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 #include "llvm/Transforms/Scalar/SROA.h"
27 #include "llvm/ADT/APInt.h"
28 #include "llvm/ADT/ArrayRef.h"
29 #include "llvm/ADT/DenseMap.h"
30 #include "llvm/ADT/PointerIntPair.h"
31 #include "llvm/ADT/STLExtras.h"
32 #include "llvm/ADT/SetVector.h"
33 #include "llvm/ADT/SmallBitVector.h"
34 #include "llvm/ADT/SmallPtrSet.h"
35 #include "llvm/ADT/SmallVector.h"
36 #include "llvm/ADT/Statistic.h"
37 #include "llvm/ADT/StringRef.h"
38 #include "llvm/ADT/Twine.h"
39 #include "llvm/ADT/iterator.h"
40 #include "llvm/ADT/iterator_range.h"
41 #include "llvm/Analysis/AssumptionCache.h"
42 #include "llvm/Analysis/GlobalsModRef.h"
43 #include "llvm/Analysis/Loads.h"
44 #include "llvm/Analysis/PtrUseVisitor.h"
45 #include "llvm/Transforms/Utils/Local.h"
46 #include "llvm/Config/llvm-config.h"
47 #include "llvm/IR/BasicBlock.h"
48 #include "llvm/IR/Constant.h"
49 #include "llvm/IR/ConstantFolder.h"
50 #include "llvm/IR/Constants.h"
51 #include "llvm/IR/DIBuilder.h"
52 #include "llvm/IR/DataLayout.h"
53 #include "llvm/IR/DebugInfoMetadata.h"
54 #include "llvm/IR/DerivedTypes.h"
55 #include "llvm/IR/Dominators.h"
56 #include "llvm/IR/Function.h"
57 #include "llvm/IR/GetElementPtrTypeIterator.h"
58 #include "llvm/IR/GlobalAlias.h"
59 #include "llvm/IR/IRBuilder.h"
60 #include "llvm/IR/InstVisitor.h"
61 #include "llvm/IR/InstrTypes.h"
62 #include "llvm/IR/Instruction.h"
63 #include "llvm/IR/Instructions.h"
64 #include "llvm/IR/IntrinsicInst.h"
65 #include "llvm/IR/Intrinsics.h"
66 #include "llvm/IR/LLVMContext.h"
67 #include "llvm/IR/Metadata.h"
68 #include "llvm/IR/Module.h"
69 #include "llvm/IR/Operator.h"
70 #include "llvm/IR/PassManager.h"
71 #include "llvm/IR/Type.h"
72 #include "llvm/IR/Use.h"
73 #include "llvm/IR/User.h"
74 #include "llvm/IR/Value.h"
75 #include "llvm/Pass.h"
76 #include "llvm/Support/Casting.h"
77 #include "llvm/Support/CommandLine.h"
78 #include "llvm/Support/Compiler.h"
79 #include "llvm/Support/Debug.h"
80 #include "llvm/Support/ErrorHandling.h"
81 #include "llvm/Support/MathExtras.h"
82 #include "llvm/Support/raw_ostream.h"
83 #include "llvm/Transforms/Scalar.h"
84 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
98 // We only use this for a debug check.
102 using namespace llvm;
103 using namespace llvm::sroa;
105 #define DEBUG_TYPE "sroa"
107 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
108 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
109 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
110 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
111 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
112 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
113 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
114 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
115 STATISTIC(NumDeleted, "Number of instructions deleted");
116 STATISTIC(NumVectorized, "Number of vectorized aggregates");
118 /// Hidden option to enable randomly shuffling the slices to help uncover
119 /// instability in their order.
120 static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
121 cl::init(false), cl::Hidden);
123 /// Hidden option to experiment with completely strict handling of inbounds
125 static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds", cl::init(false),
130 /// A custom IRBuilder inserter which prefixes all names, but only in
132 class IRBuilderPrefixedInserter : public IRBuilderDefaultInserter {
135 const Twine getNameWithPrefix(const Twine &Name) const {
136 return Name.isTriviallyEmpty() ? Name : Prefix + Name;
140 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
143 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
144 BasicBlock::iterator InsertPt) const {
145 IRBuilderDefaultInserter::InsertHelper(I, getNameWithPrefix(Name), BB,
150 /// Provide a type for IRBuilder that drops names in release builds.
151 using IRBuilderTy = IRBuilder<ConstantFolder, IRBuilderPrefixedInserter>;
153 /// A used slice of an alloca.
155 /// This structure represents a slice of an alloca used by some instruction. It
156 /// stores both the begin and end offsets of this use, a pointer to the use
157 /// itself, and a flag indicating whether we can classify the use as splittable
158 /// or not when forming partitions of the alloca.
160 /// The beginning offset of the range.
161 uint64_t BeginOffset = 0;
163 /// The ending offset, not included in the range.
164 uint64_t EndOffset = 0;
166 /// Storage for both the use of this slice and whether it can be
168 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
173 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
174 : BeginOffset(BeginOffset), EndOffset(EndOffset),
175 UseAndIsSplittable(U, IsSplittable) {}
177 uint64_t beginOffset() const { return BeginOffset; }
178 uint64_t endOffset() const { return EndOffset; }
180 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
181 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
183 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
185 bool isDead() const { return getUse() == nullptr; }
186 void kill() { UseAndIsSplittable.setPointer(nullptr); }
188 /// Support for ordering ranges.
190 /// This provides an ordering over ranges such that start offsets are
191 /// always increasing, and within equal start offsets, the end offsets are
192 /// decreasing. Thus the spanning range comes first in a cluster with the
193 /// same start position.
194 bool operator<(const Slice &RHS) const {
195 if (beginOffset() < RHS.beginOffset())
197 if (beginOffset() > RHS.beginOffset())
199 if (isSplittable() != RHS.isSplittable())
200 return !isSplittable();
201 if (endOffset() > RHS.endOffset())
206 /// Support comparison with a single offset to allow binary searches.
207 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
208 uint64_t RHSOffset) {
209 return LHS.beginOffset() < RHSOffset;
211 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
213 return LHSOffset < RHS.beginOffset();
216 bool operator==(const Slice &RHS) const {
217 return isSplittable() == RHS.isSplittable() &&
218 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
220 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
223 } // end anonymous namespace
227 template <typename T> struct isPodLike;
228 template <> struct isPodLike<Slice> { static const bool value = true; };
230 } // end namespace llvm
232 /// Representation of the alloca slices.
234 /// This class represents the slices of an alloca which are formed by its
235 /// various uses. If a pointer escapes, we can't fully build a representation
236 /// for the slices used and we reflect that in this structure. The uses are
237 /// stored, sorted by increasing beginning offset and with unsplittable slices
238 /// starting at a particular offset before splittable slices.
239 class llvm::sroa::AllocaSlices {
241 /// Construct the slices of a particular alloca.
242 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
244 /// Test whether a pointer to the allocation escapes our analysis.
246 /// If this is true, the slices are never fully built and should be
248 bool isEscaped() const { return PointerEscapingInstr; }
250 /// Support for iterating over the slices.
252 using iterator = SmallVectorImpl<Slice>::iterator;
253 using range = iterator_range<iterator>;
255 iterator begin() { return Slices.begin(); }
256 iterator end() { return Slices.end(); }
258 using const_iterator = SmallVectorImpl<Slice>::const_iterator;
259 using const_range = iterator_range<const_iterator>;
261 const_iterator begin() const { return Slices.begin(); }
262 const_iterator end() const { return Slices.end(); }
265 /// Erase a range of slices.
266 void erase(iterator Start, iterator Stop) { Slices.erase(Start, Stop); }
268 /// Insert new slices for this alloca.
270 /// This moves the slices into the alloca's slices collection, and re-sorts
271 /// everything so that the usual ordering properties of the alloca's slices
273 void insert(ArrayRef<Slice> NewSlices) {
274 int OldSize = Slices.size();
275 Slices.append(NewSlices.begin(), NewSlices.end());
276 auto SliceI = Slices.begin() + OldSize;
277 llvm::sort(SliceI, Slices.end());
278 std::inplace_merge(Slices.begin(), SliceI, Slices.end());
281 // Forward declare the iterator and range accessor for walking the
283 class partition_iterator;
284 iterator_range<partition_iterator> partitions();
286 /// Access the dead users for this alloca.
287 ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
289 /// Access the dead operands referring to this alloca.
291 /// These are operands which have cannot actually be used to refer to the
292 /// alloca as they are outside its range and the user doesn't correct for
293 /// that. These mostly consist of PHI node inputs and the like which we just
294 /// need to replace with undef.
295 ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
297 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
298 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
299 void printSlice(raw_ostream &OS, const_iterator I,
300 StringRef Indent = " ") const;
301 void printUse(raw_ostream &OS, const_iterator I,
302 StringRef Indent = " ") const;
303 void print(raw_ostream &OS) const;
304 void dump(const_iterator I) const;
309 template <typename DerivedT, typename RetT = void> class BuilderBase;
312 friend class AllocaSlices::SliceBuilder;
314 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
315 /// Handle to alloca instruction to simplify method interfaces.
319 /// The instruction responsible for this alloca not having a known set
322 /// When an instruction (potentially) escapes the pointer to the alloca, we
323 /// store a pointer to that here and abort trying to form slices of the
324 /// alloca. This will be null if the alloca slices are analyzed successfully.
325 Instruction *PointerEscapingInstr;
327 /// The slices of the alloca.
329 /// We store a vector of the slices formed by uses of the alloca here. This
330 /// vector is sorted by increasing begin offset, and then the unsplittable
331 /// slices before the splittable ones. See the Slice inner class for more
333 SmallVector<Slice, 8> Slices;
335 /// Instructions which will become dead if we rewrite the alloca.
337 /// Note that these are not separated by slice. This is because we expect an
338 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
339 /// all these instructions can simply be removed and replaced with undef as
340 /// they come from outside of the allocated space.
341 SmallVector<Instruction *, 8> DeadUsers;
343 /// Operands which will become dead if we rewrite the alloca.
345 /// These are operands that in their particular use can be replaced with
346 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
347 /// to PHI nodes and the like. They aren't entirely dead (there might be
348 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
349 /// want to swap this particular input for undef to simplify the use lists of
351 SmallVector<Use *, 8> DeadOperands;
354 /// A partition of the slices.
356 /// An ephemeral representation for a range of slices which can be viewed as
357 /// a partition of the alloca. This range represents a span of the alloca's
358 /// memory which cannot be split, and provides access to all of the slices
359 /// overlapping some part of the partition.
361 /// Objects of this type are produced by traversing the alloca's slices, but
362 /// are only ephemeral and not persistent.
363 class llvm::sroa::Partition {
365 friend class AllocaSlices;
366 friend class AllocaSlices::partition_iterator;
368 using iterator = AllocaSlices::iterator;
370 /// The beginning and ending offsets of the alloca for this
372 uint64_t BeginOffset, EndOffset;
374 /// The start and end iterators of this partition.
377 /// A collection of split slice tails overlapping the partition.
378 SmallVector<Slice *, 4> SplitTails;
380 /// Raw constructor builds an empty partition starting and ending at
381 /// the given iterator.
382 Partition(iterator SI) : SI(SI), SJ(SI) {}
385 /// The start offset of this partition.
387 /// All of the contained slices start at or after this offset.
388 uint64_t beginOffset() const { return BeginOffset; }
390 /// The end offset of this partition.
392 /// All of the contained slices end at or before this offset.
393 uint64_t endOffset() const { return EndOffset; }
395 /// The size of the partition.
397 /// Note that this can never be zero.
398 uint64_t size() const {
399 assert(BeginOffset < EndOffset && "Partitions must span some bytes!");
400 return EndOffset - BeginOffset;
403 /// Test whether this partition contains no slices, and merely spans
404 /// a region occupied by split slices.
405 bool empty() const { return SI == SJ; }
407 /// \name Iterate slices that start within the partition.
408 /// These may be splittable or unsplittable. They have a begin offset >= the
409 /// partition begin offset.
411 // FIXME: We should probably define a "concat_iterator" helper and use that
412 // to stitch together pointee_iterators over the split tails and the
413 // contiguous iterators of the partition. That would give a much nicer
414 // interface here. We could then additionally expose filtered iterators for
415 // split, unsplit, and unsplittable splices based on the usage patterns.
416 iterator begin() const { return SI; }
417 iterator end() const { return SJ; }
420 /// Get the sequence of split slice tails.
422 /// These tails are of slices which start before this partition but are
423 /// split and overlap into the partition. We accumulate these while forming
425 ArrayRef<Slice *> splitSliceTails() const { return SplitTails; }
428 /// An iterator over partitions of the alloca's slices.
430 /// This iterator implements the core algorithm for partitioning the alloca's
431 /// slices. It is a forward iterator as we don't support backtracking for
432 /// efficiency reasons, and re-use a single storage area to maintain the
433 /// current set of split slices.
435 /// It is templated on the slice iterator type to use so that it can operate
436 /// with either const or non-const slice iterators.
437 class AllocaSlices::partition_iterator
438 : public iterator_facade_base<partition_iterator, std::forward_iterator_tag,
440 friend class AllocaSlices;
442 /// Most of the state for walking the partitions is held in a class
443 /// with a nice interface for examining them.
446 /// We need to keep the end of the slices to know when to stop.
447 AllocaSlices::iterator SE;
449 /// We also need to keep track of the maximum split end offset seen.
450 /// FIXME: Do we really?
451 uint64_t MaxSplitSliceEndOffset = 0;
453 /// Sets the partition to be empty at given iterator, and sets the
455 partition_iterator(AllocaSlices::iterator SI, AllocaSlices::iterator SE)
457 // If not already at the end, advance our state to form the initial
463 /// Advance the iterator to the next partition.
465 /// Requires that the iterator not be at the end of the slices.
467 assert((P.SI != SE || !P.SplitTails.empty()) &&
468 "Cannot advance past the end of the slices!");
470 // Clear out any split uses which have ended.
471 if (!P.SplitTails.empty()) {
472 if (P.EndOffset >= MaxSplitSliceEndOffset) {
473 // If we've finished all splits, this is easy.
474 P.SplitTails.clear();
475 MaxSplitSliceEndOffset = 0;
477 // Remove the uses which have ended in the prior partition. This
478 // cannot change the max split slice end because we just checked that
479 // the prior partition ended prior to that max.
480 P.SplitTails.erase(llvm::remove_if(P.SplitTails,
482 return S->endOffset() <=
486 assert(llvm::any_of(P.SplitTails,
488 return S->endOffset() == MaxSplitSliceEndOffset;
490 "Could not find the current max split slice offset!");
491 assert(llvm::all_of(P.SplitTails,
493 return S->endOffset() <= MaxSplitSliceEndOffset;
495 "Max split slice end offset is not actually the max!");
499 // If P.SI is already at the end, then we've cleared the split tail and
500 // now have an end iterator.
502 assert(P.SplitTails.empty() && "Failed to clear the split slices!");
506 // If we had a non-empty partition previously, set up the state for
507 // subsequent partitions.
509 // Accumulate all the splittable slices which started in the old
510 // partition into the split list.
512 if (S.isSplittable() && S.endOffset() > P.EndOffset) {
513 P.SplitTails.push_back(&S);
514 MaxSplitSliceEndOffset =
515 std::max(S.endOffset(), MaxSplitSliceEndOffset);
518 // Start from the end of the previous partition.
521 // If P.SI is now at the end, we at most have a tail of split slices.
523 P.BeginOffset = P.EndOffset;
524 P.EndOffset = MaxSplitSliceEndOffset;
528 // If the we have split slices and the next slice is after a gap and is
529 // not splittable immediately form an empty partition for the split
530 // slices up until the next slice begins.
531 if (!P.SplitTails.empty() && P.SI->beginOffset() != P.EndOffset &&
532 !P.SI->isSplittable()) {
533 P.BeginOffset = P.EndOffset;
534 P.EndOffset = P.SI->beginOffset();
539 // OK, we need to consume new slices. Set the end offset based on the
540 // current slice, and step SJ past it. The beginning offset of the
541 // partition is the beginning offset of the next slice unless we have
542 // pre-existing split slices that are continuing, in which case we begin
543 // at the prior end offset.
544 P.BeginOffset = P.SplitTails.empty() ? P.SI->beginOffset() : P.EndOffset;
545 P.EndOffset = P.SI->endOffset();
548 // There are two strategies to form a partition based on whether the
549 // partition starts with an unsplittable slice or a splittable slice.
550 if (!P.SI->isSplittable()) {
551 // When we're forming an unsplittable region, it must always start at
552 // the first slice and will extend through its end.
553 assert(P.BeginOffset == P.SI->beginOffset());
555 // Form a partition including all of the overlapping slices with this
556 // unsplittable slice.
557 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
558 if (!P.SJ->isSplittable())
559 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
563 // We have a partition across a set of overlapping unsplittable
568 // If we're starting with a splittable slice, then we need to form
569 // a synthetic partition spanning it and any other overlapping splittable
571 assert(P.SI->isSplittable() && "Forming a splittable partition!");
573 // Collect all of the overlapping splittable slices.
574 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
575 P.SJ->isSplittable()) {
576 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
580 // Back upiP.EndOffset if we ended the span early when encountering an
581 // unsplittable slice. This synthesizes the early end offset of
582 // a partition spanning only splittable slices.
583 if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
584 assert(!P.SJ->isSplittable());
585 P.EndOffset = P.SJ->beginOffset();
590 bool operator==(const partition_iterator &RHS) const {
591 assert(SE == RHS.SE &&
592 "End iterators don't match between compared partition iterators!");
594 // The observed positions of partitions is marked by the P.SI iterator and
595 // the emptiness of the split slices. The latter is only relevant when
596 // P.SI == SE, as the end iterator will additionally have an empty split
597 // slices list, but the prior may have the same P.SI and a tail of split
599 if (P.SI == RHS.P.SI && P.SplitTails.empty() == RHS.P.SplitTails.empty()) {
600 assert(P.SJ == RHS.P.SJ &&
601 "Same set of slices formed two different sized partitions!");
602 assert(P.SplitTails.size() == RHS.P.SplitTails.size() &&
603 "Same slice position with differently sized non-empty split "
610 partition_iterator &operator++() {
615 Partition &operator*() { return P; }
618 /// A forward range over the partitions of the alloca's slices.
