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/GlobalVariable.h"
34 #include "llvm/IRBuilder.h"
35 #include "llvm/Instructions.h"
36 #include "llvm/IntrinsicInst.h"
37 #include "llvm/LLVMContext.h"
38 #include "llvm/Module.h"
39 #include "llvm/Operator.h"
40 #include "llvm/Pass.h"
41 #include "llvm/ADT/SetVector.h"
42 #include "llvm/ADT/SmallVector.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/STLExtras.h"
45 #include "llvm/ADT/TinyPtrVector.h"
46 #include "llvm/Analysis/Dominators.h"
47 #include "llvm/Analysis/Loads.h"
48 #include "llvm/Analysis/ValueTracking.h"
49 #include "llvm/Support/CallSite.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/GetElementPtrTypeIterator.h"
54 #include "llvm/Support/InstVisitor.h"
55 #include "llvm/Support/MathExtras.h"
56 #include "llvm/Support/ValueHandle.h"
57 #include "llvm/Support/raw_ostream.h"
58 #include "llvm/Target/TargetData.h"
59 #include "llvm/Transforms/Utils/Local.h"
60 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
61 #include "llvm/Transforms/Utils/SSAUpdater.h"
64 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
65 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
66 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
67 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
68 STATISTIC(NumDeleted, "Number of instructions deleted");
69 STATISTIC(NumVectorized, "Number of vectorized aggregates");
71 /// Hidden option to force the pass to not use DomTree and mem2reg, instead
72 /// forming SSA values through the SSAUpdater infrastructure.
74 ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden);
77 /// \brief Alloca partitioning representation.
79 /// This class represents a partitioning of an alloca into slices, and
80 /// information about the nature of uses of each slice of the alloca. The goal
81 /// is that this information is sufficient to decide if and how to split the
82 /// alloca apart and replace slices with scalars. It is also intended that this
83 /// structure can capture the relevant information needed both to decide about
84 /// and to enact these transformations.
85 class AllocaPartitioning {
87 /// \brief A common base class for representing a half-open byte range.
89 /// \brief The beginning offset of the range.
92 /// \brief The ending offset, not included in the range.
95 ByteRange() : BeginOffset(), EndOffset() {}
96 ByteRange(uint64_t BeginOffset, uint64_t EndOffset)
97 : BeginOffset(BeginOffset), EndOffset(EndOffset) {}
99 /// \brief Support for ordering ranges.
101 /// This provides an ordering over ranges such that start offsets are
102 /// always increasing, and within equal start offsets, the end offsets are
103 /// decreasing. Thus the spanning range comes first in a cluster with the
104 /// same start position.
105 bool operator<(const ByteRange &RHS) const {
106 if (BeginOffset < RHS.BeginOffset) return true;
107 if (BeginOffset > RHS.BeginOffset) return false;
108 if (EndOffset > RHS.EndOffset) return true;
112 /// \brief Support comparison with a single offset to allow binary searches.
113 friend bool operator<(const ByteRange &LHS, uint64_t RHSOffset) {
114 return LHS.BeginOffset < RHSOffset;
117 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset,
118 const ByteRange &RHS) {
119 return LHSOffset < RHS.BeginOffset;
122 bool operator==(const ByteRange &RHS) const {
123 return BeginOffset == RHS.BeginOffset && EndOffset == RHS.EndOffset;
125 bool operator!=(const ByteRange &RHS) const { return !operator==(RHS); }
128 /// \brief A partition of an alloca.
130 /// This structure represents a contiguous partition of the alloca. These are
131 /// formed by examining the uses of the alloca. During formation, they may
132 /// overlap but once an AllocaPartitioning is built, the Partitions within it
133 /// are all disjoint.
134 struct Partition : public ByteRange {
135 /// \brief Whether this partition is splittable into smaller partitions.
137 /// We flag partitions as splittable when they are formed entirely due to
138 /// accesses by trivially splittable operations such as memset and memcpy.
140 /// FIXME: At some point we should consider loads and stores of FCAs to be
141 /// splittable and eagerly split them into scalar values.
144 Partition() : ByteRange(), IsSplittable() {}
145 Partition(uint64_t BeginOffset, uint64_t EndOffset, bool IsSplittable)
146 : ByteRange(BeginOffset, EndOffset), IsSplittable(IsSplittable) {}
149 /// \brief A particular use of a partition of the alloca.
151 /// This structure is used to associate uses of a partition with it. They
152 /// mark the range of bytes which are referenced by a particular instruction,
153 /// and includes a handle to the user itself and the pointer value in use.
154 /// The bounds of these uses are determined by intersecting the bounds of the
155 /// memory use itself with a particular partition. As a consequence there is
156 /// intentionally overlap between various uses of the same partition.
157 struct PartitionUse : public ByteRange {
158 /// \brief The user of this range of the alloca.
159 AssertingVH<Instruction> User;
161 /// \brief The particular pointer value derived from this alloca in use.
162 AssertingVH<Instruction> Ptr;
164 PartitionUse() : ByteRange(), User(), Ptr() {}
165 PartitionUse(uint64_t BeginOffset, uint64_t EndOffset,
166 Instruction *User, Instruction *Ptr)
167 : ByteRange(BeginOffset, EndOffset), User(User), Ptr(Ptr) {}
170 /// \brief Construct a partitioning of a particular alloca.
172 /// Construction does most of the work for partitioning the alloca. This
173 /// performs the necessary walks of users and builds a partitioning from it.
174 AllocaPartitioning(const TargetData &TD, AllocaInst &AI);
176 /// \brief Test whether a pointer to the allocation escapes our analysis.
178 /// If this is true, the partitioning is never fully built and should be
180 bool isEscaped() const { return PointerEscapingInstr; }
182 /// \brief Support for iterating over the partitions.
184 typedef SmallVectorImpl<Partition>::iterator iterator;
185 iterator begin() { return Partitions.begin(); }
186 iterator end() { return Partitions.end(); }
188 typedef SmallVectorImpl<Partition>::const_iterator const_iterator;
189 const_iterator begin() const { return Partitions.begin(); }
190 const_iterator end() const { return Partitions.end(); }
193 /// \brief Support for iterating over and manipulating a particular
194 /// partition's uses.
196 /// The iteration support provided for uses is more limited, but also
197 /// includes some manipulation routines to support rewriting the uses of
198 /// partitions during SROA.
200 typedef SmallVectorImpl<PartitionUse>::iterator use_iterator;
201 use_iterator use_begin(unsigned Idx) { return Uses[Idx].begin(); }
202 use_iterator use_begin(const_iterator I) { return Uses[I - begin()].begin(); }
203 use_iterator use_end(unsigned Idx) { return Uses[Idx].end(); }
204 use_iterator use_end(const_iterator I) { return Uses[I - begin()].end(); }
205 void use_push_back(unsigned Idx, const PartitionUse &U) {
206 Uses[Idx].push_back(U);
208 void use_push_back(const_iterator I, const PartitionUse &U) {
209 Uses[I - begin()].push_back(U);
211 void use_erase(unsigned Idx, use_iterator UI) { Uses[Idx].erase(UI); }
212 void use_erase(const_iterator I, use_iterator UI) {
213 Uses[I - begin()].erase(UI);
216 typedef SmallVectorImpl<PartitionUse>::const_iterator const_use_iterator;
217 const_use_iterator use_begin(unsigned Idx) const { return Uses[Idx].begin(); }
218 const_use_iterator use_begin(const_iterator I) const {
219 return Uses[I - begin()].begin();
221 const_use_iterator use_end(unsigned Idx) const { return Uses[Idx].end(); }
222 const_use_iterator use_end(const_iterator I) const {
223 return Uses[I - begin()].end();
227 /// \brief Allow iterating the dead users for this alloca.
229 /// These are instructions which will never actually use the alloca as they
230 /// are outside the allocated range. They are safe to replace with undef and
233 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
234 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
235 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
238 /// \brief Allow iterating the dead expressions referring to this alloca.
240 /// These are operands which have cannot actually be used to refer to the
241 /// alloca as they are outside its range and the user doesn't correct for
242 /// that. These mostly consist of PHI node inputs and the like which we just
243 /// need to replace with undef.
245 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
246 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
247 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
250 /// \brief MemTransferInst auxiliary data.
251 /// This struct provides some auxiliary data about memory transfer
252 /// intrinsics such as memcpy and memmove. These intrinsics can use two
253 /// different ranges within the same alloca, and provide other challenges to
254 /// correctly represent. We stash extra data to help us untangle this
255 /// after the partitioning is complete.
256 struct MemTransferOffsets {
257 uint64_t DestBegin, DestEnd;
258 uint64_t SourceBegin, SourceEnd;
261 MemTransferOffsets getMemTransferOffsets(MemTransferInst &II) const {
262 return MemTransferInstData.lookup(&II);
265 /// \brief Map from a PHI or select operand back to a partition.
267 /// When manipulating PHI nodes or selects, they can use more than one
268 /// partition of an alloca. We store a special mapping to allow finding the
269 /// partition referenced by each of these operands, if any.
270 iterator findPartitionForPHIOrSelectOperand(Instruction &I, Value *Op) {
271 SmallDenseMap<std::pair<Instruction *, Value *>,
272 std::pair<unsigned, unsigned> >::const_iterator MapIt
273 = PHIOrSelectOpMap.find(std::make_pair(&I, Op));
274 if (MapIt == PHIOrSelectOpMap.end())
277 return begin() + MapIt->second.first;
280 /// \brief Map from a PHI or select operand back to the specific use of
283 /// Similar to mapping these operands back to the partitions, this maps
284 /// directly to the use structure of that partition.
285 use_iterator findPartitionUseForPHIOrSelectOperand(Instruction &I,
287 SmallDenseMap<std::pair<Instruction *, Value *>,
288 std::pair<unsigned, unsigned> >::const_iterator MapIt
289 = PHIOrSelectOpMap.find(std::make_pair(&I, Op));
290 assert(MapIt != PHIOrSelectOpMap.end());
291 return Uses[MapIt->second.first].begin() + MapIt->second.second;
294 /// \brief Compute a common type among the uses of a particular partition.
296 /// This routines walks all of the uses of a particular partition and tries
297 /// to find a common type between them. Untyped operations such as memset and
298 /// memcpy are ignored.
299 Type *getCommonType(iterator I) const;
301 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
302 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
303 void printUsers(raw_ostream &OS, const_iterator I,
304 StringRef Indent = " ") const;
305 void print(raw_ostream &OS) const;
306 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const;
307 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const;
311 template <typename DerivedT, typename RetT = void> class BuilderBase;
312 class PartitionBuilder;
313 friend class AllocaPartitioning::PartitionBuilder;
315 friend class AllocaPartitioning::UseBuilder;
318 /// \brief Handle to alloca instruction to simplify method interfaces.
322 /// \brief The instruction responsible for this alloca having no partitioning.
324 /// When an instruction (potentially) escapes the pointer to the alloca, we
325 /// store a pointer to that here and abort trying to partition the alloca.
326 /// This will be null if the alloca is partitioned successfully.
327 Instruction *PointerEscapingInstr;
329 /// \brief The partitions of the alloca.
331 /// We store a vector of the partitions over the alloca here. This vector is
332 /// sorted by increasing begin offset, and then by decreasing end offset. See
333 /// the Partition inner class for more details. Initially (during
334 /// construction) there are overlaps, but we form a disjoint sequence of
335 /// partitions while finishing construction and a fully constructed object is
336 /// expected to always have this as a disjoint space.
337 SmallVector<Partition, 8> Partitions;
339 /// \brief The uses of the partitions.
341 /// This is essentially a mapping from each partition to a list of uses of
342 /// that partition. The mapping is done with a Uses vector that has the exact
343 /// same number of entries as the partition vector. Each entry is itself
344 /// a vector of the uses.
345 SmallVector<SmallVector<PartitionUse, 2>, 8> Uses;
347 /// \brief Instructions which will become dead if we rewrite the alloca.
349 /// Note that these are not separated by partition. This is because we expect
350 /// a partitioned alloca to be completely rewritten or not rewritten at all.
351 /// If rewritten, all these instructions can simply be removed and replaced
352 /// with undef as they come from outside of the allocated space.
353 SmallVector<Instruction *, 8> DeadUsers;
355 /// \brief Operands which will become dead if we rewrite the alloca.
357 /// These are operands that in their particular use can be replaced with
358 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
359 /// to PHI nodes and the like. They aren't entirely dead (there might be
360 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
361 /// want to swap this particular input for undef to simplify the use lists of
363 SmallVector<Use *, 8> DeadOperands;
365 /// \brief The underlying storage for auxiliary memcpy and memset info.
366 SmallDenseMap<MemTransferInst *, MemTransferOffsets, 4> MemTransferInstData;
368 /// \brief A side datastructure used when building up the partitions and uses.
370 /// This mapping is only really used during the initial building of the
371 /// partitioning so that we can retain information about PHI and select nodes
373 SmallDenseMap<Instruction *, std::pair<uint64_t, bool> > PHIOrSelectSizes;
375 /// \brief Auxiliary information for particular PHI or select operands.
