SROA.cpp revision 5e8da1773c8a16a9e3f08df444420f04d71ba673
1//===- SROA.cpp - Scalar Replacement Of Aggregates ------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9/// \file 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. 15/// 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. 19/// 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. 23/// 24//===----------------------------------------------------------------------===// 25 26#define DEBUG_TYPE "sroa" 27#include "llvm/Transforms/Scalar.h" 28#include "llvm/ADT/STLExtras.h" 29#include "llvm/ADT/SetVector.h" 30#include "llvm/ADT/SmallVector.h" 31#include "llvm/ADT/Statistic.h" 32#include "llvm/Analysis/Dominators.h" 33#include "llvm/Analysis/Loads.h" 34#include "llvm/Analysis/PtrUseVisitor.h" 35#include "llvm/Analysis/ValueTracking.h" 36#include "llvm/DIBuilder.h" 37#include "llvm/DebugInfo.h" 38#include "llvm/IR/Constants.h" 39#include "llvm/IR/DataLayout.h" 40#include "llvm/IR/DerivedTypes.h" 41#include "llvm/IR/Function.h" 42#include "llvm/IR/IRBuilder.h" 43#include "llvm/IR/Instructions.h" 44#include "llvm/IR/IntrinsicInst.h" 45#include "llvm/IR/LLVMContext.h" 46#include "llvm/IR/Operator.h" 47#include "llvm/InstVisitor.h" 48#include "llvm/Pass.h" 49#include "llvm/Support/CommandLine.h" 50#include "llvm/Support/Debug.h" 51#include "llvm/Support/ErrorHandling.h" 52#include "llvm/Support/MathExtras.h" 53#include "llvm/Support/raw_ostream.h" 54#include "llvm/Transforms/Utils/Local.h" 55#include "llvm/Transforms/Utils/PromoteMemToReg.h" 56#include "llvm/Transforms/Utils/SSAUpdater.h" 57using namespace llvm; 58 59STATISTIC(NumAllocasAnalyzed, "Number of allocas analyzed for replacement"); 60STATISTIC(NumNewAllocas, "Number of new, smaller allocas introduced"); 61STATISTIC(NumPromoted, "Number of allocas promoted to SSA values"); 62STATISTIC(NumLoadsSpeculated, "Number of loads speculated to allow promotion"); 63STATISTIC(NumDeleted, "Number of instructions deleted"); 64STATISTIC(NumVectorized, "Number of vectorized aggregates"); 65 66/// Hidden option to force the pass to not use DomTree and mem2reg, instead 67/// forming SSA values through the SSAUpdater infrastructure. 68static cl::opt<bool> 69ForceSSAUpdater("force-ssa-updater", cl::init(false), cl::Hidden); 70 71namespace { 72/// \brief A common base class for representing a half-open byte range. 73struct ByteRange { 74 /// \brief The beginning offset of the range. 75 uint64_t BeginOffset; 76 77 /// \brief The ending offset, not included in the range. 78 uint64_t EndOffset; 79 80 ByteRange() : BeginOffset(), EndOffset() {} 81 ByteRange(uint64_t BeginOffset, uint64_t EndOffset) 82 : BeginOffset(BeginOffset), EndOffset(EndOffset) {} 83 84 /// \brief Support for ordering ranges. 85 /// 86 /// This provides an ordering over ranges such that start offsets are 87 /// always increasing, and within equal start offsets, the end offsets are 88 /// decreasing. Thus the spanning range comes first in a cluster with the 89 /// same start position. 90 bool operator<(const ByteRange &RHS) const { 91 if (BeginOffset < RHS.BeginOffset) return true; 92 if (BeginOffset > RHS.BeginOffset) return false; 93 if (EndOffset > RHS.EndOffset) return true; 94 return false; 95 } 96 97 /// \brief Support comparison with a single offset to allow binary searches. 98 friend bool operator<(const ByteRange &LHS, uint64_t RHSOffset) { 99 return LHS.BeginOffset < RHSOffset; 100 } 101 102 friend LLVM_ATTRIBUTE_UNUSED bool operator<(uint64_t LHSOffset, 103 const ByteRange &RHS) { 104 return LHSOffset < RHS.BeginOffset; 105 } 106 107 bool operator==(const ByteRange &RHS) const { 108 return BeginOffset == RHS.BeginOffset && EndOffset == RHS.EndOffset; 109 } 110 bool operator!=(const ByteRange &RHS) const { return !operator==(RHS); } 111}; 112 113/// \brief A partition of an alloca. 114/// 115/// This structure represents a contiguous partition of the alloca. These are 116/// formed by examining the uses of the alloca. During formation, they may 117/// overlap but once an AllocaPartitioning is built, the Partitions within it 118/// are all disjoint. 119struct Partition : public ByteRange { 120 /// \brief Whether this partition is splittable into smaller partitions. 121 /// 122 /// We flag partitions as splittable when they are formed entirely due to 123 /// accesses by trivially splittable operations such as memset and memcpy. 124 bool IsSplittable; 125 126 /// \brief Test whether a partition has been marked as dead. 127 bool isDead() const { 128 if (BeginOffset == UINT64_MAX) { 129 assert(EndOffset == UINT64_MAX); 130 return true; 131 } 132 return false; 133 } 134 135 /// \brief Kill a partition. 136 /// This is accomplished by setting both its beginning and end offset to 137 /// the maximum possible value. 138 void kill() { 139 assert(!isDead() && "He's Dead, Jim!"); 140 BeginOffset = EndOffset = UINT64_MAX; 141 } 142 143 Partition() : ByteRange(), IsSplittable() {} 144 Partition(uint64_t BeginOffset, uint64_t EndOffset, bool IsSplittable) 145 : ByteRange(BeginOffset, EndOffset), IsSplittable(IsSplittable) {} 146}; 147 148/// \brief A particular use of a partition of the alloca. 149/// 150/// This structure is used to associate uses of a partition with it. They 151/// mark the range of bytes which are referenced by a particular instruction, 152/// and includes a handle to the user itself and the pointer value in use. 153/// The bounds of these uses are determined by intersecting the bounds of the 154/// memory use itself with a particular partition. As a consequence there is 155/// intentionally overlap between various uses of the same partition. 156class PartitionUse : public ByteRange { 157 /// \brief Combined storage for both the Use* and split state. 158 PointerIntPair<Use*, 1, bool> UsePtrAndIsSplit; 159 160public: 161 PartitionUse() : ByteRange(), UsePtrAndIsSplit() {} 162 PartitionUse(uint64_t BeginOffset, uint64_t EndOffset, Use *U, 163 bool IsSplit) 164 : ByteRange(BeginOffset, EndOffset), UsePtrAndIsSplit(U, IsSplit) {} 165 166 /// \brief The use in question. Provides access to both user and used value. 167 /// 168 /// Note that this may be null if the partition use is *dead*, that is, it 169 /// should be ignored. 170 Use *getUse() const { return UsePtrAndIsSplit.getPointer(); } 171 172 /// \brief Set the use for this partition use range. 173 void setUse(Use *U) { UsePtrAndIsSplit.setPointer(U); } 174 175 /// \brief Whether this use is split across multiple partitions. 176 bool isSplit() const { return UsePtrAndIsSplit.getInt(); } 177}; 178} 179 180namespace llvm { 181template <> struct isPodLike<Partition> : llvm::true_type {}; 182template <> struct isPodLike<PartitionUse> : llvm::true_type {}; 183} 184 185namespace { 186/// \brief Alloca partitioning representation. 187/// 188/// This class represents a partitioning of an alloca into slices, and 189/// information about the nature of uses of each slice of the alloca. The goal 190/// is that this information is sufficient to decide if and how to split the 191/// alloca apart and replace slices with scalars. It is also intended that this 192/// structure can capture the relevant information needed both to decide about 193/// and to enact these transformations. 194class AllocaPartitioning { 195public: 196 /// \brief Construct a partitioning of a particular alloca. 197 /// 198 /// Construction does most of the work for partitioning the alloca. This 199 /// performs the necessary walks of users and builds a partitioning from it. 200 AllocaPartitioning(const DataLayout &TD, AllocaInst &AI); 201 202 /// \brief Test whether a pointer to the allocation escapes our analysis. 203 /// 204 /// If this is true, the partitioning is never fully built and should be 205 /// ignored. 206 bool isEscaped() const { return PointerEscapingInstr; } 207 208 /// \brief Support for iterating over the partitions. 209 /// @{ 210 typedef SmallVectorImpl<Partition>::iterator iterator; 211 iterator begin() { return Partitions.begin(); } 212 iterator end() { return Partitions.end(); } 213 214 typedef SmallVectorImpl<Partition>::const_iterator const_iterator; 215 const_iterator begin() const { return Partitions.begin(); } 216 const_iterator end() const { return Partitions.end(); } 217 /// @} 218 219 /// \brief Support for iterating over and manipulating a particular 220 /// partition's uses. 221 /// 222 /// The iteration support provided for uses is more limited, but also 223 /// includes some manipulation routines to support rewriting the uses of 224 /// partitions during SROA. 225 /// @{ 226 typedef SmallVectorImpl<PartitionUse>::iterator use_iterator; 227 use_iterator use_begin(unsigned Idx) { return Uses[Idx].begin(); } 228 use_iterator use_begin(const_iterator I) { return Uses[I - begin()].begin(); } 229 use_iterator use_end(unsigned Idx) { return Uses[Idx].end(); } 230 use_iterator use_end(const_iterator I) { return Uses[I - begin()].end(); } 231 232 typedef SmallVectorImpl<PartitionUse>::const_iterator const_use_iterator; 233 const_use_iterator use_begin(unsigned Idx) const { return Uses[Idx].begin(); } 234 const_use_iterator use_begin(const_iterator I) const { 235 return Uses[I - begin()].begin(); 236 } 237 const_use_iterator use_end(unsigned Idx) const { return Uses[Idx].end(); } 238 const_use_iterator use_end(const_iterator I) const { 239 return Uses[I - begin()].end(); 240 } 241 242 unsigned use_size(unsigned Idx) const { return Uses[Idx].size(); } 243 unsigned use_size(const_iterator I) const { return Uses[I - begin()].size(); } 244 const PartitionUse &getUse(unsigned PIdx, unsigned UIdx) const { 245 return Uses[PIdx][UIdx]; 246 } 247 const PartitionUse &getUse(const_iterator I, unsigned UIdx) const { 248 return Uses[I - begin()][UIdx]; 249 } 250 251 void use_push_back(unsigned Idx, const PartitionUse &PU) { 252 Uses[Idx].push_back(PU); 253 } 254 void use_push_back(const_iterator I, const PartitionUse &PU) { 255 Uses[I - begin()].push_back(PU); 256 } 257 /// @} 258 259 /// \brief Allow iterating the dead users for this alloca. 260 /// 261 /// These are instructions which will never actually use the alloca as they 262 /// are outside the allocated range. They are safe to replace with undef and 263 /// delete. 264 /// @{ 265 typedef SmallVectorImpl<Instruction *>::const_iterator dead_user_iterator; 266 dead_user_iterator dead_user_begin() const { return DeadUsers.begin(); } 267 dead_user_iterator dead_user_end() const { return DeadUsers.end(); } 268 /// @} 269 270 /// \brief Allow iterating the dead expressions referring to this alloca. 271 /// 272 /// These are operands which have cannot actually be used to refer to the 273 /// alloca as they are outside its range and the user doesn't correct for 274 /// that. These mostly consist of PHI node inputs and the like which we just 275 /// need to replace with undef. 276 /// @{ 277 typedef SmallVectorImpl<Use *>::const_iterator dead_op_iterator; 278 dead_op_iterator dead_op_begin() const { return DeadOperands.begin(); } 279 dead_op_iterator dead_op_end() const { return DeadOperands.end(); } 280 /// @} 281 282 /// \brief MemTransferInst auxiliary data. 283 /// This struct provides some auxiliary data about memory transfer 284 /// intrinsics such as memcpy and memmove. These intrinsics can use two 285 /// different ranges within the same alloca, and provide other challenges to 286 /// correctly represent. We stash extra data to help us untangle this 287 /// after the partitioning is complete. 288 struct MemTransferOffsets { 289 /// The destination begin and end offsets when the destination is within 290 /// this alloca. If the end offset is zero the destination is not within 291 /// this alloca. 292 uint64_t DestBegin, DestEnd; 293 294 /// The source begin and end offsets when the source is within this alloca. 295 /// If the end offset is zero, the source is not within this alloca. 296 uint64_t SourceBegin, SourceEnd; 297 298 /// Flag for whether an alloca is splittable. 299 bool IsSplittable; 300 }; 301 MemTransferOffsets getMemTransferOffsets(MemTransferInst &II) const { 302 return MemTransferInstData.lookup(&II); 303 } 304 305 /// \brief Map from a PHI or select operand back to a partition. 306 /// 307 /// When manipulating PHI nodes or selects, they can use more than one 308 /// partition of an alloca. We store a special mapping to allow finding the 309 /// partition referenced by each of these operands, if any. 310 iterator findPartitionForPHIOrSelectOperand(Use *U) { 311 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt 312 = PHIOrSelectOpMap.find(U); 313 if (MapIt == PHIOrSelectOpMap.end()) 314 return end(); 315 316 return begin() + MapIt->second.first; 317 } 318 319 /// \brief Map from a PHI or select operand back to the specific use of 320 /// a partition. 321 /// 322 /// Similar to mapping these operands back to the partitions, this maps 323 /// directly to the use structure of that partition. 324 use_iterator findPartitionUseForPHIOrSelectOperand(Use *U) { 325 SmallDenseMap<Use *, std::pair<unsigned, unsigned> >::const_iterator MapIt 326 = PHIOrSelectOpMap.find(U); 327 assert(MapIt != PHIOrSelectOpMap.end()); 328 return Uses[MapIt->second.first].begin() + MapIt->second.second; 329 } 330 331 /// \brief Compute a common type among the uses of a particular partition. 332 /// 333 /// This routines walks all of the uses of a particular partition and tries 334 /// to find a common type between them. Untyped operations such as memset and 335 /// memcpy are ignored. 336 Type *getCommonType(iterator I) const; 337 338#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 339 void print(raw_ostream &OS, const_iterator I, StringRef Indent = " ") const; 340 void printUsers(raw_ostream &OS, const_iterator I, 341 StringRef Indent = " ") const; 342 void print(raw_ostream &OS) const; 343 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump(const_iterator I) const; 344 void LLVM_ATTRIBUTE_NOINLINE LLVM_ATTRIBUTE_USED dump() const; 345#endif 346 347private: 348 template <typename DerivedT, typename RetT = void> class BuilderBase; 349 class PartitionBuilder; 350 friend class AllocaPartitioning::PartitionBuilder; 351 class UseBuilder; 352 friend class AllocaPartitioning::UseBuilder; 353 354#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 355 /// \brief Handle to alloca instruction to simplify method interfaces. 356 AllocaInst &AI; 357#endif 358 359 /// \brief The instruction responsible for this alloca having no partitioning. 360 /// 361 /// When an instruction (potentially) escapes the pointer to the alloca, we 362 /// store a pointer to that here and abort trying to partition the alloca. 363 /// This will be null if the alloca is partitioned successfully. 364 Instruction *PointerEscapingInstr; 365 366 /// \brief The partitions of the alloca. 367 /// 368 /// We store a vector of the partitions over the alloca here. This vector is 369 /// sorted by increasing begin offset, and then by decreasing end offset. See 370 /// the Partition inner class for more details. Initially (during 371 /// construction) there are overlaps, but we form a disjoint sequence of 372 /// partitions while finishing construction and a fully constructed object is 373 /// expected to always have this as a disjoint space. 374 SmallVector<Partition, 8> Partitions; 375 376 /// \brief The uses of the partitions. 377 /// 378 /// This is essentially a mapping from each partition to a list of uses of 379 /// that partition. The mapping is done with a Uses vector that has the exact 380 /// same number of entries as the partition vector. Each entry is itself 381 /// a vector of the uses. 382 SmallVector<SmallVector<PartitionUse, 2>, 8> Uses; 383 384 /// \brief Instructions which will become dead if we rewrite the alloca. 385 /// 386 /// Note that these are not separated by partition. This is because we expect 387 /// a partitioned alloca to be completely rewritten or not rewritten at all. 388 /// If rewritten, all these instructions can simply be removed and replaced 389 /// with undef as they come from outside of the allocated space. 390 SmallVector<Instruction *, 8> DeadUsers; 391 392 /// \brief Operands which will become dead if we rewrite the alloca. 393 /// 394 /// These are operands that in their particular use can be replaced with 395 /// undef when we rewrite the alloca. These show up in out-of-bounds inputs 396 /// to PHI nodes and the like. They aren't entirely dead (there might be 397 /// a GEP back into the bounds using it elsewhere) and nor is the PHI, but we 398 /// want to swap this particular input for undef to simplify the use lists of 399 /// the alloca. 400 SmallVector<Use *, 8> DeadOperands; 401 402 /// \brief The underlying storage for auxiliary memcpy and memset info. 403 SmallDenseMap<MemTransferInst *, MemTransferOffsets, 4> MemTransferInstData; 404 405 /// \brief A side datastructure used when building up the partitions and uses. 406 /// 407 /// This mapping is only really used during the initial building of the 408 /// partitioning so that we can retain information about PHI and select nodes 409 /// processed. 410 SmallDenseMap<Instruction *, std::pair<uint64_t, bool> > PHIOrSelectSizes; 411 412 /// \brief Auxiliary information for particular PHI or select operands. 413 SmallDenseMap<Use *, std::pair<unsigned, unsigned>, 4> PHIOrSelectOpMap; 414 415 /// \brief A utility routine called from the constructor. 416 /// 417 /// This does what it says on the tin. It is the key of the alloca partition 418 /// splitting and merging. After it is called we have the desired disjoint 419 /// collection of partitions. 420 void splitAndMergePartitions(); 421}; 422} 423 424static Value *foldSelectInst(SelectInst &SI) { 425 // If the condition being selected on is a constant or the same value is 426 // being selected between, fold the select. Yes this does (rarely) happen 427 // early on. 428 if (ConstantInt *CI = dyn_cast<ConstantInt>(SI.getCondition())) 429 return SI.getOperand(1+CI->isZero()); 430 if (SI.getOperand(1) == SI.getOperand(2)) 431 return SI.getOperand(1); 432 433 return 0; 434} 435 436/// \brief Builder for the alloca partitioning. 