InstCombineCalls.cpp revision a78fa8cc2dd6d2ffe5e4fe605f38aae7b3d2fb7a
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 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()) break; 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) break; 298 Size *= C->getZExtValue(); 299 } 300 } 301 } else if (CallInst *MI = extractMallocCall(Op1)) { 302 // Get allocation size. 303 Type* MallocType = getMallocAllocatedType(MI); 304 if (MallocType && MallocType->isSized()) 305 if (Value *NElems = getMallocArraySize(MI, TD, true)) 306 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems)) 307 Size = NElements->getZExtValue() * TD->getTypeAllocSize(MallocType); 308 } 309 310 // Do not return "I don't know" here. Later optimization passes could 311 // make it possible to evaluate objectsize to a constant. 312 if (Size == -1ULL) 313 break; 314 315 if (Size < Offset) { 316 // Out of bound reference? Negative index normalized to large 317 // index? Just return "I don't know". 318 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, DontKnow)); 319 } 320 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, Size-Offset)); 321 } 322 case Intrinsic::bswap: 323 // bswap(bswap(x)) -> x 324 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) 325 if (Operand->getIntrinsicID() == Intrinsic::bswap) 326 return ReplaceInstUsesWith(CI, Operand->getArgOperand(0)); 327 328 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c)) 329 if (TruncInst *TI = dyn_cast<TruncInst>(II->getArgOperand(0))) { 330 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0))) 331 if (Operand->getIntrinsicID() == Intrinsic::bswap) { 332 unsigned C = Operand->getType()->getPrimitiveSizeInBits() - 333 TI->getType()->getPrimitiveSizeInBits(); 334 Value *CV = ConstantInt::get(Operand->getType(), C); 335 Value *V = Builder->CreateLShr(Operand->getArgOperand(0), CV); 336 return new TruncInst(V, TI->getType()); 337 } 338 } 339 340 break; 341 case Intrinsic::powi: 342 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 343 // powi(x, 0) -> 1.0 344 if (Power->isZero()) 345 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0)); 346 // powi(x, 1) -> x 347 if (Power->isOne()) 348 return ReplaceInstUsesWith(CI, II->getArgOperand(0)); 349 // powi(x, -1) -> 1/x 350 if (Power->isAllOnesValue()) 351 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0), 352 II->getArgOperand(0)); 353 } 354 break; 355 case Intrinsic::cttz: { 356 // If all bits below the first known one are known zero, 357 // this value is constant. 358 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType()); 359 // FIXME: Try to simplify vectors of integers. 360 if (!IT) break; 361 uint32_t BitWidth = IT->getBitWidth(); 362 APInt KnownZero(BitWidth, 0); 363 APInt KnownOne(BitWidth, 0); 364 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth), 365 KnownZero, KnownOne); 366 unsigned TrailingZeros = KnownOne.countTrailingZeros(); 367 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros)); 368 if ((Mask & KnownZero) == Mask) 369 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 370 APInt(BitWidth, TrailingZeros))); 371 372 } 373 break; 374 case Intrinsic::ctlz: { 375 // If all bits above the first known one are known zero, 376 // this value is constant. 377 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType()); 378 // FIXME: Try to simplify vectors of integers. 379 if (!IT) break; 380 uint32_t BitWidth = IT->getBitWidth(); 381 APInt KnownZero(BitWidth, 0); 382 APInt KnownOne(BitWidth, 0); 383 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth), 384 KnownZero, KnownOne); 385 unsigned LeadingZeros = KnownOne.countLeadingZeros(); 386 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros)); 387 if ((Mask & KnownZero) == Mask) 388 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 389 APInt(BitWidth, LeadingZeros))); 390 391 } 392 break; 393 case Intrinsic::uadd_with_overflow: { 394 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 395 IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType()); 396 uint32_t BitWidth = IT->getBitWidth(); 397 APInt Mask = APInt::getSignBit(BitWidth); 398 APInt LHSKnownZero(BitWidth, 0); 399 APInt LHSKnownOne(BitWidth, 0); 400 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne); 401 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1]; 402 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1]; 403 404 if (LHSKnownNegative || LHSKnownPositive) { 405 APInt RHSKnownZero(BitWidth, 0); 406 APInt RHSKnownOne(BitWidth, 0); 407 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne); 408 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1]; 409 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1]; 410 if (LHSKnownNegative && RHSKnownNegative) { 411 // The sign bit is set in both cases: this MUST overflow. 412 // Create a simple add instruction, and insert it into the struct. 