ScalarReplAggregates.cpp revision b0e71edb6b33f822e001500dac90acf95faacea8
1//===- ScalarReplAggregates.cpp - Scalar Replacement of Aggregates --------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source 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((intptr_t)&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 static Instruction *isOnlyCopiedFromConstantGlobal(AllocationInst *AI); 122 }; 123 124 char SROA::ID = 0; 125 RegisterPass<SROA> X("scalarrepl", "Scalar Replacement of Aggregates"); 126} 127 128// Public interface to the ScalarReplAggregates pass 129FunctionPass *llvm::createScalarReplAggregatesPass(signed int Threshold) { 130 return new SROA(Threshold); 131} 132 133 134bool SROA::runOnFunction(Function &F) { 135 bool Changed = performPromotion(F); 136 while (1) { 137 bool LocalChange = performScalarRepl(F); 138 if (!LocalChange) break; // No need to repromote if no scalarrepl 139 Changed = true; 140 LocalChange = performPromotion(F); 141 if (!LocalChange) break; // No need to re-scalarrepl if no promotion 142 } 143 144 return Changed; 145} 146 147 148bool SROA::performPromotion(Function &F) { 149 std::vector<AllocaInst*> Allocas; 150 DominatorTree &DT = getAnalysis<DominatorTree>(); 151 DominanceFrontier &DF = getAnalysis<DominanceFrontier>(); 152 153 BasicBlock &BB = F.getEntryBlock(); // Get the entry node for the function 154 155 bool Changed = false; 156 157 while (1) { 158 Allocas.clear(); 159 160 // Find allocas that are safe to promote, by looking at all instructions in 161 // the entry node 162 for (BasicBlock::iterator I = BB.begin(), E = --BB.end(); I != E; ++I) 163 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) // Is it an alloca? 164 if (isAllocaPromotable(AI)) 165 Allocas.push_back(AI); 166 167 if (Allocas.empty()) break; 168 169 PromoteMemToReg(Allocas, DT, DF); 170 NumPromoted += Allocas.size(); 171 Changed = true; 172 } 173 174 return Changed; 175} 176 177// performScalarRepl - This algorithm is a simple worklist driven algorithm, 178// which runs on all of the malloc/alloca instructions in the function, removing 179// them if they are only used by getelementptr instructions. 180// 181bool SROA::performScalarRepl(Function &F) { 182 std::vector<AllocationInst*> WorkList; 183 184 // Scan the entry basic block, adding any alloca's and mallocs to the worklist 185 BasicBlock &BB = F.getEntryBlock(); 186 for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) 187 if (AllocationInst *A = dyn_cast<AllocationInst>(I)) 188 WorkList.push_back(A); 189 190 const TargetData &TD = getAnalysis<TargetData>(); 191 192 // Process the worklist 193 bool Changed = false; 194 while (!WorkList.empty()) { 195 AllocationInst *AI = WorkList.back(); 196 WorkList.pop_back(); 197 198 // Handle dead allocas trivially. These can be formed by SROA'ing arrays 199 // with unused elements. 200 if (AI->use_empty()) { 201 AI->eraseFromParent(); 202 continue; 203 } 204 205 // If we can turn this aggregate value (potentially with casts) into a 206 // simple scalar value that can be mem2reg'd into a register value. 207 bool IsNotTrivial = false; 208 if (const Type *ActualType = CanConvertToScalar(AI, IsNotTrivial)) 209 if (IsNotTrivial && ActualType != Type::VoidTy) { 210 ConvertToScalar(AI, ActualType); 211 Changed = true; 212 continue; 213 } 214 215 // Check to see if we can perform the core SROA transformation. We cannot 216 // transform the allocation instruction if it is an array allocation 217 // (allocations OF arrays are ok though), and an allocation of a scalar 218 // value cannot be decomposed at all. 219 if (!AI->isArrayAllocation() && 220 (isa<StructType>(AI->getAllocatedType()) || 221 isa<ArrayType>(AI->getAllocatedType())) && 222 AI->getAllocatedType()->isSized() && 223 TD.getTypeSize(AI->getAllocatedType()) < SRThreshold) { 224 // Check that all of the users of the allocation are capable of being 225 // transformed. 226 switch (isSafeAllocaToScalarRepl(AI)) { 227 default: assert(0 && "Unexpected value!"); 228 case 0: // Not safe to scalar replace. 229 break; 230 case 1: // Safe, but requires cleanup/canonicalizations first 231 CanonicalizeAllocaUsers(AI); 232 // FALL THROUGH. 233 case 3: // Safe to scalar replace. 234 DoScalarReplacement(AI, WorkList); 235 Changed = true; 236 continue; 237 } 238 } 239 240 // Check to see if this allocation is only modified by a memcpy/memmove from 241 // a constant global. If this is the case, we can change all users to use 242 // the constant global instead. This is commonly produced by the CFE by 243 // constructs like "void foo() { int A[] = {1,2,3,4,5,6,7,8,9...}; }" if 'A' 244 // is only subsequently read. 245 if (Instruction *TheCopy = isOnlyCopiedFromConstantGlobal(AI)) { 246 DOUT << "Found alloca equal to global: " << *AI; 247 DOUT << " memcpy = " << *TheCopy; 248 Constant *TheSrc = cast<Constant>(TheCopy->getOperand(2)); 249 AI->replaceAllUsesWith(ConstantExpr::getBitCast(TheSrc, AI->getType())); 250 TheCopy->eraseFromParent(); // Don't mutate the global. 