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 they
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 #include "llvm/Transforms/Scalar.h"
23 #include "llvm/ADT/SetVector.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/Analysis/Loads.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/IR/CallSite.h"
29 #include "llvm/IR/Constants.h"
30 #include "llvm/IR/DIBuilder.h"
31 #include "llvm/IR/DataLayout.h"
32 #include "llvm/IR/DebugInfo.h"
33 #include "llvm/IR/DerivedTypes.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/GetElementPtrTypeIterator.h"
37 #include "llvm/IR/GlobalVariable.h"
38 #include "llvm/IR/IRBuilder.h"
39 #include "llvm/IR/Instructions.h"
40 #include "llvm/IR/IntrinsicInst.h"
41 #include "llvm/IR/LLVMContext.h"
42 #include "llvm/IR/Module.h"
43 #include "llvm/IR/Operator.h"
44 #include "llvm/Pass.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/ErrorHandling.h"
47 #include "llvm/Support/MathExtras.h"
48 #include "llvm/Support/raw_ostream.h"
49 #include "llvm/Transforms/Utils/Local.h"
50 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
51 #include "llvm/Transforms/Utils/SSAUpdater.h"
54 #define DEBUG_TYPE "scalarrepl"
56 STATISTIC(NumReplaced, "Number of allocas broken up");
57 STATISTIC(NumPromoted, "Number of allocas promoted");
58 STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion");
59 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
62 struct SROA : public FunctionPass {
63 SROA(int T, bool hasDT, char &ID, int ST, int AT, int SLT)
64 : FunctionPass(ID), HasDomTree(hasDT) {
70 StructMemberThreshold = 32;
72 StructMemberThreshold = ST;
74 ArrayElementThreshold = 8;
76 ArrayElementThreshold = AT;
78 // Do not limit the scalar integer load size if no threshold is given.
79 ScalarLoadThreshold = -1;
81 ScalarLoadThreshold = SLT;
84 bool runOnFunction(Function &F) override;
86 bool performScalarRepl(Function &F);
87 bool performPromotion(Function &F);
93 /// DeadInsts - Keep track of instructions we have made dead, so that
94 /// we can remove them after we are done working.
95 SmallVector<Value*, 32> DeadInsts;
97 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
98 /// information about the uses. All these fields are initialized to false
99 /// and set to true when something is learned.
101 /// The alloca to promote.
104 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite
105 /// looping and avoid redundant work.
106 SmallPtrSet<PHINode*, 8> CheckedPHIs;
108 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
111 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
112 bool isMemCpySrc : 1;
114 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
115 bool isMemCpyDst : 1;
117 /// hasSubelementAccess - This is true if a subelement of the alloca is
118 /// ever accessed, or false if the alloca is only accessed with mem
119 /// intrinsics or load/store that only access the entire alloca at once.
120 bool hasSubelementAccess : 1;
122 /// hasALoadOrStore - This is true if there are any loads or stores to it.
123 /// The alloca may just be accessed with memcpy, for example, which would
125 bool hasALoadOrStore : 1;
127 explicit AllocaInfo(AllocaInst *ai)
128 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false),
129 hasSubelementAccess(false), hasALoadOrStore(false) {}
132 /// SRThreshold - The maximum alloca size to considered for SROA.
133 unsigned SRThreshold;
135 /// StructMemberThreshold - The maximum number of members a struct can
136 /// contain to be considered for SROA.
137 unsigned StructMemberThreshold;
139 /// ArrayElementThreshold - The maximum number of elements an array can
140 /// have to be considered for SROA.
141 unsigned ArrayElementThreshold;
143 /// ScalarLoadThreshold - The maximum size in bits of scalars to load when
144 /// converting to scalar
145 unsigned ScalarLoadThreshold;
147 void MarkUnsafe(AllocaInfo &I, Instruction *User) {
149 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n');
152 bool isSafeAllocaToScalarRepl(AllocaInst *AI);
154 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info);
155 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset,
157 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info);
158 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
159 Type *MemOpType, bool isStore, AllocaInfo &Info,
160 Instruction *TheAccess, bool AllowWholeAccess);
161 bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size);
162 uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset,
165 void DoScalarReplacement(AllocaInst *AI,
166 std::vector<AllocaInst*> &WorkList);
167 void DeleteDeadInstructions();
169 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
170 SmallVectorImpl<AllocaInst *> &NewElts);
171 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
172 SmallVectorImpl<AllocaInst *> &NewElts);
173 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
174 SmallVectorImpl<AllocaInst *> &NewElts);
175 void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI,
177 SmallVectorImpl<AllocaInst *> &NewElts);
178 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
180 SmallVectorImpl<AllocaInst *> &NewElts);
181 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
182 SmallVectorImpl<AllocaInst *> &NewElts);
183 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
184 SmallVectorImpl<AllocaInst *> &NewElts);
185 bool ShouldAttemptScalarRepl(AllocaInst *AI);
188 // SROA_DT - SROA that uses DominatorTree.
189 struct SROA_DT : public SROA {
192 SROA_DT(int T = -1, int ST = -1, int AT = -1, int SLT = -1) :
193 SROA(T, true, ID, ST, AT, SLT) {
194 initializeSROA_DTPass(*PassRegistry::getPassRegistry());
197 // getAnalysisUsage - This pass does not require any passes, but we know it
198 // will not alter the CFG, so say so.
199 void getAnalysisUsage(AnalysisUsage &AU) const override {
200 AU.addRequired<DominatorTreeWrapperPass>();
201 AU.setPreservesCFG();
205 // SROA_SSAUp - SROA that uses SSAUpdater.
206 struct SROA_SSAUp : public SROA {
209 SROA_SSAUp(int T = -1, int ST = -1, int AT = -1, int SLT = -1) :
210 SROA(T, false, ID, ST, AT, SLT) {
211 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry());
214 // getAnalysisUsage - This pass does not require any passes, but we know it
215 // will not alter the CFG, so say so.
216 void getAnalysisUsage(AnalysisUsage &AU) const override {
217 AU.setPreservesCFG();
223 char SROA_DT::ID = 0;
224 char SROA_SSAUp::ID = 0;
226 INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl",
227 "Scalar Replacement of Aggregates (DT)", false, false)
228 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
229 INITIALIZE_PASS_END(SROA_DT, "scalarrepl",
230 "Scalar Replacement of Aggregates (DT)", false, false)
232 INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa",
233 "Scalar Replacement of Aggregates (SSAUp)", false, false)
234 INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa",
235 "Scalar Replacement of Aggregates (SSAUp)", false, false)
237 // Public interface to the ScalarReplAggregates pass
238 FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold,
240 int StructMemberThreshold,
241 int ArrayElementThreshold,
242 int ScalarLoadThreshold) {
244 return new SROA_DT(Threshold, StructMemberThreshold, ArrayElementThreshold,
245 ScalarLoadThreshold);
246 return new SROA_SSAUp(Threshold, StructMemberThreshold,
247 ArrayElementThreshold, ScalarLoadThreshold);
251 //===----------------------------------------------------------------------===//
252 // Convert To Scalar Optimization.
253 //===----------------------------------------------------------------------===//
256 /// ConvertToScalarInfo - This class implements the "Convert To Scalar"
257 /// optimization, which scans the uses of an alloca and determines if it can
258 /// rewrite it in terms of a single new alloca that can be mem2reg'd.
259 class ConvertToScalarInfo {
260 /// AllocaSize - The size of the alloca being considered in bytes.
262 const DataLayout &DL;
263 unsigned ScalarLoadThreshold;
265 /// IsNotTrivial - This is set to true if there is some access to the object
266 /// which means that mem2reg can't promote it.
269 /// ScalarKind - Tracks the kind of alloca being considered for promotion,
270 /// computed based on the uses of the alloca rather than the LLVM type system.
274 // Accesses via GEPs that are consistent with element access of a vector
275 // type. This will not be converted into a vector unless there is a later
276 // access using an actual vector type.
279 // Accesses via vector operations and GEPs that are consistent with the
280 // layout of a vector type.
283 // An integer bag-of-bits with bitwise operations for insertion and
284 // extraction. Any combination of types can be converted into this kind
289 /// VectorTy - This tracks the type that we should promote the vector to if
290 /// it is possible to turn it into a vector. This starts out null, and if it
291 /// isn't possible to turn into a vector type, it gets set to VoidTy.
292 VectorType *VectorTy;
294 /// HadNonMemTransferAccess - True if there is at least one access to the
295 /// alloca that is not a MemTransferInst. We don't want to turn structs into
296 /// large integers unless there is some potential for optimization.
297 bool HadNonMemTransferAccess;
299 /// HadDynamicAccess - True if some element of this alloca was dynamic.
300 /// We don't yet have support for turning a dynamic access into a large
302 bool HadDynamicAccess;
305 explicit ConvertToScalarInfo(unsigned Size, const DataLayout &DL,
307 : AllocaSize(Size), DL(DL), ScalarLoadThreshold(SLT), IsNotTrivial(false),
308 ScalarKind(Unknown), VectorTy(0), HadNonMemTransferAccess(false),
309 HadDynamicAccess(false) { }
311 AllocaInst *TryConvert(AllocaInst *AI);
314 bool CanConvertToScalar(Value *V, uint64_t Offset, Value* NonConstantIdx);
315 void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset);
316 bool MergeInVectorType(VectorType *VInTy, uint64_t Offset);
317 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset,
318 Value *NonConstantIdx);
320 Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType,
321 uint64_t Offset, Value* NonConstantIdx,
322 IRBuilder<> &Builder);
323 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
324 uint64_t Offset, Value* NonConstantIdx,
325 IRBuilder<> &Builder);
327 } // end anonymous namespace.
330 /// TryConvert - Analyze the specified alloca, and if it is safe to do so,
331 /// rewrite it to be a new alloca which is mem2reg'able. This returns the new
332 /// alloca if possible or null if not.
333 AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
334 // If we can't convert this scalar, or if mem2reg can trivially do it, bail
336 if (!CanConvertToScalar(AI, 0, 0) || !IsNotTrivial)
339 // If an alloca has only memset / memcpy uses, it may still have an Unknown
340 // ScalarKind. Treat it as an Integer below.
