ScalarReplAggregates.cpp revision 4afc90dacf309999d8b7f6c2b4b0c56af346bab5
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/Transforms/Utils/Local.h" 35#include "llvm/Support/Debug.h" 36#include "llvm/Support/GetElementPtrTypeIterator.h" 37#include "llvm/Support/IRBuilder.h" 38#include "llvm/Support/MathExtras.h" 39#include "llvm/Support/Compiler.h" 40#include "llvm/ADT/SmallVector.h" 41#include "llvm/ADT/Statistic.h" 42#include "llvm/ADT/StringExtras.h" 43using namespace llvm; 44 45STATISTIC(NumReplaced, "Number of allocas broken up"); 46STATISTIC(NumPromoted, "Number of allocas promoted"); 47STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 48STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 49 50namespace { 51 struct VISIBILITY_HIDDEN SROA : public FunctionPass { 52 static char ID; // Pass identification, replacement for typeid 53 explicit SROA(signed T = -1) : FunctionPass(&ID) { 54 if (T == -1) 55 SRThreshold = 128; 56 else 57 SRThreshold = T; 58 } 59 60 bool runOnFunction(Function &F); 61 62 bool performScalarRepl(Function &F); 63 bool performPromotion(Function &F); 64 65 // getAnalysisUsage - This pass does not require any passes, but we know it 66 // will not alter the CFG, so say so. 67 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 68 AU.addRequired<DominatorTree>(); 69 AU.addRequired<DominanceFrontier>(); 70 AU.addRequired<TargetData>(); 71 AU.setPreservesCFG(); 72 } 73 74 private: 75 TargetData *TD; 76 77 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 78 /// information about the uses. All these fields are initialized to false 79 /// and set to true when something is learned. 80 struct AllocaInfo { 81 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 82 bool isUnsafe : 1; 83 84 /// needsCleanup - This is set to true if there is some use of the alloca 85 /// that requires cleanup. 86 bool needsCleanup : 1; 87 88 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 89 bool isMemCpySrc : 1; 90 91 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 92 bool isMemCpyDst : 1; 93 94 AllocaInfo() 95 : isUnsafe(false), needsCleanup(false), 96 isMemCpySrc(false), isMemCpyDst(false) {} 97 }; 98 99 unsigned SRThreshold; 100 101 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 102 103 int isSafeAllocaToScalarRepl(AllocationInst *AI); 104 105 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 106 AllocaInfo &Info); 107 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 108 AllocaInfo &Info); 109 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 110 unsigned OpNo, AllocaInfo &Info); 111 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI, 112 AllocaInfo &Info); 113 114 void DoScalarReplacement(AllocationInst *AI, 115 std::vector<AllocationInst*> &WorkList); 116 void CleanupGEP(GetElementPtrInst *GEP); 117 void CleanupAllocaUsers(AllocationInst *AI); 118 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base); 119 120 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 121 SmallVector<AllocaInst*, 32> &NewElts); 122 123 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 124 AllocationInst *AI, 125 SmallVector<AllocaInst*, 32> &NewElts); 126 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI, 127 SmallVector<AllocaInst*, 32> &NewElts); 128 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 129 SmallVector<AllocaInst*, 32> &NewElts); 130 131 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 132 bool &SawVec, uint64_t Offset, unsigned AllocaSize); 133 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 134 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 135 uint64_t Offset, IRBuilder<> &Builder); 136 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 137 uint64_t Offset, IRBuilder<> &Builder); 138 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI); 139 }; 140} 141 142char SROA::ID = 0; 143static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 144 145// Public interface to the ScalarReplAggregates pass 146FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 147 return new SROA(Threshold); 148} 149 150 151bool SROA::runOnFunction(Function &F) { 152 TD = &getAnalysis<TargetData>(); 153 154 bool Changed = performPromotion(F); 155 while (1) { 156 bool LocalChange = performScalarRepl(F); 157 if (!LocalChange) break; // No need to repromote if no scalarrepl 158 Changed = true; 159 LocalChange = performPromotion(F); 160 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 161 } 162 163 return Changed; 164} 165 166 167bool SROA::performPromotion(Function &F) { 168 std::vector<AllocaInst*> Allocas; 169 DominatorTree &DT = getAnalysis<DominatorTree>(); 170 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 171 172 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 173 174 bool Changed = false; 175 176 while (1) { 177 Allocas.clear(); 178 179 // Find allocas that are safe to promote, by looking at all instructions in 180 // the entry node 181 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 182 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 183 if (isAllocaPromotable(AI)) 184 Allocas.push_back(AI); 185 186 if (Allocas.empty()) break; 187 188 PromoteMemToReg(Allocas, DT, DF); 189 NumPromoted += Allocas.size(); 190 Changed = true; 191 } 192 193 return Changed; 194} 195 196/// getNumSAElements - Return the number of elements in the specific struct or 197/// array. 198static uint64_t getNumSAElements(const Type *T) { 199 if (const StructType *ST = dyn_cast<StructType>(T)) 200 return ST->getNumElements(); 201 return cast<ArrayType>(T)->getNumElements(); 202} 203 204// performScalarRepl - This algorithm is a simple worklist driven algorithm, 205// which runs on all of the malloc/alloca instructions in the function, removing 206// them if they are only used by getelementptr instructions. 207// 208bool SROA::performScalarRepl(Function &F) { 209 std::vector<AllocationInst*> WorkList; 210 211 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 212 BasicBlock &BB = F.getEntryBlock(); 213 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 214 if (AllocationInst *A = dyn_cast<AllocationInst>(I)) 215 WorkList.push_back(A); 216 217 // Process the worklist 218 bool Changed = false; 219 while (!WorkList.empty()) { 220 AllocationInst *AI = WorkList.back(); 221 WorkList.pop_back(); 222 223 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 224 // with unused elements. 225 if (AI->use_empty()) { 226 AI->eraseFromParent(); 227 continue; 228 } 229 230 // If this alloca is impossible for us to promote, reject it early. 231 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 232 continue; 233 234 // Check to see if this allocation is only modified by a memcpy/memmove from 235 // a constant global. If this is the case, we can change all users to use 236 // the constant global instead. This is commonly produced by the CFE by 237 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 238 // is only subsequently read. 239 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 240 DOUT << "Found alloca equal to global: " << *AI; 241 DOUT << " memcpy = " << *TheCopy; 242 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 243 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 244 TheCopy->eraseFromParent(); // Don't mutate the global. 245 AI->eraseFromParent(); 246 ++NumGlobals; 247 Changed = true; 248 continue; 249 } 250 251 // Check to see if we can perform the core SROA transformation. We cannot 252 // transform the allocation instruction if it is an array allocation 253 // (allocations OF arrays are ok though), and an allocation of a scalar 254 // value cannot be decomposed at all. 255 uint64_t AllocaSize = TD->getTypePaddedSize(AI->getAllocatedType()); 256 257 if ((isa<StructType>(AI->getAllocatedType()) || 258 isa<ArrayType>(AI->getAllocatedType())) && 259 // Do not promote any struct whose size is too big. 260 AllocaSize < SRThreshold && 261 // Do not promote any struct into more than "32" separate vars. 262 getNumSAElements(AI->getAllocatedType()) < SRThreshold/4) { 263 // Check that all of the users of the allocation are capable of being 264 // transformed. 