251 AI->eraseFromParent(); 252 ++NumGlobals; 253 Changed = true; 254 continue; 255 } 256 257 // Otherwise, couldn't process this. 258 } 259 260 return Changed; 261} 262 263/// DoScalarReplacement - This alloca satisfied the isSafeAllocaToScalarRepl 264/// predicate, do SROA now. 265void SROA::DoScalarReplacement(AllocationInst *AI, 266 std::vector<AllocationInst*> &WorkList) { 267 DOUT << "Found inst to SROA: " << *AI; 268 SmallVector<AllocaInst*, 32> ElementAllocas; 269 if (const StructType *ST = dyn_cast<StructType>(AI->getAllocatedType())) { 270 ElementAllocas.reserve(ST->getNumContainedTypes()); 271 for (unsigned i = 0, e = ST->getNumContainedTypes(); i != e; ++i) { 272 AllocaInst *NA = new AllocaInst(ST->getContainedType(i), 0, 273 AI->getAlignment(), 274 AI->getName() + "." + utostr(i), AI); 275 ElementAllocas.push_back(NA); 276 WorkList.push_back(NA); // Add to worklist for recursive processing 277 } 278 } else { 279 const ArrayType *AT = cast<ArrayType>(AI->getAllocatedType()); 280 ElementAllocas.reserve(AT->getNumElements()); 281 const Type *ElTy = AT->getElementType(); 282 for (unsigned i = 0, e = AT->getNumElements(); i != e; ++i) { 283 AllocaInst *NA = new AllocaInst(ElTy, 0, AI->getAlignment(), 284 AI->getName() + "." + utostr(i), AI); 285 ElementAllocas.push_back(NA); 286 WorkList.push_back(NA); // Add to worklist for recursive processing 287 } 288 } 289 290 // Now that we have created the alloca instructions that we want to use, 291 // expand the getelementptr instructions to use them. 292 // 293 while (!AI->use_empty()) { 294 Instruction *User = cast<Instruction>(AI->use_back()); 295 if (BitCastInst *BCInst = dyn_cast<BitCastInst>(User)) { 296 RewriteBitCastUserOfAlloca(BCInst, AI, ElementAllocas); 297 BCInst->eraseFromParent(); 298 continue; 299 } 300 301 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(User); 302 // We now know that the GEP is of the form: GEP <ptr>, 0, <cst> 303 unsigned Idx = 304 (unsigned)cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue(); 305 306 assert(Idx < ElementAllocas.size() && "Index out of range?"); 307 AllocaInst *AllocaToUse = ElementAllocas[Idx]; 308 309 Value *RepValue; 310 if (GEPI->getNumOperands() == 3) { 311 // Do not insert a new getelementptr instruction with zero indices, only 312 // to have it optimized out later. 313 RepValue = AllocaToUse; 314 } else { 315 // We are indexing deeply into the structure, so we still need a 316 // getelement ptr instruction to finish the indexing. This may be 317 // expanded itself once the worklist is rerun. 318 // 319 SmallVector<Value*, 8> NewArgs; 320 NewArgs.push_back(Constant::getNullValue(Type::Int32Ty)); 321 NewArgs.append(GEPI->op_begin()+3, GEPI->op_end()); 322 RepValue = new GetElementPtrInst(AllocaToUse, &NewArgs[0], 323 NewArgs.size(), "", GEPI); 324 RepValue->takeName(GEPI); 325 } 326 327 // If this GEP is to the start of the aggregate, check for memcpys. 328 if (Idx == 0) { 329 bool IsStartOfAggregateGEP = true; 330 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i) { 331 if (!isa<ConstantInt>(GEPI->getOperand(i))) { 332 IsStartOfAggregateGEP = false; 333 break; 334 } 335 if (!cast<ConstantInt>(GEPI->getOperand(i))->isZero()) { 336 IsStartOfAggregateGEP = false; 337 break; 338 } 339 } 340 341 if (IsStartOfAggregateGEP) 342 RewriteBitCastUserOfAlloca(GEPI, AI, ElementAllocas); 343 } 344 345 346 // Move all of the users over to the new GEP. 347 GEPI->replaceAllUsesWith(RepValue); 348 // Delete the old GEP 349 GEPI->eraseFromParent(); 350 } 351 352 // Finally, delete the Alloca instruction 353 AI->eraseFromParent(); 354 NumReplaced++; 355} 356 357 358/// isSafeElementUse - Check to see if this use is an allowed use for a 359/// getelementptr instruction of an array aggregate allocation. isFirstElt 360/// indicates whether Ptr is known to the start of the aggregate. 361/// 362void SROA::isSafeElementUse(Value *Ptr, bool isFirstElt, AllocationInst *AI, 363 AllocaInfo &Info) { 364 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 365 I != E; ++I) { 366 Instruction *User = cast<Instruction>(*I); 367 switch (User->getOpcode()) { 368 case Instruction::Load: break; 369 case Instruction::Store: 370 // Store is ok if storing INTO the pointer, not storing the pointer 371 if (User->getOperand(0) == Ptr) return MarkUnsafe(Info); 372 break; 373 case Instruction::GetElementPtr: { 374 GetElementPtrInst *GEP = cast<GetElementPtrInst>(User); 375 bool AreAllZeroIndices = isFirstElt; 376 if (GEP->getNumOperands() > 1) { 377 if (!isa<ConstantInt>(GEP->getOperand(1)) || 378 !cast<ConstantInt>(GEP->getOperand(1))->isZero()) 379 // Using pointer arithmetic to navigate the array. 380 return MarkUnsafe(Info); 381 382 if (AreAllZeroIndices) { 383 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i) { 384 if (!isa<ConstantInt>(GEP->getOperand(i)) || 385 !cast<ConstantInt>(GEP->getOperand(i))->isZero()) { 386 AreAllZeroIndices = false; 387 break; 388 } 389 } 390 } 391 } 392 isSafeElementUse(GEP, AreAllZeroIndices, AI, Info); 393 if (Info.isUnsafe) return; 394 break; 395 } 396 case Instruction::BitCast: 397 if (isFirstElt) { 398 isSafeUseOfBitCastedAllocation(cast<BitCastInst>(User), AI, Info); 399 if (Info.isUnsafe) return; 400 break; 401 } 402 DOUT << " Transformation preventing inst: " << *User; 403 return MarkUnsafe(Info); 404 case Instruction::Call: 405 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(User)) { 406 if (isFirstElt) { 407 isSafeMemIntrinsicOnAllocation(MI, AI, I.getOperandNo(), Info); 408 if (Info.isUnsafe) return; 409 break; 410 } 411 } 412 DOUT << " Transformation preventing inst: " << *User; 413 return MarkUnsafe(Info); 414 default: 415 DOUT << " Transformation preventing inst: " << *User; 416 return MarkUnsafe(Info); 417 } 418 } 419 return; // All users look ok :) 420} 421 422/// AllUsersAreLoads - Return true if all users of this value are loads. 423static bool AllUsersAreLoads(Value *Ptr) { 424 for (Value::use_iterator I = Ptr->use_begin(), E = Ptr->use_end(); 425 I != E; ++I) 426 if (cast<Instruction>(*I)->getOpcode() != Instruction::Load) 427 return false; 428 return true; 429} 430 431/// isSafeUseOfAllocation - Check to see if this user is an allowed use for an 432/// aggregate allocation. 433/// 434void SROA::isSafeUseOfAllocation(Instruction *User, AllocationInst *AI, 435 AllocaInfo &Info) { 436 if (BitCastInst *C = dyn_cast<BitCastInst>(User)) 437 return isSafeUseOfBitCastedAllocation(C, AI, Info); 438 439 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(User); 440 if (GEPI == 0) 441 return MarkUnsafe(Info); 442 443 gep_type_iterator I = gep_type_begin(GEPI), E = gep_type_end(GEPI); 444 445 // The GEP is not safe to transform if not of the form "GEP <ptr>, 0, <cst>". 