ScalarReplAggregates.cpp revision d93afec1dbbb1abb3df55e2e007b5f256d09f84a
1//===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This transformation implements the well known scalar replacement of 11// aggregates transformation. This xform breaks up alloca instructions of 12// aggregate type (structure or array) into individual alloca instructions for 13// each member (if possible). Then, if possible, it transforms the individual 14// alloca instructions into nice clean scalar SSA form. 15// 16// This combines a simple SRoA algorithm with the Mem2Reg algorithm because 17// often interact, especially for C++ programs. As such, iterating between 18// SRoA, then Mem2Reg until we run out of things to promote works well. 19// 20//===----------------------------------------------------------------------===// 21 22#define DEBUG_TYPE "scalarrepl" 23#include "llvm/Transforms/Scalar.h" 24#include "llvm/Constants.h" 25#include "llvm/DerivedTypes.h" 26#include "llvm/Function.h" 27#include "llvm/GlobalVariable.h" 28#include "llvm/Instructions.h" 29#include "llvm/IntrinsicInst.h" 30#include "llvm/Pass.h" 31#include "llvm/Analysis/Dominators.h" 32#include "llvm/Target/TargetData.h" 33#include "llvm/Transforms/Utils/PromoteMemToReg.h" 34#include "llvm/Support/Debug.h" 35#include "llvm/Support/GetElementPtrTypeIterator.h" 36#include "llvm/Support/MathExtras.h" 37#include "llvm/Support/Compiler.h" 38#include "llvm/ADT/SmallVector.h" 39#include "llvm/ADT/Statistic.h" 40#include "llvm/ADT/StringExtras.h" 41using namespace llvm; 42 43STATISTIC(NumReplaced, "Number of allocas broken up"); 44STATISTIC(NumPromoted, "Number of allocas promoted"); 45STATISTIC(NumConverted, "Number of aggregates converted to scalar"); 46STATISTIC(NumGlobals, "Number of allocas copied from constant global"); 47 48namespace { 49 struct VISIBILITY_HIDDEN SROA : public FunctionPass { 50 static char ID; // Pass identification, replacement for typeid 51 explicit SROA(signed T = -1) : FunctionPass(&ID) { 52 if (T == -1) 53 SRThreshold = 128; 54 else 55 SRThreshold = T; 56 } 57 58 bool runOnFunction(Function &F); 59 60 bool performScalarRepl(Function &F); 61 bool performPromotion(Function &F); 62 63 // getAnalysisUsage - This pass does not require any passes, but we know it 64 // will not alter the CFG, so say so. 65 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 66 AU.addRequired<DominatorTree>(); 67 AU.addRequired<DominanceFrontier>(); 68 AU.addRequired<TargetData>(); 69 AU.setPreservesCFG(); 70 } 71 72 private: 73 TargetData *TD; 74 75 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 76 /// information about the uses. All these fields are initialized to false 77 /// and set to true when something is learned. 78 struct AllocaInfo { 79 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 80 bool isUnsafe : 1; 81 82 /// needsCanon - This is set to true if there is some use of the alloca 83 /// that requires canonicalization. 84 bool needsCanon : 1; 85 86 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 87 bool isMemCpySrc : 1; 88 89 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 90 bool isMemCpyDst : 1; 91 92 AllocaInfo() 93 : isUnsafe(false), needsCanon(false), 94 isMemCpySrc(false), isMemCpyDst(false) {} 95 }; 96 97 unsigned SRThreshold; 98 99 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 100 101 int isSafeAllocaToScalarRepl(AllocationInst *AI); 102 103 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 104 AllocaInfo &Info); 105 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 106 AllocaInfo &Info); 107 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 108 unsigned OpNo, AllocaInfo &Info); 109 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI, 110 AllocaInfo &Info); 111 112 void DoScalarReplacement(AllocationInst *AI, 113 std::vector<AllocationInst*> &WorkList); 114 void CanonicalizeAllocaUsers(AllocationInst *AI); 115 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base); 116 117 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 118 SmallVector<AllocaInst*, 32> &NewElts); 119 120 void RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 121 AllocationInst *AI, 122 SmallVector<AllocaInst*, 32> &NewElts); 123 124 125 const Type *CanConvertToScalar(Value *V, bool &IsNotTrivial); 126 void ConvertToScalar(AllocationInst *AI, const Type *Ty); 127 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset); 128 Value *ConvertUsesOfLoadToScalar(LoadInst *LI, AllocaInst *NewAI, 129 unsigned Offset); 130 Value *ConvertUsesOfStoreToScalar(StoreInst *SI, AllocaInst *NewAI, 131 unsigned Offset); 132 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI); 133 }; 134} 135 136char SROA::ID = 0; 137static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 138 139// Public interface to the ScalarReplAggregates pass 140FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 141 return new SROA(Threshold); 142} 143 144 145bool SROA::runOnFunction(Function &F) { 146 TD = &getAnalysis<TargetData>(); 147 148 bool Changed = performPromotion(F); 149 while (1) { 150 bool LocalChange = performScalarRepl(F); 151 if (!LocalChange) break; // No need to repromote if no scalarrepl 152 Changed = true; 153 LocalChange = performPromotion(F); 154 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 155 } 156 157 return Changed; 158} 159 160 161bool SROA::performPromotion(Function &F) { 162 std::vector<AllocaInst*> Allocas; 163 DominatorTree &DT = getAnalysis<DominatorTree>(); 164 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 165 166 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 167 168 bool Changed = false; 169 170 while (1) { 171 Allocas.clear(); 172 173 // Find allocas that are safe to promote, by looking at all instructions in 174 // the entry node 175 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 176 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 177 if (isAllocaPromotable(AI)) 178 Allocas.push_back(AI); 179 180 if (Allocas.empty()) break; 181 182 PromoteMemToReg(Allocas, DT, DF); 183 NumPromoted += Allocas.size(); 184 Changed = true; 185 } 186 187 return Changed; 188} 189 190/// getNumSAElements - Return the number of elements in the specific struct or 191/// array. 192static uint64_t getNumSAElements(const Type *T) { 193 if (const StructType *ST = dyn_cast<StructType>(T)) 194 return ST->getNumElements(); 195 return cast<ArrayType>(T)->getNumElements(); 196} 197 198// performScalarRepl - This algorithm is a simple worklist driven algorithm, 199// which runs on all of the malloc/alloca instructions in the function, removing 200// them if they are only used by getelementptr instructions. 