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(NumAllocaPartitions, "Number of alloca partitions formed");
61 STATISTIC(MaxPartitionsPerAlloca, "Maximum number of partitions");
62 STATISTIC(NumAllocaPartitionUses, "Number of alloca partition uses found");
63 STATISTIC(MaxPartitionUsesPerAlloca, "Maximum number of partition uses");
64 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
65 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
66 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
67 STATISTIC(NumDeleted, "Number of instructions deleted");
68 STATISTIC(NumVectorized, "Number of vectorized aggregates");
70 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
71 /// forming SSA values through the SSAUpdater infrastructure.
73 ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
76 /// \brief A custom IRBuilder inserter which prefixes all names if they are
78 template <bool preserveNames = true>
79 class IRBuilderPrefixedInserter :
80 public IRBuilderDefaultInserter<preserveNames> {
84 void SetNamePrefix(const Twine &P) { Prefix = P.str(); }
87 void InsertHelper(Instruction *I, const Twine &Name, BasicBlock *BB,
88 BasicBlock::iterator InsertPt) const {
89 IRBuilderDefaultInserter<preserveNames>::InsertHelper(
90 I, Name.isTriviallyEmpty() ? Name : Prefix + Name, BB, InsertPt);
94 // Specialization for not preserving the name is trivial.
96 class IRBuilderPrefixedInserter<false> :
97 public IRBuilderDefaultInserter<false> {
99 void SetNamePrefix(const Twine &P) {}
102 /// \brief Provide a typedef for IRBuilder that drops names in release builds.
104 typedef llvm::IRBuilder<true, ConstantFolder,
105 IRBuilderPrefixedInserter<true> > IRBuilderTy;
107 typedef llvm::IRBuilder<false, ConstantFolder,
108 IRBuilderPrefixedInserter<false> > IRBuilderTy;
113 /// \brief A common base class for representing a half-open byte range.
115 /// \brief The beginning offset of the range.
116 uint64_t BeginOffset;
118 /// \brief The ending offset, not included in the range.
121 ByteRange() : BeginOffset(), EndOffset() {}
122 ByteRange(uint64_t BeginOffset, uint64_t EndOffset)
123 : BeginOffset(BeginOffset), EndOffset(EndOffset) {}
125 /// \brief Support for ordering ranges.
127 /// This provides an ordering over ranges such that start offsets are
128 /// always increasing, and within equal start offsets, the end offsets are
129 /// decreasing. Thus the spanning range comes first in a cluster with the
130 /// same start position.
131 bool operator<(const ByteRange &RHS) const {
132 if (BeginOffset < RHS.BeginOffset) return true;
133 if (BeginOffset > RHS.BeginOffset) return false;
134 if (EndOffset > RHS.EndOffset) return true;
138 /// \brief Support comparison with a single offset to allow binary searches.
139 friend bool operator<(const ByteRange &LHS, uint64_t RHSOffset) {
140 return LHS.BeginOffset < RHSOffset;
143 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
144 const ByteRange &RHS) {
145 return LHSOffset < RHS.BeginOffset;
148 bool operator==(const ByteRange &RHS) const {
149 return BeginOffset == RHS.BeginOffset && EndOffset == RHS.EndOffset;
151 bool operator!=(const ByteRange &RHS) const { return !operator==(RHS); }
154 /// \brief A partition of an alloca.
156 /// This structure represents a contiguous partition of the alloca. These are
157 /// formed by examining the uses of the alloca. During formation, they may
158 /// overlap but once an AllocaPartitioning is built, the Partitions within it
159 /// are all disjoint.
160 struct Partition : public ByteRange {
161 /// \brief Whether this partition is splittable into smaller partitions.
163 /// We flag partitions as splittable when they are formed entirely due to
164 /// accesses by trivially splittable operations such as memset and memcpy.
167 /// \brief Test whether a partition has been marked as dead.
168 bool isDead() const {
169 if (BeginOffset == UINT64_MAX) {
170 assert(EndOffset == UINT64_MAX);
176 /// \brief Kill a partition.
177 /// This is accomplished by setting both its beginning and end offset to
178 /// the maximum possible value.
180 assert(!isDead() && "He's Dead, Jim!");
181 BeginOffset = EndOffset = UINT64_MAX;
184 Partition() : ByteRange(), IsSplittable() {}
185 Partition(uint64_t BeginOffset, uint64_t EndOffset, bool IsSplittable)
186 : ByteRange(BeginOffset, EndOffset), IsSplittable(IsSplittable) {}
189 /// \brief A particular use of a partition of the alloca.
191 /// This structure is used to associate uses of a partition with it. They
192 /// mark the range of bytes which are referenced by a particular instruction,
193 /// and includes a handle to the user itself and the pointer value in use.
194 /// The bounds of these uses are determined by intersecting the bounds of the
195 /// memory use itself with a particular partition. As a consequence there is
196 /// intentionally overlap between various uses of the same partition.
197 class PartitionUse : public ByteRange {
198 /// \brief Combined storage for both the Use* and split state.
199 PointerIntPair<Use*, 1, bool> UsePtrAndIsSplit;
202 PartitionUse() : ByteRange(), UsePtrAndIsSplit() {}
203 PartitionUse(uint64_t BeginOffset, uint64_t EndOffset, Use *U,
205 : ByteRange(BeginOffset, EndOffset), UsePtrAndIsSplit(U, IsSplit) {}
207 /// \brief The use in question. Provides access to both user and used value.
209 /// Note that this may be null if the partition use is *dead*, that is, it
210 /// should be ignored.
211 Use *getUse() const { return UsePtrAndIsSplit.getPointer(); }
213 /// \brief Set the use for this partition use range.
214 void setUse(Use *U) { UsePtrAndIsSplit.setPointer(U); }
216 /// \brief Whether this use is split across multiple partitions.
217 bool isSplit() const { return UsePtrAndIsSplit.getInt(); }
222 template <> struct isPodLike<Partition> : llvm::true_type {};
223 template <> struct isPodLike<PartitionUse> : llvm::true_type {};
227 /// \brief Alloca partitioning representation.
229 /// This class represents a partitioning of an alloca into slices, and
230 /// information about the nature of uses of each slice of the alloca. The goal
231 /// is that this information is sufficient to decide if and how to split the
232 /// alloca apart and replace slices with scalars. It is also intended that this
233 /// structure can capture the relevant information needed both to decide about
234 /// and to enact these transformations.
235 class AllocaPartitioning {
237 /// \brief Construct a partitioning of a particular alloca.
239 /// Construction does most of the work for partitioning the alloca. This
240 /// performs the necessary walks of users and builds a partitioning from it.
241 AllocaPartitioning(const DataLayout &TD, AllocaInst &AI);
243 /// \brief Test whether a pointer to the allocation escapes our analysis.
245 /// If this is true, the partitioning is never fully built and should be
247 bool isEscaped() const { return PointerEscapingInstr; }
249 /// \brief Support for iterating over the partitions.
251 typedef SmallVectorImpl<Partition>::iterator iterator;
252 iterator begin() { return Partitions.begin(); }
253 iterator end() { return Partitions.end(); }
255 typedef SmallVectorImpl<Partition>::const_iterator const_iterator;
256 const_iterator begin() const { return Partitions.begin(); }
257 const_iterator end() const { return Partitions.end(); }
260 /// \brief Support for iterating over and manipulating a particular
261 /// partition's uses.
263 /// The iteration support provided for uses is more limited, but also
264 /// includes some manipulation routines to support rewriting the uses of
265 /// partitions during SROA.
267 typedef SmallVectorImpl<PartitionUse>::iterator use_iterator;
268 use_iterator use_begin(unsigned Idx) { return Uses[Idx].begin(); }
269 use_iterator use_begin(const_iterator I) { return Uses[I - begin()].begin(); }
270 use_iterator use_end(unsigned Idx) { return Uses[Idx].end(); }
271 use_iterator use_end(const_iterator I) { return Uses[I - begin()].end(); }
273 typedef SmallVectorImpl<PartitionUse>::const_iterator const_use_iterator;
274 const_use_iterator use_begin(unsigned Idx) const { return Uses[Idx].begin(); }
275 const_use_iterator use_begin(const_iterator I) const {
276 return Uses[I - begin()].begin();
278 const_use_iterator use_end(unsigned Idx) const { return Uses[Idx].end(); }
279 const_use_iterator use_end(const_iterator I) const {
280 return Uses[I - begin()].end();
283 unsigned use_size(unsigned Idx) const { return Uses[Idx].size(); }
284 unsigned use_size(const_iterator I) const { return Uses[I - begin()].size(); }
285 const PartitionUse &getUse(unsigned PIdx, unsigned UIdx) const {
286 return Uses[PIdx][UIdx];
288 const PartitionUse &getUse(const_iterator I, unsigned UIdx) const {
289 return Uses[I - begin()][UIdx];
292 void use_push_back(unsigned Idx, const PartitionUse &PU) {
293 Uses[Idx].push_back(PU);
295 void use_push_back(const_iterator I, const PartitionUse &PU) {
296 Uses[I - begin()].push_back(PU);
300 /// \brief Allow iterating the dead users for this alloca.
302 /// These are instructions which will never actually use the alloca as they
303 /// are outside the allocated range. They are safe to replace with undef and
306 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
307 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
308 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
311 /// \brief Allow iterating the dead expressions referring to this alloca.
313 /// These are operands which have cannot actually be used to refer to the
314 /// alloca as they are outside its range and the user doesn't correct for
315 /// that. These mostly consist of PHI node inputs and the like which we just
316 /// need to replace with undef.
318 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
319 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
320 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
323 /// \brief MemTransferInst auxiliary data.
324 /// This struct provides some auxiliary data about memory transfer
325 /// intrinsics such as memcpy and memmove. These intrinsics can use two
326 /// different ranges within the same alloca, and provide other challenges to
327 /// correctly represent. We stash extra data to help us untangle this
328 /// after the partitioning is complete.
329 struct MemTransferOffsets {
330 /// The destination begin and end offsets when the destination is within
331 /// this alloca. If the end offset is zero the destination is not within
333 uint64_t DestBegin, DestEnd;
335 /// The source begin and end offsets when the source is within this alloca.
336 /// If the end offset is zero, the source is not within this alloca.
337 uint64_t SourceBegin, SourceEnd;
339 /// Flag for whether an alloca is splittable.
342 MemTransferOffsets getMemTransferOffsets(MemTransferInst &II) const {
343 return MemTransferInstData.lookup(&II);
346 /// \brief Map from a PHI or select operand back to a partition.
348 /// When manipulating PHI nodes or selects, they can use more than one
349 /// partition of an alloca. We store a special mapping to allow finding the
350 /// partition referenced by each of these operands, if any.
351 iterator findPartitionForPHIOrSelectOperand(Use *U) {
352 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
353 = PHIOrSelectOpMap.find(U);
354 if (MapIt == PHIOrSelectOpMap.end())
357 return begin() + MapIt->second.first;
360 /// \brief Map from a PHI or select operand back to the specific use of
363 /// Similar to mapping these operands back to the partitions, this maps
364 /// directly to the use structure of that partition.
365 use_iterator findPartitionUseForPHIOrSelectOperand(Use *U) {
366 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
367 = PHIOrSelectOpMap.find(U);
368 assert(MapIt != PHIOrSelectOpMap.end());
369 return Uses[MapIt->second.first].begin() + MapIt->second.second;
372 /// \brief Compute a common type among the uses of a particular partition.
374 /// This routines walks all of the uses of a particular partition and tries
375 /// to find a common type between them. Untyped operations such as memset and
376 /// memcpy are ignored.
377 Type *getCommonType(iterator I) const;
379 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
380 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
381 void printUsers(raw_ostream &OS, const_iterator I,
382 StringRef Indent = " ") const;
383 void print(raw_ostream &OS) const;
384 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const;
385 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const;
389 template <typename DerivedT, typename RetT = void> class BuilderBase;
390 class PartitionBuilder;
391 friend class AllocaPartitioning::PartitionBuilder;
393 friend class AllocaPartitioning::UseBuilder;
395 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
396 /// \brief Handle to alloca instruction to simplify method interfaces.
400 /// \brief The instruction responsible for this alloca having no partitioning.
402 /// When an instruction (potentially) escapes the pointer to the alloca, we
403 /// store a pointer to that here and abort trying to partition the alloca.
404 /// This will be null if the alloca is partitioned successfully.
405 Instruction *PointerEscapingInstr;
407 /// \brief The partitions of the alloca.
409 /// We store a vector of the partitions over the alloca here. This vector is
410 /// sorted by increasing begin offset, and then by decreasing end offset. See
411 /// the Partition inner class for more details. Initially (during
412 /// construction) there are overlaps, but we form a disjoint sequence of
413 /// partitions while finishing construction and a fully constructed object is
414 /// expected to always have this as a disjoint space.
415 SmallVector<Partition, 8> Partitions;
417 /// \brief The uses of the partitions.
419 /// This is essentially a mapping from each partition to a list of uses of
420 /// that partition. The mapping is done with a Uses vector that has the exact
421 /// same number of entries as the partition vector. Each entry is itself
422 /// a vector of the uses.
423 SmallVector<SmallVector<PartitionUse, 2>, 8> Uses;
425 /// \brief Instructions which will become dead if we rewrite the alloca.
427 /// Note that these are not separated by partition. This is because we expect
428 /// a partitioned alloca to be completely rewritten or not rewritten at all.
429 /// If rewritten, all these instructions can simply be removed and replaced
430 /// with undef as they come from outside of the allocated space.
431 SmallVector<Instruction *, 8> DeadUsers;
433 /// \brief Operands which will become dead if we rewrite the alloca.
435 /// These are operands that in their particular use can be replaced with
436 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
437 /// to PHI nodes and the like. They aren't entirely dead (there might be
438 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
439 /// want to swap this particular input for undef to simplify the use lists of
441 SmallVector<Use *, 8> DeadOperands;
443 /// \brief The underlying storage for auxiliary memcpy and memset info.
444 SmallDenseMap<MemTransferInst *, MemTransferOffsets, 4> MemTransferInstData;
446 /// \brief A side datastructure used when building up the partitions and uses.
448 /// This mapping is only really used during the initial building of the
449 /// partitioning so that we can retain information about PHI and select nodes
451 SmallDenseMap<Instruction *, std::pair<uint64_t, bool> > PHIOrSelectSizes;
453 /// \brief Auxiliary information for particular PHI or select operands.
454 SmallDenseMap<Use *, std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap;
456 /// \brief A utility routine called from the constructor.
458 /// This does what it says on the tin. It is the key of the alloca partition
459 /// splitting and merging. After it is called we have the desired disjoint
460 /// collection of partitions.
