InstCombineCalls.cpp revision 795e70e4319bb38eb92701c2a463eeb7584a9b25
1//===- InstCombineCalls.cpp -----------------------------------------------===// 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 file implements the visitCall and visitInvoke functions. 11// 12//===----------------------------------------------------------------------===// 13 14#include "InstCombine.h" 15#include "llvm/IntrinsicInst.h" 16#include "llvm/Support/CallSite.h" 17#include "llvm/Target/TargetData.h" 18#include "llvm/Analysis/MemoryBuiltins.h" 19#include "llvm/Transforms/Utils/BuildLibCalls.h" 20using namespace llvm; 21 22/// getPromotedType - Return the specified type promoted as it would be to pass 23/// though a va_arg area. 24static const Type *getPromotedType(const Type *Ty) { 25 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) { 26 if (ITy->getBitWidth() < 32) 27 return Type::getInt32Ty(Ty->getContext()); 28 } 29 return Ty; 30} 31 32/// EnforceKnownAlignment - If the specified pointer points to an object that 33/// we control, modify the object's alignment to PrefAlign. This isn't 34/// often possible though. If alignment is important, a more reliable approach 35/// is to simply align all global variables and allocation instructions to 36/// their preferred alignment from the beginning. 37/// 38static unsigned EnforceKnownAlignment(Value *V, 39 unsigned Align, unsigned PrefAlign) { 40 41 User *U = dyn_cast<User>(V); 42 if (!U) return Align; 43 44 switch (Operator::getOpcode(U)) { 45 default: break; 46 case Instruction::BitCast: 47 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign); 48 case Instruction::GetElementPtr: { 49 // If all indexes are zero, it is just the alignment of the base pointer. 50 bool AllZeroOperands = true; 51 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i) 52 if (!isa<Constant>(*i) || 53 !cast<Constant>(*i)->isNullValue()) { 54 AllZeroOperands = false; 55 break; 56 } 57 58 if (AllZeroOperands) { 59 // Treat this like a bitcast. 60 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign); 61 } 62 return Align; 63 } 64 case Instruction::Alloca: { 65 AllocaInst *AI = cast<AllocaInst>(V); 66 // If there is a requested alignment and if this is an alloca, round up. 67 if (AI->getAlignment() >= PrefAlign) 68 return AI->getAlignment(); 69 AI->setAlignment(PrefAlign); 70 return PrefAlign; 71 } 72 } 73 74 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) { 75 // If there is a large requested alignment and we can, bump up the alignment 76 // of the global. 77 if (GV->isDeclaration()) return Align; 78 79 if (GV->getAlignment() >= PrefAlign) 80 return GV->getAlignment(); 81 // We can only increase the alignment of the global if it has no alignment 82 // specified or if it is not assigned a section. If it is assigned a 83 // section, the global could be densely packed with other objects in the 84 // section, increasing the alignment could cause padding issues. 85 if (!GV->hasSection() || GV->getAlignment() == 0) 86 GV->setAlignment(PrefAlign); 87 return GV->getAlignment(); 88 } 89 90 return Align; 91} 92 93/// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that 94/// we can determine, return it, otherwise return 0. If PrefAlign is specified, 95/// and it is more than the alignment of the ultimate object, see if we can 96/// increase the alignment of the ultimate object, making this check succeed. 97unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V, 98 unsigned PrefAlign) { 99 assert(V->getType()->isPointerTy() && 100 "GetOrEnforceKnownAlignment expects a pointer!"); 101 unsigned BitWidth = TD ? TD->getPointerSizeInBits() : 64; 102 APInt Mask = APInt::getAllOnesValue(BitWidth); 103 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 104 ComputeMaskedBits(V, Mask, KnownZero, KnownOne); 105 unsigned TrailZ = KnownZero.countTrailingOnes(); 106 107 // Avoid trouble with rediculously large TrailZ values, such as 108 // those computed from a null pointer. 109 TrailZ = std::min(TrailZ, unsigned(sizeof(unsigned) * CHAR_BIT - 1)); 110 111 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ); 112 113 // LLVM doesn't support alignments larger than this currently. 114 Align = std::min(Align, +Value::MaximumAlignment); 115 116 if (PrefAlign > Align) 117 Align = EnforceKnownAlignment(V, Align, PrefAlign); 118 119 // We don't need to make any adjustment. 120 return Align; 121} 122 123Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) { 124 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getArgOperand(0)); 125 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getArgOperand(1)); 126 unsigned MinAlign = std::min(DstAlign, SrcAlign); 127 unsigned CopyAlign = MI->getAlignment(); 128 129 if (CopyAlign < MinAlign) { 130 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), 131 MinAlign, false)); 132 return MI; 133 } 134 135 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with 136 // load/store. 137 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2)); 138 if (MemOpLength == 0) return 0; 139 140 // Source and destination pointer types are always "i8*" for intrinsic. See 141 // if the size is something we can handle with a single primitive load/store. 142 // A single load+store correctly handles overlapping memory in the memmove 143 // case. 144 unsigned Size = MemOpLength->getZExtValue(); 145 if (Size == 0) return MI; // Delete this mem transfer. 146 147 if (Size > 8 || (Size&(Size-1))) 148 return 0; // If not 1/2/4/8 bytes, exit. 149 150 // Use an integer load+store unless we can find something better. 151 unsigned SrcAddrSp = 152 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace(); 153 unsigned DstAddrSp = 154 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace(); 155 156 const IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3); 157 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp); 158 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp); 159 160 // Memcpy forces the use of i8* for the source and destination. That means 161 // that if you're using memcpy to move one double around, you'll get a cast 162 // from double* to i8*. We'd much rather use a double load+store rather than 163 // an i64 load+store, here because this improves the odds that the source or 164 // dest address will be promotable. See if we can find a better type than the 165 // integer datatype. 