413 Value *Add = Builder->CreateAdd(LHS, RHS); 414 Add->takeName(&CI); 415 Constant *V[] = { 416 UndefValue::get(LHS->getType()), 417 ConstantInt::getTrue(II->getContext()) 418 }; 419 StructType *ST = cast<StructType>(II->getType()); 420 Constant *Struct = ConstantStruct::get(ST, V); 421 return InsertValueInst::Create(Struct, Add, 0); 422 } 423 424 if (LHSKnownPositive && RHSKnownPositive) { 425 // The sign bit is clear in both cases: this CANNOT overflow. 426 // Create a simple add instruction, and insert it into the struct. 427 Value *Add = Builder->CreateNUWAdd(LHS, RHS); 428 Add->takeName(&CI); 429 Constant *V[] = { 430 UndefValue::get(LHS->getType()), 431 ConstantInt::getFalse(II->getContext()) 432 }; 433 StructType *ST = cast<StructType>(II->getType()); 434 Constant *Struct = ConstantStruct::get(ST, V); 435 return InsertValueInst::Create(Struct, Add, 0); 436 } 437 } 438 } 439 // FALL THROUGH uadd into sadd 440 case Intrinsic::sadd_with_overflow: 441 // Canonicalize constants into the RHS. 442 if (isa<Constant>(II->getArgOperand(0)) && 443 !isa<Constant>(II->getArgOperand(1))) { 444 Value *LHS = II->getArgOperand(0); 445 II->setArgOperand(0, II->getArgOperand(1)); 446 II->setArgOperand(1, LHS); 447 return II; 448 } 449 450 // X + undef -> undef 451 if (isa<UndefValue>(II->getArgOperand(1))) 452 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 453 454 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 455 // X + 0 -> {X, false} 456 if (RHS->isZero()) { 457 Constant *V[] = { 458 UndefValue::get(II->getArgOperand(0)->getType()), 459 ConstantInt::getFalse(II->getContext()) 460 }; 461 Constant *Struct = 462 ConstantStruct::get(cast<StructType>(II->getType()), V); 463 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 464 } 465 } 466 break; 467 case Intrinsic::usub_with_overflow: 468 case Intrinsic::ssub_with_overflow: 469 // undef - X -> undef 470 // X - undef -> undef 471 if (isa<UndefValue>(II->getArgOperand(0)) || 472 isa<UndefValue>(II->getArgOperand(1))) 473 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 474 475 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 476 // X - 0 -> {X, false} 477 if (RHS->isZero()) { 478 Constant *V[] = { 479 UndefValue::get(II->getArgOperand(0)->getType()), 480 ConstantInt::getFalse(II->getContext()) 481 }; 482 Constant *Struct = 483 ConstantStruct::get(cast<StructType>(II->getType()), V); 484 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 485 } 486 } 487 break; 488 case Intrinsic::umul_with_overflow: { 489 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 490 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth(); 491 APInt Mask = APInt::getAllOnesValue(BitWidth); 492 493 APInt LHSKnownZero(BitWidth, 0); 494 APInt LHSKnownOne(BitWidth, 0); 495 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne); 496 APInt RHSKnownZero(BitWidth, 0); 497 APInt RHSKnownOne(BitWidth, 0); 498 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne); 499 500 // Get the largest possible values for each operand. 501 APInt LHSMax = ~LHSKnownZero; 502 APInt RHSMax = ~RHSKnownZero; 503 504 // If multiplying the maximum values does not overflow then we can turn 505 // this into a plain NUW mul. 506 bool Overflow; 507 LHSMax.umul_ov(RHSMax, Overflow); 508 if (!Overflow) { 509 Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow"); 510 Constant *V[] = { 511 UndefValue::get(LHS->getType()), 512 Builder->getFalse() 513 }; 514 Constant *Struct = ConstantStruct::get(cast<StructType>(II->getType()),V); 515 return InsertValueInst::Create(Struct, Mul, 0); 516 } 517 } // FALL THROUGH 518 case Intrinsic::smul_with_overflow: 519 // Canonicalize constants into the RHS. 520 if (isa<Constant>(II->getArgOperand(0)) && 521 !isa<Constant>(II->getArgOperand(1))) { 522 Value *LHS = II->getArgOperand(0); 523 II->setArgOperand(0, II->getArgOperand(1)); 524 II->setArgOperand(1, LHS); 525 return II; 526 } 527 528 // X * undef -> undef 529 if (isa<UndefValue>(II->getArgOperand(1))) 530 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 531 532 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 533 // X*0 -> {0, false} 534 if (RHSI->isZero()) 535 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType())); 536 537 // X * 1 -> {X, false} 538 if (RHSI->equalsInt(1)) { 539 Constant *V[] = { 540 UndefValue::get(II->getArgOperand(0)->getType()), 541 ConstantInt::getFalse(II->getContext()) 542 }; 543 Constant *Struct = 544 ConstantStruct::get(cast<StructType>(II->getType()), V); 545 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 546 } 547 } 548 break; 549 case Intrinsic::ppc_altivec_lvx: 550 case Intrinsic::ppc_altivec_lvxl: 551 // Turn PPC lvx -> load if the pointer is known aligned. 552 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 553 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), 554 PointerType::getUnqual(II->getType())); 555 return new LoadInst(Ptr); 556 } 557 break; 558 case Intrinsic::ppc_altivec_stvx: 559 case Intrinsic::ppc_altivec_stvxl: 560 // Turn stvx -> store if the pointer is known aligned. 561 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) { 562 Type *OpPtrTy = 563 PointerType::getUnqual(II->getArgOperand(0)->getType()); 564 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); 565 return new StoreInst(II->getArgOperand(0), Ptr); 566 } 567 break; 568 case Intrinsic::x86_sse_storeu_ps: 569 case Intrinsic::x86_sse2_storeu_pd: 570 case Intrinsic::x86_sse2_storeu_dq: 571 // Turn X86 storeu -> store if the pointer is known aligned. 