ScalarReplAggregates.cpp revision e4af1cf4027d19cbe52d70858353dc1765ea3aef
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/LLVMContext.h" 31#include "llvm/Pass.h" 32#include "llvm/Analysis/Dominators.h" 33#include "llvm/Target/TargetData.h" 34#include "llvm/Transforms/Utils/PromoteMemToReg.h" 35#include "llvm/Transforms/Utils/Local.h" 36#include "llvm/Support/Debug.h" 37#include "llvm/Support/ErrorHandling.h" 38#include "llvm/Support/GetElementPtrTypeIterator.h" 39#include "llvm/Support/IRBuilder.h" 40#include "llvm/Support/MathExtras.h" 41#include "llvm/Support/Compiler.h" 42#include "llvm/ADT/SmallVector.h" 43#include "llvm/ADT/Statistic.h" 44using namespace llvm; 45 46STATISTIC(NumReplaced, "Number of allocas broken up"); 47STATISTIC(NumPromoted, "Number of allocas promoted"); 48STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 49STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 50 51namespace { 52 struct VISIBILITY_HIDDEN SROA : public FunctionPass { 53 static char ID; // Pass identification, replacement for typeid 54 explicit SROA(signed T = -1) : FunctionPass(&ID) { 55 if (T == -1) 56 SRThreshold = 128; 57 else 58 SRThreshold = T; 59 } 60 61 bool runOnFunction(Function &F); 62 63 bool performScalarRepl(Function &F); 64 bool performPromotion(Function &F); 65 66 // getAnalysisUsage - This pass does not require any passes, but we know it 67 // will not alter the CFG, so say so. 68 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 69 AU.addRequired<DominatorTree>(); 70 AU.addRequired<DominanceFrontier>(); 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 = getAnalysisIfAvailable<TargetData>(); 153 154 bool Changed = performPromotion(F); 155 156 // FIXME: ScalarRepl currently depends on TargetData more than it 157 // theoretically needs to. It should be refactored in order to support 158 // target-independent IR. Until this is done, just skip the actual 159 // scalar-replacement portion of this pass. 160 if (!TD) return Changed; 161 162 while (1) { 163 bool LocalChange = performScalarRepl(F); 164 if (!LocalChange) break; // No need to repromote if no scalarrepl 165 Changed = true; 166 LocalChange = performPromotion(F); 167 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 168 } 169 170 return Changed; 171} 172 173 174bool SROA::performPromotion(Function &F) { 175 std::vector<AllocaInst*> Allocas; 176 DominatorTree &DT = getAnalysis<DominatorTree>(); 177 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 178 179 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 180 181 bool Changed = false; 182 183 while (1) { 184 Allocas.clear(); 185 186 // Find allocas that are safe to promote, by looking at all instructions in 187 // the entry node 188 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 189 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 190 if (isAllocaPromotable(AI)) 191 Allocas.push_back(AI); 192 193 if (Allocas.empty()) break; 194 195 PromoteMemToReg(Allocas, DT, DF, F.getContext()); 196 NumPromoted += Allocas.size(); 197 Changed = true; 198 } 199 200 return Changed; 201} 202 203/// getNumSAElements - Return the number of elements in the specific struct or 204/// array. 205static uint64_t getNumSAElements(const Type *T) { 206 if (const StructType *ST = dyn_cast<StructType>(T)) 207 return ST->getNumElements(); 208 return cast<ArrayType>(T)->getNumElements(); 209} 210 211// performScalarRepl - This algorithm is a simple worklist driven algorithm, 212// which runs on all of the malloc/alloca instructions in the function, removing 213// them if they are only used by getelementptr instructions. 214// 215bool SROA::performScalarRepl(Function &F) { 216 std::vector<AllocationInst*> WorkList; 217 218 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 219 BasicBlock &BB = F.getEntryBlock(); 220 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 221 if (AllocationInst *A = dyn_cast<AllocationInst>(I)) 222 WorkList.push_back(A); 223 224 // Process the worklist 225 bool Changed = false; 226 while (!WorkList.empty()) { 227 AllocationInst *AI = WorkList.back(); 228 WorkList.pop_back(); 229 230 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 231 // with unused elements. 232 if (AI->use_empty()) { 233 AI->eraseFromParent(); 234 continue; 235 } 236 237 // If this alloca is impossible for us to promote, reject it early. 238 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 239 continue; 240 241 // Check to see if this allocation is only modified by a memcpy/memmove from 242 // a constant global. If this is the case, we can change all users to use 243 // the constant global instead. This is commonly produced by the CFE by 244 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 245 // is only subsequently read. 246 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 247 DOUT << "Found alloca equal to global: " << *AI; 248 DOUT << " memcpy = " << *TheCopy; 249 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 250 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 251 TheCopy->eraseFromParent(); // Don't mutate the global. 252 AI->eraseFromParent(); 253 ++NumGlobals; 254 Changed = true; 255 continue; 256 } 257 258 // Check to see if we can perform the core SROA transformation. We cannot 259 // transform the allocation instruction if it is an array allocation 260 // (allocations OF arrays are ok though), and an allocation of a scalar 261 // value cannot be decomposed at all. 262 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 263 264 // Do not promote [0 x %struct]. 265 if (AllocaSize == 0) continue; 266 267 // Do not promote any struct whose size is too big. 268 if (AllocaSize > SRThreshold) continue; 269 270 if ((isa<StructType>(AI->getAllocatedType()) || 271 isa<ArrayType>(AI->getAllocatedType())) && 272 // Do not promote any struct into more than "32" separate vars. 273 getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) { 274 // Check that all of the users of the allocation are capable of being 275 // transformed. 276 switch (isSafeAllocaToScalarRepl(AI)) { 277 default: llvm_unreachable("Unexpected value!"); 278 case 0: // Not safe to scalar replace. 279 break; 280 case 1: // Safe, but requires cleanup/canonicalizations first 281 CleanupAllocaUsers(AI); 282 // FALL THROUGH. 283 case 3: // Safe to scalar replace. 284 DoScalarReplacement(AI, WorkList); 285 Changed = true; 286 continue; 287 } 288 } 289 290 // If we can turn this aggregate value (potentially with casts) into a 291 // simple scalar value that can be mem2reg'd into a register value. 292 // IsNotTrivial tracks whether this is something that mem2reg could have 293 // promoted itself. If so, we don't want to transform it needlessly. Note 294 // that we can't just check based on the type: the alloca may be of an i32 295 // but that has pointer arithmetic to set byte 3 of it or something. 296 bool IsNotTrivial = false; 297 const Type *VectorTy = 0; 298 bool HadAVector = false; 299 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector, 300 0, unsigned(AllocaSize)) && IsNotTrivial) { 301 AllocaInst *NewAI; 302 // If we were able to find a vector type that can handle this with 303 // insert/extract elements, and if there was at least one use that had 304 // a vector type, promote this to a vector. We don't want to promote 305 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 306 // we just get a lot of insert/extracts. If at least one vector is 307 // involved, then we probably really do have a union of vector/array. 308 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) { 309 DOUT << "CONVERT TO VECTOR: " << *AI << " TYPE = " << *VectorTy <<"\n"; 310 311 // Create and insert the vector alloca. 312 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin()); 313 ConvertUsesToScalar(AI, NewAI, 0); 314 } else { 315 DOUT << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"; 316 317 // Create and insert the integer alloca. 318 const Type *NewTy = IntegerType::get(AI->getContext(), AllocaSize*8); 319 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 320 ConvertUsesToScalar(AI, NewAI, 0); 321 } 322 NewAI->takeName(AI); 323 AI->eraseFromParent(); 324 ++NumConverted; 325 Changed = true; 326 continue; 327 } 328 329 // Otherwise, couldn't process this alloca. 