1 //===- SROA.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. It tries to identify promotable elements of an
12 /// aggregate alloca, and promote them to registers. It will also try to
13 /// convert uses of an element (or set of elements) of an alloca into a vector
14 /// or bitfield-style integer scalar if appropriate.
16 /// It works to do this with minimal slicing of the alloca so that regions
17 /// which are merely transferred in and out of external memory remain unchanged
18 /// and are not decomposed to scalar code.
20 /// Because this also performs alloca promotion, it can be thought of as also
21 /// serving the purpose of SSA formation. The algorithm iterates on the
22 /// function until all opportunities for promotion have been realized.
24 //===----------------------------------------------------------------------===//
26 #define DEBUG_TYPE "sroa"
27 #include "llvm/Transforms/Scalar.h"
28 #include "llvm/ADT/STLExtras.h"
29 #include "llvm/ADT/SetVector.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Analysis/Loads.h"
33 #include "llvm/Analysis/PtrUseVisitor.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/DIBuilder.h"
36 #include "llvm/DebugInfo.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/DataLayout.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/LLVMContext.h"
46 #include "llvm/IR/Operator.h"
47 #include "llvm/InstVisitor.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Compiler.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Support/TimeValue.h"
55 #include "llvm/Support/raw_ostream.h"
56 #include "llvm/Transforms/Utils/Local.h"
57 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
58 #include "llvm/Transforms/Utils/SSAUpdater.h"
60 #if __cplusplus >= 201103L && !defined(NDEBUG)
61 // We only use this for a debug check in C++11
67 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
68 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
69 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
70 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
71 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
72 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
73 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
74 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
75 STATISTIC(NumDeleted, "Number of instructions deleted");
76 STATISTIC(NumVectorized, "Number of vectorized aggregates");
78 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
79 /// forming SSA values through the SSAUpdater infrastructure.
81 ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
83 /// Hidden option to enable randomly shuffling the slices to help uncover
84 /// instability in their order.
85 static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
86 cl::init(false), cl::Hidden);
88 /// Hidden option to experiment with completely strict handling of inbounds
90 static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds",
91 cl::init(false), cl::Hidden);
94 /// \brief A custom IRBuilder inserter which prefixes all names if they are
96 template <bool preserveNames = true>
97 class IRBuilderPrefixedInserter :
98 public IRBuilderDefaultInserter<preserveNames> {
102 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
105 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
106 BasicBlock::iterator InsertPt) const {
107 IRBuilderDefaultInserter<preserveNames>::InsertHelper(
108 I, Name.isTriviallyEmpty() ? Name : Prefix + Name, BB, InsertPt);
112 // Specialization for not preserving the name is trivial.
114 class IRBuilderPrefixedInserter<false> :
115 public IRBuilderDefaultInserter<false> {
117 void SetNamePrefix(const Twine &P) {}
120 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
122 typedef llvm::IRBuilder<true, ConstantFolder,
123 IRBuilderPrefixedInserter<true> > IRBuilderTy;
125 typedef llvm::IRBuilder<false, ConstantFolder,
126 IRBuilderPrefixedInserter<false> > IRBuilderTy;
131 /// \brief A used slice of an alloca.
133 /// This structure represents a slice of an alloca used by some instruction. It
134 /// stores both the begin and end offsets of this use, a pointer to the use
135 /// itself, and a flag indicating whether we can classify the use as splittable
136 /// or not when forming partitions of the alloca.
138 /// \brief The beginning offset of the range.
139 uint64_t BeginOffset;
141 /// \brief The ending offset, not included in the range.
144 /// \brief Storage for both the use of this slice and whether it can be
146 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
149 Slice() : BeginOffset(), EndOffset() {}
150 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
151 : BeginOffset(BeginOffset), EndOffset(EndOffset),
152 UseAndIsSplittable(U, IsSplittable) {}
154 uint64_t beginOffset() const { return BeginOffset; }
155 uint64_t endOffset() const { return EndOffset; }
157 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
158 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
160 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
162 bool isDead() const { return getUse() == 0; }
163 void kill() { UseAndIsSplittable.setPointer(0); }
165 /// \brief Support for ordering ranges.
167 /// This provides an ordering over ranges such that start offsets are
168 /// always increasing, and within equal start offsets, the end offsets are
169 /// decreasing. Thus the spanning range comes first in a cluster with the
170 /// same start position.
171 bool operator<(const Slice &RHS) const {
172 if (beginOffset() < RHS.beginOffset()) return true;
173 if (beginOffset() > RHS.beginOffset()) return false;
174 if (isSplittable() != RHS.isSplittable()) return !isSplittable();
175 if (endOffset() > RHS.endOffset()) return true;
179 /// \brief Support comparison with a single offset to allow binary searches.
180 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
181 uint64_t RHSOffset) {
182 return LHS.beginOffset() < RHSOffset;
184 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
186 return LHSOffset < RHS.beginOffset();
189 bool operator==(const Slice &RHS) const {
190 return isSplittable() == RHS.isSplittable() &&
191 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
193 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
195 } // end anonymous namespace
198 template <typename T> struct isPodLike;
199 template <> struct isPodLike<Slice> {
200 static const bool value = true;
205 /// \brief Representation of the alloca slices.
207 /// This class represents the slices of an alloca which are formed by its
208 /// various uses. If a pointer escapes, we can't fully build a representation
209 /// for the slices used and we reflect that in this structure. The uses are
210 /// stored, sorted by increasing beginning offset and with unsplittable slices
211 /// starting at a particular offset before splittable slices.
214 /// \brief Construct the slices of a particular alloca.
215 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
217 /// \brief Test whether a pointer to the allocation escapes our analysis.
219 /// If this is true, the slices are never fully built and should be
221 bool isEscaped() const { return PointerEscapingInstr; }
223 /// \brief Support for iterating over the slices.
225 typedef SmallVectorImpl<Slice>::iterator iterator;
226 iterator begin() { return Slices.begin(); }
227 iterator end() { return Slices.end(); }
229 typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
230 const_iterator begin() const { return Slices.begin(); }
231 const_iterator end() const { return Slices.end(); }
234 /// \brief Allow iterating the dead users for this alloca.
236 /// These are instructions which will never actually use the alloca as they
237 /// are outside the allocated range. They are safe to replace with undef and
240 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
241 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
242 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
245 /// \brief Allow iterating the dead expressions referring to this alloca.
247 /// These are operands which have cannot actually be used to refer to the
248 /// alloca as they are outside its range and the user doesn't correct for
249 /// that. These mostly consist of PHI node inputs and the like which we just
250 /// need to replace with undef.
252 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
253 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
254 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
257 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
258 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
259 void printSlice(raw_ostream &OS, const_iterator I,
260 StringRef Indent = " ") const;
261 void printUse(raw_ostream &OS, const_iterator I,
262 StringRef Indent = " ") const;
263 void print(raw_ostream &OS) const;
264 void dump(const_iterator I) const;
269 template <typename DerivedT, typename RetT = void> class BuilderBase;
271 friend class AllocaSlices::SliceBuilder;
273 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
274 /// \brief Handle to alloca instruction to simplify method interfaces.
278 /// \brief The instruction responsible for this alloca not having a known set
281 /// When an instruction (potentially) escapes the pointer to the alloca, we
282 /// store a pointer to that here and abort trying to form slices of the
283 /// alloca. This will be null if the alloca slices are analyzed successfully.
284 Instruction *PointerEscapingInstr;
286 /// \brief The slices of the alloca.
288 /// We store a vector of the slices formed by uses of the alloca here. This
289 /// vector is sorted by increasing begin offset, and then the unsplittable
290 /// slices before the splittable ones. See the Slice inner class for more
292 SmallVector<Slice, 8> Slices;
294 /// \brief Instructions which will become dead if we rewrite the alloca.
296 /// Note that these are not separated by slice. This is because we expect an
297 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
298 /// all these instructions can simply be removed and replaced with undef as
299 /// they come from outside of the allocated space.
300 SmallVector<Instruction *, 8> DeadUsers;
302 /// \brief Operands which will become dead if we rewrite the alloca.
304 /// These are operands that in their particular use can be replaced with
305 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
306 /// to PHI nodes and the like. They aren't entirely dead (there might be
307 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
308 /// want to swap this particular input for undef to simplify the use lists of
310 SmallVector<Use *, 8> DeadOperands;
314 static Value *foldSelectInst(SelectInst &SI) {
315 // If the condition being selected on is a constant or the same value is
316 // being selected between, fold the select. Yes this does (rarely) happen
318 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
319 return SI.getOperand(1+CI->isZero());
320 if (SI.getOperand(1) == SI.getOperand(2))
321 return SI.getOperand(1);
326 /// \brief Builder for the alloca slices.
328 /// This class builds a set of alloca slices by recursively visiting the uses
329 /// of an alloca and making a slice for each load and store at each offset.
330 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
331 friend class PtrUseVisitor<SliceBuilder>;
332 friend class InstVisitor<SliceBuilder>;
333 typedef PtrUseVisitor<SliceBuilder> Base;
335 const uint64_t AllocSize;
338 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
339 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
341 /// \brief Set to de-duplicate dead instructions found in the use walk.
342 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
345 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &S)
346 : PtrUseVisitor<SliceBuilder>(DL),
347 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), S(S) {}
350 void markAsDead(Instruction &I) {
351 if (VisitedDeadInsts.insert(&I))
352 S.DeadUsers.push_back(&I);
355 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
356 bool IsSplittable = false) {
357 // Completely skip uses which have a zero size or start either before or
358 // past the end of the allocation.
359 if (Size == 0 || Offset.isNegative() || Offset.uge(AllocSize)) {
360 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
361 << " which has zero size or starts outside of the "
362 << AllocSize << " byte alloca:\n"
363 << " alloca: " << S.AI << "\n"
364 << " use: " << I << "\n");
365 return markAsDead(I);
368 uint64_t BeginOffset = Offset.getZExtValue();
369 uint64_t EndOffset = BeginOffset + Size;
371 // Clamp the end offset to the end of the allocation. Note that this is
372 // formulated to handle even the case where "BeginOffset + Size" overflows.
373 // This may appear superficially to be something we could ignore entirely,
374 // but that is not so! There may be widened loads or PHI-node uses where
375 // some instructions are dead but not others. We can't completely ignore
376 // them, and so have to record at least the information here.
377 assert(AllocSize >= BeginOffset); // Established above.
378 if (Size > AllocSize - BeginOffset) {
379 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
380 << " to remain within the " << AllocSize << " byte alloca:\n"
381 << " alloca: " << S.AI << "\n"
382 << " use: " << I << "\n");
383 EndOffset = AllocSize;
386 S.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
389 void visitBitCastInst(BitCastInst &BC) {
391 return markAsDead(BC);
393 return Base::visitBitCastInst(BC);
396 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
397 if (GEPI.use_empty())
398 return markAsDead(GEPI);
400 if (SROAStrictInbounds && GEPI.isInBounds()) {
401 // FIXME: This is a manually un-factored variant of the basic code inside
402 // of GEPs with checking of the inbounds invariant specified in the
403 // langref in a very strict sense. If we ever want to enable
404 // SROAStrictInbounds, this code should be factored cleanly into
405 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
406 // by writing out the code here where we have tho underlying allocation
407 // size readily available.
408 APInt GEPOffset = Offset;
409 for (gep_type_iterator GTI = gep_type_begin(GEPI),
410 GTE = gep_type_end(GEPI);
412 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
416 // Handle a struct index, which adds its field offset to the pointer.
417 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
418 unsigned ElementIdx = OpC->getZExtValue();
419 const StructLayout *SL = DL.getStructLayout(STy);
421 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
423 // For array or vector indices, scale the index by the size of the type.
424 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
425 GEPOffset += Index * APInt(Offset.getBitWidth(),
426 DL.getTypeAllocSize(GTI.getIndexedType()));
429 // If this index has computed an intermediate pointer which is not
430 // inbounds, then the result of the GEP is a poison value and we can
431 // delete it and all uses.
432 if (GEPOffset.ugt(AllocSize))
433 return markAsDead(GEPI);
437 return Base::visitGetElementPtrInst(GEPI);
440 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
441 uint64_t Size, bool IsVolatile) {
442 // We allow splitting of loads and stores where the type is an integer type
443 // and cover the entire alloca. This prevents us from splitting over
445 // FIXME: In the great blue eventually, we should eagerly split all integer
446 // loads and stores, and then have a separate step that merges adjacent
447 // alloca partitions into a single partition suitable for integer widening.
448 // Or we should skip the merge step and rely on GVN and other passes to
449 // merge adjacent loads and stores that survive mem2reg.
451 Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize;
453 insertUse(I, Offset, Size, IsSplittable);
456 void visitLoadInst(LoadInst &LI) {
457 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
458 "All simple FCA loads should have been pre-split");
461 return PI.setAborted(&LI);
463 uint64_t Size = DL.getTypeStoreSize(LI.getType());
464 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
467 void visitStoreInst(StoreInst &SI) {
468 Value *ValOp = SI.getValueOperand();
470 return PI.setEscapedAndAborted(&SI);
472 return PI.setAborted(&SI);
474 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
476 // If this memory access can be shown to *statically* extend outside the
477 // bounds of of the allocation, it's behavior is undefined, so simply
478 // ignore it. Note that this is more strict than the generic clamping
479 // behavior of insertUse. We also try to handle cases which might run the
481 // FIXME: We should instead consider the pointer to have escaped if this
482 // function is being instrumented for addressing bugs or race conditions.
