InstCombineCalls.cpp revision 75f0d6953a8492ec21f3b82489090a1152f1f516
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" 19using namespace llvm; 20 21/// getPromotedType - Return the specified type promoted as it would be to pass 22/// though a va_arg area. 23static const Type *getPromotedType(const Type *Ty) { 24 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) { 25 if (ITy->getBitWidth() < 32) 26 return Type::getInt32Ty(Ty->getContext()); 27 } 28 return Ty; 29} 30 31/// EnforceKnownAlignment - If the specified pointer points to an object that 32/// we control, modify the object's alignment to PrefAlign. This isn't 33/// often possible though. If alignment is important, a more reliable approach 34/// is to simply align all global variables and allocation instructions to 35/// their preferred alignment from the beginning. 36/// 37static unsigned EnforceKnownAlignment(Value *V, 38 unsigned Align, unsigned PrefAlign) { 39 40 User *U = dyn_cast<User>(V); 41 if (!U) return Align; 42 43 switch (Operator::getOpcode(U)) { 44 default: break; 45 case Instruction::BitCast: 46 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign); 47 case Instruction::GetElementPtr: { 48 // If all indexes are zero, it is just the alignment of the base pointer. 49 bool AllZeroOperands = true; 50 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i) 51 if (!isa<Constant>(*i) || 52 !cast<Constant>(*i)->isNullValue()) { 53 AllZeroOperands = false; 54 break; 55 } 56 57 if (AllZeroOperands) { 58 // Treat this like a bitcast. 59 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign); 60 } 61 break; 62 } 63 } 64 65 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) { 66 // If there is a large requested alignment and we can, bump up the alignment 67 // of the global. 68 if (!GV->isDeclaration()) { 69 if (GV->getAlignment() >= PrefAlign) 70 Align = GV->getAlignment(); 71 else { 72 GV->setAlignment(PrefAlign); 73 Align = PrefAlign; 74 } 75 } 76 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) { 77 // If there is a requested alignment and if this is an alloca, round up. 78 if (AI->getAlignment() >= PrefAlign) 79 Align = AI->getAlignment(); 80 else { 81 AI->setAlignment(PrefAlign); 82 Align = PrefAlign; 83 } 84 } 85 86 return Align; 87} 88 89/// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that 90/// we can determine, return it, otherwise return 0. If PrefAlign is specified, 91/// and it is more than the alignment of the ultimate object, see if we can 92/// increase the alignment of the ultimate object, making this check succeed. 93unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V, 94 unsigned PrefAlign) { 95 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) : 96 sizeof(PrefAlign) * CHAR_BIT; 97 APInt Mask = APInt::getAllOnesValue(BitWidth); 98 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 99 ComputeMaskedBits(V, Mask, KnownZero, KnownOne); 100 unsigned TrailZ = KnownZero.countTrailingOnes(); 101 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ); 102 103 if (PrefAlign > Align) 104 Align = EnforceKnownAlignment(V, Align, PrefAlign); 105 106 // We don't need to make any adjustment. 107 return Align; 108} 109 110Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) { 111 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1)); 112 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2)); 113 unsigned MinAlign = std::min(DstAlign, SrcAlign); 114 unsigned CopyAlign = MI->getAlignment(); 115 116 if (CopyAlign < MinAlign) { 117 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), 118 MinAlign, false)); 119 return MI; 120 } 121 122 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with 123 // load/store. 124 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3)); 125 if (MemOpLength == 0) return 0; 126 127 // Source and destination pointer types are always "i8*" for intrinsic. See 128 // if the size is something we can handle with a single primitive load/store. 129 // A single load+store correctly handles overlapping memory in the memmove 130 // case. 131 unsigned Size = MemOpLength->getZExtValue(); 132 if (Size == 0) return MI; // Delete this mem transfer. 133 134 if (Size > 8 || (Size&(Size-1))) 135 return 0; // If not 1/2/4/8 bytes, exit. 136 137 // Use an integer load+store unless we can find something better. 138 Type *NewPtrTy = 139 PointerType::getUnqual(IntegerType::get(MI->getContext(), Size<<3)); 140 141 // Memcpy forces the use of i8* for the source and destination. That means 142 // that if you're using memcpy to move one double around, you'll get a cast 143 // from double* to i8*. We'd much rather use a double load+store rather than 144 // an i64 load+store, here because this improves the odds that the source or 145 // dest address will be promotable. See if we can find a better type than the 146 // integer datatype. 147 Value *StrippedDest = MI->getOperand(1)->stripPointerCasts(); 148 if (StrippedDest != MI->getOperand(1)) { 149 const Type *SrcETy = cast<PointerType>(StrippedDest->getType()) 150 ->getElementType(); 151 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) { 152 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip 153 // down through these levels if so. 154 while (!SrcETy->isSingleValueType()) { 155 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) { 156 if (STy->getNumElements() == 1) 157 SrcETy = STy->getElementType(0); 158 else 159 break; 160 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) { 161 if (ATy->getNumElements() == 1) 162 SrcETy = ATy->getElementType(); 163 else 164 break; 165 } else 166 break; 167 } 168 169 if (SrcETy->isSingleValueType()) 170 NewPtrTy = PointerType::getUnqual(SrcETy); 171 } 172 } 173 174 175 // If the memcpy/memmove provides better alignment info than we can 176 // infer, use it. 177 SrcAlign = std::max(SrcAlign, CopyAlign); 178 DstAlign = std::max(DstAlign, CopyAlign); 179 180 Value *Src = Builder->CreateBitCast(MI->getOperand(2), NewPtrTy); 181 Value *Dest = Builder->CreateBitCast(MI->getOperand(1), NewPtrTy); 182 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign); 183 InsertNewInstBefore(L, *MI); 184 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI); 185 186 // Set the size of the copy to 0, it will be deleted on the next iteration. 