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/APFloat.h" 16 #include "llvm/ADT/APInt.h" 17 #include "llvm/ADT/ArrayRef.h" 18 #include "llvm/ADT/None.h" 19 #include "llvm/ADT/Optional.h" 20 #include "llvm/ADT/STLExtras.h" 21 #include "llvm/ADT/SmallVector.h" 22 #include "llvm/ADT/Statistic.h" 23 #include "llvm/ADT/Twine.h" 24 #include "llvm/Analysis/AssumptionCache.h" 25 #include "llvm/Analysis/InstructionSimplify.h" 26 #include "llvm/Analysis/MemoryBuiltins.h" 27 #include "llvm/Transforms/Utils/Local.h" 28 #include "llvm/Analysis/ValueTracking.h" 29 #include "llvm/IR/Attributes.h" 30 #include "llvm/IR/BasicBlock.h" 31 #include "llvm/IR/CallSite.h" 32 #include "llvm/IR/Constant.h" 33 #include "llvm/IR/Constants.h" 34 #include "llvm/IR/DataLayout.h" 35 #include "llvm/IR/DerivedTypes.h" 36 #include "llvm/IR/Function.h" 37 #include "llvm/IR/GlobalVariable.h" 38 #include "llvm/IR/InstrTypes.h" 39 #include "llvm/IR/Instruction.h" 40 #include "llvm/IR/Instructions.h" 41 #include "llvm/IR/IntrinsicInst.h" 42 #include "llvm/IR/Intrinsics.h" 43 #include "llvm/IR/LLVMContext.h" 44 #include "llvm/IR/Metadata.h" 45 #include "llvm/IR/PatternMatch.h" 46 #include "llvm/IR/Statepoint.h" 47 #include "llvm/IR/Type.h" 48 #include "llvm/IR/User.h" 49 #include "llvm/IR/Value.h" 50 #include "llvm/IR/ValueHandle.h" 51 #include "llvm/Support/AtomicOrdering.h" 52 #include "llvm/Support/Casting.h" 53 #include "llvm/Support/CommandLine.h" 54 #include "llvm/Support/Compiler.h" 55 #include "llvm/Support/Debug.h" 56 #include "llvm/Support/ErrorHandling.h" 57 #include "llvm/Support/KnownBits.h" 58 #include "llvm/Support/MathExtras.h" 59 #include "llvm/Support/raw_ostream.h" 60 #include "llvm/Transforms/InstCombine/InstCombineWorklist.h" 61 #include "llvm/Transforms/Utils/SimplifyLibCalls.h" 62 #include <algorithm> 63 #include <cassert> 64 #include <cstdint> 65 #include <cstring> 66 #include <utility> 67 #include <vector> 68 69 using namespace llvm; 70 using namespace PatternMatch; 71 72 #define DEBUG_TYPE "instcombine" 73 74 STATISTIC(NumSimplified, "Number of library calls simplified"); 75 76 static cl::opt<unsigned> GuardWideningWindow( 77 "instcombine-guard-widening-window", 78 cl::init(3), 79 cl::desc("How wide an instruction window to bypass looking for " 80 "another guard")); 81 82 /// Return the specified type promoted as it would be to pass though a va_arg 83 /// area. 84 static Type *getPromotedType(Type *Ty) { 85 if (IntegerType* ITy = dyn_cast<IntegerType>(Ty)) { 86 if (ITy->getBitWidth() < 32) 87 return Type::getInt32Ty(Ty->getContext()); 88 } 89 return Ty; 90 } 91 92 /// Return a constant boolean vector that has true elements in all positions 93 /// where the input constant data vector has an element with the sign bit set. 94 static Constant *getNegativeIsTrueBoolVec(ConstantDataVector *V) { 95 SmallVector<Constant *, 32> BoolVec; 96 IntegerType *BoolTy = Type::getInt1Ty(V->getContext()); 97 for (unsigned I = 0, E = V->getNumElements(); I != E; ++I) { 98 Constant *Elt = V->getElementAsConstant(I); 99 assert((isa<ConstantInt>(Elt) || isa<ConstantFP>(Elt)) && 100 "Unexpected constant data vector element type"); 101 bool Sign = V->getElementType()->isIntegerTy() 102 ? cast<ConstantInt>(Elt)->isNegative() 103 : cast<ConstantFP>(Elt)->isNegative(); 104 BoolVec.push_back(ConstantInt::get(BoolTy, Sign)); 105 } 106 return ConstantVector::get(BoolVec); 107 } 108 109 Instruction *InstCombiner::SimplifyAnyMemTransfer(AnyMemTransferInst *MI) { 110 unsigned DstAlign = getKnownAlignment(MI->getRawDest(), DL, MI, &AC, &DT); 111 unsigned CopyDstAlign = MI->getDestAlignment(); 112 if (CopyDstAlign < DstAlign){ 113 MI->setDestAlignment(DstAlign); 114 return MI; 115 } 116 117 unsigned SrcAlign = getKnownAlignment(MI->getRawSource(), DL, MI, &AC, &DT); 118 unsigned CopySrcAlign = MI->getSourceAlignment(); 119 if (CopySrcAlign < SrcAlign) { 120 MI->setSourceAlignment(SrcAlign); 121 return MI; 122 } 123 124 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with 125 // load/store. 126 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getLength()); 127 if (!MemOpLength) return nullptr; 128 129 // Source and destination pointer types are always "i8*" for intrinsic. See 130 // if the size is something we can handle with a single primitive load/store. 131 // A single load+store correctly handles overlapping memory in the memmove 132 // case. 133 uint64_t Size = MemOpLength->getLimitedValue(); 134 assert(Size && "0-sized memory transferring should be removed already."); 135 136 if (Size > 8 || (Size&(Size-1))) 137 return nullptr; // If not 1/2/4/8 bytes, exit. 138 139 // Use an integer load+store unless we can find something better. 140 unsigned SrcAddrSp = 141 cast<PointerType>(MI->getArgOperand(1)->getType())->getAddressSpace(); 142 unsigned DstAddrSp = 143 cast<PointerType>(MI->getArgOperand(0)->getType())->getAddressSpace(); 144 145 IntegerType* IntType = IntegerType::get(MI->getContext(), Size<<3); 146 Type *NewSrcPtrTy = PointerType::get(IntType, SrcAddrSp); 147 Type *NewDstPtrTy = PointerType::get(IntType, DstAddrSp); 148 149 // If the memcpy has metadata describing the members, see if we can get the 150 // TBAA tag describing our copy. 151 MDNode *CopyMD = nullptr; 152 if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa)) { 153 CopyMD = M; 154 } else if (MDNode *M = MI->getMetadata(LLVMContext::MD_tbaa_struct)) { 155 if (M->getNumOperands() == 3 && M->getOperand(0) && 156 mdconst::hasa<ConstantInt>(M->getOperand(0)) && 157 mdconst::extract<ConstantInt>(M->getOperand(0))->isZero() && 158 M->getOperand(1) && 159 mdconst::hasa<ConstantInt>(M->getOperand(1)) && 160 mdconst::extract<ConstantInt>(M->getOperand(1))->getValue() == 161 Size && 162 M->getOperand(2) && isa<MDNode>(M->getOperand(2))) 163 CopyMD = cast<MDNode>(M->getOperand(2)); 164 } 165 166 Value *Src = Builder.CreateBitCast(MI->getArgOperand(1), NewSrcPtrTy); 167 Value *Dest = Builder.CreateBitCast(MI->getArgOperand(0), NewDstPtrTy); 168 LoadInst *L = Builder.CreateLoad(Src); 169 // Alignment from the mem intrinsic will be better, so use it. 170 L->setAlignment(CopySrcAlign); 171 if (CopyMD) 172 L->setMetadata(LLVMContext::MD_tbaa, CopyMD); 173 MDNode *LoopMemParallelMD = 174 MI->getMetadata(LLVMContext::MD_mem_parallel_loop_access); 175 if (LoopMemParallelMD) 176 L->setMetadata(LLVMContext::MD_mem_parallel_loop_access, LoopMemParallelMD); 177 178 StoreInst *S = Builder.CreateStore(L, Dest); 179 // Alignment from the mem intrinsic will be better, so use it. 180 S->setAlignment(CopyDstAlign); 181 if (CopyMD) 182 S->setMetadata(LLVMContext::MD_tbaa, CopyMD); 183 if (LoopMemParallelMD) 184 S->setMetadata(LLVMContext::MD_mem_parallel_loop_access, LoopMemParallelMD); 185 186 if (auto *MT = dyn_cast<MemTransferInst>(MI)) { 187 // non-atomics can be volatile 188 L->setVolatile(MT->isVolatile()); 189 S->setVolatile(MT->isVolatile()); 190 } 191 if (isa<AtomicMemTransferInst>(MI)) { 192 // atomics have to be unordered 193 L->setOrdering(AtomicOrdering::Unordered); 194 S->setOrdering(AtomicOrdering::Unordered); 195 } 196 197 // Set the size of the copy to 0, it will be deleted on the next iteration. 198 MI->setLength(Constant::getNullValue(MemOpLength->getType())); 199 return MI; 200 } 201 202 Instruction *InstCombiner::SimplifyAnyMemSet(AnyMemSetInst *MI) { 203 unsigned Alignment = getKnownAlignment(MI->getDest(), DL, MI, &AC, &DT); 204 if (MI->getDestAlignment() < Alignment) { 205 MI->setDestAlignment(Alignment); 206 return MI; 207 } 208 209 // Extract the length and alignment and fill if they are constant. 210 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength()); 211 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue()); 212 if (!LenC || !FillC || !FillC->getType()->isIntegerTy(8)) 213 return nullptr; 214 uint64_t Len = LenC->getLimitedValue(); 215 Alignment = MI->getDestAlignment(); 216 assert(Len && "0-sized memory setting should be removed already."); 217 218 // memset(s,c,n) -> store s, c (for n=1,2,4,8) 219 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) { 220 Type *ITy = IntegerType::get(MI->getContext(), Len*8); // n=1 -> i8. 221 222 Value *Dest = MI->getDest(); 223 unsigned DstAddrSp = cast<PointerType>(Dest->getType())->getAddressSpace(); 224 Type *NewDstPtrTy = PointerType::get(ITy, DstAddrSp); 225 Dest = Builder.CreateBitCast(Dest, NewDstPtrTy); 226 227 // Alignment 0 is identity for alignment 1 for memset, but not store. 228 if (Alignment == 0) Alignment = 1; 229 230 // Extract the fill value and store. 231 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL; 232 StoreInst *S = Builder.CreateStore(ConstantInt::get(ITy, Fill), Dest, 233 MI->isVolatile()); 234 S->setAlignment(Alignment); 235 if (isa<AtomicMemSetInst>(MI)) 236 S->setOrdering(AtomicOrdering::Unordered); 237 238 // Set the size of the copy to 0, it will be deleted on the next iteration. 239 MI->setLength(Constant::getNullValue(LenC->getType())); 240 return MI; 241 } 242 243 return nullptr; 244 } 245 246 static Value *simplifyX86immShift(const IntrinsicInst &II, 247 InstCombiner::BuilderTy &Builder) { 248 bool LogicalShift = false; 249 bool ShiftLeft = false; 250 251 switch (II.getIntrinsicID()) { 252 default: llvm_unreachable("Unexpected intrinsic!"); 253 case Intrinsic::x86_sse2_psra_d: 254 case Intrinsic::x86_sse2_psra_w: 255 case Intrinsic::x86_sse2_psrai_d: 256 case Intrinsic::x86_sse2_psrai_w: 257 case Intrinsic::x86_avx2_psra_d: 258 case Intrinsic::x86_avx2_psra_w: 259 case Intrinsic::x86_avx2_psrai_d: 260 case Intrinsic::x86_avx2_psrai_w: 261 case Intrinsic::x86_avx512_psra_q_128: 262 case Intrinsic::x86_avx512_psrai_q_128: 263 case Intrinsic::x86_avx512_psra_q_256: 264 case Intrinsic::x86_avx512_psrai_q_256: 265 case Intrinsic::x86_avx512_psra_d_512: 266 case Intrinsic::x86_avx512_psra_q_512: 267 case Intrinsic::x86_avx512_psra_w_512: 268 case Intrinsic::x86_avx512_psrai_d_512: 269 case Intrinsic::x86_avx512_psrai_q_512: 270 case Intrinsic::x86_avx512_psrai_w_512: 271 LogicalShift = false; ShiftLeft = false; 272 break; 273 case Intrinsic::x86_sse2_psrl_d: 274 case Intrinsic::x86_sse2_psrl_q: 275 case Intrinsic::x86_sse2_psrl_w: 276 case Intrinsic::x86_sse2_psrli_d: 277 case Intrinsic::x86_sse2_psrli_q: 278 case Intrinsic::x86_sse2_psrli_w: 279 case Intrinsic::x86_avx2_psrl_d: 280 case Intrinsic::x86_avx2_psrl_q: 281 case Intrinsic::x86_avx2_psrl_w: 282 case Intrinsic::x86_avx2_psrli_d: 283 case Intrinsic::x86_avx2_psrli_q: 284 case Intrinsic::x86_avx2_psrli_w: 285 case Intrinsic::x86_avx512_psrl_d_512: 286 case Intrinsic::x86_avx512_psrl_q_512: 287 case Intrinsic::x86_avx512_psrl_w_512: 288 case Intrinsic::x86_avx512_psrli_d_512: 289 case Intrinsic::x86_avx512_psrli_q_512: 290 case Intrinsic::x86_avx512_psrli_w_512: 291 LogicalShift = true; ShiftLeft = false; 292 break; 293 case Intrinsic::x86_sse2_psll_d: 294 case Intrinsic::x86_sse2_psll_q: 295 case Intrinsic::x86_sse2_psll_w: 296 case Intrinsic::x86_sse2_pslli_d: 297 case Intrinsic::x86_sse2_pslli_q: 298 case Intrinsic::x86_sse2_pslli_w: 299 case Intrinsic::x86_avx2_psll_d: 300 case Intrinsic::x86_avx2_psll_q: 301 case Intrinsic::x86_avx2_psll_w: 302 case Intrinsic::x86_avx2_pslli_d: 303 case Intrinsic::x86_avx2_pslli_q: 304 case Intrinsic::x86_avx2_pslli_w: 305 case Intrinsic::x86_avx512_psll_d_512: 306 case Intrinsic::x86_avx512_psll_q_512: 307 case Intrinsic::x86_avx512_psll_w_512: 308 case Intrinsic::x86_avx512_pslli_d_512: 309 case Intrinsic::x86_avx512_pslli_q_512: 310 case Intrinsic::x86_avx512_pslli_w_512: 311 LogicalShift = true; ShiftLeft = true; 312 break; 313 } 314 assert((LogicalShift || !ShiftLeft) && "Only logical shifts can shift left"); 315 316 // Simplify if count is constant. 317 auto Arg1 = II.getArgOperand(1); 318 auto CAZ = dyn_cast<ConstantAggregateZero>(Arg1); 319 auto CDV = dyn_cast<ConstantDataVector>(Arg1); 320 auto CInt = dyn_cast<ConstantInt>(Arg1); 321 if (!CAZ && !CDV && !CInt) 322 return nullptr; 323 324 APInt Count(64, 0); 325 if (CDV) { 326 // SSE2/AVX2 uses all the first 64-bits of the 128-bit vector 327 // operand to compute the shift amount. 328 auto VT = cast<VectorType>(CDV->getType()); 329 unsigned BitWidth = VT->getElementType()->getPrimitiveSizeInBits(); 330 assert((64 % BitWidth) == 0 && "Unexpected packed shift size"); 331 unsigned NumSubElts = 64 / BitWidth; 332 333 // Concatenate the sub-elements to create the 64-bit value. 334 for (unsigned i = 0; i != NumSubElts; ++i) { 335 unsigned SubEltIdx = (NumSubElts - 1) - i; 336 auto SubElt = cast<ConstantInt>(CDV->getElementAsConstant(SubEltIdx)); 337 Count <<= BitWidth; 338 Count |= SubElt->getValue().zextOrTrunc(64); 339 } 340 } 341 else if (CInt) 342 Count = CInt->getValue(); 343 344 auto Vec = II.getArgOperand(0); 345 auto VT = cast<VectorType>(Vec->getType()); 346 auto SVT = VT->getElementType(); 347 unsigned VWidth = VT->getNumElements(); 348 unsigned BitWidth = SVT->getPrimitiveSizeInBits(); 349 350 // If shift-by-zero then just return the original value. 351 if (Count.isNullValue()) 352 return Vec; 353 354 // Handle cases when Shift >= BitWidth. 355 if (Count.uge(BitWidth)) { 356 // If LogicalShift - just return zero. 357 if (LogicalShift) 358 return ConstantAggregateZero::get(VT); 359 360 // If ArithmeticShift - clamp Shift to (BitWidth - 1). 361 Count = APInt(64, BitWidth - 1); 362 } 363 364 // Get a constant vector of the same type as the first operand. 365 auto ShiftAmt = ConstantInt::get(SVT, Count.zextOrTrunc(BitWidth)); 366 auto ShiftVec = Builder.CreateVectorSplat(VWidth, ShiftAmt); 367 368 if (ShiftLeft) 369 return Builder.CreateShl(Vec, ShiftVec); 370 371 if (LogicalShift) 372 return Builder.CreateLShr(Vec, ShiftVec); 373 374 return Builder.CreateAShr(Vec, ShiftVec); 375 } 376 377 // Attempt to simplify AVX2 per-element shift intrinsics to a generic IR shift. 378 // Unlike the generic IR shifts, the intrinsics have defined behaviour for out 379 // of range shift amounts (logical - set to zero, arithmetic - splat sign bit). 380 static Value *simplifyX86varShift(const IntrinsicInst &II, 381 InstCombiner::BuilderTy &Builder) { 382 bool LogicalShift = false; 383 bool ShiftLeft = false; 384 385 switch (II.getIntrinsicID()) { 386 default: llvm_unreachable("Unexpected intrinsic!"); 387 case Intrinsic::x86_avx2_psrav_d: 388 case Intrinsic::x86_avx2_psrav_d_256: 389 case Intrinsic::x86_avx512_psrav_q_128: 390 case Intrinsic::x86_avx512_psrav_q_256: 391 case Intrinsic::x86_avx512_psrav_d_512: 392 case Intrinsic::x86_avx512_psrav_q_512: 393 case Intrinsic::x86_avx512_psrav_w_128: 394 case Intrinsic::x86_avx512_psrav_w_256: 395 case Intrinsic::x86_avx512_psrav_w_512: 396 LogicalShift = false; 397 ShiftLeft = false; 398 break; 399 case Intrinsic::x86_avx2_psrlv_d: 400 case Intrinsic::x86_avx2_psrlv_d_256: 401 case Intrinsic::x86_avx2_psrlv_q: 402 case Intrinsic::x86_avx2_psrlv_q_256: 403 case Intrinsic::x86_avx512_psrlv_d_512: 404 case Intrinsic::x86_avx512_psrlv_q_512: 405 case Intrinsic::x86_avx512_psrlv_w_128: 406 case Intrinsic::x86_avx512_psrlv_w_256: 407 case Intrinsic::x86_avx512_psrlv_w_512: 408 LogicalShift = true; 409 ShiftLeft = false; 410 break; 411 case Intrinsic::x86_avx2_psllv_d: 412 case Intrinsic::x86_avx2_psllv_d_256: 413 case Intrinsic::x86_avx2_psllv_q: 414 case Intrinsic::x86_avx2_psllv_q_256: 415 case Intrinsic::x86_avx512_psllv_d_512: 416 case Intrinsic::x86_avx512_psllv_q_512: 417 case Intrinsic::x86_avx512_psllv_w_128: 418 case Intrinsic::x86_avx512_psllv_w_256: 419 case Intrinsic::x86_avx512_psllv_w_512: 420 LogicalShift = true; 421 ShiftLeft = true; 422 break; 423 } 424 assert((LogicalShift || !ShiftLeft) && "Only logical shifts can shift left"); 425 426 // Simplify if all shift amounts are constant/undef. 427 auto *CShift = dyn_cast<Constant>(II.getArgOperand(1)); 428 if (!CShift) 429 return nullptr; 430 431 auto Vec = II.getArgOperand(0); 432 auto VT = cast<VectorType>(II.getType()); 433 auto SVT = VT->getVectorElementType(); 434 int NumElts = VT->getNumElements(); 435 int BitWidth = SVT->getIntegerBitWidth(); 436 437 // Collect each element's shift amount. 438 // We also collect special cases: UNDEF = -1, OUT-OF-RANGE = BitWidth. 439 bool AnyOutOfRange = false; 440 SmallVector<int, 8> ShiftAmts; 441 for (int I = 0; I < NumElts; ++I) { 442 auto *CElt = CShift->getAggregateElement(I); 443 if (CElt && isa<UndefValue>(CElt)) { 444 ShiftAmts.push_back(-1); 445 continue; 446 } 447 448 auto *COp = dyn_cast_or_null<ConstantInt>(CElt); 449 if (!COp) 450 return nullptr; 451 452 // Handle out of range shifts. 453 // If LogicalShift - set to BitWidth (special case). 454 // If ArithmeticShift - set to (BitWidth - 1) (sign splat). 455 APInt ShiftVal = COp->getValue(); 456 if (ShiftVal.uge(BitWidth)) { 457 AnyOutOfRange = LogicalShift; 458 ShiftAmts.push_back(LogicalShift ? BitWidth : BitWidth - 1); 459 continue; 460 } 461 462 ShiftAmts.push_back((int)ShiftVal.getZExtValue()); 463 } 464 465 // If all elements out of range or UNDEF, return vector of zeros/undefs. 466 // ArithmeticShift should only hit this if they are all UNDEF. 467 auto OutOfRange = [&](int Idx) { return (Idx < 0) || (BitWidth <= Idx); }; 468 if (llvm::all_of(ShiftAmts, OutOfRange)) { 469 SmallVector<Constant *, 8> ConstantVec; 470 for (int Idx : ShiftAmts) { 471 if (Idx < 0) { 472 ConstantVec.push_back(UndefValue::get(SVT)); 473 } else { 474 assert(LogicalShift && "Logical shift expected"); 475 ConstantVec.push_back(ConstantInt::getNullValue(SVT)); 476 } 477 } 478 return ConstantVector::get(ConstantVec); 479 } 480 481 // We can't handle only some out of range values with generic logical shifts. 482 if (AnyOutOfRange) 483 return nullptr; 484 485 // Build the shift amount constant vector. 486 SmallVector<Constant *, 8> ShiftVecAmts; 487 for (int Idx : ShiftAmts) { 488 if (Idx < 0) 489 ShiftVecAmts.push_back(UndefValue::get(SVT)); 490 else 491 ShiftVecAmts.push_back(ConstantInt::get(SVT, Idx)); 492 } 493 auto ShiftVec = ConstantVector::get(ShiftVecAmts); 494 495 if (ShiftLeft) 496 return Builder.CreateShl(Vec, ShiftVec); 497 498 if (LogicalShift) 499 return Builder.CreateLShr(Vec, ShiftVec); 500 501 return Builder.CreateAShr(Vec, ShiftVec); 502 } 503 504 static Value *simplifyX86pack(IntrinsicInst &II, bool IsSigned) { 505 Value *Arg0 = II.getArgOperand(0); 506 Value *Arg1 = II.getArgOperand(1); 507 Type *ResTy = II.getType(); 508 509 // Fast all undef handling. 510 if (isa<UndefValue>(Arg0) && isa<UndefValue>(Arg1)) 511 return UndefValue::get(ResTy); 512 513 Type *ArgTy = Arg0->getType(); 514 unsigned NumLanes = ResTy->getPrimitiveSizeInBits() / 128; 515 unsigned NumDstElts = ResTy->getVectorNumElements(); 516 unsigned NumSrcElts = ArgTy->getVectorNumElements(); 517 assert(NumDstElts == (2 * NumSrcElts) && "Unexpected packing types"); 518 519 unsigned NumDstEltsPerLane = NumDstElts / NumLanes; 520 unsigned NumSrcEltsPerLane = NumSrcElts / NumLanes; 521 unsigned DstScalarSizeInBits = ResTy->getScalarSizeInBits(); 522 assert(ArgTy->getScalarSizeInBits() == (2 * DstScalarSizeInBits) && 523 "Unexpected packing types"); 524 525 // Constant folding. 526 auto *Cst0 = dyn_cast<Constant>(Arg0); 527 auto *Cst1 = dyn_cast<Constant>(Arg1); 528 if (!Cst0 || !Cst1) 529 return nullptr; 530 531 SmallVector<Constant *, 32> Vals; 532 for (unsigned Lane = 0; Lane != NumLanes; ++Lane) { 533 for (unsigned Elt = 0; Elt != NumDstEltsPerLane; ++Elt) { 534 unsigned SrcIdx = Lane * NumSrcEltsPerLane + Elt % NumSrcEltsPerLane; 535 auto *Cst = (Elt >= NumSrcEltsPerLane) ? Cst1 : Cst0; 536 auto *COp = Cst->getAggregateElement(SrcIdx); 537 if (COp && isa<UndefValue>(COp)) { 538 Vals.push_back(UndefValue::get(ResTy->getScalarType())); 539 continue; 540 } 541 542 auto *CInt = dyn_cast_or_null<ConstantInt>(COp); 543 if (!CInt) 544 return nullptr; 545 546 APInt Val = CInt->getValue(); 547 assert(Val.getBitWidth() == ArgTy->getScalarSizeInBits() && 548 "Unexpected constant bitwidth"); 549 550 if (IsSigned) { 551 // PACKSS: Truncate signed value with signed saturation. 552 // Source values less than dst minint are saturated to minint. 553 // Source values greater than dst maxint are saturated to maxint. 554 if (Val.isSignedIntN(DstScalarSizeInBits)) 555 Val = Val.trunc(DstScalarSizeInBits); 556 else if (Val.isNegative()) 557 Val = APInt::getSignedMinValue(DstScalarSizeInBits); 558 else 559 Val = APInt::getSignedMaxValue(DstScalarSizeInBits); 560 } else { 561 // PACKUS: Truncate signed value with unsigned saturation. 562 // Source values less than zero are saturated to zero. 563 // Source values greater than dst maxuint are saturated to maxuint. 564 if (Val.isIntN(DstScalarSizeInBits)) 565 Val = Val.trunc(DstScalarSizeInBits); 566 else if (Val.isNegative()) 567 Val = APInt::getNullValue(DstScalarSizeInBits); 568 else 569 Val = APInt::getAllOnesValue(DstScalarSizeInBits); 570 } 571 572 Vals.push_back(ConstantInt::get(ResTy->getScalarType(), Val)); 573 } 574 } 575 576 return ConstantVector::get(Vals); 577 } 578 579 // Replace X86-specific intrinsics with generic floor-ceil where applicable. 