166 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts(); 167 if (StrippedDest != MI->getArgOperand(0)) { 168 const Type *SrcETy = cast<PointerType>(StrippedDest->getType()) 169 ->getElementType(); 170 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) { 171 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip 172 // down through these levels if so. 173 while (!SrcETy->isSingleValueType()) { 174 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) { 175 if (STy->getNumElements() == 1) 176 SrcETy = STy->getElementType(0); 177 else 178 break; 179 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) { 180 if (ATy->getNumElements() == 1) 181 SrcETy = ATy->getElementType(); 182 else 183 break; 184 } else 185 break; 186 } 187 188 if (SrcETy->isSingleValueType()) { 189 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp); 190 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp); 191 } 192 } 193 } 194 195 196 // If the memcpy/memmove provides better alignment info than we can 197 // infer, use it. 198 SrcAlign = std::max(SrcAlign, CopyAlign); 199 DstAlign = std::max(DstAlign, CopyAlign); 200 201 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy); 202 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy); 203 Instruction *L = new LoadInst(Src, "tmp", MI->isVolatile(), SrcAlign); 204 InsertNewInstBefore(L, *MI); 205 InsertNewInstBefore(new StoreInst(L, Dest, MI->isVolatile(), DstAlign), 206 *MI); 207 208 // Set the size of the copy to 0, it will be deleted on the next iteration. 209 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType())); 210 return MI; 211} 212 213Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) { 214 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest()); 215 if (MI->getAlignment() < Alignment) { 216 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), 217 Alignment, false)); 218 return MI; 219 } 220 221 // Extract the length and alignment and fill if they are constant. 222 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength()); 223 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue()); 224 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8)) 225 return 0; 226 uint64_t Len = LenC->getZExtValue(); 227 Alignment = MI->getAlignment(); 228 229 // If the length is zero, this is a no-op 230 if (Len == 0) return MI; // memset(d,c,0,a) -> noop 231 232 // memset(s,c,n) -> store s, c (for n=1,2,4,8) 233 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) { 234 const Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8. 235 236 Value *Dest = MI->getDest(); 237 Dest = Builder->CreateBitCast(Dest, PointerType::getUnqual(ITy)); 238 239 // Alignment 0 is identity for alignment 1 for memset, but not store. 240 if (Alignment == 0) Alignment = 1; 241 242 // Extract the fill value and store. 243 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL; 244 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), 245 Dest, false, Alignment), *MI); 246 247 // Set the size of the copy to 0, it will be deleted on the next iteration. 248 MI->setLength(Constant::getNullValue(LenC->getType())); 249 return MI; 250 } 251 252 return 0; 253} 254 255/// visitCallInst - CallInst simplification. This mostly only handles folding 256/// of intrinsic instructions. For normal calls, it allows visitCallSite to do 257/// the heavy lifting. 258/// 259Instruction *InstCombiner::visitCallInst(CallInst &CI) { 260 if (isFreeCall(&CI)) 261 return visitFree(CI); 262 if (isMalloc(&CI)) 263 return visitMalloc(CI); 264 265 // If the caller function is nounwind, mark the call as nounwind, even if the 266 // callee isn't. 267 if (CI.getParent()->getParent()->doesNotThrow() && 268 !CI.doesNotThrow()) { 269 CI.setDoesNotThrow(); 270 return &CI; 271 } 272 273 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI); 274 if (!II) return visitCallSite(&CI); 275 276 // Intrinsics cannot occur in an invoke, so handle them here instead of in 277 // visitCallSite. 278 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) { 279 bool Changed = false; 280 281 // memmove/cpy/set of zero bytes is a noop. 282 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) { 283 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI); 284 285 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes)) 286 if (CI->getZExtValue() == 1) { 287 // Replace the instruction with just byte operations. We would 288 // transform other cases to loads/stores, but we don't know if 289 // alignment is sufficient. 290 } 291 } 292 293 // If we have a memmove and the source operation is a constant global, 294 // then the source and dest pointers can't alias, so we can change this 295 // into a call to memcpy. 296 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) { 297 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource())) 298 if (GVSrc->isConstant()) { 299 Module *M = CI.getParent()->getParent()->getParent(); 300 Intrinsic::ID MemCpyID = Intrinsic::memcpy; 301 const Type *Tys[3] = { CI.getArgOperand(0)->getType(), 302 CI.getArgOperand(1)->getType(), 303 CI.getArgOperand(2)->getType() }; 304 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys, 3)); 305 Changed = true; 306 } 307 } 308 309 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { 310 // memmove(x,x,size) -> noop. 311 if (MTI->getSource() == MTI->getDest()) 312 return EraseInstFromFunction(CI); 313 } 314 315 // If we can determine a pointer alignment that is bigger than currently 316 // set, update the alignment. 317 if (isa<MemTransferInst>(MI)) { 318 if (Instruction *I = SimplifyMemTransfer(MI)) 319 return I; 320 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) { 321 if (Instruction *I = SimplifyMemSet(MSI)) 322 return I; 323 } 324 325 if (Changed) return II; 326 } 327 328 switch (II->getIntrinsicID()) { 329 default: break; 330 case Intrinsic::objectsize: { 331 // We need target data for just about everything so depend on it. 332 if (!TD) break; 333 334 const Type *ReturnTy = CI.getType(); 335 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1); 336 337 // Get to the real allocated thing and offset as fast as possible. 