1 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===//
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 transformation implements the well known scalar replacement of
11 // aggregates transformation. This xform breaks up alloca instructions of
12 // aggregate type (structure or array) into individual alloca instructions for
13 // each member (if possible). Then, if possible, it transforms the individual
14 // alloca instructions into nice clean scalar SSA form.
16 // This combines a simple SRoA algorithm with the Mem2Reg algorithm because
17 // often interact, especially for C++ programs. As such, iterating between
18 // SRoA, then Mem2Reg until we run out of things to promote works well.
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "scalarrepl"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/DerivedTypes.h"
26 #include "llvm/Function.h"
27 #include "llvm/GlobalVariable.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/IntrinsicInst.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/Module.h"
32 #include "llvm/Pass.h"
33 #include "llvm/Analysis/Dominators.h"
34 #include "llvm/Analysis/Loads.h"
35 #include "llvm/Analysis/ValueTracking.h"
36 #include "llvm/Target/TargetData.h"
37 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
38 #include "llvm/Transforms/Utils/Local.h"
39 #include "llvm/Transforms/Utils/SSAUpdater.h"
40 #include "llvm/Support/CallSite.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/ErrorHandling.h"
43 #include "llvm/Support/GetElementPtrTypeIterator.h"
44 #include "llvm/Support/IRBuilder.h"
45 #include "llvm/Support/MathExtras.h"
46 #include "llvm/Support/raw_ostream.h"
47 #include "llvm/ADT/SetVector.h"
48 #include "llvm/ADT/SmallVector.h"
49 #include "llvm/ADT/Statistic.h"
52 STATISTIC(NumReplaced, "Number of allocas broken up");
53 STATISTIC(NumPromoted, "Number of allocas promoted");
54 STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion");
55 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
56 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
59 struct SROA : public FunctionPass {
60 SROA(int T, bool hasDT, char &ID)
61 : FunctionPass(ID), HasDomTree(hasDT) {
68 bool runOnFunction(Function &F);
70 bool performScalarRepl(Function &F);
71 bool performPromotion(Function &F);
77 /// DeadInsts - Keep track of instructions we have made dead, so that
78 /// we can remove them after we are done working.
79 SmallVector<Value*, 32> DeadInsts;
81 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
82 /// information about the uses. All these fields are initialized to false
83 /// and set to true when something is learned.
85 /// The alloca to promote.
88 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite
89 /// looping and avoid redundant work.
90 SmallPtrSet<PHINode*, 8> CheckedPHIs;
92 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
95 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
98 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
101 /// hasSubelementAccess - This is true if a subelement of the alloca is
102 /// ever accessed, or false if the alloca is only accessed with mem
103 /// intrinsics or load/store that only access the entire alloca at once.
104 bool hasSubelementAccess : 1;
106 /// hasALoadOrStore - This is true if there are any loads or stores to it.
107 /// The alloca may just be accessed with memcpy, for example, which would
109 bool hasALoadOrStore : 1;
111 explicit AllocaInfo(AllocaInst *ai)
112 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false),
113 hasSubelementAccess(false), hasALoadOrStore(false) {}
116 unsigned SRThreshold;
118 void MarkUnsafe(AllocaInfo &I, Instruction *User) {
120 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n');
123 bool isSafeAllocaToScalarRepl(AllocaInst *AI);
125 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info);
126 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset,
128 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info);
129 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
130 const Type *MemOpType, bool isStore, AllocaInfo &Info,
131 Instruction *TheAccess, bool AllowWholeAccess);
132 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size);
133 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset,
136 void DoScalarReplacement(AllocaInst *AI,
137 std::vector<AllocaInst*> &WorkList);
138 void DeleteDeadInstructions();
140 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
141 SmallVector<AllocaInst*, 32> &NewElts);
142 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
143 SmallVector<AllocaInst*, 32> &NewElts);
144 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
145 SmallVector<AllocaInst*, 32> &NewElts);
146 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
148 SmallVector<AllocaInst*, 32> &NewElts);
149 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
150 SmallVector<AllocaInst*, 32> &NewElts);
151 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
152 SmallVector<AllocaInst*, 32> &NewElts);
154 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI);
157 // SROA_DT - SROA that uses DominatorTree.
158 struct SROA_DT : public SROA {
161 SROA_DT(int T = -1) : SROA(T, true, ID) {
162 initializeSROA_DTPass(*PassRegistry::getPassRegistry());
165 // getAnalysisUsage - This pass does not require any passes, but we know it
166 // will not alter the CFG, so say so.
167 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
168 AU.addRequired<DominatorTree>();
169 AU.setPreservesCFG();
173 // SROA_SSAUp - SROA that uses SSAUpdater.
174 struct SROA_SSAUp : public SROA {
177 SROA_SSAUp(int T = -1) : SROA(T, false, ID) {
178 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry());
181 // getAnalysisUsage - This pass does not require any passes, but we know it
182 // will not alter the CFG, so say so.
183 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
184 AU.setPreservesCFG();
190 char SROA_DT::ID = 0;
191 char SROA_SSAUp::ID = 0;
193 INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl",
194 "Scalar Replacement of Aggregates (DT)", false, false)
195 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
196 INITIALIZE_PASS_END(SROA_DT, "scalarrepl",
197 "Scalar Replacement of Aggregates (DT)", false, false)
199 INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa",
200 "Scalar Replacement of Aggregates (SSAUp)", false, false)
201 INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa",
202 "Scalar Replacement of Aggregates (SSAUp)", false, false)
204 // Public interface to the ScalarReplAggregates pass
205 FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold,
208 return new SROA_DT(Threshold);
209 return new SROA_SSAUp(Threshold);
213 //===----------------------------------------------------------------------===//
214 // Convert To Scalar Optimization.
215 //===----------------------------------------------------------------------===//
218 /// ConvertToScalarInfo - This class implements the "Convert To Scalar"
219 /// optimization, which scans the uses of an alloca and determines if it can
220 /// rewrite it in terms of a single new alloca that can be mem2reg'd.
221 class ConvertToScalarInfo {
222 /// AllocaSize - The size of the alloca being considered.
224 const TargetData &TD;
226 /// IsNotTrivial - This is set to true if there is some access to the object
227 /// which means that mem2reg can't promote it.
230 /// VectorTy - This tracks the type that we should promote the vector to if
231 /// it is possible to turn it into a vector. This starts out null, and if it
232 /// isn't possible to turn into a vector type, it gets set to VoidTy.
233 const Type *VectorTy;
235 /// HadAVector - True if there is at least one vector access to the alloca.
236 /// We don't want to turn random arrays into vectors and use vector element
237 /// insert/extract, but if there are element accesses to something that is
238 /// also declared as a vector, we do want to promote to a vector.
242 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td)
243 : AllocaSize(Size), TD(td) {
244 IsNotTrivial = false;
249 AllocaInst *TryConvert(AllocaInst *AI);
252 bool CanConvertToScalar(Value *V, uint64_t Offset);
253 void MergeInType(const Type *In, uint64_t Offset);
254 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
256 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
257 uint64_t Offset, IRBuilder<> &Builder);
258 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
259 uint64_t Offset, IRBuilder<> &Builder);
261 } // end anonymous namespace.
264 /// TryConvert - Analyze the specified alloca, and if it is safe to do so,
265 /// rewrite it to be a new alloca which is mem2reg'able. This returns the new
266 /// alloca if possible or null if not.
267 AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
268 // If we can't convert this scalar, or if mem2reg can trivially do it, bail
270 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial)
273 // If we were able to find a vector type that can handle this with
274 // insert/extract elements, and if there was at least one use that had
275 // a vector type, promote this to a vector. We don't want to promote
276 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
277 // we just get a lot of insert/extracts. If at least one vector is
278 // involved, then we probably really do have a union of vector/array.
280 if (VectorTy && VectorTy->isVectorTy() && HadAVector) {
281 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = "
282 << *VectorTy << '\n');
283 NewTy = VectorTy; // Use the vector type.
285 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
286 // Create and insert the integer alloca.
287 NewTy = IntegerType::get(AI->getContext(), AllocaSize*8);
289 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
290 ConvertUsesToScalar(AI, NewAI, 0);
294 /// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy)
295 /// so far at the offset specified by Offset (which is specified in bytes).
297 /// There are two cases we handle here:
298 /// 1) A union of vector types of the same size and potentially its elements.