187 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType())); 188 return MI; 189} 190 191Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) { 192 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest()); 193 if (MI->getAlignment() < Alignment) { 194 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), 195 Alignment, false)); 196 return MI; 197 } 198 199 // Extract the length and alignment and fill if they are constant. 200 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength()); 201 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue()); 202 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8)) 203 return 0; 204 uint64_t Len = LenC->getZExtValue(); 205 Alignment = MI->getAlignment(); 206 207 // If the length is zero, this is a no-op 208 if (Len == 0) return MI; // memset(d,c,0,a) -> noop 209 210 // memset(s,c,n) -> store s, c (for n=1,2,4,8) 211 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) { 212 const Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8. 213 214 Value *Dest = MI->getDest(); 215 Dest = Builder->CreateBitCast(Dest, PointerType::getUnqual(ITy)); 216 217 // Alignment 0 is identity for alignment 1 for memset, but not store. 218 if (Alignment == 0) Alignment = 1; 219 220 // Extract the fill value and store. 221 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL; 222 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), 223 Dest, false, Alignment), *MI); 224 225 // Set the size of the copy to 0, it will be deleted on the next iteration. 226 MI->setLength(Constant::getNullValue(LenC->getType())); 227 return MI; 228 } 229 230 return 0; 231} 232 233/// visitCallInst - CallInst simplification. This mostly only handles folding 234/// of intrinsic instructions. For normal calls, it allows visitCallSite to do 235/// the heavy lifting. 236/// 237Instruction *InstCombiner::visitCallInst(CallInst &CI) { 238 if (isFreeCall(&CI)) 239 return visitFree(CI); 240 241 // If the caller function is nounwind, mark the call as nounwind, even if the 242 // callee isn't. 243 if (CI.getParent()->getParent()->doesNotThrow() && 244 !CI.doesNotThrow()) { 245 CI.setDoesNotThrow(); 246 return &CI; 247 } 248 249 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI); 250 if (!II) return visitCallSite(&CI); 251 252 // Intrinsics cannot occur in an invoke, so handle them here instead of in 253 // visitCallSite. 254 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) { 255 bool Changed = false; 256 257 // memmove/cpy/set of zero bytes is a noop. 258 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) { 259 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI); 260 261 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes)) 262 if (CI->getZExtValue() == 1) { 263 // Replace the instruction with just byte operations. We would 264 // transform other cases to loads/stores, but we don't know if 265 // alignment is sufficient. 266 } 267 } 268 269 // If we have a memmove and the source operation is a constant global, 270 // then the source and dest pointers can't alias, so we can change this 271 // into a call to memcpy. 272 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) { 273 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource())) 274 if (GVSrc->isConstant()) { 275 Module *M = CI.getParent()->getParent()->getParent(); 276 Intrinsic::ID MemCpyID = Intrinsic::memcpy; 277 const Type *Tys[1]; 278 Tys[0] = CI.getOperand(3)->getType(); 279 CI.setOperand(0, 280 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1)); 281 Changed = true; 282 } 283 } 284 285 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { 286 // memmove(x,x,size) -> noop. 287 if (MTI->getSource() == MTI->getDest()) 288 return EraseInstFromFunction(CI); 289 } 290 291 // If we can determine a pointer alignment that is bigger than currently 292 // set, update the alignment. 293 if (isa<MemTransferInst>(MI)) { 294 if (Instruction *I = SimplifyMemTransfer(MI)) 295 return I; 296 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) { 297 if (Instruction *I = SimplifyMemSet(MSI)) 298 return I; 299 } 300 301 if (Changed) return II; 302 } 303 304 switch (II->getIntrinsicID()) { 305 default: break; 306 case Intrinsic::objectsize: { 307 // We need target data for just about everything so depend on it. 308 if (!TD) break; 309 310 const Type *ReturnTy = CI.getType(); 311 bool Min = (cast<ConstantInt>(II->getOperand(2))->getZExtValue() == 1); 312 313 // Get to the real allocated thing and offset as fast as possible. 