ScalarReplAggregates.cpp revision d356a7ee0ed7744857dcf497cb20b0128770fb0f
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 /// AllocaInfo - When analyzing uses of an alloca instruction, this captures 74 /// information about the uses. All these fields are initialized to false 75 /// and set to true when something is learned. 76 struct AllocaInfo { 77 /// isUnsafe - This is set to true if the alloca cannot be SROA'd. 78 bool isUnsafe : 1; 79 80 /// needsCanon - This is set to true if there is some use of the alloca 81 /// that requires canonicalization. 82 bool needsCanon : 1; 83 84 /// isMemCpySrc - This is true if this aggregate is memcpy'd from. 85 bool isMemCpySrc : 1; 86 87 /// isMemCpyDst - This is true if this aggregate is memcpy'd into. 88 bool isMemCpyDst : 1; 89 90 AllocaInfo() 91 : isUnsafe(false), needsCanon(false), 92 isMemCpySrc(false), isMemCpyDst(false) {} 93 }; 94 95 unsigned SRThreshold; 96 97 void MarkUnsafe(AllocaInfo &I) { I.isUnsafe = true; } 98 99 int isSafeAllocaToScalarRepl(AllocationInst *AI); 100 101 void isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 102 AllocaInfo &Info); 103 void isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 104 AllocaInfo &Info); 105 void isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 106 unsigned OpNo, AllocaInfo &Info); 107 void isSafeUseOfBitCastedAllocation(BitCastInst *User, AllocationInst *AI, 108 AllocaInfo &Info); 109 110 void DoScalarReplacement(AllocationInst *AI, 111 std::vector<AllocationInst*> &WorkList); 112 void CanonicalizeAllocaUsers(AllocationInst *AI); 113 AllocaInst *AddNewAlloca(Function &F, const Type *Ty, AllocationInst *Base); 114 115 void RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 116 SmallVector<AllocaInst*, 32> &NewElts); 117 118 const Type *CanConvertToScalar(Value *V, bool &IsNotTrivial); 119 void ConvertToScalar(AllocationInst *AI, const Type *Ty); 120 void ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset); 121 Value *ConvertUsesOfLoadToScalar(LoadInst *LI, AllocaInst *NewAI, 122 unsigned Offset); 123 Value *ConvertUsesOfStoreToScalar(StoreInst *SI, AllocaInst *NewAI, 124 unsigned Offset); 125 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI); 126 }; 127} 128 129char SROA::ID = 0; 130static RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 131 132// Public interface to the ScalarReplAggregates pass 133FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 134 return new SROA(Threshold); 135} 136 137 138bool SROA::runOnFunction(Function &F) { 139 bool Changed = performPromotion(F); 140 while (1) { 141 bool LocalChange = performScalarRepl(F); 142 if (!LocalChange) break; // No need to repromote if no scalarrepl 143 Changed = true; 144 LocalChange = performPromotion(F); 145 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 146 } 147 148 return Changed; 149} 150 151 152bool SROA::performPromotion(Function &F) { 153 std::vector<AllocaInst*> Allocas; 154 DominatorTree &DT = getAnalysis<DominatorTree>(); 155 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 156 157 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 158 159 bool Changed = false; 160 161 while (1) { 162 Allocas.clear(); 163 164 // Find allocas that are safe to promote, by looking at all instructions in 165 // the entry node 166 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 167 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 168 if (isAllocaPromotable(AI)) 169 Allocas.push_back(AI); 170 171 if (Allocas.empty()) break; 172 173 PromoteMemToReg(Allocas, DT, DF); 174 NumPromoted += Allocas.size(); 175 Changed = true; 176 } 177 178 return Changed; 179} 180 181/// getNumSAElements - Return the number of elements in the specific struct or 182/// array. 183static uint64_t getNumSAElements(const Type *T) { 184 if (const StructType *ST = dyn_cast<StructType>(T)) 185 return ST->getNumElements(); 186 return cast<ArrayType>(T)->getNumElements(); 187} 188 189// performScalarRepl - This algorithm is a simple worklist driven algorithm, 190// which runs on all of the malloc/alloca instructions in the function, removing 191// them if they are only used by getelementptr instructions. 192// 193bool SROA::performScalarRepl(Function &F) { 194 std::vector<AllocationInst*> WorkList; 195 196 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 197 BasicBlock &BB = F.getEntryBlock(); 198 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 199 if (AllocationInst *A = dyn_cast<AllocationInst>(I)) 200 WorkList.push_back(A); 201 202 const TargetData &TD = getAnalysis<TargetData>(); 203 204 // Process the worklist 205 bool Changed = false; 206 while (!WorkList.empty()) { 207 AllocationInst *AI = WorkList.back(); 208 WorkList.pop_back(); 209 210 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 211 // with unused elements. 212 if (AI->use_empty()) { 213 AI->eraseFromParent(); 214 continue; 215 } 216 217 // If we can turn this aggregate value (potentially with casts) into a 218 // simple scalar value that can be mem2reg'd into a register value. 219 bool IsNotTrivial = false; 220 if (const Type *ActualType = CanConvertToScalar(AI, IsNotTrivial)) 221 if (IsNotTrivial && ActualType != Type::VoidTy) { 222 ConvertToScalar(AI, ActualType); 223 Changed = true; 224 continue; 225 } 226 227 // Check to see if we can perform the core SROA transformation. We cannot 228 // transform the allocation instruction if it is an array allocation 229 // (allocations OF arrays are ok though), and an allocation of a scalar 230 // value cannot be decomposed at all. 231 if (!AI->isArrayAllocation() && 232 (isa<StructType>(AI->getAllocatedType()) || 233 isa<ArrayType>(AI->getAllocatedType())) && 234 AI->getAllocatedType()->isSized() && 235 // Do not promote any struct whose size is larger than "128" bytes. 236 TD.getABITypeSize(AI->getAllocatedType()) < SRThreshold && 237 // Do not promote any struct into more than "32" separate vars. 238 getNumSAElements(AI->getAllocatedType()) < SRThreshold/4) { 239 // Check that all of the users of the allocation are capable of being 240 // transformed. 241 switch (isSafeAllocaToScalarRepl(AI)) { 242 default: assert(0 && "Unexpected value!"); 243 case 0: // Not safe to scalar replace. 244 break; 245 case 1: // Safe, but requires cleanup/canonicalizations first 246 CanonicalizeAllocaUsers(AI); 247 // FALL THROUGH. 248 case 3: // Safe to scalar replace. 