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.
137 /// \brief Test whether a partition has been marked as dead.
138 bool isDead() const {
139 if (BeginOffset == UINT64_MAX) {
140 assert(EndOffset == UINT64_MAX);
146 /// \brief Kill a partition.
147 /// This is accomplished by setting both its beginning and end offset to
148 /// the maximum possible value.
150 assert(!isDead() && "He's Dead, Jim!");
151 BeginOffset = EndOffset = UINT64_MAX;
154 Partition() : ByteRange(), IsSplittable() {}
155 Partition(uint64_t BeginOffset, uint64_t EndOffset, bool IsSplittable)
156 : ByteRange(BeginOffset, EndOffset), IsSplittable(IsSplittable) {}
159 /// \brief A particular use of a partition of the alloca.
161 /// This structure is used to associate uses of a partition with it. They
162 /// mark the range of bytes which are referenced by a particular instruction,
163 /// and includes a handle to the user itself and the pointer value in use.
164 /// The bounds of these uses are determined by intersecting the bounds of the
165 /// memory use itself with a particular partition. As a consequence there is
166 /// intentionally overlap between various uses of the same partition.
167 struct PartitionUse : public ByteRange {
168 /// \brief The use in question. Provides access to both user and used value.
170 /// Note that this may be null if the partition use is *dead*, that is, it
171 /// should be ignored.
174 PartitionUse() : ByteRange(), U() {}
175 PartitionUse(uint64_t BeginOffset, uint64_t EndOffset, Use *U)
176 : ByteRange(BeginOffset, EndOffset), U(U) {}
179 /// \brief Construct a partitioning of a particular alloca.
181 /// Construction does most of the work for partitioning the alloca. This
182 /// performs the necessary walks of users and builds a partitioning from it.
183 AllocaPartitioning(const DataLayout &TD, AllocaInst &AI);
185 /// \brief Test whether a pointer to the allocation escapes our analysis.
187 /// If this is true, the partitioning is never fully built and should be
189 bool isEscaped() const { return PointerEscapingInstr; }
191 /// \brief Support for iterating over the partitions.
193 typedef SmallVectorImpl<Partition>::iterator iterator;
194 iterator begin() { return Partitions.begin(); }
195 iterator end() { return Partitions.end(); }
197 typedef SmallVectorImpl<Partition>::const_iterator const_iterator;
198 const_iterator begin() const { return Partitions.begin(); }
199 const_iterator end() const { return Partitions.end(); }
202 /// \brief Support for iterating over and manipulating a particular
203 /// partition's uses.
205 /// The iteration support provided for uses is more limited, but also
206 /// includes some manipulation routines to support rewriting the uses of
207 /// partitions during SROA.
209 typedef SmallVectorImpl<PartitionUse>::iterator use_iterator;
210 use_iterator use_begin(unsigned Idx) { return Uses[Idx].begin(); }
211 use_iterator use_begin(const_iterator I) { return Uses[I - begin()].begin(); }
212 use_iterator use_end(unsigned Idx) { return Uses[Idx].end(); }
213 use_iterator use_end(const_iterator I) { return Uses[I - begin()].end(); }
215 typedef SmallVectorImpl<PartitionUse>::const_iterator const_use_iterator;
216 const_use_iterator use_begin(unsigned Idx) const { return Uses[Idx].begin(); }
217 const_use_iterator use_begin(const_iterator I) const {
218 return Uses[I - begin()].begin();
220 const_use_iterator use_end(unsigned Idx) const { return Uses[Idx].end(); }
221 const_use_iterator use_end(const_iterator I) const {
222 return Uses[I - begin()].end();
225 unsigned use_size(unsigned Idx) const { return Uses[Idx].size(); }
226 unsigned use_size(const_iterator I) const { return Uses[I - begin()].size(); }
227 const PartitionUse &getUse(unsigned PIdx, unsigned UIdx) const {
228 return Uses[PIdx][UIdx];
230 const PartitionUse &getUse(const_iterator I, unsigned UIdx) const {
231 return Uses[I - begin()][UIdx];
234 void use_push_back(unsigned Idx, const PartitionUse &PU) {
235 Uses[Idx].push_back(PU);
237 void use_push_back(const_iterator I, const PartitionUse &PU) {
238 Uses[I - begin()].push_back(PU);
242 /// \brief Allow iterating the dead users for this alloca.
244 /// These are instructions which will never actually use the alloca as they
245 /// are outside the allocated range. They are safe to replace with undef and
248 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
249 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
250 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
253 /// \brief Allow iterating the dead expressions referring to this alloca.
255 /// These are operands which have cannot actually be used to refer to the
256 /// alloca as they are outside its range and the user doesn't correct for
257 /// that. These mostly consist of PHI node inputs and the like which we just
258 /// need to replace with undef.
260 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
261 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
262 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
265 /// \brief MemTransferInst auxiliary data.
266 /// This struct provides some auxiliary data about memory transfer
267 /// intrinsics such as memcpy and memmove. These intrinsics can use two
268 /// different ranges within the same alloca, and provide other challenges to
269 /// correctly represent. We stash extra data to help us untangle this
270 /// after the partitioning is complete.
271 struct MemTransferOffsets {
272 /// The destination begin and end offsets when the destination is within
273 /// this alloca. If the end offset is zero the destination is not within
275 uint64_t DestBegin, DestEnd;
277 /// The source begin and end offsets when the source is within this alloca.
278 /// If the end offset is zero, the source is not within this alloca.
279 uint64_t SourceBegin, SourceEnd;
281 /// Flag for whether an alloca is splittable.
284 MemTransferOffsets getMemTransferOffsets(MemTransferInst &II) const {
285 return MemTransferInstData.lookup(&II);
288 /// \brief Map from a PHI or select operand back to a partition.
290 /// When manipulating PHI nodes or selects, they can use more than one
291 /// partition of an alloca. We store a special mapping to allow finding the
292 /// partition referenced by each of these operands, if any.
293 iterator findPartitionForPHIOrSelectOperand(Use *U) {
294 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
295 = PHIOrSelectOpMap.find(U);
296 if (MapIt == PHIOrSelectOpMap.end())
299 return begin() + MapIt->second.first;
302 /// \brief Map from a PHI or select operand back to the specific use of
305 /// Similar to mapping these operands back to the partitions, this maps
306 /// directly to the use structure of that partition.
307 use_iterator findPartitionUseForPHIOrSelectOperand(Use *U) {
308 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt
309 = PHIOrSelectOpMap.find(U);
310 assert(MapIt != PHIOrSelectOpMap.end());
311 return Uses[MapIt->second.first].begin() + MapIt->second.second;
314 /// \brief Compute a common type among the uses of a particular partition.
316 /// This routines walks all of the uses of a particular partition and tries
317 /// to find a common type between them. Untyped operations such as memset and
318 /// memcpy are ignored.
319 Type *getCommonType(iterator I) const;
321 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
322 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
323 void printUsers(raw_ostream &OS, const_iterator I,
324 StringRef Indent = " ") const;
325 void print(raw_ostream &OS) const;
326 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const;
327 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const;
331 template <typename DerivedT, typename RetT = void> class BuilderBase;
332 class PartitionBuilder;
333 friend class AllocaPartitioning::PartitionBuilder;
335 friend class AllocaPartitioning::UseBuilder;
338 /// \brief Handle to alloca instruction to simplify method interfaces.
342 /// \brief The instruction responsible for this alloca having no partitioning.
344 /// When an instruction (potentially) escapes the pointer to the alloca, we
345 /// store a pointer to that here and abort trying to partition the alloca.
346 /// This will be null if the alloca is partitioned successfully.
347 Instruction *PointerEscapingInstr;
349 /// \brief The partitions of the alloca.
351 /// We store a vector of the partitions over the alloca here. This vector is
352 /// sorted by increasing begin offset, and then by decreasing end offset. See
353 /// the Partition inner class for more details. Initially (during
354 /// construction) there are overlaps, but we form a disjoint sequence of
355 /// partitions while finishing construction and a fully constructed object is
356 /// expected to always have this as a disjoint space.
357 SmallVector<Partition, 8> Partitions;
359 /// \brief The uses of the partitions.
361 /// This is essentially a mapping from each partition to a list of uses of
362 /// that partition. The mapping is done with a Uses vector that has the exact
363 /// same number of entries as the partition vector. Each entry is itself
364 /// a vector of the uses.
365 SmallVector<SmallVector<PartitionUse, 2>, 8> Uses;
367 /// \brief Instructions which will become dead if we rewrite the alloca.
369 /// Note that these are not separated by partition. This is because we expect
370 /// a partitioned alloca to be completely rewritten or not rewritten at all.
371 /// If rewritten, all these instructions can simply be removed and replaced
372 /// with undef as they come from outside of the allocated space.
373 SmallVector<Instruction *, 8> DeadUsers;
375 /// \brief Operands which will become dead if we rewrite the alloca.
377 /// These are operands that in their particular use can be replaced with
378 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
379 /// to PHI nodes and the like. They aren't entirely dead (there might be
380 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
381 /// want to swap this particular input for undef to simplify the use lists of
383 SmallVector<Use *, 8> DeadOperands;
385 /// \brief The underlying storage for auxiliary memcpy and memset info.
386 SmallDenseMap<MemTransferInst *, MemTransferOffsets, 4> MemTransferInstData;
388 /// \brief A side datastructure used when building up the partitions and uses.
390 /// This mapping is only really used during the initial building of the
391 /// partitioning so that we can retain information about PHI and select nodes
393 SmallDenseMap<Instruction *, std::pair<uint64_t, bool> > PHIOrSelectSizes;
395 /// \brief Auxiliary information for particular PHI or select operands.
396 SmallDenseMap<Use *, std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap;
398 /// \brief A utility routine called from the constructor.
400 /// This does what it says on the tin. It is the key of the alloca partition
401 /// splitting and merging. After it is called we have the desired disjoint
402 /// collection of partitions.
403 void splitAndMergePartitions();
407 template <typename DerivedT, typename RetT>
408 class AllocaPartitioning::BuilderBase
409 : public InstVisitor<DerivedT, RetT> {
411 BuilderBase(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
413 AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())),
419 const DataLayout &TD;
420 const uint64_t AllocSize;
421 AllocaPartitioning &P;
423 SmallPtrSet<Use *, 8> VisitedUses;
429 SmallVector<OffsetUse, 8> Queue;
431 // The active offset and use while visiting.
435 void enqueueUsers(Instruction &I, int64_t UserOffset) {
436 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
438 if (VisitedUses.insert(&UI.getUse())) {
439 OffsetUse OU = { &UI.getUse(), UserOffset };
445 bool computeConstantGEPOffset(GetElementPtrInst &GEPI, int64_t &GEPOffset) {
447 for (gep_type_iterator GTI = gep_type_begin(GEPI), GTE = gep_type_end(GEPI);
449 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
455 // Handle a struct index, which adds its field offset to the pointer.
456 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
457 unsigned ElementIdx = OpC->getZExtValue();
458 const StructLayout *SL = TD.getStructLayout(STy);
459 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
460 // Check that we can continue to model this GEP in a signed 64-bit offset.
461 if (ElementOffset > INT64_MAX ||
463 ((uint64_t)GEPOffset + ElementOffset) > INT64_MAX)) {
464 DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding "
465 << "what can be represented in an int64_t!\n"
466 << " alloca: " << P.AI << "\n");
470 GEPOffset = ElementOffset + (uint64_t)-GEPOffset;
472 GEPOffset += ElementOffset;
476 APInt Index = OpC->getValue().sextOrTrunc(TD.getPointerSizeInBits());
477 Index *= APInt(Index.getBitWidth(),
478 TD.getTypeAllocSize(GTI.getIndexedType()));
479 Index += APInt(Index.getBitWidth(), (uint64_t)GEPOffset,
481 // Check if the result can be stored in our int64_t offset.
482 if (!Index.isSignedIntN(sizeof(GEPOffset) * 8)) {
483 DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding "
484 << "what can be represented in an int64_t!\n"
485 << " alloca: " << P.AI << "\n");
489 GEPOffset = Index.getSExtValue();
494 Value *foldSelectInst(SelectInst &SI) {
495 // If the condition being selected on is a constant or the same value is
496 // being selected between, fold the select. Yes this does (rarely) happen
498 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
499 return SI.getOperand(1+CI->isZero());
500 if (SI.getOperand(1) == SI.getOperand(2)) {
501 assert(*U == SI.getOperand(1));
502 return SI.getOperand(1);
508 /// \brief Builder for the alloca partitioning.
510 /// This class builds an alloca partitioning by recursively visiting the uses
511 /// of an alloca and splitting the partitions for each load and store at each
513 class AllocaPartitioning::PartitionBuilder
514 : public BuilderBase<PartitionBuilder, bool> {
515 friend class InstVisitor<PartitionBuilder, bool>;
517 SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap;
520 PartitionBuilder(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
521 : BuilderBase<PartitionBuilder, bool>(TD, AI, P) {}
523 /// \brief Run the builder over the allocation.
525 // Note that we have to re-evaluate size on each trip through the loop as
526 // the queue grows at the tail.
527 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
529 Offset = Queue[Idx].Offset;
530 if (!visit(cast<Instruction>(U->getUser())))
537 bool markAsEscaping(Instruction &I) {
538 P.PointerEscapingInstr = &I;
542 void insertUse(Instruction &I, int64_t Offset, uint64_t Size,
543 bool IsSplittable = false) {
544 // Completely skip uses which have a zero size or don't overlap the
547 (Offset >= 0 && (uint64_t)Offset >= AllocSize) ||
548 (Offset < 0 && (uint64_t)-Offset >= Size)) {
549 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
550 << " which starts past the end of the " << AllocSize
552 << " alloca: " << P.AI << "\n"
553 << " use: " << I << "\n");
557 // Clamp the start to the beginning of the allocation.
559 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
560 << " to start at the beginning of the alloca:\n"
561 << " alloca: " << P.AI << "\n"
562 << " use: " << I << "\n");
563 Size -= (uint64_t)-Offset;
567 uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
569 // Clamp the end offset to the end of the allocation. Note that this is
570 // formulated to handle even the case where "BeginOffset + Size" overflows.
