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/Pass.h"
31 #include "llvm/Analysis/Dominators.h"
32 #include "llvm/Target/TargetData.h"
33 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/GetElementPtrTypeIterator.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/Compiler.h"
38 #include "llvm/ADT/SmallVector.h"
39 #include "llvm/ADT/Statistic.h"
40 #include "llvm/ADT/StringExtras.h"
43 STATISTIC(NumReplaced, "Number of allocas broken up");
44 STATISTIC(NumPromoted, "Number of allocas promoted");
45 STATISTIC(NumConverted, "Number of aggregates converted to scalar");
46 STATISTIC(NumGlobals, "Number of allocas copied from constant global");
49 struct VISIBILITY_HIDDEN SROA : public FunctionPass {
50 static char ID; // Pass identification, replacement for typeid
51 explicit SROA(signed T = -1) : FunctionPass(&ID) {
58 bool runOnFunction(Function &F);
60 bool performScalarRepl(Function &F);
61 bool performPromotion(Function &F);
63 // getAnalysisUsage - This pass does not require any passes, but we know it
64 // will not alter the CFG, so say so.
65 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
66 AU.addRequired<DominatorTree>();
67 AU.addRequired<DominanceFrontier>();
68 AU.addRequired<TargetData>();
73 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures
74 /// information about the uses. All these fields are initialized to false
75 /// and set to true when something is learned.
77 /// isUnsafe - This is set to true if the alloca cannot be SROA'd.
80 /// needsCanon - This is set to true if there is some use of the alloca
81 /// that requires canonicalization.
84 /// isMemCpySrc - This is true if this aggregate is memcpy'd from.
87 /// isMemCpyDst - This is true if this aggregate is memcpy'd into.
91 : isUnsafe(false), needsCanon(false),
92 isMemCpySrc(false), isMemCpyDst(false) {}
97 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; }
99 int isSafeAllocaToScalarRepl(AllocationInst *AI);
101 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
103 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
105 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
106 unsigned OpNo, AllocaInfo &Info);
107 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI,
110 void DoScalarReplacement(AllocationInst *AI,
111 std::vector<AllocationInst*> &WorkList);
112 void CanonicalizeAllocaUsers(AllocationInst *AI);
113 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base);
115 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
116 SmallVector<AllocaInst*, 32> &NewElts);
118 const Type *CanConvertToScalar(Value *V, bool &IsNotTrivial);
119 void ConvertToScalar(AllocationInst *AI, const Type *Ty);
120 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset);
121 Value *ConvertUsesOfLoadToScalar(LoadInst *LI, AllocaInst *NewAI,
123 Value *ConvertUsesOfStoreToScalar(StoreInst *SI, AllocaInst *NewAI,
125 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI);
130 static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates");
132 // Public interface to the ScalarReplAggregates pass
133 FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) {
134 return new SROA(Threshold);
138 bool SROA::runOnFunction(Function &F) {
139 bool Changed = performPromotion(F);
141 bool LocalChange = performScalarRepl(F);
142 if (!LocalChange) break; // No need to repromote if no scalarrepl
144 LocalChange = performPromotion(F);
145 if (!LocalChange) break; // No need to re-scalarrepl if no promotion
152 bool SROA::performPromotion(Function &F) {
153 std::vector<AllocaInst*> Allocas;
154 DominatorTree &DT = getAnalysis<DominatorTree>();
155 DominanceFrontier &DF = getAnalysis<DominanceFrontier>();
157 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function
159 bool Changed = false;
164 // Find allocas that are safe to promote, by looking at all instructions in
166 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I)
167 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca?
168 if (isAllocaPromotable(AI))
169 Allocas.push_back(AI);
171 if (Allocas.empty()) break;
173 PromoteMemToReg(Allocas, DT, DF);
174 NumPromoted += Allocas.size();
181 /// getNumSAElements - Return the number of elements in the specific struct or
183 static uint64_t getNumSAElements(const Type *T) {
184 if (const StructType *ST = dyn_cast<StructType>(T))
185 return ST->getNumElements();
186 return cast<ArrayType>(T)->getNumElements();
189 // performScalarRepl - This algorithm is a simple worklist driven algorithm,
190 // which runs on all of the malloc/alloca instructions in the function, removing
191 // them if they are only used by getelementptr instructions.
193 bool SROA::performScalarRepl(Function &F) {
194 std::vector<AllocationInst*> WorkList;
196 // Scan the entry basic block, adding any alloca's and mallocs to the worklist
197 BasicBlock &BB = F.getEntryBlock();
198 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I)
199 if (AllocationInst *A = dyn_cast<AllocationInst>(I))
200 WorkList.push_back(A);
202 const TargetData &TD = getAnalysis<TargetData>();
204 // Process the worklist
205 bool Changed = false;
206 while (!WorkList.empty()) {
207 AllocationInst *AI = WorkList.back();
210 // Handle dead allocas trivially. These can be formed by SROA'ing arrays
211 // with unused elements.
212 if (AI->use_empty()) {
213 AI->eraseFromParent();
217 // If we can turn this aggregate value (potentially with casts) into a
218 // simple scalar value that can be mem2reg'd into a register value.
219 bool IsNotTrivial = false;
220 if (const Type *ActualType = CanConvertToScalar(AI, IsNotTrivial))
221 if (IsNotTrivial && ActualType != Type::VoidTy) {
222 ConvertToScalar(AI, ActualType);
227 // Check to see if we can perform the core SROA transformation. We cannot
228 // transform the allocation instruction if it is an array allocation
229 // (allocations OF arrays are ok though), and an allocation of a scalar
230 // value cannot be decomposed at all.
231 if (!AI->isArrayAllocation() &&
232 (isa<StructType>(AI->getAllocatedType()) ||
233 isa<ArrayType>(AI->getAllocatedType())) &&
234 AI->getAllocatedType()->isSized() &&
235 // Do not promote any struct whose size is larger than "128" bytes.
