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/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/GetElementPtrTypeIterator.h"
53 #include "llvm/Support/InstVisitor.h"
54 #include "llvm/Support/MathExtras.h"
55 #include "llvm/Support/ValueHandle.h"
56 #include "llvm/Support/raw_ostream.h"
57 #include "llvm/Target/TargetData.h"
58 #include "llvm/Transforms/Utils/Local.h"
59 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
60 #include "llvm/Transforms/Utils/SSAUpdater.h"
63 STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement");
64 STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced");
65 STATISTIC(NumPromoted, "Number of allocas promoted to SSA values");
66 STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion");
67 STATISTIC(NumDeleted, "Number of instructions deleted");
68 STATISTIC(NumVectorized, "Number of vectorized aggregates");
71 /// \brief Alloca partitioning representation.
73 /// This class represents a partitioning of an alloca into slices, and
74 /// information about the nature of uses of each slice of the alloca. The goal
75 /// is that this information is sufficient to decide if and how to split the
76 /// alloca apart and replace slices with scalars. It is also intended that this
77 /// structure can capture the relevant information needed both to decide about
78 /// and to enact these transformations.
79 class AllocaPartitioning {
81 /// \brief A common base class for representing a half-open byte range.
83 /// \brief The beginning offset of the range.
86 /// \brief The ending offset, not included in the range.
89 ByteRange() : BeginOffset(), EndOffset() {}
90 ByteRange(uint64_t BeginOffset, uint64_t EndOffset)
91 : BeginOffset(BeginOffset), EndOffset(EndOffset) {}
93 /// \brief Support for ordering ranges.
95 /// This provides an ordering over ranges such that start offsets are
96 /// always increasing, and within equal start offsets, the end offsets are
97 /// decreasing. Thus the spanning range comes first in a cluster with the
98 /// same start position.
99 bool operator<(const ByteRange &RHS) const {
100 if (BeginOffset < RHS.BeginOffset) return true;
101 if (BeginOffset > RHS.BeginOffset) return false;
102 if (EndOffset > RHS.EndOffset) return true;
106 /// \brief Support comparison with a single offset to allow binary searches.
107 bool operator<(uint64_t RHSOffset) const {
108 return BeginOffset < RHSOffset;
111 bool operator==(const ByteRange &RHS) const {
112 return BeginOffset == RHS.BeginOffset && EndOffset == RHS.EndOffset;
114 bool operator!=(const ByteRange &RHS) const { return !operator==(RHS); }
117 /// \brief A partition of an alloca.
119 /// This structure represents a contiguous partition of the alloca. These are
120 /// formed by examining the uses of the alloca. During formation, they may
121 /// overlap but once an AllocaPartitioning is built, the Partitions within it
122 /// are all disjoint.
123 struct Partition : public ByteRange {
124 /// \brief Whether this partition is splittable into smaller partitions.
126 /// We flag partitions as splittable when they are formed entirely due to
127 /// accesses by trivially splittable operations such as memset and memcpy.
129 /// FIXME: At some point we should consider loads and stores of FCAs to be
130 /// splittable and eagerly split them into scalar values.
133 Partition() : ByteRange(), IsSplittable() {}
134 Partition(uint64_t BeginOffset, uint64_t EndOffset, bool IsSplittable)
135 : ByteRange(BeginOffset, EndOffset), IsSplittable(IsSplittable) {}
138 /// \brief A particular use of a partition of the alloca.
140 /// This structure is used to associate uses of a partition with it. They
141 /// mark the range of bytes which are referenced by a particular instruction,
142 /// and includes a handle to the user itself and the pointer value in use.
143 /// The bounds of these uses are determined by intersecting the bounds of the
144 /// memory use itself with a particular partition. As a consequence there is
145 /// intentionally overlap between various uses of the same partition.
146 struct PartitionUse : public ByteRange {
147 /// \brief The user of this range of the alloca.
148 AssertingVH<Instruction> User;
150 /// \brief The particular pointer value derived from this alloca in use.
151 AssertingVH<Instruction> Ptr;
153 PartitionUse() : ByteRange(), User(), Ptr() {}
154 PartitionUse(uint64_t BeginOffset, uint64_t EndOffset,
155 Instruction *User, Instruction *Ptr)
156 : ByteRange(BeginOffset, EndOffset), User(User), Ptr(Ptr) {}
159 /// \brief Construct a partitioning of a particular alloca.
161 /// Construction does most of the work for partitioning the alloca. This
162 /// performs the necessary walks of users and builds a partitioning from it.
163 AllocaPartitioning(const TargetData &TD, AllocaInst &AI);
165 /// \brief Test whether a pointer to the allocation escapes our analysis.
167 /// If this is true, the partitioning is never fully built and should be
169 bool isEscaped() const { return PointerEscapingInstr; }
171 /// \brief Support for iterating over the partitions.
173 typedef SmallVectorImpl<Partition>::iterator iterator;
174 iterator begin() { return Partitions.begin(); }
175 iterator end() { return Partitions.end(); }
177 typedef SmallVectorImpl<Partition>::const_iterator const_iterator;
178 const_iterator begin() const { return Partitions.begin(); }
179 const_iterator end() const { return Partitions.end(); }
182 /// \brief Support for iterating over and manipulating a particular
183 /// partition's uses.
185 /// The iteration support provided for uses is more limited, but also
186 /// includes some manipulation routines to support rewriting the uses of
187 /// partitions during SROA.
189 typedef SmallVectorImpl<PartitionUse>::iterator use_iterator;
190 use_iterator use_begin(unsigned Idx) { return Uses[Idx].begin(); }
191 use_iterator use_begin(const_iterator I) { return Uses[I - begin()].begin(); }
192 use_iterator use_end(unsigned Idx) { return Uses[Idx].end(); }
193 use_iterator use_end(const_iterator I) { return Uses[I - begin()].end(); }
194 void use_insert(unsigned Idx, use_iterator UI, const PartitionUse &U) {
195 Uses[Idx].insert(UI, U);
197 void use_insert(const_iterator I, use_iterator UI, const PartitionUse &U) {
198 Uses[I - begin()].insert(UI, U);
200 void use_erase(unsigned Idx, use_iterator UI) { Uses[Idx].erase(UI); }
201 void use_erase(const_iterator I, use_iterator UI) {
202 Uses[I - begin()].erase(UI);
205 typedef SmallVectorImpl<PartitionUse>::const_iterator const_use_iterator;
206 const_use_iterator use_begin(unsigned Idx) const { return Uses[Idx].begin(); }
207 const_use_iterator use_begin(const_iterator I) const {
208 return Uses[I - begin()].begin();
210 const_use_iterator use_end(unsigned Idx) const { return Uses[Idx].end(); }
211 const_use_iterator use_end(const_iterator I) const {
212 return Uses[I - begin()].end();
216 /// \brief Allow iterating the dead users for this alloca.
218 /// These are instructions which will never actually use the alloca as they
219 /// are outside the allocated range. They are safe to replace with undef and
222 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator;
223 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); }
224 dead_user_iterator dead_user_end() const { return DeadUsers.end(); }
227 /// \brief Allow iterating the dead expressions referring to this alloca.
229 /// These are operands which have cannot actually be used to refer to the
230 /// alloca as they are outside its range and the user doesn't correct for
231 /// that. These mostly consist of PHI node inputs and the like which we just
232 /// need to replace with undef.
234 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator;
235 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); }
236 dead_op_iterator dead_op_end() const { return DeadOperands.end(); }
239 /// \brief MemTransferInst auxiliary data.
240 /// This struct provides some auxiliary data about memory transfer
241 /// intrinsics such as memcpy and memmove. These intrinsics can use two
242 /// different ranges within the same alloca, and provide other challenges to
243 /// correctly represent. We stash extra data to help us untangle this
244 /// after the partitioning is complete.
245 struct MemTransferOffsets {
246 uint64_t DestBegin, DestEnd;
247 uint64_t SourceBegin, SourceEnd;
250 MemTransferOffsets getMemTransferOffsets(MemTransferInst &II) const {
251 return MemTransferInstData.lookup(&II);
254 /// \brief Map from a PHI or select operand back to a partition.
256 /// When manipulating PHI nodes or selects, they can use more than one
257 /// partition of an alloca. We store a special mapping to allow finding the
258 /// partition referenced by each of these operands, if any.
259 iterator findPartitionForPHIOrSelectOperand(Instruction &I, Value *Op) {
260 SmallDenseMap<std::pair<Instruction *, Value *>,
261 std::pair<unsigned, unsigned> >::const_iterator MapIt
262 = PHIOrSelectOpMap.find(std::make_pair(&I, Op));
263 if (MapIt == PHIOrSelectOpMap.end())
266 return begin() + MapIt->second.first;
269 /// \brief Map from a PHI or select operand back to the specific use of
272 /// Similar to mapping these operands back to the partitions, this maps
273 /// directly to the use structure of that partition.
274 use_iterator findPartitionUseForPHIOrSelectOperand(Instruction &I,
276 SmallDenseMap<std::pair<Instruction *, Value *>,
277 std::pair<unsigned, unsigned> >::const_iterator MapIt
278 = PHIOrSelectOpMap.find(std::make_pair(&I, Op));
279 assert(MapIt != PHIOrSelectOpMap.end());
280 return Uses[MapIt->second.first].begin() + MapIt->second.second;
283 /// \brief Compute a common type among the uses of a particular partition.
285 /// This routines walks all of the uses of a particular partition and tries
286 /// to find a common type between them. Untyped operations such as memset and
287 /// memcpy are ignored.
288 Type *getCommonType(iterator I) const;
290 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
291 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const;
292 void printUsers(raw_ostream &OS, const_iterator I,
293 StringRef Indent = " ") const;
294 void print(raw_ostream &OS) const;
295 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const;
296 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const;
300 template <typename DerivedT, typename RetT = void> class BuilderBase;
301 class PartitionBuilder;
302 friend class AllocaPartitioning::PartitionBuilder;
304 friend class AllocaPartitioning::UseBuilder;
306 /// \brief Handle to alloca instruction to simplify method interfaces.
309 /// \brief The instruction responsible for this alloca having no partitioning.
311 /// When an instruction (potentially) escapes the pointer to the alloca, we
312 /// store a pointer to that here and abort trying to partition the alloca.
313 /// This will be null if the alloca is partitioned successfully.
314 Instruction *PointerEscapingInstr;
316 /// \brief The partitions of the alloca.
318 /// We store a vector of the partitions over the alloca here. This vector is
319 /// sorted by increasing begin offset, and then by decreasing end offset. See
320 /// the Partition inner class for more details. Initially (during
321 /// construction) there are overlaps, but we form a disjoint sequence of
322 /// partitions while finishing construction and a fully constructed object is
323 /// expected to always have this as a disjoint space.
324 SmallVector<Partition, 8> Partitions;
326 /// \brief The uses of the partitions.
328 /// This is essentially a mapping from each partition to a list of uses of
329 /// that partition. The mapping is done with a Uses vector that has the exact
330 /// same number of entries as the partition vector. Each entry is itself
331 /// a vector of the uses.
332 SmallVector<SmallVector<PartitionUse, 2>, 8> Uses;
334 /// \brief Instructions which will become dead if we rewrite the alloca.