571 assert(AllocSize >= BeginOffset); // Established above.
572 if (Size > AllocSize - BeginOffset) {
573 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
574 << " to remain within the " << AllocSize << " byte alloca:\n"
575 << " alloca: " << P.AI << "\n"
576 << " use: " << I << "\n");
577 EndOffset = AllocSize;
580 Partition New(BeginOffset, EndOffset, IsSplittable);
581 P.Partitions.push_back(New);
584 bool handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset,
586 uint64_t Size = TD.getTypeStoreSize(Ty);
588 // If this memory access can be shown to *statically* extend outside the
589 // bounds of of the allocation, it's behavior is undefined, so simply
590 // ignore it. Note that this is more strict than the generic clamping
591 // behavior of insertUse. We also try to handle cases which might run the
593 // FIXME: We should instead consider the pointer to have escaped if this
594 // function is being instrumented for addressing bugs or race conditions.
595 if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
596 Size > (AllocSize - (uint64_t)Offset)) {
597 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte "
598 << (isa<LoadInst>(I) ? "load" : "store") << " @" << Offset
599 << " which extends past the end of the " << AllocSize
601 << " alloca: " << P.AI << "\n"
602 << " use: " << I << "\n");
606 // We allow splitting of loads and stores where the type is an integer type
607 // and which cover the entire alloca. Such integer loads and stores
608 // often require decomposition into fine grained loads and stores.
609 bool IsSplittable = false;
610 if (IntegerType *ITy = dyn_cast<IntegerType>(Ty))
611 IsSplittable = !IsVolatile && ITy->getBitWidth() == AllocSize*8;
613 insertUse(I, Offset, Size, IsSplittable);
617 bool visitBitCastInst(BitCastInst &BC) {
618 enqueueUsers(BC, Offset);
622 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
624 if (!computeConstantGEPOffset(GEPI, GEPOffset))
625 return markAsEscaping(GEPI);
627 enqueueUsers(GEPI, GEPOffset);
631 bool visitLoadInst(LoadInst &LI) {
632 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
633 "All simple FCA loads should have been pre-split");
634 return handleLoadOrStore(LI.getType(), LI, Offset, LI.isVolatile());
637 bool visitStoreInst(StoreInst &SI) {
638 Value *ValOp = SI.getValueOperand();
640 return markAsEscaping(SI);
642 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
643 "All simple FCA stores should have been pre-split");
644 return handleLoadOrStore(ValOp->getType(), SI, Offset, SI.isVolatile());
648 bool visitMemSetInst(MemSetInst &II) {
649 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
650 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
651 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
652 insertUse(II, Offset, Size, Length);
656 bool visitMemTransferInst(MemTransferInst &II) {
657 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
658 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
660 // Zero-length mem transfer intrinsics can be ignored entirely.
663 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
665 // Only intrinsics with a constant length can be split.
666 Offsets.IsSplittable = Length;
668 if (*U == II.getRawDest()) {
669 Offsets.DestBegin = Offset;
670 Offsets.DestEnd = Offset + Size;
672 if (*U == II.getRawSource()) {
673 Offsets.SourceBegin = Offset;
674 Offsets.SourceEnd = Offset + Size;
677 // If we have set up end offsets for both the source and the destination,
678 // we have found both sides of this transfer pointing at the same alloca.
679 bool SeenBothEnds = Offsets.SourceEnd && Offsets.DestEnd;
680 if (SeenBothEnds && II.getRawDest() != II.getRawSource()) {
681 unsigned PrevIdx = MemTransferPartitionMap[&II];
683 // Check if the begin offsets match and this is a non-volatile transfer.
684 // In that case, we can completely elide the transfer.
685 if (!II.isVolatile() && Offsets.SourceBegin == Offsets.DestBegin) {
686 P.Partitions[PrevIdx].kill();
690 // Otherwise we have an offset transfer within the same alloca. We can't
692 P.Partitions[PrevIdx].IsSplittable = Offsets.IsSplittable = false;
693 } else if (SeenBothEnds) {
694 // Handle the case where this exact use provides both ends of the
696 assert(II.getRawDest() == II.getRawSource());
698 // For non-volatile transfers this is a no-op.
699 if (!II.isVolatile())
702 // Otherwise just suppress splitting.
703 Offsets.IsSplittable = false;
707 // Insert the use now that we've fixed up the splittable nature.
708 insertUse(II, Offset, Size, Offsets.IsSplittable);
710 // Setup the mapping from intrinsic to partition of we've not seen both
711 // ends of this transfer.
713 unsigned NewIdx = P.Partitions.size() - 1;
715 = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx)).second;
717 "Already have intrinsic in map but haven't seen both ends");
724 // Disable SRoA for any intrinsics except for lifetime invariants.
725 // FIXME: What about debug instrinsics? This matches old behavior, but
726 // doesn't make sense.
727 bool visitIntrinsicInst(IntrinsicInst &II) {
728 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
729 II.getIntrinsicID() == Intrinsic::lifetime_end) {
730 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
731 uint64_t Size = std::min(AllocSize - Offset, Length->getLimitedValue());
732 insertUse(II, Offset, Size, true);
736 return markAsEscaping(II);
739 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
740 // We consider any PHI or select that results in a direct load or store of
741 // the same offset to be a viable use for partitioning purposes. These uses
742 // are considered unsplittable and the size is the maximum loaded or stored
744 SmallPtrSet<Instruction *, 4> Visited;
745 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
746 Visited.insert(Root);
747 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
748 // If there are no loads or stores, the access is dead. We mark that as
749 // a size zero access.
752 Instruction *I, *UsedI;
753 llvm::tie(UsedI, I) = Uses.pop_back_val();
755 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
756 Size = std::max(Size, TD.getTypeStoreSize(LI->getType()));
759 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
760 Value *Op = SI->getOperand(0);
763 Size = std::max(Size, TD.getTypeStoreSize(Op->getType()));
767 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
768 if (!GEP->hasAllZeroIndices())
770 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
771 !isa<SelectInst>(I)) {
775 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
777 if (Visited.insert(cast<Instruction>(*UI)))
778 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
779 } while (!Uses.empty());
784 bool visitPHINode(PHINode &PN) {
785 // See if we already have computed info on this node.
786 std::pair<uint64_t, bool> &PHIInfo = P.PHIOrSelectSizes[&PN];
788 PHIInfo.second = true;
789 insertUse(PN, Offset, PHIInfo.first);
793 // Check for an unsafe use of the PHI node.
794 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&PN, PHIInfo.first))
795 return markAsEscaping(*EscapingI);
797 insertUse(PN, Offset, PHIInfo.first);
801 bool visitSelectInst(SelectInst &SI) {
802 if (Value *Result = foldSelectInst(SI)) {
804 // If the result of the constant fold will be the pointer, recurse
805 // through the select as if we had RAUW'ed it.
806 enqueueUsers(SI, Offset);
811 // See if we already have computed info on this node.
812 std::pair<uint64_t, bool> &SelectInfo = P.PHIOrSelectSizes[&SI];
813 if (SelectInfo.first) {
814 SelectInfo.second = true;
815 insertUse(SI, Offset, SelectInfo.first);
819 // Check for an unsafe use of the PHI node.
820 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&SI, SelectInfo.first))
821 return markAsEscaping(*EscapingI);
823 insertUse(SI, Offset, SelectInfo.first);
827 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
828 bool visitInstruction(Instruction &I) { return markAsEscaping(I); }
832 /// \brief Use adder for the alloca partitioning.
834 /// This class adds the uses of an alloca to all of the partitions which they
835 /// use. For splittable partitions, this can end up doing essentially a linear
836 /// walk of the partitions, but the number of steps remains bounded by the
837 /// total result instruction size:
838 /// - The number of partitions is a result of the number unsplittable
839 /// instructions using the alloca.
840 /// - The number of users of each partition is at worst the total number of
841 /// splittable instructions using the alloca.
842 /// Thus we will produce N * M instructions in the end, where N are the number
843 /// of unsplittable uses and M are the number of splittable. This visitor does
844 /// the exact same number of updates to the partitioning.
846 /// In the more common case, this visitor will leverage the fact that the
847 /// partition space is pre-sorted, and do a logarithmic search for the
848 /// partition needed, making the total visit a classical ((N + M) * log(N))
849 /// complexity operation.
850 class AllocaPartitioning::UseBuilder : public BuilderBase<UseBuilder> {
851 friend class InstVisitor<UseBuilder>;
853 /// \brief Set to de-duplicate dead instructions found in the use walk.
854 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
857 UseBuilder(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P)
858 : BuilderBase<UseBuilder>(TD, AI, P) {}
860 /// \brief Run the builder over the allocation.
862 // Note that we have to re-evaluate size on each trip through the loop as
863 // the queue grows at the tail.
864 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
866 Offset = Queue[Idx].Offset;
867 this->visit(cast<Instruction>(U->getUser()));
872 void markAsDead(Instruction &I) {
873 if (VisitedDeadInsts.insert(&I))
874 P.DeadUsers.push_back(&I);
877 void insertUse(Instruction &User, int64_t Offset, uint64_t Size) {
878 // If the use has a zero size or extends outside of the allocation, record
879 // it as a dead use for elimination later.
880 if (Size == 0 || (uint64_t)Offset >= AllocSize ||
881 (Offset < 0 && (uint64_t)-Offset >= Size))
882 return markAsDead(User);
884 // Clamp the start to the beginning of the allocation.
886 Size -= (uint64_t)-Offset;
890 uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
892 // Clamp the end offset to the end of the allocation. Note that this is
893 // formulated to handle even the case where "BeginOffset + Size" overflows.
894 assert(AllocSize >= BeginOffset); // Established above.
895 if (Size > AllocSize - BeginOffset)
896 EndOffset = AllocSize;
898 // NB: This only works if we have zero overlapping partitions.
899 iterator B = std::lower_bound(P.begin(), P.end(), BeginOffset);
900 if (B != P.begin() && llvm::prior(B)->EndOffset > BeginOffset)
902 for (iterator I = B, E = P.end(); I != E && I->BeginOffset < EndOffset;
904 PartitionUse NewPU(std::max(I->BeginOffset, BeginOffset),
905 std::min(I->EndOffset, EndOffset), U);
906 P.use_push_back(I, NewPU);
907 if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser()))
908 P.PHIOrSelectOpMap[U]
909 = std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1);
913 void handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) {
914 uint64_t Size = TD.getTypeStoreSize(Ty);
916 // If this memory access can be shown to *statically* extend outside the
917 // bounds of of the allocation, it's behavior is undefined, so simply
918 // ignore it. Note that this is more strict than the generic clamping
919 // behavior of insertUse.
920 if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
921 Size > (AllocSize - (uint64_t)Offset))
922 return markAsDead(I);
924 insertUse(I, Offset, Size);
927 void visitBitCastInst(BitCastInst &BC) {
929 return markAsDead(BC);
931 enqueueUsers(BC, Offset);
934 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
935 if (GEPI.use_empty())
936 return markAsDead(GEPI);
939 if (!computeConstantGEPOffset(GEPI, GEPOffset))
940 llvm_unreachable("Unable to compute constant offset for use");
942 enqueueUsers(GEPI, GEPOffset);
945 void visitLoadInst(LoadInst &LI) {
946 handleLoadOrStore(LI.getType(), LI, Offset);
949 void visitStoreInst(StoreInst &SI) {
950 handleLoadOrStore(SI.getOperand(0)->getType(), SI, Offset);
953 void visitMemSetInst(MemSetInst &II) {
954 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
955 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
956 insertUse(II, Offset, Size);
959 void visitMemTransferInst(MemTransferInst &II) {
960 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
961 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
963 return markAsDead(II);
965 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
966 if (!II.isVolatile() && Offsets.DestEnd && Offsets.SourceEnd &&
967 Offsets.DestBegin == Offsets.SourceBegin)
968 return markAsDead(II); // Skip identity transfers without side-effects.
970 insertUse(II, Offset, Size);
973 void visitIntrinsicInst(IntrinsicInst &II) {
974 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
975 II.getIntrinsicID() == Intrinsic::lifetime_end);
977 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
978 insertUse(II, Offset,
979 std::min(AllocSize - Offset, Length->getLimitedValue()));
982 void insertPHIOrSelect(Instruction &User, uint64_t Offset) {
983 uint64_t Size = P.PHIOrSelectSizes.lookup(&User).first;
985 // For PHI and select operands outside the alloca, we can't nuke the entire
986 // phi or select -- the other side might still be relevant, so we special
987 // case them here and use a separate structure to track the operands
988 // themselves which should be replaced with undef.
989 if (Offset >= AllocSize) {
990 P.DeadOperands.push_back(U);
994 insertUse(User, Offset, Size);
996 void visitPHINode(PHINode &PN) {
998 return markAsDead(PN);
1000 insertPHIOrSelect(PN, Offset);
1002 void visitSelectInst(SelectInst &SI) {
1004 return markAsDead(SI);
1006 if (Value *Result = foldSelectInst(SI)) {
1008 // If the result of the constant fold will be the pointer, recurse
1009 // through the select as if we had RAUW'ed it.
1010 enqueueUsers(SI, Offset);
1012 // Otherwise the operand to the select is dead, and we can replace it
1014 P.DeadOperands.push_back(U);
1019 insertPHIOrSelect(SI, Offset);
1022 /// \brief Unreachable, we've already visited the alloca once.
1023 void visitInstruction(Instruction &I) {
1024 llvm_unreachable("Unhandled instruction in use builder.");
1028 void AllocaPartitioning::splitAndMergePartitions() {
1029 size_t NumDeadPartitions = 0;
1031 // Track the range of splittable partitions that we pass when accumulating
1032 // overlapping unsplittable partitions.
1033 uint64_t SplitEndOffset = 0ull;
1035 Partition New(0ull, 0ull, false);
1037 for (unsigned i = 0, j = i, e = Partitions.size(); i != e; i = j) {
1040 if (!Partitions[i].IsSplittable || New.BeginOffset == New.EndOffset) {
1041 assert(New.BeginOffset == New.EndOffset);
1042 New = Partitions[i];
1044 assert(New.IsSplittable);
1045 New.EndOffset = std::max(New.EndOffset, Partitions[i].EndOffset);
1047 assert(New.BeginOffset != New.EndOffset);
1049 // Scan the overlapping partitions.
