1 //===- InstCombineLoadStoreAlloca.cpp -------------------------------------===//
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
7 //===----------------------------------------------------------------------===//
9 // This file implements the visit functions for load, store and alloca.
11 //===----------------------------------------------------------------------===//
13 #include "InstCombineInternal.h"
14 #include "llvm/ADT/MapVector.h"
15 #include "llvm/ADT/SmallString.h"
16 #include "llvm/ADT/Statistic.h"
17 #include "llvm/Analysis/Loads.h"
18 #include "llvm/Transforms/Utils/Local.h"
19 #include "llvm/IR/ConstantRange.h"
20 #include "llvm/IR/DataLayout.h"
21 #include "llvm/IR/DebugInfoMetadata.h"
22 #include "llvm/IR/IntrinsicInst.h"
23 #include "llvm/IR/LLVMContext.h"
24 #include "llvm/IR/MDBuilder.h"
25 #include "llvm/IR/PatternMatch.h"
26 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
28 using namespace PatternMatch;
30 #define DEBUG_TYPE "instcombine"
32 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
33 STATISTIC(NumGlobalCopies, "Number of allocas copied from constant global");
35 /// pointsToConstantGlobal - Return true if V (possibly indirectly) points to
36 /// some part of a constant global variable. This intentionally only accepts
37 /// constant expressions because we can't rewrite arbitrary instructions.
38 static bool pointsToConstantGlobal(Value *V) {
39 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
40 return GV->isConstant();
42 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
43 if (CE->getOpcode() == Instruction::BitCast ||
44 CE->getOpcode() == Instruction::AddrSpaceCast ||
45 CE->getOpcode() == Instruction::GetElementPtr)
46 return pointsToConstantGlobal(CE->getOperand(0));
51 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
52 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
53 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
54 /// track of whether it moves the pointer (with IsOffset) but otherwise traverse
55 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
56 /// the alloca, and if the source pointer is a pointer to a constant global, we
57 /// can optimize this.
59 isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
60 SmallVectorImpl<Instruction *> &ToDelete) {
61 // We track lifetime intrinsics as we encounter them. If we decide to go
62 // ahead and replace the value with the global, this lets the caller quickly
63 // eliminate the markers.
65 SmallVector<std::pair<Value *, bool>, 35> ValuesToInspect;
66 ValuesToInspect.emplace_back(V, false);
67 while (!ValuesToInspect.empty()) {
68 auto ValuePair = ValuesToInspect.pop_back_val();
69 const bool IsOffset = ValuePair.second;
70 for (auto &U : ValuePair.first->uses()) {
71 auto *I = cast<Instruction>(U.getUser());
73 if (auto *LI = dyn_cast<LoadInst>(I)) {
74 // Ignore non-volatile loads, they are always ok.
75 if (!LI->isSimple()) return false;
79 if (isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I)) {
80 // If uses of the bitcast are ok, we are ok.
81 ValuesToInspect.emplace_back(I, IsOffset);
84 if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
85 // If the GEP has all zero indices, it doesn't offset the pointer. If it
87 ValuesToInspect.emplace_back(I, IsOffset || !GEP->hasAllZeroIndices());
91 if (auto *Call = dyn_cast<CallBase>(I)) {
92 // If this is the function being called then we treat it like a load and
94 if (Call->isCallee(&U))
97 unsigned DataOpNo = Call->getDataOperandNo(&U);
98 bool IsArgOperand = Call->isArgOperand(&U);
100 // Inalloca arguments are clobbered by the call.
101 if (IsArgOperand && Call->isInAllocaArgument(DataOpNo))
104 // If this is a readonly/readnone call site, then we know it is just a
105 // load (but one that potentially returns the value itself), so we can
106 // ignore it if we know that the value isn't captured.
107 if (Call->onlyReadsMemory() &&
108 (Call->use_empty() || Call->doesNotCapture(DataOpNo)))
111 // If this is being passed as a byval argument, the caller is making a
112 // copy, so it is only a read of the alloca.
113 if (IsArgOperand && Call->isByValArgument(DataOpNo))
117 // Lifetime intrinsics can be handled by the caller.
118 if (I->isLifetimeStartOrEnd()) {
119 assert(I->use_empty() && "Lifetime markers have no result to use!");
120 ToDelete.push_back(I);
124 // If this is isn't our memcpy/memmove, reject it as something we can't
126 MemTransferInst *MI = dyn_cast<MemTransferInst>(I);
130 // If the transfer is using the alloca as a source of the transfer, then
131 // ignore it since it is a load (unless the transfer is volatile).
132 if (U.getOperandNo() == 1) {
133 if (MI->isVolatile()) return false;
137 // If we already have seen a copy, reject the second one.
138 if (TheCopy) return false;
140 // If the pointer has been offset from the start of the alloca, we can't
141 // safely handle this.
142 if (IsOffset) return false;
144 // If the memintrinsic isn't using the alloca as the dest, reject it.
145 if (U.getOperandNo() != 0) return false;
147 // If the source of the memcpy/move is not a constant global, reject it.
148 if (!pointsToConstantGlobal(MI->getSource()))
151 // Otherwise, the transform is safe. Remember the copy instruction.
158 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
159 /// modified by a copy from a constant global. If we can prove this, we can
160 /// replace any uses of the alloca with uses of the global directly.
161 static MemTransferInst *
162 isOnlyCopiedFromConstantGlobal(AllocaInst *AI,
163 SmallVectorImpl<Instruction *> &ToDelete) {
164 MemTransferInst *TheCopy = nullptr;
165 if (isOnlyCopiedFromConstantGlobal(AI, TheCopy, ToDelete))
170 /// Returns true if V is dereferenceable for size of alloca.
171 static bool isDereferenceableForAllocaSize(const Value *V, const AllocaInst *AI,
172 const DataLayout &DL) {
173 if (AI->isArrayAllocation())
175 uint64_t AllocaSize = DL.getTypeStoreSize(AI->getAllocatedType());
178 return isDereferenceableAndAlignedPointer(V, AI->getAlignment(),
179 APInt(64, AllocaSize), DL);
182 static Instruction *simplifyAllocaArraySize(InstCombiner &IC, AllocaInst &AI) {
183 // Check for array size of 1 (scalar allocation).
184 if (!AI.isArrayAllocation()) {
185 // i32 1 is the canonical array size for scalar allocations.
186 if (AI.getArraySize()->getType()->isIntegerTy(32))
190 Value *V = IC.Builder.getInt32(1);
195 // Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1
196 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
197 if (C->getValue().getActiveBits() <= 64) {
198 Type *NewTy = ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
199 AllocaInst *New = IC.Builder.CreateAlloca(NewTy, nullptr, AI.getName());
200 New->setAlignment(AI.getAlignment());
202 // Scan to the end of the allocation instructions, to skip over a block of
203 // allocas if possible...also skip interleaved debug info
205 BasicBlock::iterator It(New);
206 while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It))
209 // Now that I is pointing to the first non-allocation-inst in the block,
210 // insert our getelementptr instruction...
212 Type *IdxTy = IC.getDataLayout().getIntPtrType(AI.getType());
213 Value *NullIdx = Constant::getNullValue(IdxTy);
214 Value *Idx[2] = {NullIdx, NullIdx};
215 Instruction *GEP = GetElementPtrInst::CreateInBounds(
216 NewTy, New, Idx, New->getName() + ".sub");
217 IC.InsertNewInstBefore(GEP, *It);
219 // Now make everything use the getelementptr instead of the original
221 return IC.replaceInstUsesWith(AI, GEP);
225 if (isa<UndefValue>(AI.getArraySize()))
226 return IC.replaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
228 // Ensure that the alloca array size argument has type intptr_t, so that
229 // any casting is exposed early.