446 if (I == E || 447 I.getOperand() != Constant::getNullValue(I.getOperand()->getType())) { 448 return MarkUnsafe(Info); 449 } 450 451 ++I; 452 if (I == E) return MarkUnsafe(Info); // ran out of GEP indices?? 453 454 bool IsAllZeroIndices = true; 455 456 // If this is a use of an array allocation, do a bit more checking for sanity. 457 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 458 uint64_t NumElements = AT->getNumElements(); 459 460 if (ConstantInt *Idx = dyn_cast<ConstantInt>(I.getOperand())) { 461 IsAllZeroIndices &= Idx->isZero(); 462 463 // Check to make sure that index falls within the array. If not, 464 // something funny is going on, so we won't do the optimization. 465 // 466 if (Idx->getZExtValue() >= NumElements) 467 return MarkUnsafe(Info); 468 469 // We cannot scalar repl this level of the array unless any array 470 // sub-indices are in-range constants. In particular, consider: 471 // A[0][i]. We cannot know that the user isn't doing invalid things like 472 // allowing i to index an out-of-range subscript that accesses A[1]. 473 // 474 // Scalar replacing *just* the outer index of the array is probably not 475 // going to be a win anyway, so just give up. 476 for (++I; I != E && (isa<ArrayType>(*I) || isa<VectorType>(*I)); ++I) { 477 uint64_t NumElements; 478 if (const ArrayType *SubArrayTy = dyn_cast<ArrayType>(*I)) 479 NumElements = SubArrayTy->getNumElements(); 480 else 481 NumElements = cast<VectorType>(*I)->getNumElements(); 482 483 ConstantInt *IdxVal = dyn_cast<ConstantInt>(I.getOperand()); 484 if (!IdxVal) return MarkUnsafe(Info); 485 if (IdxVal->getZExtValue() >= NumElements) 486 return MarkUnsafe(Info); 487 IsAllZeroIndices &= IdxVal->isZero(); 488 } 489 490 } else { 491 IsAllZeroIndices = 0; 492 493 // If this is an array index and the index is not constant, we cannot 494 // promote... that is unless the array has exactly one or two elements in 495 // it, in which case we CAN promote it, but we have to canonicalize this 496 // out if this is the only problem. 497 if ((NumElements == 1 || NumElements == 2) && 498 AllUsersAreLoads(GEPI)) { 499 Info.needsCanon = true; 500 return; // Canonicalization required! 501 } 502 return MarkUnsafe(Info); 503 } 504 } 505 506 // If there are any non-simple uses of this getelementptr, make sure to reject 507 // them. 508 return isSafeElementUse(GEPI, IsAllZeroIndices, AI, Info); 509} 510 511/// isSafeMemIntrinsicOnAllocation - Return true if the specified memory 512/// intrinsic can be promoted by SROA. At this point, we know that the operand 513/// of the memintrinsic is a pointer to the beginning of the allocation. 514void SROA::isSafeMemIntrinsicOnAllocation(MemIntrinsic *MI, AllocationInst *AI, 515 unsigned OpNo, AllocaInfo &Info) { 516 // If not constant length, give up. 517 ConstantInt *Length = dyn_cast<ConstantInt>(MI->getLength()); 518 if (!Length) return MarkUnsafe(Info); 519 520 // If not the whole aggregate, give up. 521 const TargetData &TD = getAnalysis<TargetData>(); 522 if (Length->getZExtValue() != TD.getTypeSize(AI->getType()->getElementType())) 523 return MarkUnsafe(Info); 524 525 // We only know about memcpy/memset/memmove. 526 if (!isa<MemCpyInst>(MI) && !isa<MemSetInst>(MI) && !isa<MemMoveInst>(MI)) 527 return MarkUnsafe(Info); 528 529 // Otherwise, we can transform it. Determine whether this is a memcpy/set 530 // into or out of the aggregate. 531 if (OpNo == 1) 532 Info.isMemCpyDst = true; 533 else { 534 assert(OpNo == 2); 535 Info.isMemCpySrc = true; 536 } 537} 538 539/// isSafeUseOfBitCastedAllocation - Return true if all users of this bitcast 540/// are 541void SROA::isSafeUseOfBitCastedAllocation(BitCastInst *BC, AllocationInst *AI, 542 AllocaInfo &Info) { 543 for (Value::use_iterator UI = BC->use_begin(), E = BC->use_end(); 544 UI != E; ++UI) { 545 if (BitCastInst *BCU = dyn_cast<BitCastInst>(UI)) { 546 isSafeUseOfBitCastedAllocation(BCU, AI, Info); 547 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(UI)) { 548 isSafeMemIntrinsicOnAllocation(MI, AI, UI.getOperandNo(), Info); 549 } else { 550 return MarkUnsafe(Info); 551 } 552 if (Info.isUnsafe) return; 553 } 554} 555 556/// RewriteBitCastUserOfAlloca - BCInst (transitively) bitcasts AI, or indexes 557/// to its first element. Transform users of the cast to use the new values 558/// instead. 559void SROA::RewriteBitCastUserOfAlloca(Instruction *BCInst, AllocationInst *AI, 560 SmallVector<AllocaInst*, 32> &NewElts) { 561 Constant *Zero = Constant::getNullValue(Type::Int32Ty); 562 const TargetData &TD = getAnalysis<TargetData>(); 563 564 Value::use_iterator UI = BCInst->use_begin(), UE = BCInst->use_end(); 565 while (UI != UE) { 566 if (BitCastInst *BCU = dyn_cast<BitCastInst>(*UI)) { 567 RewriteBitCastUserOfAlloca(BCU, AI, NewElts); 568 ++UI; 569 BCU->eraseFromParent(); 570 continue; 571 } 572 573 // Otherwise, must be memcpy/memmove/memset of the entire aggregate. Split 574 // into one per element. 575 MemIntrinsic *MI = dyn_cast<MemIntrinsic>(*UI); 576 577 // If it's not a mem intrinsic, it must be some other user of a gep of the 578 // first pointer. Just leave these alone. 