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 "InstCombine.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(Copy->getSource(),
274 if (AI.getAlignment() <= SourceAlign) {
275 DEBUG(dbgs() << "Found alloca equal to global: " << AI << '\n');
276 DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
277 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
278 EraseInstFromFunction(*ToDelete[i]);
279 Constant *TheSrc = cast<Constant>(Copy->getSource());
281 = ConstantExpr::getPointerBitCastOrAddrSpaceCast(TheSrc, AI.getType());
282 Instruction *NewI = ReplaceInstUsesWith(AI, Cast);
283 EraseInstFromFunction(*Copy);
290 // At last, use the generic allocation site handler to aggressively remove
292 return visitAllocSite(AI);
295 /// \brief Helper to combine a load to a new type.
297 /// This just does the work of combining a load to a new type. It handles
298 /// metadata, etc., and returns the new instruction. The \c NewTy should be the
299 /// loaded *value* type. This will convert it to a pointer, cast the operand to
300 /// that pointer type, load it, etc.
302 /// Note that this will create all of the instructions with whatever insert
303 /// point the \c InstCombiner currently is using.
304 static LoadInst *combineLoadToNewType(InstCombiner &IC, LoadInst &LI, Type *NewTy) {
305 Value *Ptr = LI.getPointerOperand();
306 unsigned AS = LI.getPointerAddressSpace();
307 SmallVector<std::pair<unsigned, MDNode *>, 8> MD;
308 LI.getAllMetadata(MD);
310 LoadInst *NewLoad = IC.Builder->CreateAlignedLoad(
311 IC.Builder->CreateBitCast(Ptr, NewTy->getPointerTo(AS)),
312 LI.getAlignment(), LI.getName());
313 for (const auto &MDPair : MD) {
314 unsigned ID = MDPair.first;
315 MDNode *N = MDPair.second;
316 // Note, essentially every kind of metadata should be preserved here! This
317 // routine is supposed to clone a load instruction changing *only its type*.
318 // The only metadata it makes sense to drop is metadata which is invalidated
319 // when the pointer type changes. This should essentially never be the case
320 // in LLVM, but we explicitly switch over only known metadata to be
321 // conservatively correct. If you are adding metadata to LLVM which pertains
322 // to loads, you almost certainly want to add it here.
324 case LLVMContext::MD_dbg:
325 case LLVMContext::MD_tbaa:
326 case LLVMContext::MD_prof:
327 case LLVMContext::MD_fpmath:
328 case LLVMContext::MD_tbaa_struct:
329 case LLVMContext::MD_invariant_load:
330 case LLVMContext::MD_alias_scope:
331 case LLVMContext::MD_noalias:
332 case LLVMContext::MD_nontemporal:
333 case LLVMContext::MD_mem_parallel_loop_access:
334 case LLVMContext::MD_nonnull:
335 // All of these directly apply.
336 NewLoad->setMetadata(ID, N);
339 case LLVMContext::MD_range:
340 // FIXME: It would be nice to propagate this in some way, but the type
341 // conversions make it hard.
348 /// \brief Combine loads to match the type of value their uses after looking
349 /// through intervening bitcasts.
351 /// The core idea here is that if the result of a load is used in an operation,
352 /// we should load the type most conducive to that operation. For example, when
353 /// loading an integer and converting that immediately to a pointer, we should
354 /// instead directly load a pointer.
356 /// However, this routine must never change the width of a load or the number of
357 /// loads as that would introduce a semantic change. This combine is expected to
358 /// be a semantic no-op which just allows loads to more closely model the types
359 /// of their consuming operations.
361 /// Currently, we also refuse to change the precise type used for an atomic load
362 /// or a volatile load. This is debatable, and might be reasonable to change
363 /// later. However, it is risky in case some backend or other part of LLVM is
364 /// relying on the exact type loaded to select appropriate atomic operations.
365 static Instruction *combineLoadToOperationType(InstCombiner &IC, LoadInst &LI) {
366 // FIXME: We could probably with some care handle both volatile and atomic
367 // loads here but it isn't clear that this is important.
