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/AssumptionTracker.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 // Forward declare an iterator to befriend it.
241 class partition_iterator;
243 /// \brief A partition of the slices.
245 /// An ephemeral representation for a range of slices which can be viewed as
246 /// a partition of the alloca. This range represents a span of the alloca's
247 /// memory which cannot be split, and provides access to all of the slices
248 /// overlapping some part of the partition.
250 /// Objects of this type are produced by traversing the alloca's slices, but
251 /// are only ephemeral and not persistent.
254 friend class AllocaSlices;
255 friend class AllocaSlices::partition_iterator;
257 /// \brief The begining and ending offsets of the alloca for this partition.
258 uint64_t BeginOffset, EndOffset;
260 /// \brief The start end end iterators of this partition.
263 /// \brief A collection of split slices.
264 SmallVector<Slice *, 4> SplitSlices;
266 /// \brief Raw constructor builds an empty partition starting and ending at
267 /// the given iterator.
268 Partition(iterator SI) : SI(SI), SJ(SI) {}
271 /// \brief The start offset of this partition.
273 /// All of the contained slices start at or after this offset.
274 uint64_t beginOffset() const { return BeginOffset; }
276 /// \brief The end offset of this partition.
278 /// All of the contained slices end at or before this offset.
279 uint64_t endOffset() const { return EndOffset; }
281 /// \brief The size of the partition.
283 /// Note that this can never be zero.
284 uint64_t size() const {
285 assert(BeginOffset < EndOffset && "Partitions must span some bytes!");
286 return EndOffset - BeginOffset;
289 /// \brief Test whether this partition contains no slices, and merely spans
290 /// a region occupied by split slices.
291 bool empty() const { return SI == SJ; }
293 /// \name Iterate contained slices.
294 /// All of these slices are fully contained in the partition. They may be
295 /// splittable or unsplittable.
297 iterator begin() const { return SI; }
298 iterator end() const { return SJ; }
301 /// \brief Get the sequence of split slices.
302 ArrayRef<Slice *> splitSlices() const { return SplitSlices; }
305 /// \brief An iterator over partitions of the alloca's slices.
307 /// This iterator implements the core algorithm for partitioning the alloca's
308 /// slices. It is a forward iterator as we don't support backtracking for
309 /// efficiency reasons, and re-use a single storage area to maintain the
310 /// current set of split slices.
312 /// It is templated on the slice iterator type to use so that it can operate
313 /// with either const or non-const slice iterators.
314 class partition_iterator
315 : public iterator_facade_base<partition_iterator,
316 std::forward_iterator_tag, Partition> {
317 friend class AllocaSlices;
319 /// \brief Most of the state for walking the partitions is held in a class
320 /// with a nice interface for examining them.
323 /// \brief We need to keep the end of the slices to know when to stop.
324 AllocaSlices::iterator SE;
326 /// \brief We also need to keep track of the maximum split end offset seen.
327 /// FIXME: Do we really?
328 uint64_t MaxSplitSliceEndOffset;
330 /// \brief Sets the partition to be empty at given iterator, and sets the
332 partition_iterator(AllocaSlices::iterator SI, AllocaSlices::iterator SE)
333 : P(SI), SE(SE), MaxSplitSliceEndOffset(0) {
334 // If not already at the end, advance our state to form the initial
340 /// \brief Advance the iterator to the next partition.
342 /// Requires that the iterator not be at the end of the slices.
344 assert((P.SI != SE || !P.SplitSlices.empty()) &&
345 "Cannot advance past the end of the slices!");
347 // Clear out any split uses which have ended.
348 if (!P.SplitSlices.empty()) {
349 if (P.EndOffset >= MaxSplitSliceEndOffset) {
350 // If we've finished all splits, this is easy.
351 P.SplitSlices.clear();
352 MaxSplitSliceEndOffset = 0;
354 // Remove the uses which have ended in the prior partition. This
355 // cannot change the max split slice end because we just checked that
356 // the prior partition ended prior to that max.
359 P.SplitSlices.begin(), P.SplitSlices.end(),
360 [&](Slice *S) { return S->endOffset() <= P.EndOffset; }),
361 P.SplitSlices.end());
362 assert(std::any_of(P.SplitSlices.begin(), P.SplitSlices.end(),
364 return S->endOffset() == MaxSplitSliceEndOffset;
366 "Could not find the current max split slice offset!");
367 assert(std::all_of(P.SplitSlices.begin(), P.SplitSlices.end(),
369 return S->endOffset() <= MaxSplitSliceEndOffset;
371 "Max split slice end offset is not actually the max!");
375 // If P.SI is already at the end, then we've cleared the split tail and
376 // now have an end iterator.
378 assert(P.SplitSlices.empty() && "Failed to clear the split slices!");
382 // If we had a non-empty partition previously, set up the state for
383 // subsequent partitions.
385 // Accumulate all the splittable slices which started in the old
386 // partition into the split list.
388 if (S.isSplittable() && S.endOffset() > P.EndOffset) {
389 P.SplitSlices.push_back(&S);
390 MaxSplitSliceEndOffset =
391 std::max(S.endOffset(), MaxSplitSliceEndOffset);
394 // Start from the end of the previous partition.
397 // If P.SI is now at the end, we at most have a tail of split slices.
399 P.BeginOffset = P.EndOffset;
400 P.EndOffset = MaxSplitSliceEndOffset;
404 // If the we have split slices and the next slice is after a gap and is
405 // not splittable immediately form an empty partition for the split
406 // slices up until the next slice begins.
407 if (!P.SplitSlices.empty() && P.SI->beginOffset() != P.EndOffset &&
408 !P.SI->isSplittable()) {
409 P.BeginOffset = P.EndOffset;
410 P.EndOffset = P.SI->beginOffset();
415 // OK, we need to consume new slices. Set the end offset based on the
416 // current slice, and step SJ past it. The beginning offset of the
417 // parttion is the beginning offset of the next slice unless we have
418 // pre-existing split slices that are continuing, in which case we begin
419 // at the prior end offset.
420 P.BeginOffset = P.SplitSlices.empty() ? P.SI->beginOffset() : P.EndOffset;
421 P.EndOffset = P.SI->endOffset();
424 // There are two strategies to form a partition based on whether the
425 // partition starts with an unsplittable slice or a splittable slice.
426 if (!P.SI->isSplittable()) {
427 // When we're forming an unsplittable region, it must always start at
428 // the first slice and will extend through its end.
429 assert(P.BeginOffset == P.SI->beginOffset());
431 // Form a partition including all of the overlapping slices with this
432 // unsplittable slice.
433 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
434 if (!P.SJ->isSplittable())
435 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
439 // We have a partition across a set of overlapping unsplittable
444 // If we're starting with a splittable slice, then we need to form
445 // a synthetic partition spanning it and any other overlapping splittable
447 assert(P.SI->isSplittable() && "Forming a splittable partition!");
449 // Collect all of the overlapping splittable slices.
450 while (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset &&
451 P.SJ->isSplittable()) {
452 P.EndOffset = std::max(P.EndOffset, P.SJ->endOffset());
456 // Back upiP.EndOffset if we ended the span early when encountering an
457 // unsplittable slice. This synthesizes the early end offset of
458 // a partition spanning only splittable slices.
459 if (P.SJ != SE && P.SJ->beginOffset() < P.EndOffset) {
460 assert(!P.SJ->isSplittable());
461 P.EndOffset = P.SJ->beginOffset();
466 bool operator==(const partition_iterator &RHS) const {
467 assert(SE == RHS.SE &&
468 "End iterators don't match between compared partition iterators!");
470 // The observed positions of partitions is marked by the P.SI iterator and
471 // the emptyness of the split slices. The latter is only relevant when
472 // P.SI == SE, as the end iterator will additionally have an empty split
473 // slices list, but the prior may have the same P.SI and a tail of split
475 if (P.SI == RHS.P.SI &&
476 P.SplitSlices.empty() == RHS.P.SplitSlices.empty()) {
477 assert(P.SJ == RHS.P.SJ &&
478 "Same set of slices formed two different sized partitions!");
479 assert(P.SplitSlices.size() == RHS.P.SplitSlices.size() &&
480 "Same slice position with differently sized non-empty split "
487 partition_iterator &operator++() {
492 Partition &operator*() { return P; }
495 /// \brief A forward range over the partitions of the alloca's slices.
497 /// This accesses an iterator range over the partitions of the alloca's
498 /// slices. It computes these partitions on the fly based on the overlapping
499 /// offsets of the slices and the ability to split them. It will visit "empty"
500 /// partitions to cover regions of the alloca only accessed via split
502 iterator_range<partition_iterator> partitions() {
503 return make_range(partition_iterator(begin(), end()),
504 partition_iterator(end(), end()));
507 /// \brief Access the dead users for this alloca.
508 ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
510 /// \brief Access the dead operands referring to this alloca.
512 /// These are operands which have cannot actually be used to refer to the
513 /// alloca as they are outside its range and the user doesn't correct for
514 /// that. These mostly consist of PHI node inputs and the like which we just
515 /// need to replace with undef.
516 ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
518 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
519 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
520 void printSlice(raw_ostream &OS, const_iterator I,
521 StringRef Indent = " ") const;
522 void printUse(raw_ostream &OS, const_iterator I,
523 StringRef Indent = " ") const;
524 void print(raw_ostream &OS) const;
525 void dump(const_iterator I) const;
530 template <typename DerivedT, typename RetT = void> class BuilderBase;
532 friend class AllocaSlices::SliceBuilder;
534 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
535 /// \brief Handle to alloca instruction to simplify method interfaces.
539 /// \brief The instruction responsible for this alloca not having a known set
542 /// When an instruction (potentially) escapes the pointer to the alloca, we
543 /// store a pointer to that here and abort trying to form slices of the
544 /// alloca. This will be null if the alloca slices are analyzed successfully.
545 Instruction *PointerEscapingInstr;
547 /// \brief The slices of the alloca.
549 /// We store a vector of the slices formed by uses of the alloca here. This
550 /// vector is sorted by increasing begin offset, and then the unsplittable
551 /// slices before the splittable ones. See the Slice inner class for more
553 SmallVector<Slice, 8> Slices;
555 /// \brief Instructions which will become dead if we rewrite the alloca.
557 /// Note that these are not separated by slice. This is because we expect an
558 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
559 /// all these instructions can simply be removed and replaced with undef as
560 /// they come from outside of the allocated space.
561 SmallVector<Instruction *, 8> DeadUsers;
563 /// \brief Operands which will become dead if we rewrite the alloca.
565 /// These are operands that in their particular use can be replaced with
566 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
567 /// to PHI nodes and the like. They aren't entirely dead (there might be
568 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
569 /// want to swap this particular input for undef to simplify the use lists of
571 SmallVector<Use *, 8> DeadOperands;
575 static Value *foldSelectInst(SelectInst &SI) {
576 // If the condition being selected on is a constant or the same value is
577 // being selected between, fold the select. Yes this does (rarely) happen
579 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
580 return SI.getOperand(1 + CI->isZero());
581 if (SI.getOperand(1) == SI.getOperand(2))
582 return SI.getOperand(1);
587 /// \brief A helper that folds a PHI node or a select.
588 static Value *foldPHINodeOrSelectInst(Instruction &I) {
589 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
590 // If PN merges together the same value, return that value.
591 return PN->hasConstantValue();
593 return foldSelectInst(cast<SelectInst>(I));
596 /// \brief Builder for the alloca slices.
598 /// This class builds a set of alloca slices by recursively visiting the uses
599 /// of an alloca and making a slice for each load and store at each offset.
600 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
601 friend class PtrUseVisitor<SliceBuilder>;
602 friend class InstVisitor<SliceBuilder>;
603 typedef PtrUseVisitor<SliceBuilder> Base;
605 const uint64_t AllocSize;
608 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
609 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
611 /// \brief Set to de-duplicate dead instructions found in the use walk.
612 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
615 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
616 : PtrUseVisitor<SliceBuilder>(DL),
617 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), AS(AS) {}
620 void markAsDead(Instruction &I) {
621 if (VisitedDeadInsts.insert(&I).second)
622 AS.DeadUsers.push_back(&I);
625 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
626 bool IsSplittable = false) {
627 // Completely skip uses which have a zero size or start either before or
628 // past the end of the allocation.
629 if (Size == 0 || Offset.uge(AllocSize)) {
630 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
631 << " which has zero size or starts outside of the "
632 << AllocSize << " byte alloca:\n"
633 << " alloca: " << AS.AI << "\n"
634 << " use: " << I << "\n");
635 return markAsDead(I);
638 uint64_t BeginOffset = Offset.getZExtValue();
639 uint64_t EndOffset = BeginOffset + Size;
641 // Clamp the end offset to the end of the allocation. Note that this is
642 // formulated to handle even the case where "BeginOffset + Size" overflows.
643 // This may appear superficially to be something we could ignore entirely,
644 // but that is not so! There may be widened loads or PHI-node uses where
645 // some instructions are dead but not others. We can't completely ignore
646 // them, and so have to record at least the information here.
647 assert(AllocSize >= BeginOffset); // Established above.
