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 "InstCombineInternal.h" 15 #include "llvm/ADT/Statistic.h" 16 #include "llvm/Analysis/InstructionSimplify.h" 17 #include "llvm/Analysis/MemoryBuiltins.h" 18 #include "llvm/IR/CallSite.h" 19 #include "llvm/IR/Dominators.h" 20 #include "llvm/IR/PatternMatch.h" 21 #include "llvm/IR/Statepoint.h" 22 #include "llvm/Transforms/Utils/BuildLibCalls.h" 23 #include "llvm/Transforms/Utils/Local.h" 24 #include "llvm/Transforms/Utils/SimplifyLibCalls.h" 25 using namespace llvm; 26 using namespace PatternMatch; 27 28 #define DEBUG_TYPE "instcombine" 29 30 STATISTIC(NumSimplified, "Number of library calls simplified"); 31 32 /// getPromotedType - Return the specified type promoted as it would be to pass 33 /// though a va_arg area. 34 static Type *getPromotedType(Type *Ty) { 35 if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) { 36 if (ITy->getBitWidth() < 32) 37 return Type::getInt32Ty(Ty->getContext()); 38 } 39 return Ty; 40 } 41 42 /// reduceToSingleValueType - Given an aggregate type which ultimately holds a 43 /// single scalar element, like {{{type}}} or [1 x type], return type. 44 static Type *reduceToSingleValueType(Type *T) { 45 while (!T->isSingleValueType()) { 46 if (StructType *STy = dyn_cast<StructType>(T)) { 47 if (STy->getNumElements() == 1) 48 T = STy->getElementType(0); 49 else 50 break; 51 } else if (ArrayType *ATy = dyn_cast<ArrayType>(T)) { 52 if (ATy->getNumElements() == 1) 53 T = ATy->getElementType(); 54 else 55 break; 56 } else 57 break; 58 } 59 60 return T; 61 } 62 63 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) { 64 unsigned DstAlign = getKnownAlignment(MI->getArgOperand(0), DL, MI, AC, DT); 65 unsigned SrcAlign = getKnownAlignment(MI->getArgOperand(1), DL, MI, AC, DT); 66 unsigned MinAlign = std::min(DstAlign, SrcAlign); 67 unsigned CopyAlign = MI->getAlignment(); 68 69 if (CopyAlign < MinAlign) { 70 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), MinAlign, false)); 71 return MI; 72 } 73 74 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with 75 // load/store. 76 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getArgOperand(2)); 77 if (!MemOpLength) return nullptr; 78 79 // Source and destination pointer types are always "i8*" for intrinsic. See 80 // if the size is something we can handle with a single primitive load/store. 81 // A single load+store correctly handles overlapping memory in the memmove 82 // case. 83 uint64_t Size = MemOpLength->getLimitedValue(); 84 assert(Size && "0-sized memory transferring should be removed already."); 85 86 if (Size > 8 || (Size&(Size-1))) 87 return nullptr; // If not 1/2/4/8 bytes, exit. 88 89 // Use an integer load+store unless we can find something better. 90 unsigned SrcAddrSp = 91 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace(); 92 unsigned DstAddrSp = 93 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace(); 94 95 IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3); 96 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp); 97 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp); 98 99 // Memcpy forces the use of i8* for the source and destination. That means 100 // that if you're using memcpy to move one double around, you'll get a cast 101 // from double* to i8*. We'd much rather use a double load+store rather than 102 // an i64 load+store, here because this improves the odds that the source or 103 // dest address will be promotable. See if we can find a better type than the 104 // integer datatype. 105 Value *StrippedDest = MI->getArgOperand(0)->stripPointerCasts(); 106 MDNode *CopyMD = nullptr; 107 if (StrippedDest != MI->getArgOperand(0)) { 108 Type *SrcETy = cast<PointerType>(StrippedDest->getType()) 109 ->getElementType(); 110 if (SrcETy->isSized() && DL.getTypeStoreSize(SrcETy) == Size) { 111 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip 112 // down through these levels if so. 113 SrcETy = reduceToSingleValueType(SrcETy); 114 115 if (SrcETy->isSingleValueType()) { 116 NewSrcPtrTy = PointerType::get(SrcETy, SrcAddrSp); 117 NewDstPtrTy = PointerType::get(SrcETy, DstAddrSp); 118 119 // If the memcpy has metadata describing the members, see if we can 120 // get the TBAA tag describing our copy. 121 if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa_struct)) { 122 if (M->getNumOperands() == 3 && M->getOperand(0) && 123 mdconst::hasa<ConstantInt>(M->getOperand(0)) && 124 mdconst::extract<ConstantInt>(M->getOperand(0))->isNullValue() && 125 M->getOperand(1) && 126 mdconst::hasa<ConstantInt>(M->getOperand(1)) && 127 mdconst::extract<ConstantInt>(M->getOperand(1))->getValue() == 128 Size && 129 M->getOperand(2) && isa<MDNode>(M->getOperand(2))) 130 CopyMD = cast<MDNode>(M->getOperand(2)); 131 } 132 } 133 } 134 } 135 136 // If the memcpy/memmove provides better alignment info than we can 137 // infer, use it. 138 SrcAlign = std::max(SrcAlign, CopyAlign); 139 DstAlign = std::max(DstAlign, CopyAlign); 140 141 Value *Src = Builder->CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy); 142 Value *Dest = Builder->CreateBitCast(MI->getArgOperand(0), NewDstPtrTy); 143 LoadInst *L = Builder->CreateLoad(Src, MI->isVolatile()); 144 L->setAlignment(SrcAlign); 145 if (CopyMD) 146 L->setMetadata(LLVMContext::MD_tbaa, CopyMD); 147 StoreInst *S = Builder->CreateStore(L, Dest, MI->isVolatile()); 148 S->setAlignment(DstAlign); 149 if (CopyMD) 150 S->setMetadata(LLVMContext::MD_tbaa, CopyMD); 151 152 // Set the size of the copy to 0, it will be deleted on the next iteration. 153 MI->setArgOperand(2, Constant::getNullValue(MemOpLength->getType())); 154 return MI; 155 } 156 157 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) { 158 unsigned Alignment = getKnownAlignment(MI->getDest(), DL, MI, AC, DT); 159 if (MI->getAlignment() < Alignment) { 160 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), 161 Alignment, false)); 162 return MI; 163 } 164 165 // Extract the length and alignment and fill if they are constant. 166 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength()); 167 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue()); 168 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8)) 169 return nullptr; 170 uint64_t Len = LenC->getLimitedValue(); 171 Alignment = MI->getAlignment(); 172 assert(Len && "0-sized memory setting should be removed already."); 173 174 // memset(s,c,n) -> store s, c (for n=1,2,4,8) 175 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) { 176 Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8. 177 178 Value *Dest = MI->getDest(); 179 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace(); 180 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp); 181 Dest = Builder->CreateBitCast(Dest, NewDstPtrTy); 182 183 // Alignment 0 is identity for alignment 1 for memset, but not store. 184 if (Alignment == 0) Alignment = 1; 185 186 // Extract the fill value and store. 187 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL; 188 StoreInst *S = Builder->CreateStore(ConstantInt::get(ITy, Fill), Dest, 189 MI->isVolatile()); 190 S->setAlignment(Alignment); 191 192 // Set the size of the copy to 0, it will be deleted on the next iteration. 193 MI->setLength(Constant::getNullValue(LenC->getType())); 194 return MI; 195 } 196 197 return nullptr; 198 } 199 200 static Value *SimplifyX86immshift(const IntrinsicInst &II, 201 InstCombiner::BuilderTy &Builder) { 202 bool LogicalShift = false; 203 bool ShiftLeft = false; 204 205 switch (II.getIntrinsicID()) { 206 default: 207 return nullptr; 208 case Intrinsic::x86_sse2_psra_d: 209 case Intrinsic::x86_sse2_psra_w: 210 case Intrinsic::x86_sse2_psrai_d: 211 case Intrinsic::x86_sse2_psrai_w: 212 case Intrinsic::x86_avx2_psra_d: 213 case Intrinsic::x86_avx2_psra_w: 214 case Intrinsic::x86_avx2_psrai_d: 215 case Intrinsic::x86_avx2_psrai_w: 216 LogicalShift = false; ShiftLeft = false; 217 break; 218 case Intrinsic::x86_sse2_psrl_d: 219 case Intrinsic::x86_sse2_psrl_q: 220 case Intrinsic::x86_sse2_psrl_w: 221 case Intrinsic::x86_sse2_psrli_d: 222 case Intrinsic::x86_sse2_psrli_q: 223 case Intrinsic::x86_sse2_psrli_w: 224 case Intrinsic::x86_avx2_psrl_d: 225 case Intrinsic::x86_avx2_psrl_q: 226 case Intrinsic::x86_avx2_psrl_w: 227 case Intrinsic::x86_avx2_psrli_d: 228 case Intrinsic::x86_avx2_psrli_q: 229 case Intrinsic::x86_avx2_psrli_w: 230 LogicalShift = true; ShiftLeft = false; 231 break; 232 case Intrinsic::x86_sse2_psll_d: 233 case Intrinsic::x86_sse2_psll_q: 234 case Intrinsic::x86_sse2_psll_w: 235 case Intrinsic::x86_sse2_pslli_d: 236 case Intrinsic::x86_sse2_pslli_q: 237 case Intrinsic::x86_sse2_pslli_w: 238 case Intrinsic::x86_avx2_psll_d: 239 case Intrinsic::x86_avx2_psll_q: 240 case Intrinsic::x86_avx2_psll_w: 241 case Intrinsic::x86_avx2_pslli_d: 242 case Intrinsic::x86_avx2_pslli_q: 243 case Intrinsic::x86_avx2_pslli_w: 244 LogicalShift = true; ShiftLeft = true; 245 break; 246 } 247 assert((LogicalShift || !ShiftLeft) && "Only logical shifts can shift left"); 248 249 // Simplify if count is constant. 250 auto Arg1 = II.getArgOperand(1); 251 auto CAZ = dyn_cast<ConstantAggregateZero>(Arg1); 252 auto CDV = dyn_cast<ConstantDataVector>(Arg1); 253 auto CInt = dyn_cast<ConstantInt>(Arg1); 254 if (!CAZ && !CDV && !CInt) 255 return nullptr; 256 257 APInt Count(64, 0); 258 if (CDV) { 259 // SSE2/AVX2 uses all the first 64-bits of the 128-bit vector 260 // operand to compute the shift amount. 261 auto VT = cast<VectorType>(CDV->getType()); 262 unsigned BitWidth = VT->getElementType()->getPrimitiveSizeInBits(); 263 assert((64 % BitWidth) == 0 && "Unexpected packed shift size"); 264 unsigned NumSubElts = 64 / BitWidth; 265 266 // Concatenate the sub-elements to create the 64-bit value. 267 for (unsigned i = 0; i != NumSubElts; ++i) { 268 unsigned SubEltIdx = (NumSubElts - 1) - i; 269 auto SubElt = cast<ConstantInt>(CDV->getElementAsConstant(SubEltIdx)); 270 Count = Count.shl(BitWidth); 271 Count |= SubElt->getValue().zextOrTrunc(64); 272 } 273 } 274 else if (CInt) 275 Count = CInt->getValue(); 276 277 auto Vec = II.getArgOperand(0); 278 auto VT = cast<VectorType>(Vec->getType()); 279 auto SVT = VT->getElementType(); 280 unsigned VWidth = VT->getNumElements(); 281 unsigned BitWidth = SVT->getPrimitiveSizeInBits(); 282 283 // If shift-by-zero then just return the original value. 284 if (Count == 0) 285 return Vec; 286 287 // Handle cases when Shift >= BitWidth. 288 if (Count.uge(BitWidth)) { 289 // If LogicalShift - just return zero. 290 if (LogicalShift) 291 return ConstantAggregateZero::get(VT); 292 293 // If ArithmeticShift - clamp Shift to (BitWidth - 1). 294 Count = APInt(64, BitWidth - 1); 295 } 296 297 // Get a constant vector of the same type as the first operand. 298 auto ShiftAmt = ConstantInt::get(SVT, Count.zextOrTrunc(BitWidth)); 299 auto ShiftVec = Builder.CreateVectorSplat(VWidth, ShiftAmt); 300 301 if (ShiftLeft) 302 return Builder.CreateShl(Vec, ShiftVec); 303 304 if (LogicalShift) 305 return Builder.CreateLShr(Vec, ShiftVec); 306 307 return Builder.CreateAShr(Vec, ShiftVec); 308 } 309 310 static Value *SimplifyX86extend(const IntrinsicInst &II, 311 InstCombiner::BuilderTy &Builder, 312 bool SignExtend) { 313 VectorType *SrcTy = cast<VectorType>(II.getArgOperand(0)->getType()); 314 VectorType *DstTy = cast<VectorType>(II.getType()); 315 unsigned NumDstElts = DstTy->getNumElements(); 316 317 // Extract a subvector of the first NumDstElts lanes and sign/zero extend. 