1 //===- InstCombineSelect.cpp ----------------------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements the visitSelect function. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "InstCombineInternal.h" 14 #include "llvm/ADT/APInt.h" 15 #include "llvm/ADT/Optional.h" 16 #include "llvm/ADT/STLExtras.h" 17 #include "llvm/ADT/SmallVector.h" 18 #include "llvm/Analysis/AssumptionCache.h" 19 #include "llvm/Analysis/CmpInstAnalysis.h" 20 #include "llvm/Analysis/InstructionSimplify.h" 21 #include "llvm/Analysis/ValueTracking.h" 22 #include "llvm/IR/BasicBlock.h" 23 #include "llvm/IR/Constant.h" 24 #include "llvm/IR/Constants.h" 25 #include "llvm/IR/DerivedTypes.h" 26 #include "llvm/IR/IRBuilder.h" 27 #include "llvm/IR/InstrTypes.h" 28 #include "llvm/IR/Instruction.h" 29 #include "llvm/IR/Instructions.h" 30 #include "llvm/IR/IntrinsicInst.h" 31 #include "llvm/IR/Intrinsics.h" 32 #include "llvm/IR/Operator.h" 33 #include "llvm/IR/PatternMatch.h" 34 #include "llvm/IR/Type.h" 35 #include "llvm/IR/User.h" 36 #include "llvm/IR/Value.h" 37 #include "llvm/Support/Casting.h" 38 #include "llvm/Support/ErrorHandling.h" 39 #include "llvm/Support/KnownBits.h" 40 #include "llvm/Transforms/InstCombine/InstCombineWorklist.h" 41 #include "llvm/Transforms/InstCombine/InstCombiner.h" 42 #include <cassert> 43 #include <utility> 44 45 using namespace llvm; 46 using namespace PatternMatch; 47 48 #define DEBUG_TYPE "instcombine" 49 50 /// FIXME: Enabled by default until the pattern is supported well. 51 static cl::opt<bool> EnableUnsafeSelectTransform( 52 "instcombine-unsafe-select-transform", cl::init(true), 53 cl::desc("Enable poison-unsafe select to and/or transform")); 54 55 static Value *createMinMax(InstCombiner::BuilderTy &Builder, 56 SelectPatternFlavor SPF, Value *A, Value *B) { 57 CmpInst::Predicate Pred = getMinMaxPred(SPF); 58 assert(CmpInst::isIntPredicate(Pred) && "Expected integer predicate"); 59 return Builder.CreateSelect(Builder.CreateICmp(Pred, A, B), A, B); 60 } 61 62 /// Replace a select operand based on an equality comparison with the identity 63 /// constant of a binop. 64 static Instruction *foldSelectBinOpIdentity(SelectInst &Sel, 65 const TargetLibraryInfo &TLI, 66 InstCombinerImpl &IC) { 67 // The select condition must be an equality compare with a constant operand. 68 Value *X; 69 Constant *C; 70 CmpInst::Predicate Pred; 71 if (!match(Sel.getCondition(), m_Cmp(Pred, m_Value(X), m_Constant(C)))) 72 return nullptr; 73 74 bool IsEq; 75 if (ICmpInst::isEquality(Pred)) 76 IsEq = Pred == ICmpInst::ICMP_EQ; 77 else if (Pred == FCmpInst::FCMP_OEQ) 78 IsEq = true; 79 else if (Pred == FCmpInst::FCMP_UNE) 80 IsEq = false; 81 else 82 return nullptr; 83 84 // A select operand must be a binop. 85 BinaryOperator *BO; 86 if (!match(Sel.getOperand(IsEq ? 1 : 2), m_BinOp(BO))) 87 return nullptr; 88 89 // The compare constant must be the identity constant for that binop. 90 // If this a floating-point compare with 0.0, any zero constant will do. 91 Type *Ty = BO->getType(); 92 Constant *IdC = ConstantExpr::getBinOpIdentity(BO->getOpcode(), Ty, true); 93 if (IdC != C) { 94 if (!IdC || !CmpInst::isFPPredicate(Pred)) 95 return nullptr; 96 if (!match(IdC, m_AnyZeroFP()) || !match(C, m_AnyZeroFP())) 97 return nullptr; 98 } 99 100 // Last, match the compare variable operand with a binop operand. 101 Value *Y; 102 if (!BO->isCommutative() && !match(BO, m_BinOp(m_Value(Y), m_Specific(X)))) 103 return nullptr; 104 if (!match(BO, m_c_BinOp(m_Value(Y), m_Specific(X)))) 105 return nullptr; 106 107 // +0.0 compares equal to -0.0, and so it does not behave as required for this 108 // transform. Bail out if we can not exclude that possibility. 