648 if (Size > AllocSize - BeginOffset) {
649 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
650 << " to remain within the " << AllocSize << " byte alloca:\n"
651 << " alloca: " << AS.AI << "\n"
652 << " use: " << I << "\n");
653 EndOffset = AllocSize;
656 AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
659 void visitBitCastInst(BitCastInst &BC) {
661 return markAsDead(BC);
663 return Base::visitBitCastInst(BC);
666 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
667 if (GEPI.use_empty())
668 return markAsDead(GEPI);
670 if (SROAStrictInbounds && GEPI.isInBounds()) {
671 // FIXME: This is a manually un-factored variant of the basic code inside
672 // of GEPs with checking of the inbounds invariant specified in the
673 // langref in a very strict sense. If we ever want to enable
674 // SROAStrictInbounds, this code should be factored cleanly into
675 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
676 // by writing out the code here where we have tho underlying allocation
677 // size readily available.
678 APInt GEPOffset = Offset;
679 for (gep_type_iterator GTI = gep_type_begin(GEPI),
680 GTE = gep_type_end(GEPI);
682 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
686 // Handle a struct index, which adds its field offset to the pointer.
687 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
688 unsigned ElementIdx = OpC->getZExtValue();
689 const StructLayout *SL = DL.getStructLayout(STy);
691 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
693 // For array or vector indices, scale the index by the size of the
695 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
696 GEPOffset += Index * APInt(Offset.getBitWidth(),
697 DL.getTypeAllocSize(GTI.getIndexedType()));
700 // If this index has computed an intermediate pointer which is not
701 // inbounds, then the result of the GEP is a poison value and we can
702 // delete it and all uses.
703 if (GEPOffset.ugt(AllocSize))
704 return markAsDead(GEPI);
708 return Base::visitGetElementPtrInst(GEPI);
711 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
712 uint64_t Size, bool IsVolatile) {
713 // We allow splitting of loads and stores where the type is an integer type
714 // and cover the entire alloca. This prevents us from splitting over
716 // FIXME: In the great blue eventually, we should eagerly split all integer
717 // loads and stores, and then have a separate step that merges adjacent
718 // alloca partitions into a single partition suitable for integer widening.
719 // Or we should skip the merge step and rely on GVN and other passes to
720 // merge adjacent loads and stores that survive mem2reg.
722 Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize;
724 insertUse(I, Offset, Size, IsSplittable);
727 void visitLoadInst(LoadInst &LI) {
728 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
729 "All simple FCA loads should have been pre-split");
732 return PI.setAborted(&LI);
734 uint64_t Size = DL.getTypeStoreSize(LI.getType());
735 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
738 void visitStoreInst(StoreInst &SI) {
739 Value *ValOp = SI.getValueOperand();
741 return PI.setEscapedAndAborted(&SI);
743 return PI.setAborted(&SI);
745 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
747 // If this memory access can be shown to *statically* extend outside the
748 // bounds of of the allocation, it's behavior is undefined, so simply
749 // ignore it. Note that this is more strict than the generic clamping
750 // behavior of insertUse. We also try to handle cases which might run the
752 // FIXME: We should instead consider the pointer to have escaped if this
753 // function is being instrumented for addressing bugs or race conditions.
754 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
755 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
756 << " which extends past the end of the " << AllocSize
758 << " alloca: " << AS.AI << "\n"
759 << " use: " << SI << "\n");
760 return markAsDead(SI);
763 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
764 "All simple FCA stores should have been pre-split");
765 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
768 void visitMemSetInst(MemSetInst &II) {
769 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
770 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
771 if ((Length && Length->getValue() == 0) ||
772 (IsOffsetKnown && Offset.uge(AllocSize)))
773 // Zero-length mem transfer intrinsics can be ignored entirely.
774 return markAsDead(II);
777 return PI.setAborted(&II);
779 insertUse(II, Offset, Length ? Length->getLimitedValue()
780 : AllocSize - Offset.getLimitedValue(),
784 void visitMemTransferInst(MemTransferInst &II) {
785 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
786 if (Length && Length->getValue() == 0)
787 // Zero-length mem transfer intrinsics can be ignored entirely.
788 return markAsDead(II);
790 // Because we can visit these intrinsics twice, also check to see if the
791 // first time marked this instruction as dead. If so, skip it.
792 if (VisitedDeadInsts.count(&II))
796 return PI.setAborted(&II);
798 // This side of the transfer is completely out-of-bounds, and so we can
799 // nuke the entire transfer. However, we also need to nuke the other side
800 // if already added to our partitions.
801 // FIXME: Yet another place we really should bypass this when
802 // instrumenting for ASan.
803 if (Offset.uge(AllocSize)) {
804 SmallDenseMap<Instruction *, unsigned>::iterator MTPI =
805 MemTransferSliceMap.find(&II);
806 if (MTPI != MemTransferSliceMap.end())
807 AS.Slices[MTPI->second].kill();
808 return markAsDead(II);
811 uint64_t RawOffset = Offset.getLimitedValue();
812 uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
814 // Check for the special case where the same exact value is used for both
816 if (*U == II.getRawDest() && *U == II.getRawSource()) {
817 // For non-volatile transfers this is a no-op.
818 if (!II.isVolatile())
819 return markAsDead(II);
821 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
824 // If we have seen both source and destination for a mem transfer, then
825 // they both point to the same alloca.
827 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
828 std::tie(MTPI, Inserted) =
829 MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
830 unsigned PrevIdx = MTPI->second;
832 Slice &PrevP = AS.Slices[PrevIdx];
834 // Check if the begin offsets match and this is a non-volatile transfer.
835 // In that case, we can completely elide the transfer.
836 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
838 return markAsDead(II);
841 // Otherwise we have an offset transfer within the same alloca. We can't
843 PrevP.makeUnsplittable();
846 // Insert the use now that we've fixed up the splittable nature.
847 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
849 // Check that we ended up with a valid index in the map.
850 assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&
851 "Map index doesn't point back to a slice with this user.");
854 // Disable SRoA for any intrinsics except for lifetime invariants.
855 // FIXME: What about debug intrinsics? This matches old behavior, but
856 // doesn't make sense.
857 void visitIntrinsicInst(IntrinsicInst &II) {
859 return PI.setAborted(&II);
861 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
862 II.getIntrinsicID() == Intrinsic::lifetime_end) {
863 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
864 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
865 Length->getLimitedValue());
866 insertUse(II, Offset, Size, true);
870 Base::visitIntrinsicInst(II);
873 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
874 // We consider any PHI or select that results in a direct load or store of
875 // the same offset to be a viable use for slicing purposes. These uses
876 // are considered unsplittable and the size is the maximum loaded or stored
878 SmallPtrSet<Instruction *, 4> Visited;
879 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
880 Visited.insert(Root);
881 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
882 // If there are no loads or stores, the access is dead. We mark that as
883 // a size zero access.
886 Instruction *I, *UsedI;
887 std::tie(UsedI, I) = Uses.pop_back_val();
889 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
890 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
893 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
894 Value *Op = SI->getOperand(0);
897 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
901 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
902 if (!GEP->hasAllZeroIndices())
904 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
905 !isa<SelectInst>(I)) {
909 for (User *U : I->users())
910 if (Visited.insert(cast<Instruction>(U)).second)
911 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
912 } while (!Uses.empty());
917 void visitPHINodeOrSelectInst(Instruction &I) {
918 assert(isa<PHINode>(I) || isa<SelectInst>(I));
920 return markAsDead(I);
922 // TODO: We could use SimplifyInstruction here to fold PHINodes and
923 // SelectInsts. However, doing so requires to change the current
924 // dead-operand-tracking mechanism. For instance, suppose neither loading
925 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
926 // trap either. However, if we simply replace %U with undef using the
927 // current dead-operand-tracking mechanism, "load (select undef, undef,
928 // %other)" may trap because the select may return the first operand
930 if (Value *Result = foldPHINodeOrSelectInst(I)) {
932 // If the result of the constant fold will be the pointer, recurse
933 // through the PHI/select as if we had RAUW'ed it.
936 // Otherwise the operand to the PHI/select is dead, and we can replace
938 AS.DeadOperands.push_back(U);
944 return PI.setAborted(&I);
946 // See if we already have computed info on this node.
947 uint64_t &Size = PHIOrSelectSizes[&I];
949 // This is a new PHI/Select, check for an unsafe use of it.
950 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
951 return PI.setAborted(UnsafeI);
954 // For PHI and select operands outside the alloca, we can't nuke the entire
955 // phi or select -- the other side might still be relevant, so we special
956 // case them here and use a separate structure to track the operands
957 // themselves which should be replaced with undef.
958 // FIXME: This should instead be escaped in the event we're instrumenting
959 // for address sanitization.
960 if (Offset.uge(AllocSize)) {
961 AS.DeadOperands.push_back(U);
965 insertUse(I, Offset, Size);
968 void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
970 void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
972 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
973 void visitInstruction(Instruction &I) { PI.setAborted(&I); }
976 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
978 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
981 PointerEscapingInstr(nullptr) {
982 SliceBuilder PB(DL, AI, *this);
983 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
984 if (PtrI.isEscaped() || PtrI.isAborted()) {
985 // FIXME: We should sink the escape vs. abort info into the caller nicely,
986 // possibly by just storing the PtrInfo in the AllocaSlices.
987 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
988 : PtrI.getAbortingInst();
989 assert(PointerEscapingInstr && "Did not track a bad instruction");
993 Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
999 #if __cplusplus >= 201103L && !defined(NDEBUG)
1000 if (SROARandomShuffleSlices) {
1001 std::mt19937 MT(static_cast<unsigned>(sys::TimeValue::now().msec()));
1002 std::shuffle(Slices.begin(), Slices.end(), MT);
1006 // Sort the uses. This arranges for the offsets to be in ascending order,
1007 // and the sizes to be in descending order.
1008 std::sort(Slices.begin(), Slices.end());
1011 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1013 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
1014 StringRef Indent) const {
1015 printSlice(OS, I, Indent);
1016 printUse(OS, I, Indent);
1019 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
1020 StringRef Indent) const {
1021 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
1022 << " slice #" << (I - begin())
1023 << (I->isSplittable() ? " (splittable)" : "") << "\n";
1026 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
1027 StringRef Indent) const {
1028 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
1031 void AllocaSlices::print(raw_ostream &OS) const {
1032 if (PointerEscapingInstr) {
1033 OS << "Can't analyze slices for alloca: " << AI << "\n"
1034 << " A pointer to this alloca escaped by:\n"
1035 << " " << *PointerEscapingInstr << "\n";
1039 OS << "Slices of alloca: " << AI << "\n";
1040 for (const_iterator I = begin(), E = end(); I != E; ++I)
1044 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
1047 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
1049 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1052 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1054 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1055 /// the loads and stores of an alloca instruction, as well as updating its
1056 /// debug information. This is used when a domtree is unavailable and thus
1057 /// mem2reg in its full form can't be used to handle promotion of allocas to
1059 class AllocaPromoter : public LoadAndStorePromoter {
1063 SmallVector<DbgDeclareInst *, 4> DDIs;
1064 SmallVector<DbgValueInst *, 4> DVIs;
1067 AllocaPromoter(const SmallVectorImpl<Instruction *> &Insts, SSAUpdater &S,
1068 AllocaInst &AI, DIBuilder &DIB)
1069 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1071 void run(const SmallVectorImpl<Instruction *> &Insts) {
1072 // Retain the debug information attached to the alloca for use when
1073 // rewriting loads and stores.
1074 if (auto *L = LocalAsMetadata::getIfExists(&AI)) {
1075 if (auto *DebugNode = MetadataAsValue::getIfExists(AI.getContext(), L)) {
1076 for (User *U : DebugNode->users())
1077 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
1078 DDIs.push_back(DDI);
1079 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U))
1080 DVIs.push_back(DVI);
1084 LoadAndStorePromoter::run(Insts);
1086 // While we have the debug information, clear it off of the alloca. The
1087 // caller takes care of deleting the alloca.
1088 while (!DDIs.empty())
1089 DDIs.pop_back_val()->eraseFromParent();
1090 while (!DVIs.empty())
1091 DVIs.pop_back_val()->eraseFromParent();
1095 isInstInList(Instruction *I,
1096 const SmallVectorImpl<Instruction *> &Insts) const override {
1098 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1099 Ptr = LI->getOperand(0);
1101 Ptr = cast<StoreInst>(I)->getPointerOperand();
1103 // Only used to detect cycles, which will be rare and quickly found as
1104 // we're walking up a chain of defs rather than down through uses.
1105 SmallPtrSet<Value *, 4> Visited;
1111 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr))
1112 Ptr = BCI->getOperand(0);
1113 else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr))
1114 Ptr = GEPI->getPointerOperand();
1118 } while (Visited.insert(Ptr).second);
1123 void updateDebugInfo(Instruction *Inst) const override {
1124 for (DbgDeclareInst *DDI : DDIs)
1125 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1126 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1127 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1128 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1129 for (DbgValueInst *DVI : DVIs) {
1130 Value *Arg = nullptr;
1131 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1132 // If an argument is zero extended then use argument directly. The ZExt
1133 // may be zapped by an optimization pass in future.
1134 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1135 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1136 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1137 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1139 Arg = SI->getValueOperand();
1140 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1141 Arg = LI->getPointerOperand();
1145 Instruction *DbgVal =
1146 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1147 DIExpression(DVI->getExpression()), Inst);
1148 DbgVal->setDebugLoc(DVI->getDebugLoc());
1152 } // end anon namespace
1155 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1157 /// This pass takes allocations which can be completely analyzed (that is, they
1158 /// don't escape) and tries to turn them into scalar SSA values. There are
1159 /// a few steps to this process.