580 static Value *simplifyX86round(IntrinsicInst &II, 581 InstCombiner::BuilderTy &Builder) { 582 ConstantInt *Arg = nullptr; 583 Intrinsic::ID IntrinsicID = II.getIntrinsicID(); 584 585 if (IntrinsicID == Intrinsic::x86_sse41_round_ss || 586 IntrinsicID == Intrinsic::x86_sse41_round_sd) 587 Arg = dyn_cast<ConstantInt>(II.getArgOperand(2)); 588 else if (IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ss || 589 IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_sd) 590 Arg = dyn_cast<ConstantInt>(II.getArgOperand(4)); 591 else 592 Arg = dyn_cast<ConstantInt>(II.getArgOperand(1)); 593 if (!Arg) 594 return nullptr; 595 unsigned RoundControl = Arg->getZExtValue(); 596 597 Arg = nullptr; 598 unsigned SAE = 0; 599 if (IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ps_512 || 600 IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_pd_512) 601 Arg = dyn_cast<ConstantInt>(II.getArgOperand(4)); 602 else if (IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ss || 603 IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_sd) 604 Arg = dyn_cast<ConstantInt>(II.getArgOperand(5)); 605 else 606 SAE = 4; 607 if (!SAE) { 608 if (!Arg) 609 return nullptr; 610 SAE = Arg->getZExtValue(); 611 } 612 613 if (SAE != 4 || (RoundControl != 2 /*ceil*/ && RoundControl != 1 /*floor*/)) 614 return nullptr; 615 616 Value *Src, *Dst, *Mask; 617 bool IsScalar = false; 618 if (IntrinsicID == Intrinsic::x86_sse41_round_ss || 619 IntrinsicID == Intrinsic::x86_sse41_round_sd || 620 IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ss || 621 IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_sd) { 622 IsScalar = true; 623 if (IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ss || 624 IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_sd) { 625 Mask = II.getArgOperand(3); 626 Value *Zero = Constant::getNullValue(Mask->getType()); 627 Mask = Builder.CreateAnd(Mask, 1); 628 Mask = Builder.CreateICmp(ICmpInst::ICMP_NE, Mask, Zero); 629 Dst = II.getArgOperand(2); 630 } else 631 Dst = II.getArgOperand(0); 632 Src = Builder.CreateExtractElement(II.getArgOperand(1), (uint64_t)0); 633 } else { 634 Src = II.getArgOperand(0); 635 if (IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ps_128 || 636 IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ps_256 || 637 IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ps_512 || 638 IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_pd_128 || 639 IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_pd_256 || 640 IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_pd_512) { 641 Dst = II.getArgOperand(2); 642 Mask = II.getArgOperand(3); 643 } else { 644 Dst = Src; 645 Mask = ConstantInt::getAllOnesValue( 646 Builder.getIntNTy(Src->getType()->getVectorNumElements())); 647 } 648 } 649 650 Intrinsic::ID ID = (RoundControl == 2) ? Intrinsic::ceil : Intrinsic::floor; 651 Value *Res = Builder.CreateIntrinsic(ID, {Src}, &II); 652 if (!IsScalar) { 653 if (auto *C = dyn_cast<Constant>(Mask)) 654 if (C->isAllOnesValue()) 655 return Res; 656 auto *MaskTy = VectorType::get( 657 Builder.getInt1Ty(), cast<IntegerType>(Mask->getType())->getBitWidth()); 658 Mask = Builder.CreateBitCast(Mask, MaskTy); 659 unsigned Width = Src->getType()->getVectorNumElements(); 660 if (MaskTy->getVectorNumElements() > Width) { 661 uint32_t Indices[4]; 662 for (unsigned i = 0; i != Width; ++i) 663 Indices[i] = i; 664 Mask = Builder.CreateShuffleVector(Mask, Mask, 665 makeArrayRef(Indices, Width)); 666 } 667 return Builder.CreateSelect(Mask, Res, Dst); 668 } 669 if (IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_ss || 670 IntrinsicID == Intrinsic::x86_avx512_mask_rndscale_sd) { 671 Dst = Builder.CreateExtractElement(Dst, (uint64_t)0); 672 Res = Builder.CreateSelect(Mask, Res, Dst); 673 Dst = II.getArgOperand(0); 674 } 675 return Builder.CreateInsertElement(Dst, Res, (uint64_t)0); 676 } 677 678 static Value *simplifyX86movmsk(const IntrinsicInst &II) { 679 Value *Arg = II.getArgOperand(0); 680 Type *ResTy = II.getType(); 681 Type *ArgTy = Arg->getType(); 682 683 // movmsk(undef) -> zero as we must ensure the upper bits are zero. 684 if (isa<UndefValue>(Arg)) 685 return Constant::getNullValue(ResTy); 686 687 // We can't easily peek through x86_mmx types. 688 if (!ArgTy->isVectorTy()) 689 return nullptr; 690 691 auto *C = dyn_cast<Constant>(Arg); 692 if (!C) 693 return nullptr; 694 695 // Extract signbits of the vector input and pack into integer result. 696 APInt Result(ResTy->getPrimitiveSizeInBits(), 0); 697 for (unsigned I = 0, E = ArgTy->getVectorNumElements(); I != E; ++I) { 698 auto *COp = C->getAggregateElement(I); 699 if (!COp) 700 return nullptr; 701 if (isa<UndefValue>(COp)) 702 continue; 703 704 auto *CInt = dyn_cast<ConstantInt>(COp); 705 auto *CFp = dyn_cast<ConstantFP>(COp); 706 if (!CInt && !CFp) 707 return nullptr; 708 709 if ((CInt && CInt->isNegative()) || (CFp && CFp->isNegative())) 710 Result.setBit(I); 711 } 712 713 return Constant::getIntegerValue(ResTy, Result); 714 } 715 716 static Value *simplifyX86insertps(const IntrinsicInst &II, 717 InstCombiner::BuilderTy &Builder) { 718 auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2)); 719 if (!CInt) 720 return nullptr; 721 722 VectorType *VecTy = cast<VectorType>(II.getType()); 723 assert(VecTy->getNumElements() == 4 && "insertps with wrong vector type"); 724 725 // The immediate permute control byte looks like this: 726 // [3:0] - zero mask for each 32-bit lane 727 // [5:4] - select one 32-bit destination lane 728 // [7:6] - select one 32-bit source lane 729 730 uint8_t Imm = CInt->getZExtValue(); 731 uint8_t ZMask = Imm & 0xf; 732 uint8_t DestLane = (Imm >> 4) & 0x3; 733 uint8_t SourceLane = (Imm >> 6) & 0x3; 734 735 ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy); 736 737 // If all zero mask bits are set, this was just a weird way to 738 // generate a zero vector. 739 if (ZMask == 0xf) 740 return ZeroVector; 741 742 // Initialize by passing all of the first source bits through. 743 uint32_t ShuffleMask[4] = { 0, 1, 2, 3 }; 744 745 // We may replace the second operand with the zero vector. 746 Value *V1 = II.getArgOperand(1); 747 748 if (ZMask) { 749 // If the zero mask is being used with a single input or the zero mask 750 // overrides the destination lane, this is a shuffle with the zero vector. 751 if ((II.getArgOperand(0) == II.getArgOperand(1)) || 752 (ZMask & (1 << DestLane))) { 753 V1 = ZeroVector; 754 // We may still move 32-bits of the first source vector from one lane 755 // to another. 756 ShuffleMask[DestLane] = SourceLane; 757 // The zero mask may override the previous insert operation. 758 for (unsigned i = 0; i < 4; ++i) 759 if ((ZMask >> i) & 0x1) 760 ShuffleMask[i] = i + 4; 761 } else { 762 // TODO: Model this case as 2 shuffles or a 'logical and' plus shuffle? 763 return nullptr; 764 } 765 } else { 766 // Replace the selected destination lane with the selected source lane. 767 ShuffleMask[DestLane] = SourceLane + 4; 768 } 769 770 return Builder.CreateShuffleVector(II.getArgOperand(0), V1, ShuffleMask); 771 } 772 773 /// Attempt to simplify SSE4A EXTRQ/EXTRQI instructions using constant folding 774 /// or conversion to a shuffle vector. 775 static Value *simplifyX86extrq(IntrinsicInst &II, Value *Op0, 776 ConstantInt *CILength, ConstantInt *CIIndex, 777 InstCombiner::BuilderTy &Builder) { 778 auto LowConstantHighUndef = [&](uint64_t Val) { 779 Type *IntTy64 = Type::getInt64Ty(II.getContext()); 780 Constant *Args[] = {ConstantInt::get(IntTy64, Val), 781 UndefValue::get(IntTy64)}; 782 return ConstantVector::get(Args); 783 }; 784 785 // See if we're dealing with constant values. 786 Constant *C0 = dyn_cast<Constant>(Op0); 787 ConstantInt *CI0 = 788 C0 ? dyn_cast_or_null<ConstantInt>(C0->getAggregateElement((unsigned)0)) 789 : nullptr; 790 791 // Attempt to constant fold. 792 if (CILength && CIIndex) { 793 // From AMD documentation: "The bit index and field length are each six 794 // bits in length other bits of the field are ignored." 795 APInt APIndex = CIIndex->getValue().zextOrTrunc(6); 796 APInt APLength = CILength->getValue().zextOrTrunc(6); 797 798 unsigned Index = APIndex.getZExtValue(); 799 800 // From AMD documentation: "a value of zero in the field length is 801 // defined as length of 64". 802 unsigned Length = APLength == 0 ? 64 : APLength.getZExtValue(); 803 804 // From AMD documentation: "If the sum of the bit index + length field 805 // is greater than 64, the results are undefined". 806 unsigned End = Index + Length; 807 808 // Note that both field index and field length are 8-bit quantities. 809 // Since variables 'Index' and 'Length' are unsigned values 810 // obtained from zero-extending field index and field length 811 // respectively, their sum should never wrap around. 812 if (End > 64) 813 return UndefValue::get(II.getType()); 814 815 // If we are inserting whole bytes, we can convert this to a shuffle. 816 // Lowering can recognize EXTRQI shuffle masks. 817 if ((Length % 8) == 0 && (Index % 8) == 0) { 818 // Convert bit indices to byte indices. 819 Length /= 8; 820 Index /= 8; 821 822 Type *IntTy8 = Type::getInt8Ty(II.getContext()); 823 Type *IntTy32 = Type::getInt32Ty(II.getContext()); 824 VectorType *ShufTy = VectorType::get(IntTy8, 16); 825 826 SmallVector<Constant *, 16> ShuffleMask; 827 for (int i = 0; i != (int)Length; ++i) 828 ShuffleMask.push_back( 829 Constant::getIntegerValue(IntTy32, APInt(32, i + Index))); 830 for (int i = Length; i != 8; ++i) 831 ShuffleMask.push_back( 832 Constant::getIntegerValue(IntTy32, APInt(32, i + 16))); 833 for (int i = 8; i != 16; ++i) 834 ShuffleMask.push_back(UndefValue::get(IntTy32)); 835 836 Value *SV = Builder.CreateShuffleVector( 837 Builder.CreateBitCast(Op0, ShufTy), 838 ConstantAggregateZero::get(ShufTy), ConstantVector::get(ShuffleMask)); 839 return Builder.CreateBitCast(SV, II.getType()); 840 } 841 842 // Constant Fold - shift Index'th bit to lowest position and mask off 843 // Length bits. 844 if (CI0) { 845 APInt Elt = CI0->getValue(); 846 Elt.lshrInPlace(Index); 847 Elt = Elt.zextOrTrunc(Length); 848 return LowConstantHighUndef(Elt.getZExtValue()); 849 } 850 851 // If we were an EXTRQ call, we'll save registers if we convert to EXTRQI. 852 if (II.getIntrinsicID() == Intrinsic::x86_sse4a_extrq) { 853 Value *Args[] = {Op0, CILength, CIIndex}; 854 Module *M = II.getModule(); 855 Value *F = Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_extrqi); 856 return Builder.CreateCall(F, Args); 857 } 858 } 859 860 // Constant Fold - extraction from zero is always {zero, undef}. 861 if (CI0 && CI0->isZero()) 862 return LowConstantHighUndef(0); 863 864 return nullptr; 865 } 866 867 /// Attempt to simplify SSE4A INSERTQ/INSERTQI instructions using constant 868 /// folding or conversion to a shuffle vector. 869 static Value *simplifyX86insertq(IntrinsicInst &II, Value *Op0, Value *Op1, 870 APInt APLength, APInt APIndex, 871 InstCombiner::BuilderTy &Builder) { 872 // From AMD documentation: "The bit index and field length are each six bits 873 // in length other bits of the field are ignored." 874 APIndex = APIndex.zextOrTrunc(6); 875 APLength = APLength.zextOrTrunc(6); 876 877 // Attempt to constant fold. 878 unsigned Index = APIndex.getZExtValue(); 879 880 // From AMD documentation: "a value of zero in the field length is 881 // defined as length of 64". 882 unsigned Length = APLength == 0 ? 64 : APLength.getZExtValue(); 883 884 // From AMD documentation: "If the sum of the bit index + length field 885 // is greater than 64, the results are undefined". 886 unsigned End = Index + Length; 887 888 // Note that both field index and field length are 8-bit quantities. 889 // Since variables 'Index' and 'Length' are unsigned values 890 // obtained from zero-extending field index and field length 891 // respectively, their sum should never wrap around. 892 if (End > 64) 893 return UndefValue::get(II.getType()); 894 895 // If we are inserting whole bytes, we can convert this to a shuffle. 896 // Lowering can recognize INSERTQI shuffle masks. 897 if ((Length % 8) == 0 && (Index % 8) == 0) { 898 // Convert bit indices to byte indices. 899 Length /= 8; 900 Index /= 8; 901 902 Type *IntTy8 = Type::getInt8Ty(II.getContext()); 903 Type *IntTy32 = Type::getInt32Ty(II.getContext()); 904 VectorType *ShufTy = VectorType::get(IntTy8, 16); 905 906 SmallVector<Constant *, 16> ShuffleMask; 907 for (int i = 0; i != (int)Index; ++i) 908 ShuffleMask.push_back(Constant::getIntegerValue(IntTy32, APInt(32, i))); 909 for (int i = 0; i != (int)Length; ++i) 910 ShuffleMask.push_back( 911 Constant::getIntegerValue(IntTy32, APInt(32, i + 16))); 912 for (int i = Index + Length; i != 8; ++i) 913 ShuffleMask.push_back(Constant::getIntegerValue(IntTy32, APInt(32, i))); 914 for (int i = 8; i != 16; ++i) 915 ShuffleMask.push_back(UndefValue::get(IntTy32)); 916 917 Value *SV = Builder.CreateShuffleVector(Builder.CreateBitCast(Op0, ShufTy), 918 Builder.CreateBitCast(Op1, ShufTy), 919 ConstantVector::get(ShuffleMask)); 920 return Builder.CreateBitCast(SV, II.getType()); 921 } 922 923 // See if we're dealing with constant values. 924 Constant *C0 = dyn_cast<Constant>(Op0); 925 Constant *C1 = dyn_cast<Constant>(Op1); 926 ConstantInt *CI00 = 927 C0 ? dyn_cast_or_null<ConstantInt>(C0->getAggregateElement((unsigned)0)) 928 : nullptr; 929 ConstantInt *CI10 = 930 C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)0)) 931 : nullptr; 932 933 // Constant Fold - insert bottom Length bits starting at the Index'th bit. 934 if (CI00 && CI10) { 935 APInt V00 = CI00->getValue(); 936 APInt V10 = CI10->getValue(); 937 APInt Mask = APInt::getLowBitsSet(64, Length).shl(Index); 938 V00 = V00 & ~Mask; 939 V10 = V10.zextOrTrunc(Length).zextOrTrunc(64).shl(Index); 940 APInt Val = V00 | V10; 941 Type *IntTy64 = Type::getInt64Ty(II.getContext()); 942 Constant *Args[] = {ConstantInt::get(IntTy64, Val.getZExtValue()), 943 UndefValue::get(IntTy64)}; 944 return ConstantVector::get(Args); 945 } 946 947 // If we were an INSERTQ call, we'll save demanded elements if we convert to 948 // INSERTQI. 949 if (II.getIntrinsicID() == Intrinsic::x86_sse4a_insertq) { 950 Type *IntTy8 = Type::getInt8Ty(II.getContext()); 951 Constant *CILength = ConstantInt::get(IntTy8, Length, false); 952 Constant *CIIndex = ConstantInt::get(IntTy8, Index, false); 953 954 Value *Args[] = {Op0, Op1, CILength, CIIndex}; 955 Module *M = II.getModule(); 956 Value *F = Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi); 957 return Builder.CreateCall(F, Args); 958 } 959 960 return nullptr; 961 } 962 963 /// Attempt to convert pshufb* to shufflevector if the mask is constant. 964 static Value *simplifyX86pshufb(const IntrinsicInst &II, 965 InstCombiner::BuilderTy &Builder) { 966 Constant *V = dyn_cast<Constant>(II.getArgOperand(1)); 967 if (!V) 968 return nullptr; 969 970 auto *VecTy = cast<VectorType>(II.getType()); 971 auto *MaskEltTy = Type::getInt32Ty(II.getContext()); 972 unsigned NumElts = VecTy->getNumElements(); 973 assert((NumElts == 16 || NumElts == 32 || NumElts == 64) && 974 "Unexpected number of elements in shuffle mask!"); 975 976 // Construct a shuffle mask from constant integers or UNDEFs. 977 Constant *Indexes[64] = {nullptr}; 978 979 // Each byte in the shuffle control mask forms an index to permute the 980 // corresponding byte in the destination operand. 981 for (unsigned I = 0; I < NumElts; ++I) { 982 Constant *COp = V->getAggregateElement(I); 983 if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp))) 984 return nullptr; 985 986 if (isa<UndefValue>(COp)) { 987 Indexes[I] = UndefValue::get(MaskEltTy); 988 continue; 989 } 990 991 int8_t Index = cast<ConstantInt>(COp)->getValue().getZExtValue(); 992 993 // If the most significant bit (bit[7]) of each byte of the shuffle 994 // control mask is set, then zero is written in the result byte. 995 // The zero vector is in the right-hand side of the resulting 996 // shufflevector. 997 998 // The value of each index for the high 128-bit lane is the least 999 // significant 4 bits of the respective shuffle control byte. 1000 Index = ((Index < 0) ? NumElts : Index & 0x0F) + (I & 0xF0); 1001 Indexes[I] = ConstantInt::get(MaskEltTy, Index); 1002 } 1003 1004 auto ShuffleMask = ConstantVector::get(makeArrayRef(Indexes, NumElts)); 1005 auto V1 = II.getArgOperand(0); 1006 auto V2 = Constant::getNullValue(VecTy); 1007 return Builder.CreateShuffleVector(V1, V2, ShuffleMask); 1008 } 1009 1010 /// Attempt to convert vpermilvar* to shufflevector if the mask is constant. 1011 static Value *simplifyX86vpermilvar(const IntrinsicInst &II, 1012 InstCombiner::BuilderTy &Builder) { 1013 Constant *V = dyn_cast<Constant>(II.getArgOperand(1)); 1014 if (!V) 1015 return nullptr; 1016 1017 auto *VecTy = cast<VectorType>(II.getType()); 1018 auto *MaskEltTy = Type::getInt32Ty(II.getContext()); 1019 unsigned NumElts = VecTy->getVectorNumElements(); 1020 bool IsPD = VecTy->getScalarType()->isDoubleTy(); 1021 unsigned NumLaneElts = IsPD ? 2 : 4; 1022 assert(NumElts == 16 || NumElts == 8 || NumElts == 4 || NumElts == 2); 1023 1024 // Construct a shuffle mask from constant integers or UNDEFs. 1025 Constant *Indexes[16] = {nullptr}; 1026 1027 // The intrinsics only read one or two bits, clear the rest. 1028 for (unsigned I = 0; I < NumElts; ++I) { 1029 Constant *COp = V->getAggregateElement(I); 1030 if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp))) 1031 return nullptr; 1032 1033 if (isa<UndefValue>(COp)) { 1034 Indexes[I] = UndefValue::get(MaskEltTy); 1035 continue; 1036 } 1037 1038 APInt Index = cast<ConstantInt>(COp)->getValue(); 1039 Index = Index.zextOrTrunc(32).getLoBits(2); 1040 1041 // The PD variants uses bit 1 to select per-lane element index, so 1042 // shift down to convert to generic shuffle mask index. 1043 if (IsPD) 1044 Index.lshrInPlace(1); 1045 1046 // The _256 variants are a bit trickier since the mask bits always index 1047 // into the corresponding 128 half. In order to convert to a generic 1048 // shuffle, we have to make that explicit. 1049 Index += APInt(32, (I / NumLaneElts) * NumLaneElts); 1050 1051 Indexes[I] = ConstantInt::get(MaskEltTy, Index); 1052 } 1053 1054 auto ShuffleMask = ConstantVector::get(makeArrayRef(Indexes, NumElts)); 1055 auto V1 = II.getArgOperand(0); 1056 auto V2 = UndefValue::get(V1->getType()); 1057 return Builder.CreateShuffleVector(V1, V2, ShuffleMask); 1058 } 1059 1060 /// Attempt to convert vpermd/vpermps to shufflevector if the mask is constant. 1061 static Value *simplifyX86vpermv(const IntrinsicInst &II, 1062 InstCombiner::BuilderTy &Builder) { 1063 auto *V = dyn_cast<Constant>(II.getArgOperand(1)); 1064 if (!V) 1065 return nullptr; 1066 1067 auto *VecTy = cast<VectorType>(II.getType()); 1068 auto *MaskEltTy = Type::getInt32Ty(II.getContext()); 1069 unsigned Size = VecTy->getNumElements(); 1070 assert((Size == 4 || Size == 8 || Size == 16 || Size == 32 || Size == 64) && 1071 "Unexpected shuffle mask size"); 1072 1073 // Construct a shuffle mask from constant integers or UNDEFs. 1074 Constant *Indexes[64] = {nullptr}; 1075 1076 for (unsigned I = 0; I < Size; ++I) { 1077 Constant *COp = V->getAggregateElement(I); 1078 if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp))) 1079 return nullptr; 1080 1081 if (isa<UndefValue>(COp)) { 1082 Indexes[I] = UndefValue::get(MaskEltTy); 1083 continue; 1084 } 1085 1086 uint32_t Index = cast<ConstantInt>(COp)->getZExtValue(); 1087 Index &= Size - 1; 1088 Indexes[I] = ConstantInt::get(MaskEltTy, Index); 1089 } 1090 1091 auto ShuffleMask = ConstantVector::get(makeArrayRef(Indexes, Size)); 1092 auto V1 = II.getArgOperand(0); 1093 auto V2 = UndefValue::get(VecTy); 1094 return Builder.CreateShuffleVector(V1, V2, ShuffleMask); 1095 } 1096 1097 /// Decode XOP integer vector comparison intrinsics. 