 0) |
| 5550 | return false; |
| 5551 | // When there are multiple shufflevector operations on the same input, |
| 5552 | // especially when the vector length is larger than the register size, |
| 5553 | // identical shuffle patterns may occur across different groups of elements. |
| 5554 | // To avoid overestimating the cost by counting these repeated shuffles more |
| 5555 | // than once, we only account for unique shuffle patterns. This adjustment |
| 5556 | // prevents inflated costs in the cost model for wide vectors split into |
| 5557 | // several register-sized groups. |
| 5558 | std::set<SmallVector<int, 4>> UniqueShuffles; |
| 5559 | auto AddShuffleMaskAdjustedCost = [&](InstructionCost C, ArrayRef<int> Mask) { |
| 5560 | // Compute the cost for performing the shuffle over the full vector. |
| 5561 | auto ShuffleCost = |
| 5562 | TTI.getShuffleCost(Kind: TTI::SK_PermuteTwoSrc, DstTy: VT, SrcTy: VT, Mask, CostKind); |
| 5563 | unsigned NumFullVectors = Mask.size() / MaxElementsInVector; |
| 5564 | if (NumFullVectors < 2) |
| 5565 | return C + ShuffleCost; |
| 5566 | SmallVector<int, 4> SubShuffle(MaxElementsInVector); |
| 5567 | unsigned NumUniqueGroups = 0; |
| 5568 | unsigned NumGroups = Mask.size() / MaxElementsInVector; |
| 5569 | // For each group of MaxElementsInVector contiguous elements, |
| 5570 | // collect their shuffle pattern and insert into the set of unique patterns. |
| 5571 | for (unsigned I = 0; I < NumFullVectors; ++I) { |
| 5572 | for (unsigned J = 0; J < MaxElementsInVector; ++J) |
| 5573 | SubShuffle[J] = Mask[MaxElementsInVector * I + J]; |
| 5574 | if (UniqueShuffles.insert(x: SubShuffle).second) |
| 5575 | NumUniqueGroups += 1; |
| 5576 | } |
| 5577 | return C + ShuffleCost * NumUniqueGroups / NumGroups; |
| 5578 | }; |
| 5579 | auto AddShuffleAdjustedCost = [&](InstructionCost C, Instruction *I) { |
| 5580 | auto *SV = dyn_cast<ShuffleVectorInst>(Val: I); |
| 5581 | if (!SV) |
| 5582 | return C; |
| 5583 | SmallVector<int, 16> Mask; |
| 5584 | SV->getShuffleMask(Result&: Mask); |
| 5585 | return AddShuffleMaskAdjustedCost(C, Mask); |
| 5586 | }; |
| 5587 | // Check that input consists of ShuffleVectors applied to the same input |
| 5588 | auto AllShufflesHaveSameOperands = |
| 5589 | [](SmallPtrSetImpl<Instruction *> &InputShuffles) { |
| 5590 | if (InputShuffles.size() < 2) |
| 5591 | return false; |
| 5592 | ShuffleVectorInst *FirstSV = |
| 5593 | dyn_cast<ShuffleVectorInst>(Val: *InputShuffles.begin()); |
| 5594 | if (!FirstSV) |
| 5595 | return false; |
| 5596 | |
| 5597 | Value *In0 = FirstSV->getOperand(i_nocapture: 0), *In1 = FirstSV->getOperand(i_nocapture: 1); |
| 5598 | return std::all_of( |
| 5599 | first: std::next(x: InputShuffles.begin()), last: InputShuffles.end(), |
| 5600 | pred: [&](Instruction *I) { |
| 5601 | ShuffleVectorInst *SV = dyn_cast<ShuffleVectorInst>(Val: I); |
| 5602 | return SV && SV->getOperand(i_nocapture: 0) == In0 && SV->getOperand(i_nocapture: 1) == In1; |
| 5603 | }); |
| 5604 | }; |
| 5605 | |
| 5606 | // Get the costs of the shuffles + binops before and after with the new |
| 5607 | // shuffle masks. |
| 5608 | InstructionCost CostBefore = |
| 5609 | TTI.getArithmeticInstrCost(Opcode: Op0->getOpcode(), Ty: VT, CostKind) + |
| 5610 | TTI.getArithmeticInstrCost(Opcode: Op1->getOpcode(), Ty: VT, CostKind); |
| 5611 | CostBefore += std::accumulate(first: Shuffles.begin(), last: Shuffles.end(), |
| 5612 | init: InstructionCost(0), binary_op: AddShuffleCost); |
| 5613 | if (AllShufflesHaveSameOperands(InputShuffles)) { |
| 5614 | UniqueShuffles.clear(); |
| 5615 | CostBefore += std::accumulate(first: InputShuffles.begin(), last: InputShuffles.end(), |
| 5616 | init: InstructionCost(0), binary_op: AddShuffleAdjustedCost); |
| 5617 | } else { |
| 5618 | CostBefore += std::accumulate(first: InputShuffles.begin(), last: InputShuffles.end(), |
| 5619 | init: InstructionCost(0), binary_op: AddShuffleCost); |
| 5620 | } |
| 5621 | |
| 5622 | // The new binops will be unused for lanes past the used shuffle lengths. |
| 5623 | // These types attempt to get the correct cost for that from the target. |
| 5624 | FixedVectorType *Op0SmallVT = |
| 5625 | FixedVectorType::get(ElementType: VT->getScalarType(), NumElts: V1.size()); |
| 5626 | FixedVectorType *Op1SmallVT = |
| 5627 | FixedVectorType::get(ElementType: VT->getScalarType(), NumElts: V2.size()); |
| 5628 | InstructionCost CostAfter = |
| 5629 | TTI.getArithmeticInstrCost(Opcode: Op0->getOpcode(), Ty: Op0SmallVT, CostKind) + |
| 5630 | TTI.getArithmeticInstrCost(Opcode: Op1->getOpcode(), Ty: Op1SmallVT, CostKind); |
| 5631 | UniqueShuffles.clear(); |
| 5632 | CostAfter += std::accumulate(first: ReconstructMasks.begin(), last: ReconstructMasks.end(), |
| 5633 | init: InstructionCost(0), binary_op: AddShuffleMaskAdjustedCost); |
| 5634 | std::set<SmallVector<int>> OutputShuffleMasks({V1A, V1B, V2A, V2B}); |
| 5635 | CostAfter += |
| 5636 | std::accumulate(first: OutputShuffleMasks.begin(), last: OutputShuffleMasks.end(), |
| 5637 | init: InstructionCost(0), binary_op: AddShuffleMaskCost); |
| 5638 | |
| 5639 | LLVM_DEBUG(dbgs() << "Found a binop select shuffle pattern: " << I << "\n" ); |
| 5640 | LLVM_DEBUG(dbgs() << " CostBefore: " << CostBefore |
| 5641 | << " vs CostAfter: " << CostAfter << "\n" ); |
| 5642 | if (CostBefore < CostAfter || |
| 5643 | (CostBefore == CostAfter && !feedsIntoVectorReduction(SVI))) |
| 5644 | return false; |
| 5645 | |
| 5646 | // The cost model has passed, create the new instructions. |
| 5647 | auto GetShuffleOperand = [&](Instruction *I, unsigned Op) -> Value * { |
| 5648 | auto *SV = dyn_cast<ShuffleVectorInst>(Val: I); |
| 5649 | if (!SV) |
| 5650 | return I; |
| 5651 | if (isa<UndefValue>(Val: SV->getOperand(i_nocapture: 1))) |
| 5652 | if (auto *SSV = dyn_cast<ShuffleVectorInst>(Val: SV->getOperand(i_nocapture: 0))) |
| 5653 | if (InputShuffles.contains(Ptr: SSV)) |
| 5654 | return SSV->getOperand(i_nocapture: Op); |
| 5655 | return SV->getOperand(i_nocapture: Op); |
| 5656 | }; |
| 5657 | Builder.SetInsertPoint(*SVI0A->getInsertionPointAfterDef()); |
| 5658 | Value *NSV0A = Builder.CreateShuffleVector(V1: GetShuffleOperand(SVI0A, 0), |
| 5659 | V2: GetShuffleOperand(SVI0A, 1), Mask: V1A); |
| 5660 | Builder.SetInsertPoint(*SVI0B->getInsertionPointAfterDef()); |
| 5661 | Value *NSV0B = Builder.CreateShuffleVector(V1: GetShuffleOperand(SVI0B, 0), |
| 5662 | V2: GetShuffleOperand(SVI0B, 1), Mask: V1B); |
| 5663 | Builder.SetInsertPoint(*SVI1A->getInsertionPointAfterDef()); |
| 5664 | Value *NSV1A = Builder.CreateShuffleVector(V1: GetShuffleOperand(SVI1A, 0), |
