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/Constants.h"
29 #include "llvm/DIBuilder.h"
30 #include "llvm/DebugInfo.h"
31 #include "llvm/DerivedTypes.h"
32 #include "llvm/Function.h"
33 #include "llvm/IRBuilder.h"
34 #include "llvm/Instructions.h"
35 #include "llvm/IntrinsicInst.h"
36 #include "llvm/LLVMContext.h"
37 #include "llvm/Module.h"
38 #include "llvm/Operator.h"
39 #include "llvm/Pass.h"
40 #include "llvm/ADT/SetVector.h"
41 #include "llvm/ADT/SmallVector.h"
42 #include "llvm/ADT/Statistic.h"
43 #include "llvm/ADT/STLExtras.h"
44 #include "llvm/Analysis/Dominators.h"
45 #include "llvm/Analysis/Loads.h"
46 #include "llvm/Analysis/ValueTracking.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/GetElementPtrTypeIterator.h"
51 #include "llvm/Support/InstVisitor.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include "llvm/DataLayout.h"
55 #include "llvm/Transforms/Utils/Local.h"
56 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
57 #include "llvm/Transforms/Utils/SSAUpdater.h"
60 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
61 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
62 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
63 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
64 STATISTIC(NumDeleted, "Number of instructions deleted");
65 STATISTIC(NumVectorized, "Number of vectorized aggregates");
67 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
68 /// forming SSA values through the SSAUpdater infrastructure.
70 ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
73 /// \brief Alloca partitioning representation.
75 /// This class represents a partitioning of an alloca into slices, and
76 /// information about the nature of uses of each slice of the alloca. The goal
77 /// is that this information is sufficient to decide if and how to split the
78 /// alloca apart and replace slices with scalars. It is also intended that this
79 /// structure can capture the relevant information needed both to decide about
80 /// and to enact these transformations.
81 class AllocaPartitioning {
83 /// \brief A common base class for representing a half-open byte range.
85 /// \brief The beginning offset of the range.
88 /// \brief The ending offset, not included in the range.
91 ByteRange() : BeginOffset(), EndOffset() {}
92 ByteRange(uint64_t BeginOffset, uint64_t EndOffset)
93 : BeginOffset(BeginOffset), EndOffset(EndOffset) {}
95 /// \brief Support for ordering ranges.
97 /// This provides an ordering over ranges such that start offsets are
98 /// always increasing, and within equal start offsets, the end offsets are
99 /// decreasing. Thus the spanning range comes first in a cluster with the
100 /// same start position.
101 bool operator<(const ByteRange &RHS) const {
102 if (BeginOffset < RHS.BeginOffset) return true;
103 if (BeginOffset > RHS.BeginOffset) return false;
104 if (EndOffset > RHS.EndOffset) return true;
108 /// \brief Support comparison with a single offset to allow binary searches.
109 friend bool operator<(const ByteRange &LHS, uint64_t RHSOffset) {
110 return LHS.BeginOffset < RHSOffset;
113 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
114 const ByteRange &RHS) {
115 return LHSOffset < RHS.BeginOffset;
118 bool operator==(const ByteRange &RHS) const {
119 return BeginOffset == RHS.BeginOffset && EndOffset == RHS.EndOffset;
121 bool operator!=(const ByteRange &RHS) const { return !operator==(RHS); }
124 /// \brief A partition of an alloca.
126 /// This structure represents a contiguous partition of the alloca. These are
127 /// formed by examining the uses of the alloca. During formation, they may
128 /// overlap but once an AllocaPartitioning is built, the Partitions within it
129 /// are all disjoint.
130 struct Partition : public ByteRange {
131 /// \brief Whether this partition is splittable into smaller partitions.
133 /// We flag partitions as splittable when they are formed entirely due to
134 /// accesses by trivially splittable operations such as memset and memcpy.
136 /// FIXME: At some point we should consider loads and stores of FCAs to be
137 /// splittable and eagerly split them into scalar values.
140 /// \brief Test whether a partition has been marked as dead.
141 bool isDead() const {
142 if (BeginOffset == UINT64_MAX) {
143 assert(EndOffset == UINT64_MAX);
149 /// \brief Kill a partition.
150 /// This is accomplished by setting both its beginning and end offset to
151 /// the maximum possible value.
153 assert(!isDead() && "He's Dead, Jim!");
154 BeginOffset = EndOffset = UINT64_MAX;
157 Partition() : ByteRange(), IsSplittable() {}
158 Partition(uint64_t BeginOffset, uint64_t EndOffset, bool IsSplittable)
159 : ByteRange(BeginOffset, EndOffset), IsSplittable(IsSplittable) {}
162 /// \brief A particular use of a partition of the alloca.
164 /// This structure is used to associate uses of a partition with it. They
165 /// mark the range of bytes which are referenced by a particular instruction,
166 /// and includes a handle to the user itself and the pointer value in use.
167 /// The bounds of these uses are determined by intersecting the bounds of the
168 /// memory use itself with a particular partition. As a consequence there is
169 /// intentionally overlap between various uses of the same partition.
170 struct PartitionUse : public ByteRange {
171 /// \brief The use in question. Provides access to both user and used value.
173 /// Note that this may be null if the partition use is *dead*, that is, it
174 /// should be ignored.
177 PartitionUse() : ByteRange(), U() {}
178 PartitionUse(uint64_t BeginOffset, uint64_t EndOffset, Use *U)
179 : ByteRange(BeginOffset, EndOffset), U(U) {}
182 /// \brief Construct a partitioning of a particular alloca.
184 /// Construction does most of the work for partitioning the alloca. This
185 /// performs the necessary walks of users and builds a partitioning from it.
186 AllocaPartitioning(const DataLayout &TD, AllocaInst &AI);
188 /// \brief Test whether a pointer to the allocation escapes our analysis.
190 /// If this is true, the partitioning is never fully built and should be
192 bool isEscaped() const { return PointerEscapingInstr; }
194 /// \brief Support for iterating over the partitions.
196 typedef SmallVectorImpl<Partition>::iterator iterator;
197 iterator begin() { return Partitions.begin(); }
198 iterator end() { return Partitions.end(); }
200 typedef SmallVectorImpl<Partition>::const_iterator const_iterator;
201 const_iterator begin() const { return Partitions.begin(); }
202 const_iterator end() const { return Partitions.end(); }
205 /// \brief Support for iterating over and manipulating a particular
206 /// partition's uses.
208 /// The iteration support provided for uses is more limited, but also
209 /// includes some manipulation routines to support rewriting the uses of
210 /// partitions during SROA.
212 typedef SmallVectorImpl<PartitionUse>::iterator use_iterator;
213 use_iterator use_begin(unsigned Idx) { return Uses[Idx].begin(); }
214 use_iterator use_begin(const_iterator I) { return Uses[I - begin()].begin(); }
215 use_iterator use_end(unsigned Idx) { return Uses[Idx].end(); }
216 use_iterator use_end(const_iterator I) { return Uses[I - begin()].end(); }
218 typedef SmallVectorImpl<PartitionUse>::const_iterator const_use_iterator;
219 const_use_iterator use_begin(unsigned Idx) const { return Uses[Idx].begin(); }
220 const_use_iterator use_begin(const_iterator I) const {
221 return Uses[I - begin()].begin();
223 const_use_iterator use_end(unsigned Idx) const { return Uses[Idx].end(); }
224 const_use_iterator use_end(const_iterator I) const {
225 return Uses[I - begin()].end();
228 unsigned use_size(unsigned Idx) const { return Uses[Idx].size(); }
229 unsigned use_size(const_iterator I) const { return Uses[I - begin()].size(); }
230 const PartitionUse &getUse(unsigned PIdx, unsigned UIdx) const {
231 return Uses[PIdx][UIdx];
233 const PartitionUse &getUse(const_iterator I, unsigned UIdx) const {
234 return Uses[I - begin()][UIdx];
237 void use_push_back(unsigned Idx, const PartitionUse &PU) {
238 Uses[Idx].push_back(PU);
240 void use_push_back(const_iterator I, const PartitionUse &PU) {
241 Uses[I - begin()].push_back(PU);
245 /// \brief Allow iterating the dead users for this alloca.
247 /// These are instructions which will never actually use the alloca as they
248 /// are outside the allocated range. They are safe to replace with undef and
251 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
252 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
253 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
256 /// \brief Allow iterating the dead expressions referring to this alloca.
258 /// These are operands which have cannot actually be used to refer to the
259 /// alloca as they are outside its range and the user doesn't correct for
260 /// that. These mostly consist of PHI node inputs and the like which we just
261 /// need to replace with undef.
263 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
264 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
265 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
268 /// \brief MemTransferInst auxiliary data.
269 /// This struct provides some auxiliary data about memory transfer
270 /// intrinsics such as memcpy and memmove. These intrinsics can use two
271 /// different ranges within the same alloca, and provide other challenges to
272 /// correctly represent. We stash extra data to help us untangle this
273 /// after the partitioning is complete.
274 struct MemTransferOffsets {
275 /// The destination begin and end offsets when the destination is within
276 /// this alloca. If the end offset is zero the destination is not within
278 uint64_t DestBegin, DestEnd;
280 /// The source begin and end offsets when the source is within this alloca.
281 /// If the end offset is zero, the source is not within this alloca.
282 uint64_t SourceBegin, SourceEnd;
284 /// Flag for whether an alloca is splittable.
287 MemTransferOffsets getMemTransferOffsets(MemTransferInst &II) const {
288 return MemTransferInstData.lookup(&II);
291 /// \brief Map from a PHI or select operand back to a partition.
293 /// When manipulating PHI nodes or selects, they can use more than one
294 /// partition of an alloca. We store a special mapping to allow finding the
295 /// partition referenced by each of these operands, if any.
296 iterator findPartitionForPHIOrSelectOperand(Use *U) {
297 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
298 = PHIOrSelectOpMap.find(U);
299 if (MapIt == PHIOrSelectOpMap.end())
302 return begin() + MapIt->second.first;
305 /// \brief Map from a PHI or select operand back to the specific use of
308 /// Similar to mapping these operands back to the partitions, this maps
309 /// directly to the use structure of that partition.
310 use_iterator findPartitionUseForPHIOrSelectOperand(Use *U) {
311 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
312 = PHIOrSelectOpMap.find(U);
313 assert(MapIt != PHIOrSelectOpMap.end());
314 return Uses[MapIt->second.first].begin() + MapIt->second.second;
317 /// \brief Compute a common type among the uses of a particular partition.
319 /// This routines walks all of the uses of a particular partition and tries
320 /// to find a common type between them. Untyped operations such as memset and
321 /// memcpy are ignored.
322 Type *getCommonType(iterator I) const;
324 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
325 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
326 void printUsers(raw_ostream &OS, const_iterator I,
327 StringRef Indent = " ") const;
328 void print(raw_ostream &OS) const;
329 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const;
330 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const;
334 template <typename DerivedT, typename RetT = void> class BuilderBase;
335 class PartitionBuilder;
336 friend class AllocaPartitioning::PartitionBuilder;
338 friend class AllocaPartitioning::UseBuilder;
341 /// \brief Handle to alloca instruction to simplify method interfaces.
345 /// \brief The instruction responsible for this alloca having no partitioning.
347 /// When an instruction (potentially) escapes the pointer to the alloca, we
348 /// store a pointer to that here and abort trying to partition the alloca.
349 /// This will be null if the alloca is partitioned successfully.
350 Instruction *PointerEscapingInstr;
352 /// \brief The partitions of the alloca.
354 /// We store a vector of the partitions over the alloca here. This vector is
355 /// sorted by increasing begin offset, and then by decreasing end offset. See
356 /// the Partition inner class for more details. Initially (during
357 /// construction) there are overlaps, but we form a disjoint sequence of
358 /// partitions while finishing construction and a fully constructed object is
359 /// expected to always have this as a disjoint space.
360 SmallVector<Partition, 8> Partitions;
362 /// \brief The uses of the partitions.
364 /// This is essentially a mapping from each partition to a list of uses of
365 /// that partition. The mapping is done with a Uses vector that has the exact
366 /// same number of entries as the partition vector. Each entry is itself
367 /// a vector of the uses.
368 SmallVector<SmallVector<PartitionUse, 2>, 8> Uses;
370 /// \brief Instructions which will become dead if we rewrite the alloca.
372 /// Note that these are not separated by partition. This is because we expect
373 /// a partitioned alloca to be completely rewritten or not rewritten at all.
374 /// If rewritten, all these instructions can simply be removed and replaced
375 /// with undef as they come from outside of the allocated space.
376 SmallVector<Instruction *, 8> DeadUsers;
378 /// \brief Operands which will become dead if we rewrite the alloca.
380 /// These are operands that in their particular use can be replaced with
381 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
382 /// to PHI nodes and the like. They aren't entirely dead (there might be
383 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
384 /// want to swap this particular input for undef to simplify the use lists of
386 SmallVector<Use *, 8> DeadOperands;
388 /// \brief The underlying storage for auxiliary memcpy and memset info.
389 SmallDenseMap<MemTransferInst *, MemTransferOffsets, 4> MemTransferInstData;
391 /// \brief A side datastructure used when building up the partitions and uses.
393 /// This mapping is only really used during the initial building of the
394 /// partitioning so that we can retain information about PHI and select nodes
396 SmallDenseMap<Instruction *, std::pair<uint64_t, bool> > PHIOrSelectSizes;
398 /// \brief Auxiliary information for particular PHI or select operands.
399 SmallDenseMap<Use *, std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap;
401 /// \brief A utility routine called from the constructor.
403 /// This does what it says on the tin. It is the key of the alloca partition
404 /// splitting and merging. After it is called we have the desired disjoint
405 /// collection of partitions.