318 SmallVector<int, 8> ShuffleMask; 319 for (int i = 0; i != (int)NumDstElts; ++i) 320 ShuffleMask.push_back(i); 321 322 Value *SV = Builder.CreateShuffleVector(II.getArgOperand(0), 323 UndefValue::get(SrcTy), ShuffleMask); 324 return SignExtend ? Builder.CreateSExt(SV, DstTy) 325 : Builder.CreateZExt(SV, DstTy); 326 } 327 328 static Value *SimplifyX86insertps(const IntrinsicInst &II, 329 InstCombiner::BuilderTy &Builder) { 330 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) { 331 VectorType *VecTy = cast<VectorType>(II.getType()); 332 assert(VecTy->getNumElements() == 4 && "insertps with wrong vector type"); 333 334 // The immediate permute control byte looks like this: 335 // [3:0] - zero mask for each 32-bit lane 336 // [5:4] - select one 32-bit destination lane 337 // [7:6] - select one 32-bit source lane 338 339 uint8_t Imm = CInt->getZExtValue(); 340 uint8_t ZMask = Imm & 0xf; 341 uint8_t DestLane = (Imm >> 4) & 0x3; 342 uint8_t SourceLane = (Imm >> 6) & 0x3; 343 344 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy); 345 346 // If all zero mask bits are set, this was just a weird way to 347 // generate a zero vector. 348 if (ZMask == 0xf) 349 return ZeroVector; 350 351 // Initialize by passing all of the first source bits through. 352 int ShuffleMask[4] = { 0, 1, 2, 3 }; 353 354 // We may replace the second operand with the zero vector. 355 Value *V1 = II.getArgOperand(1); 356 357 if (ZMask) { 358 // If the zero mask is being used with a single input or the zero mask 359 // overrides the destination lane, this is a shuffle with the zero vector. 360 if ((II.getArgOperand(0) == II.getArgOperand(1)) || 361 (ZMask & (1 << DestLane))) { 362 V1 = ZeroVector; 363 // We may still move 32-bits of the first source vector from one lane 364 // to another. 365 ShuffleMask[DestLane] = SourceLane; 366 // The zero mask may override the previous insert operation. 367 for (unsigned i = 0; i < 4; ++i) 368 if ((ZMask >> i) & 0x1) 369 ShuffleMask[i] = i + 4; 370 } else { 371 // TODO: Model this case as 2 shuffles or a 'logical and' plus shuffle? 372 return nullptr; 373 } 374 } else { 375 // Replace the selected destination lane with the selected source lane. 376 ShuffleMask[DestLane] = SourceLane + 4; 377 } 378 379 return Builder.CreateShuffleVector(II.getArgOperand(0), V1, ShuffleMask); 380 } 381 return nullptr; 382 } 383 384 /// Attempt to simplify SSE4A EXTRQ/EXTRQI instructions using constant folding 385 /// or conversion to a shuffle vector. 386 static Value *SimplifyX86extrq(IntrinsicInst &II, Value *Op0, 387 ConstantInt *CILength, ConstantInt *CIIndex, 388 InstCombiner::BuilderTy &Builder) { 389 auto LowConstantHighUndef = [&](uint64_t Val) { 390 Type *IntTy64 = Type::getInt64Ty(II.getContext()); 391 Constant *Args[] = {ConstantInt::get(IntTy64, Val), 392 UndefValue::get(IntTy64)}; 393 return ConstantVector::get(Args); 394 }; 395 396 // See if we're dealing with constant values. 397 Constant *C0 = dyn_cast<Constant>(Op0); 398 ConstantInt *CI0 = 399 C0 ? dyn_cast<ConstantInt>(C0->getAggregateElement((unsigned)0)) 400 : nullptr; 401 402 // Attempt to constant fold. 403 if (CILength && CIIndex) { 404 // From AMD documentation: "The bit index and field length are each six 405 // bits in length other bits of the field are ignored." 406 APInt APIndex = CIIndex->getValue().zextOrTrunc(6); 407 APInt APLength = CILength->getValue().zextOrTrunc(6); 408 409 unsigned Index = APIndex.getZExtValue(); 410 411 // From AMD documentation: "a value of zero in the field length is 412 // defined as length of 64". 413 unsigned Length = APLength == 0 ? 64 : APLength.getZExtValue(); 414 415 // From AMD documentation: "If the sum of the bit index + length field 416 // is greater than 64, the results are undefined". 417 unsigned End = Index + Length; 418 419 // Note that both field index and field length are 8-bit quantities. 420 // Since variables 'Index' and 'Length' are unsigned values 421 // obtained from zero-extending field index and field length 422 // respectively, their sum should never wrap around. 423 if (End > 64) 424 return UndefValue::get(II.getType()); 425 426 // If we are inserting whole bytes, we can convert this to a shuffle. 427 // Lowering can recognize EXTRQI shuffle masks. 428 if ((Length % 8) == 0 && (Index % 8) == 0) { 429 // Convert bit indices to byte indices. 430 Length /= 8; 431 Index /= 8; 432 433 Type *IntTy8 = Type::getInt8Ty(II.getContext()); 434 Type *IntTy32 = Type::getInt32Ty(II.getContext()); 435 VectorType *ShufTy = VectorType::get(IntTy8, 16); 436 437 SmallVector<Constant *, 16> ShuffleMask; 438 for (int i = 0; i != (int)Length; ++i) 439 ShuffleMask.push_back( 440 Constant::getIntegerValue(IntTy32, APInt(32, i + Index))); 441 for (int i = Length; i != 8; ++i) 442 ShuffleMask.push_back( 443 Constant::getIntegerValue(IntTy32, APInt(32, i + 16))); 444 for (int i = 8; i != 16; ++i) 445 ShuffleMask.push_back(UndefValue::get(IntTy32)); 446 447 Value *SV = Builder.CreateShuffleVector( 448 Builder.CreateBitCast(Op0, ShufTy), 449 ConstantAggregateZero::get(ShufTy), ConstantVector::get(ShuffleMask)); 450 return Builder.CreateBitCast(SV, II.getType()); 451 } 452 453 // Constant Fold - shift Index'th bit to lowest position and mask off 454 // Length bits. 455 if (CI0) { 456 APInt Elt = CI0->getValue(); 457 Elt = Elt.lshr(Index).zextOrTrunc(Length); 458 return LowConstantHighUndef(Elt.getZExtValue()); 459 } 460 461 // If we were an EXTRQ call, we'll save registers if we convert to EXTRQI. 462 if (II.getIntrinsicID() == Intrinsic::x86_sse4a_extrq) { 463 Value *Args[] = {Op0, CILength, CIIndex}; 464 Module *M = II.getModule(); 465 Value *F = Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_extrqi); 466 return Builder.CreateCall(F, Args); 467 } 468 } 469 470 // Constant Fold - extraction from zero is always {zero, undef}. 471 if (CI0 && CI0->equalsInt(0)) 472 return LowConstantHighUndef(0); 473 474 return nullptr; 475 } 476 477 /// Attempt to simplify SSE4A INSERTQ/INSERTQI instructions using constant 478 /// folding or conversion to a shuffle vector. 479 static Value *SimplifyX86insertq(IntrinsicInst &II, Value *Op0, Value *Op1, 480 APInt APLength, APInt APIndex, 481 InstCombiner::BuilderTy &Builder) { 482 483 // From AMD documentation: "The bit index and field length are each six bits 484 // in length other bits of the field are ignored." 485 APIndex = APIndex.zextOrTrunc(6); 486 APLength = APLength.zextOrTrunc(6); 487 488 // Attempt to constant fold. 489 unsigned Index = APIndex.getZExtValue(); 490 491 // From AMD documentation: "a value of zero in the field length is 492 // defined as length of 64". 493 unsigned Length = APLength == 0 ? 64 : APLength.getZExtValue(); 494 495 // From AMD documentation: "If the sum of the bit index + length field 496 // is greater than 64, the results are undefined". 497 unsigned End = Index + Length; 498 499 // Note that both field index and field length are 8-bit quantities. 500 // Since variables 'Index' and 'Length' are unsigned values 501 // obtained from zero-extending field index and field length 502 // respectively, their sum should never wrap around. 503 if (End > 64) 504 return UndefValue::get(II.getType()); 505 506 // If we are inserting whole bytes, we can convert this to a shuffle. 507 // Lowering can recognize INSERTQI shuffle masks. 508 if ((Length % 8) == 0 && (Index % 8) == 0) { 509 // Convert bit indices to byte indices. 510 Length /= 8; 511 Index /= 8; 512 513 Type *IntTy8 = Type::getInt8Ty(II.getContext()); 514 Type *IntTy32 = Type::getInt32Ty(II.getContext()); 515 VectorType *ShufTy = VectorType::get(IntTy8, 16); 516 517 SmallVector<Constant *, 16> ShuffleMask; 518 for (int i = 0; i != (int)Index; ++i) 519 ShuffleMask.push_back(Constant::getIntegerValue(IntTy32, APInt(32, i))); 520 for (int i = 0; i != (int)Length; ++i) 521 ShuffleMask.push_back( 522 Constant::getIntegerValue(IntTy32, APInt(32, i + 16))); 523 for (int i = Index + Length; i != 8; ++i) 524 ShuffleMask.push_back(Constant::getIntegerValue(IntTy32, APInt(32, i))); 525 for (int i = 8; i != 16; ++i) 526 ShuffleMask.push_back(UndefValue::get(IntTy32)); 527 528 Value *SV = Builder.CreateShuffleVector(Builder.CreateBitCast(Op0, ShufTy), 529 Builder.CreateBitCast(Op1, ShufTy), 530 ConstantVector::get(ShuffleMask)); 531 return Builder.CreateBitCast(SV, II.getType()); 532 } 533 534 // See if we're dealing with constant values. 535 Constant *C0 = dyn_cast<Constant>(Op0); 536 Constant *C1 = dyn_cast<Constant>(Op1); 537 ConstantInt *CI00 = 538 C0 ? dyn_cast<ConstantInt>(C0->getAggregateElement((unsigned)0)) 539 : nullptr; 540 ConstantInt *CI10 = 541 C1 ? dyn_cast<ConstantInt>(C1->getAggregateElement((unsigned)0)) 542 : nullptr; 543 544 // Constant Fold - insert bottom Length bits starting at the Index'th bit. 545 if (CI00 && CI10) { 546 APInt V00 = CI00->getValue(); 547 APInt V10 = CI10->getValue(); 548 APInt Mask = APInt::getLowBitsSet(64, Length).shl(Index); 549 V00 = V00 & ~Mask; 550 V10 = V10.zextOrTrunc(Length).zextOrTrunc(64).shl(Index); 551 APInt Val = V00 | V10; 552 Type *IntTy64 = Type::getInt64Ty(II.getContext()); 553 Constant *Args[] = {ConstantInt::get(IntTy64, Val.getZExtValue()), 554 UndefValue::get(IntTy64)}; 555 return ConstantVector::get(Args); 556 } 557 558 // If we were an INSERTQ call, we'll save demanded elements if we convert to 559 // INSERTQI. 560 if (II.getIntrinsicID() == Intrinsic::x86_sse4a_insertq) { 561 Type *IntTy8 = Type::getInt8Ty(II.getContext()); 562 Constant *CILength = ConstantInt::get(IntTy8, Length, false); 563 Constant *CIIndex = ConstantInt::get(IntTy8, Index, false); 564 565 Value *Args[] = {Op0, Op1, CILength, CIIndex}; 566 Module *M = II.getModule(); 567 Value *F = Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi); 568 return Builder.CreateCall(F, Args); 569 } 570 571 return nullptr; 572 } 573 574 /// The shuffle mask for a perm2*128 selects any two halves of two 256-bit 575 /// source vectors, unless a zero bit is set. If a zero bit is set, 576 /// then ignore that half of the mask and clear that half of the vector. 577 static Value *SimplifyX86vperm2(const IntrinsicInst &II, 578 InstCombiner::BuilderTy &Builder) { 579 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) { 580 VectorType *VecTy = cast<VectorType>(II.getType()); 581 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy); 582 583 // The immediate permute control byte looks like this: 584 // [1:0] - select 128 bits from sources for low half of destination 585 // [2] - ignore 586 // [3] - zero low half of destination 587 // [5:4] - select 128 bits from sources for high half of destination 588 // [6] - ignore 589 // [7] - zero high half of destination 590 591 uint8_t Imm = CInt->getZExtValue(); 592 593 bool LowHalfZero = Imm & 0x08; 594 bool HighHalfZero = Imm & 0x80; 595 596 // If both zero mask bits are set, this was just a weird way to 597 // generate a zero vector. 598 if (LowHalfZero && HighHalfZero) 599 return ZeroVector; 600 601 // If 0 or 1 zero mask bits are set, this is a simple shuffle. 602 unsigned NumElts = VecTy->getNumElements(); 603 unsigned HalfSize = NumElts / 2; 604 SmallVector<int, 8> ShuffleMask(NumElts); 605 606 // The high bit of the selection field chooses the 1st or 2nd operand. 607 bool LowInputSelect = Imm & 0x02; 608 bool HighInputSelect = Imm & 0x20; 609 610 // The low bit of the selection field chooses the low or high half 611 // of the selected operand. 612 bool LowHalfSelect = Imm & 0x01; 613 bool HighHalfSelect = Imm & 0x10; 614 615 // Determine which operand(s) are actually in use for this instruction. 616 Value *V0 = LowInputSelect ? II.getArgOperand(1) : II.getArgOperand(0); 617 Value *V1 = HighInputSelect ? II.getArgOperand(1) : II.getArgOperand(0); 618 619 // If needed, replace operands based on zero mask. 620 V0 = LowHalfZero ? ZeroVector : V0; 621 V1 = HighHalfZero ? ZeroVector : V1; 622 623 // Permute low half of result. 