ScalarReplAggregates.cpp revision 5ffe6acd577696a41932c7b82db06a04687e07ba
1//===- ScalarReplAggregates.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// 10// This transformation implements the well known scalar replacement of 11// aggregates transformation. This xform breaks up alloca instructions of 12// aggregate type (structure or array) into individual alloca instructions for 13// each member (if possible). Then, if possible, it transforms the individual 14// alloca instructions into nice clean scalar SSA form. 15// 16// This combines a simple SRoA algorithm with the Mem2Reg algorithm because 17// often interact, especially for C++ programs. As such, iterating between 18// SRoA, then Mem2Reg until we run out of things to promote works well. 19// 20//===----------------------------------------------------------------------===// 21 22#define DEBUG_TYPE "scalarrepl" 23#include "llvm/Transforms/Scalar.h" 24#include "llvm/Constants.h" 25#include "llvm/DerivedTypes.h" 26#include "llvm/Function.h" 27#include "llvm/GlobalVariable.h" 28#include "llvm/Instructions.h" 29#include "llvm/IntrinsicInst.h" 30#include "llvm/Pass.h" 31#include "llvm/Analysis/Dominators.h" 32#include "llvm/Target/TargetData.h" 33#include "llvm/Transforms/Utils/PromoteMemToReg.h" 34#include "llvm/Support/Debug.h" 35#include "llvm/Support/GetElementPtrTypeIterator.h" 36#include "llvm/Support/MathExtras.h" 37#include "llvm/Support/Compiler.h" 38#include "llvm/ADT/SmallVector.h" 39#include "llvm/ADT/Statistic.h" 40#include "llvm/ADT/StringExtras.h" 41using namespace llvm; 42 43STATISTIC(NumReplaced, "Number of allocas broken up"); 44STATISTIC(NumPromoted, "Number of allocas promoted"); 45STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 46STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 47 48namespace { 49 struct VISIBILITY_HIDDEN SROA : public FunctionPass { 50 static char ID; // Pass identification, replacement for typeid 51 explicit SROA(signed T = -1) : FunctionPass(&ID) { 52 if (T == -1) 53 SRThreshold = 128; 54 else 55 SRThreshold = T; 56 } 57 58 bool runOnFunction(Function &F); 59 60 bool performScalarRepl(Function &F); 61 bool performPromotion(Function &F); 62 63 // getAnalysisUsage - This pass does not require any passes, but we know it 64 // will not alter the CFG, so say so. 65 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 66 AU.addRequired<DominatorTree>(); 67 AU.addRequired<DominanceFrontier>(); 68 AU.addRequired<TargetData>(); 69 AU.setPreservesCFG(); 70 } 71 72 private: 73 TargetData *TD; 74 75 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 76 /// information about the uses. All these fields are initialized to false 77 /// and set to true when something is learned. 78 struct AllocaInfo { 79 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 80 bool isUnsafe : 1; 81 82 /// needsCanon - This is set to true if there is some use of the alloca 83 /// that requires canonicalization. 84 bool needsCanon : 1; 85 86 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 87 bool isMemCpySrc : 1; 88 89 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 90 bool isMemCpyDst : 1; 91 92 AllocaInfo() 93 : isUnsafe(false), needsCanon(false), 94 isMemCpySrc(false), isMemCpyDst(false) {} 95 }; 96 97 unsigned SRThreshold; 98 99 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 100 101 int isSafeAllocaToScalarRepl(AllocationInst *AI); 102 103 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 104 AllocaInfo &Info); 105 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 106 AllocaInfo &Info); 107 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 108 unsigned OpNo, AllocaInfo &Info); 109 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI, 110 AllocaInfo &Info); 111 112 void DoScalarReplacement(AllocationInst *AI, 113 std::vector<AllocationInst*> &WorkList); 114 void CanonicalizeAllocaUsers(AllocationInst *AI); 115 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base); 116 117 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 118 SmallVector<AllocaInst*, 32> &NewElts); 119 120 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 121 AllocationInst *AI, 122 SmallVector<AllocaInst*, 32> &NewElts); 123 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI, 124 SmallVector<AllocaInst*, 32> &NewElts); 125 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 126 SmallVector<AllocaInst*, 32> &NewElts); 127 128 const Type *CanConvertToScalar(Value *V, bool &IsNotTrivial); 129 void ConvertToScalar(AllocationInst *AI, const Type *Ty); 130 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset); 131 Value *ConvertUsesOfLoadToScalar(LoadInst *LI, AllocaInst *NewAI, 132 unsigned Offset); 133 Value *ConvertUsesOfStoreToScalar(StoreInst *SI, AllocaInst *NewAI, 134 unsigned Offset); 135 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI); 136 }; 137} 138 139char SROA::ID = 0; 140static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 141 142// Public interface to the ScalarReplAggregates pass 143FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 144 return new SROA(Threshold); 145} 146 147 148bool SROA::runOnFunction(Function &F) { 149 TD = &getAnalysis<TargetData>(); 150 151 bool Changed = performPromotion(F); 152 while (1) { 153 bool LocalChange = performScalarRepl(F); 154 if (!LocalChange) break; // No need to repromote if no scalarrepl 155 Changed = true; 156 LocalChange = performPromotion(F); 157 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 158 } 159 160 return Changed; 161} 162 163 164bool SROA::performPromotion(Function &F) { 165 std::vector<AllocaInst*> Allocas; 166 DominatorTree &DT = getAnalysis<DominatorTree>(); 167 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 168 169 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 170 171 bool Changed = false; 172 173 while (1) { 174 Allocas.clear(); 175 176 // Find allocas that are safe to promote, by looking at all instructions in 177 // the entry node 178 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 179 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 180 if (isAllocaPromotable(AI)) 181 Allocas.push_back(AI); 182 183 if (Allocas.empty()) break; 184 185 PromoteMemToReg(Allocas, DT, DF); 186 NumPromoted += Allocas.size(); 187 Changed = true; 188 } 189 190 return Changed; 191} 192 193/// getNumSAElements - Return the number of elements in the specific struct or 194/// array. 195static uint64_t getNumSAElements(const Type *T) { 196 if (const StructType *ST = dyn_cast<StructType>(T)) 197 return ST->getNumElements(); 198 return cast<ArrayType>(T)->getNumElements(); 199} 200 201// performScalarRepl - This algorithm is a simple worklist driven algorithm, 202// which runs on all of the malloc/alloca instructions in the function, removing 203// them if they are only used by getelementptr instructions. 204// 205bool SROA::performScalarRepl(Function &F) { 206 std::vector<AllocationInst*> WorkList; 207 208 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 209 BasicBlock &BB = F.getEntryBlock(); 210 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 211 if (AllocationInst *A = dyn_cast<AllocationInst>(I)) 212 WorkList.push_back(A); 213 214 // Process the worklist 215 bool Changed = false; 216 while (!WorkList.empty()) { 217 AllocationInst *AI = WorkList.back(); 218 WorkList.pop_back(); 219 220 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 221 // with unused elements. 222 if (AI->use_empty()) { 223 AI->eraseFromParent(); 224 continue; 225 } 226 227 // If we can turn this aggregate value (potentially with casts) into a 228 // simple scalar value that can be mem2reg'd into a register value. 