109 if (isa<FPMathOperator>(BO)) 110 if (!BO->hasNoSignedZeros() && !CannotBeNegativeZero(Y, &TLI)) 111 return nullptr; 112 113 // BO = binop Y, X 114 // S = { select (cmp eq X, C), BO, ? } or { select (cmp ne X, C), ?, BO } 115 // => 116 // S = { select (cmp eq X, C), Y, ? } or { select (cmp ne X, C), ?, Y } 117 return IC.replaceOperand(Sel, IsEq ? 1 : 2, Y); 118 } 119 120 /// This folds: 121 /// select (icmp eq (and X, C1)), TC, FC 122 /// iff C1 is a power 2 and the difference between TC and FC is a power-of-2. 123 /// To something like: 124 /// (shr (and (X, C1)), (log2(C1) - log2(TC-FC))) + FC 125 /// Or: 126 /// (shl (and (X, C1)), (log2(TC-FC) - log2(C1))) + FC 127 /// With some variations depending if FC is larger than TC, or the shift 128 /// isn't needed, or the bit widths don't match. 129 static Value *foldSelectICmpAnd(SelectInst &Sel, ICmpInst *Cmp, 130 InstCombiner::BuilderTy &Builder) { 131 const APInt *SelTC, *SelFC; 132 if (!match(Sel.getTrueValue(), m_APInt(SelTC)) || 133 !match(Sel.getFalseValue(), m_APInt(SelFC))) 134 return nullptr; 135 136 // If this is a vector select, we need a vector compare. 137 Type *SelType = Sel.getType(); 138 if (SelType->isVectorTy() != Cmp->getType()->isVectorTy()) 139 return nullptr; 140 141 Value *V; 142 APInt AndMask; 143 bool CreateAnd = false; 144 ICmpInst::Predicate Pred = Cmp->getPredicate(); 145 if (ICmpInst::isEquality(Pred)) { 146 if (!match(Cmp->getOperand(1), m_Zero())) 147 return nullptr; 148 149 V = Cmp->getOperand(0); 150 const APInt *AndRHS; 151 if (!match(V, m_And(m_Value(), m_Power2(AndRHS)))) 152 return nullptr; 153 154 AndMask = *AndRHS; 155 } else if (decomposeBitTestICmp(Cmp->getOperand(0), Cmp->getOperand(1), 156 Pred, V, AndMask)) { 157 assert(ICmpInst::isEquality(Pred) && "Not equality test?"); 158 if (!AndMask.isPowerOf2()) 159 return nullptr; 160 161 CreateAnd = true; 162 } else { 163 return nullptr; 164 } 165 166 // In general, when both constants are non-zero, we would need an offset to 167 // replace the select. This would require more instructions than we started 168 // with. But there's one special-case that we handle here because it can 169 // simplify/reduce the instructions. 170 APInt TC = *SelTC; 171 APInt FC = *SelFC; 172 if (!TC.isNullValue() && !FC.isNullValue()) { 173 // If the select constants differ by exactly one bit and that's the same 174 // bit that is masked and checked by the select condition, the select can 175 // be replaced by bitwise logic to set/clear one bit of the constant result. 176 if (TC.getBitWidth() != AndMask.getBitWidth() || (TC ^ FC) != AndMask) 177 return nullptr; 178 if (CreateAnd) { 179 // If we have to create an 'and', then we must kill the cmp to not 180 // increase the instruction count. 181 if (!Cmp->hasOneUse()) 182 return nullptr; 183 V = Builder.CreateAnd(V, ConstantInt::get(SelType, AndMask)); 184 } 185 bool ExtraBitInTC = TC.ugt(FC); 186 if (Pred == ICmpInst::ICMP_EQ) { 187 // If the masked bit in V is clear, clear or set the bit in the result: 188 // (V & AndMaskC) == 0 ? TC : FC --> (V & AndMaskC) ^ TC 189 // (V & AndMaskC) == 0 ? TC : FC --> (V & AndMaskC) | TC 190 Constant *C = ConstantInt::get(SelType, TC); 191 return ExtraBitInTC ? Builder.CreateXor(V, C) : Builder.CreateOr(V, C); 192 } 193 if (Pred == ICmpInst::ICMP_NE) { 194 // If the masked bit in V is set, set or clear the bit in the result: 195 // (V & AndMaskC) != 0 ? TC : FC --> (V & AndMaskC) | FC 196 // (V & AndMaskC) != 0 ? TC : FC --> (V & AndMaskC) ^ FC 197 Constant *C = ConstantInt::get(SelType, FC); 198 return ExtraBitInTC ? Builder.CreateOr(V, C) : Builder.CreateXor(V, C); 199 } 200 llvm_unreachable("Only expecting equality predicates"); 201 } 202 203 // Make sure one of the select arms is a power-of-2. 204 if (!TC.isPowerOf2() && !FC.isPowerOf2()) 205 return nullptr; 206 207 // Determine which shift is needed to transform result of the 'and' into the 208 // desired result. 