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/Dominators.h"
33 #include "llvm/Analysis/Loads.h"
34 #include "llvm/Analysis/PtrUseVisitor.h"
35 #include "llvm/Analysis/ValueTracking.h"
36 #include "llvm/DIBuilder.h"
37 #include "llvm/DebugInfo.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DerivedTypes.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/raw_ostream.h"
55 #include "llvm/Transforms/Utils/Local.h"
56 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
57 #include "llvm/Transforms/Utils/SSAUpdater.h"
60 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
61 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
62 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
63 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
64 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
65 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
66 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
67 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
68 STATISTIC(NumDeleted, "Number of instructions deleted");
69 STATISTIC(NumVectorized, "Number of vectorized aggregates");
71 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
72 /// forming SSA values through the SSAUpdater infrastructure.
74 ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
77 /// \brief A custom IRBuilder inserter which prefixes all names if they are
79 template <bool preserveNames = true>
80 class IRBuilderPrefixedInserter :
81 public IRBuilderDefaultInserter<preserveNames> {
85 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
88 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
89 BasicBlock::iterator InsertPt) const {
90 IRBuilderDefaultInserter<preserveNames>::InsertHelper(
91 I, Name.isTriviallyEmpty() ? Name : Prefix + Name, BB, InsertPt);
95 // Specialization for not preserving the name is trivial.
97 class IRBuilderPrefixedInserter<false> :
98 public IRBuilderDefaultInserter<false> {
100 void SetNamePrefix(const Twine &P) {}
103 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
105 typedef llvm::IRBuilder<true, ConstantFolder,
106 IRBuilderPrefixedInserter<true> > IRBuilderTy;
108 typedef llvm::IRBuilder<false, ConstantFolder,
109 IRBuilderPrefixedInserter<false> > IRBuilderTy;
114 /// \brief A used slice of an alloca.
116 /// This structure represents a slice of an alloca used by some instruction. It
117 /// stores both the begin and end offsets of this use, a pointer to the use
118 /// itself, and a flag indicating whether we can classify the use as splittable
119 /// or not when forming partitions of the alloca.
121 /// \brief The beginning offset of the range.
122 uint64_t BeginOffset;
124 /// \brief The ending offset, not included in the range.
127 /// \brief Storage for both the use of this slice and whether it can be
129 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
132 Slice() : BeginOffset(), EndOffset() {}
133 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
134 : BeginOffset(BeginOffset), EndOffset(EndOffset),
135 UseAndIsSplittable(U, IsSplittable) {}
137 uint64_t beginOffset() const { return BeginOffset; }
138 uint64_t endOffset() const { return EndOffset; }
140 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
141 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
143 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
145 bool isDead() const { return getUse() == 0; }
146 void kill() { UseAndIsSplittable.setPointer(0); }
148 /// \brief Support for ordering ranges.
150 /// This provides an ordering over ranges such that start offsets are
151 /// always increasing, and within equal start offsets, the end offsets are
152 /// decreasing. Thus the spanning range comes first in a cluster with the
153 /// same start position.
154 bool operator<(const Slice &RHS) const {
155 if (beginOffset() < RHS.beginOffset()) return true;
156 if (beginOffset() > RHS.beginOffset()) return false;
157 if (isSplittable() != RHS.isSplittable()) return !isSplittable();
158 if (endOffset() > RHS.endOffset()) return true;
162 /// \brief Support comparison with a single offset to allow binary searches.
163 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
164 uint64_t RHSOffset) {
165 return LHS.beginOffset() < RHSOffset;
167 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
169 return LHSOffset < RHS.beginOffset();
172 bool operator==(const Slice &RHS) const {
173 return isSplittable() == RHS.isSplittable() &&
174 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
176 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
178 } // end anonymous namespace
181 template <typename T> struct isPodLike;
182 template <> struct isPodLike<Slice> {
183 static const bool value = true;
188 /// \brief Representation of the alloca slices.
190 /// This class represents the slices of an alloca which are formed by its
191 /// various uses. If a pointer escapes, we can't fully build a representation
192 /// for the slices used and we reflect that in this structure. The uses are
193 /// stored, sorted by increasing beginning offset and with unsplittable slices
194 /// starting at a particular offset before splittable slices.
197 /// \brief Construct the slices of a particular alloca.
198 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
200 /// \brief Test whether a pointer to the allocation escapes our analysis.
202 /// If this is true, the slices are never fully built and should be
204 bool isEscaped() const { return PointerEscapingInstr; }
206 /// \brief Support for iterating over the slices.
208 typedef SmallVectorImpl<Slice>::iterator iterator;
209 iterator begin() { return Slices.begin(); }
210 iterator end() { return Slices.end(); }
212 typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
213 const_iterator begin() const { return Slices.begin(); }
214 const_iterator end() const { return Slices.end(); }
217 /// \brief Allow iterating the dead users for this alloca.
219 /// These are instructions which will never actually use the alloca as they
220 /// are outside the allocated range. They are safe to replace with undef and
223 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
224 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
225 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
228 /// \brief Allow iterating the dead expressions referring to this alloca.
230 /// These are operands which have cannot actually be used to refer to the
231 /// alloca as they are outside its range and the user doesn't correct for
232 /// that. These mostly consist of PHI node inputs and the like which we just
233 /// need to replace with undef.
235 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
236 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
237 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
240 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
241 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
242 void printSlice(raw_ostream &OS, const_iterator I,
243 StringRef Indent = " ") const;
244 void printUse(raw_ostream &OS, const_iterator I,
245 StringRef Indent = " ") const;
246 void print(raw_ostream &OS) const;
247 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const;
248 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const;
252 template <typename DerivedT, typename RetT = void> class BuilderBase;
254 friend class AllocaSlices::SliceBuilder;
256 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
257 /// \brief Handle to alloca instruction to simplify method interfaces.
261 /// \brief The instruction responsible for this alloca not having a known set
264 /// When an instruction (potentially) escapes the pointer to the alloca, we
265 /// store a pointer to that here and abort trying to form slices of the
266 /// alloca. This will be null if the alloca slices are analyzed successfully.
267 Instruction *PointerEscapingInstr;
269 /// \brief The slices of the alloca.
271 /// We store a vector of the slices formed by uses of the alloca here. This
272 /// vector is sorted by increasing begin offset, and then the unsplittable
273 /// slices before the splittable ones. See the Slice inner class for more
275 SmallVector<Slice, 8> Slices;
277 /// \brief Instructions which will become dead if we rewrite the alloca.
279 /// Note that these are not separated by slice. This is because we expect an
280 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
281 /// all these instructions can simply be removed and replaced with undef as
282 /// they come from outside of the allocated space.
283 SmallVector<Instruction *, 8> DeadUsers;
285 /// \brief Operands which will become dead if we rewrite the alloca.
287 /// These are operands that in their particular use can be replaced with
288 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
289 /// to PHI nodes and the like. They aren't entirely dead (there might be
290 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
291 /// want to swap this particular input for undef to simplify the use lists of
293 SmallVector<Use *, 8> DeadOperands;
297 static Value *foldSelectInst(SelectInst &SI) {
298 // If the condition being selected on is a constant or the same value is
299 // being selected between, fold the select. Yes this does (rarely) happen
301 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
302 return SI.getOperand(1+CI->isZero());
303 if (SI.getOperand(1) == SI.getOperand(2))
304 return SI.getOperand(1);
309 /// \brief Builder for the alloca slices.
311 /// This class builds a set of alloca slices by recursively visiting the uses
312 /// of an alloca and making a slice for each load and store at each offset.
313 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
314 friend class PtrUseVisitor<SliceBuilder>;
315 friend class InstVisitor<SliceBuilder>;
316 typedef PtrUseVisitor<SliceBuilder> Base;
318 const uint64_t AllocSize;
321 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
322 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
324 /// \brief Set to de-duplicate dead instructions found in the use walk.
325 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
328 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &S)
329 : PtrUseVisitor<SliceBuilder>(DL),
330 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), S(S) {}
333 void markAsDead(Instruction &I) {
334 if (VisitedDeadInsts.insert(&I))
335 S.DeadUsers.push_back(&I);
338 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
339 bool IsSplittable = false) {
340 // Completely skip uses which have a zero size or start either before or
341 // past the end of the allocation.
342 if (Size == 0 || Offset.isNegative() || Offset.uge(AllocSize)) {
343 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
344 << " which has zero size or starts outside of the "
345 << AllocSize << " byte alloca:\n"
346 << " alloca: " << S.AI << "\n"
347 << " use: " << I << "\n");
348 return markAsDead(I);
351 uint64_t BeginOffset = Offset.getZExtValue();
352 uint64_t EndOffset = BeginOffset + Size;
354 // Clamp the end offset to the end of the allocation. Note that this is
355 // formulated to handle even the case where "BeginOffset + Size" overflows.
356 // This may appear superficially to be something we could ignore entirely,
357 // but that is not so! There may be widened loads or PHI-node uses where
358 // some instructions are dead but not others. We can't completely ignore
359 // them, and so have to record at least the information here.
360 assert(AllocSize >= BeginOffset); // Established above.
361 if (Size > AllocSize - BeginOffset) {
362 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
363 << " to remain within the " << AllocSize << " byte alloca:\n"
364 << " alloca: " << S.AI << "\n"
365 << " use: " << I << "\n");
366 EndOffset = AllocSize;
369 S.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
372 void visitBitCastInst(BitCastInst &BC) {
374 return markAsDead(BC);
376 return Base::visitBitCastInst(BC);
379 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
380 if (GEPI.use_empty())
381 return markAsDead(GEPI);
383 return Base::visitGetElementPtrInst(GEPI);
386 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
387 uint64_t Size, bool IsVolatile) {
388 // We allow splitting of loads and stores where the type is an integer type
389 // and cover the entire alloca. This prevents us from splitting over
391 // FIXME: In the great blue eventually, we should eagerly split all integer
392 // loads and stores, and then have a separate step that merges adjacent
393 // alloca partitions into a single partition suitable for integer widening.
394 // Or we should skip the merge step and rely on GVN and other passes to
395 // merge adjacent loads and stores that survive mem2reg.
397 Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize;
399 insertUse(I, Offset, Size, IsSplittable);
402 void visitLoadInst(LoadInst &LI) {
403 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
404 "All simple FCA loads should have been pre-split");
407 return PI.setAborted(&LI);
409 uint64_t Size = DL.getTypeStoreSize(LI.getType());
410 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
413 void visitStoreInst(StoreInst &SI) {
414 Value *ValOp = SI.getValueOperand();
416 return PI.setEscapedAndAborted(&SI);
418 return PI.setAborted(&SI);
420 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
422 // If this memory access can be shown to *statically* extend outside the
423 // bounds of of the allocation, it's behavior is undefined, so simply
424 // ignore it. Note that this is more strict than the generic clamping
425 // behavior of insertUse. We also try to handle cases which might run the
427 // FIXME: We should instead consider the pointer to have escaped if this
428 // function is being instrumented for addressing bugs or race conditions.
429 if (Offset.isNegative() || Size > AllocSize ||
430 Offset.ugt(AllocSize - Size)) {
431 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
432 << " which extends past the end of the " << AllocSize
434 << " alloca: " << S.AI << "\n"
435 << " use: " << SI << "\n");
436 return markAsDead(SI);
439 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
440 "All simple FCA stores should have been pre-split");
441 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
445 void visitMemSetInst(MemSetInst &II) {
446 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
447 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
448 if ((Length && Length->getValue() == 0) ||
449 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
450 // Zero-length mem transfer intrinsics can be ignored entirely.
451 return markAsDead(II);
454 return PI.setAborted(&II);
456 insertUse(II, Offset,
457 Length ? Length->getLimitedValue()
458 : AllocSize - Offset.getLimitedValue(),
462 void visitMemTransferInst(MemTransferInst &II) {
463 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
464 if ((Length && Length->getValue() == 0) ||
465 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
466 // Zero-length mem transfer intrinsics can be ignored entirely.
467 return markAsDead(II);
470 return PI.setAborted(&II);
472 uint64_t RawOffset = Offset.getLimitedValue();
473 uint64_t Size = Length ? Length->getLimitedValue()
474 : AllocSize - RawOffset;
476 // Check for the special case where the same exact value is used for both
478 if (*U == II.getRawDest() && *U == II.getRawSource()) {
479 // For non-volatile transfers this is a no-op.
480 if (!II.isVolatile())
481 return markAsDead(II);
483 return insertUse(II, Offset, Size, /*IsSplittable=*/false);;
486 // If we have seen both source and destination for a mem transfer, then
487 // they both point to the same alloca.
489 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
490 llvm::tie(MTPI, Inserted) =
491 MemTransferSliceMap.insert(std::make_pair(&II, S.Slices.size()));
492 unsigned PrevIdx = MTPI->second;
494 Slice &PrevP = S.Slices[PrevIdx];
496 // Check if the begin offsets match and this is a non-volatile transfer.
