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 /// \brief Access the dead users for this alloca.
241 ArrayRef<Instruction *> getDeadUsers() const { return DeadUsers; }
243 /// \brief Access the dead operands referring to this alloca.
245 /// These are operands which have cannot actually be used to refer to the
246 /// alloca as they are outside its range and the user doesn't correct for
247 /// that. These mostly consist of PHI node inputs and the like which we just
248 /// need to replace with undef.
249 ArrayRef<Use *> getDeadOperands() const { return DeadOperands; }
251 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
252 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
253 void printSlice(raw_ostream &OS, const_iterator I,
254 StringRef Indent = " ") const;
255 void printUse(raw_ostream &OS, const_iterator I,
256 StringRef Indent = " ") const;
257 void print(raw_ostream &OS) const;
258 void dump(const_iterator I) const;
263 template <typename DerivedT, typename RetT = void> class BuilderBase;
265 friend class AllocaSlices::SliceBuilder;
267 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
268 /// \brief Handle to alloca instruction to simplify method interfaces.
272 /// \brief The instruction responsible for this alloca not having a known set
275 /// When an instruction (potentially) escapes the pointer to the alloca, we
276 /// store a pointer to that here and abort trying to form slices of the
277 /// alloca. This will be null if the alloca slices are analyzed successfully.
278 Instruction *PointerEscapingInstr;
280 /// \brief The slices of the alloca.
282 /// We store a vector of the slices formed by uses of the alloca here. This
283 /// vector is sorted by increasing begin offset, and then the unsplittable
284 /// slices before the splittable ones. See the Slice inner class for more
286 SmallVector<Slice, 8> Slices;
288 /// \brief Instructions which will become dead if we rewrite the alloca.
290 /// Note that these are not separated by slice. This is because we expect an
291 /// alloca to be completely rewritten or not rewritten at all. If rewritten,
292 /// all these instructions can simply be removed and replaced with undef as
293 /// they come from outside of the allocated space.
294 SmallVector<Instruction *, 8> DeadUsers;
296 /// \brief Operands which will become dead if we rewrite the alloca.
298 /// These are operands that in their particular use can be replaced with
299 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
300 /// to PHI nodes and the like. They aren't entirely dead (there might be
301 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
302 /// want to swap this particular input for undef to simplify the use lists of
304 SmallVector<Use *, 8> DeadOperands;
308 static Value *foldSelectInst(SelectInst &SI) {
309 // If the condition being selected on is a constant or the same value is
310 // being selected between, fold the select. Yes this does (rarely) happen
312 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
313 return SI.getOperand(1 + CI->isZero());
314 if (SI.getOperand(1) == SI.getOperand(2))
315 return SI.getOperand(1);
320 /// \brief A helper that folds a PHI node or a select.
321 static Value *foldPHINodeOrSelectInst(Instruction &I) {
322 if (PHINode *PN = dyn_cast<PHINode>(&I)) {
323 // If PN merges together the same value, return that value.
324 return PN->hasConstantValue();
326 return foldSelectInst(cast<SelectInst>(I));
329 /// \brief Builder for the alloca slices.
331 /// This class builds a set of alloca slices by recursively visiting the uses
332 /// of an alloca and making a slice for each load and store at each offset.
333 class AllocaSlices::SliceBuilder : public PtrUseVisitor<SliceBuilder> {
334 friend class PtrUseVisitor<SliceBuilder>;
335 friend class InstVisitor<SliceBuilder>;
336 typedef PtrUseVisitor<SliceBuilder> Base;
338 const uint64_t AllocSize;
341 SmallDenseMap<Instruction *, unsigned> MemTransferSliceMap;
342 SmallDenseMap<Instruction *, uint64_t> PHIOrSelectSizes;
344 /// \brief Set to de-duplicate dead instructions found in the use walk.
345 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
348 SliceBuilder(const DataLayout &DL, AllocaInst &AI, AllocaSlices &AS)
349 : PtrUseVisitor<SliceBuilder>(DL),
350 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), AS(AS) {}
353 void markAsDead(Instruction &I) {
354 if (VisitedDeadInsts.insert(&I).second)
355 AS.DeadUsers.push_back(&I);
358 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
359 bool IsSplittable = false) {
360 // Completely skip uses which have a zero size or start either before or
361 // past the end of the allocation.
362 if (Size == 0 || Offset.uge(AllocSize)) {
363 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
364 << " which has zero size or starts outside of the "
365 << AllocSize << " byte alloca:\n"
366 << " alloca: " << AS.AI << "\n"
367 << " use: " << I << "\n");
368 return markAsDead(I);
371 uint64_t BeginOffset = Offset.getZExtValue();
372 uint64_t EndOffset = BeginOffset + Size;
374 // Clamp the end offset to the end of the allocation. Note that this is
375 // formulated to handle even the case where "BeginOffset + Size" overflows.
376 // This may appear superficially to be something we could ignore entirely,
377 // but that is not so! There may be widened loads or PHI-node uses where
378 // some instructions are dead but not others. We can't completely ignore
379 // them, and so have to record at least the information here.
380 assert(AllocSize >= BeginOffset); // Established above.
381 if (Size > AllocSize - BeginOffset) {
382 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
383 << " to remain within the " << AllocSize << " byte alloca:\n"
384 << " alloca: " << AS.AI << "\n"
385 << " use: " << I << "\n");
386 EndOffset = AllocSize;
389 AS.Slices.push_back(Slice(BeginOffset, EndOffset, U, IsSplittable));
392 void visitBitCastInst(BitCastInst &BC) {
394 return markAsDead(BC);
396 return Base::visitBitCastInst(BC);
399 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
400 if (GEPI.use_empty())
401 return markAsDead(GEPI);
403 if (SROAStrictInbounds && GEPI.isInBounds()) {
404 // FIXME: This is a manually un-factored variant of the basic code inside
405 // of GEPs with checking of the inbounds invariant specified in the
406 // langref in a very strict sense. If we ever want to enable
407 // SROAStrictInbounds, this code should be factored cleanly into
408 // PtrUseVisitor, but it is easier to experiment with SROAStrictInbounds
409 // by writing out the code here where we have tho underlying allocation
410 // size readily available.
411 APInt GEPOffset = Offset;
412 for (gep_type_iterator GTI = gep_type_begin(GEPI),
413 GTE = gep_type_end(GEPI);
415 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
419 // Handle a struct index, which adds its field offset to the pointer.
420 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
421 unsigned ElementIdx = OpC->getZExtValue();
422 const StructLayout *SL = DL.getStructLayout(STy);
424 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx));
426 // For array or vector indices, scale the index by the size of the
428 APInt Index = OpC->getValue().sextOrTrunc(Offset.getBitWidth());
429 GEPOffset += Index * APInt(Offset.getBitWidth(),
430 DL.getTypeAllocSize(GTI.getIndexedType()));
433 // If this index has computed an intermediate pointer which is not
434 // inbounds, then the result of the GEP is a poison value and we can
435 // delete it and all uses.
436 if (GEPOffset.ugt(AllocSize))
437 return markAsDead(GEPI);
441 return Base::visitGetElementPtrInst(GEPI);
444 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
445 uint64_t Size, bool IsVolatile) {
446 // We allow splitting of loads and stores where the type is an integer type
447 // and cover the entire alloca. This prevents us from splitting over
449 // FIXME: In the great blue eventually, we should eagerly split all integer
450 // loads and stores, and then have a separate step that merges adjacent
451 // alloca partitions into a single partition suitable for integer widening.
452 // Or we should skip the merge step and rely on GVN and other passes to
453 // merge adjacent loads and stores that survive mem2reg.
455 Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize;
457 insertUse(I, Offset, Size, IsSplittable);
460 void visitLoadInst(LoadInst &LI) {
461 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
462 "All simple FCA loads should have been pre-split");
465 return PI.setAborted(&LI);
467 uint64_t Size = DL.getTypeStoreSize(LI.getType());
468 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
471 void visitStoreInst(StoreInst &SI) {
472 Value *ValOp = SI.getValueOperand();
474 return PI.setEscapedAndAborted(&SI);
476 return PI.setAborted(&SI);
478 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
480 // If this memory access can be shown to *statically* extend outside the
481 // bounds of of the allocation, it's behavior is undefined, so simply
482 // ignore it. Note that this is more strict than the generic clamping
483 // behavior of insertUse. We also try to handle cases which might run the
485 // FIXME: We should instead consider the pointer to have escaped if this
486 // function is being instrumented for addressing bugs or race conditions.
487 if (Size > AllocSize || Offset.ugt(AllocSize - Size)) {
488 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
489 << " which extends past the end of the " << AllocSize
491 << " alloca: " << AS.AI << "\n"
492 << " use: " << SI << "\n");
493 return markAsDead(SI);
496 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
497 "All simple FCA stores should have been pre-split");
498 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
501 void visitMemSetInst(MemSetInst &II) {
502 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
503 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
504 if ((Length && Length->getValue() == 0) ||
505 (IsOffsetKnown && Offset.uge(AllocSize)))
506 // Zero-length mem transfer intrinsics can be ignored entirely.
507 return markAsDead(II);
510 return PI.setAborted(&II);
512 insertUse(II, Offset, Length ? Length->getLimitedValue()
513 : AllocSize - Offset.getLimitedValue(),
517 void visitMemTransferInst(MemTransferInst &II) {
518 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
519 if (Length && Length->getValue() == 0)
520 // Zero-length mem transfer intrinsics can be ignored entirely.
521 return markAsDead(II);
523 // Because we can visit these intrinsics twice, also check to see if the
524 // first time marked this instruction as dead. If so, skip it.
525 if (VisitedDeadInsts.count(&II))
529 return PI.setAborted(&II);
531 // This side of the transfer is completely out-of-bounds, and so we can
532 // nuke the entire transfer. However, we also need to nuke the other side
533 // if already added to our partitions.
534 // FIXME: Yet another place we really should bypass this when
535 // instrumenting for ASan.
536 if (Offset.uge(AllocSize)) {
537 SmallDenseMap<Instruction *, unsigned>::iterator MTPI =
538 MemTransferSliceMap.find(&II);
539 if (MTPI != MemTransferSliceMap.end())
540 AS.Slices[MTPI->second].kill();
541 return markAsDead(II);
544 uint64_t RawOffset = Offset.getLimitedValue();
545 uint64_t Size = Length ? Length->getLimitedValue() : AllocSize - RawOffset;
547 // Check for the special case where the same exact value is used for both
549 if (*U == II.getRawDest() && *U == II.getRawSource()) {
550 // For non-volatile transfers this is a no-op.
551 if (!II.isVolatile())
552 return markAsDead(II);
554 return insertUse(II, Offset, Size, /*IsSplittable=*/false);
557 // If we have seen both source and destination for a mem transfer, then
558 // they both point to the same alloca.
560 SmallDenseMap<Instruction *, unsigned>::iterator MTPI;
561 std::tie(MTPI, Inserted) =
562 MemTransferSliceMap.insert(std::make_pair(&II, AS.Slices.size()));
563 unsigned PrevIdx = MTPI->second;
565 Slice &PrevP = AS.Slices[PrevIdx];
567 // Check if the begin offsets match and this is a non-volatile transfer.
568 // In that case, we can completely elide the transfer.
569 if (!II.isVolatile() && PrevP.beginOffset() == RawOffset) {
571 return markAsDead(II);
574 // Otherwise we have an offset transfer within the same alloca. We can't
576 PrevP.makeUnsplittable();
579 // Insert the use now that we've fixed up the splittable nature.
580 insertUse(II, Offset, Size, /*IsSplittable=*/Inserted && Length);
582 // Check that we ended up with a valid index in the map.
583 assert(AS.Slices[PrevIdx].getUse()->getUser() == &II &&
584 "Map index doesn't point back to a slice with this user.");
587 // Disable SRoA for any intrinsics except for lifetime invariants.
588 // FIXME: What about debug intrinsics? This matches old behavior, but
589 // doesn't make sense.
590 void visitIntrinsicInst(IntrinsicInst &II) {
592 return PI.setAborted(&II);
594 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
595 II.getIntrinsicID() == Intrinsic::lifetime_end) {
596 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
597 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
598 Length->getLimitedValue());
599 insertUse(II, Offset, Size, true);
603 Base::visitIntrinsicInst(II);
606 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
607 // We consider any PHI or select that results in a direct load or store of
608 // the same offset to be a viable use for slicing purposes. These uses
609 // are considered unsplittable and the size is the maximum loaded or stored
611 SmallPtrSet<Instruction *, 4> Visited;
612 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
613 Visited.insert(Root);
614 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
615 // If there are no loads or stores, the access is dead. We mark that as
616 // a size zero access.
619 Instruction *I, *UsedI;
620 std::tie(UsedI, I) = Uses.pop_back_val();
622 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
623 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
626 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
627 Value *Op = SI->getOperand(0);
630 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
634 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
635 if (!GEP->hasAllZeroIndices())
637 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
638 !isa<SelectInst>(I)) {
642 for (User *U : I->users())
643 if (Visited.insert(cast<Instruction>(U)).second)
644 Uses.push_back(std::make_pair(I, cast<Instruction>(U)));
645 } while (!Uses.empty());
650 void visitPHINodeOrSelectInst(Instruction &I) {
651 assert(isa<PHINode>(I) || isa<SelectInst>(I));
653 return markAsDead(I);
655 // TODO: We could use SimplifyInstruction here to fold PHINodes and
656 // SelectInsts. However, doing so requires to change the current
657 // dead-operand-tracking mechanism. For instance, suppose neither loading
658 // from %U nor %other traps. Then "load (select undef, %U, %other)" does not
659 // trap either. However, if we simply replace %U with undef using the
660 // current dead-operand-tracking mechanism, "load (select undef, undef,
661 // %other)" may trap because the select may return the first operand
663 if (Value *Result = foldPHINodeOrSelectInst(I)) {
665 // If the result of the constant fold will be the pointer, recurse
666 // through the PHI/select as if we had RAUW'ed it.
669 // Otherwise the operand to the PHI/select is dead, and we can replace
671 AS.DeadOperands.push_back(U);
677 return PI.setAborted(&I);
679 // See if we already have computed info on this node.
680 uint64_t &Size = PHIOrSelectSizes[&I];
682 // This is a new PHI/Select, check for an unsafe use of it.
683 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&I, Size))
684 return PI.setAborted(UnsafeI);
687 // For PHI and select operands outside the alloca, we can't nuke the entire
688 // phi or select -- the other side might still be relevant, so we special
689 // case them here and use a separate structure to track the operands
690 // themselves which should be replaced with undef.