338 Value *Op1 = II->getArgOperand(0)->stripPointerCasts(); 339 340 // If we've stripped down to a single global variable that we 341 // can know the size of then just return that. 342 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) { 343 if (GV->hasDefinitiveInitializer()) { 344 Constant *C = GV->getInitializer(); 345 uint64_t GlobalSize = TD->getTypeAllocSize(C->getType()); 346 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, GlobalSize)); 347 } else { 348 // Can't determine size of the GV. 349 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL); 350 return ReplaceInstUsesWith(CI, RetVal); 351 } 352 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) { 353 // Get alloca size. 354 if (AI->getAllocatedType()->isSized()) { 355 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 356 if (AI->isArrayAllocation()) { 357 const ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize()); 358 if (!C) break; 359 AllocaSize *= C->getZExtValue(); 360 } 361 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, AllocaSize)); 362 } 363 } else if (CallInst *MI = extractMallocCall(Op1)) { 364 const Type* MallocType = getMallocAllocatedType(MI); 365 // Get alloca size. 366 if (MallocType && MallocType->isSized()) { 367 if (Value *NElems = getMallocArraySize(MI, TD, true)) { 368 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems)) 369 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, 370 (NElements->getZExtValue() * TD->getTypeAllocSize(MallocType)))); 371 } 372 } 373 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op1)) { 374 // Only handle constant GEPs here. 375 if (CE->getOpcode() != Instruction::GetElementPtr) break; 376 GEPOperator *GEP = cast<GEPOperator>(CE); 377 378 // Make sure we're not a constant offset from an external 379 // global. 380 Value *Operand = GEP->getPointerOperand(); 381 Operand = Operand->stripPointerCasts(); 382 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Operand)) 383 if (!GV->hasDefinitiveInitializer()) break; 384 385 // Get what we're pointing to and its size. 386 const PointerType *BaseType = 387 cast<PointerType>(Operand->getType()); 388 uint64_t Size = TD->getTypeAllocSize(BaseType->getElementType()); 389 390 // Get the current byte offset into the thing. Use the original 391 // operand in case we're looking through a bitcast. 392 SmallVector<Value*, 8> Ops(CE->op_begin()+1, CE->op_end()); 393 const PointerType *OffsetType = 394 cast<PointerType>(GEP->getPointerOperand()->getType()); 395 uint64_t Offset = TD->getIndexedOffset(OffsetType, &Ops[0], Ops.size()); 396 397 if (Size < Offset) { 398 // Out of bound reference? Negative index normalized to large 399 // index? Just return "I don't know". 400 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL); 401 return ReplaceInstUsesWith(CI, RetVal); 402 } 403 404 Constant *RetVal = ConstantInt::get(ReturnTy, Size-Offset); 405 return ReplaceInstUsesWith(CI, RetVal); 406 } 407 408 // Do not return "I don't know" here. Later optimization passes could 409 // make it possible to evaluate objectsize to a constant. 410 break; 411 } 412 case Intrinsic::bswap: 413 // bswap(bswap(x)) -> x 414 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) 415 if (Operand->getIntrinsicID() == Intrinsic::bswap) 416 return ReplaceInstUsesWith(CI, Operand->getArgOperand(0)); 417 418 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c)) 419 if (TruncInst *TI = dyn_cast<TruncInst>(II->getArgOperand(0))) { 420 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0))) 421 if (Operand->getIntrinsicID() == Intrinsic::bswap) { 422 unsigned C = Operand->getType()->getPrimitiveSizeInBits() - 423 TI->getType()->getPrimitiveSizeInBits(); 424 Value *CV = ConstantInt::get(Operand->getType(), C); 425 Value *V = Builder->CreateLShr(Operand->getArgOperand(0), CV); 426 return new TruncInst(V, TI->getType()); 427 } 428 } 429 430 break; 431 case Intrinsic::powi: 432 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 433 // powi(x, 0) -> 1.0 434 if (Power->isZero()) 435 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0)); 436 // powi(x, 1) -> x 437 if (Power->isOne()) 438 return ReplaceInstUsesWith(CI, II->getArgOperand(0)); 439 // powi(x, -1) -> 1/x 440 if (Power->isAllOnesValue()) 441 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0), 442 II->getArgOperand(0)); 443 } 444 break; 445 case Intrinsic::cttz: { 446 // If all bits below the first known one are known zero, 447 // this value is constant. 448 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType()); 449 uint32_t BitWidth = IT->getBitWidth(); 450 APInt KnownZero(BitWidth, 0); 451 APInt KnownOne(BitWidth, 0); 452 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth), 453 KnownZero, KnownOne); 454 unsigned TrailingZeros = KnownOne.countTrailingZeros(); 455 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros)); 456 if ((Mask & KnownZero) == Mask) 457 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 458 APInt(BitWidth, TrailingZeros))); 459 460 } 461 break; 462 case Intrinsic::ctlz: { 463 // If all bits above the first known one are known zero, 464 // this value is constant. 