330 } 331 332 return Changed; 333} 334 335/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 336/// predicate, do SROA now. 337void SROA::DoScalarReplacement(AllocationInst *AI, 338 std::vector<AllocationInst*> &WorkList) { 339 DOUT << "Found inst to SROA: " << *AI; 340 SmallVector<AllocaInst*, 32> ElementAllocas; 341 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 342 ElementAllocas.reserve(ST->getNumContainedTypes()); 343 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 344 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 345 AI->getAlignment(), 346 AI->getName() + "." + Twine(i), AI); 347 ElementAllocas.push_back(NA); 348 WorkList.push_back(NA); // Add to worklist for recursive processing 349 } 350 } else { 351 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 352 ElementAllocas.reserve(AT->getNumElements()); 353 const Type *ElTy = AT->getElementType(); 354 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 355 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 356 AI->getName() + "." + Twine(i), AI); 357 ElementAllocas.push_back(NA); 358 WorkList.push_back(NA); // Add to worklist for recursive processing 359 } 360 } 361 362 // Now that we have created the alloca instructions that we want to use, 363 // expand the getelementptr instructions to use them. 364 // 365 while (!AI->use_empty()) { 366 Instruction *User = cast<Instruction>(AI->use_back()); 367 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) { 368 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas); 369 BCInst->eraseFromParent(); 370 continue; 371 } 372 373 // Replace: 374 // %res = load { i32, i32 }* %alloc 375 // with: 376 // %load.0 = load i32* %alloc.0 377 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 378 // %load.1 = load i32* %alloc.1 379 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 380 // (Also works for arrays instead of structs) 381 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 382 Value *Insert = UndefValue::get(LI->getType()); 383 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 384 Value *Load = new LoadInst(ElementAllocas[i], "load", LI); 385 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 386 } 387 LI->replaceAllUsesWith(Insert); 388 LI->eraseFromParent(); 389 continue; 390 } 391 392 // Replace: 393 // store { i32, i32 } %val, { i32, i32 }* %alloc 394 // with: 395 // %val.0 = extractvalue { i32, i32 } %val, 0 396 // store i32 %val.0, i32* %alloc.0 397 // %val.1 = extractvalue { i32, i32 } %val, 1 398 // store i32 %val.1, i32* %alloc.1 399 // (Also works for arrays instead of structs) 400 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 401 Value *Val = SI->getOperand(0); 402 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 403 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 404 new StoreInst(Extract, ElementAllocas[i], SI); 405 } 406 SI->eraseFromParent(); 407 continue; 408 } 409 410 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 411 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 412 unsigned Idx = 413 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 414 415 assert(Idx < ElementAllocas.size() && "Index out of range?"); 416 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 417 418 Value *RepValue; 419 if (GEPI->getNumOperands() == 3) { 420 // Do not insert a new getelementptr instruction with zero indices, only 421 // to have it optimized out later. 422 RepValue = AllocaToUse; 423 } else { 424 // We are indexing deeply into the structure, so we still need a 425 // getelement ptr instruction to finish the indexing. This may be 426 // expanded itself once the worklist is rerun. 427 // 428 SmallVector<Value*, 8> NewArgs; 429 NewArgs.push_back(Constant::getNullValue( 430 Type::getInt32Ty(AI->getContext()))); 431 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end()); 432 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(), 433 NewArgs.end(), "", GEPI); 434 RepValue->takeName(GEPI); 435 } 436 437 // If this GEP is to the start of the aggregate, check for memcpys. 438 if (Idx == 0 && GEPI->hasAllZeroIndices()) 439 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas); 440 441 // Move all of the users over to the new GEP. 442 GEPI->replaceAllUsesWith(RepValue); 443 // Delete the old GEP 444 GEPI->eraseFromParent(); 445 } 446 447 // Finally, delete the Alloca instruction 448 AI->eraseFromParent(); 449 NumReplaced++; 450} 451 452 453/// isSafeElementUse - Check to see if this use is an allowed use for a 454/// getelementptr instruction of an array aggregate allocation. isFirstElt 455/// indicates whether Ptr is known to the start of the aggregate. 456/// 457void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 458 AllocaInfo &Info) { 459 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 460 I != E; ++I) { 461 Instruction *User = cast<Instruction>(*I); 462 switch (User->getOpcode()) { 463 case Instruction::Load: break; 464 case Instruction::Store: 465 // Store is ok if storing INTO the pointer, not storing the pointer 466 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info); 467 break; 468 case Instruction::GetElementPtr: { 469 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 470 bool AreAllZeroIndices = isFirstElt; 471 if (GEP->getNumOperands() > 1) { 472 if (!isa<ConstantInt>(GEP->getOperand(1)) || 473 !cast<ConstantInt>(GEP->getOperand(1))->isZero()) 474 // Using pointer arithmetic to navigate the array. 475 return MarkUnsafe(Info); 476 477 if (AreAllZeroIndices) 478 AreAllZeroIndices = GEP->hasAllZeroIndices(); 479 } 480 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 481 if (Info.isUnsafe) return; 482 break; 483 } 484 case Instruction::BitCast: 485 if (isFirstElt) { 486 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 487 if (Info.isUnsafe) return; 488 break; 489 } 490 DOUT << " Transformation preventing inst: " << *User; 491 return MarkUnsafe(Info); 492 case Instruction::Call: 493 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 494 if (isFirstElt) { 495 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 496 if (Info.isUnsafe) return; 497 break; 498 } 499 } 500 DOUT << " Transformation preventing inst: " << *User; 501 return MarkUnsafe(Info); 502 default: 503 DOUT << " Transformation preventing inst: " << *User; 504 return MarkUnsafe(Info); 505 } 506 } 507 return; // All users look ok :) 508} 509 510/// AllUsersAreLoads - Return true if all users of this value are loads. 511static bool AllUsersAreLoads(Value *Ptr) { 512 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 513 I != E; ++I) 514 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 515 return false; 516 return true; 517} 518 519/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 520/// aggregate allocation. 521/// 522void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 523 AllocaInfo &Info) { 524 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 525 return isSafeUseOfBitCastedAllocation(C, AI, Info); 526 527 if (LoadInst *LI = dyn_cast<LoadInst>(User)) 528 if (!LI->isVolatile()) 529 return;// Loads (returning a first class aggregrate) are always rewritable 530 531 if (StoreInst *SI = dyn_cast<StoreInst>(User)) 532 if (!SI->isVolatile() && SI->getOperand(0) != AI) 533 return;// Store is ok if storing INTO the pointer, not storing the pointer 534 535 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 536 if (GEPI == 0) 537 return MarkUnsafe(Info); 538 539 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 540 541 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 542 if (I == E || 543 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) { 544 return MarkUnsafe(Info); 545 } 546 547 ++I; 548 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 549 550 bool IsAllZeroIndices = true; 551 552 // If the first index is a non-constant index into an array, see if we can 553 // handle it as a special case. 554 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 555 if (!isa<ConstantInt>(I.getOperand())) { 556 IsAllZeroIndices = 0; 557 uint64_t NumElements = AT->getNumElements(); 558 559 // If this is an array index and the index is not constant, we cannot 560 // promote... that is unless the array has exactly one or two elements in 561 // it, in which case we CAN promote it, but we have to canonicalize this 562 // out if this is the only problem. 