375 // Fold away bit casts of the loaded value by loading the desired type.
377 if (auto *BC = dyn_cast<BitCastInst>(LI.user_back())) {
378 LoadInst *NewLoad = combineLoadToNewType(IC, LI, BC->getDestTy());
379 BC->replaceAllUsesWith(NewLoad);
380 IC.EraseInstFromFunction(*BC);
384 // FIXME: We should also canonicalize loads of vectors when their elements are
385 // cast to other types.
389 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
390 Value *Op = LI.getOperand(0);
392 // Try to canonicalize the loaded type.
393 if (Instruction *Res = combineLoadToOperationType(*this, LI))
396 // Attempt to improve the alignment.
398 unsigned KnownAlign =
399 getOrEnforceKnownAlignment(Op, DL->getPrefTypeAlignment(LI.getType()),
401 unsigned LoadAlign = LI.getAlignment();
402 unsigned EffectiveLoadAlign = LoadAlign != 0 ? LoadAlign :
403 DL->getABITypeAlignment(LI.getType());
405 if (KnownAlign > EffectiveLoadAlign)
406 LI.setAlignment(KnownAlign);
407 else if (LoadAlign == 0)
408 LI.setAlignment(EffectiveLoadAlign);
411 // None of the following transforms are legal for volatile/atomic loads.
412 // FIXME: Some of it is okay for atomic loads; needs refactoring.
413 if (!LI.isSimple()) return nullptr;
415 // Do really simple store-to-load forwarding and load CSE, to catch cases
416 // where there are several consecutive memory accesses to the same location,
417 // separated by a few arithmetic operations.
418 BasicBlock::iterator BBI = &LI;
419 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
420 return ReplaceInstUsesWith(
421 LI, Builder->CreateBitCast(AvailableVal, LI.getType()));
423 // load(gep null, ...) -> unreachable
424 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
425 const Value *GEPI0 = GEPI->getOperand(0);
426 // TODO: Consider a target hook for valid address spaces for this xform.
427 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
428 // Insert a new store to null instruction before the load to indicate
429 // that this code is not reachable. We do this instead of inserting
430 // an unreachable instruction directly because we cannot modify the
432 new StoreInst(UndefValue::get(LI.getType()),
433 Constant::getNullValue(Op->getType()), &LI);
434 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
438 // load null/undef -> unreachable
439 // TODO: Consider a target hook for valid address spaces for this xform.
440 if (isa<UndefValue>(Op) ||
441 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) {
442 // Insert a new store to null instruction before the load to indicate that
443 // this code is not reachable. We do this instead of inserting an
444 // unreachable instruction directly because we cannot modify the CFG.
445 new StoreInst(UndefValue::get(LI.getType()),
446 Constant::getNullValue(Op->getType()), &LI);
447 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
450 if (Op->hasOneUse()) {
451 // Change select and PHI nodes to select values instead of addresses: this
452 // helps alias analysis out a lot, allows many others simplifications, and
453 // exposes redundancy in the code.
455 // Note that we cannot do the transformation unless we know that the
456 // introduced loads cannot trap! Something like this is valid as long as
457 // the condition is always false: load (select bool %C, int* null, int* %G),
458 // but it would not be valid if we transformed it to load from null
461 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
462 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
463 unsigned Align = LI.getAlignment();
464 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI, Align, DL) &&
465 isSafeToLoadUnconditionally(SI->getOperand(2), SI, Align, DL)) {
466 LoadInst *V1 = Builder->CreateLoad(SI->getOperand(1),
467 SI->getOperand(1)->getName()+".val");
468 LoadInst *V2 = Builder->CreateLoad(SI->getOperand(2),
469 SI->getOperand(2)->getName()+".val");
470 V1->setAlignment(Align);
471 V2->setAlignment(Align);
472 return SelectInst::Create(SI->getCondition(), V1, V2);
475 // load (select (cond, null, P)) -> load P
476 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
477 if (C->isNullValue()) {
478 LI.setOperand(0, SI->getOperand(2));
482 // load (select (cond, P, null)) -> load P
483 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
484 if (C->isNullValue()) {
485 LI.setOperand(0, SI->getOperand(1));
493 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
494 /// when possible. This makes it generally easy to do alias analysis and/or
495 /// SROA/mem2reg of the memory object.
