InstCombineCalls.cpp revision 973f72a29aeafb1fdc4f8dafc3f6c6651cbb0c99
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), KnownZero, KnownOne); 365 unsigned TrailingZeros = KnownOne.countTrailingZeros(); 366 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros)); 367 if ((Mask & KnownZero) == Mask) 368 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 369 APInt(BitWidth, TrailingZeros))); 370 371 } 372 break; 373 case Intrinsic::ctlz: { 374 // If all bits above the first known one are known zero, 375 // this value is constant. 376 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType()); 377 // FIXME: Try to simplify vectors of integers. 378 if (!IT) break; 379 uint32_t BitWidth = IT->getBitWidth(); 380 APInt KnownZero(BitWidth, 0); 381 APInt KnownOne(BitWidth, 0); 382 ComputeMaskedBits(II->getArgOperand(0), KnownZero, KnownOne); 383 unsigned LeadingZeros = KnownOne.countLeadingZeros(); 384 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros)); 385 if ((Mask & KnownZero) == Mask) 386 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 387 APInt(BitWidth, LeadingZeros))); 388 389 } 390 break; 391 case Intrinsic::uadd_with_overflow: { 392 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 393 IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType()); 394 uint32_t BitWidth = IT->getBitWidth(); 395 APInt LHSKnownZero(BitWidth, 0); 396 APInt LHSKnownOne(BitWidth, 0); 397 ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne); 398 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1]; 399 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1]; 400 401 if (LHSKnownNegative || LHSKnownPositive) { 402 APInt RHSKnownZero(BitWidth, 0); 403 APInt RHSKnownOne(BitWidth, 0); 404 ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne); 405 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1]; 406 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1]; 407 if (LHSKnownNegative && RHSKnownNegative) { 408 // The sign bit is set in both cases: this MUST overflow. 409 // Create a simple add instruction, and insert it into the struct. 410 Value *Add = Builder->CreateAdd(LHS, RHS); 411 Add->takeName(&CI); 412 Constant *V[] = { 413 UndefValue::get(LHS->getType()), 414 ConstantInt::getTrue(II->getContext()) 415 }; 416 StructType *ST = cast<StructType>(II->getType()); 417 Constant *Struct = ConstantStruct::get(ST, V); 418 return InsertValueInst::Create(Struct, Add, 0); 419 } 420 421 if (LHSKnownPositive && RHSKnownPositive) { 422 // The sign bit is clear in both cases: this CANNOT overflow. 423 // Create a simple add instruction, and insert it into the struct. 424 Value *Add = Builder->CreateNUWAdd(LHS, RHS); 425 Add->takeName(&CI); 426 Constant *V[] = { 427 UndefValue::get(LHS->getType()), 428 ConstantInt::getFalse(II->getContext()) 429 }; 430 StructType *ST = cast<StructType>(II->getType()); 431 Constant *Struct = ConstantStruct::get(ST, V); 432 return InsertValueInst::Create(Struct, Add, 0); 433 } 434 } 435 } 436 // FALL THROUGH uadd into sadd 437 case Intrinsic::sadd_with_overflow: 438 // Canonicalize constants into the RHS. 439 if (isa<Constant>(II->getArgOperand(0)) && 440 !isa<Constant>(II->getArgOperand(1))) { 441 Value *LHS = II->getArgOperand(0); 442 II->setArgOperand(0, II->getArgOperand(1)); 443 II->setArgOperand(1, LHS); 444 return II; 445 } 446 447 // X + undef -> undef 448 if (isa<UndefValue>(II->getArgOperand(1))) 449 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 450 451 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 452 // X + 0 -> {X, false} 453 if (RHS->isZero()) { 454 Constant *V[] = { 455 UndefValue::get(II->getArgOperand(0)->getType()), 456 ConstantInt::getFalse(II->getContext()) 457 }; 458 Constant *Struct = 459 ConstantStruct::get(cast<StructType>(II->getType()), V); 460 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 461 } 462 } 463 break; 464 case Intrinsic::usub_with_overflow: 465 case Intrinsic::ssub_with_overflow: 466 // undef - X -> undef 467 // X - undef -> undef 468 if (isa<UndefValue>(II->getArgOperand(0)) || 469 isa<UndefValue>(II->getArgOperand(1))) 470 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 471 472 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 473 // X - 0 -> {X, false} 474 if (RHS->isZero()) { 475 Constant *V[] = { 476 UndefValue::get(II->getArgOperand(0)->getType()), 477 ConstantInt::getFalse(II->getContext()) 478 }; 479 Constant *Struct = 480 ConstantStruct::get(cast<StructType>(II->getType()), V); 481 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 482 } 483 } 484 break; 485 case Intrinsic::umul_with_overflow: { 486 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 487 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth(); 488 489 APInt LHSKnownZero(BitWidth, 0); 490 APInt LHSKnownOne(BitWidth, 0); 491 ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne); 492 APInt RHSKnownZero(BitWidth, 0); 493 APInt RHSKnownOne(BitWidth, 0); 494 ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne); 495 496 // Get the largest possible values for each operand. 