437/// 438/// This class builds an alloca partitioning by recursively visiting the uses 439/// of an alloca and splitting the partitions for each load and store at each 440/// offset. 441class AllocaPartitioning::PartitionBuilder 442 : public PtrUseVisitor<PartitionBuilder> { 443 friend class PtrUseVisitor<PartitionBuilder>; 444 friend class InstVisitor<PartitionBuilder>; 445 typedef PtrUseVisitor<PartitionBuilder> Base; 446 447 const uint64_t AllocSize; 448 AllocaPartitioning &P; 449 450 SmallDenseMap<Instruction *, unsigned> MemTransferPartitionMap; 451 452public: 453 PartitionBuilder(const DataLayout &DL, AllocaInst &AI, AllocaPartitioning &P) 454 : PtrUseVisitor<PartitionBuilder>(DL), 455 AllocSize(DL.getTypeAllocSize(AI.getAllocatedType())), 456 P(P) {} 457 458private: 459 void insertUse(Instruction &I, const APInt &Offset, uint64_t Size, 460 bool IsSplittable = false) { 461 // Completely skip uses which have a zero size or start either before or 462 // past the end of the allocation. 463 if (Size == 0 || Offset.isNegative() || Offset.uge(AllocSize)) { 464 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte use @" << Offset 465 << " which has zero size or starts outside of the " 466 << AllocSize << " byte alloca:\n" 467 << " alloca: " << P.AI << "\n" 468 << " use: " << I << "\n"); 469 return; 470 } 471 472 uint64_t BeginOffset = Offset.getZExtValue(); 473 uint64_t EndOffset = BeginOffset + Size; 474 475 // Clamp the end offset to the end of the allocation. Note that this is 476 // formulated to handle even the case where "BeginOffset + Size" overflows. 477 // This may appear superficially to be something we could ignore entirely, 478 // but that is not so! There may be widened loads or PHI-node uses where 479 // some instructions are dead but not others. We can't completely ignore 480 // them, and so have to record at least the information here. 481 assert(AllocSize >= BeginOffset); // Established above. 482 if (Size > AllocSize - BeginOffset) { 483 DEBUG(dbgs() << "WARNING: Clamping a " << Size << " byte use @" << Offset 484 << " to remain within the " << AllocSize << " byte alloca:\n" 485 << " alloca: " << P.AI << "\n" 486 << " use: " << I << "\n"); 487 EndOffset = AllocSize; 488 } 489 490 Partition New(BeginOffset, EndOffset, IsSplittable); 491 P.Partitions.push_back(New); 492 } 493 494 void handleLoadOrStore(Type *Ty, Instruction &I, const APInt &Offset, 495 uint64_t Size, bool IsVolatile) { 496 // We allow splitting of loads and stores where the type is an integer type 497 // and cover the entire alloca. This prevents us from splitting over 498 // eagerly. 499 // FIXME: In the great blue eventually, we should eagerly split all integer 500 // loads and stores, and then have a separate step that merges adjacent 501 // alloca partitions into a single partition suitable for integer widening. 502 // Or we should skip the merge step and rely on GVN and other passes to 503 // merge adjacent loads and stores that survive mem2reg. 504 bool IsSplittable = 505 Ty->isIntegerTy() && !IsVolatile && Offset == 0 && Size >= AllocSize; 506 507 insertUse(I, Offset, Size, IsSplittable); 508 } 509 510 void visitLoadInst(LoadInst &LI) { 511 assert((!LI.isSimple() || LI.getType()->isSingleValueType()) && 512 "All simple FCA loads should have been pre-split"); 513 514 if (!IsOffsetKnown) 515 return PI.setAborted(&LI); 516 517 uint64_t Size = DL.getTypeStoreSize(LI.getType()); 518 return handleLoadOrStore(LI.getType(), LI, Offset, Size, LI.isVolatile()); 519 } 520 521 void visitStoreInst(StoreInst &SI) { 522 Value *ValOp = SI.getValueOperand(); 523 if (ValOp == *U) 524 return PI.setEscapedAndAborted(&SI); 525 if (!IsOffsetKnown) 526 return PI.setAborted(&SI); 527 528 uint64_t Size = DL.getTypeStoreSize(ValOp->getType()); 529 530 // If this memory access can be shown to *statically* extend outside the 531 // bounds of of the allocation, it's behavior is undefined, so simply 532 // ignore it. Note that this is more strict than the generic clamping 533 // behavior of insertUse. We also try to handle cases which might run the 534 // risk of overflow. 535 // FIXME: We should instead consider the pointer to have escaped if this 536 // function is being instrumented for addressing bugs or race conditions. 537 if (Offset.isNegative() || Size > AllocSize || 538 Offset.ugt(AllocSize - Size)) { 539 DEBUG(dbgs() << "WARNING: Ignoring " << Size << " byte store @" << Offset 540 << " which extends past the end of the " << AllocSize 541 << " byte alloca:\n" 542 << " alloca: " << P.AI << "\n" 543 << " use: " << SI << "\n"); 544 return; 545 } 546 547 assert((!SI.isSimple() || ValOp->getType()->isSingleValueType()) && 548 "All simple FCA stores should have been pre-split"); 549 handleLoadOrStore(ValOp->getType(), SI, Offset, Size, SI.isVolatile()); 550 } 551 552 553 void visitMemSetInst(MemSetInst &II) { 554 assert(II.getRawDest() == *U && "Pointer use is not the destination?"); 555 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength()); 556 if ((Length && Length->getValue() == 0) || 557 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize))) 558 // Zero-length mem transfer intrinsics can be ignored entirely. 559 return; 560 561 if (!IsOffsetKnown) 562 return PI.setAborted(&II); 563 564 insertUse(II, Offset, 565 Length ? Length->getLimitedValue() 566 : AllocSize - Offset.getLimitedValue(), 567 (bool)Length); 568 } 569 570 void visitMemTransferInst(MemTransferInst &II) { 571 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength()); 572 if ((Length && Length->getValue() == 0) || 573 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize))) 574 // Zero-length mem transfer intrinsics can be ignored entirely. 575 return; 576 577 if (!IsOffsetKnown) 578 return PI.setAborted(&II); 579 580 uint64_t RawOffset = Offset.getLimitedValue(); 581 uint64_t Size = Length ? Length->getLimitedValue() 582 : AllocSize - RawOffset; 583 584 MemTransferOffsets &Offsets = P.MemTransferInstData[&II]; 585 586 // Only intrinsics with a constant length can be split. 587 Offsets.IsSplittable = Length; 588 589 if (*U == II.getRawDest()) { 590 Offsets.DestBegin = RawOffset; 591 Offsets.DestEnd = RawOffset + Size; 592 } 593 if (*U == II.getRawSource()) { 594 Offsets.SourceBegin = RawOffset; 595 Offsets.SourceEnd = RawOffset + Size; 596 } 597 598 // If we have set up end offsets for both the source and the destination, 599 // we have found both sides of this transfer pointing at the same alloca. 600 bool SeenBothEnds = Offsets.SourceEnd && Offsets.DestEnd; 601 if (SeenBothEnds && II.getRawDest() != II.getRawSource()) { 602 unsigned PrevIdx = MemTransferPartitionMap[&II]; 603 604 // Check if the begin offsets match and this is a non-volatile transfer. 605 // In that case, we can completely elide the transfer. 606 if (!II.isVolatile() && Offsets.SourceBegin == Offsets.DestBegin) { 607 P.Partitions[PrevIdx].kill(); 608 return; 609 } 610 611 // Otherwise we have an offset transfer within the same alloca. We can't 612 // split those. 613 P.Partitions[PrevIdx].IsSplittable = Offsets.IsSplittable = false; 614 } else if (SeenBothEnds) { 615 // Handle the case where this exact use provides both ends of the 616 // operation. 617 assert(II.getRawDest() == II.getRawSource()); 618 619 // For non-volatile transfers this is a no-op. 620 if (!II.isVolatile()) 621 return; 622 623 // Otherwise just suppress splitting. 624 Offsets.IsSplittable = false; 625 } 626 627 628 // Insert the use now that we've fixed up the splittable nature. 629 insertUse(II, Offset, Size, Offsets.IsSplittable); 630 631 // Setup the mapping from intrinsic to partition of we've not seen both 632 // ends of this transfer. 633 if (!SeenBothEnds) { 634 unsigned NewIdx = P.Partitions.size() - 1; 635 bool Inserted 636 = MemTransferPartitionMap.insert(std::make_pair(&II, NewIdx)).second; 637 assert(Inserted && 638 "Already have intrinsic in map but haven't seen both ends"); 639 (void)Inserted; 640 } 641 } 642 643 // Disable SRoA for any intrinsics except for lifetime invariants. 644 // FIXME: What about debug intrinsics? This matches old behavior, but 645 // doesn't make sense. 646 void visitIntrinsicInst(IntrinsicInst &II) { 647 if (!IsOffsetKnown) 648 return PI.setAborted(&II); 649 650 if (II.getIntrinsicID() == Intrinsic::lifetime_start || 651 II.getIntrinsicID() == Intrinsic::lifetime_end) { 652 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0)); 653 uint64_t Size = std::min(AllocSize - Offset.getLimitedValue(), 654 Length->getLimitedValue()); 655 insertUse(II, Offset, Size, true); 656 return; 657 } 658 659 Base::visitIntrinsicInst(II); 660 } 661 662 Instruction *hasUnsafePHIOrSelectUse(Instruction *Root, uint64_t &Size) { 663 // We consider any PHI or select that results in a direct load or store of 664 // the same offset to be a viable use for partitioning purposes. These uses 665 // are considered unsplittable and the size is the maximum loaded or stored 666 // size. 667 SmallPtrSet<Instruction *, 4> Visited; 668 SmallVector<std::pair<Instruction *, Instruction *>, 4> Uses; 669 Visited.insert(Root); 670 Uses.push_back(std::make_pair(cast<Instruction>(*U), Root)); 671 // If there are no loads or stores, the access is dead. We mark that as 672 // a size zero access. 673 Size = 0; 674 do { 675 Instruction *I, *UsedI; 676 llvm::tie(UsedI, I) = Uses.pop_back_val(); 677 678 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 679 Size = std::max(Size, DL.getTypeStoreSize(LI->getType())); 680 continue; 681 } 682 if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 683 Value *Op = SI->getOperand(0); 684 if (Op == UsedI) 685 return SI; 686 Size = std::max(Size, DL.getTypeStoreSize(Op->getType())); 687 continue; 688 } 689 690 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) { 691 if (!GEP->hasAllZeroIndices()) 692 return GEP; 693 } else if (!isa<BitCastInst>(I) && !isa<PHINode>(I) && 694 !isa<SelectInst>(I)) { 695 return I; 696 } 697 698 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); UI != UE; 699 ++UI) 700 if (Visited.insert(cast<Instruction>(*UI))) 701 Uses.push_back(std::make_pair(I, cast<Instruction>(*UI))); 702 } while (!Uses.empty()); 703 704 return 0; 705 } 706 707 void visitPHINode(PHINode &PN) { 708 if (PN.use_empty()) 709 return; 710 if (!IsOffsetKnown) 711 return PI.setAborted(&PN); 712 713 // See if we already have computed info on this node. 714 std::pair<uint64_t, bool> &PHIInfo = P.PHIOrSelectSizes[&PN]; 715 if (PHIInfo.first) { 716 PHIInfo.second = true; 717 insertUse(PN, Offset, PHIInfo.first); 718 return; 719 } 720 721 // Check for an unsafe use of the PHI node. 722 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&PN, PHIInfo.first)) 723 return PI.setAborted(UnsafeI); 724 725 insertUse(PN, Offset, PHIInfo.first); 726 } 727 728 void visitSelectInst(SelectInst &SI) { 729 if (SI.use_empty()) 730 return; 731 if (Value *Result = foldSelectInst(SI)) { 732 if (Result == *U) 733 // If the result of the constant fold will be the pointer, recurse 734 // through the select as if we had RAUW'ed it. 735 enqueueUsers(SI); 736 737 return; 738 } 739 if (!IsOffsetKnown) 740 return PI.setAborted(&SI); 741 742 // See if we already have computed info on this node. 743 std::pair<uint64_t, bool> &SelectInfo = P.PHIOrSelectSizes[&SI]; 744 if (SelectInfo.first) { 745 SelectInfo.second = true; 746 insertUse(SI, Offset, SelectInfo.first); 747 return; 748 } 749 750 // Check for an unsafe use of the PHI node. 751 if (Instruction *UnsafeI = hasUnsafePHIOrSelectUse(&SI, SelectInfo.first)) 752 return PI.setAborted(UnsafeI); 753 754 insertUse(SI, Offset, SelectInfo.first); 755 } 756 757 /// \brief Disable SROA entirely if there are unhandled users of the alloca. 758 void visitInstruction(Instruction &I) { 759 PI.setAborted(&I); 760 } 761}; 762 763/// \brief Use adder for the alloca partitioning. 764/// 765/// This class adds the uses of an alloca to all of the partitions which they 766/// use. For splittable partitions, this can end up doing essentially a linear 767/// walk of the partitions, but the number of steps remains bounded by the 768/// total result instruction size: 769/// - The number of partitions is a result of the number unsplittable 770/// instructions using the alloca. 771/// - The number of users of each partition is at worst the total number of 772/// splittable instructions using the alloca. 773/// Thus we will produce N * M instructions in the end, where N are the number 774/// of unsplittable uses and M are the number of splittable. This visitor does 775/// the exact same number of updates to the partitioning. 776/// 777/// In the more common case, this visitor will leverage the fact that the 778/// partition space is pre-sorted, and do a logarithmic search for the 779/// partition needed, making the total visit a classical ((N + M) * log(N)) 780/// complexity operation. 781class AllocaPartitioning::UseBuilder : public PtrUseVisitor<UseBuilder> { 782 friend class PtrUseVisitor<UseBuilder>; 783 friend class InstVisitor<UseBuilder>; 784 typedef PtrUseVisitor<UseBuilder> Base; 785 786 const uint64_t AllocSize; 787 AllocaPartitioning &P; 788 789 /// \brief Set to de-duplicate dead instructions found in the use walk. 790 SmallPtrSet<Instruction *, 4> VisitedDeadInsts; 791 792public: 793 UseBuilder(const DataLayout &TD, AllocaInst &AI, AllocaPartitioning &P) 794 : PtrUseVisitor<UseBuilder>(TD), 795 AllocSize(TD.getTypeAllocSize(AI.getAllocatedType())), 796 P(P) {} 797 798private: 799 void markAsDead(Instruction &I) { 800 if (VisitedDeadInsts.insert(&I)) 801 P.DeadUsers.push_back(&I); 802 } 803 804 void insertUse(Instruction &User, const APInt &Offset, uint64_t Size) { 805 // If the use has a zero size or extends outside of the allocation, record 806 // it as a dead use for elimination later. 807 if (Size == 0 || Offset.isNegative() || Offset.uge(AllocSize)) 808 return markAsDead(User); 809 810 uint64_t BeginOffset = Offset.getZExtValue(); 811 uint64_t EndOffset = BeginOffset + Size; 812 813 // Clamp the end offset to the end of the allocation. Note that this is 814 // formulated to handle even the case where "BeginOffset + Size" overflows. 815 assert(AllocSize >= BeginOffset); // Established above. 816 if (Size > AllocSize - BeginOffset) 817 EndOffset = AllocSize; 818 819 // NB: This only works if we have zero overlapping partitions. 820 iterator I = std::lower_bound(P.begin(), P.end(), BeginOffset); 821 if (I != P.begin() && llvm::prior(I)->EndOffset > BeginOffset) 822 I = llvm::prior(I); 823 iterator E = P.end(); 824 bool IsSplit = llvm::next(I) != E && llvm::next(I)->BeginOffset < EndOffset; 825 for (; I != E && I->BeginOffset < EndOffset; ++I) { 826 PartitionUse NewPU(std::max(I->BeginOffset, BeginOffset), 827 std::min(I->EndOffset, EndOffset), U, IsSplit); 828 P.use_push_back(I, NewPU); 829 if (isa<PHINode>(U->getUser()) || isa<SelectInst>(U->getUser())) 830 P.PHIOrSelectOpMap[U] 831 = std::make_pair(I - P.begin(), P.Uses[I - P.begin()].size() - 1); 832 } 833 } 834 835 void visitBitCastInst(BitCastInst &BC) { 836 if (BC.use_empty()) 837 return markAsDead(BC); 838 839 return Base::visitBitCastInst(BC); 840 } 841 842 void visitGetElementPtrInst(GetElementPtrInst &GEPI) { 843 if (GEPI.use_empty()) 844 return markAsDead(GEPI); 845 846 return Base::visitGetElementPtrInst(GEPI); 847 } 848 849 void visitLoadInst(LoadInst &LI) { 850 assert(IsOffsetKnown); 851 uint64_t Size = DL.getTypeStoreSize(LI.getType()); 852 insertUse(LI, Offset, Size); 853 } 854 855 void visitStoreInst(StoreInst &SI) { 856 assert(IsOffsetKnown); 857 uint64_t Size = DL.getTypeStoreSize(SI.getOperand(0)->getType()); 858 859 // If this memory access can be shown to *statically* extend outside the 860 // bounds of of the allocation, it's behavior is undefined, so simply 861 // ignore it. Note that this is more strict than the generic clamping 862 // behavior of insertUse. 863 if (Offset.isNegative() || Size > AllocSize || 864 Offset.ugt(AllocSize - Size)) 865 return markAsDead(SI); 866 867 insertUse(SI, Offset, Size); 868 } 869 870 void visitMemSetInst(MemSetInst &II) { 871 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength()); 872 if ((Length && Length->getValue() == 0) || 873 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize))) 874 return markAsDead(II); 875 876 assert(IsOffsetKnown); 877 insertUse(II, Offset, Length ? Length->getLimitedValue() 878 : AllocSize - Offset.getLimitedValue()); 879 } 880 881 void visitMemTransferInst(MemTransferInst &II) { 882 ConstantInt *Length = dyn_cast<ConstantInt>(II.getLength()); 883 if ((Length && Length->getValue() == 0) || 884 (IsOffsetKnown && !Offset.isNegative() && Offset.uge(AllocSize))) 885 return markAsDead(II); 886 887 assert(IsOffsetKnown); 888 uint64_t Size = Length ? Length->getLimitedValue() 889 : AllocSize - Offset.getLimitedValue(); 890 891 MemTransferOffsets &Offsets = P.MemTransferInstData[&II]; 892 if (!II.isVolatile() && Offsets.DestEnd && Offsets.SourceEnd && 893 Offsets.DestBegin == Offsets.SourceBegin) 894 return markAsDead(II); // Skip identity transfers without side-effects. 895 896 insertUse(II, Offset, Size); 897 } 898 899 void visitIntrinsicInst(IntrinsicInst &II) { 900 assert(IsOffsetKnown); 901 assert(II.getIntrinsicID() == Intrinsic::lifetime_start || 902 II.getIntrinsicID() == Intrinsic::lifetime_end); 903 904 ConstantInt *Length = cast<ConstantInt>(II.getArgOperand(0)); 905 insertUse(II, Offset, std::min(Length->getLimitedValue(), 906 AllocSize - Offset.getLimitedValue())); 907 } 908 909 void insertPHIOrSelect(Instruction &User, const APInt &Offset) { 910 uint64_t Size = P.PHIOrSelectSizes.lookup(&User).first; 911 912 // For PHI and select operands outside the alloca, we can't nuke the entire 913 // phi or select -- the other side might still be relevant, so we special 914 // case them here and use a separate structure to track the operands 915 // themselves which should be replaced with undef. 