336 /// Note that these are not separated by partition. This is because we expect
337 /// a partitioned alloca to be completely rewritten or not rewritten at all.
338 /// If rewritten, all these instructions can simply be removed and replaced
339 /// with undef as they come from outside of the allocated space.
340 SmallVector<Instruction *, 8> DeadUsers;
342 /// \brief Operands which will become dead if we rewrite the alloca.
344 /// These are operands that in their particular use can be replaced with
345 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs
346 /// to PHI nodes and the like. They aren't entirely dead (there might be
347 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we
348 /// want to swap this particular input for undef to simplify the use lists of
350 SmallVector<Use *, 8> DeadOperands;
352 /// \brief The underlying storage for auxiliary memcpy and memset info.
353 SmallDenseMap<MemTransferInst *, MemTransferOffsets, 4> MemTransferInstData;
355 /// \brief A side datastructure used when building up the partitions and uses.
357 /// This mapping is only really used during the initial building of the
358 /// partitioning so that we can retain information about PHI and select nodes
360 SmallDenseMap<Instruction *, std::pair<uint64_t, bool> > PHIOrSelectSizes;
362 /// \brief Auxiliary information for particular PHI or select operands.
363 SmallDenseMap<std::pair<Instruction *, Value *>,
364 std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap;
366 /// \brief A utility routine called from the constructor.
368 /// This does what it says on the tin. It is the key of the alloca partition
369 /// splitting and merging. After it is called we have the desired disjoint
370 /// collection of partitions.
371 void splitAndMergePartitions();
375 template <typename DerivedT, typename RetT>
376 class AllocaPartitioning::BuilderBase
377 : public InstVisitor<DerivedT, RetT> {
379 BuilderBase(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
381 AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())),
387 const TargetData &TD;
388 const uint64_t AllocSize;
389 AllocaPartitioning &P;
395 SmallVector<OffsetUse, 8> Queue;
397 // The active offset and use while visiting.
401 void enqueueUsers(Instruction &I, uint64_t UserOffset) {
402 SmallPtrSet<User *, 8> UserSet;
403 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
405 if (!UserSet.insert(*UI))
408 OffsetUse OU = { &UI.getUse(), UserOffset };
413 bool computeConstantGEPOffset(GetElementPtrInst &GEPI, uint64_t &GEPOffset) {
415 for (gep_type_iterator GTI = gep_type_begin(GEPI), GTE = gep_type_end(GEPI);
417 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
423 // Handle a struct index, which adds its field offset to the pointer.
424 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
425 unsigned ElementIdx = OpC->getZExtValue();
426 const StructLayout *SL = TD.getStructLayout(STy);
427 GEPOffset += SL->getElementOffset(ElementIdx);
432 += OpC->getZExtValue() * TD.getTypeAllocSize(GTI.getIndexedType());
437 Value *foldSelectInst(SelectInst &SI) {
438 // If the condition being selected on is a constant or the same value is
439 // being selected between, fold the select. Yes this does (rarely) happen
441 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition()))
442 return SI.getOperand(1+CI->isZero());
443 if (SI.getOperand(1) == SI.getOperand(2)) {
444 assert(*U == SI.getOperand(1));
445 return SI.getOperand(1);
451 /// \brief Builder for the alloca partitioning.
453 /// This class builds an alloca partitioning by recursively visiting the uses
454 /// of an alloca and splitting the partitions for each load and store at each
456 class AllocaPartitioning::PartitionBuilder
457 : public BuilderBase<PartitionBuilder, bool> {
458 friend class InstVisitor<PartitionBuilder, bool>;
460 SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap;
463 PartitionBuilder(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
464 : BuilderBase<PartitionBuilder, bool>(TD, AI, P) {}
466 /// \brief Run the builder over the allocation.
468 // Note that we have to re-evaluate size on each trip through the loop as
469 // the queue grows at the tail.
470 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
472 Offset = Queue[Idx].Offset;
473 if (!visit(cast<Instruction>(U->getUser())))
480 bool markAsEscaping(Instruction &I) {
481 P.PointerEscapingInstr = &I;
485 void insertUse(Instruction &I, uint64_t Size, bool IsSplittable = false) {
486 uint64_t BeginOffset = Offset, EndOffset = Offset + Size;
488 // Completely skip uses which start outside of the allocation.
489 if (BeginOffset >= AllocSize) {
490 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset
491 << " which starts past the end of the " << AllocSize
493 << " alloca: " << P.AI << "\n"
494 << " use: " << I << "\n");
498 // Clamp the size to the allocation.
499 if (EndOffset > AllocSize) {
500 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset
501 << " to remain within the " << AllocSize << " byte alloca:\n"
502 << " alloca: " << P.AI << "\n"
503 << " use: " << I << "\n");
504 EndOffset = AllocSize;
507 // See if we can just add a user onto the last slot currently occupied.
508 if (!P.Partitions.empty() &&
509 P.Partitions.back().BeginOffset == BeginOffset &&
510 P.Partitions.back().EndOffset == EndOffset) {
511 P.Partitions.back().IsSplittable &= IsSplittable;
515 Partition New(BeginOffset, EndOffset, IsSplittable);
516 P.Partitions.push_back(New);
519 bool handleLoadOrStore(Type *Ty, Instruction &I) {
520 uint64_t Size = TD.getTypeStoreSize(Ty);
522 // If this memory access can be shown to *statically* extend outside the
523 // bounds of of the allocation, it's behavior is undefined, so simply
524 // ignore it. Note that this is more strict than the generic clamping
525 // behavior of insertUse. We also try to handle cases which might run the
527 // FIXME: We should instead consider the pointer to have escaped if this
528 // function is being instrumented for addressing bugs or race conditions.
529 if (Offset >= AllocSize || Size > AllocSize || Offset + Size > AllocSize) {
530 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte "
531 << (isa<LoadInst>(I) ? "load" : "store") << " @" << Offset
532 << " which extends past the end of the " << AllocSize
534 << " alloca: " << P.AI << "\n"
535 << " use: " << I << "\n");
543 bool visitBitCastInst(BitCastInst &BC) {
544 enqueueUsers(BC, Offset);
548 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) {
550 if (!computeConstantGEPOffset(GEPI, GEPOffset))
551 return markAsEscaping(GEPI);
553 enqueueUsers(GEPI, GEPOffset);
557 bool visitLoadInst(LoadInst &LI) {
558 return handleLoadOrStore(LI.getType(), LI);
561 bool visitStoreInst(StoreInst &SI) {
562 if (SI.getOperand(0) == *U)
563 return markAsEscaping(SI);
565 return handleLoadOrStore(SI.getOperand(0)->getType(), SI);
569 bool visitMemSetInst(MemSetInst &II) {
570 assert(II.getRawDest() == *U && "Pointer use is not the destination?");
571 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
572 insertUse(II, Length ? Length->getZExtValue() : AllocSize - Offset, Length);
576 bool visitMemTransferInst(MemTransferInst &II) {
577 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
578 uint64_t Size = Length ? Length->getZExtValue() : AllocSize - Offset;
580 // Zero-length mem transfer intrinsics can be ignored entirely.
583 MemTransferOffsets &Offsets = P.MemTransferInstData[&II];
585 // Only intrinsics with a constant length can be split.
586 Offsets.IsSplittable = Length;
588 if (*U != II.getRawDest()) {
589 assert(*U == II.getRawSource());
590 Offsets.SourceBegin = Offset;
591 Offsets.SourceEnd = Offset + Size;
593 Offsets.DestBegin = Offset;
594 Offsets.DestEnd = Offset + Size;
597 insertUse(II, Size, Offsets.IsSplittable);
598 unsigned NewIdx = P.Partitions.size() - 1;
600 SmallDenseMap<Instruction *, unsigned>::const_iterator PMI;
601 bool Inserted = false;
602 llvm::tie(PMI, Inserted)
603 = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx));
604 if (!Inserted && Offsets.IsSplittable) {
605 // We've found a memory transfer intrinsic which refers to the alloca as
606 // both a source and dest. We refuse to split these to simplify splitting
607 // logic. If possible, SROA will still split them into separate allocas
608 // and then re-analyze.
609 Offsets.IsSplittable = false;
610 P.Partitions[PMI->second].IsSplittable = false;
611 P.Partitions[NewIdx].IsSplittable = false;
617 // Disable SRoA for any intrinsics except for lifetime invariants.
618 // FIXME: What about debug instrinsics? This matches old behavior, but
619 // doesn't make sense.
620 bool visitIntrinsicInst(IntrinsicInst &II) {
621 if (II.getIntrinsicID() == Intrinsic::lifetime_start ||
622 II.getIntrinsicID() == Intrinsic::lifetime_end) {
623 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
624 uint64_t Size = std::min(AllocSize - Offset, Length->getLimitedValue());
625 insertUse(II, Size, true);
629 return markAsEscaping(II);
632 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) {
633 // We consider any PHI or select that results in a direct load or store of
634 // the same offset to be a viable use for partitioning purposes. These uses
635 // are considered unsplittable and the size is the maximum loaded or stored
637 SmallPtrSet<Instruction *, 4> Visited;
638 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses;
639 Visited.insert(Root);
640 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root));
642 Instruction *I, *UsedI;
643 llvm::tie(UsedI, I) = Uses.pop_back_val();
645 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
646 Size = std::max(Size, TD.getTypeStoreSize(LI->getType()));
649 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
650 Value *Op = SI->getOperand(0);
653 Size = std::max(Size, TD.getTypeStoreSize(Op->getType()));
657 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
658 if (!GEP->hasAllZeroIndices())
660 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) &&
661 !isa<SelectInst>(I)) {
665 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE;
667 if (Visited.insert(cast<Instruction>(*UI)))
668 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI)));
669 } while (!Uses.empty());
674 bool visitPHINode(PHINode &PN) {
675 // See if we already have computed info on this node.
676 std::pair<uint64_t, bool> &PHIInfo = P.PHIOrSelectSizes[&PN];
678 PHIInfo.second = true;
679 insertUse(PN, PHIInfo.first);
683 // Check for an unsafe use of the PHI node.
684 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&PN, PHIInfo.first))
685 return markAsEscaping(*EscapingI);
687 insertUse(PN, PHIInfo.first);
691 bool visitSelectInst(SelectInst &SI) {
692 if (Value *Result = foldSelectInst(SI)) {
694 // If the result of the constant fold will be the pointer, recurse
695 // through the select as if we had RAUW'ed it.
696 enqueueUsers(SI, Offset);
701 // See if we already have computed info on this node.
702 std::pair<uint64_t, bool> &SelectInfo = P.PHIOrSelectSizes[&SI];
703 if (SelectInfo.first) {
704 SelectInfo.second = true;
705 insertUse(SI, SelectInfo.first);
709 // Check for an unsafe use of the PHI node.
710 if (Instruction *EscapingI = hasUnsafePHIOrSelectUse(&SI, SelectInfo.first))
711 return markAsEscaping(*EscapingI);
713 insertUse(SI, SelectInfo.first);
717 /// \brief Disable SROA entirely if there are unhandled users of the alloca.
718 bool visitInstruction(Instruction &I) { return markAsEscaping(I); }
722 /// \brief Use adder for the alloca partitioning.