497 APInt LHSMax = ~LHSKnownZero; 498 APInt RHSMax = ~RHSKnownZero; 499 500 // If multiplying the maximum values does not overflow then we can turn 501 // this into a plain NUW mul. 502 bool Overflow; 503 LHSMax.umul_ov(RHSMax, Overflow); 504 if (!Overflow) { 505 Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow"); 506 Constant *V[] = { 507 UndefValue::get(LHS->getType()), 508 Builder->getFalse() 509 }; 510 Constant *Struct = ConstantStruct::get(cast<StructType>(II->getType()),V); 511 return InsertValueInst::Create(Struct, Mul, 0); 512 } 513 } // FALL THROUGH 514 case Intrinsic::smul_with_overflow: 515 // Canonicalize constants into the RHS. 516 if (isa<Constant>(II->getArgOperand(0)) && 517 !isa<Constant>(II->getArgOperand(1))) { 518 Value *LHS = II->getArgOperand(0); 519 II->setArgOperand(0, II->getArgOperand(1)); 520 II->setArgOperand(1, LHS); 521 return II; 522 } 523 524 // X * undef -> undef 525 if (isa<UndefValue>(II->getArgOperand(1))) 526 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 527 528 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 529 // X*0 -> {0, false} 530 if (RHSI->isZero()) 531 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType())); 532 533 // X * 1 -> {X, false} 534 if (RHSI->equalsInt(1)) { 535 Constant *V[] = { 536 UndefValue::get(II->getArgOperand(0)->getType()), 537 ConstantInt::getFalse(II->getContext()) 538 }; 539 Constant *Struct = 540 ConstantStruct::get(cast<StructType>(II->getType()), V); 541 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 542 } 543 } 544 break; 545 case Intrinsic::ppc_altivec_lvx: 546 case Intrinsic::ppc_altivec_lvxl: 547 // Turn PPC lvx -> load if the pointer is known aligned. 548 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 549 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), 550 PointerType::getUnqual(II->getType())); 551 return new LoadInst(Ptr); 552 } 553 break; 554 case Intrinsic::ppc_altivec_stvx: 555 case Intrinsic::ppc_altivec_stvxl: 556 // Turn stvx -> store if the pointer is known aligned. 557 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) { 558 Type *OpPtrTy = 559 PointerType::getUnqual(II->getArgOperand(0)->getType()); 560 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); 561 return new StoreInst(II->getArgOperand(0), Ptr); 562 } 563 break; 564 case Intrinsic::x86_sse_storeu_ps: 565 case Intrinsic::x86_sse2_storeu_pd: 566 case Intrinsic::x86_sse2_storeu_dq: 567 // Turn X86 storeu -> store if the pointer is known aligned. 568 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 569 Type *OpPtrTy = 570 PointerType::getUnqual(II->getArgOperand(1)->getType()); 571 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy); 572 return new StoreInst(II->getArgOperand(1), Ptr); 573 } 574 break; 575 576 case Intrinsic::x86_sse_cvtss2si: 577 case Intrinsic::x86_sse_cvtss2si64: 578 case Intrinsic::x86_sse_cvttss2si: 579 case Intrinsic::x86_sse_cvttss2si64: 580 case Intrinsic::x86_sse2_cvtsd2si: 581 case Intrinsic::x86_sse2_cvtsd2si64: 582 case Intrinsic::x86_sse2_cvttsd2si: 583 case Intrinsic::x86_sse2_cvttsd2si64: { 584 // These intrinsics only demand the 0th element of their input vectors. If 585 // we can simplify the input based on that, do so now. 586 unsigned VWidth = 587 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 588 APInt DemandedElts(VWidth, 1); 589 APInt UndefElts(VWidth, 0); 590 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0), 591 DemandedElts, UndefElts)) { 592 II->setArgOperand(0, V); 593 return II; 594 } 595 break; 596 } 597 598 599 case Intrinsic::x86_sse41_pmovsxbw: 600 case Intrinsic::x86_sse41_pmovsxwd: 601 case Intrinsic::x86_sse41_pmovsxdq: 602 case Intrinsic::x86_sse41_pmovzxbw: 603 case Intrinsic::x86_sse41_pmovzxwd: 604 case Intrinsic::x86_sse41_pmovzxdq: { 605 // pmov{s|z}x ignores the upper half of their input vectors. 606 unsigned VWidth = 607 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 608 unsigned LowHalfElts = VWidth / 2; 609 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts)); 610 APInt UndefElts(VWidth, 0); 611 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), 612 InputDemandedElts, 613 UndefElts)) { 614 II->setArgOperand(0, TmpV); 615 return II; 616 } 617 break; 618 } 619 620 case Intrinsic::ppc_altivec_vperm: 621 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant. 622 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) { 623 assert(Mask->getType()->getVectorNumElements() == 16 && 624 "Bad type for intrinsic!"); 625 626 // Check that all of the elements are integer constants or undefs. 627 bool AllEltsOk = true; 628 for (unsigned i = 0; i != 16; ++i) { 629 Constant *Elt = Mask->getAggregateElement(i); 630 if (Elt == 0 || 631 !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) { 632 AllEltsOk = false; 633 break; 634 } 635 } 636 637 if (AllEltsOk) { 638 // Cast the input vectors to byte vectors. 