InstCombineCalls.cpp revision 4a544a79bd735967f1d33fe675ae4566dbd17813
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/Support/CallSite.h" 16#include "llvm/Target/TargetData.h" 17#include "llvm/Analysis/MemoryBuiltins.h" 18#include "llvm/Transforms/Utils/BuildLibCalls.h" 19#include "llvm/Transforms/Utils/Local.h" 20using namespace llvm; 21 22/// getPromotedType - Return the specified type promoted as it would be to pass 23/// though a va_arg area. 24static Type *getPromotedType(Type *Ty) { 25 if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) { 26 if (ITy->getBitWidth() < 32) 27 return Type::getInt32Ty(Ty->getContext()); 28 } 29 return Ty; 30} 31 32 33Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) { 34 unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), TD); 35 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), TD); 36 unsigned MinAlign = std::min(DstAlign, SrcAlign); 37 unsigned CopyAlign = MI->getAlignment(); 38 39 if (CopyAlign < MinAlign) { 40 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), 41 MinAlign, false)); 42 return MI; 43 } 44 45 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with 46 // load/store. 47 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2)); 48 if (MemOpLength == 0) return 0; 49 50 // Source and destination pointer types are always "i8*" for intrinsic. See 51 // if the size is something we can handle with a single primitive load/store. 52 // A single load+store correctly handles overlapping memory in the memmove 53 // case. 54 unsigned Size = MemOpLength->getZExtValue(); 55 if (Size == 0) return MI; // Delete this mem transfer. 56 57 if (Size > 8 || (Size&(Size-1))) 58 return 0; // If not 1/2/4/8 bytes, exit. 59 60 // Use an integer load+store unless we can find something better. 61 unsigned SrcAddrSp = 62 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace(); 63 unsigned DstAddrSp = 64 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace(); 65 66 IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3); 67 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp); 68 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp); 69 70 // Memcpy forces the use of i8* for the source and destination. That means 71 // that if you're using memcpy to move one double around, you'll get a cast 72 // from double* to i8*. We'd much rather use a double load+store rather than 73 // an i64 load+store, here because this improves the odds that the source or 74 // dest address will be promotable. See if we can find a better type than the 75 // integer datatype. 76 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts(); 77 if (StrippedDest != MI->getArgOperand(0)) { 78 Type *SrcETy = cast<PointerType>(StrippedDest->getType()) 79 ->getElementType(); 80 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) { 81 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip 82 // down through these levels if so. 83 while (!SrcETy->isSingleValueType()) { 84 if (StructType *STy = dyn_cast<StructType>(SrcETy)) { 85 if (STy->getNumElements() == 1) 86 SrcETy = STy->getElementType(0); 87 else 88 break; 89 } else if (ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) { 90 if (ATy->getNumElements() == 1) 91 SrcETy = ATy->getElementType(); 92 else 93 break; 94 } else 95 break; 96 } 97 98 if (SrcETy->isSingleValueType()) { 99 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp); 100 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp); 101 } 102 } 103 } 104 105 106 // If the memcpy/memmove provides better alignment info than we can 107 // infer, use it. 108 SrcAlign = std::max(SrcAlign, CopyAlign); 109 DstAlign = std::max(DstAlign, CopyAlign); 110 111 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy); 112 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy); 113 LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile()); 114 L->setAlignment(SrcAlign); 115 StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile()); 116 S->setAlignment(DstAlign); 117 118 // Set the size of the copy to 0, it will be deleted on the next iteration. 119 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType())); 120 return MI; 121} 122 123Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) { 124 unsigned Alignment = getKnownAlignment(MI->getDest(), TD); 125 if (MI->getAlignment() < Alignment) { 126 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), 127 Alignment, false)); 128 return MI; 129 } 130 131 // Extract the length and alignment and fill if they are constant. 132 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength()); 133 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue()); 134 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8)) 135 return 0; 136 uint64_t Len = LenC->getZExtValue(); 137 Alignment = MI->getAlignment(); 138 139 // If the length is zero, this is a no-op 140 if (Len == 0) return MI; // memset(d,c,0,a) -> noop 141 142 // memset(s,c,n) -> store s, c (for n=1,2,4,8) 143 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) { 144 Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8. 145 146 Value *Dest = MI->getDest(); 147 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace(); 148 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp); 149 Dest = Builder->CreateBitCast(Dest, NewDstPtrTy); 150 151 // Alignment 0 is identity for alignment 1 for memset, but not store. 152 if (Alignment == 0) Alignment = 1; 153 154 // Extract the fill value and store. 155 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL; 156 StoreInst *S = Builder->CreateStore(ConstantInt::get(ITy, Fill), Dest, 157 MI->isVolatile()); 158 S->setAlignment(Alignment); 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 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)); 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 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(), Ops); 269 270 Op1 = GEP->getPointerOperand()->stripPointerCasts(); 271 272 // Make sure we're not a constant offset from an external 273 // global. 