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/IntrinsicInst.h"
16 #include "llvm/Target/TargetData.h"
17 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
18 #include "llvm/Transforms/Utils/Local.h"
19 #include "llvm/ADT/Statistic.h"
22 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
24 Instruction *InstCombiner::visitAllocaInst(AllocaInst &AI) {
25 // Convert: alloca Ty, C - where C is a constant != 1 into: alloca [C x Ty], 1
26 if (AI.isArrayAllocation()) { // Check C != 1
27 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
29 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
30 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
31 AllocaInst *New = Builder->CreateAlloca(NewTy, 0, AI.getName());
32 New->setAlignment(AI.getAlignment());
34 // Scan to the end of the allocation instructions, to skip over a block of
35 // allocas if possible...also skip interleaved debug info
37 BasicBlock::iterator It = New;
38 while (isa<AllocaInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
40 // Now that I is pointing to the first non-allocation-inst in the block,
41 // insert our getelementptr instruction...
43 Value *NullIdx =Constant::getNullValue(Type::getInt32Ty(AI.getContext()));
47 Value *V = GetElementPtrInst::CreateInBounds(New, Idx, Idx + 2,
48 New->getName()+".sub", It);
50 // Now make everything use the getelementptr instead of the original
52 return ReplaceInstUsesWith(AI, V);
53 } else if (isa<UndefValue>(AI.getArraySize())) {
54 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
58 if (TD && isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
59 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
60 // Note that we only do this for alloca's, because malloc should allocate
61 // and return a unique pointer, even for a zero byte allocation.
62 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
63 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
65 // If the alignment is 0 (unspecified), assign it the preferred alignment.
66 if (AI.getAlignment() == 0)
67 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
74 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
75 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
76 const TargetData *TD) {
77 User *CI = cast<User>(LI.getOperand(0));
78 Value *CastOp = CI->getOperand(0);
80 const PointerType *DestTy = cast<PointerType>(CI->getType());
81 const Type *DestPTy = DestTy->getElementType();
82 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
84 // If the address spaces don't match, don't eliminate the cast.
85 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
88 const Type *SrcPTy = SrcTy->getElementType();
90 if (DestPTy->isIntegerTy() || isa<PointerType>(DestPTy) ||
91 isa<VectorType>(DestPTy)) {
92 // If the source is an array, the code below will not succeed. Check to
93 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
95 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
96 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
97 if (ASrcTy->getNumElements() != 0) {
99 Idxs[0] = Constant::getNullValue(Type::getInt32Ty(LI.getContext()));
101 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
102 SrcTy = cast<PointerType>(CastOp->getType());
103 SrcPTy = SrcTy->getElementType();
106 if (IC.getTargetData() &&
107 (SrcPTy->isIntegerTy() || isa<PointerType>(SrcPTy) ||
108 isa<VectorType>(SrcPTy)) &&
109 // Do not allow turning this into a load of an integer, which is then
110 // casted to a pointer, this pessimizes pointer analysis a lot.
111 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
112 IC.getTargetData()->getTypeSizeInBits(SrcPTy) ==
113 IC.getTargetData()->getTypeSizeInBits(DestPTy)) {
115 // Okay, we are casting from one integer or pointer type to another of
116 // the same size. Instead of casting the pointer before the load, cast
117 // the result of the loaded value.
119 IC.Builder->CreateLoad(CastOp, LI.isVolatile(), CI->getName());
120 NewLoad->setAlignment(LI.getAlignment());
121 // Now cast the result of the load.
122 return new BitCastInst(NewLoad, LI.getType());
129 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
130 Value *Op = LI.getOperand(0);
132 // Attempt to improve the alignment.
134 unsigned KnownAlign =
135 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
137 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
139 LI.setAlignment(KnownAlign);
142 // load (cast X) --> cast (load X) iff safe.
143 if (isa<CastInst>(Op))
144 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
147 // None of the following transforms are legal for volatile loads.
148 if (LI.isVolatile()) return 0;
150 // Do really simple store-to-load forwarding and load CSE, to catch cases
151 // where there are several consequtive memory accesses to the same location,
152 // separated by a few arithmetic operations.
153 BasicBlock::iterator BBI = &LI;
154 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
155 return ReplaceInstUsesWith(LI, AvailableVal);
157 // load(gep null, ...) -> unreachable
158 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
159 const Value *GEPI0 = GEPI->getOperand(0);
160 // TODO: Consider a target hook for valid address spaces for this xform.
