InstCombineCalls.cpp revision 23e75da7e0622528be3c3908b5fe3ae8857cdf65
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 (extractMallocCall(&CI) || extractCallocCall(&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) return 0; 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()) return 0; 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 if (!GEP->getPointerOperandType()->isPointerTy()) 269 return 0; 270 Offset = TD->getIndexedOffset(GEP->getPointerOperandType(), Ops); 271 272 Op1 = GEP->getPointerOperand()->stripPointerCasts(); 273 274 // Make sure we're not a constant offset from an external 275 // global. 276 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) 277 if (!GV->hasDefinitiveInitializer()) return 0; 278 } 279 280 // If we've stripped down to a single global variable that we 281 // can know the size of then just return that. 282 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) { 283 if (GV->hasDefinitiveInitializer()) { 284 Constant *C = GV->getInitializer(); 285 Size = TD->getTypeAllocSize(C->getType()); 286 } else { 287 // Can't determine size of the GV. 288 Constant *RetVal = ConstantInt::get(ReturnTy, DontKnow); 289 return ReplaceInstUsesWith(CI, RetVal); 290 } 291 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) { 292 // Get alloca size. 293 if (AI->getAllocatedType()->isSized()) { 294 Size = TD->getTypeAllocSize(AI->getAllocatedType()); 295 if (AI->isArrayAllocation()) { 296 const ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize()); 297 if (!C) return 0; 298 Size *= C->getZExtValue(); 299 } 300 } 301 } else if (CallInst *MI = extractMallocCall(Op1)) { 302 // Get allocation size. 303 Value *Arg = MI->getArgOperand(0); 304 if (ConstantInt *CI = dyn_cast<ConstantInt>(Arg)) 305 Size = CI->getZExtValue(); 306 307 } else if (CallInst *MI = extractCallocCall(Op1)) { 308 // Get allocation size. 309 Value *Arg1 = MI->getArgOperand(0); 310 Value *Arg2 = MI->getArgOperand(1); 311 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(Arg1)) 312 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Arg2)) 313 Size = (CI1->getValue() * CI2->getValue()).getZExtValue(); 314 } 315 316 // Do not return "I don't know" here. Later optimization passes could 317 // make it possible to evaluate objectsize to a constant. 318 if (Size == -1ULL) 319 return 0; 320 321 if (Size < Offset) { 322 // Out of bound reference? Negative index normalized to large 323 // index? Just return "I don't know". 324 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, DontKnow)); 325 } 326 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, Size-Offset)); 327 } 328 case Intrinsic::bswap: 329 // bswap(bswap(x)) -> x 330 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) 331 if (Operand->getIntrinsicID() == Intrinsic::bswap) 332 return ReplaceInstUsesWith(CI, Operand->getArgOperand(0)); 333 334 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c)) 335 if (TruncInst *TI = dyn_cast<TruncInst>(II->getArgOperand(0))) { 336 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0))) 337 if (Operand->getIntrinsicID() == Intrinsic::bswap) { 338 unsigned C = Operand->getType()->getPrimitiveSizeInBits() - 339 TI->getType()->getPrimitiveSizeInBits(); 340 Value *CV = ConstantInt::get(Operand->getType(), C); 341 Value *V = Builder->CreateLShr(Operand->getArgOperand(0), CV); 342 return new TruncInst(V, TI->getType()); 343 } 344 } 345 346 break; 347 case Intrinsic::powi: 348 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 349 // powi(x, 0) -> 1.0 350 if (Power->isZero()) 351 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0)); 352 // powi(x, 1) -> x 353 if (Power->isOne()) 354 return ReplaceInstUsesWith(CI, II->getArgOperand(0)); 355 // powi(x, -1) -> 1/x 356 if (Power->isAllOnesValue()) 357 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0), 358 II->getArgOperand(0)); 359 } 360 break; 361 case Intrinsic::cttz: { 362 // If all bits below the first known one are known zero, 363 // this value is constant. 364 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType()); 365 // FIXME: Try to simplify vectors of integers. 366 if (!IT) break; 367 uint32_t BitWidth = IT->getBitWidth(); 368 APInt KnownZero(BitWidth, 0); 369 APInt KnownOne(BitWidth, 0); 370 ComputeMaskedBits(II->getArgOperand(0), KnownZero, KnownOne); 371 unsigned TrailingZeros = KnownOne.countTrailingZeros(); 372 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros)); 373 if ((Mask & KnownZero) == Mask) 374 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 375 APInt(BitWidth, TrailingZeros))); 376 377 } 378 break; 379 case Intrinsic::ctlz: { 380 // If all bits above the first known one are known zero, 381 // this value is constant. 382 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType()); 383 // FIXME: Try to simplify vectors of integers. 384 if (!IT) break; 385 uint32_t BitWidth = IT->getBitWidth(); 386 APInt KnownZero(BitWidth, 0); 387 APInt KnownOne(BitWidth, 0); 388 ComputeMaskedBits(II->getArgOperand(0), KnownZero, KnownOne); 389 unsigned LeadingZeros = KnownOne.countLeadingZeros(); 390 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros)); 391 if ((Mask & KnownZero) == Mask) 392 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 393 APInt(BitWidth, LeadingZeros))); 394 395 } 396 break; 397 case Intrinsic::uadd_with_overflow: { 398 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 399 IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType()); 400 uint32_t BitWidth = IT->getBitWidth(); 401 APInt LHSKnownZero(BitWidth, 0); 402 APInt LHSKnownOne(BitWidth, 0); 403 ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne); 404 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1]; 405 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1]; 406 407 if (LHSKnownNegative || LHSKnownPositive) { 408 APInt RHSKnownZero(BitWidth, 0); 409 APInt RHSKnownOne(BitWidth, 0); 410 ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne); 411 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1]; 412 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1]; 413 if (LHSKnownNegative && RHSKnownNegative) { 414 // The sign bit is set in both cases: this MUST overflow. 