299 /// Here we turn element accesses into insert/extract element operations.
300 /// This promotes a <4 x float> with a store of float to the third element
301 /// into a <4 x float> that uses insert element.
302 /// 2) A fully general blob of memory, which we turn into some (potentially
303 /// large) integer type with extract and insert operations where the loads
304 /// and stores would mutate the memory. We mark this by setting VectorTy
306 void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) {
307 // If we already decided to turn this into a blob of integer memory, there is
308 // nothing to be done.
309 if (VectorTy && VectorTy->isVoidTy())
312 // If this could be contributing to a vector, analyze it.
314 // If the In type is a vector that is the same size as the alloca, see if it
315 // matches the existing VecTy.
316 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
317 // Remember if we saw a vector type.
320 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
321 // If we're storing/loading a vector of the right size, allow it as a
322 // vector. If this the first vector we see, remember the type so that
323 // we know the element size. If this is a subsequent access, ignore it
324 // even if it is a differing type but the same size. Worst case we can
325 // bitcast the resultant vectors.
330 } else if (In->isFloatTy() || In->isDoubleTy() ||
331 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 &&
332 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
333 // If we're accessing something that could be an element of a vector, see
334 // if the implied vector agrees with what we already have and if Offset is
335 // compatible with it.
336 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
337 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 &&
339 cast<VectorType>(VectorTy)->getElementType()
340 ->getPrimitiveSizeInBits()/8 == EltSize)) {
342 VectorTy = VectorType::get(In, AllocaSize/EltSize);
347 // Otherwise, we have a case that we can't handle with an optimized vector
348 // form. We can still turn this into a large integer.
349 VectorTy = Type::getVoidTy(In->getContext());
352 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
353 /// its accesses to a single vector type, return true and set VecTy to
354 /// the new type. If we could convert the alloca into a single promotable
355 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
356 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
357 /// is the current offset from the base of the alloca being analyzed.
359 /// If we see at least one access to the value that is as a vector type, set the
361 bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) {
362 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
363 Instruction *User = cast<Instruction>(*UI);
365 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
366 // Don't break volatile loads.
367 if (LI->isVolatile())
369 // Don't touch MMX operations.
370 if (LI->getType()->isX86_MMXTy())
372 MergeInType(LI->getType(), Offset);
376 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
377 // Storing the pointer, not into the value?
378 if (SI->getOperand(0) == V || SI->isVolatile()) return false;
379 // Don't touch MMX operations.
380 if (SI->getOperand(0)->getType()->isX86_MMXTy())
382 MergeInType(SI->getOperand(0)->getType(), Offset);
386 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
387 IsNotTrivial = true; // Can't be mem2reg'd.
388 if (!CanConvertToScalar(BCI, Offset))
393 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
394 // If this is a GEP with a variable indices, we can't handle it.
395 if (!GEP->hasAllConstantIndices())
398 // Compute the offset that this GEP adds to the pointer.
399 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
400 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
401 &Indices[0], Indices.size());
402 // See if all uses can be converted.
403 if (!CanConvertToScalar(GEP, Offset+GEPOffset))
405 IsNotTrivial = true; // Can't be mem2reg'd.
409 // If this is a constant sized memset of a constant value (e.g. 0) we can
411 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
412 // Store of constant value and constant size.
413 if (!isa<ConstantInt>(MSI->getValue()) ||
414 !isa<ConstantInt>(MSI->getLength()))
416 IsNotTrivial = true; // Can't be mem2reg'd.
420 // If this is a memcpy or memmove into or out of the whole allocation, we
421 // can handle it like a load or store of the scalar type.
422 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
423 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength());
424 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0)
427 IsNotTrivial = true; // Can't be mem2reg'd.
431 // Otherwise, we cannot handle this!
438 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
439 /// directly. This happens when we are converting an "integer union" to a
440 /// single integer scalar, or when we are converting a "vector union" to a
441 /// vector with insert/extractelement instructions.
443 /// Offset is an offset from the original alloca, in bits that need to be
444 /// shifted to the right. By the end of this, there should be no uses of Ptr.
445 void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI,
447 while (!Ptr->use_empty()) {
448 Instruction *User = cast<Instruction>(Ptr->use_back());
450 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
451 ConvertUsesToScalar(CI, NewAI, Offset);
452 CI->eraseFromParent();
456 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
457 // Compute the offset that this GEP adds to the pointer.
458 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
459 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
460 &Indices[0], Indices.size());
461 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
462 GEP->eraseFromParent();
466 IRBuilder<> Builder(User);
468 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
469 // The load is a bit extract from NewAI shifted right by Offset bits.
470 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
472 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
473 LI->replaceAllUsesWith(NewLoadVal);
474 LI->eraseFromParent();
478 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
479 assert(SI->getOperand(0) != Ptr && "Consistency error!");
480 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
481 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
483 Builder.CreateStore(New, NewAI);
484 SI->eraseFromParent();
486 // If the load we just inserted is now dead, then the inserted store
487 // overwrote the entire thing.
488 if (Old->use_empty())
489 Old->eraseFromParent();
493 // If this is a constant sized memset of a constant value (e.g. 0) we can
494 // transform it into a store of the expanded constant value.
495 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
496 assert(MSI->getRawDest() == Ptr && "Consistency error!");
497 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
499 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
501 // Compute the value replicated the right number of times.
502 APInt APVal(NumBytes*8, Val);
504 // Splat the value if non-zero.
506 for (unsigned i = 1; i != NumBytes; ++i)
509 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
510 Value *New = ConvertScalar_InsertValue(
511 ConstantInt::get(User->getContext(), APVal),
512 Old, Offset, Builder);
513 Builder.CreateStore(New, NewAI);
515 // If the load we just inserted is now dead, then the memset overwrote
517 if (Old->use_empty())
518 Old->eraseFromParent();
520 MSI->eraseFromParent();
524 // If this is a memcpy or memmove into or out of the whole allocation, we
525 // can handle it like a load or store of the scalar type.
526 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
527 assert(Offset == 0 && "must be store to start of alloca");
529 // If the source and destination are both to the same alloca, then this is
530 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
532 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, 0));
534 if (GetUnderlyingObject(MTI->getSource(), 0) != OrigAI) {
535 // Dest must be OrigAI, change this to be a load from the original
536 // pointer (bitcasted), then a store to our new alloca.
537 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
538 Value *SrcPtr = MTI->getSource();
539 const PointerType* SPTy = cast<PointerType>(SrcPtr->getType());
540 const PointerType* AIPTy = cast<PointerType>(NewAI->getType());
541 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
542 AIPTy = PointerType::get(AIPTy->getElementType(),
543 SPTy->getAddressSpace());
545 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy);
547 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
548 SrcVal->setAlignment(MTI->getAlignment());
549 Builder.CreateStore(SrcVal, NewAI);
550 } else if (GetUnderlyingObject(MTI->getDest(), 0) != OrigAI) {
551 // Src must be OrigAI, change this to be a load from NewAI then a store
552 // through the original dest pointer (bitcasted).
553 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
554 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
556 const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType());
557 const PointerType* AIPTy = cast<PointerType>(NewAI->getType());
558 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
559 AIPTy = PointerType::get(AIPTy->getElementType(),
560 DPTy->getAddressSpace());
562 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy);
564 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
565 NewStore->setAlignment(MTI->getAlignment());
567 // Noop transfer. Src == Dst
570 MTI->eraseFromParent();
574 llvm_unreachable("Unsupported operation!");
578 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
579 /// or vector value FromVal, extracting the bits from the offset specified by
580 /// Offset. This returns the value, which is of type ToType.
582 /// This happens when we are converting an "integer union" to a single
583 /// integer scalar, or when we are converting a "vector union" to a vector with
584 /// insert/extractelement instructions.
586 /// Offset is an offset from the original alloca, in bits that need to be
587 /// shifted to the right.
588 Value *ConvertToScalarInfo::
589 ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
590 uint64_t Offset, IRBuilder<> &Builder) {
591 // If the load is of the whole new alloca, no conversion is needed.
592 if (FromVal->getType() == ToType && Offset == 0)
595 // If the result alloca is a vector type, this is either an element
596 // access or a bitcast to another vector type of the same size.
597 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
598 if (ToType->isVectorTy())
599 return Builder.CreateBitCast(FromVal, ToType, "tmp");
601 // Otherwise it must be an element access.
604 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
605 Elt = Offset/EltSize;
606 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
608 // Return the element extracted out of it.
