1 //===- InstCombineLoadStoreAlloca.cpp -------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visit functions for load, store and alloca.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombineInternal.h"
15 #include "llvm/ADT/Statistic.h"
16 #include "llvm/Analysis/Loads.h"
17 #include "llvm/IR/DataLayout.h"
18 #include "llvm/IR/LLVMContext.h"
19 #include "llvm/IR/IntrinsicInst.h"
20 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
21 #include "llvm/Transforms/Utils/Local.h"
24 #define DEBUG_TYPE "instcombine"
26 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
27 STATISTIC(NumGlobalCopies, "Number of allocas copied from constant global");
29 /// pointsToConstantGlobal - Return true if V (possibly indirectly) points to
30 /// some part of a constant global variable. This intentionally only accepts
31 /// constant expressions because we can't rewrite arbitrary instructions.
32 static bool pointsToConstantGlobal(Value *V) {
33 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
34 return GV->isConstant();
36 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
37 if (CE->getOpcode() == Instruction::BitCast ||
38 CE->getOpcode() == Instruction::AddrSpaceCast ||
39 CE->getOpcode() == Instruction::GetElementPtr)
40 return pointsToConstantGlobal(CE->getOperand(0));
45 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
46 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
47 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
48 /// track of whether it moves the pointer (with IsOffset) but otherwise traverse
49 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
50 /// the alloca, and if the source pointer is a pointer to a constant global, we
51 /// can optimize this.
53 isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
54 SmallVectorImpl<Instruction *> &ToDelete) {
55 // We track lifetime intrinsics as we encounter them. If we decide to go
56 // ahead and replace the value with the global, this lets the caller quickly
57 // eliminate the markers.
59 SmallVector<std::pair<Value *, bool>, 35> ValuesToInspect;
60 ValuesToInspect.push_back(std::make_pair(V, false));
61 while (!ValuesToInspect.empty()) {
62 auto ValuePair = ValuesToInspect.pop_back_val();
63 const bool IsOffset = ValuePair.second;
64 for (auto &U : ValuePair.first->uses()) {
65 Instruction *I = cast<Instruction>(U.getUser());
67 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
68 // Ignore non-volatile loads, they are always ok.
69 if (!LI->isSimple()) return false;
73 if (isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I)) {
74 // If uses of the bitcast are ok, we are ok.
75 ValuesToInspect.push_back(std::make_pair(I, IsOffset));
78 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
79 // If the GEP has all zero indices, it doesn't offset the pointer. If it
81 ValuesToInspect.push_back(
82 std::make_pair(I, IsOffset || !GEP->hasAllZeroIndices()));
86 if (CallSite CS = I) {
87 // If this is the function being called then we treat it like a load and
92 // Inalloca arguments are clobbered by the call.
93 unsigned ArgNo = CS.getArgumentNo(&U);
94 if (CS.isInAllocaArgument(ArgNo))
97 // If this is a readonly/readnone call site, then we know it is just a
98 // load (but one that potentially returns the value itself), so we can
99 // ignore it if we know that the value isn't captured.
100 if (CS.onlyReadsMemory() &&
101 (CS.getInstruction()->use_empty() || CS.doesNotCapture(ArgNo)))
104 // If this is being passed as a byval argument, the caller is making a
105 // copy, so it is only a read of the alloca.
106 if (CS.isByValArgument(ArgNo))
110 // Lifetime intrinsics can be handled by the caller.
111 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
112 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
113 II->getIntrinsicID() == Intrinsic::lifetime_end) {
114 assert(II->use_empty() && "Lifetime markers have no result to use!");
115 ToDelete.push_back(II);
120 // If this is isn't our memcpy/memmove, reject it as something we can't
122 MemTransferInst *MI = dyn_cast<MemTransferInst>(I);
126 // If the transfer is using the alloca as a source of the transfer, then
127 // ignore it since it is a load (unless the transfer is volatile).
128 if (U.getOperandNo() == 1) {
129 if (MI->isVolatile()) return false;
133 // If we already have seen a copy, reject the second one.
134 if (TheCopy) return false;
136 // If the pointer has been offset from the start of the alloca, we can't
137 // safely handle this.
138 if (IsOffset) return false;
140 // If the memintrinsic isn't using the alloca as the dest, reject it.
141 if (U.getOperandNo() != 0) return false;
143 // If the source of the memcpy/move is not a constant global, reject it.
144 if (!pointsToConstantGlobal(MI->getSource()))
147 // Otherwise, the transform is safe. Remember the copy instruction.
154 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
155 /// modified by a copy from a constant global. If we can prove this, we can
156 /// replace any uses of the alloca with uses of the global directly.
