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/SmallString.h"
16 #include "llvm/ADT/Statistic.h"
17 #include "llvm/Analysis/Loads.h"
18 #include "llvm/IR/DataLayout.h"
19 #include "llvm/IR/LLVMContext.h"
20 #include "llvm/IR/IntrinsicInst.h"
21 #include "llvm/IR/MDBuilder.h"
22 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
23 #include "llvm/Transforms/Utils/Local.h"
26 #define DEBUG_TYPE "instcombine"
28 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
29 STATISTIC(NumGlobalCopies, "Number of allocas copied from constant global");
31 /// pointsToConstantGlobal - Return true if V (possibly indirectly) points to
32 /// some part of a constant global variable. This intentionally only accepts
33 /// constant expressions because we can't rewrite arbitrary instructions.
34 static bool pointsToConstantGlobal(Value *V) {
35 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
36 return GV->isConstant();
38 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
39 if (CE->getOpcode() == Instruction::BitCast ||
40 CE->getOpcode() == Instruction::AddrSpaceCast ||
41 CE->getOpcode() == Instruction::GetElementPtr)
42 return pointsToConstantGlobal(CE->getOperand(0));
47 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
48 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
49 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
50 /// track of whether it moves the pointer (with IsOffset) but otherwise traverse
51 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
52 /// the alloca, and if the source pointer is a pointer to a constant global, we
53 /// can optimize this.
55 isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
56 SmallVectorImpl<Instruction *> &ToDelete) {
57 // We track lifetime intrinsics as we encounter them. If we decide to go
58 // ahead and replace the value with the global, this lets the caller quickly
59 // eliminate the markers.
61 SmallVector<std::pair<Value *, bool>, 35> ValuesToInspect;
62 ValuesToInspect.push_back(std::make_pair(V, false));
63 while (!ValuesToInspect.empty()) {
64 auto ValuePair = ValuesToInspect.pop_back_val();
65 const bool IsOffset = ValuePair.second;
66 for (auto &U : ValuePair.first->uses()) {
67 Instruction *I = cast<Instruction>(U.getUser());
69 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
70 // Ignore non-volatile loads, they are always ok.
71 if (!LI->isSimple()) return false;
75 if (isa<BitCastInst>(I) || isa<AddrSpaceCastInst>(I)) {
76 // If uses of the bitcast are ok, we are ok.
77 ValuesToInspect.push_back(std::make_pair(I, IsOffset));
80 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
81 // If the GEP has all zero indices, it doesn't offset the pointer. If it
83 ValuesToInspect.push_back(
84 std::make_pair(I, IsOffset || !GEP->hasAllZeroIndices()));
88 if (auto CS = CallSite(I)) {
89 // If this is the function being called then we treat it like a load and
94 // Inalloca arguments are clobbered by the call.
95 unsigned ArgNo = CS.getArgumentNo(&U);
96 if (CS.isInAllocaArgument(ArgNo))
99 // If this is a readonly/readnone call site, then we know it is just a
100 // load (but one that potentially returns the value itself), so we can
101 // ignore it if we know that the value isn't captured.
102 if (CS.onlyReadsMemory() &&
103 (CS.getInstruction()->use_empty() || CS.doesNotCapture(ArgNo)))
106 // If this is being passed as a byval argument, the caller is making a
107 // copy, so it is only a read of the alloca.
108 if (CS.isByValArgument(ArgNo))
112 // Lifetime intrinsics can be handled by the caller.
113 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
114 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
115 II->getIntrinsicID() == Intrinsic::lifetime_end) {
116 assert(II->use_empty() && "Lifetime markers have no result to use!");
117 ToDelete.push_back(II);
122 // If this is isn't our memcpy/memmove, reject it as something we can't
124 MemTransferInst *MI = dyn_cast<MemTransferInst>(I);
128 // If the transfer is using the alloca as a source of the transfer, then
129 // ignore it since it is a load (unless the transfer is volatile).
130 if (U.getOperandNo() == 1) {
131 if (MI->isVolatile()) return false;
135 // If we already have seen a copy, reject the second one.
136 if (TheCopy) return false;
138 // If the pointer has been offset from the start of the alloca, we can't
139 // safely handle this.
140 if (IsOffset) return false;
142 // If the memintrinsic isn't using the alloca as the dest, reject it.
143 if (U.getOperandNo() != 0) return false;
145 // If the source of the memcpy/move is not a constant global, reject it.
146 if (!pointsToConstantGlobal(MI->getSource()))
149 // Otherwise, the transform is safe. Remember the copy instruction.
156 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
157 /// modified by a copy from a constant global. If we can prove this, we can
158 /// replace any uses of the alloca with uses of the global directly.
159 static MemTransferInst *
160 isOnlyCopiedFromConstantGlobal(AllocaInst *AI,
161 SmallVectorImpl<Instruction *> &ToDelete) {
162 MemTransferInst *TheCopy = nullptr;
163 if (isOnlyCopiedFromConstantGlobal(AI, TheCopy, ToDelete))
168 static Instruction *simplifyAllocaArraySize(InstCombiner &IC, AllocaInst &AI) {
169 // Check for array size of 1 (scalar allocation).
170 if (!AI.isArrayAllocation()) {
171 // i32 1 is the canonical array size for scalar allocations.
172 if (AI.getArraySize()->getType()->isIntegerTy(32))
176 Value *V = IC.Builder->getInt32(1);
181 // Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1
182 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
183 Type *NewTy = ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
184 AllocaInst *New = IC.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))
194 // Now that I is pointing to the first non-allocation-inst in the block,
195 // insert our getelementptr instruction...