609 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get(
610 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp");
611 if (V->getType() != ToType)
612 V = Builder.CreateBitCast(V, ToType, "tmp");
616 // If ToType is a first class aggregate, extract out each of the pieces and
617 // use insertvalue's to form the FCA.
618 if (const StructType *ST = dyn_cast<StructType>(ToType)) {
619 const StructLayout &Layout = *TD.getStructLayout(ST);
620 Value *Res = UndefValue::get(ST);
621 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
622 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
623 Offset+Layout.getElementOffsetInBits(i),
625 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
630 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
631 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
632 Value *Res = UndefValue::get(AT);
633 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
634 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
635 Offset+i*EltSize, Builder);
636 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
641 // Otherwise, this must be a union that was converted to an integer value.
642 const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
644 // If this is a big-endian system and the load is narrower than the
645 // full alloca type, we need to do a shift to get the right bits.
647 if (TD.isBigEndian()) {
648 // On big-endian machines, the lowest bit is stored at the bit offset
649 // from the pointer given by getTypeStoreSizeInBits. This matters for
650 // integers with a bitwidth that is not a multiple of 8.
651 ShAmt = TD.getTypeStoreSizeInBits(NTy) -
652 TD.getTypeStoreSizeInBits(ToType) - Offset;
657 // Note: we support negative bitwidths (with shl) which are not defined.
658 // We do this to support (f.e.) loads off the end of a structure where
659 // only some bits are used.
660 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
661 FromVal = Builder.CreateLShr(FromVal,
662 ConstantInt::get(FromVal->getType(),
664 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
665 FromVal = Builder.CreateShl(FromVal,
666 ConstantInt::get(FromVal->getType(),
669 // Finally, unconditionally truncate the integer to the right width.
670 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType);
671 if (LIBitWidth < NTy->getBitWidth())
673 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
675 else if (LIBitWidth > NTy->getBitWidth())
677 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
680 // If the result is an integer, this is a trunc or bitcast.
681 if (ToType->isIntegerTy()) {
683 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) {
684 // Just do a bitcast, we know the sizes match up.
685 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
687 // Otherwise must be a pointer.
688 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
690 assert(FromVal->getType() == ToType && "Didn't convert right?");
694 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
695 /// or vector value "Old" at the offset specified by Offset.
697 /// This happens when we are converting an "integer union" to a
698 /// single integer scalar, or when we are converting a "vector union" to a
699 /// vector with insert/extractelement instructions.
701 /// Offset is an offset from the original alloca, in bits that need to be
702 /// shifted to the right.
703 Value *ConvertToScalarInfo::
704 ConvertScalar_InsertValue(Value *SV, Value *Old,
705 uint64_t Offset, IRBuilder<> &Builder) {
706 // Convert the stored type to the actual type, shift it left to insert
707 // then 'or' into place.
708 const Type *AllocaType = Old->getType();
709 LLVMContext &Context = Old->getContext();
711 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
712 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy);
713 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType());
715 // Changing the whole vector with memset or with an access of a different
717 if (ValSize == VecSize)
718 return Builder.CreateBitCast(SV, AllocaType, "tmp");
720 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
722 // Must be an element insertion.
723 unsigned Elt = Offset/EltSize;
725 if (SV->getType() != VTy->getElementType())
726 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
728 SV = Builder.CreateInsertElement(Old, SV,
729 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt),
734 // If SV is a first-class aggregate value, insert each value recursively.
735 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
736 const StructLayout &Layout = *TD.getStructLayout(ST);
737 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
738 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
739 Old = ConvertScalar_InsertValue(Elt, Old,
740 Offset+Layout.getElementOffsetInBits(i),
746 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
747 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
748 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
749 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
750 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
755 // If SV is a float, convert it to the appropriate integer type.
756 // If it is a pointer, do the same.
757 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType());
758 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType);
759 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType());
760 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType);
761 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
762 SV = Builder.CreateBitCast(SV,
763 IntegerType::get(SV->getContext(),SrcWidth), "tmp");
764 else if (SV->getType()->isPointerTy())
765 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp");
767 // Zero extend or truncate the value if needed.
768 if (SV->getType() != AllocaType) {
769 if (SV->getType()->getPrimitiveSizeInBits() <
770 AllocaType->getPrimitiveSizeInBits())
771 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
773 // Truncation may be needed if storing more than the alloca can hold
774 // (undefined behavior).
775 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
776 SrcWidth = DestWidth;
777 SrcStoreWidth = DestStoreWidth;
781 // If this is a big-endian system and the store is narrower than the
782 // full alloca type, we need to do a shift to get the right bits.
784 if (TD.isBigEndian()) {
785 // On big-endian machines, the lowest bit is stored at the bit offset
786 // from the pointer given by getTypeStoreSizeInBits. This matters for
787 // integers with a bitwidth that is not a multiple of 8.
788 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
793 // Note: we support negative bitwidths (with shr) which are not defined.
794 // We do this to support (f.e.) stores off the end of a structure where
795 // only some bits in the structure are set.
796 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
797 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
798 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(),
801 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
802 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(),
804 Mask = Mask.lshr(-ShAmt);
807 // Mask out the bits we are about to insert from the old value, and or
809 if (SrcWidth != DestWidth) {
810 assert(DestWidth > SrcWidth);
811 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
812 SV = Builder.CreateOr(Old, SV, "ins");
818 //===----------------------------------------------------------------------===//
820 //===----------------------------------------------------------------------===//
823 bool SROA::runOnFunction(Function &F) {
824 TD = getAnalysisIfAvailable<TargetData>();
826 bool Changed = performPromotion(F);
828 // FIXME: ScalarRepl currently depends on TargetData more than it
829 // theoretically needs to. It should be refactored in order to support
830 // target-independent IR. Until this is done, just skip the actual
831 // scalar-replacement portion of this pass.
832 if (!TD) return Changed;
835 bool LocalChange = performScalarRepl(F);
836 if (!LocalChange) break; // No need to repromote if no scalarrepl
838 LocalChange = performPromotion(F);
839 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
846 class AllocaPromoter : public LoadAndStorePromoter {
849 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S)
850 : LoadAndStorePromoter(Insts, S), AI(0) {}
852 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) {
853 // Remember which alloca we're promoting (for isInstInList).
855 LoadAndStorePromoter::run(Insts);
856 AI->eraseFromParent();
859 virtual bool isInstInList(Instruction *I,
860 const SmallVectorImpl<Instruction*> &Insts) const {
861 if (LoadInst *LI = dyn_cast<LoadInst>(I))
862 return LI->getOperand(0) == AI;
863 return cast<StoreInst>(I)->getPointerOperand() == AI;
866 } // end anon namespace
868 /// isSafeSelectToSpeculate - Select instructions that use an alloca and are
869 /// subsequently loaded can be rewritten to load both input pointers and then
870 /// select between the result, allowing the load of the alloca to be promoted.
872 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
873 /// %V = load i32* %P2
875 /// %V1 = load i32* %Alloca -> will be mem2reg'd
876 /// %V2 = load i32* %Other
877 /// %V = select i1 %cond, i32 %V1, i32 %V2
879 /// We can do this to a select if its only uses are loads and if the operand to
880 /// the select can be loaded unconditionally.
881 static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) {
882 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer();
883 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer();
885 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end();
887 LoadInst *LI = dyn_cast<LoadInst>(*UI);
888 if (LI == 0 || LI->isVolatile()) return false;
890 // Both operands to the select need to be dereferencable, either absolutely
891 // (e.g. allocas) or at this point because we can see other accesses to it.
892 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI,
893 LI->getAlignment(), TD))
895 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI,
896 LI->getAlignment(), TD))
903 /// isSafePHIToSpeculate - PHI instructions that use an alloca and are
904 /// subsequently loaded can be rewritten to load both input pointers in the pred
905 /// blocks and then PHI the results, allowing the load of the alloca to be
908 /// %P2 = phi [i32* %Alloca, i32* %Other]
909 /// %V = load i32* %P2
911 /// %V1 = load i32* %Alloca -> will be mem2reg'd
913 /// %V2 = load i32* %Other
915 /// %V = phi [i32 %V1, i32 %V2]
917 /// We can do this to a select if its only uses are loads and if the operand to
918 /// the select can be loaded unconditionally.
919 static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) {
920 // For now, we can only do this promotion if the load is in the same block as
921 // the PHI, and if there are no stores between the phi and load.
