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/ValueTracking.h"
35 #include "llvm/Target/TargetData.h"
36 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
37 #include "llvm/Transforms/Utils/Local.h"
38 #include "llvm/Support/CallSite.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/ErrorHandling.h"
41 #include "llvm/Support/GetElementPtrTypeIterator.h"
42 #include "llvm/Support/IRBuilder.h"
43 #include "llvm/Support/MathExtras.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include "llvm/ADT/SmallVector.h"
46 #include "llvm/ADT/Statistic.h"
49 STATISTIC(NumReplaced, "Number of allocas broken up");
50 STATISTIC(NumPromoted, "Number of allocas promoted");
51 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
52 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
55 struct SROA : public FunctionPass {
56 static char ID; // Pass identification, replacement for typeid
57 explicit SROA(signed T = -1) : FunctionPass(ID) {
58 initializeSROAPass(*PassRegistry::getPassRegistry());
65 bool runOnFunction(Function &F);
67 bool performScalarRepl(Function &F);
68 bool performPromotion(Function &F);
70 // getAnalysisUsage - This pass does not require any passes, but we know it
71 // will not alter the CFG, so say so.
72 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
73 AU.addRequired<DominatorTree>();
74 AU.addRequired<DominanceFrontier>();
81 /// DeadInsts - Keep track of instructions we have made dead, so that
82 /// we can remove them after we are done working.
83 SmallVector<Value*, 32> DeadInsts;
85 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
86 /// information about the uses. All these fields are initialized to false
87 /// and set to true when something is learned.
89 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
92 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
95 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
99 : isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false) {}
102 unsigned SRThreshold;
104 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; }
106 bool isSafeAllocaToScalarRepl(AllocaInst *AI);
108 void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
110 void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset,
112 void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize,
113 const Type *MemOpType, bool isStore, AllocaInfo &Info);
114 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size);
115 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset,
118 void DoScalarReplacement(AllocaInst *AI,
119 std::vector<AllocaInst*> &WorkList);
120 void DeleteDeadInstructions();
122 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
123 SmallVector<AllocaInst*, 32> &NewElts);
124 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
125 SmallVector<AllocaInst*, 32> &NewElts);
126 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
127 SmallVector<AllocaInst*, 32> &NewElts);
128 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
130 SmallVector<AllocaInst*, 32> &NewElts);
131 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
132 SmallVector<AllocaInst*, 32> &NewElts);
133 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
134 SmallVector<AllocaInst*, 32> &NewElts);
136 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI);
141 INITIALIZE_PASS_BEGIN(SROA, "scalarrepl",
142 "Scalar Replacement of Aggregates", false, false)
143 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
144 INITIALIZE_PASS_DEPENDENCY(DominanceFrontier)
145 INITIALIZE_PASS_END(SROA, "scalarrepl",
146 "Scalar Replacement of Aggregates", false, false)
148 // Public interface to the ScalarReplAggregates pass
149 FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) {
150 return new SROA(Threshold);
154 //===----------------------------------------------------------------------===//
155 // Convert To Scalar Optimization.
156 //===----------------------------------------------------------------------===//
159 /// ConvertToScalarInfo - This class implements the "Convert To Scalar"
160 /// optimization, which scans the uses of an alloca and determines if it can
161 /// rewrite it in terms of a single new alloca that can be mem2reg'd.
162 class ConvertToScalarInfo {
163 /// AllocaSize - The size of the alloca being considered.
165 const TargetData &TD;
167 /// IsNotTrivial - This is set to true if there is some access to the object
168 /// which means that mem2reg can't promote it.
171 /// VectorTy - This tracks the type that we should promote the vector to if
172 /// it is possible to turn it into a vector. This starts out null, and if it
173 /// isn't possible to turn into a vector type, it gets set to VoidTy.
174 const Type *VectorTy;
176 /// HadAVector - True if there is at least one vector access to the alloca.
177 /// We don't want to turn random arrays into vectors and use vector element
178 /// insert/extract, but if there are element accesses to something that is
179 /// also declared as a vector, we do want to promote to a vector.
183 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td)
184 : AllocaSize(Size), TD(td) {
185 IsNotTrivial = false;
190 AllocaInst *TryConvert(AllocaInst *AI);
193 bool CanConvertToScalar(Value *V, uint64_t Offset);
194 void MergeInType(const Type *In, uint64_t Offset);
195 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
197 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
198 uint64_t Offset, IRBuilder<> &Builder);
199 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
200 uint64_t Offset, IRBuilder<> &Builder);
202 } // end anonymous namespace.
205 /// IsVerbotenVectorType - Return true if this is a vector type ScalarRepl isn't
206 /// allowed to form. We do this to avoid MMX types, which is a complete hack,
207 /// but is required until the backend is fixed.
208 static bool IsVerbotenVectorType(const VectorType *VTy, const Instruction *I) {
209 StringRef Triple(I->getParent()->getParent()->getParent()->getTargetTriple());
210 if (!Triple.startswith("i386") &&
211 !Triple.startswith("x86_64"))
214 // Reject all the MMX vector types.
215 switch (VTy->getNumElements()) {
216 default: return false;
217 case 1: return VTy->getElementType()->isIntegerTy(64);
218 case 2: return VTy->getElementType()->isIntegerTy(32);
219 case 4: return VTy->getElementType()->isIntegerTy(16);
220 case 8: return VTy->getElementType()->isIntegerTy(8);
225 /// TryConvert - Analyze the specified alloca, and if it is safe to do so,
226 /// rewrite it to be a new alloca which is mem2reg'able. This returns the new
227 /// alloca if possible or null if not.
228 AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
229 // If we can't convert this scalar, or if mem2reg can trivially do it, bail
231 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial)
234 // If we were able to find a vector type that can handle this with
235 // insert/extract elements, and if there was at least one use that had
236 // a vector type, promote this to a vector. We don't want to promote
237 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
238 // we just get a lot of insert/extracts. If at least one vector is
239 // involved, then we probably really do have a union of vector/array.
241 if (VectorTy && VectorTy->isVectorTy() && HadAVector &&
242 !IsVerbotenVectorType(cast<VectorType>(VectorTy), AI)) {
243 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = "
244 << *VectorTy << '\n');
245 NewTy = VectorTy; // Use the vector type.
247 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
248 // Create and insert the integer alloca.
249 NewTy = IntegerType::get(AI->getContext(), AllocaSize*8);
251 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
252 ConvertUsesToScalar(AI, NewAI, 0);
256 /// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy)
257 /// so far at the offset specified by Offset (which is specified in bytes).
259 /// There are two cases we handle here:
260 /// 1) A union of vector types of the same size and potentially its elements.