230 Type *IntPtrTy = IC.getDataLayout().getIntPtrType(AI.getType());
231 if (AI.getArraySize()->getType() != IntPtrTy) {
232 Value *V = IC.Builder.CreateIntCast(AI.getArraySize(), IntPtrTy, false);
241 // If I and V are pointers in different address space, it is not allowed to
242 // use replaceAllUsesWith since I and V have different types. A
243 // non-target-specific transformation should not use addrspacecast on V since
244 // the two address space may be disjoint depending on target.
246 // This class chases down uses of the old pointer until reaching the load
247 // instructions, then replaces the old pointer in the load instructions with
248 // the new pointer. If during the chasing it sees bitcast or GEP, it will
249 // create new bitcast or GEP with the new pointer and use them in the load
251 class PointerReplacer {
253 PointerReplacer(InstCombiner &IC) : IC(IC) {}
254 void replacePointer(Instruction &I, Value *V);
257 void findLoadAndReplace(Instruction &I);
258 void replace(Instruction *I);
259 Value *getReplacement(Value *I);
261 SmallVector<Instruction *, 4> Path;
262 MapVector<Value *, Value *> WorkMap;
265 } // end anonymous namespace
267 void PointerReplacer::findLoadAndReplace(Instruction &I) {
268 for (auto U : I.users()) {
269 auto *Inst = dyn_cast<Instruction>(&*U);
272 LLVM_DEBUG(dbgs() << "Found pointer user: " << *U << '\n');
273 if (isa<LoadInst>(Inst)) {
277 } else if (isa<GetElementPtrInst>(Inst) || isa<BitCastInst>(Inst)) {
278 Path.push_back(Inst);
279 findLoadAndReplace(*Inst);
287 Value *PointerReplacer::getReplacement(Value *V) {
288 auto Loc = WorkMap.find(V);
289 if (Loc != WorkMap.end())
294 void PointerReplacer::replace(Instruction *I) {
295 if (getReplacement(I))
298 if (auto *LT = dyn_cast<LoadInst>(I)) {
299 auto *V = getReplacement(LT->getPointerOperand());
300 assert(V && "Operand not replaced");
301 auto *NewI = new LoadInst(I->getType(), V);
303 IC.InsertNewInstWith(NewI, *LT);
304 IC.replaceInstUsesWith(*LT, NewI);
306 } else if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
307 auto *V = getReplacement(GEP->getPointerOperand());
308 assert(V && "Operand not replaced");
309 SmallVector<Value *, 8> Indices;
310 Indices.append(GEP->idx_begin(), GEP->idx_end());
311 auto *NewI = GetElementPtrInst::Create(
312 V->getType()->getPointerElementType(), V, Indices);
313 IC.InsertNewInstWith(NewI, *GEP);
316 } else if (auto *BC = dyn_cast<BitCastInst>(I)) {
317 auto *V = getReplacement(BC->getOperand(0));
318 assert(V && "Operand not replaced");
319 auto *NewT = PointerType::get(BC->getType()->getPointerElementType(),
320 V->getType()->getPointerAddressSpace());
321 auto *NewI = new BitCastInst(V, NewT);
322 IC.InsertNewInstWith(NewI, *BC);
326 llvm_unreachable("should never reach here");
330 void PointerReplacer::replacePointer(Instruction &I, Value *V) {
332 auto *PT = cast<PointerType>(I.getType());
333 auto *NT = cast<PointerType>(V->getType());
334 assert(PT != NT && PT->getElementType() == NT->getElementType() &&
338 findLoadAndReplace(I);
341 Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
342 if (auto *I = simplifyAllocaArraySize(*this, AI))
345 if (AI.getAllocatedType()->isSized()) {
346 // If the alignment is 0 (unspecified), assign it the preferred alignment.
347 if (AI.getAlignment() == 0)
348 AI.setAlignment(DL.getPrefTypeAlignment(AI.getAllocatedType()));
350 // Move all alloca's of zero byte objects to the entry block and merge them
351 // together. Note that we only do this for alloca's, because malloc should
352 // allocate and return a unique pointer, even for a zero byte allocation.
353 if (DL.getTypeAllocSize(AI.getAllocatedType()) == 0) {
354 // For a zero sized alloca there is no point in doing an array allocation.
355 // This is helpful if the array size is a complicated expression not used
357 if (AI.isArrayAllocation()) {
358 AI.setOperand(0, ConstantInt::get(AI.getArraySize()->getType(), 1));
362 // Get the first instruction in the entry block.
363 BasicBlock &EntryBlock = AI.getParent()->getParent()->getEntryBlock();
364 Instruction *FirstInst = EntryBlock.getFirstNonPHIOrDbg();
365 if (FirstInst != &AI) {
366 // If the entry block doesn't start with a zero-size alloca then move
367 // this one to the start of the entry block. There is no problem with
368 // dominance as the array size was forced to a constant earlier already.
369 AllocaInst *EntryAI = dyn_cast<AllocaInst>(FirstInst);
370 if (!EntryAI || !EntryAI->getAllocatedType()->isSized() ||
371 DL.getTypeAllocSize(EntryAI->getAllocatedType()) != 0) {
372 AI.moveBefore(FirstInst);
376 // If the alignment of the entry block alloca is 0 (unspecified),
377 // assign it the preferred alignment.
378 if (EntryAI->getAlignment() == 0)
379 EntryAI->setAlignment(
380 DL.getPrefTypeAlignment(EntryAI->getAllocatedType()));
381 // Replace this zero-sized alloca with the one at the start of the entry
382 // block after ensuring that the address will be aligned enough for both
384 unsigned MaxAlign = std::max(EntryAI->getAlignment(),
386 EntryAI->setAlignment(MaxAlign);
387 if (AI.getType() != EntryAI->getType())
388 return new BitCastInst(EntryAI, AI.getType());
389 return replaceInstUsesWith(AI, EntryAI);
394 if (AI.getAlignment()) {
395 // Check to see if this allocation is only modified by a memcpy/memmove from
396 // a constant global whose alignment is equal to or exceeds that of the
397 // allocation. If this is the case, we can change all users to use
398 // the constant global instead. This is commonly produced by the CFE by
399 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
400 // is only subsequently read.
401 SmallVector<Instruction *, 4> ToDelete;
402 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(&AI, ToDelete)) {
403 unsigned SourceAlign = getOrEnforceKnownAlignment(
404 Copy->getSource(), AI.getAlignment(), DL, &AI, &AC, &DT);
405 if (AI.getAlignment() <= SourceAlign &&
406 isDereferenceableForAllocaSize(Copy->getSource(), &AI, DL)) {
407 LLVM_DEBUG(dbgs() << "Found alloca equal to global: " << AI << '\n');
408 LLVM_DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
409 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
410 eraseInstFromFunction(*ToDelete[i]);
411 Constant *TheSrc = cast<Constant>(Copy->getSource());
412 auto *SrcTy = TheSrc->getType();
413 auto *DestTy = PointerType::get(AI.getType()->getPointerElementType(),
414 SrcTy->getPointerAddressSpace());
416 ConstantExpr::getPointerBitCastOrAddrSpaceCast(TheSrc, DestTy);
417 if (AI.getType()->getPointerAddressSpace() ==
418 SrcTy->getPointerAddressSpace()) {
419 Instruction *NewI = replaceInstUsesWith(AI, Cast);
420 eraseInstFromFunction(*Copy);
424 PointerReplacer PtrReplacer(*this);
425 PtrReplacer.replacePointer(AI, Cast);
432 // At last, use the generic allocation site handler to aggressively remove
434 return visitAllocSite(AI);
437 // Are we allowed to form a atomic load or store of this type?