197 Type *IdxTy = IC.getDataLayout().getIntPtrType(AI.getType());
198 Value *NullIdx = Constant::getNullValue(IdxTy);
199 Value *Idx[2] = {NullIdx, NullIdx};
201 GetElementPtrInst::CreateInBounds(New, Idx, New->getName() + ".sub");
202 IC.InsertNewInstBefore(GEP, *It);
204 // Now make everything use the getelementptr instead of the original
206 return IC.ReplaceInstUsesWith(AI, GEP);
209 if (isa<UndefValue>(AI.getArraySize()))
210 return IC.ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
212 // Ensure that the alloca array size argument has type intptr_t, so that
213 // any casting is exposed early.
214 Type *IntPtrTy = IC.getDataLayout().getIntPtrType(AI.getType());
215 if (AI.getArraySize()->getType() != IntPtrTy) {
216 Value *V = IC.Builder->CreateIntCast(AI.getArraySize(), IntPtrTy, false);
224 Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
225 if (auto *I = simplifyAllocaArraySize(*this, AI))
228 if (AI.getAllocatedType()->isSized()) {
229 // If the alignment is 0 (unspecified), assign it the preferred alignment.
230 if (AI.getAlignment() == 0)
231 AI.setAlignment(DL.getPrefTypeAlignment(AI.getAllocatedType()));
233 // Move all alloca's of zero byte objects to the entry block and merge them
234 // together. Note that we only do this for alloca's, because malloc should
235 // allocate and return a unique pointer, even for a zero byte allocation.
236 if (DL.getTypeAllocSize(AI.getAllocatedType()) == 0) {
237 // For a zero sized alloca there is no point in doing an array allocation.
238 // This is helpful if the array size is a complicated expression not used
240 if (AI.isArrayAllocation()) {
241 AI.setOperand(0, ConstantInt::get(AI.getArraySize()->getType(), 1));
245 // Get the first instruction in the entry block.
246 BasicBlock &EntryBlock = AI.getParent()->getParent()->getEntryBlock();
247 Instruction *FirstInst = EntryBlock.getFirstNonPHIOrDbg();
248 if (FirstInst != &AI) {
249 // If the entry block doesn't start with a zero-size alloca then move
250 // this one to the start of the entry block. There is no problem with
251 // dominance as the array size was forced to a constant earlier already.
252 AllocaInst *EntryAI = dyn_cast<AllocaInst>(FirstInst);
253 if (!EntryAI || !EntryAI->getAllocatedType()->isSized() ||
254 DL.getTypeAllocSize(EntryAI->getAllocatedType()) != 0) {
255 AI.moveBefore(FirstInst);
259 // If the alignment of the entry block alloca is 0 (unspecified),
260 // assign it the preferred alignment.
261 if (EntryAI->getAlignment() == 0)
262 EntryAI->setAlignment(
263 DL.getPrefTypeAlignment(EntryAI->getAllocatedType()));
264 // Replace this zero-sized alloca with the one at the start of the entry
265 // block after ensuring that the address will be aligned enough for both
267 unsigned MaxAlign = std::max(EntryAI->getAlignment(),
269 EntryAI->setAlignment(MaxAlign);
270 if (AI.getType() != EntryAI->getType())
271 return new BitCastInst(EntryAI, AI.getType());
272 return ReplaceInstUsesWith(AI, EntryAI);
277 if (AI.getAlignment()) {
278 // Check to see if this allocation is only modified by a memcpy/memmove from
279 // a constant global whose alignment is equal to or exceeds that of the
280 // allocation. If this is the case, we can change all users to use
281 // the constant global instead. This is commonly produced by the CFE by
282 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
283 // is only subsequently read.
284 SmallVector<Instruction *, 4> ToDelete;
285 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(&AI, ToDelete)) {
286 unsigned SourceAlign = getOrEnforceKnownAlignment(
287 Copy->getSource(), AI.getAlignment(), DL, &AI, AC, DT);
288 if (AI.getAlignment() <= SourceAlign) {
289 DEBUG(dbgs() << "Found alloca equal to global: " << AI << '\n');
290 DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
291 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
292 EraseInstFromFunction(*ToDelete[i]);
293 Constant *TheSrc = cast<Constant>(Copy->getSource());
295 = ConstantExpr::getPointerBitCastOrAddrSpaceCast(TheSrc, AI.getType());
296 Instruction *NewI = ReplaceInstUsesWith(AI, Cast);
297 EraseInstFromFunction(*Copy);
304 // At last, use the generic allocation site handler to aggressively remove
306 return visitAllocSite(AI);
309 /// \brief Helper to combine a load to a new type.
311 /// This just does the work of combining a load to a new type. It handles
312 /// metadata, etc., and returns the new instruction. The \c NewTy should be the
313 /// loaded *value* type. This will convert it to a pointer, cast the operand to
314 /// that pointer type, load it, etc.
316 /// Note that this will create all of the instructions with whatever insert
317 /// point the \c InstCombiner currently is using.
318 static LoadInst *combineLoadToNewType(InstCombiner &IC, LoadInst &LI, Type *NewTy,
319 const Twine &Suffix = "") {
320 Value *Ptr = LI.getPointerOperand();
321 unsigned AS = LI.getPointerAddressSpace();
322 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
323 LI.getAllMetadata(MD);
325 LoadInst *NewLoad = IC.Builder->CreateAlignedLoad(
326 IC.Builder->CreateBitCast(Ptr, NewTy->getPointerTo(AS)),
327 LI.getAlignment(), LI.getName() + Suffix);
328 MDBuilder MDB(NewLoad->getContext());
329 for (const auto &MDPair : MD) {
330 unsigned ID = MDPair.first;
331 MDNode *N = MDPair.second;
332 // Note, essentially every kind of metadata should be preserved here! This
333 // routine is supposed to clone a load instruction changing *only its type*.