229 bool IsNotTrivial = false; 230 if (const Type *ActualType = CanConvertToScalar(AI, IsNotTrivial)) 231 if (IsNotTrivial && ActualType != Type::VoidTy) { 232 ConvertToScalar(AI, ActualType); 233 Changed = true; 234 continue; 235 } 236 237 // Check to see if we can perform the core SROA transformation. We cannot 238 // transform the allocation instruction if it is an array allocation 239 // (allocations OF arrays are ok though), and an allocation of a scalar 240 // value cannot be decomposed at all. 241 if (!AI->isArrayAllocation() && 242 (isa<StructType>(AI->getAllocatedType()) || 243 isa<ArrayType>(AI->getAllocatedType())) && 244 AI->getAllocatedType()->isSized() && 245 // Do not promote any struct whose size is larger than "128" bytes. 246 TD->getABITypeSize(AI->getAllocatedType()) < SRThreshold && 247 // Do not promote any struct into more than "32" separate vars. 248 getNumSAElements(AI->getAllocatedType()) < SRThreshold/4) { 249 // Check that all of the users of the allocation are capable of being 250 // transformed. 251 switch (isSafeAllocaToScalarRepl(AI)) { 252 default: assert(0 && "Unexpected value!"); 253 case 0: // Not safe to scalar replace. 254 break; 255 case 1: // Safe, but requires cleanup/canonicalizations first 256 CanonicalizeAllocaUsers(AI); 257 // FALL THROUGH. 258 case 3: // Safe to scalar replace. 259 DoScalarReplacement(AI, WorkList); 260 Changed = true; 261 continue; 262 } 263 } 264 265 // Check to see if this allocation is only modified by a memcpy/memmove from 266 // a constant global. If this is the case, we can change all users to use 267 // the constant global instead. This is commonly produced by the CFE by 268 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 269 // is only subsequently read. 270 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 271 DOUT << "Found alloca equal to global: " << *AI; 272 DOUT << " memcpy = " << *TheCopy; 273 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 274 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 275 TheCopy->eraseFromParent(); // Don't mutate the global. 276 AI->eraseFromParent(); 277 ++NumGlobals; 278 Changed = true; 279 continue; 280 } 281 282 // Otherwise, couldn't process this. 283 } 284 285 return Changed; 286} 287 288/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 289/// predicate, do SROA now. 290void SROA::DoScalarReplacement(AllocationInst *AI, 291 std::vector<AllocationInst*> &WorkList) { 292 DOUT << "Found inst to SROA: " << *AI; 293 SmallVector<AllocaInst*, 32> ElementAllocas; 294 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 295 ElementAllocas.reserve(ST->getNumContainedTypes()); 296 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 297 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 298 AI->getAlignment(), 299 AI->getName() + "." + utostr(i), AI); 300 ElementAllocas.push_back(NA); 301 WorkList.push_back(NA); // Add to worklist for recursive processing 302 } 303 } else { 304 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 305 ElementAllocas.reserve(AT->getNumElements()); 306 const Type *ElTy = AT->getElementType(); 307 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 308 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 309 AI->getName() + "." + utostr(i), AI); 310 ElementAllocas.push_back(NA); 311 WorkList.push_back(NA); // Add to worklist for recursive processing 312 } 313 } 314 315 // Now that we have created the alloca instructions that we want to use, 316 // expand the getelementptr instructions to use them. 317 // 318 while (!AI->use_empty()) { 319 Instruction *User = cast<Instruction>(AI->use_back()); 320 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) { 321 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas); 322 BCInst->eraseFromParent(); 323 continue; 324 } 325 326 // Replace: 327 // %res = load { i32, i32 }* %alloc 328 // with: 329 // %load.0 = load i32* %alloc.0 330 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 331 // %load.1 = load i32* %alloc.1 332 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 333 // (Also works for arrays instead of structs) 334 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 335 Value *Insert = UndefValue::get(LI->getType()); 336 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 337 Value *Load = new LoadInst(ElementAllocas[i], "load", LI); 338 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 339 } 340 LI->replaceAllUsesWith(Insert); 341 LI->eraseFromParent(); 342 continue; 343 } 344 345 // Replace: 346 // store { i32, i32 } %val, { i32, i32 }* %alloc 347 // with: 348 // %val.0 = extractvalue { i32, i32 } %val, 0 349 // store i32 %val.0, i32* %alloc.0 350 // %val.1 = extractvalue { i32, i32 } %val, 1 351 // store i32 %val.1, i32* %alloc.1 352 // (Also works for arrays instead of structs) 353 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 354 Value *Val = SI->getOperand(0); 355 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 356 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 357 new StoreInst(Extract, ElementAllocas[i], SI); 358 } 359 SI->eraseFromParent(); 360 continue; 361 } 362 363 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 364 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 365 unsigned Idx = 366 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 367 368 assert(Idx < ElementAllocas.size() && "Index out of range?"); 369 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 370 371 Value *RepValue; 372 if (GEPI->getNumOperands() == 3) { 373 // Do not insert a new getelementptr instruction with zero indices, only 374 // to have it optimized out later. 375 RepValue = AllocaToUse; 376 } else { 377 // We are indexing deeply into the structure, so we still need a 378 // getelement ptr instruction to finish the indexing. This may be 379 // expanded itself once the worklist is rerun. 380 // 381 SmallVector<Value*, 8> NewArgs; 382 NewArgs.push_back(Constant::getNullValue(Type::Int32Ty)); 383 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end()); 384 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(), 385 NewArgs.end(), "", GEPI); 386 RepValue->takeName(GEPI); 387 } 388 389 // If this GEP is to the start of the aggregate, check for memcpys. 390 if (Idx == 0 && GEPI->hasAllZeroIndices()) 391 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas); 392 393 // Move all of the users over to the new GEP. 394 GEPI->replaceAllUsesWith(RepValue); 395 // Delete the old GEP 396 GEPI->eraseFromParent(); 397 } 398 399 // Finally, delete the Alloca instruction 400 AI->eraseFromParent(); 401 NumReplaced++; 402} 403 404 405/// isSafeElementUse - Check to see if this use is an allowed use for a 406/// getelementptr instruction of an array aggregate allocation. isFirstElt 407/// indicates whether Ptr is known to the start of the aggregate. 408/// 409void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 410 AllocaInfo &Info) { 411 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 412 I != E; ++I) { 413 Instruction *User = cast<Instruction>(*I); 414 switch (User->getOpcode()) { 415 case Instruction::Load: break; 416 case Instruction::Store: 417 // Store is ok if storing INTO the pointer, not storing the pointer 418 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info); 419 break; 420 case Instruction::GetElementPtr: { 421 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 422 bool AreAllZeroIndices = isFirstElt; 423 if (GEP->getNumOperands() > 1) { 424 if (!isa<ConstantInt>(GEP->getOperand(1)) || 425 !cast<ConstantInt>(GEP->getOperand(1))->isZero()) 426 // Using pointer arithmetic to navigate the array. 427 return MarkUnsafe(Info); 428 429 if (AreAllZeroIndices) 430 AreAllZeroIndices = GEP->hasAllZeroIndices(); 431 } 432 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 433 if (Info.isUnsafe) return; 434 break; 435 } 436 case Instruction::BitCast: 437 if (isFirstElt) { 438 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 439 if (Info.isUnsafe) return; 440 break; 441 } 442 DOUT << " Transformation preventing inst: " << *User; 443 return MarkUnsafe(Info); 444 case Instruction::Call: 445 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 446 if (isFirstElt) { 447 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 448 if (Info.isUnsafe) return; 449 break; 450 } 451 } 452 DOUT << " Transformation preventing inst: " << *User; 453 return MarkUnsafe(Info); 454 default: 455 DOUT << " Transformation preventing inst: " << *User; 456 return MarkUnsafe(Info); 457 } 458 } 459 return; // All users look ok :) 460} 461 462/// AllUsersAreLoads - Return true if all users of this value are loads. 