461 void splitAndMergePartitions();
465 static Value *foldSelectInst(SelectInst &SI) {
466 // If the condition being selected on is a constant or the same value is
467 // being selected between, fold the select. Yes this does (rarely) happen
469 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
470 return SI.getOperand(1+CI->isZero());
471 if (SI.getOperand(1) == SI.getOperand(2))
472 return SI.getOperand(1);
477 /// \brief Builder for the alloca partitioning.
479 /// This class builds an alloca partitioning by recursively visiting the uses
480 /// of an alloca and splitting the partitions for each load and store at each
482 class AllocaPartitioning::PartitionBuilder
483 : public PtrUseVisitor<PartitionBuilder> {
484 friend class PtrUseVisitor<PartitionBuilder>;
485 friend class InstVisitor<PartitionBuilder>;
486 typedef PtrUseVisitor<PartitionBuilder> Base;
488 const uint64_t AllocSize;
489 AllocaPartitioning &P;
491 SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap;
494 PartitionBuilder(const DataLayout &DL, AllocaInst &AI, AllocaPartitioning &P)
495 : PtrUseVisitor<PartitionBuilder>(DL),
496 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())),
500 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size,
501 bool IsSplittable = false) {
502 // Completely skip uses which have a zero size or start either before or
503 // past the end of the allocation.
504 if (Size == 0 || Offset.isNegative() || Offset.uge(AllocSize)) {
505 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
506 << " which has zero size or starts outside of the "
507 << AllocSize << " byte alloca:\n"
508 << " alloca: " << P.AI << "\n"
509 << " use: " << I << "\n");
513 uint64_t BeginOffset = Offset.getZExtValue();
514 uint64_t EndOffset = BeginOffset + Size;
516 // Clamp the end offset to the end of the allocation. Note that this is
517 // formulated to handle even the case where "BeginOffset + Size" overflows.
518 // This may appear superficially to be something we could ignore entirely,
519 // but that is not so! There may be widened loads or PHI-node uses where
520 // some instructions are dead but not others. We can't completely ignore
521 // them, and so have to record at least the information here.
522 assert(AllocSize >= BeginOffset); // Established above.
523 if (Size > AllocSize - BeginOffset) {
524 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
525 << " to remain within the " << AllocSize << " byte alloca:\n"
526 << " alloca: " << P.AI << "\n"
527 << " use: " << I << "\n");
528 EndOffset = AllocSize;
531 Partition New(BeginOffset, EndOffset, IsSplittable);
532 P.Partitions.push_back(New);
535 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset,
536 uint64_t Size, bool IsVolatile) {
537 // We allow splitting of loads and stores where the type is an integer type
538 // and cover the entire alloca. This prevents us from splitting over
540 // FIXME: In the great blue eventually, we should eagerly split all integer
541 // loads and stores, and then have a separate step that merges adjacent
542 // alloca partitions into a single partition suitable for integer widening.
543 // Or we should skip the merge step and rely on GVN and other passes to
544 // merge adjacent loads and stores that survive mem2reg.
546 Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize;
548 insertUse(I, Offset, Size, IsSplittable);
551 void visitLoadInst(LoadInst &LI) {
552 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
553 "All simple FCA loads should have been pre-split");
556 return PI.setAborted(&LI);
558 uint64_t Size = DL.getTypeStoreSize(LI.getType());
559 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile());
562 void visitStoreInst(StoreInst &SI) {
563 Value *ValOp = SI.getValueOperand();
565 return PI.setEscapedAndAborted(&SI);
567 return PI.setAborted(&SI);
569 uint64_t Size = DL.getTypeStoreSize(ValOp->getType());
571 // If this memory access can be shown to *statically* extend outside the
572 // bounds of of the allocation, it's behavior is undefined, so simply
573 // ignore it. Note that this is more strict than the generic clamping
574 // behavior of insertUse. We also try to handle cases which might run the
576 // FIXME: We should instead consider the pointer to have escaped if this
577 // function is being instrumented for addressing bugs or race conditions.
578 if (Offset.isNegative() || Size > AllocSize ||
579 Offset.ugt(AllocSize - Size)) {
580 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset
581 << " which extends past the end of the " << AllocSize
583 << " alloca: " << P.AI << "\n"
584 << " use: " << SI << "\n");
588 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
589 "All simple FCA stores should have been pre-split");
590 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile());
594 void visitMemSetInst(MemSetInst &II) {
595 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
596 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
597 if ((Length && Length->getValue() == 0) ||
598 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
599 // Zero-length mem transfer intrinsics can be ignored entirely.
603 return PI.setAborted(&II);
605 insertUse(II, Offset,
606 Length ? Length->getLimitedValue()
607 : AllocSize - Offset.getLimitedValue(),
611 void visitMemTransferInst(MemTransferInst &II) {
612 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
613 if ((Length && Length->getValue() == 0) ||
614 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
615 // Zero-length mem transfer intrinsics can be ignored entirely.
619 return PI.setAborted(&II);
621 uint64_t RawOffset = Offset.getLimitedValue();
622 uint64_t Size = Length ? Length->getLimitedValue()
623 : AllocSize - RawOffset;
625 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
627 // Only intrinsics with a constant length can be split.
628 Offsets.IsSplittable = Length;
630 if (*U == II.getRawDest()) {
631 Offsets.DestBegin = RawOffset;
632 Offsets.DestEnd = RawOffset + Size;
634 if (*U == II.getRawSource()) {
635 Offsets.SourceBegin = RawOffset;
636 Offsets.SourceEnd = RawOffset + Size;
639 // If we have set up end offsets for both the source and the destination,
640 // we have found both sides of this transfer pointing at the same alloca.
641 bool SeenBothEnds = Offsets.SourceEnd && Offsets.DestEnd;
642 if (SeenBothEnds && II.getRawDest() != II.getRawSource()) {
643 unsigned PrevIdx = MemTransferPartitionMap[&II];
645 // Check if the begin offsets match and this is a non-volatile transfer.
646 // In that case, we can completely elide the transfer.
647 if (!II.isVolatile() && Offsets.SourceBegin == Offsets.DestBegin) {
648 P.Partitions[PrevIdx].kill();
652 // Otherwise we have an offset transfer within the same alloca. We can't
654 P.Partitions[PrevIdx].IsSplittable = Offsets.IsSplittable = false;
655 } else if (SeenBothEnds) {
656 // Handle the case where this exact use provides both ends of the
658 assert(II.getRawDest() == II.getRawSource());
660 // For non-volatile transfers this is a no-op.
661 if (!II.isVolatile())
664 // Otherwise just suppress splitting.
665 Offsets.IsSplittable = false;
669 // Insert the use now that we've fixed up the splittable nature.
670 insertUse(II, Offset, Size, Offsets.IsSplittable);
672 // Setup the mapping from intrinsic to partition of we've not seen both
673 // ends of this transfer.
675 unsigned NewIdx = P.Partitions.size() - 1;
677 = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx)).second;
679 "Already have intrinsic in map but haven't seen both ends");
684 // Disable SRoA for any intrinsics except for lifetime invariants.
685 // FIXME: What about debug intrinsics? This matches old behavior, but
686 // doesn't make sense.
687 void visitIntrinsicInst(IntrinsicInst &II) {
689 return PI.setAborted(&II);
691 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
692 II.getIntrinsicID() == Intrinsic::lifetime_end) {
693 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
694 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(),
695 Length->getLimitedValue());
696 insertUse(II, Offset, Size, true);
700 Base::visitIntrinsicInst(II);
703 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
704 // We consider any PHI or select that results in a direct load or store of
705 // the same offset to be a viable use for partitioning purposes. These uses
706 // are considered unsplittable and the size is the maximum loaded or stored
708 SmallPtrSet<Instruction *, 4> Visited;
709 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
710 Visited.insert(Root);
711 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
712 // If there are no loads or stores, the access is dead. We mark that as
713 // a size zero access.
716 Instruction *I, *UsedI;
717 llvm::tie(UsedI, I) = Uses.pop_back_val();
719 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
720 Size = std::max(Size, DL.getTypeStoreSize(LI->getType()));
723 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
724 Value *Op = SI->getOperand(0);
727 Size = std::max(Size, DL.getTypeStoreSize(Op->getType()));
731 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
732 if (!GEP->hasAllZeroIndices())
734 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
735 !isa<SelectInst>(I)) {
739 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
741 if (Visited.insert(cast<Instruction>(*UI)))
742 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
743 } while (!Uses.empty());
748 void visitPHINode(PHINode &PN) {
752 return PI.setAborted(&PN);
754 // See if we already have computed info on this node.
755 std::pair<uint64_t, bool> &PHIInfo = P.PHIOrSelectSizes[&PN];
757 PHIInfo.second = true;
758 insertUse(PN, Offset, PHIInfo.first);
762 // Check for an unsafe use of the PHI node.
763 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&PN, PHIInfo.first))
764 return PI.setAborted(UnsafeI);
766 insertUse(PN, Offset, PHIInfo.first);
769 void visitSelectInst(SelectInst &SI) {
772 if (Value *Result = foldSelectInst(SI)) {
774 // If the result of the constant fold will be the pointer, recurse
775 // through the select as if we had RAUW'ed it.
781 return PI.setAborted(&SI);
783 // See if we already have computed info on this node.
784 std::pair<uint64_t, bool> &SelectInfo = P.PHIOrSelectSizes[&SI];
785 if (SelectInfo.first) {
786 SelectInfo.second = true;
787 insertUse(SI, Offset, SelectInfo.first);
791 // Check for an unsafe use of the PHI node.
792 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&SI, SelectInfo.first))
793 return PI.setAborted(UnsafeI);
795 insertUse(SI, Offset, SelectInfo.first);
798 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
799 void visitInstruction(Instruction &I) {
804 /// \brief Use adder for the alloca partitioning.
806 /// This class adds the uses of an alloca to all of the partitions which they
807 /// use. For splittable partitions, this can end up doing essentially a linear
808 /// walk of the partitions, but the number of steps remains bounded by the
809 /// total result instruction size:
810 /// - The number of partitions is a result of the number unsplittable
811 /// instructions using the alloca.
812 /// - The number of users of each partition is at worst the total number of
813 /// splittable instructions using the alloca.
814 /// Thus we will produce N * M instructions in the end, where N are the number
815 /// of unsplittable uses and M are the number of splittable. This visitor does
816 /// the exact same number of updates to the partitioning.
818 /// In the more common case, this visitor will leverage the fact that the
819 /// partition space is pre-sorted, and do a logarithmic search for the
820 /// partition needed, making the total visit a classical ((N + M) * log(N))
821 /// complexity operation.
822 class AllocaPartitioning::UseBuilder : public PtrUseVisitor<UseBuilder> {
823 friend class PtrUseVisitor<UseBuilder>;
824 friend class InstVisitor<UseBuilder>;
825 typedef PtrUseVisitor<UseBuilder> Base;
827 const uint64_t AllocSize;
828 AllocaPartitioning &P;
830 /// \brief Set to de-duplicate dead instructions found in the use walk.
831 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
834 UseBuilder(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
835 : PtrUseVisitor<UseBuilder>(TD),
836 AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())),
840 void markAsDead(Instruction &I) {
841 if (VisitedDeadInsts.insert(&I))
842 P.DeadUsers.push_back(&I);
845 void insertUse(Instruction &User, const APInt &Offset, uint64_t Size) {
846 // If the use has a zero size or extends outside of the allocation, record
847 // it as a dead use for elimination later.
848 if (Size == 0 || Offset.isNegative() || Offset.uge(AllocSize))
849 return markAsDead(User);
851 uint64_t BeginOffset = Offset.getZExtValue();
852 uint64_t EndOffset = BeginOffset + Size;
854 // Clamp the end offset to the end of the allocation. Note that this is
855 // formulated to handle even the case where "BeginOffset + Size" overflows.
856 assert(AllocSize >= BeginOffset); // Established above.
857 if (Size > AllocSize - BeginOffset)
858 EndOffset = AllocSize;
860 // NB: This only works if we have zero overlapping partitions.
861 iterator I = std::lower_bound(P.begin(), P.end(), BeginOffset);
862 if (I != P.begin() && llvm::prior(I)->EndOffset > BeginOffset)
864 iterator E = P.end();
865 bool IsSplit = llvm::next(I) != E && llvm::next(I)->BeginOffset < EndOffset;
866 for (; I != E && I->BeginOffset < EndOffset; ++I) {
867 PartitionUse NewPU(std::max(I->BeginOffset, BeginOffset),
868 std::min(I->EndOffset, EndOffset), U, IsSplit);
869 P.use_push_back(I, NewPU);
870 if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser()))
871 P.PHIOrSelectOpMap[U]
872 = std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1);
876 void visitBitCastInst(BitCastInst &BC) {
878 return markAsDead(BC);
880 return Base::visitBitCastInst(BC);
883 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
884 if (GEPI.use_empty())
885 return markAsDead(GEPI);
887 return Base::visitGetElementPtrInst(GEPI);
890 void visitLoadInst(LoadInst &LI) {
891 assert(IsOffsetKnown);
892 uint64_t Size = DL.getTypeStoreSize(LI.getType());
893 insertUse(LI, Offset, Size);
896 void visitStoreInst(StoreInst &SI) {
897 assert(IsOffsetKnown);
898 uint64_t Size = DL.getTypeStoreSize(SI.getOperand(0)->getType());
900 // If this memory access can be shown to *statically* extend outside the
901 // bounds of of the allocation, it's behavior is undefined, so simply
902 // ignore it. Note that this is more strict than the generic clamping
903 // behavior of insertUse.
904 if (Offset.isNegative() || Size > AllocSize ||
905 Offset.ugt(AllocSize - Size))
906 return markAsDead(SI);
908 insertUse(SI, Offset, Size);
911 void visitMemSetInst(MemSetInst &II) {
912 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
913 if ((Length && Length->getValue() == 0) ||
914 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
915 return markAsDead(II);
917 assert(IsOffsetKnown);
918 insertUse(II, Offset, Length ? Length->getLimitedValue()
919 : AllocSize - Offset.getLimitedValue());
922 void visitMemTransferInst(MemTransferInst &II) {
923 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
924 if ((Length && Length->getValue() == 0) ||
925 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize)))
926 return markAsDead(II);
928 assert(IsOffsetKnown);
929 uint64_t Size = Length ? Length->getLimitedValue()
930 : AllocSize - Offset.getLimitedValue();
932 const MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
933 if (!II.isVolatile() && Offsets.DestEnd && Offsets.SourceEnd &&
934 Offsets.DestBegin == Offsets.SourceBegin)
935 return markAsDead(II); // Skip identity transfers without side-effects.