922 // TODO: Allow recursive phi users.
923 // TODO: Allow stores.
924 BasicBlock *BB = PN->getParent();
925 unsigned MaxAlign = 0;
926 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end();
928 LoadInst *LI = dyn_cast<LoadInst>(*UI);
929 if (LI == 0 || LI->isVolatile()) return false;
931 // For now we only allow loads in the same block as the PHI. This is a
932 // common case that happens when instcombine merges two loads through a PHI.
933 if (LI->getParent() != BB) return false;
935 // Ensure that there are no instructions between the PHI and the load that
937 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI)
938 if (BBI->mayWriteToMemory())
941 MaxAlign = std::max(MaxAlign, LI->getAlignment());
944 // Okay, we know that we have one or more loads in the same block as the PHI.
945 // We can transform this if it is safe to push the loads into the predecessor
946 // blocks. The only thing to watch out for is that we can't put a possibly
947 // trapping load in the predecessor if it is a critical edge.
948 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
949 BasicBlock *Pred = PN->getIncomingBlock(i);
951 // If the predecessor has a single successor, then the edge isn't critical.
952 if (Pred->getTerminator()->getNumSuccessors() == 1)
955 Value *InVal = PN->getIncomingValue(i);
957 // If the InVal is an invoke in the pred, we can't put a load on the edge.
958 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
959 if (II->getParent() == Pred)
962 // If this pointer is always safe to load, or if we can prove that there is
963 // already a load in the block, then we can move the load to the pred block.
964 if (InVal->isDereferenceablePointer() ||
965 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD))
975 /// tryToMakeAllocaBePromotable - This returns true if the alloca only has
976 /// direct (non-volatile) loads and stores to it. If the alloca is close but
977 /// not quite there, this will transform the code to allow promotion. As such,
978 /// it is a non-pure predicate.
979 static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) {
980 SetVector<Instruction*, SmallVector<Instruction*, 4>,
981 SmallPtrSet<Instruction*, 4> > InstsToRewrite;
983 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
986 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
987 if (LI->isVolatile())
992 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
993 if (SI->getOperand(0) == AI || SI->isVolatile())
994 return false; // Don't allow a store OF the AI, only INTO the AI.
998 if (SelectInst *SI = dyn_cast<SelectInst>(U)) {
999 // If the condition being selected on is a constant, fold the select, yes
1000 // this does (rarely) happen early on.
1001 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) {
1002 Value *Result = SI->getOperand(1+CI->isZero());
1003 SI->replaceAllUsesWith(Result);
1004 SI->eraseFromParent();
1006 // This is very rare and we just scrambled the use list of AI, start
1008 return tryToMakeAllocaBePromotable(AI, TD);
1011 // If it is safe to turn "load (select c, AI, ptr)" into a select of two
1012 // loads, then we can transform this by rewriting the select.
1013 if (!isSafeSelectToSpeculate(SI, TD))
1016 InstsToRewrite.insert(SI);
1020 if (PHINode *PN = dyn_cast<PHINode>(U)) {
1021 if (PN->use_empty()) { // Dead PHIs can be stripped.
1022 InstsToRewrite.insert(PN);
1026 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads
1027 // in the pred blocks, then we can transform this by rewriting the PHI.
1028 if (!isSafePHIToSpeculate(PN, TD))
1031 InstsToRewrite.insert(PN);
1038 // If there are no instructions to rewrite, then all uses are load/stores and
1040 if (InstsToRewrite.empty())
1043 // If we have instructions that need to be rewritten for this to be promotable
1044 // take care of it now.
1045 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) {
1046 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) {
1047 // Selects in InstsToRewrite only have load uses. Rewrite each as two
1048 // loads with a new select.
1049 while (!SI->use_empty()) {
1050 LoadInst *LI = cast<LoadInst>(SI->use_back());
1052 IRBuilder<> Builder(LI);
1053 LoadInst *TrueLoad =
1054 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t");
1055 LoadInst *FalseLoad =
1056 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".t");
1058 // Transfer alignment and TBAA info if present.
1059 TrueLoad->setAlignment(LI->getAlignment());
1060 FalseLoad->setAlignment(LI->getAlignment());
1061 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1062 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1063 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1066 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad);
1068 LI->replaceAllUsesWith(V);
1069 LI->eraseFromParent();
1072 // Now that all the loads are gone, the select is gone too.
1073 SI->eraseFromParent();
1077 // Otherwise, we have a PHI node which allows us to push the loads into the
1079 PHINode *PN = cast<PHINode>(InstsToRewrite[i]);
1080 if (PN->use_empty()) {
1081 PN->eraseFromParent();
1085 const Type *LoadTy = cast<PointerType>(PN->getType())->getElementType();
1086 PHINode *NewPN = PHINode::Create(LoadTy, PN->getName()+".ld", PN);
1088 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1089 // matter which one we get and if any differ, it doesn't matter.
1090 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back());
1091 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1092 unsigned Align = SomeLoad->getAlignment();
1094 // Rewrite all loads of the PN to use the new PHI.
1095 while (!PN->use_empty()) {
1096 LoadInst *LI = cast<LoadInst>(PN->use_back());
1097 LI->replaceAllUsesWith(NewPN);
1098 LI->eraseFromParent();
1101 // Inject loads into all of the pred blocks. Keep track of which blocks we
1102 // insert them into in case we have multiple edges from the same block.
1103 DenseMap<BasicBlock*, LoadInst*> InsertedLoads;
1105 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1106 BasicBlock *Pred = PN->getIncomingBlock(i);
1107 LoadInst *&Load = InsertedLoads[Pred];
1109 Load = new LoadInst(PN->getIncomingValue(i),
1110 PN->getName() + "." + Pred->getName(),
1111 Pred->getTerminator());
1112 Load->setAlignment(Align);
1113 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1116 NewPN->addIncoming(Load, Pred);
1119 PN->eraseFromParent();
1127 bool SROA::performPromotion(Function &F) {
1128 std::vector<AllocaInst*> Allocas;
1129 DominatorTree *DT = 0;
1131 DT = &getAnalysis<DominatorTree>();
1133 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
1135 bool Changed = false;
1136 SmallVector<Instruction*, 64> Insts;
1140 // Find allocas that are safe to promote, by looking at all instructions in
1142 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
1143 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
1144 if (tryToMakeAllocaBePromotable(AI, TD))
1145 Allocas.push_back(AI);
1147 if (Allocas.empty()) break;
1150 PromoteMemToReg(Allocas, *DT);
1153 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
1154 AllocaInst *AI = Allocas[i];
1156 // Build list of instructions to promote.
1157 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
1159 Insts.push_back(cast<Instruction>(*UI));
1161 AllocaPromoter(Insts, SSA).run(AI, Insts);
1165 NumPromoted += Allocas.size();
1173 /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for
1174 /// SROA. It must be a struct or array type with a small number of elements.
1175 static bool ShouldAttemptScalarRepl(AllocaInst *AI) {
1176 const Type *T = AI->getAllocatedType();
1177 // Do not promote any struct into more than 32 separate vars.
1178 if (const StructType *ST = dyn_cast<StructType>(T))
1179 return ST->getNumElements() <= 32;
1180 // Arrays are much less likely to be safe for SROA; only consider
1181 // them if they are very small.
1182 if (const ArrayType *AT = dyn_cast<ArrayType>(T))
1183 return AT->getNumElements() <= 8;
1188 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
1189 // which runs on all of the malloc/alloca instructions in the function, removing
1190 // them if they are only used by getelementptr instructions.
1192 bool SROA::performScalarRepl(Function &F) {
1193 std::vector<AllocaInst*> WorkList;
1195 // Scan the entry basic block, adding allocas to the worklist.
1196 BasicBlock &BB = F.getEntryBlock();
1197 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
1198 if (AllocaInst *A = dyn_cast<AllocaInst>(I))
1199 WorkList.push_back(A);
1201 // Process the worklist
1202 bool Changed = false;
1203 while (!WorkList.empty()) {
1204 AllocaInst *AI = WorkList.back();
1205 WorkList.pop_back();
1207 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
1208 // with unused elements.
1209 if (AI->use_empty()) {
1210 AI->eraseFromParent();
1215 // If this alloca is impossible for us to promote, reject it early.
1216 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
1219 // Check to see if this allocation is only modified by a memcpy/memmove from
1220 // a constant global. If this is the case, we can change all users to use
1221 // the constant global instead. This is commonly produced by the CFE by
1222 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
1223 // is only subsequently read.