161 if (isa<ConstantPointerNull>(GEPI0) && GEPI->getPointerAddressSpace() == 0){
162 // Insert a new store to null instruction before the load to indicate
163 // that this code is not reachable. We do this instead of inserting
164 // an unreachable instruction directly because we cannot modify the
166 new StoreInst(UndefValue::get(LI.getType()),
167 Constant::getNullValue(Op->getType()), &LI);
168 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
172 // load null/undef -> unreachable
173 // TODO: Consider a target hook for valid address spaces for this xform.
174 if (isa<UndefValue>(Op) ||
175 (isa<ConstantPointerNull>(Op) && LI.getPointerAddressSpace() == 0)) {
176 // Insert a new store to null instruction before the load to indicate that
177 // this code is not reachable. We do this instead of inserting an
178 // unreachable instruction directly because we cannot modify the CFG.
179 new StoreInst(UndefValue::get(LI.getType()),
180 Constant::getNullValue(Op->getType()), &LI);
181 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
184 // Instcombine load (constantexpr_cast global) -> cast (load global)
185 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
187 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
190 if (Op->hasOneUse()) {
191 // Change select and PHI nodes to select values instead of addresses: this
192 // helps alias analysis out a lot, allows many others simplifications, and
193 // exposes redundancy in the code.
195 // Note that we cannot do the transformation unless we know that the
196 // introduced loads cannot trap! Something like this is valid as long as
197 // the condition is always false: load (select bool %C, int* null, int* %G),
198 // but it would not be valid if we transformed it to load from null
201 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
202 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
203 unsigned Align = LI.getAlignment();
204 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI, Align, TD) &&
205 isSafeToLoadUnconditionally(SI->getOperand(2), SI, Align, TD)) {
206 LoadInst *V1 = Builder->CreateLoad(SI->getOperand(1),
207 SI->getOperand(1)->getName()+".val");
208 LoadInst *V2 = Builder->CreateLoad(SI->getOperand(2),
209 SI->getOperand(2)->getName()+".val");
210 V1->setAlignment(Align);
211 V2->setAlignment(Align);
212 return SelectInst::Create(SI->getCondition(), V1, V2);
215 // load (select (cond, null, P)) -> load P
216 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
217 if (C->isNullValue()) {
218 LI.setOperand(0, SI->getOperand(2));
222 // load (select (cond, P, null)) -> load P
223 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
224 if (C->isNullValue()) {
225 LI.setOperand(0, SI->getOperand(1));
233 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
234 /// when possible. This makes it generally easy to do alias analysis and/or
235 /// SROA/mem2reg of the memory object.
236 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
237 User *CI = cast<User>(SI.getOperand(1));
238 Value *CastOp = CI->getOperand(0);
240 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
241 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
242 if (SrcTy == 0) return 0;
244 const Type *SrcPTy = SrcTy->getElementType();
246 if (!DestPTy->isIntegerTy() && !isa<PointerType>(DestPTy))
249 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
250 /// to its first element. This allows us to handle things like:
251 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
253 SmallVector<Value*, 4> NewGEPIndices;
255 // If the source is an array, the code below will not succeed. Check to
256 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
258 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
259 // Index through pointer.
260 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(SI.getContext()));
261 NewGEPIndices.push_back(Zero);
264 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
265 if (!STy->getNumElements()) /* Struct can be empty {} */
267 NewGEPIndices.push_back(Zero);
268 SrcPTy = STy->getElementType(0);
269 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
270 NewGEPIndices.push_back(Zero);
271 SrcPTy = ATy->getElementType();
277 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
280 if (!SrcPTy->isIntegerTy() && !isa<PointerType>(SrcPTy))
283 // If the pointers point into different address spaces or if they point to
284 // values with different sizes, we can't do the transformation.
285 if (!IC.getTargetData() ||
286 SrcTy->getAddressSpace() !=
287 cast<PointerType>(CI->getType())->getAddressSpace() ||
288 IC.getTargetData()->getTypeSizeInBits(SrcPTy) !=
289 IC.getTargetData()->getTypeSizeInBits(DestPTy))
292 // Okay, we are casting from one integer or pointer type to another of
293 // the same size. Instead of casting the pointer before
294 // the store, cast the value to be stored.