937 insertUse(II, Offset, Size);
940 void visitIntrinsicInst(IntrinsicInst &II) {
941 assert(IsOffsetKnown);
942 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
943 II.getIntrinsicID() == Intrinsic::lifetime_end);
945 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
946 insertUse(II, Offset, std::min(Length->getLimitedValue(),
947 AllocSize - Offset.getLimitedValue()));
950 void insertPHIOrSelect(Instruction &User, const APInt &Offset) {
951 uint64_t Size = P.PHIOrSelectSizes.lookup(&User).first;
953 // For PHI and select operands outside the alloca, we can't nuke the entire
954 // phi or select -- the other side might still be relevant, so we special
955 // case them here and use a separate structure to track the operands
956 // themselves which should be replaced with undef.
957 if ((Offset.isNegative() && Offset.uge(Size)) ||
958 (!Offset.isNegative() && Offset.uge(AllocSize))) {
959 P.DeadOperands.push_back(U);
963 insertUse(User, Offset, Size);
966 void visitPHINode(PHINode &PN) {
968 return markAsDead(PN);
970 assert(IsOffsetKnown);
971 insertPHIOrSelect(PN, Offset);
974 void visitSelectInst(SelectInst &SI) {
976 return markAsDead(SI);
978 if (Value *Result = foldSelectInst(SI)) {
980 // If the result of the constant fold will be the pointer, recurse
981 // through the select as if we had RAUW'ed it.
984 // Otherwise the operand to the select is dead, and we can replace it
986 P.DeadOperands.push_back(U);
991 assert(IsOffsetKnown);
992 insertPHIOrSelect(SI, Offset);
995 /// \brief Unreachable, we've already visited the alloca once.
996 void visitInstruction(Instruction &I) {
997 llvm_unreachable("Unhandled instruction in use builder.");
1001 void AllocaPartitioning::splitAndMergePartitions() {
1002 size_t NumDeadPartitions = 0;
1004 // Track the range of splittable partitions that we pass when accumulating
1005 // overlapping unsplittable partitions.
1006 uint64_t SplitEndOffset = 0ull;
1008 Partition New(0ull, 0ull, false);
1010 for (unsigned i = 0, j = i, e = Partitions.size(); i != e; i = j) {
1013 if (!Partitions[i].IsSplittable || New.BeginOffset == New.EndOffset) {
1014 assert(New.BeginOffset == New.EndOffset);
1015 New = Partitions[i];
1017 assert(New.IsSplittable);
1018 New.EndOffset = std::max(New.EndOffset, Partitions[i].EndOffset);
1020 assert(New.BeginOffset != New.EndOffset);
1022 // Scan the overlapping partitions.
1023 while (j != e && New.EndOffset > Partitions[j].BeginOffset) {
1024 // If the new partition we are forming is splittable, stop at the first
1025 // unsplittable partition.
1026 if (New.IsSplittable && !Partitions[j].IsSplittable)
1029 // Grow the new partition to include any equally splittable range. 'j' is
1030 // always equally splittable when New is splittable, but when New is not
1031 // splittable, we may subsume some (or part of some) splitable partition
1032 // without growing the new one.
1033 if (New.IsSplittable == Partitions[j].IsSplittable) {
1034 New.EndOffset = std::max(New.EndOffset, Partitions[j].EndOffset);
1036 assert(!New.IsSplittable);
1037 assert(Partitions[j].IsSplittable);
1038 SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset);
1041 Partitions[j].kill();
1042 ++NumDeadPartitions;
1046 // If the new partition is splittable, chop off the end as soon as the
1047 // unsplittable subsequent partition starts and ensure we eventually cover
1048 // the splittable area.
1049 if (j != e && New.IsSplittable) {
1050 SplitEndOffset = std::max(SplitEndOffset, New.EndOffset);
1051 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1054 // Add the new partition if it differs from the original one and is
1055 // non-empty. We can end up with an empty partition here if it was
1056 // splittable but there is an unsplittable one that starts at the same
1058 if (New != Partitions[i]) {
1059 if (New.BeginOffset != New.EndOffset)
1060 Partitions.push_back(New);
1061 // Mark the old one for removal.
1062 Partitions[i].kill();
1063 ++NumDeadPartitions;
1066 New.BeginOffset = New.EndOffset;
1067 if (!New.IsSplittable) {
1068 New.EndOffset = std::max(New.EndOffset, SplitEndOffset);
1069 if (j != e && !Partitions[j].IsSplittable)
1070 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1071 New.IsSplittable = true;
1072 // If there is a trailing splittable partition which won't be fused into
1073 // the next splittable partition go ahead and add it onto the partitions
1075 if (New.BeginOffset < New.EndOffset &&
1076 (j == e || !Partitions[j].IsSplittable ||
1077 New.EndOffset < Partitions[j].BeginOffset)) {
1078 Partitions.push_back(New);
1079 New.BeginOffset = New.EndOffset = 0ull;
1084 // Re-sort the partitions now that they have been split and merged into
1085 // disjoint set of partitions. Also remove any of the dead partitions we've
1086 // replaced in the process.
1087 std::sort(Partitions.begin(), Partitions.end());
1088 if (NumDeadPartitions) {
1089 assert(Partitions.back().isDead());
1090 assert((ptrdiff_t)NumDeadPartitions ==
1091 std::count(Partitions.begin(), Partitions.end(), Partitions.back()));
1093 Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end());
1096 AllocaPartitioning::AllocaPartitioning(const DataLayout &TD, AllocaInst &AI)
1098 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1101 PointerEscapingInstr(0) {
1102 PartitionBuilder PB(TD, AI, *this);
1103 PartitionBuilder::PtrInfo PtrI = PB.visitPtr(AI);
1104 if (PtrI.isEscaped() || PtrI.isAborted()) {
1105 // FIXME: We should sink the escape vs. abort info into the caller nicely,
1106 // possibly by just storing the PtrInfo in the AllocaPartitioning.
1107 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst()
1108 : PtrI.getAbortingInst();
1109 assert(PointerEscapingInstr && "Did not track a bad instruction");
1113 // Sort the uses. This arranges for the offsets to be in ascending order,
1114 // and the sizes to be in descending order.
1115 std::sort(Partitions.begin(), Partitions.end());
1117 // Remove any partitions from the back which are marked as dead.
1118 while (!Partitions.empty() && Partitions.back().isDead())
1119 Partitions.pop_back();
1121 if (Partitions.size() > 1) {
1122 // Intersect splittability for all partitions with equal offsets and sizes.
1123 // Then remove all but the first so that we have a sequence of non-equal but
1124 // potentially overlapping partitions.
1125 for (iterator I = Partitions.begin(), J = I, E = Partitions.end(); I != E;
1128 while (J != E && *I == *J) {
1129 I->IsSplittable &= J->IsSplittable;
1133 Partitions.erase(std::unique(Partitions.begin(), Partitions.end()),
1136 // Split splittable and merge unsplittable partitions into a disjoint set
1137 // of partitions over the used space of the allocation.
1138 splitAndMergePartitions();
1141 // Record how many partitions we end up with.
1142 NumAllocaPartitions += Partitions.size();
1143 MaxPartitionsPerAlloca = std::max<unsigned>(Partitions.size(), MaxPartitionsPerAlloca);
1145 // Now build up the user lists for each of these disjoint partitions by
1146 // re-walking the recursive users of the alloca.
1147 Uses.resize(Partitions.size());
1148 UseBuilder UB(TD, AI, *this);
1149 PtrI = UB.visitPtr(AI);
1150 assert(!PtrI.isEscaped() && "Previously analyzed pointer now escapes!");
1151 assert(!PtrI.isAborted() && "Early aborted the visit of the pointer.");
1153 unsigned NumUses = 0;
1154 #if !defined(NDEBUG) || defined(LLVM_ENABLE_STATS)
1155 for (unsigned Idx = 0, Size = Uses.size(); Idx != Size; ++Idx)
1156 NumUses += Uses[Idx].size();
1158 NumAllocaPartitionUses += NumUses;
1159 MaxPartitionUsesPerAlloca = std::max<unsigned>(NumUses, MaxPartitionUsesPerAlloca);
1162 Type *AllocaPartitioning::getCommonType(iterator I) const {
1164 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
1165 Use *U = UI->getUse();
1167 continue; // Skip dead uses.
1168 if (isa<IntrinsicInst>(*U->getUser()))
1170 if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset)
1174 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser()))
1175 UserTy = LI->getType();
1176 else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser()))
1177 UserTy = SI->getValueOperand()->getType();
1179 return 0; // Bail if we have weird uses.
1181 if (IntegerType *ITy = dyn_cast<IntegerType>(UserTy)) {
1182 // If the type is larger than the partition, skip it. We only encounter
1183 // this for split integer operations where we want to use the type of the
1184 // entity causing the split.
1185 if (ITy->getBitWidth() > (I->EndOffset - I->BeginOffset)*8)
1188 // If we have found an integer type use covering the alloca, use that
1189 // regardless of the other types, as integers are often used for a "bucket
1194 if (Ty && Ty != UserTy)
1202 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1204 void AllocaPartitioning::print(raw_ostream &OS, const_iterator I,
1205 StringRef Indent) const {
1206 OS << Indent << "partition #" << (I - begin())
1207 << " [" << I->BeginOffset << "," << I->EndOffset << ")"
1208 << (I->IsSplittable ? " (splittable)" : "")
1209 << (Uses[I - begin()].empty() ? " (zero uses)" : "")
1213 void AllocaPartitioning::printUsers(raw_ostream &OS, const_iterator I,
1214 StringRef Indent) const {
1215 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
1217 continue; // Skip dead uses.
1218 OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") "
1219 << "used by: " << *UI->getUse()->getUser() << "\n";
1220 if (MemTransferInst *II =
1221 dyn_cast<MemTransferInst>(UI->getUse()->getUser())) {
1222 const MemTransferOffsets &MTO = MemTransferInstData.lookup(II);
1224 if (!MTO.IsSplittable)
1225 IsDest = UI->BeginOffset == MTO.DestBegin;
1227 IsDest = MTO.DestBegin != 0u;
1228 OS << Indent << " (original " << (IsDest ? "dest" : "source") << ": "
1229 << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin)
1230 << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n";
1235 void AllocaPartitioning::print(raw_ostream &OS) const {
1236 if (PointerEscapingInstr) {
1237 OS << "No partitioning for alloca: " << AI << "\n"
1238 << " A pointer to this alloca escaped by:\n"
1239 << " " << *PointerEscapingInstr << "\n";
1243 OS << "Partitioning of alloca: " << AI << "\n";
1244 for (const_iterator I = begin(), E = end(); I != E; ++I) {
1250 void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); }
1251 void AllocaPartitioning::dump() const { print(dbgs()); }
1253 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1257 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1259 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1260 /// the loads and stores of an alloca instruction, as well as updating its
1261 /// debug information. This is used when a domtree is unavailable and thus
1262 /// mem2reg in its full form can't be used to handle promotion of allocas to
1264 class AllocaPromoter : public LoadAndStorePromoter {
1268 SmallVector<DbgDeclareInst *, 4> DDIs;
1269 SmallVector<DbgValueInst *, 4> DVIs;
1272 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1273 AllocaInst &AI, DIBuilder &DIB)
1274 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1276 void run(const SmallVectorImpl<Instruction*> &Insts) {
1277 // Remember which alloca we're promoting (for isInstInList).
1278 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
1279 for (Value::use_iterator UI = DebugNode->use_begin(),
1280 UE = DebugNode->use_end();
1282 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
1283 DDIs.push_back(DDI);
1284 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
1285 DVIs.push_back(DVI);
1288 LoadAndStorePromoter::run(Insts);
1289 AI.eraseFromParent();
1290 while (!DDIs.empty())
1291 DDIs.pop_back_val()->eraseFromParent();
1292 while (!DVIs.empty())
1293 DVIs.pop_back_val()->eraseFromParent();
1296 virtual bool isInstInList(Instruction *I,
1297 const SmallVectorImpl<Instruction*> &Insts) const {
1298 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1299 return LI->getOperand(0) == &AI;
1300 return cast<StoreInst>(I)->getPointerOperand() == &AI;
1303 virtual void updateDebugInfo(Instruction *Inst) const {
1304 for (SmallVectorImpl<DbgDeclareInst *>::const_iterator I = DDIs.begin(),
1305 E = DDIs.end(); I != E; ++I) {
1306 DbgDeclareInst *DDI = *I;
1307 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1308 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1309 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1310 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1312 for (SmallVectorImpl<DbgValueInst *>::const_iterator I = DVIs.begin(),
1313 E = DVIs.end(); I != E; ++I) {
1314 DbgValueInst *DVI = *I;
1316 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1317 // If an argument is zero extended then use argument directly. The ZExt
1318 // may be zapped by an optimization pass in future.
1319 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1320 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1321 else if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1322 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1324 Arg = SI->getValueOperand();
1325 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1326 Arg = LI->getPointerOperand();
1330 Instruction *DbgVal =
1331 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1333 DbgVal->setDebugLoc(DVI->getDebugLoc());
1337 } // end anon namespace
1341 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1343 /// This pass takes allocations which can be completely analyzed (that is, they
1344 /// don't escape) and tries to turn them into scalar SSA values. There are
1345 /// a few steps to this process.
1347 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1348 /// are used to try to split them into smaller allocations, ideally of
1349 /// a single scalar data type. It will split up memcpy and memset accesses
1350 /// as necessary and try to isolate individual scalar accesses.
1351 /// 2) It will transform accesses into forms which are suitable for SSA value
1352 /// promotion. This can be replacing a memset with a scalar store of an
1353 /// integer value, or it can involve speculating operations on a PHI or
1354 /// select to be a PHI or select of the results.
1355 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1356 /// onto insert and extract operations on a vector value, and convert them to
1357 /// this form. By doing so, it will enable promotion of vector aggregates to
1358 /// SSA vector values.
1359 class SROA : public FunctionPass {
1360 const bool RequiresDomTree;
1363 const DataLayout *TD;
1366 /// \brief Worklist of alloca instructions to simplify.
1368 /// Each alloca in the function is added to this. Each new alloca formed gets
1369 /// added to it as well to recursively simplify unless that alloca can be
1370 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1371 /// the one being actively rewritten, we add it back onto the list if not
1372 /// already present to ensure it is re-visited.
1373 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
1375 /// \brief A collection of instructions to delete.
1376 /// We try to batch deletions to simplify code and make things a bit more
1378 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
1380 /// \brief Post-promotion worklist.
1382 /// Sometimes we discover an alloca which has a high probability of becoming
1383 /// viable for SROA after a round of promotion takes place. In those cases,
1384 /// the alloca is enqueued here for re-processing.
1386 /// Note that we have to be very careful to clear allocas out of this list in
1387 /// the event they are deleted.
1388 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
1390 /// \brief A collection of alloca instructions we can directly promote.
