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/DIBuilder.h"
34 #include "llvm/Analysis/Dominators.h"
35 #include "llvm/Analysis/Loads.h"
36 #include "llvm/Analysis/ValueTracking.h"
37 #include "llvm/Target/TargetData.h"
38 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
39 #include "llvm/Transforms/Utils/Local.h"
40 #include "llvm/Transforms/Utils/SSAUpdater.h"
41 #include "llvm/Support/CallSite.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/GetElementPtrTypeIterator.h"
45 #include "llvm/Support/IRBuilder.h"
46 #include "llvm/Support/MathExtras.h"
47 #include "llvm/Support/raw_ostream.h"
48 #include "llvm/ADT/SetVector.h"
49 #include "llvm/ADT/SmallVector.h"
50 #include "llvm/ADT/Statistic.h"
53 STATISTIC(NumReplaced, "Number of allocas broken up");
54 STATISTIC(NumPromoted, "Number of allocas promoted");
55 STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion");
56 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
57 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
60 struct SROA : public FunctionPass {
61 SROA(int T, bool hasDT, char &ID)
62 : FunctionPass(ID), HasDomTree(hasDT) {
69 bool runOnFunction(Function &F);
71 bool performScalarRepl(Function &F);
72 bool performPromotion(Function &F);
78 /// DeadInsts - Keep track of instructions we have made dead, so that
79 /// we can remove them after we are done working.
80 SmallVector<Value*, 32> DeadInsts;
82 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
83 /// information about the uses. All these fields are initialized to false
84 /// and set to true when something is learned.
86 /// The alloca to promote.
89 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite
90 /// looping and avoid redundant work.
91 SmallPtrSet<PHINode*, 8> CheckedPHIs;
93 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
96 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
99 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
100 bool isMemCpyDst : 1;
102 /// hasSubelementAccess - This is true if a subelement of the alloca is
103 /// ever accessed, or false if the alloca is only accessed with mem
104 /// intrinsics or load/store that only access the entire alloca at once.
105 bool hasSubelementAccess : 1;
107 /// hasALoadOrStore - This is true if there are any loads or stores to it.
108 /// The alloca may just be accessed with memcpy, for example, which would
110 bool hasALoadOrStore : 1;
112 explicit AllocaInfo(AllocaInst *ai)
113 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false),
114 hasSubelementAccess(false), hasALoadOrStore(false) {}
117 unsigned SRThreshold;
119 void MarkUnsafe(AllocaInfo &I, Instruction *User) {
121 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n');
124 bool isSafeAllocaToScalarRepl(AllocaInst *AI);
126 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info);
127 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset,
129 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info);
130 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
131 const Type *MemOpType, bool isStore, AllocaInfo &Info,
132 Instruction *TheAccess, bool AllowWholeAccess);
133 bool TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size);
134 uint64_t FindElementAndOffset(const Type *&T, uint64_t &Offset,
137 void DoScalarReplacement(AllocaInst *AI,
138 std::vector<AllocaInst*> &WorkList);
139 void DeleteDeadInstructions();
141 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
142 SmallVector<AllocaInst*, 32> &NewElts);
143 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
144 SmallVector<AllocaInst*, 32> &NewElts);
145 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
146 SmallVector<AllocaInst*, 32> &NewElts);
147 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
149 SmallVector<AllocaInst*, 32> &NewElts);
150 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
151 SmallVector<AllocaInst*, 32> &NewElts);
152 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
153 SmallVector<AllocaInst*, 32> &NewElts);
155 static MemTransferInst *isOnlyCopiedFromConstantGlobal(
156 AllocaInst *AI, SmallVector<Instruction*, 4> &ToDelete);
159 // SROA_DT - SROA that uses DominatorTree.
160 struct SROA_DT : public SROA {
163 SROA_DT(int T = -1) : SROA(T, true, ID) {
164 initializeSROA_DTPass(*PassRegistry::getPassRegistry());
167 // getAnalysisUsage - This pass does not require any passes, but we know it
168 // will not alter the CFG, so say so.
169 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
170 AU.addRequired<DominatorTree>();
171 AU.setPreservesCFG();
175 // SROA_SSAUp - SROA that uses SSAUpdater.
176 struct SROA_SSAUp : public SROA {
179 SROA_SSAUp(int T = -1) : SROA(T, false, ID) {
180 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry());
183 // getAnalysisUsage - This pass does not require any passes, but we know it
184 // will not alter the CFG, so say so.
185 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
186 AU.setPreservesCFG();
192 char SROA_DT::ID = 0;
193 char SROA_SSAUp::ID = 0;
195 INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl",
196 "Scalar Replacement of Aggregates (DT)", false, false)
197 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
198 INITIALIZE_PASS_END(SROA_DT, "scalarrepl",
199 "Scalar Replacement of Aggregates (DT)", false, false)
201 INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa",
202 "Scalar Replacement of Aggregates (SSAUp)", false, false)
203 INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa",
204 "Scalar Replacement of Aggregates (SSAUp)", false, false)
206 // Public interface to the ScalarReplAggregates pass
207 FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold,
210 return new SROA_DT(Threshold);
211 return new SROA_SSAUp(Threshold);
215 //===----------------------------------------------------------------------===//
216 // Convert To Scalar Optimization.
217 //===----------------------------------------------------------------------===//
220 /// ConvertToScalarInfo - This class implements the "Convert To Scalar"
221 /// optimization, which scans the uses of an alloca and determines if it can
222 /// rewrite it in terms of a single new alloca that can be mem2reg'd.
223 class ConvertToScalarInfo {
224 /// AllocaSize - The size of the alloca being considered in bytes.
226 const TargetData &TD;
228 /// IsNotTrivial - This is set to true if there is some access to the object
229 /// which means that mem2reg can't promote it.
232 /// ScalarKind - Tracks the kind of alloca being considered for promotion,
233 /// computed based on the uses of the alloca rather than the LLVM type system.
237 // Accesses via GEPs that are consistent with element access of a vector
238 // type. This will not be converted into a vector unless there is a later
239 // access using an actual vector type.
242 // Accesses via vector operations and GEPs that are consistent with the
243 // layout of a vector type.
246 // An integer bag-of-bits with bitwise operations for insertion and
247 // extraction. Any combination of types can be converted into this kind
252 /// VectorTy - This tracks the type that we should promote the vector to if
253 /// it is possible to turn it into a vector. This starts out null, and if it
254 /// isn't possible to turn into a vector type, it gets set to VoidTy.
255 const VectorType *VectorTy;
257 /// HadNonMemTransferAccess - True if there is at least one access to the
258 /// alloca that is not a MemTransferInst. We don't want to turn structs into
259 /// large integers unless there is some potential for optimization.
260 bool HadNonMemTransferAccess;
263 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td)
264 : AllocaSize(Size), TD(td), IsNotTrivial(false), ScalarKind(Unknown),
265 VectorTy(0), HadNonMemTransferAccess(false) { }
267 AllocaInst *TryConvert(AllocaInst *AI);
270 bool CanConvertToScalar(Value *V, uint64_t Offset);
271 void MergeInTypeForLoadOrStore(const Type *In, uint64_t Offset);
272 bool MergeInVectorType(const VectorType *VInTy, uint64_t Offset);
273 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
275 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType,
276 uint64_t Offset, IRBuilder<> &Builder);
277 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
278 uint64_t Offset, IRBuilder<> &Builder);
280 } // end anonymous namespace.
283 /// TryConvert - Analyze the specified alloca, and if it is safe to do so,
284 /// rewrite it to be a new alloca which is mem2reg'able. This returns the new
285 /// alloca if possible or null if not.
286 AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
287 // If we can't convert this scalar, or if mem2reg can trivially do it, bail
289 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial)
292 // If an alloca has only memset / memcpy uses, it may still have an Unknown
293 // ScalarKind. Treat it as an Integer below.
294 if (ScalarKind == Unknown)
295 ScalarKind = Integer;
297 // FIXME: It should be possible to promote the vector type up to the alloca's
299 if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8)
300 ScalarKind = Integer;
302 // If we were able to find a vector type that can handle this with
303 // insert/extract elements, and if there was at least one use that had
304 // a vector type, promote this to a vector. We don't want to promote
305 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
306 // we just get a lot of insert/extracts. If at least one vector is
307 // involved, then we probably really do have a union of vector/array.
309 if (ScalarKind == Vector) {
310 assert(VectorTy && "Missing type for vector scalar.");
311 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = "
312 << *VectorTy << '\n');
313 NewTy = VectorTy; // Use the vector type.
315 unsigned BitWidth = AllocaSize * 8;
316 if ((ScalarKind == ImplicitVector || ScalarKind == Integer) &&
317 !HadNonMemTransferAccess && !TD.fitsInLegalInteger(BitWidth))
320 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
321 // Create and insert the integer alloca.
322 NewTy = IntegerType::get(AI->getContext(), BitWidth);
324 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
325 ConvertUsesToScalar(AI, NewAI, 0);
329 /// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type
330 /// (VectorTy) so far at the offset specified by Offset (which is specified in
333 /// There are three cases we handle here:
334 /// 1) A union of vector types of the same size and potentially its elements.
335 /// Here we turn element accesses into insert/extract element operations.
336 /// This promotes a <4 x float> with a store of float to the third element
337 /// into a <4 x float> that uses insert element.
338 /// 2) A union of vector types with power-of-2 size differences, e.g. a float,
339 /// <2 x float> and <4 x float>. Here we turn element accesses into insert
340 /// and extract element operations, and <2 x float> accesses into a cast to
341 /// <2 x double>, an extract, and a cast back to <2 x float>.