415 // Create a simple add instruction, and insert it into the struct. 416 Value *Add = Builder->CreateAdd(LHS, RHS); 417 Add->takeName(&CI); 418 Constant *V[] = { 419 UndefValue::get(LHS->getType()), 420 ConstantInt::getTrue(II->getContext()) 421 }; 422 StructType *ST = cast<StructType>(II->getType()); 423 Constant *Struct = ConstantStruct::get(ST, V); 424 return InsertValueInst::Create(Struct, Add, 0); 425 } 426 427 if (LHSKnownPositive && RHSKnownPositive) { 428 // The sign bit is clear in both cases: this CANNOT overflow. 429 // Create a simple add instruction, and insert it into the struct. 430 Value *Add = Builder->CreateNUWAdd(LHS, RHS); 431 Add->takeName(&CI); 432 Constant *V[] = { 433 UndefValue::get(LHS->getType()), 434 ConstantInt::getFalse(II->getContext()) 435 }; 436 StructType *ST = cast<StructType>(II->getType()); 437 Constant *Struct = ConstantStruct::get(ST, V); 438 return InsertValueInst::Create(Struct, Add, 0); 439 } 440 } 441 } 442 // FALL THROUGH uadd into sadd 443 case Intrinsic::sadd_with_overflow: 444 // Canonicalize constants into the RHS. 445 if (isa<Constant>(II->getArgOperand(0)) && 446 !isa<Constant>(II->getArgOperand(1))) { 447 Value *LHS = II->getArgOperand(0); 448 II->setArgOperand(0, II->getArgOperand(1)); 449 II->setArgOperand(1, LHS); 450 return II; 451 } 452 453 // X + undef -> undef 454 if (isa<UndefValue>(II->getArgOperand(1))) 455 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 456 457 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 458 // X + 0 -> {X, false} 459 if (RHS->isZero()) { 460 Constant *V[] = { 461 UndefValue::get(II->getArgOperand(0)->getType()), 462 ConstantInt::getFalse(II->getContext()) 463 }; 464 Constant *Struct = 465 ConstantStruct::get(cast<StructType>(II->getType()), V); 466 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 467 } 468 } 469 break; 470 case Intrinsic::usub_with_overflow: 471 case Intrinsic::ssub_with_overflow: 472 // undef - X -> undef 473 // X - undef -> undef 474 if (isa<UndefValue>(II->getArgOperand(0)) || 475 isa<UndefValue>(II->getArgOperand(1))) 476 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 477 478 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 479 // X - 0 -> {X, false} 480 if (RHS->isZero()) { 481 Constant *V[] = { 482 UndefValue::get(II->getArgOperand(0)->getType()), 483 ConstantInt::getFalse(II->getContext()) 484 }; 485 Constant *Struct = 486 ConstantStruct::get(cast<StructType>(II->getType()), V); 487 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 488 } 489 } 490 break; 491 case Intrinsic::umul_with_overflow: { 492 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 493 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth(); 494 495 APInt LHSKnownZero(BitWidth, 0); 496 APInt LHSKnownOne(BitWidth, 0); 497 ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne); 498 APInt RHSKnownZero(BitWidth, 0); 499 APInt RHSKnownOne(BitWidth, 0); 500 ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne); 501 502 // Get the largest possible values for each operand. 503 APInt LHSMax = ~LHSKnownZero; 504 APInt RHSMax = ~RHSKnownZero; 505 506 // If multiplying the maximum values does not overflow then we can turn 507 // this into a plain NUW mul. 508 bool Overflow; 509 LHSMax.umul_ov(RHSMax, Overflow); 510 if (!Overflow) { 511 Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow"); 512 Constant *V[] = { 513 UndefValue::get(LHS->getType()), 514 Builder->getFalse() 515 }; 516 Constant *Struct = ConstantStruct::get(cast<StructType>(II->getType()),V); 517 return InsertValueInst::Create(Struct, Mul, 0); 518 } 519 } // FALL THROUGH 520 case Intrinsic::smul_with_overflow: 521 // Canonicalize constants into the RHS. 522 if (isa<Constant>(II->getArgOperand(0)) && 523 !isa<Constant>(II->getArgOperand(1))) { 524 Value *LHS = II->getArgOperand(0); 525 II->setArgOperand(0, II->getArgOperand(1)); 526 II->setArgOperand(1, LHS); 527 return II; 528 } 529 530 // X * undef -> undef 531 if (isa<UndefValue>(II->getArgOperand(1))) 532 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 533 534 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 535 // X*0 -> {0, false} 536 if (RHSI->isZero()) 537 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType())); 538 539 // X * 1 -> {X, false} 540 if (RHSI->equalsInt(1)) { 541 Constant *V[] = { 542 UndefValue::get(II->getArgOperand(0)->getType()), 543 ConstantInt::getFalse(II->getContext()) 544 }; 545 Constant *Struct = 546 ConstantStruct::get(cast<StructType>(II->getType()), V); 547 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 548 } 549 } 550 break; 551 case Intrinsic::ppc_altivec_lvx: 552 case Intrinsic::ppc_altivec_lvxl: 553 // Turn PPC lvx -> load if the pointer is known aligned. 554 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 555 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), 556 PointerType::getUnqual(II->getType())); 557 return new LoadInst(Ptr); 558 } 559 break; 560 case Intrinsic::ppc_altivec_stvx: 561 case Intrinsic::ppc_altivec_stvxl: 562 // Turn stvx -> store if the pointer is known aligned. 