201// 202bool SROA::performScalarRepl(Function &F) { 203 std::vector<AllocationInst*> WorkList; 204 205 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 206 BasicBlock &BB = F.getEntryBlock(); 207 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 208 if (AllocationInst *A = dyn_cast<AllocationInst>(I)) 209 WorkList.push_back(A); 210 211 // Process the worklist 212 bool Changed = false; 213 while (!WorkList.empty()) { 214 AllocationInst *AI = WorkList.back(); 215 WorkList.pop_back(); 216 217 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 218 // with unused elements. 219 if (AI->use_empty()) { 220 AI->eraseFromParent(); 221 continue; 222 } 223 224 // If we can turn this aggregate value (potentially with casts) into a 225 // simple scalar value that can be mem2reg'd into a register value. 226 bool IsNotTrivial = false; 227 if (const Type *ActualType = CanConvertToScalar(AI, IsNotTrivial)) 228 if (IsNotTrivial && ActualType != Type::VoidTy) { 229 ConvertToScalar(AI, ActualType); 230 Changed = true; 231 continue; 232 } 233 234 // Check to see if we can perform the core SROA transformation. We cannot 235 // transform the allocation instruction if it is an array allocation 236 // (allocations OF arrays are ok though), and an allocation of a scalar 237 // value cannot be decomposed at all. 238 if (!AI->isArrayAllocation() && 239 (isa<StructType>(AI->getAllocatedType()) || 240 isa<ArrayType>(AI->getAllocatedType())) && 241 AI->getAllocatedType()->isSized() && 242 // Do not promote any struct whose size is larger than "128" bytes. 243 TD->getABITypeSize(AI->getAllocatedType()) < SRThreshold && 244 // Do not promote any struct into more than "32" separate vars. 245 getNumSAElements(AI->getAllocatedType()) < SRThreshold/4) { 246 // Check that all of the users of the allocation are capable of being 247 // transformed. 248 switch (isSafeAllocaToScalarRepl(AI)) { 249 default: assert(0 && "Unexpected value!"); 250 case 0: // Not safe to scalar replace. 251 break; 252 case 1: // Safe, but requires cleanup/canonicalizations first 253 CanonicalizeAllocaUsers(AI); 254 // FALL THROUGH. 255 case 3: // Safe to scalar replace. 256 DoScalarReplacement(AI, WorkList); 257 Changed = true; 258 continue; 259 } 260 } 261 262 // Check to see if this allocation is only modified by a memcpy/memmove from 263 // a constant global. If this is the case, we can change all users to use 264 // the constant global instead. This is commonly produced by the CFE by 265 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 266 // is only subsequently read. 267 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 268 DOUT << "Found alloca equal to global: " << *AI; 269 DOUT << " memcpy = " << *TheCopy; 270 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 271 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 272 TheCopy->eraseFromParent(); // Don't mutate the global. 273 AI->eraseFromParent(); 274 ++NumGlobals; 275 Changed = true; 276 continue; 277 } 278 279 // Otherwise, couldn't process this. 280 } 281 282 return Changed; 283} 284 285/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 286/// predicate, do SROA now. 287void SROA::DoScalarReplacement(AllocationInst *AI, 288 std::vector<AllocationInst*> &WorkList) { 289 DOUT << "Found inst to SROA: " << *AI; 290 SmallVector<AllocaInst*, 32> ElementAllocas; 291 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 292 ElementAllocas.reserve(ST->getNumContainedTypes()); 293 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 294 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 295 AI->getAlignment(), 296 AI->getName() + "." + utostr(i), AI); 297 ElementAllocas.push_back(NA); 298 WorkList.push_back(NA); // Add to worklist for recursive processing 299 } 300 } else { 301 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 302 ElementAllocas.reserve(AT->getNumElements()); 303 const Type *ElTy = AT->getElementType(); 304 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 305 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 306 AI->getName() + "." + utostr(i), AI); 307 ElementAllocas.push_back(NA); 308 WorkList.push_back(NA); // Add to worklist for recursive processing 309 } 310 } 311 312 // Now that we have created the alloca instructions that we want to use, 313 // expand the getelementptr instructions to use them. 314 // 315 while (!AI->use_empty()) { 316 Instruction *User = cast<Instruction>(AI->use_back()); 317 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) { 318 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas); 319 BCInst->eraseFromParent(); 320 continue; 321 } 322 323 // Replace: 324 // %res = load { i32, i32 }* %alloc 325 // with: 326 // %load.0 = load i32* %alloc.0 327 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 328 // %load.1 = load i32* %alloc.1 329 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 330 // (Also works for arrays instead of structs) 331 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 332 Value *Insert = UndefValue::get(LI->getType()); 333 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 334 Value *Load = new LoadInst(ElementAllocas[i], "load", LI); 335 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 336 } 337 LI->replaceAllUsesWith(Insert); 338 LI->eraseFromParent(); 339 continue; 340 } 341 342 // Replace: 343 // store { i32, i32 } %val, { i32, i32 }* %alloc 344 // with: 345 // %val.0 = extractvalue { i32, i32 } %val, 0 346 // store i32 %val.0, i32* %alloc.0 347 // %val.1 = extractvalue { i32, i32 } %val, 1 348 // store i32 %val.1, i32* %alloc.1 349 // (Also works for arrays instead of structs) 350 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 351 Value *Val = SI->getOperand(0); 352 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 353 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 354 new StoreInst(Extract, ElementAllocas[i], SI); 355 } 356 SI->eraseFromParent(); 357 continue; 358 } 359 360 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 361 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 362 unsigned Idx = 363 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 364 365 assert(Idx < ElementAllocas.size() && "Index out of range?"); 366 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 367 368 Value *RepValue; 369 if (GEPI->getNumOperands() == 3) { 370 // Do not insert a new getelementptr instruction with zero indices, only 371 // to have it optimized out later. 372 RepValue = AllocaToUse; 373 } else { 374 // We are indexing deeply into the structure, so we still need a 375 // getelement ptr instruction to finish the indexing. This may be 376 // expanded itself once the worklist is rerun. 377 // 378 SmallVector<Value*, 8> NewArgs; 379 NewArgs.push_back(Constant::getNullValue(Type::Int32Ty)); 380 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end()); 381 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(), 382 NewArgs.end(), "", GEPI); 383 RepValue->takeName(GEPI); 384 } 385 386 // If this GEP is to the start of the aggregate, check for memcpys. 