314 Value *Op1 = II->getOperand(1)->stripPointerCasts(); 315 316 // If we've stripped down to a single global variable that we 317 // can know the size of then just return that. 318 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) { 319 if (GV->hasDefinitiveInitializer()) { 320 Constant *C = GV->getInitializer(); 321 uint64_t GlobalSize = TD->getTypeAllocSize(C->getType()); 322 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, GlobalSize)); 323 } else { 324 // Can't determine size of the GV. 325 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL); 326 return ReplaceInstUsesWith(CI, RetVal); 327 } 328 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) { 329 // Get alloca size. 330 if (AI->getAllocatedType()->isSized()) { 331 uint64_t AllocaSize = TD->getTypeAllocSize(AI->getAllocatedType()); 332 if (AI->isArrayAllocation()) { 333 const ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize()); 334 if (!C) break; 335 AllocaSize *= C->getZExtValue(); 336 } 337 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, AllocaSize)); 338 } 339 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op1)) { 340 // Only handle constant GEPs here. 341 if (CE->getOpcode() != Instruction::GetElementPtr) break; 342 GEPOperator *GEP = cast<GEPOperator>(CE); 343 344 // Make sure we're not a constant offset from an external 345 // global. 346 Value *Operand = GEP->getPointerOperand(); 347 Operand = Operand->stripPointerCasts(); 348 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Operand)) 349 if (!GV->hasDefinitiveInitializer()) break; 350 351 // Get what we're pointing to and its size. 352 const PointerType *BaseType = 353 cast<PointerType>(Operand->getType()); 354 uint64_t Size = TD->getTypeAllocSize(BaseType->getElementType()); 355 356 // Get the current byte offset into the thing. Use the original 357 // operand in case we're looking through a bitcast. 358 SmallVector<Value*, 8> Ops(CE->op_begin()+1, CE->op_end()); 359 const PointerType *OffsetType = 360 cast<PointerType>(GEP->getPointerOperand()->getType()); 361 uint64_t Offset = TD->getIndexedOffset(OffsetType, &Ops[0], Ops.size()); 362 363 if (Size < Offset) { 364 // Out of bound reference? Negative index normalized to large 365 // index? Just return "I don't know". 366 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL); 367 return ReplaceInstUsesWith(CI, RetVal); 368 } 369 370 Constant *RetVal = ConstantInt::get(ReturnTy, Size-Offset); 371 return ReplaceInstUsesWith(CI, RetVal); 372 373 } 374 375 // Do not return "I don't know" here. Later optimization passes could 376 // make it possible to evaluate objectsize to a constant. 377 break; 378 } 379 case Intrinsic::bswap: 380 // bswap(bswap(x)) -> x 381 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1))) 382 if (Operand->getIntrinsicID() == Intrinsic::bswap) 383 return ReplaceInstUsesWith(CI, Operand->getOperand(1)); 384 385 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c)) 386 if (TruncInst *TI = dyn_cast<TruncInst>(II->getOperand(1))) { 387 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0))) 388 if (Operand->getIntrinsicID() == Intrinsic::bswap) { 389 unsigned C = Operand->getType()->getPrimitiveSizeInBits() - 390 TI->getType()->getPrimitiveSizeInBits(); 391 Value *CV = ConstantInt::get(Operand->getType(), C); 392 Value *V = Builder->CreateLShr(Operand->getOperand(1), CV); 393 return new TruncInst(V, TI->getType()); 394 } 395 } 396 397 break; 398 case Intrinsic::powi: 399 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getOperand(2))) { 400 // powi(x, 0) -> 1.0 401 if (Power->isZero()) 402 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0)); 403 // powi(x, 1) -> x 404 if (Power->isOne()) 405 return ReplaceInstUsesWith(CI, II->getOperand(1)); 406 // powi(x, -1) -> 1/x 407 if (Power->isAllOnesValue()) 408 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0), 409 II->getOperand(1)); 410 } 411 break; 412 case Intrinsic::cttz: { 413 // If all bits below the first known one are known zero, 414 // this value is constant. 415 const IntegerType *IT = cast<IntegerType>(II->getOperand(1)->getType()); 416 uint32_t BitWidth = IT->getBitWidth(); 417 APInt KnownZero(BitWidth, 0); 418 APInt KnownOne(BitWidth, 0); 419 ComputeMaskedBits(II->getOperand(1), APInt::getAllOnesValue(BitWidth), 420 KnownZero, KnownOne); 421 unsigned TrailingZeros = KnownOne.countTrailingZeros(); 422 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros)); 423 if ((Mask & KnownZero) == Mask) 424 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 425 APInt(BitWidth, TrailingZeros))); 426 427 } 428 break; 429 case Intrinsic::ctlz: { 430 // If all bits above the first known one are known zero, 431 // this value is constant. 432 const IntegerType *IT = cast<IntegerType>(II->getOperand(1)->getType()); 433 uint32_t BitWidth = IT->getBitWidth(); 434 APInt KnownZero(BitWidth, 0); 435 APInt KnownOne(BitWidth, 0); 436 ComputeMaskedBits(II->getOperand(1), APInt::getAllOnesValue(BitWidth), 437 KnownZero, KnownOne); 438 unsigned LeadingZeros = KnownOne.countLeadingZeros(); 439 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros)); 440 if ((Mask & KnownZero) == Mask) 441 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 442 APInt(BitWidth, LeadingZeros))); 443 444 } 445 break; 446 case Intrinsic::uadd_with_overflow: { 447 Value *LHS = II->getOperand(1), *RHS = II->getOperand(2); 448 const IntegerType *IT = cast<IntegerType>(II->getOperand(1)->getType()); 449 uint32_t BitWidth = IT->getBitWidth(); 450 APInt Mask = APInt::getSignBit(BitWidth); 451 APInt LHSKnownZero(BitWidth, 0); 452 APInt LHSKnownOne(BitWidth, 0); 453 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne); 454 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1]; 455 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1]; 456 457 if (LHSKnownNegative || LHSKnownPositive) { 458 APInt RHSKnownZero(BitWidth, 0); 459 APInt RHSKnownOne(BitWidth, 0); 460 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne); 461 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1]; 462 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1]; 463 if (LHSKnownNegative && RHSKnownNegative) { 464 // The sign bit is set in both cases: this MUST overflow. 