563 if ((NumElements == 1 || NumElements == 2) && 564 AllUsersAreLoads(GEPI)) { 565 Info.needsCleanup = true; 566 return; // Canonicalization required! 567 } 568 return MarkUnsafe(Info); 569 } 570 } 571 572 // Walk through the GEP type indices, checking the types that this indexes 573 // into. 574 for (; I != E; ++I) { 575 // Ignore struct elements, no extra checking needed for these. 576 if (isa<StructType>(*I)) 577 continue; 578 579 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 580 if (!IdxVal) return MarkUnsafe(Info); 581 582 // Are all indices still zero? 583 IsAllZeroIndices &= IdxVal->isZero(); 584 585 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 586 // This GEP indexes an array. Verify that this is an in-range constant 587 // integer. Specifically, consider A[0][i]. We cannot know that the user 588 // isn't doing invalid things like allowing i to index an out-of-range 589 // subscript that accesses A[1]. Because of this, we have to reject SROA 590 // of any accesses into structs where any of the components are variables. 591 if (IdxVal->getZExtValue() >= AT->getNumElements()) 592 return MarkUnsafe(Info); 593 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) { 594 if (IdxVal->getZExtValue() >= VT->getNumElements()) 595 return MarkUnsafe(Info); 596 } 597 } 598 599 // If there are any non-simple uses of this getelementptr, make sure to reject 600 // them. 601 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 602} 603 604/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory 605/// intrinsic can be promoted by SROA. At this point, we know that the operand 606/// of the memintrinsic is a pointer to the beginning of the allocation. 607void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 608 unsigned OpNo, AllocaInfo &Info) { 609 // If not constant length, give up. 610 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 611 if (!Length) return MarkUnsafe(Info); 612 613 // If not the whole aggregate, give up. 614 if (Length->getZExtValue() != 615 TD->getTypeAllocSize(AI->getType()->getElementType())) 616 return MarkUnsafe(Info); 617 618 // We only know about memcpy/memset/memmove. 619 if (!isa<MemIntrinsic>(MI)) 620 return MarkUnsafe(Info); 621 622 // Otherwise, we can transform it. Determine whether this is a memcpy/set 623 // into or out of the aggregate. 624 if (OpNo == 1) 625 Info.isMemCpyDst = true; 626 else { 627 assert(OpNo == 2); 628 Info.isMemCpySrc = true; 629 } 630} 631 632/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast 633/// are 634void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI, 635 AllocaInfo &Info) { 636 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 637 UI != E; ++UI) { 638 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 639 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 640 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 641 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 642 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 643 if (SI->isVolatile()) 644 return MarkUnsafe(Info); 645 646 // If storing the entire alloca in one chunk through a bitcasted pointer 647 // to integer, we can transform it. This happens (for example) when you 648 // cast a {i32,i32}* to i64* and store through it. This is similar to the 649 // memcpy case and occurs in various "byval" cases and emulated memcpys. 650 if (isa<IntegerType>(SI->getOperand(0)->getType()) && 651 TD->getTypeAllocSize(SI->getOperand(0)->getType()) == 652 TD->getTypeAllocSize(AI->getType()->getElementType())) { 653 Info.isMemCpyDst = true; 654 continue; 655 } 656 return MarkUnsafe(Info); 657 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 658 if (LI->isVolatile()) 659 return MarkUnsafe(Info); 660 661 // If loading the entire alloca in one chunk through a bitcasted pointer 662 // to integer, we can transform it. This happens (for example) when you 663 // cast a {i32,i32}* to i64* and load through it. This is similar to the 664 // memcpy case and occurs in various "byval" cases and emulated memcpys. 665 if (isa<IntegerType>(LI->getType()) && 666 TD->getTypeAllocSize(LI->getType()) == 667 TD->getTypeAllocSize(AI->getType()->getElementType())) { 668 Info.isMemCpySrc = true; 669 continue; 670 } 671 return MarkUnsafe(Info); 672 } else if (isa<DbgInfoIntrinsic>(UI)) { 673 // If one user is DbgInfoIntrinsic then check if all users are 674 // DbgInfoIntrinsics. 675 if (OnlyUsedByDbgInfoIntrinsics(BC)) { 676 Info.needsCleanup = true; 677 return; 678 } 679 else 680 MarkUnsafe(Info); 681 } 682 else { 683 return MarkUnsafe(Info); 684 } 685 if (Info.isUnsafe) return; 686 } 687} 688 689/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 690/// to its first element. Transform users of the cast to use the new values 691/// instead. 692void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 693 SmallVector<AllocaInst*, 32> &NewElts) { 694 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 695 while (UI != UE) { 696 Instruction *User = cast<Instruction>(*UI++); 697 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) { 698 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 699 if (BCU->use_empty()) BCU->eraseFromParent(); 700 continue; 701 } 702 703 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 704 // This must be memcpy/memmove/memset of the entire aggregate. 705 // Split into one per element. 706 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts); 707 continue; 708 } 709 710 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 711 // If this is a store of the entire alloca from an integer, rewrite it. 712 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 713 continue; 714 } 715 716 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 717 // If this is a load of the entire alloca to an integer, rewrite it. 718 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 719 continue; 720 } 721 722 // Otherwise it must be some other user of a gep of the first pointer. Just 723 // leave these alone. 724 continue; 725 } 726} 727 728/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 729/// Rewrite it to copy or set the elements of the scalarized memory. 730void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 731 AllocationInst *AI, 732 SmallVector<AllocaInst*, 32> &NewElts) { 733 734 // If this is a memcpy/memmove, construct the other pointer as the 735 // appropriate type. The "Other" pointer is the pointer that goes to memory 736 // that doesn't have anything to do with the alloca that we are promoting. For 737 // memset, this Value* stays null. 738 Value *OtherPtr = 0; 739 LLVMContext &Context = MI->getContext(); 740 unsigned MemAlignment = MI->getAlignment(); 741 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 742 if (BCInst == MTI->getRawDest()) 743 OtherPtr = MTI->getRawSource(); 744 else { 745 assert(BCInst == MTI->getRawSource()); 746 OtherPtr = MTI->getRawDest(); 747 } 748 } 749 750 // If there is an other pointer, we want to convert it to the same pointer 751 // type as AI has, so we can GEP through it safely. 752 if (OtherPtr) { 753 // It is likely that OtherPtr is a bitcast, if so, remove it. 754 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 755 OtherPtr = BC->getOperand(0); 756 // All zero GEPs are effectively bitcasts. 757 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) 758 if (GEP->hasAllZeroIndices()) 759 OtherPtr = GEP->getOperand(0); 760 761 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 762 if (BCE->getOpcode() == Instruction::BitCast) 763 OtherPtr = BCE->getOperand(0); 764 765 // If the pointer is not the right type, insert a bitcast to the right 766 // type. 767 if (OtherPtr->getType() != AI->getType()) 768 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 769 MI); 770 } 771 772 // Process each element of the aggregate. 773 Value *TheFn = MI->getOperand(0); 774 const Type *BytePtrTy = MI->getRawDest()->getType(); 775 bool SROADest = MI->getRawDest() == BCInst; 776 777 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(MI->getContext())); 778 779 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 780 // If this is a memcpy/memmove, emit a GEP of the other element address. 