1 //===- SROA.cpp - Scalar Replacement Of Aggregates ------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This transformation implements the well known scalar replacement of
11 /// aggregates transformation. It tries to identify promotable elements of an
12 /// aggregate alloca, and promote them to registers. It will also try to
13 /// convert uses of an element (or set of elements) of an alloca into a vector
14 /// or bitfield-style integer scalar if appropriate.
16 /// It works to do this with minimal slicing of the alloca so that regions
17 /// which are merely transferred in and out of external memory remain unchanged
18 /// and are not decomposed to scalar code.
20 /// Because this also performs alloca promotion, it can be thought of as also
21 /// serving the purpose of SSA formation. The algorithm iterates on the
22 /// function until all opportunities for promotion have been realized.
24 //===----------------------------------------------------------------------===//
26 #define DEBUG_TYPE "sroa"
27 #include "llvm/Transforms/Scalar.h"
28 #include "llvm/ADT/STLExtras.h"
29 #include "llvm/ADT/SetVector.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Analysis/Dominators.h"
33 #include "llvm/Analysis/Loads.h"
34 #include "llvm/Analysis/PtrUseVisitor.h"
35 #include "llvm/Analysis/ValueTracking.h"
36 #include "llvm/DIBuilder.h"
37 #include "llvm/DebugInfo.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/IntrinsicInst.h"
45 #include "llvm/IR/LLVMContext.h"
46 #include "llvm/IR/Operator.h"
47 #include "llvm/InstVisitor.h"
48 #include "llvm/Pass.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include "llvm/Transforms/Utils/Local.h"
55 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
56 #include "llvm/Transforms/Utils/SSAUpdater.h"
59 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
60 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
61 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
62 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
63 STATISTIC(NumDeleted, "Number of instructions deleted");
64 STATISTIC(NumVectorized, "Number of vectorized aggregates");
66 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
67 /// forming SSA values through the SSAUpdater infrastructure.
69 ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
72 /// \brief Alloca partitioning representation.
74 /// This class represents a partitioning of an alloca into slices, and
75 /// information about the nature of uses of each slice of the alloca. The goal
76 /// is that this information is sufficient to decide if and how to split the
77 /// alloca apart and replace slices with scalars. It is also intended that this
78 /// structure can capture the relevant information needed both to decide about
79 /// and to enact these transformations.
80 class AllocaPartitioning {
82 /// \brief A common base class for representing a half-open byte range.
84 /// \brief The beginning offset of the range.
87 /// \brief The ending offset, not included in the range.
90 ByteRange() : BeginOffset(), EndOffset() {}
91 ByteRange(uint64_t BeginOffset, uint64_t EndOffset)
92 : BeginOffset(BeginOffset), EndOffset(EndOffset) {}
94 /// \brief Support for ordering ranges.
96 /// This provides an ordering over ranges such that start offsets are
97 /// always increasing, and within equal start offsets, the end offsets are
98 /// decreasing. Thus the spanning range comes first in a cluster with the
99 /// same start position.
100 bool operator<(const ByteRange &RHS) const {
101 if (BeginOffset < RHS.BeginOffset) return true;
102 if (BeginOffset > RHS.BeginOffset) return false;
103 if (EndOffset > RHS.EndOffset) return true;
107 /// \brief Support comparison with a single offset to allow binary searches.
108 friend bool operator<(const ByteRange &LHS, uint64_t RHSOffset) {
109 return LHS.BeginOffset < RHSOffset;
112 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
113 const ByteRange &RHS) {
114 return LHSOffset < RHS.BeginOffset;
117 bool operator==(const ByteRange &RHS) const {
118 return BeginOffset == RHS.BeginOffset && EndOffset == RHS.EndOffset;
120 bool operator!=(const ByteRange &RHS) const { return !operator==(RHS); }
123 /// \brief A partition of an alloca.
125 /// This structure represents a contiguous partition of the alloca. These are
126 /// formed by examining the uses of the alloca. During formation, they may
127 /// overlap but once an AllocaPartitioning is built, the Partitions within it
128 /// are all disjoint.
129 struct Partition : public ByteRange {
130 /// \brief Whether this partition is splittable into smaller partitions.
132 /// We flag partitions as splittable when they are formed entirely due to
133 /// accesses by trivially splittable operations such as memset and memcpy.
136 /// \brief Test whether a partition has been marked as dead.
137 bool isDead() const {
138 if (BeginOffset == UINT64_MAX) {
139 assert(EndOffset == UINT64_MAX);
145 /// \brief Kill a partition.
146 /// This is accomplished by setting both its beginning and end offset to
147 /// the maximum possible value.
149 assert(!isDead() && "He's Dead, Jim!");
150 BeginOffset = EndOffset = UINT64_MAX;
153 Partition() : ByteRange(), IsSplittable() {}
154 Partition(uint64_t BeginOffset, uint64_t EndOffset, bool IsSplittable)
155 : ByteRange(BeginOffset, EndOffset), IsSplittable(IsSplittable) {}
158 /// \brief A particular use of a partition of the alloca.
160 /// This structure is used to associate uses of a partition with it. They
161 /// mark the range of bytes which are referenced by a particular instruction,
162 /// and includes a handle to the user itself and the pointer value in use.
163 /// The bounds of these uses are determined by intersecting the bounds of the
164 /// memory use itself with a particular partition. As a consequence there is
165 /// intentionally overlap between various uses of the same partition.
166 class PartitionUse : public ByteRange {
167 /// \brief Combined storage for both the Use* and split state.
168 PointerIntPair<Use*, 1, bool> UsePtrAndIsSplit;
171 PartitionUse() : ByteRange(), UsePtrAndIsSplit() {}
172 PartitionUse(uint64_t BeginOffset, uint64_t EndOffset, Use *U,
174 : ByteRange(BeginOffset, EndOffset), UsePtrAndIsSplit(U, IsSplit) {}
176 /// \brief The use in question. Provides access to both user and used value.
178 /// Note that this may be null if the partition use is *dead*, that is, it
179 /// should be ignored.
180 Use *getUse() const { return UsePtrAndIsSplit.getPointer(); }
182 /// \brief Set the use for this partition use range.
183 void setUse(Use *U) { UsePtrAndIsSplit.setPointer(U); }
185 /// \brief Whether this use is split across multiple partitions.
186 bool isSplit() const { return UsePtrAndIsSplit.getInt(); }
189 /// \brief Construct a partitioning of a particular alloca.
191 /// Construction does most of the work for partitioning the alloca. This
192 /// performs the necessary walks of users and builds a partitioning from it.
193 AllocaPartitioning(const DataLayout &TD, AllocaInst &AI);
195 /// \brief Test whether a pointer to the allocation escapes our analysis.
197 /// If this is true, the partitioning is never fully built and should be
199 bool isEscaped() const { return PointerEscapingInstr; }
201 /// \brief Support for iterating over the partitions.
203 typedef SmallVectorImpl<Partition>::iterator iterator;
204 iterator begin() { return Partitions.begin(); }
205 iterator end() { return Partitions.end(); }
207 typedef SmallVectorImpl<Partition>::const_iterator const_iterator;
208 const_iterator begin() const { return Partitions.begin(); }
209 const_iterator end() const { return Partitions.end(); }
212 /// \brief Support for iterating over and manipulating a particular
213 /// partition's uses.
215 /// The iteration support provided for uses is more limited, but also
216 /// includes some manipulation routines to support rewriting the uses of
217 /// partitions during SROA.
219 typedef SmallVectorImpl<PartitionUse>::iterator use_iterator;
220 use_iterator use_begin(unsigned Idx) { return Uses[Idx].begin(); }
221 use_iterator use_begin(const_iterator I) { return Uses[I - begin()].begin(); }
222 use_iterator use_end(unsigned Idx) { return Uses[Idx].end(); }
223 use_iterator use_end(const_iterator I) { return Uses[I - begin()].end(); }
225 typedef SmallVectorImpl<PartitionUse>::const_iterator const_use_iterator;
226 const_use_iterator use_begin(unsigned Idx) const { return Uses[Idx].begin(); }
227 const_use_iterator use_begin(const_iterator I) const {
228 return Uses[I - begin()].begin();
230 const_use_iterator use_end(unsigned Idx) const { return Uses[Idx].end(); }
231 const_use_iterator use_end(const_iterator I) const {
232 return Uses[I - begin()].end();
235 unsigned use_size(unsigned Idx) const { return Uses[Idx].size(); }
236 unsigned use_size(const_iterator I) const { return Uses[I - begin()].size(); }
237 const PartitionUse &getUse(unsigned PIdx, unsigned UIdx) const {
238 return Uses[PIdx][UIdx];
240 const PartitionUse &getUse(const_iterator I, unsigned UIdx) const {
241 return Uses[I - begin()][UIdx];
244 void use_push_back(unsigned Idx, const PartitionUse &PU) {
245 Uses[Idx].push_back(PU);
247 void use_push_back(const_iterator I, const PartitionUse &PU) {
248 Uses[I - begin()].push_back(PU);
252 /// \brief Allow iterating the dead users for this alloca.
254 /// These are instructions which will never actually use the alloca as they
255 /// are outside the allocated range. They are safe to replace with undef and
258 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
259 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
260 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
263 /// \brief Allow iterating the dead expressions referring to this alloca.
265 /// These are operands which have cannot actually be used to refer to the
266 /// alloca as they are outside its range and the user doesn't correct for
267 /// that. These mostly consist of PHI node inputs and the like which we just
268 /// need to replace with undef.
270 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
271 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
272 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
275 /// \brief MemTransferInst auxiliary data.
276 /// This struct provides some auxiliary data about memory transfer
277 /// intrinsics such as memcpy and memmove. These intrinsics can use two
278 /// different ranges within the same alloca, and provide other challenges to
279 /// correctly represent. We stash extra data to help us untangle this
280 /// after the partitioning is complete.
281 struct MemTransferOffsets {
282 /// The destination begin and end offsets when the destination is within
283 /// this alloca. If the end offset is zero the destination is not within
285 uint64_t DestBegin, DestEnd;
287 /// The source begin and end offsets when the source is within this alloca.
288 /// If the end offset is zero, the source is not within this alloca.
289 uint64_t SourceBegin, SourceEnd;
291 /// Flag for whether an alloca is splittable.
294 MemTransferOffsets getMemTransferOffsets(MemTransferInst &II) const {
295 return MemTransferInstData.lookup(&II);
298 /// \brief Map from a PHI or select operand back to a partition.
300 /// When manipulating PHI nodes or selects, they can use more than one
301 /// partition of an alloca. We store a special mapping to allow finding the
302 /// partition referenced by each of these operands, if any.
303 iterator findPartitionForPHIOrSelectOperand(Use *U) {
304 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
305 = PHIOrSelectOpMap.find(U);
306 if (MapIt == PHIOrSelectOpMap.end())
309 return begin() + MapIt->second.first;
312 /// \brief Map from a PHI or select operand back to the specific use of
315 /// Similar to mapping these operands back to the partitions, this maps
316 /// directly to the use structure of that partition.
317 use_iterator findPartitionUseForPHIOrSelectOperand(Use *U) {
318 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
319 = PHIOrSelectOpMap.find(U);
320 assert(MapIt != PHIOrSelectOpMap.end());
321 return Uses[MapIt->second.first].begin() + MapIt->second.second;
324 /// \brief Compute a common type among the uses of a particular partition.
326 /// This routines walks all of the uses of a particular partition and tries
327 /// to find a common type between them. Untyped operations such as memset and
328 /// memcpy are ignored.
329 Type *getCommonType(iterator I) const;
331 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
332 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
333 void printUsers(raw_ostream &OS, const_iterator I,
334 StringRef Indent = " ") const;
335 void print(raw_ostream &OS) const;
336 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const;
337 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const;
341 template <typename DerivedT, typename RetT = void> class BuilderBase;
342 class PartitionBuilder;
343 friend class AllocaPartitioning::PartitionBuilder;
345 friend class AllocaPartitioning::UseBuilder;
347 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
348 /// \brief Handle to alloca instruction to simplify method interfaces.
352 /// \brief The instruction responsible for this alloca having no partitioning.
354 /// When an instruction (potentially) escapes the pointer to the alloca, we
355 /// store a pointer to that here and abort trying to partition the alloca.
356 /// This will be null if the alloca is partitioned successfully.
357 Instruction *PointerEscapingInstr;
359 /// \brief The partitions of the alloca.
361 /// We store a vector of the partitions over the alloca here. This vector is
362 /// sorted by increasing begin offset, and then by decreasing end offset. See
363 /// the Partition inner class for more details. Initially (during
364 /// construction) there are overlaps, but we form a disjoint sequence of
365 /// partitions while finishing construction and a fully constructed object is
366 /// expected to always have this as a disjoint space.
367 SmallVector<Partition, 8> Partitions;
369 /// \brief The uses of the partitions.
371 /// This is essentially a mapping from each partition to a list of uses of
372 /// that partition. The mapping is done with a Uses vector that has the exact
373 /// same number of entries as the partition vector. Each entry is itself
374 /// a vector of the uses.
375 SmallVector<SmallVector<PartitionUse, 2>, 8> Uses;
377 /// \brief Instructions which will become dead if we rewrite the alloca.
379 /// Note that these are not separated by partition. This is because we expect
380 /// a partitioned alloca to be completely rewritten or not rewritten at all.
381 /// If rewritten, all these instructions can simply be removed and replaced
382 /// with undef as they come from outside of the allocated space.
383 SmallVector<Instruction *, 8> DeadUsers;
385 /// \brief Operands which will become dead if we rewrite the alloca.
387 /// These are operands that in their particular use can be replaced with
388 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
389 /// to PHI nodes and the like. They aren't entirely dead (there might be
390 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
391 /// want to swap this particular input for undef to simplify the use lists of
393 SmallVector<Use *, 8> DeadOperands;
395 /// \brief The underlying storage for auxiliary memcpy and memset info.
396 SmallDenseMap<MemTransferInst *, MemTransferOffsets, 4> MemTransferInstData;
398 /// \brief A side datastructure used when building up the partitions and uses.
400 /// This mapping is only really used during the initial building of the
401 /// partitioning so that we can retain information about PHI and select nodes
403 SmallDenseMap<Instruction *, std::pair<uint64_t, bool> > PHIOrSelectSizes;
405 /// \brief Auxiliary information for particular PHI or select operands.
406 SmallDenseMap<Use *, std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap;
408 /// \brief A utility routine called from the constructor.
410 /// This does what it says on the tin. It is the key of the alloca partition
411 /// splitting and merging. After it is called we have the desired disjoint
412 /// collection of partitions.