579 if (!MI) { 580 ++UI; 581 continue; 582 } 583 584 // If this is a memcpy/memmove, construct the other pointer as the 585 // appropriate type. 586 Value *OtherPtr = 0; 587 if (MemCpyInst *MCI = dyn_cast<MemCpyInst>(MI)) { 588 if (BCInst == MCI->getRawDest()) 589 OtherPtr = MCI->getRawSource(); 590 else { 591 assert(BCInst == MCI->getRawSource()); 592 OtherPtr = MCI->getRawDest(); 593 } 594 } else if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) { 595 if (BCInst == MMI->getRawDest()) 596 OtherPtr = MMI->getRawSource(); 597 else { 598 assert(BCInst == MMI->getRawSource()); 599 OtherPtr = MMI->getRawDest(); 600 } 601 } 602 603 // If there is an other pointer, we want to convert it to the same pointer 604 // type as AI has, so we can GEP through it. 605 if (OtherPtr) { 606 // It is likely that OtherPtr is a bitcast, if so, remove it. 607 if (BitCastInst *BC = dyn_cast<BitCastInst>(OtherPtr)) 608 OtherPtr = BC->getOperand(0); 609 if (ConstantExpr *BCE = dyn_cast<ConstantExpr>(OtherPtr)) 610 if (BCE->getOpcode() == Instruction::BitCast) 611 OtherPtr = BCE->getOperand(0); 612 613 // If the pointer is not the right type, insert a bitcast to the right 614 // type. 615 if (OtherPtr->getType() != AI->getType()) 616 OtherPtr = new BitCastInst(OtherPtr, AI->getType(), OtherPtr->getName(), 617 MI); 618 } 619 620 // Process each element of the aggregate. 621 Value *TheFn = MI->getOperand(0); 622 const Type *BytePtrTy = MI->getRawDest()->getType(); 623 bool SROADest = MI->getRawDest() == BCInst; 624 625 for (unsigned i = 0, e = NewElts.size(); i != e; ++i) { 626 // If this is a memcpy/memmove, emit a GEP of the other element address. 627 Value *OtherElt = 0; 628 if (OtherPtr) { 629 OtherElt = new GetElementPtrInst(OtherPtr, Zero, 630 ConstantInt::get(Type::Int32Ty, i), 631 OtherPtr->getNameStr()+"."+utostr(i), 632 MI); 633 } 634 635 Value *EltPtr = NewElts[i]; 636 const Type *EltTy =cast<PointerType>(EltPtr->getType())->getElementType(); 637 638 // If we got down to a scalar, insert a load or store as appropriate. 639 if (EltTy->isFirstClassType()) { 640 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) { 641 Value *Elt = new LoadInst(SROADest ? OtherElt : EltPtr, "tmp", 642 MI); 643 new StoreInst(Elt, SROADest ? EltPtr : OtherElt, MI); 644 continue; 645 } else { 646 assert(isa<MemSetInst>(MI)); 647 648 // If the stored element is zero (common case), just store a null 649 // constant. 650 Constant *StoreVal; 651 if (ConstantInt *CI = dyn_cast<ConstantInt>(MI->getOperand(2))) { 652 if (CI->isZero()) { 653 StoreVal = Constant::getNullValue(EltTy); // 0.0, null, 0, <0,0> 654 } else { 655 // If EltTy is a vector type, get the element type. 656 const Type *ValTy = EltTy; 657 if (const VectorType *VTy = dyn_cast<VectorType>(ValTy)) 658 ValTy = VTy->getElementType(); 659 660 // Construct an integer with the right value. 661 unsigned EltSize = TD.getTypeSize(ValTy); 662 APInt OneVal(EltSize*8, CI->getZExtValue()); 663 APInt TotalVal(OneVal); 664 // Set each byte. 665 for (unsigned i = 0; i != EltSize-1; ++i) { 666 TotalVal = TotalVal.shl(8); 667 TotalVal |= OneVal; 668 } 669 670 // Convert the integer value to the appropriate type. 671 StoreVal = ConstantInt::get(TotalVal); 672 if (isa<PointerType>(ValTy)) 673 StoreVal = ConstantExpr::getIntToPtr(StoreVal, ValTy); 674 else if (ValTy->isFloatingPoint()) 675 StoreVal = ConstantExpr::getBitCast(StoreVal, ValTy); 676 assert(StoreVal->getType() == ValTy && "Type mismatch!"); 677 678 // If the requested value was a vector constant, create it. 679 if (EltTy != ValTy) { 680 unsigned NumElts = cast<VectorType>(ValTy)->getNumElements(); 681 SmallVector<Constant*, 16> Elts(NumElts, StoreVal); 682 StoreVal = ConstantVector::get(&Elts[0], NumElts); 683 } 684 } 685 new StoreInst(StoreVal, EltPtr, MI); 686 continue; 687 } 688 // Otherwise, if we're storing a byte variable, use a memset call for 689 // this element. 690 } 691 } 692 693 // Cast the element pointer to BytePtrTy. 694 if (EltPtr->getType() != BytePtrTy) 695 EltPtr = new BitCastInst(EltPtr, BytePtrTy, EltPtr->getNameStr(), MI); 696 697 // Cast the other pointer (if we have one) to BytePtrTy. 698 if (OtherElt && OtherElt->getType() != BytePtrTy) 699 OtherElt = new BitCastInst(OtherElt, BytePtrTy,OtherElt->getNameStr(), 700 MI); 701 702 unsigned EltSize = TD.getTypeSize(EltTy); 703 704 // Finally, insert the meminst for this element. 705 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) { 706 Value *Ops[] = { 707 SROADest ? EltPtr : OtherElt, // Dest ptr 708 SROADest ? OtherElt : EltPtr, // Src ptr 709 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 710 Zero // Align 711 }; 712 new CallInst(TheFn, Ops, Ops + 4, "", MI); 713 } else { 714 assert(isa<MemSetInst>(MI)); 715 Value *Ops[] = { 716 EltPtr, MI->getOperand(2), // Dest, Value, 717 ConstantInt::get(MI->getOperand(3)->getType(), EltSize), // Size 718 Zero // Align 719 }; 720 new CallInst(TheFn, Ops, Ops + 4, "", MI); 721 } 722 } 723 724 // Finally, MI is now dead, as we've modified its actions to occur on all of 725 // the elements of the aggregate. 726 ++UI; 727 MI->eraseFromParent(); 728 } 729} 730 731/// HasStructPadding - Return true if the specified type has any structure 732/// padding, false otherwise. 733static bool HasStructPadding(const Type *Ty, const TargetData &TD) { 734 if (const StructType *STy = dyn_cast<StructType>(Ty)) { 735 const StructLayout *SL = TD.getStructLayout(STy); 736 unsigned PrevFieldBitOffset = 0; 737 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 738 unsigned FieldBitOffset = SL->getElementOffset(i)*8; 739 740 // Padding in sub-elements? 741 if (HasStructPadding(STy->getElementType(i), TD)) 742 return true; 743 744 // Check to see if there is any padding between this element and the 745 // previous one. 746 if (i) { 747 unsigned PrevFieldEnd = 748 PrevFieldBitOffset+TD.getTypeSizeInBits(STy->getElementType(i-1)); 749 if (PrevFieldEnd < FieldBitOffset) 750 return true; 751 } 752 753 PrevFieldBitOffset = FieldBitOffset; 754 } 755 756 // Check for tail padding. 757 if (unsigned EltCount = STy->getNumElements()) { 758 unsigned PrevFieldEnd = PrevFieldBitOffset + 759 TD.getTypeSizeInBits(STy->getElementType(EltCount-1)); 760 if (PrevFieldEnd < SL->getSizeInBytes()*8) 761 return true; 762 } 763 764 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { 765 return HasStructPadding(ATy->getElementType(), TD); 766 } 767 return false; 768} 769 770/// isSafeStructAllocaToScalarRepl - Check to see if the specified allocation of 771/// an aggregate can be broken down into elements. Return 0 if not, 3 if safe, 772/// or 1 if safe after canonicalization has been performed. 