1 //===- LoadStoreVectorizer.cpp - GPU Load & Store Vectorizer --------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass merges loads/stores to/from sequential memory addresses into vector
10 // loads/stores. Although there's nothing GPU-specific in here, this pass is
11 // motivated by the microarchitectural quirks of nVidia and AMD GPUs.
12 //
13 // (For simplicity below we talk about loads only, but everything also applies
14 // to stores.)
15 //
16 // This pass is intended to be run late in the pipeline, after other
17 // vectorization opportunities have been exploited. So the assumption here is
18 // that immediately following our new vector load we'll need to extract out the
19 // individual elements of the load, so we can operate on them individually.
20 //
21 // On CPUs this transformation is usually not beneficial, because extracting the
22 // elements of a vector register is expensive on most architectures. It's
23 // usually better just to load each element individually into its own scalar
24 // register.
25 //
26 // However, nVidia and AMD GPUs don't have proper vector registers. Instead, a
27 // "vector load" loads directly into a series of scalar registers. In effect,
28 // extracting the elements of the vector is free. It's therefore always
29 // beneficial to vectorize a sequence of loads on these architectures.
30 //
31 // Vectorizing (perhaps a better name might be "coalescing") loads can have
32 // large performance impacts on GPU kernels, and opportunities for vectorizing
33 // are common in GPU code. This pass tries very hard to find such
34 // opportunities; its runtime is quadratic in the number of loads in a BB.
35 //
36 // Some CPU architectures, such as ARM, have instructions that load into
37 // multiple scalar registers, similar to a GPU vectorized load. In theory ARM
38 // could use this pass (with some modifications), but currently it implements
39 // its own pass to do something similar to what we do here.
40
41 #include "llvm/Transforms/Vectorize/LoadStoreVectorizer.h"
42 #include "llvm/ADT/APInt.h"
43 #include "llvm/ADT/ArrayRef.h"
44 #include "llvm/ADT/MapVector.h"
45 #include "llvm/ADT/PostOrderIterator.h"
46 #include "llvm/ADT/STLExtras.h"
47 #include "llvm/ADT/SmallPtrSet.h"
48 #include "llvm/ADT/SmallVector.h"
49 #include "llvm/ADT/Statistic.h"
50 #include "llvm/ADT/iterator_range.h"
51 #include "llvm/Analysis/AliasAnalysis.h"
52 #include "llvm/Analysis/AssumptionCache.h"
53 #include "llvm/Analysis/MemoryLocation.h"
54 #include "llvm/Analysis/ScalarEvolution.h"
55 #include "llvm/Analysis/TargetTransformInfo.h"
56 #include "llvm/Analysis/ValueTracking.h"
57 #include "llvm/Analysis/VectorUtils.h"
58 #include "llvm/IR/Attributes.h"
59 #include "llvm/IR/BasicBlock.h"
60 #include "llvm/IR/Constants.h"
61 #include "llvm/IR/DataLayout.h"
62 #include "llvm/IR/DerivedTypes.h"
63 #include "llvm/IR/Dominators.h"
64 #include "llvm/IR/Function.h"
65 #include "llvm/IR/GetElementPtrTypeIterator.h"
66 #include "llvm/IR/IRBuilder.h"
67 #include "llvm/IR/InstrTypes.h"
68 #include "llvm/IR/Instruction.h"
69 #include "llvm/IR/Instructions.h"
70 #include "llvm/IR/Module.h"
71 #include "llvm/IR/Type.h"
72 #include "llvm/IR/Value.h"
73 #include "llvm/InitializePasses.h"
74 #include "llvm/Pass.h"
75 #include "llvm/Support/Casting.h"
76 #include "llvm/Support/Debug.h"
77 #include "llvm/Support/KnownBits.h"
78 #include "llvm/Support/MathExtras.h"
79 #include "llvm/Support/raw_ostream.h"
80 #include "llvm/Transforms/Utils/Local.h"
81 #include "llvm/Transforms/Vectorize.h"
82 #include <algorithm>
83 #include <cassert>
84 #include <cstdlib>
85 #include <tuple>
86 #include <utility>
87
88 using namespace llvm;
89
90 #define DEBUG_TYPE "load-store-vectorizer"
91
92 STATISTIC(NumVectorInstructions, "Number of vector accesses generated");
93 STATISTIC(NumScalarsVectorized, "Number of scalar accesses vectorized");
94
95 // FIXME: Assuming stack alignment of 4 is always good enough
96 static const unsigned StackAdjustedAlignment = 4;
97
98 namespace {
99
100 /// ChainID is an arbitrary token that is allowed to be different only for the
101 /// accesses that are guaranteed to be considered non-consecutive by
102 /// Vectorizer::isConsecutiveAccess. It's used for grouping instructions
103 /// together and reducing the number of instructions the main search operates on
104 /// at a time, i.e. this is to reduce compile time and nothing else as the main
105 /// search has O(n^2) time complexity. The underlying type of ChainID should not
106 /// be relied upon.
107 using ChainID = const Value *;
108 using InstrList = SmallVector<Instruction *, 8>;
109 using InstrListMap = MapVector<ChainID, InstrList>;
110
111 class Vectorizer {
112 Function &F;
113 AliasAnalysis &AA;
114 AssumptionCache &AC;
115 DominatorTree &DT;
116 ScalarEvolution &SE;
117 TargetTransformInfo &TTI;
118 const DataLayout &DL;
119 IRBuilder<> Builder;
120
121 public:
Vectorizer(Function & F,AliasAnalysis & AA,AssumptionCache & AC,DominatorTree & DT,ScalarEvolution & SE,TargetTransformInfo & TTI)122 Vectorizer(Function &F, AliasAnalysis &AA, AssumptionCache &AC,
123 DominatorTree &DT, ScalarEvolution &SE, TargetTransformInfo &TTI)
124 : F(F), AA(AA), AC(AC), DT(DT), SE(SE), TTI(TTI),
125 DL(F.getParent()->getDataLayout()), Builder(SE.getContext()) {}
126
127 bool run();
128
129 private:
130 unsigned getPointerAddressSpace(Value *I);
131
132 static const unsigned MaxDepth = 3;
133
134 bool isConsecutiveAccess(Value *A, Value *B);
135 bool areConsecutivePointers(Value *PtrA, Value *PtrB, APInt PtrDelta,
136 unsigned Depth = 0) const;
137 bool lookThroughComplexAddresses(Value *PtrA, Value *PtrB, APInt PtrDelta,
138 unsigned Depth) const;
139 bool lookThroughSelects(Value *PtrA, Value *PtrB, const APInt &PtrDelta,
140 unsigned Depth) const;
141
142 /// After vectorization, reorder the instructions that I depends on
143 /// (the instructions defining its operands), to ensure they dominate I.
144 void reorder(Instruction *I);
145
146 /// Returns the first and the last instructions in Chain.
147 std::pair<BasicBlock::iterator, BasicBlock::iterator>
148 getBoundaryInstrs(ArrayRef<Instruction *> Chain);
149
150 /// Erases the original instructions after vectorizing.
151 void eraseInstructions(ArrayRef<Instruction *> Chain);
152
153 /// "Legalize" the vector type that would be produced by combining \p
154 /// ElementSizeBits elements in \p Chain. Break into two pieces such that the
155 /// total size of each piece is 1, 2 or a multiple of 4 bytes. \p Chain is
156 /// expected to have more than 4 elements.
157 std::pair<ArrayRef<Instruction *>, ArrayRef<Instruction *>>
158 splitOddVectorElts(ArrayRef<Instruction *> Chain, unsigned ElementSizeBits);
159
160 /// Finds the largest prefix of Chain that's vectorizable, checking for
161 /// intervening instructions which may affect the memory accessed by the
162 /// instructions within Chain.
163 ///
164 /// The elements of \p Chain must be all loads or all stores and must be in
165 /// address order.