1050 while (j != e && New.EndOffset > Partitions[j].BeginOffset) {
1051 // If the new partition we are forming is splittable, stop at the first
1052 // unsplittable partition.
1053 if (New.IsSplittable && !Partitions[j].IsSplittable)
1056 // Grow the new partition to include any equally splittable range. 'j' is
1057 // always equally splittable when New is splittable, but when New is not
1058 // splittable, we may subsume some (or part of some) splitable partition
1059 // without growing the new one.
1060 if (New.IsSplittable == Partitions[j].IsSplittable) {
1061 New.EndOffset = std::max(New.EndOffset, Partitions[j].EndOffset);
1063 assert(!New.IsSplittable);
1064 assert(Partitions[j].IsSplittable);
1065 SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset);
1068 Partitions[j].kill();
1069 ++NumDeadPartitions;
1073 // If the new partition is splittable, chop off the end as soon as the
1074 // unsplittable subsequent partition starts and ensure we eventually cover
1075 // the splittable area.
1076 if (j != e && New.IsSplittable) {
1077 SplitEndOffset = std::max(SplitEndOffset, New.EndOffset);
1078 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1081 // Add the new partition if it differs from the original one and is
1082 // non-empty. We can end up with an empty partition here if it was
1083 // splittable but there is an unsplittable one that starts at the same
1085 if (New != Partitions[i]) {
1086 if (New.BeginOffset != New.EndOffset)
1087 Partitions.push_back(New);
1088 // Mark the old one for removal.
1089 Partitions[i].kill();
1090 ++NumDeadPartitions;
1093 New.BeginOffset = New.EndOffset;
1094 if (!New.IsSplittable) {
1095 New.EndOffset = std::max(New.EndOffset, SplitEndOffset);
1096 if (j != e && !Partitions[j].IsSplittable)
1097 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1098 New.IsSplittable = true;
1099 // If there is a trailing splittable partition which won't be fused into
1100 // the next splittable partition go ahead and add it onto the partitions
1102 if (New.BeginOffset < New.EndOffset &&
1103 (j == e || !Partitions[j].IsSplittable ||
1104 New.EndOffset < Partitions[j].BeginOffset)) {
1105 Partitions.push_back(New);
1106 New.BeginOffset = New.EndOffset = 0ull;
1111 // Re-sort the partitions now that they have been split and merged into
1112 // disjoint set of partitions. Also remove any of the dead partitions we've
1113 // replaced in the process.
1114 std::sort(Partitions.begin(), Partitions.end());
1115 if (NumDeadPartitions) {
1116 assert(Partitions.back().isDead());
1117 assert((ptrdiff_t)NumDeadPartitions ==
1118 std::count(Partitions.begin(), Partitions.end(), Partitions.back()));
1120 Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end());
1123 AllocaPartitioning::AllocaPartitioning(const DataLayout &TD, AllocaInst &AI)
1128 PointerEscapingInstr(0) {
1129 PartitionBuilder PB(TD, AI, *this);
1133 // Sort the uses. This arranges for the offsets to be in ascending order,
1134 // and the sizes to be in descending order.
1135 std::sort(Partitions.begin(), Partitions.end());
1137 // Remove any partitions from the back which are marked as dead.
1138 while (!Partitions.empty() && Partitions.back().isDead())
1139 Partitions.pop_back();
1141 if (Partitions.size() > 1) {
1142 // Intersect splittability for all partitions with equal offsets and sizes.
1143 // Then remove all but the first so that we have a sequence of non-equal but
1144 // potentially overlapping partitions.
1145 for (iterator I = Partitions.begin(), J = I, E = Partitions.end(); I != E;
1148 while (J != E && *I == *J) {
1149 I->IsSplittable &= J->IsSplittable;
1153 Partitions.erase(std::unique(Partitions.begin(), Partitions.end()),
1156 // Split splittable and merge unsplittable partitions into a disjoint set
1157 // of partitions over the used space of the allocation.
1158 splitAndMergePartitions();
1161 // Now build up the user lists for each of these disjoint partitions by
1162 // re-walking the recursive users of the alloca.
1163 Uses.resize(Partitions.size());
1164 UseBuilder UB(TD, AI, *this);
1168 Type *AllocaPartitioning::getCommonType(iterator I) const {
1170 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
1172 continue; // Skip dead uses.
1173 if (isa<IntrinsicInst>(*UI->U->getUser()))
1175 if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset)
1179 if (LoadInst *LI = dyn_cast<LoadInst>(UI->U->getUser())) {
1180 UserTy = LI->getType();
1181 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI->U->getUser())) {
1182 UserTy = SI->getValueOperand()->getType();
1184 return 0; // Bail if we have weird uses.
1187 if (IntegerType *ITy = dyn_cast<IntegerType>(UserTy)) {
1188 // If the type is larger than the partition, skip it. We only encounter
1189 // this for split integer operations where we want to use the type of the
1190 // entity causing the split.
1191 if (ITy->getBitWidth() > (I->EndOffset - I->BeginOffset)*8)
1194 // If we have found an integer type use covering the alloca, use that
1195 // regardless of the other types, as integers are often used for a "bucket
1200 if (Ty && Ty != UserTy)
1208 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1210 void AllocaPartitioning::print(raw_ostream &OS, const_iterator I,
1211 StringRef Indent) const {
1212 OS << Indent << "partition #" << (I - begin())
1213 << " [" << I->BeginOffset << "," << I->EndOffset << ")"
1214 << (I->IsSplittable ? " (splittable)" : "")
1215 << (Uses[I - begin()].empty() ? " (zero uses)" : "")
1219 void AllocaPartitioning::printUsers(raw_ostream &OS, const_iterator I,
1220 StringRef Indent) const {
1221 for (const_use_iterator UI = use_begin(I), UE = use_end(I);
1224 continue; // Skip dead uses.
1225 OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") "
1226 << "used by: " << *UI->U->getUser() << "\n";
1227 if (MemTransferInst *II = dyn_cast<MemTransferInst>(UI->U->getUser())) {
1228 const MemTransferOffsets &MTO = MemTransferInstData.lookup(II);
1230 if (!MTO.IsSplittable)
1231 IsDest = UI->BeginOffset == MTO.DestBegin;
1233 IsDest = MTO.DestBegin != 0u;
1234 OS << Indent << " (original " << (IsDest ? "dest" : "source") << ": "
1235 << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin)
1236 << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n";
1241 void AllocaPartitioning::print(raw_ostream &OS) const {
1242 if (PointerEscapingInstr) {
1243 OS << "No partitioning for alloca: " << AI << "\n"
1244 << " A pointer to this alloca escaped by:\n"
1245 << " " << *PointerEscapingInstr << "\n";
1249 OS << "Partitioning of alloca: " << AI << "\n";
1251 for (const_iterator I = begin(), E = end(); I != E; ++I, ++Num) {
1257 void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); }
1258 void AllocaPartitioning::dump() const { print(dbgs()); }
1260 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1264 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1266 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1267 /// the loads and stores of an alloca instruction, as well as updating its
1268 /// debug information. This is used when a domtree is unavailable and thus
1269 /// mem2reg in its full form can't be used to handle promotion of allocas to
1271 class AllocaPromoter : public LoadAndStorePromoter {
1275 SmallVector<DbgDeclareInst *, 4> DDIs;
1276 SmallVector<DbgValueInst *, 4> DVIs;
1279 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1280 AllocaInst &AI, DIBuilder &DIB)
1281 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1283 void run(const SmallVectorImpl<Instruction*> &Insts) {
1284 // Remember which alloca we're promoting (for isInstInList).
1285 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
1286 for (Value::use_iterator UI = DebugNode->use_begin(),
1287 UE = DebugNode->use_end();
1289 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
1290 DDIs.push_back(DDI);
1291 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
1292 DVIs.push_back(DVI);
1295 LoadAndStorePromoter::run(Insts);
1296 AI.eraseFromParent();
1297 while (!DDIs.empty())
1298 DDIs.pop_back_val()->eraseFromParent();
1299 while (!DVIs.empty())
1300 DVIs.pop_back_val()->eraseFromParent();
1303 virtual bool isInstInList(Instruction *I,
1304 const SmallVectorImpl<Instruction*> &Insts) const {
1305 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1306 return LI->getOperand(0) == &AI;
1307 return cast<StoreInst>(I)->getPointerOperand() == &AI;
1310 virtual void updateDebugInfo(Instruction *Inst) const {
1311 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
1312 E = DDIs.end(); I != E; ++I) {
1313 DbgDeclareInst *DDI = *I;
1314 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1315 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1316 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1317 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1319 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
1320 E = DVIs.end(); I != E; ++I) {
1321 DbgValueInst *DVI = *I;
1323 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1324 // If an argument is zero extended then use argument directly. The ZExt
1325 // may be zapped by an optimization pass in future.
1326 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1327 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1328 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1329 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1331 Arg = SI->getOperand(0);
1332 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1333 Arg = LI->getOperand(0);
1337 Instruction *DbgVal =
1338 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1340 DbgVal->setDebugLoc(DVI->getDebugLoc());
1344 } // end anon namespace
1348 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1350 /// This pass takes allocations which can be completely analyzed (that is, they
1351 /// don't escape) and tries to turn them into scalar SSA values. There are
1352 /// a few steps to this process.
1354 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1355 /// are used to try to split them into smaller allocations, ideally of
1356 /// a single scalar data type. It will split up memcpy and memset accesses
1357 /// as necessary and try to isolate invidual scalar accesses.
1358 /// 2) It will transform accesses into forms which are suitable for SSA value
1359 /// promotion. This can be replacing a memset with a scalar store of an
1360 /// integer value, or it can involve speculating operations on a PHI or
1361 /// select to be a PHI or select of the results.
1362 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1363 /// onto insert and extract operations on a vector value, and convert them to
1364 /// this form. By doing so, it will enable promotion of vector aggregates to
1365 /// SSA vector values.
1366 class SROA : public FunctionPass {
1367 const bool RequiresDomTree;
1370 const DataLayout *TD;
1373 /// \brief Worklist of alloca instructions to simplify.
1375 /// Each alloca in the function is added to this. Each new alloca formed gets
1376 /// added to it as well to recursively simplify unless that alloca can be
1377 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1378 /// the one being actively rewritten, we add it back onto the list if not
1379 /// already present to ensure it is re-visited.
1380 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
1382 /// \brief A collection of instructions to delete.
1383 /// We try to batch deletions to simplify code and make things a bit more
1385 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts;
1387 /// \brief Post-promotion worklist.
1389 /// Sometimes we discover an alloca which has a high probability of becoming
1390 /// viable for SROA after a round of promotion takes place. In those cases,
1391 /// the alloca is enqueued here for re-processing.
1393 /// Note that we have to be very careful to clear allocas out of this list in
1394 /// the event they are deleted.
1395 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist;
1397 /// \brief A collection of alloca instructions we can directly promote.
1398 std::vector<AllocaInst *> PromotableAllocas;
1401 SROA(bool RequiresDomTree = true)
1402 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
1403 C(0), TD(0), DT(0) {
1404 initializeSROAPass(*PassRegistry::getPassRegistry());
1406 bool runOnFunction(Function &F);
1407 void getAnalysisUsage(AnalysisUsage &AU) const;
1409 const char *getPassName() const { return "SROA"; }
1413 friend class PHIOrSelectSpeculator;
1414 friend class AllocaPartitionRewriter;
1415 friend class AllocaPartitionVectorRewriter;
1417 bool rewriteAllocaPartition(AllocaInst &AI,
1418 AllocaPartitioning &P,
1419 AllocaPartitioning::iterator PI);
1420 bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P);
1421 bool runOnAlloca(AllocaInst &AI);
1422 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
1423 bool promoteAllocas(Function &F);
1429 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1430 return new SROA(RequiresDomTree);
1433 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1435 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
1436 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1440 /// \brief Visitor to speculate PHIs and Selects where possible.
1441 class PHIOrSelectSpeculator : public InstVisitor<PHIOrSelectSpeculator> {
1442 // Befriend the base class so it can delegate to private visit methods.
1443 friend class llvm::InstVisitor<PHIOrSelectSpeculator>;
1445 const DataLayout &TD;
1446 AllocaPartitioning &P;
1450 PHIOrSelectSpeculator(const DataLayout &TD, AllocaPartitioning &P, SROA &Pass)
1451 : TD(TD), P(P), Pass(Pass) {}
1453 /// \brief Visit the users of an alloca partition and rewrite them.
1454 void visitUsers(AllocaPartitioning::const_iterator PI) {
1455 // Note that we need to use an index here as the underlying vector of uses
1456 // may be grown during speculation. However, we never need to re-visit the
1457 // new uses, and so we can use the initial size bound.
1458 for (unsigned Idx = 0, Size = P.use_size(PI); Idx != Size; ++Idx) {
1459 const AllocaPartitioning::PartitionUse &PU = P.getUse(PI, Idx);
1461 continue; // Skip dead use.
1463 visit(cast<Instruction>(PU.U->getUser()));
1468 // By default, skip this instruction.
1469 void visitInstruction(Instruction &I) {}
1471 /// PHI instructions that use an alloca and are subsequently loaded can be
1472 /// rewritten to load both input pointers in the pred blocks and then PHI the
1473 /// results, allowing the load of the alloca to be promoted.
1475 /// %P2 = phi [i32* %Alloca, i32* %Other]
1476 /// %V = load i32* %P2
1478 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1480 /// %V2 = load i32* %Other
1482 /// %V = phi [i32 %V1, i32 %V2]
1484 /// We can do this to a select if its only uses are loads and if the operands
1485 /// to the select can be loaded unconditionally.
1487 /// FIXME: This should be hoisted into a generic utility, likely in
1488 /// Transforms/Util/Local.h
1489 bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
1490 // For now, we can only do this promotion if the load is in the same block
1491 // as the PHI, and if there are no stores between the phi and load.