249 DoScalarReplacement(AI, WorkList); 250 Changed = true; 251 continue; 252 } 253 } 254 255 // Check to see if this allocation is only modified by a memcpy/memmove from 256 // a constant global. If this is the case, we can change all users to use 257 // the constant global instead. This is commonly produced by the CFE by 258 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 259 // is only subsequently read. 260 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 261 DOUT << "Found alloca equal to global: " << *AI; 262 DOUT << " memcpy = " << *TheCopy; 263 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 264 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 265 TheCopy->eraseFromParent(); // Don't mutate the global. 266 AI->eraseFromParent(); 267 ++NumGlobals; 268 Changed = true; 269 continue; 270 } 271 272 // Otherwise, couldn't process this. 273 } 274 275 return Changed; 276} 277 278/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 279/// predicate, do SROA now. 280void SROA::DoScalarReplacement(AllocationInst *AI, 281 std::vector<AllocationInst*> &WorkList) { 282 DOUT << "Found inst to SROA: " << *AI; 283 SmallVector<AllocaInst*, 32> ElementAllocas; 284 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 285 ElementAllocas.reserve(ST->getNumContainedTypes()); 286 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 287 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 288 AI->getAlignment(), 289 AI->getName() + "." + utostr(i), AI); 290 ElementAllocas.push_back(NA); 291 WorkList.push_back(NA); // Add to worklist for recursive processing 292 } 293 } else { 294 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 295 ElementAllocas.reserve(AT->getNumElements()); 296 const Type *ElTy = AT->getElementType(); 297 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 298 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 299 AI->getName() + "." + utostr(i), AI); 300 ElementAllocas.push_back(NA); 301 WorkList.push_back(NA); // Add to worklist for recursive processing 302 } 303 } 304 305 // Now that we have created the alloca instructions that we want to use, 306 // expand the getelementptr instructions to use them. 307 // 308 while (!AI->use_empty()) { 309 Instruction *User = cast<Instruction>(AI->use_back()); 310 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) { 311 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas); 312 BCInst->eraseFromParent(); 313 continue; 314 } 315 316 // Replace: 317 // %res = load { i32, i32 }* %alloc 318 // with: 319 // %load.0 = load i32* %alloc.0 320 // %insert.0 insertvalue { i32, i32 } zeroinitializer, i32 %load.0, 0 321 // %load.1 = load i32* %alloc.1 322 // %insert = insertvalue { i32, i32 } %insert.0, i32 %load.1, 1 323 // (Also works for arrays instead of structs) 324 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 325 Value *Insert = UndefValue::get(LI->getType()); 326 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 327 Value *Load = new LoadInst(ElementAllocas[i], "load", LI); 328 Insert = InsertValueInst::Create(Insert, Load, i, "insert", LI); 329 } 330 LI->replaceAllUsesWith(Insert); 331 LI->eraseFromParent(); 332 continue; 333 } 334 335 // Replace: 336 // store { i32, i32 } %val, { i32, i32 }* %alloc 337 // with: 338 // %val.0 = extractvalue { i32, i32 } %val, 0 339 // store i32 %val.0, i32* %alloc.0 340 // %val.1 = extractvalue { i32, i32 } %val, 1 341 // store i32 %val.1, i32* %alloc.1 342 // (Also works for arrays instead of structs) 343 if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 344 Value *Val = SI->getOperand(0); 345 for (unsigned i = 0, e = ElementAllocas.size(); i != e; ++i) { 346 Value *Extract = ExtractValueInst::Create(Val, i, Val->getName(), SI); 347 new StoreInst(Extract, ElementAllocas[i], SI); 348 } 349 SI->eraseFromParent(); 350 continue; 351 } 352 353 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 354 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 355 unsigned Idx = 356 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 357 358 assert(Idx < ElementAllocas.size() && "Index out of range?"); 359 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 360 361 Value *RepValue; 362 if (GEPI->getNumOperands() == 3) { 363 // Do not insert a new getelementptr instruction with zero indices, only 364 // to have it optimized out later. 365 RepValue = AllocaToUse; 366 } else { 367 // We are indexing deeply into the structure, so we still need a 368 // getelement ptr instruction to finish the indexing. This may be 369 // expanded itself once the worklist is rerun. 370 // 371 SmallVector<Value*, 8> NewArgs; 372 NewArgs.push_back(Constant::getNullValue(Type::Int32Ty)); 373 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end()); 374 RepValue = GetElementPtrInst::Create(AllocaToUse, NewArgs.begin(), 375 NewArgs.end(), "", GEPI); 376 RepValue->takeName(GEPI); 377 } 378 379 // If this GEP is to the start of the aggregate, check for memcpys. 380 if (Idx == 0 && GEPI->hasAllZeroIndices()) 381 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas); 382 383 // Move all of the users over to the new GEP. 384 GEPI->replaceAllUsesWith(RepValue); 385 // Delete the old GEP 386 GEPI->eraseFromParent(); 387 } 388 389 // Finally, delete the Alloca instruction 390 AI->eraseFromParent(); 391 NumReplaced++; 392} 393 394 395/// isSafeElementUse - Check to see if this use is an allowed use for a 396/// getelementptr instruction of an array aggregate allocation. isFirstElt 397/// indicates whether Ptr is known to the start of the aggregate. 398/// 399void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 400 AllocaInfo &Info) { 401 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 402 I != E; ++I) { 403 Instruction *User = cast<Instruction>(*I); 404 switch (User->getOpcode()) { 405 case Instruction::Load: break; 406 case Instruction::Store: 407 // Store is ok if storing INTO the pointer, not storing the pointer 408 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info); 409 break; 410 case Instruction::GetElementPtr: { 411 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 412 bool AreAllZeroIndices = isFirstElt; 413 if (GEP->getNumOperands() > 1) { 414 if (!isa<ConstantInt>(GEP->getOperand(1)) || 415 !cast<ConstantInt>(GEP->getOperand(1))->isZero()) 416 // Using pointer arithmetic to navigate the array. 