166 ArrayRef<Instruction *> getVectorizablePrefix(ArrayRef<Instruction *> Chain);
167
168 /// Collects load and store instructions to vectorize.
169 std::pair<InstrListMap, InstrListMap> collectInstructions(BasicBlock *BB);
170
171 /// Processes the collected instructions, the \p Map. The values of \p Map
172 /// should be all loads or all stores.
173 bool vectorizeChains(InstrListMap &Map);
174
175 /// Finds the load/stores to consecutive memory addresses and vectorizes them.
176 bool vectorizeInstructions(ArrayRef<Instruction *> Instrs);
177
178 /// Vectorizes the load instructions in Chain.
179 bool
180 vectorizeLoadChain(ArrayRef<Instruction *> Chain,
181 SmallPtrSet<Instruction *, 16> *InstructionsProcessed);
182
183 /// Vectorizes the store instructions in Chain.
184 bool
185 vectorizeStoreChain(ArrayRef<Instruction *> Chain,
186 SmallPtrSet<Instruction *, 16> *InstructionsProcessed);
187
188 /// Check if this load/store access is misaligned accesses.
189 bool accessIsMisaligned(unsigned SzInBytes, unsigned AddressSpace,
190 Align Alignment);
191 };
192
193 class LoadStoreVectorizerLegacyPass : public FunctionPass {
194 public:
195 static char ID;
196
LoadStoreVectorizerLegacyPass()197 LoadStoreVectorizerLegacyPass() : FunctionPass(ID) {
198 initializeLoadStoreVectorizerLegacyPassPass(*PassRegistry::getPassRegistry());
199 }
200
201 bool runOnFunction(Function &F) override;
202
getPassName() const203 StringRef getPassName() const override {
204 return "GPU Load and Store Vectorizer";
205 }
206
getAnalysisUsage(AnalysisUsage & AU) const207 void getAnalysisUsage(AnalysisUsage &AU) const override {
208 AU.addRequired<AAResultsWrapperPass>();
209 AU.addRequired<AssumptionCacheTracker>();
210 AU.addRequired<ScalarEvolutionWrapperPass>();
211 AU.addRequired<DominatorTreeWrapperPass>();
212 AU.addRequired<TargetTransformInfoWrapperPass>();
213 AU.setPreservesCFG();
214 }
215 };
216
217 } // end anonymous namespace
218
219 char LoadStoreVectorizerLegacyPass::ID = 0;
220
221 INITIALIZE_PASS_BEGIN(LoadStoreVectorizerLegacyPass, DEBUG_TYPE,
222 "Vectorize load and Store instructions", false, false)
223 INITIALIZE_PASS_DEPENDENCY(SCEVAAWrapperPass)
224 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker);
225 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)226 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
227 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
228 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
229 INITIALIZE_PASS_END(LoadStoreVectorizerLegacyPass, DEBUG_TYPE,
230 "Vectorize load and store instructions", false, false)
231
232 Pass *llvm::createLoadStoreVectorizerPass() {
233 return new LoadStoreVectorizerLegacyPass();
234 }
235
runOnFunction(Function & F)236 bool LoadStoreVectorizerLegacyPass::runOnFunction(Function &F) {
237 // Don't vectorize when the attribute NoImplicitFloat is used.
238 if (skipFunction(F) || F.hasFnAttribute(Attribute::NoImplicitFloat))
239 return false;
240
241 AliasAnalysis &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
242 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
243 ScalarEvolution &SE = getAnalysis<ScalarEvolutionWrapperPass>().getSE();
244 TargetTransformInfo &TTI =
245 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
246
247 AssumptionCache &AC =
248 getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
249
250 Vectorizer V(F, AA, AC, DT, SE, TTI);
251 return V.run();
252 }
253
run(Function & F,FunctionAnalysisManager & AM)254 PreservedAnalyses LoadStoreVectorizerPass::run(Function &F, FunctionAnalysisManager &AM) {
255 // Don't vectorize when the attribute NoImplicitFloat is used.
256 if (F.hasFnAttribute(Attribute::NoImplicitFloat))
257 return PreservedAnalyses::all();
258
259 AliasAnalysis &AA = AM.getResult<AAManager>(F);
260 DominatorTree &DT = AM.getResult<DominatorTreeAnalysis>(F);
261 ScalarEvolution &SE = AM.getResult<ScalarEvolutionAnalysis>(F);
262 TargetTransformInfo &TTI = AM.getResult<TargetIRAnalysis>(F);
263 AssumptionCache &AC = AM.getResult<AssumptionAnalysis>(F);
264
265 Vectorizer V(F, AA, AC, DT, SE, TTI);
266 bool Changed = V.run();
267 PreservedAnalyses PA;
268 PA.preserveSet<CFGAnalyses>();
269 return Changed ? PA : PreservedAnalyses::all();
270 }
271
272 // The real propagateMetadata expects a SmallVector<Value*>, but we deal in
273 // vectors of Instructions.
propagateMetadata(Instruction * I,ArrayRef<Instruction * > IL)274 static void propagateMetadata(Instruction *I, ArrayRef<Instruction *> IL) {
275 SmallVector<Value *, 8> VL(IL.begin(), IL.end());
276 propagateMetadata(I, VL);
277 }
278
279 // Vectorizer Implementation
run()280 bool Vectorizer::run() {
281 bool Changed = false;
282
283 // Scan the blocks in the function in post order.
284 for (BasicBlock *BB : post_order(&F)) {
285 InstrListMap LoadRefs, StoreRefs;
286 std::tie(LoadRefs, StoreRefs) = collectInstructions(BB);
287 Changed |= vectorizeChains(LoadRefs);
288 Changed |= vectorizeChains(StoreRefs);
289 }
290
291 return Changed;
292 }
293
getPointerAddressSpace(Value * I)294 unsigned Vectorizer::getPointerAddressSpace(Value *I) {
295 if (LoadInst *L = dyn_cast<LoadInst>(I))
296 return L->getPointerAddressSpace();
297 if (StoreInst *S = dyn_cast<StoreInst>(I))
298 return S->getPointerAddressSpace();
299 return -1;
300 }
301
302 // FIXME: Merge with llvm::isConsecutiveAccess
isConsecutiveAccess(Value * A,Value * B)303 bool Vectorizer::isConsecutiveAccess(Value *A, Value *B) {
304 Value *PtrA = getLoadStorePointerOperand(A);
305 Value *PtrB = getLoadStorePointerOperand(B);
306 unsigned ASA = getPointerAddressSpace(A);
307 unsigned ASB = getPointerAddressSpace(B);
308
309 // Check that the address spaces match and that the pointers are valid.
310 if (!PtrA || !PtrB || (ASA != ASB))
311 return false;
312
313 // Make sure that A and B are different pointers of the same size type.
314 Type *PtrATy = getLoadStoreType(A);
315 Type *PtrBTy = getLoadStoreType(B);
316 if (PtrA == PtrB ||
317 PtrATy->isVectorTy() != PtrBTy->isVectorTy() ||
318 DL.getTypeStoreSize(PtrATy) != DL.getTypeStoreSize(PtrBTy) ||
319 DL.getTypeStoreSize(PtrATy->getScalarType()) !=
320 DL.getTypeStoreSize(PtrBTy->getScalarType()))
321 return false;
322
323 unsigned PtrBitWidth = DL.getPointerSizeInBits(ASA);
324 APInt Size(PtrBitWidth, DL.getTypeStoreSize(PtrATy));
325
326 return areConsecutivePointers(PtrA, PtrB, Size);
327 }
328
areConsecutivePointers(Value * PtrA,Value * PtrB,APInt PtrDelta,unsigned Depth) const329 bool Vectorizer::areConsecutivePointers(Value *PtrA, Value *PtrB,
330 APInt PtrDelta, unsigned Depth) const {
331 unsigned PtrBitWidth = DL.getPointerTypeSizeInBits(PtrA->getType());
332 APInt OffsetA(PtrBitWidth, 0);
333 APInt OffsetB(PtrBitWidth, 0);
334 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA);
335 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetB);
336
337 unsigned NewPtrBitWidth = DL.getTypeStoreSizeInBits(PtrA->getType());
338
339 if (NewPtrBitWidth != DL.getTypeStoreSizeInBits(PtrB->getType()))
340 return false;
341
342 // In case if we have to shrink the pointer
343 // stripAndAccumulateInBoundsConstantOffsets should properly handle a
344 // possible overflow and the value should fit into a smallest data type
345 // used in the cast/gep chain.
