1 //===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This transformation implements the well known scalar replacement of
11 // aggregates transformation. This xform breaks up alloca instructions of
12 // aggregate type (structure or array) into individual alloca instructions for
13 // each member (if possible). Then, if possible, it transforms the individual
14 // alloca instructions into nice clean scalar SSA form.
16 // This combines a simple SRoA algorithm with the Mem2Reg algorithm because
17 // often interact, especially for C++ programs. As such, iterating between
18 // SRoA, then Mem2Reg until we run out of things to promote works well.
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "scalarrepl"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/DerivedTypes.h"
26 #include "llvm/Function.h"
27 #include "llvm/GlobalVariable.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/IntrinsicInst.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/Module.h"
32 #include "llvm/Pass.h"
33 #include "llvm/Analysis/DebugInfo.h"
34 #include "llvm/Analysis/DIBuilder.h"
35 #include "llvm/Analysis/Dominators.h"
36 #include "llvm/Analysis/Loads.h"
37 #include "llvm/Analysis/ValueTracking.h"
38 #include "llvm/Target/TargetData.h"
39 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
40 #include "llvm/Transforms/Utils/Local.h"
41 #include "llvm/Transforms/Utils/SSAUpdater.h"
42 #include "llvm/Support/CallSite.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/ErrorHandling.h"
45 #include "llvm/Support/GetElementPtrTypeIterator.h"
46 #include "llvm/Support/IRBuilder.h"
47 #include "llvm/Support/MathExtras.h"
48 #include "llvm/Support/raw_ostream.h"
49 #include "llvm/ADT/SetVector.h"
50 #include "llvm/ADT/SmallVector.h"
51 #include "llvm/ADT/Statistic.h"
54 STATISTIC(NumReplaced, "Number of allocas broken up");
55 STATISTIC(NumPromoted, "Number of allocas promoted");
56 STATISTIC(NumAdjusted, "Number of scalar allocas adjusted to allow promotion");
57 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
58 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
61 struct SROA : public FunctionPass {
62 SROA(int T, bool hasDT, char &ID)
63 : FunctionPass(ID), HasDomTree(hasDT) {
70 bool runOnFunction(Function &F);
72 bool performScalarRepl(Function &F);
73 bool performPromotion(Function &F);
79 /// DeadInsts - Keep track of instructions we have made dead, so that
80 /// we can remove them after we are done working.
81 SmallVector<Value*, 32> DeadInsts;
83 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
84 /// information about the uses. All these fields are initialized to false
85 /// and set to true when something is learned.
87 /// The alloca to promote.
90 /// CheckedPHIs - This is a set of verified PHI nodes, to prevent infinite
91 /// looping and avoid redundant work.
92 SmallPtrSet<PHINode*, 8> CheckedPHIs;
94 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
97 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
100 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
101 bool isMemCpyDst : 1;
103 /// hasSubelementAccess - This is true if a subelement of the alloca is
104 /// ever accessed, or false if the alloca is only accessed with mem
105 /// intrinsics or load/store that only access the entire alloca at once.
106 bool hasSubelementAccess : 1;
108 /// hasALoadOrStore - This is true if there are any loads or stores to it.
109 /// The alloca may just be accessed with memcpy, for example, which would
111 bool hasALoadOrStore : 1;
113 explicit AllocaInfo(AllocaInst *ai)
114 : AI(ai), isUnsafe(false), isMemCpySrc(false), isMemCpyDst(false),
115 hasSubelementAccess(false), hasALoadOrStore(false) {}
118 unsigned SRThreshold;
120 void MarkUnsafe(AllocaInfo &I, Instruction *User) {
122 DEBUG(dbgs() << " Transformation preventing inst: " << *User << '\n');
125 bool isSafeAllocaToScalarRepl(AllocaInst *AI);
127 void isSafeForScalarRepl(Instruction *I, uint64_t Offset, AllocaInfo &Info);
128 void isSafePHISelectUseForScalarRepl(Instruction *User, uint64_t Offset,
130 void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info);
131 void isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
132 Type *MemOpType, bool isStore, AllocaInfo &Info,
133 Instruction *TheAccess, bool AllowWholeAccess);
134 bool TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size);
135 uint64_t FindElementAndOffset(Type *&T, uint64_t &Offset,
138 void DoScalarReplacement(AllocaInst *AI,
139 std::vector<AllocaInst*> &WorkList);
140 void DeleteDeadInstructions();
142 void RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
143 SmallVector<AllocaInst*, 32> &NewElts);
144 void RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
145 SmallVector<AllocaInst*, 32> &NewElts);
146 void RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
147 SmallVector<AllocaInst*, 32> &NewElts);
148 void RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI,
150 SmallVector<AllocaInst*, 32> &NewElts);
151 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
153 SmallVector<AllocaInst*, 32> &NewElts);
154 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
155 SmallVector<AllocaInst*, 32> &NewElts);
156 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
157 SmallVector<AllocaInst*, 32> &NewElts);
159 static MemTransferInst *isOnlyCopiedFromConstantGlobal(
160 AllocaInst *AI, SmallVector<Instruction*, 4> &ToDelete);
163 // SROA_DT - SROA that uses DominatorTree.
164 struct SROA_DT : public SROA {
167 SROA_DT(int T = -1) : SROA(T, true, ID) {
168 initializeSROA_DTPass(*PassRegistry::getPassRegistry());
171 // getAnalysisUsage - This pass does not require any passes, but we know it
172 // will not alter the CFG, so say so.
173 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
174 AU.addRequired<DominatorTree>();
175 AU.setPreservesCFG();
179 // SROA_SSAUp - SROA that uses SSAUpdater.
180 struct SROA_SSAUp : public SROA {
183 SROA_SSAUp(int T = -1) : SROA(T, false, ID) {
184 initializeSROA_SSAUpPass(*PassRegistry::getPassRegistry());
187 // getAnalysisUsage - This pass does not require any passes, but we know it
188 // will not alter the CFG, so say so.
189 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
190 AU.setPreservesCFG();
196 char SROA_DT::ID = 0;
197 char SROA_SSAUp::ID = 0;
199 INITIALIZE_PASS_BEGIN(SROA_DT, "scalarrepl",
200 "Scalar Replacement of Aggregates (DT)", false, false)
201 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
202 INITIALIZE_PASS_END(SROA_DT, "scalarrepl",
203 "Scalar Replacement of Aggregates (DT)", false, false)
205 INITIALIZE_PASS_BEGIN(SROA_SSAUp, "scalarrepl-ssa",
206 "Scalar Replacement of Aggregates (SSAUp)", false, false)
207 INITIALIZE_PASS_END(SROA_SSAUp, "scalarrepl-ssa",
208 "Scalar Replacement of Aggregates (SSAUp)", false, false)
210 // Public interface to the ScalarReplAggregates pass
211 FunctionPass *llvm::createScalarReplAggregatesPass(int Threshold,
214 return new SROA_DT(Threshold);
215 return new SROA_SSAUp(Threshold);
219 //===----------------------------------------------------------------------===//
220 // Convert To Scalar Optimization.
221 //===----------------------------------------------------------------------===//
224 /// ConvertToScalarInfo - This class implements the "Convert To Scalar"
225 /// optimization, which scans the uses of an alloca and determines if it can
226 /// rewrite it in terms of a single new alloca that can be mem2reg'd.
227 class ConvertToScalarInfo {
228 /// AllocaSize - The size of the alloca being considered in bytes.
230 const TargetData &TD;
232 /// IsNotTrivial - This is set to true if there is some access to the object
233 /// which means that mem2reg can't promote it.
236 /// ScalarKind - Tracks the kind of alloca being considered for promotion,
237 /// computed based on the uses of the alloca rather than the LLVM type system.
241 // Accesses via GEPs that are consistent with element access of a vector
242 // type. This will not be converted into a vector unless there is a later
243 // access using an actual vector type.
246 // Accesses via vector operations and GEPs that are consistent with the
247 // layout of a vector type.
250 // An integer bag-of-bits with bitwise operations for insertion and
251 // extraction. Any combination of types can be converted into this kind
256 /// VectorTy - This tracks the type that we should promote the vector to if
257 /// it is possible to turn it into a vector. This starts out null, and if it
258 /// isn't possible to turn into a vector type, it gets set to VoidTy.
259 VectorType *VectorTy;
261 /// HadNonMemTransferAccess - True if there is at least one access to the
262 /// alloca that is not a MemTransferInst. We don't want to turn structs into
263 /// large integers unless there is some potential for optimization.
264 bool HadNonMemTransferAccess;
267 explicit ConvertToScalarInfo(unsigned Size, const TargetData &td)
268 : AllocaSize(Size), TD(td), IsNotTrivial(false), ScalarKind(Unknown),
269 VectorTy(0), HadNonMemTransferAccess(false) { }
271 AllocaInst *TryConvert(AllocaInst *AI);
274 bool CanConvertToScalar(Value *V, uint64_t Offset);
275 void MergeInTypeForLoadOrStore(Type *In, uint64_t Offset);
276 bool MergeInVectorType(VectorType *VInTy, uint64_t Offset);
277 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset);
279 Value *ConvertScalar_ExtractValue(Value *NV, Type *ToType,
280 uint64_t Offset, IRBuilder<> &Builder);
281 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal,
282 uint64_t Offset, IRBuilder<> &Builder);
284 } // end anonymous namespace.
287 /// TryConvert - Analyze the specified alloca, and if it is safe to do so,
288 /// rewrite it to be a new alloca which is mem2reg'able. This returns the new
289 /// alloca if possible or null if not.
290 AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
291 // If we can't convert this scalar, or if mem2reg can trivially do it, bail
293 if (!CanConvertToScalar(AI, 0) || !IsNotTrivial)
296 // If an alloca has only memset / memcpy uses, it may still have an Unknown
297 // ScalarKind. Treat it as an Integer below.
298 if (ScalarKind == Unknown)
299 ScalarKind = Integer;
301 // FIXME: It should be possible to promote the vector type up to the alloca's
303 if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8)
304 ScalarKind = Integer;
306 // If we were able to find a vector type that can handle this with
307 // insert/extract elements, and if there was at least one use that had
308 // a vector type, promote this to a vector. We don't want to promote
309 // random stuff that doesn't use vectors (e.g. <9 x double>) because then
310 // we just get a lot of insert/extracts. If at least one vector is
311 // involved, then we probably really do have a union of vector/array.
313 if (ScalarKind == Vector) {
314 assert(VectorTy && "Missing type for vector scalar.");
315 DEBUG(dbgs() << "CONVERT TO VECTOR: " << *AI << "\n TYPE = "
316 << *VectorTy << '\n');
317 NewTy = VectorTy; // Use the vector type.
319 unsigned BitWidth = AllocaSize * 8;
320 if ((ScalarKind == ImplicitVector || ScalarKind == Integer) &&
321 !HadNonMemTransferAccess && !TD.fitsInLegalInteger(BitWidth))
324 DEBUG(dbgs() << "CONVERT TO SCALAR INTEGER: " << *AI << "\n");
325 // Create and insert the integer alloca.
326 NewTy = IntegerType::get(AI->getContext(), BitWidth);
328 AllocaInst *NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin());
329 ConvertUsesToScalar(AI, NewAI, 0);
333 /// MergeInTypeForLoadOrStore - Add the 'In' type to the accumulated vector type
334 /// (VectorTy) so far at the offset specified by Offset (which is specified in
337 /// There are three cases we handle here:
338 /// 1) A union of vector types of the same size and potentially its elements.
339 /// Here we turn element accesses into insert/extract element operations.
340 /// This promotes a <4 x float> with a store of float to the third element
341 /// into a <4 x float> that uses insert element.
342 /// 2) A union of vector types with power-of-2 size differences, e.g. a float,
343 /// <2 x float> and <4 x float>. Here we turn element accesses into insert
344 /// and extract element operations, and <2 x float> accesses into a cast to
345 /// <2 x double>, an extract, and a cast back to <2 x float>.