1492 // TODO: Allow recursive phi users.
1493 // TODO: Allow stores.
1494 BasicBlock *BB = PN.getParent();
1495 unsigned MaxAlign = 0;
1496 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
1498 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1499 if (LI == 0 || !LI->isSimple()) return false;
1501 // For now we only allow loads in the same block as the PHI. This is
1502 // a common case that happens when instcombine merges two loads through
1504 if (LI->getParent() != BB) return false;
1506 // Ensure that there are no instructions between the PHI and the load that
1508 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
1509 if (BBI->mayWriteToMemory())
1512 MaxAlign = std::max(MaxAlign, LI->getAlignment());
1513 Loads.push_back(LI);
1516 // We can only transform this if it is safe to push the loads into the
1517 // predecessor blocks. The only thing to watch out for is that we can't put
1518 // a possibly trapping load in the predecessor if it is a critical edge.
1519 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num;
1521 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
1522 Value *InVal = PN.getIncomingValue(Idx);
1524 // If the value is produced by the terminator of the predecessor (an
1525 // invoke) or it has side-effects, there is no valid place to put a load
1526 // in the predecessor.
1527 if (TI == InVal || TI->mayHaveSideEffects())
1530 // If the predecessor has a single successor, then the edge isn't
1532 if (TI->getNumSuccessors() == 1)
1535 // If this pointer is always safe to load, or if we can prove that there
1536 // is already a load in the block, then we can move the load to the pred
1538 if (InVal->isDereferenceablePointer() ||
1539 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
1548 void visitPHINode(PHINode &PN) {
1549 DEBUG(dbgs() << " original: " << PN << "\n");
1551 SmallVector<LoadInst *, 4> Loads;
1552 if (!isSafePHIToSpeculate(PN, Loads))
1555 assert(!Loads.empty());
1557 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
1558 IRBuilder<> PHIBuilder(&PN);
1559 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(),
1560 PN.getName() + ".sroa.speculated");
1562 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
1563 // matter which one we get and if any differ, it doesn't matter.
1564 LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
1565 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
1566 unsigned Align = SomeLoad->getAlignment();
1568 // Rewrite all loads of the PN to use the new PHI.
1570 LoadInst *LI = Loads.pop_back_val();
1571 LI->replaceAllUsesWith(NewPN);
1572 Pass.DeadInsts.insert(LI);
1573 } while (!Loads.empty());
1575 // Inject loads into all of the pred blocks.
1576 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
1577 BasicBlock *Pred = PN.getIncomingBlock(Idx);
1578 TerminatorInst *TI = Pred->getTerminator();
1579 Use *InUse = &PN.getOperandUse(PN.getOperandNumForIncomingValue(Idx));
1580 Value *InVal = PN.getIncomingValue(Idx);
1581 IRBuilder<> PredBuilder(TI);
1584 = PredBuilder.CreateLoad(InVal, (PN.getName() + ".sroa.speculate.load." +
1586 ++NumLoadsSpeculated;
1587 Load->setAlignment(Align);
1589 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
1590 NewPN->addIncoming(Load, Pred);
1592 Instruction *Ptr = dyn_cast<Instruction>(InVal);
1594 // No uses to rewrite.
1597 // Try to lookup and rewrite any partition uses corresponding to this phi
1599 AllocaPartitioning::iterator PI
1600 = P.findPartitionForPHIOrSelectOperand(InUse);
1604 // Replace the Use in the PartitionUse for this operand with the Use
1606 AllocaPartitioning::use_iterator UI
1607 = P.findPartitionUseForPHIOrSelectOperand(InUse);
1608 assert(isa<PHINode>(*UI->U->getUser()));
1609 UI->U = &Load->getOperandUse(Load->getPointerOperandIndex());
1611 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
1614 /// Select instructions that use an alloca and are subsequently loaded can be
1615 /// rewritten to load both input pointers and then select between the result,
1616 /// allowing the load of the alloca to be promoted.
1618 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
1619 /// %V = load i32* %P2
1621 /// %V1 = load i32* %Alloca -> will be mem2reg'd
1622 /// %V2 = load i32* %Other
1623 /// %V = select i1 %cond, i32 %V1, i32 %V2
1625 /// We can do this to a select if its only uses are loads and if the operand
1626 /// to the select can be loaded unconditionally.
1627 bool isSafeSelectToSpeculate(SelectInst &SI,
1628 SmallVectorImpl<LoadInst *> &Loads) {
1629 Value *TValue = SI.getTrueValue();
1630 Value *FValue = SI.getFalseValue();
1631 bool TDerefable = TValue->isDereferenceablePointer();
1632 bool FDerefable = FValue->isDereferenceablePointer();
1634 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
1636 LoadInst *LI = dyn_cast<LoadInst>(*UI);
1637 if (LI == 0 || !LI->isSimple()) return false;
1639 // Both operands to the select need to be dereferencable, either
1640 // absolutely (e.g. allocas) or at this point because we can see other
1642 if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
1643 LI->getAlignment(), &TD))
1645 if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
1646 LI->getAlignment(), &TD))
1648 Loads.push_back(LI);
1654 void visitSelectInst(SelectInst &SI) {
1655 DEBUG(dbgs() << " original: " << SI << "\n");
1656 IRBuilder<> IRB(&SI);
1658 // If the select isn't safe to speculate, just use simple logic to emit it.
1659 SmallVector<LoadInst *, 4> Loads;
1660 if (!isSafeSelectToSpeculate(SI, Loads))
1663 Use *Ops[2] = { &SI.getOperandUse(1), &SI.getOperandUse(2) };
1664 AllocaPartitioning::iterator PIs[2];
1665 AllocaPartitioning::PartitionUse PUs[2];
1666 for (unsigned i = 0, e = 2; i != e; ++i) {
1667 PIs[i] = P.findPartitionForPHIOrSelectOperand(Ops[i]);
1668 if (PIs[i] != P.end()) {
1669 // If the pointer is within the partitioning, remove the select from
1670 // its uses. We'll add in the new loads below.
1671 AllocaPartitioning::use_iterator UI
1672 = P.findPartitionUseForPHIOrSelectOperand(Ops[i]);
1674 // Clear out the use here so that the offsets into the use list remain
1675 // stable but this use is ignored when rewriting.
1680 Value *TV = SI.getTrueValue();
1681 Value *FV = SI.getFalseValue();
1682 // Replace the loads of the select with a select of two loads.
1683 while (!Loads.empty()) {
1684 LoadInst *LI = Loads.pop_back_val();
1686 IRB.SetInsertPoint(LI);
1688 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true");
1690 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false");
1691 NumLoadsSpeculated += 2;
1693 // Transfer alignment and TBAA info if present.
1694 TL->setAlignment(LI->getAlignment());
1695 FL->setAlignment(LI->getAlignment());
1696 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
1697 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
1698 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
1701 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL,
1702 LI->getName() + ".sroa.speculated");
1704 LoadInst *Loads[2] = { TL, FL };
1705 for (unsigned i = 0, e = 2; i != e; ++i) {
1706 if (PIs[i] != P.end()) {
1707 Use *LoadUse = &Loads[i]->getOperandUse(0);
1708 assert(PUs[i].U->get() == LoadUse->get());
1710 P.use_push_back(PIs[i], PUs[i]);
1714 DEBUG(dbgs() << " speculated to: " << *V << "\n");
1715 LI->replaceAllUsesWith(V);
1716 Pass.DeadInsts.insert(LI);
1722 /// \brief Accumulate the constant offsets in a GEP into a single APInt offset.
1724 /// If the provided GEP is all-constant, the total byte offset formed by the
1725 /// GEP is computed and Offset is set to it. If the GEP has any non-constant
1726 /// operands, the function returns false and the value of Offset is unmodified.
1727 static bool accumulateGEPOffsets(const DataLayout &TD, GEPOperator &GEP,
1729 APInt GEPOffset(Offset.getBitWidth(), 0);
1730 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1731 GTI != GTE; ++GTI) {
1732 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
1735 if (OpC->isZero()) continue;
1737 // Handle a struct index, which adds its field offset to the pointer.
1738 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1739 unsigned ElementIdx = OpC->getZExtValue();
1740 const StructLayout *SL = TD.getStructLayout(STy);
1741 GEPOffset += APInt(Offset.getBitWidth(),
1742 SL->getElementOffset(ElementIdx));
1746 APInt TypeSize(Offset.getBitWidth(),
1747 TD.getTypeAllocSize(GTI.getIndexedType()));
1748 if (VectorType *VTy = dyn_cast<VectorType>(*GTI)) {
1749 assert((VTy->getScalarSizeInBits() % 8) == 0 &&
1750 "vector element size is not a multiple of 8, cannot GEP over it");
1751 TypeSize = VTy->getScalarSizeInBits() / 8;
1754 GEPOffset += OpC->getValue().sextOrTrunc(Offset.getBitWidth()) * TypeSize;
1760 /// \brief Build a GEP out of a base pointer and indices.
1762 /// This will return the BasePtr if that is valid, or build a new GEP
1763 /// instruction using the IRBuilder if GEP-ing is needed.
1764 static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr,
1765 SmallVectorImpl<Value *> &Indices,
1766 const Twine &Prefix) {
1767 if (Indices.empty())
1770 // A single zero index is a no-op, so check for this and avoid building a GEP
1772 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1775 return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx");
1778 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1779 /// TargetTy without changing the offset of the pointer.
1781 /// This routine assumes we've already established a properly offset GEP with
1782 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1783 /// zero-indices down through type layers until we find one the same as
1784 /// TargetTy. If we can't find one with the same type, we at least try to use
1785 /// one with the same size. If none of that works, we just produce the GEP as
1786 /// indicated by Indices to have the correct offset.
1787 static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const DataLayout &TD,
1788 Value *BasePtr, Type *Ty, Type *TargetTy,
1789 SmallVectorImpl<Value *> &Indices,
1790 const Twine &Prefix) {
1792 return buildGEP(IRB, BasePtr, Indices, Prefix);
1794 // See if we can descend into a struct and locate a field with the correct
1796 unsigned NumLayers = 0;
1797 Type *ElementTy = Ty;
1799 if (ElementTy->isPointerTy())
1801 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1802 ElementTy = SeqTy->getElementType();
1803 // Note that we use the default address space as this index is over an
1804 // array or a vector, not a pointer.
1805 Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(0), 0)));
1806 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1807 if (STy->element_begin() == STy->element_end())
1808 break; // Nothing left to descend into.
1809 ElementTy = *STy->element_begin();
1810 Indices.push_back(IRB.getInt32(0));
1815 } while (ElementTy != TargetTy);
1816 if (ElementTy != TargetTy)
1817 Indices.erase(Indices.end() - NumLayers, Indices.end());
1819 return buildGEP(IRB, BasePtr, Indices, Prefix);
1822 /// \brief Recursively compute indices for a natural GEP.
1824 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1825 /// element types adding appropriate indices for the GEP.
1826 static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const DataLayout &TD,
1827 Value *Ptr, Type *Ty, APInt &Offset,
1829 SmallVectorImpl<Value *> &Indices,
1830 const Twine &Prefix) {
1832 return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix);
1834 // We can't recurse through pointer types.
1835 if (Ty->isPointerTy())
1838 // We try to analyze GEPs over vectors here, but note that these GEPs are
1839 // extremely poorly defined currently. The long-term goal is to remove GEPing
1840 // over a vector from the IR completely.
1841 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1842 unsigned ElementSizeInBits = VecTy->getScalarSizeInBits();
1843 if (ElementSizeInBits % 8)
1844 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1845 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1846 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1847 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1849 Offset -= NumSkippedElements * ElementSize;
1850 Indices.push_back(IRB.getInt(NumSkippedElements));
1851 return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
1852 Offset, TargetTy, Indices, Prefix);
1855 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1856 Type *ElementTy = ArrTy->getElementType();
1857 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1858 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1859 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1862 Offset -= NumSkippedElements * ElementSize;
1863 Indices.push_back(IRB.getInt(NumSkippedElements));
1864 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1868 StructType *STy = dyn_cast<StructType>(Ty);
1872 const StructLayout *SL = TD.getStructLayout(STy);
1873 uint64_t StructOffset = Offset.getZExtValue();
1874 if (StructOffset >= SL->getSizeInBytes())
1876 unsigned Index = SL->getElementContainingOffset(StructOffset);
1877 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1878 Type *ElementTy = STy->getElementType(Index);
1879 if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
1880 return 0; // The offset points into alignment padding.
1882 Indices.push_back(IRB.getInt32(Index));
1883 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1887 /// \brief Get a natural GEP from a base pointer to a particular offset and
1888 /// resulting in a particular type.
1890 /// The goal is to produce a "natural" looking GEP that works with the existing
1891 /// composite types to arrive at the appropriate offset and element type for
1892 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1893 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1894 /// Indices, and setting Ty to the result subtype.
1896 /// If no natural GEP can be constructed, this function returns null.
1897 static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const DataLayout &TD,
1898 Value *Ptr, APInt Offset, Type *TargetTy,
1899 SmallVectorImpl<Value *> &Indices,
1900 const Twine &Prefix) {
1901 PointerType *Ty = cast<PointerType>(Ptr->getType());
1903 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1905 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1908 Type *ElementTy = Ty->getElementType();
1909 if (!ElementTy->isSized())
1910 return 0; // We can't GEP through an unsized element.
1911 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1912 if (ElementSize == 0)
1913 return 0; // Zero-length arrays can't help us build a natural GEP.
1914 APInt NumSkippedElements = Offset.sdiv(ElementSize);
1916 Offset -= NumSkippedElements * ElementSize;
1917 Indices.push_back(IRB.getInt(NumSkippedElements));
1918 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1922 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1923 /// resulting pointer has PointerTy.
1925 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1926 /// and produces the pointer type desired. Where it cannot, it will try to use
1927 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1928 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1929 /// bitcast to the type.
1931 /// The strategy for finding the more natural GEPs is to peel off layers of the
1932 /// pointer, walking back through bit casts and GEPs, searching for a base
1933 /// pointer from which we can compute a natural GEP with the desired
1934 /// properities. The algorithm tries to fold as many constant indices into
1935 /// a single GEP as possible, thus making each GEP more independent of the
1936 /// surrounding code.