724 /// This class adds the uses of an alloca to all of the partitions which they
725 /// use. For splittable partitions, this can end up doing essentially a linear
726 /// walk of the partitions, but the number of steps remains bounded by the
727 /// total result instruction size:
728 /// - The number of partitions is a result of the number unsplittable
729 /// instructions using the alloca.
730 /// - The number of users of each partition is at worst the total number of
731 /// splittable instructions using the alloca.
732 /// Thus we will produce N * M instructions in the end, where N are the number
733 /// of unsplittable uses and M are the number of splittable. This visitor does
734 /// the exact same number of updates to the partitioning.
736 /// In the more common case, this visitor will leverage the fact that the
737 /// partition space is pre-sorted, and do a logarithmic search for the
738 /// partition needed, making the total visit a classical ((N + M) * log(N))
739 /// complexity operation.
740 class AllocaPartitioning::UseBuilder : public BuilderBase<UseBuilder> {
741 friend class InstVisitor<UseBuilder>;
743 /// \brief Set to de-duplicate dead instructions found in the use walk.
744 SmallPtrSet<Instruction *, 4> VisitedDeadInsts;
747 UseBuilder(const TargetData &TD, AllocaInst &AI, AllocaPartitioning &P)
748 : BuilderBase<UseBuilder>(TD, AI, P) {}
750 /// \brief Run the builder over the allocation.
752 // Note that we have to re-evaluate size on each trip through the loop as
753 // the queue grows at the tail.
754 for (unsigned Idx = 0; Idx < Queue.size(); ++Idx) {
756 Offset = Queue[Idx].Offset;
757 this->visit(cast<Instruction>(U->getUser()));
762 void markAsDead(Instruction &I) {
763 if (VisitedDeadInsts.insert(&I))
764 P.DeadUsers.push_back(&I);
767 void insertUse(uint64_t Size, Instruction &User) {
768 uint64_t BeginOffset = Offset, EndOffset = Offset + Size;
770 // If the use extends outside of the allocation, record it as a dead use
771 // for elimination later.
772 if (BeginOffset >= AllocSize || Size == 0)
773 return markAsDead(User);
775 // Bound the use by the size of the allocation.
776 if (EndOffset > AllocSize)
777 EndOffset = AllocSize;
779 // NB: This only works if we have zero overlapping partitions.
780 iterator B = std::lower_bound(P.begin(), P.end(), BeginOffset);
781 if (B != P.begin() && llvm::prior(B)->EndOffset > BeginOffset)
783 for (iterator I = B, E = P.end(); I != E && I->BeginOffset < EndOffset;
785 PartitionUse NewUse(std::max(I->BeginOffset, BeginOffset),
786 std::min(I->EndOffset, EndOffset),
787 &User, cast<Instruction>(*U));
788 P.Uses[I - P.begin()].push_back(NewUse);
789 if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser()))
790 P.PHIOrSelectOpMap[std::make_pair(&User, U->get())]
791 = std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1);
795 void handleLoadOrStore(Type *Ty, Instruction &I) {
796 uint64_t Size = TD.getTypeStoreSize(Ty);
798 // If this memory access can be shown to *statically* extend outside the
799 // bounds of of the allocation, it's behavior is undefined, so simply
800 // ignore it. Note that this is more strict than the generic clamping
801 // behavior of insertUse.
802 if (Offset >= AllocSize || Size > AllocSize || Offset + Size > AllocSize)
803 return markAsDead(I);
808 void visitBitCastInst(BitCastInst &BC) {
810 return markAsDead(BC);
812 enqueueUsers(BC, Offset);
815 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
816 if (GEPI.use_empty())
817 return markAsDead(GEPI);
820 if (!computeConstantGEPOffset(GEPI, GEPOffset))
821 llvm_unreachable("Unable to compute constant offset for use");
823 enqueueUsers(GEPI, GEPOffset);
826 void visitLoadInst(LoadInst &LI) {
827 handleLoadOrStore(LI.getType(), LI);
830 void visitStoreInst(StoreInst &SI) {
831 handleLoadOrStore(SI.getOperand(0)->getType(), SI);
834 void visitMemSetInst(MemSetInst &II) {
835 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
836 insertUse(Length ? Length->getZExtValue() : AllocSize - Offset, II);
839 void visitMemTransferInst(MemTransferInst &II) {
840 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength());
841 insertUse(Length ? Length->getZExtValue() : AllocSize - Offset, II);
844 void visitIntrinsicInst(IntrinsicInst &II) {
845 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
846 II.getIntrinsicID() == Intrinsic::lifetime_end);
848 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0));
849 insertUse(std::min(AllocSize - Offset, Length->getLimitedValue()), II);
852 void insertPHIOrSelect(Instruction &User) {
853 uint64_t Size = P.PHIOrSelectSizes.lookup(&User).first;
855 // For PHI and select operands outside the alloca, we can't nuke the entire
856 // phi or select -- the other side might still be relevant, so we special
857 // case them here and use a separate structure to track the operands
858 // themselves which should be replaced with undef.
859 if (Offset >= AllocSize) {
860 P.DeadOperands.push_back(U);
864 insertUse(Size, User);
866 void visitPHINode(PHINode &PN) {
868 return markAsDead(PN);
870 insertPHIOrSelect(PN);
872 void visitSelectInst(SelectInst &SI) {
874 return markAsDead(SI);
876 if (Value *Result = foldSelectInst(SI)) {
878 // If the result of the constant fold will be the pointer, recurse
879 // through the select as if we had RAUW'ed it.
880 enqueueUsers(SI, Offset);
885 insertPHIOrSelect(SI);
888 /// \brief Unreachable, we've already visited the alloca once.
889 void visitInstruction(Instruction &I) {
890 llvm_unreachable("Unhandled instruction in use builder.");
894 void AllocaPartitioning::splitAndMergePartitions() {
895 size_t NumDeadPartitions = 0;
897 // Track the range of splittable partitions that we pass when accumulating
898 // overlapping unsplittable partitions.
899 uint64_t SplitEndOffset = 0ull;
901 Partition New(0ull, 0ull, false);
903 for (unsigned i = 0, j = i, e = Partitions.size(); i != e; i = j) {
906 if (!Partitions[i].IsSplittable || New.BeginOffset == New.EndOffset) {
907 assert(New.BeginOffset == New.EndOffset);
910 assert(New.IsSplittable);
911 New.EndOffset = std::max(New.EndOffset, Partitions[i].EndOffset);
913 assert(New.BeginOffset != New.EndOffset);
915 // Scan the overlapping partitions.
916 while (j != e && New.EndOffset > Partitions[j].BeginOffset) {
917 // If the new partition we are forming is splittable, stop at the first
918 // unsplittable partition.
919 if (New.IsSplittable && !Partitions[j].IsSplittable)
922 // Grow the new partition to include any equally splittable range. 'j' is
923 // always equally splittable when New is splittable, but when New is not
924 // splittable, we may subsume some (or part of some) splitable partition
925 // without growing the new one.
926 if (New.IsSplittable == Partitions[j].IsSplittable) {
927 New.EndOffset = std::max(New.EndOffset, Partitions[j].EndOffset);
929 assert(!New.IsSplittable);
930 assert(Partitions[j].IsSplittable);
931 SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset);
934 Partitions[j].BeginOffset = Partitions[j].EndOffset = UINT64_MAX;
939 // If the new partition is splittable, chop off the end as soon as the
940 // unsplittable subsequent partition starts and ensure we eventually cover
941 // the splittable area.
942 if (j != e && New.IsSplittable) {
943 SplitEndOffset = std::max(SplitEndOffset, New.EndOffset);
944 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
947 // Add the new partition if it differs from the original one and is
948 // non-empty. We can end up with an empty partition here if it was
949 // splittable but there is an unsplittable one that starts at the same
951 if (New != Partitions[i]) {
952 if (New.BeginOffset != New.EndOffset)
953 Partitions.push_back(New);
954 // Mark the old one for removal.
955 Partitions[i].BeginOffset = Partitions[i].EndOffset = UINT64_MAX;
959 New.BeginOffset = New.EndOffset;
960 if (!New.IsSplittable) {
961 New.EndOffset = std::max(New.EndOffset, SplitEndOffset);
962 if (j != e && !Partitions[j].IsSplittable)
963 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset);
964 New.IsSplittable = true;
965 // If there is a trailing splittable partition which won't be fused into
966 // the next splittable partition go ahead and add it onto the partitions
968 if (New.BeginOffset < New.EndOffset &&
969 (j == e || !Partitions[j].IsSplittable ||
970 New.EndOffset < Partitions[j].BeginOffset)) {
971 Partitions.push_back(New);
972 New.BeginOffset = New.EndOffset = 0ull;
977 // Re-sort the partitions now that they have been split and merged into
978 // disjoint set of partitions. Also remove any of the dead partitions we've
979 // replaced in the process.
980 std::sort(Partitions.begin(), Partitions.end());
981 if (NumDeadPartitions) {
982 assert(Partitions.back().BeginOffset == UINT64_MAX);
983 assert(Partitions.back().EndOffset == UINT64_MAX);
984 assert((ptrdiff_t)NumDeadPartitions ==
985 std::count(Partitions.begin(), Partitions.end(), Partitions.back()));
987 Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end());
990 AllocaPartitioning::AllocaPartitioning(const TargetData &TD, AllocaInst &AI)
991 : AI(AI), PointerEscapingInstr(0) {
992 PartitionBuilder PB(TD, AI, *this);
996 if (Partitions.size() > 1) {
997 // Sort the uses. This arranges for the offsets to be in ascending order,
998 // and the sizes to be in descending order.
999 std::sort(Partitions.begin(), Partitions.end());
1001 // Intersect splittability for all partitions with equal offsets and sizes.
1002 // Then remove all but the first so that we have a sequence of non-equal but
1003 // potentially overlapping partitions.
1004 for (iterator I = Partitions.begin(), J = I, E = Partitions.end(); I != E;
1007 while (J != E && *I == *J) {
1008 I->IsSplittable &= J->IsSplittable;
1012 Partitions.erase(std::unique(Partitions.begin(), Partitions.end()),
1015 // Split splittable and merge unsplittable partitions into a disjoint set
1016 // of partitions over the used space of the allocation.
1017 splitAndMergePartitions();
1020 // Now build up the user lists for each of these disjoint partitions by
1021 // re-walking the recursive users of the alloca.