497 // In that case, we can completely elide the transfer.
498 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
500 return markAsDead(II);
503 // Otherwise we have an offset transfer within the same alloca. We can't
505 PrevP.makeUnsplittable();
508 // Insert the use now that we've fixed up the splittable nature.
509 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
511 // Check that we ended up with a valid index in the map.
512 assert(S.Slices[PrevIdx].getUse()->getUser() == &II &&
513 "Map index doesn't point back to a slice with this user.");
516 // Disable SRoA for any intrinsics except for lifetime invariants.
517 // FIXME: What about debug intrinsics? This matches old behavior, but
518 // doesn't make sense.
519 void visitIntrinsicInst(IntrinsicInst &II) {
521 return PI.setAborted(&II);
523 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
524 II.getIntrinsicID() == Intrinsic::lifetime_end) {
525 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
526 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
527 Length->getLimitedValue());
528 insertUse(II, Offset, Size, true);
532 Base::visitIntrinsicInst(II);
535 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
536 // We consider any PHI or select that results in a direct load or store of
537 // the same offset to be a viable use for slicing purposes. These uses
538 // are considered unsplittable and the size is the maximum loaded or stored
540 SmallPtrSet<Instruction *, 4> Visited;
541 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
542 Visited.insert(Root);
543 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
544 // If there are no loads or stores, the access is dead. We mark that as
545 // a size zero access.
548 Instruction *I, *UsedI;
549 llvm::tie(UsedI, I) = Uses.pop_back_val();
551 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
552 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
555 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
556 Value *Op = SI->getOperand(0);
559 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
563 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
564 if (!GEP->hasAllZeroIndices())
566 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
567 !isa<SelectInst>(I)) {
571 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
573 if (Visited.insert(cast<Instruction>(*UI)))
574 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
575 } while (!Uses.empty());
580 void visitPHINode(PHINode &PN) {
582 return markAsDead(PN);
584 return PI.setAborted(&PN);
586 // See if we already have computed info on this node.
587 uint64_t &PHISize = PHIOrSelectSizes[&PN];
589 // This is a new PHI node, check for an unsafe use of the PHI node.
590 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&PN, PHISize))
591 return PI.setAborted(UnsafeI);
594 // For PHI and select operands outside the alloca, we can't nuke the entire
595 // phi or select -- the other side might still be relevant, so we special
596 // case them here and use a separate structure to track the operands
597 // themselves which should be replaced with undef.
598 // FIXME: This should instead be escaped in the event we're instrumenting
599 // for address sanitization.
600 if ((Offset.isNegative() && (-Offset).uge(PHISize)) ||
601 (!Offset.isNegative() && Offset.uge(AllocSize))) {
602 S.DeadOperands.push_back(U);
606 insertUse(PN, Offset, PHISize);
609 void visitSelectInst(SelectInst &SI) {
611 return markAsDead(SI);
612 if (Value *Result = foldSelectInst(SI)) {
614 // If the result of the constant fold will be the pointer, recurse
615 // through the select as if we had RAUW'ed it.
618 // Otherwise the operand to the select is dead, and we can replace it
620 S.DeadOperands.push_back(U);
625 return PI.setAborted(&SI);
627 // See if we already have computed info on this node.
628 uint64_t &SelectSize = PHIOrSelectSizes[&SI];
630 // This is a new Select, check for an unsafe use of it.
631 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&SI, SelectSize))
632 return PI.setAborted(UnsafeI);
635 // For PHI and select operands outside the alloca, we can't nuke the entire
636 // phi or select -- the other side might still be relevant, so we special
637 // case them here and use a separate structure to track the operands
638 // themselves which should be replaced with undef.
639 // FIXME: This should instead be escaped in the event we're instrumenting
640 // for address sanitization.
641 if ((Offset.isNegative() && Offset.uge(SelectSize)) ||
642 (!Offset.isNegative() && Offset.uge(AllocSize))) {
643 S.DeadOperands.push_back(U);
647 insertUse(SI, Offset, SelectSize);
650 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
651 void visitInstruction(Instruction &I) {
658 bool operator()(const Slice &S) { return S.isDead(); }
662 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
664 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
667 PointerEscapingInstr(0) {
668 SliceBuilder PB(DL, AI, *this);
669 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
670 if (PtrI.isEscaped() || PtrI.isAborted()) {
671 // FIXME: We should sink the escape vs. abort info into the caller nicely,
672 // possibly by just storing the PtrInfo in the AllocaSlices.
673 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
674 : PtrI.getAbortingInst();
675 assert(PointerEscapingInstr && "Did not track a bad instruction");
679 // Sort the uses. This arranges for the offsets to be in ascending order,
680 // and the sizes to be in descending order.
681 std::sort(Slices.begin(), Slices.end());
683 Slices.erase(std::remove_if(Slices.begin(), Slices.end(), IsSliceDead()),
687 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
689 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
690 StringRef Indent) const {
691 printSlice(OS, I, Indent);
692 printUse(OS, I, Indent);
695 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
696 StringRef Indent) const {
697 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
698 << " slice #" << (I - begin())
699 << (I->isSplittable() ? " (splittable)" : "") << "\n";
702 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
703 StringRef Indent) const {
704 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
707 void AllocaSlices::print(raw_ostream &OS) const {
708 if (PointerEscapingInstr) {
709 OS << "Can't analyze slices for alloca: " << AI << "\n"
710 << " A pointer to this alloca escaped by:\n"
711 << " " << *PointerEscapingInstr << "\n";
715 OS << "Slices of alloca: " << AI << "\n";
716 for (const_iterator I = begin(), E = end(); I != E; ++I)
720 void AllocaSlices::dump(const_iterator I) const { print(dbgs(), I); }
721 void AllocaSlices::dump() const { print(dbgs()); }
723 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
726 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
728 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
729 /// the loads and stores of an alloca instruction, as well as updating its
730 /// debug information. This is used when a domtree is unavailable and thus
731 /// mem2reg in its full form can't be used to handle promotion of allocas to
733 class AllocaPromoter : public LoadAndStorePromoter {
737 SmallVector<DbgDeclareInst *, 4> DDIs;
738 SmallVector<DbgValueInst *, 4> DVIs;
741 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
742 AllocaInst &AI, DIBuilder &DIB)
743 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
745 void run(const SmallVectorImpl<Instruction*> &Insts) {
746 // Remember which alloca we're promoting (for isInstInList).
747 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
748 for (Value::use_iterator UI = DebugNode->use_begin(),
749 UE = DebugNode->use_end();
751 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
753 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
757 LoadAndStorePromoter::run(Insts);
758 AI.eraseFromParent();
759 while (!DDIs.empty())
760 DDIs.pop_back_val()->eraseFromParent();
761 while (!DVIs.empty())
762 DVIs.pop_back_val()->eraseFromParent();
765 virtual bool isInstInList(Instruction *I,
766 const SmallVectorImpl<Instruction*> &Insts) const {
767 if (LoadInst *LI = dyn_cast<LoadInst>(I))
768 return LI->getOperand(0) == &AI;
769 return cast<StoreInst>(I)->getPointerOperand() == &AI;
772 virtual void updateDebugInfo(Instruction *Inst) const {
773 for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(),
774 E = DDIs.end(); I != E; ++I) {
775 DbgDeclareInst *DDI = *I;
776 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
777 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
778 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
779 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
781 for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(),
782 E = DVIs.end(); I != E; ++I) {
783 DbgValueInst *DVI = *I;
785 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
786 // If an argument is zero extended then use argument directly. The ZExt
787 // may be zapped by an optimization pass in future.
788 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
789 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
790 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
791 Arg = dyn_cast<Argument>(SExt->getOperand(0));
793 Arg = SI->getValueOperand();
794 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
795 Arg = LI->getPointerOperand();
799 Instruction *DbgVal =
800 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
802 DbgVal->setDebugLoc(DVI->getDebugLoc());
806 } // end anon namespace
810 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
812 /// This pass takes allocations which can be completely analyzed (that is, they
813 /// don't escape) and tries to turn them into scalar SSA values. There are
814 /// a few steps to this process.
816 /// 1) It takes allocations of aggregates and analyzes the ways in which they
817 /// are used to try to split them into smaller allocations, ideally of
818 /// a single scalar data type. It will split up memcpy and memset accesses
819 /// as necessary and try to isolate individual scalar accesses.
820 /// 2) It will transform accesses into forms which are suitable for SSA value
821 /// promotion. This can be replacing a memset with a scalar store of an
822 /// integer value, or it can involve speculating operations on a PHI or
823 /// select to be a PHI or select of the results.
824 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
825 /// onto insert and extract operations on a vector value, and convert them to
826 /// this form. By doing so, it will enable promotion of vector aggregates to
827 /// SSA vector values.
828 class SROA : public FunctionPass {
829 const bool RequiresDomTree;
832 const DataLayout *DL;
835 /// \brief Worklist of alloca instructions to simplify.
837 /// Each alloca in the function is added to this. Each new alloca formed gets
838 /// added to it as well to recursively simplify unless that alloca can be
839 /// directly promoted. Finally, each time we rewrite a use of an alloca other
840 /// the one being actively rewritten, we add it back onto the list if not
841 /// already present to ensure it is re-visited.
842 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
844 /// \brief A collection of instructions to delete.
845 /// We try to batch deletions to simplify code and make things a bit more
847 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
849 /// \brief Post-promotion worklist.
851 /// Sometimes we discover an alloca which has a high probability of becoming
852 /// viable for SROA after a round of promotion takes place. In those cases,
853 /// the alloca is enqueued here for re-processing.
855 /// Note that we have to be very careful to clear allocas out of this list in
856 /// the event they are deleted.
857 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
859 /// \brief A collection of alloca instructions we can directly promote.
860 std::vector<AllocaInst *> PromotableAllocas;
862 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
864 /// All of these PHIs have been checked for the safety of speculation and by
865 /// being speculated will allow promoting allocas currently in the promotable
867 SetVector<PHINode *, SmallVector<PHINode *, 2> > SpeculatablePHIs;
869 /// \brief A worklist of select instructions to speculate prior to promoting
872 /// All of these select instructions have been checked for the safety of
873 /// speculation and by being speculated will allow promoting allocas
874 /// currently in the promotable queue.
875 SetVector<SelectInst *, SmallVector<SelectInst *, 2> > SpeculatableSelects;
878 SROA(bool RequiresDomTree = true)
879 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
881 initializeSROAPass(*PassRegistry::getPassRegistry());
883 bool runOnFunction(Function &F);
884 void getAnalysisUsage(AnalysisUsage &AU) const;
886 const char *getPassName() const { return "SROA"; }
890 friend class PHIOrSelectSpeculator;
891 friend class AllocaSliceRewriter;
893 bool rewritePartition(AllocaInst &AI, AllocaSlices &S,
894 AllocaSlices::iterator B, AllocaSlices::iterator E,
895 int64_t BeginOffset, int64_t EndOffset,
896 ArrayRef<AllocaSlices::iterator> SplitUses);
897 bool splitAlloca(AllocaInst &AI, AllocaSlices &S);
898 bool runOnAlloca(AllocaInst &AI);
899 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
900 bool promoteAllocas(Function &F);
906 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
907 return new SROA(RequiresDomTree);
910 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
912 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
913 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
916 /// Walk the range of a partitioning looking for a common type to cover this
917 /// sequence of slices.
918 static Type *findCommonType(AllocaSlices::const_iterator B,
919 AllocaSlices::const_iterator E,
920 uint64_t EndOffset) {
922 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
923 Use *U = I->getUse();
924 if (isa<IntrinsicInst>(*U->getUser()))
926 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
930 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser()))
931 UserTy = LI->getType();
932 else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser()))
933 UserTy = SI->getValueOperand()->getType();
935 return 0; // Bail if we have weird uses.
937 if (IntegerType *ITy = dyn_cast<IntegerType>(UserTy)) {
938 // If the type is larger than the partition, skip it. We only encounter
939 // this for split integer operations where we want to use the type of the
940 // entity causing the split.
941 if (ITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
944 // If we have found an integer type use covering the alloca, use that
945 // regardless of the other types, as integers are often used for a
951 if (Ty && Ty != UserTy)
959 /// PHI instructions that use an alloca and are subsequently loaded can be
960 /// rewritten to load both input pointers in the pred blocks and then PHI the
961 /// results, allowing the load of the alloca to be promoted.
963 /// %P2 = phi [i32* %Alloca, i32* %Other]
964 /// %V = load i32* %P2
966 /// %V1 = load i32* %Alloca -> will be mem2reg'd
968 /// %V2 = load i32* %Other
970 /// %V = phi [i32 %V1, i32 %V2]
972 /// We can do this to a select if its only uses are loads and if the operands
973 /// to the select can be loaded unconditionally.