261 /// Here we turn element accesses into insert/extract element operations.
262 /// This promotes a <4 x float> with a store of float to the third element
263 /// into a <4 x float> that uses insert element.
264 /// 2) A fully general blob of memory, which we turn into some (potentially
265 /// large) integer type with extract and insert operations where the loads
266 /// and stores would mutate the memory. We mark this by setting VectorTy
268 void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) {
269 // If we already decided to turn this into a blob of integer memory, there is
270 // nothing to be done.
271 if (VectorTy && VectorTy->isVoidTy())
274 // If this could be contributing to a vector, analyze it.
276 // If the In type is a vector that is the same size as the alloca, see if it
277 // matches the existing VecTy.
278 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
279 // Remember if we saw a vector type.
282 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
283 // If we're storing/loading a vector of the right size, allow it as a
284 // vector. If this the first vector we see, remember the type so that
285 // we know the element size. If this is a subsequent access, ignore it
286 // even if it is a differing type but the same size. Worst case we can
287 // bitcast the resultant vectors.
292 } else if (In->isFloatTy() || In->isDoubleTy() ||
293 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 &&
294 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
295 // If we're accessing something that could be an element of a vector, see
296 // if the implied vector agrees with what we already have and if Offset is
297 // compatible with it.
298 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
299 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 &&
301 cast<VectorType>(VectorTy)->getElementType()
302 ->getPrimitiveSizeInBits()/8 == EltSize)) {
304 VectorTy = VectorType::get(In, AllocaSize/EltSize);
309 // Otherwise, we have a case that we can't handle with an optimized vector
310 // form. We can still turn this into a large integer.
311 VectorTy = Type::getVoidTy(In->getContext());
314 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
315 /// its accesses to a single vector type, return true and set VecTy to
316 /// the new type. If we could convert the alloca into a single promotable
317 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
318 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
319 /// is the current offset from the base of the alloca being analyzed.
321 /// If we see at least one access to the value that is as a vector type, set the
323 bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) {
324 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
325 Instruction *User = cast<Instruction>(*UI);
327 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
328 // Don't break volatile loads.
329 if (LI->isVolatile())
331 // Don't touch MMX operations.
332 if (LI->getType()->isX86_MMXTy())
334 MergeInType(LI->getType(), Offset);
338 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
339 // Storing the pointer, not into the value?
340 if (SI->getOperand(0) == V || SI->isVolatile()) return false;
341 // Don't touch MMX operations.
342 if (SI->getOperand(0)->getType()->isX86_MMXTy())
344 MergeInType(SI->getOperand(0)->getType(), Offset);
348 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
349 IsNotTrivial = true; // Can't be mem2reg'd.
350 if (!CanConvertToScalar(BCI, Offset))
355 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
356 // If this is a GEP with a variable indices, we can't handle it.
357 if (!GEP->hasAllConstantIndices())
360 // Compute the offset that this GEP adds to the pointer.
361 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
362 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
363 &Indices[0], Indices.size());
364 // See if all uses can be converted.
365 if (!CanConvertToScalar(GEP, Offset+GEPOffset))
367 IsNotTrivial = true; // Can't be mem2reg'd.
371 // If this is a constant sized memset of a constant value (e.g. 0) we can
373 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
374 // Store of constant value and constant size.
375 if (!isa<ConstantInt>(MSI->getValue()) ||
376 !isa<ConstantInt>(MSI->getLength()))
378 IsNotTrivial = true; // Can't be mem2reg'd.
382 // If this is a memcpy or memmove into or out of the whole allocation, we
383 // can handle it like a load or store of the scalar type.
384 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
385 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength());
386 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0)
389 IsNotTrivial = true; // Can't be mem2reg'd.
393 // Otherwise, we cannot handle this!
400 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
401 /// directly. This happens when we are converting an "integer union" to a
402 /// single integer scalar, or when we are converting a "vector union" to a
403 /// vector with insert/extractelement instructions.
405 /// Offset is an offset from the original alloca, in bits that need to be
406 /// shifted to the right. By the end of this, there should be no uses of Ptr.
407 void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI,
409 while (!Ptr->use_empty()) {
410 Instruction *User = cast<Instruction>(Ptr->use_back());
412 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
413 ConvertUsesToScalar(CI, NewAI, Offset);
414 CI->eraseFromParent();
418 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
419 // Compute the offset that this GEP adds to the pointer.
420 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
421 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
422 &Indices[0], Indices.size());
423 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
424 GEP->eraseFromParent();
428 IRBuilder<> Builder(User->getParent(), User);
430 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
431 // The load is a bit extract from NewAI shifted right by Offset bits.
432 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
434 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
435 LI->replaceAllUsesWith(NewLoadVal);
436 LI->eraseFromParent();
440 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
441 assert(SI->getOperand(0) != Ptr && "Consistency error!");
442 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
443 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
445 Builder.CreateStore(New, NewAI);
446 SI->eraseFromParent();
448 // If the load we just inserted is now dead, then the inserted store
449 // overwrote the entire thing.
450 if (Old->use_empty())
451 Old->eraseFromParent();
455 // If this is a constant sized memset of a constant value (e.g. 0) we can
456 // transform it into a store of the expanded constant value.
457 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
458 assert(MSI->getRawDest() == Ptr && "Consistency error!");
459 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
461 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
463 // Compute the value replicated the right number of times.
464 APInt APVal(NumBytes*8, Val);
466 // Splat the value if non-zero.
468 for (unsigned i = 1; i != NumBytes; ++i)
471 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
472 Value *New = ConvertScalar_InsertValue(
473 ConstantInt::get(User->getContext(), APVal),
474 Old, Offset, Builder);
475 Builder.CreateStore(New, NewAI);
477 // If the load we just inserted is now dead, then the memset overwrote
479 if (Old->use_empty())
480 Old->eraseFromParent();
482 MSI->eraseFromParent();
486 // If this is a memcpy or memmove into or out of the whole allocation, we
487 // can handle it like a load or store of the scalar type.
488 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
489 assert(Offset == 0 && "must be store to start of alloca");
491 // If the source and destination are both to the same alloca, then this is
492 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
494 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, 0));
496 if (GetUnderlyingObject(MTI->getSource(), 0) != OrigAI) {
497 // Dest must be OrigAI, change this to be a load from the original
498 // pointer (bitcasted), then a store to our new alloca.
499 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
500 Value *SrcPtr = MTI->getSource();
501 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType());
503 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
504 SrcVal->setAlignment(MTI->getAlignment());
505 Builder.CreateStore(SrcVal, NewAI);
506 } else if (GetUnderlyingObject(MTI->getDest(), 0) != OrigAI) {
507 // Src must be OrigAI, change this to be a load from NewAI then a store
508 // through the original dest pointer (bitcasted).