341 if (ScalarKind == Unknown)
342 ScalarKind = Integer;
344 if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8)
345 ScalarKind = Integer;
347 // If we were able to find a vector type that can handle this with
348 // insert/extract elements, and if there was at least one use that had
349 // a vector type, promote this to a vector. We don't want to promote
350 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
351 // we just get a lot of insert/extracts. If at least one vector is
352 // involved, then we probably really do have a union of vector/array.
354 if (ScalarKind == Vector) {
355 assert(VectorTy && "Missing type for vector scalar.");
356 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = "
357 << *VectorTy << '\n');
358 NewTy = VectorTy; // Use the vector type.
360 unsigned BitWidth = AllocaSize * 8;
362 // Do not convert to scalar integer if the alloca size exceeds the
363 // scalar load threshold.
364 if (BitWidth > ScalarLoadThreshold)
367 if ((ScalarKind == ImplicitVector || ScalarKind == Integer) &&
368 !HadNonMemTransferAccess && !DL.fitsInLegalInteger(BitWidth))
370 // Dynamic accesses on integers aren't yet supported. They need us to shift
371 // by a dynamic amount which could be difficult to work out as we might not
372 // know whether to use a left or right shift.
373 if (ScalarKind == Integer && HadDynamicAccess)
376 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
377 // Create and insert the integer alloca.
378 NewTy = IntegerType::get(AI->getContext(), BitWidth);
380 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
381 ConvertUsesToScalar(AI, NewAI, 0, 0);
385 /// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type
386 /// (VectorTy) so far at the offset specified by Offset (which is specified in
389 /// There are two cases we handle here:
390 /// 1) A union of vector types of the same size and potentially its elements.
391 /// Here we turn element accesses into insert/extract element operations.
392 /// This promotes a <4 x float> with a store of float to the third element
393 /// into a <4 x float> that uses insert element.
394 /// 2) A fully general blob of memory, which we turn into some (potentially
395 /// large) integer type with extract and insert operations where the loads
396 /// and stores would mutate the memory. We mark this by setting VectorTy
398 void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In,
400 // If we already decided to turn this into a blob of integer memory, there is
401 // nothing to be done.
402 if (ScalarKind == Integer)
405 // If this could be contributing to a vector, analyze it.
407 // If the In type is a vector that is the same size as the alloca, see if it
408 // matches the existing VecTy.
409 if (VectorType *VInTy = dyn_cast<VectorType>(In)) {
410 if (MergeInVectorType(VInTy, Offset))
412 } else if (In->isFloatTy() || In->isDoubleTy() ||
413 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 &&
414 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
415 // Full width accesses can be ignored, because they can always be turned
417 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
418 if (EltSize == AllocaSize)
421 // If we're accessing something that could be an element of a vector, see
422 // if the implied vector agrees with what we already have and if Offset is
423 // compatible with it.
424 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 &&
425 (!VectorTy || EltSize == VectorTy->getElementType()
426 ->getPrimitiveSizeInBits()/8)) {
428 ScalarKind = ImplicitVector;
429 VectorTy = VectorType::get(In, AllocaSize/EltSize);
435 // Otherwise, we have a case that we can't handle with an optimized vector
436 // form. We can still turn this into a large integer.
437 ScalarKind = Integer;
440 /// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore,
441 /// returning true if the type was successfully merged and false otherwise.
442 bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy,
444 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) {
445 // If we're storing/loading a vector of the right size, allow it as a
446 // vector. If this the first vector we see, remember the type so that
447 // we know the element size. If this is a subsequent access, ignore it
448 // even if it is a differing type but the same size. Worst case we can
449 // bitcast the resultant vectors.
459 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
460 /// its accesses to a single vector type, return true and set VecTy to
461 /// the new type. If we could convert the alloca into a single promotable
462 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
463 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
464 /// is the current offset from the base of the alloca being analyzed.
466 /// If we see at least one access to the value that is as a vector type, set the
468 bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset,
469 Value* NonConstantIdx) {
470 for (User *U : V->users()) {
471 Instruction *UI = cast<Instruction>(U);
473 if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
474 // Don't break volatile loads.
477 // Don't touch MMX operations.
478 if (LI->getType()->isX86_MMXTy())
480 HadNonMemTransferAccess = true;
481 MergeInTypeForLoadOrStore(LI->getType(), Offset);
485 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
486 // Storing the pointer, not into the value?
487 if (SI->getOperand(0) == V || !SI->isSimple()) return false;
488 // Don't touch MMX operations.
489 if (SI->getOperand(0)->getType()->isX86_MMXTy())
491 HadNonMemTransferAccess = true;
492 MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset);
496 if (BitCastInst *BCI = dyn_cast<BitCastInst>(UI)) {
497 if (!onlyUsedByLifetimeMarkers(BCI))
498 IsNotTrivial = true; // Can't be mem2reg'd.
499 if (!CanConvertToScalar(BCI, Offset, NonConstantIdx))
504 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(UI)) {
505 // If this is a GEP with a variable indices, we can't handle it.
506 PointerType* PtrTy = dyn_cast<PointerType>(GEP->getPointerOperandType());
510 // Compute the offset that this GEP adds to the pointer.
511 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
512 Value *GEPNonConstantIdx = 0;
513 if (!GEP->hasAllConstantIndices()) {
514 if (!isa<VectorType>(PtrTy->getElementType()))
518 GEPNonConstantIdx = Indices.pop_back_val();
519 if (!GEPNonConstantIdx->getType()->isIntegerTy(32))
521 HadDynamicAccess = true;
523 GEPNonConstantIdx = NonConstantIdx;
524 uint64_t GEPOffset = DL.getIndexedOffset(PtrTy,
526 // See if all uses can be converted.
527 if (!CanConvertToScalar(GEP, Offset+GEPOffset, GEPNonConstantIdx))
529 IsNotTrivial = true; // Can't be mem2reg'd.
530 HadNonMemTransferAccess = true;
534 // If this is a constant sized memset of a constant value (e.g. 0) we can
536 if (MemSetInst *MSI = dyn_cast<MemSetInst>(UI)) {
537 // Store to dynamic index.
540 // Store of constant value.
541 if (!isa<ConstantInt>(MSI->getValue()))
544 // Store of constant size.
545 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength());
549 // If the size differs from the alloca, we can only convert the alloca to
550 // an integer bag-of-bits.
551 // FIXME: This should handle all of the cases that are currently accepted
552 // as vector element insertions.
553 if (Len->getZExtValue() != AllocaSize || Offset != 0)
554 ScalarKind = Integer;
556 IsNotTrivial = true; // Can't be mem2reg'd.
557 HadNonMemTransferAccess = true;
561 // If this is a memcpy or memmove into or out of the whole allocation, we
562 // can handle it like a load or store of the scalar type.
563 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(UI)) {
564 // Store to dynamic index.
567 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength());
568 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0)
571 IsNotTrivial = true; // Can't be mem2reg'd.
575 // If this is a lifetime intrinsic, we can handle it.
576 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(UI)) {
577 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
578 II->getIntrinsicID() == Intrinsic::lifetime_end) {
583 // Otherwise, we cannot handle this!
590 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
591 /// directly. This happens when we are converting an "integer union" to a
592 /// single integer scalar, or when we are converting a "vector union" to a
593 /// vector with insert/extractelement instructions.
595 /// Offset is an offset from the original alloca, in bits that need to be
596 /// shifted to the right. By the end of this, there should be no uses of Ptr.
597 void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI,
599 Value* NonConstantIdx) {
600 while (!Ptr->use_empty()) {
601 Instruction *User = cast<Instruction>(Ptr->user_back());
603 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
604 ConvertUsesToScalar(CI, NewAI, Offset, NonConstantIdx);
605 CI->eraseFromParent();
609 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
610 // Compute the offset that this GEP adds to the pointer.
611 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
612 Value* GEPNonConstantIdx = 0;
613 if (!GEP->hasAllConstantIndices()) {
614 assert(!NonConstantIdx &&
615 "Dynamic GEP reading from dynamic GEP unsupported");
616 GEPNonConstantIdx = Indices.pop_back_val();
618 GEPNonConstantIdx = NonConstantIdx;
619 uint64_t GEPOffset = DL.getIndexedOffset(GEP->getPointerOperandType(),
621 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8, GEPNonConstantIdx);
622 GEP->eraseFromParent();
626 IRBuilder<> Builder(User);
628 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
629 // The load is a bit extract from NewAI shifted right by Offset bits.
630 Value *LoadedVal = Builder.CreateLoad(NewAI);
632 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset,
633 NonConstantIdx, Builder);
634 LI->replaceAllUsesWith(NewLoadVal);
635 LI->eraseFromParent();
639 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
640 assert(SI->getOperand(0) != Ptr && "Consistency error!");
641 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
642 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
643 NonConstantIdx, Builder);
644 Builder.CreateStore(New, NewAI);
645 SI->eraseFromParent();
647 // If the load we just inserted is now dead, then the inserted store
648 // overwrote the entire thing.
649 if (Old->use_empty())
650 Old->eraseFromParent();
654 // If this is a constant sized memset of a constant value (e.g. 0) we can
655 // transform it into a store of the expanded constant value.
656 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
657 assert(MSI->getRawDest() == Ptr && "Consistency error!");
658 assert(!NonConstantIdx && "Cannot replace dynamic memset with insert");
659 int64_t SNumBytes = cast<ConstantInt>(MSI->getLength())->getSExtValue();
660 if (SNumBytes > 0 && (SNumBytes >> 32) == 0) {
661 unsigned NumBytes = static_cast<unsigned>(SNumBytes);
662 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
664 // Compute the value replicated the right number of times.
665 APInt APVal(NumBytes*8, Val);
667 // Splat the value if non-zero.