265 switch (isSafeAllocaToScalarRepl(AI)) { 266 default: assert(0 && "Unexpected value!"); 267 case 0: // Not safe to scalar replace. 268 break; 269 case 1: // Safe, but requires cleanup/canonicalizations first 270 CleanupAllocaUsers(AI); 271 // FALL THROUGH. 272 case 3: // Safe to scalar replace. 273 DoScalarReplacement(AI, WorkList); 274 Changed = true; 275 continue; 276 } 277 } 278 279 // If we can turn this aggregate value (potentially with casts) into a 280 // simple scalar value that can be mem2reg'd into a register value. 281 // IsNotTrivial tracks whether this is something that mem2reg could have 282 // promoted itself. If so, we don't want to transform it needlessly. Note 283 // that we can't just check based on the type: the alloca may be of an i32 284 // but that has pointer arithmetic to set byte 3 of it or something. 285 bool IsNotTrivial = false; 286 const Type *VectorTy = 0; 287 bool HadAVector = false; 288 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector, 289 0, unsigned(AllocaSize)) && IsNotTrivial) { 290 AllocaInst *NewAI; 291 // If we were able to find a vector type that can handle this with 292 // insert/extract elements, and if there was at least one use that had 293 // a vector type, promote this to a vector. We don't want to promote 294 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 295 // we just get a lot of insert/extracts. If at least one vector is 296 // involved, then we probably really do have a union of vector/array. 297 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) { 298 DOUT << "CONVERT TO VECTOR: " << *AI << " TYPE = " << *VectorTy <<"\n"; 299 300 // Create and insert the vector alloca. 301 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin()); 302 ConvertUsesToScalar(AI, NewAI, 0); 303 } else { 304 DOUT << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"; 305 306 // Create and insert the integer alloca. 307 const Type *NewTy = IntegerType::get(AllocaSize*8); 308 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 309 ConvertUsesToScalar(AI, NewAI, 0); 310 } 311 NewAI->takeName(AI); 312 AI->eraseFromParent(); 313 ++NumConverted; 314 Changed = true; 315 continue; 316 } 317 318 // Otherwise, couldn't process this alloca. 319 } 320 321 return Changed; 322} 323 324/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 325/// predicate, do SROA now. 326void SROA::DoScalarReplacement(AllocationInst *AI, 327 std::vector<AllocationInst*> &WorkList) { 328 DOUT << "Found inst to SROA: " << *AI; 329 SmallVector<AllocaInst*, 32> ElementAllocas; 330 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 331 ElementAllocas.reserve(ST->getNumContainedTypes()); 332 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 333 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 334 AI->getAlignment(), 335 AI->getName() + "." + utostr(i), AI); 336 ElementAllocas.push_back(NA); 337 WorkList.push_back(NA); // Add to worklist for recursive processing 338 } 339 } else { 340 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 341 ElementAllocas.reserve(AT->getNumElements()); 342 const Type *ElTy = AT->getElementType(); 343 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 344 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 345 AI->getName() + "." + utostr(i), AI); 346 ElementAllocas.push_back(NA); 347 WorkList.push_back(NA); // Add to worklist for recursive processing 348 } 349 } 350 351 // Now that we have created the alloca instructions that we want to use, 352 // expand the getelementptr instructions to use them. 353 // 354 while (!AI->use_empty()) { 355 Instruction *User = cast<Instruction>(AI->use_back()); 356 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) { 357 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas); 358 BCInst->eraseFromParent(); 359 continue; 360 } 361 362 // Replace: 363 // %res = load { i32, i32 }* %alloc 364 // with: 365 // %load.0 = load i32* %alloc.0 366 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 367 // %load.1 = load i32* %alloc.1 368 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 369 // (Also works for arrays instead of structs) 370 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 371 Value *Insert = UndefValue::get(LI->getType()); 372 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 373 Value *Load = new LoadInst(ElementAllocas[i], "load", LI); 374 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 375 } 376 LI->replaceAllUsesWith(Insert); 377 LI->eraseFromParent(); 378 continue; 379 } 380 381 // Replace: 382 // store { i32, i32 } %val, { i32, i32 }* %alloc 383 // with: 384 // %val.0 = extractvalue { i32, i32 } %val, 0 385 // store i32 %val.0, i32* %alloc.0 386 // %val.1 = extractvalue { i32, i32 } %val, 1 387 // store i32 %val.1, i32* %alloc.1 388 // (Also works for arrays instead of structs) 389 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 390 Value *Val = SI->getOperand(0); 391 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 392 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 393 new StoreInst(Extract, ElementAllocas[i], SI); 394 } 395 SI->eraseFromParent(); 396 continue; 397 } 398 399 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 400 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 401 unsigned Idx = 402 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 403 404 assert(Idx < ElementAllocas.size() && "Index out of range?"); 405 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 406 407 Value *RepValue; 408 if (GEPI->getNumOperands() == 3) { 409 // Do not insert a new getelementptr instruction with zero indices, only 410 // to have it optimized out later. 411 RepValue = AllocaToUse; 412 } else { 413 // We are indexing deeply into the structure, so we still need a 414 // getelement ptr instruction to finish the indexing. This may be 415 // expanded itself once the worklist is rerun. 416 // 417 SmallVector<Value*, 8> NewArgs; 418 NewArgs.push_back(Constant::getNullValue(Type::Int32Ty)); 419 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end()); 420 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(), 421 NewArgs.end(), "", GEPI); 422 RepValue->takeName(GEPI); 423 } 424 425 // If this GEP is to the start of the aggregate, check for memcpys. 426 if (Idx == 0 && GEPI->hasAllZeroIndices()) 427 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas); 428 429 // Move all of the users over to the new GEP. 430 GEPI->replaceAllUsesWith(RepValue); 431 // Delete the old GEP 432 GEPI->eraseFromParent(); 433 } 434 435 // Finally, delete the Alloca instruction 436 AI->eraseFromParent(); 437 NumReplaced++; 438} 439 440 441/// isSafeElementUse - Check to see if this use is an allowed use for a 442/// getelementptr instruction of an array aggregate allocation. isFirstElt 443/// indicates whether Ptr is known to the start of the aggregate. 444/// 445void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 446 AllocaInfo &Info) { 447 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 448 I != E; ++I) { 449 Instruction *User = cast<Instruction>(*I); 450 switch (User->getOpcode()) { 451 case Instruction::Load: break; 452 case Instruction::Store: 453 // Store is ok if storing INTO the pointer, not storing the pointer 454 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info); 455 break; 456 case Instruction::GetElementPtr: { 457 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 458 bool AreAllZeroIndices = isFirstElt; 459 if (GEP->getNumOperands() > 1) { 460 if (!isa<ConstantInt>(GEP->getOperand(1)) || 461 !cast<ConstantInt>(GEP->getOperand(1))->isZero()) 462 // Using pointer arithmetic to navigate the array. 