413 void splitAndMergePartitions();
417 static Value *foldSelectInst(SelectInst &SI) {
418 // If the condition being selected on is a constant or the same value is
419 // being selected between, fold the select. Yes this does (rarely) happen
421 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
422 return SI.getOperand(1+CI->isZero());
423 if (SI.getOperand(1) == SI.getOperand(2))
424 return SI.getOperand(1);
429 /// \brief Builder for the alloca partitioning.
431 /// This class builds an alloca partitioning by recursively visiting the uses
432 /// of an alloca and splitting the partitions for each load and store at each
434 class AllocaPartitioning::PartitionBuilder
435 : public PtrUseVisitor<PartitionBuilder> {
436 friend class PtrUseVisitor<PartitionBuilder>;
437 friend class InstVisitor<PartitionBuilder>;
438 typedef PtrUseVisitor<PartitionBuilder> Base;
440 const uint64_t AllocSize;
441 AllocaPartitioning &P;
443 SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap;
446 PartitionBuilder(const DataLayout &DL, AllocaInst &AI, AllocaPartitioning &P)
447 : PtrUseVisitor<PartitionBuilder>(DL),
448 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())),
452 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
453 bool IsSplittable = false) {
454 // Completely skip uses which have a zero size or start either before or
455 // past the end of the allocation.
456 if (Size == 0 || Offset.isNegative() || Offset.uge(AllocSize)) {
457 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
458 << " which has zero size or starts outside of the "
459 << AllocSize << " byte alloca:\n"
460 << " alloca: " << P.AI << "\n"
461 << " use: " << I << "\n");
465 uint64_t BeginOffset = Offset.getZExtValue();
466 uint64_t EndOffset = BeginOffset + Size;
468 // Clamp the end offset to the end of the allocation. Note that this is
469 // formulated to handle even the case where "BeginOffset + Size" overflows.
470 // This may appear superficially to be something we could ignore entirely,
471 // but that is not so! There may be widened loads or PHI-node uses where
472 // some instructions are dead but not others. We can't completely ignore
473 // them, and so have to record at least the information here.
474 assert(AllocSize >= BeginOffset); // Established above.
475 if (Size > AllocSize - BeginOffset) {
476 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
477 << " to remain within the " << AllocSize << " byte alloca:\n"
478 << " alloca: " << P.AI << "\n"
479 << " use: " << I << "\n");
480 EndOffset = AllocSize;
483 Partition New(BeginOffset, EndOffset, IsSplittable);
484 P.Partitions.push_back(New);
487 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
488 uint64_t Size, bool IsVolatile) {
489 // We allow splitting of loads and stores where the type is an integer type
490 // and cover the entire alloca. This prevents us from splitting over
492 // FIXME: In the great blue eventually, we should eagerly split all integer
493 // loads and stores, and then have a separate step that merges adjacent
494 // alloca partitions into a single partition suitable for integer widening.
495 // Or we should skip the merge step and rely on GVN and other passes to
496 // merge adjacent loads and stores that survive mem2reg.
498 Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize;
500 insertUse(I, Offset, Size, IsSplittable);
503 void visitLoadInst(LoadInst &LI) {
504 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
505 "All simple FCA loads should have been pre-split");
508 return PI.setAborted(&LI);
510 uint64_t Size = DL.getTypeStoreSize(LI.getType());
511 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
514 void visitStoreInst(StoreInst &SI) {
515 Value *ValOp = SI.getValueOperand();
517 return PI.setEscapedAndAborted(&SI);
519 return PI.setAborted(&SI);
521 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
523 // If this memory access can be shown to *statically* extend outside the
524 // bounds of of the allocation, it's behavior is undefined, so simply
525 // ignore it. Note that this is more strict than the generic clamping
526 // behavior of insertUse. We also try to handle cases which might run the
528 // FIXME: We should instead consider the pointer to have escaped if this
529 // function is being instrumented for addressing bugs or race conditions.
530 if (Offset.isNegative() || Size > AllocSize ||
531 Offset.ugt(AllocSize - Size)) {
532 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
533 << " which extends past the end of the " << AllocSize
535 << " alloca: " << P.AI << "\n"
536 << " use: " << SI << "\n");
540 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
541 "All simple FCA stores should have been pre-split");
542 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
546 void visitMemSetInst(MemSetInst &II) {
547 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
548 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
549 if ((Length && Length->getValue() == 0) ||
550 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
551 // Zero-length mem transfer intrinsics can be ignored entirely.
555 return PI.setAborted(&II);
557 insertUse(II, Offset,
558 Length ? Length->getLimitedValue()
559 : AllocSize - Offset.getLimitedValue(),
563 void visitMemTransferInst(MemTransferInst &II) {
564 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
565 if ((Length && Length->getValue() == 0) ||
566 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
567 // Zero-length mem transfer intrinsics can be ignored entirely.
571 return PI.setAborted(&II);
573 uint64_t RawOffset = Offset.getLimitedValue();
574 uint64_t Size = Length ? Length->getLimitedValue()
575 : AllocSize - RawOffset;
577 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
579 // Only intrinsics with a constant length can be split.
580 Offsets.IsSplittable = Length;
582 if (*U == II.getRawDest()) {
583 Offsets.DestBegin = RawOffset;
584 Offsets.DestEnd = RawOffset + Size;
586 if (*U == II.getRawSource()) {
587 Offsets.SourceBegin = RawOffset;
588 Offsets.SourceEnd = RawOffset + Size;
591 // If we have set up end offsets for both the source and the destination,
592 // we have found both sides of this transfer pointing at the same alloca.
593 bool SeenBothEnds = Offsets.SourceEnd && Offsets.DestEnd;
594 if (SeenBothEnds && II.getRawDest() != II.getRawSource()) {
595 unsigned PrevIdx = MemTransferPartitionMap[&II];
597 // Check if the begin offsets match and this is a non-volatile transfer.
598 // In that case, we can completely elide the transfer.
599 if (!II.isVolatile() && Offsets.SourceBegin == Offsets.DestBegin) {
600 P.Partitions[PrevIdx].kill();
604 // Otherwise we have an offset transfer within the same alloca. We can't
606 P.Partitions[PrevIdx].IsSplittable = Offsets.IsSplittable = false;
607 } else if (SeenBothEnds) {
608 // Handle the case where this exact use provides both ends of the
610 assert(II.getRawDest() == II.getRawSource());
612 // For non-volatile transfers this is a no-op.
613 if (!II.isVolatile())
616 // Otherwise just suppress splitting.
617 Offsets.IsSplittable = false;
621 // Insert the use now that we've fixed up the splittable nature.
622 insertUse(II, Offset, Size, Offsets.IsSplittable);
624 // Setup the mapping from intrinsic to partition of we've not seen both
625 // ends of this transfer.
627 unsigned NewIdx = P.Partitions.size() - 1;
629 = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx)).second;
631 "Already have intrinsic in map but haven't seen both ends");
636 // Disable SRoA for any intrinsics except for lifetime invariants.
637 // FIXME: What about debug intrinsics? This matches old behavior, but
638 // doesn't make sense.
639 void visitIntrinsicInst(IntrinsicInst &II) {
641 return PI.setAborted(&II);
643 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
644 II.getIntrinsicID() == Intrinsic::lifetime_end) {
645 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
646 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
647 Length->getLimitedValue());
648 insertUse(II, Offset, Size, true);
652 Base::visitIntrinsicInst(II);
655 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
656 // We consider any PHI or select that results in a direct load or store of
657 // the same offset to be a viable use for partitioning purposes. These uses
658 // are considered unsplittable and the size is the maximum loaded or stored
660 SmallPtrSet<Instruction *, 4> Visited;
661 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
662 Visited.insert(Root);
663 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
664 // If there are no loads or stores, the access is dead. We mark that as
665 // a size zero access.
668 Instruction *I, *UsedI;
669 llvm::tie(UsedI, I) = Uses.pop_back_val();
671 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
672 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
675 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
676 Value *Op = SI->getOperand(0);
679 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
683 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
684 if (!GEP->hasAllZeroIndices())
686 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
687 !isa<SelectInst>(I)) {
691 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
693 if (Visited.insert(cast<Instruction>(*UI)))
694 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
695 } while (!Uses.empty());
700 void visitPHINode(PHINode &PN) {
704 return PI.setAborted(&PN);
706 // See if we already have computed info on this node.
707 std::pair<uint64_t, bool> &PHIInfo = P.PHIOrSelectSizes[&PN];
709 PHIInfo.second = true;
710 insertUse(PN, Offset, PHIInfo.first);
714 // Check for an unsafe use of the PHI node.
715 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&PN, PHIInfo.first))
716 return PI.setAborted(UnsafeI);
718 insertUse(PN, Offset, PHIInfo.first);
721 void visitSelectInst(SelectInst &SI) {
724 if (Value *Result = foldSelectInst(SI)) {
726 // If the result of the constant fold will be the pointer, recurse
727 // through the select as if we had RAUW'ed it.
733 return PI.setAborted(&SI);
735 // See if we already have computed info on this node.
736 std::pair<uint64_t, bool> &SelectInfo = P.PHIOrSelectSizes[&SI];
737 if (SelectInfo.first) {
738 SelectInfo.second = true;
739 insertUse(SI, Offset, SelectInfo.first);
743 // Check for an unsafe use of the PHI node.
744 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&SI, SelectInfo.first))
745 return PI.setAborted(UnsafeI);
747 insertUse(SI, Offset, SelectInfo.first);
750 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
751 void visitInstruction(Instruction &I) {
756 /// \brief Use adder for the alloca partitioning.
758 /// This class adds the uses of an alloca to all of the partitions which they
759 /// use. For splittable partitions, this can end up doing essentially a linear
760 /// walk of the partitions, but the number of steps remains bounded by the
761 /// total result instruction size:
762 /// - The number of partitions is a result of the number unsplittable
763 /// instructions using the alloca.
764 /// - The number of users of each partition is at worst the total number of
765 /// splittable instructions using the alloca.
766 /// Thus we will produce N * M instructions in the end, where N are the number
767 /// of unsplittable uses and M are the number of splittable. This visitor does
768 /// the exact same number of updates to the partitioning.
770 /// In the more common case, this visitor will leverage the fact that the
771 /// partition space is pre-sorted, and do a logarithmic search for the
772 /// partition needed, making the total visit a classical ((N + M) * log(N))
773 /// complexity operation.
774 class AllocaPartitioning::UseBuilder : public PtrUseVisitor<UseBuilder> {
775 friend class PtrUseVisitor<UseBuilder>;
776 friend class InstVisitor<UseBuilder>;
777 typedef PtrUseVisitor<UseBuilder> Base;
779 const uint64_t AllocSize;
780 AllocaPartitioning &P;
782 /// \brief Set to de-duplicate dead instructions found in the use walk.
783 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
786 UseBuilder(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
787 : PtrUseVisitor<UseBuilder>(TD),
788 AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())),
792 void markAsDead(Instruction &I) {
793 if (VisitedDeadInsts.insert(&I))
794 P.DeadUsers.push_back(&I);
797 void insertUse(Instruction &User, const APInt &Offset, uint64_t Size) {
798 // If the use has a zero size or extends outside of the allocation, record
799 // it as a dead use for elimination later.
800 if (Size == 0 || Offset.isNegative() || Offset.uge(AllocSize))
801 return markAsDead(User);
803 uint64_t BeginOffset = Offset.getZExtValue();
804 uint64_t EndOffset = BeginOffset + Size;
806 // Clamp the end offset to the end of the allocation. Note that this is
807 // formulated to handle even the case where "BeginOffset + Size" overflows.
808 assert(AllocSize >= BeginOffset); // Established above.
809 if (Size > AllocSize - BeginOffset)
810 EndOffset = AllocSize;
812 // NB: This only works if we have zero overlapping partitions.
813 iterator I = std::lower_bound(P.begin(), P.end(), BeginOffset);
814 if (I != P.begin() && llvm::prior(I)->EndOffset > BeginOffset)
816 iterator E = P.end();
817 bool IsSplit = llvm::next(I) != E && llvm::next(I)->BeginOffset < EndOffset;
818 for (; I != E && I->BeginOffset < EndOffset; ++I) {
819 PartitionUse NewPU(std::max(I->BeginOffset, BeginOffset),
820 std::min(I->EndOffset, EndOffset), U, IsSplit);
821 P.use_push_back(I, NewPU);
822 if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser()))
823 P.PHIOrSelectOpMap[U]
824 = std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1);
828 void visitBitCastInst(BitCastInst &BC) {
830 return markAsDead(BC);
832 return Base::visitBitCastInst(BC);
835 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
836 if (GEPI.use_empty())
837 return markAsDead(GEPI);
839 return Base::visitGetElementPtrInst(GEPI);
842 void visitLoadInst(LoadInst &LI) {
843 assert(IsOffsetKnown);
844 uint64_t Size = DL.getTypeStoreSize(LI.getType());
845 insertUse(LI, Offset, Size);
848 void visitStoreInst(StoreInst &SI) {
849 assert(IsOffsetKnown);
850 uint64_t Size = DL.getTypeStoreSize(SI.getOperand(0)->getType());
852 // If this memory access can be shown to *statically* extend outside the
853 // bounds of of the allocation, it's behavior is undefined, so simply
854 // ignore it. Note that this is more strict than the generic clamping
855 // behavior of insertUse.
856 if (Offset.isNegative() || Size > AllocSize ||
857 Offset.ugt(AllocSize - Size))
858 return markAsDead(SI);
860 insertUse(SI, Offset, Size);
863 void visitMemSetInst(MemSetInst &II) {
864 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
865 if ((Length && Length->getValue() == 0) ||
866 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
867 return markAsDead(II);
869 assert(IsOffsetKnown);
870 insertUse(II, Offset, Length ? Length->getLimitedValue()
871 : AllocSize - Offset.getLimitedValue());
874 void visitMemTransferInst(MemTransferInst &II) {
875 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
876 if ((Length && Length->getValue() == 0) ||
877 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
878 return markAsDead(II);
880 assert(IsOffsetKnown);
881 uint64_t Size = Length ? Length->getLimitedValue()
882 : AllocSize - Offset.getLimitedValue();
884 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
885 if (!II.isVolatile() && Offsets.DestEnd && Offsets.SourceEnd &&
886 Offsets.DestBegin == Offsets.SourceBegin)
887 return markAsDead(II); // Skip identity transfers without side-effects.