916 if ((Offset.isNegative() && Offset.uge(Size)) || 917 (!Offset.isNegative() && Offset.uge(AllocSize))) { 918 P.DeadOperands.push_back(U); 919 return; 920 } 921 922 insertUse(User, Offset, Size); 923 } 924 925 void visitPHINode(PHINode &PN) { 926 if (PN.use_empty()) 927 return markAsDead(PN); 928 929 assert(IsOffsetKnown); 930 insertPHIOrSelect(PN, Offset); 931 } 932 933 void visitSelectInst(SelectInst &SI) { 934 if (SI.use_empty()) 935 return markAsDead(SI); 936 937 if (Value *Result = foldSelectInst(SI)) { 938 if (Result == *U) 939 // If the result of the constant fold will be the pointer, recurse 940 // through the select as if we had RAUW'ed it. 941 enqueueUsers(SI); 942 else 943 // Otherwise the operand to the select is dead, and we can replace it 944 // with undef. 945 P.DeadOperands.push_back(U); 946 947 return; 948 } 949 950 assert(IsOffsetKnown); 951 insertPHIOrSelect(SI, Offset); 952 } 953 954 /// \brief Unreachable, we've already visited the alloca once. 955 void visitInstruction(Instruction &I) { 956 llvm_unreachable("Unhandled instruction in use builder."); 957 } 958}; 959 960void AllocaPartitioning::splitAndMergePartitions() { 961 size_t NumDeadPartitions = 0; 962 963 // Track the range of splittable partitions that we pass when accumulating 964 // overlapping unsplittable partitions. 965 uint64_t SplitEndOffset = 0ull; 966 967 Partition New(0ull, 0ull, false); 968 969 for (unsigned i = 0, j = i, e = Partitions.size(); i != e; i = j) { 970 ++j; 971 972 if (!Partitions[i].IsSplittable || New.BeginOffset == New.EndOffset) { 973 assert(New.BeginOffset == New.EndOffset); 974 New = Partitions[i]; 975 } else { 976 assert(New.IsSplittable); 977 New.EndOffset = std::max(New.EndOffset, Partitions[i].EndOffset); 978 } 979 assert(New.BeginOffset != New.EndOffset); 980 981 // Scan the overlapping partitions. 982 while (j != e && New.EndOffset > Partitions[j].BeginOffset) { 983 // If the new partition we are forming is splittable, stop at the first 984 // unsplittable partition. 985 if (New.IsSplittable && !Partitions[j].IsSplittable) 986 break; 987 988 // Grow the new partition to include any equally splittable range. 'j' is 989 // always equally splittable when New is splittable, but when New is not 990 // splittable, we may subsume some (or part of some) splitable partition 991 // without growing the new one. 992 if (New.IsSplittable == Partitions[j].IsSplittable) { 993 New.EndOffset = std::max(New.EndOffset, Partitions[j].EndOffset); 994 } else { 995 assert(!New.IsSplittable); 996 assert(Partitions[j].IsSplittable); 997 SplitEndOffset = std::max(SplitEndOffset, Partitions[j].EndOffset); 998 } 999 1000 Partitions[j].kill(); 1001 ++NumDeadPartitions; 1002 ++j; 1003 } 1004 1005 // If the new partition is splittable, chop off the end as soon as the 1006 // unsplittable subsequent partition starts and ensure we eventually cover 1007 // the splittable area. 1008 if (j != e && New.IsSplittable) { 1009 SplitEndOffset = std::max(SplitEndOffset, New.EndOffset); 1010 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset); 1011 } 1012 1013 // Add the new partition if it differs from the original one and is 1014 // non-empty. We can end up with an empty partition here if it was 1015 // splittable but there is an unsplittable one that starts at the same 1016 // offset. 1017 if (New != Partitions[i]) { 1018 if (New.BeginOffset != New.EndOffset) 1019 Partitions.push_back(New); 1020 // Mark the old one for removal. 1021 Partitions[i].kill(); 1022 ++NumDeadPartitions; 1023 } 1024 1025 New.BeginOffset = New.EndOffset; 1026 if (!New.IsSplittable) { 1027 New.EndOffset = std::max(New.EndOffset, SplitEndOffset); 1028 if (j != e && !Partitions[j].IsSplittable) 1029 New.EndOffset = std::min(New.EndOffset, Partitions[j].BeginOffset); 1030 New.IsSplittable = true; 1031 // If there is a trailing splittable partition which won't be fused into 1032 // the next splittable partition go ahead and add it onto the partitions 1033 // list. 1034 if (New.BeginOffset < New.EndOffset && 1035 (j == e || !Partitions[j].IsSplittable || 1036 New.EndOffset < Partitions[j].BeginOffset)) { 1037 Partitions.push_back(New); 1038 New.BeginOffset = New.EndOffset = 0ull; 1039 } 1040 } 1041 } 1042 1043 // Re-sort the partitions now that they have been split and merged into 1044 // disjoint set of partitions. Also remove any of the dead partitions we've 1045 // replaced in the process. 1046 std::sort(Partitions.begin(), Partitions.end()); 1047 if (NumDeadPartitions) { 1048 assert(Partitions.back().isDead()); 1049 assert((ptrdiff_t)NumDeadPartitions == 1050 std::count(Partitions.begin(), Partitions.end(), Partitions.back())); 1051 } 1052 Partitions.erase(Partitions.end() - NumDeadPartitions, Partitions.end()); 1053} 1054 1055AllocaPartitioning::AllocaPartitioning(const DataLayout &TD, AllocaInst &AI) 1056 : 1057#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 1058 AI(AI), 1059#endif 1060 PointerEscapingInstr(0) { 1061 PartitionBuilder PB(TD, AI, *this); 1062 PartitionBuilder::PtrInfo PtrI = PB.visitPtr(AI); 1063 if (PtrI.isEscaped() || PtrI.isAborted()) { 1064 // FIXME: We should sink the escape vs. abort info into the caller nicely, 1065 // possibly by just storing the PtrInfo in the AllocaPartitioning. 1066 PointerEscapingInstr = PtrI.getEscapingInst() ? PtrI.getEscapingInst() 1067 : PtrI.getAbortingInst(); 1068 assert(PointerEscapingInstr && "Did not track a bad instruction"); 1069 return; 1070 } 1071 1072 // Sort the uses. This arranges for the offsets to be in ascending order, 1073 // and the sizes to be in descending order. 1074 std::sort(Partitions.begin(), Partitions.end()); 1075 1076 // Remove any partitions from the back which are marked as dead. 1077 while (!Partitions.empty() && Partitions.back().isDead()) 1078 Partitions.pop_back(); 1079 1080 if (Partitions.size() > 1) { 1081 // Intersect splittability for all partitions with equal offsets and sizes. 1082 // Then remove all but the first so that we have a sequence of non-equal but 1083 // potentially overlapping partitions. 1084 for (iterator I = Partitions.begin(), J = I, E = Partitions.end(); I != E; 1085 I = J) { 1086 ++J; 1087 while (J != E && *I == *J) { 1088 I->IsSplittable &= J->IsSplittable; 1089 ++J; 1090 } 1091 } 1092 Partitions.erase(std::unique(Partitions.begin(), Partitions.end()), 1093 Partitions.end()); 1094 1095 // Split splittable and merge unsplittable partitions into a disjoint set 1096 // of partitions over the used space of the allocation. 1097 splitAndMergePartitions(); 1098 } 1099 1100 // Now build up the user lists for each of these disjoint partitions by 1101 // re-walking the recursive users of the alloca. 1102 Uses.resize(Partitions.size()); 1103 UseBuilder UB(TD, AI, *this); 1104 PtrI = UB.visitPtr(AI); 1105 assert(!PtrI.isEscaped() && "Previously analyzed pointer now escapes!"); 1106 assert(!PtrI.isAborted() && "Early aborted the visit of the pointer."); 1107} 1108 1109Type *AllocaPartitioning::getCommonType(iterator I) const { 1110 Type *Ty = 0; 1111 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) { 1112 Use *U = UI->getUse(); 1113 if (!U) 1114 continue; // Skip dead uses. 1115 if (isa<IntrinsicInst>(*U->getUser())) 1116 continue; 1117 if (UI->BeginOffset != I->BeginOffset || UI->EndOffset != I->EndOffset) 1118 continue; 1119 1120 Type *UserTy = 0; 1121 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) 1122 UserTy = LI->getType(); 1123 else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) 1124 UserTy = SI->getValueOperand()->getType(); 1125 else 1126 return 0; // Bail if we have weird uses. 1127 1128 if (IntegerType *ITy = dyn_cast<IntegerType>(UserTy)) { 1129 // If the type is larger than the partition, skip it. We only encounter 1130 // this for split integer operations where we want to use the type of the 1131 // entity causing the split. 1132 if (ITy->getBitWidth() > (I->EndOffset - I->BeginOffset)*8) 1133 continue; 1134 1135 // If we have found an integer type use covering the alloca, use that 1136 // regardless of the other types, as integers are often used for a "bucket 1137 // of bits" type. 1138 return ITy; 1139 } 1140 1141 if (Ty && Ty != UserTy) 1142 return 0; 1143 1144 Ty = UserTy; 1145 } 1146 return Ty; 1147} 1148 1149#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 1150 1151void AllocaPartitioning::print(raw_ostream &OS, const_iterator I, 1152 StringRef Indent) const { 1153 OS << Indent << "partition #" << (I - begin()) 1154 << " [" << I->BeginOffset << "," << I->EndOffset << ")" 1155 << (I->IsSplittable ? " (splittable)" : "") 1156 << (Uses[I - begin()].empty() ? " (zero uses)" : "") 1157 << "\n"; 1158} 1159 1160void AllocaPartitioning::printUsers(raw_ostream &OS, const_iterator I, 1161 StringRef Indent) const { 1162 for (const_use_iterator UI = use_begin(I), UE = use_end(I); UI != UE; ++UI) { 1163 if (!UI->getUse()) 1164 continue; // Skip dead uses. 1165 OS << Indent << " [" << UI->BeginOffset << "," << UI->EndOffset << ") " 1166 << "used by: " << *UI->getUse()->getUser() << "\n"; 1167 if (MemTransferInst *II = 1168 dyn_cast<MemTransferInst>(UI->getUse()->getUser())) { 1169 const MemTransferOffsets &MTO = MemTransferInstData.lookup(II); 1170 bool IsDest; 1171 if (!MTO.IsSplittable) 1172 IsDest = UI->BeginOffset == MTO.DestBegin; 1173 else 1174 IsDest = MTO.DestBegin != 0u; 1175 OS << Indent << " (original " << (IsDest ? "dest" : "source") << ": " 1176 << "[" << (IsDest ? MTO.DestBegin : MTO.SourceBegin) 1177 << "," << (IsDest ? MTO.DestEnd : MTO.SourceEnd) << ")\n"; 1178 } 1179 } 1180} 1181 1182void AllocaPartitioning::print(raw_ostream &OS) const { 1183 if (PointerEscapingInstr) { 1184 OS << "No partitioning for alloca: " << AI << "\n" 1185 << " A pointer to this alloca escaped by:\n" 1186 << " " << *PointerEscapingInstr << "\n"; 1187 return; 1188 } 1189 1190 OS << "Partitioning of alloca: " << AI << "\n"; 1191 for (const_iterator I = begin(), E = end(); I != E; ++I) { 1192 print(OS, I); 1193 printUsers(OS, I); 1194 } 1195} 1196 1197void AllocaPartitioning::dump(const_iterator I) const { print(dbgs(), I); } 1198void AllocaPartitioning::dump() const { print(dbgs()); } 1199 1200#endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 1201 1202 1203namespace { 1204/// \brief Implementation of LoadAndStorePromoter for promoting allocas. 1205/// 1206/// This subclass of LoadAndStorePromoter adds overrides to handle promoting 1207/// the loads and stores of an alloca instruction, as well as updating its 1208/// debug information. This is used when a domtree is unavailable and thus 1209/// mem2reg in its full form can't be used to handle promotion of allocas to 1210/// scalar values. 1211class AllocaPromoter : public LoadAndStorePromoter { 1212 AllocaInst &AI; 1213 DIBuilder &DIB; 1214 1215 SmallVector<DbgDeclareInst *, 4> DDIs; 1216 SmallVector<DbgValueInst *, 4> DVIs; 1217 1218public: 1219 AllocaPromoter(const SmallVectorImpl<Instruction*> &Insts, SSAUpdater &S, 1220 AllocaInst &AI, DIBuilder &DIB) 1221 : LoadAndStorePromoter(Insts, S), AI(AI), DIB(DIB) {} 1222 1223 void run(const SmallVectorImpl<Instruction*> &Insts) { 1224 // Remember which alloca we're promoting (for isInstInList). 1225 if (MDNode *DebugNode = MDNode::getIfExists(AI.getContext(), &AI)) { 1226 for (Value::use_iterator UI = DebugNode->use_begin(), 1227 UE = DebugNode->use_end(); 1228 UI != UE; ++UI) 1229 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI)) 1230 DDIs.push_back(DDI); 1231 else if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(*UI)) 1232 DVIs.push_back(DVI); 1233 } 1234 1235 LoadAndStorePromoter::run(Insts); 1236 AI.eraseFromParent(); 1237 while (!DDIs.empty()) 1238 DDIs.pop_back_val()->eraseFromParent(); 1239 while (!DVIs.empty()) 1240 DVIs.pop_back_val()->eraseFromParent(); 1241 } 1242 1243 virtual bool isInstInList(Instruction *I, 1244 const SmallVectorImpl<Instruction*> &Insts) const { 1245 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 1246 return LI->getOperand(0) == &AI; 1247 return cast<StoreInst>(I)->getPointerOperand() == &AI; 1248 } 1249 1250 virtual void updateDebugInfo(Instruction *Inst) const { 1251 for (SmallVector<DbgDeclareInst *, 4>::const_iterator I = DDIs.begin(), 1252 E = DDIs.end(); I != E; ++I) { 1253 DbgDeclareInst *DDI = *I; 1254 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 1255 ConvertDebugDeclareToDebugValue(DDI, SI, DIB); 1256 else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 1257 ConvertDebugDeclareToDebugValue(DDI, LI, DIB); 1258 } 1259 for (SmallVector<DbgValueInst *, 4>::const_iterator I = DVIs.begin(), 1260 E = DVIs.end(); I != E; ++I) { 1261 DbgValueInst *DVI = *I; 1262 Value *Arg = 0; 1263 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 1264 // If an argument is zero extended then use argument directly. The ZExt 1265 // may be zapped by an optimization pass in future. 1266 if (ZExtInst *ZExt = dyn_cast<ZExtInst>(SI->getOperand(0))) 1267 Arg = dyn_cast<Argument>(ZExt->getOperand(0)); 1268 if (SExtInst *SExt = dyn_cast<SExtInst>(SI->getOperand(0))) 1269 Arg = dyn_cast<Argument>(SExt->getOperand(0)); 1270 if (!Arg) 1271 Arg = SI->getOperand(0); 1272 } else if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 1273 Arg = LI->getOperand(0); 1274 } else { 1275 continue; 1276 } 1277 Instruction *DbgVal = 1278 DIB.insertDbgValueIntrinsic(Arg, 0, DIVariable(DVI->getVariable()), 1279 Inst); 1280 DbgVal->setDebugLoc(DVI->getDebugLoc()); 1281 } 1282 } 1283}; 1284} // end anon namespace 1285 1286 1287namespace { 1288/// \brief An optimization pass providing Scalar Replacement of Aggregates. 1289/// 1290/// This pass takes allocations which can be completely analyzed (that is, they 1291/// don't escape) and tries to turn them into scalar SSA values. There are 1292/// a few steps to this process. 1293/// 1294/// 1) It takes allocations of aggregates and analyzes the ways in which they 1295/// are used to try to split them into smaller allocations, ideally of 1296/// a single scalar data type. It will split up memcpy and memset accesses 1297/// as necessary and try to isolate individual scalar accesses. 1298/// 2) It will transform accesses into forms which are suitable for SSA value 1299/// promotion. This can be replacing a memset with a scalar store of an 1300/// integer value, or it can involve speculating operations on a PHI or 1301/// select to be a PHI or select of the results. 1302/// 3) Finally, this will try to detect a pattern of accesses which map cleanly 1303/// onto insert and extract operations on a vector value, and convert them to 1304/// this form. By doing so, it will enable promotion of vector aggregates to 1305/// SSA vector values. 1306class SROA : public FunctionPass { 1307 const bool RequiresDomTree; 1308 1309 LLVMContext *C; 1310 const DataLayout *TD; 1311 DominatorTree *DT; 1312 1313 /// \brief Worklist of alloca instructions to simplify. 1314 /// 1315 /// Each alloca in the function is added to this. Each new alloca formed gets 1316 /// added to it as well to recursively simplify unless that alloca can be 1317 /// directly promoted. Finally, each time we rewrite a use of an alloca other 1318 /// the one being actively rewritten, we add it back onto the list if not 1319 /// already present to ensure it is re-visited. 1320 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > Worklist; 1321 1322 /// \brief A collection of instructions to delete. 1323 /// We try to batch deletions to simplify code and make things a bit more 1324 /// efficient. 1325 SetVector<Instruction *, SmallVector<Instruction *, 8> > DeadInsts; 1326 1327 /// \brief Post-promotion worklist. 1328 /// 1329 /// Sometimes we discover an alloca which has a high probability of becoming 1330 /// viable for SROA after a round of promotion takes place. In those cases, 1331 /// the alloca is enqueued here for re-processing. 1332 /// 1333 /// Note that we have to be very careful to clear allocas out of this list in 1334 /// the event they are deleted. 1335 SetVector<AllocaInst *, SmallVector<AllocaInst *, 16> > PostPromotionWorklist; 1336 1337 /// \brief A collection of alloca instructions we can directly promote. 1338 std::vector<AllocaInst *> PromotableAllocas; 1339 1340public: 1341 SROA(bool RequiresDomTree = true) 1342 : FunctionPass(ID), RequiresDomTree(RequiresDomTree), 1343 C(0), TD(0), DT(0) { 1344 initializeSROAPass(*PassRegistry::getPassRegistry()); 1345 } 1346 bool runOnFunction(Function &F); 1347 void getAnalysisUsage(AnalysisUsage &AU) const; 1348 1349 const char *getPassName() const { return "SROA"; } 1350 static char ID; 1351 1352private: 1353 friend class PHIOrSelectSpeculator; 1354 friend class AllocaPartitionRewriter; 1355 friend class AllocaPartitionVectorRewriter; 1356 1357 bool rewriteAllocaPartition(AllocaInst &AI, 1358 AllocaPartitioning &P, 1359 AllocaPartitioning::iterator PI); 1360 bool splitAlloca(AllocaInst &AI, AllocaPartitioning &P); 1361 bool runOnAlloca(AllocaInst &AI); 1362 void deleteDeadInstructions(SmallPtrSet<AllocaInst *, 4> &DeletedAllocas); 1363 bool promoteAllocas(Function &F); 1364}; 1365} 1366 1367char SROA::ID = 0; 1368 1369FunctionPass *llvm::createSROAPass(bool RequiresDomTree) { 1370 return new SROA(RequiresDomTree); 1371} 1372 1373INITIALIZE_PASS_BEGIN(SROA, "sroa", "Scalar Replacement Of Aggregates", 1374 false, false) 1375INITIALIZE_PASS_DEPENDENCY(DominatorTree) 1376INITIALIZE_PASS_END(SROA, "sroa", "Scalar Replacement Of Aggregates", 1377 false, false) 1378 1379namespace { 1380/// \brief Visitor to speculate PHIs and Selects where possible. 1381class PHIOrSelectSpeculator : public InstVisitor<PHIOrSelectSpeculator> { 1382 // Befriend the base class so it can delegate to private visit methods. 1383 friend class llvm::InstVisitor<PHIOrSelectSpeculator>; 1384 1385 const DataLayout &TD; 1386 AllocaPartitioning &P; 1387 SROA &Pass; 1388 1389public: 1390 PHIOrSelectSpeculator(const DataLayout &TD, AllocaPartitioning &P, SROA &Pass) 1391 : TD(TD), P(P), Pass(Pass) {} 1392 1393 /// \brief Visit the users of an alloca partition and rewrite them. 1394 void visitUsers(AllocaPartitioning::const_iterator PI) { 1395 // Note that we need to use an index here as the underlying vector of uses 1396 // may be grown during speculation. However, we never need to re-visit the 1397 // new uses, and so we can use the initial size bound. 