346 /// 3) A fully general blob of memory, which we turn into some (potentially
347 /// large) integer type with extract and insert operations where the loads
348 /// and stores would mutate the memory. We mark this by setting VectorTy
350 void ConvertToScalarInfo::MergeInTypeForLoadOrStore(Type *In,
352 // If we already decided to turn this into a blob of integer memory, there is
353 // nothing to be done.
354 if (ScalarKind == Integer)
357 // If this could be contributing to a vector, analyze it.
359 // If the In type is a vector that is the same size as the alloca, see if it
360 // matches the existing VecTy.
361 if (VectorType *VInTy = dyn_cast<VectorType>(In)) {
362 if (MergeInVectorType(VInTy, Offset))
364 } else if (In->isFloatTy() || In->isDoubleTy() ||
365 (In->isIntegerTy() && In->getPrimitiveSizeInBits() >= 8 &&
366 isPowerOf2_32(In->getPrimitiveSizeInBits()))) {
367 // Full width accesses can be ignored, because they can always be turned
369 unsigned EltSize = In->getPrimitiveSizeInBits()/8;
370 if (EltSize == AllocaSize)
373 // If we're accessing something that could be an element of a vector, see
374 // if the implied vector agrees with what we already have and if Offset is
375 // compatible with it.
376 if (Offset % EltSize == 0 && AllocaSize % EltSize == 0 &&
377 (!VectorTy || Offset * 8 < VectorTy->getPrimitiveSizeInBits())) {
379 ScalarKind = ImplicitVector;
380 VectorTy = VectorType::get(In, AllocaSize/EltSize);
384 unsigned CurrentEltSize = VectorTy->getElementType()
385 ->getPrimitiveSizeInBits()/8;
386 if (EltSize == CurrentEltSize)
389 if (In->isIntegerTy() && isPowerOf2_32(AllocaSize / EltSize))
394 // Otherwise, we have a case that we can't handle with an optimized vector
395 // form. We can still turn this into a large integer.
396 ScalarKind = Integer;
399 /// MergeInVectorType - Handles the vector case of MergeInTypeForLoadOrStore,
400 /// returning true if the type was successfully merged and false otherwise.
401 bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy,
403 // TODO: Support nonzero offsets?
407 // Only allow vectors that are a power-of-2 away from the size of the alloca.
408 if (!isPowerOf2_64(AllocaSize / (VInTy->getBitWidth() / 8)))
411 // If this the first vector we see, remember the type so that we know the
419 unsigned BitWidth = VectorTy->getBitWidth();
420 unsigned InBitWidth = VInTy->getBitWidth();
422 // Vectors of the same size can be converted using a simple bitcast.
423 if (InBitWidth == BitWidth && AllocaSize == (InBitWidth / 8)) {
428 Type *ElementTy = VectorTy->getElementType();
429 Type *InElementTy = VInTy->getElementType();
431 // If they're the same alloc size, we'll be attempting to convert between
432 // them with a vector shuffle, which requires the element types to match.
433 if (TD.getTypeAllocSize(VectorTy) == TD.getTypeAllocSize(VInTy) &&
434 ElementTy != InElementTy)
437 // Do not allow mixed integer and floating-point accesses from vectors of
439 if (ElementTy->isFloatingPointTy() != InElementTy->isFloatingPointTy())
442 if (ElementTy->isFloatingPointTy()) {
443 // Only allow floating-point vectors of different sizes if they have the
444 // same element type.
445 // TODO: This could be loosened a bit, but would anything benefit?
446 if (ElementTy != InElementTy)
449 // There are no arbitrary-precision floating-point types, which limits the
450 // number of legal vector types with larger element types that we can form
451 // to bitcast and extract a subvector.
452 // TODO: We could support some more cases with mixed fp128 and double here.
453 if (!(BitWidth == 64 || BitWidth == 128) ||
454 !(InBitWidth == 64 || InBitWidth == 128))
457 assert(ElementTy->isIntegerTy() && "Vector elements must be either integer "
458 "or floating-point.");
459 unsigned BitWidth = ElementTy->getPrimitiveSizeInBits();
460 unsigned InBitWidth = InElementTy->getPrimitiveSizeInBits();
462 // Do not allow integer types smaller than a byte or types whose widths are
463 // not a multiple of a byte.
464 if (BitWidth < 8 || InBitWidth < 8 ||
465 BitWidth % 8 != 0 || InBitWidth % 8 != 0)
469 // Pick the largest of the two vector types.
471 if (InBitWidth > BitWidth)
477 /// CanConvertToScalar - V is a pointer. If we can convert the pointee and all
478 /// its accesses to a single vector type, return true and set VecTy to
479 /// the new type. If we could convert the alloca into a single promotable
480 /// integer, return true but set VecTy to VoidTy. Further, if the use is not a
481 /// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset
482 /// is the current offset from the base of the alloca being analyzed.
484 /// If we see at least one access to the value that is as a vector type, set the
486 bool ConvertToScalarInfo::CanConvertToScalar(Value *V, uint64_t Offset) {
487 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
488 Instruction *User = cast<Instruction>(*UI);
490 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
491 // Don't break volatile loads.
494 // Don't touch MMX operations.
495 if (LI->getType()->isX86_MMXTy())
497 HadNonMemTransferAccess = true;
498 MergeInTypeForLoadOrStore(LI->getType(), Offset);
502 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
503 // Storing the pointer, not into the value?
504 if (SI->getOperand(0) == V || !SI->isSimple()) return false;
505 // Don't touch MMX operations.
506 if (SI->getOperand(0)->getType()->isX86_MMXTy())
508 HadNonMemTransferAccess = true;
509 MergeInTypeForLoadOrStore(SI->getOperand(0)->getType(), Offset);
513 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
514 if (!onlyUsedByLifetimeMarkers(BCI))
515 IsNotTrivial = true; // Can't be mem2reg'd.
516 if (!CanConvertToScalar(BCI, Offset))
521 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
522 // If this is a GEP with a variable indices, we can't handle it.
523 if (!GEP->hasAllConstantIndices())
526 // Compute the offset that this GEP adds to the pointer.
527 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
528 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
530 // See if all uses can be converted.
531 if (!CanConvertToScalar(GEP, Offset+GEPOffset))
533 IsNotTrivial = true; // Can't be mem2reg'd.
534 HadNonMemTransferAccess = true;
538 // If this is a constant sized memset of a constant value (e.g. 0) we can
540 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
541 // Store of constant value.
542 if (!isa<ConstantInt>(MSI->getValue()))
545 // Store of constant size.
546 ConstantInt *Len = dyn_cast<ConstantInt>(MSI->getLength());
550 // If the size differs from the alloca, we can only convert the alloca to
551 // an integer bag-of-bits.
552 // FIXME: This should handle all of the cases that are currently accepted
553 // as vector element insertions.
554 if (Len->getZExtValue() != AllocaSize || Offset != 0)
555 ScalarKind = Integer;
557 IsNotTrivial = true; // Can't be mem2reg'd.
558 HadNonMemTransferAccess = true;
562 // If this is a memcpy or memmove into or out of the whole allocation, we
563 // can handle it like a load or store of the scalar type.
564 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
565 ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength());
566 if (Len == 0 || Len->getZExtValue() != AllocaSize || Offset != 0)
569 IsNotTrivial = true; // Can't be mem2reg'd.
573 // If this is a lifetime intrinsic, we can handle it.
574 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
575 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
576 II->getIntrinsicID() == Intrinsic::lifetime_end) {
581 // Otherwise, we cannot handle this!
588 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
589 /// directly. This happens when we are converting an "integer union" to a
590 /// single integer scalar, or when we are converting a "vector union" to a
591 /// vector with insert/extractelement instructions.
593 /// Offset is an offset from the original alloca, in bits that need to be
594 /// shifted to the right. By the end of this, there should be no uses of Ptr.
595 void ConvertToScalarInfo::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI,
597 while (!Ptr->use_empty()) {
598 Instruction *User = cast<Instruction>(Ptr->use_back());
600 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
601 ConvertUsesToScalar(CI, NewAI, Offset);
602 CI->eraseFromParent();
606 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
607 // Compute the offset that this GEP adds to the pointer.
608 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end());
609 uint64_t GEPOffset = TD.getIndexedOffset(GEP->getPointerOperandType(),
611 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8);
612 GEP->eraseFromParent();
616 IRBuilder<> Builder(User);
618 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
619 // The load is a bit extract from NewAI shifted right by Offset bits.
620 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp");
622 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder);
623 LI->replaceAllUsesWith(NewLoadVal);
624 LI->eraseFromParent();
628 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
629 assert(SI->getOperand(0) != Ptr && "Consistency error!");
630 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
631 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset,
633 Builder.CreateStore(New, NewAI);
634 SI->eraseFromParent();
636 // If the load we just inserted is now dead, then the inserted store
637 // overwrote the entire thing.
638 if (Old->use_empty())
639 Old->eraseFromParent();
643 // If this is a constant sized memset of a constant value (e.g. 0) we can
644 // transform it into a store of the expanded constant value.
645 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) {
646 assert(MSI->getRawDest() == Ptr && "Consistency error!");
647 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue();
649 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue();
651 // Compute the value replicated the right number of times.
652 APInt APVal(NumBytes*8, Val);
654 // Splat the value if non-zero.
656 for (unsigned i = 1; i != NumBytes; ++i)
659 Instruction *Old = Builder.CreateLoad(NewAI, NewAI->getName()+".in");
660 Value *New = ConvertScalar_InsertValue(
661 ConstantInt::get(User->getContext(), APVal),
662 Old, Offset, Builder);
663 Builder.CreateStore(New, NewAI);
665 // If the load we just inserted is now dead, then the memset overwrote
667 if (Old->use_empty())
668 Old->eraseFromParent();
670 MSI->eraseFromParent();
674 // If this is a memcpy or memmove into or out of the whole allocation, we
675 // can handle it like a load or store of the scalar type.
676 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) {
677 assert(Offset == 0 && "must be store to start of alloca");
679 // If the source and destination are both to the same alloca, then this is
680 // a noop copy-to-self, just delete it. Otherwise, emit a load and store
682 AllocaInst *OrigAI = cast<AllocaInst>(GetUnderlyingObject(Ptr, &TD, 0));
684 if (GetUnderlyingObject(MTI->getSource(), &TD, 0) != OrigAI) {
685 // Dest must be OrigAI, change this to be a load from the original
686 // pointer (bitcasted), then a store to our new alloca.
687 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?");
688 Value *SrcPtr = MTI->getSource();
689 PointerType* SPTy = cast<PointerType>(SrcPtr->getType());
690 PointerType* AIPTy = cast<PointerType>(NewAI->getType());
691 if (SPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
692 AIPTy = PointerType::get(AIPTy->getElementType(),
693 SPTy->getAddressSpace());
695 SrcPtr = Builder.CreateBitCast(SrcPtr, AIPTy);
697 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval");
698 SrcVal->setAlignment(MTI->getAlignment());
699 Builder.CreateStore(SrcVal, NewAI);
700 } else if (GetUnderlyingObject(MTI->getDest(), &TD, 0) != OrigAI) {
701 // Src must be OrigAI, change this to be a load from NewAI then a store
702 // through the original dest pointer (bitcasted).
703 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?");
704 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval");
706 PointerType* DPTy = cast<PointerType>(MTI->getDest()->getType());
707 PointerType* AIPTy = cast<PointerType>(NewAI->getType());
708 if (DPTy->getAddressSpace() != AIPTy->getAddressSpace()) {
709 AIPTy = PointerType::get(AIPTy->getElementType(),
710 DPTy->getAddressSpace());
712 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), AIPTy);
714 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr);
715 NewStore->setAlignment(MTI->getAlignment());
717 // Noop transfer. Src == Dst
720 MTI->eraseFromParent();
724 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
725 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
726 II->getIntrinsicID() == Intrinsic::lifetime_end) {
727 // There's no need to preserve these, as the resulting alloca will be
728 // converted to a register anyways.