406 void splitAndMergePartitions();
410 template <typename DerivedT, typename RetT>
411 class AllocaPartitioning::BuilderBase
412 : public InstVisitor<DerivedT, RetT> {
414 BuilderBase(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
416 AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())),
422 const DataLayout &TD;
423 const uint64_t AllocSize;
424 AllocaPartitioning &P;
426 SmallPtrSet<Use *, 8> VisitedUses;
432 SmallVector<OffsetUse, 8> Queue;
434 // The active offset and use while visiting.
438 void enqueueUsers(Instruction &I, int64_t UserOffset) {
439 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
441 if (VisitedUses.insert(&UI.getUse())) {
442 OffsetUse OU = { &UI.getUse(), UserOffset };
448 bool computeConstantGEPOffset(GetElementPtrInst &GEPI, int64_t &GEPOffset) {
450 unsigned int AS = GEPI.getPointerAddressSpace();
451 for (gep_type_iterator GTI = gep_type_begin(GEPI), GTE = gep_type_end(GEPI);
453 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
459 // Handle a struct index, which adds its field offset to the pointer.
460 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
461 unsigned ElementIdx = OpC->getZExtValue();
462 const StructLayout *SL = TD.getStructLayout(STy);
463 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
464 // Check that we can continue to model this GEP in a signed 64-bit offset.
465 if (ElementOffset > INT64_MAX ||
467 ((uint64_t)GEPOffset + ElementOffset) > INT64_MAX)) {
468 DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding "
469 << "what can be represented in an int64_t!\n"
470 << " alloca: " << P.AI << "\n");
474 GEPOffset = ElementOffset + (uint64_t)-GEPOffset;
476 GEPOffset += ElementOffset;
480 APInt Index = OpC->getValue().sextOrTrunc(TD.getPointerSizeInBits(AS));
481 Index *= APInt(Index.getBitWidth(),
482 TD.getTypeAllocSize(GTI.getIndexedType()));
483 Index += APInt(Index.getBitWidth(), (uint64_t)GEPOffset,
485 // Check if the result can be stored in our int64_t offset.
486 if (!Index.isSignedIntN(sizeof(GEPOffset) * 8)) {
487 DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding "
488 << "what can be represented in an int64_t!\n"
489 << " alloca: " << P.AI << "\n");
493 GEPOffset = Index.getSExtValue();
498 Value *foldSelectInst(SelectInst &SI) {
499 // If the condition being selected on is a constant or the same value is
500 // being selected between, fold the select. Yes this does (rarely) happen
502 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
503 return SI.getOperand(1+CI->isZero());
504 if (SI.getOperand(1) == SI.getOperand(2)) {
505 assert(*U == SI.getOperand(1));
506 return SI.getOperand(1);
512 /// \brief Builder for the alloca partitioning.
514 /// This class builds an alloca partitioning by recursively visiting the uses
515 /// of an alloca and splitting the partitions for each load and store at each
517 class AllocaPartitioning::PartitionBuilder
518 : public BuilderBase<PartitionBuilder, bool> {
519 friend class InstVisitor<PartitionBuilder, bool>;
521 SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap;
524 PartitionBuilder(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
525 : BuilderBase<PartitionBuilder, bool>(TD, AI, P) {}
527 /// \brief Run the builder over the allocation.
529 // Note that we have to re-evaluate size on each trip through the loop as
530 // the queue grows at the tail.
531 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
533 Offset = Queue[Idx].Offset;
534 if (!visit(cast<Instruction>(U->getUser())))
541 bool markAsEscaping(Instruction &I) {
542 P.PointerEscapingInstr = &I;
546 void insertUse(Instruction &I, int64_t Offset, uint64_t Size,
547 bool IsSplittable = false) {
548 // Completely skip uses which have a zero size or don't overlap the
551 (Offset >= 0 && (uint64_t)Offset >= AllocSize) ||
552 (Offset < 0 && (uint64_t)-Offset >= Size)) {
553 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
554 << " which starts past the end of the " << AllocSize
556 << " alloca: " << P.AI << "\n"
557 << " use: " << I << "\n");
561 // Clamp the start to the beginning of the allocation.
563 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
564 << " to start at the beginning of the alloca:\n"
565 << " alloca: " << P.AI << "\n"
566 << " use: " << I << "\n");
567 Size -= (uint64_t)-Offset;
571 uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
573 // Clamp the end offset to the end of the allocation. Note that this is
574 // formulated to handle even the case where "BeginOffset + Size" overflows.
575 assert(AllocSize >= BeginOffset); // Established above.
576 if (Size > AllocSize - BeginOffset) {
577 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
578 << " to remain within the " << AllocSize << " byte alloca:\n"
579 << " alloca: " << P.AI << "\n"
580 << " use: " << I << "\n");
581 EndOffset = AllocSize;
584 Partition New(BeginOffset, EndOffset, IsSplittable);
585 P.Partitions.push_back(New);
588 bool handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) {
589 uint64_t Size = TD.getTypeStoreSize(Ty);
591 // If this memory access can be shown to *statically* extend outside the
592 // bounds of of the allocation, it's behavior is undefined, so simply
593 // ignore it. Note that this is more strict than the generic clamping
594 // behavior of insertUse. We also try to handle cases which might run the
596 // FIXME: We should instead consider the pointer to have escaped if this
597 // function is being instrumented for addressing bugs or race conditions.
598 if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
599 Size > (AllocSize - (uint64_t)Offset)) {
600 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte "
601 << (isa<LoadInst>(I) ? "load" : "store") << " @" << Offset
602 << " which extends past the end of the " << AllocSize
604 << " alloca: " << P.AI << "\n"
605 << " use: " << I << "\n");
609 insertUse(I, Offset, Size);
613 bool visitBitCastInst(BitCastInst &BC) {
614 enqueueUsers(BC, Offset);
618 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
620 if (!computeConstantGEPOffset(GEPI, GEPOffset))
621 return markAsEscaping(GEPI);
623 enqueueUsers(GEPI, GEPOffset);
627 bool visitLoadInst(LoadInst &LI) {
628 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
629 "All simple FCA loads should have been pre-split");
630 return handleLoadOrStore(LI.getType(), LI, Offset);
633 bool visitStoreInst(StoreInst &SI) {
634 Value *ValOp = SI.getValueOperand();
636 return markAsEscaping(SI);
638 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
639 "All simple FCA stores should have been pre-split");
640 return handleLoadOrStore(ValOp->getType(), SI, Offset);
644 bool visitMemSetInst(MemSetInst &II) {
645 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
646 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
647 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
648 insertUse(II, Offset, Size, Length);
652 bool visitMemTransferInst(MemTransferInst &II) {
653 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
654 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
656 // Zero-length mem transfer intrinsics can be ignored entirely.
659 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
661 // Only intrinsics with a constant length can be split.
662 Offsets.IsSplittable = Length;
664 if (*U == II.getRawDest()) {
665 Offsets.DestBegin = Offset;
666 Offsets.DestEnd = Offset + Size;
668 if (*U == II.getRawSource()) {
669 Offsets.SourceBegin = Offset;
670 Offsets.SourceEnd = Offset + Size;
673 // If we have set up end offsets for both the source and the destination,
674 // we have found both sides of this transfer pointing at the same alloca.
675 bool SeenBothEnds = Offsets.SourceEnd && Offsets.DestEnd;
676 if (SeenBothEnds && II.getRawDest() != II.getRawSource()) {
677 unsigned PrevIdx = MemTransferPartitionMap[&II];
679 // Check if the begin offsets match and this is a non-volatile transfer.
680 // In that case, we can completely elide the transfer.
681 if (!II.isVolatile() && Offsets.SourceBegin == Offsets.DestBegin) {
682 P.Partitions[PrevIdx].kill();
686 // Otherwise we have an offset transfer within the same alloca. We can't
688 P.Partitions[PrevIdx].IsSplittable = Offsets.IsSplittable = false;
689 } else if (SeenBothEnds) {
690 // Handle the case where this exact use provides both ends of the
692 assert(II.getRawDest() == II.getRawSource());
694 // For non-volatile transfers this is a no-op.
695 if (!II.isVolatile())
698 // Otherwise just suppress splitting.
699 Offsets.IsSplittable = false;
703 // Insert the use now that we've fixed up the splittable nature.
704 insertUse(II, Offset, Size, Offsets.IsSplittable);
706 // Setup the mapping from intrinsic to partition of we've not seen both
707 // ends of this transfer.
709 unsigned NewIdx = P.Partitions.size() - 1;
711 = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx)).second;
713 "Already have intrinsic in map but haven't seen both ends");
720 // Disable SRoA for any intrinsics except for lifetime invariants.
721 // FIXME: What about debug instrinsics? This matches old behavior, but
722 // doesn't make sense.
723 bool visitIntrinsicInst(IntrinsicInst &II) {
724 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
725 II.getIntrinsicID() == Intrinsic::lifetime_end) {
726 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
727 uint64_t Size = std::min(AllocSize - Offset, Length->getLimitedValue());
728 insertUse(II, Offset, Size, true);
732 return markAsEscaping(II);
735 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
736 // We consider any PHI or select that results in a direct load or store of
737 // the same offset to be a viable use for partitioning purposes. These uses
738 // are considered unsplittable and the size is the maximum loaded or stored
740 SmallPtrSet<Instruction *, 4> Visited;
741 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
742 Visited.insert(Root);
743 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
744 // If there are no loads or stores, the access is dead. We mark that as
745 // a size zero access.
748 Instruction *I, *UsedI;
749 llvm::tie(UsedI, I) = Uses.pop_back_val();
751 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
752 Size = std::max(Size, TD.getTypeStoreSize(LI->getType()));
755 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
756 Value *Op = SI->getOperand(0);
759 Size = std::max(Size, TD.getTypeStoreSize(Op->getType()));
763 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
764 if (!GEP->hasAllZeroIndices())
766 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
767 !isa<SelectInst>(I)) {
771 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
773 if (Visited.insert(cast<Instruction>(*UI)))
774 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
775 } while (!Uses.empty());
780 bool visitPHINode(PHINode &PN) {
781 // See if we already have computed info on this node.
782 std::pair<uint64_t, bool> &PHIInfo = P.PHIOrSelectSizes[&PN];
784 PHIInfo.second = true;
785 insertUse(PN, Offset, PHIInfo.first);
789 // Check for an unsafe use of the PHI node.
790 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&PN, PHIInfo.first))
791 return markAsEscaping(*EscapingI);
793 insertUse(PN, Offset, PHIInfo.first);
797 bool visitSelectInst(SelectInst &SI) {
798 if (Value *Result = foldSelectInst(SI)) {
800 // If the result of the constant fold will be the pointer, recurse
801 // through the select as if we had RAUW'ed it.
802 enqueueUsers(SI, Offset);
807 // See if we already have computed info on this node.
808 std::pair<uint64_t, bool> &SelectInfo = P.PHIOrSelectSizes[&SI];
809 if (SelectInfo.first) {
810 SelectInfo.second = true;
811 insertUse(SI, Offset, SelectInfo.first);
815 // Check for an unsafe use of the PHI node.
816 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&SI, SelectInfo.first))
817 return markAsEscaping(*EscapingI);
819 insertUse(SI, Offset, SelectInfo.first);
823 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
824 bool visitInstruction(Instruction &I) { return markAsEscaping(I); }
828 /// \brief Use adder for the alloca partitioning.
830 /// This class adds the uses of an alloca to all of the partitions which they
831 /// use. For splittable partitions, this can end up doing essentially a linear
832 /// walk of the partitions, but the number of steps remains bounded by the
833 /// total result instruction size:
834 /// - The number of partitions is a result of the number unsplittable
835 /// instructions using the alloca.
836 /// - The number of users of each partition is at worst the total number of
837 /// splittable instructions using the alloca.
838 /// Thus we will produce N * M instructions in the end, where N are the number
839 /// of unsplittable uses and M are the number of splittable. This visitor does
840 /// the exact same number of updates to the partitioning.
842 /// In the more common case, this visitor will leverage the fact that the
843 /// partition space is pre-sorted, and do a logarithmic search for the
844 /// partition needed, making the total visit a classical ((N + M) * log(N))
845 /// complexity operation.
846 class AllocaPartitioning::UseBuilder : public BuilderBase<UseBuilder> {
847 friend class InstVisitor<UseBuilder>;
849 /// \brief Set to de-duplicate dead instructions found in the use walk.
850 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
853 UseBuilder(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
854 : BuilderBase<UseBuilder>(TD, AI, P) {}
856 /// \brief Run the builder over the allocation.
858 // Note that we have to re-evaluate size on each trip through the loop as
859 // the queue grows at the tail.
860 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
862 Offset = Queue[Idx].Offset;
863 this->visit(cast<Instruction>(U->getUser()));
868 void markAsDead(Instruction &I) {
869 if (VisitedDeadInsts.insert(&I))
870 P.DeadUsers.push_back(&I);
873 void insertUse(Instruction &User, int64_t Offset, uint64_t Size) {
874 // If the use has a zero size or extends outside of the allocation, record
875 // it as a dead use for elimination later.
876 if (Size == 0 || (uint64_t)Offset >= AllocSize ||
877 (Offset < 0 && (uint64_t)-Offset >= Size))
878 return markAsDead(User);
880 // Clamp the start to the beginning of the allocation.
882 Size -= (uint64_t)-Offset;
886 uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
888 // Clamp the end offset to the end of the allocation. Note that this is
889 // formulated to handle even the case where "BeginOffset + Size" overflows.
890 assert(AllocSize >= BeginOffset); // Established above.
891 if (Size > AllocSize - BeginOffset)
892 EndOffset = AllocSize;
894 // NB: This only works if we have zero overlapping partitions.
