InstCombineCalls.cpp revision 783a5c2b695015a5a2dfc48ee28b9936d0077542
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 alloca 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 case Intrinsic::smul_with_overflow: 480 // Canonicalize constants into the RHS. 481 if (isa<Constant>(II->getArgOperand(0)) && 482 !isa<Constant>(II->getArgOperand(1))) { 483 Value *LHS = II->getArgOperand(0); 484 II->setArgOperand(0, II->getArgOperand(1)); 485 II->setArgOperand(1, LHS); 486 return II; 487 } 488 489 // X * undef -> undef 490 if (isa<UndefValue>(II->getArgOperand(1))) 491 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 492 493 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 494 // X*0 -> {0, false} 495 if (RHSI->isZero()) 496 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType())); 497 498 // X * 1 -> {X, false} 499 if (RHSI->equalsInt(1)) { 500 Constant *V[] = { 501 UndefValue::get(II->getArgOperand(0)->getType()), 502 ConstantInt::getFalse(II->getContext()) 503 }; 504 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 505 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 506 } 507 } 508 break; 509 case Intrinsic::ppc_altivec_lvx: 510 case Intrinsic::ppc_altivec_lvxl: 511 case Intrinsic::x86_sse_loadu_ps: 512 case Intrinsic::x86_sse2_loadu_pd: 513 case Intrinsic::x86_sse2_loadu_dq: 514 // Turn PPC lvx -> load if the pointer is known aligned. 515 // Turn X86 loadups -> load if the pointer is known aligned. 516 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 517 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), 518 PointerType::getUnqual(II->getType())); 519 return new LoadInst(Ptr); 520 } 521 break; 522 case Intrinsic::ppc_altivec_stvx: 523 case Intrinsic::ppc_altivec_stvxl: 524 // Turn stvx -> store if the pointer is known aligned. 525 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) { 526 const Type *OpPtrTy = 527 PointerType::getUnqual(II->getArgOperand(0)->getType()); 528 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); 529 return new StoreInst(II->getArgOperand(0), Ptr); 530 } 531 break; 532 case Intrinsic::x86_sse_storeu_ps: 533 case Intrinsic::x86_sse2_storeu_pd: 534 case Intrinsic::x86_sse2_storeu_dq: 535 // Turn X86 storeu -> store if the pointer is known aligned. 536 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 537 const Type *OpPtrTy = 538 PointerType::getUnqual(II->getArgOperand(1)->getType()); 539 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy); 540 return new StoreInst(II->getArgOperand(1), Ptr); 541 } 542 break; 543 544 case Intrinsic::x86_sse_cvttss2si: { 545 // These intrinsics only demands the 0th element of its input vector. If 546 // we can simplify the input based on that, do so now. 547 unsigned VWidth = 548 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 549 APInt DemandedElts(VWidth, 1); 550 APInt UndefElts(VWidth, 0); 551 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0), 552 DemandedElts, UndefElts)) { 553 II->setArgOperand(0, V); 554 return II; 555 } 556 break; 557 } 558 559 case Intrinsic::ppc_altivec_vperm: 560 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant. 561 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getArgOperand(2))) { 562 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!"); 563 564 // Check that all of the elements are integer constants or undefs. 565 bool AllEltsOk = true; 566 for (unsigned i = 0; i != 16; ++i) { 567 if (!isa<ConstantInt>(Mask->getOperand(i)) && 568 !isa<UndefValue>(Mask->getOperand(i))) { 569 AllEltsOk = false; 570 break; 571 } 572 } 573 574 if (AllEltsOk) { 575 // Cast the input vectors to byte vectors. 576 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0), 577 Mask->getType()); 578 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1), 579 Mask->getType()); 580 Value *Result = UndefValue::get(Op0->getType()); 581 582 // Only extract each element once. 583 Value *ExtractedElts[32]; 584 memset(ExtractedElts, 0, sizeof(ExtractedElts)); 585 586 for (unsigned i = 0; i != 16; ++i) { 587 if (isa<UndefValue>(Mask->getOperand(i))) 588 continue; 589 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue(); 590 Idx &= 31; // Match the hardware behavior. 