563 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) { 564 Type *OpPtrTy = 565 PointerType::getUnqual(II->getArgOperand(0)->getType()); 566 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); 567 return new StoreInst(II->getArgOperand(0), Ptr); 568 } 569 break; 570 case Intrinsic::x86_sse_storeu_ps: 571 case Intrinsic::x86_sse2_storeu_pd: 572 case Intrinsic::x86_sse2_storeu_dq: 573 // Turn X86 storeu -> store if the pointer is known aligned. 574 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 575 Type *OpPtrTy = 576 PointerType::getUnqual(II->getArgOperand(1)->getType()); 577 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy); 578 return new StoreInst(II->getArgOperand(1), Ptr); 579 } 580 break; 581 582 case Intrinsic::x86_sse_cvtss2si: 583 case Intrinsic::x86_sse_cvtss2si64: 584 case Intrinsic::x86_sse_cvttss2si: 585 case Intrinsic::x86_sse_cvttss2si64: 586 case Intrinsic::x86_sse2_cvtsd2si: 587 case Intrinsic::x86_sse2_cvtsd2si64: 588 case Intrinsic::x86_sse2_cvttsd2si: 589 case Intrinsic::x86_sse2_cvttsd2si64: { 590 // These intrinsics only demand the 0th element of their input vectors. If 591 // we can simplify the input based on that, do so now. 592 unsigned VWidth = 593 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 594 APInt DemandedElts(VWidth, 1); 595 APInt UndefElts(VWidth, 0); 596 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0), 597 DemandedElts, UndefElts)) { 598 II->setArgOperand(0, V); 599 return II; 600 } 601 break; 602 } 603 604 605 case Intrinsic::x86_sse41_pmovsxbw: 606 case Intrinsic::x86_sse41_pmovsxwd: 607 case Intrinsic::x86_sse41_pmovsxdq: 608 case Intrinsic::x86_sse41_pmovzxbw: 609 case Intrinsic::x86_sse41_pmovzxwd: 610 case Intrinsic::x86_sse41_pmovzxdq: { 611 // pmov{s|z}x ignores the upper half of their input vectors. 612 unsigned VWidth = 613 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 614 unsigned LowHalfElts = VWidth / 2; 615 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts)); 616 APInt UndefElts(VWidth, 0); 617 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), 618 InputDemandedElts, 619 UndefElts)) { 620 II->setArgOperand(0, TmpV); 621 return II; 622 } 623 break; 624 } 625 626 case Intrinsic::ppc_altivec_vperm: 627 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant. 628 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) { 629 assert(Mask->getType()->getVectorNumElements() == 16 && 630 "Bad type for intrinsic!"); 631 632 // Check that all of the elements are integer constants or undefs. 633 bool AllEltsOk = true; 634 for (unsigned i = 0; i != 16; ++i) { 635 Constant *Elt = Mask->getAggregateElement(i); 636 if (Elt == 0 || 637 !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) { 638 AllEltsOk = false; 639 break; 640 } 641 } 642 643 if (AllEltsOk) { 644 // Cast the input vectors to byte vectors. 645 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0), 646 Mask->getType()); 647 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1), 648 Mask->getType()); 649 Value *Result = UndefValue::get(Op0->getType()); 650 651 // Only extract each element once. 652 Value *ExtractedElts[32]; 653 memset(ExtractedElts, 0, sizeof(ExtractedElts)); 654 655 for (unsigned i = 0; i != 16; ++i) { 656 if (isa<UndefValue>(Mask->getAggregateElement(i))) 657 continue; 658 unsigned Idx = 659 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue(); 660 Idx &= 31; // Match the hardware behavior. 661 662 if (ExtractedElts[Idx] == 0) { 663 ExtractedElts[Idx] = 664 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1, 665 Builder->getInt32(Idx&15)); 666 } 667 668 // Insert this value into the result vector. 669 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx], 670 Builder->getInt32(i)); 671 } 672 return CastInst::Create(Instruction::BitCast, Result, CI.getType()); 673 } 674 } 675 break; 676 677 case Intrinsic::arm_neon_vld1: 678 case Intrinsic::arm_neon_vld2: 679 case Intrinsic::arm_neon_vld3: 680 case Intrinsic::arm_neon_vld4: 681 case Intrinsic::arm_neon_vld2lane: 682 case Intrinsic::arm_neon_vld3lane: 683 case Intrinsic::arm_neon_vld4lane: 684 case Intrinsic::arm_neon_vst1: 685 case Intrinsic::arm_neon_vst2: 686 case Intrinsic::arm_neon_vst3: 687 case Intrinsic::arm_neon_vst4: 688 case Intrinsic::arm_neon_vst2lane: 689 case Intrinsic::arm_neon_vst3lane: 690 case Intrinsic::arm_neon_vst4lane: { 691 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD); 692 unsigned AlignArg = II->getNumArgOperands() - 1; 693 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg)); 694 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) { 695 II->setArgOperand(AlignArg, 696 ConstantInt::get(Type::getInt32Ty(II->getContext()), 697 MemAlign, false)); 698 return II; 699 } 700 break; 701 } 702 703 case Intrinsic::arm_neon_vmulls: 704 case Intrinsic::arm_neon_vmullu: { 705 Value *Arg0 = II->getArgOperand(0); 706 Value *Arg1 = II->getArgOperand(1); 707 708 // Handle mul by zero first: 709 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) { 710 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType())); 711 } 712 713 // Check for constant LHS & RHS - in this case we just simplify. 