465 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType()); 466 uint32_t BitWidth = IT->getBitWidth(); 467 APInt KnownZero(BitWidth, 0); 468 APInt KnownOne(BitWidth, 0); 469 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth), 470 KnownZero, KnownOne); 471 unsigned LeadingZeros = KnownOne.countLeadingZeros(); 472 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros)); 473 if ((Mask & KnownZero) == Mask) 474 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 475 APInt(BitWidth, LeadingZeros))); 476 477 } 478 break; 479 case Intrinsic::uadd_with_overflow: { 480 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 481 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType()); 482 uint32_t BitWidth = IT->getBitWidth(); 483 APInt Mask = APInt::getSignBit(BitWidth); 484 APInt LHSKnownZero(BitWidth, 0); 485 APInt LHSKnownOne(BitWidth, 0); 486 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne); 487 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1]; 488 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1]; 489 490 if (LHSKnownNegative || LHSKnownPositive) { 491 APInt RHSKnownZero(BitWidth, 0); 492 APInt RHSKnownOne(BitWidth, 0); 493 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne); 494 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1]; 495 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1]; 496 if (LHSKnownNegative && RHSKnownNegative) { 497 // The sign bit is set in both cases: this MUST overflow. 498 // Create a simple add instruction, and insert it into the struct. 499 Instruction *Add = BinaryOperator::CreateAdd(LHS, RHS, "", &CI); 500 Worklist.Add(Add); 501 Constant *V[] = { 502 UndefValue::get(LHS->getType()),ConstantInt::getTrue(II->getContext()) 503 }; 504 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 505 return InsertValueInst::Create(Struct, Add, 0); 506 } 507 508 if (LHSKnownPositive && RHSKnownPositive) { 509 // The sign bit is clear in both cases: this CANNOT overflow. 510 // Create a simple add instruction, and insert it into the struct. 511 Instruction *Add = BinaryOperator::CreateNUWAdd(LHS, RHS, "", &CI); 512 Worklist.Add(Add); 513 Constant *V[] = { 514 UndefValue::get(LHS->getType()), 515 ConstantInt::getFalse(II->getContext()) 516 }; 517 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 518 return InsertValueInst::Create(Struct, Add, 0); 519 } 520 } 521 } 522 // FALL THROUGH uadd into sadd 523 case Intrinsic::sadd_with_overflow: 524 // Canonicalize constants into the RHS. 525 if (isa<Constant>(II->getArgOperand(0)) && 526 !isa<Constant>(II->getArgOperand(1))) { 527 Value *LHS = II->getArgOperand(0); 528 II->setArgOperand(0, II->getArgOperand(1)); 529 II->setArgOperand(1, LHS); 530 return II; 531 } 532 533 // X + undef -> undef 534 if (isa<UndefValue>(II->getArgOperand(1))) 535 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 536 537 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 538 // X + 0 -> {X, false} 539 if (RHS->isZero()) { 540 Constant *V[] = { 541 UndefValue::get(II->getCalledValue()->getType()), 542 ConstantInt::getFalse(II->getContext()) 543 }; 544 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 545 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 546 } 547 } 548 break; 549 case Intrinsic::usub_with_overflow: 550 case Intrinsic::ssub_with_overflow: 551 // undef - X -> undef 552 // X - undef -> undef 553 if (isa<UndefValue>(II->getArgOperand(0)) || 554 isa<UndefValue>(II->getArgOperand(1))) 555 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 556 557 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 558 // X - 0 -> {X, false} 559 if (RHS->isZero()) { 560 Constant *V[] = { 561 UndefValue::get(II->getArgOperand(0)->getType()), 562 ConstantInt::getFalse(II->getContext()) 563 }; 564 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 565 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 566 } 567 } 568 break; 569 case Intrinsic::umul_with_overflow: 570 case Intrinsic::smul_with_overflow: 571 // Canonicalize constants into the RHS. 572 if (isa<Constant>(II->getArgOperand(0)) && 573 !isa<Constant>(II->getArgOperand(1))) { 574 Value *LHS = II->getArgOperand(0); 575 II->setArgOperand(0, II->getArgOperand(1)); 576 II->setArgOperand(1, LHS); 577 return II; 578 } 579 580 // X * undef -> undef 581 if (isa<UndefValue>(II->getArgOperand(1))) 582 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 583 584 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 585 // X*0 -> {0, false} 586 if (RHSI->isZero()) 587 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType())); 588 589 // X * 1 -> {X, false} 590 if (RHSI->equalsInt(1)) { 591 Constant *V[] = { 592 UndefValue::get(II->getArgOperand(0)->getType()), 593 ConstantInt::getFalse(II->getContext()) 594 }; 595 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 596 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 597 } 598 } 599 break; 600 case Intrinsic::ppc_altivec_lvx: 601 case Intrinsic::ppc_altivec_lvxl: 602 case Intrinsic::x86_sse_loadu_ps: 603 case Intrinsic::x86_sse2_loadu_pd: 604 case Intrinsic::x86_sse2_loadu_dq: 605 // Turn PPC lvx -> load if the pointer is known aligned. 606 // Turn X86 loadups -> load if the pointer is known aligned. 607 if (GetOrEnforceKnownAlignment(II->getArgOperand(0), 16) >= 16) { 608 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), 609 PointerType::getUnqual(II->getType())); 610 return new LoadInst(Ptr); 611 } 612 break; 613 case Intrinsic::ppc_altivec_stvx: 614 case Intrinsic::ppc_altivec_stvxl: 615 // Turn stvx -> store if the pointer is known aligned. 616 if (GetOrEnforceKnownAlignment(II->getArgOperand(1), 16) >= 16) { 617 const Type *OpPtrTy = 618 PointerType::getUnqual(II->getArgOperand(0)->getType()); 619 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); 620 return new StoreInst(II->getArgOperand(0), Ptr); 621 } 622 break; 623 case Intrinsic::x86_sse_storeu_ps: 624 case Intrinsic::x86_sse2_storeu_pd: 625 case Intrinsic::x86_sse2_storeu_dq: 626 // Turn X86 storeu -> store if the pointer is known aligned. 627 if (GetOrEnforceKnownAlignment(II->getArgOperand(0), 16) >= 16) { 628 const Type *OpPtrTy = 629 PointerType::getUnqual(II->getArgOperand(1)->getType()); 630 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy); 631 return new StoreInst(II->getArgOperand(1), Ptr); 632 } 633 break; 634 635 case Intrinsic::x86_sse_cvttss2si: { 636 // These intrinsics only demands the 0th element of its input vector. If 637 // we can simplify the input based on that, do so now. 638 unsigned VWidth = 639 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 640 APInt DemandedElts(VWidth, 1); 641 APInt UndefElts(VWidth, 0); 642 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0), 643 DemandedElts, UndefElts)) { 644 II->setArgOperand(0, V); 645 return II; 646 } 647 break; 648 } 649 650 case Intrinsic::ppc_altivec_vperm: 651 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant. 652 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getArgOperand(2))) { 653 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!"); 654 655 // Check that all of the elements are integer constants or undefs. 