InstCombineCalls.cpp revision 12c807873a5988d900c8032e7d98f8f07da5c628
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/// computeAllocSize - compute the object size allocated by an allocation 169/// site. Returns 0 if the size is not constant (in SizeValue), 1 if the size 170/// is constant (in Size), and 2 if the size could not be determined within the 171/// given maximum Penalty that the computation would incurr at run-time. 172static int computeAllocSize(Value *Alloc, uint64_t &Size, Value* &SizeValue, 173 uint64_t Penalty, TargetData *TD, 174 InstCombiner::BuilderTy *Builder) { 175 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Alloc)) { 176 if (GV->hasDefinitiveInitializer()) { 177 Constant *C = GV->getInitializer(); 178 Size = TD->getTypeAllocSize(C->getType()); 179 return 1; 180 } 181 // Can't determine size of the GV. 182 return 2; 183 184 } else if (AllocaInst *AI = dyn_cast<AllocaInst>(Alloc)) { 185 if (!AI->getAllocatedType()->isSized()) 186 return 2; 187 188 Size = TD->getTypeAllocSize(AI->getAllocatedType()); 189 if (!AI->isArrayAllocation()) 190 return 1; // we are done 191 192 Value *ArraySize = AI->getArraySize(); 193 if (const ConstantInt *C = dyn_cast<ConstantInt>(ArraySize)) { 194 Size *= C->getZExtValue(); 195 return 1; 196 } 197 198 if (Penalty < 2) 199 return 2; 200 201 SizeValue = ConstantInt::get(ArraySize->getType(), Size); 202 SizeValue = Builder->CreateMul(SizeValue, ArraySize); 203 return 0; 204 205 } else if (CallInst *MI = extractMallocCall(Alloc)) { 206 SizeValue = MI->getArgOperand(0); 207 if (ConstantInt *CI = dyn_cast<ConstantInt>(SizeValue)) { 208 Size = CI->getZExtValue(); 209 return 1; 210 } 211 return Penalty >= 2 ? 0 : 2; 212 213 } else if (CallInst *MI = extractCallocCall(Alloc)) { 214 Value *Arg1 = MI->getArgOperand(0); 215 Value *Arg2 = MI->getArgOperand(1); 216 if (ConstantInt *CI1 = dyn_cast<ConstantInt>(Arg1)) { 217 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Arg2)) { 218 Size = (CI1->getValue() * CI2->getValue()).getZExtValue(); 219 return 1; 220 } 221 } 222 223 if (Penalty < 2) 224 return 2; 225 226 SizeValue = Builder->CreateMul(Arg1, Arg2); 227 return 0; 228 } 229 230 DEBUG(errs() << "computeAllocSize failed:\n"); 231 DEBUG(Alloc->dump()); 232 return 2; 233} 234 235/// visitCallInst - CallInst simplification. This mostly only handles folding 236/// of intrinsic instructions. For normal calls, it allows visitCallSite to do 237/// the heavy lifting. 238/// 239Instruction *InstCombiner::visitCallInst(CallInst &CI) { 240 if (isFreeCall(&CI)) 241 return visitFree(CI); 242 if (extractMallocCall(&CI) || extractCallocCall(&CI)) 243 return visitMalloc(CI); 244 245 // If the caller function is nounwind, mark the call as nounwind, even if the 246 // callee isn't. 247 if (CI.getParent()->getParent()->doesNotThrow() && 248 !CI.doesNotThrow()) { 249 CI.setDoesNotThrow(); 250 return &CI; 251 } 252 253 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI); 254 if (!II) return visitCallSite(&CI); 255 256 // Intrinsics cannot occur in an invoke, so handle them here instead of in 257 // visitCallSite. 258 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) { 259 bool Changed = false; 260 261 // memmove/cpy/set of zero bytes is a noop. 262 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) { 263 if (NumBytes->isNullValue()) 264 return EraseInstFromFunction(CI); 265 266 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes)) 267 if (CI->getZExtValue() == 1) { 268 // Replace the instruction with just byte operations. We would 269 // transform other cases to loads/stores, but we don't know if 270 // alignment is sufficient. 271 } 272 } 273 274 // No other transformations apply to volatile transfers. 275 if (MI->isVolatile()) 276 return 0; 277 278 // If we have a memmove and the source operation is a constant global, 279 // then the source and dest pointers can't alias, so we can change this 280 // into a call to memcpy. 281 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) { 282 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource())) 283 if (GVSrc->isConstant()) { 284 Module *M = CI.getParent()->getParent()->getParent(); 285 Intrinsic::ID MemCpyID = Intrinsic::memcpy; 286 Type *Tys[3] = { CI.getArgOperand(0)->getType(), 287 CI.getArgOperand(1)->getType(), 288 CI.getArgOperand(2)->getType() }; 289 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys)); 290 Changed = true; 291 } 292 } 293 294 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { 295 // memmove(x,x,size) -> noop. 296 if (MTI->getSource() == MTI->getDest()) 297 return EraseInstFromFunction(CI); 298 } 299 300 // If we can determine a pointer alignment that is bigger than currently 301 // set, update the alignment. 302 if (isa<MemTransferInst>(MI)) { 303 if (Instruction *I = SimplifyMemTransfer(MI)) 304 return I; 305 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) { 306 if (Instruction *I = SimplifyMemSet(MSI)) 307 return I; 308 } 309 310 if (Changed) return II; 311 } 312 313 switch (II->getIntrinsicID()) { 314 default: break; 315 case Intrinsic::objectsize: { 316 // We need target data for just about everything so depend on it. 317 if (!TD) return 0; 318 319 Type *ReturnTy = CI.getType(); 320 uint64_t Penalty = cast<ConstantInt>(II->getArgOperand(2))->getZExtValue(); 321 322 // Get to the real allocated thing and offset as fast as possible. 323 Value *Op1 = II->getArgOperand(0)->stripPointerCasts(); 324 GEPOperator *GEP; 325 326 if ((GEP = dyn_cast<GEPOperator>(Op1))) { 327 // check if we will be able to get the offset 328 if (!GEP->hasAllConstantIndices() && Penalty < 2) 329 return 0; 330 Op1 = GEP->getPointerOperand()->stripPointerCasts(); 331 } 332 333 uint64_t Size; 334 Value *SizeValue; 335 int ConstAlloc = computeAllocSize(Op1, Size, SizeValue, Penalty, TD, 336 Builder); 337 338 // Do not return "I don't know" here. Later optimization passes could 339 // make it possible to evaluate objectsize to a constant. 