1161 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1162 /// are used to try to split them into smaller allocations, ideally of
1163 /// a single scalar data type. It will split up memcpy and memset accesses
1164 /// as necessary and try to isolate individual scalar accesses.
1165 /// 2) It will transform accesses into forms which are suitable for SSA value
1166 /// promotion. This can be replacing a memset with a scalar store of an
1167 /// integer value, or it can involve speculating operations on a PHI or
1168 /// select to be a PHI or select of the results.
1169 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1170 /// onto insert and extract operations on a vector value, and convert them to
1171 /// this form. By doing so, it will enable promotion of vector aggregates to
1172 /// SSA vector values.
1173 class SROA : public FunctionPass {
1174 const bool RequiresDomTree;
1177 const DataLayout *DL;
1179 AssumptionTracker *AT;
1181 /// \brief Worklist of alloca instructions to simplify.
1183 /// Each alloca in the function is added to this. Each new alloca formed gets
1184 /// added to it as well to recursively simplify unless that alloca can be
1185 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1186 /// the one being actively rewritten, we add it back onto the list if not
1187 /// already present to ensure it is re-visited.
1188 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16>> Worklist;
1190 /// \brief A collection of instructions to delete.
1191 /// We try to batch deletions to simplify code and make things a bit more
1193 SetVector<Instruction *, SmallVector<Instruction *, 8>> DeadInsts;
1195 /// \brief Post-promotion worklist.
1197 /// Sometimes we discover an alloca which has a high probability of becoming
1198 /// viable for SROA after a round of promotion takes place. In those cases,
1199 /// the alloca is enqueued here for re-processing.
1201 /// Note that we have to be very careful to clear allocas out of this list in
1202 /// the event they are deleted.
1203 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16>> PostPromotionWorklist;
1205 /// \brief A collection of alloca instructions we can directly promote.
1206 std::vector<AllocaInst *> PromotableAllocas;
1208 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
1210 /// All of these PHIs have been checked for the safety of speculation and by
1211 /// being speculated will allow promoting allocas currently in the promotable
1213 SetVector<PHINode *, SmallVector<PHINode *, 2>> SpeculatablePHIs;
1215 /// \brief A worklist of select instructions to speculate prior to promoting
1218 /// All of these select instructions have been checked for the safety of
1219 /// speculation and by being speculated will allow promoting allocas
1220 /// currently in the promotable queue.
1221 SetVector<SelectInst *, SmallVector<SelectInst *, 2>> SpeculatableSelects;
1224 SROA(bool RequiresDomTree = true)
1225 : FunctionPass(ID), RequiresDomTree(RequiresDomTree), C(nullptr),
1226 DL(nullptr), DT(nullptr) {
1227 initializeSROAPass(*PassRegistry::getPassRegistry());
1229 bool runOnFunction(Function &F) override;
1230 void getAnalysisUsage(AnalysisUsage &AU) const override;
1232 const char *getPassName() const override { return "SROA"; }
1236 friend class PHIOrSelectSpeculator;
1237 friend class AllocaSliceRewriter;
1239 bool rewritePartition(AllocaInst &AI, AllocaSlices &AS,
1240 AllocaSlices::Partition &P);
1241 bool splitAlloca(AllocaInst &AI, AllocaSlices &AS);
1242 bool runOnAlloca(AllocaInst &AI);
1243 void clobberUse(Use &U);
1244 void deleteDeadInstructions(SmallPtrSetImpl<AllocaInst *> &DeletedAllocas);
1245 bool promoteAllocas(Function &F);
1251 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1252 return new SROA(RequiresDomTree);
1255 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates", false,
1257 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
1258 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1259 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates", false,
1262 /// Walk the range of a partitioning looking for a common type to cover this
1263 /// sequence of slices.
1264 static Type *findCommonType(AllocaSlices::const_iterator B,
1265 AllocaSlices::const_iterator E,
1266 uint64_t EndOffset) {
1268 bool TyIsCommon = true;
1269 IntegerType *ITy = nullptr;
1271 // Note that we need to look at *every* alloca slice's Use to ensure we
1272 // always get consistent results regardless of the order of slices.
1273 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1274 Use *U = I->getUse();
1275 if (isa<IntrinsicInst>(*U->getUser()))
1277 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1280 Type *UserTy = nullptr;
1281 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1282 UserTy = LI->getType();
1283 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1284 UserTy = SI->getValueOperand()->getType();
1287 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1288 // If the type is larger than the partition, skip it. We only encounter
1289 // this for split integer operations where we want to use the type of the
1290 // entity causing the split. Also skip if the type is not a byte width
1292 if (UserITy->getBitWidth() % 8 != 0 ||
1293 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1296 // Track the largest bitwidth integer type used in this way in case there
1297 // is no common type.
1298 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1302 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1303 // depend on types skipped above.
1304 if (!UserTy || (Ty && Ty != UserTy))
1305 TyIsCommon = false; // Give up on anything but an iN type.
1310 return TyIsCommon ? Ty : ITy;
1313 /// PHI instructions that use an alloca and are subsequently loaded can be
1314 /// rewritten to load both input pointers in the pred blocks and then PHI the
1315 /// results, allowing the load of the alloca to be promoted.
1317 /// %P2 = phi [i32* %Alloca, i32* %Other]
1318 /// %V = load i32* %P2
1320 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1322 /// %V2 = load i32* %Other
1324 /// %V = phi [i32 %V1, i32 %V2]
1326 /// We can do this to a select if its only uses are loads and if the operands
1327 /// to the select can be loaded unconditionally.
1329 /// FIXME: This should be hoisted into a generic utility, likely in
1330 /// Transforms/Util/Local.h
1331 static bool isSafePHIToSpeculate(PHINode &PN, const DataLayout *DL = nullptr) {
1332 // For now, we can only do this promotion if the load is in the same block
1333 // as the PHI, and if there are no stores between the phi and load.
1334 // TODO: Allow recursive phi users.
1335 // TODO: Allow stores.
1336 BasicBlock *BB = PN.getParent();
1337 unsigned MaxAlign = 0;
1338 bool HaveLoad = false;
1339 for (User *U : PN.users()) {
1340 LoadInst *LI = dyn_cast<LoadInst>(U);
1341 if (!LI || !LI->isSimple())
1344 // For now we only allow loads in the same block as the PHI. This is
1345 // a common case that happens when instcombine merges two loads through
1347 if (LI->getParent() != BB)
1350 // Ensure that there are no instructions between the PHI and the load that
1352 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1353 if (BBI->mayWriteToMemory())
1356 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1363 // We can only transform this if it is safe to push the loads into the
1364 // predecessor blocks. The only thing to watch out for is that we can't put
1365 // a possibly trapping load in the predecessor if it is a critical edge.
1366 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1367 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1368 Value *InVal = PN.getIncomingValue(Idx);
1370 // If the value is produced by the terminator of the predecessor (an
1371 // invoke) or it has side-effects, there is no valid place to put a load
1372 // in the predecessor.
1373 if (TI == InVal || TI->mayHaveSideEffects())
1376 // If the predecessor has a single successor, then the edge isn't
1378 if (TI->getNumSuccessors() == 1)
1381 // If this pointer is always safe to load, or if we can prove that there
1382 // is already a load in the block, then we can move the load to the pred
1384 if (InVal->isDereferenceablePointer(DL) ||
1385 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, DL))
1394 static void speculatePHINodeLoads(PHINode &PN) {
1395 DEBUG(dbgs() << " original: " << PN << "\n");
1397 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1398 IRBuilderTy PHIBuilder(&PN);
1399 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1400 PN.getName() + ".sroa.speculated");
1402 // Get the AA tags and alignment to use from one of the loads. It doesn't
1403 // matter which one we get and if any differ.
1404 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1407 SomeLoad->getAAMetadata(AATags);
1408 unsigned Align = SomeLoad->getAlignment();
1410 // Rewrite all loads of the PN to use the new PHI.
1411 while (!PN.use_empty()) {
1412 LoadInst *LI = cast<LoadInst>(PN.user_back());
1413 LI->replaceAllUsesWith(NewPN);
1414 LI->eraseFromParent();
1417 // Inject loads into all of the pred blocks.
1418 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1419 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1420 TerminatorInst *TI = Pred->getTerminator();
1421 Value *InVal = PN.getIncomingValue(Idx);
1422 IRBuilderTy PredBuilder(TI);
1424 LoadInst *Load = PredBuilder.CreateLoad(
1425 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1426 ++NumLoadsSpeculated;
1427 Load->setAlignment(Align);
1429 Load->setAAMetadata(AATags);
1430 NewPN->addIncoming(Load, Pred);
1433 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1434 PN.eraseFromParent();
1437 /// Select instructions that use an alloca and are subsequently loaded can be
1438 /// rewritten to load both input pointers and then select between the result,
1439 /// allowing the load of the alloca to be promoted.
1441 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1442 /// %V = load i32* %P2
1444 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1445 /// %V2 = load i32* %Other
1446 /// %V = select i1 %cond, i32 %V1, i32 %V2
1448 /// We can do this to a select if its only uses are loads and if the operand
1449 /// to the select can be loaded unconditionally.
1450 static bool isSafeSelectToSpeculate(SelectInst &SI,
1451 const DataLayout *DL = nullptr) {
1452 Value *TValue = SI.getTrueValue();
1453 Value *FValue = SI.getFalseValue();
1454 bool TDerefable = TValue->isDereferenceablePointer(DL);
1455 bool FDerefable = FValue->isDereferenceablePointer(DL);
1457 for (User *U : SI.users()) {
1458 LoadInst *LI = dyn_cast<LoadInst>(U);
1459 if (!LI || !LI->isSimple())
1462 // Both operands to the select need to be dereferencable, either
1463 // absolutely (e.g. allocas) or at this point because we can see other
1466 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment(), DL))
1469 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment(), DL))
1476 static void speculateSelectInstLoads(SelectInst &SI) {
1477 DEBUG(dbgs() << " original: " << SI << "\n");
1479 IRBuilderTy IRB(&SI);
1480 Value *TV = SI.getTrueValue();
1481 Value *FV = SI.getFalseValue();
1482 // Replace the loads of the select with a select of two loads.
1483 while (!SI.use_empty()) {
1484 LoadInst *LI = cast<LoadInst>(SI.user_back());
1485 assert(LI->isSimple() && "We only speculate simple loads");
1487 IRB.SetInsertPoint(LI);
1489 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1491 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1492 NumLoadsSpeculated += 2;
1494 // Transfer alignment and AA info if present.
1495 TL->setAlignment(LI->getAlignment());
1496 FL->setAlignment(LI->getAlignment());
1499 LI->getAAMetadata(Tags);
1501 TL->setAAMetadata(Tags);
1502 FL->setAAMetadata(Tags);
1505 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1506 LI->getName() + ".sroa.speculated");
1508 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1509 LI->replaceAllUsesWith(V);
1510 LI->eraseFromParent();
1512 SI.eraseFromParent();
1515 /// \brief Build a GEP out of a base pointer and indices.
1517 /// This will return the BasePtr if that is valid, or build a new GEP
1518 /// instruction using the IRBuilder if GEP-ing is needed.
1519 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1520 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1521 if (Indices.empty())
1524 // A single zero index is a no-op, so check for this and avoid building a GEP
1526 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1529 return IRB.CreateInBoundsGEP(BasePtr, Indices, NamePrefix + "sroa_idx");
1532 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1533 /// TargetTy without changing the offset of the pointer.
1535 /// This routine assumes we've already established a properly offset GEP with
1536 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1537 /// zero-indices down through type layers until we find one the same as
1538 /// TargetTy. If we can't find one with the same type, we at least try to use
1539 /// one with the same size. If none of that works, we just produce the GEP as
1540 /// indicated by Indices to have the correct offset.
1541 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1542 Value *BasePtr, Type *Ty, Type *TargetTy,
1543 SmallVectorImpl<Value *> &Indices,
1546 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1548 // Pointer size to use for the indices.
1549 unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1551 // See if we can descend into a struct and locate a field with the correct
1553 unsigned NumLayers = 0;
1554 Type *ElementTy = Ty;
1556 if (ElementTy->isPointerTy())
1559 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1560 ElementTy = ArrayTy->getElementType();
1561 Indices.push_back(IRB.getIntN(PtrSize, 0));
1562 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1563 ElementTy = VectorTy->getElementType();
1564 Indices.push_back(IRB.getInt32(0));
1565 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1566 if (STy->element_begin() == STy->element_end())
1567 break; // Nothing left to descend into.
1568 ElementTy = *STy->element_begin();
1569 Indices.push_back(IRB.getInt32(0));
1574 } while (ElementTy != TargetTy);
1575 if (ElementTy != TargetTy)
1576 Indices.erase(Indices.end() - NumLayers, Indices.end());
1578 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1581 /// \brief Recursively compute indices for a natural GEP.
1583 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1584 /// element types adding appropriate indices for the GEP.
1585 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1586 Value *Ptr, Type *Ty, APInt &Offset,
1588 SmallVectorImpl<Value *> &Indices,
1591 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1594 // We can't recurse through pointer types.