691 // FIXME: This should instead be escaped in the event we're instrumenting
692 // for address sanitization.
693 if (Offset.uge(AllocSize)) {
694 AS.DeadOperands.push_back(U);
698 insertUse(I, Offset, Size);
701 void visitPHINode(PHINode &PN) { visitPHINodeOrSelectInst(PN); }
703 void visitSelectInst(SelectInst &SI) { visitPHINodeOrSelectInst(SI); }
705 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
706 void visitInstruction(Instruction &I) { PI.setAborted(&I); }
709 AllocaSlices::AllocaSlices(const DataLayout &DL, AllocaInst &AI)
711 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
714 PointerEscapingInstr(nullptr) {
715 SliceBuilder PB(DL, AI, *this);
716 SliceBuilder::PtrInfo PtrI = PB.visitPtr(AI);
717 if (PtrI.isEscaped() || PtrI.isAborted()) {
718 // FIXME: We should sink the escape vs. abort info into the caller nicely,
719 // possibly by just storing the PtrInfo in the AllocaSlices.
720 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
721 : PtrI.getAbortingInst();
722 assert(PointerEscapingInstr && "Did not track a bad instruction");
726 Slices.erase(std::remove_if(Slices.begin(), Slices.end(),
732 #if __cplusplus >= 201103L && !defined(NDEBUG)
733 if (SROARandomShuffleSlices) {
734 std::mt19937 MT(static_cast<unsigned>(sys::TimeValue::now().msec()));
735 std::shuffle(Slices.begin(), Slices.end(), MT);
739 // Sort the uses. This arranges for the offsets to be in ascending order,
740 // and the sizes to be in descending order.
741 std::sort(Slices.begin(), Slices.end());
744 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
746 void AllocaSlices::print(raw_ostream &OS, const_iterator I,
747 StringRef Indent) const {
748 printSlice(OS, I, Indent);
749 printUse(OS, I, Indent);
752 void AllocaSlices::printSlice(raw_ostream &OS, const_iterator I,
753 StringRef Indent) const {
754 OS << Indent << "[" << I->beginOffset() << "," << I->endOffset() << ")"
755 << " slice #" << (I - begin())
756 << (I->isSplittable() ? " (splittable)" : "") << "\n";
759 void AllocaSlices::printUse(raw_ostream &OS, const_iterator I,
760 StringRef Indent) const {
761 OS << Indent << " used by: " << *I->getUse()->getUser() << "\n";
764 void AllocaSlices::print(raw_ostream &OS) const {
765 if (PointerEscapingInstr) {
766 OS << "Can't analyze slices for alloca: " << AI << "\n"
767 << " A pointer to this alloca escaped by:\n"
768 << " " << *PointerEscapingInstr << "\n";
772 OS << "Slices of alloca: " << AI << "\n";
773 for (const_iterator I = begin(), E = end(); I != E; ++I)
777 LLVM_DUMP_METHOD void AllocaSlices::dump(const_iterator I) const {
780 LLVM_DUMP_METHOD void AllocaSlices::dump() const { print(dbgs()); }
782 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
785 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
787 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
788 /// the loads and stores of an alloca instruction, as well as updating its
789 /// debug information. This is used when a domtree is unavailable and thus
790 /// mem2reg in its full form can't be used to handle promotion of allocas to
792 class AllocaPromoter : public LoadAndStorePromoter {
796 SmallVector<DbgDeclareInst *, 4> DDIs;
797 SmallVector<DbgValueInst *, 4> DVIs;
800 AllocaPromoter(const SmallVectorImpl<Instruction *> &Insts, SSAUpdater &S,
801 AllocaInst &AI, DIBuilder &DIB)
802 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
804 void run(const SmallVectorImpl<Instruction *> &Insts) {
805 // Retain the debug information attached to the alloca for use when
806 // rewriting loads and stores.
807 if (auto *L = LocalAsMetadata::getIfExists(&AI)) {
808 if (auto *DebugNode = MetadataAsValue::getIfExists(AI.getContext(), L)) {
809 for (User *U : DebugNode->users())
810 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(U))
812 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(U))
817 LoadAndStorePromoter::run(Insts);
819 // While we have the debug information, clear it off of the alloca. The
820 // caller takes care of deleting the alloca.
821 while (!DDIs.empty())
822 DDIs.pop_back_val()->eraseFromParent();
823 while (!DVIs.empty())
824 DVIs.pop_back_val()->eraseFromParent();
828 isInstInList(Instruction *I,
829 const SmallVectorImpl<Instruction *> &Insts) const override {
831 if (LoadInst *LI = dyn_cast<LoadInst>(I))
832 Ptr = LI->getOperand(0);
834 Ptr = cast<StoreInst>(I)->getPointerOperand();
836 // Only used to detect cycles, which will be rare and quickly found as
837 // we're walking up a chain of defs rather than down through uses.
838 SmallPtrSet<Value *, 4> Visited;
844 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr))
845 Ptr = BCI->getOperand(0);
846 else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Ptr))
847 Ptr = GEPI->getPointerOperand();
851 } while (Visited.insert(Ptr).second);
856 void updateDebugInfo(Instruction *Inst) const override {
857 for (DbgDeclareInst *DDI : DDIs)
858 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
859 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
860 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
861 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
862 for (DbgValueInst *DVI : DVIs) {
863 Value *Arg = nullptr;
864 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
865 // If an argument is zero extended then use argument directly. The ZExt
866 // may be zapped by an optimization pass in future.
867 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
868 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
869 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
870 Arg = dyn_cast<Argument>(SExt->getOperand(0));
872 Arg = SI->getValueOperand();
873 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
874 Arg = LI->getPointerOperand();
878 Instruction *DbgVal =
879 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
880 DIExpression(DVI->getExpression()), Inst);
881 DbgVal->setDebugLoc(DVI->getDebugLoc());
885 } // end anon namespace
888 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
890 /// This pass takes allocations which can be completely analyzed (that is, they
891 /// don't escape) and tries to turn them into scalar SSA values. There are
892 /// a few steps to this process.
894 /// 1) It takes allocations of aggregates and analyzes the ways in which they
895 /// are used to try to split them into smaller allocations, ideally of
896 /// a single scalar data type. It will split up memcpy and memset accesses
897 /// as necessary and try to isolate individual scalar accesses.
898 /// 2) It will transform accesses into forms which are suitable for SSA value
899 /// promotion. This can be replacing a memset with a scalar store of an
900 /// integer value, or it can involve speculating operations on a PHI or
901 /// select to be a PHI or select of the results.
902 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
903 /// onto insert and extract operations on a vector value, and convert them to
904 /// this form. By doing so, it will enable promotion of vector aggregates to
905 /// SSA vector values.
906 class SROA : public FunctionPass {
907 const bool RequiresDomTree;
910 const DataLayout *DL;
912 AssumptionTracker *AT;
914 /// \brief Worklist of alloca instructions to simplify.
916 /// Each alloca in the function is added to this. Each new alloca formed gets
917 /// added to it as well to recursively simplify unless that alloca can be
918 /// directly promoted. Finally, each time we rewrite a use of an alloca other
919 /// the one being actively rewritten, we add it back onto the list if not
920 /// already present to ensure it is re-visited.
921 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16>> Worklist;
923 /// \brief A collection of instructions to delete.
924 /// We try to batch deletions to simplify code and make things a bit more
926 SetVector<Instruction *, SmallVector<Instruction *, 8>> DeadInsts;
928 /// \brief Post-promotion worklist.
930 /// Sometimes we discover an alloca which has a high probability of becoming
931 /// viable for SROA after a round of promotion takes place. In those cases,
932 /// the alloca is enqueued here for re-processing.
934 /// Note that we have to be very careful to clear allocas out of this list in
935 /// the event they are deleted.
936 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16>> PostPromotionWorklist;
938 /// \brief A collection of alloca instructions we can directly promote.
939 std::vector<AllocaInst *> PromotableAllocas;
941 /// \brief A worklist of PHIs to speculate prior to promoting allocas.
943 /// All of these PHIs have been checked for the safety of speculation and by
944 /// being speculated will allow promoting allocas currently in the promotable
946 SetVector<PHINode *, SmallVector<PHINode *, 2>> SpeculatablePHIs;
948 /// \brief A worklist of select instructions to speculate prior to promoting
951 /// All of these select instructions have been checked for the safety of
952 /// speculation and by being speculated will allow promoting allocas
953 /// currently in the promotable queue.
954 SetVector<SelectInst *, SmallVector<SelectInst *, 2>> SpeculatableSelects;
957 SROA(bool RequiresDomTree = true)
958 : FunctionPass(ID), RequiresDomTree(RequiresDomTree), C(nullptr),
959 DL(nullptr), DT(nullptr) {
960 initializeSROAPass(*PassRegistry::getPassRegistry());
962 bool runOnFunction(Function &F) override;
963 void getAnalysisUsage(AnalysisUsage &AU) const override;
965 const char *getPassName() const override { return "SROA"; }
969 friend class PHIOrSelectSpeculator;
970 friend class AllocaSliceRewriter;
972 bool rewritePartition(AllocaInst &AI, AllocaSlices &AS,
973 AllocaSlices::iterator B, AllocaSlices::iterator E,
974 int64_t BeginOffset, int64_t EndOffset,
975 ArrayRef<AllocaSlices::iterator> SplitUses);
976 bool splitAlloca(AllocaInst &AI, AllocaSlices &AS);
977 bool runOnAlloca(AllocaInst &AI);
978 void clobberUse(Use &U);
979 void deleteDeadInstructions(SmallPtrSetImpl<AllocaInst *> &DeletedAllocas);
980 bool promoteAllocas(Function &F);
986 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
987 return new SROA(RequiresDomTree);
990 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates", false,
992 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
993 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
994 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates", false,
997 /// Walk the range of a partitioning looking for a common type to cover this
998 /// sequence of slices.
999 static Type *findCommonType(AllocaSlices::const_iterator B,
1000 AllocaSlices::const_iterator E,
1001 uint64_t EndOffset) {
1003 bool TyIsCommon = true;
1004 IntegerType *ITy = nullptr;
1006 // Note that we need to look at *every* alloca slice's Use to ensure we
1007 // always get consistent results regardless of the order of slices.
1008 for (AllocaSlices::const_iterator I = B; I != E; ++I) {
1009 Use *U = I->getUse();
1010 if (isa<IntrinsicInst>(*U->getUser()))
1012 if (I->beginOffset() != B->beginOffset() || I->endOffset() != EndOffset)
1015 Type *UserTy = nullptr;
1016 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1017 UserTy = LI->getType();
1018 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1019 UserTy = SI->getValueOperand()->getType();
1022 if (IntegerType *UserITy = dyn_cast_or_null<IntegerType>(UserTy)) {
1023 // If the type is larger than the partition, skip it. We only encounter
1024 // this for split integer operations where we want to use the type of the
1025 // entity causing the split. Also skip if the type is not a byte width
1027 if (UserITy->getBitWidth() % 8 != 0 ||
1028 UserITy->getBitWidth() / 8 > (EndOffset - B->beginOffset()))
1031 // Track the largest bitwidth integer type used in this way in case there
1032 // is no common type.
1033 if (!ITy || ITy->getBitWidth() < UserITy->getBitWidth())
1037 // To avoid depending on the order of slices, Ty and TyIsCommon must not
1038 // depend on types skipped above.
1039 if (!UserTy || (Ty && Ty != UserTy))
1040 TyIsCommon = false; // Give up on anything but an iN type.
1045 return TyIsCommon ? Ty : ITy;
1048 /// PHI instructions that use an alloca and are subsequently loaded can be
1049 /// rewritten to load both input pointers in the pred blocks and then PHI the
1050 /// results, allowing the load of the alloca to be promoted.
1052 /// %P2 = phi [i32* %Alloca, i32* %Other]
1053 /// %V = load i32* %P2
1055 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1057 /// %V2 = load i32* %Other
1059 /// %V = phi [i32 %V1, i32 %V2]
1061 /// We can do this to a select if its only uses are loads and if the operands
1062 /// to the select can be loaded unconditionally.
1064 /// FIXME: This should be hoisted into a generic utility, likely in
1065 /// Transforms/Util/Local.h
1066 static bool isSafePHIToSpeculate(PHINode &PN, const DataLayout *DL = nullptr) {
1067 // For now, we can only do this promotion if the load is in the same block
1068 // as the PHI, and if there are no stores between the phi and load.
1069 // TODO: Allow recursive phi users.
1070 // TODO: Allow stores.
1071 BasicBlock *BB = PN.getParent();
1072 unsigned MaxAlign = 0;
1073 bool HaveLoad = false;
1074 for (User *U : PN.users()) {
1075 LoadInst *LI = dyn_cast<LoadInst>(U);
1076 if (!LI || !LI->isSimple())
1079 // For now we only allow loads in the same block as the PHI. This is
1080 // a common case that happens when instcombine merges two loads through
1082 if (LI->getParent() != BB)
1085 // Ensure that there are no instructions between the PHI and the load that
1087 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1088 if (BBI->mayWriteToMemory())
1091 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1098 // We can only transform this if it is safe to push the loads into the
1099 // predecessor blocks. The only thing to watch out for is that we can't put
1100 // a possibly trapping load in the predecessor if it is a critical edge.
1101 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1102 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1103 Value *InVal = PN.getIncomingValue(Idx);
1105 // If the value is produced by the terminator of the predecessor (an
1106 // invoke) or it has side-effects, there is no valid place to put a load
1107 // in the predecessor.
1108 if (TI == InVal || TI->mayHaveSideEffects())
1111 // If the predecessor has a single successor, then the edge isn't
1113 if (TI->getNumSuccessors() == 1)
1116 // If this pointer is always safe to load, or if we can prove that there
1117 // is already a load in the block, then we can move the load to the pred
1119 if (InVal->isDereferenceablePointer(DL) ||
1120 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, DL))
1129 static void speculatePHINodeLoads(PHINode &PN) {
1130 DEBUG(dbgs() << " original: " << PN << "\n");
1132 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1133 IRBuilderTy PHIBuilder(&PN);
1134 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1135 PN.getName() + ".sroa.speculated");
1137 // Get the AA tags and alignment to use from one of the loads. It doesn't
1138 // matter which one we get and if any differ.
1139 LoadInst *SomeLoad = cast<LoadInst>(PN.user_back());
1142 SomeLoad->getAAMetadata(AATags);
1143 unsigned Align = SomeLoad->getAlignment();
1145 // Rewrite all loads of the PN to use the new PHI.