509 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
510 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
512 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType());
513 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
514 NewStore->setAlignment(MTI->getAlignment());
516 // Noop transfer. Src == Dst
519 MTI->eraseFromParent();
523 llvm_unreachable("Unsupported operation!");
527 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
528 /// or vector value FromVal, extracting the bits from the offset specified by
529 /// Offset. This returns the value, which is of type ToType.
531 /// This happens when we are converting an "integer union" to a single
532 /// integer scalar, or when we are converting a "vector union" to a vector with
533 /// insert/extractelement instructions.
535 /// Offset is an offset from the original alloca, in bits that need to be
536 /// shifted to the right.
537 Value *ConvertToScalarInfo::
538 ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
539 uint64_t Offset, IRBuilder<> &Builder) {
540 // If the load is of the whole new alloca, no conversion is needed.
541 if (FromVal->getType() == ToType && Offset == 0)
544 // If the result alloca is a vector type, this is either an element
545 // access or a bitcast to another vector type of the same size.
546 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
547 if (ToType->isVectorTy())
548 return Builder.CreateBitCast(FromVal, ToType, "tmp");
550 // Otherwise it must be an element access.
553 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
554 Elt = Offset/EltSize;
555 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
557 // Return the element extracted out of it.
558 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get(
559 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp");
560 if (V->getType() != ToType)
561 V = Builder.CreateBitCast(V, ToType, "tmp");
565 // If ToType is a first class aggregate, extract out each of the pieces and
566 // use insertvalue's to form the FCA.
567 if (const StructType *ST = dyn_cast<StructType>(ToType)) {
568 const StructLayout &Layout = *TD.getStructLayout(ST);
569 Value *Res = UndefValue::get(ST);
570 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
571 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
572 Offset+Layout.getElementOffsetInBits(i),
574 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
579 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
580 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
581 Value *Res = UndefValue::get(AT);
582 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
583 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
584 Offset+i*EltSize, Builder);
585 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
590 // Otherwise, this must be a union that was converted to an integer value.
591 const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
593 // If this is a big-endian system and the load is narrower than the
594 // full alloca type, we need to do a shift to get the right bits.
596 if (TD.isBigEndian()) {
597 // On big-endian machines, the lowest bit is stored at the bit offset
598 // from the pointer given by getTypeStoreSizeInBits. This matters for
599 // integers with a bitwidth that is not a multiple of 8.
600 ShAmt = TD.getTypeStoreSizeInBits(NTy) -
601 TD.getTypeStoreSizeInBits(ToType) - Offset;
606 // Note: we support negative bitwidths (with shl) which are not defined.
607 // We do this to support (f.e.) loads off the end of a structure where
608 // only some bits are used.
609 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
610 FromVal = Builder.CreateLShr(FromVal,
611 ConstantInt::get(FromVal->getType(),
613 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
614 FromVal = Builder.CreateShl(FromVal,
615 ConstantInt::get(FromVal->getType(),
618 // Finally, unconditionally truncate the integer to the right width.
619 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType);
620 if (LIBitWidth < NTy->getBitWidth())
622 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
624 else if (LIBitWidth > NTy->getBitWidth())
626 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
629 // If the result is an integer, this is a trunc or bitcast.
630 if (ToType->isIntegerTy()) {
632 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) {
633 // Just do a bitcast, we know the sizes match up.
634 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
636 // Otherwise must be a pointer.
637 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
639 assert(FromVal->getType() == ToType && "Didn't convert right?");
643 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
644 /// or vector value "Old" at the offset specified by Offset.
646 /// This happens when we are converting an "integer union" to a
647 /// single integer scalar, or when we are converting a "vector union" to a
648 /// vector with insert/extractelement instructions.
650 /// Offset is an offset from the original alloca, in bits that need to be
651 /// shifted to the right.
652 Value *ConvertToScalarInfo::
653 ConvertScalar_InsertValue(Value *SV, Value *Old,
654 uint64_t Offset, IRBuilder<> &Builder) {
655 // Convert the stored type to the actual type, shift it left to insert
656 // then 'or' into place.
657 const Type *AllocaType = Old->getType();
658 LLVMContext &Context = Old->getContext();
660 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
661 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy);
662 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType());
664 // Changing the whole vector with memset or with an access of a different
666 if (ValSize == VecSize)
667 return Builder.CreateBitCast(SV, AllocaType, "tmp");
669 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
671 // Must be an element insertion.
672 unsigned Elt = Offset/EltSize;
674 if (SV->getType() != VTy->getElementType())
675 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
677 SV = Builder.CreateInsertElement(Old, SV,
678 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt),
683 // If SV is a first-class aggregate value, insert each value recursively.
684 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
685 const StructLayout &Layout = *TD.getStructLayout(ST);
686 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
687 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
688 Old = ConvertScalar_InsertValue(Elt, Old,
689 Offset+Layout.getElementOffsetInBits(i),
695 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
696 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
697 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
698 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
699 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
704 // If SV is a float, convert it to the appropriate integer type.
705 // If it is a pointer, do the same.
706 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType());
707 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType);
708 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType());
709 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType);
710 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
711 SV = Builder.CreateBitCast(SV,
712 IntegerType::get(SV->getContext(),SrcWidth), "tmp");
713 else if (SV->getType()->isPointerTy())
714 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp");
716 // Zero extend or truncate the value if needed.
717 if (SV->getType() != AllocaType) {
718 if (SV->getType()->getPrimitiveSizeInBits() <
719 AllocaType->getPrimitiveSizeInBits())
720 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
722 // Truncation may be needed if storing more than the alloca can hold
723 // (undefined behavior).
724 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
725 SrcWidth = DestWidth;
726 SrcStoreWidth = DestStoreWidth;
730 // If this is a big-endian system and the store is narrower than the
731 // full alloca type, we need to do a shift to get the right bits.
733 if (TD.isBigEndian()) {
734 // On big-endian machines, the lowest bit is stored at the bit offset
735 // from the pointer given by getTypeStoreSizeInBits. This matters for
736 // integers with a bitwidth that is not a multiple of 8.
737 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
742 // Note: we support negative bitwidths (with shr) which are not defined.
743 // We do this to support (f.e.) stores off the end of a structure where
744 // only some bits in the structure are set.