342 /// 3) A fully general blob of memory, which we turn into some (potentially
343 /// large) integer type with extract and insert operations where the loads
344 /// and stores would mutate the memory. We mark this by setting VectorTy
346 void ConvertToScalarInfo::MergeInTypeForLoadOrStore(const Type *In,
348 // If we already decided to turn this into a blob of integer memory, there is
349 // nothing to be done.
350 if (ScalarKind == Integer)
353 // If this could be contributing to a vector, analyze it.
355 // If the In type is a vector that is the same size as the alloca, see if it
356 // matches the existing VecTy.
357 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) {
358 if (MergeInVectorType(VInTy, Offset))
360 } else if (In->isFloatTy() || In->isDoubleTy() ||
361 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 &&
362 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
363 // Full width accesses can be ignored, because they can always be turned
365 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
366 if (EltSize == AllocaSize)
369 // If we're accessing something that could be an element of a vector, see
370 // if the implied vector agrees with what we already have and if Offset is
371 // compatible with it.
372 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 &&
373 (!VectorTy || Offset * 8 < VectorTy->getPrimitiveSizeInBits())) {
375 ScalarKind = ImplicitVector;
376 VectorTy = VectorType::get(In, AllocaSize/EltSize);
380 unsigned CurrentEltSize = VectorTy->getElementType()
381 ->getPrimitiveSizeInBits()/8;
382 if (EltSize == CurrentEltSize)
385 if (In->isIntegerTy() && isPowerOf2_32(AllocaSize / EltSize))
390 // Otherwise, we have a case that we can't handle with an optimized vector
391 // form. We can still turn this into a large integer.
392 ScalarKind = Integer;
395 /// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore,
396 /// returning true if the type was successfully merged and false otherwise.
397 bool ConvertToScalarInfo::MergeInVectorType(const VectorType *VInTy,
399 // TODO: Support nonzero offsets?
403 // Only allow vectors that are a power-of-2 away from the size of the alloca.
404 if (!isPowerOf2_64(AllocaSize / (VInTy->getBitWidth() / 8)))
407 // If this the first vector we see, remember the type so that we know the
415 unsigned BitWidth = VectorTy->getBitWidth();
416 unsigned InBitWidth = VInTy->getBitWidth();
418 // Vectors of the same size can be converted using a simple bitcast.
419 if (InBitWidth == BitWidth && AllocaSize == (InBitWidth / 8)) {
424 const Type *ElementTy = VectorTy->getElementType();
425 const Type *InElementTy = VInTy->getElementType();
427 // Do not allow mixed integer and floating-point accesses from vectors of
429 if (ElementTy->isFloatingPointTy() != InElementTy->isFloatingPointTy())
432 if (ElementTy->isFloatingPointTy()) {
433 // Only allow floating-point vectors of different sizes if they have the
434 // same element type.
435 // TODO: This could be loosened a bit, but would anything benefit?
436 if (ElementTy != InElementTy)
439 // There are no arbitrary-precision floating-point types, which limits the
440 // number of legal vector types with larger element types that we can form
441 // to bitcast and extract a subvector.
442 // TODO: We could support some more cases with mixed fp128 and double here.
443 if (!(BitWidth == 64 || BitWidth == 128) ||
444 !(InBitWidth == 64 || InBitWidth == 128))
447 assert(ElementTy->isIntegerTy() && "Vector elements must be either integer "
448 "or floating-point.");
449 unsigned BitWidth = ElementTy->getPrimitiveSizeInBits();
450 unsigned InBitWidth = InElementTy->getPrimitiveSizeInBits();
452 // Do not allow integer types smaller than a byte or types whose widths are
453 // not a multiple of a byte.
454 if (BitWidth < 8 || InBitWidth < 8 ||
455 BitWidth % 8 != 0 || InBitWidth % 8 != 0)
459 // Pick the largest of the two vector types.
461 if (InBitWidth > BitWidth)
467 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
468 /// its accesses to a single vector type, return true and set VecTy to
469 /// the new type. If we could convert the alloca into a single promotable
470 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
471 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
472 /// is the current offset from the base of the alloca being analyzed.
474 /// If we see at least one access to the value that is as a vector type, set the
476 bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) {
477 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
478 Instruction *User = cast<Instruction>(*UI);
480 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
481 // Don't break volatile loads.
482 if (LI->isVolatile())
484 // Don't touch MMX operations.
485 if (LI->getType()->isX86_MMXTy())
487 HadNonMemTransferAccess = true;
488 MergeInTypeForLoadOrStore(LI->getType(), Offset);
492 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
493 // Storing the pointer, not into the value?
494 if (SI->getOperand(0) == V || SI->isVolatile()) return false;
495 // Don't touch MMX operations.
496 if (SI->getOperand(0)->getType()->isX86_MMXTy())
498 HadNonMemTransferAccess = true;
499 MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset);
503 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
504 IsNotTrivial = true; // Can't be mem2reg'd.
505 if (!CanConvertToScalar(BCI, Offset))
510 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
511 // If this is a GEP with a variable indices, we can't handle it.
512 if (!GEP->hasAllConstantIndices())
515 // Compute the offset that this GEP adds to the pointer.
516 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
517 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
518 &Indices[0], Indices.size());
519 // See if all uses can be converted.
520 if (!CanConvertToScalar(GEP, Offset+GEPOffset))
522 IsNotTrivial = true; // Can't be mem2reg'd.
523 HadNonMemTransferAccess = true;
527 // If this is a constant sized memset of a constant value (e.g. 0) we can
529 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
530 // Store of constant value.
531 if (!isa<ConstantInt>(MSI->getValue()))
534 // Store of constant size.
535 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength());
539 // If the size differs from the alloca, we can only convert the alloca to
540 // an integer bag-of-bits.
541 // FIXME: This should handle all of the cases that are currently accepted
542 // as vector element insertions.
543 if (Len->getZExtValue() != AllocaSize || Offset != 0)
544 ScalarKind = Integer;
546 IsNotTrivial = true; // Can't be mem2reg'd.
547 HadNonMemTransferAccess = true;
551 // If this is a memcpy or memmove into or out of the whole allocation, we
552 // can handle it like a load or store of the scalar type.
553 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
554 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength());
555 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0)
558 IsNotTrivial = true; // Can't be mem2reg'd.
562 // Otherwise, we cannot handle this!
569 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
570 /// directly. This happens when we are converting an "integer union" to a
571 /// single integer scalar, or when we are converting a "vector union" to a
572 /// vector with insert/extractelement instructions.
574 /// Offset is an offset from the original alloca, in bits that need to be
575 /// shifted to the right. By the end of this, there should be no uses of Ptr.
576 void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI,
578 while (!Ptr->use_empty()) {
579 Instruction *User = cast<Instruction>(Ptr->use_back());
581 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
582 ConvertUsesToScalar(CI, NewAI, Offset);
583 CI->eraseFromParent();
587 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
588 // Compute the offset that this GEP adds to the pointer.
589 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
590 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
591 &Indices[0], Indices.size());
592 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
593 GEP->eraseFromParent();
597 IRBuilder<> Builder(User);
599 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
600 // The load is a bit extract from NewAI shifted right by Offset bits.
601 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
603 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
604 LI->replaceAllUsesWith(NewLoadVal);
605 LI->eraseFromParent();
609 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
610 assert(SI->getOperand(0) != Ptr && "Consistency error!");
611 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
612 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
614 Builder.CreateStore(New, NewAI);
615 SI->eraseFromParent();
617 // If the load we just inserted is now dead, then the inserted store
618 // overwrote the entire thing.
619 if (Old->use_empty())
620 Old->eraseFromParent();
624 // If this is a constant sized memset of a constant value (e.g. 0) we can
625 // transform it into a store of the expanded constant value.
626 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
627 assert(MSI->getRawDest() == Ptr && "Consistency error!");
628 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
630 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
632 // Compute the value replicated the right number of times.
633 APInt APVal(NumBytes*8, Val);
635 // Splat the value if non-zero.
637 for (unsigned i = 1; i != NumBytes; ++i)
640 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
641 Value *New = ConvertScalar_InsertValue(
642 ConstantInt::get(User->getContext(), APVal),
643 Old, Offset, Builder);
644 Builder.CreateStore(New, NewAI);
646 // If the load we just inserted is now dead, then the memset overwrote
648 if (Old->use_empty())
649 Old->eraseFromParent();
651 MSI->eraseFromParent();
655 // If this is a memcpy or memmove into or out of the whole allocation, we
656 // can handle it like a load or store of the scalar type.
657 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
658 assert(Offset == 0 && "must be store to start of alloca");
660 // If the source and destination are both to the same alloca, then this is
661 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
663 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0));
665 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) {
666 // Dest must be OrigAI, change this to be a load from the original
667 // pointer (bitcasted), then a store to our new alloca.
668 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
669 Value *SrcPtr = MTI->getSource();
670 const PointerType* SPTy = cast<PointerType>(SrcPtr->getType());
671 const PointerType* AIPTy = cast<PointerType>(NewAI->getType());
672 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
673 AIPTy = PointerType::get(AIPTy->getElementType(),
674 SPTy->getAddressSpace());
676 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy);
678 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
679 SrcVal->setAlignment(MTI->getAlignment());
680 Builder.CreateStore(SrcVal, NewAI);
681 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) {
682 // Src must be OrigAI, change this to be a load from NewAI then a store
683 // through the original dest pointer (bitcasted).