157 static MemTransferInst *
158 isOnlyCopiedFromConstantGlobal(AllocaInst *AI,
159 SmallVectorImpl<Instruction *> &ToDelete) {
160 MemTransferInst *TheCopy = nullptr;
161 if (isOnlyCopiedFromConstantGlobal(AI, TheCopy, ToDelete))
166 Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
167 // Ensure that the alloca array size argument has type intptr_t, so that
168 // any casting is exposed early.
170 Type *IntPtrTy = DL->getIntPtrType(AI.getType());
171 if (AI.getArraySize()->getType() != IntPtrTy) {
172 Value *V = Builder->CreateIntCast(AI.getArraySize(),
179 // Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1
180 if (AI.isArrayAllocation()) { // Check C != 1
181 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
183 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
184 AllocaInst *New = Builder->CreateAlloca(NewTy, nullptr, AI.getName());
185 New->setAlignment(AI.getAlignment());
187 // Scan to the end of the allocation instructions, to skip over a block of
188 // allocas if possible...also skip interleaved debug info
190 BasicBlock::iterator It = New;
191 while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
193 // Now that I is pointing to the first non-allocation-inst in the block,
194 // insert our getelementptr instruction...
197 ? DL->getIntPtrType(AI.getType())
198 : Type::getInt64Ty(AI.getContext());
199 Value *NullIdx = Constant::getNullValue(IdxTy);
200 Value *Idx[2] = { NullIdx, NullIdx };
202 GetElementPtrInst::CreateInBounds(New, Idx, New->getName() + ".sub");
203 InsertNewInstBefore(GEP, *It);
205 // Now make everything use the getelementptr instead of the original
207 return ReplaceInstUsesWith(AI, GEP);
208 } else if (isa<UndefValue>(AI.getArraySize())) {
209 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
213 if (DL && AI.getAllocatedType()->isSized()) {
214 // If the alignment is 0 (unspecified), assign it the preferred alignment.
215 if (AI.getAlignment() == 0)
216 AI.setAlignment(DL->getPrefTypeAlignment(AI.getAllocatedType()));
218 // Move all alloca's of zero byte objects to the entry block and merge them
219 // together. Note that we only do this for alloca's, because malloc should
220 // allocate and return a unique pointer, even for a zero byte allocation.
221 if (DL->getTypeAllocSize(AI.getAllocatedType()) == 0) {
222 // For a zero sized alloca there is no point in doing an array allocation.
223 // This is helpful if the array size is a complicated expression not used
225 if (AI.isArrayAllocation()) {
226 AI.setOperand(0, ConstantInt::get(AI.getArraySize()->getType(), 1));
230 // Get the first instruction in the entry block.
231 BasicBlock &EntryBlock = AI.getParent()->getParent()->getEntryBlock();
232 Instruction *FirstInst = EntryBlock.getFirstNonPHIOrDbg();
233 if (FirstInst != &AI) {
234 // If the entry block doesn't start with a zero-size alloca then move
235 // this one to the start of the entry block. There is no problem with
236 // dominance as the array size was forced to a constant earlier already.
237 AllocaInst *EntryAI = dyn_cast<AllocaInst>(FirstInst);
238 if (!EntryAI || !EntryAI->getAllocatedType()->isSized() ||
239 DL->getTypeAllocSize(EntryAI->getAllocatedType()) != 0) {
240 AI.moveBefore(FirstInst);
244 // If the alignment of the entry block alloca is 0 (unspecified),
245 // assign it the preferred alignment.
246 if (EntryAI->getAlignment() == 0)
247 EntryAI->setAlignment(
248 DL->getPrefTypeAlignment(EntryAI->getAllocatedType()));
249 // Replace this zero-sized alloca with the one at the start of the entry
250 // block after ensuring that the address will be aligned enough for both
252 unsigned MaxAlign = std::max(EntryAI->getAlignment(),
254 EntryAI->setAlignment(MaxAlign);
255 if (AI.getType() != EntryAI->getType())
256 return new BitCastInst(EntryAI, AI.getType());
257 return ReplaceInstUsesWith(AI, EntryAI);
262 if (AI.getAlignment()) {
263 // Check to see if this allocation is only modified by a memcpy/memmove from
264 // a constant global whose alignment is equal to or exceeds that of the
265 // allocation. If this is the case, we can change all users to use
266 // the constant global instead. This is commonly produced by the CFE by
267 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
268 // is only subsequently read.
