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/DominanceFrontier.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/Transforms/Utils/SSAUpdater.h"
39 #include "llvm/Support/CallSite.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/ErrorHandling.h"
42 #include "llvm/Support/GetElementPtrTypeIterator.h"
43 #include "llvm/Support/IRBuilder.h"
44 #include "llvm/Support/MathExtras.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/ADT/SmallVector.h"
47 #include "llvm/ADT/Statistic.h"
50 STATISTIC(NumReplaced, "Number of allocas broken up");
51 STATISTIC(NumPromoted, "Number of allocas promoted");
52 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
53 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
56 UsePromoteMemToReg = 1
60 struct SROA : public FunctionPass {
61 static char ID; // Pass identification, replacement for typeid
62 explicit SROA(signed T = -1) : FunctionPass(ID) {
63 initializeSROAPass(*PassRegistry::getPassRegistry());
70 bool runOnFunction(Function &F);
72 bool performScalarRepl(Function &F);
73 bool performPromotion(Function &F);
75 // getAnalysisUsage - This pass does not require any passes, but we know it
76 // will not alter the CFG, so say so.
77 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
78 if (UsePromoteMemToReg) {
79 AU.addRequired<DominatorTree>();
80 AU.addRequired<DominanceFrontier>();
88 /// DeadInsts - Keep track of instructions we have made dead, so that
89 /// we can remove them after we are done working.
90 SmallVector<Value*, 32> DeadInsts;
92 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
93 /// information about the uses. All these fields are initialized to false
94 /// and set to true when something is learned.
96 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
99 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
100 bool isMemCpySrc : 1;
102 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
103 bool isMemCpyDst : 1;
106 : isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false) {}
109 unsigned SRThreshold;
111 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; }
113 bool isSafeAllocaToScalarRepl(AllocaInst *AI);
115 void isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
117 void isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t &Offset,
119 void isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize,
120 const Type *MemOpType, bool isStore, AllocaInfo &Info);
121 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size);
122 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset,
125 void DoScalarReplacement(AllocaInst *AI,
126 std::vector<AllocaInst*> &WorkList);
127 void DeleteDeadInstructions();
129 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
130 SmallVector<AllocaInst*, 32> &NewElts);
131 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
132 SmallVector<AllocaInst*, 32> &NewElts);
133 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
134 SmallVector<AllocaInst*, 32> &NewElts);
135 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
137 SmallVector<AllocaInst*, 32> &NewElts);
138 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
139 SmallVector<AllocaInst*, 32> &NewElts);
140 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
141 SmallVector<AllocaInst*, 32> &NewElts);
143 static MemTransferInst *isOnlyCopiedFromConstantGlobal(AllocaInst *AI);
148 INITIALIZE_PASS_BEGIN(SROA, "scalarrepl",
149 "Scalar Replacement of Aggregates", false, false)
150 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
151 INITIALIZE_PASS_DEPENDENCY(DominanceFrontier)
152 INITIALIZE_PASS_END(SROA, "scalarrepl",
153 "Scalar Replacement of Aggregates", false, false)
155 // Public interface to the ScalarReplAggregates pass
156 FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) {
157 return new SROA(Threshold);
161 //===----------------------------------------------------------------------===//
162 // Convert To Scalar Optimization.
163 //===----------------------------------------------------------------------===//
166 /// ConvertToScalarInfo - This class implements the "Convert To Scalar"
167 /// optimization, which scans the uses of an alloca and determines if it can
168 /// rewrite it in terms of a single new alloca that can be mem2reg'd.
169 class ConvertToScalarInfo {
170 /// AllocaSize - The size of the alloca being considered.
172 const TargetData &TD;
174 /// IsNotTrivial - This is set to true if there is some access to the object
175 /// which means that mem2reg can't promote it.
178 /// VectorTy - This tracks the type that we should promote the vector to if
179 /// it is possible to turn it into a vector. This starts out null, and if it
180 /// isn't possible to turn into a vector type, it gets set to VoidTy.
181 const Type *VectorTy;
183 /// HadAVector - True if there is at least one vector access to the alloca.
184 /// We don't want to turn random arrays into vectors and use vector element
185 /// insert/extract, but if there are element accesses to something that is
186 /// also declared as a vector, we do want to promote to a vector.
190 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td)
191 : AllocaSize(Size), TD(td) {
192 IsNotTrivial = false;
197 AllocaInst *TryConvert(AllocaInst *AI);
200 bool CanConvertToScalar(Value *V, uint64_t Offset);
201 void MergeInType(const Type *In, uint64_t Offset);
202 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
204 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
205 uint64_t Offset, IRBuilder<> &Builder);
206 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
207 uint64_t Offset, IRBuilder<> &Builder);
209 } // end anonymous namespace.
212 /// IsVerbotenVectorType - Return true if this is a vector type ScalarRepl isn't
213 /// allowed to form. We do this to avoid MMX types, which is a complete hack,
214 /// but is required until the backend is fixed.
215 static bool IsVerbotenVectorType(const VectorType *VTy, const Instruction *I) {
216 StringRef Triple(I->getParent()->getParent()->getParent()->getTargetTriple());
217 if (!Triple.startswith("i386") &&
218 !Triple.startswith("x86_64"))
221 // Reject all the MMX vector types.
222 switch (VTy->getNumElements()) {
223 default: return false;
224 case 1: return VTy->getElementType()->isIntegerTy(64);
225 case 2: return VTy->getElementType()->isIntegerTy(32);
226 case 4: return VTy->getElementType()->isIntegerTy(16);
227 case 8: return VTy->getElementType()->isIntegerTy(8);
232 /// TryConvert - Analyze the specified alloca, and if it is safe to do so,
233 /// rewrite it to be a new alloca which is mem2reg'able. This returns the new
234 /// alloca if possible or null if not.
235 AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
236 // If we can't convert this scalar, or if mem2reg can trivially do it, bail
238 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial)
241 // If we were able to find a vector type that can handle this with
242 // insert/extract elements, and if there was at least one use that had
243 // a vector type, promote this to a vector. We don't want to promote
244 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
245 // we just get a lot of insert/extracts. If at least one vector is
246 // involved, then we probably really do have a union of vector/array.
248 if (VectorTy && VectorTy->isVectorTy() && HadAVector &&
249 !IsVerbotenVectorType(cast<VectorType>(VectorTy), AI)) {
250 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = "
251 << *VectorTy << '\n');
252 NewTy = VectorTy; // Use the vector type.
254 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
255 // Create and insert the integer alloca.
256 NewTy = IntegerType::get(AI->getContext(), AllocaSize*8);
258 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
259 ConvertUsesToScalar(AI, NewAI, 0);
263 /// MergeInType - Add the 'In' type to the accumulated vector type (VectorTy)
264 /// so far at the offset specified by Offset (which is specified in bytes).
266 /// There are two cases we handle here:
267 /// 1) A union of vector types of the same size and potentially its elements.
268 /// Here we turn element accesses into insert/extract element operations.
269 /// This promotes a <4 x float> with a store of float to the third element
270 /// into a <4 x float> that uses insert element.
271 /// 2) A fully general blob of memory, which we turn into some (potentially
272 /// large) integer type with extract and insert operations where the loads
273 /// and stores would mutate the memory. We mark this by setting VectorTy
275 void ConvertToScalarInfo::MergeInType(const Type *In, uint64_t Offset) {
276 // If we already decided to turn this into a blob of integer memory, there is
277 // nothing to be done.