591 592 if (ExtractedElts[Idx] == 0) { 593 ExtractedElts[Idx] = 594 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1, 595 ConstantInt::get(Type::getInt32Ty(II->getContext()), 596 Idx&15, false), "tmp"); 597 } 598 599 // Insert this value into the result vector. 600 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx], 601 ConstantInt::get(Type::getInt32Ty(II->getContext()), 602 i, false), "tmp"); 603 } 604 return CastInst::Create(Instruction::BitCast, Result, CI.getType()); 605 } 606 } 607 break; 608 609 case Intrinsic::arm_neon_vld1: 610 case Intrinsic::arm_neon_vld2: 611 case Intrinsic::arm_neon_vld3: 612 case Intrinsic::arm_neon_vld4: 613 case Intrinsic::arm_neon_vld2lane: 614 case Intrinsic::arm_neon_vld3lane: 615 case Intrinsic::arm_neon_vld4lane: 616 case Intrinsic::arm_neon_vst1: 617 case Intrinsic::arm_neon_vst2: 618 case Intrinsic::arm_neon_vst3: 619 case Intrinsic::arm_neon_vst4: 620 case Intrinsic::arm_neon_vst2lane: 621 case Intrinsic::arm_neon_vst3lane: 622 case Intrinsic::arm_neon_vst4lane: { 623 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD); 624 unsigned AlignArg = II->getNumArgOperands() - 1; 625 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg)); 626 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) { 627 II->setArgOperand(AlignArg, 628 ConstantInt::get(Type::getInt32Ty(II->getContext()), 629 MemAlign, false)); 630 return II; 631 } 632 break; 633 } 634 635 case Intrinsic::stackrestore: { 636 // If the save is right next to the restore, remove the restore. This can 637 // happen when variable allocas are DCE'd. 638 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) { 639 if (SS->getIntrinsicID() == Intrinsic::stacksave) { 640 BasicBlock::iterator BI = SS; 641 if (&*++BI == II) 642 return EraseInstFromFunction(CI); 643 } 644 } 645 646 // Scan down this block to see if there is another stack restore in the 647 // same block without an intervening call/alloca. 648 BasicBlock::iterator BI = II; 649 TerminatorInst *TI = II->getParent()->getTerminator(); 650 bool CannotRemove = false; 651 for (++BI; &*BI != TI; ++BI) { 652 if (isa<AllocaInst>(BI) || isMalloc(BI)) { 653 CannotRemove = true; 654 break; 655 } 656 if (CallInst *BCI = dyn_cast<CallInst>(BI)) { 657 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) { 658 // If there is a stackrestore below this one, remove this one. 659 if (II->getIntrinsicID() == Intrinsic::stackrestore) 660 return EraseInstFromFunction(CI); 661 // Otherwise, ignore the intrinsic. 662 } else { 663 // If we found a non-intrinsic call, we can't remove the stack 664 // restore. 665 CannotRemove = true; 666 break; 667 } 668 } 669 } 670 671 // If the stack restore is in a return/unwind block and if there are no 672 // allocas or calls between the restore and the return, nuke the restore. 673 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI))) 674 return EraseInstFromFunction(CI); 675 break; 676 } 677 } 678 679 return visitCallSite(II); 680} 681 682// InvokeInst simplification 683// 684Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) { 685 return visitCallSite(&II); 686} 687 688/// isSafeToEliminateVarargsCast - If this cast does not affect the value 689/// passed through the varargs area, we can eliminate the use of the cast. 690static bool isSafeToEliminateVarargsCast(const CallSite CS, 691 const CastInst * const CI, 692 const TargetData * const TD, 693 const int ix) { 694 if (!CI->isLosslessCast()) 695 return false; 696 697 // The size of ByVal arguments is derived from the type, so we 698 // can't change to a type with a different size. If the size were 699 // passed explicitly we could avoid this check. 