889 insertUse(II, Offset, Size);
892 void visitIntrinsicInst(IntrinsicInst &II) {
893 assert(IsOffsetKnown);
894 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
895 II.getIntrinsicID() == Intrinsic::lifetime_end);
897 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
898 insertUse(II, Offset, std::min(Length->getLimitedValue(),
899 AllocSize - Offset.getLimitedValue()));
902 void insertPHIOrSelect(Instruction &User, const APInt &Offset) {
903 uint64_t Size = P.PHIOrSelectSizes.lookup(&User).first;
905 // For PHI and select operands outside the alloca, we can't nuke the entire
906 // phi or select -- the other side might still be relevant, so we special
907 // case them here and use a separate structure to track the operands
908 // themselves which should be replaced with undef.
909 if ((Offset.isNegative() && Offset.uge(Size)) ||
910 (!Offset.isNegative() && Offset.uge(AllocSize))) {
911 P.DeadOperands.push_back(U);
915 insertUse(User, Offset, Size);
918 void visitPHINode(PHINode &PN) {
920 return markAsDead(PN);
922 assert(IsOffsetKnown);
923 insertPHIOrSelect(PN, Offset);
926 void visitSelectInst(SelectInst &SI) {
928 return markAsDead(SI);
930 if (Value *Result = foldSelectInst(SI)) {
932 // If the result of the constant fold will be the pointer, recurse
933 // through the select as if we had RAUW'ed it.
936 // Otherwise the operand to the select is dead, and we can replace it
938 P.DeadOperands.push_back(U);
943 assert(IsOffsetKnown);
944 insertPHIOrSelect(SI, Offset);
947 /// \brief Unreachable, we've already visited the alloca once.
948 void visitInstruction(Instruction &I) {
949 llvm_unreachable("Unhandled instruction in use builder.");
953 void AllocaPartitioning::splitAndMergePartitions() {
954 size_t NumDeadPartitions = 0;
956 // Track the range of splittable partitions that we pass when accumulating
957 // overlapping unsplittable partitions.
958 uint64_t SplitEndOffset = 0ull;
960 Partition New(0ull, 0ull, false);
962 for (unsigned i = 0, j = i, e = Partitions.size(); i != e; i = j) {
965 if (!Partitions[i].IsSplittable || New.BeginOffset == New.EndOffset) {
966 assert(New.BeginOffset == New.EndOffset);
969 assert(New.IsSplittable);
970 New.EndOffset = std::max(New.EndOffset, Partitions[i].EndOffset);
972 assert(New.BeginOffset != New.EndOffset);
974 // Scan the overlapping partitions.
975 while (j != e && New.EndOffset > Partitions[j].BeginOffset) {
976 // If the new partition we are forming is splittable, stop at the first
977 // unsplittable partition.
978 if (New.IsSplittable && !Partitions[j].IsSplittable)
981 // Grow the new partition to include any equally splittable range. 'j' is
982 // always equally splittable when New is splittable, but when New is not
983 // splittable, we may subsume some (or part of some) splitable partition
984 // without growing the new one.
985 if (New.IsSplittable == Partitions[j].IsSplittable) {
986 New.EndOffset = std::max(New.EndOffset, Partitions[j].EndOffset);
988 assert(!New.IsSplittable);
989 assert(Partitions[j].IsSplittable);
990 SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset);
993 Partitions[j].kill();
998 // If the new partition is splittable, chop off the end as soon as the
999 // unsplittable subsequent partition starts and ensure we eventually cover
1000 // the splittable area.
1001 if (j != e && New.IsSplittable) {
1002 SplitEndOffset = std::max(SplitEndOffset, New.EndOffset);
1003 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1006 // Add the new partition if it differs from the original one and is
1007 // non-empty. We can end up with an empty partition here if it was
1008 // splittable but there is an unsplittable one that starts at the same
1010 if (New != Partitions[i]) {
1011 if (New.BeginOffset != New.EndOffset)
1012 Partitions.push_back(New);
1013 // Mark the old one for removal.
1014 Partitions[i].kill();
1015 ++NumDeadPartitions;
1018 New.BeginOffset = New.EndOffset;
1019 if (!New.IsSplittable) {
1020 New.EndOffset = std::max(New.EndOffset, SplitEndOffset);
1021 if (j != e && !Partitions[j].IsSplittable)
1022 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1023 New.IsSplittable = true;
1024 // If there is a trailing splittable partition which won't be fused into
1025 // the next splittable partition go ahead and add it onto the partitions
1027 if (New.BeginOffset < New.EndOffset &&
1028 (j == e || !Partitions[j].IsSplittable ||
1029 New.EndOffset < Partitions[j].BeginOffset)) {
1030 Partitions.push_back(New);
1031 New.BeginOffset = New.EndOffset = 0ull;
1036 // Re-sort the partitions now that they have been split and merged into
1037 // disjoint set of partitions. Also remove any of the dead partitions we've
1038 // replaced in the process.
1039 std::sort(Partitions.begin(), Partitions.end());
1040 if (NumDeadPartitions) {
1041 assert(Partitions.back().isDead());
1042 assert((ptrdiff_t)NumDeadPartitions ==
1043 std::count(Partitions.begin(), Partitions.end(), Partitions.back()));
1045 Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end());
1048 AllocaPartitioning::AllocaPartitioning(const DataLayout &TD, AllocaInst &AI)
1050 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1053 PointerEscapingInstr(0) {
1054 PartitionBuilder PB(TD, AI, *this);
1055 PartitionBuilder::PtrInfo PtrI = PB.visitPtr(AI);
1056 if (PtrI.isEscaped() || PtrI.isAborted()) {
1057 // FIXME: We should sink the escape vs. abort info into the caller nicely,
1058 // possibly by just storing the PtrInfo in the AllocaPartitioning.
1059 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
1060 : PtrI.getAbortingInst();
1061 assert(PointerEscapingInstr && "Did not track a bad instruction");
1065 // Sort the uses. This arranges for the offsets to be in ascending order,
1066 // and the sizes to be in descending order.
1067 std::sort(Partitions.begin(), Partitions.end());
1069 // Remove any partitions from the back which are marked as dead.
1070 while (!Partitions.empty() && Partitions.back().isDead())
1071 Partitions.pop_back();
1073 if (Partitions.size() > 1) {
1074 // Intersect splittability for all partitions with equal offsets and sizes.
1075 // Then remove all but the first so that we have a sequence of non-equal but
1076 // potentially overlapping partitions.
1077 for (iterator I = Partitions.begin(), J = I, E = Partitions.end(); I != E;
1080 while (J != E && *I == *J) {
1081 I->IsSplittable &= J->IsSplittable;
1085 Partitions.erase(std::unique(Partitions.begin(), Partitions.end()),
1088 // Split splittable and merge unsplittable partitions into a disjoint set
1089 // of partitions over the used space of the allocation.
1090 splitAndMergePartitions();
1093 // Now build up the user lists for each of these disjoint partitions by
1094 // re-walking the recursive users of the alloca.
1095 Uses.resize(Partitions.size());
1096 UseBuilder UB(TD, AI, *this);
1097 PtrI = UB.visitPtr(AI);
1098 assert(!PtrI.isEscaped() && "Previously analyzed pointer now escapes!");
1099 assert(!PtrI.isAborted() && "Early aborted the visit of the pointer.");
1102 Type *AllocaPartitioning::getCommonType(iterator I) const {
1104 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
1105 Use *U = UI->getUse();
1107 continue; // Skip dead uses.
1108 if (isa<IntrinsicInst>(*U->getUser()))
1110 if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset)
1114 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser()))
1115 UserTy = LI->getType();
1116 else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser()))
1117 UserTy = SI->getValueOperand()->getType();
1119 return 0; // Bail if we have weird uses.
1121 if (IntegerType *ITy = dyn_cast<IntegerType>(UserTy)) {
1122 // If the type is larger than the partition, skip it. We only encounter
1123 // this for split integer operations where we want to use the type of the
1124 // entity causing the split.
1125 if (ITy->getBitWidth() > (I->EndOffset - I->BeginOffset)*8)
1128 // If we have found an integer type use covering the alloca, use that
1129 // regardless of the other types, as integers are often used for a "bucket
1134 if (Ty && Ty != UserTy)
1142 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1144 void AllocaPartitioning::print(raw_ostream &OS, const_iterator I,
1145 StringRef Indent) const {
1146 OS << Indent << "partition #" << (I - begin())
1147 << " [" << I->BeginOffset << "," << I->EndOffset << ")"
1148 << (I->IsSplittable ? " (splittable)" : "")
1149 << (Uses[I - begin()].empty() ? " (zero uses)" : "")
1153 void AllocaPartitioning::printUsers(raw_ostream &OS, const_iterator I,
1154 StringRef Indent) const {
1155 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
1157 continue; // Skip dead uses.
1158 OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") "
1159 << "used by: " << *UI->getUse()->getUser() << "\n";
1160 if (MemTransferInst *II =
1161 dyn_cast<MemTransferInst>(UI->getUse()->getUser())) {
1162 const MemTransferOffsets &MTO = MemTransferInstData.lookup(II);
1164 if (!MTO.IsSplittable)
1165 IsDest = UI->BeginOffset == MTO.DestBegin;
1167 IsDest = MTO.DestBegin != 0u;
1168 OS << Indent << " (original " << (IsDest ? "dest" : "source") << ": "
1169 << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin)
1170 << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n";
1175 void AllocaPartitioning::print(raw_ostream &OS) const {
1176 if (PointerEscapingInstr) {
1177 OS << "No partitioning for alloca: " << AI << "\n"
1178 << " A pointer to this alloca escaped by:\n"
1179 << " " << *PointerEscapingInstr << "\n";
1183 OS << "Partitioning of alloca: " << AI << "\n";
1184 for (const_iterator I = begin(), E = end(); I != E; ++I) {
1190 void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); }
1191 void AllocaPartitioning::dump() const { print(dbgs()); }
1193 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1197 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1199 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1200 /// the loads and stores of an alloca instruction, as well as updating its
1201 /// debug information. This is used when a domtree is unavailable and thus
1202 /// mem2reg in its full form can't be used to handle promotion of allocas to
1204 class AllocaPromoter : public LoadAndStorePromoter {
1208 SmallVector<DbgDeclareInst *, 4> DDIs;
1209 SmallVector<DbgValueInst *, 4> DVIs;
1212 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1213 AllocaInst &AI, DIBuilder &DIB)
1214 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1216 void run(const SmallVectorImpl<Instruction*> &Insts) {
1217 // Remember which alloca we're promoting (for isInstInList).
1218 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
1219 for (Value::use_iterator UI = DebugNode->use_begin(),
1220 UE = DebugNode->use_end();
1222 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
1223 DDIs.push_back(DDI);
1224 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
1225 DVIs.push_back(DVI);
1228 LoadAndStorePromoter::run(Insts);
1229 AI.eraseFromParent();
1230 while (!DDIs.empty())
1231 DDIs.pop_back_val()->eraseFromParent();
1232 while (!DVIs.empty())
1233 DVIs.pop_back_val()->eraseFromParent();
1236 virtual bool isInstInList(Instruction *I,
1237 const SmallVectorImpl<Instruction*> &Insts) const {
1238 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1239 return LI->getOperand(0) == &AI;
1240 return cast<StoreInst>(I)->getPointerOperand() == &AI;
1243 virtual void updateDebugInfo(Instruction *Inst) const {
1244 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
1245 E = DDIs.end(); I != E; ++I) {
1246 DbgDeclareInst *DDI = *I;
1247 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1248 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1249 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1250 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1252 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
1253 E = DVIs.end(); I != E; ++I) {
1254 DbgValueInst *DVI = *I;
1256 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1257 // If an argument is zero extended then use argument directly. The ZExt
1258 // may be zapped by an optimization pass in future.
1259 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1260 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1261 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1262 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1264 Arg = SI->getOperand(0);
1265 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1266 Arg = LI->getOperand(0);
1270 Instruction *DbgVal =
1271 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1273 DbgVal->setDebugLoc(DVI->getDebugLoc());
1277 } // end anon namespace
1281 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1283 /// This pass takes allocations which can be completely analyzed (that is, they
1284 /// don't escape) and tries to turn them into scalar SSA values. There are
1285 /// a few steps to this process.
1287 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1288 /// are used to try to split them into smaller allocations, ideally of
1289 /// a single scalar data type. It will split up memcpy and memset accesses
1290 /// as necessary and try to isolate individual scalar accesses.
1291 /// 2) It will transform accesses into forms which are suitable for SSA value
1292 /// promotion. This can be replacing a memset with a scalar store of an
1293 /// integer value, or it can involve speculating operations on a PHI or
1294 /// select to be a PHI or select of the results.
1295 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1296 /// onto insert and extract operations on a vector value, and convert them to
1297 /// this form. By doing so, it will enable promotion of vector aggregates to
1298 /// SSA vector values.
1299 class SROA : public FunctionPass {
1300 const bool RequiresDomTree;
1303 const DataLayout *TD;
1306 /// \brief Worklist of alloca instructions to simplify.
1308 /// Each alloca in the function is added to this. Each new alloca formed gets
1309 /// added to it as well to recursively simplify unless that alloca can be
1310 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1311 /// the one being actively rewritten, we add it back onto the list if not
1312 /// already present to ensure it is re-visited.
1313 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
1315 /// \brief A collection of instructions to delete.
1316 /// We try to batch deletions to simplify code and make things a bit more
1318 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
1320 /// \brief Post-promotion worklist.
1322 /// Sometimes we discover an alloca which has a high probability of becoming
1323 /// viable for SROA after a round of promotion takes place. In those cases,
1324 /// the alloca is enqueued here for re-processing.
1326 /// Note that we have to be very careful to clear allocas out of this list in
1327 /// the event they are deleted.
1328 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
1330 /// \brief A collection of alloca instructions we can directly promote.
1331 std::vector<AllocaInst *> PromotableAllocas;
1334 SROA(bool RequiresDomTree = true)
1335 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
1336 C(0), TD(0), DT(0) {
1337 initializeSROAPass(*PassRegistry::getPassRegistry());
1339 bool runOnFunction(Function &F);
1340 void getAnalysisUsage(AnalysisUsage &AU) const;
1342 const char *getPassName() const { return "SROA"; }
1346 friend class PHIOrSelectSpeculator;
1347 friend class AllocaPartitionRewriter;
1348 friend class AllocaPartitionVectorRewriter;
1350 bool rewriteAllocaPartition(AllocaInst &AI,
1351 AllocaPartitioning &P,
1352 AllocaPartitioning::iterator PI);
1353 bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P);
1354 bool runOnAlloca(AllocaInst &AI);
1355 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
1356 bool promoteAllocas(Function &F);
1362 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1363 return new SROA(RequiresDomTree);
1366 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1368 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
1369 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1373 /// \brief Visitor to speculate PHIs and Selects where possible.