620 /// This accesses an iterator range over the partitions of the alloca's
621 /// slices. It computes these partitions on the fly based on the overlapping
622 /// offsets of the slices and the ability to split them. It will visit "empty"
623 /// partitions to cover regions of the alloca only accessed via split
625 iterator_range<AllocaSlices::partition_iterator> AllocaSlices::partitions() {
626 return make_range(partition_iterator(begin(), end()),
627 partition_iterator(end(), end()));
630 static Value *foldSelectInst(SelectInst &SI) {
631 // If the condition being selected on is a constant or the same value is
632 // being selected between, fold the select. Yes this does (rarely) happen
634 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
635 return SI.getOperand(1 + CI->isZero());
636 if (SI.getOperand(1) == SI.getOperand(2))
637 return SI.getOperand(1);
642 /// A helper that folds a PHI node or a select.
643 static Value *foldPHINodeOrSelectInst(Instruction &I) {
644 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
645 // If PN merges together the same value, return that value.
646 return PN->hasConstantValue();
648 return foldSelectInst(cast<SelectInst>(I));
651 /// Builder for the alloca slices.
653 /// This class builds a set of alloca slices by recursively visiting the uses
654 /// of an alloca and making a slice for each load and store at each offset.
655 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
656 friend class PtrUseVisitor<SliceBuilder>;
657 friend class InstVisitor<SliceBuilder>;
659 using Base = PtrUseVisitor<SliceBuilder>;
661 const uint64_t AllocSize;
664 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
665 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
667 /// Set to de-duplicate dead instructions found in the use walk.
668 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
671 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
672 : PtrUseVisitor<SliceBuilder>(DL),
673 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), AS(AS) {}
676 void markAsDead(Instruction &I) {
677 if (VisitedDeadInsts.insert(&I).second)
678 AS.DeadUsers.push_back(&I);
681 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
682 bool IsSplittable = false) {
683 // Completely skip uses which have a zero size or start either before or
684 // past the end of the allocation.
685 if (Size == 0 || Offset.uge(AllocSize)) {
686 LLVM_DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @"
688 << " which has zero size or starts outside of the "
689 << AllocSize << " byte alloca:\n"
690 << " alloca: " << AS.AI << "\n"
691 << " use: " << I << "\n");
692 return markAsDead(I);
695 uint64_t BeginOffset = Offset.getZExtValue();
696 uint64_t EndOffset = BeginOffset + Size;
698 // Clamp the end offset to the end of the allocation. Note that this is
699 // formulated to handle even the case where "BeginOffset + Size" overflows.
700 // This may appear superficially to be something we could ignore entirely,
701 // but that is not so! There may be widened loads or PHI-node uses where
702 // some instructions are dead but not others. We can't completely ignore
703 // them, and so have to record at least the information here.
704 assert(AllocSize >= BeginOffset); // Established above.
705 if (Size > AllocSize - BeginOffset) {
706 LLVM_DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @"
707 << Offset << " to remain within the " << AllocSize
709 << " alloca: " << AS.AI << "\n"
710 << " use: " << I << "\n");
711 EndOffset = AllocSize;
714 AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
717 void visitBitCastInst(BitCastInst &BC) {
719 return markAsDead(BC);
721 return Base::visitBitCastInst(BC);
724 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
725 if (GEPI.use_empty())
726 return markAsDead(GEPI);
728 if (SROAStrictInbounds && GEPI.isInBounds()) {
729 // FIXME: This is a manually un-factored variant of the basic code inside
730 // of GEPs with checking of the inbounds invariant specified in the
731 // langref in a very strict sense. If we ever want to enable
732 // SROAStrictInbounds, this code should be factored cleanly into
733 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
734 // by writing out the code here where we have the underlying allocation
735 // size readily available.
736 APInt GEPOffset = Offset;
737 const DataLayout &DL = GEPI.getModule()->getDataLayout();
738 for (gep_type_iterator GTI = gep_type_begin(GEPI),
739 GTE = gep_type_end(GEPI);
741 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
745 // Handle a struct index, which adds its field offset to the pointer.
746 if (StructType *STy = GTI.getStructTypeOrNull()) {
747 unsigned ElementIdx = OpC->getZExtValue();
748 const StructLayout *SL = DL.getStructLayout(STy);
750 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
752 // For array or vector indices, scale the index by the size of the
754 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
755 GEPOffset += Index * APInt(Offset.getBitWidth(),
756 DL.getTypeAllocSize(GTI.getIndexedType()));
759 // If this index has computed an intermediate pointer which is not
760 // inbounds, then the result of the GEP is a poison value and we can
761 // delete it and all uses.
762 if (GEPOffset.ugt(AllocSize))
763 return markAsDead(GEPI);
767 return Base::visitGetElementPtrInst(GEPI);
770 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
771 uint64_t Size, bool IsVolatile) {
772 // We allow splitting of non-volatile loads and stores where the type is an
773 // integer type. These may be used to implement 'memcpy' or other "transfer
774 // of bits" patterns.
775 bool IsSplittable = Ty->isIntegerTy() && !IsVolatile;
777 insertUse(I, Offset, Size, IsSplittable);
780 void visitLoadInst(LoadInst &LI) {
781 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
782 "All simple FCA loads should have been pre-split");
785 return PI.setAborted(&LI);
787 const DataLayout &DL = LI.getModule()->getDataLayout();
788 uint64_t Size = DL.getTypeStoreSize(LI.getType());
789 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
792 void visitStoreInst(StoreInst &SI) {
793 Value *ValOp = SI.getValueOperand();
795 return PI.setEscapedAndAborted(&SI);
797 return PI.setAborted(&SI);
799 const DataLayout &DL = SI.getModule()->getDataLayout();
800 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
802 // If this memory access can be shown to *statically* extend outside the
803 // bounds of the allocation, it's behavior is undefined, so simply
804 // ignore it. Note that this is more strict than the generic clamping
805 // behavior of insertUse. We also try to handle cases which might run the
807 // FIXME: We should instead consider the pointer to have escaped if this
808 // function is being instrumented for addressing bugs or race conditions.
809 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
810 LLVM_DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @"
811 << Offset << " which extends past the end of the "
812 << AllocSize << " byte alloca:\n"
813 << " alloca: " << AS.AI << "\n"
814 << " use: " << SI << "\n");
815 return markAsDead(SI);
818 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
819 "All simple FCA stores should have been pre-split");
820 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
823 void visitMemSetInst(MemSetInst &II) {
824 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
825 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
826 if ((Length && Length->getValue() == 0) ||
827 (IsOffsetKnown && Offset.uge(AllocSize)))
828 // Zero-length mem transfer intrinsics can be ignored entirely.
829 return markAsDead(II);
832 return PI.setAborted(&II);
834 insertUse(II, Offset, Length ? Length->getLimitedValue()
835 : AllocSize - Offset.getLimitedValue(),
839 void visitMemTransferInst(MemTransferInst &II) {
840 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
841 if (Length && Length->getValue() == 0)
842 // Zero-length mem transfer intrinsics can be ignored entirely.
843 return markAsDead(II);
845 // Because we can visit these intrinsics twice, also check to see if the
846 // first time marked this instruction as dead. If so, skip it.
847 if (VisitedDeadInsts.count(&II))
851 return PI.setAborted(&II);
853 // This side of the transfer is completely out-of-bounds, and so we can
854 // nuke the entire transfer. However, we also need to nuke the other side
855 // if already added to our partitions.
856 // FIXME: Yet another place we really should bypass this when
857 // instrumenting for ASan.
858 if (Offset.uge(AllocSize)) {
859 SmallDenseMap<Instruction *, unsigned>::iterator MTPI =
860 MemTransferSliceMap.find(&II);
861 if (MTPI != MemTransferSliceMap.end())
862 AS.Slices[MTPI->second].kill();
863 return markAsDead(II);
866 uint64_t RawOffset = Offset.getLimitedValue();
867 uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
869 // Check for the special case where the same exact value is used for both
871 if (*U == II.getRawDest() && *U == II.getRawSource()) {
872 // For non-volatile transfers this is a no-op.
873 if (!II.isVolatile())
874 return markAsDead(II);
876 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
879 // If we have seen both source and destination for a mem transfer, then
880 // they both point to the same alloca.
882 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
883 std::tie(MTPI, Inserted) =
884 MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
885 unsigned PrevIdx = MTPI->second;
887 Slice &PrevP = AS.Slices[PrevIdx];
889 // Check if the begin offsets match and this is a non-volatile transfer.
890 // In that case, we can completely elide the transfer.
891 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
893 return markAsDead(II);
896 // Otherwise we have an offset transfer within the same alloca. We can't
898 PrevP.makeUnsplittable();
901 // Insert the use now that we've fixed up the splittable nature.
902 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
904 // Check that we ended up with a valid index in the map.
905 assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&
906 "Map index doesn't point back to a slice with this user.");
909 // Disable SRoA for any intrinsics except for lifetime invariants.
910 // FIXME: What about debug intrinsics? This matches old behavior, but
911 // doesn't make sense.
912 void visitIntrinsicInst(IntrinsicInst &II) {
914 return PI.setAborted(&II);
916 if (II.isLifetimeStartOrEnd()) {
917 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
918 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
919 Length->getLimitedValue());
920 insertUse(II, Offset, Size, true);
924 Base::visitIntrinsicInst(II);
927 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
928 // We consider any PHI or select that results in a direct load or store of
929 // the same offset to be a viable use for slicing purposes. These uses
930 // are considered unsplittable and the size is the maximum loaded or stored
932 SmallPtrSet<Instruction *, 4> Visited;
933 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
934 Visited.insert(Root);
935 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
936 const DataLayout &DL = Root->getModule()->getDataLayout();
937 // If there are no loads or stores, the access is dead. We mark that as
938 // a size zero access.
941 Instruction *I, *UsedI;
942 std::tie(UsedI, I) = Uses.pop_back_val();
944 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
945 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
948 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
949 Value *Op = SI->getOperand(0);
952 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
956 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
957 if (!GEP->hasAllZeroIndices())
959 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
960 !isa<SelectInst>(I)) {
964 for (User *U : I->users())
965 if (Visited.insert(cast<Instruction>(U)).second)
966 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
967 } while (!Uses.empty());
972 void visitPHINodeOrSelectInst(Instruction &I) {
973 assert(isa<PHINode>(I) || isa<SelectInst>(I));
975 return markAsDead(I);
977 // TODO: We could use SimplifyInstruction here to fold PHINodes and
978 // SelectInsts. However, doing so requires to change the current
979 // dead-operand-tracking mechanism. For instance, suppose neither loading
980 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
981 // trap either. However, if we simply replace %U with undef using the
982 // current dead-operand-tracking mechanism, "load (select undef, undef,
983 // %other)" may trap because the select may return the first operand
985 if (Value *Result = foldPHINodeOrSelectInst(I)) {
987 // If the result of the constant fold will be the pointer, recurse
988 // through the PHI/select as if we had RAUW'ed it.
991 // Otherwise the operand to the PHI/select is dead, and we can replace
993 AS.DeadOperands.push_back(U);
999 return PI.setAborted(&I);
1001 // See if we already have computed info on this node.
1002 uint64_t &Size = PHIOrSelectSizes[&I];
1004 // This is a new PHI/Select, check for an unsafe use of it.
1005 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
1006 return PI.setAborted(UnsafeI);
1009 // For PHI and select operands outside the alloca, we can't nuke the entire
1010 // phi or select -- the other side might still be relevant, so we special
1011 // case them here and use a separate structure to track the operands
1012 // themselves which should be replaced with undef.
1013 // FIXME: This should instead be escaped in the event we're instrumenting
1014 // for address sanitization.
1015 if (Offset.uge(AllocSize)) {
1016 AS.DeadOperands.push_back(U);
1020 insertUse(I, Offset, Size);
1023 void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
1025 void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
1027 /// Disable SROA entirely if there are unhandled users of the alloca.
1028 void visitInstruction(Instruction &I) { PI.setAborted(&I); }
1031 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
1033 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1036 PointerEscapingInstr(nullptr) {
1037 SliceBuilder PB(DL, AI, *this);
1038 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
1039 if (PtrI.isEscaped() || PtrI.isAborted()) {
1040 // FIXME: We should sink the escape vs. abort info into the caller nicely,
1041 // possibly by just storing the PtrInfo in the AllocaSlices.
1042 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
1043 : PtrI.getAbortingInst();
1044 assert(PointerEscapingInstr && "Did not track a bad instruction");
1049 llvm::remove_if(Slices, [](const Slice &S) { return S.isDead(); }),
1053 if (SROARandomShuffleSlices) {
1054 std::mt19937 MT(static_cast<unsigned>(
1055 std::chrono::system_clock::now().time_since_epoch().count()));
1056 std::shuffle(Slices.begin(), Slices.end(), MT);
1060 // Sort the uses. This arranges for the offsets to be in ascending order,
1061 // and the sizes to be in descending order.
1065 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1067 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1068 StringRef Indent) const {
1069 printSlice(OS, I, Indent);
1071 printUse(OS, I, Indent);
1074 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1075 StringRef Indent) const {
1076 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1077 << " slice #" << (I - begin())
1078 << (I->isSplittable() ? " (splittable)" : "");
1081 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1082 StringRef Indent) const {
1083 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1086 void AllocaSlices::print(raw_ostream &OS) const {
1087 if (PointerEscapingInstr) {
1088 OS << "Can't analyze slices for alloca: " << AI << "\n"
1089 << " A pointer to this alloca escaped by:\n"
1090 << " " << *PointerEscapingInstr << "\n";
1094 OS << "Slices of alloca: " << AI << "\n";
1095 for (const_iterator I = begin(), E = end(); I != E; ++I)
1099 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
1102 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
1104 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1106 /// Walk the range of a partitioning looking for a common type to cover this
1107 /// sequence of slices.
1108 static Type *findCommonType(AllocaSlices::const_iterator B,
1109 AllocaSlices::const_iterator E,
1110 uint64_t EndOffset) {
1112 bool TyIsCommon = true;
1113 IntegerType *ITy = nullptr;
1115 // Note that we need to look at *every* alloca slice's Use to ensure we
1116 // always get consistent results regardless of the order of slices.
1117 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1118 Use *U = I->getUse();
1119 if (isa<IntrinsicInst>(*U->getUser()))
1121 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1124 Type *UserTy = nullptr;
1125 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1126 UserTy = LI->getType();
1127 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1128 UserTy = SI->getValueOperand()->getType();
1131 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1132 // If the type is larger than the partition, skip it. We only encounter
1133 // this for split integer operations where we want to use the type of the
1134 // entity causing the split. Also skip if the type is not a byte width
1136 if (UserITy->getBitWidth() % 8 != 0 ||
1137 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1140 // Track the largest bitwidth integer type used in this way in case there
1141 // is no common type.
1142 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1146 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1147 // depend on types skipped above.
1148 if (!UserTy || (Ty && Ty != UserTy))
1149 TyIsCommon = false; // Give up on anything but an iN type.
1154 return TyIsCommon ? Ty : ITy;
1157 /// PHI instructions that use an alloca and are subsequently loaded can be
1158 /// rewritten to load both input pointers in the pred blocks and then PHI the
1159 /// results, allowing the load of the alloca to be promoted.
1161 /// %P2 = phi [i32* %Alloca, i32* %Other]
1162 /// %V = load i32* %P2
1164 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1166 /// %V2 = load i32* %Other
1168 /// %V = phi [i32 %V1, i32 %V2]
1170 /// We can do this to a select if its only uses are loads and if the operands
1171 /// to the select can be loaded unconditionally.
1173 /// FIXME: This should be hoisted into a generic utility, likely in
1174 /// Transforms/Util/Local.h
1175 static bool isSafePHIToSpeculate(PHINode &PN) {
1176 // For now, we can only do this promotion if the load is in the same block
1177 // as the PHI, and if there are no stores between the phi and load.
1178 // TODO: Allow recursive phi users.
1179 // TODO: Allow stores.
1180 BasicBlock *BB = PN.getParent();
1181 unsigned MaxAlign = 0;
1182 bool HaveLoad = false;
1183 for (User *U : PN.users()) {
1184 LoadInst *LI = dyn_cast<LoadInst>(U);
1185 if (!LI || !LI->isSimple())
1188 // For now we only allow loads in the same block as the PHI. This is
1189 // a common case that happens when instcombine merges two loads through
1191 if (LI->getParent() != BB)
1194 // Ensure that there are no instructions between the PHI and the load that
1196 for (BasicBlock::iterator BBI(PN); &*BBI != LI; ++BBI)
1197 if (BBI->mayWriteToMemory())
1200 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1207 const DataLayout &DL = PN.getModule()->getDataLayout();
1209 // We can only transform this if it is safe to push the loads into the
1210 // predecessor blocks. The only thing to watch out for is that we can't put
1211 // a possibly trapping load in the predecessor if it is a critical edge.
1212 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1213 Instruction *TI = PN.getIncomingBlock(Idx)->getTerminator();
1214 Value *InVal = PN.getIncomingValue(Idx);
1216 // If the value is produced by the terminator of the predecessor (an
1217 // invoke) or it has side-effects, there is no valid place to put a load
1218 // in the predecessor.
1219 if (TI == InVal || TI->mayHaveSideEffects())
1222 // If the predecessor has a single successor, then the edge isn't
1224 if (TI->getNumSuccessors() == 1)
1227 // If this pointer is always safe to load, or if we can prove that there
1228 // is already a load in the block, then we can move the load to the pred
1230 if (isSafeToLoadUnconditionally(InVal, MaxAlign, DL, TI))
1239 static void speculatePHINodeLoads(PHINode &PN) {
1240 LLVM_DEBUG(dbgs() << " original: " << PN << "\n");
1242 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1243 IRBuilderTy PHIBuilder(&PN);
1244 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1245 PN.getName() + ".sroa.speculated");
1247 // Get the AA tags and alignment to use from one of the loads. It doesn't
1248 // matter which one we get and if any differ.
1249 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1252 SomeLoad->getAAMetadata(AATags);
1253 unsigned Align = SomeLoad->getAlignment();
1255 // Rewrite all loads of the PN to use the new PHI.
1256 while (!PN.use_empty()) {
1257 LoadInst *LI = cast<LoadInst>(PN.user_back());
1258 LI->replaceAllUsesWith(NewPN);
1259 LI->eraseFromParent();
1262 // Inject loads into all of the pred blocks.
