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);
89 /// \brief A custom IRBuilder inserter which prefixes all names if they are
91 template <bool preserveNames = true>
92 class IRBuilderPrefixedInserter :
93 public IRBuilderDefaultInserter<preserveNames> {
97 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
100 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
101 BasicBlock::iterator InsertPt) const {
102 IRBuilderDefaultInserter<preserveNames>::InsertHelper(
103 I, Name.isTriviallyEmpty() ? Name : Prefix + Name, BB, InsertPt);
107 // Specialization for not preserving the name is trivial.
109 class IRBuilderPrefixedInserter<false> :
110 public IRBuilderDefaultInserter<false> {
112 void SetNamePrefix(const Twine &P) {}
115 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
117 typedef llvm::IRBuilder<true, ConstantFolder,
118 IRBuilderPrefixedInserter<true> > IRBuilderTy;
120 typedef llvm::IRBuilder<false, ConstantFolder,
121 IRBuilderPrefixedInserter<false> > IRBuilderTy;
126 /// \brief A used slice of an alloca.
128 /// This structure represents a slice of an alloca used by some instruction. It
129 /// stores both the begin and end offsets of this use, a pointer to the use
130 /// itself, and a flag indicating whether we can classify the use as splittable
131 /// or not when forming partitions of the alloca.
133 /// \brief The beginning offset of the range.
134 uint64_t BeginOffset;
136 /// \brief The ending offset, not included in the range.
139 /// \brief Storage for both the use of this slice and whether it can be
141 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
144 Slice() : BeginOffset(), EndOffset() {}
145 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
146 : BeginOffset(BeginOffset), EndOffset(EndOffset),
147 UseAndIsSplittable(U, IsSplittable) {}
149 uint64_t beginOffset() const { return BeginOffset; }
150 uint64_t endOffset() const { return EndOffset; }
152 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
153 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
155 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
157 bool isDead() const { return getUse() == 0; }
158 void kill() { UseAndIsSplittable.setPointer(0); }
160 /// \brief Support for ordering ranges.
162 /// This provides an ordering over ranges such that start offsets are
163 /// always increasing, and within equal start offsets, the end offsets are
164 /// decreasing. Thus the spanning range comes first in a cluster with the
165 /// same start position.
166 bool operator<(const Slice &RHS) const {
167 if (beginOffset() < RHS.beginOffset()) return true;
168 if (beginOffset() > RHS.beginOffset()) return false;
169 if (isSplittable() != RHS.isSplittable()) return !isSplittable();
170 if (endOffset() > RHS.endOffset()) return true;
174 /// \brief Support comparison with a single offset to allow binary searches.
175 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
176 uint64_t RHSOffset) {
177 return LHS.beginOffset() < RHSOffset;
179 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
181 return LHSOffset < RHS.beginOffset();
184 bool operator==(const Slice &RHS) const {
185 return isSplittable() == RHS.isSplittable() &&
186 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
188 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
190 } // end anonymous namespace
193 template <typename T> struct isPodLike;
194 template <> struct isPodLike<Slice> {
195 static const bool value = true;
200 /// \brief Representation of the alloca slices.
202 /// This class represents the slices of an alloca which are formed by its
203 /// various uses. If a pointer escapes, we can't fully build a representation
204 /// for the slices used and we reflect that in this structure. The uses are
205 /// stored, sorted by increasing beginning offset and with unsplittable slices
206 /// starting at a particular offset before splittable slices.
209 /// \brief Construct the slices of a particular alloca.
210 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
212 /// \brief Test whether a pointer to the allocation escapes our analysis.
214 /// If this is true, the slices are never fully built and should be
216 bool isEscaped() const { return PointerEscapingInstr; }
218 /// \brief Support for iterating over the slices.
220 typedef SmallVectorImpl<Slice>::iterator iterator;
221 iterator begin() { return Slices.begin(); }
222 iterator end() { return Slices.end(); }
224 typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
225 const_iterator begin() const { return Slices.begin(); }
226 const_iterator end() const { return Slices.end(); }
229 /// \brief Allow iterating the dead users for this alloca.
231 /// These are instructions which will never actually use the alloca as they
232 /// are outside the allocated range. They are safe to replace with undef and
235 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
236 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
237 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
240 /// \brief Allow iterating the dead expressions referring to this alloca.
242 /// These are operands which have cannot actually be used to refer to the
243 /// alloca as they are outside its range and the user doesn't correct for
244 /// that. These mostly consist of PHI node inputs and the like which we just
245 /// need to replace with undef.
247 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
248 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
249 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
252 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
253 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
254 void printSlice(raw_ostream &OS, const_iterator I,
255 StringRef Indent = " ") const;
256 void printUse(raw_ostream &OS, const_iterator I,
257 StringRef Indent = " ") const;
258 void print(raw_ostream &OS) const;
259 void dump(const_iterator I) const;
264 template <typename DerivedT, typename RetT = void> class BuilderBase;
266 friend class AllocaSlices::SliceBuilder;
268 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
269 /// \brief Handle to alloca instruction to simplify method interfaces.
273 /// \brief The instruction responsible for this alloca not having a known set
276 /// When an instruction (potentially) escapes the pointer to the alloca, we
277 /// store a pointer to that here and abort trying to form slices of the
278 /// alloca. This will be null if the alloca slices are analyzed successfully.
279 Instruction *PointerEscapingInstr;
281 /// \brief The slices of the alloca.
283 /// We store a vector of the slices formed by uses of the alloca here. This
284 /// vector is sorted by increasing begin offset, and then the unsplittable
285 /// slices before the splittable ones. See the Slice inner class for more
287 SmallVector<Slice, 8> Slices;
289 /// \brief Instructions which will become dead if we rewrite the alloca.
291 /// Note that these are not separated by slice. This is because we expect an
292 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
293 /// all these instructions can simply be removed and replaced with undef as
294 /// they come from outside of the allocated space.
295 SmallVector<Instruction *, 8> DeadUsers;
297 /// \brief Operands which will become dead if we rewrite the alloca.
299 /// These are operands that in their particular use can be replaced with
300 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
301 /// to PHI nodes and the like. They aren't entirely dead (there might be
302 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
303 /// want to swap this particular input for undef to simplify the use lists of
305 SmallVector<Use *, 8> DeadOperands;
309 static Value *foldSelectInst(SelectInst &SI) {
310 // If the condition being selected on is a constant or the same value is
311 // being selected between, fold the select. Yes this does (rarely) happen
313 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
314 return SI.getOperand(1+CI->isZero());
315 if (SI.getOperand(1) == SI.getOperand(2))
316 return SI.getOperand(1);
321 /// \brief Builder for the alloca slices.
323 /// This class builds a set of alloca slices by recursively visiting the uses
324 /// of an alloca and making a slice for each load and store at each offset.
325 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
326 friend class PtrUseVisitor<SliceBuilder>;
327 friend class InstVisitor<SliceBuilder>;
328 typedef PtrUseVisitor<SliceBuilder> Base;
330 const uint64_t AllocSize;
333 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
334 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
336 /// \brief Set to de-duplicate dead instructions found in the use walk.
337 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
340 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &S)
341 : PtrUseVisitor<SliceBuilder>(DL),
342 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), S(S) {}
345 void markAsDead(Instruction &I) {
346 if (VisitedDeadInsts.insert(&I))
347 S.DeadUsers.push_back(&I);
350 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
351 bool IsSplittable = false) {
352 // Completely skip uses which have a zero size or start either before or
353 // past the end of the allocation.
354 if (Size == 0 || Offset.isNegative() || Offset.uge(AllocSize)) {
355 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
356 << " which has zero size or starts outside of the "
357 << AllocSize << " byte alloca:\n"
358 << " alloca: " << S.AI << "\n"
359 << " use: " << I << "\n");
360 return markAsDead(I);
363 uint64_t BeginOffset = Offset.getZExtValue();
364 uint64_t EndOffset = BeginOffset + Size;
366 // Clamp the end offset to the end of the allocation. Note that this is
367 // formulated to handle even the case where "BeginOffset + Size" overflows.
368 // This may appear superficially to be something we could ignore entirely,
369 // but that is not so! There may be widened loads or PHI-node uses where
370 // some instructions are dead but not others. We can't completely ignore
371 // them, and so have to record at least the information here.
372 assert(AllocSize >= BeginOffset); // Established above.
373 if (Size > AllocSize - BeginOffset) {
374 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
375 << " to remain within the " << AllocSize << " byte alloca:\n"
376 << " alloca: " << S.AI << "\n"
377 << " use: " << I << "\n");
378 EndOffset = AllocSize;
381 S.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
384 void visitBitCastInst(BitCastInst &BC) {
386 return markAsDead(BC);
388 return Base::visitBitCastInst(BC);
391 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
392 if (GEPI.use_empty())
393 return markAsDead(GEPI);
395 return Base::visitGetElementPtrInst(GEPI);
398 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
399 uint64_t Size, bool IsVolatile) {
400 // We allow splitting of loads and stores where the type is an integer type
401 // and cover the entire alloca. This prevents us from splitting over
403 // FIXME: In the great blue eventually, we should eagerly split all integer
404 // loads and stores, and then have a separate step that merges adjacent
405 // alloca partitions into a single partition suitable for integer widening.
406 // Or we should skip the merge step and rely on GVN and other passes to
407 // merge adjacent loads and stores that survive mem2reg.
409 Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize;
411 insertUse(I, Offset, Size, IsSplittable);
414 void visitLoadInst(LoadInst &LI) {
415 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
416 "All simple FCA loads should have been pre-split");
419 return PI.setAborted(&LI);
421 uint64_t Size = DL.getTypeStoreSize(LI.getType());
422 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
425 void visitStoreInst(StoreInst &SI) {
426 Value *ValOp = SI.getValueOperand();
428 return PI.setEscapedAndAborted(&SI);
430 return PI.setAborted(&SI);
432 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
434 // If this memory access can be shown to *statically* extend outside the
435 // bounds of of the allocation, it's behavior is undefined, so simply
436 // ignore it. Note that this is more strict than the generic clamping
437 // behavior of insertUse. We also try to handle cases which might run the
439 // FIXME: We should instead consider the pointer to have escaped if this
440 // function is being instrumented for addressing bugs or race conditions.
441 if (Offset.isNegative() || Size > AllocSize ||
442 Offset.ugt(AllocSize - Size)) {
443 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
444 << " which extends past the end of the " << AllocSize
446 << " alloca: " << S.AI << "\n"
447 << " use: " << SI << "\n");
448 return markAsDead(SI);
451 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
452 "All simple FCA stores should have been pre-split");
453 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
457 void visitMemSetInst(MemSetInst &II) {
458 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
459 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
460 if ((Length && Length->getValue() == 0) ||
461 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
462 // Zero-length mem transfer intrinsics can be ignored entirely.
463 return markAsDead(II);
466 return PI.setAborted(&II);
468 insertUse(II, Offset,
469 Length ? Length->getLimitedValue()
470 : AllocSize - Offset.getLimitedValue(),
474 void visitMemTransferInst(MemTransferInst &II) {
475 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
476 if (Length && Length->getValue() == 0)
477 // Zero-length mem transfer intrinsics can be ignored entirely.
478 return markAsDead(II);
480 // Because we can visit these intrinsics twice, also check to see if the
481 // first time marked this instruction as dead. If so, skip it.
482 if (VisitedDeadInsts.count(&II))
486 return PI.setAborted(&II);
488 // This side of the transfer is completely out-of-bounds, and so we can
489 // nuke the entire transfer. However, we also need to nuke the other side
490 // if already added to our partitions.
491 // FIXME: Yet another place we really should bypass this when
492 // instrumenting for ASan.
493 if (!Offset.isNegative() && Offset.uge(AllocSize)) {
494 SmallDenseMap<Instruction *, unsigned>::iterator MTPI = MemTransferSliceMap.find(&II);
495 if (MTPI != MemTransferSliceMap.end())
496 S.Slices[MTPI->second].kill();
497 return markAsDead(II);
500 uint64_t RawOffset = Offset.getLimitedValue();
501 uint64_t Size = Length ? Length->getLimitedValue()
502 : AllocSize - RawOffset;
504 // Check for the special case where the same exact value is used for both
506 if (*U == II.getRawDest() && *U == II.getRawSource()) {
507 // For non-volatile transfers this is a no-op.
508 if (!II.isVolatile())
509 return markAsDead(II);
511 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
514 // If we have seen both source and destination for a mem transfer, then
515 // they both point to the same alloca.
517 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
518 llvm::tie(MTPI, Inserted) =
519 MemTransferSliceMap.insert(std::make_pair(&II, S.Slices.size()));
520 unsigned PrevIdx = MTPI->second;
522 Slice &PrevP = S.Slices[PrevIdx];
524 // Check if the begin offsets match and this is a non-volatile transfer.
525 // In that case, we can completely elide the transfer.
526 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
528 return markAsDead(II);
531 // Otherwise we have an offset transfer within the same alloca. We can't
533 PrevP.makeUnsplittable();
536 // Insert the use now that we've fixed up the splittable nature.