236 TD.getABITypeSize(AI->getAllocatedType()) < SRThreshold &&
237 // Do not promote any struct into more than "32" separate vars.
238 getNumSAElements(AI->getAllocatedType()) < SRThreshold/4) {
239 // Check that all of the users of the allocation are capable of being
241 switch (isSafeAllocaToScalarRepl(AI)) {
242 default: assert(0 && "Unexpected value!");
243 case 0: // Not safe to scalar replace.
245 case 1: // Safe, but requires cleanup/canonicalizations first
246 CanonicalizeAllocaUsers(AI);
248 case 3: // Safe to scalar replace.
249 DoScalarReplacement(AI, WorkList);
255 // Check to see if this allocation is only modified by a memcpy/memmove from
256 // a constant global. If this is the case, we can change all users to use
257 // the constant global instead. This is commonly produced by the CFE by
258 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A'
259 // is only subsequently read.
260 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) {
261 DOUT << "Found alloca equal to global: " << *AI;
262 DOUT << " memcpy = " << *TheCopy;
263 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2));
264 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType()));
265 TheCopy->eraseFromParent(); // Don't mutate the global.
266 AI->eraseFromParent();
272 // Otherwise, couldn't process this.
278 /// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl
279 /// predicate, do SROA now.
280 void SROA::DoScalarReplacement(AllocationInst *AI,
281 std::vector<AllocationInst*> &WorkList) {
282 DOUT << "Found inst to SROA: " << *AI;
283 SmallVector<AllocaInst*, 32> ElementAllocas;
284 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) {
285 ElementAllocas.reserve(ST->getNumContainedTypes());
286 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) {
287 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0,
289 AI->getName() + "." + utostr(i), AI);
290 ElementAllocas.push_back(NA);
291 WorkList.push_back(NA); // Add to worklist for recursive processing
294 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType());
295 ElementAllocas.reserve(AT->getNumElements());
296 const Type *ElTy = AT->getElementType();
297 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) {
298 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(),
299 AI->getName() + "." + utostr(i), AI);
300 ElementAllocas.push_back(NA);
301 WorkList.push_back(NA); // Add to worklist for recursive processing
305 // Now that we have created the alloca instructions that we want to use,
306 // expand the getelementptr instructions to use them.
308 while (!AI->use_empty()) {
309 Instruction *User = cast<Instruction>(AI->use_back());
310 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) {
311 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas);
312 BCInst->eraseFromParent();
317 // %res = load { i32, i32 }* %alloc
319 // %load.0 = load i32* %alloc.0
320 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0
321 // %load.1 = load i32* %alloc.1
322 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1
323 // (Also works for arrays instead of structs)
324 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
325 Value *Insert = UndefValue::get(LI->getType());
326 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
327 Value *Load = new LoadInst(ElementAllocas[i], "load", LI);
328 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI);
330 LI->replaceAllUsesWith(Insert);
331 LI->eraseFromParent();
336 // store { i32, i32 } %val, { i32, i32 }* %alloc
338 // %val.0 = extractvalue { i32, i32 } %val, 0
339 // store i32 %val.0, i32* %alloc.0
340 // %val.1 = extractvalue { i32, i32 } %val, 1
341 // store i32 %val.1, i32* %alloc.1
342 // (Also works for arrays instead of structs)
343 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
344 Value *Val = SI->getOperand(0);
345 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) {
346 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI);
347 new StoreInst(Extract, ElementAllocas[i], SI);
349 SI->eraseFromParent();
353 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User);
354 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst>
356 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
358 assert(Idx < ElementAllocas.size() && "Index out of range?");
359 AllocaInst *AllocaToUse = ElementAllocas[Idx];
362 if (GEPI->getNumOperands() == 3) {
363 // Do not insert a new getelementptr instruction with zero indices, only
364 // to have it optimized out later.
365 RepValue = AllocaToUse;
367 // We are indexing deeply into the structure, so we still need a
368 // getelement ptr instruction to finish the indexing. This may be
369 // expanded itself once the worklist is rerun.
371 SmallVector<Value*, 8> NewArgs;
372 NewArgs.push_back(Constant::getNullValue(Type::Int32Ty));
373 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end());
374 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(),
375 NewArgs.end(), "", GEPI);
376 RepValue->takeName(GEPI);
379 // If this GEP is to the start of the aggregate, check for memcpys.
381 bool IsStartOfAggregateGEP = true;
382 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i) {
383 if (!isa<ConstantInt>(GEPI->getOperand(i))) {
384 IsStartOfAggregateGEP = false;
387 if (!cast<ConstantInt>(GEPI->getOperand(i))->isZero()) {
388 IsStartOfAggregateGEP = false;
393 if (IsStartOfAggregateGEP)
394 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas);
398 // Move all of the users over to the new GEP.
399 GEPI->replaceAllUsesWith(RepValue);
400 // Delete the old GEP
401 GEPI->eraseFromParent();
404 // Finally, delete the Alloca instruction
405 AI->eraseFromParent();
410 /// isSafeElementUse - Check to see if this use is an allowed use for a
411 /// getelementptr instruction of an array aggregate allocation. isFirstElt
412 /// indicates whether Ptr is known to the start of the aggregate.
414 void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI,
416 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
418 Instruction *User = cast<Instruction>(*I);
419 switch (User->getOpcode()) {
420 case Instruction::Load: break;
421 case Instruction::Store:
422 // Store is ok if storing INTO the pointer, not storing the pointer
423 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info);
425 case Instruction::GetElementPtr: {
426 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User);
427 bool AreAllZeroIndices = isFirstElt;
428 if (GEP->getNumOperands() > 1) {
429 if (!isa<ConstantInt>(GEP->getOperand(1)) ||
430 !cast<ConstantInt>(GEP->getOperand(1))->isZero())
431 // Using pointer arithmetic to navigate the array.