340 if (ConstAlloc == 2) 341 return 0; 342 343 uint64_t Offset = 0; 344 Value *OffsetValue = 0; 345 346 if (GEP) { 347 if (GEP->hasAllConstantIndices()) { 348 SmallVector<Value*, 8> Ops(GEP->idx_begin(), GEP->idx_end()); 349 assert(GEP->getPointerOperandType()->isPointerTy()); 350 Offset = TD->getIndexedOffset(GEP->getPointerOperandType(), Ops); 351 } else 352 OffsetValue = EmitGEPOffset(GEP, true /*NoNUW*/); 353 } 354 355 if (!OffsetValue && ConstAlloc) { 356 if (Size < Offset) { 357 // Out of bounds 358 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, 0)); 359 } 360 return ReplaceInstUsesWith(CI, ConstantInt::get(ReturnTy, Size-Offset)); 361 } 362 363 if (!OffsetValue) 364 OffsetValue = ConstantInt::get(ReturnTy, Offset); 365 if (ConstAlloc) 366 SizeValue = ConstantInt::get(ReturnTy, Size); 367 368 Value *Val = Builder->CreateSub(SizeValue, OffsetValue); 369 // return 0 if there's an overflow 370 Value *Cmp = Builder->CreateICmpULT(SizeValue, OffsetValue); 371 Val = Builder->CreateSelect(Cmp, ConstantInt::get(ReturnTy, 0), Val); 372 return ReplaceInstUsesWith(CI, Val); 373 } 374 case Intrinsic::bswap: 375 // bswap(bswap(x)) -> x 376 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) 377 if (Operand->getIntrinsicID() == Intrinsic::bswap) 378 return ReplaceInstUsesWith(CI, Operand->getArgOperand(0)); 379 380 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c)) 381 if (TruncInst *TI = dyn_cast<TruncInst>(II->getArgOperand(0))) { 382 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(TI->getOperand(0))) 383 if (Operand->getIntrinsicID() == Intrinsic::bswap) { 384 unsigned C = Operand->getType()->getPrimitiveSizeInBits() - 385 TI->getType()->getPrimitiveSizeInBits(); 386 Value *CV = ConstantInt::get(Operand->getType(), C); 387 Value *V = Builder->CreateLShr(Operand->getArgOperand(0), CV); 388 return new TruncInst(V, TI->getType()); 389 } 390 } 391 392 break; 393 case Intrinsic::powi: 394 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 395 // powi(x, 0) -> 1.0 396 if (Power->isZero()) 397 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0)); 398 // powi(x, 1) -> x 399 if (Power->isOne()) 400 return ReplaceInstUsesWith(CI, II->getArgOperand(0)); 401 // powi(x, -1) -> 1/x 402 if (Power->isAllOnesValue()) 403 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0), 404 II->getArgOperand(0)); 405 } 406 break; 407 case Intrinsic::cttz: { 408 // If all bits below the first known one are known zero, 409 // this value is constant. 410 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType()); 411 // FIXME: Try to simplify vectors of integers. 412 if (!IT) break; 413 uint32_t BitWidth = IT->getBitWidth(); 414 APInt KnownZero(BitWidth, 0); 415 APInt KnownOne(BitWidth, 0); 416 ComputeMaskedBits(II->getArgOperand(0), KnownZero, KnownOne); 417 unsigned TrailingZeros = KnownOne.countTrailingZeros(); 418 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros)); 419 if ((Mask & KnownZero) == Mask) 420 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 421 APInt(BitWidth, TrailingZeros))); 422 423 } 424 break; 425 case Intrinsic::ctlz: { 426 // If all bits above the first known one are known zero, 427 // this value is constant. 428 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType()); 429 // FIXME: Try to simplify vectors of integers. 430 if (!IT) break; 431 uint32_t BitWidth = IT->getBitWidth(); 432 APInt KnownZero(BitWidth, 0); 433 APInt KnownOne(BitWidth, 0); 434 ComputeMaskedBits(II->getArgOperand(0), KnownZero, KnownOne); 435 unsigned LeadingZeros = KnownOne.countLeadingZeros(); 436 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros)); 437 if ((Mask & KnownZero) == Mask) 438 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 439 APInt(BitWidth, LeadingZeros))); 440 441 } 442 break; 443 case Intrinsic::uadd_with_overflow: { 444 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 445 IntegerType *IT = cast<IntegerType>(II->getArgOperand(0)->getType()); 446 uint32_t BitWidth = IT->getBitWidth(); 447 APInt LHSKnownZero(BitWidth, 0); 448 APInt LHSKnownOne(BitWidth, 0); 449 ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne); 450 bool LHSKnownNegative = LHSKnownOne[BitWidth - 1]; 451 bool LHSKnownPositive = LHSKnownZero[BitWidth - 1]; 452 453 if (LHSKnownNegative || LHSKnownPositive) { 454 APInt RHSKnownZero(BitWidth, 0); 455 APInt RHSKnownOne(BitWidth, 0); 456 ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne); 457 bool RHSKnownNegative = RHSKnownOne[BitWidth - 1]; 458 bool RHSKnownPositive = RHSKnownZero[BitWidth - 1]; 459 if (LHSKnownNegative && RHSKnownNegative) { 460 // The sign bit is set in both cases: this MUST overflow. 461 // Create a simple add instruction, and insert it into the struct. 462 Value *Add = Builder->CreateAdd(LHS, RHS); 463 Add->takeName(&CI); 464 Constant *V[] = { 465 UndefValue::get(LHS->getType()), 466 ConstantInt::getTrue(II->getContext()) 467 }; 468 StructType *ST = cast<StructType>(II->getType()); 469 Constant *Struct = ConstantStruct::get(ST, V); 470 return InsertValueInst::Create(Struct, Add, 0); 471 } 472 473 if (LHSKnownPositive && RHSKnownPositive) { 474 // The sign bit is clear in both cases: this CANNOT overflow. 475 // Create a simple add instruction, and insert it into the struct. 476 Value *Add = Builder->CreateNUWAdd(LHS, RHS); 477 Add->takeName(&CI); 478 Constant *V[] = { 479 UndefValue::get(LHS->getType()), 480 ConstantInt::getFalse(II->getContext()) 481 }; 482 StructType *ST = cast<StructType>(II->getType()); 483 Constant *Struct = ConstantStruct::get(ST, V); 484 return InsertValueInst::Create(Struct, Add, 0); 485 } 486 } 487 } 488 // FALL THROUGH uadd into sadd 489 case Intrinsic::sadd_with_overflow: 490 // Canonicalize constants into the RHS. 491 if (isa<Constant>(II->getArgOperand(0)) && 492 !isa<Constant>(II->getArgOperand(1))) { 493 Value *LHS = II->getArgOperand(0); 494 II->setArgOperand(0, II->getArgOperand(1)); 495 II->setArgOperand(1, LHS); 496 return II; 497 } 498 499 // X + undef -> undef 500 if (isa<UndefValue>(II->getArgOperand(1))) 501 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 502 503 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 504 // X + 0 -> {X, false} 505 if (RHS->isZero()) { 506 Constant *V[] = { 507 UndefValue::get(II->getArgOperand(0)->getType()), 508 