278 if (VectorTy && VectorTy->isVoidTy())
281 // If this could be contributing to a vector, analyze it.
283 // If the In type is a vector that is the same size as the alloca, see if it
284 // matches the existing VecTy.
285 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
286 // Remember if we saw a vector type.
289 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
290 // If we're storing/loading a vector of the right size, allow it as a
291 // vector. If this the first vector we see, remember the type so that
292 // we know the element size. If this is a subsequent access, ignore it
293 // even if it is a differing type but the same size. Worst case we can
294 // bitcast the resultant vectors.
299 } else if (In->isFloatTy() || In->isDoubleTy() ||
300 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 &&
301 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
302 // If we're accessing something that could be an element of a vector, see
303 // if the implied vector agrees with what we already have and if Offset is
304 // compatible with it.
305 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
306 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 &&
308 cast<VectorType>(VectorTy)->getElementType()
309 ->getPrimitiveSizeInBits()/8 == EltSize)) {
311 VectorTy = VectorType::get(In, AllocaSize/EltSize);
316 // Otherwise, we have a case that we can't handle with an optimized vector
317 // form. We can still turn this into a large integer.
318 VectorTy = Type::getVoidTy(In->getContext());
321 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
322 /// its accesses to a single vector type, return true and set VecTy to
323 /// the new type. If we could convert the alloca into a single promotable
324 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
325 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
326 /// is the current offset from the base of the alloca being analyzed.
328 /// If we see at least one access to the value that is as a vector type, set the
330 bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) {
331 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
332 Instruction *User = cast<Instruction>(*UI);
334 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
335 // Don't break volatile loads.
336 if (LI->isVolatile())
338 // Don't touch MMX operations.
339 if (LI->getType()->isX86_MMXTy())
341 MergeInType(LI->getType(), Offset);
345 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
346 // Storing the pointer, not into the value?
347 if (SI->getOperand(0) == V || SI->isVolatile()) return false;
348 // Don't touch MMX operations.
349 if (SI->getOperand(0)->getType()->isX86_MMXTy())
351 MergeInType(SI->getOperand(0)->getType(), Offset);
355 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
356 IsNotTrivial = true; // Can't be mem2reg'd.
357 if (!CanConvertToScalar(BCI, Offset))
362 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
363 // If this is a GEP with a variable indices, we can't handle it.
364 if (!GEP->hasAllConstantIndices())
367 // Compute the offset that this GEP adds to the pointer.
368 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
369 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
370 &Indices[0], Indices.size());
371 // See if all uses can be converted.
372 if (!CanConvertToScalar(GEP, Offset+GEPOffset))
374 IsNotTrivial = true; // Can't be mem2reg'd.
378 // If this is a constant sized memset of a constant value (e.g. 0) we can
380 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
381 // Store of constant value and constant size.
382 if (!isa<ConstantInt>(MSI->getValue()) ||
383 !isa<ConstantInt>(MSI->getLength()))
385 IsNotTrivial = true; // Can't be mem2reg'd.
389 // If this is a memcpy or memmove into or out of the whole allocation, we
390 // can handle it like a load or store of the scalar type.
391 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
392 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength());
393 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0)
396 IsNotTrivial = true; // Can't be mem2reg'd.
400 // Otherwise, we cannot handle this!
407 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
408 /// directly. This happens when we are converting an "integer union" to a
409 /// single integer scalar, or when we are converting a "vector union" to a
410 /// vector with insert/extractelement instructions.
412 /// Offset is an offset from the original alloca, in bits that need to be
413 /// shifted to the right. By the end of this, there should be no uses of Ptr.
414 void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI,
416 while (!Ptr->use_empty()) {
417 Instruction *User = cast<Instruction>(Ptr->use_back());
419 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
420 ConvertUsesToScalar(CI, NewAI, Offset);
421 CI->eraseFromParent();
425 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
426 // Compute the offset that this GEP adds to the pointer.
427 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
428 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
429 &Indices[0], Indices.size());
430 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
431 GEP->eraseFromParent();
435 IRBuilder<> Builder(User);
437 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
438 // The load is a bit extract from NewAI shifted right by Offset bits.
439 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
441 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
442 LI->replaceAllUsesWith(NewLoadVal);
443 LI->eraseFromParent();
447 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
448 assert(SI->getOperand(0) != Ptr && "Consistency error!");
449 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
450 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
452 Builder.CreateStore(New, NewAI);
453 SI->eraseFromParent();
455 // If the load we just inserted is now dead, then the inserted store
456 // overwrote the entire thing.
457 if (Old->use_empty())
458 Old->eraseFromParent();
462 // If this is a constant sized memset of a constant value (e.g. 0) we can
463 // transform it into a store of the expanded constant value.
464 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
465 assert(MSI->getRawDest() == Ptr && "Consistency error!");
466 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
468 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
470 // Compute the value replicated the right number of times.
471 APInt APVal(NumBytes*8, Val);
473 // Splat the value if non-zero.
475 for (unsigned i = 1; i != NumBytes; ++i)
478 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
479 Value *New = ConvertScalar_InsertValue(
480 ConstantInt::get(User->getContext(), APVal),
481 Old, Offset, Builder);
482 Builder.CreateStore(New, NewAI);
484 // If the load we just inserted is now dead, then the memset overwrote
486 if (Old->use_empty())
487 Old->eraseFromParent();
489 MSI->eraseFromParent();
493 // If this is a memcpy or memmove into or out of the whole allocation, we
494 // can handle it like a load or store of the scalar type.
495 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
496 assert(Offset == 0 && "must be store to start of alloca");
498 // If the source and destination are both to the same alloca, then this is
499 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
501 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, 0));
503 if (GetUnderlyingObject(MTI->getSource(), 0) != OrigAI) {
504 // Dest must be OrigAI, change this to be a load from the original
505 // pointer (bitcasted), then a store to our new alloca.
506 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
507 Value *SrcPtr = MTI->getSource();
508 const PointerType* SPTy = cast<PointerType>(SrcPtr->getType());
509 const PointerType* AIPTy = cast<PointerType>(NewAI->getType());
510 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
511 AIPTy = PointerType::get(AIPTy->getElementType(),
512 SPTy->getAddressSpace());
514 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy);
516 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
517 SrcVal->setAlignment(MTI->getAlignment());
518 Builder.CreateStore(SrcVal, NewAI);
519 } else if (GetUnderlyingObject(MTI->getDest(), 0) != OrigAI) {
520 // Src must be OrigAI, change this to be a load from NewAI then a store
521 // through the original dest pointer (bitcasted).
522 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
523 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
525 const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType());
526 const PointerType* AIPTy = cast<PointerType>(NewAI->getType());
527 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
528 AIPTy = PointerType::get(AIPTy->getElementType(),
529 DPTy->getAddressSpace());
531 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy);
533 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
534 NewStore->setAlignment(MTI->getAlignment());
536 // Noop transfer. Src == Dst
539 MTI->eraseFromParent();
543 llvm_unreachable("Unsupported operation!");
547 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
548 /// or vector value FromVal, extracting the bits from the offset specified by
549 /// Offset. This returns the value, which is of type ToType.
551 /// This happens when we are converting an "integer union" to a single
552 /// integer scalar, or when we are converting a "vector union" to a vector with
553 /// insert/extractelement instructions.
555 /// Offset is an offset from the original alloca, in bits that need to be
556 /// shifted to the right.