639 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0), 640 Mask->getType()); 641 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1), 642 Mask->getType()); 643 Value *Result = UndefValue::get(Op0->getType()); 644 645 // Only extract each element once. 646 Value *ExtractedElts[32]; 647 memset(ExtractedElts, 0, sizeof(ExtractedElts)); 648 649 for (unsigned i = 0; i != 16; ++i) { 650 if (isa<UndefValue>(Mask->getAggregateElement(i))) 651 continue; 652 unsigned Idx = 653 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue(); 654 Idx &= 31; // Match the hardware behavior. 655 656 if (ExtractedElts[Idx] == 0) { 657 ExtractedElts[Idx] = 658 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1, 659 Builder->getInt32(Idx&15)); 660 } 661 662 // Insert this value into the result vector. 663 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx], 664 Builder->getInt32(i)); 665 } 666 return CastInst::Create(Instruction::BitCast, Result, CI.getType()); 667 } 668 } 669 break; 670 671 case Intrinsic::arm_neon_vld1: 672 case Intrinsic::arm_neon_vld2: 673 case Intrinsic::arm_neon_vld3: 674 case Intrinsic::arm_neon_vld4: 675 case Intrinsic::arm_neon_vld2lane: 676 case Intrinsic::arm_neon_vld3lane: 677 case Intrinsic::arm_neon_vld4lane: 678 case Intrinsic::arm_neon_vst1: 679 case Intrinsic::arm_neon_vst2: 680 case Intrinsic::arm_neon_vst3: 681 case Intrinsic::arm_neon_vst4: 682 case Intrinsic::arm_neon_vst2lane: 683 case Intrinsic::arm_neon_vst3lane: 684 case Intrinsic::arm_neon_vst4lane: { 685 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD); 686 unsigned AlignArg = II->getNumArgOperands() - 1; 687 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg)); 688 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) { 689 II->setArgOperand(AlignArg, 690 ConstantInt::get(Type::getInt32Ty(II->getContext()), 691 MemAlign, false)); 692 return II; 693 } 694 break; 695 } 696 697 case Intrinsic::arm_neon_vmulls: 698 case Intrinsic::arm_neon_vmullu: { 699 Value *Arg0 = II->getArgOperand(0); 700 Value *Arg1 = II->getArgOperand(1); 701 702 // Handle mul by zero first: 703 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) { 704 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType())); 705 } 706 707 // Check for constant LHS & RHS - in this case we just simplify. 708 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu); 709 VectorType *NewVT = cast<VectorType>(II->getType()); 710 unsigned NewWidth = NewVT->getElementType()->getIntegerBitWidth(); 711 if (ConstantDataVector *CV0 = dyn_cast<ConstantDataVector>(Arg0)) { 712 if (ConstantDataVector *CV1 = dyn_cast<ConstantDataVector>(Arg1)) { 713 VectorType* VT = cast<VectorType>(CV0->getType()); 714 SmallVector<Constant*, 4> NewElems; 715 for (unsigned i = 0; i < VT->getNumElements(); ++i) { 716 APInt CV0E = 717 (cast<ConstantInt>(CV0->getAggregateElement(i)))->getValue(); 718 CV0E = Zext ? CV0E.zext(NewWidth) : CV0E.sext(NewWidth); 719 APInt CV1E = 720 (cast<ConstantInt>(CV1->getAggregateElement(i)))->getValue(); 721 CV1E = Zext ? CV1E.zext(NewWidth) : CV1E.sext(NewWidth); 722 NewElems.push_back( 723 ConstantInt::get(NewVT->getElementType(), CV0E * CV1E)); 724 } 725 return ReplaceInstUsesWith(CI, ConstantVector::get(NewElems)); 726 } 727 728 // Couldn't simplify - cannonicalize constant to the RHS. 729 std::swap(Arg0, Arg1); 730 } 731 732 // Handle mul by one: 733 if (ConstantDataVector *CV1 = dyn_cast<ConstantDataVector>(Arg1)) { 734 if (ConstantInt *Splat = 735 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue())) { 736 if (Splat->isOne()) { 737 if (Zext) 738 return CastInst::CreateZExtOrBitCast(Arg0, II->getType()); 739 // else 740 return CastInst::CreateSExtOrBitCast(Arg0, II->getType()); 741 } 742 } 743 } 744 745 break; 746 } 747 748 case Intrinsic::stackrestore: { 749 // If the save is right next to the restore, remove the restore. This can 750 // happen when variable allocas are DCE'd. 751 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) { 752 if (SS->getIntrinsicID() == Intrinsic::stacksave) { 753 BasicBlock::iterator BI = SS; 754 if (&*++BI == II) 755 return EraseInstFromFunction(CI); 756 } 757 } 758 759 // Scan down this block to see if there is another stack restore in the 760 // same block without an intervening call/alloca. 761 BasicBlock::iterator BI = II; 762 TerminatorInst *TI = II->getParent()->getTerminator(); 763 bool CannotRemove = false; 764 for (++BI; &*BI != TI; ++BI) { 765 if (isa<AllocaInst>(BI) || isMalloc(BI)) { 766 CannotRemove = true; 767 break; 768 } 769 if (CallInst *BCI = dyn_cast<CallInst>(BI)) { 770 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) { 771 // If there is a stackrestore below this one, remove this one. 772 if (II->getIntrinsicID() == Intrinsic::stackrestore) 773 return EraseInstFromFunction(CI); 774 // Otherwise, ignore the intrinsic. 775 } else { 776 // If we found a non-intrinsic call, we can't remove the stack 777 // restore. 778 CannotRemove = true; 779 break; 780 } 781 } 782 } 783 784 // If the stack restore is in a return, resume, or unwind block and if there 785 // are no allocas or calls between the restore and the return, nuke the 786 // restore. 