ConstantInt::getFalse(II->getContext()) 509 }; 510 Constant *Struct = 511 ConstantStruct::get(cast<StructType>(II->getType()), V); 512 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 513 } 514 } 515 break; 516 case Intrinsic::usub_with_overflow: 517 case Intrinsic::ssub_with_overflow: 518 // undef - X -> undef 519 // X - undef -> undef 520 if (isa<UndefValue>(II->getArgOperand(0)) || 521 isa<UndefValue>(II->getArgOperand(1))) 522 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 523 524 if (ConstantInt *RHS = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 525 // X - 0 -> {X, false} 526 if (RHS->isZero()) { 527 Constant *V[] = { 528 UndefValue::get(II->getArgOperand(0)->getType()), 529 ConstantInt::getFalse(II->getContext()) 530 }; 531 Constant *Struct = 532 ConstantStruct::get(cast<StructType>(II->getType()), V); 533 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 534 } 535 } 536 break; 537 case Intrinsic::umul_with_overflow: { 538 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 539 unsigned BitWidth = cast<IntegerType>(LHS->getType())->getBitWidth(); 540 541 APInt LHSKnownZero(BitWidth, 0); 542 APInt LHSKnownOne(BitWidth, 0); 543 ComputeMaskedBits(LHS, LHSKnownZero, LHSKnownOne); 544 APInt RHSKnownZero(BitWidth, 0); 545 APInt RHSKnownOne(BitWidth, 0); 546 ComputeMaskedBits(RHS, RHSKnownZero, RHSKnownOne); 547 548 // Get the largest possible values for each operand. 549 APInt LHSMax = ~LHSKnownZero; 550 APInt RHSMax = ~RHSKnownZero; 551 552 // If multiplying the maximum values does not overflow then we can turn 553 // this into a plain NUW mul. 554 bool Overflow; 555 LHSMax.umul_ov(RHSMax, Overflow); 556 if (!Overflow) { 557 Value *Mul = Builder->CreateNUWMul(LHS, RHS, "umul_with_overflow"); 558 Constant *V[] = { 559 UndefValue::get(LHS->getType()), 560 Builder->getFalse() 561 }; 562 Constant *Struct = ConstantStruct::get(cast<StructType>(II->getType()),V); 563 return InsertValueInst::Create(Struct, Mul, 0); 564 } 565 } // FALL THROUGH 566 case Intrinsic::smul_with_overflow: 567 // Canonicalize constants into the RHS. 568 if (isa<Constant>(II->getArgOperand(0)) && 569 !isa<Constant>(II->getArgOperand(1))) { 570 Value *LHS = II->getArgOperand(0); 571 II->setArgOperand(0, II->getArgOperand(1)); 572 II->setArgOperand(1, LHS); 573 return II; 574 } 575 576 // X * undef -> undef 577 if (isa<UndefValue>(II->getArgOperand(1))) 578 return ReplaceInstUsesWith(CI, UndefValue::get(II->getType())); 579 580 if (ConstantInt *RHSI = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 581 // X*0 -> {0, false} 582 if (RHSI->isZero()) 583 return ReplaceInstUsesWith(CI, Constant::getNullValue(II->getType())); 584 585 // X * 1 -> {X, false} 586 if (RHSI->equalsInt(1)) { 587 Constant *V[] = { 588 UndefValue::get(II->getArgOperand(0)->getType()), 589 ConstantInt::getFalse(II->getContext()) 590 }; 591 Constant *Struct = 592 ConstantStruct::get(cast<StructType>(II->getType()), V); 593 return InsertValueInst::Create(Struct, II->getArgOperand(0), 0); 594 } 595 } 596 break; 597 case Intrinsic::ppc_altivec_lvx: 598 case Intrinsic::ppc_altivec_lvxl: 599 // Turn PPC lvx -> load if the pointer is known aligned. 600 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 601 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), 602 PointerType::getUnqual(II->getType())); 603 return new LoadInst(Ptr); 604 } 605 break; 606 case Intrinsic::ppc_altivec_stvx: 607 case Intrinsic::ppc_altivec_stvxl: 608 // Turn stvx -> store if the pointer is known aligned. 609 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, TD) >= 16) { 610 Type *OpPtrTy = 611 PointerType::getUnqual(II->getArgOperand(0)->getType()); 612 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); 613 return new StoreInst(II->getArgOperand(0), Ptr); 614 } 615 break; 616 case Intrinsic::x86_sse_storeu_ps: 617 case Intrinsic::x86_sse2_storeu_pd: 618 case Intrinsic::x86_sse2_storeu_dq: 619 // Turn X86 storeu -> store if the pointer is known aligned. 620 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, TD) >= 16) { 621 Type *OpPtrTy = 622 PointerType::getUnqual(II->getArgOperand(1)->getType()); 623 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy); 624 return new StoreInst(II->getArgOperand(1), Ptr); 625 } 626 break; 627 628 case Intrinsic::x86_sse_cvtss2si: 629 case Intrinsic::x86_sse_cvtss2si64: 630 case Intrinsic::x86_sse_cvttss2si: 631 case Intrinsic::x86_sse_cvttss2si64: 632 case Intrinsic::x86_sse2_cvtsd2si: 633 case Intrinsic::x86_sse2_cvtsd2si64: 634 case Intrinsic::x86_sse2_cvttsd2si: 635 case Intrinsic::x86_sse2_cvttsd2si64: { 636 // These intrinsics only demand the 0th element of their input vectors. If 637 // we can simplify the input based on that, do so now. 638 unsigned VWidth = 639 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 640 APInt DemandedElts(VWidth, 1); 641 APInt UndefElts(VWidth, 0); 642 if (Value *V = SimplifyDemandedVectorElts(II->getArgOperand(0), 643 DemandedElts, UndefElts)) { 644 II->setArgOperand(0, V); 645 return II; 646 } 647 break; 648 } 649 650 651 case Intrinsic::x86_sse41_pmovsxbw: 652 case Intrinsic::x86_sse41_pmovsxwd: 653 case Intrinsic::x86_sse41_pmovsxdq: 654 case Intrinsic::x86_sse41_pmovzxbw: 655 case Intrinsic::x86_sse41_pmovzxwd: 656 case Intrinsic::x86_sse41_pmovzxdq: { 657 // pmov{s|z}x ignores the upper half of their input vectors. 658 unsigned VWidth = 659 cast<VectorType>(II->getArgOperand(0)->getType())->getNumElements(); 660 unsigned LowHalfElts = VWidth / 2; 661 APInt InputDemandedElts(APInt::getBitsSet(VWidth, 0, LowHalfElts)); 662 APInt UndefElts(VWidth, 0); 663 if (Value *TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), 664 InputDemandedElts, 665 UndefElts)) { 666 II->setArgOperand(0, TmpV); 667 return II; 668 } 669 break; 670 } 671 672 case Intrinsic::ppc_altivec_vperm: 673 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant. 674 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) { 675 assert(Mask->getType()->getVectorNumElements() == 16 && 676 "Bad type for intrinsic!"); 677 678 // Check that all of the elements are integer constants or undefs. 