557 Value *ConvertToScalarInfo::
558 ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
559 uint64_t Offset, IRBuilder<> &Builder) {
560 // If the load is of the whole new alloca, no conversion is needed.
561 if (FromVal->getType() == ToType && Offset == 0)
564 // If the result alloca is a vector type, this is either an element
565 // access or a bitcast to another vector type of the same size.
566 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) {
567 if (ToType->isVectorTy())
568 return Builder.CreateBitCast(FromVal, ToType, "tmp");
570 // Otherwise it must be an element access.
573 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
574 Elt = Offset/EltSize;
575 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
577 // Return the element extracted out of it.
578 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get(
579 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp");
580 if (V->getType() != ToType)
581 V = Builder.CreateBitCast(V, ToType, "tmp");
585 // If ToType is a first class aggregate, extract out each of the pieces and
586 // use insertvalue's to form the FCA.
587 if (const StructType *ST = dyn_cast<StructType>(ToType)) {
588 const StructLayout &Layout = *TD.getStructLayout(ST);
589 Value *Res = UndefValue::get(ST);
590 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
591 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
592 Offset+Layout.getElementOffsetInBits(i),
594 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
599 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
600 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
601 Value *Res = UndefValue::get(AT);
602 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
603 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
604 Offset+i*EltSize, Builder);
605 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
610 // Otherwise, this must be a union that was converted to an integer value.
611 const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
613 // If this is a big-endian system and the load is narrower than the
614 // full alloca type, we need to do a shift to get the right bits.
616 if (TD.isBigEndian()) {
617 // On big-endian machines, the lowest bit is stored at the bit offset
618 // from the pointer given by getTypeStoreSizeInBits. This matters for
619 // integers with a bitwidth that is not a multiple of 8.
620 ShAmt = TD.getTypeStoreSizeInBits(NTy) -
621 TD.getTypeStoreSizeInBits(ToType) - Offset;
626 // Note: we support negative bitwidths (with shl) which are not defined.
627 // We do this to support (f.e.) loads off the end of a structure where
628 // only some bits are used.
629 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
630 FromVal = Builder.CreateLShr(FromVal,
631 ConstantInt::get(FromVal->getType(),
633 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
634 FromVal = Builder.CreateShl(FromVal,
635 ConstantInt::get(FromVal->getType(),
638 // Finally, unconditionally truncate the integer to the right width.
639 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType);
640 if (LIBitWidth < NTy->getBitWidth())
642 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
644 else if (LIBitWidth > NTy->getBitWidth())
646 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
649 // If the result is an integer, this is a trunc or bitcast.
650 if (ToType->isIntegerTy()) {
652 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) {
653 // Just do a bitcast, we know the sizes match up.
654 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
656 // Otherwise must be a pointer.
657 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
659 assert(FromVal->getType() == ToType && "Didn't convert right?");
663 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
664 /// or vector value "Old" at the offset specified by Offset.
666 /// This happens when we are converting an "integer union" to a
667 /// single integer scalar, or when we are converting a "vector union" to a
668 /// vector with insert/extractelement instructions.
670 /// Offset is an offset from the original alloca, in bits that need to be
671 /// shifted to the right.
672 Value *ConvertToScalarInfo::
673 ConvertScalar_InsertValue(Value *SV, Value *Old,
674 uint64_t Offset, IRBuilder<> &Builder) {
675 // Convert the stored type to the actual type, shift it left to insert
676 // then 'or' into place.
677 const Type *AllocaType = Old->getType();
678 LLVMContext &Context = Old->getContext();
680 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
681 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy);
682 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType());
684 // Changing the whole vector with memset or with an access of a different
686 if (ValSize == VecSize)
687 return Builder.CreateBitCast(SV, AllocaType, "tmp");
689 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
691 // Must be an element insertion.
692 unsigned Elt = Offset/EltSize;
694 if (SV->getType() != VTy->getElementType())
695 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp");
697 SV = Builder.CreateInsertElement(Old, SV,
698 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt),
703 // If SV is a first-class aggregate value, insert each value recursively.
704 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
705 const StructLayout &Layout = *TD.getStructLayout(ST);
706 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
707 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
708 Old = ConvertScalar_InsertValue(Elt, Old,
709 Offset+Layout.getElementOffsetInBits(i),
715 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
716 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
717 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
718 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
719 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
724 // If SV is a float, convert it to the appropriate integer type.
725 // If it is a pointer, do the same.
726 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType());
727 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType);
728 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType());
729 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType);
730 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
731 SV = Builder.CreateBitCast(SV,
732 IntegerType::get(SV->getContext(),SrcWidth), "tmp");
733 else if (SV->getType()->isPointerTy())
734 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp");
736 // Zero extend or truncate the value if needed.
737 if (SV->getType() != AllocaType) {
738 if (SV->getType()->getPrimitiveSizeInBits() <
739 AllocaType->getPrimitiveSizeInBits())
740 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
742 // Truncation may be needed if storing more than the alloca can hold
743 // (undefined behavior).
744 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
745 SrcWidth = DestWidth;
746 SrcStoreWidth = DestStoreWidth;
750 // If this is a big-endian system and the store is narrower than the
751 // full alloca type, we need to do a shift to get the right bits.
753 if (TD.isBigEndian()) {
754 // On big-endian machines, the lowest bit is stored at the bit offset
755 // from the pointer given by getTypeStoreSizeInBits. This matters for
756 // integers with a bitwidth that is not a multiple of 8.
757 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
762 // Note: we support negative bitwidths (with shr) which are not defined.
763 // We do this to support (f.e.) stores off the end of a structure where
764 // only some bits in the structure are set.
765 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
766 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
767 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(),
770 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
771 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(),
773 Mask = Mask.lshr(-ShAmt);
776 // Mask out the bits we are about to insert from the old value, and or
778 if (SrcWidth != DestWidth) {
779 assert(DestWidth > SrcWidth);
780 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
781 SV = Builder.CreateOr(Old, SV, "ins");
787 //===----------------------------------------------------------------------===//
789 //===----------------------------------------------------------------------===//
792 bool SROA::runOnFunction(Function &F) {
793 TD = getAnalysisIfAvailable<TargetData>();
795 bool Changed = performPromotion(F);
797 // FIXME: ScalarRepl currently depends on TargetData more than it
798 // theoretically needs to. It should be refactored in order to support
799 // target-independent IR. Until this is done, just skip the actual
800 // scalar-replacement portion of this pass.
801 if (!TD) return Changed;
804 bool LocalChange = performScalarRepl(F);
805 if (!LocalChange) break; // No need to repromote if no scalarrepl
807 LocalChange = performPromotion(F);
808 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
814 /// PromoteAlloca - Promote an alloca to registers, using SSAUpdater.
815 static void PromoteAlloca(AllocaInst *AI, SSAUpdater &SSA) {
816 SSA.Initialize(AI->getType()->getElementType(), AI->getName());
818 // First step: bucket up uses of the alloca by the block they occur in.
819 // This is important because we have to handle multiple defs/uses in a block
820 // ourselves: SSAUpdater is purely for cross-block references.
821 // FIXME: Want a TinyVector<Instruction*> since there is often 0/1 element.
822 DenseMap<BasicBlock*, std::vector<Instruction*> > UsesByBlock;
824 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
826 Instruction *User = cast<Instruction>(*UI);
827 UsesByBlock[User->getParent()].push_back(User);
830 // Okay, now we can iterate over all the blocks in the function with uses,
831 // processing them. Keep track of which loads are loading a live-in value.