269 SmallVector<Instruction *, 4> ToDelete;
270 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(&AI, ToDelete)) {
271 unsigned SourceAlign = getOrEnforceKnownAlignment(
272 Copy->getSource(), AI.getAlignment(), DL, AC, &AI, DT);
273 if (AI.getAlignment() <= SourceAlign) {
274 DEBUG(dbgs() << "Found alloca equal to global: " << AI << '\n');
275 DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
276 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
277 EraseInstFromFunction(*ToDelete[i]);
278 Constant *TheSrc = cast<Constant>(Copy->getSource());
280 = ConstantExpr::getPointerBitCastOrAddrSpaceCast(TheSrc, AI.getType());
281 Instruction *NewI = ReplaceInstUsesWith(AI, Cast);
282 EraseInstFromFunction(*Copy);
289 // At last, use the generic allocation site handler to aggressively remove
291 return visitAllocSite(AI);
294 /// \brief Helper to combine a load to a new type.
296 /// This just does the work of combining a load to a new type. It handles
297 /// metadata, etc., and returns the new instruction. The \c NewTy should be the
298 /// loaded *value* type. This will convert it to a pointer, cast the operand to
299 /// that pointer type, load it, etc.
301 /// Note that this will create all of the instructions with whatever insert
302 /// point the \c InstCombiner currently is using.
303 static LoadInst *combineLoadToNewType(InstCombiner &IC, LoadInst &LI, Type *NewTy) {
304 Value *Ptr = LI.getPointerOperand();
305 unsigned AS = LI.getPointerAddressSpace();
306 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
307 LI.getAllMetadata(MD);
309 LoadInst *NewLoad = IC.Builder->CreateAlignedLoad(
310 IC.Builder->CreateBitCast(Ptr, NewTy->getPointerTo(AS)),
311 LI.getAlignment(), LI.getName());
312 for (const auto &MDPair : MD) {
313 unsigned ID = MDPair.first;
314 MDNode *N = MDPair.second;
315 // Note, essentially every kind of metadata should be preserved here! This
316 // routine is supposed to clone a load instruction changing *only its type*.
317 // The only metadata it makes sense to drop is metadata which is invalidated
318 // when the pointer type changes. This should essentially never be the case
319 // in LLVM, but we explicitly switch over only known metadata to be
320 // conservatively correct. If you are adding metadata to LLVM which pertains
321 // to loads, you almost certainly want to add it here.
323 case LLVMContext::MD_dbg:
324 case LLVMContext::MD_tbaa:
325 case LLVMContext::MD_prof:
326 case LLVMContext::MD_fpmath:
327 case LLVMContext::MD_tbaa_struct:
328 case LLVMContext::MD_invariant_load:
329 case LLVMContext::MD_alias_scope:
330 case LLVMContext::MD_noalias:
331 case LLVMContext::MD_nontemporal:
332 case LLVMContext::MD_mem_parallel_loop_access:
333 // All of these directly apply.
334 NewLoad->setMetadata(ID, N);
337 case LLVMContext::MD_nonnull:
338 // FIXME: We should translate this into range metadata for integer types
340 if (NewTy->isPointerTy())
341 NewLoad->setMetadata(ID, N);
344 case LLVMContext::MD_range:
345 // FIXME: It would be nice to propagate this in some way, but the type
346 // conversions make it hard.
353 /// \brief Combine a store to a new type.
355 /// Returns the newly created store instruction.
356 static StoreInst *combineStoreToNewValue(InstCombiner &IC, StoreInst &SI, Value *V) {
357 Value *Ptr = SI.getPointerOperand();
358 unsigned AS = SI.getPointerAddressSpace();
359 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
360 SI.getAllMetadata(MD);
362 StoreInst *NewStore = IC.Builder->CreateAlignedStore(
363 V, IC.Builder->CreateBitCast(Ptr, V->getType()->getPointerTo(AS)),
365 for (const auto &MDPair : MD) {
366 unsigned ID = MDPair.first;
367 MDNode *N = MDPair.second;
368 // Note, essentially every kind of metadata should be preserved here! This
369 // routine is supposed to clone a store instruction changing *only its
370 // type*. The only metadata it makes sense to drop is metadata which is
371 // invalidated when the pointer type changes. This should essentially
372 // never be the case in LLVM, but we explicitly switch over only known
373 // metadata to be conservatively correct. If you are adding metadata to
374 // LLVM which pertains to stores, you almost certainly want to add it
377 case LLVMContext::MD_dbg:
378 case LLVMContext::MD_tbaa:
379 case LLVMContext::MD_prof:
380 case LLVMContext::MD_fpmath:
381 case LLVMContext::MD_tbaa_struct:
382 case LLVMContext::MD_alias_scope:
383 case LLVMContext::MD_noalias:
384 case LLVMContext::MD_nontemporal:
385 case LLVMContext::MD_mem_parallel_loop_access:
386 // All of these directly apply.