1595 if (Ty->isPointerTy())
1598 // We try to analyze GEPs over vectors here, but note that these GEPs are
1599 // extremely poorly defined currently. The long-term goal is to remove GEPing
1600 // over a vector from the IR completely.
1601 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1602 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1603 if (ElementSizeInBits % 8 != 0) {
1604 // GEPs over non-multiple of 8 size vector elements are invalid.
1607 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1608 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1609 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1611 Offset -= NumSkippedElements * ElementSize;
1612 Indices.push_back(IRB.getInt(NumSkippedElements));
1613 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1614 Offset, TargetTy, Indices, NamePrefix);
1617 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1618 Type *ElementTy = ArrTy->getElementType();
1619 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1620 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1621 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1624 Offset -= NumSkippedElements * ElementSize;
1625 Indices.push_back(IRB.getInt(NumSkippedElements));
1626 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1627 Indices, NamePrefix);
1630 StructType *STy = dyn_cast<StructType>(Ty);
1634 const StructLayout *SL = DL.getStructLayout(STy);
1635 uint64_t StructOffset = Offset.getZExtValue();
1636 if (StructOffset >= SL->getSizeInBytes())
1638 unsigned Index = SL->getElementContainingOffset(StructOffset);
1639 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1640 Type *ElementTy = STy->getElementType(Index);
1641 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1642 return nullptr; // The offset points into alignment padding.
1644 Indices.push_back(IRB.getInt32(Index));
1645 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1646 Indices, NamePrefix);
1649 /// \brief Get a natural GEP from a base pointer to a particular offset and
1650 /// resulting in a particular type.
1652 /// The goal is to produce a "natural" looking GEP that works with the existing
1653 /// composite types to arrive at the appropriate offset and element type for
1654 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1655 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1656 /// Indices, and setting Ty to the result subtype.
1658 /// If no natural GEP can be constructed, this function returns null.
1659 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1660 Value *Ptr, APInt Offset, Type *TargetTy,
1661 SmallVectorImpl<Value *> &Indices,
1663 PointerType *Ty = cast<PointerType>(Ptr->getType());
1665 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1667 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1670 Type *ElementTy = Ty->getElementType();
1671 if (!ElementTy->isSized())
1672 return nullptr; // We can't GEP through an unsized element.
1673 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1674 if (ElementSize == 0)
1675 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1676 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1678 Offset -= NumSkippedElements * ElementSize;
1679 Indices.push_back(IRB.getInt(NumSkippedElements));
1680 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1681 Indices, NamePrefix);
1684 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1685 /// resulting pointer has PointerTy.
1687 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1688 /// and produces the pointer type desired. Where it cannot, it will try to use
1689 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1690 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1691 /// bitcast to the type.
1693 /// The strategy for finding the more natural GEPs is to peel off layers of the
1694 /// pointer, walking back through bit casts and GEPs, searching for a base
1695 /// pointer from which we can compute a natural GEP with the desired
1696 /// properties. The algorithm tries to fold as many constant indices into
1697 /// a single GEP as possible, thus making each GEP more independent of the
1698 /// surrounding code.
1699 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1700 APInt Offset, Type *PointerTy, Twine NamePrefix) {
1701 // Even though we don't look through PHI nodes, we could be called on an
1702 // instruction in an unreachable block, which may be on a cycle.
1703 SmallPtrSet<Value *, 4> Visited;
1704 Visited.insert(Ptr);
1705 SmallVector<Value *, 4> Indices;
1707 // We may end up computing an offset pointer that has the wrong type. If we
1708 // never are able to compute one directly that has the correct type, we'll
1709 // fall back to it, so keep it around here.
1710 Value *OffsetPtr = nullptr;
1712 // Remember any i8 pointer we come across to re-use if we need to do a raw
1714 Value *Int8Ptr = nullptr;
1715 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1717 Type *TargetTy = PointerTy->getPointerElementType();
1720 // First fold any existing GEPs into the offset.
1721 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1722 APInt GEPOffset(Offset.getBitWidth(), 0);
1723 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1725 Offset += GEPOffset;
1726 Ptr = GEP->getPointerOperand();
1727 if (!Visited.insert(Ptr).second)
1731 // See if we can perform a natural GEP here.
1733 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1734 Indices, NamePrefix)) {
1735 if (P->getType() == PointerTy) {
1736 // Zap any offset pointer that we ended up computing in previous rounds.
1737 if (OffsetPtr && OffsetPtr->use_empty())
1738 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1739 I->eraseFromParent();
1747 // Stash this pointer if we've found an i8*.
1748 if (Ptr->getType()->isIntegerTy(8)) {
1750 Int8PtrOffset = Offset;
1753 // Peel off a layer of the pointer and update the offset appropriately.
1754 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1755 Ptr = cast<Operator>(Ptr)->getOperand(0);
1756 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1757 if (GA->mayBeOverridden())
1759 Ptr = GA->getAliasee();
1763 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1764 } while (Visited.insert(Ptr).second);
1768 Int8Ptr = IRB.CreateBitCast(
1769 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1770 NamePrefix + "sroa_raw_cast");
1771 Int8PtrOffset = Offset;
1774 OffsetPtr = Int8PtrOffset == 0
1776 : IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1777 NamePrefix + "sroa_raw_idx");
1781 // On the off chance we were targeting i8*, guard the bitcast here.
1782 if (Ptr->getType() != PointerTy)
1783 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1788 /// \brief Test whether we can convert a value from the old to the new type.
1790 /// This predicate should be used to guard calls to convertValue in order to
1791 /// ensure that we only try to convert viable values. The strategy is that we
1792 /// will peel off single element struct and array wrappings to get to an
1793 /// underlying value, and convert that value.
1794 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1797 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1798 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1799 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1801 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1803 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1806 // We can convert pointers to integers and vice-versa. Same for vectors
1807 // of pointers and integers.
1808 OldTy = OldTy->getScalarType();
1809 NewTy = NewTy->getScalarType();
1810 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1811 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1813 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1821 /// \brief Generic routine to convert an SSA value to a value of a different
1824 /// This will try various different casting techniques, such as bitcasts,
1825 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1826 /// two types for viability with this routine.
1827 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1829 Type *OldTy = V->getType();
1830 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1835 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1836 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1837 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1838 return IRB.CreateZExt(V, NewITy);
1840 // See if we need inttoptr for this type pair. A cast involving both scalars
1841 // and vectors requires and additional bitcast.
1842 if (OldTy->getScalarType()->isIntegerTy() &&
1843 NewTy->getScalarType()->isPointerTy()) {
1844 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1845 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1846 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1849 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1850 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1851 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1854 return IRB.CreateIntToPtr(V, NewTy);
1857 // See if we need ptrtoint for this type pair. A cast involving both scalars
1858 // and vectors requires and additional bitcast.
1859 if (OldTy->getScalarType()->isPointerTy() &&
1860 NewTy->getScalarType()->isIntegerTy()) {
1861 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1862 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1863 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1866 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1867 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1868 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1871 return IRB.CreatePtrToInt(V, NewTy);
1874 return IRB.CreateBitCast(V, NewTy);
1877 /// \brief Test whether the given slice use can be promoted to a vector.
1879 /// This function is called to test each entry in a partioning which is slated
1880 /// for a single slice.
1882 isVectorPromotionViableForSlice(const DataLayout &DL, uint64_t SliceBeginOffset,
1883 uint64_t SliceEndOffset, VectorType *Ty,
1884 uint64_t ElementSize, const Slice &S) {
1885 // First validate the slice offsets.
1886 uint64_t BeginOffset =
1887 std::max(S.beginOffset(), SliceBeginOffset) - SliceBeginOffset;
1888 uint64_t BeginIndex = BeginOffset / ElementSize;
1889 if (BeginIndex * ElementSize != BeginOffset ||
1890 BeginIndex >= Ty->getNumElements())
1892 uint64_t EndOffset =
1893 std::min(S.endOffset(), SliceEndOffset) - SliceBeginOffset;
1894 uint64_t EndIndex = EndOffset / ElementSize;
1895 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1898 assert(EndIndex > BeginIndex && "Empty vector!");
1899 uint64_t NumElements = EndIndex - BeginIndex;
1900 Type *SliceTy = (NumElements == 1)
1901 ? Ty->getElementType()
1902 : VectorType::get(Ty->getElementType(), NumElements);
1905 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1907 Use *U = S.getUse();
1909 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1910 if (MI->isVolatile())
1912 if (!S.isSplittable())
1913 return false; // Skip any unsplittable intrinsics.
1914 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1915 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1916 II->getIntrinsicID() != Intrinsic::lifetime_end)
1918 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1919 // Disable vector promotion when there are loads or stores of an FCA.
1921 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1922 if (LI->isVolatile())
1924 Type *LTy = LI->getType();
1925 if (SliceBeginOffset > S.beginOffset() || SliceEndOffset < S.endOffset()) {
1926 assert(LTy->isIntegerTy());
1929 if (!canConvertValue(DL, SliceTy, LTy))
1931 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1932 if (SI->isVolatile())
1934 Type *STy = SI->getValueOperand()->getType();
1935 if (SliceBeginOffset > S.beginOffset() || SliceEndOffset < S.endOffset()) {
1936 assert(STy->isIntegerTy());
1939 if (!canConvertValue(DL, STy, SliceTy))
1948 /// \brief Test whether the given alloca partitioning and range of slices can be
1949 /// promoted to a vector.
1951 /// This is a quick test to check whether we can rewrite a particular alloca
1952 /// partition (and its newly formed alloca) into a vector alloca with only
1953 /// whole-vector loads and stores such that it could be promoted to a vector
1954 /// SSA value. We only can ensure this for a limited set of operations, and we
1955 /// don't want to do the rewrites unless we are confident that the result will
1956 /// be promotable, so we have an early test here.
1958 isVectorPromotionViable(const DataLayout &DL, uint64_t SliceBeginOffset,
1959 uint64_t SliceEndOffset,
1960 AllocaSlices::const_range Slices,
1961 ArrayRef<AllocaSlices::iterator> SplitUses) {
1962 // Collect the candidate types for vector-based promotion. Also track whether
1963 // we have different element types.
1964 SmallVector<VectorType *, 4> CandidateTys;
1965 Type *CommonEltTy = nullptr;
1966 bool HaveCommonEltTy = true;
1967 auto CheckCandidateType = [&](Type *Ty) {
1968 if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1969 CandidateTys.push_back(VTy);
1971 CommonEltTy = VTy->getElementType();
1972 else if (CommonEltTy != VTy->getElementType())
1973 HaveCommonEltTy = false;
1976 // Consider any loads or stores that are the exact size of the slice.
1977 for (const auto &S : Slices)
1978 if (S.beginOffset() == SliceBeginOffset &&
1979 S.endOffset() == SliceEndOffset) {
1980 if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1981 CheckCandidateType(LI->getType());
1982 else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1983 CheckCandidateType(SI->getValueOperand()->getType());
1986 // If we didn't find a vector type, nothing to do here.
1987 if (CandidateTys.empty())
1990 // Remove non-integer vector types if we had multiple common element types.
1991 // FIXME: It'd be nice to replace them with integer vector types, but we can't
1992 // do that until all the backends are known to produce good code for all
1993 // integer vector types.
1994 if (!HaveCommonEltTy) {
1995 CandidateTys.erase(std::remove_if(CandidateTys.begin(), CandidateTys.end(),
1996 [](VectorType *VTy) {
1997 return !VTy->getElementType()->isIntegerTy();
1999 CandidateTys.end());
2001 // If there were no integer vector types, give up.
2002 if (CandidateTys.empty())
2005 // Rank the remaining candidate vector types. This is easy because we know
2006 // they're all integer vectors. We sort by ascending number of elements.
2007 auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
2008 assert(DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
2009 "Cannot have vector types of different sizes!");
2010 assert(RHSTy->getElementType()->isIntegerTy() &&
2011 "All non-integer types eliminated!");
2012 assert(LHSTy->getElementType()->isIntegerTy() &&
2013 "All non-integer types eliminated!");
2014 return RHSTy->getNumElements() < LHSTy->getNumElements();
2016 std::sort(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes);
2018 std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
2019 CandidateTys.end());
2021 // The only way to have the same element type in every vector type is to
2022 // have the same vector type. Check that and remove all but one.
2024 for (VectorType *VTy : CandidateTys) {
2025 assert(VTy->getElementType() == CommonEltTy &&
2026 "Unaccounted for element type!");
2027 assert(VTy == CandidateTys[0] &&
2028 "Different vector types with the same element type!");
2031 CandidateTys.resize(1);
2034 // Try each vector type, and return the one which works.
2035 auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
2036 uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
2038 // While the definition of LLVM vectors is bitpacked, we don't support sizes
2039 // that aren't byte sized.
2040 if (ElementSize % 8)
2042 assert((DL.getTypeSizeInBits(VTy) % 8) == 0 &&
2043 "vector size not a multiple of element size?");
2046 for (const auto &S : Slices)
2047 if (!isVectorPromotionViableForSlice(DL, SliceBeginOffset, SliceEndOffset,
2048 VTy, ElementSize, S))
2051 for (const auto &SI : SplitUses)
2052 if (!isVectorPromotionViableForSlice(DL, SliceBeginOffset, SliceEndOffset,
2053 VTy, ElementSize, *SI))
2058 for (VectorType *VTy : CandidateTys)
2059 if (CheckVectorTypeForPromotion(VTy))
2065 /// \brief Test whether a slice of an alloca is valid for integer widening.
