InstCombineCalls.cpp revision 6b96fe7e146f7eb594e67210c6bef511ad0a2058
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" 20#include "llvm/Transforms/Utils/Local.h" 21using namespace llvm; 22 23/// getPromotedType - Return the specified type promoted as it would be to pass 24/// though a va_arg area. 25static const Type *getPromotedType(const Type *Ty) { 26 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) { 27 if (ITy->getBitWidth() < 32) 28 return Type::getInt32Ty(Ty->getContext()); 29 } 30 return Ty; 31} 32 33 34Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) { 35 unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), TD); 36 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), TD); 37 unsigned MinAlign = std::min(DstAlign, SrcAlign); 38 unsigned CopyAlign = MI->getAlignment(); 39 40 if (CopyAlign < MinAlign) { 41 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), 42 MinAlign, false)); 43 return MI; 44 } 45 46 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with 47 // load/store. 48 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2)); 49 if (MemOpLength == 0) return 0; 50 51 // Source and destination pointer types are always "i8*" for intrinsic. See 52 // if the size is something we can handle with a single primitive load/store. 53 // A single load+store correctly handles overlapping memory in the memmove 54 // case. 55 unsigned Size = MemOpLength->getZExtValue(); 56 if (Size == 0) return MI; // Delete this mem transfer. 57 58 if (Size > 8 || (Size&(Size-1))) 59 return 0; // If not 1/2/4/8 bytes, exit. 60 61 // Use an integer load+store unless we can find something better. 62 unsigned SrcAddrSp = 63 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace(); 64 unsigned DstAddrSp = 65 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace(); 66 67 const IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3); 68 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp); 69 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp); 70 71 // Memcpy forces the use of i8* for the source and destination. That means 72 // that if you're using memcpy to move one double around, you'll get a cast 73 // from double* to i8*. We'd much rather use a double load+store rather than 74 // an i64 load+store, here because this improves the odds that the source or 75 // dest address will be promotable. See if we can find a better type than the 76 // integer datatype. 77 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts(); 78 if (StrippedDest != MI->getArgOperand(0)) { 79 const Type *SrcETy = cast<PointerType>(StrippedDest->getType()) 80 ->getElementType(); 81 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) { 82 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip 83 // down through these levels if so. 84 while (!SrcETy->isSingleValueType()) { 85 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) { 86 if (STy->getNumElements() == 1) 87 SrcETy = STy->getElementType(0); 88 else 89 break; 90 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) { 91 if (ATy->getNumElements() == 1) 92 SrcETy = ATy->getElementType(); 93 else 94 break; 95 } else 96 break; 97 } 98 99 if (SrcETy->isSingleValueType()) { 100 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp); 101 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp); 102 } 103 } 104 } 105 106 107 // If the memcpy/memmove provides better alignment info than we can 108 // infer, use it. 109 SrcAlign = std::max(SrcAlign, CopyAlign); 110 DstAlign = std::max(DstAlign, CopyAlign); 111 112 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy); 113 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy); 114 Instruction *L = new LoadInst(Src, "tmp", MI->isVolatile(), SrcAlign); 115 InsertNewInstBefore(L, *MI); 116 InsertNewInstBefore(new StoreInst(L, Dest, MI->isVolatile(), DstAlign), 117 *MI); 118 119 // Set the size of the copy to 0, it will be deleted on the next iteration. 120 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType())); 121 return MI; 122} 123 124Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) { 125 unsigned Alignment = getKnownAlignment(MI->getDest(), TD); 126 if (MI->getAlignment() < Alignment) { 127 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), 128 Alignment, false)); 129 return MI; 130 } 131 132 // Extract the length and alignment and fill if they are constant. 133 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength()); 134 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue()); 135 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8)) 136 return 0; 137 uint64_t Len = LenC->getZExtValue(); 138 Alignment = MI->getAlignment(); 139 140 // If the length is zero, this is a no-op 141 if (Len == 0) return MI; // memset(d,c,0,a) -> noop 142 143 // memset(s,c,n) -> store s, c (for n=1,2,4,8) 144 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) { 145 const Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8. 146 147 Value *Dest = MI->getDest(); 148 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace(); 149 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp); 150 Dest = Builder->CreateBitCast(Dest, NewDstPtrTy); 151 152 // Alignment 0 is identity for alignment 1 for memset, but not store. 153 if (Alignment == 0) Alignment = 1; 154 155 // Extract the fill value and store. 156 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL; 157 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), 158 Dest, false, Alignment), *MI); 159 160 // Set the size of the copy to 0, it will be deleted on the next iteration. 161 MI->setLength(Constant::getNullValue(LenC->getType())); 162 return MI; 163 } 164 165 return 0; 166} 167 168/// visitCallInst - CallInst simplification. This mostly only handles folding 169/// of intrinsic instructions. For normal calls, it allows visitCallSite to do 170/// the heavy lifting. 