624 unsigned StartIndex = LowHalfSelect ? HalfSize : 0; 625 for (unsigned i = 0; i < HalfSize; ++i) 626 ShuffleMask[i] = StartIndex + i; 627 628 // Permute high half of result. 629 StartIndex = HighHalfSelect ? HalfSize : 0; 630 StartIndex += NumElts; 631 for (unsigned i = 0; i < HalfSize; ++i) 632 ShuffleMask[i + HalfSize] = StartIndex + i; 633 634 return Builder.CreateShuffleVector(V0, V1, ShuffleMask); 635 } 636 return nullptr; 637 } 638 639 /// Decode XOP integer vector comparison intrinsics. 640 static Value *SimplifyX86vpcom(const IntrinsicInst &II, 641 InstCombiner::BuilderTy &Builder, bool IsSigned) { 642 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) { 643 uint64_t Imm = CInt->getZExtValue() & 0x7; 644 VectorType *VecTy = cast<VectorType>(II.getType()); 645 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE; 646 647 switch (Imm) { 648 case 0x0: 649 Pred = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 650 break; 651 case 0x1: 652 Pred = IsSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 653 break; 654 case 0x2: 655 Pred = IsSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 656 break; 657 case 0x3: 658 Pred = IsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 659 break; 660 case 0x4: 661 Pred = ICmpInst::ICMP_EQ; break; 662 case 0x5: 663 Pred = ICmpInst::ICMP_NE; break; 664 case 0x6: 665 return ConstantInt::getSigned(VecTy, 0); // FALSE 666 case 0x7: 667 return ConstantInt::getSigned(VecTy, -1); // TRUE 668 } 669 670 if (Value *Cmp = Builder.CreateICmp(Pred, II.getArgOperand(0), II.getArgOperand(1))) 671 return Builder.CreateSExtOrTrunc(Cmp, VecTy); 672 } 673 return nullptr; 674 } 675 676 /// visitCallInst - CallInst simplification. This mostly only handles folding 677 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do 678 /// the heavy lifting. 679 /// 680 Instruction *InstCombiner::visitCallInst(CallInst &CI) { 681 auto Args = CI.arg_operands(); 682 if (Value *V = SimplifyCall(CI.getCalledValue(), Args.begin(), Args.end(), DL, 683 TLI, DT, AC)) 684 return ReplaceInstUsesWith(CI, V); 685 686 if (isFreeCall(&CI, TLI)) 687 return visitFree(CI); 688 689 // If the caller function is nounwind, mark the call as nounwind, even if the 690 // callee isn't. 691 if (CI.getParent()->getParent()->doesNotThrow() && 692 !CI.doesNotThrow()) { 693 CI.setDoesNotThrow(); 694 return &CI; 695 } 696 697 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI); 698 if (!II) return visitCallSite(&CI); 699 700 // Intrinsics cannot occur in an invoke, so handle them here instead of in 701 // visitCallSite. 702 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) { 703 bool Changed = false; 704 705 // memmove/cpy/set of zero bytes is a noop. 706 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) { 707 if (NumBytes->isNullValue()) 708 return EraseInstFromFunction(CI); 709 710 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes)) 711 if (CI->getZExtValue() == 1) { 712 // Replace the instruction with just byte operations. We would 713 // transform other cases to loads/stores, but we don't know if 714 // alignment is sufficient. 715 } 716 } 717 718 // No other transformations apply to volatile transfers. 719 if (MI->isVolatile()) 720 return nullptr; 721 722 // If we have a memmove and the source operation is a constant global, 723 // then the source and dest pointers can't alias, so we can change this 724 // into a call to memcpy. 725 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) { 726 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource())) 727 if (GVSrc->isConstant()) { 728 Module *M = CI.getModule(); 729 Intrinsic::ID MemCpyID = Intrinsic::memcpy; 730 Type *Tys[3] = { CI.getArgOperand(0)->getType(), 731 CI.getArgOperand(1)->getType(), 732 CI.getArgOperand(2)->getType() }; 733 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys)); 734 Changed = true; 735 } 736 } 737 738 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { 739 // memmove(x,x,size) -> noop. 740 if (MTI->getSource() == MTI->getDest()) 741 return EraseInstFromFunction(CI); 742 } 743 744 // If we can determine a pointer alignment that is bigger than currently 745 // set, update the alignment. 746 if (isa<MemTransferInst>(MI)) { 747 if (Instruction *I = SimplifyMemTransfer(MI)) 748 return I; 749 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) { 750 if (Instruction *I = SimplifyMemSet(MSI)) 751 return I; 752 } 753 754 if (Changed) return II; 755 } 756 757 auto SimplifyDemandedVectorEltsLow = [this](Value *Op, unsigned Width, unsigned DemandedWidth) 758 { 759 APInt UndefElts(Width, 0); 760 APInt DemandedElts = APInt::getLowBitsSet(Width, DemandedWidth); 761 return SimplifyDemandedVectorElts(Op, DemandedElts, UndefElts); 762 }; 763 764 switch (II->getIntrinsicID()) { 765 default: break; 766 case Intrinsic::objectsize: { 767 uint64_t Size; 768 if (getObjectSize(II->getArgOperand(0), Size, DL, TLI)) 769 return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size)); 770 return nullptr; 771 } 772 case Intrinsic::bswap: { 773 Value *IIOperand = II->getArgOperand(0); 774 Value *X = nullptr; 775 776 // bswap(bswap(x)) -> x 777 if (match(IIOperand, m_BSwap(m_Value(X)))) 778 return ReplaceInstUsesWith(CI, X); 779 780 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c)) 781 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) { 782 unsigned C = X->getType()->getPrimitiveSizeInBits() - 783 IIOperand->getType()->getPrimitiveSizeInBits(); 784 Value *CV = ConstantInt::get(X->getType(), C); 785 Value *V = Builder->CreateLShr(X, CV); 786 return new TruncInst(V, IIOperand->getType()); 787 } 788 break; 789 } 790 791 case Intrinsic::bitreverse: { 792 Value *IIOperand = II->getArgOperand(0); 793 Value *X = nullptr; 794 795 // bitreverse(bitreverse(x)) -> x 796 if (match(IIOperand, m_Intrinsic<Intrinsic::bitreverse>(m_Value(X)))) 797 return ReplaceInstUsesWith(CI, X); 798 break; 799 } 800 801 case Intrinsic::powi: 802 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 803 // powi(x, 0) -> 1.0 804 if (Power->isZero()) 805 return ReplaceInstUsesWith(CI, ConstantFP::get(CI.getType(), 1.0)); 806 // powi(x, 1) -> x 807 if (Power->isOne()) 808 return ReplaceInstUsesWith(CI, II->getArgOperand(0)); 809 // powi(x, -1) -> 1/x 810 if (Power->isAllOnesValue()) 811 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0), 812 II->getArgOperand(0)); 813 } 814 break; 815 case Intrinsic::cttz: { 816 // If all bits below the first known one are known zero, 817 // this value is constant. 818 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType()); 819 // FIXME: Try to simplify vectors of integers. 820 if (!IT) break; 821 uint32_t BitWidth = IT->getBitWidth(); 822 APInt KnownZero(BitWidth, 0); 823 APInt KnownOne(BitWidth, 0); 824 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II); 825 unsigned TrailingZeros = KnownOne.countTrailingZeros(); 826 APInt Mask(APInt::getLowBitsSet(BitWidth, TrailingZeros)); 827 if ((Mask & KnownZero) == Mask) 828 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 829 APInt(BitWidth, TrailingZeros))); 830 831 } 832 break; 833 case Intrinsic::ctlz: { 834 // If all bits above the first known one are known zero, 835 // this value is constant. 836 IntegerType *IT = dyn_cast<IntegerType>(II->getArgOperand(0)->getType()); 837 // FIXME: Try to simplify vectors of integers. 838 if (!IT) break; 839 uint32_t BitWidth = IT->getBitWidth(); 840 APInt KnownZero(BitWidth, 0); 841 APInt KnownOne(BitWidth, 0); 842 computeKnownBits(II->getArgOperand(0), KnownZero, KnownOne, 0, II); 843 unsigned LeadingZeros = KnownOne.countLeadingZeros(); 844 APInt Mask(APInt::getHighBitsSet(BitWidth, LeadingZeros)); 845 if ((Mask & KnownZero) == Mask) 846 return ReplaceInstUsesWith(CI, ConstantInt::get(IT, 847 APInt(BitWidth, LeadingZeros))); 848 849 } 850 break; 851 852 case Intrinsic::uadd_with_overflow: 853 case Intrinsic::sadd_with_overflow: 854 case Intrinsic::umul_with_overflow: 855 case Intrinsic::smul_with_overflow: 856 if (isa<Constant>(II->getArgOperand(0)) && 857 !isa<Constant>(II->getArgOperand(1))) { 858 // Canonicalize constants into the RHS. 859 Value *LHS = II->getArgOperand(0); 860 II->setArgOperand(0, II->getArgOperand(1)); 861 II->setArgOperand(1, LHS); 862 return II; 863 } 864 // fall through 865 866 case Intrinsic::usub_with_overflow: 867 case Intrinsic::ssub_with_overflow: { 868 OverflowCheckFlavor OCF = 869 IntrinsicIDToOverflowCheckFlavor(II->getIntrinsicID()); 870 assert(OCF != OCF_INVALID && "unexpected!"); 871 872 Value *OperationResult = nullptr; 873 Constant *OverflowResult = nullptr; 874 if (OptimizeOverflowCheck(OCF, II->getArgOperand(0), II->getArgOperand(1), 875 *II, OperationResult, OverflowResult)) 876 return CreateOverflowTuple(II, OperationResult, OverflowResult); 877 878 break; 879 } 880 881 case Intrinsic::minnum: 882 case Intrinsic::maxnum: { 883 Value *Arg0 = II->getArgOperand(0); 884 Value *Arg1 = II->getArgOperand(1); 885 886 // fmin(x, x) -> x 887 if (Arg0 == Arg1) 888 return ReplaceInstUsesWith(CI, Arg0); 889 890 const ConstantFP *C0 = dyn_cast<ConstantFP>(Arg0); 891 const ConstantFP *C1 = dyn_cast<ConstantFP>(Arg1); 892 893 // Canonicalize constants into the RHS. 894 if (C0 && !C1) { 895 II->setArgOperand(0, Arg1); 896 II->setArgOperand(1, Arg0); 897 return II; 898 } 899 900 // fmin(x, nan) -> x 901 if (C1 && C1->isNaN()) 902 return ReplaceInstUsesWith(CI, Arg0); 903 904 // This is the value because if undef were NaN, we would return the other 905 // value and cannot return a NaN unless both operands are. 906 // 907 // fmin(undef, x) -> x 908 if (isa<UndefValue>(Arg0)) 909 return ReplaceInstUsesWith(CI, Arg1); 910 911 // fmin(x, undef) -> x 912 if (isa<UndefValue>(Arg1)) 913 return ReplaceInstUsesWith(CI, Arg0); 914 915 Value *X = nullptr; 916 Value *Y = nullptr; 917 if (II->getIntrinsicID() == Intrinsic::minnum) { 918 // fmin(x, fmin(x, y)) -> fmin(x, y) 919 // fmin(y, fmin(x, y)) -> fmin(x, y) 920 if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) { 921 if (Arg0 == X || Arg0 == Y) 922 return ReplaceInstUsesWith(CI, Arg1); 923 } 924 925 // fmin(fmin(x, y), x) -> fmin(x, y) 926 // fmin(fmin(x, y), y) -> fmin(x, y) 927 if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) { 928 if (Arg1 == X || Arg1 == Y) 929 return ReplaceInstUsesWith(CI, Arg0); 930 } 931 932 // TODO: fmin(nnan x, inf) -> x 933 // TODO: fmin(nnan ninf x, flt_max) -> x 934 if (C1 && C1->isInfinity()) { 935 // fmin(x, -inf) -> -inf 936 if (C1->isNegative()) 937 return ReplaceInstUsesWith(CI, Arg1); 938 } 939 } else { 940 assert(II->getIntrinsicID() == Intrinsic::maxnum); 941 // fmax(x, fmax(x, y)) -> fmax(x, y) 942 // fmax(y, fmax(x, y)) -> fmax(x, y) 943 if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) { 944 if (Arg0 == X || Arg0 == Y) 945 return ReplaceInstUsesWith(CI, Arg1); 946 } 947 948 // fmax(fmax(x, y), x) -> fmax(x, y) 949 // fmax(fmax(x, y), y) -> fmax(x, y) 950 if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) { 951 if (Arg1 == X || Arg1 == Y) 952 return ReplaceInstUsesWith(CI, Arg0); 953 } 954 955 // TODO: fmax(nnan x, -inf) -> x 956 // TODO: fmax(nnan ninf x, -flt_max) -> x 957 if (C1 && C1->isInfinity()) { 958 // fmax(x, inf) -> inf 959 if (!C1->isNegative()) 960 return ReplaceInstUsesWith(CI, Arg1); 961 } 962 } 963 break; 964 } 965 case Intrinsic::ppc_altivec_lvx: 966 case Intrinsic::ppc_altivec_lvxl: 967 // Turn PPC lvx -> load if the pointer is known aligned. 968 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >= 969 16) { 970 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), 971 PointerType::getUnqual(II->getType())); 972 return new LoadInst(Ptr); 973 } 974 break; 975 case Intrinsic::ppc_vsx_lxvw4x: 976 case Intrinsic::ppc_vsx_lxvd2x: { 977 // Turn PPC VSX loads into normal loads. 