537 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
539 // Check that we ended up with a valid index in the map.
540 assert(S.Slices[PrevIdx].getUse()->getUser() == &II &&
541 "Map index doesn't point back to a slice with this user.");
544 // Disable SRoA for any intrinsics except for lifetime invariants.
545 // FIXME: What about debug intrinsics? This matches old behavior, but
546 // doesn't make sense.
547 void visitIntrinsicInst(IntrinsicInst &II) {
549 return PI.setAborted(&II);
551 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
552 II.getIntrinsicID() == Intrinsic::lifetime_end) {
553 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
554 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
555 Length->getLimitedValue());
556 insertUse(II, Offset, Size, true);
560 Base::visitIntrinsicInst(II);
563 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
564 // We consider any PHI or select that results in a direct load or store of
565 // the same offset to be a viable use for slicing purposes. These uses
566 // are considered unsplittable and the size is the maximum loaded or stored
568 SmallPtrSet<Instruction *, 4> Visited;
569 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
570 Visited.insert(Root);
571 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
572 // If there are no loads or stores, the access is dead. We mark that as
573 // a size zero access.
576 Instruction *I, *UsedI;
577 llvm::tie(UsedI, I) = Uses.pop_back_val();
579 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
580 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
583 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
584 Value *Op = SI->getOperand(0);
587 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
591 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
592 if (!GEP->hasAllZeroIndices())
594 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
595 !isa<SelectInst>(I)) {
599 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
601 if (Visited.insert(cast<Instruction>(*UI)))
602 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
603 } while (!Uses.empty());
608 void visitPHINode(PHINode &PN) {
610 return markAsDead(PN);
612 return PI.setAborted(&PN);
614 // See if we already have computed info on this node.
615 uint64_t &PHISize = PHIOrSelectSizes[&PN];
617 // This is a new PHI node, check for an unsafe use of the PHI node.
618 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&PN, PHISize))
619 return PI.setAborted(UnsafeI);
622 // For PHI and select operands outside the alloca, we can't nuke the entire
623 // phi or select -- the other side might still be relevant, so we special
624 // case them here and use a separate structure to track the operands
625 // themselves which should be replaced with undef.
626 // FIXME: This should instead be escaped in the event we're instrumenting
627 // for address sanitization.
628 if ((Offset.isNegative() && (-Offset).uge(PHISize)) ||
629 (!Offset.isNegative() && Offset.uge(AllocSize))) {
630 S.DeadOperands.push_back(U);
634 insertUse(PN, Offset, PHISize);
637 void visitSelectInst(SelectInst &SI) {
639 return markAsDead(SI);
640 if (Value *Result = foldSelectInst(SI)) {
642 // If the result of the constant fold will be the pointer, recurse
643 // through the select as if we had RAUW'ed it.
646 // Otherwise the operand to the select is dead, and we can replace it
648 S.DeadOperands.push_back(U);
653 return PI.setAborted(&SI);
655 // See if we already have computed info on this node.
656 uint64_t &SelectSize = PHIOrSelectSizes[&SI];
658 // This is a new Select, check for an unsafe use of it.
659 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&SI, SelectSize))
660 return PI.setAborted(UnsafeI);
663 // For PHI and select operands outside the alloca, we can't nuke the entire
664 // phi or select -- the other side might still be relevant, so we special
665 // case them here and use a separate structure to track the operands
666 // themselves which should be replaced with undef.
667 // FIXME: This should instead be escaped in the event we're instrumenting
668 // for address sanitization.
669 if ((Offset.isNegative() && Offset.uge(SelectSize)) ||
670 (!Offset.isNegative() && Offset.uge(AllocSize))) {
671 S.DeadOperands.push_back(U);
675 insertUse(SI, Offset, SelectSize);
678 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
679 void visitInstruction(Instruction &I) {
684 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
686 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
689 PointerEscapingInstr(0) {
690 SliceBuilder PB(DL, AI, *this);
691 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
692 if (PtrI.isEscaped() || PtrI.isAborted()) {
693 // FIXME: We should sink the escape vs. abort info into the caller nicely,
694 // possibly by just storing the PtrInfo in the AllocaSlices.
695 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
696 : PtrI.getAbortingInst();
697 assert(PointerEscapingInstr && "Did not track a bad instruction");
701 Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
702 std::mem_fun_ref(&Slice::isDead)),
705 #if __cplusplus >= 201103L && !defined(NDEBUG)
706 if (SROARandomShuffleSlices) {
707 std::mt19937 MT(static_cast<unsigned>(sys::TimeValue::now().msec()));
708 std::shuffle(Slices.begin(), Slices.end(), MT);
712 // Sort the uses. This arranges for the offsets to be in ascending order,
713 // and the sizes to be in descending order.
714 std::sort(Slices.begin(), Slices.end());
717 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
719 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
720 StringRef Indent) const {
721 printSlice(OS, I, Indent);
722 printUse(OS, I, Indent);
725 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
726 StringRef Indent) const {
727 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
728 << " slice #" << (I - begin())
729 << (I->isSplittable() ? " (splittable)" : "") << "\n";
732 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
733 StringRef Indent) const {
734 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
737 void AllocaSlices::print(raw_ostream &OS) const {
738 if (PointerEscapingInstr) {
739 OS << "Can't analyze slices for alloca: " << AI << "\n"
740 << " A pointer to this alloca escaped by:\n"
741 << " " << *PointerEscapingInstr << "\n";
745 OS << "Slices of alloca: " << AI << "\n";
746 for (const_iterator I = begin(), E = end(); I != E; ++I)
750 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
753 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
755 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
758 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
760 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
761 /// the loads and stores of an alloca instruction, as well as updating its
762 /// debug information. This is used when a domtree is unavailable and thus
763 /// mem2reg in its full form can't be used to handle promotion of allocas to
765 class AllocaPromoter : public LoadAndStorePromoter {
769 SmallVector<DbgDeclareInst *, 4> DDIs;
770 SmallVector<DbgValueInst *, 4> DVIs;
773 AllocaPromoter(const SmallVectorImpl<Instruction *> &Insts, SSAUpdater &S,
774 AllocaInst &AI, DIBuilder &DIB)
775 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
777 void run(const SmallVectorImpl<Instruction*> &Insts) {
778 // Retain the debug information attached to the alloca for use when
779 // rewriting loads and stores.
780 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
781 for (Value::use_iterator UI = DebugNode->use_begin(),
782 UE = DebugNode->use_end();
784 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
786 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
790 LoadAndStorePromoter::run(Insts);
792 // While we have the debug information, clear it off of the alloca. The
793 // caller takes care of deleting the alloca.
794 while (!DDIs.empty())
795 DDIs.pop_back_val()->eraseFromParent();
796 while (!DVIs.empty())
797 DVIs.pop_back_val()->eraseFromParent();
800 virtual bool isInstInList(Instruction *I,
801 const SmallVectorImpl<Instruction*> &Insts) const {
803 if (LoadInst *LI = dyn_cast<LoadInst>(I))
804 Ptr = LI->getOperand(0);
806 Ptr = cast<StoreInst>(I)->getPointerOperand();
808 // Only used to detect cycles, which will be rare and quickly found as
809 // we're walking up a chain of defs rather than down through uses.
810 SmallPtrSet<Value *, 4> Visited;
816 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr))
817 Ptr = BCI->getOperand(0);
818 else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr))
819 Ptr = GEPI->getPointerOperand();
823 } while (Visited.insert(Ptr));
828 virtual void updateDebugInfo(Instruction *Inst) const {
829 for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(),
830 E = DDIs.end(); I != E; ++I) {
831 DbgDeclareInst *DDI = *I;
832 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
833 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
834 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
835 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
837 for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(),
838 E = DVIs.end(); I != E; ++I) {
839 DbgValueInst *DVI = *I;
841 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
842 // If an argument is zero extended then use argument directly. The ZExt
843 // may be zapped by an optimization pass in future.
844 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
845 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
846 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
847 Arg = dyn_cast<Argument>(SExt->getOperand(0));
849 Arg = SI->getValueOperand();
850 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
851 Arg = LI->getPointerOperand();
855 Instruction *DbgVal =
856 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
858 DbgVal->setDebugLoc(DVI->getDebugLoc());
862 } // end anon namespace
866 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
868 /// This pass takes allocations which can be completely analyzed (that is, they
869 /// don't escape) and tries to turn them into scalar SSA values. There are
870 /// a few steps to this process.
872 /// 1) It takes allocations of aggregates and analyzes the ways in which they
873 /// are used to try to split them into smaller allocations, ideally of
874 /// a single scalar data type. It will split up memcpy and memset accesses
875 /// as necessary and try to isolate individual scalar accesses.
876 /// 2) It will transform accesses into forms which are suitable for SSA value
877 /// promotion. This can be replacing a memset with a scalar store of an
878 /// integer value, or it can involve speculating operations on a PHI or
879 /// select to be a PHI or select of the results.
880 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
881 /// onto insert and extract operations on a vector value, and convert them to
882 /// this form. By doing so, it will enable promotion of vector aggregates to
883 /// SSA vector values.
884 class SROA : public FunctionPass {
885 const bool RequiresDomTree;
888 const DataLayout *DL;
891 /// \brief Worklist of alloca instructions to simplify.
893 /// Each alloca in the function is added to this. Each new alloca formed gets
894 /// added to it as well to recursively simplify unless that alloca can be
895 /// directly promoted. Finally, each time we rewrite a use of an alloca other
896 /// the one being actively rewritten, we add it back onto the list if not
897 /// already present to ensure it is re-visited.
898 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
900 /// \brief A collection of instructions to delete.
901 /// We try to batch deletions to simplify code and make things a bit more
903 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
905 /// \brief Post-promotion worklist.
907 /// Sometimes we discover an alloca which has a high probability of becoming
908 /// viable for SROA after a round of promotion takes place. In those cases,
909 /// the alloca is enqueued here for re-processing.
911 /// Note that we have to be very careful to clear allocas out of this list in
912 /// the event they are deleted.
913 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
915 /// \brief A collection of alloca instructions we can directly promote.
916 std::vector<AllocaInst *> PromotableAllocas;
918 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
920 /// All of these PHIs have been checked for the safety of speculation and by
921 /// being speculated will allow promoting allocas currently in the promotable
923 SetVector<PHINode *, SmallVector<PHINode *, 2> > SpeculatablePHIs;
925 /// \brief A worklist of select instructions to speculate prior to promoting
928 /// All of these select instructions have been checked for the safety of
929 /// speculation and by being speculated will allow promoting allocas
930 /// currently in the promotable queue.
931 SetVector<SelectInst *, SmallVector<SelectInst *, 2> > SpeculatableSelects;
934 SROA(bool RequiresDomTree = true)
935 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
937 initializeSROAPass(*PassRegistry::getPassRegistry());
939 bool runOnFunction(Function &F);
940 void getAnalysisUsage(AnalysisUsage &AU) const;
942 const char *getPassName() const { return "SROA"; }
946 friend class PHIOrSelectSpeculator;
947 friend class AllocaSliceRewriter;
949 bool rewritePartition(AllocaInst &AI, AllocaSlices &S,
950 AllocaSlices::iterator B, AllocaSlices::iterator E,
951 int64_t BeginOffset, int64_t EndOffset,
952 ArrayRef<AllocaSlices::iterator> SplitUses);
953 bool splitAlloca(AllocaInst &AI, AllocaSlices &S);
954 bool runOnAlloca(AllocaInst &AI);
955 void clobberUse(Use &U);
956 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
957 bool promoteAllocas(Function &F);
963 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
964 return new SROA(RequiresDomTree);
967 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
969 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
970 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
973 /// Walk the range of a partitioning looking for a common type to cover this
974 /// sequence of slices.
975 static Type *findCommonType(AllocaSlices::const_iterator B,
976 AllocaSlices::const_iterator E,
977 uint64_t EndOffset) {
979 bool TyIsCommon = true;
980 IntegerType *ITy = 0;
982 // Note that we need to look at *every* alloca slice's Use to ensure we
983 // always get consistent results regardless of the order of slices.
984 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
985 Use *U = I->getUse();
986 if (isa<IntrinsicInst>(*U->getUser()))
988 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
992 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
993 UserTy = LI->getType();
994 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
995 UserTy = SI->getValueOperand()->getType();
998 if (!UserTy || (Ty && Ty != UserTy))
999 TyIsCommon = false; // Give up on anything but an iN type.
1003 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1004 // If the type is larger than the partition, skip it. We only encounter
1005 // this for split integer operations where we want to use the type of the
1006 // entity causing the split. Also skip if the type is not a byte width
1008 if (UserITy->getBitWidth() % 8 != 0 ||
1009 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1012 // Track the largest bitwidth integer type used in this way in case there
1013 // is no common type.
1014 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1019 return TyIsCommon ? Ty : ITy;
1022 /// PHI instructions that use an alloca and are subsequently loaded can be
1023 /// rewritten to load both input pointers in the pred blocks and then PHI the
1024 /// results, allowing the load of the alloca to be promoted.