781 Value *OtherElt = 0; 782 unsigned OtherEltAlign = MemAlignment; 783 784 if (OtherPtr) { 785 Value *Idx[2] = { Zero, 786 ConstantInt::get(Type::getInt32Ty(MI->getContext()), i) }; 787 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2, 788 OtherPtr->getNameStr()+"."+Twine(i), 789 MI); 790 uint64_t EltOffset; 791 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 792 if (const StructType *ST = 793 dyn_cast<StructType>(OtherPtrTy->getElementType())) { 794 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 795 } else { 796 const Type *EltTy = 797 cast<SequentialType>(OtherPtr->getType())->getElementType(); 798 EltOffset = TD->getTypeAllocSize(EltTy)*i; 799 } 800 801 // The alignment of the other pointer is the guaranteed alignment of the 802 // element, which is affected by both the known alignment of the whole 803 // mem intrinsic and the alignment of the element. If the alignment of 804 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 805 // known alignment is just 4 bytes. 806 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 807 } 808 809 Value *EltPtr = NewElts[i]; 810 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 811 812 // If we got down to a scalar, insert a load or store as appropriate. 813 if (EltTy->isSingleValueType()) { 814 if (isa<MemTransferInst>(MI)) { 815 if (SROADest) { 816 // From Other to Alloca. 817 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 818 new StoreInst(Elt, EltPtr, MI); 819 } else { 820 // From Alloca to Other. 821 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 822 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 823 } 824 continue; 825 } 826 assert(isa<MemSetInst>(MI)); 827 828 // If the stored element is zero (common case), just store a null 829 // constant. 830 Constant *StoreVal; 831 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 832 if (CI->isZero()) { 833 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 834 } else { 835 // If EltTy is a vector type, get the element type. 836 const Type *ValTy = EltTy->getScalarType(); 837 838 // Construct an integer with the right value. 839 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 840 APInt OneVal(EltSize, CI->getZExtValue()); 841 APInt TotalVal(OneVal); 842 // Set each byte. 843 for (unsigned i = 0; 8*i < EltSize; ++i) { 844 TotalVal = TotalVal.shl(8); 845 TotalVal |= OneVal; 846 } 847 848 // Convert the integer value to the appropriate type. 849 StoreVal = ConstantInt::get(Context, TotalVal); 850 if (isa<PointerType>(ValTy)) 851 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 852 else if (ValTy->isFloatingPoint()) 853 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 854 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 855 856 // If the requested value was a vector constant, create it. 857 if (EltTy != ValTy) { 858 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 859 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 860 StoreVal = ConstantVector::get(&Elts[0], NumElts); 861 } 862 } 863 new StoreInst(StoreVal, EltPtr, MI); 864 continue; 865 } 866 // Otherwise, if we're storing a byte variable, use a memset call for 867 // this element. 868 } 869 870 // Cast the element pointer to BytePtrTy. 871 if (EltPtr->getType() != BytePtrTy) 872 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 873 874 // Cast the other pointer (if we have one) to BytePtrTy. 875 if (OtherElt && OtherElt->getType() != BytePtrTy) 876 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 877 MI); 878 879 unsigned EltSize = TD->getTypeAllocSize(EltTy); 880 881 // Finally, insert the meminst for this element. 882 if (isa<MemTransferInst>(MI)) { 883 Value *Ops[] = { 884 SROADest ? EltPtr : OtherElt, // Dest ptr 885 SROADest ? OtherElt : EltPtr, // Src ptr 886 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 887 // Align 888 ConstantInt::get(Type::getInt32Ty(MI->getContext()), OtherEltAlign) 889 }; 890 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 891 } else { 892 assert(isa<MemSetInst>(MI)); 893 Value *Ops[] = { 894 EltPtr, MI->getOperand(2), // Dest, Value, 895 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 896 Zero // Align 897 }; 898 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 899 } 900 } 901 MI->eraseFromParent(); 902} 903 904/// RewriteStoreUserOfWholeAlloca - We found an store of an integer that 905/// overwrites the entire allocation. Extract out the pieces of the stored 906/// integer and store them individually. 907void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, 908 AllocationInst *AI, 909 SmallVector<AllocaInst*, 32> &NewElts){ 910 // Extract each element out of the integer according to its structure offset 911 // and store the element value to the individual alloca. 912 Value *SrcVal = SI->getOperand(0); 913 const Type *AllocaEltTy = AI->getType()->getElementType(); 914 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 915 916 // If this isn't a store of an integer to the whole alloca, it may be a store 917 // to the first element. Just ignore the store in this case and normal SROA 918 // will handle it. 919 if (!isa<IntegerType>(SrcVal->getType()) || 920 TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits) 921 return; 922 // Handle tail padding by extending the operand 923 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 924 SrcVal = new ZExtInst(SrcVal, 925 IntegerType::get(SI->getContext(), AllocaSizeBits), 926 "", SI); 927 928 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI; 929 930 // There are two forms here: AI could be an array or struct. Both cases 931 // have different ways to compute the element offset. 932 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 933 const StructLayout *Layout = TD->getStructLayout(EltSTy); 934 935 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 936 // Get the number of bits to shift SrcVal to get the value. 937 const Type *FieldTy = EltSTy->getElementType(i); 938 uint64_t Shift = Layout->getElementOffsetInBits(i); 939 940 if (TD->isBigEndian()) 941 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 942 943 Value *EltVal = SrcVal; 944 if (Shift) { 945 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 946 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 947 "sroa.store.elt", SI); 948 } 949 950 // Truncate down to an integer of the right size. 951 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 952 953 // Ignore zero sized fields like {}, they obviously contain no data. 954 if (FieldSizeBits == 0) continue; 955 956 if (FieldSizeBits != AllocaSizeBits) 957 EltVal = new TruncInst(EltVal, 958 IntegerType::get(SI->getContext(), FieldSizeBits), 959 "", SI); 960 Value *DestField = NewElts[i]; 961 if (EltVal->getType() == FieldTy) { 962 // Storing to an integer field of this size, just do it. 963 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { 964 // Bitcast to the right element type (for fp/vector values). 965 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 966 } else { 967 // Otherwise, bitcast the dest pointer (for aggregates). 968 DestField = new BitCastInst(DestField, 969 PointerType::getUnqual(EltVal->getType()), 970 "", SI); 971 } 972 new StoreInst(EltVal, DestField, SI); 973 } 974 975 } else { 976 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 977 const Type *ArrayEltTy = ATy->getElementType(); 978 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 979 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 980 981 uint64_t Shift; 982 983 if (TD->isBigEndian()) 984 Shift = AllocaSizeBits-ElementOffset; 985 else 986 Shift = 0; 987 988 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 989 // Ignore zero sized fields like {}, they obviously contain no data. 990 if (ElementSizeBits == 0) continue; 991 992 Value *EltVal = SrcVal; 993 if (Shift) { 994 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 995 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 996 "sroa.store.elt", SI); 997 } 998 999 // Truncate down to an integer of the right size. 1000 if (ElementSizeBits != AllocaSizeBits) 1001 EltVal = new TruncInst(EltVal, 1002 IntegerType::get(SI->getContext(), 1003 ElementSizeBits),"",SI); 1004 Value *DestField = NewElts[i]; 1005 if (EltVal->getType() == ArrayEltTy) { 1006 // Storing to an integer field of this size, just do it. 1007 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { 1008 // Bitcast to the right element type (for fp/vector values). 