432 return MarkUnsafe(Info);
434 if (AreAllZeroIndices) {
435 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i) {
436 if (!isa<ConstantInt>(GEP->getOperand(i)) ||
437 !cast<ConstantInt>(GEP->getOperand(i))->isZero()) {
438 AreAllZeroIndices = false;
444 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info);
445 if (Info.isUnsafe) return;
448 case Instruction::BitCast:
450 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info);
451 if (Info.isUnsafe) return;
454 DOUT << " Transformation preventing inst: " << *User;
455 return MarkUnsafe(Info);
456 case Instruction::Call:
457 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) {
459 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info);
460 if (Info.isUnsafe) return;
464 DOUT << " Transformation preventing inst: " << *User;
465 return MarkUnsafe(Info);
467 DOUT << " Transformation preventing inst: " << *User;
468 return MarkUnsafe(Info);
471 return; // All users look ok :)
474 /// AllUsersAreLoads - Return true if all users of this value are loads.
475 static bool AllUsersAreLoads(Value *Ptr) {
476 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end();
478 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load)
483 /// isSafeUseOfAllocation - Check to see if this user is an allowed use for an
484 /// aggregate allocation.
486 void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI,
488 if (BitCastInst *C = dyn_cast<BitCastInst>(User))
489 return isSafeUseOfBitCastedAllocation(C, AI, Info);
491 if (isa<LoadInst>(User))
492 return; // Loads (returning a first class aggregrate) are always rewritable
494 if (isa<StoreInst>(User) && User->getOperand(0) != AI)
495 return; // Store is ok if storing INTO the pointer, not storing the pointer
497 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User);
499 return MarkUnsafe(Info);
501 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI);
503 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>".
505 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) {
506 return MarkUnsafe(Info);
510 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices??
512 bool IsAllZeroIndices = true;
514 // If the first index is a non-constant index into an array, see if we can
515 // handle it as a special case.
516 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
517 if (!isa<ConstantInt>(I.getOperand())) {
518 IsAllZeroIndices = 0;
519 uint64_t NumElements = AT->getNumElements();
521 // If this is an array index and the index is not constant, we cannot
522 // promote... that is unless the array has exactly one or two elements in
523 // it, in which case we CAN promote it, but we have to canonicalize this
524 // out if this is the only problem.
525 if ((NumElements == 1 || NumElements == 2) &&
526 AllUsersAreLoads(GEPI)) {
527 Info.needsCanon = true;
528 return; // Canonicalization required!
530 return MarkUnsafe(Info);
534 bool hasVector = false;
536 // Walk through the GEP type indices, checking the types that this indexes
538 for (; I != E; ++I) {
539 // Ignore struct elements, no extra checking needed for these.
540 if (isa<StructType>(*I))
543 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand());
544 if (!IdxVal) return MarkUnsafe(Info);
546 // Are all indices still zero?
547 IsAllZeroIndices &= IdxVal->isZero();
549 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
550 // This GEP indexes an array. Verify that this is an in-range constant
551 // integer. Specifically, consider A[0][i]. We cannot know that the user
552 // isn't doing invalid things like allowing i to index an out-of-range
553 // subscript that accesses A[1]. Because of this, we have to reject SROA
554 // of any accesses into structs where any of the components are variables.
555 if (IdxVal->getZExtValue() >= AT->getNumElements())
556 return MarkUnsafe(Info);
559 // Note if we've seen a vector type yet
560 hasVector |= isa<VectorType>(*I);
562 // Don't SROA pointers into vectors, unless all indices are zero. When all
563 // indices are zero, we only consider this GEP as a bitcast, but will still
564 // not consider breaking up the vector.
565 if (hasVector && !IsAllZeroIndices)
566 return MarkUnsafe(Info);
569 // If there are any non-simple uses of this getelementptr, make sure to reject
571 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info);
574 /// isSafeMemIntrinsicOnAllocation - Return true if the specified memory
575 /// intrinsic can be promoted by SROA. At this point, we know that the operand
576 /// of the memintrinsic is a pointer to the beginning of the allocation.
577 void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI,
578 unsigned OpNo, AllocaInfo &Info) {
579 // If not constant length, give up.
580 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength());
581 if (!Length) return MarkUnsafe(Info);
583 // If not the whole aggregate, give up.
584 const TargetData &TD = getAnalysis<TargetData>();
585 if (Length->getZExtValue() !=
586 TD.getABITypeSize(AI->getType()->getElementType()))
587 return MarkUnsafe(Info);
589 // We only know about memcpy/memset/memmove.
590 if (!isa<MemCpyInst>(MI) && !isa<MemSetInst>(MI) && !isa<MemMoveInst>(MI))
591 return MarkUnsafe(Info);
593 // Otherwise, we can transform it. Determine whether this is a memcpy/set
594 // into or out of the aggregate.
596 Info.isMemCpyDst = true;
599 Info.isMemCpySrc = true;
603 /// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast
605 void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI,
607 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end();
609 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) {
610 isSafeUseOfBitCastedAllocation(BCU, AI, Info);
611 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) {
612 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info);
614 return MarkUnsafe(Info);
616 if (Info.isUnsafe) return;
620 /// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes
621 /// to its first element. Transform users of the cast to use the new values
623 void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI,
624 SmallVector<AllocaInst*, 32> &NewElts) {
625 Constant *Zero = Constant::getNullValue(Type::Int32Ty);
626 const TargetData &TD = getAnalysis<TargetData>();
628 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end();
630 if (BitCastInst *BCU = dyn_cast<BitCastInst>(*UI)) {
631 RewriteBitCastUserOfAlloca(BCU, AI, NewElts);
633 BCU->eraseFromParent();
637 // Otherwise, must be memcpy/memmove/memset of the entire aggregate. Split
638 // into one per element.
639 MemIntrinsic *MI = dyn_cast<MemIntrinsic>(*UI);
641 // If it's not a mem intrinsic, it must be some other user of a gep of the
642 // first pointer. Just leave these alone.