| 5665 | V2: GetShuffleOperand(SVI1A, 1), Mask: V2A); |
| 5666 | Builder.SetInsertPoint(*SVI1B->getInsertionPointAfterDef()); |
| 5667 | Value *NSV1B = Builder.CreateShuffleVector(V1: GetShuffleOperand(SVI1B, 0), |
| 5668 | V2: GetShuffleOperand(SVI1B, 1), Mask: V2B); |
| 5669 | Builder.SetInsertPoint(Op0); |
| 5670 | Value *NOp0 = Builder.CreateBinOp(Opc: (Instruction::BinaryOps)Op0->getOpcode(), |
| 5671 | LHS: NSV0A, RHS: NSV0B); |
| 5672 | if (auto *I = dyn_cast<Instruction>(Val: NOp0)) |
| 5673 | I->copyIRFlags(V: Op0, IncludeWrapFlags: true); |
| 5674 | Builder.SetInsertPoint(Op1); |
| 5675 | Value *NOp1 = Builder.CreateBinOp(Opc: (Instruction::BinaryOps)Op1->getOpcode(), |
| 5676 | LHS: NSV1A, RHS: NSV1B); |
| 5677 | if (auto *I = dyn_cast<Instruction>(Val: NOp1)) |
| 5678 | I->copyIRFlags(V: Op1, IncludeWrapFlags: true); |
| 5679 | |
| 5680 | for (int S = 0, E = ReconstructMasks.size(); S != E; S++) { |
| 5681 | Builder.SetInsertPoint(Shuffles[S]); |
| 5682 | Value *NSV = Builder.CreateShuffleVector(V1: NOp0, V2: NOp1, Mask: ReconstructMasks[S]); |
| 5683 | replaceValue(Old&: *Shuffles[S], New&: *NSV, Erase: false); |
| 5684 | } |
| 5685 | |
| 5686 | Worklist.pushValue(V: NSV0A); |
| 5687 | Worklist.pushValue(V: NSV0B); |
| 5688 | Worklist.pushValue(V: NSV1A); |
| 5689 | Worklist.pushValue(V: NSV1B); |
| 5690 | return true; |
| 5691 | } |
| 5692 | |
| 5693 | /// Check if instruction depends on ZExt and this ZExt can be moved after the |
| 5694 | /// instruction. Move ZExt if it is profitable. For example: |
| 5695 | /// logic(zext(x),y) -> zext(logic(x,trunc(y))) |
| 5696 | /// lshr((zext(x),y) -> zext(lshr(x,trunc(y))) |
| 5697 | /// Cost model calculations takes into account if zext(x) has other users and |
| 5698 | /// whether it can be propagated through them too. |
| 5699 | bool VectorCombine::shrinkType(Instruction &I) { |
| 5700 | Value *ZExted, *OtherOperand; |
| 5701 | if (!match(V: &I, P: m_c_BitwiseLogic(L: m_ZExt(Op: m_Value(V&: ZExted)), |
| 5702 | R: m_Value(V&: OtherOperand))) && |
| 5703 | !match(V: &I, P: m_LShr(L: m_ZExt(Op: m_Value(V&: ZExted)), R: m_Value(V&: OtherOperand)))) |
| 5704 | return false; |
| 5705 | |
| 5706 | Value *ZExtOperand = I.getOperand(i: I.getOperand(i: 0) == OtherOperand ? 1 : 0); |
| 5707 | |
| 5708 | auto *BigTy = cast<FixedVectorType>(Val: I.getType()); |
| 5709 | auto *SmallTy = cast<FixedVectorType>(Val: ZExted->getType()); |
| 5710 | unsigned BW = SmallTy->getElementType()->getPrimitiveSizeInBits(); |
| 5711 | |
| 5712 | if (I.getOpcode() == Instruction::LShr) { |
| 5713 | // Check that the shift amount is less than the number of bits in the |
| 5714 | // smaller type. Otherwise, the smaller lshr will return a poison value. |
| 5715 | KnownBits ShAmtKB = computeKnownBits(V: I.getOperand(i: 1), DL: *DL); |
| 5716 | if (ShAmtKB.getMaxValue().uge(RHS: BW)) |
| 5717 | return false; |
| 5718 | } else { |
| 5719 | // Check that the expression overall uses at most the same number of bits as |
| 5720 | // ZExted |
| 5721 | KnownBits KB = computeKnownBits(V: &I, DL: *DL); |
| 5722 | if (KB.countMaxActiveBits() > BW) |
| 5723 | return false; |
| 5724 | } |
| 5725 | |
| 5726 | // Calculate costs of leaving current IR as it is and moving ZExt operation |
| 5727 | // later, along with adding truncates if needed |
| 5728 | InstructionCost ZExtCost = TTI.getCastInstrCost( |
| 5729 | Opcode: Instruction::ZExt, Dst: BigTy, Src: SmallTy, |
| 5730 | CCH: TargetTransformInfo::CastContextHint::None, CostKind); |
| 5731 | InstructionCost CurrentCost = ZExtCost; |
| 5732 | InstructionCost ShrinkCost = 0; |
| 5733 | |
| 5734 | // Calculate total cost and check that we can propagate through all ZExt users |
| 5735 | for (User *U : ZExtOperand->users()) { |
| 5736 | auto *UI = cast<Instruction>(Val: U); |
| 5737 | if (UI == &I) { |
| 5738 | CurrentCost += |
| 5739 | TTI.getArithmeticInstrCost(Opcode: UI->getOpcode(), Ty: BigTy, CostKind); |
| 5740 | ShrinkCost += |
| 5741 | TTI.getArithmeticInstrCost(Opcode: UI->getOpcode(), Ty: SmallTy, CostKind); |
| 5742 | ShrinkCost += ZExtCost; |
| 5743 | continue; |
| 5744 | } |
| 5745 | |
| 5746 | if (!Instruction::isBinaryOp(Opcode: UI->getOpcode())) |
| 5747 | return false; |
| 5748 | |
| 5749 | // Check if we can propagate ZExt through its other users |
| 5750 | KnownBits KB = computeKnownBits(V: UI, DL: *DL); |
| 5751 | if (KB.countMaxActiveBits() > BW) |
| 5752 | return false; |
| 5753 | |
| 5754 | CurrentCost += TTI.getArithmeticInstrCost(Opcode: UI->getOpcode(), Ty: BigTy, CostKind); |
| 5755 | ShrinkCost += |
| 5756 | TTI.getArithmeticInstrCost(Opcode: UI->getOpcode(), Ty: SmallTy, CostKind); |
| 5757 | ShrinkCost += ZExtCost; |
| 5758 | } |
| 5759 | |
| 5760 | // If the other instruction operand is not a constant, we'll need to |
| 5761 | // generate a truncate instruction. So we have to adjust cost |
| 5762 | if (!isa<Constant>(Val: OtherOperand)) |
| 5763 | ShrinkCost += TTI.getCastInstrCost( |
| 5764 | Opcode: Instruction::Trunc, Dst: SmallTy, Src: BigTy, |
| 5765 | CCH: TargetTransformInfo::CastContextHint::None, CostKind); |
| 5766 | |
| 5767 | // If the cost of shrinking types and leaving the IR is the same, we'll lean |
| 5768 | // towards modifying the IR because shrinking opens opportunities for other |
| 5769 | // shrinking optimisations. |
| 5770 | if (ShrinkCost > CurrentCost) |
| 5771 | return false; |
| 5772 | |
| 5773 | Builder.SetInsertPoint(&I); |
| 5774 | Value *Op0 = ZExted; |
| 5775 | Value *Op1 = Builder.CreateTrunc(V: OtherOperand, DestTy: SmallTy); |
| 5776 | // Keep the order of operands the same |
| 5777 | if (I.getOperand(i: 0) == OtherOperand) |
| 5778 | std::swap(a&: Op0, b&: Op1); |
| 5779 | Value *NewBinOp = |
| 5780 | Builder.CreateBinOp(Opc: (Instruction::BinaryOps)I.getOpcode(), LHS: Op0, RHS: Op1); |
| 5781 | cast<Instruction>(Val: NewBinOp)->copyIRFlags(V: &I); |
| 5782 | cast<Instruction>(Val: NewBinOp)->copyMetadata(SrcInst: I); |
| 5783 | Value *NewZExtr = Builder.CreateZExt(V: NewBinOp, DestTy: BigTy); |
| 5784 | replaceValue(Old&: I, New&: *NewZExtr); |
| 5785 | return true; |
| 5786 | } |
| 5787 | |
| 5788 | /// insert (DstVec, (extract SrcVec, ExtIdx), InsIdx) --> |
| 5789 | /// shuffle (DstVec, SrcVec, Mask) |
| 5790 | bool VectorCombine::foldInsExtVectorToShuffle(Instruction &I) { |
| 5791 | Value *DstVec, *SrcVec; |
| 5792 | uint64_t ExtIdx, InsIdx; |
| 5793 | if (!match(V: &I, |
| 5794 | P: m_InsertElt(Val: m_Value(V&: DstVec), |
| 5795 | Elt: m_ExtractElt(Val: m_Value(V&: SrcVec), Idx: m_ConstantInt(V&: ExtIdx)), |
| 5796 | Idx: m_ConstantInt(V&: InsIdx)))) |