376 SmallDenseMap<std::pair<Instruction *, Value *>,
377 std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap;
379 /// \brief A utility routine called from the constructor.
381 /// This does what it says on the tin. It is the key of the alloca partition
382 /// splitting and merging. After it is called we have the desired disjoint
383 /// collection of partitions.
384 void splitAndMergePartitions();
388 template <typename DerivedT, typename RetT>
389 class AllocaPartitioning::BuilderBase
390 : public InstVisitor<DerivedT, RetT> {
392 BuilderBase(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
394 AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())),
400 const TargetData &TD;
401 const uint64_t AllocSize;
402 AllocaPartitioning &P;
408 SmallVector<OffsetUse, 8> Queue;
410 // The active offset and use while visiting.
414 void enqueueUsers(Instruction &I, int64_t UserOffset) {
415 SmallPtrSet<User *, 8> UserSet;
416 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
418 if (!UserSet.insert(*UI))
421 OffsetUse OU = { &UI.getUse(), UserOffset };
426 bool computeConstantGEPOffset(GetElementPtrInst &GEPI, int64_t &GEPOffset) {
428 for (gep_type_iterator GTI = gep_type_begin(GEPI), GTE = gep_type_end(GEPI);
430 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
436 // Handle a struct index, which adds its field offset to the pointer.
437 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
438 unsigned ElementIdx = OpC->getZExtValue();
439 const StructLayout *SL = TD.getStructLayout(STy);
440 uint64_t ElementOffset = SL->getElementOffset(ElementIdx);
441 // Check that we can continue to model this GEP in a signed 64-bit offset.
442 if (ElementOffset > INT64_MAX ||
444 ((uint64_t)GEPOffset + ElementOffset) > INT64_MAX)) {
445 DEBUG(dbgs() << "WARNING: Encountered a cumulative offset exceeding "
446 << "what can be represented in an int64_t!\n"
447 << " alloca: " << P.AI << "\n");
451 GEPOffset = ElementOffset + (uint64_t)-GEPOffset;
453 GEPOffset += ElementOffset;
457 APInt Index = OpC->getValue().sextOrTrunc(TD.getPointerSizeInBits());
458 Index *= APInt(Index.getBitWidth(),
459 TD.getTypeAllocSize(GTI.getIndexedType()));
460 Index += APInt(Index.getBitWidth(), (uint64_t)GEPOffset,
462 // Check if the result can be stored in our int64_t offset.
463 if (!Index.isSignedIntN(sizeof(GEPOffset) * 8)) {
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 = Index.getSExtValue();
475 Value *foldSelectInst(SelectInst &SI) {
476 // If the condition being selected on is a constant or the same value is
477 // being selected between, fold the select. Yes this does (rarely) happen
479 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
480 return SI.getOperand(1+CI->isZero());
481 if (SI.getOperand(1) == SI.getOperand(2)) {
482 assert(*U == SI.getOperand(1));
483 return SI.getOperand(1);
489 /// \brief Builder for the alloca partitioning.
491 /// This class builds an alloca partitioning by recursively visiting the uses
492 /// of an alloca and splitting the partitions for each load and store at each
494 class AllocaPartitioning::PartitionBuilder
495 : public BuilderBase<PartitionBuilder, bool> {
496 friend class InstVisitor<PartitionBuilder, bool>;
498 SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap;
501 PartitionBuilder(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
502 : BuilderBase<PartitionBuilder, bool>(TD, AI, P) {}
504 /// \brief Run the builder over the allocation.
506 // Note that we have to re-evaluate size on each trip through the loop as
507 // the queue grows at the tail.
508 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
510 Offset = Queue[Idx].Offset;
511 if (!visit(cast<Instruction>(U->getUser())))
518 bool markAsEscaping(Instruction &I) {
519 P.PointerEscapingInstr = &I;
523 void insertUse(Instruction &I, int64_t Offset, uint64_t Size,
524 bool IsSplittable = false) {
525 // Completely skip uses which have a zero size or don't overlap the
528 (Offset >= 0 && (uint64_t)Offset >= AllocSize) ||
529 (Offset < 0 && (uint64_t)-Offset >= Size)) {
530 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
531 << " which starts past the end of the " << AllocSize
533 << " alloca: " << P.AI << "\n"
534 << " use: " << I << "\n");
538 // Clamp the start to the beginning of the allocation.
540 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
541 << " to start at the beginning of the alloca:\n"
542 << " alloca: " << P.AI << "\n"
543 << " use: " << I << "\n");
544 Size -= (uint64_t)-Offset;
548 uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
550 // Clamp the end offset to the end of the allocation. Note that this is
551 // formulated to handle even the case where "BeginOffset + Size" overflows.
552 assert(AllocSize >= BeginOffset); // Established above.
553 if (Size > AllocSize - BeginOffset) {
554 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
555 << " to remain within the " << AllocSize << " byte alloca:\n"
556 << " alloca: " << P.AI << "\n"
557 << " use: " << I << "\n");
558 EndOffset = AllocSize;
561 // See if we can just add a user onto the last slot currently occupied.
562 if (!P.Partitions.empty() &&
563 P.Partitions.back().BeginOffset == BeginOffset &&
564 P.Partitions.back().EndOffset == EndOffset) {
565 P.Partitions.back().IsSplittable &= IsSplittable;
569 Partition New(BeginOffset, EndOffset, IsSplittable);
570 P.Partitions.push_back(New);
573 bool handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) {
574 uint64_t Size = TD.getTypeStoreSize(Ty);
576 // If this memory access can be shown to *statically* extend outside the
577 // bounds of of the allocation, it's behavior is undefined, so simply
578 // ignore it. Note that this is more strict than the generic clamping
579 // behavior of insertUse. We also try to handle cases which might run the
581 // FIXME: We should instead consider the pointer to have escaped if this
582 // function is being instrumented for addressing bugs or race conditions.
583 if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
584 Size > (AllocSize - (uint64_t)Offset)) {
585 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte "
586 << (isa<LoadInst>(I) ? "load" : "store") << " @" << Offset
587 << " which extends past the end of the " << AllocSize
589 << " alloca: " << P.AI << "\n"
590 << " use: " << I << "\n");
594 insertUse(I, Offset, Size);
598 bool visitBitCastInst(BitCastInst &BC) {
599 enqueueUsers(BC, Offset);
603 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
605 if (!computeConstantGEPOffset(GEPI, GEPOffset))
606 return markAsEscaping(GEPI);
608 enqueueUsers(GEPI, GEPOffset);
612 bool visitLoadInst(LoadInst &LI) {
613 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) &&
614 "All simple FCA loads should have been pre-split");
615 return handleLoadOrStore(LI.getType(), LI, Offset);
618 bool visitStoreInst(StoreInst &SI) {
619 Value *ValOp = SI.getValueOperand();
621 return markAsEscaping(SI);
623 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) &&
624 "All simple FCA stores should have been pre-split");
625 return handleLoadOrStore(ValOp->getType(), SI, Offset);
629 bool visitMemSetInst(MemSetInst &II) {
630 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
631 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
632 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
633 insertUse(II, Offset, Size, Length);
637 bool visitMemTransferInst(MemTransferInst &II) {
638 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
639 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
641 // Zero-length mem transfer intrinsics can be ignored entirely.
644 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
646 // Only intrinsics with a constant length can be split.
647 Offsets.IsSplittable = Length;
649 if (*U != II.getRawDest()) {
650 assert(*U == II.getRawSource());
651 Offsets.SourceBegin = Offset;
652 Offsets.SourceEnd = Offset + Size;
654 Offsets.DestBegin = Offset;
655 Offsets.DestEnd = Offset + Size;
658 insertUse(II, Offset, Size, Offsets.IsSplittable);
659 unsigned NewIdx = P.Partitions.size() - 1;
661 SmallDenseMap<Instruction *, unsigned>::const_iterator PMI;
662 bool Inserted = false;
663 llvm::tie(PMI, Inserted)
664 = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx));
665 if (!Inserted && Offsets.IsSplittable) {
666 // We've found a memory transfer intrinsic which refers to the alloca as
667 // both a source and dest. We refuse to split these to simplify splitting
668 // logic. If possible, SROA will still split them into separate allocas
669 // and then re-analyze.
670 Offsets.IsSplittable = false;
671 P.Partitions[PMI->second].IsSplittable = false;
672 P.Partitions[NewIdx].IsSplittable = false;
678 // Disable SRoA for any intrinsics except for lifetime invariants.
679 // FIXME: What about debug instrinsics? This matches old behavior, but
680 // doesn't make sense.
681 bool visitIntrinsicInst(IntrinsicInst &II) {
682 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
683 II.getIntrinsicID() == Intrinsic::lifetime_end) {
684 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
685 uint64_t Size = std::min(AllocSize - Offset, Length->getLimitedValue());
686 insertUse(II, Offset, Size, true);
690 return markAsEscaping(II);
693 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
694 // We consider any PHI or select that results in a direct load or store of
695 // the same offset to be a viable use for partitioning purposes. These uses
696 // are considered unsplittable and the size is the maximum loaded or stored
698 SmallPtrSet<Instruction *, 4> Visited;
699 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
700 Visited.insert(Root);
701 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
702 // If there are no loads or stores, the access is dead. We mark that as
703 // a size zero access.
706 Instruction *I, *UsedI;
707 llvm::tie(UsedI, I) = Uses.pop_back_val();
709 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
710 Size = std::max(Size, TD.getTypeStoreSize(LI->getType()));
713 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
714 Value *Op = SI->getOperand(0);
717 Size = std::max(Size, TD.getTypeStoreSize(Op->getType()));
721 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
722 if (!GEP->hasAllZeroIndices())
724 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
725 !isa<SelectInst>(I)) {
729 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
731 if (Visited.insert(cast<Instruction>(*UI)))
732 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
733 } while (!Uses.empty());
738 bool visitPHINode(PHINode &PN) {
739 // See if we already have computed info on this node.
740 std::pair<uint64_t, bool> &PHIInfo = P.PHIOrSelectSizes[&PN];
742 PHIInfo.second = true;
743 insertUse(PN, Offset, PHIInfo.first);
747 // Check for an unsafe use of the PHI node.
748 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&PN, PHIInfo.first))
749 return markAsEscaping(*EscapingI);
751 insertUse(PN, Offset, PHIInfo.first);
755 bool visitSelectInst(SelectInst &SI) {
756 if (Value *Result = foldSelectInst(SI)) {
758 // If the result of the constant fold will be the pointer, recurse
759 // through the select as if we had RAUW'ed it.
760 enqueueUsers(SI, Offset);
765 // See if we already have computed info on this node.
766 std::pair<uint64_t, bool> &SelectInfo = P.PHIOrSelectSizes[&SI];
767 if (SelectInfo.first) {
768 SelectInfo.second = true;
769 insertUse(SI, Offset, SelectInfo.first);
773 // Check for an unsafe use of the PHI node.
774 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&SI, SelectInfo.first))
775 return markAsEscaping(*EscapingI);
777 insertUse(SI, Offset, SelectInfo.first);
781 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
782 bool visitInstruction(Instruction &I) { return markAsEscaping(I); }
786 /// \brief Use adder for the alloca partitioning.
788 /// This class adds the uses of an alloca to all of the partitions which they
789 /// use. For splittable partitions, this can end up doing essentially a linear
790 /// walk of the partitions, but the number of steps remains bounded by the
791 /// total result instruction size:
792 /// - The number of partitions is a result of the number unsplittable
793 /// instructions using the alloca.
794 /// - The number of users of each partition is at worst the total number of
795 /// splittable instructions using the alloca.
796 /// Thus we will produce N * M instructions in the end, where N are the number
797 /// of unsplittable uses and M are the number of splittable. This visitor does
798 /// the exact same number of updates to the partitioning.
800 /// In the more common case, this visitor will leverage the fact that the
801 /// partition space is pre-sorted, and do a logarithmic search for the
802 /// partition needed, making the total visit a classical ((N + M) * log(N))
803 /// complexity operation.
804 class AllocaPartitioning::UseBuilder : public BuilderBase<UseBuilder> {
805 friend class InstVisitor<UseBuilder>;
807 /// \brief Set to de-duplicate dead instructions found in the use walk.
808 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
811 UseBuilder(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
812 : BuilderBase<UseBuilder>(TD, AI, P) {}
814 /// \brief Run the builder over the allocation.
816 // Note that we have to re-evaluate size on each trip through the loop as
817 // the queue grows at the tail.
818 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
820 Offset = Queue[Idx].Offset;
821 this->visit(cast<Instruction>(U->getUser()));
826 void markAsDead(Instruction &I) {
827 if (VisitedDeadInsts.insert(&I))
828 P.DeadUsers.push_back(&I);
831 void insertUse(Instruction &User, int64_t Offset, uint64_t Size) {
832 // If the use has a zero size or extends outside of the allocation, record
833 // it as a dead use for elimination later.
834 if (Size == 0 || (uint64_t)Offset >= AllocSize ||
835 (Offset < 0 && (uint64_t)-Offset >= Size))
836 return markAsDead(User);
838 // Clamp the start to the beginning of the allocation.