1098 static Value *simplifyX86vpcom(const IntrinsicInst &II, 1099 InstCombiner::BuilderTy &Builder, 1100 bool IsSigned) { 1101 if (auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2))) { 1102 uint64_t Imm = CInt->getZExtValue() & 0x7; 1103 VectorType *VecTy = cast<VectorType>(II.getType()); 1104 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE; 1105 1106 switch (Imm) { 1107 case 0x0: 1108 Pred = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; 1109 break; 1110 case 0x1: 1111 Pred = IsSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; 1112 break; 1113 case 0x2: 1114 Pred = IsSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; 1115 break; 1116 case 0x3: 1117 Pred = IsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; 1118 break; 1119 case 0x4: 1120 Pred = ICmpInst::ICMP_EQ; break; 1121 case 0x5: 1122 Pred = ICmpInst::ICMP_NE; break; 1123 case 0x6: 1124 return ConstantInt::getSigned(VecTy, 0); // FALSE 1125 case 0x7: 1126 return ConstantInt::getSigned(VecTy, -1); // TRUE 1127 } 1128 1129 if (Value *Cmp = Builder.CreateICmp(Pred, II.getArgOperand(0), 1130 II.getArgOperand(1))) 1131 return Builder.CreateSExtOrTrunc(Cmp, VecTy); 1132 } 1133 return nullptr; 1134 } 1135 1136 static Value *simplifyMinnumMaxnum(const IntrinsicInst &II) { 1137 Value *Arg0 = II.getArgOperand(0); 1138 Value *Arg1 = II.getArgOperand(1); 1139 1140 // fmin(x, x) -> x 1141 if (Arg0 == Arg1) 1142 return Arg0; 1143 1144 const auto *C1 = dyn_cast<ConstantFP>(Arg1); 1145 1146 // fmin(x, nan) -> x 1147 if (C1 && C1->isNaN()) 1148 return Arg0; 1149 1150 // This is the value because if undef were NaN, we would return the other 1151 // value and cannot return a NaN unless both operands are. 1152 // 1153 // fmin(undef, x) -> x 1154 if (isa<UndefValue>(Arg0)) 1155 return Arg1; 1156 1157 // fmin(x, undef) -> x 1158 if (isa<UndefValue>(Arg1)) 1159 return Arg0; 1160 1161 Value *X = nullptr; 1162 Value *Y = nullptr; 1163 if (II.getIntrinsicID() == Intrinsic::minnum) { 1164 // fmin(x, fmin(x, y)) -> fmin(x, y) 1165 // fmin(y, fmin(x, y)) -> fmin(x, y) 1166 if (match(Arg1, m_FMin(m_Value(X), m_Value(Y)))) { 1167 if (Arg0 == X || Arg0 == Y) 1168 return Arg1; 1169 } 1170 1171 // fmin(fmin(x, y), x) -> fmin(x, y) 1172 // fmin(fmin(x, y), y) -> fmin(x, y) 1173 if (match(Arg0, m_FMin(m_Value(X), m_Value(Y)))) { 1174 if (Arg1 == X || Arg1 == Y) 1175 return Arg0; 1176 } 1177 1178 // TODO: fmin(nnan x, inf) -> x 1179 // TODO: fmin(nnan ninf x, flt_max) -> x 1180 if (C1 && C1->isInfinity()) { 1181 // fmin(x, -inf) -> -inf 1182 if (C1->isNegative()) 1183 return Arg1; 1184 } 1185 } else { 1186 assert(II.getIntrinsicID() == Intrinsic::maxnum); 1187 // fmax(x, fmax(x, y)) -> fmax(x, y) 1188 // fmax(y, fmax(x, y)) -> fmax(x, y) 1189 if (match(Arg1, m_FMax(m_Value(X), m_Value(Y)))) { 1190 if (Arg0 == X || Arg0 == Y) 1191 return Arg1; 1192 } 1193 1194 // fmax(fmax(x, y), x) -> fmax(x, y) 1195 // fmax(fmax(x, y), y) -> fmax(x, y) 1196 if (match(Arg0, m_FMax(m_Value(X), m_Value(Y)))) { 1197 if (Arg1 == X || Arg1 == Y) 1198 return Arg0; 1199 } 1200 1201 // TODO: fmax(nnan x, -inf) -> x 1202 // TODO: fmax(nnan ninf x, -flt_max) -> x 1203 if (C1 && C1->isInfinity()) { 1204 // fmax(x, inf) -> inf 1205 if (!C1->isNegative()) 1206 return Arg1; 1207 } 1208 } 1209 return nullptr; 1210 } 1211 1212 static bool maskIsAllOneOrUndef(Value *Mask) { 1213 auto *ConstMask = dyn_cast<Constant>(Mask); 1214 if (!ConstMask) 1215 return false; 1216 if (ConstMask->isAllOnesValue() || isa<UndefValue>(ConstMask)) 1217 return true; 1218 for (unsigned I = 0, E = ConstMask->getType()->getVectorNumElements(); I != E; 1219 ++I) { 1220 if (auto *MaskElt = ConstMask->getAggregateElement(I)) 1221 if (MaskElt->isAllOnesValue() || isa<UndefValue>(MaskElt)) 1222 continue; 1223 return false; 1224 } 1225 return true; 1226 } 1227 1228 static Value *simplifyMaskedLoad(const IntrinsicInst &II, 1229 InstCombiner::BuilderTy &Builder) { 1230 // If the mask is all ones or undefs, this is a plain vector load of the 1st 1231 // argument. 1232 if (maskIsAllOneOrUndef(II.getArgOperand(2))) { 1233 Value *LoadPtr = II.getArgOperand(0); 1234 unsigned Alignment = cast<ConstantInt>(II.getArgOperand(1))->getZExtValue(); 1235 return Builder.CreateAlignedLoad(LoadPtr, Alignment, "unmaskedload"); 1236 } 1237 1238 return nullptr; 1239 } 1240 1241 static Instruction *simplifyMaskedStore(IntrinsicInst &II, InstCombiner &IC) { 1242 auto *ConstMask = dyn_cast<Constant>(II.getArgOperand(3)); 1243 if (!ConstMask) 1244 return nullptr; 1245 1246 // If the mask is all zeros, this instruction does nothing. 1247 if (ConstMask->isNullValue()) 1248 return IC.eraseInstFromFunction(II); 1249 1250 // If the mask is all ones, this is a plain vector store of the 1st argument. 1251 if (ConstMask->isAllOnesValue()) { 1252 Value *StorePtr = II.getArgOperand(1); 1253 unsigned Alignment = cast<ConstantInt>(II.getArgOperand(2))->getZExtValue(); 1254 return new StoreInst(II.getArgOperand(0), StorePtr, false, Alignment); 1255 } 1256 1257 return nullptr; 1258 } 1259 1260 static Instruction *simplifyMaskedGather(IntrinsicInst &II, InstCombiner &IC) { 1261 // If the mask is all zeros, return the "passthru" argument of the gather. 1262 auto *ConstMask = dyn_cast<Constant>(II.getArgOperand(2)); 1263 if (ConstMask && ConstMask->isNullValue()) 1264 return IC.replaceInstUsesWith(II, II.getArgOperand(3)); 1265 1266 return nullptr; 1267 } 1268 1269 /// This function transforms launder.invariant.group and strip.invariant.group 1270 /// like: 1271 /// launder(launder(%x)) -> launder(%x) (the result is not the argument) 1272 /// launder(strip(%x)) -> launder(%x) 1273 /// strip(strip(%x)) -> strip(%x) (the result is not the argument) 1274 /// strip(launder(%x)) -> strip(%x) 1275 /// This is legal because it preserves the most recent information about 1276 /// the presence or absence of invariant.group. 1277 static Instruction *simplifyInvariantGroupIntrinsic(IntrinsicInst &II, 1278 InstCombiner &IC) { 1279 auto *Arg = II.getArgOperand(0); 1280 auto *StrippedArg = Arg->stripPointerCasts(); 1281 auto *StrippedInvariantGroupsArg = Arg->stripPointerCastsAndInvariantGroups(); 1282 if (StrippedArg == StrippedInvariantGroupsArg) 1283 return nullptr; // No launders/strips to remove. 1284 1285 Value *Result = nullptr; 1286 1287 if (II.getIntrinsicID() == Intrinsic::launder_invariant_group) 1288 Result = IC.Builder.CreateLaunderInvariantGroup(StrippedInvariantGroupsArg); 1289 else if (II.getIntrinsicID() == Intrinsic::strip_invariant_group) 1290 Result = IC.Builder.CreateStripInvariantGroup(StrippedInvariantGroupsArg); 1291 else 1292 llvm_unreachable( 1293 "simplifyInvariantGroupIntrinsic only handles launder and strip"); 1294 if (Result->getType()->getPointerAddressSpace() != 1295 II.getType()->getPointerAddressSpace()) 1296 Result = IC.Builder.CreateAddrSpaceCast(Result, II.getType()); 1297 if (Result->getType() != II.getType()) 1298 Result = IC.Builder.CreateBitCast(Result, II.getType()); 1299 1300 return cast<Instruction>(Result); 1301 } 1302 1303 static Instruction *simplifyMaskedScatter(IntrinsicInst &II, InstCombiner &IC) { 1304 // If the mask is all zeros, a scatter does nothing. 1305 auto *ConstMask = dyn_cast<Constant>(II.getArgOperand(3)); 1306 if (ConstMask && ConstMask->isNullValue()) 1307 return IC.eraseInstFromFunction(II); 1308 1309 return nullptr; 1310 } 1311 1312 static Instruction *foldCttzCtlz(IntrinsicInst &II, InstCombiner &IC) { 1313 assert((II.getIntrinsicID() == Intrinsic::cttz || 1314 II.getIntrinsicID() == Intrinsic::ctlz) && 1315 "Expected cttz or ctlz intrinsic"); 1316 Value *Op0 = II.getArgOperand(0); 1317 1318 KnownBits Known = IC.computeKnownBits(Op0, 0, &II); 1319 1320 // Create a mask for bits above (ctlz) or below (cttz) the first known one. 1321 bool IsTZ = II.getIntrinsicID() == Intrinsic::cttz; 1322 unsigned PossibleZeros = IsTZ ? Known.countMaxTrailingZeros() 1323 : Known.countMaxLeadingZeros(); 1324 unsigned DefiniteZeros = IsTZ ? Known.countMinTrailingZeros() 1325 : Known.countMinLeadingZeros(); 1326 1327 // If all bits above (ctlz) or below (cttz) the first known one are known 1328 // zero, this value is constant. 1329 // FIXME: This should be in InstSimplify because we're replacing an 1330 // instruction with a constant. 1331 if (PossibleZeros == DefiniteZeros) { 1332 auto *C = ConstantInt::get(Op0->getType(), DefiniteZeros); 1333 return IC.replaceInstUsesWith(II, C); 1334 } 1335 1336 // If the input to cttz/ctlz is known to be non-zero, 1337 // then change the 'ZeroIsUndef' parameter to 'true' 1338 // because we know the zero behavior can't affect the result. 1339 if (!Known.One.isNullValue() || 1340 isKnownNonZero(Op0, IC.getDataLayout(), 0, &IC.getAssumptionCache(), &II, 1341 &IC.getDominatorTree())) { 1342 if (!match(II.getArgOperand(1), m_One())) { 1343 II.setOperand(1, IC.Builder.getTrue()); 1344 return &II; 1345 } 1346 } 1347 1348 // Add range metadata since known bits can't completely reflect what we know. 1349 // TODO: Handle splat vectors. 1350 auto *IT = dyn_cast<IntegerType>(Op0->getType()); 1351 if (IT && IT->getBitWidth() != 1 && !II.getMetadata(LLVMContext::MD_range)) { 1352 Metadata *LowAndHigh[] = { 1353 ConstantAsMetadata::get(ConstantInt::get(IT, DefiniteZeros)), 1354 ConstantAsMetadata::get(ConstantInt::get(IT, PossibleZeros + 1))}; 1355 II.setMetadata(LLVMContext::MD_range, 1356 MDNode::get(II.getContext(), LowAndHigh)); 1357 return &II; 1358 } 1359 1360 return nullptr; 1361 } 1362 1363 static Instruction *foldCtpop(IntrinsicInst &II, InstCombiner &IC) { 1364 assert(II.getIntrinsicID() == Intrinsic::ctpop && 1365 "Expected ctpop intrinsic"); 1366 Value *Op0 = II.getArgOperand(0); 1367 // FIXME: Try to simplify vectors of integers. 1368 auto *IT = dyn_cast<IntegerType>(Op0->getType()); 1369 if (!IT) 1370 return nullptr; 1371 1372 unsigned BitWidth = IT->getBitWidth(); 1373 KnownBits Known(BitWidth); 1374 IC.computeKnownBits(Op0, Known, 0, &II); 1375 1376 unsigned MinCount = Known.countMinPopulation(); 1377 unsigned MaxCount = Known.countMaxPopulation(); 1378 1379 // Add range metadata since known bits can't completely reflect what we know. 1380 if (IT->getBitWidth() != 1 && !II.getMetadata(LLVMContext::MD_range)) { 1381 Metadata *LowAndHigh[] = { 1382 ConstantAsMetadata::get(ConstantInt::get(IT, MinCount)), 1383 ConstantAsMetadata::get(ConstantInt::get(IT, MaxCount + 1))}; 1384 II.setMetadata(LLVMContext::MD_range, 1385 MDNode::get(II.getContext(), LowAndHigh)); 1386 return &II; 1387 } 1388 1389 return nullptr; 1390 } 1391 1392 // TODO: If the x86 backend knew how to convert a bool vector mask back to an 1393 // XMM register mask efficiently, we could transform all x86 masked intrinsics 1394 // to LLVM masked intrinsics and remove the x86 masked intrinsic defs. 1395 static Instruction *simplifyX86MaskedLoad(IntrinsicInst &II, InstCombiner &IC) { 1396 Value *Ptr = II.getOperand(0); 1397 Value *Mask = II.getOperand(1); 1398 Constant *ZeroVec = Constant::getNullValue(II.getType()); 1399 1400 // Special case a zero mask since that's not a ConstantDataVector. 1401 // This masked load instruction creates a zero vector. 1402 if (isa<ConstantAggregateZero>(Mask)) 1403 return IC.replaceInstUsesWith(II, ZeroVec); 1404 1405 auto *ConstMask = dyn_cast<ConstantDataVector>(Mask); 1406 if (!ConstMask) 1407 return nullptr; 1408 1409 // The mask is constant. Convert this x86 intrinsic to the LLVM instrinsic 1410 // to allow target-independent optimizations. 1411 1412 // First, cast the x86 intrinsic scalar pointer to a vector pointer to match 1413 // the LLVM intrinsic definition for the pointer argument. 1414 unsigned AddrSpace = cast<PointerType>(Ptr->getType())->getAddressSpace(); 1415 PointerType *VecPtrTy = PointerType::get(II.getType(), AddrSpace); 1416 Value *PtrCast = IC.Builder.CreateBitCast(Ptr, VecPtrTy, "castvec"); 1417 1418 // Second, convert the x86 XMM integer vector mask to a vector of bools based 1419 // on each element's most significant bit (the sign bit). 1420 Constant *BoolMask = getNegativeIsTrueBoolVec(ConstMask); 1421 1422 // The pass-through vector for an x86 masked load is a zero vector. 1423 CallInst *NewMaskedLoad = 1424 IC.Builder.CreateMaskedLoad(PtrCast, 1, BoolMask, ZeroVec); 1425 return IC.replaceInstUsesWith(II, NewMaskedLoad); 1426 } 1427 1428 // TODO: If the x86 backend knew how to convert a bool vector mask back to an 1429 // XMM register mask efficiently, we could transform all x86 masked intrinsics 1430 // to LLVM masked intrinsics and remove the x86 masked intrinsic defs. 1431 static bool simplifyX86MaskedStore(IntrinsicInst &II, InstCombiner &IC) { 1432 Value *Ptr = II.getOperand(0); 1433 Value *Mask = II.getOperand(1); 1434 Value *Vec = II.getOperand(2); 1435 1436 // Special case a zero mask since that's not a ConstantDataVector: 1437 // this masked store instruction does nothing. 1438 if (isa<ConstantAggregateZero>(Mask)) { 1439 IC.eraseInstFromFunction(II); 1440 return true; 1441 } 1442 1443 // The SSE2 version is too weird (eg, unaligned but non-temporal) to do 1444 // anything else at this level. 1445 if (II.getIntrinsicID() == Intrinsic::x86_sse2_maskmov_dqu) 1446 return false; 1447 1448 auto *ConstMask = dyn_cast<ConstantDataVector>(Mask); 1449 if (!ConstMask) 1450 return false; 1451 1452 // The mask is constant. Convert this x86 intrinsic to the LLVM instrinsic 1453 // to allow target-independent optimizations. 1454 1455 // First, cast the x86 intrinsic scalar pointer to a vector pointer to match 1456 // the LLVM intrinsic definition for the pointer argument. 1457 unsigned AddrSpace = cast<PointerType>(Ptr->getType())->getAddressSpace(); 1458 PointerType *VecPtrTy = PointerType::get(Vec->getType(), AddrSpace); 1459 Value *PtrCast = IC.Builder.CreateBitCast(Ptr, VecPtrTy, "castvec"); 1460 1461 // Second, convert the x86 XMM integer vector mask to a vector of bools based 1462 // on each element's most significant bit (the sign bit). 1463 Constant *BoolMask = getNegativeIsTrueBoolVec(ConstMask); 1464 1465 IC.Builder.CreateMaskedStore(Vec, PtrCast, 1, BoolMask); 1466 1467 // 'Replace uses' doesn't work for stores. Erase the original masked store. 1468 IC.eraseInstFromFunction(II); 1469 return true; 1470 } 1471 1472 // Constant fold llvm.amdgcn.fmed3 intrinsics for standard inputs. 1473 // 1474 // A single NaN input is folded to minnum, so we rely on that folding for 1475 // handling NaNs. 1476 static APFloat fmed3AMDGCN(const APFloat &Src0, const APFloat &Src1, 1477 const APFloat &Src2) { 1478 APFloat Max3 = maxnum(maxnum(Src0, Src1), Src2); 1479 1480 APFloat::cmpResult Cmp0 = Max3.compare(Src0); 1481 assert(Cmp0 != APFloat::cmpUnordered && "nans handled separately"); 1482 if (Cmp0 == APFloat::cmpEqual) 1483 return maxnum(Src1, Src2); 1484 1485 APFloat::cmpResult Cmp1 = Max3.compare(Src1); 1486 assert(Cmp1 != APFloat::cmpUnordered && "nans handled separately"); 1487 if (Cmp1 == APFloat::cmpEqual) 1488 return maxnum(Src0, Src2); 1489 1490 return maxnum(Src0, Src1); 1491 } 1492 1493 /// Convert a table lookup to shufflevector if the mask is constant. 1494 /// This could benefit tbl1 if the mask is { 7,6,5,4,3,2,1,0 }, in 1495 /// which case we could lower the shufflevector with rev64 instructions 1496 /// as it's actually a byte reverse. 1497 static Value *simplifyNeonTbl1(const IntrinsicInst &II, 1498 InstCombiner::BuilderTy &Builder) { 1499 // Bail out if the mask is not a constant. 1500 auto *C = dyn_cast<Constant>(II.getArgOperand(1)); 1501 if (!C) 1502 return nullptr; 1503 1504 auto *VecTy = cast<VectorType>(II.getType()); 1505 unsigned NumElts = VecTy->getNumElements(); 1506 1507 // Only perform this transformation for <8 x i8> vector types. 1508 if (!VecTy->getElementType()->isIntegerTy(8) || NumElts != 8) 1509 return nullptr; 1510 1511 uint32_t Indexes[8]; 1512 1513 for (unsigned I = 0; I < NumElts; ++I) { 1514 Constant *COp = C->getAggregateElement(I); 1515 1516 if (!COp || !isa<ConstantInt>(COp)) 1517 return nullptr; 1518 1519 Indexes[I] = cast<ConstantInt>(COp)->getLimitedValue(); 1520 1521 // Make sure the mask indices are in range. 1522 if (Indexes[I] >= NumElts) 1523 return nullptr; 1524 } 1525 1526 auto *ShuffleMask = ConstantDataVector::get(II.getContext(), 1527 makeArrayRef(Indexes)); 1528 auto *V1 = II.getArgOperand(0); 1529 auto *V2 = Constant::getNullValue(V1->getType()); 1530 return Builder.CreateShuffleVector(V1, V2, ShuffleMask); 1531 } 1532 1533 /// Convert a vector load intrinsic into a simple llvm load instruction. 1534 /// This is beneficial when the underlying object being addressed comes 1535 /// from a constant, since we get constant-folding for free. 1536 static Value *simplifyNeonVld1(const IntrinsicInst &II, 1537 unsigned MemAlign, 1538 InstCombiner::BuilderTy &Builder) { 1539 auto *IntrAlign = dyn_cast<ConstantInt>(II.getArgOperand(1)); 1540 1541 if (!IntrAlign) 1542 return nullptr; 1543 1544 unsigned Alignment = IntrAlign->getLimitedValue() < MemAlign ? 1545 MemAlign : IntrAlign->getLimitedValue(); 1546 1547 if (!isPowerOf2_32(Alignment)) 1548 return nullptr; 1549 1550 auto *BCastInst = Builder.CreateBitCast(II.getArgOperand(0), 1551 PointerType::get(II.getType(), 0)); 1552 return Builder.CreateAlignedLoad(BCastInst, Alignment); 1553 } 1554 1555 // Returns true iff the 2 intrinsics have the same operands, limiting the 1556 // comparison to the first NumOperands. 1557 static bool haveSameOperands(const IntrinsicInst &I, const IntrinsicInst &E, 1558 unsigned NumOperands) { 1559 assert(I.getNumArgOperands() >= NumOperands && "Not enough operands"); 1560 assert(E.getNumArgOperands() >= NumOperands && "Not enough operands"); 1561 for (unsigned i = 0; i < NumOperands; i++) 1562 if (I.getArgOperand(i) != E.getArgOperand(i)) 1563 return false; 1564 return true; 1565 } 1566 1567 // Remove trivially empty start/end intrinsic ranges, i.e. a start 1568 // immediately followed by an end (ignoring debuginfo or other 1569 // start/end intrinsics in between). As this handles only the most trivial 1570 // cases, tracking the nesting level is not needed: 1571 // 1572 // call @llvm.foo.start(i1 0) ; &I 1573 // call @llvm.foo.start(i1 0) 1574 // call @llvm.foo.end(i1 0) ; This one will not be skipped: it will be removed 1575 // call @llvm.foo.end(i1 0) 1576 static bool removeTriviallyEmptyRange(IntrinsicInst &I, unsigned StartID, 1577 unsigned EndID, InstCombiner &IC) { 1578 assert(I.getIntrinsicID() == StartID && 1579 "Start intrinsic does not have expected ID"); 1580 BasicBlock::iterator BI(I), BE(I.getParent()->end()); 1581 for (++BI; BI != BE; ++BI) { 1582 if (auto *E = dyn_cast<IntrinsicInst>(BI)) { 1583 if (isa<DbgInfoIntrinsic>(E) || E->getIntrinsicID() == StartID) 1584 continue; 1585 if (E->getIntrinsicID() == EndID && 1586 haveSameOperands(I, *E, E->getNumArgOperands())) { 1587 IC.eraseInstFromFunction(*E); 1588 IC.eraseInstFromFunction(I); 1589 return true; 1590 } 1591 } 1592 break; 1593 } 1594 1595 return false; 1596 } 1597 1598 // Convert NVVM intrinsics to target-generic LLVM code where possible. 1599 static Instruction *SimplifyNVVMIntrinsic(IntrinsicInst *II, InstCombiner &IC) { 1600 // Each NVVM intrinsic we can simplify can be replaced with one of: 1601 // 1602 // * an LLVM intrinsic, 1603 // * an LLVM cast operation, 1604 // * an LLVM binary operation, or 1605 // * ad-hoc LLVM IR for the particular operation. 1606 1607 // Some transformations are only valid when the module's 1608 // flush-denormals-to-zero (ftz) setting is true/false, whereas other 1609 // transformations are valid regardless of the module's ftz setting. 1610 enum FtzRequirementTy { 1611 FTZ_Any, // Any ftz setting is ok. 1612 FTZ_MustBeOn, // Transformation is valid only if ftz is on. 1613 FTZ_MustBeOff, // Transformation is valid only if ftz is off. 1614 }; 1615 // Classes of NVVM intrinsics that can't be replaced one-to-one with a 1616 // target-generic intrinsic, cast op, or binary op but that we can nonetheless 1617 // simplify. 1618 enum SpecialCase { 1619 SPC_Reciprocal, 1620 }; 1621 1622 // SimplifyAction is a poor-man's variant (plus an additional flag) that 1623 // represents how to replace an NVVM intrinsic with target-generic LLVM IR. 1624 struct SimplifyAction { 1625 // Invariant: At most one of these Optionals has a value. 1626 Optional<Intrinsic::ID> IID; 1627 Optional<Instruction::CastOps> CastOp; 1628 Optional<Instruction::BinaryOps> BinaryOp; 1629 Optional<SpecialCase> Special; 1630 1631 FtzRequirementTy FtzRequirement = FTZ_Any; 1632 1633 SimplifyAction() = default; 1634 1635 SimplifyAction(Intrinsic::ID IID, FtzRequirementTy FtzReq) 1636 : IID(IID), FtzRequirement(FtzReq) {} 1637 1638 // Cast operations don't have anything to do with FTZ, so we skip that 1639 // argument. 1640 SimplifyAction(Instruction::CastOps CastOp) : CastOp(CastOp) {} 1641 1642 SimplifyAction(Instruction::BinaryOps BinaryOp, FtzRequirementTy FtzReq) 1643 : BinaryOp(BinaryOp), FtzRequirement(FtzReq) {} 1644 1645 SimplifyAction(SpecialCase Special, FtzRequirementTy FtzReq) 1646 : Special(Special), FtzRequirement(FtzReq) {} 1647 }; 1648 1649 // Try to generate a SimplifyAction describing how to replace our 1650 // IntrinsicInstr with target-generic LLVM IR. 1651 const SimplifyAction Action = [II]() -> SimplifyAction { 1652 switch (II->getIntrinsicID()) { 1653 // NVVM intrinsics that map directly to LLVM intrinsics. 1654 case Intrinsic::nvvm_ceil_d: 1655 return {Intrinsic::ceil, FTZ_Any}; 1656 case Intrinsic::nvvm_ceil_f: 1657 return {Intrinsic::ceil, FTZ_MustBeOff}; 1658 case Intrinsic::nvvm_ceil_ftz_f: 1659 return {Intrinsic::ceil, FTZ_MustBeOn}; 1660 case Intrinsic::nvvm_fabs_d: 1661 return {Intrinsic::fabs, FTZ_Any}; 1662 case Intrinsic::nvvm_fabs_f: 1663 return {Intrinsic::fabs, FTZ_MustBeOff}; 1664 case Intrinsic::nvvm_fabs_ftz_f: 1665 return {Intrinsic::fabs, FTZ_MustBeOn}; 1666 case Intrinsic::nvvm_floor_d: 1667 return {Intrinsic::floor, FTZ_Any}; 1668 case Intrinsic::nvvm_floor_f: 1669 return {Intrinsic::floor, FTZ_MustBeOff}; 1670 case Intrinsic::nvvm_floor_ftz_f: 1671 return {Intrinsic::floor, FTZ_MustBeOn}; 1672 case Intrinsic::nvvm_fma_rn_d: 1673 return {Intrinsic::fma, FTZ_Any}; 1674 case Intrinsic::nvvm_fma_rn_f: 1675 return {Intrinsic::fma, FTZ_MustBeOff}; 1676 case Intrinsic::nvvm_fma_rn_ftz_f: 1677 return {Intrinsic::fma, FTZ_MustBeOn}; 1678 case Intrinsic::nvvm_fmax_d: 1679 return {Intrinsic::maxnum, FTZ_Any}; 1680 case Intrinsic::nvvm_fmax_f: 1681 return {Intrinsic::maxnum, FTZ_MustBeOff}; 1682 case Intrinsic::nvvm_fmax_ftz_f: 1683 return {Intrinsic::maxnum, FTZ_MustBeOn}; 1684 case Intrinsic::nvvm_fmin_d: 1685 return {Intrinsic::minnum, FTZ_Any}; 1686 case Intrinsic::nvvm_fmin_f: 1687 return {Intrinsic::minnum, FTZ_MustBeOff}; 1688 case Intrinsic::nvvm_fmin_ftz_f: 1689 return {Intrinsic::minnum, FTZ_MustBeOn}; 1690 case Intrinsic::nvvm_round_d: 1691 return {Intrinsic::round, FTZ_Any}; 1692 case Intrinsic::nvvm_round_f: 1693 return {Intrinsic::round, FTZ_MustBeOff}; 1694 case Intrinsic::nvvm_round_ftz_f: 1695 return {Intrinsic::round, FTZ_MustBeOn}; 1696 case Intrinsic::nvvm_sqrt_rn_d: 1697 return {Intrinsic::sqrt, FTZ_Any}; 1698 case Intrinsic::nvvm_sqrt_f: 1699 // nvvm_sqrt_f is a special case. For most intrinsics, foo_ftz_f is the 1700 // ftz version, and foo_f is the non-ftz version. But nvvm_sqrt_f adopts 1701 // the ftz-ness of the surrounding code. sqrt_rn_f and sqrt_rn_ftz_f are 1702 // the versions with explicit ftz-ness. 1703 return {Intrinsic::sqrt, FTZ_Any}; 1704 case Intrinsic::nvvm_sqrt_rn_f: 1705 return {Intrinsic::sqrt, FTZ_MustBeOff}; 1706 case Intrinsic::nvvm_sqrt_rn_ftz_f: 1707 return {Intrinsic::sqrt, FTZ_MustBeOn}; 1708 case Intrinsic::nvvm_trunc_d: 1709 return {Intrinsic::trunc, FTZ_Any}; 1710 case Intrinsic::nvvm_trunc_f: 1711 return {Intrinsic::trunc, FTZ_MustBeOff}; 1712 case Intrinsic::nvvm_trunc_ftz_f: 1713 return {Intrinsic::trunc, FTZ_MustBeOn}; 1714 1715 // NVVM intrinsics that map to LLVM cast operations. 