679 bool AllEltsOk = true; 680 for (unsigned i = 0; i != 16; ++i) { 681 Constant *Elt = Mask->getAggregateElement(i); 682 if (Elt == 0 || 683 !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) { 684 AllEltsOk = false; 685 break; 686 } 687 } 688 689 if (AllEltsOk) { 690 // Cast the input vectors to byte vectors. 691 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0), 692 Mask->getType()); 693 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1), 694 Mask->getType()); 695 Value *Result = UndefValue::get(Op0->getType()); 696 697 // Only extract each element once. 698 Value *ExtractedElts[32]; 699 memset(ExtractedElts, 0, sizeof(ExtractedElts)); 700 701 for (unsigned i = 0; i != 16; ++i) { 702 if (isa<UndefValue>(Mask->getAggregateElement(i))) 703 continue; 704 unsigned Idx = 705 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue(); 706 Idx &= 31; // Match the hardware behavior. 707 708 if (ExtractedElts[Idx] == 0) { 709 ExtractedElts[Idx] = 710 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1, 711 Builder->getInt32(Idx&15)); 712 } 713 714 // Insert this value into the result vector. 715 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx], 716 Builder->getInt32(i)); 717 } 718 return CastInst::Create(Instruction::BitCast, Result, CI.getType()); 719 } 720 } 721 break; 722 723 case Intrinsic::arm_neon_vld1: 724 case Intrinsic::arm_neon_vld2: 725 case Intrinsic::arm_neon_vld3: 726 case Intrinsic::arm_neon_vld4: 727 case Intrinsic::arm_neon_vld2lane: 728 case Intrinsic::arm_neon_vld3lane: 729 case Intrinsic::arm_neon_vld4lane: 730 case Intrinsic::arm_neon_vst1: 731 case Intrinsic::arm_neon_vst2: 732 case Intrinsic::arm_neon_vst3: 733 case Intrinsic::arm_neon_vst4: 734 case Intrinsic::arm_neon_vst2lane: 735 case Intrinsic::arm_neon_vst3lane: 736 case Intrinsic::arm_neon_vst4lane: { 737 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), TD); 738 unsigned AlignArg = II->getNumArgOperands() - 1; 739 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg)); 740 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) { 741 II->setArgOperand(AlignArg, 742 ConstantInt::get(Type::getInt32Ty(II->getContext()), 743 MemAlign, false)); 744 return II; 745 } 746 break; 747 } 748 749 case Intrinsic::arm_neon_vmulls: 750 case Intrinsic::arm_neon_vmullu: { 751 Value *Arg0 = II->getArgOperand(0); 752 Value *Arg1 = II->getArgOperand(1); 753 754 // Handle mul by zero first: 755 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) { 756 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType())); 757 } 758 759 // Check for constant LHS & RHS - in this case we just simplify. 760 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu); 761 VectorType *NewVT = cast<VectorType>(II->getType()); 762 unsigned NewWidth = NewVT->getElementType()->getIntegerBitWidth(); 763 if (ConstantDataVector *CV0 = dyn_cast<ConstantDataVector>(Arg0)) { 764 if (ConstantDataVector *CV1 = dyn_cast<ConstantDataVector>(Arg1)) { 765 VectorType* VT = cast<VectorType>(CV0->getType()); 766 SmallVector<Constant*, 4> NewElems; 767 for (unsigned i = 0; i < VT->getNumElements(); ++i) { 768 APInt CV0E = 769 (cast<ConstantInt>(CV0->getAggregateElement(i)))->getValue(); 770 CV0E = Zext ? CV0E.zext(NewWidth) : CV0E.sext(NewWidth); 771 APInt CV1E = 772 (cast<ConstantInt>(CV1->getAggregateElement(i)))->getValue(); 773 CV1E = Zext ? CV1E.zext(NewWidth) : CV1E.sext(NewWidth); 774 NewElems.push_back( 775 ConstantInt::get(NewVT->getElementType(), CV0E * CV1E)); 776 } 777 return ReplaceInstUsesWith(CI, ConstantVector::get(NewElems)); 778 } 779 780 // Couldn't simplify - cannonicalize constant to the RHS. 781 std::swap(Arg0, Arg1); 782 } 783 784 // Handle mul by one: 785 if (ConstantDataVector *CV1 = dyn_cast<ConstantDataVector>(Arg1)) { 786 if (ConstantInt *Splat = 787 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue())) { 788 if (Splat->isOne()) { 789 if (Zext) 790 return CastInst::CreateZExtOrBitCast(Arg0, II->getType()); 791 // else 792 return CastInst::CreateSExtOrBitCast(Arg0, II->getType()); 793 } 794 } 795 } 796 797 break; 798 } 799 800 case Intrinsic::stackrestore: { 801 // If the save is right next to the restore, remove the restore. This can 802 // happen when variable allocas are DCE'd. 803 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) { 804 if (SS->getIntrinsicID() == Intrinsic::stacksave) { 805 BasicBlock::iterator BI = SS; 806 if (&*++BI == II) 807 return EraseInstFromFunction(CI); 808 } 809 } 810 811 // Scan down this block to see if there is another stack restore in the 812 // same block without an intervening call/alloca. 813 BasicBlock::iterator BI = II; 814 TerminatorInst *TI = II->getParent()->getTerminator(); 815 bool CannotRemove = false; 816 for (++BI; &*BI != TI; ++BI) { 817 if (isa<AllocaInst>(BI) || isMalloc(BI)) { 818 CannotRemove = true; 819 break; 820 } 821 if (CallInst *BCI = dyn_cast<CallInst>(BI)) { 822 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) { 823 // If there is a stackrestore below this one, remove this one. 824 if (II->getIntrinsicID() == Intrinsic::stackrestore) 825 return EraseInstFromFunction(CI); 826 // Otherwise, ignore the intrinsic. 827 } else { 828 // If we found a non-intrinsic call, we can't remove the stack 829 // restore. 830 CannotRemove = true; 831 break; 832 } 833 } 834 } 835 836 // If the stack restore is in a return, resume, or unwind block and if there 837 // are no allocas or calls between the restore and the return, nuke the 838 // restore. 839 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI))) 840 return EraseInstFromFunction(CI); 841 break; 842 } 843 } 844 845 return visitCallSite(II); 846} 847 848// InvokeInst simplification 849// 850Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) { 851 return visitCallSite(&II); 852} 853 854/// isSafeToEliminateVarargsCast - If this cast does not affect the value 855/// passed through the varargs area, we can eliminate the use of the cast. 856static bool isSafeToEliminateVarargsCast(const CallSite CS, 857 const CastInst * const CI, 858 const TargetData * const TD, 859 const int ix) { 860 if (!CI->isLosslessCast()) 861 return false; 862 863 // The size of ByVal arguments is derived from the type, so we 864 // can't change to a type with a different size. If the size were 865 // passed explicitly we could avoid this check. 