1391 std::vector<AllocaInst *> PromotableAllocas;
1394 SROA(bool RequiresDomTree = true)
1395 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
1396 C(0), TD(0), DT(0) {
1397 initializeSROAPass(*PassRegistry::getPassRegistry());
1399 bool runOnFunction(Function &F);
1400 void getAnalysisUsage(AnalysisUsage &AU) const;
1402 const char *getPassName() const { return "SROA"; }
1406 friend class PHIOrSelectSpeculator;
1407 friend class AllocaPartitionRewriter;
1408 friend class AllocaPartitionVectorRewriter;
1410 bool rewriteAllocaPartition(AllocaInst &AI,
1411 AllocaPartitioning &P,
1412 AllocaPartitioning::iterator PI);
1413 bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P);
1414 bool runOnAlloca(AllocaInst &AI);
1415 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
1416 bool promoteAllocas(Function &F);
1422 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1423 return new SROA(RequiresDomTree);
1426 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1428 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
1429 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1433 /// \brief Visitor to speculate PHIs and Selects where possible.
1434 class PHIOrSelectSpeculator : public InstVisitor<PHIOrSelectSpeculator> {
1435 // Befriend the base class so it can delegate to private visit methods.
1436 friend class llvm::InstVisitor<PHIOrSelectSpeculator>;
1438 const DataLayout &TD;
1439 AllocaPartitioning &P;
1443 PHIOrSelectSpeculator(const DataLayout &TD, AllocaPartitioning &P, SROA &Pass)
1444 : TD(TD), P(P), Pass(Pass) {}
1446 /// \brief Visit the users of an alloca partition and rewrite them.
1447 void visitUsers(AllocaPartitioning::const_iterator PI) {
1448 // Note that we need to use an index here as the underlying vector of uses
1449 // may be grown during speculation. However, we never need to re-visit the
1450 // new uses, and so we can use the initial size bound.
1451 for (unsigned Idx = 0, Size = P.use_size(PI); Idx != Size; ++Idx) {
1452 const PartitionUse &PU = P.getUse(PI, Idx);
1454 continue; // Skip dead use.
1456 visit(cast<Instruction>(PU.getUse()->getUser()));
1461 // By default, skip this instruction.
1462 void visitInstruction(Instruction &I) {}
1464 /// PHI instructions that use an alloca and are subsequently loaded can be
1465 /// rewritten to load both input pointers in the pred blocks and then PHI the
1466 /// results, allowing the load of the alloca to be promoted.
1468 /// %P2 = phi [i32* %Alloca, i32* %Other]
1469 /// %V = load i32* %P2
1471 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1473 /// %V2 = load i32* %Other
1475 /// %V = phi [i32 %V1, i32 %V2]
1477 /// We can do this to a select if its only uses are loads and if the operands
1478 /// to the select can be loaded unconditionally.
1480 /// FIXME: This should be hoisted into a generic utility, likely in
1481 /// Transforms/Util/Local.h
1482 bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
1483 // For now, we can only do this promotion if the load is in the same block
1484 // as the PHI, and if there are no stores between the phi and load.
1485 // TODO: Allow recursive phi users.
1486 // TODO: Allow stores.
1487 BasicBlock *BB = PN.getParent();
1488 unsigned MaxAlign = 0;
1489 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
1491 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1492 if (LI == 0 || !LI->isSimple()) return false;
1494 // For now we only allow loads in the same block as the PHI. This is
1495 // a common case that happens when instcombine merges two loads through
1497 if (LI->getParent() != BB) return false;
1499 // Ensure that there are no instructions between the PHI and the load that
1501 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1502 if (BBI->mayWriteToMemory())
1505 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1506 Loads.push_back(LI);
1509 // We can only transform this if it is safe to push the loads into the
1510 // predecessor blocks. The only thing to watch out for is that we can't put
1511 // a possibly trapping load in the predecessor if it is a critical edge.
1512 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1513 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1514 Value *InVal = PN.getIncomingValue(Idx);
1516 // If the value is produced by the terminator of the predecessor (an
1517 // invoke) or it has side-effects, there is no valid place to put a load
1518 // in the predecessor.
1519 if (TI == InVal || TI->mayHaveSideEffects())
1522 // If the predecessor has a single successor, then the edge isn't
1524 if (TI->getNumSuccessors() == 1)
1527 // If this pointer is always safe to load, or if we can prove that there
1528 // is already a load in the block, then we can move the load to the pred
1530 if (InVal->isDereferenceablePointer() ||
1531 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
1540 void visitPHINode(PHINode &PN) {
1541 DEBUG(dbgs() << " original: " << PN << "\n");
1543 SmallVector<LoadInst *, 4> Loads;
1544 if (!isSafePHIToSpeculate(PN, Loads))
1547 assert(!Loads.empty());
1549 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1550 IRBuilderTy PHIBuilder(&PN);
1551 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1552 PN.getName() + ".sroa.speculated");
1554 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1555 // matter which one we get and if any differ.
1556 LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
1557 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1558 unsigned Align = SomeLoad->getAlignment();
1560 // Rewrite all loads of the PN to use the new PHI.
1562 LoadInst *LI = Loads.pop_back_val();
1563 LI->replaceAllUsesWith(NewPN);
1564 Pass.DeadInsts.insert(LI);
1565 } while (!Loads.empty());
1567 // Inject loads into all of the pred blocks.
1568 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1569 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1570 TerminatorInst *TI = Pred->getTerminator();
1571 Use *InUse = &PN.getOperandUse(PN.getOperandNumForIncomingValue(Idx));
1572 Value *InVal = PN.getIncomingValue(Idx);
1573 IRBuilderTy PredBuilder(TI);
1576 = PredBuilder.CreateLoad(InVal, (PN.getName() + ".sroa.speculate.load." +
1578 ++NumLoadsSpeculated;
1579 Load->setAlignment(Align);
1581 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1582 NewPN->addIncoming(Load, Pred);
1584 Instruction *Ptr = dyn_cast<Instruction>(InVal);
1586 // No uses to rewrite.
1589 // Try to lookup and rewrite any partition uses corresponding to this phi
1591 AllocaPartitioning::iterator PI
1592 = P.findPartitionForPHIOrSelectOperand(InUse);
1596 // Replace the Use in the PartitionUse for this operand with the Use
1598 AllocaPartitioning::use_iterator UI
1599 = P.findPartitionUseForPHIOrSelectOperand(InUse);
1600 assert(isa<PHINode>(*UI->getUse()->getUser()));
1601 UI->setUse(&Load->getOperandUse(Load->getPointerOperandIndex()));
1603 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1606 /// Select instructions that use an alloca and are subsequently loaded can be
1607 /// rewritten to load both input pointers and then select between the result,
1608 /// allowing the load of the alloca to be promoted.
1610 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1611 /// %V = load i32* %P2
1613 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1614 /// %V2 = load i32* %Other
1615 /// %V = select i1 %cond, i32 %V1, i32 %V2
1617 /// We can do this to a select if its only uses are loads and if the operand
1618 /// to the select can be loaded unconditionally.
1619 bool isSafeSelectToSpeculate(SelectInst &SI,
1620 SmallVectorImpl<LoadInst *> &Loads) {
1621 Value *TValue = SI.getTrueValue();
1622 Value *FValue = SI.getFalseValue();
1623 bool TDerefable = TValue->isDereferenceablePointer();
1624 bool FDerefable = FValue->isDereferenceablePointer();
1626 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
1628 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1629 if (LI == 0 || !LI->isSimple()) return false;
1631 // Both operands to the select need to be dereferencable, either
1632 // absolutely (e.g. allocas) or at this point because we can see other
1634 if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
1635 LI->getAlignment(), &TD))
1637 if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
1638 LI->getAlignment(), &TD))
1640 Loads.push_back(LI);
1646 void visitSelectInst(SelectInst &SI) {
1647 DEBUG(dbgs() << " original: " << SI << "\n");
1649 // If the select isn't safe to speculate, just use simple logic to emit it.
1650 SmallVector<LoadInst *, 4> Loads;
1651 if (!isSafeSelectToSpeculate(SI, Loads))
1654 IRBuilderTy IRB(&SI);
1655 Use *Ops[2] = { &SI.getOperandUse(1), &SI.getOperandUse(2) };
1656 AllocaPartitioning::iterator PIs[2];
1657 PartitionUse PUs[2];
1658 for (unsigned i = 0, e = 2; i != e; ++i) {
1659 PIs[i] = P.findPartitionForPHIOrSelectOperand(Ops[i]);
1660 if (PIs[i] != P.end()) {
1661 // If the pointer is within the partitioning, remove the select from
1662 // its uses. We'll add in the new loads below.
1663 AllocaPartitioning::use_iterator UI
1664 = P.findPartitionUseForPHIOrSelectOperand(Ops[i]);
1666 // Clear out the use here so that the offsets into the use list remain
1667 // stable but this use is ignored when rewriting.
1672 Value *TV = SI.getTrueValue();
1673 Value *FV = SI.getFalseValue();
1674 // Replace the loads of the select with a select of two loads.
1675 while (!Loads.empty()) {
1676 LoadInst *LI = Loads.pop_back_val();
1678 IRB.SetInsertPoint(LI);
1680 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1682 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1683 NumLoadsSpeculated += 2;
1685 // Transfer alignment and TBAA info if present.
1686 TL->setAlignment(LI->getAlignment());
1687 FL->setAlignment(LI->getAlignment());
1688 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1689 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1690 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1693 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1694 LI->getName() + ".sroa.speculated");
1696 LoadInst *Loads[2] = { TL, FL };
1697 for (unsigned i = 0, e = 2; i != e; ++i) {
1698 if (PIs[i] != P.end()) {
1699 Use *LoadUse = &Loads[i]->getOperandUse(0);
1700 assert(PUs[i].getUse()->get() == LoadUse->get());
1701 PUs[i].setUse(LoadUse);
1702 P.use_push_back(PIs[i], PUs[i]);
1706 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1707 LI->replaceAllUsesWith(V);
1708 Pass.DeadInsts.insert(LI);
1714 /// \brief Build a GEP out of a base pointer and indices.
1716 /// This will return the BasePtr if that is valid, or build a new GEP
1717 /// instruction using the IRBuilder if GEP-ing is needed.
1718 static Value *buildGEP(IRBuilderTy &IRB, Value *BasePtr,
1719 SmallVectorImpl<Value *> &Indices) {
1720 if (Indices.empty())
1723 // A single zero index is a no-op, so check for this and avoid building a GEP
1725 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1728 return IRB.CreateInBoundsGEP(BasePtr, Indices, "idx");
1731 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1732 /// TargetTy without changing the offset of the pointer.
1734 /// This routine assumes we've already established a properly offset GEP with
1735 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1736 /// zero-indices down through type layers until we find one the same as
1737 /// TargetTy. If we can't find one with the same type, we at least try to use
1738 /// one with the same size. If none of that works, we just produce the GEP as
1739 /// indicated by Indices to have the correct offset.
1740 static Value *getNaturalGEPWithType(IRBuilderTy &IRB, const DataLayout &TD,
1741 Value *BasePtr, Type *Ty, Type *TargetTy,
1742 SmallVectorImpl<Value *> &Indices) {
1744 return buildGEP(IRB, BasePtr, Indices);
1746 // See if we can descend into a struct and locate a field with the correct
1748 unsigned NumLayers = 0;
1749 Type *ElementTy = Ty;
1751 if (ElementTy->isPointerTy())
1753 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1754 ElementTy = SeqTy->getElementType();
1755 // Note that we use the default address space as this index is over an
1756 // array or a vector, not a pointer.
1757 Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(0), 0)));
1758 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1759 if (STy->element_begin() == STy->element_end())
1760 break; // Nothing left to descend into.
1761 ElementTy = *STy->element_begin();
1762 Indices.push_back(IRB.getInt32(0));
1767 } while (ElementTy != TargetTy);
1768 if (ElementTy != TargetTy)
1769 Indices.erase(Indices.end() - NumLayers, Indices.end());
1771 return buildGEP(IRB, BasePtr, Indices);
1774 /// \brief Recursively compute indices for a natural GEP.
1776 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1777 /// element types adding appropriate indices for the GEP.
1778 static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &TD,
1779 Value *Ptr, Type *Ty, APInt &Offset,
1781 SmallVectorImpl<Value *> &Indices) {
1783 return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices);
1785 // We can't recurse through pointer types.
1786 if (Ty->isPointerTy())
1789 // We try to analyze GEPs over vectors here, but note that these GEPs are
1790 // extremely poorly defined currently. The long-term goal is to remove GEPing
1791 // over a vector from the IR completely.
1792 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1793 unsigned ElementSizeInBits = TD.getTypeSizeInBits(VecTy->getScalarType());
1794 if (ElementSizeInBits % 8)
1795 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1796 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1797 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1798 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1800 Offset -= NumSkippedElements * ElementSize;
1801 Indices.push_back(IRB.getInt(NumSkippedElements));
1802 return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
1803 Offset, TargetTy, Indices);
1806 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1807 Type *ElementTy = ArrTy->getElementType();
1808 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1809 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1810 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1813 Offset -= NumSkippedElements * ElementSize;
1814 Indices.push_back(IRB.getInt(NumSkippedElements));
1815 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1819 StructType *STy = dyn_cast<StructType>(Ty);
1823 const StructLayout *SL = TD.getStructLayout(STy);
1824 uint64_t StructOffset = Offset.getZExtValue();
1825 if (StructOffset >= SL->getSizeInBytes())
1827 unsigned Index = SL->getElementContainingOffset(StructOffset);
1828 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1829 Type *ElementTy = STy->getElementType(Index);
1830 if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
1831 return 0; // The offset points into alignment padding.
1833 Indices.push_back(IRB.getInt32(Index));
1834 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1838 /// \brief Get a natural GEP from a base pointer to a particular offset and
1839 /// resulting in a particular type.
1841 /// The goal is to produce a "natural" looking GEP that works with the existing
1842 /// composite types to arrive at the appropriate offset and element type for
1843 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1844 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1845 /// Indices, and setting Ty to the result subtype.
1847 /// If no natural GEP can be constructed, this function returns null.
1848 static Value *getNaturalGEPWithOffset(IRBuilderTy &IRB, const DataLayout &TD,
1849 Value *Ptr, APInt Offset, Type *TargetTy,
1850 SmallVectorImpl<Value *> &Indices) {
1851 PointerType *Ty = cast<PointerType>(Ptr->getType());
1853 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1855 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1858 Type *ElementTy = Ty->getElementType();
1859 if (!ElementTy->isSized())
1860 return 0; // We can't GEP through an unsized element.
1861 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1862 if (ElementSize == 0)
1863 return 0; // Zero-length arrays can't help us build a natural GEP.
1864 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1866 Offset -= NumSkippedElements * ElementSize;
1867 Indices.push_back(IRB.getInt(NumSkippedElements));
1868 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1872 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1873 /// resulting pointer has PointerTy.
1875 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1876 /// and produces the pointer type desired. Where it cannot, it will try to use
1877 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1878 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1879 /// bitcast to the type.