465 // Create a simple add instruction, and insert it into the struct. 466 Instruction *Add = BinaryOperator::CreateAdd(LHS, RHS, "", &CI); 467 Worklist.Add(Add); 468 Constant *V[] = { 469 UndefValue::get(LHS->getType()),ConstantInt::getTrue(II->getContext()) 470 }; 471 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 472 return InsertValueInst::Create(Struct, Add, 0); 473 } 474 475 if (LHSKnownPositive && RHSKnownPositive) { 476 // The sign bit is clear in both cases: this CANNOT overflow. 477 // Create a simple add instruction, and insert it into the struct. 478 Instruction *Add = BinaryOperator::CreateNUWAdd(LHS, RHS, "", &CI); 479 Worklist.Add(Add); 480 Constant *V[] = { 481 UndefValue::get(LHS->getType()), 482 ConstantInt::getFalse(II->getContext()) 483 }; 484 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 485 return InsertValueInst::Create(Struct, Add, 0); 486 } 487 } 488 } 489 // FALL THROUGH uadd into sadd 490 case Intrinsic::sadd_with_overflow: 491 // Canonicalize constants into the RHS. 492 if (isa<Constant>(II->getOperand(1)) && 493 !isa<Constant>(II->getOperand(2))) { 494 Value *LHS = II->getOperand(1); 495 II->setOperand(1, II->getOperand(2)); 496 II->setOperand(2, LHS); 497 return II; 498 } 499 500 // X + undef -> undef 501 if (isa<UndefValue>(II->getOperand(2))) 502 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 503 504 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getOperand(2))) { 505 // X + 0 -> {X, false} 506 if (RHS->isZero()) { 507 Constant *V[] = { 508 UndefValue::get(II->getOperand(0)->getType()), 509 ConstantInt::getFalse(II->getContext()) 510 }; 511 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 512 return InsertValueInst::Create(Struct, II->getOperand(1), 0); 513 } 514 } 515 break; 516 case Intrinsic::usub_with_overflow: 517 case Intrinsic::ssub_with_overflow: 518 // undef - X -> undef 519 // X - undef -> undef 520 if (isa<UndefValue>(II->getOperand(1)) || 521 isa<UndefValue>(II->getOperand(2))) 522 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 523 524 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getOperand(2))) { 525 // X - 0 -> {X, false} 526 if (RHS->isZero()) { 527 Constant *V[] = { 528 UndefValue::get(II->getOperand(1)->getType()), 529 ConstantInt::getFalse(II->getContext()) 530 }; 531 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 532 return InsertValueInst::Create(Struct, II->getOperand(1), 0); 533 } 534 } 535 break; 536 case Intrinsic::umul_with_overflow: 537 case Intrinsic::smul_with_overflow: 538 // Canonicalize constants into the RHS. 539 if (isa<Constant>(II->getOperand(1)) && 540 !isa<Constant>(II->getOperand(2))) { 541 Value *LHS = II->getOperand(1); 542 II->setOperand(1, II->getOperand(2)); 543 II->setOperand(2, LHS); 544 return II; 545 } 546 547 // X * undef -> undef 548 if (isa<UndefValue>(II->getOperand(2))) 549 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 550 551 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getOperand(2))) { 552 // X*0 -> {0, false} 553 if (RHSI->isZero()) 554 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType())); 555 556 // X * 1 -> {X, false} 557 if (RHSI->equalsInt(1)) { 558 Constant *V[] = { 559 UndefValue::get(II->getOperand(1)->getType()), 560 ConstantInt::getFalse(II->getContext()) 561 }; 562 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 563 return InsertValueInst::Create(Struct, II->getOperand(1), 0); 564 } 565 } 566 break; 567 case Intrinsic::ppc_altivec_lvx: 568 case Intrinsic::ppc_altivec_lvxl: 569 case Intrinsic::x86_sse_loadu_ps: 570 case Intrinsic::x86_sse2_loadu_pd: 571 case Intrinsic::x86_sse2_loadu_dq: 572 // Turn PPC lvx -> load if the pointer is known aligned. 573 // Turn X86 loadups -> load if the pointer is known aligned. 574 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) { 575 Value *Ptr = Builder->CreateBitCast(II->getOperand(1), 576 PointerType::getUnqual(II->getType())); 577 return new LoadInst(Ptr); 578 } 579 break; 580 case Intrinsic::ppc_altivec_stvx: 581 case Intrinsic::ppc_altivec_stvxl: 582 // Turn stvx -> store if the pointer is known aligned. 583 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) { 584 const Type *OpPtrTy = 585 PointerType::getUnqual(II->getOperand(1)->getType()); 586 Value *Ptr = Builder->CreateBitCast(II->getOperand(2), OpPtrTy); 587 return new StoreInst(II->getOperand(1), Ptr); 588 } 589 break; 590 case Intrinsic::x86_sse_storeu_ps: 591 case Intrinsic::x86_sse2_storeu_pd: 592 case Intrinsic::x86_sse2_storeu_dq: 593 // Turn X86 storeu -> store if the pointer is known aligned. 594 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) { 595 const Type *OpPtrTy = 596 PointerType::getUnqual(II->getOperand(2)->getType()); 597 Value *Ptr = Builder->CreateBitCast(II->getOperand(1), OpPtrTy); 598 return new StoreInst(II->getOperand(2), Ptr); 599 } 600 break; 601 602 case Intrinsic::x86_sse_cvttss2si: { 603 // These intrinsics only demands the 0th element of its input vector. If 604 // we can simplify the input based on that, do so now. 605 unsigned VWidth = 606 cast<VectorType>(II->getOperand(1)->getType())->getNumElements(); 607 APInt DemandedElts(VWidth, 1); 608 APInt UndefElts(VWidth, 0); 609 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts, 610 UndefElts)) { 611 II->setOperand(1, V); 612 return II; 613 } 614 break; 615 } 616 617 case Intrinsic::ppc_altivec_vperm: 618 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant. 619 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) { 620 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!"); 621 622 // Check that all of the elements are integer constants or undefs. 