840 Size -= (uint64_t)-Offset;
844 uint64_t BeginOffset = Offset, EndOffset = BeginOffset + Size;
846 // Clamp the end offset to the end of the allocation. Note that this is
847 // formulated to handle even the case where "BeginOffset + Size" overflows.
848 assert(AllocSize >= BeginOffset); // Established above.
849 if (Size > AllocSize - BeginOffset)
850 EndOffset = AllocSize;
852 // NB: This only works if we have zero overlapping partitions.
853 iterator B = std::lower_bound(P.begin(), P.end(), BeginOffset);
854 if (B != P.begin() && llvm::prior(B)->EndOffset > BeginOffset)
856 for (iterator I = B, E = P.end(); I != E && I->BeginOffset < EndOffset;
858 PartitionUse NewUse(std::max(I->BeginOffset, BeginOffset),
859 std::min(I->EndOffset, EndOffset),
860 &User, cast<Instruction>(*U));
861 P.use_push_back(I, NewUse);
862 if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser()))
863 P.PHIOrSelectOpMap[std::make_pair(&User, U->get())]
864 = std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1);
868 void handleLoadOrStore(Type *Ty, Instruction &I, int64_t Offset) {
869 uint64_t Size = TD.getTypeStoreSize(Ty);
871 // If this memory access can be shown to *statically* extend outside the
872 // bounds of of the allocation, it's behavior is undefined, so simply
873 // ignore it. Note that this is more strict than the generic clamping
874 // behavior of insertUse.
875 if (Offset < 0 || (uint64_t)Offset >= AllocSize ||
876 Size > (AllocSize - (uint64_t)Offset))
877 return markAsDead(I);
879 insertUse(I, Offset, Size);
882 void visitBitCastInst(BitCastInst &BC) {
884 return markAsDead(BC);
886 enqueueUsers(BC, Offset);
889 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
890 if (GEPI.use_empty())
891 return markAsDead(GEPI);
894 if (!computeConstantGEPOffset(GEPI, GEPOffset))
895 llvm_unreachable("Unable to compute constant offset for use");
897 enqueueUsers(GEPI, GEPOffset);
900 void visitLoadInst(LoadInst &LI) {
901 handleLoadOrStore(LI.getType(), LI, Offset);
904 void visitStoreInst(StoreInst &SI) {
905 handleLoadOrStore(SI.getOperand(0)->getType(), SI, Offset);
908 void visitMemSetInst(MemSetInst &II) {
909 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
910 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
911 insertUse(II, Offset, Size);
914 void visitMemTransferInst(MemTransferInst &II) {
915 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
916 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
917 insertUse(II, Offset, Size);
920 void visitIntrinsicInst(IntrinsicInst &II) {
921 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
922 II.getIntrinsicID() == Intrinsic::lifetime_end);
924 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
925 insertUse(II, Offset,
926 std::min(AllocSize - Offset, Length->getLimitedValue()));
929 void insertPHIOrSelect(Instruction &User, uint64_t Offset) {
930 uint64_t Size = P.PHIOrSelectSizes.lookup(&User).first;
932 // For PHI and select operands outside the alloca, we can't nuke the entire
933 // phi or select -- the other side might still be relevant, so we special
934 // case them here and use a separate structure to track the operands
935 // themselves which should be replaced with undef.
936 if (Offset >= AllocSize) {
937 P.DeadOperands.push_back(U);
941 insertUse(User, Offset, Size);
943 void visitPHINode(PHINode &PN) {
945 return markAsDead(PN);
947 insertPHIOrSelect(PN, Offset);
949 void visitSelectInst(SelectInst &SI) {
951 return markAsDead(SI);
953 if (Value *Result = foldSelectInst(SI)) {
955 // If the result of the constant fold will be the pointer, recurse
956 // through the select as if we had RAUW'ed it.
957 enqueueUsers(SI, Offset);
959 // Otherwise the operand to the select is dead, and we can replace it
961 P.DeadOperands.push_back(U);
966 insertPHIOrSelect(SI, Offset);
969 /// \brief Unreachable, we've already visited the alloca once.
970 void visitInstruction(Instruction &I) {
971 llvm_unreachable("Unhandled instruction in use builder.");
975 void AllocaPartitioning::splitAndMergePartitions() {
976 size_t NumDeadPartitions = 0;
978 // Track the range of splittable partitions that we pass when accumulating
979 // overlapping unsplittable partitions.
980 uint64_t SplitEndOffset = 0ull;
982 Partition New(0ull, 0ull, false);
984 for (unsigned i = 0, j = i, e = Partitions.size(); i != e; i = j) {
987 if (!Partitions[i].IsSplittable || New.BeginOffset == New.EndOffset) {
988 assert(New.BeginOffset == New.EndOffset);
991 assert(New.IsSplittable);
992 New.EndOffset = std::max(New.EndOffset, Partitions[i].EndOffset);
994 assert(New.BeginOffset != New.EndOffset);
996 // Scan the overlapping partitions.
997 while (j != e && New.EndOffset > Partitions[j].BeginOffset) {
998 // If the new partition we are forming is splittable, stop at the first
999 // unsplittable partition.
1000 if (New.IsSplittable && !Partitions[j].IsSplittable)
1003 // Grow the new partition to include any equally splittable range. 'j' is
1004 // always equally splittable when New is splittable, but when New is not
1005 // splittable, we may subsume some (or part of some) splitable partition
1006 // without growing the new one.
1007 if (New.IsSplittable == Partitions[j].IsSplittable) {
1008 New.EndOffset = std::max(New.EndOffset, Partitions[j].EndOffset);
1010 assert(!New.IsSplittable);
1011 assert(Partitions[j].IsSplittable);
1012 SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset);
1015 Partitions[j].BeginOffset = Partitions[j].EndOffset = UINT64_MAX;
1016 ++NumDeadPartitions;
1020 // If the new partition is splittable, chop off the end as soon as the
1021 // unsplittable subsequent partition starts and ensure we eventually cover
1022 // the splittable area.
1023 if (j != e && New.IsSplittable) {
1024 SplitEndOffset = std::max(SplitEndOffset, New.EndOffset);
1025 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1028 // Add the new partition if it differs from the original one and is
1029 // non-empty. We can end up with an empty partition here if it was
1030 // splittable but there is an unsplittable one that starts at the same
1032 if (New != Partitions[i]) {
1033 if (New.BeginOffset != New.EndOffset)
1034 Partitions.push_back(New);
1035 // Mark the old one for removal.
1036 Partitions[i].BeginOffset = Partitions[i].EndOffset = UINT64_MAX;
1037 ++NumDeadPartitions;
1040 New.BeginOffset = New.EndOffset;
1041 if (!New.IsSplittable) {
1042 New.EndOffset = std::max(New.EndOffset, SplitEndOffset);
1043 if (j != e && !Partitions[j].IsSplittable)
1044 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
1045 New.IsSplittable = true;
1046 // If there is a trailing splittable partition which won't be fused into
1047 // the next splittable partition go ahead and add it onto the partitions
1049 if (New.BeginOffset < New.EndOffset &&
1050 (j == e || !Partitions[j].IsSplittable ||
1051 New.EndOffset < Partitions[j].BeginOffset)) {
1052 Partitions.push_back(New);
1053 New.BeginOffset = New.EndOffset = 0ull;
1058 // Re-sort the partitions now that they have been split and merged into
1059 // disjoint set of partitions. Also remove any of the dead partitions we've
1060 // replaced in the process.
1061 std::sort(Partitions.begin(), Partitions.end());
1062 if (NumDeadPartitions) {
1063 assert(Partitions.back().BeginOffset == UINT64_MAX);
1064 assert(Partitions.back().EndOffset == UINT64_MAX);
1065 assert((ptrdiff_t)NumDeadPartitions ==
1066 std::count(Partitions.begin(), Partitions.end(), Partitions.back()));
1068 Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end());
1071 AllocaPartitioning::AllocaPartitioning(const TargetData &TD, AllocaInst &AI)
1076 PointerEscapingInstr(0) {
1077 PartitionBuilder PB(TD, AI, *this);
1081 if (Partitions.size() > 1) {
1082 // Sort the uses. This arranges for the offsets to be in ascending order,
1083 // and the sizes to be in descending order.
1084 std::sort(Partitions.begin(), Partitions.end());
1086 // Intersect splittability for all partitions with equal offsets and sizes.
1087 // Then remove all but the first so that we have a sequence of non-equal but
1088 // potentially overlapping partitions.
1089 for (iterator I = Partitions.begin(), J = I, E = Partitions.end(); I != E;
1092 while (J != E && *I == *J) {
1093 I->IsSplittable &= J->IsSplittable;
1097 Partitions.erase(std::unique(Partitions.begin(), Partitions.end()),
1100 // Split splittable and merge unsplittable partitions into a disjoint set
1101 // of partitions over the used space of the allocation.
1102 splitAndMergePartitions();
1105 // Now build up the user lists for each of these disjoint partitions by
1106 // re-walking the recursive users of the alloca.
1107 Uses.resize(Partitions.size());
1108 UseBuilder UB(TD, AI, *this);
1112 Type *AllocaPartitioning::getCommonType(iterator I) const {
1114 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
1115 if (isa<IntrinsicInst>(*UI->User))
1117 if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset)
1121 if (LoadInst *LI = dyn_cast<LoadInst>(&*UI->User)) {
1122 UserTy = LI->getType();
1123 } else if (StoreInst *SI = dyn_cast<StoreInst>(&*UI->User)) {
1124 UserTy = SI->getValueOperand()->getType();
1125 } else if (SelectInst *SI = dyn_cast<SelectInst>(&*UI->User)) {
1126 if (PointerType *PtrTy = dyn_cast<PointerType>(SI->getType()))
1127 UserTy = PtrTy->getElementType();
1128 } else if (PHINode *PN = dyn_cast<PHINode>(&*UI->User)) {
1129 if (PointerType *PtrTy = dyn_cast<PointerType>(PN->getType()))
1130 UserTy = PtrTy->getElementType();
1133 if (Ty && Ty != UserTy)
1141 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1143 void AllocaPartitioning::print(raw_ostream &OS, const_iterator I,
1144 StringRef Indent) const {
1145 OS << Indent << "partition #" << (I - begin())
1146 << " [" << I->BeginOffset << "," << I->EndOffset << ")"
1147 << (I->IsSplittable ? " (splittable)" : "")
1148 << (Uses[I - begin()].empty() ? " (zero uses)" : "")
1152 void AllocaPartitioning::printUsers(raw_ostream &OS, const_iterator I,
1153 StringRef Indent) const {
1154 for (const_use_iterator UI = use_begin(I), UE = use_end(I);
1156 OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") "
1157 << "used by: " << *UI->User << "\n";
1158 if (MemTransferInst *II = dyn_cast<MemTransferInst>(&*UI->User)) {
1159 const MemTransferOffsets &MTO = MemTransferInstData.lookup(II);
1161 if (!MTO.IsSplittable)
1162 IsDest = UI->BeginOffset == MTO.DestBegin;
1164 IsDest = MTO.DestBegin != 0u;
1165 OS << Indent << " (original " << (IsDest ? "dest" : "source") << ": "
1166 << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin)
1167 << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n";
1172 void AllocaPartitioning::print(raw_ostream &OS) const {
1173 if (PointerEscapingInstr) {
1174 OS << "No partitioning for alloca: " << AI << "\n"
1175 << " A pointer to this alloca escaped by:\n"
1176 << " " << *PointerEscapingInstr << "\n";
1180 OS << "Partitioning of alloca: " << AI << "\n";
1182 for (const_iterator I = begin(), E = end(); I != E; ++I, ++Num) {
1188 void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); }
1189 void AllocaPartitioning::dump() const { print(dbgs()); }
1191 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1195 /// \brief Implementation of LoadAndStorePromoter for promoting allocas.
1197 /// This subclass of LoadAndStorePromoter adds overrides to handle promoting
1198 /// the loads and stores of an alloca instruction, as well as updating its
1199 /// debug information. This is used when a domtree is unavailable and thus
1200 /// mem2reg in its full form can't be used to handle promotion of allocas to
1202 class AllocaPromoter : public LoadAndStorePromoter {
1206 SmallVector<DbgDeclareInst *, 4> DDIs;
1207 SmallVector<DbgValueInst *, 4> DVIs;
1210 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S,
1211 AllocaInst &AI, DIBuilder &DIB)
1212 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {}
1214 void run(const SmallVectorImpl<Instruction*> &Insts) {
1215 // Remember which alloca we're promoting (for isInstInList).