656 bool AllEltsOk = true; 657 for (unsigned i = 0; i != 16; ++i) { 658 if (!isa<ConstantInt>(Mask->getOperand(i)) && 659 !isa<UndefValue>(Mask->getOperand(i))) { 660 AllEltsOk = false; 661 break; 662 } 663 } 664 665 if (AllEltsOk) { 666 // Cast the input vectors to byte vectors. 667 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0), 668 Mask->getType()); 669 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1), 670 Mask->getType()); 671 Value *Result = UndefValue::get(Op0->getType()); 672 673 // Only extract each element once. 674 Value *ExtractedElts[32]; 675 memset(ExtractedElts, 0, sizeof(ExtractedElts)); 676 677 for (unsigned i = 0; i != 16; ++i) { 678 if (isa<UndefValue>(Mask->getOperand(i))) 679 continue; 680 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue(); 681 Idx &= 31; // Match the hardware behavior. 682 683 if (ExtractedElts[Idx] == 0) { 684 ExtractedElts[Idx] = 685 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1, 686 ConstantInt::get(Type::getInt32Ty(II->getContext()), 687 Idx&15, false), "tmp"); 688 } 689 690 // Insert this value into the result vector. 691 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx], 692 ConstantInt::get(Type::getInt32Ty(II->getContext()), 693 i, false), "tmp"); 694 } 695 return CastInst::Create(Instruction::BitCast, Result, CI.getType()); 696 } 697 } 698 break; 699 700 case Intrinsic::stackrestore: { 701 // If the save is right next to the restore, remove the restore. This can 702 // happen when variable allocas are DCE'd. 703 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) { 704 if (SS->getIntrinsicID() == Intrinsic::stacksave) { 705 BasicBlock::iterator BI = SS; 706 if (&*++BI == II) 707 return EraseInstFromFunction(CI); 708 } 709 } 710 711 // Scan down this block to see if there is another stack restore in the 712 // same block without an intervening call/alloca. 713 BasicBlock::iterator BI = II; 714 TerminatorInst *TI = II->getParent()->getTerminator(); 715 bool CannotRemove = false; 716 for (++BI; &*BI != TI; ++BI) { 717 if (isa<AllocaInst>(BI) || isMalloc(BI)) { 718 CannotRemove = true; 719 break; 720 } 721 if (CallInst *BCI = dyn_cast<CallInst>(BI)) { 722 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) { 723 // If there is a stackrestore below this one, remove this one. 724 if (II->getIntrinsicID() == Intrinsic::stackrestore) 725 return EraseInstFromFunction(CI); 726 // Otherwise, ignore the intrinsic. 727 } else { 728 // If we found a non-intrinsic call, we can't remove the stack 729 // restore. 730 CannotRemove = true; 731 break; 732 } 733 } 734 } 735 736 // If the stack restore is in a return/unwind block and if there are no 737 // allocas or calls between the restore and the return, nuke the restore. 738 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI))) 739 return EraseInstFromFunction(CI); 740 break; 741 } 742 } 743 744 return visitCallSite(II); 745} 746 747// InvokeInst simplification 748// 749Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) { 750 return visitCallSite(&II); 751} 752 753/// isSafeToEliminateVarargsCast - If this cast does not affect the value 754/// passed through the varargs area, we can eliminate the use of the cast. 755static bool isSafeToEliminateVarargsCast(const CallSite CS, 756 const CastInst * const CI, 757 const TargetData * const TD, 758 const int ix) { 759 if (!CI->isLosslessCast()) 760 return false; 761 762 // The size of ByVal arguments is derived from the type, so we 763 // can't change to a type with a different size. If the size were 764 // passed explicitly we could avoid this check. 765 if (!CS.paramHasAttr(ix, Attribute::ByVal)) 766 return true; 767 768 const Type* SrcTy = 769 cast<PointerType>(CI->getOperand(0)->getType())->getElementType(); 770 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType(); 771 if (!SrcTy->isSized() || !DstTy->isSized()) 772 return false; 773 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy)) 774 return false; 775 return true; 776} 777 778namespace { 779class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls { 780 InstCombiner *IC; 781protected: 782 void replaceCall(Value *With) { 783 NewInstruction = IC->ReplaceInstUsesWith(*CI, With); 784 } 785 bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const { 786 if (ConstantInt *SizeCI = 787 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) { 788 if (SizeCI->isAllOnesValue()) 789 return true; 790 if (isString) 791 return SizeCI->getZExtValue() >= 792 GetStringLength(CI->getArgOperand(SizeArgOp)); 793 if (ConstantInt *Arg = dyn_cast<ConstantInt>( 794 CI->getArgOperand(SizeArgOp))) 795 return SizeCI->getZExtValue() >= Arg->getZExtValue(); 796 } 797 return false; 798 } 799public: 800 InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { } 801 Instruction *NewInstruction; 802}; 803} // end anonymous namespace 804 805// Try to fold some different type of calls here. 806// Currently we're only working with the checking functions, memcpy_chk, 807// mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk, 808// strcat_chk and strncat_chk. 809Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) { 810 if (CI->getCalledFunction() == 0) return 0; 811 812 InstCombineFortifiedLibCalls Simplifier(this); 813 Simplifier.fold(CI, TD); 814 return Simplifier.NewInstruction; 815} 816 817// visitCallSite - Improvements for call and invoke instructions. 818// 819Instruction *InstCombiner::visitCallSite(CallSite CS) { 820 bool Changed = false; 821 822 // If the callee is a constexpr cast of a function, attempt to move the cast 823 // to the arguments of the call/invoke. 