1716 // 1717 // Note that llvm's target-generic conversion operators correspond to the rz 1718 // (round to zero) versions of the nvvm conversion intrinsics, even though 1719 // most everything else here uses the rn (round to nearest even) nvvm ops. 1720 case Intrinsic::nvvm_d2i_rz: 1721 case Intrinsic::nvvm_f2i_rz: 1722 case Intrinsic::nvvm_d2ll_rz: 1723 case Intrinsic::nvvm_f2ll_rz: 1724 return {Instruction::FPToSI}; 1725 case Intrinsic::nvvm_d2ui_rz: 1726 case Intrinsic::nvvm_f2ui_rz: 1727 case Intrinsic::nvvm_d2ull_rz: 1728 case Intrinsic::nvvm_f2ull_rz: 1729 return {Instruction::FPToUI}; 1730 case Intrinsic::nvvm_i2d_rz: 1731 case Intrinsic::nvvm_i2f_rz: 1732 case Intrinsic::nvvm_ll2d_rz: 1733 case Intrinsic::nvvm_ll2f_rz: 1734 return {Instruction::SIToFP}; 1735 case Intrinsic::nvvm_ui2d_rz: 1736 case Intrinsic::nvvm_ui2f_rz: 1737 case Intrinsic::nvvm_ull2d_rz: 1738 case Intrinsic::nvvm_ull2f_rz: 1739 return {Instruction::UIToFP}; 1740 1741 // NVVM intrinsics that map to LLVM binary ops. 1742 case Intrinsic::nvvm_add_rn_d: 1743 return {Instruction::FAdd, FTZ_Any}; 1744 case Intrinsic::nvvm_add_rn_f: 1745 return {Instruction::FAdd, FTZ_MustBeOff}; 1746 case Intrinsic::nvvm_add_rn_ftz_f: 1747 return {Instruction::FAdd, FTZ_MustBeOn}; 1748 case Intrinsic::nvvm_mul_rn_d: 1749 return {Instruction::FMul, FTZ_Any}; 1750 case Intrinsic::nvvm_mul_rn_f: 1751 return {Instruction::FMul, FTZ_MustBeOff}; 1752 case Intrinsic::nvvm_mul_rn_ftz_f: 1753 return {Instruction::FMul, FTZ_MustBeOn}; 1754 case Intrinsic::nvvm_div_rn_d: 1755 return {Instruction::FDiv, FTZ_Any}; 1756 case Intrinsic::nvvm_div_rn_f: 1757 return {Instruction::FDiv, FTZ_MustBeOff}; 1758 case Intrinsic::nvvm_div_rn_ftz_f: 1759 return {Instruction::FDiv, FTZ_MustBeOn}; 1760 1761 // The remainder of cases are NVVM intrinsics that map to LLVM idioms, but 1762 // need special handling. 1763 // 1764 // We seem to be missing intrinsics for rcp.approx.{ftz.}f32, which is just 1765 // as well. 1766 case Intrinsic::nvvm_rcp_rn_d: 1767 return {SPC_Reciprocal, FTZ_Any}; 1768 case Intrinsic::nvvm_rcp_rn_f: 1769 return {SPC_Reciprocal, FTZ_MustBeOff}; 1770 case Intrinsic::nvvm_rcp_rn_ftz_f: 1771 return {SPC_Reciprocal, FTZ_MustBeOn}; 1772 1773 // We do not currently simplify intrinsics that give an approximate answer. 1774 // These include: 1775 // 1776 // - nvvm_cos_approx_{f,ftz_f} 1777 // - nvvm_ex2_approx_{d,f,ftz_f} 1778 // - nvvm_lg2_approx_{d,f,ftz_f} 1779 // - nvvm_sin_approx_{f,ftz_f} 1780 // - nvvm_sqrt_approx_{f,ftz_f} 1781 // - nvvm_rsqrt_approx_{d,f,ftz_f} 1782 // - nvvm_div_approx_{ftz_d,ftz_f,f} 1783 // - nvvm_rcp_approx_ftz_d 1784 // 1785 // Ideally we'd encode them as e.g. "fast call @llvm.cos", where "fast" 1786 // means that fastmath is enabled in the intrinsic. Unfortunately only 1787 // binary operators (currently) have a fastmath bit in SelectionDAG, so this 1788 // information gets lost and we can't select on it. 1789 // 1790 // TODO: div and rcp are lowered to a binary op, so these we could in theory 1791 // lower them to "fast fdiv". 1792 1793 default: 1794 return {}; 1795 } 1796 }(); 1797 1798 // If Action.FtzRequirementTy is not satisfied by the module's ftz state, we 1799 // can bail out now. (Notice that in the case that IID is not an NVVM 1800 // intrinsic, we don't have to look up any module metadata, as 1801 // FtzRequirementTy will be FTZ_Any.) 1802 if (Action.FtzRequirement != FTZ_Any) { 1803 bool FtzEnabled = 1804 II->getFunction()->getFnAttribute("nvptx-f32ftz").getValueAsString() == 1805 "true"; 1806 1807 if (FtzEnabled != (Action.FtzRequirement == FTZ_MustBeOn)) 1808 return nullptr; 1809 } 1810 1811 // Simplify to target-generic intrinsic. 1812 if (Action.IID) { 1813 SmallVector<Value *, 4> Args(II->arg_operands()); 1814 // All the target-generic intrinsics currently of interest to us have one 1815 // type argument, equal to that of the nvvm intrinsic's argument. 1816 Type *Tys[] = {II->getArgOperand(0)->getType()}; 1817 return CallInst::Create( 1818 Intrinsic::getDeclaration(II->getModule(), *Action.IID, Tys), Args); 1819 } 1820 1821 // Simplify to target-generic binary op. 1822 if (Action.BinaryOp) 1823 return BinaryOperator::Create(*Action.BinaryOp, II->getArgOperand(0), 1824 II->getArgOperand(1), II->getName()); 1825 1826 // Simplify to target-generic cast op. 1827 if (Action.CastOp) 1828 return CastInst::Create(*Action.CastOp, II->getArgOperand(0), II->getType(), 1829 II->getName()); 1830 1831 // All that's left are the special cases. 1832 if (!Action.Special) 1833 return nullptr; 1834 1835 switch (*Action.Special) { 1836 case SPC_Reciprocal: 1837 // Simplify reciprocal. 1838 return BinaryOperator::Create( 1839 Instruction::FDiv, ConstantFP::get(II->getArgOperand(0)->getType(), 1), 1840 II->getArgOperand(0), II->getName()); 1841 } 1842 llvm_unreachable("All SpecialCase enumerators should be handled in switch."); 1843 } 1844 1845 Instruction *InstCombiner::visitVAStartInst(VAStartInst &I) { 1846 removeTriviallyEmptyRange(I, Intrinsic::vastart, Intrinsic::vaend, *this); 1847 return nullptr; 1848 } 1849 1850 Instruction *InstCombiner::visitVACopyInst(VACopyInst &I) { 1851 removeTriviallyEmptyRange(I, Intrinsic::vacopy, Intrinsic::vaend, *this); 1852 return nullptr; 1853 } 1854 1855 /// CallInst simplification. This mostly only handles folding of intrinsic 1856 /// instructions. For normal calls, it allows visitCallSite to do the heavy 1857 /// lifting. 1858 Instruction *InstCombiner::visitCallInst(CallInst &CI) { 1859 if (Value *V = SimplifyCall(&CI, SQ.getWithInstruction(&CI))) 1860 return replaceInstUsesWith(CI, V); 1861 1862 if (isFreeCall(&CI, &TLI)) 1863 return visitFree(CI); 1864 1865 // If the caller function is nounwind, mark the call as nounwind, even if the 1866 // callee isn't. 1867 if (CI.getFunction()->doesNotThrow() && !CI.doesNotThrow()) { 1868 CI.setDoesNotThrow(); 1869 return &CI; 1870 } 1871 1872 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI); 1873 if (!II) return visitCallSite(&CI); 1874 1875 // Intrinsics cannot occur in an invoke, so handle them here instead of in 1876 // visitCallSite. 1877 if (auto *MI = dyn_cast<AnyMemIntrinsic>(II)) { 1878 bool Changed = false; 1879 1880 // memmove/cpy/set of zero bytes is a noop. 1881 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) { 1882 if (NumBytes->isNullValue()) 1883 return eraseInstFromFunction(CI); 1884 1885 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes)) 1886 if (CI->getZExtValue() == 1) { 1887 // Replace the instruction with just byte operations. We would 1888 // transform other cases to loads/stores, but we don't know if 1889 // alignment is sufficient. 1890 } 1891 } 1892 1893 // No other transformations apply to volatile transfers. 1894 if (auto *M = dyn_cast<MemIntrinsic>(MI)) 1895 if (M->isVolatile()) 1896 return nullptr; 1897 1898 // If we have a memmove and the source operation is a constant global, 1899 // then the source and dest pointers can't alias, so we can change this 1900 // into a call to memcpy. 1901 if (auto *MMI = dyn_cast<AnyMemMoveInst>(MI)) { 1902 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource())) 1903 if (GVSrc->isConstant()) { 1904 Module *M = CI.getModule(); 1905 Intrinsic::ID MemCpyID = 1906 isa<AtomicMemMoveInst>(MMI) 1907 ? Intrinsic::memcpy_element_unordered_atomic 1908 : Intrinsic::memcpy; 1909 Type *Tys[3] = { CI.getArgOperand(0)->getType(), 1910 CI.getArgOperand(1)->getType(), 1911 CI.getArgOperand(2)->getType() }; 1912 CI.setCalledFunction(Intrinsic::getDeclaration(M, MemCpyID, Tys)); 1913 Changed = true; 1914 } 1915 } 1916 1917 if (AnyMemTransferInst *MTI = dyn_cast<AnyMemTransferInst>(MI)) { 1918 // memmove(x,x,size) -> noop. 1919 if (MTI->getSource() == MTI->getDest()) 1920 return eraseInstFromFunction(CI); 1921 } 1922 1923 // If we can determine a pointer alignment that is bigger than currently 1924 // set, update the alignment. 1925 if (auto *MTI = dyn_cast<AnyMemTransferInst>(MI)) { 1926 if (Instruction *I = SimplifyAnyMemTransfer(MTI)) 1927 return I; 1928 } else if (auto *MSI = dyn_cast<AnyMemSetInst>(MI)) { 1929 if (Instruction *I = SimplifyAnyMemSet(MSI)) 1930 return I; 1931 } 1932 1933 if (Changed) return II; 1934 } 1935 1936 if (Instruction *I = SimplifyNVVMIntrinsic(II, *this)) 1937 return I; 1938 1939 auto SimplifyDemandedVectorEltsLow = [this](Value *Op, unsigned Width, 1940 unsigned DemandedWidth) { 1941 APInt UndefElts(Width, 0); 1942 APInt DemandedElts = APInt::getLowBitsSet(Width, DemandedWidth); 1943 return SimplifyDemandedVectorElts(Op, DemandedElts, UndefElts); 1944 }; 1945 1946 switch (II->getIntrinsicID()) { 1947 default: break; 1948 case Intrinsic::objectsize: 1949 if (ConstantInt *N = 1950 lowerObjectSizeCall(II, DL, &TLI, /*MustSucceed=*/false)) 1951 return replaceInstUsesWith(CI, N); 1952 return nullptr; 1953 case Intrinsic::bswap: { 1954 Value *IIOperand = II->getArgOperand(0); 1955 Value *X = nullptr; 1956 1957 // bswap(trunc(bswap(x))) -> trunc(lshr(x, c)) 1958 if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) { 1959 unsigned C = X->getType()->getPrimitiveSizeInBits() - 1960 IIOperand->getType()->getPrimitiveSizeInBits(); 1961 Value *CV = ConstantInt::get(X->getType(), C); 1962 Value *V = Builder.CreateLShr(X, CV); 1963 return new TruncInst(V, IIOperand->getType()); 1964 } 1965 break; 1966 } 1967 case Intrinsic::masked_load: 1968 if (Value *SimplifiedMaskedOp = simplifyMaskedLoad(*II, Builder)) 1969 return replaceInstUsesWith(CI, SimplifiedMaskedOp); 1970 break; 1971 case Intrinsic::masked_store: 1972 return simplifyMaskedStore(*II, *this); 1973 case Intrinsic::masked_gather: 1974 return simplifyMaskedGather(*II, *this); 1975 case Intrinsic::masked_scatter: 1976 return simplifyMaskedScatter(*II, *this); 1977 case Intrinsic::launder_invariant_group: 1978 case Intrinsic::strip_invariant_group: 1979 if (auto *SkippedBarrier = simplifyInvariantGroupIntrinsic(*II, *this)) 1980 return replaceInstUsesWith(*II, SkippedBarrier); 1981 break; 1982 case Intrinsic::powi: 1983 if (ConstantInt *Power = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 1984 // 0 and 1 are handled in instsimplify 1985 1986 // powi(x, -1) -> 1/x 1987 if (Power->isMinusOne()) 1988 return BinaryOperator::CreateFDiv(ConstantFP::get(CI.getType(), 1.0), 1989 II->getArgOperand(0)); 1990 // powi(x, 2) -> x*x 1991 if (Power->equalsInt(2)) 1992 return BinaryOperator::CreateFMul(II->getArgOperand(0), 1993 II->getArgOperand(0)); 1994 } 1995 break; 1996 1997 case Intrinsic::cttz: 1998 case Intrinsic::ctlz: 1999 if (auto *I = foldCttzCtlz(*II, *this)) 2000 return I; 2001 break; 2002 2003 case Intrinsic::ctpop: 2004 if (auto *I = foldCtpop(*II, *this)) 2005 return I; 2006 break; 2007 2008 case Intrinsic::uadd_with_overflow: 2009 case Intrinsic::sadd_with_overflow: 2010 case Intrinsic::umul_with_overflow: 2011 case Intrinsic::smul_with_overflow: 2012 if (isa<Constant>(II->getArgOperand(0)) && 2013 !isa<Constant>(II->getArgOperand(1))) { 2014 // Canonicalize constants into the RHS. 2015 Value *LHS = II->getArgOperand(0); 2016 II->setArgOperand(0, II->getArgOperand(1)); 2017 II->setArgOperand(1, LHS); 2018 return II; 2019 } 2020 LLVM_FALLTHROUGH; 2021 2022 case Intrinsic::usub_with_overflow: 2023 case Intrinsic::ssub_with_overflow: { 2024 OverflowCheckFlavor OCF = 2025 IntrinsicIDToOverflowCheckFlavor(II->getIntrinsicID()); 2026 assert(OCF != OCF_INVALID && "unexpected!"); 2027 2028 Value *OperationResult = nullptr; 2029 Constant *OverflowResult = nullptr; 2030 if (OptimizeOverflowCheck(OCF, II->getArgOperand(0), II->getArgOperand(1), 2031 *II, OperationResult, OverflowResult)) 2032 return CreateOverflowTuple(II, OperationResult, OverflowResult); 2033 2034 break; 2035 } 2036 2037 case Intrinsic::minnum: 2038 case Intrinsic::maxnum: { 2039 Value *Arg0 = II->getArgOperand(0); 2040 Value *Arg1 = II->getArgOperand(1); 2041 // Canonicalize constants to the RHS. 2042 if (isa<ConstantFP>(Arg0) && !isa<ConstantFP>(Arg1)) { 2043 II->setArgOperand(0, Arg1); 2044 II->setArgOperand(1, Arg0); 2045 return II; 2046 } 2047 2048 // FIXME: Simplifications should be in instsimplify. 2049 if (Value *V = simplifyMinnumMaxnum(*II)) 2050 return replaceInstUsesWith(*II, V); 2051 2052 Value *X, *Y; 2053 if (match(Arg0, m_FNeg(m_Value(X))) && match(Arg1, m_FNeg(m_Value(Y))) && 2054 (Arg0->hasOneUse() || Arg1->hasOneUse())) { 2055 // If both operands are negated, invert the call and negate the result: 2056 // minnum(-X, -Y) --> -(maxnum(X, Y)) 2057 // maxnum(-X, -Y) --> -(minnum(X, Y)) 2058 Intrinsic::ID NewIID = II->getIntrinsicID() == Intrinsic::maxnum ? 2059 Intrinsic::minnum : Intrinsic::maxnum; 2060 Value *NewCall = Builder.CreateIntrinsic(NewIID, { X, Y }, II); 2061 Instruction *FNeg = BinaryOperator::CreateFNeg(NewCall); 2062 FNeg->copyIRFlags(II); 2063 return FNeg; 2064 } 2065 break; 2066 } 2067 case Intrinsic::fmuladd: { 2068 // Canonicalize fast fmuladd to the separate fmul + fadd. 2069 if (II->isFast()) { 2070 BuilderTy::FastMathFlagGuard Guard(Builder); 2071 Builder.setFastMathFlags(II->getFastMathFlags()); 2072 Value *Mul = Builder.CreateFMul(II->getArgOperand(0), 2073 II->getArgOperand(1)); 2074 Value *Add = Builder.CreateFAdd(Mul, II->getArgOperand(2)); 2075 Add->takeName(II); 2076 return replaceInstUsesWith(*II, Add); 2077 } 2078 2079 LLVM_FALLTHROUGH; 2080 } 2081 case Intrinsic::fma: { 2082 Value *Src0 = II->getArgOperand(0); 2083 Value *Src1 = II->getArgOperand(1); 2084 2085 // Canonicalize constant multiply operand to Src1. 2086 if (isa<Constant>(Src0) && !isa<Constant>(Src1)) { 2087 II->setArgOperand(0, Src1); 2088 II->setArgOperand(1, Src0); 2089 std::swap(Src0, Src1); 2090 } 2091 2092 // fma fneg(x), fneg(y), z -> fma x, y, z 2093 Value *X, *Y; 2094 if (match(Src0, m_FNeg(m_Value(X))) && match(Src1, m_FNeg(m_Value(Y)))) { 2095 II->setArgOperand(0, X); 2096 II->setArgOperand(1, Y); 2097 return II; 2098 } 2099 2100 // fma fabs(x), fabs(x), z -> fma x, x, z 2101 if (match(Src0, m_FAbs(m_Value(X))) && 2102 match(Src1, m_FAbs(m_Specific(X)))) { 2103 II->setArgOperand(0, X); 2104 II->setArgOperand(1, X); 2105 return II; 2106 } 2107 2108 // fma x, 1, z -> fadd x, z 2109 if (match(Src1, m_FPOne())) { 2110 auto *FAdd = BinaryOperator::CreateFAdd(Src0, II->getArgOperand(2)); 2111 FAdd->copyFastMathFlags(II); 2112 return FAdd; 2113 } 2114 2115 break; 2116 } 2117 case Intrinsic::fabs: { 2118 Value *Cond; 2119 Constant *LHS, *RHS; 2120 if (match(II->getArgOperand(0), 2121 m_Select(m_Value(Cond), m_Constant(LHS), m_Constant(RHS)))) { 2122 CallInst *Call0 = Builder.CreateCall(II->getCalledFunction(), {LHS}); 2123 CallInst *Call1 = Builder.CreateCall(II->getCalledFunction(), {RHS}); 2124 return SelectInst::Create(Cond, Call0, Call1); 2125 } 2126 2127 LLVM_FALLTHROUGH; 2128 } 2129 case Intrinsic::ceil: 2130 case Intrinsic::floor: 2131 case Intrinsic::round: 2132 case Intrinsic::nearbyint: 2133 case Intrinsic::rint: 2134 case Intrinsic::trunc: { 2135 Value *ExtSrc; 2136 if (match(II->getArgOperand(0), m_OneUse(m_FPExt(m_Value(ExtSrc))))) { 2137 // Narrow the call: intrinsic (fpext x) -> fpext (intrinsic x) 2138 Value *NarrowII = Builder.CreateIntrinsic(II->getIntrinsicID(), 2139 { ExtSrc }, II); 2140 return new FPExtInst(NarrowII, II->getType()); 2141 } 2142 break; 2143 } 2144 case Intrinsic::cos: 2145 case Intrinsic::amdgcn_cos: { 2146 Value *SrcSrc; 2147 Value *Src = II->getArgOperand(0); 2148 if (match(Src, m_FNeg(m_Value(SrcSrc))) || 2149 match(Src, m_FAbs(m_Value(SrcSrc)))) { 2150 // cos(-x) -> cos(x) 2151 // cos(fabs(x)) -> cos(x) 2152 II->setArgOperand(0, SrcSrc); 2153 return II; 2154 } 2155 2156 break; 2157 } 2158 case Intrinsic::ppc_altivec_lvx: 2159 case Intrinsic::ppc_altivec_lvxl: 2160 // Turn PPC lvx -> load if the pointer is known aligned. 2161 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, &AC, 2162 &DT) >= 16) { 2163 Value *Ptr = Builder.CreateBitCast(II->getArgOperand(0), 2164 PointerType::getUnqual(II->getType())); 2165 return new LoadInst(Ptr); 2166 } 2167 break; 2168 case Intrinsic::ppc_vsx_lxvw4x: 2169 case Intrinsic::ppc_vsx_lxvd2x: { 2170 // Turn PPC VSX loads into normal loads. 2171 Value *Ptr = Builder.CreateBitCast(II->getArgOperand(0), 2172 PointerType::getUnqual(II->getType())); 2173 return new LoadInst(Ptr, Twine(""), false, 1); 2174 } 2175 case Intrinsic::ppc_altivec_stvx: 2176 case Intrinsic::ppc_altivec_stvxl: 2177 // Turn stvx -> store if the pointer is known aligned. 2178 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, &AC, 2179 &DT) >= 16) { 2180 Type *OpPtrTy = 2181 PointerType::getUnqual(II->getArgOperand(0)->getType()); 2182 Value *Ptr = Builder.CreateBitCast(II->getArgOperand(1), OpPtrTy); 2183 return new StoreInst(II->getArgOperand(0), Ptr); 2184 } 2185 break; 2186 case Intrinsic::ppc_vsx_stxvw4x: 2187 case Intrinsic::ppc_vsx_stxvd2x: { 2188 // Turn PPC VSX stores into normal stores. 2189 Type *OpPtrTy = PointerType::getUnqual(II->getArgOperand(0)->getType()); 2190 Value *Ptr = Builder.CreateBitCast(II->getArgOperand(1), OpPtrTy); 2191 return new StoreInst(II->getArgOperand(0), Ptr, false, 1); 2192 } 2193 case Intrinsic::ppc_qpx_qvlfs: 2194 // Turn PPC QPX qvlfs -> load if the pointer is known aligned. 2195 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 16, DL, II, &AC, 2196 &DT) >= 16) { 2197 Type *VTy = VectorType::get(Builder.getFloatTy(), 2198 II->getType()->getVectorNumElements()); 2199 Value *Ptr = Builder.CreateBitCast(II->getArgOperand(0), 2200 PointerType::getUnqual(VTy)); 2201 Value *Load = Builder.CreateLoad(Ptr); 2202 return new FPExtInst(Load, II->getType()); 2203 } 2204 break; 2205 case Intrinsic::ppc_qpx_qvlfd: 2206 // Turn PPC QPX qvlfd -> load if the pointer is known aligned. 2207 if (getOrEnforceKnownAlignment(II->getArgOperand(0), 32, DL, II, &AC, 2208 &DT) >= 32) { 2209 Value *Ptr = Builder.CreateBitCast(II->getArgOperand(0), 2210 PointerType::getUnqual(II->getType())); 2211 return new LoadInst(Ptr); 2212 } 2213 break; 2214 case Intrinsic::ppc_qpx_qvstfs: 2215 // Turn PPC QPX qvstfs -> store if the pointer is known aligned. 2216 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 16, DL, II, &AC, 2217 &DT) >= 16) { 2218 Type *VTy = VectorType::get(Builder.getFloatTy(), 2219 II->getArgOperand(0)->getType()->getVectorNumElements()); 2220 Value *TOp = Builder.CreateFPTrunc(II->getArgOperand(0), VTy); 2221 Type *OpPtrTy = PointerType::getUnqual(VTy); 2222 Value *Ptr = Builder.CreateBitCast(II->getArgOperand(1), OpPtrTy); 2223 return new StoreInst(TOp, Ptr); 2224 } 2225 break; 2226 case Intrinsic::ppc_qpx_qvstfd: 2227 // Turn PPC QPX qvstfd -> store if the pointer is known aligned. 2228 if (getOrEnforceKnownAlignment(II->getArgOperand(1), 32, DL, II, &AC, 2229 &DT) >= 32) { 2230 Type *OpPtrTy = 2231 PointerType::getUnqual(II->getArgOperand(0)->getType()); 2232 Value *Ptr = Builder.CreateBitCast(II->getArgOperand(1), OpPtrTy); 2233 return new StoreInst(II->getArgOperand(0), Ptr); 2234 } 2235 break; 2236 2237 case Intrinsic::x86_bmi_bextr_32: 2238 case Intrinsic::x86_bmi_bextr_64: 2239 case Intrinsic::x86_tbm_bextri_u32: 2240 case Intrinsic::x86_tbm_bextri_u64: 2241 // If the RHS is a constant we can try some simplifications. 2242 if (auto *C = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 2243 uint64_t Shift = C->getZExtValue(); 2244 uint64_t Length = (Shift >> 8) & 0xff; 2245 Shift &= 0xff; 2246 unsigned BitWidth = II->getType()->getIntegerBitWidth(); 2247 // If the length is 0 or the shift is out of range, replace with zero. 2248 if (Length == 0 || Shift >= BitWidth) 2249 return replaceInstUsesWith(CI, ConstantInt::get(II->getType(), 0)); 2250 // If the LHS is also a constant, we can completely constant fold this. 2251 if (auto *InC = dyn_cast<ConstantInt>(II->getArgOperand(0))) { 2252 uint64_t Result = InC->getZExtValue() >> Shift; 2253 if (Length > BitWidth) 2254 Length = BitWidth; 2255 Result &= maskTrailingOnes<uint64_t>(Length); 2256 return replaceInstUsesWith(CI, ConstantInt::get(II->getType(), Result)); 2257 } 2258 // TODO should we turn this into 'and' if shift is 0? Or 'shl' if we 2259 // are only masking bits that a shift already cleared? 2260 } 2261 break; 2262 2263 case Intrinsic::x86_bmi_bzhi_32: 2264 case Intrinsic::x86_bmi_bzhi_64: 2265 // If the RHS is a constant we can try some simplifications. 