1216 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) {
1217 for (Value::use_iterator UI = DebugNode->use_begin(),
1218 UE = DebugNode->use_end();
1220 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
1221 DDIs.push_back(DDI);
1222 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI))
1223 DVIs.push_back(DVI);
1226 LoadAndStorePromoter::run(Insts);
1227 AI.eraseFromParent();
1228 while (!DDIs.empty())
1229 DDIs.pop_back_val()->eraseFromParent();
1230 while (!DVIs.empty())
1231 DVIs.pop_back_val()->eraseFromParent();
1234 virtual bool isInstInList(Instruction *I,
1235 const SmallVectorImpl<Instruction*> &Insts) const {
1236 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1237 return LI->getOperand(0) == &AI;
1238 return cast<StoreInst>(I)->getPointerOperand() == &AI;
1241 virtual void updateDebugInfo(Instruction *Inst) const {
1242 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(),
1243 E = DDIs.end(); I != E; ++I) {
1244 DbgDeclareInst *DDI = *I;
1245 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
1246 ConvertDebugDeclareToDebugValue(DDI, SI, DIB);
1247 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
1248 ConvertDebugDeclareToDebugValue(DDI, LI, DIB);
1250 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(),
1251 E = DVIs.end(); I != E; ++I) {
1252 DbgValueInst *DVI = *I;
1254 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
1255 // If an argument is zero extended then use argument directly. The ZExt
1256 // may be zapped by an optimization pass in future.
1257 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0)))
1258 Arg = dyn_cast<Argument>(ZExt->getOperand(0));
1259 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0)))
1260 Arg = dyn_cast<Argument>(SExt->getOperand(0));
1262 Arg = SI->getOperand(0);
1263 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
1264 Arg = LI->getOperand(0);
1268 Instruction *DbgVal =
1269 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()),
1271 DbgVal->setDebugLoc(DVI->getDebugLoc());
1275 } // end anon namespace
1279 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1281 /// This pass takes allocations which can be completely analyzed (that is, they
1282 /// don't escape) and tries to turn them into scalar SSA values. There are
1283 /// a few steps to this process.
1285 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1286 /// are used to try to split them into smaller allocations, ideally of
1287 /// a single scalar data type. It will split up memcpy and memset accesses
1288 /// as necessary and try to isolate invidual scalar accesses.
1289 /// 2) It will transform accesses into forms which are suitable for SSA value
1290 /// promotion. This can be replacing a memset with a scalar store of an
1291 /// integer value, or it can involve speculating operations on a PHI or
1292 /// select to be a PHI or select of the results.
1293 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1294 /// onto insert and extract operations on a vector value, and convert them to
1295 /// this form. By doing so, it will enable promotion of vector aggregates to
1296 /// SSA vector values.
1297 class SROA : public FunctionPass {
1298 const bool RequiresDomTree;
1301 const TargetData *TD;
1304 /// \brief Worklist of alloca instructions to simplify.
1306 /// Each alloca in the function is added to this. Each new alloca formed gets
1307 /// added to it as well to recursively simplify unless that alloca can be
1308 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1309 /// the one being actively rewritten, we add it back onto the list if not
1310 /// already present to ensure it is re-visited.
1311 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
1313 /// \brief A collection of instructions to delete.
1314 /// We try to batch deletions to simplify code and make things a bit more
1316 SmallVector<Instruction *, 8> DeadInsts;
1318 /// \brief A set to prevent repeatedly marking an instruction split into many
1319 /// uses as dead. Only used to guard insertion into DeadInsts.
1320 SmallPtrSet<Instruction *, 4> DeadSplitInsts;
1322 /// \brief A collection of alloca instructions we can directly promote.
1323 std::vector<AllocaInst *> PromotableAllocas;
1326 SROA(bool RequiresDomTree = true)
1327 : FunctionPass(ID), RequiresDomTree(RequiresDomTree),
1328 C(0), TD(0), DT(0) {
1329 initializeSROAPass(*PassRegistry::getPassRegistry());
1331 bool runOnFunction(Function &F);
1332 void getAnalysisUsage(AnalysisUsage &AU) const;
1334 const char *getPassName() const { return "SROA"; }
1338 friend class AllocaPartitionRewriter;
1339 friend class AllocaPartitionVectorRewriter;
1341 bool rewriteAllocaPartition(AllocaInst &AI,
1342 AllocaPartitioning &P,
1343 AllocaPartitioning::iterator PI);
1344 bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P);
1345 bool runOnAlloca(AllocaInst &AI);
1346 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
1347 bool promoteAllocas(Function &F);
1353 FunctionPass *llvm::createSROAPass(bool RequiresDomTree) {
1354 return new SROA(RequiresDomTree);
1357 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1359 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
1360 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1363 /// \brief Accumulate the constant offsets in a GEP into a single APInt offset.
1365 /// If the provided GEP is all-constant, the total byte offset formed by the
1366 /// GEP is computed and Offset is set to it. If the GEP has any non-constant
1367 /// operands, the function returns false and the value of Offset is unmodified.
1368 static bool accumulateGEPOffsets(const TargetData &TD, GEPOperator &GEP,
1370 APInt GEPOffset(Offset.getBitWidth(), 0);
1371 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1372 GTI != GTE; ++GTI) {
1373 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
1376 if (OpC->isZero()) continue;
1378 // Handle a struct index, which adds its field offset to the pointer.
1379 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1380 unsigned ElementIdx = OpC->getZExtValue();
1381 const StructLayout *SL = TD.getStructLayout(STy);
1382 GEPOffset += APInt(Offset.getBitWidth(),
1383 SL->getElementOffset(ElementIdx));
1387 APInt TypeSize(Offset.getBitWidth(),
1388 TD.getTypeAllocSize(GTI.getIndexedType()));
1389 if (VectorType *VTy = dyn_cast<VectorType>(*GTI)) {
1390 assert((VTy->getScalarSizeInBits() % 8) == 0 &&
1391 "vector element size is not a multiple of 8, cannot GEP over it");
1392 TypeSize = VTy->getScalarSizeInBits() / 8;
1395 GEPOffset += OpC->getValue().sextOrTrunc(Offset.getBitWidth()) * TypeSize;
1401 /// \brief Build a GEP out of a base pointer and indices.
1403 /// This will return the BasePtr if that is valid, or build a new GEP
1404 /// instruction using the IRBuilder if GEP-ing is needed.
1405 static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr,
1406 SmallVectorImpl<Value *> &Indices,
1407 const Twine &Prefix) {
1408 if (Indices.empty())
1411 // A single zero index is a no-op, so check for this and avoid building a GEP
1413 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1416 return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx");
1419 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1420 /// TargetTy without changing the offset of the pointer.
1422 /// This routine assumes we've already established a properly offset GEP with
1423 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1424 /// zero-indices down through type layers until we find one the same as
1425 /// TargetTy. If we can't find one with the same type, we at least try to use
1426 /// one with the same size. If none of that works, we just produce the GEP as
1427 /// indicated by Indices to have the correct offset.
1428 static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const TargetData &TD,
1429 Value *BasePtr, Type *Ty, Type *TargetTy,
1430 SmallVectorImpl<Value *> &Indices,
1431 const Twine &Prefix) {
1433 return buildGEP(IRB, BasePtr, Indices, Prefix);
1435 // See if we can descend into a struct and locate a field with the correct
1437 unsigned NumLayers = 0;
1438 Type *ElementTy = Ty;
1440 if (ElementTy->isPointerTy())
1442 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1443 ElementTy = SeqTy->getElementType();
1444 Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(), 0)));
1445 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1446 ElementTy = *STy->element_begin();
1447 Indices.push_back(IRB.getInt32(0));
1452 } while (ElementTy != TargetTy);
1453 if (ElementTy != TargetTy)
1454 Indices.erase(Indices.end() - NumLayers, Indices.end());
1456 return buildGEP(IRB, BasePtr, Indices, Prefix);
1459 /// \brief Recursively compute indices for a natural GEP.
1461 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1462 /// element types adding appropriate indices for the GEP.
1463 static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const TargetData &TD,
1464 Value *Ptr, Type *Ty, APInt &Offset,
1466 SmallVectorImpl<Value *> &Indices,
1467 const Twine &Prefix) {
1469 return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix);
1471 // We can't recurse through pointer types.
1472 if (Ty->isPointerTy())
1475 // We try to analyze GEPs over vectors here, but note that these GEPs are
1476 // extremely poorly defined currently. The long-term goal is to remove GEPing
1477 // over a vector from the IR completely.
1478 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1479 unsigned ElementSizeInBits = VecTy->getScalarSizeInBits();
1480 if (ElementSizeInBits % 8)
1481 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1482 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1483 APInt NumSkippedElements = Offset.udiv(ElementSize);
1484 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1486 Offset -= NumSkippedElements * ElementSize;
1487 Indices.push_back(IRB.getInt(NumSkippedElements));
1488 return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
1489 Offset, TargetTy, Indices, Prefix);
1492 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1493 Type *ElementTy = ArrTy->getElementType();
1494 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1495 APInt NumSkippedElements = Offset.udiv(ElementSize);
1496 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1499 Offset -= NumSkippedElements * ElementSize;
1500 Indices.push_back(IRB.getInt(NumSkippedElements));
1501 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1505 StructType *STy = dyn_cast<StructType>(Ty);
1509 const StructLayout *SL = TD.getStructLayout(STy);
1510 uint64_t StructOffset = Offset.getZExtValue();
1511 if (StructOffset >= SL->getSizeInBytes())
1513 unsigned Index = SL->getElementContainingOffset(StructOffset);
1514 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1515 Type *ElementTy = STy->getElementType(Index);
1516 if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
1517 return 0; // The offset points into alignment padding.
1519 Indices.push_back(IRB.getInt32(Index));
1520 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1524 /// \brief Get a natural GEP from a base pointer to a particular offset and
1525 /// resulting in a particular type.
1527 /// The goal is to produce a "natural" looking GEP that works with the existing
1528 /// composite types to arrive at the appropriate offset and element type for
1529 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1530 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1531 /// Indices, and setting Ty to the result subtype.
1533 /// If no natural GEP can be constructed, this function returns null.
1534 static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const TargetData &TD,
1535 Value *Ptr, APInt Offset, Type *TargetTy,
1536 SmallVectorImpl<Value *> &Indices,
1537 const Twine &Prefix) {
1538 PointerType *Ty = cast<PointerType>(Ptr->getType());
1540 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1542 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1545 Type *ElementTy = Ty->getElementType();
1546 if (!ElementTy->isSized())
1547 return 0; // We can't GEP through an unsized element.
1548 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1549 if (ElementSize == 0)
1550 return 0; // Zero-length arrays can't help us build a natural GEP.
1551 APInt NumSkippedElements = Offset.udiv(ElementSize);
1553 Offset -= NumSkippedElements * ElementSize;
1554 Indices.push_back(IRB.getInt(NumSkippedElements));
1555 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1559 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1560 /// resulting pointer has PointerTy.
1562 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1563 /// and produces the pointer type desired. Where it cannot, it will try to use
1564 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1565 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1566 /// bitcast to the type.
1568 /// The strategy for finding the more natural GEPs is to peel off layers of the
1569 /// pointer, walking back through bit casts and GEPs, searching for a base
1570 /// pointer from which we can compute a natural GEP with the desired
1571 /// properities. The algorithm tries to fold as many constant indices into
1572 /// a single GEP as possible, thus making each GEP more independent of the
1573 /// surrounding code.
1574 static Value *getAdjustedPtr(IRBuilder<> &IRB, const TargetData &TD,
1575 Value *Ptr, APInt Offset, Type *PointerTy,
1576 const Twine &Prefix) {
1577 // Even though we don't look through PHI nodes, we could be called on an
1578 // instruction in an unreachable block, which may be on a cycle.
1579 SmallPtrSet<Value *, 4> Visited;
1580 Visited.insert(Ptr);
1581 SmallVector<Value *, 4> Indices;
1583 // We may end up computing an offset pointer that has the wrong type. If we
1584 // never are able to compute one directly that has the correct type, we'll
1585 // fall back to it, so keep it around here.
1586 Value *OffsetPtr = 0;
1588 // Remember any i8 pointer we come across to re-use if we need to do a raw
1591 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1593 Type *TargetTy = PointerTy->getPointerElementType();
1596 // First fold any existing GEPs into the offset.
1597 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1598 APInt GEPOffset(Offset.getBitWidth(), 0);
1599 if (!accumulateGEPOffsets(TD, *GEP, GEPOffset))
1601 Offset += GEPOffset;
1602 Ptr = GEP->getPointerOperand();
1603 if (!Visited.insert(Ptr))
1607 // See if we can perform a natural GEP here.
1609 if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
1611 if (P->getType() == PointerTy) {
1612 // Zap any offset pointer that we ended up computing in previous rounds.
1613 if (OffsetPtr && OffsetPtr->use_empty())
1614 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1615 I->eraseFromParent();
1623 // Stash this pointer if we've found an i8*.
1624 if (Ptr->getType()->isIntegerTy(8)) {
1626 Int8PtrOffset = Offset;
1629 // Peel off a layer of the pointer and update the offset appropriately.
1630 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1631 Ptr = cast<Operator>(Ptr)->getOperand(0);
1632 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1633 if (GA->mayBeOverridden())
1635 Ptr = GA->getAliasee();
1639 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1640 } while (Visited.insert(Ptr));
1644 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1645 Prefix + ".raw_cast");
1646 Int8PtrOffset = Offset;
1649 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1650 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1651 Prefix + ".raw_idx");
1655 // On the off chance we were targeting i8*, guard the bitcast here.
1656 if (Ptr->getType() != PointerTy)
1657 Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast");
1662 /// \brief Test whether the given alloca partition can be promoted to a vector.