274 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) 275 if (!GV->hasDefinitiveInitializer()) break; 276 } 277 278 // If we've stripped down to a single global variable that we 279 // can know the size of then just return that. 280 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) { 281 if (GV->hasDefinitiveInitializer()) { 282 Constant *C = GV->getInitializer(); 283 Size = TD->getTypeAllocSize(C->getType()); 284 } else { 285 // Can't determine size of the GV. 286 Constant *RetVal = ConstantInt::get(ReturnTy, DontKnow); 287 return ReplaceInstUsesWith(CI, RetVal); 288 } 289 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) { 290 // Get alloca size. 291 if (AI->getAllocatedType()->isSized()) { 292 Size = TD->getTypeAllocSize(AI->getAllocatedType()); 293 if (AI->isArrayAllocation()) { 294 const ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize()); 295 if (!C) break; 296 Size *= C->getZExtValue(); 297 } 298 } 299 } else if (CallInst *MI = extractMallocCall(Op1)) { 300 // Get allocation size. 301 Type* MallocType = getMallocAllocatedType(MI); 302 if (MallocType && MallocType->isSized()) 303 if (Value *NElems = getMallocArraySize(MI, TD, true)) 304 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems)) 305 Size = NElements->getZExtValue() * TD->getTypeAllocSize(MallocType); 306 } 307 308 // Do not return "I don't know" here. Later optimization passes could 309 // make it possible to evaluate objectsize to a constant. 310 if (Size == -1ULL) 311 break; 312 313 if (Size < Offset) { 314 // Out of bound reference? Negative index normalized to large 315 // index? Just return "I don't know". 316 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, DontKnow)); 317 } 318 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, Size-Offset)); 319 } 320 case Intrinsic::bswap: 321 // bswap(bswap(x)) -> x 322 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) 323 if (Operand->getIntrinsicID() == Intrinsic::bswap) 324 return ReplaceInstUsesWith(CI, Operand->getArgOperand(0)); 325 326 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c)) 327 if (TruncInst *TI = dyn_cast<TruncInst>(II->getArgOperand(0))) { 328 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0))) 329 if (Operand->getIntrinsicID() == Intrinsic::bswap) { 330 unsigned C = Operand->getType()->getPrimitiveSizeInBits() - 331 TI->getType()->getPrimitiveSizeInBits(); 332 Value *CV = ConstantInt::get(Operand->getType(), C); 333 Value *V = Builder->CreateLShr(Operand->getArgOperand(0), CV); 334 return new TruncInst(V, TI->getType()); 335 } 336 } 337 338 break; 339 case Intrinsic::powi: 340 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 341 // powi(x, 0) -> 1.0 342 if (Power->isZero()) 343 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0)); 344 // powi(x, 1) -> x 345 if (Power->isOne()) 346 return ReplaceInstUsesWith(CI, II->getArgOperand(0)); 347 // powi(x, -1) -> 1/x 348 if (Power->isAllOnesValue()) 349 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0), 350 II->getArgOperand(0)); 351 } 352 break; 353 case Intrinsic::cttz: { 354 // If all bits below the first known one are known zero, 355 // this value is constant. 356 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType()); 357 // FIXME: Try to simplify vectors of integers. 358 if (!IT) break; 359 uint32_t BitWidth = IT->getBitWidth(); 360 APInt KnownZero(BitWidth, 0); 361 APInt KnownOne(BitWidth, 0); 362 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth), 363 KnownZero, KnownOne); 364 unsigned TrailingZeros = KnownOne.countTrailingZeros(); 365 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros)); 366 if ((Mask & KnownZero) == Mask) 367 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 368 APInt(BitWidth, TrailingZeros))); 369 370 } 371 break; 372 case Intrinsic::ctlz: { 373 // If all bits above the first known one are known zero, 374 // this value is constant. 375 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType()); 376 // FIXME: Try to simplify vectors of integers. 377 if (!IT) break; 378 uint32_t BitWidth = IT->getBitWidth(); 379 APInt KnownZero(BitWidth, 0); 380 APInt KnownOne(BitWidth, 0); 381 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth), 382 KnownZero, KnownOne); 383 unsigned LeadingZeros = KnownOne.countLeadingZeros(); 384 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros)); 385 if ((Mask & KnownZero) == Mask) 386 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 387 APInt(BitWidth, LeadingZeros))); 388 389 } 390 break; 391 case Intrinsic::uadd_with_overflow: { 392 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 393 IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType()); 394 uint32_t BitWidth = IT->getBitWidth(); 395 APInt Mask = APInt::getSignBit(BitWidth); 396 APInt LHSKnownZero(BitWidth, 0); 397 APInt LHSKnownOne(BitWidth, 0); 398 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne); 399 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1]; 400 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1]; 401 402 if (LHSKnownNegative || LHSKnownPositive) { 403 APInt RHSKnownZero(BitWidth, 0); 404 APInt RHSKnownOne(BitWidth, 0); 405 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne); 406 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1]; 407 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1]; 408 if (LHSKnownNegative && RHSKnownNegative) { 409 // The sign bit is set in both cases: this MUST overflow. 410 // Create a simple add instruction, and insert it into the struct. 411 Value *Add = Builder->CreateAdd(LHS, RHS); 412 Add->takeName(&CI); 413 Constant *V[] = { 414 UndefValue::get(LHS->getType()), 415 ConstantInt::getTrue(II->getContext()) 416 }; 417 StructType *ST = cast<StructType>(II->getType()); 418 Constant *Struct = ConstantStruct::get(ST, V); 419 return InsertValueInst::Create(Struct, Add, 0); 420 } 421 422 if (LHSKnownPositive && RHSKnownPositive) { 423 // The sign bit is clear in both cases: this CANNOT overflow. 