InstCombineCalls.cpp revision 3e22cb9ec30cd9b1be9b0f50e400f512124997e5
1//===- InstCombineCalls.cpp -----------------------------------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements the visitCall and visitInvoke functions. 11// 12//===----------------------------------------------------------------------===// 13 14#include "InstCombine.h" 15#include "llvm/IntrinsicInst.h" 16#include "llvm/Support/CallSite.h" 17#include "llvm/Target/TargetData.h" 18#include "llvm/Analysis/MemoryBuiltins.h" 19#include "llvm/Transforms/Utils/BuildLibCalls.h" 20#include "llvm/Transforms/Utils/Local.h" 21using namespace llvm; 22 23/// getPromotedType - Return the specified type promoted as it would be to pass 24/// though a va_arg area. 25static const Type *getPromotedType(const Type *Ty) { 26 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) { 27 if (ITy->getBitWidth() < 32) 28 return Type::getInt32Ty(Ty->getContext()); 29 } 30 return Ty; 31} 32 33 34Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) { 35 unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), TD); 36 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), TD); 37 unsigned MinAlign = std::min(DstAlign, SrcAlign); 38 unsigned CopyAlign = MI->getAlignment(); 39 40 if (CopyAlign < MinAlign) { 41 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), 42 MinAlign, false)); 43 return MI; 44 } 45 46 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with 47 // load/store. 48 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2)); 49 if (MemOpLength == 0) return 0; 50 51 // Source and destination pointer types are always "i8*" for intrinsic. See 52 // if the size is something we can handle with a single primitive load/store. 53 // A single load+store correctly handles overlapping memory in the memmove 54 // case. 55 unsigned Size = MemOpLength->getZExtValue(); 56 if (Size == 0) return MI; // Delete this mem transfer. 57 58 if (Size > 8 || (Size&(Size-1))) 59 return 0; // If not 1/2/4/8 bytes, exit. 60 61 // Use an integer load+store unless we can find something better. 62 unsigned SrcAddrSp = 63 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace(); 64 unsigned DstAddrSp = 65 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace(); 66 67 const IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3); 68 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp); 69 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp); 70 71 // Memcpy forces the use of i8* for the source and destination. That means 72 // that if you're using memcpy to move one double around, you'll get a cast 73 // from double* to i8*. We'd much rather use a double load+store rather than 74 // an i64 load+store, here because this improves the odds that the source or 75 // dest address will be promotable. See if we can find a better type than the 76 // integer datatype. 77 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts(); 78 if (StrippedDest != MI->getArgOperand(0)) { 79 const Type *SrcETy = cast<PointerType>(StrippedDest->getType()) 80 ->getElementType(); 81 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) { 82 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip 83 // down through these levels if so. 84 while (!SrcETy->isSingleValueType()) { 85 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) { 86 if (STy->getNumElements() == 1) 87 SrcETy = STy->getElementType(0); 88 else 89 break; 90 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) { 91 if (ATy->getNumElements() == 1) 92 SrcETy = ATy->getElementType(); 93 else 94 break; 95 } else 96 break; 97 } 98 99 if (SrcETy->isSingleValueType()) { 100 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp); 101 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp); 102 } 103 } 104 } 105 106 107 // If the memcpy/memmove provides better alignment info than we can 108 // infer, use it. 109 SrcAlign = std::max(SrcAlign, CopyAlign); 110 DstAlign = std::max(DstAlign, CopyAlign); 111 112 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy); 113 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy); 114 Instruction *L = new LoadInst(Src, "tmp", MI->isVolatile(), SrcAlign); 115 InsertNewInstBefore(L, *MI); 116 InsertNewInstBefore(new StoreInst(L, Dest, MI->isVolatile(), DstAlign), 117 *MI); 118 119 // Set the size of the copy to 0, it will be deleted on the next iteration. 120 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType())); 121 return MI; 122} 123 124Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) { 125 unsigned Alignment = getKnownAlignment(MI->getDest(), TD); 126 if (MI->getAlignment() < Alignment) { 127 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), 128 Alignment, false)); 129 return MI; 130 } 131 132 // Extract the length and alignment and fill if they are constant. 133 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength()); 134 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue()); 135 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8)) 136 return 0; 137 uint64_t Len = LenC->getZExtValue(); 138 Alignment = MI->getAlignment(); 139 140 // If the length is zero, this is a no-op 141 if (Len == 0) return MI; // memset(d,c,0,a) -> noop 142 143 // memset(s,c,n) -> store s, c (for n=1,2,4,8) 144 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) { 145 const Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8. 146 147 Value *Dest = MI->getDest(); 148 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace(); 149 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp); 150 Dest = Builder->CreateBitCast(Dest, NewDstPtrTy); 151 152 // Alignment 0 is identity for alignment 1 for memset, but not store. 153 if (Alignment == 0) Alignment = 1; 154 155 // Extract the fill value and store. 156 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL; 157 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), 158 Dest, false, Alignment), *MI); 159 160 // Set the size of the copy to 0, it will be deleted on the next iteration. 