ScalarReplAggregates.cpp revision 0a205a459884ec745df1c529396dd921f029dafd
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/GetElementPtrTypeIterator.h" 38#include "llvm/Support/IRBuilder.h" 39#include "llvm/Support/MathExtras.h" 40#include "llvm/Support/Compiler.h" 41#include "llvm/ADT/SmallVector.h" 42#include "llvm/ADT/Statistic.h" 43#include "llvm/ADT/StringExtras.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.addRequired<TargetData>(); 72 AU.setPreservesCFG(); 73 } 74 75 private: 76 TargetData *TD; 77 78 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 79 /// information about the uses. All these fields are initialized to false 80 /// and set to true when something is learned. 81 struct AllocaInfo { 82 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 83 bool isUnsafe : 1; 84 85 /// needsCleanup - This is set to true if there is some use of the alloca 86 /// that requires cleanup. 87 bool needsCleanup : 1; 88 89 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 90 bool isMemCpySrc : 1; 91 92 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 93 bool isMemCpyDst : 1; 94 95 AllocaInfo() 96 : isUnsafe(false), needsCleanup(false), 97 isMemCpySrc(false), isMemCpyDst(false) {} 98 }; 99 100 unsigned SRThreshold; 101 102 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 103 104 int isSafeAllocaToScalarRepl(AllocationInst *AI); 105 106 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 107 AllocaInfo &Info); 108 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 109 AllocaInfo &Info); 110 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 111 unsigned OpNo, AllocaInfo &Info); 112 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI, 113 AllocaInfo &Info); 114 115 void DoScalarReplacement(AllocationInst *AI, 116 std::vector<AllocationInst*> &WorkList); 117 void CleanupGEP(GetElementPtrInst *GEP); 118 void CleanupAllocaUsers(AllocationInst *AI); 119 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base); 120 121 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 122 SmallVector<AllocaInst*, 32> &NewElts); 123 124 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 125 AllocationInst *AI, 126 SmallVector<AllocaInst*, 32> &NewElts); 127 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI, 128 SmallVector<AllocaInst*, 32> &NewElts); 129 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 130 SmallVector<AllocaInst*, 32> &NewElts); 131 132 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 133 bool &SawVec, uint64_t Offset, unsigned AllocaSize); 134 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 135 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 136 uint64_t Offset, IRBuilder<> &Builder); 137 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 138 uint64_t Offset, IRBuilder<> &Builder); 139 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI); 140 }; 141} 142 143char SROA::ID = 0; 144static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 145 146// Public interface to the ScalarReplAggregates pass 147FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 148 return new SROA(Threshold); 149} 150 151 152bool SROA::runOnFunction(Function &F) { 153 TD = &getAnalysis<TargetData>(); 154 155 bool Changed = performPromotion(F); 156 while (1) { 157 bool LocalChange = performScalarRepl(F); 158 if (!LocalChange) break; // No need to repromote if no scalarrepl 159 Changed = true; 160 LocalChange = performPromotion(F); 161 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 162 } 163 164 return Changed; 165} 166 167 168bool SROA::performPromotion(Function &F) { 169 std::vector<AllocaInst*> Allocas; 170 DominatorTree &DT = getAnalysis<DominatorTree>(); 171 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 172 173 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 174 175 bool Changed = false; 176 177 while (1) { 178 Allocas.clear(); 179 180 // Find allocas that are safe to promote, by looking at all instructions in 181 // the entry node 182 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 183 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 184 if (isAllocaPromotable(AI)) 185 Allocas.push_back(AI); 186 187 if (Allocas.empty()) break; 188 189 PromoteMemToReg(Allocas, DT, DF, Context); 190 NumPromoted += Allocas.size(); 191 Changed = true; 192 } 193 194 return Changed; 195} 196 197/// getNumSAElements - Return the number of elements in the specific struct or 198/// array. 199static uint64_t getNumSAElements(const Type *T) { 200 if (const StructType *ST = dyn_cast<StructType>(T)) 201 return ST->getNumElements(); 202 return cast<ArrayType>(T)->getNumElements(); 203} 204 205// performScalarRepl - This algorithm is a simple worklist driven algorithm, 206// which runs on all of the malloc/alloca instructions in the function, removing 207// them if they are only used by getelementptr instructions. 208// 209bool SROA::performScalarRepl(Function &F) { 210 std::vector<AllocationInst*> WorkList; 211 212 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 213 BasicBlock &BB = F.getEntryBlock(); 214 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 215 if (AllocationInst *A = dyn_cast<AllocationInst>(I)) 216 WorkList.push_back(A); 217 218 // Process the worklist 219 bool Changed = false; 220 while (!WorkList.empty()) { 221 AllocationInst *AI = WorkList.back(); 222 WorkList.pop_back(); 223 224 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 225 // with unused elements. 226 if (AI->use_empty()) { 227 AI->eraseFromParent(); 228 continue; 229 } 230 231 // If this alloca is impossible for us to promote, reject it early. 232 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 233 continue; 234 235 // Check to see if this allocation is only modified by a memcpy/memmove from 236 // a constant global. If this is the case, we can change all users to use 237 // the constant global instead. This is commonly produced by the CFE by 238 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 239 // is only subsequently read. 240 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 241 DOUT << "Found alloca equal to global: " << *AI; 242 DOUT << " memcpy = " << *TheCopy; 243 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 244 AI->replaceAllUsesWith( 245 Context->getConstantExprBitCast(TheSrc, AI->getType())); 246 TheCopy->eraseFromParent(); // Don't mutate the global. 247 AI->eraseFromParent(); 248 ++NumGlobals; 249 Changed = true; 250 continue; 251 } 252 253 // Check to see if we can perform the core SROA transformation. We cannot 254 // transform the allocation instruction if it is an array allocation 255 // (allocations OF arrays are ok though), and an allocation of a scalar 256 // value cannot be decomposed at all. 257 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 258 259 // Do not promote any struct whose size is too big. 260 if (AllocaSize > SRThreshold) continue; 261 262 if ((isa<StructType>(AI->getAllocatedType()) || 263 isa<ArrayType>(AI->getAllocatedType())) && 264 // Do not promote any struct into more than "32" separate vars. 265 getNumSAElements(AI->getAllocatedType()) <= SRThreshold/4) { 266 // Check that all of the users of the allocation are capable of being 267 // transformed. 268 switch (isSafeAllocaToScalarRepl(AI)) { 269 default: assert(0 && "Unexpected value!"); 270 case 0: // Not safe to scalar replace. 271 break; 272 case 1: // Safe, but requires cleanup/canonicalizations first 273 CleanupAllocaUsers(AI); 274 // FALL THROUGH. 275 case 3: // Safe to scalar replace. 276 DoScalarReplacement(AI, WorkList); 277 Changed = true; 278 continue; 279 } 280 } 281 282 // If we can turn this aggregate value (potentially with casts) into a 283 // simple scalar value that can be mem2reg'd into a register value. 284 // IsNotTrivial tracks whether this is something that mem2reg could have 285 // promoted itself. If so, we don't want to transform it needlessly. Note 286 // that we can't just check based on the type: the alloca may be of an i32 287 // but that has pointer arithmetic to set byte 3 of it or something. 