496 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
497 User *CI = cast<User>(SI.getOperand(1));
498 Value *CastOp = CI->getOperand(0);
500 Type *DestPTy = CI->getType()->getPointerElementType();
501 PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
502 if (!SrcTy) return nullptr;
504 Type *SrcPTy = SrcTy->getElementType();
506 if (!DestPTy->isIntegerTy() && !DestPTy->isPointerTy())
509 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
510 /// to its first element. This allows us to handle things like:
511 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
513 SmallVector<Value*, 4> NewGEPIndices;
515 // If the source is an array, the code below will not succeed. Check to
516 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
518 if (SrcPTy->isArrayTy() || SrcPTy->isStructTy()) {
519 // Index through pointer.
520 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(SI.getContext()));
521 NewGEPIndices.push_back(Zero);
524 if (StructType *STy = dyn_cast<StructType>(SrcPTy)) {
525 if (!STy->getNumElements()) /* Struct can be empty {} */
527 NewGEPIndices.push_back(Zero);
528 SrcPTy = STy->getElementType(0);
529 } else if (ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
530 NewGEPIndices.push_back(Zero);
531 SrcPTy = ATy->getElementType();
537 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
540 if (!SrcPTy->isIntegerTy() && !SrcPTy->isPointerTy())
543 // If the pointers point into different address spaces don't do the
545 if (SrcTy->getAddressSpace() != CI->getType()->getPointerAddressSpace())
548 // If the pointers point to values of different sizes don't do the
550 if (!IC.getDataLayout() ||
551 IC.getDataLayout()->getTypeSizeInBits(SrcPTy) !=
552 IC.getDataLayout()->getTypeSizeInBits(DestPTy))
555 // If the pointers point to pointers to different address spaces don't do the
556 // transformation. It is not safe to introduce an addrspacecast instruction in
557 // this case since, depending on the target, addrspacecast may not be a no-op
559 if (SrcPTy->isPointerTy() && DestPTy->isPointerTy() &&
560 SrcPTy->getPointerAddressSpace() != DestPTy->getPointerAddressSpace())
563 // Okay, we are casting from one integer or pointer type to another of
564 // the same size. Instead of casting the pointer before
565 // the store, cast the value to be stored.
567 Instruction::CastOps opcode = Instruction::BitCast;
568 Type* CastSrcTy = DestPTy;
569 Type* CastDstTy = SrcPTy;
570 if (CastDstTy->isPointerTy()) {
571 if (CastSrcTy->isIntegerTy())
572 opcode = Instruction::IntToPtr;
573 } else if (CastDstTy->isIntegerTy()) {
574 if (CastSrcTy->isPointerTy())
575 opcode = Instruction::PtrToInt;
578 // SIOp0 is a pointer to aggregate and this is a store to the first field,
579 // emit a GEP to index into its first field.
580 if (!NewGEPIndices.empty())
581 CastOp = IC.Builder->CreateInBoundsGEP(CastOp, NewGEPIndices);
583 Value *SIOp0 = SI.getOperand(0);
584 NewCast = IC.Builder->CreateCast(opcode, SIOp0, CastDstTy,
585 SIOp0->getName()+".c");
586 SI.setOperand(0, NewCast);
587 SI.setOperand(1, CastOp);
591 /// equivalentAddressValues - Test if A and B will obviously have the same
592 /// value. This includes recognizing that %t0 and %t1 will have the same
593 /// value in code like this:
594 /// %t0 = getelementptr \@a, 0, 3
595 /// store i32 0, i32* %t0
596 /// %t1 = getelementptr \@a, 0, 3
597 /// %t2 = load i32* %t1
599 static bool equivalentAddressValues(Value *A, Value *B) {
600 // Test if the values are trivially equivalent.