1009 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 1010 } else { 1011 // Otherwise, bitcast the dest pointer (for aggregates). 1012 DestField = new BitCastInst(DestField, 1013 PointerType::getUnqual(EltVal->getType()), 1014 "", SI); 1015 } 1016 new StoreInst(EltVal, DestField, SI); 1017 1018 if (TD->isBigEndian()) 1019 Shift -= ElementOffset; 1020 else 1021 Shift += ElementOffset; 1022 } 1023 } 1024 1025 SI->eraseFromParent(); 1026} 1027 1028/// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to 1029/// an integer. Load the individual pieces to form the aggregate value. 1030void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 1031 SmallVector<AllocaInst*, 32> &NewElts) { 1032 // Extract each element out of the NewElts according to its structure offset 1033 // and form the result value. 1034 const Type *AllocaEltTy = AI->getType()->getElementType(); 1035 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1036 1037 // If this isn't a load of the whole alloca to an integer, it may be a load 1038 // of the first element. Just ignore the load in this case and normal SROA 1039 // will handle it. 1040 if (!isa<IntegerType>(LI->getType()) || 1041 TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits) 1042 return; 1043 1044 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI; 1045 1046 // There are two forms here: AI could be an array or struct. Both cases 1047 // have different ways to compute the element offset. 1048 const StructLayout *Layout = 0; 1049 uint64_t ArrayEltBitOffset = 0; 1050 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1051 Layout = TD->getStructLayout(EltSTy); 1052 } else { 1053 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1054 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1055 } 1056 1057 Value *ResultVal = 1058 Constant::getNullValue(IntegerType::get(LI->getContext(), AllocaSizeBits)); 1059 1060 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1061 // Load the value from the alloca. If the NewElt is an aggregate, cast 1062 // the pointer to an integer of the same size before doing the load. 1063 Value *SrcField = NewElts[i]; 1064 const Type *FieldTy = 1065 cast<PointerType>(SrcField->getType())->getElementType(); 1066 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1067 1068 // Ignore zero sized fields like {}, they obviously contain no data. 1069 if (FieldSizeBits == 0) continue; 1070 1071 const IntegerType *FieldIntTy = IntegerType::get(LI->getContext(), 1072 FieldSizeBits); 1073 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && 1074 !isa<VectorType>(FieldTy)) 1075 SrcField = new BitCastInst(SrcField, 1076 PointerType::getUnqual(FieldIntTy), 1077 "", LI); 1078 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1079 1080 // If SrcField is a fp or vector of the right size but that isn't an 1081 // integer type, bitcast to an integer so we can shift it. 1082 if (SrcField->getType() != FieldIntTy) 1083 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1084 1085 // Zero extend the field to be the same size as the final alloca so that 1086 // we can shift and insert it. 1087 if (SrcField->getType() != ResultVal->getType()) 1088 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1089 1090 // Determine the number of bits to shift SrcField. 1091 uint64_t Shift; 1092 if (Layout) // Struct case. 1093 Shift = Layout->getElementOffsetInBits(i); 1094 else // Array case. 1095 Shift = i*ArrayEltBitOffset; 1096 1097 if (TD->isBigEndian()) 1098 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1099 1100 if (Shift) { 1101 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1102 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1103 } 1104 1105 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1106 } 1107 1108 // Handle tail padding by truncating the result 1109 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1110 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1111 1112 LI->replaceAllUsesWith(ResultVal); 1113 LI->eraseFromParent(); 1114} 1115 1116 1117/// HasPadding - Return true if the specified type has any structure or 1118/// alignment padding, false otherwise. 1119static bool HasPadding(const Type *Ty, const TargetData &TD) { 1120 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1121 const StructLayout *SL = TD.getStructLayout(STy); 1122 unsigned PrevFieldBitOffset = 0; 1123 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1124 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1125 1126 // Padding in sub-elements? 1127 if (HasPadding(STy->getElementType(i), TD)) 1128 return true; 1129 1130 // Check to see if there is any padding between this element and the 1131 // previous one. 1132 if (i) { 1133 unsigned PrevFieldEnd = 1134 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1135 if (PrevFieldEnd < FieldBitOffset) 1136 return true; 1137 } 1138 1139 PrevFieldBitOffset = FieldBitOffset; 1140 } 1141 1142 // Check for tail padding. 1143 if (unsigned EltCount = STy->getNumElements()) { 1144 unsigned PrevFieldEnd = PrevFieldBitOffset + 1145 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1146 if (PrevFieldEnd < SL->getSizeInBits()) 1147 return true; 1148 } 1149 1150 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1151 return HasPadding(ATy->getElementType(), TD); 1152 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1153 return HasPadding(VTy->getElementType(), TD); 1154 } 1155 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1156} 1157 1158/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1159/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1160/// or 1 if safe after canonicalization has been performed. 1161/// 1162int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 1163 // Loop over the use list of the alloca. We can only transform it if all of 1164 // the users are safe to transform. 1165 AllocaInfo Info; 1166 1167 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 1168 I != E; ++I) { 1169 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 1170 if (Info.isUnsafe) { 1171 DOUT << "Cannot transform: " << *AI << " due to user: " << **I; 1172 return 0; 1173 } 1174 } 1175 1176 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1177 // source and destination, we have to be careful. In particular, the memcpy 1178 // could be moving around elements that live in structure padding of the LLVM 1179 // types, but may actually be used. In these cases, we refuse to promote the 1180 // struct. 1181 if (Info.isMemCpySrc && Info.isMemCpyDst && 1182 HasPadding(AI->getType()->getElementType(), *TD)) 1183 return 0; 1184 1185 // If we require cleanup, return 1, otherwise return 3. 1186 return Info.needsCleanup ? 1 : 3; 1187} 1188 1189/// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP 1190/// is canonicalized here. 1191void SROA::CleanupGEP(GetElementPtrInst *GEPI) { 1192 gep_type_iterator I = gep_type_begin(GEPI); 1193 ++I; 1194 1195 const ArrayType *AT = dyn_cast<ArrayType>(*I); 1196 if (!AT) 1197 return; 1198 1199 uint64_t NumElements = AT->getNumElements(); 1200 1201 if (isa<ConstantInt>(I.getOperand())) 1202 return; 1203 1204 if (NumElements == 1) { 1205 GEPI->setOperand(2, 1206 Constant::getNullValue(Type::getInt32Ty(GEPI->getContext()))); 1207 return; 1208 } 1209 1210 assert(NumElements == 2 && "Unhandled case!"); 1211 // All users of the GEP must be loads. At each use of the GEP, insert 1212 // two loads of the appropriate indexed GEP and select between them. 1213 Value *IsOne = new ICmpInst(GEPI, ICmpInst::ICMP_NE, I.getOperand(), 1214 Constant::getNullValue(I.getOperand()->getType()), 1215 "isone"); 1216 // Insert the new GEP instructions, which are properly indexed. 1217 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 1218 Indices[1] = Constant::getNullValue(Type::getInt32Ty(GEPI->getContext())); 1219 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1220 Indices.begin(), 1221 Indices.end(), 1222 GEPI->getName()+".0", GEPI); 1223 Indices[1] = ConstantInt::get(Type::getInt32Ty(GEPI->getContext()), 1); 1224 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1225 Indices.begin(), 1226 Indices.end(), 1227 GEPI->getName()+".1", GEPI); 1228 // Replace all loads of the variable index GEP with loads from both 1229 // indexes and a select. 