572 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 573 Type *OpPtrTy = 574 PointerType::getUnqual(II->getArgOperand(1)->getType()); 575 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy); 576 return new StoreInst(II->getArgOperand(1), Ptr); 577 } 578 break; 579 580 case Intrinsic::x86_sse_cvtss2si: 581 case Intrinsic::x86_sse_cvtss2si64: 582 case Intrinsic::x86_sse_cvttss2si: 583 case Intrinsic::x86_sse_cvttss2si64: 584 case Intrinsic::x86_sse2_cvtsd2si: 585 case Intrinsic::x86_sse2_cvtsd2si64: 586 case Intrinsic::x86_sse2_cvttsd2si: 587 case Intrinsic::x86_sse2_cvttsd2si64: { 588 // These intrinsics only demand the 0th element of their input vectors. If 589 // we can simplify the input based on that, do so now. 590 unsigned VWidth = 591 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 592 APInt DemandedElts(VWidth, 1); 593 APInt UndefElts(VWidth, 0); 594 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0), 595 DemandedElts, UndefElts)) { 596 II->setArgOperand(0, V); 597 return II; 598 } 599 break; 600 } 601 602 603 case Intrinsic::x86_sse41_pmovsxbw: 604 case Intrinsic::x86_sse41_pmovsxwd: 605 case Intrinsic::x86_sse41_pmovsxdq: 606 case Intrinsic::x86_sse41_pmovzxbw: 607 case Intrinsic::x86_sse41_pmovzxwd: 608 case Intrinsic::x86_sse41_pmovzxdq: { 609 // pmov{s|z}x ignores the upper half of their input vectors. 610 unsigned VWidth = 611 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 612 unsigned LowHalfElts = VWidth / 2; 613 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts)); 614 APInt UndefElts(VWidth, 0); 615 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), 616 InputDemandedElts, 617 UndefElts)) { 618 II->setArgOperand(0, TmpV); 619 return II; 620 } 621 break; 622 } 623 624 case Intrinsic::ppc_altivec_vperm: 625 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant. 626 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) { 627 assert(Mask->getType()->getVectorNumElements() == 16 && 628 "Bad type for intrinsic!"); 629 630 // Check that all of the elements are integer constants or undefs. 631 bool AllEltsOk = true; 632 for (unsigned i = 0; i != 16; ++i) { 633 Constant *Elt = Mask->getAggregateElement(i); 634 if (Elt == 0 || 635 !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) { 636 AllEltsOk = false; 637 break; 638 } 639 } 640 641 if (AllEltsOk) { 642 // Cast the input vectors to byte vectors. 643 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0), 644 Mask->getType()); 645 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1), 646 Mask->getType()); 647 Value *Result = UndefValue::get(Op0->getType()); 648 649 // Only extract each element once. 650 Value *ExtractedElts[32]; 651 memset(ExtractedElts, 0, sizeof(ExtractedElts)); 652 653 for (unsigned i = 0; i != 16; ++i) { 654 if (isa<UndefValue>(Mask->getAggregateElement(i))) 655 continue; 656 unsigned Idx = 657 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue(); 658 Idx &= 31; // Match the hardware behavior. 659 660 if (ExtractedElts[Idx] == 0) { 661 ExtractedElts[Idx] = 662 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1, 663 Builder->getInt32(Idx&15)); 664 } 665 666 // Insert this value into the result vector. 667 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx], 668 Builder->getInt32(i)); 669 } 670 return CastInst::Create(Instruction::BitCast, Result, CI.getType()); 671 } 672 } 673 break; 674 675 case Intrinsic::arm_neon_vld1: 676 case Intrinsic::arm_neon_vld2: 677 case Intrinsic::arm_neon_vld3: 678 case Intrinsic::arm_neon_vld4: 679 case Intrinsic::arm_neon_vld2lane: 680 case Intrinsic::arm_neon_vld3lane: 681 case Intrinsic::arm_neon_vld4lane: 682 case Intrinsic::arm_neon_vst1: 683 case Intrinsic::arm_neon_vst2: 684 case Intrinsic::arm_neon_vst3: 685 case Intrinsic::arm_neon_vst4: 686 case Intrinsic::arm_neon_vst2lane: 687 case Intrinsic::arm_neon_vst3lane: 688 case Intrinsic::arm_neon_vst4lane: { 689 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD); 690 unsigned AlignArg = II->getNumArgOperands() - 1; 691 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg)); 692 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) { 693 II->setArgOperand(AlignArg, 694 ConstantInt::get(Type::getInt32Ty(II->getContext()), 695 MemAlign, false)); 696 return II; 697 } 698 break; 699 } 700 701 case Intrinsic::stackrestore: { 702 // If the save is right next to the restore, remove the restore. This can 703 // happen when variable allocas are DCE'd. 704 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) { 705 if (SS->getIntrinsicID() == Intrinsic::stacksave) { 706 BasicBlock::iterator BI = SS; 707 if (&*++BI == II) 708 return EraseInstFromFunction(CI); 709 } 710 } 711 712 // Scan down this block to see if there is another stack restore in the 713 // same block without an intervening call/alloca. 