1224 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
1225 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n');
1226 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n');
1227 Constant *TheSrc = cast<Constant>(TheCopy->getSource());
1228 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
1229 TheCopy->eraseFromParent(); // Don't mutate the global.
1230 AI->eraseFromParent();
1236 // Check to see if we can perform the core SROA transformation. We cannot
1237 // transform the allocation instruction if it is an array allocation
1238 // (allocations OF arrays are ok though), and an allocation of a scalar
1239 // value cannot be decomposed at all.
1240 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
1242 // Do not promote [0 x %struct].
1243 if (AllocaSize == 0) continue;
1245 // Do not promote any struct whose size is too big.
1246 if (AllocaSize > SRThreshold) continue;
1248 // If the alloca looks like a good candidate for scalar replacement, and if
1249 // all its users can be transformed, then split up the aggregate into its
1250 // separate elements.
1251 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) {
1252 DoScalarReplacement(AI, WorkList);
1257 // If we can turn this aggregate value (potentially with casts) into a
1258 // simple scalar value that can be mem2reg'd into a register value.
1259 // IsNotTrivial tracks whether this is something that mem2reg could have
1260 // promoted itself. If so, we don't want to transform it needlessly. Note
1261 // that we can't just check based on the type: the alloca may be of an i32
1262 // but that has pointer arithmetic to set byte 3 of it or something.
1263 if (AllocaInst *NewAI =
1264 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) {
1265 NewAI->takeName(AI);
1266 AI->eraseFromParent();
1272 // Otherwise, couldn't process this alloca.
1278 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
1279 /// predicate, do SROA now.
1280 void SROA::DoScalarReplacement(AllocaInst *AI,
1281 std::vector<AllocaInst*> &WorkList) {
1282 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n');
1283 SmallVector<AllocaInst*, 32> ElementAllocas;
1284 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
1285 ElementAllocas.reserve(ST->getNumContainedTypes());
1286 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
1287 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
1289 AI->getName() + "." + Twine(i), AI);
1290 ElementAllocas.push_back(NA);
1291 WorkList.push_back(NA); // Add to worklist for recursive processing
1294 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
1295 ElementAllocas.reserve(AT->getNumElements());
1296 const Type *ElTy = AT->getElementType();
1297 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1298 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
1299 AI->getName() + "." + Twine(i), AI);
1300 ElementAllocas.push_back(NA);
1301 WorkList.push_back(NA); // Add to worklist for recursive processing
1305 // Now that we have created the new alloca instructions, rewrite all the
1306 // uses of the old alloca.
1307 RewriteForScalarRepl(AI, AI, 0, ElementAllocas);
1309 // Now erase any instructions that were made dead while rewriting the alloca.
1310 DeleteDeadInstructions();
1311 AI->eraseFromParent();
1316 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list,
1317 /// recursively including all their operands that become trivially dead.
1318 void SROA::DeleteDeadInstructions() {
1319 while (!DeadInsts.empty()) {
1320 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val());
1322 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
1323 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
1324 // Zero out the operand and see if it becomes trivially dead.
1325 // (But, don't add allocas to the dead instruction list -- they are
1326 // already on the worklist and will be deleted separately.)
1328 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U))
1329 DeadInsts.push_back(U);
1332 I->eraseFromParent();
1336 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to
1337 /// performing scalar replacement of alloca AI. The results are flagged in
1338 /// the Info parameter. Offset indicates the position within AI that is
1339 /// referenced by this instruction.
1340 void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset,
1342 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1343 Instruction *User = cast<Instruction>(*UI);
1345 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1346 isSafeForScalarRepl(BC, Offset, Info);
1347 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1348 uint64_t GEPOffset = Offset;
1349 isSafeGEP(GEPI, GEPOffset, Info);
1351 isSafeForScalarRepl(GEPI, GEPOffset, Info);
1352 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1353 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1355 return MarkUnsafe(Info, User);
1356 isSafeMemAccess(Offset, Length->getZExtValue(), 0,
1357 UI.getOperandNo() == 0, Info, MI,
1358 true /*AllowWholeAccess*/);
1359 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1360 if (LI->isVolatile())
1361 return MarkUnsafe(Info, User);
1362 const Type *LIType = LI->getType();
1363 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
1364 LIType, false, Info, LI, true /*AllowWholeAccess*/);
1365 Info.hasALoadOrStore = true;
1367 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1368 // Store is ok if storing INTO the pointer, not storing the pointer
1369 if (SI->isVolatile() || SI->getOperand(0) == I)
1370 return MarkUnsafe(Info, User);
1372 const Type *SIType = SI->getOperand(0)->getType();
1373 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
1374 SIType, true, Info, SI, true /*AllowWholeAccess*/);
1375 Info.hasALoadOrStore = true;
1376 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1377 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1379 return MarkUnsafe(Info, User);
1381 if (Info.isUnsafe) return;
1386 /// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer
1387 /// derived from the alloca, we can often still split the alloca into elements.
1388 /// This is useful if we have a large alloca where one element is phi'd
1389 /// together somewhere: we can SRoA and promote all the other elements even if
1390 /// we end up not being able to promote this one.
1392 /// All we require is that the uses of the PHI do not index into other parts of
1393 /// the alloca. The most important use case for this is single load and stores
1394 /// that are PHI'd together, which can happen due to code sinking.
1395 void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset,
1397 // If we've already checked this PHI, don't do it again.
1398 if (PHINode *PN = dyn_cast<PHINode>(I))
1399 if (!Info.CheckedPHIs.insert(PN))
1402 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1403 Instruction *User = cast<Instruction>(*UI);
1405 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1406 isSafePHISelectUseForScalarRepl(BC, Offset, Info);
1407 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1408 // Only allow "bitcast" GEPs for simplicity. We could generalize this,
1409 // but would have to prove that we're staying inside of an element being
1411 if (!GEPI->hasAllZeroIndices())
1412 return MarkUnsafe(Info, User);
1413 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info);
1414 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1415 if (LI->isVolatile())
1416 return MarkUnsafe(Info, User);
1417 const Type *LIType = LI->getType();
1418 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
1419 LIType, false, Info, LI, false /*AllowWholeAccess*/);
1420 Info.hasALoadOrStore = true;
1422 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1423 // Store is ok if storing INTO the pointer, not storing the pointer
1424 if (SI->isVolatile() || SI->getOperand(0) == I)
1425 return MarkUnsafe(Info, User);
1427 const Type *SIType = SI->getOperand(0)->getType();
1428 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
1429 SIType, true, Info, SI, false /*AllowWholeAccess*/);
1430 Info.hasALoadOrStore = true;
1431 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1432 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1434 return MarkUnsafe(Info, User);
1436 if (Info.isUnsafe) return;
1440 /// isSafeGEP - Check if a GEP instruction can be handled for scalar
1441 /// replacement. It is safe when all the indices are constant, in-bounds
1442 /// references, and when the resulting offset corresponds to an element within
1443 /// the alloca type. The results are flagged in the Info parameter. Upon
1444 /// return, Offset is adjusted as specified by the GEP indices.
1445 void SROA::isSafeGEP(GetElementPtrInst *GEPI,
1446 uint64_t &Offset, AllocaInfo &Info) {
1447 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
1451 // Walk through the GEP type indices, checking the types that this indexes
1453 for (; GEPIt != E; ++GEPIt) {
1454 // Ignore struct elements, no extra checking needed for these.
1455 if ((*GEPIt)->isStructTy())
1458 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
1460 return MarkUnsafe(Info, GEPI);
1463 // Compute the offset due to this GEP and check if the alloca has a
1464 // component element at that offset.
1465 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1466 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
1467 &Indices[0], Indices.size());
1468 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0))
1469 MarkUnsafe(Info, GEPI);
1472 /// isHomogeneousAggregate - Check if type T is a struct or array containing
1473 /// elements of the same type (which is always true for arrays). If so,
1474 /// return true with NumElts and EltTy set to the number of elements and the
1475 /// element type, respectively.
1476 static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts,
1477 const Type *&EltTy) {
1478 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
1479 NumElts = AT->getNumElements();
1480 EltTy = (NumElts == 0 ? 0 : AT->getElementType());
1483 if (const StructType *ST = dyn_cast<StructType>(T)) {
1484 NumElts = ST->getNumContainedTypes();
1485 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0));
1486 for (unsigned n = 1; n < NumElts; ++n) {
1487 if (ST->getContainedType(n) != EltTy)
1495 /// isCompatibleAggregate - Check if T1 and T2 are either the same type or are
1496 /// "homogeneous" aggregates with the same element type and number of elements.