1881 /// The strategy for finding the more natural GEPs is to peel off layers of the
1882 /// pointer, walking back through bit casts and GEPs, searching for a base
1883 /// pointer from which we can compute a natural GEP with the desired
1884 /// properties. The algorithm tries to fold as many constant indices into
1885 /// a single GEP as possible, thus making each GEP more independent of the
1886 /// surrounding code.
1887 static Value *getAdjustedPtr(IRBuilderTy &IRB, const DataLayout &TD,
1888 Value *Ptr, APInt Offset, Type *PointerTy) {
1889 // Even though we don't look through PHI nodes, we could be called on an
1890 // instruction in an unreachable block, which may be on a cycle.
1891 SmallPtrSet<Value *, 4> Visited;
1892 Visited.insert(Ptr);
1893 SmallVector<Value *, 4> Indices;
1895 // We may end up computing an offset pointer that has the wrong type. If we
1896 // never are able to compute one directly that has the correct type, we'll
1897 // fall back to it, so keep it around here.
1898 Value *OffsetPtr = 0;
1900 // Remember any i8 pointer we come across to re-use if we need to do a raw
1903 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1905 Type *TargetTy = PointerTy->getPointerElementType();
1908 // First fold any existing GEPs into the offset.
1909 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1910 APInt GEPOffset(Offset.getBitWidth(), 0);
1911 if (!GEP->accumulateConstantOffset(TD, GEPOffset))
1913 Offset += GEPOffset;
1914 Ptr = GEP->getPointerOperand();
1915 if (!Visited.insert(Ptr))
1919 // See if we can perform a natural GEP here.
1921 if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
1923 if (P->getType() == PointerTy) {
1924 // Zap any offset pointer that we ended up computing in previous rounds.
1925 if (OffsetPtr && OffsetPtr->use_empty())
1926 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1927 I->eraseFromParent();
1935 // Stash this pointer if we've found an i8*.
1936 if (Ptr->getType()->isIntegerTy(8)) {
1938 Int8PtrOffset = Offset;
1941 // Peel off a layer of the pointer and update the offset appropriately.
1942 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1943 Ptr = cast<Operator>(Ptr)->getOperand(0);
1944 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1945 if (GA->mayBeOverridden())
1947 Ptr = GA->getAliasee();
1951 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1952 } while (Visited.insert(Ptr));
1956 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1958 Int8PtrOffset = Offset;
1961 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1962 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1967 // On the off chance we were targeting i8*, guard the bitcast here.
1968 if (Ptr->getType() != PointerTy)
1969 Ptr = IRB.CreateBitCast(Ptr, PointerTy, "cast");
1974 /// \brief Test whether we can convert a value from the old to the new type.
1976 /// This predicate should be used to guard calls to convertValue in order to
1977 /// ensure that we only try to convert viable values. The strategy is that we
1978 /// will peel off single element struct and array wrappings to get to an
1979 /// underlying value, and convert that value.
1980 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
1983 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy))
1984 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy))
1985 if (NewITy->getBitWidth() >= OldITy->getBitWidth())
1987 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
1989 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
1992 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
1993 if (NewTy->isPointerTy() && OldTy->isPointerTy())
1995 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
2003 /// \brief Generic routine to convert an SSA value to a value of a different
2006 /// This will try various different casting techniques, such as bitcasts,
2007 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
2008 /// two types for viability with this routine.
2009 static Value *convertValue(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2011 assert(canConvertValue(DL, V->getType(), Ty) &&
2012 "Value not convertable to type");
2013 if (V->getType() == Ty)
2015 if (IntegerType *OldITy = dyn_cast<IntegerType>(V->getType()))
2016 if (IntegerType *NewITy = dyn_cast<IntegerType>(Ty))
2017 if (NewITy->getBitWidth() > OldITy->getBitWidth())
2018 return IRB.CreateZExt(V, NewITy);
2019 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
2020 return IRB.CreateIntToPtr(V, Ty);
2021 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
2022 return IRB.CreatePtrToInt(V, Ty);
2024 return IRB.CreateBitCast(V, Ty);
2027 /// \brief Test whether the given alloca partition can be promoted to a vector.
2029 /// This is a quick test to check whether we can rewrite a particular alloca
2030 /// partition (and its newly formed alloca) into a vector alloca with only
2031 /// whole-vector loads and stores such that it could be promoted to a vector
2032 /// SSA value. We only can ensure this for a limited set of operations, and we
2033 /// don't want to do the rewrites unless we are confident that the result will
2034 /// be promotable, so we have an early test here.
2035 static bool isVectorPromotionViable(const DataLayout &TD,
2037 AllocaPartitioning &P,
2038 uint64_t PartitionBeginOffset,
2039 uint64_t PartitionEndOffset,
2040 AllocaPartitioning::const_use_iterator I,
2041 AllocaPartitioning::const_use_iterator E) {
2042 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
2046 uint64_t ElementSize = TD.getTypeSizeInBits(Ty->getScalarType());
2048 // While the definition of LLVM vectors is bitpacked, we don't support sizes
2049 // that aren't byte sized.
2050 if (ElementSize % 8)
2052 assert((TD.getTypeSizeInBits(Ty) % 8) == 0 &&
2053 "vector size not a multiple of element size?");
2056 for (; I != E; ++I) {
2057 Use *U = I->getUse();
2059 continue; // Skip dead use.
2061 uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
2062 uint64_t BeginIndex = BeginOffset / ElementSize;
2063 if (BeginIndex * ElementSize != BeginOffset ||
2064 BeginIndex >= Ty->getNumElements())
2066 uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
2067 uint64_t EndIndex = EndOffset / ElementSize;
2068 if (EndIndex * ElementSize != EndOffset ||
2069 EndIndex > Ty->getNumElements())
2072 assert(EndIndex > BeginIndex && "Empty vector!");
2073 uint64_t NumElements = EndIndex - BeginIndex;
2075 = (NumElements == 1) ? Ty->getElementType()
2076 : VectorType::get(Ty->getElementType(), NumElements);
2078 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
2079 if (MI->isVolatile())
2081 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(U->getUser())) {
2082 const AllocaPartitioning::MemTransferOffsets &MTO
2083 = P.getMemTransferOffsets(*MTI);
2084 if (!MTO.IsSplittable)
2087 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) {
2088 // Disable vector promotion when there are loads or stores of an FCA.
2090 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
2091 if (LI->isVolatile())
2093 if (!canConvertValue(TD, PartitionTy, LI->getType()))
2095 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
2096 if (SI->isVolatile())
2098 if (!canConvertValue(TD, SI->getValueOperand()->getType(), PartitionTy))
2107 /// \brief Test whether the given alloca partition's integer operations can be
2108 /// widened to promotable ones.
2110 /// This is a quick test to check whether we can rewrite the integer loads and
2111 /// stores to a particular alloca into wider loads and stores and be able to
2112 /// promote the resulting alloca.
2113 static bool isIntegerWideningViable(const DataLayout &TD,
2115 uint64_t AllocBeginOffset,
2116 AllocaPartitioning &P,
2117 AllocaPartitioning::const_use_iterator I,
2118 AllocaPartitioning::const_use_iterator E) {
2119 uint64_t SizeInBits = TD.getTypeSizeInBits(AllocaTy);
2120 // Don't create integer types larger than the maximum bitwidth.
2121 if (SizeInBits > IntegerType::MAX_INT_BITS)
2124 // Don't try to handle allocas with bit-padding.
2125 if (SizeInBits != TD.getTypeStoreSizeInBits(AllocaTy))
2128 // We need to ensure that an integer type with the appropriate bitwidth can
2129 // be converted to the alloca type, whatever that is. We don't want to force
2130 // the alloca itself to have an integer type if there is a more suitable one.
2131 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2132 if (!canConvertValue(TD, AllocaTy, IntTy) ||
2133 !canConvertValue(TD, IntTy, AllocaTy))
2136 uint64_t Size = TD.getTypeStoreSize(AllocaTy);
2138 // Check the uses to ensure the uses are (likely) promotable integer uses.
2139 // Also ensure that the alloca has a covering load or store. We don't want
2140 // to widen the integer operations only to fail to promote due to some other
2141 // unsplittable entry (which we may make splittable later).
2142 bool WholeAllocaOp = false;
2143 for (; I != E; ++I) {
2144 Use *U = I->getUse();
2146 continue; // Skip dead use.
2148 uint64_t RelBegin = I->BeginOffset - AllocBeginOffset;
2149 uint64_t RelEnd = I->EndOffset - AllocBeginOffset;
2151 // We can't reasonably handle cases where the load or store extends past
2152 // the end of the aloca's type and into its padding.
2156 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) {
2157 if (LI->isVolatile())
2159 if (RelBegin == 0 && RelEnd == Size)
2160 WholeAllocaOp = true;
2161 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
2162 if (ITy->getBitWidth() < TD.getTypeStoreSizeInBits(ITy))
2166 // Non-integer loads need to be convertible from the alloca type so that
2167 // they are promotable.
2168 if (RelBegin != 0 || RelEnd != Size ||
2169 !canConvertValue(TD, AllocaTy, LI->getType()))
2171 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) {
2172 Type *ValueTy = SI->getValueOperand()->getType();
2173 if (SI->isVolatile())
2175 if (RelBegin == 0 && RelEnd == Size)
2176 WholeAllocaOp = true;
2177 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2178 if (ITy->getBitWidth() < TD.getTypeStoreSizeInBits(ITy))
2182 // Non-integer stores need to be convertible to the alloca type so that
2183 // they are promotable.
2184 if (RelBegin != 0 || RelEnd != Size ||
2185 !canConvertValue(TD, ValueTy, AllocaTy))
2187 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) {
2188 if (MI->isVolatile() || !isa<Constant>(MI->getLength()))
2190 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(U->getUser())) {
2191 const AllocaPartitioning::MemTransferOffsets &MTO
2192 = P.getMemTransferOffsets(*MTI);
2193 if (!MTO.IsSplittable)
2196 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) {
2197 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2198 II->getIntrinsicID() != Intrinsic::lifetime_end)
2204 return WholeAllocaOp;
2207 static Value *extractInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *V,
2208 IntegerType *Ty, uint64_t Offset,
2209 const Twine &Name) {
2210 DEBUG(dbgs() << " start: " << *V << "\n");
2211 IntegerType *IntTy = cast<IntegerType>(V->getType());
2212 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2213 "Element extends past full value");
2214 uint64_t ShAmt = 8*Offset;
2215 if (DL.isBigEndian())
2216 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2218 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2219 DEBUG(dbgs() << " shifted: " << *V << "\n");
2221 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2222 "Cannot extract to a larger integer!");
2224 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2225 DEBUG(dbgs() << " trunced: " << *V << "\n");
2230 static Value *insertInteger(const DataLayout &DL, IRBuilderTy &IRB, Value *Old,
2231 Value *V, uint64_t Offset, const Twine &Name) {
2232 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2233 IntegerType *Ty = cast<IntegerType>(V->getType());
2234 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2235 "Cannot insert a larger integer!");
2236 DEBUG(dbgs() << " start: " << *V << "\n");
2238 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2239 DEBUG(dbgs() << " extended: " << *V << "\n");
2241 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2242 "Element store outside of alloca store");
2243 uint64_t ShAmt = 8*Offset;
2244 if (DL.isBigEndian())
2245 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2247 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2248 DEBUG(dbgs() << " shifted: " << *V << "\n");
2251 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2252 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2253 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2254 DEBUG(dbgs() << " masked: " << *Old << "\n");
2255 V = IRB.CreateOr(Old, V, Name + ".insert");
2256 DEBUG(dbgs() << " inserted: " << *V << "\n");
2261 static Value *extractVector(IRBuilderTy &IRB, Value *V,
2262 unsigned BeginIndex, unsigned EndIndex,
2263 const Twine &Name) {
2264 VectorType *VecTy = cast<VectorType>(V->getType());
2265 unsigned NumElements = EndIndex - BeginIndex;
2266 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2268 if (NumElements == VecTy->getNumElements())
2271 if (NumElements == 1) {
2272 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex),
2274 DEBUG(dbgs() << " extract: " << *V << "\n");
2278 SmallVector<Constant*, 8> Mask;
2279 Mask.reserve(NumElements);
2280 for (unsigned i = BeginIndex; i != EndIndex; ++i)
2281 Mask.push_back(IRB.getInt32(i));
2282 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2283 ConstantVector::get(Mask),
2285 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2289 static Value *insertVector(IRBuilderTy &IRB, Value *Old, Value *V,
2290 unsigned BeginIndex, const Twine &Name) {
2291 VectorType *VecTy = cast<VectorType>(Old->getType());
2292 assert(VecTy && "Can only insert a vector into a vector");
2294 VectorType *Ty = dyn_cast<VectorType>(V->getType());
2296 // Single element to insert.
2297 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex),
2299 DEBUG(dbgs() << " insert: " << *V << "\n");
2303 assert(Ty->getNumElements() <= VecTy->getNumElements() &&
2304 "Too many elements!");
2305 if (Ty->getNumElements() == VecTy->getNumElements()) {
2306 assert(V->getType() == VecTy && "Vector type mismatch");
2309 unsigned EndIndex = BeginIndex + Ty->getNumElements();
2311 // When inserting a smaller vector into the larger to store, we first
2312 // use a shuffle vector to widen it with undef elements, and then
2313 // a second shuffle vector to select between the loaded vector and the
2315 SmallVector<Constant*, 8> Mask;
2316 Mask.reserve(VecTy->getNumElements());
2317 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2318 if (i >= BeginIndex && i < EndIndex)
2319 Mask.push_back(IRB.getInt32(i - BeginIndex));
2321 Mask.push_back(UndefValue::get(IRB.getInt32Ty()));
2322 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()),
2323 ConstantVector::get(Mask),
2325 DEBUG(dbgs() << " shuffle: " << *V << "\n");
2328 for (unsigned i = 0; i != VecTy->getNumElements(); ++i)
2329 Mask.push_back(IRB.getInt1(i >= BeginIndex && i < EndIndex));
2331 V = IRB.CreateSelect(ConstantVector::get(Mask), V, Old, Name + "blend");
2333 DEBUG(dbgs() << " blend: " << *V << "\n");
2338 /// \brief Visitor to rewrite instructions using a partition of an alloca to
2339 /// use a new alloca.
2341 /// Also implements the rewriting to vector-based accesses when the partition
2342 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2344 class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
2346 // Befriend the base class so it can delegate to private visit methods.
2347 friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
2349 const DataLayout &TD;
2350 AllocaPartitioning &P;
2352 AllocaInst &OldAI, &NewAI;
2353 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2356 // If we are rewriting an alloca partition which can be written as pure
2357 // vector operations, we stash extra information here. When VecTy is
2358 // non-null, we have some strict guarantees about the rewritten alloca:
2359 // - The new alloca is exactly the size of the vector type here.