| 5797 | return false; |
| 5798 | |
| 5799 | auto *DstVecTy = dyn_cast<FixedVectorType>(Val: I.getType()); |
| 5800 | auto *SrcVecTy = dyn_cast<FixedVectorType>(Val: SrcVec->getType()); |
| 5801 | // We can try combining vectors with different element sizes. |
| 5802 | if (!DstVecTy || !SrcVecTy || |
| 5803 | SrcVecTy->getElementType() != DstVecTy->getElementType()) |
| 5804 | return false; |
| 5805 | |
| 5806 | unsigned NumDstElts = DstVecTy->getNumElements(); |
| 5807 | unsigned NumSrcElts = SrcVecTy->getNumElements(); |
| 5808 | if (InsIdx >= NumDstElts || ExtIdx >= NumSrcElts || NumDstElts == 1) |
| 5809 | return false; |
| 5810 | |
| 5811 | // Insertion into poison is a cheaper single operand shuffle. |
| 5812 | TargetTransformInfo::ShuffleKind SK; |
| 5813 | SmallVector<int> Mask(NumDstElts, PoisonMaskElem); |
| 5814 | |
| 5815 | bool NeedExpOrNarrow = NumSrcElts != NumDstElts; |
| 5816 | bool NeedDstSrcSwap = isa<PoisonValue>(Val: DstVec) && !isa<UndefValue>(Val: SrcVec); |
| 5817 | if (NeedDstSrcSwap) { |
| 5818 | SK = TargetTransformInfo::SK_PermuteSingleSrc; |
| 5819 | Mask[InsIdx] = ExtIdx % NumDstElts; |
| 5820 | std::swap(a&: DstVec, b&: SrcVec); |
| 5821 | } else { |
| 5822 | SK = TargetTransformInfo::SK_PermuteTwoSrc; |
| 5823 | std::iota(first: Mask.begin(), last: Mask.end(), value: 0); |
| 5824 | Mask[InsIdx] = (ExtIdx % NumDstElts) + NumDstElts; |
| 5825 | } |
| 5826 | |
| 5827 | // Cost |
| 5828 | auto *Ins = cast<InsertElementInst>(Val: &I); |
| 5829 | auto *Ext = cast<ExtractElementInst>(Val: I.getOperand(i: 1)); |
| 5830 | InstructionCost InsCost = |
| 5831 | TTI.getVectorInstrCost(I: *Ins, Val: DstVecTy, CostKind, Index: InsIdx); |
| 5832 | InstructionCost ExtCost = |
| 5833 | TTI.getVectorInstrCost(I: *Ext, Val: DstVecTy, CostKind, Index: ExtIdx); |
| 5834 | InstructionCost OldCost = ExtCost + InsCost; |
| 5835 | |
| 5836 | InstructionCost NewCost = 0; |
| 5837 | SmallVector<int> ExtToVecMask; |
| 5838 | if (!NeedExpOrNarrow) { |
| 5839 | // Ignore 'free' identity insertion shuffle. |
| 5840 | // TODO: getShuffleCost should return TCC_Free for Identity shuffles. |
| 5841 | if (!ShuffleVectorInst::isIdentityMask(Mask, NumSrcElts)) |
| 5842 | NewCost += TTI.getShuffleCost(Kind: SK, DstTy: DstVecTy, SrcTy: DstVecTy, Mask, CostKind, Index: 0, |
| 5843 | SubTp: nullptr, Args: {DstVec, SrcVec}); |
| 5844 | } else { |
| 5845 | // When creating a length-changing-vector, always try to keep the relevant |
| 5846 | // element in an equivalent position, so that bulk shuffles are more likely |
| 5847 | // to be useful. |
| 5848 | ExtToVecMask.assign(NumElts: NumDstElts, Elt: PoisonMaskElem); |
| 5849 | ExtToVecMask[ExtIdx % NumDstElts] = ExtIdx; |
| 5850 | // Add cost for expanding or narrowing |
| 5851 | NewCost = TTI.getShuffleCost(Kind: TargetTransformInfo::SK_PermuteSingleSrc, |
| 5852 | DstTy: DstVecTy, SrcTy: SrcVecTy, Mask: ExtToVecMask, CostKind); |
| 5853 | NewCost += TTI.getShuffleCost(Kind: SK, DstTy: DstVecTy, SrcTy: DstVecTy, Mask, CostKind); |
| 5854 | } |
| 5855 | |
| 5856 | if (!Ext->hasOneUse()) |
| 5857 | NewCost += ExtCost; |
| 5858 | |
| 5859 | LLVM_DEBUG(dbgs() << "Found a insert/extract shuffle-like pair: " << I |
| 5860 | << "\n OldCost: " << OldCost << " vs NewCost: " << NewCost |
| 5861 | << "\n" ); |
| 5862 | |
| 5863 | if (OldCost < NewCost) |
| 5864 | return false; |
| 5865 | |
| 5866 | if (NeedExpOrNarrow) { |
| 5867 | if (!NeedDstSrcSwap) |
| 5868 | SrcVec = Builder.CreateShuffleVector(V: SrcVec, Mask: ExtToVecMask); |
| 5869 | else |
| 5870 | DstVec = Builder.CreateShuffleVector(V: DstVec, Mask: ExtToVecMask); |
| 5871 | } |
| 5872 | |
| 5873 | // Canonicalize undef param to RHS to help further folds. |
| 5874 | if (isa<UndefValue>(Val: DstVec) && !isa<UndefValue>(Val: SrcVec)) { |
| 5875 | ShuffleVectorInst::commuteShuffleMask(Mask, InVecNumElts: NumDstElts); |
| 5876 | std::swap(a&: DstVec, b&: SrcVec); |
| 5877 | } |
| 5878 | |
| 5879 | Value *Shuf = Builder.CreateShuffleVector(V1: DstVec, V2: SrcVec, Mask); |
| 5880 | replaceValue(Old&: I, New&: *Shuf); |
| 5881 | |
| 5882 | return true; |
| 5883 | } |
| 5884 | |
| 5885 | /// If we're interleaving 2 constant splats, for instance `<vscale x 8 x i32> |
| 5886 | /// <splat of 666>` and `<vscale x 8 x i32> <splat of 777>`, we can create a |
| 5887 | /// larger splat `<vscale x 8 x i64> <splat of ((777 << 32) | 666)>` first |
| 5888 | /// before casting it back into `<vscale x 16 x i32>`. |
| 5889 | bool VectorCombine::foldInterleaveIntrinsics(Instruction &I) { |
| 5890 | const APInt *SplatVal0, *SplatVal1; |
| 5891 | if (!match(V: &I, P: m_Intrinsic<Intrinsic::vector_interleave2>( |
| 5892 | Ops: m_APInt(Res&: SplatVal0), Ops: m_APInt(Res&: SplatVal1)))) |
| 5893 | return false; |
| 5894 | |
| 5895 | LLVM_DEBUG(dbgs() << "VC: Folding interleave2 with two splats: " << I |
| 5896 | << "\n" ); |
| 5897 | |
| 5898 | auto *VTy = |
| 5899 | cast<VectorType>(Val: cast<IntrinsicInst>(Val&: I).getArgOperand(i: 0)->getType()); |
| 5900 | auto *ExtVTy = VectorType::getExtendedElementVectorType(VTy); |
| 5901 | unsigned Width = VTy->getElementType()->getIntegerBitWidth(); |
| 5902 | |
| 5903 | // Just in case the cost of interleave2 intrinsic and bitcast are both |
| 5904 | // invalid, in which case we want to bail out, we use <= rather |
| 5905 | // than < here. Even they both have valid and equal costs, it's probably |
| 5906 | // not a good idea to emit a high-cost constant splat. |
| 5907 | if (TTI.getInstructionCost(U: &I, CostKind) <= |
| 5908 | TTI.getCastInstrCost(Opcode: Instruction::BitCast, Dst: I.getType(), Src: ExtVTy, |
| 5909 | CCH: TTI::CastContextHint::None, CostKind)) { |
| 5910 | LLVM_DEBUG(dbgs() << "VC: The cost to cast from " << *ExtVTy << " to " |
| 5911 | << *I.getType() << " is too high.\n" ); |
| 5912 | return false; |
| 5913 | } |
| 5914 | |
| 5915 | APInt NewSplatVal = SplatVal1->zext(width: Width * 2); |
| 5916 | NewSplatVal <<= Width; |
| 5917 | NewSplatVal |= SplatVal0->zext(width: Width * 2); |
| 5918 | auto *NewSplat = ConstantVector::getSplat( |
| 5919 | EC: ExtVTy->getElementCount(), Elt: ConstantInt::get(Context&: F.getContext(), V: NewSplatVal)); |
| 5920 | |
| 5921 | IRBuilder<> Builder(&I); |
| 5922 | replaceValue(Old&: I, New&: *Builder.CreateBitCast(V: NewSplat, DestTy: I.getType())); |
| 5923 | return true; |
| 5924 | } |
| 5925 | |
| 5926 | /// Given this sequence: |
| 5927 | /// ``` |
| 5928 | /// %d = llvm.vector.deinterleave2 <vscale x 16 x i32> %v |
| 5929 | /// %f0 = extractvalue { <vscale x 8 x i32>, <vscale x 8 x i32> } %d, 0 |
| 5930 | /// %f1 = extractvalue { <vscale x 8 x i32>, <vscale x 8 x i32> } %d, 1 |
