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.h"
27 #include "llvm/ADT/STLExtras.h"
28 #include "llvm/ADT/SetVector.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/Statistic.h"
31 #include "llvm/Analysis/AssumptionCache.h"
32 #include "llvm/Analysis/Loads.h"
33 #include "llvm/Analysis/PtrUseVisitor.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/IR/Constants.h"
36 #include "llvm/IR/DIBuilder.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/DebugInfo.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/InstVisitor.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/IntrinsicInst.h"
46 #include "llvm/IR/LLVMContext.h"
47 #include "llvm/IR/Operator.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Compiler.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Support/TimeValue.h"
55 #include "llvm/Support/raw_ostream.h"
56 #include "llvm/Transforms/Utils/Local.h"
57 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
58 #include "llvm/Transforms/Utils/SSAUpdater.h"
60 #if __cplusplus >= 201103L && !defined(NDEBUG)
61 // We only use this for a debug check in C++11
67 #define DEBUG_TYPE "sroa"
69 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
70 STATISTIC(NumAllocaPartitions, "Number of alloca partitions formed");
71 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions per alloca");
72 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses rewritten");
73 STATISTIC(MaxUsesPerAllocaPartition, "Maximum number of uses of a partition");
74 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
75 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
76 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
77 STATISTIC(NumDeleted, "Number of instructions deleted");
78 STATISTIC(NumVectorized, "Number of vectorized aggregates");
80 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
81 /// forming SSA values through the SSAUpdater infrastructure.
82 static cl::opt<bool> ForceSSAUpdater("force-ssa-updater", cl::init(false),
85 /// Hidden option to enable randomly shuffling the slices to help uncover
86 /// instability in their order.
87 static cl::opt<bool> SROARandomShuffleSlices("sroa-random-shuffle-slices",
88 cl::init(false), cl::Hidden);
90 /// Hidden option to experiment with completely strict handling of inbounds
92 static cl::opt<bool> SROAStrictInbounds("sroa-strict-inbounds", cl::init(false),
96 /// \brief A custom IRBuilder inserter which prefixes all names if they are
98 template <bool preserveNames = true>
99 class IRBuilderPrefixedInserter
100 : public IRBuilderDefaultInserter<preserveNames> {
104 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
107 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
108 BasicBlock::iterator InsertPt) const {
109 IRBuilderDefaultInserter<preserveNames>::InsertHelper(
110 I, Name.isTriviallyEmpty() ? Name : Prefix + Name, BB, InsertPt);
114 // Specialization for not preserving the name is trivial.
116 class IRBuilderPrefixedInserter<false>
117 : public IRBuilderDefaultInserter<false> {
119 void SetNamePrefix(const Twine &P) {}
122 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
124 typedef llvm::IRBuilder<true, ConstantFolder, IRBuilderPrefixedInserter<true>>
127 typedef llvm::IRBuilder<false, ConstantFolder, IRBuilderPrefixedInserter<false>>
133 /// \brief A used slice of an alloca.
135 /// This structure represents a slice of an alloca used by some instruction. It
136 /// stores both the begin and end offsets of this use, a pointer to the use
137 /// itself, and a flag indicating whether we can classify the use as splittable
138 /// or not when forming partitions of the alloca.
140 /// \brief The beginning offset of the range.
141 uint64_t BeginOffset;
143 /// \brief The ending offset, not included in the range.
146 /// \brief Storage for both the use of this slice and whether it can be
148 PointerIntPair<Use *, 1, bool> UseAndIsSplittable;
151 Slice() : BeginOffset(), EndOffset() {}
152 Slice(uint64_t BeginOffset, uint64_t EndOffset, Use *U, bool IsSplittable)
153 : BeginOffset(BeginOffset), EndOffset(EndOffset),
154 UseAndIsSplittable(U, IsSplittable) {}
156 uint64_t beginOffset() const { return BeginOffset; }
157 uint64_t endOffset() const { return EndOffset; }
159 bool isSplittable() const { return UseAndIsSplittable.getInt(); }
160 void makeUnsplittable() { UseAndIsSplittable.setInt(false); }
162 Use *getUse() const { return UseAndIsSplittable.getPointer(); }
164 bool isDead() const { return getUse() == nullptr; }
165 void kill() { UseAndIsSplittable.setPointer(nullptr); }
167 /// \brief Support for ordering ranges.
169 /// This provides an ordering over ranges such that start offsets are
170 /// always increasing, and within equal start offsets, the end offsets are
171 /// decreasing. Thus the spanning range comes first in a cluster with the
172 /// same start position.
173 bool operator<(const Slice &RHS) const {
174 if (beginOffset() < RHS.beginOffset())
176 if (beginOffset() > RHS.beginOffset())
178 if (isSplittable() != RHS.isSplittable())
179 return !isSplittable();
180 if (endOffset() > RHS.endOffset())
185 /// \brief Support comparison with a single offset to allow binary searches.
186 friend LLVM_ATTRIBUTE_UNUSED bool operator<(const Slice &LHS,
187 uint64_t RHSOffset) {
188 return LHS.beginOffset() < RHSOffset;
190 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
192 return LHSOffset < RHS.beginOffset();
195 bool operator==(const Slice &RHS) const {
196 return isSplittable() == RHS.isSplittable() &&
197 beginOffset() == RHS.beginOffset() && endOffset() == RHS.endOffset();
199 bool operator!=(const Slice &RHS) const { return !operator==(RHS); }
201 } // end anonymous namespace
204 template <typename T> struct isPodLike;
205 template <> struct isPodLike<Slice> { static const bool value = true; };
209 /// \brief Representation of the alloca slices.
211 /// This class represents the slices of an alloca which are formed by its
212 /// various uses. If a pointer escapes, we can't fully build a representation
213 /// for the slices used and we reflect that in this structure. The uses are
214 /// stored, sorted by increasing beginning offset and with unsplittable slices
215 /// starting at a particular offset before splittable slices.
218 /// \brief Construct the slices of a particular alloca.
219 AllocaSlices(const DataLayout &DL, AllocaInst &AI);
221 /// \brief Test whether a pointer to the allocation escapes our analysis.
223 /// If this is true, the slices are never fully built and should be
225 bool isEscaped() const { return PointerEscapingInstr; }
227 /// \brief Support for iterating over the slices.
229 typedef SmallVectorImpl<Slice>::iterator iterator;
230 typedef iterator_range<iterator> range;
231 iterator begin() { return Slices.begin(); }
232 iterator end() { return Slices.end(); }
234 typedef SmallVectorImpl<Slice>::const_iterator const_iterator;
235 typedef iterator_range<const_iterator> const_range;
236 const_iterator begin() const { return Slices.begin(); }
237 const_iterator end() const { return Slices.end(); }
240 /// \brief Erase a range of slices.
241 void erase(iterator Start, iterator Stop) { Slices.erase(Start, Stop); }
243 /// \brief Insert new slices for this alloca.
245 /// This moves the slices into the alloca's slices collection, and re-sorts
246 /// everything so that the usual ordering properties of the alloca's slices
248 void insert(ArrayRef<Slice> NewSlices) {
249 int OldSize = Slices.size();
250 Slices.append(NewSlices.begin(), NewSlices.end());
251 auto SliceI = Slices.begin() + OldSize;
252 std::sort(SliceI, Slices.end());
253 std::inplace_merge(Slices.begin(), SliceI, Slices.end());
256 // Forward declare an iterator to befriend it.
257 class partition_iterator;
259 /// \brief A partition of the slices.
261 /// An ephemeral representation for a range of slices which can be viewed as
262 /// a partition of the alloca. This range represents a span of the alloca's
263 /// memory which cannot be split, and provides access to all of the slices
264 /// overlapping some part of the partition.
266 /// Objects of this type are produced by traversing the alloca's slices, but
267 /// are only ephemeral and not persistent.
270 friend class AllocaSlices;
271 friend class AllocaSlices::partition_iterator;
273 /// \brief The beginning and ending offsets of the alloca for this
275 uint64_t BeginOffset, EndOffset;
277 /// \brief The start end end iterators of this partition.
280 /// \brief A collection of split slice tails overlapping the partition.
281 SmallVector<Slice *, 4> SplitTails;
283 /// \brief Raw constructor builds an empty partition starting and ending at
284 /// the given iterator.
285 Partition(iterator SI) : SI(SI), SJ(SI) {}
288 /// \brief The start offset of this partition.
290 /// All of the contained slices start at or after this offset.
291 uint64_t beginOffset() const { return BeginOffset; }
293 /// \brief The end offset of this partition.
295 /// All of the contained slices end at or before this offset.
296 uint64_t endOffset() const { return EndOffset; }
298 /// \brief The size of the partition.
300 /// Note that this can never be zero.
301 uint64_t size() const {
302 assert(BeginOffset < EndOffset && "Partitions must span some bytes!");
303 return EndOffset - BeginOffset;
306 /// \brief Test whether this partition contains no slices, and merely spans
307 /// a region occupied by split slices.
308 bool empty() const { return SI == SJ; }
310 /// \name Iterate slices that start within the partition.
311 /// These may be splittable or unsplittable. They have a begin offset >= the
312 /// partition begin offset.
314 // FIXME: We should probably define a "concat_iterator" helper and use that
315 // to stitch together pointee_iterators over the split tails and the
316 // contiguous iterators of the partition. That would give a much nicer
317 // interface here. We could then additionally expose filtered iterators for
318 // split, unsplit, and unsplittable splices based on the usage patterns.
319 iterator begin() const { return SI; }
320 iterator end() const { return SJ; }
323 /// \brief Get the sequence of split slice tails.
325 /// These tails are of slices which start before this partition but are
326 /// split and overlap into the partition. We accumulate these while forming
328 ArrayRef<Slice *> splitSliceTails() const { return SplitTails; }
331 /// \brief An iterator over partitions of the alloca's slices.
333 /// This iterator implements the core algorithm for partitioning the alloca's
334 /// slices. It is a forward iterator as we don't support backtracking for
335 /// efficiency reasons, and re-use a single storage area to maintain the
336 /// current set of split slices.
338 /// It is templated on the slice iterator type to use so that it can operate
339 /// with either const or non-const slice iterators.
340 class partition_iterator
341 : public iterator_facade_base<partition_iterator,
342 std::forward_iterator_tag, Partition> {
343 friend class AllocaSlices;
345 /// \brief Most of the state for walking the partitions is held in a class
346 /// with a nice interface for examining them.
349 /// \brief We need to keep the end of the slices to know when to stop.
350 AllocaSlices::iterator SE;
352 /// \brief We also need to keep track of the maximum split end offset seen.
353 /// FIXME: Do we really?
354 uint64_t MaxSplitSliceEndOffset;
356 /// \brief Sets the partition to be empty at given iterator, and sets the
358 partition_iterator(AllocaSlices::iterator SI, AllocaSlices::iterator SE)
359 : P(SI), SE(SE), MaxSplitSliceEndOffset(0) {
360 // If not already at the end, advance our state to form the initial
366 /// \brief Advance the iterator to the next partition.
368 /// Requires that the iterator not be at the end of the slices.
370 assert((P.SI != SE || !P.SplitTails.empty()) &&
371 "Cannot advance past the end of the slices!");
373 // Clear out any split uses which have ended.
374 if (!P.SplitTails.empty()) {
375 if (P.EndOffset >= MaxSplitSliceEndOffset) {
376 // If we've finished all splits, this is easy.
377 P.SplitTails.clear();
378 MaxSplitSliceEndOffset = 0;
380 // Remove the uses which have ended in the prior partition. This
381 // cannot change the max split slice end because we just checked that
382 // the prior partition ended prior to that max.
385 P.SplitTails.begin(), P.SplitTails.end(),
386 [&](Slice *S) { return S->endOffset() <= P.EndOffset; }),
388 assert(std::any_of(P.SplitTails.begin(), P.SplitTails.end(),
390 return S->endOffset() == MaxSplitSliceEndOffset;
392 "Could not find the current max split slice offset!");
393 assert(std::all_of(P.SplitTails.begin(), P.SplitTails.end(),
395 return S->endOffset() <= MaxSplitSliceEndOffset;
397 "Max split slice end offset is not actually the max!");
401 // If P.SI is already at the end, then we've cleared the split tail and
402 // now have an end iterator.
404 assert(P.SplitTails.empty() && "Failed to clear the split slices!");
408 // If we had a non-empty partition previously, set up the state for
409 // subsequent partitions.
411 // Accumulate all the splittable slices which started in the old
412 // partition into the split list.
414 if (S.isSplittable() && S.endOffset() > P.EndOffset) {
415 P.SplitTails.push_back(&S);
416 MaxSplitSliceEndOffset =
417 std::max(S.endOffset(), MaxSplitSliceEndOffset);
420 // Start from the end of the previous partition.
423 // If P.SI is now at the end, we at most have a tail of split slices.
425 P.BeginOffset = P.EndOffset;
426 P.EndOffset = MaxSplitSliceEndOffset;
430 // If the we have split slices and the next slice is after a gap and is
431 // not splittable immediately form an empty partition for the split
432 // slices up until the next slice begins.
433 if (!P.SplitTails.empty() && P.SI->beginOffset() != P.EndOffset &&
434 !P.SI->isSplittable()) {
435 P.BeginOffset = P.EndOffset;
436 P.EndOffset = P.SI->beginOffset();
441 // OK, we need to consume new slices. Set the end offset based on the
442 // current slice, and step SJ past it. The beginning offset of the
443 // partition is the beginning offset of the next slice unless we have
444 // pre-existing split slices that are continuing, in which case we begin
445 // at the prior end offset.
446 P.BeginOffset = P.SplitTails.empty() ? P.SI->beginOffset() : P.EndOffset;
447 P.EndOffset = P.SI->endOffset();
450 // There are two strategies to form a partition based on whether the
451 // partition starts with an unsplittable slice or a splittable slice.
452 if (!P.SI->isSplittable()) {
453 // When we're forming an unsplittable region, it must always start at
454 // the first slice and will extend through its end.
455 assert(P.BeginOffset == P.SI->beginOffset());
457 // Form a partition including all of the overlapping slices with this
458 // unsplittable slice.
459 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
460 if (!P.SJ->isSplittable())
461 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
465 // We have a partition across a set of overlapping unsplittable
470 // If we're starting with a splittable slice, then we need to form
471 // a synthetic partition spanning it and any other overlapping splittable
473 assert(P.SI->isSplittable() && "Forming a splittable partition!");
475 // Collect all of the overlapping splittable slices.
476 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
477 P.SJ->isSplittable()) {
478 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
482 // Back upiP.EndOffset if we ended the span early when encountering an
483 // unsplittable slice. This synthesizes the early end offset of
484 // a partition spanning only splittable slices.
485 if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
486 assert(!P.SJ->isSplittable());
487 P.EndOffset = P.SJ->beginOffset();
492 bool operator==(const partition_iterator &RHS) const {
493 assert(SE == RHS.SE &&
494 "End iterators don't match between compared partition iterators!");
496 // The observed positions of partitions is marked by the P.SI iterator and
497 // the emptiness of the split slices. The latter is only relevant when
498 // P.SI == SE, as the end iterator will additionally have an empty split
499 // slices list, but the prior may have the same P.SI and a tail of split
501 if (P.SI == RHS.P.SI &&
502 P.SplitTails.empty() == RHS.P.SplitTails.empty()) {
503 assert(P.SJ == RHS.P.SJ &&
504 "Same set of slices formed two different sized partitions!");
505 assert(P.SplitTails.size() == RHS.P.SplitTails.size() &&
506 "Same slice position with differently sized non-empty split "
513 partition_iterator &operator++() {
518 Partition &operator*() { return P; }
521 /// \brief A forward range over the partitions of the alloca's slices.
523 /// This accesses an iterator range over the partitions of the alloca's
524 /// slices. It computes these partitions on the fly based on the overlapping
525 /// offsets of the slices and the ability to split them. It will visit "empty"
526 /// partitions to cover regions of the alloca only accessed via split
528 iterator_range<partition_iterator> partitions() {
529 return make_range(partition_iterator(begin(), end()),
530 partition_iterator(end(), end()));
533 /// \brief Access the dead users for this alloca.
534 ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
536 /// \brief Access the dead operands referring to this alloca.
538 /// These are operands which have cannot actually be used to refer to the
539 /// alloca as they are outside its range and the user doesn't correct for
540 /// that. These mostly consist of PHI node inputs and the like which we just
541 /// need to replace with undef.
542 ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
544 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
545 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
546 void printSlice(raw_ostream &OS, const_iterator I,
547 StringRef Indent = " ") const;
548 void printUse(raw_ostream &OS, const_iterator I,
549 StringRef Indent = " ") const;
550 void print(raw_ostream &OS) const;
551 void dump(const_iterator I) const;
556 template <typename DerivedT, typename RetT = void> class BuilderBase;
558 friend class AllocaSlices::SliceBuilder;
560 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
561 /// \brief Handle to alloca instruction to simplify method interfaces.
565 /// \brief The instruction responsible for this alloca not having a known set
568 /// When an instruction (potentially) escapes the pointer to the alloca, we
569 /// store a pointer to that here and abort trying to form slices of the
570 /// alloca. This will be null if the alloca slices are analyzed successfully.
571 Instruction *PointerEscapingInstr;
573 /// \brief The slices of the alloca.
575 /// We store a vector of the slices formed by uses of the alloca here. This
576 /// vector is sorted by increasing begin offset, and then the unsplittable
577 /// slices before the splittable ones. See the Slice inner class for more
579 SmallVector<Slice, 8> Slices;
581 /// \brief Instructions which will become dead if we rewrite the alloca.
583 /// Note that these are not separated by slice. This is because we expect an
584 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
585 /// all these instructions can simply be removed and replaced with undef as
586 /// they come from outside of the allocated space.