161 MI->setLength(Constant::getNullValue(LenC->getType())); 162 return MI; 163 } 164 165 return 0; 166} 167 168/// visitCallInst - CallInst simplification. This mostly only handles folding 169/// of intrinsic instructions. For normal calls, it allows visitCallSite to do 170/// the heavy lifting. 171/// 172Instruction *InstCombiner::visitCallInst(CallInst &CI) { 173 if (isFreeCall(&CI)) 174 return visitFree(CI); 175 if (isMalloc(&CI)) 176 return visitMalloc(CI); 177 178 // If the caller function is nounwind, mark the call as nounwind, even if the 179 // callee isn't. 180 if (CI.getParent()->getParent()->doesNotThrow() && 181 !CI.doesNotThrow()) { 182 CI.setDoesNotThrow(); 183 return &CI; 184 } 185 186 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI); 187 if (!II) return visitCallSite(&CI); 188 189 // Intrinsics cannot occur in an invoke, so handle them here instead of in 190 // visitCallSite. 191 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) { 192 bool Changed = false; 193 194 // memmove/cpy/set of zero bytes is a noop. 195 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) { 196 if (NumBytes->isNullValue()) 197 return EraseInstFromFunction(CI); 198 199 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes)) 200 if (CI->getZExtValue() == 1) { 201 // Replace the instruction with just byte operations. We would 202 // transform other cases to loads/stores, but we don't know if 203 // alignment is sufficient. 204 } 205 } 206 207 // No other transformations apply to volatile transfers. 208 if (MI->isVolatile()) 209 return 0; 210 211 // If we have a memmove and the source operation is a constant global, 212 // then the source and dest pointers can't alias, so we can change this 213 // into a call to memcpy. 214 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) { 215 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource())) 216 if (GVSrc->isConstant()) { 217 Module *M = CI.getParent()->getParent()->getParent(); 218 Intrinsic::ID MemCpyID = Intrinsic::memcpy; 219 const Type *Tys[3] = { CI.getArgOperand(0)->getType(), 220 CI.getArgOperand(1)->getType(), 221 CI.getArgOperand(2)->getType() }; 222 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys, 3)); 223 Changed = true; 224 } 225 } 226 227 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { 228 // memmove(x,x,size) -> noop. 229 if (MTI->getSource() == MTI->getDest()) 230 return EraseInstFromFunction(CI); 231 } 232 233 // If we can determine a pointer alignment that is bigger than currently 234 // set, update the alignment. 235 if (isa<MemTransferInst>(MI)) { 236 if (Instruction *I = SimplifyMemTransfer(MI)) 237 return I; 238 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) { 239 if (Instruction *I = SimplifyMemSet(MSI)) 240 return I; 241 } 242 243 if (Changed) return II; 244 } 245 246 switch (II->getIntrinsicID()) { 247 default: break; 248 case Intrinsic::objectsize: { 249 // We need target data for just about everything so depend on it. 250 if (!TD) break; 251 252 const Type *ReturnTy = CI.getType(); 253 uint64_t DontKnow = II->getArgOperand(1) == Builder->getTrue() ? 0 : -1ULL; 254 255 // Get to the real allocated thing and offset as fast as possible. 256 Value *Op1 = II->getArgOperand(0)->stripPointerCasts(); 257 258 uint64_t Offset = 0; 259 uint64_t Size = -1ULL; 260 261 // Try to look through constant GEPs. 262 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1)) { 263 if (!GEP->hasAllConstantIndices()) break; 264 265 // Get the current byte offset into the thing. Use the original 266 // operand in case we're looking through a bitcast. 267 SmallVector<Value*, 8> Ops(GEP->idx_begin(), GEP->idx_end()); 268 Offset = TD->getIndexedOffset(GEP->getPointerOperandType(), 269 Ops.data(), Ops.size()); 270 271 Op1 = GEP->getPointerOperand()->stripPointerCasts(); 272 273 // Make sure we're not a constant offset from an external 274 // global. 275 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) 276 if (!GV->hasDefinitiveInitializer()) break; 277 } 278 279 // If we've stripped down to a single global variable that we 280 // can know the size of then just return that. 281 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op1)) { 282 if (GV->hasDefinitiveInitializer()) { 283 Constant *C = GV->getInitializer(); 284 Size = TD->getTypeAllocSize(C->getType()); 285 } else { 286 // Can't determine size of the GV. 287 Constant *RetVal = ConstantInt::get(ReturnTy, DontKnow); 288 return ReplaceInstUsesWith(CI, RetVal); 289 } 290 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(Op1)) { 291 // Get alloca size. 292 if (AI->getAllocatedType()->isSized()) { 293 Size = TD->getTypeAllocSize(AI->getAllocatedType()); 294 if (AI->isArrayAllocation()) { 295 const ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize()); 296 if (!C) break; 297 Size *= C->getZExtValue(); 298 } 299 } 300 } else if (CallInst *MI = extractMallocCall(Op1)) { 301 // Get allocation size. 302 const Type* MallocType = getMallocAllocatedType(MI); 303 if (MallocType && MallocType->isSized()) 304 if (Value *NElems = getMallocArraySize(MI, TD, true)) 305 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems)) 306 Size = NElements->getZExtValue() * TD->getTypeAllocSize(MallocType); 307 } 308 309 // Do not return "I don't know" here. Later optimization passes could 310 // make it possible to evaluate objectsize to a constant. 311 if (Size == -1ULL) 312 break; 313 314 if (Size < Offset) { 315 // Out of bound reference? Negative index normalized to large 316 // index? Just return "I don't know". 