1146 while (!PN.use_empty()) {
1147 LoadInst *LI = cast<LoadInst>(PN.user_back());
1148 LI->replaceAllUsesWith(NewPN);
1149 LI->eraseFromParent();
1152 // Inject loads into all of the pred blocks.
1153 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1154 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1155 TerminatorInst *TI = Pred->getTerminator();
1156 Value *InVal = PN.getIncomingValue(Idx);
1157 IRBuilderTy PredBuilder(TI);
1159 LoadInst *Load = PredBuilder.CreateLoad(
1160 InVal, (PN.getName() + ".sroa.speculate.load." + Pred->getName()));
1161 ++NumLoadsSpeculated;
1162 Load->setAlignment(Align);
1164 Load->setAAMetadata(AATags);
1165 NewPN->addIncoming(Load, Pred);
1168 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1169 PN.eraseFromParent();
1172 /// Select instructions that use an alloca and are subsequently loaded can be
1173 /// rewritten to load both input pointers and then select between the result,
1174 /// allowing the load of the alloca to be promoted.
1176 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1177 /// %V = load i32* %P2
1179 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1180 /// %V2 = load i32* %Other
1181 /// %V = select i1 %cond, i32 %V1, i32 %V2
1183 /// We can do this to a select if its only uses are loads and if the operand
1184 /// to the select can be loaded unconditionally.
1185 static bool isSafeSelectToSpeculate(SelectInst &SI,
1186 const DataLayout *DL = nullptr) {
1187 Value *TValue = SI.getTrueValue();
1188 Value *FValue = SI.getFalseValue();
1189 bool TDerefable = TValue->isDereferenceablePointer(DL);
1190 bool FDerefable = FValue->isDereferenceablePointer(DL);
1192 for (User *U : SI.users()) {
1193 LoadInst *LI = dyn_cast<LoadInst>(U);
1194 if (!LI || !LI->isSimple())
1197 // Both operands to the select need to be dereferencable, either
1198 // absolutely (e.g. allocas) or at this point because we can see other
1201 !isSafeToLoadUnconditionally(TValue, LI, LI->getAlignment(), DL))
1204 !isSafeToLoadUnconditionally(FValue, LI, LI->getAlignment(), DL))
1211 static void speculateSelectInstLoads(SelectInst &SI) {
1212 DEBUG(dbgs() << " original: " << SI << "\n");
1214 IRBuilderTy IRB(&SI);
1215 Value *TV = SI.getTrueValue();
1216 Value *FV = SI.getFalseValue();
1217 // Replace the loads of the select with a select of two loads.
1218 while (!SI.use_empty()) {
1219 LoadInst *LI = cast<LoadInst>(SI.user_back());
1220 assert(LI->isSimple() && "We only speculate simple loads");
1222 IRB.SetInsertPoint(LI);
1224 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1226 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1227 NumLoadsSpeculated += 2;
1229 // Transfer alignment and AA info if present.
1230 TL->setAlignment(LI->getAlignment());
1231 FL->setAlignment(LI->getAlignment());
1234 LI->getAAMetadata(Tags);
1236 TL->setAAMetadata(Tags);
1237 FL->setAAMetadata(Tags);
1240 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1241 LI->getName() + ".sroa.speculated");
1243 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1244 LI->replaceAllUsesWith(V);
1245 LI->eraseFromParent();
1247 SI.eraseFromParent();
1250 /// \brief Build a GEP out of a base pointer and indices.
1252 /// This will return the BasePtr if that is valid, or build a new GEP
1253 /// instruction using the IRBuilder if GEP-ing is needed.
1254 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1255 SmallVectorImpl<Value *> &Indices, Twine NamePrefix) {
1256 if (Indices.empty())
1259 // A single zero index is a no-op, so check for this and avoid building a GEP
1261 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1264 return IRB.CreateInBoundsGEP(BasePtr, Indices, NamePrefix + "sroa_idx");
1267 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1268 /// TargetTy without changing the offset of the pointer.
1270 /// This routine assumes we've already established a properly offset GEP with
1271 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1272 /// zero-indices down through type layers until we find one the same as
1273 /// TargetTy. If we can't find one with the same type, we at least try to use
1274 /// one with the same size. If none of that works, we just produce the GEP as
1275 /// indicated by Indices to have the correct offset.
1276 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &DL,
1277 Value *BasePtr, Type *Ty, Type *TargetTy,
1278 SmallVectorImpl<Value *> &Indices,
1281 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1283 // Pointer size to use for the indices.
1284 unsigned PtrSize = DL.getPointerTypeSizeInBits(BasePtr->getType());
1286 // See if we can descend into a struct and locate a field with the correct
1288 unsigned NumLayers = 0;
1289 Type *ElementTy = Ty;
1291 if (ElementTy->isPointerTy())
1294 if (ArrayType *ArrayTy = dyn_cast<ArrayType>(ElementTy)) {
1295 ElementTy = ArrayTy->getElementType();
1296 Indices.push_back(IRB.getIntN(PtrSize, 0));
1297 } else if (VectorType *VectorTy = dyn_cast<VectorType>(ElementTy)) {
1298 ElementTy = VectorTy->getElementType();
1299 Indices.push_back(IRB.getInt32(0));
1300 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1301 if (STy->element_begin() == STy->element_end())
1302 break; // Nothing left to descend into.
1303 ElementTy = *STy->element_begin();
1304 Indices.push_back(IRB.getInt32(0));
1309 } while (ElementTy != TargetTy);
1310 if (ElementTy != TargetTy)
1311 Indices.erase(Indices.end() - NumLayers, Indices.end());
1313 return buildGEP(IRB, BasePtr, Indices, NamePrefix);
1316 /// \brief Recursively compute indices for a natural GEP.
1318 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1319 /// element types adding appropriate indices for the GEP.
1320 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
1321 Value *Ptr, Type *Ty, APInt &Offset,
1323 SmallVectorImpl<Value *> &Indices,
1326 return getNaturalGEPWithType(IRB, DL, Ptr, Ty, TargetTy, Indices,
1329 // We can't recurse through pointer types.
1330 if (Ty->isPointerTy())
1333 // We try to analyze GEPs over vectors here, but note that these GEPs are
1334 // extremely poorly defined currently. The long-term goal is to remove GEPing
1335 // over a vector from the IR completely.
1336 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1337 unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
1338 if (ElementSizeInBits % 8 != 0) {
1339 // GEPs over non-multiple of 8 size vector elements are invalid.
1342 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1343 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1344 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1346 Offset -= NumSkippedElements * ElementSize;
1347 Indices.push_back(IRB.getInt(NumSkippedElements));
1348 return getNaturalGEPRecursively(IRB, DL, Ptr, VecTy->getElementType(),
1349 Offset, TargetTy, Indices, NamePrefix);
1352 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1353 Type *ElementTy = ArrTy->getElementType();
1354 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1355 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1356 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1359 Offset -= NumSkippedElements * ElementSize;
1360 Indices.push_back(IRB.getInt(NumSkippedElements));
1361 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1362 Indices, NamePrefix);
1365 StructType *STy = dyn_cast<StructType>(Ty);
1369 const StructLayout *SL = DL.getStructLayout(STy);
1370 uint64_t StructOffset = Offset.getZExtValue();
1371 if (StructOffset >= SL->getSizeInBytes())
1373 unsigned Index = SL->getElementContainingOffset(StructOffset);
1374 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1375 Type *ElementTy = STy->getElementType(Index);
1376 if (Offset.uge(DL.getTypeAllocSize(ElementTy)))
1377 return nullptr; // The offset points into alignment padding.
1379 Indices.push_back(IRB.getInt32(Index));
1380 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1381 Indices, NamePrefix);
1384 /// \brief Get a natural GEP from a base pointer to a particular offset and
1385 /// resulting in a particular type.
1387 /// The goal is to produce a "natural" looking GEP that works with the existing
1388 /// composite types to arrive at the appropriate offset and element type for
1389 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1390 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1391 /// Indices, and setting Ty to the result subtype.
1393 /// If no natural GEP can be constructed, this function returns null.
1394 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &DL,
1395 Value *Ptr, APInt Offset, Type *TargetTy,
1396 SmallVectorImpl<Value *> &Indices,
1398 PointerType *Ty = cast<PointerType>(Ptr->getType());
1400 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1402 if (Ty == IRB.getInt8PtrTy(Ty->getAddressSpace()) && TargetTy->isIntegerTy(8))
1405 Type *ElementTy = Ty->getElementType();
1406 if (!ElementTy->isSized())
1407 return nullptr; // We can't GEP through an unsized element.
1408 APInt ElementSize(Offset.getBitWidth(), DL.getTypeAllocSize(ElementTy));
1409 if (ElementSize == 0)
1410 return nullptr; // Zero-length arrays can't help us build a natural GEP.
1411 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1413 Offset -= NumSkippedElements * ElementSize;
1414 Indices.push_back(IRB.getInt(NumSkippedElements));
1415 return getNaturalGEPRecursively(IRB, DL, Ptr, ElementTy, Offset, TargetTy,
1416 Indices, NamePrefix);
1419 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1420 /// resulting pointer has PointerTy.
1422 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1423 /// and produces the pointer type desired. Where it cannot, it will try to use
1424 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1425 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1426 /// bitcast to the type.
1428 /// The strategy for finding the more natural GEPs is to peel off layers of the
1429 /// pointer, walking back through bit casts and GEPs, searching for a base
1430 /// pointer from which we can compute a natural GEP with the desired
1431 /// properties. The algorithm tries to fold as many constant indices into
1432 /// a single GEP as possible, thus making each GEP more independent of the
1433 /// surrounding code.
1434 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &DL, Value *Ptr,
1435 APInt Offset, Type *PointerTy, Twine NamePrefix) {
1436 // Even though we don't look through PHI nodes, we could be called on an
1437 // instruction in an unreachable block, which may be on a cycle.
1438 SmallPtrSet<Value *, 4> Visited;
1439 Visited.insert(Ptr);
1440 SmallVector<Value *, 4> Indices;
1442 // We may end up computing an offset pointer that has the wrong type. If we
1443 // never are able to compute one directly that has the correct type, we'll
1444 // fall back to it, so keep it around here.
1445 Value *OffsetPtr = nullptr;
1447 // Remember any i8 pointer we come across to re-use if we need to do a raw
1449 Value *Int8Ptr = nullptr;
1450 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1452 Type *TargetTy = PointerTy->getPointerElementType();
1455 // First fold any existing GEPs into the offset.
1456 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1457 APInt GEPOffset(Offset.getBitWidth(), 0);
1458 if (!GEP->accumulateConstantOffset(DL, GEPOffset))
1460 Offset += GEPOffset;
1461 Ptr = GEP->getPointerOperand();
1462 if (!Visited.insert(Ptr).second)
1466 // See if we can perform a natural GEP here.
1468 if (Value *P = getNaturalGEPWithOffset(IRB, DL, Ptr, Offset, TargetTy,
1469 Indices, NamePrefix)) {
1470 if (P->getType() == PointerTy) {
1471 // Zap any offset pointer that we ended up computing in previous rounds.
1472 if (OffsetPtr && OffsetPtr->use_empty())
1473 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1474 I->eraseFromParent();
1482 // Stash this pointer if we've found an i8*.
1483 if (Ptr->getType()->isIntegerTy(8)) {
1485 Int8PtrOffset = Offset;
1488 // Peel off a layer of the pointer and update the offset appropriately.
1489 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1490 Ptr = cast<Operator>(Ptr)->getOperand(0);
1491 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1492 if (GA->mayBeOverridden())
1494 Ptr = GA->getAliasee();
1498 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1499 } while (Visited.insert(Ptr).second);
1503 Int8Ptr = IRB.CreateBitCast(
1504 Ptr, IRB.getInt8PtrTy(PointerTy->getPointerAddressSpace()),
1505 NamePrefix + "sroa_raw_cast");
1506 Int8PtrOffset = Offset;
1509 OffsetPtr = Int8PtrOffset == 0
1511 : IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1512 NamePrefix + "sroa_raw_idx");
1516 // On the off chance we were targeting i8*, guard the bitcast here.
1517 if (Ptr->getType() != PointerTy)
1518 Ptr = IRB.CreateBitCast(Ptr, PointerTy, NamePrefix + "sroa_cast");
1523 /// \brief Test whether we can convert a value from the old to the new type.
1525 /// This predicate should be used to guard calls to convertValue in order to
1526 /// ensure that we only try to convert viable values. The strategy is that we
1527 /// will peel off single element struct and array wrappings to get to an
1528 /// underlying value, and convert that value.
1529 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1532 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1533 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1534 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1536 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1538 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1541 // We can convert pointers to integers and vice-versa. Same for vectors
1542 // of pointers and integers.
1543 OldTy = OldTy->getScalarType();
1544 NewTy = NewTy->getScalarType();
1545 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1546 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1548 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1556 /// \brief Generic routine to convert an SSA value to a value of a different
1559 /// This will try various different casting techniques, such as bitcasts,
1560 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1561 /// two types for viability with this routine.
1562 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1564 Type *OldTy = V->getType();
1565 assert(canConvertValue(DL, OldTy, NewTy) && "Value not convertable to type");
1570 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1571 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1572 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1573 return IRB.CreateZExt(V, NewITy);
1575 // See if we need inttoptr for this type pair. A cast involving both scalars
1576 // and vectors requires and additional bitcast.
1577 if (OldTy->getScalarType()->isIntegerTy() &&
1578 NewTy->getScalarType()->isPointerTy()) {
1579 // Expand <2 x i32> to i8* --> <2 x i32> to i64 to i8*
1580 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1581 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1584 // Expand i128 to <2 x i8*> --> i128 to <2 x i64> to <2 x i8*>
1585 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1586 return IRB.CreateIntToPtr(IRB.CreateBitCast(V, DL.getIntPtrType(NewTy)),
1589 return IRB.CreateIntToPtr(V, NewTy);
1592 // See if we need ptrtoint for this type pair. A cast involving both scalars
1593 // and vectors requires and additional bitcast.
1594 if (OldTy->getScalarType()->isPointerTy() &&
1595 NewTy->getScalarType()->isIntegerTy()) {
1596 // Expand <2 x i8*> to i128 --> <2 x i8*> to <2 x i64> to i128
1597 if (OldTy->isVectorTy() && !NewTy->isVectorTy())
1598 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1601 // Expand i8* to <2 x i32> --> i8* to i64 to <2 x i32>
1602 if (!OldTy->isVectorTy() && NewTy->isVectorTy())
1603 return IRB.CreateBitCast(IRB.CreatePtrToInt(V, DL.getIntPtrType(OldTy)),
1606 return IRB.CreatePtrToInt(V, NewTy);
1609 return IRB.CreateBitCast(V, NewTy);
1612 /// \brief Test whether the given slice use can be promoted to a vector.