866 if (!CS.isByValArgument(ix)) 867 return true; 868 869 Type* SrcTy = 870 cast<PointerType>(CI->getOperand(0)->getType())->getElementType(); 871 Type* DstTy = cast<PointerType>(CI->getType())->getElementType(); 872 if (!SrcTy->isSized() || !DstTy->isSized()) 873 return false; 874 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy)) 875 return false; 876 return true; 877} 878 879namespace { 880class InstCombineFortifiedLibCalls : public SimplifyFortifiedLibCalls { 881 InstCombiner *IC; 882protected: 883 void replaceCall(Value *With) { 884 NewInstruction = IC->ReplaceInstUsesWith(*CI, With); 885 } 886 bool isFoldable(unsigned SizeCIOp, unsigned SizeArgOp, bool isString) const { 887 if (CI->getArgOperand(SizeCIOp) == CI->getArgOperand(SizeArgOp)) 888 return true; 889 if (ConstantInt *SizeCI = 890 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp))) { 891 if (SizeCI->isAllOnesValue()) 892 return true; 893 if (isString) { 894 uint64_t Len = GetStringLength(CI->getArgOperand(SizeArgOp)); 895 // If the length is 0 we don't know how long it is and so we can't 896 // remove the check. 897 if (Len == 0) return false; 898 return SizeCI->getZExtValue() >= Len; 899 } 900 if (ConstantInt *Arg = dyn_cast<ConstantInt>( 901 CI->getArgOperand(SizeArgOp))) 902 return SizeCI->getZExtValue() >= Arg->getZExtValue(); 903 } 904 return false; 905 } 906public: 907 InstCombineFortifiedLibCalls(InstCombiner *IC) : IC(IC), NewInstruction(0) { } 908 Instruction *NewInstruction; 909}; 910} // end anonymous namespace 911 912// Try to fold some different type of calls here. 913// Currently we're only working with the checking functions, memcpy_chk, 914// mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk, 915// strcat_chk and strncat_chk. 916Instruction *InstCombiner::tryOptimizeCall(CallInst *CI, const TargetData *TD) { 917 if (CI->getCalledFunction() == 0) return 0; 918 919 InstCombineFortifiedLibCalls Simplifier(this); 920 Simplifier.fold(CI, TD); 921 return Simplifier.NewInstruction; 922} 923 924static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) { 925 // Strip off at most one level of pointer casts, looking for an alloca. This 926 // is good enough in practice and simpler than handling any number of casts. 927 Value *Underlying = TrampMem->stripPointerCasts(); 928 if (Underlying != TrampMem && 929 (!Underlying->hasOneUse() || *Underlying->use_begin() != TrampMem)) 930 return 0; 931 if (!isa<AllocaInst>(Underlying)) 932 return 0; 933 934 IntrinsicInst *InitTrampoline = 0; 935 for (Value::use_iterator I = TrampMem->use_begin(), E = TrampMem->use_end(); 936 I != E; I++) { 937 IntrinsicInst *II = dyn_cast<IntrinsicInst>(*I); 938 if (!II) 939 return 0; 940 if (II->getIntrinsicID() == Intrinsic::init_trampoline) { 941 if (InitTrampoline) 942 // More than one init_trampoline writes to this value. Give up. 943 return 0; 944 InitTrampoline = II; 945 continue; 946 } 947 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline) 948 // Allow any number of calls to adjust.trampoline. 949 continue; 950 return 0; 951 } 952 953 // No call to init.trampoline found. 954 if (!InitTrampoline) 955 return 0; 956 957 // Check that the alloca is being used in the expected way. 958 if (InitTrampoline->getOperand(0) != TrampMem) 959 return 0; 960 961 return InitTrampoline; 962} 963 964static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp, 965 Value *TrampMem) { 966 // Visit all the previous instructions in the basic block, and try to find a 967 // init.trampoline which has a direct path to the adjust.trampoline. 968 for (BasicBlock::iterator I = AdjustTramp, 969 E = AdjustTramp->getParent()->begin(); I != E; ) { 970 Instruction *Inst = --I; 971 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) 972 if (II->getIntrinsicID() == Intrinsic::init_trampoline && 973 II->getOperand(0) == TrampMem) 974 return II; 975 if (Inst->mayWriteToMemory()) 976 return 0; 977 } 978 return 0; 979} 980 981// Given a call to llvm.adjust.trampoline, find and return the corresponding 982// call to llvm.init.trampoline if the call to the trampoline can be optimized 983// to a direct call to a function. Otherwise return NULL. 984// 985static IntrinsicInst *FindInitTrampoline(Value *Callee) { 986 Callee = Callee->stripPointerCasts(); 987 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee); 988 if (!AdjustTramp || 989 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline) 990 return 0; 991 992 Value *TrampMem = AdjustTramp->getOperand(0); 993 994 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem)) 995 return IT; 996 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem)) 997 return IT; 998 return 0; 999} 1000 1001// visitCallSite - Improvements for call and invoke instructions. 1002// 1003Instruction *InstCombiner::visitCallSite(CallSite CS) { 1004 bool Changed = false; 1005 1006 // If the callee is a pointer to a function, attempt to move any casts to the 1007 // arguments of the call/invoke. 1008 Value *Callee = CS.getCalledValue(); 1009 if (!isa<Function>(Callee) && transformConstExprCastCall(CS)) 1010 return 0; 1011 1012 if (Function *CalleeF = dyn_cast<Function>(Callee)) 1013 // If the call and callee calling conventions don't match, this call must 1014 // be unreachable, as the call is undefined. 1015 if (CalleeF->getCallingConv() != CS.getCallingConv() && 1016 // Only do this for calls to a function with a body. A prototype may 1017 // not actually end up matching the implementation's calling conv for a 1018 // variety of reasons (e.g. it may be written in assembly). 1019 !CalleeF->isDeclaration()) { 1020 Instruction *OldCall = CS.getInstruction(); 1021 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 1022 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 1023 OldCall); 1024 // If OldCall dues not return void then replaceAllUsesWith undef. 1025 // This allows ValueHandlers and custom metadata to adjust itself. 