729 II->eraseFromParent();
734 llvm_unreachable("Unsupported operation!");
738 /// getScaledElementType - Gets a scaled element type for a partial vector
739 /// access of an alloca. The input types must be integer or floating-point
740 /// scalar or vector types, and the resulting type is an integer, float or
742 static Type *getScaledElementType(Type *Ty1, Type *Ty2,
743 unsigned NewBitWidth) {
744 bool IsFP1 = Ty1->isFloatingPointTy() ||
745 (Ty1->isVectorTy() &&
746 cast<VectorType>(Ty1)->getElementType()->isFloatingPointTy());
747 bool IsFP2 = Ty2->isFloatingPointTy() ||
748 (Ty2->isVectorTy() &&
749 cast<VectorType>(Ty2)->getElementType()->isFloatingPointTy());
751 LLVMContext &Context = Ty1->getContext();
753 // Prefer floating-point types over integer types, as integer types may have
754 // been created by earlier scalar replacement.
755 if (IsFP1 || IsFP2) {
756 if (NewBitWidth == 32)
757 return Type::getFloatTy(Context);
758 if (NewBitWidth == 64)
759 return Type::getDoubleTy(Context);
762 return Type::getIntNTy(Context, NewBitWidth);
765 /// CreateShuffleVectorCast - Creates a shuffle vector to convert one vector
766 /// to another vector of the same element type which has the same allocation
767 /// size but different primitive sizes (e.g. <3 x i32> and <4 x i32>).
768 static Value *CreateShuffleVectorCast(Value *FromVal, Type *ToType,
769 IRBuilder<> &Builder) {
770 Type *FromType = FromVal->getType();
771 VectorType *FromVTy = cast<VectorType>(FromType);
772 VectorType *ToVTy = cast<VectorType>(ToType);
773 assert((ToVTy->getElementType() == FromVTy->getElementType()) &&
774 "Vectors must have the same element type");
775 Value *UnV = UndefValue::get(FromType);
776 unsigned numEltsFrom = FromVTy->getNumElements();
777 unsigned numEltsTo = ToVTy->getNumElements();
779 SmallVector<Constant*, 3> Args;
780 Type* Int32Ty = Builder.getInt32Ty();
781 unsigned minNumElts = std::min(numEltsFrom, numEltsTo);
783 for (i=0; i != minNumElts; ++i)
784 Args.push_back(ConstantInt::get(Int32Ty, i));
787 Constant* UnC = UndefValue::get(Int32Ty);
788 for (; i != numEltsTo; ++i)
791 Constant *Mask = ConstantVector::get(Args);
792 return Builder.CreateShuffleVector(FromVal, UnV, Mask, "tmpV");
795 /// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer
796 /// or vector value FromVal, extracting the bits from the offset specified by
797 /// Offset. This returns the value, which is of type ToType.
799 /// This happens when we are converting an "integer union" to a single
800 /// integer scalar, or when we are converting a "vector union" to a vector with
801 /// insert/extractelement instructions.
803 /// Offset is an offset from the original alloca, in bits that need to be
804 /// shifted to the right.
805 Value *ConvertToScalarInfo::
806 ConvertScalar_ExtractValue(Value *FromVal, Type *ToType,
807 uint64_t Offset, IRBuilder<> &Builder) {
808 // If the load is of the whole new alloca, no conversion is needed.
809 Type *FromType = FromVal->getType();
810 if (FromType == ToType && Offset == 0)
813 // If the result alloca is a vector type, this is either an element
814 // access or a bitcast to another vector type of the same size.
815 if (VectorType *VTy = dyn_cast<VectorType>(FromType)) {
816 unsigned FromTypeSize = TD.getTypeAllocSize(FromType);
817 unsigned ToTypeSize = TD.getTypeAllocSize(ToType);
818 if (FromTypeSize == ToTypeSize) {
819 // If the two types have the same primitive size, use a bit cast.
820 // Otherwise, it is two vectors with the same element type that has
821 // the same allocation size but different number of elements so use
823 if (FromType->getPrimitiveSizeInBits() ==
824 ToType->getPrimitiveSizeInBits())
825 return Builder.CreateBitCast(FromVal, ToType, "tmp");
827 return CreateShuffleVectorCast(FromVal, ToType, Builder);
830 if (isPowerOf2_64(FromTypeSize / ToTypeSize)) {
831 assert(!(ToType->isVectorTy() && Offset != 0) && "Can't extract a value "
832 "of a smaller vector type at a nonzero offset.");
834 Type *CastElementTy = getScaledElementType(FromType, ToType,
836 unsigned NumCastVectorElements = FromTypeSize / ToTypeSize;
838 LLVMContext &Context = FromVal->getContext();
839 Type *CastTy = VectorType::get(CastElementTy,
840 NumCastVectorElements);
841 Value *Cast = Builder.CreateBitCast(FromVal, CastTy, "tmp");
843 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy);
844 unsigned Elt = Offset/EltSize;
845 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
846 Value *Extract = Builder.CreateExtractElement(Cast, ConstantInt::get(
847 Type::getInt32Ty(Context), Elt), "tmp");
848 return Builder.CreateBitCast(Extract, ToType, "tmp");
851 // Otherwise it must be an element access.
854 unsigned EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
855 Elt = Offset/EltSize;
856 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
858 // Return the element extracted out of it.
859 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get(
860 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp");
861 if (V->getType() != ToType)
862 V = Builder.CreateBitCast(V, ToType, "tmp");
866 // If ToType is a first class aggregate, extract out each of the pieces and
867 // use insertvalue's to form the FCA.
868 if (StructType *ST = dyn_cast<StructType>(ToType)) {
869 const StructLayout &Layout = *TD.getStructLayout(ST);
870 Value *Res = UndefValue::get(ST);
871 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
872 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i),
873 Offset+Layout.getElementOffsetInBits(i),
875 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
880 if (ArrayType *AT = dyn_cast<ArrayType>(ToType)) {
881 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
882 Value *Res = UndefValue::get(AT);
883 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
884 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(),
885 Offset+i*EltSize, Builder);
886 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp");
891 // Otherwise, this must be a union that was converted to an integer value.
892 IntegerType *NTy = cast<IntegerType>(FromVal->getType());
894 // If this is a big-endian system and the load is narrower than the
895 // full alloca type, we need to do a shift to get the right bits.
897 if (TD.isBigEndian()) {
898 // On big-endian machines, the lowest bit is stored at the bit offset
899 // from the pointer given by getTypeStoreSizeInBits. This matters for
900 // integers with a bitwidth that is not a multiple of 8.
901 ShAmt = TD.getTypeStoreSizeInBits(NTy) -
902 TD.getTypeStoreSizeInBits(ToType) - Offset;
907 // Note: we support negative bitwidths (with shl) which are not defined.
908 // We do this to support (f.e.) loads off the end of a structure where
909 // only some bits are used.
910 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
911 FromVal = Builder.CreateLShr(FromVal,
912 ConstantInt::get(FromVal->getType(),
914 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
915 FromVal = Builder.CreateShl(FromVal,
916 ConstantInt::get(FromVal->getType(),
919 // Finally, unconditionally truncate the integer to the right width.
920 unsigned LIBitWidth = TD.getTypeSizeInBits(ToType);
921 if (LIBitWidth < NTy->getBitWidth())
923 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(),
925 else if (LIBitWidth > NTy->getBitWidth())
927 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(),
930 // If the result is an integer, this is a trunc or bitcast.
931 if (ToType->isIntegerTy()) {
933 } else if (ToType->isFloatingPointTy() || ToType->isVectorTy()) {
934 // Just do a bitcast, we know the sizes match up.
935 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp");
937 // Otherwise must be a pointer.
938 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp");
940 assert(FromVal->getType() == ToType && "Didn't convert right?");
944 /// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer
945 /// or vector value "Old" at the offset specified by Offset.
947 /// This happens when we are converting an "integer union" to a
948 /// single integer scalar, or when we are converting a "vector union" to a
949 /// vector with insert/extractelement instructions.
951 /// Offset is an offset from the original alloca, in bits that need to be
952 /// shifted to the right.
953 Value *ConvertToScalarInfo::
954 ConvertScalar_InsertValue(Value *SV, Value *Old,
955 uint64_t Offset, IRBuilder<> &Builder) {
956 // Convert the stored type to the actual type, shift it left to insert
957 // then 'or' into place.
958 Type *AllocaType = Old->getType();
959 LLVMContext &Context = Old->getContext();
961 if (VectorType *VTy = dyn_cast<VectorType>(AllocaType)) {
962 uint64_t VecSize = TD.getTypeAllocSizeInBits(VTy);
963 uint64_t ValSize = TD.getTypeAllocSizeInBits(SV->getType());
965 // Changing the whole vector with memset or with an access of a different
967 if (ValSize == VecSize) {
968 // If the two types have the same primitive size, use a bit cast.
969 // Otherwise, it is two vectors with the same element type that has
970 // the same allocation size but different number of elements so use
972 if (VTy->getPrimitiveSizeInBits() ==
973 SV->getType()->getPrimitiveSizeInBits())
974 return Builder.CreateBitCast(SV, AllocaType, "tmp");
976 return CreateShuffleVectorCast(SV, VTy, Builder);
979 if (isPowerOf2_64(VecSize / ValSize)) {
980 assert(!(SV->getType()->isVectorTy() && Offset != 0) && "Can't insert a "
981 "value of a smaller vector type at a nonzero offset.");
983 Type *CastElementTy = getScaledElementType(VTy, SV->getType(),
985 unsigned NumCastVectorElements = VecSize / ValSize;
987 LLVMContext &Context = SV->getContext();
988 Type *OldCastTy = VectorType::get(CastElementTy,
989 NumCastVectorElements);
990 Value *OldCast = Builder.CreateBitCast(Old, OldCastTy, "tmp");
992 Value *SVCast = Builder.CreateBitCast(SV, CastElementTy, "tmp");
994 unsigned EltSize = TD.getTypeAllocSizeInBits(CastElementTy);
995 unsigned Elt = Offset/EltSize;
996 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking");
998 Builder.CreateInsertElement(OldCast, SVCast, ConstantInt::get(
999 Type::getInt32Ty(Context), Elt), "tmp");
1000 return Builder.CreateBitCast(Insert, AllocaType, "tmp");
1003 // Must be an element insertion.
1004 assert(SV->getType() == VTy->getElementType());
1005 uint64_t EltSize = TD.getTypeAllocSizeInBits(VTy->getElementType());
1006 unsigned Elt = Offset/EltSize;
1007 return Builder.CreateInsertElement(Old, SV,
1008 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt),
1012 // If SV is a first-class aggregate value, insert each value recursively.
1013 if (StructType *ST = dyn_cast<StructType>(SV->getType())) {
1014 const StructLayout &Layout = *TD.getStructLayout(ST);
1015 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) {
1016 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1017 Old = ConvertScalar_InsertValue(Elt, Old,
1018 Offset+Layout.getElementOffsetInBits(i),
1024 if (ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) {
1025 uint64_t EltSize = TD.getTypeAllocSizeInBits(AT->getElementType());
1026 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1027 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp");
1028 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder);
1033 // If SV is a float, convert it to the appropriate integer type.
1034 // If it is a pointer, do the same.
1035 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType());
1036 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType);
1037 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType());
1038 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType);
1039 if (SV->getType()->isFloatingPointTy() || SV->getType()->isVectorTy())
1040 SV = Builder.CreateBitCast(SV,
1041 IntegerType::get(SV->getContext(),SrcWidth), "tmp");
1042 else if (SV->getType()->isPointerTy())
1043 SV = Builder.CreatePtrToInt(SV, TD.getIntPtrType(SV->getContext()), "tmp");
1045 // Zero extend or truncate the value if needed.
1046 if (SV->getType() != AllocaType) {
1047 if (SV->getType()->getPrimitiveSizeInBits() <
1048 AllocaType->getPrimitiveSizeInBits())
1049 SV = Builder.CreateZExt(SV, AllocaType, "tmp");
1051 // Truncation may be needed if storing more than the alloca can hold
1052 // (undefined behavior).