387 NewStore->setMetadata(ID, N);
390 case LLVMContext::MD_invariant_load:
391 case LLVMContext::MD_nonnull:
392 case LLVMContext::MD_range:
393 // These don't apply for stores.
401 /// \brief Combine loads to match the type of value their uses after looking
402 /// through intervening bitcasts.
404 /// The core idea here is that if the result of a load is used in an operation,
405 /// we should load the type most conducive to that operation. For example, when
406 /// loading an integer and converting that immediately to a pointer, we should
407 /// instead directly load a pointer.
409 /// However, this routine must never change the width of a load or the number of
410 /// loads as that would introduce a semantic change. This combine is expected to
411 /// be a semantic no-op which just allows loads to more closely model the types
412 /// of their consuming operations.
414 /// Currently, we also refuse to change the precise type used for an atomic load
415 /// or a volatile load. This is debatable, and might be reasonable to change
416 /// later. However, it is risky in case some backend or other part of LLVM is
417 /// relying on the exact type loaded to select appropriate atomic operations.
418 static Instruction *combineLoadToOperationType(InstCombiner &IC, LoadInst &LI) {
419 // FIXME: We could probably with some care handle both volatile and atomic
420 // loads here but it isn't clear that this is important.
427 Type *Ty = LI.getType();
429 // Try to canonicalize loads which are only ever stored to operate over
430 // integers instead of any other type. We only do this when the loaded type
431 // is sized and has a size exactly the same as its store size and the store
432 // size is a legal integer type.
433 const DataLayout *DL = IC.getDataLayout();
434 if (!Ty->isIntegerTy() && Ty->isSized() && DL &&
435 DL->isLegalInteger(DL->getTypeStoreSizeInBits(Ty)) &&
436 DL->getTypeStoreSizeInBits(Ty) == DL->getTypeSizeInBits(Ty)) {
437 if (std::all_of(LI.user_begin(), LI.user_end(), [&LI](User *U) {
438 auto *SI = dyn_cast<StoreInst>(U);
439 return SI && SI->getPointerOperand() != &LI;
441 LoadInst *NewLoad = combineLoadToNewType(
443 Type::getIntNTy(LI.getContext(), DL->getTypeStoreSizeInBits(Ty)));
444 // Replace all the stores with stores of the newly loaded value.
445 for (auto UI = LI.user_begin(), UE = LI.user_end(); UI != UE;) {
446 auto *SI = cast<StoreInst>(*UI++);
447 IC.Builder->SetInsertPoint(SI);
448 combineStoreToNewValue(IC, *SI, NewLoad);
449 IC.EraseInstFromFunction(*SI);
451 assert(LI.use_empty() && "Failed to remove all users of the load!");
452 // Return the old load so the combiner can delete it safely.
457 // Fold away bit casts of the loaded value by loading the desired type.
459 if (auto *BC = dyn_cast<BitCastInst>(LI.user_back())) {
460 LoadInst *NewLoad = combineLoadToNewType(IC, LI, BC->getDestTy());
461 BC->replaceAllUsesWith(NewLoad);
462 IC.EraseInstFromFunction(*BC);
466 // FIXME: We should also canonicalize loads of vectors when their elements are
467 // cast to other types.
471 // If we can determine that all possible objects pointed to by the provided
472 // pointer value are, not only dereferenceable, but also definitively less than
473 // or equal to the provided maximum size, then return true. Otherwise, return
474 // false (constant global values and allocas fall into this category).
476 // FIXME: This should probably live in ValueTracking (or similar).
477 static bool isObjectSizeLessThanOrEq(Value *V, uint64_t MaxSize,
478 const DataLayout *DL) {
479 SmallPtrSet<Value *, 4> Visited;
480 SmallVector<Value *, 4> Worklist(1, V);
483 Value *P = Worklist.pop_back_val();
484 P = P->stripPointerCasts();
486 if (!Visited.insert(P).second)
489 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
490 Worklist.push_back(SI->getTrueValue());
491 Worklist.push_back(SI->getFalseValue());
495 if (PHINode *PN = dyn_cast<PHINode>(P)) {
496 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
497 Worklist.push_back(PN->getIncomingValue(i));
501 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(P)) {
502 if (GA->mayBeOverridden())
504 Worklist.push_back(GA->getAliasee());
508 // If we know how big this object is, and it is less than MaxSize, continue
509 // searching. Otherwise, return false.
510 if (AllocaInst *AI = dyn_cast<AllocaInst>(P)) {
511 if (!AI->getAllocatedType()->isSized())
514 ConstantInt *CS = dyn_cast<ConstantInt>(AI->getArraySize());
518 uint64_t TypeSize = DL->getTypeAllocSize(AI->getAllocatedType());
519 // Make sure that, even if the multiplication below would wrap as an
520 // uint64_t, we still do the right thing.