1664 /// This is a quick test to check whether we can rewrite a particular alloca
1665 /// partition (and its newly formed alloca) into a vector alloca with only
1666 /// whole-vector loads and stores such that it could be promoted to a vector
1667 /// SSA value. We only can ensure this for a limited set of operations, and we
1668 /// don't want to do the rewrites unless we are confident that the result will
1669 /// be promotable, so we have an early test here.
1670 static bool isVectorPromotionViable(const TargetData &TD,
1672 AllocaPartitioning &P,
1673 uint64_t PartitionBeginOffset,
1674 uint64_t PartitionEndOffset,
1675 AllocaPartitioning::const_use_iterator I,
1676 AllocaPartitioning::const_use_iterator E) {
1677 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1681 uint64_t VecSize = TD.getTypeSizeInBits(Ty);
1682 uint64_t ElementSize = Ty->getScalarSizeInBits();
1684 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1685 // that aren't byte sized.
1686 if (ElementSize % 8)
1688 assert((VecSize % 8) == 0 && "vector size not a multiple of element size?");
1692 for (; I != E; ++I) {
1693 uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
1694 uint64_t BeginIndex = BeginOffset / ElementSize;
1695 if (BeginIndex * ElementSize != BeginOffset ||
1696 BeginIndex >= Ty->getNumElements())
1698 uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
1699 uint64_t EndIndex = EndOffset / ElementSize;
1700 if (EndIndex * ElementSize != EndOffset ||
1701 EndIndex > Ty->getNumElements())
1704 // FIXME: We should build shuffle vector instructions to handle
1705 // non-element-sized accesses.
1706 if ((EndOffset - BeginOffset) != ElementSize &&
1707 (EndOffset - BeginOffset) != VecSize)
1710 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(&*I->User)) {
1711 if (MI->isVolatile())
1713 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(&*I->User)) {
1714 const AllocaPartitioning::MemTransferOffsets &MTO
1715 = P.getMemTransferOffsets(*MTI);
1716 if (!MTO.IsSplittable)
1719 } else if (I->Ptr->getType()->getPointerElementType()->isStructTy()) {
1720 // Disable vector promotion when there are loads or stores of an FCA.
1722 } else if (!isa<LoadInst>(*I->User) && !isa<StoreInst>(*I->User)) {
1729 /// \brief Test whether the given alloca partition can be promoted to an int.
1731 /// This is a quick test to check whether we can rewrite a particular alloca
1732 /// partition (and its newly formed alloca) into an integer alloca suitable for
1733 /// promotion to an SSA value. We only can ensure this for a limited set of
1734 /// operations, and we don't want to do the rewrites unless we are confident
1735 /// that the result will be promotable, so we have an early test here.
1736 static bool isIntegerPromotionViable(const TargetData &TD,
1738 AllocaPartitioning &P,
1739 AllocaPartitioning::const_use_iterator I,
1740 AllocaPartitioning::const_use_iterator E) {
1741 IntegerType *Ty = dyn_cast<IntegerType>(AllocaTy);
1745 // Check the uses to ensure the uses are (likely) promoteable integer uses.
1746 // Also ensure that the alloca has a covering load or store. We don't want
1747 // promote because of some other unsplittable entry (which we may make
1748 // splittable later) and lose the ability to promote each element access.
1749 bool WholeAllocaOp = false;
1750 for (; I != E; ++I) {
1751 if (LoadInst *LI = dyn_cast<LoadInst>(&*I->User)) {
1752 if (LI->isVolatile() || !LI->getType()->isIntegerTy())
1754 if (LI->getType() == Ty)
1755 WholeAllocaOp = true;
1756 } else if (StoreInst *SI = dyn_cast<StoreInst>(&*I->User)) {
1757 if (SI->isVolatile() || !SI->getValueOperand()->getType()->isIntegerTy())
1759 if (SI->getValueOperand()->getType() == Ty)
1760 WholeAllocaOp = true;
1761 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(&*I->User)) {
1762 if (MI->isVolatile())
1764 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(&*I->User)) {
1765 const AllocaPartitioning::MemTransferOffsets &MTO
1766 = P.getMemTransferOffsets(*MTI);
1767 if (!MTO.IsSplittable)
1774 return WholeAllocaOp;
1778 /// \brief Visitor to rewrite instructions using a partition of an alloca to
1779 /// use a new alloca.
1781 /// Also implements the rewriting to vector-based accesses when the partition
1782 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
1784 class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
1786 // Befriend the base class so it can delegate to private visit methods.
1787 friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
1789 const TargetData &TD;
1790 AllocaPartitioning &P;
1792 AllocaInst &OldAI, &NewAI;
1793 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
1795 // If we are rewriting an alloca partition which can be written as pure
1796 // vector operations, we stash extra information here. When VecTy is
1797 // non-null, we have some strict guarantees about the rewriten alloca:
1798 // - The new alloca is exactly the size of the vector type here.
1799 // - The accesses all either map to the entire vector or to a single
1801 // - The set of accessing instructions is only one of those handled above
1802 // in isVectorPromotionViable. Generally these are the same access kinds
1803 // which are promotable via mem2reg.
1806 uint64_t ElementSize;
1808 // This is a convenience and flag variable that will be null unless the new
1809 // alloca has a promotion-targeted integer type due to passing
1810 // isIntegerPromotionViable above. If it is non-null does, the desired
1811 // integer type will be stored here for easy access during rewriting.
1812 IntegerType *IntPromotionTy;
1814 // The offset of the partition user currently being rewritten.
1815 uint64_t BeginOffset, EndOffset;
1816 Instruction *OldPtr;
1818 // The name prefix to use when rewriting instructions for this alloca.
1819 std::string NamePrefix;
1822 AllocaPartitionRewriter(const TargetData &TD, AllocaPartitioning &P,
1823 AllocaPartitioning::iterator PI,
1824 SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
1825 uint64_t NewBeginOffset, uint64_t NewEndOffset)
1826 : TD(TD), P(P), Pass(Pass),
1827 OldAI(OldAI), NewAI(NewAI),
1828 NewAllocaBeginOffset(NewBeginOffset),
1829 NewAllocaEndOffset(NewEndOffset),
1830 VecTy(), ElementTy(), ElementSize(), IntPromotionTy(),
1831 BeginOffset(), EndOffset() {
1834 /// \brief Visit the users of the alloca partition and rewrite them.
1835 bool visitUsers(AllocaPartitioning::const_use_iterator I,
1836 AllocaPartitioning::const_use_iterator E) {
1837 if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P,
1838 NewAllocaBeginOffset, NewAllocaEndOffset,
1841 VecTy = cast<VectorType>(NewAI.getAllocatedType());
1842 ElementTy = VecTy->getElementType();
1843 assert((VecTy->getScalarSizeInBits() % 8) == 0 &&
1844 "Only multiple-of-8 sized vector elements are viable");
1845 ElementSize = VecTy->getScalarSizeInBits() / 8;
1846 } else if (isIntegerPromotionViable(TD, NewAI.getAllocatedType(),
1848 IntPromotionTy = cast<IntegerType>(NewAI.getAllocatedType());
1850 bool CanSROA = true;
1851 for (; I != E; ++I) {
1852 BeginOffset = I->BeginOffset;
1853 EndOffset = I->EndOffset;
1855 NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str();
1856 CanSROA &= visit(I->User);
1868 // Every instruction which can end up as a user must have a rewrite rule.
1869 bool visitInstruction(Instruction &I) {
1870 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
1871 llvm_unreachable("No rewrite rule for this instruction!");
1874 Twine getName(const Twine &Suffix) {
1875 return NamePrefix + Suffix;
1878 Value *getAdjustedAllocaPtr(IRBuilder<> &IRB, Type *PointerTy) {
1879 assert(BeginOffset >= NewAllocaBeginOffset);
1880 APInt Offset(TD.getPointerSizeInBits(), BeginOffset - NewAllocaBeginOffset);
1881 return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName(""));
1884 ConstantInt *getIndex(IRBuilder<> &IRB, uint64_t Offset) {
1885 assert(VecTy && "Can only call getIndex when rewriting a vector");
1886 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
1887 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
1888 uint32_t Index = RelOffset / ElementSize;
1889 assert(Index * ElementSize == RelOffset);
1890 return IRB.getInt32(Index);
1893 Value *extractInteger(IRBuilder<> &IRB, IntegerType *TargetTy,
1895 assert(IntPromotionTy && "Alloca is not an integer we can extract from");
1896 Value *V = IRB.CreateLoad(&NewAI, getName(".load"));
1897 assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
1898 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
1900 V = IRB.CreateLShr(V, RelOffset*8, getName(".shift"));
1901 if (TargetTy != IntPromotionTy) {
1902 assert(TargetTy->getBitWidth() < IntPromotionTy->getBitWidth() &&
1903 "Cannot extract to a larger integer!");
1904 V = IRB.CreateTrunc(V, TargetTy, getName(".trunc"));
1909 StoreInst *insertInteger(IRBuilder<> &IRB, Value *V, uint64_t Offset) {
1910 IntegerType *Ty = cast<IntegerType>(V->getType());
1911 if (Ty == IntPromotionTy)
1912 return IRB.CreateStore(V, &NewAI);
1914 assert(Ty->getBitWidth() < IntPromotionTy->getBitWidth() &&
1915 "Cannot insert a larger integer!");
1916 V = IRB.CreateZExt(V, IntPromotionTy, getName(".ext"));
1917 assert(Offset >= NewAllocaBeginOffset && "Out of bounds offset");
1918 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
1920 V = IRB.CreateShl(V, RelOffset*8, getName(".shift"));
1922 APInt Mask = ~Ty->getMask().zext(IntPromotionTy->getBitWidth())
1924 Value *Old = IRB.CreateAnd(IRB.CreateLoad(&NewAI, getName(".oldload")),
1925 Mask, getName(".mask"));
1926 return IRB.CreateStore(IRB.CreateOr(Old, V, getName(".insert")),
1930 void deleteIfTriviallyDead(Value *V) {
1931 Instruction *I = cast<Instruction>(V);
1932 if (isInstructionTriviallyDead(I))
1933 Pass.DeadInsts.push_back(I);
1936 Value *getValueCast(IRBuilder<> &IRB, Value *V, Type *Ty) {
1937 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
1938 return IRB.CreateIntToPtr(V, Ty);
1939 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
1940 return IRB.CreatePtrToInt(V, Ty);
1942 return IRB.CreateBitCast(V, Ty);
1945 bool rewriteVectorizedLoadInst(IRBuilder<> &IRB, LoadInst &LI, Value *OldOp) {
1947 if (LI.getType() == VecTy->getElementType() ||
1948 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
1950 = IRB.CreateExtractElement(IRB.CreateLoad(&NewAI, getName(".load")),
1951 getIndex(IRB, BeginOffset),
1952 getName(".extract"));
1954 Result = IRB.CreateLoad(&NewAI, getName(".load"));
1956 if (Result->getType() != LI.getType())
1957 Result = getValueCast(IRB, Result, LI.getType());
1958 LI.replaceAllUsesWith(Result);
1959 Pass.DeadInsts.push_back(&LI);
1961 DEBUG(dbgs() << " to: " << *Result << "\n");
1965 bool rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) {
1966 assert(!LI.isVolatile());
1967 Value *Result = extractInteger(IRB, cast<IntegerType>(LI.getType()),
1969 LI.replaceAllUsesWith(Result);
1970 Pass.DeadInsts.push_back(&LI);
1971 DEBUG(dbgs() << " to: " << *Result << "\n");
1975 bool visitLoadInst(LoadInst &LI) {
1976 DEBUG(dbgs() << " original: " << LI << "\n");
1977 Value *OldOp = LI.getOperand(0);
1978 assert(OldOp == OldPtr);
1979 IRBuilder<> IRB(&LI);
1982 return rewriteVectorizedLoadInst(IRB, LI, OldOp);
1984 return rewriteIntegerLoad(IRB, LI);
1986 Value *NewPtr = getAdjustedAllocaPtr(IRB,
1987 LI.getPointerOperand()->getType());
1988 LI.setOperand(0, NewPtr);
1989 DEBUG(dbgs() << " to: " << LI << "\n");
1991 deleteIfTriviallyDead(OldOp);
1992 return NewPtr == &NewAI && !LI.isVolatile();
1995 bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, StoreInst &SI,
1997 Value *V = SI.getValueOperand();
1998 if (V->getType() == ElementTy ||
1999 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
2000 if (V->getType() != ElementTy)
2001 V = getValueCast(IRB, V, ElementTy);
2002 V = IRB.CreateInsertElement(IRB.CreateLoad(&NewAI, getName(".load")), V,
2003 getIndex(IRB, BeginOffset),
2004 getName(".insert"));
2005 } else if (V->getType() != VecTy) {
2006 V = getValueCast(IRB, V, VecTy);
2008 StoreInst *Store = IRB.CreateStore(V, &NewAI);
2009 Pass.DeadInsts.push_back(&SI);
2012 DEBUG(dbgs() << " to: " << *Store << "\n");
2016 bool rewriteIntegerStore(IRBuilder<> &IRB, StoreInst &SI) {
2017 assert(!SI.isVolatile());
2018 StoreInst *Store = insertInteger(IRB, SI.getValueOperand(), BeginOffset);
2019 Pass.DeadInsts.push_back(&SI);
2021 DEBUG(dbgs() << " to: " << *Store << "\n");
2025 bool visitStoreInst(StoreInst &SI) {
2026 DEBUG(dbgs() << " original: " << SI << "\n");
2027 Value *OldOp = SI.getOperand(1);
2028 assert(OldOp == OldPtr);
2029 IRBuilder<> IRB(&SI);
2032 return rewriteVectorizedStoreInst(IRB, SI, OldOp);
2034 return rewriteIntegerStore(IRB, SI);
2036 Value *NewPtr = getAdjustedAllocaPtr(IRB,
2037 SI.getPointerOperand()->getType());
2038 SI.setOperand(1, NewPtr);
2039 DEBUG(dbgs() << " to: " << SI << "\n");
2041 deleteIfTriviallyDead(OldOp);
2042 return NewPtr == &NewAI && !SI.isVolatile();
2045 bool visitMemSetInst(MemSetInst &II) {
2046 DEBUG(dbgs() << " original: " << II << "\n");
2047 IRBuilder<> IRB(&II);
2048 assert(II.getRawDest() == OldPtr);
2050 // If the memset has a variable size, it cannot be split, just adjust the
2051 // pointer to the new alloca.