1263 DenseMap<BasicBlock*, Value*> InjectedLoads;
1264 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1265 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1266 Value *InVal = PN.getIncomingValue(Idx);
1268 // A PHI node is allowed to have multiple (duplicated) entries for the same
1269 // basic block, as long as the value is the same. So if we already injected
1270 // a load in the predecessor, then we should reuse the same load for all
1271 // duplicated entries.
1272 if (Value* V = InjectedLoads.lookup(Pred)) {
1273 NewPN->addIncoming(V, Pred);
1277 Instruction *TI = Pred->getTerminator();
1278 IRBuilderTy PredBuilder(TI);
1280 LoadInst *Load = PredBuilder.CreateLoad(
1281 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1282 ++NumLoadsSpeculated;
1283 Load->setAlignment(Align);
1285 Load->setAAMetadata(AATags);
1286 NewPN->addIncoming(Load, Pred);
1287 InjectedLoads[Pred] = Load;
1290 LLVM_DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1291 PN.eraseFromParent();
1294 /// Select instructions that use an alloca and are subsequently loaded can be
1295 /// rewritten to load both input pointers and then select between the result,
1296 /// allowing the load of the alloca to be promoted.
1298 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1299 /// %V = load i32* %P2
1301 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1302 /// %V2 = load i32* %Other
1303 /// %V = select i1 %cond, i32 %V1, i32 %V2
1305 /// We can do this to a select if its only uses are loads and if the operand
1306 /// to the select can be loaded unconditionally.
1307 static bool isSafeSelectToSpeculate(SelectInst &SI) {
1308 Value *TValue = SI.getTrueValue();
1309 Value *FValue = SI.getFalseValue();
1310 const DataLayout &DL = SI.getModule()->getDataLayout();
1312 for (User *U : SI.users()) {
1313 LoadInst *LI = dyn_cast<LoadInst>(U);
1314 if (!LI || !LI->isSimple())
1317 // Both operands to the select need to be dereferenceable, either
1318 // absolutely (e.g. allocas) or at this point because we can see other
1320 if (!isSafeToLoadUnconditionally(TValue, LI->getAlignment(), DL, LI))
1322 if (!isSafeToLoadUnconditionally(FValue, LI->getAlignment(), DL, LI))
1329 static void speculateSelectInstLoads(SelectInst &SI) {
1330 LLVM_DEBUG(dbgs() << " original: " << SI << "\n");
1332 IRBuilderTy IRB(&SI);
1333 Value *TV = SI.getTrueValue();
1334 Value *FV = SI.getFalseValue();
1335 // Replace the loads of the select with a select of two loads.
1336 while (!SI.use_empty()) {
1337 LoadInst *LI = cast<LoadInst>(SI.user_back());
1338 assert(LI->isSimple() && "We only speculate simple loads");
1340 IRB.SetInsertPoint(LI);
1342 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1344 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1345 NumLoadsSpeculated += 2;
1347 // Transfer alignment and AA info if present.
1348 TL->setAlignment(LI->getAlignment());
1349 FL->setAlignment(LI->getAlignment());
1352 LI->getAAMetadata(Tags);
1354 TL->setAAMetadata(Tags);
1355 FL->setAAMetadata(Tags);
1358 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1359 LI->getName() + ".sroa.speculated");
1361 LLVM_DEBUG(dbgs() << " speculated to: " << *V << "\n");
1362 LI->replaceAllUsesWith(V);
1363 LI->eraseFromParent();
1365 SI.eraseFromParent();
1368 /// Build a GEP out of a base pointer and indices.
1370 /// This will return the BasePtr if that is valid, or build a new GEP
1371 /// instruction using the IRBuilder if GEP-ing is needed.
1372 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1373 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1374 if (Indices.empty())
1377 // A single zero index is a no-op, so check for this and avoid building a GEP
1379 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1382 return IRB.CreateInBoundsGEP(nullptr, BasePtr, Indices,
1383 NamePrefix + "sroa_idx");
1386 /// Get a natural GEP off of the BasePtr walking through Ty toward
1387 /// TargetTy without changing the offset of the pointer.
1389 /// This routine assumes we've already established a properly offset GEP with
1390 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1391 /// zero-indices down through type layers until we find one the same as
1392 /// TargetTy. If we can't find one with the same type, we at least try to use
1393 /// one with the same size. If none of that works, we just produce the GEP as
1394 /// indicated by Indices to have the correct offset.
1395 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1396 Value *BasePtr, Type *Ty, Type *TargetTy,
1397 SmallVectorImpl<Value *> &Indices,
1400 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1402 // Offset size to use for the indices.
1403 unsigned OffsetSize = DL.getIndexTypeSizeInBits(BasePtr->getType());
1405 // See if we can descend into a struct and locate a field with the correct
1407 unsigned NumLayers = 0;
1408 Type *ElementTy = Ty;
1410 if (ElementTy->isPointerTy())
1413 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1414 ElementTy = ArrayTy->getElementType();
1415 Indices.push_back(IRB.getIntN(OffsetSize, 0));
1416 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1417 ElementTy = VectorTy->getElementType();
1418 Indices.push_back(IRB.getInt32(0));
1419 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1420 if (STy->element_begin() == STy->element_end())
1421 break; // Nothing left to descend into.
1422 ElementTy = *STy->element_begin();
1423 Indices.push_back(IRB.getInt32(0));
1428 } while (ElementTy != TargetTy);
1429 if (ElementTy != TargetTy)
1430 Indices.erase(Indices.end() - NumLayers, Indices.end());
1432 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1435 /// Recursively compute indices for a natural GEP.
1437 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1438 /// element types adding appropriate indices for the GEP.
1439 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1440 Value *Ptr, Type *Ty, APInt &Offset,
1442 SmallVectorImpl<Value *> &Indices,
1445 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1448 // We can't recurse through pointer types.
1449 if (Ty->isPointerTy())
1452 // We try to analyze GEPs over vectors here, but note that these GEPs are
1453 // extremely poorly defined currently. The long-term goal is to remove GEPing
1454 // over a vector from the IR completely.
1455 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1456 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1457 if (ElementSizeInBits % 8 != 0) {
1458 // GEPs over non-multiple of 8 size vector elements are invalid.
1461 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1462 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1463 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1465 Offset -= NumSkippedElements * ElementSize;
1466 Indices.push_back(IRB.getInt(NumSkippedElements));
1467 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1468 Offset, TargetTy, Indices, NamePrefix);
1471 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1472 Type *ElementTy = ArrTy->getElementType();
1473 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1474 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1475 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1478 Offset -= NumSkippedElements * ElementSize;
1479 Indices.push_back(IRB.getInt(NumSkippedElements));
1480 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1481 Indices, NamePrefix);
1484 StructType *STy = dyn_cast<StructType>(Ty);
1488 const StructLayout *SL = DL.getStructLayout(STy);
1489 uint64_t StructOffset = Offset.getZExtValue();
1490 if (StructOffset >= SL->getSizeInBytes())
1492 unsigned Index = SL->getElementContainingOffset(StructOffset);
1493 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1494 Type *ElementTy = STy->getElementType(Index);
1495 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1496 return nullptr; // The offset points into alignment padding.
1498 Indices.push_back(IRB.getInt32(Index));
1499 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1500 Indices, NamePrefix);
1503 /// Get a natural GEP from a base pointer to a particular offset and
1504 /// resulting in a particular type.
1506 /// The goal is to produce a "natural" looking GEP that works with the existing
1507 /// composite types to arrive at the appropriate offset and element type for
1508 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1509 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1510 /// Indices, and setting Ty to the result subtype.
1512 /// If no natural GEP can be constructed, this function returns null.
1513 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1514 Value *Ptr, APInt Offset, Type *TargetTy,
1515 SmallVectorImpl<Value *> &Indices,
1517 PointerType *Ty = cast<PointerType>(Ptr->getType());
1519 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1521 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1524 Type *ElementTy = Ty->getElementType();
1525 if (!ElementTy->isSized())
1526 return nullptr; // We can't GEP through an unsized element.
1527 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1528 if (ElementSize == 0)
1529 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1530 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1532 Offset -= NumSkippedElements * ElementSize;
1533 Indices.push_back(IRB.getInt(NumSkippedElements));
1534 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1535 Indices, NamePrefix);
1538 /// Compute an adjusted pointer from Ptr by Offset bytes where the
1539 /// resulting pointer has PointerTy.
1541 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1542 /// and produces the pointer type desired. Where it cannot, it will try to use
1543 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1544 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1545 /// bitcast to the type.
1547 /// The strategy for finding the more natural GEPs is to peel off layers of the
1548 /// pointer, walking back through bit casts and GEPs, searching for a base
1549 /// pointer from which we can compute a natural GEP with the desired
1550 /// properties. The algorithm tries to fold as many constant indices into
1551 /// a single GEP as possible, thus making each GEP more independent of the
1552 /// surrounding code.
1553 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1554 APInt Offset, Type *PointerTy, Twine NamePrefix) {
1555 // Even though we don't look through PHI nodes, we could be called on an
1556 // instruction in an unreachable block, which may be on a cycle.
1557 SmallPtrSet<Value *, 4> Visited;
1558 Visited.insert(Ptr);
1559 SmallVector<Value *, 4> Indices;
1561 // We may end up computing an offset pointer that has the wrong type. If we
1562 // never are able to compute one directly that has the correct type, we'll
1563 // fall back to it, so keep it and the base it was computed from around here.
1564 Value *OffsetPtr = nullptr;
1565 Value *OffsetBasePtr;
1567 // Remember any i8 pointer we come across to re-use if we need to do a raw
1569 Value *Int8Ptr = nullptr;
1570 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1572 Type *TargetTy = PointerTy->getPointerElementType();
1575 // First fold any existing GEPs into the offset.
1576 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1577 APInt GEPOffset(Offset.getBitWidth(), 0);
1578 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1580 Offset += GEPOffset;
1581 Ptr = GEP->getPointerOperand();
1582 if (!Visited.insert(Ptr).second)
1586 // See if we can perform a natural GEP here.
1588 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1589 Indices, NamePrefix)) {
1590 // If we have a new natural pointer at the offset, clear out any old
1591 // offset pointer we computed. Unless it is the base pointer or
1592 // a non-instruction, we built a GEP we don't need. Zap it.
1593 if (OffsetPtr && OffsetPtr != OffsetBasePtr)
1594 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) {
1595 assert(I->use_empty() && "Built a GEP with uses some how!");
1596 I->eraseFromParent();
1599 OffsetBasePtr = Ptr;
1600 // If we also found a pointer of the right type, we're done.
1601 if (P->getType() == PointerTy)
1605 // Stash this pointer if we've found an i8*.
1606 if (Ptr->getType()->isIntegerTy(8)) {
1608 Int8PtrOffset = Offset;
1611 // Peel off a layer of the pointer and update the offset appropriately.
1612 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1613 Ptr = cast<Operator>(Ptr)->getOperand(0);
1614 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1615 if (GA->isInterposable())
1617 Ptr = GA->getAliasee();
1621 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1622 } while (Visited.insert(Ptr).second);
1626 Int8Ptr = IRB.CreateBitCast(
1627 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1628 NamePrefix + "sroa_raw_cast");
1629 Int8PtrOffset = Offset;
1632 OffsetPtr = Int8PtrOffset == 0
1634 : IRB.CreateInBoundsGEP(IRB.getInt8Ty(), Int8Ptr,
1635 IRB.getInt(Int8PtrOffset),
1636 NamePrefix + "sroa_raw_idx");
1640 // On the off chance we were targeting i8*, guard the bitcast here.
1641 if (Ptr->getType() != PointerTy)
1642 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1647 /// Compute the adjusted alignment for a load or store from an offset.
1648 static unsigned getAdjustedAlignment(Instruction *I, uint64_t Offset,
1649 const DataLayout &DL) {
1652 if (auto *LI = dyn_cast<LoadInst>(I)) {
1653 Alignment = LI->getAlignment();
1655 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
1656 Alignment = SI->getAlignment();
1657 Ty = SI->getValueOperand()->getType();
1659 llvm_unreachable("Only loads and stores are allowed!");
1663 Alignment = DL.getABITypeAlignment(Ty);
1665 return MinAlign(Alignment, Offset);
1668 /// Test whether we can convert a value from the old to the new type.
1670 /// This predicate should be used to guard calls to convertValue in order to
1671 /// ensure that we only try to convert viable values. The strategy is that we
1672 /// will peel off single element struct and array wrappings to get to an
1673 /// underlying value, and convert that value.
1674 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1678 // For integer types, we can't handle any bit-width differences. This would
1679 // break both vector conversions with extension and introduce endianness
1680 // issues when in conjunction with loads and stores.
1681 if (isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) {
1682 assert(cast<IntegerType>(OldTy)->getBitWidth() !=
1683 cast<IntegerType>(NewTy)->getBitWidth() &&
1684 "We can't have the same bitwidth for different int types");
1688 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1690 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1693 // We can convert pointers to integers and vice-versa. Same for vectors
1694 // of pointers and integers.
1695 OldTy = OldTy->getScalarType();
1696 NewTy = NewTy->getScalarType();
1697 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1698 if (NewTy->isPointerTy() && OldTy->isPointerTy()) {
1699 return cast<PointerType>(NewTy)->getPointerAddressSpace() ==
1700 cast<PointerType>(OldTy)->getPointerAddressSpace();
1703 // We can convert integers to integral pointers, but not to non-integral
1705 if (OldTy->isIntegerTy())
1706 return !DL.isNonIntegralPointerType(NewTy);
1708 // We can convert integral pointers to integers, but non-integral pointers
1709 // need to remain pointers.
1710 if (!DL.isNonIntegralPointerType(OldTy))
1711 return NewTy->isIntegerTy();
1719 /// Generic routine to convert an SSA value to a value of a different
1722 /// This will try various different casting techniques, such as bitcasts,
1723 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1724 /// two types for viability with this routine.
1725 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1727 Type *OldTy = V->getType();
1728 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1733 assert(!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) &&
1734 "Integer types must be the exact same to convert.");
1736 // See if we need inttoptr for this type pair. A cast involving both scalars
1737 // and vectors requires and additional bitcast.
1738 if (OldTy->isIntOrIntVectorTy() && NewTy->isPtrOrPtrVectorTy()) {
1739 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1740 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1741 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1744 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1745 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1746 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1749 return IRB.CreateIntToPtr(V, NewTy);
1752 // See if we need ptrtoint for this type pair. A cast involving both scalars
1753 // and vectors requires and additional bitcast.
1754 if (OldTy->isPtrOrPtrVectorTy() && NewTy->isIntOrIntVectorTy()) {
1755 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1756 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1757 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1760 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1761 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1762 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1765 return IRB.CreatePtrToInt(V, NewTy);
1768 return IRB.CreateBitCast(V, NewTy);
1771 /// Test whether the given slice use can be promoted to a vector.
1773 /// This function is called to test each entry in a partition which is slated
1774 /// for a single slice.
1775 static bool isVectorPromotionViableForSlice(Partition &P, const Slice &S,
1777 uint64_t ElementSize,
1778 const DataLayout &DL) {
1779 // First validate the slice offsets.
1780 uint64_t BeginOffset =
1781 std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1782 uint64_t BeginIndex = BeginOffset / ElementSize;
1783 if (BeginIndex * ElementSize != BeginOffset ||
1784 BeginIndex >= Ty->getNumElements())
1786 uint64_t EndOffset =
1787 std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1788 uint64_t EndIndex = EndOffset / ElementSize;
1789 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1792 assert(EndIndex > BeginIndex && "Empty vector!");
1793 uint64_t NumElements = EndIndex - BeginIndex;
1794 Type *SliceTy = (NumElements == 1)
1795 ? Ty->getElementType()
1796 : VectorType::get(Ty->getElementType(), NumElements);
1799 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1801 Use *U = S.getUse();
1803 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1804 if (MI->isVolatile())
1806 if (!S.isSplittable())
1807 return false; // Skip any unsplittable intrinsics.
1808 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1809 if (!II->isLifetimeStartOrEnd())
1811 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1812 // Disable vector promotion when there are loads or stores of an FCA.
1814 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1815 if (LI->isVolatile())
1817 Type *LTy = LI->getType();
1818 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1819 assert(LTy->isIntegerTy());
1822 if (!canConvertValue(DL, SliceTy, LTy))
1824 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1825 if (SI->isVolatile())
1827 Type *STy = SI->getValueOperand()->getType();
1828 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1829 assert(STy->isIntegerTy());
1832 if (!canConvertValue(DL, STy, SliceTy))
1841 /// Test whether the given alloca partitioning and range of slices can be
1842 /// promoted to a vector.
1844 /// This is a quick test to check whether we can rewrite a particular alloca
1845 /// partition (and its newly formed alloca) into a vector alloca with only
1846 /// whole-vector loads and stores such that it could be promoted to a vector
1847 /// SSA value. We only can ensure this for a limited set of operations, and we
1848 /// don't want to do the rewrites unless we are confident that the result will
1849 /// be promotable, so we have an early test here.
1850 static VectorType *isVectorPromotionViable(Partition &P, const DataLayout &DL) {
1851 // Collect the candidate types for vector-based promotion. Also track whether
1852 // we have different element types.
1853 SmallVector<VectorType *, 4> CandidateTys;
1854 Type *CommonEltTy = nullptr;
1855 bool HaveCommonEltTy = true;
1856 auto CheckCandidateType = [&](Type *Ty) {
1857 if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1858 CandidateTys.push_back(VTy);
1860 CommonEltTy = VTy->getElementType();
1861 else if (CommonEltTy != VTy->getElementType())
1862 HaveCommonEltTy = false;
1865 // Consider any loads or stores that are the exact size of the slice.
1866 for (const Slice &S : P)
1867 if (S.beginOffset() == P.beginOffset() &&
1868 S.endOffset() == P.endOffset()) {
1869 if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1870 CheckCandidateType(LI->getType());
1871 else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1872 CheckCandidateType(SI->getValueOperand()->getType());
1875 // If we didn't find a vector type, nothing to do here.
1876 if (CandidateTys.empty())
1879 // Remove non-integer vector types if we had multiple common element types.