483 if (Offset.isNegative() || Size > AllocSize ||
484 Offset.ugt(AllocSize - Size)) {
485 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
486 << " which extends past the end of the " << AllocSize
488 << " alloca: " << S.AI << "\n"
489 << " use: " << SI << "\n");
490 return markAsDead(SI);
493 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
494 "All simple FCA stores should have been pre-split");
495 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
499 void visitMemSetInst(MemSetInst &II) {
500 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
501 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
502 if ((Length && Length->getValue() == 0) ||
503 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
504 // Zero-length mem transfer intrinsics can be ignored entirely.
505 return markAsDead(II);
508 return PI.setAborted(&II);
510 insertUse(II, Offset,
511 Length ? Length->getLimitedValue()
512 : AllocSize - Offset.getLimitedValue(),
516 void visitMemTransferInst(MemTransferInst &II) {
517 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
518 if (Length && Length->getValue() == 0)
519 // Zero-length mem transfer intrinsics can be ignored entirely.
520 return markAsDead(II);
522 // Because we can visit these intrinsics twice, also check to see if the
523 // first time marked this instruction as dead. If so, skip it.
524 if (VisitedDeadInsts.count(&II))
528 return PI.setAborted(&II);
530 // This side of the transfer is completely out-of-bounds, and so we can
531 // nuke the entire transfer. However, we also need to nuke the other side
532 // if already added to our partitions.
533 // FIXME: Yet another place we really should bypass this when
534 // instrumenting for ASan.
535 if (!Offset.isNegative() && Offset.uge(AllocSize)) {
536 SmallDenseMap<Instruction *, unsigned>::iterator MTPI = MemTransferSliceMap.find(&II);
537 if (MTPI != MemTransferSliceMap.end())
538 S.Slices[MTPI->second].kill();
539 return markAsDead(II);
542 uint64_t RawOffset = Offset.getLimitedValue();
543 uint64_t Size = Length ? Length->getLimitedValue()
544 : AllocSize - RawOffset;
546 // Check for the special case where the same exact value is used for both
548 if (*U == II.getRawDest() && *U == II.getRawSource()) {
549 // For non-volatile transfers this is a no-op.
550 if (!II.isVolatile())
551 return markAsDead(II);
553 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
556 // If we have seen both source and destination for a mem transfer, then
557 // they both point to the same alloca.
559 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
560 llvm::tie(MTPI, Inserted) =
561 MemTransferSliceMap.insert(std::make_pair(&II, S.Slices.size()));
562 unsigned PrevIdx = MTPI->second;
564 Slice &PrevP = S.Slices[PrevIdx];
566 // Check if the begin offsets match and this is a non-volatile transfer.
567 // In that case, we can completely elide the transfer.
568 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
570 return markAsDead(II);
573 // Otherwise we have an offset transfer within the same alloca. We can't
575 PrevP.makeUnsplittable();
578 // Insert the use now that we've fixed up the splittable nature.
579 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
581 // Check that we ended up with a valid index in the map.
582 assert(S.Slices[PrevIdx].getUse()->getUser() == &II &&
583 "Map index doesn't point back to a slice with this user.");
586 // Disable SRoA for any intrinsics except for lifetime invariants.
587 // FIXME: What about debug intrinsics? This matches old behavior, but
588 // doesn't make sense.
589 void visitIntrinsicInst(IntrinsicInst &II) {
591 return PI.setAborted(&II);
593 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
594 II.getIntrinsicID() == Intrinsic::lifetime_end) {
595 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
596 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
597 Length->getLimitedValue());
598 insertUse(II, Offset, Size, true);
602 Base::visitIntrinsicInst(II);
605 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
606 // We consider any PHI or select that results in a direct load or store of
607 // the same offset to be a viable use for slicing purposes. These uses
608 // are considered unsplittable and the size is the maximum loaded or stored
610 SmallPtrSet<Instruction *, 4> Visited;
611 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
612 Visited.insert(Root);
613 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
614 // If there are no loads or stores, the access is dead. We mark that as
615 // a size zero access.
618 Instruction *I, *UsedI;
619 llvm::tie(UsedI, I) = Uses.pop_back_val();
621 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
622 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
625 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
626 Value *Op = SI->getOperand(0);
629 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
633 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
634 if (!GEP->hasAllZeroIndices())
636 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
637 !isa<SelectInst>(I)) {
641 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
643 if (Visited.insert(cast<Instruction>(*UI)))
644 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
645 } while (!Uses.empty());
650 void visitPHINode(PHINode &PN) {
652 return markAsDead(PN);
654 return PI.setAborted(&PN);
656 // See if we already have computed info on this node.
657 uint64_t &PHISize = PHIOrSelectSizes[&PN];
659 // This is a new PHI node, check for an unsafe use of the PHI node.
660 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&PN, PHISize))
661 return PI.setAborted(UnsafeI);
664 // For PHI and select operands outside the alloca, we can't nuke the entire
665 // phi or select -- the other side might still be relevant, so we special
666 // case them here and use a separate structure to track the operands
667 // themselves which should be replaced with undef.
668 // FIXME: This should instead be escaped in the event we're instrumenting
669 // for address sanitization.
670 if ((Offset.isNegative() && (-Offset).uge(PHISize)) ||
671 (!Offset.isNegative() && Offset.uge(AllocSize))) {
672 S.DeadOperands.push_back(U);
676 insertUse(PN, Offset, PHISize);
679 void visitSelectInst(SelectInst &SI) {
681 return markAsDead(SI);
682 if (Value *Result = foldSelectInst(SI)) {
684 // If the result of the constant fold will be the pointer, recurse
685 // through the select as if we had RAUW'ed it.
688 // Otherwise the operand to the select is dead, and we can replace it
690 S.DeadOperands.push_back(U);
695 return PI.setAborted(&SI);
697 // See if we already have computed info on this node.
698 uint64_t &SelectSize = PHIOrSelectSizes[&SI];
700 // This is a new Select, check for an unsafe use of it.
701 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&SI, SelectSize))
702 return PI.setAborted(UnsafeI);
705 // For PHI and select operands outside the alloca, we can't nuke the entire
706 // phi or select -- the other side might still be relevant, so we special
707 // case them here and use a separate structure to track the operands
708 // themselves which should be replaced with undef.
709 // FIXME: This should instead be escaped in the event we're instrumenting
710 // for address sanitization.
711 if ((Offset.isNegative() && Offset.uge(SelectSize)) ||
712 (!Offset.isNegative() && Offset.uge(AllocSize))) {
713 S.DeadOperands.push_back(U);
717 insertUse(SI, Offset, SelectSize);
720 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
721 void visitInstruction(Instruction &I) {
726 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
728 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
731 PointerEscapingInstr(0) {
732 SliceBuilder PB(DL, AI, *this);
733 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
734 if (PtrI.isEscaped() || PtrI.isAborted()) {
735 // FIXME: We should sink the escape vs. abort info into the caller nicely,
736 // possibly by just storing the PtrInfo in the AllocaSlices.
737 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
738 : PtrI.getAbortingInst();
739 assert(PointerEscapingInstr && "Did not track a bad instruction");
743 Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
744 std::mem_fun_ref(&Slice::isDead)),
747 #if __cplusplus >= 201103L && !defined(NDEBUG)
748 if (SROARandomShuffleSlices) {
749 std::mt19937 MT(static_cast<unsigned>(sys::TimeValue::now().msec()));
750 std::shuffle(Slices.begin(), Slices.end(), MT);
754 // Sort the uses. This arranges for the offsets to be in ascending order,
755 // and the sizes to be in descending order.
756 std::sort(Slices.begin(), Slices.end());
759 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
761 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
762 StringRef Indent) const {
763 printSlice(OS, I, Indent);
764 printUse(OS, I, Indent);
767 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
768 StringRef Indent) const {
769 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
770 << " slice #" << (I - begin())
771 << (I->isSplittable() ? " (splittable)" : "") << "\n";
774 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
775 StringRef Indent) const {
776 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
779 void AllocaSlices::print(raw_ostream &OS) const {
780 if (PointerEscapingInstr) {
781 OS << "Can't analyze slices for alloca: " << AI << "\n"
782 << " A pointer to this alloca escaped by:\n"
783 << " " << *PointerEscapingInstr << "\n";
787 OS << "Slices of alloca: " << AI << "\n";
788 for (const_iterator I = begin(), E = end(); I != E; ++I)
792 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
795 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
797 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
800 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
802 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
803 /// the loads and stores of an alloca instruction, as well as updating its
804 /// debug information. This is used when a domtree is unavailable and thus
805 /// mem2reg in its full form can't be used to handle promotion of allocas to
807 class AllocaPromoter : public LoadAndStorePromoter {
811 SmallVector<DbgDeclareInst *, 4> DDIs;
812 SmallVector<DbgValueInst *, 4> DVIs;
815 AllocaPromoter(const SmallVectorImpl<Instruction *> &Insts, SSAUpdater &S,
816 AllocaInst &AI, DIBuilder &DIB)
817 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
819 void run(const SmallVectorImpl<Instruction*> &Insts) {
820 // Retain the debug information attached to the alloca for use when
821 // rewriting loads and stores.
822 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
823 for (Value::use_iterator UI = DebugNode->use_begin(),
824 UE = DebugNode->use_end();
826 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
828 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
832 LoadAndStorePromoter::run(Insts);
834 // While we have the debug information, clear it off of the alloca. The
835 // caller takes care of deleting the alloca.
836 while (!DDIs.empty())
837 DDIs.pop_back_val()->eraseFromParent();
838 while (!DVIs.empty())
839 DVIs.pop_back_val()->eraseFromParent();
842 virtual bool isInstInList(Instruction *I,
843 const SmallVectorImpl<Instruction*> &Insts) const {
845 if (LoadInst *LI = dyn_cast<LoadInst>(I))
846 Ptr = LI->getOperand(0);
848 Ptr = cast<StoreInst>(I)->getPointerOperand();
850 // Only used to detect cycles, which will be rare and quickly found as
851 // we're walking up a chain of defs rather than down through uses.
852 SmallPtrSet<Value *, 4> Visited;
858 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr))
859 Ptr = BCI->getOperand(0);
860 else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr))
861 Ptr = GEPI->getPointerOperand();
865 } while (Visited.insert(Ptr));
870 virtual void updateDebugInfo(Instruction *Inst) const {
871 for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(),
872 E = DDIs.end(); I != E; ++I) {
873 DbgDeclareInst *DDI = *I;
874 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
875 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
876 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
877 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
879 for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(),
880 E = DVIs.end(); I != E; ++I) {
881 DbgValueInst *DVI = *I;
883 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
884 // If an argument is zero extended then use argument directly. The ZExt
885 // may be zapped by an optimization pass in future.
886 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
887 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
888 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
889 Arg = dyn_cast<Argument>(SExt->getOperand(0));
891 Arg = SI->getValueOperand();
892 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
893 Arg = LI->getPointerOperand();
897 Instruction *DbgVal =
898 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
900 DbgVal->setDebugLoc(DVI->getDebugLoc());
904 } // end anon namespace
908 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
910 /// This pass takes allocations which can be completely analyzed (that is, they
911 /// don't escape) and tries to turn them into scalar SSA values. There are
912 /// a few steps to this process.
914 /// 1) It takes allocations of aggregates and analyzes the ways in which they
915 /// are used to try to split them into smaller allocations, ideally of
916 /// a single scalar data type. It will split up memcpy and memset accesses
917 /// as necessary and try to isolate individual scalar accesses.
918 /// 2) It will transform accesses into forms which are suitable for SSA value
919 /// promotion. This can be replacing a memset with a scalar store of an
920 /// integer value, or it can involve speculating operations on a PHI or
921 /// select to be a PHI or select of the results.
922 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
923 /// onto insert and extract operations on a vector value, and convert them to
924 /// this form. By doing so, it will enable promotion of vector aggregates to
925 /// SSA vector values.
926 class SROA : public FunctionPass {
927 const bool RequiresDomTree;
930 const DataLayout *DL;
933 /// \brief Worklist of alloca instructions to simplify.
935 /// Each alloca in the function is added to this. Each new alloca formed gets
936 /// added to it as well to recursively simplify unless that alloca can be
937 /// directly promoted. Finally, each time we rewrite a use of an alloca other
938 /// the one being actively rewritten, we add it back onto the list if not
939 /// already present to ensure it is re-visited.
940 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
942 /// \brief A collection of instructions to delete.
943 /// We try to batch deletions to simplify code and make things a bit more
945 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
947 /// \brief Post-promotion worklist.
949 /// Sometimes we discover an alloca which has a high probability of becoming
950 /// viable for SROA after a round of promotion takes place. In those cases,
951 /// the alloca is enqueued here for re-processing.
953 /// Note that we have to be very careful to clear allocas out of this list in
954 /// the event they are deleted.
955 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
957 /// \brief A collection of alloca instructions we can directly promote.
958 std::vector<AllocaInst *> PromotableAllocas;
960 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
962 /// All of these PHIs have been checked for the safety of speculation and by
963 /// being speculated will allow promoting allocas currently in the promotable
965 SetVector<PHINode *, SmallVector<PHINode *, 2> > SpeculatablePHIs;
967 /// \brief A worklist of select instructions to speculate prior to promoting
970 /// All of these select instructions have been checked for the safety of
971 /// speculation and by being speculated will allow promoting allocas
972 /// currently in the promotable queue.