1374 class PHIOrSelectSpeculator : public InstVisitor<PHIOrSelectSpeculator> {
1375 // Befriend the base class so it can delegate to private visit methods.
1376 friend class llvm::InstVisitor<PHIOrSelectSpeculator>;
1378 const DataLayout &TD;
1379 AllocaPartitioning &P;
1383 PHIOrSelectSpeculator(const DataLayout &TD, AllocaPartitioning &P, SROA &Pass)
1384 : TD(TD), P(P), Pass(Pass) {}
1386 /// \brief Visit the users of an alloca partition and rewrite them.
1387 void visitUsers(AllocaPartitioning::const_iterator PI) {
1388 // Note that we need to use an index here as the underlying vector of uses
1389 // may be grown during speculation. However, we never need to re-visit the
1390 // new uses, and so we can use the initial size bound.
1391 for (unsigned Idx = 0, Size = P.use_size(PI); Idx != Size; ++Idx) {
1392 const AllocaPartitioning::PartitionUse &PU = P.getUse(PI, Idx);
1394 continue; // Skip dead use.
1396 visit(cast<Instruction>(PU.getUse()->getUser()));
1401 // By default, skip this instruction.
1402 void visitInstruction(Instruction &I) {}
1404 /// PHI instructions that use an alloca and are subsequently loaded can be
1405 /// rewritten to load both input pointers in the pred blocks and then PHI the
1406 /// results, allowing the load of the alloca to be promoted.
1408 /// %P2 = phi [i32* %Alloca, i32* %Other]
1409 /// %V = load i32* %P2
1411 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1413 /// %V2 = load i32* %Other
1415 /// %V = phi [i32 %V1, i32 %V2]
1417 /// We can do this to a select if its only uses are loads and if the operands
1418 /// to the select can be loaded unconditionally.
1420 /// FIXME: This should be hoisted into a generic utility, likely in
1421 /// Transforms/Util/Local.h
1422 bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
1423 // For now, we can only do this promotion if the load is in the same block
1424 // as the PHI, and if there are no stores between the phi and load.
1425 // TODO: Allow recursive phi users.
1426 // TODO: Allow stores.
1427 BasicBlock *BB = PN.getParent();
1428 unsigned MaxAlign = 0;
1429 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
1431 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1432 if (LI == 0 || !LI->isSimple()) return false;
1434 // For now we only allow loads in the same block as the PHI. This is
1435 // a common case that happens when instcombine merges two loads through
1437 if (LI->getParent() != BB) return false;
1439 // Ensure that there are no instructions between the PHI and the load that
1441 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1442 if (BBI->mayWriteToMemory())
1445 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1446 Loads.push_back(LI);
1449 // We can only transform this if it is safe to push the loads into the
1450 // predecessor blocks. The only thing to watch out for is that we can't put
1451 // a possibly trapping load in the predecessor if it is a critical edge.
1452 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1453 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1454 Value *InVal = PN.getIncomingValue(Idx);
1456 // If the value is produced by the terminator of the predecessor (an
1457 // invoke) or it has side-effects, there is no valid place to put a load
1458 // in the predecessor.
1459 if (TI == InVal || TI->mayHaveSideEffects())
1462 // If the predecessor has a single successor, then the edge isn't
1464 if (TI->getNumSuccessors() == 1)
1467 // If this pointer is always safe to load, or if we can prove that there
1468 // is already a load in the block, then we can move the load to the pred
1470 if (InVal->isDereferenceablePointer() ||
1471 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
1480 void visitPHINode(PHINode &PN) {
1481 DEBUG(dbgs() << " original: " << PN << "\n");
1483 SmallVector<LoadInst *, 4> Loads;
1484 if (!isSafePHIToSpeculate(PN, Loads))
1487 assert(!Loads.empty());
1489 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1490 IRBuilder<> PHIBuilder(&PN);
1491 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1492 PN.getName() + ".sroa.speculated");
1494 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1495 // matter which one we get and if any differ.
1496 LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
1497 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1498 unsigned Align = SomeLoad->getAlignment();
1500 // Rewrite all loads of the PN to use the new PHI.
1502 LoadInst *LI = Loads.pop_back_val();
1503 LI->replaceAllUsesWith(NewPN);
1504 Pass.DeadInsts.insert(LI);
1505 } while (!Loads.empty());
1507 // Inject loads into all of the pred blocks.
1508 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1509 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1510 TerminatorInst *TI = Pred->getTerminator();
1511 Use *InUse = &PN.getOperandUse(PN.getOperandNumForIncomingValue(Idx));
1512 Value *InVal = PN.getIncomingValue(Idx);
1513 IRBuilder<> PredBuilder(TI);
1516 = PredBuilder.CreateLoad(InVal, (PN.getName() + ".sroa.speculate.load." +
1518 ++NumLoadsSpeculated;
1519 Load->setAlignment(Align);
1521 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1522 NewPN->addIncoming(Load, Pred);
1524 Instruction *Ptr = dyn_cast<Instruction>(InVal);
1526 // No uses to rewrite.
1529 // Try to lookup and rewrite any partition uses corresponding to this phi
1531 AllocaPartitioning::iterator PI
1532 = P.findPartitionForPHIOrSelectOperand(InUse);
1536 // Replace the Use in the PartitionUse for this operand with the Use
1538 AllocaPartitioning::use_iterator UI
1539 = P.findPartitionUseForPHIOrSelectOperand(InUse);
1540 assert(isa<PHINode>(*UI->getUse()->getUser()));
1541 UI->setUse(&Load->getOperandUse(Load->getPointerOperandIndex()));
1543 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1546 /// Select instructions that use an alloca and are subsequently loaded can be
1547 /// rewritten to load both input pointers and then select between the result,
1548 /// allowing the load of the alloca to be promoted.
1550 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1551 /// %V = load i32* %P2
1553 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1554 /// %V2 = load i32* %Other
1555 /// %V = select i1 %cond, i32 %V1, i32 %V2
1557 /// We can do this to a select if its only uses are loads and if the operand
1558 /// to the select can be loaded unconditionally.
1559 bool isSafeSelectToSpeculate(SelectInst &SI,
1560 SmallVectorImpl<LoadInst *> &Loads) {
1561 Value *TValue = SI.getTrueValue();
1562 Value *FValue = SI.getFalseValue();
1563 bool TDerefable = TValue->isDereferenceablePointer();
1564 bool FDerefable = FValue->isDereferenceablePointer();
1566 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
1568 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1569 if (LI == 0 || !LI->isSimple()) return false;
1571 // Both operands to the select need to be dereferencable, either
1572 // absolutely (e.g. allocas) or at this point because we can see other
1574 if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
1575 LI->getAlignment(), &TD))
1577 if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
1578 LI->getAlignment(), &TD))
1580 Loads.push_back(LI);
1586 void visitSelectInst(SelectInst &SI) {
1587 DEBUG(dbgs() << " original: " << SI << "\n");
1589 // If the select isn't safe to speculate, just use simple logic to emit it.
1590 SmallVector<LoadInst *, 4> Loads;
1591 if (!isSafeSelectToSpeculate(SI, Loads))
1594 IRBuilder<> IRB(&SI);
1595 Use *Ops[2] = { &SI.getOperandUse(1), &SI.getOperandUse(2) };
1596 AllocaPartitioning::iterator PIs[2];
1597 AllocaPartitioning::PartitionUse PUs[2];
1598 for (unsigned i = 0, e = 2; i != e; ++i) {
1599 PIs[i] = P.findPartitionForPHIOrSelectOperand(Ops[i]);
1600 if (PIs[i] != P.end()) {
1601 // If the pointer is within the partitioning, remove the select from
1602 // its uses. We'll add in the new loads below.
1603 AllocaPartitioning::use_iterator UI
1604 = P.findPartitionUseForPHIOrSelectOperand(Ops[i]);
1606 // Clear out the use here so that the offsets into the use list remain
1607 // stable but this use is ignored when rewriting.
1612 Value *TV = SI.getTrueValue();
1613 Value *FV = SI.getFalseValue();
1614 // Replace the loads of the select with a select of two loads.
1615 while (!Loads.empty()) {
1616 LoadInst *LI = Loads.pop_back_val();
1618 IRB.SetInsertPoint(LI);
1620 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1622 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1623 NumLoadsSpeculated += 2;
1625 // Transfer alignment and TBAA info if present.
1626 TL->setAlignment(LI->getAlignment());
1627 FL->setAlignment(LI->getAlignment());
1628 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1629 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1630 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1633 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1634 LI->getName() + ".sroa.speculated");
1636 LoadInst *Loads[2] = { TL, FL };
1637 for (unsigned i = 0, e = 2; i != e; ++i) {
1638 if (PIs[i] != P.end()) {
1639 Use *LoadUse = &Loads[i]->getOperandUse(0);
1640 assert(PUs[i].getUse()->get() == LoadUse->get());
1641 PUs[i].setUse(LoadUse);
1642 P.use_push_back(PIs[i], PUs[i]);
1646 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1647 LI->replaceAllUsesWith(V);
1648 Pass.DeadInsts.insert(LI);
1654 /// \brief Build a GEP out of a base pointer and indices.
1656 /// This will return the BasePtr if that is valid, or build a new GEP
1657 /// instruction using the IRBuilder if GEP-ing is needed.
1658 static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr,
1659 SmallVectorImpl<Value *> &Indices,
1660 const Twine &Prefix) {
1661 if (Indices.empty())
1664 // A single zero index is a no-op, so check for this and avoid building a GEP
1666 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1669 return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx");
1672 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1673 /// TargetTy without changing the offset of the pointer.
1675 /// This routine assumes we've already established a properly offset GEP with
1676 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1677 /// zero-indices down through type layers until we find one the same as
1678 /// TargetTy. If we can't find one with the same type, we at least try to use
1679 /// one with the same size. If none of that works, we just produce the GEP as
1680 /// indicated by Indices to have the correct offset.
1681 static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const DataLayout &TD,
1682 Value *BasePtr, Type *Ty, Type *TargetTy,
1683 SmallVectorImpl<Value *> &Indices,
1684 const Twine &Prefix) {
1686 return buildGEP(IRB, BasePtr, Indices, Prefix);
1688 // See if we can descend into a struct and locate a field with the correct
1690 unsigned NumLayers = 0;
1691 Type *ElementTy = Ty;
1693 if (ElementTy->isPointerTy())
1695 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1696 ElementTy = SeqTy->getElementType();
1697 // Note that we use the default address space as this index is over an
1698 // array or a vector, not a pointer.
1699 Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(0), 0)));
1700 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1701 if (STy->element_begin() == STy->element_end())
1702 break; // Nothing left to descend into.
1703 ElementTy = *STy->element_begin();
1704 Indices.push_back(IRB.getInt32(0));
1709 } while (ElementTy != TargetTy);
1710 if (ElementTy != TargetTy)
1711 Indices.erase(Indices.end() - NumLayers, Indices.end());
1713 return buildGEP(IRB, BasePtr, Indices, Prefix);
1716 /// \brief Recursively compute indices for a natural GEP.
1718 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1719 /// element types adding appropriate indices for the GEP.
1720 static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const DataLayout &TD,
1721 Value *Ptr, Type *Ty, APInt &Offset,
1723 SmallVectorImpl<Value *> &Indices,
1724 const Twine &Prefix) {
1726 return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix);
1728 // We can't recurse through pointer types.
1729 if (Ty->isPointerTy())
1732 // We try to analyze GEPs over vectors here, but note that these GEPs are
1733 // extremely poorly defined currently. The long-term goal is to remove GEPing
1734 // over a vector from the IR completely.
1735 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1736 unsigned ElementSizeInBits = TD.getTypeSizeInBits(VecTy->getScalarType());
1737 if (ElementSizeInBits % 8)
1738 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1739 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1740 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1741 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1743 Offset -= NumSkippedElements * ElementSize;
1744 Indices.push_back(IRB.getInt(NumSkippedElements));
1745 return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
1746 Offset, TargetTy, Indices, Prefix);
1749 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1750 Type *ElementTy = ArrTy->getElementType();
1751 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1752 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1753 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1756 Offset -= NumSkippedElements * ElementSize;
1757 Indices.push_back(IRB.getInt(NumSkippedElements));
1758 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1762 StructType *STy = dyn_cast<StructType>(Ty);
1766 const StructLayout *SL = TD.getStructLayout(STy);
1767 uint64_t StructOffset = Offset.getZExtValue();
1768 if (StructOffset >= SL->getSizeInBytes())
1770 unsigned Index = SL->getElementContainingOffset(StructOffset);
1771 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1772 Type *ElementTy = STy->getElementType(Index);
1773 if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
1774 return 0; // The offset points into alignment padding.
1776 Indices.push_back(IRB.getInt32(Index));
1777 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1781 /// \brief Get a natural GEP from a base pointer to a particular offset and
1782 /// resulting in a particular type.
1784 /// The goal is to produce a "natural" looking GEP that works with the existing
1785 /// composite types to arrive at the appropriate offset and element type for
1786 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1787 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1788 /// Indices, and setting Ty to the result subtype.
1790 /// If no natural GEP can be constructed, this function returns null.
1791 static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const DataLayout &TD,
1792 Value *Ptr, APInt Offset, Type *TargetTy,
1793 SmallVectorImpl<Value *> &Indices,
1794 const Twine &Prefix) {
1795 PointerType *Ty = cast<PointerType>(Ptr->getType());
1797 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1799 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1802 Type *ElementTy = Ty->getElementType();
1803 if (!ElementTy->isSized())
1804 return 0; // We can't GEP through an unsized element.
1805 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1806 if (ElementSize == 0)
1807 return 0; // Zero-length arrays can't help us build a natural GEP.
1808 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1810 Offset -= NumSkippedElements * ElementSize;
1811 Indices.push_back(IRB.getInt(NumSkippedElements));
1812 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1816 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1817 /// resulting pointer has PointerTy.
1819 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1820 /// and produces the pointer type desired. Where it cannot, it will try to use
1821 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1822 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1823 /// bitcast to the type.
1825 /// The strategy for finding the more natural GEPs is to peel off layers of the
1826 /// pointer, walking back through bit casts and GEPs, searching for a base
1827 /// pointer from which we can compute a natural GEP with the desired
1828 /// properties. The algorithm tries to fold as many constant indices into
1829 /// a single GEP as possible, thus making each GEP more independent of the
1830 /// surrounding code.