463static bool AllUsersAreLoads(Value *Ptr) { 464 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 465 I != E; ++I) 466 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 467 return false; 468 return true; 469} 470 471/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 472/// aggregate allocation. 473/// 474void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 475 AllocaInfo &Info) { 476 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 477 return isSafeUseOfBitCastedAllocation(C, AI, Info); 478 479 if (isa<LoadInst>(User)) 480 return; // Loads (returning a first class aggregrate) are always rewritable 481 482 if (isa<StoreInst>(User) && User->getOperand(0) != AI) 483 return; // Store is ok if storing INTO the pointer, not storing the pointer 484 485 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 486 if (GEPI == 0) 487 return MarkUnsafe(Info); 488 489 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 490 491 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 492 if (I == E || 493 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) { 494 return MarkUnsafe(Info); 495 } 496 497 ++I; 498 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 499 500 bool IsAllZeroIndices = true; 501 502 // If the first index is a non-constant index into an array, see if we can 503 // handle it as a special case. 504 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 505 if (!isa<ConstantInt>(I.getOperand())) { 506 IsAllZeroIndices = 0; 507 uint64_t NumElements = AT->getNumElements(); 508 509 // If this is an array index and the index is not constant, we cannot 510 // promote... that is unless the array has exactly one or two elements in 511 // it, in which case we CAN promote it, but we have to canonicalize this 512 // out if this is the only problem. 513 if ((NumElements == 1 || NumElements == 2) && 514 AllUsersAreLoads(GEPI)) { 515 Info.needsCanon = true; 516 return; // Canonicalization required! 517 } 518 return MarkUnsafe(Info); 519 } 520 } 521 522 // Walk through the GEP type indices, checking the types that this indexes 523 // into. 524 for (; I != E; ++I) { 525 // Ignore struct elements, no extra checking needed for these. 526 if (isa<StructType>(*I)) 527 continue; 528 529 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 530 if (!IdxVal) return MarkUnsafe(Info); 531 532 // Are all indices still zero? 533 IsAllZeroIndices &= IdxVal->isZero(); 534 535 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 536 // This GEP indexes an array. Verify that this is an in-range constant 537 // integer. Specifically, consider A[0][i]. We cannot know that the user 538 // isn't doing invalid things like allowing i to index an out-of-range 539 // subscript that accesses A[1]. Because of this, we have to reject SROA 540 // of any accesses into structs where any of the components are variables. 541 if (IdxVal->getZExtValue() >= AT->getNumElements()) 542 return MarkUnsafe(Info); 543 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) { 544 if (IdxVal->getZExtValue() >= VT->getNumElements()) 545 return MarkUnsafe(Info); 546 } 547 } 548 549 // If there are any non-simple uses of this getelementptr, make sure to reject 550 // them. 551 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 552} 553 554/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory 555/// intrinsic can be promoted by SROA. At this point, we know that the operand 556/// of the memintrinsic is a pointer to the beginning of the allocation. 557void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 558 unsigned OpNo, AllocaInfo &Info) { 559 // If not constant length, give up. 560 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 561 if (!Length) return MarkUnsafe(Info); 562 563 // If not the whole aggregate, give up. 564 if (Length->getZExtValue() != 565 TD->getABITypeSize(AI->getType()->getElementType())) 566 return MarkUnsafe(Info); 567 568 // We only know about memcpy/memset/memmove. 569 if (!isa<MemCpyInst>(MI) && !isa<MemSetInst>(MI) && !isa<MemMoveInst>(MI)) 570 return MarkUnsafe(Info); 571 572 // Otherwise, we can transform it. Determine whether this is a memcpy/set 573 // into or out of the aggregate. 574 if (OpNo == 1) 575 Info.isMemCpyDst = true; 576 else { 577 assert(OpNo == 2); 578 Info.isMemCpySrc = true; 579 } 580} 581 582/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast 583/// are 584void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI, 585 AllocaInfo &Info) { 586 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 587 UI != E; ++UI) { 588 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 589 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 590 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 591 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 592 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 593 // If storing the entire alloca in one chunk through a bitcasted pointer 594 // to integer, we can transform it. This happens (for example) when you 595 // cast a {i32,i32}* to i64* and store through it. This is similar to the 596 // memcpy case and occurs in various "byval" cases and emulated memcpys. 597 if (isa<IntegerType>(SI->getOperand(0)->getType()) && 598 TD->getABITypeSize(SI->getOperand(0)->getType()) == 599 TD->getABITypeSize(AI->getType()->getElementType())) { 600 Info.isMemCpyDst = true; 601 continue; 602 } 603 return MarkUnsafe(Info); 604 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 605 // If loading the entire alloca in one chunk through a bitcasted pointer 606 // to integer, we can transform it. This happens (for example) when you 607 // cast a {i32,i32}* to i64* and load through it. This is similar to the 608 // memcpy case and occurs in various "byval" cases and emulated memcpys. 609 if (isa<IntegerType>(LI->getType()) && 610 TD->getABITypeSize(LI->getType()) == 611 TD->getABITypeSize(AI->getType()->getElementType())) { 612 Info.isMemCpySrc = true; 613 continue; 614 } 615 return MarkUnsafe(Info); 616 } else { 617 return MarkUnsafe(Info); 618 } 619 if (Info.isUnsafe) return; 620 } 621} 622 623/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 624/// to its first element. Transform users of the cast to use the new values 625/// instead. 626void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 627 SmallVector<AllocaInst*, 32> &NewElts) { 628 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 629 while (UI != UE) { 630 Instruction *User = cast<Instruction>(*UI++); 631 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) { 632 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 633 if (BCU->use_empty()) BCU->eraseFromParent(); 634 continue; 635 } 636 637 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 638 // This must be memcpy/memmove/memset of the entire aggregate. 639 // Split into one per element. 640 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts); 641 continue; 642 } 643 644 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 645 // If this is a store of the entire alloca from an integer, rewrite it. 646 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 647 continue; 648 } 649 650 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 651 // If this is a load of the entire alloca to an integer, rewrite it. 652 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 653 continue; 654 } 655 656 // Otherwise it must be some other user of a gep of the first pointer. Just 657 // leave these alone. 658 continue; 659 } 660} 661 662/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 663/// Rewrite it to copy or set the elements of the scalarized memory. 