1230 while (!GEPI->use_empty()) { 1231 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 1232 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 1233 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 1234 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 1235 LI->replaceAllUsesWith(R); 1236 LI->eraseFromParent(); 1237 } 1238 GEPI->eraseFromParent(); 1239} 1240 1241 1242/// CleanupAllocaUsers - If SROA reported that it can promote the specified 1243/// allocation, but only if cleaned up, perform the cleanups required. 1244void SROA::CleanupAllocaUsers(AllocationInst *AI) { 1245 // At this point, we know that the end result will be SROA'd and promoted, so 1246 // we can insert ugly code if required so long as sroa+mem2reg will clean it 1247 // up. 1248 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1249 UI != E; ) { 1250 User *U = *UI++; 1251 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) 1252 CleanupGEP(GEPI); 1253 else { 1254 Instruction *I = cast<Instruction>(U); 1255 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses; 1256 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) { 1257 // Safe to remove debug info uses. 1258 while (!DbgInUses.empty()) { 1259 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back(); 1260 DI->eraseFromParent(); 1261 } 1262 I->eraseFromParent(); 1263 } 1264 } 1265 } 1266} 1267 1268/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 1269/// the offset specified by Offset (which is specified in bytes). 1270/// 1271/// There are two cases we handle here: 1272/// 1) A union of vector types of the same size and potentially its elements. 1273/// Here we turn element accesses into insert/extract element operations. 1274/// This promotes a <4 x float> with a store of float to the third element 1275/// into a <4 x float> that uses insert element. 1276/// 2) A fully general blob of memory, which we turn into some (potentially 1277/// large) integer type with extract and insert operations where the loads 1278/// and stores would mutate the memory. 1279static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy, 1280 unsigned AllocaSize, const TargetData &TD, 1281 LLVMContext &Context) { 1282 // If this could be contributing to a vector, analyze it. 1283 if (VecTy != Type::getVoidTy(Context)) { // either null or a vector type. 1284 1285 // If the In type is a vector that is the same size as the alloca, see if it 1286 // matches the existing VecTy. 1287 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 1288 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 1289 // If we're storing/loading a vector of the right size, allow it as a 1290 // vector. If this the first vector we see, remember the type so that 1291 // we know the element size. 1292 if (VecTy == 0) 1293 VecTy = VInTy; 1294 return; 1295 } 1296 } else if (In == Type::getFloatTy(Context) || 1297 In == Type::getDoubleTy(Context) || 1298 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 && 1299 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 1300 // If we're accessing something that could be an element of a vector, see 1301 // if the implied vector agrees with what we already have and if Offset is 1302 // compatible with it. 1303 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 1304 if (Offset % EltSize == 0 && 1305 AllocaSize % EltSize == 0 && 1306 (VecTy == 0 || 1307 cast<VectorType>(VecTy)->getElementType() 1308 ->getPrimitiveSizeInBits()/8 == EltSize)) { 1309 if (VecTy == 0) 1310 VecTy = VectorType::get(In, AllocaSize/EltSize); 1311 return; 1312 } 1313 } 1314 } 1315 1316 // Otherwise, we have a case that we can't handle with an optimized vector 1317 // form. We can still turn this into a large integer. 1318 VecTy = Type::getVoidTy(Context); 1319} 1320 1321/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 1322/// its accesses to use a to single vector type, return true, and set VecTy to 1323/// the new type. If we could convert the alloca into a single promotable 1324/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 1325/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 1326/// is the current offset from the base of the alloca being analyzed. 1327/// 1328/// If we see at least one access to the value that is as a vector type, set the 1329/// SawVec flag. 1330/// 1331bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 1332 bool &SawVec, uint64_t Offset, 1333 unsigned AllocaSize) { 1334 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1335 Instruction *User = cast<Instruction>(*UI); 1336 1337 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1338 // Don't break volatile loads. 1339 if (LI->isVolatile()) 1340 return false; 1341 MergeInType(LI->getType(), Offset, VecTy, 1342 AllocaSize, *TD, V->getContext()); 1343 SawVec |= isa<VectorType>(LI->getType()); 1344 continue; 1345 } 1346 1347 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1348 // Storing the pointer, not into the value? 1349 if (SI->getOperand(0) == V || SI->isVolatile()) return 0; 1350 MergeInType(SI->getOperand(0)->getType(), Offset, 1351 VecTy, AllocaSize, *TD, V->getContext()); 1352 SawVec |= isa<VectorType>(SI->getOperand(0)->getType()); 1353 continue; 1354 } 1355 1356 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 1357 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset, 1358 AllocaSize)) 1359 return false; 1360 IsNotTrivial = true; 1361 continue; 1362 } 1363 1364 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1365 // If this is a GEP with a variable indices, we can't handle it. 1366 if (!GEP->hasAllConstantIndices()) 1367 return false; 1368 1369 // Compute the offset that this GEP adds to the pointer. 1370 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1371 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1372 &Indices[0], Indices.size()); 1373 // See if all uses can be converted. 1374 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset, 1375 AllocaSize)) 1376 return false; 1377 IsNotTrivial = true; 1378 continue; 1379 } 1380 1381 // If this is a constant sized memset of a constant value (e.g. 0) we can 1382 // handle it. 1383 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1384 // Store of constant value and constant size. 1385 if (isa<ConstantInt>(MSI->getValue()) && 1386 isa<ConstantInt>(MSI->getLength())) { 1387 IsNotTrivial = true; 1388 continue; 1389 } 1390 } 1391 1392 // If this is a memcpy or memmove into or out of the whole allocation, we 1393 // can handle it like a load or store of the scalar type. 1394 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1395 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 1396 if (Len->getZExtValue() == AllocaSize && Offset == 0) { 1397 IsNotTrivial = true; 1398 continue; 1399 } 1400 } 1401 1402 // Ignore dbg intrinsic. 1403 if (isa<DbgInfoIntrinsic>(User)) 1404 continue; 1405 1406 // Otherwise, we cannot handle this! 1407 return false; 1408 } 1409 1410 return true; 1411} 1412 1413 1414/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1415/// directly. This happens when we are converting an "integer union" to a 1416/// single integer scalar, or when we are converting a "vector union" to a 1417/// vector with insert/extractelement instructions. 1418/// 1419/// Offset is an offset from the original alloca, in bits that need to be 1420/// shifted to the right. By the end of this, there should be no uses of Ptr. 1421void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 1422 while (!Ptr->use_empty()) { 1423 Instruction *User = cast<Instruction>(Ptr->use_back()); 1424 1425 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1426 ConvertUsesToScalar(CI, NewAI, Offset); 1427 CI->eraseFromParent(); 1428 continue; 1429 } 1430 1431 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1432 // Compute the offset that this GEP adds to the pointer. 