317 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, DontKnow)); 318 } 319 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, Size-Offset)); 320 } 321 case Intrinsic::bswap: 322 // bswap(bswap(x)) -> x 323 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) 324 if (Operand->getIntrinsicID() == Intrinsic::bswap) 325 return ReplaceInstUsesWith(CI, Operand->getArgOperand(0)); 326 327 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c)) 328 if (TruncInst *TI = dyn_cast<TruncInst>(II->getArgOperand(0))) { 329 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0))) 330 if (Operand->getIntrinsicID() == Intrinsic::bswap) { 331 unsigned C = Operand->getType()->getPrimitiveSizeInBits() - 332 TI->getType()->getPrimitiveSizeInBits(); 333 Value *CV = ConstantInt::get(Operand->getType(), C); 334 Value *V = Builder->CreateLShr(Operand->getArgOperand(0), CV); 335 return new TruncInst(V, TI->getType()); 336 } 337 } 338 339 break; 340 case Intrinsic::powi: 341 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 342 // powi(x, 0) -> 1.0 343 if (Power->isZero()) 344 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0)); 345 // powi(x, 1) -> x 346 if (Power->isOne()) 347 return ReplaceInstUsesWith(CI, II->getArgOperand(0)); 348 // powi(x, -1) -> 1/x 349 if (Power->isAllOnesValue()) 350 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0), 351 II->getArgOperand(0)); 352 } 353 break; 354 case Intrinsic::cttz: { 355 // If all bits below the first known one are known zero, 356 // this value is constant. 357 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType()); 358 uint32_t BitWidth = IT->getBitWidth(); 359 APInt KnownZero(BitWidth, 0); 360 APInt KnownOne(BitWidth, 0); 361 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth), 362 KnownZero, KnownOne); 363 unsigned TrailingZeros = KnownOne.countTrailingZeros(); 364 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros)); 365 if ((Mask & KnownZero) == Mask) 366 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 367 APInt(BitWidth, TrailingZeros))); 368 369 } 370 break; 371 case Intrinsic::ctlz: { 372 // If all bits above the first known one are known zero, 373 // this value is constant. 374 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType()); 375 uint32_t BitWidth = IT->getBitWidth(); 376 APInt KnownZero(BitWidth, 0); 377 APInt KnownOne(BitWidth, 0); 378 ComputeMaskedBits(II->getArgOperand(0), APInt::getAllOnesValue(BitWidth), 379 KnownZero, KnownOne); 380 unsigned LeadingZeros = KnownOne.countLeadingZeros(); 381 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros)); 382 if ((Mask & KnownZero) == Mask) 383 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 384 APInt(BitWidth, LeadingZeros))); 385 386 } 387 break; 388 case Intrinsic::uadd_with_overflow: { 389 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 390 const IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType()); 391 uint32_t BitWidth = IT->getBitWidth(); 392 APInt Mask = APInt::getSignBit(BitWidth); 393 APInt LHSKnownZero(BitWidth, 0); 394 APInt LHSKnownOne(BitWidth, 0); 395 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne); 396 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1]; 397 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1]; 398 399 if (LHSKnownNegative || LHSKnownPositive) { 400 APInt RHSKnownZero(BitWidth, 0); 401 APInt RHSKnownOne(BitWidth, 0); 402 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne); 403 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1]; 404 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1]; 405 if (LHSKnownNegative && RHSKnownNegative) { 406 // The sign bit is set in both cases: this MUST overflow. 407 // Create a simple add instruction, and insert it into the struct. 408 Instruction *Add = BinaryOperator::CreateAdd(LHS, RHS, "", &CI); 409 Worklist.Add(Add); 410 Constant *V[] = { 411 UndefValue::get(LHS->getType()),ConstantInt::getTrue(II->getContext()) 412 }; 413 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 414 return InsertValueInst::Create(Struct, Add, 0); 415 } 416 417 if (LHSKnownPositive && RHSKnownPositive) { 418 // The sign bit is clear in both cases: this CANNOT overflow. 419 // Create a simple add instruction, and insert it into the struct. 420 Instruction *Add = BinaryOperator::CreateNUWAdd(LHS, RHS, "", &CI); 421 Worklist.Add(Add); 422 Constant *V[] = { 423 UndefValue::get(LHS->getType()), 424 ConstantInt::getFalse(II->getContext()) 425 }; 426 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 427 return InsertValueInst::Create(Struct, Add, 0); 428 } 429 } 430 } 431 // FALL THROUGH uadd into sadd 432 case Intrinsic::sadd_with_overflow: 433 // Canonicalize constants into the RHS. 434 if (isa<Constant>(II->getArgOperand(0)) && 435 !isa<Constant>(II->getArgOperand(1))) { 436 Value *LHS = II->getArgOperand(0); 437 II->setArgOperand(0, II->getArgOperand(1)); 438 II->setArgOperand(1, LHS); 439 return II; 440 } 441 442 // X + undef -> undef 443 if (isa<UndefValue>(II->getArgOperand(1))) 444 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 445 446 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 447 // X + 0 -> {X, false} 448 if (RHS->isZero()) { 449 Constant *V[] = { 450 UndefValue::get(II->getArgOperand(0)->getType()), 451 ConstantInt::getFalse(II->getContext()) 452 }; 453 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 454 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 455 } 456 } 457 break; 458 case Intrinsic::usub_with_overflow: 459 case Intrinsic::ssub_with_overflow: 460 // undef - X -> undef 461 // X - undef -> undef 462 if (isa<UndefValue>(II->getArgOperand(0)) || 463 isa<UndefValue>(II->getArgOperand(1))) 464 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 465 466 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 467 // X - 0 -> {X, false} 468 if (RHS->isZero()) { 469 Constant *V[] = { 470 UndefValue::get(II->getArgOperand(0)->getType()), 471 ConstantInt::getFalse(II->getContext()) 472 }; 473 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 474 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 475 } 476 } 477 break; 478 case Intrinsic::umul_with_overflow: { 479 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 480 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth(); 481 APInt Mask = APInt::getAllOnesValue(BitWidth); 482 483 APInt LHSKnownZero(BitWidth, 0); 484 APInt LHSKnownOne(BitWidth, 0); 485 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne); 486 APInt RHSKnownZero(BitWidth, 0); 487 APInt RHSKnownOne(BitWidth, 0); 488 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne); 489 490 // Get the largest possible values for each operand. 