895 iterator B = std::lower_bound(P.begin(), P.end(), BeginOffset);
896 if (B != P.begin() && llvm::prior(B)->EndOffset > BeginOffset)
898 for (iterator I = B, E = P.end(); I != E && I->BeginOffset < EndOffset;
900 PartitionUse NewPU(std::max(I->BeginOffset, BeginOffset),
901 std::min(I->EndOffset, EndOffset), U);
902 P.use_push_back(I, NewPU);
903 if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser()))
904 P.PHIOrSelectOpMap[U]
905 = std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1);
909 void handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) {
910 uint64_t Size = TD.getTypeStoreSize(Ty);
912 // If this memory access can be shown to *statically* extend outside the
913 // bounds of of the allocation, it's behavior is undefined, so simply
914 // ignore it. Note that this is more strict than the generic clamping
915 // behavior of insertUse.
916 if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
917 Size > (AllocSize - (uint64_t)Offset))
918 return markAsDead(I);
920 insertUse(I, Offset, Size);
923 void visitBitCastInst(BitCastInst &BC) {
925 return markAsDead(BC);
927 enqueueUsers(BC, Offset);
930 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
931 if (GEPI.use_empty())
932 return markAsDead(GEPI);
935 if (!computeConstantGEPOffset(GEPI, GEPOffset))
936 llvm_unreachable("Unable to compute constant offset for use");
938 enqueueUsers(GEPI, GEPOffset);
941 void visitLoadInst(LoadInst &LI) {
942 handleLoadOrStore(LI.getType(), LI, Offset);
945 void visitStoreInst(StoreInst &SI) {
946 handleLoadOrStore(SI.getOperand(0)->getType(), SI, Offset);
949 void visitMemSetInst(MemSetInst &II) {
950 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
951 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
952 insertUse(II, Offset, Size);
955 void visitMemTransferInst(MemTransferInst &II) {
956 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
957 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
959 return markAsDead(II);
961 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
962 if (!II.isVolatile() && Offsets.DestEnd && Offsets.SourceEnd &&
963 Offsets.DestBegin == Offsets.SourceBegin)
964 return markAsDead(II); // Skip identity transfers without side-effects.
966 insertUse(II, Offset, Size);
969 void visitIntrinsicInst(IntrinsicInst &II) {
970 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
971 II.getIntrinsicID() == Intrinsic::lifetime_end);
973 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
974 insertUse(II, Offset,
975 std::min(AllocSize - Offset, Length->getLimitedValue()));
978 void insertPHIOrSelect(Instruction &User, uint64_t Offset) {
979 uint64_t Size = P.PHIOrSelectSizes.lookup(&User).first;
981 // For PHI and select operands outside the alloca, we can't nuke the entire
982 // phi or select -- the other side might still be relevant, so we special
983 // case them here and use a separate structure to track the operands
984 // themselves which should be replaced with undef.
985 if (Offset >= AllocSize) {
986 P.DeadOperands.push_back(U);
990 insertUse(User, Offset, Size);
992 void visitPHINode(PHINode &PN) {
994 return markAsDead(PN);
996 insertPHIOrSelect(PN, Offset);
998 void visitSelectInst(SelectInst &SI) {
1000 return markAsDead(SI);
1002 if (Value *Result = foldSelectInst(SI)) {
1004 // If the result of the constant fold will be the pointer, recurse
1005 // through the select as if we had RAUW'ed it.
1006 enqueueUsers(SI, Offset);
1008 // Otherwise the operand to the select is dead, and we can replace it
1010 P.DeadOperands.push_back(U);
1015 insertPHIOrSelect(SI, Offset);
1018 /// \brief Unreachable, we've already visited the alloca once.
1019 void visitInstruction(Instruction &I) {
1020 llvm_unreachable("Unhandled instruction in use builder.");
1024 void AllocaPartitioning::splitAndMergePartitions() {
1025 size_t NumDeadPartitions = 0;
1027 // Track the range of splittable partitions that we pass when accumulating
1028 // overlapping unsplittable partitions.
1029 uint64_t SplitEndOffset = 0ull;
1031 Partition New(0ull, 0ull, false);
1033 for (unsigned i = 0, j = i, e = Partitions.size(); i != e; i = j) {
1036 if (!Partitions[i].IsSplittable || New.BeginOffset == New.EndOffset) {
1037 assert(New.BeginOffset == New.EndOffset);
1038 New = Partitions[i];
1040 assert(New.IsSplittable);
1041 New.EndOffset = std::max(New.EndOffset, Partitions[i].EndOffset);
1043 assert(New.BeginOffset != New.EndOffset);
1045 // Scan the overlapping partitions.
1046 while (j != e && New.EndOffset > Partitions[j].BeginOffset) {
1047 // If the new partition we are forming is splittable, stop at the first
1048 // unsplittable partition.
1049 if (New.IsSplittable && !Partitions[j].IsSplittable)
1052 // Grow the new partition to include any equally splittable range. 'j' is
1053 // always equally splittable when New is splittable, but when New is not
1054 // splittable, we may subsume some (or part of some) splitable partition
1055 // without growing the new one.
1056 if (New.IsSplittable == Partitions[j].IsSplittable) {
1057 New.EndOffset = std::max(New.EndOffset, Partitions[j].EndOffset);
1059 assert(!New.IsSplittable);
1060 assert(Partitions[j].IsSplittable);
1061 SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset);
1064 Partitions[j].kill();
1065 ++NumDeadPartitions;
1069 // If the new partition is splittable, chop off the end as soon as the
1070 // unsplittable subsequent partition starts and ensure we eventually cover
1071 // the splittable area.
1072 if (j != e && New.IsSplittable) {
1073 SplitEndOffset = std::max(SplitEndOffset, New.EndOffset);
1074 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1077 // Add the new partition if it differs from the original one and is
1078 // non-empty. We can end up with an empty partition here if it was
1079 // splittable but there is an unsplittable one that starts at the same
1081 if (New != Partitions[i]) {
1082 if (New.BeginOffset != New.EndOffset)
1083 Partitions.push_back(New);
1084 // Mark the old one for removal.
1085 Partitions[i].kill();
1086 ++NumDeadPartitions;
1089 New.BeginOffset = New.EndOffset;
1090 if (!New.IsSplittable) {
1091 New.EndOffset = std::max(New.EndOffset, SplitEndOffset);
1092 if (j != e && !Partitions[j].IsSplittable)
1093 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1094 New.IsSplittable = true;
1095 // If there is a trailing splittable partition which won't be fused into
1096 // the next splittable partition go ahead and add it onto the partitions
1098 if (New.BeginOffset < New.EndOffset &&
1099 (j == e || !Partitions[j].IsSplittable ||
1100 New.EndOffset < Partitions[j].BeginOffset)) {
1101 Partitions.push_back(New);
1102 New.BeginOffset = New.EndOffset = 0ull;
1107 // Re-sort the partitions now that they have been split and merged into
1108 // disjoint set of partitions. Also remove any of the dead partitions we've
1109 // replaced in the process.
1110 std::sort(Partitions.begin(), Partitions.end());
1111 if (NumDeadPartitions) {
1112 assert(Partitions.back().isDead());
1113 assert((ptrdiff_t)NumDeadPartitions ==
1114 std::count(Partitions.begin(), Partitions.end(), Partitions.back()));
1116 Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end());
1119 AllocaPartitioning::AllocaPartitioning(const DataLayout &TD, AllocaInst &AI)
1124 PointerEscapingInstr(0) {
1125 PartitionBuilder PB(TD, AI, *this);
1129 // Sort the uses. This arranges for the offsets to be in ascending order,
1130 // and the sizes to be in descending order.
1131 std::sort(Partitions.begin(), Partitions.end());
1133 // Remove any partitions from the back which are marked as dead.
1134 while (!Partitions.empty() && Partitions.back().isDead())
1135 Partitions.pop_back();
1137 if (Partitions.size() > 1) {
1138 // Intersect splittability for all partitions with equal offsets and sizes.
1139 // Then remove all but the first so that we have a sequence of non-equal but
1140 // potentially overlapping partitions.
1141 for (iterator I = Partitions.begin(), J = I, E = Partitions.end(); I != E;
1144 while (J != E && *I == *J) {
1145 I->IsSplittable &= J->IsSplittable;
1149 Partitions.erase(std::unique(Partitions.begin(), Partitions.end()),
1152 // Split splittable and merge unsplittable partitions into a disjoint set
1153 // of partitions over the used space of the allocation.
1154 splitAndMergePartitions();
1157 // Now build up the user lists for each of these disjoint partitions by
1158 // re-walking the recursive users of the alloca.
1159 Uses.resize(Partitions.size());
1160 UseBuilder UB(TD, AI, *this);
1164 Type *AllocaPartitioning::getCommonType(iterator I) const {
1166 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
1168 continue; // Skip dead uses.
1169 if (isa<IntrinsicInst>(*UI->U->getUser()))
1171 if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset)
1175 if (LoadInst *LI = dyn_cast<LoadInst>(UI->U->getUser())) {
1176 UserTy = LI->getType();
1177 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI->U->getUser())) {
1178 UserTy = SI->getValueOperand()->getType();
1181 if (Ty && Ty != UserTy)
1189 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1191 void AllocaPartitioning::print(raw_ostream &OS, const_iterator I,
1192 StringRef Indent) const {
1193 OS << Indent << "partition #" << (I - begin())
1194 << " [" << I->BeginOffset << "," << I->EndOffset << ")"
1195 << (I->IsSplittable ? " (splittable)" : "")
1196 << (Uses[I - begin()].empty() ? " (zero uses)" : "")
1200 void AllocaPartitioning::printUsers(raw_ostream &OS, const_iterator I,
1201 StringRef Indent) const {
1202 for (const_use_iterator UI = use_begin(I), UE = use_end(I);
1205 continue; // Skip dead uses.
1206 OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") "
1207 << "used by: " << *UI->U->getUser() << "\n";
1208 if (MemTransferInst *II = dyn_cast<MemTransferInst>(UI->U->getUser())) {
1209 const MemTransferOffsets &MTO = MemTransferInstData.lookup(II);
1211 if (!MTO.IsSplittable)
1212 IsDest = UI->BeginOffset == MTO.DestBegin;
1214 IsDest = MTO.DestBegin != 0u;
1215 OS << Indent << " (original " << (IsDest ? "dest" : "source") << ": "
1216 << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin)
1217 << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n";
1222 void AllocaPartitioning::print(raw_ostream &OS) const {
1223 if (PointerEscapingInstr) {
1224 OS << "No partitioning for alloca: " << AI << "\n"
1225 << " A pointer to this alloca escaped by:\n"
1226 << " " << *PointerEscapingInstr << "\n";
1230 OS << "Partitioning of alloca: " << AI << "\n";
1232 for (const_iterator I = begin(), E = end(); I != E; ++I, ++Num) {
1238 void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); }
1239 void AllocaPartitioning::dump() const { print(dbgs()); }
1241 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1245 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1247 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1248 /// the loads and stores of an alloca instruction, as well as updating its
1249 /// debug information. This is used when a domtree is unavailable and thus
1250 /// mem2reg in its full form can't be used to handle promotion of allocas to
1252 class AllocaPromoter : public LoadAndStorePromoter {
1256 SmallVector<DbgDeclareInst *, 4> DDIs;
1257 SmallVector<DbgValueInst *, 4> DVIs;
1260 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1261 AllocaInst &AI, DIBuilder &DIB)
1262 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1264 void run(const SmallVectorImpl<Instruction*> &Insts) {
1265 // Remember which alloca we're promoting (for isInstInList).
1266 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
1267 for (Value::use_iterator UI = DebugNode->use_begin(),
1268 UE = DebugNode->use_end();
1270 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
1271 DDIs.push_back(DDI);
1272 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
1273 DVIs.push_back(DVI);
1276 LoadAndStorePromoter::run(Insts);
1277 AI.eraseFromParent();
1278 while (!DDIs.empty())
1279 DDIs.pop_back_val()->eraseFromParent();
1280 while (!DVIs.empty())
1281 DVIs.pop_back_val()->eraseFromParent();
1284 virtual bool isInstInList(Instruction *I,
1285 const SmallVectorImpl<Instruction*> &Insts) const {
1286 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1287 return LI->getOperand(0) == &AI;
1288 return cast<StoreInst>(I)->getPointerOperand() == &AI;
1291 virtual void updateDebugInfo(Instruction *Inst) const {
1292 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
1293 E = DDIs.end(); I != E; ++I) {
1294 DbgDeclareInst *DDI = *I;
1295 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1296 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1297 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1298 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1300 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
1301 E = DVIs.end(); I != E; ++I) {
1302 DbgValueInst *DVI = *I;
1304 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1305 // If an argument is zero extended then use argument directly. The ZExt
1306 // may be zapped by an optimization pass in future.
1307 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1308 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1309 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1310 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1312 Arg = SI->getOperand(0);
1313 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1314 Arg = LI->getOperand(0);
1318 Instruction *DbgVal =
1319 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1321 DbgVal->setDebugLoc(DVI->getDebugLoc());
1325 } // end anon namespace
1329 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1331 /// This pass takes allocations which can be completely analyzed (that is, they
1332 /// don't escape) and tries to turn them into scalar SSA values. There are
1333 /// a few steps to this process.
1335 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1336 /// are used to try to split them into smaller allocations, ideally of
1337 /// a single scalar data type. It will split up memcpy and memset accesses
1338 /// as necessary and try to isolate invidual scalar accesses.