975 /// FIXME: This should be hoisted into a generic utility, likely in
976 /// Transforms/Util/Local.h
977 static bool isSafePHIToSpeculate(PHINode &PN,
978 const DataLayout *DL = 0) {
979 // For now, we can only do this promotion if the load is in the same block
980 // as the PHI, and if there are no stores between the phi and load.
981 // TODO: Allow recursive phi users.
982 // TODO: Allow stores.
983 BasicBlock *BB = PN.getParent();
984 unsigned MaxAlign = 0;
985 bool HaveLoad = false;
986 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end(); UI != UE;
988 LoadInst *LI = dyn_cast<LoadInst>(*UI);
989 if (LI == 0 || !LI->isSimple())
992 // For now we only allow loads in the same block as the PHI. This is
993 // a common case that happens when instcombine merges two loads through
995 if (LI->getParent() != BB)
998 // Ensure that there are no instructions between the PHI and the load that
1000 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1001 if (BBI->mayWriteToMemory())
1004 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1011 // We can only transform this if it is safe to push the loads into the
1012 // predecessor blocks. The only thing to watch out for is that we can't put
1013 // a possibly trapping load in the predecessor if it is a critical edge.
1014 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1015 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1016 Value *InVal = PN.getIncomingValue(Idx);
1018 // If the value is produced by the terminator of the predecessor (an
1019 // invoke) or it has side-effects, there is no valid place to put a load
1020 // in the predecessor.
1021 if (TI == InVal || TI->mayHaveSideEffects())
1024 // If the predecessor has a single successor, then the edge isn't
1026 if (TI->getNumSuccessors() == 1)
1029 // If this pointer is always safe to load, or if we can prove that there
1030 // is already a load in the block, then we can move the load to the pred
1032 if (InVal->isDereferenceablePointer() ||
1033 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, DL))
1042 static void speculatePHINodeLoads(PHINode &PN) {
1043 DEBUG(dbgs() << " original: " << PN << "\n");
1045 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1046 IRBuilderTy PHIBuilder(&PN);
1047 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1048 PN.getName() + ".sroa.speculated");
1050 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1051 // matter which one we get and if any differ.
1052 LoadInst *SomeLoad = cast<LoadInst>(*PN.use_begin());
1053 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1054 unsigned Align = SomeLoad->getAlignment();
1056 // Rewrite all loads of the PN to use the new PHI.
1057 while (!PN.use_empty()) {
1058 LoadInst *LI = cast<LoadInst>(*PN.use_begin());
1059 LI->replaceAllUsesWith(NewPN);
1060 LI->eraseFromParent();
1063 // Inject loads into all of the pred blocks.
1064 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1065 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1066 TerminatorInst *TI = Pred->getTerminator();
1067 Value *InVal = PN.getIncomingValue(Idx);
1068 IRBuilderTy PredBuilder(TI);
1070 LoadInst *Load = PredBuilder.CreateLoad(
1071 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1072 ++NumLoadsSpeculated;
1073 Load->setAlignment(Align);
1075 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1076 NewPN->addIncoming(Load, Pred);
1079 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1080 PN.eraseFromParent();
1083 /// Select instructions that use an alloca and are subsequently loaded can be
1084 /// rewritten to load both input pointers and then select between the result,
1085 /// allowing the load of the alloca to be promoted.
1087 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1088 /// %V = load i32* %P2
1090 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1091 /// %V2 = load i32* %Other
1092 /// %V = select i1 %cond, i32 %V1, i32 %V2
1094 /// We can do this to a select if its only uses are loads and if the operand
1095 /// to the select can be loaded unconditionally.
1096 static bool isSafeSelectToSpeculate(SelectInst &SI, const DataLayout *DL = 0) {
1097 Value *TValue = SI.getTrueValue();
1098 Value *FValue = SI.getFalseValue();
1099 bool TDerefable = TValue->isDereferenceablePointer();
1100 bool FDerefable = FValue->isDereferenceablePointer();
1102 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end(); UI != UE;
1104 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1105 if (LI == 0 || !LI->isSimple())
1108 // Both operands to the select need to be dereferencable, either
1109 // absolutely (e.g. allocas) or at this point because we can see other
1112 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment(), DL))
1115 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment(), DL))
1122 static void speculateSelectInstLoads(SelectInst &SI) {
1123 DEBUG(dbgs() << " original: " << SI << "\n");
1125 IRBuilderTy IRB(&SI);
1126 Value *TV = SI.getTrueValue();
1127 Value *FV = SI.getFalseValue();
1128 // Replace the loads of the select with a select of two loads.
1129 while (!SI.use_empty()) {
1130 LoadInst *LI = cast<LoadInst>(*SI.use_begin());
1131 assert(LI->isSimple() && "We only speculate simple loads");
1133 IRB.SetInsertPoint(LI);
1135 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1137 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1138 NumLoadsSpeculated += 2;
1140 // Transfer alignment and TBAA info if present.
1141 TL->setAlignment(LI->getAlignment());
1142 FL->setAlignment(LI->getAlignment());
1143 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1144 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1145 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1148 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1149 LI->getName() + ".sroa.speculated");
1151 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1152 LI->replaceAllUsesWith(V);
1153 LI->eraseFromParent();
1155 SI.eraseFromParent();
1158 /// \brief Build a GEP out of a base pointer and indices.
1160 /// This will return the BasePtr if that is valid, or build a new GEP
1161 /// instruction using the IRBuilder if GEP-ing is needed.
1162 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1163 SmallVectorImpl<Value *> &Indices) {
1164 if (Indices.empty())
1167 // A single zero index is a no-op, so check for this and avoid building a GEP
1169 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1172 return IRB.CreateInBoundsGEP(BasePtr, Indices, "idx");
1175 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1176 /// TargetTy without changing the offset of the pointer.
1178 /// This routine assumes we've already established a properly offset GEP with
1179 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1180 /// zero-indices down through type layers until we find one the same as
1181 /// TargetTy. If we can't find one with the same type, we at least try to use
1182 /// one with the same size. If none of that works, we just produce the GEP as
1183 /// indicated by Indices to have the correct offset.
1184 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1185 Value *BasePtr, Type *Ty, Type *TargetTy,
1186 SmallVectorImpl<Value *> &Indices) {
1188 return buildGEP(IRB, BasePtr, Indices);
1190 // See if we can descend into a struct and locate a field with the correct
1192 unsigned NumLayers = 0;
1193 Type *ElementTy = Ty;
1195 if (ElementTy->isPointerTy())
1197 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1198 ElementTy = SeqTy->getElementType();
1199 // Note that we use the default address space as this index is over an
1200 // array or a vector, not a pointer.
1201 Indices.push_back(IRB.getInt(APInt(DL.getPointerSizeInBits(0), 0)));
1202 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1203 if (STy->element_begin() == STy->element_end())
1204 break; // Nothing left to descend into.
1205 ElementTy = *STy->element_begin();
1206 Indices.push_back(IRB.getInt32(0));
1211 } while (ElementTy != TargetTy);
1212 if (ElementTy != TargetTy)
1213 Indices.erase(Indices.end() - NumLayers, Indices.end());
1215 return buildGEP(IRB, BasePtr, Indices);
1218 /// \brief Recursively compute indices for a natural GEP.
1220 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1221 /// element types adding appropriate indices for the GEP.
1222 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1223 Value *Ptr, Type *Ty, APInt &Offset,
1225 SmallVectorImpl<Value *> &Indices) {
1227 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices);
1229 // We can't recurse through pointer types.
1230 if (Ty->isPointerTy())
1233 // We try to analyze GEPs over vectors here, but note that these GEPs are
1234 // extremely poorly defined currently. The long-term goal is to remove GEPing
1235 // over a vector from the IR completely.
1236 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1237 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1238 if (ElementSizeInBits % 8)
1239 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1240 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1241 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1242 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1244 Offset -= NumSkippedElements * ElementSize;
1245 Indices.push_back(IRB.getInt(NumSkippedElements));
1246 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1247 Offset, TargetTy, Indices);
1250 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1251 Type *ElementTy = ArrTy->getElementType();
1252 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1253 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1254 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1257 Offset -= NumSkippedElements * ElementSize;
1258 Indices.push_back(IRB.getInt(NumSkippedElements));
1259 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1263 StructType *STy = dyn_cast<StructType>(Ty);
1267 const StructLayout *SL = DL.getStructLayout(STy);
1268 uint64_t StructOffset = Offset.getZExtValue();
1269 if (StructOffset >= SL->getSizeInBytes())
1271 unsigned Index = SL->getElementContainingOffset(StructOffset);
1272 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1273 Type *ElementTy = STy->getElementType(Index);
1274 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1275 return 0; // The offset points into alignment padding.
1277 Indices.push_back(IRB.getInt32(Index));
1278 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1282 /// \brief Get a natural GEP from a base pointer to a particular offset and
1283 /// resulting in a particular type.
1285 /// The goal is to produce a "natural" looking GEP that works with the existing
1286 /// composite types to arrive at the appropriate offset and element type for
1287 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1288 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1289 /// Indices, and setting Ty to the result subtype.
1291 /// If no natural GEP can be constructed, this function returns null.
1292 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1293 Value *Ptr, APInt Offset, Type *TargetTy,
1294 SmallVectorImpl<Value *> &Indices) {
1295 PointerType *Ty = cast<PointerType>(Ptr->getType());
1297 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1299 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1302 Type *ElementTy = Ty->getElementType();
1303 if (!ElementTy->isSized())
1304 return 0; // We can't GEP through an unsized element.
1305 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1306 if (ElementSize == 0)
1307 return 0; // Zero-length arrays can't help us build a natural GEP.
1308 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1310 Offset -= NumSkippedElements * ElementSize;
1311 Indices.push_back(IRB.getInt(NumSkippedElements));
1312 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1316 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1317 /// resulting pointer has PointerTy.
1319 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1320 /// and produces the pointer type desired. Where it cannot, it will try to use
1321 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1322 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1323 /// bitcast to the type.
1325 /// The strategy for finding the more natural GEPs is to peel off layers of the
1326 /// pointer, walking back through bit casts and GEPs, searching for a base
1327 /// pointer from which we can compute a natural GEP with the desired
1328 /// properties. The algorithm tries to fold as many constant indices into
1329 /// a single GEP as possible, thus making each GEP more independent of the
1330 /// surrounding code.
1331 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL,
1332 Value *Ptr, APInt Offset, Type *PointerTy) {
1333 // Even though we don't look through PHI nodes, we could be called on an
1334 // instruction in an unreachable block, which may be on a cycle.
1335 SmallPtrSet<Value *, 4> Visited;
1336 Visited.insert(Ptr);
1337 SmallVector<Value *, 4> Indices;
1339 // We may end up computing an offset pointer that has the wrong type. If we
1340 // never are able to compute one directly that has the correct type, we'll
1341 // fall back to it, so keep it around here.
1342 Value *OffsetPtr = 0;
1344 // Remember any i8 pointer we come across to re-use if we need to do a raw
1347 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1349 Type *TargetTy = PointerTy->getPointerElementType();
1352 // First fold any existing GEPs into the offset.
1353 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1354 APInt GEPOffset(Offset.getBitWidth(), 0);
1355 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1357 Offset += GEPOffset;
1358 Ptr = GEP->getPointerOperand();
1359 if (!Visited.insert(Ptr))
1363 // See if we can perform a natural GEP here.
1365 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1367 if (P->getType() == PointerTy) {
1368 // Zap any offset pointer that we ended up computing in previous rounds.
1369 if (OffsetPtr && OffsetPtr->use_empty())
1370 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1371 I->eraseFromParent();
1379 // Stash this pointer if we've found an i8*.
1380 if (Ptr->getType()->isIntegerTy(8)) {
1382 Int8PtrOffset = Offset;
1385 // Peel off a layer of the pointer and update the offset appropriately.
1386 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1387 Ptr = cast<Operator>(Ptr)->getOperand(0);
1388 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1389 if (GA->mayBeOverridden())
1391 Ptr = GA->getAliasee();
1395 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1396 } while (Visited.insert(Ptr));
1400 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1402 Int8PtrOffset = Offset;
1405 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1406 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1411 // On the off chance we were targeting i8*, guard the bitcast here.
1412 if (Ptr->getType() != PointerTy)
1413 Ptr = IRB.CreateBitCast(Ptr, PointerTy, "cast");
1418 /// \brief Test whether we can convert a value from the old to the new type.
1420 /// This predicate should be used to guard calls to convertValue in order to
1421 /// ensure that we only try to convert viable values. The strategy is that we
1422 /// will peel off single element struct and array wrappings to get to an
1423 /// underlying value, and convert that value.