648 // If this is a memcpy/memmove, construct the other pointer as the
651 if (MemCpyInst *MCI = dyn_cast<MemCpyInst>(MI)) {
652 if (BCInst == MCI->getRawDest())
653 OtherPtr = MCI->getRawSource();
655 assert(BCInst == MCI->getRawSource());
656 OtherPtr = MCI->getRawDest();
658 } else if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
659 if (BCInst == MMI->getRawDest())
660 OtherPtr = MMI->getRawSource();
662 assert(BCInst == MMI->getRawSource());
663 OtherPtr = MMI->getRawDest();
667 // If there is an other pointer, we want to convert it to the same pointer
668 // type as AI has, so we can GEP through it.
670 // It is likely that OtherPtr is a bitcast, if so, remove it.
671 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr))
672 OtherPtr = BC->getOperand(0);
673 // All zero GEPs are effectively casts
674 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr))
675 if (GEP->hasAllZeroIndices())
676 OtherPtr = GEP->getOperand(0);
678 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr))
679 if (BCE->getOpcode() == Instruction::BitCast)
680 OtherPtr = BCE->getOperand(0);
682 // If the pointer is not the right type, insert a bitcast to the right
684 if (OtherPtr->getType() != AI->getType())
685 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(),
689 // Process each element of the aggregate.
690 Value *TheFn = MI->getOperand(0);
691 const Type *BytePtrTy = MI->getRawDest()->getType();
692 bool SROADest = MI->getRawDest() == BCInst;
694 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) {
695 // If this is a memcpy/memmove, emit a GEP of the other element address.
698 Value *Idx[2] = { Zero, ConstantInt::get(Type::Int32Ty, i) };
699 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2,
700 OtherPtr->getNameStr()+"."+utostr(i),
704 Value *EltPtr = NewElts[i];
705 const Type *EltTy =cast<PointerType>(EltPtr->getType())->getElementType();
707 // If we got down to a scalar, insert a load or store as appropriate.
708 if (EltTy->isSingleValueType()) {
709 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
710 Value *Elt = new LoadInst(SROADest ? OtherElt : EltPtr, "tmp",
712 new StoreInst(Elt, SROADest ? EltPtr : OtherElt, MI);
715 assert(isa<MemSetInst>(MI));
717 // If the stored element is zero (common case), just store a null
720 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) {
722 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0>
724 // If EltTy is a vector type, get the element type.
725 const Type *ValTy = EltTy;
726 if (const VectorType *VTy = dyn_cast<VectorType>(ValTy))
727 ValTy = VTy->getElementType();
729 // Construct an integer with the right value.
730 unsigned EltSize = TD.getTypeSizeInBits(ValTy);
731 APInt OneVal(EltSize, CI->getZExtValue());
732 APInt TotalVal(OneVal);
734 for (unsigned i = 0; 8*i < EltSize; ++i) {
735 TotalVal = TotalVal.shl(8);
739 // Convert the integer value to the appropriate type.
740 StoreVal = ConstantInt::get(TotalVal);
741 if (isa<PointerType>(ValTy))
742 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy);
743 else if (ValTy->isFloatingPoint())
744 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy);
745 assert(StoreVal->getType() == ValTy && "Type mismatch!");
747 // If the requested value was a vector constant, create it.
748 if (EltTy != ValTy) {
749 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements();
750 SmallVector<Constant*, 16> Elts(NumElts, StoreVal);
751 StoreVal = ConstantVector::get(&Elts[0], NumElts);
754 new StoreInst(StoreVal, EltPtr, MI);
757 // Otherwise, if we're storing a byte variable, use a memset call for
762 // Cast the element pointer to BytePtrTy.
763 if (EltPtr->getType() != BytePtrTy)
764 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI);
766 // Cast the other pointer (if we have one) to BytePtrTy.
767 if (OtherElt && OtherElt->getType() != BytePtrTy)
768 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(),
771 unsigned EltSize = TD.getABITypeSize(EltTy);
773 // Finally, insert the meminst for this element.
774 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
776 SROADest ? EltPtr : OtherElt, // Dest ptr
777 SROADest ? OtherElt : EltPtr, // Src ptr
778 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
781 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
783 assert(isa<MemSetInst>(MI));
785 EltPtr, MI->getOperand(2), // Dest, Value,
786 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size
789 CallInst::Create(TheFn, Ops, Ops + 4, "", MI);
793 // Finally, MI is now dead, as we've modified its actions to occur on all of
794 // the elements of the aggregate.
796 MI->eraseFromParent();
800 /// HasPadding - Return true if the specified type has any structure or
801 /// alignment padding, false otherwise.
802 static bool HasPadding(const Type *Ty, const TargetData &TD) {
803 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
804 const StructLayout *SL = TD.getStructLayout(STy);
805 unsigned PrevFieldBitOffset = 0;
806 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
807 unsigned FieldBitOffset = SL->getElementOffsetInBits(i);
809 // Padding in sub-elements?
810 if (HasPadding(STy->getElementType(i), TD))
813 // Check to see if there is any padding between this element and the
816 unsigned PrevFieldEnd =
817 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1));
818 if (PrevFieldEnd < FieldBitOffset)
822 PrevFieldBitOffset = FieldBitOffset;
825 // Check for tail padding.
826 if (unsigned EltCount = STy->getNumElements()) {
827 unsigned PrevFieldEnd = PrevFieldBitOffset +
828 TD.getTypeSizeInBits(STy->getElementType(EltCount-1));
829 if (PrevFieldEnd < SL->getSizeInBits())
833 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
834 return HasPadding(ATy->getElementType(), TD);
835 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
836 return HasPadding(VTy->getElementType(), TD);
838 return TD.getTypeSizeInBits(Ty) != TD.getABITypeSizeInBits(Ty);
841 /// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of
842 /// an aggregate can be broken down into elements. Return 0 if not, 3 if safe,
843 /// or 1 if safe after canonicalization has been performed.
845 int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) {
846 // Loop over the use list of the alloca. We can only transform it if all of
847 // the users are safe to transform.