714 BasicBlock::iterator BI = II; 715 TerminatorInst *TI = II->getParent()->getTerminator(); 716 bool CannotRemove = false; 717 for (++BI; &*BI != TI; ++BI) { 718 if (isa<AllocaInst>(BI) || isMalloc(BI)) { 719 CannotRemove = true; 720 break; 721 } 722 if (CallInst *BCI = dyn_cast<CallInst>(BI)) { 723 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) { 724 // If there is a stackrestore below this one, remove this one. 725 if (II->getIntrinsicID() == Intrinsic::stackrestore) 726 return EraseInstFromFunction(CI); 727 // Otherwise, ignore the intrinsic. 728 } else { 729 // If we found a non-intrinsic call, we can't remove the stack 730 // restore. 731 CannotRemove = true; 732 break; 733 } 734 } 735 } 736 737 // If the stack restore is in a return, resume, or unwind block and if there 738 // are no allocas or calls between the restore and the return, nuke the 739 // restore. 740 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI) || 741 isa<UnwindInst>(TI))) 742 return EraseInstFromFunction(CI); 743 break; 744 } 745 } 746 747 return visitCallSite(II); 748} 749 750// InvokeInst simplification 751// 752Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) { 753 return visitCallSite(&II); 754} 755 756/// isSafeToEliminateVarargsCast - If this cast does not affect the value 757/// passed through the varargs area, we can eliminate the use of the cast. 758static bool isSafeToEliminateVarargsCast(const CallSite CS, 759 const CastInst * const CI, 760 const TargetData * const TD, 761 const int ix) { 762 if (!CI->isLosslessCast()) 763 return false; 764 765 // The size of ByVal arguments is derived from the type, so we 766 // can't change to a type with a different size. If the size were 767 // passed explicitly we could avoid this check. 768 if (!CS.isByValArgument(ix)) 769 return true; 770 771 Type* SrcTy = 772 cast<PointerType>(CI->getOperand(0)->getType())->getElementType(); 773 Type* DstTy = cast<PointerType>(CI->getType())->getElementType(); 774 if (!SrcTy->isSized() || !DstTy->isSized()) 775 return false; 776 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy)) 777 return false; 778 return true; 779} 780 781namespace { 782class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls { 783 InstCombiner *IC; 784protected: 785 void replaceCall(Value *With) { 786 NewInstruction = IC->ReplaceInstUsesWith(*CI, With); 787 } 788 bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const { 789 if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp)) 790 return true; 791 if (ConstantInt *SizeCI = 792 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) { 793 if (SizeCI->isAllOnesValue()) 794 return true; 795 if (isString) { 796 uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp)); 797 // If the length is 0 we don't know how long it is and so we can't 798 // remove the check. 799 if (Len == 0) return false; 800 return SizeCI->getZExtValue() >= Len; 801 } 802 if (ConstantInt *Arg = dyn_cast<ConstantInt>( 803 CI->getArgOperand(SizeArgOp))) 804 return SizeCI->getZExtValue() >= Arg->getZExtValue(); 805 } 806 return false; 807 } 808public: 809 InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { } 810 Instruction *NewInstruction; 811}; 812} // end anonymous namespace 813 814// Try to fold some different type of calls here. 815// Currently we're only working with the checking functions, memcpy_chk, 816// mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk, 817// strcat_chk and strncat_chk. 818Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) { 819 if (CI->getCalledFunction() == 0) return 0; 820 821 InstCombineFortifiedLibCalls Simplifier(this); 822 Simplifier.fold(CI, TD); 823 return Simplifier.NewInstruction; 824} 825 826static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) { 827 // Strip off at most one level of pointer casts, looking for an alloca. This 828 // is good enough in practice and simpler than handling any number of casts. 829 Value *Underlying = TrampMem->stripPointerCasts(); 830 if (Underlying != TrampMem && 831 (!Underlying->hasOneUse() || *Underlying->use_begin() != TrampMem)) 832 return 0; 833 if (!isa<AllocaInst>(Underlying)) 834 return 0; 835 836 IntrinsicInst *InitTrampoline = 0; 837 for (Value::use_iterator I = TrampMem->use_begin(), E = TrampMem->use_end(); 838 I != E; I++) { 839 IntrinsicInst *II = dyn_cast<IntrinsicInst>(*I); 840 if (!II) 841 return 0; 842 if (II->getIntrinsicID() == Intrinsic::init_trampoline) { 843 if (InitTrampoline) 844 // More than one init_trampoline writes to this value. Give up. 845 return 0; 846 InitTrampoline = II; 847 continue; 848 } 849 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline) 850 // Allow any number of calls to adjust.trampoline. 851 continue; 852 return 0; 853 } 854 855 // No call to init.trampoline found. 856 if (!InitTrampoline) 857 return 0; 858 859 // Check that the alloca is being used in the expected way. 860 if (InitTrampoline->getOperand(0) != TrampMem) 861 return 0; 862 863 return InitTrampoline; 864} 865 866static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp, 867 Value *TrampMem) { 868 // Visit all the previous instructions in the basic block, and try to find a 869 // init.trampoline which has a direct path to the adjust.trampoline. 