664void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 665 AllocationInst *AI, 666 SmallVector<AllocaInst*, 32> &NewElts) { 667 668 // If this is a memcpy/memmove, construct the other pointer as the 669 // appropriate type. 670 Value *OtherPtr = 0; 671 if (MemCpyInst *MCI = dyn_cast<MemCpyInst>(MI)) { 672 if (BCInst == MCI->getRawDest()) 673 OtherPtr = MCI->getRawSource(); 674 else { 675 assert(BCInst == MCI->getRawSource()); 676 OtherPtr = MCI->getRawDest(); 677 } 678 } else if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) { 679 if (BCInst == MMI->getRawDest()) 680 OtherPtr = MMI->getRawSource(); 681 else { 682 assert(BCInst == MMI->getRawSource()); 683 OtherPtr = MMI->getRawDest(); 684 } 685 } 686 687 // If there is an other pointer, we want to convert it to the same pointer 688 // type as AI has, so we can GEP through it safely. 689 if (OtherPtr) { 690 // It is likely that OtherPtr is a bitcast, if so, remove it. 691 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 692 OtherPtr = BC->getOperand(0); 693 // All zero GEPs are effectively bitcasts. 694 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) 695 if (GEP->hasAllZeroIndices()) 696 OtherPtr = GEP->getOperand(0); 697 698 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 699 if (BCE->getOpcode() == Instruction::BitCast) 700 OtherPtr = BCE->getOperand(0); 701 702 // If the pointer is not the right type, insert a bitcast to the right 703 // type. 704 if (OtherPtr->getType() != AI->getType()) 705 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 706 MI); 707 } 708 709 // Process each element of the aggregate. 710 Value *TheFn = MI->getOperand(0); 711 const Type *BytePtrTy = MI->getRawDest()->getType(); 712 bool SROADest = MI->getRawDest() == BCInst; 713 714 Constant *Zero = Constant::getNullValue(Type::Int32Ty); 715 716 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 717 // If this is a memcpy/memmove, emit a GEP of the other element address. 718 Value *OtherElt = 0; 719 if (OtherPtr) { 720 Value *Idx[2] = { Zero, ConstantInt::get(Type::Int32Ty, i) }; 721 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2, 722 OtherPtr->getNameStr()+"."+utostr(i), 723 MI); 724 } 725 726 Value *EltPtr = NewElts[i]; 727 const Type *EltTy =cast<PointerType>(EltPtr->getType())->getElementType(); 728 729 // If we got down to a scalar, insert a load or store as appropriate. 730 if (EltTy->isSingleValueType()) { 731 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) { 732 Value *Elt = new LoadInst(SROADest ? OtherElt : EltPtr, "tmp", 733 MI); 734 new StoreInst(Elt, SROADest ? EltPtr : OtherElt, MI); 735 continue; 736 } 737 assert(isa<MemSetInst>(MI)); 738 739 // If the stored element is zero (common case), just store a null 740 // constant. 741 Constant *StoreVal; 742 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 743 if (CI->isZero()) { 744 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 745 } else { 746 // If EltTy is a vector type, get the element type. 747 const Type *ValTy = EltTy; 748 if (const VectorType *VTy = dyn_cast<VectorType>(ValTy)) 749 ValTy = VTy->getElementType(); 750 751 // Construct an integer with the right value. 752 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 753 APInt OneVal(EltSize, CI->getZExtValue()); 754 APInt TotalVal(OneVal); 755 // Set each byte. 756 for (unsigned i = 0; 8*i < EltSize; ++i) { 757 TotalVal = TotalVal.shl(8); 758 TotalVal |= OneVal; 759 } 760 761 // Convert the integer value to the appropriate type. 762 StoreVal = ConstantInt::get(TotalVal); 763 if (isa<PointerType>(ValTy)) 764 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 765 else if (ValTy->isFloatingPoint()) 766 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 767 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 768 769 // If the requested value was a vector constant, create it. 770 if (EltTy != ValTy) { 771 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 772 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 773 StoreVal = ConstantVector::get(&Elts[0], NumElts); 774 } 775 } 776 new StoreInst(StoreVal, EltPtr, MI); 777 continue; 778 } 779 // Otherwise, if we're storing a byte variable, use a memset call for 780 // this element. 781 } 782 783 // Cast the element pointer to BytePtrTy. 784 if (EltPtr->getType() != BytePtrTy) 785 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 786 787 // Cast the other pointer (if we have one) to BytePtrTy. 788 if (OtherElt && OtherElt->getType() != BytePtrTy) 789 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 790 MI); 791 792 unsigned EltSize = TD->getABITypeSize(EltTy); 793 794 // Finally, insert the meminst for this element. 795 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) { 796 Value *Ops[] = { 797 SROADest ? EltPtr : OtherElt, // Dest ptr 798 SROADest ? OtherElt : EltPtr, // Src ptr 799 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 800 Zero // Align 801 }; 802 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 803 } else { 804 assert(isa<MemSetInst>(MI)); 805 Value *Ops[] = { 806 EltPtr, MI->getOperand(2), // Dest, Value, 807 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 808 Zero // Align 809 }; 810 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 811 } 812 } 813 MI->eraseFromParent(); 814} 815 816/// RewriteStoreUserOfWholeAlloca - We found an store of an integer that 817/// overwrites the entire allocation. Extract out the pieces of the stored 818/// integer and store them individually. 819void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, 820 AllocationInst *AI, 821 SmallVector<AllocaInst*, 32> &NewElts){ 822 // Extract each element out of the integer according to its structure offset 823 // and store the element value to the individual alloca. 824 Value *SrcVal = SI->getOperand(0); 825 const Type *AllocaEltTy = AI->getType()->getElementType(); 826 uint64_t AllocaSizeBits = TD->getABITypeSizeInBits(AllocaEltTy); 827 828 // If this isn't a store of an integer to the whole alloca, it may be a store 829 // to the first element. Just ignore the store in this case and normal SROA 830 // will handle it. 831 if (!isa<IntegerType>(SrcVal->getType()) || 832 TD->getABITypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 833 return; 834 835 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI; 836 837 // There are two forms here: AI could be an array or struct. Both cases 838 // have different ways to compute the element offset. 839 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 840 const StructLayout *Layout = TD->getStructLayout(EltSTy); 841 842 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 843 // Get the number of bits to shift SrcVal to get the value. 844 const Type *FieldTy = EltSTy->getElementType(i); 845 uint64_t Shift = Layout->getElementOffsetInBits(i); 846 847 if (TD->isBigEndian()) 848 Shift = AllocaSizeBits-Shift-TD->getABITypeSizeInBits(FieldTy); 849 850 Value *EltVal = SrcVal; 851 if (Shift) { 852 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 853 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 854 "sroa.store.elt", SI); 855 } 856 857 // Truncate down to an integer of the right size. 858 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 859 if (FieldSizeBits != AllocaSizeBits) 860 EltVal = new TruncInst(EltVal, IntegerType::get(FieldSizeBits), "", SI); 861 Value *DestField = NewElts[i]; 862 if (EltVal->getType() == FieldTy) { 863 // Storing to an integer field of this size, just do it. 864 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { 865 // Bitcast to the right element type (for fp/vector values). 866 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 867 } else { 868 // Otherwise, bitcast the dest pointer (for aggregates). 