1937 static Value *getAdjustedPtr(IRBuilder<> &IRB, const DataLayout &TD,
1938 Value *Ptr, APInt Offset, Type *PointerTy,
1939 const Twine &Prefix) {
1940 // Even though we don't look through PHI nodes, we could be called on an
1941 // instruction in an unreachable block, which may be on a cycle.
1942 SmallPtrSet<Value *, 4> Visited;
1943 Visited.insert(Ptr);
1944 SmallVector<Value *, 4> Indices;
1946 // We may end up computing an offset pointer that has the wrong type. If we
1947 // never are able to compute one directly that has the correct type, we'll
1948 // fall back to it, so keep it around here.
1949 Value *OffsetPtr = 0;
1951 // Remember any i8 pointer we come across to re-use if we need to do a raw
1954 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1956 Type *TargetTy = PointerTy->getPointerElementType();
1959 // First fold any existing GEPs into the offset.
1960 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1961 APInt GEPOffset(Offset.getBitWidth(), 0);
1962 if (!accumulateGEPOffsets(TD, *GEP, GEPOffset))
1964 Offset += GEPOffset;
1965 Ptr = GEP->getPointerOperand();
1966 if (!Visited.insert(Ptr))
1970 // See if we can perform a natural GEP here.
1972 if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
1974 if (P->getType() == PointerTy) {
1975 // Zap any offset pointer that we ended up computing in previous rounds.
1976 if (OffsetPtr && OffsetPtr->use_empty())
1977 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1978 I->eraseFromParent();
1986 // Stash this pointer if we've found an i8*.
1987 if (Ptr->getType()->isIntegerTy(8)) {
1989 Int8PtrOffset = Offset;
1992 // Peel off a layer of the pointer and update the offset appropriately.
1993 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1994 Ptr = cast<Operator>(Ptr)->getOperand(0);
1995 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1996 if (GA->mayBeOverridden())
1998 Ptr = GA->getAliasee();
2002 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
2003 } while (Visited.insert(Ptr));
2007 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
2008 Prefix + ".raw_cast");
2009 Int8PtrOffset = Offset;
2012 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
2013 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
2014 Prefix + ".raw_idx");
2018 // On the off chance we were targeting i8*, guard the bitcast here.
2019 if (Ptr->getType() != PointerTy)
2020 Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast");
2025 /// \brief Test whether we can convert a value from the old to the new type.
2027 /// This predicate should be used to guard calls to convertValue in order to
2028 /// ensure that we only try to convert viable values. The strategy is that we
2029 /// will peel off single element struct and array wrappings to get to an
2030 /// underlying value, and convert that value.
2031 static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) {
2034 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy))
2036 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType())
2039 if (NewTy->isPointerTy() || OldTy->isPointerTy()) {
2040 if (NewTy->isPointerTy() && OldTy->isPointerTy())
2042 if (NewTy->isIntegerTy() || OldTy->isIntegerTy())
2050 /// \brief Generic routine to convert an SSA value to a value of a different
2053 /// This will try various different casting techniques, such as bitcasts,
2054 /// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test
2055 /// two types for viability with this routine.
2056 static Value *convertValue(const DataLayout &DL, IRBuilder<> &IRB, Value *V,
2058 assert(canConvertValue(DL, V->getType(), Ty) &&
2059 "Value not convertable to type");
2060 if (V->getType() == Ty)
2062 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
2063 return IRB.CreateIntToPtr(V, Ty);
2064 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
2065 return IRB.CreatePtrToInt(V, Ty);
2067 return IRB.CreateBitCast(V, Ty);
2070 /// \brief Test whether the given alloca partition can be promoted to a vector.
2072 /// This is a quick test to check whether we can rewrite a particular alloca
2073 /// partition (and its newly formed alloca) into a vector alloca with only
2074 /// whole-vector loads and stores such that it could be promoted to a vector
2075 /// SSA value. We only can ensure this for a limited set of operations, and we
2076 /// don't want to do the rewrites unless we are confident that the result will
2077 /// be promotable, so we have an early test here.
2078 static bool isVectorPromotionViable(const DataLayout &TD,
2080 AllocaPartitioning &P,
2081 uint64_t PartitionBeginOffset,
2082 uint64_t PartitionEndOffset,
2083 AllocaPartitioning::const_use_iterator I,
2084 AllocaPartitioning::const_use_iterator E) {
2085 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
2089 uint64_t VecSize = TD.getTypeSizeInBits(Ty);
2090 uint64_t ElementSize = Ty->getScalarSizeInBits();
2092 // While the definition of LLVM vectors is bitpacked, we don't support sizes
2093 // that aren't byte sized.
2094 if (ElementSize % 8)
2096 assert((VecSize % 8) == 0 && "vector size not a multiple of element size?");
2100 for (; I != E; ++I) {
2102 continue; // Skip dead use.
2104 uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
2105 uint64_t BeginIndex = BeginOffset / ElementSize;
2106 if (BeginIndex * ElementSize != BeginOffset ||
2107 BeginIndex >= Ty->getNumElements())
2109 uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
2110 uint64_t EndIndex = EndOffset / ElementSize;
2111 if (EndIndex * ElementSize != EndOffset ||
2112 EndIndex > Ty->getNumElements())
2115 // FIXME: We should build shuffle vector instructions to handle
2116 // non-element-sized accesses. See PR14055 for an example of where this
2118 if ((EndOffset - BeginOffset) != ElementSize &&
2119 (EndOffset - BeginOffset) != VecSize)
2122 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)
2131 } else if (I->U->get()->getType()->getPointerElementType()->isStructTy()) {
2132 // Disable vector promotion when there are loads or stores of an FCA.
2134 } else if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) {
2135 if (LI->isVolatile())
2137 } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) {
2138 if (SI->isVolatile())
2147 /// \brief Test whether the given alloca partition's integer operations can be
2148 /// widened to promotable ones.
2150 /// This is a quick test to check whether we can rewrite the integer loads and
2151 /// stores to a particular alloca into wider loads and stores and be able to
2152 /// promote the resulting alloca.
2153 static bool isIntegerWideningViable(const DataLayout &TD,
2155 uint64_t AllocBeginOffset,
2156 AllocaPartitioning &P,
2157 AllocaPartitioning::const_use_iterator I,
2158 AllocaPartitioning::const_use_iterator E) {
2159 uint64_t SizeInBits = TD.getTypeSizeInBits(AllocaTy);
2161 // Don't try to handle allocas with bit-padding.
2162 if (SizeInBits != TD.getTypeStoreSizeInBits(AllocaTy))
2165 // We need to ensure that an integer type with the appropriate bitwidth can
2166 // be converted to the alloca type, whatever that is. We don't want to force
2167 // the alloca itself to have an integer type if there is a more suitable one.
2168 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits);
2169 if (!canConvertValue(TD, AllocaTy, IntTy) ||
2170 !canConvertValue(TD, IntTy, AllocaTy))
2173 uint64_t Size = TD.getTypeStoreSize(AllocaTy);
2175 // Check the uses to ensure the uses are (likely) promoteable integer uses.
2176 // Also ensure that the alloca has a covering load or store. We don't want
2177 // to widen the integer operotains only to fail to promote due to some other
2178 // unsplittable entry (which we may make splittable later).
2179 bool WholeAllocaOp = false;
2180 for (; I != E; ++I) {
2182 continue; // Skip dead use.
2184 uint64_t RelBegin = I->BeginOffset - AllocBeginOffset;
2185 uint64_t RelEnd = I->EndOffset - AllocBeginOffset;
2187 // We can't reasonably handle cases where the load or store extends past
2188 // the end of the aloca's type and into its padding.
2192 if (LoadInst *LI = dyn_cast<LoadInst>(I->U->getUser())) {
2193 if (LI->isVolatile())
2195 if (RelBegin == 0 && RelEnd == Size)
2196 WholeAllocaOp = true;
2197 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
2198 if (ITy->getBitWidth() < TD.getTypeStoreSize(ITy))
2202 // Non-integer loads need to be convertible from the alloca type so that
2203 // they are promotable.
2204 if (RelBegin != 0 || RelEnd != Size ||
2205 !canConvertValue(TD, AllocaTy, LI->getType()))
2207 } else if (StoreInst *SI = dyn_cast<StoreInst>(I->U->getUser())) {
2208 Type *ValueTy = SI->getValueOperand()->getType();
2209 if (SI->isVolatile())
2211 if (RelBegin == 0 && RelEnd == Size)
2212 WholeAllocaOp = true;
2213 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) {
2214 if (ITy->getBitWidth() < TD.getTypeStoreSize(ITy))
2218 // Non-integer stores need to be convertible to the alloca type so that
2219 // they are promotable.
2220 if (RelBegin != 0 || RelEnd != Size ||
2221 !canConvertValue(TD, ValueTy, AllocaTy))
2223 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I->U->getUser())) {
2224 if (MI->isVolatile())
2226 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(I->U->getUser())) {
2227 const AllocaPartitioning::MemTransferOffsets &MTO
2228 = P.getMemTransferOffsets(*MTI);
2229 if (!MTO.IsSplittable)
2232 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I->U->getUser())) {
2233 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
2234 II->getIntrinsicID() != Intrinsic::lifetime_end)
2240 return WholeAllocaOp;
2243 static Value *extractInteger(const DataLayout &DL, IRBuilder<> &IRB, Value *V,
2244 IntegerType *Ty, uint64_t Offset,
2245 const Twine &Name) {
2246 DEBUG(dbgs() << " start: " << *V << "\n");
2247 IntegerType *IntTy = cast<IntegerType>(V->getType());
2248 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2249 "Element extends past full value");
2250 uint64_t ShAmt = 8*Offset;
2251 if (DL.isBigEndian())
2252 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2254 V = IRB.CreateLShr(V, ShAmt, Name + ".shift");
2255 DEBUG(dbgs() << " shifted: " << *V << "\n");
2257 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2258 "Cannot extract to a larger integer!");
2260 V = IRB.CreateTrunc(V, Ty, Name + ".trunc");
2261 DEBUG(dbgs() << " trunced: " << *V << "\n");
2266 static Value *insertInteger(const DataLayout &DL, IRBuilder<> &IRB, Value *Old,
2267 Value *V, uint64_t Offset, const Twine &Name) {
2268 IntegerType *IntTy = cast<IntegerType>(Old->getType());
2269 IntegerType *Ty = cast<IntegerType>(V->getType());
2270 assert(Ty->getBitWidth() <= IntTy->getBitWidth() &&
2271 "Cannot insert a larger integer!");
2272 DEBUG(dbgs() << " start: " << *V << "\n");
2274 V = IRB.CreateZExt(V, IntTy, Name + ".ext");
2275 DEBUG(dbgs() << " extended: " << *V << "\n");
2277 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) &&
2278 "Element store outside of alloca store");
2279 uint64_t ShAmt = 8*Offset;
2280 if (DL.isBigEndian())
2281 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset);
2283 V = IRB.CreateShl(V, ShAmt, Name + ".shift");
2284 DEBUG(dbgs() << " shifted: " << *V << "\n");
2287 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) {
2288 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt);
2289 Old = IRB.CreateAnd(Old, Mask, Name + ".mask");
2290 DEBUG(dbgs() << " masked: " << *Old << "\n");
2291 V = IRB.CreateOr(Old, V, Name + ".insert");
2292 DEBUG(dbgs() << " inserted: " << *V << "\n");
2298 /// \brief Visitor to rewrite instructions using a partition of an alloca to
2299 /// use a new alloca.
2301 /// Also implements the rewriting to vector-based accesses when the partition
2302 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
2304 class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
2306 // Befriend the base class so it can delegate to private visit methods.
2307 friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
2309 const DataLayout &TD;
2310 AllocaPartitioning &P;
2312 AllocaInst &OldAI, &NewAI;
2313 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
2316 // If we are rewriting an alloca partition which can be written as pure
2317 // vector operations, we stash extra information here. When VecTy is
2318 // non-null, we have some strict guarantees about the rewriten alloca:
2319 // - The new alloca is exactly the size of the vector type here.
2320 // - The accesses all either map to the entire vector or to a single
2322 // - The set of accessing instructions is only one of those handled above
2323 // in isVectorPromotionViable. Generally these are the same access kinds
2324 // which are promotable via mem2reg.
2327 uint64_t ElementSize;
2329 // This is a convenience and flag variable that will be null unless the new
2330 // alloca's integer operations should be widened to this integer type due to
2331 // passing isIntegerWideningViable above. If it is non-null, the desired
2332 // integer type will be stored here for easy access during rewriting.
2335 // The offset of the partition user currently being rewritten.
2336 uint64_t BeginOffset, EndOffset;
2338 Instruction *OldPtr;
2340 // The name prefix to use when rewriting instructions for this alloca.
2341 std::string NamePrefix;
2344 AllocaPartitionRewriter(const DataLayout &TD, AllocaPartitioning &P,
2345 AllocaPartitioning::iterator PI,
2346 SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
2347 uint64_t NewBeginOffset, uint64_t NewEndOffset)
2348 : TD(TD), P(P), Pass(Pass),
2349 OldAI(OldAI), NewAI(NewAI),
2350 NewAllocaBeginOffset(NewBeginOffset),
2351 NewAllocaEndOffset(NewEndOffset),
2352 NewAllocaTy(NewAI.getAllocatedType()),
2353 VecTy(), ElementTy(), ElementSize(), IntTy(),
2354 BeginOffset(), EndOffset() {
2357 /// \brief Visit the users of the alloca partition and rewrite them.
2358 bool visitUsers(AllocaPartitioning::const_use_iterator I,
2359 AllocaPartitioning::const_use_iterator E) {
2360 if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P,
2361 NewAllocaBeginOffset, NewAllocaEndOffset,
2364 VecTy = cast<VectorType>(NewAI.getAllocatedType());
2365 ElementTy = VecTy->getElementType();
2366 assert((VecTy->getScalarSizeInBits() % 8) == 0 &&
2367 "Only multiple-of-8 sized vector elements are viable");
2368 ElementSize = VecTy->getScalarSizeInBits() / 8;
2369 } else if (isIntegerWideningViable(TD, NewAI.getAllocatedType(),
2370 NewAllocaBeginOffset, P, I, E)) {
2371 IntTy = Type::getIntNTy(NewAI.getContext(),
2372 TD.getTypeSizeInBits(NewAI.getAllocatedType()));
2374 bool CanSROA = true;
2375 for (; I != E; ++I) {
2377 continue; // Skip dead uses.