2067 /// This implements the necessary checking for the \c isIntegerWideningViable
2068 /// test below on a single slice of the alloca.
2069 static bool isIntegerWideningViableForSlice(const DataLayout &DL,
2071 uint64_t AllocBeginOffset,
2072 uint64_t Size, const Slice &S,
2073 bool &WholeAllocaOp) {
2074 uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
2075 uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
2077 // We can't reasonably handle cases where the load or store extends past
2078 // the end of the aloca's type and into its padding.
2082 Use *U = S.getUse();
2084 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
2085 if (LI->isVolatile())
2087 // Note that we don't count vector loads or stores as whole-alloca
2088 // operations which enable integer widening because we would prefer to use
2089 // vector widening instead.
2090 if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
2091 WholeAllocaOp = true;
2092 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
2093 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
2095 } else if (RelBegin != 0 || RelEnd != Size ||
2096 !canConvertValue(DL, AllocaTy, LI->getType())) {
2097 // Non-integer loads need to be convertible from the alloca type so that
2098 // they are promotable.
2101 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
2102 Type *ValueTy = SI->getValueOperand()->getType();
2103 if (SI->isVolatile())
2105 // Note that we don't count vector loads or stores as whole-alloca
2106 // operations which enable integer widening because we would prefer to use
2107 // vector widening instead.
2108 if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
2109 WholeAllocaOp = true;
2110 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2111 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
2113 } else if (RelBegin != 0 || RelEnd != Size ||
2114 !canConvertValue(DL, ValueTy, AllocaTy)) {
2115 // Non-integer stores need to be convertible to the alloca type so that
2116 // they are promotable.
2119 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
2120 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
2122 if (!S.isSplittable())
2123 return false; // Skip any unsplittable intrinsics.
2124 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
2125 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2126 II->getIntrinsicID() != Intrinsic::lifetime_end)
2135 /// \brief Test whether the given alloca partition's integer operations can be
2136 /// widened to promotable ones.
2138 /// This is a quick test to check whether we can rewrite the integer loads and
2139 /// stores to a particular alloca into wider loads and stores and be able to
2140 /// promote the resulting alloca.
2142 isIntegerWideningViable(const DataLayout &DL, Type *AllocaTy,
2143 uint64_t AllocBeginOffset,
2144 AllocaSlices::const_range Slices,
2145 ArrayRef<AllocaSlices::iterator> SplitUses) {
2146 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
2147 // Don't create integer types larger than the maximum bitwidth.
2148 if (SizeInBits > IntegerType::MAX_INT_BITS)
2151 // Don't try to handle allocas with bit-padding.
2152 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
2155 // We need to ensure that an integer type with the appropriate bitwidth can
2156 // be converted to the alloca type, whatever that is. We don't want to force
2157 // the alloca itself to have an integer type if there is a more suitable one.
2158 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2159 if (!canConvertValue(DL, AllocaTy, IntTy) ||
2160 !canConvertValue(DL, IntTy, AllocaTy))
2163 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
2165 // While examining uses, we ensure that the alloca has a covering load or
2166 // store. We don't want to widen the integer operations only to fail to
2167 // promote due to some other unsplittable entry (which we may make splittable
2168 // later). However, if there are only splittable uses, go ahead and assume
2169 // that we cover the alloca.
2170 bool WholeAllocaOp =
2171 Slices.begin() != Slices.end() ? false : DL.isLegalInteger(SizeInBits);
2173 for (const auto &S : Slices)
2174 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
2178 for (const auto &SI : SplitUses)
2179 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
2180 *SI, WholeAllocaOp))
2183 return WholeAllocaOp;
2186 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2187 IntegerType *Ty, uint64_t Offset,
2188 const Twine &Name) {
2189 DEBUG(dbgs() << " start: " << *V << "\n");
2190 IntegerType *IntTy = cast<IntegerType>(V->getType());
2191 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2192 "Element extends past full value");
2193 uint64_t ShAmt = 8 * Offset;
2194 if (DL.isBigEndian())
2195 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2197 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2198 DEBUG(dbgs() << " shifted: " << *V << "\n");
2200 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2201 "Cannot extract to a larger integer!");
2203 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2204 DEBUG(dbgs() << " trunced: " << *V << "\n");
2209 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2210 Value *V, uint64_t Offset, const Twine &Name) {
2211 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2212 IntegerType *Ty = cast<IntegerType>(V->getType());
2213 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2214 "Cannot insert a larger integer!");
2215 DEBUG(dbgs() << " start: " << *V << "\n");
2217 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2218 DEBUG(dbgs() << " extended: " << *V << "\n");
2220 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2221 "Element store outside of alloca store");
2222 uint64_t ShAmt = 8 * Offset;
2223 if (DL.isBigEndian())
2224 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2226 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2227 DEBUG(dbgs() << " shifted: " << *V << "\n");
2230 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2231 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2232 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2233 DEBUG(dbgs() << " masked: " << *Old << "\n");
2234 V = IRB.CreateOr(Old, V, Name + ".insert");
2235 DEBUG(dbgs() << " inserted: " << *V << "\n");
2240 static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
2241 unsigned EndIndex, const Twine &Name) {
2242 VectorType *VecTy = cast<VectorType>(V->getType());
2243 unsigned NumElements = EndIndex - BeginIndex;
2244 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2246 if (NumElements == VecTy->getNumElements())
2249 if (NumElements == 1) {
2250 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2252 DEBUG(dbgs() << " extract: " << *V << "\n");
2256 SmallVector<Constant *, 8> Mask;
2257 Mask.reserve(NumElements);
2258 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2259 Mask.push_back(IRB.getInt32(i));
2260 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2261 ConstantVector::get(Mask), Name + ".extract");
2262 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2266 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2267 unsigned BeginIndex, const Twine &Name) {
2268 VectorType *VecTy = cast<VectorType>(Old->getType());
2269 assert(VecTy && "Can only insert a vector into a vector");
2271 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2273 // Single element to insert.
2274 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2276 DEBUG(dbgs() << " insert: " << *V << "\n");
2280 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2281 "Too many elements!");
2282 if (Ty->getNumElements() == VecTy->getNumElements()) {
2283 assert(V->getType() == VecTy && "Vector type mismatch");
2286 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2288 // When inserting a smaller vector into the larger to store, we first
2289 // use a shuffle vector to widen it with undef elements, and then
2290 // a second shuffle vector to select between the loaded vector and the
2292 SmallVector<Constant *, 8> Mask;
2293 Mask.reserve(VecTy->getNumElements());
2294 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2295 if (i >= BeginIndex && i < EndIndex)
2296 Mask.push_back(IRB.getInt32(i - BeginIndex));
2298 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2299 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2300 ConstantVector::get(Mask), Name + ".expand");
2301 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2304 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2305 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2307 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2309 DEBUG(dbgs() << " blend: " << *V << "\n");
2314 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
2315 /// to use a new alloca.
2317 /// Also implements the rewriting to vector-based accesses when the partition
2318 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2320 class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
2321 // Befriend the base class so it can delegate to private visit methods.
2322 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
2323 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
2325 const DataLayout &DL;
2328 AllocaInst &OldAI, &NewAI;
2329 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2332 // This is a convenience and flag variable that will be null unless the new
2333 // alloca's integer operations should be widened to this integer type due to
2334 // passing isIntegerWideningViable above. If it is non-null, the desired
2335 // integer type will be stored here for easy access during rewriting.
2338 // If we are rewriting an alloca partition which can be written as pure
2339 // vector operations, we stash extra information here. When VecTy is
2340 // non-null, we have some strict guarantees about the rewritten alloca:
2341 // - The new alloca is exactly the size of the vector type here.
2342 // - The accesses all either map to the entire vector or to a single
2344 // - The set of accessing instructions is only one of those handled above
2345 // in isVectorPromotionViable. Generally these are the same access kinds
2346 // which are promotable via mem2reg.
2349 uint64_t ElementSize;
2351 // The original offset of the slice currently being rewritten relative to
2352 // the original alloca.
2353 uint64_t BeginOffset, EndOffset;
2354 // The new offsets of the slice currently being rewritten relative to the
2356 uint64_t NewBeginOffset, NewEndOffset;
2362 Instruction *OldPtr;
2364 // Track post-rewrite users which are PHI nodes and Selects.
2365 SmallPtrSetImpl<PHINode *> &PHIUsers;
2366 SmallPtrSetImpl<SelectInst *> &SelectUsers;
2368 // Utility IR builder, whose name prefix is setup for each visited use, and
2369 // the insertion point is set to point to the user.
2373 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROA &Pass,
2374 AllocaInst &OldAI, AllocaInst &NewAI,
2375 uint64_t NewAllocaBeginOffset,
2376 uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2377 VectorType *PromotableVecTy,
2378 SmallPtrSetImpl<PHINode *> &PHIUsers,
2379 SmallPtrSetImpl<SelectInst *> &SelectUsers)
2380 : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2381 NewAllocaBeginOffset(NewAllocaBeginOffset),
2382 NewAllocaEndOffset(NewAllocaEndOffset),
2383 NewAllocaTy(NewAI.getAllocatedType()),
2384 IntTy(IsIntegerPromotable
2387 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2389 VecTy(PromotableVecTy),
2390 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2391 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2392 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2393 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2394 IRB(NewAI.getContext(), ConstantFolder()) {
2396 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2397 "Only multiple-of-8 sized vector elements are viable");
2400 assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy));
2403 bool visit(AllocaSlices::const_iterator I) {
2404 bool CanSROA = true;
2405 BeginOffset = I->beginOffset();
2406 EndOffset = I->endOffset();
2407 IsSplittable = I->isSplittable();
2409 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2411 // Compute the intersecting offset range.
2412 assert(BeginOffset < NewAllocaEndOffset);
2413 assert(EndOffset > NewAllocaBeginOffset);
2414 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2415 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2417 SliceSize = NewEndOffset - NewBeginOffset;
2419 OldUse = I->getUse();
2420 OldPtr = cast<Instruction>(OldUse->get());
2422 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2423 IRB.SetInsertPoint(OldUserI);
2424 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2425 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2427 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2434 // Make sure the other visit overloads are visible.
2437 // Every instruction which can end up as a user must have a rewrite rule.
2438 bool visitInstruction(Instruction &I) {
2439 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2440 llvm_unreachable("No rewrite rule for this instruction!");
2443 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2444 // Note that the offset computation can use BeginOffset or NewBeginOffset
2445 // interchangeably for unsplit slices.
2446 assert(IsSplit || BeginOffset == NewBeginOffset);
2447 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2450 StringRef OldName = OldPtr->getName();
2451 // Skip through the last '.sroa.' component of the name.
2452 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2453 if (LastSROAPrefix != StringRef::npos) {
2454 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2455 // Look for an SROA slice index.
2456 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2457 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2458 // Strip the index and look for the offset.
2459 OldName = OldName.substr(IndexEnd + 1);
2460 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2461 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2462 // Strip the offset.
2463 OldName = OldName.substr(OffsetEnd + 1);
2466 // Strip any SROA suffixes as well.
2467 OldName = OldName.substr(0, OldName.find(".sroa_"));
2470 return getAdjustedPtr(IRB, DL, &NewAI,
2471 APInt(DL.getPointerSizeInBits(), Offset), PointerTy,
2473 Twine(OldName) + "."
2480 /// \brief Compute suitable alignment to access this slice of the *new*
2483 /// You can optionally pass a type to this routine and if that type's ABI
2484 /// alignment is itself suitable, this will return zero.
2485 unsigned getSliceAlign(Type *Ty = nullptr) {
2486 unsigned NewAIAlign = NewAI.getAlignment();
2488 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2490 MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2491 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2494 unsigned getIndex(uint64_t Offset) {
2495 assert(VecTy && "Can only call getIndex when rewriting a vector");
2496 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2497 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2498 uint32_t Index = RelOffset / ElementSize;
2499 assert(Index * ElementSize == RelOffset);
2503 void deleteIfTriviallyDead(Value *V) {
2504 Instruction *I = cast<Instruction>(V);
2505 if (isInstructionTriviallyDead(I))
2506 Pass.DeadInsts.insert(I);
2509 Value *rewriteVectorizedLoadInst() {
2510 unsigned BeginIndex = getIndex(NewBeginOffset);
2511 unsigned EndIndex = getIndex(NewEndOffset);
2512 assert(EndIndex > BeginIndex && "Empty vector!");
2514 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2515 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2518 Value *rewriteIntegerLoad(LoadInst &LI) {
2519 assert(IntTy && "We cannot insert an integer to the alloca");
2520 assert(!LI.isVolatile());
2521 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2522 V = convertValue(DL, IRB, V, IntTy);
2523 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2524 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2525 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2526 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2531 bool visitLoadInst(LoadInst &LI) {
2532 DEBUG(dbgs() << " original: " << LI << "\n");
2533 Value *OldOp = LI.getOperand(0);
2534 assert(OldOp == OldPtr);
2536 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2538 bool IsPtrAdjusted = false;
2541 V = rewriteVectorizedLoadInst();
2542 } else if (IntTy && LI.getType()->isIntegerTy()) {
2543 V = rewriteIntegerLoad(LI);
2544 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2545 canConvertValue(DL, NewAllocaTy, LI.getType())) {
2546 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), LI.isVolatile(),
2549 Type *LTy = TargetTy->getPointerTo();
2550 V = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2551 getSliceAlign(TargetTy), LI.isVolatile(),
2553 IsPtrAdjusted = true;
2555 V = convertValue(DL, IRB, V, TargetTy);
2558 assert(!LI.isVolatile());
2559 assert(LI.getType()->isIntegerTy() &&
2560 "Only integer type loads and stores are split");
2561 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2562 "Split load isn't smaller than original load");
2563 assert(LI.getType()->getIntegerBitWidth() ==
2564 DL.getTypeStoreSizeInBits(LI.getType()) &&
2565 "Non-byte-multiple bit width");
2566 // Move the insertion point just past the load so that we can refer to it.