973 SetVector<SelectInst *, SmallVector<SelectInst *, 2> > SpeculatableSelects;
976 SROA(bool RequiresDomTree = true)
977 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
979 initializeSROAPass(*PassRegistry::getPassRegistry());
981 bool runOnFunction(Function &F);
982 void getAnalysisUsage(AnalysisUsage &AU) const;
984 const char *getPassName() const { return "SROA"; }
988 friend class PHIOrSelectSpeculator;
989 friend class AllocaSliceRewriter;
991 bool rewritePartition(AllocaInst &AI, AllocaSlices &S,
992 AllocaSlices::iterator B, AllocaSlices::iterator E,
993 int64_t BeginOffset, int64_t EndOffset,
994 ArrayRef<AllocaSlices::iterator> SplitUses);
995 bool splitAlloca(AllocaInst &AI, AllocaSlices &S);
996 bool runOnAlloca(AllocaInst &AI);
997 void clobberUse(Use &U);
998 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
999 bool promoteAllocas(Function &F);
1005 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1006 return new SROA(RequiresDomTree);
1009 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1011 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1012 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1015 /// Walk the range of a partitioning looking for a common type to cover this
1016 /// sequence of slices.
1017 static Type *findCommonType(AllocaSlices::const_iterator B,
1018 AllocaSlices::const_iterator E,
1019 uint64_t EndOffset) {
1021 bool TyIsCommon = true;
1022 IntegerType *ITy = 0;
1024 // Note that we need to look at *every* alloca slice's Use to ensure we
1025 // always get consistent results regardless of the order of slices.
1026 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1027 Use *U = I->getUse();
1028 if (isa<IntrinsicInst>(*U->getUser()))
1030 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1034 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1035 UserTy = LI->getType();
1036 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1037 UserTy = SI->getValueOperand()->getType();
1040 if (!UserTy || (Ty && Ty != UserTy))
1041 TyIsCommon = false; // Give up on anything but an iN type.
1045 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1046 // If the type is larger than the partition, skip it. We only encounter
1047 // this for split integer operations where we want to use the type of the
1048 // entity causing the split. Also skip if the type is not a byte width
1050 if (UserITy->getBitWidth() % 8 != 0 ||
1051 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1054 // Track the largest bitwidth integer type used in this way in case there
1055 // is no common type.
1056 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1061 return TyIsCommon ? Ty : ITy;
1064 /// PHI instructions that use an alloca and are subsequently loaded can be
1065 /// rewritten to load both input pointers in the pred blocks and then PHI the
1066 /// results, allowing the load of the alloca to be promoted.
1068 /// %P2 = phi [i32* %Alloca, i32* %Other]
1069 /// %V = load i32* %P2
1071 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1073 /// %V2 = load i32* %Other
1075 /// %V = phi [i32 %V1, i32 %V2]
1077 /// We can do this to a select if its only uses are loads and if the operands
1078 /// to the select can be loaded unconditionally.
1080 /// FIXME: This should be hoisted into a generic utility, likely in
1081 /// Transforms/Util/Local.h
1082 static bool isSafePHIToSpeculate(PHINode &PN,
1083 const DataLayout *DL = 0) {
1084 // For now, we can only do this promotion if the load is in the same block
1085 // as the PHI, and if there are no stores between the phi and load.
1086 // TODO: Allow recursive phi users.
1087 // TODO: Allow stores.
1088 BasicBlock *BB = PN.getParent();
1089 unsigned MaxAlign = 0;
1090 bool HaveLoad = false;
1091 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end(); UI != UE;
1093 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1094 if (LI == 0 || !LI->isSimple())
1097 // For now we only allow loads in the same block as the PHI. This is
1098 // a common case that happens when instcombine merges two loads through
1100 if (LI->getParent() != BB)
1103 // Ensure that there are no instructions between the PHI and the load that
1105 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1106 if (BBI->mayWriteToMemory())
1109 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1116 // We can only transform this if it is safe to push the loads into the
1117 // predecessor blocks. The only thing to watch out for is that we can't put
1118 // a possibly trapping load in the predecessor if it is a critical edge.
1119 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1120 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1121 Value *InVal = PN.getIncomingValue(Idx);
1123 // If the value is produced by the terminator of the predecessor (an
1124 // invoke) or it has side-effects, there is no valid place to put a load
1125 // in the predecessor.
1126 if (TI == InVal || TI->mayHaveSideEffects())
1129 // If the predecessor has a single successor, then the edge isn't
1131 if (TI->getNumSuccessors() == 1)
1134 // If this pointer is always safe to load, or if we can prove that there
1135 // is already a load in the block, then we can move the load to the pred
1137 if (InVal->isDereferenceablePointer() ||
1138 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, DL))
1147 static void speculatePHINodeLoads(PHINode &PN) {
1148 DEBUG(dbgs() << " original: " << PN << "\n");
1150 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1151 IRBuilderTy PHIBuilder(&PN);
1152 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1153 PN.getName() + ".sroa.speculated");
1155 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1156 // matter which one we get and if any differ.
1157 LoadInst *SomeLoad = cast<LoadInst>(*PN.use_begin());
1158 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1159 unsigned Align = SomeLoad->getAlignment();
1161 // Rewrite all loads of the PN to use the new PHI.
1162 while (!PN.use_empty()) {
1163 LoadInst *LI = cast<LoadInst>(*PN.use_begin());
1164 LI->replaceAllUsesWith(NewPN);
1165 LI->eraseFromParent();
1168 // Inject loads into all of the pred blocks.
1169 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1170 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1171 TerminatorInst *TI = Pred->getTerminator();
1172 Value *InVal = PN.getIncomingValue(Idx);
1173 IRBuilderTy PredBuilder(TI);
1175 LoadInst *Load = PredBuilder.CreateLoad(
1176 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1177 ++NumLoadsSpeculated;
1178 Load->setAlignment(Align);
1180 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1181 NewPN->addIncoming(Load, Pred);
1184 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1185 PN.eraseFromParent();
1188 /// Select instructions that use an alloca and are subsequently loaded can be
1189 /// rewritten to load both input pointers and then select between the result,
1190 /// allowing the load of the alloca to be promoted.
1192 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1193 /// %V = load i32* %P2
1195 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1196 /// %V2 = load i32* %Other
1197 /// %V = select i1 %cond, i32 %V1, i32 %V2
1199 /// We can do this to a select if its only uses are loads and if the operand
1200 /// to the select can be loaded unconditionally.
1201 static bool isSafeSelectToSpeculate(SelectInst &SI, const DataLayout *DL = 0) {
1202 Value *TValue = SI.getTrueValue();
1203 Value *FValue = SI.getFalseValue();
1204 bool TDerefable = TValue->isDereferenceablePointer();
1205 bool FDerefable = FValue->isDereferenceablePointer();
1207 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end(); UI != UE;
1209 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1210 if (LI == 0 || !LI->isSimple())
1213 // Both operands to the select need to be dereferencable, either
1214 // absolutely (e.g. allocas) or at this point because we can see other
1217 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment(), DL))
1220 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment(), DL))
1227 static void speculateSelectInstLoads(SelectInst &SI) {
1228 DEBUG(dbgs() << " original: " << SI << "\n");
1230 IRBuilderTy IRB(&SI);
1231 Value *TV = SI.getTrueValue();
1232 Value *FV = SI.getFalseValue();
1233 // Replace the loads of the select with a select of two loads.
1234 while (!SI.use_empty()) {
1235 LoadInst *LI = cast<LoadInst>(*SI.use_begin());
1236 assert(LI->isSimple() && "We only speculate simple loads");
1238 IRB.SetInsertPoint(LI);
1240 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1242 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1243 NumLoadsSpeculated += 2;
1245 // Transfer alignment and TBAA info if present.
1246 TL->setAlignment(LI->getAlignment());
1247 FL->setAlignment(LI->getAlignment());
1248 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1249 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1250 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1253 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1254 LI->getName() + ".sroa.speculated");
1256 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1257 LI->replaceAllUsesWith(V);
1258 LI->eraseFromParent();
1260 SI.eraseFromParent();
1263 /// \brief Build a GEP out of a base pointer and indices.
1265 /// This will return the BasePtr if that is valid, or build a new GEP
1266 /// instruction using the IRBuilder if GEP-ing is needed.
1267 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1268 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1269 if (Indices.empty())
1272 // A single zero index is a no-op, so check for this and avoid building a GEP
1274 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1277 return IRB.CreateInBoundsGEP(BasePtr, Indices, NamePrefix + "sroa_idx");
1280 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1281 /// TargetTy without changing the offset of the pointer.
1283 /// This routine assumes we've already established a properly offset GEP with
1284 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1285 /// zero-indices down through type layers until we find one the same as
1286 /// TargetTy. If we can't find one with the same type, we at least try to use
1287 /// one with the same size. If none of that works, we just produce the GEP as
1288 /// indicated by Indices to have the correct offset.
1289 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1290 Value *BasePtr, Type *Ty, Type *TargetTy,
1291 SmallVectorImpl<Value *> &Indices,
1294 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1296 // See if we can descend into a struct and locate a field with the correct
1298 unsigned NumLayers = 0;
1299 Type *ElementTy = Ty;
1301 if (ElementTy->isPointerTy())
1303 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1304 ElementTy = SeqTy->getElementType();
1305 // Note that we use the default address space as this index is over an
1306 // array or a vector, not a pointer.
1307 Indices.push_back(IRB.getInt(APInt(DL.getPointerSizeInBits(0), 0)));
1308 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1309 if (STy->element_begin() == STy->element_end())
1310 break; // Nothing left to descend into.
1311 ElementTy = *STy->element_begin();
1312 Indices.push_back(IRB.getInt32(0));
1317 } while (ElementTy != TargetTy);
1318 if (ElementTy != TargetTy)
1319 Indices.erase(Indices.end() - NumLayers, Indices.end());
1321 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1324 /// \brief Recursively compute indices for a natural GEP.
1326 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1327 /// element types adding appropriate indices for the GEP.
1328 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1329 Value *Ptr, Type *Ty, APInt &Offset,
1331 SmallVectorImpl<Value *> &Indices,
1334 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices, NamePrefix);
1336 // We can't recurse through pointer types.
1337 if (Ty->isPointerTy())
1340 // We try to analyze GEPs over vectors here, but note that these GEPs are
1341 // extremely poorly defined currently. The long-term goal is to remove GEPing
1342 // over a vector from the IR completely.
1343 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1344 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1345 if (ElementSizeInBits % 8)
1346 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1347 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1348 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1349 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1351 Offset -= NumSkippedElements * ElementSize;
1352 Indices.push_back(IRB.getInt(NumSkippedElements));
1353 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1354 Offset, TargetTy, Indices, NamePrefix);
1357 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1358 Type *ElementTy = ArrTy->getElementType();
1359 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1360 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1361 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1364 Offset -= NumSkippedElements * ElementSize;
1365 Indices.push_back(IRB.getInt(NumSkippedElements));
1366 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1367 Indices, NamePrefix);
1370 StructType *STy = dyn_cast<StructType>(Ty);
1374 const StructLayout *SL = DL.getStructLayout(STy);
1375 uint64_t StructOffset = Offset.getZExtValue();
1376 if (StructOffset >= SL->getSizeInBytes())
1378 unsigned Index = SL->getElementContainingOffset(StructOffset);
1379 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1380 Type *ElementTy = STy->getElementType(Index);
1381 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1382 return 0; // The offset points into alignment padding.
1384 Indices.push_back(IRB.getInt32(Index));
1385 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1386 Indices, NamePrefix);
1389 /// \brief Get a natural GEP from a base pointer to a particular offset and
1390 /// resulting in a particular type.
1392 /// The goal is to produce a "natural" looking GEP that works with the existing
1393 /// composite types to arrive at the appropriate offset and element type for
1394 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1395 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1396 /// Indices, and setting Ty to the result subtype.
1398 /// If no natural GEP can be constructed, this function returns null.
1399 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1400 Value *Ptr, APInt Offset, Type *TargetTy,
1401 SmallVectorImpl<Value *> &Indices,
1403 PointerType *Ty = cast<PointerType>(Ptr->getType());
1405 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1407 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1410 Type *ElementTy = Ty->getElementType();
1411 if (!ElementTy->isSized())
1412 return 0; // We can't GEP through an unsized element.
1413 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1414 if (ElementSize == 0)
1415 return 0; // Zero-length arrays can't help us build a natural GEP.
1416 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1418 Offset -= NumSkippedElements * ElementSize;
1419 Indices.push_back(IRB.getInt(NumSkippedElements));
1420 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1421 Indices, NamePrefix);
1424 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1425 /// resulting pointer has PointerTy.
1427 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1428 /// and produces the pointer type desired. Where it cannot, it will try to use
1429 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1430 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1431 /// bitcast to the type.
1433 /// The strategy for finding the more natural GEPs is to peel off layers of the
1434 /// pointer, walking back through bit casts and GEPs, searching for a base
1435 /// pointer from which we can compute a natural GEP with the desired
1436 /// properties. The algorithm tries to fold as many constant indices into
1437 /// a single GEP as possible, thus making each GEP more independent of the
1438 /// surrounding code.
1439 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1440 APInt Offset, Type *PointerTy,
1442 // Even though we don't look through PHI nodes, we could be called on an
1443 // instruction in an unreachable block, which may be on a cycle.
1444 SmallPtrSet<Value *, 4> Visited;
1445 Visited.insert(Ptr);
1446 SmallVector<Value *, 4> Indices;
1448 // We may end up computing an offset pointer that has the wrong type. If we
1449 // never are able to compute one directly that has the correct type, we'll
1450 // fall back to it, so keep it around here.
1451 Value *OffsetPtr = 0;
1453 // Remember any i8 pointer we come across to re-use if we need to do a raw
1456 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1458 Type *TargetTy = PointerTy->getPointerElementType();
1461 // First fold any existing GEPs into the offset.