171/// 172Instruction *InstCombiner::visitCallInst(CallInst &CI) { 173 if (isFreeCall(&CI)) 174 return visitFree(CI); 175 if (isMalloc(&CI)) 176 return visitMalloc(CI); 177 178 // If the caller function is nounwind, mark the call as nounwind, even if the 179 // callee isn't. 180 if (CI.getParent()->getParent()->doesNotThrow() && 181 !CI.doesNotThrow()) { 182 CI.setDoesNotThrow(); 183 return &CI; 184 } 185 186 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI); 187 if (!II) return visitCallSite(&CI); 188 189 // Intrinsics cannot occur in an invoke, so handle them here instead of in 190 // visitCallSite. 191 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) { 192 bool Changed = false; 193 194 // memmove/cpy/set of zero bytes is a noop. 195 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) { 196 if (NumBytes->isNullValue()) 197 return EraseInstFromFunction(CI); 198 199 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes)) 200 if (CI->getZExtValue() == 1) { 201 // Replace the instruction with just byte operations. We would 202 // transform other cases to loads/stores, but we don't know if 203 // alignment is sufficient. 204 } 205 } 206 207 // No other transformations apply to volatile transfers. 208 if (MI->isVolatile()) 209 return 0; 210 211 // If we have a memmove and the source operation is a constant global, 212 // then the source and dest pointers can't alias, so we can change this 213 // into a call to memcpy. 214 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) { 215 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource())) 216 if (GVSrc->isConstant()) { 217 Module *M = CI.getParent()->getParent()->getParent(); 218 Intrinsic::ID MemCpyID = Intrinsic::memcpy; 219 const Type *Tys[3] = { CI.getArgOperand(0)->getType(), 220 CI.getArgOperand(1)->getType(), 221 CI.getArgOperand(2)->getType() }; 222 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys, 3)); 223 Changed = true; 224 } 225 } 226 227 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { 228 // memmove(x,x,size) -> noop. 229 if (MTI->getSource() == MTI->getDest()) 230 return EraseInstFromFunction(CI); 231 } 232 233 // If we can determine a pointer alignment that is bigger than currently 234 // set, update the alignment. 235 if (isa<MemTransferInst>(MI)) { 236 if (Instruction *I = SimplifyMemTransfer(MI)) 237 return I; 238 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) { 239 if (Instruction *I = SimplifyMemSet(MSI)) 240 return I; 241 } 242 243 if (Changed) return II; 244 } 245 246 switch (II->getIntrinsicID()) { 247 default: break; 248 case Intrinsic::objectsize: { 249 // We need target data for just about everything so depend on it. 250 if (!TD) break; 251 252 const Type *ReturnTy = CI.getType(); 253 uint64_t DontKnow = II->getArgOperand(1) == Builder->getTrue() ? 0 : -1ULL; 254 255 // Get to the real allocated thing and offset as fast as possible. 256 Value *Op1 = II->getArgOperand(0)->stripPointerCasts(); 257 258 uint64_t Offset = 0; 259 uint64_t Size = -1ULL; 260 261 // Try to look through constant GEPs. 262 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1)) { 263 if (!GEP->hasAllConstantIndices()) break; 264 265 // Get the current byte offset into the thing. Use the original 266 // operand in case we're looking through a bitcast. 267 SmallVector<Value*, 8> Ops(GEP->idx_begin(), GEP->idx_end()); 268 Offset = TD->getIndexedOffset(GEP->getPointerOperandType(), 269 Ops.data(), Ops.size()); 270 271 Op1 = GEP->getPointerOperand()->stripPointerCasts(); 272 273 // Make sure we're not a constant offset from an external 274 // global. 275 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) 276 if (!GV->hasDefinitiveInitializer()) break; 277 } 278 279 // If we've stripped down to a single global variable that we 280 // can know the size of then just return that. 281 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) { 282 if (GV->hasDefinitiveInitializer()) { 283 Constant *C = GV->getInitializer(); 284 Size = TD->getTypeAllocSize(C->getType()); 285 } else { 286 // Can't determine size of the GV. 287 Constant *RetVal = ConstantInt::get(ReturnTy, DontKnow); 288 return ReplaceInstUsesWith(CI, RetVal); 289 } 290 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) { 291 // Get alloca size. 292 if (AI->getAllocatedType()->isSized()) { 293 Size = TD->getTypeAllocSize(AI->getAllocatedType()); 294 if (AI->isArrayAllocation()) { 295 const ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize()); 296 if (!C) break; 297 Size *= C->getZExtValue(); 298 } 299 } 300 } else if (CallInst *MI = extractMallocCall(Op1)) { 301 // Get allocation size. 302 const Type* MallocType = getMallocAllocatedType(MI); 303 if (MallocType && MallocType->isSized()) 304 if (Value *NElems = getMallocArraySize(MI, TD, true)) 305 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems)) 306 Size = NElements->getZExtValue() * TD->getTypeAllocSize(MallocType); 307 } 308 309 // Do not return "I don't know" here. Later optimization passes could 310 // make it possible to evaluate objectsize to a constant. 311 if (Size == -1ULL) 312 break; 313 314 if (Size < Offset) { 315 // Out of bound reference? Negative index normalized to large 316 // index? Just return "I don't know". 317 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, DontKnow)); 318 } 319 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, Size-Offset)); 320 } 321 case Intrinsic::bswap: 322 // bswap(bswap(x)) -> x 323 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) 324 if (Operand->getIntrinsicID() == Intrinsic::bswap) 325 return ReplaceInstUsesWith(CI, Operand->getArgOperand(0)); 326 327 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c)) 328 if (TruncInst *TI = dyn_cast<TruncInst>(II->getArgOperand(0))) { 329 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0))) 330 if (Operand->getIntrinsicID() == Intrinsic::bswap) { 331 unsigned C = Operand->getType()->getPrimitiveSizeInBits() - 332 TI->getType()->getPrimitiveSizeInBits(); 333 Value *CV = ConstantInt::get(Operand->getType(), C); 334 Value *V = Builder->CreateLShr(Operand->getArgOperand(0), CV); 335 return new TruncInst(V, TI->getType()); 336 } 337 } 338 339 break; 340 case Intrinsic::powi: 341 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 342 // powi(x, 0) -> 1.0 343 if (Power->isZero()) 344 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0)); 345 // powi(x, 1) -> x 346 if (Power->isOne()) 347 return ReplaceInstUsesWith(CI, II->getArgOperand(0)); 348 // powi(x, -1) -> 1/x 349 if (Power->isAllOnesValue()) 350 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0), 351 II->getArgOperand(0)); 352 } 353 break; 354 case Intrinsic::cttz: { 355 // If all bits below the first known one are known zero, 356 // this value is constant. 