1614 /// This function is called to test each entry in a partioning which is slated
1615 /// for a single slice.
1617 isVectorPromotionViableForSlice(const DataLayout &DL, uint64_t SliceBeginOffset,
1618 uint64_t SliceEndOffset, VectorType *Ty,
1619 uint64_t ElementSize, const Slice &S) {
1620 // First validate the slice offsets.
1621 uint64_t BeginOffset =
1622 std::max(S.beginOffset(), SliceBeginOffset) - SliceBeginOffset;
1623 uint64_t BeginIndex = BeginOffset / ElementSize;
1624 if (BeginIndex * ElementSize != BeginOffset ||
1625 BeginIndex >= Ty->getNumElements())
1627 uint64_t EndOffset =
1628 std::min(S.endOffset(), SliceEndOffset) - SliceBeginOffset;
1629 uint64_t EndIndex = EndOffset / ElementSize;
1630 if (EndIndex * ElementSize != EndOffset || EndIndex > Ty->getNumElements())
1633 assert(EndIndex > BeginIndex && "Empty vector!");
1634 uint64_t NumElements = EndIndex - BeginIndex;
1635 Type *SliceTy = (NumElements == 1)
1636 ? Ty->getElementType()
1637 : VectorType::get(Ty->getElementType(), NumElements);
1640 Type::getIntNTy(Ty->getContext(), NumElements * ElementSize * 8);
1642 Use *U = S.getUse();
1644 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1645 if (MI->isVolatile())
1647 if (!S.isSplittable())
1648 return false; // Skip any unsplittable intrinsics.
1649 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1650 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1651 II->getIntrinsicID() != Intrinsic::lifetime_end)
1653 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
1654 // Disable vector promotion when there are loads or stores of an FCA.
1656 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1657 if (LI->isVolatile())
1659 Type *LTy = LI->getType();
1660 if (SliceBeginOffset > S.beginOffset() || SliceEndOffset < S.endOffset()) {
1661 assert(LTy->isIntegerTy());
1664 if (!canConvertValue(DL, SliceTy, LTy))
1666 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1667 if (SI->isVolatile())
1669 Type *STy = SI->getValueOperand()->getType();
1670 if (SliceBeginOffset > S.beginOffset() || SliceEndOffset < S.endOffset()) {
1671 assert(STy->isIntegerTy());
1674 if (!canConvertValue(DL, STy, SliceTy))
1683 /// \brief Test whether the given alloca partitioning and range of slices can be
1684 /// promoted to a vector.
1686 /// This is a quick test to check whether we can rewrite a particular alloca
1687 /// partition (and its newly formed alloca) into a vector alloca with only
1688 /// whole-vector loads and stores such that it could be promoted to a vector
1689 /// SSA value. We only can ensure this for a limited set of operations, and we
1690 /// don't want to do the rewrites unless we are confident that the result will
1691 /// be promotable, so we have an early test here.
1693 isVectorPromotionViable(const DataLayout &DL, uint64_t SliceBeginOffset,
1694 uint64_t SliceEndOffset,
1695 AllocaSlices::const_range Slices,
1696 ArrayRef<AllocaSlices::iterator> SplitUses) {
1697 // Collect the candidate types for vector-based promotion. Also track whether
1698 // we have different element types.
1699 SmallVector<VectorType *, 4> CandidateTys;
1700 Type *CommonEltTy = nullptr;
1701 bool HaveCommonEltTy = true;
1702 auto CheckCandidateType = [&](Type *Ty) {
1703 if (auto *VTy = dyn_cast<VectorType>(Ty)) {
1704 CandidateTys.push_back(VTy);
1706 CommonEltTy = VTy->getElementType();
1707 else if (CommonEltTy != VTy->getElementType())
1708 HaveCommonEltTy = false;
1711 // Consider any loads or stores that are the exact size of the slice.
1712 for (const auto &S : Slices)
1713 if (S.beginOffset() == SliceBeginOffset &&
1714 S.endOffset() == SliceEndOffset) {
1715 if (auto *LI = dyn_cast<LoadInst>(S.getUse()->getUser()))
1716 CheckCandidateType(LI->getType());
1717 else if (auto *SI = dyn_cast<StoreInst>(S.getUse()->getUser()))
1718 CheckCandidateType(SI->getValueOperand()->getType());
1721 // If we didn't find a vector type, nothing to do here.
1722 if (CandidateTys.empty())
1725 // Remove non-integer vector types if we had multiple common element types.
1726 // FIXME: It'd be nice to replace them with integer vector types, but we can't
1727 // do that until all the backends are known to produce good code for all
1728 // integer vector types.
1729 if (!HaveCommonEltTy) {
1730 CandidateTys.erase(std::remove_if(CandidateTys.begin(), CandidateTys.end(),
1731 [](VectorType *VTy) {
1732 return !VTy->getElementType()->isIntegerTy();
1734 CandidateTys.end());
1736 // If there were no integer vector types, give up.
1737 if (CandidateTys.empty())
1740 // Rank the remaining candidate vector types. This is easy because we know
1741 // they're all integer vectors. We sort by ascending number of elements.
1742 auto RankVectorTypes = [&DL](VectorType *RHSTy, VectorType *LHSTy) {
1743 assert(DL.getTypeSizeInBits(RHSTy) == DL.getTypeSizeInBits(LHSTy) &&
1744 "Cannot have vector types of different sizes!");
1745 assert(RHSTy->getElementType()->isIntegerTy() &&
1746 "All non-integer types eliminated!");
1747 assert(LHSTy->getElementType()->isIntegerTy() &&
1748 "All non-integer types eliminated!");
1749 return RHSTy->getNumElements() < LHSTy->getNumElements();
1751 std::sort(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes);
1753 std::unique(CandidateTys.begin(), CandidateTys.end(), RankVectorTypes),
1754 CandidateTys.end());
1756 // The only way to have the same element type in every vector type is to
1757 // have the same vector type. Check that and remove all but one.
1759 for (VectorType *VTy : CandidateTys) {
1760 assert(VTy->getElementType() == CommonEltTy &&
1761 "Unaccounted for element type!");
1762 assert(VTy == CandidateTys[0] &&
1763 "Different vector types with the same element type!");
1766 CandidateTys.resize(1);
1769 // Try each vector type, and return the one which works.
1770 auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
1771 uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
1773 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1774 // that aren't byte sized.
1775 if (ElementSize % 8)
1777 assert((DL.getTypeSizeInBits(VTy) % 8) == 0 &&
1778 "vector size not a multiple of element size?");
1781 for (const auto &S : Slices)
1782 if (!isVectorPromotionViableForSlice(DL, SliceBeginOffset, SliceEndOffset,
1783 VTy, ElementSize, S))
1786 for (const auto &SI : SplitUses)
1787 if (!isVectorPromotionViableForSlice(DL, SliceBeginOffset, SliceEndOffset,
1788 VTy, ElementSize, *SI))
1793 for (VectorType *VTy : CandidateTys)
1794 if (CheckVectorTypeForPromotion(VTy))
1800 /// \brief Test whether a slice of an alloca is valid for integer widening.
1802 /// This implements the necessary checking for the \c isIntegerWideningViable
1803 /// test below on a single slice of the alloca.
1804 static bool isIntegerWideningViableForSlice(const DataLayout &DL,
1806 uint64_t AllocBeginOffset,
1807 uint64_t Size, const Slice &S,
1808 bool &WholeAllocaOp) {
1809 uint64_t RelBegin = S.beginOffset() - AllocBeginOffset;
1810 uint64_t RelEnd = S.endOffset() - AllocBeginOffset;
1812 // We can't reasonably handle cases where the load or store extends past
1813 // the end of the aloca's type and into its padding.
1817 Use *U = S.getUse();
1819 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
1820 if (LI->isVolatile())
1822 // Note that we don't count vector loads or stores as whole-alloca
1823 // operations which enable integer widening because we would prefer to use
1824 // vector widening instead.
1825 if (!isa<VectorType>(LI->getType()) && RelBegin == 0 && RelEnd == Size)
1826 WholeAllocaOp = true;
1827 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
1828 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1830 } else if (RelBegin != 0 || RelEnd != Size ||
1831 !canConvertValue(DL, AllocaTy, LI->getType())) {
1832 // Non-integer loads need to be convertible from the alloca type so that
1833 // they are promotable.
1836 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
1837 Type *ValueTy = SI->getValueOperand()->getType();
1838 if (SI->isVolatile())
1840 // Note that we don't count vector loads or stores as whole-alloca
1841 // operations which enable integer widening because we would prefer to use
1842 // vector widening instead.
1843 if (!isa<VectorType>(ValueTy) && RelBegin == 0 && RelEnd == Size)
1844 WholeAllocaOp = true;
1845 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
1846 if (ITy->getBitWidth() < DL.getTypeStoreSizeInBits(ITy))
1848 } else if (RelBegin != 0 || RelEnd != Size ||
1849 !canConvertValue(DL, ValueTy, AllocaTy)) {
1850 // Non-integer stores need to be convertible to the alloca type so that
1851 // they are promotable.
1854 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
1855 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
1857 if (!S.isSplittable())
1858 return false; // Skip any unsplittable intrinsics.
1859 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
1860 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
1861 II->getIntrinsicID() != Intrinsic::lifetime_end)
1870 /// \brief Test whether the given alloca partition's integer operations can be
1871 /// widened to promotable ones.
1873 /// This is a quick test to check whether we can rewrite the integer loads and
1874 /// stores to a particular alloca into wider loads and stores and be able to
1875 /// promote the resulting alloca.
1877 isIntegerWideningViable(const DataLayout &DL, Type *AllocaTy,
1878 uint64_t AllocBeginOffset,
1879 AllocaSlices::const_range Slices,
1880 ArrayRef<AllocaSlices::iterator> SplitUses) {
1881 uint64_t SizeInBits = DL.getTypeSizeInBits(AllocaTy);
1882 // Don't create integer types larger than the maximum bitwidth.
1883 if (SizeInBits > IntegerType::MAX_INT_BITS)
1886 // Don't try to handle allocas with bit-padding.
1887 if (SizeInBits != DL.getTypeStoreSizeInBits(AllocaTy))
1890 // We need to ensure that an integer type with the appropriate bitwidth can
1891 // be converted to the alloca type, whatever that is. We don't want to force
1892 // the alloca itself to have an integer type if there is a more suitable one.
1893 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
1894 if (!canConvertValue(DL, AllocaTy, IntTy) ||
1895 !canConvertValue(DL, IntTy, AllocaTy))
1898 uint64_t Size = DL.getTypeStoreSize(AllocaTy);
1900 // While examining uses, we ensure that the alloca has a covering load or
1901 // store. We don't want to widen the integer operations only to fail to
1902 // promote due to some other unsplittable entry (which we may make splittable
1903 // later). However, if there are only splittable uses, go ahead and assume
1904 // that we cover the alloca.
1905 bool WholeAllocaOp =
1906 Slices.begin() != Slices.end() ? false : DL.isLegalInteger(SizeInBits);
1908 for (const auto &S : Slices)
1909 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1913 for (const auto &SI : SplitUses)
1914 if (!isIntegerWideningViableForSlice(DL, AllocaTy, AllocBeginOffset, Size,
1915 *SI, WholeAllocaOp))
1918 return WholeAllocaOp;
1921 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
1922 IntegerType *Ty, uint64_t Offset,
1923 const Twine &Name) {
1924 DEBUG(dbgs() << " start: " << *V << "\n");
1925 IntegerType *IntTy = cast<IntegerType>(V->getType());
1926 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1927 "Element extends past full value");
1928 uint64_t ShAmt = 8 * Offset;
1929 if (DL.isBigEndian())
1930 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1932 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
1933 DEBUG(dbgs() << " shifted: " << *V << "\n");
1935 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1936 "Cannot extract to a larger integer!");
1938 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
1939 DEBUG(dbgs() << " trunced: " << *V << "\n");
1944 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
1945 Value *V, uint64_t Offset, const Twine &Name) {
1946 IntegerType *IntTy = cast<IntegerType>(Old->getType());
1947 IntegerType *Ty = cast<IntegerType>(V->getType());
1948 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
1949 "Cannot insert a larger integer!");
1950 DEBUG(dbgs() << " start: " << *V << "\n");
1952 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
1953 DEBUG(dbgs() << " extended: " << *V << "\n");
1955 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
1956 "Element store outside of alloca store");
1957 uint64_t ShAmt = 8 * Offset;
1958 if (DL.isBigEndian())
1959 ShAmt = 8 * (DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
1961 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
1962 DEBUG(dbgs() << " shifted: " << *V << "\n");
1965 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
1966 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
1967 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
1968 DEBUG(dbgs() << " masked: " << *Old << "\n");
1969 V = IRB.CreateOr(Old, V, Name + ".insert");
1970 DEBUG(dbgs() << " inserted: " << *V << "\n");
1975 static Value *extractVector(IRBuilderTy &IRB, Value *V, unsigned BeginIndex,
1976 unsigned EndIndex, const Twine &Name) {
1977 VectorType *VecTy = cast<VectorType>(V->getType());
1978 unsigned NumElements = EndIndex - BeginIndex;
1979 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
1981 if (NumElements == VecTy->getNumElements())
1984 if (NumElements == 1) {
1985 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
1987 DEBUG(dbgs() << " extract: " << *V << "\n");
1991 SmallVector<Constant *, 8> Mask;
1992 Mask.reserve(NumElements);
1993 for (unsigned i = BeginIndex; i != EndIndex; ++i)
1994 Mask.push_back(IRB.getInt32(i));
1995 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
1996 ConstantVector::get(Mask), Name + ".extract");
1997 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2001 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2002 unsigned BeginIndex, const Twine &Name) {
2003 VectorType *VecTy = cast<VectorType>(Old->getType());
2004 assert(VecTy && "Can only insert a vector into a vector");
2006 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2008 // Single element to insert.
2009 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2011 DEBUG(dbgs() << " insert: " << *V << "\n");
2015 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2016 "Too many elements!");
2017 if (Ty->getNumElements() == VecTy->getNumElements()) {
2018 assert(V->getType() == VecTy && "Vector type mismatch");
2021 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2023 // When inserting a smaller vector into the larger to store, we first
2024 // use a shuffle vector to widen it with undef elements, and then
2025 // a second shuffle vector to select between the loaded vector and the
2027 SmallVector<Constant *, 8> Mask;
2028 Mask.reserve(VecTy->getNumElements());
2029 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2030 if (i >= BeginIndex && i < EndIndex)
2031 Mask.push_back(IRB.getInt32(i - BeginIndex));
2033 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2034 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2035 ConstantVector::get(Mask), Name + ".expand");
2036 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2039 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2040 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2042 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2044 DEBUG(dbgs() << " blend: " << *V << "\n");
2049 /// \brief Visitor to rewrite instructions using p particular slice of an alloca
2050 /// to use a new alloca.