1339 /// 2) It will transform accesses into forms which are suitable for SSA value
1340 /// promotion. This can be replacing a memset with a scalar store of an
1341 /// integer value, or it can involve speculating operations on a PHI or
1342 /// select to be a PHI or select of the results.
1343 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1344 /// onto insert and extract operations on a vector value, and convert them to
1345 /// this form. By doing so, it will enable promotion of vector aggregates to
1346 /// SSA vector values.
1347 class SROA : public FunctionPass {
1348 const bool RequiresDomTree;
1351 const DataLayout *TD;
1354 /// \brief Worklist of alloca instructions to simplify.
1356 /// Each alloca in the function is added to this. Each new alloca formed gets
1357 /// added to it as well to recursively simplify unless that alloca can be
1358 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1359 /// the one being actively rewritten, we add it back onto the list if not
1360 /// already present to ensure it is re-visited.
1361 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
1363 /// \brief A collection of instructions to delete.
1364 /// We try to batch deletions to simplify code and make things a bit more
1366 SmallVector<Instruction *, 8> DeadInsts;
1368 /// \brief A set to prevent repeatedly marking an instruction split into many
1369 /// uses as dead. Only used to guard insertion into DeadInsts.
1370 SmallPtrSet<Instruction *, 4> DeadSplitInsts;
1372 /// \brief Post-promotion worklist.
1374 /// Sometimes we discover an alloca which has a high probability of becoming
1375 /// viable for SROA after a round of promotion takes place. In those cases,
1376 /// the alloca is enqueued here for re-processing.
1378 /// Note that we have to be very careful to clear allocas out of this list in
1379 /// the event they are deleted.
1380 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
1382 /// \brief A collection of alloca instructions we can directly promote.
1383 std::vector<AllocaInst *> PromotableAllocas;
1386 SROA(bool RequiresDomTree = true)
1387 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
1388 C(0), TD(0), DT(0) {
1389 initializeSROAPass(*PassRegistry::getPassRegistry());
1391 bool runOnFunction(Function &F);
1392 void getAnalysisUsage(AnalysisUsage &AU) const;
1394 const char *getPassName() const { return "SROA"; }
1398 friend class PHIOrSelectSpeculator;
1399 friend class AllocaPartitionRewriter;
1400 friend class AllocaPartitionVectorRewriter;
1402 bool rewriteAllocaPartition(AllocaInst &AI,
1403 AllocaPartitioning &P,
1404 AllocaPartitioning::iterator PI);
1405 bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P);
1406 bool runOnAlloca(AllocaInst &AI);
1407 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
1408 bool promoteAllocas(Function &F);
1414 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1415 return new SROA(RequiresDomTree);
1418 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1420 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
1421 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1425 /// \brief Visitor to speculate PHIs and Selects where possible.
1426 class PHIOrSelectSpeculator : public InstVisitor<PHIOrSelectSpeculator> {
1427 // Befriend the base class so it can delegate to private visit methods.
1428 friend class llvm::InstVisitor<PHIOrSelectSpeculator>;
1430 const DataLayout &TD;
1431 AllocaPartitioning &P;
1435 PHIOrSelectSpeculator(const DataLayout &TD, AllocaPartitioning &P, SROA &Pass)
1436 : TD(TD), P(P), Pass(Pass) {}
1438 /// \brief Visit the users of an alloca partition and rewrite them.
1439 void visitUsers(AllocaPartitioning::const_iterator PI) {
1440 // Note that we need to use an index here as the underlying vector of uses
1441 // may be grown during speculation. However, we never need to re-visit the
1442 // new uses, and so we can use the initial size bound.
1443 for (unsigned Idx = 0, Size = P.use_size(PI); Idx != Size; ++Idx) {
1444 const AllocaPartitioning::PartitionUse &PU = P.getUse(PI, Idx);
1446 continue; // Skip dead use.
1448 visit(cast<Instruction>(PU.U->getUser()));
1453 // By default, skip this instruction.
1454 void visitInstruction(Instruction &I) {}
1456 /// PHI instructions that use an alloca and are subsequently loaded can be
1457 /// rewritten to load both input pointers in the pred blocks and then PHI the
1458 /// results, allowing the load of the alloca to be promoted.
1460 /// %P2 = phi [i32* %Alloca, i32* %Other]
1461 /// %V = load i32* %P2
1463 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1465 /// %V2 = load i32* %Other
1467 /// %V = phi [i32 %V1, i32 %V2]
1469 /// We can do this to a select if its only uses are loads and if the operands
1470 /// to the select can be loaded unconditionally.
1472 /// FIXME: This should be hoisted into a generic utility, likely in
1473 /// Transforms/Util/Local.h
1474 bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
1475 // For now, we can only do this promotion if the load is in the same block
1476 // as the PHI, and if there are no stores between the phi and load.
1477 // TODO: Allow recursive phi users.
1478 // TODO: Allow stores.
1479 BasicBlock *BB = PN.getParent();
1480 unsigned MaxAlign = 0;
1481 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
1483 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1484 if (LI == 0 || !LI->isSimple()) return false;
1486 // For now we only allow loads in the same block as the PHI. This is
1487 // a common case that happens when instcombine merges two loads through
1489 if (LI->getParent() != BB) return false;
1491 // Ensure that there are no instructions between the PHI and the load that
1493 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1494 if (BBI->mayWriteToMemory())
1497 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1498 Loads.push_back(LI);
1501 // We can only transform this if it is safe to push the loads into the
1502 // predecessor blocks. The only thing to watch out for is that we can't put
1503 // a possibly trapping load in the predecessor if it is a critical edge.
1504 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num;
1506 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1507 Value *InVal = PN.getIncomingValue(Idx);
1509 // If the value is produced by the terminator of the predecessor (an
1510 // invoke) or it has side-effects, there is no valid place to put a load
1511 // in the predecessor.
1512 if (TI == InVal || TI->mayHaveSideEffects())
1515 // If the predecessor has a single successor, then the edge isn't
1517 if (TI->getNumSuccessors() == 1)
1520 // If this pointer is always safe to load, or if we can prove that there
1521 // is already a load in the block, then we can move the load to the pred
1523 if (InVal->isDereferenceablePointer() ||
1524 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
1533 void visitPHINode(PHINode &PN) {
1534 DEBUG(dbgs() << " original: " << PN << "\n");
1536 SmallVector<LoadInst *, 4> Loads;
1537 if (!isSafePHIToSpeculate(PN, Loads))
1540 assert(!Loads.empty());
1542 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1543 IRBuilder<> PHIBuilder(&PN);
1544 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1545 PN.getName() + ".sroa.speculated");
1547 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1548 // matter which one we get and if any differ, it doesn't matter.
1549 LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
1550 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1551 unsigned Align = SomeLoad->getAlignment();
1553 // Rewrite all loads of the PN to use the new PHI.
1555 LoadInst *LI = Loads.pop_back_val();
1556 LI->replaceAllUsesWith(NewPN);
1557 Pass.DeadInsts.push_back(LI);
1558 } while (!Loads.empty());
1560 // Inject loads into all of the pred blocks.
1561 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1562 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1563 TerminatorInst *TI = Pred->getTerminator();
1564 Use *InUse = &PN.getOperandUse(PN.getOperandNumForIncomingValue(Idx));
1565 Value *InVal = PN.getIncomingValue(Idx);
1566 IRBuilder<> PredBuilder(TI);
1569 = PredBuilder.CreateLoad(InVal, (PN.getName() + ".sroa.speculate.load." +
1571 ++NumLoadsSpeculated;
1572 Load->setAlignment(Align);
1574 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1575 NewPN->addIncoming(Load, Pred);
1577 Instruction *Ptr = dyn_cast<Instruction>(InVal);
1579 // No uses to rewrite.
1582 // Try to lookup and rewrite any partition uses corresponding to this phi
1584 AllocaPartitioning::iterator PI
1585 = P.findPartitionForPHIOrSelectOperand(InUse);
1589 // Replace the Use in the PartitionUse for this operand with the Use
1591 AllocaPartitioning::use_iterator UI
1592 = P.findPartitionUseForPHIOrSelectOperand(InUse);
1593 assert(isa<PHINode>(*UI->U->getUser()));
1594 UI->U = &Load->getOperandUse(Load->getPointerOperandIndex());
1596 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1599 /// Select instructions that use an alloca and are subsequently loaded can be
1600 /// rewritten to load both input pointers and then select between the result,
1601 /// allowing the load of the alloca to be promoted.
1603 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1604 /// %V = load i32* %P2
1606 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1607 /// %V2 = load i32* %Other
1608 /// %V = select i1 %cond, i32 %V1, i32 %V2
1610 /// We can do this to a select if its only uses are loads and if the operand
1611 /// to the select can be loaded unconditionally.
1612 bool isSafeSelectToSpeculate(SelectInst &SI,
1613 SmallVectorImpl<LoadInst *> &Loads) {
1614 Value *TValue = SI.getTrueValue();
1615 Value *FValue = SI.getFalseValue();
1616 bool TDerefable = TValue->isDereferenceablePointer();
1617 bool FDerefable = FValue->isDereferenceablePointer();
1619 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
1621 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1622 if (LI == 0 || !LI->isSimple()) return false;
1624 // Both operands to the select need to be dereferencable, either
1625 // absolutely (e.g. allocas) or at this point because we can see other
1627 if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
1628 LI->getAlignment(), &TD))
1630 if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
1631 LI->getAlignment(), &TD))
1633 Loads.push_back(LI);
1639 void visitSelectInst(SelectInst &SI) {
1640 DEBUG(dbgs() << " original: " << SI << "\n");
1641 IRBuilder<> IRB(&SI);
1643 // If the select isn't safe to speculate, just use simple logic to emit it.
1644 SmallVector<LoadInst *, 4> Loads;
1645 if (!isSafeSelectToSpeculate(SI, Loads))
1648 Use *Ops[2] = { &SI.getOperandUse(1), &SI.getOperandUse(2) };
1649 AllocaPartitioning::iterator PIs[2];
1650 AllocaPartitioning::PartitionUse PUs[2];
1651 for (unsigned i = 0, e = 2; i != e; ++i) {
1652 PIs[i] = P.findPartitionForPHIOrSelectOperand(Ops[i]);
1653 if (PIs[i] != P.end()) {
1654 // If the pointer is within the partitioning, remove the select from
1655 // its uses. We'll add in the new loads below.
1656 AllocaPartitioning::use_iterator UI
1657 = P.findPartitionUseForPHIOrSelectOperand(Ops[i]);
1659 // Clear out the use here so that the offsets into the use list remain
1660 // stable but this use is ignored when rewriting.
1665 Value *TV = SI.getTrueValue();
1666 Value *FV = SI.getFalseValue();
1667 // Replace the loads of the select with a select of two loads.
1668 while (!Loads.empty()) {
1669 LoadInst *LI = Loads.pop_back_val();
1671 IRB.SetInsertPoint(LI);
1673 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1675 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1676 NumLoadsSpeculated += 2;
1678 // Transfer alignment and TBAA info if present.
1679 TL->setAlignment(LI->getAlignment());
1680 FL->setAlignment(LI->getAlignment());
1681 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1682 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1683 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1686 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1687 LI->getName() + ".sroa.speculated");
1689 LoadInst *Loads[2] = { TL, FL };
1690 for (unsigned i = 0, e = 2; i != e; ++i) {
1691 if (PIs[i] != P.end()) {
1692 Use *LoadUse = &Loads[i]->getOperandUse(0);
1693 assert(PUs[i].U->get() == LoadUse->get());
1695 P.use_push_back(PIs[i], PUs[i]);
1699 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1700 LI->replaceAllUsesWith(V);
1701 Pass.DeadInsts.push_back(LI);
1707 /// \brief Accumulate the constant offsets in a GEP into a single APInt offset.
1709 /// If the provided GEP is all-constant, the total byte offset formed by the
1710 /// GEP is computed and Offset is set to it. If the GEP has any non-constant
1711 /// operands, the function returns false and the value of Offset is unmodified.
1712 static bool accumulateGEPOffsets(const DataLayout &TD, GEPOperator &GEP,
1714 APInt GEPOffset(Offset.getBitWidth(), 0);
1715 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1716 GTI != GTE; ++GTI) {
1717 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
1720 if (OpC->isZero()) continue;
1722 // Handle a struct index, which adds its field offset to the pointer.
1723 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1724 unsigned ElementIdx = OpC->getZExtValue();
1725 const StructLayout *SL = TD.getStructLayout(STy);
1726 GEPOffset += APInt(Offset.getBitWidth(),
1727 SL->getElementOffset(ElementIdx));
1731 APInt TypeSize(Offset.getBitWidth(),
1732 TD.getTypeAllocSize(GTI.getIndexedType()));
1733 if (VectorType *VTy = dyn_cast<VectorType>(*GTI)) {
1734 assert((VTy->getScalarSizeInBits() % 8) == 0 &&
1735 "vector element size is not a multiple of 8, cannot GEP over it");
1736 TypeSize = VTy->getScalarSizeInBits() / 8;
1739 GEPOffset += OpC->getValue().sextOrTrunc(Offset.getBitWidth()) * TypeSize;
1745 /// \brief Build a GEP out of a base pointer and indices.
1747 /// This will return the BasePtr if that is valid, or build a new GEP
1748 /// instruction using the IRBuilder if GEP-ing is needed.
1749 static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr,
1750 SmallVectorImpl<Value *> &Indices,
1751 const Twine &Prefix) {
1752 if (Indices.empty())
1755 // A single zero index is a no-op, so check for this and avoid building a GEP
1757 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1760 return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx");
1763 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1764 /// TargetTy without changing the offset of the pointer.