424 // Create a simple add instruction, and insert it into the struct. 425 Value *Add = Builder->CreateNUWAdd(LHS, RHS); 426 Add->takeName(&CI); 427 Constant *V[] = { 428 UndefValue::get(LHS->getType()), 429 ConstantInt::getFalse(II->getContext()) 430 }; 431 StructType *ST = cast<StructType>(II->getType()); 432 Constant *Struct = ConstantStruct::get(ST, V); 433 return InsertValueInst::Create(Struct, Add, 0); 434 } 435 } 436 } 437 // FALL THROUGH uadd into sadd 438 case Intrinsic::sadd_with_overflow: 439 // Canonicalize constants into the RHS. 440 if (isa<Constant>(II->getArgOperand(0)) && 441 !isa<Constant>(II->getArgOperand(1))) { 442 Value *LHS = II->getArgOperand(0); 443 II->setArgOperand(0, II->getArgOperand(1)); 444 II->setArgOperand(1, LHS); 445 return II; 446 } 447 448 // X + undef -> undef 449 if (isa<UndefValue>(II->getArgOperand(1))) 450 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 451 452 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 453 // X + 0 -> {X, false} 454 if (RHS->isZero()) { 455 Constant *V[] = { 456 UndefValue::get(II->getArgOperand(0)->getType()), 457 ConstantInt::getFalse(II->getContext()) 458 }; 459 Constant *Struct = 460 ConstantStruct::get(cast<StructType>(II->getType()), V); 461 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 462 } 463 } 464 break; 465 case Intrinsic::usub_with_overflow: 466 case Intrinsic::ssub_with_overflow: 467 // undef - X -> undef 468 // X - undef -> undef 469 if (isa<UndefValue>(II->getArgOperand(0)) || 470 isa<UndefValue>(II->getArgOperand(1))) 471 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 472 473 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 474 // X - 0 -> {X, false} 475 if (RHS->isZero()) { 476 Constant *V[] = { 477 UndefValue::get(II->getArgOperand(0)->getType()), 478 ConstantInt::getFalse(II->getContext()) 479 }; 480 Constant *Struct = 481 ConstantStruct::get(cast<StructType>(II->getType()), V); 482 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 483 } 484 } 485 break; 486 case Intrinsic::umul_with_overflow: { 487 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 488 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth(); 489 APInt Mask = APInt::getAllOnesValue(BitWidth); 490 491 APInt LHSKnownZero(BitWidth, 0); 492 APInt LHSKnownOne(BitWidth, 0); 493 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne); 494 APInt RHSKnownZero(BitWidth, 0); 495 APInt RHSKnownOne(BitWidth, 0); 496 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne); 497 498 // Get the largest possible values for each operand. 499 APInt LHSMax = ~LHSKnownZero; 500 APInt RHSMax = ~RHSKnownZero; 501 502 // If multiplying the maximum values does not overflow then we can turn 503 // this into a plain NUW mul. 504 bool Overflow; 505 LHSMax.umul_ov(RHSMax, Overflow); 506 if (!Overflow) { 507 Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow"); 508 Constant *V[] = { 509 UndefValue::get(LHS->getType()), 510 Builder->getFalse() 511 }; 512 Constant *Struct = ConstantStruct::get(cast<StructType>(II->getType()),V); 513 return InsertValueInst::Create(Struct, Mul, 0); 514 } 515 } // FALL THROUGH 516 case Intrinsic::smul_with_overflow: 517 // Canonicalize constants into the RHS. 518 if (isa<Constant>(II->getArgOperand(0)) && 519 !isa<Constant>(II->getArgOperand(1))) { 520 Value *LHS = II->getArgOperand(0); 521 II->setArgOperand(0, II->getArgOperand(1)); 522 II->setArgOperand(1, LHS); 523 return II; 524 } 525 526 // X * undef -> undef 527 if (isa<UndefValue>(II->getArgOperand(1))) 528 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 529 530 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 531 // X*0 -> {0, false} 532 if (RHSI->isZero()) 533 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType())); 534 535 // X * 1 -> {X, false} 536 if (RHSI->equalsInt(1)) { 537 Constant *V[] = { 538 UndefValue::get(II->getArgOperand(0)->getType()), 539 ConstantInt::getFalse(II->getContext()) 540 }; 541 Constant *Struct = 542 ConstantStruct::get(cast<StructType>(II->getType()), V); 543 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 544 } 545 } 546 break; 547 case Intrinsic::ppc_altivec_lvx: 548 case Intrinsic::ppc_altivec_lvxl: 549 // Turn PPC lvx -> load if the pointer is known aligned. 550 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 551 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), 552 PointerType::getUnqual(II->getType())); 553 return new LoadInst(Ptr); 554 } 555 break; 556 case Intrinsic::ppc_altivec_stvx: 557 case Intrinsic::ppc_altivec_stvxl: 558 // Turn stvx -> store if the pointer is known aligned. 559 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) { 560 Type *OpPtrTy = 561 PointerType::getUnqual(II->getArgOperand(0)->getType()); 562 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); 563 return new StoreInst(II->getArgOperand(0), Ptr); 564 } 565 break; 566 case Intrinsic::x86_sse_storeu_ps: 567 case Intrinsic::x86_sse2_storeu_pd: 568 case Intrinsic::x86_sse2_storeu_dq: 569 // Turn X86 storeu -> store if the pointer is known aligned. 570 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 571 Type *OpPtrTy = 572 PointerType::getUnqual(II->getArgOperand(1)->getType()); 573 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy); 574 return new StoreInst(II->getArgOperand(1), Ptr); 575 } 576 break; 577 578 case Intrinsic::x86_sse_cvtss2si: 579 case Intrinsic::x86_sse_cvtss2si64: 580 case Intrinsic::x86_sse_cvttss2si: 581 case Intrinsic::x86_sse_cvttss2si64: 582 case Intrinsic::x86_sse2_cvtsd2si: 583 case Intrinsic::x86_sse2_cvtsd2si64: 584 case Intrinsic::x86_sse2_cvttsd2si: 585 case Intrinsic::x86_sse2_cvttsd2si64: { 586 // These intrinsics only demand the 0th element of their input vectors. If 587 // we can simplify the input based on that, do so now. 588 unsigned VWidth = 589 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 590 APInt DemandedElts(VWidth, 1); 591 APInt UndefElts(VWidth, 0); 592 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0), 593 DemandedElts, UndefElts)) { 594 II->setArgOperand(0, V); 595 return II; 596 } 597 break; 598 } 599 600 601 case Intrinsic::x86_sse41_pmovsxbw: 602 case Intrinsic::x86_sse41_pmovsxwd: 603 case Intrinsic::x86_sse41_pmovsxdq: 604 case Intrinsic::x86_sse41_pmovzxbw: 605 case Intrinsic::x86_sse41_pmovzxwd: 606 case Intrinsic::x86_sse41_pmovzxdq: { 607 // pmov{s|z}x ignores the upper half of their input vectors. 