1462 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1463 APInt GEPOffset(Offset.getBitWidth(), 0);
1464 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1466 Offset += GEPOffset;
1467 Ptr = GEP->getPointerOperand();
1468 if (!Visited.insert(Ptr))
1472 // See if we can perform a natural GEP here.
1474 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1475 Indices, NamePrefix)) {
1476 if (P->getType() == PointerTy) {
1477 // Zap any offset pointer that we ended up computing in previous rounds.
1478 if (OffsetPtr && OffsetPtr->use_empty())
1479 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1480 I->eraseFromParent();
1488 // Stash this pointer if we've found an i8*.
1489 if (Ptr->getType()->isIntegerTy(8)) {
1491 Int8PtrOffset = Offset;
1494 // Peel off a layer of the pointer and update the offset appropriately.
1495 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1496 Ptr = cast<Operator>(Ptr)->getOperand(0);
1497 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1498 if (GA->mayBeOverridden())
1500 Ptr = GA->getAliasee();
1504 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1505 } while (Visited.insert(Ptr));
1509 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1510 NamePrefix + "sroa_raw_cast");
1511 Int8PtrOffset = Offset;
1514 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1515 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1516 NamePrefix + "sroa_raw_idx");
1520 // On the off chance we were targeting i8*, guard the bitcast here.
1521 if (Ptr->getType() != PointerTy)
1522 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1527 /// \brief Test whether we can convert a value from the old to the new type.
1529 /// This predicate should be used to guard calls to convertValue in order to
1530 /// ensure that we only try to convert viable values. The strategy is that we
1531 /// will peel off single element struct and array wrappings to get to an
1532 /// underlying value, and convert that value.
1533 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1536 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1537 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1538 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1540 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1542 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1545 // We can convert pointers to integers and vice-versa. Same for vectors
1546 // of pointers and integers.
1547 OldTy = OldTy->getScalarType();
1548 NewTy = NewTy->getScalarType();
1549 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1550 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1552 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1560 /// \brief Generic routine to convert an SSA value to a value of a different
1563 /// This will try various different casting techniques, such as bitcasts,
1564 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1565 /// two types for viability with this routine.
1566 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1568 Type *OldTy = V->getType();
1569 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1574 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1575 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1576 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1577 return IRB.CreateZExt(V, NewITy);
1579 // See if we need inttoptr for this type pair. A cast involving both scalars
1580 // and vectors requires and additional bitcast.
1581 if (OldTy->getScalarType()->isIntegerTy() &&
1582 NewTy->getScalarType()->isPointerTy()) {
1583 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1584 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1585 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1588 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1589 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1590 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1593 return IRB.CreateIntToPtr(V, NewTy);
1596 // See if we need ptrtoint for this type pair. A cast involving both scalars
1597 // and vectors requires and additional bitcast.
1598 if (OldTy->getScalarType()->isPointerTy() &&
1599 NewTy->getScalarType()->isIntegerTy()) {
1600 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1601 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1602 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1605 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1606 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1607 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1610 return IRB.CreatePtrToInt(V, NewTy);
1613 return IRB.CreateBitCast(V, NewTy);
1616 /// \brief Test whether the given slice use can be promoted to a vector.
1618 /// This function is called to test each entry in a partioning which is slated
1619 /// for a single slice.
1620 static bool isVectorPromotionViableForSlice(
1621 const DataLayout &DL, AllocaSlices &S, uint64_t SliceBeginOffset,
1622 uint64_t SliceEndOffset, VectorType *Ty, uint64_t ElementSize,
1623 AllocaSlices::const_iterator I) {
1624 // First validate the slice offsets.
1625 uint64_t BeginOffset =
1626 std::max(I->beginOffset(), SliceBeginOffset) - SliceBeginOffset;
1627 uint64_t BeginIndex = BeginOffset / ElementSize;
1628 if (BeginIndex * ElementSize != BeginOffset ||
1629 BeginIndex >= Ty->getNumElements())
1631 uint64_t EndOffset =
1632 std::min(I->endOffset(), SliceEndOffset) - SliceBeginOffset;
1633 uint64_t EndIndex = EndOffset / ElementSize;
1634 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1637 assert(EndIndex > BeginIndex && "Empty vector!");
1638 uint64_t NumElements = EndIndex - BeginIndex;
1640 (NumElements == 1) ? Ty->getElementType()
1641 : VectorType::get(Ty->getElementType(), NumElements);
1644 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1646 Use *U = I->getUse();
1648 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1649 if (MI->isVolatile())
1651 if (!I->isSplittable())
1652 return false; // Skip any unsplittable intrinsics.
1653 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1654 // Disable vector promotion when there are loads or stores of an FCA.
1656 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1657 if (LI->isVolatile())
1659 Type *LTy = LI->getType();
1660 if (SliceBeginOffset > I->beginOffset() ||
1661 SliceEndOffset < I->endOffset()) {
1662 assert(LTy->isIntegerTy());
1665 if (!canConvertValue(DL, SliceTy, LTy))
1667 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1668 if (SI->isVolatile())
1670 Type *STy = SI->getValueOperand()->getType();
1671 if (SliceBeginOffset > I->beginOffset() ||
1672 SliceEndOffset < I->endOffset()) {
1673 assert(STy->isIntegerTy());
1676 if (!canConvertValue(DL, STy, SliceTy))
1685 /// \brief Test whether the given alloca partitioning and range of slices can be
1686 /// promoted to a vector.
1688 /// This is a quick test to check whether we can rewrite a particular alloca
1689 /// partition (and its newly formed alloca) into a vector alloca with only
1690 /// whole-vector loads and stores such that it could be promoted to a vector
1691 /// SSA value. We only can ensure this for a limited set of operations, and we
1692 /// don't want to do the rewrites unless we are confident that the result will
1693 /// be promotable, so we have an early test here.
1695 isVectorPromotionViable(const DataLayout &DL, Type *AllocaTy, AllocaSlices &S,
1696 uint64_t SliceBeginOffset, uint64_t SliceEndOffset,
1697 AllocaSlices::const_iterator I,
1698 AllocaSlices::const_iterator E,
1699 ArrayRef<AllocaSlices::iterator> SplitUses) {
1700 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1704 uint64_t ElementSize = DL.getTypeSizeInBits(Ty->getScalarType());
1706 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1707 // that aren't byte sized.
1708 if (ElementSize % 8)
1710 assert((DL.getTypeSizeInBits(Ty) % 8) == 0 &&
1711 "vector size not a multiple of element size?");
1715 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1716 SliceEndOffset, Ty, ElementSize, I))
1719 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1720 SUE = SplitUses.end();
1722 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1723 SliceEndOffset, Ty, ElementSize, *SUI))
1729 /// \brief Test whether a slice of an alloca is valid for integer widening.
1731 /// This implements the necessary checking for the \c isIntegerWideningViable
1732 /// test below on a single slice of the alloca.
1733 static bool isIntegerWideningViableForSlice(const DataLayout &DL,
1735 uint64_t AllocBeginOffset,
1736 uint64_t Size, AllocaSlices &S,
1737 AllocaSlices::const_iterator I,
1738 bool &WholeAllocaOp) {
1739 uint64_t RelBegin = I->beginOffset() - AllocBeginOffset;
1740 uint64_t RelEnd = I->endOffset() - AllocBeginOffset;
1742 // We can't reasonably handle cases where the load or store extends past
1743 // the end of the aloca's type and into its padding.
1747 Use *U = I->getUse();
1749 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1750 if (LI->isVolatile())
1752 if (RelBegin == 0 && RelEnd == Size)
1753 WholeAllocaOp = true;
1754 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1755 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1757 } else if (RelBegin != 0 || RelEnd != Size ||
1758 !canConvertValue(DL, AllocaTy, LI->getType())) {
1759 // Non-integer loads need to be convertible from the alloca type so that
1760 // they are promotable.
1763 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1764 Type *ValueTy = SI->getValueOperand()->getType();
1765 if (SI->isVolatile())
1767 if (RelBegin == 0 && RelEnd == Size)
1768 WholeAllocaOp = true;
1769 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1770 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1772 } else if (RelBegin != 0 || RelEnd != Size ||
1773 !canConvertValue(DL, ValueTy, AllocaTy)) {
1774 // Non-integer stores need to be convertible to the alloca type so that
1775 // they are promotable.
1778 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1779 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1781 if (!I->isSplittable())
1782 return false; // Skip any unsplittable intrinsics.
1783 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1784 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1785 II->getIntrinsicID() != Intrinsic::lifetime_end)
1794 /// \brief Test whether the given alloca partition's integer operations can be
1795 /// widened to promotable ones.
1797 /// This is a quick test to check whether we can rewrite the integer loads and
1798 /// stores to a particular alloca into wider loads and stores and be able to
1799 /// promote the resulting alloca.
1801 isIntegerWideningViable(const DataLayout &DL, Type *AllocaTy,
1802 uint64_t AllocBeginOffset, AllocaSlices &S,
1803 AllocaSlices::const_iterator I,
1804 AllocaSlices::const_iterator E,
1805 ArrayRef<AllocaSlices::iterator> SplitUses) {
1806 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1807 // Don't create integer types larger than the maximum bitwidth.
1808 if (SizeInBits > IntegerType::MAX_INT_BITS)
1811 // Don't try to handle allocas with bit-padding.
1812 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1815 // We need to ensure that an integer type with the appropriate bitwidth can
1816 // be converted to the alloca type, whatever that is. We don't want to force
1817 // the alloca itself to have an integer type if there is a more suitable one.
1818 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1819 if (!canConvertValue(DL, AllocaTy, IntTy) ||
1820 !canConvertValue(DL, IntTy, AllocaTy))
1823 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1825 // While examining uses, we ensure that the alloca has a covering load or
1826 // store. We don't want to widen the integer operations only to fail to
1827 // promote due to some other unsplittable entry (which we may make splittable
1828 // later). However, if there are only splittable uses, go ahead and assume
1829 // that we cover the alloca.
1830 bool WholeAllocaOp = (I != E) ? false : DL.isLegalInteger(SizeInBits);
1833 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1834 S, I, WholeAllocaOp))
1837 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1838 SUE = SplitUses.end();
1840 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1841 S, *SUI, WholeAllocaOp))
1844 return WholeAllocaOp;
1847 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1848 IntegerType *Ty, uint64_t Offset,
1849 const Twine &Name) {
1850 DEBUG(dbgs() << " start: " << *V << "\n");
1851 IntegerType *IntTy = cast<IntegerType>(V->getType());
1852 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1853 "Element extends past full value");
1854 uint64_t ShAmt = 8*Offset;
1855 if (DL.isBigEndian())
1856 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1858 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
1859 DEBUG(dbgs() << " shifted: " << *V << "\n");
1861 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1862 "Cannot extract to a larger integer!");
1864 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
1865 DEBUG(dbgs() << " trunced: " << *V << "\n");
1870 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
1871 Value *V, uint64_t Offset, const Twine &Name) {
1872 IntegerType *IntTy = cast<IntegerType>(Old->getType());
1873 IntegerType *Ty = cast<IntegerType>(V->getType());
1874 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1875 "Cannot insert a larger integer!");
1876 DEBUG(dbgs() << " start: " << *V << "\n");
1878 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
1879 DEBUG(dbgs() << " extended: " << *V << "\n");
1881 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1882 "Element store outside of alloca store");
1883 uint64_t ShAmt = 8*Offset;
1884 if (DL.isBigEndian())
1885 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1887 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
1888 DEBUG(dbgs() << " shifted: " << *V << "\n");
1891 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
1892 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
1893 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
1894 DEBUG(dbgs() << " masked: " << *Old << "\n");
1895 V = IRB.CreateOr(Old, V, Name + ".insert");
1896 DEBUG(dbgs() << " inserted: " << *V << "\n");
1901 static Value *extractVector(IRBuilderTy &IRB, Value *V,
1902 unsigned BeginIndex, unsigned EndIndex,
1903 const Twine &Name) {
1904 VectorType *VecTy = cast<VectorType>(V->getType());
1905 unsigned NumElements = EndIndex - BeginIndex;
1906 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
1908 if (NumElements == VecTy->getNumElements())
1911 if (NumElements == 1) {
1912 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
1914 DEBUG(dbgs() << " extract: " << *V << "\n");
1918 SmallVector<Constant*, 8> Mask;
1919 Mask.reserve(NumElements);
1920 for (unsigned i = BeginIndex; i != EndIndex; ++i)
1921 Mask.push_back(IRB.getInt32(i));
1922 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1923 ConstantVector::get(Mask),
1925 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1929 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
1930 unsigned BeginIndex, const Twine &Name) {
1931 VectorType *VecTy = cast<VectorType>(Old->getType());
1932 assert(VecTy && "Can only insert a vector into a vector");
1934 VectorType *Ty = dyn_cast<VectorType>(V->getType());
1936 // Single element to insert.
1937 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
1939 DEBUG(dbgs() << " insert: " << *V << "\n");
1943 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
1944 "Too many elements!");
1945 if (Ty->getNumElements() == VecTy->getNumElements()) {
1946 assert(V->getType() == VecTy && "Vector type mismatch");
1949 unsigned EndIndex = BeginIndex + Ty->getNumElements();
1951 // When inserting a smaller vector into the larger to store, we first
1952 // use a shuffle vector to widen it with undef elements, and then
1953 // a second shuffle vector to select between the loaded vector and the
1955 SmallVector<Constant*, 8> Mask;
1956 Mask.reserve(VecTy->getNumElements());
1957 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1958 if (i >= BeginIndex && i < EndIndex)
1959 Mask.push_back(IRB.getInt32(i - BeginIndex));
1961 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
1962 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1963 ConstantVector::get(Mask),
1965 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1968 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1969 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
1971 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
1973 DEBUG(dbgs() << " blend: " << *V << "\n");
1978 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
1979 /// to use a new alloca.