714 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu); 715 VectorType *NewVT = cast<VectorType>(II->getType()); 716 unsigned NewWidth = NewVT->getElementType()->getIntegerBitWidth(); 717 if (ConstantDataVector *CV0 = dyn_cast<ConstantDataVector>(Arg0)) { 718 if (ConstantDataVector *CV1 = dyn_cast<ConstantDataVector>(Arg1)) { 719 VectorType* VT = cast<VectorType>(CV0->getType()); 720 SmallVector<Constant*, 4> NewElems; 721 for (unsigned i = 0; i < VT->getNumElements(); ++i) { 722 APInt CV0E = 723 (cast<ConstantInt>(CV0->getAggregateElement(i)))->getValue(); 724 CV0E = Zext ? CV0E.zext(NewWidth) : CV0E.sext(NewWidth); 725 APInt CV1E = 726 (cast<ConstantInt>(CV1->getAggregateElement(i)))->getValue(); 727 CV1E = Zext ? CV1E.zext(NewWidth) : CV1E.sext(NewWidth); 728 NewElems.push_back( 729 ConstantInt::get(NewVT->getElementType(), CV0E * CV1E)); 730 } 731 return ReplaceInstUsesWith(CI, ConstantVector::get(NewElems)); 732 } 733 734 // Couldn't simplify - cannonicalize constant to the RHS. 735 std::swap(Arg0, Arg1); 736 } 737 738 // Handle mul by one: 739 if (ConstantDataVector *CV1 = dyn_cast<ConstantDataVector>(Arg1)) { 740 if (ConstantInt *Splat = 741 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue())) { 742 if (Splat->isOne()) { 743 if (Zext) 744 return CastInst::CreateZExtOrBitCast(Arg0, II->getType()); 745 // else 746 return CastInst::CreateSExtOrBitCast(Arg0, II->getType()); 747 } 748 } 749 } 750 751 break; 752 } 753 754 case Intrinsic::stackrestore: { 755 // If the save is right next to the restore, remove the restore. This can 756 // happen when variable allocas are DCE'd. 757 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) { 758 if (SS->getIntrinsicID() == Intrinsic::stacksave) { 759 BasicBlock::iterator BI = SS; 760 if (&*++BI == II) 761 return EraseInstFromFunction(CI); 762 } 763 } 764 765 // Scan down this block to see if there is another stack restore in the 766 // same block without an intervening call/alloca. 767 BasicBlock::iterator BI = II; 768 TerminatorInst *TI = II->getParent()->getTerminator(); 769 bool CannotRemove = false; 770 for (++BI; &*BI != TI; ++BI) { 771 if (isa<AllocaInst>(BI) || isMalloc(BI)) { 772 CannotRemove = true; 773 break; 774 } 775 if (CallInst *BCI = dyn_cast<CallInst>(BI)) { 776 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) { 777 // If there is a stackrestore below this one, remove this one. 778 if (II->getIntrinsicID() == Intrinsic::stackrestore) 779 return EraseInstFromFunction(CI); 780 // Otherwise, ignore the intrinsic. 781 } else { 782 // If we found a non-intrinsic call, we can't remove the stack 783 // restore. 784 CannotRemove = true; 785 break; 786 } 787 } 788 } 789 790 // If the stack restore is in a return, resume, or unwind block and if there 791 // are no allocas or calls between the restore and the return, nuke the 792 // restore. 793 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI))) 794 return EraseInstFromFunction(CI); 795 break; 796 } 797 } 798 799 return visitCallSite(II); 800} 801 802// InvokeInst simplification 803// 804Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) { 805 return visitCallSite(&II); 806} 807 808/// isSafeToEliminateVarargsCast - If this cast does not affect the value 809/// passed through the varargs area, we can eliminate the use of the cast. 810static bool isSafeToEliminateVarargsCast(const CallSite CS, 811 const CastInst * const CI, 812 const TargetData * const TD, 813 const int ix) { 814 if (!CI->isLosslessCast()) 815 return false; 816 817 // The size of ByVal arguments is derived from the type, so we 818 // can't change to a type with a different size. If the size were 819 // passed explicitly we could avoid this check. 820 if (!CS.isByValArgument(ix)) 821 return true; 822 823 Type* SrcTy = 824 cast<PointerType>(CI->getOperand(0)->getType())->getElementType(); 825 Type* DstTy = cast<PointerType>(CI->getType())->getElementType(); 826 if (!SrcTy->isSized() || !DstTy->isSized()) 827 return false; 828 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy)) 829 return false; 830 return true; 831} 832 833namespace { 834class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls { 835 InstCombiner *IC; 836protected: 837 void replaceCall(Value *With) { 838 NewInstruction = IC->ReplaceInstUsesWith(*CI, With); 839 } 840 bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const { 841 if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp)) 842 return true; 843 if (ConstantInt *SizeCI = 844 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) { 845 if (SizeCI->isAllOnesValue()) 846 return true; 847 if (isString) { 848 uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp)); 849 // If the length is 0 we don't know how long it is and so we can't 850 // remove the check. 851 if (Len == 0) return false; 852 return SizeCI->getZExtValue() >= Len; 853 } 854 if (ConstantInt *Arg = dyn_cast<ConstantInt>( 855 CI->getArgOperand(SizeArgOp))) 856 return SizeCI->getZExtValue() >= Arg->getZExtValue(); 857 } 858 return false; 859 } 860public: 861 InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { } 862 Instruction *NewInstruction; 863}; 864} // end anonymous namespace 865 866// Try to fold some different type of calls here. 867// Currently we're only working with the checking functions, memcpy_chk, 868// mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk, 869// strcat_chk and strncat_chk. 870Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) { 871 if (CI->getCalledFunction() == 0) return 0; 872 873 InstCombineFortifiedLibCalls Simplifier(this); 874 Simplifier.fold(CI, TD); 875 return Simplifier.NewInstruction; 876} 877 878static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) { 879 // Strip off at most one level of pointer casts, looking for an alloca. This 880 // is good enough in practice and simpler than handling any number of casts. 