787 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI))) 788 return EraseInstFromFunction(CI); 789 break; 790 } 791 } 792 793 return visitCallSite(II); 794} 795 796// InvokeInst simplification 797// 798Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) { 799 return visitCallSite(&II); 800} 801 802/// isSafeToEliminateVarargsCast - If this cast does not affect the value 803/// passed through the varargs area, we can eliminate the use of the cast. 804static bool isSafeToEliminateVarargsCast(const CallSite CS, 805 const CastInst * const CI, 806 const TargetData * const TD, 807 const int ix) { 808 if (!CI->isLosslessCast()) 809 return false; 810 811 // The size of ByVal arguments is derived from the type, so we 812 // can't change to a type with a different size. If the size were 813 // passed explicitly we could avoid this check. 814 if (!CS.isByValArgument(ix)) 815 return true; 816 817 Type* SrcTy = 818 cast<PointerType>(CI->getOperand(0)->getType())->getElementType(); 819 Type* DstTy = cast<PointerType>(CI->getType())->getElementType(); 820 if (!SrcTy->isSized() || !DstTy->isSized()) 821 return false; 822 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy)) 823 return false; 824 return true; 825} 826 827namespace { 828class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls { 829 InstCombiner *IC; 830protected: 831 void replaceCall(Value *With) { 832 NewInstruction = IC->ReplaceInstUsesWith(*CI, With); 833 } 834 bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const { 835 if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp)) 836 return true; 837 if (ConstantInt *SizeCI = 838 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) { 839 if (SizeCI->isAllOnesValue()) 840 return true; 841 if (isString) { 842 uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp)); 843 // If the length is 0 we don't know how long it is and so we can't 844 // remove the check. 845 if (Len == 0) return false; 846 return SizeCI->getZExtValue() >= Len; 847 } 848 if (ConstantInt *Arg = dyn_cast<ConstantInt>( 849 CI->getArgOperand(SizeArgOp))) 850 return SizeCI->getZExtValue() >= Arg->getZExtValue(); 851 } 852 return false; 853 } 854public: 855 InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { } 856 Instruction *NewInstruction; 857}; 858} // end anonymous namespace 859 860// Try to fold some different type of calls here. 861// Currently we're only working with the checking functions, memcpy_chk, 862// mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk, 863// strcat_chk and strncat_chk. 864Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) { 865 if (CI->getCalledFunction() == 0) return 0; 866 867 InstCombineFortifiedLibCalls Simplifier(this); 868 Simplifier.fold(CI, TD); 869 return Simplifier.NewInstruction; 870} 871 872static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) { 873 // Strip off at most one level of pointer casts, looking for an alloca. This 874 // is good enough in practice and simpler than handling any number of casts. 875 Value *Underlying = TrampMem->stripPointerCasts(); 876 if (Underlying != TrampMem && 877 (!Underlying->hasOneUse() || *Underlying->use_begin() != TrampMem)) 878 return 0; 879 if (!isa<AllocaInst>(Underlying)) 880 return 0; 881 882 IntrinsicInst *InitTrampoline = 0; 883 for (Value::use_iterator I = TrampMem->use_begin(), E = TrampMem->use_end(); 884 I != E; I++) { 885 IntrinsicInst *II = dyn_cast<IntrinsicInst>(*I); 886 if (!II) 887 return 0; 888 if (II->getIntrinsicID() == Intrinsic::init_trampoline) { 889 if (InitTrampoline) 890 // More than one init_trampoline writes to this value. Give up. 891 return 0; 892 InitTrampoline = II; 893 continue; 894 } 895 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline) 896 // Allow any number of calls to adjust.trampoline. 897 continue; 898 return 0; 899 } 900 901 // No call to init.trampoline found. 902 if (!InitTrampoline) 903 return 0; 904 905 // Check that the alloca is being used in the expected way. 906 if (InitTrampoline->getOperand(0) != TrampMem) 907 return 0; 908 909 return InitTrampoline; 910} 911 912static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp, 913 Value *TrampMem) { 914 // Visit all the previous instructions in the basic block, and try to find a 915 // init.trampoline which has a direct path to the adjust.trampoline. 916 for (BasicBlock::iterator I = AdjustTramp, 917 E = AdjustTramp->getParent()->begin(); I != E; ) { 918 Instruction *Inst = --I; 919 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) 920 if (II->getIntrinsicID() == Intrinsic::init_trampoline && 921 II->getOperand(0) == TrampMem) 922 return II; 923 if (Inst->mayWriteToMemory()) 924 return 0; 925 } 926 return 0; 927} 928 929// Given a call to llvm.adjust.trampoline, find and return the corresponding 930// call to llvm.init.trampoline if the call to the trampoline can be optimized 931// to a direct call to a function. Otherwise return NULL. 932// 933static IntrinsicInst *FindInitTrampoline(Value *Callee) { 934 Callee = Callee->stripPointerCasts(); 935 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee); 936 if (!AdjustTramp || 937 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline) 938 return 0; 939 940 Value *TrampMem = AdjustTramp->getOperand(0); 941 942 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem)) 943 return IT; 944 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem)) 945 return IT; 946 return 0; 947} 948 949// visitCallSite - Improvements for call and invoke instructions. 