334 // The only metadata it makes sense to drop is metadata which is invalidated
335 // when the pointer type changes. This should essentially never be the case
336 // in LLVM, but we explicitly switch over only known metadata to be
337 // conservatively correct. If you are adding metadata to LLVM which pertains
338 // to loads, you almost certainly want to add it here.
340 case LLVMContext::MD_dbg:
341 case LLVMContext::MD_tbaa:
342 case LLVMContext::MD_prof:
343 case LLVMContext::MD_fpmath:
344 case LLVMContext::MD_tbaa_struct:
345 case LLVMContext::MD_invariant_load:
346 case LLVMContext::MD_alias_scope:
347 case LLVMContext::MD_noalias:
348 case LLVMContext::MD_nontemporal:
349 case LLVMContext::MD_mem_parallel_loop_access:
350 // All of these directly apply.
351 NewLoad->setMetadata(ID, N);
354 case LLVMContext::MD_nonnull:
355 // This only directly applies if the new type is also a pointer.
356 if (NewTy->isPointerTy()) {
357 NewLoad->setMetadata(ID, N);
360 // If it's integral now, translate it to !range metadata.
361 if (NewTy->isIntegerTy()) {
362 auto *ITy = cast<IntegerType>(NewTy);
363 auto *NullInt = ConstantExpr::getPtrToInt(
364 ConstantPointerNull::get(cast<PointerType>(Ptr->getType())), ITy);
366 ConstantExpr::getAdd(NullInt, ConstantInt::get(ITy, 1));
367 NewLoad->setMetadata(LLVMContext::MD_range,
368 MDB.createRange(NonNullInt, NullInt));
371 case LLVMContext::MD_align:
372 case LLVMContext::MD_dereferenceable:
373 case LLVMContext::MD_dereferenceable_or_null:
374 // These only directly apply if the new type is also a pointer.
375 if (NewTy->isPointerTy())
376 NewLoad->setMetadata(ID, N);
378 case LLVMContext::MD_range:
379 // FIXME: It would be nice to propagate this in some way, but the type
380 // conversions make it hard. If the new type is a pointer, we could
381 // translate it to !nonnull metadata.
388 /// \brief Combine a store to a new type.
390 /// Returns the newly created store instruction.
391 static StoreInst *combineStoreToNewValue(InstCombiner &IC, StoreInst &SI, Value *V) {
392 Value *Ptr = SI.getPointerOperand();
393 unsigned AS = SI.getPointerAddressSpace();
394 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
395 SI.getAllMetadata(MD);
397 StoreInst *NewStore = IC.Builder->CreateAlignedStore(
398 V, IC.Builder->CreateBitCast(Ptr, V->getType()->getPointerTo(AS)),
400 for (const auto &MDPair : MD) {
401 unsigned ID = MDPair.first;
402 MDNode *N = MDPair.second;
403 // Note, essentially every kind of metadata should be preserved here! This
404 // routine is supposed to clone a store instruction changing *only its
405 // type*. The only metadata it makes sense to drop is metadata which is
406 // invalidated when the pointer type changes. This should essentially
407 // never be the case in LLVM, but we explicitly switch over only known
408 // metadata to be conservatively correct. If you are adding metadata to
409 // LLVM which pertains to stores, you almost certainly want to add it
412 case LLVMContext::MD_dbg:
413 case LLVMContext::MD_tbaa:
414 case LLVMContext::MD_prof:
415 case LLVMContext::MD_fpmath:
416 case LLVMContext::MD_tbaa_struct:
417 case LLVMContext::MD_alias_scope:
418 case LLVMContext::MD_noalias:
419 case LLVMContext::MD_nontemporal:
420 case LLVMContext::MD_mem_parallel_loop_access:
421 // All of these directly apply.
422 NewStore->setMetadata(ID, N);
425 case LLVMContext::MD_invariant_load:
426 case LLVMContext::MD_nonnull:
427 case LLVMContext::MD_range:
428 case LLVMContext::MD_align:
429 case LLVMContext::MD_dereferenceable:
430 case LLVMContext::MD_dereferenceable_or_null:
431 // These don't apply for stores.
439 /// \brief Combine loads to match the type of value their uses after looking
440 /// through intervening bitcasts.
442 /// The core idea here is that if the result of a load is used in an operation,
443 /// we should load the type most conducive to that operation. For example, when
444 /// loading an integer and converting that immediately to a pointer, we should
445 /// instead directly load a pointer.
447 /// However, this routine must never change the width of a load or the number of
448 /// loads as that would introduce a semantic change. This combine is expected to
449 /// be a semantic no-op which just allows loads to more closely model the types
450 /// of their consuming operations.
452 /// Currently, we also refuse to change the precise type used for an atomic load
453 /// or a volatile load. This is debatable, and might be reasonable to change
454 /// later. However, it is risky in case some backend or other part of LLVM is
455 /// relying on the exact type loaded to select appropriate atomic operations.
456 static Instruction *combineLoadToOperationType(InstCombiner &IC, LoadInst &LI) {
457 // FIXME: We could probably with some care handle both volatile and atomic
458 // loads here but it isn't clear that this is important.
465 Type *Ty = LI.getType();
466 const DataLayout &DL = IC.getDataLayout();
468 // Try to canonicalize loads which are only ever stored to operate over
469 // integers instead of any other type. We only do this when the loaded type
470 // is sized and has a size exactly the same as its store size and the store
471 // size is a legal integer type.