832 // Walk the uses in the use-list order to be determinstic.
833 SmallVector<LoadInst*, 32> LiveInLoads;
834 DenseMap<Value*, Value*> ReplacedLoads;
836 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
838 Instruction *User = cast<Instruction>(*UI);
839 BasicBlock *BB = User->getParent();
840 std::vector<Instruction*> &BlockUses = UsesByBlock[BB];
842 // If this block has already been processed, ignore this repeat use.
843 if (BlockUses.empty()) continue;
845 // Okay, this is the first use in the block. If this block just has a
846 // single user in it, we can rewrite it trivially.
847 if (BlockUses.size() == 1) {
848 // If it is a store, it is a trivial def of the value in the block.
849 if (StoreInst *SI = dyn_cast<StoreInst>(User))
850 SSA.AddAvailableValue(BB, SI->getOperand(0));
852 // Otherwise it is a load, queue it to rewrite as a live-in load.
853 LiveInLoads.push_back(cast<LoadInst>(User));
858 // Otherwise, check to see if this block is all loads.
859 bool HasStore = false;
860 for (unsigned i = 0, e = BlockUses.size(); i != e; ++i) {
861 if (isa<StoreInst>(BlockUses[i])) {
867 // If so, we can queue them all as live in loads. We don't have an
868 // efficient way to tell which on is first in the block and don't want to
869 // scan large blocks, so just add all loads as live ins.
871 for (unsigned i = 0, e = BlockUses.size(); i != e; ++i)
872 LiveInLoads.push_back(cast<LoadInst>(BlockUses[i]));
877 // Otherwise, we have mixed loads and stores (or just a bunch of stores).
878 // Since SSAUpdater is purely for cross-block values, we need to determine
879 // the order of these instructions in the block. If the first use in the
880 // block is a load, then it uses the live in value. The last store defines
881 // the live out value. We handle this by doing a linear scan of the block.
882 Value *StoredValue = 0;
883 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ++II) {
884 if (LoadInst *L = dyn_cast<LoadInst>(II)) {
885 // If this is a load from an unrelated pointer, ignore it.
886 if (L->getOperand(0) != AI) continue;
888 // If we haven't seen a store yet, this is a live in use, otherwise
889 // use the stored value.
891 L->replaceAllUsesWith(StoredValue);
892 ReplacedLoads[L] = StoredValue;
894 LiveInLoads.push_back(L);
899 if (StoreInst *S = dyn_cast<StoreInst>(II)) {
900 // If this is a store to an unrelated pointer, ignore it.
901 if (S->getPointerOperand() != AI) continue;
903 // Remember that this is the active value in the block.
904 StoredValue = S->getOperand(0);
908 // The last stored value that happened is the live-out for the block.
909 assert(StoredValue && "Already checked that there is a store in block");
910 SSA.AddAvailableValue(BB, StoredValue);
914 // Okay, now we rewrite all loads that use live-in values in the loop,
915 // inserting PHI nodes as necessary.
916 for (unsigned i = 0, e = LiveInLoads.size(); i != e; ++i) {
917 LoadInst *ALoad = LiveInLoads[i];
918 Value *NewVal = SSA.GetValueInMiddleOfBlock(ALoad->getParent());
919 ALoad->replaceAllUsesWith(NewVal);
920 ReplacedLoads[ALoad] = NewVal;
923 // Now that everything is rewritten, delete the old instructions from the
924 // function. They should all be dead now.
925 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; ) {
926 Instruction *User = cast<Instruction>(*UI++);
928 // If this is a load that still has uses, then the load must have been added
929 // as a live value in the SSAUpdate data structure for a block (e.g. because
930 // the loaded value was stored later). In this case, we need to recursively
931 // propagate the updates until we get to the real value.
932 if (!User->use_empty()) {
933 Value *NewVal = ReplacedLoads[User];
934 assert(NewVal && "not a replaced load?");
936 // Propagate down to the ultimate replacee. The intermediately loads
937 // could theoretically already have been deleted, so we don't want to
938 // dereference the Value*'s.
939 DenseMap<Value*, Value*>::iterator RLI = ReplacedLoads.find(NewVal);
940 while (RLI != ReplacedLoads.end()) {
941 NewVal = RLI->second;
942 RLI = ReplacedLoads.find(NewVal);
945 User->replaceAllUsesWith(NewVal);
948 User->eraseFromParent();
953 bool SROA::performPromotion(Function &F) {
954 std::vector<AllocaInst*> Allocas;
955 DominatorTree *DT = 0;
956 DominanceFrontier *DF = 0;
957 if (UsePromoteMemToReg) {
958 DT = &getAnalysis<DominatorTree>();
959 DF = &getAnalysis<DominanceFrontier>();
962 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
964 bool Changed = false;
969 // Find allocas that are safe to promote, by looking at all instructions in
971 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
972 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
973 if (isAllocaPromotable(AI))
974 Allocas.push_back(AI);
976 if (Allocas.empty()) break;
978 if (UsePromoteMemToReg)
979 PromoteMemToReg(Allocas, *DT, *DF);
982 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
983 PromoteAlloca(Allocas[i], SSA);
984 Allocas[i]->eraseFromParent();
987 NumPromoted += Allocas.size();
995 /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for
996 /// SROA. It must be a struct or array type with a small number of elements.
997 static bool ShouldAttemptScalarRepl(AllocaInst *AI) {
998 const Type *T = AI->getAllocatedType();
999 // Do not promote any struct into more than 32 separate vars.
1000 if (const StructType *ST = dyn_cast<StructType>(T))
1001 return ST->getNumElements() <= 32;
1002 // Arrays are much less likely to be safe for SROA; only consider
1003 // them if they are very small.
1004 if (const ArrayType *AT = dyn_cast<ArrayType>(T))
1005 return AT->getNumElements() <= 8;
1010 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
1011 // which runs on all of the malloc/alloca instructions in the function, removing
1012 // them if they are only used by getelementptr instructions.
1014 bool SROA::performScalarRepl(Function &F) {
1015 std::vector<AllocaInst*> WorkList;
1017 // Scan the entry basic block, adding allocas to the worklist.
1018 BasicBlock &BB = F.getEntryBlock();
1019 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
1020 if (AllocaInst *A = dyn_cast<AllocaInst>(I))
1021 WorkList.push_back(A);
1023 // Process the worklist
1024 bool Changed = false;
1025 while (!WorkList.empty()) {
1026 AllocaInst *AI = WorkList.back();
1027 WorkList.pop_back();
1029 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
1030 // with unused elements.
1031 if (AI->use_empty()) {
1032 AI->eraseFromParent();
1037 // If this alloca is impossible for us to promote, reject it early.
1038 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
1041 // Check to see if this allocation is only modified by a memcpy/memmove from
1042 // a constant global. If this is the case, we can change all users to use
1043 // the constant global instead. This is commonly produced by the CFE by
1044 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
1045 // is only subsequently read.
1046 if (MemTransferInst *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
1047 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n');
1048 DEBUG(dbgs() << " memcpy = " << *TheCopy << '\n');
1049 Constant *TheSrc = cast<Constant>(TheCopy->getSource());
1050 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
1051 TheCopy->eraseFromParent(); // Don't mutate the global.
1052 AI->eraseFromParent();
1058 // Check to see if we can perform the core SROA transformation. We cannot
1059 // transform the allocation instruction if it is an array allocation
1060 // (allocations OF arrays are ok though), and an allocation of a scalar
1061 // value cannot be decomposed at all.