2378 BeginOffset = I->BeginOffset;
2379 EndOffset = I->EndOffset;
2381 OldPtr = cast<Instruction>(I->U->get());
2382 NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str();
2383 CanSROA &= visit(cast<Instruction>(I->U->getUser()));
2399 // Every instruction which can end up as a user must have a rewrite rule.
2400 bool visitInstruction(Instruction &I) {
2401 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
2402 llvm_unreachable("No rewrite rule for this instruction!");
2405 Twine getName(const Twine &Suffix) {
2406 return NamePrefix + Suffix;
2409 Value *getAdjustedAllocaPtr(IRBuilder<> &IRB, Type *PointerTy) {
2410 assert(BeginOffset >= NewAllocaBeginOffset);
2411 APInt Offset(TD.getPointerSizeInBits(), BeginOffset - NewAllocaBeginOffset);
2412 return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName(""));
2415 /// \brief Compute suitable alignment to access an offset into the new alloca.
2416 unsigned getOffsetAlign(uint64_t Offset) {
2417 unsigned NewAIAlign = NewAI.getAlignment();
2419 NewAIAlign = TD.getABITypeAlignment(NewAI.getAllocatedType());
2420 return MinAlign(NewAIAlign, Offset);
2423 /// \brief Compute suitable alignment to access this partition of the new
2425 unsigned getPartitionAlign() {
2426 return getOffsetAlign(BeginOffset - NewAllocaBeginOffset);
2429 /// \brief Compute suitable alignment to access a type at an offset of the
2432 /// \returns zero if the type's ABI alignment is a suitable alignment,
2433 /// otherwise returns the maximal suitable alignment.
2434 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) {
2435 unsigned Align = getOffsetAlign(Offset);
2436 return Align == TD.getABITypeAlignment(Ty) ? 0 : Align;
2439 /// \brief Compute suitable alignment to access a type at the beginning of
2440 /// this partition of the new alloca.
2442 /// See \c getOffsetTypeAlign for details; this routine delegates to it.
2443 unsigned getPartitionTypeAlign(Type *Ty) {
2444 return getOffsetTypeAlign(Ty, BeginOffset - NewAllocaBeginOffset);
2447 ConstantInt *getIndex(IRBuilder<> &IRB, uint64_t Offset) {
2448 assert(VecTy && "Can only call getIndex when rewriting a vector");
2449 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
2450 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
2451 uint32_t Index = RelOffset / ElementSize;
2452 assert(Index * ElementSize == RelOffset);
2453 return IRB.getInt32(Index);
2456 void deleteIfTriviallyDead(Value *V) {
2457 Instruction *I = cast<Instruction>(V);
2458 if (isInstructionTriviallyDead(I))
2459 Pass.DeadInsts.insert(I);
2462 Value *rewriteVectorizedLoadInst(IRBuilder<> &IRB, LoadInst &LI, Value *OldOp) {
2463 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2465 if (LI.getType() == VecTy->getElementType() ||
2466 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
2467 V = IRB.CreateExtractElement(V, getIndex(IRB, BeginOffset),
2468 getName(".extract"));
2473 Value *rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) {
2474 assert(IntTy && "We cannot insert an integer to the alloca");
2475 assert(!LI.isVolatile());
2476 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2478 V = convertValue(TD, IRB, V, IntTy);
2479 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2480 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2481 if (Offset > 0 || EndOffset < NewAllocaEndOffset)
2482 V = extractInteger(TD, IRB, V, cast<IntegerType>(LI.getType()), Offset,
2483 getName(".extract"));
2487 bool visitLoadInst(LoadInst &LI) {
2488 DEBUG(dbgs() << " original: " << LI << "\n");
2489 Value *OldOp = LI.getOperand(0);
2490 assert(OldOp == OldPtr);
2491 IRBuilder<> IRB(&LI);
2493 uint64_t Size = EndOffset - BeginOffset;
2494 bool IsSplitIntLoad = Size < TD.getTypeStoreSize(LI.getType());
2495 Type *TargetTy = IsSplitIntLoad ? Type::getIntNTy(LI.getContext(), Size * 8)
2497 bool IsPtrAdjusted = false;
2500 V = rewriteVectorizedLoadInst(IRB, LI, OldOp);
2501 } else if (IntTy && LI.getType()->isIntegerTy()) {
2502 V = rewriteIntegerLoad(IRB, LI);
2503 } else if (BeginOffset == NewAllocaBeginOffset &&
2504 canConvertValue(TD, NewAllocaTy, LI.getType())) {
2505 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2506 LI.isVolatile(), getName(".load"));
2508 Type *LTy = TargetTy->getPointerTo();
2509 V = IRB.CreateAlignedLoad(getAdjustedAllocaPtr(IRB, LTy),
2510 getPartitionTypeAlign(TargetTy),
2511 LI.isVolatile(), getName(".load"));
2512 IsPtrAdjusted = true;
2514 V = convertValue(TD, IRB, V, TargetTy);
2516 if (IsSplitIntLoad) {
2517 assert(!LI.isVolatile());
2518 assert(LI.getType()->isIntegerTy() &&
2519 "Only integer type loads and stores are split");
2520 assert(LI.getType()->getIntegerBitWidth() ==
2521 TD.getTypeStoreSizeInBits(LI.getType()) &&
2522 "Non-byte-multiple bit width");
2523 assert(LI.getType()->getIntegerBitWidth() ==
2524 TD.getTypeAllocSizeInBits(OldAI.getAllocatedType()) &&
2525 "Only alloca-wide loads can be split and recomposed");
2526 // Move the insertion point just past the load so that we can refer to it.
2527 IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI)));
2528 // Create a placeholder value with the same type as LI to use as the
2529 // basis for the new value. This allows us to replace the uses of LI with
2530 // the computed value, and then replace the placeholder with LI, leaving
2531 // LI only used for this computation.
2533 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo()));
2534 V = insertInteger(TD, IRB, Placeholder, V, BeginOffset,
2535 getName(".insert"));
2536 LI.replaceAllUsesWith(V);
2537 Placeholder->replaceAllUsesWith(&LI);
2540 LI.replaceAllUsesWith(V);
2543 Pass.DeadInsts.insert(&LI);
2544 deleteIfTriviallyDead(OldOp);
2545 DEBUG(dbgs() << " to: " << *V << "\n");
2546 return !LI.isVolatile() && !IsPtrAdjusted;
2549 bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, Value *V,
2550 StoreInst &SI, Value *OldOp) {
2551 if (V->getType() == ElementTy ||
2552 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
2553 if (V->getType() != ElementTy)
2554 V = convertValue(TD, IRB, V, ElementTy);
2555 LoadInst *LI = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2557 V = IRB.CreateInsertElement(LI, V, getIndex(IRB, BeginOffset),
2558 getName(".insert"));
2559 } else if (V->getType() != VecTy) {
2560 V = convertValue(TD, IRB, V, VecTy);
2562 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2563 Pass.DeadInsts.insert(&SI);
2566 DEBUG(dbgs() << " to: " << *Store << "\n");
2570 bool rewriteIntegerStore(IRBuilder<> &IRB, Value *V, StoreInst &SI) {
2571 assert(IntTy && "We cannot extract an integer from the alloca");
2572 assert(!SI.isVolatile());
2573 if (TD.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) {
2574 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2575 getName(".oldload"));
2576 Old = convertValue(TD, IRB, Old, IntTy);
2577 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2578 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2579 V = insertInteger(TD, IRB, Old, SI.getValueOperand(), Offset,
2580 getName(".insert"));
2582 V = convertValue(TD, IRB, V, NewAllocaTy);
2583 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment());
2584 Pass.DeadInsts.insert(&SI);
2586 DEBUG(dbgs() << " to: " << *Store << "\n");
2590 bool visitStoreInst(StoreInst &SI) {
2591 DEBUG(dbgs() << " original: " << SI << "\n");
2592 Value *OldOp = SI.getOperand(1);
2593 assert(OldOp == OldPtr);
2594 IRBuilder<> IRB(&SI);
2596 Value *V = SI.getValueOperand();
2598 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2599 // alloca that should be re-examined after promoting this alloca.
2600 if (V->getType()->isPointerTy())
2601 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets()))
2602 Pass.PostPromotionWorklist.insert(AI);
2604 uint64_t Size = EndOffset - BeginOffset;
2605 if (Size < TD.getTypeStoreSize(V->getType())) {
2606 assert(!SI.isVolatile());
2607 assert(V->getType()->isIntegerTy() &&
2608 "Only integer type loads and stores are split");
2609 assert(V->getType()->getIntegerBitWidth() ==
2610 TD.getTypeStoreSizeInBits(V->getType()) &&
2611 "Non-byte-multiple bit width");
2612 assert(V->getType()->getIntegerBitWidth() ==
2613 TD.getTypeSizeInBits(OldAI.getAllocatedType()) &&
2614 "Only alloca-wide stores can be split and recomposed");
2615 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8);
2616 V = extractInteger(TD, IRB, V, NarrowTy, BeginOffset,
2617 getName(".extract"));
2621 return rewriteVectorizedStoreInst(IRB, V, SI, OldOp);
2622 if (IntTy && V->getType()->isIntegerTy())
2623 return rewriteIntegerStore(IRB, V, SI);
2626 if (BeginOffset == NewAllocaBeginOffset &&
2627 canConvertValue(TD, V->getType(), NewAllocaTy)) {
2628 V = convertValue(TD, IRB, V, NewAllocaTy);
2629 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2632 Value *NewPtr = getAdjustedAllocaPtr(IRB, V->getType()->getPointerTo());
2633 NewSI = IRB.CreateAlignedStore(V, NewPtr,
2634 getPartitionTypeAlign(V->getType()),
2638 Pass.DeadInsts.insert(&SI);
2639 deleteIfTriviallyDead(OldOp);
2641 DEBUG(dbgs() << " to: " << *NewSI << "\n");
2642 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile();
2645 bool visitMemSetInst(MemSetInst &II) {
2646 DEBUG(dbgs() << " original: " << II << "\n");
2647 IRBuilder<> IRB(&II);
2648 assert(II.getRawDest() == OldPtr);
2650 // If the memset has a variable size, it cannot be split, just adjust the
2651 // pointer to the new alloca.
2652 if (!isa<Constant>(II.getLength())) {
2653 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2654 Type *CstTy = II.getAlignmentCst()->getType();
2655 II.setAlignment(ConstantInt::get(CstTy, getPartitionAlign()));
2657 deleteIfTriviallyDead(OldPtr);
2661 // Record this instruction for deletion.
2662 Pass.DeadInsts.insert(&II);
2664 Type *AllocaTy = NewAI.getAllocatedType();
2665 Type *ScalarTy = AllocaTy->getScalarType();
2667 // If this doesn't map cleanly onto the alloca type, and that type isn't
2668 // a single value type, just emit a memset.
2669 if (!VecTy && !IntTy &&
2670 (BeginOffset != NewAllocaBeginOffset ||
2671 EndOffset != NewAllocaEndOffset ||
2672 !AllocaTy->isSingleValueType() ||
2673 !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)))) {
2674 Type *SizeTy = II.getLength()->getType();
2675 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2677 = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB,
2678 II.getRawDest()->getType()),
2679 II.getValue(), Size, getPartitionAlign(),
2682 DEBUG(dbgs() << " to: " << *New << "\n");
2686 // If we can represent this as a simple value, we have to build the actual
2687 // value to store, which requires expanding the byte present in memset to
2688 // a sensible representation for the alloca type. This is essentially
2689 // splatting the byte to a sufficiently wide integer, bitcasting to the
2690 // desired scalar type, and splatting it across any desired vector type.
2691 uint64_t Size = EndOffset - BeginOffset;
2692 Value *V = II.getValue();
2693 IntegerType *VTy = cast<IntegerType>(V->getType());
2694 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8);
2695 if (Size*8 > VTy->getBitWidth())
2696 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, getName(".zext")),
2697 ConstantExpr::getUDiv(
2698 Constant::getAllOnesValue(SplatIntTy),
2699 ConstantExpr::getZExt(
2700 Constant::getAllOnesValue(V->getType()),
2702 getName(".isplat"));
2704 // If this is an element-wide memset of a vectorizable alloca, insert it.
2705 if (VecTy && (BeginOffset > NewAllocaBeginOffset ||
2706 EndOffset < NewAllocaEndOffset)) {
2707 if (V->getType() != ScalarTy)
2708 V = convertValue(TD, IRB, V, ScalarTy);
2709 StoreInst *Store = IRB.CreateAlignedStore(
2710 IRB.CreateInsertElement(IRB.CreateAlignedLoad(&NewAI,
2711 NewAI.getAlignment(),
2713 V, getIndex(IRB, BeginOffset),
2714 getName(".insert")),
2715 &NewAI, NewAI.getAlignment());
2717 DEBUG(dbgs() << " to: " << *Store << "\n");
2721 // If this is a memset on an alloca where we can widen stores, insert the
2723 if (IntTy && (BeginOffset > NewAllocaBeginOffset ||
2724 EndOffset < NewAllocaEndOffset)) {
2725 assert(!II.isVolatile());
2726 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2727 getName(".oldload"));
2728 Old = convertValue(TD, IRB, Old, IntTy);
2729 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2730 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2731 V = insertInteger(TD, IRB, Old, V, Offset, getName(".insert"));
2734 if (V->getType() != AllocaTy)
2735 V = convertValue(TD, IRB, V, AllocaTy);
2737 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(),
2740 DEBUG(dbgs() << " to: " << *New << "\n");
2741 return !II.isVolatile();
2744 bool visitMemTransferInst(MemTransferInst &II) {
2745 // Rewriting of memory transfer instructions can be a bit tricky. We break
2746 // them into two categories: split intrinsics and unsplit intrinsics.