1497 static bool isCompatibleAggregate(const Type *T1, const Type *T2) {
1501 unsigned NumElts1, NumElts2;
1502 const Type *EltTy1, *EltTy2;
1503 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) &&
1504 isHomogeneousAggregate(T2, NumElts2, EltTy2) &&
1505 NumElts1 == NumElts2 &&
1512 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI
1513 /// alloca or has an offset and size that corresponds to a component element
1514 /// within it. The offset checked here may have been formed from a GEP with a
1515 /// pointer bitcasted to a different type.
1517 /// If AllowWholeAccess is true, then this allows uses of the entire alloca as a
1518 /// unit. If false, it only allows accesses known to be in a single element.
1519 void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
1520 const Type *MemOpType, bool isStore,
1521 AllocaInfo &Info, Instruction *TheAccess,
1522 bool AllowWholeAccess) {
1523 // Check if this is a load/store of the entire alloca.
1524 if (Offset == 0 && AllowWholeAccess &&
1525 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) {
1526 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer
1527 // loads/stores (which are essentially the same as the MemIntrinsics with
1528 // regard to copying padding between elements). But, if an alloca is
1529 // flagged as both a source and destination of such operations, we'll need
1530 // to check later for padding between elements.
1531 if (!MemOpType || MemOpType->isIntegerTy()) {
1533 Info.isMemCpyDst = true;
1535 Info.isMemCpySrc = true;
1538 // This is also safe for references using a type that is compatible with
1539 // the type of the alloca, so that loads/stores can be rewritten using
1540 // insertvalue/extractvalue.
1541 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) {
1542 Info.hasSubelementAccess = true;
1546 // Check if the offset/size correspond to a component within the alloca type.
1547 const Type *T = Info.AI->getAllocatedType();
1548 if (TypeHasComponent(T, Offset, MemSize)) {
1549 Info.hasSubelementAccess = true;
1553 return MarkUnsafe(Info, TheAccess);
1556 /// TypeHasComponent - Return true if T has a component type with the
1557 /// specified offset and size. If Size is zero, do not check the size.
1558 bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) {
1561 if (const StructType *ST = dyn_cast<StructType>(T)) {
1562 const StructLayout *Layout = TD->getStructLayout(ST);
1563 unsigned EltIdx = Layout->getElementContainingOffset(Offset);
1564 EltTy = ST->getContainedType(EltIdx);
1565 EltSize = TD->getTypeAllocSize(EltTy);
1566 Offset -= Layout->getElementOffset(EltIdx);
1567 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
1568 EltTy = AT->getElementType();
1569 EltSize = TD->getTypeAllocSize(EltTy);
1570 if (Offset >= AT->getNumElements() * EltSize)
1576 if (Offset == 0 && (Size == 0 || EltSize == Size))
1578 // Check if the component spans multiple elements.
1579 if (Offset + Size > EltSize)
1581 return TypeHasComponent(EltTy, Offset, Size);
1584 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite
1585 /// the instruction I, which references it, to use the separate elements.
1586 /// Offset indicates the position within AI that is referenced by this
1588 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
1589 SmallVector<AllocaInst*, 32> &NewElts) {
1590 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) {
1591 Use &TheUse = UI.getUse();
1592 Instruction *User = cast<Instruction>(*UI++);
1594 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1595 RewriteBitCast(BC, AI, Offset, NewElts);
1599 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1600 RewriteGEP(GEPI, AI, Offset, NewElts);
1604 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1605 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1606 uint64_t MemSize = Length->getZExtValue();
1608 MemSize == TD->getTypeAllocSize(AI->getAllocatedType()))
1609 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
1610 // Otherwise the intrinsic can only touch a single element and the
1611 // address operand will be updated, so nothing else needs to be done.
1615 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1616 const Type *LIType = LI->getType();
1618 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) {
1620 // %res = load { i32, i32 }* %alloc
1622 // %load.0 = load i32* %alloc.0
1623 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
1624 // %load.1 = load i32* %alloc.1
1625 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
1626 // (Also works for arrays instead of structs)
1627 Value *Insert = UndefValue::get(LIType);
1628 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1629 Value *Load = new LoadInst(NewElts[i], "load", LI);
1630 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
1632 LI->replaceAllUsesWith(Insert);
1633 DeadInsts.push_back(LI);
1634 } else if (LIType->isIntegerTy() &&
1635 TD->getTypeAllocSize(LIType) ==
1636 TD->getTypeAllocSize(AI->getAllocatedType())) {
1637 // If this is a load of the entire alloca to an integer, rewrite it.
1638 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
1643 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1644 Value *Val = SI->getOperand(0);
1645 const Type *SIType = Val->getType();
1646 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) {
1648 // store { i32, i32 } %val, { i32, i32 }* %alloc
1650 // %val.0 = extractvalue { i32, i32 } %val, 0
1651 // store i32 %val.0, i32* %alloc.0
1652 // %val.1 = extractvalue { i32, i32 } %val, 1
1653 // store i32 %val.1, i32* %alloc.1
1654 // (Also works for arrays instead of structs)
1655 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1656 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
1657 new StoreInst(Extract, NewElts[i], SI);
1659 DeadInsts.push_back(SI);
1660 } else if (SIType->isIntegerTy() &&
1661 TD->getTypeAllocSize(SIType) ==
1662 TD->getTypeAllocSize(AI->getAllocatedType())) {
1663 // If this is a store of the entire alloca from an integer, rewrite it.
1664 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
1669 if (isa<SelectInst>(User) || isa<PHINode>(User)) {
1670 // If we have a PHI user of the alloca itself (as opposed to a GEP or
1671 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to
1673 if (!isa<AllocaInst>(I)) continue;
1675 assert(Offset == 0 && NewElts[0] &&
1676 "Direct alloca use should have a zero offset");
1678 // If we have a use of the alloca, we know the derived uses will be
1679 // utilizing just the first element of the scalarized result. Insert a
1680 // bitcast of the first alloca before the user as required.
1681 AllocaInst *NewAI = NewElts[0];
1682 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI);
1683 NewAI->moveBefore(BCI);
1690 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced
1691 /// and recursively continue updating all of its uses.
1692 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
1693 SmallVector<AllocaInst*, 32> &NewElts) {
1694 RewriteForScalarRepl(BC, AI, Offset, NewElts);
1695 if (BC->getOperand(0) != AI)
1698 // The bitcast references the original alloca. Replace its uses with
1699 // references to the first new element alloca.
1700 Instruction *Val = NewElts[0];
1701 if (Val->getType() != BC->getDestTy()) {
1702 Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
1705 BC->replaceAllUsesWith(Val);
1706 DeadInsts.push_back(BC);
1709 /// FindElementAndOffset - Return the index of the element containing Offset
1710 /// within the specified type, which must be either a struct or an array.
1711 /// Sets T to the type of the element and Offset to the offset within that
1712 /// element. IdxTy is set to the type of the index result to be used in a
1713 /// GEP instruction.
1714 uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset,
1715 const Type *&IdxTy) {
1717 if (const StructType *ST = dyn_cast<StructType>(T)) {
1718 const StructLayout *Layout = TD->getStructLayout(ST);
1719 Idx = Layout->getElementContainingOffset(Offset);
1720 T = ST->getContainedType(Idx);
1721 Offset -= Layout->getElementOffset(Idx);
1722 IdxTy = Type::getInt32Ty(T->getContext());
1725 const ArrayType *AT = cast<ArrayType>(T);
1726 T = AT->getElementType();
1727 uint64_t EltSize = TD->getTypeAllocSize(T);
1728 Idx = Offset / EltSize;
1729 Offset -= Idx * EltSize;
1730 IdxTy = Type::getInt64Ty(T->getContext());
1734 /// RewriteGEP - Check if this GEP instruction moves the pointer across
1735 /// elements of the alloca that are being split apart, and if so, rewrite
1736 /// the GEP to be relative to the new element.