870 for (BasicBlock::iterator I = AdjustTramp, 871 E = AdjustTramp->getParent()->begin(); I != E; ) { 872 Instruction *Inst = --I; 873 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) 874 if (II->getIntrinsicID() == Intrinsic::init_trampoline && 875 II->getOperand(0) == TrampMem) 876 return II; 877 if (Inst->mayWriteToMemory()) 878 return 0; 879 } 880 return 0; 881} 882 883// Given a call to llvm.adjust.trampoline, find and return the corresponding 884// call to llvm.init.trampoline if the call to the trampoline can be optimized 885// to a direct call to a function. Otherwise return NULL. 886// 887static IntrinsicInst *FindInitTrampoline(Value *Callee) { 888 Callee = Callee->stripPointerCasts(); 889 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee); 890 if (!AdjustTramp || 891 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline) 892 return 0; 893 894 Value *TrampMem = AdjustTramp->getOperand(0); 895 896 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem)) 897 return IT; 898 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem)) 899 return IT; 900 return 0; 901} 902 903// visitCallSite - Improvements for call and invoke instructions. 904// 905Instruction *InstCombiner::visitCallSite(CallSite CS) { 906 bool Changed = false; 907 908 // If the callee is a pointer to a function, attempt to move any casts to the 909 // arguments of the call/invoke. 910 Value *Callee = CS.getCalledValue(); 911 if (!isa<Function>(Callee) && transformConstExprCastCall(CS)) 912 return 0; 913 914 if (Function *CalleeF = dyn_cast<Function>(Callee)) 915 // If the call and callee calling conventions don't match, this call must 916 // be unreachable, as the call is undefined. 917 if (CalleeF->getCallingConv() != CS.getCallingConv() && 918 // Only do this for calls to a function with a body. A prototype may 919 // not actually end up matching the implementation's calling conv for a 920 // variety of reasons (e.g. it may be written in assembly). 921 !CalleeF->isDeclaration()) { 922 Instruction *OldCall = CS.getInstruction(); 923 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 924 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 925 OldCall); 926 // If OldCall dues not return void then replaceAllUsesWith undef. 927 // This allows ValueHandlers and custom metadata to adjust itself. 928 if (!OldCall->getType()->isVoidTy()) 929 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType())); 930 if (isa<CallInst>(OldCall)) 931 return EraseInstFromFunction(*OldCall); 932 933 // We cannot remove an invoke, because it would change the CFG, just 934 // change the callee to a null pointer. 935 cast<InvokeInst>(OldCall)->setCalledFunction( 936 Constant::getNullValue(CalleeF->getType())); 937 return 0; 938 } 939 940 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 941 // This instruction is not reachable, just remove it. We insert a store to 942 // undef so that we know that this code is not reachable, despite the fact 943 // that we can't modify the CFG here. 944 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 945 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 946 CS.getInstruction()); 947 948 // If CS does not return void then replaceAllUsesWith undef. 949 // This allows ValueHandlers and custom metadata to adjust itself. 950 if (!CS.getInstruction()->getType()->isVoidTy()) 951 ReplaceInstUsesWith(*CS.getInstruction(), 952 UndefValue::get(CS.getInstruction()->getType())); 953 954 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) { 955 // Don't break the CFG, insert a dummy cond branch. 956 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(), 957 ConstantInt::getTrue(Callee->getContext()), II); 958 } 959 return EraseInstFromFunction(*CS.getInstruction()); 960 } 961 962 if (IntrinsicInst *II = FindInitTrampoline(Callee)) 963 return transformCallThroughTrampoline(CS, II); 964 965 PointerType *PTy = cast<PointerType>(Callee->getType()); 966 FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 967 if (FTy->isVarArg()) { 968 int ix = FTy->getNumParams(); 969 // See if we can optimize any arguments passed through the varargs area of 970 // the call. 971 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(), 972 E = CS.arg_end(); I != E; ++I, ++ix) { 973 CastInst *CI = dyn_cast<CastInst>(*I); 974 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) { 975 *I = CI->getOperand(0); 976 Changed = true; 977 } 978 } 979 } 980 981 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) { 982 // Inline asm calls cannot throw - mark them 'nounwind'. 983 CS.setDoesNotThrow(); 984 Changed = true; 985 } 986 987 // Try to optimize the call if possible, we require TargetData for most of 988 // this. None of these calls are seen as possibly dead so go ahead and 989 // delete the instruction now. 990 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) { 991 Instruction *I = tryOptimizeCall(CI, TD); 992 // If we changed something return the result, etc. Otherwise let 993 // the fallthrough check. 