346 assert(OffsetA.getMinSignedBits() <= NewPtrBitWidth &&
347 OffsetB.getMinSignedBits() <= NewPtrBitWidth);
348
349 OffsetA = OffsetA.sextOrTrunc(NewPtrBitWidth);
350 OffsetB = OffsetB.sextOrTrunc(NewPtrBitWidth);
351 PtrDelta = PtrDelta.sextOrTrunc(NewPtrBitWidth);
352
353 APInt OffsetDelta = OffsetB - OffsetA;
354
355 // Check if they are based on the same pointer. That makes the offsets
356 // sufficient.
357 if (PtrA == PtrB)
358 return OffsetDelta == PtrDelta;
359
360 // Compute the necessary base pointer delta to have the necessary final delta
361 // equal to the pointer delta requested.
362 APInt BaseDelta = PtrDelta - OffsetDelta;
363
364 // Compute the distance with SCEV between the base pointers.
365 const SCEV *PtrSCEVA = SE.getSCEV(PtrA);
366 const SCEV *PtrSCEVB = SE.getSCEV(PtrB);
367 const SCEV *C = SE.getConstant(BaseDelta);
368 const SCEV *X = SE.getAddExpr(PtrSCEVA, C);
369 if (X == PtrSCEVB)
370 return true;
371
372 // The above check will not catch the cases where one of the pointers is
373 // factorized but the other one is not, such as (C + (S * (A + B))) vs
374 // (AS + BS). Get the minus scev. That will allow re-combining the expresions
375 // and getting the simplified difference.
376 const SCEV *Dist = SE.getMinusSCEV(PtrSCEVB, PtrSCEVA);
377 if (C == Dist)
378 return true;
379
380 // Sometimes even this doesn't work, because SCEV can't always see through
381 // patterns that look like (gep (ext (add (shl X, C1), C2))). Try checking
382 // things the hard way.
383 return lookThroughComplexAddresses(PtrA, PtrB, BaseDelta, Depth);
384 }
385
checkNoWrapFlags(Instruction * I,bool Signed)386 static bool checkNoWrapFlags(Instruction *I, bool Signed) {
387 BinaryOperator *BinOpI = cast<BinaryOperator>(I);
388 return (Signed && BinOpI->hasNoSignedWrap()) ||
389 (!Signed && BinOpI->hasNoUnsignedWrap());
390 }
391
checkIfSafeAddSequence(const APInt & IdxDiff,Instruction * AddOpA,unsigned MatchingOpIdxA,Instruction * AddOpB,unsigned MatchingOpIdxB,bool Signed)392 static bool checkIfSafeAddSequence(const APInt &IdxDiff, Instruction *AddOpA,
393 unsigned MatchingOpIdxA, Instruction *AddOpB,
394 unsigned MatchingOpIdxB, bool Signed) {
395 // If both OpA and OpB is an add with NSW/NUW and with
396 // one of the operands being the same, we can guarantee that the
397 // transformation is safe if we can prove that OpA won't overflow when
398 // IdxDiff added to the other operand of OpA.
399 // For example:
400 // %tmp7 = add nsw i32 %tmp2, %v0
401 // %tmp8 = sext i32 %tmp7 to i64
402 // ...
403 // %tmp11 = add nsw i32 %v0, 1
404 // %tmp12 = add nsw i32 %tmp2, %tmp11
405 // %tmp13 = sext i32 %tmp12 to i64
406 //
407 // Both %tmp7 and %tmp2 has the nsw flag and the first operand
408 // is %tmp2. It's guaranteed that adding 1 to %tmp7 won't overflow
409 // because %tmp11 adds 1 to %v0 and both %tmp11 and %tmp12 has the
410 // nsw flag.
411 assert(AddOpA->getOpcode() == Instruction::Add &&
412 AddOpB->getOpcode() == Instruction::Add &&
413 checkNoWrapFlags(AddOpA, Signed) && checkNoWrapFlags(AddOpB, Signed));
414 if (AddOpA->getOperand(MatchingOpIdxA) ==
415 AddOpB->getOperand(MatchingOpIdxB)) {
416 Value *OtherOperandA = AddOpA->getOperand(MatchingOpIdxA == 1 ? 0 : 1);
417 Value *OtherOperandB = AddOpB->getOperand(MatchingOpIdxB == 1 ? 0 : 1);
418 Instruction *OtherInstrA = dyn_cast<Instruction>(OtherOperandA);
419 Instruction *OtherInstrB = dyn_cast<Instruction>(OtherOperandB);
420 // Match `x +nsw/nuw y` and `x +nsw/nuw (y +nsw/nuw IdxDiff)`.
421 if (OtherInstrB && OtherInstrB->getOpcode() == Instruction::Add &&
422 checkNoWrapFlags(OtherInstrB, Signed) &&
423 isa<ConstantInt>(OtherInstrB->getOperand(1))) {
424 int64_t CstVal =
425 cast<ConstantInt>(OtherInstrB->getOperand(1))->getSExtValue();
426 if (OtherInstrB->getOperand(0) == OtherOperandA &&
427 IdxDiff.getSExtValue() == CstVal)
428 return true;
429 }
430 // Match `x +nsw/nuw (y +nsw/nuw -Idx)` and `x +nsw/nuw (y +nsw/nuw x)`.
431 if (OtherInstrA && OtherInstrA->getOpcode() == Instruction::Add &&
432 checkNoWrapFlags(OtherInstrA, Signed) &&
433 isa<ConstantInt>(OtherInstrA->getOperand(1))) {
434 int64_t CstVal =
435 cast<ConstantInt>(OtherInstrA->getOperand(1))->getSExtValue();
436 if (OtherInstrA->getOperand(0) == OtherOperandB &&
437 IdxDiff.getSExtValue() == -CstVal)
438 return true;
439 }
440 // Match `x +nsw/nuw (y +nsw/nuw c)` and
441 // `x +nsw/nuw (y +nsw/nuw (c + IdxDiff))`.
442 if (OtherInstrA && OtherInstrB &&
443 OtherInstrA->getOpcode() == Instruction::Add &&
444 OtherInstrB->getOpcode() == Instruction::Add &&
445 checkNoWrapFlags(OtherInstrA, Signed) &&
446 checkNoWrapFlags(OtherInstrB, Signed) &&
447 isa<ConstantInt>(OtherInstrA->getOperand(1)) &&
448 isa<ConstantInt>(OtherInstrB->getOperand(1))) {
449 int64_t CstValA =
450 cast<ConstantInt>(OtherInstrA->getOperand(1))->getSExtValue();
451 int64_t CstValB =
452 cast<ConstantInt>(OtherInstrB->getOperand(1))->getSExtValue();
453 if (OtherInstrA->getOperand(0) == OtherInstrB->getOperand(0) &&
454 IdxDiff.getSExtValue() == (CstValB - CstValA))
455 return true;
456 }
457 }
458 return false;
459 }
460
lookThroughComplexAddresses(Value * PtrA,Value * PtrB,APInt PtrDelta,unsigned Depth) const461 bool Vectorizer::lookThroughComplexAddresses(Value *PtrA, Value *PtrB,
462 APInt PtrDelta,
463 unsigned Depth) const {
464 auto *GEPA = dyn_cast<GetElementPtrInst>(PtrA);
465 auto *GEPB = dyn_cast<GetElementPtrInst>(PtrB);
466 if (!GEPA || !GEPB)
467 return lookThroughSelects(PtrA, PtrB, PtrDelta, Depth);
468
469 // Look through GEPs after checking they're the same except for the last
470 // index.