472 if (!Ty->isIntegerTy() && Ty->isSized() &&
473 DL.isLegalInteger(DL.getTypeStoreSizeInBits(Ty)) &&
474 DL.getTypeStoreSizeInBits(Ty) == DL.getTypeSizeInBits(Ty)) {
475 if (std::all_of(LI.user_begin(), LI.user_end(), [&LI](User *U) {
476 auto *SI = dyn_cast<StoreInst>(U);
477 return SI && SI->getPointerOperand() != &LI;
479 LoadInst *NewLoad = combineLoadToNewType(
481 Type::getIntNTy(LI.getContext(), DL.getTypeStoreSizeInBits(Ty)));
482 // Replace all the stores with stores of the newly loaded value.
483 for (auto UI = LI.user_begin(), UE = LI.user_end(); UI != UE;) {
484 auto *SI = cast<StoreInst>(*UI++);
485 IC.Builder->SetInsertPoint(SI);
486 combineStoreToNewValue(IC, *SI, NewLoad);
487 IC.EraseInstFromFunction(*SI);
489 assert(LI.use_empty() && "Failed to remove all users of the load!");
490 // Return the old load so the combiner can delete it safely.
495 // Fold away bit casts of the loaded value by loading the desired type.
496 // We can do this for BitCastInsts as well as casts from and to pointer types,
497 // as long as those are noops (i.e., the source or dest type have the same
498 // bitwidth as the target's pointers).
500 if (auto* CI = dyn_cast<CastInst>(LI.user_back())) {
501 if (CI->isNoopCast(DL)) {
502 LoadInst *NewLoad = combineLoadToNewType(IC, LI, CI->getDestTy());
503 CI->replaceAllUsesWith(NewLoad);
504 IC.EraseInstFromFunction(*CI);
509 // FIXME: We should also canonicalize loads of vectors when their elements are
510 // cast to other types.
514 static Instruction *unpackLoadToAggregate(InstCombiner &IC, LoadInst &LI) {
515 // FIXME: We could probably with some care handle both volatile and atomic
516 // stores here but it isn't clear that this is important.
520 Type *T = LI.getType();
521 if (!T->isAggregateType())
524 assert(LI.getAlignment() && "Alignment must be set at this point");
526 if (auto *ST = dyn_cast<StructType>(T)) {
527 // If the struct only have one element, we unpack.
528 unsigned Count = ST->getNumElements();
530 LoadInst *NewLoad = combineLoadToNewType(IC, LI, ST->getTypeAtIndex(0U),
532 return IC.ReplaceInstUsesWith(LI, IC.Builder->CreateInsertValue(
533 UndefValue::get(T), NewLoad, 0, LI.getName()));
536 // We don't want to break loads with padding here as we'd loose
537 // the knowledge that padding exists for the rest of the pipeline.
538 const DataLayout &DL = IC.getDataLayout();
539 auto *SL = DL.getStructLayout(ST);
540 if (SL->hasPadding())
543 auto Name = LI.getName();
544 SmallString<16> LoadName = Name;
545 LoadName += ".unpack";
546 SmallString<16> EltName = Name;
548 auto *Addr = LI.getPointerOperand();
549 Value *V = UndefValue::get(T);
550 auto *IdxType = Type::getInt32Ty(ST->getContext());
551 auto *Zero = ConstantInt::get(IdxType, 0);
552 for (unsigned i = 0; i < Count; i++) {
553 Value *Indices[2] = {
555 ConstantInt::get(IdxType, i),
557 auto *Ptr = IC.Builder->CreateInBoundsGEP(ST, Addr, makeArrayRef(Indices), EltName);
558 auto *L = IC.Builder->CreateLoad(ST->getTypeAtIndex(i), Ptr, LoadName);
559 V = IC.Builder->CreateInsertValue(V, L, i);
563 return IC.ReplaceInstUsesWith(LI, V);
566 if (auto *AT = dyn_cast<ArrayType>(T)) {
567 // If the array only have one element, we unpack.
568 if (AT->getNumElements() == 1) {
569 LoadInst *NewLoad = combineLoadToNewType(IC, LI, AT->getElementType(),
571 return IC.ReplaceInstUsesWith(LI, IC.Builder->CreateInsertValue(
572 UndefValue::get(T), NewLoad, 0, LI.getName()));
579 // If we can determine that all possible objects pointed to by the provided
580 // pointer value are, not only dereferenceable, but also definitively less than
581 // or equal to the provided maximum size, then return true. Otherwise, return
582 // false (constant global values and allocas fall into this category).
584 // FIXME: This should probably live in ValueTracking (or similar).
585 static bool isObjectSizeLessThanOrEq(Value *V, uint64_t MaxSize,
586 const DataLayout &DL) {
587 SmallPtrSet<Value *, 4> Visited;
588 SmallVector<Value *, 4> Worklist(1, V);
591 Value *P = Worklist.pop_back_val();
592 P = P->stripPointerCasts();
594 if (!Visited.insert(P).second)
597 if (SelectInst *SI = dyn_cast<SelectInst>(P)) {
598 Worklist.push_back(SI->getTrueValue());
599 Worklist.push_back(SI->getFalseValue());
603 if (PHINode *PN = dyn_cast<PHINode>(P)) {
604 for (Value *IncValue : PN->incoming_values())
605 Worklist.push_back(IncValue);
609 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(P)) {
610 if (GA->mayBeOverridden())
612 Worklist.push_back(GA->getAliasee());
616 // If we know how big this object is, and it is less than MaxSize, continue
617 // searching. Otherwise, return false.