684 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
685 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
687 const PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType());
688 const PointerType* AIPTy = cast<PointerType>(NewAI->getType());
689 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
690 AIPTy = PointerType::get(AIPTy->getElementType(),
691 DPTy->getAddressSpace());
693 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy);
695 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
696 NewStore->setAlignment(MTI->getAlignment());
698 // Noop transfer. Src == Dst
701 MTI->eraseFromParent();
705 llvm_unreachable("Unsupported operation!");
709 /// getScaledElementType - Gets a scaled element type for a partial vector
710 /// access of an alloca. The input types must be integer or floating-point
711 /// scalar or vector types, and the resulting type is an integer, float or
713 static const Type *getScaledElementType(const Type *Ty1, const Type *Ty2,
714 unsigned NewBitWidth) {
715 bool IsFP1 = Ty1->isFloatingPointTy() ||
716 (Ty1->isVectorTy() &&
717 cast<VectorType>(Ty1)->getElementType()->isFloatingPointTy());
718 bool IsFP2 = Ty2->isFloatingPointTy() ||
719 (Ty2->isVectorTy() &&
720 cast<VectorType>(Ty2)->getElementType()->isFloatingPointTy());
722 LLVMContext &Context = Ty1->getContext();
724 // Prefer floating-point types over integer types, as integer types may have
725 // been created by earlier scalar replacement.
726 if (IsFP1 || IsFP2) {
727 if (NewBitWidth == 32)
728 return Type::getFloatTy(Context);
729 if (NewBitWidth == 64)
730 return Type::getDoubleTy(Context);
733 return Type::getIntNTy(Context, NewBitWidth);
736 /// CreateShuffleVectorCast - Creates a shuffle vector to convert one vector
737 /// to another vector of the same element type which has the same allocation
738 /// size but different primitive sizes (e.g. <3 x i32> and <4 x i32>).
739 static Value *CreateShuffleVectorCast(Value *FromVal, const Type *ToType,
740 IRBuilder<> &Builder) {
741 const Type *FromType = FromVal->getType();
742 const VectorType *FromVTy = cast<VectorType>(FromType);
743 const VectorType *ToVTy = cast<VectorType>(ToType);
744 assert((ToVTy->getElementType() == FromVTy->getElementType()) &&
745 "Vectors must have the same element type");
746 Value *UnV = UndefValue::get(FromType);
747 unsigned numEltsFrom = FromVTy->getNumElements();
748 unsigned numEltsTo = ToVTy->getNumElements();
750 SmallVector<Constant*, 3> Args;
751 const Type* Int32Ty = Builder.getInt32Ty();
752 unsigned minNumElts = std::min(numEltsFrom, numEltsTo);
754 for (i=0; i != minNumElts; ++i)
755 Args.push_back(ConstantInt::get(Int32Ty, i));
758 Constant* UnC = UndefValue::get(Int32Ty);
759 for (; i != numEltsTo; ++i)
762 Constant *Mask = ConstantVector::get(Args);
763 return Builder.CreateShuffleVector(FromVal, UnV, Mask, "tmpV");
766 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
767 /// or vector value FromVal, extracting the bits from the offset specified by
768 /// Offset. This returns the value, which is of type ToType.
770 /// This happens when we are converting an "integer union" to a single
771 /// integer scalar, or when we are converting a "vector union" to a vector with
772 /// insert/extractelement instructions.
774 /// Offset is an offset from the original alloca, in bits that need to be
775 /// shifted to the right.
776 Value *ConvertToScalarInfo::
777 ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType,
778 uint64_t Offset, IRBuilder<> &Builder) {
779 // If the load is of the whole new alloca, no conversion is needed.
780 const Type *FromType = FromVal->getType();
781 if (FromType == ToType && Offset == 0)
784 // If the result alloca is a vector type, this is either an element
785 // access or a bitcast to another vector type of the same size.
786 if (const VectorType *VTy = dyn_cast<VectorType>(FromType)) {
787 unsigned FromTypeSize = TD.getTypeAllocSize(FromType);
788 unsigned ToTypeSize = TD.getTypeAllocSize(ToType);
789 if (FromTypeSize == ToTypeSize) {
790 // If the two types have the same primitive size, use a bit cast.
791 // Otherwise, it is two vectors with the same element type that has
792 // the same allocation size but different number of elements so use
794 if (FromType->getPrimitiveSizeInBits() ==
795 ToType->getPrimitiveSizeInBits())
796 return Builder.CreateBitCast(FromVal, ToType, "tmp");
798 return CreateShuffleVectorCast(FromVal, ToType, Builder);
801 if (isPowerOf2_64(FromTypeSize / ToTypeSize)) {
802 assert(!(ToType->isVectorTy() && Offset != 0) && "Can't extract a value "
803 "of a smaller vector type at a nonzero offset.");
805 const Type *CastElementTy = getScaledElementType(FromType, ToType,
807 unsigned NumCastVectorElements = FromTypeSize / ToTypeSize;
809 LLVMContext &Context = FromVal->getContext();
810 const Type *CastTy = VectorType::get(CastElementTy,
811 NumCastVectorElements);
812 Value *Cast = Builder.CreateBitCast(FromVal, CastTy, "tmp");
814 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy);
815 unsigned Elt = Offset/EltSize;
816 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
817 Value *Extract = Builder.CreateExtractElement(Cast, ConstantInt::get(
818 Type::getInt32Ty(Context), Elt), "tmp");
819 return Builder.CreateBitCast(Extract, ToType, "tmp");
822 // Otherwise it must be an element access.
825 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
826 Elt = Offset/EltSize;
827 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
829 // Return the element extracted out of it.
830 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get(
831 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp");
832 if (V->getType() != ToType)
833 V = Builder.CreateBitCast(V, ToType, "tmp");
837 // If ToType is a first class aggregate, extract out each of the pieces and
838 // use insertvalue's to form the FCA.
839 if (const StructType *ST = dyn_cast<StructType>(ToType)) {
840 const StructLayout &Layout = *TD.getStructLayout(ST);
841 Value *Res = UndefValue::get(ST);
842 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
843 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
844 Offset+Layout.getElementOffsetInBits(i),
846 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
851 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
852 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
853 Value *Res = UndefValue::get(AT);
854 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
855 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
856 Offset+i*EltSize, Builder);
857 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
862 // Otherwise, this must be a union that was converted to an integer value.
863 const IntegerType *NTy = cast<IntegerType>(FromVal->getType());
865 // If this is a big-endian system and the load is narrower than the
866 // full alloca type, we need to do a shift to get the right bits.
868 if (TD.isBigEndian()) {
869 // On big-endian machines, the lowest bit is stored at the bit offset
870 // from the pointer given by getTypeStoreSizeInBits. This matters for
871 // integers with a bitwidth that is not a multiple of 8.
872 ShAmt = TD.getTypeStoreSizeInBits(NTy) -
873 TD.getTypeStoreSizeInBits(ToType) - Offset;
878 // Note: we support negative bitwidths (with shl) which are not defined.
879 // We do this to support (f.e.) loads off the end of a structure where
880 // only some bits are used.
881 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
882 FromVal = Builder.CreateLShr(FromVal,
883 ConstantInt::get(FromVal->getType(),
885 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
886 FromVal = Builder.CreateShl(FromVal,
887 ConstantInt::get(FromVal->getType(),
890 // Finally, unconditionally truncate the integer to the right width.
891 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType);
892 if (LIBitWidth < NTy->getBitWidth())
894 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
896 else if (LIBitWidth > NTy->getBitWidth())
898 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
901 // If the result is an integer, this is a trunc or bitcast.
902 if (ToType->isIntegerTy()) {
904 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) {
905 // Just do a bitcast, we know the sizes match up.
906 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
908 // Otherwise must be a pointer.
909 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
911 assert(FromVal->getType() == ToType && "Didn't convert right?");
915 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
916 /// or vector value "Old" at the offset specified by Offset.
918 /// This happens when we are converting an "integer union" to a
919 /// single integer scalar, or when we are converting a "vector union" to a
920 /// vector with insert/extractelement instructions.
922 /// Offset is an offset from the original alloca, in bits that need to be
923 /// shifted to the right.
924 Value *ConvertToScalarInfo::
925 ConvertScalar_InsertValue(Value *SV, Value *Old,
926 uint64_t Offset, IRBuilder<> &Builder) {
927 // Convert the stored type to the actual type, shift it left to insert
928 // then 'or' into place.
929 const Type *AllocaType = Old->getType();
930 LLVMContext &Context = Old->getContext();
932 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
933 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy);
934 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType());
936 // Changing the whole vector with memset or with an access of a different
938 if (ValSize == VecSize) {
939 // If the two types have the same primitive size, use a bit cast.
940 // Otherwise, it is two vectors with the same element type that has
941 // the same allocation size but different number of elements so use
943 if (VTy->getPrimitiveSizeInBits() ==
944 SV->getType()->getPrimitiveSizeInBits())
945 return Builder.CreateBitCast(SV, AllocaType, "tmp");
947 return CreateShuffleVectorCast(SV, VTy, Builder);
950 if (isPowerOf2_64(VecSize / ValSize)) {
951 assert(!(SV->getType()->isVectorTy() && Offset != 0) && "Can't insert a "
952 "value of a smaller vector type at a nonzero offset.");
954 const Type *CastElementTy = getScaledElementType(VTy, SV->getType(),
956 unsigned NumCastVectorElements = VecSize / ValSize;
958 LLVMContext &Context = SV->getContext();
959 const Type *OldCastTy = VectorType::get(CastElementTy,
960 NumCastVectorElements);
961 Value *OldCast = Builder.CreateBitCast(Old, OldCastTy, "tmp");
963 Value *SVCast = Builder.CreateBitCast(SV, CastElementTy, "tmp");
965 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy);
966 unsigned Elt = Offset/EltSize;
967 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
969 Builder.CreateInsertElement(OldCast, SVCast, ConstantInt::get(
970 Type::getInt32Ty(Context), Elt), "tmp");
971 return Builder.CreateBitCast(Insert, AllocaType, "tmp");
974 // Must be an element insertion.