1062 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
1064 // Do not promote [0 x %struct].
1065 if (AllocaSize == 0) continue;
1067 // Do not promote any struct whose size is too big.
1068 if (AllocaSize > SRThreshold) continue;
1070 // If the alloca looks like a good candidate for scalar replacement, and if
1071 // all its users can be transformed, then split up the aggregate into its
1072 // separate elements.
1073 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) {
1074 DoScalarReplacement(AI, WorkList);
1079 // If we can turn this aggregate value (potentially with casts) into a
1080 // simple scalar value that can be mem2reg'd into a register value.
1081 // IsNotTrivial tracks whether this is something that mem2reg could have
1082 // promoted itself. If so, we don't want to transform it needlessly. Note
1083 // that we can't just check based on the type: the alloca may be of an i32
1084 // but that has pointer arithmetic to set byte 3 of it or something.
1085 if (AllocaInst *NewAI =
1086 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) {
1087 NewAI->takeName(AI);
1088 AI->eraseFromParent();
1094 // Otherwise, couldn't process this alloca.
1100 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
1101 /// predicate, do SROA now.
1102 void SROA::DoScalarReplacement(AllocaInst *AI,
1103 std::vector<AllocaInst*> &WorkList) {
1104 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n');
1105 SmallVector<AllocaInst*, 32> ElementAllocas;
1106 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
1107 ElementAllocas.reserve(ST->getNumContainedTypes());
1108 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
1109 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
1111 AI->getName() + "." + Twine(i), AI);
1112 ElementAllocas.push_back(NA);
1113 WorkList.push_back(NA); // Add to worklist for recursive processing
1116 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
1117 ElementAllocas.reserve(AT->getNumElements());
1118 const Type *ElTy = AT->getElementType();
1119 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1120 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
1121 AI->getName() + "." + Twine(i), AI);
1122 ElementAllocas.push_back(NA);
1123 WorkList.push_back(NA); // Add to worklist for recursive processing
1127 // Now that we have created the new alloca instructions, rewrite all the
1128 // uses of the old alloca.
1129 RewriteForScalarRepl(AI, AI, 0, ElementAllocas);
1131 // Now erase any instructions that were made dead while rewriting the alloca.
1132 DeleteDeadInstructions();
1133 AI->eraseFromParent();
1138 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list,
1139 /// recursively including all their operands that become trivially dead.
1140 void SROA::DeleteDeadInstructions() {
1141 while (!DeadInsts.empty()) {
1142 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val());
1144 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
1145 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
1146 // Zero out the operand and see if it becomes trivially dead.
1147 // (But, don't add allocas to the dead instruction list -- they are
1148 // already on the worklist and will be deleted separately.)
1150 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U))
1151 DeadInsts.push_back(U);
1154 I->eraseFromParent();
1158 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to
1159 /// performing scalar replacement of alloca AI. The results are flagged in
1160 /// the Info parameter. Offset indicates the position within AI that is
1161 /// referenced by this instruction.
1162 void SROA::isSafeForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
1164 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1165 Instruction *User = cast<Instruction>(*UI);
1167 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1168 isSafeForScalarRepl(BC, AI, Offset, Info);
1169 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1170 uint64_t GEPOffset = Offset;
1171 isSafeGEP(GEPI, AI, GEPOffset, Info);
1173 isSafeForScalarRepl(GEPI, AI, GEPOffset, Info);
1174 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1175 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1177 isSafeMemAccess(AI, Offset, Length->getZExtValue(), 0,
1178 UI.getOperandNo() == 0, Info);
1181 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1182 if (!LI->isVolatile()) {
1183 const Type *LIType = LI->getType();
1184 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(LIType),
1185 LIType, false, Info);
1188 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1189 // Store is ok if storing INTO the pointer, not storing the pointer
1190 if (!SI->isVolatile() && SI->getOperand(0) != I) {
1191 const Type *SIType = SI->getOperand(0)->getType();
1192 isSafeMemAccess(AI, Offset, TD->getTypeAllocSize(SIType),
1193 SIType, true, Info);
1197 DEBUG(errs() << " Transformation preventing inst: " << *User << '\n');
1200 if (Info.isUnsafe) return;
1204 /// isSafeGEP - Check if a GEP instruction can be handled for scalar
1205 /// replacement. It is safe when all the indices are constant, in-bounds
1206 /// references, and when the resulting offset corresponds to an element within
1207 /// the alloca type. The results are flagged in the Info parameter. Upon
1208 /// return, Offset is adjusted as specified by the GEP indices.
1209 void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI,
1210 uint64_t &Offset, AllocaInfo &Info) {
1211 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
1215 // Walk through the GEP type indices, checking the types that this indexes
1217 for (; GEPIt != E; ++GEPIt) {
1218 // Ignore struct elements, no extra checking needed for these.
1219 if ((*GEPIt)->isStructTy())
1222 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
1224 return MarkUnsafe(Info);
1227 // Compute the offset due to this GEP and check if the alloca has a
1228 // component element at that offset.
1229 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1230 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
1231 &Indices[0], Indices.size());
1232 if (!TypeHasComponent(AI->getAllocatedType(), Offset, 0))
1236 /// isHomogeneousAggregate - Check if type T is a struct or array containing
1237 /// elements of the same type (which is always true for arrays). If so,
1238 /// return true with NumElts and EltTy set to the number of elements and the
1239 /// element type, respectively.
1240 static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts,
1241 const Type *&EltTy) {
1242 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
1243 NumElts = AT->getNumElements();
1244 EltTy = (NumElts == 0 ? 0 : AT->getElementType());
1247 if (const StructType *ST = dyn_cast<StructType>(T)) {
1248 NumElts = ST->getNumContainedTypes();
1249 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0));
1250 for (unsigned n = 1; n < NumElts; ++n) {
1251 if (ST->getContainedType(n) != EltTy)
1259 /// isCompatibleAggregate - Check if T1 and T2 are either the same type or are
1260 /// "homogeneous" aggregates with the same element type and number of elements.
1261 static bool isCompatibleAggregate(const Type *T1, const Type *T2) {
1265 unsigned NumElts1, NumElts2;
1266 const Type *EltTy1, *EltTy2;
1267 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) &&
1268 isHomogeneousAggregate(T2, NumElts2, EltTy2) &&
1269 NumElts1 == NumElts2 &&
1276 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI
1277 /// alloca or has an offset and size that corresponds to a component element
1278 /// within it. The offset checked here may have been formed from a GEP with a
1279 /// pointer bitcasted to a different type.
1280 void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize,
1281 const Type *MemOpType, bool isStore,
1283 // Check if this is a load/store of the entire alloca.
1284 if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) {
1285 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer
1286 // loads/stores (which are essentially the same as the MemIntrinsics with
1287 // regard to copying padding between elements). But, if an alloca is
1288 // flagged as both a source and destination of such operations, we'll need
1289 // to check later for padding between elements.
1290 if (!MemOpType || MemOpType->isIntegerTy()) {
1292 Info.isMemCpyDst = true;
1294 Info.isMemCpySrc = true;
1297 // This is also safe for references using a type that is compatible with
1298 // the type of the alloca, so that loads/stores can be rewritten using
1299 // insertvalue/extractvalue.