618 if (AllocaInst *AI = dyn_cast<AllocaInst>(P)) {
619 if (!AI->getAllocatedType()->isSized())
622 ConstantInt *CS = dyn_cast<ConstantInt>(AI->getArraySize());
626 uint64_t TypeSize = DL.getTypeAllocSize(AI->getAllocatedType());
627 // Make sure that, even if the multiplication below would wrap as an
628 // uint64_t, we still do the right thing.
629 if ((CS->getValue().zextOrSelf(128)*APInt(128, TypeSize)).ugt(MaxSize))
634 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(P)) {
635 if (!GV->hasDefinitiveInitializer() || !GV->isConstant())
638 uint64_t InitSize = DL.getTypeAllocSize(GV->getType()->getElementType());
639 if (InitSize > MaxSize)
645 } while (!Worklist.empty());
650 // If we're indexing into an object of a known size, and the outer index is
651 // not a constant, but having any value but zero would lead to undefined
652 // behavior, replace it with zero.
654 // For example, if we have:
655 // @f.a = private unnamed_addr constant [1 x i32] [i32 12], align 4
657 // %arrayidx = getelementptr inbounds [1 x i32]* @f.a, i64 0, i64 %x
658 // ... = load i32* %arrayidx, align 4
659 // Then we know that we can replace %x in the GEP with i64 0.
661 // FIXME: We could fold any GEP index to zero that would cause UB if it were
662 // not zero. Currently, we only handle the first such index. Also, we could
663 // also search through non-zero constant indices if we kept track of the
664 // offsets those indices implied.
665 static bool canReplaceGEPIdxWithZero(InstCombiner &IC, GetElementPtrInst *GEPI,
666 Instruction *MemI, unsigned &Idx) {
667 if (GEPI->getNumOperands() < 2)
670 // Find the first non-zero index of a GEP. If all indices are zero, return
671 // one past the last index.
672 auto FirstNZIdx = [](const GetElementPtrInst *GEPI) {
674 for (unsigned IE = GEPI->getNumOperands(); I != IE; ++I) {
675 Value *V = GEPI->getOperand(I);
676 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V))
686 // Skip through initial 'zero' indices, and find the corresponding pointer
687 // type. See if the next index is not a constant.
688 Idx = FirstNZIdx(GEPI);
689 if (Idx == GEPI->getNumOperands())
691 if (isa<Constant>(GEPI->getOperand(Idx)))
694 SmallVector<Value *, 4> Ops(GEPI->idx_begin(), GEPI->idx_begin() + Idx);
695 Type *AllocTy = GetElementPtrInst::getIndexedType(
696 cast<PointerType>(GEPI->getOperand(0)->getType()->getScalarType())
699 if (!AllocTy || !AllocTy->isSized())
701 const DataLayout &DL = IC.getDataLayout();
702 uint64_t TyAllocSize = DL.getTypeAllocSize(AllocTy);
704 // If there are more indices after the one we might replace with a zero, make
705 // sure they're all non-negative. If any of them are negative, the overall
706 // address being computed might be before the base address determined by the
707 // first non-zero index.
708 auto IsAllNonNegative = [&]() {
709 for (unsigned i = Idx+1, e = GEPI->getNumOperands(); i != e; ++i) {
710 bool KnownNonNegative, KnownNegative;
711 IC.ComputeSignBit(GEPI->getOperand(i), KnownNonNegative,
712 KnownNegative, 0, MemI);
713 if (KnownNonNegative)
721 // FIXME: If the GEP is not inbounds, and there are extra indices after the
722 // one we'll replace, those could cause the address computation to wrap
723 // (rendering the IsAllNonNegative() check below insufficient). We can do
724 // better, ignoring zero indices (and other indices we can prove small
725 // enough not to wrap).
726 if (Idx+1 != GEPI->getNumOperands() && !GEPI->isInBounds())
729 // Note that isObjectSizeLessThanOrEq will return true only if the pointer is
730 // also known to be dereferenceable.
731 return isObjectSizeLessThanOrEq(GEPI->getOperand(0), TyAllocSize, DL) &&
735 // If we're indexing into an object with a variable index for the memory
736 // access, but the object has only one element, we can assume that the index
737 // will always be zero. If we replace the GEP, return it.
738 template <typename T>
739 static Instruction *replaceGEPIdxWithZero(InstCombiner &IC, Value *Ptr,
741 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr)) {
743 if (canReplaceGEPIdxWithZero(IC, GEPI, &MemI, Idx)) {
744 Instruction *NewGEPI = GEPI->clone();
745 NewGEPI->setOperand(Idx,
746 ConstantInt::get(GEPI->getOperand(Idx)->getType(), 0));
747 NewGEPI->insertBefore(GEPI);
748 MemI.setOperand(MemI.getPointerOperandIndex(), NewGEPI);
756 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
757 Value *Op = LI.getOperand(0);
759 // Try to canonicalize the loaded type.
760 if (Instruction *Res = combineLoadToOperationType(*this, LI))
763 // Attempt to improve the alignment.
764 unsigned KnownAlign = getOrEnforceKnownAlignment(
765 Op, DL.getPrefTypeAlignment(LI.getType()), DL, &LI, AC, DT);
766 unsigned LoadAlign = LI.getAlignment();
767 unsigned EffectiveLoadAlign =
768 LoadAlign != 0 ? LoadAlign : DL.getABITypeAlignment(LI.getType());
770 if (KnownAlign > EffectiveLoadAlign)
771 LI.setAlignment(KnownAlign);
772 else if (LoadAlign == 0)
773 LI.setAlignment(EffectiveLoadAlign);
775 // Replace GEP indices if possible.