2360 // - The accesses all either map to the entire vector or to a single
2362 // - The set of accessing instructions is only one of those handled above
2363 // in isVectorPromotionViable. Generally these are the same access kinds
2364 // which are promotable via mem2reg.
2367 uint64_t ElementSize;
2369 // This is a convenience and flag variable that will be null unless the new
2370 // alloca's integer operations should be widened to this integer type due to
2371 // passing isIntegerWideningViable above. If it is non-null, the desired
2372 // integer type will be stored here for easy access during rewriting.
2375 // The offset of the partition user currently being rewritten.
2376 uint64_t BeginOffset, EndOffset;
2379 Instruction *OldPtr;
2381 // Utility IR builder, whose name prefix is setup for each visited use, and
2382 // the insertion point is set to point to the user.
2386 AllocaPartitionRewriter(const DataLayout &TD, AllocaPartitioning &P,
2387 AllocaPartitioning::iterator PI,
2388 SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
2389 uint64_t NewBeginOffset, uint64_t NewEndOffset)
2390 : TD(TD), P(P), Pass(Pass),
2391 OldAI(OldAI), NewAI(NewAI),
2392 NewAllocaBeginOffset(NewBeginOffset),
2393 NewAllocaEndOffset(NewEndOffset),
2394 NewAllocaTy(NewAI.getAllocatedType()),
2395 VecTy(), ElementTy(), ElementSize(), IntTy(),
2396 BeginOffset(), EndOffset(), IsSplit(), OldUse(), OldPtr(),
2397 IRB(NewAI.getContext(), ConstantFolder()) {
2400 /// \brief Visit the users of the alloca partition and rewrite them.
2401 bool visitUsers(AllocaPartitioning::const_use_iterator I,
2402 AllocaPartitioning::const_use_iterator E) {
2403 if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P,
2404 NewAllocaBeginOffset, NewAllocaEndOffset,
2407 VecTy = cast<VectorType>(NewAI.getAllocatedType());
2408 ElementTy = VecTy->getElementType();
2409 assert((TD.getTypeSizeInBits(VecTy->getScalarType()) % 8) == 0 &&
2410 "Only multiple-of-8 sized vector elements are viable");
2411 ElementSize = TD.getTypeSizeInBits(VecTy->getScalarType()) / 8;
2412 } else if (isIntegerWideningViable(TD, NewAI.getAllocatedType(),
2413 NewAllocaBeginOffset, P, I, E)) {
2414 IntTy = Type::getIntNTy(NewAI.getContext(),
2415 TD.getTypeSizeInBits(NewAI.getAllocatedType()));
2417 bool CanSROA = true;
2418 for (; I != E; ++I) {
2420 continue; // Skip dead uses.
2421 BeginOffset = I->BeginOffset;
2422 EndOffset = I->EndOffset;
2423 IsSplit = I->isSplit();
2424 OldUse = I->getUse();
2425 OldPtr = cast<Instruction>(OldUse->get());
2427 Instruction *OldUserI = cast<Instruction>(OldUse->getUser());
2428 IRB.SetInsertPoint(OldUserI);
2429 IRB.SetCurrentDebugLocation(OldUserI->getDebugLoc());
2430 IRB.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) +
2433 CanSROA &= visit(cast<Instruction>(OldUse->getUser()));
2449 // Every instruction which can end up as a user must have a rewrite rule.
2450 bool visitInstruction(Instruction &I) {
2451 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2452 llvm_unreachable("No rewrite rule for this instruction!");
2455 Value *getAdjustedAllocaPtr(IRBuilderTy &IRB, Type *PointerTy) {
2456 assert(BeginOffset >= NewAllocaBeginOffset);
2457 APInt Offset(TD.getPointerSizeInBits(), BeginOffset - NewAllocaBeginOffset);
2458 return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy);
2461 /// \brief Compute suitable alignment to access an offset into the new alloca.
2462 unsigned getOffsetAlign(uint64_t Offset) {
2463 unsigned NewAIAlign = NewAI.getAlignment();
2465 NewAIAlign = TD.getABITypeAlignment(NewAI.getAllocatedType());
2466 return MinAlign(NewAIAlign, Offset);
2469 /// \brief Compute suitable alignment to access this partition of the new
2471 unsigned getPartitionAlign() {
2472 return getOffsetAlign(BeginOffset - NewAllocaBeginOffset);
2475 /// \brief Compute suitable alignment to access a type at an offset of the
2478 /// \returns zero if the type's ABI alignment is a suitable alignment,
2479 /// otherwise returns the maximal suitable alignment.
2480 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
2481 unsigned Align = getOffsetAlign(Offset);
2482 return Align == TD.getABITypeAlignment(Ty) ? 0 : Align;
2485 /// \brief Compute suitable alignment to access a type at the beginning of
2486 /// this partition of the new alloca.
2488 /// See \c getOffsetTypeAlign for details; this routine delegates to it.
2489 unsigned getPartitionTypeAlign(Type *Ty) {
2490 return getOffsetTypeAlign(Ty, BeginOffset - NewAllocaBeginOffset);
2493 unsigned getIndex(uint64_t Offset) {
2494 assert(VecTy && "Can only call getIndex when rewriting a vector");
2495 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2496 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2497 uint32_t Index = RelOffset / ElementSize;
2498 assert(Index * ElementSize == RelOffset);
2502 void deleteIfTriviallyDead(Value *V) {
2503 Instruction *I = cast<Instruction>(V);
2504 if (isInstructionTriviallyDead(I))
2505 Pass.DeadInsts.insert(I);
2508 Value *rewriteVectorizedLoadInst() {
2509 unsigned BeginIndex = getIndex(BeginOffset);
2510 unsigned EndIndex = getIndex(EndOffset);
2511 assert(EndIndex > BeginIndex && "Empty vector!");
2513 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2515 return extractVector(IRB, V, BeginIndex, EndIndex, "vec");
2518 Value *rewriteIntegerLoad(LoadInst &LI) {
2519 assert(IntTy && "We cannot insert an integer to the alloca");
2520 assert(!LI.isVolatile());
2521 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2523 V = convertValue(TD, IRB, V, IntTy);
2524 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2525 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2526 if (Offset > 0 || EndOffset < NewAllocaEndOffset)
2527 V = extractInteger(TD, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2532 bool visitLoadInst(LoadInst &LI) {
2533 DEBUG(dbgs() << " original: " << LI << "\n");
2534 Value *OldOp = LI.getOperand(0);
2535 assert(OldOp == OldPtr);
2537 uint64_t Size = EndOffset - BeginOffset;
2539 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), Size * 8)
2541 bool IsPtrAdjusted = false;
2544 V = rewriteVectorizedLoadInst();
2545 } else if (IntTy && LI.getType()->isIntegerTy()) {
2546 V = rewriteIntegerLoad(LI);
2547 } else if (BeginOffset == NewAllocaBeginOffset &&
2548 canConvertValue(TD, NewAllocaTy, LI.getType())) {
2549 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2550 LI.isVolatile(), "load");
2552 Type *LTy = TargetTy->getPointerTo();
2553 V = IRB.CreateAlignedLoad(getAdjustedAllocaPtr(IRB, LTy),
2554 getPartitionTypeAlign(TargetTy),
2555 LI.isVolatile(), "load");
2556 IsPtrAdjusted = true;
2558 V = convertValue(TD, IRB, V, TargetTy);
2561 assert(!LI.isVolatile());
2562 assert(LI.getType()->isIntegerTy() &&
2563 "Only integer type loads and stores are split");
2564 assert(Size < TD.getTypeStoreSize(LI.getType()) &&
2565 "Split load isn't smaller than original load");
2566 assert(LI.getType()->getIntegerBitWidth() ==
2567 TD.getTypeStoreSizeInBits(LI.getType()) &&
2568 "Non-byte-multiple bit width");
2569 // Move the insertion point just past the load so that we can refer to it.
2570 IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI)));
2571 // Create a placeholder value with the same type as LI to use as the
2572 // basis for the new value. This allows us to replace the uses of LI with
2573 // the computed value, and then replace the placeholder with LI, leaving
2574 // LI only used for this computation.
2576 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2577 V = insertInteger(TD, IRB, Placeholder, V, BeginOffset,
2579 LI.replaceAllUsesWith(V);
2580 Placeholder->replaceAllUsesWith(&LI);
2583 LI.replaceAllUsesWith(V);
2586 Pass.DeadInsts.insert(&LI);
2587 deleteIfTriviallyDead(OldOp);
2588 DEBUG(dbgs() << " to: " << *V << "\n");
2589 return !LI.isVolatile() && !IsPtrAdjusted;
2592 bool rewriteVectorizedStoreInst(Value *V,
2593 StoreInst &SI, Value *OldOp) {
2594 if (V->getType() != VecTy) {
2595 unsigned BeginIndex = getIndex(BeginOffset);
2596 unsigned EndIndex = getIndex(EndOffset);
2597 assert(EndIndex > BeginIndex && "Empty vector!");
2598 unsigned NumElements = EndIndex - BeginIndex;
2599 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2601 = (NumElements == 1) ? ElementTy
2602 : VectorType::get(ElementTy, NumElements);
2603 if (V->getType() != PartitionTy)
2604 V = convertValue(TD, IRB, V, PartitionTy);
2606 // Mix in the existing elements.
2607 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2609 V = insertVector(IRB, Old, V, BeginIndex, "vec");
2611 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2612 Pass.DeadInsts.insert(&SI);
2615 DEBUG(dbgs() << " to: " << *Store << "\n");
2619 bool rewriteIntegerStore(Value *V, StoreInst &SI) {
2620 assert(IntTy && "We cannot extract an integer from the alloca");
2621 assert(!SI.isVolatile());
2622 if (TD.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2623 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2625 Old = convertValue(TD, IRB, Old, IntTy);
2626 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2627 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2628 V = insertInteger(TD, IRB, Old, SI.getValueOperand(), Offset,
2631 V = convertValue(TD, IRB, V, NewAllocaTy);
2632 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2633 Pass.DeadInsts.insert(&SI);
2635 DEBUG(dbgs() << " to: " << *Store << "\n");
2639 bool visitStoreInst(StoreInst &SI) {
2640 DEBUG(dbgs() << " original: " << SI << "\n");
2641 Value *OldOp = SI.getOperand(1);
2642 assert(OldOp == OldPtr);
2644 Value *V = SI.getValueOperand();
2646 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2647 // alloca that should be re-examined after promoting this alloca.
2648 if (V->getType()->isPointerTy())
2649 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2650 Pass.PostPromotionWorklist.insert(AI);
2652 uint64_t Size = EndOffset - BeginOffset;
2653 if (Size < TD.getTypeStoreSize(V->getType())) {
2654 assert(!SI.isVolatile());
2655 assert(IsSplit && "A seemingly split store isn't splittable");
2656 assert(V->getType()->isIntegerTy() &&
2657 "Only integer type loads and stores are split");
2658 assert(V->getType()->getIntegerBitWidth() ==
2659 TD.getTypeStoreSizeInBits(V->getType()) &&
2660 "Non-byte-multiple bit width");
2661 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8);
2662 V = extractInteger(TD, IRB, V, NarrowTy, BeginOffset,
2667 return rewriteVectorizedStoreInst(V, SI, OldOp);
2668 if (IntTy && V->getType()->isIntegerTy())
2669 return rewriteIntegerStore(V, SI);
2672 if (BeginOffset == NewAllocaBeginOffset &&
2673 EndOffset == NewAllocaEndOffset &&
2674 canConvertValue(TD, V->getType(), NewAllocaTy)) {
2675 V = convertValue(TD, IRB, V, NewAllocaTy);
2676 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2679 Value *NewPtr = getAdjustedAllocaPtr(IRB, V->getType()->getPointerTo());
2680 NewSI = IRB.CreateAlignedStore(V, NewPtr,
2681 getPartitionTypeAlign(V->getType()),
2685 Pass.DeadInsts.insert(&SI);
2686 deleteIfTriviallyDead(OldOp);
2688 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2689 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2692 /// \brief Compute an integer value from splatting an i8 across the given
2693 /// number of bytes.
2695 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't
2696 /// call this routine.
2697 /// FIXME: Heed the advice above.
2699 /// \param V The i8 value to splat.
2700 /// \param Size The number of bytes in the output (assuming i8 is one byte)
2701 Value *getIntegerSplat(Value *V, unsigned Size) {
2702 assert(Size > 0 && "Expected a positive number of bytes.");
2703 IntegerType *VTy = cast<IntegerType>(V->getType());
2704 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte");
2708 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2709 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, "zext"),
2710 ConstantExpr::getUDiv(
2711 Constant::getAllOnesValue(SplatIntTy),
2712 ConstantExpr::getZExt(
2713 Constant::getAllOnesValue(V->getType()),
2719 /// \brief Compute a vector splat for a given element value.
2720 Value *getVectorSplat(Value *V, unsigned NumElements) {
2721 V = IRB.CreateVectorSplat(NumElements, V, "vsplat");
2722 DEBUG(dbgs() << " splat: " << *V << "\n");
2726 bool visitMemSetInst(MemSetInst &II) {
2727 DEBUG(dbgs() << " original: " << II << "\n");
2728 assert(II.getRawDest() == OldPtr);
2730 // If the memset has a variable size, it cannot be split, just adjust the
2731 // pointer to the new alloca.
2732 if (!isa<Constant>(II.getLength())) {
2733 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2734 Type *CstTy = II.getAlignmentCst()->getType();
2735 II.setAlignment(ConstantInt::get(CstTy, getPartitionAlign()));
2737 deleteIfTriviallyDead(OldPtr);
2741 // Record this instruction for deletion.
2742 Pass.DeadInsts.insert(&II);
2744 Type *AllocaTy = NewAI.getAllocatedType();
2745 Type *ScalarTy = AllocaTy->getScalarType();
2747 // If this doesn't map cleanly onto the alloca type, and that type isn't
2748 // a single value type, just emit a memset.
2749 if (!VecTy && !IntTy &&
2750 (BeginOffset != NewAllocaBeginOffset ||
2751 EndOffset != NewAllocaEndOffset ||
2752 !AllocaTy->isSingleValueType() ||
2753 !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)) ||
2754 TD.getTypeSizeInBits(ScalarTy)%8 != 0)) {
2755 Type *SizeTy = II.getLength()->getType();
2756 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2758 = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB,
2759 II.getRawDest()->getType()),
2760 II.getValue(), Size, getPartitionAlign(),
2763 DEBUG(dbgs() << " to: " << *New << "\n");
2767 // If we can represent this as a simple value, we have to build the actual
2768 // value to store, which requires expanding the byte present in memset to
2769 // a sensible representation for the alloca type. This is essentially
2770 // splatting the byte to a sufficiently wide integer, splatting it across
2771 // any desired vector width, and bitcasting to the final type.