357 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType()); 358 uint32_t BitWidth = IT->getBitWidth(); 359 APInt KnownZero(BitWidth, 0); 360 APInt KnownOne(BitWidth, 0); 361 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth), 362 KnownZero, KnownOne); 363 unsigned TrailingZeros = KnownOne.countTrailingZeros(); 364 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros)); 365 if ((Mask & KnownZero) == Mask) 366 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 367 APInt(BitWidth, TrailingZeros))); 368 369 } 370 break; 371 case Intrinsic::ctlz: { 372 // If all bits above the first known one are known zero, 373 // this value is constant. 374 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType()); 375 uint32_t BitWidth = IT->getBitWidth(); 376 APInt KnownZero(BitWidth, 0); 377 APInt KnownOne(BitWidth, 0); 378 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth), 379 KnownZero, KnownOne); 380 unsigned LeadingZeros = KnownOne.countLeadingZeros(); 381 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros)); 382 if ((Mask & KnownZero) == Mask) 383 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 384 APInt(BitWidth, LeadingZeros))); 385 386 } 387 break; 388 case Intrinsic::uadd_with_overflow: { 389 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 390 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType()); 391 uint32_t BitWidth = IT->getBitWidth(); 392 APInt Mask = APInt::getSignBit(BitWidth); 393 APInt LHSKnownZero(BitWidth, 0); 394 APInt LHSKnownOne(BitWidth, 0); 395 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne); 396 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1]; 397 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1]; 398 399 if (LHSKnownNegative || LHSKnownPositive) { 400 APInt RHSKnownZero(BitWidth, 0); 401 APInt RHSKnownOne(BitWidth, 0); 402 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne); 403 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1]; 404 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1]; 405 if (LHSKnownNegative && RHSKnownNegative) { 406 // The sign bit is set in both cases: this MUST overflow. 407 // Create a simple add instruction, and insert it into the struct. 408 Instruction *Add = BinaryOperator::CreateAdd(LHS, RHS, "", &CI); 409 Worklist.Add(Add); 410 Constant *V[] = { 411 UndefValue::get(LHS->getType()),ConstantInt::getTrue(II->getContext()) 412 }; 413 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 414 return InsertValueInst::Create(Struct, Add, 0); 415 } 416 417 if (LHSKnownPositive && RHSKnownPositive) { 418 // The sign bit is clear in both cases: this CANNOT overflow. 419 // Create a simple add instruction, and insert it into the struct. 420 Instruction *Add = BinaryOperator::CreateNUWAdd(LHS, RHS, "", &CI); 421 Worklist.Add(Add); 422 Constant *V[] = { 423 UndefValue::get(LHS->getType()), 424 ConstantInt::getFalse(II->getContext()) 425 }; 426 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 427 return InsertValueInst::Create(Struct, Add, 0); 428 } 429 } 430 } 431 // FALL THROUGH uadd into sadd 432 case Intrinsic::sadd_with_overflow: 433 // Canonicalize constants into the RHS. 434 if (isa<Constant>(II->getArgOperand(0)) && 435 !isa<Constant>(II->getArgOperand(1))) { 436 Value *LHS = II->getArgOperand(0); 437 II->setArgOperand(0, II->getArgOperand(1)); 438 II->setArgOperand(1, LHS); 439 return II; 440 } 441 442 // X + undef -> undef 443 if (isa<UndefValue>(II->getArgOperand(1))) 444 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 445 446 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 447 // X + 0 -> {X, false} 448 if (RHS->isZero()) { 449 Constant *V[] = { 450 UndefValue::get(II->getArgOperand(0)->getType()), 451 ConstantInt::getFalse(II->getContext()) 452 }; 453 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 454 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 455 } 456 } 457 break; 458 case Intrinsic::usub_with_overflow: 459 case Intrinsic::ssub_with_overflow: 460 // undef - X -> undef 461 // X - undef -> undef 462 if (isa<UndefValue>(II->getArgOperand(0)) || 463 isa<UndefValue>(II->getArgOperand(1))) 464 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 465 466 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 467 // X - 0 -> {X, false} 468 if (RHS->isZero()) { 469 Constant *V[] = { 470 UndefValue::get(II->getArgOperand(0)->getType()), 471 ConstantInt::getFalse(II->getContext()) 472 }; 473 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 474 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 475 } 476 } 477 break; 478 case Intrinsic::umul_with_overflow: { 479 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 480 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth(); 481 APInt Mask = APInt::getAllOnesValue(BitWidth); 482 483 APInt LHSKnownZero(BitWidth, 0); 484 APInt LHSKnownOne(BitWidth, 0); 485 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne); 486 APInt RHSKnownZero(BitWidth, 0); 487 APInt RHSKnownOne(BitWidth, 0); 488 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne); 489 490 // Get the largest possible values for each operand, extended to be large 491 // enough so that every possible product of two BitWidth-sized ints fits. 492 APInt LHSMax = (~LHSKnownZero).zext(BitWidth*2); 493 APInt RHSMax = (~RHSKnownZero).zext(BitWidth*2); 494 495 // If multiplying the maximum values does not overflow then we can turn 496 // this into a plain NUW mul. 497 if ((LHSMax * RHSMax).getActiveBits() <= BitWidth) { 498 Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow"); 499 Constant *V[] = { 500 UndefValue::get(LHS->getType()), 501 Builder->getFalse() 502 }; 503 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 504 return InsertValueInst::Create(Struct, Mul, 0); 505 } 506 } // FALL THROUGH 507 case Intrinsic::smul_with_overflow: 508 // Canonicalize constants into the RHS. 509 if (isa<Constant>(II->getArgOperand(0)) && 510 !isa<Constant>(II->getArgOperand(1))) { 511 Value *LHS = II->getArgOperand(0); 512 II->setArgOperand(0, II->getArgOperand(1)); 513 II->setArgOperand(1, LHS); 514 return II; 515 } 516 517 // X * undef -> undef 518 if (isa<UndefValue>(II->getArgOperand(1))) 519 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 520 521 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 522 // X*0 -> {0, false} 523 if (RHSI->isZero()) 524 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType())); 525 526 // X * 1 -> {X, false} 527 if (RHSI->equalsInt(1)) { 528 Constant *V[] = { 529 UndefValue::get(II->getArgOperand(0)->getType()), 530 ConstantInt::getFalse(II->getContext()) 531 }; 532 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 533 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 534 } 535 } 536 break; 537 case Intrinsic::ppc_altivec_lvx: 538 case Intrinsic::ppc_altivec_lvxl: 539 case Intrinsic::x86_sse_loadu_ps: 540 case Intrinsic::x86_sse2_loadu_pd: 541 case Intrinsic::x86_sse2_loadu_dq: 542 // Turn PPC lvx -> load if the pointer is known aligned. 543 // Turn X86 loadups -> load if the pointer is known aligned. 544 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 545 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), 546 PointerType::getUnqual(II->getType())); 547 return new LoadInst(Ptr); 548 } 549 break; 550 case Intrinsic::ppc_altivec_stvx: 551 case Intrinsic::ppc_altivec_stvxl: 552 // Turn stvx -> store if the pointer is known aligned. 553 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) { 554 const Type *OpPtrTy = 555 PointerType::getUnqual(II->getArgOperand(0)->getType()); 556 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); 557 return new StoreInst(II->getArgOperand(0), Ptr); 558 } 559 break; 560 case Intrinsic::x86_sse_storeu_ps: 561 case Intrinsic::x86_sse2_storeu_pd: 562 case Intrinsic::x86_sse2_storeu_dq: 563 // Turn X86 storeu -> store if the pointer is known aligned. 564 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 565 const Type *OpPtrTy = 566 PointerType::getUnqual(II->getArgOperand(1)->getType()); 567 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy); 568 return new StoreInst(II->getArgOperand(1), Ptr); 569 } 570 break; 571 572 case Intrinsic::x86_sse_cvtss2si: 573 case Intrinsic::x86_sse_cvtss2si64: 574 case Intrinsic::x86_sse_cvttss2si: 575 case Intrinsic::x86_sse_cvttss2si64: 576 case Intrinsic::x86_sse2_cvtsd2si: 577 case Intrinsic::x86_sse2_cvtsd2si64: 578 case Intrinsic::x86_sse2_cvttsd2si: 579 case Intrinsic::x86_sse2_cvttsd2si64: { 580 // These intrinsics only demand the 0th element of their input vectors. If 581 // we can simplify the input based on that, do so now. 582 unsigned VWidth = 583 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 584 APInt DemandedElts(VWidth, 1); 585 APInt UndefElts(VWidth, 0); 586 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0), 587 DemandedElts, UndefElts)) { 588 II->setArgOperand(0, V); 589 return II; 590 } 591 break; 592 } 593 594 case Intrinsic::ppc_altivec_vperm: 595 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant. 596 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getArgOperand(2))) { 597 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!"); 598 599 // Check that all of the elements are integer constants or undefs. 600 bool AllEltsOk = true; 601 for (unsigned i = 0; i != 16; ++i) { 602 if (!isa<ConstantInt>(Mask->getOperand(i)) && 603 !isa<UndefValue>(Mask->getOperand(i))) { 604 AllEltsOk = false; 605 break; 606 } 607 } 608 609 if (AllEltsOk) { 610 // Cast the input vectors to byte vectors. 611 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0), 612 Mask->getType()); 613 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1), 614 Mask->getType()); 615 Value *Result = UndefValue::get(Op0->getType()); 616 617 // Only extract each element once. 618 Value *ExtractedElts[32]; 619 memset(ExtractedElts, 0, sizeof(ExtractedElts)); 620 621 for (unsigned i = 0; i != 16; ++i) { 622 if (isa<UndefValue>(Mask->getOperand(i))) 623 continue; 624 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue(); 625 Idx &= 31; // Match the hardware behavior. 626 627 if (ExtractedElts[Idx] == 0) { 628 ExtractedElts[Idx] = 629 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1, 630 ConstantInt::get(Type::getInt32Ty(II->getContext()), 631 Idx&15, false), "tmp"); 632 } 633 634 // Insert this value into the result vector. 635 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx], 636 ConstantInt::get(Type::getInt32Ty(II->getContext()), 637 i, false), "tmp"); 638 } 639 return CastInst::Create(Instruction::BitCast, Result, CI.getType()); 640 } 641 } 642 break; 643 644 case Intrinsic::arm_neon_vld1: 645 case Intrinsic::arm_neon_vld2: 646 case Intrinsic::arm_neon_vld3: 647 case Intrinsic::arm_neon_vld4: 648 case Intrinsic::arm_neon_vld2lane: 649 case Intrinsic::arm_neon_vld3lane: 650 case Intrinsic::arm_neon_vld4lane: 651 case Intrinsic::arm_neon_vst1: 652 case Intrinsic::arm_neon_vst2: 653 case Intrinsic::arm_neon_vst3: 654 case Intrinsic::arm_neon_vst4: 655 case Intrinsic::arm_neon_vst2lane: 656 case Intrinsic::arm_neon_vst3lane: 657 case Intrinsic::arm_neon_vst4lane: { 658 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD); 659 unsigned AlignArg = II->getNumArgOperands() - 1; 660 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg)); 661 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) { 662 II->setArgOperand(AlignArg, 663 ConstantInt::get(Type::getInt32Ty(II->getContext()), 664 MemAlign, false)); 665 return II; 666 } 667 break; 668 } 669 670 case Intrinsic::stackrestore: { 671 // If the save is right next to the restore, remove the restore. This can 672 // happen when variable allocas are DCE'd. 673 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) { 674 if (SS->getIntrinsicID() == Intrinsic::stacksave) { 675 BasicBlock::iterator BI = SS; 676 if (&*++BI == II) 677 return EraseInstFromFunction(CI); 678 } 679 } 680 681 // Scan down this block to see if there is another stack restore in the 682 // same block without an intervening call/alloca. 