471 if (GEPA->getNumOperands() != GEPB->getNumOperands() ||
472 GEPA->getPointerOperand() != GEPB->getPointerOperand())
473 return false;
474 gep_type_iterator GTIA = gep_type_begin(GEPA);
475 gep_type_iterator GTIB = gep_type_begin(GEPB);
476 for (unsigned I = 0, E = GEPA->getNumIndices() - 1; I < E; ++I) {
477 if (GTIA.getOperand() != GTIB.getOperand())
478 return false;
479 ++GTIA;
480 ++GTIB;
481 }
482
483 Instruction *OpA = dyn_cast<Instruction>(GTIA.getOperand());
484 Instruction *OpB = dyn_cast<Instruction>(GTIB.getOperand());
485 if (!OpA || !OpB || OpA->getOpcode() != OpB->getOpcode() ||
486 OpA->getType() != OpB->getType())
487 return false;
488
489 if (PtrDelta.isNegative()) {
490 if (PtrDelta.isMinSignedValue())
491 return false;
492 PtrDelta.negate();
493 std::swap(OpA, OpB);
494 }
495 uint64_t Stride = DL.getTypeAllocSize(GTIA.getIndexedType());
496 if (PtrDelta.urem(Stride) != 0)
497 return false;
498 unsigned IdxBitWidth = OpA->getType()->getScalarSizeInBits();
499 APInt IdxDiff = PtrDelta.udiv(Stride).zext(IdxBitWidth);
500
501 // Only look through a ZExt/SExt.
502 if (!isa<SExtInst>(OpA) && !isa<ZExtInst>(OpA))
503 return false;
504
505 bool Signed = isa<SExtInst>(OpA);
506
507 // At this point A could be a function parameter, i.e. not an instruction
508 Value *ValA = OpA->getOperand(0);
509 OpB = dyn_cast<Instruction>(OpB->getOperand(0));
510 if (!OpB || ValA->getType() != OpB->getType())
511 return false;
512
513 // Now we need to prove that adding IdxDiff to ValA won't overflow.
514 bool Safe = false;
515
516 // First attempt: if OpB is an add with NSW/NUW, and OpB is IdxDiff added to
517 // ValA, we're okay.
518 if (OpB->getOpcode() == Instruction::Add &&
519 isa<ConstantInt>(OpB->getOperand(1)) &&
520 IdxDiff.sle(cast<ConstantInt>(OpB->getOperand(1))->getSExtValue()) &&
521 checkNoWrapFlags(OpB, Signed))
522 Safe = true;
523
524 // Second attempt: check if we have eligible add NSW/NUW instruction
525 // sequences.
526 OpA = dyn_cast<Instruction>(ValA);
527 if (!Safe && OpA && OpA->getOpcode() == Instruction::Add &&
528 OpB->getOpcode() == Instruction::Add && checkNoWrapFlags(OpA, Signed) &&
529 checkNoWrapFlags(OpB, Signed)) {
530 // In the checks below a matching operand in OpA and OpB is
531 // an operand which is the same in those two instructions.
532 // Below we account for possible orders of the operands of
533 // these add instructions.
534 for (unsigned MatchingOpIdxA : {0, 1})
535 for (unsigned MatchingOpIdxB : {0, 1})
536 if (!Safe)
537 Safe = checkIfSafeAddSequence(IdxDiff, OpA, MatchingOpIdxA, OpB,
538 MatchingOpIdxB, Signed);
539 }
540
541 unsigned BitWidth = ValA->getType()->getScalarSizeInBits();
542
543 // Third attempt:
544 // If all set bits of IdxDiff or any higher order bit other than the sign bit
545 // are known to be zero in ValA, we can add Diff to it while guaranteeing no
546 // overflow of any sort.
547 if (!Safe) {
548 KnownBits Known(BitWidth);
549 computeKnownBits(ValA, Known, DL, 0, &AC, OpB, &DT);
550 APInt BitsAllowedToBeSet = Known.Zero.zext(IdxDiff.getBitWidth());
551 if (Signed)
552 BitsAllowedToBeSet.clearBit(BitWidth - 1);
553 if (BitsAllowedToBeSet.ult(IdxDiff))
554 return false;
555 }
556
557 const SCEV *OffsetSCEVA = SE.getSCEV(ValA);
558 const SCEV *OffsetSCEVB = SE.getSCEV(OpB);
559 const SCEV *C = SE.getConstant(IdxDiff.trunc(BitWidth));
560 const SCEV *X = SE.getAddExpr(OffsetSCEVA, C);
561 return X == OffsetSCEVB;
562 }
563
lookThroughSelects(Value * PtrA,Value * PtrB,const APInt & PtrDelta,unsigned Depth) const564 bool Vectorizer::lookThroughSelects(Value *PtrA, Value *PtrB,
565 const APInt &PtrDelta,
566 unsigned Depth) const {
567 if (Depth++ == MaxDepth)
568 return false;
569
570 if (auto *SelectA = dyn_cast<SelectInst>(PtrA)) {
571 if (auto *SelectB = dyn_cast<SelectInst>(PtrB)) {
572 return SelectA->getCondition() == SelectB->getCondition() &&
573 areConsecutivePointers(SelectA->getTrueValue(),
574 SelectB->getTrueValue(), PtrDelta, Depth) &&
575 areConsecutivePointers(SelectA->getFalseValue(),
576 SelectB->getFalseValue(), PtrDelta, Depth);
577 }
578 }
579 return false;
580 }
581
reorder(Instruction * I)582 void Vectorizer::reorder(Instruction *I) {
583 SmallPtrSet<Instruction *, 16> InstructionsToMove;
584 SmallVector<Instruction *, 16> Worklist;
585
586 Worklist.push_back(I);
587 while (!Worklist.empty()) {
588 Instruction *IW = Worklist.pop_back_val();
589 int NumOperands = IW->getNumOperands();
590 for (int i = 0; i < NumOperands; i++) {
591 Instruction *IM = dyn_cast<Instruction>(IW->getOperand(i));
592 if (!IM || IM->getOpcode() == Instruction::PHI)
593 continue;
594
595 // If IM is in another BB, no need to move it, because this pass only
596 // vectorizes instructions within one BB.
597 if (IM->getParent() != I->getParent())
598 continue;
599
600 if (!IM->comesBefore(I)) {
601 InstructionsToMove.insert(IM);
602 Worklist.push_back(IM);
603 }
604 }
605 }
606
607 // All instructions to move should follow I. Start from I, not from begin().
608 for (auto BBI = I->getIterator(), E = I->getParent()->end(); BBI != E;
609 ++BBI) {
610 if (!InstructionsToMove.count(&*BBI))
611 continue;
612 Instruction *IM = &*BBI;
613 --BBI;
614 IM->removeFromParent();
615 IM->insertBefore(I);
616 }
617 }
618
619 std::pair<BasicBlock::iterator, BasicBlock::iterator>
getBoundaryInstrs(ArrayRef<Instruction * > Chain)620 Vectorizer::getBoundaryInstrs(ArrayRef<Instruction *> Chain) {
621 Instruction *C0 = Chain[0];
622 BasicBlock::iterator FirstInstr = C0->getIterator();
623 BasicBlock::iterator LastInstr = C0->getIterator();
624
625 BasicBlock *BB = C0->getParent();
626 unsigned NumFound = 0;
627 for (Instruction &I : *BB) {
628 if (!is_contained(Chain, &I))
629 continue;
630
631 ++NumFound;
632 if (NumFound == 1) {
633 FirstInstr = I.getIterator();
634 }
635 if (NumFound == Chain.size()) {
636 LastInstr = I.getIterator();
637 break;
638 }
639 }
640
641 // Range is [first, last).