994 if (I) return EraseInstFromFunction(*I); 995 } 996 997 return Changed ? CS.getInstruction() : 0; 998} 999 1000// transformConstExprCastCall - If the callee is a constexpr cast of a function, 1001// attempt to move the cast to the arguments of the call/invoke. 1002// 1003bool InstCombiner::transformConstExprCastCall(CallSite CS) { 1004 Function *Callee = 1005 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts()); 1006 if (Callee == 0) 1007 return false; 1008 Instruction *Caller = CS.getInstruction(); 1009 const AttrListPtr &CallerPAL = CS.getAttributes(); 1010 1011 // Okay, this is a cast from a function to a different type. Unless doing so 1012 // would cause a type conversion of one of our arguments, change this call to 1013 // be a direct call with arguments casted to the appropriate types. 1014 // 1015 FunctionType *FT = Callee->getFunctionType(); 1016 Type *OldRetTy = Caller->getType(); 1017 Type *NewRetTy = FT->getReturnType(); 1018 1019 if (NewRetTy->isStructTy()) 1020 return false; // TODO: Handle multiple return values. 1021 1022 // Check to see if we are changing the return type... 1023 if (OldRetTy != NewRetTy) { 1024 if (Callee->isDeclaration() && 1025 // Conversion is ok if changing from one pointer type to another or from 1026 // a pointer to an integer of the same size. 1027 !((OldRetTy->isPointerTy() || !TD || 1028 OldRetTy == TD->getIntPtrType(Caller->getContext())) && 1029 (NewRetTy->isPointerTy() || !TD || 1030 NewRetTy == TD->getIntPtrType(Caller->getContext())))) 1031 return false; // Cannot transform this return value. 1032 1033 if (!Caller->use_empty() && 1034 // void -> non-void is handled specially 1035 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy)) 1036 return false; // Cannot transform this return value. 1037 1038 if (!CallerPAL.isEmpty() && !Caller->use_empty()) { 1039 Attributes RAttrs = CallerPAL.getRetAttributes(); 1040 if (RAttrs & Attribute::typeIncompatible(NewRetTy)) 1041 return false; // Attribute not compatible with transformed value. 1042 } 1043 1044 // If the callsite is an invoke instruction, and the return value is used by 1045 // a PHI node in a successor, we cannot change the return type of the call 1046 // because there is no place to put the cast instruction (without breaking 1047 // the critical edge). Bail out in this case. 1048 if (!Caller->use_empty()) 1049 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) 1050 for (Value::use_iterator UI = II->use_begin(), E = II->use_end(); 1051 UI != E; ++UI) 1052 if (PHINode *PN = dyn_cast<PHINode>(*UI)) 1053 if (PN->getParent() == II->getNormalDest() || 1054 PN->getParent() == II->getUnwindDest()) 1055 return false; 1056 } 1057 1058 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin()); 1059 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs); 1060 1061 CallSite::arg_iterator AI = CS.arg_begin(); 1062 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) { 1063 Type *ParamTy = FT->getParamType(i); 1064 Type *ActTy = (*AI)->getType(); 1065 1066 if (!CastInst::isCastable(ActTy, ParamTy)) 1067 return false; // Cannot transform this parameter value. 1068 1069 Attributes Attrs = CallerPAL.getParamAttributes(i + 1); 1070 if (Attrs & Attribute::typeIncompatible(ParamTy)) 1071 return false; // Attribute not compatible with transformed value. 1072 1073 // If the parameter is passed as a byval argument, then we have to have a 1074 // sized type and the sized type has to have the same size as the old type. 1075 if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) { 1076 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy); 1077 if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0) 1078 return false; 1079 1080 Type *CurElTy = cast<PointerType>(ActTy)->getElementType(); 1081 if (TD->getTypeAllocSize(CurElTy) != 1082 TD->getTypeAllocSize(ParamPTy->getElementType())) 1083 return false; 1084 } 1085 1086 // Converting from one pointer type to another or between a pointer and an 1087 // integer of the same size is safe even if we do not have a body. 1088 bool isConvertible = ActTy == ParamTy || 1089 (TD && ((ParamTy->isPointerTy() || 1090 ParamTy == TD->getIntPtrType(Caller->getContext())) && 1091 (ActTy->isPointerTy() || 1092 ActTy == TD->getIntPtrType(Caller->getContext())))); 1093 if (Callee->isDeclaration() && !isConvertible) return false; 1094 } 1095 1096 if (Callee->isDeclaration()) { 1097 // Do not delete arguments unless we have a function body. 1098 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg()) 1099 return false; 1100 1101 // If the callee is just a declaration, don't change the varargsness of the 1102 // call. We don't want to introduce a varargs call where one doesn't 1103 // already exist. 1104 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType()); 1105 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg()) 1106 return false; 1107 } 1108 1109 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && 1110 !CallerPAL.isEmpty()) 1111 // In this case we have more arguments than the new function type, but we 1112 // won't be dropping them. Check that these extra arguments have attributes 1113 // that are compatible with being a vararg call argument. 