2775 // If this is a memset of a vectorized alloca, insert it.
2776 assert(ElementTy == ScalarTy);
2778 unsigned BeginIndex = getIndex(BeginOffset);
2779 unsigned EndIndex = getIndex(EndOffset);
2780 assert(EndIndex > BeginIndex && "Empty vector!");
2781 unsigned NumElements = EndIndex - BeginIndex;
2782 assert(NumElements <= VecTy->getNumElements() && "Too many elements!");
2785 getIntegerSplat(II.getValue(), TD.getTypeSizeInBits(ElementTy) / 8);
2786 Splat = convertValue(TD, IRB, Splat, ElementTy);
2787 if (NumElements > 1)
2788 Splat = getVectorSplat(Splat, NumElements);
2790 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2792 V = insertVector(IRB, Old, Splat, BeginIndex, "vec");
2794 // If this is a memset on an alloca where we can widen stores, insert the
2796 assert(!II.isVolatile());
2798 uint64_t Size = EndOffset - BeginOffset;
2799 V = getIntegerSplat(II.getValue(), Size);
2801 if (IntTy && (BeginOffset != NewAllocaBeginOffset ||
2802 EndOffset != NewAllocaBeginOffset)) {
2803 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2805 Old = convertValue(TD, IRB, Old, IntTy);
2806 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2807 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2808 V = insertInteger(TD, IRB, Old, V, Offset, "insert");
2810 assert(V->getType() == IntTy &&
2811 "Wrong type for an alloca wide integer!");
2813 V = convertValue(TD, IRB, V, AllocaTy);
2815 // Established these invariants above.
2816 assert(BeginOffset == NewAllocaBeginOffset);
2817 assert(EndOffset == NewAllocaEndOffset);
2819 V = getIntegerSplat(II.getValue(), TD.getTypeSizeInBits(ScalarTy) / 8);
2820 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy))
2821 V = getVectorSplat(V, AllocaVecTy->getNumElements());
2823 V = convertValue(TD, IRB, V, AllocaTy);
2826 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2829 DEBUG(dbgs() << " to: " << *New << "\n");
2830 return !II.isVolatile();
2833 bool visitMemTransferInst(MemTransferInst &II) {
2834 // Rewriting of memory transfer instructions can be a bit tricky. We break
2835 // them into two categories: split intrinsics and unsplit intrinsics.
2837 DEBUG(dbgs() << " original: " << II << "\n");
2839 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2840 bool IsDest = II.getRawDest() == OldPtr;
2842 const AllocaPartitioning::MemTransferOffsets &MTO
2843 = P.getMemTransferOffsets(II);
2845 // Compute the relative offset within the transfer.
2846 unsigned IntPtrWidth = TD.getPointerSizeInBits();
2847 APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
2848 : MTO.SourceBegin));
2850 unsigned Align = II.getAlignment();
2852 Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2853 MinAlign(II.getAlignment(), getPartitionAlign()));
2855 // For unsplit intrinsics, we simply modify the source and destination
2856 // pointers in place. This isn't just an optimization, it is a matter of
2857 // correctness. With unsplit intrinsics we may be dealing with transfers
2858 // within a single alloca before SROA ran, or with transfers that have
2859 // a variable length. We may also be dealing with memmove instead of
2860 // memcpy, and so simply updating the pointers is the necessary for us to
2861 // update both source and dest of a single call.
2862 if (!MTO.IsSplittable) {
2863 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2865 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2867 II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType()));
2869 Type *CstTy = II.getAlignmentCst()->getType();
2870 II.setAlignment(ConstantInt::get(CstTy, Align));
2872 DEBUG(dbgs() << " to: " << II << "\n");
2873 deleteIfTriviallyDead(OldOp);
2876 // For split transfer intrinsics we have an incredibly useful assurance:
2877 // the source and destination do not reside within the same alloca, and at
2878 // least one of them does not escape. This means that we can replace
2879 // memmove with memcpy, and we don't need to worry about all manner of
2880 // downsides to splitting and transforming the operations.
2882 // If this doesn't map cleanly onto the alloca type, and that type isn't
2883 // a single value type, just emit a memcpy.
2885 = !VecTy && !IntTy && (BeginOffset != NewAllocaBeginOffset ||
2886 EndOffset != NewAllocaEndOffset ||
2887 !NewAI.getAllocatedType()->isSingleValueType());
2889 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2890 // size hasn't been shrunk based on analysis of the viable range, this is
2892 if (EmitMemCpy && &OldAI == &NewAI) {
2893 uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin;
2894 uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd;
2895 // Ensure the start lines up.
2896 assert(BeginOffset == OrigBegin);
2899 // Rewrite the size as needed.
2900 if (EndOffset != OrigEnd)
2901 II.setLength(ConstantInt::get(II.getLength()->getType(),
2902 EndOffset - BeginOffset));
2905 // Record this instruction for deletion.
2906 Pass.DeadInsts.insert(&II);
2908 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2909 // alloca that should be re-examined after rewriting this instruction.
2910 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2912 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2913 Pass.Worklist.insert(AI);
2916 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2917 : II.getRawDest()->getType();
2919 // Compute the other pointer, folding as much as possible to produce
2920 // a single, simple GEP in most cases.
2921 OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy);
2924 = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType()
2925 : II.getRawSource()->getType());
2926 Type *SizeTy = II.getLength()->getType();
2927 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2929 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2930 IsDest ? OtherPtr : OurPtr,
2931 Size, Align, II.isVolatile());
2933 DEBUG(dbgs() << " to: " << *New << "\n");
2937 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2938 // is equivalent to 1, but that isn't true if we end up rewriting this as
2943 bool IsWholeAlloca = BeginOffset == NewAllocaBeginOffset &&
2944 EndOffset == NewAllocaEndOffset;
2945 uint64_t Size = EndOffset - BeginOffset;
2946 unsigned BeginIndex = VecTy ? getIndex(BeginOffset) : 0;
2947 unsigned EndIndex = VecTy ? getIndex(EndOffset) : 0;
2948 unsigned NumElements = EndIndex - BeginIndex;
2949 IntegerType *SubIntTy
2950 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0;
2952 Type *OtherPtrTy = NewAI.getType();
2953 if (VecTy && !IsWholeAlloca) {
2954 if (NumElements == 1)
2955 OtherPtrTy = VecTy->getElementType();
2957 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements);
2959 OtherPtrTy = OtherPtrTy->getPointerTo();
2960 } else if (IntTy && !IsWholeAlloca) {
2961 OtherPtrTy = SubIntTy->getPointerTo();
2964 Value *SrcPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy);
2965 Value *DstPtr = &NewAI;
2967 std::swap(SrcPtr, DstPtr);
2970 if (VecTy && !IsWholeAlloca && !IsDest) {
2971 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2973 Src = extractVector(IRB, Src, BeginIndex, EndIndex, "vec");
2974 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2975 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2977 Src = convertValue(TD, IRB, Src, IntTy);
2978 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2979 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2980 Src = extractInteger(TD, IRB, Src, SubIntTy, Offset, "extract");
2982 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2986 if (VecTy && !IsWholeAlloca && IsDest) {
2987 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2989 Src = insertVector(IRB, Old, Src, BeginIndex, "vec");
2990 } else if (IntTy && !IsWholeAlloca && IsDest) {
2991 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2993 Old = convertValue(TD, IRB, Old, IntTy);
2994 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2995 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2996 Src = insertInteger(TD, IRB, Old, Src, Offset, "insert");
2997 Src = convertValue(TD, IRB, Src, NewAllocaTy);
3000 StoreInst *Store = cast<StoreInst>(
3001 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
3003 DEBUG(dbgs() << " to: " << *Store << "\n");
3004 return !II.isVolatile();
3007 bool visitIntrinsicInst(IntrinsicInst &II) {
3008 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
3009 II.getIntrinsicID() == Intrinsic::lifetime_end);
3010 DEBUG(dbgs() << " original: " << II << "\n");
3011 assert(II.getArgOperand(1) == OldPtr);
3013 // Record this instruction for deletion.
3014 Pass.DeadInsts.insert(&II);
3017 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
3018 EndOffset - BeginOffset);
3019 Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType());
3021 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
3022 New = IRB.CreateLifetimeStart(Ptr, Size);
3024 New = IRB.CreateLifetimeEnd(Ptr, Size);
3027 DEBUG(dbgs() << " to: " << *New << "\n");
3031 bool visitPHINode(PHINode &PN) {
3032 DEBUG(dbgs() << " original: " << PN << "\n");
3034 // We would like to compute a new pointer in only one place, but have it be
3035 // as local as possible to the PHI. To do that, we re-use the location of
3036 // the old pointer, which necessarily must be in the right position to
3037 // dominate the PHI.
3038 IRBuilderTy PtrBuilder(cast<Instruction>(OldPtr));
3039 PtrBuilder.SetNamePrefix(Twine(NewAI.getName()) + "." + Twine(BeginOffset) +
3042 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
3043 // Replace the operands which were using the old pointer.
3044 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
3046 DEBUG(dbgs() << " to: " << PN << "\n");
3047 deleteIfTriviallyDead(OldPtr);
3051 bool visitSelectInst(SelectInst &SI) {
3052 DEBUG(dbgs() << " original: " << SI << "\n");
3053 assert((SI.getTrueValue() == OldPtr || SI.getFalseValue() == OldPtr) &&
3054 "Pointer isn't an operand!");
3056 Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
3057 // Replace the operands which were using the old pointer.
3058 if (SI.getOperand(1) == OldPtr)
3059 SI.setOperand(1, NewPtr);
3060 if (SI.getOperand(2) == OldPtr)
3061 SI.setOperand(2, NewPtr);
3063 DEBUG(dbgs() << " to: " << SI << "\n");
3064 deleteIfTriviallyDead(OldPtr);
3072 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
3074 /// This pass aggressively rewrites all aggregate loads and stores on
3075 /// a particular pointer (or any pointer derived from it which we can identify)
3076 /// with scalar loads and stores.
3077 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
3078 // Befriend the base class so it can delegate to private visit methods.
3079 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
3081 const DataLayout &TD;
3083 /// Queue of pointer uses to analyze and potentially rewrite.
3084 SmallVector<Use *, 8> Queue;
3086 /// Set to prevent us from cycling with phi nodes and loops.
3087 SmallPtrSet<User *, 8> Visited;
3089 /// The current pointer use being rewritten. This is used to dig up the used
3090 /// value (as opposed to the user).
3094 AggLoadStoreRewriter(const DataLayout &TD) : TD(TD) {}
3096 /// Rewrite loads and stores through a pointer and all pointers derived from
3098 bool rewrite(Instruction &I) {
3099 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
3101 bool Changed = false;
3102 while (!Queue.empty()) {
3103 U = Queue.pop_back_val();
3104 Changed |= visit(cast<Instruction>(U->getUser()));
3110 /// Enqueue all the users of the given instruction for further processing.
3111 /// This uses a set to de-duplicate users.
3112 void enqueueUsers(Instruction &I) {
3113 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
3115 if (Visited.insert(*UI))
3116 Queue.push_back(&UI.getUse());
3119 // Conservative default is to not rewrite anything.
3120 bool visitInstruction(Instruction &I) { return false; }
3122 /// \brief Generic recursive split emission class.
3123 template <typename Derived>
3126 /// The builder used to form new instructions.
3128 /// The indices which to be used with insert- or extractvalue to select the
3129 /// appropriate value within the aggregate.
3130 SmallVector<unsigned, 4> Indices;
3131 /// The indices to a GEP instruction which will move Ptr to the correct slot
3132 /// within the aggregate.
3133 SmallVector<Value *, 4> GEPIndices;
3134 /// The base pointer of the original op, used as a base for GEPing the
3135 /// split operations.
3138 /// Initialize the splitter with an insertion point, Ptr and start with a
3139 /// single zero GEP index.
3140 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3141 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3144 /// \brief Generic recursive split emission routine.
3146 /// This method recursively splits an aggregate op (load or store) into
3147 /// scalar or vector ops. It splits recursively until it hits a single value
3148 /// and emits that single value operation via the template argument.
3150 /// The logic of this routine relies on GEPs and insertvalue and
3151 /// extractvalue all operating with the same fundamental index list, merely
3152 /// formatted differently (GEPs need actual values).
3154 /// \param Ty The type being split recursively into smaller ops.
3155 /// \param Agg The aggregate value being built up or stored, depending on
3156 /// whether this is splitting a load or a store respectively.
3157 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3158 if (Ty->isSingleValueType())
3159 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3161 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3162 unsigned OldSize = Indices.size();
3164 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3166 assert(Indices.size() == OldSize && "Did not return to the old size");
3167 Indices.push_back(Idx);
3168 GEPIndices.push_back(IRB.getInt32(Idx));
3169 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3170 GEPIndices.pop_back();
3176 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3177 unsigned OldSize = Indices.size();
3179 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3181 assert(Indices.size() == OldSize && "Did not return to the old size");
3182 Indices.push_back(Idx);
3183 GEPIndices.push_back(IRB.getInt32(Idx));
3184 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3185 GEPIndices.pop_back();
3191 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3195 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3196 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3197 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3199 /// Emit a leaf load of a single value. This is called at the leaves of the
3200 /// recursive emission to actually load values.
3201 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3202 assert(Ty->isSingleValueType());
3203 // Load the single value and insert it using the indices.
3204 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep");
3205 Value *Load = IRB.CreateLoad(GEP, Name + ".load");
3206 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3207 DEBUG(dbgs() << " to: " << *Load << "\n");
3211 bool visitLoadInst(LoadInst &LI) {
3212 assert(LI.getPointerOperand() == *U);
3213 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3216 // We have an aggregate being loaded, split it apart.
3217 DEBUG(dbgs() << " original: " << LI << "\n");
3218 LoadOpSplitter Splitter(&LI, *U);
3219 Value *V = UndefValue::get(LI.getType());
3220 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3221 LI.replaceAllUsesWith(V);
3222 LI.eraseFromParent();
3226 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3227 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3228 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3230 /// Emit a leaf store of a single value. This is called at the leaves of the
3231 /// recursive emission to actually produce stores.
3232 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3233 assert(Ty->isSingleValueType());
3234 // Extract the single value and store it using the indices.
3235 Value *Store = IRB.CreateStore(
3236 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
3237 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
3239 DEBUG(dbgs() << " to: " << *Store << "\n");
3243 bool visitStoreInst(StoreInst &SI) {
3244 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3246 Value *V = SI.getValueOperand();
3247 if (V->getType()->isSingleValueType())
3250 // We have an aggregate being stored, split it apart.