642 return std::make_pair(FirstInstr, ++LastInstr);
643 }
644
eraseInstructions(ArrayRef<Instruction * > Chain)645 void Vectorizer::eraseInstructions(ArrayRef<Instruction *> Chain) {
646 SmallVector<Instruction *, 16> Instrs;
647 for (Instruction *I : Chain) {
648 Value *PtrOperand = getLoadStorePointerOperand(I);
649 assert(PtrOperand && "Instruction must have a pointer operand.");
650 Instrs.push_back(I);
651 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(PtrOperand))
652 Instrs.push_back(GEP);
653 }
654
655 // Erase instructions.
656 for (Instruction *I : Instrs)
657 if (I->use_empty())
658 I->eraseFromParent();
659 }
660
661 std::pair<ArrayRef<Instruction *>, ArrayRef<Instruction *>>
splitOddVectorElts(ArrayRef<Instruction * > Chain,unsigned ElementSizeBits)662 Vectorizer::splitOddVectorElts(ArrayRef<Instruction *> Chain,
663 unsigned ElementSizeBits) {
664 unsigned ElementSizeBytes = ElementSizeBits / 8;
665 unsigned SizeBytes = ElementSizeBytes * Chain.size();
666 unsigned NumLeft = (SizeBytes - (SizeBytes % 4)) / ElementSizeBytes;
667 if (NumLeft == Chain.size()) {
668 if ((NumLeft & 1) == 0)
669 NumLeft /= 2; // Split even in half
670 else
671 --NumLeft; // Split off last element
672 } else if (NumLeft == 0)
673 NumLeft = 1;
674 return std::make_pair(Chain.slice(0, NumLeft), Chain.slice(NumLeft));
675 }
676
677 ArrayRef<Instruction *>
getVectorizablePrefix(ArrayRef<Instruction * > Chain)678 Vectorizer::getVectorizablePrefix(ArrayRef<Instruction *> Chain) {
679 // These are in BB order, unlike Chain, which is in address order.
680 SmallVector<Instruction *, 16> MemoryInstrs;
681 SmallVector<Instruction *, 16> ChainInstrs;
682
683 bool IsLoadChain = isa<LoadInst>(Chain[0]);
684 LLVM_DEBUG({
685 for (Instruction *I : Chain) {
686 if (IsLoadChain)
687 assert(isa<LoadInst>(I) &&
688 "All elements of Chain must be loads, or all must be stores.");
689 else
690 assert(isa<StoreInst>(I) &&
691 "All elements of Chain must be loads, or all must be stores.");
692 }
693 });
694
695 for (Instruction &I : make_range(getBoundaryInstrs(Chain))) {
696 if ((isa<LoadInst>(I) || isa<StoreInst>(I)) && is_contained(Chain, &I)) {
697 ChainInstrs.push_back(&I);
698 continue;
699 }
700 if (!isGuaranteedToTransferExecutionToSuccessor(&I)) {
701 LLVM_DEBUG(dbgs() << "LSV: Found instruction may not transfer execution: "
702 << I << '\n');
703 break;
704 }
705 if (I.mayReadOrWriteMemory())
706 MemoryInstrs.push_back(&I);
707 }
708
709 // Loop until we find an instruction in ChainInstrs that we can't vectorize.
710 unsigned ChainInstrIdx = 0;
711 Instruction *BarrierMemoryInstr = nullptr;
712
713 for (unsigned E = ChainInstrs.size(); ChainInstrIdx < E; ++ChainInstrIdx) {
714 Instruction *ChainInstr = ChainInstrs[ChainInstrIdx];
715
716 // If a barrier memory instruction was found, chain instructions that follow
717 // will not be added to the valid prefix.
718 if (BarrierMemoryInstr && BarrierMemoryInstr->comesBefore(ChainInstr))
719 break;
720
721 // Check (in BB order) if any instruction prevents ChainInstr from being
722 // vectorized. Find and store the first such "conflicting" instruction.
723 for (Instruction *MemInstr : MemoryInstrs) {
724 // If a barrier memory instruction was found, do not check past it.
725 if (BarrierMemoryInstr && BarrierMemoryInstr->comesBefore(MemInstr))
726 break;
727
728 auto *MemLoad = dyn_cast<LoadInst>(MemInstr);
729 auto *ChainLoad = dyn_cast<LoadInst>(ChainInstr);
730 if (MemLoad && ChainLoad)
731 continue;
732
733 // We can ignore the alias if the we have a load store pair and the load
734 // is known to be invariant. The load cannot be clobbered by the store.
735 auto IsInvariantLoad = [](const LoadInst *LI) -> bool {
736 return LI->hasMetadata(LLVMContext::MD_invariant_load);
737 };
738
739 if (IsLoadChain) {
740 // We can ignore the alias as long as the load comes before the store,
741 // because that means we won't be moving the load past the store to
742 // vectorize it (the vectorized load is inserted at the location of the
743 // first load in the chain).
744 if (ChainInstr->comesBefore(MemInstr) ||
745 (ChainLoad && IsInvariantLoad(ChainLoad)))
746 continue;
747 } else {
748 // Same case, but in reverse.
749 if (MemInstr->comesBefore(ChainInstr) ||
750 (MemLoad && IsInvariantLoad(MemLoad)))
751 continue;
752 }
753
754 ModRefInfo MR =
755 AA.getModRefInfo(MemInstr, MemoryLocation::get(ChainInstr));
756 if (IsLoadChain ? isModSet(MR) : isModOrRefSet(MR)) {
757 LLVM_DEBUG({
758 dbgs() << "LSV: Found alias:\n"
759 " Aliasing instruction:\n"
760 << " " << *MemInstr << '\n'
761 << " Aliased instruction and pointer:\n"
762 << " " << *ChainInstr << '\n'
763 << " " << *getLoadStorePointerOperand(ChainInstr) << '\n';
764 });
765 // Save this aliasing memory instruction as a barrier, but allow other
766 // instructions that precede the barrier to be vectorized with this one.
767 BarrierMemoryInstr = MemInstr;
768 break;
769 }
770 }
771 // Continue the search only for store chains, since vectorizing stores that
772 // precede an aliasing load is valid. Conversely, vectorizing loads is valid
773 // up to an aliasing store, but should not pull loads from further down in
774 // the basic block.
775 if (IsLoadChain && BarrierMemoryInstr) {
776 // The BarrierMemoryInstr is a store that precedes ChainInstr.
777 assert(BarrierMemoryInstr->comesBefore(ChainInstr));
778 break;
779 }
780 }
781
782 // Find the largest prefix of Chain whose elements are all in
783 // ChainInstrs[0, ChainInstrIdx). This is the largest vectorizable prefix of
784 // Chain. (Recall that Chain is in address order, but ChainInstrs is in BB
785 // order.)
786 SmallPtrSet<Instruction *, 8> VectorizableChainInstrs(
787 ChainInstrs.begin(), ChainInstrs.begin() + ChainInstrIdx);
788 unsigned ChainIdx = 0;
789 for (unsigned ChainLen = Chain.size(); ChainIdx < ChainLen; ++ChainIdx) {
790 if (!VectorizableChainInstrs.count(Chain[ChainIdx]))
791 break;
792 }
793 return Chain.slice(0, ChainIdx);
794 }
795
getChainID(const Value * Ptr)796 static ChainID getChainID(const Value *Ptr) {
797 const Value *ObjPtr = getUnderlyingObject(Ptr);
798 if (const auto *Sel = dyn_cast<SelectInst>(ObjPtr)) {
799 // The select's themselves are distinct instructions even if they share the
800 // same condition and evaluate to consecutive pointers for true and false
801 // values of the condition. Therefore using the select's themselves for
802 // grouping instructions would put consecutive accesses into different lists
803 // and they won't be even checked for being consecutive, and won't be
804 // vectorized.