608 unsigned VWidth = 609 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 610 unsigned LowHalfElts = VWidth / 2; 611 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts)); 612 APInt UndefElts(VWidth, 0); 613 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), 614 InputDemandedElts, 615 UndefElts)) { 616 II->setArgOperand(0, TmpV); 617 return II; 618 } 619 break; 620 } 621 622 case Intrinsic::ppc_altivec_vperm: 623 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant. 624 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getArgOperand(2))) { 625 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!"); 626 627 // Check that all of the elements are integer constants or undefs. 628 bool AllEltsOk = true; 629 for (unsigned i = 0; i != 16; ++i) { 630 if (!isa<ConstantInt>(Mask->getOperand(i)) && 631 !isa<UndefValue>(Mask->getOperand(i))) { 632 AllEltsOk = false; 633 break; 634 } 635 } 636 637 if (AllEltsOk) { 638 // Cast the input vectors to byte vectors. 639 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0), 640 Mask->getType()); 641 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1), 642 Mask->getType()); 643 Value *Result = UndefValue::get(Op0->getType()); 644 645 // Only extract each element once. 646 Value *ExtractedElts[32]; 647 memset(ExtractedElts, 0, sizeof(ExtractedElts)); 648 649 for (unsigned i = 0; i != 16; ++i) { 650 if (isa<UndefValue>(Mask->getOperand(i))) 651 continue; 652 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue(); 653 Idx &= 31; // Match the hardware behavior. 654 655 if (ExtractedElts[Idx] == 0) { 656 ExtractedElts[Idx] = 657 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1, 658 ConstantInt::get(Type::getInt32Ty(II->getContext()), 659 Idx&15, false), "tmp"); 660 } 661 662 // Insert this value into the result vector. 663 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx], 664 ConstantInt::get(Type::getInt32Ty(II->getContext()), 665 i, false), "tmp"); 666 } 667 return CastInst::Create(Instruction::BitCast, Result, CI.getType()); 668 } 669 } 670 break; 671 672 case Intrinsic::arm_neon_vld1: 673 case Intrinsic::arm_neon_vld2: 674 case Intrinsic::arm_neon_vld3: 675 case Intrinsic::arm_neon_vld4: 676 case Intrinsic::arm_neon_vld2lane: 677 case Intrinsic::arm_neon_vld3lane: 678 case Intrinsic::arm_neon_vld4lane: 679 case Intrinsic::arm_neon_vst1: 680 case Intrinsic::arm_neon_vst2: 681 case Intrinsic::arm_neon_vst3: 682 case Intrinsic::arm_neon_vst4: 683 case Intrinsic::arm_neon_vst2lane: 684 case Intrinsic::arm_neon_vst3lane: 685 case Intrinsic::arm_neon_vst4lane: { 686 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD); 687 unsigned AlignArg = II->getNumArgOperands() - 1; 688 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg)); 689 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) { 690 II->setArgOperand(AlignArg, 691 ConstantInt::get(Type::getInt32Ty(II->getContext()), 692 MemAlign, false)); 693 return II; 694 } 695 break; 696 } 697 698 case Intrinsic::stackrestore: { 699 // If the save is right next to the restore, remove the restore. This can 700 // happen when variable allocas are DCE'd. 701 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) { 702 if (SS->getIntrinsicID() == Intrinsic::stacksave) { 703 BasicBlock::iterator BI = SS; 704 if (&*++BI == II) 705 return EraseInstFromFunction(CI); 706 } 707 } 708 709 // Scan down this block to see if there is another stack restore in the 710 // same block without an intervening call/alloca. 711 BasicBlock::iterator BI = II; 712 TerminatorInst *TI = II->getParent()->getTerminator(); 713 bool CannotRemove = false; 714 for (++BI; &*BI != TI; ++BI) { 715 if (isa<AllocaInst>(BI) || isMalloc(BI)) { 716 CannotRemove = true; 717 break; 718 } 719 if (CallInst *BCI = dyn_cast<CallInst>(BI)) { 720 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) { 721 // If there is a stackrestore below this one, remove this one. 722 if (II->getIntrinsicID() == Intrinsic::stackrestore) 723 return EraseInstFromFunction(CI); 724 // Otherwise, ignore the intrinsic. 725 } else { 726 // If we found a non-intrinsic call, we can't remove the stack 727 // restore. 728 CannotRemove = true; 729 break; 730 } 731 } 732 } 733 734 // If the stack restore is in a return, resume, or unwind block and if there 735 // are no allocas or calls between the restore and the return, nuke the 736 // restore. 737 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI) || 738 isa<UnwindInst>(TI))) 739 return EraseInstFromFunction(CI); 740 break; 741 } 742 } 743 744 return visitCallSite(II); 745} 746 747// InvokeInst simplification 748// 749Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) { 750 return visitCallSite(&II); 751} 752 753/// isSafeToEliminateVarargsCast - If this cast does not affect the value 754/// passed through the varargs area, we can eliminate the use of the cast. 755static bool isSafeToEliminateVarargsCast(const CallSite CS, 756 const CastInst * const CI, 757 const TargetData * const TD, 758 const int ix) { 759 if (!CI->isLosslessCast()) 760 return false; 761 762 // The size of ByVal arguments is derived from the type, so we 763 // can't change to a type with a different size. If the size were 764 // passed explicitly we could avoid this check. 