1831 static Value *getAdjustedPtr(IRBuilder<> &IRB, const DataLayout &TD,
1832 Value *Ptr, APInt Offset, Type *PointerTy,
1833 const Twine &Prefix) {
1834 // Even though we don't look through PHI nodes, we could be called on an
1835 // instruction in an unreachable block, which may be on a cycle.
1836 SmallPtrSet<Value *, 4> Visited;
1837 Visited.insert(Ptr);
1838 SmallVector<Value *, 4> Indices;
1840 // We may end up computing an offset pointer that has the wrong type. If we
1841 // never are able to compute one directly that has the correct type, we'll
1842 // fall back to it, so keep it around here.
1843 Value *OffsetPtr = 0;
1845 // Remember any i8 pointer we come across to re-use if we need to do a raw
1848 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1850 Type *TargetTy = PointerTy->getPointerElementType();
1853 // First fold any existing GEPs into the offset.
1854 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1855 APInt GEPOffset(Offset.getBitWidth(), 0);
1856 if (!GEP->accumulateConstantOffset(TD, GEPOffset))
1858 Offset += GEPOffset;
1859 Ptr = GEP->getPointerOperand();
1860 if (!Visited.insert(Ptr))
1864 // See if we can perform a natural GEP here.
1866 if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
1868 if (P->getType() == PointerTy) {
1869 // Zap any offset pointer that we ended up computing in previous rounds.
1870 if (OffsetPtr && OffsetPtr->use_empty())
1871 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1872 I->eraseFromParent();
1880 // Stash this pointer if we've found an i8*.
1881 if (Ptr->getType()->isIntegerTy(8)) {
1883 Int8PtrOffset = Offset;
1886 // Peel off a layer of the pointer and update the offset appropriately.
1887 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1888 Ptr = cast<Operator>(Ptr)->getOperand(0);
1889 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1890 if (GA->mayBeOverridden())
1892 Ptr = GA->getAliasee();
1896 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1897 } while (Visited.insert(Ptr));
1901 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1902 Prefix + ".raw_cast");
1903 Int8PtrOffset = Offset;
1906 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1907 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1908 Prefix + ".raw_idx");
1912 // On the off chance we were targeting i8*, guard the bitcast here.
1913 if (Ptr->getType() != PointerTy)
1914 Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast");
1919 /// \brief Test whether we can convert a value from the old to the new type.
1921 /// This predicate should be used to guard calls to convertValue in order to
1922 /// ensure that we only try to convert viable values. The strategy is that we
1923 /// will peel off single element struct and array wrappings to get to an
1924 /// underlying value, and convert that value.
1925 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1928 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1929 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1930 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1932 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1934 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1937 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1938 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1940 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
1948 /// \brief Generic routine to convert an SSA value to a value of a different
1951 /// This will try various different casting techniques, such as bitcasts,
1952 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
1953 /// two types for viability with this routine.
1954 static Value *convertValue(const DataLayout &DL, IRBuilder<> &IRB, Value *V,
1956 assert(canConvertValue(DL, V->getType(), Ty) &&
1957 "Value not convertable to type");
1958 if (V->getType() == Ty)
1960 if (IntegerType *OldITy = dyn_cast<IntegerType>(V->getType()))
1961 if (IntegerType *NewITy = dyn_cast<IntegerType>(Ty))
1962 if (NewITy->getBitWidth() > OldITy->getBitWidth())
1963 return IRB.CreateZExt(V, NewITy);
1964 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
1965 return IRB.CreateIntToPtr(V, Ty);
1966 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
1967 return IRB.CreatePtrToInt(V, Ty);
1969 return IRB.CreateBitCast(V, Ty);
1972 /// \brief Test whether the given alloca partition can be promoted to a vector.
1974 /// This is a quick test to check whether we can rewrite a particular alloca
1975 /// partition (and its newly formed alloca) into a vector alloca with only
1976 /// whole-vector loads and stores such that it could be promoted to a vector
1977 /// SSA value. We only can ensure this for a limited set of operations, and we
1978 /// don't want to do the rewrites unless we are confident that the result will
1979 /// be promotable, so we have an early test here.
1980 static bool isVectorPromotionViable(const DataLayout &TD,
1982 AllocaPartitioning &P,
1983 uint64_t PartitionBeginOffset,
1984 uint64_t PartitionEndOffset,
1985 AllocaPartitioning::const_use_iterator I,
1986 AllocaPartitioning::const_use_iterator E) {
1987 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1991 uint64_t ElementSize = TD.getTypeSizeInBits(Ty->getScalarType());
1993 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1994 // that aren't byte sized.
1995 if (ElementSize % 8)
1997 assert((TD.getTypeSizeInBits(Ty) % 8) == 0 &&
1998 "vector size not a multiple of element size?");
2001 for (; I != E; ++I) {
2002 Use *U = I->getUse();
2004 continue; // Skip dead use.
2006 uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
2007 uint64_t BeginIndex = BeginOffset / ElementSize;
2008 if (BeginIndex * ElementSize != BeginOffset ||
2009 BeginIndex >= Ty->getNumElements())
2011 uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
2012 uint64_t EndIndex = EndOffset / ElementSize;
2013 if (EndIndex * ElementSize != EndOffset ||
2014 EndIndex > Ty->getNumElements())
2017 assert(EndIndex > BeginIndex && "Empty vector!");
2018 uint64_t NumElements = EndIndex - BeginIndex;
2020 = (NumElements == 1) ? Ty->getElementType()
2021 : VectorType::get(Ty->getElementType(), NumElements);
2023 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
2024 if (MI->isVolatile())
2026 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(U->getUser())) {
2027 const AllocaPartitioning::MemTransferOffsets &MTO
2028 = P.getMemTransferOffsets(*MTI);
2029 if (!MTO.IsSplittable)
2032 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
2033 // Disable vector promotion when there are loads or stores of an FCA.
2035 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
2036 if (LI->isVolatile())
2038 if (!canConvertValue(TD, PartitionTy, LI->getType()))
2040 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
2041 if (SI->isVolatile())
2043 if (!canConvertValue(TD, SI->getValueOperand()->getType(), PartitionTy))
2052 /// \brief Test whether the given alloca partition's integer operations can be
2053 /// widened to promotable ones.
2055 /// This is a quick test to check whether we can rewrite the integer loads and
2056 /// stores to a particular alloca into wider loads and stores and be able to
2057 /// promote the resulting alloca.
2058 static bool isIntegerWideningViable(const DataLayout &TD,
2060 uint64_t AllocBeginOffset,
2061 AllocaPartitioning &P,
2062 AllocaPartitioning::const_use_iterator I,
2063 AllocaPartitioning::const_use_iterator E) {
2064 uint64_t SizeInBits = TD.getTypeSizeInBits(AllocaTy);
2065 // Don't create integer types larger than the maximum bitwidth.
2066 if (SizeInBits > IntegerType::MAX_INT_BITS)
2069 // Don't try to handle allocas with bit-padding.
2070 if (SizeInBits != TD.getTypeStoreSizeInBits(AllocaTy))
2073 // We need to ensure that an integer type with the appropriate bitwidth can
2074 // be converted to the alloca type, whatever that is. We don't want to force
2075 // the alloca itself to have an integer type if there is a more suitable one.
2076 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2077 if (!canConvertValue(TD, AllocaTy, IntTy) ||
2078 !canConvertValue(TD, IntTy, AllocaTy))
2081 uint64_t Size = TD.getTypeStoreSize(AllocaTy);
2083 // Check the uses to ensure the uses are (likely) promotable integer uses.
2084 // Also ensure that the alloca has a covering load or store. We don't want
2085 // to widen the integer operations only to fail to promote due to some other
2086 // unsplittable entry (which we may make splittable later).
2087 bool WholeAllocaOp = false;
2088 for (; I != E; ++I) {
2089 Use *U = I->getUse();
2091 continue; // Skip dead use.
2093 uint64_t RelBegin = I->BeginOffset - AllocBeginOffset;
2094 uint64_t RelEnd = I->EndOffset - AllocBeginOffset;
2096 // We can't reasonably handle cases where the load or store extends past
2097 // the end of the aloca's type and into its padding.
2101 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
2102 if (LI->isVolatile())
2104 if (RelBegin == 0 && RelEnd == Size)
2105 WholeAllocaOp = true;
2106 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
2107 if (ITy->getBitWidth() < TD.getTypeStoreSizeInBits(ITy))
2111 // Non-integer loads need to be convertible from the alloca type so that
2112 // they are promotable.
2113 if (RelBegin != 0 || RelEnd != Size ||
2114 !canConvertValue(TD, AllocaTy, LI->getType()))
2116 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
2117 Type *ValueTy = SI->getValueOperand()->getType();
2118 if (SI->isVolatile())
2120 if (RelBegin == 0 && RelEnd == Size)
2121 WholeAllocaOp = true;
2122 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2123 if (ITy->getBitWidth() < TD.getTypeStoreSizeInBits(ITy))
2127 // Non-integer stores need to be convertible to the alloca type so that
2128 // they are promotable.
2129 if (RelBegin != 0 || RelEnd != Size ||
2130 !canConvertValue(TD, ValueTy, AllocaTy))
2132 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
2133 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
2135 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(U->getUser())) {
2136 const AllocaPartitioning::MemTransferOffsets &MTO
2137 = P.getMemTransferOffsets(*MTI);
2138 if (!MTO.IsSplittable)
2141 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
2142 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2143 II->getIntrinsicID() != Intrinsic::lifetime_end)
2149 return WholeAllocaOp;
2152 static Value *extractInteger(const DataLayout &DL, IRBuilder<> &IRB, Value *V,
2153 IntegerType *Ty, uint64_t Offset,
2154 const Twine &Name) {
2155 DEBUG(dbgs() << " start: " << *V << "\n");
2156 IntegerType *IntTy = cast<IntegerType>(V->getType());
2157 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2158 "Element extends past full value");
2159 uint64_t ShAmt = 8*Offset;
2160 if (DL.isBigEndian())
2161 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2163 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2164 DEBUG(dbgs() << " shifted: " << *V << "\n");
2166 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2167 "Cannot extract to a larger integer!");
2169 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2170 DEBUG(dbgs() << " trunced: " << *V << "\n");
2175 static Value *insertInteger(const DataLayout &DL, IRBuilder<> &IRB, Value *Old,
2176 Value *V, uint64_t Offset, const Twine &Name) {
2177 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2178 IntegerType *Ty = cast<IntegerType>(V->getType());
2179 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2180 "Cannot insert a larger integer!");
2181 DEBUG(dbgs() << " start: " << *V << "\n");
2183 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2184 DEBUG(dbgs() << " extended: " << *V << "\n");
2186 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2187 "Element store outside of alloca store");
2188 uint64_t ShAmt = 8*Offset;
2189 if (DL.isBigEndian())
2190 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2192 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2193 DEBUG(dbgs() << " shifted: " << *V << "\n");
2196 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2197 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2198 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2199 DEBUG(dbgs() << " masked: " << *Old << "\n");
2200 V = IRB.CreateOr(Old, V, Name + ".insert");
2201 DEBUG(dbgs() << " inserted: " << *V << "\n");
2206 static Value *extractVector(IRBuilder<> &IRB, Value *V,
2207 unsigned BeginIndex, unsigned EndIndex,
2208 const Twine &Name) {
2209 VectorType *VecTy = cast<VectorType>(V->getType());
2210 unsigned NumElements = EndIndex - BeginIndex;
2211 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2213 if (NumElements == VecTy->getNumElements())
2216 if (NumElements == 1) {
2217 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2219 DEBUG(dbgs() << " extract: " << *V << "\n");
2223 SmallVector<Constant*, 8> Mask;
2224 Mask.reserve(NumElements);
2225 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2226 Mask.push_back(IRB.getInt32(i));
2227 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2228 ConstantVector::get(Mask),
2230 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2234 static Value *insertVector(IRBuilder<> &IRB, Value *Old, Value *V,
2235 unsigned BeginIndex, const Twine &Name) {
2236 VectorType *VecTy = cast<VectorType>(Old->getType());
2237 assert(VecTy && "Can only insert a vector into a vector");
2239 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2241 // Single element to insert.
2242 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2244 DEBUG(dbgs() << " insert: " << *V << "\n");
2248 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2249 "Too many elements!");
2250 if (Ty->getNumElements() == VecTy->getNumElements()) {
2251 assert(V->getType() == VecTy && "Vector type mismatch");
2254 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2256 // When inserting a smaller vector into the larger to store, we first
2257 // use a shuffle vector to widen it with undef elements, and then
2258 // a second shuffle vector to select between the loaded vector and the
2260 SmallVector<Constant*, 8> Mask;
2261 Mask.reserve(VecTy->getNumElements());
2262 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2263 if (i >= BeginIndex && i < EndIndex)
2264 Mask.push_back(IRB.getInt32(i - BeginIndex));
2266 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2267 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2268 ConstantVector::get(Mask),
2270 DEBUG(dbgs() << " shuffle1: " << *V << "\n");
2273 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2274 if (i >= BeginIndex && i < EndIndex)
2275 Mask.push_back(IRB.getInt32(i));
2277 Mask.push_back(IRB.getInt32(i + VecTy->getNumElements()));
2278 V = IRB.CreateShuffleVector(V, Old, ConstantVector::get(Mask),
2280 DEBUG(dbgs() << " shuffle2: " << *V << "\n");
2285 /// \brief Visitor to rewrite instructions using a partition of an alloca to
2286 /// use a new alloca.
2288 /// Also implements the rewriting to vector-based accesses when the partition
2289 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2291 class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
2293 // Befriend the base class so it can delegate to private visit methods.
2294 friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
2296 const DataLayout &TD;
2297 AllocaPartitioning &P;
2299 AllocaInst &OldAI, &NewAI;
2300 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2303 // If we are rewriting an alloca partition which can be written as pure
2304 // vector operations, we stash extra information here. When VecTy is
2305 // non-null, we have some strict guarantees about the rewritten alloca:
2306 // - The new alloca is exactly the size of the vector type here.
2307 // - The accesses all either map to the entire vector or to a single
2309 // - The set of accessing instructions is only one of those handled above
2310 // in isVectorPromotionViable. Generally these are the same access kinds
2311 // which are promotable via mem2reg.
2314 uint64_t ElementSize;
2316 // This is a convenience and flag variable that will be null unless the new
2317 // alloca's integer operations should be widened to this integer type due to
2318 // passing isIntegerWideningViable above. If it is non-null, the desired
2319 // integer type will be stored here for easy access during rewriting.
2322 // The offset of the partition user currently being rewritten.