3251 DEBUG(dbgs() << " original: " << SI << "\n");
3252 StoreOpSplitter Splitter(&SI, *U);
3253 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3254 SI.eraseFromParent();
3258 bool visitBitCastInst(BitCastInst &BC) {
3263 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3268 bool visitPHINode(PHINode &PN) {
3273 bool visitSelectInst(SelectInst &SI) {
3280 /// \brief Strip aggregate type wrapping.
3282 /// This removes no-op aggregate types wrapping an underlying type. It will
3283 /// strip as many layers of types as it can without changing either the type
3284 /// size or the allocated size.
3285 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3286 if (Ty->isSingleValueType())
3289 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3290 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3293 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3294 InnerTy = ArrTy->getElementType();
3295 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3296 const StructLayout *SL = DL.getStructLayout(STy);
3297 unsigned Index = SL->getElementContainingOffset(0);
3298 InnerTy = STy->getElementType(Index);
3303 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3304 TypeSize > DL.getTypeSizeInBits(InnerTy))
3307 return stripAggregateTypeWrapping(DL, InnerTy);
3310 /// \brief Try to find a partition of the aggregate type passed in for a given
3311 /// offset and size.
3313 /// This recurses through the aggregate type and tries to compute a subtype
3314 /// based on the offset and size. When the offset and size span a sub-section
3315 /// of an array, it will even compute a new array type for that sub-section,
3316 /// and the same for structs.
3318 /// Note that this routine is very strict and tries to find a partition of the
3319 /// type which produces the *exact* right offset and size. It is not forgiving
3320 /// when the size or offset cause either end of type-based partition to be off.
3321 /// Also, this is a best-effort routine. It is reasonable to give up and not
3322 /// return a type if necessary.
3323 static Type *getTypePartition(const DataLayout &TD, Type *Ty,
3324 uint64_t Offset, uint64_t Size) {
3325 if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size)
3326 return stripAggregateTypeWrapping(TD, Ty);
3327 if (Offset > TD.getTypeAllocSize(Ty) ||
3328 (TD.getTypeAllocSize(Ty) - Offset) < Size)
3331 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3332 // We can't partition pointers...
3333 if (SeqTy->isPointerTy())
3336 Type *ElementTy = SeqTy->getElementType();
3337 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3338 uint64_t NumSkippedElements = Offset / ElementSize;
3339 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) {
3340 if (NumSkippedElements >= ArrTy->getNumElements())
3342 } else if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) {
3343 if (NumSkippedElements >= VecTy->getNumElements())
3346 Offset -= NumSkippedElements * ElementSize;
3348 // First check if we need to recurse.
3349 if (Offset > 0 || Size < ElementSize) {
3350 // Bail if the partition ends in a different array element.
3351 if ((Offset + Size) > ElementSize)
3353 // Recurse through the element type trying to peel off offset bytes.
3354 return getTypePartition(TD, ElementTy, Offset, Size);
3356 assert(Offset == 0);
3358 if (Size == ElementSize)
3359 return stripAggregateTypeWrapping(TD, ElementTy);
3360 assert(Size > ElementSize);
3361 uint64_t NumElements = Size / ElementSize;
3362 if (NumElements * ElementSize != Size)
3364 return ArrayType::get(ElementTy, NumElements);
3367 StructType *STy = dyn_cast<StructType>(Ty);
3371 const StructLayout *SL = TD.getStructLayout(STy);
3372 if (Offset >= SL->getSizeInBytes())
3374 uint64_t EndOffset = Offset + Size;
3375 if (EndOffset > SL->getSizeInBytes())
3378 unsigned Index = SL->getElementContainingOffset(Offset);
3379 Offset -= SL->getElementOffset(Index);
3381 Type *ElementTy = STy->getElementType(Index);
3382 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3383 if (Offset >= ElementSize)
3384 return 0; // The offset points into alignment padding.
3386 // See if any partition must be contained by the element.
3387 if (Offset > 0 || Size < ElementSize) {
3388 if ((Offset + Size) > ElementSize)
3390 return getTypePartition(TD, ElementTy, Offset, Size);
3392 assert(Offset == 0);
3394 if (Size == ElementSize)
3395 return stripAggregateTypeWrapping(TD, ElementTy);
3397 StructType::element_iterator EI = STy->element_begin() + Index,
3398 EE = STy->element_end();
3399 if (EndOffset < SL->getSizeInBytes()) {
3400 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3401 if (Index == EndIndex)
3402 return 0; // Within a single element and its padding.
3404 // Don't try to form "natural" types if the elements don't line up with the
3406 // FIXME: We could potentially recurse down through the last element in the
3407 // sub-struct to find a natural end point.
3408 if (SL->getElementOffset(EndIndex) != EndOffset)
3411 assert(Index < EndIndex);
3412 EE = STy->element_begin() + EndIndex;
3415 // Try to build up a sub-structure.
3416 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
3418 const StructLayout *SubSL = TD.getStructLayout(SubTy);
3419 if (Size != SubSL->getSizeInBytes())
3420 return 0; // The sub-struct doesn't have quite the size needed.
3425 /// \brief Rewrite an alloca partition's users.
3427 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3428 /// to rewrite uses of an alloca partition to be conducive for SSA value
3429 /// promotion. If the partition needs a new, more refined alloca, this will
3430 /// build that new alloca, preserving as much type information as possible, and
3431 /// rewrite the uses of the old alloca to point at the new one and have the
3432 /// appropriate new offsets. It also evaluates how successful the rewrite was
3433 /// at enabling promotion and if it was successful queues the alloca to be
3435 bool SROA::rewriteAllocaPartition(AllocaInst &AI,
3436 AllocaPartitioning &P,
3437 AllocaPartitioning::iterator PI) {
3438 uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset;
3439 bool IsLive = false;
3440 for (AllocaPartitioning::use_iterator UI = P.use_begin(PI),
3442 UI != UE && !IsLive; ++UI)
3446 return false; // No live uses left of this partition.
3448 DEBUG(dbgs() << "Speculating PHIs and selects in partition "
3449 << "[" << PI->BeginOffset << "," << PI->EndOffset << ")\n");
3451 PHIOrSelectSpeculator Speculator(*TD, P, *this);
3452 DEBUG(dbgs() << " speculating ");
3453 DEBUG(P.print(dbgs(), PI, ""));
3454 Speculator.visitUsers(PI);
3456 // Try to compute a friendly type for this partition of the alloca. This
3457 // won't always succeed, in which case we fall back to a legal integer type
3458 // or an i8 array of an appropriate size.
3460 if (Type *PartitionTy = P.getCommonType(PI))
3461 if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize)
3462 AllocaTy = PartitionTy;
3464 if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(),
3465 PI->BeginOffset, AllocaSize))
3466 AllocaTy = PartitionTy;
3468 (AllocaTy->isArrayTy() &&
3469 AllocaTy->getArrayElementType()->isIntegerTy())) &&
3470 TD->isLegalInteger(AllocaSize * 8))
3471 AllocaTy = Type::getIntNTy(*C, AllocaSize * 8);
3473 AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize);
3474 assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize);
3476 // Check for the case where we're going to rewrite to a new alloca of the
3477 // exact same type as the original, and with the same access offsets. In that
3478 // case, re-use the existing alloca, but still run through the rewriter to
3479 // perform phi and select speculation.
3481 if (AllocaTy == AI.getAllocatedType()) {
3482 assert(PI->BeginOffset == 0 &&
3483 "Non-zero begin offset but same alloca type");
3484 assert(PI == P.begin() && "Begin offset is zero on later partition");
3487 unsigned Alignment = AI.getAlignment();
3489 // The minimum alignment which users can rely on when the explicit
3490 // alignment is omitted or zero is that required by the ABI for this
3492 Alignment = TD->getABITypeAlignment(AI.getAllocatedType());
3494 Alignment = MinAlign(Alignment, PI->BeginOffset);
3495 // If we will get at least this much alignment from the type alone, leave
3496 // the alloca's alignment unconstrained.
3497 if (Alignment <= TD->getABITypeAlignment(AllocaTy))
3499 NewAI = new AllocaInst(AllocaTy, 0, Alignment,
3500 AI.getName() + ".sroa." + Twine(PI - P.begin()),
3505 DEBUG(dbgs() << "Rewriting alloca partition "
3506 << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: "
3509 // Track the high watermark of the post-promotion worklist. We will reset it
3510 // to this point if the alloca is not in fact scheduled for promotion.
3511 unsigned PPWOldSize = PostPromotionWorklist.size();
3513 AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI,
3514 PI->BeginOffset, PI->EndOffset);
3515 DEBUG(dbgs() << " rewriting ");
3516 DEBUG(P.print(dbgs(), PI, ""));
3517 bool Promotable = Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI));
3519 DEBUG(dbgs() << " and queuing for promotion\n");
3520 PromotableAllocas.push_back(NewAI);
3521 } else if (NewAI != &AI) {
3522 // If we can't promote the alloca, iterate on it to check for new
3523 // refinements exposed by splitting the current alloca. Don't iterate on an
3524 // alloca which didn't actually change and didn't get promoted.
3525 Worklist.insert(NewAI);
3528 // Drop any post-promotion work items if promotion didn't happen.
3530 while (PostPromotionWorklist.size() > PPWOldSize)
3531 PostPromotionWorklist.pop_back();
3536 /// \brief Walks the partitioning of an alloca rewriting uses of each partition.
3537 bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
3538 bool Changed = false;
3539 for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE;
3541 Changed |= rewriteAllocaPartition(AI, P, PI);
3546 /// \brief Analyze an alloca for SROA.
3548 /// This analyzes the alloca to ensure we can reason about it, builds
3549 /// a partitioning of the alloca, and then hands it off to be split and
3550 /// rewritten as needed.
3551 bool SROA::runOnAlloca(AllocaInst &AI) {
3552 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3553 ++NumAllocasAnalyzed;
3555 // Special case dead allocas, as they're trivial.
3556 if (AI.use_empty()) {
3557 AI.eraseFromParent();
3561 // Skip alloca forms that this analysis can't handle.
3562 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3563 TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
3566 bool Changed = false;
3568 // First, split any FCA loads and stores touching this alloca to promote
3569 // better splitting and promotion opportunities.
3570 AggLoadStoreRewriter AggRewriter(*TD);
3571 Changed |= AggRewriter.rewrite(AI);
3573 // Build the partition set using a recursive instruction-visiting builder.
3574 AllocaPartitioning P(*TD, AI);
3575 DEBUG(P.print(dbgs()));
3579 // Delete all the dead users of this alloca before splitting and rewriting it.
3580 for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
3581 DE = P.dead_user_end();
3584 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3585 DeadInsts.insert(*DI);
3587 for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(),
3588 DE = P.dead_op_end();
3591 // Clobber the use with an undef value.
3592 **DO = UndefValue::get(OldV->getType());
3593 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3594 if (isInstructionTriviallyDead(OldI)) {
3596 DeadInsts.insert(OldI);
3600 // No partitions to split. Leave the dead alloca for a later pass to clean up.
3601 if (P.begin() == P.end())
3604 return splitAlloca(AI, P) || Changed;
3607 /// \brief Delete the dead instructions accumulated in this run.
3609 /// Recursively deletes the dead instructions we've accumulated. This is done
3610 /// at the very end to maximize locality of the recursive delete and to
3611 /// minimize the problems of invalidated instruction pointers as such pointers
3612 /// are used heavily in the intermediate stages of the algorithm.
3614 /// We also record the alloca instructions deleted here so that they aren't
3615 /// subsequently handed to mem2reg to promote.
3616 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3617 while (!DeadInsts.empty()) {
3618 Instruction *I = DeadInsts.pop_back_val();
3619 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3621 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3623 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3624 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3625 // Zero out the operand and see if it becomes trivially dead.
3627 if (isInstructionTriviallyDead(U))
3628 DeadInsts.insert(U);
3631 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3632 DeletedAllocas.insert(AI);
3635 I->eraseFromParent();
3639 /// \brief Promote the allocas, using the best available technique.
3641 /// This attempts to promote whatever allocas have been identified as viable in
3642 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3643 /// If there is a domtree available, we attempt to promote using the full power
3644 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3645 /// based on the SSAUpdater utilities. This function returns whether any
3646 /// promotion occurred.
3647 bool SROA::promoteAllocas(Function &F) {
3648 if (PromotableAllocas.empty())
3651 NumPromoted += PromotableAllocas.size();
3653 if (DT && !ForceSSAUpdater) {
3654 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3655 PromoteMemToReg(PromotableAllocas, *DT);
3656 PromotableAllocas.clear();
3660 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3662 DIBuilder DIB(*F.getParent());
3663 SmallVector<Instruction*, 64> Insts;
3665 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3666 AllocaInst *AI = PromotableAllocas[Idx];
3667 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
3669 Instruction *I = cast<Instruction>(*UI++);
3670 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3671 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3672 // leading to them) here. Eventually it should use them to optimize the
3673 // scalar values produced.
3674 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
3675 assert(onlyUsedByLifetimeMarkers(I) &&
3676 "Found a bitcast used outside of a lifetime marker.");
3677 while (!I->use_empty())
3678 cast<Instruction>(*I->use_begin())->eraseFromParent();
3679 I->eraseFromParent();
3682 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3683 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3684 II->getIntrinsicID() == Intrinsic::lifetime_end);
3685 II->eraseFromParent();
3691 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3695 PromotableAllocas.clear();
3700 /// \brief A predicate to test whether an alloca belongs to a set.
3701 class IsAllocaInSet {
3702 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3706 typedef AllocaInst *argument_type;
3708 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3709 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3713 bool SROA::runOnFunction(Function &F) {
3714 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3715 C = &F.getContext();
3716 TD = getAnalysisIfAvailable<DataLayout>();
3718 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3721 DT = getAnalysisIfAvailable<DominatorTree>();
3723 BasicBlock &EntryBB = F.getEntryBlock();
3724 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3726 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3727 Worklist.insert(AI);
3729 bool Changed = false;
3730 // A set of deleted alloca instruction pointers which should be removed from
3731 // the list of promotable allocas.
3732 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3735 while (!Worklist.empty()) {
3736 Changed |= runOnAlloca(*Worklist.pop_back_val());
3737 deleteDeadInstructions(DeletedAllocas);
3739 // Remove the deleted allocas from various lists so that we don't try to
3740 // continue processing them.
3741 if (!DeletedAllocas.empty()) {
3742 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3743 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3744 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3745 PromotableAllocas.end(),
3746 IsAllocaInSet(DeletedAllocas)),
3747 PromotableAllocas.end());
3748 DeletedAllocas.clear();
3752 Changed |= promoteAllocas(F);
3754 Worklist = PostPromotionWorklist;
3755 PostPromotionWorklist.clear();
3756 } while (!Worklist.empty());
3761 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3762 if (RequiresDomTree)
3763 AU.addRequired<DominatorTree>();
3764 AU.setPreservesCFG();