438 static bool isSupportedAtomicType(Type *Ty) {
439 return Ty->isIntOrPtrTy() || Ty->isFloatingPointTy();
442 /// Helper to combine a load to a new type.
444 /// This just does the work of combining a load to a new type. It handles
445 /// metadata, etc., and returns the new instruction. The \c NewTy should be the
446 /// loaded *value* type. This will convert it to a pointer, cast the operand to
447 /// that pointer type, load it, etc.
449 /// Note that this will create all of the instructions with whatever insert
450 /// point the \c InstCombiner currently is using.
451 static LoadInst *combineLoadToNewType(InstCombiner &IC, LoadInst &LI, Type *NewTy,
452 const Twine &Suffix = "") {
453 assert((!LI.isAtomic() || isSupportedAtomicType(NewTy)) &&
454 "can't fold an atomic load to requested type");
456 Value *Ptr = LI.getPointerOperand();
457 unsigned AS = LI.getPointerAddressSpace();
458 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
459 LI.getAllMetadata(MD);
461 Value *NewPtr = nullptr;
462 if (!(match(Ptr, m_BitCast(m_Value(NewPtr))) &&
463 NewPtr->getType()->getPointerElementType() == NewTy &&
464 NewPtr->getType()->getPointerAddressSpace() == AS))
465 NewPtr = IC.Builder.CreateBitCast(Ptr, NewTy->getPointerTo(AS));
467 LoadInst *NewLoad = IC.Builder.CreateAlignedLoad(
468 NewTy, NewPtr, LI.getAlignment(), LI.isVolatile(), LI.getName() + Suffix);
469 NewLoad->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
470 MDBuilder MDB(NewLoad->getContext());
471 for (const auto &MDPair : MD) {
472 unsigned ID = MDPair.first;
473 MDNode *N = MDPair.second;
474 // Note, essentially every kind of metadata should be preserved here! This
475 // routine is supposed to clone a load instruction changing *only its type*.
476 // The only metadata it makes sense to drop is metadata which is invalidated
477 // when the pointer type changes. This should essentially never be the case
478 // in LLVM, but we explicitly switch over only known metadata to be
479 // conservatively correct. If you are adding metadata to LLVM which pertains
480 // to loads, you almost certainly want to add it here.
482 case LLVMContext::MD_dbg:
483 case LLVMContext::MD_tbaa:
484 case LLVMContext::MD_prof:
485 case LLVMContext::MD_fpmath:
486 case LLVMContext::MD_tbaa_struct:
487 case LLVMContext::MD_invariant_load:
488 case LLVMContext::MD_alias_scope:
489 case LLVMContext::MD_noalias:
490 case LLVMContext::MD_nontemporal:
491 case LLVMContext::MD_mem_parallel_loop_access:
492 case LLVMContext::MD_access_group:
493 // All of these directly apply.
494 NewLoad->setMetadata(ID, N);
497 case LLVMContext::MD_nonnull:
498 copyNonnullMetadata(LI, N, *NewLoad);
500 case LLVMContext::MD_align:
501 case LLVMContext::MD_dereferenceable:
502 case LLVMContext::MD_dereferenceable_or_null:
503 // These only directly apply if the new type is also a pointer.
504 if (NewTy->isPointerTy())
505 NewLoad->setMetadata(ID, N);
507 case LLVMContext::MD_range:
508 copyRangeMetadata(IC.getDataLayout(), LI, N, *NewLoad);
515 /// Combine a store to a new type.
517 /// Returns the newly created store instruction.
518 static StoreInst *combineStoreToNewValue(InstCombiner &IC, StoreInst &SI, Value *V) {
519 assert((!SI.isAtomic() || isSupportedAtomicType(V->getType())) &&
520 "can't fold an atomic store of requested type");
522 Value *Ptr = SI.getPointerOperand();
523 unsigned AS = SI.getPointerAddressSpace();
524 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
525 SI.getAllMetadata(MD);
527 StoreInst *NewStore = IC.Builder.CreateAlignedStore(
528 V, IC.Builder.CreateBitCast(Ptr, V->getType()->getPointerTo(AS)),
529 SI.getAlignment(), SI.isVolatile());
530 NewStore->setAtomic(SI.getOrdering(), SI.getSyncScopeID());
531 for (const auto &MDPair : MD) {
532 unsigned ID = MDPair.first;
533 MDNode *N = MDPair.second;
534 // Note, essentially every kind of metadata should be preserved here! This
535 // routine is supposed to clone a store instruction changing *only its
536 // type*. The only metadata it makes sense to drop is metadata which is
537 // invalidated when the pointer type changes. This should essentially
538 // never be the case in LLVM, but we explicitly switch over only known
539 // metadata to be conservatively correct. If you are adding metadata to
540 // LLVM which pertains to stores, you almost certainly want to add it
543 case LLVMContext::MD_dbg:
544 case LLVMContext::MD_tbaa:
545 case LLVMContext::MD_prof:
546 case LLVMContext::MD_fpmath:
547 case LLVMContext::MD_tbaa_struct:
548 case LLVMContext::MD_alias_scope:
549 case LLVMContext::MD_noalias:
550 case LLVMContext::MD_nontemporal:
551 case LLVMContext::MD_mem_parallel_loop_access:
552 case LLVMContext::MD_access_group:
553 // All of these directly apply.
554 NewStore->setMetadata(ID, N);
556 case LLVMContext::MD_invariant_load:
557 case LLVMContext::MD_nonnull:
558 case LLVMContext::MD_range:
559 case LLVMContext::MD_align:
560 case LLVMContext::MD_dereferenceable:
561 case LLVMContext::MD_dereferenceable_or_null:
562 // These don't apply for stores.
570 /// Returns true if instruction represent minmax pattern like:
571 /// select ((cmp load V1, load V2), V1, V2).
572 static bool isMinMaxWithLoads(Value *V) {
573 assert(V->getType()->isPointerTy() && "Expected pointer type.");
574 // Ignore possible ty* to ixx* bitcast.
575 V = peekThroughBitcast(V);
576 // Check that select is select ((cmp load V1, load V2), V1, V2) - minmax
578 CmpInst::Predicate Pred;
583 if (!match(V, m_Select(m_Cmp(Pred, m_Instruction(L1), m_Instruction(L2)),
584 m_Value(LHS), m_Value(RHS))))
586 return (match(L1, m_Load(m_Specific(LHS))) &&
587 match(L2, m_Load(m_Specific(RHS)))) ||
588 (match(L1, m_Load(m_Specific(RHS))) &&
589 match(L2, m_Load(m_Specific(LHS))));
592 /// Combine loads to match the type of their uses' value after looking
593 /// through intervening bitcasts.
595 /// The core idea here is that if the result of a load is used in an operation,
596 /// we should load the type most conducive to that operation. For example, when
597 /// loading an integer and converting that immediately to a pointer, we should
598 /// instead directly load a pointer.
600 /// However, this routine must never change the width of a load or the number of
601 /// loads as that would introduce a semantic change. This combine is expected to
602 /// be a semantic no-op which just allows loads to more closely model the types
603 /// of their consuming operations.
605 /// Currently, we also refuse to change the precise type used for an atomic load
606 /// or a volatile load. This is debatable, and might be reasonable to change
607 /// later. However, it is risky in case some backend or other part of LLVM is
608 /// relying on the exact type loaded to select appropriate atomic operations.