765 if (!CS.paramHasAttr(ix, Attribute::ByVal)) 766 return true; 767 768 Type* SrcTy = 769 cast<PointerType>(CI->getOperand(0)->getType())->getElementType(); 770 Type* DstTy = cast<PointerType>(CI->getType())->getElementType(); 771 if (!SrcTy->isSized() || !DstTy->isSized()) 772 return false; 773 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy)) 774 return false; 775 return true; 776} 777 778namespace { 779class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls { 780 InstCombiner *IC; 781protected: 782 void replaceCall(Value *With) { 783 NewInstruction = IC->ReplaceInstUsesWith(*CI, With); 784 } 785 bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const { 786 if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp)) 787 return true; 788 if (ConstantInt *SizeCI = 789 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) { 790 if (SizeCI->isAllOnesValue()) 791 return true; 792 if (isString) { 793 uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp)); 794 // If the length is 0 we don't know how long it is and so we can't 795 // remove the check. 796 if (Len == 0) return false; 797 return SizeCI->getZExtValue() >= Len; 798 } 799 if (ConstantInt *Arg = dyn_cast<ConstantInt>( 800 CI->getArgOperand(SizeArgOp))) 801 return SizeCI->getZExtValue() >= Arg->getZExtValue(); 802 } 803 return false; 804 } 805public: 806 InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { } 807 Instruction *NewInstruction; 808}; 809} // end anonymous namespace 810 811// Try to fold some different type of calls here. 812// Currently we're only working with the checking functions, memcpy_chk, 813// mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk, 814// strcat_chk and strncat_chk. 815Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) { 816 if (CI->getCalledFunction() == 0) return 0; 817 818 InstCombineFortifiedLibCalls Simplifier(this); 819 Simplifier.fold(CI, TD); 820 return Simplifier.NewInstruction; 821} 822 823static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) { 824 // Strip off at most one level of pointer casts, looking for an alloca. This 825 // is good enough in practice and simpler than handling any number of casts. 826 Value *Underlying = TrampMem->stripPointerCasts(); 827 if (Underlying != TrampMem && 828 (!Underlying->hasOneUse() || *Underlying->use_begin() != TrampMem)) 829 return 0; 830 if (!isa<AllocaInst>(Underlying)) 831 return 0; 832 833 IntrinsicInst *InitTrampoline = 0; 834 for (Value::use_iterator I = TrampMem->use_begin(), E = TrampMem->use_end(); 835 I != E; I++) { 836 IntrinsicInst *II = dyn_cast<IntrinsicInst>(*I); 837 if (!II) 838 return 0; 839 if (II->getIntrinsicID() == Intrinsic::init_trampoline) { 840 if (InitTrampoline) 841 // More than one init_trampoline writes to this value. Give up. 842 return 0; 843 InitTrampoline = II; 844 continue; 845 } 846 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline) 847 // Allow any number of calls to adjust.trampoline. 848 continue; 849 return 0; 850 } 851 852 // No call to init.trampoline found. 853 if (!InitTrampoline) 854 return 0; 855 856 // Check that the alloca is being used in the expected way. 857 if (InitTrampoline->getOperand(0) != TrampMem) 858 return 0; 859 860 return InitTrampoline; 861} 862 863static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp, 864 Value *TrampMem) { 865 // Visit all the previous instructions in the basic block, and try to find a 866 // init.trampoline which has a direct path to the adjust.trampoline. 867 for (BasicBlock::iterator I = AdjustTramp, 868 E = AdjustTramp->getParent()->begin(); I != E; ) { 869 Instruction *Inst = --I; 870 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) 871 if (II->getIntrinsicID() == Intrinsic::init_trampoline && 872 II->getOperand(0) == TrampMem) 873 return II; 874 if (Inst->mayWriteToMemory()) 875 return 0; 876 } 877 return 0; 878} 879 880// Given a call to llvm.adjust.trampoline, find and return the corresponding 881// call to llvm.init.trampoline if the call to the trampoline can be optimized 882// to a direct call to a function. Otherwise return NULL. 883// 884static IntrinsicInst *FindInitTrampoline(Value *Callee) { 885 Callee = Callee->stripPointerCasts(); 886 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee); 887 if (!AdjustTramp || 888 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline) 889 return 0; 890 891 Value *TrampMem = AdjustTramp->getOperand(0); 892 893 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem)) 894 return IT; 895 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem)) 896 return IT; 897 return 0; 898} 899 900// visitCallSite - Improvements for call and invoke instructions. 901// 902Instruction *InstCombiner::visitCallSite(CallSite CS) { 903 bool Changed = false; 904 905 // If the callee is a pointer to a function, attempt to move any casts to the 906 // arguments of the call/invoke. 907 Value *Callee = CS.getCalledValue(); 908 if (!isa<Function>(Callee) && transformConstExprCastCall(CS)) 909 return 0; 910 911 if (Function *CalleeF = dyn_cast<Function>(Callee)) 912 // If the call and callee calling conventions don't match, this call must 913 // be unreachable, as the call is undefined. 914 if (CalleeF->getCallingConv() != CS.getCallingConv() && 915 // Only do this for calls to a function with a body. A prototype may 916 // not actually end up matching the implementation's calling conv for a 917 // variety of reasons (e.g. it may be written in assembly). 918 !CalleeF->isDeclaration()) { 919 Instruction *OldCall = CS.getInstruction(); 920 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 921 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 922 OldCall); 923 // If OldCall dues not return void then replaceAllUsesWith undef. 924 // This allows ValueHandlers and custom metadata to adjust itself. 925 if (!OldCall->getType()->isVoidTy()) 926 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType())); 927 if (isa<CallInst>(OldCall)) 928 return EraseInstFromFunction(*OldCall); 929 930 // We cannot remove an invoke, because it would change the CFG, just 931 // change the callee to a null pointer. 