587 SmallVector<Instruction *, 8> DeadUsers;
589 /// \brief Operands which will become dead if we rewrite the alloca.
591 /// These are operands that in their particular use can be replaced with
592 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
593 /// to PHI nodes and the like. They aren't entirely dead (there might be
594 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
595 /// want to swap this particular input for undef to simplify the use lists of
597 SmallVector<Use *, 8> DeadOperands;
601 static Value *foldSelectInst(SelectInst &SI) {
602 // If the condition being selected on is a constant or the same value is
603 // being selected between, fold the select. Yes this does (rarely) happen
605 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
606 return SI.getOperand(1 + CI->isZero());
607 if (SI.getOperand(1) == SI.getOperand(2))
608 return SI.getOperand(1);
613 /// \brief A helper that folds a PHI node or a select.
614 static Value *foldPHINodeOrSelectInst(Instruction &I) {
615 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
616 // If PN merges together the same value, return that value.
617 return PN->hasConstantValue();
619 return foldSelectInst(cast<SelectInst>(I));
622 /// \brief Builder for the alloca slices.
624 /// This class builds a set of alloca slices by recursively visiting the uses
625 /// of an alloca and making a slice for each load and store at each offset.
626 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
627 friend class PtrUseVisitor<SliceBuilder>;
628 friend class InstVisitor<SliceBuilder>;
629 typedef PtrUseVisitor<SliceBuilder> Base;
631 const uint64_t AllocSize;
634 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
635 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
637 /// \brief Set to de-duplicate dead instructions found in the use walk.
638 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
641 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
642 : PtrUseVisitor<SliceBuilder>(DL),
643 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), AS(AS) {}
646 void markAsDead(Instruction &I) {
647 if (VisitedDeadInsts.insert(&I).second)
648 AS.DeadUsers.push_back(&I);
651 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
652 bool IsSplittable = false) {
653 // Completely skip uses which have a zero size or start either before or
654 // past the end of the allocation.
655 if (Size == 0 || Offset.uge(AllocSize)) {
656 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
657 << " which has zero size or starts outside of the "
658 << AllocSize << " byte alloca:\n"
659 << " alloca: " << AS.AI << "\n"
660 << " use: " << I << "\n");
661 return markAsDead(I);
664 uint64_t BeginOffset = Offset.getZExtValue();
665 uint64_t EndOffset = BeginOffset + Size;
667 // Clamp the end offset to the end of the allocation. Note that this is
668 // formulated to handle even the case where "BeginOffset + Size" overflows.
669 // This may appear superficially to be something we could ignore entirely,
670 // but that is not so! There may be widened loads or PHI-node uses where
671 // some instructions are dead but not others. We can't completely ignore
672 // them, and so have to record at least the information here.
673 assert(AllocSize >= BeginOffset); // Established above.
674 if (Size > AllocSize - BeginOffset) {
675 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
676 << " to remain within the " << AllocSize << " byte alloca:\n"
677 << " alloca: " << AS.AI << "\n"
678 << " use: " << I << "\n");
679 EndOffset = AllocSize;
682 AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
685 void visitBitCastInst(BitCastInst &BC) {
687 return markAsDead(BC);
689 return Base::visitBitCastInst(BC);
692 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
693 if (GEPI.use_empty())
694 return markAsDead(GEPI);
696 if (SROAStrictInbounds && GEPI.isInBounds()) {
697 // FIXME: This is a manually un-factored variant of the basic code inside
698 // of GEPs with checking of the inbounds invariant specified in the
699 // langref in a very strict sense. If we ever want to enable
700 // SROAStrictInbounds, this code should be factored cleanly into
701 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
702 // by writing out the code here where we have tho underlying allocation
703 // size readily available.
704 APInt GEPOffset = Offset;
705 const DataLayout &DL = GEPI.getModule()->getDataLayout();
706 for (gep_type_iterator GTI = gep_type_begin(GEPI),
707 GTE = gep_type_end(GEPI);
709 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
713 // Handle a struct index, which adds its field offset to the pointer.
714 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
715 unsigned ElementIdx = OpC->getZExtValue();
716 const StructLayout *SL = DL.getStructLayout(STy);
718 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
720 // For array or vector indices, scale the index by the size of the
722 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
723 GEPOffset += Index * APInt(Offset.getBitWidth(),
724 DL.getTypeAllocSize(GTI.getIndexedType()));
727 // If this index has computed an intermediate pointer which is not
728 // inbounds, then the result of the GEP is a poison value and we can
729 // delete it and all uses.
730 if (GEPOffset.ugt(AllocSize))
731 return markAsDead(GEPI);
735 return Base::visitGetElementPtrInst(GEPI);
738 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
739 uint64_t Size, bool IsVolatile) {
740 // We allow splitting of non-volatile loads and stores where the type is an
741 // integer type. These may be used to implement 'memcpy' or other "transfer
742 // of bits" patterns.
743 bool IsSplittable = Ty->isIntegerTy() && !IsVolatile;
745 insertUse(I, Offset, Size, IsSplittable);
748 void visitLoadInst(LoadInst &LI) {
749 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
750 "All simple FCA loads should have been pre-split");
753 return PI.setAborted(&LI);
755 const DataLayout &DL = LI.getModule()->getDataLayout();
756 uint64_t Size = DL.getTypeStoreSize(LI.getType());
757 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
760 void visitStoreInst(StoreInst &SI) {
761 Value *ValOp = SI.getValueOperand();
763 return PI.setEscapedAndAborted(&SI);
765 return PI.setAborted(&SI);
767 const DataLayout &DL = SI.getModule()->getDataLayout();
768 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
770 // If this memory access can be shown to *statically* extend outside the
771 // bounds of of the allocation, it's behavior is undefined, so simply
772 // ignore it. Note that this is more strict than the generic clamping
773 // behavior of insertUse. We also try to handle cases which might run the
775 // FIXME: We should instead consider the pointer to have escaped if this
776 // function is being instrumented for addressing bugs or race conditions.
777 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
778 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
779 << " which extends past the end of the " << AllocSize
781 << " alloca: " << AS.AI << "\n"
782 << " use: " << SI << "\n");
783 return markAsDead(SI);
786 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
787 "All simple FCA stores should have been pre-split");
788 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
791 void visitMemSetInst(MemSetInst &II) {
792 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
793 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
794 if ((Length && Length->getValue() == 0) ||
795 (IsOffsetKnown && Offset.uge(AllocSize)))
796 // Zero-length mem transfer intrinsics can be ignored entirely.
797 return markAsDead(II);
800 return PI.setAborted(&II);
802 insertUse(II, Offset, Length ? Length->getLimitedValue()
803 : AllocSize - Offset.getLimitedValue(),
807 void visitMemTransferInst(MemTransferInst &II) {
808 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
809 if (Length && Length->getValue() == 0)
810 // Zero-length mem transfer intrinsics can be ignored entirely.
811 return markAsDead(II);
813 // Because we can visit these intrinsics twice, also check to see if the
814 // first time marked this instruction as dead. If so, skip it.
815 if (VisitedDeadInsts.count(&II))
819 return PI.setAborted(&II);
821 // This side of the transfer is completely out-of-bounds, and so we can
822 // nuke the entire transfer. However, we also need to nuke the other side
823 // if already added to our partitions.
824 // FIXME: Yet another place we really should bypass this when
825 // instrumenting for ASan.
826 if (Offset.uge(AllocSize)) {
827 SmallDenseMap<Instruction *, unsigned>::iterator MTPI =
828 MemTransferSliceMap.find(&II);
829 if (MTPI != MemTransferSliceMap.end())
830 AS.Slices[MTPI->second].kill();
831 return markAsDead(II);
834 uint64_t RawOffset = Offset.getLimitedValue();
835 uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
837 // Check for the special case where the same exact value is used for both
839 if (*U == II.getRawDest() && *U == II.getRawSource()) {
840 // For non-volatile transfers this is a no-op.
841 if (!II.isVolatile())
842 return markAsDead(II);
844 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
847 // If we have seen both source and destination for a mem transfer, then
848 // they both point to the same alloca.
850 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
851 std::tie(MTPI, Inserted) =
852 MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
853 unsigned PrevIdx = MTPI->second;
855 Slice &PrevP = AS.Slices[PrevIdx];
857 // Check if the begin offsets match and this is a non-volatile transfer.
858 // In that case, we can completely elide the transfer.
859 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
861 return markAsDead(II);
864 // Otherwise we have an offset transfer within the same alloca. We can't
866 PrevP.makeUnsplittable();
869 // Insert the use now that we've fixed up the splittable nature.
870 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
872 // Check that we ended up with a valid index in the map.
873 assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&
874 "Map index doesn't point back to a slice with this user.");
877 // Disable SRoA for any intrinsics except for lifetime invariants.
878 // FIXME: What about debug intrinsics? This matches old behavior, but
879 // doesn't make sense.
880 void visitIntrinsicInst(IntrinsicInst &II) {
882 return PI.setAborted(&II);
884 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
885 II.getIntrinsicID() == Intrinsic::lifetime_end) {
886 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
887 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
888 Length->getLimitedValue());
889 insertUse(II, Offset, Size, true);
893 Base::visitIntrinsicInst(II);
896 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
897 // We consider any PHI or select that results in a direct load or store of
898 // the same offset to be a viable use for slicing purposes. These uses
899 // are considered unsplittable and the size is the maximum loaded or stored
901 SmallPtrSet<Instruction *, 4> Visited;
902 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
903 Visited.insert(Root);
904 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
905 const DataLayout &DL = Root->getModule()->getDataLayout();
906 // If there are no loads or stores, the access is dead. We mark that as
907 // a size zero access.
910 Instruction *I, *UsedI;
911 std::tie(UsedI, I) = Uses.pop_back_val();
913 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
914 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
917 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
918 Value *Op = SI->getOperand(0);
921 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
925 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
926 if (!GEP->hasAllZeroIndices())
928 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
929 !isa<SelectInst>(I)) {
933 for (User *U : I->users())
934 if (Visited.insert(cast<Instruction>(U)).second)
935 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
936 } while (!Uses.empty());
941 void visitPHINodeOrSelectInst(Instruction &I) {
942 assert(isa<PHINode>(I) || isa<SelectInst>(I));
944 return markAsDead(I);
946 // TODO: We could use SimplifyInstruction here to fold PHINodes and
947 // SelectInsts. However, doing so requires to change the current
948 // dead-operand-tracking mechanism. For instance, suppose neither loading
949 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
950 // trap either. However, if we simply replace %U with undef using the
951 // current dead-operand-tracking mechanism, "load (select undef, undef,
952 // %other)" may trap because the select may return the first operand
954 if (Value *Result = foldPHINodeOrSelectInst(I)) {
956 // If the result of the constant fold will be the pointer, recurse
957 // through the PHI/select as if we had RAUW'ed it.
960 // Otherwise the operand to the PHI/select is dead, and we can replace
962 AS.DeadOperands.push_back(U);
968 return PI.setAborted(&I);
970 // See if we already have computed info on this node.
971 uint64_t &Size = PHIOrSelectSizes[&I];
973 // This is a new PHI/Select, check for an unsafe use of it.
974 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
975 return PI.setAborted(UnsafeI);
978 // For PHI and select operands outside the alloca, we can't nuke the entire
979 // phi or select -- the other side might still be relevant, so we special
980 // case them here and use a separate structure to track the operands
981 // themselves which should be replaced with undef.
982 // FIXME: This should instead be escaped in the event we're instrumenting
983 // for address sanitization.
984 if (Offset.uge(AllocSize)) {
985 AS.DeadOperands.push_back(U);
989 insertUse(I, Offset, Size);
992 void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
994 void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
996 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
997 void visitInstruction(Instruction &I) { PI.setAborted(&I); }
1000 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
1002 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1005 PointerEscapingInstr(nullptr) {
1006 SliceBuilder PB(DL, AI, *this);
1007 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
1008 if (PtrI.isEscaped() || PtrI.isAborted()) {
1009 // FIXME: We should sink the escape vs. abort info into the caller nicely,
1010 // possibly by just storing the PtrInfo in the AllocaSlices.
1011 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
1012 : PtrI.getAbortingInst();
1013 assert(PointerEscapingInstr && "Did not track a bad instruction");
1017 Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
1018 [](const Slice &S) {
1023 #if __cplusplus >= 201103L && !defined(NDEBUG)
1024 if (SROARandomShuffleSlices) {
1025 std::mt19937 MT(static_cast<unsigned>(sys::TimeValue::now().msec()));
1026 std::shuffle(Slices.begin(), Slices.end(), MT);
1030 // Sort the uses. This arranges for the offsets to be in ascending order,
1031 // and the sizes to be in descending order.
1032 std::sort(Slices.begin(), Slices.end());
1035 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1037 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1038 StringRef Indent) const {
1039 printSlice(OS, I, Indent);
1041 printUse(OS, I, Indent);
1044 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1045 StringRef Indent) const {
1046 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1047 << " slice #" << (I - begin())
1048 << (I->isSplittable() ? " (splittable)" : "");
1051 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1052 StringRef Indent) const {
1053 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1056 void AllocaSlices::print(raw_ostream &OS) const {
1057 if (PointerEscapingInstr) {
1058 OS << "Can't analyze slices for alloca: " << AI << "\n"
1059 << " A pointer to this alloca escaped by:\n"
1060 << " " << *PointerEscapingInstr << "\n";
1064 OS << "Slices of alloca: " << AI << "\n";
1065 for (const_iterator I = begin(), E = end(); I != E; ++I)
1069 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
1072 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
1074 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1077 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1079 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1080 /// the loads and stores of an alloca instruction, as well as updating its
1081 /// debug information. This is used when a domtree is unavailable and thus
1082 /// mem2reg in its full form can't be used to handle promotion of allocas to
1084 class AllocaPromoter : public LoadAndStorePromoter {
1088 SmallVector<DbgDeclareInst *, 4> DDIs;
1089 SmallVector<DbgValueInst *, 4> DVIs;
1092 AllocaPromoter(ArrayRef<const Instruction *> Insts,
1094 AllocaInst &AI, DIBuilder &DIB)
1095 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1097 void run(const SmallVectorImpl<Instruction *> &Insts) {
1098 // Retain the debug information attached to the alloca for use when
1099 // rewriting loads and stores.
1100 if (auto *L = LocalAsMetadata::getIfExists(&AI)) {
1101 if (auto *DINode = MetadataAsValue::getIfExists(AI.getContext(), L)) {
1102 for (User *U : DINode->users())
1103 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
1104 DDIs.push_back(DDI);
1105 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U))
1106 DVIs.push_back(DVI);
1110 LoadAndStorePromoter::run(Insts);
1112 // While we have the debug information, clear it off of the alloca. The
1113 // caller takes care of deleting the alloca.
1114 while (!DDIs.empty())
1115 DDIs.pop_back_val()->eraseFromParent();
1116 while (!DVIs.empty())
1117 DVIs.pop_back_val()->eraseFromParent();
1121 isInstInList(Instruction *I,
1122 const SmallVectorImpl<Instruction *> &Insts) const override {
1124 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1125 Ptr = LI->getOperand(0);
1127 Ptr = cast<StoreInst>(I)->getPointerOperand();
1129 // Only used to detect cycles, which will be rare and quickly found as
1130 // we're walking up a chain of defs rather than down through uses.
1131 SmallPtrSet<Value *, 4> Visited;
1137 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr))
1138 Ptr = BCI->getOperand(0);
1139 else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr))
1140 Ptr = GEPI->getPointerOperand();
1144 } while (Visited.insert(Ptr).second);
1149 void updateDebugInfo(Instruction *Inst) const override {
1150 for (DbgDeclareInst *DDI : DDIs)
1151 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1152 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1153 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1154 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1155 for (DbgValueInst *DVI : DVIs) {
1156 Value *Arg = nullptr;
1157 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1158 // If an argument is zero extended then use argument directly. The ZExt
1159 // may be zapped by an optimization pass in future.
1160 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1161 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1162 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1163 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1165 Arg = SI->getValueOperand();
1166 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1167 Arg = LI->getPointerOperand();
1171 DIB.insertDbgValueIntrinsic(Arg, 0, DVI->getVariable(),
1172 DVI->getExpression(), DVI->getDebugLoc(),
1177 } // end anon namespace
1180 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1182 /// This pass takes allocations which can be completely analyzed (that is, they
1183 /// don't escape) and tries to turn them into scalar SSA values. There are
1184 /// a few steps to this process.
1186 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1187 /// are used to try to split them into smaller allocations, ideally of
1188 /// a single scalar data type. It will split up memcpy and memset accesses
1189 /// as necessary and try to isolate individual scalar accesses.
1190 /// 2) It will transform accesses into forms which are suitable for SSA value
1191 /// promotion. This can be replacing a memset with a scalar store of an
1192 /// integer value, or it can involve speculating operations on a PHI or
1193 /// select to be a PHI or select of the results.
1194 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1195 /// onto insert and extract operations on a vector value, and convert them to
1196 /// this form. By doing so, it will enable promotion of vector aggregates to
1197 /// SSA vector values.
1198 class SROA : public FunctionPass {
1199 const bool RequiresDomTree;
1203 AssumptionCache *AC;
1205 /// \brief Worklist of alloca instructions to simplify.
1207 /// Each alloca in the function is added to this. Each new alloca formed gets
1208 /// added to it as well to recursively simplify unless that alloca can be
1209 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1210 /// the one being actively rewritten, we add it back onto the list if not
1211 /// already present to ensure it is re-visited.
1212 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16>> Worklist;
1214 /// \brief A collection of instructions to delete.