1022 Uses.resize(Partitions.size());
1023 UseBuilder UB(TD, AI, *this);
1025 for (iterator I = Partitions.begin(), E = Partitions.end(); I != E; ++I)
1026 std::stable_sort(use_begin(I), use_end(I));
1029 Type *AllocaPartitioning::getCommonType(iterator I) const {
1031 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) {
1032 if (isa<MemIntrinsic>(*UI->User))
1034 if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset)
1038 if (LoadInst *LI = dyn_cast<LoadInst>(&*UI->User)) {
1039 UserTy = LI->getType();
1040 } else if (StoreInst *SI = dyn_cast<StoreInst>(&*UI->User)) {
1041 UserTy = SI->getValueOperand()->getType();
1042 } else if (SelectInst *SI = dyn_cast<SelectInst>(&*UI->User)) {
1043 if (PointerType *PtrTy = dyn_cast<PointerType>(SI->getType()))
1044 UserTy = PtrTy->getElementType();
1045 } else if (PHINode *PN = dyn_cast<PHINode>(&*UI->User)) {
1046 if (PointerType *PtrTy = dyn_cast<PointerType>(PN->getType()))
1047 UserTy = PtrTy->getElementType();
1050 if (Ty && Ty != UserTy)
1058 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1060 void AllocaPartitioning::print(raw_ostream &OS, const_iterator I,
1061 StringRef Indent) const {
1062 OS << Indent << "partition #" << (I - begin())
1063 << " [" << I->BeginOffset << "," << I->EndOffset << ")"
1064 << (I->IsSplittable ? " (splittable)" : "")
1065 << (Uses[I - begin()].empty() ? " (zero uses)" : "")
1069 void AllocaPartitioning::printUsers(raw_ostream &OS, const_iterator I,
1070 StringRef Indent) const {
1071 for (const_use_iterator UI = use_begin(I), UE = use_end(I);
1073 OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") "
1074 << "used by: " << *UI->User << "\n";
1075 if (MemTransferInst *II = dyn_cast<MemTransferInst>(&*UI->User)) {
1076 const MemTransferOffsets &MTO = MemTransferInstData.lookup(II);
1078 if (!MTO.IsSplittable)
1079 IsDest = UI->BeginOffset == MTO.DestBegin;
1081 IsDest = MTO.DestBegin != 0u;
1082 OS << Indent << " (original " << (IsDest ? "dest" : "source") << ": "
1083 << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin)
1084 << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n";
1089 void AllocaPartitioning::print(raw_ostream &OS) const {
1090 if (PointerEscapingInstr) {
1091 OS << "No partitioning for alloca: " << AI << "\n"
1092 << " A pointer to this alloca escaped by:\n"
1093 << " " << *PointerEscapingInstr << "\n";
1097 OS << "Partitioning of alloca: " << AI << "\n";
1099 for (const_iterator I = begin(), E = end(); I != E; ++I, ++Num) {
1105 void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); }
1106 void AllocaPartitioning::dump() const { print(dbgs()); }
1108 #endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1112 /// \brief An optimization pass providing Scalar Replacement of Aggregates.
1114 /// This pass takes allocations which can be completely analyzed (that is, they
1115 /// don't escape) and tries to turn them into scalar SSA values. There are
1116 /// a few steps to this process.
1118 /// 1) It takes allocations of aggregates and analyzes the ways in which they
1119 /// are used to try to split them into smaller allocations, ideally of
1120 /// a single scalar data type. It will split up memcpy and memset accesses
1121 /// as necessary and try to isolate invidual scalar accesses.
1122 /// 2) It will transform accesses into forms which are suitable for SSA value
1123 /// promotion. This can be replacing a memset with a scalar store of an
1124 /// integer value, or it can involve speculating operations on a PHI or
1125 /// select to be a PHI or select of the results.
1126 /// 3) Finally, this will try to detect a pattern of accesses which map cleanly
1127 /// onto insert and extract operations on a vector value, and convert them to
1128 /// this form. By doing so, it will enable promotion of vector aggregates to
1129 /// SSA vector values.
1130 class SROA : public FunctionPass {
1132 const TargetData *TD;
1135 /// \brief Worklist of alloca instructions to simplify.
1137 /// Each alloca in the function is added to this. Each new alloca formed gets
1138 /// added to it as well to recursively simplify unless that alloca can be
1139 /// directly promoted. Finally, each time we rewrite a use of an alloca other
1140 /// the one being actively rewritten, we add it back onto the list if not
1141 /// already present to ensure it is re-visited.
1142 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist;
1144 /// \brief A collection of instructions to delete.
1145 /// We try to batch deletions to simplify code and make things a bit more
1147 SmallVector<Instruction *, 8> DeadInsts;
1149 /// \brief A set to prevent repeatedly marking an instruction split into many
1150 /// uses as dead. Only used to guard insertion into DeadInsts.
1151 SmallPtrSet<Instruction *, 4> DeadSplitInsts;
1153 /// \brief A collection of alloca instructions we can directly promote.
1154 std::vector<AllocaInst *> PromotableAllocas;
1157 SROA() : FunctionPass(ID), C(0), TD(0), DT(0) {
1158 initializeSROAPass(*PassRegistry::getPassRegistry());
1160 bool runOnFunction(Function &F);
1161 void getAnalysisUsage(AnalysisUsage &AU) const;
1163 const char *getPassName() const { return "SROA"; }
1167 friend class AllocaPartitionRewriter;
1168 friend class AllocaPartitionVectorRewriter;
1170 bool rewriteAllocaPartition(AllocaInst &AI,
1171 AllocaPartitioning &P,
1172 AllocaPartitioning::iterator PI);
1173 bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P);
1174 bool runOnAlloca(AllocaInst &AI);
1175 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas);
1181 FunctionPass *llvm::createSROAPass() {
1185 INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates",
1187 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
1188 INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates",
1191 /// \brief Accumulate the constant offsets in a GEP into a single APInt offset.
1193 /// If the provided GEP is all-constant, the total byte offset formed by the
1194 /// GEP is computed and Offset is set to it. If the GEP has any non-constant
1195 /// operands, the function returns false and the value of Offset is unmodified.
1196 static bool accumulateGEPOffsets(const TargetData &TD, GEPOperator &GEP,
1198 APInt GEPOffset(Offset.getBitWidth(), 0);
1199 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP);
1200 GTI != GTE; ++GTI) {
1201 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
1204 if (OpC->isZero()) continue;
1206 // Handle a struct index, which adds its field offset to the pointer.
1207 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1208 unsigned ElementIdx = OpC->getZExtValue();
1209 const StructLayout *SL = TD.getStructLayout(STy);
1210 GEPOffset += APInt(Offset.getBitWidth(),
1211 SL->getElementOffset(ElementIdx));
1215 APInt TypeSize(Offset.getBitWidth(),
1216 TD.getTypeAllocSize(GTI.getIndexedType()));
1217 if (VectorType *VTy = dyn_cast<VectorType>(*GTI)) {
1218 assert((VTy->getScalarSizeInBits() % 8) == 0 &&
1219 "vector element size is not a multiple of 8, cannot GEP over it");
1220 TypeSize = VTy->getScalarSizeInBits() / 8;
1223 GEPOffset += OpC->getValue().sextOrTrunc(Offset.getBitWidth()) * TypeSize;
1229 /// \brief Build a GEP out of a base pointer and indices.
1231 /// This will return the BasePtr if that is valid, or build a new GEP
1232 /// instruction using the IRBuilder if GEP-ing is needed.
1233 static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr,
1234 SmallVectorImpl<Value *> &Indices,
1235 const Twine &Prefix) {
1236 if (Indices.empty())
1239 // A single zero index is a no-op, so check for this and avoid building a GEP
1241 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero())
1244 return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx");
1247 /// \brief Get a natural GEP off of the BasePtr walking through Ty toward
1248 /// TargetTy without changing the offset of the pointer.
1250 /// This routine assumes we've already established a properly offset GEP with
1251 /// Indices, and arrived at the Ty type. The goal is to continue to GEP with
1252 /// zero-indices down through type layers until we find one the same as
1253 /// TargetTy. If we can't find one with the same type, we at least try to use
1254 /// one with the same size. If none of that works, we just produce the GEP as
1255 /// indicated by Indices to have the correct offset.
1256 static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const TargetData &TD,
1257 Value *BasePtr, Type *Ty, Type *TargetTy,
1258 SmallVectorImpl<Value *> &Indices,
1259 const Twine &Prefix) {
1261 return buildGEP(IRB, BasePtr, Indices, Prefix);
1263 // See if we can descend into a struct and locate a field with the correct
1265 unsigned NumLayers = 0;
1266 Type *ElementTy = Ty;
1268 if (ElementTy->isPointerTy())
1270 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) {
1271 ElementTy = SeqTy->getElementType();
1272 Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(), 0)));
1273 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) {
1274 ElementTy = *STy->element_begin();
1275 Indices.push_back(IRB.getInt32(0));
1280 } while (ElementTy != TargetTy);
1281 if (ElementTy != TargetTy)
1282 Indices.erase(Indices.end() - NumLayers, Indices.end());
1284 return buildGEP(IRB, BasePtr, Indices, Prefix);
1287 /// \brief Recursively compute indices for a natural GEP.
1289 /// This is the recursive step for getNaturalGEPWithOffset that walks down the
1290 /// element types adding appropriate indices for the GEP.
1291 static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const TargetData &TD,
1292 Value *Ptr, Type *Ty, APInt &Offset,
1294 SmallVectorImpl<Value *> &Indices,
1295 const Twine &Prefix) {
1297 return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix);
1299 // We can't recurse through pointer types.
1300 if (Ty->isPointerTy())
1303 // We try to analyze GEPs over vectors here, but note that these GEPs are
1304 // extremely poorly defined currently. The long-term goal is to remove GEPing
1305 // over a vector from the IR completely.
1306 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
1307 unsigned ElementSizeInBits = VecTy->getScalarSizeInBits();
1308 if (ElementSizeInBits % 8)
1309 return 0; // GEPs over non-multiple of 8 size vector elements are invalid.
1310 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8);
1311 APInt NumSkippedElements = Offset.udiv(ElementSize);
1312 if (NumSkippedElements.ugt(VecTy->getNumElements()))
1314 Offset -= NumSkippedElements * ElementSize;
1315 Indices.push_back(IRB.getInt(NumSkippedElements));
1316 return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(),
1317 Offset, TargetTy, Indices, Prefix);
1320 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) {
1321 Type *ElementTy = ArrTy->getElementType();
1322 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1323 APInt NumSkippedElements = Offset.udiv(ElementSize);
1324 if (NumSkippedElements.ugt(ArrTy->getNumElements()))
1327 Offset -= NumSkippedElements * ElementSize;
1328 Indices.push_back(IRB.getInt(NumSkippedElements));
1329 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1333 StructType *STy = dyn_cast<StructType>(Ty);
1337 const StructLayout *SL = TD.getStructLayout(STy);
1338 uint64_t StructOffset = Offset.getZExtValue();
1339 if (StructOffset >= SL->getSizeInBytes())
1341 unsigned Index = SL->getElementContainingOffset(StructOffset);
1342 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index));
1343 Type *ElementTy = STy->getElementType(Index);
1344 if (Offset.uge(TD.getTypeAllocSize(ElementTy)))
1345 return 0; // The offset points into alignment padding.
1347 Indices.push_back(IRB.getInt32(Index));
1348 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1352 /// \brief Get a natural GEP from a base pointer to a particular offset and
1353 /// resulting in a particular type.
1355 /// The goal is to produce a "natural" looking GEP that works with the existing
1356 /// composite types to arrive at the appropriate offset and element type for
1357 /// a pointer. TargetTy is the element type the returned GEP should point-to if
1358 /// possible. We recurse by decreasing Offset, adding the appropriate index to
1359 /// Indices, and setting Ty to the result subtype.
1361 /// If no natural GEP can be constructed, this function returns null.
1362 static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const TargetData &TD,
1363 Value *Ptr, APInt Offset, Type *TargetTy,
1364 SmallVectorImpl<Value *> &Indices,
1365 const Twine &Prefix) {
1366 PointerType *Ty = cast<PointerType>(Ptr->getType());
1368 // Don't consider any GEPs through an i8* as natural unless the TargetTy is
1370 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8))
1373 Type *ElementTy = Ty->getElementType();
1374 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy));
1375 if (ElementSize == 0)
1376 return 0; // Zero-length arrays can't help us build a natural GEP.