850 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end();
852 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info);
854 DOUT << "Cannot transform: " << *AI << " due to user: " << **I;
859 // Okay, we know all the users are promotable. If the aggregate is a memcpy
860 // source and destination, we have to be careful. In particular, the memcpy
861 // could be moving around elements that live in structure padding of the LLVM
862 // types, but may actually be used. In these cases, we refuse to promote the
864 if (Info.isMemCpySrc && Info.isMemCpyDst &&
865 HasPadding(AI->getType()->getElementType(), getAnalysis<TargetData>()))
868 // If we require cleanup, return 1, otherwise return 3.
869 return Info.needsCanon ? 1 : 3;
872 /// CanonicalizeAllocaUsers - If SROA reported that it can promote the specified
873 /// allocation, but only if cleaned up, perform the cleanups required.
874 void SROA::CanonicalizeAllocaUsers(AllocationInst *AI) {
875 // At this point, we know that the end result will be SROA'd and promoted, so
876 // we can insert ugly code if required so long as sroa+mem2reg will clean it
878 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
880 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI++);
882 gep_type_iterator I = gep_type_begin(GEPI);
885 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) {
886 uint64_t NumElements = AT->getNumElements();
888 if (!isa<ConstantInt>(I.getOperand())) {
889 if (NumElements == 1) {
890 GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty));
892 assert(NumElements == 2 && "Unhandled case!");
893 // All users of the GEP must be loads. At each use of the GEP, insert
894 // two loads of the appropriate indexed GEP and select between them.
895 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(),
896 Constant::getNullValue(I.getOperand()->getType()),
898 // Insert the new GEP instructions, which are properly indexed.
899 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end());
900 Indices[1] = Constant::getNullValue(Type::Int32Ty);
901 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
904 GEPI->getName()+".0", GEPI);
905 Indices[1] = ConstantInt::get(Type::Int32Ty, 1);
906 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0),
909 GEPI->getName()+".1", GEPI);
910 // Replace all loads of the variable index GEP with loads from both
911 // indexes and a select.
912 while (!GEPI->use_empty()) {
913 LoadInst *LI = cast<LoadInst>(GEPI->use_back());
914 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI);
915 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI);
916 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI);
917 LI->replaceAllUsesWith(R);
918 LI->eraseFromParent();
920 GEPI->eraseFromParent();
927 /// MergeInType - Add the 'In' type to the accumulated type so far. If the
928 /// types are incompatible, return true, otherwise update Accum and return
931 /// There are three cases we handle here:
932 /// 1) An effectively-integer union, where the pieces are stored into as
933 /// smaller integers (common with byte swap and other idioms).
934 /// 2) A union of vector types of the same size and potentially its elements.
935 /// Here we turn element accesses into insert/extract element operations.
936 /// 3) A union of scalar types, such as int/float or int/pointer. Here we
937 /// merge together into integers, allowing the xform to work with #1 as
939 static bool MergeInType(const Type *In, const Type *&Accum,
940 const TargetData &TD) {
941 // If this is our first type, just use it.
942 const VectorType *PTy;
943 if (Accum == Type::VoidTy || In == Accum) {
945 } else if (In == Type::VoidTy) {
947 } else if (In->isInteger() && Accum->isInteger()) { // integer union.
948 // Otherwise pick whichever type is larger.
949 if (cast<IntegerType>(In)->getBitWidth() >
950 cast<IntegerType>(Accum)->getBitWidth())
952 } else if (isa<PointerType>(In) && isa<PointerType>(Accum)) {
953 // Pointer unions just stay as one of the pointers.
954 } else if (isa<VectorType>(In) || isa<VectorType>(Accum)) {
955 if ((PTy = dyn_cast<VectorType>(Accum)) &&
956 PTy->getElementType() == In) {
957 // Accum is a vector, and we are accessing an element: ok.
958 } else if ((PTy = dyn_cast<VectorType>(In)) &&
959 PTy->getElementType() == Accum) {
960 // In is a vector, and accum is an element: ok, remember In.
962 } else if ((PTy = dyn_cast<VectorType>(In)) && isa<VectorType>(Accum) &&
963 PTy->getBitWidth() == cast<VectorType>(Accum)->getBitWidth()) {
964 // Two vectors of the same size: keep Accum.
966 // Cannot insert an short into a <4 x int> or handle
967 // <2 x int> -> <4 x int>
971 // Pointer/FP/Integer unions merge together as integers.
972 switch (Accum->getTypeID()) {
973 case Type::PointerTyID: Accum = TD.getIntPtrType(); break;
974 case Type::FloatTyID: Accum = Type::Int32Ty; break;
975 case Type::DoubleTyID: Accum = Type::Int64Ty; break;
976 case Type::X86_FP80TyID: return true;
977 case Type::FP128TyID: return true;
978 case Type::PPC_FP128TyID: return true;
980 assert(Accum->isInteger() && "Unknown FP type!");
984 switch (In->getTypeID()) {
985 case Type::PointerTyID: In = TD.getIntPtrType(); break;
986 case Type::FloatTyID: In = Type::Int32Ty; break;
987 case Type::DoubleTyID: In = Type::Int64Ty; break;
988 case Type::X86_FP80TyID: return true;
989 case Type::FP128TyID: return true;
990 case Type::PPC_FP128TyID: return true;
992 assert(In->isInteger() && "Unknown FP type!");
995 return MergeInType(In, Accum, TD);
1000 /// getUIntAtLeastAsBigAs - Return an unsigned integer type that is at least
1001 /// as big as the specified type. If there is no suitable type, this returns
1003 const Type *getUIntAtLeastAsBigAs(unsigned NumBits) {
1004 if (NumBits > 64) return 0;
1005 if (NumBits > 32) return Type::Int64Ty;
1006 if (NumBits > 16) return Type::Int32Ty;
1007 if (NumBits > 8) return Type::Int16Ty;
1008 return Type::Int8Ty;
1011 /// CanConvertToScalar - V is a pointer. If we can convert the pointee to a
1012 /// single scalar integer type, return that type. Further, if the use is not
1013 /// a completely trivial use that mem2reg could promote, set IsNotTrivial. If
1014 /// there are no uses of this pointer, return Type::VoidTy to differentiate from
1017 const Type *SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial) {
1018 const Type *UsedType = Type::VoidTy; // No uses, no forced type.