824 if (transformConstExprCastCall(CS)) return 0; 825 826 Value *Callee = CS.getCalledValue(); 827 828 if (Function *CalleeF = dyn_cast<Function>(Callee)) 829 // If the call and callee calling conventions don't match, this call must 830 // be unreachable, as the call is undefined. 831 if (CalleeF->getCallingConv() != CS.getCallingConv() && 832 // Only do this for calls to a function with a body. A prototype may 833 // not actually end up matching the implementation's calling conv for a 834 // variety of reasons (e.g. it may be written in assembly). 835 !CalleeF->isDeclaration()) { 836 Instruction *OldCall = CS.getInstruction(); 837 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 838 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 839 OldCall); 840 // If OldCall dues not return void then replaceAllUsesWith undef. 841 // This allows ValueHandlers and custom metadata to adjust itself. 842 if (!OldCall->getType()->isVoidTy()) 843 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType())); 844 if (isa<CallInst>(OldCall)) 845 return EraseInstFromFunction(*OldCall); 846 847 // We cannot remove an invoke, because it would change the CFG, just 848 // change the callee to a null pointer. 849 cast<InvokeInst>(OldCall)->setCalledFunction( 850 Constant::getNullValue(CalleeF->getType())); 851 return 0; 852 } 853 854 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 855 // This instruction is not reachable, just remove it. We insert a store to 856 // undef so that we know that this code is not reachable, despite the fact 857 // that we can't modify the CFG here. 858 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 859 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 860 CS.getInstruction()); 861 862 // If CS does not return void then replaceAllUsesWith undef. 863 // This allows ValueHandlers and custom metadata to adjust itself. 864 if (!CS.getInstruction()->getType()->isVoidTy()) 865 CS.getInstruction()-> 866 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType())); 867 868 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) { 869 // Don't break the CFG, insert a dummy cond branch. 870 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(), 871 ConstantInt::getTrue(Callee->getContext()), II); 872 } 873 return EraseInstFromFunction(*CS.getInstruction()); 874 } 875 876 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee)) 877 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0))) 878 if (In->getIntrinsicID() == Intrinsic::init_trampoline) 879 return transformCallThroughTrampoline(CS); 880 881 const PointerType *PTy = cast<PointerType>(Callee->getType()); 882 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 883 if (FTy->isVarArg()) { 884 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1); 885 // See if we can optimize any arguments passed through the varargs area of 886 // the call. 887 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(), 888 E = CS.arg_end(); I != E; ++I, ++ix) { 889 CastInst *CI = dyn_cast<CastInst>(*I); 890 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) { 891 *I = CI->getOperand(0); 892 Changed = true; 893 } 894 } 895 } 896 897 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) { 898 // Inline asm calls cannot throw - mark them 'nounwind'. 899 CS.setDoesNotThrow(); 900 Changed = true; 901 } 902 903 // Try to optimize the call if possible, we require TargetData for most of 904 // this. None of these calls are seen as possibly dead so go ahead and 905 // delete the instruction now. 906 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) { 907 Instruction *I = tryOptimizeCall(CI, TD); 908 // If we changed something return the result, etc. Otherwise let 909 // the fallthrough check. 910 if (I) return EraseInstFromFunction(*I); 911 } 912 913 return Changed ? CS.getInstruction() : 0; 914} 915 916// transformConstExprCastCall - If the callee is a constexpr cast of a function, 917// attempt to move the cast to the arguments of the call/invoke. 918// 919bool InstCombiner::transformConstExprCastCall(CallSite CS) { 920 if (!isa<ConstantExpr>(CS.getCalledValue())) return false; 921 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue()); 922 if (CE->getOpcode() != Instruction::BitCast || 923 !isa<Function>(CE->getOperand(0))) 924 return false; 925 Function *Callee = cast<Function>(CE->getOperand(0)); 926 Instruction *Caller = CS.getInstruction(); 927 const AttrListPtr &CallerPAL = CS.getAttributes(); 928 929 // Okay, this is a cast from a function to a different type. Unless doing so 930 // would cause a type conversion of one of our arguments, change this call to 931 // be a direct call with arguments casted to the appropriate types. 932 // 933 const FunctionType *FT = Callee->getFunctionType(); 934 const Type *OldRetTy = Caller->getType(); 935 const Type *NewRetTy = FT->getReturnType(); 936 937 if (NewRetTy->isStructTy()) 938 return false; // TODO: Handle multiple return values. 939 940 // Check to see if we are changing the return type... 941 if (OldRetTy != NewRetTy) { 942 if (Callee->isDeclaration() && 943 // Conversion is ok if changing from one pointer type to another or from 944 // a pointer to an integer of the same size. 945 !((OldRetTy->isPointerTy() || !TD || 946 OldRetTy == TD->getIntPtrType(Caller->getContext())) && 947 (NewRetTy->isPointerTy() || !TD || 948 NewRetTy == TD->getIntPtrType(Caller->getContext())))) 949 return false; // Cannot transform this return value. 950 951 if (!Caller->use_empty() && 952 // void -> non-void is handled specially 953 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy)) 954 return false; // Cannot transform this return value. 955 956 if (!CallerPAL.isEmpty() && !Caller->use_empty()) { 957 Attributes RAttrs = CallerPAL.getRetAttributes(); 958 if (RAttrs & Attribute::typeIncompatible(NewRetTy)) 959 return false; // Attribute not compatible with transformed value. 