1300 if (isCompatibleAggregate(MemOpType, AI->getAllocatedType()))
1303 // Check if the offset/size correspond to a component within the alloca type.
1304 const Type *T = AI->getAllocatedType();
1305 if (TypeHasComponent(T, Offset, MemSize))
1308 return MarkUnsafe(Info);
1311 /// TypeHasComponent - Return true if T has a component type with the
1312 /// specified offset and size. If Size is zero, do not check the size.
1313 bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) {
1316 if (const StructType *ST = dyn_cast<StructType>(T)) {
1317 const StructLayout *Layout = TD->getStructLayout(ST);
1318 unsigned EltIdx = Layout->getElementContainingOffset(Offset);
1319 EltTy = ST->getContainedType(EltIdx);
1320 EltSize = TD->getTypeAllocSize(EltTy);
1321 Offset -= Layout->getElementOffset(EltIdx);
1322 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
1323 EltTy = AT->getElementType();
1324 EltSize = TD->getTypeAllocSize(EltTy);
1325 if (Offset >= AT->getNumElements() * EltSize)
1331 if (Offset == 0 && (Size == 0 || EltSize == Size))
1333 // Check if the component spans multiple elements.
1334 if (Offset + Size > EltSize)
1336 return TypeHasComponent(EltTy, Offset, Size);
1339 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite
1340 /// the instruction I, which references it, to use the separate elements.
1341 /// Offset indicates the position within AI that is referenced by this
1343 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
1344 SmallVector<AllocaInst*, 32> &NewElts) {
1345 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1346 Instruction *User = cast<Instruction>(*UI);
1348 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1349 RewriteBitCast(BC, AI, Offset, NewElts);
1350 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1351 RewriteGEP(GEPI, AI, Offset, NewElts);
1352 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1353 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1354 uint64_t MemSize = Length->getZExtValue();
1356 MemSize == TD->getTypeAllocSize(AI->getAllocatedType()))
1357 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
1358 // Otherwise the intrinsic can only touch a single element and the
1359 // address operand will be updated, so nothing else needs to be done.
1360 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1361 const Type *LIType = LI->getType();
1362 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) {
1364 // %res = load { i32, i32 }* %alloc
1366 // %load.0 = load i32* %alloc.0
1367 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
1368 // %load.1 = load i32* %alloc.1
1369 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
1370 // (Also works for arrays instead of structs)
1371 Value *Insert = UndefValue::get(LIType);
1372 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1373 Value *Load = new LoadInst(NewElts[i], "load", LI);
1374 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
1376 LI->replaceAllUsesWith(Insert);
1377 DeadInsts.push_back(LI);
1378 } else if (LIType->isIntegerTy() &&
1379 TD->getTypeAllocSize(LIType) ==
1380 TD->getTypeAllocSize(AI->getAllocatedType())) {
1381 // If this is a load of the entire alloca to an integer, rewrite it.
1382 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
1384 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1385 Value *Val = SI->getOperand(0);
1386 const Type *SIType = Val->getType();
1387 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) {
1389 // store { i32, i32 } %val, { i32, i32 }* %alloc
1391 // %val.0 = extractvalue { i32, i32 } %val, 0
1392 // store i32 %val.0, i32* %alloc.0
1393 // %val.1 = extractvalue { i32, i32 } %val, 1
1394 // store i32 %val.1, i32* %alloc.1
1395 // (Also works for arrays instead of structs)
1396 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1397 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
1398 new StoreInst(Extract, NewElts[i], SI);
1400 DeadInsts.push_back(SI);
1401 } else if (SIType->isIntegerTy() &&
1402 TD->getTypeAllocSize(SIType) ==
1403 TD->getTypeAllocSize(AI->getAllocatedType())) {
1404 // If this is a store of the entire alloca from an integer, rewrite it.
1405 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
1411 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced
1412 /// and recursively continue updating all of its uses.
1413 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
1414 SmallVector<AllocaInst*, 32> &NewElts) {
1415 RewriteForScalarRepl(BC, AI, Offset, NewElts);
1416 if (BC->getOperand(0) != AI)
1419 // The bitcast references the original alloca. Replace its uses with
1420 // references to the first new element alloca.
1421 Instruction *Val = NewElts[0];
1422 if (Val->getType() != BC->getDestTy()) {
1423 Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
1426 BC->replaceAllUsesWith(Val);
1427 DeadInsts.push_back(BC);
1430 /// FindElementAndOffset - Return the index of the element containing Offset
1431 /// within the specified type, which must be either a struct or an array.
1432 /// Sets T to the type of the element and Offset to the offset within that
1433 /// element. IdxTy is set to the type of the index result to be used in a
1434 /// GEP instruction.
1435 uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset,
1436 const Type *&IdxTy) {
1438 if (const StructType *ST = dyn_cast<StructType>(T)) {
1439 const StructLayout *Layout = TD->getStructLayout(ST);
1440 Idx = Layout->getElementContainingOffset(Offset);
1441 T = ST->getContainedType(Idx);
1442 Offset -= Layout->getElementOffset(Idx);
1443 IdxTy = Type::getInt32Ty(T->getContext());
1446 const ArrayType *AT = cast<ArrayType>(T);
1447 T = AT->getElementType();
1448 uint64_t EltSize = TD->getTypeAllocSize(T);
1449 Idx = Offset / EltSize;
1450 Offset -= Idx * EltSize;
1451 IdxTy = Type::getInt64Ty(T->getContext());
1455 /// RewriteGEP - Check if this GEP instruction moves the pointer across
1456 /// elements of the alloca that are being split apart, and if so, rewrite
1457 /// the GEP to be relative to the new element.
1458 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
1459 SmallVector<AllocaInst*, 32> &NewElts) {
1460 uint64_t OldOffset = Offset;
1461 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1462 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
1463 &Indices[0], Indices.size());
1465 RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
1467 const Type *T = AI->getAllocatedType();
1469 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy);
1470 if (GEPI->getOperand(0) == AI)
1471 OldIdx = ~0ULL; // Force the GEP to be rewritten.
1473 T = AI->getAllocatedType();
1474 uint64_t EltOffset = Offset;
1475 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
1477 // If this GEP does not move the pointer across elements of the alloca
1478 // being split, then it does not needs to be rewritten.
1482 const Type *i32Ty = Type::getInt32Ty(AI->getContext());
1483 SmallVector<Value*, 8> NewArgs;
1484 NewArgs.push_back(Constant::getNullValue(i32Ty));
1485 while (EltOffset != 0) {
1486 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
1487 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
1489 Instruction *Val = NewElts[Idx];
1490 if (NewArgs.size() > 1) {
1491 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(),
1492 NewArgs.end(), "", GEPI);
1493 Val->takeName(GEPI);
1495 if (Val->getType() != GEPI->getType())
1496 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI);
1497 GEPI->replaceAllUsesWith(Val);
1498 DeadInsts.push_back(GEPI);
1501 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
1502 /// Rewrite it to copy or set the elements of the scalarized memory.
1503 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
1505 SmallVector<AllocaInst*, 32> &NewElts) {
1506 // If this is a memcpy/memmove, construct the other pointer as the
1507 // appropriate type. The "Other" pointer is the pointer that goes to memory
1508 // that doesn't have anything to do with the alloca that we are promoting. For
1509 // memset, this Value* stays null.