2266 if (auto *C = dyn_cast<ConstantInt>(II->getArgOperand(1))) { 2267 uint64_t Index = C->getZExtValue() & 0xff; 2268 unsigned BitWidth = II->getType()->getIntegerBitWidth(); 2269 if (Index >= BitWidth) 2270 return replaceInstUsesWith(CI, II->getArgOperand(0)); 2271 if (Index == 0) 2272 return replaceInstUsesWith(CI, ConstantInt::get(II->getType(), 0)); 2273 // If the LHS is also a constant, we can completely constant fold this. 2274 if (auto *InC = dyn_cast<ConstantInt>(II->getArgOperand(0))) { 2275 uint64_t Result = InC->getZExtValue(); 2276 Result &= maskTrailingOnes<uint64_t>(Index); 2277 return replaceInstUsesWith(CI, ConstantInt::get(II->getType(), Result)); 2278 } 2279 // TODO should we convert this to an AND if the RHS is constant? 2280 } 2281 break; 2282 2283 case Intrinsic::x86_vcvtph2ps_128: 2284 case Intrinsic::x86_vcvtph2ps_256: { 2285 auto Arg = II->getArgOperand(0); 2286 auto ArgType = cast<VectorType>(Arg->getType()); 2287 auto RetType = cast<VectorType>(II->getType()); 2288 unsigned ArgWidth = ArgType->getNumElements(); 2289 unsigned RetWidth = RetType->getNumElements(); 2290 assert(RetWidth <= ArgWidth && "Unexpected input/return vector widths"); 2291 assert(ArgType->isIntOrIntVectorTy() && 2292 ArgType->getScalarSizeInBits() == 16 && 2293 "CVTPH2PS input type should be 16-bit integer vector"); 2294 assert(RetType->getScalarType()->isFloatTy() && 2295 "CVTPH2PS output type should be 32-bit float vector"); 2296 2297 // Constant folding: Convert to generic half to single conversion. 2298 if (isa<ConstantAggregateZero>(Arg)) 2299 return replaceInstUsesWith(*II, ConstantAggregateZero::get(RetType)); 2300 2301 if (isa<ConstantDataVector>(Arg)) { 2302 auto VectorHalfAsShorts = Arg; 2303 if (RetWidth < ArgWidth) { 2304 SmallVector<uint32_t, 8> SubVecMask; 2305 for (unsigned i = 0; i != RetWidth; ++i) 2306 SubVecMask.push_back((int)i); 2307 VectorHalfAsShorts = Builder.CreateShuffleVector( 2308 Arg, UndefValue::get(ArgType), SubVecMask); 2309 } 2310 2311 auto VectorHalfType = 2312 VectorType::get(Type::getHalfTy(II->getContext()), RetWidth); 2313 auto VectorHalfs = 2314 Builder.CreateBitCast(VectorHalfAsShorts, VectorHalfType); 2315 auto VectorFloats = Builder.CreateFPExt(VectorHalfs, RetType); 2316 return replaceInstUsesWith(*II, VectorFloats); 2317 } 2318 2319 // We only use the lowest lanes of the argument. 2320 if (Value *V = SimplifyDemandedVectorEltsLow(Arg, ArgWidth, RetWidth)) { 2321 II->setArgOperand(0, V); 2322 return II; 2323 } 2324 break; 2325 } 2326 2327 case Intrinsic::x86_sse_cvtss2si: 2328 case Intrinsic::x86_sse_cvtss2si64: 2329 case Intrinsic::x86_sse_cvttss2si: 2330 case Intrinsic::x86_sse_cvttss2si64: 2331 case Intrinsic::x86_sse2_cvtsd2si: 2332 case Intrinsic::x86_sse2_cvtsd2si64: 2333 case Intrinsic::x86_sse2_cvttsd2si: 2334 case Intrinsic::x86_sse2_cvttsd2si64: 2335 case Intrinsic::x86_avx512_vcvtss2si32: 2336 case Intrinsic::x86_avx512_vcvtss2si64: 2337 case Intrinsic::x86_avx512_vcvtss2usi32: 2338 case Intrinsic::x86_avx512_vcvtss2usi64: 2339 case Intrinsic::x86_avx512_vcvtsd2si32: 2340 case Intrinsic::x86_avx512_vcvtsd2si64: 2341 case Intrinsic::x86_avx512_vcvtsd2usi32: 2342 case Intrinsic::x86_avx512_vcvtsd2usi64: 2343 case Intrinsic::x86_avx512_cvttss2si: 2344 case Intrinsic::x86_avx512_cvttss2si64: 2345 case Intrinsic::x86_avx512_cvttss2usi: 2346 case Intrinsic::x86_avx512_cvttss2usi64: 2347 case Intrinsic::x86_avx512_cvttsd2si: 2348 case Intrinsic::x86_avx512_cvttsd2si64: 2349 case Intrinsic::x86_avx512_cvttsd2usi: 2350 case Intrinsic::x86_avx512_cvttsd2usi64: { 2351 // These intrinsics only demand the 0th element of their input vectors. If 2352 // we can simplify the input based on that, do so now. 2353 Value *Arg = II->getArgOperand(0); 2354 unsigned VWidth = Arg->getType()->getVectorNumElements(); 2355 if (Value *V = SimplifyDemandedVectorEltsLow(Arg, VWidth, 1)) { 2356 II->setArgOperand(0, V); 2357 return II; 2358 } 2359 break; 2360 } 2361 2362 case Intrinsic::x86_sse41_round_ps: 2363 case Intrinsic::x86_sse41_round_pd: 2364 case Intrinsic::x86_avx_round_ps_256: 2365 case Intrinsic::x86_avx_round_pd_256: 2366 case Intrinsic::x86_avx512_mask_rndscale_ps_128: 2367 case Intrinsic::x86_avx512_mask_rndscale_ps_256: 2368 case Intrinsic::x86_avx512_mask_rndscale_ps_512: 2369 case Intrinsic::x86_avx512_mask_rndscale_pd_128: 2370 case Intrinsic::x86_avx512_mask_rndscale_pd_256: 2371 case Intrinsic::x86_avx512_mask_rndscale_pd_512: 2372 case Intrinsic::x86_avx512_mask_rndscale_ss: 2373 case Intrinsic::x86_avx512_mask_rndscale_sd: 2374 if (Value *V = simplifyX86round(*II, Builder)) 2375 return replaceInstUsesWith(*II, V); 2376 break; 2377 2378 case Intrinsic::x86_mmx_pmovmskb: 2379 case Intrinsic::x86_sse_movmsk_ps: 2380 case Intrinsic::x86_sse2_movmsk_pd: 2381 case Intrinsic::x86_sse2_pmovmskb_128: 2382 case Intrinsic::x86_avx_movmsk_pd_256: 2383 case Intrinsic::x86_avx_movmsk_ps_256: 2384 case Intrinsic::x86_avx2_pmovmskb: 2385 if (Value *V = simplifyX86movmsk(*II)) 2386 return replaceInstUsesWith(*II, V); 2387 break; 2388 2389 case Intrinsic::x86_sse_comieq_ss: 2390 case Intrinsic::x86_sse_comige_ss: 2391 case Intrinsic::x86_sse_comigt_ss: 2392 case Intrinsic::x86_sse_comile_ss: 2393 case Intrinsic::x86_sse_comilt_ss: 2394 case Intrinsic::x86_sse_comineq_ss: 2395 case Intrinsic::x86_sse_ucomieq_ss: 2396 case Intrinsic::x86_sse_ucomige_ss: 2397 case Intrinsic::x86_sse_ucomigt_ss: 2398 case Intrinsic::x86_sse_ucomile_ss: 2399 case Intrinsic::x86_sse_ucomilt_ss: 2400 case Intrinsic::x86_sse_ucomineq_ss: 2401 case Intrinsic::x86_sse2_comieq_sd: 2402 case Intrinsic::x86_sse2_comige_sd: 2403 case Intrinsic::x86_sse2_comigt_sd: 2404 case Intrinsic::x86_sse2_comile_sd: 2405 case Intrinsic::x86_sse2_comilt_sd: 2406 case Intrinsic::x86_sse2_comineq_sd: 2407 case Intrinsic::x86_sse2_ucomieq_sd: 2408 case Intrinsic::x86_sse2_ucomige_sd: 2409 case Intrinsic::x86_sse2_ucomigt_sd: 2410 case Intrinsic::x86_sse2_ucomile_sd: 2411 case Intrinsic::x86_sse2_ucomilt_sd: 2412 case Intrinsic::x86_sse2_ucomineq_sd: 2413 case Intrinsic::x86_avx512_vcomi_ss: 2414 case Intrinsic::x86_avx512_vcomi_sd: 2415 case Intrinsic::x86_avx512_mask_cmp_ss: 2416 case Intrinsic::x86_avx512_mask_cmp_sd: { 2417 // These intrinsics only demand the 0th element of their input vectors. If 2418 // we can simplify the input based on that, do so now. 2419 bool MadeChange = false; 2420 Value *Arg0 = II->getArgOperand(0); 2421 Value *Arg1 = II->getArgOperand(1); 2422 unsigned VWidth = Arg0->getType()->getVectorNumElements(); 2423 if (Value *V = SimplifyDemandedVectorEltsLow(Arg0, VWidth, 1)) { 2424 II->setArgOperand(0, V); 2425 MadeChange = true; 2426 } 2427 if (Value *V = SimplifyDemandedVectorEltsLow(Arg1, VWidth, 1)) { 2428 II->setArgOperand(1, V); 2429 MadeChange = true; 2430 } 2431 if (MadeChange) 2432 return II; 2433 break; 2434 } 2435 case Intrinsic::x86_avx512_cmp_pd_128: 2436 case Intrinsic::x86_avx512_cmp_pd_256: 2437 case Intrinsic::x86_avx512_cmp_pd_512: 2438 case Intrinsic::x86_avx512_cmp_ps_128: 2439 case Intrinsic::x86_avx512_cmp_ps_256: 2440 case Intrinsic::x86_avx512_cmp_ps_512: { 2441 // Folding cmp(sub(a,b),0) -> cmp(a,b) and cmp(0,sub(a,b)) -> cmp(b,a) 2442 Value *Arg0 = II->getArgOperand(0); 2443 Value *Arg1 = II->getArgOperand(1); 2444 bool Arg0IsZero = match(Arg0, m_PosZeroFP()); 2445 if (Arg0IsZero) 2446 std::swap(Arg0, Arg1); 2447 Value *A, *B; 2448 // This fold requires only the NINF(not +/- inf) since inf minus 2449 // inf is nan. 2450 // NSZ(No Signed Zeros) is not needed because zeros of any sign are 2451 // equal for both compares. 2452 // NNAN is not needed because nans compare the same for both compares. 2453 // The compare intrinsic uses the above assumptions and therefore 2454 // doesn't require additional flags. 2455 if ((match(Arg0, m_OneUse(m_FSub(m_Value(A), m_Value(B)))) && 2456 match(Arg1, m_PosZeroFP()) && isa<Instruction>(Arg0) && 2457 cast<Instruction>(Arg0)->getFastMathFlags().noInfs())) { 2458 if (Arg0IsZero) 2459 std::swap(A, B); 2460 II->setArgOperand(0, A); 2461 II->setArgOperand(1, B); 2462 return II; 2463 } 2464 break; 2465 } 2466 2467 case Intrinsic::x86_avx512_add_ps_512: 2468 case Intrinsic::x86_avx512_div_ps_512: 2469 case Intrinsic::x86_avx512_mul_ps_512: 2470 case Intrinsic::x86_avx512_sub_ps_512: 2471 case Intrinsic::x86_avx512_add_pd_512: 2472 case Intrinsic::x86_avx512_div_pd_512: 2473 case Intrinsic::x86_avx512_mul_pd_512: 2474 case Intrinsic::x86_avx512_sub_pd_512: 2475 // If the rounding mode is CUR_DIRECTION(4) we can turn these into regular 2476 // IR operations. 2477 if (auto *R = dyn_cast<ConstantInt>(II->getArgOperand(2))) { 2478 if (R->getValue() == 4) { 2479 Value *Arg0 = II->getArgOperand(0); 2480 Value *Arg1 = II->getArgOperand(1); 2481 2482 Value *V; 2483 switch (II->getIntrinsicID()) { 2484 default: llvm_unreachable("Case stmts out of sync!"); 2485 case Intrinsic::x86_avx512_add_ps_512: 2486 case Intrinsic::x86_avx512_add_pd_512: 2487 V = Builder.CreateFAdd(Arg0, Arg1); 2488 break; 2489 case Intrinsic::x86_avx512_sub_ps_512: 2490 case Intrinsic::x86_avx512_sub_pd_512: 2491 V = Builder.CreateFSub(Arg0, Arg1); 2492 break; 2493 case Intrinsic::x86_avx512_mul_ps_512: 2494 case Intrinsic::x86_avx512_mul_pd_512: 2495 V = Builder.CreateFMul(Arg0, Arg1); 2496 break; 2497 case Intrinsic::x86_avx512_div_ps_512: 2498 case Intrinsic::x86_avx512_div_pd_512: 2499 V = Builder.CreateFDiv(Arg0, Arg1); 2500 break; 2501 } 2502 2503 return replaceInstUsesWith(*II, V); 2504 } 2505 } 2506 break; 2507 2508 case Intrinsic::x86_avx512_mask_add_ss_round: 2509 case Intrinsic::x86_avx512_mask_div_ss_round: 2510 case Intrinsic::x86_avx512_mask_mul_ss_round: 2511 case Intrinsic::x86_avx512_mask_sub_ss_round: 2512 case Intrinsic::x86_avx512_mask_add_sd_round: 2513 case Intrinsic::x86_avx512_mask_div_sd_round: 2514 case Intrinsic::x86_avx512_mask_mul_sd_round: 2515 case Intrinsic::x86_avx512_mask_sub_sd_round: 2516 // If the rounding mode is CUR_DIRECTION(4) we can turn these into regular 2517 // IR operations. 2518 if (auto *R = dyn_cast<ConstantInt>(II->getArgOperand(4))) { 2519 if (R->getValue() == 4) { 2520 // Extract the element as scalars. 2521 Value *Arg0 = II->getArgOperand(0); 2522 Value *Arg1 = II->getArgOperand(1); 2523 Value *LHS = Builder.CreateExtractElement(Arg0, (uint64_t)0); 2524 Value *RHS = Builder.CreateExtractElement(Arg1, (uint64_t)0); 2525 2526 Value *V; 2527 switch (II->getIntrinsicID()) { 2528 default: llvm_unreachable("Case stmts out of sync!"); 2529 case Intrinsic::x86_avx512_mask_add_ss_round: 2530 case Intrinsic::x86_avx512_mask_add_sd_round: 2531 V = Builder.CreateFAdd(LHS, RHS); 2532 break; 2533 case Intrinsic::x86_avx512_mask_sub_ss_round: 2534 case Intrinsic::x86_avx512_mask_sub_sd_round: 2535 V = Builder.CreateFSub(LHS, RHS); 2536 break; 2537 case Intrinsic::x86_avx512_mask_mul_ss_round: 2538 case Intrinsic::x86_avx512_mask_mul_sd_round: 2539 V = Builder.CreateFMul(LHS, RHS); 2540 break; 2541 case Intrinsic::x86_avx512_mask_div_ss_round: 2542 case Intrinsic::x86_avx512_mask_div_sd_round: 2543 V = Builder.CreateFDiv(LHS, RHS); 2544 break; 2545 } 2546 2547 // Handle the masking aspect of the intrinsic. 2548 Value *Mask = II->getArgOperand(3); 2549 auto *C = dyn_cast<ConstantInt>(Mask); 2550 // We don't need a select if we know the mask bit is a 1. 2551 if (!C || !C->getValue()[0]) { 2552 // Cast the mask to an i1 vector and then extract the lowest element. 2553 auto *MaskTy = VectorType::get(Builder.getInt1Ty(), 2554 cast<IntegerType>(Mask->getType())->getBitWidth()); 2555 Mask = Builder.CreateBitCast(Mask, MaskTy); 2556 Mask = Builder.CreateExtractElement(Mask, (uint64_t)0); 2557 // Extract the lowest element from the passthru operand. 2558 Value *Passthru = Builder.CreateExtractElement(II->getArgOperand(2), 2559 (uint64_t)0); 2560 V = Builder.CreateSelect(Mask, V, Passthru); 2561 } 2562 2563 // Insert the result back into the original argument 0. 2564 V = Builder.CreateInsertElement(Arg0, V, (uint64_t)0); 2565 2566 return replaceInstUsesWith(*II, V); 2567 } 2568 } 2569 LLVM_FALLTHROUGH; 2570 2571 // X86 scalar intrinsics simplified with SimplifyDemandedVectorElts. 2572 case Intrinsic::x86_avx512_mask_max_ss_round: 2573 case Intrinsic::x86_avx512_mask_min_ss_round: 2574 case Intrinsic::x86_avx512_mask_max_sd_round: 2575 case Intrinsic::x86_avx512_mask_min_sd_round: 2576 case Intrinsic::x86_sse_cmp_ss: 2577 case Intrinsic::x86_sse_min_ss: 2578 case Intrinsic::x86_sse_max_ss: 2579 case Intrinsic::x86_sse2_cmp_sd: 2580 case Intrinsic::x86_sse2_min_sd: 2581 case Intrinsic::x86_sse2_max_sd: 2582 case Intrinsic::x86_xop_vfrcz_ss: 2583 case Intrinsic::x86_xop_vfrcz_sd: { 2584 unsigned VWidth = II->getType()->getVectorNumElements(); 2585 APInt UndefElts(VWidth, 0); 2586 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); 2587 if (Value *V = SimplifyDemandedVectorElts(II, AllOnesEltMask, UndefElts)) { 2588 if (V != II) 2589 return replaceInstUsesWith(*II, V); 2590 return II; 2591 } 2592 break; 2593 } 2594 case Intrinsic::x86_sse41_round_ss: 2595 case Intrinsic::x86_sse41_round_sd: { 2596 unsigned VWidth = II->getType()->getVectorNumElements(); 2597 APInt UndefElts(VWidth, 0); 2598 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth)); 2599 if (Value *V = SimplifyDemandedVectorElts(II, AllOnesEltMask, UndefElts)) { 2600 if (V != II) 2601 return replaceInstUsesWith(*II, V); 2602 return II; 2603 } else if (Value *V = simplifyX86round(*II, Builder)) 2604 return replaceInstUsesWith(*II, V); 2605 break; 2606 } 2607 2608 // Constant fold ashr( <A x Bi>, Ci ). 2609 // Constant fold lshr( <A x Bi>, Ci ). 2610 // Constant fold shl( <A x Bi>, Ci ). 2611 case Intrinsic::x86_sse2_psrai_d: 2612 case Intrinsic::x86_sse2_psrai_w: 2613 case Intrinsic::x86_avx2_psrai_d: 2614 case Intrinsic::x86_avx2_psrai_w: 2615 case Intrinsic::x86_avx512_psrai_q_128: 2616 case Intrinsic::x86_avx512_psrai_q_256: 2617 case Intrinsic::x86_avx512_psrai_d_512: 2618 case Intrinsic::x86_avx512_psrai_q_512: 2619 case Intrinsic::x86_avx512_psrai_w_512: 2620 case Intrinsic::x86_sse2_psrli_d: 2621 case Intrinsic::x86_sse2_psrli_q: 2622 case Intrinsic::x86_sse2_psrli_w: 2623 case Intrinsic::x86_avx2_psrli_d: 2624 case Intrinsic::x86_avx2_psrli_q: 2625 case Intrinsic::x86_avx2_psrli_w: 2626 case Intrinsic::x86_avx512_psrli_d_512: 2627 case Intrinsic::x86_avx512_psrli_q_512: 2628 case Intrinsic::x86_avx512_psrli_w_512: 2629 case Intrinsic::x86_sse2_pslli_d: 2630 case Intrinsic::x86_sse2_pslli_q: 2631 case Intrinsic::x86_sse2_pslli_w: 2632 case Intrinsic::x86_avx2_pslli_d: 2633 case Intrinsic::x86_avx2_pslli_q: 2634 case Intrinsic::x86_avx2_pslli_w: 2635 case Intrinsic::x86_avx512_pslli_d_512: 2636 case Intrinsic::x86_avx512_pslli_q_512: 2637 case Intrinsic::x86_avx512_pslli_w_512: 2638 if (Value *V = simplifyX86immShift(*II, Builder)) 2639 return replaceInstUsesWith(*II, V); 2640 break; 2641 2642 case Intrinsic::x86_sse2_psra_d: 2643 case Intrinsic::x86_sse2_psra_w: 2644 case Intrinsic::x86_avx2_psra_d: 2645 case Intrinsic::x86_avx2_psra_w: 2646 case Intrinsic::x86_avx512_psra_q_128: 2647 case Intrinsic::x86_avx512_psra_q_256: 2648 case Intrinsic::x86_avx512_psra_d_512: 2649 case Intrinsic::x86_avx512_psra_q_512: 2650 case Intrinsic::x86_avx512_psra_w_512: 2651 case Intrinsic::x86_sse2_psrl_d: 2652 case Intrinsic::x86_sse2_psrl_q: 2653 case Intrinsic::x86_sse2_psrl_w: 2654 case Intrinsic::x86_avx2_psrl_d: 2655 case Intrinsic::x86_avx2_psrl_q: 2656 case Intrinsic::x86_avx2_psrl_w: 2657 case Intrinsic::x86_avx512_psrl_d_512: 2658 case Intrinsic::x86_avx512_psrl_q_512: 2659 case Intrinsic::x86_avx512_psrl_w_512: 2660 case Intrinsic::x86_sse2_psll_d: 2661 case Intrinsic::x86_sse2_psll_q: 2662 case Intrinsic::x86_sse2_psll_w: 2663 case Intrinsic::x86_avx2_psll_d: 2664 case Intrinsic::x86_avx2_psll_q: 2665 case Intrinsic::x86_avx2_psll_w: 2666 case Intrinsic::x86_avx512_psll_d_512: 2667 case Intrinsic::x86_avx512_psll_q_512: 2668 case Intrinsic::x86_avx512_psll_w_512: { 2669 if (Value *V = simplifyX86immShift(*II, Builder)) 2670 return replaceInstUsesWith(*II, V); 2671 2672 // SSE2/AVX2 uses only the first 64-bits of the 128-bit vector 2673 // operand to compute the shift amount. 2674 Value *Arg1 = II->getArgOperand(1); 2675 assert(Arg1->getType()->getPrimitiveSizeInBits() == 128 && 2676 "Unexpected packed shift size"); 2677 unsigned VWidth = Arg1->getType()->getVectorNumElements(); 2678 2679 if (Value *V = SimplifyDemandedVectorEltsLow(Arg1, VWidth, VWidth / 2)) { 2680 II->setArgOperand(1, V); 2681 return II; 2682 } 2683 break; 2684 } 2685 2686 case Intrinsic::x86_avx2_psllv_d: 2687 case Intrinsic::x86_avx2_psllv_d_256: 2688 case Intrinsic::x86_avx2_psllv_q: 2689 case Intrinsic::x86_avx2_psllv_q_256: 2690 case Intrinsic::x86_avx512_psllv_d_512: 2691 case Intrinsic::x86_avx512_psllv_q_512: 2692 case Intrinsic::x86_avx512_psllv_w_128: 2693 case Intrinsic::x86_avx512_psllv_w_256: 2694 case Intrinsic::x86_avx512_psllv_w_512: 2695 case Intrinsic::x86_avx2_psrav_d: 2696 case Intrinsic::x86_avx2_psrav_d_256: 2697 case Intrinsic::x86_avx512_psrav_q_128: 2698 case Intrinsic::x86_avx512_psrav_q_256: 2699 case Intrinsic::x86_avx512_psrav_d_512: 2700 case Intrinsic::x86_avx512_psrav_q_512: 2701 case Intrinsic::x86_avx512_psrav_w_128: 2702 case Intrinsic::x86_avx512_psrav_w_256: 2703 case Intrinsic::x86_avx512_psrav_w_512: 2704 case Intrinsic::x86_avx2_psrlv_d: 2705 case Intrinsic::x86_avx2_psrlv_d_256: 2706 case Intrinsic::x86_avx2_psrlv_q: 2707 case Intrinsic::x86_avx2_psrlv_q_256: 2708 case Intrinsic::x86_avx512_psrlv_d_512: 2709 case Intrinsic::x86_avx512_psrlv_q_512: 2710 case Intrinsic::x86_avx512_psrlv_w_128: 2711 case Intrinsic::x86_avx512_psrlv_w_256: 2712 case Intrinsic::x86_avx512_psrlv_w_512: 2713 if (Value *V = simplifyX86varShift(*II, Builder)) 2714 return replaceInstUsesWith(*II, V); 2715 break; 2716 2717 case Intrinsic::x86_sse2_packssdw_128: 2718 case Intrinsic::x86_sse2_packsswb_128: 2719 case Intrinsic::x86_avx2_packssdw: 2720 case Intrinsic::x86_avx2_packsswb: 2721 case Intrinsic::x86_avx512_packssdw_512: 2722 case Intrinsic::x86_avx512_packsswb_512: 2723 if (Value *V = simplifyX86pack(*II, true)) 2724 return replaceInstUsesWith(*II, V); 2725 break; 2726 2727 case Intrinsic::x86_sse2_packuswb_128: 2728 case Intrinsic::x86_sse41_packusdw: 2729 case Intrinsic::x86_avx2_packusdw: 2730 case Intrinsic::x86_avx2_packuswb: 2731 case Intrinsic::x86_avx512_packusdw_512: 2732 case Intrinsic::x86_avx512_packuswb_512: 2733 if (Value *V = simplifyX86pack(*II, false)) 2734 return replaceInstUsesWith(*II, V); 2735 break; 2736 2737 case Intrinsic::x86_pclmulqdq: 2738 case Intrinsic::x86_pclmulqdq_256: 2739 case Intrinsic::x86_pclmulqdq_512: { 2740 if (auto *C = dyn_cast<ConstantInt>(II->getArgOperand(2))) { 2741 unsigned Imm = C->getZExtValue(); 2742 2743 bool MadeChange = false; 2744 Value *Arg0 = II->getArgOperand(0); 2745 Value *Arg1 = II->getArgOperand(1); 2746 unsigned VWidth = Arg0->getType()->getVectorNumElements(); 2747 2748 APInt UndefElts1(VWidth, 0); 2749 APInt DemandedElts1 = APInt::getSplat(VWidth, 2750 APInt(2, (Imm & 0x01) ? 2 : 1)); 2751 if (Value *V = SimplifyDemandedVectorElts(Arg0, DemandedElts1, 2752 UndefElts1)) { 2753 II->setArgOperand(0, V); 2754 MadeChange = true; 2755 } 2756 2757 APInt UndefElts2(VWidth, 0); 2758 APInt DemandedElts2 = APInt::getSplat(VWidth, 2759 APInt(2, (Imm & 0x10) ? 2 : 1)); 2760 if (Value *V = SimplifyDemandedVectorElts(Arg1, DemandedElts2, 2761 UndefElts2)) { 2762 II->setArgOperand(1, V); 2763 MadeChange = true; 2764 } 2765 2766 // If either input elements are undef, the result is zero. 2767 if (DemandedElts1.isSubsetOf(UndefElts1) || 2768 DemandedElts2.isSubsetOf(UndefElts2)) 2769 return replaceInstUsesWith(*II, 2770 ConstantAggregateZero::get(II->getType())); 2771 2772 if (MadeChange) 2773 return II; 2774 } 2775 break; 2776 } 2777 2778 case Intrinsic::x86_sse41_insertps: 2779 if (Value *V = simplifyX86insertps(*II, Builder)) 2780 return replaceInstUsesWith(*II, V); 2781 break; 2782 2783 case Intrinsic::x86_sse4a_extrq: { 2784 Value *Op0 = II->getArgOperand(0); 2785 Value *Op1 = II->getArgOperand(1); 2786 unsigned VWidth0 = Op0->getType()->getVectorNumElements(); 2787 unsigned VWidth1 = Op1->getType()->getVectorNumElements(); 2788 assert(Op0->getType()->getPrimitiveSizeInBits() == 128 && 2789 Op1->getType()->getPrimitiveSizeInBits() == 128 && VWidth0 == 2 && 2790 VWidth1 == 16 && "Unexpected operand sizes"); 2791 2792 // See if we're dealing with constant values. 2793 Constant *C1 = dyn_cast<Constant>(Op1); 2794 ConstantInt *CILength = 2795 C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)0)) 2796 : nullptr; 2797 ConstantInt *CIIndex = 2798 C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)1)) 2799 : nullptr; 2800 2801 // Attempt to simplify to a constant, shuffle vector or EXTRQI call. 2802 if (Value *V = simplifyX86extrq(*II, Op0, CILength, CIIndex, Builder)) 2803 return replaceInstUsesWith(*II, V); 2804 2805 // EXTRQ only uses the lowest 64-bits of the first 128-bit vector 2806 // operands and the lowest 16-bits of the second. 2807 bool MadeChange = false; 2808 if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) { 2809 II->setArgOperand(0, V); 2810 MadeChange = true; 2811 } 2812 if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 2)) { 2813 II->setArgOperand(1, V); 2814 MadeChange = true; 2815 } 2816 if (MadeChange) 2817 return II; 2818 break; 2819 } 2820 2821 case Intrinsic::x86_sse4a_extrqi: { 2822 // EXTRQI: Extract Length bits starting from Index. Zero pad the remaining 2823 // bits of the lower 64-bits. The upper 64-bits are undefined. 