950// 951Instruction *InstCombiner::visitCallSite(CallSite CS) { 952 bool Changed = false; 953 954 // If the callee is a pointer to a function, attempt to move any casts to the 955 // arguments of the call/invoke. 956 Value *Callee = CS.getCalledValue(); 957 if (!isa<Function>(Callee) && transformConstExprCastCall(CS)) 958 return 0; 959 960 if (Function *CalleeF = dyn_cast<Function>(Callee)) 961 // If the call and callee calling conventions don't match, this call must 962 // be unreachable, as the call is undefined. 963 if (CalleeF->getCallingConv() != CS.getCallingConv() && 964 // Only do this for calls to a function with a body. A prototype may 965 // not actually end up matching the implementation's calling conv for a 966 // variety of reasons (e.g. it may be written in assembly). 967 !CalleeF->isDeclaration()) { 968 Instruction *OldCall = CS.getInstruction(); 969 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 970 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 971 OldCall); 972 // If OldCall dues not return void then replaceAllUsesWith undef. 973 // This allows ValueHandlers and custom metadata to adjust itself. 974 if (!OldCall->getType()->isVoidTy()) 975 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType())); 976 if (isa<CallInst>(OldCall)) 977 return EraseInstFromFunction(*OldCall); 978 979 // We cannot remove an invoke, because it would change the CFG, just 980 // change the callee to a null pointer. 981 cast<InvokeInst>(OldCall)->setCalledFunction( 982 Constant::getNullValue(CalleeF->getType())); 983 return 0; 984 } 985 986 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 987 // This instruction is not reachable, just remove it. We insert a store to 988 // undef so that we know that this code is not reachable, despite the fact 989 // that we can't modify the CFG here. 990 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 991 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 992 CS.getInstruction()); 993 994 // If CS does not return void then replaceAllUsesWith undef. 995 // This allows ValueHandlers and custom metadata to adjust itself. 996 if (!CS.getInstruction()->getType()->isVoidTy()) 997 ReplaceInstUsesWith(*CS.getInstruction(), 998 UndefValue::get(CS.getInstruction()->getType())); 999 1000 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) { 1001 // Don't break the CFG, insert a dummy cond branch. 1002 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(), 1003 ConstantInt::getTrue(Callee->getContext()), II); 1004 } 1005 return EraseInstFromFunction(*CS.getInstruction()); 1006 } 1007 1008 if (IntrinsicInst *II = FindInitTrampoline(Callee)) 1009 return transformCallThroughTrampoline(CS, II); 1010 1011 PointerType *PTy = cast<PointerType>(Callee->getType()); 1012 FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 1013 if (FTy->isVarArg()) { 1014 int ix = FTy->getNumParams(); 1015 // See if we can optimize any arguments passed through the varargs area of 1016 // the call. 1017 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(), 1018 E = CS.arg_end(); I != E; ++I, ++ix) { 1019 CastInst *CI = dyn_cast<CastInst>(*I); 1020 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) { 1021 *I = CI->getOperand(0); 1022 Changed = true; 1023 } 1024 } 1025 } 1026 1027 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) { 1028 // Inline asm calls cannot throw - mark them 'nounwind'. 1029 CS.setDoesNotThrow(); 1030 Changed = true; 1031 } 1032 1033 // Try to optimize the call if possible, we require TargetData for most of 1034 // this. None of these calls are seen as possibly dead so go ahead and 1035 // delete the instruction now. 1036 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) { 1037 Instruction *I = tryOptimizeCall(CI, TD); 1038 // If we changed something return the result, etc. Otherwise let 1039 // the fallthrough check. 1040 if (I) return EraseInstFromFunction(*I); 1041 } 1042 1043 return Changed ? CS.getInstruction() : 0; 1044} 1045 1046// transformConstExprCastCall - If the callee is a constexpr cast of a function, 1047// attempt to move the cast to the arguments of the call/invoke. 1048// 1049bool InstCombiner::transformConstExprCastCall(CallSite CS) { 1050 Function *Callee = 1051 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts()); 1052 if (Callee == 0) 1053 return false; 1054 Instruction *Caller = CS.getInstruction(); 1055 const AttrListPtr &CallerPAL = CS.getAttributes(); 1056 1057 // Okay, this is a cast from a function to a different type. Unless doing so 1058 // would cause a type conversion of one of our arguments, change this call to 1059 // be a direct call with arguments casted to the appropriate types. 