975 assert(SV->getType() == VTy->getElementType());
976 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
977 unsigned Elt = Offset/EltSize;
978 return Builder.CreateInsertElement(Old, SV,
979 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt),
983 // If SV is a first-class aggregate value, insert each value recursively.
984 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) {
985 const StructLayout &Layout = *TD.getStructLayout(ST);
986 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
987 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
988 Old = ConvertScalar_InsertValue(Elt, Old,
989 Offset+Layout.getElementOffsetInBits(i),
995 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
996 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
997 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
998 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
999 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
1004 // If SV is a float, convert it to the appropriate integer type.
1005 // If it is a pointer, do the same.
1006 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType());
1007 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType);
1008 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType());
1009 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType);
1010 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
1011 SV = Builder.CreateBitCast(SV,
1012 IntegerType::get(SV->getContext(),SrcWidth), "tmp");
1013 else if (SV->getType()->isPointerTy())
1014 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp");
1016 // Zero extend or truncate the value if needed.
1017 if (SV->getType() != AllocaType) {
1018 if (SV->getType()->getPrimitiveSizeInBits() <
1019 AllocaType->getPrimitiveSizeInBits())
1020 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
1022 // Truncation may be needed if storing more than the alloca can hold
1023 // (undefined behavior).
1024 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
1025 SrcWidth = DestWidth;
1026 SrcStoreWidth = DestStoreWidth;
1030 // If this is a big-endian system and the store is narrower than the
1031 // full alloca type, we need to do a shift to get the right bits.
1033 if (TD.isBigEndian()) {
1034 // On big-endian machines, the lowest bit is stored at the bit offset
1035 // from the pointer given by getTypeStoreSizeInBits. This matters for
1036 // integers with a bitwidth that is not a multiple of 8.
1037 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
1042 // Note: we support negative bitwidths (with shr) which are not defined.
1043 // We do this to support (f.e.) stores off the end of a structure where
1044 // only some bits in the structure are set.
1045 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
1046 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1047 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(),
1050 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1051 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(),
1053 Mask = Mask.lshr(-ShAmt);
1056 // Mask out the bits we are about to insert from the old value, and or
1058 if (SrcWidth != DestWidth) {
1059 assert(DestWidth > SrcWidth);
1060 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
1061 SV = Builder.CreateOr(Old, SV, "ins");
1067 //===----------------------------------------------------------------------===//
1069 //===----------------------------------------------------------------------===//
1072 bool SROA::runOnFunction(Function &F) {
1073 TD = getAnalysisIfAvailable<TargetData>();
1075 bool Changed = performPromotion(F);
1077 // FIXME: ScalarRepl currently depends on TargetData more than it
1078 // theoretically needs to. It should be refactored in order to support
1079 // target-independent IR. Until this is done, just skip the actual
1080 // scalar-replacement portion of this pass.
1081 if (!TD) return Changed;
1084 bool LocalChange = performScalarRepl(F);
1085 if (!LocalChange) break; // No need to repromote if no scalarrepl
1087 LocalChange = performPromotion(F);
1088 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
1095 class AllocaPromoter : public LoadAndStorePromoter {
1098 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1099 DbgDeclareInst *DD, DIBuilder *&DB)
1100 : LoadAndStorePromoter(Insts, S, DD, DB), AI(0) {}
1102 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) {
1103 // Remember which alloca we're promoting (for isInstInList).
1105 LoadAndStorePromoter::run(Insts);
1106 AI->eraseFromParent();
1109 virtual bool isInstInList(Instruction *I,
1110 const SmallVectorImpl<Instruction*> &Insts) const {
1111 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1112 return LI->getOperand(0) == AI;
1113 return cast<StoreInst>(I)->getPointerOperand() == AI;
1116 } // end anon namespace
1118 /// isSafeSelectToSpeculate - Select instructions that use an alloca and are
1119 /// subsequently loaded can be rewritten to load both input pointers and then
1120 /// select between the result, allowing the load of the alloca to be promoted.
1122 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1123 /// %V = load i32* %P2
1125 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1126 /// %V2 = load i32* %Other
1127 /// %V = select i1 %cond, i32 %V1, i32 %V2
1129 /// We can do this to a select if its only uses are loads and if the operand to
1130 /// the select can be loaded unconditionally.
1131 static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) {
1132 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer();
1133 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer();
1135 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end();
1137 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1138 if (LI == 0 || LI->isVolatile()) return false;
1140 // Both operands to the select need to be dereferencable, either absolutely
1141 // (e.g. allocas) or at this point because we can see other accesses to it.
1142 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI,
1143 LI->getAlignment(), TD))
1145 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI,
1146 LI->getAlignment(), TD))
1153 /// isSafePHIToSpeculate - PHI instructions that use an alloca and are
1154 /// subsequently loaded can be rewritten to load both input pointers in the pred
1155 /// blocks and then PHI the results, allowing the load of the alloca to be
1158 /// %P2 = phi [i32* %Alloca, i32* %Other]
1159 /// %V = load i32* %P2
1161 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1163 /// %V2 = load i32* %Other
1165 /// %V = phi [i32 %V1, i32 %V2]
1167 /// We can do this to a select if its only uses are loads and if the operand to
1168 /// the select can be loaded unconditionally.
1169 static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) {
1170 // For now, we can only do this promotion if the load is in the same block as
1171 // the PHI, and if there are no stores between the phi and load.
1172 // TODO: Allow recursive phi users.
1173 // TODO: Allow stores.
1174 BasicBlock *BB = PN->getParent();
1175 unsigned MaxAlign = 0;
1176 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end();
1178 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1179 if (LI == 0 || LI->isVolatile()) return false;
1181 // For now we only allow loads in the same block as the PHI. This is a
1182 // common case that happens when instcombine merges two loads through a PHI.
1183 if (LI->getParent() != BB) return false;
1185 // Ensure that there are no instructions between the PHI and the load that
1187 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI)
1188 if (BBI->mayWriteToMemory())
1191 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1194 // Okay, we know that we have one or more loads in the same block as the PHI.
1195 // We can transform this if it is safe to push the loads into the predecessor
1196 // blocks. The only thing to watch out for is that we can't put a possibly
1197 // trapping load in the predecessor if it is a critical edge.
1198 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1199 BasicBlock *Pred = PN->getIncomingBlock(i);
1201 // If the predecessor has a single successor, then the edge isn't critical.
1202 if (Pred->getTerminator()->getNumSuccessors() == 1)
1205 Value *InVal = PN->getIncomingValue(i);
1207 // If the InVal is an invoke in the pred, we can't put a load on the edge.
1208 if (InvokeInst *II = dyn_cast<InvokeInst>(InVal))
1209 if (II->getParent() == Pred)
1212 // If this pointer is always safe to load, or if we can prove that there is
1213 // already a load in the block, then we can move the load to the pred block.
1214 if (InVal->isDereferenceablePointer() ||
1215 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD))
1225 /// tryToMakeAllocaBePromotable - This returns true if the alloca only has
1226 /// direct (non-volatile) loads and stores to it. If the alloca is close but
1227 /// not quite there, this will transform the code to allow promotion. As such,
1228 /// it is a non-pure predicate.
1229 static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) {
1230 SetVector<Instruction*, SmallVector<Instruction*, 4>,
1231 SmallPtrSet<Instruction*, 4> > InstsToRewrite;
1233 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
1236 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1237 if (LI->isVolatile())
1242 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1243 if (SI->getOperand(0) == AI || SI->isVolatile())
1244 return false; // Don't allow a store OF the AI, only INTO the AI.
1248 if (SelectInst *SI = dyn_cast<SelectInst>(U)) {
1249 // If the condition being selected on is a constant, fold the select, yes
1250 // this does (rarely) happen early on.
1251 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) {
1252 Value *Result = SI->getOperand(1+CI->isZero());
1253 SI->replaceAllUsesWith(Result);
1254 SI->eraseFromParent();
1256 // This is very rare and we just scrambled the use list of AI, start
1258 return tryToMakeAllocaBePromotable(AI, TD);
1261 // If it is safe to turn "load (select c, AI, ptr)" into a select of two
1262 // loads, then we can transform this by rewriting the select.
1263 if (!isSafeSelectToSpeculate(SI, TD))
1266 InstsToRewrite.insert(SI);
1270 if (PHINode *PN = dyn_cast<PHINode>(U)) {
1271 if (PN->use_empty()) { // Dead PHIs can be stripped.
1272 InstsToRewrite.insert(PN);
1276 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads
1277 // in the pred blocks, then we can transform this by rewriting the PHI.
1278 if (!isSafePHIToSpeculate(PN, TD))
1281 InstsToRewrite.insert(PN);
1288 // If there are no instructions to rewrite, then all uses are load/stores and
1290 if (InstsToRewrite.empty())
1293 // If we have instructions that need to be rewritten for this to be promotable
1294 // take care of it now.
1295 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) {
1296 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) {
1297 // Selects in InstsToRewrite only have load uses. Rewrite each as two
1298 // loads with a new select.
1299 while (!SI->use_empty()) {
1300 LoadInst *LI = cast<LoadInst>(SI->use_back());
1302 IRBuilder<> Builder(LI);
1303 LoadInst *TrueLoad =
1304 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t");
1305 LoadInst *FalseLoad =
1306 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f");
1308 // Transfer alignment and TBAA info if present.
1309 TrueLoad->setAlignment(LI->getAlignment());
1310 FalseLoad->setAlignment(LI->getAlignment());
1311 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1312 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1313 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1316 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad);
1318 LI->replaceAllUsesWith(V);
1319 LI->eraseFromParent();
1322 // Now that all the loads are gone, the select is gone too.