1737 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
1738 SmallVector<AllocaInst*, 32> &NewElts) {
1739 uint64_t OldOffset = Offset;
1740 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1741 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
1742 &Indices[0], Indices.size());
1744 RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
1746 const Type *T = AI->getAllocatedType();
1748 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy);
1749 if (GEPI->getOperand(0) == AI)
1750 OldIdx = ~0ULL; // Force the GEP to be rewritten.
1752 T = AI->getAllocatedType();
1753 uint64_t EltOffset = Offset;
1754 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
1756 // If this GEP does not move the pointer across elements of the alloca
1757 // being split, then it does not needs to be rewritten.
1761 const Type *i32Ty = Type::getInt32Ty(AI->getContext());
1762 SmallVector<Value*, 8> NewArgs;
1763 NewArgs.push_back(Constant::getNullValue(i32Ty));
1764 while (EltOffset != 0) {
1765 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
1766 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
1768 Instruction *Val = NewElts[Idx];
1769 if (NewArgs.size() > 1) {
1770 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(),
1771 NewArgs.end(), "", GEPI);
1772 Val->takeName(GEPI);
1774 if (Val->getType() != GEPI->getType())
1775 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI);
1776 GEPI->replaceAllUsesWith(Val);
1777 DeadInsts.push_back(GEPI);
1780 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
1781 /// Rewrite it to copy or set the elements of the scalarized memory.
1782 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
1784 SmallVector<AllocaInst*, 32> &NewElts) {
1785 // If this is a memcpy/memmove, construct the other pointer as the
1786 // appropriate type. The "Other" pointer is the pointer that goes to memory
1787 // that doesn't have anything to do with the alloca that we are promoting. For
1788 // memset, this Value* stays null.
1789 Value *OtherPtr = 0;
1790 unsigned MemAlignment = MI->getAlignment();
1791 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
1792 if (Inst == MTI->getRawDest())
1793 OtherPtr = MTI->getRawSource();
1795 assert(Inst == MTI->getRawSource());
1796 OtherPtr = MTI->getRawDest();
1800 // If there is an other pointer, we want to convert it to the same pointer
1801 // type as AI has, so we can GEP through it safely.
1803 unsigned AddrSpace =
1804 cast<PointerType>(OtherPtr->getType())->getAddressSpace();
1806 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an
1807 // optimization, but it's also required to detect the corner case where
1808 // both pointer operands are referencing the same memory, and where
1809 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This
1810 // function is only called for mem intrinsics that access the whole
1811 // aggregate, so non-zero GEPs are not an issue here.)
1812 OtherPtr = OtherPtr->stripPointerCasts();
1814 // Copying the alloca to itself is a no-op: just delete it.
1815 if (OtherPtr == AI || OtherPtr == NewElts[0]) {
1816 // This code will run twice for a no-op memcpy -- once for each operand.
1817 // Put only one reference to MI on the DeadInsts list.
1818 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(),
1819 E = DeadInsts.end(); I != E; ++I)
1820 if (*I == MI) return;
1821 DeadInsts.push_back(MI);
1825 // If the pointer is not the right type, insert a bitcast to the right
1828 PointerType::get(AI->getType()->getElementType(), AddrSpace);
1830 if (OtherPtr->getType() != NewTy)
1831 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI);
1834 // Process each element of the aggregate.
1835 bool SROADest = MI->getRawDest() == Inst;
1837 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
1839 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1840 // If this is a memcpy/memmove, emit a GEP of the other element address.
1841 Value *OtherElt = 0;
1842 unsigned OtherEltAlign = MemAlignment;
1845 Value *Idx[2] = { Zero,
1846 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
1847 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2,
1848 OtherPtr->getName()+"."+Twine(i),
1851 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
1852 const Type *OtherTy = OtherPtrTy->getElementType();
1853 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) {
1854 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
1856 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
1857 EltOffset = TD->getTypeAllocSize(EltTy)*i;
1860 // The alignment of the other pointer is the guaranteed alignment of the
1861 // element, which is affected by both the known alignment of the whole
1862 // mem intrinsic and the alignment of the element. If the alignment of
1863 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
1864 // known alignment is just 4 bytes.
1865 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
1868 Value *EltPtr = NewElts[i];
1869 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
1871 // If we got down to a scalar, insert a load or store as appropriate.
1872 if (EltTy->isSingleValueType()) {
1873 if (isa<MemTransferInst>(MI)) {
1875 // From Other to Alloca.
1876 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
1877 new StoreInst(Elt, EltPtr, MI);
1879 // From Alloca to Other.
1880 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
1881 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
1885 assert(isa<MemSetInst>(MI));
1887 // If the stored element is zero (common case), just store a null
1890 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) {
1892 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
1894 // If EltTy is a vector type, get the element type.
1895 const Type *ValTy = EltTy->getScalarType();
1897 // Construct an integer with the right value.
1898 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
1899 APInt OneVal(EltSize, CI->getZExtValue());
1900 APInt TotalVal(OneVal);
1902 for (unsigned i = 0; 8*i < EltSize; ++i) {
1903 TotalVal = TotalVal.shl(8);
1907 // Convert the integer value to the appropriate type.
1908 StoreVal = ConstantInt::get(CI->getContext(), TotalVal);
1909 if (ValTy->isPointerTy())
1910 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
1911 else if (ValTy->isFloatingPointTy())
1912 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
1913 assert(StoreVal->getType() == ValTy && "Type mismatch!");
1915 // If the requested value was a vector constant, create it.
1916 if (EltTy != ValTy) {
1917 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
1918 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
1919 StoreVal = ConstantVector::get(&Elts[0], NumElts);
1922 new StoreInst(StoreVal, EltPtr, MI);
1925 // Otherwise, if we're storing a byte variable, use a memset call for
1929 unsigned EltSize = TD->getTypeAllocSize(EltTy);
1931 IRBuilder<> Builder(MI);
1933 // Finally, insert the meminst for this element.
1934 if (isa<MemSetInst>(MI)) {
1935 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize,
1938 assert(isa<MemTransferInst>(MI));
1939 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr
1940 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr
1942 if (isa<MemCpyInst>(MI))
1943 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile());
1945 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile());
1948 DeadInsts.push_back(MI);
1951 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that
1952 /// overwrites the entire allocation. Extract out the pieces of the stored
1953 /// integer and store them individually.
1954 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
1955 SmallVector<AllocaInst*, 32> &NewElts){
1956 // Extract each element out of the integer according to its structure offset
1957 // and store the element value to the individual alloca.
1958 Value *SrcVal = SI->getOperand(0);
1959 const Type *AllocaEltTy = AI->getAllocatedType();
1960 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
1962 IRBuilder<> Builder(SI);
1964 // Handle tail padding by extending the operand
1965 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
1966 SrcVal = Builder.CreateZExt(SrcVal,
1967 IntegerType::get(SI->getContext(), AllocaSizeBits));
1969 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI
1972 // There are two forms here: AI could be an array or struct. Both cases
1973 // have different ways to compute the element offset.
1974 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1975 const StructLayout *Layout = TD->getStructLayout(EltSTy);
1977 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1978 // Get the number of bits to shift SrcVal to get the value.
1979 const Type *FieldTy = EltSTy->getElementType(i);
1980 uint64_t Shift = Layout->getElementOffsetInBits(i);
1982 if (TD->isBigEndian())
1983 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
1985 Value *EltVal = SrcVal;
1987 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
1988 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
1991 // Truncate down to an integer of the right size.
1992 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1994 // Ignore zero sized fields like {}, they obviously contain no data.
1995 if (FieldSizeBits == 0) continue;
1997 if (FieldSizeBits != AllocaSizeBits)
1998 EltVal = Builder.CreateTrunc(EltVal,
1999 IntegerType::get(SI->getContext(), FieldSizeBits));
2000 Value *DestField = NewElts[i];
2001 if (EltVal->getType() == FieldTy) {
2002 // Storing to an integer field of this size, just do it.
2003 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) {
2004 // Bitcast to the right element type (for fp/vector values).
2005 EltVal = Builder.CreateBitCast(EltVal, FieldTy);
2007 // Otherwise, bitcast the dest pointer (for aggregates).
2008 DestField = Builder.CreateBitCast(DestField,
2009 PointerType::getUnqual(EltVal->getType()));
2011 new StoreInst(EltVal, DestField, SI);
2015 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
2016 const Type *ArrayEltTy = ATy->getElementType();
2017 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
2018 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
2022 if (TD->isBigEndian())
2023 Shift = AllocaSizeBits-ElementOffset;
2027 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2028 // Ignore zero sized fields like {}, they obviously contain no data.