417 return MarkUnsafe(Info); 418 419 if (AreAllZeroIndices) 420 AreAllZeroIndices = GEP->hasAllZeroIndices(); 421 } 422 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 423 if (Info.isUnsafe) return; 424 break; 425 } 426 case Instruction::BitCast: 427 if (isFirstElt) { 428 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 429 if (Info.isUnsafe) return; 430 break; 431 } 432 DOUT << " Transformation preventing inst: " << *User; 433 return MarkUnsafe(Info); 434 case Instruction::Call: 435 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 436 if (isFirstElt) { 437 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 438 if (Info.isUnsafe) return; 439 break; 440 } 441 } 442 DOUT << " Transformation preventing inst: " << *User; 443 return MarkUnsafe(Info); 444 default: 445 DOUT << " Transformation preventing inst: " << *User; 446 return MarkUnsafe(Info); 447 } 448 } 449 return; // All users look ok :) 450} 451 452/// AllUsersAreLoads - Return true if all users of this value are loads. 453static bool AllUsersAreLoads(Value *Ptr) { 454 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 455 I != E; ++I) 456 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 457 return false; 458 return true; 459} 460 461/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 462/// aggregate allocation. 463/// 464void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 465 AllocaInfo &Info) { 466 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 467 return isSafeUseOfBitCastedAllocation(C, AI, Info); 468 469 if (isa<LoadInst>(User)) 470 return; // Loads (returning a first class aggregrate) are always rewritable 471 472 if (isa<StoreInst>(User) && User->getOperand(0) != AI) 473 return; // Store is ok if storing INTO the pointer, not storing the pointer 474 475 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 476 if (GEPI == 0) 477 return MarkUnsafe(Info); 478 479 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 480 481 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 482 if (I == E || 483 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) { 484 return MarkUnsafe(Info); 485 } 486 487 ++I; 488 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 489 490 bool IsAllZeroIndices = true; 491 492 // If the first index is a non-constant index into an array, see if we can 493 // handle it as a special case. 494 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 495 if (!isa<ConstantInt>(I.getOperand())) { 496 IsAllZeroIndices = 0; 497 uint64_t NumElements = AT->getNumElements(); 498 499 // If this is an array index and the index is not constant, we cannot 500 // promote... that is unless the array has exactly one or two elements in 501 // it, in which case we CAN promote it, but we have to canonicalize this 502 // out if this is the only problem. 503 if ((NumElements == 1 || NumElements == 2) && 504 AllUsersAreLoads(GEPI)) { 505 Info.needsCanon = true; 506 return; // Canonicalization required! 507 } 508 return MarkUnsafe(Info); 509 } 510 } 511 512 // Walk through the GEP type indices, checking the types that this indexes 513 // into. 514 for (; I != E; ++I) { 515 // Ignore struct elements, no extra checking needed for these. 516 if (isa<StructType>(*I)) 517 continue; 518 519 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 520 if (!IdxVal) return MarkUnsafe(Info); 521 522 // Are all indices still zero? 523 IsAllZeroIndices &= IdxVal->isZero(); 524 525 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 526 // This GEP indexes an array. Verify that this is an in-range constant 527 // integer. Specifically, consider A[0][i]. We cannot know that the user 528 // isn't doing invalid things like allowing i to index an out-of-range 529 // subscript that accesses A[1]. Because of this, we have to reject SROA 530 // of any accesses into structs where any of the components are variables. 531 if (IdxVal->getZExtValue() >= AT->getNumElements()) 532 return MarkUnsafe(Info); 533 } else if (const VectorType *VT = dyn_cast<VectorType>(*I)) { 534 if (IdxVal->getZExtValue() >= VT->getNumElements()) 535 return MarkUnsafe(Info); 536 } 537 } 538 539 // If there are any non-simple uses of this getelementptr, make sure to reject 540 // them. 541 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 542} 543 544/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory 545/// intrinsic can be promoted by SROA. At this point, we know that the operand 546/// of the memintrinsic is a pointer to the beginning of the allocation. 547void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 548 unsigned OpNo, AllocaInfo &Info) { 549 // If not constant length, give up. 550 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 551 if (!Length) return MarkUnsafe(Info); 552 553 // If not the whole aggregate, give up. 554 const TargetData &TD = getAnalysis<TargetData>(); 555 if (Length->getZExtValue() != 556 TD.getABITypeSize(AI->getType()->getElementType())) 557 return MarkUnsafe(Info); 558 559 // We only know about memcpy/memset/memmove. 560 if (!isa<MemCpyInst>(MI) && !isa<MemSetInst>(MI) && !isa<MemMoveInst>(MI)) 561 return MarkUnsafe(Info); 562 563 // Otherwise, we can transform it. Determine whether this is a memcpy/set 564 // into or out of the aggregate. 565 if (OpNo == 1) 566 Info.isMemCpyDst = true; 567 else { 568 assert(OpNo == 2); 569 Info.isMemCpySrc = true; 570 } 571} 572 573/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast 574/// are 575void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI, 576 AllocaInfo &Info) { 577 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 578 UI != E; ++UI) { 579 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 580 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 581 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 582 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 583 } else { 584 return MarkUnsafe(Info); 585 } 586 if (Info.isUnsafe) return; 587 } 588} 589 590/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 591/// to its first element. Transform users of the cast to use the new values 592/// instead. 593void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 594 SmallVector<AllocaInst*, 32> &NewElts) { 595 Constant *Zero = Constant::getNullValue(Type::Int32Ty); 596 const TargetData &TD = getAnalysis<TargetData>(); 597 598 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 599 while (UI != UE) { 600 if (BitCastInst *BCU = dyn_cast<BitCastInst>(*UI)) { 601 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 602 ++UI; 603 BCU->eraseFromParent(); 604 continue; 605 } 606 607 // Otherwise, must be memcpy/memmove/memset of the entire aggregate. Split 608 // into one per element. 609 MemIntrinsic *MI = dyn_cast<MemIntrinsic>(*UI); 610 611 // If it's not a mem intrinsic, it must be some other user of a gep of the 612 // first pointer. Just leave these alone. 