700 if (!CS.paramHasAttr(ix, Attribute::ByVal)) 701 return true; 702 703 const Type* SrcTy = 704 cast<PointerType>(CI->getOperand(0)->getType())->getElementType(); 705 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType(); 706 if (!SrcTy->isSized() || !DstTy->isSized()) 707 return false; 708 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy)) 709 return false; 710 return true; 711} 712 713namespace { 714class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls { 715 InstCombiner *IC; 716protected: 717 void replaceCall(Value *With) { 718 NewInstruction = IC->ReplaceInstUsesWith(*CI, With); 719 } 720 bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const { 721 if (ConstantInt *SizeCI = 722 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) { 723 if (SizeCI->isAllOnesValue()) 724 return true; 725 if (isString) 726 return SizeCI->getZExtValue() >= 727 GetStringLength(CI->getArgOperand(SizeArgOp)); 728 if (ConstantInt *Arg = dyn_cast<ConstantInt>( 729 CI->getArgOperand(SizeArgOp))) 730 return SizeCI->getZExtValue() >= Arg->getZExtValue(); 731 } 732 return false; 733 } 734public: 735 InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { } 736 Instruction *NewInstruction; 737}; 738} // end anonymous namespace 739 740// Try to fold some different type of calls here. 741// Currently we're only working with the checking functions, memcpy_chk, 742// mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk, 743// strcat_chk and strncat_chk. 744Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) { 745 if (CI->getCalledFunction() == 0) return 0; 746 747 InstCombineFortifiedLibCalls Simplifier(this); 748 Simplifier.fold(CI, TD); 749 return Simplifier.NewInstruction; 750} 751 752// visitCallSite - Improvements for call and invoke instructions. 753// 754Instruction *InstCombiner::visitCallSite(CallSite CS) { 755 bool Changed = false; 756 757 // If the callee is a pointer to a function, attempt to move any casts to the 758 // arguments of the call/invoke. 759 Value *Callee = CS.getCalledValue(); 760 if (!isa<Function>(Callee) && transformConstExprCastCall(CS)) 761 return 0; 762 763 if (Function *CalleeF = dyn_cast<Function>(Callee)) 764 // If the call and callee calling conventions don't match, this call must 765 // be unreachable, as the call is undefined. 766 if (CalleeF->getCallingConv() != CS.getCallingConv() && 767 // Only do this for calls to a function with a body. A prototype may 768 // not actually end up matching the implementation's calling conv for a 769 // variety of reasons (e.g. it may be written in assembly). 770 !CalleeF->isDeclaration()) { 771 Instruction *OldCall = CS.getInstruction(); 772 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 773 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 774 OldCall); 775 // If OldCall dues not return void then replaceAllUsesWith undef. 776 // This allows ValueHandlers and custom metadata to adjust itself. 777 if (!OldCall->getType()->isVoidTy()) 778 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType())); 779 if (isa<CallInst>(OldCall)) 780 return EraseInstFromFunction(*OldCall); 781 782 // We cannot remove an invoke, because it would change the CFG, just 783 // change the callee to a null pointer. 784 cast<InvokeInst>(OldCall)->setCalledFunction( 785 Constant::getNullValue(CalleeF->getType())); 786 return 0; 787 } 788 789 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 790 // This instruction is not reachable, just remove it. We insert a store to 791 // undef so that we know that this code is not reachable, despite the fact 792 // that we can't modify the CFG here. 793 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 794 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 795 CS.getInstruction()); 796 797 // If CS does not return void then replaceAllUsesWith undef. 798 // This allows ValueHandlers and custom metadata to adjust itself. 799 if (!CS.getInstruction()->getType()->isVoidTy()) 800 CS.getInstruction()-> 801 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType())); 802 803 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) { 804 // Don't break the CFG, insert a dummy cond branch. 