683 BasicBlock::iterator BI = II; 684 TerminatorInst *TI = II->getParent()->getTerminator(); 685 bool CannotRemove = false; 686 for (++BI; &*BI != TI; ++BI) { 687 if (isa<AllocaInst>(BI) || isMalloc(BI)) { 688 CannotRemove = true; 689 break; 690 } 691 if (CallInst *BCI = dyn_cast<CallInst>(BI)) { 692 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) { 693 // If there is a stackrestore below this one, remove this one. 694 if (II->getIntrinsicID() == Intrinsic::stackrestore) 695 return EraseInstFromFunction(CI); 696 // Otherwise, ignore the intrinsic. 697 } else { 698 // If we found a non-intrinsic call, we can't remove the stack 699 // restore. 700 CannotRemove = true; 701 break; 702 } 703 } 704 } 705 706 // If the stack restore is in a return/unwind block and if there are no 707 // allocas or calls between the restore and the return, nuke the restore. 708 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI))) 709 return EraseInstFromFunction(CI); 710 break; 711 } 712 } 713 714 return visitCallSite(II); 715} 716 717// InvokeInst simplification 718// 719Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) { 720 return visitCallSite(&II); 721} 722 723/// isSafeToEliminateVarargsCast - If this cast does not affect the value 724/// passed through the varargs area, we can eliminate the use of the cast. 725static bool isSafeToEliminateVarargsCast(const CallSite CS, 726 const CastInst * const CI, 727 const TargetData * const TD, 728 const int ix) { 729 if (!CI->isLosslessCast()) 730 return false; 731 732 // The size of ByVal arguments is derived from the type, so we 733 // can't change to a type with a different size. If the size were 734 // passed explicitly we could avoid this check. 735 if (!CS.paramHasAttr(ix, Attribute::ByVal)) 736 return true; 737 738 const Type* SrcTy = 739 cast<PointerType>(CI->getOperand(0)->getType())->getElementType(); 740 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType(); 741 if (!SrcTy->isSized() || !DstTy->isSized()) 742 return false; 743 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy)) 744 return false; 745 return true; 746} 747 748namespace { 749class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls { 750 InstCombiner *IC; 751protected: 752 void replaceCall(Value *With) { 753 NewInstruction = IC->ReplaceInstUsesWith(*CI, With); 754 } 755 bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const { 756 if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp)) 757 return true; 758 if (ConstantInt *SizeCI = 759 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) { 760 if (SizeCI->isAllOnesValue()) 761 return true; 762 if (isString) 763 return SizeCI->getZExtValue() >= 764 GetStringLength(CI->getArgOperand(SizeArgOp)); 765 if (ConstantInt *Arg = dyn_cast<ConstantInt>( 766 CI->getArgOperand(SizeArgOp))) 767 return SizeCI->getZExtValue() >= Arg->getZExtValue(); 768 } 769 return false; 770 } 771public: 772 InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { } 773 Instruction *NewInstruction; 774}; 775} // end anonymous namespace 776 777// Try to fold some different type of calls here. 778// Currently we're only working with the checking functions, memcpy_chk, 779// mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk, 780// strcat_chk and strncat_chk. 781Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) { 782 if (CI->getCalledFunction() == 0) return 0; 783 784 InstCombineFortifiedLibCalls Simplifier(this); 785 Simplifier.fold(CI, TD); 786 return Simplifier.NewInstruction; 787} 788 789// visitCallSite - Improvements for call and invoke instructions. 790// 791Instruction *InstCombiner::visitCallSite(CallSite CS) { 792 bool Changed = false; 793 794 // If the callee is a pointer to a function, attempt to move any casts to the 795 // arguments of the call/invoke. 796 Value *Callee = CS.getCalledValue(); 797 if (!isa<Function>(Callee) && transformConstExprCastCall(CS)) 798 return 0; 799 800 if (Function *CalleeF = dyn_cast<Function>(Callee)) 801 // If the call and callee calling conventions don't match, this call must 802 // be unreachable, as the call is undefined. 803 if (CalleeF->getCallingConv() != CS.getCallingConv() && 804 // Only do this for calls to a function with a body. A prototype may 805 // not actually end up matching the implementation's calling conv for a 806 // variety of reasons (e.g. it may be written in assembly). 807 !CalleeF->isDeclaration()) { 808 Instruction *OldCall = CS.getInstruction(); 809 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 810 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 811 OldCall); 812 // If OldCall dues not return void then replaceAllUsesWith undef. 813 // This allows ValueHandlers and custom metadata to adjust itself. 814 if (!OldCall->getType()->isVoidTy()) 815 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType())); 816 if (isa<CallInst>(OldCall)) 817 return EraseInstFromFunction(*OldCall); 818 819 // We cannot remove an invoke, because it would change the CFG, just 820 // change the callee to a null pointer. 