1215 /// We try to batch deletions to simplify code and make things a bit more
1217 SetVector<Instruction *, SmallVector<Instruction *, 8>> DeadInsts;
1219 /// \brief Post-promotion worklist.
1221 /// Sometimes we discover an alloca which has a high probability of becoming
1222 /// viable for SROA after a round of promotion takes place. In those cases,
1223 /// the alloca is enqueued here for re-processing.
1225 /// Note that we have to be very careful to clear allocas out of this list in
1226 /// the event they are deleted.
1227 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16>> PostPromotionWorklist;
1229 /// \brief A collection of alloca instructions we can directly promote.
1230 std::vector<AllocaInst *> PromotableAllocas;
1232 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
1234 /// All of these PHIs have been checked for the safety of speculation and by
1235 /// being speculated will allow promoting allocas currently in the promotable
1237 SetVector<PHINode *, SmallVector<PHINode *, 2>> SpeculatablePHIs;
1239 /// \brief A worklist of select instructions to speculate prior to promoting
1242 /// All of these select instructions have been checked for the safety of
1243 /// speculation and by being speculated will allow promoting allocas
1244 /// currently in the promotable queue.
1245 SetVector<SelectInst *, SmallVector<SelectInst *, 2>> SpeculatableSelects;
1248 SROA(bool RequiresDomTree = true)
1249 : FunctionPass(ID), RequiresDomTree(RequiresDomTree), C(nullptr),
1251 initializeSROAPass(*PassRegistry::getPassRegistry());
1253 bool runOnFunction(Function &F) override;
1254 void getAnalysisUsage(AnalysisUsage &AU) const override;
1256 const char *getPassName() const override { return "SROA"; }
1260 friend class PHIOrSelectSpeculator;
1261 friend class AllocaSliceRewriter;
1263 bool presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS);
1264 AllocaInst *rewritePartition(AllocaInst &AI, AllocaSlices &AS,
1265 AllocaSlices::Partition &P);
1266 bool splitAlloca(AllocaInst &AI, AllocaSlices &AS);
1267 bool runOnAlloca(AllocaInst &AI);
1268 void clobberUse(Use &U);
1269 void deleteDeadInstructions(SmallPtrSetImpl<AllocaInst *> &DeletedAllocas);
1270 bool promoteAllocas(Function &F);
1276 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1277 return new SROA(RequiresDomTree);
1280 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates", false,
1282 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1283 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1284 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates", false,
1287 /// Walk the range of a partitioning looking for a common type to cover this
1288 /// sequence of slices.
1289 static Type *findCommonType(AllocaSlices::const_iterator B,
1290 AllocaSlices::const_iterator E,
1291 uint64_t EndOffset) {
1293 bool TyIsCommon = true;
1294 IntegerType *ITy = nullptr;
1296 // Note that we need to look at *every* alloca slice's Use to ensure we
1297 // always get consistent results regardless of the order of slices.
1298 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1299 Use *U = I->getUse();
1300 if (isa<IntrinsicInst>(*U->getUser()))
1302 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1305 Type *UserTy = nullptr;
1306 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1307 UserTy = LI->getType();
1308 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1309 UserTy = SI->getValueOperand()->getType();
1312 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1313 // If the type is larger than the partition, skip it. We only encounter
1314 // this for split integer operations where we want to use the type of the
1315 // entity causing the split. Also skip if the type is not a byte width
1317 if (UserITy->getBitWidth() % 8 != 0 ||
1318 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1321 // Track the largest bitwidth integer type used in this way in case there
1322 // is no common type.
1323 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1327 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1328 // depend on types skipped above.
1329 if (!UserTy || (Ty && Ty != UserTy))
1330 TyIsCommon = false; // Give up on anything but an iN type.
1335 return TyIsCommon ? Ty : ITy;
1338 /// PHI instructions that use an alloca and are subsequently loaded can be
1339 /// rewritten to load both input pointers in the pred blocks and then PHI the
1340 /// results, allowing the load of the alloca to be promoted.
1342 /// %P2 = phi [i32* %Alloca, i32* %Other]
1343 /// %V = load i32* %P2
1345 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1347 /// %V2 = load i32* %Other
1349 /// %V = phi [i32 %V1, i32 %V2]
1351 /// We can do this to a select if its only uses are loads and if the operands
1352 /// to the select can be loaded unconditionally.
1354 /// FIXME: This should be hoisted into a generic utility, likely in
1355 /// Transforms/Util/Local.h
1356 static bool isSafePHIToSpeculate(PHINode &PN) {
1357 // For now, we can only do this promotion if the load is in the same block
1358 // as the PHI, and if there are no stores between the phi and load.
1359 // TODO: Allow recursive phi users.
1360 // TODO: Allow stores.
1361 BasicBlock *BB = PN.getParent();
1362 unsigned MaxAlign = 0;
1363 bool HaveLoad = false;
1364 for (User *U : PN.users()) {
1365 LoadInst *LI = dyn_cast<LoadInst>(U);
1366 if (!LI || !LI->isSimple())
1369 // For now we only allow loads in the same block as the PHI. This is
1370 // a common case that happens when instcombine merges two loads through
1372 if (LI->getParent() != BB)
1375 // Ensure that there are no instructions between the PHI and the load that
1377 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1378 if (BBI->mayWriteToMemory())
1381 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1388 const DataLayout &DL = PN.getModule()->getDataLayout();
1390 // We can only transform this if it is safe to push the loads into the
1391 // predecessor blocks. The only thing to watch out for is that we can't put
1392 // a possibly trapping load in the predecessor if it is a critical edge.
1393 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1394 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1395 Value *InVal = PN.getIncomingValue(Idx);
1397 // If the value is produced by the terminator of the predecessor (an
1398 // invoke) or it has side-effects, there is no valid place to put a load
1399 // in the predecessor.
1400 if (TI == InVal || TI->mayHaveSideEffects())
1403 // If the predecessor has a single successor, then the edge isn't
1405 if (TI->getNumSuccessors() == 1)
1408 // If this pointer is always safe to load, or if we can prove that there
1409 // is already a load in the block, then we can move the load to the pred
1411 if (isDereferenceablePointer(InVal, DL) ||
1412 isSafeToLoadUnconditionally(InVal, TI, MaxAlign))
1421 static void speculatePHINodeLoads(PHINode &PN) {
1422 DEBUG(dbgs() << " original: " << PN << "\n");
1424 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1425 IRBuilderTy PHIBuilder(&PN);
1426 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1427 PN.getName() + ".sroa.speculated");
1429 // Get the AA tags and alignment to use from one of the loads. It doesn't
1430 // matter which one we get and if any differ.
1431 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1434 SomeLoad->getAAMetadata(AATags);
1435 unsigned Align = SomeLoad->getAlignment();
1437 // Rewrite all loads of the PN to use the new PHI.
1438 while (!PN.use_empty()) {
1439 LoadInst *LI = cast<LoadInst>(PN.user_back());
1440 LI->replaceAllUsesWith(NewPN);
1441 LI->eraseFromParent();
1444 // Inject loads into all of the pred blocks.
1445 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1446 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1447 TerminatorInst *TI = Pred->getTerminator();
1448 Value *InVal = PN.getIncomingValue(Idx);
1449 IRBuilderTy PredBuilder(TI);
1451 LoadInst *Load = PredBuilder.CreateLoad(
1452 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1453 ++NumLoadsSpeculated;
1454 Load->setAlignment(Align);
1456 Load->setAAMetadata(AATags);
1457 NewPN->addIncoming(Load, Pred);
1460 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1461 PN.eraseFromParent();
1464 /// Select instructions that use an alloca and are subsequently loaded can be
1465 /// rewritten to load both input pointers and then select between the result,
1466 /// allowing the load of the alloca to be promoted.
1468 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1469 /// %V = load i32* %P2
1471 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1472 /// %V2 = load i32* %Other
1473 /// %V = select i1 %cond, i32 %V1, i32 %V2
1475 /// We can do this to a select if its only uses are loads and if the operand
1476 /// to the select can be loaded unconditionally.
1477 static bool isSafeSelectToSpeculate(SelectInst &SI) {
1478 Value *TValue = SI.getTrueValue();
1479 Value *FValue = SI.getFalseValue();
1480 const DataLayout &DL = SI.getModule()->getDataLayout();
1481 bool TDerefable = isDereferenceablePointer(TValue, DL);
1482 bool FDerefable = isDereferenceablePointer(FValue, DL);
1484 for (User *U : SI.users()) {
1485 LoadInst *LI = dyn_cast<LoadInst>(U);
1486 if (!LI || !LI->isSimple())
1489 // Both operands to the select need to be dereferencable, either
1490 // absolutely (e.g. allocas) or at this point because we can see other
1493 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment()))
1496 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment()))
1503 static void speculateSelectInstLoads(SelectInst &SI) {
1504 DEBUG(dbgs() << " original: " << SI << "\n");
1506 IRBuilderTy IRB(&SI);
1507 Value *TV = SI.getTrueValue();
1508 Value *FV = SI.getFalseValue();
1509 // Replace the loads of the select with a select of two loads.
1510 while (!SI.use_empty()) {
1511 LoadInst *LI = cast<LoadInst>(SI.user_back());
1512 assert(LI->isSimple() && "We only speculate simple loads");
1514 IRB.SetInsertPoint(LI);
1516 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1518 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1519 NumLoadsSpeculated += 2;
1521 // Transfer alignment and AA info if present.
1522 TL->setAlignment(LI->getAlignment());
1523 FL->setAlignment(LI->getAlignment());
1526 LI->getAAMetadata(Tags);
1528 TL->setAAMetadata(Tags);
1529 FL->setAAMetadata(Tags);
1532 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1533 LI->getName() + ".sroa.speculated");
1535 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1536 LI->replaceAllUsesWith(V);
1537 LI->eraseFromParent();
1539 SI.eraseFromParent();
1542 /// \brief Build a GEP out of a base pointer and indices.
1544 /// This will return the BasePtr if that is valid, or build a new GEP
1545 /// instruction using the IRBuilder if GEP-ing is needed.
1546 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1547 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1548 if (Indices.empty())
1551 // A single zero index is a no-op, so check for this and avoid building a GEP
1553 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1556 return IRB.CreateInBoundsGEP(nullptr, BasePtr, Indices,
1557 NamePrefix + "sroa_idx");
1560 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1561 /// TargetTy without changing the offset of the pointer.
1563 /// This routine assumes we've already established a properly offset GEP with
1564 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1565 /// zero-indices down through type layers until we find one the same as
1566 /// TargetTy. If we can't find one with the same type, we at least try to use
1567 /// one with the same size. If none of that works, we just produce the GEP as
1568 /// indicated by Indices to have the correct offset.
1569 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1570 Value *BasePtr, Type *Ty, Type *TargetTy,
1571 SmallVectorImpl<Value *> &Indices,
1574 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1576 // Pointer size to use for the indices.
1577 unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1579 // See if we can descend into a struct and locate a field with the correct
1581 unsigned NumLayers = 0;
1582 Type *ElementTy = Ty;
1584 if (ElementTy->isPointerTy())
1587 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1588 ElementTy = ArrayTy->getElementType();
1589 Indices.push_back(IRB.getIntN(PtrSize, 0));
1590 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1591 ElementTy = VectorTy->getElementType();
1592 Indices.push_back(IRB.getInt32(0));
1593 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1594 if (STy->element_begin() == STy->element_end())
1595 break; // Nothing left to descend into.
1596 ElementTy = *STy->element_begin();
1597 Indices.push_back(IRB.getInt32(0));
1602 } while (ElementTy != TargetTy);
1603 if (ElementTy != TargetTy)
1604 Indices.erase(Indices.end() - NumLayers, Indices.end());
1606 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1609 /// \brief Recursively compute indices for a natural GEP.
1611 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1612 /// element types adding appropriate indices for the GEP.
1613 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1614 Value *Ptr, Type *Ty, APInt &Offset,
1616 SmallVectorImpl<Value *> &Indices,
1619 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1622 // We can't recurse through pointer types.
1623 if (Ty->isPointerTy())
1626 // We try to analyze GEPs over vectors here, but note that these GEPs are
1627 // extremely poorly defined currently. The long-term goal is to remove GEPing
1628 // over a vector from the IR completely.
1629 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1630 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1631 if (ElementSizeInBits % 8 != 0) {
1632 // GEPs over non-multiple of 8 size vector elements are invalid.
1635 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1636 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1637 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1639 Offset -= NumSkippedElements * ElementSize;
1640 Indices.push_back(IRB.getInt(NumSkippedElements));
1641 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1642 Offset, TargetTy, Indices, NamePrefix);
1645 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1646 Type *ElementTy = ArrTy->getElementType();
1647 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1648 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1649 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1652 Offset -= NumSkippedElements * ElementSize;
1653 Indices.push_back(IRB.getInt(NumSkippedElements));
1654 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1655 Indices, NamePrefix);
1658 StructType *STy = dyn_cast<StructType>(Ty);
1662 const StructLayout *SL = DL.getStructLayout(STy);
1663 uint64_t StructOffset = Offset.getZExtValue();
1664 if (StructOffset >= SL->getSizeInBytes())
1666 unsigned Index = SL->getElementContainingOffset(StructOffset);
1667 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1668 Type *ElementTy = STy->getElementType(Index);
1669 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1670 return nullptr; // The offset points into alignment padding.
1672 Indices.push_back(IRB.getInt32(Index));
1673 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1674 Indices, NamePrefix);
1677 /// \brief Get a natural GEP from a base pointer to a particular offset and
1678 /// resulting in a particular type.
1680 /// The goal is to produce a "natural" looking GEP that works with the existing
1681 /// composite types to arrive at the appropriate offset and element type for
1682 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1683 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1684 /// Indices, and setting Ty to the result subtype.
1686 /// If no natural GEP can be constructed, this function returns null.
1687 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1688 Value *Ptr, APInt Offset, Type *TargetTy,
1689 SmallVectorImpl<Value *> &Indices,
1691 PointerType *Ty = cast<PointerType>(Ptr->getType());
1693 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1695 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1698 Type *ElementTy = Ty->getElementType();
1699 if (!ElementTy->isSized())
1700 return nullptr; // We can't GEP through an unsized element.
1701 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1702 if (ElementSize == 0)
1703 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1704 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1706 Offset -= NumSkippedElements * ElementSize;
1707 Indices.push_back(IRB.getInt(NumSkippedElements));
1708 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1709 Indices, NamePrefix);
1712 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1713 /// resulting pointer has PointerTy.
1715 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1716 /// and produces the pointer type desired. Where it cannot, it will try to use
1717 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1718 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1719 /// bitcast to the type.
1721 /// The strategy for finding the more natural GEPs is to peel off layers of the
1722 /// pointer, walking back through bit casts and GEPs, searching for a base
1723 /// pointer from which we can compute a natural GEP with the desired
1724 /// properties. The algorithm tries to fold as many constant indices into
1725 /// a single GEP as possible, thus making each GEP more independent of the
1726 /// surrounding code.
1727 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1728 APInt Offset, Type *PointerTy, Twine NamePrefix) {
1729 // Even though we don't look through PHI nodes, we could be called on an
1730 // instruction in an unreachable block, which may be on a cycle.
1731 SmallPtrSet<Value *, 4> Visited;
1732 Visited.insert(Ptr);
1733 SmallVector<Value *, 4> Indices;
1735 // We may end up computing an offset pointer that has the wrong type. If we
1736 // never are able to compute one directly that has the correct type, we'll
1737 // fall back to it, so keep it and the base it was computed from around here.
1738 Value *OffsetPtr = nullptr;
1739 Value *OffsetBasePtr;
1741 // Remember any i8 pointer we come across to re-use if we need to do a raw
1743 Value *Int8Ptr = nullptr;
1744 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1746 Type *TargetTy = PointerTy->getPointerElementType();
1749 // First fold any existing GEPs into the offset.
1750 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1751 APInt GEPOffset(Offset.getBitWidth(), 0);
1752 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1754 Offset += GEPOffset;
1755 Ptr = GEP->getPointerOperand();
1756 if (!Visited.insert(Ptr).second)
1760 // See if we can perform a natural GEP here.
1762 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1763 Indices, NamePrefix)) {
1764 // If we have a new natural pointer at the offset, clear out any old
1765 // offset pointer we computed. Unless it is the base pointer or
1766 // a non-instruction, we built a GEP we don't need. Zap it.
1767 if (OffsetPtr && OffsetPtr != OffsetBasePtr)
1768 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) {
1769 assert(I->use_empty() && "Built a GEP with uses some how!");
1770 I->eraseFromParent();
1773 OffsetBasePtr = Ptr;
1774 // If we also found a pointer of the right type, we're done.
1775 if (P->getType() == PointerTy)
1779 // Stash this pointer if we've found an i8*.
1780 if (Ptr->getType()->isIntegerTy(8)) {
1782 Int8PtrOffset = Offset;
1785 // Peel off a layer of the pointer and update the offset appropriately.
1786 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1787 Ptr = cast<Operator>(Ptr)->getOperand(0);
1788 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1789 if (GA->mayBeOverridden())
1791 Ptr = GA->getAliasee();
1795 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1796 } while (Visited.insert(Ptr).second);
1800 Int8Ptr = IRB.CreateBitCast(
1801 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1802 NamePrefix + "sroa_raw_cast");
1803 Int8PtrOffset = Offset;
1806 OffsetPtr = Int8PtrOffset == 0
1808 : IRB.CreateInBoundsGEP(IRB.getInt8Ty(), Int8Ptr,
1809 IRB.getInt(Int8PtrOffset),
1810 NamePrefix + "sroa_raw_idx");
1814 // On the off chance we were targeting i8*, guard the bitcast here.
1815 if (Ptr->getType() != PointerTy)
1816 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1821 /// \brief Compute the adjusted alignment for a load or store from an offset.