1026 /// %P2 = phi [i32* %Alloca, i32* %Other]
1027 /// %V = load i32* %P2
1029 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1031 /// %V2 = load i32* %Other
1033 /// %V = phi [i32 %V1, i32 %V2]
1035 /// We can do this to a select if its only uses are loads and if the operands
1036 /// to the select can be loaded unconditionally.
1038 /// FIXME: This should be hoisted into a generic utility, likely in
1039 /// Transforms/Util/Local.h
1040 static bool isSafePHIToSpeculate(PHINode &PN,
1041 const DataLayout *DL = 0) {
1042 // For now, we can only do this promotion if the load is in the same block
1043 // as the PHI, and if there are no stores between the phi and load.
1044 // TODO: Allow recursive phi users.
1045 // TODO: Allow stores.
1046 BasicBlock *BB = PN.getParent();
1047 unsigned MaxAlign = 0;
1048 bool HaveLoad = false;
1049 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end(); UI != UE;
1051 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1052 if (LI == 0 || !LI->isSimple())
1055 // For now we only allow loads in the same block as the PHI. This is
1056 // a common case that happens when instcombine merges two loads through
1058 if (LI->getParent() != BB)
1061 // Ensure that there are no instructions between the PHI and the load that
1063 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1064 if (BBI->mayWriteToMemory())
1067 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1074 // We can only transform this if it is safe to push the loads into the
1075 // predecessor blocks. The only thing to watch out for is that we can't put
1076 // a possibly trapping load in the predecessor if it is a critical edge.
1077 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1078 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1079 Value *InVal = PN.getIncomingValue(Idx);
1081 // If the value is produced by the terminator of the predecessor (an
1082 // invoke) or it has side-effects, there is no valid place to put a load
1083 // in the predecessor.
1084 if (TI == InVal || TI->mayHaveSideEffects())
1087 // If the predecessor has a single successor, then the edge isn't
1089 if (TI->getNumSuccessors() == 1)
1092 // If this pointer is always safe to load, or if we can prove that there
1093 // is already a load in the block, then we can move the load to the pred
1095 if (InVal->isDereferenceablePointer() ||
1096 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, DL))
1105 static void speculatePHINodeLoads(PHINode &PN) {
1106 DEBUG(dbgs() << " original: " << PN << "\n");
1108 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1109 IRBuilderTy PHIBuilder(&PN);
1110 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1111 PN.getName() + ".sroa.speculated");
1113 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1114 // matter which one we get and if any differ.
1115 LoadInst *SomeLoad = cast<LoadInst>(*PN.use_begin());
1116 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1117 unsigned Align = SomeLoad->getAlignment();
1119 // Rewrite all loads of the PN to use the new PHI.
1120 while (!PN.use_empty()) {
1121 LoadInst *LI = cast<LoadInst>(*PN.use_begin());
1122 LI->replaceAllUsesWith(NewPN);
1123 LI->eraseFromParent();
1126 // Inject loads into all of the pred blocks.
1127 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1128 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1129 TerminatorInst *TI = Pred->getTerminator();
1130 Value *InVal = PN.getIncomingValue(Idx);
1131 IRBuilderTy PredBuilder(TI);
1133 LoadInst *Load = PredBuilder.CreateLoad(
1134 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1135 ++NumLoadsSpeculated;
1136 Load->setAlignment(Align);
1138 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1139 NewPN->addIncoming(Load, Pred);
1142 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1143 PN.eraseFromParent();
1146 /// Select instructions that use an alloca and are subsequently loaded can be
1147 /// rewritten to load both input pointers and then select between the result,
1148 /// allowing the load of the alloca to be promoted.
1150 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1151 /// %V = load i32* %P2
1153 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1154 /// %V2 = load i32* %Other
1155 /// %V = select i1 %cond, i32 %V1, i32 %V2
1157 /// We can do this to a select if its only uses are loads and if the operand
1158 /// to the select can be loaded unconditionally.
1159 static bool isSafeSelectToSpeculate(SelectInst &SI, const DataLayout *DL = 0) {
1160 Value *TValue = SI.getTrueValue();
1161 Value *FValue = SI.getFalseValue();
1162 bool TDerefable = TValue->isDereferenceablePointer();
1163 bool FDerefable = FValue->isDereferenceablePointer();
1165 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end(); UI != UE;
1167 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1168 if (LI == 0 || !LI->isSimple())
1171 // Both operands to the select need to be dereferencable, either
1172 // absolutely (e.g. allocas) or at this point because we can see other
1175 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment(), DL))
1178 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment(), DL))
1185 static void speculateSelectInstLoads(SelectInst &SI) {
1186 DEBUG(dbgs() << " original: " << SI << "\n");
1188 IRBuilderTy IRB(&SI);
1189 Value *TV = SI.getTrueValue();
1190 Value *FV = SI.getFalseValue();
1191 // Replace the loads of the select with a select of two loads.
1192 while (!SI.use_empty()) {
1193 LoadInst *LI = cast<LoadInst>(*SI.use_begin());
1194 assert(LI->isSimple() && "We only speculate simple loads");
1196 IRB.SetInsertPoint(LI);
1198 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1200 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1201 NumLoadsSpeculated += 2;
1203 // Transfer alignment and TBAA info if present.
1204 TL->setAlignment(LI->getAlignment());
1205 FL->setAlignment(LI->getAlignment());
1206 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1207 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1208 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1211 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1212 LI->getName() + ".sroa.speculated");
1214 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1215 LI->replaceAllUsesWith(V);
1216 LI->eraseFromParent();
1218 SI.eraseFromParent();
1221 /// \brief Build a GEP out of a base pointer and indices.
1223 /// This will return the BasePtr if that is valid, or build a new GEP
1224 /// instruction using the IRBuilder if GEP-ing is needed.
1225 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1226 SmallVectorImpl<Value *> &Indices) {
1227 if (Indices.empty())
1230 // A single zero index is a no-op, so check for this and avoid building a GEP
1232 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1235 return IRB.CreateInBoundsGEP(BasePtr, Indices, "idx");
1238 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1239 /// TargetTy without changing the offset of the pointer.
1241 /// This routine assumes we've already established a properly offset GEP with
1242 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1243 /// zero-indices down through type layers until we find one the same as
1244 /// TargetTy. If we can't find one with the same type, we at least try to use
1245 /// one with the same size. If none of that works, we just produce the GEP as
1246 /// indicated by Indices to have the correct offset.
1247 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1248 Value *BasePtr, Type *Ty, Type *TargetTy,
1249 SmallVectorImpl<Value *> &Indices) {
1251 return buildGEP(IRB, BasePtr, Indices);
1253 // See if we can descend into a struct and locate a field with the correct
1255 unsigned NumLayers = 0;
1256 Type *ElementTy = Ty;
1258 if (ElementTy->isPointerTy())
1260 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1261 ElementTy = SeqTy->getElementType();
1262 // Note that we use the default address space as this index is over an
1263 // array or a vector, not a pointer.
1264 Indices.push_back(IRB.getInt(APInt(DL.getPointerSizeInBits(0), 0)));
1265 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1266 if (STy->element_begin() == STy->element_end())
1267 break; // Nothing left to descend into.
1268 ElementTy = *STy->element_begin();
1269 Indices.push_back(IRB.getInt32(0));
1274 } while (ElementTy != TargetTy);
1275 if (ElementTy != TargetTy)
1276 Indices.erase(Indices.end() - NumLayers, Indices.end());
1278 return buildGEP(IRB, BasePtr, Indices);
1281 /// \brief Recursively compute indices for a natural GEP.
1283 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1284 /// element types adding appropriate indices for the GEP.
1285 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1286 Value *Ptr, Type *Ty, APInt &Offset,
1288 SmallVectorImpl<Value *> &Indices) {
1290 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices);
1292 // We can't recurse through pointer types.
1293 if (Ty->isPointerTy())
1296 // We try to analyze GEPs over vectors here, but note that these GEPs are
1297 // extremely poorly defined currently. The long-term goal is to remove GEPing
1298 // over a vector from the IR completely.
1299 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1300 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1301 if (ElementSizeInBits % 8)
1302 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1303 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1304 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1305 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1307 Offset -= NumSkippedElements * ElementSize;
1308 Indices.push_back(IRB.getInt(NumSkippedElements));
1309 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1310 Offset, TargetTy, Indices);
1313 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1314 Type *ElementTy = ArrTy->getElementType();
1315 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1316 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1317 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1320 Offset -= NumSkippedElements * ElementSize;
1321 Indices.push_back(IRB.getInt(NumSkippedElements));
1322 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1326 StructType *STy = dyn_cast<StructType>(Ty);
1330 const StructLayout *SL = DL.getStructLayout(STy);
1331 uint64_t StructOffset = Offset.getZExtValue();
1332 if (StructOffset >= SL->getSizeInBytes())
1334 unsigned Index = SL->getElementContainingOffset(StructOffset);
1335 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1336 Type *ElementTy = STy->getElementType(Index);
1337 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1338 return 0; // The offset points into alignment padding.
1340 Indices.push_back(IRB.getInt32(Index));
1341 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1345 /// \brief Get a natural GEP from a base pointer to a particular offset and
1346 /// resulting in a particular type.
1348 /// The goal is to produce a "natural" looking GEP that works with the existing
1349 /// composite types to arrive at the appropriate offset and element type for
1350 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1351 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1352 /// Indices, and setting Ty to the result subtype.
1354 /// If no natural GEP can be constructed, this function returns null.
1355 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1356 Value *Ptr, APInt Offset, Type *TargetTy,
1357 SmallVectorImpl<Value *> &Indices) {
1358 PointerType *Ty = cast<PointerType>(Ptr->getType());
1360 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1362 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1365 Type *ElementTy = Ty->getElementType();
1366 if (!ElementTy->isSized())
1367 return 0; // We can't GEP through an unsized element.
1368 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1369 if (ElementSize == 0)
1370 return 0; // Zero-length arrays can't help us build a natural GEP.
1371 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1373 Offset -= NumSkippedElements * ElementSize;
1374 Indices.push_back(IRB.getInt(NumSkippedElements));
1375 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1379 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1380 /// resulting pointer has PointerTy.
1382 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1383 /// and produces the pointer type desired. Where it cannot, it will try to use
1384 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1385 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1386 /// bitcast to the type.
1388 /// The strategy for finding the more natural GEPs is to peel off layers of the
1389 /// pointer, walking back through bit casts and GEPs, searching for a base
1390 /// pointer from which we can compute a natural GEP with the desired
1391 /// properties. The algorithm tries to fold as many constant indices into
1392 /// a single GEP as possible, thus making each GEP more independent of the
1393 /// surrounding code.
1394 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL,
1395 Value *Ptr, APInt Offset, Type *PointerTy) {
1396 // Even though we don't look through PHI nodes, we could be called on an
1397 // instruction in an unreachable block, which may be on a cycle.
1398 SmallPtrSet<Value *, 4> Visited;
1399 Visited.insert(Ptr);
1400 SmallVector<Value *, 4> Indices;
1402 // We may end up computing an offset pointer that has the wrong type. If we
1403 // never are able to compute one directly that has the correct type, we'll
1404 // fall back to it, so keep it around here.
1405 Value *OffsetPtr = 0;
1407 // Remember any i8 pointer we come across to re-use if we need to do a raw
1410 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1412 Type *TargetTy = PointerTy->getPointerElementType();
1415 // First fold any existing GEPs into the offset.
1416 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1417 APInt GEPOffset(Offset.getBitWidth(), 0);
1418 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1420 Offset += GEPOffset;
1421 Ptr = GEP->getPointerOperand();
1422 if (!Visited.insert(Ptr))
1426 // See if we can perform a natural GEP here.
1428 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1430 if (P->getType() == PointerTy) {
1431 // Zap any offset pointer that we ended up computing in previous rounds.
1432 if (OffsetPtr && OffsetPtr->use_empty())
1433 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1434 I->eraseFromParent();
1442 // Stash this pointer if we've found an i8*.
1443 if (Ptr->getType()->isIntegerTy(8)) {
1445 Int8PtrOffset = Offset;
1448 // Peel off a layer of the pointer and update the offset appropriately.
1449 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1450 Ptr = cast<Operator>(Ptr)->getOperand(0);
1451 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1452 if (GA->mayBeOverridden())
1454 Ptr = GA->getAliasee();
1458 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1459 } while (Visited.insert(Ptr));
1463 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1465 Int8PtrOffset = Offset;
1468 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1469 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1474 // On the off chance we were targeting i8*, guard the bitcast here.
1475 if (Ptr->getType() != PointerTy)
1476 Ptr = IRB.CreateBitCast(Ptr, PointerTy, "cast");
1481 /// \brief Test whether we can convert a value from the old to the new type.
1483 /// This predicate should be used to guard calls to convertValue in order to
1484 /// ensure that we only try to convert viable values. The strategy is that we
1485 /// will peel off single element struct and array wrappings to get to an
1486 /// underlying value, and convert that value.
1487 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1490 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1491 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1492 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1494 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1496 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1499 // We can convert pointers to integers and vice-versa. Same for vectors
1500 // of pointers and integers.