296 Value *SIOp0 = SI.getOperand(0);
297 Instruction::CastOps opcode = Instruction::BitCast;
298 const Type* CastSrcTy = SIOp0->getType();
299 const Type* CastDstTy = SrcPTy;
300 if (isa<PointerType>(CastDstTy)) {
301 if (CastSrcTy->isIntegerTy())
302 opcode = Instruction::IntToPtr;
303 } else if (isa<IntegerType>(CastDstTy)) {
304 if (isa<PointerType>(SIOp0->getType()))
305 opcode = Instruction::PtrToInt;
308 // SIOp0 is a pointer to aggregate and this is a store to the first field,
309 // emit a GEP to index into its first field.
310 if (!NewGEPIndices.empty())
311 CastOp = IC.Builder->CreateInBoundsGEP(CastOp, NewGEPIndices.begin(),
312 NewGEPIndices.end());
314 NewCast = IC.Builder->CreateCast(opcode, SIOp0, CastDstTy,
315 SIOp0->getName()+".c");
316 return new StoreInst(NewCast, CastOp);
319 /// equivalentAddressValues - Test if A and B will obviously have the same
320 /// value. This includes recognizing that %t0 and %t1 will have the same
321 /// value in code like this:
322 /// %t0 = getelementptr \@a, 0, 3
323 /// store i32 0, i32* %t0
324 /// %t1 = getelementptr \@a, 0, 3
325 /// %t2 = load i32* %t1
327 static bool equivalentAddressValues(Value *A, Value *B) {
328 // Test if the values are trivially equivalent.
329 if (A == B) return true;
331 // Test if the values come form identical arithmetic instructions.
332 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
333 // its only used to compare two uses within the same basic block, which
334 // means that they'll always either have the same value or one of them
335 // will have an undefined value.
336 if (isa<BinaryOperator>(A) ||
339 isa<GetElementPtrInst>(A))
340 if (Instruction *BI = dyn_cast<Instruction>(B))
341 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
344 // Otherwise they may not be equivalent.
348 // If this instruction has two uses, one of which is a llvm.dbg.declare,
349 // return the llvm.dbg.declare.
350 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
353 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
355 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
357 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
358 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
365 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
366 Value *Val = SI.getOperand(0);
367 Value *Ptr = SI.getOperand(1);
369 // If the RHS is an alloca with a single use, zapify the store, making the
371 // If the RHS is an alloca with a two uses, the other one being a
372 // llvm.dbg.declare, zapify the store and the declare, making the
373 // alloca dead. We must do this to prevent declares from affecting
375 if (!SI.isVolatile()) {
376 if (Ptr->hasOneUse()) {
377 if (isa<AllocaInst>(Ptr))
378 return EraseInstFromFunction(SI);
379 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
380 if (isa<AllocaInst>(GEP->getOperand(0))) {
381 if (GEP->getOperand(0)->hasOneUse())
382 return EraseInstFromFunction(SI);
383 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
384 EraseInstFromFunction(*DI);
385 return EraseInstFromFunction(SI);
390 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
391 EraseInstFromFunction(*DI);
392 return EraseInstFromFunction(SI);
396 // Attempt to improve the alignment.
398 unsigned KnownAlign =
399 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
401 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
403 SI.setAlignment(KnownAlign);
406 // Do really simple DSE, to catch cases where there are several consecutive
407 // stores to the same location, separated by a few arithmetic operations. This
408 // situation often occurs with bitfield accesses.
409 BasicBlock::iterator BBI = &SI;
410 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
413 // Don't count debug info directives, lest they affect codegen,
414 // and we skip pointer-to-pointer bitcasts, which are NOPs.
415 if (isa<DbgInfoIntrinsic>(BBI) ||
416 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
421 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
422 // Prev store isn't volatile, and stores to the same location?
423 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
427 EraseInstFromFunction(*PrevSI);
433 // If this is a load, we have to stop. However, if the loaded value is from
434 // the pointer we're loading and is producing the pointer we're storing,
435 // then *this* store is dead (X = load P; store X -> P).
436 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
437 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
439 return EraseInstFromFunction(SI);
441 // Otherwise, this is a load from some other location. Stores before it
446 // Don't skip over loads or things that can modify memory.
447 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
452 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
454 // store X, null -> turns into 'unreachable' in SimplifyCFG
455 if (isa<ConstantPointerNull>(Ptr) && SI.getPointerAddressSpace() == 0) {
456 if (!isa<UndefValue>(Val)) {
457 SI.setOperand(0, UndefValue::get(Val->getType()));
458 if (Instruction *U = dyn_cast<Instruction>(Val))
459 Worklist.Add(U); // Dropped a use.
461 return 0; // Do not modify these!