805 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(), 806 ConstantInt::getTrue(Callee->getContext()), II); 807 } 808 return EraseInstFromFunction(*CS.getInstruction()); 809 } 810 811 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee)) 812 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0))) 813 if (In->getIntrinsicID() == Intrinsic::init_trampoline) 814 return transformCallThroughTrampoline(CS); 815 816 const PointerType *PTy = cast<PointerType>(Callee->getType()); 817 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 818 if (FTy->isVarArg()) { 819 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1); 820 // See if we can optimize any arguments passed through the varargs area of 821 // the call. 822 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(), 823 E = CS.arg_end(); I != E; ++I, ++ix) { 824 CastInst *CI = dyn_cast<CastInst>(*I); 825 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) { 826 *I = CI->getOperand(0); 827 Changed = true; 828 } 829 } 830 } 831 832 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) { 833 // Inline asm calls cannot throw - mark them 'nounwind'. 834 CS.setDoesNotThrow(); 835 Changed = true; 836 } 837 838 // Try to optimize the call if possible, we require TargetData for most of 839 // this. None of these calls are seen as possibly dead so go ahead and 840 // delete the instruction now. 841 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) { 842 Instruction *I = tryOptimizeCall(CI, TD); 843 // If we changed something return the result, etc. Otherwise let 844 // the fallthrough check. 845 if (I) return EraseInstFromFunction(*I); 846 } 847 848 return Changed ? CS.getInstruction() : 0; 849} 850 851// transformConstExprCastCall - If the callee is a constexpr cast of a function, 852// attempt to move the cast to the arguments of the call/invoke. 853// 854bool InstCombiner::transformConstExprCastCall(CallSite CS) { 855 Function *Callee = 856 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts()); 857 if (Callee == 0) 858 return false; 859 Instruction *Caller = CS.getInstruction(); 860 const AttrListPtr &CallerPAL = CS.getAttributes(); 861 862 // Okay, this is a cast from a function to a different type. Unless doing so 863 // would cause a type conversion of one of our arguments, change this call to 864 // be a direct call with arguments casted to the appropriate types. 865 // 866 const FunctionType *FT = Callee->getFunctionType(); 867 const Type *OldRetTy = Caller->getType(); 868 const Type *NewRetTy = FT->getReturnType(); 869 870 if (NewRetTy->isStructTy()) 871 return false; // TODO: Handle multiple return values. 872 873 // Check to see if we are changing the return type... 874 if (OldRetTy != NewRetTy) { 875 if (Callee->isDeclaration() && 876 // Conversion is ok if changing from one pointer type to another or from 877 // a pointer to an integer of the same size. 878 !((OldRetTy->isPointerTy() || !TD || 879 OldRetTy == TD->getIntPtrType(Caller->getContext())) && 880 (NewRetTy->isPointerTy() || !TD || 881 NewRetTy == TD->getIntPtrType(Caller->getContext())))) 882 return false; // Cannot transform this return value. 883 884 if (!Caller->use_empty() && 885 // void -> non-void is handled specially 886 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy)) 887 return false; // Cannot transform this return value. 888 889 if (!CallerPAL.isEmpty() && !Caller->use_empty()) { 890 Attributes RAttrs = CallerPAL.getRetAttributes(); 891 if (RAttrs & Attribute::typeIncompatible(NewRetTy)) 892 return false; // Attribute not compatible with transformed value. 893 } 894 895 // If the callsite is an invoke instruction, and the return value is used by 896 // a PHI node in a successor, we cannot change the return type of the call 897 // because there is no place to put the cast instruction (without breaking 898 // the critical edge). Bail out in this case. 899 if (!Caller->use_empty()) 900 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) 901 for (Value::use_iterator UI = II->use_begin(), E = II->use_end(); 902 UI != E; ++UI) 903 if (PHINode *PN = dyn_cast<PHINode>(*UI)) 904 if (PN->getParent() == II->getNormalDest() || 905 PN->getParent() == II->getUnwindDest()) 906 return false; 907 } 908 909 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin()); 910 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs); 911 912 CallSite::arg_iterator AI = CS.arg_begin(); 913 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) { 914 const Type *ParamTy = FT->getParamType(i); 915 const Type *ActTy = (*AI)->getType(); 916 917 if (!CastInst::isCastable(ActTy, ParamTy)) 918 return false; // Cannot transform this parameter value. 