932 cast<InvokeInst>(OldCall)->setCalledFunction( 933 Constant::getNullValue(CalleeF->getType())); 934 return 0; 935 } 936 937 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 938 // This instruction is not reachable, just remove it. We insert a store to 939 // undef so that we know that this code is not reachable, despite the fact 940 // that we can't modify the CFG here. 941 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 942 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 943 CS.getInstruction()); 944 945 // If CS does not return void then replaceAllUsesWith undef. 946 // This allows ValueHandlers and custom metadata to adjust itself. 947 if (!CS.getInstruction()->getType()->isVoidTy()) 948 ReplaceInstUsesWith(*CS.getInstruction(), 949 UndefValue::get(CS.getInstruction()->getType())); 950 951 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) { 952 // Don't break the CFG, insert a dummy cond branch. 953 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(), 954 ConstantInt::getTrue(Callee->getContext()), II); 955 } 956 return EraseInstFromFunction(*CS.getInstruction()); 957 } 958 959 if (IntrinsicInst *II = FindInitTrampoline(Callee)) 960 return transformCallThroughTrampoline(CS, II); 961 962 PointerType *PTy = cast<PointerType>(Callee->getType()); 963 FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 964 if (FTy->isVarArg()) { 965 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1); 966 // See if we can optimize any arguments passed through the varargs area of 967 // the call. 968 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(), 969 E = CS.arg_end(); I != E; ++I, ++ix) { 970 CastInst *CI = dyn_cast<CastInst>(*I); 971 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) { 972 *I = CI->getOperand(0); 973 Changed = true; 974 } 975 } 976 } 977 978 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) { 979 // Inline asm calls cannot throw - mark them 'nounwind'. 980 CS.setDoesNotThrow(); 981 Changed = true; 982 } 983 984 // Try to optimize the call if possible, we require TargetData for most of 985 // this. None of these calls are seen as possibly dead so go ahead and 986 // delete the instruction now. 987 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) { 988 Instruction *I = tryOptimizeCall(CI, TD); 989 // If we changed something return the result, etc. Otherwise let 990 // the fallthrough check. 991 if (I) return EraseInstFromFunction(*I); 992 } 993 994 return Changed ? CS.getInstruction() : 0; 995} 996 997// transformConstExprCastCall - If the callee is a constexpr cast of a function, 998// attempt to move the cast to the arguments of the call/invoke. 999// 1000bool InstCombiner::transformConstExprCastCall(CallSite CS) { 1001 Function *Callee = 1002 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts()); 1003 if (Callee == 0) 1004 return false; 1005 Instruction *Caller = CS.getInstruction(); 1006 const AttrListPtr &CallerPAL = CS.getAttributes(); 1007 1008 // Okay, this is a cast from a function to a different type. Unless doing so 1009 // would cause a type conversion of one of our arguments, change this call to 1010 // be a direct call with arguments casted to the appropriate types. 1011 // 1012 FunctionType *FT = Callee->getFunctionType(); 1013 Type *OldRetTy = Caller->getType(); 1014 Type *NewRetTy = FT->getReturnType(); 1015 1016 if (NewRetTy->isStructTy()) 1017 return false; // TODO: Handle multiple return values. 1018 1019 // Check to see if we are changing the return type... 1020 if (OldRetTy != NewRetTy) { 1021 if (Callee->isDeclaration() && 1022 // Conversion is ok if changing from one pointer type to another or from 1023 // a pointer to an integer of the same size. 1024 !((OldRetTy->isPointerTy() || !TD || 1025 OldRetTy == TD->getIntPtrType(Caller->getContext())) && 1026 (NewRetTy->isPointerTy() || !TD || 1027 NewRetTy == TD->getIntPtrType(Caller->getContext())))) 1028 return false; // Cannot transform this return value. 1029 1030 if (!Caller->use_empty() && 1031 // void -> non-void is handled specially 1032 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy)) 1033 return false; // Cannot transform this return value. 1034 1035 if (!CallerPAL.isEmpty() && !Caller->use_empty()) { 1036 Attributes RAttrs = CallerPAL.getRetAttributes(); 1037 if (RAttrs & Attribute::typeIncompatible(NewRetTy)) 1038 return false; // Attribute not compatible with transformed value. 1039 } 1040 1041 // If the callsite is an invoke instruction, and the return value is used by 1042 // a PHI node in a successor, we cannot change the return type of the call 1043 // because there is no place to put the cast instruction (without breaking 1044 // the critical edge). Bail out in this case. 1045 if (!Caller->use_empty()) 1046 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) 1047 for (Value::use_iterator UI = II->use_begin(), E = II->use_end(); 1048 UI != E; ++UI) 1049 if (PHINode *PN = dyn_cast<PHINode>(*UI)) 1050 if (PN->getParent() == II->getNormalDest() || 1051 PN->getParent() == II->getUnwindDest()) 1052 return false; 1053 } 1054 1055 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin()); 1056 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs); 1057 1058 CallSite::arg_iterator AI = CS.arg_begin(); 1059 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) { 1060 Type *ParamTy = FT->getParamType(i); 1061 Type *ActTy = (*AI)->getType(); 1062 1063 if (!CastInst::isCastable(ActTy, ParamTy)) 1064 return false; // Cannot transform this parameter value. 1065 1066 unsigned Attrs = CallerPAL.getParamAttributes(i + 1); 1067 if (Attrs & Attribute::typeIncompatible(ParamTy)) 1068 return false; // Attribute not compatible with transformed value. 