1053 SV = Builder.CreateTrunc(SV, AllocaType, "tmp");
1054 SrcWidth = DestWidth;
1055 SrcStoreWidth = DestStoreWidth;
1059 // If this is a big-endian system and the store is narrower than the
1060 // full alloca type, we need to do a shift to get the right bits.
1062 if (TD.isBigEndian()) {
1063 // On big-endian machines, the lowest bit is stored at the bit offset
1064 // from the pointer given by getTypeStoreSizeInBits. This matters for
1065 // integers with a bitwidth that is not a multiple of 8.
1066 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
1071 // Note: we support negative bitwidths (with shr) which are not defined.
1072 // We do this to support (f.e.) stores off the end of a structure where
1073 // only some bits in the structure are set.
1074 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
1075 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1076 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(),
1079 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1080 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(),
1082 Mask = Mask.lshr(-ShAmt);
1085 // Mask out the bits we are about to insert from the old value, and or
1087 if (SrcWidth != DestWidth) {
1088 assert(DestWidth > SrcWidth);
1089 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask");
1090 SV = Builder.CreateOr(Old, SV, "ins");
1096 //===----------------------------------------------------------------------===//
1098 //===----------------------------------------------------------------------===//
1101 bool SROA::runOnFunction(Function &F) {
1102 TD = getAnalysisIfAvailable<TargetData>();
1104 bool Changed = performPromotion(F);
1106 // FIXME: ScalarRepl currently depends on TargetData more than it
1107 // theoretically needs to. It should be refactored in order to support
1108 // target-independent IR. Until this is done, just skip the actual
1109 // scalar-replacement portion of this pass.
1110 if (!TD) return Changed;
1113 bool LocalChange = performScalarRepl(F);
1114 if (!LocalChange) break; // No need to repromote if no scalarrepl
1116 LocalChange = performPromotion(F);
1117 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
1124 class AllocaPromoter : public LoadAndStorePromoter {
1127 SmallVector<DbgDeclareInst *, 4> DDIs;
1128 SmallVector<DbgValueInst *, 4> DVIs;
1130 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1132 : LoadAndStorePromoter(Insts, S), AI(0), DIB(DB) {}
1134 void run(AllocaInst *AI, const SmallVectorImpl<Instruction*> &Insts) {
1135 // Remember which alloca we're promoting (for isInstInList).
1137 if (MDNode *DebugNode = MDNode::getIfExists(AI->getContext(), AI))
1138 for (Value::use_iterator UI = DebugNode->use_begin(),
1139 E = DebugNode->use_end(); UI != E; ++UI)
1140 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
1141 DDIs.push_back(DDI);
1142 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
1143 DVIs.push_back(DVI);
1145 LoadAndStorePromoter::run(Insts);
1146 AI->eraseFromParent();
1147 for (SmallVector<DbgDeclareInst *, 4>::iterator I = DDIs.begin(),
1148 E = DDIs.end(); I != E; ++I) {
1149 DbgDeclareInst *DDI = *I;
1150 DDI->eraseFromParent();
1152 for (SmallVector<DbgValueInst *, 4>::iterator I = DVIs.begin(),
1153 E = DVIs.end(); I != E; ++I) {
1154 DbgValueInst *DVI = *I;
1155 DVI->eraseFromParent();
1159 virtual bool isInstInList(Instruction *I,
1160 const SmallVectorImpl<Instruction*> &Insts) const {
1161 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1162 return LI->getOperand(0) == AI;
1163 return cast<StoreInst>(I)->getPointerOperand() == AI;
1166 virtual void updateDebugInfo(Instruction *Inst) const {
1167 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
1168 E = DDIs.end(); I != E; ++I) {
1169 DbgDeclareInst *DDI = *I;
1170 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1171 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB);
1172 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1173 ConvertDebugDeclareToDebugValue(DDI, LI, *DIB);
1175 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
1176 E = DVIs.end(); I != E; ++I) {
1177 DbgValueInst *DVI = *I;
1178 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1179 Instruction *DbgVal = NULL;
1180 // If an argument is zero extended then use argument directly. The ZExt
1181 // may be zapped by an optimization pass in future.
1182 Argument *ExtendedArg = NULL;
1183 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1184 ExtendedArg = dyn_cast<Argument>(ZExt->getOperand(0));
1185 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1186 ExtendedArg = dyn_cast<Argument>(SExt->getOperand(0));
1188 DbgVal = DIB->insertDbgValueIntrinsic(ExtendedArg, 0,
1189 DIVariable(DVI->getVariable()),
1192 DbgVal = DIB->insertDbgValueIntrinsic(SI->getOperand(0), 0,
1193 DIVariable(DVI->getVariable()),
1195 DbgVal->setDebugLoc(DVI->getDebugLoc());
1196 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1197 Instruction *DbgVal =
1198 DIB->insertDbgValueIntrinsic(LI->getOperand(0), 0,
1199 DIVariable(DVI->getVariable()), LI);
1200 DbgVal->setDebugLoc(DVI->getDebugLoc());
1205 } // end anon namespace
1207 /// isSafeSelectToSpeculate - Select instructions that use an alloca and are
1208 /// subsequently loaded can be rewritten to load both input pointers and then
1209 /// select between the result, allowing the load of the alloca to be promoted.
1211 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1212 /// %V = load i32* %P2
1214 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1215 /// %V2 = load i32* %Other
1216 /// %V = select i1 %cond, i32 %V1, i32 %V2
1218 /// We can do this to a select if its only uses are loads and if the operand to
1219 /// the select can be loaded unconditionally.
1220 static bool isSafeSelectToSpeculate(SelectInst *SI, const TargetData *TD) {
1221 bool TDerefable = SI->getTrueValue()->isDereferenceablePointer();
1222 bool FDerefable = SI->getFalseValue()->isDereferenceablePointer();
1224 for (Value::use_iterator UI = SI->use_begin(), UE = SI->use_end();
1226 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1227 if (LI == 0 || !LI->isSimple()) return false;
1229 // Both operands to the select need to be dereferencable, either absolutely
1230 // (e.g. allocas) or at this point because we can see other accesses to it.
1231 if (!TDerefable && !isSafeToLoadUnconditionally(SI->getTrueValue(), LI,
1232 LI->getAlignment(), TD))
1234 if (!FDerefable && !isSafeToLoadUnconditionally(SI->getFalseValue(), LI,
1235 LI->getAlignment(), TD))
1242 /// isSafePHIToSpeculate - PHI instructions that use an alloca and are
1243 /// subsequently loaded can be rewritten to load both input pointers in the pred
1244 /// blocks and then PHI the results, allowing the load of the alloca to be
1247 /// %P2 = phi [i32* %Alloca, i32* %Other]
1248 /// %V = load i32* %P2
1250 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1252 /// %V2 = load i32* %Other
1254 /// %V = phi [i32 %V1, i32 %V2]
1256 /// We can do this to a select if its only uses are loads and if the operand to
1257 /// the select can be loaded unconditionally.
1258 static bool isSafePHIToSpeculate(PHINode *PN, const TargetData *TD) {
1259 // For now, we can only do this promotion if the load is in the same block as
1260 // the PHI, and if there are no stores between the phi and load.
1261 // TODO: Allow recursive phi users.
1262 // TODO: Allow stores.
1263 BasicBlock *BB = PN->getParent();
1264 unsigned MaxAlign = 0;
1265 for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end();
1267 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1268 if (LI == 0 || !LI->isSimple()) return false;
1270 // For now we only allow loads in the same block as the PHI. This is a
1271 // common case that happens when instcombine merges two loads through a PHI.
1272 if (LI->getParent() != BB) return false;
1274 // Ensure that there are no instructions between the PHI and the load that
1276 for (BasicBlock::iterator BBI = PN; &*BBI != LI; ++BBI)
1277 if (BBI->mayWriteToMemory())
1280 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1283 // Okay, we know that we have one or more loads in the same block as the PHI.
1284 // We can transform this if it is safe to push the loads into the predecessor
1285 // blocks. The only thing to watch out for is that we can't put a possibly
1286 // trapping load in the predecessor if it is a critical edge.
1287 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1288 BasicBlock *Pred = PN->getIncomingBlock(i);
1289 Value *InVal = PN->getIncomingValue(i);
1291 // If the terminator of the predecessor has side-effects (an invoke),
1292 // there is no safe place to put a load in the predecessor.
1293 if (Pred->getTerminator()->mayHaveSideEffects())
1296 // If the value is produced by the terminator of the predecessor
1297 // (an invoke), there is no valid place to put a load in the predecessor.
1298 if (Pred->getTerminator() == InVal)
1301 // If the predecessor has a single successor, then the edge isn't critical.
1302 if (Pred->getTerminator()->getNumSuccessors() == 1)
1305 // If this pointer is always safe to load, or if we can prove that there is
1306 // already a load in the block, then we can move the load to the pred block.
1307 if (InVal->isDereferenceablePointer() ||
1308 isSafeToLoadUnconditionally(InVal, Pred->getTerminator(), MaxAlign, TD))
1318 /// tryToMakeAllocaBePromotable - This returns true if the alloca only has
1319 /// direct (non-volatile) loads and stores to it. If the alloca is close but
1320 /// not quite there, this will transform the code to allow promotion. As such,
1321 /// it is a non-pure predicate.
1322 static bool tryToMakeAllocaBePromotable(AllocaInst *AI, const TargetData *TD) {
1323 SetVector<Instruction*, SmallVector<Instruction*, 4>,
1324 SmallPtrSet<Instruction*, 4> > InstsToRewrite;
1326 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
1329 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1330 if (!LI->isSimple())
1335 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1336 if (SI->getOperand(0) == AI || !SI->isSimple())
1337 return false; // Don't allow a store OF the AI, only INTO the AI.
1341 if (SelectInst *SI = dyn_cast<SelectInst>(U)) {
1342 // If the condition being selected on is a constant, fold the select, yes
1343 // this does (rarely) happen early on.
1344 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition())) {
1345 Value *Result = SI->getOperand(1+CI->isZero());
1346 SI->replaceAllUsesWith(Result);
1347 SI->eraseFromParent();
1349 // This is very rare and we just scrambled the use list of AI, start
1351 return tryToMakeAllocaBePromotable(AI, TD);
1354 // If it is safe to turn "load (select c, AI, ptr)" into a select of two
1355 // loads, then we can transform this by rewriting the select.
1356 if (!isSafeSelectToSpeculate(SI, TD))
1359 InstsToRewrite.insert(SI);
1363 if (PHINode *PN = dyn_cast<PHINode>(U)) {
1364 if (PN->use_empty()) { // Dead PHIs can be stripped.
1365 InstsToRewrite.insert(PN);
1369 // If it is safe to turn "load (phi [AI, ptr, ...])" into a PHI of loads
1370 // in the pred blocks, then we can transform this by rewriting the PHI.
1371 if (!isSafePHIToSpeculate(PN, TD))
1374 InstsToRewrite.insert(PN);
1378 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
1379 if (onlyUsedByLifetimeMarkers(BCI)) {
1380 InstsToRewrite.insert(BCI);
1388 // If there are no instructions to rewrite, then all uses are load/stores and
1390 if (InstsToRewrite.empty())
1393 // If we have instructions that need to be rewritten for this to be promotable
1394 // take care of it now.
1395 for (unsigned i = 0, e = InstsToRewrite.size(); i != e; ++i) {
1396 if (BitCastInst *BCI = dyn_cast<BitCastInst>(InstsToRewrite[i])) {
1397 // This could only be a bitcast used by nothing but lifetime intrinsics.
1398 for (BitCastInst::use_iterator I = BCI->use_begin(), E = BCI->use_end();
1400 Use &U = I.getUse();
1402 cast<Instruction>(U.getUser())->eraseFromParent();
1404 BCI->eraseFromParent();
1408 if (SelectInst *SI = dyn_cast<SelectInst>(InstsToRewrite[i])) {
1409 // Selects in InstsToRewrite only have load uses. Rewrite each as two
1410 // loads with a new select.