623 bool AllEltsOk = true; 624 for (unsigned i = 0; i != 16; ++i) { 625 if (!isa<ConstantInt>(Mask->getOperand(i)) && 626 !isa<UndefValue>(Mask->getOperand(i))) { 627 AllEltsOk = false; 628 break; 629 } 630 } 631 632 if (AllEltsOk) { 633 // Cast the input vectors to byte vectors. 634 Value *Op0 = Builder->CreateBitCast(II->getOperand(1), Mask->getType()); 635 Value *Op1 = Builder->CreateBitCast(II->getOperand(2), Mask->getType()); 636 Value *Result = UndefValue::get(Op0->getType()); 637 638 // Only extract each element once. 639 Value *ExtractedElts[32]; 640 memset(ExtractedElts, 0, sizeof(ExtractedElts)); 641 642 for (unsigned i = 0; i != 16; ++i) { 643 if (isa<UndefValue>(Mask->getOperand(i))) 644 continue; 645 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue(); 646 Idx &= 31; // Match the hardware behavior. 647 648 if (ExtractedElts[Idx] == 0) { 649 ExtractedElts[Idx] = 650 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1, 651 ConstantInt::get(Type::getInt32Ty(II->getContext()), 652 Idx&15, false), "tmp"); 653 } 654 655 // Insert this value into the result vector. 656 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx], 657 ConstantInt::get(Type::getInt32Ty(II->getContext()), 658 i, false), "tmp"); 659 } 660 return CastInst::Create(Instruction::BitCast, Result, CI.getType()); 661 } 662 } 663 break; 664 665 case Intrinsic::stackrestore: { 666 // If the save is right next to the restore, remove the restore. This can 667 // happen when variable allocas are DCE'd. 668 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) { 669 if (SS->getIntrinsicID() == Intrinsic::stacksave) { 670 BasicBlock::iterator BI = SS; 671 if (&*++BI == II) 672 return EraseInstFromFunction(CI); 673 } 674 } 675 676 // Scan down this block to see if there is another stack restore in the 677 // same block without an intervening call/alloca. 678 BasicBlock::iterator BI = II; 679 TerminatorInst *TI = II->getParent()->getTerminator(); 680 bool CannotRemove = false; 681 for (++BI; &*BI != TI; ++BI) { 682 if (isa<AllocaInst>(BI) || isMalloc(BI)) { 683 CannotRemove = true; 684 break; 685 } 686 if (CallInst *BCI = dyn_cast<CallInst>(BI)) { 687 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) { 688 // If there is a stackrestore below this one, remove this one. 689 if (II->getIntrinsicID() == Intrinsic::stackrestore) 690 return EraseInstFromFunction(CI); 691 // Otherwise, ignore the intrinsic. 692 } else { 693 // If we found a non-intrinsic call, we can't remove the stack 694 // restore. 695 CannotRemove = true; 696 break; 697 } 698 } 699 } 700 701 // If the stack restore is in a return/unwind block and if there are no 702 // allocas or calls between the restore and the return, nuke the restore. 703 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI))) 704 return EraseInstFromFunction(CI); 705 break; 706 } 707 } 708 709 return visitCallSite(II); 710} 711 712// InvokeInst simplification 713// 714Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) { 715 return visitCallSite(&II); 716} 717 718/// isSafeToEliminateVarargsCast - If this cast does not affect the value 719/// passed through the varargs area, we can eliminate the use of the cast. 720static bool isSafeToEliminateVarargsCast(const CallSite CS, 721 const CastInst * const CI, 722 const TargetData * const TD, 723 const int ix) { 724 if (!CI->isLosslessCast()) 725 return false; 726 727 // The size of ByVal arguments is derived from the type, so we 728 // can't change to a type with a different size. If the size were 729 // passed explicitly we could avoid this check. 730 if (!CS.paramHasAttr(ix, Attribute::ByVal)) 731 return true; 732 733 const Type* SrcTy = 734 cast<PointerType>(CI->getOperand(0)->getType())->getElementType(); 735 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType(); 736 if (!SrcTy->isSized() || !DstTy->isSized()) 737 return false; 738 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy)) 739 return false; 740 return true; 741} 742 743// visitCallSite - Improvements for call and invoke instructions. 744// 745Instruction *InstCombiner::visitCallSite(CallSite CS) { 746 bool Changed = false; 747 748 // If the callee is a constexpr cast of a function, attempt to move the cast 749 // to the arguments of the call/invoke. 750 if (transformConstExprCastCall(CS)) return 0; 751 752 Value *Callee = CS.getCalledValue(); 753 754 if (Function *CalleeF = dyn_cast<Function>(Callee)) 755 // If the call and callee calling conventions don't match, this call must 756 // be unreachable, as the call is undefined. 757 if (CalleeF->getCallingConv() != CS.getCallingConv() && 758 // Only do this for calls to a function with a body. A prototype may 759 // not actually end up matching the implementation's calling conv for a 760 // variety of reasons (e.g. it may be written in assembly). 761 !CalleeF->isDeclaration()) { 762 Instruction *OldCall = CS.getInstruction(); 763 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 764 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 765 OldCall); 766 // If OldCall dues not return void then replaceAllUsesWith undef. 767 // This allows ValueHandlers and custom metadata to adjust itself. 768 if (!OldCall->getType()->isVoidTy()) 769 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType())); 770 if (isa<CallInst>(OldCall)) 771 return EraseInstFromFunction(*OldCall); 772 773 // We cannot remove an invoke, because it would change the CFG, just 774 // change the callee to a null pointer. 