1069 1070 // If the parameter is passed as a byval argument, then we have to have a 1071 // sized type and the sized type has to have the same size as the old type. 1072 if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) { 1073 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy); 1074 if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0) 1075 return false; 1076 1077 Type *CurElTy = cast<PointerType>(ActTy)->getElementType(); 1078 if (TD->getTypeAllocSize(CurElTy) != 1079 TD->getTypeAllocSize(ParamPTy->getElementType())) 1080 return false; 1081 } 1082 1083 // Converting from one pointer type to another or between a pointer and an 1084 // integer of the same size is safe even if we do not have a body. 1085 bool isConvertible = ActTy == ParamTy || 1086 (TD && ((ParamTy->isPointerTy() || 1087 ParamTy == TD->getIntPtrType(Caller->getContext())) && 1088 (ActTy->isPointerTy() || 1089 ActTy == TD->getIntPtrType(Caller->getContext())))); 1090 if (Callee->isDeclaration() && !isConvertible) return false; 1091 } 1092 1093 if (Callee->isDeclaration()) { 1094 // Do not delete arguments unless we have a function body. 1095 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg()) 1096 return false; 1097 1098 // If the callee is just a declaration, don't change the varargsness of the 1099 // call. We don't want to introduce a varargs call where one doesn't 1100 // already exist. 1101 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType()); 1102 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg()) 1103 return false; 1104 } 1105 1106 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && 1107 !CallerPAL.isEmpty()) 1108 // In this case we have more arguments than the new function type, but we 1109 // won't be dropping them. Check that these extra arguments have attributes 1110 // that are compatible with being a vararg call argument. 1111 for (unsigned i = CallerPAL.getNumSlots(); i; --i) { 1112 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams()) 1113 break; 1114 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs; 1115 if (PAttrs & Attribute::VarArgsIncompatible) 1116 return false; 1117 } 1118 1119 1120 // Okay, we decided that this is a safe thing to do: go ahead and start 1121 // inserting cast instructions as necessary. 1122 std::vector<Value*> Args; 1123 Args.reserve(NumActualArgs); 1124 SmallVector<AttributeWithIndex, 8> attrVec; 1125 attrVec.reserve(NumCommonArgs); 1126 1127 // Get any return attributes. 1128 Attributes RAttrs = CallerPAL.getRetAttributes(); 1129 1130 // If the return value is not being used, the type may not be compatible 1131 // with the existing attributes. Wipe out any problematic attributes. 1132 RAttrs &= ~Attribute::typeIncompatible(NewRetTy); 1133 1134 // Add the new return attributes. 1135 if (RAttrs) 1136 attrVec.push_back(AttributeWithIndex::get(0, RAttrs)); 1137 1138 AI = CS.arg_begin(); 1139 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) { 1140 Type *ParamTy = FT->getParamType(i); 1141 if ((*AI)->getType() == ParamTy) { 1142 Args.push_back(*AI); 1143 } else { 1144 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, 1145 false, ParamTy, false); 1146 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp")); 1147 } 1148 1149 // Add any parameter attributes. 1150 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1151 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1152 } 1153 1154 // If the function takes more arguments than the call was taking, add them 1155 // now. 1156 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) 1157 Args.push_back(Constant::getNullValue(FT->getParamType(i))); 1158 1159 // If we are removing arguments to the function, emit an obnoxious warning. 1160 if (FT->getNumParams() < NumActualArgs) { 1161 if (!FT->isVarArg()) { 1162 errs() << "WARNING: While resolving call to function '" 1163 << Callee->getName() << "' arguments were dropped!\n"; 1164 } else { 1165 // Add all of the arguments in their promoted form to the arg list. 1166 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) { 1167 Type *PTy = getPromotedType((*AI)->getType()); 1168 if (PTy != (*AI)->getType()) { 1169 // Must promote to pass through va_arg area! 1170 Instruction::CastOps opcode = 1171 CastInst::getCastOpcode(*AI, false, PTy, false); 1172 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp")); 1173 } else { 1174 Args.push_back(*AI); 1175 } 1176 1177 // Add any parameter attributes. 1178 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1179 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1180 } 1181 } 1182 } 1183 1184 if (Attributes FnAttrs = CallerPAL.getFnAttributes()) 1185 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); 1186 1187 if (NewRetTy->isVoidTy()) 1188 Caller->setName(""); // Void type should not have a name. 1189 1190 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(), 1191 attrVec.end()); 1192 1193 Instruction *NC; 1194 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1195 NC = Builder->CreateInvoke(Callee, II->getNormalDest(), 1196 II->getUnwindDest(), Args); 1197 NC->takeName(II); 1198 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv()); 1199 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL); 1200 } else { 1201 CallInst *CI = cast<CallInst>(Caller); 1202 NC = Builder->CreateCall(Callee, Args); 1203 NC->takeName(CI); 1204 if (CI->isTailCall()) 1205 cast<CallInst>(NC)->setTailCall(); 1206 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv()); 1207 cast<CallInst>(NC)->setAttributes(NewCallerPAL); 1208 } 1209 1210 // Insert a cast of the return type as necessary. 1211 Value *NV = NC; 1212 if (OldRetTy != NV->getType() && !Caller->use_empty()) { 1213 if (!NV->getType()->isVoidTy()) { 1214 Instruction::CastOps opcode = 1215 CastInst::getCastOpcode(NC, false, OldRetTy, false); 1216 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp"); 1217 NC->setDebugLoc(Caller->getDebugLoc()); 1218 1219 // If this is an invoke instruction, we should insert it after the first 1220 // non-phi, instruction in the normal successor block. 