919 920 unsigned Attrs = CallerPAL.getParamAttributes(i + 1); 921 if (Attrs & Attribute::typeIncompatible(ParamTy)) 922 return false; // Attribute not compatible with transformed value. 923 924 // If the parameter is passed as a byval argument, then we have to have a 925 // sized type and the sized type has to have the same size as the old type. 926 if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) { 927 const PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy); 928 if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0) 929 return false; 930 931 const Type *CurElTy = cast<PointerType>(ActTy)->getElementType(); 932 if (TD->getTypeAllocSize(CurElTy) != 933 TD->getTypeAllocSize(ParamPTy->getElementType())) 934 return false; 935 } 936 937 // Converting from one pointer type to another or between a pointer and an 938 // integer of the same size is safe even if we do not have a body. 939 bool isConvertible = ActTy == ParamTy || 940 (TD && ((ParamTy->isPointerTy() || 941 ParamTy == TD->getIntPtrType(Caller->getContext())) && 942 (ActTy->isPointerTy() || 943 ActTy == TD->getIntPtrType(Caller->getContext())))); 944 if (Callee->isDeclaration() && !isConvertible) return false; 945 } 946 947 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() && 948 Callee->isDeclaration()) 949 return false; // Do not delete arguments unless we have a function body. 950 951 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && 952 !CallerPAL.isEmpty()) 953 // In this case we have more arguments than the new function type, but we 954 // won't be dropping them. Check that these extra arguments have attributes 955 // that are compatible with being a vararg call argument. 956 for (unsigned i = CallerPAL.getNumSlots(); i; --i) { 957 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams()) 958 break; 959 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs; 960 if (PAttrs & Attribute::VarArgsIncompatible) 961 return false; 962 } 963 964 // Okay, we decided that this is a safe thing to do: go ahead and start 965 // inserting cast instructions as necessary... 966 std::vector<Value*> Args; 967 Args.reserve(NumActualArgs); 968 SmallVector<AttributeWithIndex, 8> attrVec; 969 attrVec.reserve(NumCommonArgs); 970 971 // Get any return attributes. 972 Attributes RAttrs = CallerPAL.getRetAttributes(); 973 974 // If the return value is not being used, the type may not be compatible 975 // with the existing attributes. Wipe out any problematic attributes. 976 RAttrs &= ~Attribute::typeIncompatible(NewRetTy); 977 978 // Add the new return attributes. 979 if (RAttrs) 980 attrVec.push_back(AttributeWithIndex::get(0, RAttrs)); 981 982 AI = CS.arg_begin(); 983 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) { 984 const Type *ParamTy = FT->getParamType(i); 985 if ((*AI)->getType() == ParamTy) { 986 Args.push_back(*AI); 987 } else { 988 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, 989 false, ParamTy, false); 990 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp")); 991 } 992 993 // Add any parameter attributes. 994 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 995 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 996 } 997 998 // If the function takes more arguments than the call was taking, add them 999 // now. 1000 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) 1001 Args.push_back(Constant::getNullValue(FT->getParamType(i))); 1002 1003 // If we are removing arguments to the function, emit an obnoxious warning. 1004 if (FT->getNumParams() < NumActualArgs) { 1005 if (!FT->isVarArg()) { 1006 errs() << "WARNING: While resolving call to function '" 1007 << Callee->getName() << "' arguments were dropped!