1323 SI->eraseFromParent();
1327 // Otherwise, we have a PHI node which allows us to push the loads into the
1329 PHINode *PN = cast<PHINode>(InstsToRewrite[i]);
1330 if (PN->use_empty()) {
1331 PN->eraseFromParent();
1335 const Type *LoadTy = cast<PointerType>(PN->getType())->getElementType();
1336 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(),
1337 PN->getName()+".ld", PN);
1339 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1340 // matter which one we get and if any differ, it doesn't matter.
1341 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back());
1342 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1343 unsigned Align = SomeLoad->getAlignment();
1345 // Rewrite all loads of the PN to use the new PHI.
1346 while (!PN->use_empty()) {
1347 LoadInst *LI = cast<LoadInst>(PN->use_back());
1348 LI->replaceAllUsesWith(NewPN);
1349 LI->eraseFromParent();
1352 // Inject loads into all of the pred blocks. Keep track of which blocks we
1353 // insert them into in case we have multiple edges from the same block.
1354 DenseMap<BasicBlock*, LoadInst*> InsertedLoads;
1356 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1357 BasicBlock *Pred = PN->getIncomingBlock(i);
1358 LoadInst *&Load = InsertedLoads[Pred];
1360 Load = new LoadInst(PN->getIncomingValue(i),
1361 PN->getName() + "." + Pred->getName(),
1362 Pred->getTerminator());
1363 Load->setAlignment(Align);
1364 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1367 NewPN->addIncoming(Load, Pred);
1370 PN->eraseFromParent();
1377 bool SROA::performPromotion(Function &F) {
1378 std::vector<AllocaInst*> Allocas;
1379 DominatorTree *DT = 0;
1381 DT = &getAnalysis<DominatorTree>();
1383 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
1385 bool Changed = false;
1386 SmallVector<Instruction*, 64> Insts;
1391 // Find allocas that are safe to promote, by looking at all instructions in
1393 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
1394 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
1395 if (tryToMakeAllocaBePromotable(AI, TD))
1396 Allocas.push_back(AI);
1398 if (Allocas.empty()) break;
1401 PromoteMemToReg(Allocas, *DT);
1404 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
1405 AllocaInst *AI = Allocas[i];
1407 // Build list of instructions to promote.
1408 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
1410 Insts.push_back(cast<Instruction>(*UI));
1412 DbgDeclareInst *DDI = FindAllocaDbgDeclare(AI);
1414 DIB = new DIBuilder(*AI->getParent()->getParent()->getParent());
1415 AllocaPromoter(Insts, SSA, DDI, DIB).run(AI, Insts);
1419 NumPromoted += Allocas.size();
1423 // FIXME: Is there a better way to handle the lazy initialization of DIB
1424 // so that there doesn't need to be an explicit delete?
1431 /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for
1432 /// SROA. It must be a struct or array type with a small number of elements.
1433 static bool ShouldAttemptScalarRepl(AllocaInst *AI) {
1434 const Type *T = AI->getAllocatedType();
1435 // Do not promote any struct into more than 32 separate vars.
1436 if (const StructType *ST = dyn_cast<StructType>(T))
1437 return ST->getNumElements() <= 32;
1438 // Arrays are much less likely to be safe for SROA; only consider
1439 // them if they are very small.
1440 if (const ArrayType *AT = dyn_cast<ArrayType>(T))
1441 return AT->getNumElements() <= 8;
1446 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
1447 // which runs on all of the alloca instructions in the function, removing them
1448 // if they are only used by getelementptr instructions.
1450 bool SROA::performScalarRepl(Function &F) {
1451 std::vector<AllocaInst*> WorkList;
1453 // Scan the entry basic block, adding allocas to the worklist.
1454 BasicBlock &BB = F.getEntryBlock();
1455 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
1456 if (AllocaInst *A = dyn_cast<AllocaInst>(I))
1457 WorkList.push_back(A);
1459 // Process the worklist
1460 bool Changed = false;
1461 while (!WorkList.empty()) {
1462 AllocaInst *AI = WorkList.back();
1463 WorkList.pop_back();
1465 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
1466 // with unused elements.
1467 if (AI->use_empty()) {
1468 AI->eraseFromParent();
1473 // If this alloca is impossible for us to promote, reject it early.
1474 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
1477 // Check to see if this allocation is only modified by a memcpy/memmove from
1478 // a constant global. If this is the case, we can change all users to use
1479 // the constant global instead. This is commonly produced by the CFE by
1480 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
1481 // is only subsequently read.
1482 SmallVector<Instruction *, 4> ToDelete;
1483 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(AI, ToDelete)) {
1484 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n');
1485 DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
1486 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
1487 ToDelete[i]->eraseFromParent();
1488 Constant *TheSrc = cast<Constant>(Copy->getSource());
1489 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
1490 Copy->eraseFromParent(); // Don't mutate the global.
1491 AI->eraseFromParent();
1497 // Check to see if we can perform the core SROA transformation. We cannot
1498 // transform the allocation instruction if it is an array allocation
1499 // (allocations OF arrays are ok though), and an allocation of a scalar
1500 // value cannot be decomposed at all.
1501 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
1503 // Do not promote [0 x %struct].
1504 if (AllocaSize == 0) continue;
1506 // Do not promote any struct whose size is too big.
1507 if (AllocaSize > SRThreshold) continue;
1509 // If the alloca looks like a good candidate for scalar replacement, and if
1510 // all its users can be transformed, then split up the aggregate into its
1511 // separate elements.
1512 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) {
1513 DoScalarReplacement(AI, WorkList);
1518 // If we can turn this aggregate value (potentially with casts) into a
1519 // simple scalar value that can be mem2reg'd into a register value.
1520 // IsNotTrivial tracks whether this is something that mem2reg could have
1521 // promoted itself. If so, we don't want to transform it needlessly. Note
1522 // that we can't just check based on the type: the alloca may be of an i32
1523 // but that has pointer arithmetic to set byte 3 of it or something.
1524 if (AllocaInst *NewAI =
1525 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) {
1526 NewAI->takeName(AI);
1527 AI->eraseFromParent();
1533 // Otherwise, couldn't process this alloca.
1539 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
1540 /// predicate, do SROA now.
1541 void SROA::DoScalarReplacement(AllocaInst *AI,
1542 std::vector<AllocaInst*> &WorkList) {
1543 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n');
1544 SmallVector<AllocaInst*, 32> ElementAllocas;
1545 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
1546 ElementAllocas.reserve(ST->getNumContainedTypes());
1547 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
1548 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
1550 AI->getName() + "." + Twine(i), AI);
1551 ElementAllocas.push_back(NA);
1552 WorkList.push_back(NA); // Add to worklist for recursive processing
1555 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
1556 ElementAllocas.reserve(AT->getNumElements());
1557 const Type *ElTy = AT->getElementType();
1558 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1559 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
1560 AI->getName() + "." + Twine(i), AI);
1561 ElementAllocas.push_back(NA);
1562 WorkList.push_back(NA); // Add to worklist for recursive processing
1566 // Now that we have created the new alloca instructions, rewrite all the
1567 // uses of the old alloca.
1568 RewriteForScalarRepl(AI, AI, 0, ElementAllocas);
1570 // Now erase any instructions that were made dead while rewriting the alloca.
1571 DeleteDeadInstructions();
1572 AI->eraseFromParent();
1577 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list,
1578 /// recursively including all their operands that become trivially dead.
1579 void SROA::DeleteDeadInstructions() {
1580 while (!DeadInsts.empty()) {
1581 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val());
1583 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
1584 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
1585 // Zero out the operand and see if it becomes trivially dead.
1586 // (But, don't add allocas to the dead instruction list -- they are
1587 // already on the worklist and will be deleted separately.)
1589 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U))
1590 DeadInsts.push_back(U);
1593 I->eraseFromParent();
1597 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to
1598 /// performing scalar replacement of alloca AI. The results are flagged in
1599 /// the Info parameter. Offset indicates the position within AI that is
1600 /// referenced by this instruction.
1601 void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset,
1603 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1604 Instruction *User = cast<Instruction>(*UI);
1606 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1607 isSafeForScalarRepl(BC, Offset, Info);
1608 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1609 uint64_t GEPOffset = Offset;
1610 isSafeGEP(GEPI, GEPOffset, Info);
1612 isSafeForScalarRepl(GEPI, GEPOffset, Info);
1613 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1614 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1616 return MarkUnsafe(Info, User);
1617 isSafeMemAccess(Offset, Length->getZExtValue(), 0,
1618 UI.getOperandNo() == 0, Info, MI,
1619 true /*AllowWholeAccess*/);
1620 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1621 if (LI->isVolatile())
1622 return MarkUnsafe(Info, User);
1623 const Type *LIType = LI->getType();
1624 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
1625 LIType, false, Info, LI, true /*AllowWholeAccess*/);
1626 Info.hasALoadOrStore = true;
1628 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1629 // Store is ok if storing INTO the pointer, not storing the pointer
1630 if (SI->isVolatile() || SI->getOperand(0) == I)
1631 return MarkUnsafe(Info, User);
1633 const Type *SIType = SI->getOperand(0)->getType();
1634 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
1635 SIType, true, Info, SI, true /*AllowWholeAccess*/);
1636 Info.hasALoadOrStore = true;
1637 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1638 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1640 return MarkUnsafe(Info, User);
1642 if (Info.isUnsafe) return;
1647 /// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer
1648 /// derived from the alloca, we can often still split the alloca into elements.
1649 /// This is useful if we have a large alloca where one element is phi'd
1650 /// together somewhere: we can SRoA and promote all the other elements even if
1651 /// we end up not being able to promote this one.
1653 /// All we require is that the uses of the PHI do not index into other parts of
1654 /// the alloca. The most important use case for this is single load and stores
1655 /// that are PHI'd together, which can happen due to code sinking.
1656 void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset,
1658 // If we've already checked this PHI, don't do it again.