387 if (Idx == 0 && GEPI->hasAllZeroIndices()) 388 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas); 389 390 // Move all of the users over to the new GEP. 391 GEPI->replaceAllUsesWith(RepValue); 392 // Delete the old GEP 393 GEPI->eraseFromParent(); 394 } 395 396 // Finally, delete the Alloca instruction 397 AI->eraseFromParent(); 398 NumReplaced++; 399} 400 401 402/// isSafeElementUse - Check to see if this use is an allowed use for a 403/// getelementptr instruction of an array aggregate allocation. isFirstElt 404/// indicates whether Ptr is known to the start of the aggregate. 405/// 406void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 407 AllocaInfo &Info) { 408 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 409 I != E; ++I) { 410 Instruction *User = cast<Instruction>(*I); 411 switch (User->getOpcode()) { 412 case Instruction::Load: break; 413 case Instruction::Store: 414 // Store is ok if storing INTO the pointer, not storing the pointer 415 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info); 416 break; 417 case Instruction::GetElementPtr: { 418 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 419 bool AreAllZeroIndices = isFirstElt; 420 if (GEP->getNumOperands() > 1) { 421 if (!isa<ConstantInt>(GEP->getOperand(1)) || 422 !cast<ConstantInt>(GEP->getOperand(1))->isZero()) 423 // Using pointer arithmetic to navigate the array. 424 return MarkUnsafe(Info); 425 426 if (AreAllZeroIndices) 427 AreAllZeroIndices = GEP->hasAllZeroIndices(); 428 } 429 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 430 if (Info.isUnsafe) return; 431 break; 432 } 433 case Instruction::BitCast: 434 if (isFirstElt) { 435 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 436 if (Info.isUnsafe) return; 437 break; 438 } 439 DOUT << " Transformation preventing inst: " << *User; 440 return MarkUnsafe(Info); 441 case Instruction::Call: 442 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 443 if (isFirstElt) { 444 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 445 if (Info.isUnsafe) return; 446 break; 447 } 448 } 449 DOUT << " Transformation preventing inst: " << *User; 450 return MarkUnsafe(Info); 451 default: 452 DOUT << " Transformation preventing inst: " << *User; 453 return MarkUnsafe(Info); 454 } 455 } 456 return; // All users look ok :) 457} 458 459/// AllUsersAreLoads - Return true if all users of this value are loads. 460static bool AllUsersAreLoads(Value *Ptr) { 461 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 462 I != E; ++I) 463 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 464 return false; 465 return true; 466} 467 468/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 469/// aggregate allocation. 470/// 471void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 472 AllocaInfo &Info) { 473 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 474 return isSafeUseOfBitCastedAllocation(C, AI, Info); 475 476 if (isa<LoadInst>(User)) 477 return; // Loads (returning a first class aggregrate) are always rewritable 478 479 if (isa<StoreInst>(User) && User->getOperand(0) != AI) 480 return; // Store is ok if storing INTO the pointer, not storing the pointer 481 482 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 483 if (GEPI == 0) 484 return MarkUnsafe(Info); 485 486 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 487 488 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 489 if (I == E || 490 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) { 491 return MarkUnsafe(Info); 492 } 493 494 ++I; 495 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 496 497 bool IsAllZeroIndices = true; 498 499 // If the first index is a non-constant index into an array, see if we can 500 // handle it as a special case. 501 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 502 if (!isa<ConstantInt>(I.getOperand())) { 503 IsAllZeroIndices = 0; 504 uint64_t NumElements = AT->getNumElements(); 505 506 // If this is an array index and the index is not constant, we cannot 507 // promote... that is unless the array has exactly one or two elements in 508 // it, in which case we CAN promote it, but we have to canonicalize this 509 // out if this is the only problem. 510 if ((NumElements == 1 || NumElements == 2) && 511 AllUsersAreLoads(GEPI)) { 512 Info.needsCanon = true; 513 return; // Canonicalization required! 514 } 515 return MarkUnsafe(Info); 516 } 517 } 518 519 // Walk through the GEP type indices, checking the types that this indexes 520 // into. 521 for (; I != E; ++I) { 522 // Ignore struct elements, no extra checking needed for these. 523 if (isa<StructType>(*I)) 524 continue; 525 526 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 527 if (!IdxVal) return MarkUnsafe(Info); 528 529 // Are all indices still zero? 530 IsAllZeroIndices &= IdxVal->isZero(); 531 532 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 533 // This GEP indexes an array. Verify that this is an in-range constant 534 // integer. Specifically, consider A[0][i]. We cannot know that the user 535 // isn't doing invalid things like allowing i to index an out-of-range 536 // subscript that accesses A[1]. Because of this, we have to reject SROA 537 // of any accesses into structs where any of the components are variables. 538 if (IdxVal->getZExtValue() >= AT->getNumElements()) 539 return MarkUnsafe(Info); 540 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) { 541 if (IdxVal->getZExtValue() >= VT->getNumElements()) 542 return MarkUnsafe(Info); 543 } 544 } 545 546 // If there are any non-simple uses of this getelementptr, make sure to reject 547 // them. 548 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 549} 550 551/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory 552/// intrinsic can be promoted by SROA. At this point, we know that the operand 553/// of the memintrinsic is a pointer to the beginning of the allocation. 554void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 555 unsigned OpNo, AllocaInfo &Info) { 556 // If not constant length, give up. 557 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 558 if (!Length) return MarkUnsafe(Info); 559 560 // If not the whole aggregate, give up. 561 if (Length->getZExtValue() != 562 TD->getABITypeSize(AI->getType()->getElementType())) 563 return MarkUnsafe(Info); 564 565 // We only know about memcpy/memset/memmove. 566 if (!isa<MemCpyInst>(MI) && !isa<MemSetInst>(MI) && !isa<MemMoveInst>(MI)) 567 return MarkUnsafe(Info); 568 569 // Otherwise, we can transform it. Determine whether this is a memcpy/set 570 // into or out of the aggregate. 