1377 APInt NumSkippedElements = Offset.udiv(ElementSize);
1379 Offset -= NumSkippedElements * ElementSize;
1380 Indices.push_back(IRB.getInt(NumSkippedElements));
1381 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy,
1385 /// \brief Compute an adjusted pointer from Ptr by Offset bytes where the
1386 /// resulting pointer has PointerTy.
1388 /// This tries very hard to compute a "natural" GEP which arrives at the offset
1389 /// and produces the pointer type desired. Where it cannot, it will try to use
1390 /// the natural GEP to arrive at the offset and bitcast to the type. Where that
1391 /// fails, it will try to use an existing i8* and GEP to the byte offset and
1392 /// bitcast to the type.
1394 /// The strategy for finding the more natural GEPs is to peel off layers of the
1395 /// pointer, walking back through bit casts and GEPs, searching for a base
1396 /// pointer from which we can compute a natural GEP with the desired
1397 /// properities. The algorithm tries to fold as many constant indices into
1398 /// a single GEP as possible, thus making each GEP more independent of the
1399 /// surrounding code.
1400 static Value *getAdjustedPtr(IRBuilder<> &IRB, const TargetData &TD,
1401 Value *Ptr, APInt Offset, Type *PointerTy,
1402 const Twine &Prefix) {
1403 // Even though we don't look through PHI nodes, we could be called on an
1404 // instruction in an unreachable block, which may be on a cycle.
1405 SmallPtrSet<Value *, 4> Visited;
1406 Visited.insert(Ptr);
1407 SmallVector<Value *, 4> Indices;
1409 // We may end up computing an offset pointer that has the wrong type. If we
1410 // never are able to compute one directly that has the correct type, we'll
1411 // fall back to it, so keep it around here.
1412 Value *OffsetPtr = 0;
1414 // Remember any i8 pointer we come across to re-use if we need to do a raw
1417 APInt Int8PtrOffset(Offset.getBitWidth(), 0);
1419 Type *TargetTy = PointerTy->getPointerElementType();
1422 // First fold any existing GEPs into the offset.
1423 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) {
1424 APInt GEPOffset(Offset.getBitWidth(), 0);
1425 if (!accumulateGEPOffsets(TD, *GEP, GEPOffset))
1427 Offset += GEPOffset;
1428 Ptr = GEP->getPointerOperand();
1429 if (!Visited.insert(Ptr))
1433 // See if we can perform a natural GEP here.
1435 if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy,
1437 if (P->getType() == PointerTy) {
1438 // Zap any offset pointer that we ended up computing in previous rounds.
1439 if (OffsetPtr && OffsetPtr->use_empty())
1440 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr))
1441 I->eraseFromParent();
1449 // Stash this pointer if we've found an i8*.
1450 if (Ptr->getType()->isIntegerTy(8)) {
1452 Int8PtrOffset = Offset;
1455 // Peel off a layer of the pointer and update the offset appropriately.
1456 if (Operator::getOpcode(Ptr) == Instruction::BitCast) {
1457 Ptr = cast<Operator>(Ptr)->getOperand(0);
1458 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) {
1459 if (GA->mayBeOverridden())
1461 Ptr = GA->getAliasee();
1465 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!");
1466 } while (Visited.insert(Ptr));
1470 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(),
1471 Prefix + ".raw_cast");
1472 Int8PtrOffset = Offset;
1475 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr :
1476 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset),
1477 Prefix + ".raw_idx");
1481 // On the off chance we were targeting i8*, guard the bitcast here.
1482 if (Ptr->getType() != PointerTy)
1483 Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast");
1488 /// \brief Test whether the given alloca partition can be promoted to a vector.
1490 /// This is a quick test to check whether we can rewrite a particular alloca
1491 /// partition (and its newly formed alloca) into a vector alloca with only
1492 /// whole-vector loads and stores such that it could be promoted to a vector
1493 /// SSA value. We only can ensure this for a limited set of operations, and we
1494 /// don't want to do the rewrites unless we are confident that the result will
1495 /// be promotable, so we have an early test here.
1496 static bool isVectorPromotionViable(const TargetData &TD,
1498 AllocaPartitioning &P,
1499 uint64_t PartitionBeginOffset,
1500 uint64_t PartitionEndOffset,
1501 AllocaPartitioning::const_use_iterator I,
1502 AllocaPartitioning::const_use_iterator E) {
1503 VectorType *Ty = dyn_cast<VectorType>(AllocaTy);
1507 uint64_t VecSize = TD.getTypeSizeInBits(Ty);
1508 uint64_t ElementSize = Ty->getScalarSizeInBits();
1510 // While the definition of LLVM vectors is bitpacked, we don't support sizes
1511 // that aren't byte sized.
1512 if (ElementSize % 8)
1514 assert((VecSize % 8) == 0 && "vector size not a multiple of element size?");
1518 for (; I != E; ++I) {
1519 uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset;
1520 uint64_t BeginIndex = BeginOffset / ElementSize;
1521 if (BeginIndex * ElementSize != BeginOffset ||
1522 BeginIndex >= Ty->getNumElements())
1524 uint64_t EndOffset = I->EndOffset - PartitionBeginOffset;
1525 uint64_t EndIndex = EndOffset / ElementSize;
1526 if (EndIndex * ElementSize != EndOffset ||
1527 EndIndex > Ty->getNumElements())
1530 // FIXME: We should build shuffle vector instructions to handle
1531 // non-element-sized accesses.
1532 if ((EndOffset - BeginOffset) != ElementSize &&
1533 (EndOffset - BeginOffset) != VecSize)
1536 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(&*I->User)) {
1537 if (MI->isVolatile())
1539 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(&*I->User)) {
1540 const AllocaPartitioning::MemTransferOffsets &MTO
1541 = P.getMemTransferOffsets(*MTI);
1542 if (!MTO.IsSplittable)
1545 } else if (I->Ptr->getType()->getPointerElementType()->isStructTy()) {
1546 // Disable vector promotion when there are loads or stores of an FCA.
1548 } else if (!isa<LoadInst>(*I->User) && !isa<StoreInst>(*I->User)) {
1556 /// \brief Visitor to rewrite instructions using a partition of an alloca to
1557 /// use a new alloca.
1559 /// Also implements the rewriting to vector-based accesses when the partition
1560 /// passes the isVectorPromotionViable predicate. Most of the rewriting logic
1562 class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter,
1564 // Befriend the base class so it can delegate to private visit methods.
1565 friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>;
1567 const TargetData &TD;
1568 AllocaPartitioning &P;
1570 AllocaInst &OldAI, &NewAI;
1571 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset;
1573 // If we are rewriting an alloca partition which can be written as pure
1574 // vector operations, we stash extra information here. When VecTy is
1575 // non-null, we have some strict guarantees about the rewriten alloca:
1576 // - The new alloca is exactly the size of the vector type here.
1577 // - The accesses all either map to the entire vector or to a single
1579 // - The set of accessing instructions is only one of those handled above
1580 // in isVectorPromotionViable. Generally these are the same access kinds
1581 // which are promotable via mem2reg.
1584 uint64_t ElementSize;
1586 // The offset of the partition user currently being rewritten.
1587 uint64_t BeginOffset, EndOffset;
1588 Instruction *OldPtr;
1590 // The name prefix to use when rewriting instructions for this alloca.
1591 std::string NamePrefix;
1594 AllocaPartitionRewriter(const TargetData &TD, AllocaPartitioning &P,
1595 AllocaPartitioning::iterator PI,
1596 SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI,
1597 uint64_t NewBeginOffset, uint64_t NewEndOffset)
1598 : TD(TD), P(P), Pass(Pass),
1599 OldAI(OldAI), NewAI(NewAI),
1600 NewAllocaBeginOffset(NewBeginOffset),
1601 NewAllocaEndOffset(NewEndOffset),
1602 VecTy(), ElementTy(), ElementSize(),
1603 BeginOffset(), EndOffset() {
1606 /// \brief Visit the users of the alloca partition and rewrite them.
1607 bool visitUsers(AllocaPartitioning::const_use_iterator I,
1608 AllocaPartitioning::const_use_iterator E) {
1609 if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P,
1610 NewAllocaBeginOffset, NewAllocaEndOffset,
1613 VecTy = cast<VectorType>(NewAI.getAllocatedType());
1614 ElementTy = VecTy->getElementType();
1615 assert((VecTy->getScalarSizeInBits() % 8) == 0 &&
1616 "Only multiple-of-8 sized vector elements are viable");
1617 ElementSize = VecTy->getScalarSizeInBits() / 8;
1619 bool CanSROA = true;
1620 for (; I != E; ++I) {
1621 BeginOffset = I->BeginOffset;
1622 EndOffset = I->EndOffset;
1624 NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str();
1625 CanSROA &= visit(I->User);
1637 // Every instruction which can end up as a user must have a rewrite rule.