463 return MarkUnsafe(Info); 464 465 if (AreAllZeroIndices) 466 AreAllZeroIndices = GEP->hasAllZeroIndices(); 467 } 468 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 469 if (Info.isUnsafe) return; 470 break; 471 } 472 case Instruction::BitCast: 473 if (isFirstElt) { 474 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 475 if (Info.isUnsafe) return; 476 break; 477 } 478 DOUT << " Transformation preventing inst: " << *User; 479 return MarkUnsafe(Info); 480 case Instruction::Call: 481 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 482 if (isFirstElt) { 483 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 484 if (Info.isUnsafe) return; 485 break; 486 } 487 } 488 DOUT << " Transformation preventing inst: " << *User; 489 return MarkUnsafe(Info); 490 default: 491 DOUT << " Transformation preventing inst: " << *User; 492 return MarkUnsafe(Info); 493 } 494 } 495 return; // All users look ok :) 496} 497 498/// AllUsersAreLoads - Return true if all users of this value are loads. 499static bool AllUsersAreLoads(Value *Ptr) { 500 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 501 I != E; ++I) 502 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 503 return false; 504 return true; 505} 506 507/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 508/// aggregate allocation. 509/// 510void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 511 AllocaInfo &Info) { 512 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 513 return isSafeUseOfBitCastedAllocation(C, AI, Info); 514 515 if (LoadInst *LI = dyn_cast<LoadInst>(User)) 516 if (!LI->isVolatile()) 517 return;// Loads (returning a first class aggregrate) are always rewritable 518 519 if (StoreInst *SI = dyn_cast<StoreInst>(User)) 520 if (!SI->isVolatile() && SI->getOperand(0) != AI) 521 return;// Store is ok if storing INTO the pointer, not storing the pointer 522 523 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 524 if (GEPI == 0) 525 return MarkUnsafe(Info); 526 527 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 528 529 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 530 if (I == E || 531 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) { 532 return MarkUnsafe(Info); 533 } 534 535 ++I; 536 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 537 538 bool IsAllZeroIndices = true; 539 540 // If the first index is a non-constant index into an array, see if we can 541 // handle it as a special case. 542 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 543 if (!isa<ConstantInt>(I.getOperand())) { 544 IsAllZeroIndices = 0; 545 uint64_t NumElements = AT->getNumElements(); 546 547 // If this is an array index and the index is not constant, we cannot 548 // promote... that is unless the array has exactly one or two elements in 549 // it, in which case we CAN promote it, but we have to canonicalize this 550 // out if this is the only problem. 551 if ((NumElements == 1 || NumElements == 2) && 552 AllUsersAreLoads(GEPI)) { 553 Info.needsCleanup = true; 554 return; // Canonicalization required! 555 } 556 return MarkUnsafe(Info); 557 } 558 } 559 560 // Walk through the GEP type indices, checking the types that this indexes 561 // into. 562 for (; I != E; ++I) { 563 // Ignore struct elements, no extra checking needed for these. 564 if (isa<StructType>(*I)) 565 continue; 566 567 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 568 if (!IdxVal) return MarkUnsafe(Info); 569 570 // Are all indices still zero? 571 IsAllZeroIndices &= IdxVal->isZero(); 572 573 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 574 // This GEP indexes an array. Verify that this is an in-range constant 575 // integer. Specifically, consider A[0][i]. We cannot know that the user 576 // isn't doing invalid things like allowing i to index an out-of-range 577 // subscript that accesses A[1]. Because of this, we have to reject SROA 578 // of any accesses into structs where any of the components are variables. 579 if (IdxVal->getZExtValue() >= AT->getNumElements()) 580 return MarkUnsafe(Info); 581 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) { 582 if (IdxVal->getZExtValue() >= VT->getNumElements()) 583 return MarkUnsafe(Info); 584 } 585 } 586 587 // If there are any non-simple uses of this getelementptr, make sure to reject 588 // them. 589 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 590} 591 592/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory 593/// intrinsic can be promoted by SROA. At this point, we know that the operand 594/// of the memintrinsic is a pointer to the beginning of the allocation. 595void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 596 unsigned OpNo, AllocaInfo &Info) { 597 // If not constant length, give up. 598 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 599 if (!Length) return MarkUnsafe(Info); 600 601 // If not the whole aggregate, give up. 602 if (Length->getZExtValue() != 603 TD->getTypePaddedSize(AI->getType()->getElementType())) 604 return MarkUnsafe(Info); 605 606 // We only know about memcpy/memset/memmove. 607 if (!isa<MemCpyInst>(MI) && !isa<MemSetInst>(MI) && !isa<MemMoveInst>(MI)) 608 return MarkUnsafe(Info); 609 610 // Otherwise, we can transform it. Determine whether this is a memcpy/set 611 // into or out of the aggregate. 612 if (OpNo == 1) 613 Info.isMemCpyDst = true; 614 else { 615 assert(OpNo == 2); 616 Info.isMemCpySrc = true; 617 } 618} 619 620/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast 621/// are 622void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI, 623 AllocaInfo &Info) { 624 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 625 UI != E; ++UI) { 626 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 627 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 628 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 629 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 630 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 631 if (SI->isVolatile()) 632 return MarkUnsafe(Info); 633 634 // If storing the entire alloca in one chunk through a bitcasted pointer 635 // to integer, we can transform it. This happens (for example) when you 636 // cast a {i32,i32}* to i64* and store through it. This is similar to the 637 // memcpy case and occurs in various "byval" cases and emulated memcpys. 638 if (isa<IntegerType>(SI->getOperand(0)->getType()) && 639 TD->getTypePaddedSize(SI->getOperand(0)->getType()) == 640 TD->getTypePaddedSize(AI->getType()->getElementType())) { 641 Info.isMemCpyDst = true; 642 continue; 643 } 644 return MarkUnsafe(Info); 645 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 646 if (LI->isVolatile()) 647 return MarkUnsafe(Info); 648 649 // If loading the entire alloca in one chunk through a bitcasted pointer 650 // to integer, we can transform it. This happens (for example) when you 651 // cast a {i32,i32}* to i64* and load through it. This is similar to the 652 // memcpy case and occurs in various "byval" cases and emulated memcpys. 653 if (isa<IntegerType>(LI->getType()) && 654 TD->getTypePaddedSize(LI->getType()) == 655 TD->getTypePaddedSize(AI->getType()->getElementType())) { 656 Info.isMemCpySrc = true; 657 continue; 658 } 659 return MarkUnsafe(Info); 660 } else if (isa<DbgInfoIntrinsic>(UI)) { 661 // If one user is DbgInfoIntrinsic then check if all users are 662 // DbgInfoIntrinsics. 663 if (OnlyUsedByDbgInfoIntrinsics(BC)) { 664 Info.needsCleanup = true; 665 return; 666 } 667 else 668 MarkUnsafe(Info); 669 } 670 else { 671 return MarkUnsafe(Info); 672 } 673 if (Info.isUnsafe) return; 674 } 675} 676 677/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 678/// to its first element. Transform users of the cast to use the new values 679/// instead. 680void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 681 SmallVector<AllocaInst*, 32> &NewElts) { 682 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 683 while (UI != UE) { 684 Instruction *User = cast<Instruction>(*UI++); 685 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) { 686 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 687 if (BCU->use_empty()) BCU->eraseFromParent(); 688 continue; 689 } 690 691 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 692 // This must be memcpy/memmove/memset of the entire aggregate. 