1766 /// This routine assumes we've already established a properly offset GEP with
1767 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1768 /// zero-indices down through type layers until we find one the same as
1769 /// TargetTy. If we can't find one with the same type, we at least try to use
1770 /// one with the same size. If none of that works, we just produce the GEP as
1771 /// indicated by Indices to have the correct offset.
1772 static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const DataLayout &TD,
1773 Value *BasePtr, Type *Ty, Type *TargetTy,
1774 SmallVectorImpl<Value *> &Indices,
1775 const Twine &Prefix) {
1777 return buildGEP(IRB, BasePtr, Indices, Prefix);
1779 // See if we can descend into a struct and locate a field with the correct
1781 unsigned NumLayers = 0;
1782 Type *ElementTy = Ty;
1784 if (ElementTy->isPointerTy())
1786 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1787 ElementTy = SeqTy->getElementType();
1788 Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(
1789 ElementTy->isPointerTy() ?
1790 cast<PointerType>(ElementTy)->getAddressSpace(): 0), 0)));
1791 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1792 if (STy->element_begin() == STy->element_end())
1793 break; // Nothing left to descend into.
1794 ElementTy = *STy->element_begin();
1795 Indices.push_back(IRB.getInt32(0));
1800 } while (ElementTy != TargetTy);
1801 if (ElementTy != TargetTy)
1802 Indices.erase(Indices.end() - NumLayers, Indices.end());
1804 return buildGEP(IRB, BasePtr, Indices, Prefix);
1807 /// \brief Recursively compute indices for a natural GEP.
1809 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1810 /// element types adding appropriate indices for the GEP.
1811 static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const DataLayout &TD,
1812 Value *Ptr, Type *Ty, APInt &Offset,
1814 SmallVectorImpl<Value *> &Indices,
1815 const Twine &Prefix) {
1817 return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix);
1819 // We can't recurse through pointer types.
1820 if (Ty->isPointerTy())
1823 // We try to analyze GEPs over vectors here, but note that these GEPs are
1824 // extremely poorly defined currently. The long-term goal is to remove GEPing
1825 // over a vector from the IR completely.
1826 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1827 unsigned ElementSizeInBits = VecTy->getScalarSizeInBits();
1828 if (ElementSizeInBits % 8)
1829 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1830 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1831 APInt NumSkippedElements = Offset.udiv(ElementSize);
1832 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1834 Offset -= NumSkippedElements * ElementSize;
1835 Indices.push_back(IRB.getInt(NumSkippedElements));
1836 return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
1837 Offset, TargetTy, Indices, Prefix);
1840 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1841 Type *ElementTy = ArrTy->getElementType();
1842 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1843 APInt NumSkippedElements = Offset.udiv(ElementSize);
1844 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1847 Offset -= NumSkippedElements * ElementSize;
1848 Indices.push_back(IRB.getInt(NumSkippedElements));
1849 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1853 StructType *STy = dyn_cast<StructType>(Ty);
1857 const StructLayout *SL = TD.getStructLayout(STy);
1858 uint64_t StructOffset = Offset.getZExtValue();
1859 if (StructOffset >= SL->getSizeInBytes())
1861 unsigned Index = SL->getElementContainingOffset(StructOffset);
1862 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1863 Type *ElementTy = STy->getElementType(Index);
1864 if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
1865 return 0; // The offset points into alignment padding.
1867 Indices.push_back(IRB.getInt32(Index));
1868 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1872 /// \brief Get a natural GEP from a base pointer to a particular offset and
1873 /// resulting in a particular type.
1875 /// The goal is to produce a "natural" looking GEP that works with the existing
1876 /// composite types to arrive at the appropriate offset and element type for
1877 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1878 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1879 /// Indices, and setting Ty to the result subtype.
1881 /// If no natural GEP can be constructed, this function returns null.
1882 static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const DataLayout &TD,
1883 Value *Ptr, APInt Offset, Type *TargetTy,
1884 SmallVectorImpl<Value *> &Indices,
1885 const Twine &Prefix) {
1886 PointerType *Ty = cast<PointerType>(Ptr->getType());
1888 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1890 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1893 Type *ElementTy = Ty->getElementType();
1894 if (!ElementTy->isSized())
1895 return 0; // We can't GEP through an unsized element.
1896 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1897 if (ElementSize == 0)
1898 return 0; // Zero-length arrays can't help us build a natural GEP.
1899 APInt NumSkippedElements = Offset.udiv(ElementSize);
1901 Offset -= NumSkippedElements * ElementSize;
1902 Indices.push_back(IRB.getInt(NumSkippedElements));
1903 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1907 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1908 /// resulting pointer has PointerTy.
1910 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1911 /// and produces the pointer type desired. Where it cannot, it will try to use
1912 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1913 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1914 /// bitcast to the type.
1916 /// The strategy for finding the more natural GEPs is to peel off layers of the
1917 /// pointer, walking back through bit casts and GEPs, searching for a base
1918 /// pointer from which we can compute a natural GEP with the desired
1919 /// properities. The algorithm tries to fold as many constant indices into
1920 /// a single GEP as possible, thus making each GEP more independent of the
1921 /// surrounding code.
1922 static Value *getAdjustedPtr(IRBuilder<> &IRB, const DataLayout &TD,
1923 Value *Ptr, APInt Offset, Type *PointerTy,
1924 const Twine &Prefix) {
1925 // Even though we don't look through PHI nodes, we could be called on an
1926 // instruction in an unreachable block, which may be on a cycle.
1927 SmallPtrSet<Value *, 4> Visited;
1928 Visited.insert(Ptr);
1929 SmallVector<Value *, 4> Indices;
1931 // We may end up computing an offset pointer that has the wrong type. If we
1932 // never are able to compute one directly that has the correct type, we'll
1933 // fall back to it, so keep it around here.
1934 Value *OffsetPtr = 0;
1936 // Remember any i8 pointer we come across to re-use if we need to do a raw
1939 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1941 Type *TargetTy = PointerTy->getPointerElementType();
1944 // First fold any existing GEPs into the offset.
1945 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1946 APInt GEPOffset(Offset.getBitWidth(), 0);
1947 if (!accumulateGEPOffsets(TD, *GEP, GEPOffset))
1949 Offset += GEPOffset;
1950 Ptr = GEP->getPointerOperand();
1951 if (!Visited.insert(Ptr))
1955 // See if we can perform a natural GEP here.
1957 if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
1959 if (P->getType() == PointerTy) {
1960 // Zap any offset pointer that we ended up computing in previous rounds.
1961 if (OffsetPtr && OffsetPtr->use_empty())
1962 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1963 I->eraseFromParent();
1971 // Stash this pointer if we've found an i8*.
1972 if (Ptr->getType()->isIntegerTy(8)) {
1974 Int8PtrOffset = Offset;
1977 // Peel off a layer of the pointer and update the offset appropriately.
1978 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1979 Ptr = cast<Operator>(Ptr)->getOperand(0);
1980 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1981 if (GA->mayBeOverridden())
1983 Ptr = GA->getAliasee();
1987 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1988 } while (Visited.insert(Ptr));
1992 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1993 Prefix + ".raw_cast");
1994 Int8PtrOffset = Offset;
1997 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1998 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1999 Prefix + ".raw_idx");
2003 // On the off chance we were targeting i8*, guard the bitcast here.
2004 if (Ptr->getType() != PointerTy)
2005 Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast");
2010 /// \brief Test whether the given alloca partition can be promoted to a vector.
2012 /// This is a quick test to check whether we can rewrite a particular alloca
2013 /// partition (and its newly formed alloca) into a vector alloca with only
2014 /// whole-vector loads and stores such that it could be promoted to a vector
2015 /// SSA value. We only can ensure this for a limited set of operations, and we
2016 /// don't want to do the rewrites unless we are confident that the result will
2017 /// be promotable, so we have an early test here.
2018 static bool isVectorPromotionViable(const DataLayout &TD,
2020 AllocaPartitioning &P,
2021 uint64_t PartitionBeginOffset,
2022 uint64_t PartitionEndOffset,
2023 AllocaPartitioning::const_use_iterator I,
2024 AllocaPartitioning::const_use_iterator E) {
2025 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
2029 uint64_t VecSize = TD.getTypeSizeInBits(Ty);
2030 uint64_t ElementSize = Ty->getScalarSizeInBits();
2032 // While the definition of LLVM vectors is bitpacked, we don't support sizes
2033 // that aren't byte sized.
2034 if (ElementSize % 8)
2036 assert((VecSize % 8) == 0 && "vector size not a multiple of element size?");
2040 for (; I != E; ++I) {
2042 continue; // Skip dead use.
2044 uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
2045 uint64_t BeginIndex = BeginOffset / ElementSize;
2046 if (BeginIndex * ElementSize != BeginOffset ||
2047 BeginIndex >= Ty->getNumElements())
2049 uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
2050 uint64_t EndIndex = EndOffset / ElementSize;
2051 if (EndIndex * ElementSize != EndOffset ||
2052 EndIndex > Ty->getNumElements())
2055 // FIXME: We should build shuffle vector instructions to handle
2056 // non-element-sized accesses.
2057 if ((EndOffset - BeginOffset) != ElementSize &&
2058 (EndOffset - BeginOffset) != VecSize)
2061 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
2062 if (MI->isVolatile())
2064 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
2065 const AllocaPartitioning::MemTransferOffsets &MTO
2066 = P.getMemTransferOffsets(*MTI);
2067 if (!MTO.IsSplittable)
2070 } else if (I->U->get()->getType()->getPointerElementType()->isStructTy()) {
2071 // Disable vector promotion when there are loads or stores of an FCA.
2073 } else if (!isa<LoadInst>(I->U->getUser()) &&
2074 !isa<StoreInst>(I->U->getUser())) {
2081 /// \brief Test whether the given alloca partition can be promoted to an int.
2083 /// This is a quick test to check whether we can rewrite a particular alloca
2084 /// partition (and its newly formed alloca) into an integer alloca suitable for
2085 /// promotion to an SSA value. We only can ensure this for a limited set of
2086 /// operations, and we don't want to do the rewrites unless we are confident
2087 /// that the result will be promotable, so we have an early test here.
2088 static bool isIntegerPromotionViable(const DataLayout &TD,
2090 uint64_t AllocBeginOffset,
2091 AllocaPartitioning &P,
2092 AllocaPartitioning::const_use_iterator I,
2093 AllocaPartitioning::const_use_iterator E) {
2094 IntegerType *Ty = dyn_cast<IntegerType>(AllocaTy);
2095 if (!Ty || 8*TD.getTypeStoreSize(Ty) != Ty->getBitWidth())
2098 // Check the uses to ensure the uses are (likely) promoteable integer uses.
2099 // Also ensure that the alloca has a covering load or store. We don't want
2100 // promote because of some other unsplittable entry (which we may make
2101 // splittable later) and lose the ability to promote each element access.
2102 bool WholeAllocaOp = false;
2103 for (; I != E; ++I) {
2105 continue; // Skip dead use.
2107 // We can't reasonably handle cases where the load or store extends past
2108 // the end of the aloca's type and into its padding.
2109 if ((I->EndOffset - AllocBeginOffset) > TD.getTypeStoreSize(Ty))
2112 if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) {
2113 if (LI->isVolatile() || !LI->getType()->isIntegerTy())
2115 if (LI->getType() == Ty)
2116 WholeAllocaOp = true;
2117 } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) {
2118 if (SI->isVolatile() || !SI->getValueOperand()->getType()->isIntegerTy())
2120 if (SI->getValueOperand()->getType() == Ty)
2121 WholeAllocaOp = true;
2122 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
2123 if (MI->isVolatile())
2125 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
2126 const AllocaPartitioning::MemTransferOffsets &MTO
2127 = P.getMemTransferOffsets(*MTI);
2128 if (!MTO.IsSplittable)
2135 return WholeAllocaOp;
2139 /// \brief Visitor to rewrite instructions using a partition of an alloca to
2140 /// use a new alloca.
2142 /// Also implements the rewriting to vector-based accesses when the partition
2143 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2145 class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
2147 // Befriend the base class so it can delegate to private visit methods.
2148 friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
2150 const DataLayout &TD;
2151 AllocaPartitioning &P;
2153 AllocaInst &OldAI, &NewAI;
2154 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2156 // If we are rewriting an alloca partition which can be written as pure
2157 // vector operations, we stash extra information here. When VecTy is
2158 // non-null, we have some strict guarantees about the rewriten alloca:
2159 // - The new alloca is exactly the size of the vector type here.
2160 // - The accesses all either map to the entire vector or to a single
2162 // - The set of accessing instructions is only one of those handled above
2163 // in isVectorPromotionViable. Generally these are the same access kinds
2164 // which are promotable via mem2reg.
2167 uint64_t ElementSize;
2169 // This is a convenience and flag variable that will be null unless the new
2170 // alloca has a promotion-targeted integer type due to passing
2171 // isIntegerPromotionViable above. If it is non-null does, the desired
2172 // integer type will be stored here for easy access during rewriting.
2173 IntegerType *IntPromotionTy;
2175 // The offset of the partition user currently being rewritten.
2176 uint64_t BeginOffset, EndOffset;
2178 Instruction *OldPtr;
2180 // The name prefix to use when rewriting instructions for this alloca.
2181 std::string NamePrefix;
2184 AllocaPartitionRewriter(const DataLayout &TD, AllocaPartitioning &P,
2185 AllocaPartitioning::iterator PI,
2186 SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
2187 uint64_t NewBeginOffset, uint64_t NewEndOffset)
2188 : TD(TD), P(P), Pass(Pass),
2189 OldAI(OldAI), NewAI(NewAI),
2190 NewAllocaBeginOffset(NewBeginOffset),
2191 NewAllocaEndOffset(NewEndOffset),
2192 VecTy(), ElementTy(), ElementSize(), IntPromotionTy(),
2193 BeginOffset(), EndOffset() {
2196 /// \brief Visit the users of the alloca partition and rewrite them.