571 if (OpNo == 1) 572 Info.isMemCpyDst = true; 573 else { 574 assert(OpNo == 2); 575 Info.isMemCpySrc = true; 576 } 577} 578 579/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast 580/// are 581void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI, 582 AllocaInfo &Info) { 583 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 584 UI != E; ++UI) { 585 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 586 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 587 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 588 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 589 } else { 590 return MarkUnsafe(Info); 591 } 592 if (Info.isUnsafe) return; 593 } 594} 595 596/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 597/// to its first element. Transform users of the cast to use the new values 598/// instead. 599void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 600 SmallVector<AllocaInst*, 32> &NewElts) { 601 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 602 while (UI != UE) { 603 Instruction *User = cast<Instruction>(*UI++); 604 if (BitCastInst *BCU = dyn_cast<BitCastInst>(User)) { 605 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 606 BCU->eraseFromParent(); 607 continue; 608 } 609 610 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 611 // This must be memcpy/memmove/memset of the entire aggregate. 612 // Split into one per element. 613 RewriteMemIntrinUserOfAlloca(MI, BCInst, AI, NewElts); 614 MI->eraseFromParent(); 615 continue; 616 } 617 618 // If it's not a mem intrinsic, it must be some other user of a gep of the 619 // first pointer. Just leave these alone. 620 continue; 621 } 622} 623 624/// RewriteMemIntrinUserOfAlloca - MI is a memcpy/memset/memmove from or to AI. 625/// Rewrite it to copy or set the elements of the scalarized memory. 626void SROA::RewriteMemIntrinUserOfAlloca(MemIntrinsic *MI, Instruction *BCInst, 627 AllocationInst *AI, 628 SmallVector<AllocaInst*, 32> &NewElts) { 629 630 // If this is a memcpy/memmove, construct the other pointer as the 631 // appropriate type. 632 Value *OtherPtr = 0; 633 if (MemCpyInst *MCI = dyn_cast<MemCpyInst>(MI)) { 634 if (BCInst == MCI->getRawDest()) 635 OtherPtr = MCI->getRawSource(); 636 else { 637 assert(BCInst == MCI->getRawSource()); 638 OtherPtr = MCI->getRawDest(); 639 } 640 } else if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) { 641 if (BCInst == MMI->getRawDest()) 642 OtherPtr = MMI->getRawSource(); 643 else { 644 assert(BCInst == MMI->getRawSource()); 645 OtherPtr = MMI->getRawDest(); 646 } 647 } 648 649 // If there is an other pointer, we want to convert it to the same pointer 650 // type as AI has, so we can GEP through it safely. 651 if (OtherPtr) { 652 // It is likely that OtherPtr is a bitcast, if so, remove it. 653 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 654 OtherPtr = BC->getOperand(0); 655 // All zero GEPs are effectively bitcasts. 656 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) 657 if (GEP->hasAllZeroIndices()) 658 OtherPtr = GEP->getOperand(0); 659 660 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 661 if (BCE->getOpcode() == Instruction::BitCast) 662 OtherPtr = BCE->getOperand(0); 663 664 // If the pointer is not the right type, insert a bitcast to the right 665 // type. 666 if (OtherPtr->getType() != AI->getType()) 667 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 668 MI); 669 } 670 671 // Process each element of the aggregate. 672 Value *TheFn = MI->getOperand(0); 673 const Type *BytePtrTy = MI->getRawDest()->getType(); 674 bool SROADest = MI->getRawDest() == BCInst; 675 676 Constant *Zero = Constant::getNullValue(Type::Int32Ty); 677 678 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 679 // If this is a memcpy/memmove, emit a GEP of the other element address. 680 Value *OtherElt = 0; 681 if (OtherPtr) { 682 Value *Idx[2] = { Zero, ConstantInt::get(Type::Int32Ty, i) }; 683 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2, 684 OtherPtr->getNameStr()+"."+utostr(i), 685 MI); 686 } 687 688 Value *EltPtr = NewElts[i]; 689 const Type *EltTy =cast<PointerType>(EltPtr->getType())->getElementType(); 690 691 // If we got down to a scalar, insert a load or store as appropriate. 692 if (EltTy->isSingleValueType()) { 693 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) { 694 Value *Elt = new LoadInst(SROADest ? OtherElt : EltPtr, "tmp", 695 MI); 696 new StoreInst(Elt, SROADest ? EltPtr : OtherElt, MI); 697 continue; 698 } 699 assert(isa<MemSetInst>(MI)); 700 701 // If the stored element is zero (common case), just store a null 702 // constant. 703 Constant *StoreVal; 704 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 705 if (CI->isZero()) { 706 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 707 } else { 708 // If EltTy is a vector type, get the element type. 709 const Type *ValTy = EltTy; 710 if (const VectorType *VTy = dyn_cast<VectorType>(ValTy)) 711 ValTy = VTy->getElementType(); 712 713 // Construct an integer with the right value. 714 unsigned EltSize = TD->getTypeSizeInBits(ValTy); 715 APInt OneVal(EltSize, CI->getZExtValue()); 716 APInt TotalVal(OneVal); 717 // Set each byte. 718 for (unsigned i = 0; 8*i < EltSize; ++i) { 719 TotalVal = TotalVal.shl(8); 720 TotalVal |= OneVal; 721 } 722 723 // Convert the integer value to the appropriate type. 724 StoreVal = ConstantInt::get(TotalVal); 725 if (isa<PointerType>(ValTy)) 726 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 727 else if (ValTy->isFloatingPoint()) 728 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 729 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 730 731 // If the requested value was a vector constant, create it. 732 if (EltTy != ValTy) { 733 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 734 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 735 StoreVal = ConstantVector::get(&Elts[0], NumElts); 736 } 737 } 738 new StoreInst(StoreVal, EltPtr, MI); 739 continue; 740 } 741 // Otherwise, if we're storing a byte variable, use a memset call for 742 // this element. 743 } 744 745 // Cast the element pointer to BytePtrTy. 746 if (EltPtr->getType() != BytePtrTy) 747 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 748 749 // Cast the other pointer (if we have one) to BytePtrTy. 750 if (OtherElt && OtherElt->getType() != BytePtrTy) 751 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 752 MI); 753 754 unsigned EltSize = TD->getABITypeSize(EltTy); 755 756 // Finally, insert the meminst for this element. 757 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) { 758 Value *Ops[] = { 759 SROADest ? EltPtr : OtherElt, // Dest ptr 760 SROADest ? OtherElt : EltPtr, // Src ptr 761 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 762 Zero // Align 763 }; 764 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 765 } else { 766 assert(isa<MemSetInst>(MI)); 767 Value *Ops[] = { 768 EltPtr, MI->getOperand(2), // Dest, Value, 769 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 770 Zero // Align 771 }; 772 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 773 } 774 } 775} 776 777 778/// HasPadding - Return true if the specified type has any structure or 779/// alignment padding, false otherwise. 