| 5931 | /// |
| 5932 | /// %low0 = and <vscale x 8 x i32> %f0, splat (i32 65535) |
| 5933 | /// %low1 = shl <vscale x 8 x i32> %f1, splat (i32 16) |
| 5934 | /// %merge0 = or disjoint <vscale x 8 x i32> %low0, %low1 |
| 5935 | /// |
| 5936 | /// %high0 = and <vscale x 8 x i32> %f1, splat (i32 -65536) |
| 5937 | /// %high1 = lshr <vscale x 8 x i32> %f0, splat (i32 16) |
| 5938 | /// %merge1 = or disjoint <vscale x 8 x i32> %high0, %high1 |
| 5939 | /// ``` |
| 5940 | /// It is actually just de-interleaving a 16-bit vector with double the |
| 5941 | /// vector length. More generally speaking, it's de-interleaving on a vector |
| 5942 | /// with half the element width as the original vector. |
| 5943 | /// |
| 5944 | /// Therefore, we can turn it into: |
| 5945 | /// ``` |
| 5946 | /// %narrow.v = bitcast <vscale x 16 x i32> %v to <vscale x 32 x i16> |
| 5947 | /// %d = llvm.vector.deinterleave2 <vscale x 32 x i16> %narrow.v |
| 5948 | /// %f0 = extractvalue { <vscale x 16 x i16>, <vscale x 16 x i16> } %d, 0 |
| 5949 | /// %f1 = extractvalue { <vscale x 16 x i16>, <vscale x 16 x i16> } %d, 1 |
| 5950 | /// |
| 5951 | /// %merge0 = bitcast <vscale x 16 x i16> %f0 to <vscale x 8 x i32> |
| 5952 | /// %merge1 = bitcast <vscale x 16 x i16> %f1 to <vscale x 8 x i32> |
| 5953 | /// ``` |
| 5954 | bool VectorCombine::foldDeinterleaveIntrinsics(Instruction &I) { |
| 5955 | // This pattern involves bitcast that is not compatible with big endian. |
| 5956 | if (DL->isBigEndian()) |
| 5957 | return false; |
| 5958 | |
| 5959 | using namespace PatternMatch; |
| 5960 | Value *DeinterleavedVal; |
| 5961 | if (!match(V: &I, P: m_Deinterleave2(Op: m_Value(V&: DeinterleavedVal)))) |
| 5962 | return false; |
| 5963 | |
| 5964 | VectorType *VecTy = cast<VectorType>(Val: DeinterleavedVal->getType()); |
| 5965 | IntegerType *ElementTy = dyn_cast<IntegerType>(Val: VecTy->getElementType()); |
| 5966 | if (!ElementTy) |
| 5967 | return false; |
| 5968 | unsigned ElementWidth = ElementTy->getBitWidth(); |
| 5969 | if (ElementWidth < 2 || !isPowerOf2_32(Value: ElementWidth)) |
| 5970 | return false; |
| 5971 | unsigned HalfElementWidth = ElementWidth / 2; |
| 5972 | |
| 5973 | if (!I.hasNUses(N: 2)) |
| 5974 | return false; |
| 5975 | std::array<ExtractValueInst *, 2> OrigFields{}; |
| 5976 | for (User *Usr : I.users()) { |
| 5977 | auto *E = dyn_cast<ExtractValueInst>(Val: Usr); |
| 5978 | // The deinterleave result can only be used by extractions. |
| 5979 | if (!E || E->getNumIndices() != 1) |
| 5980 | return false; |
| 5981 | unsigned Idx = *E->idx_begin(); |
| 5982 | // A single field cannot be extracted more than once. |
| 5983 | if (Idx >= 2 || OrigFields[Idx] || !E->hasNUses(N: 2)) |
| 5984 | return false; |
| 5985 | OrigFields[Idx] = E; |
| 5986 | } |
| 5987 | |
| 5988 | // Find the merge instruction (i.e. OR) first. |
| 5989 | SmallVector<Instruction *, 2> MergeInsts; |
| 5990 | for (auto *FieldUsr : OrigFields[0]->users()) { |
| 5991 | if (!FieldUsr->hasOneUse() || !isa<Instruction>(Val: FieldUsr->user_back())) |
| 5992 | return false; |
| 5993 | MergeInsts.push_back(Elt: cast<Instruction>(Val: FieldUsr->user_back())); |
| 5994 | } |
| 5995 | assert(MergeInsts.size() == 2); |
| 5996 | |
| 5997 | // Pattern match bottom-up from the merge instructions. |
| 5998 | auto MatchMerge = [&](void) -> bool { |
| 5999 | APInt LoMask = APInt::getLowBitsSet(numBits: ElementWidth, loBitsSet: HalfElementWidth); |
| 6000 | APInt HiMask = APInt::getHighBitsSet(numBits: ElementWidth, hiBitsSet: HalfElementWidth); |
| 6001 | return match(V: MergeInsts[0], |
| 6002 | P: m_c_Or(L: m_And(L: m_Specific(V: OrigFields[0]), R: m_SpecificInt(V: LoMask)), |
| 6003 | R: m_Shl(L: m_Specific(V: OrigFields[1]), |
| 6004 | R: m_SpecificInt(V: HalfElementWidth)))) && |
| 6005 | match(V: MergeInsts[1], |
| 6006 | P: m_c_Or(L: m_And(L: m_Specific(V: OrigFields[1]), R: m_SpecificInt(V: HiMask)), |
| 6007 | R: m_LShr(L: m_Specific(V: OrigFields[0]), |
| 6008 | R: m_SpecificInt(V: HalfElementWidth)))); |
| 6009 | }; |
| 6010 | if (!MatchMerge()) { |
| 6011 | std::swap(a&: MergeInsts[0], b&: MergeInsts[1]); |
| 6012 | if (!MatchMerge()) |
| 6013 | return false; |
| 6014 | } |
| 6015 | |
| 6016 | // Profitability check. |
| 6017 | InstructionCost OldCost = |
| 6018 | TTI.getInstructionCost(U: MergeInsts[0], CostKind) + |
| 6019 | TTI.getInstructionCost(U: cast<Instruction>(Val: MergeInsts[0]->getOperand(i: 0)), |
| 6020 | CostKind) + |
| 6021 | TTI.getInstructionCost(U: cast<Instruction>(Val: MergeInsts[0]->getOperand(i: 1)), |
| 6022 | CostKind); |
| 6023 | // There are two fields (assuming SHL has the same cost as LSHR). |
| 6024 | OldCost *= 2; |
| 6025 | |
| 6026 | auto *NewFieldTy = VecTy->getWithNewBitWidth(NewBitWidth: HalfElementWidth); |
| 6027 | auto *NewVecTy = |
| 6028 | VectorType::getDoubleElementsVectorType(VTy: cast<VectorType>(Val: NewFieldTy)); |
| 6029 | InstructionCost NewCost = |
| 6030 | TTI.getCastInstrCost(Opcode: Instruction::BitCast, Dst: VecTy, Src: NewVecTy, |
| 6031 | CCH: TTI::CastContextHint::None, CostKind) + |
| 6032 | TTI.getCastInstrCost(Opcode: Instruction::BitCast, Dst: NewFieldTy, |
| 6033 | Src: MergeInsts[0]->getType(), CCH: TTI::CastContextHint::None, |
| 6034 | CostKind) * |
| 6035 | 2; |
| 6036 | if (OldCost <= NewCost || !NewCost.isValid()) { |
| 6037 | LLVM_DEBUG( |
| 6038 | dbgs() << "VC: New deinterleave2 sequence cost (" << NewCost << ")" |
| 6039 | << " is higher than that of the old one (" << OldCost << ")\n" ); |
| 6040 | return false; |
| 6041 | } |
| 6042 | |
| 6043 | // Do the replacement. |
| 6044 | IRBuilder<> Builder(&I); |
| 6045 | Value *NewVecCast = Builder.CreateBitCast(V: DeinterleavedVal, DestTy: NewVecTy); |
| 6046 | Value *NewDeinterleave = Builder.CreateIntrinsic( |
| 6047 | ID: Intrinsic::vector_deinterleave2, OverloadTypes: {NewVecTy}, Args: {NewVecCast}); |
| 6048 | for (auto [Idx, MergeInst] : enumerate(First&: MergeInsts)) { |
| 6049 | Value *NewField = Builder.CreateExtractValue(Agg: NewDeinterleave, Idxs: Idx); |
| 6050 | NewField = Builder.CreateBitCast(V: NewField, DestTy: MergeInst->getType()); |
| 6051 | replaceValue(Old&: *MergeInst, New&: *NewField); |
| 6052 | } |
| 6053 | |
| 6054 | return true; |
| 6055 | } |
| 6056 | |
| 6057 | bool VectorCombine::foldBitcastOfVPLoad(Instruction &I) { |
| 6058 | const DataLayout &DL = I.getDataLayout(); |
| 6059 | auto *Cast = dyn_cast<CastInst>(Val: &I); |
| 6060 | if (!Cast || !Cast->isNoopCast(DL) || !isa<VectorType>(Val: Cast->getDestTy())) |
| 6061 | return false; |
| 6062 | |