1019 const TargetData &TD = getAnalysis<TargetData>();
1020 const PointerType *PTy = cast<PointerType>(V->getType());
1022 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1023 Instruction *User = cast<Instruction>(*UI);
1025 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1026 // FIXME: Loads of a first class aggregrate value could be converted to a
1027 // series of loads and insertvalues
1028 if (!LI->getType()->isSingleValueType())
1031 if (MergeInType(LI->getType(), UsedType, TD))
1034 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1035 // Storing the pointer, not into the value?
1036 if (SI->getOperand(0) == V) return 0;
1038 // FIXME: Stores of a first class aggregrate value could be converted to a
1039 // series of extractvalues and stores
1040 if (!SI->getOperand(0)->getType()->isSingleValueType())
1043 // NOTE: We could handle storing of FP imms into integers here!
1045 if (MergeInType(SI->getOperand(0)->getType(), UsedType, TD))
1047 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
1048 IsNotTrivial = true;
1049 const Type *SubTy = CanConvertToScalar(CI, IsNotTrivial);
1050 if (!SubTy || MergeInType(SubTy, UsedType, TD)) return 0;
1051 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1052 // Check to see if this is stepping over an element: GEP Ptr, int C
1053 if (GEP->getNumOperands() == 2 && isa<ConstantInt>(GEP->getOperand(1))) {
1054 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue();
1055 unsigned ElSize = TD.getABITypeSize(PTy->getElementType());
1056 unsigned BitOffset = Idx*ElSize*8;
1057 if (BitOffset > 64 || !isPowerOf2_32(ElSize)) return 0;
1059 IsNotTrivial = true;
1060 const Type *SubElt = CanConvertToScalar(GEP, IsNotTrivial);
1061 if (SubElt == 0) return 0;
1062 if (SubElt != Type::VoidTy && SubElt->isInteger()) {
1064 getUIntAtLeastAsBigAs(TD.getABITypeSizeInBits(SubElt)+BitOffset);
1065 if (NewTy == 0 || MergeInType(NewTy, UsedType, TD)) return 0;
1068 } else if (GEP->getNumOperands() == 3 &&
1069 isa<ConstantInt>(GEP->getOperand(1)) &&
1070 isa<ConstantInt>(GEP->getOperand(2)) &&
1071 cast<ConstantInt>(GEP->getOperand(1))->isZero()) {
1072 // We are stepping into an element, e.g. a structure or an array:
1073 // GEP Ptr, int 0, uint C
1074 const Type *AggTy = PTy->getElementType();
1075 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
1077 if (const ArrayType *ATy = dyn_cast<ArrayType>(AggTy)) {
1078 if (Idx >= ATy->getNumElements()) return 0; // Out of range.
1079 } else if (const VectorType *VectorTy = dyn_cast<VectorType>(AggTy)) {
1080 // Getting an element of the vector.
1081 if (Idx >= VectorTy->getNumElements()) return 0; // Out of range.
1083 // Merge in the vector type.
1084 if (MergeInType(VectorTy, UsedType, TD)) return 0;
1086 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial);
1087 if (SubTy == 0) return 0;
1089 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, TD))
1092 // We'll need to change this to an insert/extract element operation.
1093 IsNotTrivial = true;
1094 continue; // Everything looks ok
1096 } else if (isa<StructType>(AggTy)) {
1097 // Structs are always ok.
1101 const Type *NTy = getUIntAtLeastAsBigAs(TD.getABITypeSizeInBits(AggTy));
1102 if (NTy == 0 || MergeInType(NTy, UsedType, TD)) return 0;
1103 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial);
1104 if (SubTy == 0) return 0;
1105 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, TD))
1107 continue; // Everything looks ok
1111 // Cannot handle this!
1119 /// ConvertToScalar - The specified alloca passes the CanConvertToScalar
1120 /// predicate and is non-trivial. Convert it to something that can be trivially
1121 /// promoted into a register by mem2reg.
1122 void SROA::ConvertToScalar(AllocationInst *AI, const Type *ActualTy) {
1123 DOUT << "CONVERT TO SCALAR: " << *AI << " TYPE = "
1124 << *ActualTy << "\n";
1127 BasicBlock *EntryBlock = AI->getParent();
1128 assert(EntryBlock == &EntryBlock->getParent()->getEntryBlock() &&
1129 "Not in the entry block!");
1130 EntryBlock->getInstList().remove(AI); // Take the alloca out of the program.
1132 // Create and insert the alloca.
1133 AllocaInst *NewAI = new AllocaInst(ActualTy, 0, AI->getName(),
1134 EntryBlock->begin());
1135 ConvertUsesToScalar(AI, NewAI, 0);
1140 /// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca
1141 /// directly. This happens when we are converting an "integer union" to a
1142 /// single integer scalar, or when we are converting a "vector union" to a
1143 /// vector with insert/extractelement instructions.
1145 /// Offset is an offset from the original alloca, in bits that need to be
1146 /// shifted to the right. By the end of this, there should be no uses of Ptr.