1411 while (!SI->use_empty()) {
1412 LoadInst *LI = cast<LoadInst>(SI->use_back());
1414 IRBuilder<> Builder(LI);
1415 LoadInst *TrueLoad =
1416 Builder.CreateLoad(SI->getTrueValue(), LI->getName()+".t");
1417 LoadInst *FalseLoad =
1418 Builder.CreateLoad(SI->getFalseValue(), LI->getName()+".f");
1420 // Transfer alignment and TBAA info if present.
1421 TrueLoad->setAlignment(LI->getAlignment());
1422 FalseLoad->setAlignment(LI->getAlignment());
1423 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1424 TrueLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1425 FalseLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1428 Value *V = Builder.CreateSelect(SI->getCondition(), TrueLoad, FalseLoad);
1430 LI->replaceAllUsesWith(V);
1431 LI->eraseFromParent();
1434 // Now that all the loads are gone, the select is gone too.
1435 SI->eraseFromParent();
1439 // Otherwise, we have a PHI node which allows us to push the loads into the
1441 PHINode *PN = cast<PHINode>(InstsToRewrite[i]);
1442 if (PN->use_empty()) {
1443 PN->eraseFromParent();
1447 Type *LoadTy = cast<PointerType>(PN->getType())->getElementType();
1448 PHINode *NewPN = PHINode::Create(LoadTy, PN->getNumIncomingValues(),
1449 PN->getName()+".ld", PN);
1451 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1452 // matter which one we get and if any differ, it doesn't matter.
1453 LoadInst *SomeLoad = cast<LoadInst>(PN->use_back());
1454 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1455 unsigned Align = SomeLoad->getAlignment();
1457 // Rewrite all loads of the PN to use the new PHI.
1458 while (!PN->use_empty()) {
1459 LoadInst *LI = cast<LoadInst>(PN->use_back());
1460 LI->replaceAllUsesWith(NewPN);
1461 LI->eraseFromParent();
1464 // Inject loads into all of the pred blocks. Keep track of which blocks we
1465 // insert them into in case we have multiple edges from the same block.
1466 DenseMap<BasicBlock*, LoadInst*> InsertedLoads;
1468 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1469 BasicBlock *Pred = PN->getIncomingBlock(i);
1470 LoadInst *&Load = InsertedLoads[Pred];
1472 Load = new LoadInst(PN->getIncomingValue(i),
1473 PN->getName() + "." + Pred->getName(),
1474 Pred->getTerminator());
1475 Load->setAlignment(Align);
1476 if (TBAATag) Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1479 NewPN->addIncoming(Load, Pred);
1482 PN->eraseFromParent();
1489 bool SROA::performPromotion(Function &F) {
1490 std::vector<AllocaInst*> Allocas;
1491 DominatorTree *DT = 0;
1493 DT = &getAnalysis<DominatorTree>();
1495 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
1496 DIBuilder DIB(*F.getParent());
1497 bool Changed = false;
1498 SmallVector<Instruction*, 64> Insts;
1502 // Find allocas that are safe to promote, by looking at all instructions in
1504 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
1505 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
1506 if (tryToMakeAllocaBePromotable(AI, TD))
1507 Allocas.push_back(AI);
1509 if (Allocas.empty()) break;
1512 PromoteMemToReg(Allocas, *DT);
1515 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
1516 AllocaInst *AI = Allocas[i];
1518 // Build list of instructions to promote.
1519 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
1521 Insts.push_back(cast<Instruction>(*UI));
1522 AllocaPromoter(Insts, SSA, &DIB).run(AI, Insts);
1526 NumPromoted += Allocas.size();
1534 /// ShouldAttemptScalarRepl - Decide if an alloca is a good candidate for
1535 /// SROA. It must be a struct or array type with a small number of elements.
1536 static bool ShouldAttemptScalarRepl(AllocaInst *AI) {
1537 Type *T = AI->getAllocatedType();
1538 // Do not promote any struct into more than 32 separate vars.
1539 if (StructType *ST = dyn_cast<StructType>(T))
1540 return ST->getNumElements() <= 32;
1541 // Arrays are much less likely to be safe for SROA; only consider
1542 // them if they are very small.
1543 if (ArrayType *AT = dyn_cast<ArrayType>(T))
1544 return AT->getNumElements() <= 8;
1549 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
1550 // which runs on all of the alloca instructions in the function, removing them
1551 // if they are only used by getelementptr instructions.
1553 bool SROA::performScalarRepl(Function &F) {
1554 std::vector<AllocaInst*> WorkList;
1556 // Scan the entry basic block, adding allocas to the worklist.
1557 BasicBlock &BB = F.getEntryBlock();
1558 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
1559 if (AllocaInst *A = dyn_cast<AllocaInst>(I))
1560 WorkList.push_back(A);
1562 // Process the worklist
1563 bool Changed = false;
1564 while (!WorkList.empty()) {
1565 AllocaInst *AI = WorkList.back();
1566 WorkList.pop_back();
1568 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
1569 // with unused elements.
1570 if (AI->use_empty()) {
1571 AI->eraseFromParent();
1576 // If this alloca is impossible for us to promote, reject it early.
1577 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized())
1580 // Check to see if this allocation is only modified by a memcpy/memmove from
1581 // a constant global. If this is the case, we can change all users to use
1582 // the constant global instead. This is commonly produced by the CFE by
1583 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
1584 // is only subsequently read.
1585 SmallVector<Instruction *, 4> ToDelete;
1586 if (MemTransferInst *Copy = isOnlyCopiedFromConstantGlobal(AI, ToDelete)) {
1587 DEBUG(dbgs() << "Found alloca equal to global: " << *AI << '\n');
1588 DEBUG(dbgs() << " memcpy = " << *Copy << '\n');
1589 for (unsigned i = 0, e = ToDelete.size(); i != e; ++i)
1590 ToDelete[i]->eraseFromParent();
1591 Constant *TheSrc = cast<Constant>(Copy->getSource());
1592 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
1593 Copy->eraseFromParent(); // Don't mutate the global.
1594 AI->eraseFromParent();
1600 // Check to see if we can perform the core SROA transformation. We cannot
1601 // transform the allocation instruction if it is an array allocation
1602 // (allocations OF arrays are ok though), and an allocation of a scalar
1603 // value cannot be decomposed at all.
1604 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType());
1606 // Do not promote [0 x %struct].
1607 if (AllocaSize == 0) continue;
1609 // Do not promote any struct whose size is too big.
1610 if (AllocaSize > SRThreshold) continue;
1612 // If the alloca looks like a good candidate for scalar replacement, and if
1613 // all its users can be transformed, then split up the aggregate into its
1614 // separate elements.
1615 if (ShouldAttemptScalarRepl(AI) && isSafeAllocaToScalarRepl(AI)) {
1616 DoScalarReplacement(AI, WorkList);
1621 // If we can turn this aggregate value (potentially with casts) into a
1622 // simple scalar value that can be mem2reg'd into a register value.
1623 // IsNotTrivial tracks whether this is something that mem2reg could have
1624 // promoted itself. If so, we don't want to transform it needlessly. Note
1625 // that we can't just check based on the type: the alloca may be of an i32
1626 // but that has pointer arithmetic to set byte 3 of it or something.
1627 if (AllocaInst *NewAI =
1628 ConvertToScalarInfo((unsigned)AllocaSize, *TD).TryConvert(AI)) {
1629 NewAI->takeName(AI);
1630 AI->eraseFromParent();
1636 // Otherwise, couldn't process this alloca.
1642 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
1643 /// predicate, do SROA now.
1644 void SROA::DoScalarReplacement(AllocaInst *AI,
1645 std::vector<AllocaInst*> &WorkList) {
1646 DEBUG(dbgs() << "Found inst to SROA: " << *AI << '\n');
1647 SmallVector<AllocaInst*, 32> ElementAllocas;
1648 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
1649 ElementAllocas.reserve(ST->getNumContainedTypes());
1650 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
1651 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
1653 AI->getName() + "." + Twine(i), AI);
1654 ElementAllocas.push_back(NA);
1655 WorkList.push_back(NA); // Add to worklist for recursive processing
1658 ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
1659 ElementAllocas.reserve(AT->getNumElements());
1660 Type *ElTy = AT->getElementType();
1661 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
1662 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
1663 AI->getName() + "." + Twine(i), AI);
1664 ElementAllocas.push_back(NA);
1665 WorkList.push_back(NA); // Add to worklist for recursive processing
1669 // Now that we have created the new alloca instructions, rewrite all the
1670 // uses of the old alloca.
1671 RewriteForScalarRepl(AI, AI, 0, ElementAllocas);
1673 // Now erase any instructions that were made dead while rewriting the alloca.
1674 DeleteDeadInstructions();
1675 AI->eraseFromParent();
1680 /// DeleteDeadInstructions - Erase instructions on the DeadInstrs list,
1681 /// recursively including all their operands that become trivially dead.
1682 void SROA::DeleteDeadInstructions() {
1683 while (!DeadInsts.empty()) {
1684 Instruction *I = cast<Instruction>(DeadInsts.pop_back_val());
1686 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
1687 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
1688 // Zero out the operand and see if it becomes trivially dead.
1689 // (But, don't add allocas to the dead instruction list -- they are
1690 // already on the worklist and will be deleted separately.)
1692 if (isInstructionTriviallyDead(U) && !isa<AllocaInst>(U))
1693 DeadInsts.push_back(U);
1696 I->eraseFromParent();
1700 /// isSafeForScalarRepl - Check if instruction I is a safe use with regard to
1701 /// performing scalar replacement of alloca AI. The results are flagged in
1702 /// the Info parameter. Offset indicates the position within AI that is
1703 /// referenced by this instruction.
1704 void SROA::isSafeForScalarRepl(Instruction *I, uint64_t Offset,
1706 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1707 Instruction *User = cast<Instruction>(*UI);
1709 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1710 isSafeForScalarRepl(BC, Offset, Info);
1711 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1712 uint64_t GEPOffset = Offset;
1713 isSafeGEP(GEPI, GEPOffset, Info);
1715 isSafeForScalarRepl(GEPI, GEPOffset, Info);
1716 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1717 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1719 return MarkUnsafe(Info, User);
1720 isSafeMemAccess(Offset, Length->getZExtValue(), 0,
1721 UI.getOperandNo() == 0, Info, MI,
1722 true /*AllowWholeAccess*/);
1723 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1724 if (!LI->isSimple())
1725 return MarkUnsafe(Info, User);
1726 Type *LIType = LI->getType();
1727 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
1728 LIType, false, Info, LI, true /*AllowWholeAccess*/);
1729 Info.hasALoadOrStore = true;
1731 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1732 // Store is ok if storing INTO the pointer, not storing the pointer
1733 if (!SI->isSimple() || SI->getOperand(0) == I)
1734 return MarkUnsafe(Info, User);
1736 Type *SIType = SI->getOperand(0)->getType();
1737 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
1738 SIType, true, Info, SI, true /*AllowWholeAccess*/);
1739 Info.hasALoadOrStore = true;
1740 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
1741 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1742 II->getIntrinsicID() != Intrinsic::lifetime_end)
1743 return MarkUnsafe(Info, User);
1744 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1745 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1747 return MarkUnsafe(Info, User);
1749 if (Info.isUnsafe) return;
1754 /// isSafePHIUseForScalarRepl - If we see a PHI node or select using a pointer
1755 /// derived from the alloca, we can often still split the alloca into elements.
1756 /// This is useful if we have a large alloca where one element is phi'd
1757 /// together somewhere: we can SRoA and promote all the other elements even if
1758 /// we end up not being able to promote this one.
1760 /// All we require is that the uses of the PHI do not index into other parts of
1761 /// the alloca. The most important use case for this is single load and stores
1762 /// that are PHI'd together, which can happen due to code sinking.
1763 void SROA::isSafePHISelectUseForScalarRepl(Instruction *I, uint64_t Offset,
1765 // If we've already checked this PHI, don't do it again.