1026 if (!OldCall->getType()->isVoidTy()) 1027 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType())); 1028 if (isa<CallInst>(OldCall)) 1029 return EraseInstFromFunction(*OldCall); 1030 1031 // We cannot remove an invoke, because it would change the CFG, just 1032 // change the callee to a null pointer. 1033 cast<InvokeInst>(OldCall)->setCalledFunction( 1034 Constant::getNullValue(CalleeF->getType())); 1035 return 0; 1036 } 1037 1038 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 1039 // This instruction is not reachable, just remove it. We insert a store to 1040 // undef so that we know that this code is not reachable, despite the fact 1041 // that we can't modify the CFG here. 1042 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 1043 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 1044 CS.getInstruction()); 1045 1046 // If CS does not return void then replaceAllUsesWith undef. 1047 // This allows ValueHandlers and custom metadata to adjust itself. 1048 if (!CS.getInstruction()->getType()->isVoidTy()) 1049 ReplaceInstUsesWith(*CS.getInstruction(), 1050 UndefValue::get(CS.getInstruction()->getType())); 1051 1052 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) { 1053 // Don't break the CFG, insert a dummy cond branch. 1054 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(), 1055 ConstantInt::getTrue(Callee->getContext()), II); 1056 } 1057 return EraseInstFromFunction(*CS.getInstruction()); 1058 } 1059 1060 if (IntrinsicInst *II = FindInitTrampoline(Callee)) 1061 return transformCallThroughTrampoline(CS, II); 1062 1063 PointerType *PTy = cast<PointerType>(Callee->getType()); 1064 FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 1065 if (FTy->isVarArg()) { 1066 int ix = FTy->getNumParams(); 1067 // See if we can optimize any arguments passed through the varargs area of 1068 // the call. 1069 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(), 1070 E = CS.arg_end(); I != E; ++I, ++ix) { 1071 CastInst *CI = dyn_cast<CastInst>(*I); 1072 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) { 1073 *I = CI->getOperand(0); 1074 Changed = true; 1075 } 1076 } 1077 } 1078 1079 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) { 1080 // Inline asm calls cannot throw - mark them 'nounwind'. 1081 CS.setDoesNotThrow(); 1082 Changed = true; 1083 } 1084 1085 // Try to optimize the call if possible, we require TargetData for most of 1086 // this. None of these calls are seen as possibly dead so go ahead and 1087 // delete the instruction now. 1088 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) { 1089 Instruction *I = tryOptimizeCall(CI, TD); 1090 // If we changed something return the result, etc. Otherwise let 1091 // the fallthrough check. 1092 if (I) return EraseInstFromFunction(*I); 1093 } 1094 1095 return Changed ? CS.getInstruction() : 0; 1096} 1097 1098// transformConstExprCastCall - If the callee is a constexpr cast of a function, 1099// attempt to move the cast to the arguments of the call/invoke. 1100// 1101bool InstCombiner::transformConstExprCastCall(CallSite CS) { 1102 Function *Callee = 1103 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts()); 1104 if (Callee == 0) 1105 return false; 1106 Instruction *Caller = CS.getInstruction(); 1107 const AttrListPtr &CallerPAL = CS.getAttributes(); 1108 1109 // Okay, this is a cast from a function to a different type. Unless doing so 1110 // would cause a type conversion of one of our arguments, change this call to 1111 // be a direct call with arguments casted to the appropriate types. 1112 // 1113 FunctionType *FT = Callee->getFunctionType(); 1114 Type *OldRetTy = Caller->getType(); 1115 Type *NewRetTy = FT->getReturnType(); 1116 1117 if (NewRetTy->isStructTy()) 1118 return false; // TODO: Handle multiple return values. 1119 1120 // Check to see if we are changing the return type... 1121 if (OldRetTy != NewRetTy) { 1122 if (Callee->isDeclaration() && 1123 // Conversion is ok if changing from one pointer type to another or from 1124 // a pointer to an integer of the same size. 1125 !((OldRetTy->isPointerTy() || !TD || 1126 OldRetTy == TD->getIntPtrType(Caller->getContext())) && 1127 (NewRetTy->isPointerTy() || !TD || 1128 NewRetTy == TD->getIntPtrType(Caller->getContext())))) 1129 return false; // Cannot transform this return value. 1130 1131 if (!Caller->use_empty() && 1132 // void -> non-void is handled specially 1133 !NewRetTy->isVoidTy() && !CastInst::isCastable(NewRetTy, OldRetTy)) 1134 return false; // Cannot transform this return value. 1135 1136 if (!CallerPAL.isEmpty() && !Caller->use_empty()) { 1137 Attributes RAttrs = CallerPAL.getRetAttributes(); 1138 if (RAttrs & Attribute::typeIncompatible(NewRetTy)) 1139 return false; // Attribute not compatible with transformed value. 1140 } 1141 1142 // If the callsite is an invoke instruction, and the return value is used by 1143 // a PHI node in a successor, we cannot change the return type of the call 1144 // because there is no place to put the cast instruction (without breaking 1145 // the critical edge). Bail out in this case. 1146 if (!Caller->use_empty()) 1147 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) 1148 for (Value::use_iterator UI = II->use_begin(), E = II->use_end(); 1149 UI != E; ++UI) 1150 if (PHINode *PN = dyn_cast<PHINode>(*UI)) 1151 if (PN->getParent() == II->getNormalDest() || 1152 PN->getParent() == II->getUnwindDest()) 1153 return false; 1154 } 1155 1156 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin()); 1157 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs); 1158 1159 CallSite::arg_iterator AI = CS.arg_begin(); 1160 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) { 1161 Type *ParamTy = FT->getParamType(i); 1162 Type *ActTy = (*AI)->getType(); 1163 1164 if (!CastInst::isCastable(ActTy, ParamTy)) 1165 return false; // Cannot transform this parameter value. 1166 1167 Attributes Attrs = CallerPAL.getParamAttributes(i + 1); 1168 if (Attrs & Attribute::typeIncompatible(ParamTy)) 1169 return false; // Attribute not compatible with transformed value. 