1981 /// Also implements the rewriting to vector-based accesses when the partition
1982 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
1984 class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
1985 // Befriend the base class so it can delegate to private visit methods.
1986 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
1987 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
1989 const DataLayout &DL;
1992 AllocaInst &OldAI, &NewAI;
1993 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
1996 // If we are rewriting an alloca partition which can be written as pure
1997 // vector operations, we stash extra information here. When VecTy is
1998 // non-null, we have some strict guarantees about the rewritten alloca:
1999 // - The new alloca is exactly the size of the vector type here.
2000 // - The accesses all either map to the entire vector or to a single
2002 // - The set of accessing instructions is only one of those handled above
2003 // in isVectorPromotionViable. Generally these are the same access kinds
2004 // which are promotable via mem2reg.
2007 uint64_t ElementSize;
2009 // This is a convenience and flag variable that will be null unless the new
2010 // alloca's integer operations should be widened to this integer type due to
2011 // passing isIntegerWideningViable above. If it is non-null, the desired
2012 // integer type will be stored here for easy access during rewriting.
2015 // The offset of the slice currently being rewritten.
2016 uint64_t BeginOffset, EndOffset;
2020 Instruction *OldPtr;
2022 // Track post-rewrite users which are PHI nodes and Selects.
2023 SmallPtrSetImpl<PHINode *> &PHIUsers;
2024 SmallPtrSetImpl<SelectInst *> &SelectUsers;
2026 // Utility IR builder, whose name prefix is setup for each visited use, and
2027 // the insertion point is set to point to the user.
2031 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &S, SROA &Pass,
2032 AllocaInst &OldAI, AllocaInst &NewAI,
2033 uint64_t NewBeginOffset, uint64_t NewEndOffset,
2034 bool IsVectorPromotable, bool IsIntegerPromotable,
2035 SmallPtrSetImpl<PHINode *> &PHIUsers,
2036 SmallPtrSetImpl<SelectInst *> &SelectUsers)
2037 : DL(DL), S(S), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2038 NewAllocaBeginOffset(NewBeginOffset), NewAllocaEndOffset(NewEndOffset),
2039 NewAllocaTy(NewAI.getAllocatedType()),
2040 VecTy(IsVectorPromotable ? cast<VectorType>(NewAllocaTy) : 0),
2041 ElementTy(VecTy ? VecTy->getElementType() : 0),
2042 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2043 IntTy(IsIntegerPromotable
2046 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2048 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2049 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2050 IRB(NewAI.getContext(), ConstantFolder()) {
2052 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2053 "Only multiple-of-8 sized vector elements are viable");
2056 assert((!IsVectorPromotable && !IsIntegerPromotable) ||
2057 IsVectorPromotable != IsIntegerPromotable);
2060 bool visit(AllocaSlices::const_iterator I) {
2061 bool CanSROA = true;
2062 BeginOffset = I->beginOffset();
2063 EndOffset = I->endOffset();
2064 IsSplittable = I->isSplittable();
2066 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2068 OldUse = I->getUse();
2069 OldPtr = cast<Instruction>(OldUse->get());
2071 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2072 IRB.SetInsertPoint(OldUserI);
2073 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2074 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2076 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2083 // Make sure the other visit overloads are visible.
2086 // Every instruction which can end up as a user must have a rewrite rule.
2087 bool visitInstruction(Instruction &I) {
2088 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2089 llvm_unreachable("No rewrite rule for this instruction!");
2092 Value *getAdjustedAllocaPtr(IRBuilderTy &IRB, uint64_t Offset,
2094 assert(Offset >= NewAllocaBeginOffset);
2096 StringRef OldName = OldPtr->getName();
2097 // Skip through the last '.sroa.' component of the name.
2098 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2099 if (LastSROAPrefix != StringRef::npos) {
2100 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2101 // Look for an SROA slice index.
2102 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2103 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2104 // Strip the index and look for the offset.
2105 OldName = OldName.substr(IndexEnd + 1);
2106 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2107 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2108 // Strip the offset.
2109 OldName = OldName.substr(OffsetEnd + 1);
2112 // Strip any SROA suffixes as well.
2113 OldName = OldName.substr(0, OldName.find(".sroa_"));
2115 return getAdjustedPtr(IRB, DL, &NewAI, APInt(DL.getPointerSizeInBits(),
2116 Offset - NewAllocaBeginOffset),
2119 Twine(OldName) + "."
2126 /// \brief Compute suitable alignment to access an offset into the new alloca.
2127 unsigned getOffsetAlign(uint64_t Offset) {
2128 unsigned NewAIAlign = NewAI.getAlignment();
2130 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2131 return MinAlign(NewAIAlign, Offset);
2134 /// \brief Compute suitable alignment to access a type at an offset of the
2137 /// \returns zero if the type's ABI alignment is a suitable alignment,
2138 /// otherwise returns the maximal suitable alignment.
2139 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
2140 unsigned Align = getOffsetAlign(Offset);
2141 return Align == DL.getABITypeAlignment(Ty) ? 0 : Align;
2144 unsigned getIndex(uint64_t Offset) {
2145 assert(VecTy && "Can only call getIndex when rewriting a vector");
2146 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2147 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2148 uint32_t Index = RelOffset / ElementSize;
2149 assert(Index * ElementSize == RelOffset);
2153 void deleteIfTriviallyDead(Value *V) {
2154 Instruction *I = cast<Instruction>(V);
2155 if (isInstructionTriviallyDead(I))
2156 Pass.DeadInsts.insert(I);
2159 Value *rewriteVectorizedLoadInst(uint64_t NewBeginOffset,
2160 uint64_t NewEndOffset) {
2161 unsigned BeginIndex = getIndex(NewBeginOffset);
2162 unsigned EndIndex = getIndex(NewEndOffset);
2163 assert(EndIndex > BeginIndex && "Empty vector!");
2165 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2167 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2170 Value *rewriteIntegerLoad(LoadInst &LI, uint64_t NewBeginOffset,
2171 uint64_t NewEndOffset) {
2172 assert(IntTy && "We cannot insert an integer to the alloca");
2173 assert(!LI.isVolatile());
2174 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2176 V = convertValue(DL, IRB, V, IntTy);
2177 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2178 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2179 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2180 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2185 bool visitLoadInst(LoadInst &LI) {
2186 DEBUG(dbgs() << " original: " << LI << "\n");
2187 Value *OldOp = LI.getOperand(0);
2188 assert(OldOp == OldPtr);
2190 // Compute the intersecting offset range.
2191 assert(BeginOffset < NewAllocaEndOffset);
2192 assert(EndOffset > NewAllocaBeginOffset);
2193 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2194 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2196 uint64_t Size = NewEndOffset - NewBeginOffset;
2198 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), Size * 8)
2200 bool IsPtrAdjusted = false;
2203 V = rewriteVectorizedLoadInst(NewBeginOffset, NewEndOffset);
2204 } else if (IntTy && LI.getType()->isIntegerTy()) {
2205 V = rewriteIntegerLoad(LI, NewBeginOffset, NewEndOffset);
2206 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2207 canConvertValue(DL, NewAllocaTy, LI.getType())) {
2208 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2209 LI.isVolatile(), LI.getName());
2211 Type *LTy = TargetTy->getPointerTo();
2212 V = IRB.CreateAlignedLoad(
2213 getAdjustedAllocaPtr(IRB, NewBeginOffset, LTy),
2214 getOffsetTypeAlign(TargetTy, NewBeginOffset - NewAllocaBeginOffset),
2215 LI.isVolatile(), LI.getName());
2216 IsPtrAdjusted = true;
2218 V = convertValue(DL, IRB, V, TargetTy);
2221 assert(!LI.isVolatile());
2222 assert(LI.getType()->isIntegerTy() &&
2223 "Only integer type loads and stores are split");
2224 assert(Size < DL.getTypeStoreSize(LI.getType()) &&
2225 "Split load isn't smaller than original load");
2226 assert(LI.getType()->getIntegerBitWidth() ==
2227 DL.getTypeStoreSizeInBits(LI.getType()) &&
2228 "Non-byte-multiple bit width");
2229 // Move the insertion point just past the load so that we can refer to it.
2230 IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI)));
2231 // Create a placeholder value with the same type as LI to use as the
2232 // basis for the new value. This allows us to replace the uses of LI with
2233 // the computed value, and then replace the placeholder with LI, leaving
2234 // LI only used for this computation.
2236 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2237 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset,
2239 LI.replaceAllUsesWith(V);
2240 Placeholder->replaceAllUsesWith(&LI);
2243 LI.replaceAllUsesWith(V);
2246 Pass.DeadInsts.insert(&LI);
2247 deleteIfTriviallyDead(OldOp);
2248 DEBUG(dbgs() << " to: " << *V << "\n");
2249 return !LI.isVolatile() && !IsPtrAdjusted;
2252 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp,
2253 uint64_t NewBeginOffset,
2254 uint64_t NewEndOffset) {
2255 if (V->getType() != VecTy) {
2256 unsigned BeginIndex = getIndex(NewBeginOffset);
2257 unsigned EndIndex = getIndex(NewEndOffset);
2258 assert(EndIndex > BeginIndex && "Empty vector!");
2259 unsigned NumElements = EndIndex - BeginIndex;
2260 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2262 (NumElements == 1) ? ElementTy
2263 : VectorType::get(ElementTy, NumElements);
2264 if (V->getType() != SliceTy)
2265 V = convertValue(DL, IRB, V, SliceTy);
2267 // Mix in the existing elements.
2268 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2270 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2272 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2273 Pass.DeadInsts.insert(&SI);
2276 DEBUG(dbgs() << " to: " << *Store << "\n");
2280 bool rewriteIntegerStore(Value *V, StoreInst &SI,
2281 uint64_t NewBeginOffset, uint64_t NewEndOffset) {
2282 assert(IntTy && "We cannot extract an integer from the alloca");
2283 assert(!SI.isVolatile());
2284 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2285 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2287 Old = convertValue(DL, IRB, Old, IntTy);
2288 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2289 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2290 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset,
2293 V = convertValue(DL, IRB, V, NewAllocaTy);
2294 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2295 Pass.DeadInsts.insert(&SI);
2297 DEBUG(dbgs() << " to: " << *Store << "\n");
2301 bool visitStoreInst(StoreInst &SI) {
2302 DEBUG(dbgs() << " original: " << SI << "\n");
2303 Value *OldOp = SI.getOperand(1);
2304 assert(OldOp == OldPtr);
2306 Value *V = SI.getValueOperand();
2308 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2309 // alloca that should be re-examined after promoting this alloca.
2310 if (V->getType()->isPointerTy())
2311 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2312 Pass.PostPromotionWorklist.insert(AI);
2314 // Compute the intersecting offset range.
2315 assert(BeginOffset < NewAllocaEndOffset);
2316 assert(EndOffset > NewAllocaBeginOffset);
2317 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2318 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2320 uint64_t Size = NewEndOffset - NewBeginOffset;
2321 if (Size < DL.getTypeStoreSize(V->getType())) {
2322 assert(!SI.isVolatile());
2323 assert(V->getType()->isIntegerTy() &&
2324 "Only integer type loads and stores are split");
2325 assert(V->getType()->getIntegerBitWidth() ==
2326 DL.getTypeStoreSizeInBits(V->getType()) &&
2327 "Non-byte-multiple bit width");
2328 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8);
2329 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset,
2334 return rewriteVectorizedStoreInst(V, SI, OldOp, NewBeginOffset,
2336 if (IntTy && V->getType()->isIntegerTy())
2337 return rewriteIntegerStore(V, SI, NewBeginOffset, NewEndOffset);
2340 if (NewBeginOffset == NewAllocaBeginOffset &&
2341 NewEndOffset == NewAllocaEndOffset &&
2342 canConvertValue(DL, V->getType(), NewAllocaTy)) {
2343 V = convertValue(DL, IRB, V, NewAllocaTy);
2344 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2347 Value *NewPtr = getAdjustedAllocaPtr(IRB, NewBeginOffset,
2348 V->getType()->getPointerTo());
2349 NewSI = IRB.CreateAlignedStore(
2350 V, NewPtr, getOffsetTypeAlign(V->getType(),
2351 NewBeginOffset - NewAllocaBeginOffset),
2355 Pass.DeadInsts.insert(&SI);
2356 deleteIfTriviallyDead(OldOp);
2358 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2359 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2362 /// \brief Compute an integer value from splatting an i8 across the given
2363 /// number of bytes.
2365 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2366 /// call this routine.
2367 /// FIXME: Heed the advice above.
2369 /// \param V The i8 value to splat.
2370 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2371 Value *getIntegerSplat(Value *V, unsigned Size) {
2372 assert(Size > 0 && "Expected a positive number of bytes.");
2373 IntegerType *VTy = cast<IntegerType>(V->getType());
2374 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2378 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2379 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, "zext"),
2380 ConstantExpr::getUDiv(
2381 Constant::getAllOnesValue(SplatIntTy),
2382 ConstantExpr::getZExt(
2383 Constant::getAllOnesValue(V->getType()),
2389 /// \brief Compute a vector splat for a given element value.