669 for (unsigned i = 1; i != NumBytes; ++i)
672 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
673 Value *New = ConvertScalar_InsertValue(
674 ConstantInt::get(User->getContext(), APVal),
675 Old, Offset, 0, Builder);
676 Builder.CreateStore(New, NewAI);
678 // If the load we just inserted is now dead, then the memset overwrote
680 if (Old->use_empty())
681 Old->eraseFromParent();
683 MSI->eraseFromParent();
687 // If this is a memcpy or memmove into or out of the whole allocation, we
688 // can handle it like a load or store of the scalar type.
689 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
690 assert(Offset == 0 && "must be store to start of alloca");
691 assert(!NonConstantIdx && "Cannot replace dynamic transfer with insert");
693 // If the source and destination are both to the same alloca, then this is
694 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
696 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &DL, 0));
698 if (GetUnderlyingObject(MTI->getSource(), &DL, 0) != OrigAI) {
699 // Dest must be OrigAI, change this to be a load from the original
700 // pointer (bitcasted), then a store to our new alloca.
701 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
702 Value *SrcPtr = MTI->getSource();
703 PointerType* SPTy = cast<PointerType>(SrcPtr->getType());
704 PointerType* AIPTy = cast<PointerType>(NewAI->getType());
705 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
706 AIPTy = PointerType::get(AIPTy->getElementType(),
707 SPTy->getAddressSpace());
709 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy);
711 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
712 SrcVal->setAlignment(MTI->getAlignment());
713 Builder.CreateStore(SrcVal, NewAI);
714 } else if (GetUnderlyingObject(MTI->getDest(), &DL, 0) != OrigAI) {
715 // Src must be OrigAI, change this to be a load from NewAI then a store
716 // through the original dest pointer (bitcasted).
717 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
718 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
720 PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType());
721 PointerType* AIPTy = cast<PointerType>(NewAI->getType());
722 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
723 AIPTy = PointerType::get(AIPTy->getElementType(),
724 DPTy->getAddressSpace());
726 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy);
728 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
729 NewStore->setAlignment(MTI->getAlignment());
731 // Noop transfer. Src == Dst
734 MTI->eraseFromParent();
738 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
739 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
740 II->getIntrinsicID() == Intrinsic::lifetime_end) {
741 // There's no need to preserve these, as the resulting alloca will be
742 // converted to a register anyways.
743 II->eraseFromParent();
748 llvm_unreachable("Unsupported operation!");
752 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
753 /// or vector value FromVal, extracting the bits from the offset specified by
754 /// Offset. This returns the value, which is of type ToType.
756 /// This happens when we are converting an "integer union" to a single
757 /// integer scalar, or when we are converting a "vector union" to a vector with
758 /// insert/extractelement instructions.
760 /// Offset is an offset from the original alloca, in bits that need to be
761 /// shifted to the right.
762 Value *ConvertToScalarInfo::
763 ConvertScalar_ExtractValue(Value *FromVal, Type *ToType,
764 uint64_t Offset, Value* NonConstantIdx,
765 IRBuilder<> &Builder) {
766 // If the load is of the whole new alloca, no conversion is needed.
767 Type *FromType = FromVal->getType();
768 if (FromType == ToType && Offset == 0)
771 // If the result alloca is a vector type, this is either an element
772 // access or a bitcast to another vector type of the same size.
773 if (VectorType *VTy = dyn_cast<VectorType>(FromType)) {
774 unsigned FromTypeSize = DL.getTypeAllocSize(FromType);
775 unsigned ToTypeSize = DL.getTypeAllocSize(ToType);
776 if (FromTypeSize == ToTypeSize)
777 return Builder.CreateBitCast(FromVal, ToType);
779 // Otherwise it must be an element access.
782 unsigned EltSize = DL.getTypeAllocSizeInBits(VTy->getElementType());
783 Elt = Offset/EltSize;
784 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
786 // Return the element extracted out of it.
788 if (NonConstantIdx) {
790 Idx = Builder.CreateAdd(NonConstantIdx,
791 Builder.getInt32(Elt),
794 Idx = NonConstantIdx;
796 Idx = Builder.getInt32(Elt);
797 Value *V = Builder.CreateExtractElement(FromVal, Idx);
798 if (V->getType() != ToType)
799 V = Builder.CreateBitCast(V, ToType);
803 // If ToType is a first class aggregate, extract out each of the pieces and
804 // use insertvalue's to form the FCA.
805 if (StructType *ST = dyn_cast<StructType>(ToType)) {
806 assert(!NonConstantIdx &&
807 "Dynamic indexing into struct types not supported");
808 const StructLayout &Layout = *DL.getStructLayout(ST);
809 Value *Res = UndefValue::get(ST);
810 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
811 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
812 Offset+Layout.getElementOffsetInBits(i),
814 Res = Builder.CreateInsertValue(Res, Elt, i);
819 if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
820 assert(!NonConstantIdx &&
821 "Dynamic indexing into array types not supported");
822 uint64_t EltSize = DL.getTypeAllocSizeInBits(AT->getElementType());
823 Value *Res = UndefValue::get(AT);
824 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
825 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
826 Offset+i*EltSize, 0, Builder);
827 Res = Builder.CreateInsertValue(Res, Elt, i);
832 // Otherwise, this must be a union that was converted to an integer value.
833 IntegerType *NTy = cast<IntegerType>(FromVal->getType());
835 // If this is a big-endian system and the load is narrower than the
836 // full alloca type, we need to do a shift to get the right bits.
838 if (DL.isBigEndian()) {
839 // On big-endian machines, the lowest bit is stored at the bit offset
840 // from the pointer given by getTypeStoreSizeInBits. This matters for
841 // integers with a bitwidth that is not a multiple of 8.
842 ShAmt = DL.getTypeStoreSizeInBits(NTy) -
843 DL.getTypeStoreSizeInBits(ToType) - Offset;
848 // Note: we support negative bitwidths (with shl) which are not defined.
849 // We do this to support (f.e.) loads off the end of a structure where
850 // only some bits are used.
851 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
852 FromVal = Builder.CreateLShr(FromVal,
853 ConstantInt::get(FromVal->getType(), ShAmt));
854 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
855 FromVal = Builder.CreateShl(FromVal,
856 ConstantInt::get(FromVal->getType(), -ShAmt));
858 // Finally, unconditionally truncate the integer to the right width.
859 unsigned LIBitWidth = DL.getTypeSizeInBits(ToType);
860 if (LIBitWidth < NTy->getBitWidth())
862 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
864 else if (LIBitWidth > NTy->getBitWidth())
866 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
869 // If the result is an integer, this is a trunc or bitcast.
870 if (ToType->isIntegerTy()) {
872 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) {
873 // Just do a bitcast, we know the sizes match up.
874 FromVal = Builder.CreateBitCast(FromVal, ToType);
876 // Otherwise must be a pointer.
877 FromVal = Builder.CreateIntToPtr(FromVal, ToType);
879 assert(FromVal->getType() == ToType && "Didn't convert right?");
883 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
884 /// or vector value "Old" at the offset specified by Offset.
886 /// This happens when we are converting an "integer union" to a
887 /// single integer scalar, or when we are converting a "vector union" to a
888 /// vector with insert/extractelement instructions.
890 /// Offset is an offset from the original alloca, in bits that need to be
891 /// shifted to the right.
893 /// NonConstantIdx is an index value if there was a GEP with a non-constant
894 /// index value. If this is 0 then all GEPs used to find this insert address
896 Value *ConvertToScalarInfo::
897 ConvertScalar_InsertValue(Value *SV, Value *Old,
898 uint64_t Offset, Value* NonConstantIdx,
899 IRBuilder<> &Builder) {
900 // Convert the stored type to the actual type, shift it left to insert
901 // then 'or' into place.
902 Type *AllocaType = Old->getType();
903 LLVMContext &Context = Old->getContext();
905 if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
906 uint64_t VecSize = DL.getTypeAllocSizeInBits(VTy);
907 uint64_t ValSize = DL.getTypeAllocSizeInBits(SV->getType());
909 // Changing the whole vector with memset or with an access of a different
911 if (ValSize == VecSize)
912 return Builder.CreateBitCast(SV, AllocaType);
914 // Must be an element insertion.
915 Type *EltTy = VTy->getElementType();
916 if (SV->getType() != EltTy)
917 SV = Builder.CreateBitCast(SV, EltTy);
918 uint64_t EltSize = DL.getTypeAllocSizeInBits(EltTy);
919 unsigned Elt = Offset/EltSize;
921 if (NonConstantIdx) {
923 Idx = Builder.CreateAdd(NonConstantIdx,
924 Builder.getInt32(Elt),
927 Idx = NonConstantIdx;
929 Idx = Builder.getInt32(Elt);
930 return Builder.CreateInsertElement(Old, SV, Idx);
933 // If SV is a first-class aggregate value, insert each value recursively.
934 if (StructType *ST = dyn_cast<StructType>(SV->getType())) {
935 assert(!NonConstantIdx &&
936 "Dynamic indexing into struct types not supported");
937 const StructLayout &Layout = *DL.getStructLayout(ST);
938 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
939 Value *Elt = Builder.CreateExtractValue(SV, i);
940 Old = ConvertScalar_InsertValue(Elt, Old,
941 Offset+Layout.getElementOffsetInBits(i),
947 if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
948 assert(!NonConstantIdx &&
949 "Dynamic indexing into array types not supported");
950 uint64_t EltSize = DL.getTypeAllocSizeInBits(AT->getElementType());
951 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
952 Value *Elt = Builder.CreateExtractValue(SV, i);
953 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, 0, Builder);
958 // If SV is a float, convert it to the appropriate integer type.
959 // If it is a pointer, do the same.
960 unsigned SrcWidth = DL.getTypeSizeInBits(SV->getType());
961 unsigned DestWidth = DL.getTypeSizeInBits(AllocaType);
962 unsigned SrcStoreWidth = DL.getTypeStoreSizeInBits(SV->getType());
963 unsigned DestStoreWidth = DL.getTypeStoreSizeInBits(AllocaType);
964 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
965 SV = Builder.CreateBitCast(SV, IntegerType::get(SV->getContext(),SrcWidth));
966 else if (SV->getType()->isPointerTy())
967 SV = Builder.CreatePtrToInt(SV, DL.getIntPtrType(SV->getType()));
969 // Zero extend or truncate the value if needed.