1114 for (unsigned i = CallerPAL.getNumSlots(); i; --i) { 1115 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams()) 1116 break; 1117 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs; 1118 if (PAttrs & Attribute::VarArgsIncompatible) 1119 return false; 1120 } 1121 1122 1123 // Okay, we decided that this is a safe thing to do: go ahead and start 1124 // inserting cast instructions as necessary. 1125 std::vector<Value*> Args; 1126 Args.reserve(NumActualArgs); 1127 SmallVector<AttributeWithIndex, 8> attrVec; 1128 attrVec.reserve(NumCommonArgs); 1129 1130 // Get any return attributes. 1131 Attributes RAttrs = CallerPAL.getRetAttributes(); 1132 1133 // If the return value is not being used, the type may not be compatible 1134 // with the existing attributes. Wipe out any problematic attributes. 1135 RAttrs &= ~Attribute::typeIncompatible(NewRetTy); 1136 1137 // Add the new return attributes. 1138 if (RAttrs) 1139 attrVec.push_back(AttributeWithIndex::get(0, RAttrs)); 1140 1141 AI = CS.arg_begin(); 1142 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) { 1143 Type *ParamTy = FT->getParamType(i); 1144 if ((*AI)->getType() == ParamTy) { 1145 Args.push_back(*AI); 1146 } else { 1147 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, 1148 false, ParamTy, false); 1149 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy)); 1150 } 1151 1152 // Add any parameter attributes. 1153 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1154 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1155 } 1156 1157 // If the function takes more arguments than the call was taking, add them 1158 // now. 1159 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) 1160 Args.push_back(Constant::getNullValue(FT->getParamType(i))); 1161 1162 // If we are removing arguments to the function, emit an obnoxious warning. 1163 if (FT->getNumParams() < NumActualArgs) { 1164 if (!FT->isVarArg()) { 1165 errs() << "WARNING: While resolving call to function '" 1166 << Callee->getName() << "' arguments were dropped!\n"; 1167 } else { 1168 // Add all of the arguments in their promoted form to the arg list. 1169 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) { 1170 Type *PTy = getPromotedType((*AI)->getType()); 1171 if (PTy != (*AI)->getType()) { 1172 // Must promote to pass through va_arg area! 1173 Instruction::CastOps opcode = 1174 CastInst::getCastOpcode(*AI, false, PTy, false); 1175 Args.push_back(Builder->CreateCast(opcode, *AI, PTy)); 1176 } else { 1177 Args.push_back(*AI); 1178 } 1179 1180 // Add any parameter attributes. 1181 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1182 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1183 } 1184 } 1185 } 1186 1187 if (Attributes FnAttrs = CallerPAL.getFnAttributes()) 1188 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); 1189 1190 if (NewRetTy->isVoidTy()) 1191 Caller->setName(""); // Void type should not have a name. 1192 1193 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(), 1194 attrVec.end()); 1195 1196 Instruction *NC; 1197 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1198 NC = Builder->CreateInvoke(Callee, II->getNormalDest(), 1199 II->getUnwindDest(), Args); 1200 NC->takeName(II); 1201 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv()); 1202 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL); 1203 } else { 1204 CallInst *CI = cast<CallInst>(Caller); 1205 NC = Builder->CreateCall(Callee, Args); 1206 NC->takeName(CI); 1207 if (CI->isTailCall()) 1208 cast<CallInst>(NC)->setTailCall(); 1209 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv()); 1210 cast<CallInst>(NC)->setAttributes(NewCallerPAL); 1211 } 1212 1213 // Insert a cast of the return type as necessary. 1214 Value *NV = NC; 1215 if (OldRetTy != NV->getType() && !Caller->use_empty()) { 1216 if (!NV->getType()->isVoidTy()) { 1217 Instruction::CastOps opcode = 1218 CastInst::getCastOpcode(NC, false, OldRetTy, false); 1219 NV = NC = CastInst::Create(opcode, NC, OldRetTy); 1220 NC->setDebugLoc(Caller->getDebugLoc()); 1221 1222 // If this is an invoke instruction, we should insert it after the first 1223 // non-phi, instruction in the normal successor block. 1224 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1225 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt(); 1226 InsertNewInstBefore(NC, *I); 1227 } else { 1228 // Otherwise, it's a call, just insert cast right after the call. 1229 InsertNewInstBefore(NC, *Caller); 1230 } 1231 Worklist.AddUsersToWorkList(*Caller); 1232 } else { 1233 NV = UndefValue::get(Caller->getType()); 1234 } 1235 } 1236 1237 if (!Caller->use_empty()) 1238 ReplaceInstUsesWith(*Caller, NV); 1239 1240 EraseInstFromFunction(*Caller); 1241 return true; 1242} 1243 1244// transformCallThroughTrampoline - Turn a call to a function created by 1245// init_trampoline / adjust_trampoline intrinsic pair into a direct call to the 1246// underlying function. 