821 cast<InvokeInst>(OldCall)->setCalledFunction( 822 Constant::getNullValue(CalleeF->getType())); 823 return 0; 824 } 825 826 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 827 // This instruction is not reachable, just remove it. We insert a store to 828 // undef so that we know that this code is not reachable, despite the fact 829 // that we can't modify the CFG here. 830 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 831 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 832 CS.getInstruction()); 833 834 // If CS does not return void then replaceAllUsesWith undef. 835 // This allows ValueHandlers and custom metadata to adjust itself. 836 if (!CS.getInstruction()->getType()->isVoidTy()) 837 CS.getInstruction()-> 838 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType())); 839 840 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) { 841 // Don't break the CFG, insert a dummy cond branch. 842 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(), 843 ConstantInt::getTrue(Callee->getContext()), II); 844 } 845 return EraseInstFromFunction(*CS.getInstruction()); 846 } 847 848 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee)) 849 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0))) 850 if (In->getIntrinsicID() == Intrinsic::init_trampoline) 851 return transformCallThroughTrampoline(CS); 852 853 const PointerType *PTy = cast<PointerType>(Callee->getType()); 854 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 855 if (FTy->isVarArg()) { 856 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1); 857 // See if we can optimize any arguments passed through the varargs area of 858 // the call. 859 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(), 860 E = CS.arg_end(); I != E; ++I, ++ix) { 861 CastInst *CI = dyn_cast<CastInst>(*I); 862 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) { 863 *I = CI->getOperand(0); 864 Changed = true; 865 } 866 } 867 } 868 869 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) { 870 // Inline asm calls cannot throw - mark them 'nounwind'. 871 CS.setDoesNotThrow(); 872 Changed = true; 873 } 874 875 // Try to optimize the call if possible, we require TargetData for most of 876 // this. None of these calls are seen as possibly dead so go ahead and 877 // delete the instruction now. 878 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) { 879 Instruction *I = tryOptimizeCall(CI, TD); 880 // If we changed something return the result, etc. Otherwise let 881 // the fallthrough check. 882 if (I) return EraseInstFromFunction(*I); 883 } 884 885 return Changed ? CS.getInstruction() : 0; 886} 887 888// transformConstExprCastCall - If the callee is a constexpr cast of a function, 889// attempt to move the cast to the arguments of the call/invoke. 890// 891bool InstCombiner::transformConstExprCastCall(CallSite CS) { 892 Function *Callee = 893 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts()); 894 if (Callee == 0) 895 return false; 896 Instruction *Caller = CS.getInstruction(); 897 const AttrListPtr &CallerPAL = CS.getAttributes(); 898 899 // Okay, this is a cast from a function to a different type. Unless doing so 900 // would cause a type conversion of one of our arguments, change this call to 901 // be a direct call with arguments casted to the appropriate types. 902 // 903 const FunctionType *FT = Callee->getFunctionType(); 904 const Type *OldRetTy = Caller->getType(); 905 const Type *NewRetTy = FT->getReturnType(); 906 907 if (NewRetTy->isStructTy()) 908 return false; // TODO: Handle multiple return values. 909 910 // Check to see if we are changing the return type... 911 if (OldRetTy != NewRetTy) { 912 if (Callee->isDeclaration() && 913 // Conversion is ok if changing from one pointer type to another or from 914 // a pointer to an integer of the same size. 915 !((OldRetTy->isPointerTy() || !TD || 916 OldRetTy == TD->getIntPtrType(Caller->getContext())) && 917 (NewRetTy->isPointerTy() || !TD || 918 NewRetTy == TD->getIntPtrType(Caller->getContext())))) 919 return false; // Cannot transform this return value. 920 921 if (!Caller->use_empty() && 922 // void -> non-void is handled specially 923 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy)) 924 return false; // Cannot transform this return value. 925 926 if (!CallerPAL.isEmpty() && !Caller->use_empty()) { 927 Attributes RAttrs = CallerPAL.getRetAttributes(); 928 if (RAttrs & Attribute::typeIncompatible(NewRetTy)) 929 return false; // Attribute not compatible with transformed value. 930 } 931 932 // If the callsite is an invoke instruction, and the return value is used by 933 // a PHI node in a successor, we cannot change the return type of the call 934 // because there is no place to put the cast instruction (without breaking 935 // the critical edge). Bail out in this case. 936 if (!Caller->use_empty()) 937 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) 938 for (Value::use_iterator UI = II->use_begin(), E = II->use_end(); 939 UI != E; ++UI) 940 if (PHINode *PN = dyn_cast<PHINode>(*UI)) 941 if (PN->getParent() == II->getNormalDest() || 942 PN->getParent() == II->getUnwindDest()) 943 return false; 944 } 945 946 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin()); 947 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs); 948 949 CallSite::arg_iterator AI = CS.arg_begin(); 950 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) { 951 const Type *ParamTy = FT->getParamType(i); 952 const Type *ActTy = (*AI)->getType(); 953 954 if (!CastInst::isCastable(ActTy, ParamTy)) 955 return false; // Cannot transform this parameter value. 956 957 unsigned Attrs = CallerPAL.getParamAttributes(i + 1); 958 if (Attrs & Attribute::typeIncompatible(ParamTy)) 959 return false; // Attribute not compatible with transformed value. 