1060 // 1061 FunctionType *FT = Callee->getFunctionType(); 1062 Type *OldRetTy = Caller->getType(); 1063 Type *NewRetTy = FT->getReturnType(); 1064 1065 if (NewRetTy->isStructTy()) 1066 return false; // TODO: Handle multiple return values. 1067 1068 // Check to see if we are changing the return type... 1069 if (OldRetTy != NewRetTy) { 1070 if (Callee->isDeclaration() && 1071 // Conversion is ok if changing from one pointer type to another or from 1072 // a pointer to an integer of the same size. 1073 !((OldRetTy->isPointerTy() || !TD || 1074 OldRetTy == TD->getIntPtrType(Caller->getContext())) && 1075 (NewRetTy->isPointerTy() || !TD || 1076 NewRetTy == TD->getIntPtrType(Caller->getContext())))) 1077 return false; // Cannot transform this return value. 1078 1079 if (!Caller->use_empty() && 1080 // void -> non-void is handled specially 1081 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy)) 1082 return false; // Cannot transform this return value. 1083 1084 if (!CallerPAL.isEmpty() && !Caller->use_empty()) { 1085 Attributes RAttrs = CallerPAL.getRetAttributes(); 1086 if (RAttrs & Attribute::typeIncompatible(NewRetTy)) 1087 return false; // Attribute not compatible with transformed value. 1088 } 1089 1090 // If the callsite is an invoke instruction, and the return value is used by 1091 // a PHI node in a successor, we cannot change the return type of the call 1092 // because there is no place to put the cast instruction (without breaking 1093 // the critical edge). Bail out in this case. 1094 if (!Caller->use_empty()) 1095 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) 1096 for (Value::use_iterator UI = II->use_begin(), E = II->use_end(); 1097 UI != E; ++UI) 1098 if (PHINode *PN = dyn_cast<PHINode>(*UI)) 1099 if (PN->getParent() == II->getNormalDest() || 1100 PN->getParent() == II->getUnwindDest()) 1101 return false; 1102 } 1103 1104 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin()); 1105 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs); 1106 1107 CallSite::arg_iterator AI = CS.arg_begin(); 1108 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) { 1109 Type *ParamTy = FT->getParamType(i); 1110 Type *ActTy = (*AI)->getType(); 1111 1112 if (!CastInst::isCastable(ActTy, ParamTy)) 1113 return false; // Cannot transform this parameter value. 1114 1115 Attributes Attrs = CallerPAL.getParamAttributes(i + 1); 1116 if (Attrs & Attribute::typeIncompatible(ParamTy)) 1117 return false; // Attribute not compatible with transformed value. 1118 1119 // If the parameter is passed as a byval argument, then we have to have a 1120 // sized type and the sized type has to have the same size as the old type. 1121 if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) { 1122 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy); 1123 if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0) 1124 return false; 1125 1126 Type *CurElTy = cast<PointerType>(ActTy)->getElementType(); 1127 if (TD->getTypeAllocSize(CurElTy) != 1128 TD->getTypeAllocSize(ParamPTy->getElementType())) 1129 return false; 1130 } 1131 1132 // Converting from one pointer type to another or between a pointer and an 1133 // integer of the same size is safe even if we do not have a body. 1134 bool isConvertible = ActTy == ParamTy || 1135 (TD && ((ParamTy->isPointerTy() || 1136 ParamTy == TD->getIntPtrType(Caller->getContext())) && 1137 (ActTy->isPointerTy() || 1138 ActTy == TD->getIntPtrType(Caller->getContext())))); 1139 if (Callee->isDeclaration() && !isConvertible) return false; 1140 } 1141 1142 if (Callee->isDeclaration()) { 1143 // Do not delete arguments unless we have a function body. 1144 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg()) 1145 return false; 1146 1147 // If the callee is just a declaration, don't change the varargsness of the 1148 // call. We don't want to introduce a varargs call where one doesn't 1149 // already exist. 1150 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType()); 1151 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg()) 1152 return false; 1153 1154 // If both the callee and the cast type are varargs, we still have to make 1155 // sure the number of fixed parameters are the same or we have the same 1156 // ABI issues as if we introduce a varargs call. 1157 if (FT->isVarArg() && 1158 cast<FunctionType>(APTy->getElementType())->isVarArg() && 1159 FT->getNumParams() != 1160 cast<FunctionType>(APTy->getElementType())->getNumParams()) 1161 return false; 1162 } 1163 1164 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && 1165 !CallerPAL.isEmpty()) 1166 // In this case we have more arguments than the new function type, but we 1167 // won't be dropping them. Check that these extra arguments have attributes 1168 // that are compatible with being a vararg call argument. 