1822 static unsigned getAdjustedAlignment(Instruction *I, uint64_t Offset,
1823 const DataLayout &DL) {
1826 if (auto *LI = dyn_cast<LoadInst>(I)) {
1827 Alignment = LI->getAlignment();
1829 } else if (auto *SI = dyn_cast<StoreInst>(I)) {
1830 Alignment = SI->getAlignment();
1831 Ty = SI->getValueOperand()->getType();
1833 llvm_unreachable("Only loads and stores are allowed!");
1837 Alignment = DL.getABITypeAlignment(Ty);
1839 return MinAlign(Alignment, Offset);
1842 /// \brief Test whether we can convert a value from the old to the new type.
1844 /// This predicate should be used to guard calls to convertValue in order to
1845 /// ensure that we only try to convert viable values. The strategy is that we
1846 /// will peel off single element struct and array wrappings to get to an
1847 /// underlying value, and convert that value.
1848 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1852 // For integer types, we can't handle any bit-width differences. This would
1853 // break both vector conversions with extension and introduce endianness
1854 // issues when in conjunction with loads and stores.
1855 if (isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) {
1856 assert(cast<IntegerType>(OldTy)->getBitWidth() !=
1857 cast<IntegerType>(NewTy)->getBitWidth() &&
1858 "We can't have the same bitwidth for different int types");
1862 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1864 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1867 // We can convert pointers to integers and vice-versa. Same for vectors
1868 // of pointers and integers.
1869 OldTy = OldTy->getScalarType();
1870 NewTy = NewTy->getScalarType();
1871 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1872 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1874 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1882 /// \brief Generic routine to convert an SSA value to a value of a different
1885 /// This will try various different casting techniques, such as bitcasts,
1886 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1887 /// two types for viability with this routine.
1888 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1890 Type *OldTy = V->getType();
1891 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1896 assert(!(isa<IntegerType>(OldTy) && isa<IntegerType>(NewTy)) &&
1897 "Integer types must be the exact same to convert.");
1899 // See if we need inttoptr for this type pair. A cast involving both scalars
1900 // and vectors requires and additional bitcast.
1901 if (OldTy->getScalarType()->isIntegerTy() &&
1902 NewTy->getScalarType()->isPointerTy()) {
1903 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1904 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1905 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1908 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1909 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1910 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1913 return IRB.CreateIntToPtr(V, NewTy);
1916 // See if we need ptrtoint for this type pair. A cast involving both scalars
1917 // and vectors requires and additional bitcast.
1918 if (OldTy->getScalarType()->isPointerTy() &&
1919 NewTy->getScalarType()->isIntegerTy()) {
1920 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1921 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1922 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1925 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1926 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1927 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1930 return IRB.CreatePtrToInt(V, NewTy);
1933 return IRB.CreateBitCast(V, NewTy);
1936 /// \brief Test whether the given slice use can be promoted to a vector.
1938 /// This function is called to test each entry in a partition which is slated
1939 /// for a single slice.
1940 static bool isVectorPromotionViableForSlice(AllocaSlices::Partition &P,
1941 const Slice &S, VectorType *Ty,
1942 uint64_t ElementSize,
1943 const DataLayout &DL) {
1944 // First validate the slice offsets.
1945 uint64_t BeginOffset =
1946 std::max(S.beginOffset(), P.beginOffset()) - P.beginOffset();
1947 uint64_t BeginIndex = BeginOffset / ElementSize;
1948 if (BeginIndex * ElementSize != BeginOffset ||
1949 BeginIndex >= Ty->getNumElements())
1951 uint64_t EndOffset =
1952 std::min(S.endOffset(), P.endOffset()) - P.beginOffset();
1953 uint64_t EndIndex = EndOffset / ElementSize;
1954 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1957 assert(EndIndex > BeginIndex && "Empty vector!");
1958 uint64_t NumElements = EndIndex - BeginIndex;
1959 Type *SliceTy = (NumElements == 1)
1960 ? Ty->getElementType()
1961 : VectorType::get(Ty->getElementType(), NumElements);
1964 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1966 Use *U = S.getUse();
1968 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1969 if (MI->isVolatile())
1971 if (!S.isSplittable())
1972 return false; // Skip any unsplittable intrinsics.
1973 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1974 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1975 II->getIntrinsicID() != Intrinsic::lifetime_end)
1977 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1978 // Disable vector promotion when there are loads or stores of an FCA.
1980 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1981 if (LI->isVolatile())
1983 Type *LTy = LI->getType();
1984 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1985 assert(LTy->isIntegerTy());
1988 if (!canConvertValue(DL, SliceTy, LTy))
1990 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1991 if (SI->isVolatile())
1993 Type *STy = SI->getValueOperand()->getType();
1994 if (P.beginOffset() > S.beginOffset() || P.endOffset() < S.endOffset()) {
1995 assert(STy->isIntegerTy());
1998 if (!canConvertValue(DL, STy, SliceTy))
2007 /// \brief Test whether the given alloca partitioning and range of slices can be
2008 /// promoted to a vector.
2010 /// This is a quick test to check whether we can rewrite a particular alloca
2011 /// partition (and its newly formed alloca) into a vector alloca with only
2012 /// whole-vector loads and stores such that it could be promoted to a vector
2013 /// SSA value. We only can ensure this for a limited set of operations, and we
2014 /// don't want to do the rewrites unless we are confident that the result will
2015 /// be promotable, so we have an early test here.
2016 static VectorType *isVectorPromotionViable(AllocaSlices::Partition &P,
2017 const DataLayout &DL) {
2018 // Collect the candidate types for vector-based promotion. Also track whether
2019 // we have different element types.
2020 SmallVector<VectorType *, 4> CandidateTys;
2021 Type *CommonEltTy = nullptr;
2022 bool HaveCommonEltTy = true;
2023 auto CheckCandidateType = [&](Type *Ty) {
2024 if (auto *VTy = dyn_cast<VectorType>(Ty)) {
2025 CandidateTys.push_back(VTy);
2027 CommonEltTy = VTy->getElementType();
2028 else if (CommonEltTy != VTy->getElementType())
2029 HaveCommonEltTy = false;
2032 // Consider any loads or stores that are the exact size of the slice.
2033 for (const Slice &S : P)
2034 if (S.beginOffset() == P.beginOffset() &&
2035 S.endOffset() == P.endOffset()) {
2036 if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
2037 CheckCandidateType(LI->getType());
2038 else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
2039 CheckCandidateType(SI->getValueOperand()->getType());
2042 // If we didn't find a vector type, nothing to do here.
2043 if (CandidateTys.empty())
2046 // Remove non-integer vector types if we had multiple common element types.
2047 // FIXME: It'd be nice to replace them with integer vector types, but we can't
2048 // do that until all the backends are known to produce good code for all
2049 // integer vector types.
2050 if (!HaveCommonEltTy) {
2051 CandidateTys.erase(std::remove_if(CandidateTys.begin(), CandidateTys.end(),
2052 [](VectorType *VTy) {
2053 return !VTy->getElementType()->isIntegerTy();
2055 CandidateTys.end());
2057 // If there were no integer vector types, give up.
2058 if (CandidateTys.empty())
2061 // Rank the remaining candidate vector types. This is easy because we know
2062 // they're all integer vectors. We sort by ascending number of elements.
2063 auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
2064 assert(DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
2065 "Cannot have vector types of different sizes!");
2066 assert(RHSTy->getElementType()->isIntegerTy() &&
2067 "All non-integer types eliminated!");
2068 assert(LHSTy->getElementType()->isIntegerTy() &&
2069 "All non-integer types eliminated!");
2070 return RHSTy->getNumElements() < LHSTy->getNumElements();
2072 std::sort(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes);
2074 std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
2075 CandidateTys.end());
2077 // The only way to have the same element type in every vector type is to
2078 // have the same vector type. Check that and remove all but one.
2080 for (VectorType *VTy : CandidateTys) {
2081 assert(VTy->getElementType() == CommonEltTy &&
2082 "Unaccounted for element type!");
2083 assert(VTy == CandidateTys[0] &&
2084 "Different vector types with the same element type!");
2087 CandidateTys.resize(1);
2090 // Try each vector type, and return the one which works.
2091 auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
2092 uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
2094 // While the definition of LLVM vectors is bitpacked, we don't support sizes
2095 // that aren't byte sized.
2096 if (ElementSize % 8)
2098 assert((DL.getTypeSizeInBits(VTy) % 8) == 0 &&
2099 "vector size not a multiple of element size?");
2102 for (const Slice &S : P)
2103 if (!isVectorPromotionViableForSlice(P, S, VTy, ElementSize, DL))
2106 for (const Slice *S : P.splitSliceTails())
2107 if (!isVectorPromotionViableForSlice(P, *S, VTy, ElementSize, DL))
2112 for (VectorType *VTy : CandidateTys)
2113 if (CheckVectorTypeForPromotion(VTy))
2119 /// \brief Test whether a slice of an alloca is valid for integer widening.
2121 /// This implements the necessary checking for the \c isIntegerWideningViable
2122 /// test below on a single slice of the alloca.
2123 static bool isIntegerWideningViableForSlice(const Slice &S,
2124 uint64_t AllocBeginOffset,
2126 const DataLayout &DL,
2127 bool &WholeAllocaOp) {
2128 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
2130 uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
2131 uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
2133 // We can't reasonably handle cases where the load or store extends past
2134 // the end of the alloca's type and into its padding.
2138 Use *U = S.getUse();
2140 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
2141 if (LI->isVolatile())
2143 // We can't handle loads that extend past the allocated memory.
2144 if (DL.getTypeStoreSize(LI->getType()) > Size)
2146 // Note that we don't count vector loads or stores as whole-alloca
2147 // operations which enable integer widening because we would prefer to use
2148 // vector widening instead.
2149 if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
2150 WholeAllocaOp = true;
2151 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
2152 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
2154 } else if (RelBegin != 0 || RelEnd != Size ||
2155 !canConvertValue(DL, AllocaTy, LI->getType())) {
2156 // Non-integer loads need to be convertible from the alloca type so that
2157 // they are promotable.
2160 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
2161 Type *ValueTy = SI->getValueOperand()->getType();
2162 if (SI->isVolatile())
2164 // We can't handle stores that extend past the allocated memory.
2165 if (DL.getTypeStoreSize(ValueTy) > Size)
2167 // Note that we don't count vector loads or stores as whole-alloca
2168 // operations which enable integer widening because we would prefer to use
2169 // vector widening instead.
2170 if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
2171 WholeAllocaOp = true;
2172 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2173 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
2175 } else if (RelBegin != 0 || RelEnd != Size ||
2176 !canConvertValue(DL, ValueTy, AllocaTy)) {
2177 // Non-integer stores need to be convertible to the alloca type so that
2178 // they are promotable.
2181 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
2182 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
2184 if (!S.isSplittable())
2185 return false; // Skip any unsplittable intrinsics.
2186 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
2187 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2188 II->getIntrinsicID() != Intrinsic::lifetime_end)
2197 /// \brief Test whether the given alloca partition's integer operations can be
2198 /// widened to promotable ones.
2200 /// This is a quick test to check whether we can rewrite the integer loads and
2201 /// stores to a particular alloca into wider loads and stores and be able to
2202 /// promote the resulting alloca.
2203 static bool isIntegerWideningViable(AllocaSlices::Partition &P, Type *AllocaTy,
2204 const DataLayout &DL) {
2205 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
2206 // Don't create integer types larger than the maximum bitwidth.
2207 if (SizeInBits > IntegerType::MAX_INT_BITS)
2210 // Don't try to handle allocas with bit-padding.
2211 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
2214 // We need to ensure that an integer type with the appropriate bitwidth can
2215 // be converted to the alloca type, whatever that is. We don't want to force
2216 // the alloca itself to have an integer type if there is a more suitable one.
2217 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2218 if (!canConvertValue(DL, AllocaTy, IntTy) ||
2219 !canConvertValue(DL, IntTy, AllocaTy))
2222 // While examining uses, we ensure that the alloca has a covering load or
2223 // store. We don't want to widen the integer operations only to fail to
2224 // promote due to some other unsplittable entry (which we may make splittable
2225 // later). However, if there are only splittable uses, go ahead and assume
2226 // that we cover the alloca.
2227 // FIXME: We shouldn't consider split slices that happen to start in the
2228 // partition here...
2229 bool WholeAllocaOp =
2230 P.begin() != P.end() ? false : DL.isLegalInteger(SizeInBits);
2232 for (const Slice &S : P)
2233 if (!isIntegerWideningViableForSlice(S, P.beginOffset(), AllocaTy, DL,
2237 for (const Slice *S : P.splitSliceTails())
2238 if (!isIntegerWideningViableForSlice(*S, P.beginOffset(), AllocaTy, DL,
2242 return WholeAllocaOp;
2245 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2246 IntegerType *Ty, uint64_t Offset,
2247 const Twine &Name) {
2248 DEBUG(dbgs() << " start: " << *V << "\n");
2249 IntegerType *IntTy = cast<IntegerType>(V->getType());
2250 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2251 "Element extends past full value");
2252 uint64_t ShAmt = 8 * Offset;
2253 if (DL.isBigEndian())
2254 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2256 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2257 DEBUG(dbgs() << " shifted: " << *V << "\n");
2259 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2260 "Cannot extract to a larger integer!");
2262 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2263 DEBUG(dbgs() << " trunced: " << *V << "\n");
2268 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2269 Value *V, uint64_t Offset, const Twine &Name) {
2270 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2271 IntegerType *Ty = cast<IntegerType>(V->getType());
2272 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2273 "Cannot insert a larger integer!");
2274 DEBUG(dbgs() << " start: " << *V << "\n");
2276 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2277 DEBUG(dbgs() << " extended: " << *V << "\n");
2279 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2280 "Element store outside of alloca store");
2281 uint64_t ShAmt = 8 * Offset;
2282 if (DL.isBigEndian())
2283 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2285 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2286 DEBUG(dbgs() << " shifted: " << *V << "\n");
2289 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2290 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2291 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2292 DEBUG(dbgs() << " masked: " << *Old << "\n");
2293 V = IRB.CreateOr(Old, V, Name + ".insert");
2294 DEBUG(dbgs() << " inserted: " << *V << "\n");
2299 static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2300 unsigned EndIndex, const Twine &Name) {
2301 VectorType *VecTy = cast<VectorType>(V->getType());
2302 unsigned NumElements = EndIndex - BeginIndex;
2303 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2305 if (NumElements == VecTy->getNumElements())
2308 if (NumElements == 1) {
2309 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2311 DEBUG(dbgs() << " extract: " << *V << "\n");
2315 SmallVector<Constant *, 8> Mask;
2316 Mask.reserve(NumElements);
2317 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2318 Mask.push_back(IRB.getInt32(i));
2319 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2320 ConstantVector::get(Mask), Name + ".extract");
2321 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2325 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2326 unsigned BeginIndex, const Twine &Name) {
2327 VectorType *VecTy = cast<VectorType>(Old->getType());
2328 assert(VecTy && "Can only insert a vector into a vector");
2330 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2332 // Single element to insert.
2333 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2335 DEBUG(dbgs() << " insert: " << *V << "\n");
2339 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2340 "Too many elements!");
2341 if (Ty->getNumElements() == VecTy->getNumElements()) {
2342 assert(V->getType() == VecTy && "Vector type mismatch");
2345 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2347 // When inserting a smaller vector into the larger to store, we first
2348 // use a shuffle vector to widen it with undef elements, and then
2349 // a second shuffle vector to select between the loaded vector and the
2351 SmallVector<Constant *, 8> Mask;
2352 Mask.reserve(VecTy->getNumElements());
2353 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2354 if (i >= BeginIndex && i < EndIndex)
2355 Mask.push_back(IRB.getInt32(i - BeginIndex));
2357 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2358 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2359 ConstantVector::get(Mask), Name + ".expand");
2360 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2363 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2364 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2366 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2368 DEBUG(dbgs() << " blend: " << *V << "\n");
2373 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
2374 /// to use a new alloca.
2376 /// Also implements the rewriting to vector-based accesses when the partition
2377 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2379 class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
2380 // Befriend the base class so it can delegate to private visit methods.
2381 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
2382 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
2384 const DataLayout &DL;
2387 AllocaInst &OldAI, &NewAI;
2388 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2391 // This is a convenience and flag variable that will be null unless the new
2392 // alloca's integer operations should be widened to this integer type due to
2393 // passing isIntegerWideningViable above. If it is non-null, the desired
2394 // integer type will be stored here for easy access during rewriting.
2397 // If we are rewriting an alloca partition which can be written as pure
2398 // vector operations, we stash extra information here. When VecTy is
2399 // non-null, we have some strict guarantees about the rewritten alloca:
2400 // - The new alloca is exactly the size of the vector type here.
2401 // - The accesses all either map to the entire vector or to a single
2403 // - The set of accessing instructions is only one of those handled above
2404 // in isVectorPromotionViable. Generally these are the same access kinds
2405 // which are promotable via mem2reg.
2408 uint64_t ElementSize;
2410 // The original offset of the slice currently being rewritten relative to
2411 // the original alloca.
2412 uint64_t BeginOffset, EndOffset;
2413 // The new offsets of the slice currently being rewritten relative to the
2415 uint64_t NewBeginOffset, NewEndOffset;
2421 Instruction *OldPtr;
2423 // Track post-rewrite users which are PHI nodes and Selects.
2424 SmallPtrSetImpl<PHINode *> &PHIUsers;
2425 SmallPtrSetImpl<SelectInst *> &SelectUsers;
2427 // Utility IR builder, whose name prefix is setup for each visited use, and
2428 // the insertion point is set to point to the user.