491 APInt LHSMax = ~LHSKnownZero; 492 APInt RHSMax = ~RHSKnownZero; 493 494 // If multiplying the maximum values does not overflow then we can turn 495 // this into a plain NUW mul. 496 bool Overflow; 497 LHSMax.umul_ov(RHSMax, Overflow); 498 if (!Overflow) { 499 Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow"); 500 Constant *V[] = { 501 UndefValue::get(LHS->getType()), 502 Builder->getFalse() 503 }; 504 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 505 return InsertValueInst::Create(Struct, Mul, 0); 506 } 507 } // FALL THROUGH 508 case Intrinsic::smul_with_overflow: 509 // Canonicalize constants into the RHS. 510 if (isa<Constant>(II->getArgOperand(0)) && 511 !isa<Constant>(II->getArgOperand(1))) { 512 Value *LHS = II->getArgOperand(0); 513 II->setArgOperand(0, II->getArgOperand(1)); 514 II->setArgOperand(1, LHS); 515 return II; 516 } 517 518 // X * undef -> undef 519 if (isa<UndefValue>(II->getArgOperand(1))) 520 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 521 522 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 523 // X*0 -> {0, false} 524 if (RHSI->isZero()) 525 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType())); 526 527 // X * 1 -> {X, false} 528 if (RHSI->equalsInt(1)) { 529 Constant *V[] = { 530 UndefValue::get(II->getArgOperand(0)->getType()), 531 ConstantInt::getFalse(II->getContext()) 532 }; 533 Constant *Struct = ConstantStruct::get(II->getContext(), V, 2, false); 534 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 535 } 536 } 537 break; 538 case Intrinsic::ppc_altivec_lvx: 539 case Intrinsic::ppc_altivec_lvxl: 540 // Turn PPC lvx -> load if the pointer is known aligned. 541 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 542 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), 543 PointerType::getUnqual(II->getType())); 544 return new LoadInst(Ptr); 545 } 546 break; 547 case Intrinsic::ppc_altivec_stvx: 548 case Intrinsic::ppc_altivec_stvxl: 549 // Turn stvx -> store if the pointer is known aligned. 550 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) { 551 const Type *OpPtrTy = 552 PointerType::getUnqual(II->getArgOperand(0)->getType()); 553 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); 554 return new StoreInst(II->getArgOperand(0), Ptr); 555 } 556 break; 557 case Intrinsic::x86_sse_storeu_ps: 558 case Intrinsic::x86_sse2_storeu_pd: 559 case Intrinsic::x86_sse2_storeu_dq: 560 // Turn X86 storeu -> store if the pointer is known aligned. 561 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 562 const Type *OpPtrTy = 563 PointerType::getUnqual(II->getArgOperand(1)->getType()); 564 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy); 565 return new StoreInst(II->getArgOperand(1), Ptr); 566 } 567 break; 568 569 case Intrinsic::x86_sse_cvtss2si: 570 case Intrinsic::x86_sse_cvtss2si64: 571 case Intrinsic::x86_sse_cvttss2si: 572 case Intrinsic::x86_sse_cvttss2si64: 573 case Intrinsic::x86_sse2_cvtsd2si: 574 case Intrinsic::x86_sse2_cvtsd2si64: 575 case Intrinsic::x86_sse2_cvttsd2si: 576 case Intrinsic::x86_sse2_cvttsd2si64: { 577 // These intrinsics only demand the 0th element of their input vectors. If 578 // we can simplify the input based on that, do so now. 579 unsigned VWidth = 580 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 581 APInt DemandedElts(VWidth, 1); 582 APInt UndefElts(VWidth, 0); 583 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0), 584 DemandedElts, UndefElts)) { 585 II->setArgOperand(0, V); 586 return II; 587 } 588 break; 589 } 590 591 592 case Intrinsic::x86_sse41_pmovsxbw: 593 case Intrinsic::x86_sse41_pmovsxwd: 594 case Intrinsic::x86_sse41_pmovsxdq: 595 case Intrinsic::x86_sse41_pmovzxbw: 596 case Intrinsic::x86_sse41_pmovzxwd: 597 case Intrinsic::x86_sse41_pmovzxdq: { 598 unsigned VWidth = 599 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 600 unsigned LowHalfElts = VWidth / 2; 601 APInt InputDemandedElts(VWidth, 0); 602 InputDemandedElts = InputDemandedElts.getBitsSet(VWidth, 0, LowHalfElts); 603 APInt UndefElts(VWidth, 0); 604 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), 605 InputDemandedElts, 606 UndefElts)) { 607 II->setArgOperand(0, TmpV); 608 return II; 609 } 610 break; 611 } 612 613 case Intrinsic::ppc_altivec_vperm: 614 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant. 615 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getArgOperand(2))) { 616 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!"); 617 618 // Check that all of the elements are integer constants or undefs. 619 bool AllEltsOk = true; 620 for (unsigned i = 0; i != 16; ++i) { 621 if (!isa<ConstantInt>(Mask->getOperand(i)) && 622 !isa<UndefValue>(Mask->getOperand(i))) { 623 AllEltsOk = false; 624 break; 625 } 626 } 627 628 if (AllEltsOk) { 629 // Cast the input vectors to byte vectors. 630 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0), 631 Mask->getType()); 632 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1), 633 Mask->getType()); 634 Value *Result = UndefValue::get(Op0->getType()); 635 636 // Only extract each element once. 637 Value *ExtractedElts[32]; 638 memset(ExtractedElts, 0, sizeof(ExtractedElts)); 639 640 for (unsigned i = 0; i != 16; ++i) { 641 if (isa<UndefValue>(Mask->getOperand(i))) 642 continue; 643 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue(); 644 Idx &= 31; // Match the hardware behavior. 645 646 if (ExtractedElts[Idx] == 0) { 647 ExtractedElts[Idx] = 648 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1, 649 ConstantInt::get(Type::getInt32Ty(II->getContext()), 650 Idx&15, false), "tmp"); 651 } 652 653 // Insert this value into the result vector. 