1398 for (unsigned Idx = 0, Size = P.use_size(PI); Idx != Size; ++Idx) { 1399 const PartitionUse &PU = P.getUse(PI, Idx); 1400 if (!PU.getUse()) 1401 continue; // Skip dead use. 1402 1403 visit(cast<Instruction>(PU.getUse()->getUser())); 1404 } 1405 } 1406 1407private: 1408 // By default, skip this instruction. 1409 void visitInstruction(Instruction &I) {} 1410 1411 /// PHI instructions that use an alloca and are subsequently loaded can be 1412 /// rewritten to load both input pointers in the pred blocks and then PHI the 1413 /// results, allowing the load of the alloca to be promoted. 1414 /// From this: 1415 /// %P2 = phi [i32* %Alloca, i32* %Other] 1416 /// %V = load i32* %P2 1417 /// to: 1418 /// %V1 = load i32* %Alloca -> will be mem2reg'd 1419 /// ... 1420 /// %V2 = load i32* %Other 1421 /// ... 1422 /// %V = phi [i32 %V1, i32 %V2] 1423 /// 1424 /// We can do this to a select if its only uses are loads and if the operands 1425 /// to the select can be loaded unconditionally. 1426 /// 1427 /// FIXME: This should be hoisted into a generic utility, likely in 1428 /// Transforms/Util/Local.h 1429 bool isSafePHIToSpeculate(PHINode &PN, SmallVectorImpl<LoadInst *> &Loads) { 1430 // For now, we can only do this promotion if the load is in the same block 1431 // as the PHI, and if there are no stores between the phi and load. 1432 // TODO: Allow recursive phi users. 1433 // TODO: Allow stores. 1434 BasicBlock *BB = PN.getParent(); 1435 unsigned MaxAlign = 0; 1436 for (Value::use_iterator UI = PN.use_begin(), UE = PN.use_end(); 1437 UI != UE; ++UI) { 1438 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1439 if (LI == 0 || !LI->isSimple()) return false; 1440 1441 // For now we only allow loads in the same block as the PHI. This is 1442 // a common case that happens when instcombine merges two loads through 1443 // a PHI. 1444 if (LI->getParent() != BB) return false; 1445 1446 // Ensure that there are no instructions between the PHI and the load that 1447 // could store. 1448 for (BasicBlock::iterator BBI = &PN; &*BBI != LI; ++BBI) 1449 if (BBI->mayWriteToMemory()) 1450 return false; 1451 1452 MaxAlign = std::max(MaxAlign, LI->getAlignment()); 1453 Loads.push_back(LI); 1454 } 1455 1456 // We can only transform this if it is safe to push the loads into the 1457 // predecessor blocks. The only thing to watch out for is that we can't put 1458 // a possibly trapping load in the predecessor if it is a critical edge. 1459 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) { 1460 TerminatorInst *TI = PN.getIncomingBlock(Idx)->getTerminator(); 1461 Value *InVal = PN.getIncomingValue(Idx); 1462 1463 // If the value is produced by the terminator of the predecessor (an 1464 // invoke) or it has side-effects, there is no valid place to put a load 1465 // in the predecessor. 1466 if (TI == InVal || TI->mayHaveSideEffects()) 1467 return false; 1468 1469 // If the predecessor has a single successor, then the edge isn't 1470 // critical. 1471 if (TI->getNumSuccessors() == 1) 1472 continue; 1473 1474 // If this pointer is always safe to load, or if we can prove that there 1475 // is already a load in the block, then we can move the load to the pred 1476 // block. 1477 if (InVal->isDereferenceablePointer() || 1478 isSafeToLoadUnconditionally(InVal, TI, MaxAlign, &TD)) 1479 continue; 1480 1481 return false; 1482 } 1483 1484 return true; 1485 } 1486 1487 void visitPHINode(PHINode &PN) { 1488 DEBUG(dbgs() << " original: " << PN << "\n"); 1489 1490 SmallVector<LoadInst *, 4> Loads; 1491 if (!isSafePHIToSpeculate(PN, Loads)) 1492 return; 1493 1494 assert(!Loads.empty()); 1495 1496 Type *LoadTy = cast<PointerType>(PN.getType())->getElementType(); 1497 IRBuilder<> PHIBuilder(&PN); 1498 PHINode *NewPN = PHIBuilder.CreatePHI(LoadTy, PN.getNumIncomingValues(), 1499 PN.getName() + ".sroa.speculated"); 1500 1501 // Get the TBAA tag and alignment to use from one of the loads. It doesn't 1502 // matter which one we get and if any differ. 1503 LoadInst *SomeLoad = cast<LoadInst>(Loads.back()); 1504 MDNode *TBAATag = SomeLoad->getMetadata(LLVMContext::MD_tbaa); 1505 unsigned Align = SomeLoad->getAlignment(); 1506 1507 // Rewrite all loads of the PN to use the new PHI. 1508 do { 1509 LoadInst *LI = Loads.pop_back_val(); 1510 LI->replaceAllUsesWith(NewPN); 1511 Pass.DeadInsts.insert(LI); 1512 } while (!Loads.empty()); 1513 1514 // Inject loads into all of the pred blocks. 1515 for (unsigned Idx = 0, Num = PN.getNumIncomingValues(); Idx != Num; ++Idx) { 1516 BasicBlock *Pred = PN.getIncomingBlock(Idx); 1517 TerminatorInst *TI = Pred->getTerminator(); 1518 Use *InUse = &PN.getOperandUse(PN.getOperandNumForIncomingValue(Idx)); 1519 Value *InVal = PN.getIncomingValue(Idx); 1520 IRBuilder<> PredBuilder(TI); 1521 1522 LoadInst *Load 1523 = PredBuilder.CreateLoad(InVal, (PN.getName() + ".sroa.speculate.load." + 1524 Pred->getName())); 1525 ++NumLoadsSpeculated; 1526 Load->setAlignment(Align); 1527 if (TBAATag) 1528 Load->setMetadata(LLVMContext::MD_tbaa, TBAATag); 1529 NewPN->addIncoming(Load, Pred); 1530 1531 Instruction *Ptr = dyn_cast<Instruction>(InVal); 1532 if (!Ptr) 1533 // No uses to rewrite. 1534 continue; 1535 1536 // Try to lookup and rewrite any partition uses corresponding to this phi 1537 // input. 1538 AllocaPartitioning::iterator PI 1539 = P.findPartitionForPHIOrSelectOperand(InUse); 1540 if (PI == P.end()) 1541 continue; 1542 1543 // Replace the Use in the PartitionUse for this operand with the Use 1544 // inside the load. 1545 AllocaPartitioning::use_iterator UI 1546 = P.findPartitionUseForPHIOrSelectOperand(InUse); 1547 assert(isa<PHINode>(*UI->getUse()->getUser())); 1548 UI->setUse(&Load->getOperandUse(Load->getPointerOperandIndex())); 1549 } 1550 DEBUG(dbgs() << " speculated to: " << *NewPN << "\n"); 1551 } 1552 1553 /// Select instructions that use an alloca and are subsequently loaded can be 1554 /// rewritten to load both input pointers and then select between the result, 1555 /// allowing the load of the alloca to be promoted. 1556 /// From this: 1557 /// %P2 = select i1 %cond, i32* %Alloca, i32* %Other 1558 /// %V = load i32* %P2 1559 /// to: 1560 /// %V1 = load i32* %Alloca -> will be mem2reg'd 1561 /// %V2 = load i32* %Other 1562 /// %V = select i1 %cond, i32 %V1, i32 %V2 1563 /// 1564 /// We can do this to a select if its only uses are loads and if the operand 1565 /// to the select can be loaded unconditionally. 1566 bool isSafeSelectToSpeculate(SelectInst &SI, 1567 SmallVectorImpl<LoadInst *> &Loads) { 1568 Value *TValue = SI.getTrueValue(); 1569 Value *FValue = SI.getFalseValue(); 1570 bool TDerefable = TValue->isDereferenceablePointer(); 1571 bool FDerefable = FValue->isDereferenceablePointer(); 1572 1573 for (Value::use_iterator UI = SI.use_begin(), UE = SI.use_end(); 1574 UI != UE; ++UI) { 1575 LoadInst *LI = dyn_cast<LoadInst>(*UI); 1576 if (LI == 0 || !LI->isSimple()) return false; 1577 1578 // Both operands to the select need to be dereferencable, either 1579 // absolutely (e.g. allocas) or at this point because we can see other 1580 // accesses to it. 1581 if (!TDerefable && !isSafeToLoadUnconditionally(TValue, LI, 1582 LI->getAlignment(), &TD)) 1583 return false; 1584 if (!FDerefable && !isSafeToLoadUnconditionally(FValue, LI, 1585 LI->getAlignment(), &TD)) 1586 return false; 1587 Loads.push_back(LI); 1588 } 1589 1590 return true; 1591 } 1592 1593 void visitSelectInst(SelectInst &SI) { 1594 DEBUG(dbgs() << " original: " << SI << "\n"); 1595 1596 // If the select isn't safe to speculate, just use simple logic to emit it. 1597 SmallVector<LoadInst *, 4> Loads; 1598 if (!isSafeSelectToSpeculate(SI, Loads)) 1599 return; 1600 1601 IRBuilder<> IRB(&SI); 1602 Use *Ops[2] = { &SI.getOperandUse(1), &SI.getOperandUse(2) }; 1603 AllocaPartitioning::iterator PIs[2]; 1604 PartitionUse PUs[2]; 1605 for (unsigned i = 0, e = 2; i != e; ++i) { 1606 PIs[i] = P.findPartitionForPHIOrSelectOperand(Ops[i]); 1607 if (PIs[i] != P.end()) { 1608 // If the pointer is within the partitioning, remove the select from 1609 // its uses. We'll add in the new loads below. 1610 AllocaPartitioning::use_iterator UI 1611 = P.findPartitionUseForPHIOrSelectOperand(Ops[i]); 1612 PUs[i] = *UI; 1613 // Clear out the use here so that the offsets into the use list remain 1614 // stable but this use is ignored when rewriting. 1615 UI->setUse(0); 1616 } 1617 } 1618 1619 Value *TV = SI.getTrueValue(); 1620 Value *FV = SI.getFalseValue(); 1621 // Replace the loads of the select with a select of two loads. 1622 while (!Loads.empty()) { 1623 LoadInst *LI = Loads.pop_back_val(); 1624 1625 IRB.SetInsertPoint(LI); 1626 LoadInst *TL = 1627 IRB.CreateLoad(TV, LI->getName() + ".sroa.speculate.load.true"); 1628 LoadInst *FL = 1629 IRB.CreateLoad(FV, LI->getName() + ".sroa.speculate.load.false"); 1630 NumLoadsSpeculated += 2; 1631 1632 // Transfer alignment and TBAA info if present. 1633 TL->setAlignment(LI->getAlignment()); 1634 FL->setAlignment(LI->getAlignment()); 1635 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) { 1636 TL->setMetadata(LLVMContext::MD_tbaa, Tag); 1637 FL->setMetadata(LLVMContext::MD_tbaa, Tag); 1638 } 1639 1640 Value *V = IRB.CreateSelect(SI.getCondition(), TL, FL, 1641 LI->getName() + ".sroa.speculated"); 1642 1643 LoadInst *Loads[2] = { TL, FL }; 1644 for (unsigned i = 0, e = 2; i != e; ++i) { 1645 if (PIs[i] != P.end()) { 1646 Use *LoadUse = &Loads[i]->getOperandUse(0); 1647 assert(PUs[i].getUse()->get() == LoadUse->get()); 1648 PUs[i].setUse(LoadUse); 1649 P.use_push_back(PIs[i], PUs[i]); 1650 } 1651 } 1652 1653 DEBUG(dbgs() << " speculated to: " << *V << "\n"); 1654 LI->replaceAllUsesWith(V); 1655 Pass.DeadInsts.insert(LI); 1656 } 1657 } 1658}; 1659} 1660 1661/// \brief Build a GEP out of a base pointer and indices. 1662/// 1663/// This will return the BasePtr if that is valid, or build a new GEP 1664/// instruction using the IRBuilder if GEP-ing is needed. 1665static Value *buildGEP(IRBuilder<> &IRB, Value *BasePtr, 1666 SmallVectorImpl<Value *> &Indices, 1667 const Twine &Prefix) { 1668 if (Indices.empty()) 1669 return BasePtr; 1670 1671 // A single zero index is a no-op, so check for this and avoid building a GEP 1672 // in that case. 1673 if (Indices.size() == 1 && cast<ConstantInt>(Indices.back())->isZero()) 1674 return BasePtr; 1675 1676 return IRB.CreateInBoundsGEP(BasePtr, Indices, Prefix + ".idx"); 1677} 1678 1679/// \brief Get a natural GEP off of the BasePtr walking through Ty toward 1680/// TargetTy without changing the offset of the pointer. 1681/// 1682/// This routine assumes we've already established a properly offset GEP with 1683/// Indices, and arrived at the Ty type. The goal is to continue to GEP with 1684/// zero-indices down through type layers until we find one the same as 1685/// TargetTy. If we can't find one with the same type, we at least try to use 1686/// one with the same size. If none of that works, we just produce the GEP as 1687/// indicated by Indices to have the correct offset. 1688static Value *getNaturalGEPWithType(IRBuilder<> &IRB, const DataLayout &TD, 1689 Value *BasePtr, Type *Ty, Type *TargetTy, 1690 SmallVectorImpl<Value *> &Indices, 1691 const Twine &Prefix) { 1692 if (Ty == TargetTy) 1693 return buildGEP(IRB, BasePtr, Indices, Prefix); 1694 1695 // See if we can descend into a struct and locate a field with the correct 1696 // type. 1697 unsigned NumLayers = 0; 1698 Type *ElementTy = Ty; 1699 do { 1700 if (ElementTy->isPointerTy()) 1701 break; 1702 if (SequentialType *SeqTy = dyn_cast<SequentialType>(ElementTy)) { 1703 ElementTy = SeqTy->getElementType(); 1704 // Note that we use the default address space as this index is over an 1705 // array or a vector, not a pointer. 1706 Indices.push_back(IRB.getInt(APInt(TD.getPointerSizeInBits(0), 0))); 1707 } else if (StructType *STy = dyn_cast<StructType>(ElementTy)) { 1708 if (STy->element_begin() == STy->element_end()) 1709 break; // Nothing left to descend into. 1710 ElementTy = *STy->element_begin(); 1711 Indices.push_back(IRB.getInt32(0)); 1712 } else { 1713 break; 1714 } 1715 ++NumLayers; 1716 } while (ElementTy != TargetTy); 1717 if (ElementTy != TargetTy) 1718 Indices.erase(Indices.end() - NumLayers, Indices.end()); 1719 1720 return buildGEP(IRB, BasePtr, Indices, Prefix); 1721} 1722 1723/// \brief Recursively compute indices for a natural GEP. 1724/// 1725/// This is the recursive step for getNaturalGEPWithOffset that walks down the 1726/// element types adding appropriate indices for the GEP. 1727static Value *getNaturalGEPRecursively(IRBuilder<> &IRB, const DataLayout &TD, 1728 Value *Ptr, Type *Ty, APInt &Offset, 1729 Type *TargetTy, 1730 SmallVectorImpl<Value *> &Indices, 1731 const Twine &Prefix) { 1732 if (Offset == 0) 1733 return getNaturalGEPWithType(IRB, TD, Ptr, Ty, TargetTy, Indices, Prefix); 1734 1735 // We can't recurse through pointer types. 1736 if (Ty->isPointerTy()) 1737 return 0; 1738 1739 // We try to analyze GEPs over vectors here, but note that these GEPs are 1740 // extremely poorly defined currently. The long-term goal is to remove GEPing 1741 // over a vector from the IR completely. 1742 if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) { 1743 unsigned ElementSizeInBits = TD.getTypeSizeInBits(VecTy->getScalarType()); 1744 if (ElementSizeInBits % 8) 1745 return 0; // GEPs over non-multiple of 8 size vector elements are invalid. 1746 APInt ElementSize(Offset.getBitWidth(), ElementSizeInBits / 8); 1747 APInt NumSkippedElements = Offset.sdiv(ElementSize); 1748 if (NumSkippedElements.ugt(VecTy->getNumElements())) 1749 return 0; 1750 Offset -= NumSkippedElements * ElementSize; 1751 Indices.push_back(IRB.getInt(NumSkippedElements)); 1752 return getNaturalGEPRecursively(IRB, TD, Ptr, VecTy->getElementType(), 1753 Offset, TargetTy, Indices, Prefix); 1754 } 1755 1756 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) { 1757 Type *ElementTy = ArrTy->getElementType(); 1758 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy)); 1759 APInt NumSkippedElements = Offset.sdiv(ElementSize); 1760 if (NumSkippedElements.ugt(ArrTy->getNumElements())) 1761 return 0; 1762 1763 Offset -= NumSkippedElements * ElementSize; 1764 Indices.push_back(IRB.getInt(NumSkippedElements)); 1765 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy, 1766 Indices, Prefix); 1767 } 1768 1769 StructType *STy = dyn_cast<StructType>(Ty); 1770 if (!STy) 1771 return 0; 1772 1773 const StructLayout *SL = TD.getStructLayout(STy); 1774 uint64_t StructOffset = Offset.getZExtValue(); 1775 if (StructOffset >= SL->getSizeInBytes()) 1776 return 0; 1777 unsigned Index = SL->getElementContainingOffset(StructOffset); 1778 Offset -= APInt(Offset.getBitWidth(), SL->getElementOffset(Index)); 1779 Type *ElementTy = STy->getElementType(Index); 1780 if (Offset.uge(TD.getTypeAllocSize(ElementTy))) 1781 return 0; // The offset points into alignment padding. 1782 1783 Indices.push_back(IRB.getInt32(Index)); 1784 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy, 1785 Indices, Prefix); 1786} 1787 1788/// \brief Get a natural GEP from a base pointer to a particular offset and 1789/// resulting in a particular type. 1790/// 1791/// The goal is to produce a "natural" looking GEP that works with the existing 1792/// composite types to arrive at the appropriate offset and element type for 1793/// a pointer. TargetTy is the element type the returned GEP should point-to if 1794/// possible. We recurse by decreasing Offset, adding the appropriate index to 1795/// Indices, and setting Ty to the result subtype. 1796/// 1797/// If no natural GEP can be constructed, this function returns null. 1798static Value *getNaturalGEPWithOffset(IRBuilder<> &IRB, const DataLayout &TD, 1799 Value *Ptr, APInt Offset, Type *TargetTy, 1800 SmallVectorImpl<Value *> &Indices, 1801 const Twine &Prefix) { 1802 PointerType *Ty = cast<PointerType>(Ptr->getType()); 1803 1804 // Don't consider any GEPs through an i8* as natural unless the TargetTy is 1805 // an i8. 1806 if (Ty == IRB.getInt8PtrTy() && TargetTy->isIntegerTy(8)) 1807 return 0; 1808 1809 Type *ElementTy = Ty->getElementType(); 1810 if (!ElementTy->isSized()) 1811 return 0; // We can't GEP through an unsized element. 1812 APInt ElementSize(Offset.getBitWidth(), TD.getTypeAllocSize(ElementTy)); 1813 if (ElementSize == 0) 1814 return 0; // Zero-length arrays can't help us build a natural GEP. 1815 APInt NumSkippedElements = Offset.sdiv(ElementSize); 1816 1817 Offset -= NumSkippedElements * ElementSize; 1818 Indices.push_back(IRB.getInt(NumSkippedElements)); 1819 return getNaturalGEPRecursively(IRB, TD, Ptr, ElementTy, Offset, TargetTy, 1820 Indices, Prefix); 1821} 1822 1823/// \brief Compute an adjusted pointer from Ptr by Offset bytes where the 1824/// resulting pointer has PointerTy. 1825/// 1826/// This tries very hard to compute a "natural" GEP which arrives at the offset 1827/// and produces the pointer type desired. Where it cannot, it will try to use 1828/// the natural GEP to arrive at the offset and bitcast to the type. Where that 1829/// fails, it will try to use an existing i8* and GEP to the byte offset and 1830/// bitcast to the type. 1831/// 1832/// The strategy for finding the more natural GEPs is to peel off layers of the 1833/// pointer, walking back through bit casts and GEPs, searching for a base 1834/// pointer from which we can compute a natural GEP with the desired 1835/// properties. The algorithm tries to fold as many constant indices into 1836/// a single GEP as possible, thus making each GEP more independent of the 1837/// surrounding code. 1838static Value *getAdjustedPtr(IRBuilder<> &IRB, const DataLayout &TD, 1839 Value *Ptr, APInt Offset, Type *PointerTy, 1840 const Twine &Prefix) { 1841 // Even though we don't look through PHI nodes, we could be called on an 1842 // instruction in an unreachable block, which may be on a cycle. 