1659 if (PHINode *PN = dyn_cast<PHINode>(I))
1660 if (!Info.CheckedPHIs.insert(PN))
1663 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1664 Instruction *User = cast<Instruction>(*UI);
1666 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1667 isSafePHISelectUseForScalarRepl(BC, Offset, Info);
1668 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1669 // Only allow "bitcast" GEPs for simplicity. We could generalize this,
1670 // but would have to prove that we're staying inside of an element being
1672 if (!GEPI->hasAllZeroIndices())
1673 return MarkUnsafe(Info, User);
1674 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info);
1675 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1676 if (LI->isVolatile())
1677 return MarkUnsafe(Info, User);
1678 const Type *LIType = LI->getType();
1679 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
1680 LIType, false, Info, LI, false /*AllowWholeAccess*/);
1681 Info.hasALoadOrStore = true;
1683 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1684 // Store is ok if storing INTO the pointer, not storing the pointer
1685 if (SI->isVolatile() || SI->getOperand(0) == I)
1686 return MarkUnsafe(Info, User);
1688 const Type *SIType = SI->getOperand(0)->getType();
1689 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
1690 SIType, true, Info, SI, false /*AllowWholeAccess*/);
1691 Info.hasALoadOrStore = true;
1692 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1693 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1695 return MarkUnsafe(Info, User);
1697 if (Info.isUnsafe) return;
1701 /// isSafeGEP - Check if a GEP instruction can be handled for scalar
1702 /// replacement. It is safe when all the indices are constant, in-bounds
1703 /// references, and when the resulting offset corresponds to an element within
1704 /// the alloca type. The results are flagged in the Info parameter. Upon
1705 /// return, Offset is adjusted as specified by the GEP indices.
1706 void SROA::isSafeGEP(GetElementPtrInst *GEPI,
1707 uint64_t &Offset, AllocaInfo &Info) {
1708 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
1712 // Walk through the GEP type indices, checking the types that this indexes
1714 for (; GEPIt != E; ++GEPIt) {
1715 // Ignore struct elements, no extra checking needed for these.
1716 if ((*GEPIt)->isStructTy())
1719 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
1721 return MarkUnsafe(Info, GEPI);
1724 // Compute the offset due to this GEP and check if the alloca has a
1725 // component element at that offset.
1726 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1727 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
1728 &Indices[0], Indices.size());
1729 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0))
1730 MarkUnsafe(Info, GEPI);
1733 /// isHomogeneousAggregate - Check if type T is a struct or array containing
1734 /// elements of the same type (which is always true for arrays). If so,
1735 /// return true with NumElts and EltTy set to the number of elements and the
1736 /// element type, respectively.
1737 static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts,
1738 const Type *&EltTy) {
1739 if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
1740 NumElts = AT->getNumElements();
1741 EltTy = (NumElts == 0 ? 0 : AT->getElementType());
1744 if (const StructType *ST = dyn_cast<StructType>(T)) {
1745 NumElts = ST->getNumContainedTypes();
1746 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0));
1747 for (unsigned n = 1; n < NumElts; ++n) {
1748 if (ST->getContainedType(n) != EltTy)
1756 /// isCompatibleAggregate - Check if T1 and T2 are either the same type or are
1757 /// "homogeneous" aggregates with the same element type and number of elements.
1758 static bool isCompatibleAggregate(const Type *T1, const Type *T2) {
1762 unsigned NumElts1, NumElts2;
1763 const Type *EltTy1, *EltTy2;
1764 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) &&
1765 isHomogeneousAggregate(T2, NumElts2, EltTy2) &&
1766 NumElts1 == NumElts2 &&
1773 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI
1774 /// alloca or has an offset and size that corresponds to a component element
1775 /// within it. The offset checked here may have been formed from a GEP with a
1776 /// pointer bitcasted to a different type.
1778 /// If AllowWholeAccess is true, then this allows uses of the entire alloca as a
1779 /// unit. If false, it only allows accesses known to be in a single element.
1780 void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
1781 const Type *MemOpType, bool isStore,
1782 AllocaInfo &Info, Instruction *TheAccess,
1783 bool AllowWholeAccess) {
1784 // Check if this is a load/store of the entire alloca.
1785 if (Offset == 0 && AllowWholeAccess &&
1786 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) {
1787 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer
1788 // loads/stores (which are essentially the same as the MemIntrinsics with
1789 // regard to copying padding between elements). But, if an alloca is
1790 // flagged as both a source and destination of such operations, we'll need
1791 // to check later for padding between elements.
1792 if (!MemOpType || MemOpType->isIntegerTy()) {
1794 Info.isMemCpyDst = true;
1796 Info.isMemCpySrc = true;
1799 // This is also safe for references using a type that is compatible with
1800 // the type of the alloca, so that loads/stores can be rewritten using
1801 // insertvalue/extractvalue.
1802 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) {
1803 Info.hasSubelementAccess = true;
1807 // Check if the offset/size correspond to a component within the alloca type.
1808 const Type *T = Info.AI->getAllocatedType();
1809 if (TypeHasComponent(T, Offset, MemSize)) {
1810 Info.hasSubelementAccess = true;
1814 return MarkUnsafe(Info, TheAccess);
1817 /// TypeHasComponent - Return true if T has a component type with the
1818 /// specified offset and size. If Size is zero, do not check the size.
1819 bool SROA::TypeHasComponent(const Type *T, uint64_t Offset, uint64_t Size) {
1822 if (const StructType *ST = dyn_cast<StructType>(T)) {
1823 const StructLayout *Layout = TD->getStructLayout(ST);
1824 unsigned EltIdx = Layout->getElementContainingOffset(Offset);
1825 EltTy = ST->getContainedType(EltIdx);
1826 EltSize = TD->getTypeAllocSize(EltTy);
1827 Offset -= Layout->getElementOffset(EltIdx);
1828 } else if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
1829 EltTy = AT->getElementType();
1830 EltSize = TD->getTypeAllocSize(EltTy);
1831 if (Offset >= AT->getNumElements() * EltSize)
1837 if (Offset == 0 && (Size == 0 || EltSize == Size))
1839 // Check if the component spans multiple elements.
1840 if (Offset + Size > EltSize)
1842 return TypeHasComponent(EltTy, Offset, Size);
1845 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite
1846 /// the instruction I, which references it, to use the separate elements.
1847 /// Offset indicates the position within AI that is referenced by this
1849 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
1850 SmallVector<AllocaInst*, 32> &NewElts) {
1851 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) {
1852 Use &TheUse = UI.getUse();
1853 Instruction *User = cast<Instruction>(*UI++);
1855 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1856 RewriteBitCast(BC, AI, Offset, NewElts);
1860 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1861 RewriteGEP(GEPI, AI, Offset, NewElts);
1865 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1866 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1867 uint64_t MemSize = Length->getZExtValue();
1869 MemSize == TD->getTypeAllocSize(AI->getAllocatedType()))
1870 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
1871 // Otherwise the intrinsic can only touch a single element and the
1872 // address operand will be updated, so nothing else needs to be done.
1876 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1877 const Type *LIType = LI->getType();
1879 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) {
1881 // %res = load { i32, i32 }* %alloc
1883 // %load.0 = load i32* %alloc.0
1884 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
1885 // %load.1 = load i32* %alloc.1
1886 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
1887 // (Also works for arrays instead of structs)
1888 Value *Insert = UndefValue::get(LIType);
1889 IRBuilder<> Builder(LI);
1890 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1891 Value *Load = Builder.CreateLoad(NewElts[i], "load");
1892 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert");
1894 LI->replaceAllUsesWith(Insert);
1895 DeadInsts.push_back(LI);
1896 } else if (LIType->isIntegerTy() &&
1897 TD->getTypeAllocSize(LIType) ==
1898 TD->getTypeAllocSize(AI->getAllocatedType())) {
1899 // If this is a load of the entire alloca to an integer, rewrite it.
1900 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
1905 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1906 Value *Val = SI->getOperand(0);
1907 const Type *SIType = Val->getType();
1908 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) {
1910 // store { i32, i32 } %val, { i32, i32 }* %alloc
1912 // %val.0 = extractvalue { i32, i32 } %val, 0
1913 // store i32 %val.0, i32* %alloc.0
1914 // %val.1 = extractvalue { i32, i32 } %val, 1
1915 // store i32 %val.1, i32* %alloc.1
1916 // (Also works for arrays instead of structs)
1917 IRBuilder<> Builder(SI);
1918 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1919 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName());
1920 Builder.CreateStore(Extract, NewElts[i]);
1922 DeadInsts.push_back(SI);
1923 } else if (SIType->isIntegerTy() &&
1924 TD->getTypeAllocSize(SIType) ==
1925 TD->getTypeAllocSize(AI->getAllocatedType())) {
1926 // If this is a store of the entire alloca from an integer, rewrite it.
1927 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
1932 if (isa<SelectInst>(User) || isa<PHINode>(User)) {
1933 // If we have a PHI user of the alloca itself (as opposed to a GEP or
1934 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to
1936 if (!isa<AllocaInst>(I)) continue;
1938 assert(Offset == 0 && NewElts[0] &&
1939 "Direct alloca use should have a zero offset");
1941 // If we have a use of the alloca, we know the derived uses will be
1942 // utilizing just the first element of the scalarized result. Insert a
1943 // bitcast of the first alloca before the user as required.
1944 AllocaInst *NewAI = NewElts[0];
1945 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI);
1946 NewAI->moveBefore(BCI);
1953 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced
1954 /// and recursively continue updating all of its uses.
1955 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
1956 SmallVector<AllocaInst*, 32> &NewElts) {
1957 RewriteForScalarRepl(BC, AI, Offset, NewElts);
1958 if (BC->getOperand(0) != AI)
1961 // The bitcast references the original alloca. Replace its uses with
1962 // references to the first new element alloca.