881 Value *Underlying = TrampMem->stripPointerCasts(); 882 if (Underlying != TrampMem && 883 (!Underlying->hasOneUse() || *Underlying->use_begin() != TrampMem)) 884 return 0; 885 if (!isa<AllocaInst>(Underlying)) 886 return 0; 887 888 IntrinsicInst *InitTrampoline = 0; 889 for (Value::use_iterator I = TrampMem->use_begin(), E = TrampMem->use_end(); 890 I != E; I++) { 891 IntrinsicInst *II = dyn_cast<IntrinsicInst>(*I); 892 if (!II) 893 return 0; 894 if (II->getIntrinsicID() == Intrinsic::init_trampoline) { 895 if (InitTrampoline) 896 // More than one init_trampoline writes to this value. Give up. 897 return 0; 898 InitTrampoline = II; 899 continue; 900 } 901 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline) 902 // Allow any number of calls to adjust.trampoline. 903 continue; 904 return 0; 905 } 906 907 // No call to init.trampoline found. 908 if (!InitTrampoline) 909 return 0; 910 911 // Check that the alloca is being used in the expected way. 912 if (InitTrampoline->getOperand(0) != TrampMem) 913 return 0; 914 915 return InitTrampoline; 916} 917 918static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp, 919 Value *TrampMem) { 920 // Visit all the previous instructions in the basic block, and try to find a 921 // init.trampoline which has a direct path to the adjust.trampoline. 922 for (BasicBlock::iterator I = AdjustTramp, 923 E = AdjustTramp->getParent()->begin(); I != E; ) { 924 Instruction *Inst = --I; 925 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) 926 if (II->getIntrinsicID() == Intrinsic::init_trampoline && 927 II->getOperand(0) == TrampMem) 928 return II; 929 if (Inst->mayWriteToMemory()) 930 return 0; 931 } 932 return 0; 933} 934 935// Given a call to llvm.adjust.trampoline, find and return the corresponding 936// call to llvm.init.trampoline if the call to the trampoline can be optimized 937// to a direct call to a function. Otherwise return NULL. 938// 939static IntrinsicInst *FindInitTrampoline(Value *Callee) { 940 Callee = Callee->stripPointerCasts(); 941 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee); 942 if (!AdjustTramp || 943 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline) 944 return 0; 945 946 Value *TrampMem = AdjustTramp->getOperand(0); 947 948 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem)) 949 return IT; 950 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem)) 951 return IT; 952 return 0; 953} 954 955// visitCallSite - Improvements for call and invoke instructions. 956// 957Instruction *InstCombiner::visitCallSite(CallSite CS) { 958 bool Changed = false; 959 960 // If the callee is a pointer to a function, attempt to move any casts to the 961 // arguments of the call/invoke. 962 Value *Callee = CS.getCalledValue(); 963 if (!isa<Function>(Callee) && transformConstExprCastCall(CS)) 964 return 0; 965 966 if (Function *CalleeF = dyn_cast<Function>(Callee)) 967 // If the call and callee calling conventions don't match, this call must 968 // be unreachable, as the call is undefined. 969 if (CalleeF->getCallingConv() != CS.getCallingConv() && 970 // Only do this for calls to a function with a body. A prototype may 971 // not actually end up matching the implementation's calling conv for a 972 // variety of reasons (e.g. it may be written in assembly). 973 !CalleeF->isDeclaration()) { 974 Instruction *OldCall = CS.getInstruction(); 975 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 976 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 977 OldCall); 978 // If OldCall dues not return void then replaceAllUsesWith undef. 979 // This allows ValueHandlers and custom metadata to adjust itself. 980 if (!OldCall->getType()->isVoidTy()) 981 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType())); 982 if (isa<CallInst>(OldCall)) 983 return EraseInstFromFunction(*OldCall); 984 985 // We cannot remove an invoke, because it would change the CFG, just 986 // change the callee to a null pointer. 987 cast<InvokeInst>(OldCall)->setCalledFunction( 988 Constant::getNullValue(CalleeF->getType())); 989 return 0; 990 } 991 992 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 993 // This instruction is not reachable, just remove it. We insert a store to 994 // undef so that we know that this code is not reachable, despite the fact 995 // that we can't modify the CFG here. 996 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 997 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 998 CS.getInstruction()); 999 1000 // If CS does not return void then replaceAllUsesWith undef. 1001 // This allows ValueHandlers and custom metadata to adjust itself. 1002 if (!CS.getInstruction()->getType()->isVoidTy()) 1003 ReplaceInstUsesWith(*CS.getInstruction(), 1004 UndefValue::get(CS.getInstruction()->getType())); 1005 1006 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) { 1007 // Don't break the CFG, insert a dummy cond branch. 1008 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(), 1009 ConstantInt::getTrue(Callee->getContext()), II); 1010 } 1011 return EraseInstFromFunction(*CS.getInstruction()); 1012 } 1013 1014 if (IntrinsicInst *II = FindInitTrampoline(Callee)) 1015 return transformCallThroughTrampoline(CS, II); 1016 1017 PointerType *PTy = cast<PointerType>(Callee->getType()); 1018 FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 1019 if (FTy->isVarArg()) { 1020 int ix = FTy->getNumParams(); 1021 // See if we can optimize any arguments passed through the varargs area of 1022 // the call. 1023 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(), 1024 E = CS.arg_end(); I != E; ++I, ++ix) { 1025 CastInst *CI = dyn_cast<CastInst>(*I); 1026 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) { 1027 *I = CI->getOperand(0); 1028 Changed = true; 1029 } 1030 } 1031 } 1032 1033 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) { 1034 // Inline asm calls cannot throw - mark them 'nounwind'. 