745 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
746 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
747 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(),
750 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
751 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(),
753 Mask = Mask.lshr(-ShAmt);
756 // Mask out the bits we are about to insert from the old value, and or
758 if (SrcWidth != DestWidth) {
759 assert(DestWidth > SrcWidth);
760 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
761 SV = Builder.CreateOr(Old, SV, "ins");
767 //===----------------------------------------------------------------------===//
769 //===----------------------------------------------------------------------===//
772 bool SROA::runOnFunction(Function &F) {
773 TD = getAnalysisIfAvailable<TargetData>();
775 bool Changed = performPromotion(F);
777 // FIXME: ScalarRepl currently depends on TargetData more than it
778 // theoretically needs to. It should be refactored in order to support
779 // target-independent IR. Until this is done, just skip the actual
780 // scalar-replacement portion of this pass.
781 if (!TD) return Changed;
784 bool LocalChange = performScalarRepl(F);
785 if (!LocalChange) break; // No need to repromote if no scalarrepl
787 LocalChange = performPromotion(F);
788 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
795 bool SROA::performPromotion(Function &F) {
796 std::vector<AllocaInst*> Allocas;
797 DominatorTree &DT = getAnalysis<DominatorTree>();
798 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
800 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
802 bool Changed = false;
807 // Find allocas that are safe to promote, by looking at all instructions in
809 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
810 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
811 if (isAllocaPromotable(AI))
812 Allocas.push_back(AI);
814 if (Allocas.empty()) break;
816 PromoteMemToReg(Allocas, DT, DF);
817 NumPromoted += Allocas.size();
825 /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for
826 /// SROA. It must be a struct or array type with a small number of elements.
827 static bool ShouldAttemptScalarRepl(AllocaInst *AI) {
828 const Type *T = AI->getAllocatedType();
829 // Do not promote any struct into more than 32 separate vars.
830 if (const StructType *ST = dyn_cast<StructType>(T))
831 return ST->getNumElements() <= 32;
832 // Arrays are much less likely to be safe for SROA; only consider
833 // them if they are very small.
834 if (const ArrayType *AT = dyn_cast<ArrayType>(T))
835 return AT->getNumElements() <= 8;
840 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
841 // which runs on all of the malloc/alloca instructions in the function, removing
842 // them if they are only used by getelementptr instructions.
844 bool SROA::performScalarRepl(Function &F) {
845 std::vector<AllocaInst*> WorkList;
847 // Scan the entry basic block, adding allocas to the worklist.
848 BasicBlock &BB = F.getEntryBlock();
849 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
850 if (AllocaInst *A = dyn_cast<AllocaInst>(I))
851 WorkList.push_back(A);
853 // Process the worklist
854 bool Changed = false;
855 while (!WorkList.empty()) {
856 AllocaInst *AI = WorkList.back();
859 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
860 // with unused elements.
861 if (AI->use_empty()) {
862 AI->eraseFromParent();
867 // If this alloca is impossible for us to promote, reject it early.
868 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
871 // Check to see if this allocation is only modified by a memcpy/memmove from
872 // a constant global. If this is the case, we can change all users to use
873 // the constant global instead. This is commonly produced by the CFE by
874 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
875 // is only subsequently read.
876 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
877 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n');
878 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n');
879 Constant *TheSrc = cast<Constant>(TheCopy->getSource());
880 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
881 TheCopy->eraseFromParent(); // Don't mutate the global.
882 AI->eraseFromParent();
888 // Check to see if we can perform the core SROA transformation. We cannot
889 // transform the allocation instruction if it is an array allocation
890 // (allocations OF arrays are ok though), and an allocation of a scalar
891 // value cannot be decomposed at all.
892 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
894 // Do not promote [0 x %struct].
895 if (AllocaSize == 0) continue;
897 // Do not promote any struct whose size is too big.
898 if (AllocaSize > SRThreshold) continue;
900 // If the alloca looks like a good candidate for scalar replacement, and if
901 // all its users can be transformed, then split up the aggregate into its
902 // separate elements.
903 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) {
904 DoScalarReplacement(AI, WorkList);
909 // If we can turn this aggregate value (potentially with casts) into a
910 // simple scalar value that can be mem2reg'd into a register value.
911 // IsNotTrivial tracks whether this is something that mem2reg could have
912 // promoted itself. If so, we don't want to transform it needlessly. Note
913 // that we can't just check based on the type: the alloca may be of an i32
914 // but that has pointer arithmetic to set byte 3 of it or something.
915 if (AllocaInst *NewAI =
916 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) {
918 AI->eraseFromParent();
924 // Otherwise, couldn't process this alloca.
930 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
931 /// predicate, do SROA now.
932 void SROA::DoScalarReplacement(AllocaInst *AI,
933 std::vector<AllocaInst*> &WorkList) {
934 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n');
935 SmallVector<AllocaInst*, 32> ElementAllocas;
936 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
937 ElementAllocas.reserve(ST->getNumContainedTypes());
938 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
939 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
941 AI->getName() + "." + Twine(i), AI);
942 ElementAllocas.push_back(NA);
943 WorkList.push_back(NA); // Add to worklist for recursive processing
946 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
947 ElementAllocas.reserve(AT->getNumElements());
948 const Type *ElTy = AT->getElementType();
949 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
950 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
951 AI->getName() + "." + Twine(i), AI);
952 ElementAllocas.push_back(NA);
953 WorkList.push_back(NA); // Add to worklist for recursive processing
957 // Now that we have created the new alloca instructions, rewrite all the
958 // uses of the old alloca.
959 RewriteForScalarRepl(AI, AI, 0, ElementAllocas);
961 // Now erase any instructions that were made dead while rewriting the alloca.
962 DeleteDeadInstructions();
963 AI->eraseFromParent();
968 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list,
969 /// recursively including all their operands that become trivially dead.
970 void SROA::DeleteDeadInstructions() {
971 while (!DeadInsts.empty()) {
972 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val());
974 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
975 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
976 // Zero out the operand and see if it becomes trivially dead.
977 // (But, don't add allocas to the dead instruction list -- they are
978 // already on the worklist and will be deleted separately.)
980 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U))
981 DeadInsts.push_back(U);
984 I->eraseFromParent();
988 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to
989 /// performing scalar replacement of alloca AI. The results are flagged in
990 /// the Info parameter. Offset indicates the position within AI that is
991 /// referenced by this instruction.