288 bool IsNotTrivial = false; 289 const Type *VectorTy = 0; 290 bool HadAVector = false; 291 if (CanConvertToScalar(AI, IsNotTrivial, VectorTy, HadAVector, 292 0, unsigned(AllocaSize)) && IsNotTrivial) { 293 AllocaInst *NewAI; 294 // If we were able to find a vector type that can handle this with 295 // insert/extract elements, and if there was at least one use that had 296 // a vector type, promote this to a vector. We don't want to promote 297 // random stuff that doesn't use vectors (e.g. <9 x double>) because then 298 // we just get a lot of insert/extracts. If at least one vector is 299 // involved, then we probably really do have a union of vector/array. 300 if (VectorTy && isa<VectorType>(VectorTy) && HadAVector) { 301 DOUT << "CONVERT TO VECTOR: " << *AI << " TYPE = " << *VectorTy <<"\n"; 302 303 // Create and insert the vector alloca. 304 NewAI = new AllocaInst(VectorTy, 0, "", AI->getParent()->begin()); 305 ConvertUsesToScalar(AI, NewAI, 0); 306 } else { 307 DOUT << "CONVERT TO SCALAR INTEGER: " << *AI << "\n"; 308 309 // Create and insert the integer alloca. 310 const Type *NewTy = Context->getIntegerType(AllocaSize*8); 311 NewAI = new AllocaInst(NewTy, 0, "", AI->getParent()->begin()); 312 ConvertUsesToScalar(AI, NewAI, 0); 313 } 314 NewAI->takeName(AI); 315 AI->eraseFromParent(); 316 ++NumConverted; 317 Changed = true; 318 continue; 319 } 320 321 // Otherwise, couldn't process this alloca. 322 } 323 324 return Changed; 325} 326 327/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 328/// predicate, do SROA now. 329void SROA::DoScalarReplacement(AllocationInst *AI, 330 std::vector<AllocationInst*> &WorkList) { 331 DOUT << "Found inst to SROA: " << *AI; 332 SmallVector<AllocaInst*, 32> ElementAllocas; 333 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 334 ElementAllocas.reserve(ST->getNumContainedTypes()); 335 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 336 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 337 AI->getAlignment(), 338 AI->getName() + "." + utostr(i), AI); 339 ElementAllocas.push_back(NA); 340 WorkList.push_back(NA); // Add to worklist for recursive processing 341 } 342 } else { 343 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 344 ElementAllocas.reserve(AT->getNumElements()); 345 const Type *ElTy = AT->getElementType(); 346 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 347 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 348 AI->getName() + "." + utostr(i), AI); 349 ElementAllocas.push_back(NA); 350 WorkList.push_back(NA); // Add to worklist for recursive processing 351 } 352 } 353 354 // Now that we have created the alloca instructions that we want to use, 355 // expand the getelementptr instructions to use them. 356 // 357 while (!AI->use_empty()) { 358 Instruction *User = cast<Instruction>(AI->use_back()); 359 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) { 360 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas); 361 BCInst->eraseFromParent(); 362 continue; 363 } 364 365 // Replace: 366 // %res = load { i32, i32 }* %alloc 367 // with: 368 // %load.0 = load i32* %alloc.0 369 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 370 // %load.1 = load i32* %alloc.1 371 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 372 // (Also works for arrays instead of structs) 373 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 374 Value *Insert = Context->getUndef(LI->getType()); 375 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 376 Value *Load = new LoadInst(ElementAllocas[i], "load", LI); 377 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 378 } 379 LI->replaceAllUsesWith(Insert); 380 LI->eraseFromParent(); 381 continue; 382 } 383 384 // Replace: 385 // store { i32, i32 } %val, { i32, i32 }* %alloc 386 // with: 387 // %val.0 = extractvalue { i32, i32 } %val, 0 388 // store i32 %val.0, i32* %alloc.0 389 // %val.1 = extractvalue { i32, i32 } %val, 1 390 // store i32 %val.1, i32* %alloc.1 391 // (Also works for arrays instead of structs) 392 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 393 Value *Val = SI->getOperand(0); 394 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 395 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 396 new StoreInst(Extract, ElementAllocas[i], SI); 397 } 398 SI->eraseFromParent(); 399 continue; 400 } 401 402 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 403 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 404 unsigned Idx = 405 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 406 407 assert(Idx < ElementAllocas.size() && "Index out of range?"); 408 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 409 410 Value *RepValue; 411 if (GEPI->getNumOperands() == 3) { 412 // Do not insert a new getelementptr instruction with zero indices, only 413 // to have it optimized out later. 414 RepValue = AllocaToUse; 415 } else { 416 // We are indexing deeply into the structure, so we still need a 417 // getelement ptr instruction to finish the indexing. This may be 418 // expanded itself once the worklist is rerun. 419 // 420 SmallVector<Value*, 8> NewArgs; 421 NewArgs.push_back(Context->getNullValue(Type::Int32Ty)); 422 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end()); 423 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(), 424 NewArgs.end(), "", GEPI); 425 RepValue->takeName(GEPI); 426 } 427 428 // If this GEP is to the start of the aggregate, check for memcpys. 429 if (Idx == 0 && GEPI->hasAllZeroIndices()) 430 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas); 431 432 // Move all of the users over to the new GEP. 433 GEPI->replaceAllUsesWith(RepValue); 434 // Delete the old GEP 435 GEPI->eraseFromParent(); 436 } 437 438 // Finally, delete the Alloca instruction 439 AI->eraseFromParent(); 440 NumReplaced++; 441} 442 443 444/// isSafeElementUse - Check to see if this use is an allowed use for a 445/// getelementptr instruction of an array aggregate allocation. isFirstElt 446/// indicates whether Ptr is known to the start of the aggregate. 447/// 448void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 449 AllocaInfo &Info) { 450 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 451 I != E; ++I) { 452 Instruction *User = cast<Instruction>(*I); 453 switch (User->getOpcode()) { 454 case Instruction::Load: break; 455 case Instruction::Store: 456 // Store is ok if storing INTO the pointer, not storing the pointer 457 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info); 458 break; 459 case Instruction::GetElementPtr: { 460 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 461 bool AreAllZeroIndices = isFirstElt; 462 if (GEP->getNumOperands() > 1) { 463 if (!isa<ConstantInt>(GEP->getOperand(1)) || 464 !cast<ConstantInt>(GEP->getOperand(1))->isZero()) 465 // Using pointer arithmetic to navigate the array. 466 return MarkUnsafe(Info); 467 468 if (AreAllZeroIndices) 469 AreAllZeroIndices = GEP->hasAllZeroIndices(); 470 } 471 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 472 if (Info.isUnsafe) return; 473 break; 474 } 475 case Instruction::BitCast: 476 if (isFirstElt) { 477 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 478 if (Info.isUnsafe) return; 479 break; 480 } 481 DOUT << " Transformation preventing inst: " << *User; 482 return MarkUnsafe(Info); 483 case Instruction::Call: 484 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 485 if (isFirstElt) { 486 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 487 if (Info.isUnsafe) return; 488 break; 489 } 490 } 491 DOUT << " Transformation preventing inst: " << *User; 492 return MarkUnsafe(Info); 493 default: 494 DOUT << " Transformation preventing inst: " << *User; 495 return MarkUnsafe(Info); 496 } 497 } 498 return; // All users look ok :) 499} 500 501/// AllUsersAreLoads - Return true if all users of this value are loads. 502static bool AllUsersAreLoads(Value *Ptr) { 503 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 504 I != E; ++I) 505 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 506 return false; 507 return true; 508} 509 510/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 511/// aggregate allocation. 