| 6063 | // Fold away bit casts of the loaded value by loading the desired type, |
| 6064 | // if the mask is all-ones. |
| 6065 | Value *EVL; |
| 6066 | auto *II = dyn_cast<VPIntrinsic>(Val: I.getOperand(i: 0)); |
| 6067 | if (!II || !match(V: II, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::vp_load>( |
| 6068 | Ops: m_Value(), Ops: m_AllOnes(), Ops: m_Value(V&: EVL))))) |
| 6069 | return false; |
| 6070 | |
| 6071 | VectorType *OrigVecTy = cast<VectorType>(Val: II->getType()); |
| 6072 | Align OrigAlign = |
| 6073 | DL.getValueOrABITypeAlignment(Alignment: II->getPointerAlignment(), Ty: OrigVecTy); |
| 6074 | ElementCount OrigVecCnt = OrigVecTy->getElementCount(); |
| 6075 | VectorType *NewVecTy = cast<VectorType>(Val: Cast->getDestTy()); |
| 6076 | ElementCount NewVecCnt = NewVecTy->getElementCount(); |
| 6077 | |
| 6078 | // Right now we only support cases where the NewVec is longer, because for |
| 6079 | // cases where it's shorter, we have to be sure that EVL can be exactly |
| 6080 | // divided, otherwise it might yield incorrect results or even page faults |
| 6081 | // (if we round-up during the division). |
| 6082 | if (!(OrigVecCnt.isScalable() == NewVecCnt.isScalable() && |
| 6083 | NewVecCnt.hasKnownScalarFactor(RHS: OrigVecCnt))) |
| 6084 | return false; |
| 6085 | |
| 6086 | InstructionCost OldCost = |
| 6087 | TTI.getMemIntrinsicInstrCost(MICA: {Intrinsic::vp_load, OrigVecTy, |
| 6088 | II->getMemoryPointerParam(), false, |
| 6089 | OrigAlign}, |
| 6090 | CostKind) + |
| 6091 | TTI.getCastInstrCost(Opcode: Instruction::BitCast, Dst: Cast->getType(), Src: OrigVecTy, |
| 6092 | CCH: TTI::CastContextHint::None, CostKind); |
| 6093 | InstructionCost NewCost = TTI.getMemIntrinsicInstrCost( |
| 6094 | MICA: {Intrinsic::vp_load, NewVecTy, II->getMemoryPointerParam(), false, |
| 6095 | OrigAlign}, |
| 6096 | CostKind); |
| 6097 | LLVM_DEBUG(dbgs() << "foldBitcastOfVPLoad: OldCost=" << OldCost |
| 6098 | << " NewCost=" << NewCost << "\n" ); |
| 6099 | if (NewCost > OldCost || !NewCost.isValid()) |
| 6100 | return false; |
| 6101 | |
| 6102 | unsigned Factor = NewVecCnt.getKnownScalarFactor(RHS: OrigVecCnt); |
| 6103 | Value *NewEVL = Builder.CreateNUWMul(LHS: EVL, RHS: Builder.getInt32(C: Factor)); |
| 6104 | Value *NewMask = Builder.CreateVectorSplat(EC: NewVecCnt, V: Builder.getTrue()); |
| 6105 | CallInst *NewVP = Builder.CreateIntrinsicWithoutFolding( |
| 6106 | RetTy: NewVecTy, ID: Intrinsic::vp_load, |
| 6107 | Args: {II->getMemoryPointerParam(), NewMask, NewEVL}); |
| 6108 | // Preserve the original alignment. |
| 6109 | NewVP->addParamAttrs( |
| 6110 | ArgNo: 0, B: AttrBuilder(II->getContext()).addAlignmentAttr(Align: OrigAlign)); |
| 6111 | replaceValue(Old&: *Cast, New&: *NewVP); |
| 6112 | return true; |
| 6113 | } |
| 6114 | /// Fold the following cases into a single byte-level bit-reverse operation |
| 6115 | /// and accepts bswap and bitreverse intrinsics: |
| 6116 | /// bswap(bitreverse(x)) <--> bitcast(bitreverse(bitcast(x))) |
| 6117 | /// bitreverse(bswap(x)) <--> bitcast(bitreverse(bitcast(x))) |
| 6118 | /// The direction of the fold is cost-model driven. |
| 6119 | bool VectorCombine::foldBitOrderReverseAndSwap(Instruction &I) { |
| 6120 | Value *X; |
| 6121 | |
| 6122 | if (match(V: &I, P: m_BitCast(Op: m_BitReverse(Op0: m_BitCast(Op: m_Value(V&: X)))))) { |
| 6123 | Type *Ty = X->getType(); |
| 6124 | Type *VecTy = I.getOperand(i: 0)->getType(); |
| 6125 | if (Ty->isIntegerTy() && Ty == I.getType() && isa<FixedVectorType>(Val: VecTy) && |
| 6126 | cast<FixedVectorType>(Val: VecTy)->getElementType()->isIntegerTy(BitWidth: 8) && |
| 6127 | Ty->getIntegerBitWidth() % 16 == 0) { |
| 6128 | auto *InnerCall = dyn_cast<Instruction>(Val: I.getOperand(i: 0)); |
| 6129 | if (!InnerCall) |
| 6130 | return false; |
| 6131 | auto *InnerBitCast = dyn_cast<BitCastInst>(Val: InnerCall->getOperand(i: 0)); |
| 6132 | if (!InnerBitCast) |
| 6133 | return false; |
| 6134 | InstructionCost OldCost = TTI.getInstructionCost(U: InnerBitCast, CostKind) + |
| 6135 | TTI.getInstructionCost(U: InnerCall, CostKind) + |
| 6136 | TTI.getInstructionCost(U: &I, CostKind); |
| 6137 | IntrinsicCostAttributes ICABSwap(Intrinsic::bswap, Ty, {Ty}); |
| 6138 | IntrinsicCostAttributes ICABRev(Intrinsic::bitreverse, Ty, {Ty}); |
| 6139 | InstructionCost NewCost = TTI.getIntrinsicInstrCost(ICA: ICABSwap, CostKind) + |
| 6140 | TTI.getIntrinsicInstrCost(ICA: ICABRev, CostKind); |
| 6141 | if (!InnerCall->hasOneUse()) |
| 6142 | NewCost += TTI.getInstructionCost(U: InnerCall, CostKind) + |
| 6143 | TTI.getInstructionCost(U: InnerBitCast, CostKind); |
| 6144 | else if (!InnerBitCast->hasOneUse()) |
| 6145 | NewCost += TTI.getInstructionCost(U: InnerBitCast, CostKind); |
| 6146 | LLVM_DEBUG(dbgs() << "Found bitreverse vector roundtrip: " << I |
| 6147 | << "\n OldCost: " << OldCost |
| 6148 | << " vs NewCost: " << NewCost << "\n" ); |
| 6149 | if (NewCost.isValid() && NewCost < OldCost) { |
| 6150 | Builder.SetInsertPoint(&I); |
| 6151 | Value *BSwap = Builder.CreateUnaryIntrinsic(ID: Intrinsic::bswap, Op: X); |
| 6152 | Worklist.pushValue(V: BSwap); |
| 6153 | Value *BRev = |
| 6154 | Builder.CreateUnaryIntrinsic(ID: Intrinsic::bitreverse, Op: BSwap); |
| 6155 | replaceValue(Old&: I, New&: *BRev); |
| 6156 | return true; |
| 6157 | } |
| 6158 | } |
| 6159 | } |
| 6160 | |
| 6161 | if (!match(V: &I, P: m_BitReverse(Op0: m_BSwap(Op0: m_Value(V&: X)))) && |
| 6162 | !match(V: &I, P: m_BSwap(Op0: m_BitReverse(Op0: m_Value(V&: X))))) |
| 6163 | return false; |
| 6164 | Type *Ty = I.getType(); |
| 6165 | Type *I8Ty = Builder.getInt8Ty(); |
| 6166 | TypeSize ElementSize = DL->getTypeStoreSize(Ty); |
| 6167 | ElementCount NewVecCnt = ElementCount::get(MinVal: ElementSize.getKnownMinValue(), |
| 6168 | Scalable: ElementSize.isScalable()); |
| 6169 | Type *NewVecTy = VectorType::get(ElementType: I8Ty, EC: NewVecCnt); |
| 6170 | auto *II = cast<IntrinsicInst>(Val: &I); |
| 6171 | auto *InnerII = cast<IntrinsicInst>(Val: II->getArgOperand(i: 0)); |
| 6172 | // OldCost = cost of bitreverse/bswap + cost of bswap/bitreverse |
| 6173 | InstructionCost OldCost = TTI.getInstructionCost(U: II, CostKind) + |
| 6174 | TTI.getInstructionCost(U: InnerII, CostKind); |
| 6175 | // NewCost = cost of bitcast to byte vector + |
| 6176 | // cost of bitreverse/bswap on byte vector + |
| 6177 | // cost of bitcast back to original type |
| 6178 | InstructionCost CastToVecCost = TTI.getCastInstrCost( |
| 6179 | Opcode: Instruction::BitCast, Dst: NewVecTy, Src: Ty, CCH: TTI::CastContextHint::None, CostKind); |
| 6180 | InstructionCost CastToOrigCost = TTI.getCastInstrCost( |