609 static Instruction *combineLoadToOperationType(InstCombiner &IC, LoadInst &LI) {
610 // FIXME: We could probably with some care handle both volatile and ordered
611 // atomic loads here but it isn't clear that this is important.
612 if (!LI.isUnordered())
618 // swifterror values can't be bitcasted.
619 if (LI.getPointerOperand()->isSwiftError())
622 Type *Ty = LI.getType();
623 const DataLayout &DL = IC.getDataLayout();
625 // Try to canonicalize loads which are only ever stored to operate over
626 // integers instead of any other type. We only do this when the loaded type
627 // is sized and has a size exactly the same as its store size and the store
628 // size is a legal integer type.
629 // Do not perform canonicalization if minmax pattern is found (to avoid
631 if (!Ty->isIntegerTy() && Ty->isSized() &&
632 DL.isLegalInteger(DL.getTypeStoreSizeInBits(Ty)) &&
633 DL.typeSizeEqualsStoreSize(Ty) &&
634 !DL.isNonIntegralPointerType(Ty) &&
636 peekThroughBitcast(LI.getPointerOperand(), /*OneUseOnly=*/true))) {
637 if (all_of(LI.users(), [&LI](User *U) {
638 auto *SI = dyn_cast<StoreInst>(U);
639 return SI && SI->getPointerOperand() != &LI &&
640 !SI->getPointerOperand()->isSwiftError();
642 LoadInst *NewLoad = combineLoadToNewType(
644 Type::getIntNTy(LI.getContext(), DL.getTypeStoreSizeInBits(Ty)));
645 // Replace all the stores with stores of the newly loaded value.
646 for (auto UI = LI.user_begin(), UE = LI.user_end(); UI != UE;) {
647 auto *SI = cast<StoreInst>(*UI++);
648 IC.Builder.SetInsertPoint(SI);
649 combineStoreToNewValue(IC, *SI, NewLoad);
650 IC.eraseInstFromFunction(*SI);
652 assert(LI.use_empty() && "Failed to remove all users of the load!");
653 // Return the old load so the combiner can delete it safely.
658 // Fold away bit casts of the loaded value by loading the desired type.
659 // We can do this for BitCastInsts as well as casts from and to pointer types,
660 // as long as those are noops (i.e., the source or dest type have the same
661 // bitwidth as the target's pointers).
663 if (auto* CI = dyn_cast<CastInst>(LI.user_back()))
664 if (CI->isNoopCast(DL))
665 if (!LI.isAtomic() || isSupportedAtomicType(CI->getDestTy())) {
666 LoadInst *NewLoad = combineLoadToNewType(IC, LI, CI->getDestTy());
667 CI->replaceAllUsesWith(NewLoad);
668 IC.eraseInstFromFunction(*CI);
672 // FIXME: We should also canonicalize loads of vectors when their elements are
673 // cast to other types.
677 static Instruction *unpackLoadToAggregate(InstCombiner &IC, LoadInst &LI) {
678 // FIXME: We could probably with some care handle both volatile and atomic
679 // stores here but it isn't clear that this is important.
683 Type *T = LI.getType();
684 if (!T->isAggregateType())
687 StringRef Name = LI.getName();
688 assert(LI.getAlignment() && "Alignment must be set at this point");
690 if (auto *ST = dyn_cast<StructType>(T)) {
691 // If the struct only have one element, we unpack.
692 auto NumElements = ST->getNumElements();
693 if (NumElements == 1) {
694 LoadInst *NewLoad = combineLoadToNewType(IC, LI, ST->getTypeAtIndex(0U),
697 LI.getAAMetadata(AAMD);
698 NewLoad->setAAMetadata(AAMD);
699 return IC.replaceInstUsesWith(LI, IC.Builder.CreateInsertValue(
700 UndefValue::get(T), NewLoad, 0, Name));
703 // We don't want to break loads with padding here as we'd loose
704 // the knowledge that padding exists for the rest of the pipeline.
705 const DataLayout &DL = IC.getDataLayout();
706 auto *SL = DL.getStructLayout(ST);
707 if (SL->hasPadding())
710 auto Align = LI.getAlignment();
712 Align = DL.getABITypeAlignment(ST);
714 auto *Addr = LI.getPointerOperand();
715 auto *IdxType = Type::getInt32Ty(T->getContext());
716 auto *Zero = ConstantInt::get(IdxType, 0);
718 Value *V = UndefValue::get(T);
719 for (unsigned i = 0; i < NumElements; i++) {
720 Value *Indices[2] = {
722 ConstantInt::get(IdxType, i),
724 auto *Ptr = IC.Builder.CreateInBoundsGEP(ST, Addr, makeArrayRef(Indices),
726 auto EltAlign = MinAlign(Align, SL->getElementOffset(i));
727 auto *L = IC.Builder.CreateAlignedLoad(ST->getElementType(i), Ptr,
728 EltAlign, Name + ".unpack");
729 // Propagate AA metadata. It'll still be valid on the narrowed load.
731 LI.getAAMetadata(AAMD);
732 L->setAAMetadata(AAMD);
733 V = IC.Builder.CreateInsertValue(V, L, i);
737 return IC.replaceInstUsesWith(LI, V);
740 if (auto *AT = dyn_cast<ArrayType>(T)) {
741 auto *ET = AT->getElementType();
742 auto NumElements = AT->getNumElements();
743 if (NumElements == 1) {
744 LoadInst *NewLoad = combineLoadToNewType(IC, LI, ET, ".unpack");
746 LI.getAAMetadata(AAMD);
747 NewLoad->setAAMetadata(AAMD);
748 return IC.replaceInstUsesWith(LI, IC.Builder.CreateInsertValue(
749 UndefValue::get(T), NewLoad, 0, Name));
752 // Bail out if the array is too large. Ideally we would like to optimize
753 // arrays of arbitrary size but this has a terrible impact on compile time.
754 // The threshold here is chosen arbitrarily, maybe needs a little bit of
756 if (NumElements > IC.MaxArraySizeForCombine)
759 const DataLayout &DL = IC.getDataLayout();
760 auto EltSize = DL.getTypeAllocSize(ET);
761 auto Align = LI.getAlignment();
763 Align = DL.getABITypeAlignment(T);
765 auto *Addr = LI.getPointerOperand();
766 auto *IdxType = Type::getInt64Ty(T->getContext());
767 auto *Zero = ConstantInt::get(IdxType, 0);
769 Value *V = UndefValue::get(T);
771 for (uint64_t i = 0; i < NumElements; i++) {
772 Value *Indices[2] = {
774 ConstantInt::get(IdxType, i),
776 auto *Ptr = IC.Builder.CreateInBoundsGEP(AT, Addr, makeArrayRef(Indices),
778 auto *L = IC.Builder.CreateAlignedLoad(
779 AT->getElementType(), Ptr, MinAlign(Align, Offset), Name + ".unpack");
781 LI.getAAMetadata(AAMD);
782 L->setAAMetadata(AAMD);
783 V = IC.Builder.CreateInsertValue(V, L, i);
788 return IC.replaceInstUsesWith(LI, V);
794 // If we can determine that all possible objects pointed to by the provided
795 // pointer value are, not only dereferenceable, but also definitively less than
796 // or equal to the provided maximum size, then return true. Otherwise, return
797 // false (constant global values and allocas fall into this category).
799 // FIXME: This should probably live in ValueTracking (or similar).