2432 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROA &Pass,
2433 AllocaInst &OldAI, AllocaInst &NewAI,
2434 uint64_t NewAllocaBeginOffset,
2435 uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2436 VectorType *PromotableVecTy,
2437 SmallPtrSetImpl<PHINode *> &PHIUsers,
2438 SmallPtrSetImpl<SelectInst *> &SelectUsers)
2439 : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2440 NewAllocaBeginOffset(NewAllocaBeginOffset),
2441 NewAllocaEndOffset(NewAllocaEndOffset),
2442 NewAllocaTy(NewAI.getAllocatedType()),
2443 IntTy(IsIntegerPromotable
2446 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2448 VecTy(PromotableVecTy),
2449 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2450 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2451 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2452 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2453 IRB(NewAI.getContext(), ConstantFolder()) {
2455 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2456 "Only multiple-of-8 sized vector elements are viable");
2459 assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy));
2462 bool visit(AllocaSlices::const_iterator I) {
2463 bool CanSROA = true;
2464 BeginOffset = I->beginOffset();
2465 EndOffset = I->endOffset();
2466 IsSplittable = I->isSplittable();
2468 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2469 DEBUG(dbgs() << " rewriting " << (IsSplit ? "split " : ""));
2470 DEBUG(AS.printSlice(dbgs(), I, ""));
2471 DEBUG(dbgs() << "\n");
2473 // Compute the intersecting offset range.
2474 assert(BeginOffset < NewAllocaEndOffset);
2475 assert(EndOffset > NewAllocaBeginOffset);
2476 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2477 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2479 SliceSize = NewEndOffset - NewBeginOffset;
2481 OldUse = I->getUse();
2482 OldPtr = cast<Instruction>(OldUse->get());
2484 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2485 IRB.SetInsertPoint(OldUserI);
2486 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2487 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2489 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2496 // Make sure the other visit overloads are visible.
2499 // Every instruction which can end up as a user must have a rewrite rule.
2500 bool visitInstruction(Instruction &I) {
2501 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2502 llvm_unreachable("No rewrite rule for this instruction!");
2505 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2506 // Note that the offset computation can use BeginOffset or NewBeginOffset
2507 // interchangeably for unsplit slices.
2508 assert(IsSplit || BeginOffset == NewBeginOffset);
2509 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2512 StringRef OldName = OldPtr->getName();
2513 // Skip through the last '.sroa.' component of the name.
2514 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2515 if (LastSROAPrefix != StringRef::npos) {
2516 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2517 // Look for an SROA slice index.
2518 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2519 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2520 // Strip the index and look for the offset.
2521 OldName = OldName.substr(IndexEnd + 1);
2522 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2523 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2524 // Strip the offset.
2525 OldName = OldName.substr(OffsetEnd + 1);
2528 // Strip any SROA suffixes as well.
2529 OldName = OldName.substr(0, OldName.find(".sroa_"));
2532 return getAdjustedPtr(IRB, DL, &NewAI,
2533 APInt(DL.getPointerSizeInBits(), Offset), PointerTy,
2535 Twine(OldName) + "."
2542 /// \brief Compute suitable alignment to access this slice of the *new*
2545 /// You can optionally pass a type to this routine and if that type's ABI
2546 /// alignment is itself suitable, this will return zero.
2547 unsigned getSliceAlign(Type *Ty = nullptr) {
2548 unsigned NewAIAlign = NewAI.getAlignment();
2550 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2552 MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2553 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2556 unsigned getIndex(uint64_t Offset) {
2557 assert(VecTy && "Can only call getIndex when rewriting a vector");
2558 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2559 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2560 uint32_t Index = RelOffset / ElementSize;
2561 assert(Index * ElementSize == RelOffset);
2565 void deleteIfTriviallyDead(Value *V) {
2566 Instruction *I = cast<Instruction>(V);
2567 if (isInstructionTriviallyDead(I))
2568 Pass.DeadInsts.insert(I);
2571 Value *rewriteVectorizedLoadInst() {
2572 unsigned BeginIndex = getIndex(NewBeginOffset);
2573 unsigned EndIndex = getIndex(NewEndOffset);
2574 assert(EndIndex > BeginIndex && "Empty vector!");
2576 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2577 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2580 Value *rewriteIntegerLoad(LoadInst &LI) {
2581 assert(IntTy && "We cannot insert an integer to the alloca");
2582 assert(!LI.isVolatile());
2583 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2584 V = convertValue(DL, IRB, V, IntTy);
2585 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2586 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2587 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2588 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2593 bool visitLoadInst(LoadInst &LI) {
2594 DEBUG(dbgs() << " original: " << LI << "\n");
2595 Value *OldOp = LI.getOperand(0);
2596 assert(OldOp == OldPtr);
2598 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2600 const bool IsLoadPastEnd = DL.getTypeStoreSize(TargetTy) > SliceSize;
2601 bool IsPtrAdjusted = false;
2604 V = rewriteVectorizedLoadInst();
2605 } else if (IntTy && LI.getType()->isIntegerTy()) {
2606 V = rewriteIntegerLoad(LI);
2607 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2608 NewEndOffset == NewAllocaEndOffset &&
2609 (canConvertValue(DL, NewAllocaTy, TargetTy) ||
2610 (IsLoadPastEnd && NewAllocaTy->isIntegerTy() &&
2611 TargetTy->isIntegerTy()))) {
2612 LoadInst *NewLI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2613 LI.isVolatile(), LI.getName());
2614 if (LI.isVolatile())
2615 NewLI->setAtomic(LI.getOrdering(), LI.getSynchScope());
2618 // If this is an integer load past the end of the slice (which means the
2619 // bytes outside the slice are undef or this load is dead) just forcibly
2620 // fix the integer size with correct handling of endianness.
2621 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2622 if (auto *TITy = dyn_cast<IntegerType>(TargetTy))
2623 if (AITy->getBitWidth() < TITy->getBitWidth()) {
2624 V = IRB.CreateZExt(V, TITy, "load.ext");
2625 if (DL.isBigEndian())
2626 V = IRB.CreateShl(V, TITy->getBitWidth() - AITy->getBitWidth(),
2630 Type *LTy = TargetTy->getPointerTo();
2631 LoadInst *NewLI = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2632 getSliceAlign(TargetTy),
2633 LI.isVolatile(), LI.getName());
2634 if (LI.isVolatile())
2635 NewLI->setAtomic(LI.getOrdering(), LI.getSynchScope());
2638 IsPtrAdjusted = true;
2640 V = convertValue(DL, IRB, V, TargetTy);
2643 assert(!LI.isVolatile());
2644 assert(LI.getType()->isIntegerTy() &&
2645 "Only integer type loads and stores are split");
2646 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2647 "Split load isn't smaller than original load");
2648 assert(LI.getType()->getIntegerBitWidth() ==
2649 DL.getTypeStoreSizeInBits(LI.getType()) &&
2650 "Non-byte-multiple bit width");
2651 // Move the insertion point just past the load so that we can refer to it.
2652 IRB.SetInsertPoint(std::next(BasicBlock::iterator(&LI)));
2653 // Create a placeholder value with the same type as LI to use as the
2654 // basis for the new value. This allows us to replace the uses of LI with
2655 // the computed value, and then replace the placeholder with LI, leaving
2656 // LI only used for this computation.
2657 Value *Placeholder =
2658 new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2659 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset - BeginOffset,
2661 LI.replaceAllUsesWith(V);
2662 Placeholder->replaceAllUsesWith(&LI);
2665 LI.replaceAllUsesWith(V);
2668 Pass.DeadInsts.insert(&LI);
2669 deleteIfTriviallyDead(OldOp);
2670 DEBUG(dbgs() << " to: " << *V << "\n");
2671 return !LI.isVolatile() && !IsPtrAdjusted;
2674 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2675 if (V->getType() != VecTy) {
2676 unsigned BeginIndex = getIndex(NewBeginOffset);
2677 unsigned EndIndex = getIndex(NewEndOffset);
2678 assert(EndIndex > BeginIndex && "Empty vector!");
2679 unsigned NumElements = EndIndex - BeginIndex;
2680 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2681 Type *SliceTy = (NumElements == 1)
2683 : VectorType::get(ElementTy, NumElements);
2684 if (V->getType() != SliceTy)
2685 V = convertValue(DL, IRB, V, SliceTy);
2687 // Mix in the existing elements.
2688 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2689 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2691 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2692 Pass.DeadInsts.insert(&SI);
2695 DEBUG(dbgs() << " to: " << *Store << "\n");
2699 bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2700 assert(IntTy && "We cannot extract an integer from the alloca");
2701 assert(!SI.isVolatile());
2702 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2704 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2705 Old = convertValue(DL, IRB, Old, IntTy);
2706 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2707 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2708 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2710 V = convertValue(DL, IRB, V, NewAllocaTy);
2711 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2712 Pass.DeadInsts.insert(&SI);
2714 DEBUG(dbgs() << " to: " << *Store << "\n");
2718 bool visitStoreInst(StoreInst &SI) {
2719 DEBUG(dbgs() << " original: " << SI << "\n");
2720 Value *OldOp = SI.getOperand(1);
2721 assert(OldOp == OldPtr);
2723 Value *V = SI.getValueOperand();
2725 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2726 // alloca that should be re-examined after promoting this alloca.
2727 if (V->getType()->isPointerTy())
2728 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2729 Pass.PostPromotionWorklist.insert(AI);
2731 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2732 assert(!SI.isVolatile());
2733 assert(V->getType()->isIntegerTy() &&
2734 "Only integer type loads and stores are split");
2735 assert(V->getType()->getIntegerBitWidth() ==
2736 DL.getTypeStoreSizeInBits(V->getType()) &&
2737 "Non-byte-multiple bit width");
2738 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2739 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset - BeginOffset,
2744 return rewriteVectorizedStoreInst(V, SI, OldOp);
2745 if (IntTy && V->getType()->isIntegerTy())
2746 return rewriteIntegerStore(V, SI);
2748 const bool IsStorePastEnd = DL.getTypeStoreSize(V->getType()) > SliceSize;
2750 if (NewBeginOffset == NewAllocaBeginOffset &&
2751 NewEndOffset == NewAllocaEndOffset &&
2752 (canConvertValue(DL, V->getType(), NewAllocaTy) ||
2753 (IsStorePastEnd && NewAllocaTy->isIntegerTy() &&
2754 V->getType()->isIntegerTy()))) {
2755 // If this is an integer store past the end of slice (and thus the bytes
2756 // past that point are irrelevant or this is unreachable), truncate the
2757 // value prior to storing.
2758 if (auto *VITy = dyn_cast<IntegerType>(V->getType()))
2759 if (auto *AITy = dyn_cast<IntegerType>(NewAllocaTy))
2760 if (VITy->getBitWidth() > AITy->getBitWidth()) {
2761 if (DL.isBigEndian())
2762 V = IRB.CreateLShr(V, VITy->getBitWidth() - AITy->getBitWidth(),
2764 V = IRB.CreateTrunc(V, AITy, "load.trunc");
2767 V = convertValue(DL, IRB, V, NewAllocaTy);
2768 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2771 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo());
2772 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2775 if (SI.isVolatile())
2776 NewSI->setAtomic(SI.getOrdering(), SI.getSynchScope());
2777 Pass.DeadInsts.insert(&SI);
2778 deleteIfTriviallyDead(OldOp);
2780 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2781 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2784 /// \brief Compute an integer value from splatting an i8 across the given
2785 /// number of bytes.
2787 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2788 /// call this routine.
2789 /// FIXME: Heed the advice above.
2791 /// \param V The i8 value to splat.
2792 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2793 Value *getIntegerSplat(Value *V, unsigned Size) {
2794 assert(Size > 0 && "Expected a positive number of bytes.");
2795 IntegerType *VTy = cast<IntegerType>(V->getType());
2796 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2800 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2802 IRB.CreateZExt(V, SplatIntTy, "zext"),
2803 ConstantExpr::getUDiv(
2804 Constant::getAllOnesValue(SplatIntTy),
2805 ConstantExpr::getZExt(Constant::getAllOnesValue(V->getType()),
2811 /// \brief Compute a vector splat for a given element value.
2812 Value *getVectorSplat(Value *V, unsigned NumElements) {
2813 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2814 DEBUG(dbgs() << " splat: " << *V << "\n");
2818 bool visitMemSetInst(MemSetInst &II) {
2819 DEBUG(dbgs() << " original: " << II << "\n");
2820 assert(II.getRawDest() == OldPtr);
2822 // If the memset has a variable size, it cannot be split, just adjust the
2823 // pointer to the new alloca.
2824 if (!isa<Constant>(II.getLength())) {
2826 assert(NewBeginOffset == BeginOffset);
2827 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2828 Type *CstTy = II.getAlignmentCst()->getType();
2829 II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2831 deleteIfTriviallyDead(OldPtr);
2835 // Record this instruction for deletion.
2836 Pass.DeadInsts.insert(&II);
2838 Type *AllocaTy = NewAI.getAllocatedType();
2839 Type *ScalarTy = AllocaTy->getScalarType();
2841 // If this doesn't map cleanly onto the alloca type, and that type isn't
2842 // a single value type, just emit a memset.
2843 if (!VecTy && !IntTy &&
2844 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2845 SliceSize != DL.getTypeStoreSize(AllocaTy) ||
2846 !AllocaTy->isSingleValueType() ||
2847 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2848 DL.getTypeSizeInBits(ScalarTy) % 8 != 0)) {
2849 Type *SizeTy = II.getLength()->getType();
2850 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2851 CallInst *New = IRB.CreateMemSet(
2852 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2853 getSliceAlign(), II.isVolatile());
2855 DEBUG(dbgs() << " to: " << *New << "\n");
2859 // If we can represent this as a simple value, we have to build the actual
2860 // value to store, which requires expanding the byte present in memset to
2861 // a sensible representation for the alloca type. This is essentially
2862 // splatting the byte to a sufficiently wide integer, splatting it across
2863 // any desired vector width, and bitcasting to the final type.
2867 // If this is a memset of a vectorized alloca, insert it.
2868 assert(ElementTy == ScalarTy);
2870 unsigned BeginIndex = getIndex(NewBeginOffset);
2871 unsigned EndIndex = getIndex(NewEndOffset);
2872 assert(EndIndex > BeginIndex && "Empty vector!");
2873 unsigned NumElements = EndIndex - BeginIndex;
2874 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2877 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2878 Splat = convertValue(DL, IRB, Splat, ElementTy);
2879 if (NumElements > 1)
2880 Splat = getVectorSplat(Splat, NumElements);
2883 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2884 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2886 // If this is a memset on an alloca where we can widen stores, insert the
2888 assert(!II.isVolatile());
2890 uint64_t Size = NewEndOffset - NewBeginOffset;
2891 V = getIntegerSplat(II.getValue(), Size);
2893 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2894 EndOffset != NewAllocaBeginOffset)) {
2896 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2897 Old = convertValue(DL, IRB, Old, IntTy);
2898 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2899 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2901 assert(V->getType() == IntTy &&
2902 "Wrong type for an alloca wide integer!");
2904 V = convertValue(DL, IRB, V, AllocaTy);
2906 // Established these invariants above.
2907 assert(NewBeginOffset == NewAllocaBeginOffset);
2908 assert(NewEndOffset == NewAllocaEndOffset);
2910 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2911 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2912 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2914 V = convertValue(DL, IRB, V, AllocaTy);
2917 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2920 DEBUG(dbgs() << " to: " << *New << "\n");
2921 return !II.isVolatile();
2924 bool visitMemTransferInst(MemTransferInst &II) {
2925 // Rewriting of memory transfer instructions can be a bit tricky. We break
2926 // them into two categories: split intrinsics and unsplit intrinsics.
2928 DEBUG(dbgs() << " original: " << II << "\n");
2930 bool IsDest = &II.getRawDestUse() == OldUse;
2931 assert((IsDest && II.getRawDest() == OldPtr) ||
2932 (!IsDest && II.getRawSource() == OldPtr));
2934 unsigned SliceAlign = getSliceAlign();
2936 // For unsplit intrinsics, we simply modify the source and destination
2937 // pointers in place. This isn't just an optimization, it is a matter of
2938 // correctness. With unsplit intrinsics we may be dealing with transfers
2939 // within a single alloca before SROA ran, or with transfers that have
2940 // a variable length. We may also be dealing with memmove instead of
2941 // memcpy, and so simply updating the pointers is the necessary for us to
2942 // update both source and dest of a single call.
2943 if (!IsSplittable) {
2944 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2946 II.setDest(AdjustedPtr);
2948 II.setSource(AdjustedPtr);
2950 if (II.getAlignment() > SliceAlign) {
2951 Type *CstTy = II.getAlignmentCst()->getType();
2953 ConstantInt::get(CstTy, MinAlign(II.getAlignment(), SliceAlign)));
2956 DEBUG(dbgs() << " to: " << II << "\n");
2957 deleteIfTriviallyDead(OldPtr);
2960 // For split transfer intrinsics we have an incredibly useful assurance:
2961 // the source and destination do not reside within the same alloca, and at
2962 // least one of them does not escape. This means that we can replace
2963 // memmove with memcpy, and we don't need to worry about all manner of
2964 // downsides to splitting and transforming the operations.
2966 // If this doesn't map cleanly onto the alloca type, and that type isn't
2967 // a single value type, just emit a memcpy.
2970 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2971 SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2972 !NewAI.getAllocatedType()->isSingleValueType());
2974 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2975 // size hasn't been shrunk based on analysis of the viable range, this is
2977 if (EmitMemCpy && &OldAI == &NewAI) {
2978 // Ensure the start lines up.
2979 assert(NewBeginOffset == BeginOffset);
2981 // Rewrite the size as needed.