521 if ((CS->getValue().zextOrSelf(128)*APInt(128, TypeSize)).ugt(MaxSize))
526 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
527 if (!GV->hasDefinitiveInitializer() || !GV->isConstant())
530 uint64_t InitSize = DL->getTypeAllocSize(GV->getType()->getElementType());
531 if (InitSize > MaxSize)
537 } while (!Worklist.empty());
542 // If we're indexing into an object of a known size, and the outer index is
543 // not a constant, but having any value but zero would lead to undefined
544 // behavior, replace it with zero.
546 // For example, if we have:
547 // @f.a = private unnamed_addr constant [1 x i32] [i32 12], align 4
549 // %arrayidx = getelementptr inbounds [1 x i32]* @f.a, i64 0, i64 %x
550 // ... = load i32* %arrayidx, align 4
551 // Then we know that we can replace %x in the GEP with i64 0.
553 // FIXME: We could fold any GEP index to zero that would cause UB if it were
554 // not zero. Currently, we only handle the first such index. Also, we could
555 // also search through non-zero constant indices if we kept track of the
556 // offsets those indices implied.
557 static bool canReplaceGEPIdxWithZero(InstCombiner &IC, GetElementPtrInst *GEPI,
558 Instruction *MemI, unsigned &Idx) {
559 const DataLayout *DL = IC.getDataLayout();
560 if (GEPI->getNumOperands() < 2 || !DL)
563 // Find the first non-zero index of a GEP. If all indices are zero, return
564 // one past the last index.
565 auto FirstNZIdx = [](const GetElementPtrInst *GEPI) {
567 for (unsigned IE = GEPI->getNumOperands(); I != IE; ++I) {
568 Value *V = GEPI->getOperand(I);
569 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V))
579 // Skip through initial 'zero' indices, and find the corresponding pointer
580 // type. See if the next index is not a constant.
581 Idx = FirstNZIdx(GEPI);
582 if (Idx == GEPI->getNumOperands())
584 if (isa<Constant>(GEPI->getOperand(Idx)))
587 SmallVector<Value *, 4> Ops(GEPI->idx_begin(), GEPI->idx_begin() + Idx);
589 GetElementPtrInst::getIndexedType(GEPI->getOperand(0)->getType(), Ops);
590 if (!AllocTy || !AllocTy->isSized())
592 uint64_t TyAllocSize = DL->getTypeAllocSize(AllocTy);
594 // If there are more indices after the one we might replace with a zero, make
595 // sure they're all non-negative. If any of them are negative, the overall
596 // address being computed might be before the base address determined by the
597 // first non-zero index.
598 auto IsAllNonNegative = [&]() {
599 for (unsigned i = Idx+1, e = GEPI->getNumOperands(); i != e; ++i) {
600 bool KnownNonNegative, KnownNegative;
601 IC.ComputeSignBit(GEPI->getOperand(i), KnownNonNegative,
602 KnownNegative, 0, MemI);
603 if (KnownNonNegative)
611 // FIXME: If the GEP is not inbounds, and there are extra indices after the
612 // one we'll replace, those could cause the address computation to wrap
613 // (rendering the IsAllNonNegative() check below insufficient). We can do
614 // better, ignoring zero indicies (and other indicies we can prove small
615 // enough not to wrap).
616 if (Idx+1 != GEPI->getNumOperands() && !GEPI->isInBounds())
619 // Note that isObjectSizeLessThanOrEq will return true only if the pointer is
620 // also known to be dereferenceable.
621 return isObjectSizeLessThanOrEq(GEPI->getOperand(0), TyAllocSize, DL) &&
625 // If we're indexing into an object with a variable index for the memory
626 // access, but the object has only one element, we can assume that the index
627 // will always be zero. If we replace the GEP, return it.
628 template <typename T>
629 static Instruction *replaceGEPIdxWithZero(InstCombiner &IC, Value *Ptr,
631 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr)) {
633 if (canReplaceGEPIdxWithZero(IC, GEPI, &MemI, Idx)) {
634 Instruction *NewGEPI = GEPI->clone();
635 NewGEPI->setOperand(Idx,
636 ConstantInt::get(GEPI->getOperand(Idx)->getType(), 0));
637 NewGEPI->insertBefore(GEPI);
638 MemI.setOperand(MemI.getPointerOperandIndex(), NewGEPI);
646 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
647 Value *Op = LI.getOperand(0);
649 // Try to canonicalize the loaded type.
650 if (Instruction *Res = combineLoadToOperationType(*this, LI))
653 // Attempt to improve the alignment.