1501 OldTy = OldTy->getScalarType();
1502 NewTy = NewTy->getScalarType();
1503 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1504 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1506 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1514 /// \brief Generic routine to convert an SSA value to a value of a different
1517 /// This will try various different casting techniques, such as bitcasts,
1518 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1519 /// two types for viability with this routine.
1520 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1522 Type *OldTy = V->getType();
1523 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1528 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1529 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1530 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1531 return IRB.CreateZExt(V, NewITy);
1533 // See if we need inttoptr for this type pair. A cast involving both scalars
1534 // and vectors requires and additional bitcast.
1535 if (OldTy->getScalarType()->isIntegerTy() &&
1536 NewTy->getScalarType()->isPointerTy()) {
1537 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1538 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1539 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1542 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1543 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1544 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1547 return IRB.CreateIntToPtr(V, NewTy);
1550 // See if we need ptrtoint for this type pair. A cast involving both scalars
1551 // and vectors requires and additional bitcast.
1552 if (OldTy->getScalarType()->isPointerTy() &&
1553 NewTy->getScalarType()->isIntegerTy()) {
1554 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1555 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1556 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1559 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1560 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1561 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1564 return IRB.CreatePtrToInt(V, NewTy);
1567 return IRB.CreateBitCast(V, NewTy);
1570 /// \brief Test whether the given slice use can be promoted to a vector.
1572 /// This function is called to test each entry in a partioning which is slated
1573 /// for a single slice.
1574 static bool isVectorPromotionViableForSlice(
1575 const DataLayout &DL, AllocaSlices &S, uint64_t SliceBeginOffset,
1576 uint64_t SliceEndOffset, VectorType *Ty, uint64_t ElementSize,
1577 AllocaSlices::const_iterator I) {
1578 // First validate the slice offsets.
1579 uint64_t BeginOffset =
1580 std::max(I->beginOffset(), SliceBeginOffset) - SliceBeginOffset;
1581 uint64_t BeginIndex = BeginOffset / ElementSize;
1582 if (BeginIndex * ElementSize != BeginOffset ||
1583 BeginIndex >= Ty->getNumElements())
1585 uint64_t EndOffset =
1586 std::min(I->endOffset(), SliceEndOffset) - SliceBeginOffset;
1587 uint64_t EndIndex = EndOffset / ElementSize;
1588 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1591 assert(EndIndex > BeginIndex && "Empty vector!");
1592 uint64_t NumElements = EndIndex - BeginIndex;
1594 (NumElements == 1) ? Ty->getElementType()
1595 : VectorType::get(Ty->getElementType(), NumElements);
1598 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1600 Use *U = I->getUse();
1602 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1603 if (MI->isVolatile())
1605 if (!I->isSplittable())
1606 return false; // Skip any unsplittable intrinsics.
1607 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1608 // Disable vector promotion when there are loads or stores of an FCA.
1610 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1611 if (LI->isVolatile())
1613 Type *LTy = LI->getType();
1614 if (SliceBeginOffset > I->beginOffset() ||
1615 SliceEndOffset < I->endOffset()) {
1616 assert(LTy->isIntegerTy());
1619 if (!canConvertValue(DL, SliceTy, LTy))
1621 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1622 if (SI->isVolatile())
1624 Type *STy = SI->getValueOperand()->getType();
1625 if (SliceBeginOffset > I->beginOffset() ||
1626 SliceEndOffset < I->endOffset()) {
1627 assert(STy->isIntegerTy());
1630 if (!canConvertValue(DL, STy, SliceTy))
1639 /// \brief Test whether the given alloca partitioning and range of slices can be
1640 /// promoted to a vector.
1642 /// This is a quick test to check whether we can rewrite a particular alloca
1643 /// partition (and its newly formed alloca) into a vector alloca with only
1644 /// whole-vector loads and stores such that it could be promoted to a vector
1645 /// SSA value. We only can ensure this for a limited set of operations, and we
1646 /// don't want to do the rewrites unless we are confident that the result will
1647 /// be promotable, so we have an early test here.
1649 isVectorPromotionViable(const DataLayout &DL, Type *AllocaTy, AllocaSlices &S,
1650 uint64_t SliceBeginOffset, uint64_t SliceEndOffset,
1651 AllocaSlices::const_iterator I,
1652 AllocaSlices::const_iterator E,
1653 ArrayRef<AllocaSlices::iterator> SplitUses) {
1654 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1658 uint64_t ElementSize = DL.getTypeSizeInBits(Ty->getScalarType());
1660 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1661 // that aren't byte sized.
1662 if (ElementSize % 8)
1664 assert((DL.getTypeSizeInBits(Ty) % 8) == 0 &&
1665 "vector size not a multiple of element size?");
1669 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1670 SliceEndOffset, Ty, ElementSize, I))
1673 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1674 SUE = SplitUses.end();
1676 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1677 SliceEndOffset, Ty, ElementSize, *SUI))
1683 /// \brief Test whether a slice of an alloca is valid for integer widening.
1685 /// This implements the necessary checking for the \c isIntegerWideningViable
1686 /// test below on a single slice of the alloca.
1687 static bool isIntegerWideningViableForSlice(const DataLayout &DL,
1689 uint64_t AllocBeginOffset,
1690 uint64_t Size, AllocaSlices &S,
1691 AllocaSlices::const_iterator I,
1692 bool &WholeAllocaOp) {
1693 uint64_t RelBegin = I->beginOffset() - AllocBeginOffset;
1694 uint64_t RelEnd = I->endOffset() - AllocBeginOffset;
1696 // We can't reasonably handle cases where the load or store extends past
1697 // the end of the aloca's type and into its padding.
1701 Use *U = I->getUse();
1703 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1704 if (LI->isVolatile())
1706 if (RelBegin == 0 && RelEnd == Size)
1707 WholeAllocaOp = true;
1708 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1709 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1711 } else if (RelBegin != 0 || RelEnd != Size ||
1712 !canConvertValue(DL, AllocaTy, LI->getType())) {
1713 // Non-integer loads need to be convertible from the alloca type so that
1714 // they are promotable.
1717 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1718 Type *ValueTy = SI->getValueOperand()->getType();
1719 if (SI->isVolatile())
1721 if (RelBegin == 0 && RelEnd == Size)
1722 WholeAllocaOp = true;
1723 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1724 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1726 } else if (RelBegin != 0 || RelEnd != Size ||
1727 !canConvertValue(DL, ValueTy, AllocaTy)) {
1728 // Non-integer stores need to be convertible to the alloca type so that
1729 // they are promotable.
1732 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1733 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1735 if (!I->isSplittable())
1736 return false; // Skip any unsplittable intrinsics.
1737 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1738 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1739 II->getIntrinsicID() != Intrinsic::lifetime_end)
1748 /// \brief Test whether the given alloca partition's integer operations can be
1749 /// widened to promotable ones.
1751 /// This is a quick test to check whether we can rewrite the integer loads and
1752 /// stores to a particular alloca into wider loads and stores and be able to
1753 /// promote the resulting alloca.
1755 isIntegerWideningViable(const DataLayout &DL, Type *AllocaTy,
1756 uint64_t AllocBeginOffset, AllocaSlices &S,
1757 AllocaSlices::const_iterator I,
1758 AllocaSlices::const_iterator E,
1759 ArrayRef<AllocaSlices::iterator> SplitUses) {
1760 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1761 // Don't create integer types larger than the maximum bitwidth.
1762 if (SizeInBits > IntegerType::MAX_INT_BITS)
1765 // Don't try to handle allocas with bit-padding.
1766 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1769 // We need to ensure that an integer type with the appropriate bitwidth can
1770 // be converted to the alloca type, whatever that is. We don't want to force
1771 // the alloca itself to have an integer type if there is a more suitable one.
1772 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1773 if (!canConvertValue(DL, AllocaTy, IntTy) ||
1774 !canConvertValue(DL, IntTy, AllocaTy))
1777 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1779 // While examining uses, we ensure that the alloca has a covering load or
1780 // store. We don't want to widen the integer operations only to fail to
1781 // promote due to some other unsplittable entry (which we may make splittable
1782 // later). However, if there are only splittable uses, go ahead and assume
1783 // that we cover the alloca.
1784 bool WholeAllocaOp = (I != E) ? false : DL.isLegalInteger(SizeInBits);
1787 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1788 S, I, WholeAllocaOp))
1791 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1792 SUE = SplitUses.end();
1794 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1795 S, *SUI, WholeAllocaOp))
1798 return WholeAllocaOp;
1801 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1802 IntegerType *Ty, uint64_t Offset,
1803 const Twine &Name) {
1804 DEBUG(dbgs() << " start: " << *V << "\n");
1805 IntegerType *IntTy = cast<IntegerType>(V->getType());
1806 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1807 "Element extends past full value");
1808 uint64_t ShAmt = 8*Offset;
1809 if (DL.isBigEndian())
1810 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1812 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
1813 DEBUG(dbgs() << " shifted: " << *V << "\n");
1815 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1816 "Cannot extract to a larger integer!");
1818 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
1819 DEBUG(dbgs() << " trunced: " << *V << "\n");
1824 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
1825 Value *V, uint64_t Offset, const Twine &Name) {
1826 IntegerType *IntTy = cast<IntegerType>(Old->getType());
1827 IntegerType *Ty = cast<IntegerType>(V->getType());
1828 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1829 "Cannot insert a larger integer!");
1830 DEBUG(dbgs() << " start: " << *V << "\n");
1832 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
1833 DEBUG(dbgs() << " extended: " << *V << "\n");
1835 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1836 "Element store outside of alloca store");
1837 uint64_t ShAmt = 8*Offset;
1838 if (DL.isBigEndian())
1839 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1841 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
1842 DEBUG(dbgs() << " shifted: " << *V << "\n");
1845 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
1846 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
1847 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
1848 DEBUG(dbgs() << " masked: " << *Old << "\n");
1849 V = IRB.CreateOr(Old, V, Name + ".insert");
1850 DEBUG(dbgs() << " inserted: " << *V << "\n");
1855 static Value *extractVector(IRBuilderTy &IRB, Value *V,
1856 unsigned BeginIndex, unsigned EndIndex,
1857 const Twine &Name) {
1858 VectorType *VecTy = cast<VectorType>(V->getType());
1859 unsigned NumElements = EndIndex - BeginIndex;
1860 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
1862 if (NumElements == VecTy->getNumElements())
1865 if (NumElements == 1) {
1866 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
1868 DEBUG(dbgs() << " extract: " << *V << "\n");
1872 SmallVector<Constant*, 8> Mask;
1873 Mask.reserve(NumElements);
1874 for (unsigned i = BeginIndex; i != EndIndex; ++i)
1875 Mask.push_back(IRB.getInt32(i));
1876 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1877 ConstantVector::get(Mask),
1879 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1883 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
1884 unsigned BeginIndex, const Twine &Name) {
1885 VectorType *VecTy = cast<VectorType>(Old->getType());
1886 assert(VecTy && "Can only insert a vector into a vector");
1888 VectorType *Ty = dyn_cast<VectorType>(V->getType());
1890 // Single element to insert.
1891 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
1893 DEBUG(dbgs() << " insert: " << *V << "\n");
1897 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
1898 "Too many elements!");
1899 if (Ty->getNumElements() == VecTy->getNumElements()) {
1900 assert(V->getType() == VecTy && "Vector type mismatch");
1903 unsigned EndIndex = BeginIndex + Ty->getNumElements();
1905 // When inserting a smaller vector into the larger to store, we first
1906 // use a shuffle vector to widen it with undef elements, and then
1907 // a second shuffle vector to select between the loaded vector and the
1909 SmallVector<Constant*, 8> Mask;
1910 Mask.reserve(VecTy->getNumElements());
1911 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1912 if (i >= BeginIndex && i < EndIndex)
1913 Mask.push_back(IRB.getInt32(i - BeginIndex));
1915 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
1916 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1917 ConstantVector::get(Mask),
1919 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1922 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1923 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
1925 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
1927 DEBUG(dbgs() << " blend: " << *V << "\n");
1932 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
1933 /// to use a new alloca.
1935 /// Also implements the rewriting to vector-based accesses when the partition
1936 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
1938 class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
1939 // Befriend the base class so it can delegate to private visit methods.
1940 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
1941 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
1943 const DataLayout &DL;
1946 AllocaInst &OldAI, &NewAI;
1947 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
1950 // If we are rewriting an alloca partition which can be written as pure
1951 // vector operations, we stash extra information here. When VecTy is
1952 // non-null, we have some strict guarantees about the rewritten alloca:
1953 // - The new alloca is exactly the size of the vector type here.
1954 // - The accesses all either map to the entire vector or to a single
1956 // - The set of accessing instructions is only one of those handled above
1957 // in isVectorPromotionViable. Generally these are the same access kinds
1958 // which are promotable via mem2reg.
1961 uint64_t ElementSize;
1963 // This is a convenience and flag variable that will be null unless the new
1964 // alloca's integer operations should be widened to this integer type due to
1965 // passing isIntegerWideningViable above. If it is non-null, the desired
1966 // integer type will be stored here for easy access during rewriting.