2052 /// Also implements the rewriting to vector-based accesses when the partition
2053 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2055 class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
2056 // Befriend the base class so it can delegate to private visit methods.
2057 friend class llvm::InstVisitor<AllocaSliceRewriter, bool>;
2058 typedef llvm::InstVisitor<AllocaSliceRewriter, bool> Base;
2060 const DataLayout &DL;
2063 AllocaInst &OldAI, &NewAI;
2064 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2067 // This is a convenience and flag variable that will be null unless the new
2068 // alloca's integer operations should be widened to this integer type due to
2069 // passing isIntegerWideningViable above. If it is non-null, the desired
2070 // integer type will be stored here for easy access during rewriting.
2073 // If we are rewriting an alloca partition which can be written as pure
2074 // vector operations, we stash extra information here. When VecTy is
2075 // non-null, we have some strict guarantees about the rewritten alloca:
2076 // - The new alloca is exactly the size of the vector type here.
2077 // - The accesses all either map to the entire vector or to a single
2079 // - The set of accessing instructions is only one of those handled above
2080 // in isVectorPromotionViable. Generally these are the same access kinds
2081 // which are promotable via mem2reg.
2084 uint64_t ElementSize;
2086 // The original offset of the slice currently being rewritten relative to
2087 // the original alloca.
2088 uint64_t BeginOffset, EndOffset;
2089 // The new offsets of the slice currently being rewritten relative to the
2091 uint64_t NewBeginOffset, NewEndOffset;
2097 Instruction *OldPtr;
2099 // Track post-rewrite users which are PHI nodes and Selects.
2100 SmallPtrSetImpl<PHINode *> &PHIUsers;
2101 SmallPtrSetImpl<SelectInst *> &SelectUsers;
2103 // Utility IR builder, whose name prefix is setup for each visited use, and
2104 // the insertion point is set to point to the user.
2108 AllocaSliceRewriter(const DataLayout &DL, AllocaSlices &AS, SROA &Pass,
2109 AllocaInst &OldAI, AllocaInst &NewAI,
2110 uint64_t NewAllocaBeginOffset,
2111 uint64_t NewAllocaEndOffset, bool IsIntegerPromotable,
2112 VectorType *PromotableVecTy,
2113 SmallPtrSetImpl<PHINode *> &PHIUsers,
2114 SmallPtrSetImpl<SelectInst *> &SelectUsers)
2115 : DL(DL), AS(AS), Pass(Pass), OldAI(OldAI), NewAI(NewAI),
2116 NewAllocaBeginOffset(NewAllocaBeginOffset),
2117 NewAllocaEndOffset(NewAllocaEndOffset),
2118 NewAllocaTy(NewAI.getAllocatedType()),
2119 IntTy(IsIntegerPromotable
2122 DL.getTypeSizeInBits(NewAI.getAllocatedType()))
2124 VecTy(PromotableVecTy),
2125 ElementTy(VecTy ? VecTy->getElementType() : nullptr),
2126 ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
2127 BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
2128 OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
2129 IRB(NewAI.getContext(), ConstantFolder()) {
2131 assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
2132 "Only multiple-of-8 sized vector elements are viable");
2135 assert((!IntTy && !VecTy) || (IntTy && !VecTy) || (!IntTy && VecTy));
2138 bool visit(AllocaSlices::const_iterator I) {
2139 bool CanSROA = true;
2140 BeginOffset = I->beginOffset();
2141 EndOffset = I->endOffset();
2142 IsSplittable = I->isSplittable();
2144 BeginOffset < NewAllocaBeginOffset || EndOffset > NewAllocaEndOffset;
2146 // Compute the intersecting offset range.
2147 assert(BeginOffset < NewAllocaEndOffset);
2148 assert(EndOffset > NewAllocaBeginOffset);
2149 NewBeginOffset = std::max(BeginOffset, NewAllocaBeginOffset);
2150 NewEndOffset = std::min(EndOffset, NewAllocaEndOffset);
2152 SliceSize = NewEndOffset - NewBeginOffset;
2154 OldUse = I->getUse();
2155 OldPtr = cast<Instruction>(OldUse->get());
2157 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2158 IRB.SetInsertPoint(OldUserI);
2159 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2160 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) + ".");
2162 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2169 // Make sure the other visit overloads are visible.
2172 // Every instruction which can end up as a user must have a rewrite rule.
2173 bool visitInstruction(Instruction &I) {
2174 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2175 llvm_unreachable("No rewrite rule for this instruction!");
2178 Value *getNewAllocaSlicePtr(IRBuilderTy &IRB, Type *PointerTy) {
2179 // Note that the offset computation can use BeginOffset or NewBeginOffset
2180 // interchangeably for unsplit slices.
2181 assert(IsSplit || BeginOffset == NewBeginOffset);
2182 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2185 StringRef OldName = OldPtr->getName();
2186 // Skip through the last '.sroa.' component of the name.
2187 size_t LastSROAPrefix = OldName.rfind(".sroa.");
2188 if (LastSROAPrefix != StringRef::npos) {
2189 OldName = OldName.substr(LastSROAPrefix + strlen(".sroa."));
2190 // Look for an SROA slice index.
2191 size_t IndexEnd = OldName.find_first_not_of("0123456789");
2192 if (IndexEnd != StringRef::npos && OldName[IndexEnd] == '.') {
2193 // Strip the index and look for the offset.
2194 OldName = OldName.substr(IndexEnd + 1);
2195 size_t OffsetEnd = OldName.find_first_not_of("0123456789");
2196 if (OffsetEnd != StringRef::npos && OldName[OffsetEnd] == '.')
2197 // Strip the offset.
2198 OldName = OldName.substr(OffsetEnd + 1);
2201 // Strip any SROA suffixes as well.
2202 OldName = OldName.substr(0, OldName.find(".sroa_"));
2205 return getAdjustedPtr(IRB, DL, &NewAI,
2206 APInt(DL.getPointerSizeInBits(), Offset), PointerTy,
2208 Twine(OldName) + "."
2215 /// \brief Compute suitable alignment to access this slice of the *new*
2218 /// You can optionally pass a type to this routine and if that type's ABI
2219 /// alignment is itself suitable, this will return zero.
2220 unsigned getSliceAlign(Type *Ty = nullptr) {
2221 unsigned NewAIAlign = NewAI.getAlignment();
2223 NewAIAlign = DL.getABITypeAlignment(NewAI.getAllocatedType());
2225 MinAlign(NewAIAlign, NewBeginOffset - NewAllocaBeginOffset);
2226 return (Ty && Align == DL.getABITypeAlignment(Ty)) ? 0 : Align;
2229 unsigned getIndex(uint64_t Offset) {
2230 assert(VecTy && "Can only call getIndex when rewriting a vector");
2231 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2232 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2233 uint32_t Index = RelOffset / ElementSize;
2234 assert(Index * ElementSize == RelOffset);
2238 void deleteIfTriviallyDead(Value *V) {
2239 Instruction *I = cast<Instruction>(V);
2240 if (isInstructionTriviallyDead(I))
2241 Pass.DeadInsts.insert(I);
2244 Value *rewriteVectorizedLoadInst() {
2245 unsigned BeginIndex = getIndex(NewBeginOffset);
2246 unsigned EndIndex = getIndex(NewEndOffset);
2247 assert(EndIndex > BeginIndex && "Empty vector!");
2249 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2250 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2253 Value *rewriteIntegerLoad(LoadInst &LI) {
2254 assert(IntTy && "We cannot insert an integer to the alloca");
2255 assert(!LI.isVolatile());
2256 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2257 V = convertValue(DL, IRB, V, IntTy);
2258 assert(NewBeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2259 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2260 if (Offset > 0 || NewEndOffset < NewAllocaEndOffset)
2261 V = extractInteger(DL, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2266 bool visitLoadInst(LoadInst &LI) {
2267 DEBUG(dbgs() << " original: " << LI << "\n");
2268 Value *OldOp = LI.getOperand(0);
2269 assert(OldOp == OldPtr);
2271 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), SliceSize * 8)
2273 bool IsPtrAdjusted = false;
2276 V = rewriteVectorizedLoadInst();
2277 } else if (IntTy && LI.getType()->isIntegerTy()) {
2278 V = rewriteIntegerLoad(LI);
2279 } else if (NewBeginOffset == NewAllocaBeginOffset &&
2280 canConvertValue(DL, NewAllocaTy, LI.getType())) {
2281 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), LI.isVolatile(),
2284 Type *LTy = TargetTy->getPointerTo();
2285 V = IRB.CreateAlignedLoad(getNewAllocaSlicePtr(IRB, LTy),
2286 getSliceAlign(TargetTy), LI.isVolatile(),
2288 IsPtrAdjusted = true;
2290 V = convertValue(DL, IRB, V, TargetTy);
2293 assert(!LI.isVolatile());
2294 assert(LI.getType()->isIntegerTy() &&
2295 "Only integer type loads and stores are split");
2296 assert(SliceSize < DL.getTypeStoreSize(LI.getType()) &&
2297 "Split load isn't smaller than original load");
2298 assert(LI.getType()->getIntegerBitWidth() ==
2299 DL.getTypeStoreSizeInBits(LI.getType()) &&
2300 "Non-byte-multiple bit width");
2301 // Move the insertion point just past the load so that we can refer to it.
2302 IRB.SetInsertPoint(std::next(BasicBlock::iterator(&LI)));
2303 // Create a placeholder value with the same type as LI to use as the
2304 // basis for the new value. This allows us to replace the uses of LI with
2305 // the computed value, and then replace the placeholder with LI, leaving
2306 // LI only used for this computation.
2307 Value *Placeholder =
2308 new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2309 V = insertInteger(DL, IRB, Placeholder, V, NewBeginOffset, "insert");
2310 LI.replaceAllUsesWith(V);
2311 Placeholder->replaceAllUsesWith(&LI);
2314 LI.replaceAllUsesWith(V);
2317 Pass.DeadInsts.insert(&LI);
2318 deleteIfTriviallyDead(OldOp);
2319 DEBUG(dbgs() << " to: " << *V << "\n");
2320 return !LI.isVolatile() && !IsPtrAdjusted;
2323 bool rewriteVectorizedStoreInst(Value *V, StoreInst &SI, Value *OldOp) {
2324 if (V->getType() != VecTy) {
2325 unsigned BeginIndex = getIndex(NewBeginOffset);
2326 unsigned EndIndex = getIndex(NewEndOffset);
2327 assert(EndIndex > BeginIndex && "Empty vector!");
2328 unsigned NumElements = EndIndex - BeginIndex;
2329 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2330 Type *SliceTy = (NumElements == 1)
2332 : VectorType::get(ElementTy, NumElements);
2333 if (V->getType() != SliceTy)
2334 V = convertValue(DL, IRB, V, SliceTy);
2336 // Mix in the existing elements.
2337 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2338 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2340 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2341 Pass.DeadInsts.insert(&SI);
2344 DEBUG(dbgs() << " to: " << *Store << "\n");
2348 bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2349 assert(IntTy && "We cannot extract an integer from the alloca");
2350 assert(!SI.isVolatile());
2351 if (DL.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2353 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2354 Old = convertValue(DL, IRB, Old, IntTy);
2355 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2356 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2357 V = insertInteger(DL, IRB, Old, SI.getValueOperand(), Offset, "insert");
2359 V = convertValue(DL, IRB, V, NewAllocaTy);
2360 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2361 Pass.DeadInsts.insert(&SI);
2363 DEBUG(dbgs() << " to: " << *Store << "\n");
2367 bool visitStoreInst(StoreInst &SI) {
2368 DEBUG(dbgs() << " original: " << SI << "\n");
2369 Value *OldOp = SI.getOperand(1);
2370 assert(OldOp == OldPtr);
2372 Value *V = SI.getValueOperand();
2374 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2375 // alloca that should be re-examined after promoting this alloca.
2376 if (V->getType()->isPointerTy())
2377 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2378 Pass.PostPromotionWorklist.insert(AI);
2380 if (SliceSize < DL.getTypeStoreSize(V->getType())) {
2381 assert(!SI.isVolatile());
2382 assert(V->getType()->isIntegerTy() &&
2383 "Only integer type loads and stores are split");
2384 assert(V->getType()->getIntegerBitWidth() ==
2385 DL.getTypeStoreSizeInBits(V->getType()) &&
2386 "Non-byte-multiple bit width");
2387 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), SliceSize * 8);
2388 V = extractInteger(DL, IRB, V, NarrowTy, NewBeginOffset, "extract");
2392 return rewriteVectorizedStoreInst(V, SI, OldOp);
2393 if (IntTy && V->getType()->isIntegerTy())
2394 return rewriteIntegerStore(V, SI);
2397 if (NewBeginOffset == NewAllocaBeginOffset &&
2398 NewEndOffset == NewAllocaEndOffset &&
2399 canConvertValue(DL, V->getType(), NewAllocaTy)) {
2400 V = convertValue(DL, IRB, V, NewAllocaTy);
2401 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2404 Value *NewPtr = getNewAllocaSlicePtr(IRB, V->getType()->getPointerTo());
2405 NewSI = IRB.CreateAlignedStore(V, NewPtr, getSliceAlign(V->getType()),
2409 Pass.DeadInsts.insert(&SI);
2410 deleteIfTriviallyDead(OldOp);
2412 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2413 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2416 /// \brief Compute an integer value from splatting an i8 across the given
2417 /// number of bytes.
2419 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2420 /// call this routine.
2421 /// FIXME: Heed the advice above.
2423 /// \param V The i8 value to splat.
2424 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2425 Value *getIntegerSplat(Value *V, unsigned Size) {
2426 assert(Size > 0 && "Expected a positive number of bytes.");
2427 IntegerType *VTy = cast<IntegerType>(V->getType());
2428 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2432 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size * 8);
2434 IRB.CreateZExt(V, SplatIntTy, "zext"),
2435 ConstantExpr::getUDiv(
2436 Constant::getAllOnesValue(SplatIntTy),
2437 ConstantExpr::getZExt(Constant::getAllOnesValue(V->getType()),
2443 /// \brief Compute a vector splat for a given element value.