773/// 774int SROA::isSafeAllocaToScalarRepl(AllocationInst *AI) { 775 // Loop over the use list of the alloca. We can only transform it if all of 776 // the users are safe to transform. 777 AllocaInfo Info; 778 779 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); 780 I != E; ++I) { 781 isSafeUseOfAllocation(cast<Instruction>(*I), AI, Info); 782 if (Info.isUnsafe) { 783 DOUT << "Cannot transform: " << *AI << " due to user: " << **I; 784 return 0; 785 } 786 } 787 788 // Okay, we know all the users are promotable. If the aggregate is a memcpy 789 // source and destination, we have to be careful. In particular, the memcpy 790 // could be moving around elements that live in structure padding of the LLVM 791 // types, but may actually be used. In these cases, we refuse to promote the 792 // struct. 793 if (Info.isMemCpySrc && Info.isMemCpyDst && 794 HasStructPadding(AI->getType()->getElementType(), 795 getAnalysis<TargetData>())) 796 return 0; 797 798 // If we require cleanup, return 1, otherwise return 3. 799 return Info.needsCanon ? 1 : 3; 800} 801 802/// CanonicalizeAllocaUsers - If SROA reported that it can promote the specified 803/// allocation, but only if cleaned up, perform the cleanups required. 804void SROA::CanonicalizeAllocaUsers(AllocationInst *AI) { 805 // At this point, we know that the end result will be SROA'd and promoted, so 806 // we can insert ugly code if required so long as sroa+mem2reg will clean it 807 // up. 808 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); 809 UI != E; ) { 810 GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI++); 811 if (!GEPI) continue; 812 gep_type_iterator I = gep_type_begin(GEPI); 813 ++I; 814 815 if (const ArrayType *AT = dyn_cast<ArrayType>(*I)) { 816 uint64_t NumElements = AT->getNumElements(); 817 818 if (!isa<ConstantInt>(I.getOperand())) { 819 if (NumElements == 1) { 820 GEPI->setOperand(2, Constant::getNullValue(Type::Int32Ty)); 821 } else { 822 assert(NumElements == 2 && "Unhandled case!"); 823 // All users of the GEP must be loads. At each use of the GEP, insert 824 // two loads of the appropriate indexed GEP and select between them. 825 Value *IsOne = new ICmpInst(ICmpInst::ICMP_NE, I.getOperand(), 826 Constant::getNullValue(I.getOperand()->getType()), 827 "isone", GEPI); 828 // Insert the new GEP instructions, which are properly indexed. 829 SmallVector<Value*, 8> Indices(GEPI->op_begin()+1, GEPI->op_end()); 830 Indices[1] = Constant::getNullValue(Type::Int32Ty); 831 Value *ZeroIdx = new GetElementPtrInst(GEPI->getOperand(0), 832 &Indices[0], Indices.size(), 833 GEPI->getName()+".0", GEPI); 834 Indices[1] = ConstantInt::get(Type::Int32Ty, 1); 835 Value *OneIdx = new GetElementPtrInst(GEPI->getOperand(0), 836 &Indices[0], Indices.size(), 837 GEPI->getName()+".1", GEPI); 838 // Replace all loads of the variable index GEP with loads from both 839 // indexes and a select. 840 while (!GEPI->use_empty()) { 841 LoadInst *LI = cast<LoadInst>(GEPI->use_back()); 842 Value *Zero = new LoadInst(ZeroIdx, LI->getName()+".0", LI); 843 Value *One = new LoadInst(OneIdx , LI->getName()+".1", LI); 844 Value *R = new SelectInst(IsOne, One, Zero, LI->getName(), LI); 845 LI->replaceAllUsesWith(R); 846 LI->eraseFromParent(); 847 } 848 GEPI->eraseFromParent(); 849 } 850 } 851 } 852 } 853} 854 855/// MergeInType - Add the 'In' type to the accumulated type so far. If the 856/// types are incompatible, return true, otherwise update Accum and return 857/// false. 858/// 859/// There are three cases we handle here: 860/// 1) An effectively-integer union, where the pieces are stored into as 861/// smaller integers (common with byte swap and other idioms). 862/// 2) A union of vector types of the same size and potentially its elements. 863/// Here we turn element accesses into insert/extract element operations. 864/// 3) A union of scalar types, such as int/float or int/pointer. Here we 865/// merge together into integers, allowing the xform to work with #1 as 866/// well. 867static bool MergeInType(const Type *In, const Type *&Accum, 868 const TargetData &TD) { 869 // If this is our first type, just use it. 870 const VectorType *PTy; 871 if (Accum == Type::VoidTy || In == Accum) { 872 Accum = In; 873 } else if (In == Type::VoidTy) { 874 // Noop. 875 } else if (In->isInteger() && Accum->isInteger()) { // integer union. 876 // Otherwise pick whichever type is larger. 877 if (cast<IntegerType>(In)->getBitWidth() > 878 cast<IntegerType>(Accum)->getBitWidth()) 879 Accum = In; 880 } else if (isa<PointerType>(In) && isa<PointerType>(Accum)) { 881 // Pointer unions just stay as one of the pointers. 882 } else if (isa<VectorType>(In) || isa<VectorType>(Accum)) { 883 if ((PTy = dyn_cast<VectorType>(Accum)) && 884 PTy->getElementType() == In) { 885 // Accum is a vector, and we are accessing an element: ok. 886 } else if ((PTy = dyn_cast<VectorType>(In)) && 887 PTy->getElementType() == Accum) { 888 // In is a vector, and accum is an element: ok, remember In. 889 Accum = In; 890 } else if ((PTy = dyn_cast<VectorType>(In)) && isa<VectorType>(Accum) && 891 PTy->getBitWidth() == cast<VectorType>(Accum)->getBitWidth()) { 892 // Two vectors of the same size: keep Accum. 893 } else { 894 // Cannot insert an short into a <4 x int> or handle 895 // <2 x int> -> <4 x int> 896 return true; 897 } 898 } else { 899 // Pointer/FP/Integer unions merge together as integers. 