1766 if (PHINode *PN = dyn_cast<PHINode>(I))
1767 if (!Info.CheckedPHIs.insert(PN))
1770 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E; ++UI) {
1771 Instruction *User = cast<Instruction>(*UI);
1773 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1774 isSafePHISelectUseForScalarRepl(BC, Offset, Info);
1775 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1776 // Only allow "bitcast" GEPs for simplicity. We could generalize this,
1777 // but would have to prove that we're staying inside of an element being
1779 if (!GEPI->hasAllZeroIndices())
1780 return MarkUnsafe(Info, User);
1781 isSafePHISelectUseForScalarRepl(GEPI, Offset, Info);
1782 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1783 if (!LI->isSimple())
1784 return MarkUnsafe(Info, User);
1785 Type *LIType = LI->getType();
1786 isSafeMemAccess(Offset, TD->getTypeAllocSize(LIType),
1787 LIType, false, Info, LI, false /*AllowWholeAccess*/);
1788 Info.hasALoadOrStore = true;
1790 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1791 // Store is ok if storing INTO the pointer, not storing the pointer
1792 if (!SI->isSimple() || SI->getOperand(0) == I)
1793 return MarkUnsafe(Info, User);
1795 Type *SIType = SI->getOperand(0)->getType();
1796 isSafeMemAccess(Offset, TD->getTypeAllocSize(SIType),
1797 SIType, true, Info, SI, false /*AllowWholeAccess*/);
1798 Info.hasALoadOrStore = true;
1799 } else if (isa<PHINode>(User) || isa<SelectInst>(User)) {
1800 isSafePHISelectUseForScalarRepl(User, Offset, Info);
1802 return MarkUnsafe(Info, User);
1804 if (Info.isUnsafe) return;
1808 /// isSafeGEP - Check if a GEP instruction can be handled for scalar
1809 /// replacement. It is safe when all the indices are constant, in-bounds
1810 /// references, and when the resulting offset corresponds to an element within
1811 /// the alloca type. The results are flagged in the Info parameter. Upon
1812 /// return, Offset is adjusted as specified by the GEP indices.
1813 void SROA::isSafeGEP(GetElementPtrInst *GEPI,
1814 uint64_t &Offset, AllocaInfo &Info) {
1815 gep_type_iterator GEPIt = gep_type_begin(GEPI), E = gep_type_end(GEPI);
1819 // Walk through the GEP type indices, checking the types that this indexes
1821 for (; GEPIt != E; ++GEPIt) {
1822 // Ignore struct elements, no extra checking needed for these.
1823 if ((*GEPIt)->isStructTy())
1826 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
1828 return MarkUnsafe(Info, GEPI);
1831 // Compute the offset due to this GEP and check if the alloca has a
1832 // component element at that offset.
1833 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
1834 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
1835 if (!TypeHasComponent(Info.AI->getAllocatedType(), Offset, 0))
1836 MarkUnsafe(Info, GEPI);
1839 /// isHomogeneousAggregate - Check if type T is a struct or array containing
1840 /// elements of the same type (which is always true for arrays). If so,
1841 /// return true with NumElts and EltTy set to the number of elements and the
1842 /// element type, respectively.
1843 static bool isHomogeneousAggregate(Type *T, unsigned &NumElts,
1845 if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
1846 NumElts = AT->getNumElements();
1847 EltTy = (NumElts == 0 ? 0 : AT->getElementType());
1850 if (StructType *ST = dyn_cast<StructType>(T)) {
1851 NumElts = ST->getNumContainedTypes();
1852 EltTy = (NumElts == 0 ? 0 : ST->getContainedType(0));
1853 for (unsigned n = 1; n < NumElts; ++n) {
1854 if (ST->getContainedType(n) != EltTy)
1862 /// isCompatibleAggregate - Check if T1 and T2 are either the same type or are
1863 /// "homogeneous" aggregates with the same element type and number of elements.
1864 static bool isCompatibleAggregate(Type *T1, Type *T2) {
1868 unsigned NumElts1, NumElts2;
1869 Type *EltTy1, *EltTy2;
1870 if (isHomogeneousAggregate(T1, NumElts1, EltTy1) &&
1871 isHomogeneousAggregate(T2, NumElts2, EltTy2) &&
1872 NumElts1 == NumElts2 &&
1879 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI
1880 /// alloca or has an offset and size that corresponds to a component element
1881 /// within it. The offset checked here may have been formed from a GEP with a
1882 /// pointer bitcasted to a different type.
1884 /// If AllowWholeAccess is true, then this allows uses of the entire alloca as a
1885 /// unit. If false, it only allows accesses known to be in a single element.
1886 void SROA::isSafeMemAccess(uint64_t Offset, uint64_t MemSize,
1887 Type *MemOpType, bool isStore,
1888 AllocaInfo &Info, Instruction *TheAccess,
1889 bool AllowWholeAccess) {
1890 // Check if this is a load/store of the entire alloca.
1891 if (Offset == 0 && AllowWholeAccess &&
1892 MemSize == TD->getTypeAllocSize(Info.AI->getAllocatedType())) {
1893 // This can be safe for MemIntrinsics (where MemOpType is 0) and integer
1894 // loads/stores (which are essentially the same as the MemIntrinsics with
1895 // regard to copying padding between elements). But, if an alloca is
1896 // flagged as both a source and destination of such operations, we'll need
1897 // to check later for padding between elements.
1898 if (!MemOpType || MemOpType->isIntegerTy()) {
1900 Info.isMemCpyDst = true;
1902 Info.isMemCpySrc = true;
1905 // This is also safe for references using a type that is compatible with
1906 // the type of the alloca, so that loads/stores can be rewritten using
1907 // insertvalue/extractvalue.
1908 if (isCompatibleAggregate(MemOpType, Info.AI->getAllocatedType())) {
1909 Info.hasSubelementAccess = true;
1913 // Check if the offset/size correspond to a component within the alloca type.
1914 Type *T = Info.AI->getAllocatedType();
1915 if (TypeHasComponent(T, Offset, MemSize)) {
1916 Info.hasSubelementAccess = true;
1920 return MarkUnsafe(Info, TheAccess);
1923 /// TypeHasComponent - Return true if T has a component type with the
1924 /// specified offset and size. If Size is zero, do not check the size.
1925 bool SROA::TypeHasComponent(Type *T, uint64_t Offset, uint64_t Size) {
1928 if (StructType *ST = dyn_cast<StructType>(T)) {
1929 const StructLayout *Layout = TD->getStructLayout(ST);
1930 unsigned EltIdx = Layout->getElementContainingOffset(Offset);
1931 EltTy = ST->getContainedType(EltIdx);
1932 EltSize = TD->getTypeAllocSize(EltTy);
1933 Offset -= Layout->getElementOffset(EltIdx);
1934 } else if (ArrayType *AT = dyn_cast<ArrayType>(T)) {
1935 EltTy = AT->getElementType();
1936 EltSize = TD->getTypeAllocSize(EltTy);
1937 if (Offset >= AT->getNumElements() * EltSize)
1943 if (Offset == 0 && (Size == 0 || EltSize == Size))
1945 // Check if the component spans multiple elements.
1946 if (Offset + Size > EltSize)
1948 return TypeHasComponent(EltTy, Offset, Size);
1951 /// RewriteForScalarRepl - Alloca AI is being split into NewElts, so rewrite
1952 /// the instruction I, which references it, to use the separate elements.
1953 /// Offset indicates the position within AI that is referenced by this
1955 void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
1956 SmallVector<AllocaInst*, 32> &NewElts) {
1957 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI!=E;) {
1958 Use &TheUse = UI.getUse();
1959 Instruction *User = cast<Instruction>(*UI++);
1961 if (BitCastInst *BC = dyn_cast<BitCastInst>(User)) {
1962 RewriteBitCast(BC, AI, Offset, NewElts);
1966 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User)) {
1967 RewriteGEP(GEPI, AI, Offset, NewElts);
1971 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
1972 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
1973 uint64_t MemSize = Length->getZExtValue();
1975 MemSize == TD->getTypeAllocSize(AI->getAllocatedType()))
1976 RewriteMemIntrinUserOfAlloca(MI, I, AI, NewElts);
1977 // Otherwise the intrinsic can only touch a single element and the
1978 // address operand will be updated, so nothing else needs to be done.
1982 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(User)) {
1983 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
1984 II->getIntrinsicID() == Intrinsic::lifetime_end) {
1985 RewriteLifetimeIntrinsic(II, AI, Offset, NewElts);
1990 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1991 Type *LIType = LI->getType();
1993 if (isCompatibleAggregate(LIType, AI->getAllocatedType())) {
1995 // %res = load { i32, i32 }* %alloc
1997 // %load.0 = load i32* %alloc.0
1998 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
1999 // %load.1 = load i32* %alloc.1
2000 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
2001 // (Also works for arrays instead of structs)
2002 Value *Insert = UndefValue::get(LIType);
2003 IRBuilder<> Builder(LI);
2004 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2005 Value *Load = Builder.CreateLoad(NewElts[i], "load");
2006 Insert = Builder.CreateInsertValue(Insert, Load, i, "insert");
2008 LI->replaceAllUsesWith(Insert);
2009 DeadInsts.push_back(LI);
2010 } else if (LIType->isIntegerTy() &&
2011 TD->getTypeAllocSize(LIType) ==
2012 TD->getTypeAllocSize(AI->getAllocatedType())) {
2013 // If this is a load of the entire alloca to an integer, rewrite it.
2014 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts);
2019 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
2020 Value *Val = SI->getOperand(0);
2021 Type *SIType = Val->getType();
2022 if (isCompatibleAggregate(SIType, AI->getAllocatedType())) {
2024 // store { i32, i32 } %val, { i32, i32 }* %alloc
2026 // %val.0 = extractvalue { i32, i32 } %val, 0
2027 // store i32 %val.0, i32* %alloc.0
2028 // %val.1 = extractvalue { i32, i32 } %val, 1
2029 // store i32 %val.1, i32* %alloc.1
2030 // (Also works for arrays instead of structs)
2031 IRBuilder<> Builder(SI);
2032 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2033 Value *Extract = Builder.CreateExtractValue(Val, i, Val->getName());
2034 Builder.CreateStore(Extract, NewElts[i]);
2036 DeadInsts.push_back(SI);
2037 } else if (SIType->isIntegerTy() &&
2038 TD->getTypeAllocSize(SIType) ==
2039 TD->getTypeAllocSize(AI->getAllocatedType())) {
2040 // If this is a store of the entire alloca from an integer, rewrite it.
2041 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts);
2046 if (isa<SelectInst>(User) || isa<PHINode>(User)) {
2047 // If we have a PHI user of the alloca itself (as opposed to a GEP or
2048 // bitcast) we have to rewrite it. GEP and bitcast uses will be RAUW'd to
2050 if (!isa<AllocaInst>(I)) continue;
2052 assert(Offset == 0 && NewElts[0] &&
2053 "Direct alloca use should have a zero offset");
2055 // If we have a use of the alloca, we know the derived uses will be
2056 // utilizing just the first element of the scalarized result. Insert a
2057 // bitcast of the first alloca before the user as required.
2058 AllocaInst *NewAI = NewElts[0];
2059 BitCastInst *BCI = new BitCastInst(NewAI, AI->getType(), "", NewAI);
2060 NewAI->moveBefore(BCI);
2067 /// RewriteBitCast - Update a bitcast reference to the alloca being replaced
2068 /// and recursively continue updating all of its uses.
2069 void SROA::RewriteBitCast(BitCastInst *BC, AllocaInst *AI, uint64_t Offset,
2070 SmallVector<AllocaInst*, 32> &NewElts) {
2071 RewriteForScalarRepl(BC, AI, Offset, NewElts);
2072 if (BC->getOperand(0) != AI)
2075 // The bitcast references the original alloca. Replace its uses with
2076 // references to the first new element alloca.
2077 Instruction *Val = NewElts[0];
2078 if (Val->getType() != BC->getDestTy()) {
2079 Val = new BitCastInst(Val, BC->getDestTy(), "", BC);
2082 BC->replaceAllUsesWith(Val);
2083 DeadInsts.push_back(BC);
2086 /// FindElementAndOffset - Return the index of the element containing Offset
2087 /// within the specified type, which must be either a struct or an array.