655 unsigned KnownAlign = getOrEnforceKnownAlignment(
656 Op, DL->getPrefTypeAlignment(LI.getType()), DL, AC, &LI, DT);
657 unsigned LoadAlign = LI.getAlignment();
658 unsigned EffectiveLoadAlign = LoadAlign != 0 ? LoadAlign :
659 DL->getABITypeAlignment(LI.getType());
661 if (KnownAlign > EffectiveLoadAlign)
662 LI.setAlignment(KnownAlign);
663 else if (LoadAlign == 0)
664 LI.setAlignment(EffectiveLoadAlign);
667 // Replace GEP indices if possible.
668 if (Instruction *NewGEPI = replaceGEPIdxWithZero(*this, Op, LI)) {
669 Worklist.Add(NewGEPI);
673 // None of the following transforms are legal for volatile/atomic loads.
674 // FIXME: Some of it is okay for atomic loads; needs refactoring.
675 if (!LI.isSimple()) return nullptr;
677 // Do really simple store-to-load forwarding and load CSE, to catch cases
678 // where there are several consecutive memory accesses to the same location,
679 // separated by a few arithmetic operations.
680 BasicBlock::iterator BBI = &LI;
681 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
682 return ReplaceInstUsesWith(
683 LI, Builder->CreateBitOrPointerCast(AvailableVal, LI.getType(),
684 LI.getName() + ".cast"));
686 // load(gep null, ...) -> unreachable
687 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
688 const Value *GEPI0 = GEPI->getOperand(0);
689 // TODO: Consider a target hook for valid address spaces for this xform.
690 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
691 // Insert a new store to null instruction before the load to indicate
692 // that this code is not reachable. We do this instead of inserting
693 // an unreachable instruction directly because we cannot modify the
695 new StoreInst(UndefValue::get(LI.getType()),
696 Constant::getNullValue(Op->getType()), &LI);
697 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
701 // load null/undef -> unreachable
702 // TODO: Consider a target hook for valid address spaces for this xform.
703 if (isa<UndefValue>(Op) ||
704 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) {
705 // Insert a new store to null instruction before the load to indicate that
706 // this code is not reachable. We do this instead of inserting an
707 // unreachable instruction directly because we cannot modify the CFG.
708 new StoreInst(UndefValue::get(LI.getType()),
709 Constant::getNullValue(Op->getType()), &LI);
710 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
713 if (Op->hasOneUse()) {
714 // Change select and PHI nodes to select values instead of addresses: this
715 // helps alias analysis out a lot, allows many others simplifications, and
716 // exposes redundancy in the code.
718 // Note that we cannot do the transformation unless we know that the
719 // introduced loads cannot trap! Something like this is valid as long as
720 // the condition is always false: load (select bool %C, int* null, int* %G),
721 // but it would not be valid if we transformed it to load from null
724 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
725 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
726 unsigned Align = LI.getAlignment();
727 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI, Align, DL) &&
728 isSafeToLoadUnconditionally(SI->getOperand(2), SI, Align, DL)) {
729 LoadInst *V1 = Builder->CreateLoad(SI->getOperand(1),
730 SI->getOperand(1)->getName()+".val");
731 LoadInst *V2 = Builder->CreateLoad(SI->getOperand(2),
732 SI->getOperand(2)->getName()+".val");
733 V1->setAlignment(Align);
734 V2->setAlignment(Align);
735 return SelectInst::Create(SI->getCondition(), V1, V2);
738 // load (select (cond, null, P)) -> load P
739 if (isa<ConstantPointerNull>(SI->getOperand(1)) &&
740 LI.getPointerAddressSpace() == 0) {
741 LI.setOperand(0, SI->getOperand(2));
745 // load (select (cond, P, null)) -> load P
746 if (isa<ConstantPointerNull>(SI->getOperand(2)) &&
747 LI.getPointerAddressSpace() == 0) {
748 LI.setOperand(0, SI->getOperand(1));
756 /// \brief Combine stores to match the type of value being stored.
758 /// The core idea here is that the memory does not have any intrinsic type and
759 /// where we can we should match the type of a store to the type of value being
762 /// However, this routine must never change the width of a store or the number of
763 /// stores as that would introduce a semantic change. This combine is expected to
764 /// be a semantic no-op which just allows stores to more closely model the types
765 /// of their incoming values.
767 /// Currently, we also refuse to change the precise type used for an atomic or
768 /// volatile store. This is debatable, and might be reasonable to change later.
769 /// However, it is risky in case some backend or other part of LLVM is relying
770 /// on the exact type stored to select appropriate atomic operations.
772 /// \returns true if the store was successfully combined away. This indicates
773 /// the caller must erase the store instruction. We have to let the caller erase
774 /// the store instruction sas otherwise there is no way to signal whether it was
775 /// combined or not: IC.EraseInstFromFunction returns a null pointer.