2052 if (!isa<Constant>(II.getLength())) {
2053 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2054 deleteIfTriviallyDead(OldPtr);
2058 // Record this instruction for deletion.
2059 if (Pass.DeadSplitInsts.insert(&II))
2060 Pass.DeadInsts.push_back(&II);
2062 Type *AllocaTy = NewAI.getAllocatedType();
2063 Type *ScalarTy = AllocaTy->getScalarType();
2065 // If this doesn't map cleanly onto the alloca type, and that type isn't
2066 // a single value type, just emit a memset.
2067 if (!VecTy && (BeginOffset != NewAllocaBeginOffset ||
2068 EndOffset != NewAllocaEndOffset ||
2069 !AllocaTy->isSingleValueType() ||
2070 !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)))) {
2071 Type *SizeTy = II.getLength()->getType();
2072 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2075 = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB,
2076 II.getRawDest()->getType()),
2077 II.getValue(), Size, II.getAlignment(),
2080 DEBUG(dbgs() << " to: " << *New << "\n");
2084 // If we can represent this as a simple value, we have to build the actual
2085 // value to store, which requires expanding the byte present in memset to
2086 // a sensible representation for the alloca type. This is essentially
2087 // splatting the byte to a sufficiently wide integer, bitcasting to the
2088 // desired scalar type, and splatting it across any desired vector type.
2089 Value *V = II.getValue();
2090 IntegerType *VTy = cast<IntegerType>(V->getType());
2091 Type *IntTy = Type::getIntNTy(VTy->getContext(),
2092 TD.getTypeSizeInBits(ScalarTy));
2093 if (TD.getTypeSizeInBits(ScalarTy) > VTy->getBitWidth())
2094 V = IRB.CreateMul(IRB.CreateZExt(V, IntTy, getName(".zext")),
2095 ConstantExpr::getUDiv(
2096 Constant::getAllOnesValue(IntTy),
2097 ConstantExpr::getZExt(
2098 Constant::getAllOnesValue(V->getType()),
2100 getName(".isplat"));
2101 if (V->getType() != ScalarTy) {
2102 if (ScalarTy->isPointerTy())
2103 V = IRB.CreateIntToPtr(V, ScalarTy);
2104 else if (ScalarTy->isPrimitiveType() || ScalarTy->isVectorTy())
2105 V = IRB.CreateBitCast(V, ScalarTy);
2106 else if (ScalarTy->isIntegerTy())
2107 llvm_unreachable("Computed different integer types with equal widths");
2109 llvm_unreachable("Invalid scalar type");
2112 // If this is an element-wide memset of a vectorizable alloca, insert it.
2113 if (VecTy && (BeginOffset > NewAllocaBeginOffset ||
2114 EndOffset < NewAllocaEndOffset)) {
2115 StoreInst *Store = IRB.CreateStore(
2116 IRB.CreateInsertElement(IRB.CreateLoad(&NewAI, getName(".load")), V,
2117 getIndex(IRB, BeginOffset),
2118 getName(".insert")),
2121 DEBUG(dbgs() << " to: " << *Store << "\n");
2125 // Splat to a vector if needed.
2126 if (VectorType *VecTy = dyn_cast<VectorType>(AllocaTy)) {
2127 VectorType *SplatSourceTy = VectorType::get(V->getType(), 1);
2128 V = IRB.CreateShuffleVector(
2129 IRB.CreateInsertElement(UndefValue::get(SplatSourceTy), V,
2130 IRB.getInt32(0), getName(".vsplat.insert")),
2131 UndefValue::get(SplatSourceTy),
2132 ConstantVector::getSplat(VecTy->getNumElements(), IRB.getInt32(0)),
2133 getName(".vsplat.shuffle"));
2134 assert(V->getType() == VecTy);
2137 Value *New = IRB.CreateStore(V, &NewAI, II.isVolatile());
2139 DEBUG(dbgs() << " to: " << *New << "\n");
2140 return !II.isVolatile();
2143 bool visitMemTransferInst(MemTransferInst &II) {
2144 // Rewriting of memory transfer instructions can be a bit tricky. We break
2145 // them into two categories: split intrinsics and unsplit intrinsics.
2147 DEBUG(dbgs() << " original: " << II << "\n");
2148 IRBuilder<> IRB(&II);
2150 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
2151 bool IsDest = II.getRawDest() == OldPtr;
2153 const AllocaPartitioning::MemTransferOffsets &MTO
2154 = P.getMemTransferOffsets(II);
2156 // For unsplit intrinsics, we simply modify the source and destination
2157 // pointers in place. This isn't just an optimization, it is a matter of
2158 // correctness. With unsplit intrinsics we may be dealing with transfers
2159 // within a single alloca before SROA ran, or with transfers that have
2160 // a variable length. We may also be dealing with memmove instead of
2161 // memcpy, and so simply updating the pointers is the necessary for us to
2162 // update both source and dest of a single call.
2163 if (!MTO.IsSplittable) {
2164 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
2166 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
2168 II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType()));
2170 DEBUG(dbgs() << " to: " << II << "\n");
2171 deleteIfTriviallyDead(OldOp);
2174 // For split transfer intrinsics we have an incredibly useful assurance:
2175 // the source and destination do not reside within the same alloca, and at
2176 // least one of them does not escape. This means that we can replace
2177 // memmove with memcpy, and we don't need to worry about all manner of
2178 // downsides to splitting and transforming the operations.
2180 // Compute the relative offset within the transfer.
2181 unsigned IntPtrWidth = TD.getPointerSizeInBits();
2182 APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
2183 : MTO.SourceBegin));
2185 // If this doesn't map cleanly onto the alloca type, and that type isn't
2186 // a single value type, just emit a memcpy.
2188 = !VecTy && (BeginOffset != NewAllocaBeginOffset ||
2189 EndOffset != NewAllocaEndOffset ||
2190 !NewAI.getAllocatedType()->isSingleValueType());
2192 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
2193 // size hasn't been shrunk based on analysis of the viable range, this is
2195 if (EmitMemCpy && &OldAI == &NewAI) {
2196 uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin;
2197 uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd;
2198 // Ensure the start lines up.
2199 assert(BeginOffset == OrigBegin);
2202 // Rewrite the size as needed.
2203 if (EndOffset != OrigEnd)
2204 II.setLength(ConstantInt::get(II.getLength()->getType(),
2205 EndOffset - BeginOffset));
2208 // Record this instruction for deletion.
2209 if (Pass.DeadSplitInsts.insert(&II))
2210 Pass.DeadInsts.push_back(&II);
2212 bool IsVectorElement = VecTy && (BeginOffset > NewAllocaBeginOffset ||
2213 EndOffset < NewAllocaEndOffset);
2215 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
2216 : II.getRawDest()->getType();
2218 OtherPtrTy = IsVectorElement ? VecTy->getElementType()->getPointerTo()
2221 // Compute the other pointer, folding as much as possible to produce
2222 // a single, simple GEP in most cases.
2223 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
2224 OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
2225 getName("." + OtherPtr->getName()));
2227 // Strip all inbounds GEPs and pointer casts to try to dig out any root
2228 // alloca that should be re-examined after rewriting this instruction.
2230 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
2231 Pass.Worklist.insert(AI);
2235 = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType()
2236 : II.getRawSource()->getType());
2237 Type *SizeTy = II.getLength()->getType();
2238 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
2240 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
2241 IsDest ? OtherPtr : OurPtr,
2242 Size, II.getAlignment(),
2245 DEBUG(dbgs() << " to: " << *New << "\n");
2249 Value *SrcPtr = OtherPtr;
2250 Value *DstPtr = &NewAI;
2252 std::swap(SrcPtr, DstPtr);
2255 if (IsVectorElement && !IsDest) {
2256 // We have to extract rather than load.
2257 Src = IRB.CreateExtractElement(IRB.CreateLoad(SrcPtr,
2258 getName(".copyload")),
2259 getIndex(IRB, BeginOffset),
2260 getName(".copyextract"));
2262 Src = IRB.CreateLoad(SrcPtr, II.isVolatile(), getName(".copyload"));
2265 if (IsVectorElement && IsDest) {
2266 // We have to insert into a loaded copy before storing.
2267 Src = IRB.CreateInsertElement(IRB.CreateLoad(&NewAI, getName(".load")),
2268 Src, getIndex(IRB, BeginOffset),
2269 getName(".insert"));
2272 Value *Store = IRB.CreateStore(Src, DstPtr, II.isVolatile());
2274 DEBUG(dbgs() << " to: " << *Store << "\n");
2275 return !II.isVolatile();
2278 bool visitIntrinsicInst(IntrinsicInst &II) {
2279 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
2280 II.getIntrinsicID() == Intrinsic::lifetime_end);
2281 DEBUG(dbgs() << " original: " << II << "\n");
2282 IRBuilder<> IRB(&II);
2283 assert(II.getArgOperand(1) == OldPtr);
2285 // Record this instruction for deletion.
2286 if (Pass.DeadSplitInsts.insert(&II))
2287 Pass.DeadInsts.push_back(&II);
2290 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
2291 EndOffset - BeginOffset);
2292 Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType());
2294 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2295 New = IRB.CreateLifetimeStart(Ptr, Size);
2297 New = IRB.CreateLifetimeEnd(Ptr, Size);
2299 DEBUG(dbgs() << " to: " << *New << "\n");
2303 /// PHI instructions that use an alloca and are subsequently loaded can be
2304 /// rewritten to load both input pointers in the pred blocks and then PHI the
2305 /// results, allowing the load of the alloca to be promoted.
2307 /// %P2 = phi [i32* %Alloca, i32* %Other]
2308 /// %V = load i32* %P2
2310 /// %V1 = load i32* %Alloca -> will be mem2reg'd
2312 /// %V2 = load i32* %Other
2314 /// %V = phi [i32 %V1, i32 %V2]
2316 /// We can do this to a select if its only uses are loads and if the operand
2317 /// to the select can be loaded unconditionally.
2319 /// FIXME: This should be hoisted into a generic utility, likely in
2320 /// Transforms/Util/Local.h
2321 bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
2322 // For now, we can only do this promotion if the load is in the same block
2323 // as the PHI, and if there are no stores between the phi and load.
2324 // TODO: Allow recursive phi users.
2325 // TODO: Allow stores.
2326 BasicBlock *BB = PN.getParent();
2327 unsigned MaxAlign = 0;
2328 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
2330 LoadInst *LI = dyn_cast<LoadInst>(*UI);
2331 if (LI == 0 || !LI->isSimple()) return false;
2333 // For now we only allow loads in the same block as the PHI. This is
2334 // a common case that happens when instcombine merges two loads through
2336 if (LI->getParent() != BB) return false;
2338 // Ensure that there are no instructions between the PHI and the load that
2340 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
2341 if (BBI->mayWriteToMemory())
2344 MaxAlign = std::max(MaxAlign, LI->getAlignment());
2345 Loads.push_back(LI);
2348 // We can only transform this if it is safe to push the loads into the
2349 // predecessor blocks. The only thing to watch out for is that we can't put
2350 // a possibly trapping load in the predecessor if it is a critical edge.
2351 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num;
2353 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
2354 Value *InVal = PN.getIncomingValue(Idx);
2356 // If the value is produced by the terminator of the predecessor (an
2357 // invoke) or it has side-effects, there is no valid place to put a load
2358 // in the predecessor.
2359 if (TI == InVal || TI->mayHaveSideEffects())
2362 // If the predecessor has a single successor, then the edge isn't
2364 if (TI->getNumSuccessors() == 1)
2367 // If this pointer is always safe to load, or if we can prove that there
2368 // is already a load in the block, then we can move the load to the pred
2370 if (InVal->isDereferenceablePointer() ||
2371 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
2380 bool visitPHINode(PHINode &PN) {
2381 DEBUG(dbgs() << " original: " << PN << "\n");
2382 // We would like to compute a new pointer in only one place, but have it be
2383 // as local as possible to the PHI. To do that, we re-use the location of
2384 // the old pointer, which necessarily must be in the right position to
2385 // dominate the PHI.