2444 Value *getVectorSplat(Value *V, unsigned NumElements) {
2445 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2446 DEBUG(dbgs() << " splat: " << *V << "\n");
2450 bool visitMemSetInst(MemSetInst &II) {
2451 DEBUG(dbgs() << " original: " << II << "\n");
2452 assert(II.getRawDest() == OldPtr);
2454 // If the memset has a variable size, it cannot be split, just adjust the
2455 // pointer to the new alloca.
2456 if (!isa<Constant>(II.getLength())) {
2458 assert(NewBeginOffset == BeginOffset);
2459 II.setDest(getNewAllocaSlicePtr(IRB, OldPtr->getType()));
2460 Type *CstTy = II.getAlignmentCst()->getType();
2461 II.setAlignment(ConstantInt::get(CstTy, getSliceAlign()));
2463 deleteIfTriviallyDead(OldPtr);
2467 // Record this instruction for deletion.
2468 Pass.DeadInsts.insert(&II);
2470 Type *AllocaTy = NewAI.getAllocatedType();
2471 Type *ScalarTy = AllocaTy->getScalarType();
2473 // If this doesn't map cleanly onto the alloca type, and that type isn't
2474 // a single value type, just emit a memset.
2475 if (!VecTy && !IntTy &&
2476 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2477 SliceSize != DL.getTypeStoreSize(AllocaTy) ||
2478 !AllocaTy->isSingleValueType() ||
2479 !DL.isLegalInteger(DL.getTypeSizeInBits(ScalarTy)) ||
2480 DL.getTypeSizeInBits(ScalarTy) % 8 != 0)) {
2481 Type *SizeTy = II.getLength()->getType();
2482 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2483 CallInst *New = IRB.CreateMemSet(
2484 getNewAllocaSlicePtr(IRB, OldPtr->getType()), II.getValue(), Size,
2485 getSliceAlign(), II.isVolatile());
2487 DEBUG(dbgs() << " to: " << *New << "\n");
2491 // If we can represent this as a simple value, we have to build the actual
2492 // value to store, which requires expanding the byte present in memset to
2493 // a sensible representation for the alloca type. This is essentially
2494 // splatting the byte to a sufficiently wide integer, splatting it across
2495 // any desired vector width, and bitcasting to the final type.
2499 // If this is a memset of a vectorized alloca, insert it.
2500 assert(ElementTy == ScalarTy);
2502 unsigned BeginIndex = getIndex(NewBeginOffset);
2503 unsigned EndIndex = getIndex(NewEndOffset);
2504 assert(EndIndex > BeginIndex && "Empty vector!");
2505 unsigned NumElements = EndIndex - BeginIndex;
2506 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2509 getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ElementTy) / 8);
2510 Splat = convertValue(DL, IRB, Splat, ElementTy);
2511 if (NumElements > 1)
2512 Splat = getVectorSplat(Splat, NumElements);
2515 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2516 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2518 // If this is a memset on an alloca where we can widen stores, insert the
2520 assert(!II.isVolatile());
2522 uint64_t Size = NewEndOffset - NewBeginOffset;
2523 V = getIntegerSplat(II.getValue(), Size);
2525 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2526 EndOffset != NewAllocaBeginOffset)) {
2528 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2529 Old = convertValue(DL, IRB, Old, IntTy);
2530 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2531 V = insertInteger(DL, IRB, Old, V, Offset, "insert");
2533 assert(V->getType() == IntTy &&
2534 "Wrong type for an alloca wide integer!");
2536 V = convertValue(DL, IRB, V, AllocaTy);
2538 // Established these invariants above.
2539 assert(NewBeginOffset == NewAllocaBeginOffset);
2540 assert(NewEndOffset == NewAllocaEndOffset);
2542 V = getIntegerSplat(II.getValue(), DL.getTypeSizeInBits(ScalarTy) / 8);
2543 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2544 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2546 V = convertValue(DL, IRB, V, AllocaTy);
2549 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2552 DEBUG(dbgs() << " to: " << *New << "\n");
2553 return !II.isVolatile();
2556 bool visitMemTransferInst(MemTransferInst &II) {
2557 // Rewriting of memory transfer instructions can be a bit tricky. We break
2558 // them into two categories: split intrinsics and unsplit intrinsics.
2560 DEBUG(dbgs() << " original: " << II << "\n");
2562 bool IsDest = &II.getRawDestUse() == OldUse;
2563 assert((IsDest && II.getRawDest() == OldPtr) ||
2564 (!IsDest && II.getRawSource() == OldPtr));
2566 unsigned SliceAlign = getSliceAlign();
2568 // For unsplit intrinsics, we simply modify the source and destination
2569 // pointers in place. This isn't just an optimization, it is a matter of
2570 // correctness. With unsplit intrinsics we may be dealing with transfers
2571 // within a single alloca before SROA ran, or with transfers that have
2572 // a variable length. We may also be dealing with memmove instead of
2573 // memcpy, and so simply updating the pointers is the necessary for us to
2574 // update both source and dest of a single call.
2575 if (!IsSplittable) {
2576 Value *AdjustedPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2578 II.setDest(AdjustedPtr);
2580 II.setSource(AdjustedPtr);
2582 if (II.getAlignment() > SliceAlign) {
2583 Type *CstTy = II.getAlignmentCst()->getType();
2585 ConstantInt::get(CstTy, MinAlign(II.getAlignment(), SliceAlign)));
2588 DEBUG(dbgs() << " to: " << II << "\n");
2589 deleteIfTriviallyDead(OldPtr);
2592 // For split transfer intrinsics we have an incredibly useful assurance:
2593 // the source and destination do not reside within the same alloca, and at
2594 // least one of them does not escape. This means that we can replace
2595 // memmove with memcpy, and we don't need to worry about all manner of
2596 // downsides to splitting and transforming the operations.
2598 // If this doesn't map cleanly onto the alloca type, and that type isn't
2599 // a single value type, just emit a memcpy.
2602 (BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset ||
2603 SliceSize != DL.getTypeStoreSize(NewAI.getAllocatedType()) ||
2604 !NewAI.getAllocatedType()->isSingleValueType());
2606 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2607 // size hasn't been shrunk based on analysis of the viable range, this is
2609 if (EmitMemCpy && &OldAI == &NewAI) {
2610 // Ensure the start lines up.
2611 assert(NewBeginOffset == BeginOffset);
2613 // Rewrite the size as needed.
2614 if (NewEndOffset != EndOffset)
2615 II.setLength(ConstantInt::get(II.getLength()->getType(),
2616 NewEndOffset - NewBeginOffset));
2619 // Record this instruction for deletion.
2620 Pass.DeadInsts.insert(&II);
2622 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2623 // alloca that should be re-examined after rewriting this instruction.
2624 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2625 if (AllocaInst *AI =
2626 dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) {
2627 assert(AI != &OldAI && AI != &NewAI &&
2628 "Splittable transfers cannot reach the same alloca on both ends.");
2629 Pass.Worklist.insert(AI);
2632 Type *OtherPtrTy = OtherPtr->getType();
2633 unsigned OtherAS = OtherPtrTy->getPointerAddressSpace();
2635 // Compute the relative offset for the other pointer within the transfer.
2636 unsigned IntPtrWidth = DL.getPointerSizeInBits(OtherAS);
2637 APInt OtherOffset(IntPtrWidth, NewBeginOffset - BeginOffset);
2638 unsigned OtherAlign = MinAlign(II.getAlignment() ? II.getAlignment() : 1,
2639 OtherOffset.zextOrTrunc(64).getZExtValue());
2642 // Compute the other pointer, folding as much as possible to produce
2643 // a single, simple GEP in most cases.
2644 OtherPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2645 OtherPtr->getName() + ".");
2647 Value *OurPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2648 Type *SizeTy = II.getLength()->getType();
2649 Constant *Size = ConstantInt::get(SizeTy, NewEndOffset - NewBeginOffset);
2651 CallInst *New = IRB.CreateMemCpy(
2652 IsDest ? OurPtr : OtherPtr, IsDest ? OtherPtr : OurPtr, Size,
2653 MinAlign(SliceAlign, OtherAlign), II.isVolatile());
2655 DEBUG(dbgs() << " to: " << *New << "\n");
2659 bool IsWholeAlloca = NewBeginOffset == NewAllocaBeginOffset &&
2660 NewEndOffset == NewAllocaEndOffset;
2661 uint64_t Size = NewEndOffset - NewBeginOffset;
2662 unsigned BeginIndex = VecTy ? getIndex(NewBeginOffset) : 0;
2663 unsigned EndIndex = VecTy ? getIndex(NewEndOffset) : 0;
2664 unsigned NumElements = EndIndex - BeginIndex;
2665 IntegerType *SubIntTy =
2666 IntTy ? Type::getIntNTy(IntTy->getContext(), Size * 8) : nullptr;
2668 // Reset the other pointer type to match the register type we're going to
2669 // use, but using the address space of the original other pointer.
2670 if (VecTy && !IsWholeAlloca) {
2671 if (NumElements == 1)
2672 OtherPtrTy = VecTy->getElementType();
2674 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2676 OtherPtrTy = OtherPtrTy->getPointerTo(OtherAS);
2677 } else if (IntTy && !IsWholeAlloca) {
2678 OtherPtrTy = SubIntTy->getPointerTo(OtherAS);
2680 OtherPtrTy = NewAllocaTy->getPointerTo(OtherAS);
2683 Value *SrcPtr = getAdjustedPtr(IRB, DL, OtherPtr, OtherOffset, OtherPtrTy,
2684 OtherPtr->getName() + ".");
2685 unsigned SrcAlign = OtherAlign;
2686 Value *DstPtr = &NewAI;
2687 unsigned DstAlign = SliceAlign;
2689 std::swap(SrcPtr, DstPtr);
2690 std::swap(SrcAlign, DstAlign);
2694 if (VecTy && !IsWholeAlloca && !IsDest) {
2695 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2696 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2697 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2698 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "load");
2699 Src = convertValue(DL, IRB, Src, IntTy);
2700 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2701 Src = extractInteger(DL, IRB, Src, SubIntTy, Offset, "extract");
2704 IRB.CreateAlignedLoad(SrcPtr, SrcAlign, II.isVolatile(), "copyload");
2707 if (VecTy && !IsWholeAlloca && IsDest) {
2709 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2710 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2711 } else if (IntTy && !IsWholeAlloca && IsDest) {
2713 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), "oldload");
2714 Old = convertValue(DL, IRB, Old, IntTy);
2715 uint64_t Offset = NewBeginOffset - NewAllocaBeginOffset;
2716 Src = insertInteger(DL, IRB, Old, Src, Offset, "insert");
2717 Src = convertValue(DL, IRB, Src, NewAllocaTy);
2720 StoreInst *Store = cast<StoreInst>(
2721 IRB.CreateAlignedStore(Src, DstPtr, DstAlign, II.isVolatile()));
2723 DEBUG(dbgs() << " to: " << *Store << "\n");
2724 return !II.isVolatile();
2727 bool visitIntrinsicInst(IntrinsicInst &II) {
2728 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2729 II.getIntrinsicID() == Intrinsic::lifetime_end);
2730 DEBUG(dbgs() << " original: " << II << "\n");
2731 assert(II.getArgOperand(1) == OldPtr);
2733 // Record this instruction for deletion.
2734 Pass.DeadInsts.insert(&II);
2737 ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2738 NewEndOffset - NewBeginOffset);
2739 Value *Ptr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2741 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2742 New = IRB.CreateLifetimeStart(Ptr, Size);
2744 New = IRB.CreateLifetimeEnd(Ptr, Size);
2747 DEBUG(dbgs() << " to: " << *New << "\n");
2751 bool visitPHINode(PHINode &PN) {
2752 DEBUG(dbgs() << " original: " << PN << "\n");
2753 assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
2754 assert(EndOffset <= NewAllocaEndOffset && "PHIs are unsplittable");
2756 // We would like to compute a new pointer in only one place, but have it be
2757 // as local as possible to the PHI. To do that, we re-use the location of
2758 // the old pointer, which necessarily must be in the right position to
2759 // dominate the PHI.
2760 IRBuilderTy PtrBuilder(IRB);
2761 if (isa<PHINode>(OldPtr))
2762 PtrBuilder.SetInsertPoint(OldPtr->getParent()->getFirstInsertionPt());
2764 PtrBuilder.SetInsertPoint(OldPtr);
2765 PtrBuilder.SetCurrentDebugLocation(OldPtr->getDebugLoc());
2767 Value *NewPtr = getNewAllocaSlicePtr(PtrBuilder, OldPtr->getType());
2768 // Replace the operands which were using the old pointer.
2769 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2771 DEBUG(dbgs() << " to: " << PN << "\n");
2772 deleteIfTriviallyDead(OldPtr);
2774 // PHIs can't be promoted on their own, but often can be speculated. We
2775 // check the speculation outside of the rewriter so that we see the
2776 // fully-rewritten alloca.
2777 PHIUsers.insert(&PN);
2781 bool visitSelectInst(SelectInst &SI) {
2782 DEBUG(dbgs() << " original: " << SI << "\n");
2783 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
2784 "Pointer isn't an operand!");
2785 assert(BeginOffset >= NewAllocaBeginOffset && "Selects are unsplittable");
2786 assert(EndOffset <= NewAllocaEndOffset && "Selects are unsplittable");
2788 Value *NewPtr = getNewAllocaSlicePtr(IRB, OldPtr->getType());
2789 // Replace the operands which were using the old pointer.
2790 if (SI.getOperand(1) == OldPtr)
2791 SI.setOperand(1, NewPtr);
2792 if (SI.getOperand(2) == OldPtr)
2793 SI.setOperand(2, NewPtr);
2795 DEBUG(dbgs() << " to: " << SI << "\n");
2796 deleteIfTriviallyDead(OldPtr);
2798 // Selects can't be promoted on their own, but often can be speculated. We
2799 // check the speculation outside of the rewriter so that we see the
2800 // fully-rewritten alloca.
2801 SelectUsers.insert(&SI);
2808 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2810 /// This pass aggressively rewrites all aggregate loads and stores on
2811 /// a particular pointer (or any pointer derived from it which we can identify)
2812 /// with scalar loads and stores.
2813 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2814 // Befriend the base class so it can delegate to private visit methods.
2815 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2817 const DataLayout &DL;
2819 /// Queue of pointer uses to analyze and potentially rewrite.
2820 SmallVector<Use *, 8> Queue;
2822 /// Set to prevent us from cycling with phi nodes and loops.
2823 SmallPtrSet<User *, 8> Visited;
2825 /// The current pointer use being rewritten. This is used to dig up the used
2826 /// value (as opposed to the user).