2982 if (NewEndOffset != EndOffset)
2983 II.setLength(ConstantInt::get(II.getLength()->getType(),
2984 NewEndOffset - NewBeginOffset));
2987 // Record this instruction for deletion.
2988 Pass.DeadInsts.insert(&II);
2990 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2991 // alloca that should be re-examined after rewriting this instruction.
2992 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2993 if (AllocaInst *AI =
2994 dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2995 assert(AI != &OldAI && AI != &NewAI &&
2996 "Splittable transfers cannot reach the same alloca on both ends.");
2997 Pass.Worklist.insert(AI);
3000 Type *OtherPtrTy = OtherPtr->getType();
3001 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
3003 // Compute the relative offset for the other pointer within the transfer.
3004 unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
3005 APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
3006 unsigned OtherAlign = MinAlign(II.getAlignment() ? II.getAlignment() : 1,
3007 OtherOffset.zextOrTrunc(64).getZExtValue());
3010 // Compute the other pointer, folding as much as possible to produce
3011 // a single, simple GEP in most cases.
3012 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
3013 OtherPtr->getName() + ".");
3015 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3016 Type *SizeTy = II.getLength()->getType();
3017 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
3019 CallInst *New = IRB.CreateMemCpy(
3020 IsDest ? OurPtr : OtherPtr, IsDest ? OtherPtr : OurPtr, Size,
3021 MinAlign(SliceAlign, OtherAlign), II.isVolatile());
3023 DEBUG(dbgs() << " to: " << *New << "\n");
3027 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
3028 NewEndOffset == NewAllocaEndOffset;
3029 uint64_t Size = NewEndOffset - NewBeginOffset;
3030 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
3031 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
3032 unsigned NumElements = EndIndex - BeginIndex;
3033 IntegerType *SubIntTy =
3034 IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
3036 // Reset the other pointer type to match the register type we're going to
3037 // use, but using the address space of the original other pointer.
3038 if (VecTy && !IsWholeAlloca) {
3039 if (NumElements == 1)
3040 OtherPtrTy = VecTy->getElementType();
3042 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
3044 OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
3045 } else if (IntTy && !IsWholeAlloca) {
3046 OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
3048 OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
3051 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
3052 OtherPtr->getName() + ".");
3053 unsigned SrcAlign = OtherAlign;
3054 Value *DstPtr = &NewAI;
3055 unsigned DstAlign = SliceAlign;
3057 std::swap(SrcPtr, DstPtr);
3058 std::swap(SrcAlign, DstAlign);
3062 if (VecTy && !IsWholeAlloca && !IsDest) {
3063 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
3064 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
3065 } else if (IntTy && !IsWholeAlloca && !IsDest) {
3066 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
3067 Src = convertValue(DL, IRB, Src, IntTy);
3068 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
3069 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
3072 IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(), "copyload");
3075 if (VecTy && !IsWholeAlloca && IsDest) {
3077 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
3078 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
3079 } else if (IntTy && !IsWholeAlloca && IsDest) {
3081 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
3082 Old = convertValue(DL, IRB, Old, IntTy);
3083 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
3084 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
3085 Src = convertValue(DL, IRB, Src, NewAllocaTy);
3088 StoreInst *Store = cast<StoreInst>(
3089 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
3091 DEBUG(dbgs() << " to: " << *Store << "\n");
3092 return !II.isVolatile();
3095 bool visitIntrinsicInst(IntrinsicInst &II) {
3096 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
3097 II.getIntrinsicID() == Intrinsic::lifetime_end);
3098 DEBUG(dbgs() << " original: " << II << "\n");
3099 assert(II.getArgOperand(1) == OldPtr);
3101 // Record this instruction for deletion.
3102 Pass.DeadInsts.insert(&II);
3105 ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
3106 NewEndOffset - NewBeginOffset);
3107 Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3109 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
3110 New = IRB.CreateLifetimeStart(Ptr, Size);
3112 New = IRB.CreateLifetimeEnd(Ptr, Size);
3115 DEBUG(dbgs() << " to: " << *New << "\n");
3119 bool visitPHINode(PHINode &PN) {
3120 DEBUG(dbgs() << " original: " << PN << "\n");
3121 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
3122 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
3124 // We would like to compute a new pointer in only one place, but have it be
3125 // as local as possible to the PHI. To do that, we re-use the location of
3126 // the old pointer, which necessarily must be in the right position to
3127 // dominate the PHI.
3128 IRBuilderTy PtrBuilder(IRB);
3129 if (isa<PHINode>(OldPtr))
3130 PtrBuilder.SetInsertPoint(OldPtr->getParent()->getFirstInsertionPt());
3132 PtrBuilder.SetInsertPoint(OldPtr);
3133 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
3135 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
3136 // Replace the operands which were using the old pointer.
3137 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
3139 DEBUG(dbgs() << " to: " << PN << "\n");
3140 deleteIfTriviallyDead(OldPtr);
3142 // PHIs can't be promoted on their own, but often can be speculated. We
3143 // check the speculation outside of the rewriter so that we see the
3144 // fully-rewritten alloca.
3145 PHIUsers.insert(&PN);
3149 bool visitSelectInst(SelectInst &SI) {
3150 DEBUG(dbgs() << " original: " << SI << "\n");
3151 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
3152 "Pointer isn't an operand!");
3153 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
3154 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
3156 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3157 // Replace the operands which were using the old pointer.
3158 if (SI.getOperand(1) == OldPtr)
3159 SI.setOperand(1, NewPtr);
3160 if (SI.getOperand(2) == OldPtr)
3161 SI.setOperand(2, NewPtr);
3163 DEBUG(dbgs() << " to: " << SI << "\n");
3164 deleteIfTriviallyDead(OldPtr);
3166 // Selects can't be promoted on their own, but often can be speculated. We
3167 // check the speculation outside of the rewriter so that we see the
3168 // fully-rewritten alloca.
3169 SelectUsers.insert(&SI);
3176 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
3178 /// This pass aggressively rewrites all aggregate loads and stores on
3179 /// a particular pointer (or any pointer derived from it which we can identify)
3180 /// with scalar loads and stores.
3181 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3182 // Befriend the base class so it can delegate to private visit methods.
3183 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
3185 const DataLayout &DL;
3187 /// Queue of pointer uses to analyze and potentially rewrite.
3188 SmallVector<Use *, 8> Queue;
3190 /// Set to prevent us from cycling with phi nodes and loops.
3191 SmallPtrSet<User *, 8> Visited;
3193 /// The current pointer use being rewritten. This is used to dig up the used
3194 /// value (as opposed to the user).
3198 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
3200 /// Rewrite loads and stores through a pointer and all pointers derived from
3202 bool rewrite(Instruction &I) {
3203 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
3205 bool Changed = false;
3206 while (!Queue.empty()) {
3207 U = Queue.pop_back_val();
3208 Changed |= visit(cast<Instruction>(U->getUser()));
3214 /// Enqueue all the users of the given instruction for further processing.
3215 /// This uses a set to de-duplicate users.
3216 void enqueueUsers(Instruction &I) {
3217 for (Use &U : I.uses())
3218 if (Visited.insert(U.getUser()).second)
3219 Queue.push_back(&U);
3222 // Conservative default is to not rewrite anything.
3223 bool visitInstruction(Instruction &I) { return false; }
3225 /// \brief Generic recursive split emission class.
3226 template <typename Derived> class OpSplitter {
3228 /// The builder used to form new instructions.
3230 /// The indices which to be used with insert- or extractvalue to select the
3231 /// appropriate value within the aggregate.
3232 SmallVector<unsigned, 4> Indices;
3233 /// The indices to a GEP instruction which will move Ptr to the correct slot
3234 /// within the aggregate.
3235 SmallVector<Value *, 4> GEPIndices;
3236 /// The base pointer of the original op, used as a base for GEPing the
3237 /// split operations.
3240 /// Initialize the splitter with an insertion point, Ptr and start with a
3241 /// single zero GEP index.
3242 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3243 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3246 /// \brief Generic recursive split emission routine.
3248 /// This method recursively splits an aggregate op (load or store) into
3249 /// scalar or vector ops. It splits recursively until it hits a single value
3250 /// and emits that single value operation via the template argument.
3252 /// The logic of this routine relies on GEPs and insertvalue and
3253 /// extractvalue all operating with the same fundamental index list, merely
3254 /// formatted differently (GEPs need actual values).
3256 /// \param Ty The type being split recursively into smaller ops.
3257 /// \param Agg The aggregate value being built up or stored, depending on
3258 /// whether this is splitting a load or a store respectively.
3259 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3260 if (Ty->isSingleValueType())
3261 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3263 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3264 unsigned OldSize = Indices.size();
3266 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3268 assert(Indices.size() == OldSize && "Did not return to the old size");
3269 Indices.push_back(Idx);
3270 GEPIndices.push_back(IRB.getInt32(Idx));
3271 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3272 GEPIndices.pop_back();
3278 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3279 unsigned OldSize = Indices.size();
3281 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3283 assert(Indices.size() == OldSize && "Did not return to the old size");
3284 Indices.push_back(Idx);
3285 GEPIndices.push_back(IRB.getInt32(Idx));
3286 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3287 GEPIndices.pop_back();
3293 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3297 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3298 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3299 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3301 /// Emit a leaf load of a single value. This is called at the leaves of the
3302 /// recursive emission to actually load values.
3303 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3304 assert(Ty->isSingleValueType());
3305 // Load the single value and insert it using the indices.
3307 IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep");
3308 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
3309 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3310 DEBUG(dbgs() << " to: " << *Load << "\n");
3314 bool visitLoadInst(LoadInst &LI) {
3315 assert(LI.getPointerOperand() == *U);
3316 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3319 // We have an aggregate being loaded, split it apart.
3320 DEBUG(dbgs() << " original: " << LI << "\n");
3321 LoadOpSplitter Splitter(&LI, *U);
3322 Value *V = UndefValue::get(LI.getType());
3323 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3324 LI.replaceAllUsesWith(V);
3325 LI.eraseFromParent();
3329 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3330 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3331 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3333 /// Emit a leaf store of a single value. This is called at the leaves of the
3334 /// recursive emission to actually produce stores.
3335 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3336 assert(Ty->isSingleValueType());
3337 // Extract the single value and store it using the indices.
3338 Value *Store = IRB.CreateStore(
3339 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
3340 IRB.CreateInBoundsGEP(nullptr, Ptr, GEPIndices, Name + ".gep"));
3342 DEBUG(dbgs() << " to: " << *Store << "\n");
3346 bool visitStoreInst(StoreInst &SI) {
3347 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3349 Value *V = SI.getValueOperand();
3350 if (V->getType()->isSingleValueType())
3353 // We have an aggregate being stored, split it apart.
3354 DEBUG(dbgs() << " original: " << SI << "\n");
3355 StoreOpSplitter Splitter(&SI, *U);
3356 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3357 SI.eraseFromParent();
3361 bool visitBitCastInst(BitCastInst &BC) {
3366 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3371 bool visitPHINode(PHINode &PN) {
3376 bool visitSelectInst(SelectInst &SI) {
3383 /// \brief Strip aggregate type wrapping.
3385 /// This removes no-op aggregate types wrapping an underlying type. It will
3386 /// strip as many layers of types as it can without changing either the type
3387 /// size or the allocated size.
3388 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3389 if (Ty->isSingleValueType())
3392 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3393 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3396 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3397 InnerTy = ArrTy->getElementType();
3398 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3399 const StructLayout *SL = DL.getStructLayout(STy);
3400 unsigned Index = SL->getElementContainingOffset(0);
3401 InnerTy = STy->getElementType(Index);
3406 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3407 TypeSize > DL.getTypeSizeInBits(InnerTy))
3410 return stripAggregateTypeWrapping(DL, InnerTy);
3413 /// \brief Try to find a partition of the aggregate type passed in for a given
3414 /// offset and size.
3416 /// This recurses through the aggregate type and tries to compute a subtype
3417 /// based on the offset and size. When the offset and size span a sub-section
3418 /// of an array, it will even compute a new array type for that sub-section,
3419 /// and the same for structs.
3421 /// Note that this routine is very strict and tries to find a partition of the
3422 /// type which produces the *exact* right offset and size. It is not forgiving
3423 /// when the size or offset cause either end of type-based partition to be off.
3424 /// Also, this is a best-effort routine. It is reasonable to give up and not
3425 /// return a type if necessary.
3426 static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3428 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3429 return stripAggregateTypeWrapping(DL, Ty);
3430 if (Offset > DL.getTypeAllocSize(Ty) ||
3431 (DL.getTypeAllocSize(Ty) - Offset) < Size)
3434 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3435 // We can't partition pointers...
3436 if (SeqTy->isPointerTy())
3439 Type *ElementTy = SeqTy->getElementType();
3440 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3441 uint64_t NumSkippedElements = Offset / ElementSize;
3442 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
3443 if (NumSkippedElements >= ArrTy->getNumElements())
3445 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
3446 if (NumSkippedElements >= VecTy->getNumElements())
3449 Offset -= NumSkippedElements * ElementSize;
3451 // First check if we need to recurse.
3452 if (Offset > 0 || Size < ElementSize) {
3453 // Bail if the partition ends in a different array element.
3454 if ((Offset + Size) > ElementSize)
3456 // Recurse through the element type trying to peel off offset bytes.
3457 return getTypePartition(DL, ElementTy, Offset, Size);
3459 assert(Offset == 0);
3461 if (Size == ElementSize)
3462 return stripAggregateTypeWrapping(DL, ElementTy);
3463 assert(Size > ElementSize);
3464 uint64_t NumElements = Size / ElementSize;
3465 if (NumElements * ElementSize != Size)
3467 return ArrayType::get(ElementTy, NumElements);
3470 StructType *STy = dyn_cast<StructType>(Ty);
3474 const StructLayout *SL = DL.getStructLayout(STy);
3475 if (Offset >= SL->getSizeInBytes())
3477 uint64_t EndOffset = Offset + Size;
3478 if (EndOffset > SL->getSizeInBytes())
3481 unsigned Index = SL->getElementContainingOffset(Offset);
3482 Offset -= SL->getElementOffset(Index);
3484 Type *ElementTy = STy->getElementType(Index);
3485 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3486 if (Offset >= ElementSize)
3487 return nullptr; // The offset points into alignment padding.
3489 // See if any partition must be contained by the element.
3490 if (Offset > 0 || Size < ElementSize) {
3491 if ((Offset + Size) > ElementSize)
3493 return getTypePartition(DL, ElementTy, Offset, Size);
3495 assert(Offset == 0);
3497 if (Size == ElementSize)
3498 return stripAggregateTypeWrapping(DL, ElementTy);
3500 StructType::element_iterator EI = STy->element_begin() + Index,
3501 EE = STy->element_end();
3502 if (EndOffset < SL->getSizeInBytes()) {
3503 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3504 if (Index == EndIndex)
3505 return nullptr; // Within a single element and its padding.
3507 // Don't try to form "natural" types if the elements don't line up with the
3509 // FIXME: We could potentially recurse down through the last element in the
3510 // sub-struct to find a natural end point.
3511 if (SL->getElementOffset(EndIndex) != EndOffset)
3514 assert(Index < EndIndex);
3515 EE = STy->element_begin() + EndIndex;
3518 // Try to build up a sub-structure.
3520 StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3521 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3522 if (Size != SubSL->getSizeInBytes())
3523 return nullptr; // The sub-struct doesn't have quite the size needed.
3528 /// \brief Pre-split loads and stores to simplify rewriting.
3530 /// We want to break up the splittable load+store pairs as much as
3531 /// possible. This is important to do as a preprocessing step, as once we
3532 /// start rewriting the accesses to partitions of the alloca we lose the
3533 /// necessary information to correctly split apart paired loads and stores
3534 /// which both point into this alloca. The case to consider is something like
3537 /// %a = alloca [12 x i8]
3538 /// %gep1 = getelementptr [12 x i8]* %a, i32 0, i32 0
3539 /// %gep2 = getelementptr [12 x i8]* %a, i32 0, i32 4
3540 /// %gep3 = getelementptr [12 x i8]* %a, i32 0, i32 8
3541 /// %iptr1 = bitcast i8* %gep1 to i64*
3542 /// %iptr2 = bitcast i8* %gep2 to i64*
3543 /// %fptr1 = bitcast i8* %gep1 to float*
3544 /// %fptr2 = bitcast i8* %gep2 to float*
3545 /// %fptr3 = bitcast i8* %gep3 to float*
3546 /// store float 0.0, float* %fptr1
3547 /// store float 1.0, float* %fptr2
3548 /// %v = load i64* %iptr1
3549 /// store i64 %v, i64* %iptr2
3550 /// %f1 = load float* %fptr2
3551 /// %f2 = load float* %fptr3
3553 /// Here we want to form 3 partitions of the alloca, each 4 bytes large, and
3554 /// promote everything so we recover the 2 SSA values that should have been
3555 /// there all along.
3557 /// \returns true if any changes are made.
3558 bool SROA::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
3559 DEBUG(dbgs() << "Pre-splitting loads and stores\n");
3561 // Track the loads and stores which are candidates for pre-splitting here, in
3562 // the order they first appear during the partition scan. These give stable
3563 // iteration order and a basis for tracking which loads and stores we
3565 SmallVector<LoadInst *, 4> Loads;
3566 SmallVector<StoreInst *, 4> Stores;
3568 // We need to accumulate the splits required of each load or store where we
3569 // can find them via a direct lookup. This is important to cross-check loads
3570 // and stores against each other. We also track the slice so that we can kill
3571 // all the slices that end up split.
3572 struct SplitOffsets {
3574 std::vector<uint64_t> Splits;
3576 SmallDenseMap<Instruction *, SplitOffsets, 8> SplitOffsetsMap;
3578 // Track loads out of this alloca which cannot, for any reason, be pre-split.