780static bool HasPadding(const Type *Ty, const TargetData &TD) { 781 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 782 const StructLayout *SL = TD.getStructLayout(STy); 783 unsigned PrevFieldBitOffset = 0; 784 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 785 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 786 787 // Padding in sub-elements? 788 if (HasPadding(STy->getElementType(i), TD)) 789 return true; 790 791 // Check to see if there is any padding between this element and the 792 // previous one. 793 if (i) { 794 unsigned PrevFieldEnd = 795 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 796 if (PrevFieldEnd < FieldBitOffset) 797 return true; 798 } 799 800 PrevFieldBitOffset = FieldBitOffset; 801 } 802 803 // Check for tail padding. 804 if (unsigned EltCount = STy->getNumElements()) { 805 unsigned PrevFieldEnd = PrevFieldBitOffset + 806 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 807 if (PrevFieldEnd < SL->getSizeInBits()) 808 return true; 809 } 810 811 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 812 return HasPadding(ATy->getElementType(), TD); 813 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 814 return HasPadding(VTy->getElementType(), TD); 815 } 816 return TD.getTypeSizeInBits(Ty) != TD.getABITypeSizeInBits(Ty); 817} 818 819/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 820/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 821/// or 1 if safe after canonicalization has been performed. 822/// 823int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 824 // Loop over the use list of the alloca. We can only transform it if all of 825 // the users are safe to transform. 826 AllocaInfo Info; 827 828 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 829 I != E; ++I) { 830 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 831 if (Info.isUnsafe) { 832 DOUT << "Cannot transform: " << *AI << " due to user: " << **I; 833 return 0; 834 } 835 } 836 837 // Okay, we know all the users are promotable. If the aggregate is a memcpy 838 // source and destination, we have to be careful. In particular, the memcpy 839 // could be moving around elements that live in structure padding of the LLVM 840 // types, but may actually be used. In these cases, we refuse to promote the 841 // struct. 842 if (Info.isMemCpySrc && Info.isMemCpyDst && 843 HasPadding(AI->getType()->getElementType(), *TD)) 844 return 0; 845 846 // If we require cleanup, return 1, otherwise return 3. 847 return Info.needsCanon ? 1 : 3; 848} 849 850/// CanonicalizeAllocaUsers - If SROA reported that it can promote the specified 851/// allocation, but only if cleaned up, perform the cleanups required. 852void SROA::CanonicalizeAllocaUsers(AllocationInst *AI) { 853 // At this point, we know that the end result will be SROA'd and promoted, so 854 // we can insert ugly code if required so long as sroa+mem2reg will clean it 855 // up. 856 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 857 UI != E; ) { 858 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI++); 859 if (!GEPI) continue; 860 gep_type_iterator I = gep_type_begin(GEPI); 861 ++I; 862 863 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 864 uint64_t NumElements = AT->getNumElements(); 865 866 if (!isa<ConstantInt>(I.getOperand())) { 867 if (NumElements == 1) { 868 GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty)); 869 } else { 870 assert(NumElements == 2 && "Unhandled case!"); 871 // All users of the GEP must be loads. At each use of the GEP, insert 872 // two loads of the appropriate indexed GEP and select between them. 873 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(), 874 Constant::getNullValue(I.getOperand()->getType()), 875 "isone", GEPI); 876 // Insert the new GEP instructions, which are properly indexed. 877 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 878 Indices[1] = Constant::getNullValue(Type::Int32Ty); 879 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 880 Indices.begin(), 881 Indices.end(), 882 GEPI->getName()+".0", GEPI); 883 Indices[1] = ConstantInt::get(Type::Int32Ty, 1); 884 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 885 Indices.begin(), 886 Indices.end(), 887 GEPI->getName()+".1", GEPI); 888 // Replace all loads of the variable index GEP with loads from both 889 // indexes and a select. 890 while (!GEPI->use_empty()) { 891 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 892 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 893 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 894 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 895 LI->replaceAllUsesWith(R); 896 LI->eraseFromParent(); 897 } 898 GEPI->eraseFromParent(); 899 } 900 } 901 } 902 } 903} 904 905/// MergeInType - Add the 'In' type to the accumulated type so far. If the 906/// types are incompatible, return true, otherwise update Accum and return 907/// false. 908/// 909/// There are three cases we handle here: 910/// 1) An effectively-integer union, where the pieces are stored into as 911/// smaller integers (common with byte swap and other idioms). 912/// 2) A union of vector types of the same size and potentially its elements. 913/// Here we turn element accesses into insert/extract element operations. 914/// 3) A union of scalar types, such as int/float or int/pointer. Here we 915/// merge together into integers, allowing the xform to work with #1 as 916/// well. 917static bool MergeInType(const Type *In, const Type *&Accum, 918 const TargetData &TD) { 919 // If this is our first type, just use it. 920 const VectorType *PTy; 921 if (Accum == Type::VoidTy || In == Accum) { 922 Accum = In; 923 } else if (In == Type::VoidTy) { 924 // Noop. 925 } else if (In->isInteger() && Accum->isInteger()) { // integer union. 926 // Otherwise pick whichever type is larger. 927 if (cast<IntegerType>(In)->getBitWidth() > 928 cast<IntegerType>(Accum)->getBitWidth()) 929 Accum = In; 930 } else if (isa<PointerType>(In) && isa<PointerType>(Accum)) { 931 // Pointer unions just stay as one of the pointers. 932 } else if (isa<VectorType>(In) || isa<VectorType>(Accum)) { 933 if ((PTy = dyn_cast<VectorType>(Accum)) && 934 PTy->getElementType() == In) { 935 // Accum is a vector, and we are accessing an element: ok. 936 } else if ((PTy = dyn_cast<VectorType>(In)) && 937 PTy->getElementType() == Accum) { 938 // In is a vector, and accum is an element: ok, remember In. 939 Accum = In; 940 } else if ((PTy = dyn_cast<VectorType>(In)) && isa<VectorType>(Accum) && 941 PTy->getBitWidth() == cast<VectorType>(Accum)->getBitWidth()) { 942 // Two vectors of the same size: keep Accum. 943 } else { 944 // Cannot insert an short into a <4 x int> or handle 945 // <2 x int> -> <4 x int> 946 return true; 947 } 948 } else { 949 // Pointer/FP/Integer unions merge together as integers. 