978 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), 979 PointerType::getUnqual(II->getType())); 980 return new LoadInst(Ptr, Twine(""), false, 1); 981 } 982 case Intrinsic::ppc_altivec_stvx: 983 case Intrinsic::ppc_altivec_stvxl: 984 // Turn stvx -> store if the pointer is known aligned. 985 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >= 986 16) { 987 Type *OpPtrTy = 988 PointerType::getUnqual(II->getArgOperand(0)->getType()); 989 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); 990 return new StoreInst(II->getArgOperand(0), Ptr); 991 } 992 break; 993 case Intrinsic::ppc_vsx_stxvw4x: 994 case Intrinsic::ppc_vsx_stxvd2x: { 995 // Turn PPC VSX stores into normal stores. 996 Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType()); 997 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); 998 return new StoreInst(II->getArgOperand(0), Ptr, false, 1); 999 } 1000 case Intrinsic::ppc_qpx_qvlfs: 1001 // Turn PPC QPX qvlfs -> load if the pointer is known aligned. 1002 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >= 1003 16) { 1004 Type *VTy = VectorType::get(Builder->getFloatTy(), 1005 II->getType()->getVectorNumElements()); 1006 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), 1007 PointerType::getUnqual(VTy)); 1008 Value *Load = Builder->CreateLoad(Ptr); 1009 return new FPExtInst(Load, II->getType()); 1010 } 1011 break; 1012 case Intrinsic::ppc_qpx_qvlfd: 1013 // Turn PPC QPX qvlfd -> load if the pointer is known aligned. 1014 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 32, DL, II, AC, DT) >= 1015 32) { 1016 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), 1017 PointerType::getUnqual(II->getType())); 1018 return new LoadInst(Ptr); 1019 } 1020 break; 1021 case Intrinsic::ppc_qpx_qvstfs: 1022 // Turn PPC QPX qvstfs -> store if the pointer is known aligned. 1023 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, AC, DT) >= 1024 16) { 1025 Type *VTy = VectorType::get(Builder->getFloatTy(), 1026 II->getArgOperand(0)->getType()->getVectorNumElements()); 1027 Value *TOp = Builder->CreateFPTrunc(II->getArgOperand(0), VTy); 1028 Type *OpPtrTy = PointerType::getUnqual(VTy); 1029 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); 1030 return new StoreInst(TOp, Ptr); 1031 } 1032 break; 1033 case Intrinsic::ppc_qpx_qvstfd: 1034 // Turn PPC QPX qvstfd -> store if the pointer is known aligned. 1035 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, II, AC, DT) >= 1036 32) { 1037 Type *OpPtrTy = 1038 PointerType::getUnqual(II->getArgOperand(0)->getType()); 1039 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(1), OpPtrTy); 1040 return new StoreInst(II->getArgOperand(0), Ptr); 1041 } 1042 break; 1043 1044 case Intrinsic::x86_sse_storeu_ps: 1045 case Intrinsic::x86_sse2_storeu_pd: 1046 case Intrinsic::x86_sse2_storeu_dq: 1047 // Turn X86 storeu -> store if the pointer is known aligned. 1048 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, AC, DT) >= 1049 16) { 1050 Type *OpPtrTy = 1051 PointerType::getUnqual(II->getArgOperand(1)->getType()); 1052 Value *Ptr = Builder->CreateBitCast(II->getArgOperand(0), OpPtrTy); 1053 return new StoreInst(II->getArgOperand(1), Ptr); 1054 } 1055 break; 1056 1057 case Intrinsic::x86_vcvtph2ps_128: 1058 case Intrinsic::x86_vcvtph2ps_256: { 1059 auto Arg = II->getArgOperand(0); 1060 auto ArgType = cast<VectorType>(Arg->getType()); 1061 auto RetType = cast<VectorType>(II->getType()); 1062 unsigned ArgWidth = ArgType->getNumElements(); 1063 unsigned RetWidth = RetType->getNumElements(); 1064 assert(RetWidth <= ArgWidth && "Unexpected input/return vector widths"); 1065 assert(ArgType->isIntOrIntVectorTy() && 1066 ArgType->getScalarSizeInBits() == 16 && 1067 "CVTPH2PS input type should be 16-bit integer vector"); 1068 assert(RetType->getScalarType()->isFloatTy() && 1069 "CVTPH2PS output type should be 32-bit float vector"); 1070 1071 // Constant folding: Convert to generic half to single conversion. 1072 if (isa<ConstantAggregateZero>(Arg)) 1073 return ReplaceInstUsesWith(*II, ConstantAggregateZero::get(RetType)); 1074 1075 if (isa<ConstantDataVector>(Arg)) { 1076 auto VectorHalfAsShorts = Arg; 1077 if (RetWidth < ArgWidth) { 1078 SmallVector<int, 8> SubVecMask; 1079 for (unsigned i = 0; i != RetWidth; ++i) 1080 SubVecMask.push_back((int)i); 1081 VectorHalfAsShorts = Builder->CreateShuffleVector( 1082 Arg, UndefValue::get(ArgType), SubVecMask); 1083 } 1084 1085 auto VectorHalfType = 1086 VectorType::get(Type::getHalfTy(II->getContext()), RetWidth); 1087 auto VectorHalfs = 1088 Builder->CreateBitCast(VectorHalfAsShorts, VectorHalfType); 1089 auto VectorFloats = Builder->CreateFPExt(VectorHalfs, RetType); 1090 return ReplaceInstUsesWith(*II, VectorFloats); 1091 } 1092 1093 // We only use the lowest lanes of the argument. 1094 if (Value *V = SimplifyDemandedVectorEltsLow(Arg, ArgWidth, RetWidth)) { 1095 II->setArgOperand(0, V); 1096 return II; 1097 } 1098 break; 1099 } 1100 1101 case Intrinsic::x86_sse_cvtss2si: 1102 case Intrinsic::x86_sse_cvtss2si64: 1103 case Intrinsic::x86_sse_cvttss2si: 1104 case Intrinsic::x86_sse_cvttss2si64: 1105 case Intrinsic::x86_sse2_cvtsd2si: 1106 case Intrinsic::x86_sse2_cvtsd2si64: 1107 case Intrinsic::x86_sse2_cvttsd2si: 1108 case Intrinsic::x86_sse2_cvttsd2si64: { 1109 // These intrinsics only demand the 0th element of their input vectors. If 1110 // we can simplify the input based on that, do so now. 1111 Value *Arg = II->getArgOperand(0); 1112 unsigned VWidth = Arg->getType()->getVectorNumElements(); 1113 if (Value *V = SimplifyDemandedVectorEltsLow(Arg, VWidth, 1)) { 1114 II->setArgOperand(0, V); 1115 return II; 1116 } 1117 break; 1118 } 1119 1120 // Constant fold ashr( <A x Bi>, Ci ). 1121 // Constant fold lshr( <A x Bi>, Ci ). 1122 // Constant fold shl( <A x Bi>, Ci ). 1123 case Intrinsic::x86_sse2_psrai_d: 1124 case Intrinsic::x86_sse2_psrai_w: 1125 case Intrinsic::x86_avx2_psrai_d: 1126 case Intrinsic::x86_avx2_psrai_w: 1127 case Intrinsic::x86_sse2_psrli_d: 1128 case Intrinsic::x86_sse2_psrli_q: 1129 case Intrinsic::x86_sse2_psrli_w: 1130 case Intrinsic::x86_avx2_psrli_d: 1131 case Intrinsic::x86_avx2_psrli_q: 1132 case Intrinsic::x86_avx2_psrli_w: 1133 case Intrinsic::x86_sse2_pslli_d: 1134 case Intrinsic::x86_sse2_pslli_q: 1135 case Intrinsic::x86_sse2_pslli_w: 1136 case Intrinsic::x86_avx2_pslli_d: 1137 case Intrinsic::x86_avx2_pslli_q: 1138 case Intrinsic::x86_avx2_pslli_w: 1139 if (Value *V = SimplifyX86immshift(*II, *Builder)) 1140 return ReplaceInstUsesWith(*II, V); 1141 break; 1142 1143 case Intrinsic::x86_sse2_psra_d: 1144 case Intrinsic::x86_sse2_psra_w: 1145 case Intrinsic::x86_avx2_psra_d: 1146 case Intrinsic::x86_avx2_psra_w: 1147 case Intrinsic::x86_sse2_psrl_d: 1148 case Intrinsic::x86_sse2_psrl_q: 1149 case Intrinsic::x86_sse2_psrl_w: 1150 case Intrinsic::x86_avx2_psrl_d: 1151 case Intrinsic::x86_avx2_psrl_q: 1152 case Intrinsic::x86_avx2_psrl_w: 1153 case Intrinsic::x86_sse2_psll_d: 1154 case Intrinsic::x86_sse2_psll_q: 1155 case Intrinsic::x86_sse2_psll_w: 1156 case Intrinsic::x86_avx2_psll_d: 1157 case Intrinsic::x86_avx2_psll_q: 1158 case Intrinsic::x86_avx2_psll_w: { 1159 if (Value *V = SimplifyX86immshift(*II, *Builder)) 1160 return ReplaceInstUsesWith(*II, V); 1161 1162 // SSE2/AVX2 uses only the first 64-bits of the 128-bit vector 1163 // operand to compute the shift amount. 1164 Value *Arg1 = II->getArgOperand(1); 1165 assert(Arg1->getType()->getPrimitiveSizeInBits() == 128 && 1166 "Unexpected packed shift size"); 1167 unsigned VWidth = Arg1->getType()->getVectorNumElements(); 1168 1169 if (Value *V = SimplifyDemandedVectorEltsLow(Arg1, VWidth, VWidth / 2)) { 1170 II->setArgOperand(1, V); 1171 return II; 1172 } 1173 break; 1174 } 1175 1176 case Intrinsic::x86_avx2_pmovsxbd: 1177 case Intrinsic::x86_avx2_pmovsxbq: 1178 case Intrinsic::x86_avx2_pmovsxbw: 1179 case Intrinsic::x86_avx2_pmovsxdq: 1180 case Intrinsic::x86_avx2_pmovsxwd: 1181 case Intrinsic::x86_avx2_pmovsxwq: 1182 if (Value *V = SimplifyX86extend(*II, *Builder, true)) 1183 return ReplaceInstUsesWith(*II, V); 1184 break; 1185 1186 case Intrinsic::x86_sse41_pmovzxbd: 1187 case Intrinsic::x86_sse41_pmovzxbq: 1188 case Intrinsic::x86_sse41_pmovzxbw: 1189 case Intrinsic::x86_sse41_pmovzxdq: 1190 case Intrinsic::x86_sse41_pmovzxwd: 1191 case Intrinsic::x86_sse41_pmovzxwq: 1192 case Intrinsic::x86_avx2_pmovzxbd: 1193 case Intrinsic::x86_avx2_pmovzxbq: 1194 case Intrinsic::x86_avx2_pmovzxbw: 1195 case Intrinsic::x86_avx2_pmovzxdq: 1196 case Intrinsic::x86_avx2_pmovzxwd: 1197 case Intrinsic::x86_avx2_pmovzxwq: 1198 if (Value *V = SimplifyX86extend(*II, *Builder, false)) 1199 return ReplaceInstUsesWith(*II, V); 1200 break; 1201 1202 case Intrinsic::x86_sse41_insertps: 1203 if (Value *V = SimplifyX86insertps(*II, *Builder)) 1204 return ReplaceInstUsesWith(*II, V); 1205 break; 1206 1207 case Intrinsic::x86_sse4a_extrq: { 1208 Value *Op0 = II->getArgOperand(0); 1209 Value *Op1 = II->getArgOperand(1); 1210 unsigned VWidth0 = Op0->getType()->getVectorNumElements(); 1211 unsigned VWidth1 = Op1->getType()->getVectorNumElements(); 1212 assert(Op0->getType()->getPrimitiveSizeInBits() == 128 && 1213 Op1->getType()->getPrimitiveSizeInBits() == 128 && VWidth0 == 2 && 1214 VWidth1 == 16 && "Unexpected operand sizes"); 1215 1216 // See if we're dealing with constant values. 1217 Constant *C1 = dyn_cast<Constant>(Op1); 1218 ConstantInt *CILength = 1219 C1 ? dyn_cast<ConstantInt>(C1->getAggregateElement((unsigned)0)) 1220 : nullptr; 1221 ConstantInt *CIIndex = 1222 C1 ? dyn_cast<ConstantInt>(C1->getAggregateElement((unsigned)1)) 1223 : nullptr; 1224 1225 // Attempt to simplify to a constant, shuffle vector or EXTRQI call. 1226 if (Value *V = SimplifyX86extrq(*II, Op0, CILength, CIIndex, *Builder)) 1227 return ReplaceInstUsesWith(*II, V); 1228 1229 // EXTRQ only uses the lowest 64-bits of the first 128-bit vector 1230 // operands and the lowest 16-bits of the second. 1231 if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) { 1232 II->setArgOperand(0, V); 1233 return II; 1234 } 1235 if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 2)) { 1236 II->setArgOperand(1, V); 1237 return II; 1238 } 1239 break; 1240 } 1241 1242 case Intrinsic::x86_sse4a_extrqi: { 1243 // EXTRQI: Extract Length bits starting from Index. Zero pad the remaining 1244 // bits of the lower 64-bits. The upper 64-bits are undefined. 1245 Value *Op0 = II->getArgOperand(0); 1246 unsigned VWidth = Op0->getType()->getVectorNumElements(); 1247 assert(Op0->getType()->getPrimitiveSizeInBits() == 128 && VWidth == 2 && 1248 "Unexpected operand size"); 1249 1250 // See if we're dealing with constant values. 1251 ConstantInt *CILength = dyn_cast<ConstantInt>(II->getArgOperand(1)); 1252 ConstantInt *CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(2)); 1253 1254 // Attempt to simplify to a constant or shuffle vector. 1255 if (Value *V = SimplifyX86extrq(*II, Op0, CILength, CIIndex, *Builder)) 1256 return ReplaceInstUsesWith(*II, V); 1257 1258 // EXTRQI only uses the lowest 64-bits of the first 128-bit vector 1259 // operand. 1260 if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth, 1)) { 1261 II->setArgOperand(0, V); 1262 return II; 1263 } 1264 break; 1265 } 1266 1267 case Intrinsic::x86_sse4a_insertq: { 1268 Value *Op0 = II->getArgOperand(0); 1269 Value *Op1 = II->getArgOperand(1); 1270 unsigned VWidth = Op0->getType()->getVectorNumElements(); 1271 assert(Op0->getType()->getPrimitiveSizeInBits() == 128 && 1272 Op1->getType()->getPrimitiveSizeInBits() == 128 && VWidth == 2 && 1273 Op1->getType()->getVectorNumElements() == 2 && 1274 "Unexpected operand size"); 1275 1276 // See if we're dealing with constant values. 