\n"; 1008 } else { 1009 // Add all of the arguments in their promoted form to the arg list. 1010 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) { 1011 const Type *PTy = getPromotedType((*AI)->getType()); 1012 if (PTy != (*AI)->getType()) { 1013 // Must promote to pass through va_arg area! 1014 Instruction::CastOps opcode = 1015 CastInst::getCastOpcode(*AI, false, PTy, false); 1016 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp")); 1017 } else { 1018 Args.push_back(*AI); 1019 } 1020 1021 // Add any parameter attributes. 1022 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1023 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1024 } 1025 } 1026 } 1027 1028 if (Attributes FnAttrs = CallerPAL.getFnAttributes()) 1029 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); 1030 1031 if (NewRetTy->isVoidTy()) 1032 Caller->setName(""); // Void type should not have a name. 1033 1034 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(), 1035 attrVec.end()); 1036 1037 Instruction *NC; 1038 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1039 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(), 1040 Args.begin(), Args.end(), 1041 Caller->getName(), Caller); 1042 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv()); 1043 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL); 1044 } else { 1045 NC = CallInst::Create(Callee, Args.begin(), Args.end(), 1046 Caller->getName(), Caller); 1047 CallInst *CI = cast<CallInst>(Caller); 1048 if (CI->isTailCall()) 1049 cast<CallInst>(NC)->setTailCall(); 1050 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv()); 1051 cast<CallInst>(NC)->setAttributes(NewCallerPAL); 1052 } 1053 1054 // Insert a cast of the return type as necessary. 1055 Value *NV = NC; 1056 if (OldRetTy != NV->getType() && !Caller->use_empty()) { 1057 if (!NV->getType()->isVoidTy()) { 1058 Instruction::CastOps opcode = 1059 CastInst::getCastOpcode(NC, false, OldRetTy, false); 1060 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp"); 1061 1062 // If this is an invoke instruction, we should insert it after the first 1063 // non-phi, instruction in the normal successor block. 1064 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1065 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI(); 1066 InsertNewInstBefore(NC, *I); 1067 } else { 1068 // Otherwise, it's a call, just insert cast right after the call. 1069 InsertNewInstBefore(NC, *Caller); 1070 } 1071 Worklist.AddUsersToWorkList(*Caller); 1072 } else { 1073 NV = UndefValue::get(Caller->getType()); 1074 } 1075 } 1076 1077 if (!Caller->use_empty()) 1078 Caller->replaceAllUsesWith(NV); 1079 1080 EraseInstFromFunction(*Caller); 1081 return true; 1082} 1083 1084// transformCallThroughTrampoline - Turn a call to a function created by the 1085// init_trampoline intrinsic into a direct call to the underlying function. 1086// 1087Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) { 1088 Value *Callee = CS.getCalledValue(); 1089 const PointerType *PTy = cast<PointerType>(Callee->getType()); 1090 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 1091 const AttrListPtr &Attrs = CS.getAttributes(); 1092 1093 // If the call already has the 'nest' attribute somewhere then give up - 1094 // otherwise 'nest' would occur twice after splicing in the chain. 1095 if (Attrs.hasAttrSomewhere(Attribute::Nest)) 1096 return 0; 1097 1098 IntrinsicInst *Tramp = 1099 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0)); 1100 1101 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts()); 1102 const PointerType *NestFPTy = cast<PointerType>(NestF->getType()); 1103 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType()); 1104 1105 const AttrListPtr &NestAttrs = NestF->getAttributes(); 1106 if (!NestAttrs.isEmpty()) { 1107 unsigned NestIdx = 1; 1108 const Type *NestTy = 0; 1109 Attributes NestAttr = Attribute::None; 1110 1111 // Look for a parameter marked with the 'nest' attribute. 1112 for (FunctionType::param_iterator I = NestFTy->param_begin(), 1113 E = NestFTy->param_end(); I != E; ++NestIdx, ++I) 1114 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) { 1115 // Record the parameter type and any other attributes. 