2824 Value *Op0 = II->getArgOperand(0); 2825 unsigned VWidth = Op0->getType()->getVectorNumElements(); 2826 assert(Op0->getType()->getPrimitiveSizeInBits() == 128 && VWidth == 2 && 2827 "Unexpected operand size"); 2828 2829 // See if we're dealing with constant values. 2830 ConstantInt *CILength = dyn_cast<ConstantInt>(II->getArgOperand(1)); 2831 ConstantInt *CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(2)); 2832 2833 // Attempt to simplify to a constant or shuffle vector. 2834 if (Value *V = simplifyX86extrq(*II, Op0, CILength, CIIndex, Builder)) 2835 return replaceInstUsesWith(*II, V); 2836 2837 // EXTRQI only uses the lowest 64-bits of the first 128-bit vector 2838 // operand. 2839 if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth, 1)) { 2840 II->setArgOperand(0, V); 2841 return II; 2842 } 2843 break; 2844 } 2845 2846 case Intrinsic::x86_sse4a_insertq: { 2847 Value *Op0 = II->getArgOperand(0); 2848 Value *Op1 = II->getArgOperand(1); 2849 unsigned VWidth = Op0->getType()->getVectorNumElements(); 2850 assert(Op0->getType()->getPrimitiveSizeInBits() == 128 && 2851 Op1->getType()->getPrimitiveSizeInBits() == 128 && VWidth == 2 && 2852 Op1->getType()->getVectorNumElements() == 2 && 2853 "Unexpected operand size"); 2854 2855 // See if we're dealing with constant values. 2856 Constant *C1 = dyn_cast<Constant>(Op1); 2857 ConstantInt *CI11 = 2858 C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)1)) 2859 : nullptr; 2860 2861 // Attempt to simplify to a constant, shuffle vector or INSERTQI call. 2862 if (CI11) { 2863 const APInt &V11 = CI11->getValue(); 2864 APInt Len = V11.zextOrTrunc(6); 2865 APInt Idx = V11.lshr(8).zextOrTrunc(6); 2866 if (Value *V = simplifyX86insertq(*II, Op0, Op1, Len, Idx, Builder)) 2867 return replaceInstUsesWith(*II, V); 2868 } 2869 2870 // INSERTQ only uses the lowest 64-bits of the first 128-bit vector 2871 // operand. 2872 if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth, 1)) { 2873 II->setArgOperand(0, V); 2874 return II; 2875 } 2876 break; 2877 } 2878 2879 case Intrinsic::x86_sse4a_insertqi: { 2880 // INSERTQI: Extract lowest Length bits from lower half of second source and 2881 // insert over first source starting at Index bit. The upper 64-bits are 2882 // undefined. 2883 Value *Op0 = II->getArgOperand(0); 2884 Value *Op1 = II->getArgOperand(1); 2885 unsigned VWidth0 = Op0->getType()->getVectorNumElements(); 2886 unsigned VWidth1 = Op1->getType()->getVectorNumElements(); 2887 assert(Op0->getType()->getPrimitiveSizeInBits() == 128 && 2888 Op1->getType()->getPrimitiveSizeInBits() == 128 && VWidth0 == 2 && 2889 VWidth1 == 2 && "Unexpected operand sizes"); 2890 2891 // See if we're dealing with constant values. 2892 ConstantInt *CILength = dyn_cast<ConstantInt>(II->getArgOperand(2)); 2893 ConstantInt *CIIndex = dyn_cast<ConstantInt>(II->getArgOperand(3)); 2894 2895 // Attempt to simplify to a constant or shuffle vector. 2896 if (CILength && CIIndex) { 2897 APInt Len = CILength->getValue().zextOrTrunc(6); 2898 APInt Idx = CIIndex->getValue().zextOrTrunc(6); 2899 if (Value *V = simplifyX86insertq(*II, Op0, Op1, Len, Idx, Builder)) 2900 return replaceInstUsesWith(*II, V); 2901 } 2902 2903 // INSERTQI only uses the lowest 64-bits of the first two 128-bit vector 2904 // operands. 2905 bool MadeChange = false; 2906 if (Value *V = SimplifyDemandedVectorEltsLow(Op0, VWidth0, 1)) { 2907 II->setArgOperand(0, V); 2908 MadeChange = true; 2909 } 2910 if (Value *V = SimplifyDemandedVectorEltsLow(Op1, VWidth1, 1)) { 2911 II->setArgOperand(1, V); 2912 MadeChange = true; 2913 } 2914 if (MadeChange) 2915 return II; 2916 break; 2917 } 2918 2919 case Intrinsic::x86_sse41_pblendvb: 2920 case Intrinsic::x86_sse41_blendvps: 2921 case Intrinsic::x86_sse41_blendvpd: 2922 case Intrinsic::x86_avx_blendv_ps_256: 2923 case Intrinsic::x86_avx_blendv_pd_256: 2924 case Intrinsic::x86_avx2_pblendvb: { 2925 // Convert blendv* to vector selects if the mask is constant. 2926 // This optimization is convoluted because the intrinsic is defined as 2927 // getting a vector of floats or doubles for the ps and pd versions. 2928 // FIXME: That should be changed. 2929 2930 Value *Op0 = II->getArgOperand(0); 2931 Value *Op1 = II->getArgOperand(1); 2932 Value *Mask = II->getArgOperand(2); 2933 2934 // fold (blend A, A, Mask) -> A 2935 if (Op0 == Op1) 2936 return replaceInstUsesWith(CI, Op0); 2937 2938 // Zero Mask - select 1st argument. 2939 if (isa<ConstantAggregateZero>(Mask)) 2940 return replaceInstUsesWith(CI, Op0); 2941 2942 // Constant Mask - select 1st/2nd argument lane based on top bit of mask. 2943 if (auto *ConstantMask = dyn_cast<ConstantDataVector>(Mask)) { 2944 Constant *NewSelector = getNegativeIsTrueBoolVec(ConstantMask); 2945 return SelectInst::Create(NewSelector, Op1, Op0, "blendv"); 2946 } 2947 break; 2948 } 2949 2950 case Intrinsic::x86_ssse3_pshuf_b_128: 2951 case Intrinsic::x86_avx2_pshuf_b: 2952 case Intrinsic::x86_avx512_pshuf_b_512: 2953 if (Value *V = simplifyX86pshufb(*II, Builder)) 2954 return replaceInstUsesWith(*II, V); 2955 break; 2956 2957 case Intrinsic::x86_avx_vpermilvar_ps: 2958 case Intrinsic::x86_avx_vpermilvar_ps_256: 2959 case Intrinsic::x86_avx512_vpermilvar_ps_512: 2960 case Intrinsic::x86_avx_vpermilvar_pd: 2961 case Intrinsic::x86_avx_vpermilvar_pd_256: 2962 case Intrinsic::x86_avx512_vpermilvar_pd_512: 2963 if (Value *V = simplifyX86vpermilvar(*II, Builder)) 2964 return replaceInstUsesWith(*II, V); 2965 break; 2966 2967 case Intrinsic::x86_avx2_permd: 2968 case Intrinsic::x86_avx2_permps: 2969 case Intrinsic::x86_avx512_permvar_df_256: 2970 case Intrinsic::x86_avx512_permvar_df_512: 2971 case Intrinsic::x86_avx512_permvar_di_256: 2972 case Intrinsic::x86_avx512_permvar_di_512: 2973 case Intrinsic::x86_avx512_permvar_hi_128: 2974 case Intrinsic::x86_avx512_permvar_hi_256: 2975 case Intrinsic::x86_avx512_permvar_hi_512: 2976 case Intrinsic::x86_avx512_permvar_qi_128: 2977 case Intrinsic::x86_avx512_permvar_qi_256: 2978 case Intrinsic::x86_avx512_permvar_qi_512: 2979 case Intrinsic::x86_avx512_permvar_sf_512: 2980 case Intrinsic::x86_avx512_permvar_si_512: 2981 if (Value *V = simplifyX86vpermv(*II, Builder)) 2982 return replaceInstUsesWith(*II, V); 2983 break; 2984 2985 case Intrinsic::x86_avx_maskload_ps: 2986 case Intrinsic::x86_avx_maskload_pd: 2987 case Intrinsic::x86_avx_maskload_ps_256: 2988 case Intrinsic::x86_avx_maskload_pd_256: 2989 case Intrinsic::x86_avx2_maskload_d: 2990 case Intrinsic::x86_avx2_maskload_q: 2991 case Intrinsic::x86_avx2_maskload_d_256: 2992 case Intrinsic::x86_avx2_maskload_q_256: 2993 if (Instruction *I = simplifyX86MaskedLoad(*II, *this)) 2994 return I; 2995 break; 2996 2997 case Intrinsic::x86_sse2_maskmov_dqu: 2998 case Intrinsic::x86_avx_maskstore_ps: 2999 case Intrinsic::x86_avx_maskstore_pd: 3000 case Intrinsic::x86_avx_maskstore_ps_256: 3001 case Intrinsic::x86_avx_maskstore_pd_256: 3002 case Intrinsic::x86_avx2_maskstore_d: 3003 case Intrinsic::x86_avx2_maskstore_q: 3004 case Intrinsic::x86_avx2_maskstore_d_256: 3005 case Intrinsic::x86_avx2_maskstore_q_256: 3006 if (simplifyX86MaskedStore(*II, *this)) 3007 return nullptr; 3008 break; 3009 3010 case Intrinsic::x86_xop_vpcomb: 3011 case Intrinsic::x86_xop_vpcomd: 3012 case Intrinsic::x86_xop_vpcomq: 3013 case Intrinsic::x86_xop_vpcomw: 3014 if (Value *V = simplifyX86vpcom(*II, Builder, true)) 3015 return replaceInstUsesWith(*II, V); 3016 break; 3017 3018 case Intrinsic::x86_xop_vpcomub: 3019 case Intrinsic::x86_xop_vpcomud: 3020 case Intrinsic::x86_xop_vpcomuq: 3021 case Intrinsic::x86_xop_vpcomuw: 3022 if (Value *V = simplifyX86vpcom(*II, Builder, false)) 3023 return replaceInstUsesWith(*II, V); 3024 break; 3025 3026 case Intrinsic::ppc_altivec_vperm: 3027 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant. 3028 // Note that ppc_altivec_vperm has a big-endian bias, so when creating 3029 // a vectorshuffle for little endian, we must undo the transformation 3030 // performed on vec_perm in altivec.h. That is, we must complement 3031 // the permutation mask with respect to 31 and reverse the order of 3032 // V1 and V2. 3033 if (Constant *Mask = dyn_cast<Constant>(II->getArgOperand(2))) { 3034 assert(Mask->getType()->getVectorNumElements() == 16 && 3035 "Bad type for intrinsic!"); 3036 3037 // Check that all of the elements are integer constants or undefs. 3038 bool AllEltsOk = true; 3039 for (unsigned i = 0; i != 16; ++i) { 3040 Constant *Elt = Mask->getAggregateElement(i); 3041 if (!Elt || !(isa<ConstantInt>(Elt) || isa<UndefValue>(Elt))) { 3042 AllEltsOk = false; 3043 break; 3044 } 3045 } 3046 3047 if (AllEltsOk) { 3048 // Cast the input vectors to byte vectors. 3049 Value *Op0 = Builder.CreateBitCast(II->getArgOperand(0), 3050 Mask->getType()); 3051 Value *Op1 = Builder.CreateBitCast(II->getArgOperand(1), 3052 Mask->getType()); 3053 Value *Result = UndefValue::get(Op0->getType()); 3054 3055 // Only extract each element once. 3056 Value *ExtractedElts[32]; 3057 memset(ExtractedElts, 0, sizeof(ExtractedElts)); 3058 3059 for (unsigned i = 0; i != 16; ++i) { 3060 if (isa<UndefValue>(Mask->getAggregateElement(i))) 3061 continue; 3062 unsigned Idx = 3063 cast<ConstantInt>(Mask->getAggregateElement(i))->getZExtValue(); 3064 Idx &= 31; // Match the hardware behavior. 3065 if (DL.isLittleEndian()) 3066 Idx = 31 - Idx; 3067 3068 if (!ExtractedElts[Idx]) { 3069 Value *Op0ToUse = (DL.isLittleEndian()) ? Op1 : Op0; 3070 Value *Op1ToUse = (DL.isLittleEndian()) ? Op0 : Op1; 3071 ExtractedElts[Idx] = 3072 Builder.CreateExtractElement(Idx < 16 ? Op0ToUse : Op1ToUse, 3073 Builder.getInt32(Idx&15)); 3074 } 3075 3076 // Insert this value into the result vector. 3077 Result = Builder.CreateInsertElement(Result, ExtractedElts[Idx], 3078 Builder.getInt32(i)); 3079 } 3080 return CastInst::Create(Instruction::BitCast, Result, CI.getType()); 3081 } 3082 } 3083 break; 3084 3085 case Intrinsic::arm_neon_vld1: { 3086 unsigned MemAlign = getKnownAlignment(II->getArgOperand(0), 3087 DL, II, &AC, &DT); 3088 if (Value *V = simplifyNeonVld1(*II, MemAlign, Builder)) 3089 return replaceInstUsesWith(*II, V); 3090 break; 3091 } 3092 3093 case Intrinsic::arm_neon_vld2: 3094 case Intrinsic::arm_neon_vld3: 3095 case Intrinsic::arm_neon_vld4: 3096 case Intrinsic::arm_neon_vld2lane: 3097 case Intrinsic::arm_neon_vld3lane: 3098 case Intrinsic::arm_neon_vld4lane: 3099 case Intrinsic::arm_neon_vst1: 3100 case Intrinsic::arm_neon_vst2: 3101 case Intrinsic::arm_neon_vst3: 3102 case Intrinsic::arm_neon_vst4: 3103 case Intrinsic::arm_neon_vst2lane: 3104 case Intrinsic::arm_neon_vst3lane: 3105 case Intrinsic::arm_neon_vst4lane: { 3106 unsigned MemAlign = 3107 getKnownAlignment(II->getArgOperand(0), DL, II, &AC, &DT); 3108 unsigned AlignArg = II->getNumArgOperands() - 1; 3109 ConstantInt *IntrAlign = dyn_cast<ConstantInt>(II->getArgOperand(AlignArg)); 3110 if (IntrAlign && IntrAlign->getZExtValue() < MemAlign) { 3111 II->setArgOperand(AlignArg, 3112 ConstantInt::get(Type::getInt32Ty(II->getContext()), 3113 MemAlign, false)); 3114 return II; 3115 } 3116 break; 3117 } 3118 3119 case Intrinsic::arm_neon_vtbl1: 3120 case Intrinsic::aarch64_neon_tbl1: 3121 if (Value *V = simplifyNeonTbl1(*II, Builder)) 3122 return replaceInstUsesWith(*II, V); 3123 break; 3124 3125 case Intrinsic::arm_neon_vmulls: 3126 case Intrinsic::arm_neon_vmullu: 3127 case Intrinsic::aarch64_neon_smull: 3128 case Intrinsic::aarch64_neon_umull: { 3129 Value *Arg0 = II->getArgOperand(0); 3130 Value *Arg1 = II->getArgOperand(1); 3131 3132 // Handle mul by zero first: 3133 if (isa<ConstantAggregateZero>(Arg0) || isa<ConstantAggregateZero>(Arg1)) { 3134 return replaceInstUsesWith(CI, ConstantAggregateZero::get(II->getType())); 3135 } 3136 3137 // Check for constant LHS & RHS - in this case we just simplify. 3138 bool Zext = (II->getIntrinsicID() == Intrinsic::arm_neon_vmullu || 3139 II->getIntrinsicID() == Intrinsic::aarch64_neon_umull); 3140 VectorType *NewVT = cast<VectorType>(II->getType()); 3141 if (Constant *CV0 = dyn_cast<Constant>(Arg0)) { 3142 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) { 3143 CV0 = ConstantExpr::getIntegerCast(CV0, NewVT, /*isSigned=*/!Zext); 3144 CV1 = ConstantExpr::getIntegerCast(CV1, NewVT, /*isSigned=*/!Zext); 3145 3146 return replaceInstUsesWith(CI, ConstantExpr::getMul(CV0, CV1)); 3147 } 3148 3149 // Couldn't simplify - canonicalize constant to the RHS. 3150 std::swap(Arg0, Arg1); 3151 } 3152 3153 // Handle mul by one: 3154 if (Constant *CV1 = dyn_cast<Constant>(Arg1)) 3155 if (ConstantInt *Splat = 3156 dyn_cast_or_null<ConstantInt>(CV1->getSplatValue())) 3157 if (Splat->isOne()) 3158 return CastInst::CreateIntegerCast(Arg0, II->getType(), 3159 /*isSigned=*/!Zext); 3160 3161 break; 3162 } 3163 case Intrinsic::arm_neon_aesd: 3164 case Intrinsic::arm_neon_aese: 3165 case Intrinsic::aarch64_crypto_aesd: 3166 case Intrinsic::aarch64_crypto_aese: { 3167 Value *DataArg = II->getArgOperand(0); 3168 Value *KeyArg = II->getArgOperand(1); 3169 3170 // Try to use the builtin XOR in AESE and AESD to eliminate a prior XOR 3171 Value *Data, *Key; 3172 if (match(KeyArg, m_ZeroInt()) && 3173 match(DataArg, m_Xor(m_Value(Data), m_Value(Key)))) { 3174 II->setArgOperand(0, Data); 3175 II->setArgOperand(1, Key); 3176 return II; 3177 } 3178 break; 3179 } 3180 case Intrinsic::amdgcn_rcp: { 3181 Value *Src = II->getArgOperand(0); 3182 3183 // TODO: Move to ConstantFolding/InstSimplify? 3184 if (isa<UndefValue>(Src)) 3185 return replaceInstUsesWith(CI, Src); 3186 3187 if (const ConstantFP *C = dyn_cast<ConstantFP>(Src)) { 3188 const APFloat &ArgVal = C->getValueAPF(); 3189 APFloat Val(ArgVal.getSemantics(), 1.0); 3190 APFloat::opStatus Status = Val.divide(ArgVal, 3191 APFloat::rmNearestTiesToEven); 3192 // Only do this if it was exact and therefore not dependent on the 3193 // rounding mode. 3194 if (Status == APFloat::opOK) 3195 return replaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val)); 3196 } 3197 3198 break; 3199 } 3200 case Intrinsic::amdgcn_rsq: { 3201 Value *Src = II->getArgOperand(0); 3202 3203 // TODO: Move to ConstantFolding/InstSimplify? 3204 if (isa<UndefValue>(Src)) 3205 return replaceInstUsesWith(CI, Src); 3206 break; 3207 } 3208 case Intrinsic::amdgcn_frexp_mant: 3209 case Intrinsic::amdgcn_frexp_exp: { 3210 Value *Src = II->getArgOperand(0); 3211 if (const ConstantFP *C = dyn_cast<ConstantFP>(Src)) { 3212 int Exp; 3213 APFloat Significand = frexp(C->getValueAPF(), Exp, 3214 APFloat::rmNearestTiesToEven); 3215 3216 if (II->getIntrinsicID() == Intrinsic::amdgcn_frexp_mant) { 3217 return replaceInstUsesWith(CI, ConstantFP::get(II->getContext(), 3218 Significand)); 3219 } 3220 3221 // Match instruction special case behavior. 3222 if (Exp == APFloat::IEK_NaN || Exp == APFloat::IEK_Inf) 3223 Exp = 0; 3224 3225 return replaceInstUsesWith(CI, ConstantInt::get(II->getType(), Exp)); 3226 } 3227 3228 if (isa<UndefValue>(Src)) 3229 return replaceInstUsesWith(CI, UndefValue::get(II->getType())); 3230 3231 break; 3232 } 3233 case Intrinsic::amdgcn_class: { 3234 enum { 3235 S_NAN = 1 << 0, // Signaling NaN 3236 Q_NAN = 1 << 1, // Quiet NaN 3237 N_INFINITY = 1 << 2, // Negative infinity 3238 N_NORMAL = 1 << 3, // Negative normal 3239 N_SUBNORMAL = 1 << 4, // Negative subnormal 3240 N_ZERO = 1 << 5, // Negative zero 3241 P_ZERO = 1 << 6, // Positive zero 3242 P_SUBNORMAL = 1 << 7, // Positive subnormal 3243 P_NORMAL = 1 << 8, // Positive normal 3244 P_INFINITY = 1 << 9 // Positive infinity 3245 }; 3246 3247 const uint32_t FullMask = S_NAN | Q_NAN | N_INFINITY | N_NORMAL | 3248 N_SUBNORMAL | N_ZERO | P_ZERO | P_SUBNORMAL | P_NORMAL | P_INFINITY; 3249 3250 Value *Src0 = II->getArgOperand(0); 3251 Value *Src1 = II->getArgOperand(1); 3252 const ConstantInt *CMask = dyn_cast<ConstantInt>(Src1); 3253 if (!CMask) { 3254 if (isa<UndefValue>(Src0)) 3255 return replaceInstUsesWith(*II, UndefValue::get(II->getType())); 3256 3257 if (isa<UndefValue>(Src1)) 3258 return replaceInstUsesWith(*II, ConstantInt::get(II->getType(), false)); 3259 break; 3260 } 3261 3262 uint32_t Mask = CMask->getZExtValue(); 3263 3264 // If all tests are made, it doesn't matter what the value is. 3265 if ((Mask & FullMask) == FullMask) 3266 return replaceInstUsesWith(*II, ConstantInt::get(II->getType(), true)); 3267 3268 if ((Mask & FullMask) == 0) 3269 return replaceInstUsesWith(*II, ConstantInt::get(II->getType(), false)); 3270 3271 if (Mask == (S_NAN | Q_NAN)) { 3272 // Equivalent of isnan. Replace with standard fcmp. 3273 Value *FCmp = Builder.CreateFCmpUNO(Src0, Src0); 3274 FCmp->takeName(II); 3275 return replaceInstUsesWith(*II, FCmp); 3276 } 3277 3278 const ConstantFP *CVal = dyn_cast<ConstantFP>(Src0); 3279 if (!CVal) { 3280 if (isa<UndefValue>(Src0)) 3281 return replaceInstUsesWith(*II, UndefValue::get(II->getType())); 3282 3283 // Clamp mask to used bits 3284 if ((Mask & FullMask) != Mask) { 3285 CallInst *NewCall = Builder.CreateCall(II->getCalledFunction(), 3286 { Src0, ConstantInt::get(Src1->getType(), Mask & FullMask) } 3287 ); 3288 3289 NewCall->takeName(II); 3290 return replaceInstUsesWith(*II, NewCall); 3291 } 3292 3293 break; 3294 } 3295 3296 const APFloat &Val = CVal->getValueAPF(); 3297 3298 bool Result = 3299 ((Mask & S_NAN) && Val.isNaN() && Val.isSignaling()) || 3300 ((Mask & Q_NAN) && Val.isNaN() && !Val.isSignaling()) || 3301 ((Mask & N_INFINITY) && Val.isInfinity() && Val.isNegative()) || 3302 ((Mask & N_NORMAL) && Val.isNormal() && Val.isNegative()) || 3303 ((Mask & N_SUBNORMAL) && Val.isDenormal() && Val.isNegative()) || 3304 ((Mask & N_ZERO) && Val.isZero() && Val.isNegative()) || 3305 ((Mask & P_ZERO) && Val.isZero() && !Val.isNegative()) || 3306 ((Mask & P_SUBNORMAL) && Val.isDenormal() && !Val.isNegative()) || 3307 ((Mask & P_NORMAL) && Val.isNormal() && !Val.isNegative()) || 3308 ((Mask & P_INFINITY) && Val.isInfinity() && !Val.isNegative()); 3309 3310 return replaceInstUsesWith(*II, ConstantInt::get(II->getType(), Result)); 3311 } 3312 case Intrinsic::amdgcn_cvt_pkrtz: { 3313 Value *Src0 = II->getArgOperand(0); 3314 Value *Src1 = II->getArgOperand(1); 3315 if (const ConstantFP *C0 = dyn_cast<ConstantFP>(Src0)) { 3316 if (const ConstantFP *C1 = dyn_cast<ConstantFP>(Src1)) { 3317 const fltSemantics &HalfSem 3318 = II->getType()->getScalarType()->getFltSemantics(); 3319 bool LosesInfo; 3320 APFloat Val0 = C0->getValueAPF(); 3321 APFloat Val1 = C1->getValueAPF(); 3322 Val0.convert(HalfSem, APFloat::rmTowardZero, &LosesInfo); 3323 Val1.convert(HalfSem, APFloat::rmTowardZero, &LosesInfo); 3324 3325 Constant *Folded = ConstantVector::get({ 3326 ConstantFP::get(II->getContext(), Val0), 3327 ConstantFP::get(II->getContext(), Val1) }); 3328 return replaceInstUsesWith(*II, Folded); 3329 } 3330 } 3331 3332 if (isa<UndefValue>(Src0) && isa<UndefValue>(Src1)) 3333 return replaceInstUsesWith(*II, UndefValue::get(II->getType())); 3334 3335 break; 3336 } 3337 case Intrinsic::amdgcn_cvt_pknorm_i16: 3338 case Intrinsic::amdgcn_cvt_pknorm_u16: 3339 case Intrinsic::amdgcn_cvt_pk_i16: 3340 case Intrinsic::amdgcn_cvt_pk_u16: { 3341 Value *Src0 = II->getArgOperand(0); 3342 Value *Src1 = II->getArgOperand(1); 3343 3344 if (isa<UndefValue>(Src0) && isa<UndefValue>(Src1)) 3345 return replaceInstUsesWith(*II, UndefValue::get(II->getType())); 3346 3347 break; 3348 } 3349 case Intrinsic::amdgcn_ubfe: 3350 case Intrinsic::amdgcn_sbfe: { 3351 // Decompose simple cases into standard shifts. 3352 Value *Src = II->getArgOperand(0); 3353 if (isa<UndefValue>(Src)) 3354 return replaceInstUsesWith(*II, Src); 3355 3356 unsigned Width; 3357 Type *Ty = II->getType(); 3358 unsigned IntSize = Ty->getIntegerBitWidth(); 3359 3360 ConstantInt *CWidth = dyn_cast<ConstantInt>(II->getArgOperand(2)); 3361 if (CWidth) { 3362 Width = CWidth->getZExtValue(); 3363 if ((Width & (IntSize - 1)) == 0) 3364 return replaceInstUsesWith(*II, ConstantInt::getNullValue(Ty)); 3365 3366 if (Width >= IntSize) { 3367 // Hardware ignores high bits, so remove those. 3368 II->setArgOperand(2, ConstantInt::get(CWidth->getType(), 3369 Width & (IntSize - 1))); 3370 return II; 3371 } 3372 } 3373 3374 unsigned Offset; 3375 ConstantInt *COffset = dyn_cast<ConstantInt>(II->getArgOperand(1)); 3376 if (COffset) { 3377 Offset = COffset->getZExtValue(); 3378 if (Offset >= IntSize) { 3379 II->setArgOperand(1, ConstantInt::get(COffset->getType(), 3380 Offset & (IntSize - 1))); 3381 return II; 3382 } 3383 } 3384 3385 bool Signed = II->getIntrinsicID() == Intrinsic::amdgcn_sbfe; 3386 3387 // TODO: Also emit sub if only width is constant. 3388 if (!CWidth && COffset && Offset == 0) { 3389 Constant *KSize = ConstantInt::get(COffset->getType(), IntSize); 3390 Value *ShiftVal = Builder.CreateSub(KSize, II->getArgOperand(2)); 3391 ShiftVal = Builder.CreateZExt(ShiftVal, II->getType()); 3392 3393 Value *Shl = Builder.CreateShl(Src, ShiftVal); 3394 Value *RightShift = Signed ? Builder.CreateAShr(Shl, ShiftVal) 3395 : Builder.CreateLShr(Shl, ShiftVal); 3396 RightShift->takeName(II); 3397 return replaceInstUsesWith(*II, RightShift); 3398 } 3399 3400 if (!CWidth || !COffset) 3401 break; 3402 3403 // TODO: This allows folding to undef when the hardware has specific 3404 // behavior? 