2386 IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr));
2388 SmallVector<LoadInst *, 4> Loads;
2389 if (!isSafePHIToSpeculate(PN, Loads)) {
2390 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
2391 // Replace the operands which were using the old pointer.
2392 User::op_iterator OI = PN.op_begin(), OE = PN.op_end();
2393 for (; OI != OE; ++OI)
2397 DEBUG(dbgs() << " to: " << PN << "\n");
2398 deleteIfTriviallyDead(OldPtr);
2401 assert(!Loads.empty());
2403 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
2404 IRBuilder<> PHIBuilder(&PN);
2405 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues());
2406 NewPN->takeName(&PN);
2408 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
2409 // matter which one we get and if any differ, it doesn't matter.
2410 LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
2411 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
2412 unsigned Align = SomeLoad->getAlignment();
2413 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
2415 // Rewrite all loads of the PN to use the new PHI.
2417 LoadInst *LI = Loads.pop_back_val();
2418 LI->replaceAllUsesWith(NewPN);
2419 Pass.DeadInsts.push_back(LI);
2420 } while (!Loads.empty());
2422 // Inject loads into all of the pred blocks.
2423 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
2424 BasicBlock *Pred = PN.getIncomingBlock(Idx);
2425 TerminatorInst *TI = Pred->getTerminator();
2426 Value *InVal = PN.getIncomingValue(Idx);
2427 IRBuilder<> PredBuilder(TI);
2429 // Map the value to the new alloca pointer if this was the old alloca
2431 bool ThisOperand = InVal == OldPtr;
2436 = PredBuilder.CreateLoad(InVal, getName(".sroa.speculate." +
2438 ++NumLoadsSpeculated;
2439 Load->setAlignment(Align);
2441 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
2442 NewPN->addIncoming(Load, Pred);
2446 Instruction *OtherPtr = dyn_cast<Instruction>(InVal);
2448 // No uses to rewrite.
2451 // Try to lookup and rewrite any partition uses corresponding to this phi
2453 AllocaPartitioning::iterator PI
2454 = P.findPartitionForPHIOrSelectOperand(PN, OtherPtr);
2455 if (PI != P.end()) {
2456 // If the other pointer is within the partitioning, replace the PHI in
2457 // its uses with the load we just speculated, or add another load for
2458 // it to rewrite if we've already replaced the PHI.
2459 AllocaPartitioning::use_iterator UI
2460 = P.findPartitionUseForPHIOrSelectOperand(PN, OtherPtr);
2461 if (isa<PHINode>(*UI->User))
2464 AllocaPartitioning::PartitionUse OtherUse = *UI;
2465 OtherUse.User = Load;
2466 P.use_push_back(PI, OtherUse);
2470 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
2471 return NewPtr == &NewAI;
2474 /// Select instructions that use an alloca and are subsequently loaded can be
2475 /// rewritten to load both input pointers and then select between the result,
2476 /// allowing the load of the alloca to be promoted.
2478 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
2479 /// %V = load i32* %P2
2481 /// %V1 = load i32* %Alloca -> will be mem2reg'd
2482 /// %V2 = load i32* %Other
2483 /// %V = select i1 %cond, i32 %V1, i32 %V2
2485 /// We can do this to a select if its only uses are loads and if the operand
2486 /// to the select can be loaded unconditionally.
2487 bool isSafeSelectToSpeculate(SelectInst &SI,
2488 SmallVectorImpl<LoadInst *> &Loads) {
2489 Value *TValue = SI.getTrueValue();
2490 Value *FValue = SI.getFalseValue();
2491 bool TDerefable = TValue->isDereferenceablePointer();
2492 bool FDerefable = FValue->isDereferenceablePointer();
2494 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
2496 LoadInst *LI = dyn_cast<LoadInst>(*UI);
2497 if (LI == 0 || !LI->isSimple()) return false;
2499 // Both operands to the select need to be dereferencable, either
2500 // absolutely (e.g. allocas) or at this point because we can see other
2502 if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
2503 LI->getAlignment(), &TD))
2505 if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
2506 LI->getAlignment(), &TD))
2508 Loads.push_back(LI);
2514 bool visitSelectInst(SelectInst &SI) {
2515 DEBUG(dbgs() << " original: " << SI << "\n");
2516 IRBuilder<> IRB(&SI);
2518 // Find the operand we need to rewrite here.
2519 bool IsTrueVal = SI.getTrueValue() == OldPtr;
2521 assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!");
2523 assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!");
2524 Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
2526 // If the select isn't safe to speculate, just use simple logic to emit it.
2527 SmallVector<LoadInst *, 4> Loads;
2528 if (!isSafeSelectToSpeculate(SI, Loads)) {
2529 SI.setOperand(IsTrueVal ? 1 : 2, NewPtr);
2530 DEBUG(dbgs() << " to: " << SI << "\n");
2531 deleteIfTriviallyDead(OldPtr);
2535 Value *OtherPtr = IsTrueVal ? SI.getFalseValue() : SI.getTrueValue();
2536 AllocaPartitioning::iterator PI
2537 = P.findPartitionForPHIOrSelectOperand(SI, OtherPtr);
2538 AllocaPartitioning::PartitionUse OtherUse;
2539 if (PI != P.end()) {
2540 // If the other pointer is within the partitioning, remove the select
2541 // from its uses. We'll add in the new loads below.
2542 AllocaPartitioning::use_iterator UI
2543 = P.findPartitionUseForPHIOrSelectOperand(SI, OtherPtr);
2545 P.use_erase(PI, UI);
2548 Value *TV = IsTrueVal ? NewPtr : SI.getTrueValue();
2549 Value *FV = IsTrueVal ? SI.getFalseValue() : NewPtr;
2550 // Replace the loads of the select with a select of two loads.
2551 while (!Loads.empty()) {
2552 LoadInst *LI = Loads.pop_back_val();
2554 IRB.SetInsertPoint(LI);
2556 IRB.CreateLoad(TV, getName("." + LI->getName() + ".true"));
2558 IRB.CreateLoad(FV, getName("." + LI->getName() + ".false"));
2559 NumLoadsSpeculated += 2;
2560 if (PI != P.end()) {
2561 LoadInst *OtherLoad = IsTrueVal ? FL : TL;
2562 assert(OtherUse.Ptr == OtherLoad->getOperand(0));
2563 OtherUse.User = OtherLoad;
2564 P.use_push_back(PI, OtherUse);
2567 // Transfer alignment and TBAA info if present.
2568 TL->setAlignment(LI->getAlignment());
2569 FL->setAlignment(LI->getAlignment());
2570 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
2571 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
2572 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
2575 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL);
2577 DEBUG(dbgs() << " speculated to: " << *V << "\n");
2578 LI->replaceAllUsesWith(V);
2579 Pass.DeadInsts.push_back(LI);
2582 deleteIfTriviallyDead(OldPtr);
2583 return NewPtr == &NewAI;
2590 /// \brief Visitor to rewrite aggregate loads and stores as scalar.
2592 /// This pass aggressively rewrites all aggregate loads and stores on
2593 /// a particular pointer (or any pointer derived from it which we can identify)
2594 /// with scalar loads and stores.
2595 class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> {
2596 // Befriend the base class so it can delegate to private visit methods.
2597 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>;
2599 const TargetData &TD;
2601 /// Queue of pointer uses to analyze and potentially rewrite.
2602 SmallVector<Use *, 8> Queue;
2604 /// Set to prevent us from cycling with phi nodes and loops.
2605 SmallPtrSet<User *, 8> Visited;
2607 /// The current pointer use being rewritten. This is used to dig up the used
2608 /// value (as opposed to the user).
2612 AggLoadStoreRewriter(const TargetData &TD) : TD(TD) {}
2614 /// Rewrite loads and stores through a pointer and all pointers derived from
2616 bool rewrite(Instruction &I) {
2617 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n");
2619 bool Changed = false;
2620 while (!Queue.empty()) {
2621 U = Queue.pop_back_val();
2622 Changed |= visit(cast<Instruction>(U->getUser()));
2628 /// Enqueue all the users of the given instruction for further processing.
2629 /// This uses a set to de-duplicate users.
2630 void enqueueUsers(Instruction &I) {
2631 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE;
2633 if (Visited.insert(*UI))
2634 Queue.push_back(&UI.getUse());
2637 // Conservative default is to not rewrite anything.
2638 bool visitInstruction(Instruction &I) { return false; }
2640 /// \brief Generic recursive split emission class.
2641 template <typename Derived>
2644 /// The builder used to form new instructions.
2646 /// The indices which to be used with insert- or extractvalue to select the
2647 /// appropriate value within the aggregate.
2648 SmallVector<unsigned, 4> Indices;
2649 /// The indices to a GEP instruction which will move Ptr to the correct slot
2650 /// within the aggregate.
2651 SmallVector<Value *, 4> GEPIndices;
2652 /// The base pointer of the original op, used as a base for GEPing the
2653 /// split operations.
2656 /// Initialize the splitter with an insertion point, Ptr and start with a
2657 /// single zero GEP index.
2658 OpSplitter(Instruction *InsertionPoint, Value *Ptr)
2659 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {}
2662 /// \brief Generic recursive split emission routine.
2664 /// This method recursively splits an aggregate op (load or store) into
2665 /// scalar or vector ops. It splits recursively until it hits a single value
2666 /// and emits that single value operation via the template argument.
2668 /// The logic of this routine relies on GEPs and insertvalue and
2669 /// extractvalue all operating with the same fundamental index list, merely
2670 /// formatted differently (GEPs need actual values).
2672 /// \param Ty The type being split recursively into smaller ops.
2673 /// \param Agg The aggregate value being built up or stored, depending on
2674 /// whether this is splitting a load or a store respectively.
2675 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) {
2676 if (Ty->isSingleValueType())
2677 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name);
2679 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
2680 unsigned OldSize = Indices.size();
2682 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size;
2684 assert(Indices.size() == OldSize && "Did not return to the old size");
2685 Indices.push_back(Idx);
2686 GEPIndices.push_back(IRB.getInt32(Idx));
2687 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx));
2688 GEPIndices.pop_back();
2694 if (StructType *STy = dyn_cast<StructType>(Ty)) {
2695 unsigned OldSize = Indices.size();
2697 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size;
2699 assert(Indices.size() == OldSize && "Did not return to the old size");
2700 Indices.push_back(Idx);
2701 GEPIndices.push_back(IRB.getInt32(Idx));
2702 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx));
2703 GEPIndices.pop_back();
2709 llvm_unreachable("Only arrays and structs are aggregate loadable types");
2713 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> {
2714 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2715 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {}
2717 /// Emit a leaf load of a single value. This is called at the leaves of the
2718 /// recursive emission to actually load values.
2719 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2720 assert(Ty->isSingleValueType());
2721 // Load the single value and insert it using the indices.
2722 Value *Load = IRB.CreateLoad(IRB.CreateInBoundsGEP(Ptr, GEPIndices,
2725 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert");
2726 DEBUG(dbgs() << " to: " << *Load << "\n");
2730 bool visitLoadInst(LoadInst &LI) {
2731 assert(LI.getPointerOperand() == *U);
2732 if (!LI.isSimple() || LI.getType()->isSingleValueType())
2735 // We have an aggregate being loaded, split it apart.
2736 DEBUG(dbgs() << " original: " << LI << "\n");
2737 LoadOpSplitter Splitter(&LI, *U);
2738 Value *V = UndefValue::get(LI.getType());
2739 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca");
2740 LI.replaceAllUsesWith(V);
2741 LI.eraseFromParent();
2745 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> {
2746 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr)
2747 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {}
2749 /// Emit a leaf store of a single value. This is called at the leaves of the
2750 /// recursive emission to actually produce stores.
2751 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) {
2752 assert(Ty->isSingleValueType());
2753 // Extract the single value and store it using the indices.
2754 Value *Store = IRB.CreateStore(
2755 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"),
2756 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"));
2758 DEBUG(dbgs() << " to: " << *Store << "\n");
2762 bool visitStoreInst(StoreInst &SI) {
2763 if (!SI.isSimple() || SI.getPointerOperand() != *U)
2765 Value *V = SI.getValueOperand();
2766 if (V->getType()->isSingleValueType())
2769 // We have an aggregate being stored, split it apart.
2770 DEBUG(dbgs() << " original: " << SI << "\n");
2771 StoreOpSplitter Splitter(&SI, *U);
2772 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca");
2773 SI.eraseFromParent();
2777 bool visitBitCastInst(BitCastInst &BC) {
2782 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
2787 bool visitPHINode(PHINode &PN) {
2792 bool visitSelectInst(SelectInst &SI) {
2799 /// \brief Try to find a partition of the aggregate type passed in for a given
2800 /// offset and size.