776 static bool combineStoreToValueType(InstCombiner &IC, StoreInst &SI) {
777 // FIXME: We could probably with some care handle both volatile and atomic
778 // stores here but it isn't clear that this is important.
782 Value *V = SI.getValueOperand();
784 // Fold away bit casts of the stored value by storing the original type.
785 if (auto *BC = dyn_cast<BitCastInst>(V)) {
786 V = BC->getOperand(0);
787 combineStoreToNewValue(IC, SI, V);
791 // FIXME: We should also canonicalize loads of vectors when their elements are
792 // cast to other types.
796 /// equivalentAddressValues - Test if A and B will obviously have the same
797 /// value. This includes recognizing that %t0 and %t1 will have the same
798 /// value in code like this:
799 /// %t0 = getelementptr \@a, 0, 3
800 /// store i32 0, i32* %t0
801 /// %t1 = getelementptr \@a, 0, 3
802 /// %t2 = load i32* %t1
804 static bool equivalentAddressValues(Value *A, Value *B) {
805 // Test if the values are trivially equivalent.
806 if (A == B) return true;
808 // Test if the values come form identical arithmetic instructions.
809 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
810 // its only used to compare two uses within the same basic block, which
811 // means that they'll always either have the same value or one of them
812 // will have an undefined value.
813 if (isa<BinaryOperator>(A) ||
816 isa<GetElementPtrInst>(A))
817 if (Instruction *BI = dyn_cast<Instruction>(B))
818 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
821 // Otherwise they may not be equivalent.
825 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
826 Value *Val = SI.getOperand(0);
827 Value *Ptr = SI.getOperand(1);
829 // Try to canonicalize the stored type.
830 if (combineStoreToValueType(*this, SI))
831 return EraseInstFromFunction(SI);
833 // Attempt to improve the alignment.
835 unsigned KnownAlign = getOrEnforceKnownAlignment(
836 Ptr, DL->getPrefTypeAlignment(Val->getType()), DL, AC, &SI, DT);
837 unsigned StoreAlign = SI.getAlignment();
838 unsigned EffectiveStoreAlign = StoreAlign != 0 ? StoreAlign :
839 DL->getABITypeAlignment(Val->getType());
841 if (KnownAlign > EffectiveStoreAlign)
842 SI.setAlignment(KnownAlign);
843 else if (StoreAlign == 0)
844 SI.setAlignment(EffectiveStoreAlign);
847 // Replace GEP indices if possible.
848 if (Instruction *NewGEPI = replaceGEPIdxWithZero(*this, Ptr, SI)) {
849 Worklist.Add(NewGEPI);
853 // Don't hack volatile/atomic stores.
854 // FIXME: Some bits are legal for atomic stores; needs refactoring.
855 if (!SI.isSimple()) return nullptr;
857 // If the RHS is an alloca with a single use, zapify the store, making the
859 if (Ptr->hasOneUse()) {
860 if (isa<AllocaInst>(Ptr))
861 return EraseInstFromFunction(SI);
862 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
863 if (isa<AllocaInst>(GEP->getOperand(0))) {
864 if (GEP->getOperand(0)->hasOneUse())
865 return EraseInstFromFunction(SI);
870 // Do really simple DSE, to catch cases where there are several consecutive
871 // stores to the same location, separated by a few arithmetic operations. This
872 // situation often occurs with bitfield accesses.
873 BasicBlock::iterator BBI = &SI;
874 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
877 // Don't count debug info directives, lest they affect codegen,
878 // and we skip pointer-to-pointer bitcasts, which are NOPs.
879 if (isa<DbgInfoIntrinsic>(BBI) ||
880 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
885 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
886 // Prev store isn't volatile, and stores to the same location?
887 if (PrevSI->isSimple() && equivalentAddressValues(PrevSI->getOperand(1),
891 EraseInstFromFunction(*PrevSI);
897 // If this is a load, we have to stop. However, if the loaded value is from
898 // the pointer we're loading and is producing the pointer we're storing,
899 // then *this* store is dead (X = load P; store X -> P).
900 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
901 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
903 return EraseInstFromFunction(SI);
905 // Otherwise, this is a load from some other location. Stores before it
910 // Don't skip over loads or things that can modify memory.
911 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
915 // store X, null -> turns into 'unreachable' in SimplifyCFG
916 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
917 if (!isa<UndefValue>(Val)) {
918 SI.setOperand(0, UndefValue::get(Val->getType()));
919 if (Instruction *U = dyn_cast<Instruction>(Val))
920 Worklist.Add(U); // Dropped a use.
922 return nullptr; // Do not modify these!