950 switch (Accum->getTypeID()) { 951 case Type::PointerTyID: Accum = TD.getIntPtrType(); break; 952 case Type::FloatTyID: Accum = Type::Int32Ty; break; 953 case Type::DoubleTyID: Accum = Type::Int64Ty; break; 954 case Type::X86_FP80TyID: return true; 955 case Type::FP128TyID: return true; 956 case Type::PPC_FP128TyID: return true; 957 default: 958 assert(Accum->isInteger() && "Unknown FP type!"); 959 break; 960 } 961 962 switch (In->getTypeID()) { 963 case Type::PointerTyID: In = TD.getIntPtrType(); break; 964 case Type::FloatTyID: In = Type::Int32Ty; break; 965 case Type::DoubleTyID: In = Type::Int64Ty; break; 966 case Type::X86_FP80TyID: return true; 967 case Type::FP128TyID: return true; 968 case Type::PPC_FP128TyID: return true; 969 default: 970 assert(In->isInteger() && "Unknown FP type!"); 971 break; 972 } 973 return MergeInType(In, Accum, TD); 974 } 975 return false; 976} 977 978/// getIntAtLeastAsBigAs - Return an integer type that is at least as big as the 979/// specified type. If there is no suitable type, this returns null. 980const Type *getIntAtLeastAsBigAs(unsigned NumBits) { 981 if (NumBits > 64) return 0; 982 if (NumBits > 32) return Type::Int64Ty; 983 if (NumBits > 16) return Type::Int32Ty; 984 if (NumBits > 8) return Type::Int16Ty; 985 return Type::Int8Ty; 986} 987 988/// CanConvertToScalar - V is a pointer. If we can convert the pointee to a 989/// single scalar integer type, return that type. Further, if the use is not 990/// a completely trivial use that mem2reg could promote, set IsNotTrivial. If 991/// there are no uses of this pointer, return Type::VoidTy to differentiate from 992/// failure. 993/// 994const Type *SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial) { 995 const Type *UsedType = Type::VoidTy; // No uses, no forced type. 996 const PointerType *PTy = cast<PointerType>(V->getType()); 997 998 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 999 Instruction *User = cast<Instruction>(*UI); 1000 1001 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1002 // FIXME: Loads of a first class aggregrate value could be converted to a 1003 // series of loads and insertvalues 1004 if (!LI->getType()->isSingleValueType()) 1005 return 0; 1006 1007 if (MergeInType(LI->getType(), UsedType, *TD)) 1008 return 0; 1009 continue; 1010 } 1011 1012 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1013 // Storing the pointer, not into the value? 1014 if (SI->getOperand(0) == V) return 0; 1015 1016 // FIXME: Stores of a first class aggregrate value could be converted to a 1017 // series of extractvalues and stores 1018 if (!SI->getOperand(0)->getType()->isSingleValueType()) 1019 return 0; 1020 1021 // NOTE: We could handle storing of FP imms into integers here! 1022 1023 if (MergeInType(SI->getOperand(0)->getType(), UsedType, *TD)) 1024 return 0; 1025 continue; 1026 } 1027 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1028 IsNotTrivial = true; 1029 const Type *SubTy = CanConvertToScalar(CI, IsNotTrivial); 1030 if (!SubTy || MergeInType(SubTy, UsedType, *TD)) return 0; 1031 continue; 1032 } 1033 1034 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1035 // Check to see if this is stepping over an element: GEP Ptr, int C 1036 if (GEP->getNumOperands() == 2 && isa<ConstantInt>(GEP->getOperand(1))) { 1037 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue(); 1038 unsigned ElSize = TD->getABITypeSize(PTy->getElementType()); 1039 unsigned BitOffset = Idx*ElSize*8; 1040 if (BitOffset > 64 || !isPowerOf2_32(ElSize)) return 0; 1041 1042 IsNotTrivial = true; 1043 const Type *SubElt = CanConvertToScalar(GEP, IsNotTrivial); 1044 if (SubElt == 0) return 0; 1045 if (SubElt != Type::VoidTy && SubElt->isInteger()) { 1046 const Type *NewTy = 1047 getIntAtLeastAsBigAs(TD->getABITypeSizeInBits(SubElt)+BitOffset); 1048 if (NewTy == 0 || MergeInType(NewTy, UsedType, *TD)) return 0; 1049 continue; 1050 } 1051 // Cannot handle this! 1052 return 0; 1053 } 1054 1055 if (GEP->getNumOperands() == 3 && 1056 isa<ConstantInt>(GEP->getOperand(1)) && 1057 isa<ConstantInt>(GEP->getOperand(2)) && 1058 cast<ConstantInt>(GEP->getOperand(1))->isZero()) { 1059 // We are stepping into an element, e.g. a structure or an array: 1060 // GEP Ptr, i32 0, i32 Cst 1061 const Type *AggTy = PTy->getElementType(); 1062 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 1063 1064 if (const ArrayType *ATy = dyn_cast<ArrayType>(AggTy)) { 1065 if (Idx >= ATy->getNumElements()) return 0; // Out of range. 1066 } else if (const VectorType *VectorTy = dyn_cast<VectorType>(AggTy)) { 1067 // Getting an element of the vector. 1068 if (Idx >= VectorTy->getNumElements()) return 0; // Out of range. 1069 1070 // Merge in the vector type. 1071 if (MergeInType(VectorTy, UsedType, *TD)) return 0; 1072 1073 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial); 1074 if (SubTy == 0) return 0; 1075 1076 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, *TD)) 1077 return 0; 1078 1079 // We'll need to change this to an insert/extract element operation. 1080 IsNotTrivial = true; 1081 continue; // Everything looks ok 1082 1083 } else if (isa<StructType>(AggTy)) { 1084 // Structs are always ok. 1085 } else { 1086 return 0; 1087 } 1088 const Type *NTy = getIntAtLeastAsBigAs(TD->getABITypeSizeInBits(AggTy)); 1089 if (NTy == 0 || MergeInType(NTy, UsedType, *TD)) return 0; 1090 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial); 1091 if (SubTy == 0) return 0; 1092 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, *TD)) 1093 return 0; 1094 continue; // Everything looks ok 1095 } 1096 return 0; 1097 } 1098 1099 // Cannot handle this! 1100 return 0; 1101 } 1102 1103 return UsedType; 1104} 1105 1106/// ConvertToScalar - The specified alloca passes the CanConvertToScalar 1107/// predicate and is non-trivial. Convert it to something that can be trivially 1108/// promoted into a register by mem2reg. 1109void SROA::ConvertToScalar(AllocationInst *AI, const Type *ActualTy) { 1110 DOUT << "CONVERT TO SCALAR: " << *AI << " TYPE = " 1111 << *ActualTy << "\n"; 1112 ++NumConverted; 1113 1114 BasicBlock *EntryBlock = AI->getParent(); 1115 assert(EntryBlock == &EntryBlock->getParent()->getEntryBlock() && 1116 "Not in the entry block!"); 1117 EntryBlock->getInstList().remove(AI); // Take the alloca out of the program. 1118 1119 // Create and insert the alloca. 1120 AllocaInst *NewAI = new AllocaInst(ActualTy, 0, AI->getName(), 1121 EntryBlock->begin()); 1122 ConvertUsesToScalar(AI, NewAI, 0); 1123 delete AI; 1124} 1125 1126 1127/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1128/// directly. This happens when we are converting an "integer union" to a 1129/// single integer scalar, or when we are converting a "vector union" to a 1130/// vector with insert/extractelement instructions. 