2830 AggLoadStoreRewriter(const DataLayout &DL) : DL(DL) {}
2832 /// Rewrite loads and stores through a pointer and all pointers derived from
2834 bool rewrite(Instruction &I) {
2835 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2837 bool Changed = false;
2838 while (!Queue.empty()) {
2839 U = Queue.pop_back_val();
2840 Changed |= visit(cast<Instruction>(U->getUser()));
2846 /// Enqueue all the users of the given instruction for further processing.
2847 /// This uses a set to de-duplicate users.
2848 void enqueueUsers(Instruction &I) {
2849 for (Use &U : I.uses())
2850 if (Visited.insert(U.getUser()).second)
2851 Queue.push_back(&U);
2854 // Conservative default is to not rewrite anything.
2855 bool visitInstruction(Instruction &I) { return false; }
2857 /// \brief Generic recursive split emission class.
2858 template <typename Derived> class OpSplitter {
2860 /// The builder used to form new instructions.
2862 /// The indices which to be used with insert- or extractvalue to select the
2863 /// appropriate value within the aggregate.
2864 SmallVector<unsigned, 4> Indices;
2865 /// The indices to a GEP instruction which will move Ptr to the correct slot
2866 /// within the aggregate.
2867 SmallVector<Value *, 4> GEPIndices;
2868 /// The base pointer of the original op, used as a base for GEPing the
2869 /// split operations.
2872 /// Initialize the splitter with an insertion point, Ptr and start with a
2873 /// single zero GEP index.
2874 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2875 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2878 /// \brief Generic recursive split emission routine.
2880 /// This method recursively splits an aggregate op (load or store) into
2881 /// scalar or vector ops. It splits recursively until it hits a single value
2882 /// and emits that single value operation via the template argument.
2884 /// The logic of this routine relies on GEPs and insertvalue and
2885 /// extractvalue all operating with the same fundamental index list, merely
2886 /// formatted differently (GEPs need actual values).
2888 /// \param Ty The type being split recursively into smaller ops.
2889 /// \param Agg The aggregate value being built up or stored, depending on
2890 /// whether this is splitting a load or a store respectively.
2891 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2892 if (Ty->isSingleValueType())
2893 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2895 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2896 unsigned OldSize = Indices.size();
2898 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2900 assert(Indices.size() == OldSize && "Did not return to the old size");
2901 Indices.push_back(Idx);
2902 GEPIndices.push_back(IRB.getInt32(Idx));
2903 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2904 GEPIndices.pop_back();
2910 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2911 unsigned OldSize = Indices.size();
2913 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2915 assert(Indices.size() == OldSize && "Did not return to the old size");
2916 Indices.push_back(Idx);
2917 GEPIndices.push_back(IRB.getInt32(Idx));
2918 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2919 GEPIndices.pop_back();
2925 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2929 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2930 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2931 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2933 /// Emit a leaf load of a single value. This is called at the leaves of the
2934 /// recursive emission to actually load values.
2935 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2936 assert(Ty->isSingleValueType());
2937 // Load the single value and insert it using the indices.
2938 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
2939 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
2940 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2941 DEBUG(dbgs() << " to: " << *Load << "\n");
2945 bool visitLoadInst(LoadInst &LI) {
2946 assert(LI.getPointerOperand() == *U);
2947 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2950 // We have an aggregate being loaded, split it apart.
2951 DEBUG(dbgs() << " original: " << LI << "\n");
2952 LoadOpSplitter Splitter(&LI, *U);
2953 Value *V = UndefValue::get(LI.getType());
2954 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2955 LI.replaceAllUsesWith(V);
2956 LI.eraseFromParent();
2960 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2961 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2962 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2964 /// Emit a leaf store of a single value. This is called at the leaves of the
2965 /// recursive emission to actually produce stores.
2966 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2967 assert(Ty->isSingleValueType());
2968 // Extract the single value and store it using the indices.
2969 Value *Store = IRB.CreateStore(
2970 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2971 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2973 DEBUG(dbgs() << " to: " << *Store << "\n");
2977 bool visitStoreInst(StoreInst &SI) {
2978 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2980 Value *V = SI.getValueOperand();
2981 if (V->getType()->isSingleValueType())
2984 // We have an aggregate being stored, split it apart.
2985 DEBUG(dbgs() << " original: " << SI << "\n");
2986 StoreOpSplitter Splitter(&SI, *U);
2987 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2988 SI.eraseFromParent();
2992 bool visitBitCastInst(BitCastInst &BC) {
2997 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3002 bool visitPHINode(PHINode &PN) {
3007 bool visitSelectInst(SelectInst &SI) {
3014 /// \brief Strip aggregate type wrapping.
3016 /// This removes no-op aggregate types wrapping an underlying type. It will
3017 /// strip as many layers of types as it can without changing either the type
3018 /// size or the allocated size.
3019 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3020 if (Ty->isSingleValueType())
3023 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3024 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3027 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3028 InnerTy = ArrTy->getElementType();
3029 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3030 const StructLayout *SL = DL.getStructLayout(STy);
3031 unsigned Index = SL->getElementContainingOffset(0);
3032 InnerTy = STy->getElementType(Index);
3037 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3038 TypeSize > DL.getTypeSizeInBits(InnerTy))
3041 return stripAggregateTypeWrapping(DL, InnerTy);
3044 /// \brief Try to find a partition of the aggregate type passed in for a given
3045 /// offset and size.
3047 /// This recurses through the aggregate type and tries to compute a subtype
3048 /// based on the offset and size. When the offset and size span a sub-section
3049 /// of an array, it will even compute a new array type for that sub-section,
3050 /// and the same for structs.
3052 /// Note that this routine is very strict and tries to find a partition of the
3053 /// type which produces the *exact* right offset and size. It is not forgiving
3054 /// when the size or offset cause either end of type-based partition to be off.
3055 /// Also, this is a best-effort routine. It is reasonable to give up and not
3056 /// return a type if necessary.
3057 static Type *getTypePartition(const DataLayout &DL, Type *Ty, uint64_t Offset,
3059 if (Offset == 0 && DL.getTypeAllocSize(Ty) == Size)
3060 return stripAggregateTypeWrapping(DL, Ty);
3061 if (Offset > DL.getTypeAllocSize(Ty) ||
3062 (DL.getTypeAllocSize(Ty) - Offset) < Size)
3065 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3066 // We can't partition pointers...
3067 if (SeqTy->isPointerTy())
3070 Type *ElementTy = SeqTy->getElementType();
3071 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3072 uint64_t NumSkippedElements = Offset / ElementSize;
3073 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
3074 if (NumSkippedElements >= ArrTy->getNumElements())
3076 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
3077 if (NumSkippedElements >= VecTy->getNumElements())
3080 Offset -= NumSkippedElements * ElementSize;
3082 // First check if we need to recurse.
3083 if (Offset > 0 || Size < ElementSize) {
3084 // Bail if the partition ends in a different array element.
3085 if ((Offset + Size) > ElementSize)
3087 // Recurse through the element type trying to peel off offset bytes.
3088 return getTypePartition(DL, ElementTy, Offset, Size);
3090 assert(Offset == 0);
3092 if (Size == ElementSize)
3093 return stripAggregateTypeWrapping(DL, ElementTy);
3094 assert(Size > ElementSize);
3095 uint64_t NumElements = Size / ElementSize;
3096 if (NumElements * ElementSize != Size)
3098 return ArrayType::get(ElementTy, NumElements);
3101 StructType *STy = dyn_cast<StructType>(Ty);
3105 const StructLayout *SL = DL.getStructLayout(STy);
3106 if (Offset >= SL->getSizeInBytes())
3108 uint64_t EndOffset = Offset + Size;
3109 if (EndOffset > SL->getSizeInBytes())
3112 unsigned Index = SL->getElementContainingOffset(Offset);
3113 Offset -= SL->getElementOffset(Index);
3115 Type *ElementTy = STy->getElementType(Index);
3116 uint64_t ElementSize = DL.getTypeAllocSize(ElementTy);
3117 if (Offset >= ElementSize)
3118 return nullptr; // The offset points into alignment padding.
3120 // See if any partition must be contained by the element.
3121 if (Offset > 0 || Size < ElementSize) {
3122 if ((Offset + Size) > ElementSize)
3124 return getTypePartition(DL, ElementTy, Offset, Size);
3126 assert(Offset == 0);
3128 if (Size == ElementSize)
3129 return stripAggregateTypeWrapping(DL, ElementTy);
3131 StructType::element_iterator EI = STy->element_begin() + Index,
3132 EE = STy->element_end();
3133 if (EndOffset < SL->getSizeInBytes()) {
3134 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3135 if (Index == EndIndex)
3136 return nullptr; // Within a single element and its padding.
3138 // Don't try to form "natural" types if the elements don't line up with the
3140 // FIXME: We could potentially recurse down through the last element in the
3141 // sub-struct to find a natural end point.
3142 if (SL->getElementOffset(EndIndex) != EndOffset)
3145 assert(Index < EndIndex);
3146 EE = STy->element_begin() + EndIndex;
3149 // Try to build up a sub-structure.
3151 StructType::get(STy->getContext(), makeArrayRef(EI, EE), STy->isPacked());
3152 const StructLayout *SubSL = DL.getStructLayout(SubTy);
3153 if (Size != SubSL->getSizeInBytes())
3154 return nullptr; // The sub-struct doesn't have quite the size needed.
3159 /// \brief Rewrite an alloca partition's users.
3161 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3162 /// to rewrite uses of an alloca partition to be conducive for SSA value
3163 /// promotion. If the partition needs a new, more refined alloca, this will
3164 /// build that new alloca, preserving as much type information as possible, and
3165 /// rewrite the uses of the old alloca to point at the new one and have the
3166 /// appropriate new offsets. It also evaluates how successful the rewrite was
3167 /// at enabling promotion and if it was successful queues the alloca to be
3169 bool SROA::rewritePartition(AllocaInst &AI, AllocaSlices &AS,
3170 AllocaSlices::iterator B, AllocaSlices::iterator E,
3171 int64_t BeginOffset, int64_t EndOffset,
3172 ArrayRef<AllocaSlices::iterator> SplitUses) {
3173 assert(BeginOffset < EndOffset);
3174 uint64_t SliceSize = EndOffset - BeginOffset;
3176 // Try to compute a friendly type for this partition of the alloca. This
3177 // won't always succeed, in which case we fall back to a legal integer type
3178 // or an i8 array of an appropriate size.
3179 Type *SliceTy = nullptr;
3180 if (Type *CommonUseTy = findCommonType(B, E, EndOffset))
3181 if (DL->getTypeAllocSize(CommonUseTy) >= SliceSize)
3182 SliceTy = CommonUseTy;
3184 if (Type *TypePartitionTy = getTypePartition(*DL, AI.getAllocatedType(),
3185 BeginOffset, SliceSize))
3186 SliceTy = TypePartitionTy;
3187 if ((!SliceTy || (SliceTy->isArrayTy() &&
3188 SliceTy->getArrayElementType()->isIntegerTy())) &&
3189 DL->isLegalInteger(SliceSize * 8))
3190 SliceTy = Type::getIntNTy(*C, SliceSize * 8);
3192 SliceTy = ArrayType::get(Type::getInt8Ty(*C), SliceSize);
3193 assert(DL->getTypeAllocSize(SliceTy) >= SliceSize);
3195 bool IsIntegerPromotable = isIntegerWideningViable(
3196 *DL, SliceTy, BeginOffset, AllocaSlices::const_range(B, E), SplitUses);
3201 : isVectorPromotionViable(*DL, BeginOffset, EndOffset,
3202 AllocaSlices::const_range(B, E), SplitUses);
3206 // Check for the case where we're going to rewrite to a new alloca of the
3207 // exact same type as the original, and with the same access offsets. In that
3208 // case, re-use the existing alloca, but still run through the rewriter to
3209 // perform phi and select speculation.
3211 if (SliceTy == AI.getAllocatedType()) {
3212 assert(BeginOffset == 0 && "Non-zero begin offset but same alloca type");
3214 // FIXME: We should be able to bail at this point with "nothing changed".
3215 // FIXME: We might want to defer PHI speculation until after here.
3217 unsigned Alignment = AI.getAlignment();
3219 // The minimum alignment which users can rely on when the explicit
3220 // alignment is omitted or zero is that required by the ABI for this
3222 Alignment = DL->getABITypeAlignment(AI.getAllocatedType());
3224 Alignment = MinAlign(Alignment, BeginOffset);
3225 // If we will get at least this much alignment from the type alone, leave
3226 // the alloca's alignment unconstrained.
3227 if (Alignment <= DL->getABITypeAlignment(SliceTy))
3230 new AllocaInst(SliceTy, nullptr, Alignment,
3231 AI.getName() + ".sroa." + Twine(B - AS.begin()), &AI);
3235 DEBUG(dbgs() << "Rewriting alloca partition "
3236 << "[" << BeginOffset << "," << EndOffset << ") to: " << *NewAI
3239 // Track the high watermark on the worklist as it is only relevant for
3240 // promoted allocas. We will reset it to this point if the alloca is not in
3241 // fact scheduled for promotion.
3242 unsigned PPWOldSize = PostPromotionWorklist.size();
3243 unsigned NumUses = 0;
3244 SmallPtrSet<PHINode *, 8> PHIUsers;
3245 SmallPtrSet<SelectInst *, 8> SelectUsers;
3247 AllocaSliceRewriter Rewriter(*DL, AS, *this, AI, *NewAI, BeginOffset,
3248 EndOffset, IsIntegerPromotable, VecTy, PHIUsers,
3250 bool Promotable = true;
3251 for (auto &SplitUse : SplitUses) {
3252 DEBUG(dbgs() << " rewriting split ");
3253 DEBUG(AS.printSlice(dbgs(), SplitUse, ""));
3254 Promotable &= Rewriter.visit(SplitUse);
3257 for (AllocaSlices::iterator I = B; I != E; ++I) {
3258 DEBUG(dbgs() << " rewriting ");
3259 DEBUG(AS.printSlice(dbgs(), I, ""));
3260 Promotable &= Rewriter.visit(I);
3264 NumAllocaPartitionUses += NumUses;
3265 MaxUsesPerAllocaPartition =
3266 std::max<unsigned>(NumUses, MaxUsesPerAllocaPartition);
3268 // Now that we've processed all the slices in the new partition, check if any
3269 // PHIs or Selects would block promotion.