693 // Split into one per element. 694 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts); 695 continue; 696 } 697 698 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 699 // If this is a store of the entire alloca from an integer, rewrite it. 700 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 701 continue; 702 } 703 704 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 705 // If this is a load of the entire alloca to an integer, rewrite it. 706 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 707 continue; 708 } 709 710 // Otherwise it must be some other user of a gep of the first pointer. Just 711 // leave these alone. 712 continue; 713 } 714} 715 716/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 717/// Rewrite it to copy or set the elements of the scalarized memory. 718void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 719 AllocationInst *AI, 720 SmallVector<AllocaInst*, 32> &NewElts) { 721 722 // If this is a memcpy/memmove, construct the other pointer as the 723 // appropriate type. 724 Value *OtherPtr = 0; 725 if (MemCpyInst *MCI = dyn_cast<MemCpyInst>(MI)) { 726 if (BCInst == MCI->getRawDest()) 727 OtherPtr = MCI->getRawSource(); 728 else { 729 assert(BCInst == MCI->getRawSource()); 730 OtherPtr = MCI->getRawDest(); 731 } 732 } else if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) { 733 if (BCInst == MMI->getRawDest()) 734 OtherPtr = MMI->getRawSource(); 735 else { 736 assert(BCInst == MMI->getRawSource()); 737 OtherPtr = MMI->getRawDest(); 738 } 739 } 740 741 // If there is an other pointer, we want to convert it to the same pointer 742 // type as AI has, so we can GEP through it safely. 743 if (OtherPtr) { 744 // It is likely that OtherPtr is a bitcast, if so, remove it. 745 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 746 OtherPtr = BC->getOperand(0); 747 // All zero GEPs are effectively bitcasts. 748 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) 749 if (GEP->hasAllZeroIndices()) 750 OtherPtr = GEP->getOperand(0); 751 752 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 753 if (BCE->getOpcode() == Instruction::BitCast) 754 OtherPtr = BCE->getOperand(0); 755 756 // If the pointer is not the right type, insert a bitcast to the right 757 // type. 758 if (OtherPtr->getType() != AI->getType()) 759 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 760 MI); 761 } 762 763 // Process each element of the aggregate. 764 Value *TheFn = MI->getOperand(0); 765 const Type *BytePtrTy = MI->getRawDest()->getType(); 766 bool SROADest = MI->getRawDest() == BCInst; 767 768 Constant *Zero = Constant::getNullValue(Type::Int32Ty); 769 770 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 771 // If this is a memcpy/memmove, emit a GEP of the other element address. 772 Value *OtherElt = 0; 773 if (OtherPtr) { 774 Value *Idx[2] = { Zero, ConstantInt::get(Type::Int32Ty, i) }; 775 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2, 776 OtherPtr->getNameStr()+"."+utostr(i), 777 MI); 778 } 779 780 Value *EltPtr = NewElts[i]; 781 const Type *EltTy =cast<PointerType>(EltPtr->getType())->getElementType(); 782 783 // If we got down to a scalar, insert a load or store as appropriate. 784 if (EltTy->isSingleValueType()) { 785 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) { 786 Value *Elt = new LoadInst(SROADest ? OtherElt : EltPtr, "tmp", 787 MI); 788 new StoreInst(Elt, SROADest ? EltPtr : OtherElt, MI); 789 continue; 790 } 791 assert(isa<MemSetInst>(MI)); 792 793 // If the stored element is zero (common case), just store a null 794 // constant. 795 Constant *StoreVal; 796 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 797 if (CI->isZero()) { 798 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 799 } else { 800 // If EltTy is a vector type, get the element type. 801 const Type *ValTy = EltTy; 802 if (const VectorType *VTy = dyn_cast<VectorType>(ValTy)) 803 ValTy = VTy->getElementType(); 804 805 // Construct an integer with the right value. 806 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 807 APInt OneVal(EltSize, CI->getZExtValue()); 808 APInt TotalVal(OneVal); 809 // Set each byte. 810 for (unsigned i = 0; 8*i < EltSize; ++i) { 811 TotalVal = TotalVal.shl(8); 812 TotalVal |= OneVal; 813 } 814 815 // Convert the integer value to the appropriate type. 816 StoreVal = ConstantInt::get(TotalVal); 817 if (isa<PointerType>(ValTy)) 818 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 819 else if (ValTy->isFloatingPoint()) 820 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 821 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 822 823 // If the requested value was a vector constant, create it. 824 if (EltTy != ValTy) { 825 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 826 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 827 StoreVal = ConstantVector::get(&Elts[0], NumElts); 828 } 829 } 830 new StoreInst(StoreVal, EltPtr, MI); 831 continue; 832 } 833 // Otherwise, if we're storing a byte variable, use a memset call for 834 // this element. 835 } 836 837 // Cast the element pointer to BytePtrTy. 838 if (EltPtr->getType() != BytePtrTy) 839 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 840 841 // Cast the other pointer (if we have one) to BytePtrTy. 842 if (OtherElt && OtherElt->getType() != BytePtrTy) 843 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 844 MI); 845 846 unsigned EltSize = TD->getTypePaddedSize(EltTy); 847 848 // Finally, insert the meminst for this element. 849 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) { 850 Value *Ops[] = { 851 SROADest ? EltPtr : OtherElt, // Dest ptr 852 SROADest ? OtherElt : EltPtr, // Src ptr 853 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 854 Zero // Align 855 }; 856 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 857 } else { 858 assert(isa<MemSetInst>(MI)); 859 Value *Ops[] = { 860 EltPtr, MI->getOperand(2), // Dest, Value, 861 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 862 Zero // Align 863 }; 864 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 865 } 866 } 867 MI->eraseFromParent(); 868} 869 870/// RewriteStoreUserOfWholeAlloca - We found an store of an integer that 871/// overwrites the entire allocation. Extract out the pieces of the stored 872/// integer and store them individually. 873void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, 874 AllocationInst *AI, 875 SmallVector<AllocaInst*, 32> &NewElts){ 876 // Extract each element out of the integer according to its structure offset 877 // and store the element value to the individual alloca. 878 Value *SrcVal = SI->getOperand(0); 879 const Type *AllocaEltTy = AI->getType()->getElementType(); 880 uint64_t AllocaSizeBits = TD->getTypePaddedSizeInBits(AllocaEltTy); 881 882 // If this isn't a store of an integer to the whole alloca, it may be a store 883 // to the first element. Just ignore the store in this case and normal SROA 884 // will handle it. 885 if (!isa<IntegerType>(SrcVal->getType()) || 886 TD->getTypePaddedSizeInBits(SrcVal->getType()) != AllocaSizeBits) 887 return; 888 889 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI; 890 891 // There are two forms here: AI could be an array or struct. Both cases 892 // have different ways to compute the element offset. 893 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 894 const StructLayout *Layout = TD->getStructLayout(EltSTy); 895 896 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 897 // Get the number of bits to shift SrcVal to get the value. 