2748 DEBUG(dbgs() << " original: " << II << "\n");
2749 IRBuilder<> IRB(&II);
2751 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2752 bool IsDest = II.getRawDest() == OldPtr;
2754 const AllocaPartitioning::MemTransferOffsets &MTO
2755 = P.getMemTransferOffsets(II);
2757 // Compute the relative offset within the transfer.
2758 unsigned IntPtrWidth = TD.getPointerSizeInBits();
2759 APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
2760 : MTO.SourceBegin));
2762 unsigned Align = II.getAlignment();
2764 Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(),
2765 MinAlign(II.getAlignment(), getPartitionAlign()));
2767 // For unsplit intrinsics, we simply modify the source and destination
2768 // pointers in place. This isn't just an optimization, it is a matter of
2769 // correctness. With unsplit intrinsics we may be dealing with transfers
2770 // within a single alloca before SROA ran, or with transfers that have
2771 // a variable length. We may also be dealing with memmove instead of
2772 // memcpy, and so simply updating the pointers is the necessary for us to
2773 // update both source and dest of a single call.
2774 if (!MTO.IsSplittable) {
2775 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2777 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2779 II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType()));
2781 Type *CstTy = II.getAlignmentCst()->getType();
2782 II.setAlignment(ConstantInt::get(CstTy, Align));
2784 DEBUG(dbgs() << " to: " << II << "\n");
2785 deleteIfTriviallyDead(OldOp);
2788 // For split transfer intrinsics we have an incredibly useful assurance:
2789 // the source and destination do not reside within the same alloca, and at
2790 // least one of them does not escape. This means that we can replace
2791 // memmove with memcpy, and we don't need to worry about all manner of
2792 // downsides to splitting and transforming the operations.
2794 // If this doesn't map cleanly onto the alloca type, and that type isn't
2795 // a single value type, just emit a memcpy.
2797 = !VecTy && !IntTy && (BeginOffset != NewAllocaBeginOffset ||
2798 EndOffset != NewAllocaEndOffset ||
2799 !NewAI.getAllocatedType()->isSingleValueType());
2801 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2802 // size hasn't been shrunk based on analysis of the viable range, this is
2804 if (EmitMemCpy && &OldAI == &NewAI) {
2805 uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin;
2806 uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd;
2807 // Ensure the start lines up.
2808 assert(BeginOffset == OrigBegin);
2811 // Rewrite the size as needed.
2812 if (EndOffset != OrigEnd)
2813 II.setLength(ConstantInt::get(II.getLength()->getType(),
2814 EndOffset - BeginOffset));
2817 // Record this instruction for deletion.
2818 Pass.DeadInsts.insert(&II);
2820 bool IsWholeAlloca = BeginOffset == NewAllocaBeginOffset &&
2821 EndOffset == NewAllocaEndOffset;
2822 bool IsVectorElement = VecTy && !IsWholeAlloca;
2823 uint64_t Size = EndOffset - BeginOffset;
2824 IntegerType *SubIntTy
2825 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0;
2827 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2828 : II.getRawDest()->getType();
2830 if (IsVectorElement)
2831 OtherPtrTy = VecTy->getElementType()->getPointerTo();
2832 else if (IntTy && !IsWholeAlloca)
2833 OtherPtrTy = SubIntTy->getPointerTo();
2835 OtherPtrTy = NewAI.getType();
2838 // Compute the other pointer, folding as much as possible to produce
2839 // a single, simple GEP in most cases.
2840 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2841 OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
2842 getName("." + OtherPtr->getName()));
2844 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2845 // alloca that should be re-examined after rewriting this instruction.
2847 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2848 Pass.Worklist.insert(AI);
2852 = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType()
2853 : II.getRawSource()->getType());
2854 Type *SizeTy = II.getLength()->getType();
2855 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2857 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2858 IsDest ? OtherPtr : OurPtr,
2859 Size, Align, II.isVolatile());
2861 DEBUG(dbgs() << " to: " << *New << "\n");
2865 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy
2866 // is equivalent to 1, but that isn't true if we end up rewriting this as
2871 Value *SrcPtr = OtherPtr;
2872 Value *DstPtr = &NewAI;
2874 std::swap(SrcPtr, DstPtr);
2877 if (IsVectorElement && !IsDest) {
2878 // We have to extract rather than load.
2879 Src = IRB.CreateExtractElement(
2880 IRB.CreateAlignedLoad(SrcPtr, Align, getName(".copyload")),
2881 getIndex(IRB, BeginOffset),
2882 getName(".copyextract"));
2883 } else if (IntTy && !IsWholeAlloca && !IsDest) {
2884 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2886 Src = convertValue(TD, IRB, Src, IntTy);
2887 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2888 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2889 Src = extractInteger(TD, IRB, Src, SubIntTy, Offset, getName(".extract"));
2891 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(),
2892 getName(".copyload"));
2895 if (IntTy && !IsWholeAlloca && IsDest) {
2896 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(),
2897 getName(".oldload"));
2898 Old = convertValue(TD, IRB, Old, IntTy);
2899 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset");
2900 uint64_t Offset = BeginOffset - NewAllocaBeginOffset;
2901 Src = insertInteger(TD, IRB, Old, Src, Offset, getName(".insert"));
2902 Src = convertValue(TD, IRB, Src, NewAllocaTy);
2905 if (IsVectorElement && IsDest) {
2906 // We have to insert into a loaded copy before storing.
2907 Src = IRB.CreateInsertElement(
2908 IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), getName(".load")),
2909 Src, getIndex(IRB, BeginOffset),
2910 getName(".insert"));
2913 StoreInst *Store = cast<StoreInst>(
2914 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile()));
2916 DEBUG(dbgs() << " to: " << *Store << "\n");
2917 return !II.isVolatile();
2920 bool visitIntrinsicInst(IntrinsicInst &II) {
2921 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2922 II.getIntrinsicID() == Intrinsic::lifetime_end);
2923 DEBUG(dbgs() << " original: " << II << "\n");
2924 IRBuilder<> IRB(&II);
2925 assert(II.getArgOperand(1) == OldPtr);
2927 // Record this instruction for deletion.
2928 Pass.DeadInsts.insert(&II);
2931 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2932 EndOffset - BeginOffset);
2933 Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType());
2935 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2936 New = IRB.CreateLifetimeStart(Ptr, Size);
2938 New = IRB.CreateLifetimeEnd(Ptr, Size);
2940 DEBUG(dbgs() << " to: " << *New << "\n");
2944 bool visitPHINode(PHINode &PN) {
2945 DEBUG(dbgs() << " original: " << PN << "\n");
2947 // We would like to compute a new pointer in only one place, but have it be
2948 // as local as possible to the PHI. To do that, we re-use the location of
2949 // the old pointer, which necessarily must be in the right position to
2950 // dominate the PHI.
2951 IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr));
2953 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
2954 // Replace the operands which were using the old pointer.
2955 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr);
2957 DEBUG(dbgs() << " to: " << PN << "\n");
2958 deleteIfTriviallyDead(OldPtr);
2962 bool visitSelectInst(SelectInst &SI) {
2963 DEBUG(dbgs() << " original: " << SI << "\n");
2964 IRBuilder<> IRB(&SI);
2966 // Find the operand we need to rewrite here.
2967 bool IsTrueVal = SI.getTrueValue() == OldPtr;
2969 assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!");
2971 assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!");
2973 Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
2974 SI.setOperand(IsTrueVal ? 1 : 2, NewPtr);
2975 DEBUG(dbgs() << " to: " << SI << "\n");
2976 deleteIfTriviallyDead(OldPtr);
2984 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2986 /// This pass aggressively rewrites all aggregate loads and stores on
2987 /// a particular pointer (or any pointer derived from it which we can identify)
2988 /// with scalar loads and stores.
2989 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2990 // Befriend the base class so it can delegate to private visit methods.
2991 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2993 const DataLayout &TD;
2995 /// Queue of pointer uses to analyze and potentially rewrite.
2996 SmallVector<Use *, 8> Queue;
2998 /// Set to prevent us from cycling with phi nodes and loops.
2999 SmallPtrSet<User *, 8> Visited;
3001 /// The current pointer use being rewritten. This is used to dig up the used
3002 /// value (as opposed to the user).
3006 AggLoadStoreRewriter(const DataLayout &TD) : TD(TD) {}
3008 /// Rewrite loads and stores through a pointer and all pointers derived from
3010 bool rewrite(Instruction &I) {
3011 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
3013 bool Changed = false;
3014 while (!Queue.empty()) {
3015 U = Queue.pop_back_val();
3016 Changed |= visit(cast<Instruction>(U->getUser()));
3022 /// Enqueue all the users of the given instruction for further processing.
3023 /// This uses a set to de-duplicate users.
3024 void enqueueUsers(Instruction &I) {
3025 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
3027 if (Visited.insert(*UI))
3028 Queue.push_back(&UI.getUse());
3031 // Conservative default is to not rewrite anything.
3032 bool visitInstruction(Instruction &I) { return false; }
3034 /// \brief Generic recursive split emission class.
3035 template <typename Derived>
3038 /// The builder used to form new instructions.
3040 /// The indices which to be used with insert- or extractvalue to select the
3041 /// appropriate value within the aggregate.
3042 SmallVector<unsigned, 4> Indices;
3043 /// The indices to a GEP instruction which will move Ptr to the correct slot
3044 /// within the aggregate.
3045 SmallVector<Value *, 4> GEPIndices;
3046 /// The base pointer of the original op, used as a base for GEPing the
3047 /// split operations.
3050 /// Initialize the splitter with an insertion point, Ptr and start with a
3051 /// single zero GEP index.
3052 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
3053 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
3056 /// \brief Generic recursive split emission routine.
3058 /// This method recursively splits an aggregate op (load or store) into
3059 /// scalar or vector ops. It splits recursively until it hits a single value
3060 /// and emits that single value operation via the template argument.
3062 /// The logic of this routine relies on GEPs and insertvalue and
3063 /// extractvalue all operating with the same fundamental index list, merely
3064 /// formatted differently (GEPs need actual values).
3066 /// \param Ty The type being split recursively into smaller ops.
3067 /// \param Agg The aggregate value being built up or stored, depending on
3068 /// whether this is splitting a load or a store respectively.
3069 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
3070 if (Ty->isSingleValueType())
3071 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
3073 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
3074 unsigned OldSize = Indices.size();
3076 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
3078 assert(Indices.size() == OldSize && "Did not return to the old size");
3079 Indices.push_back(Idx);
3080 GEPIndices.push_back(IRB.getInt32(Idx));
3081 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
3082 GEPIndices.pop_back();
3088 if (StructType *STy = dyn_cast<StructType>(Ty)) {
3089 unsigned OldSize = Indices.size();
3091 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
3093 assert(Indices.size() == OldSize && "Did not return to the old size");
3094 Indices.push_back(Idx);
3095 GEPIndices.push_back(IRB.getInt32(Idx));
3096 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
3097 GEPIndices.pop_back();
3103 llvm_unreachable("Only arrays and structs are aggregate loadable types");
3107 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
3108 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3109 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
3111 /// Emit a leaf load of a single value. This is called at the leaves of the
3112 /// recursive emission to actually load values.
3113 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3114 assert(Ty->isSingleValueType());
3115 // Load the single value and insert it using the indices.
3116 Value *Load = IRB.CreateLoad(IRB.CreateInBoundsGEP(Ptr, GEPIndices,
3119 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
3120 DEBUG(dbgs() << " to: " << *Load << "\n");
3124 bool visitLoadInst(LoadInst &LI) {
3125 assert(LI.getPointerOperand() == *U);
3126 if (!LI.isSimple() || LI.getType()->isSingleValueType())
3129 // We have an aggregate being loaded, split it apart.
3130 DEBUG(dbgs() << " original: " << LI << "\n");
3131 LoadOpSplitter Splitter(&LI, *U);
3132 Value *V = UndefValue::get(LI.getType());
3133 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
3134 LI.replaceAllUsesWith(V);
3135 LI.eraseFromParent();
3139 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
3140 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
3141 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
3143 /// Emit a leaf store of a single value. This is called at the leaves of the
3144 /// recursive emission to actually produce stores.
3145 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
3146 assert(Ty->isSingleValueType());
3147 // Extract the single value and store it using the indices.
3148 Value *Store = IRB.CreateStore(
3149 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
3150 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
3152 DEBUG(dbgs() << " to: " << *Store << "\n");
3156 bool visitStoreInst(StoreInst &SI) {
3157 if (!SI.isSimple() || SI.getPointerOperand() != *U)
3159 Value *V = SI.getValueOperand();
3160 if (V->getType()->isSingleValueType())
3163 // We have an aggregate being stored, split it apart.
3164 DEBUG(dbgs() << " original: " << SI << "\n");
3165 StoreOpSplitter Splitter(&SI, *U);
3166 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
3167 SI.eraseFromParent();
3171 bool visitBitCastInst(BitCastInst &BC) {
3176 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
3181 bool visitPHINode(PHINode &PN) {
3186 bool visitSelectInst(SelectInst &SI) {
3193 /// \brief Strip aggregate type wrapping.
3195 /// This removes no-op aggregate types wrapping an underlying type. It will
3196 /// strip as many layers of types as it can without changing either the type
3197 /// size or the allocated size.
3198 static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) {
3199 if (Ty->isSingleValueType())
3202 uint64_t AllocSize = DL.getTypeAllocSize(Ty);
3203 uint64_t TypeSize = DL.getTypeSizeInBits(Ty);
3206 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
3207 InnerTy = ArrTy->getElementType();
3208 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
3209 const StructLayout *SL = DL.getStructLayout(STy);
3210 unsigned Index = SL->getElementContainingOffset(0);
3211 InnerTy = STy->getElementType(Index);
3216 if (AllocSize > DL.getTypeAllocSize(InnerTy) ||
3217 TypeSize > DL.getTypeSizeInBits(InnerTy))
3220 return stripAggregateTypeWrapping(DL, InnerTy);
3223 /// \brief Try to find a partition of the aggregate type passed in for a given
3224 /// offset and size.