613 if (!MI) { 614 ++UI; 615 continue; 616 } 617 618 // If this is a memcpy/memmove, construct the other pointer as the 619 // appropriate type. 620 Value *OtherPtr = 0; 621 if (MemCpyInst *MCI = dyn_cast<MemCpyInst>(MI)) { 622 if (BCInst == MCI->getRawDest()) 623 OtherPtr = MCI->getRawSource(); 624 else { 625 assert(BCInst == MCI->getRawSource()); 626 OtherPtr = MCI->getRawDest(); 627 } 628 } else if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) { 629 if (BCInst == MMI->getRawDest()) 630 OtherPtr = MMI->getRawSource(); 631 else { 632 assert(BCInst == MMI->getRawSource()); 633 OtherPtr = MMI->getRawDest(); 634 } 635 } 636 637 // If there is an other pointer, we want to convert it to the same pointer 638 // type as AI has, so we can GEP through it. 639 if (OtherPtr) { 640 // It is likely that OtherPtr is a bitcast, if so, remove it. 641 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 642 OtherPtr = BC->getOperand(0); 643 // All zero GEPs are effectively bitcasts. 644 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(OtherPtr)) 645 if (GEP->hasAllZeroIndices()) 646 OtherPtr = GEP->getOperand(0); 647 648 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 649 if (BCE->getOpcode() == Instruction::BitCast) 650 OtherPtr = BCE->getOperand(0); 651 652 // If the pointer is not the right type, insert a bitcast to the right 653 // type. 654 if (OtherPtr->getType() != AI->getType()) 655 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 656 MI); 657 } 658 659 // Process each element of the aggregate. 660 Value *TheFn = MI->getOperand(0); 661 const Type *BytePtrTy = MI->getRawDest()->getType(); 662 bool SROADest = MI->getRawDest() == BCInst; 663 664 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 665 // If this is a memcpy/memmove, emit a GEP of the other element address. 666 Value *OtherElt = 0; 667 if (OtherPtr) { 668 Value *Idx[2] = { Zero, ConstantInt::get(Type::Int32Ty, i) }; 669 OtherElt = GetElementPtrInst::Create(OtherPtr, Idx, Idx + 2, 670 OtherPtr->getNameStr()+"."+utostr(i), 671 MI); 672 } 673 674 Value *EltPtr = NewElts[i]; 675 const Type *EltTy =cast<PointerType>(EltPtr->getType())->getElementType(); 676 677 // If we got down to a scalar, insert a load or store as appropriate. 678 if (EltTy->isSingleValueType()) { 679 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) { 680 Value *Elt = new LoadInst(SROADest ? OtherElt : EltPtr, "tmp", 681 MI); 682 new StoreInst(Elt, SROADest ? EltPtr : OtherElt, MI); 683 continue; 684 } else { 685 assert(isa<MemSetInst>(MI)); 686 687 // If the stored element is zero (common case), just store a null 688 // constant. 689 Constant *StoreVal; 690 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 691 if (CI->isZero()) { 692 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 693 } else { 694 // If EltTy is a vector type, get the element type. 695 const Type *ValTy = EltTy; 696 if (const VectorType *VTy = dyn_cast<VectorType>(ValTy)) 697 ValTy = VTy->getElementType(); 698 699 // Construct an integer with the right value. 700 unsigned EltSize = TD.getTypeSizeInBits(ValTy); 701 APInt OneVal(EltSize, CI->getZExtValue()); 702 APInt TotalVal(OneVal); 703 // Set each byte. 704 for (unsigned i = 0; 8*i < EltSize; ++i) { 705 TotalVal = TotalVal.shl(8); 706 TotalVal |= OneVal; 707 } 708 709 // Convert the integer value to the appropriate type. 710 StoreVal = ConstantInt::get(TotalVal); 711 if (isa<PointerType>(ValTy)) 712 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 713 else if (ValTy->isFloatingPoint()) 714 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 715 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 716 717 // If the requested value was a vector constant, create it. 718 if (EltTy != ValTy) { 719 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 720 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 721 StoreVal = ConstantVector::get(&Elts[0], NumElts); 722 } 723 } 724 new StoreInst(StoreVal, EltPtr, MI); 725 continue; 726 } 727 // Otherwise, if we're storing a byte variable, use a memset call for 728 // this element. 729 } 730 } 731 732 // Cast the element pointer to BytePtrTy. 733 if (EltPtr->getType() != BytePtrTy) 734 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 735 736 // Cast the other pointer (if we have one) to BytePtrTy. 737 if (OtherElt && OtherElt->getType() != BytePtrTy) 738 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 739 MI); 740 741 unsigned EltSize = TD.getABITypeSize(EltTy); 742 743 // Finally, insert the meminst for this element. 744 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) { 745 Value *Ops[] = { 746 SROADest ? EltPtr : OtherElt, // Dest ptr 747 SROADest ? OtherElt : EltPtr, // Src ptr 748 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 749 Zero // Align 750 }; 751 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 752 } else { 753 assert(isa<MemSetInst>(MI)); 754 Value *Ops[] = { 755 EltPtr, MI->getOperand(2), // Dest, Value, 756 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 757 Zero // Align 758 }; 759 CallInst::Create(TheFn, Ops, Ops + 4, "", MI); 760 } 761 } 762 763 // Finally, MI is now dead, as we've modified its actions to occur on all of 764 // the elements of the aggregate. 765 ++UI; 766 MI->eraseFromParent(); 767 } 768} 769 770/// HasPadding - Return true if the specified type has any structure or 771/// alignment padding, false otherwise. 772static bool HasPadding(const Type *Ty, const TargetData &TD) { 773 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 774 const StructLayout *SL = TD.getStructLayout(STy); 775 unsigned PrevFieldBitOffset = 0; 776 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 777 unsigned FieldBitOffset = SL->getElementOffsetInBits(i); 778 779 // Padding in sub-elements? 780 if (HasPadding(STy->getElementType(i), TD)) 781 return true; 782 783 // Check to see if there is any padding between this element and the 784 // previous one. 785 if (i) { 786 unsigned PrevFieldEnd = 787 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 788 if (PrevFieldEnd < FieldBitOffset) 789 return true; 790 } 791 792 PrevFieldBitOffset = FieldBitOffset; 793 } 794 795 // Check for tail padding. 796 if (unsigned EltCount = STy->getNumElements()) { 797 unsigned PrevFieldEnd = PrevFieldBitOffset + 798 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 799 if (PrevFieldEnd < SL->getSizeInBits()) 800 return true; 801 } 802 803 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 804 return HasPadding(ATy->getElementType(), TD); 805 } else if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) { 806 return HasPadding(VTy->getElementType(), TD); 807 } 808 return TD.getTypeSizeInBits(Ty) != TD.getABITypeSizeInBits(Ty); 809} 810 811/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 812/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 813/// or 1 if safe after canonicalization has been performed. 