654 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx], 655 ConstantInt::get(Type::getInt32Ty(II->getContext()), 656 i, false), "tmp"); 657 } 658 return CastInst::Create(Instruction::BitCast, Result, CI.getType()); 659 } 660 } 661 break; 662 663 case Intrinsic::arm_neon_vld1: 664 case Intrinsic::arm_neon_vld2: 665 case Intrinsic::arm_neon_vld3: 666 case Intrinsic::arm_neon_vld4: 667 case Intrinsic::arm_neon_vld2lane: 668 case Intrinsic::arm_neon_vld3lane: 669 case Intrinsic::arm_neon_vld4lane: 670 case Intrinsic::arm_neon_vst1: 671 case Intrinsic::arm_neon_vst2: 672 case Intrinsic::arm_neon_vst3: 673 case Intrinsic::arm_neon_vst4: 674 case Intrinsic::arm_neon_vst2lane: 675 case Intrinsic::arm_neon_vst3lane: 676 case Intrinsic::arm_neon_vst4lane: { 677 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD); 678 unsigned AlignArg = II->getNumArgOperands() - 1; 679 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg)); 680 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) { 681 II->setArgOperand(AlignArg, 682 ConstantInt::get(Type::getInt32Ty(II->getContext()), 683 MemAlign, false)); 684 return II; 685 } 686 break; 687 } 688 689 case Intrinsic::stackrestore: { 690 // If the save is right next to the restore, remove the restore. This can 691 // happen when variable allocas are DCE'd. 692 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) { 693 if (SS->getIntrinsicID() == Intrinsic::stacksave) { 694 BasicBlock::iterator BI = SS; 695 if (&*++BI == II) 696 return EraseInstFromFunction(CI); 697 } 698 } 699 700 // Scan down this block to see if there is another stack restore in the 701 // same block without an intervening call/alloca. 702 BasicBlock::iterator BI = II; 703 TerminatorInst *TI = II->getParent()->getTerminator(); 704 bool CannotRemove = false; 705 for (++BI; &*BI != TI; ++BI) { 706 if (isa<AllocaInst>(BI) || isMalloc(BI)) { 707 CannotRemove = true; 708 break; 709 } 710 if (CallInst *BCI = dyn_cast<CallInst>(BI)) { 711 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) { 712 // If there is a stackrestore below this one, remove this one. 713 if (II->getIntrinsicID() == Intrinsic::stackrestore) 714 return EraseInstFromFunction(CI); 715 // Otherwise, ignore the intrinsic. 716 } else { 717 // If we found a non-intrinsic call, we can't remove the stack 718 // restore. 719 CannotRemove = true; 720 break; 721 } 722 } 723 } 724 725 // If the stack restore is in a return/unwind block and if there are no 726 // allocas or calls between the restore and the return, nuke the restore. 727 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI))) 728 return EraseInstFromFunction(CI); 729 break; 730 } 731 } 732 733 return visitCallSite(II); 734} 735 736// InvokeInst simplification 737// 738Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) { 739 return visitCallSite(&II); 740} 741 742/// isSafeToEliminateVarargsCast - If this cast does not affect the value 743/// passed through the varargs area, we can eliminate the use of the cast. 744static bool isSafeToEliminateVarargsCast(const CallSite CS, 745 const CastInst * const CI, 746 const TargetData * const TD, 747 const int ix) { 748 if (!CI->isLosslessCast()) 749 return false; 750 751 // The size of ByVal arguments is derived from the type, so we 752 // can't change to a type with a different size. If the size were 753 // passed explicitly we could avoid this check. 754 if (!CS.paramHasAttr(ix, Attribute::ByVal)) 755 return true; 756 757 const Type* SrcTy = 758 cast<PointerType>(CI->getOperand(0)->getType())->getElementType(); 759 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType(); 760 if (!SrcTy->isSized() || !DstTy->isSized()) 761 return false; 762 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy)) 763 return false; 764 return true; 765} 766 767namespace { 768class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls { 769 InstCombiner *IC; 770protected: 771 void replaceCall(Value *With) { 772 NewInstruction = IC->ReplaceInstUsesWith(*CI, With); 773 } 774 bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const { 775 if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp)) 776 return true; 777 if (ConstantInt *SizeCI = 778 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) { 779 if (SizeCI->isAllOnesValue()) 780 return true; 781 if (isString) { 782 uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp)); 783 // If the length is 0 we don't know how long it is and so we can't 784 // remove the check. 785 if (Len == 0) return false; 786 return SizeCI->getZExtValue() >= Len; 787 } 788 if (ConstantInt *Arg = dyn_cast<ConstantInt>( 789 CI->getArgOperand(SizeArgOp))) 790 return SizeCI->getZExtValue() >= Arg->getZExtValue(); 791 } 792 return false; 793 } 794public: 795 InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { } 796 Instruction *NewInstruction; 797}; 798} // end anonymous namespace 799 800// Try to fold some different type of calls here. 801// Currently we're only working with the checking functions, memcpy_chk, 802// mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk, 803// strcat_chk and strncat_chk. 804Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) { 805 if (CI->getCalledFunction() == 0) return 0; 806 807 InstCombineFortifiedLibCalls Simplifier(this); 808 Simplifier.fold(CI, TD); 809 return Simplifier.NewInstruction; 810} 811 812// visitCallSite - Improvements for call and invoke instructions. 813// 814Instruction *InstCombiner::visitCallSite(CallSite CS) { 815 bool Changed = false; 816 817 // If the callee is a pointer to a function, attempt to move any casts to the 818 // arguments of the call/invoke. 819 Value *Callee = CS.getCalledValue(); 820 if (!isa<Function>(Callee) && transformConstExprCastCall(CS)) 821 return 0; 822 823 if (Function *CalleeF = dyn_cast<Function>(Callee)) 824 // If the call and callee calling conventions don't match, this call must 825 // be unreachable, as the call is undefined. 826 if (CalleeF->getCallingConv() != CS.getCallingConv() && 827 // Only do this for calls to a function with a body. A prototype may 828 // not actually end up matching the implementation's calling conv for a 829 // variety of reasons (e.g. it may be written in assembly). 830 !CalleeF->isDeclaration()) { 831 Instruction *OldCall = CS.getInstruction(); 832 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 833 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 834 OldCall); 835 // If OldCall dues not return void then replaceAllUsesWith undef. 836 // This allows ValueHandlers and custom metadata to adjust itself. 