1147 void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset) {
1148 while (!Ptr->use_empty()) {
1149 Instruction *User = cast<Instruction>(Ptr->use_back());
1151 if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
1152 Value *NV = ConvertUsesOfLoadToScalar(LI, NewAI, Offset);
1153 LI->replaceAllUsesWith(NV);
1154 LI->eraseFromParent();
1155 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
1156 assert(SI->getOperand(0) != Ptr && "Consistency error!");
1158 Value *SV = ConvertUsesOfStoreToScalar(SI, NewAI, Offset);
1159 new StoreInst(SV, NewAI, SI);
1160 SI->eraseFromParent();
1162 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) {
1163 ConvertUsesToScalar(CI, NewAI, Offset);
1164 CI->eraseFromParent();
1165 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) {
1166 const PointerType *AggPtrTy =
1167 cast<PointerType>(GEP->getOperand(0)->getType());
1168 const TargetData &TD = getAnalysis<TargetData>();
1169 unsigned AggSizeInBits =
1170 TD.getABITypeSizeInBits(AggPtrTy->getElementType());
1172 // Check to see if this is stepping over an element: GEP Ptr, int C
1173 unsigned NewOffset = Offset;
1174 if (GEP->getNumOperands() == 2) {
1175 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue();
1176 unsigned BitOffset = Idx*AggSizeInBits;
1178 NewOffset += BitOffset;
1179 } else if (GEP->getNumOperands() == 3) {
1180 // We know that operand #2 is zero.
1181 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
1182 const Type *AggTy = AggPtrTy->getElementType();
1183 if (const SequentialType *SeqTy = dyn_cast<SequentialType>(AggTy)) {
1184 unsigned ElSizeBits =
1185 TD.getABITypeSizeInBits(SeqTy->getElementType());
1187 NewOffset += ElSizeBits*Idx;
1188 } else if (const StructType *STy = dyn_cast<StructType>(AggTy)) {
1189 unsigned EltBitOffset =
1190 TD.getStructLayout(STy)->getElementOffsetInBits(Idx);
1192 NewOffset += EltBitOffset;
1194 assert(0 && "Unsupported operation!");
1198 assert(0 && "Unsupported operation!");
1201 ConvertUsesToScalar(GEP, NewAI, NewOffset);
1202 GEP->eraseFromParent();
1204 assert(0 && "Unsupported operation!");
1210 /// ConvertUsesOfLoadToScalar - Convert all of the users the specified load to
1211 /// use the new alloca directly, returning the value that should replace the
1212 /// load. This happens when we are converting an "integer union" to a
1213 /// single integer scalar, or when we are converting a "vector union" to a
1214 /// vector with insert/extractelement instructions.
1216 /// Offset is an offset from the original alloca, in bits that need to be
1217 /// shifted to the right. By the end of this, there should be no uses of Ptr.
1218 Value *SROA::ConvertUsesOfLoadToScalar(LoadInst *LI, AllocaInst *NewAI,
1220 // The load is a bit extract from NewAI shifted right by Offset bits.
1221 Value *NV = new LoadInst(NewAI, LI->getName(), LI);
1223 if (NV->getType() == LI->getType() && Offset == 0) {
1224 // We win, no conversion needed.
1228 // If the result type of the 'union' is a pointer, then this must be ptr->ptr
1229 // cast. Anything else would result in NV being an integer.
1230 if (isa<PointerType>(NV->getType())) {
1231 assert(isa<PointerType>(LI->getType()));
1232 return new BitCastInst(NV, LI->getType(), LI->getName(), LI);
1235 if (const VectorType *VTy = dyn_cast<VectorType>(NV->getType())) {
1236 // If the result alloca is a vector type, this is either an element
1237 // access or a bitcast to another vector type.
1238 if (isa<VectorType>(LI->getType()))
1239 return new BitCastInst(NV, LI->getType(), LI->getName(), LI);
1241 // Otherwise it must be an element access.
1242 const TargetData &TD = getAnalysis<TargetData>();
1245 unsigned EltSize = TD.getABITypeSizeInBits(VTy->getElementType());
1246 Elt = Offset/EltSize;
1247 Offset -= EltSize*Elt;
1249 NV = new ExtractElementInst(NV, ConstantInt::get(Type::Int32Ty, Elt),
1252 // If we're done, return this element.
1253 if (NV->getType() == LI->getType() && Offset == 0)
1257 const IntegerType *NTy = cast<IntegerType>(NV->getType());
1259 // If this is a big-endian system and the load is narrower than the
1260 // full alloca type, we need to do a shift to get the right bits.
1262 const TargetData &TD = getAnalysis<TargetData>();
1263 if (TD.isBigEndian()) {
1264 // On big-endian machines, the lowest bit is stored at the bit offset
1265 // from the pointer given by getTypeStoreSizeInBits. This matters for
1266 // integers with a bitwidth that is not a multiple of 8.
1267 ShAmt = TD.getTypeStoreSizeInBits(NTy) -
1268 TD.getTypeStoreSizeInBits(LI->getType()) - Offset;
1273 // Note: we support negative bitwidths (with shl) which are not defined.
1274 // We do this to support (f.e.) loads off the end of a structure where
1275 // only some bits are used.
1276 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth())
1277 NV = BinaryOperator::CreateLShr(NV,
1278 ConstantInt::get(NV->getType(),ShAmt),
1280 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth())
1281 NV = BinaryOperator::CreateShl(NV,
1282 ConstantInt::get(NV->getType(),-ShAmt),
1285 // Finally, unconditionally truncate the integer to the right width.
1286 unsigned LIBitWidth = TD.getTypeSizeInBits(LI->getType());
1287 if (LIBitWidth < NTy->getBitWidth())
1288 NV = new TruncInst(NV, IntegerType::get(LIBitWidth),
1291 // If the result is an integer, this is a trunc or bitcast.
1292 if (isa<IntegerType>(LI->getType())) {
1294 } else if (LI->getType()->isFloatingPoint()) {
1295 // Just do a bitcast, we know the sizes match up.
1296 NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI);
1298 // Otherwise must be a pointer.
1299 NV = new IntToPtrInst(NV, LI->getType(), LI->getName(), LI);
1301 assert(NV->getType() == LI->getType() && "Didn't convert right?");
1306 /// ConvertUsesOfStoreToScalar - Convert the specified store to a load+store
1307 /// pair of the new alloca directly, returning the value that should be stored
1308 /// to the alloca. This happens when we are converting an "integer union" to a
1309 /// single integer scalar, or when we are converting a "vector union" to a
1310 /// vector with insert/extractelement instructions.
1312 /// Offset is an offset from the original alloca, in bits that need to be
1313 /// shifted to the right. By the end of this, there should be no uses of Ptr.