1277 Constant *C1 = dyn_cast<Constant>(Op1); 1278 ConstantInt *CI11 = 1279 C1 ? dyn_cast<ConstantInt>(C1->getAggregateElement((unsigned)1)) 1280 : nullptr; 1281 1282 // Attempt to simplify to a constant, shuffle vector or INSERTQI call. 1283 if (CI11) { 1284 APInt V11 = CI11->getValue(); 1285 APInt Len = V11.zextOrTrunc(6); 1286 APInt Idx = V11.lshr(8).zextOrTrunc(6); 1287 if (Value *V = SimplifyX86insertq(*II, Op0, Op1, Len, Idx, *Builder)) 1288 return ReplaceInstUsesWith(*II, V); 1289 } 1290 1291 // INSERTQ only uses the lowest 64-bits of the first 128-bit vector 1292 // operand. 1293 if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth, 1)) { 1294 II->setArgOperand(0, V); 1295 return II; 1296 } 1297 break; 1298 } 1299 1300 case Intrinsic::x86_sse4a_insertqi: { 1301 // INSERTQI: Extract lowest Length bits from lower half of second source and 1302 // insert over first source starting at Index bit. The upper 64-bits are 1303 // undefined. 1304 Value *Op0 = II->getArgOperand(0); 1305 Value *Op1 = II->getArgOperand(1); 1306 unsigned VWidth0 = Op0->getType()->getVectorNumElements(); 1307 unsigned VWidth1 = Op1->getType()->getVectorNumElements(); 1308 assert(Op0->getType()->getPrimitiveSizeInBits() == 128 && 1309 Op1->getType()->getPrimitiveSizeInBits() == 128 && VWidth0 == 2 && 1310 VWidth1 == 2 && "Unexpected operand sizes"); 1311 1312 // See if we're dealing with constant values. 1313 ConstantInt *CILength = dyn_cast<ConstantInt>(II->getArgOperand(2)); 1314 ConstantInt *CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(3)); 1315 1316 // Attempt to simplify to a constant or shuffle vector. 1317 if (CILength && CIIndex) { 1318 APInt Len = CILength->getValue().zextOrTrunc(6); 1319 APInt Idx = CIIndex->getValue().zextOrTrunc(6); 1320 if (Value *V = SimplifyX86insertq(*II, Op0, Op1, Len, Idx, *Builder)) 1321 return ReplaceInstUsesWith(*II, V); 1322 } 1323 1324 // INSERTQI only uses the lowest 64-bits of the first two 128-bit vector 1325 // operands. 1326 if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) { 1327 II->setArgOperand(0, V); 1328 return II; 1329 } 1330 1331 if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 1)) { 1332 II->setArgOperand(1, V); 1333 return II; 1334 } 1335 break; 1336 } 1337 1338 case Intrinsic::x86_sse41_pblendvb: 1339 case Intrinsic::x86_sse41_blendvps: 1340 case Intrinsic::x86_sse41_blendvpd: 1341 case Intrinsic::x86_avx_blendv_ps_256: 1342 case Intrinsic::x86_avx_blendv_pd_256: 1343 case Intrinsic::x86_avx2_pblendvb: { 1344 // Convert blendv* to vector selects if the mask is constant. 1345 // This optimization is convoluted because the intrinsic is defined as 1346 // getting a vector of floats or doubles for the ps and pd versions. 1347 // FIXME: That should be changed. 1348 1349 Value *Op0 = II->getArgOperand(0); 1350 Value *Op1 = II->getArgOperand(1); 1351 Value *Mask = II->getArgOperand(2); 1352 1353 // fold (blend A, A, Mask) -> A 1354 if (Op0 == Op1) 1355 return ReplaceInstUsesWith(CI, Op0); 1356 1357 // Zero Mask - select 1st argument. 1358 if (isa<ConstantAggregateZero>(Mask)) 1359 return ReplaceInstUsesWith(CI, Op0); 1360 1361 // Constant Mask - select 1st/2nd argument lane based on top bit of mask. 1362 if (auto C = dyn_cast<ConstantDataVector>(Mask)) { 1363 auto Tyi1 = Builder->getInt1Ty(); 1364 auto SelectorType = cast<VectorType>(Mask->getType()); 1365 auto EltTy = SelectorType->getElementType(); 1366 unsigned Size = SelectorType->getNumElements(); 1367 unsigned BitWidth = 1368 EltTy->isFloatTy() 1369 ? 32 1370 : (EltTy->isDoubleTy() ? 64 : EltTy->getIntegerBitWidth()); 1371 assert((BitWidth == 64 || BitWidth == 32 || BitWidth == 8) && 1372 "Wrong arguments for variable blend intrinsic"); 1373 SmallVector<Constant *, 32> Selectors; 1374 for (unsigned I = 0; I < Size; ++I) { 1375 // The intrinsics only read the top bit 1376 uint64_t Selector; 1377 if (BitWidth == 8) 1378 Selector = C->getElementAsInteger(I); 1379 else 1380 Selector = C->getElementAsAPFloat(I).bitcastToAPInt().getZExtValue(); 1381 Selectors.push_back(ConstantInt::get(Tyi1, Selector >> (BitWidth - 1))); 1382 } 1383 auto NewSelector = ConstantVector::get(Selectors); 1384 return SelectInst::Create(NewSelector, Op1, Op0, "blendv"); 1385 } 1386 break; 1387 } 1388 1389 case Intrinsic::x86_ssse3_pshuf_b_128: 1390 case Intrinsic::x86_avx2_pshuf_b: { 1391 // Turn pshufb(V1,mask) -> shuffle(V1,Zero,mask) if mask is a constant. 1392 auto *V = II->getArgOperand(1); 1393 auto *VTy = cast<VectorType>(V->getType()); 1394 unsigned NumElts = VTy->getNumElements(); 1395 assert((NumElts == 16 || NumElts == 32) && 1396 "Unexpected number of elements in shuffle mask!"); 1397 // Initialize the resulting shuffle mask to all zeroes. 1398 uint32_t Indexes[32] = {0}; 1399 1400 if (auto *Mask = dyn_cast<ConstantDataVector>(V)) { 1401 // Each byte in the shuffle control mask forms an index to permute the 1402 // corresponding byte in the destination operand. 1403 for (unsigned I = 0; I < NumElts; ++I) { 1404 int8_t Index = Mask->getElementAsInteger(I); 1405 // If the most significant bit (bit[7]) of each byte of the shuffle 1406 // control mask is set, then zero is written in the result byte. 1407 // The zero vector is in the right-hand side of the resulting 1408 // shufflevector. 1409 1410 // The value of each index is the least significant 4 bits of the 1411 // shuffle control byte. 1412 Indexes[I] = (Index < 0) ? NumElts : Index & 0xF; 1413 } 1414 } else if (!isa<ConstantAggregateZero>(V)) 1415 break; 1416 1417 // The value of each index for the high 128-bit lane is the least 1418 // significant 4 bits of the respective shuffle control byte. 1419 for (unsigned I = 16; I < NumElts; ++I) 1420 Indexes[I] += I & 0xF0; 1421 1422 auto NewC = ConstantDataVector::get(V->getContext(), 1423 makeArrayRef(Indexes, NumElts)); 1424 auto V1 = II->getArgOperand(0); 1425 auto V2 = Constant::getNullValue(II->getType()); 1426 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC); 1427 return ReplaceInstUsesWith(CI, Shuffle); 1428 } 1429 1430 case Intrinsic::x86_avx_vpermilvar_ps: 1431 case Intrinsic::x86_avx_vpermilvar_ps_256: 1432 case Intrinsic::x86_avx_vpermilvar_pd: 1433 case Intrinsic::x86_avx_vpermilvar_pd_256: { 1434 // Convert vpermil* to shufflevector if the mask is constant. 1435 Value *V = II->getArgOperand(1); 1436 unsigned Size = cast<VectorType>(V->getType())->getNumElements(); 1437 assert(Size == 8 || Size == 4 || Size == 2); 1438 uint32_t Indexes[8]; 1439 if (auto C = dyn_cast<ConstantDataVector>(V)) { 1440 // The intrinsics only read one or two bits, clear the rest. 1441 for (unsigned I = 0; I < Size; ++I) { 1442 uint32_t Index = C->getElementAsInteger(I) & 0x3; 1443 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd || 1444 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) 1445 Index >>= 1; 1446 Indexes[I] = Index; 1447 } 1448 } else if (isa<ConstantAggregateZero>(V)) { 1449 for (unsigned I = 0; I < Size; ++I) 1450 Indexes[I] = 0; 1451 } else { 1452 break; 1453 } 1454 // The _256 variants are a bit trickier since the mask bits always index 1455 // into the corresponding 128 half. In order to convert to a generic 1456 // shuffle, we have to make that explicit. 1457 if (II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_ps_256 || 1458 II->getIntrinsicID() == Intrinsic::x86_avx_vpermilvar_pd_256) { 1459 for (unsigned I = Size / 2; I < Size; ++I) 1460 Indexes[I] += Size / 2; 1461 } 1462 auto NewC = 1463 ConstantDataVector::get(V->getContext(), makeArrayRef(Indexes, Size)); 1464 auto V1 = II->getArgOperand(0); 1465 auto V2 = UndefValue::get(V1->getType()); 1466 auto Shuffle = Builder->CreateShuffleVector(V1, V2, NewC); 1467 return ReplaceInstUsesWith(CI, Shuffle); 1468 } 1469 1470 case Intrinsic::x86_avx_vperm2f128_pd_256: 1471 case Intrinsic::x86_avx_vperm2f128_ps_256: 1472 case Intrinsic::x86_avx_vperm2f128_si_256: 1473 case Intrinsic::x86_avx2_vperm2i128: 1474 if (Value *V = SimplifyX86vperm2(*II, *Builder)) 1475 return ReplaceInstUsesWith(*II, V); 1476 break; 1477 1478 case Intrinsic::x86_xop_vpcomb: 1479 case Intrinsic::x86_xop_vpcomd: 1480 case Intrinsic::x86_xop_vpcomq: 1481 case Intrinsic::x86_xop_vpcomw: 1482 if (Value *V = SimplifyX86vpcom(*II, *Builder, true)) 1483 return ReplaceInstUsesWith(*II, V); 1484 break; 1485 1486 case Intrinsic::x86_xop_vpcomub: 1487 case Intrinsic::x86_xop_vpcomud: 1488 case Intrinsic::x86_xop_vpcomuq: 1489 case Intrinsic::x86_xop_vpcomuw: 1490 if (Value *V = SimplifyX86vpcom(*II, *Builder, false)) 1491 return ReplaceInstUsesWith(*II, V); 1492 break; 1493 1494 case Intrinsic::ppc_altivec_vperm: 1495 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant. 1496 // Note that ppc_altivec_vperm has a big-endian bias, so when creating 1497 // a vectorshuffle for little endian, we must undo the transformation 1498 // performed on vec_perm in altivec.h. That is, we must complement 1499 // the permutation mask with respect to 31 and reverse the order of 1500 // V1 and V2. 1501 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) { 1502 assert(Mask->getType()->getVectorNumElements() == 16 && 1503 "Bad type for intrinsic!"); 1504 1505 // Check that all of the elements are integer constants or undefs. 1506 bool AllEltsOk = true; 1507 for (unsigned i = 0; i != 16; ++i) { 1508 Constant *Elt = Mask->getAggregateElement(i); 1509 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) { 1510 AllEltsOk = false; 1511 break; 1512 } 1513 } 1514 1515 if (AllEltsOk) { 1516 // Cast the input vectors to byte vectors. 1517 Value *Op0 = Builder->CreateBitCast(II->getArgOperand(0), 1518 Mask->getType()); 1519 Value *Op1 = Builder->CreateBitCast(II->getArgOperand(1), 1520 Mask->getType()); 1521 Value *Result = UndefValue::get(Op0->getType()); 1522 1523 // Only extract each element once. 1524 Value *ExtractedElts[32]; 1525 memset(ExtractedElts, 0, sizeof(ExtractedElts)); 1526 1527 for (unsigned i = 0; i != 16; ++i) { 1528 if (isa<UndefValue>(Mask->getAggregateElement(i))) 1529 continue; 1530 unsigned Idx = 1531 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue(); 1532 Idx &= 31; // Match the hardware behavior. 1533 if (DL.isLittleEndian()) 1534 Idx = 31 - Idx; 1535 1536 if (!ExtractedElts[Idx]) { 1537 Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0; 1538 Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1; 1539 ExtractedElts[Idx] = 1540 Builder->CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse, 1541 Builder->getInt32(Idx&15)); 1542 } 1543 1544 // Insert this value into the result vector. 1545 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx], 1546 Builder->getInt32(i)); 1547 } 1548 return CastInst::Create(Instruction::BitCast, Result, CI.getType()); 1549 } 1550 } 1551 break; 1552 1553 case Intrinsic::arm_neon_vld1: 1554 case Intrinsic::arm_neon_vld2: 1555 case Intrinsic::arm_neon_vld3: 1556 case Intrinsic::arm_neon_vld4: 1557 case Intrinsic::arm_neon_vld2lane: 1558 case Intrinsic::arm_neon_vld3lane: 1559 case Intrinsic::arm_neon_vld4lane: 1560 case Intrinsic::arm_neon_vst1: 1561 case Intrinsic::arm_neon_vst2: 1562 case Intrinsic::arm_neon_vst3: 1563 case Intrinsic::arm_neon_vst4: 1564 case Intrinsic::arm_neon_vst2lane: 1565 case Intrinsic::arm_neon_vst3lane: 1566 case Intrinsic::arm_neon_vst4lane: { 1567 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), DL, II, AC, DT); 1568 unsigned AlignArg = II->getNumArgOperands() - 1; 1569 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg)); 1570 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) { 1571 II->setArgOperand(AlignArg, 1572 ConstantInt::get(Type::getInt32Ty(II->getContext()), 1573 MemAlign, false)); 1574 return II; 1575 } 1576 break; 1577 } 1578 1579 case Intrinsic::arm_neon_vmulls: 1580 case Intrinsic::arm_neon_vmullu: 1581 case Intrinsic::aarch64_neon_smull: 1582 case Intrinsic::aarch64_neon_umull: { 1583 Value *Arg0 = II->getArgOperand(0); 1584 Value *Arg1 = II->getArgOperand(1); 1585 1586 // Handle mul by zero first: 1587 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) { 1588 return ReplaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType())); 1589 } 1590 1591 // Check for constant LHS & RHS - in this case we just simplify. 