1963 Instruction *Val = NewElts[0];
1964 if (Val->getType() != BC->getDestTy()) {
1965 Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
1968 BC->replaceAllUsesWith(Val);
1969 DeadInsts.push_back(BC);
1972 /// FindElementAndOffset - Return the index of the element containing Offset
1973 /// within the specified type, which must be either a struct or an array.
1974 /// Sets T to the type of the element and Offset to the offset within that
1975 /// element. IdxTy is set to the type of the index result to be used in a
1976 /// GEP instruction.
1977 uint64_t SROA::FindElementAndOffset(const Type *&T, uint64_t &Offset,
1978 const Type *&IdxTy) {
1980 if (const StructType *ST = dyn_cast<StructType>(T)) {
1981 const StructLayout *Layout = TD->getStructLayout(ST);
1982 Idx = Layout->getElementContainingOffset(Offset);
1983 T = ST->getContainedType(Idx);
1984 Offset -= Layout->getElementOffset(Idx);
1985 IdxTy = Type::getInt32Ty(T->getContext());
1988 const ArrayType *AT = cast<ArrayType>(T);
1989 T = AT->getElementType();
1990 uint64_t EltSize = TD->getTypeAllocSize(T);
1991 Idx = Offset / EltSize;
1992 Offset -= Idx * EltSize;
1993 IdxTy = Type::getInt64Ty(T->getContext());
1997 /// RewriteGEP - Check if this GEP instruction moves the pointer across
1998 /// elements of the alloca that are being split apart, and if so, rewrite
1999 /// the GEP to be relative to the new element.
2000 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
2001 SmallVector<AllocaInst*, 32> &NewElts) {
2002 uint64_t OldOffset = Offset;
2003 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
2004 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(),
2005 &Indices[0], Indices.size());
2007 RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
2009 const Type *T = AI->getAllocatedType();
2011 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy);
2012 if (GEPI->getOperand(0) == AI)
2013 OldIdx = ~0ULL; // Force the GEP to be rewritten.
2015 T = AI->getAllocatedType();
2016 uint64_t EltOffset = Offset;
2017 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
2019 // If this GEP does not move the pointer across elements of the alloca
2020 // being split, then it does not needs to be rewritten.
2024 const Type *i32Ty = Type::getInt32Ty(AI->getContext());
2025 SmallVector<Value*, 8> NewArgs;
2026 NewArgs.push_back(Constant::getNullValue(i32Ty));
2027 while (EltOffset != 0) {
2028 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
2029 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
2031 Instruction *Val = NewElts[Idx];
2032 if (NewArgs.size() > 1) {
2033 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs.begin(),
2034 NewArgs.end(), "", GEPI);
2035 Val->takeName(GEPI);
2037 if (Val->getType() != GEPI->getType())
2038 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI);
2039 GEPI->replaceAllUsesWith(Val);
2040 DeadInsts.push_back(GEPI);
2043 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
2044 /// Rewrite it to copy or set the elements of the scalarized memory.
2045 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
2047 SmallVector<AllocaInst*, 32> &NewElts) {
2048 // If this is a memcpy/memmove, construct the other pointer as the
2049 // appropriate type. The "Other" pointer is the pointer that goes to memory
2050 // that doesn't have anything to do with the alloca that we are promoting. For
2051 // memset, this Value* stays null.
2052 Value *OtherPtr = 0;
2053 unsigned MemAlignment = MI->getAlignment();
2054 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
2055 if (Inst == MTI->getRawDest())
2056 OtherPtr = MTI->getRawSource();
2058 assert(Inst == MTI->getRawSource());
2059 OtherPtr = MTI->getRawDest();
2063 // If there is an other pointer, we want to convert it to the same pointer
2064 // type as AI has, so we can GEP through it safely.
2066 unsigned AddrSpace =
2067 cast<PointerType>(OtherPtr->getType())->getAddressSpace();
2069 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an
2070 // optimization, but it's also required to detect the corner case where
2071 // both pointer operands are referencing the same memory, and where
2072 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This
2073 // function is only called for mem intrinsics that access the whole
2074 // aggregate, so non-zero GEPs are not an issue here.)
2075 OtherPtr = OtherPtr->stripPointerCasts();
2077 // Copying the alloca to itself is a no-op: just delete it.
2078 if (OtherPtr == AI || OtherPtr == NewElts[0]) {
2079 // This code will run twice for a no-op memcpy -- once for each operand.
2080 // Put only one reference to MI on the DeadInsts list.
2081 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(),
2082 E = DeadInsts.end(); I != E; ++I)
2083 if (*I == MI) return;
2084 DeadInsts.push_back(MI);
2088 // If the pointer is not the right type, insert a bitcast to the right
2091 PointerType::get(AI->getType()->getElementType(), AddrSpace);
2093 if (OtherPtr->getType() != NewTy)
2094 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI);
2097 // Process each element of the aggregate.
2098 bool SROADest = MI->getRawDest() == Inst;
2100 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
2102 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2103 // If this is a memcpy/memmove, emit a GEP of the other element address.
2104 Value *OtherElt = 0;
2105 unsigned OtherEltAlign = MemAlignment;
2108 Value *Idx[2] = { Zero,
2109 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
2110 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx, Idx + 2,
2111 OtherPtr->getName()+"."+Twine(i),
2114 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
2115 const Type *OtherTy = OtherPtrTy->getElementType();
2116 if (const StructType *ST = dyn_cast<StructType>(OtherTy)) {
2117 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
2119 const Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
2120 EltOffset = TD->getTypeAllocSize(EltTy)*i;
2123 // The alignment of the other pointer is the guaranteed alignment of the
2124 // element, which is affected by both the known alignment of the whole
2125 // mem intrinsic and the alignment of the element. If the alignment of
2126 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
2127 // known alignment is just 4 bytes.
2128 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
2131 Value *EltPtr = NewElts[i];
2132 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
2134 // If we got down to a scalar, insert a load or store as appropriate.
2135 if (EltTy->isSingleValueType()) {
2136 if (isa<MemTransferInst>(MI)) {
2138 // From Other to Alloca.
2139 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
2140 new StoreInst(Elt, EltPtr, MI);
2142 // From Alloca to Other.
2143 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
2144 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
2148 assert(isa<MemSetInst>(MI));
2150 // If the stored element is zero (common case), just store a null
2153 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) {
2155 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
2157 // If EltTy is a vector type, get the element type.
2158 const Type *ValTy = EltTy->getScalarType();
2160 // Construct an integer with the right value.
2161 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
2162 APInt OneVal(EltSize, CI->getZExtValue());
2163 APInt TotalVal(OneVal);
2165 for (unsigned i = 0; 8*i < EltSize; ++i) {
2166 TotalVal = TotalVal.shl(8);
2170 // Convert the integer value to the appropriate type.
2171 StoreVal = ConstantInt::get(CI->getContext(), TotalVal);
2172 if (ValTy->isPointerTy())
2173 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
2174 else if (ValTy->isFloatingPointTy())
2175 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
2176 assert(StoreVal->getType() == ValTy && "Type mismatch!");
2178 // If the requested value was a vector constant, create it.
2179 if (EltTy != ValTy) {
2180 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
2181 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
2182 StoreVal = ConstantVector::get(Elts);
2185 new StoreInst(StoreVal, EltPtr, MI);
2188 // Otherwise, if we're storing a byte variable, use a memset call for
2192 unsigned EltSize = TD->getTypeAllocSize(EltTy);
2194 IRBuilder<> Builder(MI);
2196 // Finally, insert the meminst for this element.
2197 if (isa<MemSetInst>(MI)) {
2198 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize,
2201 assert(isa<MemTransferInst>(MI));
2202 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr
2203 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr
2205 if (isa<MemCpyInst>(MI))
2206 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile());
2208 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile());
2211 DeadInsts.push_back(MI);
2214 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that
2215 /// overwrites the entire allocation. Extract out the pieces of the stored
2216 /// integer and store them individually.
2217 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
2218 SmallVector<AllocaInst*, 32> &NewElts){
2219 // Extract each element out of the integer according to its structure offset
2220 // and store the element value to the individual alloca.
2221 Value *SrcVal = SI->getOperand(0);
2222 const Type *AllocaEltTy = AI->getAllocatedType();
2223 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
2225 IRBuilder<> Builder(SI);
2227 // Handle tail padding by extending the operand
2228 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
2229 SrcVal = Builder.CreateZExt(SrcVal,
2230 IntegerType::get(SI->getContext(), AllocaSizeBits));
2232 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI
2235 // There are two forms here: AI could be an array or struct. Both cases
2236 // have different ways to compute the element offset.
2237 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2238 const StructLayout *Layout = TD->getStructLayout(EltSTy);
2240 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2241 // Get the number of bits to shift SrcVal to get the value.
2242 const Type *FieldTy = EltSTy->getElementType(i);
2243 uint64_t Shift = Layout->getElementOffsetInBits(i);
2245 if (TD->isBigEndian())
2246 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
2248 Value *EltVal = SrcVal;
2250 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2251 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2254 // Truncate down to an integer of the right size.
2255 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
2257 // Ignore zero sized fields like {}, they obviously contain no data.
2258 if (FieldSizeBits == 0) continue;
2260 if (FieldSizeBits != AllocaSizeBits)
2261 EltVal = Builder.CreateTrunc(EltVal,
2262 IntegerType::get(SI->getContext(), FieldSizeBits));
2263 Value *DestField = NewElts[i];
2264 if (EltVal->getType() == FieldTy) {
2265 // Storing to an integer field of this size, just do it.
2266 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) {
2267 // Bitcast to the right element type (for fp/vector values).