2088 /// Sets T to the type of the element and Offset to the offset within that
2089 /// element. IdxTy is set to the type of the index result to be used in a
2090 /// GEP instruction.
2091 uint64_t SROA::FindElementAndOffset(Type *&T, uint64_t &Offset,
2094 if (StructType *ST = dyn_cast<StructType>(T)) {
2095 const StructLayout *Layout = TD->getStructLayout(ST);
2096 Idx = Layout->getElementContainingOffset(Offset);
2097 T = ST->getContainedType(Idx);
2098 Offset -= Layout->getElementOffset(Idx);
2099 IdxTy = Type::getInt32Ty(T->getContext());
2102 ArrayType *AT = cast<ArrayType>(T);
2103 T = AT->getElementType();
2104 uint64_t EltSize = TD->getTypeAllocSize(T);
2105 Idx = Offset / EltSize;
2106 Offset -= Idx * EltSize;
2107 IdxTy = Type::getInt64Ty(T->getContext());
2111 /// RewriteGEP - Check if this GEP instruction moves the pointer across
2112 /// elements of the alloca that are being split apart, and if so, rewrite
2113 /// the GEP to be relative to the new element.
2114 void SROA::RewriteGEP(GetElementPtrInst *GEPI, AllocaInst *AI, uint64_t Offset,
2115 SmallVector<AllocaInst*, 32> &NewElts) {
2116 uint64_t OldOffset = Offset;
2117 SmallVector<Value*, 8> Indices(GEPI->op_begin() + 1, GEPI->op_end());
2118 Offset += TD->getIndexedOffset(GEPI->getPointerOperandType(), Indices);
2120 RewriteForScalarRepl(GEPI, AI, Offset, NewElts);
2122 Type *T = AI->getAllocatedType();
2124 uint64_t OldIdx = FindElementAndOffset(T, OldOffset, IdxTy);
2125 if (GEPI->getOperand(0) == AI)
2126 OldIdx = ~0ULL; // Force the GEP to be rewritten.
2128 T = AI->getAllocatedType();
2129 uint64_t EltOffset = Offset;
2130 uint64_t Idx = FindElementAndOffset(T, EltOffset, IdxTy);
2132 // If this GEP does not move the pointer across elements of the alloca
2133 // being split, then it does not needs to be rewritten.
2137 Type *i32Ty = Type::getInt32Ty(AI->getContext());
2138 SmallVector<Value*, 8> NewArgs;
2139 NewArgs.push_back(Constant::getNullValue(i32Ty));
2140 while (EltOffset != 0) {
2141 uint64_t EltIdx = FindElementAndOffset(T, EltOffset, IdxTy);
2142 NewArgs.push_back(ConstantInt::get(IdxTy, EltIdx));
2144 Instruction *Val = NewElts[Idx];
2145 if (NewArgs.size() > 1) {
2146 Val = GetElementPtrInst::CreateInBounds(Val, NewArgs, "", GEPI);
2147 Val->takeName(GEPI);
2149 if (Val->getType() != GEPI->getType())
2150 Val = new BitCastInst(Val, GEPI->getType(), Val->getName(), GEPI);
2151 GEPI->replaceAllUsesWith(Val);
2152 DeadInsts.push_back(GEPI);
2155 /// RewriteLifetimeIntrinsic - II is a lifetime.start/lifetime.end. Rewrite it
2156 /// to mark the lifetime of the scalarized memory.
2157 void SROA::RewriteLifetimeIntrinsic(IntrinsicInst *II, AllocaInst *AI,
2159 SmallVector<AllocaInst*, 32> &NewElts) {
2160 ConstantInt *OldSize = cast<ConstantInt>(II->getArgOperand(0));
2161 // Put matching lifetime markers on everything from Offset up to
2163 Type *AIType = AI->getAllocatedType();
2164 uint64_t NewOffset = Offset;
2166 uint64_t Idx = FindElementAndOffset(AIType, NewOffset, IdxTy);
2168 IRBuilder<> Builder(II);
2169 uint64_t Size = OldSize->getLimitedValue();
2172 // Splice the first element and index 'NewOffset' bytes in. SROA will
2173 // split the alloca again later.
2174 Value *V = Builder.CreateBitCast(NewElts[Idx], Builder.getInt8PtrTy());
2175 V = Builder.CreateGEP(V, Builder.getInt64(NewOffset));
2177 IdxTy = NewElts[Idx]->getAllocatedType();
2178 uint64_t EltSize = TD->getTypeAllocSize(IdxTy) - NewOffset;
2179 if (EltSize > Size) {
2185 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
2186 Builder.CreateLifetimeStart(V, Builder.getInt64(EltSize));
2188 Builder.CreateLifetimeEnd(V, Builder.getInt64(EltSize));
2192 for (; Idx != NewElts.size() && Size; ++Idx) {
2193 IdxTy = NewElts[Idx]->getAllocatedType();
2194 uint64_t EltSize = TD->getTypeAllocSize(IdxTy);
2195 if (EltSize > Size) {
2201 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
2202 Builder.CreateLifetimeStart(NewElts[Idx],
2203 Builder.getInt64(EltSize));
2205 Builder.CreateLifetimeEnd(NewElts[Idx],
2206 Builder.getInt64(EltSize));
2208 DeadInsts.push_back(II);
2211 /// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI.
2212 /// Rewrite it to copy or set the elements of the scalarized memory.
2213 void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *Inst,
2215 SmallVector<AllocaInst*, 32> &NewElts) {
2216 // If this is a memcpy/memmove, construct the other pointer as the
2217 // appropriate type. The "Other" pointer is the pointer that goes to memory
2218 // that doesn't have anything to do with the alloca that we are promoting. For
2219 // memset, this Value* stays null.
2220 Value *OtherPtr = 0;
2221 unsigned MemAlignment = MI->getAlignment();
2222 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy
2223 if (Inst == MTI->getRawDest())
2224 OtherPtr = MTI->getRawSource();
2226 assert(Inst == MTI->getRawSource());
2227 OtherPtr = MTI->getRawDest();
2231 // If there is an other pointer, we want to convert it to the same pointer
2232 // type as AI has, so we can GEP through it safely.
2234 unsigned AddrSpace =
2235 cast<PointerType>(OtherPtr->getType())->getAddressSpace();
2237 // Remove bitcasts and all-zero GEPs from OtherPtr. This is an
2238 // optimization, but it's also required to detect the corner case where
2239 // both pointer operands are referencing the same memory, and where
2240 // OtherPtr may be a bitcast or GEP that currently being rewritten. (This
2241 // function is only called for mem intrinsics that access the whole
2242 // aggregate, so non-zero GEPs are not an issue here.)
2243 OtherPtr = OtherPtr->stripPointerCasts();
2245 // Copying the alloca to itself is a no-op: just delete it.
2246 if (OtherPtr == AI || OtherPtr == NewElts[0]) {
2247 // This code will run twice for a no-op memcpy -- once for each operand.
2248 // Put only one reference to MI on the DeadInsts list.
2249 for (SmallVector<Value*, 32>::const_iterator I = DeadInsts.begin(),
2250 E = DeadInsts.end(); I != E; ++I)
2251 if (*I == MI) return;
2252 DeadInsts.push_back(MI);
2256 // If the pointer is not the right type, insert a bitcast to the right
2259 PointerType::get(AI->getType()->getElementType(), AddrSpace);
2261 if (OtherPtr->getType() != NewTy)
2262 OtherPtr = new BitCastInst(OtherPtr, NewTy, OtherPtr->getName(), MI);
2265 // Process each element of the aggregate.
2266 bool SROADest = MI->getRawDest() == Inst;
2268 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext()));
2270 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2271 // If this is a memcpy/memmove, emit a GEP of the other element address.
2272 Value *OtherElt = 0;
2273 unsigned OtherEltAlign = MemAlignment;
2276 Value *Idx[2] = { Zero,
2277 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) };
2278 OtherElt = GetElementPtrInst::CreateInBounds(OtherPtr, Idx,
2279 OtherPtr->getName()+"."+Twine(i),
2282 PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType());
2283 Type *OtherTy = OtherPtrTy->getElementType();
2284 if (StructType *ST = dyn_cast<StructType>(OtherTy)) {
2285 EltOffset = TD->getStructLayout(ST)->getElementOffset(i);
2287 Type *EltTy = cast<SequentialType>(OtherTy)->getElementType();
2288 EltOffset = TD->getTypeAllocSize(EltTy)*i;
2291 // The alignment of the other pointer is the guaranteed alignment of the
2292 // element, which is affected by both the known alignment of the whole
2293 // mem intrinsic and the alignment of the element. If the alignment of
2294 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the
2295 // known alignment is just 4 bytes.
2296 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset);
2299 Value *EltPtr = NewElts[i];
2300 Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType();
2302 // If we got down to a scalar, insert a load or store as appropriate.
2303 if (EltTy->isSingleValueType()) {
2304 if (isa<MemTransferInst>(MI)) {
2306 // From Other to Alloca.
2307 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI);
2308 new StoreInst(Elt, EltPtr, MI);
2310 // From Alloca to Other.
2311 Value *Elt = new LoadInst(EltPtr, "tmp", MI);
2312 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI);
2316 assert(isa<MemSetInst>(MI));
2318 // If the stored element is zero (common case), just store a null
2321 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getArgOperand(1))) {
2323 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
2325 // If EltTy is a vector type, get the element type.
2326 Type *ValTy = EltTy->getScalarType();
2328 // Construct an integer with the right value.
2329 unsigned EltSize = TD->getTypeSizeInBits(ValTy);
2330 APInt OneVal(EltSize, CI->getZExtValue());
2331 APInt TotalVal(OneVal);
2333 for (unsigned i = 0; 8*i < EltSize; ++i) {
2334 TotalVal = TotalVal.shl(8);
2338 // Convert the integer value to the appropriate type.
2339 StoreVal = ConstantInt::get(CI->getContext(), TotalVal);
2340 if (ValTy->isPointerTy())
2341 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
2342 else if (ValTy->isFloatingPointTy())
2343 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
2344 assert(StoreVal->getType() == ValTy && "Type mismatch!");
2346 // If the requested value was a vector constant, create it.
2347 if (EltTy != ValTy) {
2348 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
2349 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
2350 StoreVal = ConstantVector::get(Elts);
2353 new StoreInst(StoreVal, EltPtr, MI);
2356 // Otherwise, if we're storing a byte variable, use a memset call for
2360 unsigned EltSize = TD->getTypeAllocSize(EltTy);
2362 IRBuilder<> Builder(MI);
2364 // Finally, insert the meminst for this element.
2365 if (isa<MemSetInst>(MI)) {
2366 Builder.CreateMemSet(EltPtr, MI->getArgOperand(1), EltSize,
2369 assert(isa<MemTransferInst>(MI));
2370 Value *Dst = SROADest ? EltPtr : OtherElt; // Dest ptr
2371 Value *Src = SROADest ? OtherElt : EltPtr; // Src ptr
2373 if (isa<MemCpyInst>(MI))
2374 Builder.CreateMemCpy(Dst, Src, EltSize, OtherEltAlign,MI->isVolatile());
2376 Builder.CreateMemMove(Dst, Src, EltSize,OtherEltAlign,MI->isVolatile());
2379 DeadInsts.push_back(MI);
2382 /// RewriteStoreUserOfWholeAlloca - We found a store of an integer that
2383 /// overwrites the entire allocation. Extract out the pieces of the stored
2384 /// integer and store them individually.
2385 void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocaInst *AI,
2386 SmallVector<AllocaInst*, 32> &NewElts){
2387 // Extract each element out of the integer according to its structure offset
2388 // and store the element value to the individual alloca.
2389 Value *SrcVal = SI->getOperand(0);
2390 Type *AllocaEltTy = AI->getAllocatedType();
2391 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
2393 IRBuilder<> Builder(SI);
2395 // Handle tail padding by extending the operand
2396 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits)
2397 SrcVal = Builder.CreateZExt(SrcVal,
2398 IntegerType::get(SI->getContext(), AllocaSizeBits));
2400 DEBUG(dbgs() << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << '\n' << *SI
2403 // There are two forms here: AI could be an array or struct. Both cases
2404 // have different ways to compute the element offset.