3270 for (SmallPtrSetImpl<PHINode *>::iterator I = PHIUsers.begin(),
3273 if (!isSafePHIToSpeculate(**I, DL)) {
3276 SelectUsers.clear();
3279 for (SmallPtrSetImpl<SelectInst *>::iterator I = SelectUsers.begin(),
3280 E = SelectUsers.end();
3282 if (!isSafeSelectToSpeculate(**I, DL)) {
3285 SelectUsers.clear();
3290 if (PHIUsers.empty() && SelectUsers.empty()) {
3291 // Promote the alloca.
3292 PromotableAllocas.push_back(NewAI);
3294 // If we have either PHIs or Selects to speculate, add them to those
3295 // worklists and re-queue the new alloca so that we promote in on the
3297 for (PHINode *PHIUser : PHIUsers)
3298 SpeculatablePHIs.insert(PHIUser);
3299 for (SelectInst *SelectUser : SelectUsers)
3300 SpeculatableSelects.insert(SelectUser);
3301 Worklist.insert(NewAI);
3304 // If we can't promote the alloca, iterate on it to check for new
3305 // refinements exposed by splitting the current alloca. Don't iterate on an
3306 // alloca which didn't actually change and didn't get promoted.
3308 Worklist.insert(NewAI);
3310 // Drop any post-promotion work items if promotion didn't happen.
3311 while (PostPromotionWorklist.size() > PPWOldSize)
3312 PostPromotionWorklist.pop_back();
3319 removeFinishedSplitUses(SmallVectorImpl<AllocaSlices::iterator> &SplitUses,
3320 uint64_t &MaxSplitUseEndOffset, uint64_t Offset) {
3321 if (Offset >= MaxSplitUseEndOffset) {
3323 MaxSplitUseEndOffset = 0;
3327 size_t SplitUsesOldSize = SplitUses.size();
3328 SplitUses.erase(std::remove_if(
3329 SplitUses.begin(), SplitUses.end(),
3330 [Offset](const AllocaSlices::iterator &I) {
3331 return I->endOffset() <= Offset;
3334 if (SplitUsesOldSize == SplitUses.size())
3337 // Recompute the max. While this is linear, so is remove_if.
3338 MaxSplitUseEndOffset = 0;
3339 for (AllocaSlices::iterator SplitUse : SplitUses)
3340 MaxSplitUseEndOffset =
3341 std::max(SplitUse->endOffset(), MaxSplitUseEndOffset);
3344 /// \brief Walks the slices of an alloca and form partitions based on them,
3345 /// rewriting each of their uses.
3346 bool SROA::splitAlloca(AllocaInst &AI, AllocaSlices &AS) {
3347 if (AS.begin() == AS.end())
3350 unsigned NumPartitions = 0;
3351 bool Changed = false;
3352 SmallVector<AllocaSlices::iterator, 4> SplitUses;
3353 uint64_t MaxSplitUseEndOffset = 0;
3355 uint64_t BeginOffset = AS.begin()->beginOffset();
3357 for (AllocaSlices::iterator SI = AS.begin(), SJ = std::next(SI),
3359 SI != SE; SI = SJ) {
3360 uint64_t MaxEndOffset = SI->endOffset();
3362 if (!SI->isSplittable()) {
3363 // When we're forming an unsplittable region, it must always start at the
3364 // first slice and will extend through its end.
3365 assert(BeginOffset == SI->beginOffset());
3367 // Form a partition including all of the overlapping slices with this
3368 // unsplittable slice.
3369 while (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3370 if (!SJ->isSplittable())
3371 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3375 assert(SI->isSplittable()); // Established above.
3377 // Collect all of the overlapping splittable slices.
3378 while (SJ != SE && SJ->beginOffset() < MaxEndOffset &&
3379 SJ->isSplittable()) {
3380 MaxEndOffset = std::max(MaxEndOffset, SJ->endOffset());
3384 // Back up MaxEndOffset and SJ if we ended the span early when
3385 // encountering an unsplittable slice.
3386 if (SJ != SE && SJ->beginOffset() < MaxEndOffset) {
3387 assert(!SJ->isSplittable());
3388 MaxEndOffset = SJ->beginOffset();
3392 // Check if we have managed to move the end offset forward yet. If so,
3393 // we'll have to rewrite uses and erase old split uses.
3394 if (BeginOffset < MaxEndOffset) {
3395 // Rewrite a sequence of overlapping slices.
3396 Changed |= rewritePartition(AI, AS, SI, SJ, BeginOffset, MaxEndOffset,
3400 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset, MaxEndOffset);
3403 // Accumulate all the splittable slices from the [SI,SJ) region which
3404 // overlap going forward.
3405 for (AllocaSlices::iterator SK = SI; SK != SJ; ++SK)
3406 if (SK->isSplittable() && SK->endOffset() > MaxEndOffset) {
3407 SplitUses.push_back(SK);
3408 MaxSplitUseEndOffset = std::max(SK->endOffset(), MaxSplitUseEndOffset);
3411 // If we're already at the end and we have no split uses, we're done.
3412 if (SJ == SE && SplitUses.empty())
3415 // If we have no split uses or no gap in offsets, we're ready to move to
3417 if (SplitUses.empty() || (SJ != SE && MaxEndOffset == SJ->beginOffset())) {
3418 BeginOffset = SJ->beginOffset();
3422 // Even if we have split slices, if the next slice is splittable and the
3423 // split slices reach it, we can simply set up the beginning offset of the
3424 // next iteration to bridge between them.
3425 if (SJ != SE && SJ->isSplittable() &&
3426 MaxSplitUseEndOffset > SJ->beginOffset()) {
3427 BeginOffset = MaxEndOffset;
3431 // Otherwise, we have a tail of split slices. Rewrite them with an empty
3433 uint64_t PostSplitEndOffset =
3434 SJ == SE ? MaxSplitUseEndOffset : SJ->beginOffset();
3436 Changed |= rewritePartition(AI, AS, SJ, SJ, MaxEndOffset,
3437 PostSplitEndOffset, SplitUses);
3441 break; // Skip the rest, we don't need to do any cleanup.
3443 removeFinishedSplitUses(SplitUses, MaxSplitUseEndOffset,
3444 PostSplitEndOffset);
3446 // Now just reset the begin offset for the next iteration.
3447 BeginOffset = SJ->beginOffset();
3450 NumAllocaPartitions += NumPartitions;
3451 MaxPartitionsPerAlloca =
3452 std::max<unsigned>(NumPartitions, MaxPartitionsPerAlloca);
3457 /// \brief Clobber a use with undef, deleting the used value if it becomes dead.
3458 void SROA::clobberUse(Use &U) {
3460 // Replace the use with an undef value.
3461 U = UndefValue::get(OldV->getType());
3463 // Check for this making an instruction dead. We have to garbage collect
3464 // all the dead instructions to ensure the uses of any alloca end up being
3466 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3467 if (isInstructionTriviallyDead(OldI)) {
3468 DeadInsts.insert(OldI);
3472 /// \brief Analyze an alloca for SROA.
3474 /// This analyzes the alloca to ensure we can reason about it, builds
3475 /// the slices of the alloca, and then hands it off to be split and
3476 /// rewritten as needed.
3477 bool SROA::runOnAlloca(AllocaInst &AI) {
3478 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3479 ++NumAllocasAnalyzed;
3481 // Special case dead allocas, as they're trivial.
3482 if (AI.use_empty()) {
3483 AI.eraseFromParent();
3487 // Skip alloca forms that this analysis can't handle.
3488 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3489 DL->getTypeAllocSize(AI.getAllocatedType()) == 0)
3492 bool Changed = false;
3494 // First, split any FCA loads and stores touching this alloca to promote
3495 // better splitting and promotion opportunities.
3496 AggLoadStoreRewriter AggRewriter(*DL);
3497 Changed |= AggRewriter.rewrite(AI);
3499 // Build the slices using a recursive instruction-visiting builder.
3500 AllocaSlices AS(*DL, AI);
3501 DEBUG(AS.print(dbgs()));
3505 // Delete all the dead users of this alloca before splitting and rewriting it.
3506 for (Instruction *DeadUser : AS.getDeadUsers()) {
3507 // Free up everything used by this instruction.
3508 for (Use &DeadOp : DeadUser->operands())
3511 // Now replace the uses of this instruction.
3512 DeadUser->replaceAllUsesWith(UndefValue::get(DeadUser->getType()));
3514 // And mark it for deletion.
3515 DeadInsts.insert(DeadUser);
3518 for (Use *DeadOp : AS.getDeadOperands()) {
3519 clobberUse(*DeadOp);
3523 // No slices to split. Leave the dead alloca for a later pass to clean up.
3524 if (AS.begin() == AS.end())
3527 Changed |= splitAlloca(AI, AS);
3529 DEBUG(dbgs() << " Speculating PHIs\n");
3530 while (!SpeculatablePHIs.empty())
3531 speculatePHINodeLoads(*SpeculatablePHIs.pop_back_val());
3533 DEBUG(dbgs() << " Speculating Selects\n");
3534 while (!SpeculatableSelects.empty())
3535 speculateSelectInstLoads(*SpeculatableSelects.pop_back_val());
3540 /// \brief Delete the dead instructions accumulated in this run.
3542 /// Recursively deletes the dead instructions we've accumulated. This is done
3543 /// at the very end to maximize locality of the recursive delete and to
3544 /// minimize the problems of invalidated instruction pointers as such pointers
3545 /// are used heavily in the intermediate stages of the algorithm.
3547 /// We also record the alloca instructions deleted here so that they aren't
3548 /// subsequently handed to mem2reg to promote.
3549 void SROA::deleteDeadInstructions(
3550 SmallPtrSetImpl<AllocaInst *> &DeletedAllocas) {
3551 while (!DeadInsts.empty()) {
3552 Instruction *I = DeadInsts.pop_back_val();
3553 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3555 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3557 for (Use &Operand : I->operands())
3558 if (Instruction *U = dyn_cast<Instruction>(Operand)) {
3559 // Zero out the operand and see if it becomes trivially dead.
3561 if (isInstructionTriviallyDead(U))
3562 DeadInsts.insert(U);
3565 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3566 DeletedAllocas.insert(AI);
3569 I->eraseFromParent();
3573 static void enqueueUsersInWorklist(Instruction &I,
3574 SmallVectorImpl<Instruction *> &Worklist,
3575 SmallPtrSetImpl<Instruction *> &Visited) {
3576 for (User *U : I.users())
3577 if (Visited.insert(cast<Instruction>(U)).second)
3578 Worklist.push_back(cast<Instruction>(U));
3581 /// \brief Promote the allocas, using the best available technique.
3583 /// This attempts to promote whatever allocas have been identified as viable in
3584 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3585 /// If there is a domtree available, we attempt to promote using the full power
3586 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3587 /// based on the SSAUpdater utilities. This function returns whether any
3588 /// promotion occurred.
3589 bool SROA::promoteAllocas(Function &F) {
3590 if (PromotableAllocas.empty())
3593 NumPromoted += PromotableAllocas.size();
3595 if (DT && !ForceSSAUpdater) {
3596 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3597 PromoteMemToReg(PromotableAllocas, *DT, nullptr, AT);
3598 PromotableAllocas.clear();
3602 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3604 DIBuilder DIB(*F.getParent(), /*AllowUnresolved*/ false);
3605 SmallVector<Instruction *, 64> Insts;
3607 // We need a worklist to walk the uses of each alloca.
3608 SmallVector<Instruction *, 8> Worklist;
3609 SmallPtrSet<Instruction *, 8> Visited;
3610 SmallVector<Instruction *, 32> DeadInsts;
3612 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3613 AllocaInst *AI = PromotableAllocas[Idx];
3618 enqueueUsersInWorklist(*AI, Worklist, Visited);
3620 while (!Worklist.empty()) {
3621 Instruction *I = Worklist.pop_back_val();
3623 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3624 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3625 // leading to them) here. Eventually it should use them to optimize the
3626 // scalar values produced.
3627 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3628 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3629 II->getIntrinsicID() == Intrinsic::lifetime_end);
3630 II->eraseFromParent();
3634 // Push the loads and stores we find onto the list. SROA will already
3635 // have validated that all loads and stores are viable candidates for
3637 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
3638 assert(LI->getType() == AI->getAllocatedType());
3639 Insts.push_back(LI);
3642 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
3643 assert(SI->getValueOperand()->getType() == AI->getAllocatedType());
3644 Insts.push_back(SI);
3648 // For everything else, we know that only no-op bitcasts and GEPs will
3649 // make it this far, just recurse through them and recall them for later
3651 DeadInsts.push_back(I);
3652 enqueueUsersInWorklist(*I, Worklist, Visited);
3654 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3655 while (!DeadInsts.empty())
3656 DeadInsts.pop_back_val()->eraseFromParent();
3657 AI->eraseFromParent();
3660 PromotableAllocas.clear();
3664 bool SROA::runOnFunction(Function &F) {
3665 if (skipOptnoneFunction(F))
3668 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3669 C = &F.getContext();
3670 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
3672 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3675 DL = &DLP->getDataLayout();
3676 DominatorTreeWrapperPass *DTWP =
3677 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
3678 DT = DTWP ? &DTWP->getDomTree() : nullptr;
3679 AT = &getAnalysis<AssumptionTracker>();
3681 BasicBlock &EntryBB = F.getEntryBlock();
3682 for (BasicBlock::iterator I = EntryBB.begin(), E = std::prev(EntryBB.end());
3684 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3685 Worklist.insert(AI);
3687 bool Changed = false;
3688 // A set of deleted alloca instruction pointers which should be removed from
3689 // the list of promotable allocas.
3690 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3693 while (!Worklist.empty()) {
3694 Changed |= runOnAlloca(*Worklist.pop_back_val());
3695 deleteDeadInstructions(DeletedAllocas);
3697 // Remove the deleted allocas from various lists so that we don't try to
3698 // continue processing them.
3699 if (!DeletedAllocas.empty()) {
3700 auto IsInSet = [&](AllocaInst *AI) { return DeletedAllocas.count(AI); };
3701 Worklist.remove_if(IsInSet);
3702 PostPromotionWorklist.remove_if(IsInSet);
3703 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3704 PromotableAllocas.end(),
3706 PromotableAllocas.end());
3707 DeletedAllocas.clear();
3711 Changed |= promoteAllocas(F);
3713 Worklist = PostPromotionWorklist;
3714 PostPromotionWorklist.clear();
3715 } while (!Worklist.empty());
3720 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3721 AU.addRequired<AssumptionTracker>();
3722 if (RequiresDomTree)
3723 AU.addRequired<DominatorTreeWrapperPass>();
3724 AU.setPreservesCFG();