800 static bool isObjectSizeLessThanOrEq(Value *V, uint64_t MaxSize,
801 const DataLayout &DL) {
802 SmallPtrSet<Value *, 4> Visited;
803 SmallVector<Value *, 4> Worklist(1, V);
806 Value *P = Worklist.pop_back_val();
807 P = P->stripPointerCasts();
809 if (!Visited.insert(P).second)
812 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
813 Worklist.push_back(SI->getTrueValue());
814 Worklist.push_back(SI->getFalseValue());
818 if (PHINode *PN = dyn_cast<PHINode>(P)) {
819 for (Value *IncValue : PN->incoming_values())
820 Worklist.push_back(IncValue);
824 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(P)) {
825 if (GA->isInterposable())
827 Worklist.push_back(GA->getAliasee());
831 // If we know how big this object is, and it is less than MaxSize, continue
832 // searching. Otherwise, return false.
833 if (AllocaInst *AI = dyn_cast<AllocaInst>(P)) {
834 if (!AI->getAllocatedType()->isSized())
837 ConstantInt *CS = dyn_cast<ConstantInt>(AI->getArraySize());
841 uint64_t TypeSize = DL.getTypeAllocSize(AI->getAllocatedType());
842 // Make sure that, even if the multiplication below would wrap as an
843 // uint64_t, we still do the right thing.
844 if ((CS->getValue().zextOrSelf(128)*APInt(128, TypeSize)).ugt(MaxSize))
849 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
850 if (!GV->hasDefinitiveInitializer() || !GV->isConstant())
853 uint64_t InitSize = DL.getTypeAllocSize(GV->getValueType());
854 if (InitSize > MaxSize)
860 } while (!Worklist.empty());
865 // If we're indexing into an object of a known size, and the outer index is
866 // not a constant, but having any value but zero would lead to undefined
867 // behavior, replace it with zero.
869 // For example, if we have:
870 // @f.a = private unnamed_addr constant [1 x i32] [i32 12], align 4
872 // %arrayidx = getelementptr inbounds [1 x i32]* @f.a, i64 0, i64 %x
873 // ... = load i32* %arrayidx, align 4
874 // Then we know that we can replace %x in the GEP with i64 0.
876 // FIXME: We could fold any GEP index to zero that would cause UB if it were
877 // not zero. Currently, we only handle the first such index. Also, we could
878 // also search through non-zero constant indices if we kept track of the
879 // offsets those indices implied.
880 static bool canReplaceGEPIdxWithZero(InstCombiner &IC, GetElementPtrInst *GEPI,
881 Instruction *MemI, unsigned &Idx) {
882 if (GEPI->getNumOperands() < 2)
885 // Find the first non-zero index of a GEP. If all indices are zero, return
886 // one past the last index.
887 auto FirstNZIdx = [](const GetElementPtrInst *GEPI) {
889 for (unsigned IE = GEPI->getNumOperands(); I != IE; ++I) {
890 Value *V = GEPI->getOperand(I);
891 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V))
901 // Skip through initial 'zero' indices, and find the corresponding pointer
902 // type. See if the next index is not a constant.
903 Idx = FirstNZIdx(GEPI);
904 if (Idx == GEPI->getNumOperands())
906 if (isa<Constant>(GEPI->getOperand(Idx)))
909 SmallVector<Value *, 4> Ops(GEPI->idx_begin(), GEPI->idx_begin() + Idx);
911 GetElementPtrInst::getIndexedType(GEPI->getSourceElementType(), Ops);
912 if (!AllocTy || !AllocTy->isSized())
914 const DataLayout &DL = IC.getDataLayout();
915 uint64_t TyAllocSize = DL.getTypeAllocSize(AllocTy);
917 // If there are more indices after the one we might replace with a zero, make
918 // sure they're all non-negative. If any of them are negative, the overall
919 // address being computed might be before the base address determined by the
920 // first non-zero index.
921 auto IsAllNonNegative = [&]() {
922 for (unsigned i = Idx+1, e = GEPI->getNumOperands(); i != e; ++i) {
923 KnownBits Known = IC.computeKnownBits(GEPI->getOperand(i), 0, MemI);
924 if (Known.isNonNegative())
932 // FIXME: If the GEP is not inbounds, and there are extra indices after the
933 // one we'll replace, those could cause the address computation to wrap
934 // (rendering the IsAllNonNegative() check below insufficient). We can do
935 // better, ignoring zero indices (and other indices we can prove small
936 // enough not to wrap).
937 if (Idx+1 != GEPI->getNumOperands() && !GEPI->isInBounds())
940 // Note that isObjectSizeLessThanOrEq will return true only if the pointer is
941 // also known to be dereferenceable.
942 return isObjectSizeLessThanOrEq(GEPI->getOperand(0), TyAllocSize, DL) &&
946 // If we're indexing into an object with a variable index for the memory
947 // access, but the object has only one element, we can assume that the index
948 // will always be zero. If we replace the GEP, return it.
949 template <typename T>
950 static Instruction *replaceGEPIdxWithZero(InstCombiner &IC, Value *Ptr,
952 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr)) {
954 if (canReplaceGEPIdxWithZero(IC, GEPI, &MemI, Idx)) {
955 Instruction *NewGEPI = GEPI->clone();
956 NewGEPI->setOperand(Idx,
957 ConstantInt::get(GEPI->getOperand(Idx)->getType(), 0));
958 NewGEPI->insertBefore(GEPI);
959 MemI.setOperand(MemI.getPointerOperandIndex(), NewGEPI);
967 static bool canSimplifyNullStoreOrGEP(StoreInst &SI) {
968 if (NullPointerIsDefined(SI.getFunction(), SI.getPointerAddressSpace()))
971 auto *Ptr = SI.getPointerOperand();
972 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr))
973 Ptr = GEPI->getOperand(0);
974 return (isa<ConstantPointerNull>(Ptr) &&
975 !NullPointerIsDefined(SI.getFunction(), SI.getPointerAddressSpace()));
978 static bool canSimplifyNullLoadOrGEP(LoadInst &LI, Value *Op) {
979 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
980 const Value *GEPI0 = GEPI->getOperand(0);
981 if (isa<ConstantPointerNull>(GEPI0) &&
982 !NullPointerIsDefined(LI.getFunction(), GEPI->getPointerAddressSpace()))
985 if (isa<UndefValue>(Op) ||
986 (isa<ConstantPointerNull>(Op) &&
987 !NullPointerIsDefined(LI.getFunction(), LI.getPointerAddressSpace())))
992 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
993 Value *Op = LI.getOperand(0);
995 // Try to canonicalize the loaded type.
996 if (Instruction *Res = combineLoadToOperationType(*this, LI))
999 // Attempt to improve the alignment.
1000 unsigned KnownAlign = getOrEnforceKnownAlignment(
1001 Op, DL.getPrefTypeAlignment(LI.getType()), DL, &LI, &AC, &DT);
1002 unsigned LoadAlign = LI.getAlignment();
1003 unsigned EffectiveLoadAlign =
1004 LoadAlign != 0 ? LoadAlign : DL.getABITypeAlignment(LI.getType());
1006 if (KnownAlign > EffectiveLoadAlign)
1007 LI.setAlignment(KnownAlign);
1008 else if (LoadAlign == 0)
1009 LI.setAlignment(EffectiveLoadAlign);
1011 // Replace GEP indices if possible.
1012 if (Instruction *NewGEPI = replaceGEPIdxWithZero(*this, Op, LI)) {
1013 Worklist.Add(NewGEPI);
1017 if (Instruction *Res = unpackLoadToAggregate(*this, LI))
1020 // Do really simple store-to-load forwarding and load CSE, to catch cases
1021 // where there are several consecutive memory accesses to the same location,
1022 // separated by a few arithmetic operations.