3405 if (Offset + Width < IntSize) { 3406 Value *Shl = Builder.CreateShl(Src, IntSize - Offset - Width); 3407 Value *RightShift = Signed ? Builder.CreateAShr(Shl, IntSize - Width) 3408 : Builder.CreateLShr(Shl, IntSize - Width); 3409 RightShift->takeName(II); 3410 return replaceInstUsesWith(*II, RightShift); 3411 } 3412 3413 Value *RightShift = Signed ? Builder.CreateAShr(Src, Offset) 3414 : Builder.CreateLShr(Src, Offset); 3415 3416 RightShift->takeName(II); 3417 return replaceInstUsesWith(*II, RightShift); 3418 } 3419 case Intrinsic::amdgcn_exp: 3420 case Intrinsic::amdgcn_exp_compr: { 3421 ConstantInt *En = dyn_cast<ConstantInt>(II->getArgOperand(1)); 3422 if (!En) // Illegal. 3423 break; 3424 3425 unsigned EnBits = En->getZExtValue(); 3426 if (EnBits == 0xf) 3427 break; // All inputs enabled. 3428 3429 bool IsCompr = II->getIntrinsicID() == Intrinsic::amdgcn_exp_compr; 3430 bool Changed = false; 3431 for (int I = 0; I < (IsCompr ? 2 : 4); ++I) { 3432 if ((!IsCompr && (EnBits & (1 << I)) == 0) || 3433 (IsCompr && ((EnBits & (0x3 << (2 * I))) == 0))) { 3434 Value *Src = II->getArgOperand(I + 2); 3435 if (!isa<UndefValue>(Src)) { 3436 II->setArgOperand(I + 2, UndefValue::get(Src->getType())); 3437 Changed = true; 3438 } 3439 } 3440 } 3441 3442 if (Changed) 3443 return II; 3444 3445 break; 3446 } 3447 case Intrinsic::amdgcn_fmed3: { 3448 // Note this does not preserve proper sNaN behavior if IEEE-mode is enabled 3449 // for the shader. 3450 3451 Value *Src0 = II->getArgOperand(0); 3452 Value *Src1 = II->getArgOperand(1); 3453 Value *Src2 = II->getArgOperand(2); 3454 3455 // Checking for NaN before canonicalization provides better fidelity when 3456 // mapping other operations onto fmed3 since the order of operands is 3457 // unchanged. 3458 CallInst *NewCall = nullptr; 3459 if (match(Src0, m_NaN()) || isa<UndefValue>(Src0)) { 3460 NewCall = Builder.CreateMinNum(Src1, Src2); 3461 } else if (match(Src1, m_NaN()) || isa<UndefValue>(Src1)) { 3462 NewCall = Builder.CreateMinNum(Src0, Src2); 3463 } else if (match(Src2, m_NaN()) || isa<UndefValue>(Src2)) { 3464 NewCall = Builder.CreateMaxNum(Src0, Src1); 3465 } 3466 3467 if (NewCall) { 3468 NewCall->copyFastMathFlags(II); 3469 NewCall->takeName(II); 3470 return replaceInstUsesWith(*II, NewCall); 3471 } 3472 3473 bool Swap = false; 3474 // Canonicalize constants to RHS operands. 3475 // 3476 // fmed3(c0, x, c1) -> fmed3(x, c0, c1) 3477 if (isa<Constant>(Src0) && !isa<Constant>(Src1)) { 3478 std::swap(Src0, Src1); 3479 Swap = true; 3480 } 3481 3482 if (isa<Constant>(Src1) && !isa<Constant>(Src2)) { 3483 std::swap(Src1, Src2); 3484 Swap = true; 3485 } 3486 3487 if (isa<Constant>(Src0) && !isa<Constant>(Src1)) { 3488 std::swap(Src0, Src1); 3489 Swap = true; 3490 } 3491 3492 if (Swap) { 3493 II->setArgOperand(0, Src0); 3494 II->setArgOperand(1, Src1); 3495 II->setArgOperand(2, Src2); 3496 return II; 3497 } 3498 3499 if (const ConstantFP *C0 = dyn_cast<ConstantFP>(Src0)) { 3500 if (const ConstantFP *C1 = dyn_cast<ConstantFP>(Src1)) { 3501 if (const ConstantFP *C2 = dyn_cast<ConstantFP>(Src2)) { 3502 APFloat Result = fmed3AMDGCN(C0->getValueAPF(), C1->getValueAPF(), 3503 C2->getValueAPF()); 3504 return replaceInstUsesWith(*II, 3505 ConstantFP::get(Builder.getContext(), Result)); 3506 } 3507 } 3508 } 3509 3510 break; 3511 } 3512 case Intrinsic::amdgcn_icmp: 3513 case Intrinsic::amdgcn_fcmp: { 3514 const ConstantInt *CC = dyn_cast<ConstantInt>(II->getArgOperand(2)); 3515 if (!CC) 3516 break; 3517 3518 // Guard against invalid arguments. 3519 int64_t CCVal = CC->getZExtValue(); 3520 bool IsInteger = II->getIntrinsicID() == Intrinsic::amdgcn_icmp; 3521 if ((IsInteger && (CCVal < CmpInst::FIRST_ICMP_PREDICATE || 3522 CCVal > CmpInst::LAST_ICMP_PREDICATE)) || 3523 (!IsInteger && (CCVal < CmpInst::FIRST_FCMP_PREDICATE || 3524 CCVal > CmpInst::LAST_FCMP_PREDICATE))) 3525 break; 3526 3527 Value *Src0 = II->getArgOperand(0); 3528 Value *Src1 = II->getArgOperand(1); 3529 3530 if (auto *CSrc0 = dyn_cast<Constant>(Src0)) { 3531 if (auto *CSrc1 = dyn_cast<Constant>(Src1)) { 3532 Constant *CCmp = ConstantExpr::getCompare(CCVal, CSrc0, CSrc1); 3533 if (CCmp->isNullValue()) { 3534 return replaceInstUsesWith( 3535 *II, ConstantExpr::getSExt(CCmp, II->getType())); 3536 } 3537 3538 // The result of V_ICMP/V_FCMP assembly instructions (which this 3539 // intrinsic exposes) is one bit per thread, masked with the EXEC 3540 // register (which contains the bitmask of live threads). So a 3541 // comparison that always returns true is the same as a read of the 3542 // EXEC register. 3543 Value *NewF = Intrinsic::getDeclaration( 3544 II->getModule(), Intrinsic::read_register, II->getType()); 3545 Metadata *MDArgs[] = {MDString::get(II->getContext(), "exec")}; 3546 MDNode *MD = MDNode::get(II->getContext(), MDArgs); 3547 Value *Args[] = {MetadataAsValue::get(II->getContext(), MD)}; 3548 CallInst *NewCall = Builder.CreateCall(NewF, Args); 3549 NewCall->addAttribute(AttributeList::FunctionIndex, 3550 Attribute::Convergent); 3551 NewCall->takeName(II); 3552 return replaceInstUsesWith(*II, NewCall); 3553 } 3554 3555 // Canonicalize constants to RHS. 3556 CmpInst::Predicate SwapPred 3557 = CmpInst::getSwappedPredicate(static_cast<CmpInst::Predicate>(CCVal)); 3558 II->setArgOperand(0, Src1); 3559 II->setArgOperand(1, Src0); 3560 II->setArgOperand(2, ConstantInt::get(CC->getType(), 3561 static_cast<int>(SwapPred))); 3562 return II; 3563 } 3564 3565 if (CCVal != CmpInst::ICMP_EQ && CCVal != CmpInst::ICMP_NE) 3566 break; 3567 3568 // Canonicalize compare eq with true value to compare != 0 3569 // llvm.amdgcn.icmp(zext (i1 x), 1, eq) 3570 // -> llvm.amdgcn.icmp(zext (i1 x), 0, ne) 3571 // llvm.amdgcn.icmp(sext (i1 x), -1, eq) 3572 // -> llvm.amdgcn.icmp(sext (i1 x), 0, ne) 3573 Value *ExtSrc; 3574 if (CCVal == CmpInst::ICMP_EQ && 3575 ((match(Src1, m_One()) && match(Src0, m_ZExt(m_Value(ExtSrc)))) || 3576 (match(Src1, m_AllOnes()) && match(Src0, m_SExt(m_Value(ExtSrc))))) && 3577 ExtSrc->getType()->isIntegerTy(1)) { 3578 II->setArgOperand(1, ConstantInt::getNullValue(Src1->getType())); 3579 II->setArgOperand(2, ConstantInt::get(CC->getType(), CmpInst::ICMP_NE)); 3580 return II; 3581 } 3582 3583 CmpInst::Predicate SrcPred; 3584 Value *SrcLHS; 3585 Value *SrcRHS; 3586 3587 // Fold compare eq/ne with 0 from a compare result as the predicate to the 3588 // intrinsic. The typical use is a wave vote function in the library, which 3589 // will be fed from a user code condition compared with 0. Fold in the 3590 // redundant compare. 3591 3592 // llvm.amdgcn.icmp([sz]ext ([if]cmp pred a, b), 0, ne) 3593 // -> llvm.amdgcn.[if]cmp(a, b, pred) 3594 // 3595 // llvm.amdgcn.icmp([sz]ext ([if]cmp pred a, b), 0, eq) 3596 // -> llvm.amdgcn.[if]cmp(a, b, inv pred) 3597 if (match(Src1, m_Zero()) && 3598 match(Src0, 3599 m_ZExtOrSExt(m_Cmp(SrcPred, m_Value(SrcLHS), m_Value(SrcRHS))))) { 3600 if (CCVal == CmpInst::ICMP_EQ) 3601 SrcPred = CmpInst::getInversePredicate(SrcPred); 3602 3603 Intrinsic::ID NewIID = CmpInst::isFPPredicate(SrcPred) ? 3604 Intrinsic::amdgcn_fcmp : Intrinsic::amdgcn_icmp; 3605 3606 Value *NewF = Intrinsic::getDeclaration(II->getModule(), NewIID, 3607 SrcLHS->getType()); 3608 Value *Args[] = { SrcLHS, SrcRHS, 3609 ConstantInt::get(CC->getType(), SrcPred) }; 3610 CallInst *NewCall = Builder.CreateCall(NewF, Args); 3611 NewCall->takeName(II); 3612 return replaceInstUsesWith(*II, NewCall); 3613 } 3614 3615 break; 3616 } 3617 case Intrinsic::amdgcn_wqm_vote: { 3618 // wqm_vote is identity when the argument is constant. 3619 if (!isa<Constant>(II->getArgOperand(0))) 3620 break; 3621 3622 return replaceInstUsesWith(*II, II->getArgOperand(0)); 3623 } 3624 case Intrinsic::amdgcn_kill: { 3625 const ConstantInt *C = dyn_cast<ConstantInt>(II->getArgOperand(0)); 3626 if (!C || !C->getZExtValue()) 3627 break; 3628 3629 // amdgcn.kill(i1 1) is a no-op 3630 return eraseInstFromFunction(CI); 3631 } 3632 case Intrinsic::amdgcn_update_dpp: { 3633 Value *Old = II->getArgOperand(0); 3634 3635 auto BC = dyn_cast<ConstantInt>(II->getArgOperand(5)); 3636 auto RM = dyn_cast<ConstantInt>(II->getArgOperand(3)); 3637 auto BM = dyn_cast<ConstantInt>(II->getArgOperand(4)); 3638 if (!BC || !RM || !BM || 3639 BC->isZeroValue() || 3640 RM->getZExtValue() != 0xF || 3641 BM->getZExtValue() != 0xF || 3642 isa<UndefValue>(Old)) 3643 break; 3644 3645 // If bound_ctrl = 1, row mask = bank mask = 0xf we can omit old value. 3646 II->setOperand(0, UndefValue::get(Old->getType())); 3647 return II; 3648 } 3649 case Intrinsic::stackrestore: { 3650 // If the save is right next to the restore, remove the restore. This can 3651 // happen when variable allocas are DCE'd. 3652 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getArgOperand(0))) { 3653 if (SS->getIntrinsicID() == Intrinsic::stacksave) { 3654 // Skip over debug info. 3655 if (SS->getNextNonDebugInstruction() == II) { 3656 return eraseInstFromFunction(CI); 3657 } 3658 } 3659 } 3660 3661 // Scan down this block to see if there is another stack restore in the 3662 // same block without an intervening call/alloca. 3663 BasicBlock::iterator BI(II); 3664 TerminatorInst *TI = II->getParent()->getTerminator(); 3665 bool CannotRemove = false; 3666 for (++BI; &*BI != TI; ++BI) { 3667 if (isa<AllocaInst>(BI)) { 3668 CannotRemove = true; 3669 break; 3670 } 3671 if (CallInst *BCI = dyn_cast<CallInst>(BI)) { 3672 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) { 3673 // If there is a stackrestore below this one, remove this one. 3674 if (II->getIntrinsicID() == Intrinsic::stackrestore) 3675 return eraseInstFromFunction(CI); 3676 3677 // Bail if we cross over an intrinsic with side effects, such as 3678 // llvm.stacksave, llvm.read_register, or llvm.setjmp. 3679 if (II->mayHaveSideEffects()) { 3680 CannotRemove = true; 3681 break; 3682 } 3683 } else { 3684 // If we found a non-intrinsic call, we can't remove the stack 3685 // restore. 3686 CannotRemove = true; 3687 break; 3688 } 3689 } 3690 } 3691 3692 // If the stack restore is in a return, resume, or unwind block and if there 3693 // are no allocas or calls between the restore and the return, nuke the 3694 // restore. 3695 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<ResumeInst>(TI))) 3696 return eraseInstFromFunction(CI); 3697 break; 3698 } 3699 case Intrinsic::lifetime_start: 3700 // Asan needs to poison memory to detect invalid access which is possible 3701 // even for empty lifetime range. 3702 if (II->getFunction()->hasFnAttribute(Attribute::SanitizeAddress) || 3703 II->getFunction()->hasFnAttribute(Attribute::SanitizeHWAddress)) 3704 break; 3705 3706 if (removeTriviallyEmptyRange(*II, Intrinsic::lifetime_start, 3707 Intrinsic::lifetime_end, *this)) 3708 return nullptr; 3709 break; 3710 case Intrinsic::assume: { 3711 Value *IIOperand = II->getArgOperand(0); 3712 // Remove an assume if it is followed by an identical assume. 3713 // TODO: Do we need this? Unless there are conflicting assumptions, the 3714 // computeKnownBits(IIOperand) below here eliminates redundant assumes. 3715 Instruction *Next = II->getNextNonDebugInstruction(); 3716 if (match(Next, m_Intrinsic<Intrinsic::assume>(m_Specific(IIOperand)))) 3717 return eraseInstFromFunction(CI); 3718 3719 // Canonicalize assume(a && b) -> assume(a); assume(b); 3720 // Note: New assumption intrinsics created here are registered by 3721 // the InstCombineIRInserter object. 3722 Value *AssumeIntrinsic = II->getCalledValue(), *A, *B; 3723 if (match(IIOperand, m_And(m_Value(A), m_Value(B)))) { 3724 Builder.CreateCall(AssumeIntrinsic, A, II->getName()); 3725 Builder.CreateCall(AssumeIntrinsic, B, II->getName()); 3726 return eraseInstFromFunction(*II); 3727 } 3728 // assume(!(a || b)) -> assume(!a); assume(!b); 3729 if (match(IIOperand, m_Not(m_Or(m_Value(A), m_Value(B))))) { 3730 Builder.CreateCall(AssumeIntrinsic, Builder.CreateNot(A), II->getName()); 3731 Builder.CreateCall(AssumeIntrinsic, Builder.CreateNot(B), II->getName()); 3732 return eraseInstFromFunction(*II); 3733 } 3734 3735 // assume( (load addr) != null ) -> add 'nonnull' metadata to load 3736 // (if assume is valid at the load) 3737 CmpInst::Predicate Pred; 3738 Instruction *LHS; 3739 if (match(IIOperand, m_ICmp(Pred, m_Instruction(LHS), m_Zero())) && 3740 Pred == ICmpInst::ICMP_NE && LHS->getOpcode() == Instruction::Load && 3741 LHS->getType()->isPointerTy() && 3742 isValidAssumeForContext(II, LHS, &DT)) { 3743 MDNode *MD = MDNode::get(II->getContext(), None); 3744 LHS->setMetadata(LLVMContext::MD_nonnull, MD); 3745 return eraseInstFromFunction(*II); 3746 3747 // TODO: apply nonnull return attributes to calls and invokes 3748 // TODO: apply range metadata for range check patterns? 3749 } 3750 3751 // If there is a dominating assume with the same condition as this one, 3752 // then this one is redundant, and should be removed. 3753 KnownBits Known(1); 3754 computeKnownBits(IIOperand, Known, 0, II); 3755 if (Known.isAllOnes()) 3756 return eraseInstFromFunction(*II); 3757 3758 // Update the cache of affected values for this assumption (we might be 3759 // here because we just simplified the condition). 3760 AC.updateAffectedValues(II); 3761 break; 3762 } 3763 case Intrinsic::experimental_gc_relocate: { 3764 // Translate facts known about a pointer before relocating into 3765 // facts about the relocate value, while being careful to 3766 // preserve relocation semantics. 3767 Value *DerivedPtr = cast<GCRelocateInst>(II)->getDerivedPtr(); 3768 3769 // Remove the relocation if unused, note that this check is required 3770 // to prevent the cases below from looping forever. 3771 if (II->use_empty()) 3772 return eraseInstFromFunction(*II); 3773 3774 // Undef is undef, even after relocation. 3775 // TODO: provide a hook for this in GCStrategy. This is clearly legal for 3776 // most practical collectors, but there was discussion in the review thread 3777 // about whether it was legal for all possible collectors. 3778 if (isa<UndefValue>(DerivedPtr)) 3779 // Use undef of gc_relocate's type to replace it. 3780 return replaceInstUsesWith(*II, UndefValue::get(II->getType())); 3781 3782 if (auto *PT = dyn_cast<PointerType>(II->getType())) { 3783 // The relocation of null will be null for most any collector. 3784 // TODO: provide a hook for this in GCStrategy. There might be some 3785 // weird collector this property does not hold for. 3786 if (isa<ConstantPointerNull>(DerivedPtr)) 3787 // Use null-pointer of gc_relocate's type to replace it. 3788 return replaceInstUsesWith(*II, ConstantPointerNull::get(PT)); 3789 3790 // isKnownNonNull -> nonnull attribute 3791 if (isKnownNonZero(DerivedPtr, DL, 0, &AC, II, &DT)) 3792 II->addAttribute(AttributeList::ReturnIndex, Attribute::NonNull); 3793 } 3794 3795 // TODO: bitcast(relocate(p)) -> relocate(bitcast(p)) 3796 // Canonicalize on the type from the uses to the defs 3797 3798 // TODO: relocate((gep p, C, C2, ...)) -> gep(relocate(p), C, C2, ...) 3799 break; 3800 } 3801 3802 case Intrinsic::experimental_guard: { 3803 // Is this guard followed by another guard? We scan forward over a small 3804 // fixed window of instructions to handle common cases with conditions 3805 // computed between guards. 3806 Instruction *NextInst = II->getNextNode(); 3807 for (unsigned i = 0; i < GuardWideningWindow; i++) { 3808 // Note: Using context-free form to avoid compile time blow up 3809 if (!isSafeToSpeculativelyExecute(NextInst)) 3810 break; 3811 NextInst = NextInst->getNextNode(); 3812 } 3813 Value *NextCond = nullptr; 3814 if (match(NextInst, 3815 m_Intrinsic<Intrinsic::experimental_guard>(m_Value(NextCond)))) { 3816 Value *CurrCond = II->getArgOperand(0); 3817 3818 // Remove a guard that it is immediately preceded by an identical guard. 3819 if (CurrCond == NextCond) 3820 return eraseInstFromFunction(*NextInst); 3821 3822 // Otherwise canonicalize guard(a); guard(b) -> guard(a & b). 3823 Instruction* MoveI = II->getNextNode(); 3824 while (MoveI != NextInst) { 3825 auto *Temp = MoveI; 3826 MoveI = MoveI->getNextNode(); 3827 Temp->moveBefore(II); 3828 } 3829 II->setArgOperand(0, Builder.CreateAnd(CurrCond, NextCond)); 3830 return eraseInstFromFunction(*NextInst); 3831 } 3832 break; 3833 } 3834 } 3835 return visitCallSite(II); 3836 } 3837 3838 // Fence instruction simplification 3839 Instruction *InstCombiner::visitFenceInst(FenceInst &FI) { 3840 // Remove identical consecutive fences. 3841 Instruction *Next = FI.getNextNonDebugInstruction(); 3842 if (auto *NFI = dyn_cast<FenceInst>(Next)) 3843 if (FI.isIdenticalTo(NFI)) 3844 return eraseInstFromFunction(FI); 3845 return nullptr; 3846 } 3847 3848 // InvokeInst simplification 3849 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) { 3850 return visitCallSite(&II); 3851 } 3852 3853 /// If this cast does not affect the value passed through the varargs area, we 3854 /// can eliminate the use of the cast. 3855 static bool isSafeToEliminateVarargsCast(const CallSite CS, 3856 const DataLayout &DL, 3857 const CastInst *const CI, 3858 const int ix) { 3859 if (!CI->isLosslessCast()) 3860 return false; 3861 3862 // If this is a GC intrinsic, avoid munging types. We need types for 3863 // statepoint reconstruction in SelectionDAG. 3864 // TODO: This is probably something which should be expanded to all 3865 // intrinsics since the entire point of intrinsics is that 3866 // they are understandable by the optimizer. 3867 if (isStatepoint(CS) || isGCRelocate(CS) || isGCResult(CS)) 3868 return false; 3869 3870 // The size of ByVal or InAlloca arguments is derived from the type, so we 3871 // can't change to a type with a different size. If the size were 3872 // passed explicitly we could avoid this check. 3873 if (!CS.isByValOrInAllocaArgument(ix)) 3874 return true; 3875 3876 Type* SrcTy = 3877 cast<PointerType>(CI->getOperand(0)->getType())->getElementType(); 3878 Type* DstTy = cast<PointerType>(CI->getType())->getElementType(); 3879 if (!SrcTy->isSized() || !DstTy->isSized()) 3880 return false; 3881 if (DL.getTypeAllocSize(SrcTy) != DL.getTypeAllocSize(DstTy)) 3882 return false; 3883 return true; 3884 } 3885 3886 Instruction *InstCombiner::tryOptimizeCall(CallInst *CI) { 3887 if (!CI->getCalledFunction()) return nullptr; 3888 3889 auto InstCombineRAUW = [this](Instruction *From, Value *With) { 3890 replaceInstUsesWith(*From, With); 3891 }; 3892 LibCallSimplifier Simplifier(DL, &TLI, ORE, InstCombineRAUW); 3893 if (Value *With = Simplifier.optimizeCall(CI)) { 3894 ++NumSimplified; 3895 return CI->use_empty() ? CI : replaceInstUsesWith(*CI, With); 3896 } 3897 3898 return nullptr; 3899 } 3900 3901 static IntrinsicInst *findInitTrampolineFromAlloca(Value *TrampMem) { 3902 // Strip off at most one level of pointer casts, looking for an alloca. This 3903 // is good enough in practice and simpler than handling any number of casts. 3904 Value *Underlying = TrampMem->stripPointerCasts(); 3905 if (Underlying != TrampMem && 3906 (!Underlying->hasOneUse() || Underlying->user_back() != TrampMem)) 3907 return nullptr; 3908 if (!isa<AllocaInst>(Underlying)) 3909 return nullptr; 3910 3911 IntrinsicInst *InitTrampoline = nullptr; 3912 for (User *U : TrampMem->users()) { 3913 IntrinsicInst *II = dyn_cast<IntrinsicInst>(U); 3914 if (!II) 3915 return nullptr; 3916 if (II->getIntrinsicID() == Intrinsic::init_trampoline) { 3917 if (InitTrampoline) 3918 // More than one init_trampoline writes to this value. Give up. 3919 return nullptr; 3920 InitTrampoline = II; 3921 continue; 3922 } 3923 if (II->getIntrinsicID() == Intrinsic::adjust_trampoline) 3924 // Allow any number of calls to adjust.trampoline. 3925 continue; 3926 return nullptr; 3927 } 3928 3929 // No call to init.trampoline found. 3930 if (!InitTrampoline) 3931 return nullptr; 3932 3933 // Check that the alloca is being used in the expected way. 3934 if (InitTrampoline->getOperand(0) != TrampMem) 3935 return nullptr; 3936 3937 return InitTrampoline; 3938 } 3939 3940 static IntrinsicInst *findInitTrampolineFromBB(IntrinsicInst *AdjustTramp, 3941 Value *TrampMem) { 3942 // Visit all the previous instructions in the basic block, and try to find a 3943 // init.trampoline which has a direct path to the adjust.trampoline. 3944 for (BasicBlock::iterator I = AdjustTramp->getIterator(), 3945 E = AdjustTramp->getParent()->begin(); 3946 I != E;) { 3947 Instruction *Inst = &*--I; 3948 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) 3949 if (II->getIntrinsicID() == Intrinsic::init_trampoline && 3950 II->getOperand(0) == TrampMem) 3951 return II; 3952 if (Inst->mayWriteToMemory()) 3953 return nullptr; 3954 } 3955 return nullptr; 3956 } 3957 3958 // Given a call to llvm.adjust.trampoline, find and return the corresponding 3959 // call to llvm.init.trampoline if the call to the trampoline can be optimized 3960 // to a direct call to a function. Otherwise return NULL. 3961 static IntrinsicInst *findInitTrampoline(Value *Callee) { 3962 Callee = Callee->stripPointerCasts(); 3963 IntrinsicInst *AdjustTramp = dyn_cast<IntrinsicInst>(Callee); 3964 if (!AdjustTramp || 3965 AdjustTramp->getIntrinsicID() != Intrinsic::adjust_trampoline) 3966 return nullptr; 3967 3968 Value *TrampMem = AdjustTramp->getOperand(0); 3969 3970 if (IntrinsicInst *IT = findInitTrampolineFromAlloca(TrampMem)) 3971 return IT; 3972 if (IntrinsicInst *IT = findInitTrampolineFromBB(AdjustTramp, TrampMem)) 3973 return IT; 3974 return nullptr; 3975 } 3976 3977 /// Improvements for call and invoke instructions. 