2197 bool visitUsers(AllocaPartitioning::const_use_iterator I,
2198 AllocaPartitioning::const_use_iterator E) {
2199 if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P,
2200 NewAllocaBeginOffset, NewAllocaEndOffset,
2203 VecTy = cast<VectorType>(NewAI.getAllocatedType());
2204 ElementTy = VecTy->getElementType();
2205 assert((VecTy->getScalarSizeInBits() % 8) == 0 &&
2206 "Only multiple-of-8 sized vector elements are viable");
2207 ElementSize = VecTy->getScalarSizeInBits() / 8;
2208 } else if (isIntegerPromotionViable(TD, NewAI.getAllocatedType(),
2209 NewAllocaBeginOffset, P, I, E)) {
2210 IntPromotionTy = cast<IntegerType>(NewAI.getAllocatedType());
2212 bool CanSROA = true;
2213 for (; I != E; ++I) {
2215 continue; // Skip dead uses.
2216 BeginOffset = I->BeginOffset;
2217 EndOffset = I->EndOffset;
2219 OldPtr = cast<Instruction>(I->U->get());
2220 NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str();
2221 CanSROA &= visit(cast<Instruction>(I->U->getUser()));
2233 // Every instruction which can end up as a user must have a rewrite rule.
2234 bool visitInstruction(Instruction &I) {
2235 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2236 llvm_unreachable("No rewrite rule for this instruction!");
2239 Twine getName(const Twine &Suffix) {
2240 return NamePrefix + Suffix;
2243 Value *getAdjustedAllocaPtr(IRBuilder<> &IRB, Type *PointerTy) {
2244 assert(BeginOffset >= NewAllocaBeginOffset);
2245 unsigned AS = cast<PointerType>(PointerTy)->getAddressSpace();
2246 APInt Offset(TD.getPointerSizeInBits(AS), BeginOffset - NewAllocaBeginOffset);
2247 return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName(""));
2250 /// \brief Compute suitable alignment to access an offset into the new alloca.
2251 unsigned getOffsetAlign(uint64_t Offset) {
2252 unsigned NewAIAlign = NewAI.getAlignment();
2254 NewAIAlign = TD.getABITypeAlignment(NewAI.getAllocatedType());
2255 return MinAlign(NewAIAlign, Offset);
2258 /// \brief Compute suitable alignment to access this partition of the new
2260 unsigned getPartitionAlign() {
2261 return getOffsetAlign(BeginOffset - NewAllocaBeginOffset);
2264 /// \brief Compute suitable alignment to access a type at an offset of the
2267 /// \returns zero if the type's ABI alignment is a suitable alignment,
2268 /// otherwise returns the maximal suitable alignment.
2269 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
2270 unsigned Align = getOffsetAlign(Offset);
2271 return Align == TD.getABITypeAlignment(Ty) ? 0 : Align;
2274 /// \brief Compute suitable alignment to access a type at the beginning of
2275 /// this partition of the new alloca.
2277 /// See \c getOffsetTypeAlign for details; this routine delegates to it.
2278 unsigned getPartitionTypeAlign(Type *Ty) {
2279 return getOffsetTypeAlign(Ty, BeginOffset - NewAllocaBeginOffset);
2282 ConstantInt *getIndex(IRBuilder<> &IRB, uint64_t Offset) {
2283 assert(VecTy && "Can only call getIndex when rewriting a vector");
2284 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2285 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2286 uint32_t Index = RelOffset / ElementSize;
2287 assert(Index * ElementSize == RelOffset);
2288 return IRB.getInt32(Index);
2291 Value *extractInteger(IRBuilder<> &IRB, IntegerType *TargetTy,
2293 assert(IntPromotionTy && "Alloca is not an integer we can extract from");
2294 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2296 assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
2297 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2298 assert(TD.getTypeStoreSize(TargetTy) + RelOffset <=
2299 TD.getTypeStoreSize(IntPromotionTy) &&
2300 "Element load outside of alloca store");
2301 uint64_t ShAmt = 8*RelOffset;
2302 if (TD.isBigEndian())
2303 ShAmt = 8*(TD.getTypeStoreSize(IntPromotionTy) -
2304 TD.getTypeStoreSize(TargetTy) - RelOffset);
2306 V = IRB.CreateLShr(V, ShAmt, getName(".shift"));
2307 if (TargetTy != IntPromotionTy) {
2308 assert(TargetTy->getBitWidth() < IntPromotionTy->getBitWidth() &&
2309 "Cannot extract to a larger integer!");
2310 V = IRB.CreateTrunc(V, TargetTy, getName(".trunc"));
2315 StoreInst *insertInteger(IRBuilder<> &IRB, Value *V, uint64_t Offset) {
2316 IntegerType *Ty = cast<IntegerType>(V->getType());
2317 if (Ty == IntPromotionTy)
2318 return IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2320 assert(Ty->getBitWidth() < IntPromotionTy->getBitWidth() &&
2321 "Cannot insert a larger integer!");
2322 V = IRB.CreateZExt(V, IntPromotionTy, getName(".ext"));
2323 assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
2324 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2325 assert(TD.getTypeStoreSize(Ty) + RelOffset <=
2326 TD.getTypeStoreSize(IntPromotionTy) &&
2327 "Element store outside of alloca store");
2328 uint64_t ShAmt = 8*RelOffset;
2329 if (TD.isBigEndian())
2330 ShAmt = 8*(TD.getTypeStoreSize(IntPromotionTy) - TD.getTypeStoreSize(Ty)
2333 V = IRB.CreateShl(V, ShAmt, getName(".shift"));
2335 APInt Mask = ~Ty->getMask().zext(IntPromotionTy->getBitWidth()).shl(ShAmt);
2336 Value *Old = IRB.CreateAnd(IRB.CreateAlignedLoad(&NewAI,
2337 NewAI.getAlignment(),
2338 getName(".oldload")),
2339 Mask, getName(".mask"));
2340 return IRB.CreateAlignedStore(IRB.CreateOr(Old, V, getName(".insert")),
2341 &NewAI, NewAI.getAlignment());
2344 void deleteIfTriviallyDead(Value *V) {
2345 Instruction *I = cast<Instruction>(V);
2346 if (isInstructionTriviallyDead(I))
2347 Pass.DeadInsts.push_back(I);
2350 Value *getValueCast(IRBuilder<> &IRB, Value *V, Type *Ty) {
2351 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
2352 return IRB.CreateIntToPtr(V, Ty);
2353 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
2354 return IRB.CreatePtrToInt(V, Ty);
2356 return IRB.CreateBitCast(V, Ty);
2359 bool rewriteVectorizedLoadInst(IRBuilder<> &IRB, LoadInst &LI, Value *OldOp) {
2361 if (LI.getType() == VecTy->getElementType() ||
2362 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
2363 Result = IRB.CreateExtractElement(
2364 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
2365 getIndex(IRB, BeginOffset), getName(".extract"));
2367 Result = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2370 if (Result->getType() != LI.getType())
2371 Result = getValueCast(IRB, Result, LI.getType());
2372 LI.replaceAllUsesWith(Result);
2373 Pass.DeadInsts.push_back(&LI);
2375 DEBUG(dbgs() << " to: " << *Result << "\n");
2379 bool rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) {
2380 assert(!LI.isVolatile());
2381 Value *Result = extractInteger(IRB, cast<IntegerType>(LI.getType()),
2383 LI.replaceAllUsesWith(Result);
2384 Pass.DeadInsts.push_back(&LI);
2385 DEBUG(dbgs() << " to: " << *Result << "\n");
2389 bool visitLoadInst(LoadInst &LI) {
2390 DEBUG(dbgs() << " original: " << LI << "\n");
2391 Value *OldOp = LI.getOperand(0);
2392 assert(OldOp == OldPtr);
2393 IRBuilder<> IRB(&LI);
2396 return rewriteVectorizedLoadInst(IRB, LI, OldOp);
2398 return rewriteIntegerLoad(IRB, LI);
2400 Value *NewPtr = getAdjustedAllocaPtr(IRB,
2401 LI.getPointerOperand()->getType());
2402 LI.setOperand(0, NewPtr);
2403 LI.setAlignment(getPartitionTypeAlign(LI.getType()));
2404 DEBUG(dbgs() << " to: " << LI << "\n");
2406 deleteIfTriviallyDead(OldOp);
2407 return NewPtr == &NewAI && !LI.isVolatile();
2410 bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, StoreInst &SI,
2412 Value *V = SI.getValueOperand();
2413 if (V->getType() == ElementTy ||
2414 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
2415 if (V->getType() != ElementTy)
2416 V = getValueCast(IRB, V, ElementTy);
2417 LoadInst *LI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2419 V = IRB.CreateInsertElement(LI, V, getIndex(IRB, BeginOffset),
2420 getName(".insert"));
2421 } else if (V->getType() != VecTy) {
2422 V = getValueCast(IRB, V, VecTy);
2424 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2425 Pass.DeadInsts.push_back(&SI);
2428 DEBUG(dbgs() << " to: " << *Store << "\n");
2432 bool rewriteIntegerStore(IRBuilder<> &IRB, StoreInst &SI) {
2433 assert(!SI.isVolatile());
2434 StoreInst *Store = insertInteger(IRB, SI.getValueOperand(), BeginOffset);
2435 Pass.DeadInsts.push_back(&SI);
2437 DEBUG(dbgs() << " to: " << *Store << "\n");
2441 bool visitStoreInst(StoreInst &SI) {
2442 DEBUG(dbgs() << " original: " << SI << "\n");
2443 Value *OldOp = SI.getOperand(1);
2444 assert(OldOp == OldPtr);
2445 IRBuilder<> IRB(&SI);
2448 return rewriteVectorizedStoreInst(IRB, SI, OldOp);
2450 return rewriteIntegerStore(IRB, SI);
2452 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2453 // alloca that should be re-examined after promoting this alloca.
2454 if (SI.getValueOperand()->getType()->isPointerTy())
2455 if (AllocaInst *AI = dyn_cast<AllocaInst>(SI.getValueOperand()
2456 ->stripInBoundsOffsets()))
2457 Pass.PostPromotionWorklist.insert(AI);
2459 Value *NewPtr = getAdjustedAllocaPtr(IRB,
2460 SI.getPointerOperand()->getType());
2461 SI.setOperand(1, NewPtr);
2462 SI.setAlignment(getPartitionTypeAlign(SI.getValueOperand()->getType()));
2463 DEBUG(dbgs() << " to: " << SI << "\n");
2465 deleteIfTriviallyDead(OldOp);
2466 return NewPtr == &NewAI && !SI.isVolatile();
2469 bool visitMemSetInst(MemSetInst &II) {
2470 DEBUG(dbgs() << " original: " << II << "\n");
2471 IRBuilder<> IRB(&II);
2472 assert(II.getRawDest() == OldPtr);
2474 // If the memset has a variable size, it cannot be split, just adjust the
2475 // pointer to the new alloca.
2476 if (!isa<Constant>(II.getLength())) {
2477 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2478 Type *CstTy = II.getAlignmentCst()->getType();
2479 II.setAlignment(ConstantInt::get(CstTy, getPartitionAlign()));
2481 deleteIfTriviallyDead(OldPtr);
2485 // Record this instruction for deletion.
2486 if (Pass.DeadSplitInsts.insert(&II))
2487 Pass.DeadInsts.push_back(&II);
2489 Type *AllocaTy = NewAI.getAllocatedType();
2490 Type *ScalarTy = AllocaTy->getScalarType();
2492 // If this doesn't map cleanly onto the alloca type, and that type isn't
2493 // a single value type, just emit a memset.
2494 if (!VecTy && (BeginOffset != NewAllocaBeginOffset ||
2495 EndOffset != NewAllocaEndOffset ||
2496 !AllocaTy->isSingleValueType() ||
2497 !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)))) {
2498 Type *SizeTy = II.getLength()->getType();
2499 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2501 = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB,
2502 II.getRawDest()->getType()),
2503 II.getValue(), Size, getPartitionAlign(),
2506 DEBUG(dbgs() << " to: " << *New << "\n");
2510 // If we can represent this as a simple value, we have to build the actual
2511 // value to store, which requires expanding the byte present in memset to
2512 // a sensible representation for the alloca type. This is essentially
2513 // splatting the byte to a sufficiently wide integer, bitcasting to the
2514 // desired scalar type, and splatting it across any desired vector type.
2515 Value *V = II.getValue();
2516 IntegerType *VTy = cast<IntegerType>(V->getType());
2517 Type *IntTy = Type::getIntNTy(VTy->getContext(),
2518 TD.getTypeSizeInBits(ScalarTy));
2519 if (TD.getTypeSizeInBits(ScalarTy) > VTy->getBitWidth())
2520 V = IRB.CreateMul(IRB.CreateZExt(V, IntTy, getName(".zext")),
2521 ConstantExpr::getUDiv(
2522 Constant::getAllOnesValue(IntTy),
2523 ConstantExpr::getZExt(
2524 Constant::getAllOnesValue(V->getType()),
2526 getName(".isplat"));
2527 if (V->getType() != ScalarTy) {
2528 if (ScalarTy->isPointerTy())
2529 V = IRB.CreateIntToPtr(V, ScalarTy);
2530 else if (ScalarTy->isPrimitiveType() || ScalarTy->isVectorTy())
2531 V = IRB.CreateBitCast(V, ScalarTy);
2532 else if (ScalarTy->isIntegerTy())
2533 llvm_unreachable("Computed different integer types with equal widths");
2535 llvm_unreachable("Invalid scalar type");
2538 // If this is an element-wide memset of a vectorizable alloca, insert it.