2323 uint64_t BeginOffset, EndOffset;
2326 Instruction *OldPtr;
2328 // The name prefix to use when rewriting instructions for this alloca.
2329 std::string NamePrefix;
2332 AllocaPartitionRewriter(const DataLayout &TD, AllocaPartitioning &P,
2333 AllocaPartitioning::iterator PI,
2334 SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
2335 uint64_t NewBeginOffset, uint64_t NewEndOffset)
2336 : TD(TD), P(P), Pass(Pass),
2337 OldAI(OldAI), NewAI(NewAI),
2338 NewAllocaBeginOffset(NewBeginOffset),
2339 NewAllocaEndOffset(NewEndOffset),
2340 NewAllocaTy(NewAI.getAllocatedType()),
2341 VecTy(), ElementTy(), ElementSize(), IntTy(),
2342 BeginOffset(), EndOffset(), IsSplit(), OldUse(), OldPtr() {
2345 /// \brief Visit the users of the alloca partition and rewrite them.
2346 bool visitUsers(AllocaPartitioning::const_use_iterator I,
2347 AllocaPartitioning::const_use_iterator E) {
2348 if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P,
2349 NewAllocaBeginOffset, NewAllocaEndOffset,
2352 VecTy = cast<VectorType>(NewAI.getAllocatedType());
2353 ElementTy = VecTy->getElementType();
2354 assert((TD.getTypeSizeInBits(VecTy->getScalarType()) % 8) == 0 &&
2355 "Only multiple-of-8 sized vector elements are viable");
2356 ElementSize = TD.getTypeSizeInBits(VecTy->getScalarType()) / 8;
2357 } else if (isIntegerWideningViable(TD, NewAI.getAllocatedType(),
2358 NewAllocaBeginOffset, P, I, E)) {
2359 IntTy = Type::getIntNTy(NewAI.getContext(),
2360 TD.getTypeSizeInBits(NewAI.getAllocatedType()));
2362 bool CanSROA = true;
2363 for (; I != E; ++I) {
2365 continue; // Skip dead uses.
2366 BeginOffset = I->BeginOffset;
2367 EndOffset = I->EndOffset;
2368 IsSplit = I->isSplit();
2369 OldUse = I->getUse();
2370 OldPtr = cast<Instruction>(OldUse->get());
2371 NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str();
2372 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2388 // Every instruction which can end up as a user must have a rewrite rule.
2389 bool visitInstruction(Instruction &I) {
2390 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2391 llvm_unreachable("No rewrite rule for this instruction!");
2394 Twine getName(const Twine &Suffix) {
2395 return NamePrefix + Suffix;
2398 Value *getAdjustedAllocaPtr(IRBuilder<> &IRB, Type *PointerTy) {
2399 assert(BeginOffset >= NewAllocaBeginOffset);
2400 APInt Offset(TD.getPointerSizeInBits(), BeginOffset - NewAllocaBeginOffset);
2401 return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName(""));
2404 /// \brief Compute suitable alignment to access an offset into the new alloca.
2405 unsigned getOffsetAlign(uint64_t Offset) {
2406 unsigned NewAIAlign = NewAI.getAlignment();
2408 NewAIAlign = TD.getABITypeAlignment(NewAI.getAllocatedType());
2409 return MinAlign(NewAIAlign, Offset);
2412 /// \brief Compute suitable alignment to access this partition of the new
2414 unsigned getPartitionAlign() {
2415 return getOffsetAlign(BeginOffset - NewAllocaBeginOffset);
2418 /// \brief Compute suitable alignment to access a type at an offset of the
2421 /// \returns zero if the type's ABI alignment is a suitable alignment,
2422 /// otherwise returns the maximal suitable alignment.
2423 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
2424 unsigned Align = getOffsetAlign(Offset);
2425 return Align == TD.getABITypeAlignment(Ty) ? 0 : Align;
2428 /// \brief Compute suitable alignment to access a type at the beginning of
2429 /// this partition of the new alloca.
2431 /// See \c getOffsetTypeAlign for details; this routine delegates to it.
2432 unsigned getPartitionTypeAlign(Type *Ty) {
2433 return getOffsetTypeAlign(Ty, BeginOffset - NewAllocaBeginOffset);
2436 unsigned getIndex(uint64_t Offset) {
2437 assert(VecTy && "Can only call getIndex when rewriting a vector");
2438 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2439 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2440 uint32_t Index = RelOffset / ElementSize;
2441 assert(Index * ElementSize == RelOffset);
2445 void deleteIfTriviallyDead(Value *V) {
2446 Instruction *I = cast<Instruction>(V);
2447 if (isInstructionTriviallyDead(I))
2448 Pass.DeadInsts.insert(I);
2451 Value *rewriteVectorizedLoadInst(IRBuilder<> &IRB) {
2452 unsigned BeginIndex = getIndex(BeginOffset);
2453 unsigned EndIndex = getIndex(EndOffset);
2454 assert(EndIndex > BeginIndex && "Empty vector!");
2456 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2458 return extractVector(IRB, V, BeginIndex, EndIndex, getName(".vec"));
2461 Value *rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) {
2462 assert(IntTy && "We cannot insert an integer to the alloca");
2463 assert(!LI.isVolatile());
2464 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2466 V = convertValue(TD, IRB, V, IntTy);
2467 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2468 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2469 if (Offset > 0 || EndOffset < NewAllocaEndOffset)
2470 V = extractInteger(TD, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2471 getName(".extract"));
2475 bool visitLoadInst(LoadInst &LI) {
2476 DEBUG(dbgs() << " original: " << LI << "\n");
2477 Value *OldOp = LI.getOperand(0);
2478 assert(OldOp == OldPtr);
2480 uint64_t Size = EndOffset - BeginOffset;
2482 IRBuilder<> IRB(&LI);
2483 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), Size * 8)
2485 bool IsPtrAdjusted = false;
2488 V = rewriteVectorizedLoadInst(IRB);
2489 } else if (IntTy && LI.getType()->isIntegerTy()) {
2490 V = rewriteIntegerLoad(IRB, LI);
2491 } else if (BeginOffset == NewAllocaBeginOffset &&
2492 canConvertValue(TD, NewAllocaTy, LI.getType())) {
2493 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2494 LI.isVolatile(), getName(".load"));
2496 Type *LTy = TargetTy->getPointerTo();
2497 V = IRB.CreateAlignedLoad(getAdjustedAllocaPtr(IRB, LTy),
2498 getPartitionTypeAlign(TargetTy),
2499 LI.isVolatile(), getName(".load"));
2500 IsPtrAdjusted = true;
2502 V = convertValue(TD, IRB, V, TargetTy);
2505 assert(!LI.isVolatile());
2506 assert(LI.getType()->isIntegerTy() &&
2507 "Only integer type loads and stores are split");
2508 assert(Size < TD.getTypeStoreSize(LI.getType()) &&
2509 "Split load isn't smaller than original load");
2510 assert(LI.getType()->getIntegerBitWidth() ==
2511 TD.getTypeStoreSizeInBits(LI.getType()) &&
2512 "Non-byte-multiple bit width");
2513 // Move the insertion point just past the load so that we can refer to it.
2514 IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI)));
2515 // Create a placeholder value with the same type as LI to use as the
2516 // basis for the new value. This allows us to replace the uses of LI with
2517 // the computed value, and then replace the placeholder with LI, leaving
2518 // LI only used for this computation.
2520 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2521 V = insertInteger(TD, IRB, Placeholder, V, BeginOffset,
2522 getName(".insert"));
2523 LI.replaceAllUsesWith(V);
2524 Placeholder->replaceAllUsesWith(&LI);
2527 LI.replaceAllUsesWith(V);
2530 Pass.DeadInsts.insert(&LI);
2531 deleteIfTriviallyDead(OldOp);
2532 DEBUG(dbgs() << " to: " << *V << "\n");
2533 return !LI.isVolatile() && !IsPtrAdjusted;
2536 bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, Value *V,
2537 StoreInst &SI, Value *OldOp) {
2538 unsigned BeginIndex = getIndex(BeginOffset);
2539 unsigned EndIndex = getIndex(EndOffset);
2540 assert(EndIndex > BeginIndex && "Empty vector!");
2541 unsigned NumElements = EndIndex - BeginIndex;
2542 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2544 = (NumElements == 1) ? ElementTy
2545 : VectorType::get(ElementTy, NumElements);
2546 if (V->getType() != PartitionTy)
2547 V = convertValue(TD, IRB, V, PartitionTy);
2549 // Mix in the existing elements.
2550 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2552 V = insertVector(IRB, Old, V, BeginIndex, getName(".vec"));
2554 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2555 Pass.DeadInsts.insert(&SI);
2558 DEBUG(dbgs() << " to: " << *Store << "\n");
2562 bool rewriteIntegerStore(IRBuilder<> &IRB, Value *V, StoreInst &SI) {
2563 assert(IntTy && "We cannot extract an integer from the alloca");
2564 assert(!SI.isVolatile());
2565 if (TD.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2566 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2567 getName(".oldload"));
2568 Old = convertValue(TD, IRB, Old, IntTy);
2569 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2570 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2571 V = insertInteger(TD, IRB, Old, SI.getValueOperand(), Offset,
2572 getName(".insert"));
2574 V = convertValue(TD, IRB, V, NewAllocaTy);
2575 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2576 Pass.DeadInsts.insert(&SI);
2578 DEBUG(dbgs() << " to: " << *Store << "\n");
2582 bool visitStoreInst(StoreInst &SI) {
2583 DEBUG(dbgs() << " original: " << SI << "\n");
2584 Value *OldOp = SI.getOperand(1);
2585 assert(OldOp == OldPtr);
2586 IRBuilder<> IRB(&SI);
2588 Value *V = SI.getValueOperand();
2590 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2591 // alloca that should be re-examined after promoting this alloca.
2592 if (V->getType()->isPointerTy())
2593 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2594 Pass.PostPromotionWorklist.insert(AI);
2596 uint64_t Size = EndOffset - BeginOffset;
2597 if (Size < TD.getTypeStoreSize(V->getType())) {
2598 assert(!SI.isVolatile());
2599 assert(IsSplit && "A seemingly split store isn't splittable");
2600 assert(V->getType()->isIntegerTy() &&
2601 "Only integer type loads and stores are split");
2602 assert(V->getType()->getIntegerBitWidth() ==
2603 TD.getTypeStoreSizeInBits(V->getType()) &&
2604 "Non-byte-multiple bit width");
2605 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8);
2606 V = extractInteger(TD, IRB, V, NarrowTy, BeginOffset,
2607 getName(".extract"));
2611 return rewriteVectorizedStoreInst(IRB, V, SI, OldOp);
2612 if (IntTy && V->getType()->isIntegerTy())
2613 return rewriteIntegerStore(IRB, V, SI);
2616 if (BeginOffset == NewAllocaBeginOffset &&
2617 canConvertValue(TD, V->getType(), NewAllocaTy)) {
2618 V = convertValue(TD, IRB, V, NewAllocaTy);
2619 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2622 Value *NewPtr = getAdjustedAllocaPtr(IRB, V->getType()->getPointerTo());
2623 NewSI = IRB.CreateAlignedStore(V, NewPtr,
2624 getPartitionTypeAlign(V->getType()),
2628 Pass.DeadInsts.insert(&SI);
2629 deleteIfTriviallyDead(OldOp);
2631 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2632 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2635 /// \brief Compute an integer value from splatting an i8 across the given
2636 /// number of bytes.
2638 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2639 /// call this routine.
2640 /// FIXME: Heed the advice above.
2642 /// \param V The i8 value to splat.
2643 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2644 Value *getIntegerSplat(IRBuilder<> &IRB, Value *V, unsigned Size) {
2645 assert(Size > 0 && "Expected a positive number of bytes.");
2646 IntegerType *VTy = cast<IntegerType>(V->getType());
2647 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2651 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2652 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, getName(".zext")),
2653 ConstantExpr::getUDiv(
2654 Constant::getAllOnesValue(SplatIntTy),
2655 ConstantExpr::getZExt(
2656 Constant::getAllOnesValue(V->getType()),
2658 getName(".isplat"));
2662 /// \brief Compute a vector splat for a given element value.
2663 Value *getVectorSplat(IRBuilder<> &IRB, Value *V, unsigned NumElements) {
2664 V = IRB.CreateVectorSplat(NumElements, V, NamePrefix);
2665 DEBUG(dbgs() << " splat: " << *V << "\n");
2669 bool visitMemSetInst(MemSetInst &II) {
2670 DEBUG(dbgs() << " original: " << II << "\n");
2671 IRBuilder<> IRB(&II);
2672 assert(II.getRawDest() == OldPtr);
2674 // If the memset has a variable size, it cannot be split, just adjust the
2675 // pointer to the new alloca.
2676 if (!isa<Constant>(II.getLength())) {
2677 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2678 Type *CstTy = II.getAlignmentCst()->getType();
2679 II.setAlignment(ConstantInt::get(CstTy, getPartitionAlign()));
2681 deleteIfTriviallyDead(OldPtr);
2685 // Record this instruction for deletion.
2686 Pass.DeadInsts.insert(&II);
2688 Type *AllocaTy = NewAI.getAllocatedType();
2689 Type *ScalarTy = AllocaTy->getScalarType();
2691 // If this doesn't map cleanly onto the alloca type, and that type isn't
2692 // a single value type, just emit a memset.
2693 if (!VecTy && !IntTy &&
2694 (BeginOffset != NewAllocaBeginOffset ||
2695 EndOffset != NewAllocaEndOffset ||
2696 !AllocaTy->isSingleValueType() ||
2697 !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)) ||
2698 TD.getTypeSizeInBits(ScalarTy)%8 != 0)) {
2699 Type *SizeTy = II.getLength()->getType();
2700 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2702 = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB,
2703 II.getRawDest()->getType()),
2704 II.getValue(), Size, getPartitionAlign(),
2707 DEBUG(dbgs() << " to: " << *New << "\n");
2711 // If we can represent this as a simple value, we have to build the actual
2712 // value to store, which requires expanding the byte present in memset to
2713 // a sensible representation for the alloca type. This is essentially
2714 // splatting the byte to a sufficiently wide integer, splatting it across
2715 // any desired vector width, and bitcasting to the final type.
2719 // If this is a memset of a vectorized alloca, insert it.