2567 IRB.SetInsertPoint(std::next(BasicBlock::iterator(&LI)));
2568 // Create a placeholder value with the same type as LI to use as the
2569 // basis for the new value. This allows us to replace the uses of LI with
2570 // the computed value, and then replace the placeholder with LI, leaving
2571 // LI only used for this computation.
2572 Value *Placeholder =
2573 new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2574 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset, "insert");
2575 LI.replaceAllUsesWith(V);
2576 Placeholder->replaceAllUsesWith(&LI);
2579 LI.replaceAllUsesWith(V);
2582 Pass.DeadInsts.insert(&LI);
2583 deleteIfTriviallyDead(OldOp);
2584 DEBUG(dbgs() << " to: " << *V << "\n");
2585 return !LI.isVolatile() && !IsPtrAdjusted;
2588 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2589 if (V->getType() != VecTy) {
2590 unsigned BeginIndex = getIndex(NewBeginOffset);
2591 unsigned EndIndex = getIndex(NewEndOffset);
2592 assert(EndIndex > BeginIndex && "Empty vector!");
2593 unsigned NumElements = EndIndex - BeginIndex;
2594 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2595 Type *SliceTy = (NumElements == 1)
2597 : VectorType::get(ElementTy, NumElements);
2598 if (V->getType() != SliceTy)
2599 V = convertValue(DL, IRB, V, SliceTy);
2601 // Mix in the existing elements.
2602 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2603 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2605 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2606 Pass.DeadInsts.insert(&SI);
2609 DEBUG(dbgs() << " to: " << *Store << "\n");
2613 bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2614 assert(IntTy && "We cannot extract an integer from the alloca");
2615 assert(!SI.isVolatile());
2616 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2618 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2619 Old = convertValue(DL, IRB, Old, IntTy);
2620 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2621 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2622 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2624 V = convertValue(DL, IRB, V, NewAllocaTy);
2625 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2626 Pass.DeadInsts.insert(&SI);
2628 DEBUG(dbgs() << " to: " << *Store << "\n");
2632 bool visitStoreInst(StoreInst &SI) {
2633 DEBUG(dbgs() << " original: " << SI << "\n");
2634 Value *OldOp = SI.getOperand(1);
2635 assert(OldOp == OldPtr);
2637 Value *V = SI.getValueOperand();
2639 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2640 // alloca that should be re-examined after promoting this alloca.
2641 if (V->getType()->isPointerTy())
2642 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2643 Pass.PostPromotionWorklist.insert(AI);
2645 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2646 assert(!SI.isVolatile());
2647 assert(V->getType()->isIntegerTy() &&
2648 "Only integer type loads and stores are split");
2649 assert(V->getType()->getIntegerBitWidth() ==
2650 DL.getTypeStoreSizeInBits(V->getType()) &&
2651 "Non-byte-multiple bit width");
2652 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2653 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset, "extract");
2657 return rewriteVectorizedStoreInst(V, SI, OldOp);
2658 if (IntTy && V->getType()->isIntegerTy())
2659 return rewriteIntegerStore(V, SI);
2662 if (NewBeginOffset == NewAllocaBeginOffset &&
2663 NewEndOffset == NewAllocaEndOffset &&
2664 canConvertValue(DL, V->getType(), NewAllocaTy)) {
2665 V = convertValue(DL, IRB, V, NewAllocaTy);
2666 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2669 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo());
2670 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2674 Pass.DeadInsts.insert(&SI);
2675 deleteIfTriviallyDead(OldOp);
2677 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2678 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2681 /// \brief Compute an integer value from splatting an i8 across the given
2682 /// number of bytes.
2684 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2685 /// call this routine.
2686 /// FIXME: Heed the advice above.
2688 /// \param V The i8 value to splat.
2689 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2690 Value *getIntegerSplat(Value *V, unsigned Size) {
2691 assert(Size > 0 && "Expected a positive number of bytes.");
2692 IntegerType *VTy = cast<IntegerType>(V->getType());
2693 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2697 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2699 IRB.CreateZExt(V, SplatIntTy, "zext"),
2700 ConstantExpr::getUDiv(
2701 Constant::getAllOnesValue(SplatIntTy),
2702 ConstantExpr::getZExt(Constant::getAllOnesValue(V->getType()),
2708 /// \brief Compute a vector splat for a given element value.
2709 Value *getVectorSplat(Value *V, unsigned NumElements) {
2710 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2711 DEBUG(dbgs() << " splat: " << *V << "\n");
2715 bool visitMemSetInst(MemSetInst &II) {
2716 DEBUG(dbgs() << " original: " << II << "\n");
2717 assert(II.getRawDest() == OldPtr);
2719 // If the memset has a variable size, it cannot be split, just adjust the
2720 // pointer to the new alloca.
2721 if (!isa<Constant>(II.getLength())) {
2723 assert(NewBeginOffset == BeginOffset);
2724 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2725 Type *CstTy = II.getAlignmentCst()->getType();
2726 II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2728 deleteIfTriviallyDead(OldPtr);
2732 // Record this instruction for deletion.
2733 Pass.DeadInsts.insert(&II);
2735 Type *AllocaTy = NewAI.getAllocatedType();
2736 Type *ScalarTy = AllocaTy->getScalarType();
2738 // If this doesn't map cleanly onto the alloca type, and that type isn't
2739 // a single value type, just emit a memset.
2740 if (!VecTy && !IntTy &&
2741 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2742 SliceSize != DL.getTypeStoreSize(AllocaTy) ||
2743 !AllocaTy->isSingleValueType() ||
2744 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2745 DL.getTypeSizeInBits(ScalarTy) % 8 != 0)) {
2746 Type *SizeTy = II.getLength()->getType();
2747 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2748 CallInst *New = IRB.CreateMemSet(
2749 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2750 getSliceAlign(), II.isVolatile());
2752 DEBUG(dbgs() << " to: " << *New << "\n");
2756 // If we can represent this as a simple value, we have to build the actual
2757 // value to store, which requires expanding the byte present in memset to
2758 // a sensible representation for the alloca type. This is essentially
2759 // splatting the byte to a sufficiently wide integer, splatting it across
2760 // any desired vector width, and bitcasting to the final type.
2764 // If this is a memset of a vectorized alloca, insert it.
2765 assert(ElementTy == ScalarTy);
2767 unsigned BeginIndex = getIndex(NewBeginOffset);
2768 unsigned EndIndex = getIndex(NewEndOffset);
2769 assert(EndIndex > BeginIndex && "Empty vector!");
2770 unsigned NumElements = EndIndex - BeginIndex;
2771 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2774 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2775 Splat = convertValue(DL, IRB, Splat, ElementTy);
2776 if (NumElements > 1)
2777 Splat = getVectorSplat(Splat, NumElements);
2780 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2781 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2783 // If this is a memset on an alloca where we can widen stores, insert the
2785 assert(!II.isVolatile());
2787 uint64_t Size = NewEndOffset - NewBeginOffset;
2788 V = getIntegerSplat(II.getValue(), Size);
2790 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2791 EndOffset != NewAllocaBeginOffset)) {
2793 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2794 Old = convertValue(DL, IRB, Old, IntTy);
2795 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2796 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2798 assert(V->getType() == IntTy &&
2799 "Wrong type for an alloca wide integer!");
2801 V = convertValue(DL, IRB, V, AllocaTy);
2803 // Established these invariants above.
2804 assert(NewBeginOffset == NewAllocaBeginOffset);
2805 assert(NewEndOffset == NewAllocaEndOffset);
2807 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2808 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2809 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2811 V = convertValue(DL, IRB, V, AllocaTy);
2814 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2817 DEBUG(dbgs() << " to: " << *New << "\n");
2818 return !II.isVolatile();
2821 bool visitMemTransferInst(MemTransferInst &II) {
2822 // Rewriting of memory transfer instructions can be a bit tricky. We break
2823 // them into two categories: split intrinsics and unsplit intrinsics.
2825 DEBUG(dbgs() << " original: " << II << "\n");
2827 bool IsDest = &II.getRawDestUse() == OldUse;
2828 assert((IsDest && II.getRawDest() == OldPtr) ||
2829 (!IsDest && II.getRawSource() == OldPtr));
2831 unsigned SliceAlign = getSliceAlign();
2833 // For unsplit intrinsics, we simply modify the source and destination
2834 // pointers in place. This isn't just an optimization, it is a matter of
2835 // correctness. With unsplit intrinsics we may be dealing with transfers
2836 // within a single alloca before SROA ran, or with transfers that have
2837 // a variable length. We may also be dealing with memmove instead of
2838 // memcpy, and so simply updating the pointers is the necessary for us to
2839 // update both source and dest of a single call.
2840 if (!IsSplittable) {
2841 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2843 II.setDest(AdjustedPtr);
2845 II.setSource(AdjustedPtr);
2847 if (II.getAlignment() > SliceAlign) {
2848 Type *CstTy = II.getAlignmentCst()->getType();
2850 ConstantInt::get(CstTy, MinAlign(II.getAlignment(), SliceAlign)));
2853 DEBUG(dbgs() << " to: " << II << "\n");
2854 deleteIfTriviallyDead(OldPtr);
2857 // For split transfer intrinsics we have an incredibly useful assurance:
2858 // the source and destination do not reside within the same alloca, and at
2859 // least one of them does not escape. This means that we can replace
2860 // memmove with memcpy, and we don't need to worry about all manner of
2861 // downsides to splitting and transforming the operations.
2863 // If this doesn't map cleanly onto the alloca type, and that type isn't
2864 // a single value type, just emit a memcpy.
2867 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2868 SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2869 !NewAI.getAllocatedType()->isSingleValueType());
2871 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2872 // size hasn't been shrunk based on analysis of the viable range, this is
2874 if (EmitMemCpy && &OldAI == &NewAI) {
2875 // Ensure the start lines up.
2876 assert(NewBeginOffset == BeginOffset);
2878 // Rewrite the size as needed.
2879 if (NewEndOffset != EndOffset)
2880 II.setLength(ConstantInt::get(II.getLength()->getType(),
2881 NewEndOffset - NewBeginOffset));
2884 // Record this instruction for deletion.
2885 Pass.DeadInsts.insert(&II);
2887 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2888 // alloca that should be re-examined after rewriting this instruction.
2889 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2890 if (AllocaInst *AI =
2891 dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2892 assert(AI != &OldAI && AI != &NewAI &&
2893 "Splittable transfers cannot reach the same alloca on both ends.");
2894 Pass.Worklist.insert(AI);
2897 Type *OtherPtrTy = OtherPtr->getType();
2898 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2900 // Compute the relative offset for the other pointer within the transfer.
2901 unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
2902 APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2903 unsigned OtherAlign = MinAlign(II.getAlignment() ? II.getAlignment() : 1,
2904 OtherOffset.zextOrTrunc(64).getZExtValue());
2907 // Compute the other pointer, folding as much as possible to produce
2908 // a single, simple GEP in most cases.
2909 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2910 OtherPtr->getName() + ".");
2912 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2913 Type *SizeTy = II.getLength()->getType();
2914 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2916 CallInst *New = IRB.CreateMemCpy(
2917 IsDest ? OurPtr : OtherPtr, IsDest ? OtherPtr : OurPtr, Size,
2918 MinAlign(SliceAlign, OtherAlign), II.isVolatile());
2920 DEBUG(dbgs() << " to: " << *New << "\n");
2924 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2925 NewEndOffset == NewAllocaEndOffset;
2926 uint64_t Size = NewEndOffset - NewBeginOffset;
2927 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2928 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2929 unsigned NumElements = EndIndex - BeginIndex;
2930 IntegerType *SubIntTy =
2931 IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
2933 // Reset the other pointer type to match the register type we're going to
2934 // use, but using the address space of the original other pointer.