775 cast<InvokeInst>(OldCall)->setOperand(0, 776 Constant::getNullValue(CalleeF->getType())); 777 return 0; 778 } 779 780 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 781 // This instruction is not reachable, just remove it. We insert a store to 782 // undef so that we know that this code is not reachable, despite the fact 783 // that we can't modify the CFG here. 784 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 785 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 786 CS.getInstruction()); 787 788 // If CS dues not return void then replaceAllUsesWith undef. 789 // This allows ValueHandlers and custom metadata to adjust itself. 790 if (!CS.getInstruction()->getType()->isVoidTy()) 791 CS.getInstruction()-> 792 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType())); 793 794 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) { 795 // Don't break the CFG, insert a dummy cond branch. 796 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(), 797 ConstantInt::getTrue(Callee->getContext()), II); 798 } 799 return EraseInstFromFunction(*CS.getInstruction()); 800 } 801 802 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee)) 803 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0))) 804 if (In->getIntrinsicID() == Intrinsic::init_trampoline) 805 return transformCallThroughTrampoline(CS); 806 807 const PointerType *PTy = cast<PointerType>(Callee->getType()); 808 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 809 if (FTy->isVarArg()) { 810 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1); 811 // See if we can optimize any arguments passed through the varargs area of 812 // the call. 813 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(), 814 E = CS.arg_end(); I != E; ++I, ++ix) { 815 CastInst *CI = dyn_cast<CastInst>(*I); 816 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) { 817 *I = CI->getOperand(0); 818 Changed = true; 819 } 820 } 821 } 822 823 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) { 824 // Inline asm calls cannot throw - mark them 'nounwind'. 825 CS.setDoesNotThrow(); 826 Changed = true; 827 } 828 829 return Changed ? CS.getInstruction() : 0; 830} 831 832// transformConstExprCastCall - If the callee is a constexpr cast of a function, 833// attempt to move the cast to the arguments of the call/invoke. 834// 835bool InstCombiner::transformConstExprCastCall(CallSite CS) { 836 if (!isa<ConstantExpr>(CS.getCalledValue())) return false; 837 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue()); 838 if (CE->getOpcode() != Instruction::BitCast || 839 !isa<Function>(CE->getOperand(0))) 840 return false; 841 Function *Callee = cast<Function>(CE->getOperand(0)); 842 Instruction *Caller = CS.getInstruction(); 843 const AttrListPtr &CallerPAL = CS.getAttributes(); 844 845 // Okay, this is a cast from a function to a different type. Unless doing so 846 // would cause a type conversion of one of our arguments, change this call to 847 // be a direct call with arguments casted to the appropriate types. 848 // 849 const FunctionType *FT = Callee->getFunctionType(); 850 const Type *OldRetTy = Caller->getType(); 851 const Type *NewRetTy = FT->getReturnType(); 852 853 if (NewRetTy->isStructTy()) 854 return false; // TODO: Handle multiple return values. 855 856 // Check to see if we are changing the return type... 857 if (OldRetTy != NewRetTy) { 858 if (Callee->isDeclaration() && 859 // Conversion is ok if changing from one pointer type to another or from 860 // a pointer to an integer of the same size. 861 !((OldRetTy->isPointerTy() || !TD || 862 OldRetTy == TD->getIntPtrType(Caller->getContext())) && 863 (NewRetTy->isPointerTy() || !TD || 864 NewRetTy == TD->getIntPtrType(Caller->getContext())))) 865 return false; // Cannot transform this return value. 866 867 if (!Caller->use_empty() && 868 // void -> non-void is handled specially 869 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy)) 870 return false; // Cannot transform this return value. 871 872 if (!CallerPAL.isEmpty() && !Caller->use_empty()) { 873 Attributes RAttrs = CallerPAL.getRetAttributes(); 874 if (RAttrs & Attribute::typeIncompatible(NewRetTy)) 875 return false; // Attribute not compatible with transformed value. 876 } 877 878 // If the callsite is an invoke instruction, and the return value is used by 879 // a PHI node in a successor, we cannot change the return type of the call 880 // because there is no place to put the cast instruction (without breaking 881 // the critical edge). Bail out in this case. 882 if (!Caller->use_empty()) 883 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) 884 for (Value::use_iterator UI = II->use_begin(), E = II->use_end(); 885 UI != E; ++UI) 886 if (PHINode *PN = dyn_cast<PHINode>(*UI)) 887 if (PN->getParent() == II->getNormalDest() || 888 PN->getParent() == II->getUnwindDest()) 889 return false; 890 } 891 892 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin()); 893 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs); 894 895 CallSite::arg_iterator AI = CS.arg_begin(); 896 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) { 897 const Type *ParamTy = FT->getParamType(i); 898 const Type *ActTy = (*AI)->getType(); 899 900 if (!CastInst::isCastable(ActTy, ParamTy)) 901 return false; // Cannot transform this parameter value. 902 903 if (CallerPAL.getParamAttributes(i + 1) 904 & Attribute::typeIncompatible(ParamTy)) 905 return false; // Attribute not compatible with transformed value. 