1131/// 1132/// Offset is an offset from the original alloca, in bits that need to be 1133/// shifted to the right. By the end of this, there should be no uses of Ptr. 1134void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset) { 1135 while (!Ptr->use_empty()) { 1136 Instruction *User = cast<Instruction>(Ptr->use_back()); 1137 1138 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1139 Value *NV = ConvertUsesOfLoadToScalar(LI, NewAI, Offset); 1140 LI->replaceAllUsesWith(NV); 1141 LI->eraseFromParent(); 1142 continue; 1143 } 1144 1145 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1146 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1147 1148 Value *SV = ConvertUsesOfStoreToScalar(SI, NewAI, Offset); 1149 new StoreInst(SV, NewAI, SI); 1150 SI->eraseFromParent(); 1151 continue; 1152 } 1153 1154 if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1155 ConvertUsesToScalar(CI, NewAI, Offset); 1156 CI->eraseFromParent(); 1157 continue; 1158 } 1159 1160 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1161 const PointerType *AggPtrTy = 1162 cast<PointerType>(GEP->getOperand(0)->getType()); 1163 unsigned AggSizeInBits = 1164 TD->getABITypeSizeInBits(AggPtrTy->getElementType()); 1165 1166 // Check to see if this is stepping over an element: GEP Ptr, int C 1167 unsigned NewOffset = Offset; 1168 if (GEP->getNumOperands() == 2) { 1169 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue(); 1170 unsigned BitOffset = Idx*AggSizeInBits; 1171 1172 NewOffset += BitOffset; 1173 ConvertUsesToScalar(GEP, NewAI, NewOffset); 1174 GEP->eraseFromParent(); 1175 continue; 1176 } 1177 1178 assert(GEP->getNumOperands() == 3 && "Unsupported operation"); 1179 1180 // We know that operand #2 is zero. 1181 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 1182 const Type *AggTy = AggPtrTy->getElementType(); 1183 if (const SequentialType *SeqTy = dyn_cast<SequentialType>(AggTy)) { 1184 unsigned ElSizeBits = 1185 TD->getABITypeSizeInBits(SeqTy->getElementType()); 1186 1187 NewOffset += ElSizeBits*Idx; 1188 } else { 1189 const StructType *STy = cast<StructType>(AggTy); 1190 unsigned EltBitOffset = 1191 TD->getStructLayout(STy)->getElementOffsetInBits(Idx); 1192 1193 NewOffset += EltBitOffset; 1194 } 1195 ConvertUsesToScalar(GEP, NewAI, NewOffset); 1196 GEP->eraseFromParent(); 1197 continue; 1198 } 1199 1200 assert(0 && "Unsupported operation!"); 1201 abort(); 1202 } 1203} 1204 1205/// ConvertUsesOfLoadToScalar - Convert all of the users the specified load to 1206/// use the new alloca directly, returning the value that should replace the 1207/// load. This happens when we are converting an "integer union" to a 1208/// single integer scalar, or when we are converting a "vector union" to a 1209/// vector with insert/extractelement instructions. 1210/// 1211/// Offset is an offset from the original alloca, in bits that need to be 1212/// shifted to the right. By the end of this, there should be no uses of Ptr. 1213Value *SROA::ConvertUsesOfLoadToScalar(LoadInst *LI, AllocaInst *NewAI, 1214 unsigned Offset) { 1215 // The load is a bit extract from NewAI shifted right by Offset bits. 1216 Value *NV = new LoadInst(NewAI, LI->getName(), LI); 1217 1218 if (NV->getType() == LI->getType() && Offset == 0) { 1219 // We win, no conversion needed. 1220 return NV; 1221 } 1222 1223 // If the result type of the 'union' is a pointer, then this must be ptr->ptr 1224 // cast. Anything else would result in NV being an integer. 1225 if (isa<PointerType>(NV->getType())) { 1226 assert(isa<PointerType>(LI->getType())); 1227 return new BitCastInst(NV, LI->getType(), LI->getName(), LI); 1228 } 1229 1230 if (const VectorType *VTy = dyn_cast<VectorType>(NV->getType())) { 1231 // If the result alloca is a vector type, this is either an element 1232 // access or a bitcast to another vector type. 1233 if (isa<VectorType>(LI->getType())) 1234 return new BitCastInst(NV, LI->getType(), LI->getName(), LI); 1235 1236 // Otherwise it must be an element access. 1237 unsigned Elt = 0; 1238 if (Offset) { 1239 unsigned EltSize = TD->getABITypeSizeInBits(VTy->getElementType()); 1240 Elt = Offset/EltSize; 1241 Offset -= EltSize*Elt; 1242 } 1243 NV = new ExtractElementInst(NV, ConstantInt::get(Type::Int32Ty, Elt), 1244 "tmp", LI); 1245 1246 // If we're done, return this element. 1247 if (NV->getType() == LI->getType() && Offset == 0) 1248 return NV; 1249 } 1250 1251 const IntegerType *NTy = cast<IntegerType>(NV->getType()); 1252 1253 // If this is a big-endian system and the load is narrower than the 1254 // full alloca type, we need to do a shift to get the right bits. 1255 int ShAmt = 0; 1256 if (TD->isBigEndian()) { 1257 // On big-endian machines, the lowest bit is stored at the bit offset 1258 // from the pointer given by getTypeStoreSizeInBits. This matters for 1259 // integers with a bitwidth that is not a multiple of 8. 1260 ShAmt = TD->getTypeStoreSizeInBits(NTy) - 1261 TD->getTypeStoreSizeInBits(LI->getType()) - Offset; 1262 } else { 1263 ShAmt = Offset; 1264 } 1265 1266 // Note: we support negative bitwidths (with shl) which are not defined. 1267 // We do this to support (f.e.) loads off the end of a structure where 1268 // only some bits are used. 1269 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1270 NV = BinaryOperator::CreateLShr(NV, 1271 ConstantInt::get(NV->getType(),ShAmt), 1272 LI->getName(), LI); 1273 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1274 NV = BinaryOperator::CreateShl(NV, 1275 ConstantInt::get(NV->getType(),-ShAmt), 1276 LI->getName(), LI); 1277 1278 // Finally, unconditionally truncate the integer to the right width. 1279 unsigned LIBitWidth = TD->getTypeSizeInBits(LI->getType()); 1280 if (LIBitWidth < NTy->getBitWidth()) 1281 NV = new TruncInst(NV, IntegerType::get(LIBitWidth), 1282 LI->getName(), LI); 1283 1284 // If the result is an integer, this is a trunc or bitcast. 1285 if (isa<IntegerType>(LI->getType())) { 1286 // Should be done. 1287 } else if (LI->getType()->isFloatingPoint()) { 1288 // Just do a bitcast, we know the sizes match up. 1289 NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI); 1290 } else { 1291 // Otherwise must be a pointer. 1292 NV = new IntToPtrInst(NV, LI->getType(), LI->getName(), LI); 1293 } 1294 assert(NV->getType() == LI->getType() && "Didn't convert right?"); 1295 return NV; 1296} 1297 1298 1299/// ConvertUsesOfStoreToScalar - Convert the specified store to a load+store 1300/// pair of the new alloca directly, returning the value that should be stored 1301/// to the alloca. This happens when we are converting an "integer union" to a 1302/// single integer scalar, or when we are converting a "vector union" to a 1303/// vector with insert/extractelement instructions. 