992 void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
994 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
995 Instruction *User = cast<Instruction>(*UI);
997 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
998 isSafeForScalarRepl(BC, AI, Offset, Info);
999 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1000 uint64_t GEPOffset = Offset;
1001 isSafeGEP(GEPI, AI, GEPOffset, Info);
1003 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info);
1004 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1005 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1007 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0,
1008 UI.getOperandNo() == 0, Info);
1011 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1012 if (!LI->isVolatile()) {
1013 const Type *LIType = LI->getType();
1014 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType),
1015 LIType, false, Info);
1018 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1019 // Store is ok if storing INTO the pointer, not storing the pointer
1020 if (!SI->isVolatile() && SI->getOperand(0) != I) {
1021 const Type *SIType = SI->getOperand(0)->getType();
1022 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType),
1023 SIType, true, Info);
1027 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n');
1030 if (Info.isUnsafe) return;
1034 /// isSafeGEP - Check if a GEP instruction can be handled for scalar
1035 /// replacement. It is safe when all the indices are constant, in-bounds
1036 /// references, and when the resulting offset corresponds to an element within
1037 /// the alloca type. The results are flagged in the Info parameter. Upon
1038 /// return, Offset is adjusted as specified by the GEP indices.
1039 void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI,
1040 uint64_t &Offset, AllocaInfo &Info) {
1041 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
1045 // Walk through the GEP type indices, checking the types that this indexes
1047 for (; GEPIt != E; ++GEPIt) {
1048 // Ignore struct elements, no extra checking needed for these.
1049 if ((*GEPIt)->isStructTy())
1052 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
1054 return MarkUnsafe(Info);
1057 // Compute the offset due to this GEP and check if the alloca has a
1058 // component element at that offset.
1059 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1060 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
1061 &Indices[0], Indices.size());
1062 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0))
1066 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI
1067 /// alloca or has an offset and size that corresponds to a component element
1068 /// within it. The offset checked here may have been formed from a GEP with a
1069 /// pointer bitcasted to a different type.
1070 void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize,
1071 const Type *MemOpType, bool isStore,
1073 // Check if this is a load/store of the entire alloca.
1074 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) {
1075 bool UsesAggregateType = (MemOpType == AI->getAllocatedType());
1076 // This is safe for MemIntrinsics (where MemOpType is 0), integer types
1077 // (which are essentially the same as the MemIntrinsics, especially with
1078 // regard to copying padding between elements), or references using the
1079 // aggregate type of the alloca.
1080 if (!MemOpType || MemOpType->isIntegerTy() || UsesAggregateType) {
1081 if (!UsesAggregateType) {
1083 Info.isMemCpyDst = true;
1085 Info.isMemCpySrc = true;
1090 // Check if the offset/size correspond to a component within the alloca type.
1091 const Type *T = AI->getAllocatedType();
1092 if (TypeHasComponent(T, Offset, MemSize))
1095 return MarkUnsafe(Info);
1098 /// TypeHasComponent - Return true if T has a component type with the
1099 /// specified offset and size. If Size is zero, do not check the size.
1100 bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) {
1103 if (const StructType *ST = dyn_cast<StructType>(T)) {
1104 const StructLayout *Layout = TD->getStructLayout(ST);
1105 unsigned EltIdx = Layout->getElementContainingOffset(Offset);
1106 EltTy = ST->getContainedType(EltIdx);
1107 EltSize = TD->getTypeAllocSize(EltTy);
1108 Offset -= Layout->getElementOffset(EltIdx);
1109 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
1110 EltTy = AT->getElementType();
1111 EltSize = TD->getTypeAllocSize(EltTy);
1112 if (Offset >= AT->getNumElements() * EltSize)
1118 if (Offset == 0 && (Size == 0 || EltSize == Size))
1120 // Check if the component spans multiple elements.
1121 if (Offset + Size > EltSize)
1123 return TypeHasComponent(EltTy, Offset, Size);
1126 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite
1127 /// the instruction I, which references it, to use the separate elements.
1128 /// Offset indicates the position within AI that is referenced by this
1130 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
1131 SmallVector<AllocaInst*, 32> &NewElts) {
1132 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1133 Instruction *User = cast<Instruction>(*UI);
1135 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1136 RewriteBitCast(BC, AI, Offset, NewElts);
1137 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1138 RewriteGEP(GEPI, AI, Offset, NewElts);
1139 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1140 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1141 uint64_t MemSize = Length->getZExtValue();
1143 MemSize == TD->getTypeAllocSize(AI->getAllocatedType()))
1144 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
1145 // Otherwise the intrinsic can only touch a single element and the
1146 // address operand will be updated, so nothing else needs to be done.
1147 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1148 const Type *LIType = LI->getType();
1149 if (LIType == AI->getAllocatedType()) {
1151 // %res = load { i32, i32 }* %alloc
1153 // %load.0 = load i32* %alloc.0
1154 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
1155 // %load.1 = load i32* %alloc.1
1156 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
1157 // (Also works for arrays instead of structs)
1158 Value *Insert = UndefValue::get(LIType);
1159 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1160 Value *Load = new LoadInst(NewElts[i], "load", LI);
1161 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
1163 LI->replaceAllUsesWith(Insert);
1164 DeadInsts.push_back(LI);
1165 } else if (LIType->isIntegerTy() &&
1166 TD->getTypeAllocSize(LIType) ==
1167 TD->getTypeAllocSize(AI->getAllocatedType())) {
1168 // If this is a load of the entire alloca to an integer, rewrite it.
1169 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
1171 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1172 Value *Val = SI->getOperand(0);
1173 const Type *SIType = Val->getType();
1174 if (SIType == AI->getAllocatedType()) {
1176 // store { i32, i32 } %val, { i32, i32 }* %alloc
1178 // %val.0 = extractvalue { i32, i32 } %val, 0
1179 // store i32 %val.0, i32* %alloc.0
1180 // %val.1 = extractvalue { i32, i32 } %val, 1
1181 // store i32 %val.1, i32* %alloc.1
1182 // (Also works for arrays instead of structs)
1183 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1184 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
1185 new StoreInst(Extract, NewElts[i], SI);
1187 DeadInsts.push_back(SI);
1188 } else if (SIType->isIntegerTy() &&
1189 TD->getTypeAllocSize(SIType) ==
1190 TD->getTypeAllocSize(AI->getAllocatedType())) {
1191 // If this is a store of the entire alloca from an integer, rewrite it.
1192 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
1198 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced
1199 /// and recursively continue updating all of its uses.
1200 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
1201 SmallVector<AllocaInst*, 32> &NewElts) {
1202 RewriteForScalarRepl(BC, AI, Offset, NewElts);
1203 if (BC->getOperand(0) != AI)
1206 // The bitcast references the original alloca. Replace its uses with
1207 // references to the first new element alloca.
1208 Instruction *Val = NewElts[0];
1209 if (Val->getType() != BC->getDestTy()) {
1210 Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
1213 BC->replaceAllUsesWith(Val);
1214 DeadInsts.push_back(BC);
1217 /// FindElementAndOffset - Return the index of the element containing Offset
1218 /// within the specified type, which must be either a struct or an array.