1247// 1248Instruction * 1249InstCombiner::transformCallThroughTrampoline(CallSite CS, 1250 IntrinsicInst *Tramp) { 1251 Value *Callee = CS.getCalledValue(); 1252 PointerType *PTy = cast<PointerType>(Callee->getType()); 1253 FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 1254 const AttrListPtr &Attrs = CS.getAttributes(); 1255 1256 // If the call already has the 'nest' attribute somewhere then give up - 1257 // otherwise 'nest' would occur twice after splicing in the chain. 1258 if (Attrs.hasAttrSomewhere(Attribute::Nest)) 1259 return 0; 1260 1261 assert(Tramp && 1262 "transformCallThroughTrampoline called with incorrect CallSite."); 1263 1264 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts()); 1265 PointerType *NestFPTy = cast<PointerType>(NestF->getType()); 1266 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType()); 1267 1268 const AttrListPtr &NestAttrs = NestF->getAttributes(); 1269 if (!NestAttrs.isEmpty()) { 1270 unsigned NestIdx = 1; 1271 Type *NestTy = 0; 1272 Attributes NestAttr = Attribute::None; 1273 1274 // Look for a parameter marked with the 'nest' attribute. 1275 for (FunctionType::param_iterator I = NestFTy->param_begin(), 1276 E = NestFTy->param_end(); I != E; ++NestIdx, ++I) 1277 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) { 1278 // Record the parameter type and any other attributes. 1279 NestTy = *I; 1280 NestAttr = NestAttrs.getParamAttributes(NestIdx); 1281 break; 1282 } 1283 1284 if (NestTy) { 1285 Instruction *Caller = CS.getInstruction(); 1286 std::vector<Value*> NewArgs; 1287 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1); 1288 1289 SmallVector<AttributeWithIndex, 8> NewAttrs; 1290 NewAttrs.reserve(Attrs.getNumSlots() + 1); 1291 1292 // Insert the nest argument into the call argument list, which may 1293 // mean appending it. Likewise for attributes. 1294 1295 // Add any result attributes. 1296 if (Attributes Attr = Attrs.getRetAttributes()) 1297 NewAttrs.push_back(AttributeWithIndex::get(0, Attr)); 1298 1299 { 1300 unsigned Idx = 1; 1301 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 1302 do { 1303 if (Idx == NestIdx) { 1304 // Add the chain argument and attributes. 1305 Value *NestVal = Tramp->getArgOperand(2); 1306 if (NestVal->getType() != NestTy) 1307 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest"); 1308 NewArgs.push_back(NestVal); 1309 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr)); 1310 } 1311 1312 if (I == E) 1313 break; 1314 1315 // Add the original argument and attributes. 1316 NewArgs.push_back(*I); 1317 if (Attributes Attr = Attrs.getParamAttributes(Idx)) 1318 NewAttrs.push_back 1319 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr)); 1320 1321 ++Idx, ++I; 1322 } while (1); 1323 } 1324 1325 // Add any function attributes. 1326 if (Attributes Attr = Attrs.getFnAttributes()) 1327 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr)); 1328 1329 // The trampoline may have been bitcast to a bogus type (FTy). 1330 // Handle this by synthesizing a new function type, equal to FTy 1331 // with the chain parameter inserted. 1332 1333 std::vector<Type*> NewTypes; 1334 NewTypes.reserve(FTy->getNumParams()+1); 1335 1336 // Insert the chain's type into the list of parameter types, which may 1337 // mean appending it. 1338 { 1339 unsigned Idx = 1; 1340 FunctionType::param_iterator I = FTy->param_begin(), 1341 E = FTy->param_end(); 1342 1343 do { 1344 if (Idx == NestIdx) 1345 // Add the chain's type. 1346 NewTypes.push_back(NestTy); 1347 1348 if (I == E) 1349 break; 1350 1351 // Add the original type. 1352 NewTypes.push_back(*I); 1353 1354 ++Idx, ++I; 1355 } while (1); 1356 } 1357 1358 // Replace the trampoline call with a direct call. Let the generic 1359 // code sort out any function type mismatches. 1360 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes, 1361 FTy->isVarArg()); 1362 Constant *NewCallee = 1363 NestF->getType() == PointerType::getUnqual(NewFTy) ? 1364 NestF : ConstantExpr::getBitCast(NestF, 1365 PointerType::getUnqual(NewFTy)); 1366 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(), 1367 NewAttrs.end()); 1368 1369 Instruction *NewCaller; 1370 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1371 NewCaller = InvokeInst::Create(NewCallee, 1372 II->getNormalDest(), II->getUnwindDest(), 1373 NewArgs); 1374 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv()); 1375 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL); 1376 } else { 1377 NewCaller = CallInst::Create(NewCallee, NewArgs); 1378 if (cast<CallInst>(Caller)->isTailCall()) 1379 cast<CallInst>(NewCaller)->setTailCall(); 1380 cast<CallInst>(NewCaller)-> 1381 setCallingConv(cast<CallInst>(Caller)->getCallingConv()); 1382 cast<CallInst>(NewCaller)->setAttributes(NewPAL); 1383 } 1384 1385 return NewCaller; 1386 } 1387 } 1388 1389 // Replace the trampoline call with a direct call. Since there is no 'nest' 1390 // parameter, there is no need to adjust the argument list. Let the generic 1391 // code sort out any function type mismatches. 1392 Constant *NewCallee = 1393 NestF->getType() == PTy ? NestF : 1394 ConstantExpr::getBitCast(NestF, PTy); 1395 CS.setCalledFunction(NewCallee); 1396 return CS.getInstruction(); 1397} 1398