1433 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1434 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1435 &Indices[0], Indices.size()); 1436 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 1437 GEP->eraseFromParent(); 1438 continue; 1439 } 1440 1441 IRBuilder<> Builder(User->getParent(), User); 1442 1443 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1444 // The load is a bit extract from NewAI shifted right by Offset bits. 1445 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 1446 Value *NewLoadVal 1447 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 1448 LI->replaceAllUsesWith(NewLoadVal); 1449 LI->eraseFromParent(); 1450 continue; 1451 } 1452 1453 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1454 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1455 // FIXME: Remove once builder has Twine API. 1456 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").str().c_str()); 1457 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 1458 Builder); 1459 Builder.CreateStore(New, NewAI); 1460 SI->eraseFromParent(); 1461 continue; 1462 } 1463 1464 // If this is a constant sized memset of a constant value (e.g. 0) we can 1465 // transform it into a store of the expanded constant value. 1466 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1467 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 1468 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 1469 if (NumBytes != 0) { 1470 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 1471 1472 // Compute the value replicated the right number of times. 1473 APInt APVal(NumBytes*8, Val); 1474 1475 // Splat the value if non-zero. 1476 if (Val) 1477 for (unsigned i = 1; i != NumBytes; ++i) 1478 APVal |= APVal << 8; 1479 1480 // FIXME: Remove once builder has Twine API. 1481 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").str().c_str()); 1482 Value *New = ConvertScalar_InsertValue( 1483 ConstantInt::get(User->getContext(), APVal), 1484 Old, Offset, Builder); 1485 Builder.CreateStore(New, NewAI); 1486 } 1487 MSI->eraseFromParent(); 1488 continue; 1489 } 1490 1491 // If this is a memcpy or memmove into or out of the whole allocation, we 1492 // can handle it like a load or store of the scalar type. 1493 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1494 assert(Offset == 0 && "must be store to start of alloca"); 1495 1496 // If the source and destination are both to the same alloca, then this is 1497 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 1498 // as appropriate. 1499 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject()); 1500 1501 if (MTI->getSource()->getUnderlyingObject() != OrigAI) { 1502 // Dest must be OrigAI, change this to be a load from the original 1503 // pointer (bitcasted), then a store to our new alloca. 1504 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 1505 Value *SrcPtr = MTI->getSource(); 1506 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 1507 1508 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 1509 SrcVal->setAlignment(MTI->getAlignment()); 1510 Builder.CreateStore(SrcVal, NewAI); 1511 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) { 1512 // Src must be OrigAI, change this to be a load from NewAI then a store 1513 // through the original dest pointer (bitcasted). 1514 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 1515 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 1516 1517 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 1518 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 1519 NewStore->setAlignment(MTI->getAlignment()); 1520 } else { 1521 // Noop transfer. Src == Dst 1522 } 1523 1524 1525 MTI->eraseFromParent(); 1526 continue; 1527 } 1528 1529 // If user is a dbg info intrinsic then it is safe to remove it. 1530 if (isa<DbgInfoIntrinsic>(User)) { 1531 User->eraseFromParent(); 1532 continue; 1533 } 1534 1535 llvm_unreachable("Unsupported operation!"); 1536 } 1537} 1538 1539/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 1540/// or vector value FromVal, extracting the bits from the offset specified by 1541/// Offset. This returns the value, which is of type ToType. 1542/// 1543/// This happens when we are converting an "integer union" to a single 1544/// integer scalar, or when we are converting a "vector union" to a vector with 1545/// insert/extractelement instructions. 1546/// 1547/// Offset is an offset from the original alloca, in bits that need to be 1548/// shifted to the right. 1549Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 1550 uint64_t Offset, IRBuilder<> &Builder) { 1551 // If the load is of the whole new alloca, no conversion is needed. 1552 if (FromVal->getType() == ToType && Offset == 0) 1553 return FromVal; 1554 1555 // If the result alloca is a vector type, this is either an element 1556 // access or a bitcast to another vector type of the same size. 1557 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 1558 if (isa<VectorType>(ToType)) 1559 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 1560 1561 // Otherwise it must be an element access. 1562 unsigned Elt = 0; 1563 if (Offset) { 1564 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1565 Elt = Offset/EltSize; 1566 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 1567 } 1568 // Return the element extracted out of it. 1569 Value *V = Builder.CreateExtractElement(FromVal, ConstantInt::get( 1570 Type::getInt32Ty(FromVal->getContext()), Elt), "tmp"); 1571 if (V->getType() != ToType) 1572 V = Builder.CreateBitCast(V, ToType, "tmp"); 1573 return V; 1574 } 1575 1576 // If ToType is a first class aggregate, extract out each of the pieces and 1577 // use insertvalue's to form the FCA. 1578 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 1579 const StructLayout &Layout = *TD->getStructLayout(ST); 1580 Value *Res = UndefValue::get(ST); 1581 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1582 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 1583 Offset+Layout.getElementOffsetInBits(i), 1584 Builder); 1585 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1586 } 1587 return Res; 1588 } 1589 1590 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 1591 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1592 Value *Res = UndefValue::get(AT); 1593 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1594 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 1595 Offset+i*EltSize, Builder); 1596 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1597 } 1598 return Res; 1599 } 1600 1601 // Otherwise, this must be a union that was converted to an integer value. 1602 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 1603 1604 // If this is a big-endian system and the load is narrower than the 1605 // full alloca type, we need to do a shift to get the right bits. 1606 int ShAmt = 0; 1607 if (TD->isBigEndian()) { 1608 // On big-endian machines, the lowest bit is stored at the bit offset 1609 // from the pointer given by getTypeStoreSizeInBits. This matters for 1610 // integers with a bitwidth that is not a multiple of 8. 1611 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1612 TD->getTypeStoreSizeInBits(ToType) - Offset; 1613 } else { 1614 ShAmt = Offset; 1615 } 1616 1617 // Note: we support negative bitwidths (with shl) which are not defined. 1618 // We do this to support (f.e.) loads off the end of a structure where 1619 // only some bits are used. 1620 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1621 FromVal = Builder.CreateLShr(FromVal, 1622 ConstantInt::get(FromVal->getType(), 1623 ShAmt), "tmp"); 1624 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1625 FromVal = Builder.CreateShl(FromVal, 1626 ConstantInt::get(FromVal->getType(), 1627 -ShAmt), "tmp"); 1628 1629 // Finally, unconditionally truncate the integer to the right width. 1630 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 1631 if (LIBitWidth < NTy->getBitWidth()) 1632 FromVal = 1633 Builder.CreateTrunc(FromVal, IntegerType::get(FromVal->getContext(), 1634 LIBitWidth), "tmp"); 1635 else if (LIBitWidth > NTy->getBitWidth()) 1636 FromVal = 1637 Builder.CreateZExt(FromVal, IntegerType::get(FromVal->getContext(), 1638 LIBitWidth), "tmp"); 1639 1640 // If the result is an integer, this is a trunc or bitcast. 1641 if (isa<IntegerType>(ToType)) { 1642 // Should be done. 1643 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) { 1644 // Just do a bitcast, we know the sizes match up. 1645 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 1646 } else { 1647 // Otherwise must be a pointer. 