2029 if (ElementSizeBits == 0) continue;
2031 Value *EltVal = SrcVal;
2033 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2034 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2037 // Truncate down to an integer of the right size.
2038 if (ElementSizeBits != AllocaSizeBits)
2039 EltVal = Builder.CreateTrunc(EltVal,
2040 IntegerType::get(SI->getContext(),
2042 Value *DestField = NewElts[i];
2043 if (EltVal->getType() == ArrayEltTy) {
2044 // Storing to an integer field of this size, just do it.
2045 } else if (ArrayEltTy->isFloatingPointTy() ||
2046 ArrayEltTy->isVectorTy()) {
2047 // Bitcast to the right element type (for fp/vector values).
2048 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy);
2050 // Otherwise, bitcast the dest pointer (for aggregates).
2051 DestField = Builder.CreateBitCast(DestField,
2052 PointerType::getUnqual(EltVal->getType()));
2054 new StoreInst(EltVal, DestField, SI);
2056 if (TD->isBigEndian())
2057 Shift -= ElementOffset;
2059 Shift += ElementOffset;
2063 DeadInsts.push_back(SI);
2066 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to
2067 /// an integer. Load the individual pieces to form the aggregate value.
2068 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
2069 SmallVector<AllocaInst*, 32> &NewElts) {
2070 // Extract each element out of the NewElts according to its structure offset
2071 // and form the result value.
2072 const Type *AllocaEltTy = AI->getAllocatedType();
2073 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
2075 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
2078 // There are two forms here: AI could be an array or struct. Both cases
2079 // have different ways to compute the element offset.
2080 const StructLayout *Layout = 0;
2081 uint64_t ArrayEltBitOffset = 0;
2082 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2083 Layout = TD->getStructLayout(EltSTy);
2085 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
2086 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
2090 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
2092 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2093 // Load the value from the alloca. If the NewElt is an aggregate, cast
2094 // the pointer to an integer of the same size before doing the load.
2095 Value *SrcField = NewElts[i];
2096 const Type *FieldTy =
2097 cast<PointerType>(SrcField->getType())->getElementType();
2098 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
2100 // Ignore zero sized fields like {}, they obviously contain no data.
2101 if (FieldSizeBits == 0) continue;
2103 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
2105 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() &&
2106 !FieldTy->isVectorTy())
2107 SrcField = new BitCastInst(SrcField,
2108 PointerType::getUnqual(FieldIntTy),
2110 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
2112 // If SrcField is a fp or vector of the right size but that isn't an
2113 // integer type, bitcast to an integer so we can shift it.
2114 if (SrcField->getType() != FieldIntTy)
2115 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
2117 // Zero extend the field to be the same size as the final alloca so that
2118 // we can shift and insert it.
2119 if (SrcField->getType() != ResultVal->getType())
2120 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
2122 // Determine the number of bits to shift SrcField.
2124 if (Layout) // Struct case.
2125 Shift = Layout->getElementOffsetInBits(i);
2127 Shift = i*ArrayEltBitOffset;
2129 if (TD->isBigEndian())
2130 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
2133 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
2134 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
2137 // Don't create an 'or x, 0' on the first iteration.
2138 if (!isa<Constant>(ResultVal) ||
2139 !cast<Constant>(ResultVal)->isNullValue())
2140 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
2142 ResultVal = SrcField;
2145 // Handle tail padding by truncating the result
2146 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
2147 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
2149 LI->replaceAllUsesWith(ResultVal);
2150 DeadInsts.push_back(LI);
2153 /// HasPadding - Return true if the specified type has any structure or
2154 /// alignment padding in between the elements that would be split apart
2155 /// by SROA; return false otherwise.
2156 static bool HasPadding(const Type *Ty, const TargetData &TD) {
2157 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2158 Ty = ATy->getElementType();
2159 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
2162 // SROA currently handles only Arrays and Structs.
2163 const StructType *STy = cast<StructType>(Ty);
2164 const StructLayout *SL = TD.getStructLayout(STy);
2165 unsigned PrevFieldBitOffset = 0;
2166 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2167 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
2169 // Check to see if there is any padding between this element and the
2172 unsigned PrevFieldEnd =
2173 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
2174 if (PrevFieldEnd < FieldBitOffset)
2177 PrevFieldBitOffset = FieldBitOffset;
2179 // Check for tail padding.
2180 if (unsigned EltCount = STy->getNumElements()) {
2181 unsigned PrevFieldEnd = PrevFieldBitOffset +
2182 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
2183 if (PrevFieldEnd < SL->getSizeInBits())
2189 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
2190 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
2191 /// or 1 if safe after canonicalization has been performed.
2192 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) {
2193 // Loop over the use list of the alloca. We can only transform it if all of
2194 // the users are safe to transform.
2195 AllocaInfo Info(AI);
2197 isSafeForScalarRepl(AI, 0, Info);
2198 if (Info.isUnsafe) {
2199 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n');
2203 // Okay, we know all the users are promotable. If the aggregate is a memcpy
2204 // source and destination, we have to be careful. In particular, the memcpy
2205 // could be moving around elements that live in structure padding of the LLVM
2206 // types, but may actually be used. In these cases, we refuse to promote the
2208 if (Info.isMemCpySrc && Info.isMemCpyDst &&
2209 HasPadding(AI->getAllocatedType(), *TD))
2212 // If the alloca never has an access to just *part* of it, but is accessed
2213 // via loads and stores, then we should use ConvertToScalarInfo to promote
2214 // the alloca instead of promoting each piece at a time and inserting fission
2216 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) {
2217 // If the struct/array just has one element, use basic SRoA.
2218 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
2219 if (ST->getNumElements() > 1) return false;
2221 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1)
2231 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
2232 /// some part of a constant global variable. This intentionally only accepts
2233 /// constant expressions because we don't can't rewrite arbitrary instructions.
2234 static bool PointsToConstantGlobal(Value *V) {
2235 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
2236 return GV->isConstant();
2237 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
2238 if (CE->getOpcode() == Instruction::BitCast ||
2239 CE->getOpcode() == Instruction::GetElementPtr)
2240 return PointsToConstantGlobal(CE->getOperand(0));
2244 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
2245 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
2246 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
2247 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
2248 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
2249 /// the alloca, and if the source pointer is a pointer to a constant global, we
2250 /// can optimize this.
2251 static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
2253 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
2254 User *U = cast<Instruction>(*UI);
2256 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
2257 // Ignore non-volatile loads, they are always ok.
2258 if (LI->isVolatile()) return false;
2262 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
2263 // If uses of the bitcast are ok, we are ok.
2264 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
2268 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
2269 // If the GEP has all zero indices, it doesn't offset the pointer. If it
2270 // doesn't, it does.
2271 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
2272 isOffset || !GEP->hasAllZeroIndices()))
2277 if (CallSite CS = U) {
2278 // If this is a readonly/readnone call site, then we know it is just a
2279 // load and we can ignore it.
2280 if (CS.onlyReadsMemory())
2283 // If this is the function being called then we treat it like a load and
2285 if (CS.isCallee(UI))
2288 // If this is being passed as a byval argument, the caller is making a
2289 // copy, so it is only a read of the alloca.
2290 unsigned ArgNo = CS.getArgumentNo(UI);
2291 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal))
2295 // If this is isn't our memcpy/memmove, reject it as something we can't
2297 MemTransferInst *MI = dyn_cast<MemTransferInst>(U);
2301 // If the transfer is using the alloca as a source of the transfer, then
2302 // ignore it since it is a load (unless the transfer is volatile).
2303 if (UI.getOperandNo() == 1) {
2304 if (MI->isVolatile()) return false;
2308 // If we already have seen a copy, reject the second one.
2309 if (TheCopy) return false;
2311 // If the pointer has been offset from the start of the alloca, we can't
2312 // safely handle this.
2313 if (isOffset) return false;
2315 // If the memintrinsic isn't using the alloca as the dest, reject it.
2316 if (UI.getOperandNo() != 0) return false;
2318 // If the source of the memcpy/move is not a constant global, reject it.
2319 if (!PointsToConstantGlobal(MI->getSource()))
2322 // Otherwise, the transform is safe. Remember the copy instruction.
2328 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
2329 /// modified by a copy from a constant global. If we can prove this, we can
2330 /// replace any uses of the alloca with uses of the global directly.
2331 MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) {
2332 MemTransferInst *TheCopy = 0;
2333 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))