601 if (A == B) return true;
603 // Test if the values come form identical arithmetic instructions.
604 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
605 // its only used to compare two uses within the same basic block, which
606 // means that they'll always either have the same value or one of them
607 // will have an undefined value.
608 if (isa<BinaryOperator>(A) ||
611 isa<GetElementPtrInst>(A))
612 if (Instruction *BI = dyn_cast<Instruction>(B))
613 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
616 // Otherwise they may not be equivalent.
620 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
621 Value *Val = SI.getOperand(0);
622 Value *Ptr = SI.getOperand(1);
624 // Attempt to improve the alignment.
626 unsigned KnownAlign =
627 getOrEnforceKnownAlignment(Ptr, DL->getPrefTypeAlignment(Val->getType()),
629 unsigned StoreAlign = SI.getAlignment();
630 unsigned EffectiveStoreAlign = StoreAlign != 0 ? StoreAlign :
631 DL->getABITypeAlignment(Val->getType());
633 if (KnownAlign > EffectiveStoreAlign)
634 SI.setAlignment(KnownAlign);
635 else if (StoreAlign == 0)
636 SI.setAlignment(EffectiveStoreAlign);
639 // Don't hack volatile/atomic stores.
640 // FIXME: Some bits are legal for atomic stores; needs refactoring.
641 if (!SI.isSimple()) return nullptr;
643 // If the RHS is an alloca with a single use, zapify the store, making the
645 if (Ptr->hasOneUse()) {
646 if (isa<AllocaInst>(Ptr))
647 return EraseInstFromFunction(SI);
648 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
649 if (isa<AllocaInst>(GEP->getOperand(0))) {
650 if (GEP->getOperand(0)->hasOneUse())
651 return EraseInstFromFunction(SI);
656 // Do really simple DSE, to catch cases where there are several consecutive
657 // stores to the same location, separated by a few arithmetic operations. This
658 // situation often occurs with bitfield accesses.
659 BasicBlock::iterator BBI = &SI;
660 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
663 // Don't count debug info directives, lest they affect codegen,
664 // and we skip pointer-to-pointer bitcasts, which are NOPs.
665 if (isa<DbgInfoIntrinsic>(BBI) ||
666 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
671 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
672 // Prev store isn't volatile, and stores to the same location?
673 if (PrevSI->isSimple() && equivalentAddressValues(PrevSI->getOperand(1),
677 EraseInstFromFunction(*PrevSI);
683 // If this is a load, we have to stop. However, if the loaded value is from
684 // the pointer we're loading and is producing the pointer we're storing,
685 // then *this* store is dead (X = load P; store X -> P).
686 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
687 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
689 return EraseInstFromFunction(SI);
691 // Otherwise, this is a load from some other location. Stores before it
696 // Don't skip over loads or things that can modify memory.
697 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
701 // store X, null -> turns into 'unreachable' in SimplifyCFG
702 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
703 if (!isa<UndefValue>(Val)) {
704 SI.setOperand(0, UndefValue::get(Val->getType()));
705 if (Instruction *U = dyn_cast<Instruction>(Val))
706 Worklist.Add(U); // Dropped a use.
708 return nullptr; // Do not modify these!
711 // store undef, Ptr -> noop
712 if (isa<UndefValue>(Val))
713 return EraseInstFromFunction(SI);
715 // If the pointer destination is a cast, see if we can fold the cast into the
717 if (isa<CastInst>(Ptr))
718 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
720 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
722 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
726 // If this store is the last instruction in the basic block (possibly
727 // excepting debug info instructions), and if the block ends with an
728 // unconditional branch, try to move it to the successor block.
732 } while (isa<DbgInfoIntrinsic>(BBI) ||
733 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy()));
734 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
735 if (BI->isUnconditional())
736 if (SimplifyStoreAtEndOfBlock(SI))
737 return nullptr; // xform done!