3978 Instruction *InstCombiner::visitCallSite(CallSite CS) { 3979 if (isAllocLikeFn(CS.getInstruction(), &TLI)) 3980 return visitAllocSite(*CS.getInstruction()); 3981 3982 bool Changed = false; 3983 3984 // Mark any parameters that are known to be non-null with the nonnull 3985 // attribute. This is helpful for inlining calls to functions with null 3986 // checks on their arguments. 3987 SmallVector<unsigned, 4> ArgNos; 3988 unsigned ArgNo = 0; 3989 3990 for (Value *V : CS.args()) { 3991 if (V->getType()->isPointerTy() && 3992 !CS.paramHasAttr(ArgNo, Attribute::NonNull) && 3993 isKnownNonZero(V, DL, 0, &AC, CS.getInstruction(), &DT)) 3994 ArgNos.push_back(ArgNo); 3995 ArgNo++; 3996 } 3997 3998 assert(ArgNo == CS.arg_size() && "sanity check"); 3999 4000 if (!ArgNos.empty()) { 4001 AttributeList AS = CS.getAttributes(); 4002 LLVMContext &Ctx = CS.getInstruction()->getContext(); 4003 AS = AS.addParamAttribute(Ctx, ArgNos, 4004 Attribute::get(Ctx, Attribute::NonNull)); 4005 CS.setAttributes(AS); 4006 Changed = true; 4007 } 4008 4009 // If the callee is a pointer to a function, attempt to move any casts to the 4010 // arguments of the call/invoke. 4011 Value *Callee = CS.getCalledValue(); 4012 if (!isa<Function>(Callee) && transformConstExprCastCall(CS)) 4013 return nullptr; 4014 4015 if (Function *CalleeF = dyn_cast<Function>(Callee)) { 4016 // Remove the convergent attr on calls when the callee is not convergent. 4017 if (CS.isConvergent() && !CalleeF->isConvergent() && 4018 !CalleeF->isIntrinsic()) { 4019 LLVM_DEBUG(dbgs() << "Removing convergent attr from instr " 4020 << CS.getInstruction() << "\n"); 4021 CS.setNotConvergent(); 4022 return CS.getInstruction(); 4023 } 4024 4025 // If the call and callee calling conventions don't match, this call must 4026 // be unreachable, as the call is undefined. 4027 if (CalleeF->getCallingConv() != CS.getCallingConv() && 4028 // Only do this for calls to a function with a body. A prototype may 4029 // not actually end up matching the implementation's calling conv for a 4030 // variety of reasons (e.g. it may be written in assembly). 4031 !CalleeF->isDeclaration()) { 4032 Instruction *OldCall = CS.getInstruction(); 4033 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 4034 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 4035 OldCall); 4036 // If OldCall does not return void then replaceAllUsesWith undef. 4037 // This allows ValueHandlers and custom metadata to adjust itself. 4038 if (!OldCall->getType()->isVoidTy()) 4039 replaceInstUsesWith(*OldCall, UndefValue::get(OldCall->getType())); 4040 if (isa<CallInst>(OldCall)) 4041 return eraseInstFromFunction(*OldCall); 4042 4043 // We cannot remove an invoke, because it would change the CFG, just 4044 // change the callee to a null pointer. 4045 cast<InvokeInst>(OldCall)->setCalledFunction( 4046 Constant::getNullValue(CalleeF->getType())); 4047 return nullptr; 4048 } 4049 } 4050 4051 if ((isa<ConstantPointerNull>(Callee) && 4052 !NullPointerIsDefined(CS.getInstruction()->getFunction())) || 4053 isa<UndefValue>(Callee)) { 4054 // If CS does not return void then replaceAllUsesWith undef. 4055 // This allows ValueHandlers and custom metadata to adjust itself. 4056 if (!CS.getInstruction()->getType()->isVoidTy()) 4057 replaceInstUsesWith(*CS.getInstruction(), 4058 UndefValue::get(CS.getInstruction()->getType())); 4059 4060 if (isa<InvokeInst>(CS.getInstruction())) { 4061 // Can't remove an invoke because we cannot change the CFG. 4062 return nullptr; 4063 } 4064 4065 // This instruction is not reachable, just remove it. We insert a store to 4066 // undef so that we know that this code is not reachable, despite the fact 4067 // that we can't modify the CFG here. 4068 new StoreInst(ConstantInt::getTrue(Callee->getContext()), 4069 UndefValue::get(Type::getInt1PtrTy(Callee->getContext())), 4070 CS.getInstruction()); 4071 4072 return eraseInstFromFunction(*CS.getInstruction()); 4073 } 4074 4075 if (IntrinsicInst *II = findInitTrampoline(Callee)) 4076 return transformCallThroughTrampoline(CS, II); 4077 4078 PointerType *PTy = cast<PointerType>(Callee->getType()); 4079 FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 4080 if (FTy->isVarArg()) { 4081 int ix = FTy->getNumParams(); 4082 // See if we can optimize any arguments passed through the varargs area of 4083 // the call. 4084 for (CallSite::arg_iterator I = CS.arg_begin() + FTy->getNumParams(), 4085 E = CS.arg_end(); I != E; ++I, ++ix) { 4086 CastInst *CI = dyn_cast<CastInst>(*I); 4087 if (CI && isSafeToEliminateVarargsCast(CS, DL, CI, ix)) { 4088 *I = CI->getOperand(0); 4089 Changed = true; 4090 } 4091 } 4092 } 4093 4094 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) { 4095 // Inline asm calls cannot throw - mark them 'nounwind'. 4096 CS.setDoesNotThrow(); 4097 Changed = true; 4098 } 4099 4100 // Try to optimize the call if possible, we require DataLayout for most of 4101 // this. None of these calls are seen as possibly dead so go ahead and 4102 // delete the instruction now. 4103 if (CallInst *CI = dyn_cast<CallInst>(CS.getInstruction())) { 4104 Instruction *I = tryOptimizeCall(CI); 4105 // If we changed something return the result, etc. Otherwise let 4106 // the fallthrough check. 4107 if (I) return eraseInstFromFunction(*I); 4108 } 4109 4110 return Changed ? CS.getInstruction() : nullptr; 4111 } 4112 4113 /// If the callee is a constexpr cast of a function, attempt to move the cast to 4114 /// the arguments of the call/invoke. 4115 bool InstCombiner::transformConstExprCastCall(CallSite CS) { 4116 auto *Callee = dyn_cast<Function>(CS.getCalledValue()->stripPointerCasts()); 4117 if (!Callee) 4118 return false; 4119 4120 // If this is a call to a thunk function, don't remove the cast. Thunks are 4121 // used to transparently forward all incoming parameters and outgoing return 4122 // values, so it's important to leave the cast in place. 4123 if (Callee->hasFnAttribute("thunk")) 4124 return false; 4125 4126 // If this is a musttail call, the callee's prototype must match the caller's 4127 // prototype with the exception of pointee types. The code below doesn't 4128 // implement that, so we can't do this transform. 4129 // TODO: Do the transform if it only requires adding pointer casts. 4130 if (CS.isMustTailCall()) 4131 return false; 4132 4133 Instruction *Caller = CS.getInstruction(); 4134 const AttributeList &CallerPAL = CS.getAttributes(); 4135 4136 // Okay, this is a cast from a function to a different type. Unless doing so 4137 // would cause a type conversion of one of our arguments, change this call to 4138 // be a direct call with arguments casted to the appropriate types. 4139 FunctionType *FT = Callee->getFunctionType(); 4140 Type *OldRetTy = Caller->getType(); 4141 Type *NewRetTy = FT->getReturnType(); 4142 4143 // Check to see if we are changing the return type... 4144 if (OldRetTy != NewRetTy) { 4145 4146 if (NewRetTy->isStructTy()) 4147 return false; // TODO: Handle multiple return values. 4148 4149 if (!CastInst::isBitOrNoopPointerCastable(NewRetTy, OldRetTy, DL)) { 4150 if (Callee->isDeclaration()) 4151 return false; // Cannot transform this return value. 4152 4153 if (!Caller->use_empty() && 4154 // void -> non-void is handled specially 4155 !NewRetTy->isVoidTy()) 4156 return false; // Cannot transform this return value. 4157 } 4158 4159 if (!CallerPAL.isEmpty() && !Caller->use_empty()) { 4160 AttrBuilder RAttrs(CallerPAL, AttributeList::ReturnIndex); 4161 if (RAttrs.overlaps(AttributeFuncs::typeIncompatible(NewRetTy))) 4162 return false; // Attribute not compatible with transformed value. 4163 } 4164 4165 // If the callsite is an invoke instruction, and the return value is used by 4166 // a PHI node in a successor, we cannot change the return type of the call 4167 // because there is no place to put the cast instruction (without breaking 4168 // the critical edge). Bail out in this case. 4169 if (!Caller->use_empty()) 4170 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) 4171 for (User *U : II->users()) 4172 if (PHINode *PN = dyn_cast<PHINode>(U)) 4173 if (PN->getParent() == II->getNormalDest() || 4174 PN->getParent() == II->getUnwindDest()) 4175 return false; 4176 } 4177 4178 unsigned NumActualArgs = CS.arg_size(); 4179 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs); 4180 4181 // Prevent us turning: 4182 // declare void @takes_i32_inalloca(i32* inalloca) 4183 // call void bitcast (void (i32*)* @takes_i32_inalloca to void (i32)*)(i32 0) 4184 // 4185 // into: 4186 // call void @takes_i32_inalloca(i32* null) 4187 // 4188 // Similarly, avoid folding away bitcasts of byval calls. 4189 if (Callee->getAttributes().hasAttrSomewhere(Attribute::InAlloca) || 4190 Callee->getAttributes().hasAttrSomewhere(Attribute::ByVal)) 4191 return false; 4192 4193 CallSite::arg_iterator AI = CS.arg_begin(); 4194 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) { 4195 Type *ParamTy = FT->getParamType(i); 4196 Type *ActTy = (*AI)->getType(); 4197 4198 if (!CastInst::isBitOrNoopPointerCastable(ActTy, ParamTy, DL)) 4199 return false; // Cannot transform this parameter value. 4200 4201 if (AttrBuilder(CallerPAL.getParamAttributes(i)) 4202 .overlaps(AttributeFuncs::typeIncompatible(ParamTy))) 4203 return false; // Attribute not compatible with transformed value. 4204 4205 if (CS.isInAllocaArgument(i)) 4206 return false; // Cannot transform to and from inalloca. 4207 4208 // If the parameter is passed as a byval argument, then we have to have a 4209 // sized type and the sized type has to have the same size as the old type. 4210 if (ParamTy != ActTy && CallerPAL.hasParamAttribute(i, Attribute::ByVal)) { 4211 PointerType *ParamPTy = dyn_cast<PointerType>(ParamTy); 4212 if (!ParamPTy || !ParamPTy->getElementType()->isSized()) 4213 return false; 4214 4215 Type *CurElTy = ActTy->getPointerElementType(); 4216 if (DL.getTypeAllocSize(CurElTy) != 4217 DL.getTypeAllocSize(ParamPTy->getElementType())) 4218 return false; 4219 } 4220 } 4221 4222 if (Callee->isDeclaration()) { 4223 // Do not delete arguments unless we have a function body. 4224 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg()) 4225 return false; 4226 4227 // If the callee is just a declaration, don't change the varargsness of the 4228 // call. We don't want to introduce a varargs call where one doesn't 4229 // already exist. 4230 PointerType *APTy = cast<PointerType>(CS.getCalledValue()->getType()); 4231 if (FT->isVarArg()!=cast<FunctionType>(APTy->getElementType())->isVarArg()) 4232 return false; 4233 4234 // If both the callee and the cast type are varargs, we still have to make 4235 // sure the number of fixed parameters are the same or we have the same 4236 // ABI issues as if we introduce a varargs call. 4237 if (FT->isVarArg() && 4238 cast<FunctionType>(APTy->getElementType())->isVarArg() && 4239 FT->getNumParams() != 4240 cast<FunctionType>(APTy->getElementType())->getNumParams()) 4241 return false; 4242 } 4243 4244 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && 4245 !CallerPAL.isEmpty()) { 4246 // In this case we have more arguments than the new function type, but we 4247 // won't be dropping them. Check that these extra arguments have attributes 4248 // that are compatible with being a vararg call argument. 4249 unsigned SRetIdx; 4250 if (CallerPAL.hasAttrSomewhere(Attribute::StructRet, &SRetIdx) && 4251 SRetIdx > FT->getNumParams()) 4252 return false; 4253 } 4254 4255 // Okay, we decided that this is a safe thing to do: go ahead and start 4256 // inserting cast instructions as necessary. 4257 SmallVector<Value *, 8> Args; 4258 SmallVector<AttributeSet, 8> ArgAttrs; 4259 Args.reserve(NumActualArgs); 4260 ArgAttrs.reserve(NumActualArgs); 4261 4262 // Get any return attributes. 4263 AttrBuilder RAttrs(CallerPAL, AttributeList::ReturnIndex); 4264 4265 // If the return value is not being used, the type may not be compatible 4266 // with the existing attributes. Wipe out any problematic attributes. 4267 RAttrs.remove(AttributeFuncs::typeIncompatible(NewRetTy)); 4268 4269 AI = CS.arg_begin(); 4270 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) { 4271 Type *ParamTy = FT->getParamType(i); 4272 4273 Value *NewArg = *AI; 4274 if ((*AI)->getType() != ParamTy) 4275 NewArg = Builder.CreateBitOrPointerCast(*AI, ParamTy); 4276 Args.push_back(NewArg); 4277 4278 // Add any parameter attributes. 4279 ArgAttrs.push_back(CallerPAL.getParamAttributes(i)); 4280 } 4281 4282 // If the function takes more arguments than the call was taking, add them 4283 // now. 4284 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i) { 4285 Args.push_back(Constant::getNullValue(FT->getParamType(i))); 4286 ArgAttrs.push_back(AttributeSet()); 4287 } 4288 4289 // If we are removing arguments to the function, emit an obnoxious warning. 4290 if (FT->getNumParams() < NumActualArgs) { 4291 // TODO: if (!FT->isVarArg()) this call may be unreachable. PR14722 4292 if (FT->isVarArg()) { 4293 // Add all of the arguments in their promoted form to the arg list. 4294 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) { 4295 Type *PTy = getPromotedType((*AI)->getType()); 4296 Value *NewArg = *AI; 4297 if (PTy != (*AI)->getType()) { 4298 // Must promote to pass through va_arg area! 4299 Instruction::CastOps opcode = 4300 CastInst::getCastOpcode(*AI, false, PTy, false); 4301 NewArg = Builder.CreateCast(opcode, *AI, PTy); 4302 } 4303 Args.push_back(NewArg); 4304 4305 // Add any parameter attributes. 4306 ArgAttrs.push_back(CallerPAL.getParamAttributes(i)); 4307 } 4308 } 4309 } 4310 4311 AttributeSet FnAttrs = CallerPAL.getFnAttributes(); 4312 4313 if (NewRetTy->isVoidTy()) 4314 Caller->setName(""); // Void type should not have a name. 4315 4316 assert((ArgAttrs.size() == FT->getNumParams() || FT->isVarArg()) && 4317 "missing argument attributes"); 4318 LLVMContext &Ctx = Callee->getContext(); 4319 AttributeList NewCallerPAL = AttributeList::get( 4320 Ctx, FnAttrs, AttributeSet::get(Ctx, RAttrs), ArgAttrs); 4321 4322 SmallVector<OperandBundleDef, 1> OpBundles; 4323 CS.getOperandBundlesAsDefs(OpBundles); 4324 4325 CallSite NewCS; 4326 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 4327 NewCS = Builder.CreateInvoke(Callee, II->getNormalDest(), 4328 II->getUnwindDest(), Args, OpBundles); 4329 } else { 4330 NewCS = Builder.CreateCall(Callee, Args, OpBundles); 4331 cast<CallInst>(NewCS.getInstruction()) 4332 ->setTailCallKind(cast<CallInst>(Caller)->getTailCallKind()); 4333 } 4334 NewCS->takeName(Caller); 4335 NewCS.setCallingConv(CS.getCallingConv()); 4336 NewCS.setAttributes(NewCallerPAL); 4337 4338 // Preserve the weight metadata for the new call instruction. The metadata 4339 // is used by SamplePGO to check callsite's hotness. 4340 uint64_t W; 4341 if (Caller->extractProfTotalWeight(W)) 4342 NewCS->setProfWeight(W); 4343 4344 // Insert a cast of the return type as necessary. 4345 Instruction *NC = NewCS.getInstruction(); 4346 Value *NV = NC; 4347 if (OldRetTy != NV->getType() && !Caller->use_empty()) { 4348 if (!NV->getType()->isVoidTy()) { 4349 NV = NC = CastInst::CreateBitOrPointerCast(NC, OldRetTy); 4350 NC->setDebugLoc(Caller->getDebugLoc()); 4351 4352 // If this is an invoke instruction, we should insert it after the first 4353 // non-phi, instruction in the normal successor block. 4354 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 4355 BasicBlock::iterator I = II->getNormalDest()->getFirstInsertionPt(); 4356 InsertNewInstBefore(NC, *I); 4357 } else { 4358 // Otherwise, it's a call, just insert cast right after the call. 4359 InsertNewInstBefore(NC, *Caller); 4360 } 4361 Worklist.AddUsersToWorkList(*Caller); 4362 } else { 4363 NV = UndefValue::get(Caller->getType()); 4364 } 4365 } 4366 4367 if (!Caller->use_empty()) 4368 replaceInstUsesWith(*Caller, NV); 4369 else if (Caller->hasValueHandle()) { 4370 if (OldRetTy == NV->getType()) 4371 ValueHandleBase::ValueIsRAUWd(Caller, NV); 4372 else 4373 // We cannot call ValueIsRAUWd with a different type, and the 4374 // actual tracked value will disappear. 4375 ValueHandleBase::ValueIsDeleted(Caller); 4376 } 4377 4378 eraseInstFromFunction(*Caller); 4379 return true; 4380 } 4381 4382 /// Turn a call to a function created by init_trampoline / adjust_trampoline 4383 /// intrinsic pair into a direct call to the underlying function. 4384 Instruction * 4385 InstCombiner::transformCallThroughTrampoline(CallSite CS, 4386 IntrinsicInst *Tramp) { 4387 Value *Callee = CS.getCalledValue(); 4388 PointerType *PTy = cast<PointerType>(Callee->getType()); 4389 FunctionType *FTy = cast<FunctionType>(PTy->getElementType()); 4390 AttributeList Attrs = CS.getAttributes(); 4391 4392 // If the call already has the 'nest' attribute somewhere then give up - 4393 // otherwise 'nest' would occur twice after splicing in the chain. 4394 if (Attrs.hasAttrSomewhere(Attribute::Nest)) 4395 return nullptr; 4396 4397 assert(Tramp && 4398 "transformCallThroughTrampoline called with incorrect CallSite."); 4399 4400 Function *NestF =cast<Function>(Tramp->getArgOperand(1)->stripPointerCasts()); 4401 FunctionType *NestFTy = cast<FunctionType>(NestF->getValueType()); 4402 4403 AttributeList NestAttrs = NestF->getAttributes(); 4404 if (!NestAttrs.isEmpty()) { 4405 unsigned NestArgNo = 0; 4406 Type *NestTy = nullptr; 4407 AttributeSet NestAttr; 4408 4409 // Look for a parameter marked with the 'nest' attribute. 4410 for (FunctionType::param_iterator I = NestFTy->param_begin(), 4411 E = NestFTy->param_end(); 4412 I != E; ++NestArgNo, ++I) { 4413 AttributeSet AS = NestAttrs.getParamAttributes(NestArgNo); 4414 if (AS.hasAttribute(Attribute::Nest)) { 4415 // Record the parameter type and any other attributes. 4416 NestTy = *I; 4417 NestAttr = AS; 4418 break; 4419 } 4420 } 4421 4422 if (NestTy) { 4423 Instruction *Caller = CS.getInstruction(); 4424 std::vector<Value*> NewArgs; 4425 std::vector<AttributeSet> NewArgAttrs; 4426 NewArgs.reserve(CS.arg_size() + 1); 4427 NewArgAttrs.reserve(CS.arg_size()); 4428 4429 // Insert the nest argument into the call argument list, which may 4430 // mean appending it. Likewise for attributes. 4431 4432 { 4433 unsigned ArgNo = 0; 4434 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 4435 do { 4436 if (ArgNo == NestArgNo) { 4437 // Add the chain argument and attributes. 4438 Value *NestVal = Tramp->getArgOperand(2); 4439 if (NestVal->getType() != NestTy) 4440 NestVal = Builder.CreateBitCast(NestVal, NestTy, "nest"); 4441 NewArgs.push_back(NestVal); 4442 NewArgAttrs.push_back(NestAttr); 4443 } 4444 4445 if (I == E) 4446 break; 4447 4448 // Add the original argument and attributes. 4449 NewArgs.push_back(*I); 4450 NewArgAttrs.push_back(Attrs.getParamAttributes(ArgNo)); 4451 4452 ++ArgNo; 4453 ++I; 4454 } while (true); 4455 } 4456 4457 // The trampoline may have been bitcast to a bogus type (FTy). 4458 // Handle this by synthesizing a new function type, equal to FTy 4459 // with the chain parameter inserted. 4460 4461 std::vector<Type*> NewTypes; 4462 NewTypes.reserve(FTy->getNumParams()+1); 4463 4464 // Insert the chain's type into the list of parameter types, which may 4465 // mean appending it. 4466 { 4467 unsigned ArgNo = 0; 4468 FunctionType::param_iterator I = FTy->param_begin(), 4469 E = FTy->param_end(); 4470 4471 do { 4472 if (ArgNo == NestArgNo) 4473 // Add the chain's type. 4474 NewTypes.push_back(NestTy); 4475 4476 if (I == E) 4477 break; 4478 4479 // Add the original type. 4480 NewTypes.push_back(*I); 4481 4482 ++ArgNo; 4483 ++I; 4484 } while (true); 4485 } 4486 4487 // Replace the trampoline call with a direct call. Let the generic 4488 // code sort out any function type mismatches. 4489 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes, 4490 FTy->isVarArg()); 4491 Constant *NewCallee = 4492 NestF->getType() == PointerType::getUnqual(NewFTy) ? 4493 NestF : ConstantExpr::getBitCast(NestF, 4494 PointerType::getUnqual(NewFTy)); 4495 AttributeList NewPAL = 4496 AttributeList::get(FTy->getContext(), Attrs.getFnAttributes(), 4497 Attrs.getRetAttributes(), NewArgAttrs); 4498 4499 SmallVector<OperandBundleDef, 1> OpBundles; 4500 CS.getOperandBundlesAsDefs(OpBundles); 4501 4502 Instruction *NewCaller; 4503 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) { 4504 NewCaller = InvokeInst::Create(NewCallee, 4505 II->getNormalDest(), II->getUnwindDest(), 4506 NewArgs, OpBundles); 4507 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv()); 4508 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL); 4509 } else { 4510 NewCaller = CallInst::Create(NewCallee, NewArgs, OpBundles); 4511 cast<CallInst>(NewCaller)->setTailCallKind( 4512 cast<CallInst>(Caller)->getTailCallKind()); 4513 cast<CallInst>(NewCaller)->setCallingConv( 4514 cast<CallInst>(Caller)->getCallingConv()); 4515 cast<CallInst>(NewCaller)->setAttributes(NewPAL); 4516 } 4517 NewCaller->setDebugLoc(Caller->getDebugLoc()); 4518 4519 return NewCaller; 4520 } 4521 } 4522 4523 // Replace the trampoline call with a direct call. Since there is no 'nest' 4524 // parameter, there is no need to adjust the argument list. Let the generic 4525 // code sort out any function type mismatches. 4526 Constant *NewCallee = 4527 NestF->getType() == PTy ? NestF : 4528 ConstantExpr::getBitCast(NestF, PTy); 4529 CS.setCalledFunction(NewCallee); 4530 return CS.getInstruction(); 4531 } 4532