2390 Value *getVectorSplat(Value *V, unsigned NumElements) {
2391 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2392 DEBUG(dbgs() << " splat: " << *V << "\n");
2396 bool visitMemSetInst(MemSetInst &II) {
2397 DEBUG(dbgs() << " original: " << II << "\n");
2398 assert(II.getRawDest() == OldPtr);
2400 // If the memset has a variable size, it cannot be split, just adjust the
2401 // pointer to the new alloca.
2402 if (!isa<Constant>(II.getLength())) {
2404 assert(BeginOffset >= NewAllocaBeginOffset);
2405 II.setDest(getAdjustedAllocaPtr(IRB, BeginOffset, OldPtr->getType()));
2406 Type *CstTy = II.getAlignmentCst()->getType();
2407 II.setAlignment(ConstantInt::get(CstTy, getOffsetAlign(BeginOffset)));
2409 deleteIfTriviallyDead(OldPtr);
2413 // Record this instruction for deletion.
2414 Pass.DeadInsts.insert(&II);
2416 Type *AllocaTy = NewAI.getAllocatedType();
2417 Type *ScalarTy = AllocaTy->getScalarType();
2419 // Compute the intersecting offset range.
2420 assert(BeginOffset < NewAllocaEndOffset);
2421 assert(EndOffset > NewAllocaBeginOffset);
2422 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2423 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2424 uint64_t SliceOffset = NewBeginOffset - NewAllocaBeginOffset;
2426 // If this doesn't map cleanly onto the alloca type, and that type isn't
2427 // a single value type, just emit a memset.
2428 if (!VecTy && !IntTy &&
2429 (BeginOffset > NewAllocaBeginOffset ||
2430 EndOffset < NewAllocaEndOffset ||
2431 !AllocaTy->isSingleValueType() ||
2432 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2433 DL.getTypeSizeInBits(ScalarTy)%8 != 0)) {
2434 Type *SizeTy = II.getLength()->getType();
2435 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2436 CallInst *New = IRB.CreateMemSet(
2437 getAdjustedAllocaPtr(IRB, NewBeginOffset, OldPtr->getType()),
2438 II.getValue(), Size, getOffsetAlign(SliceOffset), II.isVolatile());
2440 DEBUG(dbgs() << " to: " << *New << "\n");
2444 // If we can represent this as a simple value, we have to build the actual
2445 // value to store, which requires expanding the byte present in memset to
2446 // a sensible representation for the alloca type. This is essentially
2447 // splatting the byte to a sufficiently wide integer, splatting it across
2448 // any desired vector width, and bitcasting to the final type.
2452 // If this is a memset of a vectorized alloca, insert it.
2453 assert(ElementTy == ScalarTy);
2455 unsigned BeginIndex = getIndex(NewBeginOffset);
2456 unsigned EndIndex = getIndex(NewEndOffset);
2457 assert(EndIndex > BeginIndex && "Empty vector!");
2458 unsigned NumElements = EndIndex - BeginIndex;
2459 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2462 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2463 Splat = convertValue(DL, IRB, Splat, ElementTy);
2464 if (NumElements > 1)
2465 Splat = getVectorSplat(Splat, NumElements);
2467 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2469 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2471 // If this is a memset on an alloca where we can widen stores, insert the
2473 assert(!II.isVolatile());
2475 uint64_t Size = NewEndOffset - NewBeginOffset;
2476 V = getIntegerSplat(II.getValue(), Size);
2478 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2479 EndOffset != NewAllocaBeginOffset)) {
2480 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2482 Old = convertValue(DL, IRB, Old, IntTy);
2483 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2484 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2486 assert(V->getType() == IntTy &&
2487 "Wrong type for an alloca wide integer!");
2489 V = convertValue(DL, IRB, V, AllocaTy);
2491 // Established these invariants above.
2492 assert(NewBeginOffset == NewAllocaBeginOffset);
2493 assert(NewEndOffset == NewAllocaEndOffset);
2495 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2496 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2497 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2499 V = convertValue(DL, IRB, V, AllocaTy);
2502 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2505 DEBUG(dbgs() << " to: " << *New << "\n");
2506 return !II.isVolatile();
2509 bool visitMemTransferInst(MemTransferInst &II) {
2510 // Rewriting of memory transfer instructions can be a bit tricky. We break
2511 // them into two categories: split intrinsics and unsplit intrinsics.
2513 DEBUG(dbgs() << " original: " << II << "\n");
2515 // Compute the intersecting offset range.
2516 assert(BeginOffset < NewAllocaEndOffset);
2517 assert(EndOffset > NewAllocaBeginOffset);
2518 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2519 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2521 bool IsDest = &II.getRawDestUse() == OldUse;
2522 assert((IsDest && II.getRawDest() == OldPtr) ||
2523 (!IsDest && II.getRawSource() == OldPtr));
2525 // Compute the relative offset within the transfer.
2526 unsigned IntPtrWidth = DL.getPointerSizeInBits();
2527 APInt RelOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2529 unsigned Align = II.getAlignment();
2530 uint64_t SliceOffset = NewBeginOffset - NewAllocaBeginOffset;
2533 MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2534 MinAlign(II.getAlignment(), getOffsetAlign(SliceOffset)));
2536 // For unsplit intrinsics, we simply modify the source and destination
2537 // pointers in place. This isn't just an optimization, it is a matter of
2538 // correctness. With unsplit intrinsics we may be dealing with transfers
2539 // within a single alloca before SROA ran, or with transfers that have
2540 // a variable length. We may also be dealing with memmove instead of
2541 // memcpy, and so simply updating the pointers is the necessary for us to
2542 // update both source and dest of a single call.
2543 if (!IsSplittable) {
2544 Value *AdjustedPtr =
2545 getAdjustedAllocaPtr(IRB, BeginOffset, OldPtr->getType());
2547 II.setDest(AdjustedPtr);
2549 II.setSource(AdjustedPtr);
2551 Type *CstTy = II.getAlignmentCst()->getType();
2552 II.setAlignment(ConstantInt::get(CstTy, Align));
2554 DEBUG(dbgs() << " to: " << II << "\n");
2555 deleteIfTriviallyDead(OldPtr);
2558 // For split transfer intrinsics we have an incredibly useful assurance:
2559 // the source and destination do not reside within the same alloca, and at
2560 // least one of them does not escape. This means that we can replace
2561 // memmove with memcpy, and we don't need to worry about all manner of
2562 // downsides to splitting and transforming the operations.
2564 // If this doesn't map cleanly onto the alloca type, and that type isn't
2565 // a single value type, just emit a memcpy.
2567 = !VecTy && !IntTy && (BeginOffset > NewAllocaBeginOffset ||
2568 EndOffset < NewAllocaEndOffset ||
2569 !NewAI.getAllocatedType()->isSingleValueType());
2571 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2572 // size hasn't been shrunk based on analysis of the viable range, this is
2574 if (EmitMemCpy && &OldAI == &NewAI) {
2575 // Ensure the start lines up.
2576 assert(NewBeginOffset == BeginOffset);
2578 // Rewrite the size as needed.
2579 if (NewEndOffset != EndOffset)
2580 II.setLength(ConstantInt::get(II.getLength()->getType(),
2581 NewEndOffset - NewBeginOffset));
2584 // Record this instruction for deletion.
2585 Pass.DeadInsts.insert(&II);
2587 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2588 // alloca that should be re-examined after rewriting this instruction.
2589 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2591 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2592 assert(AI != &OldAI && AI != &NewAI &&
2593 "Splittable transfers cannot reach the same alloca on both ends.");
2594 Pass.Worklist.insert(AI);
2598 Type *OtherPtrTy = OtherPtr->getType();
2600 // Compute the other pointer, folding as much as possible to produce
2601 // a single, simple GEP in most cases.
2602 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy,
2603 OtherPtr->getName() + ".");
2606 getAdjustedAllocaPtr(IRB, NewBeginOffset, OldPtr->getType());
2607 Type *SizeTy = II.getLength()->getType();
2608 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2610 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2611 IsDest ? OtherPtr : OurPtr,
2612 Size, Align, II.isVolatile());
2614 DEBUG(dbgs() << " to: " << *New << "\n");
2618 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2619 // is equivalent to 1, but that isn't true if we end up rewriting this as
2624 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2625 NewEndOffset == NewAllocaEndOffset;
2626 uint64_t Size = NewEndOffset - NewBeginOffset;
2627 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2628 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2629 unsigned NumElements = EndIndex - BeginIndex;
2630 IntegerType *SubIntTy
2631 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0;
2633 Type *OtherPtrTy = NewAI.getType();
2634 if (VecTy && !IsWholeAlloca) {
2635 if (NumElements == 1)
2636 OtherPtrTy = VecTy->getElementType();
2638 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2640 OtherPtrTy = OtherPtrTy->getPointerTo();
2641 } else if (IntTy && !IsWholeAlloca) {
2642 OtherPtrTy = SubIntTy->getPointerTo();
2645 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy,
2646 OtherPtr->getName() + ".");
2647 Value *DstPtr = &NewAI;
2649 std::swap(SrcPtr, DstPtr);
2652 if (VecTy && !IsWholeAlloca && !IsDest) {
2653 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2655 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2656 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2657 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2659 Src = convertValue(DL, IRB, Src, IntTy);
2660 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2661 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2663 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2667 if (VecTy && !IsWholeAlloca && IsDest) {
2668 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2670 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2671 } else if (IntTy && !IsWholeAlloca && IsDest) {
2672 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2674 Old = convertValue(DL, IRB, Old, IntTy);
2675 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2676 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2677 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2680 StoreInst *Store = cast<StoreInst>(
2681 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2683 DEBUG(dbgs() << " to: " << *Store << "\n");
2684 return !II.isVolatile();
2687 bool visitIntrinsicInst(IntrinsicInst &II) {
2688 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2689 II.getIntrinsicID() == Intrinsic::lifetime_end);
2690 DEBUG(dbgs() << " original: " << II << "\n");
2691 assert(II.getArgOperand(1) == OldPtr);
2693 // Compute the intersecting offset range.
2694 assert(BeginOffset < NewAllocaEndOffset);
2695 assert(EndOffset > NewAllocaBeginOffset);
2696 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2697 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2699 // Record this instruction for deletion.
2700 Pass.DeadInsts.insert(&II);
2703 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2704 NewEndOffset - NewBeginOffset);
2705 Value *Ptr = getAdjustedAllocaPtr(IRB, NewBeginOffset, OldPtr->getType());
2707 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2708 New = IRB.CreateLifetimeStart(Ptr, Size);
2710 New = IRB.CreateLifetimeEnd(Ptr, Size);
2713 DEBUG(dbgs() << " to: " << *New << "\n");
2717 bool visitPHINode(PHINode &PN) {
2718 DEBUG(dbgs() << " original: " << PN << "\n");
2719 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2720 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2722 // We would like to compute a new pointer in only one place, but have it be
2723 // as local as possible to the PHI. To do that, we re-use the location of
2724 // the old pointer, which necessarily must be in the right position to
2725 // dominate the PHI.
2726 IRBuilderTy PtrBuilder(IRB);
2727 PtrBuilder.SetInsertPoint(OldPtr);
2728 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
2731 getAdjustedAllocaPtr(PtrBuilder, BeginOffset, OldPtr->getType());
2732 // Replace the operands which were using the old pointer.
2733 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2735 DEBUG(dbgs() << " to: " << PN << "\n");
2736 deleteIfTriviallyDead(OldPtr);
2738 // PHIs can't be promoted on their own, but often can be speculated. We
2739 // check the speculation outside of the rewriter so that we see the
2740 // fully-rewritten alloca.
2741 PHIUsers.insert(&PN);
2745 bool visitSelectInst(SelectInst &SI) {
2746 DEBUG(dbgs() << " original: " << SI << "\n");
2747 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2748 "Pointer isn't an operand!");
2749 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2750 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2752 Value *NewPtr = getAdjustedAllocaPtr(IRB, BeginOffset, OldPtr->getType());
2753 // Replace the operands which were using the old pointer.
2754 if (SI.getOperand(1) == OldPtr)
2755 SI.setOperand(1, NewPtr);
2756 if (SI.getOperand(2) == OldPtr)
2757 SI.setOperand(2, NewPtr);
2759 DEBUG(dbgs() << " to: " << SI << "\n");
2760 deleteIfTriviallyDead(OldPtr);
2762 // Selects can't be promoted on their own, but often can be speculated. We
2763 // check the speculation outside of the rewriter so that we see the
2764 // fully-rewritten alloca.
2765 SelectUsers.insert(&SI);
2773 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2775 /// This pass aggressively rewrites all aggregate loads and stores on
2776 /// a particular pointer (or any pointer derived from it which we can identify)
2777 /// with scalar loads and stores.
2778 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2779 // Befriend the base class so it can delegate to private visit methods.
2780 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2782 const DataLayout &DL;
2784 /// Queue of pointer uses to analyze and potentially rewrite.
2785 SmallVector<Use *, 8> Queue;
2787 /// Set to prevent us from cycling with phi nodes and loops.
2788 SmallPtrSet<User *, 8> Visited;
2790 /// The current pointer use being rewritten. This is used to dig up the used
2791 /// value (as opposed to the user).
2795 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
2797 /// Rewrite loads and stores through a pointer and all pointers derived from
2799 bool rewrite(Instruction &I) {
2800 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2802 bool Changed = false;
2803 while (!Queue.empty()) {
2804 U = Queue.pop_back_val();
2805 Changed |= visit(cast<Instruction>(U->getUser()));
2811 /// Enqueue all the users of the given instruction for further processing.
2812 /// This uses a set to de-duplicate users.