1219 /// Sets T to the type of the element and Offset to the offset within that
1220 /// element. IdxTy is set to the type of the index result to be used in a
1221 /// GEP instruction.
1222 uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset,
1223 const Type *&IdxTy) {
1225 if (const StructType *ST = dyn_cast<StructType>(T)) {
1226 const StructLayout *Layout = TD->getStructLayout(ST);
1227 Idx = Layout->getElementContainingOffset(Offset);
1228 T = ST->getContainedType(Idx);
1229 Offset -= Layout->getElementOffset(Idx);
1230 IdxTy = Type::getInt32Ty(T->getContext());
1233 const ArrayType *AT = cast<ArrayType>(T);
1234 T = AT->getElementType();
1235 uint64_t EltSize = TD->getTypeAllocSize(T);
1236 Idx = Offset / EltSize;
1237 Offset -= Idx * EltSize;
1238 IdxTy = Type::getInt64Ty(T->getContext());
1242 /// RewriteGEP - Check if this GEP instruction moves the pointer across
1243 /// elements of the alloca that are being split apart, and if so, rewrite
1244 /// the GEP to be relative to the new element.
1245 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
1246 SmallVector<AllocaInst*, 32> &NewElts) {
1247 uint64_t OldOffset = Offset;
1248 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1249 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
1250 &Indices[0], Indices.size());
1252 RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
1254 const Type *T = AI->getAllocatedType();
1256 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy);
1257 if (GEPI->getOperand(0) == AI)
1258 OldIdx = ~0ULL; // Force the GEP to be rewritten.
1260 T = AI->getAllocatedType();
1261 uint64_t EltOffset = Offset;
1262 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
1264 // If this GEP does not move the pointer across elements of the alloca
1265 // being split, then it does not needs to be rewritten.
1269 const Type *i32Ty = Type::getInt32Ty(AI->getContext());
1270 SmallVector<Value*, 8> NewArgs;
1271 NewArgs.push_back(Constant::getNullValue(i32Ty));
1272 while (EltOffset != 0) {
1273 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
1274 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
1276 Instruction *Val = NewElts[Idx];
1277 if (NewArgs.size() > 1) {
1278 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(),
1279 NewArgs.end(), "", GEPI);
1280 Val->takeName(GEPI);
1282 if (Val->getType() != GEPI->getType())
1283 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI);
1284 GEPI->replaceAllUsesWith(Val);
1285 DeadInsts.push_back(GEPI);
1288 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
1289 /// Rewrite it to copy or set the elements of the scalarized memory.
1290 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
1292 SmallVector<AllocaInst*, 32> &NewElts) {
1293 // If this is a memcpy/memmove, construct the other pointer as the
1294 // appropriate type. The "Other" pointer is the pointer that goes to memory
1295 // that doesn't have anything to do with the alloca that we are promoting. For
1296 // memset, this Value* stays null.
1297 Value *OtherPtr = 0;
1298 unsigned MemAlignment = MI->getAlignment();
1299 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
1300 if (Inst == MTI->getRawDest())
1301 OtherPtr = MTI->getRawSource();
1303 assert(Inst == MTI->getRawSource());
1304 OtherPtr = MTI->getRawDest();
1308 // If there is an other pointer, we want to convert it to the same pointer
1309 // type as AI has, so we can GEP through it safely.
1311 unsigned AddrSpace =
1312 cast<PointerType>(OtherPtr->getType())->getAddressSpace();
1314 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an
1315 // optimization, but it's also required to detect the corner case where
1316 // both pointer operands are referencing the same memory, and where
1317 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This
1318 // function is only called for mem intrinsics that access the whole
1319 // aggregate, so non-zero GEPs are not an issue here.)
1320 OtherPtr = OtherPtr->stripPointerCasts();
1322 // Copying the alloca to itself is a no-op: just delete it.
1323 if (OtherPtr == AI || OtherPtr == NewElts[0]) {
1324 // This code will run twice for a no-op memcpy -- once for each operand.
1325 // Put only one reference to MI on the DeadInsts list.
1326 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(),
1327 E = DeadInsts.end(); I != E; ++I)
1328 if (*I == MI) return;
1329 DeadInsts.push_back(MI);
1333 // If the pointer is not the right type, insert a bitcast to the right
1336 PointerType::get(AI->getType()->getElementType(), AddrSpace);
1338 if (OtherPtr->getType() != NewTy)
1339 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI);
1342 // Process each element of the aggregate.
1343 Value *TheFn = MI->getCalledValue();
1344 const Type *BytePtrTy = MI->getRawDest()->getType();
1345 bool SROADest = MI->getRawDest() == Inst;
1347 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
1349 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1350 // If this is a memcpy/memmove, emit a GEP of the other element address.
1351 Value *OtherElt = 0;
1352 unsigned OtherEltAlign = MemAlignment;
1355 Value *Idx[2] = { Zero,
1356 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
1357 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2,
1358 OtherPtr->getName()+"."+Twine(i),
1361 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
1362 const Type *OtherTy = OtherPtrTy->getElementType();
1363 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) {
1364 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
1366 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
1367 EltOffset = TD->getTypeAllocSize(EltTy)*i;
1370 // The alignment of the other pointer is the guaranteed alignment of the
1371 // element, which is affected by both the known alignment of the whole
1372 // mem intrinsic and the alignment of the element. If the alignment of
1373 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
1374 // known alignment is just 4 bytes.
1375 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
1378 Value *EltPtr = NewElts[i];
1379 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
1381 // If we got down to a scalar, insert a load or store as appropriate.
1382 if (EltTy->isSingleValueType()) {
1383 if (isa<MemTransferInst>(MI)) {
1385 // From Other to Alloca.
1386 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
1387 new StoreInst(Elt, EltPtr, MI);
1389 // From Alloca to Other.
1390 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
1391 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
1395 assert(isa<MemSetInst>(MI));
1397 // If the stored element is zero (common case), just store a null
1400 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) {
1402 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
1404 // If EltTy is a vector type, get the element type.
1405 const Type *ValTy = EltTy->getScalarType();
1407 // Construct an integer with the right value.
1408 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
1409 APInt OneVal(EltSize, CI->getZExtValue());
1410 APInt TotalVal(OneVal);
1412 for (unsigned i = 0; 8*i < EltSize; ++i) {
1413 TotalVal = TotalVal.shl(8);
1417 // Convert the integer value to the appropriate type.
1418 StoreVal = ConstantInt::get(CI->getContext(), TotalVal);
1419 if (ValTy->isPointerTy())
1420 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
1421 else if (ValTy->isFloatingPointTy())
1422 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
1423 assert(StoreVal->getType() == ValTy && "Type mismatch!");
1425 // If the requested value was a vector constant, create it.