805 return Sel->getCondition();
806 }
807 return ObjPtr;
808 }
809
810 std::pair<InstrListMap, InstrListMap>
collectInstructions(BasicBlock * BB)811 Vectorizer::collectInstructions(BasicBlock *BB) {
812 InstrListMap LoadRefs;
813 InstrListMap StoreRefs;
814
815 for (Instruction &I : *BB) {
816 if (!I.mayReadOrWriteMemory())
817 continue;
818
819 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
820 if (!LI->isSimple())
821 continue;
822
823 // Skip if it's not legal.
824 if (!TTI.isLegalToVectorizeLoad(LI))
825 continue;
826
827 Type *Ty = LI->getType();
828 if (!VectorType::isValidElementType(Ty->getScalarType()))
829 continue;
830
831 // Skip weird non-byte sizes. They probably aren't worth the effort of
832 // handling correctly.
833 unsigned TySize = DL.getTypeSizeInBits(Ty);
834 if ((TySize % 8) != 0)
835 continue;
836
837 // Skip vectors of pointers. The vectorizeLoadChain/vectorizeStoreChain
838 // functions are currently using an integer type for the vectorized
839 // load/store, and does not support casting between the integer type and a
840 // vector of pointers (e.g. i64 to <2 x i16*>)
841 if (Ty->isVectorTy() && Ty->isPtrOrPtrVectorTy())
842 continue;
843
844 Value *Ptr = LI->getPointerOperand();
845 unsigned AS = Ptr->getType()->getPointerAddressSpace();
846 unsigned VecRegSize = TTI.getLoadStoreVecRegBitWidth(AS);
847
848 unsigned VF = VecRegSize / TySize;
849 VectorType *VecTy = dyn_cast<VectorType>(Ty);
850
851 // No point in looking at these if they're too big to vectorize.
852 if (TySize > VecRegSize / 2 ||
853 (VecTy && TTI.getLoadVectorFactor(VF, TySize, TySize / 8, VecTy) == 0))
854 continue;
855
856 // Save the load locations.
857 const ChainID ID = getChainID(Ptr);
858 LoadRefs[ID].push_back(LI);
859 } else if (StoreInst *SI = dyn_cast<StoreInst>(&I)) {
860 if (!SI->isSimple())
861 continue;
862
863 // Skip if it's not legal.
864 if (!TTI.isLegalToVectorizeStore(SI))
865 continue;
866
867 Type *Ty = SI->getValueOperand()->getType();
868 if (!VectorType::isValidElementType(Ty->getScalarType()))
869 continue;
870
871 // Skip vectors of pointers. The vectorizeLoadChain/vectorizeStoreChain
872 // functions are currently using an integer type for the vectorized
873 // load/store, and does not support casting between the integer type and a
874 // vector of pointers (e.g. i64 to <2 x i16*>)
875 if (Ty->isVectorTy() && Ty->isPtrOrPtrVectorTy())
876 continue;
877
878 // Skip weird non-byte sizes. They probably aren't worth the effort of
879 // handling correctly.
880 unsigned TySize = DL.getTypeSizeInBits(Ty);
881 if ((TySize % 8) != 0)
882 continue;
883
884 Value *Ptr = SI->getPointerOperand();
885 unsigned AS = Ptr->getType()->getPointerAddressSpace();
886 unsigned VecRegSize = TTI.getLoadStoreVecRegBitWidth(AS);
887
888 unsigned VF = VecRegSize / TySize;
889 VectorType *VecTy = dyn_cast<VectorType>(Ty);
890
891 // No point in looking at these if they're too big to vectorize.
892 if (TySize > VecRegSize / 2 ||
893 (VecTy && TTI.getStoreVectorFactor(VF, TySize, TySize / 8, VecTy) == 0))
894 continue;
895
896 // Save store location.
897 const ChainID ID = getChainID(Ptr);
898 StoreRefs[ID].push_back(SI);
899 }
900 }
901
902 return {LoadRefs, StoreRefs};
903 }
904
vectorizeChains(InstrListMap & Map)905 bool Vectorizer::vectorizeChains(InstrListMap &Map) {
906 bool Changed = false;
907
908 for (const std::pair<ChainID, InstrList> &Chain : Map) {
909 unsigned Size = Chain.second.size();
910 if (Size < 2)
911 continue;
912
913 LLVM_DEBUG(dbgs() << "LSV: Analyzing a chain of length " << Size << ".\n");
914
915 // Process the stores in chunks of 64.
916 for (unsigned CI = 0, CE = Size; CI < CE; CI += 64) {
917 unsigned Len = std::min<unsigned>(CE - CI, 64);
918 ArrayRef<Instruction *> Chunk(&Chain.second[CI], Len);
919 Changed |= vectorizeInstructions(Chunk);
920 }
921 }
922
923 return Changed;
924 }
925
vectorizeInstructions(ArrayRef<Instruction * > Instrs)926 bool Vectorizer::vectorizeInstructions(ArrayRef<Instruction *> Instrs) {
927 LLVM_DEBUG(dbgs() << "LSV: Vectorizing " << Instrs.size()
928 << " instructions.\n");
929 SmallVector<int, 16> Heads, Tails;
930 int ConsecutiveChain[64];
931
932 // Do a quadratic search on all of the given loads/stores and find all of the
933 // pairs of loads/stores that follow each other.
934 for (int i = 0, e = Instrs.size(); i < e; ++i) {
935 ConsecutiveChain[i] = -1;
936 for (int j = e - 1; j >= 0; --j) {
937 if (i == j)
938 continue;
939
940 if (isConsecutiveAccess(Instrs[i], Instrs[j])) {
941 if (ConsecutiveChain[i] != -1) {
942 int CurDistance = std::abs(ConsecutiveChain[i] - i);
943 int NewDistance = std::abs(ConsecutiveChain[i] - j);
944 if (j < i || NewDistance > CurDistance)
945 continue; // Should not insert.
946 }
947
948 Tails.push_back(j);
949 Heads.push_back(i);
950 ConsecutiveChain[i] = j;
951 }
952 }
953 }
954
955 bool Changed = false;
956 SmallPtrSet<Instruction *, 16> InstructionsProcessed;
957
958 for (int Head : Heads) {
959 if (InstructionsProcessed.count(Instrs[Head]))
960 continue;
961 bool LongerChainExists = false;
962 for (unsigned TIt = 0; TIt < Tails.size(); TIt++)
963 if (Head == Tails[TIt] &&
964 !InstructionsProcessed.count(Instrs[Heads[TIt]])) {
965 LongerChainExists = true;
966 break;
967 }
968 if (LongerChainExists)
969 continue;
970
971 // We found an instr that starts a chain. Now follow the chain and try to
972 // vectorize it.
973 SmallVector<Instruction *, 16> Operands;
974 int I = Head;
975 while (I != -1 && (is_contained(Tails, I) || is_contained(Heads, I))) {
976 if (InstructionsProcessed.count(Instrs[I]))
977 break;
978
979 Operands.push_back(Instrs[I]);
980 I = ConsecutiveChain[I];
981 }
982
983 bool Vectorized = false;
984 if (isa<LoadInst>(*Operands.begin()))
985 Vectorized = vectorizeLoadChain(Operands, &InstructionsProcessed);
986 else
987 Vectorized = vectorizeStoreChain(Operands, &InstructionsProcessed);
988
989 Changed |= Vectorized;
990 }
991
992 return Changed;
993 }
994
vectorizeStoreChain(ArrayRef<Instruction * > Chain,SmallPtrSet<Instruction *,16> * InstructionsProcessed)995 bool Vectorizer::vectorizeStoreChain(
996 ArrayRef<Instruction *> Chain,
997 SmallPtrSet<Instruction *, 16> *InstructionsProcessed) {
998 StoreInst *S0 = cast<StoreInst>(Chain[0]);
999
1000 // If the vector has an int element, default to int for the whole store.