898 const Type *FieldTy = EltSTy->getElementType(i); 899 uint64_t Shift = Layout->getElementOffsetInBits(i); 900 901 if (TD->isBigEndian()) 902 Shift = AllocaSizeBits-Shift-TD->getTypePaddedSizeInBits(FieldTy); 903 904 Value *EltVal = SrcVal; 905 if (Shift) { 906 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 907 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 908 "sroa.store.elt", SI); 909 } 910 911 // Truncate down to an integer of the right size. 912 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 913 914 // Ignore zero sized fields like {}, they obviously contain no data. 915 if (FieldSizeBits == 0) continue; 916 917 if (FieldSizeBits != AllocaSizeBits) 918 EltVal = new TruncInst(EltVal, IntegerType::get(FieldSizeBits), "", SI); 919 Value *DestField = NewElts[i]; 920 if (EltVal->getType() == FieldTy) { 921 // Storing to an integer field of this size, just do it. 922 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { 923 // Bitcast to the right element type (for fp/vector values). 924 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 925 } else { 926 // Otherwise, bitcast the dest pointer (for aggregates). 927 DestField = new BitCastInst(DestField, 928 PointerType::getUnqual(EltVal->getType()), 929 "", SI); 930 } 931 new StoreInst(EltVal, DestField, SI); 932 } 933 934 } else { 935 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 936 const Type *ArrayEltTy = ATy->getElementType(); 937 uint64_t ElementOffset = TD->getTypePaddedSizeInBits(ArrayEltTy); 938 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 939 940 uint64_t Shift; 941 942 if (TD->isBigEndian()) 943 Shift = AllocaSizeBits-ElementOffset; 944 else 945 Shift = 0; 946 947 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 948 // Ignore zero sized fields like {}, they obviously contain no data. 949 if (ElementSizeBits == 0) continue; 950 951 Value *EltVal = SrcVal; 952 if (Shift) { 953 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 954 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 955 "sroa.store.elt", SI); 956 } 957 958 // Truncate down to an integer of the right size. 959 if (ElementSizeBits != AllocaSizeBits) 960 EltVal = new TruncInst(EltVal, IntegerType::get(ElementSizeBits),"",SI); 961 Value *DestField = NewElts[i]; 962 if (EltVal->getType() == ArrayEltTy) { 963 // Storing to an integer field of this size, just do it. 964 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { 965 // Bitcast to the right element type (for fp/vector values). 966 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 967 } else { 968 // Otherwise, bitcast the dest pointer (for aggregates). 969 DestField = new BitCastInst(DestField, 970 PointerType::getUnqual(EltVal->getType()), 971 "", SI); 972 } 973 new StoreInst(EltVal, DestField, SI); 974 975 if (TD->isBigEndian()) 976 Shift -= ElementOffset; 977 else 978 Shift += ElementOffset; 979 } 980 } 981 982 SI->eraseFromParent(); 983} 984 985/// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to 986/// an integer. Load the individual pieces to form the aggregate value. 987void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 988 SmallVector<AllocaInst*, 32> &NewElts) { 989 // Extract each element out of the NewElts according to its structure offset 990 // and form the result value. 991 const Type *AllocaEltTy = AI->getType()->getElementType(); 992 uint64_t AllocaSizeBits = TD->getTypePaddedSizeInBits(AllocaEltTy); 993 994 // If this isn't a load of the whole alloca to an integer, it may be a load 995 // of the first element. Just ignore the load in this case and normal SROA 996 // will handle it. 997 if (!isa<IntegerType>(LI->getType()) || 998 TD->getTypePaddedSizeInBits(LI->getType()) != AllocaSizeBits) 999 return; 1000 1001 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI; 1002 1003 // There are two forms here: AI could be an array or struct. Both cases 1004 // have different ways to compute the element offset. 1005 const StructLayout *Layout = 0; 1006 uint64_t ArrayEltBitOffset = 0; 1007 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1008 Layout = TD->getStructLayout(EltSTy); 1009 } else { 1010 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1011 ArrayEltBitOffset = TD->getTypePaddedSizeInBits(ArrayEltTy); 1012 } 1013 1014 Value *ResultVal = Constant::getNullValue(LI->getType()); 1015 1016 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1017 // Load the value from the alloca. If the NewElt is an aggregate, cast 1018 // the pointer to an integer of the same size before doing the load. 1019 Value *SrcField = NewElts[i]; 1020 const Type *FieldTy = 1021 cast<PointerType>(SrcField->getType())->getElementType(); 1022 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1023 1024 // Ignore zero sized fields like {}, they obviously contain no data. 1025 if (FieldSizeBits == 0) continue; 1026 1027 const IntegerType *FieldIntTy = IntegerType::get(FieldSizeBits); 1028 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && 1029 !isa<VectorType>(FieldTy)) 1030 SrcField = new BitCastInst(SrcField, PointerType::getUnqual(FieldIntTy), 1031 "", LI); 1032 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1033 1034 // If SrcField is a fp or vector of the right size but that isn't an 1035 // integer type, bitcast to an integer so we can shift it. 1036 if (SrcField->getType() != FieldIntTy) 1037 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1038 1039 // Zero extend the field to be the same size as the final alloca so that 1040 // we can shift and insert it. 1041 if (SrcField->getType() != ResultVal->getType()) 1042 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1043 1044 // Determine the number of bits to shift SrcField. 1045 uint64_t Shift; 1046 if (Layout) // Struct case. 1047 Shift = Layout->getElementOffsetInBits(i); 1048 else // Array case. 1049 Shift = i*ArrayEltBitOffset; 1050 1051 if (TD->isBigEndian()) 1052 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1053 1054 if (Shift) { 1055 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1056 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1057 } 1058 1059 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1060 } 1061 1062 LI->replaceAllUsesWith(ResultVal); 1063 LI->eraseFromParent(); 1064} 1065 1066 1067/// HasPadding - Return true if the specified type has any structure or 1068/// alignment padding, false otherwise. 1069static bool HasPadding(const Type *Ty, const TargetData &TD) { 1070 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1071 const StructLayout *SL = TD.getStructLayout(STy); 1072 unsigned PrevFieldBitOffset = 0; 1073 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1074 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1075 1076 // Padding in sub-elements? 1077 if (HasPadding(STy->getElementType(i), TD)) 1078 return true; 1079 1080 // Check to see if there is any padding between this element and the 1081 // previous one. 1082 if (i) { 1083 unsigned PrevFieldEnd = 1084 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1085 if (PrevFieldEnd < FieldBitOffset) 1086 return true; 1087 } 1088 1089 PrevFieldBitOffset = FieldBitOffset; 1090 } 1091 1092 // Check for tail padding. 1093 if (unsigned EltCount = STy->getNumElements()) { 1094 unsigned PrevFieldEnd = PrevFieldBitOffset + 1095 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1096 if (PrevFieldEnd < SL->getSizeInBits()) 1097 return true; 1098 } 1099 1100 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1101 return HasPadding(ATy->getElementType(), TD); 1102 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1103 return HasPadding(VTy->getElementType(), TD); 1104 } 1105 return TD.getTypeSizeInBits(Ty) != TD.getTypePaddedSizeInBits(Ty); 1106} 1107 1108/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1109/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1110/// or 1 if safe after canonicalization has been performed. 1111/// 1112int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 1113 // Loop over the use list of the alloca. We can only transform it if all of 1114 // the users are safe to transform. 