2720 assert(ElementTy == ScalarTy);
2722 unsigned BeginIndex = getIndex(BeginOffset);
2723 unsigned EndIndex = getIndex(EndOffset);
2724 assert(EndIndex > BeginIndex && "Empty vector!");
2725 unsigned NumElements = EndIndex - BeginIndex;
2726 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2728 Value *Splat = getIntegerSplat(IRB, II.getValue(),
2729 TD.getTypeSizeInBits(ElementTy)/8);
2730 Splat = convertValue(TD, IRB, Splat, ElementTy);
2731 if (NumElements > 1)
2732 Splat = getVectorSplat(IRB, Splat, NumElements);
2734 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2735 getName(".oldload"));
2736 V = insertVector(IRB, Old, Splat, BeginIndex, getName(".vec"));
2738 // If this is a memset on an alloca where we can widen stores, insert the
2740 assert(!II.isVolatile());
2742 uint64_t Size = EndOffset - BeginOffset;
2743 V = getIntegerSplat(IRB, II.getValue(), Size);
2745 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2746 EndOffset != NewAllocaBeginOffset)) {
2747 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2748 getName(".oldload"));
2749 Old = convertValue(TD, IRB, Old, IntTy);
2750 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2751 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2752 V = insertInteger(TD, IRB, Old, V, Offset, getName(".insert"));
2754 assert(V->getType() == IntTy &&
2755 "Wrong type for an alloca wide integer!");
2757 V = convertValue(TD, IRB, V, AllocaTy);
2759 // Established these invariants above.
2760 assert(BeginOffset == NewAllocaBeginOffset);
2761 assert(EndOffset == NewAllocaEndOffset);
2763 V = getIntegerSplat(IRB, II.getValue(),
2764 TD.getTypeSizeInBits(ScalarTy)/8);
2765 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2766 V = getVectorSplat(IRB, V, AllocaVecTy->getNumElements());
2768 V = convertValue(TD, IRB, V, AllocaTy);
2771 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2774 DEBUG(dbgs() << " to: " << *New << "\n");
2775 return !II.isVolatile();
2778 bool visitMemTransferInst(MemTransferInst &II) {
2779 // Rewriting of memory transfer instructions can be a bit tricky. We break
2780 // them into two categories: split intrinsics and unsplit intrinsics.
2782 DEBUG(dbgs() << " original: " << II << "\n");
2783 IRBuilder<> IRB(&II);
2785 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2786 bool IsDest = II.getRawDest() == OldPtr;
2788 const AllocaPartitioning::MemTransferOffsets &MTO
2789 = P.getMemTransferOffsets(II);
2791 // Compute the relative offset within the transfer.
2792 unsigned IntPtrWidth = TD.getPointerSizeInBits();
2793 APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
2794 : MTO.SourceBegin));
2796 unsigned Align = II.getAlignment();
2798 Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2799 MinAlign(II.getAlignment(), getPartitionAlign()));
2801 // For unsplit intrinsics, we simply modify the source and destination
2802 // pointers in place. This isn't just an optimization, it is a matter of
2803 // correctness. With unsplit intrinsics we may be dealing with transfers
2804 // within a single alloca before SROA ran, or with transfers that have
2805 // a variable length. We may also be dealing with memmove instead of
2806 // memcpy, and so simply updating the pointers is the necessary for us to
2807 // update both source and dest of a single call.
2808 if (!MTO.IsSplittable) {
2809 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2811 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2813 II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType()));
2815 Type *CstTy = II.getAlignmentCst()->getType();
2816 II.setAlignment(ConstantInt::get(CstTy, Align));
2818 DEBUG(dbgs() << " to: " << II << "\n");
2819 deleteIfTriviallyDead(OldOp);
2822 // For split transfer intrinsics we have an incredibly useful assurance:
2823 // the source and destination do not reside within the same alloca, and at
2824 // least one of them does not escape. This means that we can replace
2825 // memmove with memcpy, and we don't need to worry about all manner of
2826 // downsides to splitting and transforming the operations.
2828 // If this doesn't map cleanly onto the alloca type, and that type isn't
2829 // a single value type, just emit a memcpy.
2831 = !VecTy && !IntTy && (BeginOffset != NewAllocaBeginOffset ||
2832 EndOffset != NewAllocaEndOffset ||
2833 !NewAI.getAllocatedType()->isSingleValueType());
2835 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2836 // size hasn't been shrunk based on analysis of the viable range, this is
2838 if (EmitMemCpy && &OldAI == &NewAI) {
2839 uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin;
2840 uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd;
2841 // Ensure the start lines up.
2842 assert(BeginOffset == OrigBegin);
2845 // Rewrite the size as needed.
2846 if (EndOffset != OrigEnd)
2847 II.setLength(ConstantInt::get(II.getLength()->getType(),
2848 EndOffset - BeginOffset));
2851 // Record this instruction for deletion.
2852 Pass.DeadInsts.insert(&II);
2854 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2855 // alloca that should be re-examined after rewriting this instruction.
2856 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2858 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2859 Pass.Worklist.insert(AI);
2862 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2863 : II.getRawDest()->getType();
2865 // Compute the other pointer, folding as much as possible to produce
2866 // a single, simple GEP in most cases.
2867 OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
2868 getName("." + OtherPtr->getName()));
2871 = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType()
2872 : II.getRawSource()->getType());
2873 Type *SizeTy = II.getLength()->getType();
2874 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2876 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2877 IsDest ? OtherPtr : OurPtr,
2878 Size, Align, II.isVolatile());
2880 DEBUG(dbgs() << " to: " << *New << "\n");
2884 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2885 // is equivalent to 1, but that isn't true if we end up rewriting this as
2890 bool IsWholeAlloca = BeginOffset == NewAllocaBeginOffset &&
2891 EndOffset == NewAllocaEndOffset;
2892 uint64_t Size = EndOffset - BeginOffset;
2893 unsigned BeginIndex = VecTy ? getIndex(BeginOffset) : 0;
2894 unsigned EndIndex = VecTy ? getIndex(EndOffset) : 0;
2895 unsigned NumElements = EndIndex - BeginIndex;
2896 IntegerType *SubIntTy
2897 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0;
2899 Type *OtherPtrTy = NewAI.getType();
2900 if (VecTy && !IsWholeAlloca) {
2901 if (NumElements == 1)
2902 OtherPtrTy = VecTy->getElementType();
2904 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2906 OtherPtrTy = OtherPtrTy->getPointerTo();
2907 } else if (IntTy && !IsWholeAlloca) {
2908 OtherPtrTy = SubIntTy->getPointerTo();
2911 Value *SrcPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
2912 getName("." + OtherPtr->getName()));
2913 Value *DstPtr = &NewAI;
2915 std::swap(SrcPtr, DstPtr);
2918 if (VecTy && !IsWholeAlloca && !IsDest) {
2919 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2921 Src = extractVector(IRB, Src, BeginIndex, EndIndex, getName(".vec"));
2922 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2923 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2925 Src = convertValue(TD, IRB, Src, IntTy);
2926 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2927 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2928 Src = extractInteger(TD, IRB, Src, SubIntTy, Offset, getName(".extract"));
2930 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2931 getName(".copyload"));
2934 if (VecTy && !IsWholeAlloca && IsDest) {
2935 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2936 getName(".oldload"));
2937 Src = insertVector(IRB, Old, Src, BeginIndex, getName(".vec"));
2938 } else if (IntTy && !IsWholeAlloca && IsDest) {
2939 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2940 getName(".oldload"));
2941 Old = convertValue(TD, IRB, Old, IntTy);
2942 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2943 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2944 Src = insertInteger(TD, IRB, Old, Src, Offset, getName(".insert"));
2945 Src = convertValue(TD, IRB, Src, NewAllocaTy);
2948 StoreInst *Store = cast<StoreInst>(
2949 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2951 DEBUG(dbgs() << " to: " << *Store << "\n");
2952 return !II.isVolatile();
2955 bool visitIntrinsicInst(IntrinsicInst &II) {
2956 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2957 II.getIntrinsicID() == Intrinsic::lifetime_end);
2958 DEBUG(dbgs() << " original: " << II << "\n");
2959 IRBuilder<> IRB(&II);
2960 assert(II.getArgOperand(1) == OldPtr);
2962 // Record this instruction for deletion.
2963 Pass.DeadInsts.insert(&II);
2966 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2967 EndOffset - BeginOffset);
2968 Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType());
2970 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2971 New = IRB.CreateLifetimeStart(Ptr, Size);
2973 New = IRB.CreateLifetimeEnd(Ptr, Size);
2976 DEBUG(dbgs() << " to: " << *New << "\n");
2980 bool visitPHINode(PHINode &PN) {
2981 DEBUG(dbgs() << " original: " << PN << "\n");
2983 // We would like to compute a new pointer in only one place, but have it be
2984 // as local as possible to the PHI. To do that, we re-use the location of
2985 // the old pointer, which necessarily must be in the right position to
2986 // dominate the PHI.
2987 IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr));
2989 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
2990 // Replace the operands which were using the old pointer.
2991 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2993 DEBUG(dbgs() << " to: " << PN << "\n");
2994 deleteIfTriviallyDead(OldPtr);
2998 bool visitSelectInst(SelectInst &SI) {
2999 DEBUG(dbgs() << " original: " << SI << "\n");
3000 IRBuilder<> IRB(&SI);
3002 // Find the operand we need to rewrite here.
3003 bool IsTrueVal = SI.getTrueValue() == OldPtr;
3005 assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!");
3007 assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!");
3009 Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
3010 SI.setOperand(IsTrueVal ? 1 : 2, NewPtr);
3011 DEBUG(dbgs() << " to: " << SI << "\n");
3012 deleteIfTriviallyDead(OldPtr);
3020 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
3022 /// This pass aggressively rewrites all aggregate loads and stores on
3023 /// a particular pointer (or any pointer derived from it which we can identify)
3024 /// with scalar loads and stores.
3025 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3026 // Befriend the base class so it can delegate to private visit methods.
3027 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
3029 const DataLayout &TD;
3031 /// Queue of pointer uses to analyze and potentially rewrite.
3032 SmallVector<Use *, 8> Queue;
3034 /// Set to prevent us from cycling with phi nodes and loops.
3035 SmallPtrSet<User *, 8> Visited;
3037 /// The current pointer use being rewritten. This is used to dig up the used
3038 /// value (as opposed to the user).
3042 AggLoadStoreRewriter(const DataLayout &TD) : TD(TD) {}
3044 /// Rewrite loads and stores through a pointer and all pointers derived from
3046 bool rewrite(Instruction &I) {
3047 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
3049 bool Changed = false;
3050 while (!Queue.empty()) {
3051 U = Queue.pop_back_val();
3052 Changed |= visit(cast<Instruction>(U->getUser()));
3058 /// Enqueue all the users of the given instruction for further processing.
3059 /// This uses a set to de-duplicate users.
3060 void enqueueUsers(Instruction &I) {
3061 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
3063 if (Visited.insert(*UI))
3064 Queue.push_back(&UI.getUse());
3067 // Conservative default is to not rewrite anything.
3068 bool visitInstruction(Instruction &I) { return false; }
3070 /// \brief Generic recursive split emission class.
3071 template <typename Derived>
3074 /// The builder used to form new instructions.
3076 /// The indices which to be used with insert- or extractvalue to select the
3077 /// appropriate value within the aggregate.
3078 SmallVector<unsigned, 4> Indices;
3079 /// The indices to a GEP instruction which will move Ptr to the correct slot
3080 /// within the aggregate.
3081 SmallVector<Value *, 4> GEPIndices;
3082 /// The base pointer of the original op, used as a base for GEPing the
3083 /// split operations.
3086 /// Initialize the splitter with an insertion point, Ptr and start with a
3087 /// single zero GEP index.
3088 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3089 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3092 /// \brief Generic recursive split emission routine.
3094 /// This method recursively splits an aggregate op (load or store) into
3095 /// scalar or vector ops. It splits recursively until it hits a single value
3096 /// and emits that single value operation via the template argument.
3098 /// The logic of this routine relies on GEPs and insertvalue and
3099 /// extractvalue all operating with the same fundamental index list, merely
3100 /// formatted differently (GEPs need actual values).
3102 /// \param Ty The type being split recursively into smaller ops.
3103 /// \param Agg The aggregate value being built up or stored, depending on
3104 /// whether this is splitting a load or a store respectively.
3105 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3106 if (Ty->isSingleValueType())
3107 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3109 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3110 unsigned OldSize = Indices.size();
3112 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3114 assert(Indices.size() == OldSize && "Did not return to the old size");
3115 Indices.push_back(Idx);
3116 GEPIndices.push_back(IRB.getInt32(Idx));
3117 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3118 GEPIndices.pop_back();
3124 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3125 unsigned OldSize = Indices.size();
3127 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3129 assert(Indices.size() == OldSize && "Did not return to the old size");
3130 Indices.push_back(Idx);
3131 GEPIndices.push_back(IRB.getInt32(Idx));
3132 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3133 GEPIndices.pop_back();
3139 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3143 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3144 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3145 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3147 /// Emit a leaf load of a single value. This is called at the leaves of the
3148 /// recursive emission to actually load values.
3149 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3150 assert(Ty->isSingleValueType());
3151 // Load the single value and insert it using the indices.
3152 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
3153 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
3154 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3155 DEBUG(dbgs() << " to: " << *Load << "\n");
3159 bool visitLoadInst(LoadInst &LI) {
3160 assert(LI.getPointerOperand() == *U);
3161 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3164 // We have an aggregate being loaded, split it apart.
3165 DEBUG(dbgs() << " original: " << LI << "\n");
3166 LoadOpSplitter Splitter(&LI, *U);
3167 Value *V = UndefValue::get(LI.getType());
3168 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3169 LI.replaceAllUsesWith(V);
3170 LI.eraseFromParent();
3174 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3175 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3176 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3178 /// Emit a leaf store of a single value. This is called at the leaves of the
3179 /// recursive emission to actually produce stores.
3180 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3181 assert(Ty->isSingleValueType());
3182 // Extract the single value and store it using the indices.
3183 Value *Store = IRB.CreateStore(
3184 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
3185 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
3187 DEBUG(dbgs() << " to: " << *Store << "\n");
3191 bool visitStoreInst(StoreInst &SI) {
3192 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3194 Value *V = SI.getValueOperand();
3195 if (V->getType()->isSingleValueType())
3198 // We have an aggregate being stored, split it apart.
3199 DEBUG(dbgs() << " original: " << SI << "\n");
3200 StoreOpSplitter Splitter(&SI, *U);
3201 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3202 SI.eraseFromParent();
3206 bool visitBitCastInst(BitCastInst &BC) {
3211 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3216 bool visitPHINode(PHINode &PN) {
3221 bool visitSelectInst(SelectInst &SI) {
3228 /// \brief Strip aggregate type wrapping.