3226 /// This recurses through the aggregate type and tries to compute a subtype
3227 /// based on the offset and size. When the offset and size span a sub-section
3228 /// of an array, it will even compute a new array type for that sub-section,
3229 /// and the same for structs.
3231 /// Note that this routine is very strict and tries to find a partition of the
3232 /// type which produces the *exact* right offset and size. It is not forgiving
3233 /// when the size or offset cause either end of type-based partition to be off.
3234 /// Also, this is a best-effort routine. It is reasonable to give up and not
3235 /// return a type if necessary.
3236 static Type *getTypePartition(const DataLayout &TD, Type *Ty,
3237 uint64_t Offset, uint64_t Size) {
3238 if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size)
3239 return stripAggregateTypeWrapping(TD, Ty);
3240 if (Offset > TD.getTypeAllocSize(Ty) ||
3241 (TD.getTypeAllocSize(Ty) - Offset) < Size)
3244 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
3245 // We can't partition pointers...
3246 if (SeqTy->isPointerTy())
3249 Type *ElementTy = SeqTy->getElementType();
3250 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3251 uint64_t NumSkippedElements = Offset / ElementSize;
3252 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy))
3253 if (NumSkippedElements >= ArrTy->getNumElements())
3255 if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy))
3256 if (NumSkippedElements >= VecTy->getNumElements())
3258 Offset -= NumSkippedElements * ElementSize;
3260 // First check if we need to recurse.
3261 if (Offset > 0 || Size < ElementSize) {
3262 // Bail if the partition ends in a different array element.
3263 if ((Offset + Size) > ElementSize)
3265 // Recurse through the element type trying to peel off offset bytes.
3266 return getTypePartition(TD, ElementTy, Offset, Size);
3268 assert(Offset == 0);
3270 if (Size == ElementSize)
3271 return stripAggregateTypeWrapping(TD, ElementTy);
3272 assert(Size > ElementSize);
3273 uint64_t NumElements = Size / ElementSize;
3274 if (NumElements * ElementSize != Size)
3276 return ArrayType::get(ElementTy, NumElements);
3279 StructType *STy = dyn_cast<StructType>(Ty);
3283 const StructLayout *SL = TD.getStructLayout(STy);
3284 if (Offset >= SL->getSizeInBytes())
3286 uint64_t EndOffset = Offset + Size;
3287 if (EndOffset > SL->getSizeInBytes())
3290 unsigned Index = SL->getElementContainingOffset(Offset);
3291 Offset -= SL->getElementOffset(Index);
3293 Type *ElementTy = STy->getElementType(Index);
3294 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
3295 if (Offset >= ElementSize)
3296 return 0; // The offset points into alignment padding.
3298 // See if any partition must be contained by the element.
3299 if (Offset > 0 || Size < ElementSize) {
3300 if ((Offset + Size) > ElementSize)
3302 return getTypePartition(TD, ElementTy, Offset, Size);
3304 assert(Offset == 0);
3306 if (Size == ElementSize)
3307 return stripAggregateTypeWrapping(TD, ElementTy);
3309 StructType::element_iterator EI = STy->element_begin() + Index,
3310 EE = STy->element_end();
3311 if (EndOffset < SL->getSizeInBytes()) {
3312 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
3313 if (Index == EndIndex)
3314 return 0; // Within a single element and its padding.
3316 // Don't try to form "natural" types if the elements don't line up with the
3318 // FIXME: We could potentially recurse down through the last element in the
3319 // sub-struct to find a natural end point.
3320 if (SL->getElementOffset(EndIndex) != EndOffset)
3323 assert(Index < EndIndex);
3324 EE = STy->element_begin() + EndIndex;
3327 // Try to build up a sub-structure.
3328 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE),
3330 const StructLayout *SubSL = TD.getStructLayout(SubTy);
3331 if (Size != SubSL->getSizeInBytes())
3332 return 0; // The sub-struct doesn't have quite the size needed.
3337 /// \brief Rewrite an alloca partition's users.
3339 /// This routine drives both of the rewriting goals of the SROA pass. It tries
3340 /// to rewrite uses of an alloca partition to be conducive for SSA value
3341 /// promotion. If the partition needs a new, more refined alloca, this will
3342 /// build that new alloca, preserving as much type information as possible, and
3343 /// rewrite the uses of the old alloca to point at the new one and have the
3344 /// appropriate new offsets. It also evaluates how successful the rewrite was
3345 /// at enabling promotion and if it was successful queues the alloca to be
3347 bool SROA::rewriteAllocaPartition(AllocaInst &AI,
3348 AllocaPartitioning &P,
3349 AllocaPartitioning::iterator PI) {
3350 uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset;
3351 bool IsLive = false;
3352 for (AllocaPartitioning::use_iterator UI = P.use_begin(PI),
3354 UI != UE && !IsLive; ++UI)
3358 return false; // No live uses left of this partition.
3360 DEBUG(dbgs() << "Speculating PHIs and selects in partition "
3361 << "[" << PI->BeginOffset << "," << PI->EndOffset << ")\n");
3363 PHIOrSelectSpeculator Speculator(*TD, P, *this);
3364 DEBUG(dbgs() << " speculating ");
3365 DEBUG(P.print(dbgs(), PI, ""));
3366 Speculator.visitUsers(PI);
3368 // Try to compute a friendly type for this partition of the alloca. This
3369 // won't always succeed, in which case we fall back to a legal integer type
3370 // or an i8 array of an appropriate size.
3372 if (Type *PartitionTy = P.getCommonType(PI))
3373 if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize)
3374 AllocaTy = PartitionTy;
3376 if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(),
3377 PI->BeginOffset, AllocaSize))
3378 AllocaTy = PartitionTy;
3380 (AllocaTy->isArrayTy() &&
3381 AllocaTy->getArrayElementType()->isIntegerTy())) &&
3382 TD->isLegalInteger(AllocaSize * 8))
3383 AllocaTy = Type::getIntNTy(*C, AllocaSize * 8);
3385 AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize);
3386 assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize);
3388 // Check for the case where we're going to rewrite to a new alloca of the
3389 // exact same type as the original, and with the same access offsets. In that
3390 // case, re-use the existing alloca, but still run through the rewriter to
3391 // performe phi and select speculation.
3393 if (AllocaTy == AI.getAllocatedType()) {
3394 assert(PI->BeginOffset == 0 &&
3395 "Non-zero begin offset but same alloca type");
3396 assert(PI == P.begin() && "Begin offset is zero on later partition");
3399 unsigned Alignment = AI.getAlignment();
3401 // The minimum alignment which users can rely on when the explicit
3402 // alignment is omitted or zero is that required by the ABI for this
3404 Alignment = TD->getABITypeAlignment(AI.getAllocatedType());
3406 Alignment = MinAlign(Alignment, PI->BeginOffset);
3407 // If we will get at least this much alignment from the type alone, leave
3408 // the alloca's alignment unconstrained.
3409 if (Alignment <= TD->getABITypeAlignment(AllocaTy))
3411 NewAI = new AllocaInst(AllocaTy, 0, Alignment,
3412 AI.getName() + ".sroa." + Twine(PI - P.begin()),
3417 DEBUG(dbgs() << "Rewriting alloca partition "
3418 << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: "
3421 // Track the high watermark of the post-promotion worklist. We will reset it
3422 // to this point if the alloca is not in fact scheduled for promotion.
3423 unsigned PPWOldSize = PostPromotionWorklist.size();
3425 AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI,
3426 PI->BeginOffset, PI->EndOffset);
3427 DEBUG(dbgs() << " rewriting ");
3428 DEBUG(P.print(dbgs(), PI, ""));
3429 bool Promotable = Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI));
3431 DEBUG(dbgs() << " and queuing for promotion\n");
3432 PromotableAllocas.push_back(NewAI);
3433 } else if (NewAI != &AI) {
3434 // If we can't promote the alloca, iterate on it to check for new
3435 // refinements exposed by splitting the current alloca. Don't iterate on an
3436 // alloca which didn't actually change and didn't get promoted.
3437 Worklist.insert(NewAI);
3440 // Drop any post-promotion work items if promotion didn't happen.
3442 while (PostPromotionWorklist.size() > PPWOldSize)
3443 PostPromotionWorklist.pop_back();
3448 /// \brief Walks the partitioning of an alloca rewriting uses of each partition.
3449 bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
3450 bool Changed = false;
3451 for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE;
3453 Changed |= rewriteAllocaPartition(AI, P, PI);
3458 /// \brief Analyze an alloca for SROA.
3460 /// This analyzes the alloca to ensure we can reason about it, builds
3461 /// a partitioning of the alloca, and then hands it off to be split and
3462 /// rewritten as needed.
3463 bool SROA::runOnAlloca(AllocaInst &AI) {
3464 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3465 ++NumAllocasAnalyzed;
3467 // Special case dead allocas, as they're trivial.
3468 if (AI.use_empty()) {
3469 AI.eraseFromParent();
3473 // Skip alloca forms that this analysis can't handle.
3474 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3475 TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
3478 bool Changed = false;
3480 // First, split any FCA loads and stores touching this alloca to promote
3481 // better splitting and promotion opportunities.
3482 AggLoadStoreRewriter AggRewriter(*TD);
3483 Changed |= AggRewriter.rewrite(AI);
3485 // Build the partition set using a recursive instruction-visiting builder.
3486 AllocaPartitioning P(*TD, AI);
3487 DEBUG(P.print(dbgs()));
3491 // Delete all the dead users of this alloca before splitting and rewriting it.
3492 for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
3493 DE = P.dead_user_end();
3496 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3497 DeadInsts.insert(*DI);
3499 for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(),
3500 DE = P.dead_op_end();
3503 // Clobber the use with an undef value.
3504 **DO = UndefValue::get(OldV->getType());
3505 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3506 if (isInstructionTriviallyDead(OldI)) {
3508 DeadInsts.insert(OldI);
3512 // No partitions to split. Leave the dead alloca for a later pass to clean up.
3513 if (P.begin() == P.end())
3516 return splitAlloca(AI, P) || Changed;
3519 /// \brief Delete the dead instructions accumulated in this run.
3521 /// Recursively deletes the dead instructions we've accumulated. This is done
3522 /// at the very end to maximize locality of the recursive delete and to
3523 /// minimize the problems of invalidated instruction pointers as such pointers
3524 /// are used heavily in the intermediate stages of the algorithm.
3526 /// We also record the alloca instructions deleted here so that they aren't
3527 /// subsequently handed to mem2reg to promote.
3528 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3529 while (!DeadInsts.empty()) {
3530 Instruction *I = DeadInsts.pop_back_val();
3531 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3533 I->replaceAllUsesWith(UndefValue::get(I->getType()));
3535 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3536 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3537 // Zero out the operand and see if it becomes trivially dead.
3539 if (isInstructionTriviallyDead(U))
3540 DeadInsts.insert(U);
3543 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3544 DeletedAllocas.insert(AI);
3547 I->eraseFromParent();
3551 /// \brief Promote the allocas, using the best available technique.
3553 /// This attempts to promote whatever allocas have been identified as viable in
3554 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3555 /// If there is a domtree available, we attempt to promote using the full power
3556 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3557 /// based on the SSAUpdater utilities. This function returns whether any
3558 /// promotion occured.
3559 bool SROA::promoteAllocas(Function &F) {
3560 if (PromotableAllocas.empty())
3563 NumPromoted += PromotableAllocas.size();
3565 if (DT && !ForceSSAUpdater) {
3566 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3567 PromoteMemToReg(PromotableAllocas, *DT);
3568 PromotableAllocas.clear();
3572 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3574 DIBuilder DIB(*F.getParent());
3575 SmallVector<Instruction*, 64> Insts;
3577 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3578 AllocaInst *AI = PromotableAllocas[Idx];
3579 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
3581 Instruction *I = cast<Instruction>(*UI++);
3582 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3583 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3584 // leading to them) here. Eventually it should use them to optimize the
3585 // scalar values produced.
3586 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
3587 assert(onlyUsedByLifetimeMarkers(I) &&
3588 "Found a bitcast used outside of a lifetime marker.");
3589 while (!I->use_empty())
3590 cast<Instruction>(*I->use_begin())->eraseFromParent();
3591 I->eraseFromParent();
3594 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3595 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3596 II->getIntrinsicID() == Intrinsic::lifetime_end);
3597 II->eraseFromParent();
3603 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3607 PromotableAllocas.clear();
3612 /// \brief A predicate to test whether an alloca belongs to a set.
3613 class IsAllocaInSet {
3614 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3618 typedef AllocaInst *argument_type;
3620 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3621 bool operator()(AllocaInst *AI) const { return Set.count(AI); }
3625 bool SROA::runOnFunction(Function &F) {
3626 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3627 C = &F.getContext();
3628 TD = getAnalysisIfAvailable<DataLayout>();
3630 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3633 DT = getAnalysisIfAvailable<DominatorTree>();
3635 BasicBlock &EntryBB = F.getEntryBlock();
3636 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3638 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3639 Worklist.insert(AI);
3641 bool Changed = false;
3642 // A set of deleted alloca instruction pointers which should be removed from
3643 // the list of promotable allocas.
3644 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3647 while (!Worklist.empty()) {
3648 Changed |= runOnAlloca(*Worklist.pop_back_val());
3649 deleteDeadInstructions(DeletedAllocas);
3651 // Remove the deleted allocas from various lists so that we don't try to
3652 // continue processing them.
3653 if (!DeletedAllocas.empty()) {
3654 Worklist.remove_if(IsAllocaInSet(DeletedAllocas));
3655 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas));
3656 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3657 PromotableAllocas.end(),
3658 IsAllocaInSet(DeletedAllocas)),
3659 PromotableAllocas.end());
3660 DeletedAllocas.clear();
3664 Changed |= promoteAllocas(F);
3666 Worklist = PostPromotionWorklist;
3667 PostPromotionWorklist.clear();
3668 } while (!Worklist.empty());
3673 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3674 if (RequiresDomTree)
3675 AU.addRequired<DominatorTree>();
3676 AU.setPreservesCFG();