1023 BasicBlock::iterator BBI(LI);
1024 bool IsLoadCSE = false;
1025 if (Value *AvailableVal = FindAvailableLoadedValue(
1026 &LI, LI.getParent(), BBI, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1028 combineMetadataForCSE(cast<LoadInst>(AvailableVal), &LI, false);
1030 return replaceInstUsesWith(
1031 LI, Builder.CreateBitOrPointerCast(AvailableVal, LI.getType(),
1032 LI.getName() + ".cast"));
1035 // None of the following transforms are legal for volatile/ordered atomic
1036 // loads. Most of them do apply for unordered atomics.
1037 if (!LI.isUnordered()) return nullptr;
1039 // load(gep null, ...) -> unreachable
1040 // load null/undef -> unreachable
1041 // TODO: Consider a target hook for valid address spaces for this xforms.
1042 if (canSimplifyNullLoadOrGEP(LI, Op)) {
1043 // Insert a new store to null instruction before the load to indicate
1044 // that this code is not reachable. We do this instead of inserting
1045 // an unreachable instruction directly because we cannot modify the
1047 StoreInst *SI = new StoreInst(UndefValue::get(LI.getType()),
1048 Constant::getNullValue(Op->getType()), &LI);
1049 SI->setDebugLoc(LI.getDebugLoc());
1050 return replaceInstUsesWith(LI, UndefValue::get(LI.getType()));
1053 if (Op->hasOneUse()) {
1054 // Change select and PHI nodes to select values instead of addresses: this
1055 // helps alias analysis out a lot, allows many others simplifications, and
1056 // exposes redundancy in the code.
1058 // Note that we cannot do the transformation unless we know that the
1059 // introduced loads cannot trap! Something like this is valid as long as
1060 // the condition is always false: load (select bool %C, int* null, int* %G),
1061 // but it would not be valid if we transformed it to load from null
1064 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
1065 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
1066 unsigned Align = LI.getAlignment();
1067 if (isSafeToLoadUnconditionally(SI->getOperand(1), LI.getType(), Align,
1069 isSafeToLoadUnconditionally(SI->getOperand(2), LI.getType(), Align,
1072 Builder.CreateLoad(LI.getType(), SI->getOperand(1),
1073 SI->getOperand(1)->getName() + ".val");
1075 Builder.CreateLoad(LI.getType(), SI->getOperand(2),
1076 SI->getOperand(2)->getName() + ".val");
1077 assert(LI.isUnordered() && "implied by above");
1078 V1->setAlignment(Align);
1079 V1->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
1080 V2->setAlignment(Align);
1081 V2->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
1082 return SelectInst::Create(SI->getCondition(), V1, V2);
1085 // load (select (cond, null, P)) -> load P
1086 if (isa<ConstantPointerNull>(SI->getOperand(1)) &&
1087 !NullPointerIsDefined(SI->getFunction(),
1088 LI.getPointerAddressSpace())) {
1089 LI.setOperand(0, SI->getOperand(2));
1093 // load (select (cond, P, null)) -> load P
1094 if (isa<ConstantPointerNull>(SI->getOperand(2)) &&
1095 !NullPointerIsDefined(SI->getFunction(),
1096 LI.getPointerAddressSpace())) {
1097 LI.setOperand(0, SI->getOperand(1));
1105 /// Look for extractelement/insertvalue sequence that acts like a bitcast.
1107 /// \returns underlying value that was "cast", or nullptr otherwise.
1109 /// For example, if we have:
1111 /// %E0 = extractelement <2 x double> %U, i32 0
1112 /// %V0 = insertvalue [2 x double] undef, double %E0, 0
1113 /// %E1 = extractelement <2 x double> %U, i32 1
1114 /// %V1 = insertvalue [2 x double] %V0, double %E1, 1
1116 /// and the layout of a <2 x double> is isomorphic to a [2 x double],
1117 /// then %V1 can be safely approximated by a conceptual "bitcast" of %U.
1118 /// Note that %U may contain non-undef values where %V1 has undef.
1119 static Value *likeBitCastFromVector(InstCombiner &IC, Value *V) {
1121 while (auto *IV = dyn_cast<InsertValueInst>(V)) {
1122 auto *E = dyn_cast<ExtractElementInst>(IV->getInsertedValueOperand());
1125 auto *W = E->getVectorOperand();
1130 auto *CI = dyn_cast<ConstantInt>(E->getIndexOperand());
1131 if (!CI || IV->getNumIndices() != 1 || CI->getZExtValue() != *IV->idx_begin())
1133 V = IV->getAggregateOperand();
1135 if (!isa<UndefValue>(V) ||!U)
1138 auto *UT = cast<VectorType>(U->getType());
1139 auto *VT = V->getType();
1140 // Check that types UT and VT are bitwise isomorphic.
1141 const auto &DL = IC.getDataLayout();
1142 if (DL.getTypeStoreSizeInBits(UT) != DL.getTypeStoreSizeInBits(VT)) {
1145 if (auto *AT = dyn_cast<ArrayType>(VT)) {
1146 if (AT->getNumElements() != UT->getNumElements())
1149 auto *ST = cast<StructType>(VT);
1150 if (ST->getNumElements() != UT->getNumElements())
1152 for (const auto *EltT : ST->elements()) {
1153 if (EltT != UT->getElementType())
1160 /// Combine stores to match the type of value being stored.
1162 /// The core idea here is that the memory does not have any intrinsic type and
1163 /// where we can we should match the type of a store to the type of value being
1166 /// However, this routine must never change the width of a store or the number of
1167 /// stores as that would introduce a semantic change. This combine is expected to
1168 /// be a semantic no-op which just allows stores to more closely model the types
1169 /// of their incoming values.
1171 /// Currently, we also refuse to change the precise type used for an atomic or
1172 /// volatile store. This is debatable, and might be reasonable to change later.
1173 /// However, it is risky in case some backend or other part of LLVM is relying
1174 /// on the exact type stored to select appropriate atomic operations.
1176 /// \returns true if the store was successfully combined away. This indicates
1177 /// the caller must erase the store instruction. We have to let the caller erase
1178 /// the store instruction as otherwise there is no way to signal whether it was
1179 /// combined or not: IC.EraseInstFromFunction returns a null pointer.
1180 static bool combineStoreToValueType(InstCombiner &IC, StoreInst &SI) {
1181 // FIXME: We could probably with some care handle both volatile and ordered
1182 // atomic stores here but it isn't clear that this is important.
1183 if (!SI.isUnordered())
1186 // swifterror values can't be bitcasted.
1187 if (SI.getPointerOperand()->isSwiftError())
1190 Value *V = SI.getValueOperand();
1192 // Fold away bit casts of the stored value by storing the original type.
1193 if (auto *BC = dyn_cast<BitCastInst>(V)) {
1194 V = BC->getOperand(0);
1195 if (!SI.isAtomic() || isSupportedAtomicType(V->getType())) {
1196 combineStoreToNewValue(IC, SI, V);
1201 if (Value *U = likeBitCastFromVector(IC, V))
1202 if (!SI.isAtomic() || isSupportedAtomicType(U->getType())) {
1203 combineStoreToNewValue(IC, SI, U);
1207 // FIXME: We should also canonicalize stores of vectors when their elements
1208 // are cast to other types.
1212 static bool unpackStoreToAggregate(InstCombiner &IC, StoreInst &SI) {
1213 // FIXME: We could probably with some care handle both volatile and atomic
1214 // stores here but it isn't clear that this is important.
1218 Value *V = SI.getValueOperand();
1219 Type *T = V->getType();
1221 if (!T->isAggregateType())
1224 if (auto *ST = dyn_cast<StructType>(T)) {
1225 // If the struct only have one element, we unpack.