1314 Value *SROA::ConvertUsesOfStoreToScalar(StoreInst *SI, AllocaInst *NewAI,
1317 // Convert the stored type to the actual type, shift it left to insert
1318 // then 'or' into place.
1319 Value *SV = SI->getOperand(0);
1320 const Type *AllocaType = NewAI->getType()->getElementType();
1321 if (SV->getType() == AllocaType && Offset == 0) {
1323 } else if (const VectorType *PTy = dyn_cast<VectorType>(AllocaType)) {
1324 Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI);
1326 // If the result alloca is a vector type, this is either an element
1327 // access or a bitcast to another vector type.
1328 if (isa<VectorType>(SV->getType())) {
1329 SV = new BitCastInst(SV, AllocaType, SV->getName(), SI);
1331 // Must be an element insertion.
1332 const TargetData &TD = getAnalysis<TargetData>();
1333 unsigned Elt = Offset/TD.getABITypeSizeInBits(PTy->getElementType());
1334 SV = InsertElementInst::Create(Old, SV,
1335 ConstantInt::get(Type::Int32Ty, Elt),
1338 } else if (isa<PointerType>(AllocaType)) {
1339 // If the alloca type is a pointer, then all the elements must be
1341 if (SV->getType() != AllocaType)
1342 SV = new BitCastInst(SV, AllocaType, SV->getName(), SI);
1344 Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI);
1346 // If SV is a float, convert it to the appropriate integer type.
1347 // If it is a pointer, do the same, and also handle ptr->ptr casts
1349 const TargetData &TD = getAnalysis<TargetData>();
1350 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType());
1351 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType);
1352 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType());
1353 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType);
1354 if (SV->getType()->isFloatingPoint())
1355 SV = new BitCastInst(SV, IntegerType::get(SrcWidth),
1357 else if (isa<PointerType>(SV->getType()))
1358 SV = new PtrToIntInst(SV, TD.getIntPtrType(), SV->getName(), SI);
1360 // Always zero extend the value if needed.
1361 if (SV->getType() != AllocaType)
1362 SV = new ZExtInst(SV, AllocaType, SV->getName(), SI);
1364 // If this is a big-endian system and the store is narrower than the
1365 // full alloca type, we need to do a shift to get the right bits.
1367 if (TD.isBigEndian()) {
1368 // On big-endian machines, the lowest bit is stored at the bit offset
1369 // from the pointer given by getTypeStoreSizeInBits. This matters for
1370 // integers with a bitwidth that is not a multiple of 8.
1371 ShAmt = DestStoreWidth - SrcStoreWidth - Offset;
1376 // Note: we support negative bitwidths (with shr) which are not defined.
1377 // We do this to support (f.e.) stores off the end of a structure where
1378 // only some bits in the structure are set.
1379 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth));
1380 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) {
1381 SV = BinaryOperator::CreateShl(SV,
1382 ConstantInt::get(SV->getType(), ShAmt),
1385 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) {
1386 SV = BinaryOperator::CreateLShr(SV,
1387 ConstantInt::get(SV->getType(),-ShAmt),
1389 Mask = Mask.lshr(ShAmt);
1392 // Mask out the bits we are about to insert from the old value, and or
1394 if (SrcWidth != DestWidth) {
1395 assert(DestWidth > SrcWidth);
1396 Old = BinaryOperator::CreateAnd(Old, ConstantInt::get(~Mask),
1397 Old->getName()+".mask", SI);
1398 SV = BinaryOperator::CreateOr(Old, SV, SV->getName()+".ins", SI);
1406 /// PointsToConstantGlobal - Return true if V (possibly indirectly) points to
1407 /// some part of a constant global variable. This intentionally only accepts
1408 /// constant expressions because we don't can't rewrite arbitrary instructions.
1409 static bool PointsToConstantGlobal(Value *V) {
1410 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
1411 return GV->isConstant();
1412 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
1413 if (CE->getOpcode() == Instruction::BitCast ||
1414 CE->getOpcode() == Instruction::GetElementPtr)
1415 return PointsToConstantGlobal(CE->getOperand(0));
1419 /// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
1420 /// pointer to an alloca. Ignore any reads of the pointer, return false if we
1421 /// see any stores or other unknown uses. If we see pointer arithmetic, keep
1422 /// track of whether it moves the pointer (with isOffset) but otherwise traverse
1423 /// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
1424 /// the alloca, and if the source pointer is a pointer to a constant global, we
1425 /// can optimize this.
1426 static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy,
1428 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
1429 if (isa<LoadInst>(*UI)) {
1430 // Ignore loads, they are always ok.
1433 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) {
1434 // If uses of the bitcast are ok, we are ok.
1435 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset))
1439 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
1440 // If the GEP has all zero indices, it doesn't offset the pointer. If it
1441 // doesn't, it does.
1442 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy,
1443 isOffset || !GEP->hasAllZeroIndices()))
1448 // If this is isn't our memcpy/memmove, reject it as something we can't
1450 if (!isa<MemCpyInst>(*UI) && !isa<MemMoveInst>(*UI))
1453 // If we already have seen a copy, reject the second one.
1454 if (TheCopy) return false;
1456 // If the pointer has been offset from the start of the alloca, we can't
1457 // safely handle this.
1458 if (isOffset) return false;
1460 // If the memintrinsic isn't using the alloca as the dest, reject it.
1461 if (UI.getOperandNo() != 1) return false;
1463 MemIntrinsic *MI = cast<MemIntrinsic>(*UI);
1465 // If the source of the memcpy/move is not a constant global, reject it.
1466 if (!PointsToConstantGlobal(MI->getOperand(2)))
1469 // Otherwise, the transform is safe. Remember the copy instruction.
1475 /// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only
1476 /// modified by a copy from a constant global. If we can prove this, we can
1477 /// replace any uses of the alloca with uses of the global directly.
1478 Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) {
1479 Instruction *TheCopy = 0;
1480 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false))