906 907 // Converting from one pointer type to another or between a pointer and an 908 // integer of the same size is safe even if we do not have a body. 909 bool isConvertible = ActTy == ParamTy || 910 (TD && ((ParamTy->isPointerTy() || 911 ParamTy == TD->getIntPtrType(Caller->getContext())) && 912 (ActTy->isPointerTy() || 913 ActTy == TD->getIntPtrType(Caller->getContext())))); 914 if (Callee->isDeclaration() && !isConvertible) return false; 915 } 916 917 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() && 918 Callee->isDeclaration()) 919 return false; // Do not delete arguments unless we have a function body. 920 921 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && 922 !CallerPAL.isEmpty()) 923 // In this case we have more arguments than the new function type, but we 924 // won't be dropping them. Check that these extra arguments have attributes 925 // that are compatible with being a vararg call argument. 926 for (unsigned i = CallerPAL.getNumSlots(); i; --i) { 927 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams()) 928 break; 929 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs; 930 if (PAttrs & Attribute::VarArgsIncompatible) 931 return false; 932 } 933 934 // Okay, we decided that this is a safe thing to do: go ahead and start 935 // inserting cast instructions as necessary... 936 std::vector<Value*> Args; 937 Args.reserve(NumActualArgs); 938 SmallVector<AttributeWithIndex, 8> attrVec; 939 attrVec.reserve(NumCommonArgs); 940 941 // Get any return attributes. 942 Attributes RAttrs = CallerPAL.getRetAttributes(); 943 944 // If the return value is not being used, the type may not be compatible 945 // with the existing attributes. Wipe out any problematic attributes. 946 RAttrs &= ~Attribute::typeIncompatible(NewRetTy); 947 948 // Add the new return attributes. 949 if (RAttrs) 950 attrVec.push_back(AttributeWithIndex::get(0, RAttrs)); 951 952 AI = CS.arg_begin(); 953 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) { 954 const Type *ParamTy = FT->getParamType(i); 955 if ((*AI)->getType() == ParamTy) { 956 Args.push_back(*AI); 957 } else { 958 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, 959 false, ParamTy, false); 960 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp")); 961 } 962 963 // Add any parameter attributes. 964 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 965 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 966 } 967 968 // If the function takes more arguments than the call was taking, add them 969 // now. 970 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) 971 Args.push_back(Constant::getNullValue(FT->getParamType(i))); 972 973 // If we are removing arguments to the function, emit an obnoxious warning. 974 if (FT->getNumParams() < NumActualArgs) { 975 if (!FT->isVarArg()) { 976 errs() << "WARNING: While resolving call to function '" 977 << Callee->getName() << "' arguments were dropped!\n"; 978 } else { 979 // Add all of the arguments in their promoted form to the arg list. 980 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) { 981 const Type *PTy = getPromotedType((*AI)->getType()); 982 if (PTy != (*AI)->getType()) { 983 // Must promote to pass through va_arg area! 984 Instruction::CastOps opcode = 985 CastInst::getCastOpcode(*AI, false, PTy, false); 986 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp")); 987 } else { 988 Args.push_back(*AI); 989 } 990 991 // Add any parameter attributes. 992 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 993 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 994 } 995 } 996 } 997 998 if (Attributes FnAttrs = CallerPAL.getFnAttributes()) 999 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); 1000 1001 if (NewRetTy->isVoidTy()) 1002 Caller->setName(""); // Void type should not have a name. 1003 1004 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(), 1005 attrVec.end()); 1006 1007 Instruction *NC; 1008 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1009 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(), 1010 Args.begin(), Args.end(), 1011 Caller->getName(), Caller); 1012 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv()); 1013 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL); 1014 } else { 1015 NC = CallInst::Create(Callee, Args.begin(), Args.end(), 1016 Caller->getName(), Caller); 1017 CallInst *CI = cast<CallInst>(Caller); 1018 if (CI->isTailCall()) 1019 cast<CallInst>(NC)->setTailCall(); 1020 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv()); 1021 cast<CallInst>(NC)->setAttributes(NewCallerPAL); 1022 } 1023 1024 // Insert a cast of the return type as necessary. 1025 Value *NV = NC; 1026 if (OldRetTy != NV->getType() && !Caller->use_empty()) { 1027 if (!NV->getType()->isVoidTy()) { 1028 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false, 1029 OldRetTy, false); 1030 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp"); 1031 1032 // If this is an invoke instruction, we should insert it after the first 1033 // non-phi, instruction in the normal successor block. 1034 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1035 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI(); 1036 InsertNewInstBefore(NC, *I); 1037 } else { 1038 // Otherwise, it's a call, just insert cast right after the call instr 1039 InsertNewInstBefore(NC, *Caller); 1040 } 1041 Worklist.AddUsersToWorkList(*Caller); 1042 } else { 1043 NV = UndefValue::get(Caller->getType()); 1044 } 1045 } 1046 1047 1048 if (!Caller->use_empty()) 1049 Caller->replaceAllUsesWith(NV); 1050 1051 EraseInstFromFunction(*Caller); 1052 return true; 1053} 1054 1055// transformCallThroughTrampoline - Turn a call to a function created by the 1056// init_trampoline intrinsic into a direct call to the underlying function. 