1424 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1427 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1428 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1429 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1431 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1433 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1436 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1437 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1439 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1447 /// \brief Generic routine to convert an SSA value to a value of a different
1450 /// This will try various different casting techniques, such as bitcasts,
1451 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1452 /// two types for viability with this routine.
1453 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1455 assert(canConvertValue(DL, V->getType(), Ty) &&
1456 "Value not convertable to type");
1457 if (V->getType() == Ty)
1459 if (IntegerType *OldITy = dyn_cast<IntegerType>(V->getType()))
1460 if (IntegerType *NewITy = dyn_cast<IntegerType>(Ty))
1461 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1462 return IRB.CreateZExt(V, NewITy);
1463 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
1464 return IRB.CreateIntToPtr(V, Ty);
1465 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
1466 return IRB.CreatePtrToInt(V, Ty);
1468 return IRB.CreateBitCast(V, Ty);
1471 /// \brief Test whether the given slice use can be promoted to a vector.
1473 /// This function is called to test each entry in a partioning which is slated
1474 /// for a single slice.
1475 static bool isVectorPromotionViableForSlice(
1476 const DataLayout &DL, AllocaSlices &S, uint64_t SliceBeginOffset,
1477 uint64_t SliceEndOffset, VectorType *Ty, uint64_t ElementSize,
1478 AllocaSlices::const_iterator I) {
1479 // First validate the slice offsets.
1480 uint64_t BeginOffset =
1481 std::max(I->beginOffset(), SliceBeginOffset) - SliceBeginOffset;
1482 uint64_t BeginIndex = BeginOffset / ElementSize;
1483 if (BeginIndex * ElementSize != BeginOffset ||
1484 BeginIndex >= Ty->getNumElements())
1486 uint64_t EndOffset =
1487 std::min(I->endOffset(), SliceEndOffset) - SliceBeginOffset;
1488 uint64_t EndIndex = EndOffset / ElementSize;
1489 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1492 assert(EndIndex > BeginIndex && "Empty vector!");
1493 uint64_t NumElements = EndIndex - BeginIndex;
1495 (NumElements == 1) ? Ty->getElementType()
1496 : VectorType::get(Ty->getElementType(), NumElements);
1499 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1501 Use *U = I->getUse();
1503 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1504 if (MI->isVolatile())
1506 if (!I->isSplittable())
1507 return false; // Skip any unsplittable intrinsics.
1508 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1509 // Disable vector promotion when there are loads or stores of an FCA.
1511 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1512 if (LI->isVolatile())
1514 Type *LTy = LI->getType();
1515 if (SliceBeginOffset > I->beginOffset() ||
1516 SliceEndOffset < I->endOffset()) {
1517 assert(LTy->isIntegerTy());
1520 if (!canConvertValue(DL, SliceTy, LTy))
1522 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1523 if (SI->isVolatile())
1525 Type *STy = SI->getValueOperand()->getType();
1526 if (SliceBeginOffset > I->beginOffset() ||
1527 SliceEndOffset < I->endOffset()) {
1528 assert(STy->isIntegerTy());
1531 if (!canConvertValue(DL, STy, SliceTy))
1540 /// \brief Test whether the given alloca partitioning and range of slices can be
1541 /// promoted to a vector.
1543 /// This is a quick test to check whether we can rewrite a particular alloca
1544 /// partition (and its newly formed alloca) into a vector alloca with only
1545 /// whole-vector loads and stores such that it could be promoted to a vector
1546 /// SSA value. We only can ensure this for a limited set of operations, and we
1547 /// don't want to do the rewrites unless we are confident that the result will
1548 /// be promotable, so we have an early test here.
1550 isVectorPromotionViable(const DataLayout &DL, Type *AllocaTy, AllocaSlices &S,
1551 uint64_t SliceBeginOffset, uint64_t SliceEndOffset,
1552 AllocaSlices::const_iterator I,
1553 AllocaSlices::const_iterator E,
1554 ArrayRef<AllocaSlices::iterator> SplitUses) {
1555 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1559 uint64_t ElementSize = DL.getTypeSizeInBits(Ty->getScalarType());
1561 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1562 // that aren't byte sized.
1563 if (ElementSize % 8)
1565 assert((DL.getTypeSizeInBits(Ty) % 8) == 0 &&
1566 "vector size not a multiple of element size?");
1570 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1571 SliceEndOffset, Ty, ElementSize, I))
1574 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1575 SUE = SplitUses.end();
1577 if (!isVectorPromotionViableForSlice(DL, S, SliceBeginOffset,
1578 SliceEndOffset, Ty, ElementSize, *SUI))
1584 /// \brief Test whether a slice of an alloca is valid for integer widening.
1586 /// This implements the necessary checking for the \c isIntegerWideningViable
1587 /// test below on a single slice of the alloca.
1588 static bool isIntegerWideningViableForSlice(const DataLayout &DL,
1590 uint64_t AllocBeginOffset,
1591 uint64_t Size, AllocaSlices &S,
1592 AllocaSlices::const_iterator I,
1593 bool &WholeAllocaOp) {
1594 uint64_t RelBegin = I->beginOffset() - AllocBeginOffset;
1595 uint64_t RelEnd = I->endOffset() - AllocBeginOffset;
1597 // We can't reasonably handle cases where the load or store extends past
1598 // the end of the aloca's type and into its padding.
1602 Use *U = I->getUse();
1604 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1605 if (LI->isVolatile())
1607 if (RelBegin == 0 && RelEnd == Size)
1608 WholeAllocaOp = true;
1609 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1610 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1612 } else if (RelBegin != 0 || RelEnd != Size ||
1613 !canConvertValue(DL, AllocaTy, LI->getType())) {
1614 // Non-integer loads need to be convertible from the alloca type so that
1615 // they are promotable.
1618 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1619 Type *ValueTy = SI->getValueOperand()->getType();
1620 if (SI->isVolatile())
1622 if (RelBegin == 0 && RelEnd == Size)
1623 WholeAllocaOp = true;
1624 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1625 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1627 } else if (RelBegin != 0 || RelEnd != Size ||
1628 !canConvertValue(DL, ValueTy, AllocaTy)) {
1629 // Non-integer stores need to be convertible to the alloca type so that
1630 // they are promotable.
1633 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1634 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1636 if (!I->isSplittable())
1637 return false; // Skip any unsplittable intrinsics.
1638 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1639 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1640 II->getIntrinsicID() != Intrinsic::lifetime_end)
1649 /// \brief Test whether the given alloca partition's integer operations can be
1650 /// widened to promotable ones.
1652 /// This is a quick test to check whether we can rewrite the integer loads and
1653 /// stores to a particular alloca into wider loads and stores and be able to
1654 /// promote the resulting alloca.
1656 isIntegerWideningViable(const DataLayout &DL, Type *AllocaTy,
1657 uint64_t AllocBeginOffset, AllocaSlices &S,
1658 AllocaSlices::const_iterator I,
1659 AllocaSlices::const_iterator E,
1660 ArrayRef<AllocaSlices::iterator> SplitUses) {
1661 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1662 // Don't create integer types larger than the maximum bitwidth.
1663 if (SizeInBits > IntegerType::MAX_INT_BITS)
1666 // Don't try to handle allocas with bit-padding.
1667 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1670 // We need to ensure that an integer type with the appropriate bitwidth can
1671 // be converted to the alloca type, whatever that is. We don't want to force
1672 // the alloca itself to have an integer type if there is a more suitable one.
1673 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1674 if (!canConvertValue(DL, AllocaTy, IntTy) ||
1675 !canConvertValue(DL, IntTy, AllocaTy))
1678 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1680 // While examining uses, we ensure that the alloca has a covering load or
1681 // store. We don't want to widen the integer operations only to fail to
1682 // promote due to some other unsplittable entry (which we may make splittable
1683 // later). However, if there are only splittable uses, go ahead and assume
1684 // that we cover the alloca.
1685 bool WholeAllocaOp = (I != E) ? false : DL.isLegalInteger(SizeInBits);
1688 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1689 S, I, WholeAllocaOp))
1692 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
1693 SUE = SplitUses.end();
1695 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1696 S, *SUI, WholeAllocaOp))
1699 return WholeAllocaOp;
1702 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1703 IntegerType *Ty, uint64_t Offset,
1704 const Twine &Name) {
1705 DEBUG(dbgs() << " start: " << *V << "\n");
1706 IntegerType *IntTy = cast<IntegerType>(V->getType());
1707 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1708 "Element extends past full value");
1709 uint64_t ShAmt = 8*Offset;
1710 if (DL.isBigEndian())
1711 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1713 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
1714 DEBUG(dbgs() << " shifted: " << *V << "\n");
1716 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1717 "Cannot extract to a larger integer!");
1719 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
1720 DEBUG(dbgs() << " trunced: " << *V << "\n");
1725 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
1726 Value *V, uint64_t Offset, const Twine &Name) {
1727 IntegerType *IntTy = cast<IntegerType>(Old->getType());
1728 IntegerType *Ty = cast<IntegerType>(V->getType());
1729 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1730 "Cannot insert a larger integer!");
1731 DEBUG(dbgs() << " start: " << *V << "\n");
1733 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
1734 DEBUG(dbgs() << " extended: " << *V << "\n");
1736 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1737 "Element store outside of alloca store");
1738 uint64_t ShAmt = 8*Offset;
1739 if (DL.isBigEndian())
1740 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1742 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
1743 DEBUG(dbgs() << " shifted: " << *V << "\n");
1746 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
1747 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
1748 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
1749 DEBUG(dbgs() << " masked: " << *Old << "\n");
1750 V = IRB.CreateOr(Old, V, Name + ".insert");
1751 DEBUG(dbgs() << " inserted: " << *V << "\n");
1756 static Value *extractVector(IRBuilderTy &IRB, Value *V,
1757 unsigned BeginIndex, unsigned EndIndex,
1758 const Twine &Name) {
1759 VectorType *VecTy = cast<VectorType>(V->getType());
1760 unsigned NumElements = EndIndex - BeginIndex;
1761 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
1763 if (NumElements == VecTy->getNumElements())
1766 if (NumElements == 1) {
1767 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
1769 DEBUG(dbgs() << " extract: " << *V << "\n");
1773 SmallVector<Constant*, 8> Mask;
1774 Mask.reserve(NumElements);
1775 for (unsigned i = BeginIndex; i != EndIndex; ++i)
1776 Mask.push_back(IRB.getInt32(i));
1777 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1778 ConstantVector::get(Mask),
1780 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1784 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
1785 unsigned BeginIndex, const Twine &Name) {
1786 VectorType *VecTy = cast<VectorType>(Old->getType());
1787 assert(VecTy && "Can only insert a vector into a vector");
1789 VectorType *Ty = dyn_cast<VectorType>(V->getType());
1791 // Single element to insert.
1792 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
1794 DEBUG(dbgs() << " insert: " << *V << "\n");
1798 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
1799 "Too many elements!");
1800 if (Ty->getNumElements() == VecTy->getNumElements()) {
1801 assert(V->getType() == VecTy && "Vector type mismatch");
1804 unsigned EndIndex = BeginIndex + Ty->getNumElements();
1806 // When inserting a smaller vector into the larger to store, we first
1807 // use a shuffle vector to widen it with undef elements, and then
1808 // a second shuffle vector to select between the loaded vector and the
1810 SmallVector<Constant*, 8> Mask;
1811 Mask.reserve(VecTy->getNumElements());
1812 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1813 if (i >= BeginIndex && i < EndIndex)
1814 Mask.push_back(IRB.getInt32(i - BeginIndex));
1816 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
1817 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1818 ConstantVector::get(Mask),
1820 DEBUG(dbgs() << " shuffle: " << *V << "\n");
1823 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
1824 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
1826 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
1828 DEBUG(dbgs() << " blend: " << *V << "\n");
1833 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
1834 /// to use a new alloca.
1836 /// Also implements the rewriting to vector-based accesses when the partition
1837 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
1839 class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
1840 // Befriend the base class so it can delegate to private visit methods.
1841 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
1842 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
1844 const DataLayout &DL;
1847 AllocaInst &OldAI, &NewAI;
1848 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
1851 // If we are rewriting an alloca partition which can be written as pure
1852 // vector operations, we stash extra information here. When VecTy is
1853 // non-null, we have some strict guarantees about the rewritten alloca:
1854 // - The new alloca is exactly the size of the vector type here.
1855 // - The accesses all either map to the entire vector or to a single
1857 // - The set of accessing instructions is only one of those handled above
1858 // in isVectorPromotionViable. Generally these are the same access kinds
1859 // which are promotable via mem2reg.
1862 uint64_t ElementSize;
1864 // This is a convenience and flag variable that will be null unless the new
1865 // alloca's integer operations should be widened to this integer type due to
1866 // passing isIntegerWideningViable above. If it is non-null, the desired
1867 // integer type will be stored here for easy access during rewriting.
1870 // The offset of the slice currently being rewritten.