1035 CS.setDoesNotThrow(); 1036 Changed = true; 1037 } 1038 1039 // Try to optimize the call if possible, we require TargetData for most of 1040 // this. None of these calls are seen as possibly dead so go ahead and 1041 // delete the instruction now. 1042 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) { 1043 Instruction *I = tryOptimizeCall(CI, TD); 1044 // If we changed something return the result, etc. Otherwise let 1045 // the fallthrough check. 1046 if (I) return EraseInstFromFunction(*I); 1047 } 1048 1049 return Changed ? CS.getInstruction() : 0; 1050} 1051 1052// transformConstExprCastCall - If the callee is a constexpr cast of a function, 1053// attempt to move the cast to the arguments of the call/invoke. 1054// 1055bool InstCombiner::transformConstExprCastCall(CallSite CS) { 1056 Function *Callee = 1057 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts()); 1058 if (Callee == 0) 1059 return false; 1060 Instruction *Caller = CS.getInstruction(); 1061 const AttrListPtr &CallerPAL = CS.getAttributes(); 1062 1063 // Okay, this is a cast from a function to a different type. Unless doing so 1064 // would cause a type conversion of one of our arguments, change this call to 1065 // be a direct call with arguments casted to the appropriate types. 1066 // 1067 FunctionType *FT = Callee->getFunctionType(); 1068 Type *OldRetTy = Caller->getType(); 1069 Type *NewRetTy = FT->getReturnType(); 1070 1071 if (NewRetTy->isStructTy()) 1072 return false; // TODO: Handle multiple return values. 1073 1074 // Check to see if we are changing the return type... 1075 if (OldRetTy != NewRetTy) { 1076 if (Callee->isDeclaration() && 1077 // Conversion is ok if changing from one pointer type to another or from 1078 // a pointer to an integer of the same size. 1079 !((OldRetTy->isPointerTy() || !TD || 1080 OldRetTy == TD->getIntPtrType(Caller->getContext())) && 1081 (NewRetTy->isPointerTy() || !TD || 1082 NewRetTy == TD->getIntPtrType(Caller->getContext())))) 1083 return false; // Cannot transform this return value. 1084 1085 if (!Caller->use_empty() && 1086 // void -> non-void is handled specially 1087 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy)) 1088 return false; // Cannot transform this return value. 1089 1090 if (!CallerPAL.isEmpty() && !Caller->use_empty()) { 1091 Attributes RAttrs = CallerPAL.getRetAttributes(); 1092 if (RAttrs & Attribute::typeIncompatible(NewRetTy)) 1093 return false; // Attribute not compatible with transformed value. 1094 } 1095 1096 // If the callsite is an invoke instruction, and the return value is used by 1097 // a PHI node in a successor, we cannot change the return type of the call 1098 // because there is no place to put the cast instruction (without breaking 1099 // the critical edge). Bail out in this case. 1100 if (!Caller->use_empty()) 1101 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) 1102 for (Value::use_iterator UI = II->use_begin(), E = II->use_end(); 1103 UI != E; ++UI) 1104 if (PHINode *PN = dyn_cast<PHINode>(*UI)) 1105 if (PN->getParent() == II->getNormalDest() || 1106 PN->getParent() == II->getUnwindDest()) 1107 return false; 1108 } 1109 1110 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin()); 1111 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs); 1112 1113 CallSite::arg_iterator AI = CS.arg_begin(); 1114 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) { 1115 Type *ParamTy = FT->getParamType(i); 1116 Type *ActTy = (*AI)->getType(); 1117 1118 if (!CastInst::isCastable(ActTy, ParamTy)) 1119 return false; // Cannot transform this parameter value. 1120 1121 Attributes Attrs = CallerPAL.getParamAttributes(i + 1); 1122 if (Attrs & Attribute::typeIncompatible(ParamTy)) 1123 return false; // Attribute not compatible with transformed value. 1124 1125 // If the parameter is passed as a byval argument, then we have to have a 1126 // sized type and the sized type has to have the same size as the old type. 1127 if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) { 1128 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy); 1129 if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0) 1130 return false; 1131 1132 Type *CurElTy = cast<PointerType>(ActTy)->getElementType(); 1133 if (TD->getTypeAllocSize(CurElTy) != 1134 TD->getTypeAllocSize(ParamPTy->getElementType())) 1135 return false; 1136 } 1137 1138 // Converting from one pointer type to another or between a pointer and an 1139 // integer of the same size is safe even if we do not have a body. 1140 bool isConvertible = ActTy == ParamTy || 1141 (TD && ((ParamTy->isPointerTy() || 1142 ParamTy == TD->getIntPtrType(Caller->getContext())) && 1143 (ActTy->isPointerTy() || 1144 ActTy == TD->getIntPtrType(Caller->getContext())))); 1145 if (Callee->isDeclaration() && !isConvertible) return false; 1146 } 1147 1148 if (Callee->isDeclaration()) { 1149 // Do not delete arguments unless we have a function body. 1150 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg()) 1151 return false; 1152 1153 // If the callee is just a declaration, don't change the varargsness of the 1154 // call. We don't want to introduce a varargs call where one doesn't 1155 // already exist. 