776 if (Instruction *NewGEPI = replaceGEPIdxWithZero(*this, Op, LI)) {
777 Worklist.Add(NewGEPI);
781 // None of the following transforms are legal for volatile/atomic loads.
782 // FIXME: Some of it is okay for atomic loads; needs refactoring.
783 if (!LI.isSimple()) return nullptr;
785 if (Instruction *Res = unpackLoadToAggregate(*this, LI))
788 // Do really simple store-to-load forwarding and load CSE, to catch cases
789 // where there are several consecutive memory accesses to the same location,
790 // separated by a few arithmetic operations.
791 BasicBlock::iterator BBI(LI);
793 if (Value *AvailableVal =
794 FindAvailableLoadedValue(Op, LI.getParent(), BBI,
795 DefMaxInstsToScan, AA, &AATags)) {
796 if (LoadInst *NLI = dyn_cast<LoadInst>(AvailableVal)) {
797 unsigned KnownIDs[] = {
798 LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
799 LLVMContext::MD_noalias, LLVMContext::MD_range,
800 LLVMContext::MD_invariant_load, LLVMContext::MD_nonnull,
801 LLVMContext::MD_invariant_group, LLVMContext::MD_align,
802 LLVMContext::MD_dereferenceable,
803 LLVMContext::MD_dereferenceable_or_null};
804 combineMetadata(NLI, &LI, KnownIDs);
807 return ReplaceInstUsesWith(
808 LI, Builder->CreateBitOrPointerCast(AvailableVal, LI.getType(),
809 LI.getName() + ".cast"));
812 // load(gep null, ...) -> unreachable
813 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
814 const Value *GEPI0 = GEPI->getOperand(0);
815 // TODO: Consider a target hook for valid address spaces for this xform.
816 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
817 // Insert a new store to null instruction before the load to indicate
818 // that this code is not reachable. We do this instead of inserting
819 // an unreachable instruction directly because we cannot modify the
821 new StoreInst(UndefValue::get(LI.getType()),
822 Constant::getNullValue(Op->getType()), &LI);
823 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
827 // load null/undef -> unreachable
828 // TODO: Consider a target hook for valid address spaces for this xform.
829 if (isa<UndefValue>(Op) ||
830 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) {
831 // Insert a new store to null instruction before the load to indicate that
832 // this code is not reachable. We do this instead of inserting an
833 // unreachable instruction directly because we cannot modify the CFG.
834 new StoreInst(UndefValue::get(LI.getType()),
835 Constant::getNullValue(Op->getType()), &LI);
836 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
839 if (Op->hasOneUse()) {
840 // Change select and PHI nodes to select values instead of addresses: this
841 // helps alias analysis out a lot, allows many others simplifications, and
842 // exposes redundancy in the code.
844 // Note that we cannot do the transformation unless we know that the
845 // introduced loads cannot trap! Something like this is valid as long as
846 // the condition is always false: load (select bool %C, int* null, int* %G),
847 // but it would not be valid if we transformed it to load from null
850 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
851 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
852 unsigned Align = LI.getAlignment();
853 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI, Align) &&
854 isSafeToLoadUnconditionally(SI->getOperand(2), SI, Align)) {
855 LoadInst *V1 = Builder->CreateLoad(SI->getOperand(1),
856 SI->getOperand(1)->getName()+".val");
857 LoadInst *V2 = Builder->CreateLoad(SI->getOperand(2),
858 SI->getOperand(2)->getName()+".val");
859 V1->setAlignment(Align);
860 V2->setAlignment(Align);
861 return SelectInst::Create(SI->getCondition(), V1, V2);
864 // load (select (cond, null, P)) -> load P
865 if (isa<ConstantPointerNull>(SI->getOperand(1)) &&
866 LI.getPointerAddressSpace() == 0) {
867 LI.setOperand(0, SI->getOperand(2));
871 // load (select (cond, P, null)) -> load P
872 if (isa<ConstantPointerNull>(SI->getOperand(2)) &&
873 LI.getPointerAddressSpace() == 0) {
874 LI.setOperand(0, SI->getOperand(1));
882 /// \brief Combine stores to match the type of value being stored.
884 /// The core idea here is that the memory does not have any intrinsic type and
885 /// where we can we should match the type of a store to the type of value being
888 /// However, this routine must never change the width of a store or the number of
889 /// stores as that would introduce a semantic change. This combine is expected to
890 /// be a semantic no-op which just allows stores to more closely model the types
891 /// of their incoming values.
893 /// Currently, we also refuse to change the precise type used for an atomic or
894 /// volatile store. This is debatable, and might be reasonable to change later.
895 /// However, it is risky in case some backend or other part of LLVM is relying
896 /// on the exact type stored to select appropriate atomic operations.
898 /// \returns true if the store was successfully combined away. This indicates
899 /// the caller must erase the store instruction. We have to let the caller erase
900 /// the store instruction as otherwise there is no way to signal whether it was
901 /// combined or not: IC.EraseInstFromFunction returns a null pointer.
902 static bool combineStoreToValueType(InstCombiner &IC, StoreInst &SI) {
903 // FIXME: We could probably with some care handle both volatile and atomic
904 // stores here but it isn't clear that this is important.
908 Value *V = SI.getValueOperand();
910 // Fold away bit casts of the stored value by storing the original type.
911 if (auto *BC = dyn_cast<BitCastInst>(V)) {
912 V = BC->getOperand(0);
913 combineStoreToNewValue(IC, SI, V);
917 // FIXME: We should also canonicalize loads of vectors when their elements are
918 // cast to other types.