837 if (!OldCall->getType()->isVoidTy()) 838 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType())); 839 if (isa<CallInst>(OldCall)) 840 return EraseInstFromFunction(*OldCall); 841 842 // We cannot remove an invoke, because it would change the CFG, just 843 // change the callee to a null pointer. 844 cast<InvokeInst>(OldCall)->setCalledFunction( 845 Constant::getNullValue(CalleeF->getType())); 846 return 0; 847 } 848 849 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 850 // This instruction is not reachable, just remove it. We insert a store to 851 // undef so that we know that this code is not reachable, despite the fact 852 // that we can't modify the CFG here. 853 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 854 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 855 CS.getInstruction()); 856 857 // If CS does not return void then replaceAllUsesWith undef. 858 // This allows ValueHandlers and custom metadata to adjust itself. 859 if (!CS.getInstruction()->getType()->isVoidTy()) 860 ReplaceInstUsesWith(*CS.getInstruction(), 861 UndefValue::get(CS.getInstruction()->getType())); 862 863 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) { 864 // Don't break the CFG, insert a dummy cond branch. 865 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(), 866 ConstantInt::getTrue(Callee->getContext()), II); 867 } 868 return EraseInstFromFunction(*CS.getInstruction()); 869 } 870 871 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee)) 872 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0))) 873 if (In->getIntrinsicID() == Intrinsic::init_trampoline) 874 return transformCallThroughTrampoline(CS); 875 876 const PointerType *PTy = cast<PointerType>(Callee->getType()); 877 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 878 if (FTy->isVarArg()) { 879 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1); 880 // See if we can optimize any arguments passed through the varargs area of 881 // the call. 882 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(), 883 E = CS.arg_end(); I != E; ++I, ++ix) { 884 CastInst *CI = dyn_cast<CastInst>(*I); 885 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) { 886 *I = CI->getOperand(0); 887 Changed = true; 888 } 889 } 890 } 891 892 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) { 893 // Inline asm calls cannot throw - mark them 'nounwind'. 894 CS.setDoesNotThrow(); 895 Changed = true; 896 } 897 898 // Try to optimize the call if possible, we require TargetData for most of 899 // this. None of these calls are seen as possibly dead so go ahead and 900 // delete the instruction now. 901 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) { 902 Instruction *I = tryOptimizeCall(CI, TD); 903 // If we changed something return the result, etc. Otherwise let 904 // the fallthrough check. 905 if (I) return EraseInstFromFunction(*I); 906 } 907 908 return Changed ? CS.getInstruction() : 0; 909} 910 911// transformConstExprCastCall - If the callee is a constexpr cast of a function, 912// attempt to move the cast to the arguments of the call/invoke. 913// 914bool InstCombiner::transformConstExprCastCall(CallSite CS) { 915 Function *Callee = 916 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts()); 917 if (Callee == 0) 918 return false; 919 Instruction *Caller = CS.getInstruction(); 920 const AttrListPtr &CallerPAL = CS.getAttributes(); 921 922 // Okay, this is a cast from a function to a different type. Unless doing so 923 // would cause a type conversion of one of our arguments, change this call to 924 // be a direct call with arguments casted to the appropriate types. 925 // 926 const FunctionType *FT = Callee->getFunctionType(); 927 const Type *OldRetTy = Caller->getType(); 928 const Type *NewRetTy = FT->getReturnType(); 929 930 if (NewRetTy->isStructTy()) 931 return false; // TODO: Handle multiple return values. 932 933 // Check to see if we are changing the return type... 934 if (OldRetTy != NewRetTy) { 935 if (Callee->isDeclaration() && 936 // Conversion is ok if changing from one pointer type to another or from 937 // a pointer to an integer of the same size. 938 !((OldRetTy->isPointerTy() || !TD || 939 OldRetTy == TD->getIntPtrType(Caller->getContext())) && 940 (NewRetTy->isPointerTy() || !TD || 941 NewRetTy == TD->getIntPtrType(Caller->getContext())))) 942 return false; // Cannot transform this return value. 943 944 if (!Caller->use_empty() && 945 // void -> non-void is handled specially 946 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy)) 947 return false; // Cannot transform this return value. 948 949 if (!CallerPAL.isEmpty() && !Caller->use_empty()) { 950 Attributes RAttrs = CallerPAL.getRetAttributes(); 951 if (RAttrs & Attribute::typeIncompatible(NewRetTy)) 952 return false; // Attribute not compatible with transformed value. 953 } 954 955 // If the callsite is an invoke instruction, and the return value is used by 956 // a PHI node in a successor, we cannot change the return type of the call 957 // because there is no place to put the cast instruction (without breaking 958 // the critical edge). Bail out in this case. 959 if (!Caller->use_empty()) 960 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) 961 for (Value::use_iterator UI = II->use_begin(), E = II->use_end(); 962 UI != E; ++UI) 963 if (PHINode *PN = dyn_cast<PHINode>(*UI)) 964 if (PN->getParent() == II->getNormalDest() || 965 PN->getParent() == II->getUnwindDest()) 966 return false; 967 } 968 969 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin()); 970 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs); 971 972 CallSite::arg_iterator AI = CS.arg_begin(); 973 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) { 974 const Type *ParamTy = FT->getParamType(i); 975 const Type *ActTy = (*AI)->getType(); 976 977 if (!CastInst::isCastable(ActTy, ParamTy)) 978 return false; // Cannot transform this parameter value. 979 980 unsigned Attrs = CallerPAL.getParamAttributes(i + 1); 981 if (Attrs & Attribute::typeIncompatible(ParamTy)) 982 return false; // Attribute not compatible with transformed value. 