970 if (SV->getType() != AllocaType) {
971 if (SV->getType()->getPrimitiveSizeInBits() <
972 AllocaType->getPrimitiveSizeInBits())
973 SV = Builder.CreateZExt(SV, AllocaType);
975 // Truncation may be needed if storing more than the alloca can hold
976 // (undefined behavior).
977 SV = Builder.CreateTrunc(SV, AllocaType);
978 SrcWidth = DestWidth;
979 SrcStoreWidth = DestStoreWidth;
983 // If this is a big-endian system and the store is narrower than the
984 // full alloca type, we need to do a shift to get the right bits.
986 if (DL.isBigEndian()) {
987 // On big-endian machines, the lowest bit is stored at the bit offset
988 // from the pointer given by getTypeStoreSizeInBits. This matters for
989 // integers with a bitwidth that is not a multiple of 8.
990 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
995 // Note: we support negative bitwidths (with shr) which are not defined.
996 // We do this to support (f.e.) stores off the end of a structure where
997 // only some bits in the structure are set.
998 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
999 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1000 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt));
1002 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1003 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt));
1004 Mask = Mask.lshr(-ShAmt);
1007 // Mask out the bits we are about to insert from the old value, and or
1009 if (SrcWidth != DestWidth) {
1010 assert(DestWidth > SrcWidth);
1011 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
1012 SV = Builder.CreateOr(Old, SV, "ins");
1018 //===----------------------------------------------------------------------===//
1020 //===----------------------------------------------------------------------===//
1023 bool SROA::runOnFunction(Function &F) {
1024 if (skipOptnoneFunction(F))
1027 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1028 DL = DLP ? &DLP->getDataLayout() : 0;
1030 bool Changed = performPromotion(F);
1032 // FIXME: ScalarRepl currently depends on DataLayout more than it
1033 // theoretically needs to. It should be refactored in order to support
1034 // target-independent IR. Until this is done, just skip the actual
1035 // scalar-replacement portion of this pass.
1036 if (!DL) return Changed;
1039 bool LocalChange = performScalarRepl(F);
1040 if (!LocalChange) break; // No need to repromote if no scalarrepl
1042 LocalChange = performPromotion(F);
1043 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
1050 class AllocaPromoter : public LoadAndStorePromoter {
1053 SmallVector<DbgDeclareInst *, 4> DDIs;
1054 SmallVector<DbgValueInst *, 4> DVIs;
1056 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1058 : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {}
1060 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) {
1061 // Remember which alloca we're promoting (for isInstInList).
1063 if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI)) {
1064 for (User *U : DebugNode->users())
1065 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
1066 DDIs.push_back(DDI);
1067 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U))
1068 DVIs.push_back(DVI);
1071 LoadAndStorePromoter::run(Insts);
1072 AI->eraseFromParent();
1073 for (SmallVectorImpl<DbgDeclareInst *>::iterator I = DDIs.begin(),
1074 E = DDIs.end(); I != E; ++I) {
1075 DbgDeclareInst *DDI = *I;
1076 DDI->eraseFromParent();
1078 for (SmallVectorImpl<DbgValueInst *>::iterator I = DVIs.begin(),
1079 E = DVIs.end(); I != E; ++I) {
1080 DbgValueInst *DVI = *I;
1081 DVI->eraseFromParent();
1085 bool isInstInList(Instruction *I,
1086 const SmallVectorImpl<Instruction*> &Insts) const override {
1087 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1088 return LI->getOperand(0) == AI;
1089 return cast<StoreInst>(I)->getPointerOperand() == AI;
1092 void updateDebugInfo(Instruction *Inst) const override {
1093 for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(),
1094 E = DDIs.end(); I != E; ++I) {
1095 DbgDeclareInst *DDI = *I;
1096 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1097 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB);
1098 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1099 ConvertDebugDeclareToDebugValue(DDI, LI, *DIB);
1101 for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(),
1102 E = DVIs.end(); I != E; ++I) {
1103 DbgValueInst *DVI = *I;
1105 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1106 // If an argument is zero extended then use argument directly. The ZExt
1107 // may be zapped by an optimization pass in future.
1108 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1109 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1110 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1111 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1113 Arg = SI->getOperand(0);
1114 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1115 Arg = LI->getOperand(0);
1119 Instruction *DbgVal =
1120 DIB->insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1122 DbgVal->setDebugLoc(DVI->getDebugLoc());
1126 } // end anon namespace
1128 /// isSafeSelectToSpeculate - Select instructions that use an alloca and are
1129 /// subsequently loaded can be rewritten to load both input pointers and then
1130 /// select between the result, allowing the load of the alloca to be promoted.
1132 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1133 /// %V = load i32* %P2
1135 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1136 /// %V2 = load i32* %Other
1137 /// %V = select i1 %cond, i32 %V1, i32 %V2
1139 /// We can do this to a select if its only uses are loads and if the operand to
1140 /// the select can be loaded unconditionally.
1141 static bool isSafeSelectToSpeculate(SelectInst *SI, const DataLayout *DL) {
1142 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer();
1143 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer();
1145 for (User *U : SI->users()) {
1146 LoadInst *LI = dyn_cast<LoadInst>(U);
1147 if (LI == 0 || !LI->isSimple()) return false;
1149 // Both operands to the select need to be dereferencable, either absolutely
1150 // (e.g. allocas) or at this point because we can see other accesses to it.
1151 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI,
1152 LI->getAlignment(), DL))
1154 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI,
1155 LI->getAlignment(), DL))
1162 /// isSafePHIToSpeculate - PHI instructions that use an alloca and are
1163 /// subsequently loaded can be rewritten to load both input pointers in the pred
1164 /// blocks and then PHI the results, allowing the load of the alloca to be
1167 /// %P2 = phi [i32* %Alloca, i32* %Other]
1168 /// %V = load i32* %P2
1170 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1172 /// %V2 = load i32* %Other
1174 /// %V = phi [i32 %V1, i32 %V2]
1176 /// We can do this to a select if its only uses are loads and if the operand to
1177 /// the select can be loaded unconditionally.
1178 static bool isSafePHIToSpeculate(PHINode *PN, const DataLayout *DL) {
1179 // For now, we can only do this promotion if the load is in the same block as
1180 // the PHI, and if there are no stores between the phi and load.
1181 // TODO: Allow recursive phi users.
1182 // TODO: Allow stores.
1183 BasicBlock *BB = PN->getParent();
1184 unsigned MaxAlign = 0;
1185 for (User *U : PN->users()) {
1186 LoadInst *LI = dyn_cast<LoadInst>(U);
1187 if (LI == 0 || !LI->isSimple()) return false;
1189 // For now we only allow loads in the same block as the PHI. This is a
1190 // common case that happens when instcombine merges two loads through a PHI.
1191 if (LI->getParent() != BB) return false;
1193 // Ensure that there are no instructions between the PHI and the load that
1195 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI)
1196 if (BBI->mayWriteToMemory())
1199 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1202 // Okay, we know that we have one or more loads in the same block as the PHI.
1203 // We can transform this if it is safe to push the loads into the predecessor
1204 // blocks. The only thing to watch out for is that we can't put a possibly
1205 // trapping load in the predecessor if it is a critical edge.
1206 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1207 BasicBlock *Pred = PN->getIncomingBlock(i);
1208 Value *InVal = PN->getIncomingValue(i);
1210 // If the terminator of the predecessor has side-effects (an invoke),
1211 // there is no safe place to put a load in the predecessor.
1212 if (Pred->getTerminator()->mayHaveSideEffects())
1215 // If the value is produced by the terminator of the predecessor
1216 // (an invoke), there is no valid place to put a load in the predecessor.
1217 if (Pred->getTerminator() == InVal)
1220 // If the predecessor has a single successor, then the edge isn't critical.
1221 if (Pred->getTerminator()->getNumSuccessors() == 1)
1224 // If this pointer is always safe to load, or if we can prove that there is
1225 // already a load in the block, then we can move the load to the pred block.
1226 if (InVal->isDereferenceablePointer() ||
1227 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, DL))
1237 /// tryToMakeAllocaBePromotable - This returns true if the alloca only has
1238 /// direct (non-volatile) loads and stores to it. If the alloca is close but
1239 /// not quite there, this will transform the code to allow promotion. As such,
1240 /// it is a non-pure predicate.
1241 static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const DataLayout *DL) {
1242 SetVector<Instruction*, SmallVector<Instruction*, 4>,
1243 SmallPtrSet<Instruction*, 4> > InstsToRewrite;
1244 for (User *U : AI->users()) {
1245 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1246 if (!LI->isSimple())
1251 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1252 if (SI->getOperand(0) == AI || !SI->isSimple())
1253 return false; // Don't allow a store OF the AI, only INTO the AI.
1257 if (SelectInst *SI = dyn_cast<SelectInst>(U)) {
1258 // If the condition being selected on is a constant, fold the select, yes
1259 // this does (rarely) happen early on.
1260 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) {
1261 Value *Result = SI->getOperand(1+CI->isZero());
1262 SI->replaceAllUsesWith(Result);
1263 SI->eraseFromParent();
1265 // This is very rare and we just scrambled the use list of AI, start
1267 return tryToMakeAllocaBePromotable(AI, DL);
1270 // If it is safe to turn "load (select c, AI, ptr)" into a select of two
1271 // loads, then we can transform this by rewriting the select.
1272 if (!isSafeSelectToSpeculate(SI, DL))
1275 InstsToRewrite.insert(SI);
1279 if (PHINode *PN = dyn_cast<PHINode>(U)) {
1280 if (PN->use_empty()) { // Dead PHIs can be stripped.
1281 InstsToRewrite.insert(PN);
1285 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads
1286 // in the pred blocks, then we can transform this by rewriting the PHI.