1880 // FIXME: It'd be nice to replace them with integer vector types, but we can't
1881 // do that until all the backends are known to produce good code for all
1882 // integer vector types.
1883 if (!HaveCommonEltTy) {
1885 llvm::remove_if(CandidateTys,
1886 [](VectorType *VTy) {
1887 return !VTy->getElementType()->isIntegerTy();
1889 CandidateTys.end());
1891 // If there were no integer vector types, give up.
1892 if (CandidateTys.empty())
1895 // Rank the remaining candidate vector types. This is easy because we know
1896 // they're all integer vectors. We sort by ascending number of elements.
1897 auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
1899 assert(DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
1900 "Cannot have vector types of different sizes!");
1901 assert(RHSTy->getElementType()->isIntegerTy() &&
1902 "All non-integer types eliminated!");
1903 assert(LHSTy->getElementType()->isIntegerTy() &&
1904 "All non-integer types eliminated!");
1905 return RHSTy->getNumElements() < LHSTy->getNumElements();
1907 llvm::sort(CandidateTys, RankVectorTypes);
1909 std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
1910 CandidateTys.end());
1912 // The only way to have the same element type in every vector type is to
1913 // have the same vector type. Check that and remove all but one.
1915 for (VectorType *VTy : CandidateTys) {
1916 assert(VTy->getElementType() == CommonEltTy &&
1917 "Unaccounted for element type!");
1918 assert(VTy == CandidateTys[0] &&
1919 "Different vector types with the same element type!");
1922 CandidateTys.resize(1);
1925 // Try each vector type, and return the one which works.
1926 auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
1927 uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
1929 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1930 // that aren't byte sized.
1931 if (ElementSize % 8)
1933 assert((DL.getTypeSizeInBits(VTy) % 8) == 0 &&
1934 "vector size not a multiple of element size?");
1937 for (const Slice &S : P)
1938 if (!isVectorPromotionViableForSlice(P, S, VTy, ElementSize, DL))
1941 for (const Slice *S : P.splitSliceTails())
1942 if (!isVectorPromotionViableForSlice(P, *S, VTy, ElementSize, DL))
1947 for (VectorType *VTy : CandidateTys)
1948 if (CheckVectorTypeForPromotion(VTy))
1954 /// Test whether a slice of an alloca is valid for integer widening.
1956 /// This implements the necessary checking for the \c isIntegerWideningViable
1957 /// test below on a single slice of the alloca.
1958 static bool isIntegerWideningViableForSlice(const Slice &S,
1959 uint64_t AllocBeginOffset,
1961 const DataLayout &DL,
1962 bool &WholeAllocaOp) {
1963 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1965 uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
1966 uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
1968 // We can't reasonably handle cases where the load or store extends past
1969 // the end of the alloca's type and into its padding.
1973 Use *U = S.getUse();
1975 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1976 if (LI->isVolatile())
1978 // We can't handle loads that extend past the allocated memory.
1979 if (DL.getTypeStoreSize(LI->getType()) > Size)
1981 // So far, AllocaSliceRewriter does not support widening split slice tails
1982 // in rewriteIntegerLoad.
1983 if (S.beginOffset() < AllocBeginOffset)
1985 // Note that we don't count vector loads or stores as whole-alloca
1986 // operations which enable integer widening because we would prefer to use
1987 // vector widening instead.
1988 if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
1989 WholeAllocaOp = true;
1990 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1991 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1993 } else if (RelBegin != 0 || RelEnd != Size ||
1994 !canConvertValue(DL, AllocaTy, LI->getType())) {
1995 // Non-integer loads need to be convertible from the alloca type so that
1996 // they are promotable.
1999 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
2000 Type *ValueTy = SI->getValueOperand()->getType();
2001 if (SI->isVolatile())
2003 // We can't handle stores that extend past the allocated memory.
2004 if (DL.getTypeStoreSize(ValueTy) > Size)
2006 // So far, AllocaSliceRewriter does not support widening split slice tails
2007 // in rewriteIntegerStore.
2008 if (S.beginOffset() < AllocBeginOffset)
2010 // Note that we don't count vector loads or stores as whole-alloca
2011 // operations which enable integer widening because we would prefer to use
2012 // vector widening instead.
2013 if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
2014 WholeAllocaOp = true;
2015 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2016 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
2018 } else if (RelBegin != 0 || RelEnd != Size ||
2019 !canConvertValue(DL, ValueTy, AllocaTy)) {
2020 // Non-integer stores need to be convertible to the alloca type so that
2021 // they are promotable.
2024 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
2025 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
2027 if (!S.isSplittable())
2028 return false; // Skip any unsplittable intrinsics.
2029 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
2030 if (!II->isLifetimeStartOrEnd())
2039 /// Test whether the given alloca partition's integer operations can be
2040 /// widened to promotable ones.
2042 /// This is a quick test to check whether we can rewrite the integer loads and
2043 /// stores to a particular alloca into wider loads and stores and be able to
2044 /// promote the resulting alloca.
2045 static bool isIntegerWideningViable(Partition &P, Type *AllocaTy,
2046 const DataLayout &DL) {
2047 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
2048 // Don't create integer types larger than the maximum bitwidth.
2049 if (SizeInBits > IntegerType::MAX_INT_BITS)
2052 // Don't try to handle allocas with bit-padding.
2053 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
2056 // We need to ensure that an integer type with the appropriate bitwidth can
2057 // be converted to the alloca type, whatever that is. We don't want to force
2058 // the alloca itself to have an integer type if there is a more suitable one.
2059 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2060 if (!canConvertValue(DL, AllocaTy, IntTy) ||
2061 !canConvertValue(DL, IntTy, AllocaTy))
2064 // While examining uses, we ensure that the alloca has a covering load or
2065 // store. We don't want to widen the integer operations only to fail to
2066 // promote due to some other unsplittable entry (which we may make splittable
2067 // later). However, if there are only splittable uses, go ahead and assume
2068 // that we cover the alloca.
2069 // FIXME: We shouldn't consider split slices that happen to start in the
2070 // partition here...
2071 bool WholeAllocaOp =
2072 P.begin() != P.end() ? false : DL.isLegalInteger(SizeInBits);
2074 for (const Slice &S : P)
2075 if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2079 for (const Slice *S : P.splitSliceTails())
2080 if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2084 return WholeAllocaOp;
2087 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2088 IntegerType *Ty, uint64_t Offset,
2089 const Twine &Name) {
2090 LLVM_DEBUG(dbgs() << " start: " << *V << "\n");
2091 IntegerType *IntTy = cast<IntegerType>(V->getType());
2092 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2093 "Element extends past full value");
2094 uint64_t ShAmt = 8 * Offset;
2095 if (DL.isBigEndian())
2096 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2098 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2099 LLVM_DEBUG(dbgs() << " shifted: " << *V << "\n");
2101 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2102 "Cannot extract to a larger integer!");
2104 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2105 LLVM_DEBUG(dbgs() << " trunced: " << *V << "\n");
2110 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2111 Value *V, uint64_t Offset, const Twine &Name) {
2112 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2113 IntegerType *Ty = cast<IntegerType>(V->getType());
2114 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2115 "Cannot insert a larger integer!");
2116 LLVM_DEBUG(dbgs() << " start: " << *V << "\n");
2118 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2119 LLVM_DEBUG(dbgs() << " extended: " << *V << "\n");
2121 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2122 "Element store outside of alloca store");
2123 uint64_t ShAmt = 8 * Offset;
2124 if (DL.isBigEndian())
2125 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2127 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2128 LLVM_DEBUG(dbgs() << " shifted: " << *V << "\n");
2131 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2132 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2133 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2134 LLVM_DEBUG(dbgs() << " masked: " << *Old << "\n");
2135 V = IRB.CreateOr(Old, V, Name + ".insert");
2136 LLVM_DEBUG(dbgs() << " inserted: " << *V << "\n");
2141 static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2142 unsigned EndIndex, const Twine &Name) {
2143 VectorType *VecTy = cast<VectorType>(V->getType());
2144 unsigned NumElements = EndIndex - BeginIndex;
2145 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2147 if (NumElements == VecTy->getNumElements())
2150 if (NumElements == 1) {
2151 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2153 LLVM_DEBUG(dbgs() << " extract: " << *V << "\n");
2157 SmallVector<Constant *, 8> Mask;
2158 Mask.reserve(NumElements);
2159 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2160 Mask.push_back(IRB.getInt32(i));
2161 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2162 ConstantVector::get(Mask), Name + ".extract");
2163 LLVM_DEBUG(dbgs() << " shuffle: " << *V << "\n");
2167 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2168 unsigned BeginIndex, const Twine &Name) {
2169 VectorType *VecTy = cast<VectorType>(Old->getType());
2170 assert(VecTy && "Can only insert a vector into a vector");
2172 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2174 // Single element to insert.
2175 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2177 LLVM_DEBUG(dbgs() << " insert: " << *V << "\n");
2181 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2182 "Too many elements!");
2183 if (Ty->getNumElements() == VecTy->getNumElements()) {
2184 assert(V->getType() == VecTy && "Vector type mismatch");
2187 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2189 // When inserting a smaller vector into the larger to store, we first
2190 // use a shuffle vector to widen it with undef elements, and then
2191 // a second shuffle vector to select between the loaded vector and the
2193 SmallVector<Constant *, 8> Mask;
2194 Mask.reserve(VecTy->getNumElements());
2195 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2196 if (i >= BeginIndex && i < EndIndex)
2197 Mask.push_back(IRB.getInt32(i - BeginIndex));
2199 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2200 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2201 ConstantVector::get(Mask), Name + ".expand");
2202 LLVM_DEBUG(dbgs() << " shuffle: " << *V << "\n");
2205 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2206 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2208 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2210 LLVM_DEBUG(dbgs() << " blend: " << *V << "\n");
2214 /// Visitor to rewrite instructions using p particular slice of an alloca
2215 /// to use a new alloca.
2217 /// Also implements the rewriting to vector-based accesses when the partition
2218 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2220 class llvm::sroa::AllocaSliceRewriter
2221 : public InstVisitor<AllocaSliceRewriter, bool> {
2222 // Befriend the base class so it can delegate to private visit methods.
2223 friend class InstVisitor<AllocaSliceRewriter, bool>;
2225 using Base = InstVisitor<AllocaSliceRewriter, bool>;
2227 const DataLayout &DL;
2230 AllocaInst &OldAI, &NewAI;
2231 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2234 // This is a convenience and flag variable that will be null unless the new
2235 // alloca's integer operations should be widened to this integer type due to
2236 // passing isIntegerWideningViable above. If it is non-null, the desired
2237 // integer type will be stored here for easy access during rewriting.
2240 // If we are rewriting an alloca partition which can be written as pure
2241 // vector operations, we stash extra information here. When VecTy is
2242 // non-null, we have some strict guarantees about the rewritten alloca:
2243 // - The new alloca is exactly the size of the vector type here.
2244 // - The accesses all either map to the entire vector or to a single
2246 // - The set of accessing instructions is only one of those handled above
2247 // in isVectorPromotionViable. Generally these are the same access kinds
2248 // which are promotable via mem2reg.
2251 uint64_t ElementSize;
2253 // The original offset of the slice currently being rewritten relative to
2254 // the original alloca.
2255 uint64_t BeginOffset = 0;
2256 uint64_t EndOffset = 0;
2258 // The new offsets of the slice currently being rewritten relative to the
2260 uint64_t NewBeginOffset, NewEndOffset;
2263 bool IsSplittable = false;
2264 bool IsSplit = false;
2265 Use *OldUse = nullptr;
2266 Instruction *OldPtr = nullptr;
2268 // Track post-rewrite users which are PHI nodes and Selects.
2269 SmallSetVector<PHINode *, 8> &PHIUsers;
2270 SmallSetVector<SelectInst *, 8> &SelectUsers;
2272 // Utility IR builder, whose name prefix is setup for each visited use, and
2273 // the insertion point is set to point to the user.
2277 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROA &Pass,
2278 AllocaInst &OldAI, AllocaInst &NewAI,
2279 uint64_t NewAllocaBeginOffset,
2280 uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2281 VectorType *PromotableVecTy,
2282 SmallSetVector<PHINode *, 8> &PHIUsers,
2283 SmallSetVector<SelectInst *, 8> &SelectUsers)
2284 : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2285 NewAllocaBeginOffset(NewAllocaBeginOffset),
2286 NewAllocaEndOffset(NewAllocaEndOffset),
2287 NewAllocaTy(NewAI.getAllocatedType()),
2288 IntTy(IsIntegerPromotable
2291 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2293 VecTy(PromotableVecTy),
2294 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2295 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2296 PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2297 IRB(NewAI.getContext(), ConstantFolder()) {
2299 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2300 "Only multiple-of-8 sized vector elements are viable");
2303 assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy));
2306 bool visit(AllocaSlices::const_iterator I) {
2307 bool CanSROA = true;
2308 BeginOffset = I->beginOffset();
2309 EndOffset = I->endOffset();
2310 IsSplittable = I->isSplittable();
2312 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2313 LLVM_DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""));
2314 LLVM_DEBUG(AS.printSlice(dbgs(), I, ""));
2315 LLVM_DEBUG(dbgs() << "\n");
2317 // Compute the intersecting offset range.
2318 assert(BeginOffset < NewAllocaEndOffset);
2319 assert(EndOffset > NewAllocaBeginOffset);
2320 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2321 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2323 SliceSize = NewEndOffset - NewBeginOffset;
2325 OldUse = I->getUse();
2326 OldPtr = cast<Instruction>(OldUse->get());
2328 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2329 IRB.SetInsertPoint(OldUserI);
2330 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2331 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2333 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2340 // Make sure the other visit overloads are visible.
2343 // Every instruction which can end up as a user must have a rewrite rule.
2344 bool visitInstruction(Instruction &I) {
2345 LLVM_DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2346 llvm_unreachable("No rewrite rule for this instruction!");
2349 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2350 // Note that the offset computation can use BeginOffset or NewBeginOffset
2351 // interchangeably for unsplit slices.
2352 assert(IsSplit || BeginOffset == NewBeginOffset);
2353 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2356 StringRef OldName = OldPtr->getName();
2357 // Skip through the last '.sroa.' component of the name.
2358 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2359 if (LastSROAPrefix != StringRef::npos) {
2360 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2361 // Look for an SROA slice index.
2362 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2363 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2364 // Strip the index and look for the offset.
2365 OldName = OldName.substr(IndexEnd + 1);
2366 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2367 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2368 // Strip the offset.
2369 OldName = OldName.substr(OffsetEnd + 1);
2372 // Strip any SROA suffixes as well.
2373 OldName = OldName.substr(0, OldName.find(".sroa_"));
2376 return getAdjustedPtr(IRB, DL, &NewAI,
2377 APInt(DL.getIndexTypeSizeInBits(PointerTy), Offset),
2380 Twine(OldName) + "."
2387 /// Compute suitable alignment to access this slice of the *new*
2390 /// You can optionally pass a type to this routine and if that type's ABI
2391 /// alignment is itself suitable, this will return zero.
2392 unsigned getSliceAlign(Type *Ty = nullptr) {
2393 unsigned NewAIAlign = NewAI.getAlignment();
2395 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2397 MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2398 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2401 unsigned getIndex(uint64_t Offset) {
2402 assert(VecTy && "Can only call getIndex when rewriting a vector");
2403 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2404 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2405 uint32_t Index = RelOffset / ElementSize;
2406 assert(Index * ElementSize == RelOffset);
2410 void deleteIfTriviallyDead(Value *V) {
2411 Instruction *I = cast<Instruction>(V);
2412 if (isInstructionTriviallyDead(I))
2413 Pass.DeadInsts.insert(I);
2416 Value *rewriteVectorizedLoadInst() {
2417 unsigned BeginIndex = getIndex(NewBeginOffset);
2418 unsigned EndIndex = getIndex(NewEndOffset);
2419 assert(EndIndex > BeginIndex && "Empty vector!");
2421 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2422 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2425 Value *rewriteIntegerLoad(LoadInst &LI) {
2426 assert(IntTy && "We cannot insert an integer to the alloca");
2427 assert(!LI.isVolatile());
2428 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2429 V = convertValue(DL, IRB, V, IntTy);
2430 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2431 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2432 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset) {
2433 IntegerType *ExtractTy = Type::getIntNTy(LI.getContext(), SliceSize * 8);
2434 V = extractInteger(DL, IRB, V, ExtractTy, Offset, "extract");
2436 // It is possible that the extracted type is not the load type. This
2437 // happens if there is a load past the end of the alloca, and as
2438 // a consequence the slice is narrower but still a candidate for integer
2439 // lowering. To handle this case, we just zero extend the extracted
2441 assert(cast<IntegerType>(LI.getType())->getBitWidth() >= SliceSize * 8 &&
2442 "Can only handle an extract for an overly wide load");
2443 if (cast<IntegerType>(LI.getType())->getBitWidth() > SliceSize * 8)
2444 V = IRB.CreateZExt(V, LI.getType());
2448 bool visitLoadInst(LoadInst &LI) {
2449 LLVM_DEBUG(dbgs() << " original: " << LI << "\n");
2450 Value *OldOp = LI.getOperand(0);
2451 assert(OldOp == OldPtr);
2454 LI.getAAMetadata(AATags);
2456 unsigned AS = LI.getPointerAddressSpace();
2458 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2460 const bool IsLoadPastEnd = DL.getTypeStoreSize(TargetTy) > SliceSize;
2461 bool IsPtrAdjusted = false;
2464 V = rewriteVectorizedLoadInst();
2465 } else if (IntTy && LI.getType()->isIntegerTy()) {
2466 V = rewriteIntegerLoad(LI);
2467 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2468 NewEndOffset == NewAllocaEndOffset &&
2469 (canConvertValue(DL, NewAllocaTy, TargetTy) ||
2470 (IsLoadPastEnd && NewAllocaTy->isIntegerTy() &&
2471 TargetTy->isIntegerTy()))) {
2472 LoadInst *NewLI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2473 LI.isVolatile(), LI.getName());
2475 NewLI->setAAMetadata(AATags);
2476 if (LI.isVolatile())
2477 NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2479 // Any !nonnull metadata or !range metadata on the old load is also valid
2480 // on the new load. This is even true in some cases even when the loads
2481 // are different types, for example by mapping !nonnull metadata to
2482 // !range metadata by modeling the null pointer constant converted to the
2484 // FIXME: Add support for range metadata here. Currently the utilities
2485 // for this don't propagate range metadata in trivial cases from one
2486 // integer load to another, don't handle non-addrspace-0 null pointers
2487 // correctly, and don't have any support for mapping ranges as the
2488 // integer type becomes winder or narrower.