1592 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu || 1593 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull); 1594 VectorType *NewVT = cast<VectorType>(II->getType()); 1595 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) { 1596 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) { 1597 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext); 1598 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext); 1599 1600 return ReplaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1)); 1601 } 1602 1603 // Couldn't simplify - canonicalize constant to the RHS. 1604 std::swap(Arg0, Arg1); 1605 } 1606 1607 // Handle mul by one: 1608 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) 1609 if (ConstantInt *Splat = 1610 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue())) 1611 if (Splat->isOne()) 1612 return CastInst::CreateIntegerCast(Arg0, II->getType(), 1613 /*isSigned=*/!Zext); 1614 1615 break; 1616 } 1617 1618 case Intrinsic::AMDGPU_rcp: { 1619 if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) { 1620 const APFloat &ArgVal = C->getValueAPF(); 1621 APFloat Val(ArgVal.getSemantics(), 1.0); 1622 APFloat::opStatus Status = Val.divide(ArgVal, 1623 APFloat::rmNearestTiesToEven); 1624 // Only do this if it was exact and therefore not dependent on the 1625 // rounding mode. 1626 if (Status == APFloat::opOK) 1627 return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val)); 1628 } 1629 1630 break; 1631 } 1632 case Intrinsic::stackrestore: { 1633 // If the save is right next to the restore, remove the restore. This can 1634 // happen when variable allocas are DCE'd. 1635 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) { 1636 if (SS->getIntrinsicID() == Intrinsic::stacksave) { 1637 if (&*++SS->getIterator() == II) 1638 return EraseInstFromFunction(CI); 1639 } 1640 } 1641 1642 // Scan down this block to see if there is another stack restore in the 1643 // same block without an intervening call/alloca. 1644 BasicBlock::iterator BI(II); 1645 TerminatorInst *TI = II->getParent()->getTerminator(); 1646 bool CannotRemove = false; 1647 for (++BI; &*BI != TI; ++BI) { 1648 if (isa<AllocaInst>(BI)) { 1649 CannotRemove = true; 1650 break; 1651 } 1652 if (CallInst *BCI = dyn_cast<CallInst>(BI)) { 1653 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) { 1654 // If there is a stackrestore below this one, remove this one. 1655 if (II->getIntrinsicID() == Intrinsic::stackrestore) 1656 return EraseInstFromFunction(CI); 1657 // Otherwise, ignore the intrinsic. 1658 } else { 1659 // If we found a non-intrinsic call, we can't remove the stack 1660 // restore. 1661 CannotRemove = true; 1662 break; 1663 } 1664 } 1665 } 1666 1667 // If the stack restore is in a return, resume, or unwind block and if there 1668 // are no allocas or calls between the restore and the return, nuke the 1669 // restore. 1670 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI))) 1671 return EraseInstFromFunction(CI); 1672 break; 1673 } 1674 case Intrinsic::lifetime_start: { 1675 // Remove trivially empty lifetime_start/end ranges, i.e. a start 1676 // immediately followed by an end (ignoring debuginfo or other 1677 // lifetime markers in between). 1678 BasicBlock::iterator BI = II->getIterator(), BE = II->getParent()->end(); 1679 for (++BI; BI != BE; ++BI) { 1680 if (IntrinsicInst *LTE = dyn_cast<IntrinsicInst>(BI)) { 1681 if (isa<DbgInfoIntrinsic>(LTE) || 1682 LTE->getIntrinsicID() == Intrinsic::lifetime_start) 1683 continue; 1684 if (LTE->getIntrinsicID() == Intrinsic::lifetime_end) { 1685 if (II->getOperand(0) == LTE->getOperand(0) && 1686 II->getOperand(1) == LTE->getOperand(1)) { 1687 EraseInstFromFunction(*LTE); 1688 return EraseInstFromFunction(*II); 1689 } 1690 continue; 1691 } 1692 } 1693 break; 1694 } 1695 break; 1696 } 1697 case Intrinsic::assume: { 1698 // Canonicalize assume(a && b) -> assume(a); assume(b); 1699 // Note: New assumption intrinsics created here are registered by 1700 // the InstCombineIRInserter object. 1701 Value *IIOperand = II->getArgOperand(0), *A, *B, 1702 *AssumeIntrinsic = II->getCalledValue(); 1703 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) { 1704 Builder->CreateCall(AssumeIntrinsic, A, II->getName()); 1705 Builder->CreateCall(AssumeIntrinsic, B, II->getName()); 1706 return EraseInstFromFunction(*II); 1707 } 1708 // assume(!(a || b)) -> assume(!a); assume(!b); 1709 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) { 1710 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(A), 1711 II->getName()); 1712 Builder->CreateCall(AssumeIntrinsic, Builder->CreateNot(B), 1713 II->getName()); 1714 return EraseInstFromFunction(*II); 1715 } 1716 1717 // assume( (load addr) != null ) -> add 'nonnull' metadata to load 1718 // (if assume is valid at the load) 1719 if (ICmpInst* ICmp = dyn_cast<ICmpInst>(IIOperand)) { 1720 Value *LHS = ICmp->getOperand(0); 1721 Value *RHS = ICmp->getOperand(1); 1722 if (ICmpInst::ICMP_NE == ICmp->getPredicate() && 1723 isa<LoadInst>(LHS) && 1724 isa<Constant>(RHS) && 1725 RHS->getType()->isPointerTy() && 1726 cast<Constant>(RHS)->isNullValue()) { 1727 LoadInst* LI = cast<LoadInst>(LHS); 1728 if (isValidAssumeForContext(II, LI, DT)) { 1729 MDNode *MD = MDNode::get(II->getContext(), None); 1730 LI->setMetadata(LLVMContext::MD_nonnull, MD); 1731 return EraseInstFromFunction(*II); 1732 } 1733 } 1734 // TODO: apply nonnull return attributes to calls and invokes 1735 // TODO: apply range metadata for range check patterns? 1736 } 1737 // If there is a dominating assume with the same condition as this one, 1738 // then this one is redundant, and should be removed. 1739 APInt KnownZero(1, 0), KnownOne(1, 0); 1740 computeKnownBits(IIOperand, KnownZero, KnownOne, 0, II); 1741 if (KnownOne.isAllOnesValue()) 1742 return EraseInstFromFunction(*II); 1743 1744 break; 1745 } 1746 case Intrinsic::experimental_gc_relocate: { 1747 // Translate facts known about a pointer before relocating into 1748 // facts about the relocate value, while being careful to 1749 // preserve relocation semantics. 1750 GCRelocateOperands Operands(II); 1751 Value *DerivedPtr = Operands.getDerivedPtr(); 1752 auto *GCRelocateType = cast<PointerType>(II->getType()); 1753 1754 // Remove the relocation if unused, note that this check is required 1755 // to prevent the cases below from looping forever. 1756 if (II->use_empty()) 1757 return EraseInstFromFunction(*II); 1758 1759 // Undef is undef, even after relocation. 1760 // TODO: provide a hook for this in GCStrategy. This is clearly legal for 1761 // most practical collectors, but there was discussion in the review thread 1762 // about whether it was legal for all possible collectors. 1763 if (isa<UndefValue>(DerivedPtr)) { 1764 // gc_relocate is uncasted. Use undef of gc_relocate's type to replace it. 1765 return ReplaceInstUsesWith(*II, UndefValue::get(GCRelocateType)); 1766 } 1767 1768 // The relocation of null will be null for most any collector. 1769 // TODO: provide a hook for this in GCStrategy. There might be some weird 1770 // collector this property does not hold for. 1771 if (isa<ConstantPointerNull>(DerivedPtr)) { 1772 // gc_relocate is uncasted. Use null-pointer of gc_relocate's type to replace it. 1773 return ReplaceInstUsesWith(*II, ConstantPointerNull::get(GCRelocateType)); 1774 } 1775 1776 // isKnownNonNull -> nonnull attribute 1777 if (isKnownNonNullAt(DerivedPtr, II, DT, TLI)) 1778 II->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull); 1779 1780 // isDereferenceablePointer -> deref attribute 1781 if (isDereferenceablePointer(DerivedPtr, DL)) { 1782 if (Argument *A = dyn_cast<Argument>(DerivedPtr)) { 1783 uint64_t Bytes = A->getDereferenceableBytes(); 1784 II->addDereferenceableAttr(AttributeSet::ReturnIndex, Bytes); 1785 } 1786 } 1787 1788 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p)) 1789 // Canonicalize on the type from the uses to the defs 1790 1791 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...) 1792 } 1793 } 1794 1795 return visitCallSite(II); 1796 } 1797 1798 // InvokeInst simplification 1799 // 1800 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) { 1801 return visitCallSite(&II); 1802 } 1803 1804 /// isSafeToEliminateVarargsCast - If this cast does not affect the value 1805 /// passed through the varargs area, we can eliminate the use of the cast. 1806 static bool isSafeToEliminateVarargsCast(const CallSite CS, 1807 const DataLayout &DL, 1808 const CastInst *const CI, 1809 const int ix) { 1810 if (!CI->isLosslessCast()) 1811 return false; 1812 1813 // If this is a GC intrinsic, avoid munging types. We need types for 1814 // statepoint reconstruction in SelectionDAG. 1815 // TODO: This is probably something which should be expanded to all 1816 // intrinsics since the entire point of intrinsics is that 1817 // they are understandable by the optimizer. 1818 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS)) 1819 return false; 1820 1821 // The size of ByVal or InAlloca arguments is derived from the type, so we 1822 // can't change to a type with a different size. If the size were 1823 // passed explicitly we could avoid this check. 1824 if (!CS.isByValOrInAllocaArgument(ix)) 1825 return true; 1826 1827 Type* SrcTy = 1828 cast<PointerType>(CI->getOperand(0)->getType())->getElementType(); 1829 Type* DstTy = cast<PointerType>(CI->getType())->getElementType(); 1830 if (!SrcTy->isSized() || !DstTy->isSized()) 1831 return false; 1832 if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy)) 1833 return false; 1834 return true; 1835 } 1836 1837 // Try to fold some different type of calls here. 1838 // Currently we're only working with the checking functions, memcpy_chk, 1839 // mempcpy_chk, memmove_chk, memset_chk, strcpy_chk, stpcpy_chk, strncpy_chk, 1840 // strcat_chk and strncat_chk. 1841 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) { 1842 if (!CI->getCalledFunction()) return nullptr; 1843 1844 auto InstCombineRAUW = [this](Instruction *From, Value *With) { 1845 ReplaceInstUsesWith(*From, With); 1846 }; 1847 LibCallSimplifier Simplifier(DL, TLI, InstCombineRAUW); 1848 if (Value *With = Simplifier.optimizeCall(CI)) { 1849 ++NumSimplified; 1850 return CI->use_empty() ? CI : ReplaceInstUsesWith(*CI, With); 1851 } 1852 1853 return nullptr; 1854 } 1855 1856 static IntrinsicInst *FindInitTrampolineFromAlloca(Value *TrampMem) { 1857 // Strip off at most one level of pointer casts, looking for an alloca. This 1858 // is good enough in practice and simpler than handling any number of casts. 1859 Value *Underlying = TrampMem->stripPointerCasts(); 1860 if (Underlying != TrampMem && 1861 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem)) 1862 return nullptr; 1863 if (!isa<AllocaInst>(Underlying)) 1864 return nullptr; 1865 1866 IntrinsicInst *InitTrampoline = nullptr; 1867 for (User *U : TrampMem->users()) { 1868 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U); 1869 if (!II) 1870 return nullptr; 1871 if (II->getIntrinsicID() == Intrinsic::init_trampoline) { 1872 if (InitTrampoline) 1873 // More than one init_trampoline writes to this value. Give up. 1874 return nullptr; 1875 InitTrampoline = II; 1876 continue; 1877 } 1878 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline) 1879 // Allow any number of calls to adjust.trampoline. 