1843 SmallPtrSet<Value *, 4> Visited; 1844 Visited.insert(Ptr); 1845 SmallVector<Value *, 4> Indices; 1846 1847 // We may end up computing an offset pointer that has the wrong type. If we 1848 // never are able to compute one directly that has the correct type, we'll 1849 // fall back to it, so keep it around here. 1850 Value *OffsetPtr = 0; 1851 1852 // Remember any i8 pointer we come across to re-use if we need to do a raw 1853 // byte offset. 1854 Value *Int8Ptr = 0; 1855 APInt Int8PtrOffset(Offset.getBitWidth(), 0); 1856 1857 Type *TargetTy = PointerTy->getPointerElementType(); 1858 1859 do { 1860 // First fold any existing GEPs into the offset. 1861 while (GEPOperator *GEP = dyn_cast<GEPOperator>(Ptr)) { 1862 APInt GEPOffset(Offset.getBitWidth(), 0); 1863 if (!GEP->accumulateConstantOffset(TD, GEPOffset)) 1864 break; 1865 Offset += GEPOffset; 1866 Ptr = GEP->getPointerOperand(); 1867 if (!Visited.insert(Ptr)) 1868 break; 1869 } 1870 1871 // See if we can perform a natural GEP here. 1872 Indices.clear(); 1873 if (Value *P = getNaturalGEPWithOffset(IRB, TD, Ptr, Offset, TargetTy, 1874 Indices, Prefix)) { 1875 if (P->getType() == PointerTy) { 1876 // Zap any offset pointer that we ended up computing in previous rounds. 1877 if (OffsetPtr && OffsetPtr->use_empty()) 1878 if (Instruction *I = dyn_cast<Instruction>(OffsetPtr)) 1879 I->eraseFromParent(); 1880 return P; 1881 } 1882 if (!OffsetPtr) { 1883 OffsetPtr = P; 1884 } 1885 } 1886 1887 // Stash this pointer if we've found an i8*. 1888 if (Ptr->getType()->isIntegerTy(8)) { 1889 Int8Ptr = Ptr; 1890 Int8PtrOffset = Offset; 1891 } 1892 1893 // Peel off a layer of the pointer and update the offset appropriately. 1894 if (Operator::getOpcode(Ptr) == Instruction::BitCast) { 1895 Ptr = cast<Operator>(Ptr)->getOperand(0); 1896 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(Ptr)) { 1897 if (GA->mayBeOverridden()) 1898 break; 1899 Ptr = GA->getAliasee(); 1900 } else { 1901 break; 1902 } 1903 assert(Ptr->getType()->isPointerTy() && "Unexpected operand type!"); 1904 } while (Visited.insert(Ptr)); 1905 1906 if (!OffsetPtr) { 1907 if (!Int8Ptr) { 1908 Int8Ptr = IRB.CreateBitCast(Ptr, IRB.getInt8PtrTy(), 1909 Prefix + ".raw_cast"); 1910 Int8PtrOffset = Offset; 1911 } 1912 1913 OffsetPtr = Int8PtrOffset == 0 ? Int8Ptr : 1914 IRB.CreateInBoundsGEP(Int8Ptr, IRB.getInt(Int8PtrOffset), 1915 Prefix + ".raw_idx"); 1916 } 1917 Ptr = OffsetPtr; 1918 1919 // On the off chance we were targeting i8*, guard the bitcast here. 1920 if (Ptr->getType() != PointerTy) 1921 Ptr = IRB.CreateBitCast(Ptr, PointerTy, Prefix + ".cast"); 1922 1923 return Ptr; 1924} 1925 1926/// \brief Test whether we can convert a value from the old to the new type. 1927/// 1928/// This predicate should be used to guard calls to convertValue in order to 1929/// ensure that we only try to convert viable values. The strategy is that we 1930/// will peel off single element struct and array wrappings to get to an 1931/// underlying value, and convert that value. 1932static bool canConvertValue(const DataLayout &DL, Type *OldTy, Type *NewTy) { 1933 if (OldTy == NewTy) 1934 return true; 1935 if (IntegerType *OldITy = dyn_cast<IntegerType>(OldTy)) 1936 if (IntegerType *NewITy = dyn_cast<IntegerType>(NewTy)) 1937 if (NewITy->getBitWidth() >= OldITy->getBitWidth()) 1938 return true; 1939 if (DL.getTypeSizeInBits(NewTy) != DL.getTypeSizeInBits(OldTy)) 1940 return false; 1941 if (!NewTy->isSingleValueType() || !OldTy->isSingleValueType()) 1942 return false; 1943 1944 if (NewTy->isPointerTy() || OldTy->isPointerTy()) { 1945 if (NewTy->isPointerTy() && OldTy->isPointerTy()) 1946 return true; 1947 if (NewTy->isIntegerTy() || OldTy->isIntegerTy()) 1948 return true; 1949 return false; 1950 } 1951 1952 return true; 1953} 1954 1955/// \brief Generic routine to convert an SSA value to a value of a different 1956/// type. 1957/// 1958/// This will try various different casting techniques, such as bitcasts, 1959/// inttoptr, and ptrtoint casts. Use the \c canConvertValue predicate to test 1960/// two types for viability with this routine. 1961static Value *convertValue(const DataLayout &DL, IRBuilder<> &IRB, Value *V, 1962 Type *Ty) { 1963 assert(canConvertValue(DL, V->getType(), Ty) && 1964 "Value not convertable to type"); 1965 if (V->getType() == Ty) 1966 return V; 1967 if (IntegerType *OldITy = dyn_cast<IntegerType>(V->getType())) 1968 if (IntegerType *NewITy = dyn_cast<IntegerType>(Ty)) 1969 if (NewITy->getBitWidth() > OldITy->getBitWidth()) 1970 return IRB.CreateZExt(V, NewITy); 1971 if (V->getType()->isIntegerTy() && Ty->isPointerTy()) 1972 return IRB.CreateIntToPtr(V, Ty); 1973 if (V->getType()->isPointerTy() && Ty->isIntegerTy()) 1974 return IRB.CreatePtrToInt(V, Ty); 1975 1976 return IRB.CreateBitCast(V, Ty); 1977} 1978 1979/// \brief Test whether the given alloca partition can be promoted to a vector. 1980/// 1981/// This is a quick test to check whether we can rewrite a particular alloca 1982/// partition (and its newly formed alloca) into a vector alloca with only 1983/// whole-vector loads and stores such that it could be promoted to a vector 1984/// SSA value. We only can ensure this for a limited set of operations, and we 1985/// don't want to do the rewrites unless we are confident that the result will 1986/// be promotable, so we have an early test here. 1987static bool isVectorPromotionViable(const DataLayout &TD, 1988 Type *AllocaTy, 1989 AllocaPartitioning &P, 1990 uint64_t PartitionBeginOffset, 1991 uint64_t PartitionEndOffset, 1992 AllocaPartitioning::const_use_iterator I, 1993 AllocaPartitioning::const_use_iterator E) { 1994 VectorType *Ty = dyn_cast<VectorType>(AllocaTy); 1995 if (!Ty) 1996 return false; 1997 1998 uint64_t ElementSize = TD.getTypeSizeInBits(Ty->getScalarType()); 1999 2000 // While the definition of LLVM vectors is bitpacked, we don't support sizes 2001 // that aren't byte sized. 2002 if (ElementSize % 8) 2003 return false; 2004 assert((TD.getTypeSizeInBits(Ty) % 8) == 0 && 2005 "vector size not a multiple of element size?"); 2006 ElementSize /= 8; 2007 2008 for (; I != E; ++I) { 2009 Use *U = I->getUse(); 2010 if (!U) 2011 continue; // Skip dead use. 2012 2013 uint64_t BeginOffset = I->BeginOffset - PartitionBeginOffset; 2014 uint64_t BeginIndex = BeginOffset / ElementSize; 2015 if (BeginIndex * ElementSize != BeginOffset || 2016 BeginIndex >= Ty->getNumElements()) 2017 return false; 2018 uint64_t EndOffset = I->EndOffset - PartitionBeginOffset; 2019 uint64_t EndIndex = EndOffset / ElementSize; 2020 if (EndIndex * ElementSize != EndOffset || 2021 EndIndex > Ty->getNumElements()) 2022 return false; 2023 2024 assert(EndIndex > BeginIndex && "Empty vector!"); 2025 uint64_t NumElements = EndIndex - BeginIndex; 2026 Type *PartitionTy 2027 = (NumElements == 1) ? Ty->getElementType() 2028 : VectorType::get(Ty->getElementType(), NumElements); 2029 2030 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) { 2031 if (MI->isVolatile()) 2032 return false; 2033 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(U->getUser())) { 2034 const AllocaPartitioning::MemTransferOffsets &MTO 2035 = P.getMemTransferOffsets(*MTI); 2036 if (!MTO.IsSplittable) 2037 return false; 2038 } 2039 } else if (U->get()->getType()->getPointerElementType()->isStructTy()) { 2040 // Disable vector promotion when there are loads or stores of an FCA. 2041 return false; 2042 } else if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) { 2043 if (LI->isVolatile()) 2044 return false; 2045 if (!canConvertValue(TD, PartitionTy, LI->getType())) 2046 return false; 2047 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) { 2048 if (SI->isVolatile()) 2049 return false; 2050 if (!canConvertValue(TD, SI->getValueOperand()->getType(), PartitionTy)) 2051 return false; 2052 } else { 2053 return false; 2054 } 2055 } 2056 return true; 2057} 2058 2059/// \brief Test whether the given alloca partition's integer operations can be 2060/// widened to promotable ones. 2061/// 2062/// This is a quick test to check whether we can rewrite the integer loads and 2063/// stores to a particular alloca into wider loads and stores and be able to 2064/// promote the resulting alloca. 2065static bool isIntegerWideningViable(const DataLayout &TD, 2066 Type *AllocaTy, 2067 uint64_t AllocBeginOffset, 2068 AllocaPartitioning &P, 2069 AllocaPartitioning::const_use_iterator I, 2070 AllocaPartitioning::const_use_iterator E) { 2071 uint64_t SizeInBits = TD.getTypeSizeInBits(AllocaTy); 2072 // Don't create integer types larger than the maximum bitwidth. 2073 if (SizeInBits > IntegerType::MAX_INT_BITS) 2074 return false; 2075 2076 // Don't try to handle allocas with bit-padding. 2077 if (SizeInBits != TD.getTypeStoreSizeInBits(AllocaTy)) 2078 return false; 2079 2080 // We need to ensure that an integer type with the appropriate bitwidth can 2081 // be converted to the alloca type, whatever that is. We don't want to force 2082 // the alloca itself to have an integer type if there is a more suitable one. 2083 Type *IntTy = Type::getIntNTy(AllocaTy->getContext(), SizeInBits); 2084 if (!canConvertValue(TD, AllocaTy, IntTy) || 2085 !canConvertValue(TD, IntTy, AllocaTy)) 2086 return false; 2087 2088 uint64_t Size = TD.getTypeStoreSize(AllocaTy); 2089 2090 // Check the uses to ensure the uses are (likely) promotable integer uses. 2091 // Also ensure that the alloca has a covering load or store. We don't want 2092 // to widen the integer operations only to fail to promote due to some other 2093 // unsplittable entry (which we may make splittable later). 2094 bool WholeAllocaOp = false; 2095 for (; I != E; ++I) { 2096 Use *U = I->getUse(); 2097 if (!U) 2098 continue; // Skip dead use. 2099 2100 uint64_t RelBegin = I->BeginOffset - AllocBeginOffset; 2101 uint64_t RelEnd = I->EndOffset - AllocBeginOffset; 2102 2103 // We can't reasonably handle cases where the load or store extends past 2104 // the end of the aloca's type and into its padding. 2105 if (RelEnd > Size) 2106 return false; 2107 2108 if (LoadInst *LI = dyn_cast<LoadInst>(U->getUser())) { 2109 if (LI->isVolatile()) 2110 return false; 2111 if (RelBegin == 0 && RelEnd == Size) 2112 WholeAllocaOp = true; 2113 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) { 2114 if (ITy->getBitWidth() < TD.getTypeStoreSizeInBits(ITy)) 2115 return false; 2116 continue; 2117 } 2118 // Non-integer loads need to be convertible from the alloca type so that 2119 // they are promotable. 2120 if (RelBegin != 0 || RelEnd != Size || 2121 !canConvertValue(TD, AllocaTy, LI->getType())) 2122 return false; 2123 } else if (StoreInst *SI = dyn_cast<StoreInst>(U->getUser())) { 2124 Type *ValueTy = SI->getValueOperand()->getType(); 2125 if (SI->isVolatile()) 2126 return false; 2127 if (RelBegin == 0 && RelEnd == Size) 2128 WholeAllocaOp = true; 2129 if (IntegerType *ITy = dyn_cast<IntegerType>(ValueTy)) { 2130 if (ITy->getBitWidth() < TD.getTypeStoreSizeInBits(ITy)) 2131 return false; 2132 continue; 2133 } 2134 // Non-integer stores need to be convertible to the alloca type so that 2135 // they are promotable. 2136 if (RelBegin != 0 || RelEnd != Size || 2137 !canConvertValue(TD, ValueTy, AllocaTy)) 2138 return false; 2139 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U->getUser())) { 2140 if (MI->isVolatile() || !isa<Constant>(MI->getLength())) 2141 return false; 2142 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(U->getUser())) { 2143 const AllocaPartitioning::MemTransferOffsets &MTO 2144 = P.getMemTransferOffsets(*MTI); 2145 if (!MTO.IsSplittable) 2146 return false; 2147 } 2148 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U->getUser())) { 2149 if (II->getIntrinsicID() != Intrinsic::lifetime_start && 2150 II->getIntrinsicID() != Intrinsic::lifetime_end) 2151 return false; 2152 } else { 2153 return false; 2154 } 2155 } 2156 return WholeAllocaOp; 2157} 2158 2159static Value *extractInteger(const DataLayout &DL, IRBuilder<> &IRB, Value *V, 2160 IntegerType *Ty, uint64_t Offset, 2161 const Twine &Name) { 2162 DEBUG(dbgs() << " start: " << *V << "\n"); 2163 IntegerType *IntTy = cast<IntegerType>(V->getType()); 2164 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) && 2165 "Element extends past full value"); 2166 uint64_t ShAmt = 8*Offset; 2167 if (DL.isBigEndian()) 2168 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset); 2169 if (ShAmt) { 2170 V = IRB.CreateLShr(V, ShAmt, Name + ".shift"); 2171 DEBUG(dbgs() << " shifted: " << *V << "\n"); 2172 } 2173 assert(Ty->getBitWidth() <= IntTy->getBitWidth() && 2174 "Cannot extract to a larger integer!"); 2175 if (Ty != IntTy) { 2176 V = IRB.CreateTrunc(V, Ty, Name + ".trunc"); 2177 DEBUG(dbgs() << " trunced: " << *V << "\n"); 2178 } 2179 return V; 2180} 2181 2182static Value *insertInteger(const DataLayout &DL, IRBuilder<> &IRB, Value *Old, 2183 Value *V, uint64_t Offset, const Twine &Name) { 2184 IntegerType *IntTy = cast<IntegerType>(Old->getType()); 2185 IntegerType *Ty = cast<IntegerType>(V->getType()); 2186 assert(Ty->getBitWidth() <= IntTy->getBitWidth() && 2187 "Cannot insert a larger integer!"); 2188 DEBUG(dbgs() << " start: " << *V << "\n"); 2189 if (Ty != IntTy) { 2190 V = IRB.CreateZExt(V, IntTy, Name + ".ext"); 2191 DEBUG(dbgs() << " extended: " << *V << "\n"); 2192 } 2193 assert(DL.getTypeStoreSize(Ty) + Offset <= DL.getTypeStoreSize(IntTy) && 2194 "Element store outside of alloca store"); 2195 uint64_t ShAmt = 8*Offset; 2196 if (DL.isBigEndian()) 2197 ShAmt = 8*(DL.getTypeStoreSize(IntTy) - DL.getTypeStoreSize(Ty) - Offset); 2198 if (ShAmt) { 2199 V = IRB.CreateShl(V, ShAmt, Name + ".shift"); 2200 DEBUG(dbgs() << " shifted: " << *V << "\n"); 2201 } 2202 2203 if (ShAmt || Ty->getBitWidth() < IntTy->getBitWidth()) { 2204 APInt Mask = ~Ty->getMask().zext(IntTy->getBitWidth()).shl(ShAmt); 2205 Old = IRB.CreateAnd(Old, Mask, Name + ".mask"); 2206 DEBUG(dbgs() << " masked: " << *Old << "\n"); 2207 V = IRB.CreateOr(Old, V, Name + ".insert"); 2208 DEBUG(dbgs() << " inserted: " << *V << "\n"); 2209 } 2210 return V; 2211} 2212 2213static Value *extractVector(IRBuilder<> &IRB, Value *V, 2214 unsigned BeginIndex, unsigned EndIndex, 2215 const Twine &Name) { 2216 VectorType *VecTy = cast<VectorType>(V->getType()); 2217 unsigned NumElements = EndIndex - BeginIndex; 2218 assert(NumElements <= VecTy->getNumElements() && "Too many elements!"); 2219 2220 if (NumElements == VecTy->getNumElements()) 2221 return V; 2222 2223 if (NumElements == 1) { 2224 V = IRB.CreateExtractElement(V, IRB.getInt32(BeginIndex), 2225 Name + ".extract"); 2226 DEBUG(dbgs() << " extract: " << *V << "\n"); 2227 return V; 2228 } 2229 2230 SmallVector<Constant*, 8> Mask; 2231 Mask.reserve(NumElements); 2232 for (unsigned i = BeginIndex; i != EndIndex; ++i) 2233 Mask.push_back(IRB.getInt32(i)); 2234 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()), 2235 ConstantVector::get(Mask), 2236 Name + ".extract"); 2237 DEBUG(dbgs() << " shuffle: " << *V << "\n"); 2238 return V; 2239} 2240 2241static Value *insertVector(IRBuilder<> &IRB, Value *Old, Value *V, 2242 unsigned BeginIndex, const Twine &Name) { 2243 VectorType *VecTy = cast<VectorType>(Old->getType()); 2244 assert(VecTy && "Can only insert a vector into a vector"); 2245 2246 VectorType *Ty = dyn_cast<VectorType>(V->getType()); 2247 if (!Ty) { 2248 // Single element to insert. 2249 V = IRB.CreateInsertElement(Old, V, IRB.getInt32(BeginIndex), 2250 Name + ".insert"); 2251 DEBUG(dbgs() << " insert: " << *V << "\n"); 2252 return V; 2253 } 2254 2255 assert(Ty->getNumElements() <= VecTy->getNumElements() && 2256 "Too many elements!"); 2257 if (Ty->getNumElements() == VecTy->getNumElements()) { 2258 assert(V->getType() == VecTy && "Vector type mismatch"); 2259 return V; 2260 } 2261 unsigned EndIndex = BeginIndex + Ty->getNumElements(); 2262 2263 // When inserting a smaller vector into the larger to store, we first 2264 // use a shuffle vector to widen it with undef elements, and then 2265 // a second shuffle vector to select between the loaded vector and the 2266 // incoming vector. 2267 SmallVector<Constant*, 8> Mask; 2268 Mask.reserve(VecTy->getNumElements()); 2269 for (unsigned i = 0; i != VecTy->getNumElements(); ++i) 2270 if (i >= BeginIndex && i < EndIndex) 2271 Mask.push_back(IRB.getInt32(i - BeginIndex)); 2272 else 2273 Mask.push_back(UndefValue::get(IRB.getInt32Ty())); 2274 V = IRB.CreateShuffleVector(V, UndefValue::get(V->getType()), 2275 ConstantVector::get(Mask), 2276 Name + ".expand"); 2277 DEBUG(dbgs() << " shuffle1: " << *V << "\n"); 2278 2279 Mask.clear(); 2280 for (unsigned i = 0; i != VecTy->getNumElements(); ++i) 2281 if (i >= BeginIndex && i < EndIndex) 2282 Mask.push_back(IRB.getInt32(i)); 2283 else 2284 Mask.push_back(IRB.getInt32(i + VecTy->getNumElements())); 2285 V = IRB.CreateShuffleVector(V, Old, ConstantVector::get(Mask), 2286 Name + "insert"); 2287 DEBUG(dbgs() << " shuffle2: " << *V << "\n"); 2288 return V; 2289} 2290 2291namespace { 2292/// \brief Visitor to rewrite instructions using a partition of an alloca to 2293/// use a new alloca. 2294/// 2295/// Also implements the rewriting to vector-based accesses when the partition 2296/// passes the isVectorPromotionViable predicate. Most of the rewriting logic 2297/// lives here. 2298class AllocaPartitionRewriter : public InstVisitor<AllocaPartitionRewriter, 2299 bool> { 2300 // Befriend the base class so it can delegate to private visit methods. 2301 friend class llvm::InstVisitor<AllocaPartitionRewriter, bool>; 2302 2303 const DataLayout &TD; 2304 AllocaPartitioning &P; 2305 SROA &Pass; 2306 AllocaInst &OldAI, &NewAI; 2307 const uint64_t NewAllocaBeginOffset, NewAllocaEndOffset; 2308 Type *NewAllocaTy; 2309 2310 // If we are rewriting an alloca partition which can be written as pure 2311 // vector operations, we stash extra information here. When VecTy is 2312 // non-null, we have some strict guarantees about the rewritten alloca: 2313 // - The new alloca is exactly the size of the vector type here. 2314 // - The accesses all either map to the entire vector or to a single 2315 // element. 