1221 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1222 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt(); 1223 InsertNewInstBefore(NC, *I); 1224 } else { 1225 // Otherwise, it's a call, just insert cast right after the call. 1226 InsertNewInstBefore(NC, *Caller); 1227 } 1228 Worklist.AddUsersToWorkList(*Caller); 1229 } else { 1230 NV = UndefValue::get(Caller->getType()); 1231 } 1232 } 1233 1234 if (!Caller->use_empty()) 1235 ReplaceInstUsesWith(*Caller, NV); 1236 1237 EraseInstFromFunction(*Caller); 1238 return true; 1239} 1240 1241// transformCallThroughTrampoline - Turn a call to a function created by 1242// init_trampoline / adjust_trampoline intrinsic pair into a direct call to the 1243// underlying function. 1244// 1245Instruction * 1246InstCombiner::transformCallThroughTrampoline(CallSite CS, 1247 IntrinsicInst *Tramp) { 1248 Value *Callee = CS.getCalledValue(); 1249 PointerType *PTy = cast<PointerType>(Callee->getType()); 1250 FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 1251 const AttrListPtr &Attrs = CS.getAttributes(); 1252 1253 // If the call already has the 'nest' attribute somewhere then give up - 1254 // otherwise 'nest' would occur twice after splicing in the chain. 1255 if (Attrs.hasAttrSomewhere(Attribute::Nest)) 1256 return 0; 1257 1258 assert(Tramp && 1259 "transformCallThroughTrampoline called with incorrect CallSite."); 1260 1261 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts()); 1262 PointerType *NestFPTy = cast<PointerType>(NestF->getType()); 1263 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType()); 1264 1265 const AttrListPtr &NestAttrs = NestF->getAttributes(); 1266 if (!NestAttrs.isEmpty()) { 1267 unsigned NestIdx = 1; 1268 Type *NestTy = 0; 1269 Attributes NestAttr = Attribute::None; 1270 1271 // Look for a parameter marked with the 'nest' attribute. 1272 for (FunctionType::param_iterator I = NestFTy->param_begin(), 1273 E = NestFTy->param_end(); I != E; ++NestIdx, ++I) 1274 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) { 1275 // Record the parameter type and any other attributes. 1276 NestTy = *I; 1277 NestAttr = NestAttrs.getParamAttributes(NestIdx); 1278 break; 1279 } 1280 1281 if (NestTy) { 1282 Instruction *Caller = CS.getInstruction(); 1283 std::vector<Value*> NewArgs; 1284 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1); 1285 1286 SmallVector<AttributeWithIndex, 8> NewAttrs; 1287 NewAttrs.reserve(Attrs.getNumSlots() + 1); 1288 1289 // Insert the nest argument into the call argument list, which may 1290 // mean appending it. Likewise for attributes. 1291 1292 // Add any result attributes. 1293 if (Attributes Attr = Attrs.getRetAttributes()) 1294 NewAttrs.push_back(AttributeWithIndex::get(0, Attr)); 1295 1296 { 1297 unsigned Idx = 1; 1298 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 1299 do { 1300 if (Idx == NestIdx) { 1301 // Add the chain argument and attributes. 1302 Value *NestVal = Tramp->getArgOperand(2); 1303 if (NestVal->getType() != NestTy) 1304 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest"); 1305 NewArgs.push_back(NestVal); 1306 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr)); 1307 } 1308 1309 if (I == E) 1310 break; 1311 1312 // Add the original argument and attributes. 1313 NewArgs.push_back(*I); 1314 if (Attributes Attr = Attrs.getParamAttributes(Idx)) 1315 NewAttrs.push_back 1316 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr)); 1317 1318 ++Idx, ++I; 1319 } while (1); 1320 } 1321 1322 // Add any function attributes. 1323 if (Attributes Attr = Attrs.getFnAttributes()) 1324 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr)); 1325 1326 // The trampoline may have been bitcast to a bogus type (FTy). 1327 // Handle this by synthesizing a new function type, equal to FTy 1328 // with the chain parameter inserted. 1329 1330 std::vector<Type*> NewTypes; 1331 NewTypes.reserve(FTy->getNumParams()+1); 1332 1333 // Insert the chain's type into the list of parameter types, which may 1334 // mean appending it. 1335 { 1336 unsigned Idx = 1; 1337 FunctionType::param_iterator I = FTy->param_begin(), 1338 E = FTy->param_end(); 1339 1340 do { 1341 if (Idx == NestIdx) 1342 // Add the chain's type. 1343 NewTypes.push_back(NestTy); 1344 1345 if (I == E) 1346 break; 1347 1348 // Add the original type. 1349 NewTypes.push_back(*I); 1350 1351 ++Idx, ++I; 1352 } while (1); 1353 } 1354 1355 // Replace the trampoline call with a direct call. Let the generic 1356 // code sort out any function type mismatches. 1357 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes, 1358 FTy->isVarArg()); 1359 Constant *NewCallee = 1360 NestF->getType() == PointerType::getUnqual(NewFTy) ? 1361 NestF : ConstantExpr::getBitCast(NestF, 1362 PointerType::getUnqual(NewFTy)); 1363 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(), 1364 NewAttrs.end()); 1365 1366 Instruction *NewCaller; 1367 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1368 NewCaller = InvokeInst::Create(NewCallee, 1369 II->getNormalDest(), II->getUnwindDest(), 1370 NewArgs); 1371 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv()); 1372 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL); 1373 } else { 1374 NewCaller = CallInst::Create(NewCallee, NewArgs); 1375 if (cast<CallInst>(Caller)->isTailCall()) 1376 cast<CallInst>(NewCaller)->setTailCall(); 1377 cast<CallInst>(NewCaller)-> 1378 setCallingConv(cast<CallInst>(Caller)->getCallingConv()); 1379 cast<CallInst>(NewCaller)->setAttributes(NewPAL); 1380 } 1381 1382 return NewCaller; 1383 } 1384 } 1385 1386 // Replace the trampoline call with a direct call. Since there is no 'nest' 1387 // parameter, there is no need to adjust the argument list. Let the generic 1388 // code sort out any function type mismatches. 1389 Constant *NewCallee = 1390 NestF->getType() == PTy ? NestF : 1391 ConstantExpr::getBitCast(NestF, PTy); 1392 CS.setCalledFunction(NewCallee); 1393 return CS.getInstruction(); 1394} 1395 1396