2405 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2406 const StructLayout *Layout = TD->getStructLayout(EltSTy);
2408 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2409 // Get the number of bits to shift SrcVal to get the value.
2410 Type *FieldTy = EltSTy->getElementType(i);
2411 uint64_t Shift = Layout->getElementOffsetInBits(i);
2413 if (TD->isBigEndian())
2414 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy);
2416 Value *EltVal = SrcVal;
2418 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2419 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2422 // Truncate down to an integer of the right size.
2423 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
2425 // Ignore zero sized fields like {}, they obviously contain no data.
2426 if (FieldSizeBits == 0) continue;
2428 if (FieldSizeBits != AllocaSizeBits)
2429 EltVal = Builder.CreateTrunc(EltVal,
2430 IntegerType::get(SI->getContext(), FieldSizeBits));
2431 Value *DestField = NewElts[i];
2432 if (EltVal->getType() == FieldTy) {
2433 // Storing to an integer field of this size, just do it.
2434 } else if (FieldTy->isFloatingPointTy() || FieldTy->isVectorTy()) {
2435 // Bitcast to the right element type (for fp/vector values).
2436 EltVal = Builder.CreateBitCast(EltVal, FieldTy);
2438 // Otherwise, bitcast the dest pointer (for aggregates).
2439 DestField = Builder.CreateBitCast(DestField,
2440 PointerType::getUnqual(EltVal->getType()));
2442 new StoreInst(EltVal, DestField, SI);
2446 ArrayType *ATy = cast<ArrayType>(AllocaEltTy);
2447 Type *ArrayEltTy = ATy->getElementType();
2448 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
2449 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy);
2453 if (TD->isBigEndian())
2454 Shift = AllocaSizeBits-ElementOffset;
2458 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2459 // Ignore zero sized fields like {}, they obviously contain no data.
2460 if (ElementSizeBits == 0) continue;
2462 Value *EltVal = SrcVal;
2464 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift);
2465 EltVal = Builder.CreateLShr(EltVal, ShiftVal, "sroa.store.elt");
2468 // Truncate down to an integer of the right size.
2469 if (ElementSizeBits != AllocaSizeBits)
2470 EltVal = Builder.CreateTrunc(EltVal,
2471 IntegerType::get(SI->getContext(),
2473 Value *DestField = NewElts[i];
2474 if (EltVal->getType() == ArrayEltTy) {
2475 // Storing to an integer field of this size, just do it.
2476 } else if (ArrayEltTy->isFloatingPointTy() ||
2477 ArrayEltTy->isVectorTy()) {
2478 // Bitcast to the right element type (for fp/vector values).
2479 EltVal = Builder.CreateBitCast(EltVal, ArrayEltTy);
2481 // Otherwise, bitcast the dest pointer (for aggregates).
2482 DestField = Builder.CreateBitCast(DestField,
2483 PointerType::getUnqual(EltVal->getType()));
2485 new StoreInst(EltVal, DestField, SI);
2487 if (TD->isBigEndian())
2488 Shift -= ElementOffset;
2490 Shift += ElementOffset;
2494 DeadInsts.push_back(SI);
2497 /// RewriteLoadUserOfWholeAlloca - We found a load of the entire allocation to
2498 /// an integer. Load the individual pieces to form the aggregate value.
2499 void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocaInst *AI,
2500 SmallVector<AllocaInst*, 32> &NewElts) {
2501 // Extract each element out of the NewElts according to its structure offset
2502 // and form the result value.
2503 Type *AllocaEltTy = AI->getAllocatedType();
2504 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy);
2506 DEBUG(dbgs() << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << '\n' << *LI
2509 // There are two forms here: AI could be an array or struct. Both cases
2510 // have different ways to compute the element offset.
2511 const StructLayout *Layout = 0;
2512 uint64_t ArrayEltBitOffset = 0;
2513 if (StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) {
2514 Layout = TD->getStructLayout(EltSTy);
2516 Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType();
2517 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy);
2521 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits));
2523 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
2524 // Load the value from the alloca. If the NewElt is an aggregate, cast
2525 // the pointer to an integer of the same size before doing the load.
2526 Value *SrcField = NewElts[i];
2528 cast<PointerType>(SrcField->getType())->getElementType();
2529 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy);
2531 // Ignore zero sized fields like {}, they obviously contain no data.
2532 if (FieldSizeBits == 0) continue;
2534 IntegerType *FieldIntTy = IntegerType::get(LI->getContext(),
2536 if (!FieldTy->isIntegerTy() && !FieldTy->isFloatingPointTy() &&
2537 !FieldTy->isVectorTy())
2538 SrcField = new BitCastInst(SrcField,
2539 PointerType::getUnqual(FieldIntTy),
2541 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI);
2543 // If SrcField is a fp or vector of the right size but that isn't an
2544 // integer type, bitcast to an integer so we can shift it.
2545 if (SrcField->getType() != FieldIntTy)
2546 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI);
2548 // Zero extend the field to be the same size as the final alloca so that
2549 // we can shift and insert it.
2550 if (SrcField->getType() != ResultVal->getType())
2551 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI);
2553 // Determine the number of bits to shift SrcField.
2555 if (Layout) // Struct case.
2556 Shift = Layout->getElementOffsetInBits(i);
2558 Shift = i*ArrayEltBitOffset;
2560 if (TD->isBigEndian())
2561 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth();
2564 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift);
2565 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI);
2568 // Don't create an 'or x, 0' on the first iteration.
2569 if (!isa<Constant>(ResultVal) ||
2570 !cast<Constant>(ResultVal)->isNullValue())
2571 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI);
2573 ResultVal = SrcField;
2576 // Handle tail padding by truncating the result
2577 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits)
2578 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI);
2580 LI->replaceAllUsesWith(ResultVal);
2581 DeadInsts.push_back(LI);
2584 /// HasPadding - Return true if the specified type has any structure or
2585 /// alignment padding in between the elements that would be split apart
2586 /// by SROA; return false otherwise.
2587 static bool HasPadding(Type *Ty, const TargetData &TD) {
2588 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2589 Ty = ATy->getElementType();
2590 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty);
2593 // SROA currently handles only Arrays and Structs.
2594 StructType *STy = cast<StructType>(Ty);
2595 const StructLayout *SL = TD.getStructLayout(STy);
2596 unsigned PrevFieldBitOffset = 0;
2597 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2598 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
2600 // Check to see if there is any padding between this element and the
2603 unsigned PrevFieldEnd =
2604 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
2605 if (PrevFieldEnd < FieldBitOffset)
2608 PrevFieldBitOffset = FieldBitOffset;
2610 // Check for tail padding.
2611 if (unsigned EltCount = STy->getNumElements()) {
2612 unsigned PrevFieldEnd = PrevFieldBitOffset +
2613 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
2614 if (PrevFieldEnd < SL->getSizeInBits())
2620 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
2621 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
2622 /// or 1 if safe after canonicalization has been performed.
2623 bool SROA::isSafeAllocaToScalarRepl(AllocaInst *AI) {
2624 // Loop over the use list of the alloca. We can only transform it if all of
2625 // the users are safe to transform.
2626 AllocaInfo Info(AI);
2628 isSafeForScalarRepl(AI, 0, Info);
2629 if (Info.isUnsafe) {
2630 DEBUG(dbgs() << "Cannot transform: " << *AI << '\n');
2634 // Okay, we know all the users are promotable. If the aggregate is a memcpy
2635 // source and destination, we have to be careful. In particular, the memcpy
2636 // could be moving around elements that live in structure padding of the LLVM
2637 // types, but may actually be used. In these cases, we refuse to promote the
2639 if (Info.isMemCpySrc && Info.isMemCpyDst &&
2640 HasPadding(AI->getAllocatedType(), *TD))
2643 // If the alloca never has an access to just *part* of it, but is accessed
2644 // via loads and stores, then we should use ConvertToScalarInfo to promote
2645 // the alloca instead of promoting each piece at a time and inserting fission
2647 if (!Info.hasSubelementAccess && Info.hasALoadOrStore) {
2648 // If the struct/array just has one element, use basic SRoA.
2649 if (StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
2650 if (ST->getNumElements() > 1) return false;
2652 if (cast<ArrayType>(AI->getAllocatedType())->getNumElements() > 1)
2662 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
2663 /// some part of a constant global variable. This intentionally only accepts
2664 /// constant expressions because we don't can't rewrite arbitrary instructions.
2665 static bool PointsToConstantGlobal(Value *V) {
2666 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
2667 return GV->isConstant();
2668 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
2669 if (CE->getOpcode() == Instruction::BitCast ||
2670 CE->getOpcode() == Instruction::GetElementPtr)
2671 return PointsToConstantGlobal(CE->getOperand(0));
2675 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
2676 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
2677 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
2678 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
2679 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
2680 /// the alloca, and if the source pointer is a pointer to a constant global, we
2681 /// can optimize this.
2683 isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
2685 SmallVector<Instruction *, 4> &LifetimeMarkers) {
2686 // We track lifetime intrinsics as we encounter them. If we decide to go
2687 // ahead and replace the value with the global, this lets the caller quickly
2688 // eliminate the markers.
2690 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
2691 User *U = cast<Instruction>(*UI);
2693 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
2694 // Ignore non-volatile loads, they are always ok.
2695 if (!LI->isSimple()) return false;
2699 if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
2700 // If uses of the bitcast are ok, we are ok.
2701 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset,
2706 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
2707 // If the GEP has all zero indices, it doesn't offset the pointer. If it
2708 // doesn't, it does.
2709 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
2710 isOffset || !GEP->hasAllZeroIndices(),
2716 if (CallSite CS = U) {
2717 // If this is the function being called then we treat it like a load and
2719 if (CS.isCallee(UI))
2722 // If this is a readonly/readnone call site, then we know it is just a
2723 // load (but one that potentially returns the value itself), so we can
2724 // ignore it if we know that the value isn't captured.
2725 unsigned ArgNo = CS.getArgumentNo(UI);
2726 if (CS.onlyReadsMemory() &&
2727 (CS.getInstruction()->use_empty() ||
2728 CS.paramHasAttr(ArgNo+1, Attribute::NoCapture)))
2731 // If this is being passed as a byval argument, the caller is making a
2732 // copy, so it is only a read of the alloca.
2733 if (CS.paramHasAttr(ArgNo+1, Attribute::ByVal))
2737 // Lifetime intrinsics can be handled by the caller.
2738 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
2739 if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
2740 II->getIntrinsicID() == Intrinsic::lifetime_end) {
2741 assert(II->use_empty() && "Lifetime markers have no result to use!");
2742 LifetimeMarkers.push_back(II);
2747 // If this is isn't our memcpy/memmove, reject it as something we can't
2749 MemTransferInst *MI = dyn_cast<MemTransferInst>(U);
2753 // If the transfer is using the alloca as a source of the transfer, then
2754 // ignore it since it is a load (unless the transfer is volatile).
2755 if (UI.getOperandNo() == 1) {
2756 if (MI->isVolatile()) return false;
2760 // If we already have seen a copy, reject the second one.
2761 if (TheCopy) return false;
2763 // If the pointer has been offset from the start of the alloca, we can't
2764 // safely handle this.
2765 if (isOffset) return false;
2767 // If the memintrinsic isn't using the alloca as the dest, reject it.
2768 if (UI.getOperandNo() != 0) return false;
2770 // If the source of the memcpy/move is not a constant global, reject it.
2771 if (!PointsToConstantGlobal(MI->getSource()))
2774 // Otherwise, the transform is safe. Remember the copy instruction.
2780 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
2781 /// modified by a copy from a constant global. If we can prove this, we can
2782 /// replace any uses of the alloca with uses of the global directly.
2784 SROA::isOnlyCopiedFromConstantGlobal(AllocaInst *AI,
2785 SmallVector<Instruction*, 4> &ToDelete) {
2786 MemTransferInst *TheCopy = 0;
2787 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false, ToDelete))