1638 bool visitInstruction(Instruction &I) {
1639 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n");
1640 llvm_unreachable("No rewrite rule for this instruction!");
1643 Twine getName(const Twine &Suffix) {
1644 return NamePrefix + Suffix;
1647 Value *getAdjustedAllocaPtr(IRBuilder<> &IRB, Type *PointerTy) {
1648 assert(BeginOffset >= NewAllocaBeginOffset);
1649 APInt Offset(TD.getPointerSizeInBits(), BeginOffset - NewAllocaBeginOffset);
1650 return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName(""));
1653 ConstantInt *getIndex(IRBuilder<> &IRB, uint64_t Offset) {
1654 assert(VecTy && "Can only call getIndex when rewriting a vector");
1655 uint64_t RelOffset = Offset - NewAllocaBeginOffset;
1656 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds");
1657 uint32_t Index = RelOffset / ElementSize;
1658 assert(Index * ElementSize == RelOffset);
1659 return IRB.getInt32(Index);
1662 void deleteIfTriviallyDead(Value *V) {
1663 Instruction *I = cast<Instruction>(V);
1664 if (isInstructionTriviallyDead(I))
1665 Pass.DeadInsts.push_back(I);
1668 Value *getValueCast(IRBuilder<> &IRB, Value *V, Type *Ty) {
1669 if (V->getType()->isIntegerTy() && Ty->isPointerTy())
1670 return IRB.CreateIntToPtr(V, Ty);
1671 if (V->getType()->isPointerTy() && Ty->isIntegerTy())
1672 return IRB.CreatePtrToInt(V, Ty);
1674 return IRB.CreateBitCast(V, Ty);
1677 bool rewriteVectorizedLoadInst(IRBuilder<> &IRB, LoadInst &LI, Value *OldOp) {
1679 if (LI.getType() == VecTy->getElementType() ||
1680 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
1682 = IRB.CreateExtractElement(IRB.CreateLoad(&NewAI, getName(".load")),
1683 getIndex(IRB, BeginOffset),
1684 getName(".extract"));
1686 Result = IRB.CreateLoad(&NewAI, getName(".load"));
1688 if (Result->getType() != LI.getType())
1689 Result = getValueCast(IRB, Result, LI.getType());
1690 LI.replaceAllUsesWith(Result);
1691 Pass.DeadInsts.push_back(&LI);
1693 DEBUG(dbgs() << " to: " << *Result << "\n");
1697 bool visitLoadInst(LoadInst &LI) {
1698 DEBUG(dbgs() << " original: " << LI << "\n");
1699 Value *OldOp = LI.getOperand(0);
1700 assert(OldOp == OldPtr);
1701 IRBuilder<> IRB(&LI);
1704 return rewriteVectorizedLoadInst(IRB, LI, OldOp);
1706 Value *NewPtr = getAdjustedAllocaPtr(IRB,
1707 LI.getPointerOperand()->getType());
1708 LI.setOperand(0, NewPtr);
1709 DEBUG(dbgs() << " to: " << LI << "\n");
1711 deleteIfTriviallyDead(OldOp);
1712 return NewPtr == &NewAI && !LI.isVolatile();
1715 bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, StoreInst &SI,
1717 Value *V = SI.getValueOperand();
1718 if (V->getType() == ElementTy ||
1719 BeginOffset > NewAllocaBeginOffset || EndOffset < NewAllocaEndOffset) {
1720 if (V->getType() != ElementTy)
1721 V = getValueCast(IRB, V, ElementTy);
1722 V = IRB.CreateInsertElement(IRB.CreateLoad(&NewAI, getName(".load")), V,
1723 getIndex(IRB, BeginOffset),
1724 getName(".insert"));
1725 } else if (V->getType() != VecTy) {
1726 V = getValueCast(IRB, V, VecTy);
1728 StoreInst *Store = IRB.CreateStore(V, &NewAI);
1729 Pass.DeadInsts.push_back(&SI);
1732 DEBUG(dbgs() << " to: " << *Store << "\n");
1736 bool visitStoreInst(StoreInst &SI) {
1737 DEBUG(dbgs() << " original: " << SI << "\n");
1738 Value *OldOp = SI.getOperand(1);
1739 assert(OldOp == OldPtr);
1740 IRBuilder<> IRB(&SI);
1743 return rewriteVectorizedStoreInst(IRB, SI, OldOp);
1745 Value *NewPtr = getAdjustedAllocaPtr(IRB,
1746 SI.getPointerOperand()->getType());
1747 SI.setOperand(1, NewPtr);
1748 DEBUG(dbgs() << " to: " << SI << "\n");
1750 deleteIfTriviallyDead(OldOp);
1751 return NewPtr == &NewAI && !SI.isVolatile();
1754 bool visitMemSetInst(MemSetInst &II) {
1755 DEBUG(dbgs() << " original: " << II << "\n");
1756 IRBuilder<> IRB(&II);
1757 assert(II.getRawDest() == OldPtr);
1759 // If the memset has a variable size, it cannot be split, just adjust the
1760 // pointer to the new alloca.
1761 if (!isa<Constant>(II.getLength())) {
1762 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
1763 deleteIfTriviallyDead(OldPtr);
1767 // Record this instruction for deletion.
1768 if (Pass.DeadSplitInsts.insert(&II))
1769 Pass.DeadInsts.push_back(&II);
1771 Type *AllocaTy = NewAI.getAllocatedType();
1772 Type *ScalarTy = AllocaTy->getScalarType();
1774 // If this doesn't map cleanly onto the alloca type, and that type isn't
1775 // a single value type, just emit a memset.
1776 if (!VecTy && (BeginOffset != NewAllocaBeginOffset ||
1777 EndOffset != NewAllocaEndOffset ||
1778 !AllocaTy->isSingleValueType() ||
1779 !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)))) {
1780 Type *SizeTy = II.getLength()->getType();
1781 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
1784 = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB,
1785 II.getRawDest()->getType()),
1786 II.getValue(), Size, II.getAlignment(),
1789 DEBUG(dbgs() << " to: " << *New << "\n");
1793 // If we can represent this as a simple value, we have to build the actual
1794 // value to store, which requires expanding the byte present in memset to
1795 // a sensible representation for the alloca type. This is essentially
1796 // splatting the byte to a sufficiently wide integer, bitcasting to the
1797 // desired scalar type, and splatting it across any desired vector type.
1798 Value *V = II.getValue();
1799 IntegerType *VTy = cast<IntegerType>(V->getType());
1800 Type *IntTy = Type::getIntNTy(VTy->getContext(),
1801 TD.getTypeSizeInBits(ScalarTy));
1802 if (TD.getTypeSizeInBits(ScalarTy) > VTy->getBitWidth())
1803 V = IRB.CreateMul(IRB.CreateZExt(V, IntTy, getName(".zext")),
1804 ConstantExpr::getUDiv(
1805 Constant::getAllOnesValue(IntTy),
1806 ConstantExpr::getZExt(
1807 Constant::getAllOnesValue(V->getType()),
1809 getName(".isplat"));
1810 if (V->getType() != ScalarTy) {
1811 if (ScalarTy->isPointerTy())
1812 V = IRB.CreateIntToPtr(V, ScalarTy);
1813 else if (ScalarTy->isPrimitiveType() || ScalarTy->isVectorTy())
1814 V = IRB.CreateBitCast(V, ScalarTy);
1815 else if (ScalarTy->isIntegerTy())
1816 llvm_unreachable("Computed different integer types with equal widths");
1818 llvm_unreachable("Invalid scalar type");
1821 // If this is an element-wide memset of a vectorizable alloca, insert it.
1822 if (VecTy && (BeginOffset > NewAllocaBeginOffset ||
1823 EndOffset < NewAllocaEndOffset)) {
1824 StoreInst *Store = IRB.CreateStore(
1825 IRB.CreateInsertElement(IRB.CreateLoad(&NewAI, getName(".load")), V,
1826 getIndex(IRB, BeginOffset),
1827 getName(".insert")),
1830 DEBUG(dbgs() << " to: " << *Store << "\n");
1834 // Splat to a vector if needed.
1835 if (VectorType *VecTy = dyn_cast<VectorType>(AllocaTy)) {
1836 VectorType *SplatSourceTy = VectorType::get(V->getType(), 1);
1837 V = IRB.CreateShuffleVector(
1838 IRB.CreateInsertElement(UndefValue::get(SplatSourceTy), V,
1839 IRB.getInt32(0), getName(".vsplat.insert")),
1840 UndefValue::get(SplatSourceTy),
1841 ConstantVector::getSplat(VecTy->getNumElements(), IRB.getInt32(0)),
1842 getName(".vsplat.shuffle"));
1843 assert(V->getType() == VecTy);
1846 Value *New = IRB.CreateStore(V, &NewAI, II.isVolatile());
1848 DEBUG(dbgs() << " to: " << *New << "\n");
1849 return !II.isVolatile();
1852 bool visitMemTransferInst(MemTransferInst &II) {
1853 // Rewriting of memory transfer instructions can be a bit tricky. We break
1854 // them into two categories: split intrinsics and unsplit intrinsics.
1856 DEBUG(dbgs() << " original: " << II << "\n");
1857 IRBuilder<> IRB(&II);
1859 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr);
1860 bool IsDest = II.getRawDest() == OldPtr;
1862 const AllocaPartitioning::MemTransferOffsets &MTO
1863 = P.getMemTransferOffsets(II);
1865 // For unsplit intrinsics, we simply modify the source and destination
1866 // pointers in place. This isn't just an optimization, it is a matter of
1867 // correctness. With unsplit intrinsics we may be dealing with transfers
1868 // within a single alloca before SROA ran, or with transfers that have
1869 // a variable length. We may also be dealing with memmove instead of
1870 // memcpy, and so simply updating the pointers is the necessary for us to
1871 // update both source and dest of a single call.
1872 if (!MTO.IsSplittable) {
1873 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource();
1875 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType()));
1877 II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType()));
1879 DEBUG(dbgs() << " to: " << II << "\n");
1880 deleteIfTriviallyDead(OldOp);
1883 // For split transfer intrinsics we have an incredibly useful assurance:
1884 // the source and destination do not reside within the same alloca, and at
1885 // least one of them does not escape. This means that we can replace
1886 // memmove with memcpy, and we don't need to worry about all manner of
1887 // downsides to splitting and transforming the operations.
1889 // Compute the relative offset within the transfer.
1890 unsigned IntPtrWidth = TD.getPointerSizeInBits();
1891 APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin
1892 : MTO.SourceBegin));
1894 // If this doesn't map cleanly onto the alloca type, and that type isn't
1895 // a single value type, just emit a memcpy.
1897 = !VecTy && (BeginOffset != NewAllocaBeginOffset ||
1898 EndOffset != NewAllocaEndOffset ||
1899 !NewAI.getAllocatedType()->isSingleValueType());
1901 // If we're just going to emit a memcpy, the alloca hasn't changed, and the
1902 // size hasn't been shrunk based on analysis of the viable range, this is
1904 if (EmitMemCpy && &OldAI == &NewAI) {
1905 uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin;
1906 uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd;
1907 // Ensure the start lines up.
1908 assert(BeginOffset == OrigBegin);
1910 // Rewrite the size as needed.
1911 if (EndOffset != OrigEnd)
1912 II.setLength(ConstantInt::get(II.getLength()->getType(),
1913 EndOffset - BeginOffset));
1916 // Record this instruction for deletion.
1917 if (Pass.DeadSplitInsts.insert(&II))
1918 Pass.DeadInsts.push_back(&II);
1920 bool IsVectorElement = VecTy && (BeginOffset > NewAllocaBeginOffset ||
1921 EndOffset < NewAllocaEndOffset);
1923 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType()
1924 : II.getRawDest()->getType();
1926 OtherPtrTy = IsVectorElement ? VecTy->getElementType()->getPointerTo()
1929 // Compute the other pointer, folding as much as possible to produce
1930 // a single, simple GEP in most cases.
1931 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest();
1932 OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy,
1933 getName("." + OtherPtr->getName()));
1935 // Strip all inbounds GEPs and pointer casts to try to dig out any root
1936 // alloca that should be re-examined after rewriting this instruction.
1938 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets()))
1939 Pass.Worklist.insert(AI);
1943 = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType()
1944 : II.getRawSource()->getType());
1945 Type *SizeTy = II.getLength()->getType();
1946 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset);
1948 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr,
1949 IsDest ? OtherPtr : OurPtr,
1950 Size, II.getAlignment(),
1953 DEBUG(dbgs() << " to: " << *New << "\n");
1957 Value *SrcPtr = OtherPtr;
1958 Value *DstPtr = &NewAI;
1960 std::swap(SrcPtr, DstPtr);
1963 if (IsVectorElement && !IsDest) {
1964 // We have to extract rather than load.
1965 Src = IRB.CreateExtractElement(IRB.CreateLoad(SrcPtr,
1966 getName(".copyload")),
1967 getIndex(IRB, BeginOffset),
1968 getName(".copyextract"));
1970 Src = IRB.CreateLoad(SrcPtr, II.isVolatile(), getName(".copyload"));
1973 if (IsVectorElement && IsDest) {
1974 // We have to insert into a loaded copy before storing.