960 961 // If the parameter is passed as a byval argument, then we have to have a 962 // sized type and the sized type has to have the same size as the old type. 963 if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) { 964 const PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy); 965 if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0) 966 return false; 967 968 const Type *CurElTy = cast<PointerType>(ActTy)->getElementType(); 969 if (TD->getTypeAllocSize(CurElTy) != 970 TD->getTypeAllocSize(ParamPTy->getElementType())) 971 return false; 972 } 973 974 // Converting from one pointer type to another or between a pointer and an 975 // integer of the same size is safe even if we do not have a body. 976 bool isConvertible = ActTy == ParamTy || 977 (TD && ((ParamTy->isPointerTy() || 978 ParamTy == TD->getIntPtrType(Caller->getContext())) && 979 (ActTy->isPointerTy() || 980 ActTy == TD->getIntPtrType(Caller->getContext())))); 981 if (Callee->isDeclaration() && !isConvertible) return false; 982 } 983 984 if (Callee->isDeclaration()) { 985 // Do not delete arguments unless we have a function body. 986 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg()) 987 return false; 988 989 // If the callee is just a declaration, don't change the varargsness of the 990 // call. We don't want to introduce a varargs call where one doesn't 991 // already exist. 992 const PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType()); 993 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg()) 994 return false; 995 } 996 997 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && 998 !CallerPAL.isEmpty()) 999 // In this case we have more arguments than the new function type, but we 1000 // won't be dropping them. Check that these extra arguments have attributes 1001 // that are compatible with being a vararg call argument. 1002 for (unsigned i = CallerPAL.getNumSlots(); i; --i) { 1003 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams()) 1004 break; 1005 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs; 1006 if (PAttrs & Attribute::VarArgsIncompatible) 1007 return false; 1008 } 1009 1010 1011 // Okay, we decided that this is a safe thing to do: go ahead and start 1012 // inserting cast instructions as necessary. 1013 std::vector<Value*> Args; 1014 Args.reserve(NumActualArgs); 1015 SmallVector<AttributeWithIndex, 8> attrVec; 1016 attrVec.reserve(NumCommonArgs); 1017 1018 // Get any return attributes. 1019 Attributes RAttrs = CallerPAL.getRetAttributes(); 1020 1021 // If the return value is not being used, the type may not be compatible 1022 // with the existing attributes. Wipe out any problematic attributes. 1023 RAttrs &= ~Attribute::typeIncompatible(NewRetTy); 1024 1025 // Add the new return attributes. 1026 if (RAttrs) 1027 attrVec.push_back(AttributeWithIndex::get(0, RAttrs)); 1028 1029 AI = CS.arg_begin(); 1030 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) { 1031 const Type *ParamTy = FT->getParamType(i); 1032 if ((*AI)->getType() == ParamTy) { 1033 Args.push_back(*AI); 1034 } else { 1035 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, 1036 false, ParamTy, false); 1037 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp")); 1038 } 1039 1040 // Add any parameter attributes. 1041 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1042 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1043 } 1044 1045 // If the function takes more arguments than the call was taking, add them 1046 // now. 1047 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) 1048 Args.push_back(Constant::getNullValue(FT->getParamType(i))); 1049 1050 // If we are removing arguments to the function, emit an obnoxious warning. 1051 if (FT->getNumParams() < NumActualArgs) { 1052 if (!FT->isVarArg()) { 1053 errs() << "WARNING: While resolving call to function '" 1054 << Callee->getName() << "' arguments were dropped!\n"; 1055 } else { 1056 // Add all of the arguments in their promoted form to the arg list. 1057 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) { 1058 const Type *PTy = getPromotedType((*AI)->getType()); 1059 if (PTy != (*AI)->getType()) { 1060 // Must promote to pass through va_arg area! 1061 Instruction::CastOps opcode = 1062 CastInst::getCastOpcode(*AI, false, PTy, false); 1063 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp")); 1064 } else { 1065 Args.push_back(*AI); 1066 } 1067 1068 // Add any parameter attributes. 1069 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1070 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1071 } 1072 } 1073 } 1074 1075 if (Attributes FnAttrs = CallerPAL.getFnAttributes()) 1076 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); 1077 1078 if (NewRetTy->isVoidTy()) 1079 Caller->setName(""); // Void type should not have a name. 1080 1081 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(), 1082 attrVec.end()); 1083 1084 Instruction *NC; 1085 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1086 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(), 1087 Args.begin(), Args.end(), 1088 Caller->getName(), Caller); 1089 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv()); 1090 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL); 1091 } else { 1092 NC = CallInst::Create(Callee, Args.begin(), Args.end(), 1093 Caller->getName(), Caller); 1094 CallInst *CI = cast<CallInst>(Caller); 1095 if (CI->isTailCall()) 1096 cast<CallInst>(NC)->setTailCall(); 1097 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv()); 1098 cast<CallInst>(NC)->setAttributes(NewCallerPAL); 1099 } 1100 1101 // Insert a cast of the return type as necessary. 1102 Value *NV = NC; 1103 if (OldRetTy != NV->getType() && !Caller->use_empty()) { 1104 if (!NV->getType()->isVoidTy()) { 1105 Instruction::CastOps opcode = 1106 CastInst::getCastOpcode(NC, false, OldRetTy, false); 1107 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp"); 1108 1109 // If this is an invoke instruction, we should insert it after the first 1110 // non-phi, instruction in the normal successor block. 