1115 AllocaInfo Info; 1116 1117 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 1118 I != E; ++I) { 1119 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 1120 if (Info.isUnsafe) { 1121 DOUT << "Cannot transform: " << *AI << " due to user: " << **I; 1122 return 0; 1123 } 1124 } 1125 1126 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1127 // source and destination, we have to be careful. In particular, the memcpy 1128 // could be moving around elements that live in structure padding of the LLVM 1129 // types, but may actually be used. In these cases, we refuse to promote the 1130 // struct. 1131 if (Info.isMemCpySrc && Info.isMemCpyDst && 1132 HasPadding(AI->getType()->getElementType(), *TD)) 1133 return 0; 1134 1135 // If we require cleanup, return 1, otherwise return 3. 1136 return Info.needsCleanup ? 1 : 3; 1137} 1138 1139/// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP 1140/// is canonicalized here. 1141void SROA::CleanupGEP(GetElementPtrInst *GEPI) { 1142 gep_type_iterator I = gep_type_begin(GEPI); 1143 ++I; 1144 1145 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 1146 uint64_t NumElements = AT->getNumElements(); 1147 1148 if (!isa<ConstantInt>(I.getOperand())) { 1149 if (NumElements == 1) { 1150 GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty)); 1151 } else { 1152 assert(NumElements == 2 && "Unhandled case!"); 1153 // All users of the GEP must be loads. At each use of the GEP, insert 1154 // two loads of the appropriate indexed GEP and select between them. 1155 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(), 1156 Constant::getNullValue(I.getOperand()->getType()), 1157 "isone", GEPI); 1158 // Insert the new GEP instructions, which are properly indexed. 1159 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 1160 Indices[1] = Constant::getNullValue(Type::Int32Ty); 1161 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1162 Indices.begin(), 1163 Indices.end(), 1164 GEPI->getName()+".0", GEPI); 1165 Indices[1] = ConstantInt::get(Type::Int32Ty, 1); 1166 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1167 Indices.begin(), 1168 Indices.end(), 1169 GEPI->getName()+".1", GEPI); 1170 // Replace all loads of the variable index GEP with loads from both 1171 // indexes and a select. 1172 while (!GEPI->use_empty()) { 1173 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 1174 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 1175 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 1176 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 1177 LI->replaceAllUsesWith(R); 1178 LI->eraseFromParent(); 1179 } 1180 GEPI->eraseFromParent(); 1181 } 1182 } 1183 } 1184} 1185 1186/// CleanupAllocaUsers - If SROA reported that it can promote the specified 1187/// allocation, but only if cleaned up, perform the cleanups required. 1188void SROA::CleanupAllocaUsers(AllocationInst *AI) { 1189 // At this point, we know that the end result will be SROA'd and promoted, so 1190 // we can insert ugly code if required so long as sroa+mem2reg will clean it 1191 // up. 1192 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1193 UI != E; ) { 1194 User *U = *UI++; 1195 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) 1196 CleanupGEP(GEPI); 1197 else if (Instruction *I = dyn_cast<Instruction>(U)) { 1198 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses; 1199 if (OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) { 1200 // Safe to remove debug info uses. 1201 while (!DbgInUses.empty()) { 1202 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back(); 1203 DI->eraseFromParent(); 1204 } 1205 I->eraseFromParent(); 1206 } 1207 } 1208 } 1209} 1210 1211/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 1212/// the offset specified by Offset (which is specified in bytes). 1213/// 1214/// There are two cases we handle here: 1215/// 1) A union of vector types of the same size and potentially its elements. 1216/// Here we turn element accesses into insert/extract element operations. 1217/// This promotes a <4 x float> with a store of float to the third element 1218/// into a <4 x float> that uses insert element. 1219/// 2) A fully general blob of memory, which we turn into some (potentially 1220/// large) integer type with extract and insert operations where the loads 1221/// and stores would mutate the memory. 1222static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy, 1223 unsigned AllocaSize, const TargetData &TD) { 1224 // If this could be contributing to a vector, analyze it. 1225 if (VecTy != Type::VoidTy) { // either null or a vector type. 1226 1227 // If the In type is a vector that is the same size as the alloca, see if it 1228 // matches the existing VecTy. 1229 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 1230 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 1231 // If we're storing/loading a vector of the right size, allow it as a 1232 // vector. If this the first vector we see, remember the type so that 1233 // we know the element size. 1234 if (VecTy == 0) 1235 VecTy = VInTy; 1236 return; 1237 } 1238 } else if (In == Type::FloatTy || In == Type::DoubleTy || 1239 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 && 1240 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 1241 // If we're accessing something that could be an element of a vector, see 1242 // if the implied vector agrees with what we already have and if Offset is 1243 // compatible with it. 1244 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 1245 if (Offset % EltSize == 0 && 1246 AllocaSize % EltSize == 0 && 1247 (VecTy == 0 || 1248 cast<VectorType>(VecTy)->getElementType() 1249 ->getPrimitiveSizeInBits()/8 == EltSize)) { 1250 if (VecTy == 0) 1251 VecTy = VectorType::get(In, AllocaSize/EltSize); 1252 return; 1253 } 1254 } 1255 } 1256 1257 // Otherwise, we have a case that we can't handle with an optimized vector 1258 // form. We can still turn this into a large integer. 1259 VecTy = Type::VoidTy; 1260} 1261 1262/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 1263/// its accesses to use a to single vector type, return true, and set VecTy to 1264/// the new type. If we could convert the alloca into a single promotable 1265/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 1266/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 1267/// is the current offset from the base of the alloca being analyzed. 1268/// 1269/// If we see at least one access to the value that is as a vector type, set the 1270/// SawVec flag. 1271/// 1272bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 1273 bool &SawVec, uint64_t Offset, 1274 unsigned AllocaSize) { 1275 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1276 Instruction *User = cast<Instruction>(*UI); 1277 1278 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1279 // Don't break volatile loads. 1280 if (LI->isVolatile()) 1281 return false; 1282 MergeInType(LI->getType(), Offset, VecTy, AllocaSize, *TD); 1283 SawVec |= isa<VectorType>(LI->getType()); 1284 continue; 1285 } 1286 1287 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1288 // Storing the pointer, not into the value? 1289 if (SI->getOperand(0) == V || SI->isVolatile()) return 0; 1290 MergeInType(SI->getOperand(0)->getType(), Offset, VecTy, AllocaSize, *TD); 1291 SawVec |= isa<VectorType>(SI->getOperand(0)->getType()); 1292 continue; 1293 } 1294 1295 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 1296 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset, 1297 AllocaSize)) 1298 return false; 1299 IsNotTrivial = true; 1300 continue; 1301 } 1302 1303 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1304 // If this is a GEP with a variable indices, we can't handle it. 1305 if (!GEP->hasAllConstantIndices()) 1306 return false; 1307 1308 // Compute the offset that this GEP adds to the pointer. 1309 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1310 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1311 &Indices[0], Indices.size()); 1312 // See if all uses can be converted. 