1001 Type *StoreTy = nullptr;
1002 for (Instruction *I : Chain) {
1003 StoreTy = cast<StoreInst>(I)->getValueOperand()->getType();
1004 if (StoreTy->isIntOrIntVectorTy())
1005 break;
1006
1007 if (StoreTy->isPtrOrPtrVectorTy()) {
1008 StoreTy = Type::getIntNTy(F.getParent()->getContext(),
1009 DL.getTypeSizeInBits(StoreTy));
1010 break;
1011 }
1012 }
1013 assert(StoreTy && "Failed to find store type");
1014
1015 unsigned Sz = DL.getTypeSizeInBits(StoreTy);
1016 unsigned AS = S0->getPointerAddressSpace();
1017 unsigned VecRegSize = TTI.getLoadStoreVecRegBitWidth(AS);
1018 unsigned VF = VecRegSize / Sz;
1019 unsigned ChainSize = Chain.size();
1020 Align Alignment = S0->getAlign();
1021
1022 if (!isPowerOf2_32(Sz) || VF < 2 || ChainSize < 2) {
1023 InstructionsProcessed->insert(Chain.begin(), Chain.end());
1024 return false;
1025 }
1026
1027 ArrayRef<Instruction *> NewChain = getVectorizablePrefix(Chain);
1028 if (NewChain.empty()) {
1029 // No vectorization possible.
1030 InstructionsProcessed->insert(Chain.begin(), Chain.end());
1031 return false;
1032 }
1033 if (NewChain.size() == 1) {
1034 // Failed after the first instruction. Discard it and try the smaller chain.
1035 InstructionsProcessed->insert(NewChain.front());
1036 return false;
1037 }
1038
1039 // Update Chain to the valid vectorizable subchain.
1040 Chain = NewChain;
1041 ChainSize = Chain.size();
1042
1043 // Check if it's legal to vectorize this chain. If not, split the chain and
1044 // try again.
1045 unsigned EltSzInBytes = Sz / 8;
1046 unsigned SzInBytes = EltSzInBytes * ChainSize;
1047
1048 FixedVectorType *VecTy;
1049 auto *VecStoreTy = dyn_cast<FixedVectorType>(StoreTy);
1050 if (VecStoreTy)
1051 VecTy = FixedVectorType::get(StoreTy->getScalarType(),
1052 Chain.size() * VecStoreTy->getNumElements());
1053 else
1054 VecTy = FixedVectorType::get(StoreTy, Chain.size());
1055
1056 // If it's more than the max vector size or the target has a better
1057 // vector factor, break it into two pieces.
1058 unsigned TargetVF = TTI.getStoreVectorFactor(VF, Sz, SzInBytes, VecTy);
1059 if (ChainSize > VF || (VF != TargetVF && TargetVF < ChainSize)) {
1060 LLVM_DEBUG(dbgs() << "LSV: Chain doesn't match with the vector factor."
1061 " Creating two separate arrays.\n");
1062 bool Vectorized = false;
1063 Vectorized |=
1064 vectorizeStoreChain(Chain.slice(0, TargetVF), InstructionsProcessed);
1065 Vectorized |=
1066 vectorizeStoreChain(Chain.slice(TargetVF), InstructionsProcessed);
1067 return Vectorized;
1068 }
1069
1070 LLVM_DEBUG({
1071 dbgs() << "LSV: Stores to vectorize:\n";
1072 for (Instruction *I : Chain)
1073 dbgs() << " " << *I << "\n";
1074 });
1075
1076 // We won't try again to vectorize the elements of the chain, regardless of
1077 // whether we succeed below.
1078 InstructionsProcessed->insert(Chain.begin(), Chain.end());
1079
1080 // If the store is going to be misaligned, don't vectorize it.
1081 if (accessIsMisaligned(SzInBytes, AS, Alignment)) {
1082 if (S0->getPointerAddressSpace() != DL.getAllocaAddrSpace()) {
1083 auto Chains = splitOddVectorElts(Chain, Sz);
1084 bool Vectorized = false;
1085 Vectorized |= vectorizeStoreChain(Chains.first, InstructionsProcessed);
1086 Vectorized |= vectorizeStoreChain(Chains.second, InstructionsProcessed);
1087 return Vectorized;
1088 }
1089
1090 Align NewAlign = getOrEnforceKnownAlignment(S0->getPointerOperand(),
1091 Align(StackAdjustedAlignment),
1092 DL, S0, nullptr, &DT);
1093 if (NewAlign >= Alignment)
1094 Alignment = NewAlign;
1095 else
1096 return false;
1097 }
1098
1099 if (!TTI.isLegalToVectorizeStoreChain(SzInBytes, Alignment, AS)) {
1100 auto Chains = splitOddVectorElts(Chain, Sz);
1101 bool Vectorized = false;
1102 Vectorized |= vectorizeStoreChain(Chains.first, InstructionsProcessed);
1103 Vectorized |= vectorizeStoreChain(Chains.second, InstructionsProcessed);
1104 return Vectorized;
1105 }
1106
1107 BasicBlock::iterator First, Last;
1108 std::tie(First, Last) = getBoundaryInstrs(Chain);
1109 Builder.SetInsertPoint(&*Last);
1110
1111 Value *Vec = PoisonValue::get(VecTy);
1112
1113 if (VecStoreTy) {
1114 unsigned VecWidth = VecStoreTy->getNumElements();
1115 for (unsigned I = 0, E = Chain.size(); I != E; ++I) {
1116 StoreInst *Store = cast<StoreInst>(Chain[I]);
1117 for (unsigned J = 0, NE = VecStoreTy->getNumElements(); J != NE; ++J) {
1118 unsigned NewIdx = J + I * VecWidth;
1119 Value *Extract = Builder.CreateExtractElement(Store->getValueOperand(),
1120 Builder.getInt32(J));
1121 if (Extract->getType() != StoreTy->getScalarType())
1122 Extract = Builder.CreateBitCast(Extract, StoreTy->getScalarType());
1123
1124 Value *Insert =
1125 Builder.CreateInsertElement(Vec, Extract, Builder.getInt32(NewIdx));
1126 Vec = Insert;
1127 }
1128 }
1129 } else {
1130 for (unsigned I = 0, E = Chain.size(); I != E; ++I) {
1131 StoreInst *Store = cast<StoreInst>(Chain[I]);
1132 Value *Extract = Store->getValueOperand();
1133 if (Extract->getType() != StoreTy->getScalarType())
1134 Extract =
1135 Builder.CreateBitOrPointerCast(Extract, StoreTy->getScalarType());
1136
1137 Value *Insert =
1138 Builder.CreateInsertElement(Vec, Extract, Builder.getInt32(I));
1139 Vec = Insert;
1140 }
1141 }
1142
1143 StoreInst *SI = Builder.CreateAlignedStore(
1144 Vec,
1145 Builder.CreateBitCast(S0->getPointerOperand(), VecTy->getPointerTo(AS)),
1146 Alignment);
1147 propagateMetadata(SI, Chain);
1148
1149 eraseInstructions(Chain);
1150 ++NumVectorInstructions;
1151 NumScalarsVectorized += Chain.size();
1152 return true;
1153 }
1154
vectorizeLoadChain(ArrayRef<Instruction * > Chain,SmallPtrSet<Instruction *,16> * InstructionsProcessed)1155 bool Vectorizer::vectorizeLoadChain(
1156 ArrayRef<Instruction *> Chain,
1157 SmallPtrSet<Instruction *, 16> *InstructionsProcessed) {
1158 LoadInst *L0 = cast<LoadInst>(Chain[0]);
1159
1160 // If the vector has an int element, default to int for the whole load.