2802 /// This recurses through the aggregate type and tries to compute a subtype
2803 /// based on the offset and size. When the offset and size span a sub-section
2804 /// of an array, it will even compute a new array type for that sub-section,
2805 /// and the same for structs.
2807 /// Note that this routine is very strict and tries to find a partition of the
2808 /// type which produces the *exact* right offset and size. It is not forgiving
2809 /// when the size or offset cause either end of type-based partition to be off.
2810 /// Also, this is a best-effort routine. It is reasonable to give up and not
2811 /// return a type if necessary.
2812 static Type *getTypePartition(const TargetData &TD, Type *Ty,
2813 uint64_t Offset, uint64_t Size) {
2814 if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size)
2817 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
2818 // We can't partition pointers...
2819 if (SeqTy->isPointerTy())
2822 Type *ElementTy = SeqTy->getElementType();
2823 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
2824 uint64_t NumSkippedElements = Offset / ElementSize;
2825 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy))
2826 if (NumSkippedElements >= ArrTy->getNumElements())
2828 if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy))
2829 if (NumSkippedElements >= VecTy->getNumElements())
2831 Offset -= NumSkippedElements * ElementSize;
2833 // First check if we need to recurse.
2834 if (Offset > 0 || Size < ElementSize) {
2835 // Bail if the partition ends in a different array element.
2836 if ((Offset + Size) > ElementSize)
2838 // Recurse through the element type trying to peel off offset bytes.
2839 return getTypePartition(TD, ElementTy, Offset, Size);
2841 assert(Offset == 0);
2843 if (Size == ElementSize)
2845 assert(Size > ElementSize);
2846 uint64_t NumElements = Size / ElementSize;
2847 if (NumElements * ElementSize != Size)
2849 return ArrayType::get(ElementTy, NumElements);
2852 StructType *STy = dyn_cast<StructType>(Ty);
2856 const StructLayout *SL = TD.getStructLayout(STy);
2857 if (Offset >= SL->getSizeInBytes())
2859 uint64_t EndOffset = Offset + Size;
2860 if (EndOffset > SL->getSizeInBytes())
2863 unsigned Index = SL->getElementContainingOffset(Offset);
2864 Offset -= SL->getElementOffset(Index);
2866 Type *ElementTy = STy->getElementType(Index);
2867 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
2868 if (Offset >= ElementSize)
2869 return 0; // The offset points into alignment padding.
2871 // See if any partition must be contained by the element.
2872 if (Offset > 0 || Size < ElementSize) {
2873 if ((Offset + Size) > ElementSize)
2875 return getTypePartition(TD, ElementTy, Offset, Size);
2877 assert(Offset == 0);
2879 if (Size == ElementSize)
2882 StructType::element_iterator EI = STy->element_begin() + Index,
2883 EE = STy->element_end();
2884 if (EndOffset < SL->getSizeInBytes()) {
2885 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
2886 if (Index == EndIndex)
2887 return 0; // Within a single element and its padding.
2889 // Don't try to form "natural" types if the elements don't line up with the
2891 // FIXME: We could potentially recurse down through the last element in the
2892 // sub-struct to find a natural end point.
2893 if (SL->getElementOffset(EndIndex) != EndOffset)
2896 assert(Index < EndIndex);
2897 EE = STy->element_begin() + EndIndex;
2900 // Try to build up a sub-structure.
2901 SmallVector<Type *, 4> ElementTys;
2903 ElementTys.push_back(*EI++);
2905 StructType *SubTy = StructType::get(STy->getContext(), ElementTys,
2907 const StructLayout *SubSL = TD.getStructLayout(SubTy);
2908 if (Size != SubSL->getSizeInBytes())
2909 return 0; // The sub-struct doesn't have quite the size needed.
2914 /// \brief Rewrite an alloca partition's users.
2916 /// This routine drives both of the rewriting goals of the SROA pass. It tries
2917 /// to rewrite uses of an alloca partition to be conducive for SSA value
2918 /// promotion. If the partition needs a new, more refined alloca, this will
2919 /// build that new alloca, preserving as much type information as possible, and
2920 /// rewrite the uses of the old alloca to point at the new one and have the
2921 /// appropriate new offsets. It also evaluates how successful the rewrite was
2922 /// at enabling promotion and if it was successful queues the alloca to be
2924 bool SROA::rewriteAllocaPartition(AllocaInst &AI,
2925 AllocaPartitioning &P,
2926 AllocaPartitioning::iterator PI) {
2927 uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset;
2928 if (P.use_begin(PI) == P.use_end(PI))
2929 return false; // No live uses left of this partition.
2931 // Try to compute a friendly type for this partition of the alloca. This
2932 // won't always succeed, in which case we fall back to a legal integer type
2933 // or an i8 array of an appropriate size.
2935 if (Type *PartitionTy = P.getCommonType(PI))
2936 if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize)
2937 AllocaTy = PartitionTy;
2939 if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(),
2940 PI->BeginOffset, AllocaSize))
2941 AllocaTy = PartitionTy;
2943 (AllocaTy->isArrayTy() &&
2944 AllocaTy->getArrayElementType()->isIntegerTy())) &&
2945 TD->isLegalInteger(AllocaSize * 8))
2946 AllocaTy = Type::getIntNTy(*C, AllocaSize * 8);
2948 AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize);
2949 assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize);
2951 // Check for the case where we're going to rewrite to a new alloca of the
2952 // exact same type as the original, and with the same access offsets. In that
2953 // case, re-use the existing alloca, but still run through the rewriter to
2954 // performe phi and select speculation.
2956 if (AllocaTy == AI.getAllocatedType()) {
2957 assert(PI->BeginOffset == 0 &&
2958 "Non-zero begin offset but same alloca type");
2959 assert(PI == P.begin() && "Begin offset is zero on later partition");
2962 // FIXME: The alignment here is overly conservative -- we could in many
2963 // cases get away with much weaker alignment constraints.
2964 NewAI = new AllocaInst(AllocaTy, 0, AI.getAlignment(),
2965 AI.getName() + ".sroa." + Twine(PI - P.begin()),
2970 DEBUG(dbgs() << "Rewriting alloca partition "
2971 << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: "
2974 AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI,
2975 PI->BeginOffset, PI->EndOffset);
2976 DEBUG(dbgs() << " rewriting ");
2977 DEBUG(P.print(dbgs(), PI, ""));
2978 if (Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI))) {
2979 DEBUG(dbgs() << " and queuing for promotion\n");
2980 PromotableAllocas.push_back(NewAI);
2981 } else if (NewAI != &AI) {
2982 // If we can't promote the alloca, iterate on it to check for new
2983 // refinements exposed by splitting the current alloca. Don't iterate on an
2984 // alloca which didn't actually change and didn't get promoted.
2985 Worklist.insert(NewAI);
2990 /// \brief Walks the partitioning of an alloca rewriting uses of each partition.
2991 bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
2992 bool Changed = false;
2993 for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE;
2995 Changed |= rewriteAllocaPartition(AI, P, PI);
3000 /// \brief Analyze an alloca for SROA.
3002 /// This analyzes the alloca to ensure we can reason about it, builds
3003 /// a partitioning of the alloca, and then hands it off to be split and
3004 /// rewritten as needed.
3005 bool SROA::runOnAlloca(AllocaInst &AI) {
3006 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
3007 ++NumAllocasAnalyzed;
3009 // Special case dead allocas, as they're trivial.
3010 if (AI.use_empty()) {
3011 AI.eraseFromParent();
3015 // Skip alloca forms that this analysis can't handle.
3016 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
3017 TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
3020 // First check if this is a non-aggregate type that we should simply promote.
3021 if (!AI.getAllocatedType()->isAggregateType() && isAllocaPromotable(&AI)) {
3022 DEBUG(dbgs() << " Trivially scalar type, queuing for promotion...\n");
3023 PromotableAllocas.push_back(&AI);
3027 bool Changed = false;
3029 // First, split any FCA loads and stores touching this alloca to promote
3030 // better splitting and promotion opportunities.
3031 AggLoadStoreRewriter AggRewriter(*TD);
3032 Changed |= AggRewriter.rewrite(AI);
3034 // Build the partition set using a recursive instruction-visiting builder.
3035 AllocaPartitioning P(*TD, AI);
3036 DEBUG(P.print(dbgs()));
3040 // No partitions to split. Leave the dead alloca for a later pass to clean up.
3041 if (P.begin() == P.end())
3044 // Delete all the dead users of this alloca before splitting and rewriting it.
3045 for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
3046 DE = P.dead_user_end();
3049 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
3050 DeadInsts.push_back(*DI);
3052 for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(),
3053 DE = P.dead_op_end();
3056 // Clobber the use with an undef value.
3057 **DO = UndefValue::get(OldV->getType());
3058 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
3059 if (isInstructionTriviallyDead(OldI)) {
3061 DeadInsts.push_back(OldI);
3065 return splitAlloca(AI, P) || Changed;
3068 /// \brief Delete the dead instructions accumulated in this run.
3070 /// Recursively deletes the dead instructions we've accumulated. This is done
3071 /// at the very end to maximize locality of the recursive delete and to
3072 /// minimize the problems of invalidated instruction pointers as such pointers
3073 /// are used heavily in the intermediate stages of the algorithm.
3075 /// We also record the alloca instructions deleted here so that they aren't
3076 /// subsequently handed to mem2reg to promote.
3077 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
3078 DeadSplitInsts.clear();
3079 while (!DeadInsts.empty()) {
3080 Instruction *I = DeadInsts.pop_back_val();
3081 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
3083 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
3084 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
3085 // Zero out the operand and see if it becomes trivially dead.
3087 if (isInstructionTriviallyDead(U))
3088 DeadInsts.push_back(U);
3091 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3092 DeletedAllocas.insert(AI);
3095 I->eraseFromParent();
3099 /// \brief Promote the allocas, using the best available technique.
3101 /// This attempts to promote whatever allocas have been identified as viable in
3102 /// the PromotableAllocas list. If that list is empty, there is nothing to do.
3103 /// If there is a domtree available, we attempt to promote using the full power
3104 /// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is
3105 /// based on the SSAUpdater utilities. This function returns whether any
3106 /// promotion occured.
3107 bool SROA::promoteAllocas(Function &F) {
3108 if (PromotableAllocas.empty())
3111 // Ensure that the list is unique.
3112 std::sort(PromotableAllocas.begin(), PromotableAllocas.end());
3113 PromotableAllocas.erase(std::unique(PromotableAllocas.begin(),
3114 PromotableAllocas.end()),
3115 PromotableAllocas.end());
3117 NumPromoted += PromotableAllocas.size();
3119 if (DT && !ForceSSAUpdater) {
3120 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
3121 PromoteMemToReg(PromotableAllocas, *DT);
3122 PromotableAllocas.clear();
3126 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n");
3128 DIBuilder DIB(*F.getParent());
3129 SmallVector<Instruction*, 64> Insts;
3131 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) {
3132 AllocaInst *AI = PromotableAllocas[Idx];
3133 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
3135 Instruction *I = cast<Instruction>(*UI++);
3136 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about
3137 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs
3138 // leading to them) here. Eventually it should use them to optimize the
3139 // scalar values produced.
3140 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) {
3141 assert(onlyUsedByLifetimeMarkers(I) &&
3142 "Found a bitcast used outside of a lifetime marker.");
3143 while (!I->use_empty())
3144 cast<Instruction>(*I->use_begin())->eraseFromParent();
3145 I->eraseFromParent();
3148 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
3149 assert(II->getIntrinsicID() == Intrinsic::lifetime_start ||
3150 II->getIntrinsicID() == Intrinsic::lifetime_end);
3151 II->eraseFromParent();
3157 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts);
3161 PromotableAllocas.clear();
3166 /// \brief A predicate to test whether an alloca belongs to a set.
3167 class IsAllocaInSet {
3168 typedef SmallPtrSet<AllocaInst *, 4> SetType;
3172 IsAllocaInSet(const SetType &Set) : Set(Set) {}
3173 bool operator()(AllocaInst *AI) { return Set.count(AI); }
3177 bool SROA::runOnFunction(Function &F) {
3178 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
3179 C = &F.getContext();
3180 TD = getAnalysisIfAvailable<TargetData>();
3182 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
3185 DT = getAnalysisIfAvailable<DominatorTree>();
3187 BasicBlock &EntryBB = F.getEntryBlock();
3188 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
3190 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
3191 Worklist.insert(AI);
3193 bool Changed = false;
3194 // A set of deleted alloca instruction pointers which should be removed from
3195 // the list of promotable allocas.
3196 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
3198 while (!Worklist.empty()) {
3199 Changed |= runOnAlloca(*Worklist.pop_back_val());
3200 deleteDeadInstructions(DeletedAllocas);
3201 if (!DeletedAllocas.empty()) {
3202 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
3203 PromotableAllocas.end(),
3204 IsAllocaInSet(DeletedAllocas)),
3205 PromotableAllocas.end());
3206 DeletedAllocas.clear();
3210 Changed |= promoteAllocas(F);
3215 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
3216 if (RequiresDomTree)
3217 AU.addRequired<DominatorTree>();
3218 AU.setPreservesCFG();