| 6181 | Opcode: Instruction::BitCast, Dst: Ty, Src: NewVecTy, CCH: TTI::CastContextHint::None, CostKind); |
| 6182 | IntrinsicCostAttributes ICANew(Intrinsic::bitreverse, NewVecTy, {NewVecTy}); |
| 6183 | InstructionCost NewIntrinsicCost = |
| 6184 | TTI.getIntrinsicInstrCost(ICA: ICANew, CostKind); |
| 6185 | InstructionCost NewCost = CastToVecCost + NewIntrinsicCost + CastToOrigCost; |
| 6186 | if (!InnerII->hasOneUse()) |
| 6187 | NewCost += TTI.getInstructionCost(U: InnerII, CostKind); |
| 6188 | LLVM_DEBUG(dbgs() << "Found bitorder reverse and swap: " << I |
| 6189 | << "\n OldCost: " << OldCost << " vs NewCost: " << NewCost |
| 6190 | << "\n" ); |
| 6191 | if (!NewCost.isValid() || NewCost >= OldCost) |
| 6192 | return false; |
| 6193 | // Perform transform: bitcast(arg, <N x i8>), bitreverse, bitcast back |
| 6194 | Builder.SetInsertPoint(II); |
| 6195 | Value *CastToVec = Builder.CreateBitCast(V: X, DestTy: NewVecTy); |
| 6196 | Value *NewCall = |
| 6197 | Builder.CreateUnaryIntrinsic(ID: Intrinsic::bitreverse, Op: CastToVec); |
| 6198 | Value *CastToOrig = Builder.CreateBitCast(V: NewCall, DestTy: Ty); |
| 6199 | replaceValue(Old&: I, New&: *CastToOrig); |
| 6200 | return true; |
| 6201 | } |
| 6202 | // Attempt to shrink loads that are only used by shufflevector instructions. |
| 6203 | bool VectorCombine::shrinkLoadForShuffles(Instruction &I) { |
| 6204 | auto *OldLoad = dyn_cast<LoadInst>(Val: &I); |
| 6205 | if (!OldLoad || !OldLoad->isSimple()) |
| 6206 | return false; |
| 6207 | |
| 6208 | auto *OldLoadTy = dyn_cast<FixedVectorType>(Val: OldLoad->getType()); |
| 6209 | if (!OldLoadTy) |
| 6210 | return false; |
| 6211 | |
| 6212 | unsigned const OldNumElements = OldLoadTy->getNumElements(); |
| 6213 | |
| 6214 | // Search all uses of load. If all uses are shufflevector instructions, and |
| 6215 | // the second operands are all poison values, find the minimum and maximum |
| 6216 | // indices of the vector elements referenced by all shuffle masks. |
| 6217 | // Otherwise return `std::nullopt`. |
| 6218 | using IndexRange = std::pair<int, int>; |
| 6219 | auto GetIndexRangeInShuffles = [&]() -> std::optional<IndexRange> { |
| 6220 | IndexRange OutputRange = IndexRange(OldNumElements, -1); |
| 6221 | for (llvm::Use &Use : I.uses()) { |
| 6222 | // Ensure all uses match the required pattern. |
| 6223 | User *Shuffle = Use.getUser(); |
| 6224 | ArrayRef<int> Mask; |
| 6225 | |
| 6226 | if (!match(V: Shuffle, |
| 6227 | P: m_Shuffle(v1: m_Specific(V: OldLoad), v2: m_Undef(), mask: m_Mask(Mask)))) |
| 6228 | return std::nullopt; |
| 6229 | |
| 6230 | // Ignore shufflevector instructions that have no uses. |
| 6231 | if (Shuffle->use_empty()) |
| 6232 | continue; |
| 6233 | |
| 6234 | // Find the min and max indices used by the shufflevector instruction. |
| 6235 | for (int Index : Mask) { |
| 6236 | if (Index >= 0 && Index < static_cast<int>(OldNumElements)) { |
| 6237 | OutputRange.first = std::min(a: Index, b: OutputRange.first); |
| 6238 | OutputRange.second = std::max(a: Index, b: OutputRange.second); |
| 6239 | } |
| 6240 | } |
| 6241 | } |
| 6242 | |
| 6243 | if (OutputRange.second < OutputRange.first) |
| 6244 | return std::nullopt; |
| 6245 | |
| 6246 | return OutputRange; |
| 6247 | }; |
| 6248 | |
| 6249 | // Get the range of vector elements used by shufflevector instructions. |
| 6250 | if (std::optional<IndexRange> Indices = GetIndexRangeInShuffles()) { |
| 6251 | unsigned const NewNumElements = Indices->second + 1u; |
| 6252 | |
| 6253 | // If the range of vector elements is smaller than the full load, attempt |
| 6254 | // to create a smaller load. |
| 6255 | if (NewNumElements < OldNumElements) { |
| 6256 | IRBuilder Builder(&I); |
| 6257 | Builder.SetCurrentDebugLocation(I.getDebugLoc()); |
| 6258 | |
| 6259 | // Calculate costs of old and new ops. |
| 6260 | Type *ElemTy = OldLoadTy->getElementType(); |
| 6261 | FixedVectorType *NewLoadTy = FixedVectorType::get(ElementType: ElemTy, NumElts: NewNumElements); |
| 6262 | Value *PtrOp = OldLoad->getPointerOperand(); |
| 6263 | |
| 6264 | InstructionCost OldCost = TTI.getMemoryOpCost( |
| 6265 | Opcode: Instruction::Load, Src: OldLoad->getType(), Alignment: OldLoad->getAlign(), |
| 6266 | AddressSpace: OldLoad->getPointerAddressSpace(), CostKind); |
| 6267 | InstructionCost NewCost = |
| 6268 | TTI.getMemoryOpCost(Opcode: Instruction::Load, Src: NewLoadTy, Alignment: OldLoad->getAlign(), |
| 6269 | AddressSpace: OldLoad->getPointerAddressSpace(), CostKind); |
| 6270 | |
| 6271 | using UseEntry = std::pair<ShuffleVectorInst *, std::vector<int>>; |
| 6272 | SmallVector<UseEntry, 4u> NewUses; |
| 6273 | unsigned const MaxIndex = NewNumElements * 2u; |
| 6274 | |
| 6275 | for (llvm::Use &Use : I.uses()) { |
| 6276 | auto *Shuffle = cast<ShuffleVectorInst>(Val: Use.getUser()); |
| 6277 | |
| 6278 | // Ignore shufflevector instructions that have no uses. |
| 6279 | if (Shuffle->use_empty()) |
| 6280 | continue; |
| 6281 | |
| 6282 | ArrayRef<int> OldMask = Shuffle->getShuffleMask(); |
| 6283 | |
| 6284 | // Create entry for new use. |
| 6285 | NewUses.push_back(Elt: {Shuffle, OldMask}); |
| 6286 | |
| 6287 | // Validate mask indices. |
| 6288 | for (int Index : OldMask) { |
| 6289 | if (Index >= static_cast<int>(MaxIndex)) |
| 6290 | return false; |
| 6291 | } |
| 6292 | |
| 6293 | // Update costs. |
| 6294 | OldCost += |
| 6295 | TTI.getShuffleCost(Kind: TTI::SK_PermuteSingleSrc, DstTy: Shuffle->getType(), |
| 6296 | SrcTy: OldLoadTy, Mask: OldMask, CostKind); |
| 6297 | NewCost += |
| 6298 | TTI.getShuffleCost(Kind: TTI::SK_PermuteSingleSrc, DstTy: Shuffle->getType(), |
| 6299 | SrcTy: NewLoadTy, Mask: OldMask, CostKind); |
| 6300 | } |
| 6301 | |
| 6302 | LLVM_DEBUG( |
| 6303 | dbgs() << "Found a load used only by shufflevector instructions: " |
| 6304 | << I << "\n OldCost: " << OldCost |
| 6305 | << " vs NewCost: " << NewCost << "\n" ); |
| 6306 | |
| 6307 | if (OldCost < NewCost || !NewCost.isValid()) |
| 6308 | return false; |
| 6309 | |
| 6310 | // Create new load of smaller vector. |
| 6311 | auto *NewLoad = cast<LoadInst>( |
| 6312 | Val: Builder.CreateAlignedLoad(Ty: NewLoadTy, Ptr: PtrOp, Align: OldLoad->getAlign())); |
| 6313 | NewLoad->copyMetadata(SrcInst: I); |
| 6314 | |
| 6315 | // Replace all uses. |
| 6316 | for (UseEntry &Use : NewUses) { |
| 6317 | ShuffleVectorInst *Shuffle = Use.first; |
| 6318 | std::vector<int> &NewMask = Use.second; |
| 6319 | |
| 6320 | Builder.SetInsertPoint(Shuffle); |
| 6321 | Builder.SetCurrentDebugLocation(Shuffle->getDebugLoc()); |
| 6322 | Value *NewShuffle = Builder.CreateShuffleVector( |