1169 for (unsigned i = CallerPAL.getNumSlots(); i; --i) { 1170 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams()) 1171 break; 1172 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs; 1173 if (PAttrs & Attribute::VarArgsIncompatible) 1174 return false; 1175 } 1176 1177 1178 // Okay, we decided that this is a safe thing to do: go ahead and start 1179 // inserting cast instructions as necessary. 1180 std::vector<Value*> Args; 1181 Args.reserve(NumActualArgs); 1182 SmallVector<AttributeWithIndex, 8> attrVec; 1183 attrVec.reserve(NumCommonArgs); 1184 1185 // Get any return attributes. 1186 Attributes RAttrs = CallerPAL.getRetAttributes(); 1187 1188 // If the return value is not being used, the type may not be compatible 1189 // with the existing attributes. Wipe out any problematic attributes. 1190 RAttrs &= ~Attribute::typeIncompatible(NewRetTy); 1191 1192 // Add the new return attributes. 1193 if (RAttrs) 1194 attrVec.push_back(AttributeWithIndex::get(0, RAttrs)); 1195 1196 AI = CS.arg_begin(); 1197 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) { 1198 Type *ParamTy = FT->getParamType(i); 1199 if ((*AI)->getType() == ParamTy) { 1200 Args.push_back(*AI); 1201 } else { 1202 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, 1203 false, ParamTy, false); 1204 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy)); 1205 } 1206 1207 // Add any parameter attributes. 1208 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1209 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1210 } 1211 1212 // If the function takes more arguments than the call was taking, add them 1213 // now. 1214 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) 1215 Args.push_back(Constant::getNullValue(FT->getParamType(i))); 1216 1217 // If we are removing arguments to the function, emit an obnoxious warning. 1218 if (FT->getNumParams() < NumActualArgs) { 1219 if (!FT->isVarArg()) { 1220 errs() << "WARNING: While resolving call to function '" 1221 << Callee->getName() << "' arguments were dropped!\n"; 1222 } else { 1223 // Add all of the arguments in their promoted form to the arg list. 1224 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) { 1225 Type *PTy = getPromotedType((*AI)->getType()); 1226 if (PTy != (*AI)->getType()) { 1227 // Must promote to pass through va_arg area! 1228 Instruction::CastOps opcode = 1229 CastInst::getCastOpcode(*AI, false, PTy, false); 1230 Args.push_back(Builder->CreateCast(opcode, *AI, PTy)); 1231 } else { 1232 Args.push_back(*AI); 1233 } 1234 1235 // Add any parameter attributes. 1236 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1237 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1238 } 1239 } 1240 } 1241 1242 if (Attributes FnAttrs = CallerPAL.getFnAttributes()) 1243 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); 1244 1245 if (NewRetTy->isVoidTy()) 1246 Caller->setName(""); // Void type should not have a name. 1247 1248 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(), 1249 attrVec.end()); 1250 1251 Instruction *NC; 1252 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1253 NC = Builder->CreateInvoke(Callee, II->getNormalDest(), 1254 II->getUnwindDest(), Args); 1255 NC->takeName(II); 1256 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv()); 1257 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL); 1258 } else { 1259 CallInst *CI = cast<CallInst>(Caller); 1260 NC = Builder->CreateCall(Callee, Args); 1261 NC->takeName(CI); 1262 if (CI->isTailCall()) 1263 cast<CallInst>(NC)->setTailCall(); 1264 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv()); 1265 cast<CallInst>(NC)->setAttributes(NewCallerPAL); 1266 } 1267 1268 // Insert a cast of the return type as necessary. 1269 Value *NV = NC; 1270 if (OldRetTy != NV->getType() && !Caller->use_empty()) { 1271 if (!NV->getType()->isVoidTy()) { 1272 Instruction::CastOps opcode = 1273 CastInst::getCastOpcode(NC, false, OldRetTy, false); 1274 NV = NC = CastInst::Create(opcode, NC, OldRetTy); 1275 NC->setDebugLoc(Caller->getDebugLoc()); 1276 1277 // If this is an invoke instruction, we should insert it after the first 1278 // non-phi, instruction in the normal successor block. 1279 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1280 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt(); 1281 InsertNewInstBefore(NC, *I); 1282 } else { 1283 // Otherwise, it's a call, just insert cast right after the call. 1284 InsertNewInstBefore(NC, *Caller); 1285 } 1286 Worklist.AddUsersToWorkList(*Caller); 1287 } else { 1288 NV = UndefValue::get(Caller->getType()); 1289 } 1290 } 1291 1292 if (!Caller->use_empty()) 1293 ReplaceInstUsesWith(*Caller, NV); 1294 1295 EraseInstFromFunction(*Caller); 1296 return true; 1297} 1298 1299// transformCallThroughTrampoline - Turn a call to a function created by 1300// init_trampoline / adjust_trampoline intrinsic pair into a direct call to the 1301// underlying function. 