1111 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1112 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI(); 1113 InsertNewInstBefore(NC, *I); 1114 } else { 1115 // Otherwise, it's a call, just insert cast right after the call. 1116 InsertNewInstBefore(NC, *Caller); 1117 } 1118 Worklist.AddUsersToWorkList(*Caller); 1119 } else { 1120 NV = UndefValue::get(Caller->getType()); 1121 } 1122 } 1123 1124 if (!Caller->use_empty()) 1125 Caller->replaceAllUsesWith(NV); 1126 1127 EraseInstFromFunction(*Caller); 1128 return true; 1129} 1130 1131// transformCallThroughTrampoline - Turn a call to a function created by the 1132// init_trampoline intrinsic into a direct call to the underlying function. 1133// 1134Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) { 1135 Value *Callee = CS.getCalledValue(); 1136 const PointerType *PTy = cast<PointerType>(Callee->getType()); 1137 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 1138 const AttrListPtr &Attrs = CS.getAttributes(); 1139 1140 // If the call already has the 'nest' attribute somewhere then give up - 1141 // otherwise 'nest' would occur twice after splicing in the chain. 1142 if (Attrs.hasAttrSomewhere(Attribute::Nest)) 1143 return 0; 1144 1145 IntrinsicInst *Tramp = 1146 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0)); 1147 1148 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts()); 1149 const PointerType *NestFPTy = cast<PointerType>(NestF->getType()); 1150 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType()); 1151 1152 const AttrListPtr &NestAttrs = NestF->getAttributes(); 1153 if (!NestAttrs.isEmpty()) { 1154 unsigned NestIdx = 1; 1155 const Type *NestTy = 0; 1156 Attributes NestAttr = Attribute::None; 1157 1158 // Look for a parameter marked with the 'nest' attribute. 1159 for (FunctionType::param_iterator I = NestFTy->param_begin(), 1160 E = NestFTy->param_end(); I != E; ++NestIdx, ++I) 1161 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) { 1162 // Record the parameter type and any other attributes. 1163 NestTy = *I; 1164 NestAttr = NestAttrs.getParamAttributes(NestIdx); 1165 break; 1166 } 1167 1168 if (NestTy) { 1169 Instruction *Caller = CS.getInstruction(); 1170 std::vector<Value*> NewArgs; 1171 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1); 1172 1173 SmallVector<AttributeWithIndex, 8> NewAttrs; 1174 NewAttrs.reserve(Attrs.getNumSlots() + 1); 1175 1176 // Insert the nest argument into the call argument list, which may 1177 // mean appending it. Likewise for attributes. 1178 1179 // Add any result attributes. 1180 if (Attributes Attr = Attrs.getRetAttributes()) 1181 NewAttrs.push_back(AttributeWithIndex::get(0, Attr)); 1182 1183 { 1184 unsigned Idx = 1; 1185 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 1186 do { 1187 if (Idx == NestIdx) { 1188 // Add the chain argument and attributes. 1189 Value *NestVal = Tramp->getArgOperand(2); 1190 if (NestVal->getType() != NestTy) 1191 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller); 1192 NewArgs.push_back(NestVal); 1193 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr)); 1194 } 1195 1196 if (I == E) 1197 break; 1198 1199 // Add the original argument and attributes. 1200 NewArgs.push_back(*I); 1201 if (Attributes Attr = Attrs.getParamAttributes(Idx)) 1202 NewAttrs.push_back 1203 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr)); 1204 1205 ++Idx, ++I; 1206 } while (1); 1207 } 1208 1209 // Add any function attributes. 1210 if (Attributes Attr = Attrs.getFnAttributes()) 1211 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr)); 1212 1213 // The trampoline may have been bitcast to a bogus type (FTy). 1214 // Handle this by synthesizing a new function type, equal to FTy 1215 // with the chain parameter inserted. 1216 1217 std::vector<const Type*> NewTypes; 1218 NewTypes.reserve(FTy->getNumParams()+1); 1219 1220 // Insert the chain's type into the list of parameter types, which may 1221 // mean appending it. 1222 { 1223 unsigned Idx = 1; 1224 FunctionType::param_iterator I = FTy->param_begin(), 1225 E = FTy->param_end(); 1226 1227 do { 1228 if (Idx == NestIdx) 1229 // Add the chain's type. 1230 NewTypes.push_back(NestTy); 1231 1232 if (I == E) 1233 break; 1234 1235 // Add the original type. 1236 NewTypes.push_back(*I); 1237 1238 ++Idx, ++I; 1239 } while (1); 1240 } 1241 1242 // Replace the trampoline call with a direct call. Let the generic 1243 // code sort out any function type mismatches. 1244 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes, 1245 FTy->isVarArg()); 1246 Constant *NewCallee = 1247 NestF->getType() == PointerType::getUnqual(NewFTy) ? 1248 NestF : ConstantExpr::getBitCast(NestF, 1249 PointerType::getUnqual(NewFTy)); 1250 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(), 1251 NewAttrs.end()); 1252 1253 Instruction *NewCaller; 1254 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1255 NewCaller = InvokeInst::Create(NewCallee, 1256 II->getNormalDest(), II->getUnwindDest(), 1257 NewArgs.begin(), NewArgs.end(), 1258 Caller->getName(), Caller); 1259 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv()); 1260 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL); 1261 } else { 1262 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(), 1263 Caller->getName(), Caller); 1264 if (cast<CallInst>(Caller)->isTailCall()) 1265 cast<CallInst>(NewCaller)->setTailCall(); 1266 cast<CallInst>(NewCaller)-> 1267 setCallingConv(cast<CallInst>(Caller)->getCallingConv()); 1268 cast<CallInst>(NewCaller)->setAttributes(NewPAL); 1269 } 1270 if (!Caller->getType()->isVoidTy()) 1271 Caller->replaceAllUsesWith(NewCaller); 1272 Caller->eraseFromParent(); 1273 Worklist.Remove(Caller); 1274 return 0; 1275 } 1276 } 1277 1278 // Replace the trampoline call with a direct call. Since there is no 'nest' 1279 // parameter, there is no need to adjust the argument list. Let the generic 1280 // code sort out any function type mismatches. 1281 Constant *NewCallee = 1282 NestF->getType() == PTy ? NestF : 1283 ConstantExpr::getBitCast(NestF, PTy); 1284 CS.setCalledFunction(NewCallee); 1285 return CS.getInstruction(); 1286} 1287 1288