900 switch (Accum->getTypeID()) { 901 case Type::PointerTyID: Accum = TD.getIntPtrType(); break; 902 case Type::FloatTyID: Accum = Type::Int32Ty; break; 903 case Type::DoubleTyID: Accum = Type::Int64Ty; break; 904 default: 905 assert(Accum->isInteger() && "Unknown FP type!"); 906 break; 907 } 908 909 switch (In->getTypeID()) { 910 case Type::PointerTyID: In = TD.getIntPtrType(); break; 911 case Type::FloatTyID: In = Type::Int32Ty; break; 912 case Type::DoubleTyID: In = Type::Int64Ty; break; 913 default: 914 assert(In->isInteger() && "Unknown FP type!"); 915 break; 916 } 917 return MergeInType(In, Accum, TD); 918 } 919 return false; 920} 921 922/// getUIntAtLeastAsBitAs - Return an unsigned integer type that is at least 923/// as big as the specified type. If there is no suitable type, this returns 924/// null. 925const Type *getUIntAtLeastAsBitAs(unsigned NumBits) { 926 if (NumBits > 64) return 0; 927 if (NumBits > 32) return Type::Int64Ty; 928 if (NumBits > 16) return Type::Int32Ty; 929 if (NumBits > 8) return Type::Int16Ty; 930 return Type::Int8Ty; 931} 932 933/// CanConvertToScalar - V is a pointer. If we can convert the pointee to a 934/// single scalar integer type, return that type. Further, if the use is not 935/// a completely trivial use that mem2reg could promote, set IsNotTrivial. If 936/// there are no uses of this pointer, return Type::VoidTy to differentiate from 937/// failure. 938/// 939const Type *SROA::CanConvertToScalar(Value *V, bool &IsNotTrivial) { 940 const Type *UsedType = Type::VoidTy; // No uses, no forced type. 941 const TargetData &TD = getAnalysis<TargetData>(); 942 const PointerType *PTy = cast<PointerType>(V->getType()); 943 944 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 945 Instruction *User = cast<Instruction>(*UI); 946 947 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 948 if (MergeInType(LI->getType(), UsedType, TD)) 949 return 0; 950 951 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 952 // Storing the pointer, not into the value? 953 if (SI->getOperand(0) == V) return 0; 954 955 // NOTE: We could handle storing of FP imms into integers here! 956 957 if (MergeInType(SI->getOperand(0)->getType(), UsedType, TD)) 958 return 0; 959 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 960 IsNotTrivial = true; 961 const Type *SubTy = CanConvertToScalar(CI, IsNotTrivial); 962 if (!SubTy || MergeInType(SubTy, UsedType, TD)) return 0; 963 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 964 // Check to see if this is stepping over an element: GEP Ptr, int C 965 if (GEP->getNumOperands() == 2 && isa<ConstantInt>(GEP->getOperand(1))) { 966 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue(); 967 unsigned ElSize = TD.getTypeSize(PTy->getElementType()); 968 unsigned BitOffset = Idx*ElSize*8; 969 if (BitOffset > 64 || !isPowerOf2_32(ElSize)) return 0; 970 971 IsNotTrivial = true; 972 const Type *SubElt = CanConvertToScalar(GEP, IsNotTrivial); 973 if (SubElt == 0) return 0; 974 if (SubElt != Type::VoidTy && SubElt->isInteger()) { 975 const Type *NewTy = 976 getUIntAtLeastAsBitAs(TD.getTypeSize(SubElt)*8+BitOffset); 977 if (NewTy == 0 || MergeInType(NewTy, UsedType, TD)) return 0; 978 continue; 979 } 980 } else if (GEP->getNumOperands() == 3 && 981 isa<ConstantInt>(GEP->getOperand(1)) && 982 isa<ConstantInt>(GEP->getOperand(2)) && 983 cast<ConstantInt>(GEP->getOperand(1))->isZero()) { 984 // We are stepping into an element, e.g. a structure or an array: 985 // GEP Ptr, int 0, uint C 986 const Type *AggTy = PTy->getElementType(); 987 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 988 989 if (const ArrayType *ATy = dyn_cast<ArrayType>(AggTy)) { 990 if (Idx >= ATy->getNumElements()) return 0; // Out of range. 991 } else if (const VectorType *VectorTy = dyn_cast<VectorType>(AggTy)) { 992 // Getting an element of the vector. 993 if (Idx >= VectorTy->getNumElements()) return 0; // Out of range. 994 995 // Merge in the vector type. 996 if (MergeInType(VectorTy, UsedType, TD)) return 0; 997 998 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial); 999 if (SubTy == 0) return 0; 1000 1001 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, TD)) 1002 return 0; 1003 1004 // We'll need to change this to an insert/extract element operation. 1005 IsNotTrivial = true; 1006 continue; // Everything looks ok 1007 1008 } else if (isa<StructType>(AggTy)) { 1009 // Structs are always ok. 1010 } else { 1011 return 0; 1012 } 1013 const Type *NTy = getUIntAtLeastAsBitAs(TD.getTypeSize(AggTy)*8); 1014 if (NTy == 0 || MergeInType(NTy, UsedType, TD)) return 0; 1015 const Type *SubTy = CanConvertToScalar(GEP, IsNotTrivial); 1016 if (SubTy == 0) return 0; 1017 if (SubTy != Type::VoidTy && MergeInType(SubTy, UsedType, TD)) 1018 return 0; 1019 continue; // Everything looks ok 1020 } 1021 return 0; 1022 } else { 1023 // Cannot handle this! 1024 return 0; 1025 } 1026 } 1027 1028 return UsedType; 1029} 1030 1031/// ConvertToScalar - The specified alloca passes the CanConvertToScalar 1032/// predicate and is non-trivial. Convert it to something that can be trivially 1033/// promoted into a register by mem2reg. 1034void SROA::ConvertToScalar(AllocationInst *AI, const Type *ActualTy) { 1035 DOUT << "CONVERT TO SCALAR: " << *AI << " TYPE = " 1036 << *ActualTy << "\n"; 1037 ++NumConverted; 1038 1039 BasicBlock *EntryBlock = AI->getParent(); 1040 assert(EntryBlock == &EntryBlock->getParent()->getEntryBlock() && 1041 "Not in the entry block!"); 1042 EntryBlock->getInstList().remove(AI); // Take the alloca out of the program. 1043 1044 // Create and insert the alloca. 1045 AllocaInst *NewAI = new AllocaInst(ActualTy, 0, AI->getName(), 1046 EntryBlock->begin()); 1047 ConvertUsesToScalar(AI, NewAI, 0); 1048 delete AI; 1049} 1050 1051 1052/// ConvertUsesToScalar - Convert all of the users of Ptr to use the new alloca 1053/// directly. This happens when we are converting an "integer union" to a 1054/// single integer scalar, or when we are converting a "vector union" to a 1055/// vector with insert/extractelement instructions. 