1302// 1303Instruction * 1304InstCombiner::transformCallThroughTrampoline(CallSite CS, 1305 IntrinsicInst *Tramp) { 1306 Value *Callee = CS.getCalledValue(); 1307 PointerType *PTy = cast<PointerType>(Callee->getType()); 1308 FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 1309 const AttrListPtr &Attrs = CS.getAttributes(); 1310 1311 // If the call already has the 'nest' attribute somewhere then give up - 1312 // otherwise 'nest' would occur twice after splicing in the chain. 1313 if (Attrs.hasAttrSomewhere(Attribute::Nest)) 1314 return 0; 1315 1316 assert(Tramp && 1317 "transformCallThroughTrampoline called with incorrect CallSite."); 1318 1319 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts()); 1320 PointerType *NestFPTy = cast<PointerType>(NestF->getType()); 1321 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType()); 1322 1323 const AttrListPtr &NestAttrs = NestF->getAttributes(); 1324 if (!NestAttrs.isEmpty()) { 1325 unsigned NestIdx = 1; 1326 Type *NestTy = 0; 1327 Attributes NestAttr = Attribute::None; 1328 1329 // Look for a parameter marked with the 'nest' attribute. 1330 for (FunctionType::param_iterator I = NestFTy->param_begin(), 1331 E = NestFTy->param_end(); I != E; ++NestIdx, ++I) 1332 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) { 1333 // Record the parameter type and any other attributes. 1334 NestTy = *I; 1335 NestAttr = NestAttrs.getParamAttributes(NestIdx); 1336 break; 1337 } 1338 1339 if (NestTy) { 1340 Instruction *Caller = CS.getInstruction(); 1341 std::vector<Value*> NewArgs; 1342 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1); 1343 1344 SmallVector<AttributeWithIndex, 8> NewAttrs; 1345 NewAttrs.reserve(Attrs.getNumSlots() + 1); 1346 1347 // Insert the nest argument into the call argument list, which may 1348 // mean appending it. Likewise for attributes. 1349 1350 // Add any result attributes. 1351 if (Attributes Attr = Attrs.getRetAttributes()) 1352 NewAttrs.push_back(AttributeWithIndex::get(0, Attr)); 1353 1354 { 1355 unsigned Idx = 1; 1356 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 1357 do { 1358 if (Idx == NestIdx) { 1359 // Add the chain argument and attributes. 1360 Value *NestVal = Tramp->getArgOperand(2); 1361 if (NestVal->getType() != NestTy) 1362 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest"); 1363 NewArgs.push_back(NestVal); 1364 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr)); 1365 } 1366 1367 if (I == E) 1368 break; 1369 1370 // Add the original argument and attributes. 1371 NewArgs.push_back(*I); 1372 if (Attributes Attr = Attrs.getParamAttributes(Idx)) 1373 NewAttrs.push_back 1374 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr)); 1375 1376 ++Idx, ++I; 1377 } while (1); 1378 } 1379 1380 // Add any function attributes. 1381 if (Attributes Attr = Attrs.getFnAttributes()) 1382 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr)); 1383 1384 // The trampoline may have been bitcast to a bogus type (FTy). 1385 // Handle this by synthesizing a new function type, equal to FTy 1386 // with the chain parameter inserted. 1387 1388 std::vector<Type*> NewTypes; 1389 NewTypes.reserve(FTy->getNumParams()+1); 1390 1391 // Insert the chain's type into the list of parameter types, which may 1392 // mean appending it. 1393 { 1394 unsigned Idx = 1; 1395 FunctionType::param_iterator I = FTy->param_begin(), 1396 E = FTy->param_end(); 1397 1398 do { 1399 if (Idx == NestIdx) 1400 // Add the chain's type. 1401 NewTypes.push_back(NestTy); 1402 1403 if (I == E) 1404 break; 1405 1406 // Add the original type. 1407 NewTypes.push_back(*I); 1408 1409 ++Idx, ++I; 1410 } while (1); 1411 } 1412 1413 // Replace the trampoline call with a direct call. Let the generic 1414 // code sort out any function type mismatches. 1415 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes, 1416 FTy->isVarArg()); 1417 Constant *NewCallee = 1418 NestF->getType() == PointerType::getUnqual(NewFTy) ? 1419 NestF : ConstantExpr::getBitCast(NestF, 1420 PointerType::getUnqual(NewFTy)); 1421 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(), 1422 NewAttrs.end()); 1423 1424 Instruction *NewCaller; 1425 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1426 NewCaller = InvokeInst::Create(NewCallee, 1427 II->getNormalDest(), II->getUnwindDest(), 1428 NewArgs); 1429 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv()); 1430 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL); 1431 } else { 1432 NewCaller = CallInst::Create(NewCallee, NewArgs); 1433 if (cast<CallInst>(Caller)->isTailCall()) 1434 cast<CallInst>(NewCaller)->setTailCall(); 1435 cast<CallInst>(NewCaller)-> 1436 setCallingConv(cast<CallInst>(Caller)->getCallingConv()); 1437 cast<CallInst>(NewCaller)->setAttributes(NewPAL); 1438 } 1439 1440 return NewCaller; 1441 } 1442 } 1443 1444 // Replace the trampoline call with a direct call. Since there is no 'nest' 1445 // parameter, there is no need to adjust the argument list. Let the generic 1446 // code sort out any function type mismatches. 1447 Constant *NewCallee = 1448 NestF->getType() == PTy ? NestF : 1449 ConstantExpr::getBitCast(NestF, PTy); 1450 CS.setCalledFunction(NewCallee); 1451 return CS.getInstruction(); 1452} 1453