2489 if (MDNode *N = LI.getMetadata(LLVMContext::MD_nonnull))
2490 copyNonnullMetadata(LI, N, *NewLI);
2492 // Try to preserve nonnull metadata
2495 // If this is an integer load past the end of the slice (which means the
2496 // bytes outside the slice are undef or this load is dead) just forcibly
2497 // fix the integer size with correct handling of endianness.
2498 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2499 if (auto *TITy = dyn_cast<IntegerType>(TargetTy))
2500 if (AITy->getBitWidth() < TITy->getBitWidth()) {
2501 V = IRB.CreateZExt(V, TITy, "load.ext");
2502 if (DL.isBigEndian())
2503 V = IRB.CreateShl(V, TITy->getBitWidth() - AITy->getBitWidth(),
2507 Type *LTy = TargetTy->getPointerTo(AS);
2508 LoadInst *NewLI = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2509 getSliceAlign(TargetTy),
2510 LI.isVolatile(), LI.getName());
2512 NewLI->setAAMetadata(AATags);
2513 if (LI.isVolatile())
2514 NewLI->setAtomic(LI.getOrdering(), LI.getSyncScopeID());
2517 IsPtrAdjusted = true;
2519 V = convertValue(DL, IRB, V, TargetTy);
2522 assert(!LI.isVolatile());
2523 assert(LI.getType()->isIntegerTy() &&
2524 "Only integer type loads and stores are split");
2525 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2526 "Split load isn't smaller than original load");
2527 assert(LI.getType()->getIntegerBitWidth() ==
2528 DL.getTypeStoreSizeInBits(LI.getType()) &&
2529 "Non-byte-multiple bit width");
2530 // Move the insertion point just past the load so that we can refer to it.
2531 IRB.SetInsertPoint(&*std::next(BasicBlock::iterator(&LI)));
2532 // Create a placeholder value with the same type as LI to use as the
2533 // basis for the new value. This allows us to replace the uses of LI with
2534 // the computed value, and then replace the placeholder with LI, leaving
2535 // LI only used for this computation.
2536 Value *Placeholder =
2537 new LoadInst(UndefValue::get(LI.getType()->getPointerTo(AS)));
2538 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset - BeginOffset,
2540 LI.replaceAllUsesWith(V);
2541 Placeholder->replaceAllUsesWith(&LI);
2542 Placeholder->deleteValue();
2544 LI.replaceAllUsesWith(V);
2547 Pass.DeadInsts.insert(&LI);
2548 deleteIfTriviallyDead(OldOp);
2549 LLVM_DEBUG(dbgs() << " to: " << *V << "\n");
2550 return !LI.isVolatile() && !IsPtrAdjusted;
2553 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp,
2555 if (V->getType() != VecTy) {
2556 unsigned BeginIndex = getIndex(NewBeginOffset);
2557 unsigned EndIndex = getIndex(NewEndOffset);
2558 assert(EndIndex > BeginIndex && "Empty vector!");
2559 unsigned NumElements = EndIndex - BeginIndex;
2560 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2561 Type *SliceTy = (NumElements == 1)
2563 : VectorType::get(ElementTy, NumElements);
2564 if (V->getType() != SliceTy)
2565 V = convertValue(DL, IRB, V, SliceTy);
2567 // Mix in the existing elements.
2568 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2569 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2571 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2573 Store->setAAMetadata(AATags);
2574 Pass.DeadInsts.insert(&SI);
2576 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n");
2580 bool rewriteIntegerStore(Value *V, StoreInst &SI, AAMDNodes AATags) {
2581 assert(IntTy && "We cannot extract an integer from the alloca");
2582 assert(!SI.isVolatile());
2583 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2585 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2586 Old = convertValue(DL, IRB, Old, IntTy);
2587 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2588 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2589 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2591 V = convertValue(DL, IRB, V, NewAllocaTy);
2592 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2593 Store->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
2594 LLVMContext::MD_access_group});
2596 Store->setAAMetadata(AATags);
2597 Pass.DeadInsts.insert(&SI);
2598 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n");
2602 bool visitStoreInst(StoreInst &SI) {
2603 LLVM_DEBUG(dbgs() << " original: " << SI << "\n");
2604 Value *OldOp = SI.getOperand(1);
2605 assert(OldOp == OldPtr);
2608 SI.getAAMetadata(AATags);
2610 Value *V = SI.getValueOperand();
2612 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2613 // alloca that should be re-examined after promoting this alloca.
2614 if (V->getType()->isPointerTy())
2615 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2616 Pass.PostPromotionWorklist.insert(AI);
2618 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2619 assert(!SI.isVolatile());
2620 assert(V->getType()->isIntegerTy() &&
2621 "Only integer type loads and stores are split");
2622 assert(V->getType()->getIntegerBitWidth() ==
2623 DL.getTypeStoreSizeInBits(V->getType()) &&
2624 "Non-byte-multiple bit width");
2625 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2626 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset - BeginOffset,
2631 return rewriteVectorizedStoreInst(V, SI, OldOp, AATags);
2632 if (IntTy && V->getType()->isIntegerTy())
2633 return rewriteIntegerStore(V, SI, AATags);
2635 const bool IsStorePastEnd = DL.getTypeStoreSize(V->getType()) > SliceSize;
2637 if (NewBeginOffset == NewAllocaBeginOffset &&
2638 NewEndOffset == NewAllocaEndOffset &&
2639 (canConvertValue(DL, V->getType(), NewAllocaTy) ||
2640 (IsStorePastEnd && NewAllocaTy->isIntegerTy() &&
2641 V->getType()->isIntegerTy()))) {
2642 // If this is an integer store past the end of slice (and thus the bytes
2643 // past that point are irrelevant or this is unreachable), truncate the
2644 // value prior to storing.
2645 if (auto *VITy = dyn_cast<IntegerType>(V->getType()))
2646 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2647 if (VITy->getBitWidth() > AITy->getBitWidth()) {
2648 if (DL.isBigEndian())
2649 V = IRB.CreateLShr(V, VITy->getBitWidth() - AITy->getBitWidth(),
2651 V = IRB.CreateTrunc(V, AITy, "load.trunc");
2654 V = convertValue(DL, IRB, V, NewAllocaTy);
2655 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2658 unsigned AS = SI.getPointerAddressSpace();
2659 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo(AS));
2660 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2663 NewSI->copyMetadata(SI, {LLVMContext::MD_mem_parallel_loop_access,
2664 LLVMContext::MD_access_group});
2666 NewSI->setAAMetadata(AATags);
2667 if (SI.isVolatile())
2668 NewSI->setAtomic(SI.getOrdering(), SI.getSyncScopeID());
2669 Pass.DeadInsts.insert(&SI);
2670 deleteIfTriviallyDead(OldOp);
2672 LLVM_DEBUG(dbgs() << " to: " << *NewSI << "\n");
2673 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2676 /// Compute an integer value from splatting an i8 across the given
2677 /// number of bytes.
2679 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2680 /// call this routine.
2681 /// FIXME: Heed the advice above.
2683 /// \param V The i8 value to splat.
2684 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2685 Value *getIntegerSplat(Value *V, unsigned Size) {
2686 assert(Size > 0 && "Expected a positive number of bytes.");
2687 IntegerType *VTy = cast<IntegerType>(V->getType());
2688 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2692 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2694 IRB.CreateZExt(V, SplatIntTy, "zext"),
2695 ConstantExpr::getUDiv(
2696 Constant::getAllOnesValue(SplatIntTy),
2697 ConstantExpr::getZExt(Constant::getAllOnesValue(V->getType()),
2703 /// Compute a vector splat for a given element value.
2704 Value *getVectorSplat(Value *V, unsigned NumElements) {
2705 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2706 LLVM_DEBUG(dbgs() << " splat: " << *V << "\n");
2710 bool visitMemSetInst(MemSetInst &II) {
2711 LLVM_DEBUG(dbgs() << " original: " << II << "\n");
2712 assert(II.getRawDest() == OldPtr);
2715 II.getAAMetadata(AATags);
2717 // If the memset has a variable size, it cannot be split, just adjust the
2718 // pointer to the new alloca.
2719 if (!isa<Constant>(II.getLength())) {
2721 assert(NewBeginOffset == BeginOffset);
2722 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2723 II.setDestAlignment(getSliceAlign());
2725 deleteIfTriviallyDead(OldPtr);
2729 // Record this instruction for deletion.
2730 Pass.DeadInsts.insert(&II);
2732 Type *AllocaTy = NewAI.getAllocatedType();
2733 Type *ScalarTy = AllocaTy->getScalarType();
2735 // If this doesn't map cleanly onto the alloca type, and that type isn't
2736 // a single value type, just emit a memset.
2737 if (!VecTy && !IntTy &&
2738 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2739 SliceSize != DL.getTypeStoreSize(AllocaTy) ||
2740 !AllocaTy->isSingleValueType() ||
2741 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2742 DL.getTypeSizeInBits(ScalarTy) % 8 != 0)) {
2743 Type *SizeTy = II.getLength()->getType();
2744 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2745 CallInst *New = IRB.CreateMemSet(
2746 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2747 getSliceAlign(), II.isVolatile());
2749 New->setAAMetadata(AATags);
2750 LLVM_DEBUG(dbgs() << " to: " << *New << "\n");
2754 // If we can represent this as a simple value, we have to build the actual
2755 // value to store, which requires expanding the byte present in memset to
2756 // a sensible representation for the alloca type. This is essentially
2757 // splatting the byte to a sufficiently wide integer, splatting it across
2758 // any desired vector width, and bitcasting to the final type.
2762 // If this is a memset of a vectorized alloca, insert it.
2763 assert(ElementTy == ScalarTy);
2765 unsigned BeginIndex = getIndex(NewBeginOffset);
2766 unsigned EndIndex = getIndex(NewEndOffset);
2767 assert(EndIndex > BeginIndex && "Empty vector!");
2768 unsigned NumElements = EndIndex - BeginIndex;
2769 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2772 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2773 Splat = convertValue(DL, IRB, Splat, ElementTy);
2774 if (NumElements > 1)
2775 Splat = getVectorSplat(Splat, NumElements);
2778 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2779 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2781 // If this is a memset on an alloca where we can widen stores, insert the
2783 assert(!II.isVolatile());
2785 uint64_t Size = NewEndOffset - NewBeginOffset;
2786 V = getIntegerSplat(II.getValue(), Size);
2788 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2789 EndOffset != NewAllocaBeginOffset)) {
2791 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2792 Old = convertValue(DL, IRB, Old, IntTy);
2793 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2794 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2796 assert(V->getType() == IntTy &&
2797 "Wrong type for an alloca wide integer!");
2799 V = convertValue(DL, IRB, V, AllocaTy);
2801 // Established these invariants above.
2802 assert(NewBeginOffset == NewAllocaBeginOffset);
2803 assert(NewEndOffset == NewAllocaEndOffset);
2805 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2806 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2807 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2809 V = convertValue(DL, IRB, V, AllocaTy);
2812 StoreInst *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2815 New->setAAMetadata(AATags);
2816 LLVM_DEBUG(dbgs() << " to: " << *New << "\n");
2817 return !II.isVolatile();
2820 bool visitMemTransferInst(MemTransferInst &II) {
2821 // Rewriting of memory transfer instructions can be a bit tricky. We break
2822 // them into two categories: split intrinsics and unsplit intrinsics.
2824 LLVM_DEBUG(dbgs() << " original: " << II << "\n");
2827 II.getAAMetadata(AATags);
2829 bool IsDest = &II.getRawDestUse() == OldUse;
2830 assert((IsDest && II.getRawDest() == OldPtr) ||
2831 (!IsDest && II.getRawSource() == OldPtr));
2833 unsigned SliceAlign = getSliceAlign();
2835 // For unsplit intrinsics, we simply modify the source and destination
2836 // pointers in place. This isn't just an optimization, it is a matter of
2837 // correctness. With unsplit intrinsics we may be dealing with transfers
2838 // within a single alloca before SROA ran, or with transfers that have
2839 // a variable length. We may also be dealing with memmove instead of
2840 // memcpy, and so simply updating the pointers is the necessary for us to
2841 // update both source and dest of a single call.
2842 if (!IsSplittable) {
2843 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2845 II.setDest(AdjustedPtr);
2846 II.setDestAlignment(SliceAlign);
2849 II.setSource(AdjustedPtr);
2850 II.setSourceAlignment(SliceAlign);
2853 LLVM_DEBUG(dbgs() << " to: " << II << "\n");
2854 deleteIfTriviallyDead(OldPtr);
2857 // For split transfer intrinsics we have an incredibly useful assurance:
2858 // the source and destination do not reside within the same alloca, and at
2859 // least one of them does not escape. This means that we can replace
2860 // memmove with memcpy, and we don't need to worry about all manner of
2861 // downsides to splitting and transforming the operations.
2863 // If this doesn't map cleanly onto the alloca type, and that type isn't
2864 // a single value type, just emit a memcpy.
2867 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2868 SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2869 !NewAI.getAllocatedType()->isSingleValueType());
2871 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2872 // size hasn't been shrunk based on analysis of the viable range, this is
2874 if (EmitMemCpy && &OldAI == &NewAI) {
2875 // Ensure the start lines up.
2876 assert(NewBeginOffset == BeginOffset);
2878 // Rewrite the size as needed.
2879 if (NewEndOffset != EndOffset)
2880 II.setLength(ConstantInt::get(II.getLength()->getType(),
2881 NewEndOffset - NewBeginOffset));
2884 // Record this instruction for deletion.
2885 Pass.DeadInsts.insert(&II);
2887 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2888 // alloca that should be re-examined after rewriting this instruction.
2889 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2890 if (AllocaInst *AI =
2891 dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2892 assert(AI != &OldAI && AI != &NewAI &&
2893 "Splittable transfers cannot reach the same alloca on both ends.");
2894 Pass.Worklist.insert(AI);
2897 Type *OtherPtrTy = OtherPtr->getType();
2898 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2900 // Compute the relative offset for the other pointer within the transfer.
2901 unsigned OffsetWidth = DL.getIndexSizeInBits(OtherAS);
2902 APInt OtherOffset(OffsetWidth, NewBeginOffset - BeginOffset);
2903 unsigned OtherAlign =
2904 IsDest ? II.getSourceAlignment() : II.getDestAlignment();
2905 OtherAlign = MinAlign(OtherAlign ? OtherAlign : 1,
2906 OtherOffset.zextOrTrunc(64).getZExtValue());
2909 // Compute the other pointer, folding as much as possible to produce
2910 // a single, simple GEP in most cases.
2911 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2912 OtherPtr->getName() + ".");
2914 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2915 Type *SizeTy = II.getLength()->getType();
2916 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2918 Value *DestPtr, *SrcPtr;
2919 unsigned DestAlign, SrcAlign;
2920 // Note: IsDest is true iff we're copying into the new alloca slice
2923 DestAlign = SliceAlign;
2925 SrcAlign = OtherAlign;
2928 DestAlign = OtherAlign;
2930 SrcAlign = SliceAlign;
2932 CallInst *New = IRB.CreateMemCpy(DestPtr, DestAlign, SrcPtr, SrcAlign,
2933 Size, II.isVolatile());
2935 New->setAAMetadata(AATags);
2936 LLVM_DEBUG(dbgs() << " to: " << *New << "\n");
2940 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2941 NewEndOffset == NewAllocaEndOffset;
2942 uint64_t Size = NewEndOffset - NewBeginOffset;
2943 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2944 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2945 unsigned NumElements = EndIndex - BeginIndex;
2946 IntegerType *SubIntTy =
2947 IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
2949 // Reset the other pointer type to match the register type we're going to
2950 // use, but using the address space of the original other pointer.
2951 if (VecTy && !IsWholeAlloca) {
2952 if (NumElements == 1)
2953 OtherPtrTy = VecTy->getElementType();
2955 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2957 OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2958 } else if (IntTy && !IsWholeAlloca) {
2959 OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2961 OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2964 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2965 OtherPtr->getName() + ".");
2966 unsigned SrcAlign = OtherAlign;
2967 Value *DstPtr = &NewAI;
2968 unsigned DstAlign = SliceAlign;
2970 std::swap(SrcPtr, DstPtr);
2971 std::swap(SrcAlign, DstAlign);
2975 if (VecTy && !IsWholeAlloca && !IsDest) {
2976 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2977 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2978 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2979 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2980 Src = convertValue(DL, IRB, Src, IntTy);
2981 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2982 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2984 LoadInst *Load = IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(),
2987 Load->setAAMetadata(AATags);
2991 if (VecTy && !IsWholeAlloca && IsDest) {
2993 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2994 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2995 } else if (IntTy && !IsWholeAlloca && IsDest) {
2997 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2998 Old = convertValue(DL, IRB, Old, IntTy);
2999 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
3000 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
3001 Src = convertValue(DL, IRB, Src, NewAllocaTy);
3004 StoreInst *Store = cast<StoreInst>(
3005 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
3007 Store->setAAMetadata(AATags);
3008 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n");
3009 return !II.isVolatile();
3012 bool visitIntrinsicInst(IntrinsicInst &II) {
3013 assert(II.isLifetimeStartOrEnd());
3014 LLVM_DEBUG(dbgs() << " original: " << II << "\n");
3015 assert(II.getArgOperand(1) == OldPtr);
3017 // Record this instruction for deletion.
3018 Pass.DeadInsts.insert(&II);
3020 // Lifetime intrinsics are only promotable if they cover the whole alloca.
3021 // Therefore, we drop lifetime intrinsics which don't cover the whole
3023 // (In theory, intrinsics which partially cover an alloca could be
3024 // promoted, but PromoteMemToReg doesn't handle that case.)
3025 // FIXME: Check whether the alloca is promotable before dropping the
3026 // lifetime intrinsics?
3027 if (NewBeginOffset != NewAllocaBeginOffset ||
3028 NewEndOffset != NewAllocaEndOffset)
3032 ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
3033 NewEndOffset - NewBeginOffset);
3034 // Lifetime intrinsics always expect an i8* so directly get such a pointer
3035 // for the new alloca slice.