1871 uint64_t BeginOffset, EndOffset;
1875 Instruction *OldPtr;
1877 // Utility IR builder, whose name prefix is setup for each visited use, and
1878 // the insertion point is set to point to the user.
1882 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &S, SROA &Pass,
1883 AllocaInst &OldAI, AllocaInst &NewAI,
1884 uint64_t NewBeginOffset, uint64_t NewEndOffset,
1885 bool IsVectorPromotable = false,
1886 bool IsIntegerPromotable = false)
1887 : DL(DL), S(S), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
1888 NewAllocaBeginOffset(NewBeginOffset), NewAllocaEndOffset(NewEndOffset),
1889 NewAllocaTy(NewAI.getAllocatedType()),
1890 VecTy(IsVectorPromotable ? cast<VectorType>(NewAllocaTy) : 0),
1891 ElementTy(VecTy ? VecTy->getElementType() : 0),
1892 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
1893 IntTy(IsIntegerPromotable
1896 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
1898 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
1899 OldPtr(), IRB(NewAI.getContext(), ConstantFolder()) {
1901 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
1902 "Only multiple-of-8 sized vector elements are viable");
1905 assert((!IsVectorPromotable && !IsIntegerPromotable) ||
1906 IsVectorPromotable != IsIntegerPromotable);
1909 bool visit(AllocaSlices::const_iterator I) {
1910 bool CanSROA = true;
1911 BeginOffset = I->beginOffset();
1912 EndOffset = I->endOffset();
1913 IsSplittable = I->isSplittable();
1915 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
1917 OldUse = I->getUse();
1918 OldPtr = cast<Instruction>(OldUse->get());
1920 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
1921 IRB.SetInsertPoint(OldUserI);
1922 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
1923 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
1925 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
1932 // Make sure the other visit overloads are visible.
1935 // Every instruction which can end up as a user must have a rewrite rule.
1936 bool visitInstruction(Instruction &I) {
1937 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
1938 llvm_unreachable("No rewrite rule for this instruction!");
1941 Value *getAdjustedAllocaPtr(IRBuilderTy &IRB, uint64_t Offset,
1943 assert(Offset >= NewAllocaBeginOffset);
1944 return getAdjustedPtr(IRB, DL, &NewAI, APInt(DL.getPointerSizeInBits(),
1945 Offset - NewAllocaBeginOffset),
1949 /// \brief Compute suitable alignment to access an offset into the new alloca.
1950 unsigned getOffsetAlign(uint64_t Offset) {
1951 unsigned NewAIAlign = NewAI.getAlignment();
1953 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
1954 return MinAlign(NewAIAlign, Offset);
1957 /// \brief Compute suitable alignment to access a type at an offset of the
1960 /// \returns zero if the type's ABI alignment is a suitable alignment,
1961 /// otherwise returns the maximal suitable alignment.
1962 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
1963 unsigned Align = getOffsetAlign(Offset);
1964 return Align == DL.getABITypeAlignment(Ty) ? 0 : Align;
1967 unsigned getIndex(uint64_t Offset) {
1968 assert(VecTy && "Can only call getIndex when rewriting a vector");
1969 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
1970 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
1971 uint32_t Index = RelOffset / ElementSize;
1972 assert(Index * ElementSize == RelOffset);
1976 void deleteIfTriviallyDead(Value *V) {
1977 Instruction *I = cast<Instruction>(V);
1978 if (isInstructionTriviallyDead(I))
1979 Pass.DeadInsts.insert(I);
1982 Value *rewriteVectorizedLoadInst(uint64_t NewBeginOffset,
1983 uint64_t NewEndOffset) {
1984 unsigned BeginIndex = getIndex(NewBeginOffset);
1985 unsigned EndIndex = getIndex(NewEndOffset);
1986 assert(EndIndex > BeginIndex && "Empty vector!");
1988 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
1990 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
1993 Value *rewriteIntegerLoad(LoadInst &LI, uint64_t NewBeginOffset,
1994 uint64_t NewEndOffset) {
1995 assert(IntTy && "We cannot insert an integer to the alloca");
1996 assert(!LI.isVolatile());
1997 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
1999 V = convertValue(DL, IRB, V, IntTy);
2000 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2001 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2002 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2003 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2008 bool visitLoadInst(LoadInst &LI) {
2009 DEBUG(dbgs() << " original: " << LI << "\n");
2010 Value *OldOp = LI.getOperand(0);
2011 assert(OldOp == OldPtr);
2013 // Compute the intersecting offset range.
2014 assert(BeginOffset < NewAllocaEndOffset);
2015 assert(EndOffset > NewAllocaBeginOffset);
2016 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2017 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2019 uint64_t Size = NewEndOffset - NewBeginOffset;
2021 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), Size * 8)
2023 bool IsPtrAdjusted = false;
2026 V = rewriteVectorizedLoadInst(NewBeginOffset, NewEndOffset);
2027 } else if (IntTy && LI.getType()->isIntegerTy()) {
2028 V = rewriteIntegerLoad(LI, NewBeginOffset, NewEndOffset);
2029 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2030 canConvertValue(DL, NewAllocaTy, LI.getType())) {
2031 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2032 LI.isVolatile(), "load");
2034 Type *LTy = TargetTy->getPointerTo();
2035 V = IRB.CreateAlignedLoad(
2036 getAdjustedAllocaPtr(IRB, NewBeginOffset, LTy),
2037 getOffsetTypeAlign(TargetTy, NewBeginOffset - NewAllocaBeginOffset),
2038 LI.isVolatile(), "load");
2039 IsPtrAdjusted = true;
2041 V = convertValue(DL, IRB, V, TargetTy);
2044 assert(!LI.isVolatile());
2045 assert(LI.getType()->isIntegerTy() &&
2046 "Only integer type loads and stores are split");
2047 assert(Size < DL.getTypeStoreSize(LI.getType()) &&
2048 "Split load isn't smaller than original load");
2049 assert(LI.getType()->getIntegerBitWidth() ==
2050 DL.getTypeStoreSizeInBits(LI.getType()) &&
2051 "Non-byte-multiple bit width");
2052 // Move the insertion point just past the load so that we can refer to it.
2053 IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI)));
2054 // Create a placeholder value with the same type as LI to use as the
2055 // basis for the new value. This allows us to replace the uses of LI with
2056 // the computed value, and then replace the placeholder with LI, leaving
2057 // LI only used for this computation.
2059 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2060 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset,
2062 LI.replaceAllUsesWith(V);
2063 Placeholder->replaceAllUsesWith(&LI);
2066 LI.replaceAllUsesWith(V);
2069 Pass.DeadInsts.insert(&LI);
2070 deleteIfTriviallyDead(OldOp);
2071 DEBUG(dbgs() << " to: " << *V << "\n");
2072 return !LI.isVolatile() && !IsPtrAdjusted;
2075 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp,
2076 uint64_t NewBeginOffset,
2077 uint64_t NewEndOffset) {
2078 if (V->getType() != VecTy) {
2079 unsigned BeginIndex = getIndex(NewBeginOffset);
2080 unsigned EndIndex = getIndex(NewEndOffset);
2081 assert(EndIndex > BeginIndex && "Empty vector!");
2082 unsigned NumElements = EndIndex - BeginIndex;
2083 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2085 (NumElements == 1) ? ElementTy
2086 : VectorType::get(ElementTy, NumElements);
2087 if (V->getType() != SliceTy)
2088 V = convertValue(DL, IRB, V, SliceTy);
2090 // Mix in the existing elements.
2091 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2093 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2095 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2096 Pass.DeadInsts.insert(&SI);
2099 DEBUG(dbgs() << " to: " << *Store << "\n");
2103 bool rewriteIntegerStore(Value *V, StoreInst &SI,
2104 uint64_t NewBeginOffset, uint64_t NewEndOffset) {
2105 assert(IntTy && "We cannot extract an integer from the alloca");
2106 assert(!SI.isVolatile());
2107 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2108 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2110 Old = convertValue(DL, IRB, Old, IntTy);
2111 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2112 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2113 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset,
2116 V = convertValue(DL, IRB, V, NewAllocaTy);
2117 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2118 Pass.DeadInsts.insert(&SI);
2120 DEBUG(dbgs() << " to: " << *Store << "\n");
2124 bool visitStoreInst(StoreInst &SI) {
2125 DEBUG(dbgs() << " original: " << SI << "\n");
2126 Value *OldOp = SI.getOperand(1);
2127 assert(OldOp == OldPtr);
2129 Value *V = SI.getValueOperand();
2131 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2132 // alloca that should be re-examined after promoting this alloca.
2133 if (V->getType()->isPointerTy())
2134 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2135 Pass.PostPromotionWorklist.insert(AI);
2137 // Compute the intersecting offset range.
2138 assert(BeginOffset < NewAllocaEndOffset);
2139 assert(EndOffset > NewAllocaBeginOffset);
2140 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2141 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2143 uint64_t Size = NewEndOffset - NewBeginOffset;
2144 if (Size < DL.getTypeStoreSize(V->getType())) {
2145 assert(!SI.isVolatile());
2146 assert(V->getType()->isIntegerTy() &&
2147 "Only integer type loads and stores are split");
2148 assert(V->getType()->getIntegerBitWidth() ==
2149 DL.getTypeStoreSizeInBits(V->getType()) &&
2150 "Non-byte-multiple bit width");
2151 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8);
2152 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset,
2157 return rewriteVectorizedStoreInst(V, SI, OldOp, NewBeginOffset,
2159 if (IntTy && V->getType()->isIntegerTy())
2160 return rewriteIntegerStore(V, SI, NewBeginOffset, NewEndOffset);
2163 if (NewBeginOffset == NewAllocaBeginOffset &&
2164 NewEndOffset == NewAllocaEndOffset &&
2165 canConvertValue(DL, V->getType(), NewAllocaTy)) {
2166 V = convertValue(DL, IRB, V, NewAllocaTy);
2167 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2170 Value *NewPtr = getAdjustedAllocaPtr(IRB, NewBeginOffset,
2171 V->getType()->getPointerTo());
2172 NewSI = IRB.CreateAlignedStore(
2173 V, NewPtr, getOffsetTypeAlign(
2174 V->getType(), NewBeginOffset - NewAllocaBeginOffset),
2178 Pass.DeadInsts.insert(&SI);
2179 deleteIfTriviallyDead(OldOp);
2181 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2182 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2185 /// \brief Compute an integer value from splatting an i8 across the given
2186 /// number of bytes.
2188 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2189 /// call this routine.
2190 /// FIXME: Heed the advice above.
2192 /// \param V The i8 value to splat.
2193 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2194 Value *getIntegerSplat(Value *V, unsigned Size) {
2195 assert(Size > 0 && "Expected a positive number of bytes.");
2196 IntegerType *VTy = cast<IntegerType>(V->getType());
2197 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2201 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2202 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, "zext"),
2203 ConstantExpr::getUDiv(
2204 Constant::getAllOnesValue(SplatIntTy),
2205 ConstantExpr::getZExt(
2206 Constant::getAllOnesValue(V->getType()),
2212 /// \brief Compute a vector splat for a given element value.
2213 Value *getVectorSplat(Value *V, unsigned NumElements) {
2214 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2215 DEBUG(dbgs() << " splat: " << *V << "\n");
2219 bool visitMemSetInst(MemSetInst &II) {
2220 DEBUG(dbgs() << " original: " << II << "\n");
2221 assert(II.getRawDest() == OldPtr);
2223 // If the memset has a variable size, it cannot be split, just adjust the
2224 // pointer to the new alloca.
2225 if (!isa<Constant>(II.getLength())) {
2227 assert(BeginOffset >= NewAllocaBeginOffset);
2229 getAdjustedAllocaPtr(IRB, BeginOffset, II.getRawDest()->getType()));
2230 Type *CstTy = II.getAlignmentCst()->getType();
2231 II.setAlignment(ConstantInt::get(CstTy, getOffsetAlign(BeginOffset)));
2233 deleteIfTriviallyDead(OldPtr);
2237 // Record this instruction for deletion.
2238 Pass.DeadInsts.insert(&II);
2240 Type *AllocaTy = NewAI.getAllocatedType();
2241 Type *ScalarTy = AllocaTy->getScalarType();
2243 // Compute the intersecting offset range.
2244 assert(BeginOffset < NewAllocaEndOffset);
2245 assert(EndOffset > NewAllocaBeginOffset);
2246 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2247 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2248 uint64_t SliceOffset = NewBeginOffset - NewAllocaBeginOffset;
2250 // If this doesn't map cleanly onto the alloca type, and that type isn't
2251 // a single value type, just emit a memset.