1426 if (EltTy != ValTy) {
1427 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
1428 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
1429 StoreVal = ConstantVector::get(&Elts[0], NumElts);
1432 new StoreInst(StoreVal, EltPtr, MI);
1435 // Otherwise, if we're storing a byte variable, use a memset call for
1439 // Cast the element pointer to BytePtrTy.
1440 if (EltPtr->getType() != BytePtrTy)
1441 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getName(), MI);
1443 // Cast the other pointer (if we have one) to BytePtrTy.
1444 if (OtherElt && OtherElt->getType() != BytePtrTy) {
1445 // Preserve address space of OtherElt
1446 const PointerType* OtherPTy = cast<PointerType>(OtherElt->getType());
1447 const PointerType* PTy = cast<PointerType>(BytePtrTy);
1448 if (OtherPTy->getElementType() != PTy->getElementType()) {
1449 Type *NewOtherPTy = PointerType::get(PTy->getElementType(),
1450 OtherPTy->getAddressSpace());
1451 OtherElt = new BitCastInst(OtherElt, NewOtherPTy,
1452 OtherElt->getName(), MI);
1456 unsigned EltSize = TD->getTypeAllocSize(EltTy);
1458 // Finally, insert the meminst for this element.
1459 if (isa<MemTransferInst>(MI)) {
1461 SROADest ? EltPtr : OtherElt, // Dest ptr
1462 SROADest ? OtherElt : EltPtr, // Src ptr
1463 ConstantInt::get(MI->getArgOperand(2)->getType(), EltSize), // Size
1465 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign),
1466 MI->getVolatileCst()
1468 // In case we fold the address space overloaded memcpy of A to B
1469 // with memcpy of B to C, change the function to be a memcpy of A to C.
1470 const Type *Tys[] = { Ops[0]->getType(), Ops[1]->getType(),
1471 Ops[2]->getType() };
1472 Module *M = MI->getParent()->getParent()->getParent();
1473 TheFn = Intrinsic::getDeclaration(M, MI->getIntrinsicID(), Tys, 3);
1474 CallInst::Create(TheFn, Ops, Ops + 5, "", MI);
1476 assert(isa<MemSetInst>(MI));
1478 EltPtr, MI->getArgOperand(1), // Dest, Value,
1479 ConstantInt::get(MI->getArgOperand(2)->getType(), EltSize), // Size
1481 ConstantInt::getFalse(MI->getContext()) // isVolatile
1483 const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() };
1484 Module *M = MI->getParent()->getParent()->getParent();
1485 TheFn = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2);
1486 CallInst::Create(TheFn, Ops, Ops + 5, "", MI);
1489 DeadInsts.push_back(MI);
1492 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that
1493 /// overwrites the entire allocation. Extract out the pieces of the stored
1494 /// integer and store them individually.
1495 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
1496 SmallVector<AllocaInst*, 32> &NewElts){
1497 // Extract each element out of the integer according to its structure offset
1498 // and store the element value to the individual alloca.
1499 Value *SrcVal = SI->getOperand(0);
1500 const Type *AllocaEltTy = AI->getAllocatedType();
1501 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
1503 // Handle tail padding by extending the operand
1504 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
1505 SrcVal = new ZExtInst(SrcVal,
1506 IntegerType::get(SI->getContext(), AllocaSizeBits),
1509 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI
1512 // There are two forms here: AI could be an array or struct. Both cases
1513 // have different ways to compute the element offset.
1514 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1515 const StructLayout *Layout = TD->getStructLayout(EltSTy);
1517 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1518 // Get the number of bits to shift SrcVal to get the value.
1519 const Type *FieldTy = EltSTy->getElementType(i);
1520 uint64_t Shift = Layout->getElementOffsetInBits(i);
1522 if (TD->isBigEndian())
1523 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
1525 Value *EltVal = SrcVal;
1527 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
1528 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
1529 "sroa.store.elt", SI);
1532 // Truncate down to an integer of the right size.
1533 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1535 // Ignore zero sized fields like {}, they obviously contain no data.
1536 if (FieldSizeBits == 0) continue;
1538 if (FieldSizeBits != AllocaSizeBits)
1539 EltVal = new TruncInst(EltVal,
1540 IntegerType::get(SI->getContext(), FieldSizeBits),
1542 Value *DestField = NewElts[i];
1543 if (EltVal->getType() == FieldTy) {
1544 // Storing to an integer field of this size, just do it.
1545 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) {
1546 // Bitcast to the right element type (for fp/vector values).
1547 EltVal = new BitCastInst(EltVal, FieldTy, "", SI);
1549 // Otherwise, bitcast the dest pointer (for aggregates).
1550 DestField = new BitCastInst(DestField,
1551 PointerType::getUnqual(EltVal->getType()),
1554 new StoreInst(EltVal, DestField, SI);
1558 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
1559 const Type *ArrayEltTy = ATy->getElementType();
1560 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
1561 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
1565 if (TD->isBigEndian())
1566 Shift = AllocaSizeBits-ElementOffset;
1570 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1571 // Ignore zero sized fields like {}, they obviously contain no data.
1572 if (ElementSizeBits == 0) continue;
1574 Value *EltVal = SrcVal;
1576 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
1577 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
1578 "sroa.store.elt", SI);
1581 // Truncate down to an integer of the right size.
1582 if (ElementSizeBits != AllocaSizeBits)
1583 EltVal = new TruncInst(EltVal,
1584 IntegerType::get(SI->getContext(),
1585 ElementSizeBits),"",SI);
1586 Value *DestField = NewElts[i];
1587 if (EltVal->getType() == ArrayEltTy) {
1588 // Storing to an integer field of this size, just do it.
1589 } else if (ArrayEltTy->isFloatingPointTy() ||
1590 ArrayEltTy->isVectorTy()) {
1591 // Bitcast to the right element type (for fp/vector values).
1592 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI);
1594 // Otherwise, bitcast the dest pointer (for aggregates).
1595 DestField = new BitCastInst(DestField,
1596 PointerType::getUnqual(EltVal->getType()),
1599 new StoreInst(EltVal, DestField, SI);
1601 if (TD->isBigEndian())
1602 Shift -= ElementOffset;
1604 Shift += ElementOffset;
1608 DeadInsts.push_back(SI);
1611 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to
1612 /// an integer. Load the individual pieces to form the aggregate value.
1613 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
1614 SmallVector<AllocaInst*, 32> &NewElts) {
1615 // Extract each element out of the NewElts according to its structure offset
1616 // and form the result value.
1617 const Type *AllocaEltTy = AI->getAllocatedType();
1618 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
1620 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
1623 // There are two forms here: AI could be an array or struct. Both cases
1624 // have different ways to compute the element offset.