1880 continue; 1881 return nullptr; 1882 } 1883 1884 // No call to init.trampoline found. 1885 if (!InitTrampoline) 1886 return nullptr; 1887 1888 // Check that the alloca is being used in the expected way. 1889 if (InitTrampoline->getOperand(0) != TrampMem) 1890 return nullptr; 1891 1892 return InitTrampoline; 1893 } 1894 1895 static IntrinsicInst *FindInitTrampolineFromBB(IntrinsicInst *AdjustTramp, 1896 Value *TrampMem) { 1897 // Visit all the previous instructions in the basic block, and try to find a 1898 // init.trampoline which has a direct path to the adjust.trampoline. 1899 for (BasicBlock::iterator I = AdjustTramp->getIterator(), 1900 E = AdjustTramp->getParent()->begin(); 1901 I != E;) { 1902 Instruction *Inst = &*--I; 1903 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) 1904 if (II->getIntrinsicID() == Intrinsic::init_trampoline && 1905 II->getOperand(0) == TrampMem) 1906 return II; 1907 if (Inst->mayWriteToMemory()) 1908 return nullptr; 1909 } 1910 return nullptr; 1911 } 1912 1913 // Given a call to llvm.adjust.trampoline, find and return the corresponding 1914 // call to llvm.init.trampoline if the call to the trampoline can be optimized 1915 // to a direct call to a function. Otherwise return NULL. 1916 // 1917 static IntrinsicInst *FindInitTrampoline(Value *Callee) { 1918 Callee = Callee->stripPointerCasts(); 1919 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee); 1920 if (!AdjustTramp || 1921 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline) 1922 return nullptr; 1923 1924 Value *TrampMem = AdjustTramp->getOperand(0); 1925 1926 if (IntrinsicInst *IT = FindInitTrampolineFromAlloca(TrampMem)) 1927 return IT; 1928 if (IntrinsicInst *IT = FindInitTrampolineFromBB(AdjustTramp, TrampMem)) 1929 return IT; 1930 return nullptr; 1931 } 1932 1933 // visitCallSite - Improvements for call and invoke instructions. 1934 // 1935 Instruction *InstCombiner::visitCallSite(CallSite CS) { 1936 1937 if (isAllocLikeFn(CS.getInstruction(), TLI)) 1938 return visitAllocSite(*CS.getInstruction()); 1939 1940 bool Changed = false; 1941 1942 // Mark any parameters that are known to be non-null with the nonnull 1943 // attribute. This is helpful for inlining calls to functions with null 1944 // checks on their arguments. 1945 SmallVector<unsigned, 4> Indices; 1946 unsigned ArgNo = 0; 1947 1948 for (Value *V : CS.args()) { 1949 if (V->getType()->isPointerTy() && !CS.paramHasAttr(ArgNo+1, Attribute::NonNull) && 1950 isKnownNonNullAt(V, CS.getInstruction(), DT, TLI)) 1951 Indices.push_back(ArgNo + 1); 1952 ArgNo++; 1953 } 1954 1955 assert(ArgNo == CS.arg_size() && "sanity check"); 1956 1957 if (!Indices.empty()) { 1958 AttributeSet AS = CS.getAttributes(); 1959 LLVMContext &Ctx = CS.getInstruction()->getContext(); 1960 AS = AS.addAttribute(Ctx, Indices, 1961 Attribute::get(Ctx, Attribute::NonNull)); 1962 CS.setAttributes(AS); 1963 Changed = true; 1964 } 1965 1966 // If the callee is a pointer to a function, attempt to move any casts to the 1967 // arguments of the call/invoke. 1968 Value *Callee = CS.getCalledValue(); 1969 if (!isa<Function>(Callee) && transformConstExprCastCall(CS)) 1970 return nullptr; 1971 1972 if (Function *CalleeF = dyn_cast<Function>(Callee)) 1973 // If the call and callee calling conventions don't match, this call must 1974 // be unreachable, as the call is undefined. 1975 if (CalleeF->getCallingConv() != CS.getCallingConv() && 1976 // Only do this for calls to a function with a body. A prototype may 1977 // not actually end up matching the implementation's calling conv for a 1978 // variety of reasons (e.g. it may be written in assembly). 1979 !CalleeF->isDeclaration()) { 1980 Instruction *OldCall = CS.getInstruction(); 1981 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 1982 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 1983 OldCall); 1984 // If OldCall does not return void then replaceAllUsesWith undef. 1985 // This allows ValueHandlers and custom metadata to adjust itself. 1986 if (!OldCall->getType()->isVoidTy()) 1987 ReplaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType())); 1988 if (isa<CallInst>(OldCall)) 1989 return EraseInstFromFunction(*OldCall); 1990 1991 // We cannot remove an invoke, because it would change the CFG, just 1992 // change the callee to a null pointer. 1993 cast<InvokeInst>(OldCall)->setCalledFunction( 1994 Constant::getNullValue(CalleeF->getType())); 1995 return nullptr; 1996 } 1997 1998 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) { 1999 // If CS does not return void then replaceAllUsesWith undef. 2000 // This allows ValueHandlers and custom metadata to adjust itself. 2001 if (!CS.getInstruction()->getType()->isVoidTy()) 2002 ReplaceInstUsesWith(*CS.getInstruction(), 2003 UndefValue::get(CS.getInstruction()->getType())); 2004 2005 if (isa<InvokeInst>(CS.getInstruction())) { 2006 // Can't remove an invoke because we cannot change the CFG. 2007 return nullptr; 2008 } 2009 2010 // This instruction is not reachable, just remove it. We insert a store to 2011 // undef so that we know that this code is not reachable, despite the fact 2012 // that we can't modify the CFG here. 2013 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 2014 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 2015 CS.getInstruction()); 2016 2017 return EraseInstFromFunction(*CS.getInstruction()); 2018 } 2019 2020 if (IntrinsicInst *II = FindInitTrampoline(Callee)) 2021 return transformCallThroughTrampoline(CS, II); 2022 2023 PointerType *PTy = cast<PointerType>(Callee->getType()); 2024 FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 2025 if (FTy->isVarArg()) { 2026 int ix = FTy->getNumParams(); 2027 // See if we can optimize any arguments passed through the varargs area of 2028 // the call. 2029 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(), 2030 E = CS.arg_end(); I != E; ++I, ++ix) { 2031 CastInst *CI = dyn_cast<CastInst>(*I); 2032 if (CI && isSafeToEliminateVarargsCast(CS, DL, CI, ix)) { 2033 *I = CI->getOperand(0); 2034 Changed = true; 2035 } 2036 } 2037 } 2038 2039 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) { 2040 // Inline asm calls cannot throw - mark them 'nounwind'. 2041 CS.setDoesNotThrow(); 2042 Changed = true; 2043 } 2044 2045 // Try to optimize the call if possible, we require DataLayout for most of 2046 // this. None of these calls are seen as possibly dead so go ahead and 2047 // delete the instruction now. 2048 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) { 2049 Instruction *I = tryOptimizeCall(CI); 2050 // If we changed something return the result, etc. Otherwise let 2051 // the fallthrough check. 2052 if (I) return EraseInstFromFunction(*I); 2053 } 2054 2055 return Changed ? CS.getInstruction() : nullptr; 2056 } 2057 2058 // transformConstExprCastCall - If the callee is a constexpr cast of a function, 2059 // attempt to move the cast to the arguments of the call/invoke. 2060 // 2061 bool InstCombiner::transformConstExprCastCall(CallSite CS) { 2062 Function *Callee = 2063 dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts()); 2064 if (!Callee) 2065 return false; 2066 // The prototype of thunks are a lie, don't try to directly call such 2067 // functions. 2068 if (Callee->hasFnAttribute("thunk")) 2069 return false; 2070 Instruction *Caller = CS.getInstruction(); 2071 const AttributeSet &CallerPAL = CS.getAttributes(); 2072 2073 // Okay, this is a cast from a function to a different type. Unless doing so 2074 // would cause a type conversion of one of our arguments, change this call to 2075 // be a direct call with arguments casted to the appropriate types. 2076 // 2077 FunctionType *FT = Callee->getFunctionType(); 2078 Type *OldRetTy = Caller->getType(); 2079 Type *NewRetTy = FT->getReturnType(); 2080 2081 // Check to see if we are changing the return type... 2082 if (OldRetTy != NewRetTy) { 2083 2084 if (NewRetTy->isStructTy()) 2085 return false; // TODO: Handle multiple return values. 2086 2087 if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) { 2088 if (Callee->isDeclaration()) 2089 return false; // Cannot transform this return value. 2090 2091 if (!Caller->use_empty() && 2092 // void -> non-void is handled specially 2093 !NewRetTy->isVoidTy()) 2094 return false; // Cannot transform this return value. 2095 } 2096 2097 if (!CallerPAL.isEmpty() && !Caller->use_empty()) { 2098 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex); 2099 if (RAttrs.overlaps(AttributeFuncs::typeIncompatible(NewRetTy))) 2100 return false; // Attribute not compatible with transformed value. 2101 } 2102 2103 // If the callsite is an invoke instruction, and the return value is used by 2104 // a PHI node in a successor, we cannot change the return type of the call 2105 // because there is no place to put the cast instruction (without breaking 2106 // the critical edge). Bail out in this case. 2107 if (!Caller->use_empty()) 2108 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) 2109 for (User *U : II->users()) 2110 if (PHINode *PN = dyn_cast<PHINode>(U)) 2111 if (PN->getParent() == II->getNormalDest() || 2112 PN->getParent() == II->getUnwindDest()) 2113 return false; 2114 } 2115 2116 unsigned NumActualArgs = CS.arg_size(); 2117 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs); 2118 2119 // Prevent us turning: 2120 // declare void @takes_i32_inalloca(i32* inalloca) 2121 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0) 2122 // 2123 // into: 2124 // call void @takes_i32_inalloca(i32* null) 2125 // 2126 // Similarly, avoid folding away bitcasts of byval calls. 2127 if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) || 2128 Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal)) 2129 return false; 2130 2131 CallSite::arg_iterator AI = CS.arg_begin(); 2132 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) { 2133 Type *ParamTy = FT->getParamType(i); 2134 Type *ActTy = (*AI)->getType(); 2135 2136 if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL)) 2137 return false; // Cannot transform this parameter value. 2138 2139 if (AttrBuilder(CallerPAL.getParamAttributes(i + 1), i + 1). 2140 overlaps(AttributeFuncs::typeIncompatible(ParamTy))) 2141 return false; // Attribute not compatible with transformed value. 2142 2143 if (CS.isInAllocaArgument(i)) 2144 return false; // Cannot transform to and from inalloca. 2145 2146 // If the parameter is passed as a byval argument, then we have to have a 2147 // sized type and the sized type has to have the same size as the old type. 2148 if (ParamTy != ActTy && 2149 CallerPAL.getParamAttributes(i + 1).hasAttribute(i + 1, 2150 Attribute::ByVal)) { 2151 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy); 2152 if (!ParamPTy || !ParamPTy->getElementType()->isSized()) 2153 return false; 2154 2155 Type *CurElTy = ActTy->getPointerElementType(); 2156 if (DL.getTypeAllocSize(CurElTy) != 2157 DL.getTypeAllocSize(ParamPTy->getElementType())) 2158 return false; 2159 } 2160 } 2161 2162 if (Callee->isDeclaration()) { 2163 // Do not delete arguments unless we have a function body. 2164 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg()) 2165 return false; 2166 2167 // If the callee is just a declaration, don't change the varargsness of the 2168 // call. We don't want to introduce a varargs call where one doesn't 2169 // already exist. 