1510 Value *OtherPtr = 0;
1511 unsigned MemAlignment = MI->getAlignment();
1512 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
1513 if (Inst == MTI->getRawDest())
1514 OtherPtr = MTI->getRawSource();
1516 assert(Inst == MTI->getRawSource());
1517 OtherPtr = MTI->getRawDest();
1521 // If there is an other pointer, we want to convert it to the same pointer
1522 // type as AI has, so we can GEP through it safely.
1524 unsigned AddrSpace =
1525 cast<PointerType>(OtherPtr->getType())->getAddressSpace();
1527 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an
1528 // optimization, but it's also required to detect the corner case where
1529 // both pointer operands are referencing the same memory, and where
1530 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This
1531 // function is only called for mem intrinsics that access the whole
1532 // aggregate, so non-zero GEPs are not an issue here.)
1533 OtherPtr = OtherPtr->stripPointerCasts();
1535 // Copying the alloca to itself is a no-op: just delete it.
1536 if (OtherPtr == AI || OtherPtr == NewElts[0]) {
1537 // This code will run twice for a no-op memcpy -- once for each operand.
1538 // Put only one reference to MI on the DeadInsts list.
1539 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(),
1540 E = DeadInsts.end(); I != E; ++I)
1541 if (*I == MI) return;
1542 DeadInsts.push_back(MI);
1546 // If the pointer is not the right type, insert a bitcast to the right
1549 PointerType::get(AI->getType()->getElementType(), AddrSpace);
1551 if (OtherPtr->getType() != NewTy)
1552 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI);
1555 // Process each element of the aggregate.
1556 bool SROADest = MI->getRawDest() == Inst;
1558 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
1560 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1561 // If this is a memcpy/memmove, emit a GEP of the other element address.
1562 Value *OtherElt = 0;
1563 unsigned OtherEltAlign = MemAlignment;
1566 Value *Idx[2] = { Zero,
1567 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
1568 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2,
1569 OtherPtr->getName()+"."+Twine(i),
1572 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
1573 const Type *OtherTy = OtherPtrTy->getElementType();
1574 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) {
1575 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
1577 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
1578 EltOffset = TD->getTypeAllocSize(EltTy)*i;
1581 // The alignment of the other pointer is the guaranteed alignment of the
1582 // element, which is affected by both the known alignment of the whole
1583 // mem intrinsic and the alignment of the element. If the alignment of
1584 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
1585 // known alignment is just 4 bytes.
1586 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
1589 Value *EltPtr = NewElts[i];
1590 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
1592 // If we got down to a scalar, insert a load or store as appropriate.
1593 if (EltTy->isSingleValueType()) {
1594 if (isa<MemTransferInst>(MI)) {
1596 // From Other to Alloca.
1597 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
1598 new StoreInst(Elt, EltPtr, MI);
1600 // From Alloca to Other.
1601 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
1602 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
1606 assert(isa<MemSetInst>(MI));
1608 // If the stored element is zero (common case), just store a null
1611 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) {
1613 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
1615 // If EltTy is a vector type, get the element type.
1616 const Type *ValTy = EltTy->getScalarType();
1618 // Construct an integer with the right value.
1619 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
1620 APInt OneVal(EltSize, CI->getZExtValue());
1621 APInt TotalVal(OneVal);
1623 for (unsigned i = 0; 8*i < EltSize; ++i) {
1624 TotalVal = TotalVal.shl(8);
1628 // Convert the integer value to the appropriate type.
1629 StoreVal = ConstantInt::get(CI->getContext(), TotalVal);
1630 if (ValTy->isPointerTy())
1631 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
1632 else if (ValTy->isFloatingPointTy())
1633 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
1634 assert(StoreVal->getType() == ValTy && "Type mismatch!");
1636 // If the requested value was a vector constant, create it.
1637 if (EltTy != ValTy) {
1638 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
1639 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
1640 StoreVal = ConstantVector::get(&Elts[0], NumElts);
1643 new StoreInst(StoreVal, EltPtr, MI);
1646 // Otherwise, if we're storing a byte variable, use a memset call for
1650 unsigned EltSize = TD->getTypeAllocSize(EltTy);
1652 IRBuilder<> Builder(MI);
1654 // Finally, insert the meminst for this element.
1655 if (isa<MemSetInst>(MI)) {
1656 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize,
1659 assert(isa<MemTransferInst>(MI));
1660 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr
1661 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr
1663 if (isa<MemCpyInst>(MI))
1664 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile());
1666 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile());
1669 DeadInsts.push_back(MI);
1672 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that
1673 /// overwrites the entire allocation. Extract out the pieces of the stored
1674 /// integer and store them individually.
1675 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
1676 SmallVector<AllocaInst*, 32> &NewElts){
1677 // Extract each element out of the integer according to its structure offset
1678 // and store the element value to the individual alloca.
1679 Value *SrcVal = SI->getOperand(0);
1680 const Type *AllocaEltTy = AI->getAllocatedType();
1681 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
1683 // Handle tail padding by extending the operand
1684 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
1685 SrcVal = new ZExtInst(SrcVal,
1686 IntegerType::get(SI->getContext(), AllocaSizeBits),
1689 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI
1692 // There are two forms here: AI could be an array or struct. Both cases
1693 // have different ways to compute the element offset.
1694 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1695 const StructLayout *Layout = TD->getStructLayout(EltSTy);
1697 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1698 // Get the number of bits to shift SrcVal to get the value.
1699 const Type *FieldTy = EltSTy->getElementType(i);
1700 uint64_t Shift = Layout->getElementOffsetInBits(i);
1702 if (TD->isBigEndian())
1703 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
1705 Value *EltVal = SrcVal;
1707 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
1708 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
1709 "sroa.store.elt", SI);
1712 // Truncate down to an integer of the right size.
1713 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1715 // Ignore zero sized fields like {}, they obviously contain no data.
1716 if (FieldSizeBits == 0) continue;
1718 if (FieldSizeBits != AllocaSizeBits)
1719 EltVal = new TruncInst(EltVal,
1720 IntegerType::get(SI->getContext(), FieldSizeBits),
1722 Value *DestField = NewElts[i];
1723 if (EltVal->getType() == FieldTy) {
1724 // Storing to an integer field of this size, just do it.
1725 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) {
1726 // Bitcast to the right element type (for fp/vector values).
1727 EltVal = new BitCastInst(EltVal, FieldTy, "", SI);
1729 // Otherwise, bitcast the dest pointer (for aggregates).
1730 DestField = new BitCastInst(DestField,
1731 PointerType::getUnqual(EltVal->getType()),
1734 new StoreInst(EltVal, DestField, SI);
1738 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
1739 const Type *ArrayEltTy = ATy->getElementType();
1740 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
1741 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
1745 if (TD->isBigEndian())
1746 Shift = AllocaSizeBits-ElementOffset;
1750 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1751 // Ignore zero sized fields like {}, they obviously contain no data.
1752 if (ElementSizeBits == 0) continue;
1754 Value *EltVal = SrcVal;
1756 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
1757 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal,
1758 "sroa.store.elt", SI);
1761 // Truncate down to an integer of the right size.
1762 if (ElementSizeBits != AllocaSizeBits)
1763 EltVal = new TruncInst(EltVal,
1764 IntegerType::get(SI->getContext(),
1765 ElementSizeBits), "", SI);
1766 Value *DestField = NewElts[i];
1767 if (EltVal->getType() == ArrayEltTy) {
1768 // Storing to an integer field of this size, just do it.
1769 } else if (ArrayEltTy->isFloatingPointTy() ||
1770 ArrayEltTy->isVectorTy()) {
1771 // Bitcast to the right element type (for fp/vector values).