512/// 513void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 514 AllocaInfo &Info) { 515 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 516 return isSafeUseOfBitCastedAllocation(C, AI, Info); 517 518 if (LoadInst *LI = dyn_cast<LoadInst>(User)) 519 if (!LI->isVolatile()) 520 return;// Loads (returning a first class aggregrate) are always rewritable 521 522 if (StoreInst *SI = dyn_cast<StoreInst>(User)) 523 if (!SI->isVolatile() && SI->getOperand(0) != AI) 524 return;// Store is ok if storing INTO the pointer, not storing the pointer 525 526 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 527 if (GEPI == 0) 528 return MarkUnsafe(Info); 529 530 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 531 532 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 533 if (I == E || 534 I.getOperand() != Context->getNullValue(I.getOperand()->getType())) { 535 return MarkUnsafe(Info); 536 } 537 538 ++I; 539 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 540 541 bool IsAllZeroIndices = true; 542 543 // If the first index is a non-constant index into an array, see if we can 544 // handle it as a special case. 545 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 546 if (!isa<ConstantInt>(I.getOperand())) { 547 IsAllZeroIndices = 0; 548 uint64_t NumElements = AT->getNumElements(); 549 550 // If this is an array index and the index is not constant, we cannot 551 // promote... that is unless the array has exactly one or two elements in 552 // it, in which case we CAN promote it, but we have to canonicalize this 553 // out if this is the only problem. 554 if ((NumElements == 1 || NumElements == 2) && 555 AllUsersAreLoads(GEPI)) { 556 Info.needsCleanup = true; 557 return; // Canonicalization required! 558 } 559 return MarkUnsafe(Info); 560 } 561 } 562 563 // Walk through the GEP type indices, checking the types that this indexes 564 // into. 565 for (; I != E; ++I) { 566 // Ignore struct elements, no extra checking needed for these. 567 if (isa<StructType>(*I)) 568 continue; 569 570 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 571 if (!IdxVal) return MarkUnsafe(Info); 572 573 // Are all indices still zero? 574 IsAllZeroIndices &= IdxVal->isZero(); 575 576 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 577 // This GEP indexes an array. Verify that this is an in-range constant 578 // integer. Specifically, consider A[0][i]. We cannot know that the user 579 // isn't doing invalid things like allowing i to index an out-of-range 580 // subscript that accesses A[1]. Because of this, we have to reject SROA 581 // of any accesses into structs where any of the components are variables. 582 if (IdxVal->getZExtValue() >= AT->getNumElements()) 583 return MarkUnsafe(Info); 584 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) { 585 if (IdxVal->getZExtValue() >= VT->getNumElements()) 586 return MarkUnsafe(Info); 587 } 588 } 589 590 // If there are any non-simple uses of this getelementptr, make sure to reject 591 // them. 592 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 593} 594 595/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory 596/// intrinsic can be promoted by SROA. At this point, we know that the operand 597/// of the memintrinsic is a pointer to the beginning of the allocation. 598void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 599 unsigned OpNo, AllocaInfo &Info) { 600 // If not constant length, give up. 601 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 602 if (!Length) return MarkUnsafe(Info); 603 604 // If not the whole aggregate, give up. 605 if (Length->getZExtValue() != 606 TD->getTypeAllocSize(AI->getType()->getElementType())) 607 return MarkUnsafe(Info); 608 609 // We only know about memcpy/memset/memmove. 610 if (!isa<MemIntrinsic>(MI)) 611 return MarkUnsafe(Info); 612 613 // Otherwise, we can transform it. Determine whether this is a memcpy/set 614 // into or out of the aggregate. 615 if (OpNo == 1) 616 Info.isMemCpyDst = true; 617 else { 618 assert(OpNo == 2); 619 Info.isMemCpySrc = true; 620 } 621} 622 623/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast 624/// are 625void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI, 626 AllocaInfo &Info) { 627 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 628 UI != E; ++UI) { 629 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 630 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 631 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 632 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 633 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 634 if (SI->isVolatile()) 635 return MarkUnsafe(Info); 636 637 // If storing the entire alloca in one chunk through a bitcasted pointer 638 // to integer, we can transform it. This happens (for example) when you 639 // cast a {i32,i32}* to i64* and store through it. This is similar to the 640 // memcpy case and occurs in various "byval" cases and emulated memcpys. 641 if (isa<IntegerType>(SI->getOperand(0)->getType()) && 642 TD->getTypeAllocSize(SI->getOperand(0)->getType()) == 643 TD->getTypeAllocSize(AI->getType()->getElementType())) { 644 Info.isMemCpyDst = true; 645 continue; 646 } 647 return MarkUnsafe(Info); 648 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 649 if (LI->isVolatile()) 650 return MarkUnsafe(Info); 651 652 // If loading the entire alloca in one chunk through a bitcasted pointer 653 // to integer, we can transform it. This happens (for example) when you 654 // cast a {i32,i32}* to i64* and load through it. This is similar to the 655 // memcpy case and occurs in various "byval" cases and emulated memcpys. 656 if (isa<IntegerType>(LI->getType()) && 657 TD->getTypeAllocSize(LI->getType()) == 658 TD->getTypeAllocSize(AI->getType()->getElementType())) { 659 Info.isMemCpySrc = true; 660 continue; 661 } 662 return MarkUnsafe(Info); 663 } else if (isa<DbgInfoIntrinsic>(UI)) { 664 // If one user is DbgInfoIntrinsic then check if all users are 665 // DbgInfoIntrinsics. 666 if (OnlyUsedByDbgInfoIntrinsics(BC)) { 667 Info.needsCleanup = true; 668 return; 669 } 670 else 671 MarkUnsafe(Info); 672 } 673 else { 674 return MarkUnsafe(Info); 675 } 676 if (Info.isUnsafe) return; 677 } 678} 679 680/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 681/// to its first element. Transform users of the cast to use the new values 682/// instead. 683void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 684 SmallVector<AllocaInst*, 32> &NewElts) { 685 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 686 while (UI != UE) { 687 Instruction *User = cast<Instruction>(*UI++); 688 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) { 689 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 690 if (BCU->use_empty()) BCU->eraseFromParent(); 691 continue; 692 } 693 694 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 695 // This must be memcpy/memmove/memset of the entire aggregate. 696 // Split into one per element. 697 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts); 698 continue; 699 } 700 701 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 702 // If this is a store of the entire alloca from an integer, rewrite it. 703 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 704 continue; 705 } 706 707 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 708 // If this is a load of the entire alloca to an integer, rewrite it. 709 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 710 continue; 711 } 712 713 // Otherwise it must be some other user of a gep of the first pointer. Just 714 // leave these alone. 715 continue; 716 } 717} 718 719/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 720/// Rewrite it to copy or set the elements of the scalarized memory. 721void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 722 AllocationInst *AI, 723 SmallVector<AllocaInst*, 32> &NewElts) { 724 725 // If this is a memcpy/memmove, construct the other pointer as the 726 // appropriate type. The "Other" pointer is the pointer that goes to memory 727 // that doesn't have anything to do with the alloca that we are promoting. For 728 // memset, this Value* stays null. 729 Value *OtherPtr = 0; 730 unsigned MemAlignment = MI->getAlignment(); 731 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 732 if (BCInst == MTI->getRawDest()) 733 OtherPtr = MTI->getRawSource(); 734 else { 735 assert(BCInst == MTI->getRawSource()); 736 OtherPtr = MTI->getRawDest(); 737 } 738 } 739 740 // If there is an other pointer, we want to convert it to the same pointer 741 // type as AI has, so we can GEP through it safely. 742 if (OtherPtr) { 743 // It is likely that OtherPtr is a bitcast, if so, remove it. 744 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 745 OtherPtr = BC->getOperand(0); 746 // All zero GEPs are effectively bitcasts. 747 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) 748 if (GEP->hasAllZeroIndices()) 749 OtherPtr = GEP->getOperand(0); 750 751 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 752 if (BCE->getOpcode() == Instruction::BitCast) 753 OtherPtr = BCE->getOperand(0); 754 755 // If the pointer is not the right type, insert a bitcast to the right 756 // type. 757 if (OtherPtr->getType() != AI->getType()) 758 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 759 MI); 760 } 761 762 // Process each element of the aggregate. 763 Value *TheFn = MI->getOperand(0); 764 const Type *BytePtrTy = MI->getRawDest()->getType(); 765 bool SROADest = MI->getRawDest() == BCInst; 766 767 Constant *Zero = Context->getNullValue(Type::Int32Ty); 768 769 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 770 // If this is a memcpy/memmove, emit a GEP of the other element address. 771 Value *OtherElt = 0; 772 unsigned OtherEltAlign = MemAlignment; 773 774 if (OtherPtr) { 775 Value *Idx[2] = { Zero, Context->getConstantInt(Type::Int32Ty, i) }; 776 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2, 777 OtherPtr->getNameStr()+"."+utostr(i), 778 MI); 779 uint64_t EltOffset; 780 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 781 if (const StructType *ST = 782 dyn_cast<StructType>(OtherPtrTy->getElementType())) { 783 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 784 } else { 785 const Type *EltTy = 786 cast<SequentialType>(OtherPtr->getType())->getElementType(); 787 EltOffset = TD->getTypeAllocSize(EltTy)*i; 788 } 789 790 // The alignment of the other pointer is the guaranteed alignment of the 791 // element, which is affected by both the known alignment of the whole 792 // mem intrinsic and the alignment of the element. If the alignment of 793 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 794 // known alignment is just 4 bytes. 795 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 796 } 797 798 Value *EltPtr = NewElts[i]; 799 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 800 801 // If we got down to a scalar, insert a load or store as appropriate. 802 if (EltTy->isSingleValueType()) { 803 if (isa<MemTransferInst>(MI)) { 804 if (SROADest) { 805 // From Other to Alloca. 806 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 807 new StoreInst(Elt, EltPtr, MI); 808 } else { 809 // From Alloca to Other. 810 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 811 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 812 } 813 continue; 814 } 815 assert(isa<MemSetInst>(MI)); 816 817 // If the stored element is zero (common case), just store a null 818 // constant. 819 Constant *StoreVal; 820 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 821 if (CI->isZero()) { 822 StoreVal = Context->getNullValue(EltTy); // 0.0, null, 0, <0,0> 823 } else { 824 // If EltTy is a vector type, get the element type. 825 const Type *ValTy = EltTy->getScalarType(); 826 827 // Construct an integer with the right value. 828 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 829 APInt OneVal(EltSize, CI->getZExtValue()); 830 APInt TotalVal(OneVal); 831 // Set each byte. 832 for (unsigned i = 0; 8*i < EltSize; ++i) { 833 TotalVal = TotalVal.shl(8); 834 TotalVal |= OneVal; 835 } 836 837 // Convert the integer value to the appropriate type. 838 StoreVal = Context->getConstantInt(TotalVal); 839 if (isa<PointerType>(ValTy)) 840 StoreVal = Context->getConstantExprIntToPtr(StoreVal, ValTy); 841 else if (ValTy->isFloatingPoint()) 842 StoreVal = Context->getConstantExprBitCast(StoreVal, ValTy); 843 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 844 845 // If the requested value was a vector constant, create it. 846 if (EltTy != ValTy) { 847 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 848 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 849 StoreVal = Context->getConstantVector(&Elts[0], NumElts); 850 } 851 } 852 new StoreInst(StoreVal, EltPtr, MI); 853 continue; 854 } 855 // Otherwise, if we're storing a byte variable, use a memset call for 856 // this element. 857 } 858 859 // Cast the element pointer to BytePtrTy. 860 if (EltPtr->getType() != BytePtrTy) 861 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 862 863 // Cast the other pointer (if we have one) to BytePtrTy. 864 if (OtherElt && OtherElt->getType() != BytePtrTy) 865 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 866 MI); 867 868 unsigned EltSize = TD->getTypeAllocSize(EltTy); 869 870 // Finally, insert the meminst for this element. 871 if (isa<MemTransferInst>(MI)) { 872 Value *Ops[] = { 873 SROADest ? EltPtr : OtherElt, // Dest ptr 874 SROADest ? OtherElt : EltPtr, // Src ptr 875 Context->getConstantInt(MI->getOperand(3)->getType(), EltSize), // Size 876 Context->getConstantInt(Type::Int32Ty, OtherEltAlign) // Align 877 }; 878 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 879 } else { 880 assert(isa<MemSetInst>(MI)); 881 Value *Ops[] = { 882 EltPtr, MI->getOperand(2), // Dest, Value, 883 Context->getConstantInt(MI->getOperand(3)->getType(), EltSize), // Size 884 Zero // Align 885 }; 886 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 887 } 888 } 889 MI->eraseFromParent(); 890} 891 892/// RewriteStoreUserOfWholeAlloca - We found an store of an integer that 893/// overwrites the entire allocation. Extract out the pieces of the stored 894/// integer and store them individually. 895void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, 896 AllocationInst *AI, 897 SmallVector<AllocaInst*, 32> &NewElts){ 898 // Extract each element out of the integer according to its structure offset 899 // and store the element value to the individual alloca. 900 Value *SrcVal = SI->getOperand(0); 901 const Type *AllocaEltTy = AI->getType()->getElementType(); 902 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 903 904 // If this isn't a store of an integer to the whole alloca, it may be a store 905 // to the first element. Just ignore the store in this case and normal SROA 906 // will handle it. 907 if (!isa<IntegerType>(SrcVal->getType()) || 908 TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits) 909 return; 910 // Handle tail padding by extending the operand 911 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 912 SrcVal = new ZExtInst(SrcVal, 913 Context->getIntegerType(AllocaSizeBits), "", SI); 914 915 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI; 916 917 // There are two forms here: AI could be an array or struct. Both cases 918 // have different ways to compute the element offset. 919 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 920 const StructLayout *Layout = TD->getStructLayout(EltSTy); 921 922 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 923 // Get the number of bits to shift SrcVal to get the value. 924 const Type *FieldTy = EltSTy->getElementType(i); 925 uint64_t Shift = Layout->getElementOffsetInBits(i); 926 927 if (TD->isBigEndian()) 928 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 929 930 Value *EltVal = SrcVal; 931 if (Shift) { 932 Value *ShiftVal = Context->getConstantInt(EltVal->getType(), Shift); 933 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 934 "sroa.store.elt", SI); 935 } 936 937 // Truncate down to an integer of the right size. 938 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 939 940 // Ignore zero sized fields like {}, they obviously contain no data. 941 if (FieldSizeBits == 0) continue; 942 943 if (FieldSizeBits != AllocaSizeBits) 944 EltVal = new TruncInst(EltVal, 945 Context->getIntegerType(FieldSizeBits), "", SI); 946 Value *DestField = NewElts[i]; 947 if (EltVal->getType() == FieldTy) { 948 // Storing to an integer field of this size, just do it. 949 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { 950 // Bitcast to the right element type (for fp/vector values). 951 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 952 } else { 953 // Otherwise, bitcast the dest pointer (for aggregates). 954 DestField = new BitCastInst(DestField, 955 Context->getPointerTypeUnqual(EltVal->getType()), 956 "", SI); 957 } 958 new StoreInst(EltVal, DestField, SI); 959 } 960 961 } else { 962 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 963 const Type *ArrayEltTy = ATy->getElementType(); 964 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 965 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 966 967 uint64_t Shift; 968 969 if (TD->isBigEndian()) 970 Shift = AllocaSizeBits-ElementOffset; 971 else 972 Shift = 0; 973 974 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 975 // Ignore zero sized fields like {}, they obviously contain no data. 976 if (ElementSizeBits == 0) continue; 977 978 Value *EltVal = SrcVal; 979 if (Shift) { 980 Value *ShiftVal = Context->getConstantInt(EltVal->getType(), Shift); 981 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 982 "sroa.store.elt", SI); 983 } 984 985 // Truncate down to an integer of the right size. 986 if (ElementSizeBits != AllocaSizeBits) 987 EltVal = new TruncInst(EltVal, 988 Context->getIntegerType(ElementSizeBits),"",SI); 989 Value *DestField = NewElts[i]; 990 if (EltVal->getType() == ArrayEltTy) { 991 // Storing to an integer field of this size, just do it. 992 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { 993 // Bitcast to the right element type (for fp/vector values). 994 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 995 } else { 996 // Otherwise, bitcast the dest pointer (for aggregates). 997 DestField = new BitCastInst(DestField, 998 Context->getPointerTypeUnqual(EltVal->getType()), 999 "", SI); 1000 } 1001 new StoreInst(EltVal, DestField, SI); 1002 1003 if (TD->isBigEndian()) 1004 Shift -= ElementOffset; 1005 else 1006 Shift += ElementOffset; 1007 } 1008 } 1009 1010 SI->eraseFromParent(); 1011} 1012 1013/// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to 1014/// an integer. Load the individual pieces to form the aggregate value. 1015void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 1016 SmallVector<AllocaInst*, 32> &NewElts) { 1017 // Extract each element out of the NewElts according to its structure offset 1018 // and form the result value. 1019 const Type *AllocaEltTy = AI->getType()->getElementType(); 1020 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1021 1022 // If this isn't a load of the whole alloca to an integer, it may be a load 1023 // of the first element. Just ignore the load in this case and normal SROA 1024 // will handle it. 1025 if (!isa<IntegerType>(LI->getType()) || 1026 TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits) 1027 return; 1028 1029 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI; 1030 1031 // There are two forms here: AI could be an array or struct. Both cases 1032 // have different ways to compute the element offset. 1033 const StructLayout *Layout = 0; 1034 uint64_t ArrayEltBitOffset = 0; 1035 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1036 Layout = TD->getStructLayout(EltSTy); 1037 } else { 1038 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1039 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1040 } 1041 1042 Value *ResultVal = 1043 Context->getNullValue(Context->getIntegerType(AllocaSizeBits)); 1044 1045 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1046 // Load the value from the alloca. If the NewElt is an aggregate, cast 1047 // the pointer to an integer of the same size before doing the load. 1048 Value *SrcField = NewElts[i]; 1049 const Type *FieldTy = 1050 cast<PointerType>(SrcField->getType())->getElementType(); 1051 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1052 1053 // Ignore zero sized fields like {}, they obviously contain no data. 1054 if (FieldSizeBits == 0) continue; 1055 1056 const IntegerType *FieldIntTy = Context->getIntegerType(FieldSizeBits); 1057 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && 1058 !isa<VectorType>(FieldTy)) 1059 SrcField = new BitCastInst(SrcField, 1060 Context->getPointerTypeUnqual(FieldIntTy), 1061 "", LI); 1062 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1063 1064 // If SrcField is a fp or vector of the right size but that isn't an 1065 // integer type, bitcast to an integer so we can shift it. 1066 if (SrcField->getType() != FieldIntTy) 1067 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1068 1069 // Zero extend the field to be the same size as the final alloca so that 1070 // we can shift and insert it. 1071 if (SrcField->getType() != ResultVal->getType()) 1072 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1073 1074 // Determine the number of bits to shift SrcField. 1075 uint64_t Shift; 1076 if (Layout) // Struct case. 1077 Shift = Layout->getElementOffsetInBits(i); 1078 else // Array case. 1079 Shift = i*ArrayEltBitOffset; 1080 1081 if (TD->isBigEndian()) 1082 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1083 1084 if (Shift) { 1085 Value *ShiftVal = Context->getConstantInt(SrcField->getType(), Shift); 1086 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1087 } 1088 1089 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1090 } 1091 1092 // Handle tail padding by truncating the result 1093 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1094 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1095 1096 LI->replaceAllUsesWith(ResultVal); 1097 LI->eraseFromParent(); 1098} 1099 1100 1101/// HasPadding - Return true if the specified type has any structure or 1102/// alignment padding, false otherwise. 1103static bool HasPadding(const Type *Ty, const TargetData &TD) { 1104 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1105 const StructLayout *SL = TD.getStructLayout(STy); 1106 unsigned PrevFieldBitOffset = 0; 1107 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1108 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1109 1110 // Padding in sub-elements? 1111 if (HasPadding(STy->getElementType(i), TD)) 1112 return true; 1113 1114 // Check to see if there is any padding between this element and the 1115 // previous one. 1116 if (i) { 1117 unsigned PrevFieldEnd = 1118 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1119 if (PrevFieldEnd < FieldBitOffset) 1120 return true; 1121 } 1122 1123 PrevFieldBitOffset = FieldBitOffset; 1124 } 1125 1126 // Check for tail padding. 1127 if (unsigned EltCount = STy->getNumElements()) { 1128 unsigned PrevFieldEnd = PrevFieldBitOffset + 1129 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1130 if (PrevFieldEnd < SL->getSizeInBits()) 1131 return true; 1132 } 1133 1134 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1135 return HasPadding(ATy->getElementType(), TD); 1136 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1137 return HasPadding(VTy->getElementType(), TD); 1138 } 1139 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1140} 1141 1142/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1143/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1144/// or 1 if safe after canonicalization has been performed. 1145/// 1146int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 1147 // Loop over the use list of the alloca. We can only transform it if all of 1148 // the users are safe to transform. 1149 AllocaInfo Info; 1150 1151 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 1152 I != E; ++I) { 1153 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 1154 if (Info.isUnsafe) { 1155 DOUT << "Cannot transform: " << *AI << " due to user: " << **I; 1156 return 0; 1157 } 1158 } 1159 1160 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1161 // source and destination, we have to be careful. In particular, the memcpy 1162 // could be moving around elements that live in structure padding of the LLVM 1163 // types, but may actually be used. In these cases, we refuse to promote the 1164 // struct. 1165 if (Info.isMemCpySrc && Info.isMemCpyDst && 1166 HasPadding(AI->getType()->getElementType(), *TD)) 1167 return 0; 1168 1169 // If we require cleanup, return 1, otherwise return 3. 1170 return Info.needsCleanup ? 1 : 3; 1171} 1172 1173/// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP 1174/// is canonicalized here. 1175void SROA::CleanupGEP(GetElementPtrInst *GEPI) { 1176 gep_type_iterator I = gep_type_begin(GEPI); 1177 ++I; 1178 1179 const ArrayType *AT = dyn_cast<ArrayType>(*I); 1180 if (!AT) 1181 return; 1182 1183 uint64_t NumElements = AT->getNumElements(); 1184 1185 if (isa<ConstantInt>(I.getOperand())) 1186 return; 1187 1188 if (NumElements == 1) { 1189 GEPI->setOperand(2, Context->getNullValue(Type::Int32Ty)); 1190 return; 1191 } 1192 1193 assert(NumElements == 2 && "Unhandled case!"); 1194 // All users of the GEP must be loads. At each use of the GEP, insert 1195 // two loads of the appropriate indexed GEP and select between them. 1196 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(), 1197 Context->getNullValue(I.getOperand()->getType()), 1198 "isone", GEPI); 1199 // Insert the new GEP instructions, which are properly indexed. 1200 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 1201 Indices[1] = Context->getNullValue(Type::Int32Ty); 1202 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1203 Indices.begin(), 1204 Indices.end(), 1205 GEPI->getName()+".0", GEPI); 1206 Indices[1] = Context->getConstantInt(Type::Int32Ty, 1); 1207 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1208 Indices.begin(), 1209 Indices.end(), 1210 GEPI->getName()+".1", GEPI); 1211 // Replace all loads of the variable index GEP with loads from both 1212 // indexes and a select. 1213 while (!GEPI->use_empty()) { 1214 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 1215 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 1216 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 1217 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 1218 LI->replaceAllUsesWith(R); 1219 LI->eraseFromParent(); 1220 } 1221 GEPI->eraseFromParent(); 1222} 1223 1224 1225/// CleanupAllocaUsers - If SROA reported that it can promote the specified 1226/// allocation, but only if cleaned up, perform the cleanups required. 1227void SROA::CleanupAllocaUsers(AllocationInst *AI) { 1228 // At this point, we know that the end result will be SROA'd and promoted, so 1229 // we can insert ugly code if required so long as sroa+mem2reg will clean it 1230 // up. 1231 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1232 UI != E; ) { 1233 User *U = *UI++; 1234 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) 1235 CleanupGEP(GEPI); 1236 else { 1237 Instruction *I = cast<Instruction>(U); 1238 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses; 1239 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) { 1240 // Safe to remove debug info uses. 1241 while (!DbgInUses.empty()) { 1242 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back(); 1243 DI->eraseFromParent(); 1244 } 1245 I->eraseFromParent(); 1246 } 1247 } 1248 } 1249} 1250 1251/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 1252/// the offset specified by Offset (which is specified in bytes). 1253/// 1254/// There are two cases we handle here: 1255/// 1) A union of vector types of the same size and potentially its elements. 1256/// Here we turn element accesses into insert/extract element operations. 1257/// This promotes a <4 x float> with a store of float to the third element 1258/// into a <4 x float> that uses insert element. 1259/// 2) A fully general blob of memory, which we turn into some (potentially 1260/// large) integer type with extract and insert operations where the loads 1261/// and stores would mutate the memory. 1262static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy, 1263 unsigned AllocaSize, const TargetData &TD, 1264 LLVMContext* Context) { 1265 // If this could be contributing to a vector, analyze it. 1266 if (VecTy != Type::VoidTy) { // either null or a vector type. 1267 1268 // If the In type is a vector that is the same size as the alloca, see if it 1269 // matches the existing VecTy. 1270 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 1271 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 1272 // If we're storing/loading a vector of the right size, allow it as a 1273 // vector. If this the first vector we see, remember the type so that 1274 // we know the element size. 1275 if (VecTy == 0) 1276 VecTy = VInTy; 1277 return; 1278 } 1279 } else if (In == Type::FloatTy || In == Type::DoubleTy || 1280 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 && 1281 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 1282 // If we're accessing something that could be an element of a vector, see 1283 // if the implied vector agrees with what we already have and if Offset is 1284 // compatible with it. 1285 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 1286 if (Offset % EltSize == 0 && 1287 AllocaSize % EltSize == 0 && 1288 (VecTy == 0 || 1289 cast<VectorType>(VecTy)->getElementType() 1290 ->getPrimitiveSizeInBits()/8 == EltSize)) { 1291 if (VecTy == 0) 1292 VecTy = Context->getVectorType(In, AllocaSize/EltSize); 1293 return; 1294 } 1295 } 1296 } 1297 1298 // Otherwise, we have a case that we can't handle with an optimized vector 1299 // form. We can still turn this into a large integer. 1300 VecTy = Type::VoidTy; 1301} 1302 1303/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 1304/// its accesses to use a to single vector type, return true, and set VecTy to 1305/// the new type. If we could convert the alloca into a single promotable 1306/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 1307/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 1308/// is the current offset from the base of the alloca being analyzed. 1309/// 1310/// If we see at least one access to the value that is as a vector type, set the 1311/// SawVec flag. 1312/// 1313bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 1314 bool &SawVec, uint64_t Offset, 1315 unsigned AllocaSize) { 1316 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1317 Instruction *User = cast<Instruction>(*UI); 1318 1319 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1320 // Don't break volatile loads. 1321 if (LI->isVolatile()) 1322 return false; 1323 MergeInType(LI->getType(), Offset, VecTy, AllocaSize, *TD, Context); 1324 SawVec |= isa<VectorType>(LI->getType()); 1325 continue; 1326 } 1327 1328 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1329 // Storing the pointer, not into the value? 1330 if (SI->getOperand(0) == V || SI->isVolatile()) return 0; 1331 MergeInType(SI->getOperand(0)->getType(), Offset, 1332 VecTy, AllocaSize, *TD, Context); 1333 SawVec |= isa<VectorType>(SI->getOperand(0)->getType()); 1334 continue; 1335 } 1336 1337 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 1338 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset, 1339 AllocaSize)) 1340 return false; 1341 IsNotTrivial = true; 1342 continue; 1343 } 1344 1345 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1346 // If this is a GEP with a variable indices, we can't handle it. 1347 if (!GEP->hasAllConstantIndices()) 1348 return false; 1349 1350 // Compute the offset that this GEP adds to the pointer. 1351 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1352 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1353 &Indices[0], Indices.size()); 1354 // See if all uses can be converted. 1355 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset, 1356 AllocaSize)) 1357 return false; 1358 IsNotTrivial = true; 1359 continue; 1360 } 1361 1362 // If this is a constant sized memset of a constant value (e.g. 0) we can 1363 // handle it. 1364 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1365 // Store of constant value and constant size. 1366 if (isa<ConstantInt>(MSI->getValue()) && 1367 isa<ConstantInt>(MSI->getLength())) { 1368 IsNotTrivial = true; 1369 continue; 1370 } 1371 } 1372 1373 // If this is a memcpy or memmove into or out of the whole allocation, we 1374 // can handle it like a load or store of the scalar type. 1375 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1376 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 1377 if (Len->getZExtValue() == AllocaSize && Offset == 0) { 1378 IsNotTrivial = true; 1379 continue; 1380 } 1381 } 1382 1383 // Ignore dbg intrinsic. 1384 if (isa<DbgInfoIntrinsic>(User)) 1385 continue; 1386 1387 // Otherwise, we cannot handle this! 1388 return false; 1389 } 1390 1391 return true; 1392} 1393 1394 1395/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1396/// directly. This happens when we are converting an "integer union" to a 1397/// single integer scalar, or when we are converting a "vector union" to a 1398/// vector with insert/extractelement instructions. 1399/// 1400/// Offset is an offset from the original alloca, in bits that need to be 1401/// shifted to the right. By the end of this, there should be no uses of Ptr. 1402void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 1403 while (!Ptr->use_empty()) { 1404 Instruction *User = cast<Instruction>(Ptr->use_back()); 1405 1406 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1407 ConvertUsesToScalar(CI, NewAI, Offset); 1408 CI->eraseFromParent(); 1409 continue; 1410 } 1411 1412 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1413 // Compute the offset that this GEP adds to the pointer. 1414 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1415 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1416 &Indices[0], Indices.size()); 1417 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 1418 GEP->eraseFromParent(); 1419 continue; 1420 } 1421 1422 IRBuilder<> Builder(User->getParent(), User); 1423 1424 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1425 // The load is a bit extract from NewAI shifted right by Offset bits. 1426 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 1427 Value *NewLoadVal 1428 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 1429 LI->replaceAllUsesWith(NewLoadVal); 1430 LI->eraseFromParent(); 1431 continue; 1432 } 1433 1434 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1435 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1436 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str()); 1437 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 1438 Builder); 1439 Builder.CreateStore(New, NewAI); 1440 SI->eraseFromParent(); 1441 continue; 1442 } 1443 1444 // If this is a constant sized memset of a constant value (e.g. 0) we can 1445 // transform it into a store of the expanded constant value. 1446 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1447 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 1448 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 1449 if (NumBytes != 0) { 1450 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 1451 1452 // Compute the value replicated the right number of times. 1453 APInt APVal(NumBytes*8, Val); 1454 1455 // Splat the value if non-zero. 1456 if (Val) 1457 for (unsigned i = 1; i != NumBytes; ++i) 1458 APVal |= APVal << 8; 1459 1460 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").c_str()); 1461 Value *New = ConvertScalar_InsertValue(Context->getConstantInt(APVal), 1462 Old, Offset, Builder); 1463 Builder.CreateStore(New, NewAI); 1464 } 1465 MSI->eraseFromParent(); 1466 continue; 1467 } 1468 1469 // If this is a memcpy or memmove into or out of the whole allocation, we 1470 // can handle it like a load or store of the scalar type. 1471 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1472 assert(Offset == 0 && "must be store to start of alloca"); 1473 1474 // If the source and destination are both to the same alloca, then this is 1475 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 1476 // as appropriate. 1477 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject()); 1478 1479 if (MTI->getSource()->getUnderlyingObject() != OrigAI) { 1480 // Dest must be OrigAI, change this to be a load from the original 1481 // pointer (bitcasted), then a store to our new alloca. 1482 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 1483 Value *SrcPtr = MTI->getSource(); 1484 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 1485 1486 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 1487 SrcVal->setAlignment(MTI->getAlignment()); 1488 Builder.CreateStore(SrcVal, NewAI); 1489 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) { 1490 // Src must be OrigAI, change this to be a load from NewAI then a store 1491 // through the original dest pointer (bitcasted). 1492 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 1493 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 1494 1495 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 1496 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 1497 NewStore->setAlignment(MTI->getAlignment()); 1498 } else { 1499 // Noop transfer. Src == Dst 1500 } 1501 1502 1503 MTI->eraseFromParent(); 1504 continue; 1505 } 1506 1507 // If user is a dbg info intrinsic then it is safe to remove it. 1508 if (isa<DbgInfoIntrinsic>(User)) { 1509 User->eraseFromParent(); 1510 continue; 1511 } 1512 1513 assert(0 && "Unsupported operation!"); 1514 abort(); 1515 } 1516} 1517 1518/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 1519/// or vector value FromVal, extracting the bits from the offset specified by 1520/// Offset. This returns the value, which is of type ToType. 1521/// 1522/// This happens when we are converting an "integer union" to a single 1523/// integer scalar, or when we are converting a "vector union" to a vector with 1524/// insert/extractelement instructions. 1525/// 1526/// Offset is an offset from the original alloca, in bits that need to be 1527/// shifted to the right. 1528Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 1529 uint64_t Offset, IRBuilder<> &Builder) { 1530 // If the load is of the whole new alloca, no conversion is needed. 1531 if (FromVal->getType() == ToType && Offset == 0) 1532 return FromVal; 1533 1534 // If the result alloca is a vector type, this is either an element 1535 // access or a bitcast to another vector type of the same size. 1536 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 1537 if (isa<VectorType>(ToType)) 1538 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 1539 1540 // Otherwise it must be an element access. 1541 unsigned Elt = 0; 1542 if (Offset) { 1543 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1544 Elt = Offset/EltSize; 1545 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 1546 } 1547 // Return the element extracted out of it. 1548 Value *V = Builder.CreateExtractElement(FromVal, 1549 Context->getConstantInt(Type::Int32Ty,Elt), 1550 "tmp"); 1551 if (V->getType() != ToType) 1552 V = Builder.CreateBitCast(V, ToType, "tmp"); 1553 return V; 1554 } 1555 1556 // If ToType is a first class aggregate, extract out each of the pieces and 1557 // use insertvalue's to form the FCA. 1558 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 1559 const StructLayout &Layout = *TD->getStructLayout(ST); 1560 Value *Res = Context->getUndef(ST); 1561 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1562 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 1563 Offset+Layout.getElementOffsetInBits(i), 1564 Builder); 1565 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1566 } 1567 return Res; 1568 } 1569 1570 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 1571 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1572 Value *Res = Context->getUndef(AT); 1573 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1574 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 1575 Offset+i*EltSize, Builder); 1576 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1577 } 1578 return Res; 1579 } 1580 1581 // Otherwise, this must be a union that was converted to an integer value. 1582 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 1583 1584 // If this is a big-endian system and the load is narrower than the 1585 // full alloca type, we need to do a shift to get the right bits. 1586 int ShAmt = 0; 1587 if (TD->isBigEndian()) { 1588 // On big-endian machines, the lowest bit is stored at the bit offset 1589 // from the pointer given by getTypeStoreSizeInBits. This matters for 1590 // integers with a bitwidth that is not a multiple of 8. 1591 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1592 TD->getTypeStoreSizeInBits(ToType) - Offset; 1593 } else { 1594 ShAmt = Offset; 1595 } 1596 1597 // Note: we support negative bitwidths (with shl) which are not defined. 1598 // We do this to support (f.e.) loads off the end of a structure where 1599 // only some bits are used. 1600 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1601 FromVal = Builder.CreateLShr(FromVal, 1602 Context->getConstantInt(FromVal->getType(), 1603 ShAmt), "tmp"); 1604 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1605 FromVal = Builder.CreateShl(FromVal, 1606 Context->getConstantInt(FromVal->getType(), 1607 -ShAmt), "tmp"); 1608 1609 // Finally, unconditionally truncate the integer to the right width. 1610 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 1611 if (LIBitWidth < NTy->getBitWidth()) 1612 FromVal = 1613 Builder.CreateTrunc(FromVal, Context->getIntegerType(LIBitWidth), "tmp"); 1614 else if (LIBitWidth > NTy->getBitWidth()) 1615 FromVal = 1616 Builder.CreateZExt(FromVal, Context->getIntegerType(LIBitWidth), "tmp"); 1617 1618 // If the result is an integer, this is a trunc or bitcast. 1619 if (isa<IntegerType>(ToType)) { 1620 // Should be done. 1621 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) { 1622 // Just do a bitcast, we know the sizes match up. 1623 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 1624 } else { 1625 // Otherwise must be a pointer. 1626 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 1627 } 1628 assert(FromVal->getType() == ToType && "Didn't convert right?"); 1629 return FromVal; 1630} 1631 1632 1633/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 1634/// or vector value "Old" at the offset specified by Offset. 1635/// 1636/// This happens when we are converting an "integer union" to a 1637/// single integer scalar, or when we are converting a "vector union" to a 1638/// vector with insert/extractelement instructions. 1639/// 1640/// Offset is an offset from the original alloca, in bits that need to be 1641/// shifted to the right. 1642Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 1643 uint64_t Offset, IRBuilder<> &Builder) { 1644 1645 // Convert the stored type to the actual type, shift it left to insert 1646 // then 'or' into place. 1647 const Type *AllocaType = Old->getType(); 1648 1649 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 1650 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy); 1651 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType()); 1652 1653 // Changing the whole vector with memset or with an access of a different 1654 // vector type? 1655 if (ValSize == VecSize) 1656 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 1657 1658 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1659 1660 // Must be an element insertion. 1661 unsigned Elt = Offset/EltSize; 1662 1663 if (SV->getType() != VTy->getElementType()) 1664 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 1665 1666 SV = Builder.CreateInsertElement(Old, SV, 1667 Context->getConstantInt(Type::Int32Ty, Elt), 1668 "tmp"); 1669 return SV; 1670 } 1671 1672 // If SV is a first-class aggregate value, insert each value recursively. 1673 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 1674 const StructLayout &Layout = *TD->getStructLayout(ST); 1675 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1676 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1677 Old = ConvertScalar_InsertValue(Elt, Old, 1678 Offset+Layout.getElementOffsetInBits(i), 1679 Builder); 1680 } 1681 return Old; 1682 } 1683 1684 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1685 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1686 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1687 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1688 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1689 } 1690 return Old; 1691 } 1692 1693 // If SV is a float, convert it to the appropriate integer type. 1694 // If it is a pointer, do the same. 1695 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1696 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1697 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1698 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1699 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType())) 1700 SV = Builder.CreateBitCast(SV, Context->getIntegerType(SrcWidth), "tmp"); 1701 else if (isa<PointerType>(SV->getType())) 1702 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(), "tmp"); 1703 1704 // Zero extend or truncate the value if needed. 1705 if (SV->getType() != AllocaType) { 1706 if (SV->getType()->getPrimitiveSizeInBits() < 1707 AllocaType->getPrimitiveSizeInBits()) 1708 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1709 else { 1710 // Truncation may be needed if storing more than the alloca can hold 1711 // (undefined behavior). 1712 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1713 SrcWidth = DestWidth; 1714 SrcStoreWidth = DestStoreWidth; 1715 } 1716 } 1717 1718 // If this is a big-endian system and the store is narrower than the 1719 // full alloca type, we need to do a shift to get the right bits. 1720 int ShAmt = 0; 1721 if (TD->isBigEndian()) { 1722 // On big-endian machines, the lowest bit is stored at the bit offset 1723 // from the pointer given by getTypeStoreSizeInBits. This matters for 1724 // integers with a bitwidth that is not a multiple of 8. 1725 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1726 } else { 1727 ShAmt = Offset; 1728 } 1729 1730 // Note: we support negative bitwidths (with shr) which are not defined. 1731 // We do this to support (f.e.) stores off the end of a structure where 1732 // only some bits in the structure are set. 1733 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1734 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1735 SV = Builder.CreateShl(SV, Context->getConstantInt(SV->getType(), 1736 ShAmt), "tmp"); 1737 Mask <<= ShAmt; 1738 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1739 SV = Builder.CreateLShr(SV, Context->getConstantInt(SV->getType(), 1740 -ShAmt), "tmp"); 1741 Mask = Mask.lshr(-ShAmt); 1742 } 1743 1744 // Mask out the bits we are about to insert from the old value, and or 1745 // in the new bits. 1746 if (SrcWidth != DestWidth) { 1747 assert(DestWidth > SrcWidth); 1748 Old = Builder.CreateAnd(Old, Context->getConstantInt(~Mask), "mask"); 1749 SV = Builder.CreateOr(Old, SV, "ins"); 1750 } 1751 return SV; 1752} 1753 1754 1755 1756/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1757/// some part of a constant global variable. This intentionally only accepts 1758/// constant expressions because we don't can't rewrite arbitrary instructions. 1759static bool PointsToConstantGlobal(Value *V) { 1760 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1761 return GV->isConstant(); 1762 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1763 if (CE->getOpcode() == Instruction::BitCast || 1764 CE->getOpcode() == Instruction::GetElementPtr) 1765 return PointsToConstantGlobal(CE->getOperand(0)); 1766 return false; 1767} 1768 1769/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1770/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1771/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1772/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1773/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1774/// the alloca, and if the source pointer is a pointer to a constant global, we 1775/// can optimize this. 1776static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1777 bool isOffset) { 1778 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1779 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) 1780 // Ignore non-volatile loads, they are always ok. 1781 if (!LI->isVolatile()) 1782 continue; 1783 1784 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1785 // If uses of the bitcast are ok, we are ok. 1786 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1787 return false; 1788 continue; 1789 } 1790 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1791 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1792 // doesn't, it does. 1793 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1794 isOffset || !GEP->hasAllZeroIndices())) 1795 return false; 1796 continue; 1797 } 1798 1799 // If this is isn't our memcpy/memmove, reject it as something we can't 1800 // handle. 1801 if (!isa<MemTransferInst>(*UI)) 1802 return false; 1803 1804 // If we already have seen a copy, reject the second one. 1805 if (TheCopy) return false; 1806 1807 // If the pointer has been offset from the start of the alloca, we can't 1808 // safely handle this. 1809 if (isOffset) return false; 1810 1811 // If the memintrinsic isn't using the alloca as the dest, reject it. 1812 if (UI.getOperandNo() != 1) return false; 1813 1814 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1815 1816 // If the source of the memcpy/move is not a constant global, reject it. 1817 if (!PointsToConstantGlobal(MI->getOperand(2))) 1818 return false; 1819 1820 // Otherwise, the transform is safe. Remember the copy instruction. 1821 TheCopy = MI; 1822 } 1823 return true; 1824} 1825 1826/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1827/// modified by a copy from a constant global. If we can prove this, we can 1828/// replace any uses of the alloca with uses of the global directly. 1829Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) { 1830 Instruction *TheCopy = 0; 1831 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1832 return TheCopy; 1833 return 0; 1834} 1835