1170 1171 // If the parameter is passed as a byval argument, then we have to have a 1172 // sized type and the sized type has to have the same size as the old type. 1173 if (ParamTy != ActTy && (Attrs & Attribute::ByVal)) { 1174 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy); 1175 if (ParamPTy == 0 || !ParamPTy->getElementType()->isSized() || TD == 0) 1176 return false; 1177 1178 Type *CurElTy = cast<PointerType>(ActTy)->getElementType(); 1179 if (TD->getTypeAllocSize(CurElTy) != 1180 TD->getTypeAllocSize(ParamPTy->getElementType())) 1181 return false; 1182 } 1183 1184 // Converting from one pointer type to another or between a pointer and an 1185 // integer of the same size is safe even if we do not have a body. 1186 bool isConvertible = ActTy == ParamTy || 1187 (TD && ((ParamTy->isPointerTy() || 1188 ParamTy == TD->getIntPtrType(Caller->getContext())) && 1189 (ActTy->isPointerTy() || 1190 ActTy == TD->getIntPtrType(Caller->getContext())))); 1191 if (Callee->isDeclaration() && !isConvertible) return false; 1192 } 1193 1194 if (Callee->isDeclaration()) { 1195 // Do not delete arguments unless we have a function body. 1196 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg()) 1197 return false; 1198 1199 // If the callee is just a declaration, don't change the varargsness of the 1200 // call. We don't want to introduce a varargs call where one doesn't 1201 // already exist. 1202 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType()); 1203 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg()) 1204 return false; 1205 1206 // If both the callee and the cast type are varargs, we still have to make 1207 // sure the number of fixed parameters are the same or we have the same 1208 // ABI issues as if we introduce a varargs call. 1209 if (FT->isVarArg() && 1210 cast<FunctionType>(APTy->getElementType())->isVarArg() && 1211 FT->getNumParams() != 1212 cast<FunctionType>(APTy->getElementType())->getNumParams()) 1213 return false; 1214 } 1215 1216 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && 1217 !CallerPAL.isEmpty()) 1218 // In this case we have more arguments than the new function type, but we 1219 // won't be dropping them. Check that these extra arguments have attributes 1220 // that are compatible with being a vararg call argument. 1221 for (unsigned i = CallerPAL.getNumSlots(); i; --i) { 1222 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams()) 1223 break; 1224 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs; 1225 if (PAttrs & Attribute::VarArgsIncompatible) 1226 return false; 1227 } 1228 1229 1230 // Okay, we decided that this is a safe thing to do: go ahead and start 1231 // inserting cast instructions as necessary. 1232 std::vector<Value*> Args; 1233 Args.reserve(NumActualArgs); 1234 SmallVector<AttributeWithIndex, 8> attrVec; 1235 attrVec.reserve(NumCommonArgs); 1236 1237 // Get any return attributes. 1238 Attributes RAttrs = CallerPAL.getRetAttributes(); 1239 1240 // If the return value is not being used, the type may not be compatible 1241 // with the existing attributes. Wipe out any problematic attributes. 1242 RAttrs &= ~Attribute::typeIncompatible(NewRetTy); 1243 1244 // Add the new return attributes. 1245 if (RAttrs) 1246 attrVec.push_back(AttributeWithIndex::get(0, RAttrs)); 1247 1248 AI = CS.arg_begin(); 1249 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) { 1250 Type *ParamTy = FT->getParamType(i); 1251 if ((*AI)->getType() == ParamTy) { 1252 Args.push_back(*AI); 1253 } else { 1254 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, 1255 false, ParamTy, false); 1256 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy)); 1257 } 1258 1259 // Add any parameter attributes. 1260 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1261 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1262 } 1263 1264 // If the function takes more arguments than the call was taking, add them 1265 // now. 1266 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) 1267 Args.push_back(Constant::getNullValue(FT->getParamType(i))); 1268 1269 // If we are removing arguments to the function, emit an obnoxious warning. 1270 if (FT->getNumParams() < NumActualArgs) { 1271 if (!FT->isVarArg()) { 1272 errs() << "WARNING: While resolving call to function '" 1273 << Callee->getName() << "' arguments were dropped!\n"; 1274 } else { 1275 // Add all of the arguments in their promoted form to the arg list. 1276 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) { 1277 Type *PTy = getPromotedType((*AI)->getType()); 1278 if (PTy != (*AI)->getType()) { 1279 // Must promote to pass through va_arg area! 1280 Instruction::CastOps opcode = 1281 CastInst::getCastOpcode(*AI, false, PTy, false); 1282 Args.push_back(Builder->CreateCast(opcode, *AI, PTy)); 1283 } else { 1284 Args.push_back(*AI); 1285 } 1286 1287 // Add any parameter attributes. 1288 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1)) 1289 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs)); 1290 } 1291 } 1292 } 1293 1294 if (Attributes FnAttrs = CallerPAL.getFnAttributes()) 1295 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs)); 1296 1297 if (NewRetTy->isVoidTy()) 1298 Caller->setName(""); // Void type should not have a name. 1299 1300 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(), 1301 attrVec.end()); 1302 1303 Instruction *NC; 1304 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1305 NC = Builder->CreateInvoke(Callee, II->getNormalDest(), 1306 II->getUnwindDest(), Args); 1307 NC->takeName(II); 1308 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv()); 1309 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL); 1310 } else { 1311 CallInst *CI = cast<CallInst>(Caller); 1312 NC = Builder->CreateCall(Callee, Args); 1313 NC->takeName(CI); 1314 if (CI->isTailCall()) 1315 cast<CallInst>(NC)->setTailCall(); 1316 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv()); 1317 cast<CallInst>(NC)->setAttributes(NewCallerPAL); 1318 } 1319 1320 // Insert a cast of the return type as necessary. 1321 Value *NV = NC; 1322 if (OldRetTy != NV->getType() && !Caller->use_empty()) { 1323 if (!NV->getType()->isVoidTy()) { 1324 Instruction::CastOps opcode = 1325 CastInst::getCastOpcode(NC, false, OldRetTy, false); 1326 NV = NC = CastInst::Create(opcode, NC, OldRetTy); 1327 NC->setDebugLoc(Caller->getDebugLoc()); 1328 1329 // If this is an invoke instruction, we should insert it after the first 1330 // non-phi, instruction in the normal successor block. 