1156 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType()); 1157 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg()) 1158 return false; 1159 1160 // If both the callee and the cast type are varargs, we still have to make 1161 // sure the number of fixed parameters are the same or we have the same 1162 // ABI issues as if we introduce a varargs call. 1163 if (FT->isVarArg() && 1164 cast<FunctionType>(APTy->getElementType())->isVarArg() && 1165 FT->getNumParams() != 1166 cast<FunctionType>(APTy->getElementType())->getNumParams()) 1167 return false; 1168 } 1169 1170 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && 1171 !CallerPAL.isEmpty()) 1172 // In this case we have more arguments than the new function type, but we 1173 // won't be dropping them. Check that these extra arguments have attributes 1174 // that are compatible with being a vararg call argument. 1175 for (unsigned i = CallerPAL.getNumSlots(); i; --i) { 1176 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams()) 1177 break; 1178 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs; 1179 if (PAttrs & Attribute::VarArgsIncompatible) 1180 return false; 1181 } 1182 1183 1184 // Okay, we decided that this is a safe thing to do: go ahead and start 1185 // inserting cast instructions as necessary. 1186 std::vector<Value*> Args; 1187 Args.reserve(NumActualArgs); 1188 SmallVector<AttributeWithIndex, 8> attrVec; 1189 attrVec.reserve(NumCommonArgs); 1190 1191 // Get any return attributes. 1192 Attributes RAttrs = CallerPAL.getRetAttributes(); 1193 1194 // If the return value is not being used, the type may not be compatible 1195 // with the existing attributes. Wipe out any problematic attributes. 1196 RAttrs &= ~Attribute::typeIncompatible(NewRetTy); 1197 1198 // Add the new return attributes. 1199 if (RAttrs) 1200 attrVec.push_back(AttributeWithIndex::get(0, RAttrs)); 1201 1202 AI = CS.arg_begin(); 1203 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) { 1204 Type *ParamTy = FT->getParamType(i); 1205 if ((*AI)->getType() == ParamTy) { 1206 Args.push_back(*AI); 1207 } else { 1208 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, 1209 false, ParamTy, false); 1210 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy)); 1211 } 1212 1213 // Add any parameter attributes. 1214 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1215 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1216 } 1217 1218 // If the function takes more arguments than the call was taking, add them 1219 // now. 1220 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) 1221 Args.push_back(Constant::getNullValue(FT->getParamType(i))); 1222 1223 // If we are removing arguments to the function, emit an obnoxious warning. 1224 if (FT->getNumParams() < NumActualArgs) { 1225 if (!FT->isVarArg()) { 1226 errs() << "WARNING: While resolving call to function '" 1227 << Callee->getName() << "' arguments were dropped!\n"; 1228 } else { 1229 // Add all of the arguments in their promoted form to the arg list. 1230 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) { 1231 Type *PTy = getPromotedType((*AI)->getType()); 1232 if (PTy != (*AI)->getType()) { 1233 // Must promote to pass through va_arg area! 1234 Instruction::CastOps opcode = 1235 CastInst::getCastOpcode(*AI, false, PTy, false); 1236 Args.push_back(Builder->CreateCast(opcode, *AI, PTy)); 1237 } else { 1238 Args.push_back(*AI); 1239 } 1240 1241 // Add any parameter attributes. 1242 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1243 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1244 } 1245 } 1246 } 1247 1248 if (Attributes FnAttrs = CallerPAL.getFnAttributes()) 1249 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); 1250 1251 if (NewRetTy->isVoidTy()) 1252 Caller->setName(""); // Void type should not have a name. 1253 1254 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(), 1255 attrVec.end()); 1256 1257 Instruction *NC; 1258 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1259 NC = Builder->CreateInvoke(Callee, II->getNormalDest(), 1260 II->getUnwindDest(), Args); 1261 NC->takeName(II); 1262 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv()); 1263 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL); 1264 } else { 1265 CallInst *CI = cast<CallInst>(Caller); 1266 NC = Builder->CreateCall(Callee, Args); 1267 NC->takeName(CI); 1268 if (CI->isTailCall()) 1269 cast<CallInst>(NC)->setTailCall(); 1270 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv()); 1271 cast<CallInst>(NC)->setAttributes(NewCallerPAL); 1272 } 1273 1274 // Insert a cast of the return type as necessary. 1275 Value *NV = NC; 1276 if (OldRetTy != NV->getType() && !Caller->use_empty()) { 1277 if (!NV->getType()->isVoidTy()) { 1278 Instruction::CastOps opcode = 1279 CastInst::getCastOpcode(NC, false, OldRetTy, false); 1280 NV = NC = CastInst::Create(opcode, NC, OldRetTy); 1281 NC->setDebugLoc(Caller->getDebugLoc()); 1282 1283 // If this is an invoke instruction, we should insert it after the first 1284 // non-phi, instruction in the normal successor block. 1285 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1286 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt(); 1287 InsertNewInstBefore(NC, *I); 1288 } else { 1289 // Otherwise, it's a call, just insert cast right after the call. 1290 InsertNewInstBefore(NC, *Caller); 1291 } 1292 Worklist.AddUsersToWorkList(*Caller); 1293 } else { 1294 NV = UndefValue::get(Caller->getType()); 1295 } 1296 } 1297 1298 if (!Caller->use_empty()) 1299 ReplaceInstUsesWith(*Caller, NV); 1300 1301 EraseInstFromFunction(*Caller); 1302 return true; 1303} 1304 1305// transformCallThroughTrampoline - Turn a call to a function created by 1306// init_trampoline / adjust_trampoline intrinsic pair into a direct call to the 1307// underlying function. 