869 DestField = new BitCastInst(DestField, 870 PointerType::getUnqual(EltVal->getType()), 871 "", SI); 872 } 873 new StoreInst(EltVal, DestField, SI); 874 } 875 876 } else { 877 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 878 const Type *ArrayEltTy = ATy->getElementType(); 879 uint64_t ElementOffset = TD->getABITypeSizeInBits(ArrayEltTy); 880 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 881 882 uint64_t Shift; 883 884 if (TD->isBigEndian()) 885 Shift = AllocaSizeBits-ElementOffset; 886 else 887 Shift = 0; 888 889 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 890 891 Value *EltVal = SrcVal; 892 if (Shift) { 893 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 894 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 895 "sroa.store.elt", SI); 896 } 897 898 // Truncate down to an integer of the right size. 899 if (ElementSizeBits != AllocaSizeBits) 900 EltVal = new TruncInst(EltVal, IntegerType::get(ElementSizeBits),"",SI); 901 Value *DestField = NewElts[i]; 902 if (EltVal->getType() == ArrayEltTy) { 903 // Storing to an integer field of this size, just do it. 904 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { 905 // Bitcast to the right element type (for fp/vector values). 906 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 907 } else { 908 // Otherwise, bitcast the dest pointer (for aggregates). 909 DestField = new BitCastInst(DestField, 910 PointerType::getUnqual(EltVal->getType()), 911 "", SI); 912 } 913 new StoreInst(EltVal, DestField, SI); 914 915 if (TD->isBigEndian()) 916 Shift -= ElementOffset; 917 else 918 Shift += ElementOffset; 919 } 920 } 921 922 SI->eraseFromParent(); 923} 924 925/// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to 926/// an integer. Load the individual pieces to form the aggregate value. 927void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 928 SmallVector<AllocaInst*, 32> &NewElts) { 929 // Extract each element out of the NewElts according to its structure offset 930 // and form the result value. 931 const Type *AllocaEltTy = AI->getType()->getElementType(); 932 uint64_t AllocaSizeBits = TD->getABITypeSizeInBits(AllocaEltTy); 933 934 // If this isn't a load of the whole alloca to an integer, it may be a load 935 // of the first element. Just ignore the load in this case and normal SROA 936 // will handle it. 937 if (!isa<IntegerType>(LI->getType()) || 938 TD->getABITypeSizeInBits(LI->getType()) != AllocaSizeBits) 939 return; 940 941 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI; 942 943 // There are two forms here: AI could be an array or struct. Both cases 944 // have different ways to compute the element offset. 945 const StructLayout *Layout = 0; 946 uint64_t ArrayEltBitOffset = 0; 947 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 948 Layout = TD->getStructLayout(EltSTy); 949 } else { 950 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 951 ArrayEltBitOffset = TD->getABITypeSizeInBits(ArrayEltTy); 952 } 953 954 Value *ResultVal = Constant::getNullValue(LI->getType()); 955 956 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 957 // Load the value from the alloca. If the NewElt is an aggregate, cast 958 // the pointer to an integer of the same size before doing the load. 959 Value *SrcField = NewElts[i]; 960 const Type *FieldTy = 961 cast<PointerType>(SrcField->getType())->getElementType(); 962 const IntegerType *FieldIntTy = 963 IntegerType::get(TD->getTypeSizeInBits(FieldTy)); 964 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && 965 !isa<VectorType>(FieldTy)) 966 SrcField = new BitCastInst(SrcField, PointerType::getUnqual(FieldIntTy), 967 "", LI); 968 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 969 970 // If SrcField is a fp or vector of the right size but that isn't an 971 // integer type, bitcast to an integer so we can shift it. 972 if (SrcField->getType() != FieldIntTy) 973 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 974 975 // Zero extend the field to be the same size as the final alloca so that 976 // we can shift and insert it. 977 if (SrcField->getType() != ResultVal->getType()) 978 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 979 980 // Determine the number of bits to shift SrcField. 981 uint64_t Shift; 982 if (Layout) // Struct case. 983 Shift = Layout->getElementOffsetInBits(i); 984 else // Array case. 985 Shift = i*ArrayEltBitOffset; 986 987 if (TD->isBigEndian()) 988 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 989 990 if (Shift) { 991 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 992 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 993 } 994 995 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 996 } 997 998 LI->replaceAllUsesWith(ResultVal); 999 LI->eraseFromParent(); 1000} 1001 1002 1003/// HasPadding - Return true if the specified type has any structure or 1004/// alignment padding, false otherwise. 1005static bool HasPadding(const Type *Ty, const TargetData &TD) { 1006 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1007 const StructLayout *SL = TD.getStructLayout(STy); 1008 unsigned PrevFieldBitOffset = 0; 1009 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1010 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1011 1012 // Padding in sub-elements? 1013 if (HasPadding(STy->getElementType(i), TD)) 1014 return true; 1015 1016 // Check to see if there is any padding between this element and the 1017 // previous one. 1018 if (i) { 1019 unsigned PrevFieldEnd = 1020 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1021 if (PrevFieldEnd < FieldBitOffset) 1022 return true; 1023 } 1024 1025 PrevFieldBitOffset = FieldBitOffset; 1026 } 1027 1028 // Check for tail padding. 1029 if (unsigned EltCount = STy->getNumElements()) { 1030 unsigned PrevFieldEnd = PrevFieldBitOffset + 1031 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1032 if (PrevFieldEnd < SL->getSizeInBits()) 1033 return true; 1034 } 1035 1036 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1037 return HasPadding(ATy->getElementType(), TD); 1038 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1039 return HasPadding(VTy->getElementType(), TD); 1040 } 1041 return TD.getTypeSizeInBits(Ty) != TD.getABITypeSizeInBits(Ty); 1042} 1043 1044/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1045/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1046/// or 1 if safe after canonicalization has been performed. 1047/// 1048int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 1049 // Loop over the use list of the alloca. We can only transform it if all of 1050 // the users are safe to transform. 1051 AllocaInfo Info; 1052 1053 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 1054 I != E; ++I) { 1055 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 1056 if (Info.isUnsafe) { 1057 DOUT << "Cannot transform: " << *AI << " due to user: " << **I; 1058 return 0; 1059 } 1060 } 1061 1062 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1063 // source and destination, we have to be careful. In particular, the memcpy 1064 // could be moving around elements that live in structure padding of the LLVM 1065 // types, but may actually be used. In these cases, we refuse to promote the 1066 // struct. 1067 if (Info.isMemCpySrc && Info.isMemCpyDst && 1068 HasPadding(AI->getType()->getElementType(), *TD)) 1069 return 0; 1070 1071 // If we require cleanup, return 1, otherwise return 3. 1072 return Info.needsCanon ? 1 : 3; 1073} 1074 1075/// CanonicalizeAllocaUsers - If SROA reported that it can promote the specified 1076/// allocation, but only if cleaned up, perform the cleanups required. 1077void SROA::CanonicalizeAllocaUsers(AllocationInst *AI) { 1078 // At this point, we know that the end result will be SROA'd and promoted, so 1079 // we can insert ugly code if required so long as sroa+mem2reg will clean it 1080 // up. 1081 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1082 UI != E; ) { 1083 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI++); 1084 if (!GEPI) continue; 1085 gep_type_iterator I = gep_type_begin(GEPI); 1086 ++I; 1087 1088 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 1089 uint64_t NumElements = AT->getNumElements(); 1090 1091 if (!isa<ConstantInt>(I.getOperand())) { 1092 if (NumElements == 1) { 1093 GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty)); 1094 } else { 1095 assert(NumElements == 2 && "Unhandled case!"); 1096 // All users of the GEP must be loads. At each use of the GEP, insert 1097 // two loads of the appropriate indexed GEP and select between them. 1098 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(), 1099 Constant::getNullValue(I.getOperand()->getType()), 1100 "isone", GEPI); 1101 // Insert the new GEP instructions, which are properly indexed. 