1287 if (!isSafePHIToSpeculate(PN, DL))
1290 InstsToRewrite.insert(PN);
1294 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
1295 if (onlyUsedByLifetimeMarkers(BCI)) {
1296 InstsToRewrite.insert(BCI);
1304 // If there are no instructions to rewrite, then all uses are load/stores and
1306 if (InstsToRewrite.empty())
1309 // If we have instructions that need to be rewritten for this to be promotable
1310 // take care of it now.
1311 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) {
1312 if (BitCastInst *BCI = dyn_cast<BitCastInst>(InstsToRewrite[i])) {
1313 // This could only be a bitcast used by nothing but lifetime intrinsics.
1314 for (BitCastInst::user_iterator I = BCI->user_begin(), E = BCI->user_end();
1316 cast<Instruction>(*I++)->eraseFromParent();
1317 BCI->eraseFromParent();
1321 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) {
1322 // Selects in InstsToRewrite only have load uses. Rewrite each as two
1323 // loads with a new select.
1324 while (!SI->use_empty()) {
1325 LoadInst *LI = cast<LoadInst>(SI->user_back());
1327 IRBuilder<> Builder(LI);
1328 LoadInst *TrueLoad =
1329 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t");
1330 LoadInst *FalseLoad =
1331 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f");
1333 // Transfer alignment and TBAA info if present.
1334 TrueLoad->setAlignment(LI->getAlignment());
1335 FalseLoad->setAlignment(LI->getAlignment());
1336 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1337 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1338 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1341 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad);
1343 LI->replaceAllUsesWith(V);
1344 LI->eraseFromParent();
1347 // Now that all the loads are gone, the select is gone too.
1348 SI->eraseFromParent();
1352 // Otherwise, we have a PHI node which allows us to push the loads into the
1354 PHINode *PN = cast<PHINode>(InstsToRewrite[i]);
1355 if (PN->use_empty()) {
1356 PN->eraseFromParent();
1360 Type *LoadTy = cast<PointerType>(PN->getType())->getElementType();
1361 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(),
1362 PN->getName()+".ld", PN);
1364 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1365 // matter which one we get and if any differ, it doesn't matter.
1366 LoadInst *SomeLoad = cast<LoadInst>(PN->user_back());
1367 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1368 unsigned Align = SomeLoad->getAlignment();
1370 // Rewrite all loads of the PN to use the new PHI.
1371 while (!PN->use_empty()) {
1372 LoadInst *LI = cast<LoadInst>(PN->user_back());
1373 LI->replaceAllUsesWith(NewPN);
1374 LI->eraseFromParent();
1377 // Inject loads into all of the pred blocks. Keep track of which blocks we
1378 // insert them into in case we have multiple edges from the same block.
1379 DenseMap<BasicBlock*, LoadInst*> InsertedLoads;
1381 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1382 BasicBlock *Pred = PN->getIncomingBlock(i);
1383 LoadInst *&Load = InsertedLoads[Pred];
1385 Load = new LoadInst(PN->getIncomingValue(i),
1386 PN->getName() + "." + Pred->getName(),
1387 Pred->getTerminator());
1388 Load->setAlignment(Align);
1389 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1392 NewPN->addIncoming(Load, Pred);
1395 PN->eraseFromParent();
1402 bool SROA::performPromotion(Function &F) {
1403 std::vector<AllocaInst*> Allocas;
1404 DominatorTree *DT = 0;
1406 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1408 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
1409 DIBuilder DIB(*F.getParent());
1410 bool Changed = false;
1411 SmallVector<Instruction*, 64> Insts;
1415 // Find allocas that are safe to promote, by looking at all instructions in
1417 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
1418 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
1419 if (tryToMakeAllocaBePromotable(AI, DL))
1420 Allocas.push_back(AI);
1422 if (Allocas.empty()) break;
1425 PromoteMemToReg(Allocas, *DT);
1428 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
1429 AllocaInst *AI = Allocas[i];
1431 // Build list of instructions to promote.
1432 for (User *U : AI->users())
1433 Insts.push_back(cast<Instruction>(U));
1434 AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts);
1438 NumPromoted += Allocas.size();
1446 /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for
1447 /// SROA. It must be a struct or array type with a small number of elements.
1448 bool SROA::ShouldAttemptScalarRepl(AllocaInst *AI) {
1449 Type *T = AI->getAllocatedType();
1450 // Do not promote any struct that has too many members.
1451 if (StructType *ST = dyn_cast<StructType>(T))
1452 return ST->getNumElements() <= StructMemberThreshold;
1453 // Do not promote any array that has too many elements.
1454 if (ArrayType *AT = dyn_cast<ArrayType>(T))
1455 return AT->getNumElements() <= ArrayElementThreshold;
1459 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
1460 // which runs on all of the alloca instructions in the entry block, removing
1461 // them if they are only used by getelementptr instructions.
1463 bool SROA::performScalarRepl(Function &F) {
1464 std::vector<AllocaInst*> WorkList;
1466 // Scan the entry basic block, adding allocas to the worklist.
1467 BasicBlock &BB = F.getEntryBlock();
1468 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
1469 if (AllocaInst *A = dyn_cast<AllocaInst>(I))
1470 WorkList.push_back(A);
1472 // Process the worklist
1473 bool Changed = false;
1474 while (!WorkList.empty()) {
1475 AllocaInst *AI = WorkList.back();
1476 WorkList.pop_back();
1478 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
1479 // with unused elements.
1480 if (AI->use_empty()) {
1481 AI->eraseFromParent();
1486 // If this alloca is impossible for us to promote, reject it early.
1487 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
1490 // Check to see if we can perform the core SROA transformation. We cannot
1491 // transform the allocation instruction if it is an array allocation
1492 // (allocations OF arrays are ok though), and an allocation of a scalar
1493 // value cannot be decomposed at all.
1494 uint64_t AllocaSize = DL->getTypeAllocSize(AI->getAllocatedType());
1496 // Do not promote [0 x %struct].
1497 if (AllocaSize == 0) continue;
1499 // Do not promote any struct whose size is too big.
1500 if (AllocaSize > SRThreshold) continue;
1502 // If the alloca looks like a good candidate for scalar replacement, and if
1503 // all its users can be transformed, then split up the aggregate into its
1504 // separate elements.
1505 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) {
1506 DoScalarReplacement(AI, WorkList);
1511 // If we can turn this aggregate value (potentially with casts) into a
1512 // simple scalar value that can be mem2reg'd into a register value.
1513 // IsNotTrivial tracks whether this is something that mem2reg could have
1514 // promoted itself. If so, we don't want to transform it needlessly. Note
1515 // that we can't just check based on the type: the alloca may be of an i32
1516 // but that has pointer arithmetic to set byte 3 of it or something.
1517 if (AllocaInst *NewAI = ConvertToScalarInfo(
1518 (unsigned)AllocaSize, *DL, ScalarLoadThreshold).TryConvert(AI)) {
1519 NewAI->takeName(AI);
1520 AI->eraseFromParent();
1526 // Otherwise, couldn't process this alloca.
1532 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
1533 /// predicate, do SROA now.
1534 void SROA::DoScalarReplacement(AllocaInst *AI,
1535 std::vector<AllocaInst*> &WorkList) {
1536 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n');
1537 SmallVector<AllocaInst*, 32> ElementAllocas;
1538 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
1539 ElementAllocas.reserve(ST->getNumContainedTypes());
1540 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
1541 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
1543 AI->getName() + "." + Twine(i), AI);
1544 ElementAllocas.push_back(NA);
1545 WorkList.push_back(NA); // Add to worklist for recursive processing
1548 ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
1549 ElementAllocas.reserve(AT->getNumElements());
1550 Type *ElTy = AT->getElementType();
1551 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1552 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
1553 AI->getName() + "." + Twine(i), AI);
1554 ElementAllocas.push_back(NA);
1555 WorkList.push_back(NA); // Add to worklist for recursive processing
1559 // Now that we have created the new alloca instructions, rewrite all the
1560 // uses of the old alloca.
1561 RewriteForScalarRepl(AI, AI, 0, ElementAllocas);
1563 // Now erase any instructions that were made dead while rewriting the alloca.
1564 DeleteDeadInstructions();
1565 AI->eraseFromParent();
1570 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list,
1571 /// recursively including all their operands that become trivially dead.
1572 void SROA::DeleteDeadInstructions() {
1573 while (!DeadInsts.empty()) {
1574 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val());
1576 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
1577 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
1578 // Zero out the operand and see if it becomes trivially dead.
1579 // (But, don't add allocas to the dead instruction list -- they are
1580 // already on the worklist and will be deleted separately.)
1582 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U))
1583 DeadInsts.push_back(U);
1586 I->eraseFromParent();
1590 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to
1591 /// performing scalar replacement of alloca AI. The results are flagged in
1592 /// the Info parameter. Offset indicates the position within AI that is
1593 /// referenced by this instruction.