922 static bool unpackStoreToAggregate(InstCombiner &IC, StoreInst &SI) {
923 // FIXME: We could probably with some care handle both volatile and atomic
924 // stores here but it isn't clear that this is important.
928 Value *V = SI.getValueOperand();
929 Type *T = V->getType();
931 if (!T->isAggregateType())
934 if (auto *ST = dyn_cast<StructType>(T)) {
935 // If the struct only have one element, we unpack.
936 unsigned Count = ST->getNumElements();
938 V = IC.Builder->CreateExtractValue(V, 0);
939 combineStoreToNewValue(IC, SI, V);
943 // We don't want to break loads with padding here as we'd loose
944 // the knowledge that padding exists for the rest of the pipeline.
945 const DataLayout &DL = IC.getDataLayout();
946 auto *SL = DL.getStructLayout(ST);
947 if (SL->hasPadding())
950 SmallString<16> EltName = V->getName();
952 auto *Addr = SI.getPointerOperand();
953 SmallString<16> AddrName = Addr->getName();
954 AddrName += ".repack";
955 auto *IdxType = Type::getInt32Ty(ST->getContext());
956 auto *Zero = ConstantInt::get(IdxType, 0);
957 for (unsigned i = 0; i < Count; i++) {
958 Value *Indices[2] = {
960 ConstantInt::get(IdxType, i),
962 auto *Ptr = IC.Builder->CreateInBoundsGEP(ST, Addr, makeArrayRef(Indices), AddrName);
963 auto *Val = IC.Builder->CreateExtractValue(V, i, EltName);
964 IC.Builder->CreateStore(Val, Ptr);
970 if (auto *AT = dyn_cast<ArrayType>(T)) {
971 // If the array only have one element, we unpack.
972 if (AT->getNumElements() == 1) {
973 V = IC.Builder->CreateExtractValue(V, 0);
974 combineStoreToNewValue(IC, SI, V);
982 /// equivalentAddressValues - Test if A and B will obviously have the same
983 /// value. This includes recognizing that %t0 and %t1 will have the same
984 /// value in code like this:
985 /// %t0 = getelementptr \@a, 0, 3
986 /// store i32 0, i32* %t0
987 /// %t1 = getelementptr \@a, 0, 3
988 /// %t2 = load i32* %t1
990 static bool equivalentAddressValues(Value *A, Value *B) {
991 // Test if the values are trivially equivalent.
992 if (A == B) return true;
994 // Test if the values come form identical arithmetic instructions.
995 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
996 // its only used to compare two uses within the same basic block, which
997 // means that they'll always either have the same value or one of them
998 // will have an undefined value.
999 if (isa<BinaryOperator>(A) ||
1002 isa<GetElementPtrInst>(A))
1003 if (Instruction *BI = dyn_cast<Instruction>(B))
1004 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
1007 // Otherwise they may not be equivalent.
1011 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
1012 Value *Val = SI.getOperand(0);
1013 Value *Ptr = SI.getOperand(1);
1015 // Try to canonicalize the stored type.
1016 if (combineStoreToValueType(*this, SI))
1017 return EraseInstFromFunction(SI);
1019 // Attempt to improve the alignment.
1020 unsigned KnownAlign = getOrEnforceKnownAlignment(
1021 Ptr, DL.getPrefTypeAlignment(Val->getType()), DL, &SI, AC, DT);
1022 unsigned StoreAlign = SI.getAlignment();
1023 unsigned EffectiveStoreAlign =
1024 StoreAlign != 0 ? StoreAlign : DL.getABITypeAlignment(Val->getType());
1026 if (KnownAlign > EffectiveStoreAlign)
1027 SI.setAlignment(KnownAlign);
1028 else if (StoreAlign == 0)
1029 SI.setAlignment(EffectiveStoreAlign);
1031 // Try to canonicalize the stored type.
1032 if (unpackStoreToAggregate(*this, SI))
1033 return EraseInstFromFunction(SI);
1035 // Replace GEP indices if possible.
1036 if (Instruction *NewGEPI = replaceGEPIdxWithZero(*this, Ptr, SI)) {
1037 Worklist.Add(NewGEPI);
1041 // Don't hack volatile/ordered stores.
1042 // FIXME: Some bits are legal for ordered atomic stores; needs refactoring.
1043 if (!SI.isUnordered()) return nullptr;
1045 // If the RHS is an alloca with a single use, zapify the store, making the
1047 if (Ptr->hasOneUse()) {
1048 if (isa<AllocaInst>(Ptr))
1049 return EraseInstFromFunction(SI);
1050 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
1051 if (isa<AllocaInst>(GEP->getOperand(0))) {
1052 if (GEP->getOperand(0)->hasOneUse())
1053 return EraseInstFromFunction(SI);
1058 // Do really simple DSE, to catch cases where there are several consecutive
1059 // stores to the same location, separated by a few arithmetic operations. This
1060 // situation often occurs with bitfield accesses.
1061 BasicBlock::iterator BBI(SI);
1062 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
1065 // Don't count debug info directives, lest they affect codegen,
1066 // and we skip pointer-to-pointer bitcasts, which are NOPs.
1067 if (isa<DbgInfoIntrinsic>(BBI) ||
1068 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
1073 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
1074 // Prev store isn't volatile, and stores to the same location?
1075 if (PrevSI->isUnordered() && equivalentAddressValues(PrevSI->getOperand(1),
1076 SI.getOperand(1))) {
1079 EraseInstFromFunction(*PrevSI);
1085 // If this is a load, we have to stop. However, if the loaded value is from
1086 // the pointer we're loading and is producing the pointer we're storing,
1087 // then *this* store is dead (X = load P; store X -> P).