1102 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 1103 Indices[1] = Constant::getNullValue(Type::Int32Ty); 1104 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1105 Indices.begin(), 1106 Indices.end(), 1107 GEPI->getName()+".0", GEPI); 1108 Indices[1] = ConstantInt::get(Type::Int32Ty, 1); 1109 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1110 Indices.begin(), 1111 Indices.end(), 1112 GEPI->getName()+".1", GEPI); 1113 // Replace all loads of the variable index GEP with loads from both 1114 // indexes and a select. 1115 while (!GEPI->use_empty()) { 1116 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 1117 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 1118 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 1119 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 1120 LI->replaceAllUsesWith(R); 1121 LI->eraseFromParent(); 1122 } 1123 GEPI->eraseFromParent(); 1124 } 1125 } 1126 } 1127 } 1128} 1129 1130/// MergeInType - Add the 'In' type to the accumulated type so far. If the 1131/// types are incompatible, return true, otherwise update Accum and return 1132/// false. 1133/// 1134/// There are three cases we handle here: 1135/// 1) An effectively-integer union, where the pieces are stored into as 1136/// smaller integers (common with byte swap and other idioms). 1137/// 2) A union of vector types of the same size and potentially its elements. 1138/// Here we turn element accesses into insert/extract element operations. 1139/// 3) A union of scalar types, such as int/float or int/pointer. Here we 1140/// merge together into integers, allowing the xform to work with #1 as 1141/// well. 1142static bool MergeInType(const Type *In, const Type *&Accum, 1143 const TargetData &TD) { 1144 // If this is our first type, just use it. 1145 const VectorType *PTy; 1146 if (Accum == Type::VoidTy || In == Accum) { 1147 Accum = In; 1148 } else if (In == Type::VoidTy) { 1149 // Noop. 1150 } else if (In->isInteger() && Accum->isInteger()) { // integer union. 1151 // Otherwise pick whichever type is larger. 1152 if (cast<IntegerType>(In)->getBitWidth() > 1153 cast<IntegerType>(Accum)->getBitWidth()) 1154 Accum = In; 1155 } else if (isa<PointerType>(In) && isa<PointerType>(Accum)) { 1156 // Pointer unions just stay as one of the pointers. 1157 } else if (isa<VectorType>(In) || isa<VectorType>(Accum)) { 1158 if ((PTy = dyn_cast<VectorType>(Accum)) && 1159 PTy->getElementType() == In) { 1160 // Accum is a vector, and we are accessing an element: ok. 1161 } else if ((PTy = dyn_cast<VectorType>(In)) && 1162 PTy->getElementType() == Accum) { 1163 // In is a vector, and accum is an element: ok, remember In. 1164 Accum = In; 1165 } else if ((PTy = dyn_cast<VectorType>(In)) && isa<VectorType>(Accum) && 1166 PTy->getBitWidth() == cast<VectorType>(Accum)->getBitWidth()) { 1167 // Two vectors of the same size: keep Accum. 1168 } else { 1169 // Cannot insert an short into a <4 x int> or handle 1170 // <2 x int> -> <4 x int> 1171 return true; 1172 } 1173 } else { 1174 // Pointer/FP/Integer unions merge together as integers. 1175 switch (Accum->getTypeID()) { 1176 case Type::PointerTyID: Accum = TD.getIntPtrType(); break; 1177 case Type::FloatTyID: Accum = Type::Int32Ty; break; 1178 case Type::DoubleTyID: Accum = Type::Int64Ty; break; 1179 case Type::X86_FP80TyID: return true; 1180 case Type::FP128TyID: return true; 1181 case Type::PPC_FP128TyID: return true; 1182 default: 1183 assert(Accum->isInteger() && "Unknown FP type!"); 1184 break; 1185 } 1186 1187 switch (In->getTypeID()) { 1188 case Type::PointerTyID: In = TD.getIntPtrType(); break; 1189 case Type::FloatTyID: In = Type::Int32Ty; break; 1190 case Type::DoubleTyID: In = Type::Int64Ty; break; 1191 case Type::X86_FP80TyID: return true; 1192 case Type::FP128TyID: return true; 1193 case Type::PPC_FP128TyID: return true; 1194 default: 1195 assert(In->isInteger() && "Unknown FP type!"); 1196 break; 1197 } 1198 return MergeInType(In, Accum, TD); 1199 } 1200 return false; 1201} 1202 1203/// getIntAtLeastAsBigAs - Return an integer type that is at least as big as the 1204/// specified type. If there is no suitable type, this returns null. 1205const Type *getIntAtLeastAsBigAs(unsigned NumBits) { 1206 if (NumBits > 64) return 0; 1207 if (NumBits > 32) return Type::Int64Ty; 1208 if (NumBits > 16) return Type::Int32Ty; 1209 if (NumBits > 8) return Type::Int16Ty; 1210 return Type::Int8Ty; 1211} 1212 1213/// CanConvertToScalar - V is a pointer. If we can convert the pointee to a 1214/// single scalar integer type, return that type. Further, if the use is not 1215/// a completely trivial use that mem2reg could promote, set IsNotTrivial. If 1216/// there are no uses of this pointer, return Type::VoidTy to differentiate from 1217/// failure. 1218/// 1219const Type *SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial) { 1220 const Type *UsedType = Type::VoidTy; // No uses, no forced type. 1221 const PointerType *PTy = cast<PointerType>(V->getType()); 1222 1223 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1224 Instruction *User = cast<Instruction>(*UI); 1225 1226 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1227 // FIXME: Loads of a first class aggregrate value could be converted to a 1228 // series of loads and insertvalues 1229 if (!LI->getType()->isSingleValueType()) 1230 return 0; 1231 1232 if (MergeInType(LI->getType(), UsedType, *TD)) 1233 return 0; 1234 continue; 1235 } 1236 1237 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1238 // Storing the pointer, not into the value? 1239 if (SI->getOperand(0) == V) return 0; 1240 1241 // FIXME: Stores of a first class aggregrate value could be converted to a 1242 // series of extractvalues and stores 1243 if (!SI->getOperand(0)->getType()->isSingleValueType()) 1244 return 0; 1245 1246 // NOTE: We could handle storing of FP imms into integers here! 1247 1248 if (MergeInType(SI->getOperand(0)->getType(), UsedType, *TD)) 1249 return 0; 1250 continue; 1251 } 1252 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1253 IsNotTrivial = true; 1254 const Type *SubTy = CanConvertToScalar(CI, IsNotTrivial); 1255 if (!SubTy || MergeInType(SubTy, UsedType, *TD)) return 0; 1256 continue; 1257 } 1258 1259 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1260 // Check to see if this is stepping over an element: GEP Ptr, int C 1261 if (GEP->getNumOperands() == 2 && isa<ConstantInt>(GEP->getOperand(1))) { 1262 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue(); 1263 unsigned ElSize = TD->getABITypeSize(PTy->getElementType()); 1264 unsigned BitOffset = Idx*ElSize*8; 1265 if (BitOffset > 64 || !isPowerOf2_32(ElSize)) return 0; 1266 1267 IsNotTrivial = true; 1268 const Type *SubElt = CanConvertToScalar(GEP, IsNotTrivial); 1269 if (SubElt == 0) return 0; 1270 if (SubElt != Type::VoidTy && SubElt->isInteger()) { 1271 const Type *NewTy = 1272 getIntAtLeastAsBigAs(TD->getABITypeSizeInBits(SubElt)+BitOffset); 1273 if (NewTy == 0 || MergeInType(NewTy, UsedType, *TD)) return 0; 1274 continue; 1275 } 1276 // Cannot handle this! 1277 return 0; 1278 } 1279 1280 if (GEP->getNumOperands() == 3 && 1281 isa<ConstantInt>(GEP->getOperand(1)) && 1282 isa<ConstantInt>(GEP->getOperand(2)) && 1283 cast<ConstantInt>(GEP->getOperand(1))->isZero()) { 1284 // We are stepping into an element, e.g. a structure or an array: 1285 // GEP Ptr, i32 0, i32 Cst 1286 const Type *AggTy = PTy->getElementType(); 1287 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 1288 1289 if (const ArrayType *ATy = dyn_cast<ArrayType>(AggTy)) { 1290 if (Idx >= ATy->getNumElements()) return 0; // Out of range. 