1625 const StructLayout *Layout = 0;
1626 uint64_t ArrayEltBitOffset = 0;
1627 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1628 Layout = TD->getStructLayout(EltSTy);
1630 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
1631 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
1635 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
1637 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1638 // Load the value from the alloca. If the NewElt is an aggregate, cast
1639 // the pointer to an integer of the same size before doing the load.
1640 Value *SrcField = NewElts[i];
1641 const Type *FieldTy =
1642 cast<PointerType>(SrcField->getType())->getElementType();
1643 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1645 // Ignore zero sized fields like {}, they obviously contain no data.
1646 if (FieldSizeBits == 0) continue;
1648 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
1650 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() &&
1651 !FieldTy->isVectorTy())
1652 SrcField = new BitCastInst(SrcField,
1653 PointerType::getUnqual(FieldIntTy),
1655 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
1657 // If SrcField is a fp or vector of the right size but that isn't an
1658 // integer type, bitcast to an integer so we can shift it.
1659 if (SrcField->getType() != FieldIntTy)
1660 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
1662 // Zero extend the field to be the same size as the final alloca so that
1663 // we can shift and insert it.
1664 if (SrcField->getType() != ResultVal->getType())
1665 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
1667 // Determine the number of bits to shift SrcField.
1669 if (Layout) // Struct case.
1670 Shift = Layout->getElementOffsetInBits(i);
1672 Shift = i*ArrayEltBitOffset;
1674 if (TD->isBigEndian())
1675 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
1678 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
1679 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
1682 // Don't create an 'or x, 0' on the first iteration.
1683 if (!isa<Constant>(ResultVal) ||
1684 !cast<Constant>(ResultVal)->isNullValue())
1685 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
1687 ResultVal = SrcField;
1690 // Handle tail padding by truncating the result
1691 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
1692 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
1694 LI->replaceAllUsesWith(ResultVal);
1695 DeadInsts.push_back(LI);
1698 /// HasPadding - Return true if the specified type has any structure or
1699 /// alignment padding, false otherwise.
1700 static bool HasPadding(const Type *Ty, const TargetData &TD) {
1701 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty))
1702 return HasPadding(ATy->getElementType(), TD);
1704 if (const VectorType *VTy = dyn_cast<VectorType>(Ty))
1705 return HasPadding(VTy->getElementType(), TD);
1707 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1708 const StructLayout *SL = TD.getStructLayout(STy);
1709 unsigned PrevFieldBitOffset = 0;
1710 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1711 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
1713 // Padding in sub-elements?
1714 if (HasPadding(STy->getElementType(i), TD))
1717 // Check to see if there is any padding between this element and the
1720 unsigned PrevFieldEnd =
1721 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
1722 if (PrevFieldEnd < FieldBitOffset)
1726 PrevFieldBitOffset = FieldBitOffset;
1729 // Check for tail padding.
1730 if (unsigned EltCount = STy->getNumElements()) {
1731 unsigned PrevFieldEnd = PrevFieldBitOffset +
1732 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
1733 if (PrevFieldEnd < SL->getSizeInBits())
1738 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
1741 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
1742 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
1743 /// or 1 if safe after canonicalization has been performed.
1744 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) {
1745 // Loop over the use list of the alloca. We can only transform it if all of
1746 // the users are safe to transform.
1749 isSafeForScalarRepl(AI, AI, 0, Info);
1750 if (Info.isUnsafe) {
1751 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n');
1755 // Okay, we know all the users are promotable. If the aggregate is a memcpy
1756 // source and destination, we have to be careful. In particular, the memcpy
1757 // could be moving around elements that live in structure padding of the LLVM
1758 // types, but may actually be used. In these cases, we refuse to promote the
1760 if (Info.isMemCpySrc && Info.isMemCpyDst &&
1761 HasPadding(AI->getAllocatedType(), *TD))
1769 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
1770 /// some part of a constant global variable. This intentionally only accepts
1771 /// constant expressions because we don't can't rewrite arbitrary instructions.
1772 static bool PointsToConstantGlobal(Value *V) {
1773 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
1774 return GV->isConstant();
1775 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1776 if (CE->getOpcode() == Instruction::BitCast ||
1777 CE->getOpcode() == Instruction::GetElementPtr)
1778 return PointsToConstantGlobal(CE->getOperand(0));
1782 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
1783 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
1784 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
1785 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
1786 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
1787 /// the alloca, and if the source pointer is a pointer to a constant global, we
1788 /// can optimize this.
1789 static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
1791 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1792 User *U = cast<Instruction>(*UI);
1794 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1795 // Ignore non-volatile loads, they are always ok.
1796 if (LI->isVolatile()) return false;
1800 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
1801 // If uses of the bitcast are ok, we are ok.
1802 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
1806 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
1807 // If the GEP has all zero indices, it doesn't offset the pointer. If it
1808 // doesn't, it does.
1809 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
1810 isOffset || !GEP->hasAllZeroIndices()))
1815 if (CallSite CS = U) {
1816 // If this is a readonly/readnone call site, then we know it is just a
1817 // load and we can ignore it.
1818 if (CS.onlyReadsMemory())
1821 // If this is the function being called then we treat it like a load and
1823 if (CS.isCallee(UI))
1826 // If this is being passed as a byval argument, the caller is making a
1827 // copy, so it is only a read of the alloca.
1828 unsigned ArgNo = CS.getArgumentNo(UI);
1829 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal))
1833 // If this is isn't our memcpy/memmove, reject it as something we can't
1835 MemTransferInst *MI = dyn_cast<MemTransferInst>(U);
1839 // If the transfer is using the alloca as a source of the transfer, then
1840 // ignore it since it is a load (unless the transfer is volatile).
1841 if (UI.getOperandNo() == 1) {
1842 if (MI->isVolatile()) return false;
1846 // If we already have seen a copy, reject the second one.
1847 if (TheCopy) return false;
1849 // If the pointer has been offset from the start of the alloca, we can't
1850 // safely handle this.
1851 if (isOffset) return false;
1853 // If the memintrinsic isn't using the alloca as the dest, reject it.
1854 if (UI.getOperandNo() != 0) return false;
1856 // If the source of the memcpy/move is not a constant global, reject it.
1857 if (!PointsToConstantGlobal(MI->getSource()))
1860 // Otherwise, the transform is safe. Remember the copy instruction.
1866 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
1867 /// modified by a copy from a constant global. If we can prove this, we can
1868 /// replace any uses of the alloca with uses of the global directly.
1869 MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) {
1870 MemTransferInst *TheCopy = 0;
1871 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))