2268 EltVal = Builder.CreateBitCast(EltVal, FieldTy);
2270 // Otherwise, bitcast the dest pointer (for aggregates).
2271 DestField = Builder.CreateBitCast(DestField,
2272 PointerType::getUnqual(EltVal->getType()));
2274 new StoreInst(EltVal, DestField, SI);
2278 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
2279 const Type *ArrayEltTy = ATy->getElementType();
2280 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
2281 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
2285 if (TD->isBigEndian())
2286 Shift = AllocaSizeBits-ElementOffset;
2290 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2291 // Ignore zero sized fields like {}, they obviously contain no data.
2292 if (ElementSizeBits == 0) continue;
2294 Value *EltVal = SrcVal;
2296 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2297 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2300 // Truncate down to an integer of the right size.
2301 if (ElementSizeBits != AllocaSizeBits)
2302 EltVal = Builder.CreateTrunc(EltVal,
2303 IntegerType::get(SI->getContext(),
2305 Value *DestField = NewElts[i];
2306 if (EltVal->getType() == ArrayEltTy) {
2307 // Storing to an integer field of this size, just do it.
2308 } else if (ArrayEltTy->isFloatingPointTy() ||
2309 ArrayEltTy->isVectorTy()) {
2310 // Bitcast to the right element type (for fp/vector values).
2311 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy);
2313 // Otherwise, bitcast the dest pointer (for aggregates).
2314 DestField = Builder.CreateBitCast(DestField,
2315 PointerType::getUnqual(EltVal->getType()));
2317 new StoreInst(EltVal, DestField, SI);
2319 if (TD->isBigEndian())
2320 Shift -= ElementOffset;
2322 Shift += ElementOffset;
2326 DeadInsts.push_back(SI);
2329 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to
2330 /// an integer. Load the individual pieces to form the aggregate value.
2331 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
2332 SmallVector<AllocaInst*, 32> &NewElts) {
2333 // Extract each element out of the NewElts according to its structure offset
2334 // and form the result value.
2335 const Type *AllocaEltTy = AI->getAllocatedType();
2336 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
2338 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
2341 // There are two forms here: AI could be an array or struct. Both cases
2342 // have different ways to compute the element offset.
2343 const StructLayout *Layout = 0;
2344 uint64_t ArrayEltBitOffset = 0;
2345 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2346 Layout = TD->getStructLayout(EltSTy);
2348 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
2349 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
2353 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
2355 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2356 // Load the value from the alloca. If the NewElt is an aggregate, cast
2357 // the pointer to an integer of the same size before doing the load.
2358 Value *SrcField = NewElts[i];
2359 const Type *FieldTy =
2360 cast<PointerType>(SrcField->getType())->getElementType();
2361 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
2363 // Ignore zero sized fields like {}, they obviously contain no data.
2364 if (FieldSizeBits == 0) continue;
2366 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
2368 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() &&
2369 !FieldTy->isVectorTy())
2370 SrcField = new BitCastInst(SrcField,
2371 PointerType::getUnqual(FieldIntTy),
2373 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
2375 // If SrcField is a fp or vector of the right size but that isn't an
2376 // integer type, bitcast to an integer so we can shift it.
2377 if (SrcField->getType() != FieldIntTy)
2378 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
2380 // Zero extend the field to be the same size as the final alloca so that
2381 // we can shift and insert it.
2382 if (SrcField->getType() != ResultVal->getType())
2383 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
2385 // Determine the number of bits to shift SrcField.
2387 if (Layout) // Struct case.
2388 Shift = Layout->getElementOffsetInBits(i);
2390 Shift = i*ArrayEltBitOffset;
2392 if (TD->isBigEndian())
2393 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
2396 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
2397 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
2400 // Don't create an 'or x, 0' on the first iteration.
2401 if (!isa<Constant>(ResultVal) ||
2402 !cast<Constant>(ResultVal)->isNullValue())
2403 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
2405 ResultVal = SrcField;
2408 // Handle tail padding by truncating the result
2409 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
2410 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
2412 LI->replaceAllUsesWith(ResultVal);
2413 DeadInsts.push_back(LI);
2416 /// HasPadding - Return true if the specified type has any structure or
2417 /// alignment padding in between the elements that would be split apart
2418 /// by SROA; return false otherwise.
2419 static bool HasPadding(const Type *Ty, const TargetData &TD) {
2420 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2421 Ty = ATy->getElementType();
2422 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
2425 // SROA currently handles only Arrays and Structs.
2426 const StructType *STy = cast<StructType>(Ty);
2427 const StructLayout *SL = TD.getStructLayout(STy);
2428 unsigned PrevFieldBitOffset = 0;
2429 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2430 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
2432 // Check to see if there is any padding between this element and the
2435 unsigned PrevFieldEnd =
2436 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
2437 if (PrevFieldEnd < FieldBitOffset)
2440 PrevFieldBitOffset = FieldBitOffset;
2442 // Check for tail padding.
2443 if (unsigned EltCount = STy->getNumElements()) {
2444 unsigned PrevFieldEnd = PrevFieldBitOffset +
2445 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
2446 if (PrevFieldEnd < SL->getSizeInBits())
2452 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
2453 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
2454 /// or 1 if safe after canonicalization has been performed.
2455 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) {
2456 // Loop over the use list of the alloca. We can only transform it if all of
2457 // the users are safe to transform.
2458 AllocaInfo Info(AI);
2460 isSafeForScalarRepl(AI, 0, Info);
2461 if (Info.isUnsafe) {
2462 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n');
2466 // Okay, we know all the users are promotable. If the aggregate is a memcpy
2467 // source and destination, we have to be careful. In particular, the memcpy
2468 // could be moving around elements that live in structure padding of the LLVM
2469 // types, but may actually be used. In these cases, we refuse to promote the
2471 if (Info.isMemCpySrc && Info.isMemCpyDst &&
2472 HasPadding(AI->getAllocatedType(), *TD))
2475 // If the alloca never has an access to just *part* of it, but is accessed
2476 // via loads and stores, then we should use ConvertToScalarInfo to promote
2477 // the alloca instead of promoting each piece at a time and inserting fission
2479 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) {
2480 // If the struct/array just has one element, use basic SRoA.
2481 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
2482 if (ST->getNumElements() > 1) return false;
2484 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1)
2494 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
2495 /// some part of a constant global variable. This intentionally only accepts
2496 /// constant expressions because we don't can't rewrite arbitrary instructions.
2497 static bool PointsToConstantGlobal(Value *V) {
2498 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
2499 return GV->isConstant();
2500 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
2501 if (CE->getOpcode() == Instruction::BitCast ||
2502 CE->getOpcode() == Instruction::GetElementPtr)
2503 return PointsToConstantGlobal(CE->getOperand(0));
2507 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
2508 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
2509 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
2510 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
2511 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
2512 /// the alloca, and if the source pointer is a pointer to a constant global, we
2513 /// can optimize this.
2515 isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
2517 SmallVector<Instruction *, 4> &LifetimeMarkers) {
2518 // We track lifetime intrinsics as we encounter them. If we decide to go
2519 // ahead and replace the value with the global, this lets the caller quickly
2520 // eliminate the markers.
2522 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
2523 User *U = cast<Instruction>(*UI);
2525 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
2526 // Ignore non-volatile loads, they are always ok.
2527 if (LI->isVolatile()) return false;
2531 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
2532 // If uses of the bitcast are ok, we are ok.
2533 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset,
2538 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
2539 // If the GEP has all zero indices, it doesn't offset the pointer. If it
2540 // doesn't, it does.
2541 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
2542 isOffset || !GEP->hasAllZeroIndices(),
2548 if (CallSite CS = U) {
2549 // If this is the function being called then we treat it like a load and
2551 if (CS.isCallee(UI))
2554 // If this is a readonly/readnone call site, then we know it is just a
2555 // load (but one that potentially returns the value itself), so we can
2556 // ignore it if we know that the value isn't captured.
2557 unsigned ArgNo = CS.getArgumentNo(UI);
2558 if (CS.onlyReadsMemory() &&
2559 (CS.getInstruction()->use_empty() ||
2560 CS.paramHasAttr(ArgNo+1, Attribute::NoCapture)))
2563 // If this is being passed as a byval argument, the caller is making a
2564 // copy, so it is only a read of the alloca.
2565 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal))
2569 // Lifetime intrinsics can be handled by the caller.
2570 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
2571 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
2572 II->getIntrinsicID() == Intrinsic::lifetime_end) {
2573 assert(II->use_empty() && "Lifetime markers have no result to use!");
2574 LifetimeMarkers.push_back(II);
2579 // If this is isn't our memcpy/memmove, reject it as something we can't
2581 MemTransferInst *MI = dyn_cast<MemTransferInst>(U);
2585 // If the transfer is using the alloca as a source of the transfer, then
2586 // ignore it since it is a load (unless the transfer is volatile).
2587 if (UI.getOperandNo() == 1) {
2588 if (MI->isVolatile()) return false;
2592 // If we already have seen a copy, reject the second one.
2593 if (TheCopy) return false;
2595 // If the pointer has been offset from the start of the alloca, we can't
2596 // safely handle this.
2597 if (isOffset) return false;
2599 // If the memintrinsic isn't using the alloca as the dest, reject it.
2600 if (UI.getOperandNo() != 0) return false;
2602 // If the source of the memcpy/move is not a constant global, reject it.
2603 if (!PointsToConstantGlobal(MI->getSource()))
2606 // Otherwise, the transform is safe. Remember the copy instruction.
2612 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
2613 /// modified by a copy from a constant global. If we can prove this, we can
2614 /// replace any uses of the alloca with uses of the global directly.
2616 SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI,
2617 SmallVector<Instruction*, 4> &ToDelete) {
2618 MemTransferInst *TheCopy = 0;
2619 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete))