ScalarReplAggregates.cpp revision debcb01b0f0a15f568ca69e8f288fade4bfc7297
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" 44#include "llvm/ADT/StringExtras.h" 45using namespace llvm; 46 47STATISTIC(NumReplaced, "Number of allocas broken up"); 48STATISTIC(NumPromoted, "Number of allocas promoted"); 49STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 50STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 51 52namespace { 53 struct VISIBILITY_HIDDEN SROA : public FunctionPass { 54 static char ID; // Pass identification, replacement for typeid 55 explicit SROA(signed T = -1) : FunctionPass(&ID) { 56 if (T == -1) 57 SRThreshold = 128; 58 else 59 SRThreshold = T; 60 } 61 62 bool runOnFunction(Function &F); 63 64 bool performScalarRepl(Function &F); 65 bool performPromotion(Function &F); 66 67 // getAnalysisUsage - This pass does not require any passes, but we know it 68 // will not alter the CFG, so say so. 69 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 70 AU.addRequired<DominatorTree>(); 71 AU.addRequired<DominanceFrontier>(); 72 AU.addRequired<TargetData>(); 73 AU.setPreservesCFG(); 74 } 75 76 private: 77 TargetData *TD; 78 79 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 80 /// information about the uses. All these fields are initialized to false 81 /// and set to true when something is learned. 82 struct AllocaInfo { 83 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 84 bool isUnsafe : 1; 85 86 /// needsCleanup - This is set to true if there is some use of the alloca 87 /// that requires cleanup. 88 bool needsCleanup : 1; 89 90 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 91 bool isMemCpySrc : 1; 92 93 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 94 bool isMemCpyDst : 1; 95 96 AllocaInfo() 97 : isUnsafe(false), needsCleanup(false), 98 isMemCpySrc(false), isMemCpyDst(false) {} 99 }; 100 101 unsigned SRThreshold; 102 103 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 104 105 int isSafeAllocaToScalarRepl(AllocationInst *AI); 106 107 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 108 AllocaInfo &Info); 109 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 110 AllocaInfo &Info); 111 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 112 unsigned OpNo, AllocaInfo &Info); 113 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI, 114 AllocaInfo &Info); 115 116 void DoScalarReplacement(AllocationInst *AI, 117 std::vector<AllocationInst*> &WorkList); 118 void CleanupGEP(GetElementPtrInst *GEP); 119 void CleanupAllocaUsers(AllocationInst *AI); 120 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base); 121 122 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 123 SmallVector<AllocaInst*, 32> &NewElts); 124 125 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 126 AllocationInst *AI, 127 SmallVector<AllocaInst*, 32> &NewElts); 128 void RewriteStoreUserOfWholeAlloca(StoreInst *SI, AllocationInst *AI, 129 SmallVector<AllocaInst*, 32> &NewElts); 130 void RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 131 SmallVector<AllocaInst*, 32> &NewElts); 132 133 bool CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 134 bool &SawVec, uint64_t Offset, unsigned AllocaSize); 135 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset); 136 Value *ConvertScalar_ExtractValue(Value *NV, const Type *ToType, 137 uint64_t Offset, IRBuilder<> &Builder); 138 Value *ConvertScalar_InsertValue(Value *StoredVal, Value *ExistingVal, 139 uint64_t Offset, IRBuilder<> &Builder); 140 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI); 141 }; 142} 143 144char SROA::ID = 0; 145static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 146 147// Public interface to the ScalarReplAggregates pass 148FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 149 return new SROA(Threshold); 150} 151 152 153bool SROA::runOnFunction(Function &F) { 154 TD = &getAnalysis<TargetData>(); 155 156 bool Changed = performPromotion(F); 157 while (1) { 158 bool LocalChange = performScalarRepl(F); 159 if (!LocalChange) break; // No need to repromote if no scalarrepl 160 Changed = true; 161 LocalChange = performPromotion(F); 162 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 163 } 164 165 return Changed; 166} 167 168 169bool SROA::performPromotion(Function &F) { 170 std::vector<AllocaInst*> Allocas; 171 DominatorTree &DT = getAnalysis<DominatorTree>(); 172 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 173 174 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 175 176 bool Changed = false; 177 178 while (1) { 179 Allocas.clear(); 180 181 // Find allocas that are safe to promote, by looking at all instructions in 182 // the entry node 183 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 184 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 185 if (isAllocaPromotable(AI)) 186 Allocas.push_back(AI); 187 188 if (Allocas.empty()) break; 189 190 PromoteMemToReg(Allocas, DT, DF, F.getContext()); 191 NumPromoted += Allocas.size(); 192 Changed = true; 193 } 194 195 return Changed; 196} 197 198/// getNumSAElements - Return the number of elements in the specific struct or 199/// array. 200static uint64_t getNumSAElements(const Type *T) { 201 if (const StructType *ST = dyn_cast<StructType>(T)) 202 return ST->getNumElements(); 203 return cast<ArrayType>(T)->getNumElements(); 204} 205 206// performScalarRepl - This algorithm is a simple worklist driven algorithm, 207// which runs on all of the malloc/alloca instructions in the function, removing 208// them if they are only used by getelementptr instructions. 209// 210bool SROA::performScalarRepl(Function &F) { 211 std::vector<AllocationInst*> WorkList; 212 213 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 214 BasicBlock &BB = F.getEntryBlock(); 215 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 216 if (AllocationInst *A = dyn_cast<AllocationInst>(I)) 217 WorkList.push_back(A); 218 219 // Process the worklist 220 bool Changed = false; 221 while (!WorkList.empty()) { 222 AllocationInst *AI = WorkList.back(); 223 WorkList.pop_back(); 224 225 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 226 // with unused elements. 227 if (AI->use_empty()) { 228 AI->eraseFromParent(); 229 continue; 230 } 231 232 // If this alloca is impossible for us to promote, reject it early. 233 if (AI->isArrayAllocation() || !AI->getAllocatedType()->isSized()) 234 continue; 235 236 // Check to see if this allocation is only modified by a memcpy/memmove from 237 // a constant global. If this is the case, we can change all users to use 238 // the constant global instead. This is commonly produced by the CFE by 239 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 240 // is only subsequently read. 241 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 242 DOUT << "Found alloca equal to global: " << *AI; 243 DOUT << " memcpy = " << *TheCopy; 244 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 245 AI->replaceAllUsesWith(ConstantExpr::getBitCast(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: llvm_unreachable("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 = IntegerType::get(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 LLVMContext &Context = AI->getContext(); 334 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 335 ElementAllocas.reserve(ST->getNumContainedTypes()); 336 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 337 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 338 AI->getAlignment(), 339 AI->getName() + "." + utostr(i), AI); 340 ElementAllocas.push_back(NA); 341 WorkList.push_back(NA); // Add to worklist for recursive processing 342 } 343 } else { 344 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 345 ElementAllocas.reserve(AT->getNumElements()); 346 const Type *ElTy = AT->getElementType(); 347 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 348 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 349 AI->getName() + "." + utostr(i), AI); 350 ElementAllocas.push_back(NA); 351 WorkList.push_back(NA); // Add to worklist for recursive processing 352 } 353 } 354 355 // Now that we have created the alloca instructions that we want to use, 356 // expand the getelementptr instructions to use them. 357 // 358 while (!AI->use_empty()) { 359 Instruction *User = cast<Instruction>(AI->use_back()); 360 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) { 361 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas); 362 BCInst->eraseFromParent(); 363 continue; 364 } 365 366 // Replace: 367 // %res = load { i32, i32 }* %alloc 368 // with: 369 // %load.0 = load i32* %alloc.0 370 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 371 // %load.1 = load i32* %alloc.1 372 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 373 // (Also works for arrays instead of structs) 374 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 375 Value *Insert = Context.getUndef(LI->getType()); 376 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 377 Value *Load = new LoadInst(ElementAllocas[i], "load", LI); 378 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 379 } 380 LI->replaceAllUsesWith(Insert); 381 LI->eraseFromParent(); 382 continue; 383 } 384 385 // Replace: 386 // store { i32, i32 } %val, { i32, i32 }* %alloc 387 // with: 388 // %val.0 = extractvalue { i32, i32 } %val, 0 389 // store i32 %val.0, i32* %alloc.0 390 // %val.1 = extractvalue { i32, i32 } %val, 1 391 // store i32 %val.1, i32* %alloc.1 392 // (Also works for arrays instead of structs) 393 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 394 Value *Val = SI->getOperand(0); 395 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 396 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 397 new StoreInst(Extract, ElementAllocas[i], SI); 398 } 399 SI->eraseFromParent(); 400 continue; 401 } 402 403 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 404 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 405 unsigned Idx = 406 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 407 408 assert(Idx < ElementAllocas.size() && "Index out of range?"); 409 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 410 411 Value *RepValue; 412 if (GEPI->getNumOperands() == 3) { 413 // Do not insert a new getelementptr instruction with zero indices, only 414 // to have it optimized out later. 415 RepValue = AllocaToUse; 416 } else { 417 // We are indexing deeply into the structure, so we still need a 418 // getelement ptr instruction to finish the indexing. This may be 419 // expanded itself once the worklist is rerun. 420 // 421 SmallVector<Value*, 8> NewArgs; 422 NewArgs.push_back(Context.getNullValue(Type::Int32Ty)); 423 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end()); 424 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(), 425 NewArgs.end(), "", GEPI); 426 RepValue->takeName(GEPI); 427 } 428 429 // If this GEP is to the start of the aggregate, check for memcpys. 430 if (Idx == 0 && GEPI->hasAllZeroIndices()) 431 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas); 432 433 // Move all of the users over to the new GEP. 434 GEPI->replaceAllUsesWith(RepValue); 435 // Delete the old GEP 436 GEPI->eraseFromParent(); 437 } 438 439 // Finally, delete the Alloca instruction 440 AI->eraseFromParent(); 441 NumReplaced++; 442} 443 444 445/// isSafeElementUse - Check to see if this use is an allowed use for a 446/// getelementptr instruction of an array aggregate allocation. isFirstElt 447/// indicates whether Ptr is known to the start of the aggregate. 448/// 449void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 450 AllocaInfo &Info) { 451 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 452 I != E; ++I) { 453 Instruction *User = cast<Instruction>(*I); 454 switch (User->getOpcode()) { 455 case Instruction::Load: break; 456 case Instruction::Store: 457 // Store is ok if storing INTO the pointer, not storing the pointer 458 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info); 459 break; 460 case Instruction::GetElementPtr: { 461 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 462 bool AreAllZeroIndices = isFirstElt; 463 if (GEP->getNumOperands() > 1) { 464 if (!isa<ConstantInt>(GEP->getOperand(1)) || 465 !cast<ConstantInt>(GEP->getOperand(1))->isZero()) 466 // Using pointer arithmetic to navigate the array. 467 return MarkUnsafe(Info); 468 469 if (AreAllZeroIndices) 470 AreAllZeroIndices = GEP->hasAllZeroIndices(); 471 } 472 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 473 if (Info.isUnsafe) return; 474 break; 475 } 476 case Instruction::BitCast: 477 if (isFirstElt) { 478 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 479 if (Info.isUnsafe) return; 480 break; 481 } 482 DOUT << " Transformation preventing inst: " << *User; 483 return MarkUnsafe(Info); 484 case Instruction::Call: 485 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 486 if (isFirstElt) { 487 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 488 if (Info.isUnsafe) return; 489 break; 490 } 491 } 492 DOUT << " Transformation preventing inst: " << *User; 493 return MarkUnsafe(Info); 494 default: 495 DOUT << " Transformation preventing inst: " << *User; 496 return MarkUnsafe(Info); 497 } 498 } 499 return; // All users look ok :) 500} 501 502/// AllUsersAreLoads - Return true if all users of this value are loads. 503static bool AllUsersAreLoads(Value *Ptr) { 504 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 505 I != E; ++I) 506 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 507 return false; 508 return true; 509} 510 511/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 512/// aggregate allocation. 513/// 514void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 515 AllocaInfo &Info) { 516 LLVMContext &Context = User->getContext(); 517 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 518 return isSafeUseOfBitCastedAllocation(C, AI, Info); 519 520 if (LoadInst *LI = dyn_cast<LoadInst>(User)) 521 if (!LI->isVolatile()) 522 return;// Loads (returning a first class aggregrate) are always rewritable 523 524 if (StoreInst *SI = dyn_cast<StoreInst>(User)) 525 if (!SI->isVolatile() && SI->getOperand(0) != AI) 526 return;// Store is ok if storing INTO the pointer, not storing the pointer 527 528 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 529 if (GEPI == 0) 530 return MarkUnsafe(Info); 531 532 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 533 534 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 535 if (I == E || 536 I.getOperand() != Context.getNullValue(I.getOperand()->getType())) { 537 return MarkUnsafe(Info); 538 } 539 540 ++I; 541 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 542 543 bool IsAllZeroIndices = true; 544 545 // If the first index is a non-constant index into an array, see if we can 546 // handle it as a special case. 547 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 548 if (!isa<ConstantInt>(I.getOperand())) { 549 IsAllZeroIndices = 0; 550 uint64_t NumElements = AT->getNumElements(); 551 552 // If this is an array index and the index is not constant, we cannot 553 // promote... that is unless the array has exactly one or two elements in 554 // it, in which case we CAN promote it, but we have to canonicalize this 555 // out if this is the only problem. 556 if ((NumElements == 1 || NumElements == 2) && 557 AllUsersAreLoads(GEPI)) { 558 Info.needsCleanup = true; 559 return; // Canonicalization required! 560 } 561 return MarkUnsafe(Info); 562 } 563 } 564 565 // Walk through the GEP type indices, checking the types that this indexes 566 // into. 567 for (; I != E; ++I) { 568 // Ignore struct elements, no extra checking needed for these. 569 if (isa<StructType>(*I)) 570 continue; 571 572 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 573 if (!IdxVal) return MarkUnsafe(Info); 574 575 // Are all indices still zero? 576 IsAllZeroIndices &= IdxVal->isZero(); 577 578 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 579 // This GEP indexes an array. Verify that this is an in-range constant 580 // integer. Specifically, consider A[0][i]. We cannot know that the user 581 // isn't doing invalid things like allowing i to index an out-of-range 582 // subscript that accesses A[1]. Because of this, we have to reject SROA 583 // of any accesses into structs where any of the components are variables. 584 if (IdxVal->getZExtValue() >= AT->getNumElements()) 585 return MarkUnsafe(Info); 586 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) { 587 if (IdxVal->getZExtValue() >= VT->getNumElements()) 588 return MarkUnsafe(Info); 589 } 590 } 591 592 // If there are any non-simple uses of this getelementptr, make sure to reject 593 // them. 594 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 595} 596 597/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory 598/// intrinsic can be promoted by SROA. At this point, we know that the operand 599/// of the memintrinsic is a pointer to the beginning of the allocation. 600void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 601 unsigned OpNo, AllocaInfo &Info) { 602 // If not constant length, give up. 603 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 604 if (!Length) return MarkUnsafe(Info); 605 606 // If not the whole aggregate, give up. 607 if (Length->getZExtValue() != 608 TD->getTypeAllocSize(AI->getType()->getElementType())) 609 return MarkUnsafe(Info); 610 611 // We only know about memcpy/memset/memmove. 612 if (!isa<MemIntrinsic>(MI)) 613 return MarkUnsafe(Info); 614 615 // Otherwise, we can transform it. Determine whether this is a memcpy/set 616 // into or out of the aggregate. 617 if (OpNo == 1) 618 Info.isMemCpyDst = true; 619 else { 620 assert(OpNo == 2); 621 Info.isMemCpySrc = true; 622 } 623} 624 625/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast 626/// are 627void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI, 628 AllocaInfo &Info) { 629 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 630 UI != E; ++UI) { 631 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 632 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 633 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 634 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 635 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) { 636 if (SI->isVolatile()) 637 return MarkUnsafe(Info); 638 639 // If storing the entire alloca in one chunk through a bitcasted pointer 640 // to integer, we can transform it. This happens (for example) when you 641 // cast a {i32,i32}* to i64* and store through it. This is similar to the 642 // memcpy case and occurs in various "byval" cases and emulated memcpys. 643 if (isa<IntegerType>(SI->getOperand(0)->getType()) && 644 TD->getTypeAllocSize(SI->getOperand(0)->getType()) == 645 TD->getTypeAllocSize(AI->getType()->getElementType())) { 646 Info.isMemCpyDst = true; 647 continue; 648 } 649 return MarkUnsafe(Info); 650 } else if (LoadInst *LI = dyn_cast<LoadInst>(UI)) { 651 if (LI->isVolatile()) 652 return MarkUnsafe(Info); 653 654 // If loading the entire alloca in one chunk through a bitcasted pointer 655 // to integer, we can transform it. This happens (for example) when you 656 // cast a {i32,i32}* to i64* and load through it. This is similar to the 657 // memcpy case and occurs in various "byval" cases and emulated memcpys. 658 if (isa<IntegerType>(LI->getType()) && 659 TD->getTypeAllocSize(LI->getType()) == 660 TD->getTypeAllocSize(AI->getType()->getElementType())) { 661 Info.isMemCpySrc = true; 662 continue; 663 } 664 return MarkUnsafe(Info); 665 } else if (isa<DbgInfoIntrinsic>(UI)) { 666 // If one user is DbgInfoIntrinsic then check if all users are 667 // DbgInfoIntrinsics. 668 if (OnlyUsedByDbgInfoIntrinsics(BC)) { 669 Info.needsCleanup = true; 670 return; 671 } 672 else 673 MarkUnsafe(Info); 674 } 675 else { 676 return MarkUnsafe(Info); 677 } 678 if (Info.isUnsafe) return; 679 } 680} 681 682/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 683/// to its first element. Transform users of the cast to use the new values 684/// instead. 685void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 686 SmallVector<AllocaInst*, 32> &NewElts) { 687 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 688 while (UI != UE) { 689 Instruction *User = cast<Instruction>(*UI++); 690 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) { 691 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 692 if (BCU->use_empty()) BCU->eraseFromParent(); 693 continue; 694 } 695 696 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 697 // This must be memcpy/memmove/memset of the entire aggregate. 698 // Split into one per element. 699 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts); 700 continue; 701 } 702 703 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 704 // If this is a store of the entire alloca from an integer, rewrite it. 705 RewriteStoreUserOfWholeAlloca(SI, AI, NewElts); 706 continue; 707 } 708 709 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 710 // If this is a load of the entire alloca to an integer, rewrite it. 711 RewriteLoadUserOfWholeAlloca(LI, AI, NewElts); 712 continue; 713 } 714 715 // Otherwise it must be some other user of a gep of the first pointer. Just 716 // leave these alone. 717 continue; 718 } 719} 720 721/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 722/// Rewrite it to copy or set the elements of the scalarized memory. 723void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 724 AllocationInst *AI, 725 SmallVector<AllocaInst*, 32> &NewElts) { 726 727 // If this is a memcpy/memmove, construct the other pointer as the 728 // appropriate type. The "Other" pointer is the pointer that goes to memory 729 // that doesn't have anything to do with the alloca that we are promoting. For 730 // memset, this Value* stays null. 731 Value *OtherPtr = 0; 732 LLVMContext &Context = MI->getContext(); 733 unsigned MemAlignment = MI->getAlignment(); 734 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { // memmove/memcopy 735 if (BCInst == MTI->getRawDest()) 736 OtherPtr = MTI->getRawSource(); 737 else { 738 assert(BCInst == MTI->getRawSource()); 739 OtherPtr = MTI->getRawDest(); 740 } 741 } 742 743 // If there is an other pointer, we want to convert it to the same pointer 744 // type as AI has, so we can GEP through it safely. 745 if (OtherPtr) { 746 // It is likely that OtherPtr is a bitcast, if so, remove it. 747 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 748 OtherPtr = BC->getOperand(0); 749 // All zero GEPs are effectively bitcasts. 750 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) 751 if (GEP->hasAllZeroIndices()) 752 OtherPtr = GEP->getOperand(0); 753 754 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 755 if (BCE->getOpcode() == Instruction::BitCast) 756 OtherPtr = BCE->getOperand(0); 757 758 // If the pointer is not the right type, insert a bitcast to the right 759 // type. 760 if (OtherPtr->getType() != AI->getType()) 761 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 762 MI); 763 } 764 765 // Process each element of the aggregate. 766 Value *TheFn = MI->getOperand(0); 767 const Type *BytePtrTy = MI->getRawDest()->getType(); 768 bool SROADest = MI->getRawDest() == BCInst; 769 770 Constant *Zero = Context.getNullValue(Type::Int32Ty); 771 772 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 773 // If this is a memcpy/memmove, emit a GEP of the other element address. 774 Value *OtherElt = 0; 775 unsigned OtherEltAlign = MemAlignment; 776 777 if (OtherPtr) { 778 Value *Idx[2] = { Zero, ConstantInt::get(Type::Int32Ty, i) }; 779 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2, 780 OtherPtr->getNameStr()+"."+utostr(i), 781 MI); 782 uint64_t EltOffset; 783 const PointerType *OtherPtrTy = cast<PointerType>(OtherPtr->getType()); 784 if (const StructType *ST = 785 dyn_cast<StructType>(OtherPtrTy->getElementType())) { 786 EltOffset = TD->getStructLayout(ST)->getElementOffset(i); 787 } else { 788 const Type *EltTy = 789 cast<SequentialType>(OtherPtr->getType())->getElementType(); 790 EltOffset = TD->getTypeAllocSize(EltTy)*i; 791 } 792 793 // The alignment of the other pointer is the guaranteed alignment of the 794 // element, which is affected by both the known alignment of the whole 795 // mem intrinsic and the alignment of the element. If the alignment of 796 // the memcpy (f.e.) is 32 but the element is at a 4-byte offset, then the 797 // known alignment is just 4 bytes. 798 OtherEltAlign = (unsigned)MinAlign(OtherEltAlign, EltOffset); 799 } 800 801 Value *EltPtr = NewElts[i]; 802 const Type *EltTy = cast<PointerType>(EltPtr->getType())->getElementType(); 803 804 // If we got down to a scalar, insert a load or store as appropriate. 805 if (EltTy->isSingleValueType()) { 806 if (isa<MemTransferInst>(MI)) { 807 if (SROADest) { 808 // From Other to Alloca. 809 Value *Elt = new LoadInst(OtherElt, "tmp", false, OtherEltAlign, MI); 810 new StoreInst(Elt, EltPtr, MI); 811 } else { 812 // From Alloca to Other. 813 Value *Elt = new LoadInst(EltPtr, "tmp", MI); 814 new StoreInst(Elt, OtherElt, false, OtherEltAlign, MI); 815 } 816 continue; 817 } 818 assert(isa<MemSetInst>(MI)); 819 820 // If the stored element is zero (common case), just store a null 821 // constant. 822 Constant *StoreVal; 823 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 824 if (CI->isZero()) { 825 StoreVal = Context.getNullValue(EltTy); // 0.0, null, 0, <0,0> 826 } else { 827 // If EltTy is a vector type, get the element type. 828 const Type *ValTy = EltTy->getScalarType(); 829 830 // Construct an integer with the right value. 831 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 832 APInt OneVal(EltSize, CI->getZExtValue()); 833 APInt TotalVal(OneVal); 834 // Set each byte. 835 for (unsigned i = 0; 8*i < EltSize; ++i) { 836 TotalVal = TotalVal.shl(8); 837 TotalVal |= OneVal; 838 } 839 840 // Convert the integer value to the appropriate type. 841 StoreVal = ConstantInt::get(Context, TotalVal); 842 if (isa<PointerType>(ValTy)) 843 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 844 else if (ValTy->isFloatingPoint()) 845 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 846 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 847 848 // If the requested value was a vector constant, create it. 849 if (EltTy != ValTy) { 850 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 851 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 852 StoreVal = ConstantVector::get(&Elts[0], NumElts); 853 } 854 } 855 new StoreInst(StoreVal, EltPtr, MI); 856 continue; 857 } 858 // Otherwise, if we're storing a byte variable, use a memset call for 859 // this element. 860 } 861 862 // Cast the element pointer to BytePtrTy. 863 if (EltPtr->getType() != BytePtrTy) 864 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 865 866 // Cast the other pointer (if we have one) to BytePtrTy. 867 if (OtherElt && OtherElt->getType() != BytePtrTy) 868 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 869 MI); 870 871 unsigned EltSize = TD->getTypeAllocSize(EltTy); 872 873 // Finally, insert the meminst for this element. 874 if (isa<MemTransferInst>(MI)) { 875 Value *Ops[] = { 876 SROADest ? EltPtr : OtherElt, // Dest ptr 877 SROADest ? OtherElt : EltPtr, // Src ptr 878 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 879 ConstantInt::get(Type::Int32Ty, OtherEltAlign) // Align 880 }; 881 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 882 } else { 883 assert(isa<MemSetInst>(MI)); 884 Value *Ops[] = { 885 EltPtr, MI->getOperand(2), // Dest, Value, 886 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 887 Zero // Align 888 }; 889 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 890 } 891 } 892 MI->eraseFromParent(); 893} 894 895/// RewriteStoreUserOfWholeAlloca - We found an store of an integer that 896/// overwrites the entire allocation. Extract out the pieces of the stored 897/// integer and store them individually. 898void SROA::RewriteStoreUserOfWholeAlloca(StoreInst *SI, 899 AllocationInst *AI, 900 SmallVector<AllocaInst*, 32> &NewElts){ 901 // Extract each element out of the integer according to its structure offset 902 // and store the element value to the individual alloca. 903 Value *SrcVal = SI->getOperand(0); 904 const Type *AllocaEltTy = AI->getType()->getElementType(); 905 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 906 907 // If this isn't a store of an integer to the whole alloca, it may be a store 908 // to the first element. Just ignore the store in this case and normal SROA 909 // will handle it. 910 if (!isa<IntegerType>(SrcVal->getType()) || 911 TD->getTypeAllocSizeInBits(SrcVal->getType()) != AllocaSizeBits) 912 return; 913 // Handle tail padding by extending the operand 914 if (TD->getTypeSizeInBits(SrcVal->getType()) != AllocaSizeBits) 915 SrcVal = new ZExtInst(SrcVal, 916 IntegerType::get(AllocaSizeBits), "", SI); 917 918 DOUT << "PROMOTING STORE TO WHOLE ALLOCA: " << *AI << *SI; 919 920 // There are two forms here: AI could be an array or struct. Both cases 921 // have different ways to compute the element offset. 922 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 923 const StructLayout *Layout = TD->getStructLayout(EltSTy); 924 925 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 926 // Get the number of bits to shift SrcVal to get the value. 927 const Type *FieldTy = EltSTy->getElementType(i); 928 uint64_t Shift = Layout->getElementOffsetInBits(i); 929 930 if (TD->isBigEndian()) 931 Shift = AllocaSizeBits-Shift-TD->getTypeAllocSizeInBits(FieldTy); 932 933 Value *EltVal = SrcVal; 934 if (Shift) { 935 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 936 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 937 "sroa.store.elt", SI); 938 } 939 940 // Truncate down to an integer of the right size. 941 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 942 943 // Ignore zero sized fields like {}, they obviously contain no data. 944 if (FieldSizeBits == 0) continue; 945 946 if (FieldSizeBits != AllocaSizeBits) 947 EltVal = new TruncInst(EltVal, 948 IntegerType::get(FieldSizeBits), "", SI); 949 Value *DestField = NewElts[i]; 950 if (EltVal->getType() == FieldTy) { 951 // Storing to an integer field of this size, just do it. 952 } else if (FieldTy->isFloatingPoint() || isa<VectorType>(FieldTy)) { 953 // Bitcast to the right element type (for fp/vector values). 954 EltVal = new BitCastInst(EltVal, FieldTy, "", SI); 955 } else { 956 // Otherwise, bitcast the dest pointer (for aggregates). 957 DestField = new BitCastInst(DestField, 958 PointerType::getUnqual(EltVal->getType()), 959 "", SI); 960 } 961 new StoreInst(EltVal, DestField, SI); 962 } 963 964 } else { 965 const ArrayType *ATy = cast<ArrayType>(AllocaEltTy); 966 const Type *ArrayEltTy = ATy->getElementType(); 967 uint64_t ElementOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 968 uint64_t ElementSizeBits = TD->getTypeSizeInBits(ArrayEltTy); 969 970 uint64_t Shift; 971 972 if (TD->isBigEndian()) 973 Shift = AllocaSizeBits-ElementOffset; 974 else 975 Shift = 0; 976 977 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 978 // Ignore zero sized fields like {}, they obviously contain no data. 979 if (ElementSizeBits == 0) continue; 980 981 Value *EltVal = SrcVal; 982 if (Shift) { 983 Value *ShiftVal = ConstantInt::get(EltVal->getType(), Shift); 984 EltVal = BinaryOperator::CreateLShr(EltVal, ShiftVal, 985 "sroa.store.elt", SI); 986 } 987 988 // Truncate down to an integer of the right size. 989 if (ElementSizeBits != AllocaSizeBits) 990 EltVal = new TruncInst(EltVal, 991 IntegerType::get(ElementSizeBits),"",SI); 992 Value *DestField = NewElts[i]; 993 if (EltVal->getType() == ArrayEltTy) { 994 // Storing to an integer field of this size, just do it. 995 } else if (ArrayEltTy->isFloatingPoint() || isa<VectorType>(ArrayEltTy)) { 996 // Bitcast to the right element type (for fp/vector values). 997 EltVal = new BitCastInst(EltVal, ArrayEltTy, "", SI); 998 } else { 999 // Otherwise, bitcast the dest pointer (for aggregates). 1000 DestField = new BitCastInst(DestField, 1001 PointerType::getUnqual(EltVal->getType()), 1002 "", SI); 1003 } 1004 new StoreInst(EltVal, DestField, SI); 1005 1006 if (TD->isBigEndian()) 1007 Shift -= ElementOffset; 1008 else 1009 Shift += ElementOffset; 1010 } 1011 } 1012 1013 SI->eraseFromParent(); 1014} 1015 1016/// RewriteLoadUserOfWholeAlloca - We found an load of the entire allocation to 1017/// an integer. Load the individual pieces to form the aggregate value. 1018void SROA::RewriteLoadUserOfWholeAlloca(LoadInst *LI, AllocationInst *AI, 1019 SmallVector<AllocaInst*, 32> &NewElts) { 1020 // Extract each element out of the NewElts according to its structure offset 1021 // and form the result value. 1022 const Type *AllocaEltTy = AI->getType()->getElementType(); 1023 uint64_t AllocaSizeBits = TD->getTypeAllocSizeInBits(AllocaEltTy); 1024 1025 // If this isn't a load of the whole alloca to an integer, it may be a load 1026 // of the first element. Just ignore the load in this case and normal SROA 1027 // will handle it. 1028 if (!isa<IntegerType>(LI->getType()) || 1029 TD->getTypeAllocSizeInBits(LI->getType()) != AllocaSizeBits) 1030 return; 1031 1032 DOUT << "PROMOTING LOAD OF WHOLE ALLOCA: " << *AI << *LI; 1033 1034 // There are two forms here: AI could be an array or struct. Both cases 1035 // have different ways to compute the element offset. 1036 const StructLayout *Layout = 0; 1037 uint64_t ArrayEltBitOffset = 0; 1038 if (const StructType *EltSTy = dyn_cast<StructType>(AllocaEltTy)) { 1039 Layout = TD->getStructLayout(EltSTy); 1040 } else { 1041 const Type *ArrayEltTy = cast<ArrayType>(AllocaEltTy)->getElementType(); 1042 ArrayEltBitOffset = TD->getTypeAllocSizeInBits(ArrayEltTy); 1043 } 1044 1045 LLVMContext &Context = LI->getContext(); 1046 1047 Value *ResultVal = 1048 Context.getNullValue(IntegerType::get(AllocaSizeBits)); 1049 1050 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 1051 // Load the value from the alloca. If the NewElt is an aggregate, cast 1052 // the pointer to an integer of the same size before doing the load. 1053 Value *SrcField = NewElts[i]; 1054 const Type *FieldTy = 1055 cast<PointerType>(SrcField->getType())->getElementType(); 1056 uint64_t FieldSizeBits = TD->getTypeSizeInBits(FieldTy); 1057 1058 // Ignore zero sized fields like {}, they obviously contain no data. 1059 if (FieldSizeBits == 0) continue; 1060 1061 const IntegerType *FieldIntTy = IntegerType::get(FieldSizeBits); 1062 if (!isa<IntegerType>(FieldTy) && !FieldTy->isFloatingPoint() && 1063 !isa<VectorType>(FieldTy)) 1064 SrcField = new BitCastInst(SrcField, 1065 PointerType::getUnqual(FieldIntTy), 1066 "", LI); 1067 SrcField = new LoadInst(SrcField, "sroa.load.elt", LI); 1068 1069 // If SrcField is a fp or vector of the right size but that isn't an 1070 // integer type, bitcast to an integer so we can shift it. 1071 if (SrcField->getType() != FieldIntTy) 1072 SrcField = new BitCastInst(SrcField, FieldIntTy, "", LI); 1073 1074 // Zero extend the field to be the same size as the final alloca so that 1075 // we can shift and insert it. 1076 if (SrcField->getType() != ResultVal->getType()) 1077 SrcField = new ZExtInst(SrcField, ResultVal->getType(), "", LI); 1078 1079 // Determine the number of bits to shift SrcField. 1080 uint64_t Shift; 1081 if (Layout) // Struct case. 1082 Shift = Layout->getElementOffsetInBits(i); 1083 else // Array case. 1084 Shift = i*ArrayEltBitOffset; 1085 1086 if (TD->isBigEndian()) 1087 Shift = AllocaSizeBits-Shift-FieldIntTy->getBitWidth(); 1088 1089 if (Shift) { 1090 Value *ShiftVal = ConstantInt::get(SrcField->getType(), Shift); 1091 SrcField = BinaryOperator::CreateShl(SrcField, ShiftVal, "", LI); 1092 } 1093 1094 ResultVal = BinaryOperator::CreateOr(SrcField, ResultVal, "", LI); 1095 } 1096 1097 // Handle tail padding by truncating the result 1098 if (TD->getTypeSizeInBits(LI->getType()) != AllocaSizeBits) 1099 ResultVal = new TruncInst(ResultVal, LI->getType(), "", LI); 1100 1101 LI->replaceAllUsesWith(ResultVal); 1102 LI->eraseFromParent(); 1103} 1104 1105 1106/// HasPadding - Return true if the specified type has any structure or 1107/// alignment padding, false otherwise. 1108static bool HasPadding(const Type *Ty, const TargetData &TD) { 1109 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 1110 const StructLayout *SL = TD.getStructLayout(STy); 1111 unsigned PrevFieldBitOffset = 0; 1112 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1113 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 1114 1115 // Padding in sub-elements? 1116 if (HasPadding(STy->getElementType(i), TD)) 1117 return true; 1118 1119 // Check to see if there is any padding between this element and the 1120 // previous one. 1121 if (i) { 1122 unsigned PrevFieldEnd = 1123 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 1124 if (PrevFieldEnd < FieldBitOffset) 1125 return true; 1126 } 1127 1128 PrevFieldBitOffset = FieldBitOffset; 1129 } 1130 1131 // Check for tail padding. 1132 if (unsigned EltCount = STy->getNumElements()) { 1133 unsigned PrevFieldEnd = PrevFieldBitOffset + 1134 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 1135 if (PrevFieldEnd < SL->getSizeInBits()) 1136 return true; 1137 } 1138 1139 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 1140 return HasPadding(ATy->getElementType(), TD); 1141 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 1142 return HasPadding(VTy->getElementType(), TD); 1143 } 1144 return TD.getTypeSizeInBits(Ty) != TD.getTypeAllocSizeInBits(Ty); 1145} 1146 1147/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 1148/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 1149/// or 1 if safe after canonicalization has been performed. 1150/// 1151int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 1152 // Loop over the use list of the alloca. We can only transform it if all of 1153 // the users are safe to transform. 1154 AllocaInfo Info; 1155 1156 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 1157 I != E; ++I) { 1158 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 1159 if (Info.isUnsafe) { 1160 DOUT << "Cannot transform: " << *AI << " due to user: " << **I; 1161 return 0; 1162 } 1163 } 1164 1165 // Okay, we know all the users are promotable. If the aggregate is a memcpy 1166 // source and destination, we have to be careful. In particular, the memcpy 1167 // could be moving around elements that live in structure padding of the LLVM 1168 // types, but may actually be used. In these cases, we refuse to promote the 1169 // struct. 1170 if (Info.isMemCpySrc && Info.isMemCpyDst && 1171 HasPadding(AI->getType()->getElementType(), *TD)) 1172 return 0; 1173 1174 // If we require cleanup, return 1, otherwise return 3. 1175 return Info.needsCleanup ? 1 : 3; 1176} 1177 1178/// CleanupGEP - GEP is used by an Alloca, which can be prompted after the GEP 1179/// is canonicalized here. 1180void SROA::CleanupGEP(GetElementPtrInst *GEPI) { 1181 gep_type_iterator I = gep_type_begin(GEPI); 1182 ++I; 1183 1184 const ArrayType *AT = dyn_cast<ArrayType>(*I); 1185 if (!AT) 1186 return; 1187 1188 uint64_t NumElements = AT->getNumElements(); 1189 1190 if (isa<ConstantInt>(I.getOperand())) 1191 return; 1192 1193 LLVMContext &Context = GEPI->getContext(); 1194 1195 if (NumElements == 1) { 1196 GEPI->setOperand(2, Context.getNullValue(Type::Int32Ty)); 1197 return; 1198 } 1199 1200 assert(NumElements == 2 && "Unhandled case!"); 1201 // All users of the GEP must be loads. At each use of the GEP, insert 1202 // two loads of the appropriate indexed GEP and select between them. 1203 Value *IsOne = new ICmpInst(GEPI, ICmpInst::ICMP_NE, I.getOperand(), 1204 Context.getNullValue(I.getOperand()->getType()), 1205 "isone"); 1206 // Insert the new GEP instructions, which are properly indexed. 1207 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 1208 Indices[1] = Context.getNullValue(Type::Int32Ty); 1209 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1210 Indices.begin(), 1211 Indices.end(), 1212 GEPI->getName()+".0", GEPI); 1213 Indices[1] = ConstantInt::get(Type::Int32Ty, 1); 1214 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 1215 Indices.begin(), 1216 Indices.end(), 1217 GEPI->getName()+".1", GEPI); 1218 // Replace all loads of the variable index GEP with loads from both 1219 // indexes and a select. 1220 while (!GEPI->use_empty()) { 1221 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 1222 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 1223 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 1224 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 1225 LI->replaceAllUsesWith(R); 1226 LI->eraseFromParent(); 1227 } 1228 GEPI->eraseFromParent(); 1229} 1230 1231 1232/// CleanupAllocaUsers - If SROA reported that it can promote the specified 1233/// allocation, but only if cleaned up, perform the cleanups required. 1234void SROA::CleanupAllocaUsers(AllocationInst *AI) { 1235 // At this point, we know that the end result will be SROA'd and promoted, so 1236 // we can insert ugly code if required so long as sroa+mem2reg will clean it 1237 // up. 1238 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 1239 UI != E; ) { 1240 User *U = *UI++; 1241 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) 1242 CleanupGEP(GEPI); 1243 else { 1244 Instruction *I = cast<Instruction>(U); 1245 SmallVector<DbgInfoIntrinsic *, 2> DbgInUses; 1246 if (!isa<StoreInst>(I) && OnlyUsedByDbgInfoIntrinsics(I, &DbgInUses)) { 1247 // Safe to remove debug info uses. 1248 while (!DbgInUses.empty()) { 1249 DbgInfoIntrinsic *DI = DbgInUses.back(); DbgInUses.pop_back(); 1250 DI->eraseFromParent(); 1251 } 1252 I->eraseFromParent(); 1253 } 1254 } 1255 } 1256} 1257 1258/// MergeInType - Add the 'In' type to the accumulated type (Accum) so far at 1259/// the offset specified by Offset (which is specified in bytes). 1260/// 1261/// There are two cases we handle here: 1262/// 1) A union of vector types of the same size and potentially its elements. 1263/// Here we turn element accesses into insert/extract element operations. 1264/// This promotes a <4 x float> with a store of float to the third element 1265/// into a <4 x float> that uses insert element. 1266/// 2) A fully general blob of memory, which we turn into some (potentially 1267/// large) integer type with extract and insert operations where the loads 1268/// and stores would mutate the memory. 1269static void MergeInType(const Type *In, uint64_t Offset, const Type *&VecTy, 1270 unsigned AllocaSize, const TargetData &TD, 1271 LLVMContext &Context) { 1272 // If this could be contributing to a vector, analyze it. 1273 if (VecTy != Type::VoidTy) { // either null or a vector type. 1274 1275 // If the In type is a vector that is the same size as the alloca, see if it 1276 // matches the existing VecTy. 1277 if (const VectorType *VInTy = dyn_cast<VectorType>(In)) { 1278 if (VInTy->getBitWidth()/8 == AllocaSize && Offset == 0) { 1279 // If we're storing/loading a vector of the right size, allow it as a 1280 // vector. If this the first vector we see, remember the type so that 1281 // we know the element size. 1282 if (VecTy == 0) 1283 VecTy = VInTy; 1284 return; 1285 } 1286 } else if (In == Type::FloatTy || In == Type::DoubleTy || 1287 (isa<IntegerType>(In) && In->getPrimitiveSizeInBits() >= 8 && 1288 isPowerOf2_32(In->getPrimitiveSizeInBits()))) { 1289 // If we're accessing something that could be an element of a vector, see 1290 // if the implied vector agrees with what we already have and if Offset is 1291 // compatible with it. 1292 unsigned EltSize = In->getPrimitiveSizeInBits()/8; 1293 if (Offset % EltSize == 0 && 1294 AllocaSize % EltSize == 0 && 1295 (VecTy == 0 || 1296 cast<VectorType>(VecTy)->getElementType() 1297 ->getPrimitiveSizeInBits()/8 == EltSize)) { 1298 if (VecTy == 0) 1299 VecTy = VectorType::get(In, AllocaSize/EltSize); 1300 return; 1301 } 1302 } 1303 } 1304 1305 // Otherwise, we have a case that we can't handle with an optimized vector 1306 // form. We can still turn this into a large integer. 1307 VecTy = Type::VoidTy; 1308} 1309 1310/// CanConvertToScalar - V is a pointer. If we can convert the pointee and all 1311/// its accesses to use a to single vector type, return true, and set VecTy to 1312/// the new type. If we could convert the alloca into a single promotable 1313/// integer, return true but set VecTy to VoidTy. Further, if the use is not a 1314/// completely trivial use that mem2reg could promote, set IsNotTrivial. Offset 1315/// is the current offset from the base of the alloca being analyzed. 1316/// 1317/// If we see at least one access to the value that is as a vector type, set the 1318/// SawVec flag. 1319/// 1320bool SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial, const Type *&VecTy, 1321 bool &SawVec, uint64_t Offset, 1322 unsigned AllocaSize) { 1323 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1324 Instruction *User = cast<Instruction>(*UI); 1325 1326 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1327 // Don't break volatile loads. 1328 if (LI->isVolatile()) 1329 return false; 1330 MergeInType(LI->getType(), Offset, VecTy, 1331 AllocaSize, *TD, V->getContext()); 1332 SawVec |= isa<VectorType>(LI->getType()); 1333 continue; 1334 } 1335 1336 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1337 // Storing the pointer, not into the value? 1338 if (SI->getOperand(0) == V || SI->isVolatile()) return 0; 1339 MergeInType(SI->getOperand(0)->getType(), Offset, 1340 VecTy, AllocaSize, *TD, V->getContext()); 1341 SawVec |= isa<VectorType>(SI->getOperand(0)->getType()); 1342 continue; 1343 } 1344 1345 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) { 1346 if (!CanConvertToScalar(BCI, IsNotTrivial, VecTy, SawVec, Offset, 1347 AllocaSize)) 1348 return false; 1349 IsNotTrivial = true; 1350 continue; 1351 } 1352 1353 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1354 // If this is a GEP with a variable indices, we can't handle it. 1355 if (!GEP->hasAllConstantIndices()) 1356 return false; 1357 1358 // Compute the offset that this GEP adds to the pointer. 1359 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1360 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1361 &Indices[0], Indices.size()); 1362 // See if all uses can be converted. 1363 if (!CanConvertToScalar(GEP, IsNotTrivial, VecTy, SawVec,Offset+GEPOffset, 1364 AllocaSize)) 1365 return false; 1366 IsNotTrivial = true; 1367 continue; 1368 } 1369 1370 // If this is a constant sized memset of a constant value (e.g. 0) we can 1371 // handle it. 1372 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1373 // Store of constant value and constant size. 1374 if (isa<ConstantInt>(MSI->getValue()) && 1375 isa<ConstantInt>(MSI->getLength())) { 1376 IsNotTrivial = true; 1377 continue; 1378 } 1379 } 1380 1381 // If this is a memcpy or memmove into or out of the whole allocation, we 1382 // can handle it like a load or store of the scalar type. 1383 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1384 if (ConstantInt *Len = dyn_cast<ConstantInt>(MTI->getLength())) 1385 if (Len->getZExtValue() == AllocaSize && Offset == 0) { 1386 IsNotTrivial = true; 1387 continue; 1388 } 1389 } 1390 1391 // Ignore dbg intrinsic. 1392 if (isa<DbgInfoIntrinsic>(User)) 1393 continue; 1394 1395 // Otherwise, we cannot handle this! 1396 return false; 1397 } 1398 1399 return true; 1400} 1401 1402 1403/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1404/// directly. This happens when we are converting an "integer union" to a 1405/// single integer scalar, or when we are converting a "vector union" to a 1406/// vector with insert/extractelement instructions. 1407/// 1408/// Offset is an offset from the original alloca, in bits that need to be 1409/// shifted to the right. By the end of this, there should be no uses of Ptr. 1410void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, uint64_t Offset) { 1411 while (!Ptr->use_empty()) { 1412 Instruction *User = cast<Instruction>(Ptr->use_back()); 1413 1414 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1415 ConvertUsesToScalar(CI, NewAI, Offset); 1416 CI->eraseFromParent(); 1417 continue; 1418 } 1419 1420 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1421 // Compute the offset that this GEP adds to the pointer. 1422 SmallVector<Value*, 8> Indices(GEP->op_begin()+1, GEP->op_end()); 1423 uint64_t GEPOffset = TD->getIndexedOffset(GEP->getOperand(0)->getType(), 1424 &Indices[0], Indices.size()); 1425 ConvertUsesToScalar(GEP, NewAI, Offset+GEPOffset*8); 1426 GEP->eraseFromParent(); 1427 continue; 1428 } 1429 1430 IRBuilder<> Builder(User->getParent(), User); 1431 1432 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1433 // The load is a bit extract from NewAI shifted right by Offset bits. 1434 Value *LoadedVal = Builder.CreateLoad(NewAI, "tmp"); 1435 Value *NewLoadVal 1436 = ConvertScalar_ExtractValue(LoadedVal, LI->getType(), Offset, Builder); 1437 LI->replaceAllUsesWith(NewLoadVal); 1438 LI->eraseFromParent(); 1439 continue; 1440 } 1441 1442 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1443 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1444 // FIXME: Remove once builder has Twine API. 1445 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").str().c_str()); 1446 Value *New = ConvertScalar_InsertValue(SI->getOperand(0), Old, Offset, 1447 Builder); 1448 Builder.CreateStore(New, NewAI); 1449 SI->eraseFromParent(); 1450 continue; 1451 } 1452 1453 // If this is a constant sized memset of a constant value (e.g. 0) we can 1454 // transform it into a store of the expanded constant value. 