814/// 815int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 816 // Loop over the use list of the alloca. We can only transform it if all of 817 // the users are safe to transform. 818 AllocaInfo Info; 819 820 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 821 I != E; ++I) { 822 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 823 if (Info.isUnsafe) { 824 DOUT << "Cannot transform: " << *AI << " due to user: " << **I; 825 return 0; 826 } 827 } 828 829 // Okay, we know all the users are promotable. If the aggregate is a memcpy 830 // source and destination, we have to be careful. In particular, the memcpy 831 // could be moving around elements that live in structure padding of the LLVM 832 // types, but may actually be used. In these cases, we refuse to promote the 833 // struct. 834 if (Info.isMemCpySrc && Info.isMemCpyDst && 835 HasPadding(AI->getType()->getElementType(), getAnalysis<TargetData>())) 836 return 0; 837 838 // If we require cleanup, return 1, otherwise return 3. 839 return Info.needsCanon ? 1 : 3; 840} 841 842/// CanonicalizeAllocaUsers - If SROA reported that it can promote the specified 843/// allocation, but only if cleaned up, perform the cleanups required. 844void SROA::CanonicalizeAllocaUsers(AllocationInst *AI) { 845 // At this point, we know that the end result will be SROA'd and promoted, so 846 // we can insert ugly code if required so long as sroa+mem2reg will clean it 847 // up. 848 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 849 UI != E; ) { 850 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI++); 851 if (!GEPI) continue; 852 gep_type_iterator I = gep_type_begin(GEPI); 853 ++I; 854 855 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 856 uint64_t NumElements = AT->getNumElements(); 857 858 if (!isa<ConstantInt>(I.getOperand())) { 859 if (NumElements == 1) { 860 GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty)); 861 } else { 862 assert(NumElements == 2 && "Unhandled case!"); 863 // All users of the GEP must be loads. At each use of the GEP, insert 864 // two loads of the appropriate indexed GEP and select between them. 865 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(), 866 Constant::getNullValue(I.getOperand()->getType()), 867 "isone", GEPI); 868 // Insert the new GEP instructions, which are properly indexed. 869 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 870 Indices[1] = Constant::getNullValue(Type::Int32Ty); 871 Value *ZeroIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 872 Indices.begin(), 873 Indices.end(), 874 GEPI->getName()+".0", GEPI); 875 Indices[1] = ConstantInt::get(Type::Int32Ty, 1); 876 Value *OneIdx = GetElementPtrInst::Create(GEPI->getOperand(0), 877 Indices.begin(), 878 Indices.end(), 879 GEPI->getName()+".1", GEPI); 880 // Replace all loads of the variable index GEP with loads from both 881 // indexes and a select. 882 while (!GEPI->use_empty()) { 883 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 884 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 885 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 886 Value *R = SelectInst::Create(IsOne, One, Zero, LI->getName(), LI); 887 LI->replaceAllUsesWith(R); 888 LI->eraseFromParent(); 889 } 890 GEPI->eraseFromParent(); 891 } 892 } 893 } 894 } 895} 896 897/// MergeInType - Add the 'In' type to the accumulated type so far. If the 898/// types are incompatible, return true, otherwise update Accum and return 899/// false. 900/// 901/// There are three cases we handle here: 902/// 1) An effectively-integer union, where the pieces are stored into as 903/// smaller integers (common with byte swap and other idioms). 904/// 2) A union of vector types of the same size and potentially its elements. 905/// Here we turn element accesses into insert/extract element operations. 906/// 3) A union of scalar types, such as int/float or int/pointer. Here we 907/// merge together into integers, allowing the xform to work with #1 as 908/// well. 909static bool MergeInType(const Type *In, const Type *&Accum, 910 const TargetData &TD) { 911 // If this is our first type, just use it. 912 const VectorType *PTy; 913 if (Accum == Type::VoidTy || In == Accum) { 914 Accum = In; 915 } else if (In == Type::VoidTy) { 916 // Noop. 917 } else if (In->isInteger() && Accum->isInteger()) { // integer union. 918 // Otherwise pick whichever type is larger. 919 if (cast<IntegerType>(In)->getBitWidth() > 920 cast<IntegerType>(Accum)->getBitWidth()) 921 Accum = In; 922 } else if (isa<PointerType>(In) && isa<PointerType>(Accum)) { 923 // Pointer unions just stay as one of the pointers. 924 } else if (isa<VectorType>(In) || isa<VectorType>(Accum)) { 925 if ((PTy = dyn_cast<VectorType>(Accum)) && 926 PTy->getElementType() == In) { 927 // Accum is a vector, and we are accessing an element: ok. 928 } else if ((PTy = dyn_cast<VectorType>(In)) && 929 PTy->getElementType() == Accum) { 930 // In is a vector, and accum is an element: ok, remember In. 931 Accum = In; 932 } else if ((PTy = dyn_cast<VectorType>(In)) && isa<VectorType>(Accum) && 933 PTy->getBitWidth() == cast<VectorType>(Accum)->getBitWidth()) { 934 // Two vectors of the same size: keep Accum. 935 } else { 936 // Cannot insert an short into a <4 x int> or handle 937 // <2 x int> -> <4 x int> 938 return true; 939 } 940 } else { 941 // Pointer/FP/Integer unions merge together as integers. 942 switch (Accum->getTypeID()) { 943 case Type::PointerTyID: Accum = TD.getIntPtrType(); break; 944 case Type::FloatTyID: Accum = Type::Int32Ty; break; 945 case Type::DoubleTyID: Accum = Type::Int64Ty; break; 946 case Type::X86_FP80TyID: return true; 947 case Type::FP128TyID: return true; 948 case Type::PPC_FP128TyID: return true; 949 default: 950 assert(Accum->isInteger() && "Unknown FP type!"); 951 break; 952 } 953 954 switch (In->getTypeID()) { 955 case Type::PointerTyID: In = TD.getIntPtrType(); break; 956 case Type::FloatTyID: In = Type::Int32Ty; break; 957 case Type::DoubleTyID: In = Type::Int64Ty; break; 958 case Type::X86_FP80TyID: return true; 959 case Type::FP128TyID: return true; 960 case Type::PPC_FP128TyID: return true; 961 default: 962 assert(In->isInteger() && "Unknown FP type!"); 963 break; 964 } 965 return MergeInType(In, Accum, TD); 966 } 967 return false; 968} 969 970/// getUIntAtLeastAsBigAs - Return an unsigned integer type that is at least 971/// as big as the specified type. If there is no suitable type, this returns 972/// null. 973const Type *getUIntAtLeastAsBigAs(unsigned NumBits) { 974 if (NumBits > 64) return 0; 975 if (NumBits > 32) return Type::Int64Ty; 976 if (NumBits > 16) return Type::Int32Ty; 977 if (NumBits > 8) return Type::Int16Ty; 978 return Type::Int8Ty; 979} 980 981/// CanConvertToScalar - V is a pointer. If we can convert the pointee to a 982/// single scalar integer type, return that type. Further, if the use is not 983/// a completely trivial use that mem2reg could promote, set IsNotTrivial. If 984/// there are no uses of this pointer, return Type::VoidTy to differentiate from 985/// failure. 