1648 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 1649 } 1650 assert(FromVal->getType() == ToType && "Didn't convert right?"); 1651 return FromVal; 1652} 1653 1654 1655/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 1656/// or vector value "Old" at the offset specified by Offset. 1657/// 1658/// This happens when we are converting an "integer union" to a 1659/// single integer scalar, or when we are converting a "vector union" to a 1660/// vector with insert/extractelement instructions. 1661/// 1662/// Offset is an offset from the original alloca, in bits that need to be 1663/// shifted to the right. 1664Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 1665 uint64_t Offset, IRBuilder<> &Builder) { 1666 1667 // Convert the stored type to the actual type, shift it left to insert 1668 // then 'or' into place. 1669 const Type *AllocaType = Old->getType(); 1670 LLVMContext &Context = Old->getContext(); 1671 1672 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 1673 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy); 1674 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType()); 1675 1676 // Changing the whole vector with memset or with an access of a different 1677 // vector type? 1678 if (ValSize == VecSize) 1679 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 1680 1681 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1682 1683 // Must be an element insertion. 1684 unsigned Elt = Offset/EltSize; 1685 1686 if (SV->getType() != VTy->getElementType()) 1687 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 1688 1689 SV = Builder.CreateInsertElement(Old, SV, 1690 ConstantInt::get(Type::getInt32Ty(SV->getContext()), Elt), 1691 "tmp"); 1692 return SV; 1693 } 1694 1695 // If SV is a first-class aggregate value, insert each value recursively. 1696 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 1697 const StructLayout &Layout = *TD->getStructLayout(ST); 1698 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1699 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1700 Old = ConvertScalar_InsertValue(Elt, Old, 1701 Offset+Layout.getElementOffsetInBits(i), 1702 Builder); 1703 } 1704 return Old; 1705 } 1706 1707 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1708 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1709 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1710 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1711 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1712 } 1713 return Old; 1714 } 1715 1716 // If SV is a float, convert it to the appropriate integer type. 1717 // If it is a pointer, do the same. 1718 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1719 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1720 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1721 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1722 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType())) 1723 SV = Builder.CreateBitCast(SV, 1724 IntegerType::get(SV->getContext(),SrcWidth), "tmp"); 1725 else if (isa<PointerType>(SV->getType())) 1726 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(SV->getContext()), "tmp"); 1727 1728 // Zero extend or truncate the value if needed. 1729 if (SV->getType() != AllocaType) { 1730 if (SV->getType()->getPrimitiveSizeInBits() < 1731 AllocaType->getPrimitiveSizeInBits()) 1732 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1733 else { 1734 // Truncation may be needed if storing more than the alloca can hold 1735 // (undefined behavior). 1736 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1737 SrcWidth = DestWidth; 1738 SrcStoreWidth = DestStoreWidth; 1739 } 1740 } 1741 1742 // If this is a big-endian system and the store is narrower than the 1743 // full alloca type, we need to do a shift to get the right bits. 1744 int ShAmt = 0; 1745 if (TD->isBigEndian()) { 1746 // On big-endian machines, the lowest bit is stored at the bit offset 1747 // from the pointer given by getTypeStoreSizeInBits. This matters for 1748 // integers with a bitwidth that is not a multiple of 8. 1749 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1750 } else { 1751 ShAmt = Offset; 1752 } 1753 1754 // Note: we support negative bitwidths (with shr) which are not defined. 1755 // We do this to support (f.e.) stores off the end of a structure where 1756 // only some bits in the structure are set. 1757 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1758 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1759 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1760 ShAmt), "tmp"); 1761 Mask <<= ShAmt; 1762 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1763 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1764 -ShAmt), "tmp"); 1765 Mask = Mask.lshr(-ShAmt); 1766 } 1767 1768 // Mask out the bits we are about to insert from the old value, and or 1769 // in the new bits. 1770 if (SrcWidth != DestWidth) { 1771 assert(DestWidth > SrcWidth); 1772 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1773 SV = Builder.CreateOr(Old, SV, "ins"); 1774 } 1775 return SV; 1776} 1777 1778 1779 1780/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1781/// some part of a constant global variable. This intentionally only accepts 1782/// constant expressions because we don't can't rewrite arbitrary instructions. 1783static bool PointsToConstantGlobal(Value *V) { 1784 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1785 return GV->isConstant(); 1786 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1787 if (CE->getOpcode() == Instruction::BitCast || 1788 CE->getOpcode() == Instruction::GetElementPtr) 1789 return PointsToConstantGlobal(CE->getOperand(0)); 1790 return false; 1791} 1792 1793/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1794/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1795/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1796/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1797/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1798/// the alloca, and if the source pointer is a pointer to a constant global, we 1799/// can optimize this. 1800static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1801 bool isOffset) { 1802 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1803 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) 1804 // Ignore non-volatile loads, they are always ok. 1805 if (!LI->isVolatile()) 1806 continue; 1807 1808 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1809 // If uses of the bitcast are ok, we are ok. 1810 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1811 return false; 1812 continue; 1813 } 1814 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1815 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1816 // doesn't, it does. 1817 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1818 isOffset || !GEP->hasAllZeroIndices())) 1819 return false; 1820 continue; 1821 } 1822 1823 // If this is isn't our memcpy/memmove, reject it as something we can't 1824 // handle. 1825 if (!isa<MemTransferInst>(*UI)) 1826 return false; 1827 1828 // If we already have seen a copy, reject the second one. 1829 if (TheCopy) return false; 1830 1831 // If the pointer has been offset from the start of the alloca, we can't 1832 // safely handle this. 1833 if (isOffset) return false; 1834 1835 // If the memintrinsic isn't using the alloca as the dest, reject it. 1836 if (UI.getOperandNo() != 1) return false; 1837 1838 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1839 1840 // If the source of the memcpy/move is not a constant global, reject it. 1841 if (!PointsToConstantGlobal(MI->getOperand(2))) 1842 return false; 1843 1844 // Otherwise, the transform is safe. Remember the copy instruction. 1845 TheCopy = MI; 1846 } 1847 return true; 1848} 1849 1850/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1851/// modified by a copy from a constant global. If we can prove this, we can 1852/// replace any uses of the alloca with uses of the global directly. 1853Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) { 1854 Instruction *TheCopy = 0; 1855 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1856 return TheCopy; 1857 return 0; 1858} 1859