2316 // - The set of accessing instructions is only one of those handled above 2317 // in isVectorPromotionViable. Generally these are the same access kinds 2318 // which are promotable via mem2reg. 2319 VectorType *VecTy; 2320 Type *ElementTy; 2321 uint64_t ElementSize; 2322 2323 // This is a convenience and flag variable that will be null unless the new 2324 // alloca's integer operations should be widened to this integer type due to 2325 // passing isIntegerWideningViable above. If it is non-null, the desired 2326 // integer type will be stored here for easy access during rewriting. 2327 IntegerType *IntTy; 2328 2329 // The offset of the partition user currently being rewritten. 2330 uint64_t BeginOffset, EndOffset; 2331 bool IsSplit; 2332 Use *OldUse; 2333 Instruction *OldPtr; 2334 2335 // The name prefix to use when rewriting instructions for this alloca. 2336 std::string NamePrefix; 2337 2338public: 2339 AllocaPartitionRewriter(const DataLayout &TD, AllocaPartitioning &P, 2340 AllocaPartitioning::iterator PI, 2341 SROA &Pass, AllocaInst &OldAI, AllocaInst &NewAI, 2342 uint64_t NewBeginOffset, uint64_t NewEndOffset) 2343 : TD(TD), P(P), Pass(Pass), 2344 OldAI(OldAI), NewAI(NewAI), 2345 NewAllocaBeginOffset(NewBeginOffset), 2346 NewAllocaEndOffset(NewEndOffset), 2347 NewAllocaTy(NewAI.getAllocatedType()), 2348 VecTy(), ElementTy(), ElementSize(), IntTy(), 2349 BeginOffset(), EndOffset(), IsSplit(), OldUse(), OldPtr() { 2350 } 2351 2352 /// \brief Visit the users of the alloca partition and rewrite them. 2353 bool visitUsers(AllocaPartitioning::const_use_iterator I, 2354 AllocaPartitioning::const_use_iterator E) { 2355 if (isVectorPromotionViable(TD, NewAI.getAllocatedType(), P, 2356 NewAllocaBeginOffset, NewAllocaEndOffset, 2357 I, E)) { 2358 ++NumVectorized; 2359 VecTy = cast<VectorType>(NewAI.getAllocatedType()); 2360 ElementTy = VecTy->getElementType(); 2361 assert((TD.getTypeSizeInBits(VecTy->getScalarType()) % 8) == 0 && 2362 "Only multiple-of-8 sized vector elements are viable"); 2363 ElementSize = TD.getTypeSizeInBits(VecTy->getScalarType()) / 8; 2364 } else if (isIntegerWideningViable(TD, NewAI.getAllocatedType(), 2365 NewAllocaBeginOffset, P, I, E)) { 2366 IntTy = Type::getIntNTy(NewAI.getContext(), 2367 TD.getTypeSizeInBits(NewAI.getAllocatedType())); 2368 } 2369 bool CanSROA = true; 2370 for (; I != E; ++I) { 2371 if (!I->getUse()) 2372 continue; // Skip dead uses. 2373 BeginOffset = I->BeginOffset; 2374 EndOffset = I->EndOffset; 2375 IsSplit = I->isSplit(); 2376 OldUse = I->getUse(); 2377 OldPtr = cast<Instruction>(OldUse->get()); 2378 NamePrefix = (Twine(NewAI.getName()) + "." + Twine(BeginOffset)).str(); 2379 CanSROA &= visit(cast<Instruction>(OldUse->getUser())); 2380 } 2381 if (VecTy) { 2382 assert(CanSROA); 2383 VecTy = 0; 2384 ElementTy = 0; 2385 ElementSize = 0; 2386 } 2387 if (IntTy) { 2388 assert(CanSROA); 2389 IntTy = 0; 2390 } 2391 return CanSROA; 2392 } 2393 2394private: 2395 // Every instruction which can end up as a user must have a rewrite rule. 2396 bool visitInstruction(Instruction &I) { 2397 DEBUG(dbgs() << " !!!! Cannot rewrite: " << I << "\n"); 2398 llvm_unreachable("No rewrite rule for this instruction!"); 2399 } 2400 2401 Twine getName(const Twine &Suffix) { 2402 return NamePrefix + Suffix; 2403 } 2404 2405 Value *getAdjustedAllocaPtr(IRBuilder<> &IRB, Type *PointerTy) { 2406 assert(BeginOffset >= NewAllocaBeginOffset); 2407 APInt Offset(TD.getPointerSizeInBits(), BeginOffset - NewAllocaBeginOffset); 2408 return getAdjustedPtr(IRB, TD, &NewAI, Offset, PointerTy, getName("")); 2409 } 2410 2411 /// \brief Compute suitable alignment to access an offset into the new alloca. 2412 unsigned getOffsetAlign(uint64_t Offset) { 2413 unsigned NewAIAlign = NewAI.getAlignment(); 2414 if (!NewAIAlign) 2415 NewAIAlign = TD.getABITypeAlignment(NewAI.getAllocatedType()); 2416 return MinAlign(NewAIAlign, Offset); 2417 } 2418 2419 /// \brief Compute suitable alignment to access this partition of the new 2420 /// alloca. 2421 unsigned getPartitionAlign() { 2422 return getOffsetAlign(BeginOffset - NewAllocaBeginOffset); 2423 } 2424 2425 /// \brief Compute suitable alignment to access a type at an offset of the 2426 /// new alloca. 2427 /// 2428 /// \returns zero if the type's ABI alignment is a suitable alignment, 2429 /// otherwise returns the maximal suitable alignment. 2430 unsigned getOffsetTypeAlign(Type *Ty, uint64_t Offset) { 2431 unsigned Align = getOffsetAlign(Offset); 2432 return Align == TD.getABITypeAlignment(Ty) ? 0 : Align; 2433 } 2434 2435 /// \brief Compute suitable alignment to access a type at the beginning of 2436 /// this partition of the new alloca. 2437 /// 2438 /// See \c getOffsetTypeAlign for details; this routine delegates to it. 2439 unsigned getPartitionTypeAlign(Type *Ty) { 2440 return getOffsetTypeAlign(Ty, BeginOffset - NewAllocaBeginOffset); 2441 } 2442 2443 unsigned getIndex(uint64_t Offset) { 2444 assert(VecTy && "Can only call getIndex when rewriting a vector"); 2445 uint64_t RelOffset = Offset - NewAllocaBeginOffset; 2446 assert(RelOffset / ElementSize < UINT32_MAX && "Index out of bounds"); 2447 uint32_t Index = RelOffset / ElementSize; 2448 assert(Index * ElementSize == RelOffset); 2449 return Index; 2450 } 2451 2452 void deleteIfTriviallyDead(Value *V) { 2453 Instruction *I = cast<Instruction>(V); 2454 if (isInstructionTriviallyDead(I)) 2455 Pass.DeadInsts.insert(I); 2456 } 2457 2458 Value *rewriteVectorizedLoadInst(IRBuilder<> &IRB) { 2459 unsigned BeginIndex = getIndex(BeginOffset); 2460 unsigned EndIndex = getIndex(EndOffset); 2461 assert(EndIndex > BeginIndex && "Empty vector!"); 2462 2463 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), 2464 getName(".load")); 2465 return extractVector(IRB, V, BeginIndex, EndIndex, getName(".vec")); 2466 } 2467 2468 Value *rewriteIntegerLoad(IRBuilder<> &IRB, LoadInst &LI) { 2469 assert(IntTy && "We cannot insert an integer to the alloca"); 2470 assert(!LI.isVolatile()); 2471 Value *V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), 2472 getName(".load")); 2473 V = convertValue(TD, IRB, V, IntTy); 2474 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset"); 2475 uint64_t Offset = BeginOffset - NewAllocaBeginOffset; 2476 if (Offset > 0 || EndOffset < NewAllocaEndOffset) 2477 V = extractInteger(TD, IRB, V, cast<IntegerType>(LI.getType()), Offset, 2478 getName(".extract")); 2479 return V; 2480 } 2481 2482 bool visitLoadInst(LoadInst &LI) { 2483 DEBUG(dbgs() << " original: " << LI << "\n"); 2484 Value *OldOp = LI.getOperand(0); 2485 assert(OldOp == OldPtr); 2486 2487 uint64_t Size = EndOffset - BeginOffset; 2488 2489 IRBuilder<> IRB(&LI); 2490 Type *TargetTy = IsSplit ? Type::getIntNTy(LI.getContext(), Size * 8) 2491 : LI.getType(); 2492 bool IsPtrAdjusted = false; 2493 Value *V; 2494 if (VecTy) { 2495 V = rewriteVectorizedLoadInst(IRB); 2496 } else if (IntTy && LI.getType()->isIntegerTy()) { 2497 V = rewriteIntegerLoad(IRB, LI); 2498 } else if (BeginOffset == NewAllocaBeginOffset && 2499 canConvertValue(TD, NewAllocaTy, LI.getType())) { 2500 V = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), 2501 LI.isVolatile(), getName(".load")); 2502 } else { 2503 Type *LTy = TargetTy->getPointerTo(); 2504 V = IRB.CreateAlignedLoad(getAdjustedAllocaPtr(IRB, LTy), 2505 getPartitionTypeAlign(TargetTy), 2506 LI.isVolatile(), getName(".load")); 2507 IsPtrAdjusted = true; 2508 } 2509 V = convertValue(TD, IRB, V, TargetTy); 2510 2511 if (IsSplit) { 2512 assert(!LI.isVolatile()); 2513 assert(LI.getType()->isIntegerTy() && 2514 "Only integer type loads and stores are split"); 2515 assert(Size < TD.getTypeStoreSize(LI.getType()) && 2516 "Split load isn't smaller than original load"); 2517 assert(LI.getType()->getIntegerBitWidth() == 2518 TD.getTypeStoreSizeInBits(LI.getType()) && 2519 "Non-byte-multiple bit width"); 2520 // Move the insertion point just past the load so that we can refer to it. 2521 IRB.SetInsertPoint(llvm::next(BasicBlock::iterator(&LI))); 2522 // Create a placeholder value with the same type as LI to use as the 2523 // basis for the new value. This allows us to replace the uses of LI with 2524 // the computed value, and then replace the placeholder with LI, leaving 2525 // LI only used for this computation. 2526 Value *Placeholder 2527 = new LoadInst(UndefValue::get(LI.getType()->getPointerTo())); 2528 V = insertInteger(TD, IRB, Placeholder, V, BeginOffset, 2529 getName(".insert")); 2530 LI.replaceAllUsesWith(V); 2531 Placeholder->replaceAllUsesWith(&LI); 2532 delete Placeholder; 2533 } else { 2534 LI.replaceAllUsesWith(V); 2535 } 2536 2537 Pass.DeadInsts.insert(&LI); 2538 deleteIfTriviallyDead(OldOp); 2539 DEBUG(dbgs() << " to: " << *V << "\n"); 2540 return !LI.isVolatile() && !IsPtrAdjusted; 2541 } 2542 2543 bool rewriteVectorizedStoreInst(IRBuilder<> &IRB, Value *V, 2544 StoreInst &SI, Value *OldOp) { 2545 unsigned BeginIndex = getIndex(BeginOffset); 2546 unsigned EndIndex = getIndex(EndOffset); 2547 assert(EndIndex > BeginIndex && "Empty vector!"); 2548 unsigned NumElements = EndIndex - BeginIndex; 2549 assert(NumElements <= VecTy->getNumElements() && "Too many elements!"); 2550 Type *PartitionTy 2551 = (NumElements == 1) ? ElementTy 2552 : VectorType::get(ElementTy, NumElements); 2553 if (V->getType() != PartitionTy) 2554 V = convertValue(TD, IRB, V, PartitionTy); 2555 2556 // Mix in the existing elements. 2557 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), 2558 getName(".load")); 2559 V = insertVector(IRB, Old, V, BeginIndex, getName(".vec")); 2560 2561 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment()); 2562 Pass.DeadInsts.insert(&SI); 2563 2564 (void)Store; 2565 DEBUG(dbgs() << " to: " << *Store << "\n"); 2566 return true; 2567 } 2568 2569 bool rewriteIntegerStore(IRBuilder<> &IRB, Value *V, StoreInst &SI) { 2570 assert(IntTy && "We cannot extract an integer from the alloca"); 2571 assert(!SI.isVolatile()); 2572 if (TD.getTypeSizeInBits(V->getType()) != IntTy->getBitWidth()) { 2573 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), 2574 getName(".oldload")); 2575 Old = convertValue(TD, IRB, Old, IntTy); 2576 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset"); 2577 uint64_t Offset = BeginOffset - NewAllocaBeginOffset; 2578 V = insertInteger(TD, IRB, Old, SI.getValueOperand(), Offset, 2579 getName(".insert")); 2580 } 2581 V = convertValue(TD, IRB, V, NewAllocaTy); 2582 StoreInst *Store = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment()); 2583 Pass.DeadInsts.insert(&SI); 2584 (void)Store; 2585 DEBUG(dbgs() << " to: " << *Store << "\n"); 2586 return true; 2587 } 2588 2589 bool visitStoreInst(StoreInst &SI) { 2590 DEBUG(dbgs() << " original: " << SI << "\n"); 2591 Value *OldOp = SI.getOperand(1); 2592 assert(OldOp == OldPtr); 2593 IRBuilder<> IRB(&SI); 2594 2595 Value *V = SI.getValueOperand(); 2596 2597 // Strip all inbounds GEPs and pointer casts to try to dig out any root 2598 // alloca that should be re-examined after promoting this alloca. 2599 if (V->getType()->isPointerTy()) 2600 if (AllocaInst *AI = dyn_cast<AllocaInst>(V->stripInBoundsOffsets())) 2601 Pass.PostPromotionWorklist.insert(AI); 2602 2603 uint64_t Size = EndOffset - BeginOffset; 2604 if (Size < TD.getTypeStoreSize(V->getType())) { 2605 assert(!SI.isVolatile()); 2606 assert(IsSplit && "A seemingly split store isn't splittable"); 2607 assert(V->getType()->isIntegerTy() && 2608 "Only integer type loads and stores are split"); 2609 assert(V->getType()->getIntegerBitWidth() == 2610 TD.getTypeStoreSizeInBits(V->getType()) && 2611 "Non-byte-multiple bit width"); 2612 IntegerType *NarrowTy = Type::getIntNTy(SI.getContext(), Size * 8); 2613 V = extractInteger(TD, IRB, V, NarrowTy, BeginOffset, 2614 getName(".extract")); 2615 } 2616 2617 if (VecTy) 2618 return rewriteVectorizedStoreInst(IRB, V, SI, OldOp); 2619 if (IntTy && V->getType()->isIntegerTy()) 2620 return rewriteIntegerStore(IRB, V, SI); 2621 2622 StoreInst *NewSI; 2623 if (BeginOffset == NewAllocaBeginOffset && 2624 canConvertValue(TD, V->getType(), NewAllocaTy)) { 2625 V = convertValue(TD, IRB, V, NewAllocaTy); 2626 NewSI = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(), 2627 SI.isVolatile()); 2628 } else { 2629 Value *NewPtr = getAdjustedAllocaPtr(IRB, V->getType()->getPointerTo()); 2630 NewSI = IRB.CreateAlignedStore(V, NewPtr, 2631 getPartitionTypeAlign(V->getType()), 2632 SI.isVolatile()); 2633 } 2634 (void)NewSI; 2635 Pass.DeadInsts.insert(&SI); 2636 deleteIfTriviallyDead(OldOp); 2637 2638 DEBUG(dbgs() << " to: " << *NewSI << "\n"); 2639 return NewSI->getPointerOperand() == &NewAI && !SI.isVolatile(); 2640 } 2641 2642 /// \brief Compute an integer value from splatting an i8 across the given 2643 /// number of bytes. 2644 /// 2645 /// Note that this routine assumes an i8 is a byte. If that isn't true, don't 2646 /// call this routine. 2647 /// FIXME: Heed the advice above. 2648 /// 2649 /// \param V The i8 value to splat. 2650 /// \param Size The number of bytes in the output (assuming i8 is one byte) 2651 Value *getIntegerSplat(IRBuilder<> &IRB, Value *V, unsigned Size) { 2652 assert(Size > 0 && "Expected a positive number of bytes."); 2653 IntegerType *VTy = cast<IntegerType>(V->getType()); 2654 assert(VTy->getBitWidth() == 8 && "Expected an i8 value for the byte"); 2655 if (Size == 1) 2656 return V; 2657 2658 Type *SplatIntTy = Type::getIntNTy(VTy->getContext(), Size*8); 2659 V = IRB.CreateMul(IRB.CreateZExt(V, SplatIntTy, getName(".zext")), 2660 ConstantExpr::getUDiv( 2661 Constant::getAllOnesValue(SplatIntTy), 2662 ConstantExpr::getZExt( 2663 Constant::getAllOnesValue(V->getType()), 2664 SplatIntTy)), 2665 getName(".isplat")); 2666 return V; 2667 } 2668 2669 /// \brief Compute a vector splat for a given element value. 2670 Value *getVectorSplat(IRBuilder<> &IRB, Value *V, unsigned NumElements) { 2671 V = IRB.CreateVectorSplat(NumElements, V, NamePrefix); 2672 DEBUG(dbgs() << " splat: " << *V << "\n"); 2673 return V; 2674 } 2675 2676 bool visitMemSetInst(MemSetInst &II) { 2677 DEBUG(dbgs() << " original: " << II << "\n"); 2678 IRBuilder<> IRB(&II); 2679 assert(II.getRawDest() == OldPtr); 2680 2681 // If the memset has a variable size, it cannot be split, just adjust the 2682 // pointer to the new alloca. 2683 if (!isa<Constant>(II.getLength())) { 2684 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType())); 2685 Type *CstTy = II.getAlignmentCst()->getType(); 2686 II.setAlignment(ConstantInt::get(CstTy, getPartitionAlign())); 2687 2688 deleteIfTriviallyDead(OldPtr); 2689 return false; 2690 } 2691 2692 // Record this instruction for deletion. 2693 Pass.DeadInsts.insert(&II); 2694 2695 Type *AllocaTy = NewAI.getAllocatedType(); 2696 Type *ScalarTy = AllocaTy->getScalarType(); 2697 2698 // If this doesn't map cleanly onto the alloca type, and that type isn't 2699 // a single value type, just emit a memset. 2700 if (!VecTy && !IntTy && 2701 (BeginOffset != NewAllocaBeginOffset || 2702 EndOffset != NewAllocaEndOffset || 2703 !AllocaTy->isSingleValueType() || 2704 !TD.isLegalInteger(TD.getTypeSizeInBits(ScalarTy)) || 2705 TD.getTypeSizeInBits(ScalarTy)%8 != 0)) { 2706 Type *SizeTy = II.getLength()->getType(); 2707 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset); 2708 CallInst *New 2709 = IRB.CreateMemSet(getAdjustedAllocaPtr(IRB, 2710 II.getRawDest()->getType()), 2711 II.getValue(), Size, getPartitionAlign(), 2712 II.isVolatile()); 2713 (void)New; 2714 DEBUG(dbgs() << " to: " << *New << "\n"); 2715 return false; 2716 } 2717 2718 // If we can represent this as a simple value, we have to build the actual 2719 // value to store, which requires expanding the byte present in memset to 2720 // a sensible representation for the alloca type. This is essentially 2721 // splatting the byte to a sufficiently wide integer, splatting it across 2722 // any desired vector width, and bitcasting to the final type. 2723 Value *V; 2724 2725 if (VecTy) { 2726 // If this is a memset of a vectorized alloca, insert it. 2727 assert(ElementTy == ScalarTy); 2728 2729 unsigned BeginIndex = getIndex(BeginOffset); 2730 unsigned EndIndex = getIndex(EndOffset); 2731 assert(EndIndex > BeginIndex && "Empty vector!"); 2732 unsigned NumElements = EndIndex - BeginIndex; 2733 assert(NumElements <= VecTy->getNumElements() && "Too many elements!"); 2734 2735 Value *Splat = getIntegerSplat(IRB, II.getValue(), 2736 TD.getTypeSizeInBits(ElementTy)/8); 2737 Splat = convertValue(TD, IRB, Splat, ElementTy); 2738 if (NumElements > 1) 2739 Splat = getVectorSplat(IRB, Splat, NumElements); 2740 2741 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), 2742 getName(".oldload")); 2743 V = insertVector(IRB, Old, Splat, BeginIndex, getName(".vec")); 2744 } else if (IntTy) { 2745 // If this is a memset on an alloca where we can widen stores, insert the 2746 // set integer. 2747 assert(!II.isVolatile()); 2748 2749 uint64_t Size = EndOffset - BeginOffset; 2750 V = getIntegerSplat(IRB, II.getValue(), Size); 2751 2752 if (IntTy && (BeginOffset != NewAllocaBeginOffset || 2753 EndOffset != NewAllocaBeginOffset)) { 2754 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), 2755 getName(".oldload")); 2756 Old = convertValue(TD, IRB, Old, IntTy); 2757 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset"); 2758 uint64_t Offset = BeginOffset - NewAllocaBeginOffset; 2759 V = insertInteger(TD, IRB, Old, V, Offset, getName(".insert")); 2760 } else { 2761 assert(V->getType() == IntTy && 2762 "Wrong type for an alloca wide integer!"); 2763 } 2764 V = convertValue(TD, IRB, V, AllocaTy); 2765 } else { 2766 // Established these invariants above. 