925 // store undef, Ptr -> noop
926 if (isa<UndefValue>(Val))
927 return EraseInstFromFunction(SI);
929 // If this store is the last instruction in the basic block (possibly
930 // excepting debug info instructions), and if the block ends with an
931 // unconditional branch, try to move it to the successor block.
935 } while (isa<DbgInfoIntrinsic>(BBI) ||
936 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy()));
937 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
938 if (BI->isUnconditional())
939 if (SimplifyStoreAtEndOfBlock(SI))
940 return nullptr; // xform done!
945 /// SimplifyStoreAtEndOfBlock - Turn things like:
946 /// if () { *P = v1; } else { *P = v2 }
947 /// into a phi node with a store in the successor.
949 /// Simplify things like:
950 /// *P = v1; if () { *P = v2; }
951 /// into a phi node with a store in the successor.
953 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
954 BasicBlock *StoreBB = SI.getParent();
956 // Check to see if the successor block has exactly two incoming edges. If
957 // so, see if the other predecessor contains a store to the same location.
958 // if so, insert a PHI node (if needed) and move the stores down.
959 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
961 // Determine whether Dest has exactly two predecessors and, if so, compute
962 // the other predecessor.
963 pred_iterator PI = pred_begin(DestBB);
965 BasicBlock *OtherBB = nullptr;
970 if (++PI == pred_end(DestBB))
979 if (++PI != pred_end(DestBB))
982 // Bail out if all the relevant blocks aren't distinct (this can happen,
983 // for example, if SI is in an infinite loop)
984 if (StoreBB == DestBB || OtherBB == DestBB)
987 // Verify that the other block ends in a branch and is not otherwise empty.
988 BasicBlock::iterator BBI = OtherBB->getTerminator();
989 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
990 if (!OtherBr || BBI == OtherBB->begin())
993 // If the other block ends in an unconditional branch, check for the 'if then
994 // else' case. there is an instruction before the branch.
995 StoreInst *OtherStore = nullptr;
996 if (OtherBr->isUnconditional()) {
998 // Skip over debugging info.
999 while (isa<DbgInfoIntrinsic>(BBI) ||
1000 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
1001 if (BBI==OtherBB->begin())
1005 // If this isn't a store, isn't a store to the same location, or is not the
1006 // right kind of store, bail out.
1007 OtherStore = dyn_cast<StoreInst>(BBI);
1008 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
1009 !SI.isSameOperationAs(OtherStore))
1012 // Otherwise, the other block ended with a conditional branch. If one of the
1013 // destinations is StoreBB, then we have the if/then case.
1014 if (OtherBr->getSuccessor(0) != StoreBB &&
1015 OtherBr->getSuccessor(1) != StoreBB)
1018 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
1019 // if/then triangle. See if there is a store to the same ptr as SI that
1020 // lives in OtherBB.
1022 // Check to see if we find the matching store.
1023 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
1024 if (OtherStore->getOperand(1) != SI.getOperand(1) ||
1025 !SI.isSameOperationAs(OtherStore))
1029 // If we find something that may be using or overwriting the stored
1030 // value, or if we run out of instructions, we can't do the xform.
1031 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
1032 BBI == OtherBB->begin())
1036 // In order to eliminate the store in OtherBr, we have to
1037 // make sure nothing reads or overwrites the stored value in
1039 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
1040 // FIXME: This should really be AA driven.
1041 if (I->mayReadFromMemory() || I->mayWriteToMemory())
1046 // Insert a PHI node now if we need it.
1047 Value *MergedVal = OtherStore->getOperand(0);
1048 if (MergedVal != SI.getOperand(0)) {
1049 PHINode *PN = PHINode::Create(MergedVal->getType(), 2, "storemerge");
1050 PN->addIncoming(SI.getOperand(0), SI.getParent());
1051 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
1052 MergedVal = InsertNewInstBefore(PN, DestBB->front());
1055 // Advance to a place where it is safe to insert the new store and
1057 BBI = DestBB->getFirstInsertionPt();
1058 StoreInst *NewSI = new StoreInst(MergedVal, SI.getOperand(1),
1062 SI.getSynchScope());
1063 InsertNewInstBefore(NewSI, *BBI);
1064 NewSI->setDebugLoc(OtherStore->getDebugLoc());
1066 // If the two stores had AA tags, merge them.
1068 SI.getAAMetadata(AATags);
1070 OtherStore->getAAMetadata(AATags, /* Merge = */ true);
1071 NewSI->setAAMetadata(AATags);
1074 // Nuke the old stores.
1075 EraseInstFromFunction(SI);
1076 EraseInstFromFunction(*OtherStore);