1308// 1309Instruction * 1310InstCombiner::transformCallThroughTrampoline(CallSite CS, 1311 IntrinsicInst *Tramp) { 1312 Value *Callee = CS.getCalledValue(); 1313 PointerType *PTy = cast<PointerType>(Callee->getType()); 1314 FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 1315 const AttrListPtr &Attrs = CS.getAttributes(); 1316 1317 // If the call already has the 'nest' attribute somewhere then give up - 1318 // otherwise 'nest' would occur twice after splicing in the chain. 1319 if (Attrs.hasAttrSomewhere(Attribute::Nest)) 1320 return 0; 1321 1322 assert(Tramp && 1323 "transformCallThroughTrampoline called with incorrect CallSite."); 1324 1325 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts()); 1326 PointerType *NestFPTy = cast<PointerType>(NestF->getType()); 1327 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType()); 1328 1329 const AttrListPtr &NestAttrs = NestF->getAttributes(); 1330 if (!NestAttrs.isEmpty()) { 1331 unsigned NestIdx = 1; 1332 Type *NestTy = 0; 1333 Attributes NestAttr = Attribute::None; 1334 1335 // Look for a parameter marked with the 'nest' attribute. 1336 for (FunctionType::param_iterator I = NestFTy->param_begin(), 1337 E = NestFTy->param_end(); I != E; ++NestIdx, ++I) 1338 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) { 1339 // Record the parameter type and any other attributes. 1340 NestTy = *I; 1341 NestAttr = NestAttrs.getParamAttributes(NestIdx); 1342 break; 1343 } 1344 1345 if (NestTy) { 1346 Instruction *Caller = CS.getInstruction(); 1347 std::vector<Value*> NewArgs; 1348 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1); 1349 1350 SmallVector<AttributeWithIndex, 8> NewAttrs; 1351 NewAttrs.reserve(Attrs.getNumSlots() + 1); 1352 1353 // Insert the nest argument into the call argument list, which may 1354 // mean appending it. Likewise for attributes. 1355 1356 // Add any result attributes. 1357 if (Attributes Attr = Attrs.getRetAttributes()) 1358 NewAttrs.push_back(AttributeWithIndex::get(0, Attr)); 1359 1360 { 1361 unsigned Idx = 1; 1362 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 1363 do { 1364 if (Idx == NestIdx) { 1365 // Add the chain argument and attributes. 1366 Value *NestVal = Tramp->getArgOperand(2); 1367 if (NestVal->getType() != NestTy) 1368 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest"); 1369 NewArgs.push_back(NestVal); 1370 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr)); 1371 } 1372 1373 if (I == E) 1374 break; 1375 1376 // Add the original argument and attributes. 1377 NewArgs.push_back(*I); 1378 if (Attributes Attr = Attrs.getParamAttributes(Idx)) 1379 NewAttrs.push_back 1380 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr)); 1381 1382 ++Idx, ++I; 1383 } while (1); 1384 } 1385 1386 // Add any function attributes. 1387 if (Attributes Attr = Attrs.getFnAttributes()) 1388 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr)); 1389 1390 // The trampoline may have been bitcast to a bogus type (FTy). 1391 // Handle this by synthesizing a new function type, equal to FTy 1392 // with the chain parameter inserted. 1393 1394 std::vector<Type*> NewTypes; 1395 NewTypes.reserve(FTy->getNumParams()+1); 1396 1397 // Insert the chain's type into the list of parameter types, which may 1398 // mean appending it. 1399 { 1400 unsigned Idx = 1; 1401 FunctionType::param_iterator I = FTy->param_begin(), 1402 E = FTy->param_end(); 1403 1404 do { 1405 if (Idx == NestIdx) 1406 // Add the chain's type. 1407 NewTypes.push_back(NestTy); 1408 1409 if (I == E) 1410 break; 1411 1412 // Add the original type. 1413 NewTypes.push_back(*I); 1414 1415 ++Idx, ++I; 1416 } while (1); 1417 } 1418 1419 // Replace the trampoline call with a direct call. Let the generic 1420 // code sort out any function type mismatches. 1421 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes, 1422 FTy->isVarArg()); 1423 Constant *NewCallee = 1424 NestF->getType() == PointerType::getUnqual(NewFTy) ? 1425 NestF : ConstantExpr::getBitCast(NestF, 1426 PointerType::getUnqual(NewFTy)); 1427 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(), 1428 NewAttrs.end()); 1429 1430 Instruction *NewCaller; 1431 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1432 NewCaller = InvokeInst::Create(NewCallee, 1433 II->getNormalDest(), II->getUnwindDest(), 1434 NewArgs); 1435 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv()); 1436 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL); 1437 } else { 1438 NewCaller = CallInst::Create(NewCallee, NewArgs); 1439 if (cast<CallInst>(Caller)->isTailCall()) 1440 cast<CallInst>(NewCaller)->setTailCall(); 1441 cast<CallInst>(NewCaller)-> 1442 setCallingConv(cast<CallInst>(Caller)->getCallingConv()); 1443 cast<CallInst>(NewCaller)->setAttributes(NewPAL); 1444 } 1445 1446 return NewCaller; 1447 } 1448 } 1449 1450 // Replace the trampoline call with a direct call. Since there is no 'nest' 1451 // parameter, there is no need to adjust the argument list. Let the generic 1452 // code sort out any function type mismatches. 1453 Constant *NewCallee = 1454 NestF->getType() == PTy ? NestF : 1455 ConstantExpr::getBitCast(NestF, PTy); 1456 CS.setCalledFunction(NewCallee); 1457 return CS.getInstruction(); 1458} 1459