986/// 987const Type *SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial) { 988 const Type *UsedType = Type::VoidTy; // No uses, no forced type. 989 const TargetData &TD = getAnalysis<TargetData>(); 990 const PointerType *PTy = cast<PointerType>(V->getType()); 991 992 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 993 Instruction *User = cast<Instruction>(*UI); 994 995 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 996 // FIXME: Loads of a first class aggregrate value could be converted to a 997 // series of loads and insertvalues 998 if (!LI->getType()->isSingleValueType()) 999 return 0; 1000 1001 if (MergeInType(LI->getType(), UsedType, TD)) 1002 return 0; 1003 1004 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1005 // Storing the pointer, not into the value? 1006 if (SI->getOperand(0) == V) return 0; 1007 1008 // FIXME: Stores of a first class aggregrate value could be converted to a 1009 // series of extractvalues and stores 1010 if (!SI->getOperand(0)->getType()->isSingleValueType()) 1011 return 0; 1012 1013 // NOTE: We could handle storing of FP imms into integers here! 1014 1015 if (MergeInType(SI->getOperand(0)->getType(), UsedType, TD)) 1016 return 0; 1017 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1018 IsNotTrivial = true; 1019 const Type *SubTy = CanConvertToScalar(CI, IsNotTrivial); 1020 if (!SubTy || MergeInType(SubTy, UsedType, TD)) return 0; 1021 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1022 // Check to see if this is stepping over an element: GEP Ptr, int C 1023 if (GEP->getNumOperands() == 2 && isa<ConstantInt>(GEP->getOperand(1))) { 1024 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue(); 1025 unsigned ElSize = TD.getABITypeSize(PTy->getElementType()); 1026 unsigned BitOffset = Idx*ElSize*8; 1027 if (BitOffset > 64 || !isPowerOf2_32(ElSize)) return 0; 1028 1029 IsNotTrivial = true; 1030 const Type *SubElt = CanConvertToScalar(GEP, IsNotTrivial); 1031 if (SubElt == 0) return 0; 1032 if (SubElt != Type::VoidTy && SubElt->isInteger()) { 1033 const Type *NewTy = 1034 getUIntAtLeastAsBigAs(TD.getABITypeSizeInBits(SubElt)+BitOffset); 1035 if (NewTy == 0 || MergeInType(NewTy, UsedType, TD)) return 0; 1036 continue; 1037 } 1038 } else if (GEP->getNumOperands() == 3 && 1039 isa<ConstantInt>(GEP->getOperand(1)) && 1040 isa<ConstantInt>(GEP->getOperand(2)) && 1041 cast<ConstantInt>(GEP->getOperand(1))->isZero()) { 1042 // We are stepping into an element, e.g. a structure or an array: 1043 // GEP Ptr, i32 0, i32 Cst 1044 const Type *AggTy = PTy->getElementType(); 1045 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 1046 1047 if (const ArrayType *ATy = dyn_cast<ArrayType>(AggTy)) { 1048 if (Idx >= ATy->getNumElements()) return 0; // Out of range. 1049 } else if (const VectorType *VectorTy = dyn_cast<VectorType>(AggTy)) { 1050 // Getting an element of the vector. 1051 if (Idx >= VectorTy->getNumElements()) return 0; // Out of range. 1052 1053 // Merge in the vector type. 1054 if (MergeInType(VectorTy, UsedType, TD)) return 0; 1055 1056 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial); 1057 if (SubTy == 0) return 0; 1058 1059 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, TD)) 1060 return 0; 1061 1062 // We'll need to change this to an insert/extract element operation. 1063 IsNotTrivial = true; 1064 continue; // Everything looks ok 1065 1066 } else if (isa<StructType>(AggTy)) { 1067 // Structs are always ok. 1068 } else { 1069 return 0; 1070 } 1071 const Type *NTy = getUIntAtLeastAsBigAs(TD.getABITypeSizeInBits(AggTy)); 1072 if (NTy == 0 || MergeInType(NTy, UsedType, TD)) return 0; 1073 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial); 1074 if (SubTy == 0) return 0; 1075 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, TD)) 1076 return 0; 1077 continue; // Everything looks ok 1078 } 1079 return 0; 1080 } else { 1081 // Cannot handle this! 1082 return 0; 1083 } 1084 } 1085 1086 return UsedType; 1087} 1088 1089/// ConvertToScalar - The specified alloca passes the CanConvertToScalar 1090/// predicate and is non-trivial. Convert it to something that can be trivially 1091/// promoted into a register by mem2reg. 1092void SROA::ConvertToScalar(AllocationInst *AI, const Type *ActualTy) { 1093 DOUT << "CONVERT TO SCALAR: " << *AI << " TYPE = " 1094 << *ActualTy << "\n"; 1095 ++NumConverted; 1096 1097 BasicBlock *EntryBlock = AI->getParent(); 1098 assert(EntryBlock == &EntryBlock->getParent()->getEntryBlock() && 1099 "Not in the entry block!"); 1100 EntryBlock->getInstList().remove(AI); // Take the alloca out of the program. 1101 1102 // Create and insert the alloca. 1103 AllocaInst *NewAI = new AllocaInst(ActualTy, 0, AI->getName(), 1104 EntryBlock->begin()); 1105 ConvertUsesToScalar(AI, NewAI, 0); 1106 delete AI; 1107} 1108 1109 1110/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1111/// directly. This happens when we are converting an "integer union" to a 1112/// single integer scalar, or when we are converting a "vector union" to a 1113/// vector with insert/extractelement instructions. 