1969 // The offset of the slice currently being rewritten.
1970 uint64_t BeginOffset, EndOffset;
1974 Instruction *OldPtr;
1976 // Track post-rewrite users which are PHI nodes and Selects.
1977 SmallPtrSetImpl<PHINode *> &PHIUsers;
1978 SmallPtrSetImpl<SelectInst *> &SelectUsers;
1980 // Utility IR builder, whose name prefix is setup for each visited use, and
1981 // the insertion point is set to point to the user.
1985 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &S, SROA &Pass,
1986 AllocaInst &OldAI, AllocaInst &NewAI,
1987 uint64_t NewBeginOffset, uint64_t NewEndOffset,
1988 bool IsVectorPromotable, bool IsIntegerPromotable,
1989 SmallPtrSetImpl<PHINode *> &PHIUsers,
1990 SmallPtrSetImpl<SelectInst *> &SelectUsers)
1991 : DL(DL), S(S), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
1992 NewAllocaBeginOffset(NewBeginOffset), NewAllocaEndOffset(NewEndOffset),
1993 NewAllocaTy(NewAI.getAllocatedType()),
1994 VecTy(IsVectorPromotable ? cast<VectorType>(NewAllocaTy) : 0),
1995 ElementTy(VecTy ? VecTy->getElementType() : 0),
1996 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
1997 IntTy(IsIntegerPromotable
2000 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2002 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2003 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2004 IRB(NewAI.getContext(), ConstantFolder()) {
2006 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2007 "Only multiple-of-8 sized vector elements are viable");
2010 assert((!IsVectorPromotable && !IsIntegerPromotable) ||
2011 IsVectorPromotable != IsIntegerPromotable);
2014 bool visit(AllocaSlices::const_iterator I) {
2015 bool CanSROA = true;
2016 BeginOffset = I->beginOffset();
2017 EndOffset = I->endOffset();
2018 IsSplittable = I->isSplittable();
2020 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2022 OldUse = I->getUse();
2023 OldPtr = cast<Instruction>(OldUse->get());
2025 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2026 IRB.SetInsertPoint(OldUserI);
2027 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2028 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2030 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2037 // Make sure the other visit overloads are visible.
2040 // Every instruction which can end up as a user must have a rewrite rule.
2041 bool visitInstruction(Instruction &I) {
2042 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2043 llvm_unreachable("No rewrite rule for this instruction!");
2046 Value *getAdjustedAllocaPtr(IRBuilderTy &IRB, uint64_t Offset,
2048 assert(Offset >= NewAllocaBeginOffset);
2049 return getAdjustedPtr(IRB, DL, &NewAI, APInt(DL.getPointerSizeInBits(),
2050 Offset - NewAllocaBeginOffset),
2054 /// \brief Compute suitable alignment to access an offset into the new alloca.
2055 unsigned getOffsetAlign(uint64_t Offset) {
2056 unsigned NewAIAlign = NewAI.getAlignment();
2058 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2059 return MinAlign(NewAIAlign, Offset);
2062 /// \brief Compute suitable alignment to access a type at an offset of the
2065 /// \returns zero if the type's ABI alignment is a suitable alignment,
2066 /// otherwise returns the maximal suitable alignment.
2067 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
2068 unsigned Align = getOffsetAlign(Offset);
2069 return Align == DL.getABITypeAlignment(Ty) ? 0 : Align;
2072 unsigned getIndex(uint64_t Offset) {
2073 assert(VecTy && "Can only call getIndex when rewriting a vector");
2074 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2075 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2076 uint32_t Index = RelOffset / ElementSize;
2077 assert(Index * ElementSize == RelOffset);
2081 void deleteIfTriviallyDead(Value *V) {
2082 Instruction *I = cast<Instruction>(V);
2083 if (isInstructionTriviallyDead(I))
2084 Pass.DeadInsts.insert(I);
2087 Value *rewriteVectorizedLoadInst(uint64_t NewBeginOffset,
2088 uint64_t NewEndOffset) {
2089 unsigned BeginIndex = getIndex(NewBeginOffset);
2090 unsigned EndIndex = getIndex(NewEndOffset);
2091 assert(EndIndex > BeginIndex && "Empty vector!");
2093 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2095 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2098 Value *rewriteIntegerLoad(LoadInst &LI, uint64_t NewBeginOffset,
2099 uint64_t NewEndOffset) {
2100 assert(IntTy && "We cannot insert an integer to the alloca");
2101 assert(!LI.isVolatile());
2102 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2104 V = convertValue(DL, IRB, V, IntTy);
2105 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2106 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2107 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2108 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2113 bool visitLoadInst(LoadInst &LI) {
2114 DEBUG(dbgs() << " original: " << LI << "\n");
2115 Value *OldOp = LI.getOperand(0);
2116 assert(OldOp == OldPtr);
2118 // Compute the intersecting offset range.
2119 assert(BeginOffset < NewAllocaEndOffset);
2120 assert(EndOffset > NewAllocaBeginOffset);
2121 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2122 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2124 uint64_t Size = NewEndOffset - NewBeginOffset;
2126 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), Size * 8)
2128 bool IsPtrAdjusted = false;
2131 V = rewriteVectorizedLoadInst(NewBeginOffset, NewEndOffset);
2132 } else if (IntTy && LI.getType()->isIntegerTy()) {
2133 V = rewriteIntegerLoad(LI, NewBeginOffset, NewEndOffset);
2134 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2135 canConvertValue(DL, NewAllocaTy, LI.getType())) {
2136 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2137 LI.isVolatile(), "load");
2139 Type *LTy = TargetTy->getPointerTo();
2140 V = IRB.CreateAlignedLoad(
2141 getAdjustedAllocaPtr(IRB, NewBeginOffset, LTy),
2142 getOffsetTypeAlign(TargetTy, NewBeginOffset - NewAllocaBeginOffset),
2143 LI.isVolatile(), "load");
2144 IsPtrAdjusted = true;
2146 V = convertValue(DL, IRB, V, TargetTy);
2149 assert(!LI.isVolatile());
2150 assert(LI.getType()->isIntegerTy() &&
2151 "Only integer type loads and stores are split");
2152 assert(Size < DL.getTypeStoreSize(LI.getType()) &&
2153 "Split load isn't smaller than original load");
2154 assert(LI.getType()->getIntegerBitWidth() ==
2155 DL.getTypeStoreSizeInBits(LI.getType()) &&
2156 "Non-byte-multiple bit width");
2157 // Move the insertion point just past the load so that we can refer to it.
2158 IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI)));
2159 // Create a placeholder value with the same type as LI to use as the
2160 // basis for the new value. This allows us to replace the uses of LI with
2161 // the computed value, and then replace the placeholder with LI, leaving
2162 // LI only used for this computation.
2164 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2165 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset,
2167 LI.replaceAllUsesWith(V);
2168 Placeholder->replaceAllUsesWith(&LI);
2171 LI.replaceAllUsesWith(V);
2174 Pass.DeadInsts.insert(&LI);
2175 deleteIfTriviallyDead(OldOp);
2176 DEBUG(dbgs() << " to: " << *V << "\n");
2177 return !LI.isVolatile() && !IsPtrAdjusted;
2180 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp,
2181 uint64_t NewBeginOffset,
2182 uint64_t NewEndOffset) {
2183 if (V->getType() != VecTy) {
2184 unsigned BeginIndex = getIndex(NewBeginOffset);
2185 unsigned EndIndex = getIndex(NewEndOffset);
2186 assert(EndIndex > BeginIndex && "Empty vector!");
2187 unsigned NumElements = EndIndex - BeginIndex;
2188 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2190 (NumElements == 1) ? ElementTy
2191 : VectorType::get(ElementTy, NumElements);
2192 if (V->getType() != SliceTy)
2193 V = convertValue(DL, IRB, V, SliceTy);
2195 // Mix in the existing elements.
2196 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2198 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2200 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2201 Pass.DeadInsts.insert(&SI);
2204 DEBUG(dbgs() << " to: " << *Store << "\n");
2208 bool rewriteIntegerStore(Value *V, StoreInst &SI,
2209 uint64_t NewBeginOffset, uint64_t NewEndOffset) {
2210 assert(IntTy && "We cannot extract an integer from the alloca");
2211 assert(!SI.isVolatile());
2212 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2213 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2215 Old = convertValue(DL, IRB, Old, IntTy);
2216 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2217 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2218 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset,
2221 V = convertValue(DL, IRB, V, NewAllocaTy);
2222 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2223 Pass.DeadInsts.insert(&SI);
2225 DEBUG(dbgs() << " to: " << *Store << "\n");
2229 bool visitStoreInst(StoreInst &SI) {
2230 DEBUG(dbgs() << " original: " << SI << "\n");
2231 Value *OldOp = SI.getOperand(1);
2232 assert(OldOp == OldPtr);
2234 Value *V = SI.getValueOperand();
2236 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2237 // alloca that should be re-examined after promoting this alloca.
2238 if (V->getType()->isPointerTy())
2239 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2240 Pass.PostPromotionWorklist.insert(AI);
2242 // Compute the intersecting offset range.
2243 assert(BeginOffset < NewAllocaEndOffset);
2244 assert(EndOffset > NewAllocaBeginOffset);
2245 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2246 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2248 uint64_t Size = NewEndOffset - NewBeginOffset;
2249 if (Size < DL.getTypeStoreSize(V->getType())) {
2250 assert(!SI.isVolatile());
2251 assert(V->getType()->isIntegerTy() &&
2252 "Only integer type loads and stores are split");
2253 assert(V->getType()->getIntegerBitWidth() ==
2254 DL.getTypeStoreSizeInBits(V->getType()) &&
2255 "Non-byte-multiple bit width");
2256 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8);
2257 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset,
2262 return rewriteVectorizedStoreInst(V, SI, OldOp, NewBeginOffset,
2264 if (IntTy && V->getType()->isIntegerTy())
2265 return rewriteIntegerStore(V, SI, NewBeginOffset, NewEndOffset);
2268 if (NewBeginOffset == NewAllocaBeginOffset &&
2269 NewEndOffset == NewAllocaEndOffset &&
2270 canConvertValue(DL, V->getType(), NewAllocaTy)) {
2271 V = convertValue(DL, IRB, V, NewAllocaTy);
2272 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2275 Value *NewPtr = getAdjustedAllocaPtr(IRB, NewBeginOffset,
2276 V->getType()->getPointerTo());
2277 NewSI = IRB.CreateAlignedStore(
2278 V, NewPtr, getOffsetTypeAlign(
2279 V->getType(), NewBeginOffset - NewAllocaBeginOffset),
2283 Pass.DeadInsts.insert(&SI);
2284 deleteIfTriviallyDead(OldOp);
2286 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2287 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2290 /// \brief Compute an integer value from splatting an i8 across the given
2291 /// number of bytes.
2293 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2294 /// call this routine.
2295 /// FIXME: Heed the advice above.
2297 /// \param V The i8 value to splat.
2298 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2299 Value *getIntegerSplat(Value *V, unsigned Size) {
2300 assert(Size > 0 && "Expected a positive number of bytes.");
2301 IntegerType *VTy = cast<IntegerType>(V->getType());
2302 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2306 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2307 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, "zext"),
2308 ConstantExpr::getUDiv(
2309 Constant::getAllOnesValue(SplatIntTy),
2310 ConstantExpr::getZExt(
2311 Constant::getAllOnesValue(V->getType()),
2317 /// \brief Compute a vector splat for a given element value.
2318 Value *getVectorSplat(Value *V, unsigned NumElements) {
2319 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2320 DEBUG(dbgs() << " splat: " << *V << "\n");
2324 bool visitMemSetInst(MemSetInst &II) {
2325 DEBUG(dbgs() << " original: " << II << "\n");
2326 assert(II.getRawDest() == OldPtr);
2328 // If the memset has a variable size, it cannot be split, just adjust the
2329 // pointer to the new alloca.
2330 if (!isa<Constant>(II.getLength())) {
2332 assert(BeginOffset >= NewAllocaBeginOffset);
2334 getAdjustedAllocaPtr(IRB, BeginOffset, II.getRawDest()->getType()));
2335 Type *CstTy = II.getAlignmentCst()->getType();
2336 II.setAlignment(ConstantInt::get(CstTy, getOffsetAlign(BeginOffset)));
2338 deleteIfTriviallyDead(OldPtr);
2342 // Record this instruction for deletion.
2343 Pass.DeadInsts.insert(&II);
2345 Type *AllocaTy = NewAI.getAllocatedType();
2346 Type *ScalarTy = AllocaTy->getScalarType();
2348 // Compute the intersecting offset range.
2349 assert(BeginOffset < NewAllocaEndOffset);
2350 assert(EndOffset > NewAllocaBeginOffset);
2351 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2352 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2353 uint64_t SliceOffset = NewBeginOffset - NewAllocaBeginOffset;
2355 // If this doesn't map cleanly onto the alloca type, and that type isn't
2356 // a single value type, just emit a memset.