1975 Src = IRB.CreateInsertElement(IRB.CreateLoad(&NewAI, getName(".load")),
1976 Src, getIndex(IRB, BeginOffset),
1977 getName(".insert"));
1980 Value *Store = IRB.CreateStore(Src, DstPtr, II.isVolatile());
1982 DEBUG(dbgs() << " to: " << *Store << "\n");
1983 return !II.isVolatile();
1986 bool visitIntrinsicInst(IntrinsicInst &II) {
1987 assert(II.getIntrinsicID() == Intrinsic::lifetime_start ||
1988 II.getIntrinsicID() == Intrinsic::lifetime_end);
1989 DEBUG(dbgs() << " original: " << II << "\n");
1990 IRBuilder<> IRB(&II);
1991 assert(II.getArgOperand(1) == OldPtr);
1993 // Record this instruction for deletion.
1994 if (Pass.DeadSplitInsts.insert(&II))
1995 Pass.DeadInsts.push_back(&II);
1998 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()),
1999 EndOffset - BeginOffset);
2000 Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType());
2002 if (II.getIntrinsicID() == Intrinsic::lifetime_start)
2003 New = IRB.CreateLifetimeStart(Ptr, Size);
2005 New = IRB.CreateLifetimeEnd(Ptr, Size);
2007 DEBUG(dbgs() << " to: " << *New << "\n");
2011 /// PHI instructions that use an alloca and are subsequently loaded can be
2012 /// rewritten to load both input pointers in the pred blocks and then PHI the
2013 /// results, allowing the load of the alloca to be promoted.
2015 /// %P2 = phi [i32* %Alloca, i32* %Other]
2016 /// %V = load i32* %P2
2018 /// %V1 = load i32* %Alloca -> will be mem2reg'd
2020 /// %V2 = load i32* %Other
2022 /// %V = phi [i32 %V1, i32 %V2]
2024 /// We can do this to a select if its only uses are loads and if the operand
2025 /// to the select can be loaded unconditionally.
2027 /// FIXME: This should be hoisted into a generic utility, likely in
2028 /// Transforms/Util/Local.h
2029 bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) {
2030 // For now, we can only do this promotion if the load is in the same block
2031 // as the PHI, and if there are no stores between the phi and load.
2032 // TODO: Allow recursive phi users.
2033 // TODO: Allow stores.
2034 BasicBlock *BB = PN.getParent();
2035 unsigned MaxAlign = 0;
2036 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end();
2038 LoadInst *LI = dyn_cast<LoadInst>(*UI);
2039 if (LI == 0 || !LI->isSimple()) return false;
2041 // For now we only allow loads in the same block as the PHI. This is
2042 // a common case that happens when instcombine merges two loads through
2044 if (LI->getParent() != BB) return false;
2046 // Ensure that there are no instructions between the PHI and the load that
2048 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI)
2049 if (BBI->mayWriteToMemory())
2052 MaxAlign = std::max(MaxAlign, LI->getAlignment());
2053 Loads.push_back(LI);
2056 // We can only transform this if it is safe to push the loads into the
2057 // predecessor blocks. The only thing to watch out for is that we can't put
2058 // a possibly trapping load in the predecessor if it is a critical edge.
2059 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num;
2061 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator();
2062 Value *InVal = PN.getIncomingValue(Idx);
2064 // If the value is produced by the terminator of the predecessor (an
2065 // invoke) or it has side-effects, there is no valid place to put a load
2066 // in the predecessor.
2067 if (TI == InVal || TI->mayHaveSideEffects())
2070 // If the predecessor has a single successor, then the edge isn't
2072 if (TI->getNumSuccessors() == 1)
2075 // If this pointer is always safe to load, or if we can prove that there
2076 // is already a load in the block, then we can move the load to the pred
2078 if (InVal->isDereferenceablePointer() ||
2079 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD))
2088 bool visitPHINode(PHINode &PN) {
2089 DEBUG(dbgs() << " original: " << PN << "\n");
2090 // We would like to compute a new pointer in only one place, but have it be
2091 // as local as possible to the PHI. To do that, we re-use the location of
2092 // the old pointer, which necessarily must be in the right position to
2093 // dominate the PHI.
2094 IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr));
2096 SmallVector<LoadInst *, 4> Loads;
2097 if (!isSafePHIToSpeculate(PN, Loads)) {
2098 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
2099 // Replace the operands which were using the old pointer.
2100 User::op_iterator OI = PN.op_begin(), OE = PN.op_end();
2101 for (; OI != OE; ++OI)
2105 DEBUG(dbgs() << " to: " << PN << "\n");
2106 deleteIfTriviallyDead(OldPtr);
2109 assert(!Loads.empty());
2111 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType();
2112 IRBuilder<> PHIBuilder(&PN);
2113 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues());
2114 NewPN->takeName(&PN);
2116 // Get the TBAA tag and alignment to use from one of the loads. It doesn't
2117 // matter which one we get and if any differ, it doesn't matter.
2118 LoadInst *SomeLoad = cast<LoadInst>(Loads.back());
2119 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa);
2120 unsigned Align = SomeLoad->getAlignment();
2121 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType());
2123 // Rewrite all loads of the PN to use the new PHI.
2125 LoadInst *LI = Loads.pop_back_val();
2126 LI->replaceAllUsesWith(NewPN);
2127 Pass.DeadInsts.push_back(LI);
2128 } while (!Loads.empty());
2130 // Inject loads into all of the pred blocks.
2131 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) {
2132 BasicBlock *Pred = PN.getIncomingBlock(Idx);
2133 TerminatorInst *TI = Pred->getTerminator();
2134 Value *InVal = PN.getIncomingValue(Idx);
2135 IRBuilder<> PredBuilder(TI);
2137 // Map the value to the new alloca pointer if this was the old alloca
2139 bool ThisOperand = InVal == OldPtr;
2144 = PredBuilder.CreateLoad(InVal, getName(".sroa.speculate." +
2146 ++NumLoadsSpeculated;
2147 Load->setAlignment(Align);
2149 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag);
2150 NewPN->addIncoming(Load, Pred);
2154 Instruction *OtherPtr = dyn_cast<Instruction>(InVal);
2156 // No uses to rewrite.
2159 // Try to lookup and rewrite any partition uses corresponding to this phi
2161 AllocaPartitioning::iterator PI
2162 = P.findPartitionForPHIOrSelectOperand(PN, OtherPtr);
2163 if (PI != P.end()) {
2164 // If the other pointer is within the partitioning, replace the PHI in
2165 // its uses with the load we just speculated, or add another load for
2166 // it to rewrite if we've already replaced the PHI.
2167 AllocaPartitioning::use_iterator UI
2168 = P.findPartitionUseForPHIOrSelectOperand(PN, OtherPtr);
2169 if (isa<PHINode>(*UI->User))
2172 AllocaPartitioning::PartitionUse OtherUse = *UI;
2173 OtherUse.User = Load;
2174 P.use_insert(PI, std::upper_bound(UI, P.use_end(PI), OtherUse),
2179 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n");
2180 return NewPtr == &NewAI;
2183 /// Select instructions that use an alloca and are subsequently loaded can be
2184 /// rewritten to load both input pointers and then select between the result,
2185 /// allowing the load of the alloca to be promoted.
2187 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other
2188 /// %V = load i32* %P2
2190 /// %V1 = load i32* %Alloca -> will be mem2reg'd
2191 /// %V2 = load i32* %Other
2192 /// %V = select i1 %cond, i32 %V1, i32 %V2
2194 /// We can do this to a select if its only uses are loads and if the operand
2195 /// to the select can be loaded unconditionally.
2196 bool isSafeSelectToSpeculate(SelectInst &SI,
2197 SmallVectorImpl<LoadInst *> &Loads) {
2198 Value *TValue = SI.getTrueValue();
2199 Value *FValue = SI.getFalseValue();
2200 bool TDerefable = TValue->isDereferenceablePointer();
2201 bool FDerefable = FValue->isDereferenceablePointer();
2203 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end();
2205 LoadInst *LI = dyn_cast<LoadInst>(*UI);
2206 if (LI == 0 || !LI->isSimple()) return false;
2208 // Both operands to the select need to be dereferencable, either
2209 // absolutely (e.g. allocas) or at this point because we can see other
2211 if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI,
2212 LI->getAlignment(), &TD))
2214 if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI,
2215 LI->getAlignment(), &TD))
2217 Loads.push_back(LI);
2223 bool visitSelectInst(SelectInst &SI) {
2224 DEBUG(dbgs() << " original: " << SI << "\n");
2225 IRBuilder<> IRB(&SI);
2227 // Find the operand we need to rewrite here.
2228 bool IsTrueVal = SI.getTrueValue() == OldPtr;
2230 assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!");
2232 assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!");
2233 Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType());
2235 // If the select isn't safe to speculate, just use simple logic to emit it.
2236 SmallVector<LoadInst *, 4> Loads;
2237 if (!isSafeSelectToSpeculate(SI, Loads)) {
2238 SI.setOperand(IsTrueVal ? 1 : 2, NewPtr);
2239 DEBUG(dbgs() << " to: " << SI << "\n");
2240 deleteIfTriviallyDead(OldPtr);
2244 Value *OtherPtr = IsTrueVal ? SI.getFalseValue() : SI.getTrueValue();
2245 AllocaPartitioning::iterator PI
2246 = P.findPartitionForPHIOrSelectOperand(SI, OtherPtr);
2247 AllocaPartitioning::PartitionUse OtherUse;
2248 if (PI != P.end()) {
2249 // If the other pointer is within the partitioning, remove the select
2250 // from its uses. We'll add in the new loads below.
2251 AllocaPartitioning::use_iterator UI
2252 = P.findPartitionUseForPHIOrSelectOperand(SI, OtherPtr);
2254 P.use_erase(PI, UI);
2257 Value *TV = IsTrueVal ? NewPtr : SI.getTrueValue();
2258 Value *FV = IsTrueVal ? SI.getFalseValue() : NewPtr;
2259 // Replace the loads of the select with a select of two loads.
2260 while (!Loads.empty()) {
2261 LoadInst *LI = Loads.pop_back_val();
2263 IRB.SetInsertPoint(LI);
2265 IRB.CreateLoad(TV, getName("." + LI->getName() + ".true"));
2267 IRB.CreateLoad(FV, getName("." + LI->getName() + ".false"));
2268 NumLoadsSpeculated += 2;
2269 if (PI != P.end()) {
2270 LoadInst *OtherLoad = IsTrueVal ? FL : TL;
2271 assert(OtherUse.Ptr == OtherLoad->getOperand(0));
2272 OtherUse.User = OtherLoad;
2273 P.use_insert(PI, P.use_end(PI), OtherUse);
2276 // Transfer alignment and TBAA info if present.
2277 TL->setAlignment(LI->getAlignment());
2278 FL->setAlignment(LI->getAlignment());
2279 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) {
2280 TL->setMetadata(LLVMContext::MD_tbaa, Tag);
2281 FL->setMetadata(LLVMContext::MD_tbaa, Tag);
2284 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL);
2286 DEBUG(dbgs() << " speculated to: " << *V << "\n");
2287 LI->replaceAllUsesWith(V);
2288 Pass.DeadInsts.push_back(LI);
2291 std::stable_sort(P.use_begin(PI), P.use_end(PI));
2293 deleteIfTriviallyDead(OldPtr);
2294 return NewPtr == &NewAI;
2300 /// \brief Try to find a partition of the aggregate type passed in for a given
2301 /// offset and size.
2303 /// This recurses through the aggregate type and tries to compute a subtype
2304 /// based on the offset and size. When the offset and size span a sub-section
2305 /// of an array, it will even compute a new array type for that sub-section.