1594 void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset,
1596 for (Use &U : I->uses()) {
1597 Instruction *User = cast<Instruction>(U.getUser());
1599 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1600 isSafeForScalarRepl(BC, Offset, Info);
1601 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1602 uint64_t GEPOffset = Offset;
1603 isSafeGEP(GEPI, GEPOffset, Info);
1605 isSafeForScalarRepl(GEPI, GEPOffset, Info);
1606 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1607 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1609 return MarkUnsafe(Info, User);
1610 if (Length->isNegative())
1611 return MarkUnsafe(Info, User);
1613 isSafeMemAccess(Offset, Length->getZExtValue(), 0,
1614 U.getOperandNo() == 0, Info, MI,
1615 true /*AllowWholeAccess*/);
1616 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1617 if (!LI->isSimple())
1618 return MarkUnsafe(Info, User);
1619 Type *LIType = LI->getType();
1620 isSafeMemAccess(Offset, DL->getTypeAllocSize(LIType),
1621 LIType, false, Info, LI, true /*AllowWholeAccess*/);
1622 Info.hasALoadOrStore = true;
1624 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1625 // Store is ok if storing INTO the pointer, not storing the pointer
1626 if (!SI->isSimple() || SI->getOperand(0) == I)
1627 return MarkUnsafe(Info, User);
1629 Type *SIType = SI->getOperand(0)->getType();
1630 isSafeMemAccess(Offset, DL->getTypeAllocSize(SIType),
1631 SIType, true, Info, SI, true /*AllowWholeAccess*/);
1632 Info.hasALoadOrStore = true;
1633 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
1634 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1635 II->getIntrinsicID() != Intrinsic::lifetime_end)
1636 return MarkUnsafe(Info, User);
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 (User *U : I->users()) {
1664 Instruction *UI = cast<Instruction>(U);
1666 if (BitCastInst *BC = dyn_cast<BitCastInst>(UI)) {
1667 isSafePHISelectUseForScalarRepl(BC, Offset, Info);
1668 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(UI)) {
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, UI);
1674 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info);
1675 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
1676 if (!LI->isSimple())
1677 return MarkUnsafe(Info, UI);
1678 Type *LIType = LI->getType();
1679 isSafeMemAccess(Offset, DL->getTypeAllocSize(LIType),
1680 LIType, false, Info, LI, false /*AllowWholeAccess*/);
1681 Info.hasALoadOrStore = true;
1683 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1684 // Store is ok if storing INTO the pointer, not storing the pointer
1685 if (!SI->isSimple() || SI->getOperand(0) == I)
1686 return MarkUnsafe(Info, UI);
1688 Type *SIType = SI->getOperand(0)->getType();
1689 isSafeMemAccess(Offset, DL->getTypeAllocSize(SIType),
1690 SIType, true, Info, SI, false /*AllowWholeAccess*/);
1691 Info.hasALoadOrStore = true;
1692 } else if (isa<PHINode>(UI) || isa<SelectInst>(UI)) {
1693 isSafePHISelectUseForScalarRepl(UI, Offset, Info);
1695 return MarkUnsafe(Info, UI);
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);
1711 bool NonConstant = false;
1712 unsigned NonConstantIdxSize = 0;
1714 // Walk through the GEP type indices, checking the types that this indexes
1716 for (; GEPIt != E; ++GEPIt) {
1717 // Ignore struct elements, no extra checking needed for these.
1718 if ((*GEPIt)->isStructTy())
1721 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
1723 return MarkUnsafe(Info, GEPI);
1726 // Compute the offset due to this GEP and check if the alloca has a
1727 // component element at that offset.
1728 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1729 // If this GEP is non-constant then the last operand must have been a
1730 // dynamic index into a vector. Pop this now as it has no impact on the
1731 // constant part of the offset.
1734 Offset += DL->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
1735 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset,
1736 NonConstantIdxSize))
1737 MarkUnsafe(Info, GEPI);
1740 /// isHomogeneousAggregate - Check if type T is a struct or array containing
1741 /// elements of the same type (which is always true for arrays). If so,
1742 /// return true with NumElts and EltTy set to the number of elements and the
1743 /// element type, respectively.
1744 static bool isHomogeneousAggregate(Type *T, unsigned &NumElts,
1746 if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
1747 NumElts = AT->getNumElements();
1748 EltTy = (NumElts == 0 ? 0 : AT->getElementType());
1751 if (StructType *ST = dyn_cast<StructType>(T)) {
1752 NumElts = ST->getNumContainedTypes();
1753 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0));
1754 for (unsigned n = 1; n < NumElts; ++n) {
1755 if (ST->getContainedType(n) != EltTy)
1763 /// isCompatibleAggregate - Check if T1 and T2 are either the same type or are
1764 /// "homogeneous" aggregates with the same element type and number of elements.
1765 static bool isCompatibleAggregate(Type *T1, Type *T2) {
1769 unsigned NumElts1, NumElts2;
1770 Type *EltTy1, *EltTy2;
1771 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) &&
1772 isHomogeneousAggregate(T2, NumElts2, EltTy2) &&
1773 NumElts1 == NumElts2 &&
1780 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI
1781 /// alloca or has an offset and size that corresponds to a component element
1782 /// within it. The offset checked here may have been formed from a GEP with a
1783 /// pointer bitcasted to a different type.
1785 /// If AllowWholeAccess is true, then this allows uses of the entire alloca as a
1786 /// unit. If false, it only allows accesses known to be in a single element.
1787 void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
1788 Type *MemOpType, bool isStore,
1789 AllocaInfo &Info, Instruction *TheAccess,
1790 bool AllowWholeAccess) {
1791 // Check if this is a load/store of the entire alloca.
1792 if (Offset == 0 && AllowWholeAccess &&
1793 MemSize == DL->getTypeAllocSize(Info.AI->getAllocatedType())) {
1794 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer
1795 // loads/stores (which are essentially the same as the MemIntrinsics with
1796 // regard to copying padding between elements). But, if an alloca is
1797 // flagged as both a source and destination of such operations, we'll need
1798 // to check later for padding between elements.
1799 if (!MemOpType || MemOpType->isIntegerTy()) {
1801 Info.isMemCpyDst = true;
1803 Info.isMemCpySrc = true;
1806 // This is also safe for references using a type that is compatible with
1807 // the type of the alloca, so that loads/stores can be rewritten using
1808 // insertvalue/extractvalue.
1809 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) {
1810 Info.hasSubelementAccess = true;
1814 // Check if the offset/size correspond to a component within the alloca type.
1815 Type *T = Info.AI->getAllocatedType();
1816 if (TypeHasComponent(T, Offset, MemSize)) {
1817 Info.hasSubelementAccess = true;
1821 return MarkUnsafe(Info, TheAccess);
1824 /// TypeHasComponent - Return true if T has a component type with the
1825 /// specified offset and size. If Size is zero, do not check the size.
1826 bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) {
1829 if (StructType *ST = dyn_cast<StructType>(T)) {
1830 const StructLayout *Layout = DL->getStructLayout(ST);
1831 unsigned EltIdx = Layout->getElementContainingOffset(Offset);
1832 EltTy = ST->getContainedType(EltIdx);
1833 EltSize = DL->getTypeAllocSize(EltTy);
1834 Offset -= Layout->getElementOffset(EltIdx);
1835 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
1836 EltTy = AT->getElementType();
1837 EltSize = DL->getTypeAllocSize(EltTy);
1838 if (Offset >= AT->getNumElements() * EltSize)
1841 } else if (VectorType *VT = dyn_cast<VectorType>(T)) {
1842 EltTy = VT->getElementType();
1843 EltSize = DL->getTypeAllocSize(EltTy);
1844 if (Offset >= VT->getNumElements() * EltSize)
1850 if (Offset == 0 && (Size == 0 || EltSize == Size))
1852 // Check if the component spans multiple elements.
1853 if (Offset + Size > EltSize)
1855 return TypeHasComponent(EltTy, Offset, Size);
1858 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite
1859 /// the instruction I, which references it, to use the separate elements.
1860 /// Offset indicates the position within AI that is referenced by this
1862 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
1863 SmallVectorImpl<AllocaInst *> &NewElts) {
1864 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) {
1865 Use &TheUse = *UI++;
1866 Instruction *User = cast<Instruction>(TheUse.getUser());
1868 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1869 RewriteBitCast(BC, AI, Offset, NewElts);
1873 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1874 RewriteGEP(GEPI, AI, Offset, NewElts);
1878 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1879 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1880 uint64_t MemSize = Length->getZExtValue();
1882 MemSize == DL->getTypeAllocSize(AI->getAllocatedType()))
1883 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
1884 // Otherwise the intrinsic can only touch a single element and the
1885 // address operand will be updated, so nothing else needs to be done.
1889 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
1890 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
1891 II->getIntrinsicID() == Intrinsic::lifetime_end) {
1892 RewriteLifetimeIntrinsic(II, AI, Offset, NewElts);
1897 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1898 Type *LIType = LI->getType();
1900 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) {
1902 // %res = load { i32, i32 }* %alloc
1904 // %load.0 = load i32* %alloc.0
1905 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
1906 // %load.1 = load i32* %alloc.1
1907 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
1908 // (Also works for arrays instead of structs)
1909 Value *Insert = UndefValue::get(LIType);
1910 IRBuilder<> Builder(LI);
1911 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1912 Value *Load = Builder.CreateLoad(NewElts[i], "load");
1913 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert");
1915 LI->replaceAllUsesWith(Insert);
1916 DeadInsts.push_back(LI);
1917 } else if (LIType->isIntegerTy() &&
1918 DL->getTypeAllocSize(LIType) ==
1919 DL->getTypeAllocSize(AI->getAllocatedType())) {
1920 // If this is a load of the entire alloca to an integer, rewrite it.
1921 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
1926 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1927 Value *Val = SI->getOperand(0);
1928 Type *SIType = Val->getType();
1929 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) {
1931 // store { i32, i32 } %val, { i32, i32 }* %alloc
1933 // %val.0 = extractvalue { i32, i32 } %val, 0
1934 // store i32 %val.0, i32* %alloc.0
1935 // %val.1 = extractvalue { i32, i32 } %val, 1
1936 // store i32 %val.1, i32* %alloc.1
1937 // (Also works for arrays instead of structs)
1938 IRBuilder<> Builder(SI);
1939 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
1940 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName());
1941 Builder.CreateStore(Extract, NewElts[i]);
1943 DeadInsts.push_back(SI);
1944 } else if (SIType->isIntegerTy() &&
1945 DL->getTypeAllocSize(SIType) ==
1946 DL->getTypeAllocSize(AI->getAllocatedType())) {
1947 // If this is a store of the entire alloca from an integer, rewrite it.
1948 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
1953 if (isa<SelectInst>(User) || isa<PHINode>(User)) {
1954 // If we have a PHI user of the alloca itself (as opposed to a GEP or
1955 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to
1957 if (!isa<AllocaInst>(I)) continue;
1959 assert(Offset == 0 && NewElts[0] &&
1960 "Direct alloca use should have a zero offset");
1962 // If we have a use of the alloca, we know the derived uses will be
1963 // utilizing just the first element of the scalarized result. Insert a
1964 // bitcast of the first alloca before the user as required.
1965 AllocaInst *NewAI = NewElts[0];
1966 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI);
1967 NewAI->moveBefore(BCI);
1974 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced
1975 /// and recursively continue updating all of its uses.