2813 void enqueueUsers(Instruction &I) {
2814 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
2816 if (Visited.insert(*UI))
2817 Queue.push_back(&UI.getUse());
2820 // Conservative default is to not rewrite anything.
2821 bool visitInstruction(Instruction &I) { return false; }
2823 /// \brief Generic recursive split emission class.
2824 template <typename Derived>
2827 /// The builder used to form new instructions.
2829 /// The indices which to be used with insert- or extractvalue to select the
2830 /// appropriate value within the aggregate.
2831 SmallVector<unsigned, 4> Indices;
2832 /// The indices to a GEP instruction which will move Ptr to the correct slot
2833 /// within the aggregate.
2834 SmallVector<Value *, 4> GEPIndices;
2835 /// The base pointer of the original op, used as a base for GEPing the
2836 /// split operations.
2839 /// Initialize the splitter with an insertion point, Ptr and start with a
2840 /// single zero GEP index.
2841 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2842 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2845 /// \brief Generic recursive split emission routine.
2847 /// This method recursively splits an aggregate op (load or store) into
2848 /// scalar or vector ops. It splits recursively until it hits a single value
2849 /// and emits that single value operation via the template argument.
2851 /// The logic of this routine relies on GEPs and insertvalue and
2852 /// extractvalue all operating with the same fundamental index list, merely
2853 /// formatted differently (GEPs need actual values).
2855 /// \param Ty The type being split recursively into smaller ops.
2856 /// \param Agg The aggregate value being built up or stored, depending on
2857 /// whether this is splitting a load or a store respectively.
2858 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2859 if (Ty->isSingleValueType())
2860 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2862 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2863 unsigned OldSize = Indices.size();
2865 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2867 assert(Indices.size() == OldSize && "Did not return to the old size");
2868 Indices.push_back(Idx);
2869 GEPIndices.push_back(IRB.getInt32(Idx));
2870 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2871 GEPIndices.pop_back();
2877 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2878 unsigned OldSize = Indices.size();
2880 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2882 assert(Indices.size() == OldSize && "Did not return to the old size");
2883 Indices.push_back(Idx);
2884 GEPIndices.push_back(IRB.getInt32(Idx));
2885 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2886 GEPIndices.pop_back();
2892 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2896 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2897 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2898 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2900 /// Emit a leaf load of a single value. This is called at the leaves of the
2901 /// recursive emission to actually load values.
2902 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2903 assert(Ty->isSingleValueType());
2904 // Load the single value and insert it using the indices.
2905 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
2906 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
2907 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2908 DEBUG(dbgs() << " to: " << *Load << "\n");
2912 bool visitLoadInst(LoadInst &LI) {
2913 assert(LI.getPointerOperand() == *U);
2914 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2917 // We have an aggregate being loaded, split it apart.
2918 DEBUG(dbgs() << " original: " << LI << "\n");
2919 LoadOpSplitter Splitter(&LI, *U);
2920 Value *V = UndefValue::get(LI.getType());
2921 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2922 LI.replaceAllUsesWith(V);
2923 LI.eraseFromParent();
2927 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2928 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2929 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2931 /// Emit a leaf store of a single value. This is called at the leaves of the
2932 /// recursive emission to actually produce stores.
2933 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2934 assert(Ty->isSingleValueType());
2935 // Extract the single value and store it using the indices.
2936 Value *Store = IRB.CreateStore(
2937 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2938 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2940 DEBUG(dbgs() << " to: " << *Store << "\n");
2944 bool visitStoreInst(StoreInst &SI) {
2945 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2947 Value *V = SI.getValueOperand();
2948 if (V->getType()->isSingleValueType())
2951 // We have an aggregate being stored, split it apart.
2952 DEBUG(dbgs() << " original: " << SI << "\n");
2953 StoreOpSplitter Splitter(&SI, *U);
2954 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2955 SI.eraseFromParent();
2959 bool visitBitCastInst(BitCastInst &BC) {
2964 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2969 bool visitPHINode(PHINode &PN) {
2974 bool visitSelectInst(SelectInst &SI) {
2981 /// \brief Strip aggregate type wrapping.
2983 /// This removes no-op aggregate types wrapping an underlying type. It will
2984 /// strip as many layers of types as it can without changing either the type
2985 /// size or the allocated size.
2986 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
2987 if (Ty->isSingleValueType())
2990 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
2991 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
2994 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
2995 InnerTy = ArrTy->getElementType();
2996 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
2997 const StructLayout *SL = DL.getStructLayout(STy);
2998 unsigned Index = SL->getElementContainingOffset(0);
2999 InnerTy = STy->getElementType(Index);
3004 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3005 TypeSize > DL.getTypeSizeInBits(InnerTy))
3008 return stripAggregateTypeWrapping(DL, InnerTy);
3011 /// \brief Try to find a partition of the aggregate type passed in for a given
3012 /// offset and size.
3014 /// This recurses through the aggregate type and tries to compute a subtype
3015 /// based on the offset and size. When the offset and size span a sub-section
3016 /// of an array, it will even compute a new array type for that sub-section,
3017 /// and the same for structs.
3019 /// Note that this routine is very strict and tries to find a partition of the
3020 /// type which produces the *exact* right offset and size. It is not forgiving
3021 /// when the size or offset cause either end of type-based partition to be off.
3022 /// Also, this is a best-effort routine. It is reasonable to give up and not
3023 /// return a type if necessary.
3024 static Type *getTypePartition(const DataLayout &DL, Type *Ty,
3025 uint64_t Offset, uint64_t Size) {
3026 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3027 return stripAggregateTypeWrapping(DL, Ty);
3028 if (Offset > DL.getTypeAllocSize(Ty) ||
3029 (DL.getTypeAllocSize(Ty) - Offset) < Size)
3032 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3033 // We can't partition pointers...
3034 if (SeqTy->isPointerTy())
3037 Type *ElementTy = SeqTy->getElementType();
3038 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3039 uint64_t NumSkippedElements = Offset / ElementSize;
3040 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
3041 if (NumSkippedElements >= ArrTy->getNumElements())
3043 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
3044 if (NumSkippedElements >= VecTy->getNumElements())
3047 Offset -= NumSkippedElements * ElementSize;
3049 // First check if we need to recurse.
3050 if (Offset > 0 || Size < ElementSize) {
3051 // Bail if the partition ends in a different array element.
3052 if ((Offset + Size) > ElementSize)
3054 // Recurse through the element type trying to peel off offset bytes.
3055 return getTypePartition(DL, ElementTy, Offset, Size);
3057 assert(Offset == 0);
3059 if (Size == ElementSize)
3060 return stripAggregateTypeWrapping(DL, ElementTy);
3061 assert(Size > ElementSize);
3062 uint64_t NumElements = Size / ElementSize;
3063 if (NumElements * ElementSize != Size)
3065 return ArrayType::get(ElementTy, NumElements);
3068 StructType *STy = dyn_cast<StructType>(Ty);
3072 const StructLayout *SL = DL.getStructLayout(STy);
3073 if (Offset >= SL->getSizeInBytes())
3075 uint64_t EndOffset = Offset + Size;
3076 if (EndOffset > SL->getSizeInBytes())
3079 unsigned Index = SL->getElementContainingOffset(Offset);
3080 Offset -= SL->getElementOffset(Index);
3082 Type *ElementTy = STy->getElementType(Index);
3083 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3084 if (Offset >= ElementSize)
3085 return 0; // The offset points into alignment padding.
3087 // See if any partition must be contained by the element.
3088 if (Offset > 0 || Size < ElementSize) {
3089 if ((Offset + Size) > ElementSize)
3091 return getTypePartition(DL, ElementTy, Offset, Size);
3093 assert(Offset == 0);
3095 if (Size == ElementSize)
3096 return stripAggregateTypeWrapping(DL, ElementTy);
3098 StructType::element_iterator EI = STy->element_begin() + Index,
3099 EE = STy->element_end();
3100 if (EndOffset < SL->getSizeInBytes()) {
3101 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3102 if (Index == EndIndex)
3103 return 0; // Within a single element and its padding.
3105 // Don't try to form "natural" types if the elements don't line up with the
3107 // FIXME: We could potentially recurse down through the last element in the
3108 // sub-struct to find a natural end point.
3109 if (SL->getElementOffset(EndIndex) != EndOffset)
3112 assert(Index < EndIndex);
3113 EE = STy->element_begin() + EndIndex;
3116 // Try to build up a sub-structure.
3117 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
3119 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3120 if (Size != SubSL->getSizeInBytes())
3121 return 0; // The sub-struct doesn't have quite the size needed.
3126 /// \brief Rewrite an alloca partition's users.
3128 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3129 /// to rewrite uses of an alloca partition to be conducive for SSA value
3130 /// promotion. If the partition needs a new, more refined alloca, this will
3131 /// build that new alloca, preserving as much type information as possible, and
3132 /// rewrite the uses of the old alloca to point at the new one and have the
3133 /// appropriate new offsets. It also evaluates how successful the rewrite was
3134 /// at enabling promotion and if it was successful queues the alloca to be
3136 bool SROA::rewritePartition(AllocaInst &AI, AllocaSlices &S,
3137 AllocaSlices::iterator B, AllocaSlices::iterator E,
3138 int64_t BeginOffset, int64_t EndOffset,
3139 ArrayRef<AllocaSlices::iterator> SplitUses) {
3140 assert(BeginOffset < EndOffset);
3141 uint64_t SliceSize = EndOffset - BeginOffset;
3143 // Try to compute a friendly type for this partition of the alloca. This
3144 // won't always succeed, in which case we fall back to a legal integer type
3145 // or an i8 array of an appropriate size.
3147 if (Type *CommonUseTy = findCommonType(B, E, EndOffset))
3148 if (DL->getTypeAllocSize(CommonUseTy) >= SliceSize)
3149 SliceTy = CommonUseTy;
3151 if (Type *TypePartitionTy = getTypePartition(*DL, AI.getAllocatedType(),
3152 BeginOffset, SliceSize))
3153 SliceTy = TypePartitionTy;
3154 if ((!SliceTy || (SliceTy->isArrayTy() &&
3155 SliceTy->getArrayElementType()->isIntegerTy())) &&
3156 DL->isLegalInteger(SliceSize * 8))
3157 SliceTy = Type::getIntNTy(*C, SliceSize * 8);
3159 SliceTy = ArrayType::get(Type::getInt8Ty(*C), SliceSize);
3160 assert(DL->getTypeAllocSize(SliceTy) >= SliceSize);
3162 bool IsVectorPromotable = isVectorPromotionViable(
3163 *DL, SliceTy, S, BeginOffset, EndOffset, B, E, SplitUses);
3165 bool IsIntegerPromotable =
3166 !IsVectorPromotable &&
3167 isIntegerWideningViable(*DL, SliceTy, BeginOffset, S, B, E, SplitUses);
3169 // Check for the case where we're going to rewrite to a new alloca of the
3170 // exact same type as the original, and with the same access offsets. In that
3171 // case, re-use the existing alloca, but still run through the rewriter to
3172 // perform phi and select speculation.
3174 if (SliceTy == AI.getAllocatedType()) {
3175 assert(BeginOffset == 0 &&
3176 "Non-zero begin offset but same alloca type");
3178 // FIXME: We should be able to bail at this point with "nothing changed".
3179 // FIXME: We might want to defer PHI speculation until after here.
3181 unsigned Alignment = AI.getAlignment();
3183 // The minimum alignment which users can rely on when the explicit
3184 // alignment is omitted or zero is that required by the ABI for this
3186 Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
3188 Alignment = MinAlign(Alignment, BeginOffset);
3189 // If we will get at least this much alignment from the type alone, leave
3190 // the alloca's alignment unconstrained.
3191 if (Alignment <= DL->getABITypeAlignment(SliceTy))
3193 NewAI = new AllocaInst(SliceTy, 0, Alignment,
3194 AI.getName() + ".sroa." + Twine(B - S.begin()), &AI);
3198 DEBUG(dbgs() << "Rewriting alloca partition "
3199 << "[" << BeginOffset << "," << EndOffset << ") to: " << *NewAI
3202 // Track the high watermark on the worklist as it is only relevant for
3203 // promoted allocas. We will reset it to this point if the alloca is not in
3204 // fact scheduled for promotion.
3205 unsigned PPWOldSize = PostPromotionWorklist.size();
3206 unsigned NumUses = 0;
3207 SmallPtrSet<PHINode *, 8> PHIUsers;
3208 SmallPtrSet<SelectInst *, 8> SelectUsers;
3210 AllocaSliceRewriter Rewriter(*DL, S, *this, AI, *NewAI, BeginOffset,
3211 EndOffset, IsVectorPromotable,
3212 IsIntegerPromotable, PHIUsers, SelectUsers);
3213 bool Promotable = true;
3214 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
3215 SUE = SplitUses.end();
3216 SUI != SUE; ++SUI) {
3217 DEBUG(dbgs() << " rewriting split ");
3218 DEBUG(S.printSlice(dbgs(), *SUI, ""));
3219 Promotable &= Rewriter.visit(*SUI);
3222 for (AllocaSlices::iterator I = B; I != E; ++I) {
3223 DEBUG(dbgs() << " rewriting ");
3224 DEBUG(S.printSlice(dbgs(), I, ""));
3225 Promotable &= Rewriter.visit(I);
3229 NumAllocaPartitionUses += NumUses;
3230 MaxUsesPerAllocaPartition =
3231 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3233 // Now that we've processed all the slices in the new partition, check if any
3234 // PHIs or Selects would block promotion.