2252 if (!VecTy && !IntTy &&
2253 (BeginOffset > NewAllocaBeginOffset ||
2254 EndOffset < NewAllocaEndOffset ||
2255 !AllocaTy->isSingleValueType() ||
2256 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2257 DL.getTypeSizeInBits(ScalarTy)%8 != 0)) {
2258 Type *SizeTy = II.getLength()->getType();
2259 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2260 CallInst *New = IRB.CreateMemSet(
2261 getAdjustedAllocaPtr(IRB, NewBeginOffset, II.getRawDest()->getType()),
2262 II.getValue(), Size, getOffsetAlign(SliceOffset), II.isVolatile());
2264 DEBUG(dbgs() << " to: " << *New << "\n");
2268 // If we can represent this as a simple value, we have to build the actual
2269 // value to store, which requires expanding the byte present in memset to
2270 // a sensible representation for the alloca type. This is essentially
2271 // splatting the byte to a sufficiently wide integer, splatting it across
2272 // any desired vector width, and bitcasting to the final type.
2276 // If this is a memset of a vectorized alloca, insert it.
2277 assert(ElementTy == ScalarTy);
2279 unsigned BeginIndex = getIndex(NewBeginOffset);
2280 unsigned EndIndex = getIndex(NewEndOffset);
2281 assert(EndIndex > BeginIndex && "Empty vector!");
2282 unsigned NumElements = EndIndex - BeginIndex;
2283 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2286 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2287 Splat = convertValue(DL, IRB, Splat, ElementTy);
2288 if (NumElements > 1)
2289 Splat = getVectorSplat(Splat, NumElements);
2291 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2293 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2295 // If this is a memset on an alloca where we can widen stores, insert the
2297 assert(!II.isVolatile());
2299 uint64_t Size = NewEndOffset - NewBeginOffset;
2300 V = getIntegerSplat(II.getValue(), Size);
2302 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2303 EndOffset != NewAllocaBeginOffset)) {
2304 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2306 Old = convertValue(DL, IRB, Old, IntTy);
2307 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2308 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2310 assert(V->getType() == IntTy &&
2311 "Wrong type for an alloca wide integer!");
2313 V = convertValue(DL, IRB, V, AllocaTy);
2315 // Established these invariants above.
2316 assert(NewBeginOffset == NewAllocaBeginOffset);
2317 assert(NewEndOffset == NewAllocaEndOffset);
2319 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2320 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2321 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2323 V = convertValue(DL, IRB, V, AllocaTy);
2326 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2329 DEBUG(dbgs() << " to: " << *New << "\n");
2330 return !II.isVolatile();
2333 bool visitMemTransferInst(MemTransferInst &II) {
2334 // Rewriting of memory transfer instructions can be a bit tricky. We break
2335 // them into two categories: split intrinsics and unsplit intrinsics.
2337 DEBUG(dbgs() << " original: " << II << "\n");
2339 // Compute the intersecting offset range.
2340 assert(BeginOffset < NewAllocaEndOffset);
2341 assert(EndOffset > NewAllocaBeginOffset);
2342 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2343 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2345 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2346 bool IsDest = II.getRawDest() == OldPtr;
2348 // Compute the relative offset within the transfer.
2349 unsigned IntPtrWidth = DL.getPointerSizeInBits();
2350 APInt RelOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2352 unsigned Align = II.getAlignment();
2353 uint64_t SliceOffset = NewBeginOffset - NewAllocaBeginOffset;
2356 MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2357 MinAlign(II.getAlignment(), getOffsetAlign(SliceOffset)));
2359 // For unsplit intrinsics, we simply modify the source and destination
2360 // pointers in place. This isn't just an optimization, it is a matter of
2361 // correctness. With unsplit intrinsics we may be dealing with transfers
2362 // within a single alloca before SROA ran, or with transfers that have
2363 // a variable length. We may also be dealing with memmove instead of
2364 // memcpy, and so simply updating the pointers is the necessary for us to
2365 // update both source and dest of a single call.
2366 if (!IsSplittable) {
2367 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2370 getAdjustedAllocaPtr(IRB, BeginOffset, II.getRawDest()->getType()));
2372 II.setSource(getAdjustedAllocaPtr(IRB, BeginOffset,
2373 II.getRawSource()->getType()));
2375 Type *CstTy = II.getAlignmentCst()->getType();
2376 II.setAlignment(ConstantInt::get(CstTy, Align));
2378 DEBUG(dbgs() << " to: " << II << "\n");
2379 deleteIfTriviallyDead(OldOp);
2382 // For split transfer intrinsics we have an incredibly useful assurance:
2383 // the source and destination do not reside within the same alloca, and at
2384 // least one of them does not escape. This means that we can replace
2385 // memmove with memcpy, and we don't need to worry about all manner of
2386 // downsides to splitting and transforming the operations.
2388 // If this doesn't map cleanly onto the alloca type, and that type isn't
2389 // a single value type, just emit a memcpy.
2391 = !VecTy && !IntTy && (BeginOffset > NewAllocaBeginOffset ||
2392 EndOffset < NewAllocaEndOffset ||
2393 !NewAI.getAllocatedType()->isSingleValueType());
2395 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2396 // size hasn't been shrunk based on analysis of the viable range, this is
2398 if (EmitMemCpy && &OldAI == &NewAI) {
2399 // Ensure the start lines up.
2400 assert(NewBeginOffset == BeginOffset);
2402 // Rewrite the size as needed.
2403 if (NewEndOffset != EndOffset)
2404 II.setLength(ConstantInt::get(II.getLength()->getType(),
2405 NewEndOffset - NewBeginOffset));
2408 // Record this instruction for deletion.
2409 Pass.DeadInsts.insert(&II);
2411 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2412 // alloca that should be re-examined after rewriting this instruction.
2413 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2415 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2416 Pass.Worklist.insert(AI);
2419 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2420 : II.getRawDest()->getType();
2422 // Compute the other pointer, folding as much as possible to produce
2423 // a single, simple GEP in most cases.
2424 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy);
2426 Value *OurPtr = getAdjustedAllocaPtr(
2427 IRB, NewBeginOffset,
2428 IsDest ? II.getRawDest()->getType() : II.getRawSource()->getType());
2429 Type *SizeTy = II.getLength()->getType();
2430 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2432 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2433 IsDest ? OtherPtr : OurPtr,
2434 Size, Align, II.isVolatile());
2436 DEBUG(dbgs() << " to: " << *New << "\n");
2440 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2441 // is equivalent to 1, but that isn't true if we end up rewriting this as
2446 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2447 NewEndOffset == NewAllocaEndOffset;
2448 uint64_t Size = NewEndOffset - NewBeginOffset;
2449 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2450 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2451 unsigned NumElements = EndIndex - BeginIndex;
2452 IntegerType *SubIntTy
2453 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0;
2455 Type *OtherPtrTy = NewAI.getType();
2456 if (VecTy && !IsWholeAlloca) {
2457 if (NumElements == 1)
2458 OtherPtrTy = VecTy->getElementType();
2460 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2462 OtherPtrTy = OtherPtrTy->getPointerTo();
2463 } else if (IntTy && !IsWholeAlloca) {
2464 OtherPtrTy = SubIntTy->getPointerTo();
2467 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, RelOffset, OtherPtrTy);
2468 Value *DstPtr = &NewAI;
2470 std::swap(SrcPtr, DstPtr);
2473 if (VecTy && !IsWholeAlloca && !IsDest) {
2474 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2476 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2477 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2478 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2480 Src = convertValue(DL, IRB, Src, IntTy);
2481 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2482 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2484 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2488 if (VecTy && !IsWholeAlloca && IsDest) {
2489 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2491 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2492 } else if (IntTy && !IsWholeAlloca && IsDest) {
2493 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2495 Old = convertValue(DL, IRB, Old, IntTy);
2496 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2497 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2498 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2501 StoreInst *Store = cast<StoreInst>(
2502 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2504 DEBUG(dbgs() << " to: " << *Store << "\n");
2505 return !II.isVolatile();
2508 bool visitIntrinsicInst(IntrinsicInst &II) {
2509 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2510 II.getIntrinsicID() == Intrinsic::lifetime_end);
2511 DEBUG(dbgs() << " original: " << II << "\n");
2512 assert(II.getArgOperand(1) == OldPtr);
2514 // Compute the intersecting offset range.
2515 assert(BeginOffset < NewAllocaEndOffset);
2516 assert(EndOffset > NewAllocaBeginOffset);
2517 uint64_t NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2518 uint64_t NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2520 // Record this instruction for deletion.
2521 Pass.DeadInsts.insert(&II);
2524 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2525 NewEndOffset - NewBeginOffset);
2527 getAdjustedAllocaPtr(IRB, NewBeginOffset, II.getArgOperand(1)->getType());
2529 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2530 New = IRB.CreateLifetimeStart(Ptr, Size);
2532 New = IRB.CreateLifetimeEnd(Ptr, Size);
2535 DEBUG(dbgs() << " to: " << *New << "\n");
2539 bool visitPHINode(PHINode &PN) {
2540 DEBUG(dbgs() << " original: " << PN << "\n");
2541 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2542 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2544 // We would like to compute a new pointer in only one place, but have it be
2545 // as local as possible to the PHI. To do that, we re-use the location of
2546 // the old pointer, which necessarily must be in the right position to
2547 // dominate the PHI.
2548 IRBuilderTy PtrBuilder(cast<Instruction>(OldPtr));
2549 PtrBuilder.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) +
2553 getAdjustedAllocaPtr(PtrBuilder, BeginOffset, OldPtr->getType());
2554 // Replace the operands which were using the old pointer.
2555 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2557 DEBUG(dbgs() << " to: " << PN << "\n");
2558 deleteIfTriviallyDead(OldPtr);
2560 // Check whether we can speculate this PHI node, and if so remember that
2561 // fact and return that this alloca remains viable for promotion to an SSA
2563 if (isSafePHIToSpeculate(PN, &DL)) {
2564 Pass.SpeculatablePHIs.insert(&PN);
2568 return false; // PHIs can't be promoted on their own.
2571 bool visitSelectInst(SelectInst &SI) {
2572 DEBUG(dbgs() << " original: " << SI << "\n");
2573 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2574 "Pointer isn't an operand!");
2575 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2576 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2578 Value *NewPtr = getAdjustedAllocaPtr(IRB, BeginOffset, OldPtr->getType());
2579 // Replace the operands which were using the old pointer.
2580 if (SI.getOperand(1) == OldPtr)
2581 SI.setOperand(1, NewPtr);
2582 if (SI.getOperand(2) == OldPtr)
2583 SI.setOperand(2, NewPtr);
2585 DEBUG(dbgs() << " to: " << SI << "\n");
2586 deleteIfTriviallyDead(OldPtr);
2588 // Check whether we can speculate this select instruction, and if so
2589 // remember that fact and return that this alloca remains viable for
2590 // promotion to an SSA value.
2591 if (isSafeSelectToSpeculate(SI, &DL)) {
2592 Pass.SpeculatableSelects.insert(&SI);
2596 return false; // Selects can't be promoted on their own.
2603 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2605 /// This pass aggressively rewrites all aggregate loads and stores on
2606 /// a particular pointer (or any pointer derived from it which we can identify)
2607 /// with scalar loads and stores.
2608 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2609 // Befriend the base class so it can delegate to private visit methods.
2610 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2612 const DataLayout &DL;
2614 /// Queue of pointer uses to analyze and potentially rewrite.
2615 SmallVector<Use *, 8> Queue;
2617 /// Set to prevent us from cycling with phi nodes and loops.
2618 SmallPtrSet<User *, 8> Visited;
2620 /// The current pointer use being rewritten. This is used to dig up the used
2621 /// value (as opposed to the user).
2625 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
2627 /// Rewrite loads and stores through a pointer and all pointers derived from
2629 bool rewrite(Instruction &I) {
2630 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2632 bool Changed = false;
2633 while (!Queue.empty()) {
2634 U = Queue.pop_back_val();
2635 Changed |= visit(cast<Instruction>(U->getUser()));
2641 /// Enqueue all the users of the given instruction for further processing.
2642 /// This uses a set to de-duplicate users.
2643 void enqueueUsers(Instruction &I) {
2644 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
2646 if (Visited.insert(*UI))
2647 Queue.push_back(&UI.getUse());
2650 // Conservative default is to not rewrite anything.
2651 bool visitInstruction(Instruction &I) { return false; }
2653 /// \brief Generic recursive split emission class.
2654 template <typename Derived>
2657 /// The builder used to form new instructions.
2659 /// The indices which to be used with insert- or extractvalue to select the
2660 /// appropriate value within the aggregate.
2661 SmallVector<unsigned, 4> Indices;
2662 /// The indices to a GEP instruction which will move Ptr to the correct slot
2663 /// within the aggregate.
2664 SmallVector<Value *, 4> GEPIndices;
2665 /// The base pointer of the original op, used as a base for GEPing the
2666 /// split operations.
2669 /// Initialize the splitter with an insertion point, Ptr and start with a
2670 /// single zero GEP index.