2306 static Type *getTypePartition(const TargetData &TD, Type *Ty,
2307 uint64_t Offset, uint64_t Size) {
2308 if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size)
2311 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) {
2312 // We can't partition pointers...
2313 if (SeqTy->isPointerTy())
2316 Type *ElementTy = SeqTy->getElementType();
2317 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
2318 uint64_t NumSkippedElements = Offset / ElementSize;
2319 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy))
2320 if (NumSkippedElements >= ArrTy->getNumElements())
2322 if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy))
2323 if (NumSkippedElements >= VecTy->getNumElements())
2325 Offset -= NumSkippedElements * ElementSize;
2327 // First check if we need to recurse.
2328 if (Offset > 0 || Size < ElementSize) {
2329 // Bail if the partition ends in a different array element.
2330 if ((Offset + Size) > ElementSize)
2332 // Recurse through the element type trying to peel off offset bytes.
2333 return getTypePartition(TD, ElementTy, Offset, Size);
2335 assert(Offset == 0);
2337 if (Size == ElementSize)
2339 assert(Size > ElementSize);
2340 uint64_t NumElements = Size / ElementSize;
2341 if (NumElements * ElementSize != Size)
2343 return ArrayType::get(ElementTy, NumElements);
2346 StructType *STy = dyn_cast<StructType>(Ty);
2350 const StructLayout *SL = TD.getStructLayout(STy);
2351 if (Offset > SL->getSizeInBytes())
2353 uint64_t EndOffset = Offset + Size;
2354 if (EndOffset > SL->getSizeInBytes())
2357 unsigned Index = SL->getElementContainingOffset(Offset);
2358 if (SL->getElementOffset(Index) != Offset)
2359 return 0; // Inside of padding.
2360 Offset -= SL->getElementOffset(Index);
2362 Type *ElementTy = STy->getElementType(Index);
2363 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy);
2364 if (Offset >= ElementSize)
2365 return 0; // The offset points into alignment padding.
2367 // See if any partition must be contained by the element.
2368 if (Offset > 0 || Size < ElementSize) {
2369 if ((Offset + Size) > ElementSize)
2371 // Bail if this is a poniter element, we can't recurse through them.
2372 if (ElementTy->isPointerTy())
2374 return getTypePartition(TD, ElementTy, Offset, Size);
2376 assert(Offset == 0);
2378 if (Size == ElementSize)
2381 StructType::element_iterator EI = STy->element_begin() + Index,
2382 EE = STy->element_end();
2383 if (EndOffset < SL->getSizeInBytes()) {
2384 unsigned EndIndex = SL->getElementContainingOffset(EndOffset);
2385 if (Index == EndIndex)
2386 return 0; // Within a single element and its padding.
2387 assert(Index < EndIndex);
2388 assert(Index + EndIndex <= STy->getNumElements());
2389 EE = STy->element_begin() + EndIndex;
2392 // Try to build up a sub-structure.
2393 SmallVector<Type *, 4> ElementTys;
2395 ElementTys.push_back(*EI++);
2397 StructType *SubTy = StructType::get(STy->getContext(), ElementTys,
2399 const StructLayout *SubSL = TD.getStructLayout(SubTy);
2400 if (Size == SubSL->getSizeInBytes())
2403 // FIXME: We could potentially recurse down through the last element in the
2404 // sub-struct to find a natural end point.
2408 /// \brief Rewrite an alloca partition's users.
2410 /// This routine drives both of the rewriting goals of the SROA pass. It tries
2411 /// to rewrite uses of an alloca partition to be conducive for SSA value
2412 /// promotion. If the partition needs a new, more refined alloca, this will
2413 /// build that new alloca, preserving as much type information as possible, and
2414 /// rewrite the uses of the old alloca to point at the new one and have the
2415 /// appropriate new offsets. It also evaluates how successful the rewrite was
2416 /// at enabling promotion and if it was successful queues the alloca to be
2418 bool SROA::rewriteAllocaPartition(AllocaInst &AI,
2419 AllocaPartitioning &P,
2420 AllocaPartitioning::iterator PI) {
2421 uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset;
2422 if (P.use_begin(PI) == P.use_end(PI))
2423 return false; // No live uses left of this partition.
2425 // Try to compute a friendly type for this partition of the alloca. This
2426 // won't always succeed, in which case we fall back to a legal integer type
2427 // or an i8 array of an appropriate size.
2429 if (Type *PartitionTy = P.getCommonType(PI))
2430 if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize)
2431 AllocaTy = PartitionTy;
2433 if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(),
2434 PI->BeginOffset, AllocaSize))
2435 AllocaTy = PartitionTy;
2437 (AllocaTy->isArrayTy() &&
2438 AllocaTy->getArrayElementType()->isIntegerTy())) &&
2439 TD->isLegalInteger(AllocaSize * 8))
2440 AllocaTy = Type::getIntNTy(*C, AllocaSize * 8);
2442 AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize);
2443 assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize);
2445 // Check for the case where we're going to rewrite to a new alloca of the
2446 // exact same type as the original, and with the same access offsets. In that
2447 // case, re-use the existing alloca, but still run through the rewriter to
2448 // performe phi and select speculation.
2450 if (AllocaTy == AI.getAllocatedType()) {
2451 assert(PI->BeginOffset == 0 &&
2452 "Non-zero begin offset but same alloca type");
2453 assert(PI == P.begin() && "Begin offset is zero on later partition");
2456 // FIXME: The alignment here is overly conservative -- we could in many
2457 // cases get away with much weaker alignment constraints.
2458 NewAI = new AllocaInst(AllocaTy, 0, AI.getAlignment(),
2459 AI.getName() + ".sroa." + Twine(PI - P.begin()),
2464 DEBUG(dbgs() << "Rewriting alloca partition "
2465 << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: "
2468 AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI,
2469 PI->BeginOffset, PI->EndOffset);
2470 DEBUG(dbgs() << " rewriting ");
2471 DEBUG(P.print(dbgs(), PI, ""));
2472 if (Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI))) {
2473 DEBUG(dbgs() << " and queuing for promotion\n");
2474 PromotableAllocas.push_back(NewAI);
2475 } else if (NewAI != &AI) {
2476 // If we can't promote the alloca, iterate on it to check for new
2477 // refinements exposed by splitting the current alloca. Don't iterate on an
2478 // alloca which didn't actually change and didn't get promoted.
2479 Worklist.insert(NewAI);
2484 /// \brief Walks the partitioning of an alloca rewriting uses of each partition.
2485 bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) {
2486 bool Changed = false;
2487 for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE;
2489 Changed |= rewriteAllocaPartition(AI, P, PI);
2494 /// \brief Analyze an alloca for SROA.
2496 /// This analyzes the alloca to ensure we can reason about it, builds
2497 /// a partitioning of the alloca, and then hands it off to be split and
2498 /// rewritten as needed.
2499 bool SROA::runOnAlloca(AllocaInst &AI) {
2500 DEBUG(dbgs() << "SROA alloca: " << AI << "\n");
2501 ++NumAllocasAnalyzed;
2503 // Special case dead allocas, as they're trivial.
2504 if (AI.use_empty()) {
2505 AI.eraseFromParent();
2509 // Skip alloca forms that this analysis can't handle.
2510 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() ||
2511 TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
2514 // First check if this is a non-aggregate type that we should simply promote.
2515 if (!AI.getAllocatedType()->isAggregateType() && isAllocaPromotable(&AI)) {
2516 DEBUG(dbgs() << " Trivially scalar type, queuing for promotion...\n");
2517 PromotableAllocas.push_back(&AI);
2521 // Build the partition set using a recursive instruction-visiting builder.
2522 AllocaPartitioning P(*TD, AI);
2523 DEBUG(P.print(dbgs()));
2527 // No partitions to split. Leave the dead alloca for a later pass to clean up.
2528 if (P.begin() == P.end())
2531 // Delete all the dead users of this alloca before splitting and rewriting it.
2532 bool Changed = false;
2533 for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(),
2534 DE = P.dead_user_end();
2537 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType()));
2538 DeadInsts.push_back(*DI);
2540 for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(),
2541 DE = P.dead_op_end();
2544 // Clobber the use with an undef value.
2545 **DO = UndefValue::get(OldV->getType());
2546 if (Instruction *OldI = dyn_cast<Instruction>(OldV))
2547 if (isInstructionTriviallyDead(OldI)) {
2549 DeadInsts.push_back(OldI);
2553 return splitAlloca(AI, P) || Changed;
2556 /// \brief Delete the dead instructions accumulated in this run.
2558 /// Recursively deletes the dead instructions we've accumulated. This is done
2559 /// at the very end to maximize locality of the recursive delete and to
2560 /// minimize the problems of invalidated instruction pointers as such pointers
2561 /// are used heavily in the intermediate stages of the algorithm.
2563 /// We also record the alloca instructions deleted here so that they aren't
2564 /// subsequently handed to mem2reg to promote.
2565 void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) {
2566 DeadSplitInsts.clear();
2567 while (!DeadInsts.empty()) {
2568 Instruction *I = DeadInsts.pop_back_val();
2569 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n");
2571 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI)
2572 if (Instruction *U = dyn_cast<Instruction>(*OI)) {
2573 // Zero out the operand and see if it becomes trivially dead.
2575 if (isInstructionTriviallyDead(U))
2576 DeadInsts.push_back(U);
2579 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
2580 DeletedAllocas.insert(AI);
2583 I->eraseFromParent();
2588 /// \brief A predicate to test whether an alloca belongs to a set.
2589 class IsAllocaInSet {
2590 typedef SmallPtrSet<AllocaInst *, 4> SetType;
2594 IsAllocaInSet(const SetType &Set) : Set(Set) {}
2595 bool operator()(AllocaInst *AI) { return Set.count(AI); }
2599 bool SROA::runOnFunction(Function &F) {
2600 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n");
2601 C = &F.getContext();
2602 TD = getAnalysisIfAvailable<TargetData>();
2604 DEBUG(dbgs() << " Skipping SROA -- no target data!\n");
2607 DT = &getAnalysis<DominatorTree>();
2609 BasicBlock &EntryBB = F.getEntryBlock();
2610 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end());
2612 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
2613 Worklist.insert(AI);
2615 bool Changed = false;
2616 // A set of deleted alloca instruction pointers which should be removed from
2617 // the list of promotable allocas.
2618 SmallPtrSet<AllocaInst *, 4> DeletedAllocas;
2620 while (!Worklist.empty()) {
2621 Changed |= runOnAlloca(*Worklist.pop_back_val());
2622 deleteDeadInstructions(DeletedAllocas);
2623 if (!DeletedAllocas.empty()) {
2624 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(),
2625 PromotableAllocas.end(),
2626 IsAllocaInSet(DeletedAllocas)),
2627 PromotableAllocas.end());
2628 DeletedAllocas.clear();
2632 if (!PromotableAllocas.empty()) {
2633 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n");
2634 PromoteMemToReg(PromotableAllocas, *DT);
2636 NumPromoted += PromotableAllocas.size();
2637 PromotableAllocas.clear();
2643 void SROA::getAnalysisUsage(AnalysisUsage &AU) const {
2644 AU.addRequired<DominatorTree>();
2645 AU.setPreservesCFG();