3579 // This is important as we also cannot pre-split stores of those loads!
3580 // FIXME: This is all pretty gross. It means that we can be more aggressive
3581 // in pre-splitting when the load feeding the store happens to come from
3582 // a separate alloca. Put another way, the effectiveness of SROA would be
3583 // decreased by a frontend which just concatenated all of its local allocas
3584 // into one big flat alloca. But defeating such patterns is exactly the job
3585 // SROA is tasked with! Sadly, to not have this discrepancy we would have
3586 // change store pre-splitting to actually force pre-splitting of the load
3587 // that feeds it *and all stores*. That makes pre-splitting much harder, but
3588 // maybe it would make it more principled?
3589 SmallPtrSet<LoadInst *, 8> UnsplittableLoads;
3591 DEBUG(dbgs() << " Searching for candidate loads and stores\n");
3592 for (auto &P : AS.partitions()) {
3593 for (Slice &S : P) {
3594 Instruction *I = cast<Instruction>(S.getUse()->getUser());
3595 if (!S.isSplittable() ||S.endOffset() <= P.endOffset()) {
3596 // If this was a load we have to track that it can't participate in any
3598 if (auto *LI = dyn_cast<LoadInst>(I))
3599 UnsplittableLoads.insert(LI);
3602 assert(P.endOffset() > S.beginOffset() &&
3603 "Empty or backwards partition!");
3605 // Determine if this is a pre-splittable slice.
3606 if (auto *LI = dyn_cast<LoadInst>(I)) {
3607 assert(!LI->isVolatile() && "Cannot split volatile loads!");
3609 // The load must be used exclusively to store into other pointers for
3610 // us to be able to arbitrarily pre-split it. The stores must also be
3611 // simple to avoid changing semantics.
3612 auto IsLoadSimplyStored = [](LoadInst *LI) {
3613 for (User *LU : LI->users()) {
3614 auto *SI = dyn_cast<StoreInst>(LU);
3615 if (!SI || !SI->isSimple())
3620 if (!IsLoadSimplyStored(LI)) {
3621 UnsplittableLoads.insert(LI);
3625 Loads.push_back(LI);
3626 } else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser())) {
3628 S.getUse() != &SI->getOperandUse(SI->getPointerOperandIndex()))
3630 auto *StoredLoad = dyn_cast<LoadInst>(SI->getValueOperand());
3631 if (!StoredLoad || !StoredLoad->isSimple())
3633 assert(!SI->isVolatile() && "Cannot split volatile stores!");
3635 Stores.push_back(SI);
3637 // Other uses cannot be pre-split.
3641 // Record the initial split.
3642 DEBUG(dbgs() << " Candidate: " << *I << "\n");
3643 auto &Offsets = SplitOffsetsMap[I];
3644 assert(Offsets.Splits.empty() &&
3645 "Should not have splits the first time we see an instruction!");
3647 Offsets.Splits.push_back(P.endOffset() - S.beginOffset());
3650 // Now scan the already split slices, and add a split for any of them which
3651 // we're going to pre-split.
3652 for (Slice *S : P.splitSliceTails()) {
3653 auto SplitOffsetsMapI =
3654 SplitOffsetsMap.find(cast<Instruction>(S->getUse()->getUser()));
3655 if (SplitOffsetsMapI == SplitOffsetsMap.end())
3657 auto &Offsets = SplitOffsetsMapI->second;
3659 assert(Offsets.S == S && "Found a mismatched slice!");
3660 assert(!Offsets.Splits.empty() &&
3661 "Cannot have an empty set of splits on the second partition!");
3662 assert(Offsets.Splits.back() ==
3663 P.beginOffset() - Offsets.S->beginOffset() &&
3664 "Previous split does not end where this one begins!");
3666 // Record each split. The last partition's end isn't needed as the size
3667 // of the slice dictates that.
3668 if (S->endOffset() > P.endOffset())
3669 Offsets.Splits.push_back(P.endOffset() - Offsets.S->beginOffset());
3673 // We may have split loads where some of their stores are split stores. For
3674 // such loads and stores, we can only pre-split them if their splits exactly
3675 // match relative to their starting offset. We have to verify this prior to
3678 std::remove_if(Stores.begin(), Stores.end(),
3679 [&UnsplittableLoads, &SplitOffsetsMap](StoreInst *SI) {
3680 // Lookup the load we are storing in our map of split
3682 auto *LI = cast<LoadInst>(SI->getValueOperand());
3683 // If it was completely unsplittable, then we're done,
3684 // and this store can't be pre-split.
3685 if (UnsplittableLoads.count(LI))
3688 auto LoadOffsetsI = SplitOffsetsMap.find(LI);
3689 if (LoadOffsetsI == SplitOffsetsMap.end())
3690 return false; // Unrelated loads are definitely safe.
3691 auto &LoadOffsets = LoadOffsetsI->second;
3693 // Now lookup the store's offsets.
3694 auto &StoreOffsets = SplitOffsetsMap[SI];
3696 // If the relative offsets of each split in the load and
3697 // store match exactly, then we can split them and we
3698 // don't need to remove them here.
3699 if (LoadOffsets.Splits == StoreOffsets.Splits)
3703 << " Mismatched splits for load and store:\n"
3704 << " " << *LI << "\n"
3705 << " " << *SI << "\n");
3707 // We've found a store and load that we need to split
3708 // with mismatched relative splits. Just give up on them
3709 // and remove both instructions from our list of
3711 UnsplittableLoads.insert(LI);
3715 // Now we have to go *back* through all the stores, because a later store may
3716 // have caused an earlier store's load to become unsplittable and if it is
3717 // unsplittable for the later store, then we can't rely on it being split in
3718 // the earlier store either.
3719 Stores.erase(std::remove_if(Stores.begin(), Stores.end(),
3720 [&UnsplittableLoads](StoreInst *SI) {
3722 cast<LoadInst>(SI->getValueOperand());
3723 return UnsplittableLoads.count(LI);
3726 // Once we've established all the loads that can't be split for some reason,
3727 // filter any that made it into our list out.
3728 Loads.erase(std::remove_if(Loads.begin(), Loads.end(),
3729 [&UnsplittableLoads](LoadInst *LI) {
3730 return UnsplittableLoads.count(LI);
3735 // If no loads or stores are left, there is no pre-splitting to be done for
3737 if (Loads.empty() && Stores.empty())
3740 // From here on, we can't fail and will be building new accesses, so rig up
3742 IRBuilderTy IRB(&AI);
3744 // Collect the new slices which we will merge into the alloca slices.
3745 SmallVector<Slice, 4> NewSlices;
3747 // Track any allocas we end up splitting loads and stores for so we iterate
3749 SmallPtrSet<AllocaInst *, 4> ResplitPromotableAllocas;
3751 // At this point, we have collected all of the loads and stores we can
3752 // pre-split, and the specific splits needed for them. We actually do the
3753 // splitting in a specific order in order to handle when one of the loads in
3754 // the value operand to one of the stores.
3756 // First, we rewrite all of the split loads, and just accumulate each split
3757 // load in a parallel structure. We also build the slices for them and append
3758 // them to the alloca slices.
3759 SmallDenseMap<LoadInst *, std::vector<LoadInst *>, 1> SplitLoadsMap;
3760 std::vector<LoadInst *> SplitLoads;
3761 const DataLayout &DL = AI.getModule()->getDataLayout();
3762 for (LoadInst *LI : Loads) {
3765 IntegerType *Ty = cast<IntegerType>(LI->getType());
3766 uint64_t LoadSize = Ty->getBitWidth() / 8;
3767 assert(LoadSize > 0 && "Cannot have a zero-sized integer load!");
3769 auto &Offsets = SplitOffsetsMap[LI];
3770 assert(LoadSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3771 "Slice size should always match load size exactly!");
3772 uint64_t BaseOffset = Offsets.S->beginOffset();
3773 assert(BaseOffset + LoadSize > BaseOffset &&
3774 "Cannot represent alloca access size using 64-bit integers!");
3776 Instruction *BasePtr = cast<Instruction>(LI->getPointerOperand());
3777 IRB.SetInsertPoint(BasicBlock::iterator(LI));
3779 DEBUG(dbgs() << " Splitting load: " << *LI << "\n");
3781 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3782 int Idx = 0, Size = Offsets.Splits.size();
3784 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3785 auto *PartPtrTy = PartTy->getPointerTo(LI->getPointerAddressSpace());
3786 LoadInst *PLoad = IRB.CreateAlignedLoad(
3787 getAdjustedPtr(IRB, DL, BasePtr,
3788 APInt(DL.getPointerSizeInBits(), PartOffset),
3789 PartPtrTy, BasePtr->getName() + "."),
3790 getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3793 // Append this load onto the list of split loads so we can find it later
3794 // to rewrite the stores.
3795 SplitLoads.push_back(PLoad);
3797 // Now build a new slice for the alloca.
3798 NewSlices.push_back(
3799 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3800 &PLoad->getOperandUse(PLoad->getPointerOperandIndex()),
3801 /*IsSplittable*/ false));
3802 DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3803 << ", " << NewSlices.back().endOffset() << "): " << *PLoad
3806 // See if we've handled all the splits.
3810 // Setup the next partition.
3811 PartOffset = Offsets.Splits[Idx];
3813 PartSize = (Idx < Size ? Offsets.Splits[Idx] : LoadSize) - PartOffset;
3816 // Now that we have the split loads, do the slow walk over all uses of the
3817 // load and rewrite them as split stores, or save the split loads to use
3818 // below if the store is going to be split there anyways.
3819 bool DeferredStores = false;
3820 for (User *LU : LI->users()) {
3821 StoreInst *SI = cast<StoreInst>(LU);
3822 if (!Stores.empty() && SplitOffsetsMap.count(SI)) {
3823 DeferredStores = true;
3824 DEBUG(dbgs() << " Deferred splitting of store: " << *SI << "\n");
3828 Value *StoreBasePtr = SI->getPointerOperand();
3829 IRB.SetInsertPoint(BasicBlock::iterator(SI));
3831 DEBUG(dbgs() << " Splitting store of load: " << *SI << "\n");
3833 for (int Idx = 0, Size = SplitLoads.size(); Idx < Size; ++Idx) {
3834 LoadInst *PLoad = SplitLoads[Idx];
3835 uint64_t PartOffset = Idx == 0 ? 0 : Offsets.Splits[Idx - 1];
3837 PLoad->getType()->getPointerTo(SI->getPointerAddressSpace());
3839 StoreInst *PStore = IRB.CreateAlignedStore(
3840 PLoad, getAdjustedPtr(IRB, DL, StoreBasePtr,
3841 APInt(DL.getPointerSizeInBits(), PartOffset),
3842 PartPtrTy, StoreBasePtr->getName() + "."),
3843 getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3845 DEBUG(dbgs() << " +" << PartOffset << ":" << *PStore << "\n");
3848 // We want to immediately iterate on any allocas impacted by splitting
3849 // this store, and we have to track any promotable alloca (indicated by
3850 // a direct store) as needing to be resplit because it is no longer
3852 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(StoreBasePtr)) {
3853 ResplitPromotableAllocas.insert(OtherAI);
3854 Worklist.insert(OtherAI);
3855 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3856 StoreBasePtr->stripInBoundsOffsets())) {
3857 Worklist.insert(OtherAI);
3860 // Mark the original store as dead.
3861 DeadInsts.insert(SI);
3864 // Save the split loads if there are deferred stores among the users.
3866 SplitLoadsMap.insert(std::make_pair(LI, std::move(SplitLoads)));
3868 // Mark the original load as dead and kill the original slice.
3869 DeadInsts.insert(LI);
3873 // Second, we rewrite all of the split stores. At this point, we know that
3874 // all loads from this alloca have been split already. For stores of such
3875 // loads, we can simply look up the pre-existing split loads. For stores of
3876 // other loads, we split those loads first and then write split stores of
3878 for (StoreInst *SI : Stores) {
3879 auto *LI = cast<LoadInst>(SI->getValueOperand());
3880 IntegerType *Ty = cast<IntegerType>(LI->getType());
3881 uint64_t StoreSize = Ty->getBitWidth() / 8;
3882 assert(StoreSize > 0 && "Cannot have a zero-sized integer store!");
3884 auto &Offsets = SplitOffsetsMap[SI];
3885 assert(StoreSize == Offsets.S->endOffset() - Offsets.S->beginOffset() &&
3886 "Slice size should always match load size exactly!");
3887 uint64_t BaseOffset = Offsets.S->beginOffset();
3888 assert(BaseOffset + StoreSize > BaseOffset &&
3889 "Cannot represent alloca access size using 64-bit integers!");
3891 Value *LoadBasePtr = LI->getPointerOperand();
3892 Instruction *StoreBasePtr = cast<Instruction>(SI->getPointerOperand());
3894 DEBUG(dbgs() << " Splitting store: " << *SI << "\n");
3896 // Check whether we have an already split load.
3897 auto SplitLoadsMapI = SplitLoadsMap.find(LI);
3898 std::vector<LoadInst *> *SplitLoads = nullptr;
3899 if (SplitLoadsMapI != SplitLoadsMap.end()) {
3900 SplitLoads = &SplitLoadsMapI->second;
3901 assert(SplitLoads->size() == Offsets.Splits.size() + 1 &&
3902 "Too few split loads for the number of splits in the store!");
3904 DEBUG(dbgs() << " of load: " << *LI << "\n");
3907 uint64_t PartOffset = 0, PartSize = Offsets.Splits.front();
3908 int Idx = 0, Size = Offsets.Splits.size();
3910 auto *PartTy = Type::getIntNTy(Ty->getContext(), PartSize * 8);
3911 auto *PartPtrTy = PartTy->getPointerTo(SI->getPointerAddressSpace());
3913 // Either lookup a split load or create one.
3916 PLoad = (*SplitLoads)[Idx];
3918 IRB.SetInsertPoint(BasicBlock::iterator(LI));
3919 PLoad = IRB.CreateAlignedLoad(
3920 getAdjustedPtr(IRB, DL, LoadBasePtr,
3921 APInt(DL.getPointerSizeInBits(), PartOffset),
3922 PartPtrTy, LoadBasePtr->getName() + "."),
3923 getAdjustedAlignment(LI, PartOffset, DL), /*IsVolatile*/ false,
3927 // And store this partition.
3928 IRB.SetInsertPoint(BasicBlock::iterator(SI));
3929 StoreInst *PStore = IRB.CreateAlignedStore(
3930 PLoad, getAdjustedPtr(IRB, DL, StoreBasePtr,
3931 APInt(DL.getPointerSizeInBits(), PartOffset),
3932 PartPtrTy, StoreBasePtr->getName() + "."),
3933 getAdjustedAlignment(SI, PartOffset, DL), /*IsVolatile*/ false);
3935 // Now build a new slice for the alloca.
3936 NewSlices.push_back(
3937 Slice(BaseOffset + PartOffset, BaseOffset + PartOffset + PartSize,
3938 &PStore->getOperandUse(PStore->getPointerOperandIndex()),
3939 /*IsSplittable*/ false));
3940 DEBUG(dbgs() << " new slice [" << NewSlices.back().beginOffset()
3941 << ", " << NewSlices.back().endOffset() << "): " << *PStore
3944 DEBUG(dbgs() << " of split load: " << *PLoad << "\n");
3947 // See if we've finished all the splits.
3951 // Setup the next partition.
3952 PartOffset = Offsets.Splits[Idx];
3954 PartSize = (Idx < Size ? Offsets.Splits[Idx] : StoreSize) - PartOffset;
3957 // We want to immediately iterate on any allocas impacted by splitting
3958 // this load, which is only relevant if it isn't a load of this alloca and
3959 // thus we didn't already split the loads above. We also have to keep track
3960 // of any promotable allocas we split loads on as they can no longer be
3963 if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(LoadBasePtr)) {
3964 assert(OtherAI != &AI && "We can't re-split our own alloca!");
3965 ResplitPromotableAllocas.insert(OtherAI);
3966 Worklist.insert(OtherAI);
3967 } else if (AllocaInst *OtherAI = dyn_cast<AllocaInst>(
3968 LoadBasePtr->stripInBoundsOffsets())) {
3969 assert(OtherAI != &AI && "We can't re-split our own alloca!");
3970 Worklist.insert(OtherAI);
3974 // Mark the original store as dead now that we've split it up and kill its
3975 // slice. Note that we leave the original load in place unless this store
3976 // was its only use. It may in turn be split up if it is an alloca load
3977 // for some other alloca, but it may be a normal load. This may introduce
3978 // redundant loads, but where those can be merged the rest of the optimizer
3979 // should handle the merging, and this uncovers SSA splits which is more
3980 // important. In practice, the original loads will almost always be fully
3981 // split and removed eventually, and the splits will be merged by any
3982 // trivial CSE, including instcombine.
3983 if (LI->hasOneUse()) {
3984 assert(*LI->user_begin() == SI && "Single use isn't this store!");
3985 DeadInsts.insert(LI);
3987 DeadInsts.insert(SI);
3991 // Remove the killed slices that have ben pre-split.
3992 AS.erase(std::remove_if(AS.begin(), AS.end(), [](const Slice &S) {
3996 // Insert our new slices. This will sort and merge them into the sorted
3998 AS.insert(NewSlices);
4000 DEBUG(dbgs() << " Pre-split slices:\n");
4002 for (auto I = AS.begin(), E = AS.end(); I != E; ++I)
4003 DEBUG(AS.print(dbgs(), I, " "));
4006 // Finally, don't try to promote any allocas that new require re-splitting.
4007 // They have already been added to the worklist above.