2170 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType()); 2171 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg()) 2172 return false; 2173 2174 // If both the callee and the cast type are varargs, we still have to make 2175 // sure the number of fixed parameters are the same or we have the same 2176 // ABI issues as if we introduce a varargs call. 2177 if (FT->isVarArg() && 2178 cast<FunctionType>(APTy->getElementType())->isVarArg() && 2179 FT->getNumParams() != 2180 cast<FunctionType>(APTy->getElementType())->getNumParams()) 2181 return false; 2182 } 2183 2184 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && 2185 !CallerPAL.isEmpty()) 2186 // In this case we have more arguments than the new function type, but we 2187 // won't be dropping them. Check that these extra arguments have attributes 2188 // that are compatible with being a vararg call argument. 2189 for (unsigned i = CallerPAL.getNumSlots(); i; --i) { 2190 unsigned Index = CallerPAL.getSlotIndex(i - 1); 2191 if (Index <= FT->getNumParams()) 2192 break; 2193 2194 // Check if it has an attribute that's incompatible with varargs. 2195 AttributeSet PAttrs = CallerPAL.getSlotAttributes(i - 1); 2196 if (PAttrs.hasAttribute(Index, Attribute::StructRet)) 2197 return false; 2198 } 2199 2200 2201 // Okay, we decided that this is a safe thing to do: go ahead and start 2202 // inserting cast instructions as necessary. 2203 std::vector<Value*> Args; 2204 Args.reserve(NumActualArgs); 2205 SmallVector<AttributeSet, 8> attrVec; 2206 attrVec.reserve(NumCommonArgs); 2207 2208 // Get any return attributes. 2209 AttrBuilder RAttrs(CallerPAL, AttributeSet::ReturnIndex); 2210 2211 // If the return value is not being used, the type may not be compatible 2212 // with the existing attributes. Wipe out any problematic attributes. 2213 RAttrs.remove(AttributeFuncs::typeIncompatible(NewRetTy)); 2214 2215 // Add the new return attributes. 2216 if (RAttrs.hasAttributes()) 2217 attrVec.push_back(AttributeSet::get(Caller->getContext(), 2218 AttributeSet::ReturnIndex, RAttrs)); 2219 2220 AI = CS.arg_begin(); 2221 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) { 2222 Type *ParamTy = FT->getParamType(i); 2223 2224 if ((*AI)->getType() == ParamTy) { 2225 Args.push_back(*AI); 2226 } else { 2227 Args.push_back(Builder->CreateBitOrPointerCast(*AI, ParamTy)); 2228 } 2229 2230 // Add any parameter attributes. 2231 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1); 2232 if (PAttrs.hasAttributes()) 2233 attrVec.push_back(AttributeSet::get(Caller->getContext(), i + 1, 2234 PAttrs)); 2235 } 2236 2237 // If the function takes more arguments than the call was taking, add them 2238 // now. 2239 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) 2240 Args.push_back(Constant::getNullValue(FT->getParamType(i))); 2241 2242 // If we are removing arguments to the function, emit an obnoxious warning. 2243 if (FT->getNumParams() < NumActualArgs) { 2244 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722 2245 if (FT->isVarArg()) { 2246 // Add all of the arguments in their promoted form to the arg list. 2247 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) { 2248 Type *PTy = getPromotedType((*AI)->getType()); 2249 if (PTy != (*AI)->getType()) { 2250 // Must promote to pass through va_arg area! 2251 Instruction::CastOps opcode = 2252 CastInst::getCastOpcode(*AI, false, PTy, false); 2253 Args.push_back(Builder->CreateCast(opcode, *AI, PTy)); 2254 } else { 2255 Args.push_back(*AI); 2256 } 2257 2258 // Add any parameter attributes. 2259 AttrBuilder PAttrs(CallerPAL.getParamAttributes(i + 1), i + 1); 2260 if (PAttrs.hasAttributes()) 2261 attrVec.push_back(AttributeSet::get(FT->getContext(), i + 1, 2262 PAttrs)); 2263 } 2264 } 2265 } 2266 2267 AttributeSet FnAttrs = CallerPAL.getFnAttributes(); 2268 if (CallerPAL.hasAttributes(AttributeSet::FunctionIndex)) 2269 attrVec.push_back(AttributeSet::get(Callee->getContext(), FnAttrs)); 2270 2271 if (NewRetTy->isVoidTy()) 2272 Caller->setName(""); // Void type should not have a name. 2273 2274 const AttributeSet &NewCallerPAL = AttributeSet::get(Callee->getContext(), 2275 attrVec); 2276 2277 SmallVector<OperandBundleDef, 1> OpBundles; 2278 CS.getOperandBundlesAsDefs(OpBundles); 2279 2280 Instruction *NC; 2281 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 2282 NC = Builder->CreateInvoke(Callee, II->getNormalDest(), II->getUnwindDest(), 2283 Args, OpBundles); 2284 NC->takeName(II); 2285 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv()); 2286 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL); 2287 } else { 2288 CallInst *CI = cast<CallInst>(Caller); 2289 NC = Builder->CreateCall(Callee, Args, OpBundles); 2290 NC->takeName(CI); 2291 if (CI->isTailCall()) 2292 cast<CallInst>(NC)->setTailCall(); 2293 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv()); 2294 cast<CallInst>(NC)->setAttributes(NewCallerPAL); 2295 } 2296 2297 // Insert a cast of the return type as necessary. 2298 Value *NV = NC; 2299 if (OldRetTy != NV->getType() && !Caller->use_empty()) { 2300 if (!NV->getType()->isVoidTy()) { 2301 NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy); 2302 NC->setDebugLoc(Caller->getDebugLoc()); 2303 2304 // If this is an invoke instruction, we should insert it after the first 2305 // non-phi, instruction in the normal successor block. 2306 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 2307 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt(); 2308 InsertNewInstBefore(NC, *I); 2309 } else { 2310 // Otherwise, it's a call, just insert cast right after the call. 2311 InsertNewInstBefore(NC, *Caller); 2312 } 2313 Worklist.AddUsersToWorkList(*Caller); 2314 } else { 2315 NV = UndefValue::get(Caller->getType()); 2316 } 2317 } 2318 2319 if (!Caller->use_empty()) 2320 ReplaceInstUsesWith(*Caller, NV); 2321 else if (Caller->hasValueHandle()) { 2322 if (OldRetTy == NV->getType()) 2323 ValueHandleBase::ValueIsRAUWd(Caller, NV); 2324 else 2325 // We cannot call ValueIsRAUWd with a different type, and the 2326 // actual tracked value will disappear. 2327 ValueHandleBase::ValueIsDeleted(Caller); 2328 } 2329 2330 EraseInstFromFunction(*Caller); 2331 return true; 2332 } 2333 2334 // transformCallThroughTrampoline - Turn a call to a function created by 2335 // init_trampoline / adjust_trampoline intrinsic pair into a direct call to the 2336 // underlying function. 2337 // 2338 Instruction * 2339 InstCombiner::transformCallThroughTrampoline(CallSite CS, 2340 IntrinsicInst *Tramp) { 2341 Value *Callee = CS.getCalledValue(); 2342 PointerType *PTy = cast<PointerType>(Callee->getType()); 2343 FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 2344 const AttributeSet &Attrs = CS.getAttributes(); 2345 2346 // If the call already has the 'nest' attribute somewhere then give up - 2347 // otherwise 'nest' would occur twice after splicing in the chain. 2348 if (Attrs.hasAttrSomewhere(Attribute::Nest)) 2349 return nullptr; 2350 2351 assert(Tramp && 2352 "transformCallThroughTrampoline called with incorrect CallSite."); 2353 2354 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts()); 2355 PointerType *NestFPTy = cast<PointerType>(NestF->getType()); 2356 FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType()); 2357 2358 const AttributeSet &NestAttrs = NestF->getAttributes(); 2359 if (!NestAttrs.isEmpty()) { 2360 unsigned NestIdx = 1; 2361 Type *NestTy = nullptr; 2362 AttributeSet NestAttr; 2363 2364 // Look for a parameter marked with the 'nest' attribute. 2365 for (FunctionType::param_iterator I = NestFTy->param_begin(), 2366 E = NestFTy->param_end(); I != E; ++NestIdx, ++I) 2367 if (NestAttrs.hasAttribute(NestIdx, Attribute::Nest)) { 2368 // Record the parameter type and any other attributes. 2369 NestTy = *I; 2370 NestAttr = NestAttrs.getParamAttributes(NestIdx); 2371 break; 2372 } 2373 2374 if (NestTy) { 2375 Instruction *Caller = CS.getInstruction(); 2376 std::vector<Value*> NewArgs; 2377 NewArgs.reserve(CS.arg_size() + 1); 2378 2379 SmallVector<AttributeSet, 8> NewAttrs; 2380 NewAttrs.reserve(Attrs.getNumSlots() + 1); 2381 2382 // Insert the nest argument into the call argument list, which may 2383 // mean appending it. Likewise for attributes. 2384 2385 // Add any result attributes. 2386 if (Attrs.hasAttributes(AttributeSet::ReturnIndex)) 2387 NewAttrs.push_back(AttributeSet::get(Caller->getContext(), 2388 Attrs.getRetAttributes())); 2389 2390 { 2391 unsigned Idx = 1; 2392 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 2393 do { 2394 if (Idx == NestIdx) { 2395 // Add the chain argument and attributes. 2396 Value *NestVal = Tramp->getArgOperand(2); 2397 if (NestVal->getType() != NestTy) 2398 NestVal = Builder->CreateBitCast(NestVal, NestTy, "nest"); 2399 NewArgs.push_back(NestVal); 2400 NewAttrs.push_back(AttributeSet::get(Caller->getContext(), 2401 NestAttr)); 2402 } 2403 2404 if (I == E) 2405 break; 2406 2407 // Add the original argument and attributes. 2408 NewArgs.push_back(*I); 2409 AttributeSet Attr = Attrs.getParamAttributes(Idx); 2410 if (Attr.hasAttributes(Idx)) { 2411 AttrBuilder B(Attr, Idx); 2412 NewAttrs.push_back(AttributeSet::get(Caller->getContext(), 2413 Idx + (Idx >= NestIdx), B)); 2414 } 2415 2416 ++Idx, ++I; 2417 } while (1); 2418 } 2419 2420 // Add any function attributes. 2421 if (Attrs.hasAttributes(AttributeSet::FunctionIndex)) 2422 NewAttrs.push_back(AttributeSet::get(FTy->getContext(), 2423 Attrs.getFnAttributes())); 2424 2425 // The trampoline may have been bitcast to a bogus type (FTy). 2426 // Handle this by synthesizing a new function type, equal to FTy 2427 // with the chain parameter inserted. 2428 2429 std::vector<Type*> NewTypes; 2430 NewTypes.reserve(FTy->getNumParams()+1); 2431 2432 // Insert the chain's type into the list of parameter types, which may 2433 // mean appending it. 2434 { 2435 unsigned Idx = 1; 2436 FunctionType::param_iterator I = FTy->param_begin(), 2437 E = FTy->param_end(); 2438 2439 do { 2440 if (Idx == NestIdx) 2441 // Add the chain's type. 2442 NewTypes.push_back(NestTy); 2443 2444 if (I == E) 2445 break; 2446 2447 // Add the original type. 2448 NewTypes.push_back(*I); 2449 2450 ++Idx, ++I; 2451 } while (1); 2452 } 2453 2454 // Replace the trampoline call with a direct call. Let the generic 2455 // code sort out any function type mismatches. 2456 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes, 2457 FTy->isVarArg()); 2458 Constant *NewCallee = 2459 NestF->getType() == PointerType::getUnqual(NewFTy) ? 2460 NestF : ConstantExpr::getBitCast(NestF, 2461 PointerType::getUnqual(NewFTy)); 2462 const AttributeSet &NewPAL = 2463 AttributeSet::get(FTy->getContext(), NewAttrs); 2464 2465 Instruction *NewCaller; 2466 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 2467 NewCaller = InvokeInst::Create(NewCallee, 2468 II->getNormalDest(), II->getUnwindDest(), 2469 NewArgs); 2470 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv()); 2471 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL); 2472 } else { 2473 NewCaller = CallInst::Create(NewCallee, NewArgs); 2474 if (cast<CallInst>(Caller)->isTailCall()) 2475 cast<CallInst>(NewCaller)->setTailCall(); 2476 cast<CallInst>(NewCaller)-> 2477 setCallingConv(cast<CallInst>(Caller)->getCallingConv()); 2478 cast<CallInst>(NewCaller)->setAttributes(NewPAL); 2479 } 2480 2481 return NewCaller; 2482 } 2483 } 2484 2485 // Replace the trampoline call with a direct call. Since there is no 'nest' 2486 // parameter, there is no need to adjust the argument list. Let the generic 2487 // code sort out any function type mismatches. 2488 Constant *NewCallee = 2489 NestF->getType() == PTy ? NestF : 2490 ConstantExpr::getBitCast(NestF, PTy); 2491 CS.setCalledFunction(NewCallee); 2492 return CS.getInstruction(); 2493 } 2494