960 } 961 962 // If the callsite is an invoke instruction, and the return value is used by 963 // a PHI node in a successor, we cannot change the return type of the call 964 // because there is no place to put the cast instruction (without breaking 965 // the critical edge). Bail out in this case. 966 if (!Caller->use_empty()) 967 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) 968 for (Value::use_iterator UI = II->use_begin(), E = II->use_end(); 969 UI != E; ++UI) 970 if (PHINode *PN = dyn_cast<PHINode>(*UI)) 971 if (PN->getParent() == II->getNormalDest() || 972 PN->getParent() == II->getUnwindDest()) 973 return false; 974 } 975 976 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin()); 977 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs); 978 979 CallSite::arg_iterator AI = CS.arg_begin(); 980 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) { 981 const Type *ParamTy = FT->getParamType(i); 982 const Type *ActTy = (*AI)->getType(); 983 984 if (!CastInst::isCastable(ActTy, ParamTy)) 985 return false; // Cannot transform this parameter value. 986 987 if (CallerPAL.getParamAttributes(i + 1) 988 & Attribute::typeIncompatible(ParamTy)) 989 return false; // Attribute not compatible with transformed value. 990 991 // Converting from one pointer type to another or between a pointer and an 992 // integer of the same size is safe even if we do not have a body. 993 bool isConvertible = ActTy == ParamTy || 994 (TD && ((ParamTy->isPointerTy() || 995 ParamTy == TD->getIntPtrType(Caller->getContext())) && 996 (ActTy->isPointerTy() || 997 ActTy == TD->getIntPtrType(Caller->getContext())))); 998 if (Callee->isDeclaration() && !isConvertible) return false; 999 } 1000 1001 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() && 1002 Callee->isDeclaration()) 1003 return false; // Do not delete arguments unless we have a function body. 1004 1005 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && 1006 !CallerPAL.isEmpty()) 1007 // In this case we have more arguments than the new function type, but we 1008 // won't be dropping them. Check that these extra arguments have attributes 1009 // that are compatible with being a vararg call argument. 1010 for (unsigned i = CallerPAL.getNumSlots(); i; --i) { 1011 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams()) 1012 break; 1013 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs; 1014 if (PAttrs & Attribute::VarArgsIncompatible) 1015 return false; 1016 } 1017 1018 // Okay, we decided that this is a safe thing to do: go ahead and start 1019 // inserting cast instructions as necessary... 1020 std::vector<Value*> Args; 1021 Args.reserve(NumActualArgs); 1022 SmallVector<AttributeWithIndex, 8> attrVec; 1023 attrVec.reserve(NumCommonArgs); 1024 1025 // Get any return attributes. 1026 Attributes RAttrs = CallerPAL.getRetAttributes(); 1027 1028 // If the return value is not being used, the type may not be compatible 1029 // with the existing attributes. Wipe out any problematic attributes. 1030 RAttrs &= ~Attribute::typeIncompatible(NewRetTy); 1031 1032 // Add the new return attributes. 1033 if (RAttrs) 1034 attrVec.push_back(AttributeWithIndex::get(0, RAttrs)); 1035 1036 AI = CS.arg_begin(); 1037 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) { 1038 const Type *ParamTy = FT->getParamType(i); 1039 if ((*AI)->getType() == ParamTy) { 1040 Args.push_back(*AI); 1041 } else { 1042 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, 1043 false, ParamTy, false); 1044 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp")); 1045 } 1046 1047 // Add any parameter attributes. 1048 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1049 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1050 } 1051 1052 // If the function takes more arguments than the call was taking, add them 1053 // now. 1054 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) 1055 Args.push_back(Constant::getNullValue(FT->getParamType(i))); 1056 1057 // If we are removing arguments to the function, emit an obnoxious warning. 1058 if (FT->getNumParams() < NumActualArgs) { 1059 if (!FT->isVarArg()) { 1060 errs() << "WARNING: While resolving call to function '" 1061 << Callee->getName() << "' arguments were dropped!\n"; 1062 } else { 1063 // Add all of the arguments in their promoted form to the arg list. 1064 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) { 1065 const Type *PTy = getPromotedType((*AI)->getType()); 1066 if (PTy != (*AI)->getType()) { 1067 // Must promote to pass through va_arg area! 1068 Instruction::CastOps opcode = 1069 CastInst::getCastOpcode(*AI, false, PTy, false); 1070 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp")); 1071 } else { 1072 Args.push_back(*AI); 1073 } 1074 1075 // Add any parameter attributes. 1076 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1077 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1078 } 1079 } 1080 } 1081 1082 if (Attributes FnAttrs = CallerPAL.getFnAttributes()) 1083 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); 1084 1085 if (NewRetTy->isVoidTy()) 1086 Caller->setName(""); // Void type should not have a name. 1087 1088 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(), 1089 attrVec.end()); 1090 1091 Instruction *NC; 1092 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1093 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(), 1094 Args.begin(), Args.end(), 1095 Caller->getName(), Caller); 1096 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv()); 1097 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL); 1098 } else { 1099 NC = CallInst::Create(Callee, Args.begin(), Args.end(), 1100 Caller->getName(), Caller); 1101 CallInst *CI = cast<CallInst>(Caller); 1102 if (CI->isTailCall()) 1103 cast<CallInst>(NC)->setTailCall(); 1104 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv()); 1105 cast<CallInst>(NC)->setAttributes(NewCallerPAL); 1106 } 1107 1108 // Insert a cast of the return type as necessary. 1109 Value *NV = NC; 1110 if (OldRetTy != NV->getType() && !Caller->use_empty()) { 1111 if (!NV->getType()->isVoidTy()) { 1112 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false, 1113 OldRetTy, false); 1114 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp"); 1115 1116 // If this is an invoke instruction, we should insert it after the first 1117 // non-phi, instruction in the normal successor block. 