2539 if (VecTy && (BeginOffset > NewAllocaBeginOffset ||
2540 EndOffset < NewAllocaEndOffset)) {
2541 StoreInst *Store = IRB.CreateAlignedStore(
2542 IRB.CreateInsertElement(IRB.CreateAlignedLoad(&NewAI,
2543 NewAI.getAlignment(),
2545 V, getIndex(IRB, BeginOffset),
2546 getName(".insert")),
2547 &NewAI, NewAI.getAlignment());
2549 DEBUG(dbgs() << " to: " << *Store << "\n");
2553 // Splat to a vector if needed.
2554 if (VectorType *VecTy = dyn_cast<VectorType>(AllocaTy)) {
2555 VectorType *SplatSourceTy = VectorType::get(V->getType(), 1);
2556 V = IRB.CreateShuffleVector(
2557 IRB.CreateInsertElement(UndefValue::get(SplatSourceTy), V,
2558 IRB.getInt32(0), getName(".vsplat.insert")),
2559 UndefValue::get(SplatSourceTy),
2560 ConstantVector::getSplat(VecTy->getNumElements(), IRB.getInt32(0)),
2561 getName(".vsplat.shuffle"));
2562 assert(V->getType() == VecTy);
2565 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2568 DEBUG(dbgs() << " to: " << *New << "\n");
2569 return !II.isVolatile();
2572 bool visitMemTransferInst(MemTransferInst &II) {
2573 // Rewriting of memory transfer instructions can be a bit tricky. We break
2574 // them into two categories: split intrinsics and unsplit intrinsics.
2576 DEBUG(dbgs() << " original: " << II << "\n");
2577 IRBuilder<> IRB(&II);
2579 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2580 bool IsDest = II.getRawDest() == OldPtr;
2582 const AllocaPartitioning::MemTransferOffsets &MTO
2583 = P.getMemTransferOffsets(II);
2585 assert(OldPtr->getType()->isPointerTy() && "Must be a pointer type!");
2586 unsigned AS = cast<PointerType>(OldPtr->getType())->getAddressSpace();
2587 // Compute the relative offset within the transfer.
2588 unsigned IntPtrWidth = TD.getPointerSizeInBits(AS);
2589 APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
2590 : MTO.SourceBegin));
2592 unsigned Align = II.getAlignment();
2594 Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2595 MinAlign(II.getAlignment(), getPartitionAlign()));
2597 // For unsplit intrinsics, we simply modify the source and destination
2598 // pointers in place. This isn't just an optimization, it is a matter of
2599 // correctness. With unsplit intrinsics we may be dealing with transfers
2600 // within a single alloca before SROA ran, or with transfers that have
2601 // a variable length. We may also be dealing with memmove instead of
2602 // memcpy, and so simply updating the pointers is the necessary for us to
2603 // update both source and dest of a single call.
2604 if (!MTO.IsSplittable) {
2605 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2607 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2609 II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType()));
2611 Type *CstTy = II.getAlignmentCst()->getType();
2612 II.setAlignment(ConstantInt::get(CstTy, Align));
2614 DEBUG(dbgs() << " to: " << II << "\n");
2615 deleteIfTriviallyDead(OldOp);
2618 // For split transfer intrinsics we have an incredibly useful assurance:
2619 // the source and destination do not reside within the same alloca, and at
2620 // least one of them does not escape. This means that we can replace
2621 // memmove with memcpy, and we don't need to worry about all manner of
2622 // downsides to splitting and transforming the operations.
2624 // If this doesn't map cleanly onto the alloca type, and that type isn't
2625 // a single value type, just emit a memcpy.
2627 = !VecTy && (BeginOffset != NewAllocaBeginOffset ||
2628 EndOffset != NewAllocaEndOffset ||
2629 !NewAI.getAllocatedType()->isSingleValueType());
2631 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2632 // size hasn't been shrunk based on analysis of the viable range, this is
2634 if (EmitMemCpy && &OldAI == &NewAI) {
2635 uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin;
2636 uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd;
2637 // Ensure the start lines up.
2638 assert(BeginOffset == OrigBegin);
2641 // Rewrite the size as needed.
2642 if (EndOffset != OrigEnd)
2643 II.setLength(ConstantInt::get(II.getLength()->getType(),
2644 EndOffset - BeginOffset));
2647 // Record this instruction for deletion.
2648 if (Pass.DeadSplitInsts.insert(&II))
2649 Pass.DeadInsts.push_back(&II);
2651 bool IsVectorElement = VecTy && (BeginOffset > NewAllocaBeginOffset ||
2652 EndOffset < NewAllocaEndOffset);
2654 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2655 : II.getRawDest()->getType();
2657 OtherPtrTy = IsVectorElement ? VecTy->getElementType()->getPointerTo()
2660 // Compute the other pointer, folding as much as possible to produce
2661 // a single, simple GEP in most cases.
2662 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2663 OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
2664 getName("." + OtherPtr->getName()));
2666 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2667 // alloca that should be re-examined after rewriting this instruction.
2669 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2670 Pass.Worklist.insert(AI);
2674 = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType()
2675 : II.getRawSource()->getType());
2676 Type *SizeTy = II.getLength()->getType();
2677 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2679 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2680 IsDest ? OtherPtr : OurPtr,
2681 Size, Align, II.isVolatile());
2683 DEBUG(dbgs() << " to: " << *New << "\n");
2687 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2688 // is equivalent to 1, but that isn't true if we end up rewriting this as
2693 Value *SrcPtr = OtherPtr;
2694 Value *DstPtr = &NewAI;
2696 std::swap(SrcPtr, DstPtr);
2699 if (IsVectorElement && !IsDest) {
2700 // We have to extract rather than load.
2701 Src = IRB.CreateExtractElement(
2702 IRB.CreateAlignedLoad(SrcPtr, Align, getName(".copyload")),
2703 getIndex(IRB, BeginOffset),
2704 getName(".copyextract"));
2706 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2707 getName(".copyload"));
2710 if (IsVectorElement && IsDest) {
2711 // We have to insert into a loaded copy before storing.
2712 Src = IRB.CreateInsertElement(
2713 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
2714 Src, getIndex(IRB, BeginOffset),
2715 getName(".insert"));
2718 StoreInst *Store = cast<StoreInst>(
2719 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2721 DEBUG(dbgs() << " to: " << *Store << "\n");
2722 return !II.isVolatile();
2725 bool visitIntrinsicInst(IntrinsicInst &II) {
2726 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2727 II.getIntrinsicID() == Intrinsic::lifetime_end);
2728 DEBUG(dbgs() << " original: " << II << "\n");
2729 IRBuilder<> IRB(&II);
2730 assert(II.getArgOperand(1) == OldPtr);
2732 // Record this instruction for deletion.
2733 if (Pass.DeadSplitInsts.insert(&II))
2734 Pass.DeadInsts.push_back(&II);
2737 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2738 EndOffset - BeginOffset);
2739 Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType());
2741 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2742 New = IRB.CreateLifetimeStart(Ptr, Size);
2744 New = IRB.CreateLifetimeEnd(Ptr, Size);
2746 DEBUG(dbgs() << " to: " << *New << "\n");
2750 bool visitPHINode(PHINode &PN) {
2751 DEBUG(dbgs() << " original: " << PN << "\n");
2753 // We would like to compute a new pointer in only one place, but have it be
2754 // as local as possible to the PHI. To do that, we re-use the location of
2755 // the old pointer, which necessarily must be in the right position to
2756 // dominate the PHI.
2757 IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr));
2759 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
2760 // Replace the operands which were using the old pointer.
2761 User::op_iterator OI = PN.op_begin(), OE = PN.op_end();
2762 for (; OI != OE; ++OI)
2766 DEBUG(dbgs() << " to: " << PN << "\n");
2767 deleteIfTriviallyDead(OldPtr);
2771 bool visitSelectInst(SelectInst &SI) {
2772 DEBUG(dbgs() << " original: " << SI << "\n");
2773 IRBuilder<> IRB(&SI);
2775 // Find the operand we need to rewrite here.
2776 bool IsTrueVal = SI.getTrueValue() == OldPtr;
2778 assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!");
2780 assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!");
2782 Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
2783 SI.setOperand(IsTrueVal ? 1 : 2, NewPtr);
2784 DEBUG(dbgs() << " to: " << SI << "\n");
2785 deleteIfTriviallyDead(OldPtr);
2793 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2795 /// This pass aggressively rewrites all aggregate loads and stores on
2796 /// a particular pointer (or any pointer derived from it which we can identify)
2797 /// with scalar loads and stores.
2798 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2799 // Befriend the base class so it can delegate to private visit methods.
2800 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2802 const DataLayout &TD;
2804 /// Queue of pointer uses to analyze and potentially rewrite.
2805 SmallVector<Use *, 8> Queue;
2807 /// Set to prevent us from cycling with phi nodes and loops.
2808 SmallPtrSet<User *, 8> Visited;
2810 /// The current pointer use being rewritten. This is used to dig up the used
2811 /// value (as opposed to the user).
2815 AggLoadStoreRewriter(const DataLayout &TD) : TD(TD) {}
2817 /// Rewrite loads and stores through a pointer and all pointers derived from
2819 bool rewrite(Instruction &I) {
2820 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2822 bool Changed = false;
2823 while (!Queue.empty()) {
2824 U = Queue.pop_back_val();
2825 Changed |= visit(cast<Instruction>(U->getUser()));
2831 /// Enqueue all the users of the given instruction for further processing.
2832 /// This uses a set to de-duplicate users.
2833 void enqueueUsers(Instruction &I) {
2834 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
2836 if (Visited.insert(*UI))
2837 Queue.push_back(&UI.getUse());
2840 // Conservative default is to not rewrite anything.
2841 bool visitInstruction(Instruction &I) { return false; }
2843 /// \brief Generic recursive split emission class.
2844 template <typename Derived>
2847 /// The builder used to form new instructions.
2849 /// The indices which to be used with insert- or extractvalue to select the
2850 /// appropriate value within the aggregate.
2851 SmallVector<unsigned, 4> Indices;
2852 /// The indices to a GEP instruction which will move Ptr to the correct slot
2853 /// within the aggregate.
2854 SmallVector<Value *, 4> GEPIndices;
2855 /// The base pointer of the original op, used as a base for GEPing the
2856 /// split operations.
2859 /// Initialize the splitter with an insertion point, Ptr and start with a
2860 /// single zero GEP index.
2861 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2862 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2865 /// \brief Generic recursive split emission routine.
2867 /// This method recursively splits an aggregate op (load or store) into
2868 /// scalar or vector ops. It splits recursively until it hits a single value
2869 /// and emits that single value operation via the template argument.
2871 /// The logic of this routine relies on GEPs and insertvalue and
2872 /// extractvalue all operating with the same fundamental index list, merely
2873 /// formatted differently (GEPs need actual values).
2875 /// \param Ty The type being split recursively into smaller ops.
2876 /// \param Agg The aggregate value being built up or stored, depending on
2877 /// whether this is splitting a load or a store respectively.
2878 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2879 if (Ty->isSingleValueType())
2880 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2882 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2883 unsigned OldSize = Indices.size();
2885 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2887 assert(Indices.size() == OldSize && "Did not return to the old size");
2888 Indices.push_back(Idx);
2889 GEPIndices.push_back(IRB.getInt32(Idx));
2890 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2891 GEPIndices.pop_back();
2897 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2898 unsigned OldSize = Indices.size();
2900 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2902 assert(Indices.size() == OldSize && "Did not return to the old size");
2903 Indices.push_back(Idx);
2904 GEPIndices.push_back(IRB.getInt32(Idx));
2905 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2906 GEPIndices.pop_back();
2912 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2916 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2917 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2918 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2920 /// Emit a leaf load of a single value. This is called at the leaves of the
2921 /// recursive emission to actually load values.
2922 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2923 assert(Ty->isSingleValueType());
2924 // Load the single value and insert it using the indices.
2925 Value *Load = IRB.CreateLoad(IRB.CreateInBoundsGEP(Ptr, GEPIndices,
2928 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2929 DEBUG(dbgs() << " to: " << *Load << "\n");
2933 bool visitLoadInst(LoadInst &LI) {
2934 assert(LI.getPointerOperand() == *U);
2935 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2938 // We have an aggregate being loaded, split it apart.
2939 DEBUG(dbgs() << " original: " << LI << "\n");
2940 LoadOpSplitter Splitter(&LI, *U);
2941 Value *V = UndefValue::get(LI.getType());
2942 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2943 LI.replaceAllUsesWith(V);
2944 LI.eraseFromParent();
2948 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2949 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2950 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2952 /// Emit a leaf store of a single value. This is called at the leaves of the
2953 /// recursive emission to actually produce stores.
2954 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2955 assert(Ty->isSingleValueType());
2956 // Extract the single value and store it using the indices.
2957 Value *Store = IRB.CreateStore(
2958 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2959 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2961 DEBUG(dbgs() << " to: " << *Store << "\n");
2965 bool visitStoreInst(StoreInst &SI) {
2966 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2968 Value *V = SI.getValueOperand();
2969 if (V->getType()->isSingleValueType())
2972 // We have an aggregate being stored, split it apart.
2973 DEBUG(dbgs() << " original: " << SI << "\n");
2974 StoreOpSplitter Splitter(&SI, *U);
2975 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2976 SI.eraseFromParent();
2980 bool visitBitCastInst(BitCastInst &BC) {
2985 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2990 bool visitPHINode(PHINode &PN) {
2995 bool visitSelectInst(SelectInst &SI) {
3002 /// \brief Try to find a partition of the aggregate type passed in for a given
3003 /// offset and size.