1226 unsigned Count = ST->getNumElements();
1228 V = IC.Builder.CreateExtractValue(V, 0);
1229 combineStoreToNewValue(IC, SI, V);
1233 // We don't want to break loads with padding here as we'd loose
1234 // the knowledge that padding exists for the rest of the pipeline.
1235 const DataLayout &DL = IC.getDataLayout();
1236 auto *SL = DL.getStructLayout(ST);
1237 if (SL->hasPadding())
1240 auto Align = SI.getAlignment();
1242 Align = DL.getABITypeAlignment(ST);
1244 SmallString<16> EltName = V->getName();
1246 auto *Addr = SI.getPointerOperand();
1247 SmallString<16> AddrName = Addr->getName();
1248 AddrName += ".repack";
1250 auto *IdxType = Type::getInt32Ty(ST->getContext());
1251 auto *Zero = ConstantInt::get(IdxType, 0);
1252 for (unsigned i = 0; i < Count; i++) {
1253 Value *Indices[2] = {
1255 ConstantInt::get(IdxType, i),
1257 auto *Ptr = IC.Builder.CreateInBoundsGEP(ST, Addr, makeArrayRef(Indices),
1259 auto *Val = IC.Builder.CreateExtractValue(V, i, EltName);
1260 auto EltAlign = MinAlign(Align, SL->getElementOffset(i));
1261 llvm::Instruction *NS = IC.Builder.CreateAlignedStore(Val, Ptr, EltAlign);
1263 SI.getAAMetadata(AAMD);
1264 NS->setAAMetadata(AAMD);
1270 if (auto *AT = dyn_cast<ArrayType>(T)) {
1271 // If the array only have one element, we unpack.
1272 auto NumElements = AT->getNumElements();
1273 if (NumElements == 1) {
1274 V = IC.Builder.CreateExtractValue(V, 0);
1275 combineStoreToNewValue(IC, SI, V);
1279 // Bail out if the array is too large. Ideally we would like to optimize
1280 // arrays of arbitrary size but this has a terrible impact on compile time.
1281 // The threshold here is chosen arbitrarily, maybe needs a little bit of
1283 if (NumElements > IC.MaxArraySizeForCombine)
1286 const DataLayout &DL = IC.getDataLayout();
1287 auto EltSize = DL.getTypeAllocSize(AT->getElementType());
1288 auto Align = SI.getAlignment();
1290 Align = DL.getABITypeAlignment(T);
1292 SmallString<16> EltName = V->getName();
1294 auto *Addr = SI.getPointerOperand();
1295 SmallString<16> AddrName = Addr->getName();
1296 AddrName += ".repack";
1298 auto *IdxType = Type::getInt64Ty(T->getContext());
1299 auto *Zero = ConstantInt::get(IdxType, 0);
1301 uint64_t Offset = 0;
1302 for (uint64_t i = 0; i < NumElements; i++) {
1303 Value *Indices[2] = {
1305 ConstantInt::get(IdxType, i),
1307 auto *Ptr = IC.Builder.CreateInBoundsGEP(AT, Addr, makeArrayRef(Indices),
1309 auto *Val = IC.Builder.CreateExtractValue(V, i, EltName);
1310 auto EltAlign = MinAlign(Align, Offset);
1311 Instruction *NS = IC.Builder.CreateAlignedStore(Val, Ptr, EltAlign);
1313 SI.getAAMetadata(AAMD);
1314 NS->setAAMetadata(AAMD);
1324 /// equivalentAddressValues - Test if A and B will obviously have the same
1325 /// value. This includes recognizing that %t0 and %t1 will have the same
1326 /// value in code like this:
1327 /// %t0 = getelementptr \@a, 0, 3
1328 /// store i32 0, i32* %t0
1329 /// %t1 = getelementptr \@a, 0, 3
1330 /// %t2 = load i32* %t1
1332 static bool equivalentAddressValues(Value *A, Value *B) {
1333 // Test if the values are trivially equivalent.
1334 if (A == B) return true;
1336 // Test if the values come form identical arithmetic instructions.
1337 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
1338 // its only used to compare two uses within the same basic block, which
1339 // means that they'll always either have the same value or one of them
1340 // will have an undefined value.
1341 if (isa<BinaryOperator>(A) ||
1344 isa<GetElementPtrInst>(A))
1345 if (Instruction *BI = dyn_cast<Instruction>(B))
1346 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
1349 // Otherwise they may not be equivalent.
1353 /// Converts store (bitcast (load (bitcast (select ...)))) to
1354 /// store (load (select ...)), where select is minmax:
1355 /// select ((cmp load V1, load V2), V1, V2).
1356 static bool removeBitcastsFromLoadStoreOnMinMax(InstCombiner &IC,
1359 if (!match(SI.getPointerOperand(), m_BitCast(m_Value())))
1363 if (!match(SI.getValueOperand(), m_Load(m_BitCast(m_Value(LoadAddr)))))
1365 auto *LI = cast<LoadInst>(SI.getValueOperand());
1366 if (!LI->getType()->isIntegerTy())
1368 if (!isMinMaxWithLoads(LoadAddr))
1371 if (!all_of(LI->users(), [LI, LoadAddr](User *U) {
1372 auto *SI = dyn_cast<StoreInst>(U);
1373 return SI && SI->getPointerOperand() != LI &&
1374 peekThroughBitcast(SI->getPointerOperand()) != LoadAddr &&
1375 !SI->getPointerOperand()->isSwiftError();
1379 IC.Builder.SetInsertPoint(LI);
1380 LoadInst *NewLI = combineLoadToNewType(
1381 IC, *LI, LoadAddr->getType()->getPointerElementType());
1382 // Replace all the stores with stores of the newly loaded value.
1383 for (auto *UI : LI->users()) {
1384 auto *USI = cast<StoreInst>(UI);
1385 IC.Builder.SetInsertPoint(USI);
1386 combineStoreToNewValue(IC, *USI, NewLI);
1388 IC.replaceInstUsesWith(*LI, UndefValue::get(LI->getType()));
1389 IC.eraseInstFromFunction(*LI);
1393 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
1394 Value *Val = SI.getOperand(0);
1395 Value *Ptr = SI.getOperand(1);
1397 // Try to canonicalize the stored type.
1398 if (combineStoreToValueType(*this, SI))
1399 return eraseInstFromFunction(SI);
1401 // Attempt to improve the alignment.
1402 unsigned KnownAlign = getOrEnforceKnownAlignment(
1403 Ptr, DL.getPrefTypeAlignment(Val->getType()), DL, &SI, &AC, &DT);
1404 unsigned StoreAlign = SI.getAlignment();
1405 unsigned EffectiveStoreAlign =
1406 StoreAlign != 0 ? StoreAlign : DL.getABITypeAlignment(Val->getType());
1408 if (KnownAlign > EffectiveStoreAlign)
1409 SI.setAlignment(KnownAlign);
1410 else if (StoreAlign == 0)
1411 SI.setAlignment(EffectiveStoreAlign);
1413 // Try to canonicalize the stored type.
1414 if (unpackStoreToAggregate(*this, SI))
1415 return eraseInstFromFunction(SI);
1417 if (removeBitcastsFromLoadStoreOnMinMax(*this, SI))
1418 return eraseInstFromFunction(SI);
1420 // Replace GEP indices if possible.
1421 if (Instruction *NewGEPI = replaceGEPIdxWithZero(*this, Ptr, SI)) {
1422 Worklist.Add(NewGEPI);
1426 // Don't hack volatile/ordered stores.
1427 // FIXME: Some bits are legal for ordered atomic stores; needs refactoring.