1291 } else if (const VectorType *VectorTy = dyn_cast<VectorType>(AggTy)) { 1292 // Getting an element of the vector. 1293 if (Idx >= VectorTy->getNumElements()) return 0; // Out of range. 1294 1295 // Merge in the vector type. 1296 if (MergeInType(VectorTy, UsedType, *TD)) return 0; 1297 1298 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial); 1299 if (SubTy == 0) return 0; 1300 1301 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, *TD)) 1302 return 0; 1303 1304 // We'll need to change this to an insert/extract element operation. 1305 IsNotTrivial = true; 1306 continue; // Everything looks ok 1307 1308 } else if (isa<StructType>(AggTy)) { 1309 // Structs are always ok. 1310 } else { 1311 return 0; 1312 } 1313 const Type *NTy = getIntAtLeastAsBigAs(TD->getABITypeSizeInBits(AggTy)); 1314 if (NTy == 0 || MergeInType(NTy, UsedType, *TD)) return 0; 1315 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial); 1316 if (SubTy == 0) return 0; 1317 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, *TD)) 1318 return 0; 1319 continue; // Everything looks ok 1320 } 1321 return 0; 1322 } 1323 1324 // Cannot handle this! 1325 return 0; 1326 } 1327 1328 return UsedType; 1329} 1330 1331/// ConvertToScalar - The specified alloca passes the CanConvertToScalar 1332/// predicate and is non-trivial. Convert it to something that can be trivially 1333/// promoted into a register by mem2reg. 1334void SROA::ConvertToScalar(AllocationInst *AI, const Type *ActualTy) { 1335 DOUT << "CONVERT TO SCALAR: " << *AI << " TYPE = " 1336 << *ActualTy << "\n"; 1337 ++NumConverted; 1338 1339 BasicBlock *EntryBlock = AI->getParent(); 1340 assert(EntryBlock == &EntryBlock->getParent()->getEntryBlock() && 1341 "Not in the entry block!"); 1342 EntryBlock->getInstList().remove(AI); // Take the alloca out of the program. 1343 1344 // Create and insert the alloca. 1345 AllocaInst *NewAI = new AllocaInst(ActualTy, 0, AI->getName(), 1346 EntryBlock->begin()); 1347 ConvertUsesToScalar(AI, NewAI, 0); 1348 delete AI; 1349} 1350 1351 1352/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1353/// directly. This happens when we are converting an "integer union" to a 1354/// single integer scalar, or when we are converting a "vector union" to a 1355/// vector with insert/extractelement instructions. 1356/// 1357/// Offset is an offset from the original alloca, in bits that need to be 1358/// shifted to the right. By the end of this, there should be no uses of Ptr. 1359void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset) { 1360 while (!Ptr->use_empty()) { 1361 Instruction *User = cast<Instruction>(Ptr->use_back()); 1362 1363 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1364 Value *NV = ConvertUsesOfLoadToScalar(LI, NewAI, Offset); 1365 LI->replaceAllUsesWith(NV); 1366 LI->eraseFromParent(); 1367 continue; 1368 } 1369 1370 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1371 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1372 1373 Value *SV = ConvertUsesOfStoreToScalar(SI, NewAI, Offset); 1374 new StoreInst(SV, NewAI, SI); 1375 SI->eraseFromParent(); 1376 continue; 1377 } 1378 1379 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1380 ConvertUsesToScalar(CI, NewAI, Offset); 1381 CI->eraseFromParent(); 1382 continue; 1383 } 1384 1385 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1386 const PointerType *AggPtrTy = 1387 cast<PointerType>(GEP->getOperand(0)->getType()); 1388 unsigned AggSizeInBits = 1389 TD->getABITypeSizeInBits(AggPtrTy->getElementType()); 1390 1391 // Check to see if this is stepping over an element: GEP Ptr, int C 1392 unsigned NewOffset = Offset; 1393 if (GEP->getNumOperands() == 2) { 1394 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue(); 1395 unsigned BitOffset = Idx*AggSizeInBits; 1396 1397 NewOffset += BitOffset; 1398 ConvertUsesToScalar(GEP, NewAI, NewOffset); 1399 GEP->eraseFromParent(); 1400 continue; 1401 } 1402 1403 assert(GEP->getNumOperands() == 3 && "Unsupported operation"); 1404 1405 // We know that operand #2 is zero. 1406 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 1407 const Type *AggTy = AggPtrTy->getElementType(); 1408 if (const SequentialType *SeqTy = dyn_cast<SequentialType>(AggTy)) { 1409 unsigned ElSizeBits = 1410 TD->getABITypeSizeInBits(SeqTy->getElementType()); 1411 1412 NewOffset += ElSizeBits*Idx; 1413 } else { 1414 const StructType *STy = cast<StructType>(AggTy); 1415 unsigned EltBitOffset = 1416 TD->getStructLayout(STy)->getElementOffsetInBits(Idx); 1417 1418 NewOffset += EltBitOffset; 1419 } 1420 ConvertUsesToScalar(GEP, NewAI, NewOffset); 1421 GEP->eraseFromParent(); 1422 continue; 1423 } 1424 1425 assert(0 && "Unsupported operation!"); 1426 abort(); 1427 } 1428} 1429 1430/// ConvertUsesOfLoadToScalar - Convert all of the users the specified load to 1431/// use the new alloca directly, returning the value that should replace the 1432/// load. This happens when we are converting an "integer union" to a 1433/// single integer scalar, or when we are converting a "vector union" to a 1434/// vector with insert/extractelement instructions. 1435/// 1436/// Offset is an offset from the original alloca, in bits that need to be 1437/// shifted to the right. By the end of this, there should be no uses of Ptr. 1438Value *SROA::ConvertUsesOfLoadToScalar(LoadInst *LI, AllocaInst *NewAI, 1439 unsigned Offset) { 1440 // The load is a bit extract from NewAI shifted right by Offset bits. 1441 Value *NV = new LoadInst(NewAI, LI->getName(), LI); 1442 1443 if (NV->getType() == LI->getType() && Offset == 0) { 1444 // We win, no conversion needed. 1445 return NV; 1446 } 1447 1448 // If the result type of the 'union' is a pointer, then this must be ptr->ptr 1449 // cast. Anything else would result in NV being an integer. 1450 if (isa<PointerType>(NV->getType())) { 1451 assert(isa<PointerType>(LI->getType())); 1452 return new BitCastInst(NV, LI->getType(), LI->getName(), LI); 1453 } 1454 1455 if (const VectorType *VTy = dyn_cast<VectorType>(NV->getType())) { 1456 // If the result alloca is a vector type, this is either an element 1457 // access or a bitcast to another vector type. 1458 if (isa<VectorType>(LI->getType())) 1459 return new BitCastInst(NV, LI->getType(), LI->getName(), LI); 1460 1461 // Otherwise it must be an element access. 1462 unsigned Elt = 0; 1463 if (Offset) { 1464 unsigned EltSize = TD->getABITypeSizeInBits(VTy->getElementType()); 1465 Elt = Offset/EltSize; 1466 Offset -= EltSize*Elt; 1467 } 1468 NV = new ExtractElementInst(NV, ConstantInt::get(Type::Int32Ty, Elt), 1469 "tmp", LI); 1470 1471 // If we're done, return this element. 1472 if (NV->getType() == LI->getType() && Offset == 0) 1473 return NV; 1474 } 1475 1476 const IntegerType *NTy = cast<IntegerType>(NV->getType()); 1477 1478 // If this is a big-endian system and the load is narrower than the 1479 // full alloca type, we need to do a shift to get the right bits. 1480 int ShAmt = 0; 1481 if (TD->isBigEndian()) { 1482 // On big-endian machines, the lowest bit is stored at the bit offset 1483 // from the pointer given by getTypeStoreSizeInBits. This matters for 1484 // integers with a bitwidth that is not a multiple of 8. 1485 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1486 TD->getTypeStoreSizeInBits(LI->getType()) - Offset; 1487 } else { 1488 ShAmt = Offset; 1489 } 1490 1491 // Note: we support negative bitwidths (with shl) which are not defined. 1492 // We do this to support (f.e.) loads off the end of a structure where 1493 // only some bits are used. 1494 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1495 NV = BinaryOperator::CreateLShr(NV, 1496 ConstantInt::get(NV->getType(),ShAmt), 1497 LI->getName(), LI); 1498 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1499 NV = BinaryOperator::CreateShl(NV, 1500 ConstantInt::get(NV->getType(),-ShAmt), 1501 LI->getName(), LI); 1502 1503 // Finally, unconditionally truncate the integer to the right width. 