1455 if (MemSetInst *MSI = dyn_cast<MemSetInst>(User)) { 1456 assert(MSI->getRawDest() == Ptr && "Consistency error!"); 1457 unsigned NumBytes = cast<ConstantInt>(MSI->getLength())->getZExtValue(); 1458 if (NumBytes != 0) { 1459 unsigned Val = cast<ConstantInt>(MSI->getValue())->getZExtValue(); 1460 1461 // Compute the value replicated the right number of times. 1462 APInt APVal(NumBytes*8, Val); 1463 1464 // Splat the value if non-zero. 1465 if (Val) 1466 for (unsigned i = 1; i != NumBytes; ++i) 1467 APVal |= APVal << 8; 1468 1469 // FIXME: Remove once builder has Twine API. 1470 Value *Old = Builder.CreateLoad(NewAI, (NewAI->getName()+".in").str().c_str()); 1471 Value *New = ConvertScalar_InsertValue( 1472 ConstantInt::get(User->getContext(), APVal), 1473 Old, Offset, Builder); 1474 Builder.CreateStore(New, NewAI); 1475 } 1476 MSI->eraseFromParent(); 1477 continue; 1478 } 1479 1480 // If this is a memcpy or memmove into or out of the whole allocation, we 1481 // can handle it like a load or store of the scalar type. 1482 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(User)) { 1483 assert(Offset == 0 && "must be store to start of alloca"); 1484 1485 // If the source and destination are both to the same alloca, then this is 1486 // a noop copy-to-self, just delete it. Otherwise, emit a load and store 1487 // as appropriate. 1488 AllocaInst *OrigAI = cast<AllocaInst>(Ptr->getUnderlyingObject()); 1489 1490 if (MTI->getSource()->getUnderlyingObject() != OrigAI) { 1491 // Dest must be OrigAI, change this to be a load from the original 1492 // pointer (bitcasted), then a store to our new alloca. 1493 assert(MTI->getRawDest() == Ptr && "Neither use is of pointer?"); 1494 Value *SrcPtr = MTI->getSource(); 1495 SrcPtr = Builder.CreateBitCast(SrcPtr, NewAI->getType()); 1496 1497 LoadInst *SrcVal = Builder.CreateLoad(SrcPtr, "srcval"); 1498 SrcVal->setAlignment(MTI->getAlignment()); 1499 Builder.CreateStore(SrcVal, NewAI); 1500 } else if (MTI->getDest()->getUnderlyingObject() != OrigAI) { 1501 // Src must be OrigAI, change this to be a load from NewAI then a store 1502 // through the original dest pointer (bitcasted). 1503 assert(MTI->getRawSource() == Ptr && "Neither use is of pointer?"); 1504 LoadInst *SrcVal = Builder.CreateLoad(NewAI, "srcval"); 1505 1506 Value *DstPtr = Builder.CreateBitCast(MTI->getDest(), NewAI->getType()); 1507 StoreInst *NewStore = Builder.CreateStore(SrcVal, DstPtr); 1508 NewStore->setAlignment(MTI->getAlignment()); 1509 } else { 1510 // Noop transfer. Src == Dst 1511 } 1512 1513 1514 MTI->eraseFromParent(); 1515 continue; 1516 } 1517 1518 // If user is a dbg info intrinsic then it is safe to remove it. 1519 if (isa<DbgInfoIntrinsic>(User)) { 1520 User->eraseFromParent(); 1521 continue; 1522 } 1523 1524 llvm_unreachable("Unsupported operation!"); 1525 } 1526} 1527 1528/// ConvertScalar_ExtractValue - Extract a value of type ToType from an integer 1529/// or vector value FromVal, extracting the bits from the offset specified by 1530/// Offset. This returns the value, which is of type ToType. 1531/// 1532/// This happens when we are converting an "integer union" to a single 1533/// integer scalar, or when we are converting a "vector union" to a vector with 1534/// insert/extractelement instructions. 1535/// 1536/// Offset is an offset from the original alloca, in bits that need to be 1537/// shifted to the right. 1538Value *SROA::ConvertScalar_ExtractValue(Value *FromVal, const Type *ToType, 1539 uint64_t Offset, IRBuilder<> &Builder) { 1540 // If the load is of the whole new alloca, no conversion is needed. 1541 if (FromVal->getType() == ToType && Offset == 0) 1542 return FromVal; 1543 1544 LLVMContext &Context = FromVal->getContext(); 1545 1546 // If the result alloca is a vector type, this is either an element 1547 // access or a bitcast to another vector type of the same size. 1548 if (const VectorType *VTy = dyn_cast<VectorType>(FromVal->getType())) { 1549 if (isa<VectorType>(ToType)) 1550 return Builder.CreateBitCast(FromVal, ToType, "tmp"); 1551 1552 // Otherwise it must be an element access. 1553 unsigned Elt = 0; 1554 if (Offset) { 1555 unsigned EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1556 Elt = Offset/EltSize; 1557 assert(EltSize*Elt == Offset && "Invalid modulus in validity checking"); 1558 } 1559 // Return the element extracted out of it. 1560 Value *V = Builder.CreateExtractElement(FromVal, 1561 ConstantInt::get(Type::Int32Ty,Elt), 1562 "tmp"); 1563 if (V->getType() != ToType) 1564 V = Builder.CreateBitCast(V, ToType, "tmp"); 1565 return V; 1566 } 1567 1568 // If ToType is a first class aggregate, extract out each of the pieces and 1569 // use insertvalue's to form the FCA. 1570 if (const StructType *ST = dyn_cast<StructType>(ToType)) { 1571 const StructLayout &Layout = *TD->getStructLayout(ST); 1572 Value *Res = Context.getUndef(ST); 1573 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1574 Value *Elt = ConvertScalar_ExtractValue(FromVal, ST->getElementType(i), 1575 Offset+Layout.getElementOffsetInBits(i), 1576 Builder); 1577 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1578 } 1579 return Res; 1580 } 1581 1582 if (const ArrayType *AT = dyn_cast<ArrayType>(ToType)) { 1583 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1584 Value *Res = Context.getUndef(AT); 1585 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1586 Value *Elt = ConvertScalar_ExtractValue(FromVal, AT->getElementType(), 1587 Offset+i*EltSize, Builder); 1588 Res = Builder.CreateInsertValue(Res, Elt, i, "tmp"); 1589 } 1590 return Res; 1591 } 1592 1593 // Otherwise, this must be a union that was converted to an integer value. 1594 const IntegerType *NTy = cast<IntegerType>(FromVal->getType()); 1595 1596 // If this is a big-endian system and the load is narrower than the 1597 // full alloca type, we need to do a shift to get the right bits. 1598 int ShAmt = 0; 1599 if (TD->isBigEndian()) { 1600 // On big-endian machines, the lowest bit is stored at the bit offset 1601 // from the pointer given by getTypeStoreSizeInBits. This matters for 1602 // integers with a bitwidth that is not a multiple of 8. 1603 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1604 TD->getTypeStoreSizeInBits(ToType) - Offset; 1605 } else { 1606 ShAmt = Offset; 1607 } 1608 1609 // Note: we support negative bitwidths (with shl) which are not defined. 1610 // We do this to support (f.e.) loads off the end of a structure where 1611 // only some bits are used. 1612 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1613 FromVal = Builder.CreateLShr(FromVal, 1614 ConstantInt::get(FromVal->getType(), 1615 ShAmt), "tmp"); 1616 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1617 FromVal = Builder.CreateShl(FromVal, 1618 ConstantInt::get(FromVal->getType(), 1619 -ShAmt), "tmp"); 1620 1621 // Finally, unconditionally truncate the integer to the right width. 1622 unsigned LIBitWidth = TD->getTypeSizeInBits(ToType); 1623 if (LIBitWidth < NTy->getBitWidth()) 1624 FromVal = 1625 Builder.CreateTrunc(FromVal, IntegerType::get(LIBitWidth), "tmp"); 1626 else if (LIBitWidth > NTy->getBitWidth()) 1627 FromVal = 1628 Builder.CreateZExt(FromVal, IntegerType::get(LIBitWidth), "tmp"); 1629 1630 // If the result is an integer, this is a trunc or bitcast. 1631 if (isa<IntegerType>(ToType)) { 1632 // Should be done. 1633 } else if (ToType->isFloatingPoint() || isa<VectorType>(ToType)) { 1634 // Just do a bitcast, we know the sizes match up. 1635 FromVal = Builder.CreateBitCast(FromVal, ToType, "tmp"); 1636 } else { 1637 // Otherwise must be a pointer. 