3036 Type *PointerTy = IRB.getInt8PtrTy(OldPtr->getType()->getPointerAddressSpace());
3037 Value *Ptr = getNewAllocaSlicePtr(IRB, PointerTy);
3039 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
3040 New = IRB.CreateLifetimeStart(Ptr, Size);
3042 New = IRB.CreateLifetimeEnd(Ptr, Size);
3045 LLVM_DEBUG(dbgs() << " to: " << *New << "\n");
3050 void fixLoadStoreAlign(Instruction &Root) {
3051 // This algorithm implements the same visitor loop as
3052 // hasUnsafePHIOrSelectUse, and fixes the alignment of each load
3054 SmallPtrSet<Instruction *, 4> Visited;
3055 SmallVector<Instruction *, 4> Uses;
3056 Visited.insert(&Root);
3057 Uses.push_back(&Root);
3059 Instruction *I = Uses.pop_back_val();
3061 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3062 unsigned LoadAlign = LI->getAlignment();
3064 LoadAlign = DL.getABITypeAlignment(LI->getType());
3065 LI->setAlignment(std::min(LoadAlign, getSliceAlign()));
3068 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3069 unsigned StoreAlign = SI->getAlignment();
3071 Value *Op = SI->getOperand(0);
3072 StoreAlign = DL.getABITypeAlignment(Op->getType());
3074 SI->setAlignment(std::min(StoreAlign, getSliceAlign()));
3078 assert(isa<BitCastInst>(I) || isa<PHINode>(I) ||
3079 isa<SelectInst>(I) || isa<GetElementPtrInst>(I));
3080 for (User *U : I->users())
3081 if (Visited.insert(cast<Instruction>(U)).second)
3082 Uses.push_back(cast<Instruction>(U));
3083 } while (!Uses.empty());
3086 bool visitPHINode(PHINode &PN) {
3087 LLVM_DEBUG(dbgs() << " original: " << PN << "\n");
3088 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
3089 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
3091 // We would like to compute a new pointer in only one place, but have it be
3092 // as local as possible to the PHI. To do that, we re-use the location of
3093 // the old pointer, which necessarily must be in the right position to
3094 // dominate the PHI.
3095 IRBuilderTy PtrBuilder(IRB);
3096 if (isa<PHINode>(OldPtr))
3097 PtrBuilder.SetInsertPoint(&*OldPtr->getParent()->getFirstInsertionPt());
3099 PtrBuilder.SetInsertPoint(OldPtr);
3100 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
3102 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
3103 // Replace the operands which were using the old pointer.
3104 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
3106 LLVM_DEBUG(dbgs() << " to: " << PN << "\n");
3107 deleteIfTriviallyDead(OldPtr);
3109 // Fix the alignment of any loads or stores using this PHI node.
3110 fixLoadStoreAlign(PN);
3112 // PHIs can't be promoted on their own, but often can be speculated. We
3113 // check the speculation outside of the rewriter so that we see the
3114 // fully-rewritten alloca.
3115 PHIUsers.insert(&PN);
3119 bool visitSelectInst(SelectInst &SI) {
3120 LLVM_DEBUG(dbgs() << " original: " << SI << "\n");
3121 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
3122 "Pointer isn't an operand!");
3123 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
3124 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
3126 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3127 // Replace the operands which were using the old pointer.
3128 if (SI.getOperand(1) == OldPtr)
3129 SI.setOperand(1, NewPtr);
3130 if (SI.getOperand(2) == OldPtr)
3131 SI.setOperand(2, NewPtr);
3133 LLVM_DEBUG(dbgs() << " to: " << SI << "\n");
3134 deleteIfTriviallyDead(OldPtr);
3136 // Fix the alignment of any loads or stores using this select.
3137 fixLoadStoreAlign(SI);
3139 // Selects can't be promoted on their own, but often can be speculated. We
3140 // check the speculation outside of the rewriter so that we see the
3141 // fully-rewritten alloca.
3142 SelectUsers.insert(&SI);
3149 /// Visitor to rewrite aggregate loads and stores as scalar.
3151 /// This pass aggressively rewrites all aggregate loads and stores on
3152 /// a particular pointer (or any pointer derived from it which we can identify)
3153 /// with scalar loads and stores.
3154 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3155 // Befriend the base class so it can delegate to private visit methods.
3156 friend class InstVisitor<AggLoadStoreRewriter, bool>;
3158 /// Queue of pointer uses to analyze and potentially rewrite.
3159 SmallVector<Use *, 8> Queue;
3161 /// Set to prevent us from cycling with phi nodes and loops.
3162 SmallPtrSet<User *, 8> Visited;
3164 /// The current pointer use being rewritten. This is used to dig up the used
3165 /// value (as opposed to the user).
3168 /// Used to calculate offsets, and hence alignment, of subobjects.
3169 const DataLayout &DL;
3172 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
3174 /// Rewrite loads and stores through a pointer and all pointers derived from
3176 bool rewrite(Instruction &I) {
3177 LLVM_DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
3179 bool Changed = false;
3180 while (!Queue.empty()) {
3181 U = Queue.pop_back_val();
3182 Changed |= visit(cast<Instruction>(U->getUser()));
3188 /// Enqueue all the users of the given instruction for further processing.
3189 /// This uses a set to de-duplicate users.
3190 void enqueueUsers(Instruction &I) {
3191 for (Use &U : I.uses())
3192 if (Visited.insert(U.getUser()).second)
3193 Queue.push_back(&U);
3196 // Conservative default is to not rewrite anything.
3197 bool visitInstruction(Instruction &I) { return false; }
3199 /// Generic recursive split emission class.
3200 template <typename Derived> class OpSplitter {
3202 /// The builder used to form new instructions.
3205 /// The indices which to be used with insert- or extractvalue to select the
3206 /// appropriate value within the aggregate.
3207 SmallVector<unsigned, 4> Indices;
3209 /// The indices to a GEP instruction which will move Ptr to the correct slot
3210 /// within the aggregate.
3211 SmallVector<Value *, 4> GEPIndices;
3213 /// The base pointer of the original op, used as a base for GEPing the
3214 /// split operations.
3217 /// The base pointee type being GEPed into.
3220 /// Known alignment of the base pointer.
3223 /// To calculate offset of each component so we can correctly deduce
3225 const DataLayout &DL;
3227 /// Initialize the splitter with an insertion point, Ptr and start with a
3228 /// single zero GEP index.
3229 OpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3230 unsigned BaseAlign, const DataLayout &DL)
3231 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr),
3232 BaseTy(BaseTy), BaseAlign(BaseAlign), DL(DL) {}
3235 /// Generic recursive split emission routine.
3237 /// This method recursively splits an aggregate op (load or store) into
3238 /// scalar or vector ops. It splits recursively until it hits a single value
3239 /// and emits that single value operation via the template argument.
3241 /// The logic of this routine relies on GEPs and insertvalue and
3242 /// extractvalue all operating with the same fundamental index list, merely
3243 /// formatted differently (GEPs need actual values).
3245 /// \param Ty The type being split recursively into smaller ops.
3246 /// \param Agg The aggregate value being built up or stored, depending on
3247 /// whether this is splitting a load or a store respectively.
3248 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3249 if (Ty->isSingleValueType()) {
3250 unsigned Offset = DL.getIndexedOffsetInType(BaseTy, GEPIndices);
3251 return static_cast<Derived *>(this)->emitFunc(
3252 Ty, Agg, MinAlign(BaseAlign, Offset), Name);
3255 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3256 unsigned OldSize = Indices.size();
3258 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3260 assert(Indices.size() == OldSize && "Did not return to the old size");
3261 Indices.push_back(Idx);
3262 GEPIndices.push_back(IRB.getInt32(Idx));
3263 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3264 GEPIndices.pop_back();
3270 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3271 unsigned OldSize = Indices.size();
3273 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3275 assert(Indices.size() == OldSize && "Did not return to the old size");
3276 Indices.push_back(Idx);
3277 GEPIndices.push_back(IRB.getInt32(Idx));
3278 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3279 GEPIndices.pop_back();
3285 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3289 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3292 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3293 AAMDNodes AATags, unsigned BaseAlign, const DataLayout &DL)
3294 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr, BaseTy, BaseAlign,
3295 DL), AATags(AATags) {}
3297 /// Emit a leaf load of a single value. This is called at the leaves of the
3298 /// recursive emission to actually load values.
3299 void emitFunc(Type *Ty, Value *&Agg, unsigned Align, const Twine &Name) {
3300 assert(Ty->isSingleValueType());
3301 // Load the single value and insert it using the indices.
3303 IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3304 LoadInst *Load = IRB.CreateAlignedLoad(GEP, Align, Name + ".load");
3306 Load->setAAMetadata(AATags);
3307 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3308 LLVM_DEBUG(dbgs() << " to: " << *Load << "\n");
3312 bool visitLoadInst(LoadInst &LI) {
3313 assert(LI.getPointerOperand() == *U);
3314 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3317 // We have an aggregate being loaded, split it apart.
3318 LLVM_DEBUG(dbgs() << " original: " << LI << "\n");
3320 LI.getAAMetadata(AATags);
3321 LoadOpSplitter Splitter(&LI, *U, LI.getType(), AATags,
3322 getAdjustedAlignment(&LI, 0, DL), DL);
3323 Value *V = UndefValue::get(LI.getType());
3324 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3325 LI.replaceAllUsesWith(V);
3326 LI.eraseFromParent();
3330 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3331 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr, Type *BaseTy,
3332 AAMDNodes AATags, unsigned BaseAlign, const DataLayout &DL)
3333 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr, BaseTy, BaseAlign,
3337 /// Emit a leaf store of a single value. This is called at the leaves of the
3338 /// recursive emission to actually produce stores.
3339 void emitFunc(Type *Ty, Value *&Agg, unsigned Align, const Twine &Name) {
3340 assert(Ty->isSingleValueType());
3341 // Extract the single value and store it using the indices.
3343 // The gep and extractvalue values are factored out of the CreateStore
3344 // call to make the output independent of the argument evaluation order.
3345 Value *ExtractValue =
3346 IRB.CreateExtractValue(Agg, Indices, Name + ".extract");
3347 Value *InBoundsGEP =
3348 IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3350 IRB.CreateAlignedStore(ExtractValue, InBoundsGEP, Align);
3352 Store->setAAMetadata(AATags);
3353 LLVM_DEBUG(dbgs() << " to: " << *Store << "\n");
3357 bool visitStoreInst(StoreInst &SI) {
3358 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3360 Value *V = SI.getValueOperand();
3361 if (V->getType()->isSingleValueType())
3364 // We have an aggregate being stored, split it apart.
3365 LLVM_DEBUG(dbgs() << " original: " << SI << "\n");
3367 SI.getAAMetadata(AATags);
3368 StoreOpSplitter Splitter(&SI, *U, V->getType(), AATags,
3369 getAdjustedAlignment(&SI, 0, DL), DL);
3370 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3371 SI.eraseFromParent();
3375 bool visitBitCastInst(BitCastInst &BC) {
3380 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3385 bool visitPHINode(PHINode &PN) {
3390 bool visitSelectInst(SelectInst &SI) {
3396 } // end anonymous namespace
3398 /// Strip aggregate type wrapping.
3400 /// This removes no-op aggregate types wrapping an underlying type. It will
3401 /// strip as many layers of types as it can without changing either the type
3402 /// size or the allocated size.
3403 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3404 if (Ty->isSingleValueType())
3407 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3408 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3411 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3412 InnerTy = ArrTy->getElementType();
3413 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3414 const StructLayout *SL = DL.getStructLayout(STy);
3415 unsigned Index = SL->getElementContainingOffset(0);
3416 InnerTy = STy->getElementType(Index);
3421 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3422 TypeSize > DL.getTypeSizeInBits(InnerTy))
3425 return stripAggregateTypeWrapping(DL, InnerTy);
3428 /// Try to find a partition of the aggregate type passed in for a given
3429 /// offset and size.
3431 /// This recurses through the aggregate type and tries to compute a subtype
3432 /// based on the offset and size. When the offset and size span a sub-section
3433 /// of an array, it will even compute a new array type for that sub-section,
3434 /// and the same for structs.
3436 /// Note that this routine is very strict and tries to find a partition of the
3437 /// type which produces the *exact* right offset and size. It is not forgiving
3438 /// when the size or offset cause either end of type-based partition to be off.
3439 /// Also, this is a best-effort routine. It is reasonable to give up and not
3440 /// return a type if necessary.
3441 static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3443 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3444 return stripAggregateTypeWrapping(DL, Ty);
3445 if (Offset > DL.getTypeAllocSize(Ty) ||
3446 (DL.getTypeAllocSize(Ty) - Offset) < Size)
3449 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3450 Type *ElementTy = SeqTy->getElementType();
3451 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3452 uint64_t NumSkippedElements = Offset / ElementSize;
3453 if (NumSkippedElements >= SeqTy->getNumElements())
3455 Offset -= NumSkippedElements * ElementSize;
3457 // First check if we need to recurse.
3458 if (Offset > 0 || Size < ElementSize) {
3459 // Bail if the partition ends in a different array element.
3460 if ((Offset + Size) > ElementSize)
3462 // Recurse through the element type trying to peel off offset bytes.
3463 return getTypePartition(DL, ElementTy, Offset, Size);
3465 assert(Offset == 0);
3467 if (Size == ElementSize)
3468 return stripAggregateTypeWrapping(DL, ElementTy);
3469 assert(Size > ElementSize);
3470 uint64_t NumElements = Size / ElementSize;
3471 if (NumElements * ElementSize != Size)
3473 return ArrayType::get(ElementTy, NumElements);
3476 StructType *STy = dyn_cast<StructType>(Ty);
3480 const StructLayout *SL = DL.getStructLayout(STy);
3481 if (Offset >= SL->getSizeInBytes())
3483 uint64_t EndOffset = Offset + Size;
3484 if (EndOffset > SL->getSizeInBytes())
3487 unsigned Index = SL->getElementContainingOffset(Offset);
3488 Offset -= SL->getElementOffset(Index);
3490 Type *ElementTy = STy->getElementType(Index);
3491 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3492 if (Offset >= ElementSize)
3493 return nullptr; // The offset points into alignment padding.
3495 // See if any partition must be contained by the element.
3496 if (Offset > 0 || Size < ElementSize) {
3497 if ((Offset + Size) > ElementSize)
3499 return getTypePartition(DL, ElementTy, Offset, Size);
3501 assert(Offset == 0);
3503 if (Size == ElementSize)
3504 return stripAggregateTypeWrapping(DL, ElementTy);
3506 StructType::element_iterator EI = STy->element_begin() + Index,
3507 EE = STy->element_end();
3508 if (EndOffset < SL->getSizeInBytes()) {
3509 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3510 if (Index == EndIndex)
3511 return nullptr; // Within a single element and its padding.
3513 // Don't try to form "natural" types if the elements don't line up with the
3515 // FIXME: We could potentially recurse down through the last element in the
3516 // sub-struct to find a natural end point.
3517 if (SL->getElementOffset(EndIndex) != EndOffset)
3520 assert(Index < EndIndex);
3521 EE = STy->element_begin() + EndIndex;
3524 // Try to build up a sub-structure.
3526 StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3527 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3528 if (Size != SubSL->getSizeInBytes())
3529 return nullptr; // The sub-struct doesn't have quite the size needed.
3534 /// Pre-split loads and stores to simplify rewriting.
3536 /// We want to break up the splittable load+store pairs as much as
3537 /// possible. This is important to do as a preprocessing step, as once we
3538 /// start rewriting the accesses to partitions of the alloca we lose the
3539 /// necessary information to correctly split apart paired loads and stores
3540 /// which both point into this alloca. The case to consider is something like
3543 /// %a = alloca [12 x i8]
3544 /// %gep1 = getelementptr [12 x i8]* %a, i32 0, i32 0
3545 /// %gep2 = getelementptr [12 x i8]* %a, i32 0, i32 4
3546 /// %gep3 = getelementptr [12 x i8]* %a, i32 0, i32 8
3547 /// %iptr1 = bitcast i8* %gep1 to i64*
3548 /// %iptr2 = bitcast i8* %gep2 to i64*
3549 /// %fptr1 = bitcast i8* %gep1 to float*
3550 /// %fptr2 = bitcast i8* %gep2 to float*
3551 /// %fptr3 = bitcast i8* %gep3 to float*
3552 /// store float 0.0, float* %fptr1
3553 /// store float 1.0, float* %fptr2
3554 /// %v = load i64* %iptr1
3555 /// store i64 %v, i64* %iptr2
3556 /// %f1 = load float* %fptr2
3557 /// %f2 = load float* %fptr3
3559 /// Here we want to form 3 partitions of the alloca, each 4 bytes large, and
3560 /// promote everything so we recover the 2 SSA values that should have been
3561 /// there all along.
3563 /// \returns true if any changes are made.
3564 bool SROA::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
3565 LLVM_DEBUG(dbgs() << "Pre-splitting loads and stores\n");
3567 // Track the loads and stores which are candidates for pre-splitting here, in
3568 // the order they first appear during the partition scan. These give stable
3569 // iteration order and a basis for tracking which loads and stores we
3571 SmallVector<LoadInst *, 4> Loads;
3572 SmallVector<StoreInst *, 4> Stores;
3574 // We need to accumulate the splits required of each load or store where we
3575 // can find them via a direct lookup. This is important to cross-check loads
3576 // and stores against each other. We also track the slice so that we can kill
3577 // all the slices that end up split.
3578 struct SplitOffsets {
3580 std::vector<uint64_t> Splits;
3582 SmallDenseMap<Instruction *, SplitOffsets, 8> SplitOffsetsMap;
3584 // Track loads out of this alloca which cannot, for any reason, be pre-split.
3585 // This is important as we also cannot pre-split stores of those loads!
3586 // FIXME: This is all pretty gross. It means that we can be more aggressive
3587 // in pre-splitting when the load feeding the store happens to come from
3588 // a separate alloca. Put another way, the effectiveness of SROA would be
3589 // decreased by a frontend which just concatenated all of its local allocas
3590 // into one big flat alloca. But defeating such patterns is exactly the job
3591 // SROA is tasked with! Sadly, to not have this discrepancy we would have
3592 // change store pre-splitting to actually force pre-splitting of the load
3593 // that feeds it *and all stores*. That makes pre-splitting much harder, but
3594 // maybe it would make it more principled?