1976 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
1977 SmallVectorImpl<AllocaInst *> &NewElts) {
1978 RewriteForScalarRepl(BC, AI, Offset, NewElts);
1979 if (BC->getOperand(0) != AI)
1982 // The bitcast references the original alloca. Replace its uses with
1983 // references to the alloca containing offset zero (which is normally at
1984 // index zero, but might not be in cases involving structs with elements
1986 Type *T = AI->getAllocatedType();
1987 uint64_t EltOffset = 0;
1989 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
1990 Instruction *Val = NewElts[Idx];
1991 if (Val->getType() != BC->getDestTy()) {
1992 Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
1995 BC->replaceAllUsesWith(Val);
1996 DeadInsts.push_back(BC);
1999 /// FindElementAndOffset - Return the index of the element containing Offset
2000 /// within the specified type, which must be either a struct or an array.
2001 /// Sets T to the type of the element and Offset to the offset within that
2002 /// element. IdxTy is set to the type of the index result to be used in a
2003 /// GEP instruction.
2004 uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset,
2007 if (StructType *ST = dyn_cast<StructType>(T)) {
2008 const StructLayout *Layout = DL->getStructLayout(ST);
2009 Idx = Layout->getElementContainingOffset(Offset);
2010 T = ST->getContainedType(Idx);
2011 Offset -= Layout->getElementOffset(Idx);
2012 IdxTy = Type::getInt32Ty(T->getContext());
2014 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
2015 T = AT->getElementType();
2016 uint64_t EltSize = DL->getTypeAllocSize(T);
2017 Idx = Offset / EltSize;
2018 Offset -= Idx * EltSize;
2019 IdxTy = Type::getInt64Ty(T->getContext());
2022 VectorType *VT = cast<VectorType>(T);
2023 T = VT->getElementType();
2024 uint64_t EltSize = DL->getTypeAllocSize(T);
2025 Idx = Offset / EltSize;
2026 Offset -= Idx * EltSize;
2027 IdxTy = Type::getInt64Ty(T->getContext());
2031 /// RewriteGEP - Check if this GEP instruction moves the pointer across
2032 /// elements of the alloca that are being split apart, and if so, rewrite
2033 /// the GEP to be relative to the new element.
2034 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
2035 SmallVectorImpl<AllocaInst *> &NewElts) {
2036 uint64_t OldOffset = Offset;
2037 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
2038 // If the GEP was dynamic then it must have been a dynamic vector lookup.
2039 // In this case, it must be the last GEP operand which is dynamic so keep that
2040 // aside until we've found the constant GEP offset then add it back in at the
2042 Value* NonConstantIdx = 0;
2043 if (!GEPI->hasAllConstantIndices())
2044 NonConstantIdx = Indices.pop_back_val();
2045 Offset += DL->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
2047 RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
2049 Type *T = AI->getAllocatedType();
2051 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy);
2052 if (GEPI->getOperand(0) == AI)
2053 OldIdx = ~0ULL; // Force the GEP to be rewritten.
2055 T = AI->getAllocatedType();
2056 uint64_t EltOffset = Offset;
2057 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
2059 // If this GEP does not move the pointer across elements of the alloca
2060 // being split, then it does not needs to be rewritten.
2064 Type *i32Ty = Type::getInt32Ty(AI->getContext());
2065 SmallVector<Value*, 8> NewArgs;
2066 NewArgs.push_back(Constant::getNullValue(i32Ty));
2067 while (EltOffset != 0) {
2068 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
2069 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
2071 if (NonConstantIdx) {
2073 // This GEP has a dynamic index. We need to add "i32 0" to index through
2074 // any structs or arrays in the original type until we get to the vector
2076 while (!isa<VectorType>(GepTy)) {
2077 NewArgs.push_back(Constant::getNullValue(i32Ty));
2078 GepTy = cast<CompositeType>(GepTy)->getTypeAtIndex(0U);
2080 NewArgs.push_back(NonConstantIdx);
2082 Instruction *Val = NewElts[Idx];
2083 if (NewArgs.size() > 1) {
2084 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI);
2085 Val->takeName(GEPI);
2087 if (Val->getType() != GEPI->getType())
2088 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI);
2089 GEPI->replaceAllUsesWith(Val);
2090 DeadInsts.push_back(GEPI);
2093 /// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it
2094 /// to mark the lifetime of the scalarized memory.
2095 void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI,
2097 SmallVectorImpl<AllocaInst *> &NewElts) {
2098 ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0));
2099 // Put matching lifetime markers on everything from Offset up to
2101 Type *AIType = AI->getAllocatedType();
2102 uint64_t NewOffset = Offset;
2104 uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy);
2106 IRBuilder<> Builder(II);
2107 uint64_t Size = OldSize->getLimitedValue();
2110 // Splice the first element and index 'NewOffset' bytes in. SROA will
2111 // split the alloca again later.
2112 unsigned AS = AI->getType()->getAddressSpace();
2113 Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy(AS));
2114 V = Builder.CreateGEP(V, Builder.getInt64(NewOffset));
2116 IdxTy = NewElts[Idx]->getAllocatedType();
2117 uint64_t EltSize = DL->getTypeAllocSize(IdxTy) - NewOffset;
2118 if (EltSize > Size) {
2124 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
2125 Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize));
2127 Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize));
2131 for (; Idx != NewElts.size() && Size; ++Idx) {
2132 IdxTy = NewElts[Idx]->getAllocatedType();
2133 uint64_t EltSize = DL->getTypeAllocSize(IdxTy);
2134 if (EltSize > Size) {
2140 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
2141 Builder.CreateLifetimeStart(NewElts[Idx],
2142 Builder.getInt64(EltSize));
2144 Builder.CreateLifetimeEnd(NewElts[Idx],
2145 Builder.getInt64(EltSize));
2147 DeadInsts.push_back(II);
2150 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
2151 /// Rewrite it to copy or set the elements of the scalarized memory.
2153 SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
2155 SmallVectorImpl<AllocaInst *> &NewElts) {
2156 // If this is a memcpy/memmove, construct the other pointer as the
2157 // appropriate type. The "Other" pointer is the pointer that goes to memory
2158 // that doesn't have anything to do with the alloca that we are promoting. For
2159 // memset, this Value* stays null.
2160 Value *OtherPtr = 0;
2161 unsigned MemAlignment = MI->getAlignment();
2162 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
2163 if (Inst == MTI->getRawDest())
2164 OtherPtr = MTI->getRawSource();
2166 assert(Inst == MTI->getRawSource());
2167 OtherPtr = MTI->getRawDest();
2171 // If there is an other pointer, we want to convert it to the same pointer
2172 // type as AI has, so we can GEP through it safely.
2174 unsigned AddrSpace =
2175 cast<PointerType>(OtherPtr->getType())->getAddressSpace();
2177 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an
2178 // optimization, but it's also required to detect the corner case where
2179 // both pointer operands are referencing the same memory, and where
2180 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This
2181 // function is only called for mem intrinsics that access the whole
2182 // aggregate, so non-zero GEPs are not an issue here.)
2183 OtherPtr = OtherPtr->stripPointerCasts();
2185 // Copying the alloca to itself is a no-op: just delete it.
2186 if (OtherPtr == AI || OtherPtr == NewElts[0]) {
2187 // This code will run twice for a no-op memcpy -- once for each operand.
2188 // Put only one reference to MI on the DeadInsts list.
2189 for (SmallVectorImpl<Value *>::const_iterator I = DeadInsts.begin(),
2190 E = DeadInsts.end(); I != E; ++I)
2191 if (*I == MI) return;
2192 DeadInsts.push_back(MI);
2196 // If the pointer is not the right type, insert a bitcast to the right
2199 PointerType::get(AI->getType()->getElementType(), AddrSpace);
2201 if (OtherPtr->getType() != NewTy)
2202 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI);
2205 // Process each element of the aggregate.
2206 bool SROADest = MI->getRawDest() == Inst;
2208 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
2210 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2211 // If this is a memcpy/memmove, emit a GEP of the other element address.
2212 Value *OtherElt = 0;
2213 unsigned OtherEltAlign = MemAlignment;
2216 Value *Idx[2] = { Zero,
2217 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
2218 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx,
2219 OtherPtr->getName()+"."+Twine(i),
2222 PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
2223 Type *OtherTy = OtherPtrTy->getElementType();
2224 if (StructType *ST = dyn_cast<StructType>(OtherTy)) {
2225 EltOffset = DL->getStructLayout(ST)->getElementOffset(i);
2227 Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
2228 EltOffset = DL->getTypeAllocSize(EltTy)*i;
2231 // The alignment of the other pointer is the guaranteed alignment of the
2232 // element, which is affected by both the known alignment of the whole
2233 // mem intrinsic and the alignment of the element. If the alignment of
2234 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
2235 // known alignment is just 4 bytes.
2236 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
2239 Value *EltPtr = NewElts[i];
2240 Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
2242 // If we got down to a scalar, insert a load or store as appropriate.
2243 if (EltTy->isSingleValueType()) {
2244 if (isa<MemTransferInst>(MI)) {
2246 // From Other to Alloca.
2247 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
2248 new StoreInst(Elt, EltPtr, MI);
2250 // From Alloca to Other.
2251 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
2252 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
2256 assert(isa<MemSetInst>(MI));
2258 // If the stored element is zero (common case), just store a null
2261 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) {
2263 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
2265 // If EltTy is a vector type, get the element type.
2266 Type *ValTy = EltTy->getScalarType();
2268 // Construct an integer with the right value.
2269 unsigned EltSize = DL->getTypeSizeInBits(ValTy);
2270 APInt OneVal(EltSize, CI->getZExtValue());
2271 APInt TotalVal(OneVal);
2273 for (unsigned i = 0; 8*i < EltSize; ++i) {
2274 TotalVal = TotalVal.shl(8);
2278 // Convert the integer value to the appropriate type.