742 /// SimplifyStoreAtEndOfBlock - Turn things like:
743 /// if () { *P = v1; } else { *P = v2 }
744 /// into a phi node with a store in the successor.
746 /// Simplify things like:
747 /// *P = v1; if () { *P = v2; }
748 /// into a phi node with a store in the successor.
750 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
751 BasicBlock *StoreBB = SI.getParent();
753 // Check to see if the successor block has exactly two incoming edges. If
754 // so, see if the other predecessor contains a store to the same location.
755 // if so, insert a PHI node (if needed) and move the stores down.
756 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
758 // Determine whether Dest has exactly two predecessors and, if so, compute
759 // the other predecessor.
760 pred_iterator PI = pred_begin(DestBB);
762 BasicBlock *OtherBB = nullptr;
767 if (++PI == pred_end(DestBB))
776 if (++PI != pred_end(DestBB))
779 // Bail out if all the relevant blocks aren't distinct (this can happen,
780 // for example, if SI is in an infinite loop)
781 if (StoreBB == DestBB || OtherBB == DestBB)
784 // Verify that the other block ends in a branch and is not otherwise empty.
785 BasicBlock::iterator BBI = OtherBB->getTerminator();
786 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
787 if (!OtherBr || BBI == OtherBB->begin())
790 // If the other block ends in an unconditional branch, check for the 'if then
791 // else' case. there is an instruction before the branch.
792 StoreInst *OtherStore = nullptr;
793 if (OtherBr->isUnconditional()) {
795 // Skip over debugging info.
796 while (isa<DbgInfoIntrinsic>(BBI) ||
797 (isa<BitCastInst>(BBI) && BBI->getType()->isPointerTy())) {
798 if (BBI==OtherBB->begin())
802 // If this isn't a store, isn't a store to the same location, or is not the
803 // right kind of store, bail out.
804 OtherStore = dyn_cast<StoreInst>(BBI);
805 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
806 !SI.isSameOperationAs(OtherStore))
809 // Otherwise, the other block ended with a conditional branch. If one of the
810 // destinations is StoreBB, then we have the if/then case.
811 if (OtherBr->getSuccessor(0) != StoreBB &&
812 OtherBr->getSuccessor(1) != StoreBB)
815 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
816 // if/then triangle. See if there is a store to the same ptr as SI that
819 // Check to see if we find the matching store.
820 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
821 if (OtherStore->getOperand(1) != SI.getOperand(1) ||
822 !SI.isSameOperationAs(OtherStore))
826 // If we find something that may be using or overwriting the stored
827 // value, or if we run out of instructions, we can't do the xform.
828 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
829 BBI == OtherBB->begin())
833 // In order to eliminate the store in OtherBr, we have to
834 // make sure nothing reads or overwrites the stored value in
836 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
837 // FIXME: This should really be AA driven.
838 if (I->mayReadFromMemory() || I->mayWriteToMemory())
843 // Insert a PHI node now if we need it.
844 Value *MergedVal = OtherStore->getOperand(0);
845 if (MergedVal != SI.getOperand(0)) {
846 PHINode *PN = PHINode::Create(MergedVal->getType(), 2, "storemerge");
847 PN->addIncoming(SI.getOperand(0), SI.getParent());
848 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
849 MergedVal = InsertNewInstBefore(PN, DestBB->front());
852 // Advance to a place where it is safe to insert the new store and
854 BBI = DestBB->getFirstInsertionPt();
855 StoreInst *NewSI = new StoreInst(MergedVal, SI.getOperand(1),
860 InsertNewInstBefore(NewSI, *BBI);
861 NewSI->setDebugLoc(OtherStore->getDebugLoc());
863 // If the two stores had AA tags, merge them.
865 SI.getAAMetadata(AATags);
867 OtherStore->getAAMetadata(AATags, /* Merge = */ true);
868 NewSI->setAAMetadata(AATags);
871 // Nuke the old stores.
872 EraseInstFromFunction(SI);
873 EraseInstFromFunction(*OtherStore);