1772 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI);
1774 // Otherwise, bitcast the dest pointer (for aggregates).
1775 DestField = new BitCastInst(DestField,
1776 PointerType::getUnqual(EltVal->getType()),
1779 new StoreInst(EltVal, DestField, SI);
1781 if (TD->isBigEndian())
1782 Shift -= ElementOffset;
1784 Shift += ElementOffset;
1788 DeadInsts.push_back(SI);
1791 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to
1792 /// an integer. Load the individual pieces to form the aggregate value.
1793 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
1794 SmallVector<AllocaInst*, 32> &NewElts) {
1795 // Extract each element out of the NewElts according to its structure offset
1796 // and form the result value.
1797 const Type *AllocaEltTy = AI->getAllocatedType();
1798 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
1800 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
1803 // There are two forms here: AI could be an array or struct. Both cases
1804 // have different ways to compute the element offset.
1805 const StructLayout *Layout = 0;
1806 uint64_t ArrayEltBitOffset = 0;
1807 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
1808 Layout = TD->getStructLayout(EltSTy);
1810 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
1811 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
1815 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
1817 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1818 // Load the value from the alloca. If the NewElt is an aggregate, cast
1819 // the pointer to an integer of the same size before doing the load.
1820 Value *SrcField = NewElts[i];
1821 const Type *FieldTy =
1822 cast<PointerType>(SrcField->getType())->getElementType();
1823 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
1825 // Ignore zero sized fields like {}, they obviously contain no data.
1826 if (FieldSizeBits == 0) continue;
1828 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
1830 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() &&
1831 !FieldTy->isVectorTy())
1832 SrcField = new BitCastInst(SrcField,
1833 PointerType::getUnqual(FieldIntTy),
1835 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
1837 // If SrcField is a fp or vector of the right size but that isn't an
1838 // integer type, bitcast to an integer so we can shift it.
1839 if (SrcField->getType() != FieldIntTy)
1840 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
1842 // Zero extend the field to be the same size as the final alloca so that
1843 // we can shift and insert it.
1844 if (SrcField->getType() != ResultVal->getType())
1845 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
1847 // Determine the number of bits to shift SrcField.
1849 if (Layout) // Struct case.
1850 Shift = Layout->getElementOffsetInBits(i);
1852 Shift = i*ArrayEltBitOffset;
1854 if (TD->isBigEndian())
1855 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
1858 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
1859 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
1862 // Don't create an 'or x, 0' on the first iteration.
1863 if (!isa<Constant>(ResultVal) ||
1864 !cast<Constant>(ResultVal)->isNullValue())
1865 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
1867 ResultVal = SrcField;
1870 // Handle tail padding by truncating the result
1871 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
1872 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
1874 LI->replaceAllUsesWith(ResultVal);
1875 DeadInsts.push_back(LI);
1878 /// HasPadding - Return true if the specified type has any structure or
1879 /// alignment padding in between the elements that would be split apart
1880 /// by SROA; return false otherwise.
1881 static bool HasPadding(const Type *Ty, const TargetData &TD) {
1882 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1883 Ty = ATy->getElementType();
1884 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
1887 // SROA currently handles only Arrays and Structs.
1888 const StructType *STy = cast<StructType>(Ty);
1889 const StructLayout *SL = TD.getStructLayout(STy);
1890 unsigned PrevFieldBitOffset = 0;
1891 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1892 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
1894 // Check to see if there is any padding between this element and the
1897 unsigned PrevFieldEnd =
1898 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
1899 if (PrevFieldEnd < FieldBitOffset)
1902 PrevFieldBitOffset = FieldBitOffset;
1904 // Check for tail padding.
1905 if (unsigned EltCount = STy->getNumElements()) {
1906 unsigned PrevFieldEnd = PrevFieldBitOffset +
1907 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
1908 if (PrevFieldEnd < SL->getSizeInBits())
1914 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
1915 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
1916 /// or 1 if safe after canonicalization has been performed.
1917 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) {
1918 // Loop over the use list of the alloca. We can only transform it if all of
1919 // the users are safe to transform.
1922 isSafeForScalarRepl(AI, AI, 0, Info);
1923 if (Info.isUnsafe) {
1924 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n');
1928 // Okay, we know all the users are promotable. If the aggregate is a memcpy
1929 // source and destination, we have to be careful. In particular, the memcpy
1930 // could be moving around elements that live in structure padding of the LLVM
1931 // types, but may actually be used. In these cases, we refuse to promote the
1933 if (Info.isMemCpySrc && Info.isMemCpyDst &&
1934 HasPadding(AI->getAllocatedType(), *TD))
1942 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
1943 /// some part of a constant global variable. This intentionally only accepts
1944 /// constant expressions because we don't can't rewrite arbitrary instructions.
1945 static bool PointsToConstantGlobal(Value *V) {
1946 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
1947 return GV->isConstant();
1948 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1949 if (CE->getOpcode() == Instruction::BitCast ||
1950 CE->getOpcode() == Instruction::GetElementPtr)
1951 return PointsToConstantGlobal(CE->getOperand(0));
1955 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
1956 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
1957 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
1958 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
1959 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
1960 /// the alloca, and if the source pointer is a pointer to a constant global, we
1961 /// can optimize this.
1962 static bool isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
1964 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1965 User *U = cast<Instruction>(*UI);
1967 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1968 // Ignore non-volatile loads, they are always ok.
1969 if (LI->isVolatile()) return false;
1973 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
1974 // If uses of the bitcast are ok, we are ok.
1975 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
1979 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
1980 // If the GEP has all zero indices, it doesn't offset the pointer. If it
1981 // doesn't, it does.
1982 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
1983 isOffset || !GEP->hasAllZeroIndices()))
1988 if (CallSite CS = U) {
1989 // If this is a readonly/readnone call site, then we know it is just a
1990 // load and we can ignore it.
1991 if (CS.onlyReadsMemory())
1994 // If this is the function being called then we treat it like a load and
1996 if (CS.isCallee(UI))
1999 // If this is being passed as a byval argument, the caller is making a
2000 // copy, so it is only a read of the alloca.
2001 unsigned ArgNo = CS.getArgumentNo(UI);
2002 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal))
2006 // If this is isn't our memcpy/memmove, reject it as something we can't
2008 MemTransferInst *MI = dyn_cast<MemTransferInst>(U);
2012 // If the transfer is using the alloca as a source of the transfer, then
2013 // ignore it since it is a load (unless the transfer is volatile).
2014 if (UI.getOperandNo() == 1) {
2015 if (MI->isVolatile()) return false;
2019 // If we already have seen a copy, reject the second one.
2020 if (TheCopy) return false;
2022 // If the pointer has been offset from the start of the alloca, we can't
2023 // safely handle this.
2024 if (isOffset) return false;
2026 // If the memintrinsic isn't using the alloca as the dest, reject it.
2027 if (UI.getOperandNo() != 0) return false;
2029 // If the source of the memcpy/move is not a constant global, reject it.
2030 if (!PointsToConstantGlobal(MI->getSource()))
2033 // Otherwise, the transform is safe. Remember the copy instruction.
2039 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
2040 /// modified by a copy from a constant global. If we can prove this, we can
2041 /// replace any uses of the alloca with uses of the global directly.
2042 MemTransferInst *SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI) {
2043 MemTransferInst *TheCopy = 0;
2044 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))