2671 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2672 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2675 /// \brief Generic recursive split emission routine.
2677 /// This method recursively splits an aggregate op (load or store) into
2678 /// scalar or vector ops. It splits recursively until it hits a single value
2679 /// and emits that single value operation via the template argument.
2681 /// The logic of this routine relies on GEPs and insertvalue and
2682 /// extractvalue all operating with the same fundamental index list, merely
2683 /// formatted differently (GEPs need actual values).
2685 /// \param Ty The type being split recursively into smaller ops.
2686 /// \param Agg The aggregate value being built up or stored, depending on
2687 /// whether this is splitting a load or a store respectively.
2688 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2689 if (Ty->isSingleValueType())
2690 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2692 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2693 unsigned OldSize = Indices.size();
2695 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2697 assert(Indices.size() == OldSize && "Did not return to the old size");
2698 Indices.push_back(Idx);
2699 GEPIndices.push_back(IRB.getInt32(Idx));
2700 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2701 GEPIndices.pop_back();
2707 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2708 unsigned OldSize = Indices.size();
2710 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2712 assert(Indices.size() == OldSize && "Did not return to the old size");
2713 Indices.push_back(Idx);
2714 GEPIndices.push_back(IRB.getInt32(Idx));
2715 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2716 GEPIndices.pop_back();
2722 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2726 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2727 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2728 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2730 /// Emit a leaf load of a single value. This is called at the leaves of the
2731 /// recursive emission to actually load values.
2732 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2733 assert(Ty->isSingleValueType());
2734 // Load the single value and insert it using the indices.
2735 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
2736 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
2737 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2738 DEBUG(dbgs() << " to: " << *Load << "\n");
2742 bool visitLoadInst(LoadInst &LI) {
2743 assert(LI.getPointerOperand() == *U);
2744 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2747 // We have an aggregate being loaded, split it apart.
2748 DEBUG(dbgs() << " original: " << LI << "\n");
2749 LoadOpSplitter Splitter(&LI, *U);
2750 Value *V = UndefValue::get(LI.getType());
2751 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2752 LI.replaceAllUsesWith(V);
2753 LI.eraseFromParent();
2757 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2758 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2759 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2761 /// Emit a leaf store of a single value. This is called at the leaves of the
2762 /// recursive emission to actually produce stores.
2763 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2764 assert(Ty->isSingleValueType());
2765 // Extract the single value and store it using the indices.
2766 Value *Store = IRB.CreateStore(
2767 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2768 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2770 DEBUG(dbgs() << " to: " << *Store << "\n");
2774 bool visitStoreInst(StoreInst &SI) {
2775 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2777 Value *V = SI.getValueOperand();
2778 if (V->getType()->isSingleValueType())
2781 // We have an aggregate being stored, split it apart.
2782 DEBUG(dbgs() << " original: " << SI << "\n");
2783 StoreOpSplitter Splitter(&SI, *U);
2784 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2785 SI.eraseFromParent();
2789 bool visitBitCastInst(BitCastInst &BC) {
2794 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2799 bool visitPHINode(PHINode &PN) {
2804 bool visitSelectInst(SelectInst &SI) {
2811 /// \brief Strip aggregate type wrapping.
2813 /// This removes no-op aggregate types wrapping an underlying type. It will
2814 /// strip as many layers of types as it can without changing either the type
2815 /// size or the allocated size.
2816 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
2817 if (Ty->isSingleValueType())
2820 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
2821 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
2824 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
2825 InnerTy = ArrTy->getElementType();
2826 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
2827 const StructLayout *SL = DL.getStructLayout(STy);
2828 unsigned Index = SL->getElementContainingOffset(0);
2829 InnerTy = STy->getElementType(Index);
2834 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
2835 TypeSize > DL.getTypeSizeInBits(InnerTy))
2838 return stripAggregateTypeWrapping(DL, InnerTy);
2841 /// \brief Try to find a partition of the aggregate type passed in for a given
2842 /// offset and size.
2844 /// This recurses through the aggregate type and tries to compute a subtype
2845 /// based on the offset and size. When the offset and size span a sub-section
2846 /// of an array, it will even compute a new array type for that sub-section,
2847 /// and the same for structs.
2849 /// Note that this routine is very strict and tries to find a partition of the
2850 /// type which produces the *exact* right offset and size. It is not forgiving
2851 /// when the size or offset cause either end of type-based partition to be off.
2852 /// Also, this is a best-effort routine. It is reasonable to give up and not
2853 /// return a type if necessary.
2854 static Type *getTypePartition(const DataLayout &DL, Type *Ty,
2855 uint64_t Offset, uint64_t Size) {
2856 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
2857 return stripAggregateTypeWrapping(DL, Ty);
2858 if (Offset > DL.getTypeAllocSize(Ty) ||
2859 (DL.getTypeAllocSize(Ty) - Offset) < Size)
2862 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
2863 // We can't partition pointers...
2864 if (SeqTy->isPointerTy())
2867 Type *ElementTy = SeqTy->getElementType();
2868 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
2869 uint64_t NumSkippedElements = Offset / ElementSize;
2870 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
2871 if (NumSkippedElements >= ArrTy->getNumElements())
2873 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
2874 if (NumSkippedElements >= VecTy->getNumElements())
2877 Offset -= NumSkippedElements * ElementSize;
2879 // First check if we need to recurse.
2880 if (Offset > 0 || Size < ElementSize) {
2881 // Bail if the partition ends in a different array element.
2882 if ((Offset + Size) > ElementSize)
2884 // Recurse through the element type trying to peel off offset bytes.
2885 return getTypePartition(DL, ElementTy, Offset, Size);
2887 assert(Offset == 0);
2889 if (Size == ElementSize)
2890 return stripAggregateTypeWrapping(DL, ElementTy);
2891 assert(Size > ElementSize);
2892 uint64_t NumElements = Size / ElementSize;
2893 if (NumElements * ElementSize != Size)
2895 return ArrayType::get(ElementTy, NumElements);
2898 StructType *STy = dyn_cast<StructType>(Ty);
2902 const StructLayout *SL = DL.getStructLayout(STy);
2903 if (Offset >= SL->getSizeInBytes())
2905 uint64_t EndOffset = Offset + Size;
2906 if (EndOffset > SL->getSizeInBytes())
2909 unsigned Index = SL->getElementContainingOffset(Offset);
2910 Offset -= SL->getElementOffset(Index);
2912 Type *ElementTy = STy->getElementType(Index);
2913 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
2914 if (Offset >= ElementSize)
2915 return 0; // The offset points into alignment padding.
2917 // See if any partition must be contained by the element.
2918 if (Offset > 0 || Size < ElementSize) {
2919 if ((Offset + Size) > ElementSize)
2921 return getTypePartition(DL, ElementTy, Offset, Size);
2923 assert(Offset == 0);
2925 if (Size == ElementSize)
2926 return stripAggregateTypeWrapping(DL, ElementTy);
2928 StructType::element_iterator EI = STy->element_begin() + Index,
2929 EE = STy->element_end();
2930 if (EndOffset < SL->getSizeInBytes()) {
2931 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
2932 if (Index == EndIndex)
2933 return 0; // Within a single element and its padding.
2935 // Don't try to form "natural" types if the elements don't line up with the
2937 // FIXME: We could potentially recurse down through the last element in the
2938 // sub-struct to find a natural end point.
2939 if (SL->getElementOffset(EndIndex) != EndOffset)
2942 assert(Index < EndIndex);
2943 EE = STy->element_begin() + EndIndex;
2946 // Try to build up a sub-structure.
2947 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
2949 const StructLayout *SubSL = DL.getStructLayout(SubTy);
2950 if (Size != SubSL->getSizeInBytes())
2951 return 0; // The sub-struct doesn't have quite the size needed.
2956 /// \brief Rewrite an alloca partition's users.
2958 /// This routine drives both of the rewriting goals of the SROA pass. It tries
2959 /// to rewrite uses of an alloca partition to be conducive for SSA value
2960 /// promotion. If the partition needs a new, more refined alloca, this will
2961 /// build that new alloca, preserving as much type information as possible, and
2962 /// rewrite the uses of the old alloca to point at the new one and have the
2963 /// appropriate new offsets. It also evaluates how successful the rewrite was
2964 /// at enabling promotion and if it was successful queues the alloca to be
2966 bool SROA::rewritePartition(AllocaInst &AI, AllocaSlices &S,
2967 AllocaSlices::iterator B, AllocaSlices::iterator E,
2968 int64_t BeginOffset, int64_t EndOffset,
2969 ArrayRef<AllocaSlices::iterator> SplitUses) {
2970 assert(BeginOffset < EndOffset);
2971 uint64_t SliceSize = EndOffset - BeginOffset;
2973 // Try to compute a friendly type for this partition of the alloca. This
2974 // won't always succeed, in which case we fall back to a legal integer type
2975 // or an i8 array of an appropriate size.
2977 if (Type *CommonUseTy = findCommonType(B, E, EndOffset))
2978 if (DL->getTypeAllocSize(CommonUseTy) >= SliceSize)
2979 SliceTy = CommonUseTy;
2981 if (Type *TypePartitionTy = getTypePartition(*DL, AI.getAllocatedType(),
2982 BeginOffset, SliceSize))
2983 SliceTy = TypePartitionTy;
2984 if ((!SliceTy || (SliceTy->isArrayTy() &&
2985 SliceTy->getArrayElementType()->isIntegerTy())) &&
2986 DL->isLegalInteger(SliceSize * 8))
2987 SliceTy = Type::getIntNTy(*C, SliceSize * 8);
2989 SliceTy = ArrayType::get(Type::getInt8Ty(*C), SliceSize);
2990 assert(DL->getTypeAllocSize(SliceTy) >= SliceSize);
2992 bool IsVectorPromotable = isVectorPromotionViable(
2993 *DL, SliceTy, S, BeginOffset, EndOffset, B, E, SplitUses);
2995 bool IsIntegerPromotable =
2996 !IsVectorPromotable &&
2997 isIntegerWideningViable(*DL, SliceTy, BeginOffset, S, B, E, SplitUses);
2999 // Check for the case where we're going to rewrite to a new alloca of the
3000 // exact same type as the original, and with the same access offsets. In that
3001 // case, re-use the existing alloca, but still run through the rewriter to
3002 // perform phi and select speculation.
3004 if (SliceTy == AI.getAllocatedType()) {
3005 assert(BeginOffset == 0 &&
3006 "Non-zero begin offset but same alloca type");
3008 // FIXME: We should be able to bail at this point with "nothing changed".
3009 // FIXME: We might want to defer PHI speculation until after here.
3011 unsigned Alignment = AI.getAlignment();
3013 // The minimum alignment which users can rely on when the explicit
3014 // alignment is omitted or zero is that required by the ABI for this
3016 Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
3018 Alignment = MinAlign(Alignment, BeginOffset);
3019 // If we will get at least this much alignment from the type alone, leave
3020 // the alloca's alignment unconstrained.
3021 if (Alignment <= DL->getABITypeAlignment(SliceTy))
3023 NewAI = new AllocaInst(SliceTy, 0, Alignment,
3024 AI.getName() + ".sroa." + Twine(B - S.begin()), &AI);
3028 DEBUG(dbgs() << "Rewriting alloca partition "
3029 << "[" << BeginOffset << "," << EndOffset << ") to: " << *NewAI
3032 // Track the high watermark on several worklists that are only relevant for
3033 // promoted allocas. We will reset it to this point if the alloca is not in
3034 // fact scheduled for promotion.
3035 unsigned PPWOldSize = PostPromotionWorklist.size();
3036 unsigned SPOldSize = SpeculatablePHIs.size();
3037 unsigned SSOldSize = SpeculatableSelects.size();
3039 #if !defined(NDEBUG) || defined(LLVM_ENABLE_STATS)
3040 unsigned NumUses = 0;
3043 AllocaSliceRewriter Rewriter(*DL, S, *this, AI, *NewAI, BeginOffset,
3044 EndOffset, IsVectorPromotable,
3045 IsIntegerPromotable);
3046 bool Promotable = true;
3047 for (ArrayRef<AllocaSlices::iterator>::const_iterator SUI = SplitUses.begin(),
3048 SUE = SplitUses.end();
3049 SUI != SUE; ++SUI) {
3050 DEBUG(dbgs() << " rewriting split ");
3051 DEBUG(S.printSlice(dbgs(), *SUI, ""));
3052 Promotable &= Rewriter.visit(*SUI);
3053 #if !defined(NDEBUG) || defined(LLVM_ENABLE_STATS)
3057 for (AllocaSlices::iterator I = B; I != E; ++I) {
3058 DEBUG(dbgs() << " rewriting ");
3059 DEBUG(S.printSlice(dbgs(), I, ""));
3060 Promotable &= Rewriter.visit(I);
3061 #if !defined(NDEBUG) || defined(LLVM_ENABLE_STATS)
3066 #if !defined(NDEBUG) || defined(LLVM_ENABLE_STATS)
3067 NumAllocaPartitionUses += NumUses;
3068 MaxUsesPerAllocaPartition =
3069 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3072 if (Promotable && (SpeculatablePHIs.size() > SPOldSize ||
3073 SpeculatableSelects.size() > SSOldSize)) {
3074 // If we have a promotable alloca except for some unspeculated loads below
3075 // PHIs or Selects, iterate once. We will speculate the loads and on the
3076 // next iteration rewrite them into a promotable form.