2357 if (!VecTy && !IntTy &&
2358 (BeginOffset > NewAllocaBeginOffset ||
2359 EndOffset < NewAllocaEndOffset ||
2360 !AllocaTy->isSingleValueType() ||
2361 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2362 DL.getTypeSizeInBits(ScalarTy)%8 != 0)) {
2363 Type *SizeTy = II.getLength()->getType();
2364 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2365 CallInst *New = IRB.CreateMemSet(
2366 getAdjustedAllocaPtr(IRB, NewBeginOffset, II.getRawDest()->getType()),
2367 II.getValue(), Size, getOffsetAlign(SliceOffset), II.isVolatile());
2369 DEBUG(dbgs() << " to: " << *New << "\n");
2373 // If we can represent this as a simple value, we have to build the actual
2374 // value to store, which requires expanding the byte present in memset to
2375 // a sensible representation for the alloca type. This is essentially
2376 // splatting the byte to a sufficiently wide integer, splatting it across
2377 // any desired vector width, and bitcasting to the final type.
2381 // If this is a memset of a vectorized alloca, insert it.
2382 assert(ElementTy == ScalarTy);
2384 unsigned BeginIndex = getIndex(NewBeginOffset);
2385 unsigned EndIndex = getIndex(NewEndOffset);
2386 assert(EndIndex > BeginIndex && "Empty vector!");
2387 unsigned NumElements = EndIndex - BeginIndex;
2388 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2391 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2392 Splat = convertValue(DL, IRB, Splat, ElementTy);
2393 if (NumElements > 1)
2394 Splat = getVectorSplat(Splat, NumElements);
2396 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2398 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2400 // If this is a memset on an alloca where we can widen stores, insert the
2402 assert(!II.isVolatile());
2404 uint64_t Size = NewEndOffset - NewBeginOffset;
2405 V = getIntegerSplat(II.getValue(), Size);
2407 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2408 EndOffset != NewAllocaBeginOffset)) {
2409 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2411 Old = convertValue(DL, IRB, Old, IntTy);
2412 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2413 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2415 assert(V->getType() == IntTy &&
2416 "Wrong type for an alloca wide integer!");
2418 V = convertValue(DL, IRB, V, AllocaTy);
2420 // Established these invariants above.
2421 assert(NewBeginOffset == NewAllocaBeginOffset);
2422 assert(NewEndOffset == NewAllocaEndOffset);
2424 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2425 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2426 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2428 V = convertValue(DL, IRB, V, AllocaTy);
2431 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2434 DEBUG(dbgs() << " to: " << *New << "\n");
2435 return !II.isVolatile();
2438 bool visitMemTransferInst(MemTransferInst &II) {
2439 // Rewriting of memory transfer instructions can be a bit tricky. We break
2440 // them into two categories: split intrinsics and unsplit intrinsics.
2442 DEBUG(dbgs() << " original: " << II << "\n");
2444 // Compute the intersecting offset range.
2445 assert(BeginOffset < NewAllocaEndOffset);
2446 assert(EndOffset > NewAllocaBeginOffset);
2447 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2448 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2450 bool IsDest = &II.getRawDestUse() == OldUse;
2451 assert((IsDest && II.getRawDest() == OldPtr) ||
2452 (!IsDest && II.getRawSource() == OldPtr));
2454 // Compute the relative offset within the transfer.
2455 unsigned IntPtrWidth = DL.getPointerSizeInBits();
2456 APInt RelOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2458 unsigned Align = II.getAlignment();
2459 uint64_t SliceOffset = NewBeginOffset - NewAllocaBeginOffset;
2462 MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2463 MinAlign(II.getAlignment(), getOffsetAlign(SliceOffset)));
2465 // For unsplit intrinsics, we simply modify the source and destination
2466 // pointers in place. This isn't just an optimization, it is a matter of
2467 // correctness. With unsplit intrinsics we may be dealing with transfers
2468 // within a single alloca before SROA ran, or with transfers that have
2469 // a variable length. We may also be dealing with memmove instead of
2470 // memcpy, and so simply updating the pointers is the necessary for us to
2471 // update both source and dest of a single call.
2472 if (!IsSplittable) {
2473 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2476 getAdjustedAllocaPtr(IRB, BeginOffset, II.getRawDest()->getType()));
2478 II.setSource(getAdjustedAllocaPtr(IRB, BeginOffset,
2479 II.getRawSource()->getType()));
2481 Type *CstTy = II.getAlignmentCst()->getType();
2482 II.setAlignment(ConstantInt::get(CstTy, Align));
2484 DEBUG(dbgs() << " to: " << II << "\n");
2485 deleteIfTriviallyDead(OldOp);
2488 // For split transfer intrinsics we have an incredibly useful assurance:
2489 // the source and destination do not reside within the same alloca, and at
2490 // least one of them does not escape. This means that we can replace
2491 // memmove with memcpy, and we don't need to worry about all manner of
2492 // downsides to splitting and transforming the operations.
2494 // If this doesn't map cleanly onto the alloca type, and that type isn't
2495 // a single value type, just emit a memcpy.
2497 = !VecTy && !IntTy && (BeginOffset > NewAllocaBeginOffset ||
2498 EndOffset < NewAllocaEndOffset ||
2499 !NewAI.getAllocatedType()->isSingleValueType());
2501 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2502 // size hasn't been shrunk based on analysis of the viable range, this is
2504 if (EmitMemCpy && &OldAI == &NewAI) {
2505 // Ensure the start lines up.
2506 assert(NewBeginOffset == BeginOffset);
2508 // Rewrite the size as needed.
2509 if (NewEndOffset != EndOffset)
2510 II.setLength(ConstantInt::get(II.getLength()->getType(),
2511 NewEndOffset - NewBeginOffset));
2514 // Record this instruction for deletion.
2515 Pass.DeadInsts.insert(&II);
2517 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2518 // alloca that should be re-examined after rewriting this instruction.
2519 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2521 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2522 assert(AI != &OldAI && AI != &NewAI &&
2523 "Splittable transfers cannot reach the same alloca on both ends.");
2524 Pass.Worklist.insert(AI);
2528 Type *OtherPtrTy = OtherPtr->getType();
2530 // Compute the other pointer, folding as much as possible to produce
2531 // a single, simple GEP in most cases.
2532 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy);
2534 Value *OurPtr = getAdjustedAllocaPtr(
2535 IRB, NewBeginOffset,
2536 IsDest ? II.getRawDest()->getType() : II.getRawSource()->getType());
2537 Type *SizeTy = II.getLength()->getType();
2538 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2540 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2541 IsDest ? OtherPtr : OurPtr,
2542 Size, Align, II.isVolatile());
2544 DEBUG(dbgs() << " to: " << *New << "\n");
2548 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2549 // is equivalent to 1, but that isn't true if we end up rewriting this as
2554 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2555 NewEndOffset == NewAllocaEndOffset;
2556 uint64_t Size = NewEndOffset - NewBeginOffset;
2557 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2558 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2559 unsigned NumElements = EndIndex - BeginIndex;
2560 IntegerType *SubIntTy
2561 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0;
2563 Type *OtherPtrTy = NewAI.getType();
2564 if (VecTy && !IsWholeAlloca) {
2565 if (NumElements == 1)
2566 OtherPtrTy = VecTy->getElementType();
2568 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2570 OtherPtrTy = OtherPtrTy->getPointerTo();
2571 } else if (IntTy && !IsWholeAlloca) {
2572 OtherPtrTy = SubIntTy->getPointerTo();
2575 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy);
2576 Value *DstPtr = &NewAI;
2578 std::swap(SrcPtr, DstPtr);
2581 if (VecTy && !IsWholeAlloca && !IsDest) {
2582 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2584 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2585 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2586 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2588 Src = convertValue(DL, IRB, Src, IntTy);
2589 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2590 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2592 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2596 if (VecTy && !IsWholeAlloca && IsDest) {
2597 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2599 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2600 } else if (IntTy && !IsWholeAlloca && IsDest) {
2601 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2603 Old = convertValue(DL, IRB, Old, IntTy);
2604 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2605 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2606 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2609 StoreInst *Store = cast<StoreInst>(
2610 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2612 DEBUG(dbgs() << " to: " << *Store << "\n");
2613 return !II.isVolatile();
2616 bool visitIntrinsicInst(IntrinsicInst &II) {
2617 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2618 II.getIntrinsicID() == Intrinsic::lifetime_end);
2619 DEBUG(dbgs() << " original: " << II << "\n");
2620 assert(II.getArgOperand(1) == OldPtr);
2622 // Compute the intersecting offset range.
2623 assert(BeginOffset < NewAllocaEndOffset);
2624 assert(EndOffset > NewAllocaBeginOffset);
2625 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2626 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2628 // Record this instruction for deletion.
2629 Pass.DeadInsts.insert(&II);
2632 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2633 NewEndOffset - NewBeginOffset);
2635 getAdjustedAllocaPtr(IRB, NewBeginOffset, II.getArgOperand(1)->getType());
2637 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2638 New = IRB.CreateLifetimeStart(Ptr, Size);
2640 New = IRB.CreateLifetimeEnd(Ptr, Size);
2643 DEBUG(dbgs() << " to: " << *New << "\n");
2647 bool visitPHINode(PHINode &PN) {
2648 DEBUG(dbgs() << " original: " << PN << "\n");
2649 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2650 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2652 // We would like to compute a new pointer in only one place, but have it be
2653 // as local as possible to the PHI. To do that, we re-use the location of
2654 // the old pointer, which necessarily must be in the right position to
2655 // dominate the PHI.
2656 IRBuilderTy PtrBuilder(OldPtr);
2657 PtrBuilder.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) +
2661 getAdjustedAllocaPtr(PtrBuilder, BeginOffset, OldPtr->getType());
2662 // Replace the operands which were using the old pointer.
2663 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2665 DEBUG(dbgs() << " to: " << PN << "\n");
2666 deleteIfTriviallyDead(OldPtr);
2668 // PHIs can't be promoted on their own, but often can be speculated. We
2669 // check the speculation outside of the rewriter so that we see the
2670 // fully-rewritten alloca.
2671 PHIUsers.insert(&PN);
2675 bool visitSelectInst(SelectInst &SI) {
2676 DEBUG(dbgs() << " original: " << SI << "\n");
2677 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2678 "Pointer isn't an operand!");
2679 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2680 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2682 Value *NewPtr = getAdjustedAllocaPtr(IRB, BeginOffset, OldPtr->getType());
2683 // Replace the operands which were using the old pointer.
2684 if (SI.getOperand(1) == OldPtr)
2685 SI.setOperand(1, NewPtr);
2686 if (SI.getOperand(2) == OldPtr)
2687 SI.setOperand(2, NewPtr);
2689 DEBUG(dbgs() << " to: " << SI << "\n");
2690 deleteIfTriviallyDead(OldPtr);
2692 // Selects can't be promoted on their own, but often can be speculated. We
2693 // check the speculation outside of the rewriter so that we see the
2694 // fully-rewritten alloca.
2695 SelectUsers.insert(&SI);
2703 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2705 /// This pass aggressively rewrites all aggregate loads and stores on
2706 /// a particular pointer (or any pointer derived from it which we can identify)
2707 /// with scalar loads and stores.
2708 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2709 // Befriend the base class so it can delegate to private visit methods.
2710 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2712 const DataLayout &DL;
2714 /// Queue of pointer uses to analyze and potentially rewrite.
2715 SmallVector<Use *, 8> Queue;
2717 /// Set to prevent us from cycling with phi nodes and loops.
2718 SmallPtrSet<User *, 8> Visited;
2720 /// The current pointer use being rewritten. This is used to dig up the used
2721 /// value (as opposed to the user).
2725 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
2727 /// Rewrite loads and stores through a pointer and all pointers derived from
2729 bool rewrite(Instruction &I) {
2730 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2732 bool Changed = false;
2733 while (!Queue.empty()) {
2734 U = Queue.pop_back_val();
2735 Changed |= visit(cast<Instruction>(U->getUser()));
2741 /// Enqueue all the users of the given instruction for further processing.
2742 /// This uses a set to de-duplicate users.
2743 void enqueueUsers(Instruction &I) {
2744 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
2746 if (Visited.insert(*UI))
2747 Queue.push_back(&UI.getUse());
2750 // Conservative default is to not rewrite anything.
2751 bool visitInstruction(Instruction &I) { return false; }
2753 /// \brief Generic recursive split emission class.
2754 template <typename Derived>
2757 /// The builder used to form new instructions.
2759 /// The indices which to be used with insert- or extractvalue to select the
2760 /// appropriate value within the aggregate.
2761 SmallVector<unsigned, 4> Indices;
2762 /// The indices to a GEP instruction which will move Ptr to the correct slot
2763 /// within the aggregate.
2764 SmallVector<Value *, 4> GEPIndices;
2765 /// The base pointer of the original op, used as a base for GEPing the
2766 /// split operations.
2769 /// Initialize the splitter with an insertion point, Ptr and start with a
2770 /// single zero GEP index.
2771 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2772 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2775 /// \brief Generic recursive split emission routine.
2777 /// This method recursively splits an aggregate op (load or store) into
2778 /// scalar or vector ops. It splits recursively until it hits a single value
2779 /// and emits that single value operation via the template argument.