3595 SmallPtrSet<LoadInst *, 8> UnsplittableLoads;
3597 LLVM_DEBUG(dbgs() << " Searching for candidate loads and stores\n");
3598 for (auto &P : AS.partitions()) {
3599 for (Slice &S : P) {
3600 Instruction *I = cast<Instruction>(S.getUse()->getUser());
3601 if (!S.isSplittable() || S.endOffset() <= P.endOffset()) {
3602 // If this is a load we have to track that it can't participate in any
3603 // pre-splitting. If this is a store of a load we have to track that
3604 // that load also can't participate in any pre-splitting.
3605 if (auto *LI = dyn_cast<LoadInst>(I))
3606 UnsplittableLoads.insert(LI);
3607 else if (auto *SI = dyn_cast<StoreInst>(I))
3608 if (auto *LI = dyn_cast<LoadInst>(SI->getValueOperand()))
3609 UnsplittableLoads.insert(LI);
3612 assert(P.endOffset() > S.beginOffset() &&
3613 "Empty or backwards partition!");
3615 // Determine if this is a pre-splittable slice.
3616 if (auto *LI = dyn_cast<LoadInst>(I)) {
3617 assert(!LI->isVolatile() && "Cannot split volatile loads!");
3619 // The load must be used exclusively to store into other pointers for
3620 // us to be able to arbitrarily pre-split it. The stores must also be
3621 // simple to avoid changing semantics.
3622 auto IsLoadSimplyStored = [](LoadInst *LI) {
3623 for (User *LU : LI->users()) {
3624 auto *SI = dyn_cast<StoreInst>(LU);
3625 if (!SI || !SI->isSimple())
3630 if (!IsLoadSimplyStored(LI)) {
3631 UnsplittableLoads.insert(LI);
3635 Loads.push_back(LI);
3636 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
3637 if (S.getUse() != &SI->getOperandUse(SI->getPointerOperandIndex()))
3638 // Skip stores *of* pointers. FIXME: This shouldn't even be possible!
3640 auto *StoredLoad = dyn_cast<LoadInst>(SI->getValueOperand());
3641 if (!StoredLoad || !StoredLoad->isSimple())
3643 assert(!SI->isVolatile() && "Cannot split volatile stores!");
3645 Stores.push_back(SI);
3647 // Other uses cannot be pre-split.
3651 // Record the initial split.
3652 LLVM_DEBUG(dbgs() << " Candidate: " << *I << "\n");
3653 auto &Offsets = SplitOffsetsMap[I];
3654 assert(Offsets.Splits.empty() &&
3655 "Should not have splits the first time we see an instruction!");
3657 Offsets.Splits.push_back(P.endOffset() - S.beginOffset());
3660 // Now scan the already split slices, and add a split for any of them which
3661 // we're going to pre-split.
3662 for (Slice *S : P.splitSliceTails()) {
3663 auto SplitOffsetsMapI =
3664 SplitOffsetsMap.find(cast<Instruction>(S->getUse()->getUser()));
3665 if (SplitOffsetsMapI == SplitOffsetsMap.end())
3667 auto &Offsets = SplitOffsetsMapI->second;
3669 assert(Offsets.S == S && "Found a mismatched slice!");
3670 assert(!Offsets.Splits.empty() &&
3671 "Cannot have an empty set of splits on the second partition!");
3672 assert(Offsets.Splits.back() ==
3673 P.beginOffset() - Offsets.S->beginOffset() &&
3674 "Previous split does not end where this one begins!");
3676 // Record each split. The last partition's end isn't needed as the size
3677 // of the slice dictates that.
3678 if (S->endOffset() > P.endOffset())
3679 Offsets.Splits.push_back(P.endOffset() - Offsets.S->beginOffset());
3683 // We may have split loads where some of their stores are split stores. For
3684 // such loads and stores, we can only pre-split them if their splits exactly
3685 // match relative to their starting offset. We have to verify this prior to
3688 llvm::remove_if(Stores,
3689 [&UnsplittableLoads, &SplitOffsetsMap](StoreInst *SI) {
3690 // Lookup the load we are storing in our map of split
3692 auto *LI = cast<LoadInst>(SI->getValueOperand());
3693 // If it was completely unsplittable, then we're done,
3694 // and this store can't be pre-split.
3695 if (UnsplittableLoads.count(LI))
3698 auto LoadOffsetsI = SplitOffsetsMap.find(LI);
3699 if (LoadOffsetsI == SplitOffsetsMap.end())
3700 return false; // Unrelated loads are definitely safe.
3701 auto &LoadOffsets = LoadOffsetsI->second;
3703 // Now lookup the store's offsets.
3704 auto &StoreOffsets = SplitOffsetsMap[SI];
3706 // If the relative offsets of each split in the load and
3707 // store match exactly, then we can split them and we
3708 // don't need to remove them here.
3709 if (LoadOffsets.Splits == StoreOffsets.Splits)
3714 << " Mismatched splits for load and store:\n"
3715 << " " << *LI << "\n"
3716 << " " << *SI << "\n");
3718 // We've found a store and load that we need to split
3719 // with mismatched relative splits. Just give up on them
3720 // and remove both instructions from our list of
3722 UnsplittableLoads.insert(LI);
3726 // Now we have to go *back* through all the stores, because a later store may
3727 // have caused an earlier store's load to become unsplittable and if it is
3728 // unsplittable for the later store, then we can't rely on it being split in
3729 // the earlier store either.
3730 Stores.erase(llvm::remove_if(Stores,
3731 [&UnsplittableLoads](StoreInst *SI) {
3733 cast<LoadInst>(SI->getValueOperand());
3734 return UnsplittableLoads.count(LI);
3737 // Once we've established all the loads that can't be split for some reason,
3738 // filter any that made it into our list out.
3739 Loads.erase(llvm::remove_if(Loads,
3740 [&UnsplittableLoads](LoadInst *LI) {
3741 return UnsplittableLoads.count(LI);
3745 // If no loads or stores are left, there is no pre-splitting to be done for
3747 if (Loads.empty() && Stores.empty())
3750 // From here on, we can't fail and will be building new accesses, so rig up
3752 IRBuilderTy IRB(&AI);
3754 // Collect the new slices which we will merge into the alloca slices.
3755 SmallVector<Slice, 4> NewSlices;
3757 // Track any allocas we end up splitting loads and stores for so we iterate
3759 SmallPtrSet<AllocaInst *, 4> ResplitPromotableAllocas;
3761 // At this point, we have collected all of the loads and stores we can
3762 // pre-split, and the specific splits needed for them. We actually do the
3763 // splitting in a specific order in order to handle when one of the loads in
3764 // the value operand to one of the stores.
3766 // First, we rewrite all of the split loads, and just accumulate each split
3767 // load in a parallel structure. We also build the slices for them and append
3768 // them to the alloca slices.
3769 SmallDenseMap<LoadInst *, std::vector<LoadInst *>, 1> SplitLoadsMap;
3770 std::vector<LoadInst *> SplitLoads;
3771 const DataLayout &DL = AI.getModule()->getDataLayout();
3772 for (LoadInst *LI : Loads) {
3775 IntegerType *Ty = cast<IntegerType>(LI->getType());
3776 uint64_t LoadSize = Ty->getBitWidth() / 8;
3777 assert(LoadSize > 0 && "Cannot have a zero-sized integer load!");
3779 auto &Offsets = SplitOffsetsMap[LI];
3780 assert(LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3781 "Slice size should always match load size exactly!");
3782 uint64_t BaseOffset = Offsets.S->beginOffset();
3783 assert(BaseOffset + LoadSize > BaseOffset &&
3784 "Cannot represent alloca access size using 64-bit integers!");
3786 Instruction *BasePtr = cast<Instruction>(LI->getPointerOperand());
3787 IRB.SetInsertPoint(LI);
3789 LLVM_DEBUG(dbgs() << " Splitting load: " << *LI << "\n");
3791 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3792 int Idx = 0, Size = Offsets.Splits.size();
3794 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3795 auto AS = LI->getPointerAddressSpace();
3796 auto *PartPtrTy = PartTy->getPointerTo(AS);
3797 LoadInst *PLoad = IRB.CreateAlignedLoad(
3798 getAdjustedPtr(IRB, DL, BasePtr,
3799 APInt(DL.getIndexSizeInBits(AS), PartOffset),
3800 PartPtrTy, BasePtr->getName() + "."),
3801 getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3803 PLoad->copyMetadata(*LI, {LLVMContext::MD_mem_parallel_loop_access,
3804 LLVMContext::MD_access_group});
3806 // Append this load onto the list of split loads so we can find it later
3807 // to rewrite the stores.
3808 SplitLoads.push_back(PLoad);
3810 // Now build a new slice for the alloca.
3811 NewSlices.push_back(
3812 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3813 &PLoad->getOperandUse(PLoad->getPointerOperandIndex()),
3814 /*IsSplittable*/ false));
3815 LLVM_DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3816 << ", " << NewSlices.back().endOffset()
3817 << "): " << *PLoad << "\n");
3819 // See if we've handled all the splits.
3823 // Setup the next partition.
3824 PartOffset = Offsets.Splits[Idx];
3826 PartSize = (Idx < Size ? Offsets.Splits[Idx] : LoadSize) - PartOffset;
3829 // Now that we have the split loads, do the slow walk over all uses of the
3830 // load and rewrite them as split stores, or save the split loads to use
3831 // below if the store is going to be split there anyways.
3832 bool DeferredStores = false;
3833 for (User *LU : LI->users()) {
3834 StoreInst *SI = cast<StoreInst>(LU);
3835 if (!Stores.empty() && SplitOffsetsMap.count(SI)) {
3836 DeferredStores = true;
3837 LLVM_DEBUG(dbgs() << " Deferred splitting of store: " << *SI
3842 Value *StoreBasePtr = SI->getPointerOperand();
3843 IRB.SetInsertPoint(SI);
3845 LLVM_DEBUG(dbgs() << " Splitting store of load: " << *SI << "\n");
3847 for (int Idx = 0, Size = SplitLoads.size(); Idx < Size; ++Idx) {
3848 LoadInst *PLoad = SplitLoads[Idx];
3849 uint64_t PartOffset = Idx == 0 ? 0 : Offsets.Splits[Idx - 1];
3851 PLoad->getType()->getPointerTo(SI->getPointerAddressSpace());
3853 auto AS = SI->getPointerAddressSpace();
3854 StoreInst *PStore = IRB.CreateAlignedStore(
3856 getAdjustedPtr(IRB, DL, StoreBasePtr,
3857 APInt(DL.getIndexSizeInBits(AS), PartOffset),
3858 PartPtrTy, StoreBasePtr->getName() + "."),
3859 getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3860 PStore->copyMetadata(*LI, {LLVMContext::MD_mem_parallel_loop_access,
3861 LLVMContext::MD_access_group});
3862 LLVM_DEBUG(dbgs() << " +" << PartOffset << ":" << *PStore << "\n");
3865 // We want to immediately iterate on any allocas impacted by splitting
3866 // this store, and we have to track any promotable alloca (indicated by
3867 // a direct store) as needing to be resplit because it is no longer
3869 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {
3870 ResplitPromotableAllocas.insert(OtherAI);
3871 Worklist.insert(OtherAI);
3872 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3873 StoreBasePtr->stripInBoundsOffsets())) {
3874 Worklist.insert(OtherAI);
3877 // Mark the original store as dead.
3878 DeadInsts.insert(SI);
3881 // Save the split loads if there are deferred stores among the users.
3883 SplitLoadsMap.insert(std::make_pair(LI, std::move(SplitLoads)));
3885 // Mark the original load as dead and kill the original slice.
3886 DeadInsts.insert(LI);
3890 // Second, we rewrite all of the split stores. At this point, we know that
3891 // all loads from this alloca have been split already. For stores of such
3892 // loads, we can simply look up the pre-existing split loads. For stores of
3893 // other loads, we split those loads first and then write split stores of
3895 for (StoreInst *SI : Stores) {
3896 auto *LI = cast<LoadInst>(SI->getValueOperand());
3897 IntegerType *Ty = cast<IntegerType>(LI->getType());
3898 uint64_t StoreSize = Ty->getBitWidth() / 8;
3899 assert(StoreSize > 0 && "Cannot have a zero-sized integer store!");
3901 auto &Offsets = SplitOffsetsMap[SI];
3902 assert(StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3903 "Slice size should always match load size exactly!");
3904 uint64_t BaseOffset = Offsets.S->beginOffset();
3905 assert(BaseOffset + StoreSize > BaseOffset &&
3906 "Cannot represent alloca access size using 64-bit integers!");
3908 Value *LoadBasePtr = LI->getPointerOperand();
3909 Instruction *StoreBasePtr = cast<Instruction>(SI->getPointerOperand());
3911 LLVM_DEBUG(dbgs() << " Splitting store: " << *SI << "\n");
3913 // Check whether we have an already split load.
3914 auto SplitLoadsMapI = SplitLoadsMap.find(LI);
3915 std::vector<LoadInst *> *SplitLoads = nullptr;
3916 if (SplitLoadsMapI != SplitLoadsMap.end()) {
3917 SplitLoads = &SplitLoadsMapI->second;
3918 assert(SplitLoads->size() == Offsets.Splits.size() + 1 &&
3919 "Too few split loads for the number of splits in the store!");
3921 LLVM_DEBUG(dbgs() << " of load: " << *LI << "\n");
3924 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3925 int Idx = 0, Size = Offsets.Splits.size();
3927 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3928 auto *LoadPartPtrTy = PartTy->getPointerTo(LI->getPointerAddressSpace());
3929 auto *StorePartPtrTy = PartTy->getPointerTo(SI->getPointerAddressSpace());
3931 // Either lookup a split load or create one.
3934 PLoad = (*SplitLoads)[Idx];
3936 IRB.SetInsertPoint(LI);
3937 auto AS = LI->getPointerAddressSpace();
3938 PLoad = IRB.CreateAlignedLoad(
3939 getAdjustedPtr(IRB, DL, LoadBasePtr,
3940 APInt(DL.getIndexSizeInBits(AS), PartOffset),
3941 LoadPartPtrTy, LoadBasePtr->getName() + "."),
3942 getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3946 // And store this partition.
3947 IRB.SetInsertPoint(SI);
3948 auto AS = SI->getPointerAddressSpace();
3949 StoreInst *PStore = IRB.CreateAlignedStore(
3951 getAdjustedPtr(IRB, DL, StoreBasePtr,
3952 APInt(DL.getIndexSizeInBits(AS), PartOffset),
3953 StorePartPtrTy, StoreBasePtr->getName() + "."),
3954 getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3956 // Now build a new slice for the alloca.
3957 NewSlices.push_back(
3958 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3959 &PStore->getOperandUse(PStore->getPointerOperandIndex()),
3960 /*IsSplittable*/ false));
3961 LLVM_DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3962 << ", " << NewSlices.back().endOffset()
3963 << "): " << *PStore << "\n");
3965 LLVM_DEBUG(dbgs() << " of split load: " << *PLoad << "\n");
3968 // See if we've finished all the splits.
3972 // Setup the next partition.
3973 PartOffset = Offsets.Splits[Idx];
3975 PartSize = (Idx < Size ? Offsets.Splits[Idx] : StoreSize) - PartOffset;
3978 // We want to immediately iterate on any allocas impacted by splitting
3979 // this load, which is only relevant if it isn't a load of this alloca and
3980 // thus we didn't already split the loads above. We also have to keep track
3981 // of any promotable allocas we split loads on as they can no longer be
3984 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(LoadBasePtr)) {
3985 assert(OtherAI != &AI && "We can't re-split our own alloca!");
3986 ResplitPromotableAllocas.insert(OtherAI);
3987 Worklist.insert(OtherAI);
3988 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3989 LoadBasePtr->stripInBoundsOffsets())) {
3990 assert(OtherAI != &AI && "We can't re-split our own alloca!");
3991 Worklist.insert(OtherAI);
3995 // Mark the original store as dead now that we've split it up and kill its
3996 // slice. Note that we leave the original load in place unless this store
3997 // was its only use. It may in turn be split up if it is an alloca load
3998 // for some other alloca, but it may be a normal load. This may introduce
3999 // redundant loads, but where those can be merged the rest of the optimizer
4000 // should handle the merging, and this uncovers SSA splits which is more
4001 // important. In practice, the original loads will almost always be fully
4002 // split and removed eventually, and the splits will be merged by any
4003 // trivial CSE, including instcombine.
4004 if (LI->hasOneUse()) {
4005 assert(*LI->user_begin() == SI && "Single use isn't this store!");
4006 DeadInsts.insert(LI);
4008 DeadInsts.insert(SI);
4012 // Remove the killed slices that have ben pre-split.
4013 AS.erase(llvm::remove_if(AS, [](const Slice &S) { return S.isDead(); }),
4016 // Insert our new slices. This will sort and merge them into the sorted
4018 AS.insert(NewSlices);
4020 LLVM_DEBUG(dbgs() << " Pre-split slices:\n");
4022 for (auto I = AS.begin(), E = AS.end(); I != E; ++I)
4023 LLVM_DEBUG(AS.print(dbgs(), I, " "));
4026 // Finally, don't try to promote any allocas that new require re-splitting.
4027 // They have already been added to the worklist above.
4028 PromotableAllocas.erase(
4031 [&](AllocaInst *AI) { return ResplitPromotableAllocas.count(AI); }),
4032 PromotableAllocas.end());
4037 /// Rewrite an alloca partition's users.
4039 /// This routine drives both of the rewriting goals of the SROA pass. It tries
4040 /// to rewrite uses of an alloca partition to be conducive for SSA value
4041 /// promotion. If the partition needs a new, more refined alloca, this will
4042 /// build that new alloca, preserving as much type information as possible, and
4043 /// rewrite the uses of the old alloca to point at the new one and have the
4044 /// appropriate new offsets. It also evaluates how successful the rewrite was
4045 /// at enabling promotion and if it was successful queues the alloca to be
4047 AllocaInst *SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
4049 // Try to compute a friendly type for this partition of the alloca. This
4050 // won't always succeed, in which case we fall back to a legal integer type
4051 // or an i8 array of an appropriate size.