1428 if (!SI.isUnordered()) return nullptr;
1430 // If the RHS is an alloca with a single use, zapify the store, making the
1432 if (Ptr->hasOneUse()) {
1433 if (isa<AllocaInst>(Ptr))
1434 return eraseInstFromFunction(SI);
1435 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
1436 if (isa<AllocaInst>(GEP->getOperand(0))) {
1437 if (GEP->getOperand(0)->hasOneUse())
1438 return eraseInstFromFunction(SI);
1443 // If we have a store to a location which is known constant, we can conclude
1444 // that the store must be storing the constant value (else the memory
1445 // wouldn't be constant), and this must be a noop.
1446 if (AA->pointsToConstantMemory(Ptr))
1447 return eraseInstFromFunction(SI);
1449 // Do really simple DSE, to catch cases where there are several consecutive
1450 // stores to the same location, separated by a few arithmetic operations. This
1451 // situation often occurs with bitfield accesses.
1452 BasicBlock::iterator BBI(SI);
1453 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
1456 // Don't count debug info directives, lest they affect codegen,
1457 // and we skip pointer-to-pointer bitcasts, which are NOPs.
1458 if (isa<DbgInfoIntrinsic>(BBI) ||
1459 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
1464 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
1465 // Prev store isn't volatile, and stores to the same location?
1466 if (PrevSI->isUnordered() && equivalentAddressValues(PrevSI->getOperand(1),
1467 SI.getOperand(1))) {
1470 eraseInstFromFunction(*PrevSI);
1476 // If this is a load, we have to stop. However, if the loaded value is from
1477 // the pointer we're loading and is producing the pointer we're storing,
1478 // then *this* store is dead (X = load P; store X -> P).
1479 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
1480 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr)) {
1481 assert(SI.isUnordered() && "can't eliminate ordering operation");
1482 return eraseInstFromFunction(SI);
1485 // Otherwise, this is a load from some other location. Stores before it
1490 // Don't skip over loads, throws or things that can modify memory.
1491 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory() || BBI->mayThrow())
1495 // store X, null -> turns into 'unreachable' in SimplifyCFG
1496 // store X, GEP(null, Y) -> turns into 'unreachable' in SimplifyCFG
1497 if (canSimplifyNullStoreOrGEP(SI)) {
1498 if (!isa<UndefValue>(Val)) {
1499 SI.setOperand(0, UndefValue::get(Val->getType()));
1500 if (Instruction *U = dyn_cast<Instruction>(Val))
1501 Worklist.Add(U); // Dropped a use.
1503 return nullptr; // Do not modify these!
1506 // store undef, Ptr -> noop
1507 if (isa<UndefValue>(Val))
1508 return eraseInstFromFunction(SI);
1510 // If this store is the second-to-last instruction in the basic block
1511 // (excluding debug info and bitcasts of pointers) and if the block ends with
1512 // an unconditional branch, try to move the store to the successor block.
1513 BBI = SI.getIterator();
1516 } while (isa<DbgInfoIntrinsic>(BBI) ||
1517 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy()));
1519 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
1520 if (BI->isUnconditional())
1521 mergeStoreIntoSuccessor(SI);
1526 /// Try to transform:
1527 /// if () { *P = v1; } else { *P = v2 }
1529 /// *P = v1; if () { *P = v2; }
1530 /// into a phi node with a store in the successor.
1531 bool InstCombiner::mergeStoreIntoSuccessor(StoreInst &SI) {
1532 assert(SI.isUnordered() &&
1533 "This code has not been audited for volatile or ordered store case.");
1535 // Check if the successor block has exactly 2 incoming edges.
1536 BasicBlock *StoreBB = SI.getParent();
1537 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
1538 if (!DestBB->hasNPredecessors(2))
1541 // Capture the other block (the block that doesn't contain our store).
1542 pred_iterator PredIter = pred_begin(DestBB);
1543 if (*PredIter == StoreBB)
1545 BasicBlock *OtherBB = *PredIter;
1547 // Bail out if all of the relevant blocks aren't distinct. This can happen,
1548 // for example, if SI is in an infinite loop.
1549 if (StoreBB == DestBB || OtherBB == DestBB)
1552 // Verify that the other block ends in a branch and is not otherwise empty.
1553 BasicBlock::iterator BBI(OtherBB->getTerminator());
1554 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
1555 if (!OtherBr || BBI == OtherBB->begin())
1558 // If the other block ends in an unconditional branch, check for the 'if then
1559 // else' case. There is an instruction before the branch.
1560 StoreInst *OtherStore = nullptr;
1561 if (OtherBr->isUnconditional()) {
1563 // Skip over debugging info.
1564 while (isa<DbgInfoIntrinsic>(BBI) ||
1565 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
1566 if (BBI==OtherBB->begin())
1570 // If this isn't a store, isn't a store to the same location, or is not the
1571 // right kind of store, bail out.
1572 OtherStore = dyn_cast<StoreInst>(BBI);
1573 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
1574 !SI.isSameOperationAs(OtherStore))
1577 // Otherwise, the other block ended with a conditional branch. If one of the
1578 // destinations is StoreBB, then we have the if/then case.
1579 if (OtherBr->getSuccessor(0) != StoreBB &&
1580 OtherBr->getSuccessor(1) != StoreBB)
1583 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
1584 // if/then triangle. See if there is a store to the same ptr as SI that
1585 // lives in OtherBB.
1587 // Check to see if we find the matching store.
1588 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
1589 if (OtherStore->getOperand(1) != SI.getOperand(1) ||
1590 !SI.isSameOperationAs(OtherStore))
1594 // If we find something that may be using or overwriting the stored
1595 // value, or if we run out of instructions, we can't do the transform.
1596 if (BBI->mayReadFromMemory() || BBI->mayThrow() ||
1597 BBI->mayWriteToMemory() || BBI == OtherBB->begin())
1601 // In order to eliminate the store in OtherBr, we have to make sure nothing
1602 // reads or overwrites the stored value in StoreBB.
1603 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
1604 // FIXME: This should really be AA driven.
1605 if (I->mayReadFromMemory() || I->mayThrow() || I->mayWriteToMemory())
1610 // Insert a PHI node now if we need it.
1611 Value *MergedVal = OtherStore->getOperand(0);
1612 // The debug locations of the original instructions might differ. Merge them.
1613 DebugLoc MergedLoc = DILocation::getMergedLocation(SI.getDebugLoc(),
1614 OtherStore->getDebugLoc());
1615 if (MergedVal != SI.getOperand(0)) {
1616 PHINode *PN = PHINode::Create(MergedVal->getType(), 2, "storemerge");
1617 PN->addIncoming(SI.getOperand(0), SI.getParent());
1618 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
1619 MergedVal = InsertNewInstBefore(PN, DestBB->front());
1620 PN->setDebugLoc(MergedLoc);
1623 // Advance to a place where it is safe to insert the new store and insert it.
1624 BBI = DestBB->getFirstInsertionPt();
1625 StoreInst *NewSI = new StoreInst(MergedVal, SI.getOperand(1),
1626 SI.isVolatile(), SI.getAlignment(),
1627 SI.getOrdering(), SI.getSyncScopeID());
1628 InsertNewInstBefore(NewSI, *BBI);
1629 NewSI->setDebugLoc(MergedLoc);
1631 // If the two stores had AA tags, merge them.
1633 SI.getAAMetadata(AATags);
1635 OtherStore->getAAMetadata(AATags, /* Merge = */ true);
1636 NewSI->setAAMetadata(AATags);
1639 // Nuke the old stores.
1640 eraseInstFromFunction(SI);
1641 eraseInstFromFunction(*OtherStore);