1304/// 1305/// Offset is an offset from the original alloca, in bits that need to be 1306/// shifted to the right. By the end of this, there should be no uses of Ptr. 1307Value *SROA::ConvertUsesOfStoreToScalar(StoreInst *SI, AllocaInst *NewAI, 1308 unsigned Offset) { 1309 1310 // Convert the stored type to the actual type, shift it left to insert 1311 // then 'or' into place. 1312 Value *SV = SI->getOperand(0); 1313 const Type *AllocaType = NewAI->getType()->getElementType(); 1314 if (SV->getType() == AllocaType && Offset == 0) { 1315 // All is well. 1316 } else if (const VectorType *PTy = dyn_cast<VectorType>(AllocaType)) { 1317 Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI); 1318 1319 // If the result alloca is a vector type, this is either an element 1320 // access or a bitcast to another vector type. 1321 if (isa<VectorType>(SV->getType())) { 1322 SV = new BitCastInst(SV, AllocaType, SV->getName(), SI); 1323 } else { 1324 // Must be an element insertion. 1325 unsigned Elt = Offset/TD->getABITypeSizeInBits(PTy->getElementType()); 1326 SV = InsertElementInst::Create(Old, SV, 1327 ConstantInt::get(Type::Int32Ty, Elt), 1328 "tmp", SI); 1329 } 1330 } else if (isa<PointerType>(AllocaType)) { 1331 // If the alloca type is a pointer, then all the elements must be 1332 // pointers. 1333 if (SV->getType() != AllocaType) 1334 SV = new BitCastInst(SV, AllocaType, SV->getName(), SI); 1335 } else { 1336 Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI); 1337 1338 // If SV is a float, convert it to the appropriate integer type. 1339 // If it is a pointer, do the same, and also handle ptr->ptr casts 1340 // here. 1341 unsigned SrcWidth = TD->getTypeSizeInBits(SV->getType()); 1342 unsigned DestWidth = TD->getTypeSizeInBits(AllocaType); 1343 unsigned SrcStoreWidth = TD->getTypeStoreSizeInBits(SV->getType()); 1344 unsigned DestStoreWidth = TD->getTypeStoreSizeInBits(AllocaType); 1345 if (SV->getType()->isFloatingPoint()) 1346 SV = new BitCastInst(SV, IntegerType::get(SrcWidth), 1347 SV->getName(), SI); 1348 else if (isa<PointerType>(SV->getType())) 1349 SV = new PtrToIntInst(SV, TD->getIntPtrType(), SV->getName(), SI); 1350 1351 // Always zero extend the value if needed. 1352 if (SV->getType() != AllocaType) 1353 SV = new ZExtInst(SV, AllocaType, SV->getName(), SI); 1354 1355 // If this is a big-endian system and the store is narrower than the 1356 // full alloca type, we need to do a shift to get the right bits. 1357 int ShAmt = 0; 1358 if (TD->isBigEndian()) { 1359 // On big-endian machines, the lowest bit is stored at the bit offset 1360 // from the pointer given by getTypeStoreSizeInBits. This matters for 1361 // integers with a bitwidth that is not a multiple of 8. 1362 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1363 } else { 1364 ShAmt = Offset; 1365 } 1366 1367 // Note: we support negative bitwidths (with shr) which are not defined. 1368 // We do this to support (f.e.) stores off the end of a structure where 1369 // only some bits in the structure are set. 1370 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1371 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1372 SV = BinaryOperator::CreateShl(SV, 1373 ConstantInt::get(SV->getType(), ShAmt), 1374 SV->getName(), SI); 1375 Mask <<= ShAmt; 1376 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1377 SV = BinaryOperator::CreateLShr(SV, 1378 ConstantInt::get(SV->getType(),-ShAmt), 1379 SV->getName(), SI); 1380 Mask = Mask.lshr(ShAmt); 1381 } 1382 1383 // Mask out the bits we are about to insert from the old value, and or 1384 // in the new bits. 1385 if (SrcWidth != DestWidth) { 1386 assert(DestWidth > SrcWidth); 1387 Old = BinaryOperator::CreateAnd(Old, ConstantInt::get(~Mask), 1388 Old->getName()+".mask", SI); 1389 SV = BinaryOperator::CreateOr(Old, SV, SV->getName()+".ins", SI); 1390 } 1391 } 1392 return SV; 1393} 1394 1395 1396 1397/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1398/// some part of a constant global variable. This intentionally only accepts 1399/// constant expressions because we don't can't rewrite arbitrary instructions. 1400static bool PointsToConstantGlobal(Value *V) { 1401 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1402 return GV->isConstant(); 1403 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1404 if (CE->getOpcode() == Instruction::BitCast || 1405 CE->getOpcode() == Instruction::GetElementPtr) 1406 return PointsToConstantGlobal(CE->getOperand(0)); 1407 return false; 1408} 1409 1410/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1411/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1412/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1413/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1414/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1415/// the alloca, and if the source pointer is a pointer to a constant global, we 1416/// can optimize this. 1417static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1418 bool isOffset) { 1419 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1420 if (isa<LoadInst>(*UI)) { 1421 // Ignore loads, they are always ok. 1422 continue; 1423 } 1424 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1425 // If uses of the bitcast are ok, we are ok. 1426 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1427 return false; 1428 continue; 1429 } 1430 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1431 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1432 // doesn't, it does. 1433 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1434 isOffset || !GEP->hasAllZeroIndices())) 1435 return false; 1436 continue; 1437 } 1438 1439 // If this is isn't our memcpy/memmove, reject it as something we can't 1440 // handle. 1441 if (!isa<MemCpyInst>(*UI) && !isa<MemMoveInst>(*UI)) 1442 return false; 1443 1444 // If we already have seen a copy, reject the second one. 1445 if (TheCopy) return false; 1446 1447 // If the pointer has been offset from the start of the alloca, we can't 1448 // safely handle this. 1449 if (isOffset) return false; 1450 1451 // If the memintrinsic isn't using the alloca as the dest, reject it. 1452 if (UI.getOperandNo() != 1) return false; 1453 1454 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1455 1456 // If the source of the memcpy/move is not a constant global, reject it. 1457 if (!PointsToConstantGlobal(MI->getOperand(2))) 1458 return false; 1459 1460 // Otherwise, the transform is safe. Remember the copy instruction. 1461 TheCopy = MI; 1462 } 1463 return true; 1464} 1465 1466/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1467/// modified by a copy from a constant global. If we can prove this, we can 1468/// replace any uses of the alloca with uses of the global directly. 1469Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) { 1470 Instruction *TheCopy = 0; 1471 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1472 return TheCopy; 1473 return 0; 1474} 1475