3235 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3238 if (!isSafePHIToSpeculate(**I, DL)) {
3241 SelectUsers.clear();
3244 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3245 E = SelectUsers.end();
3247 if (!isSafeSelectToSpeculate(**I, DL)) {
3250 SelectUsers.clear();
3255 if (PHIUsers.empty() && SelectUsers.empty()) {
3256 // Promote the alloca.
3257 PromotableAllocas.push_back(NewAI);
3259 // If we have either PHIs or Selects to speculate, add them to those
3260 // worklists and re-queue the new alloca so that we promote in on the
3262 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3265 SpeculatablePHIs.insert(*I);
3266 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3267 E = SelectUsers.end();
3269 SpeculatableSelects.insert(*I);
3270 Worklist.insert(NewAI);
3273 // If we can't promote the alloca, iterate on it to check for new
3274 // refinements exposed by splitting the current alloca. Don't iterate on an
3275 // alloca which didn't actually change and didn't get promoted.
3277 Worklist.insert(NewAI);
3279 // Drop any post-promotion work items if promotion didn't happen.
3280 while (PostPromotionWorklist.size() > PPWOldSize)
3281 PostPromotionWorklist.pop_back();
3288 struct IsSliceEndLessOrEqualTo {
3289 uint64_t UpperBound;
3291 IsSliceEndLessOrEqualTo(uint64_t UpperBound) : UpperBound(UpperBound) {}
3293 bool operator()(const AllocaSlices::iterator &I) {
3294 return I->endOffset() <= UpperBound;
3300 removeFinishedSplitUses(SmallVectorImpl<AllocaSlices::iterator> &SplitUses,
3301 uint64_t &MaxSplitUseEndOffset, uint64_t Offset) {
3302 if (Offset >= MaxSplitUseEndOffset) {
3304 MaxSplitUseEndOffset = 0;
3308 size_t SplitUsesOldSize = SplitUses.size();
3309 SplitUses.erase(std::remove_if(SplitUses.begin(), SplitUses.end(),
3310 IsSliceEndLessOrEqualTo(Offset)),
3312 if (SplitUsesOldSize == SplitUses.size())
3315 // Recompute the max. While this is linear, so is remove_if.
3316 MaxSplitUseEndOffset = 0;
3317 for (SmallVectorImpl<AllocaSlices::iterator>::iterator
3318 SUI = SplitUses.begin(),
3319 SUE = SplitUses.end();
3321 MaxSplitUseEndOffset = std::max((*SUI)->endOffset(), MaxSplitUseEndOffset);
3324 /// \brief Walks the slices of an alloca and form partitions based on them,
3325 /// rewriting each of their uses.
3326 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &S) {
3327 if (S.begin() == S.end())
3330 unsigned NumPartitions = 0;
3331 bool Changed = false;
3332 SmallVector<AllocaSlices::iterator, 4> SplitUses;
3333 uint64_t MaxSplitUseEndOffset = 0;
3335 uint64_t BeginOffset = S.begin()->beginOffset();
3337 for (AllocaSlices::iterator SI = S.begin(), SJ = llvm::next(SI), SE = S.end();
3338 SI != SE; SI = SJ) {
3339 uint64_t MaxEndOffset = SI->endOffset();
3341 if (!SI->isSplittable()) {
3342 // When we're forming an unsplittable region, it must always start at the
3343 // first slice and will extend through its end.
3344 assert(BeginOffset == SI->beginOffset());
3346 // Form a partition including all of the overlapping slices with this
3347 // unsplittable slice.
3348 while (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3349 if (!SJ->isSplittable())
3350 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3354 assert(SI->isSplittable()); // Established above.
3356 // Collect all of the overlapping splittable slices.
3357 while (SJ != SE && SJ->beginOffset() < MaxEndOffset &&
3358 SJ->isSplittable()) {
3359 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3363 // Back up MaxEndOffset and SJ if we ended the span early when
3364 // encountering an unsplittable slice.
3365 if (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3366 assert(!SJ->isSplittable());
3367 MaxEndOffset = SJ->beginOffset();
3371 // Check if we have managed to move the end offset forward yet. If so,
3372 // we'll have to rewrite uses and erase old split uses.
3373 if (BeginOffset < MaxEndOffset) {
3374 // Rewrite a sequence of overlapping slices.
3376 rewritePartition(AI, S, SI, SJ, BeginOffset, MaxEndOffset, SplitUses);
3379 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset, MaxEndOffset);
3382 // Accumulate all the splittable slices from the [SI,SJ) region which
3383 // overlap going forward.
3384 for (AllocaSlices::iterator SK = SI; SK != SJ; ++SK)
3385 if (SK->isSplittable() && SK->endOffset() > MaxEndOffset) {
3386 SplitUses.push_back(SK);
3387 MaxSplitUseEndOffset = std::max(SK->endOffset(), MaxSplitUseEndOffset);
3390 // If we're already at the end and we have no split uses, we're done.
3391 if (SJ == SE && SplitUses.empty())
3394 // If we have no split uses or no gap in offsets, we're ready to move to
3396 if (SplitUses.empty() || (SJ != SE && MaxEndOffset == SJ->beginOffset())) {
3397 BeginOffset = SJ->beginOffset();
3401 // Even if we have split slices, if the next slice is splittable and the
3402 // split slices reach it, we can simply set up the beginning offset of the
3403 // next iteration to bridge between them.
3404 if (SJ != SE && SJ->isSplittable() &&
3405 MaxSplitUseEndOffset > SJ->beginOffset()) {
3406 BeginOffset = MaxEndOffset;
3410 // Otherwise, we have a tail of split slices. Rewrite them with an empty
3412 uint64_t PostSplitEndOffset =
3413 SJ == SE ? MaxSplitUseEndOffset : SJ->beginOffset();
3415 Changed |= rewritePartition(AI, S, SJ, SJ, MaxEndOffset, PostSplitEndOffset,
3420 break; // Skip the rest, we don't need to do any cleanup.
3422 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset,
3423 PostSplitEndOffset);
3425 // Now just reset the begin offset for the next iteration.
3426 BeginOffset = SJ->beginOffset();
3429 NumAllocaPartitions += NumPartitions;
3430 MaxPartitionsPerAlloca =
3431 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
3436 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
3437 void SROA::clobberUse(Use &U) {
3439 // Replace the use with an undef value.
3440 U = UndefValue::get(OldV->getType());
3442 // Check for this making an instruction dead. We have to garbage collect
3443 // all the dead instructions to ensure the uses of any alloca end up being
3445 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3446 if (isInstructionTriviallyDead(OldI)) {
3447 DeadInsts.insert(OldI);
3451 /// \brief Analyze an alloca for SROA.
3453 /// This analyzes the alloca to ensure we can reason about it, builds
3454 /// the slices of the alloca, and then hands it off to be split and
3455 /// rewritten as needed.
3456 bool SROA::runOnAlloca(AllocaInst &AI) {
3457 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3458 ++NumAllocasAnalyzed;
3460 // Special case dead allocas, as they're trivial.
3461 if (AI.use_empty()) {
3462 AI.eraseFromParent();
3466 // Skip alloca forms that this analysis can't handle.
3467 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3468 DL->getTypeAllocSize(AI.getAllocatedType()) == 0)
3471 bool Changed = false;
3473 // First, split any FCA loads and stores touching this alloca to promote
3474 // better splitting and promotion opportunities.
3475 AggLoadStoreRewriter AggRewriter(*DL);
3476 Changed |= AggRewriter.rewrite(AI);
3478 // Build the slices using a recursive instruction-visiting builder.
3479 AllocaSlices S(*DL, AI);
3480 DEBUG(S.print(dbgs()));
3484 // Delete all the dead users of this alloca before splitting and rewriting it.
3485 for (AllocaSlices::dead_user_iterator DI = S.dead_user_begin(),
3486 DE = S.dead_user_end();
3488 // Free up everything used by this instruction.
3489 for (User::op_iterator DOI = (*DI)->op_begin(), DOE = (*DI)->op_end();
3493 // Now replace the uses of this instruction.
3494 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3496 // And mark it for deletion.
3497 DeadInsts.insert(*DI);
3500 for (AllocaSlices::dead_op_iterator DO = S.dead_op_begin(),
3501 DE = S.dead_op_end();
3507 // No slices to split. Leave the dead alloca for a later pass to clean up.
3508 if (S.begin() == S.end())
3511 Changed |= splitAlloca(AI, S);
3513 DEBUG(dbgs() << " Speculating PHIs\n");
3514 while (!SpeculatablePHIs.empty())
3515 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
3517 DEBUG(dbgs() << " Speculating Selects\n");
3518 while (!SpeculatableSelects.empty())
3519 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
3524 /// \brief Delete the dead instructions accumulated in this run.
3526 /// Recursively deletes the dead instructions we've accumulated. This is done
3527 /// at the very end to maximize locality of the recursive delete and to
3528 /// minimize the problems of invalidated instruction pointers as such pointers
3529 /// are used heavily in the intermediate stages of the algorithm.
3531 /// We also record the alloca instructions deleted here so that they aren't
3532 /// subsequently handed to mem2reg to promote.
3533 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3534 while (!DeadInsts.empty()) {
3535 Instruction *I = DeadInsts.pop_back_val();
3536 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3538 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3540 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3541 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3542 // Zero out the operand and see if it becomes trivially dead.
3544 if (isInstructionTriviallyDead(U))
3545 DeadInsts.insert(U);
3548 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3549 DeletedAllocas.insert(AI);
3552 I->eraseFromParent();
3556 static void enqueueUsersInWorklist(Instruction &I,
3557 SmallVectorImpl<Instruction *> &Worklist,
3558 SmallPtrSet<Instruction *, 8> &Visited) {
3559 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
3561 if (Visited.insert(cast<Instruction>(*UI)))
3562 Worklist.push_back(cast<Instruction>(*UI));
3565 /// \brief Promote the allocas, using the best available technique.
3567 /// This attempts to promote whatever allocas have been identified as viable in
3568 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3569 /// If there is a domtree available, we attempt to promote using the full power
3570 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3571 /// based on the SSAUpdater utilities. This function returns whether any
3572 /// promotion occurred.
3573 bool SROA::promoteAllocas(Function &F) {
3574 if (PromotableAllocas.empty())
3577 NumPromoted += PromotableAllocas.size();
3579 if (DT && !ForceSSAUpdater) {
3580 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3581 PromoteMemToReg(PromotableAllocas, *DT);
3582 PromotableAllocas.clear();
3586 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3588 DIBuilder DIB(*F.getParent());
3589 SmallVector<Instruction *, 64> Insts;
3591 // We need a worklist to walk the uses of each alloca.
3592 SmallVector<Instruction *, 8> Worklist;
3593 SmallPtrSet<Instruction *, 8> Visited;
3594 SmallVector<Instruction *, 32> DeadInsts;
3596 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3597 AllocaInst *AI = PromotableAllocas[Idx];
3602 enqueueUsersInWorklist(*AI, Worklist, Visited);
3604 while (!Worklist.empty()) {
3605 Instruction *I = Worklist.pop_back_val();
3607 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3608 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3609 // leading to them) here. Eventually it should use them to optimize the
3610 // scalar values produced.
3611 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3612 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3613 II->getIntrinsicID() == Intrinsic::lifetime_end);
3614 II->eraseFromParent();
3618 // Push the loads and stores we find onto the list. SROA will already
3619 // have validated that all loads and stores are viable candidates for
3621 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3622 assert(LI->getType() == AI->getAllocatedType());
3623 Insts.push_back(LI);
3626 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3627 assert(SI->getValueOperand()->getType() == AI->getAllocatedType());
3628 Insts.push_back(SI);
3632 // For everything else, we know that only no-op bitcasts and GEPs will
3633 // make it this far, just recurse through them and recall them for later
3635 DeadInsts.push_back(I);
3636 enqueueUsersInWorklist(*I, Worklist, Visited);
3638 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3639 while (!DeadInsts.empty())
3640 DeadInsts.pop_back_val()->eraseFromParent();
3641 AI->eraseFromParent();
3644 PromotableAllocas.clear();
3649 /// \brief A predicate to test whether an alloca belongs to a set.
3650 class IsAllocaInSet {
3651 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3655 typedef AllocaInst *argument_type;
3657 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3658 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3662 bool SROA::runOnFunction(Function &F) {
3663 if (skipOptnoneFunction(F))
3666 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3667 C = &F.getContext();
3668 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
3670 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3673 DL = &DLP->getDataLayout();
3674 DominatorTreeWrapperPass *DTWP =
3675 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
3676 DT = DTWP ? &DTWP->getDomTree() : 0;
3678 BasicBlock &EntryBB = F.getEntryBlock();
3679 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3681 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3682 Worklist.insert(AI);
3684 bool Changed = false;
3685 // A set of deleted alloca instruction pointers which should be removed from
3686 // the list of promotable allocas.
3687 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3690 while (!Worklist.empty()) {
3691 Changed |= runOnAlloca(*Worklist.pop_back_val());
3692 deleteDeadInstructions(DeletedAllocas);
3694 // Remove the deleted allocas from various lists so that we don't try to
3695 // continue processing them.
3696 if (!DeletedAllocas.empty()) {
3697 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3698 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3699 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3700 PromotableAllocas.end(),
3701 IsAllocaInSet(DeletedAllocas)),
3702 PromotableAllocas.end());
3703 DeletedAllocas.clear();
3707 Changed |= promoteAllocas(F);
3709 Worklist = PostPromotionWorklist;
3710 PostPromotionWorklist.clear();
3711 } while (!Worklist.empty());
3716 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3717 if (RequiresDomTree)
3718 AU.addRequired<DominatorTreeWrapperPass>();
3719 AU.setPreservesCFG();