| 6323 | V1: NewLoad, V2: PoisonValue::get(T: NewLoadTy), Mask: NewMask); |
| 6324 | |
| 6325 | replaceValue(Old&: *Shuffle, New&: *NewShuffle, Erase: false); |
| 6326 | } |
| 6327 | |
| 6328 | return true; |
| 6329 | } |
| 6330 | } |
| 6331 | return false; |
| 6332 | } |
| 6333 | |
| 6334 | // Attempt to narrow a phi of shufflevector instructions where the two incoming |
| 6335 | // values have the same operands but different masks. If the two shuffle masks |
| 6336 | // are offsets of one another we can use one branch to rotate the incoming |
| 6337 | // vector and perform one larger shuffle after the phi. |
| 6338 | bool VectorCombine::shrinkPhiOfShuffles(Instruction &I) { |
| 6339 | auto *Phi = dyn_cast<PHINode>(Val: &I); |
| 6340 | if (!Phi || Phi->getNumIncomingValues() != 2u) |
| 6341 | return false; |
| 6342 | |
| 6343 | Value *Op = nullptr; |
| 6344 | ArrayRef<int> Mask0; |
| 6345 | ArrayRef<int> Mask1; |
| 6346 | |
| 6347 | if (!match(V: Phi->getOperand(i_nocapture: 0u), |
| 6348 | P: m_OneUse(SubPattern: m_Shuffle(v1: m_Value(V&: Op), v2: m_Poison(), mask: m_Mask(Mask0)))) || |
| 6349 | !match(V: Phi->getOperand(i_nocapture: 1u), |
| 6350 | P: m_OneUse(SubPattern: m_Shuffle(v1: m_Specific(V: Op), v2: m_Poison(), mask: m_Mask(Mask1))))) |
| 6351 | return false; |
| 6352 | |
| 6353 | auto *Shuf = cast<ShuffleVectorInst>(Val: Phi->getOperand(i_nocapture: 0u)); |
| 6354 | |
| 6355 | // Ensure result vectors are wider than the argument vector. |
| 6356 | auto *InputVT = cast<FixedVectorType>(Val: Op->getType()); |
| 6357 | auto *ResultVT = cast<FixedVectorType>(Val: Shuf->getType()); |
| 6358 | auto const InputNumElements = InputVT->getNumElements(); |
| 6359 | |
| 6360 | if (InputNumElements >= ResultVT->getNumElements()) |
| 6361 | return false; |
| 6362 | |
| 6363 | // Take the difference of the two shuffle masks at each index. Ignore poison |
| 6364 | // values at the same index in both masks. |
| 6365 | SmallVector<int, 16> NewMask; |
| 6366 | NewMask.reserve(N: Mask0.size()); |
| 6367 | |
| 6368 | for (auto [M0, M1] : zip(t&: Mask0, u&: Mask1)) { |
| 6369 | if (M0 >= 0 && M1 >= 0) |
| 6370 | NewMask.push_back(Elt: M0 - M1); |
| 6371 | else if (M0 == -1 && M1 == -1) |
| 6372 | continue; |
| 6373 | else |
| 6374 | return false; |
| 6375 | } |
| 6376 | |
| 6377 | // Ensure all elements of the new mask are equal. If the difference between |
| 6378 | // the incoming mask elements is the same, the two must be constant offsets |
| 6379 | // of one another. |
| 6380 | if (NewMask.empty() || !all_equal(Range&: NewMask)) |
| 6381 | return false; |
| 6382 | |
| 6383 | // Create new mask using difference of the two incoming masks. |
| 6384 | int MaskOffset = NewMask[0u]; |
| 6385 | unsigned Index = (InputNumElements + MaskOffset) % InputNumElements; |
| 6386 | NewMask.clear(); |
| 6387 | |
| 6388 | for (unsigned I = 0u; I < InputNumElements; ++I) { |
| 6389 | NewMask.push_back(Elt: Index); |
| 6390 | Index = (Index + 1u) % InputNumElements; |
| 6391 | } |
| 6392 | |
| 6393 | // Calculate costs for worst cases and compare. |
| 6394 | auto const Kind = TTI::SK_PermuteSingleSrc; |
| 6395 | auto OldCost = |
| 6396 | std::max(a: TTI.getShuffleCost(Kind, DstTy: ResultVT, SrcTy: InputVT, Mask: Mask0, CostKind), |
| 6397 | b: TTI.getShuffleCost(Kind, DstTy: ResultVT, SrcTy: InputVT, Mask: Mask1, CostKind)); |
| 6398 | auto NewCost = TTI.getShuffleCost(Kind, DstTy: InputVT, SrcTy: InputVT, Mask: NewMask, CostKind) + |
| 6399 | TTI.getShuffleCost(Kind, DstTy: ResultVT, SrcTy: InputVT, Mask: Mask1, CostKind); |
| 6400 | |
| 6401 | LLVM_DEBUG(dbgs() << "Found a phi of mergeable shuffles: " << I |
| 6402 | << "\n OldCost: " << OldCost << " vs NewCost: " << NewCost |
| 6403 | << "\n" ); |
| 6404 | |
| 6405 | if (NewCost > OldCost) |
| 6406 | return false; |
| 6407 | |
| 6408 | // Create new shuffles and narrowed phi. |
| 6409 | auto Builder = IRBuilder(Shuf); |
| 6410 | Builder.SetCurrentDebugLocation(Shuf->getDebugLoc()); |
| 6411 | auto *PoisonVal = PoisonValue::get(T: InputVT); |
| 6412 | auto *NewShuf0 = Builder.CreateShuffleVector(V1: Op, V2: PoisonVal, Mask: NewMask); |
| 6413 | Worklist.push(I: cast<Instruction>(Val: NewShuf0)); |
| 6414 | |
| 6415 | Builder.SetInsertPoint(Phi); |
| 6416 | Builder.SetCurrentDebugLocation(Phi->getDebugLoc()); |
| 6417 | auto *NewPhi = Builder.CreatePHI(Ty: NewShuf0->getType(), NumReservedValues: 2u); |
| 6418 | NewPhi->addIncoming(V: NewShuf0, BB: Phi->getIncomingBlock(i: 0u)); |
| 6419 | NewPhi->addIncoming(V: Op, BB: Phi->getIncomingBlock(i: 1u)); |
| 6420 | |
| 6421 | Builder.SetInsertPoint(*NewPhi->getInsertionPointAfterDef()); |
| 6422 | PoisonVal = PoisonValue::get(T: NewPhi->getType()); |
| 6423 | auto *NewShuf1 = Builder.CreateShuffleVector(V1: NewPhi, V2: PoisonVal, Mask: Mask1); |
| 6424 | |
| 6425 | replaceValue(Old&: *Phi, New&: *NewShuf1); |
| 6426 | return true; |
| 6427 | } |
| 6428 | |
| 6429 | /// This is the entry point for all transforms. Pass manager differences are |
| 6430 | /// handled in the callers of this function. |
| 6431 | bool VectorCombine::run() { |
| 6432 | if (DisableVectorCombine) |
| 6433 | return false; |
| 6434 | |
| 6435 | // Don't attempt vectorization if the target does not support vectors. |
| 6436 | if (!TTI.getNumberOfRegisters(ClassID: TTI.getRegisterClassForType(/*Vector*/ true))) |
| 6437 | return false; |
| 6438 | |
| 6439 | LLVM_DEBUG(dbgs() << "\n\nVECTORCOMBINE on " << F.getName() << "\n" ); |
| 6440 | |
| 6441 | auto FoldInst = [this](Instruction &I) { |
| 6442 | Builder.SetInsertPoint(&I); |
| 6443 | bool IsVectorType = isa<VectorType>(Val: I.getType()); |
| 6444 | bool IsFixedVectorType = isa<FixedVectorType>(Val: I.getType()); |
| 6445 | auto Opcode = I.getOpcode(); |
| 6446 | |
| 6447 | LLVM_DEBUG(dbgs() << "VC: Visiting: " << I << '\n'); |
| 6448 | |
| 6449 | // These folds should be beneficial regardless of when this pass is run |
| 6450 | // in the optimization pipeline. |
| 6451 | // The type checking is for run-time efficiency. We can avoid wasting time |
| 6452 | // dispatching to folding functions if there's no chance of matching. |
| 6453 | if (IsFixedVectorType) { |
| 6454 | switch (Opcode) { |
| 6455 | case Instruction::InsertElement: |
| 6456 | if (vectorizeLoadInsert(I)) |
| 6457 | return true; |
| 6458 | break; |
| 6459 | case Instruction::ShuffleVector: |
| 6460 | if (widenSubvectorLoad(I)) |
| 6461 | return true; |
| 6462 | break; |
| 6463 | default: |
| 6464 | break; |
| 6465 | } |
| 6466 | } |
| 6467 | |
| 6468 | // This transform works with scalable and fixed vectors |