3005 /// This recurses through the aggregate type and tries to compute a subtype
3006 /// based on the offset and size. When the offset and size span a sub-section
3007 /// of an array, it will even compute a new array type for that sub-section,
3008 /// and the same for structs.
3010 /// Note that this routine is very strict and tries to find a partition of the
3011 /// type which produces the *exact* right offset and size. It is not forgiving
3012 /// when the size or offset cause either end of type-based partition to be off.
3013 /// Also, this is a best-effort routine. It is reasonable to give up and not
3014 /// return a type if necessary.
3015 static Type *getTypePartition(const DataLayout &TD, Type *Ty,
3016 uint64_t Offset, uint64_t Size) {
3017 if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size)
3020 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3021 // We can't partition pointers...
3022 if (SeqTy->isPointerTy())
3025 Type *ElementTy = SeqTy->getElementType();
3026 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3027 uint64_t NumSkippedElements = Offset / ElementSize;
3028 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy))
3029 if (NumSkippedElements >= ArrTy->getNumElements())
3031 if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy))
3032 if (NumSkippedElements >= VecTy->getNumElements())
3034 Offset -= NumSkippedElements * ElementSize;
3036 // First check if we need to recurse.
3037 if (Offset > 0 || Size < ElementSize) {
3038 // Bail if the partition ends in a different array element.
3039 if ((Offset + Size) > ElementSize)
3041 // Recurse through the element type trying to peel off offset bytes.
3042 return getTypePartition(TD, ElementTy, Offset, Size);
3044 assert(Offset == 0);
3046 if (Size == ElementSize)
3048 assert(Size > ElementSize);
3049 uint64_t NumElements = Size / ElementSize;
3050 if (NumElements * ElementSize != Size)
3052 return ArrayType::get(ElementTy, NumElements);
3055 StructType *STy = dyn_cast<StructType>(Ty);
3059 const StructLayout *SL = TD.getStructLayout(STy);
3060 if (Offset >= SL->getSizeInBytes())
3062 uint64_t EndOffset = Offset + Size;
3063 if (EndOffset > SL->getSizeInBytes())
3066 unsigned Index = SL->getElementContainingOffset(Offset);
3067 Offset -= SL->getElementOffset(Index);
3069 Type *ElementTy = STy->getElementType(Index);
3070 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3071 if (Offset >= ElementSize)
3072 return 0; // The offset points into alignment padding.
3074 // See if any partition must be contained by the element.
3075 if (Offset > 0 || Size < ElementSize) {
3076 if ((Offset + Size) > ElementSize)
3078 return getTypePartition(TD, ElementTy, Offset, Size);
3080 assert(Offset == 0);
3082 if (Size == ElementSize)
3085 StructType::element_iterator EI = STy->element_begin() + Index,
3086 EE = STy->element_end();
3087 if (EndOffset < SL->getSizeInBytes()) {
3088 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3089 if (Index == EndIndex)
3090 return 0; // Within a single element and its padding.
3092 // Don't try to form "natural" types if the elements don't line up with the
3094 // FIXME: We could potentially recurse down through the last element in the
3095 // sub-struct to find a natural end point.
3096 if (SL->getElementOffset(EndIndex) != EndOffset)
3099 assert(Index < EndIndex);
3100 EE = STy->element_begin() + EndIndex;
3103 // Try to build up a sub-structure.
3104 SmallVector<Type *, 4> ElementTys;
3106 ElementTys.push_back(*EI++);
3108 StructType *SubTy = StructType::get(STy->getContext(), ElementTys,
3110 const StructLayout *SubSL = TD.getStructLayout(SubTy);
3111 if (Size != SubSL->getSizeInBytes())
3112 return 0; // The sub-struct doesn't have quite the size needed.
3117 /// \brief Rewrite an alloca partition's users.
3119 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3120 /// to rewrite uses of an alloca partition to be conducive for SSA value
3121 /// promotion. If the partition needs a new, more refined alloca, this will
3122 /// build that new alloca, preserving as much type information as possible, and
3123 /// rewrite the uses of the old alloca to point at the new one and have the
3124 /// appropriate new offsets. It also evaluates how successful the rewrite was
3125 /// at enabling promotion and if it was successful queues the alloca to be
3127 bool SROA::rewriteAllocaPartition(AllocaInst &AI,
3128 AllocaPartitioning &P,
3129 AllocaPartitioning::iterator PI) {
3130 uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset;
3131 bool IsLive = false;
3132 for (AllocaPartitioning::use_iterator UI = P.use_begin(PI),
3134 UI != UE && !IsLive; ++UI)
3138 return false; // No live uses left of this partition.
3140 DEBUG(dbgs() << "Speculating PHIs and selects in partition "
3141 << "[" << PI->BeginOffset << "," << PI->EndOffset << ")\n");
3143 PHIOrSelectSpeculator Speculator(*TD, P, *this);
3144 DEBUG(dbgs() << " speculating ");
3145 DEBUG(P.print(dbgs(), PI, ""));
3146 Speculator.visitUsers(PI);
3148 // Try to compute a friendly type for this partition of the alloca. This
3149 // won't always succeed, in which case we fall back to a legal integer type
3150 // or an i8 array of an appropriate size.
3152 if (Type *PartitionTy = P.getCommonType(PI))
3153 if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize)
3154 AllocaTy = PartitionTy;
3156 if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(),
3157 PI->BeginOffset, AllocaSize))
3158 AllocaTy = PartitionTy;
3160 (AllocaTy->isArrayTy() &&
3161 AllocaTy->getArrayElementType()->isIntegerTy())) &&
3162 TD->isLegalInteger(AllocaSize * 8))
3163 AllocaTy = Type::getIntNTy(*C, AllocaSize * 8);
3165 AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize);
3166 assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize);
3168 // Check for the case where we're going to rewrite to a new alloca of the
3169 // exact same type as the original, and with the same access offsets. In that
3170 // case, re-use the existing alloca, but still run through the rewriter to
3171 // performe phi and select speculation.
3173 if (AllocaTy == AI.getAllocatedType()) {
3174 assert(PI->BeginOffset == 0 &&
3175 "Non-zero begin offset but same alloca type");
3176 assert(PI == P.begin() && "Begin offset is zero on later partition");
3179 unsigned Alignment = AI.getAlignment();
3181 // The minimum alignment which users can rely on when the explicit
3182 // alignment is omitted or zero is that required by the ABI for this
3184 Alignment = TD->getABITypeAlignment(AI.getAllocatedType());
3186 Alignment = MinAlign(Alignment, PI->BeginOffset);
3187 // If we will get at least this much alignment from the type alone, leave
3188 // the alloca's alignment unconstrained.
3189 if (Alignment <= TD->getABITypeAlignment(AllocaTy))
3191 NewAI = new AllocaInst(AllocaTy, 0, Alignment,
3192 AI.getName() + ".sroa." + Twine(PI - P.begin()),
3197 DEBUG(dbgs() << "Rewriting alloca partition "
3198 << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: "
3201 // Track the high watermark of the post-promotion worklist. We will reset it
3202 // to this point if the alloca is not in fact scheduled for promotion.
3203 unsigned PPWOldSize = PostPromotionWorklist.size();
3205 AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI,
3206 PI->BeginOffset, PI->EndOffset);
3207 DEBUG(dbgs() << " rewriting ");
3208 DEBUG(P.print(dbgs(), PI, ""));
3209 bool Promotable = Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI));
3211 DEBUG(dbgs() << " and queuing for promotion\n");
3212 PromotableAllocas.push_back(NewAI);
3213 } else if (NewAI != &AI) {
3214 // If we can't promote the alloca, iterate on it to check for new
3215 // refinements exposed by splitting the current alloca. Don't iterate on an
3216 // alloca which didn't actually change and didn't get promoted.
3217 Worklist.insert(NewAI);
3220 // Drop any post-promotion work items if promotion didn't happen.
3222 while (PostPromotionWorklist.size() > PPWOldSize)
3223 PostPromotionWorklist.pop_back();
3228 /// \brief Walks the partitioning of an alloca rewriting uses of each partition.
3229 bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
3230 bool Changed = false;
3231 for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE;
3233 Changed |= rewriteAllocaPartition(AI, P, PI);
3238 /// \brief Analyze an alloca for SROA.
3240 /// This analyzes the alloca to ensure we can reason about it, builds
3241 /// a partitioning of the alloca, and then hands it off to be split and
3242 /// rewritten as needed.
3243 bool SROA::runOnAlloca(AllocaInst &AI) {
3244 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3245 ++NumAllocasAnalyzed;
3247 // Special case dead allocas, as they're trivial.
3248 if (AI.use_empty()) {
3249 AI.eraseFromParent();
3253 // Skip alloca forms that this analysis can't handle.
3254 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3255 TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
3258 bool Changed = false;
3260 // First, split any FCA loads and stores touching this alloca to promote
3261 // better splitting and promotion opportunities.
3262 AggLoadStoreRewriter AggRewriter(*TD);
3263 Changed |= AggRewriter.rewrite(AI);
3265 // Build the partition set using a recursive instruction-visiting builder.
3266 AllocaPartitioning P(*TD, AI);
3267 DEBUG(P.print(dbgs()));
3271 // Delete all the dead users of this alloca before splitting and rewriting it.
3272 for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
3273 DE = P.dead_user_end();
3276 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3277 DeadInsts.push_back(*DI);
3279 for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(),
3280 DE = P.dead_op_end();
3283 // Clobber the use with an undef value.
3284 **DO = UndefValue::get(OldV->getType());
3285 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3286 if (isInstructionTriviallyDead(OldI)) {
3288 DeadInsts.push_back(OldI);
3292 // No partitions to split. Leave the dead alloca for a later pass to clean up.
3293 if (P.begin() == P.end())
3296 return splitAlloca(AI, P) || Changed;
3299 /// \brief Delete the dead instructions accumulated in this run.
3301 /// Recursively deletes the dead instructions we've accumulated. This is done
3302 /// at the very end to maximize locality of the recursive delete and to
3303 /// minimize the problems of invalidated instruction pointers as such pointers
3304 /// are used heavily in the intermediate stages of the algorithm.
3306 /// We also record the alloca instructions deleted here so that they aren't
3307 /// subsequently handed to mem2reg to promote.
3308 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3309 DeadSplitInsts.clear();
3310 while (!DeadInsts.empty()) {
3311 Instruction *I = DeadInsts.pop_back_val();
3312 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3314 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3315 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3316 // Zero out the operand and see if it becomes trivially dead.
3318 if (isInstructionTriviallyDead(U))
3319 DeadInsts.push_back(U);
3322 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3323 DeletedAllocas.insert(AI);
3326 I->eraseFromParent();
3330 /// \brief Promote the allocas, using the best available technique.
3332 /// This attempts to promote whatever allocas have been identified as viable in
3333 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3334 /// If there is a domtree available, we attempt to promote using the full power
3335 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3336 /// based on the SSAUpdater utilities. This function returns whether any
3337 /// promotion occured.
3338 bool SROA::promoteAllocas(Function &F) {
3339 if (PromotableAllocas.empty())
3342 NumPromoted += PromotableAllocas.size();
3344 if (DT && !ForceSSAUpdater) {
3345 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3346 PromoteMemToReg(PromotableAllocas, *DT);
3347 PromotableAllocas.clear();
3351 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3353 DIBuilder DIB(*F.getParent());
3354 SmallVector<Instruction*, 64> Insts;
3356 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3357 AllocaInst *AI = PromotableAllocas[Idx];
3358 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
3360 Instruction *I = cast<Instruction>(*UI++);
3361 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3362 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3363 // leading to them) here. Eventually it should use them to optimize the
3364 // scalar values produced.
3365 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
3366 assert(onlyUsedByLifetimeMarkers(I) &&
3367 "Found a bitcast used outside of a lifetime marker.");
3368 while (!I->use_empty())
3369 cast<Instruction>(*I->use_begin())->eraseFromParent();
3370 I->eraseFromParent();
3373 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3374 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3375 II->getIntrinsicID() == Intrinsic::lifetime_end);
3376 II->eraseFromParent();
3382 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3386 PromotableAllocas.clear();
3391 /// \brief A predicate to test whether an alloca belongs to a set.
3392 class IsAllocaInSet {
3393 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3397 typedef AllocaInst *argument_type;
3399 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3400 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3404 bool SROA::runOnFunction(Function &F) {
3405 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3406 C = &F.getContext();
3407 TD = getAnalysisIfAvailable<DataLayout>();
3409 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3412 DT = getAnalysisIfAvailable<DominatorTree>();
3414 BasicBlock &EntryBB = F.getEntryBlock();
3415 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3417 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3418 Worklist.insert(AI);
3420 bool Changed = false;
3421 // A set of deleted alloca instruction pointers which should be removed from
3422 // the list of promotable allocas.
3423 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3426 while (!Worklist.empty()) {
3427 Changed |= runOnAlloca(*Worklist.pop_back_val());
3428 deleteDeadInstructions(DeletedAllocas);
3430 // Remove the deleted allocas from various lists so that we don't try to
3431 // continue processing them.
3432 if (!DeletedAllocas.empty()) {
3433 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3434 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3435 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3436 PromotableAllocas.end(),
3437 IsAllocaInSet(DeletedAllocas)),
3438 PromotableAllocas.end());
3439 DeletedAllocas.clear();
3443 Changed |= promoteAllocas(F);
3445 Worklist = PostPromotionWorklist;
3446 PostPromotionWorklist.clear();
3447 } while (!Worklist.empty());
3452 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3453 if (RequiresDomTree)
3454 AU.addRequired<DominatorTree>();
3455 AU.setPreservesCFG();