1116 NestTy = *I; 1117 NestAttr = NestAttrs.getParamAttributes(NestIdx); 1118 break; 1119 } 1120 1121 if (NestTy) { 1122 Instruction *Caller = CS.getInstruction(); 1123 std::vector<Value*> NewArgs; 1124 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1); 1125 1126 SmallVector<AttributeWithIndex, 8> NewAttrs; 1127 NewAttrs.reserve(Attrs.getNumSlots() + 1); 1128 1129 // Insert the nest argument into the call argument list, which may 1130 // mean appending it. Likewise for attributes. 1131 1132 // Add any result attributes. 1133 if (Attributes Attr = Attrs.getRetAttributes()) 1134 NewAttrs.push_back(AttributeWithIndex::get(0, Attr)); 1135 1136 { 1137 unsigned Idx = 1; 1138 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 1139 do { 1140 if (Idx == NestIdx) { 1141 // Add the chain argument and attributes. 1142 Value *NestVal = Tramp->getArgOperand(2); 1143 if (NestVal->getType() != NestTy) 1144 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller); 1145 NewArgs.push_back(NestVal); 1146 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr)); 1147 } 1148 1149 if (I == E) 1150 break; 1151 1152 // Add the original argument and attributes. 1153 NewArgs.push_back(*I); 1154 if (Attributes Attr = Attrs.getParamAttributes(Idx)) 1155 NewAttrs.push_back 1156 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr)); 1157 1158 ++Idx, ++I; 1159 } while (1); 1160 } 1161 1162 // Add any function attributes. 1163 if (Attributes Attr = Attrs.getFnAttributes()) 1164 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr)); 1165 1166 // The trampoline may have been bitcast to a bogus type (FTy). 1167 // Handle this by synthesizing a new function type, equal to FTy 1168 // with the chain parameter inserted. 1169 1170 std::vector<const Type*> NewTypes; 1171 NewTypes.reserve(FTy->getNumParams()+1); 1172 1173 // Insert the chain's type into the list of parameter types, which may 1174 // mean appending it. 1175 { 1176 unsigned Idx = 1; 1177 FunctionType::param_iterator I = FTy->param_begin(), 1178 E = FTy->param_end(); 1179 1180 do { 1181 if (Idx == NestIdx) 1182 // Add the chain's type. 1183 NewTypes.push_back(NestTy); 1184 1185 if (I == E) 1186 break; 1187 1188 // Add the original type. 1189 NewTypes.push_back(*I); 1190 1191 ++Idx, ++I; 1192 } while (1); 1193 } 1194 1195 // Replace the trampoline call with a direct call. Let the generic 1196 // code sort out any function type mismatches. 1197 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes, 1198 FTy->isVarArg()); 1199 Constant *NewCallee = 1200 NestF->getType() == PointerType::getUnqual(NewFTy) ? 1201 NestF : ConstantExpr::getBitCast(NestF, 1202 PointerType::getUnqual(NewFTy)); 1203 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(), 1204 NewAttrs.end()); 1205 1206 Instruction *NewCaller; 1207 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1208 NewCaller = InvokeInst::Create(NewCallee, 1209 II->getNormalDest(), II->getUnwindDest(), 1210 NewArgs.begin(), NewArgs.end(), 1211 Caller->getName(), Caller); 1212 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv()); 1213 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL); 1214 } else { 1215 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(), 1216 Caller->getName(), Caller); 1217 if (cast<CallInst>(Caller)->isTailCall()) 1218 cast<CallInst>(NewCaller)->setTailCall(); 1219 cast<CallInst>(NewCaller)-> 1220 setCallingConv(cast<CallInst>(Caller)->getCallingConv()); 1221 cast<CallInst>(NewCaller)->setAttributes(NewPAL); 1222 } 1223 if (!Caller->getType()->isVoidTy()) 1224 Caller->replaceAllUsesWith(NewCaller); 1225 Caller->eraseFromParent(); 1226 Worklist.Remove(Caller); 1227 return 0; 1228 } 1229 } 1230 1231 // Replace the trampoline call with a direct call. Since there is no 'nest' 1232 // parameter, there is no need to adjust the argument list. Let the generic 1233 // code sort out any function type mismatches. 1234 Constant *NewCallee = 1235 NestF->getType() == PTy ? NestF : 1236 ConstantExpr::getBitCast(NestF, PTy); 1237 CS.setCalledFunction(NewCallee); 1238 return CS.getInstruction(); 1239} 1240 1241