2767 assert(BeginOffset == NewAllocaBeginOffset); 2768 assert(EndOffset == NewAllocaEndOffset); 2769 2770 V = getIntegerSplat(IRB, II.getValue(), 2771 TD.getTypeSizeInBits(ScalarTy)/8); 2772 if (VectorType *AllocaVecTy = dyn_cast<VectorType>(AllocaTy)) 2773 V = getVectorSplat(IRB, V, AllocaVecTy->getNumElements()); 2774 2775 V = convertValue(TD, IRB, V, AllocaTy); 2776 } 2777 2778 Value *New = IRB.CreateAlignedStore(V, &NewAI, NewAI.getAlignment(), 2779 II.isVolatile()); 2780 (void)New; 2781 DEBUG(dbgs() << " to: " << *New << "\n"); 2782 return !II.isVolatile(); 2783 } 2784 2785 bool visitMemTransferInst(MemTransferInst &II) { 2786 // Rewriting of memory transfer instructions can be a bit tricky. We break 2787 // them into two categories: split intrinsics and unsplit intrinsics. 2788 2789 DEBUG(dbgs() << " original: " << II << "\n"); 2790 IRBuilder<> IRB(&II); 2791 2792 assert(II.getRawSource() == OldPtr || II.getRawDest() == OldPtr); 2793 bool IsDest = II.getRawDest() == OldPtr; 2794 2795 const AllocaPartitioning::MemTransferOffsets &MTO 2796 = P.getMemTransferOffsets(II); 2797 2798 // Compute the relative offset within the transfer. 2799 unsigned IntPtrWidth = TD.getPointerSizeInBits(); 2800 APInt RelOffset(IntPtrWidth, BeginOffset - (IsDest ? MTO.DestBegin 2801 : MTO.SourceBegin)); 2802 2803 unsigned Align = II.getAlignment(); 2804 if (Align > 1) 2805 Align = MinAlign(RelOffset.zextOrTrunc(64).getZExtValue(), 2806 MinAlign(II.getAlignment(), getPartitionAlign())); 2807 2808 // For unsplit intrinsics, we simply modify the source and destination 2809 // pointers in place. This isn't just an optimization, it is a matter of 2810 // correctness. With unsplit intrinsics we may be dealing with transfers 2811 // within a single alloca before SROA ran, or with transfers that have 2812 // a variable length. We may also be dealing with memmove instead of 2813 // memcpy, and so simply updating the pointers is the necessary for us to 2814 // update both source and dest of a single call. 2815 if (!MTO.IsSplittable) { 2816 Value *OldOp = IsDest ? II.getRawDest() : II.getRawSource(); 2817 if (IsDest) 2818 II.setDest(getAdjustedAllocaPtr(IRB, II.getRawDest()->getType())); 2819 else 2820 II.setSource(getAdjustedAllocaPtr(IRB, II.getRawSource()->getType())); 2821 2822 Type *CstTy = II.getAlignmentCst()->getType(); 2823 II.setAlignment(ConstantInt::get(CstTy, Align)); 2824 2825 DEBUG(dbgs() << " to: " << II << "\n"); 2826 deleteIfTriviallyDead(OldOp); 2827 return false; 2828 } 2829 // For split transfer intrinsics we have an incredibly useful assurance: 2830 // the source and destination do not reside within the same alloca, and at 2831 // least one of them does not escape. This means that we can replace 2832 // memmove with memcpy, and we don't need to worry about all manner of 2833 // downsides to splitting and transforming the operations. 2834 2835 // If this doesn't map cleanly onto the alloca type, and that type isn't 2836 // a single value type, just emit a memcpy. 2837 bool EmitMemCpy 2838 = !VecTy && !IntTy && (BeginOffset != NewAllocaBeginOffset || 2839 EndOffset != NewAllocaEndOffset || 2840 !NewAI.getAllocatedType()->isSingleValueType()); 2841 2842 // If we're just going to emit a memcpy, the alloca hasn't changed, and the 2843 // size hasn't been shrunk based on analysis of the viable range, this is 2844 // a no-op. 2845 if (EmitMemCpy && &OldAI == &NewAI) { 2846 uint64_t OrigBegin = IsDest ? MTO.DestBegin : MTO.SourceBegin; 2847 uint64_t OrigEnd = IsDest ? MTO.DestEnd : MTO.SourceEnd; 2848 // Ensure the start lines up. 2849 assert(BeginOffset == OrigBegin); 2850 (void)OrigBegin; 2851 2852 // Rewrite the size as needed. 2853 if (EndOffset != OrigEnd) 2854 II.setLength(ConstantInt::get(II.getLength()->getType(), 2855 EndOffset - BeginOffset)); 2856 return false; 2857 } 2858 // Record this instruction for deletion. 2859 Pass.DeadInsts.insert(&II); 2860 2861 // Strip all inbounds GEPs and pointer casts to try to dig out any root 2862 // alloca that should be re-examined after rewriting this instruction. 2863 Value *OtherPtr = IsDest ? II.getRawSource() : II.getRawDest(); 2864 if (AllocaInst *AI 2865 = dyn_cast<AllocaInst>(OtherPtr->stripInBoundsOffsets())) 2866 Pass.Worklist.insert(AI); 2867 2868 if (EmitMemCpy) { 2869 Type *OtherPtrTy = IsDest ? II.getRawSource()->getType() 2870 : II.getRawDest()->getType(); 2871 2872 // Compute the other pointer, folding as much as possible to produce 2873 // a single, simple GEP in most cases. 2874 OtherPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy, 2875 getName("." + OtherPtr->getName())); 2876 2877 Value *OurPtr 2878 = getAdjustedAllocaPtr(IRB, IsDest ? II.getRawDest()->getType() 2879 : II.getRawSource()->getType()); 2880 Type *SizeTy = II.getLength()->getType(); 2881 Constant *Size = ConstantInt::get(SizeTy, EndOffset - BeginOffset); 2882 2883 CallInst *New = IRB.CreateMemCpy(IsDest ? OurPtr : OtherPtr, 2884 IsDest ? OtherPtr : OurPtr, 2885 Size, Align, II.isVolatile()); 2886 (void)New; 2887 DEBUG(dbgs() << " to: " << *New << "\n"); 2888 return false; 2889 } 2890 2891 // Note that we clamp the alignment to 1 here as a 0 alignment for a memcpy 2892 // is equivalent to 1, but that isn't true if we end up rewriting this as 2893 // a load or store. 2894 if (!Align) 2895 Align = 1; 2896 2897 bool IsWholeAlloca = BeginOffset == NewAllocaBeginOffset && 2898 EndOffset == NewAllocaEndOffset; 2899 uint64_t Size = EndOffset - BeginOffset; 2900 unsigned BeginIndex = VecTy ? getIndex(BeginOffset) : 0; 2901 unsigned EndIndex = VecTy ? getIndex(EndOffset) : 0; 2902 unsigned NumElements = EndIndex - BeginIndex; 2903 IntegerType *SubIntTy 2904 = IntTy ? Type::getIntNTy(IntTy->getContext(), Size*8) : 0; 2905 2906 Type *OtherPtrTy = NewAI.getType(); 2907 if (VecTy && !IsWholeAlloca) { 2908 if (NumElements == 1) 2909 OtherPtrTy = VecTy->getElementType(); 2910 else 2911 OtherPtrTy = VectorType::get(VecTy->getElementType(), NumElements); 2912 2913 OtherPtrTy = OtherPtrTy->getPointerTo(); 2914 } else if (IntTy && !IsWholeAlloca) { 2915 OtherPtrTy = SubIntTy->getPointerTo(); 2916 } 2917 2918 Value *SrcPtr = getAdjustedPtr(IRB, TD, OtherPtr, RelOffset, OtherPtrTy, 2919 getName("." + OtherPtr->getName())); 2920 Value *DstPtr = &NewAI; 2921 if (!IsDest) 2922 std::swap(SrcPtr, DstPtr); 2923 2924 Value *Src; 2925 if (VecTy && !IsWholeAlloca && !IsDest) { 2926 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), 2927 getName(".load")); 2928 Src = extractVector(IRB, Src, BeginIndex, EndIndex, getName(".vec")); 2929 } else if (IntTy && !IsWholeAlloca && !IsDest) { 2930 Src = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), 2931 getName(".load")); 2932 Src = convertValue(TD, IRB, Src, IntTy); 2933 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset"); 2934 uint64_t Offset = BeginOffset - NewAllocaBeginOffset; 2935 Src = extractInteger(TD, IRB, Src, SubIntTy, Offset, getName(".extract")); 2936 } else { 2937 Src = IRB.CreateAlignedLoad(SrcPtr, Align, II.isVolatile(), 2938 getName(".copyload")); 2939 } 2940 2941 if (VecTy && !IsWholeAlloca && IsDest) { 2942 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), 2943 getName(".oldload")); 2944 Src = insertVector(IRB, Old, Src, BeginIndex, getName(".vec")); 2945 } else if (IntTy && !IsWholeAlloca && IsDest) { 2946 Value *Old = IRB.CreateAlignedLoad(&NewAI, NewAI.getAlignment(), 2947 getName(".oldload")); 2948 Old = convertValue(TD, IRB, Old, IntTy); 2949 assert(BeginOffset >= NewAllocaBeginOffset && "Out of bounds offset"); 2950 uint64_t Offset = BeginOffset - NewAllocaBeginOffset; 2951 Src = insertInteger(TD, IRB, Old, Src, Offset, getName(".insert")); 2952 Src = convertValue(TD, IRB, Src, NewAllocaTy); 2953 } 2954 2955 StoreInst *Store = cast<StoreInst>( 2956 IRB.CreateAlignedStore(Src, DstPtr, Align, II.isVolatile())); 2957 (void)Store; 2958 DEBUG(dbgs() << " to: " << *Store << "\n"); 2959 return !II.isVolatile(); 2960 } 2961 2962 bool visitIntrinsicInst(IntrinsicInst &II) { 2963 assert(II.getIntrinsicID() == Intrinsic::lifetime_start || 2964 II.getIntrinsicID() == Intrinsic::lifetime_end); 2965 DEBUG(dbgs() << " original: " << II << "\n"); 2966 IRBuilder<> IRB(&II); 2967 assert(II.getArgOperand(1) == OldPtr); 2968 2969 // Record this instruction for deletion. 2970 Pass.DeadInsts.insert(&II); 2971 2972 ConstantInt *Size 2973 = ConstantInt::get(cast<IntegerType>(II.getArgOperand(0)->getType()), 2974 EndOffset - BeginOffset); 2975 Value *Ptr = getAdjustedAllocaPtr(IRB, II.getArgOperand(1)->getType()); 2976 Value *New; 2977 if (II.getIntrinsicID() == Intrinsic::lifetime_start) 2978 New = IRB.CreateLifetimeStart(Ptr, Size); 2979 else 2980 New = IRB.CreateLifetimeEnd(Ptr, Size); 2981 2982 (void)New; 2983 DEBUG(dbgs() << " to: " << *New << "\n"); 2984 return true; 2985 } 2986 2987 bool visitPHINode(PHINode &PN) { 2988 DEBUG(dbgs() << " original: " << PN << "\n"); 2989 2990 // We would like to compute a new pointer in only one place, but have it be 2991 // as local as possible to the PHI. To do that, we re-use the location of 2992 // the old pointer, which necessarily must be in the right position to 2993 // dominate the PHI. 2994 IRBuilder<> PtrBuilder(cast<Instruction>(OldPtr)); 2995 2996 Value *NewPtr = getAdjustedAllocaPtr(PtrBuilder, OldPtr->getType()); 2997 // Replace the operands which were using the old pointer. 2998 std::replace(PN.op_begin(), PN.op_end(), cast<Value>(OldPtr), NewPtr); 2999 3000 DEBUG(dbgs() << " to: " << PN << "\n"); 3001 deleteIfTriviallyDead(OldPtr); 3002 return false; 3003 } 3004 3005 bool visitSelectInst(SelectInst &SI) { 3006 DEBUG(dbgs() << " original: " << SI << "\n"); 3007 IRBuilder<> IRB(&SI); 3008 3009 // Find the operand we need to rewrite here. 3010 bool IsTrueVal = SI.getTrueValue() == OldPtr; 3011 if (IsTrueVal) 3012 assert(SI.getFalseValue() != OldPtr && "Pointer is both operands!"); 3013 else 3014 assert(SI.getFalseValue() == OldPtr && "Pointer isn't an operand!"); 3015 3016 Value *NewPtr = getAdjustedAllocaPtr(IRB, OldPtr->getType()); 3017 SI.setOperand(IsTrueVal ? 1 : 2, NewPtr); 3018 DEBUG(dbgs() << " to: " << SI << "\n"); 3019 deleteIfTriviallyDead(OldPtr); 3020 return false; 3021 } 3022 3023}; 3024} 3025 3026namespace { 3027/// \brief Visitor to rewrite aggregate loads and stores as scalar. 3028/// 3029/// This pass aggressively rewrites all aggregate loads and stores on 3030/// a particular pointer (or any pointer derived from it which we can identify) 3031/// with scalar loads and stores. 3032class AggLoadStoreRewriter : public InstVisitor<AggLoadStoreRewriter, bool> { 3033 // Befriend the base class so it can delegate to private visit methods. 3034 friend class llvm::InstVisitor<AggLoadStoreRewriter, bool>; 3035 3036 const DataLayout &TD; 3037 3038 /// Queue of pointer uses to analyze and potentially rewrite. 3039 SmallVector<Use *, 8> Queue; 3040 3041 /// Set to prevent us from cycling with phi nodes and loops. 3042 SmallPtrSet<User *, 8> Visited; 3043 3044 /// The current pointer use being rewritten. This is used to dig up the used 3045 /// value (as opposed to the user). 3046 Use *U; 3047 3048public: 3049 AggLoadStoreRewriter(const DataLayout &TD) : TD(TD) {} 3050 3051 /// Rewrite loads and stores through a pointer and all pointers derived from 3052 /// it. 3053 bool rewrite(Instruction &I) { 3054 DEBUG(dbgs() << " Rewriting FCA loads and stores...\n"); 3055 enqueueUsers(I); 3056 bool Changed = false; 3057 while (!Queue.empty()) { 3058 U = Queue.pop_back_val(); 3059 Changed |= visit(cast<Instruction>(U->getUser())); 3060 } 3061 return Changed; 3062 } 3063 3064private: 3065 /// Enqueue all the users of the given instruction for further processing. 3066 /// This uses a set to de-duplicate users. 3067 void enqueueUsers(Instruction &I) { 3068 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end(); UI != UE; 3069 ++UI) 3070 if (Visited.insert(*UI)) 3071 Queue.push_back(&UI.getUse()); 3072 } 3073 3074 // Conservative default is to not rewrite anything. 3075 bool visitInstruction(Instruction &I) { return false; } 3076 3077 /// \brief Generic recursive split emission class. 3078 template <typename Derived> 3079 class OpSplitter { 3080 protected: 3081 /// The builder used to form new instructions. 3082 IRBuilder<> IRB; 3083 /// The indices which to be used with insert- or extractvalue to select the 3084 /// appropriate value within the aggregate. 3085 SmallVector<unsigned, 4> Indices; 3086 /// The indices to a GEP instruction which will move Ptr to the correct slot 3087 /// within the aggregate. 3088 SmallVector<Value *, 4> GEPIndices; 3089 /// The base pointer of the original op, used as a base for GEPing the 3090 /// split operations. 3091 Value *Ptr; 3092 3093 /// Initialize the splitter with an insertion point, Ptr and start with a 3094 /// single zero GEP index. 3095 OpSplitter(Instruction *InsertionPoint, Value *Ptr) 3096 : IRB(InsertionPoint), GEPIndices(1, IRB.getInt32(0)), Ptr(Ptr) {} 3097 3098 public: 3099 /// \brief Generic recursive split emission routine. 3100 /// 3101 /// This method recursively splits an aggregate op (load or store) into 3102 /// scalar or vector ops. It splits recursively until it hits a single value 3103 /// and emits that single value operation via the template argument. 3104 /// 3105 /// The logic of this routine relies on GEPs and insertvalue and 3106 /// extractvalue all operating with the same fundamental index list, merely 3107 /// formatted differently (GEPs need actual values). 3108 /// 3109 /// \param Ty The type being split recursively into smaller ops. 3110 /// \param Agg The aggregate value being built up or stored, depending on 3111 /// whether this is splitting a load or a store respectively. 3112 void emitSplitOps(Type *Ty, Value *&Agg, const Twine &Name) { 3113 if (Ty->isSingleValueType()) 3114 return static_cast<Derived *>(this)->emitFunc(Ty, Agg, Name); 3115 3116 if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 3117 unsigned OldSize = Indices.size(); 3118 (void)OldSize; 3119 for (unsigned Idx = 0, Size = ATy->getNumElements(); Idx != Size; 3120 ++Idx) { 3121 assert(Indices.size() == OldSize && "Did not return to the old size"); 3122 Indices.push_back(Idx); 3123 GEPIndices.push_back(IRB.getInt32(Idx)); 3124 emitSplitOps(ATy->getElementType(), Agg, Name + "." + Twine(Idx)); 3125 GEPIndices.pop_back(); 3126 Indices.pop_back(); 3127 } 3128 return; 3129 } 3130 3131 if (StructType *STy = dyn_cast<StructType>(Ty)) { 3132 unsigned OldSize = Indices.size(); 3133 (void)OldSize; 3134 for (unsigned Idx = 0, Size = STy->getNumElements(); Idx != Size; 3135 ++Idx) { 3136 assert(Indices.size() == OldSize && "Did not return to the old size"); 3137 Indices.push_back(Idx); 3138 GEPIndices.push_back(IRB.getInt32(Idx)); 3139 emitSplitOps(STy->getElementType(Idx), Agg, Name + "." + Twine(Idx)); 3140 GEPIndices.pop_back(); 3141 Indices.pop_back(); 3142 } 3143 return; 3144 } 3145 3146 llvm_unreachable("Only arrays and structs are aggregate loadable types"); 3147 } 3148 }; 3149 3150 struct LoadOpSplitter : public OpSplitter<LoadOpSplitter> { 3151 LoadOpSplitter(Instruction *InsertionPoint, Value *Ptr) 3152 : OpSplitter<LoadOpSplitter>(InsertionPoint, Ptr) {} 3153 3154 /// Emit a leaf load of a single value. This is called at the leaves of the 3155 /// recursive emission to actually load values. 3156 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) { 3157 assert(Ty->isSingleValueType()); 3158 // Load the single value and insert it using the indices. 3159 Value *GEP = IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep"); 3160 Value *Load = IRB.CreateLoad(GEP, Name + ".load"); 3161 Agg = IRB.CreateInsertValue(Agg, Load, Indices, Name + ".insert"); 3162 DEBUG(dbgs() << " to: " << *Load << "\n"); 3163 } 3164 }; 3165 3166 bool visitLoadInst(LoadInst &LI) { 3167 assert(LI.getPointerOperand() == *U); 3168 if (!LI.isSimple() || LI.getType()->isSingleValueType()) 3169 return false; 3170 3171 // We have an aggregate being loaded, split it apart. 3172 DEBUG(dbgs() << " original: " << LI << "\n"); 3173 LoadOpSplitter Splitter(&LI, *U); 3174 Value *V = UndefValue::get(LI.getType()); 3175 Splitter.emitSplitOps(LI.getType(), V, LI.getName() + ".fca"); 3176 LI.replaceAllUsesWith(V); 3177 LI.eraseFromParent(); 3178 return true; 3179 } 3180 3181 struct StoreOpSplitter : public OpSplitter<StoreOpSplitter> { 3182 StoreOpSplitter(Instruction *InsertionPoint, Value *Ptr) 3183 : OpSplitter<StoreOpSplitter>(InsertionPoint, Ptr) {} 3184 3185 /// Emit a leaf store of a single value. This is called at the leaves of the 3186 /// recursive emission to actually produce stores. 3187 void emitFunc(Type *Ty, Value *&Agg, const Twine &Name) { 3188 assert(Ty->isSingleValueType()); 3189 // Extract the single value and store it using the indices. 3190 Value *Store = IRB.CreateStore( 3191 IRB.CreateExtractValue(Agg, Indices, Name + ".extract"), 3192 IRB.CreateInBoundsGEP(Ptr, GEPIndices, Name + ".gep")); 3193 (void)Store; 3194 DEBUG(dbgs() << " to: " << *Store << "\n"); 3195 } 3196 }; 3197 3198 bool visitStoreInst(StoreInst &SI) { 3199 if (!SI.isSimple() || SI.getPointerOperand() != *U) 3200 return false; 3201 Value *V = SI.getValueOperand(); 3202 if (V->getType()->isSingleValueType()) 3203 return false; 3204 3205 // We have an aggregate being stored, split it apart. 3206 DEBUG(dbgs() << " original: " << SI << "\n"); 3207 StoreOpSplitter Splitter(&SI, *U); 3208 Splitter.emitSplitOps(V->getType(), V, V->getName() + ".fca"); 3209 SI.eraseFromParent(); 3210 return true; 3211 } 3212 3213 bool visitBitCastInst(BitCastInst &BC) { 3214 enqueueUsers(BC); 3215 return false; 3216 } 3217 3218 bool visitGetElementPtrInst(GetElementPtrInst &GEPI) { 3219 enqueueUsers(GEPI); 3220 return false; 3221 } 3222 3223 bool visitPHINode(PHINode &PN) { 3224 enqueueUsers(PN); 3225 return false; 3226 } 3227 3228 bool visitSelectInst(SelectInst &SI) { 3229 enqueueUsers(SI); 3230 return false; 3231 } 3232}; 3233} 3234 3235/// \brief Strip aggregate type wrapping. 