1114/// 1115/// Offset is an offset from the original alloca, in bits that need to be 1116/// shifted to the right. By the end of this, there should be no uses of Ptr. 1117void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset) { 1118 while (!Ptr->use_empty()) { 1119 Instruction *User = cast<Instruction>(Ptr->use_back()); 1120 1121 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1122 Value *NV = ConvertUsesOfLoadToScalar(LI, NewAI, Offset); 1123 LI->replaceAllUsesWith(NV); 1124 LI->eraseFromParent(); 1125 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1126 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1127 1128 Value *SV = ConvertUsesOfStoreToScalar(SI, NewAI, Offset); 1129 new StoreInst(SV, NewAI, SI); 1130 SI->eraseFromParent(); 1131 1132 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1133 ConvertUsesToScalar(CI, NewAI, Offset); 1134 CI->eraseFromParent(); 1135 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1136 const PointerType *AggPtrTy = 1137 cast<PointerType>(GEP->getOperand(0)->getType()); 1138 const TargetData &TD = getAnalysis<TargetData>(); 1139 unsigned AggSizeInBits = 1140 TD.getABITypeSizeInBits(AggPtrTy->getElementType()); 1141 1142 // Check to see if this is stepping over an element: GEP Ptr, int C 1143 unsigned NewOffset = Offset; 1144 if (GEP->getNumOperands() == 2) { 1145 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue(); 1146 unsigned BitOffset = Idx*AggSizeInBits; 1147 1148 NewOffset += BitOffset; 1149 } else if (GEP->getNumOperands() == 3) { 1150 // We know that operand #2 is zero. 1151 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 1152 const Type *AggTy = AggPtrTy->getElementType(); 1153 if (const SequentialType *SeqTy = dyn_cast<SequentialType>(AggTy)) { 1154 unsigned ElSizeBits = 1155 TD.getABITypeSizeInBits(SeqTy->getElementType()); 1156 1157 NewOffset += ElSizeBits*Idx; 1158 } else if (const StructType *STy = dyn_cast<StructType>(AggTy)) { 1159 unsigned EltBitOffset = 1160 TD.getStructLayout(STy)->getElementOffsetInBits(Idx); 1161 1162 NewOffset += EltBitOffset; 1163 } else { 1164 assert(0 && "Unsupported operation!"); 1165 abort(); 1166 } 1167 } else { 1168 assert(0 && "Unsupported operation!"); 1169 abort(); 1170 } 1171 ConvertUsesToScalar(GEP, NewAI, NewOffset); 1172 GEP->eraseFromParent(); 1173 } else { 1174 assert(0 && "Unsupported operation!"); 1175 abort(); 1176 } 1177 } 1178} 1179 1180/// ConvertUsesOfLoadToScalar - Convert all of the users the specified load to 1181/// use the new alloca directly, returning the value that should replace the 1182/// load. This happens when we are converting an "integer union" to a 1183/// single integer scalar, or when we are converting a "vector union" to a 1184/// vector with insert/extractelement instructions. 1185/// 1186/// Offset is an offset from the original alloca, in bits that need to be 1187/// shifted to the right. By the end of this, there should be no uses of Ptr. 1188Value *SROA::ConvertUsesOfLoadToScalar(LoadInst *LI, AllocaInst *NewAI, 1189 unsigned Offset) { 1190 // The load is a bit extract from NewAI shifted right by Offset bits. 1191 Value *NV = new LoadInst(NewAI, LI->getName(), LI); 1192 1193 if (NV->getType() == LI->getType() && Offset == 0) { 1194 // We win, no conversion needed. 1195 return NV; 1196 } 1197 1198 // If the result type of the 'union' is a pointer, then this must be ptr->ptr 1199 // cast. Anything else would result in NV being an integer. 1200 if (isa<PointerType>(NV->getType())) { 1201 assert(isa<PointerType>(LI->getType())); 1202 return new BitCastInst(NV, LI->getType(), LI->getName(), LI); 1203 } 1204 1205 if (const VectorType *VTy = dyn_cast<VectorType>(NV->getType())) { 1206 // If the result alloca is a vector type, this is either an element 1207 // access or a bitcast to another vector type. 1208 if (isa<VectorType>(LI->getType())) 1209 return new BitCastInst(NV, LI->getType(), LI->getName(), LI); 1210 1211 // Otherwise it must be an element access. 1212 const TargetData &TD = getAnalysis<TargetData>(); 1213 unsigned Elt = 0; 1214 if (Offset) { 1215 unsigned EltSize = TD.getABITypeSizeInBits(VTy->getElementType()); 1216 Elt = Offset/EltSize; 1217 Offset -= EltSize*Elt; 1218 } 1219 NV = new ExtractElementInst(NV, ConstantInt::get(Type::Int32Ty, Elt), 1220 "tmp", LI); 1221 1222 // If we're done, return this element. 1223 if (NV->getType() == LI->getType() && Offset == 0) 1224 return NV; 1225 } 1226 1227 const IntegerType *NTy = cast<IntegerType>(NV->getType()); 1228 1229 // If this is a big-endian system and the load is narrower than the 1230 // full alloca type, we need to do a shift to get the right bits. 1231 int ShAmt = 0; 1232 const TargetData &TD = getAnalysis<TargetData>(); 1233 if (TD.isBigEndian()) { 1234 // On big-endian machines, the lowest bit is stored at the bit offset 1235 // from the pointer given by getTypeStoreSizeInBits. This matters for 1236 // integers with a bitwidth that is not a multiple of 8. 1237 ShAmt = TD.getTypeStoreSizeInBits(NTy) - 1238 TD.getTypeStoreSizeInBits(LI->getType()) - Offset; 1239 } else { 1240 ShAmt = Offset; 1241 } 1242 1243 // Note: we support negative bitwidths (with shl) which are not defined. 1244 // We do this to support (f.e.) loads off the end of a structure where 1245 // only some bits are used. 1246 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1247 NV = BinaryOperator::CreateLShr(NV, 1248 ConstantInt::get(NV->getType(),ShAmt), 1249 LI->getName(), LI); 1250 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1251 NV = BinaryOperator::CreateShl(NV, 1252 ConstantInt::get(NV->getType(),-ShAmt), 1253 LI->getName(), LI); 1254 1255 // Finally, unconditionally truncate the integer to the right width. 1256 unsigned LIBitWidth = TD.getTypeSizeInBits(LI->getType()); 1257 if (LIBitWidth < NTy->getBitWidth()) 1258 NV = new TruncInst(NV, IntegerType::get(LIBitWidth), 1259 LI->getName(), LI); 1260 1261 // If the result is an integer, this is a trunc or bitcast. 1262 if (isa<IntegerType>(LI->getType())) { 1263 // Should be done. 1264 } else if (LI->getType()->isFloatingPoint()) { 1265 // Just do a bitcast, we know the sizes match up. 1266 NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI); 1267 } else { 1268 // Otherwise must be a pointer. 1269 NV = new IntToPtrInst(NV, LI->getType(), LI->getName(), LI); 1270 } 1271 assert(NV->getType() == LI->getType() && "Didn't convert right?"); 1272 return NV; 1273} 1274 1275 1276/// ConvertUsesOfStoreToScalar - Convert the specified store to a load+store 1277/// pair of the new alloca directly, returning the value that should be stored 1278/// to the alloca. This happens when we are converting an "integer union" to a 1279/// single integer scalar, or when we are converting a "vector union" to a 1280/// vector with insert/extractelement instructions. 1281/// 1282/// Offset is an offset from the original alloca, in bits that need to be 1283/// shifted to the right. By the end of this, there should be no uses of Ptr. 1284Value *SROA::ConvertUsesOfStoreToScalar(StoreInst *SI, AllocaInst *NewAI, 1285 unsigned Offset) { 1286 1287 // Convert the stored type to the actual type, shift it left to insert 1288 // then 'or' into place. 