1088 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
1089 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr)) {
1090 assert(SI.isUnordered() && "can't eliminate ordering operation");
1091 return EraseInstFromFunction(SI);
1094 // Otherwise, this is a load from some other location. Stores before it
1099 // Don't skip over loads or things that can modify memory.
1100 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
1104 // store X, null -> turns into 'unreachable' in SimplifyCFG
1105 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
1106 if (!isa<UndefValue>(Val)) {
1107 SI.setOperand(0, UndefValue::get(Val->getType()));
1108 if (Instruction *U = dyn_cast<Instruction>(Val))
1109 Worklist.Add(U); // Dropped a use.
1111 return nullptr; // Do not modify these!
1114 // store undef, Ptr -> noop
1115 if (isa<UndefValue>(Val))
1116 return EraseInstFromFunction(SI);
1118 // The code below needs to be audited and adjusted for unordered atomics
1122 // If this store is the last instruction in the basic block (possibly
1123 // excepting debug info instructions), and if the block ends with an
1124 // unconditional branch, try to move it to the successor block.
1125 BBI = SI.getIterator();
1128 } while (isa<DbgInfoIntrinsic>(BBI) ||
1129 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy()));
1130 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
1131 if (BI->isUnconditional())
1132 if (SimplifyStoreAtEndOfBlock(SI))
1133 return nullptr; // xform done!
1138 /// SimplifyStoreAtEndOfBlock - Turn things like:
1139 /// if () { *P = v1; } else { *P = v2 }
1140 /// into a phi node with a store in the successor.
1142 /// Simplify things like:
1143 /// *P = v1; if () { *P = v2; }
1144 /// into a phi node with a store in the successor.
1146 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
1147 BasicBlock *StoreBB = SI.getParent();
1149 // Check to see if the successor block has exactly two incoming edges. If
1150 // so, see if the other predecessor contains a store to the same location.
1151 // if so, insert a PHI node (if needed) and move the stores down.
1152 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
1154 // Determine whether Dest has exactly two predecessors and, if so, compute
1155 // the other predecessor.
1156 pred_iterator PI = pred_begin(DestBB);
1157 BasicBlock *P = *PI;
1158 BasicBlock *OtherBB = nullptr;
1163 if (++PI == pred_end(DestBB))
1172 if (++PI != pred_end(DestBB))
1175 // Bail out if all the relevant blocks aren't distinct (this can happen,
1176 // for example, if SI is in an infinite loop)
1177 if (StoreBB == DestBB || OtherBB == DestBB)
1180 // Verify that the other block ends in a branch and is not otherwise empty.
1181 BasicBlock::iterator BBI(OtherBB->getTerminator());
1182 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
1183 if (!OtherBr || BBI == OtherBB->begin())
1186 // If the other block ends in an unconditional branch, check for the 'if then
1187 // else' case. there is an instruction before the branch.
1188 StoreInst *OtherStore = nullptr;
1189 if (OtherBr->isUnconditional()) {
1191 // Skip over debugging info.
1192 while (isa<DbgInfoIntrinsic>(BBI) ||
1193 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
1194 if (BBI==OtherBB->begin())
1198 // If this isn't a store, isn't a store to the same location, or is not the
1199 // right kind of store, bail out.
1200 OtherStore = dyn_cast<StoreInst>(BBI);
1201 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
1202 !SI.isSameOperationAs(OtherStore))
1205 // Otherwise, the other block ended with a conditional branch. If one of the
1206 // destinations is StoreBB, then we have the if/then case.
1207 if (OtherBr->getSuccessor(0) != StoreBB &&
1208 OtherBr->getSuccessor(1) != StoreBB)
1211 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
1212 // if/then triangle. See if there is a store to the same ptr as SI that
1213 // lives in OtherBB.
1215 // Check to see if we find the matching store.
1216 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
1217 if (OtherStore->getOperand(1) != SI.getOperand(1) ||
1218 !SI.isSameOperationAs(OtherStore))
1222 // If we find something that may be using or overwriting the stored
1223 // value, or if we run out of instructions, we can't do the xform.
1224 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
1225 BBI == OtherBB->begin())
1229 // In order to eliminate the store in OtherBr, we have to
1230 // make sure nothing reads or overwrites the stored value in
1232 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
1233 // FIXME: This should really be AA driven.
1234 if (I->mayReadFromMemory() || I->mayWriteToMemory())
1239 // Insert a PHI node now if we need it.
1240 Value *MergedVal = OtherStore->getOperand(0);
1241 if (MergedVal != SI.getOperand(0)) {
1242 PHINode *PN = PHINode::Create(MergedVal->getType(), 2, "storemerge");
1243 PN->addIncoming(SI.getOperand(0), SI.getParent());
1244 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
1245 MergedVal = InsertNewInstBefore(PN, DestBB->front());
1248 // Advance to a place where it is safe to insert the new store and
1250 BBI = DestBB->getFirstInsertionPt();
1251 StoreInst *NewSI = new StoreInst(MergedVal, SI.getOperand(1),
1255 SI.getSynchScope());
1256 InsertNewInstBefore(NewSI, *BBI);
1257 NewSI->setDebugLoc(OtherStore->getDebugLoc());
1259 // If the two stores had AA tags, merge them.
1261 SI.getAAMetadata(AATags);
1263 OtherStore->getAAMetadata(AATags, /* Merge = */ true);
1264 NewSI->setAAMetadata(AATags);
1267 // Nuke the old stores.
1268 EraseInstFromFunction(SI);
1269 EraseInstFromFunction(*OtherStore);