2935 if (VecTy && !IsWholeAlloca) {
2936 if (NumElements == 1)
2937 OtherPtrTy = VecTy->getElementType();
2939 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2941 OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2942 } else if (IntTy && !IsWholeAlloca) {
2943 OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2945 OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2948 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2949 OtherPtr->getName() + ".");
2950 unsigned SrcAlign = OtherAlign;
2951 Value *DstPtr = &NewAI;
2952 unsigned DstAlign = SliceAlign;
2954 std::swap(SrcPtr, DstPtr);
2955 std::swap(SrcAlign, DstAlign);
2959 if (VecTy && !IsWholeAlloca && !IsDest) {
2960 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2961 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2962 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2963 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2964 Src = convertValue(DL, IRB, Src, IntTy);
2965 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2966 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2969 IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(), "copyload");
2972 if (VecTy && !IsWholeAlloca && IsDest) {
2974 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2975 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2976 } else if (IntTy && !IsWholeAlloca && IsDest) {
2978 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2979 Old = convertValue(DL, IRB, Old, IntTy);
2980 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2981 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2982 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2985 StoreInst *Store = cast<StoreInst>(
2986 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
2988 DEBUG(dbgs() << " to: " << *Store << "\n");
2989 return !II.isVolatile();
2992 bool visitIntrinsicInst(IntrinsicInst &II) {
2993 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2994 II.getIntrinsicID() == Intrinsic::lifetime_end);
2995 DEBUG(dbgs() << " original: " << II << "\n");
2996 assert(II.getArgOperand(1) == OldPtr);
2998 // Record this instruction for deletion.
2999 Pass.DeadInsts.insert(&II);
3002 ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
3003 NewEndOffset - NewBeginOffset);
3004 Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3006 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
3007 New = IRB.CreateLifetimeStart(Ptr, Size);
3009 New = IRB.CreateLifetimeEnd(Ptr, Size);
3012 DEBUG(dbgs() << " to: " << *New << "\n");
3016 bool visitPHINode(PHINode &PN) {
3017 DEBUG(dbgs() << " original: " << PN << "\n");
3018 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
3019 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
3021 // We would like to compute a new pointer in only one place, but have it be
3022 // as local as possible to the PHI. To do that, we re-use the location of
3023 // the old pointer, which necessarily must be in the right position to
3024 // dominate the PHI.
3025 IRBuilderTy PtrBuilder(IRB);
3026 if (isa<PHINode>(OldPtr))
3027 PtrBuilder.SetInsertPoint(OldPtr->getParent()->getFirstInsertionPt());
3029 PtrBuilder.SetInsertPoint(OldPtr);
3030 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
3032 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
3033 // Replace the operands which were using the old pointer.
3034 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
3036 DEBUG(dbgs() << " to: " << PN << "\n");
3037 deleteIfTriviallyDead(OldPtr);
3039 // PHIs can't be promoted on their own, but often can be speculated. We
3040 // check the speculation outside of the rewriter so that we see the
3041 // fully-rewritten alloca.
3042 PHIUsers.insert(&PN);
3046 bool visitSelectInst(SelectInst &SI) {
3047 DEBUG(dbgs() << " original: " << SI << "\n");
3048 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
3049 "Pointer isn't an operand!");
3050 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
3051 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
3053 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
3054 // Replace the operands which were using the old pointer.
3055 if (SI.getOperand(1) == OldPtr)
3056 SI.setOperand(1, NewPtr);
3057 if (SI.getOperand(2) == OldPtr)
3058 SI.setOperand(2, NewPtr);
3060 DEBUG(dbgs() << " to: " << SI << "\n");
3061 deleteIfTriviallyDead(OldPtr);
3063 // Selects can't be promoted on their own, but often can be speculated. We
3064 // check the speculation outside of the rewriter so that we see the
3065 // fully-rewritten alloca.
3066 SelectUsers.insert(&SI);
3073 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
3075 /// This pass aggressively rewrites all aggregate loads and stores on
3076 /// a particular pointer (or any pointer derived from it which we can identify)
3077 /// with scalar loads and stores.
3078 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3079 // Befriend the base class so it can delegate to private visit methods.
3080 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
3082 const DataLayout &DL;
3084 /// Queue of pointer uses to analyze and potentially rewrite.
3085 SmallVector<Use *, 8> Queue;
3087 /// Set to prevent us from cycling with phi nodes and loops.
3088 SmallPtrSet<User *, 8> Visited;
3090 /// The current pointer use being rewritten. This is used to dig up the used
3091 /// value (as opposed to the user).
3095 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
3097 /// Rewrite loads and stores through a pointer and all pointers derived from
3099 bool rewrite(Instruction &I) {
3100 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
3102 bool Changed = false;
3103 while (!Queue.empty()) {
3104 U = Queue.pop_back_val();
3105 Changed |= visit(cast<Instruction>(U->getUser()));
3111 /// Enqueue all the users of the given instruction for further processing.
3112 /// This uses a set to de-duplicate users.
3113 void enqueueUsers(Instruction &I) {
3114 for (Use &U : I.uses())
3115 if (Visited.insert(U.getUser()).second)
3116 Queue.push_back(&U);
3119 // Conservative default is to not rewrite anything.
3120 bool visitInstruction(Instruction &I) { return false; }
3122 /// \brief Generic recursive split emission class.
3123 template <typename Derived> class OpSplitter {
3125 /// The builder used to form new instructions.
3127 /// The indices which to be used with insert- or extractvalue to select the
3128 /// appropriate value within the aggregate.
3129 SmallVector<unsigned, 4> Indices;
3130 /// The indices to a GEP instruction which will move Ptr to the correct slot
3131 /// within the aggregate.
3132 SmallVector<Value *, 4> GEPIndices;
3133 /// The base pointer of the original op, used as a base for GEPing the
3134 /// split operations.
3137 /// Initialize the splitter with an insertion point, Ptr and start with a
3138 /// single zero GEP index.
3139 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3140 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3143 /// \brief Generic recursive split emission routine.
3145 /// This method recursively splits an aggregate op (load or store) into
3146 /// scalar or vector ops. It splits recursively until it hits a single value
3147 /// and emits that single value operation via the template argument.
3149 /// The logic of this routine relies on GEPs and insertvalue and
3150 /// extractvalue all operating with the same fundamental index list, merely
3151 /// formatted differently (GEPs need actual values).
3153 /// \param Ty The type being split recursively into smaller ops.
3154 /// \param Agg The aggregate value being built up or stored, depending on
3155 /// whether this is splitting a load or a store respectively.
3156 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3157 if (Ty->isSingleValueType())
3158 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3160 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3161 unsigned OldSize = Indices.size();
3163 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3165 assert(Indices.size() == OldSize && "Did not return to the old size");
3166 Indices.push_back(Idx);
3167 GEPIndices.push_back(IRB.getInt32(Idx));
3168 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3169 GEPIndices.pop_back();
3175 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3176 unsigned OldSize = Indices.size();
3178 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3180 assert(Indices.size() == OldSize && "Did not return to the old size");
3181 Indices.push_back(Idx);
3182 GEPIndices.push_back(IRB.getInt32(Idx));
3183 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3184 GEPIndices.pop_back();
3190 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3194 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3195 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3196 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3198 /// Emit a leaf load of a single value. This is called at the leaves of the
3199 /// recursive emission to actually load values.
3200 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3201 assert(Ty->isSingleValueType());
3202 // Load the single value and insert it using the indices.
3203 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
3204 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
3205 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3206 DEBUG(dbgs() << " to: " << *Load << "\n");
3210 bool visitLoadInst(LoadInst &LI) {
3211 assert(LI.getPointerOperand() == *U);
3212 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3215 // We have an aggregate being loaded, split it apart.
3216 DEBUG(dbgs() << " original: " << LI << "\n");
3217 LoadOpSplitter Splitter(&LI, *U);
3218 Value *V = UndefValue::get(LI.getType());
3219 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3220 LI.replaceAllUsesWith(V);
3221 LI.eraseFromParent();
3225 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3226 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3227 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3229 /// Emit a leaf store of a single value. This is called at the leaves of the
3230 /// recursive emission to actually produce stores.
3231 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3232 assert(Ty->isSingleValueType());
3233 // Extract the single value and store it using the indices.
3234 Value *Store = IRB.CreateStore(
3235 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
3236 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
3238 DEBUG(dbgs() << " to: " << *Store << "\n");
3242 bool visitStoreInst(StoreInst &SI) {
3243 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3245 Value *V = SI.getValueOperand();
3246 if (V->getType()->isSingleValueType())
3249 // We have an aggregate being stored, split it apart.
3250 DEBUG(dbgs() << " original: " << SI << "\n");
3251 StoreOpSplitter Splitter(&SI, *U);
3252 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3253 SI.eraseFromParent();
3257 bool visitBitCastInst(BitCastInst &BC) {
3262 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3267 bool visitPHINode(PHINode &PN) {
3272 bool visitSelectInst(SelectInst &SI) {
3279 /// \brief Strip aggregate type wrapping.
3281 /// This removes no-op aggregate types wrapping an underlying type. It will
3282 /// strip as many layers of types as it can without changing either the type
3283 /// size or the allocated size.
3284 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3285 if (Ty->isSingleValueType())
3288 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3289 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3292 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3293 InnerTy = ArrTy->getElementType();
3294 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3295 const StructLayout *SL = DL.getStructLayout(STy);
3296 unsigned Index = SL->getElementContainingOffset(0);
3297 InnerTy = STy->getElementType(Index);
3302 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3303 TypeSize > DL.getTypeSizeInBits(InnerTy))
3306 return stripAggregateTypeWrapping(DL, InnerTy);
3309 /// \brief Try to find a partition of the aggregate type passed in for a given
3310 /// offset and size.
3312 /// This recurses through the aggregate type and tries to compute a subtype
3313 /// based on the offset and size. When the offset and size span a sub-section
3314 /// of an array, it will even compute a new array type for that sub-section,
3315 /// and the same for structs.
3317 /// Note that this routine is very strict and tries to find a partition of the
3318 /// type which produces the *exact* right offset and size. It is not forgiving
3319 /// when the size or offset cause either end of type-based partition to be off.
3320 /// Also, this is a best-effort routine. It is reasonable to give up and not
3321 /// return a type if necessary.
3322 static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3324 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3325 return stripAggregateTypeWrapping(DL, Ty);
3326 if (Offset > DL.getTypeAllocSize(Ty) ||
3327 (DL.getTypeAllocSize(Ty) - Offset) < Size)
3330 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3331 // We can't partition pointers...
3332 if (SeqTy->isPointerTy())
3335 Type *ElementTy = SeqTy->getElementType();
3336 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3337 uint64_t NumSkippedElements = Offset / ElementSize;
3338 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
3339 if (NumSkippedElements >= ArrTy->getNumElements())
3341 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
3342 if (NumSkippedElements >= VecTy->getNumElements())
3345 Offset -= NumSkippedElements * ElementSize;
3347 // First check if we need to recurse.
3348 if (Offset > 0 || Size < ElementSize) {
3349 // Bail if the partition ends in a different array element.
3350 if ((Offset + Size) > ElementSize)
3352 // Recurse through the element type trying to peel off offset bytes.
3353 return getTypePartition(DL, ElementTy, Offset, Size);
3355 assert(Offset == 0);
3357 if (Size == ElementSize)
3358 return stripAggregateTypeWrapping(DL, ElementTy);
3359 assert(Size > ElementSize);
3360 uint64_t NumElements = Size / ElementSize;
3361 if (NumElements * ElementSize != Size)
3363 return ArrayType::get(ElementTy, NumElements);
3366 StructType *STy = dyn_cast<StructType>(Ty);
3370 const StructLayout *SL = DL.getStructLayout(STy);
3371 if (Offset >= SL->getSizeInBytes())
3373 uint64_t EndOffset = Offset + Size;
3374 if (EndOffset > SL->getSizeInBytes())
3377 unsigned Index = SL->getElementContainingOffset(Offset);
3378 Offset -= SL->getElementOffset(Index);
3380 Type *ElementTy = STy->getElementType(Index);
3381 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3382 if (Offset >= ElementSize)
3383 return nullptr; // The offset points into alignment padding.
3385 // See if any partition must be contained by the element.
3386 if (Offset > 0 || Size < ElementSize) {
3387 if ((Offset + Size) > ElementSize)
3389 return getTypePartition(DL, ElementTy, Offset, Size);
3391 assert(Offset == 0);
3393 if (Size == ElementSize)
3394 return stripAggregateTypeWrapping(DL, ElementTy);
3396 StructType::element_iterator EI = STy->element_begin() + Index,
3397 EE = STy->element_end();
3398 if (EndOffset < SL->getSizeInBytes()) {
3399 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3400 if (Index == EndIndex)
3401 return nullptr; // Within a single element and its padding.
3403 // Don't try to form "natural" types if the elements don't line up with the
3405 // FIXME: We could potentially recurse down through the last element in the
3406 // sub-struct to find a natural end point.
3407 if (SL->getElementOffset(EndIndex) != EndOffset)
3410 assert(Index < EndIndex);
3411 EE = STy->element_begin() + EndIndex;
3414 // Try to build up a sub-structure.
3416 StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3417 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3418 if (Size != SubSL->getSizeInBytes())
3419 return nullptr; // The sub-struct doesn't have quite the size needed.
3424 /// \brief Rewrite an alloca partition's users.
3426 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3427 /// to rewrite uses of an alloca partition to be conducive for SSA value
3428 /// promotion. If the partition needs a new, more refined alloca, this will
3429 /// build that new alloca, preserving as much type information as possible, and
3430 /// rewrite the uses of the old alloca to point at the new one and have the
3431 /// appropriate new offsets. It also evaluates how successful the rewrite was
3432 /// at enabling promotion and if it was successful queues the alloca to be
3434 bool SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
3435 AllocaSlices::Partition &P) {
3436 // Try to compute a friendly type for this partition of the alloca. This
3437 // won't always succeed, in which case we fall back to a legal integer type
3438 // or an i8 array of an appropriate size.