4052 Type *SliceTy = nullptr;
4053 const DataLayout &DL = AI.getModule()->getDataLayout();
4054 if (Type *CommonUseTy = findCommonType(P.begin(), P.end(), P.endOffset()))
4055 if (DL.getTypeAllocSize(CommonUseTy) >= P.size())
4056 SliceTy = CommonUseTy;
4058 if (Type *TypePartitionTy = getTypePartition(DL, AI.getAllocatedType(),
4059 P.beginOffset(), P.size()))
4060 SliceTy = TypePartitionTy;
4061 if ((!SliceTy || (SliceTy->isArrayTy() &&
4062 SliceTy->getArrayElementType()->isIntegerTy())) &&
4063 DL.isLegalInteger(P.size() * 8))
4064 SliceTy = Type::getIntNTy(*C, P.size() * 8);
4066 SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
4067 assert(DL.getTypeAllocSize(SliceTy) >= P.size());
4069 bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, DL);
4072 IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, DL);
4076 // Check for the case where we're going to rewrite to a new alloca of the
4077 // exact same type as the original, and with the same access offsets. In that
4078 // case, re-use the existing alloca, but still run through the rewriter to
4079 // perform phi and select speculation.
4080 // P.beginOffset() can be non-zero even with the same type in a case with
4081 // out-of-bounds access (e.g. @PR35657 function in SROA/basictest.ll).
4083 if (SliceTy == AI.getAllocatedType() && P.beginOffset() == 0) {
4085 // FIXME: We should be able to bail at this point with "nothing changed".
4086 // FIXME: We might want to defer PHI speculation until after here.
4087 // FIXME: return nullptr;
4089 unsigned Alignment = AI.getAlignment();
4091 // The minimum alignment which users can rely on when the explicit
4092 // alignment is omitted or zero is that required by the ABI for this
4094 Alignment = DL.getABITypeAlignment(AI.getAllocatedType());
4096 Alignment = MinAlign(Alignment, P.beginOffset());
4097 // If we will get at least this much alignment from the type alone, leave
4098 // the alloca's alignment unconstrained.
4099 if (Alignment <= DL.getABITypeAlignment(SliceTy))
4101 NewAI = new AllocaInst(
4102 SliceTy, AI.getType()->getAddressSpace(), nullptr, Alignment,
4103 AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
4104 // Copy the old AI debug location over to the new one.
4105 NewAI->setDebugLoc(AI.getDebugLoc());
4109 LLVM_DEBUG(dbgs() << "Rewriting alloca partition "
4110 << "[" << P.beginOffset() << "," << P.endOffset()
4111 << ") to: " << *NewAI << "\n");
4113 // Track the high watermark on the worklist as it is only relevant for
4114 // promoted allocas. We will reset it to this point if the alloca is not in
4115 // fact scheduled for promotion.
4116 unsigned PPWOldSize = PostPromotionWorklist.size();
4117 unsigned NumUses = 0;
4118 SmallSetVector<PHINode *, 8> PHIUsers;
4119 SmallSetVector<SelectInst *, 8> SelectUsers;
4121 AllocaSliceRewriter Rewriter(DL, AS, *this, AI, *NewAI, P.beginOffset(),
4122 P.endOffset(), IsIntegerPromotable, VecTy,
4123 PHIUsers, SelectUsers);
4124 bool Promotable = true;
4125 for (Slice *S : P.splitSliceTails()) {
4126 Promotable &= Rewriter.visit(S);
4129 for (Slice &S : P) {
4130 Promotable &= Rewriter.visit(&S);
4134 NumAllocaPartitionUses += NumUses;
4135 MaxUsesPerAllocaPartition.updateMax(NumUses);
4137 // Now that we've processed all the slices in the new partition, check if any
4138 // PHIs or Selects would block promotion.
4139 for (PHINode *PHI : PHIUsers)
4140 if (!isSafePHIToSpeculate(*PHI)) {
4143 SelectUsers.clear();
4147 for (SelectInst *Sel : SelectUsers)
4148 if (!isSafeSelectToSpeculate(*Sel)) {
4151 SelectUsers.clear();
4156 if (PHIUsers.empty() && SelectUsers.empty()) {
4157 // Promote the alloca.
4158 PromotableAllocas.push_back(NewAI);
4160 // If we have either PHIs or Selects to speculate, add them to those
4161 // worklists and re-queue the new alloca so that we promote in on the
4163 for (PHINode *PHIUser : PHIUsers)
4164 SpeculatablePHIs.insert(PHIUser);
4165 for (SelectInst *SelectUser : SelectUsers)
4166 SpeculatableSelects.insert(SelectUser);
4167 Worklist.insert(NewAI);
4170 // Drop any post-promotion work items if promotion didn't happen.
4171 while (PostPromotionWorklist.size() > PPWOldSize)
4172 PostPromotionWorklist.pop_back();
4174 // We couldn't promote and we didn't create a new partition, nothing
4179 // If we can't promote the alloca, iterate on it to check for new
4180 // refinements exposed by splitting the current alloca. Don't iterate on an
4181 // alloca which didn't actually change and didn't get promoted.
4182 Worklist.insert(NewAI);
4188 /// Walks the slices of an alloca and form partitions based on them,
4189 /// rewriting each of their uses.
4190 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
4191 if (AS.begin() == AS.end())
4194 unsigned NumPartitions = 0;
4195 bool Changed = false;
4196 const DataLayout &DL = AI.getModule()->getDataLayout();
4198 // First try to pre-split loads and stores.
4199 Changed |= presplitLoadsAndStores(AI, AS);
4201 // Now that we have identified any pre-splitting opportunities,
4202 // mark loads and stores unsplittable except for the following case.
4203 // We leave a slice splittable if all other slices are disjoint or fully
4204 // included in the slice, such as whole-alloca loads and stores.
4205 // If we fail to split these during pre-splitting, we want to force them
4206 // to be rewritten into a partition.
4207 bool IsSorted = true;
4209 uint64_t AllocaSize = DL.getTypeAllocSize(AI.getAllocatedType());
4210 const uint64_t MaxBitVectorSize = 1024;
4211 if (AllocaSize <= MaxBitVectorSize) {
4212 // If a byte boundary is included in any load or store, a slice starting or
4213 // ending at the boundary is not splittable.
4214 SmallBitVector SplittableOffset(AllocaSize + 1, true);
4216 for (unsigned O = S.beginOffset() + 1;
4217 O < S.endOffset() && O < AllocaSize; O++)
4218 SplittableOffset.reset(O);
4220 for (Slice &S : AS) {
4221 if (!S.isSplittable())
4224 if ((S.beginOffset() > AllocaSize || SplittableOffset[S.beginOffset()]) &&
4225 (S.endOffset() > AllocaSize || SplittableOffset[S.endOffset()]))
4228 if (isa<LoadInst>(S.getUse()->getUser()) ||
4229 isa<StoreInst>(S.getUse()->getUser())) {
4230 S.makeUnsplittable();
4236 // We only allow whole-alloca splittable loads and stores
4237 // for a large alloca to avoid creating too large BitVector.
4238 for (Slice &S : AS) {
4239 if (!S.isSplittable())
4242 if (S.beginOffset() == 0 && S.endOffset() >= AllocaSize)
4245 if (isa<LoadInst>(S.getUse()->getUser()) ||
4246 isa<StoreInst>(S.getUse()->getUser())) {
4247 S.makeUnsplittable();
4256 /// Describes the allocas introduced by rewritePartition in order to migrate
4262 Fragment(AllocaInst *AI, uint64_t O, uint64_t S)
4263 : Alloca(AI), Offset(O), Size(S) {}
4265 SmallVector<Fragment, 4> Fragments;
4267 // Rewrite each partition.
4268 for (auto &P : AS.partitions()) {
4269 if (AllocaInst *NewAI = rewritePartition(AI, AS, P)) {
4272 uint64_t SizeOfByte = 8;
4273 uint64_t AllocaSize = DL.getTypeSizeInBits(NewAI->getAllocatedType());
4274 // Don't include any padding.
4275 uint64_t Size = std::min(AllocaSize, P.size() * SizeOfByte);
4276 Fragments.push_back(Fragment(NewAI, P.beginOffset() * SizeOfByte, Size));
4282 NumAllocaPartitions += NumPartitions;
4283 MaxPartitionsPerAlloca.updateMax(NumPartitions);
4285 // Migrate debug information from the old alloca to the new alloca(s)
4286 // and the individual partitions.
4287 TinyPtrVector<DbgVariableIntrinsic *> DbgDeclares = FindDbgAddrUses(&AI);
4288 if (!DbgDeclares.empty()) {
4289 auto *Var = DbgDeclares.front()->getVariable();
4290 auto *Expr = DbgDeclares.front()->getExpression();
4291 auto VarSize = Var->getSizeInBits();
4292 DIBuilder DIB(*AI.getModule(), /*AllowUnresolved*/ false);
4293 uint64_t AllocaSize = DL.getTypeSizeInBits(AI.getAllocatedType());
4294 for (auto Fragment : Fragments) {
4295 // Create a fragment expression describing the new partition or reuse AI's
4296 // expression if there is only one partition.
4297 auto *FragmentExpr = Expr;
4298 if (Fragment.Size < AllocaSize || Expr->isFragment()) {
4299 // If this alloca is already a scalar replacement of a larger aggregate,
4300 // Fragment.Offset describes the offset inside the scalar.
4301 auto ExprFragment = Expr->getFragmentInfo();
4302 uint64_t Offset = ExprFragment ? ExprFragment->OffsetInBits : 0;
4303 uint64_t Start = Offset + Fragment.Offset;
4304 uint64_t Size = Fragment.Size;
4307 ExprFragment->OffsetInBits + ExprFragment->SizeInBits;
4308 if (Start >= AbsEnd)
4309 // No need to describe a SROAed padding.
4311 Size = std::min(Size, AbsEnd - Start);
4313 // The new, smaller fragment is stenciled out from the old fragment.
4314 if (auto OrigFragment = FragmentExpr->getFragmentInfo()) {
4315 assert(Start >= OrigFragment->OffsetInBits &&
4316 "new fragment is outside of original fragment");
4317 Start -= OrigFragment->OffsetInBits;
4320 // The alloca may be larger than the variable.
4322 if (Size > *VarSize)
4324 if (Size == 0 || Start + Size > *VarSize)
4328 // Avoid creating a fragment expression that covers the entire variable.
4329 if (!VarSize || *VarSize != Size) {
4331 DIExpression::createFragmentExpression(Expr, Start, Size))
4338 // Remove any existing intrinsics describing the same alloca.
4339 for (DbgVariableIntrinsic *OldDII : FindDbgAddrUses(Fragment.Alloca))
4340 OldDII->eraseFromParent();
4342 DIB.insertDeclare(Fragment.Alloca, Var, FragmentExpr,
4343 DbgDeclares.front()->getDebugLoc(), &AI);
4349 /// Clobber a use with undef, deleting the used value if it becomes dead.
4350 void SROA::clobberUse(Use &U) {
4352 // Replace the use with an undef value.
4353 U = UndefValue::get(OldV->getType());
4355 // Check for this making an instruction dead. We have to garbage collect
4356 // all the dead instructions to ensure the uses of any alloca end up being
4358 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
4359 if (isInstructionTriviallyDead(OldI)) {
4360 DeadInsts.insert(OldI);
4364 /// Analyze an alloca for SROA.
4366 /// This analyzes the alloca to ensure we can reason about it, builds
4367 /// the slices of the alloca, and then hands it off to be split and
4368 /// rewritten as needed.
4369 bool SROA::runOnAlloca(AllocaInst &AI) {
4370 LLVM_DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
4371 ++NumAllocasAnalyzed;
4373 // Special case dead allocas, as they're trivial.
4374 if (AI.use_empty()) {
4375 AI.eraseFromParent();
4378 const DataLayout &DL = AI.getModule()->getDataLayout();
4380 // Skip alloca forms that this analysis can't handle.
4381 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
4382 DL.getTypeAllocSize(AI.getAllocatedType()) == 0)
4385 bool Changed = false;
4387 // First, split any FCA loads and stores touching this alloca to promote
4388 // better splitting and promotion opportunities.
4389 AggLoadStoreRewriter AggRewriter(DL);
4390 Changed |= AggRewriter.rewrite(AI);
4392 // Build the slices using a recursive instruction-visiting builder.
4393 AllocaSlices AS(DL, AI);
4394 LLVM_DEBUG(AS.print(dbgs()));
4398 // Delete all the dead users of this alloca before splitting and rewriting it.
4399 for (Instruction *DeadUser : AS.getDeadUsers()) {
4400 // Free up everything used by this instruction.
4401 for (Use &DeadOp : DeadUser->operands())
4404 // Now replace the uses of this instruction.
4405 DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
4407 // And mark it for deletion.
4408 DeadInsts.insert(DeadUser);
4411 for (Use *DeadOp : AS.getDeadOperands()) {
4412 clobberUse(*DeadOp);
4416 // No slices to split. Leave the dead alloca for a later pass to clean up.
4417 if (AS.begin() == AS.end())
4420 Changed |= splitAlloca(AI, AS);
4422 LLVM_DEBUG(dbgs() << " Speculating PHIs\n");
4423 while (!SpeculatablePHIs.empty())
4424 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
4426 LLVM_DEBUG(dbgs() << " Speculating Selects\n");
4427 while (!SpeculatableSelects.empty())
4428 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
4433 /// Delete the dead instructions accumulated in this run.
4435 /// Recursively deletes the dead instructions we've accumulated. This is done
4436 /// at the very end to maximize locality of the recursive delete and to
4437 /// minimize the problems of invalidated instruction pointers as such pointers
4438 /// are used heavily in the intermediate stages of the algorithm.
4440 /// We also record the alloca instructions deleted here so that they aren't
4441 /// subsequently handed to mem2reg to promote.
4442 bool SROA::deleteDeadInstructions(
4443 SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
4444 bool Changed = false;
4445 while (!DeadInsts.empty()) {
4446 Instruction *I = DeadInsts.pop_back_val();
4447 LLVM_DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
4449 // If the instruction is an alloca, find the possible dbg.declare connected
4450 // to it, and remove it too. We must do this before calling RAUW or we will
4451 // not be able to find it.
4452 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4453 DeletedAllocas.insert(AI);
4454 for (DbgVariableIntrinsic *OldDII : FindDbgAddrUses(AI))
4455 OldDII->eraseFromParent();
4458 I->replaceAllUsesWith(UndefValue::get(I->getType()));
4460 for (Use &Operand : I->operands())
4461 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
4462 // Zero out the operand and see if it becomes trivially dead.
4464 if (isInstructionTriviallyDead(U))
4465 DeadInsts.insert(U);
4469 I->eraseFromParent();
4475 /// Promote the allocas, using the best available technique.
4477 /// This attempts to promote whatever allocas have been identified as viable in
4478 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
4479 /// This function returns whether any promotion occurred.
4480 bool SROA::promoteAllocas(Function &F) {
4481 if (PromotableAllocas.empty())
4484 NumPromoted += PromotableAllocas.size();
4486 LLVM_DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
4487 PromoteMemToReg(PromotableAllocas, *DT, AC);
4488 PromotableAllocas.clear();
4492 PreservedAnalyses SROA::runImpl(Function &F, DominatorTree &RunDT,
4493 AssumptionCache &RunAC) {
4494 LLVM_DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
4495 C = &F.getContext();
4499 BasicBlock &EntryBB = F.getEntryBlock();
4500 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
4502 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
4503 Worklist.insert(AI);
4506 bool Changed = false;
4507 // A set of deleted alloca instruction pointers which should be removed from
4508 // the list of promotable allocas.
4509 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
4512 while (!Worklist.empty()) {
4513 Changed |= runOnAlloca(*Worklist.pop_back_val());
4514 Changed |= deleteDeadInstructions(DeletedAllocas);
4516 // Remove the deleted allocas from various lists so that we don't try to
4517 // continue processing them.
4518 if (!DeletedAllocas.empty()) {
4519 auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
4520 Worklist.remove_if(IsInSet);
4521 PostPromotionWorklist.remove_if(IsInSet);
4522 PromotableAllocas.erase(llvm::remove_if(PromotableAllocas, IsInSet),
4523 PromotableAllocas.end());
4524 DeletedAllocas.clear();
4528 Changed |= promoteAllocas(F);
4530 Worklist = PostPromotionWorklist;
4531 PostPromotionWorklist.clear();
4532 } while (!Worklist.empty());
4535 return PreservedAnalyses::all();
4537 PreservedAnalyses PA;
4538 PA.preserveSet<CFGAnalyses>();
4539 PA.preserve<GlobalsAA>();
4543 PreservedAnalyses SROA::run(Function &F, FunctionAnalysisManager &AM) {
4544 return runImpl(F, AM.getResult<DominatorTreeAnalysis>(F),
4545 AM.getResult<AssumptionAnalysis>(F));
4548 /// A legacy pass for the legacy pass manager that wraps the \c SROA pass.
4550 /// This is in the llvm namespace purely to allow it to be a friend of the \c
4552 class llvm::sroa::SROALegacyPass : public FunctionPass {
4553 /// The SROA implementation.
4559 SROALegacyPass() : FunctionPass(ID) {
4560 initializeSROALegacyPassPass(*PassRegistry::getPassRegistry());
4563 bool runOnFunction(Function &F) override {
4564 if (skipFunction(F))
4567 auto PA = Impl.runImpl(
4568 F, getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
4569 getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F));
4570 return !PA.areAllPreserved();
4573 void getAnalysisUsage(AnalysisUsage &AU) const override {
4574 AU.addRequired<AssumptionCacheTracker>();
4575 AU.addRequired<DominatorTreeWrapperPass>();
4576 AU.addPreserved<GlobalsAAWrapperPass>();
4577 AU.setPreservesCFG();
4580 StringRef getPassName() const override { return "SROA"; }
4583 char SROALegacyPass::ID = 0;
4585 FunctionPass *llvm::createSROAPass() { return new SROALegacyPass(); }
4587 INITIALIZE_PASS_BEGIN(SROALegacyPass, "sroa",
4588 "Scalar Replacement Of Aggregates", false, false)
4589 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
4590 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
4591 INITIALIZE_PASS_END(SROALegacyPass, "sroa", "Scalar Replacement Of Aggregates",