1161 Type *LoadTy = nullptr;
1162 for (const auto &V : Chain) {
1163 LoadTy = cast<LoadInst>(V)->getType();
1164 if (LoadTy->isIntOrIntVectorTy())
1165 break;
1166
1167 if (LoadTy->isPtrOrPtrVectorTy()) {
1168 LoadTy = Type::getIntNTy(F.getParent()->getContext(),
1169 DL.getTypeSizeInBits(LoadTy));
1170 break;
1171 }
1172 }
1173 assert(LoadTy && "Can't determine LoadInst type from chain");
1174
1175 unsigned Sz = DL.getTypeSizeInBits(LoadTy);
1176 unsigned AS = L0->getPointerAddressSpace();
1177 unsigned VecRegSize = TTI.getLoadStoreVecRegBitWidth(AS);
1178 unsigned VF = VecRegSize / Sz;
1179 unsigned ChainSize = Chain.size();
1180 Align Alignment = L0->getAlign();
1181
1182 if (!isPowerOf2_32(Sz) || VF < 2 || ChainSize < 2) {
1183 InstructionsProcessed->insert(Chain.begin(), Chain.end());
1184 return false;
1185 }
1186
1187 ArrayRef<Instruction *> NewChain = getVectorizablePrefix(Chain);
1188 if (NewChain.empty()) {
1189 // No vectorization possible.
1190 InstructionsProcessed->insert(Chain.begin(), Chain.end());
1191 return false;
1192 }
1193 if (NewChain.size() == 1) {
1194 // Failed after the first instruction. Discard it and try the smaller chain.
1195 InstructionsProcessed->insert(NewChain.front());
1196 return false;
1197 }
1198
1199 // Update Chain to the valid vectorizable subchain.
1200 Chain = NewChain;
1201 ChainSize = Chain.size();
1202
1203 // Check if it's legal to vectorize this chain. If not, split the chain and
1204 // try again.
1205 unsigned EltSzInBytes = Sz / 8;
1206 unsigned SzInBytes = EltSzInBytes * ChainSize;
1207 VectorType *VecTy;
1208 auto *VecLoadTy = dyn_cast<FixedVectorType>(LoadTy);
1209 if (VecLoadTy)
1210 VecTy = FixedVectorType::get(LoadTy->getScalarType(),
1211 Chain.size() * VecLoadTy->getNumElements());
1212 else
1213 VecTy = FixedVectorType::get(LoadTy, Chain.size());
1214
1215 // If it's more than the max vector size or the target has a better
1216 // vector factor, break it into two pieces.
1217 unsigned TargetVF = TTI.getLoadVectorFactor(VF, Sz, SzInBytes, VecTy);
1218 if (ChainSize > VF || (VF != TargetVF && TargetVF < ChainSize)) {
1219 LLVM_DEBUG(dbgs() << "LSV: Chain doesn't match with the vector factor."
1220 " Creating two separate arrays.\n");
1221 bool Vectorized = false;
1222 Vectorized |=
1223 vectorizeLoadChain(Chain.slice(0, TargetVF), InstructionsProcessed);
1224 Vectorized |=
1225 vectorizeLoadChain(Chain.slice(TargetVF), InstructionsProcessed);
1226 return Vectorized;
1227 }
1228
1229 // We won't try again to vectorize the elements of the chain, regardless of
1230 // whether we succeed below.
1231 InstructionsProcessed->insert(Chain.begin(), Chain.end());
1232
1233 // If the load is going to be misaligned, don't vectorize it.
1234 if (accessIsMisaligned(SzInBytes, AS, Alignment)) {
1235 if (L0->getPointerAddressSpace() != DL.getAllocaAddrSpace()) {
1236 auto Chains = splitOddVectorElts(Chain, Sz);
1237 bool Vectorized = false;
1238 Vectorized |= vectorizeLoadChain(Chains.first, InstructionsProcessed);
1239 Vectorized |= vectorizeLoadChain(Chains.second, InstructionsProcessed);
1240 return Vectorized;
1241 }
1242
1243 Align NewAlign = getOrEnforceKnownAlignment(L0->getPointerOperand(),
1244 Align(StackAdjustedAlignment),
1245 DL, L0, nullptr, &DT);
1246 if (NewAlign >= Alignment)
1247 Alignment = NewAlign;
1248 else
1249 return false;
1250 }
1251
1252 if (!TTI.isLegalToVectorizeLoadChain(SzInBytes, Alignment, AS)) {
1253 auto Chains = splitOddVectorElts(Chain, Sz);
1254 bool Vectorized = false;
1255 Vectorized |= vectorizeLoadChain(Chains.first, InstructionsProcessed);
1256 Vectorized |= vectorizeLoadChain(Chains.second, InstructionsProcessed);
1257 return Vectorized;
1258 }
1259
1260 LLVM_DEBUG({
1261 dbgs() << "LSV: Loads to vectorize:\n";
1262 for (Instruction *I : Chain)
1263 I->dump();
1264 });
1265
1266 // getVectorizablePrefix already computed getBoundaryInstrs. The value of
1267 // Last may have changed since then, but the value of First won't have. If it
1268 // matters, we could compute getBoundaryInstrs only once and reuse it here.
1269 BasicBlock::iterator First, Last;
1270 std::tie(First, Last) = getBoundaryInstrs(Chain);
1271 Builder.SetInsertPoint(&*First);
1272
1273 Value *Bitcast =
1274 Builder.CreateBitCast(L0->getPointerOperand(), VecTy->getPointerTo(AS));
1275 LoadInst *LI =
1276 Builder.CreateAlignedLoad(VecTy, Bitcast, MaybeAlign(Alignment));
1277 propagateMetadata(LI, Chain);
1278
1279 for (unsigned I = 0, E = Chain.size(); I != E; ++I) {
1280 Value *CV = Chain[I];
1281 Value *V;
1282 if (VecLoadTy) {
1283 // Extract a subvector using shufflevector.
1284 unsigned VecWidth = VecLoadTy->getNumElements();
1285 auto Mask =
1286 llvm::to_vector<8>(llvm::seq<int>(I * VecWidth, (I + 1) * VecWidth));
1287 V = Builder.CreateShuffleVector(LI, Mask, CV->getName());
1288 } else {
1289 V = Builder.CreateExtractElement(LI, Builder.getInt32(I), CV->getName());
1290 }
1291
1292 if (V->getType() != CV->getType()) {
1293 V = Builder.CreateBitOrPointerCast(V, CV->getType());
1294 }
1295
1296 // Replace the old instruction.
1297 CV->replaceAllUsesWith(V);
1298 }
1299
1300 // Since we might have opaque pointers we might end up using the pointer
1301 // operand of the first load (wrt. memory loaded) for the vector load. Since
1302 // this first load might not be the first in the block we potentially need to
1303 // reorder the pointer operand (and its operands). If we have a bitcast though
1304 // it might be before the load and should be the reorder start instruction.
1305 // "Might" because for opaque pointers the "bitcast" is just the first loads
1306 // pointer operand, as oppposed to something we inserted at the right position
1307 // ourselves.
1308 Instruction *BCInst = dyn_cast<Instruction>(Bitcast);
1309 reorder((BCInst && BCInst != L0->getPointerOperand()) ? BCInst : LI);
1310
1311 eraseInstructions(Chain);
1312
1313 ++NumVectorInstructions;
1314 NumScalarsVectorized += Chain.size();
1315 return true;
1316 }
1317
accessIsMisaligned(unsigned SzInBytes,unsigned AddressSpace,Align Alignment)1318 bool Vectorizer::accessIsMisaligned(unsigned SzInBytes, unsigned AddressSpace,
1319 Align Alignment) {
1320 if (Alignment.value() % SzInBytes == 0)
1321 return false;
1322
1323 bool Fast = false;
1324 bool Allows = TTI.allowsMisalignedMemoryAccesses(F.getParent()->getContext(),
1325 SzInBytes * 8, AddressSpace,
1326 Alignment, &Fast);
1327 LLVM_DEBUG(dbgs() << "LSV: Target said misaligned is allowed? " << Allows
1328 << " and fast? " << Fast << "\n";);
1329 return !Allows || !Fast;
1330 }
1331