983 984 // If the parameter is passed as a byval argument, then we have to have a 985 // sized type and the sized type has to have the same size as the old type. 986 if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) { 987 const PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy); 988 if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0) 989 return false; 990 991 const Type *CurElTy = cast<PointerType>(ActTy)->getElementType(); 992 if (TD->getTypeAllocSize(CurElTy) != 993 TD->getTypeAllocSize(ParamPTy->getElementType())) 994 return false; 995 } 996 997 // Converting from one pointer type to another or between a pointer and an 998 // integer of the same size is safe even if we do not have a body. 999 bool isConvertible = ActTy == ParamTy || 1000 (TD && ((ParamTy->isPointerTy() || 1001 ParamTy == TD->getIntPtrType(Caller->getContext())) && 1002 (ActTy->isPointerTy() || 1003 ActTy == TD->getIntPtrType(Caller->getContext())))); 1004 if (Callee->isDeclaration() && !isConvertible) return false; 1005 } 1006 1007 if (Callee->isDeclaration()) { 1008 // Do not delete arguments unless we have a function body. 1009 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg()) 1010 return false; 1011 1012 // If the callee is just a declaration, don't change the varargsness of the 1013 // call. We don't want to introduce a varargs call where one doesn't 1014 // already exist. 1015 const PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType()); 1016 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg()) 1017 return false; 1018 } 1019 1020 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && 1021 !CallerPAL.isEmpty()) 1022 // In this case we have more arguments than the new function type, but we 1023 // won't be dropping them. Check that these extra arguments have attributes 1024 // that are compatible with being a vararg call argument. 1025 for (unsigned i = CallerPAL.getNumSlots(); i; --i) { 1026 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams()) 1027 break; 1028 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs; 1029 if (PAttrs & Attribute::VarArgsIncompatible) 1030 return false; 1031 } 1032 1033 1034 // Okay, we decided that this is a safe thing to do: go ahead and start 1035 // inserting cast instructions as necessary. 1036 std::vector<Value*> Args; 1037 Args.reserve(NumActualArgs); 1038 SmallVector<AttributeWithIndex, 8> attrVec; 1039 attrVec.reserve(NumCommonArgs); 1040 1041 // Get any return attributes. 1042 Attributes RAttrs = CallerPAL.getRetAttributes(); 1043 1044 // If the return value is not being used, the type may not be compatible 1045 // with the existing attributes. Wipe out any problematic attributes. 1046 RAttrs &= ~Attribute::typeIncompatible(NewRetTy); 1047 1048 // Add the new return attributes. 1049 if (RAttrs) 1050 attrVec.push_back(AttributeWithIndex::get(0, RAttrs)); 1051 1052 AI = CS.arg_begin(); 1053 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) { 1054 const Type *ParamTy = FT->getParamType(i); 1055 if ((*AI)->getType() == ParamTy) { 1056 Args.push_back(*AI); 1057 } else { 1058 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, 1059 false, ParamTy, false); 1060 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp")); 1061 } 1062 1063 // Add any parameter attributes. 1064 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1065 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1066 } 1067 1068 // If the function takes more arguments than the call was taking, add them 1069 // now. 1070 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) 1071 Args.push_back(Constant::getNullValue(FT->getParamType(i))); 1072 1073 // If we are removing arguments to the function, emit an obnoxious warning. 1074 if (FT->getNumParams() < NumActualArgs) { 1075 if (!FT->isVarArg()) { 1076 errs() << "WARNING: While resolving call to function '" 1077 << Callee->getName() << "' arguments were dropped!\n"; 1078 } else { 1079 // Add all of the arguments in their promoted form to the arg list. 1080 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) { 1081 const Type *PTy = getPromotedType((*AI)->getType()); 1082 if (PTy != (*AI)->getType()) { 1083 // Must promote to pass through va_arg area! 1084 Instruction::CastOps opcode = 1085 CastInst::getCastOpcode(*AI, false, PTy, false); 1086 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp")); 1087 } else { 1088 Args.push_back(*AI); 1089 } 1090 1091 // Add any parameter attributes. 1092 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1093 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1094 } 1095 } 1096 } 1097 1098 if (Attributes FnAttrs = CallerPAL.getFnAttributes()) 1099 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); 1100 1101 if (NewRetTy->isVoidTy()) 1102 Caller->setName(""); // Void type should not have a name. 1103 1104 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(), 1105 attrVec.end()); 1106 1107 Instruction *NC; 1108 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1109 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(), 1110 Args.begin(), Args.end(), 1111 Caller->getName(), Caller); 1112 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv()); 1113 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL); 1114 } else { 1115 NC = CallInst::Create(Callee, Args.begin(), Args.end(), 1116 Caller->getName(), Caller); 1117 CallInst *CI = cast<CallInst>(Caller); 1118 if (CI->isTailCall()) 1119 cast<CallInst>(NC)->setTailCall(); 1120 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv()); 1121 cast<CallInst>(NC)->setAttributes(NewCallerPAL); 1122 } 1123 1124 // Insert a cast of the return type as necessary. 1125 Value *NV = NC; 1126 if (OldRetTy != NV->getType() && !Caller->use_empty()) { 1127 if (!NV->getType()->isVoidTy()) { 1128 Instruction::CastOps opcode = 1129 CastInst::getCastOpcode(NC, false, OldRetTy, false); 1130 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp"); 1131 1132 // If this is an invoke instruction, we should insert it after the first 1133 // non-phi, instruction in the normal successor block. 