1313 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset, 1314 AllocaSize)) 1315 return false; 1316 IsNotTrivial = true; 1317 continue; 1318 } 1319 1320 // If this is a constant sized memset of a constant value (e.g. 0) we can 1321 // handle it. 1322 if (isa<MemSetInst>(User) && 1323 // Store of constant value. 1324 isa<ConstantInt>(User->getOperand(2)) && 1325 // Store with constant size. 1326 isa<ConstantInt>(User->getOperand(3))) { 1327 VecTy = Type::VoidTy; 1328 IsNotTrivial = true; 1329 continue; 1330 } 1331 1332 // Otherwise, we cannot handle this! 1333 return false; 1334 } 1335 1336 return true; 1337} 1338 1339 1340/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1341/// directly. This happens when we are converting an "integer union" to a 1342/// single integer scalar, or when we are converting a "vector union" to a 1343/// vector with insert/extractelement instructions. 1344/// 1345/// Offset is an offset from the original alloca, in bits that need to be 1346/// shifted to the right. By the end of this, there should be no uses of Ptr. 1347void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 1348 while (!Ptr->use_empty()) { 1349 Instruction *User = cast<Instruction>(Ptr->use_back()); 1350 1351 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1352 ConvertUsesToScalar(CI, NewAI, Offset); 1353 CI->eraseFromParent(); 1354 continue; 1355 } 1356 1357 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1358 // Compute the offset that this GEP adds to the pointer. 1359 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1360 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1361 &Indices[0], Indices.size()); 1362 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 1363 GEP->eraseFromParent(); 1364 continue; 1365 } 1366 1367 IRBuilder<> Builder(User->getParent(), User); 1368 1369 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1370 // The load is a bit extract from NewAI shifted right by Offset bits. 1371 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 1372 Value *NewLoadVal 1373 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 1374 LI->replaceAllUsesWith(NewLoadVal); 1375 LI->eraseFromParent(); 1376 continue; 1377 } 1378 1379 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1380 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1381 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str()); 1382 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 1383 Builder); 1384 Builder.CreateStore(New, NewAI); 1385 SI->eraseFromParent(); 1386 continue; 1387 } 1388 1389 // If this is a constant sized memset of a constant value (e.g. 0) we can 1390 // transform it into a store of the expanded constant value. 1391 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1392 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 1393 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 1394 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 1395 1396 // Compute the value replicated the right number of times. 1397 APInt APVal(NumBytes*8, Val); 1398 1399 // Splat the value if non-zero. 1400 if (Val) 1401 for (unsigned i = 1; i != NumBytes; ++i) 1402 APVal |= APVal << 8; 1403 1404 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str()); 1405 Value *New = ConvertScalar_InsertValue(ConstantInt::get(APVal), Old, 1406 Offset, Builder); 1407 Builder.CreateStore(New, NewAI); 1408 MSI->eraseFromParent(); 1409 continue; 1410 } 1411 1412 1413 assert(0 && "Unsupported operation!"); 1414 abort(); 1415 } 1416} 1417 1418/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 1419/// or vector value FromVal, extracting the bits from the offset specified by 1420/// Offset. This returns the value, which is of type ToType. 1421/// 1422/// This happens when we are converting an "integer union" to a single 1423/// integer scalar, or when we are converting a "vector union" to a vector with 1424/// insert/extractelement instructions. 1425/// 1426/// Offset is an offset from the original alloca, in bits that need to be 1427/// shifted to the right. 1428Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 1429 uint64_t Offset, IRBuilder<> &Builder) { 1430 // If the load is of the whole new alloca, no conversion is needed. 1431 if (FromVal->getType() == ToType && Offset == 0) 1432 return FromVal; 1433 1434 // If the result alloca is a vector type, this is either an element 1435 // access or a bitcast to another vector type of the same size. 1436 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 1437 if (isa<VectorType>(ToType)) 1438 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 1439 1440 // Otherwise it must be an element access. 1441 unsigned Elt = 0; 1442 if (Offset) { 1443 unsigned EltSize = TD->getTypePaddedSizeInBits(VTy->getElementType()); 1444 Elt = Offset/EltSize; 1445 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 1446 } 1447 // Return the element extracted out of it. 1448 Value *V = Builder.CreateExtractElement(FromVal, 1449 ConstantInt::get(Type::Int32Ty,Elt), 1450 "tmp"); 1451 if (V->getType() != ToType) 1452 V = Builder.CreateBitCast(V, ToType, "tmp"); 1453 return V; 1454 } 1455 1456 // If ToType is a first class aggregate, extract out each of the pieces and 1457 // use insertvalue's to form the FCA. 1458 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 1459 const StructLayout &Layout = *TD->getStructLayout(ST); 1460 Value *Res = UndefValue::get(ST); 1461 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1462 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 1463 Offset+Layout.getElementOffsetInBits(i), 1464 Builder); 1465 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1466 } 1467 return Res; 1468 } 1469 1470 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 1471 uint64_t EltSize = TD->getTypePaddedSizeInBits(AT->getElementType()); 1472 Value *Res = UndefValue::get(AT); 1473 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1474 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 1475 Offset+i*EltSize, Builder); 1476 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1477 } 1478 return Res; 1479 } 1480 1481 // Otherwise, this must be a union that was converted to an integer value. 1482 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 1483 1484 // If this is a big-endian system and the load is narrower than the 1485 // full alloca type, we need to do a shift to get the right bits. 1486 int ShAmt = 0; 1487 if (TD->isBigEndian()) { 1488 // On big-endian machines, the lowest bit is stored at the bit offset 1489 // from the pointer given by getTypeStoreSizeInBits. This matters for 1490 // integers with a bitwidth that is not a multiple of 8. 1491 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1492 TD->getTypeStoreSizeInBits(ToType) - Offset; 1493 } else { 1494 ShAmt = Offset; 1495 } 1496 1497 // Note: we support negative bitwidths (with shl) which are not defined. 1498 // We do this to support (f.e.) loads off the end of a structure where 1499 // only some bits are used. 1500 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1501 FromVal = Builder.CreateLShr(FromVal, ConstantInt::get(FromVal->getType(), 1502 ShAmt), "tmp"); 1503 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1504 FromVal = Builder.CreateShl(FromVal, ConstantInt::get(FromVal->getType(), 1505 -ShAmt), "tmp"); 1506 1507 // Finally, unconditionally truncate the integer to the right width. 1508 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 1509 if (LIBitWidth < NTy->getBitWidth()) 1510 FromVal = Builder.CreateTrunc(FromVal, IntegerType::get(LIBitWidth), "tmp"); 1511 else if (LIBitWidth > NTy->getBitWidth()) 1512 FromVal = Builder.CreateZExt(FromVal, IntegerType::get(LIBitWidth), "tmp"); 1513 1514 // If the result is an integer, this is a trunc or bitcast. 1515 if (isa<IntegerType>(ToType)) { 1516 // Should be done. 1517 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) { 1518 // Just do a bitcast, we know the sizes match up. 1519 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 1520 } else { 1521 // Otherwise must be a pointer. 