3236/// 3237/// This removes no-op aggregate types wrapping an underlying type. It will 3238/// strip as many layers of types as it can without changing either the type 3239/// size or the allocated size. 3240static Type *stripAggregateTypeWrapping(const DataLayout &DL, Type *Ty) { 3241 if (Ty->isSingleValueType()) 3242 return Ty; 3243 3244 uint64_t AllocSize = DL.getTypeAllocSize(Ty); 3245 uint64_t TypeSize = DL.getTypeSizeInBits(Ty); 3246 3247 Type *InnerTy; 3248 if (ArrayType *ArrTy = dyn_cast<ArrayType>(Ty)) { 3249 InnerTy = ArrTy->getElementType(); 3250 } else if (StructType *STy = dyn_cast<StructType>(Ty)) { 3251 const StructLayout *SL = DL.getStructLayout(STy); 3252 unsigned Index = SL->getElementContainingOffset(0); 3253 InnerTy = STy->getElementType(Index); 3254 } else { 3255 return Ty; 3256 } 3257 3258 if (AllocSize > DL.getTypeAllocSize(InnerTy) || 3259 TypeSize > DL.getTypeSizeInBits(InnerTy)) 3260 return Ty; 3261 3262 return stripAggregateTypeWrapping(DL, InnerTy); 3263} 3264 3265/// \brief Try to find a partition of the aggregate type passed in for a given 3266/// offset and size. 3267/// 3268/// This recurses through the aggregate type and tries to compute a subtype 3269/// based on the offset and size. When the offset and size span a sub-section 3270/// of an array, it will even compute a new array type for that sub-section, 3271/// and the same for structs. 3272/// 3273/// Note that this routine is very strict and tries to find a partition of the 3274/// type which produces the *exact* right offset and size. It is not forgiving 3275/// when the size or offset cause either end of type-based partition to be off. 3276/// Also, this is a best-effort routine. It is reasonable to give up and not 3277/// return a type if necessary. 3278static Type *getTypePartition(const DataLayout &TD, Type *Ty, 3279 uint64_t Offset, uint64_t Size) { 3280 if (Offset == 0 && TD.getTypeAllocSize(Ty) == Size) 3281 return stripAggregateTypeWrapping(TD, Ty); 3282 if (Offset > TD.getTypeAllocSize(Ty) || 3283 (TD.getTypeAllocSize(Ty) - Offset) < Size) 3284 return 0; 3285 3286 if (SequentialType *SeqTy = dyn_cast<SequentialType>(Ty)) { 3287 // We can't partition pointers... 3288 if (SeqTy->isPointerTy()) 3289 return 0; 3290 3291 Type *ElementTy = SeqTy->getElementType(); 3292 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy); 3293 uint64_t NumSkippedElements = Offset / ElementSize; 3294 if (ArrayType *ArrTy = dyn_cast<ArrayType>(SeqTy)) 3295 if (NumSkippedElements >= ArrTy->getNumElements()) 3296 return 0; 3297 if (VectorType *VecTy = dyn_cast<VectorType>(SeqTy)) 3298 if (NumSkippedElements >= VecTy->getNumElements()) 3299 return 0; 3300 Offset -= NumSkippedElements * ElementSize; 3301 3302 // First check if we need to recurse. 3303 if (Offset > 0 || Size < ElementSize) { 3304 // Bail if the partition ends in a different array element. 3305 if ((Offset + Size) > ElementSize) 3306 return 0; 3307 // Recurse through the element type trying to peel off offset bytes. 3308 return getTypePartition(TD, ElementTy, Offset, Size); 3309 } 3310 assert(Offset == 0); 3311 3312 if (Size == ElementSize) 3313 return stripAggregateTypeWrapping(TD, ElementTy); 3314 assert(Size > ElementSize); 3315 uint64_t NumElements = Size / ElementSize; 3316 if (NumElements * ElementSize != Size) 3317 return 0; 3318 return ArrayType::get(ElementTy, NumElements); 3319 } 3320 3321 StructType *STy = dyn_cast<StructType>(Ty); 3322 if (!STy) 3323 return 0; 3324 3325 const StructLayout *SL = TD.getStructLayout(STy); 3326 if (Offset >= SL->getSizeInBytes()) 3327 return 0; 3328 uint64_t EndOffset = Offset + Size; 3329 if (EndOffset > SL->getSizeInBytes()) 3330 return 0; 3331 3332 unsigned Index = SL->getElementContainingOffset(Offset); 3333 Offset -= SL->getElementOffset(Index); 3334 3335 Type *ElementTy = STy->getElementType(Index); 3336 uint64_t ElementSize = TD.getTypeAllocSize(ElementTy); 3337 if (Offset >= ElementSize) 3338 return 0; // The offset points into alignment padding. 3339 3340 // See if any partition must be contained by the element. 3341 if (Offset > 0 || Size < ElementSize) { 3342 if ((Offset + Size) > ElementSize) 3343 return 0; 3344 return getTypePartition(TD, ElementTy, Offset, Size); 3345 } 3346 assert(Offset == 0); 3347 3348 if (Size == ElementSize) 3349 return stripAggregateTypeWrapping(TD, ElementTy); 3350 3351 StructType::element_iterator EI = STy->element_begin() + Index, 3352 EE = STy->element_end(); 3353 if (EndOffset < SL->getSizeInBytes()) { 3354 unsigned EndIndex = SL->getElementContainingOffset(EndOffset); 3355 if (Index == EndIndex) 3356 return 0; // Within a single element and its padding. 3357 3358 // Don't try to form "natural" types if the elements don't line up with the 3359 // expected size. 3360 // FIXME: We could potentially recurse down through the last element in the 3361 // sub-struct to find a natural end point. 3362 if (SL->getElementOffset(EndIndex) != EndOffset) 3363 return 0; 3364 3365 assert(Index < EndIndex); 3366 EE = STy->element_begin() + EndIndex; 3367 } 3368 3369 // Try to build up a sub-structure. 3370 StructType *SubTy = StructType::get(STy->getContext(), makeArrayRef(EI, EE), 3371 STy->isPacked()); 3372 const StructLayout *SubSL = TD.getStructLayout(SubTy); 3373 if (Size != SubSL->getSizeInBytes()) 3374 return 0; // The sub-struct doesn't have quite the size needed. 3375 3376 return SubTy; 3377} 3378 3379/// \brief Rewrite an alloca partition's users. 3380/// 3381/// This routine drives both of the rewriting goals of the SROA pass. It tries 3382/// to rewrite uses of an alloca partition to be conducive for SSA value 3383/// promotion. If the partition needs a new, more refined alloca, this will 3384/// build that new alloca, preserving as much type information as possible, and 3385/// rewrite the uses of the old alloca to point at the new one and have the 3386/// appropriate new offsets. It also evaluates how successful the rewrite was 3387/// at enabling promotion and if it was successful queues the alloca to be 3388/// promoted. 3389bool SROA::rewriteAllocaPartition(AllocaInst &AI, 3390 AllocaPartitioning &P, 3391 AllocaPartitioning::iterator PI) { 3392 uint64_t AllocaSize = PI->EndOffset - PI->BeginOffset; 3393 bool IsLive = false; 3394 for (AllocaPartitioning::use_iterator UI = P.use_begin(PI), 3395 UE = P.use_end(PI); 3396 UI != UE && !IsLive; ++UI) 3397 if (UI->getUse()) 3398 IsLive = true; 3399 if (!IsLive) 3400 return false; // No live uses left of this partition. 3401 3402 DEBUG(dbgs() << "Speculating PHIs and selects in partition " 3403 << "[" << PI->BeginOffset << "," << PI->EndOffset << ")\n"); 3404 3405 PHIOrSelectSpeculator Speculator(*TD, P, *this); 3406 DEBUG(dbgs() << " speculating "); 3407 DEBUG(P.print(dbgs(), PI, "")); 3408 Speculator.visitUsers(PI); 3409 3410 // Try to compute a friendly type for this partition of the alloca. This 3411 // won't always succeed, in which case we fall back to a legal integer type 3412 // or an i8 array of an appropriate size. 3413 Type *AllocaTy = 0; 3414 if (Type *PartitionTy = P.getCommonType(PI)) 3415 if (TD->getTypeAllocSize(PartitionTy) >= AllocaSize) 3416 AllocaTy = PartitionTy; 3417 if (!AllocaTy) 3418 if (Type *PartitionTy = getTypePartition(*TD, AI.getAllocatedType(), 3419 PI->BeginOffset, AllocaSize)) 3420 AllocaTy = PartitionTy; 3421 if ((!AllocaTy || 3422 (AllocaTy->isArrayTy() && 3423 AllocaTy->getArrayElementType()->isIntegerTy())) && 3424 TD->isLegalInteger(AllocaSize * 8)) 3425 AllocaTy = Type::getIntNTy(*C, AllocaSize * 8); 3426 if (!AllocaTy) 3427 AllocaTy = ArrayType::get(Type::getInt8Ty(*C), AllocaSize); 3428 assert(TD->getTypeAllocSize(AllocaTy) >= AllocaSize); 3429 3430 // Check for the case where we're going to rewrite to a new alloca of the 3431 // exact same type as the original, and with the same access offsets. In that 3432 // case, re-use the existing alloca, but still run through the rewriter to 3433 // perform phi and select speculation. 3434 AllocaInst *NewAI; 3435 if (AllocaTy == AI.getAllocatedType()) { 3436 assert(PI->BeginOffset == 0 && 3437 "Non-zero begin offset but same alloca type"); 3438 assert(PI == P.begin() && "Begin offset is zero on later partition"); 3439 NewAI = &AI; 3440 } else { 3441 unsigned Alignment = AI.getAlignment(); 3442 if (!Alignment) { 3443 // The minimum alignment which users can rely on when the explicit 3444 // alignment is omitted or zero is that required by the ABI for this 3445 // type. 3446 Alignment = TD->getABITypeAlignment(AI.getAllocatedType()); 3447 } 3448 Alignment = MinAlign(Alignment, PI->BeginOffset); 3449 // If we will get at least this much alignment from the type alone, leave 3450 // the alloca's alignment unconstrained. 3451 if (Alignment <= TD->getABITypeAlignment(AllocaTy)) 3452 Alignment = 0; 3453 NewAI = new AllocaInst(AllocaTy, 0, Alignment, 3454 AI.getName() + ".sroa." + Twine(PI - P.begin()), 3455 &AI); 3456 ++NumNewAllocas; 3457 } 3458 3459 DEBUG(dbgs() << "Rewriting alloca partition " 3460 << "[" << PI->BeginOffset << "," << PI->EndOffset << ") to: " 3461 << *NewAI << "\n"); 3462 3463 // Track the high watermark of the post-promotion worklist. We will reset it 3464 // to this point if the alloca is not in fact scheduled for promotion. 3465 unsigned PPWOldSize = PostPromotionWorklist.size(); 3466 3467 AllocaPartitionRewriter Rewriter(*TD, P, PI, *this, AI, *NewAI, 3468 PI->BeginOffset, PI->EndOffset); 3469 DEBUG(dbgs() << " rewriting "); 3470 DEBUG(P.print(dbgs(), PI, "")); 3471 bool Promotable = Rewriter.visitUsers(P.use_begin(PI), P.use_end(PI)); 3472 if (Promotable) { 3473 DEBUG(dbgs() << " and queuing for promotion\n"); 3474 PromotableAllocas.push_back(NewAI); 3475 } else if (NewAI != &AI) { 3476 // If we can't promote the alloca, iterate on it to check for new 3477 // refinements exposed by splitting the current alloca. Don't iterate on an 3478 // alloca which didn't actually change and didn't get promoted. 3479 Worklist.insert(NewAI); 3480 } 3481 3482 // Drop any post-promotion work items if promotion didn't happen. 3483 if (!Promotable) 3484 while (PostPromotionWorklist.size() > PPWOldSize) 3485 PostPromotionWorklist.pop_back(); 3486 3487 return true; 3488} 3489 3490/// \brief Walks the partitioning of an alloca rewriting uses of each partition. 3491bool SROA::splitAlloca(AllocaInst &AI, AllocaPartitioning &P) { 3492 bool Changed = false; 3493 for (AllocaPartitioning::iterator PI = P.begin(), PE = P.end(); PI != PE; 3494 ++PI) 3495 Changed |= rewriteAllocaPartition(AI, P, PI); 3496 3497 return Changed; 3498} 3499 3500/// \brief Analyze an alloca for SROA. 3501/// 3502/// This analyzes the alloca to ensure we can reason about it, builds 3503/// a partitioning of the alloca, and then hands it off to be split and 3504/// rewritten as needed. 3505bool SROA::runOnAlloca(AllocaInst &AI) { 3506 DEBUG(dbgs() << "SROA alloca: " << AI << "\n"); 3507 ++NumAllocasAnalyzed; 3508 3509 // Special case dead allocas, as they're trivial. 3510 if (AI.use_empty()) { 3511 AI.eraseFromParent(); 3512 return true; 3513 } 3514 3515 // Skip alloca forms that this analysis can't handle. 3516 if (AI.isArrayAllocation() || !AI.getAllocatedType()->isSized() || 3517 TD->getTypeAllocSize(AI.getAllocatedType()) == 0) 3518 return false; 3519 3520 bool Changed = false; 3521 3522 // First, split any FCA loads and stores touching this alloca to promote 3523 // better splitting and promotion opportunities. 3524 AggLoadStoreRewriter AggRewriter(*TD); 3525 Changed |= AggRewriter.rewrite(AI); 3526 3527 // Build the partition set using a recursive instruction-visiting builder. 3528 AllocaPartitioning P(*TD, AI); 3529 DEBUG(P.print(dbgs())); 3530 if (P.isEscaped()) 3531 return Changed; 3532 3533 // Delete all the dead users of this alloca before splitting and rewriting it. 3534 for (AllocaPartitioning::dead_user_iterator DI = P.dead_user_begin(), 3535 DE = P.dead_user_end(); 3536 DI != DE; ++DI) { 3537 Changed = true; 3538 (*DI)->replaceAllUsesWith(UndefValue::get((*DI)->getType())); 3539 DeadInsts.insert(*DI); 3540 } 3541 for (AllocaPartitioning::dead_op_iterator DO = P.dead_op_begin(), 3542 DE = P.dead_op_end(); 3543 DO != DE; ++DO) { 3544 Value *OldV = **DO; 3545 // Clobber the use with an undef value. 3546 **DO = UndefValue::get(OldV->getType()); 3547 if (Instruction *OldI = dyn_cast<Instruction>(OldV)) 3548 if (isInstructionTriviallyDead(OldI)) { 3549 Changed = true; 3550 DeadInsts.insert(OldI); 3551 } 3552 } 3553 3554 // No partitions to split. Leave the dead alloca for a later pass to clean up. 3555 if (P.begin() == P.end()) 3556 return Changed; 3557 3558 return splitAlloca(AI, P) || Changed; 3559} 3560 3561/// \brief Delete the dead instructions accumulated in this run. 3562/// 3563/// Recursively deletes the dead instructions we've accumulated. This is done 3564/// at the very end to maximize locality of the recursive delete and to 3565/// minimize the problems of invalidated instruction pointers as such pointers 3566/// are used heavily in the intermediate stages of the algorithm. 3567/// 3568/// We also record the alloca instructions deleted here so that they aren't 3569/// subsequently handed to mem2reg to promote. 3570void SROA::deleteDeadInstructions(SmallPtrSet<AllocaInst*, 4> &DeletedAllocas) { 3571 while (!DeadInsts.empty()) { 3572 Instruction *I = DeadInsts.pop_back_val(); 3573 DEBUG(dbgs() << "Deleting dead instruction: " << *I << "\n"); 3574 3575 I->replaceAllUsesWith(UndefValue::get(I->getType())); 3576 3577 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 3578 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 3579 // Zero out the operand and see if it becomes trivially dead. 3580 *OI = 0; 3581 if (isInstructionTriviallyDead(U)) 3582 DeadInsts.insert(U); 3583 } 3584 3585 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) 3586 DeletedAllocas.insert(AI); 3587 3588 ++NumDeleted; 3589 I->eraseFromParent(); 3590 } 3591} 3592 3593/// \brief Promote the allocas, using the best available technique. 3594/// 3595/// This attempts to promote whatever allocas have been identified as viable in 3596/// the PromotableAllocas list. If that list is empty, there is nothing to do. 3597/// If there is a domtree available, we attempt to promote using the full power 3598/// of mem2reg. Otherwise, we build and use the AllocaPromoter above which is 3599/// based on the SSAUpdater utilities. This function returns whether any 3600/// promotion occurred. 3601bool SROA::promoteAllocas(Function &F) { 3602 if (PromotableAllocas.empty()) 3603 return false; 3604 3605 NumPromoted += PromotableAllocas.size(); 3606 3607 if (DT && !ForceSSAUpdater) { 3608 DEBUG(dbgs() << "Promoting allocas with mem2reg...\n"); 3609 PromoteMemToReg(PromotableAllocas, *DT); 3610 PromotableAllocas.clear(); 3611 return true; 3612 } 3613 3614 DEBUG(dbgs() << "Promoting allocas with SSAUpdater...\n"); 3615 SSAUpdater SSA; 3616 DIBuilder DIB(*F.getParent()); 3617 SmallVector<Instruction*, 64> Insts; 3618 3619 for (unsigned Idx = 0, Size = PromotableAllocas.size(); Idx != Size; ++Idx) { 3620 AllocaInst *AI = PromotableAllocas[Idx]; 3621 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end(); 3622 UI != UE;) { 3623 Instruction *I = cast<Instruction>(*UI++); 3624 // FIXME: Currently the SSAUpdater infrastructure doesn't reason about 3625 // lifetime intrinsics and so we strip them (and the bitcasts+GEPs 3626 // leading to them) here. Eventually it should use them to optimize the 3627 // scalar values produced. 3628 if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I)) { 3629 assert(onlyUsedByLifetimeMarkers(I) && 3630 "Found a bitcast used outside of a lifetime marker."); 3631 while (!I->use_empty()) 3632 cast<Instruction>(*I->use_begin())->eraseFromParent(); 3633 I->eraseFromParent(); 3634 continue; 3635 } 3636 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 3637 assert(II->getIntrinsicID() == Intrinsic::lifetime_start || 3638 II->getIntrinsicID() == Intrinsic::lifetime_end); 3639 II->eraseFromParent(); 3640 continue; 3641 } 3642 3643 Insts.push_back(I); 3644 } 3645 AllocaPromoter(Insts, SSA, *AI, DIB).run(Insts); 3646 Insts.clear(); 3647 } 3648 3649 PromotableAllocas.clear(); 3650 return true; 3651} 3652 3653namespace { 3654 /// \brief A predicate to test whether an alloca belongs to a set. 3655 class IsAllocaInSet { 3656 typedef SmallPtrSet<AllocaInst *, 4> SetType; 3657 const SetType &Set; 3658 3659 public: 3660 typedef AllocaInst *argument_type; 3661 3662 IsAllocaInSet(const SetType &Set) : Set(Set) {} 3663 bool operator()(AllocaInst *AI) const { return Set.count(AI); } 3664 }; 3665} 3666 3667bool SROA::runOnFunction(Function &F) { 3668 DEBUG(dbgs() << "SROA function: " << F.getName() << "\n"); 3669 C = &F.getContext(); 3670 TD = getAnalysisIfAvailable<DataLayout>(); 3671 if (!TD) { 3672 DEBUG(dbgs() << " Skipping SROA -- no target data!\n"); 3673 return false; 3674 } 3675 DT = getAnalysisIfAvailable<DominatorTree>(); 3676 3677 BasicBlock &EntryBB = F.getEntryBlock(); 3678 for (BasicBlock::iterator I = EntryBB.begin(), E = llvm::prior(EntryBB.end()); 3679 I != E; ++I) 3680 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) 3681 Worklist.insert(AI); 3682 3683 bool Changed = false; 3684 // A set of deleted alloca instruction pointers which should be removed from 3685 // the list of promotable allocas. 3686 SmallPtrSet<AllocaInst *, 4> DeletedAllocas; 3687 3688 do { 3689 while (!Worklist.empty()) { 3690 Changed |= runOnAlloca(*Worklist.pop_back_val()); 3691 deleteDeadInstructions(DeletedAllocas); 3692 3693 // Remove the deleted allocas from various lists so that we don't try to 3694 // continue processing them. 3695 if (!DeletedAllocas.empty()) { 3696 Worklist.remove_if(IsAllocaInSet(DeletedAllocas)); 3697 PostPromotionWorklist.remove_if(IsAllocaInSet(DeletedAllocas)); 3698 PromotableAllocas.erase(std::remove_if(PromotableAllocas.begin(), 3699 PromotableAllocas.end(), 3700 IsAllocaInSet(DeletedAllocas)), 3701 PromotableAllocas.end()); 3702 DeletedAllocas.clear(); 3703 } 3704 } 3705 3706 Changed |= promoteAllocas(F); 3707 3708 Worklist = PostPromotionWorklist; 3709 PostPromotionWorklist.clear(); 3710 } while (!Worklist.empty()); 3711 3712 return Changed; 3713} 3714 3715void SROA::getAnalysisUsage(AnalysisUsage &AU) const { 3716 if (RequiresDomTree) 3717 AU.addRequired<DominatorTree>(); 3718 AU.setPreservesCFG(); 3719} 3720