4008 PromotableAllocas.erase(
4010 PromotableAllocas.begin(), PromotableAllocas.end(),
4011 [&](AllocaInst *AI) { return ResplitPromotableAllocas.count(AI); }),
4012 PromotableAllocas.end());
4017 /// \brief Rewrite an alloca partition's users.
4019 /// This routine drives both of the rewriting goals of the SROA pass. It tries
4020 /// to rewrite uses of an alloca partition to be conducive for SSA value
4021 /// promotion. If the partition needs a new, more refined alloca, this will
4022 /// build that new alloca, preserving as much type information as possible, and
4023 /// rewrite the uses of the old alloca to point at the new one and have the
4024 /// appropriate new offsets. It also evaluates how successful the rewrite was
4025 /// at enabling promotion and if it was successful queues the alloca to be
4027 AllocaInst *SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
4028 AllocaSlices::Partition &P) {
4029 // Try to compute a friendly type for this partition of the alloca. This
4030 // won't always succeed, in which case we fall back to a legal integer type
4031 // or an i8 array of an appropriate size.
4032 Type *SliceTy = nullptr;
4033 const DataLayout &DL = AI.getModule()->getDataLayout();
4034 if (Type *CommonUseTy = findCommonType(P.begin(), P.end(), P.endOffset()))
4035 if (DL.getTypeAllocSize(CommonUseTy) >= P.size())
4036 SliceTy = CommonUseTy;
4038 if (Type *TypePartitionTy = getTypePartition(DL, AI.getAllocatedType(),
4039 P.beginOffset(), P.size()))
4040 SliceTy = TypePartitionTy;
4041 if ((!SliceTy || (SliceTy->isArrayTy() &&
4042 SliceTy->getArrayElementType()->isIntegerTy())) &&
4043 DL.isLegalInteger(P.size() * 8))
4044 SliceTy = Type::getIntNTy(*C, P.size() * 8);
4046 SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
4047 assert(DL.getTypeAllocSize(SliceTy) >= P.size());
4049 bool IsIntegerPromotable = isIntegerWideningViable(P, SliceTy, DL);
4052 IsIntegerPromotable ? nullptr : isVectorPromotionViable(P, DL);
4056 // Check for the case where we're going to rewrite to a new alloca of the
4057 // exact same type as the original, and with the same access offsets. In that
4058 // case, re-use the existing alloca, but still run through the rewriter to
4059 // perform phi and select speculation.
4061 if (SliceTy == AI.getAllocatedType()) {
4062 assert(P.beginOffset() == 0 &&
4063 "Non-zero begin offset but same alloca type");
4065 // FIXME: We should be able to bail at this point with "nothing changed".
4066 // FIXME: We might want to defer PHI speculation until after here.
4067 // FIXME: return nullptr;
4069 unsigned Alignment = AI.getAlignment();
4071 // The minimum alignment which users can rely on when the explicit
4072 // alignment is omitted or zero is that required by the ABI for this
4074 Alignment = DL.getABITypeAlignment(AI.getAllocatedType());
4076 Alignment = MinAlign(Alignment, P.beginOffset());
4077 // If we will get at least this much alignment from the type alone, leave
4078 // the alloca's alignment unconstrained.
4079 if (Alignment <= DL.getABITypeAlignment(SliceTy))
4081 NewAI = new AllocaInst(
4082 SliceTy, nullptr, Alignment,
4083 AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
4087 DEBUG(dbgs() << "Rewriting alloca partition "
4088 << "[" << P.beginOffset() << "," << P.endOffset()
4089 << ") to: " << *NewAI << "\n");
4091 // Track the high watermark on the worklist as it is only relevant for
4092 // promoted allocas. We will reset it to this point if the alloca is not in
4093 // fact scheduled for promotion.
4094 unsigned PPWOldSize = PostPromotionWorklist.size();
4095 unsigned NumUses = 0;
4096 SmallPtrSet<PHINode *, 8> PHIUsers;
4097 SmallPtrSet<SelectInst *, 8> SelectUsers;
4099 AllocaSliceRewriter Rewriter(DL, AS, *this, AI, *NewAI, P.beginOffset(),
4100 P.endOffset(), IsIntegerPromotable, VecTy,
4101 PHIUsers, SelectUsers);
4102 bool Promotable = true;
4103 for (Slice *S : P.splitSliceTails()) {
4104 Promotable &= Rewriter.visit(S);
4107 for (Slice &S : P) {
4108 Promotable &= Rewriter.visit(&S);
4112 NumAllocaPartitionUses += NumUses;
4113 MaxUsesPerAllocaPartition =
4114 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
4116 // Now that we've processed all the slices in the new partition, check if any
4117 // PHIs or Selects would block promotion.
4118 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
4121 if (!isSafePHIToSpeculate(**I)) {
4124 SelectUsers.clear();
4127 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
4128 E = SelectUsers.end();
4130 if (!isSafeSelectToSpeculate(**I)) {
4133 SelectUsers.clear();
4138 if (PHIUsers.empty() && SelectUsers.empty()) {
4139 // Promote the alloca.
4140 PromotableAllocas.push_back(NewAI);
4142 // If we have either PHIs or Selects to speculate, add them to those
4143 // worklists and re-queue the new alloca so that we promote in on the
4145 for (PHINode *PHIUser : PHIUsers)
4146 SpeculatablePHIs.insert(PHIUser);
4147 for (SelectInst *SelectUser : SelectUsers)
4148 SpeculatableSelects.insert(SelectUser);
4149 Worklist.insert(NewAI);
4152 // If we can't promote the alloca, iterate on it to check for new
4153 // refinements exposed by splitting the current alloca. Don't iterate on an
4154 // alloca which didn't actually change and didn't get promoted.
4156 Worklist.insert(NewAI);
4158 // Drop any post-promotion work items if promotion didn't happen.
4159 while (PostPromotionWorklist.size() > PPWOldSize)
4160 PostPromotionWorklist.pop_back();
4166 /// \brief Walks the slices of an alloca and form partitions based on them,
4167 /// rewriting each of their uses.
4168 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
4169 if (AS.begin() == AS.end())
4172 unsigned NumPartitions = 0;
4173 bool Changed = false;
4174 const DataLayout &DL = AI.getModule()->getDataLayout();
4176 // First try to pre-split loads and stores.
4177 Changed |= presplitLoadsAndStores(AI, AS);
4179 // Now that we have identified any pre-splitting opportunities, mark any
4180 // splittable (non-whole-alloca) loads and stores as unsplittable. If we fail
4181 // to split these during pre-splitting, we want to force them to be
4182 // rewritten into a partition.
4183 bool IsSorted = true;
4184 for (Slice &S : AS) {
4185 if (!S.isSplittable())
4187 // FIXME: We currently leave whole-alloca splittable loads and stores. This
4188 // used to be the only splittable loads and stores and we need to be
4189 // confident that the above handling of splittable loads and stores is
4190 // completely sufficient before we forcibly disable the remaining handling.
4191 if (S.beginOffset() == 0 &&
4192 S.endOffset() >= DL.getTypeAllocSize(AI.getAllocatedType()))
4194 if (isa<LoadInst>(S.getUse()->getUser()) ||
4195 isa<StoreInst>(S.getUse()->getUser())) {
4196 S.makeUnsplittable();
4201 std::sort(AS.begin(), AS.end());
4203 /// \brief Describes the allocas introduced by rewritePartition
4204 /// in order to migrate the debug info.
4209 Piece(AllocaInst *AI, uint64_t O, uint64_t S)
4210 : Alloca(AI), Offset(O), Size(S) {}
4212 SmallVector<Piece, 4> Pieces;
4214 // Rewrite each partition.
4215 for (auto &P : AS.partitions()) {
4216 if (AllocaInst *NewAI = rewritePartition(AI, AS, P)) {
4219 uint64_t SizeOfByte = 8;
4220 uint64_t AllocaSize = DL.getTypeSizeInBits(NewAI->getAllocatedType());
4221 // Don't include any padding.
4222 uint64_t Size = std::min(AllocaSize, P.size() * SizeOfByte);
4223 Pieces.push_back(Piece(NewAI, P.beginOffset() * SizeOfByte, Size));
4229 NumAllocaPartitions += NumPartitions;
4230 MaxPartitionsPerAlloca =
4231 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
4233 // Migrate debug information from the old alloca to the new alloca(s)
4234 // and the individual partitions.
4235 if (DbgDeclareInst *DbgDecl = FindAllocaDbgDeclare(&AI)) {
4236 auto *Var = DbgDecl->getVariable();
4237 auto *Expr = DbgDecl->getExpression();
4238 DIBuilder DIB(*AI.getParent()->getParent()->getParent(),
4239 /*AllowUnresolved*/ false);
4240 bool IsSplit = Pieces.size() > 1;
4241 for (auto Piece : Pieces) {
4242 // Create a piece expression describing the new partition or reuse AI's
4243 // expression if there is only one partition.
4244 auto *PieceExpr = Expr;
4245 if (IsSplit || Expr->isBitPiece()) {
4246 // If this alloca is already a scalar replacement of a larger aggregate,
4247 // Piece.Offset describes the offset inside the scalar.
4248 uint64_t Offset = Expr->isBitPiece() ? Expr->getBitPieceOffset() : 0;
4249 uint64_t Start = Offset + Piece.Offset;
4250 uint64_t Size = Piece.Size;
4251 if (Expr->isBitPiece()) {
4252 uint64_t AbsEnd = Expr->getBitPieceOffset() + Expr->getBitPieceSize();
4253 if (Start >= AbsEnd)
4254 // No need to describe a SROAed padding.
4256 Size = std::min(Size, AbsEnd - Start);
4258 PieceExpr = DIB.createBitPieceExpression(Start, Size);
4261 // Remove any existing dbg.declare intrinsic describing the same alloca.
4262 if (DbgDeclareInst *OldDDI = FindAllocaDbgDeclare(Piece.Alloca))
4263 OldDDI->eraseFromParent();
4265 DIB.insertDeclare(Piece.Alloca, Var, PieceExpr, DbgDecl->getDebugLoc(),
4272 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
4273 void SROA::clobberUse(Use &U) {
4275 // Replace the use with an undef value.
4276 U = UndefValue::get(OldV->getType());
4278 // Check for this making an instruction dead. We have to garbage collect
4279 // all the dead instructions to ensure the uses of any alloca end up being
4281 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
4282 if (isInstructionTriviallyDead(OldI)) {
4283 DeadInsts.insert(OldI);
4287 /// \brief Analyze an alloca for SROA.
4289 /// This analyzes the alloca to ensure we can reason about it, builds
4290 /// the slices of the alloca, and then hands it off to be split and
4291 /// rewritten as needed.
4292 bool SROA::runOnAlloca(AllocaInst &AI) {
4293 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
4294 ++NumAllocasAnalyzed;
4296 // Special case dead allocas, as they're trivial.
4297 if (AI.use_empty()) {
4298 AI.eraseFromParent();
4301 const DataLayout &DL = AI.getModule()->getDataLayout();
4303 // Skip alloca forms that this analysis can't handle.
4304 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
4305 DL.getTypeAllocSize(AI.getAllocatedType()) == 0)
4308 bool Changed = false;
4310 // First, split any FCA loads and stores touching this alloca to promote
4311 // better splitting and promotion opportunities.
4312 AggLoadStoreRewriter AggRewriter(DL);
4313 Changed |= AggRewriter.rewrite(AI);
4315 // Build the slices using a recursive instruction-visiting builder.
4316 AllocaSlices AS(DL, AI);
4317 DEBUG(AS.print(dbgs()));
4321 // Delete all the dead users of this alloca before splitting and rewriting it.
4322 for (Instruction *DeadUser : AS.getDeadUsers()) {
4323 // Free up everything used by this instruction.
4324 for (Use &DeadOp : DeadUser->operands())
4327 // Now replace the uses of this instruction.
4328 DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
4330 // And mark it for deletion.
4331 DeadInsts.insert(DeadUser);
4334 for (Use *DeadOp : AS.getDeadOperands()) {
4335 clobberUse(*DeadOp);
4339 // No slices to split. Leave the dead alloca for a later pass to clean up.
4340 if (AS.begin() == AS.end())
4343 Changed |= splitAlloca(AI, AS);
4345 DEBUG(dbgs() << " Speculating PHIs\n");
4346 while (!SpeculatablePHIs.empty())
4347 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
4349 DEBUG(dbgs() << " Speculating Selects\n");
4350 while (!SpeculatableSelects.empty())
4351 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
4356 /// \brief Delete the dead instructions accumulated in this run.
4358 /// Recursively deletes the dead instructions we've accumulated. This is done
4359 /// at the very end to maximize locality of the recursive delete and to
4360 /// minimize the problems of invalidated instruction pointers as such pointers
4361 /// are used heavily in the intermediate stages of the algorithm.
4363 /// We also record the alloca instructions deleted here so that they aren't
4364 /// subsequently handed to mem2reg to promote.
4365 void SROA::deleteDeadInstructions(
4366 SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
4367 while (!DeadInsts.empty()) {
4368 Instruction *I = DeadInsts.pop_back_val();
4369 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
4371 I->replaceAllUsesWith(UndefValue::get(I->getType()));
4373 for (Use &Operand : I->operands())
4374 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
4375 // Zero out the operand and see if it becomes trivially dead.
4377 if (isInstructionTriviallyDead(U))
4378 DeadInsts.insert(U);
4381 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
4382 DeletedAllocas.insert(AI);
4383 if (DbgDeclareInst *DbgDecl = FindAllocaDbgDeclare(AI))
4384 DbgDecl->eraseFromParent();
4388 I->eraseFromParent();
4392 static void enqueueUsersInWorklist(Instruction &I,
4393 SmallVectorImpl<Instruction *> &Worklist,
4394 SmallPtrSetImpl<Instruction *> &Visited) {
4395 for (User *U : I.users())
4396 if (Visited.insert(cast<Instruction>(U)).second)
4397 Worklist.push_back(cast<Instruction>(U));
4400 /// \brief Promote the allocas, using the best available technique.
4402 /// This attempts to promote whatever allocas have been identified as viable in
4403 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
4404 /// If there is a domtree available, we attempt to promote using the full power
4405 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
4406 /// based on the SSAUpdater utilities. This function returns whether any
4407 /// promotion occurred.
4408 bool SROA::promoteAllocas(Function &F) {
4409 if (PromotableAllocas.empty())
4412 NumPromoted += PromotableAllocas.size();
4414 if (DT && !ForceSSAUpdater) {
4415 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
4416 PromoteMemToReg(PromotableAllocas, *DT, nullptr, AC);
4417 PromotableAllocas.clear();
4421 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
4423 DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false);
4424 SmallVector<Instruction *, 64> Insts;
4426 // We need a worklist to walk the uses of each alloca.
4427 SmallVector<Instruction *, 8> Worklist;
4428 SmallPtrSet<Instruction *, 8> Visited;
4429 SmallVector<Instruction *, 32> DeadInsts;
4431 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
4432 AllocaInst *AI = PromotableAllocas[Idx];
4437 enqueueUsersInWorklist(*AI, Worklist, Visited);
4439 while (!Worklist.empty()) {
4440 Instruction *I = Worklist.pop_back_val();
4442 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
4443 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
4444 // leading to them) here. Eventually it should use them to optimize the
4445 // scalar values produced.
4446 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
4447 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
4448 II->getIntrinsicID() == Intrinsic::lifetime_end);
4449 II->eraseFromParent();
4453 // Push the loads and stores we find onto the list. SROA will already
4454 // have validated that all loads and stores are viable candidates for
4456 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
4457 assert(LI->getType() == AI->getAllocatedType());
4458 Insts.push_back(LI);
4461 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
4462 assert(SI->getValueOperand()->getType() == AI->getAllocatedType());
4463 Insts.push_back(SI);
4467 // For everything else, we know that only no-op bitcasts and GEPs will
4468 // make it this far, just recurse through them and recall them for later
4470 DeadInsts.push_back(I);
4471 enqueueUsersInWorklist(*I, Worklist, Visited);
4473 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
4474 while (!DeadInsts.empty())
4475 DeadInsts.pop_back_val()->eraseFromParent();
4476 AI->eraseFromParent();
4479 PromotableAllocas.clear();
4483 bool SROA::runOnFunction(Function &F) {
4484 if (skipOptnoneFunction(F))
4487 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
4488 C = &F.getContext();
4489 DominatorTreeWrapperPass *DTWP =
4490 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
4491 DT = DTWP ? &DTWP->getDomTree() : nullptr;
4492 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
4494 BasicBlock &EntryBB = F.getEntryBlock();
4495 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
4497 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
4498 Worklist.insert(AI);
4501 bool Changed = false;
4502 // A set of deleted alloca instruction pointers which should be removed from
4503 // the list of promotable allocas.
4504 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
4507 while (!Worklist.empty()) {
4508 Changed |= runOnAlloca(*Worklist.pop_back_val());
4509 deleteDeadInstructions(DeletedAllocas);
4511 // Remove the deleted allocas from various lists so that we don't try to
4512 // continue processing them.
4513 if (!DeletedAllocas.empty()) {
4514 auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
4515 Worklist.remove_if(IsInSet);
4516 PostPromotionWorklist.remove_if(IsInSet);
4517 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
4518 PromotableAllocas.end(),
4520 PromotableAllocas.end());
4521 DeletedAllocas.clear();
4525 Changed |= promoteAllocas(F);
4527 Worklist = PostPromotionWorklist;
4528 PostPromotionWorklist.clear();
4529 } while (!Worklist.empty());
4534 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
4535 AU.addRequired<AssumptionCacheTracker>();
4536 if (RequiresDomTree)
4537 AU.addRequired<DominatorTreeWrapperPass>();
4538 AU.setPreservesCFG();