1522 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 1523 } 1524 assert(FromVal->getType() == ToType && "Didn't convert right?"); 1525 return FromVal; 1526} 1527 1528 1529/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 1530/// or vector value "Old" at the offset specified by Offset. 1531/// 1532/// This happens when we are converting an "integer union" to a 1533/// single integer scalar, or when we are converting a "vector union" to a 1534/// vector with insert/extractelement instructions. 1535/// 1536/// Offset is an offset from the original alloca, in bits that need to be 1537/// shifted to the right. 1538Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 1539 uint64_t Offset, IRBuilder<> &Builder) { 1540 1541 // Convert the stored type to the actual type, shift it left to insert 1542 // then 'or' into place. 1543 const Type *AllocaType = Old->getType(); 1544 1545 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 1546 // If the result alloca is a vector type, this is either an element 1547 // access or a bitcast to another vector type. 1548 if (isa<VectorType>(SV->getType())) { 1549 SV = Builder.CreateBitCast(SV, AllocaType, "tmp"); 1550 } else { 1551 // Must be an element insertion. 1552 unsigned Elt = Offset/TD->getTypePaddedSizeInBits(VTy->getElementType()); 1553 1554 if (SV->getType() != VTy->getElementType()) 1555 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 1556 1557 SV = Builder.CreateInsertElement(Old, SV, 1558 ConstantInt::get(Type::Int32Ty, Elt), 1559 "tmp"); 1560 } 1561 return SV; 1562 } 1563 1564 // If SV is a first-class aggregate value, insert each value recursively. 1565 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 1566 const StructLayout &Layout = *TD->getStructLayout(ST); 1567 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1568 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1569 Old = ConvertScalar_InsertValue(Elt, Old, 1570 Offset+Layout.getElementOffsetInBits(i), 1571 Builder); 1572 } 1573 return Old; 1574 } 1575 1576 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1577 uint64_t EltSize = TD->getTypePaddedSizeInBits(AT->getElementType()); 1578 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1579 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1580 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1581 } 1582 return Old; 1583 } 1584 1585 // If SV is a float, convert it to the appropriate integer type. 1586 // If it is a pointer, do the same. 1587 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1588 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1589 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1590 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1591 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType())) 1592 SV = Builder.CreateBitCast(SV, IntegerType::get(SrcWidth), "tmp"); 1593 else if (isa<PointerType>(SV->getType())) 1594 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(), "tmp"); 1595 1596 // Zero extend or truncate the value if needed. 1597 if (SV->getType() != AllocaType) { 1598 if (SV->getType()->getPrimitiveSizeInBits() < 1599 AllocaType->getPrimitiveSizeInBits()) 1600 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1601 else { 1602 // Truncation may be needed if storing more than the alloca can hold 1603 // (undefined behavior). 1604 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1605 SrcWidth = DestWidth; 1606 SrcStoreWidth = DestStoreWidth; 1607 } 1608 } 1609 1610 // If this is a big-endian system and the store is narrower than the 1611 // full alloca type, we need to do a shift to get the right bits. 1612 int ShAmt = 0; 1613 if (TD->isBigEndian()) { 1614 // On big-endian machines, the lowest bit is stored at the bit offset 1615 // from the pointer given by getTypeStoreSizeInBits. This matters for 1616 // integers with a bitwidth that is not a multiple of 8. 1617 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1618 } else { 1619 ShAmt = Offset; 1620 } 1621 1622 // Note: we support negative bitwidths (with shr) which are not defined. 1623 // We do this to support (f.e.) stores off the end of a structure where 1624 // only some bits in the structure are set. 1625 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1626 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1627 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), ShAmt), "tmp"); 1628 Mask <<= ShAmt; 1629 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1630 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), -ShAmt), "tmp"); 1631 Mask = Mask.lshr(-ShAmt); 1632 } 1633 1634 // Mask out the bits we are about to insert from the old value, and or 1635 // in the new bits. 1636 if (SrcWidth != DestWidth) { 1637 assert(DestWidth > SrcWidth); 1638 Old = Builder.CreateAnd(Old, ConstantInt::get(~Mask), "mask"); 1639 SV = Builder.CreateOr(Old, SV, "ins"); 1640 } 1641 return SV; 1642} 1643 1644 1645 1646/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1647/// some part of a constant global variable. This intentionally only accepts 1648/// constant expressions because we don't can't rewrite arbitrary instructions. 1649static bool PointsToConstantGlobal(Value *V) { 1650 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1651 return GV->isConstant(); 1652 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1653 if (CE->getOpcode() == Instruction::BitCast || 1654 CE->getOpcode() == Instruction::GetElementPtr) 1655 return PointsToConstantGlobal(CE->getOperand(0)); 1656 return false; 1657} 1658 1659/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1660/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1661/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1662/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1663/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1664/// the alloca, and if the source pointer is a pointer to a constant global, we 1665/// can optimize this. 1666static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1667 bool isOffset) { 1668 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1669 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) 1670 // Ignore non-volatile loads, they are always ok. 1671 if (!LI->isVolatile()) 1672 continue; 1673 1674 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1675 // If uses of the bitcast are ok, we are ok. 1676 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1677 return false; 1678 continue; 1679 } 1680 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1681 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1682 // doesn't, it does. 1683 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1684 isOffset || !GEP->hasAllZeroIndices())) 1685 return false; 1686 continue; 1687 } 1688 1689 // If this is isn't our memcpy/memmove, reject it as something we can't 1690 // handle. 1691 if (!isa<MemCpyInst>(*UI) && !isa<MemMoveInst>(*UI)) 1692 return false; 1693 1694 // If we already have seen a copy, reject the second one. 1695 if (TheCopy) return false; 1696 1697 // If the pointer has been offset from the start of the alloca, we can't 1698 // safely handle this. 1699 if (isOffset) return false; 1700 1701 // If the memintrinsic isn't using the alloca as the dest, reject it. 1702 if (UI.getOperandNo() != 1) return false; 1703 1704 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1705 1706 // If the source of the memcpy/move is not a constant global, reject it. 1707 if (!PointsToConstantGlobal(MI->getOperand(2))) 1708 return false; 1709 1710 // Otherwise, the transform is safe. Remember the copy instruction. 1711 TheCopy = MI; 1712 } 1713 return true; 1714} 1715 1716/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1717/// modified by a copy from a constant global. If we can prove this, we can 1718/// replace any uses of the alloca with uses of the global directly. 1719Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) { 1720 Instruction *TheCopy = 0; 1721 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1722 return TheCopy; 1723 return 0; 1724} 1725