2781 /// The logic of this routine relies on GEPs and insertvalue and
2782 /// extractvalue all operating with the same fundamental index list, merely
2783 /// formatted differently (GEPs need actual values).
2785 /// \param Ty The type being split recursively into smaller ops.
2786 /// \param Agg The aggregate value being built up or stored, depending on
2787 /// whether this is splitting a load or a store respectively.
2788 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2789 if (Ty->isSingleValueType())
2790 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2792 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2793 unsigned OldSize = Indices.size();
2795 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2797 assert(Indices.size() == OldSize && "Did not return to the old size");
2798 Indices.push_back(Idx);
2799 GEPIndices.push_back(IRB.getInt32(Idx));
2800 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2801 GEPIndices.pop_back();
2807 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2808 unsigned OldSize = Indices.size();
2810 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2812 assert(Indices.size() == OldSize && "Did not return to the old size");
2813 Indices.push_back(Idx);
2814 GEPIndices.push_back(IRB.getInt32(Idx));
2815 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2816 GEPIndices.pop_back();
2822 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2826 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2827 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2828 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2830 /// Emit a leaf load of a single value. This is called at the leaves of the
2831 /// recursive emission to actually load values.
2832 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2833 assert(Ty->isSingleValueType());
2834 // Load the single value and insert it using the indices.
2835 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
2836 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
2837 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2838 DEBUG(dbgs() << " to: " << *Load << "\n");
2842 bool visitLoadInst(LoadInst &LI) {
2843 assert(LI.getPointerOperand() == *U);
2844 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2847 // We have an aggregate being loaded, split it apart.
2848 DEBUG(dbgs() << " original: " << LI << "\n");
2849 LoadOpSplitter Splitter(&LI, *U);
2850 Value *V = UndefValue::get(LI.getType());
2851 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2852 LI.replaceAllUsesWith(V);
2853 LI.eraseFromParent();
2857 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2858 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2859 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2861 /// Emit a leaf store of a single value. This is called at the leaves of the
2862 /// recursive emission to actually produce stores.
2863 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2864 assert(Ty->isSingleValueType());
2865 // Extract the single value and store it using the indices.
2866 Value *Store = IRB.CreateStore(
2867 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2868 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2870 DEBUG(dbgs() << " to: " << *Store << "\n");
2874 bool visitStoreInst(StoreInst &SI) {
2875 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2877 Value *V = SI.getValueOperand();
2878 if (V->getType()->isSingleValueType())
2881 // We have an aggregate being stored, split it apart.
2882 DEBUG(dbgs() << " original: " << SI << "\n");
2883 StoreOpSplitter Splitter(&SI, *U);
2884 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2885 SI.eraseFromParent();
2889 bool visitBitCastInst(BitCastInst &BC) {
2894 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2899 bool visitPHINode(PHINode &PN) {
2904 bool visitSelectInst(SelectInst &SI) {
2911 /// \brief Strip aggregate type wrapping.
2913 /// This removes no-op aggregate types wrapping an underlying type. It will
2914 /// strip as many layers of types as it can without changing either the type
2915 /// size or the allocated size.
2916 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
2917 if (Ty->isSingleValueType())
2920 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
2921 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
2924 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
2925 InnerTy = ArrTy->getElementType();
2926 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
2927 const StructLayout *SL = DL.getStructLayout(STy);
2928 unsigned Index = SL->getElementContainingOffset(0);
2929 InnerTy = STy->getElementType(Index);
2934 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
2935 TypeSize > DL.getTypeSizeInBits(InnerTy))
2938 return stripAggregateTypeWrapping(DL, InnerTy);
2941 /// \brief Try to find a partition of the aggregate type passed in for a given
2942 /// offset and size.
2944 /// This recurses through the aggregate type and tries to compute a subtype
2945 /// based on the offset and size. When the offset and size span a sub-section
2946 /// of an array, it will even compute a new array type for that sub-section,
2947 /// and the same for structs.
2949 /// Note that this routine is very strict and tries to find a partition of the
2950 /// type which produces the *exact* right offset and size. It is not forgiving
2951 /// when the size or offset cause either end of type-based partition to be off.
2952 /// Also, this is a best-effort routine. It is reasonable to give up and not
2953 /// return a type if necessary.
2954 static Type *getTypePartition(const DataLayout &DL, Type *Ty,
2955 uint64_t Offset, uint64_t Size) {
2956 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
2957 return stripAggregateTypeWrapping(DL, Ty);
2958 if (Offset > DL.getTypeAllocSize(Ty) ||
2959 (DL.getTypeAllocSize(Ty) - Offset) < Size)
2962 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
2963 // We can't partition pointers...
2964 if (SeqTy->isPointerTy())
2967 Type *ElementTy = SeqTy->getElementType();
2968 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
2969 uint64_t NumSkippedElements = Offset / ElementSize;
2970 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
2971 if (NumSkippedElements >= ArrTy->getNumElements())
2973 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
2974 if (NumSkippedElements >= VecTy->getNumElements())
2977 Offset -= NumSkippedElements * ElementSize;
2979 // First check if we need to recurse.
2980 if (Offset > 0 || Size < ElementSize) {
2981 // Bail if the partition ends in a different array element.
2982 if ((Offset + Size) > ElementSize)
2984 // Recurse through the element type trying to peel off offset bytes.
2985 return getTypePartition(DL, ElementTy, Offset, Size);
2987 assert(Offset == 0);
2989 if (Size == ElementSize)
2990 return stripAggregateTypeWrapping(DL, ElementTy);
2991 assert(Size > ElementSize);
2992 uint64_t NumElements = Size / ElementSize;
2993 if (NumElements * ElementSize != Size)
2995 return ArrayType::get(ElementTy, NumElements);
2998 StructType *STy = dyn_cast<StructType>(Ty);
3002 const StructLayout *SL = DL.getStructLayout(STy);
3003 if (Offset >= SL->getSizeInBytes())
3005 uint64_t EndOffset = Offset + Size;
3006 if (EndOffset > SL->getSizeInBytes())
3009 unsigned Index = SL->getElementContainingOffset(Offset);
3010 Offset -= SL->getElementOffset(Index);
3012 Type *ElementTy = STy->getElementType(Index);
3013 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3014 if (Offset >= ElementSize)
3015 return 0; // The offset points into alignment padding.
3017 // See if any partition must be contained by the element.
3018 if (Offset > 0 || Size < ElementSize) {
3019 if ((Offset + Size) > ElementSize)
3021 return getTypePartition(DL, ElementTy, Offset, Size);
3023 assert(Offset == 0);
3025 if (Size == ElementSize)
3026 return stripAggregateTypeWrapping(DL, ElementTy);
3028 StructType::element_iterator EI = STy->element_begin() + Index,
3029 EE = STy->element_end();
3030 if (EndOffset < SL->getSizeInBytes()) {
3031 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3032 if (Index == EndIndex)
3033 return 0; // Within a single element and its padding.
3035 // Don't try to form "natural" types if the elements don't line up with the
3037 // FIXME: We could potentially recurse down through the last element in the
3038 // sub-struct to find a natural end point.
3039 if (SL->getElementOffset(EndIndex) != EndOffset)
3042 assert(Index < EndIndex);
3043 EE = STy->element_begin() + EndIndex;
3046 // Try to build up a sub-structure.
3047 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
3049 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3050 if (Size != SubSL->getSizeInBytes())
3051 return 0; // The sub-struct doesn't have quite the size needed.
3056 /// \brief Rewrite an alloca partition's users.
3058 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3059 /// to rewrite uses of an alloca partition to be conducive for SSA value
3060 /// promotion. If the partition needs a new, more refined alloca, this will
3061 /// build that new alloca, preserving as much type information as possible, and
3062 /// rewrite the uses of the old alloca to point at the new one and have the
3063 /// appropriate new offsets. It also evaluates how successful the rewrite was
3064 /// at enabling promotion and if it was successful queues the alloca to be
3066 bool SROA::rewritePartition(AllocaInst &AI, AllocaSlices &S,
3067 AllocaSlices::iterator B, AllocaSlices::iterator E,
3068 int64_t BeginOffset, int64_t EndOffset,
3069 ArrayRef<AllocaSlices::iterator> SplitUses) {
3070 assert(BeginOffset < EndOffset);
3071 uint64_t SliceSize = EndOffset - BeginOffset;
3073 // Try to compute a friendly type for this partition of the alloca. This
3074 // won't always succeed, in which case we fall back to a legal integer type
3075 // or an i8 array of an appropriate size.
3077 if (Type *CommonUseTy = findCommonType(B, E, EndOffset))
3078 if (DL->getTypeAllocSize(CommonUseTy) >= SliceSize)
3079 SliceTy = CommonUseTy;
3081 if (Type *TypePartitionTy = getTypePartition(*DL, AI.getAllocatedType(),
3082 BeginOffset, SliceSize))
3083 SliceTy = TypePartitionTy;
3084 if ((!SliceTy || (SliceTy->isArrayTy() &&
3085 SliceTy->getArrayElementType()->isIntegerTy())) &&
3086 DL->isLegalInteger(SliceSize * 8))
3087 SliceTy = Type::getIntNTy(*C, SliceSize * 8);
3089 SliceTy = ArrayType::get(Type::getInt8Ty(*C), SliceSize);
3090 assert(DL->getTypeAllocSize(SliceTy) >= SliceSize);
3092 bool IsVectorPromotable = isVectorPromotionViable(
3093 *DL, SliceTy, S, BeginOffset, EndOffset, B, E, SplitUses);
3095 bool IsIntegerPromotable =
3096 !IsVectorPromotable &&
3097 isIntegerWideningViable(*DL, SliceTy, BeginOffset, S, B, E, SplitUses);
3099 // Check for the case where we're going to rewrite to a new alloca of the
3100 // exact same type as the original, and with the same access offsets. In that
3101 // case, re-use the existing alloca, but still run through the rewriter to
3102 // perform phi and select speculation.
3104 if (SliceTy == AI.getAllocatedType()) {
3105 assert(BeginOffset == 0 &&
3106 "Non-zero begin offset but same alloca type");
3108 // FIXME: We should be able to bail at this point with "nothing changed".
3109 // FIXME: We might want to defer PHI speculation until after here.
3111 unsigned Alignment = AI.getAlignment();
3113 // The minimum alignment which users can rely on when the explicit
3114 // alignment is omitted or zero is that required by the ABI for this
3116 Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
3118 Alignment = MinAlign(Alignment, BeginOffset);
3119 // If we will get at least this much alignment from the type alone, leave
3120 // the alloca's alignment unconstrained.
3121 if (Alignment <= DL->getABITypeAlignment(SliceTy))
3123 NewAI = new AllocaInst(SliceTy, 0, Alignment,
3124 AI.getName() + ".sroa." + Twine(B - S.begin()), &AI);
3128 DEBUG(dbgs() << "Rewriting alloca partition "
3129 << "[" << BeginOffset << "," << EndOffset << ") to: " << *NewAI
3132 // Track the high watermark on the worklist as it is only relevant for
3133 // promoted allocas. We will reset it to this point if the alloca is not in
3134 // fact scheduled for promotion.
3135 unsigned PPWOldSize = PostPromotionWorklist.size();
3136 unsigned NumUses = 0;
3137 SmallPtrSet<PHINode *, 8> PHIUsers;
3138 SmallPtrSet<SelectInst *, 8> SelectUsers;
3140 AllocaSliceRewriter Rewriter(*DL, S, *this, AI, *NewAI, BeginOffset,
3141 EndOffset, IsVectorPromotable,
3142 IsIntegerPromotable, PHIUsers, SelectUsers);
3143 bool Promotable = true;
3144 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
3145 SUE = SplitUses.end();
3146 SUI != SUE; ++SUI) {
3147 DEBUG(dbgs() << " rewriting split ");
3148 DEBUG(S.printSlice(dbgs(), *SUI, ""));
3149 Promotable &= Rewriter.visit(*SUI);
3152 for (AllocaSlices::iterator I = B; I != E; ++I) {
3153 DEBUG(dbgs() << " rewriting ");
3154 DEBUG(S.printSlice(dbgs(), I, ""));
3155 Promotable &= Rewriter.visit(I);
3159 NumAllocaPartitionUses += NumUses;
3160 MaxUsesPerAllocaPartition =
3161 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3163 // Now that we've processed all the slices in the new partition, check if any
3164 // PHIs or Selects would block promotion.
3165 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3168 if (!isSafePHIToSpeculate(**I, DL)) {
3171 SelectUsers.clear();
3173 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3174 E = SelectUsers.end();
3176 if (!isSafeSelectToSpeculate(**I, DL)) {
3179 SelectUsers.clear();
3183 if (PHIUsers.empty() && SelectUsers.empty()) {
3184 // Promote the alloca.