1289 Value *SV = SI->getOperand(0); 1290 const Type *AllocaType = NewAI->getType()->getElementType(); 1291 if (SV->getType() == AllocaType && Offset == 0) { 1292 // All is well. 1293 } else if (const VectorType *PTy = dyn_cast<VectorType>(AllocaType)) { 1294 Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI); 1295 1296 // If the result alloca is a vector type, this is either an element 1297 // access or a bitcast to another vector type. 1298 if (isa<VectorType>(SV->getType())) { 1299 SV = new BitCastInst(SV, AllocaType, SV->getName(), SI); 1300 } else { 1301 // Must be an element insertion. 1302 const TargetData &TD = getAnalysis<TargetData>(); 1303 unsigned Elt = Offset/TD.getABITypeSizeInBits(PTy->getElementType()); 1304 SV = InsertElementInst::Create(Old, SV, 1305 ConstantInt::get(Type::Int32Ty, Elt), 1306 "tmp", SI); 1307 } 1308 } else if (isa<PointerType>(AllocaType)) { 1309 // If the alloca type is a pointer, then all the elements must be 1310 // pointers. 1311 if (SV->getType() != AllocaType) 1312 SV = new BitCastInst(SV, AllocaType, SV->getName(), SI); 1313 } else { 1314 Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI); 1315 1316 // If SV is a float, convert it to the appropriate integer type. 1317 // If it is a pointer, do the same, and also handle ptr->ptr casts 1318 // here. 1319 const TargetData &TD = getAnalysis<TargetData>(); 1320 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 1321 unsigned DestWidth = TD.getTypeSizeInBits(AllocaType); 1322 unsigned SrcStoreWidth = TD.getTypeStoreSizeInBits(SV->getType()); 1323 unsigned DestStoreWidth = TD.getTypeStoreSizeInBits(AllocaType); 1324 if (SV->getType()->isFloatingPoint()) 1325 SV = new BitCastInst(SV, IntegerType::get(SrcWidth), 1326 SV->getName(), SI); 1327 else if (isa<PointerType>(SV->getType())) 1328 SV = new PtrToIntInst(SV, TD.getIntPtrType(), SV->getName(), SI); 1329 1330 // Always zero extend the value if needed. 1331 if (SV->getType() != AllocaType) 1332 SV = new ZExtInst(SV, AllocaType, SV->getName(), SI); 1333 1334 // If this is a big-endian system and the store is narrower than the 1335 // full alloca type, we need to do a shift to get the right bits. 1336 int ShAmt = 0; 1337 if (TD.isBigEndian()) { 1338 // On big-endian machines, the lowest bit is stored at the bit offset 1339 // from the pointer given by getTypeStoreSizeInBits. This matters for 1340 // integers with a bitwidth that is not a multiple of 8. 1341 ShAmt = DestStoreWidth - SrcStoreWidth - Offset; 1342 } else { 1343 ShAmt = Offset; 1344 } 1345 1346 // Note: we support negative bitwidths (with shr) which are not defined. 1347 // We do this to support (f.e.) stores off the end of a structure where 1348 // only some bits in the structure are set. 1349 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1350 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1351 SV = BinaryOperator::CreateShl(SV, 1352 ConstantInt::get(SV->getType(), ShAmt), 1353 SV->getName(), SI); 1354 Mask <<= ShAmt; 1355 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1356 SV = BinaryOperator::CreateLShr(SV, 1357 ConstantInt::get(SV->getType(),-ShAmt), 1358 SV->getName(), SI); 1359 Mask = Mask.lshr(ShAmt); 1360 } 1361 1362 // Mask out the bits we are about to insert from the old value, and or 1363 // in the new bits. 1364 if (SrcWidth != DestWidth) { 1365 assert(DestWidth > SrcWidth); 1366 Old = BinaryOperator::CreateAnd(Old, ConstantInt::get(~Mask), 1367 Old->getName()+".mask", SI); 1368 SV = BinaryOperator::CreateOr(Old, SV, SV->getName()+".ins", SI); 1369 } 1370 } 1371 return SV; 1372} 1373 1374 1375 1376/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1377/// some part of a constant global variable. This intentionally only accepts 1378/// constant expressions because we don't can't rewrite arbitrary instructions. 1379static bool PointsToConstantGlobal(Value *V) { 1380 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1381 return GV->isConstant(); 1382 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1383 if (CE->getOpcode() == Instruction::BitCast || 1384 CE->getOpcode() == Instruction::GetElementPtr) 1385 return PointsToConstantGlobal(CE->getOperand(0)); 1386 return false; 1387} 1388 1389/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1390/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1391/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1392/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1393/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1394/// the alloca, and if the source pointer is a pointer to a constant global, we 1395/// can optimize this. 1396static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1397 bool isOffset) { 1398 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1399 if (isa<LoadInst>(*UI)) { 1400 // Ignore loads, they are always ok. 1401 continue; 1402 } 1403 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1404 // If uses of the bitcast are ok, we are ok. 1405 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1406 return false; 1407 continue; 1408 } 1409 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1410 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1411 // doesn't, it does. 1412 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1413 isOffset || !GEP->hasAllZeroIndices())) 1414 return false; 1415 continue; 1416 } 1417 1418 // If this is isn't our memcpy/memmove, reject it as something we can't 1419 // handle. 1420 if (!isa<MemCpyInst>(*UI) && !isa<MemMoveInst>(*UI)) 1421 return false; 1422 1423 // If we already have seen a copy, reject the second one. 1424 if (TheCopy) return false; 1425 1426 // If the pointer has been offset from the start of the alloca, we can't 1427 // safely handle this. 1428 if (isOffset) return false; 1429 1430 // If the memintrinsic isn't using the alloca as the dest, reject it. 1431 if (UI.getOperandNo() != 1) return false; 1432 1433 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1434 1435 // If the source of the memcpy/move is not a constant global, reject it. 1436 if (!PointsToConstantGlobal(MI->getOperand(2))) 1437 return false; 1438 1439 // Otherwise, the transform is safe. Remember the copy instruction. 1440 TheCopy = MI; 1441 } 1442 return true; 1443} 1444 1445/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1446/// modified by a copy from a constant global. If we can prove this, we can 1447/// replace any uses of the alloca with uses of the global directly. 1448Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) { 1449 Instruction *TheCopy = 0; 1450 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1451 return TheCopy; 1452 return 0; 1453} 1454