3077 Worklist.insert(NewAI);
3078 } else if (Promotable) {
3079 DEBUG(dbgs() << " and queuing for promotion\n");
3080 PromotableAllocas.push_back(NewAI);
3081 } else if (NewAI != &AI) {
3082 // If we can't promote the alloca, iterate on it to check for new
3083 // refinements exposed by splitting the current alloca. Don't iterate on an
3084 // alloca which didn't actually change and didn't get promoted.
3085 // FIXME: We should actually track whether the rewriter changed anything.
3086 Worklist.insert(NewAI);
3089 // Drop any post-promotion work items if promotion didn't happen.
3091 while (PostPromotionWorklist.size() > PPWOldSize)
3092 PostPromotionWorklist.pop_back();
3093 while (SpeculatablePHIs.size() > SPOldSize)
3094 SpeculatablePHIs.pop_back();
3095 while (SpeculatableSelects.size() > SSOldSize)
3096 SpeculatableSelects.pop_back();
3103 struct IsSliceEndLessOrEqualTo {
3104 uint64_t UpperBound;
3106 IsSliceEndLessOrEqualTo(uint64_t UpperBound) : UpperBound(UpperBound) {}
3108 bool operator()(const AllocaSlices::iterator &I) {
3109 return I->endOffset() <= UpperBound;
3115 removeFinishedSplitUses(SmallVectorImpl<AllocaSlices::iterator> &SplitUses,
3116 uint64_t &MaxSplitUseEndOffset, uint64_t Offset) {
3117 if (Offset >= MaxSplitUseEndOffset) {
3119 MaxSplitUseEndOffset = 0;
3123 size_t SplitUsesOldSize = SplitUses.size();
3124 SplitUses.erase(std::remove_if(SplitUses.begin(), SplitUses.end(),
3125 IsSliceEndLessOrEqualTo(Offset)),
3127 if (SplitUsesOldSize == SplitUses.size())
3130 // Recompute the max. While this is linear, so is remove_if.
3131 MaxSplitUseEndOffset = 0;
3132 for (SmallVectorImpl<AllocaSlices::iterator>::iterator
3133 SUI = SplitUses.begin(),
3134 SUE = SplitUses.end();
3136 MaxSplitUseEndOffset = std::max((*SUI)->endOffset(), MaxSplitUseEndOffset);
3139 /// \brief Walks the slices of an alloca and form partitions based on them,
3140 /// rewriting each of their uses.
3141 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &S) {
3142 if (S.begin() == S.end())
3145 #if !defined(NDEBUG) || defined(LLVM_ENABLE_STATS)
3146 unsigned NumPartitions = 0;
3149 bool Changed = false;
3150 SmallVector<AllocaSlices::iterator, 4> SplitUses;
3151 uint64_t MaxSplitUseEndOffset = 0;
3153 uint64_t BeginOffset = S.begin()->beginOffset();
3155 for (AllocaSlices::iterator SI = S.begin(), SJ = llvm::next(SI), SE = S.end();
3156 SI != SE; SI = SJ) {
3157 uint64_t MaxEndOffset = SI->endOffset();
3159 if (!SI->isSplittable()) {
3160 // When we're forming an unsplittable region, it must always start at the
3161 // first slice and will extend through its end.
3162 assert(BeginOffset == SI->beginOffset());
3164 // Form a partition including all of the overlapping slices with this
3165 // unsplittable slice.
3166 while (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3167 if (!SJ->isSplittable())
3168 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3172 assert(SI->isSplittable()); // Established above.
3174 // Collect all of the overlapping splittable slices.
3175 while (SJ != SE && SJ->beginOffset() < MaxEndOffset &&
3176 SJ->isSplittable()) {
3177 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3181 // Back up MaxEndOffset and SJ if we ended the span early when
3182 // encountering an unsplittable slice.
3183 if (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3184 assert(!SJ->isSplittable());
3185 MaxEndOffset = SJ->beginOffset();
3189 // Check if we have managed to move the end offset forward yet. If so,
3190 // we'll have to rewrite uses and erase old split uses.
3191 if (BeginOffset < MaxEndOffset) {
3192 // Rewrite a sequence of overlapping slices.
3194 rewritePartition(AI, S, SI, SJ, BeginOffset, MaxEndOffset, SplitUses);
3195 #if !defined(NDEBUG) || defined(LLVM_ENABLE_STATS)
3199 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset, MaxEndOffset);
3202 // Accumulate all the splittable slices from the [SI,SJ) region which
3203 // overlap going forward.
3204 for (AllocaSlices::iterator SK = SI; SK != SJ; ++SK)
3205 if (SK->isSplittable() && SK->endOffset() > MaxEndOffset) {
3206 SplitUses.push_back(SK);
3207 MaxSplitUseEndOffset = std::max(SK->endOffset(), MaxSplitUseEndOffset);
3210 // If we're already at the end and we have no split uses, we're done.
3211 if (SJ == SE && SplitUses.empty())
3214 // If we have no split uses or no gap in offsets, we're ready to move to
3216 if (SplitUses.empty() || (SJ != SE && MaxEndOffset == SJ->beginOffset())) {
3217 BeginOffset = SJ->beginOffset();
3221 // Even if we have split slices, if the next slice is splittable and the
3222 // split slices reach it, we can simply set up the beginning offset of the
3223 // next iteration to bridge between them.
3224 if (SJ != SE && SJ->isSplittable() &&
3225 MaxSplitUseEndOffset > SJ->beginOffset()) {
3226 BeginOffset = MaxEndOffset;
3230 // Otherwise, we have a tail of split slices. Rewrite them with an empty
3232 uint64_t PostSplitEndOffset =
3233 SJ == SE ? MaxSplitUseEndOffset : SJ->beginOffset();
3235 Changed |= rewritePartition(AI, S, SJ, SJ, MaxEndOffset, PostSplitEndOffset,
3237 #if !defined(NDEBUG) || defined(LLVM_ENABLE_STATS)
3242 break; // Skip the rest, we don't need to do any cleanup.
3244 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset,
3245 PostSplitEndOffset);
3247 // Now just reset the begin offset for the next iteration.
3248 BeginOffset = SJ->beginOffset();
3251 #if !defined(NDEBUG) || defined(LLVM_ENABLE_STATS)
3252 NumAllocaPartitions += NumPartitions;
3253 MaxPartitionsPerAlloca =
3254 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
3260 /// \brief Analyze an alloca for SROA.
3262 /// This analyzes the alloca to ensure we can reason about it, builds
3263 /// the slices of the alloca, and then hands it off to be split and
3264 /// rewritten as needed.
3265 bool SROA::runOnAlloca(AllocaInst &AI) {
3266 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3267 ++NumAllocasAnalyzed;
3269 // Special case dead allocas, as they're trivial.
3270 if (AI.use_empty()) {
3271 AI.eraseFromParent();
3275 // Skip alloca forms that this analysis can't handle.
3276 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3277 DL->getTypeAllocSize(AI.getAllocatedType()) == 0)
3280 bool Changed = false;
3282 // First, split any FCA loads and stores touching this alloca to promote
3283 // better splitting and promotion opportunities.
3284 AggLoadStoreRewriter AggRewriter(*DL);
3285 Changed |= AggRewriter.rewrite(AI);
3287 // Build the slices using a recursive instruction-visiting builder.
3288 AllocaSlices S(*DL, AI);
3289 DEBUG(S.print(dbgs()));
3293 // Delete all the dead users of this alloca before splitting and rewriting it.
3294 for (AllocaSlices::dead_user_iterator DI = S.dead_user_begin(),
3295 DE = S.dead_user_end();
3298 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3299 DeadInsts.insert(*DI);
3301 for (AllocaSlices::dead_op_iterator DO = S.dead_op_begin(),
3302 DE = S.dead_op_end();
3305 // Clobber the use with an undef value.
3306 **DO = UndefValue::get(OldV->getType());
3307 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3308 if (isInstructionTriviallyDead(OldI)) {
3310 DeadInsts.insert(OldI);
3314 // No slices to split. Leave the dead alloca for a later pass to clean up.
3315 if (S.begin() == S.end())
3318 Changed |= splitAlloca(AI, S);
3320 DEBUG(dbgs() << " Speculating PHIs\n");
3321 while (!SpeculatablePHIs.empty())
3322 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
3324 DEBUG(dbgs() << " Speculating Selects\n");
3325 while (!SpeculatableSelects.empty())
3326 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
3331 /// \brief Delete the dead instructions accumulated in this run.
3333 /// Recursively deletes the dead instructions we've accumulated. This is done
3334 /// at the very end to maximize locality of the recursive delete and to
3335 /// minimize the problems of invalidated instruction pointers as such pointers
3336 /// are used heavily in the intermediate stages of the algorithm.
3338 /// We also record the alloca instructions deleted here so that they aren't
3339 /// subsequently handed to mem2reg to promote.
3340 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3341 while (!DeadInsts.empty()) {
3342 Instruction *I = DeadInsts.pop_back_val();
3343 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3345 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3347 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3348 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3349 // Zero out the operand and see if it becomes trivially dead.
3351 if (isInstructionTriviallyDead(U))
3352 DeadInsts.insert(U);
3355 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3356 DeletedAllocas.insert(AI);
3359 I->eraseFromParent();
3363 /// \brief Promote the allocas, using the best available technique.
3365 /// This attempts to promote whatever allocas have been identified as viable in
3366 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3367 /// If there is a domtree available, we attempt to promote using the full power
3368 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3369 /// based on the SSAUpdater utilities. This function returns whether any
3370 /// promotion occurred.
3371 bool SROA::promoteAllocas(Function &F) {
3372 if (PromotableAllocas.empty())
3375 NumPromoted += PromotableAllocas.size();
3377 if (DT && !ForceSSAUpdater) {
3378 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3379 PromoteMemToReg(PromotableAllocas, *DT);
3380 PromotableAllocas.clear();
3384 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3386 DIBuilder DIB(*F.getParent());
3387 SmallVector<Instruction*, 64> Insts;
3389 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3390 AllocaInst *AI = PromotableAllocas[Idx];
3391 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
3393 Instruction *I = cast<Instruction>(*UI++);
3394 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3395 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3396 // leading to them) here. Eventually it should use them to optimize the
3397 // scalar values produced.
3398 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
3399 assert(onlyUsedByLifetimeMarkers(I) &&
3400 "Found a bitcast used outside of a lifetime marker.");
3401 while (!I->use_empty())
3402 cast<Instruction>(*I->use_begin())->eraseFromParent();
3403 I->eraseFromParent();
3406 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3407 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3408 II->getIntrinsicID() == Intrinsic::lifetime_end);
3409 II->eraseFromParent();
3415 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3419 PromotableAllocas.clear();
3424 /// \brief A predicate to test whether an alloca belongs to a set.
3425 class IsAllocaInSet {
3426 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3430 typedef AllocaInst *argument_type;
3432 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3433 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3437 bool SROA::runOnFunction(Function &F) {
3438 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3439 C = &F.getContext();
3440 DL = getAnalysisIfAvailable<DataLayout>();
3442 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3445 DT = getAnalysisIfAvailable<DominatorTree>();
3447 BasicBlock &EntryBB = F.getEntryBlock();
3448 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3450 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3451 Worklist.insert(AI);
3453 bool Changed = false;
3454 // A set of deleted alloca instruction pointers which should be removed from
3455 // the list of promotable allocas.
3456 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3459 while (!Worklist.empty()) {
3460 Changed |= runOnAlloca(*Worklist.pop_back_val());
3461 deleteDeadInstructions(DeletedAllocas);
3463 // Remove the deleted allocas from various lists so that we don't try to
3464 // continue processing them.
3465 if (!DeletedAllocas.empty()) {
3466 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3467 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3468 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3469 PromotableAllocas.end(),
3470 IsAllocaInSet(DeletedAllocas)),
3471 PromotableAllocas.end());
3472 DeletedAllocas.clear();
3476 Changed |= promoteAllocas(F);
3478 Worklist = PostPromotionWorklist;
3479 PostPromotionWorklist.clear();
3480 } while (!Worklist.empty());
3485 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3486 if (RequiresDomTree)
3487 AU.addRequired<DominatorTree>();
3488 AU.setPreservesCFG();