1057// 1058Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) { 1059 Value *Callee = CS.getCalledValue(); 1060 const PointerType *PTy = cast<PointerType>(Callee->getType()); 1061 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 1062 const AttrListPtr &Attrs = CS.getAttributes(); 1063 1064 // If the call already has the 'nest' attribute somewhere then give up - 1065 // otherwise 'nest' would occur twice after splicing in the chain. 1066 if (Attrs.hasAttrSomewhere(Attribute::Nest)) 1067 return 0; 1068 1069 IntrinsicInst *Tramp = 1070 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0)); 1071 1072 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts()); 1073 const PointerType *NestFPTy = cast<PointerType>(NestF->getType()); 1074 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType()); 1075 1076 const AttrListPtr &NestAttrs = NestF->getAttributes(); 1077 if (!NestAttrs.isEmpty()) { 1078 unsigned NestIdx = 1; 1079 const Type *NestTy = 0; 1080 Attributes NestAttr = Attribute::None; 1081 1082 // Look for a parameter marked with the 'nest' attribute. 1083 for (FunctionType::param_iterator I = NestFTy->param_begin(), 1084 E = NestFTy->param_end(); I != E; ++NestIdx, ++I) 1085 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) { 1086 // Record the parameter type and any other attributes. 1087 NestTy = *I; 1088 NestAttr = NestAttrs.getParamAttributes(NestIdx); 1089 break; 1090 } 1091 1092 if (NestTy) { 1093 Instruction *Caller = CS.getInstruction(); 1094 std::vector<Value*> NewArgs; 1095 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1); 1096 1097 SmallVector<AttributeWithIndex, 8> NewAttrs; 1098 NewAttrs.reserve(Attrs.getNumSlots() + 1); 1099 1100 // Insert the nest argument into the call argument list, which may 1101 // mean appending it. Likewise for attributes. 1102 1103 // Add any result attributes. 1104 if (Attributes Attr = Attrs.getRetAttributes()) 1105 NewAttrs.push_back(AttributeWithIndex::get(0, Attr)); 1106 1107 { 1108 unsigned Idx = 1; 1109 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 1110 do { 1111 if (Idx == NestIdx) { 1112 // Add the chain argument and attributes. 1113 Value *NestVal = Tramp->getOperand(3); 1114 if (NestVal->getType() != NestTy) 1115 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller); 1116 NewArgs.push_back(NestVal); 1117 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr)); 1118 } 1119 1120 if (I == E) 1121 break; 1122 1123 // Add the original argument and attributes. 1124 NewArgs.push_back(*I); 1125 if (Attributes Attr = Attrs.getParamAttributes(Idx)) 1126 NewAttrs.push_back 1127 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr)); 1128 1129 ++Idx, ++I; 1130 } while (1); 1131 } 1132 1133 // Add any function attributes. 1134 if (Attributes Attr = Attrs.getFnAttributes()) 1135 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr)); 1136 1137 // The trampoline may have been bitcast to a bogus type (FTy). 1138 // Handle this by synthesizing a new function type, equal to FTy 1139 // with the chain parameter inserted. 1140 1141 std::vector<const Type*> NewTypes; 1142 NewTypes.reserve(FTy->getNumParams()+1); 1143 1144 // Insert the chain's type into the list of parameter types, which may 1145 // mean appending it. 1146 { 1147 unsigned Idx = 1; 1148 FunctionType::param_iterator I = FTy->param_begin(), 1149 E = FTy->param_end(); 1150 1151 do { 1152 if (Idx == NestIdx) 1153 // Add the chain's type. 1154 NewTypes.push_back(NestTy); 1155 1156 if (I == E) 1157 break; 1158 1159 // Add the original type. 1160 NewTypes.push_back(*I); 1161 1162 ++Idx, ++I; 1163 } while (1); 1164 } 1165 1166 // Replace the trampoline call with a direct call. Let the generic 1167 // code sort out any function type mismatches. 1168 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes, 1169 FTy->isVarArg()); 1170 Constant *NewCallee = 1171 NestF->getType() == PointerType::getUnqual(NewFTy) ? 1172 NestF : ConstantExpr::getBitCast(NestF, 1173 PointerType::getUnqual(NewFTy)); 1174 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(), 1175 NewAttrs.end()); 1176 1177 Instruction *NewCaller; 1178 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1179 NewCaller = InvokeInst::Create(NewCallee, 1180 II->getNormalDest(), II->getUnwindDest(), 1181 NewArgs.begin(), NewArgs.end(), 1182 Caller->getName(), Caller); 1183 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv()); 1184 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL); 1185 } else { 1186 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(), 1187 Caller->getName(), Caller); 1188 if (cast<CallInst>(Caller)->isTailCall()) 1189 cast<CallInst>(NewCaller)->setTailCall(); 1190 cast<CallInst>(NewCaller)-> 1191 setCallingConv(cast<CallInst>(Caller)->getCallingConv()); 1192 cast<CallInst>(NewCaller)->setAttributes(NewPAL); 1193 } 1194 if (!Caller->getType()->isVoidTy()) 1195 Caller->replaceAllUsesWith(NewCaller); 1196 Caller->eraseFromParent(); 1197 Worklist.Remove(Caller); 1198 return 0; 1199 } 1200 } 1201 1202 // Replace the trampoline call with a direct call. Since there is no 'nest' 1203 // parameter, there is no need to adjust the argument list. Let the generic 1204 // code sort out any function type mismatches. 1205 Constant *NewCallee = 1206 NestF->getType() == PTy ? NestF : 1207 ConstantExpr::getBitCast(NestF, PTy); 1208 CS.setCalledFunction(NewCallee); 1209 return CS.getInstruction(); 1210} 1211 1212