1638 FromVal = Builder.CreateIntToPtr(FromVal, ToType, "tmp"); 1639 } 1640 assert(FromVal->getType() == ToType && "Didn't convert right?"); 1641 return FromVal; 1642} 1643 1644 1645/// ConvertScalar_InsertValue - Insert the value "SV" into the existing integer 1646/// or vector value "Old" at the offset specified by Offset. 1647/// 1648/// This happens when we are converting an "integer union" to a 1649/// single integer scalar, or when we are converting a "vector union" to a 1650/// vector with insert/extractelement instructions. 1651/// 1652/// Offset is an offset from the original alloca, in bits that need to be 1653/// shifted to the right. 1654Value *SROA::ConvertScalar_InsertValue(Value *SV, Value *Old, 1655 uint64_t Offset, IRBuilder<> &Builder) { 1656 1657 // Convert the stored type to the actual type, shift it left to insert 1658 // then 'or' into place. 1659 const Type *AllocaType = Old->getType(); 1660 LLVMContext &Context = Old->getContext(); 1661 1662 if (const VectorType *VTy = dyn_cast<VectorType>(AllocaType)) { 1663 uint64_t VecSize = TD->getTypeAllocSizeInBits(VTy); 1664 uint64_t ValSize = TD->getTypeAllocSizeInBits(SV->getType()); 1665 1666 // Changing the whole vector with memset or with an access of a different 1667 // vector type? 1668 if (ValSize == VecSize) 1669 return Builder.CreateBitCast(SV, AllocaType, "tmp"); 1670 1671 uint64_t EltSize = TD->getTypeAllocSizeInBits(VTy->getElementType()); 1672 1673 // Must be an element insertion. 1674 unsigned Elt = Offset/EltSize; 1675 1676 if (SV->getType() != VTy->getElementType()) 1677 SV = Builder.CreateBitCast(SV, VTy->getElementType(), "tmp"); 1678 1679 SV = Builder.CreateInsertElement(Old, SV, 1680 ConstantInt::get(Type::Int32Ty, Elt), 1681 "tmp"); 1682 return SV; 1683 } 1684 1685 // If SV is a first-class aggregate value, insert each value recursively. 1686 if (const StructType *ST = dyn_cast<StructType>(SV->getType())) { 1687 const StructLayout &Layout = *TD->getStructLayout(ST); 1688 for (unsigned i = 0, e = ST->getNumElements(); i != e; ++i) { 1689 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1690 Old = ConvertScalar_InsertValue(Elt, Old, 1691 Offset+Layout.getElementOffsetInBits(i), 1692 Builder); 1693 } 1694 return Old; 1695 } 1696 1697 if (const ArrayType *AT = dyn_cast<ArrayType>(SV->getType())) { 1698 uint64_t EltSize = TD->getTypeAllocSizeInBits(AT->getElementType()); 1699 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 1700 Value *Elt = Builder.CreateExtractValue(SV, i, "tmp"); 1701 Old = ConvertScalar_InsertValue(Elt, Old, Offset+i*EltSize, Builder); 1702 } 1703 return Old; 1704 } 1705 1706 // If SV is a float, convert it to the appropriate integer type. 1707 // If it is a pointer, do the same. 1708 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1709 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1710 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1711 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1712 if (SV->getType()->isFloatingPoint() || isa<VectorType>(SV->getType())) 1713 SV = Builder.CreateBitCast(SV, IntegerType::get(SrcWidth), "tmp"); 1714 else if (isa<PointerType>(SV->getType())) 1715 SV = Builder.CreatePtrToInt(SV, TD->getIntPtrType(), "tmp"); 1716 1717 // Zero extend or truncate the value if needed. 1718 if (SV->getType() != AllocaType) { 1719 if (SV->getType()->getPrimitiveSizeInBits() < 1720 AllocaType->getPrimitiveSizeInBits()) 1721 SV = Builder.CreateZExt(SV, AllocaType, "tmp"); 1722 else { 1723 // Truncation may be needed if storing more than the alloca can hold 1724 // (undefined behavior). 1725 SV = Builder.CreateTrunc(SV, AllocaType, "tmp"); 1726 SrcWidth = DestWidth; 1727 SrcStoreWidth = DestStoreWidth; 1728 } 1729 } 1730 1731 // If this is a big-endian system and the store is narrower than the 1732 // full alloca type, we need to do a shift to get the right bits. 1733 int ShAmt = 0; 1734 if (TD->isBigEndian()) { 1735 // On big-endian machines, the lowest bit is stored at the bit offset 1736 // from the pointer given by getTypeStoreSizeInBits. This matters for 1737 // integers with a bitwidth that is not a multiple of 8. 1738 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1739 } else { 1740 ShAmt = Offset; 1741 } 1742 1743 // Note: we support negative bitwidths (with shr) which are not defined. 1744 // We do this to support (f.e.) stores off the end of a structure where 1745 // only some bits in the structure are set. 1746 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1747 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1748 SV = Builder.CreateShl(SV, ConstantInt::get(SV->getType(), 1749 ShAmt), "tmp"); 1750 Mask <<= ShAmt; 1751 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1752 SV = Builder.CreateLShr(SV, ConstantInt::get(SV->getType(), 1753 -ShAmt), "tmp"); 1754 Mask = Mask.lshr(-ShAmt); 1755 } 1756 1757 // Mask out the bits we are about to insert from the old value, and or 1758 // in the new bits. 1759 if (SrcWidth != DestWidth) { 1760 assert(DestWidth > SrcWidth); 1761 Old = Builder.CreateAnd(Old, ConstantInt::get(Context, ~Mask), "mask"); 1762 SV = Builder.CreateOr(Old, SV, "ins"); 1763 } 1764 return SV; 1765} 1766 1767 1768 1769/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1770/// some part of a constant global variable. This intentionally only accepts 1771/// constant expressions because we don't can't rewrite arbitrary instructions. 1772static bool PointsToConstantGlobal(Value *V) { 1773 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1774 return GV->isConstant(); 1775 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1776 if (CE->getOpcode() == Instruction::BitCast || 1777 CE->getOpcode() == Instruction::GetElementPtr) 1778 return PointsToConstantGlobal(CE->getOperand(0)); 1779 return false; 1780} 1781 1782/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1783/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1784/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1785/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1786/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1787/// the alloca, and if the source pointer is a pointer to a constant global, we 1788/// can optimize this. 1789static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1790 bool isOffset) { 1791 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1792 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) 1793 // Ignore non-volatile loads, they are always ok. 1794 if (!LI->isVolatile()) 1795 continue; 1796 1797 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1798 // If uses of the bitcast are ok, we are ok. 1799 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1800 return false; 1801 continue; 1802 } 1803 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1804 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1805 // doesn't, it does. 1806 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1807 isOffset || !GEP->hasAllZeroIndices())) 1808 return false; 1809 continue; 1810 } 1811 1812 // If this is isn't our memcpy/memmove, reject it as something we can't 1813 // handle. 1814 if (!isa<MemTransferInst>(*UI)) 1815 return false; 1816 1817 // If we already have seen a copy, reject the second one. 1818 if (TheCopy) return false; 1819 1820 // If the pointer has been offset from the start of the alloca, we can't 1821 // safely handle this. 1822 if (isOffset) return false; 1823 1824 // If the memintrinsic isn't using the alloca as the dest, reject it. 1825 if (UI.getOperandNo() != 1) return false; 1826 1827 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1828 1829 // If the source of the memcpy/move is not a constant global, reject it. 1830 if (!PointsToConstantGlobal(MI->getOperand(2))) 1831 return false; 1832 1833 // Otherwise, the transform is safe. Remember the copy instruction. 1834 TheCopy = MI; 1835 } 1836 return true; 1837} 1838 1839/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1840/// modified by a copy from a constant global. If we can prove this, we can 1841/// replace any uses of the alloca with uses of the global directly. 1842Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) { 1843 Instruction *TheCopy = 0; 1844 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1845 return TheCopy; 1846 return 0; 1847} 1848