3185 PromotableAllocas.push_back(NewAI);
3187 // If we have either PHIs or Selects to speculate, add them to those
3188 // worklists and re-queue the new alloca so that we promote in on the
3190 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3193 SpeculatablePHIs.insert(*I);
3194 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3195 E = SelectUsers.end();
3197 SpeculatableSelects.insert(*I);
3198 Worklist.insert(NewAI);
3201 // If we can't promote the alloca, iterate on it to check for new
3202 // refinements exposed by splitting the current alloca. Don't iterate on an
3203 // alloca which didn't actually change and didn't get promoted.
3205 Worklist.insert(NewAI);
3207 // Drop any post-promotion work items if promotion didn't happen.
3208 while (PostPromotionWorklist.size() > PPWOldSize)
3209 PostPromotionWorklist.pop_back();
3216 struct IsSliceEndLessOrEqualTo {
3217 uint64_t UpperBound;
3219 IsSliceEndLessOrEqualTo(uint64_t UpperBound) : UpperBound(UpperBound) {}
3221 bool operator()(const AllocaSlices::iterator &I) {
3222 return I->endOffset() <= UpperBound;
3228 removeFinishedSplitUses(SmallVectorImpl<AllocaSlices::iterator> &SplitUses,
3229 uint64_t &MaxSplitUseEndOffset, uint64_t Offset) {
3230 if (Offset >= MaxSplitUseEndOffset) {
3232 MaxSplitUseEndOffset = 0;
3236 size_t SplitUsesOldSize = SplitUses.size();
3237 SplitUses.erase(std::remove_if(SplitUses.begin(), SplitUses.end(),
3238 IsSliceEndLessOrEqualTo(Offset)),
3240 if (SplitUsesOldSize == SplitUses.size())
3243 // Recompute the max. While this is linear, so is remove_if.
3244 MaxSplitUseEndOffset = 0;
3245 for (SmallVectorImpl<AllocaSlices::iterator>::iterator
3246 SUI = SplitUses.begin(),
3247 SUE = SplitUses.end();
3249 MaxSplitUseEndOffset = std::max((*SUI)->endOffset(), MaxSplitUseEndOffset);
3252 /// \brief Walks the slices of an alloca and form partitions based on them,
3253 /// rewriting each of their uses.
3254 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &S) {
3255 if (S.begin() == S.end())
3258 unsigned NumPartitions = 0;
3259 bool Changed = false;
3260 SmallVector<AllocaSlices::iterator, 4> SplitUses;
3261 uint64_t MaxSplitUseEndOffset = 0;
3263 uint64_t BeginOffset = S.begin()->beginOffset();
3265 for (AllocaSlices::iterator SI = S.begin(), SJ = llvm::next(SI), SE = S.end();
3266 SI != SE; SI = SJ) {
3267 uint64_t MaxEndOffset = SI->endOffset();
3269 if (!SI->isSplittable()) {
3270 // When we're forming an unsplittable region, it must always start at the
3271 // first slice and will extend through its end.
3272 assert(BeginOffset == SI->beginOffset());
3274 // Form a partition including all of the overlapping slices with this
3275 // unsplittable slice.
3276 while (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3277 if (!SJ->isSplittable())
3278 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3282 assert(SI->isSplittable()); // Established above.
3284 // Collect all of the overlapping splittable slices.
3285 while (SJ != SE && SJ->beginOffset() < MaxEndOffset &&
3286 SJ->isSplittable()) {
3287 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3291 // Back up MaxEndOffset and SJ if we ended the span early when
3292 // encountering an unsplittable slice.
3293 if (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3294 assert(!SJ->isSplittable());
3295 MaxEndOffset = SJ->beginOffset();
3299 // Check if we have managed to move the end offset forward yet. If so,
3300 // we'll have to rewrite uses and erase old split uses.
3301 if (BeginOffset < MaxEndOffset) {
3302 // Rewrite a sequence of overlapping slices.
3304 rewritePartition(AI, S, SI, SJ, BeginOffset, MaxEndOffset, SplitUses);
3307 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset, MaxEndOffset);
3310 // Accumulate all the splittable slices from the [SI,SJ) region which
3311 // overlap going forward.
3312 for (AllocaSlices::iterator SK = SI; SK != SJ; ++SK)
3313 if (SK->isSplittable() && SK->endOffset() > MaxEndOffset) {
3314 SplitUses.push_back(SK);
3315 MaxSplitUseEndOffset = std::max(SK->endOffset(), MaxSplitUseEndOffset);
3318 // If we're already at the end and we have no split uses, we're done.
3319 if (SJ == SE && SplitUses.empty())
3322 // If we have no split uses or no gap in offsets, we're ready to move to
3324 if (SplitUses.empty() || (SJ != SE && MaxEndOffset == SJ->beginOffset())) {
3325 BeginOffset = SJ->beginOffset();
3329 // Even if we have split slices, if the next slice is splittable and the
3330 // split slices reach it, we can simply set up the beginning offset of the
3331 // next iteration to bridge between them.
3332 if (SJ != SE && SJ->isSplittable() &&
3333 MaxSplitUseEndOffset > SJ->beginOffset()) {
3334 BeginOffset = MaxEndOffset;
3338 // Otherwise, we have a tail of split slices. Rewrite them with an empty
3340 uint64_t PostSplitEndOffset =
3341 SJ == SE ? MaxSplitUseEndOffset : SJ->beginOffset();
3343 Changed |= rewritePartition(AI, S, SJ, SJ, MaxEndOffset, PostSplitEndOffset,
3348 break; // Skip the rest, we don't need to do any cleanup.
3350 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset,
3351 PostSplitEndOffset);
3353 // Now just reset the begin offset for the next iteration.
3354 BeginOffset = SJ->beginOffset();
3357 NumAllocaPartitions += NumPartitions;
3358 MaxPartitionsPerAlloca =
3359 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
3364 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
3365 void SROA::clobberUse(Use &U) {
3367 // Replace the use with an undef value.
3368 U = UndefValue::get(OldV->getType());
3370 // Check for this making an instruction dead. We have to garbage collect
3371 // all the dead instructions to ensure the uses of any alloca end up being
3373 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3374 if (isInstructionTriviallyDead(OldI)) {
3375 DeadInsts.insert(OldI);
3379 /// \brief Analyze an alloca for SROA.
3381 /// This analyzes the alloca to ensure we can reason about it, builds
3382 /// the slices of the alloca, and then hands it off to be split and
3383 /// rewritten as needed.
3384 bool SROA::runOnAlloca(AllocaInst &AI) {
3385 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3386 ++NumAllocasAnalyzed;
3388 // Special case dead allocas, as they're trivial.
3389 if (AI.use_empty()) {
3390 AI.eraseFromParent();
3394 // Skip alloca forms that this analysis can't handle.
3395 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3396 DL->getTypeAllocSize(AI.getAllocatedType()) == 0)
3399 bool Changed = false;
3401 // First, split any FCA loads and stores touching this alloca to promote
3402 // better splitting and promotion opportunities.
3403 AggLoadStoreRewriter AggRewriter(*DL);
3404 Changed |= AggRewriter.rewrite(AI);
3406 // Build the slices using a recursive instruction-visiting builder.
3407 AllocaSlices S(*DL, AI);
3408 DEBUG(S.print(dbgs()));
3412 // Delete all the dead users of this alloca before splitting and rewriting it.
3413 for (AllocaSlices::dead_user_iterator DI = S.dead_user_begin(),
3414 DE = S.dead_user_end();
3416 // Free up everything used by this instruction.
3417 for (User::op_iterator DOI = (*DI)->op_begin(), DOE = (*DI)->op_end();
3421 // Now replace the uses of this instruction.
3422 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3424 // And mark it for deletion.
3425 DeadInsts.insert(*DI);
3428 for (AllocaSlices::dead_op_iterator DO = S.dead_op_begin(),
3429 DE = S.dead_op_end();
3435 // No slices to split. Leave the dead alloca for a later pass to clean up.
3436 if (S.begin() == S.end())
3439 Changed |= splitAlloca(AI, S);
3441 DEBUG(dbgs() << " Speculating PHIs\n");
3442 while (!SpeculatablePHIs.empty())
3443 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
3445 DEBUG(dbgs() << " Speculating Selects\n");
3446 while (!SpeculatableSelects.empty())
3447 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
3452 /// \brief Delete the dead instructions accumulated in this run.
3454 /// Recursively deletes the dead instructions we've accumulated. This is done
3455 /// at the very end to maximize locality of the recursive delete and to
3456 /// minimize the problems of invalidated instruction pointers as such pointers
3457 /// are used heavily in the intermediate stages of the algorithm.
3459 /// We also record the alloca instructions deleted here so that they aren't
3460 /// subsequently handed to mem2reg to promote.
3461 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3462 while (!DeadInsts.empty()) {
3463 Instruction *I = DeadInsts.pop_back_val();
3464 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3466 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3468 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3469 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3470 // Zero out the operand and see if it becomes trivially dead.
3472 if (isInstructionTriviallyDead(U))
3473 DeadInsts.insert(U);
3476 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3477 DeletedAllocas.insert(AI);
3480 I->eraseFromParent();
3484 static void enqueueUsersInWorklist(Instruction &I,
3485 SmallVectorImpl<Instruction *> &Worklist,
3486 SmallPtrSet<Instruction *, 8> &Visited) {
3487 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
3489 if (Visited.insert(cast<Instruction>(*UI)))
3490 Worklist.push_back(cast<Instruction>(*UI));
3493 /// \brief Promote the allocas, using the best available technique.
3495 /// This attempts to promote whatever allocas have been identified as viable in
3496 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3497 /// If there is a domtree available, we attempt to promote using the full power
3498 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3499 /// based on the SSAUpdater utilities. This function returns whether any
3500 /// promotion occurred.
3501 bool SROA::promoteAllocas(Function &F) {
3502 if (PromotableAllocas.empty())
3505 NumPromoted += PromotableAllocas.size();
3507 if (DT && !ForceSSAUpdater) {
3508 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3509 PromoteMemToReg(PromotableAllocas, *DT);
3510 PromotableAllocas.clear();
3514 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3516 DIBuilder DIB(*F.getParent());
3517 SmallVector<Instruction *, 64> Insts;
3519 // We need a worklist to walk the uses of each alloca.
3520 SmallVector<Instruction *, 8> Worklist;
3521 SmallPtrSet<Instruction *, 8> Visited;
3522 SmallVector<Instruction *, 32> DeadInsts;
3524 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3525 AllocaInst *AI = PromotableAllocas[Idx];
3530 enqueueUsersInWorklist(*AI, Worklist, Visited);
3532 while (!Worklist.empty()) {
3533 Instruction *I = Worklist.pop_back_val();
3535 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3536 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3537 // leading to them) here. Eventually it should use them to optimize the
3538 // scalar values produced.
3539 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3540 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3541 II->getIntrinsicID() == Intrinsic::lifetime_end);
3542 II->eraseFromParent();
3546 // Push the loads and stores we find onto the list. SROA will already
3547 // have validated that all loads and stores are viable candidates for
3549 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3550 assert(LI->getType() == AI->getAllocatedType());
3551 Insts.push_back(LI);
3554 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3555 assert(SI->getValueOperand()->getType() == AI->getAllocatedType());
3556 Insts.push_back(SI);
3560 // For everything else, we know that only no-op bitcasts and GEPs will
3561 // make it this far, just recurse through them and recall them for later
3563 DeadInsts.push_back(I);
3564 enqueueUsersInWorklist(*I, Worklist, Visited);
3566 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3567 while (!DeadInsts.empty())
3568 DeadInsts.pop_back_val()->eraseFromParent();
3569 AI->eraseFromParent();
3572 PromotableAllocas.clear();
3577 /// \brief A predicate to test whether an alloca belongs to a set.
3578 class IsAllocaInSet {
3579 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3583 typedef AllocaInst *argument_type;
3585 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3586 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3590 bool SROA::runOnFunction(Function &F) {
3591 if (skipOptnoneFunction(F))
3594 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3595 C = &F.getContext();
3596 DL = getAnalysisIfAvailable<DataLayout>();
3598 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3601 DominatorTreeWrapperPass *DTWP =
3602 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
3603 DT = DTWP ? &DTWP->getDomTree() : 0;
3605 BasicBlock &EntryBB = F.getEntryBlock();
3606 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3608 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3609 Worklist.insert(AI);
3611 bool Changed = false;
3612 // A set of deleted alloca instruction pointers which should be removed from
3613 // the list of promotable allocas.
3614 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3617 while (!Worklist.empty()) {
3618 Changed |= runOnAlloca(*Worklist.pop_back_val());
3619 deleteDeadInstructions(DeletedAllocas);
3621 // Remove the deleted allocas from various lists so that we don't try to
3622 // continue processing them.
3623 if (!DeletedAllocas.empty()) {
3624 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3625 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3626 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3627 PromotableAllocas.end(),
3628 IsAllocaInSet(DeletedAllocas)),
3629 PromotableAllocas.end());
3630 DeletedAllocas.clear();
3634 Changed |= promoteAllocas(F);
3636 Worklist = PostPromotionWorklist;
3637 PostPromotionWorklist.clear();
3638 } while (!Worklist.empty());
3643 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3644 if (RequiresDomTree)
3645 AU.addRequired<DominatorTreeWrapperPass>();
3646 AU.setPreservesCFG();