2279 StoreVal = ConstantInt::get(CI->getContext(), TotalVal);
2280 if (ValTy->isPointerTy())
2281 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
2282 else if (ValTy->isFloatingPointTy())
2283 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
2284 assert(StoreVal->getType() == ValTy && "Type mismatch!");
2286 // If the requested value was a vector constant, create it.
2287 if (EltTy->isVectorTy()) {
2288 unsigned NumElts = cast<VectorType>(EltTy)->getNumElements();
2289 StoreVal = ConstantVector::getSplat(NumElts, StoreVal);
2292 new StoreInst(StoreVal, EltPtr, MI);
2295 // Otherwise, if we're storing a byte variable, use a memset call for
2299 unsigned EltSize = DL->getTypeAllocSize(EltTy);
2303 IRBuilder<> Builder(MI);
2305 // Finally, insert the meminst for this element.
2306 if (isa<MemSetInst>(MI)) {
2307 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize,
2310 assert(isa<MemTransferInst>(MI));
2311 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr
2312 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr
2314 if (isa<MemCpyInst>(MI))
2315 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile());
2317 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile());
2320 DeadInsts.push_back(MI);
2323 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that
2324 /// overwrites the entire allocation. Extract out the pieces of the stored
2325 /// integer and store them individually.
2327 SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
2328 SmallVectorImpl<AllocaInst *> &NewElts) {
2329 // Extract each element out of the integer according to its structure offset
2330 // and store the element value to the individual alloca.
2331 Value *SrcVal = SI->getOperand(0);
2332 Type *AllocaEltTy = AI->getAllocatedType();
2333 uint64_t AllocaSizeBits = DL->getTypeAllocSizeInBits(AllocaEltTy);
2335 IRBuilder<> Builder(SI);
2337 // Handle tail padding by extending the operand
2338 if (DL->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
2339 SrcVal = Builder.CreateZExt(SrcVal,
2340 IntegerType::get(SI->getContext(), AllocaSizeBits));
2342 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI
2345 // There are two forms here: AI could be an array or struct. Both cases
2346 // have different ways to compute the element offset.
2347 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2348 const StructLayout *Layout = DL->getStructLayout(EltSTy);
2350 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2351 // Get the number of bits to shift SrcVal to get the value.
2352 Type *FieldTy = EltSTy->getElementType(i);
2353 uint64_t Shift = Layout->getElementOffsetInBits(i);
2355 if (DL->isBigEndian())
2356 Shift = AllocaSizeBits-Shift-DL->getTypeAllocSizeInBits(FieldTy);
2358 Value *EltVal = SrcVal;
2360 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2361 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2364 // Truncate down to an integer of the right size.
2365 uint64_t FieldSizeBits = DL->getTypeSizeInBits(FieldTy);
2367 // Ignore zero sized fields like {}, they obviously contain no data.
2368 if (FieldSizeBits == 0) continue;
2370 if (FieldSizeBits != AllocaSizeBits)
2371 EltVal = Builder.CreateTrunc(EltVal,
2372 IntegerType::get(SI->getContext(), FieldSizeBits));
2373 Value *DestField = NewElts[i];
2374 if (EltVal->getType() == FieldTy) {
2375 // Storing to an integer field of this size, just do it.
2376 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) {
2377 // Bitcast to the right element type (for fp/vector values).
2378 EltVal = Builder.CreateBitCast(EltVal, FieldTy);
2380 // Otherwise, bitcast the dest pointer (for aggregates).
2381 DestField = Builder.CreateBitCast(DestField,
2382 PointerType::getUnqual(EltVal->getType()));
2384 new StoreInst(EltVal, DestField, SI);
2388 ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
2389 Type *ArrayEltTy = ATy->getElementType();
2390 uint64_t ElementOffset = DL->getTypeAllocSizeInBits(ArrayEltTy);
2391 uint64_t ElementSizeBits = DL->getTypeSizeInBits(ArrayEltTy);
2395 if (DL->isBigEndian())
2396 Shift = AllocaSizeBits-ElementOffset;
2400 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2401 // Ignore zero sized fields like {}, they obviously contain no data.
2402 if (ElementSizeBits == 0) continue;
2404 Value *EltVal = SrcVal;
2406 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2407 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2410 // Truncate down to an integer of the right size.
2411 if (ElementSizeBits != AllocaSizeBits)
2412 EltVal = Builder.CreateTrunc(EltVal,
2413 IntegerType::get(SI->getContext(),
2415 Value *DestField = NewElts[i];
2416 if (EltVal->getType() == ArrayEltTy) {
2417 // Storing to an integer field of this size, just do it.
2418 } else if (ArrayEltTy->isFloatingPointTy() ||
2419 ArrayEltTy->isVectorTy()) {
2420 // Bitcast to the right element type (for fp/vector values).
2421 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy);
2423 // Otherwise, bitcast the dest pointer (for aggregates).
2424 DestField = Builder.CreateBitCast(DestField,
2425 PointerType::getUnqual(EltVal->getType()));
2427 new StoreInst(EltVal, DestField, SI);
2429 if (DL->isBigEndian())
2430 Shift -= ElementOffset;
2432 Shift += ElementOffset;
2436 DeadInsts.push_back(SI);
2439 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to
2440 /// an integer. Load the individual pieces to form the aggregate value.
2442 SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
2443 SmallVectorImpl<AllocaInst *> &NewElts) {
2444 // Extract each element out of the NewElts according to its structure offset
2445 // and form the result value.
2446 Type *AllocaEltTy = AI->getAllocatedType();
2447 uint64_t AllocaSizeBits = DL->getTypeAllocSizeInBits(AllocaEltTy);
2449 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
2452 // There are two forms here: AI could be an array or struct. Both cases
2453 // have different ways to compute the element offset.
2454 const StructLayout *Layout = 0;
2455 uint64_t ArrayEltBitOffset = 0;
2456 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2457 Layout = DL->getStructLayout(EltSTy);
2459 Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
2460 ArrayEltBitOffset = DL->getTypeAllocSizeInBits(ArrayEltTy);
2464 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
2466 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2467 // Load the value from the alloca. If the NewElt is an aggregate, cast
2468 // the pointer to an integer of the same size before doing the load.
2469 Value *SrcField = NewElts[i];
2471 cast<PointerType>(SrcField->getType())->getElementType();
2472 uint64_t FieldSizeBits = DL->getTypeSizeInBits(FieldTy);
2474 // Ignore zero sized fields like {}, they obviously contain no data.
2475 if (FieldSizeBits == 0) continue;
2477 IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
2479 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() &&
2480 !FieldTy->isVectorTy())
2481 SrcField = new BitCastInst(SrcField,
2482 PointerType::getUnqual(FieldIntTy),
2484 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
2486 // If SrcField is a fp or vector of the right size but that isn't an
2487 // integer type, bitcast to an integer so we can shift it.
2488 if (SrcField->getType() != FieldIntTy)
2489 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
2491 // Zero extend the field to be the same size as the final alloca so that
2492 // we can shift and insert it.
2493 if (SrcField->getType() != ResultVal->getType())
2494 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
2496 // Determine the number of bits to shift SrcField.
2498 if (Layout) // Struct case.
2499 Shift = Layout->getElementOffsetInBits(i);
2501 Shift = i*ArrayEltBitOffset;
2503 if (DL->isBigEndian())
2504 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
2507 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
2508 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
2511 // Don't create an 'or x, 0' on the first iteration.
2512 if (!isa<Constant>(ResultVal) ||
2513 !cast<Constant>(ResultVal)->isNullValue())
2514 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
2516 ResultVal = SrcField;
2519 // Handle tail padding by truncating the result
2520 if (DL->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
2521 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
2523 LI->replaceAllUsesWith(ResultVal);
2524 DeadInsts.push_back(LI);
2527 /// HasPadding - Return true if the specified type has any structure or
2528 /// alignment padding in between the elements that would be split apart
2529 /// by SROA; return false otherwise.
2530 static bool HasPadding(Type *Ty, const DataLayout &DL) {
2531 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2532 Ty = ATy->getElementType();
2533 return DL.getTypeSizeInBits(Ty) != DL.getTypeAllocSizeInBits(Ty);
2536 // SROA currently handles only Arrays and Structs.
2537 StructType *STy = cast<StructType>(Ty);
2538 const StructLayout *SL = DL.getStructLayout(STy);
2539 unsigned PrevFieldBitOffset = 0;
2540 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2541 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
2543 // Check to see if there is any padding between this element and the
2546 unsigned PrevFieldEnd =
2547 PrevFieldBitOffset+DL.getTypeSizeInBits(STy->getElementType(i-1));
2548 if (PrevFieldEnd < FieldBitOffset)
2551 PrevFieldBitOffset = FieldBitOffset;
2553 // Check for tail padding.
2554 if (unsigned EltCount = STy->getNumElements()) {
2555 unsigned PrevFieldEnd = PrevFieldBitOffset +
2556 DL.getTypeSizeInBits(STy->getElementType(EltCount-1));
2557 if (PrevFieldEnd < SL->getSizeInBits())
2563 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
2564 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
2565 /// or 1 if safe after canonicalization has been performed.
2566 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) {
2567 // Loop over the use list of the alloca. We can only transform it if all of
2568 // the users are safe to transform.
2569 AllocaInfo Info(AI);
2571 isSafeForScalarRepl(AI, 0, Info);
2572 if (Info.isUnsafe) {
2573 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n');
2577 // Okay, we know all the users are promotable. If the aggregate is a memcpy
2578 // source and destination, we have to be careful. In particular, the memcpy
2579 // could be moving around elements that live in structure padding of the LLVM
2580 // types, but may actually be used. In these cases, we refuse to promote the
2582 if (Info.isMemCpySrc && Info.isMemCpyDst &&
2583 HasPadding(AI->getAllocatedType(), *DL))
2586 // If the alloca never has an access to just *part* of it, but is accessed
2587 // via loads and stores, then we should use ConvertToScalarInfo to promote
2588 // the alloca instead of promoting each piece at a time and inserting fission
2590 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) {
2591 // If the struct/array just has one element, use basic SRoA.
2592 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
2593 if (ST->getNumElements() > 1) return false;
2595 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1)