1504 unsigned LIBitWidth = TD->getTypeSizeInBits(LI->getType()); 1505 if (LIBitWidth < NTy->getBitWidth()) 1506 NV = new TruncInst(NV, IntegerType::get(LIBitWidth), 1507 LI->getName(), LI); 1508 1509 // If the result is an integer, this is a trunc or bitcast. 1510 if (isa<IntegerType>(LI->getType())) { 1511 // Should be done. 1512 } else if (LI->getType()->isFloatingPoint()) { 1513 // Just do a bitcast, we know the sizes match up. 1514 NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI); 1515 } else { 1516 // Otherwise must be a pointer. 1517 NV = new IntToPtrInst(NV, LI->getType(), LI->getName(), LI); 1518 } 1519 assert(NV->getType() == LI->getType() && "Didn't convert right?"); 1520 return NV; 1521} 1522 1523 1524/// ConvertUsesOfStoreToScalar - Convert the specified store to a load+store 1525/// pair of the new alloca directly, returning the value that should be stored 1526/// to the alloca. This happens when we are converting an "integer union" to a 1527/// single integer scalar, or when we are converting a "vector union" to a 1528/// vector with insert/extractelement instructions. 1529/// 1530/// Offset is an offset from the original alloca, in bits that need to be 1531/// shifted to the right. By the end of this, there should be no uses of Ptr. 1532Value *SROA::ConvertUsesOfStoreToScalar(StoreInst *SI, AllocaInst *NewAI, 1533 unsigned Offset) { 1534 1535 // Convert the stored type to the actual type, shift it left to insert 1536 // then 'or' into place. 1537 Value *SV = SI->getOperand(0); 1538 const Type *AllocaType = NewAI->getType()->getElementType(); 1539 if (SV->getType() == AllocaType && Offset == 0) { 1540 // All is well. 1541 } else if (const VectorType *PTy = dyn_cast<VectorType>(AllocaType)) { 1542 Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI); 1543 1544 // If the result alloca is a vector type, this is either an element 1545 // access or a bitcast to another vector type. 1546 if (isa<VectorType>(SV->getType())) { 1547 SV = new BitCastInst(SV, AllocaType, SV->getName(), SI); 1548 } else { 1549 // Must be an element insertion. 1550 unsigned Elt = Offset/TD->getABITypeSizeInBits(PTy->getElementType()); 1551 SV = InsertElementInst::Create(Old, SV, 1552 ConstantInt::get(Type::Int32Ty, Elt), 1553 "tmp", SI); 1554 } 1555 } else if (isa<PointerType>(AllocaType)) { 1556 // If the alloca type is a pointer, then all the elements must be 1557 // pointers. 1558 if (SV->getType() != AllocaType) 1559 SV = new BitCastInst(SV, AllocaType, SV->getName(), SI); 1560 } else { 1561 Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI); 1562 1563 // If SV is a float, convert it to the appropriate integer type. 1564 // If it is a pointer, do the same, and also handle ptr->ptr casts 1565 // here. 1566 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1567 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1568 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1569 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1570 if (SV->getType()->isFloatingPoint()) 1571 SV = new BitCastInst(SV, IntegerType::get(SrcWidth), 1572 SV->getName(), SI); 1573 else if (isa<PointerType>(SV->getType())) 1574 SV = new PtrToIntInst(SV, TD->getIntPtrType(), SV->getName(), SI); 1575 1576 // Always zero extend the value if needed. 1577 if (SV->getType() != AllocaType) 1578 SV = new ZExtInst(SV, AllocaType, SV->getName(), SI); 1579 1580 // If this is a big-endian system and the store is narrower than the 1581 // full alloca type, we need to do a shift to get the right bits. 1582 int ShAmt = 0; 1583 if (TD->isBigEndian()) { 1584 // On big-endian machines, the lowest bit is stored at the bit offset 1585 // from the pointer given by getTypeStoreSizeInBits. This matters for 1586 // integers with a bitwidth that is not a multiple of 8. 1587 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1588 } else { 1589 ShAmt = Offset; 1590 } 1591 1592 // Note: we support negative bitwidths (with shr) which are not defined. 1593 // We do this to support (f.e.) stores off the end of a structure where 1594 // only some bits in the structure are set. 1595 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1596 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1597 SV = BinaryOperator::CreateShl(SV, 1598 ConstantInt::get(SV->getType(), ShAmt), 1599 SV->getName(), SI); 1600 Mask <<= ShAmt; 1601 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1602 SV = BinaryOperator::CreateLShr(SV, 1603 ConstantInt::get(SV->getType(),-ShAmt), 1604 SV->getName(), SI); 1605 Mask = Mask.lshr(ShAmt); 1606 } 1607 1608 // Mask out the bits we are about to insert from the old value, and or 1609 // in the new bits. 1610 if (SrcWidth != DestWidth) { 1611 assert(DestWidth > SrcWidth); 1612 Old = BinaryOperator::CreateAnd(Old, ConstantInt::get(~Mask), 1613 Old->getName()+".mask", SI); 1614 SV = BinaryOperator::CreateOr(Old, SV, SV->getName()+".ins", SI); 1615 } 1616 } 1617 return SV; 1618} 1619 1620 1621 1622/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1623/// some part of a constant global variable. This intentionally only accepts 1624/// constant expressions because we don't can't rewrite arbitrary instructions. 1625static bool PointsToConstantGlobal(Value *V) { 1626 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1627 return GV->isConstant(); 1628 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1629 if (CE->getOpcode() == Instruction::BitCast || 1630 CE->getOpcode() == Instruction::GetElementPtr) 1631 return PointsToConstantGlobal(CE->getOperand(0)); 1632 return false; 1633} 1634 1635/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1636/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1637/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1638/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1639/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1640/// the alloca, and if the source pointer is a pointer to a constant global, we 1641/// can optimize this. 1642static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1643 bool isOffset) { 1644 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1645 if (isa<LoadInst>(*UI)) { 1646 // Ignore loads, they are always ok. 1647 continue; 1648 } 1649 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1650 // If uses of the bitcast are ok, we are ok. 1651 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1652 return false; 1653 continue; 1654 } 1655 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1656 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1657 // doesn't, it does. 1658 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1659 isOffset || !GEP->hasAllZeroIndices())) 1660 return false; 1661 continue; 1662 } 1663 1664 // If this is isn't our memcpy/memmove, reject it as something we can't 1665 // handle. 1666 if (!isa<MemCpyInst>(*UI) && !isa<MemMoveInst>(*UI)) 1667 return false; 1668 1669 // If we already have seen a copy, reject the second one. 1670 if (TheCopy) return false; 1671 1672 // If the pointer has been offset from the start of the alloca, we can't 1673 // safely handle this. 1674 if (isOffset) return false; 1675 1676 // If the memintrinsic isn't using the alloca as the dest, reject it. 1677 if (UI.getOperandNo() != 1) return false; 1678 1679 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1680 1681 // If the source of the memcpy/move is not a constant global, reject it. 1682 if (!PointsToConstantGlobal(MI->getOperand(2))) 1683 return false; 1684 1685 // Otherwise, the transform is safe. Remember the copy instruction. 1686 TheCopy = MI; 1687 } 1688 return true; 1689} 1690 1691/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1692/// modified by a copy from a constant global. If we can prove this, we can 1693/// replace any uses of the alloca with uses of the global directly. 1694Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) { 1695 Instruction *TheCopy = 0; 1696 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1697 return TheCopy; 1698 return 0; 1699} 1700