1118 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1119 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI(); 1120 InsertNewInstBefore(NC, *I); 1121 } else { 1122 // Otherwise, it's a call, just insert cast right after the call instr 1123 InsertNewInstBefore(NC, *Caller); 1124 } 1125 Worklist.AddUsersToWorkList(*Caller); 1126 } else { 1127 NV = UndefValue::get(Caller->getType()); 1128 } 1129 } 1130 1131 1132 if (!Caller->use_empty()) 1133 Caller->replaceAllUsesWith(NV); 1134 1135 EraseInstFromFunction(*Caller); 1136 return true; 1137} 1138 1139// transformCallThroughTrampoline - Turn a call to a function created by the 1140// init_trampoline intrinsic into a direct call to the underlying function. 1141// 1142Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) { 1143 Value *Callee = CS.getCalledValue(); 1144 const PointerType *PTy = cast<PointerType>(Callee->getType()); 1145 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 1146 const AttrListPtr &Attrs = CS.getAttributes(); 1147 1148 // If the call already has the 'nest' attribute somewhere then give up - 1149 // otherwise 'nest' would occur twice after splicing in the chain. 1150 if (Attrs.hasAttrSomewhere(Attribute::Nest)) 1151 return 0; 1152 1153 IntrinsicInst *Tramp = 1154 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0)); 1155 1156 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts()); 1157 const PointerType *NestFPTy = cast<PointerType>(NestF->getType()); 1158 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType()); 1159 1160 const AttrListPtr &NestAttrs = NestF->getAttributes(); 1161 if (!NestAttrs.isEmpty()) { 1162 unsigned NestIdx = 1; 1163 const Type *NestTy = 0; 1164 Attributes NestAttr = Attribute::None; 1165 1166 // Look for a parameter marked with the 'nest' attribute. 1167 for (FunctionType::param_iterator I = NestFTy->param_begin(), 1168 E = NestFTy->param_end(); I != E; ++NestIdx, ++I) 1169 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) { 1170 // Record the parameter type and any other attributes. 1171 NestTy = *I; 1172 NestAttr = NestAttrs.getParamAttributes(NestIdx); 1173 break; 1174 } 1175 1176 if (NestTy) { 1177 Instruction *Caller = CS.getInstruction(); 1178 std::vector<Value*> NewArgs; 1179 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1); 1180 1181 SmallVector<AttributeWithIndex, 8> NewAttrs; 1182 NewAttrs.reserve(Attrs.getNumSlots() + 1); 1183 1184 // Insert the nest argument into the call argument list, which may 1185 // mean appending it. Likewise for attributes. 1186 1187 // Add any result attributes. 1188 if (Attributes Attr = Attrs.getRetAttributes()) 1189 NewAttrs.push_back(AttributeWithIndex::get(0, Attr)); 1190 1191 { 1192 unsigned Idx = 1; 1193 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 1194 do { 1195 if (Idx == NestIdx) { 1196 // Add the chain argument and attributes. 1197 Value *NestVal = Tramp->getArgOperand(2); 1198 if (NestVal->getType() != NestTy) 1199 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller); 1200 NewArgs.push_back(NestVal); 1201 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr)); 1202 } 1203 1204 if (I == E) 1205 break; 1206 1207 // Add the original argument and attributes. 1208 NewArgs.push_back(*I); 1209 if (Attributes Attr = Attrs.getParamAttributes(Idx)) 1210 NewAttrs.push_back 1211 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr)); 1212 1213 ++Idx, ++I; 1214 } while (1); 1215 } 1216 1217 // Add any function attributes. 1218 if (Attributes Attr = Attrs.getFnAttributes()) 1219 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr)); 1220 1221 // The trampoline may have been bitcast to a bogus type (FTy). 1222 // Handle this by synthesizing a new function type, equal to FTy 1223 // with the chain parameter inserted. 1224 1225 std::vector<const Type*> NewTypes; 1226 NewTypes.reserve(FTy->getNumParams()+1); 1227 1228 // Insert the chain's type into the list of parameter types, which may 1229 // mean appending it. 1230 { 1231 unsigned Idx = 1; 1232 FunctionType::param_iterator I = FTy->param_begin(), 1233 E = FTy->param_end(); 1234 1235 do { 1236 if (Idx == NestIdx) 1237 // Add the chain's type. 1238 NewTypes.push_back(NestTy); 1239 1240 if (I == E) 1241 break; 1242 1243 // Add the original type. 1244 NewTypes.push_back(*I); 1245 1246 ++Idx, ++I; 1247 } while (1); 1248 } 1249 1250 // Replace the trampoline call with a direct call. Let the generic 1251 // code sort out any function type mismatches. 1252 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes, 1253 FTy->isVarArg()); 1254 Constant *NewCallee = 1255 NestF->getType() == PointerType::getUnqual(NewFTy) ? 1256 NestF : ConstantExpr::getBitCast(NestF, 1257 PointerType::getUnqual(NewFTy)); 1258 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(), 1259 NewAttrs.end()); 1260 1261 Instruction *NewCaller; 1262 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1263 NewCaller = InvokeInst::Create(NewCallee, 1264 II->getNormalDest(), II->getUnwindDest(), 1265 NewArgs.begin(), NewArgs.end(), 1266 Caller->getName(), Caller); 1267 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv()); 1268 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL); 1269 } else { 1270 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(), 1271 Caller->getName(), Caller); 1272 if (cast<CallInst>(Caller)->isTailCall()) 1273 cast<CallInst>(NewCaller)->setTailCall(); 1274 cast<CallInst>(NewCaller)-> 1275 setCallingConv(cast<CallInst>(Caller)->getCallingConv()); 1276 cast<CallInst>(NewCaller)->setAttributes(NewPAL); 1277 } 1278 if (!Caller->getType()->isVoidTy()) 1279 Caller->replaceAllUsesWith(NewCaller); 1280 Caller->eraseFromParent(); 1281 Worklist.Remove(Caller); 1282 return 0; 1283 } 1284 } 1285 1286 // Replace the trampoline call with a direct call. Since there is no 'nest' 1287 // parameter, there is no need to adjust the argument list. Let the generic 1288 // code sort out any function type mismatches. 1289 Constant *NewCallee = 1290 NestF->getType() == PTy ? NestF : 1291 ConstantExpr::getBitCast(NestF, PTy); 1292 CS.setCalledFunction(NewCallee); 1293 return CS.getInstruction(); 1294} 1295 1296