1056/// 1057/// Offset is an offset from the original alloca, in bits that need to be 1058/// shifted to the right. By the end of this, there should be no uses of Ptr. 1059void SROA::ConvertUsesToScalar(Value *Ptr, AllocaInst *NewAI, unsigned Offset) { 1060 const TargetData &TD = getAnalysis<TargetData>(); 1061 while (!Ptr->use_empty()) { 1062 Instruction *User = cast<Instruction>(Ptr->use_back()); 1063 1064 if (LoadInst *LI = dyn_cast<LoadInst>(User)) { 1065 // The load is a bit extract from NewAI shifted right by Offset bits. 1066 Value *NV = new LoadInst(NewAI, LI->getName(), LI); 1067 if (NV->getType() == LI->getType()) { 1068 // We win, no conversion needed. 1069 } else if (const VectorType *PTy = dyn_cast<VectorType>(NV->getType())) { 1070 // If the result alloca is a vector type, this is either an element 1071 // access or a bitcast to another vector type. 1072 if (isa<VectorType>(LI->getType())) { 1073 NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI); 1074 } else { 1075 // Must be an element access. 1076 unsigned Elt = Offset/(TD.getTypeSize(PTy->getElementType())*8); 1077 NV = new ExtractElementInst( 1078 NV, ConstantInt::get(Type::Int32Ty, Elt), "tmp", LI); 1079 } 1080 } else if (isa<PointerType>(NV->getType())) { 1081 assert(isa<PointerType>(LI->getType())); 1082 // Must be ptr->ptr cast. Anything else would result in NV being 1083 // an integer. 1084 NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI); 1085 } else { 1086 const IntegerType *NTy = cast<IntegerType>(NV->getType()); 1087 unsigned LIBitWidth = TD.getTypeSizeInBits(LI->getType()); 1088 1089 // If this is a big-endian system and the load is narrower than the 1090 // full alloca type, we need to do a shift to get the right bits. 1091 int ShAmt = 0; 1092 if (TD.isBigEndian()) { 1093 ShAmt = NTy->getBitWidth()-LIBitWidth-Offset; 1094 } else { 1095 ShAmt = Offset; 1096 } 1097 1098 // Note: we support negative bitwidths (with shl) which are not defined. 1099 // We do this to support (f.e.) loads off the end of a structure where 1100 // only some bits are used. 1101 if (ShAmt > 0 && (unsigned)ShAmt < NTy->getBitWidth()) 1102 NV = BinaryOperator::createLShr(NV, 1103 ConstantInt::get(NV->getType(),ShAmt), 1104 LI->getName(), LI); 1105 else if (ShAmt < 0 && (unsigned)-ShAmt < NTy->getBitWidth()) 1106 NV = BinaryOperator::createShl(NV, 1107 ConstantInt::get(NV->getType(),-ShAmt), 1108 LI->getName(), LI); 1109 1110 // Finally, unconditionally truncate the integer to the right width. 1111 if (LIBitWidth < NTy->getBitWidth()) 1112 NV = new TruncInst(NV, IntegerType::get(LIBitWidth), 1113 LI->getName(), LI); 1114 1115 // If the result is an integer, this is a trunc or bitcast. 1116 if (isa<IntegerType>(LI->getType())) { 1117 assert(NV->getType() == LI->getType() && "Truncate wasn't enough?"); 1118 } else if (LI->getType()->isFloatingPoint()) { 1119 // Just do a bitcast, we know the sizes match up. 1120 NV = new BitCastInst(NV, LI->getType(), LI->getName(), LI); 1121 } else { 1122 // Otherwise must be a pointer. 1123 NV = new IntToPtrInst(NV, LI->getType(), LI->getName(), LI); 1124 } 1125 } 1126 LI->replaceAllUsesWith(NV); 1127 LI->eraseFromParent(); 1128 } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) { 1129 assert(SI->getOperand(0) != Ptr && "Consistency error!"); 1130 1131 // Convert the stored type to the actual type, shift it left to insert 1132 // then 'or' into place. 1133 Value *SV = SI->getOperand(0); 1134 const Type *AllocaType = NewAI->getType()->getElementType(); 1135 if (SV->getType() == AllocaType) { 1136 // All is well. 1137 } else if (const VectorType *PTy = dyn_cast<VectorType>(AllocaType)) { 1138 Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI); 1139 1140 // If the result alloca is a vector type, this is either an element 1141 // access or a bitcast to another vector type. 1142 if (isa<VectorType>(SV->getType())) { 1143 SV = new BitCastInst(SV, AllocaType, SV->getName(), SI); 1144 } else { 1145 // Must be an element insertion. 1146 unsigned Elt = Offset/(TD.getTypeSize(PTy->getElementType())*8); 1147 SV = new InsertElementInst(Old, SV, 1148 ConstantInt::get(Type::Int32Ty, Elt), 1149 "tmp", SI); 1150 } 1151 } else if (isa<PointerType>(AllocaType)) { 1152 // If the alloca type is a pointer, then all the elements must be 1153 // pointers. 1154 if (SV->getType() != AllocaType) 1155 SV = new BitCastInst(SV, AllocaType, SV->getName(), SI); 1156 } else { 1157 Value *Old = new LoadInst(NewAI, NewAI->getName()+".in", SI); 1158 1159 // If SV is a float, convert it to the appropriate integer type. 1160 // If it is a pointer, do the same, and also handle ptr->ptr casts 1161 // here. 1162 unsigned SrcWidth = TD.getTypeSizeInBits(SV->getType()); 1163 unsigned DestWidth = AllocaType->getPrimitiveSizeInBits(); 1164 if (SV->getType()->isFloatingPoint()) 1165 SV = new BitCastInst(SV, IntegerType::get(SrcWidth), 1166 SV->getName(), SI); 1167 else if (isa<PointerType>(SV->getType())) 1168 SV = new PtrToIntInst(SV, TD.getIntPtrType(), SV->getName(), SI); 1169 1170 // Always zero extend the value if needed. 1171 if (SV->getType() != AllocaType) 1172 SV = new ZExtInst(SV, AllocaType, SV->getName(), SI); 1173 1174 // If this is a big-endian system and the store is narrower than the 1175 // full alloca type, we need to do a shift to get the right bits. 1176 int ShAmt = 0; 1177 if (TD.isBigEndian()) { 1178 ShAmt = DestWidth-SrcWidth-Offset; 1179 } else { 1180 ShAmt = Offset; 1181 } 1182 1183 // Note: we support negative bitwidths (with shr) which are not defined. 