1331 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1332 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt(); 1333 InsertNewInstBefore(NC, *I); 1334 } else { 1335 // Otherwise, it's a call, just insert cast right after the call. 1336 InsertNewInstBefore(NC, *Caller); 1337 } 1338 Worklist.AddUsersToWorkList(*Caller); 1339 } else { 1340 NV = UndefValue::get(Caller->getType()); 1341 } 1342 } 1343 1344 if (!Caller->use_empty()) 1345 ReplaceInstUsesWith(*Caller, NV); 1346 1347 EraseInstFromFunction(*Caller); 1348 return true; 1349} 1350 1351// transformCallThroughTrampoline - Turn a call to a function created by 1352// init_trampoline / adjust_trampoline intrinsic pair into a direct call to the 1353// underlying function. 1354// 1355Instruction * 1356InstCombiner::transformCallThroughTrampoline(CallSite CS, 1357 IntrinsicInst *Tramp) { 1358 Value *Callee = CS.getCalledValue(); 1359 PointerType *PTy = cast<PointerType>(Callee->getType()); 1360 FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 1361 const AttrListPtr &Attrs = CS.getAttributes(); 1362 1363 // If the call already has the 'nest' attribute somewhere then give up - 1364 // otherwise 'nest' would occur twice after splicing in the chain. 1365 if (Attrs.hasAttrSomewhere(Attribute::Nest)) 1366 return 0; 1367 1368 assert(Tramp && 1369 "transformCallThroughTrampoline called with incorrect CallSite."); 1370 1371 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts()); 1372 PointerType *NestFPTy = cast<PointerType>(NestF->getType()); 1373 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType()); 1374 1375 const AttrListPtr &NestAttrs = NestF->getAttributes(); 1376 if (!NestAttrs.isEmpty()) { 1377 unsigned NestIdx = 1; 1378 Type *NestTy = 0; 1379 Attributes NestAttr = Attribute::None; 1380 1381 // Look for a parameter marked with the 'nest' attribute. 1382 for (FunctionType::param_iterator I = NestFTy->param_begin(), 1383 E = NestFTy->param_end(); I != E; ++NestIdx, ++I) 1384 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) { 1385 // Record the parameter type and any other attributes. 1386 NestTy = *I; 1387 NestAttr = NestAttrs.getParamAttributes(NestIdx); 1388 break; 1389 } 1390 1391 if (NestTy) { 1392 Instruction *Caller = CS.getInstruction(); 1393 std::vector<Value*> NewArgs; 1394 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1); 1395 1396 SmallVector<AttributeWithIndex, 8> NewAttrs; 1397 NewAttrs.reserve(Attrs.getNumSlots() + 1); 1398 1399 // Insert the nest argument into the call argument list, which may 1400 // mean appending it. Likewise for attributes. 1401 1402 // Add any result attributes. 1403 if (Attributes Attr = Attrs.getRetAttributes()) 1404 NewAttrs.push_back(AttributeWithIndex::get(0, Attr)); 1405 1406 { 1407 unsigned Idx = 1; 1408 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 1409 do { 1410 if (Idx == NestIdx) { 1411 // Add the chain argument and attributes. 1412 Value *NestVal = Tramp->getArgOperand(2); 1413 if (NestVal->getType() != NestTy) 1414 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest"); 1415 NewArgs.push_back(NestVal); 1416 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr)); 1417 } 1418 1419 if (I == E) 1420 break; 1421 1422 // Add the original argument and attributes. 1423 NewArgs.push_back(*I); 1424 if (Attributes Attr = Attrs.getParamAttributes(Idx)) 1425 NewAttrs.push_back 1426 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr)); 1427 1428 ++Idx, ++I; 1429 } while (1); 1430 } 1431 1432 // Add any function attributes. 1433 if (Attributes Attr = Attrs.getFnAttributes()) 1434 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr)); 1435 1436 // The trampoline may have been bitcast to a bogus type (FTy). 1437 // Handle this by synthesizing a new function type, equal to FTy 1438 // with the chain parameter inserted. 1439 1440 std::vector<Type*> NewTypes; 1441 NewTypes.reserve(FTy->getNumParams()+1); 1442 1443 // Insert the chain's type into the list of parameter types, which may 1444 // mean appending it. 1445 { 1446 unsigned Idx = 1; 1447 FunctionType::param_iterator I = FTy->param_begin(), 1448 E = FTy->param_end(); 1449 1450 do { 1451 if (Idx == NestIdx) 1452 // Add the chain's type. 1453 NewTypes.push_back(NestTy); 1454 1455 if (I == E) 1456 break; 1457 1458 // Add the original type. 1459 NewTypes.push_back(*I); 1460 1461 ++Idx, ++I; 1462 } while (1); 1463 } 1464 1465 // Replace the trampoline call with a direct call. Let the generic 1466 // code sort out any function type mismatches. 1467 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes, 1468 FTy->isVarArg()); 1469 Constant *NewCallee = 1470 NestF->getType() == PointerType::getUnqual(NewFTy) ? 1471 NestF : ConstantExpr::getBitCast(NestF, 1472 PointerType::getUnqual(NewFTy)); 1473 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(), 1474 NewAttrs.end()); 1475 1476 Instruction *NewCaller; 1477 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 1478 NewCaller = InvokeInst::Create(NewCallee, 1479 II->getNormalDest(), II->getUnwindDest(), 1480 NewArgs); 1481 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv()); 1482 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL); 1483 } else { 1484 NewCaller = CallInst::Create(NewCallee, NewArgs); 1485 if (cast<CallInst>(Caller)->isTailCall()) 1486 cast<CallInst>(NewCaller)->setTailCall(); 1487 cast<CallInst>(NewCaller)-> 1488 setCallingConv(cast<CallInst>(Caller)->getCallingConv()); 1489 cast<CallInst>(NewCaller)->setAttributes(NewPAL); 1490 } 1491 1492 return NewCaller; 1493 } 1494 } 1495 1496 // Replace the trampoline call with a direct call. Since there is no 'nest' 1497 // parameter, there is no need to adjust the argument list. Let the generic 1498 // code sort out any function type mismatches. 1499 Constant *NewCallee = 1500 NestF->getType() == PTy ? NestF : 1501 ConstantExpr::getBitCast(NestF, PTy); 1502 CS.setCalledFunction(NewCallee); 1503 return CS.getInstruction(); 1504} 1505