| 6469 | // TODO: Identify and allow other scalable transforms |
| 6470 | if (IsVectorType) { |
| 6471 | if (scalarizeOpOrCmp(I)) |
| 6472 | return true; |
| 6473 | if (scalarizeLoad(I)) |
| 6474 | return true; |
| 6475 | if (scalarizeExtExtract(I)) |
| 6476 | return true; |
| 6477 | if (scalarizeVPIntrinsic(I)) |
| 6478 | return true; |
| 6479 | if (foldInterleaveIntrinsics(I)) |
| 6480 | return true; |
| 6481 | if (foldBitcastOfVPLoad(I)) |
| 6482 | return true; |
| 6483 | } |
| 6484 | |
| 6485 | if (foldDeinterleaveIntrinsics(I)) |
| 6486 | return true; |
| 6487 | |
| 6488 | if (Opcode == Instruction::Store) |
| 6489 | if (foldSingleElementStore(I)) |
| 6490 | return true; |
| 6491 | |
| 6492 | // If this is an early pipeline invocation of this pass, we are done. |
| 6493 | if (TryEarlyFoldsOnly) |
| 6494 | return false; |
| 6495 | |
| 6496 | if (Opcode == Instruction::Call) |
| 6497 | if (foldBitOrderReverseAndSwap(I)) |
| 6498 | return true; |
| 6499 | if (Opcode == Instruction::BitCast) |
| 6500 | if (foldBitOrderReverseAndSwap(I)) |
| 6501 | return true; |
| 6502 | |
| 6503 | // Otherwise, try folds that improve codegen but may interfere with |
| 6504 | // early IR canonicalizations. |
| 6505 | // The type checking is for run-time efficiency. We can avoid wasting time |
| 6506 | // dispatching to folding functions if there's no chance of matching. |
| 6507 | if (IsFixedVectorType) { |
| 6508 | switch (Opcode) { |
| 6509 | case Instruction::InsertElement: |
| 6510 | if (foldInsExtFNeg(I)) |
| 6511 | return true; |
| 6512 | if (foldInsExtBinop(I)) |
| 6513 | return true; |
| 6514 | if (foldInsExtVectorToShuffle(I)) |
| 6515 | return true; |
| 6516 | break; |
| 6517 | case Instruction::ShuffleVector: |
| 6518 | if (foldPermuteOfBinops(I)) |
| 6519 | return true; |
| 6520 | if (foldShuffleOfBinops(I)) |
| 6521 | return true; |
| 6522 | if (foldShuffleOfSelects(I)) |
| 6523 | return true; |
| 6524 | if (foldShuffleOfCastops(I)) |
| 6525 | return true; |
| 6526 | if (foldShuffleOfShuffles(I)) |
| 6527 | return true; |
| 6528 | if (foldPermuteOfIntrinsic(I)) |
| 6529 | return true; |
| 6530 | if (foldShufflesOfLengthChangingShuffles(I)) |
| 6531 | return true; |
| 6532 | if (foldShuffleOfIntrinsics(I)) |
| 6533 | return true; |
| 6534 | if (foldSelectShuffle(I)) |
| 6535 | return true; |
| 6536 | if (foldShuffleToIdentity(I)) |
| 6537 | return true; |
| 6538 | break; |
| 6539 | case Instruction::Load: |
| 6540 | if (shrinkLoadForShuffles(I)) |
| 6541 | return true; |
| 6542 | break; |
| 6543 | case Instruction::BitCast: |
| 6544 | if (foldBitcastShuffle(I)) |
| 6545 | return true; |
| 6546 | if (foldSelectsFromBitcast(I)) |
| 6547 | return true; |
| 6548 | break; |
| 6549 | case Instruction::And: |
| 6550 | case Instruction::Or: |
| 6551 | case Instruction::Xor: |
| 6552 | if (foldBitOpOfCastops(I)) |
| 6553 | return true; |
| 6554 | if (foldBitOpOfCastConstant(I)) |
| 6555 | return true; |
| 6556 | break; |
| 6557 | case Instruction::PHI: |
| 6558 | if (shrinkPhiOfShuffles(I)) |
| 6559 | return true; |
| 6560 | break; |
| 6561 | default: |
| 6562 | if (shrinkType(I)) |
| 6563 | return true; |
| 6564 | break; |
| 6565 | } |
| 6566 | } else { |
| 6567 | switch (Opcode) { |
| 6568 | case Instruction::Call: |
| 6569 | if (foldShuffleFromReductions(I)) |
| 6570 | return true; |
| 6571 | if (foldCastFromReductions(I)) |
| 6572 | return true; |
| 6573 | break; |
| 6574 | case Instruction::ExtractElement: |
| 6575 | if (foldShuffleChainsToReduce(I)) |
| 6576 | return true; |
| 6577 | break; |
| 6578 | case Instruction::ICmp: |
| 6579 | if (foldSignBitReductionCmp(I)) |
| 6580 | return true; |
| 6581 | if (foldICmpEqZeroVectorReduce(I)) |
| 6582 | return true; |
| 6583 | if (foldReductionZeroTest(I)) |
| 6584 | return true; |
| 6585 | if (foldEquivalentReductionCmp(I)) |
| 6586 | return true; |
| 6587 | if (foldReduceAddCmpZero(I)) |
| 6588 | return true; |
| 6589 | [[fallthrough]]; |
| 6590 | case Instruction::FCmp: |
| 6591 | if (foldExtractExtract(I)) |
| 6592 | return true; |
| 6593 | break; |
| 6594 | case Instruction::Or: |
| 6595 | if (foldConcatOfBoolMasks(I)) |
| 6596 | return true; |
| 6597 | [[fallthrough]]; |
| 6598 | default: |
| 6599 | if (Instruction::isBinaryOp(Opcode)) { |
| 6600 | if (foldExtractExtract(I)) |
| 6601 | return true; |
| 6602 | if (foldExtractedCmps(I)) |
| 6603 | return true; |
| 6604 | if (foldBinopOfReductions(I)) |
| 6605 | return true; |
| 6606 | } |
| 6607 | break; |
| 6608 | } |
| 6609 | } |
| 6610 | return false; |
| 6611 | }; |
| 6612 | |
| 6613 | bool MadeChange = false; |
| 6614 | for (BasicBlock &BB : F) { |
| 6615 | // Ignore unreachable basic blocks. |
| 6616 | if (!DT.isReachableFromEntry(A: &BB)) |
| 6617 | continue; |
| 6618 | // Use early increment range so that we can erase instructions in loop. |
| 6619 | // make_early_inc_range is not applicable here, as the next iterator may |
| 6620 | // be invalidated by RecursivelyDeleteTriviallyDeadInstructions. |
| 6621 | // We manually maintain the next instruction and update it when it is about |
| 6622 | // to be deleted. |
| 6623 | Instruction *I = &BB.front(); |
| 6624 | while (I) { |
| 6625 | NextInst = I->getNextNode(); |
| 6626 | if (!I->isDebugOrPseudoInst()) |
| 6627 | MadeChange |= FoldInst(*I); |
| 6628 | I = NextInst; |
| 6629 | } |
| 6630 | } |
| 6631 | |
| 6632 | NextInst = nullptr; |
| 6633 | |
| 6634 | while (!Worklist.isEmpty()) { |
| 6635 | Instruction *I = Worklist.removeOne(); |
| 6636 | if (!I) |
| 6637 | continue; |
| 6638 | |
| 6639 | if (isInstructionTriviallyDead(I)) { |
| 6640 | eraseInstruction(I&: *I); |
| 6641 | continue; |
| 6642 | } |
| 6643 | |
| 6644 | MadeChange |= FoldInst(*I); |
| 6645 | } |
| 6646 | |
| 6647 | return MadeChange; |
| 6648 | } |
| 6649 | |
| 6650 | PreservedAnalyses VectorCombinePass::run(Function &F, |
| 6651 | FunctionAnalysisManager &FAM) { |
| 6652 | auto &AC = FAM.getResult<AssumptionAnalysis>(IR&: F); |
| 6653 | TargetTransformInfo &TTI = FAM.getResult<TargetIRAnalysis>(IR&: F); |
| 6654 | DominatorTree &DT = FAM.getResult<DominatorTreeAnalysis>(IR&: F); |
| 6655 | AAResults &AA = FAM.getResult<AAManager>(IR&: F); |
| 6656 | const DataLayout *DL = &F.getDataLayout(); |
| 6657 | TTI::TargetCostKind CostKind = |
| 6658 | F.hasOptSize() ? TTI::TCK_CodeSize : TTI::TCK_RecipThroughput; |
| 6659 | VectorCombine Combiner(F, TTI, DT, AA, AC, DL, CostKind, TryEarlyFoldsOnly); |
| 6660 | if (!Combiner.run()) |
| 6661 | return PreservedAnalyses::all(); |
| 6662 | PreservedAnalyses PA; |
| 6663 | PA.preserveSet<CFGAnalyses>(); |
| 6664 | return PA; |
| 6665 | } |
| 6666 | |