3230 /// This removes no-op aggregate types wrapping an underlying type. It will
3231 /// strip as many layers of types as it can without changing either the type
3232 /// size or the allocated size.
3233 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3234 if (Ty->isSingleValueType())
3237 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3238 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3241 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3242 InnerTy = ArrTy->getElementType();
3243 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3244 const StructLayout *SL = DL.getStructLayout(STy);
3245 unsigned Index = SL->getElementContainingOffset(0);
3246 InnerTy = STy->getElementType(Index);
3251 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3252 TypeSize > DL.getTypeSizeInBits(InnerTy))
3255 return stripAggregateTypeWrapping(DL, InnerTy);
3258 /// \brief Try to find a partition of the aggregate type passed in for a given
3259 /// offset and size.
3261 /// This recurses through the aggregate type and tries to compute a subtype
3262 /// based on the offset and size. When the offset and size span a sub-section
3263 /// of an array, it will even compute a new array type for that sub-section,
3264 /// and the same for structs.
3266 /// Note that this routine is very strict and tries to find a partition of the
3267 /// type which produces the *exact* right offset and size. It is not forgiving
3268 /// when the size or offset cause either end of type-based partition to be off.
3269 /// Also, this is a best-effort routine. It is reasonable to give up and not
3270 /// return a type if necessary.
3271 static Type *getTypePartition(const DataLayout &TD, Type *Ty,
3272 uint64_t Offset, uint64_t Size) {
3273 if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size)
3274 return stripAggregateTypeWrapping(TD, Ty);
3275 if (Offset > TD.getTypeAllocSize(Ty) ||
3276 (TD.getTypeAllocSize(Ty) - Offset) < Size)
3279 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3280 // We can't partition pointers...
3281 if (SeqTy->isPointerTy())
3284 Type *ElementTy = SeqTy->getElementType();
3285 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3286 uint64_t NumSkippedElements = Offset / ElementSize;
3287 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy))
3288 if (NumSkippedElements >= ArrTy->getNumElements())
3290 if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy))
3291 if (NumSkippedElements >= VecTy->getNumElements())
3293 Offset -= NumSkippedElements * ElementSize;
3295 // First check if we need to recurse.
3296 if (Offset > 0 || Size < ElementSize) {
3297 // Bail if the partition ends in a different array element.
3298 if ((Offset + Size) > ElementSize)
3300 // Recurse through the element type trying to peel off offset bytes.
3301 return getTypePartition(TD, ElementTy, Offset, Size);
3303 assert(Offset == 0);
3305 if (Size == ElementSize)
3306 return stripAggregateTypeWrapping(TD, ElementTy);
3307 assert(Size > ElementSize);
3308 uint64_t NumElements = Size / ElementSize;
3309 if (NumElements * ElementSize != Size)
3311 return ArrayType::get(ElementTy, NumElements);
3314 StructType *STy = dyn_cast<StructType>(Ty);
3318 const StructLayout *SL = TD.getStructLayout(STy);
3319 if (Offset >= SL->getSizeInBytes())
3321 uint64_t EndOffset = Offset + Size;
3322 if (EndOffset > SL->getSizeInBytes())
3325 unsigned Index = SL->getElementContainingOffset(Offset);
3326 Offset -= SL->getElementOffset(Index);
3328 Type *ElementTy = STy->getElementType(Index);
3329 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3330 if (Offset >= ElementSize)
3331 return 0; // The offset points into alignment padding.
3333 // See if any partition must be contained by the element.
3334 if (Offset > 0 || Size < ElementSize) {
3335 if ((Offset + Size) > ElementSize)
3337 return getTypePartition(TD, ElementTy, Offset, Size);
3339 assert(Offset == 0);
3341 if (Size == ElementSize)
3342 return stripAggregateTypeWrapping(TD, ElementTy);
3344 StructType::element_iterator EI = STy->element_begin() + Index,
3345 EE = STy->element_end();
3346 if (EndOffset < SL->getSizeInBytes()) {
3347 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3348 if (Index == EndIndex)
3349 return 0; // Within a single element and its padding.
3351 // Don't try to form "natural" types if the elements don't line up with the
3353 // FIXME: We could potentially recurse down through the last element in the
3354 // sub-struct to find a natural end point.
3355 if (SL->getElementOffset(EndIndex) != EndOffset)
3358 assert(Index < EndIndex);
3359 EE = STy->element_begin() + EndIndex;
3362 // Try to build up a sub-structure.
3363 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
3365 const StructLayout *SubSL = TD.getStructLayout(SubTy);
3366 if (Size != SubSL->getSizeInBytes())
3367 return 0; // The sub-struct doesn't have quite the size needed.
3372 /// \brief Rewrite an alloca partition's users.
3374 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3375 /// to rewrite uses of an alloca partition to be conducive for SSA value
3376 /// promotion. If the partition needs a new, more refined alloca, this will
3377 /// build that new alloca, preserving as much type information as possible, and
3378 /// rewrite the uses of the old alloca to point at the new one and have the
3379 /// appropriate new offsets. It also evaluates how successful the rewrite was
3380 /// at enabling promotion and if it was successful queues the alloca to be
3382 bool SROA::rewriteAllocaPartition(AllocaInst &AI,
3383 AllocaPartitioning &P,
3384 AllocaPartitioning::iterator PI) {
3385 uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset;
3386 bool IsLive = false;
3387 for (AllocaPartitioning::use_iterator UI = P.use_begin(PI),
3389 UI != UE && !IsLive; ++UI)
3393 return false; // No live uses left of this partition.
3395 DEBUG(dbgs() << "Speculating PHIs and selects in partition "
3396 << "[" << PI->BeginOffset << "," << PI->EndOffset << ")\n");
3398 PHIOrSelectSpeculator Speculator(*TD, P, *this);
3399 DEBUG(dbgs() << " speculating ");
3400 DEBUG(P.print(dbgs(), PI, ""));
3401 Speculator.visitUsers(PI);
3403 // Try to compute a friendly type for this partition of the alloca. This
3404 // won't always succeed, in which case we fall back to a legal integer type
3405 // or an i8 array of an appropriate size.
3407 if (Type *PartitionTy = P.getCommonType(PI))
3408 if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize)
3409 AllocaTy = PartitionTy;
3411 if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(),
3412 PI->BeginOffset, AllocaSize))
3413 AllocaTy = PartitionTy;
3415 (AllocaTy->isArrayTy() &&
3416 AllocaTy->getArrayElementType()->isIntegerTy())) &&
3417 TD->isLegalInteger(AllocaSize * 8))
3418 AllocaTy = Type::getIntNTy(*C, AllocaSize * 8);
3420 AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize);
3421 assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize);
3423 // Check for the case where we're going to rewrite to a new alloca of the
3424 // exact same type as the original, and with the same access offsets. In that
3425 // case, re-use the existing alloca, but still run through the rewriter to
3426 // perform phi and select speculation.
3428 if (AllocaTy == AI.getAllocatedType()) {
3429 assert(PI->BeginOffset == 0 &&
3430 "Non-zero begin offset but same alloca type");
3431 assert(PI == P.begin() && "Begin offset is zero on later partition");
3434 unsigned Alignment = AI.getAlignment();
3436 // The minimum alignment which users can rely on when the explicit
3437 // alignment is omitted or zero is that required by the ABI for this
3439 Alignment = TD->getABITypeAlignment(AI.getAllocatedType());
3441 Alignment = MinAlign(Alignment, PI->BeginOffset);
3442 // If we will get at least this much alignment from the type alone, leave
3443 // the alloca's alignment unconstrained.
3444 if (Alignment <= TD->getABITypeAlignment(AllocaTy))
3446 NewAI = new AllocaInst(AllocaTy, 0, Alignment,
3447 AI.getName() + ".sroa." + Twine(PI - P.begin()),
3452 DEBUG(dbgs() << "Rewriting alloca partition "
3453 << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: "
3456 // Track the high watermark of the post-promotion worklist. We will reset it
3457 // to this point if the alloca is not in fact scheduled for promotion.
3458 unsigned PPWOldSize = PostPromotionWorklist.size();
3460 AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI,
3461 PI->BeginOffset, PI->EndOffset);
3462 DEBUG(dbgs() << " rewriting ");
3463 DEBUG(P.print(dbgs(), PI, ""));
3464 bool Promotable = Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI));
3466 DEBUG(dbgs() << " and queuing for promotion\n");
3467 PromotableAllocas.push_back(NewAI);
3468 } else if (NewAI != &AI) {
3469 // If we can't promote the alloca, iterate on it to check for new
3470 // refinements exposed by splitting the current alloca. Don't iterate on an
3471 // alloca which didn't actually change and didn't get promoted.
3472 Worklist.insert(NewAI);
3475 // Drop any post-promotion work items if promotion didn't happen.
3477 while (PostPromotionWorklist.size() > PPWOldSize)
3478 PostPromotionWorklist.pop_back();
3483 /// \brief Walks the partitioning of an alloca rewriting uses of each partition.
3484 bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
3485 bool Changed = false;
3486 for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE;
3488 Changed |= rewriteAllocaPartition(AI, P, PI);
3493 /// \brief Analyze an alloca for SROA.
3495 /// This analyzes the alloca to ensure we can reason about it, builds
3496 /// a partitioning of the alloca, and then hands it off to be split and
3497 /// rewritten as needed.
3498 bool SROA::runOnAlloca(AllocaInst &AI) {
3499 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3500 ++NumAllocasAnalyzed;
3502 // Special case dead allocas, as they're trivial.
3503 if (AI.use_empty()) {
3504 AI.eraseFromParent();
3508 // Skip alloca forms that this analysis can't handle.
3509 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3510 TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
3513 bool Changed = false;
3515 // First, split any FCA loads and stores touching this alloca to promote
3516 // better splitting and promotion opportunities.
3517 AggLoadStoreRewriter AggRewriter(*TD);
3518 Changed |= AggRewriter.rewrite(AI);
3520 // Build the partition set using a recursive instruction-visiting builder.
3521 AllocaPartitioning P(*TD, AI);
3522 DEBUG(P.print(dbgs()));
3526 // Delete all the dead users of this alloca before splitting and rewriting it.
3527 for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
3528 DE = P.dead_user_end();
3531 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3532 DeadInsts.insert(*DI);
3534 for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(),
3535 DE = P.dead_op_end();
3538 // Clobber the use with an undef value.
3539 **DO = UndefValue::get(OldV->getType());
3540 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3541 if (isInstructionTriviallyDead(OldI)) {
3543 DeadInsts.insert(OldI);
3547 // No partitions to split. Leave the dead alloca for a later pass to clean up.
3548 if (P.begin() == P.end())
3551 return splitAlloca(AI, P) || Changed;
3554 /// \brief Delete the dead instructions accumulated in this run.
3556 /// Recursively deletes the dead instructions we've accumulated. This is done
3557 /// at the very end to maximize locality of the recursive delete and to
3558 /// minimize the problems of invalidated instruction pointers as such pointers
3559 /// are used heavily in the intermediate stages of the algorithm.
3561 /// We also record the alloca instructions deleted here so that they aren't
3562 /// subsequently handed to mem2reg to promote.
3563 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3564 while (!DeadInsts.empty()) {
3565 Instruction *I = DeadInsts.pop_back_val();
3566 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3568 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3570 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3571 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3572 // Zero out the operand and see if it becomes trivially dead.
3574 if (isInstructionTriviallyDead(U))
3575 DeadInsts.insert(U);
3578 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3579 DeletedAllocas.insert(AI);
3582 I->eraseFromParent();
3586 /// \brief Promote the allocas, using the best available technique.
3588 /// This attempts to promote whatever allocas have been identified as viable in
3589 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3590 /// If there is a domtree available, we attempt to promote using the full power
3591 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3592 /// based on the SSAUpdater utilities. This function returns whether any
3593 /// promotion occurred.
3594 bool SROA::promoteAllocas(Function &F) {
3595 if (PromotableAllocas.empty())
3598 NumPromoted += PromotableAllocas.size();
3600 if (DT && !ForceSSAUpdater) {
3601 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3602 PromoteMemToReg(PromotableAllocas, *DT);
3603 PromotableAllocas.clear();
3607 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3609 DIBuilder DIB(*F.getParent());
3610 SmallVector<Instruction*, 64> Insts;
3612 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3613 AllocaInst *AI = PromotableAllocas[Idx];
3614 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
3616 Instruction *I = cast<Instruction>(*UI++);
3617 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3618 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3619 // leading to them) here. Eventually it should use them to optimize the
3620 // scalar values produced.
3621 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
3622 assert(onlyUsedByLifetimeMarkers(I) &&
3623 "Found a bitcast used outside of a lifetime marker.");
3624 while (!I->use_empty())
3625 cast<Instruction>(*I->use_begin())->eraseFromParent();
3626 I->eraseFromParent();
3629 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3630 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3631 II->getIntrinsicID() == Intrinsic::lifetime_end);
3632 II->eraseFromParent();
3638 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3642 PromotableAllocas.clear();
3647 /// \brief A predicate to test whether an alloca belongs to a set.
3648 class IsAllocaInSet {
3649 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3653 typedef AllocaInst *argument_type;
3655 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3656 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3660 bool SROA::runOnFunction(Function &F) {
3661 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3662 C = &F.getContext();
3663 TD = getAnalysisIfAvailable<DataLayout>();
3665 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3668 DT = getAnalysisIfAvailable<DominatorTree>();
3670 BasicBlock &EntryBB = F.getEntryBlock();
3671 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3673 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3674 Worklist.insert(AI);
3676 bool Changed = false;
3677 // A set of deleted alloca instruction pointers which should be removed from
3678 // the list of promotable allocas.
3679 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3682 while (!Worklist.empty()) {
3683 Changed |= runOnAlloca(*Worklist.pop_back_val());
3684 deleteDeadInstructions(DeletedAllocas);
3686 // Remove the deleted allocas from various lists so that we don't try to
3687 // continue processing them.
3688 if (!DeletedAllocas.empty()) {
3689 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3690 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3691 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3692 PromotableAllocas.end(),
3693 IsAllocaInSet(DeletedAllocas)),
3694 PromotableAllocas.end());
3695 DeletedAllocas.clear();
3699 Changed |= promoteAllocas(F);
3701 Worklist = PostPromotionWorklist;
3702 PostPromotionWorklist.clear();
3703 } while (!Worklist.empty());
3708 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3709 if (RequiresDomTree)
3710 AU.addRequired<DominatorTree>();
3711 AU.setPreservesCFG();