464 // store undef, Ptr -> noop
465 if (isa<UndefValue>(Val))
466 return EraseInstFromFunction(SI);
468 // If the pointer destination is a cast, see if we can fold the cast into the
470 if (isa<CastInst>(Ptr))
471 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
473 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
475 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
479 // If this store is the last instruction in the basic block (possibly
480 // excepting debug info instructions), and if the block ends with an
481 // unconditional branch, try to move it to the successor block.
485 } while (isa<DbgInfoIntrinsic>(BBI) ||
486 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
487 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
488 if (BI->isUnconditional())
489 if (SimplifyStoreAtEndOfBlock(SI))
490 return 0; // xform done!
495 /// SimplifyStoreAtEndOfBlock - Turn things like:
496 /// if () { *P = v1; } else { *P = v2 }
497 /// into a phi node with a store in the successor.
499 /// Simplify things like:
500 /// *P = v1; if () { *P = v2; }
501 /// into a phi node with a store in the successor.
503 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
504 BasicBlock *StoreBB = SI.getParent();
506 // Check to see if the successor block has exactly two incoming edges. If
507 // so, see if the other predecessor contains a store to the same location.
508 // if so, insert a PHI node (if needed) and move the stores down.
509 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
511 // Determine whether Dest has exactly two predecessors and, if so, compute
512 // the other predecessor.
513 pred_iterator PI = pred_begin(DestBB);
514 BasicBlock *OtherBB = 0;
518 if (PI == pred_end(DestBB))
521 if (*PI != StoreBB) {
526 if (++PI != pred_end(DestBB))
529 // Bail out if all the relevant blocks aren't distinct (this can happen,
530 // for example, if SI is in an infinite loop)
531 if (StoreBB == DestBB || OtherBB == DestBB)
534 // Verify that the other block ends in a branch and is not otherwise empty.
535 BasicBlock::iterator BBI = OtherBB->getTerminator();
536 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
537 if (!OtherBr || BBI == OtherBB->begin())
540 // If the other block ends in an unconditional branch, check for the 'if then
541 // else' case. there is an instruction before the branch.
542 StoreInst *OtherStore = 0;
543 if (OtherBr->isUnconditional()) {
545 // Skip over debugging info.
546 while (isa<DbgInfoIntrinsic>(BBI) ||
547 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
548 if (BBI==OtherBB->begin())
552 // If this isn't a store, isn't a store to the same location, or if the
553 // alignments differ, bail out.
554 OtherStore = dyn_cast<StoreInst>(BBI);
555 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1) ||
556 OtherStore->getAlignment() != SI.getAlignment())
559 // Otherwise, the other block ended with a conditional branch. If one of the
560 // destinations is StoreBB, then we have the if/then case.
561 if (OtherBr->getSuccessor(0) != StoreBB &&
562 OtherBr->getSuccessor(1) != StoreBB)
565 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
566 // if/then triangle. See if there is a store to the same ptr as SI that
569 // Check to see if we find the matching store.
570 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
571 if (OtherStore->getOperand(1) != SI.getOperand(1) ||
572 OtherStore->getAlignment() != SI.getAlignment())
576 // If we find something that may be using or overwriting the stored
577 // value, or if we run out of instructions, we can't do the xform.
578 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
579 BBI == OtherBB->begin())
583 // In order to eliminate the store in OtherBr, we have to
584 // make sure nothing reads or overwrites the stored value in
586 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
587 // FIXME: This should really be AA driven.
588 if (I->mayReadFromMemory() || I->mayWriteToMemory())
593 // Insert a PHI node now if we need it.
594 Value *MergedVal = OtherStore->getOperand(0);
595 if (MergedVal != SI.getOperand(0)) {
596 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
597 PN->reserveOperandSpace(2);
598 PN->addIncoming(SI.getOperand(0), SI.getParent());
599 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
600 MergedVal = InsertNewInstBefore(PN, DestBB->front());
603 // Advance to a place where it is safe to insert the new store and
605 BBI = DestBB->getFirstNonPHI();
606 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
607 OtherStore->isVolatile(),
608 SI.getAlignment()), *BBI);
610 // Nuke the old stores.
611 EraseInstFromFunction(SI);
612 EraseInstFromFunction(*OtherStore);