3439 Type *SliceTy = nullptr;
3440 if (Type *CommonUseTy = findCommonType(P.begin(), P.end(), P.endOffset()))
3441 if (DL->getTypeAllocSize(CommonUseTy) >= P.size())
3442 SliceTy = CommonUseTy;
3444 if (Type *TypePartitionTy = getTypePartition(*DL, AI.getAllocatedType(),
3445 P.beginOffset(), P.size()))
3446 SliceTy = TypePartitionTy;
3447 if ((!SliceTy || (SliceTy->isArrayTy() &&
3448 SliceTy->getArrayElementType()->isIntegerTy())) &&
3449 DL->isLegalInteger(P.size() * 8))
3450 SliceTy = Type::getIntNTy(*C, P.size() * 8);
3452 SliceTy = ArrayType::get(Type::getInt8Ty(*C), P.size());
3453 assert(DL->getTypeAllocSize(SliceTy) >= P.size());
3455 bool IsIntegerPromotable = isIntegerWideningViable(
3456 *DL, SliceTy, P.beginOffset(),
3457 AllocaSlices::const_range(P.begin(), P.end()), P.splitSlices());
3462 : isVectorPromotionViable(
3463 *DL, P.beginOffset(), P.endOffset(),
3464 AllocaSlices::const_range(P.begin(), P.end()), P.splitSlices());
3468 // Check for the case where we're going to rewrite to a new alloca of the
3469 // exact same type as the original, and with the same access offsets. In that
3470 // case, re-use the existing alloca, but still run through the rewriter to
3471 // perform phi and select speculation.
3473 if (SliceTy == AI.getAllocatedType()) {
3474 assert(P.beginOffset() == 0 &&
3475 "Non-zero begin offset but same alloca type");
3477 // FIXME: We should be able to bail at this point with "nothing changed".
3478 // FIXME: We might want to defer PHI speculation until after here.
3480 unsigned Alignment = AI.getAlignment();
3482 // The minimum alignment which users can rely on when the explicit
3483 // alignment is omitted or zero is that required by the ABI for this
3485 Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
3487 Alignment = MinAlign(Alignment, P.beginOffset());
3488 // If we will get at least this much alignment from the type alone, leave
3489 // the alloca's alignment unconstrained.
3490 if (Alignment <= DL->getABITypeAlignment(SliceTy))
3492 NewAI = new AllocaInst(
3493 SliceTy, nullptr, Alignment,
3494 AI.getName() + ".sroa." + Twine(P.begin() - AS.begin()), &AI);
3498 DEBUG(dbgs() << "Rewriting alloca partition "
3499 << "[" << P.beginOffset() << "," << P.endOffset()
3500 << ") to: " << *NewAI << "\n");
3502 // Track the high watermark on the worklist as it is only relevant for
3503 // promoted allocas. We will reset it to this point if the alloca is not in
3504 // fact scheduled for promotion.
3505 unsigned PPWOldSize = PostPromotionWorklist.size();
3506 unsigned NumUses = 0;
3507 SmallPtrSet<PHINode *, 8> PHIUsers;
3508 SmallPtrSet<SelectInst *, 8> SelectUsers;
3510 AllocaSliceRewriter Rewriter(*DL, AS, *this, AI, *NewAI, P.beginOffset(),
3511 P.endOffset(), IsIntegerPromotable, VecTy,
3512 PHIUsers, SelectUsers);
3513 bool Promotable = true;
3514 for (Slice *S : P.splitSlices()) {
3515 DEBUG(dbgs() << " rewriting split ");
3516 DEBUG(AS.printSlice(dbgs(), S, ""));
3517 Promotable &= Rewriter.visit(S);
3520 for (Slice &S : P) {
3521 DEBUG(dbgs() << " rewriting ");
3522 DEBUG(AS.printSlice(dbgs(), &S, ""));
3523 Promotable &= Rewriter.visit(&S);
3527 NumAllocaPartitionUses += NumUses;
3528 MaxUsesPerAllocaPartition =
3529 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3531 // Now that we've processed all the slices in the new partition, check if any
3532 // PHIs or Selects would block promotion.
3533 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3536 if (!isSafePHIToSpeculate(**I, DL)) {
3539 SelectUsers.clear();
3542 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3543 E = SelectUsers.end();
3545 if (!isSafeSelectToSpeculate(**I, DL)) {
3548 SelectUsers.clear();
3553 if (PHIUsers.empty() && SelectUsers.empty()) {
3554 // Promote the alloca.
3555 PromotableAllocas.push_back(NewAI);
3557 // If we have either PHIs or Selects to speculate, add them to those
3558 // worklists and re-queue the new alloca so that we promote in on the
3560 for (PHINode *PHIUser : PHIUsers)
3561 SpeculatablePHIs.insert(PHIUser);
3562 for (SelectInst *SelectUser : SelectUsers)
3563 SpeculatableSelects.insert(SelectUser);
3564 Worklist.insert(NewAI);
3567 // If we can't promote the alloca, iterate on it to check for new
3568 // refinements exposed by splitting the current alloca. Don't iterate on an
3569 // alloca which didn't actually change and didn't get promoted.
3571 Worklist.insert(NewAI);
3573 // Drop any post-promotion work items if promotion didn't happen.
3574 while (PostPromotionWorklist.size() > PPWOldSize)
3575 PostPromotionWorklist.pop_back();
3581 /// \brief Walks the slices of an alloca and form partitions based on them,
3582 /// rewriting each of their uses.
3583 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
3584 if (AS.begin() == AS.end())
3587 unsigned NumPartitions = 0;
3588 bool Changed = false;
3590 // Rewrite each parttion.
3591 for (auto &P : AS.partitions()) {
3592 Changed |= rewritePartition(AI, AS, P);
3596 NumAllocaPartitions += NumPartitions;
3597 MaxPartitionsPerAlloca =
3598 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
3603 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
3604 void SROA::clobberUse(Use &U) {
3606 // Replace the use with an undef value.
3607 U = UndefValue::get(OldV->getType());
3609 // Check for this making an instruction dead. We have to garbage collect
3610 // all the dead instructions to ensure the uses of any alloca end up being
3612 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3613 if (isInstructionTriviallyDead(OldI)) {
3614 DeadInsts.insert(OldI);
3618 /// \brief Analyze an alloca for SROA.
3620 /// This analyzes the alloca to ensure we can reason about it, builds
3621 /// the slices of the alloca, and then hands it off to be split and
3622 /// rewritten as needed.
3623 bool SROA::runOnAlloca(AllocaInst &AI) {
3624 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3625 ++NumAllocasAnalyzed;
3627 // Special case dead allocas, as they're trivial.
3628 if (AI.use_empty()) {
3629 AI.eraseFromParent();
3633 // Skip alloca forms that this analysis can't handle.
3634 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3635 DL->getTypeAllocSize(AI.getAllocatedType()) == 0)
3638 bool Changed = false;
3640 // First, split any FCA loads and stores touching this alloca to promote
3641 // better splitting and promotion opportunities.
3642 AggLoadStoreRewriter AggRewriter(*DL);
3643 Changed |= AggRewriter.rewrite(AI);
3645 // Build the slices using a recursive instruction-visiting builder.
3646 AllocaSlices AS(*DL, AI);
3647 DEBUG(AS.print(dbgs()));
3651 // Delete all the dead users of this alloca before splitting and rewriting it.
3652 for (Instruction *DeadUser : AS.getDeadUsers()) {
3653 // Free up everything used by this instruction.
3654 for (Use &DeadOp : DeadUser->operands())
3657 // Now replace the uses of this instruction.
3658 DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
3660 // And mark it for deletion.
3661 DeadInsts.insert(DeadUser);
3664 for (Use *DeadOp : AS.getDeadOperands()) {
3665 clobberUse(*DeadOp);
3669 // No slices to split. Leave the dead alloca for a later pass to clean up.
3670 if (AS.begin() == AS.end())
3673 Changed |= splitAlloca(AI, AS);
3675 DEBUG(dbgs() << " Speculating PHIs\n");
3676 while (!SpeculatablePHIs.empty())
3677 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
3679 DEBUG(dbgs() << " Speculating Selects\n");
3680 while (!SpeculatableSelects.empty())
3681 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
3686 /// \brief Delete the dead instructions accumulated in this run.
3688 /// Recursively deletes the dead instructions we've accumulated. This is done
3689 /// at the very end to maximize locality of the recursive delete and to
3690 /// minimize the problems of invalidated instruction pointers as such pointers
3691 /// are used heavily in the intermediate stages of the algorithm.
3693 /// We also record the alloca instructions deleted here so that they aren't
3694 /// subsequently handed to mem2reg to promote.
3695 void SROA::deleteDeadInstructions(
3696 SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
3697 while (!DeadInsts.empty()) {
3698 Instruction *I = DeadInsts.pop_back_val();
3699 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3701 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3703 for (Use &Operand : I->operands())
3704 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
3705 // Zero out the operand and see if it becomes trivially dead.
3707 if (isInstructionTriviallyDead(U))
3708 DeadInsts.insert(U);
3711 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3712 DeletedAllocas.insert(AI);
3715 I->eraseFromParent();
3719 static void enqueueUsersInWorklist(Instruction &I,
3720 SmallVectorImpl<Instruction *> &Worklist,
3721 SmallPtrSetImpl<Instruction *> &Visited) {
3722 for (User *U : I.users())
3723 if (Visited.insert(cast<Instruction>(U)).second)
3724 Worklist.push_back(cast<Instruction>(U));
3727 /// \brief Promote the allocas, using the best available technique.
3729 /// This attempts to promote whatever allocas have been identified as viable in
3730 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3731 /// If there is a domtree available, we attempt to promote using the full power
3732 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3733 /// based on the SSAUpdater utilities. This function returns whether any
3734 /// promotion occurred.
3735 bool SROA::promoteAllocas(Function &F) {
3736 if (PromotableAllocas.empty())
3739 NumPromoted += PromotableAllocas.size();
3741 if (DT && !ForceSSAUpdater) {
3742 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3743 PromoteMemToReg(PromotableAllocas, *DT, nullptr, AT);
3744 PromotableAllocas.clear();
3748 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3750 DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false);
3751 SmallVector<Instruction *, 64> Insts;
3753 // We need a worklist to walk the uses of each alloca.
3754 SmallVector<Instruction *, 8> Worklist;
3755 SmallPtrSet<Instruction *, 8> Visited;
3756 SmallVector<Instruction *, 32> DeadInsts;
3758 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3759 AllocaInst *AI = PromotableAllocas[Idx];
3764 enqueueUsersInWorklist(*AI, Worklist, Visited);
3766 while (!Worklist.empty()) {
3767 Instruction *I = Worklist.pop_back_val();
3769 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3770 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3771 // leading to them) here. Eventually it should use them to optimize the
3772 // scalar values produced.
3773 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3774 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3775 II->getIntrinsicID() == Intrinsic::lifetime_end);
3776 II->eraseFromParent();
3780 // Push the loads and stores we find onto the list. SROA will already
3781 // have validated that all loads and stores are viable candidates for
3783 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3784 assert(LI->getType() == AI->getAllocatedType());
3785 Insts.push_back(LI);
3788 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3789 assert(SI->getValueOperand()->getType() == AI->getAllocatedType());
3790 Insts.push_back(SI);
3794 // For everything else, we know that only no-op bitcasts and GEPs will
3795 // make it this far, just recurse through them and recall them for later
3797 DeadInsts.push_back(I);
3798 enqueueUsersInWorklist(*I, Worklist, Visited);
3800 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3801 while (!DeadInsts.empty())
3802 DeadInsts.pop_back_val()->eraseFromParent();
3803 AI->eraseFromParent();
3806 PromotableAllocas.clear();
3810 bool SROA::runOnFunction(Function &F) {
3811 if (skipOptnoneFunction(F))
3814 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3815 C = &F.getContext();
3816 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
3818 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3821 DL = &DLP->getDataLayout();
3822 DominatorTreeWrapperPass *DTWP =
3823 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
3824 DT = DTWP ? &DTWP->getDomTree() : nullptr;
3825 AT = &getAnalysis<AssumptionTracker>();
3827 BasicBlock &EntryBB = F.getEntryBlock();
3828 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
3830 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3831 Worklist.insert(AI);
3833 bool Changed = false;
3834 // A set of deleted alloca instruction pointers which should be removed from
3835 // the list of promotable allocas.
3836 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3839 while (!Worklist.empty()) {
3840 Changed |= runOnAlloca(*Worklist.pop_back_val());
3841 deleteDeadInstructions(DeletedAllocas);
3843 // Remove the deleted allocas from various lists so that we don't try to
3844 // continue processing them.
3845 if (!DeletedAllocas.empty()) {
3846 auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
3847 Worklist.remove_if(IsInSet);
3848 PostPromotionWorklist.remove_if(IsInSet);
3849 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3850 PromotableAllocas.end(),
3852 PromotableAllocas.end());
3853 DeletedAllocas.clear();
3857 Changed |= promoteAllocas(F);
3859 Worklist = PostPromotionWorklist;
3860 PostPromotionWorklist.clear();
3861 } while (!Worklist.empty());
3866 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3867 AU.addRequired<AssumptionTracker>();
3868 if (RequiresDomTree)
3869 AU.addRequired<DominatorTreeWrapperPass>();
3870 AU.setPreservesCFG();