1184 // We do this to support (f.e.) stores off the end of a structure where 1185 // only some bits in the structure are set. 1186 APInt Mask(APInt::getLowBitsSet(DestWidth, SrcWidth)); 1187 if (ShAmt > 0 && (unsigned)ShAmt < DestWidth) { 1188 SV = BinaryOperator::createShl(SV, 1189 ConstantInt::get(SV->getType(), ShAmt), 1190 SV->getName(), SI); 1191 Mask <<= ShAmt; 1192 } else if (ShAmt < 0 && (unsigned)-ShAmt < DestWidth) { 1193 SV = BinaryOperator::createLShr(SV, 1194 ConstantInt::get(SV->getType(),-ShAmt), 1195 SV->getName(), SI); 1196 Mask = Mask.lshr(ShAmt); 1197 } 1198 1199 // Mask out the bits we are about to insert from the old value, and or 1200 // in the new bits. 1201 if (SrcWidth != DestWidth) { 1202 assert(DestWidth > SrcWidth); 1203 Old = BinaryOperator::createAnd(Old, ConstantInt::get(~Mask), 1204 Old->getName()+".mask", SI); 1205 SV = BinaryOperator::createOr(Old, SV, SV->getName()+".ins", SI); 1206 } 1207 } 1208 new StoreInst(SV, NewAI, SI); 1209 SI->eraseFromParent(); 1210 1211 } else if (BitCastInst *CI = dyn_cast<BitCastInst>(User)) { 1212 ConvertUsesToScalar(CI, NewAI, Offset); 1213 CI->eraseFromParent(); 1214 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(User)) { 1215 const PointerType *AggPtrTy = 1216 cast<PointerType>(GEP->getOperand(0)->getType()); 1217 const TargetData &TD = getAnalysis<TargetData>(); 1218 unsigned AggSizeInBits = TD.getTypeSize(AggPtrTy->getElementType())*8; 1219 1220 // Check to see if this is stepping over an element: GEP Ptr, int C 1221 unsigned NewOffset = Offset; 1222 if (GEP->getNumOperands() == 2) { 1223 unsigned Idx = cast<ConstantInt>(GEP->getOperand(1))->getZExtValue(); 1224 unsigned BitOffset = Idx*AggSizeInBits; 1225 1226 NewOffset += BitOffset; 1227 } else if (GEP->getNumOperands() == 3) { 1228 // We know that operand #2 is zero. 1229 unsigned Idx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue(); 1230 const Type *AggTy = AggPtrTy->getElementType(); 1231 if (const SequentialType *SeqTy = dyn_cast<SequentialType>(AggTy)) { 1232 unsigned ElSizeBits = TD.getTypeSize(SeqTy->getElementType())*8; 1233 1234 NewOffset += ElSizeBits*Idx; 1235 } else if (const StructType *STy = dyn_cast<StructType>(AggTy)) { 1236 unsigned EltBitOffset = 1237 TD.getStructLayout(STy)->getElementOffset(Idx)*8; 1238 1239 NewOffset += EltBitOffset; 1240 } else { 1241 assert(0 && "Unsupported operation!"); 1242 abort(); 1243 } 1244 } else { 1245 assert(0 && "Unsupported operation!"); 1246 abort(); 1247 } 1248 ConvertUsesToScalar(GEP, NewAI, NewOffset); 1249 GEP->eraseFromParent(); 1250 } else { 1251 assert(0 && "Unsupported operation!"); 1252 abort(); 1253 } 1254 } 1255} 1256 1257 1258/// PointsToConstantGlobal - Return true if V (possibly indirectly) points to 1259/// some part of a constant global variable. This intentionally only accepts 1260/// constant expressions because we don't can't rewrite arbitrary instructions. 1261static bool PointsToConstantGlobal(Value *V) { 1262 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) 1263 return GV->isConstant(); 1264 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 1265 if (CE->getOpcode() == Instruction::BitCast || 1266 CE->getOpcode() == Instruction::GetElementPtr) 1267 return PointsToConstantGlobal(CE->getOperand(0)); 1268 return false; 1269} 1270 1271/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived) 1272/// pointer to an alloca. Ignore any reads of the pointer, return false if we 1273/// see any stores or other unknown uses. If we see pointer arithmetic, keep 1274/// track of whether it moves the pointer (with isOffset) but otherwise traverse 1275/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to 1276/// the alloca, and if the source pointer is a pointer to a constant global, we 1277/// can optimize this. 1278static bool isOnlyCopiedFromConstantGlobal(Value *V, Instruction *&TheCopy, 1279 bool isOffset) { 1280 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) { 1281 if (isa<LoadInst>(*UI)) { 1282 // Ignore loads, they are always ok. 1283 continue; 1284 } 1285 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) { 1286 // If uses of the bitcast are ok, we are ok. 1287 if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, isOffset)) 1288 return false; 1289 continue; 1290 } 1291 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) { 1292 // If the GEP has all zero indices, it doesn't offset the pointer. If it 1293 // doesn't, it does. 1294 if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, 1295 isOffset || !GEP->hasAllZeroIndices())) 1296 return false; 1297 continue; 1298 } 1299 1300 // If this is isn't our memcpy/memmove, reject it as something we can't 1301 // handle. 1302 if (!isa<MemCpyInst>(*UI) && !isa<MemMoveInst>(*UI)) 1303 return false; 1304 1305 // If we already have seen a copy, reject the second one. 1306 if (TheCopy) return false; 1307 1308 // If the pointer has been offset from the start of the alloca, we can't 1309 // safely handle this. 1310 if (isOffset) return false; 1311 1312 // If the memintrinsic isn't using the alloca as the dest, reject it. 1313 if (UI.getOperandNo() != 1) return false; 1314 1315 MemIntrinsic *MI = cast<MemIntrinsic>(*UI); 1316 1317 // If the source of the memcpy/move is not a constant global, reject it. 1318 if (!PointsToConstantGlobal(MI->getOperand(2))) 1319 return false; 1320 1321 // Otherwise, the transform is safe. Remember the copy instruction. 1322 TheCopy = MI; 1323 } 1324 return true; 1325} 1326 1327/// isOnlyCopiedFromConstantGlobal - Return true if the specified alloca is only 1328/// modified by a copy from a constant global. If we can prove this, we can 1329/// replace any uses of the alloca with uses of the global directly. 1330Instruction *SROA::isOnlyCopiedFromConstantGlobal(AllocationInst *AI) { 1331 Instruction *TheCopy = 0; 1332 if (::isOnlyCopiedFromConstantGlobal(AI, TheCopy, false)) 1333 return TheCopy; 1334 return 0; 1335} 1336