1134 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1135 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI(); 1136 InsertNewInstBefore(NC, *I); 1137 } else { 1138 // Otherwise, it's a call, just insert cast right after the call. 1139 InsertNewInstBefore(NC, *Caller); 1140 } 1141 Worklist.AddUsersToWorkList(*Caller); 1142 } else { 1143 NV = UndefValue::get(Caller->getType()); 1144 } 1145 } 1146 1147 if (!Caller->use_empty()) 1148 ReplaceInstUsesWith(*Caller, NV); 1149 1150 EraseInstFromFunction(*Caller); 1151 return true; 1152} 1153 1154// transformCallThroughTrampoline - Turn a call to a function created by the 1155// init_trampoline intrinsic into a direct call to the underlying function. 1156// 1157Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) { 1158 Value *Callee = CS.getCalledValue(); 1159 const PointerType *PTy = cast<PointerType>(Callee->getType()); 1160 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 1161 const AttrListPtr &Attrs = CS.getAttributes(); 1162 1163 // If the call already has the 'nest' attribute somewhere then give up - 1164 // otherwise 'nest' would occur twice after splicing in the chain. 1165 if (Attrs.hasAttrSomewhere(Attribute::Nest)) 1166 return 0; 1167 1168 IntrinsicInst *Tramp = 1169 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0)); 1170 1171 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts()); 1172 const PointerType *NestFPTy = cast<PointerType>(NestF->getType()); 1173 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType()); 1174 1175 const AttrListPtr &NestAttrs = NestF->getAttributes(); 1176 if (!NestAttrs.isEmpty()) { 1177 unsigned NestIdx = 1; 1178 const Type *NestTy = 0; 1179 Attributes NestAttr = Attribute::None; 1180 1181 // Look for a parameter marked with the 'nest' attribute. 1182 for (FunctionType::param_iterator I = NestFTy->param_begin(), 1183 E = NestFTy->param_end(); I != E; ++NestIdx, ++I) 1184 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) { 1185 // Record the parameter type and any other attributes. 1186 NestTy = *I; 1187 NestAttr = NestAttrs.getParamAttributes(NestIdx); 1188 break; 1189 } 1190 1191 if (NestTy) { 1192 Instruction *Caller = CS.getInstruction(); 1193 std::vector<Value*> NewArgs; 1194 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1); 1195 1196 SmallVector<AttributeWithIndex, 8> NewAttrs; 1197 NewAttrs.reserve(Attrs.getNumSlots() + 1); 1198 1199 // Insert the nest argument into the call argument list, which may 1200 // mean appending it. Likewise for attributes. 1201 1202 // Add any result attributes. 1203 if (Attributes Attr = Attrs.getRetAttributes()) 1204 NewAttrs.push_back(AttributeWithIndex::get(0, Attr)); 1205 1206 { 1207 unsigned Idx = 1; 1208 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 1209 do { 1210 if (Idx == NestIdx) { 1211 // Add the chain argument and attributes. 1212 Value *NestVal = Tramp->getArgOperand(2); 1213 if (NestVal->getType() != NestTy) 1214 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller); 1215 NewArgs.push_back(NestVal); 1216 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr)); 1217 } 1218 1219 if (I == E) 1220 break; 1221 1222 // Add the original argument and attributes. 1223 NewArgs.push_back(*I); 1224 if (Attributes Attr = Attrs.getParamAttributes(Idx)) 1225 NewAttrs.push_back 1226 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr)); 1227 1228 ++Idx, ++I; 1229 } while (1); 1230 } 1231 1232 // Add any function attributes. 1233 if (Attributes Attr = Attrs.getFnAttributes()) 1234 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr)); 1235 1236 // The trampoline may have been bitcast to a bogus type (FTy). 1237 // Handle this by synthesizing a new function type, equal to FTy 1238 // with the chain parameter inserted. 1239 1240 std::vector<const Type*> NewTypes; 1241 NewTypes.reserve(FTy->getNumParams()+1); 1242 1243 // Insert the chain's type into the list of parameter types, which may 1244 // mean appending it. 1245 { 1246 unsigned Idx = 1; 1247 FunctionType::param_iterator I = FTy->param_begin(), 1248 E = FTy->param_end(); 1249 1250 do { 1251 if (Idx == NestIdx) 1252 // Add the chain's type. 1253 NewTypes.push_back(NestTy); 1254 1255 if (I == E) 1256 break; 1257 1258 // Add the original type. 1259 NewTypes.push_back(*I); 1260 1261 ++Idx, ++I; 1262 } while (1); 1263 } 1264 1265 // Replace the trampoline call with a direct call. Let the generic 1266 // code sort out any function type mismatches. 1267 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes, 1268 FTy->isVarArg()); 1269 Constant *NewCallee = 1270 NestF->getType() == PointerType::getUnqual(NewFTy) ? 1271 NestF : ConstantExpr::getBitCast(NestF, 1272 PointerType::getUnqual(NewFTy)); 1273 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(), 1274 NewAttrs.end()); 1275 1276 Instruction *NewCaller; 1277 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1278 NewCaller = InvokeInst::Create(NewCallee, 1279 II->getNormalDest(), II->getUnwindDest(), 1280 NewArgs.begin(), NewArgs.end(), 1281 Caller->getName(), Caller); 1282 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv()); 1283 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL); 1284 } else { 1285 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(), 1286 Caller->getName(), Caller); 1287 if (cast<CallInst>(Caller)->isTailCall()) 1288 cast<CallInst>(NewCaller)->setTailCall(); 1289 cast<CallInst>(NewCaller)-> 1290 setCallingConv(cast<CallInst>(Caller)->getCallingConv()); 1291 cast<CallInst>(NewCaller)->setAttributes(NewPAL); 1292 } 1293 if (!Caller->getType()->isVoidTy()) 1294 ReplaceInstUsesWith(*Caller, NewCaller); 1295 Caller->eraseFromParent(); 1296 Worklist.Remove(Caller); 1297 return 0; 1298 } 1299 } 1300 1301 // Replace the trampoline call with a direct call. Since there is no 'nest' 1302 // parameter, there is no need to adjust the argument list. Let the generic 1303 // code sort out any function type mismatches. 1304 Constant *NewCallee = 1305 NestF->getType() == PTy ? NestF : 1306 ConstantExpr::getBitCast(NestF, PTy); 1307 CS.setCalledFunction(NewCallee); 1308 return CS.getInstruction(); 1309} 1310 1311