209 const APInt &ValC = !TC.isNullValue() ? TC : FC; 210 unsigned ValZeros = ValC.logBase2(); 211 unsigned AndZeros = AndMask.logBase2(); 212 213 // Insert the 'and' instruction on the input to the truncate. 214 if (CreateAnd) 215 V = Builder.CreateAnd(V, ConstantInt::get(V->getType(), AndMask)); 216 217 // If types don't match, we can still convert the select by introducing a zext 218 // or a trunc of the 'and'. 219 if (ValZeros > AndZeros) { 220 V = Builder.CreateZExtOrTrunc(V, SelType); 221 V = Builder.CreateShl(V, ValZeros - AndZeros); 222 } else if (ValZeros < AndZeros) { 223 V = Builder.CreateLShr(V, AndZeros - ValZeros); 224 V = Builder.CreateZExtOrTrunc(V, SelType); 225 } else { 226 V = Builder.CreateZExtOrTrunc(V, SelType); 227 } 228 229 // Okay, now we know that everything is set up, we just don't know whether we 230 // have a icmp_ne or icmp_eq and whether the true or false val is the zero. 231 bool ShouldNotVal = !TC.isNullValue(); 232 ShouldNotVal ^= Pred == ICmpInst::ICMP_NE; 233 if (ShouldNotVal) 234 V = Builder.CreateXor(V, ValC); 235 236 return V; 237 } 238 239 /// We want to turn code that looks like this: 240 /// %C = or %A, %B 241 /// %D = select %cond, %C, %A 242 /// into: 243 /// %C = select %cond, %B, 0 244 /// %D = or %A, %C 245 /// 246 /// Assuming that the specified instruction is an operand to the select, return 247 /// a bitmask indicating which operands of this instruction are foldable if they 248 /// equal the other incoming value of the select. 249 static unsigned getSelectFoldableOperands(BinaryOperator *I) { 250 switch (I->getOpcode()) { 251 case Instruction::Add: 252 case Instruction::Mul: 253 case Instruction::And: 254 case Instruction::Or: 255 case Instruction::Xor: 256 return 3; // Can fold through either operand. 257 case Instruction::Sub: // Can only fold on the amount subtracted. 258 case Instruction::Shl: // Can only fold on the shift amount. 259 case Instruction::LShr: 260 case Instruction::AShr: 261 return 1; 262 default: 263 return 0; // Cannot fold 264 } 265 } 266 267 /// We have (select c, TI, FI), and we know that TI and FI have the same opcode. 268 Instruction *InstCombinerImpl::foldSelectOpOp(SelectInst &SI, Instruction *TI, 269 Instruction *FI) { 270 // Don't break up min/max patterns. The hasOneUse checks below prevent that 271 // for most cases, but vector min/max with bitcasts can be transformed. If the 272 // one-use restrictions are eased for other patterns, we still don't want to 273 // obfuscate min/max. 274 if ((match(&SI, m_SMin(m_Value(), m_Value())) || 275 match(&SI, m_SMax(m_Value(), m_Value())) || 276 match(&SI, m_UMin(m_Value(), m_Value())) || 277 match(&SI, m_UMax(m_Value(), m_Value())))) 278 return nullptr; 279 280 // If this is a cast from the same type, merge. 281 Value *Cond = SI.getCondition(); 282 Type *CondTy = Cond->getType(); 283 if (TI->getNumOperands() == 1 && TI->isCast()) { 284 Type *FIOpndTy = FI->getOperand(0)->getType(); 285 if (TI->getOperand(0)->getType() != FIOpndTy) 286 return nullptr; 287 288 // The select condition may be a vector. We may only change the operand 289 // type if the vector width remains the same (and matches the condition). 290 if (auto *CondVTy = dyn_cast<VectorType>(CondTy)) { 291 if (!FIOpndTy->isVectorTy() || 292 CondVTy->getElementCount() != 293 cast<VectorType>(FIOpndTy)->getElementCount()) 294 return nullptr; 295 296 // TODO: If the backend knew how to deal with casts better, we could 297 // remove this limitation. For now, there's too much potential to create 298 // worse codegen by promoting the select ahead of size-altering casts 299 // (PR28160). 300 // 301 // Note that ValueTracking's matchSelectPattern() looks through casts 302 // without checking 'hasOneUse' when it matches min/max patterns, so this 303 // transform may end up happening anyway. 304 if (TI->getOpcode() != Instruction::BitCast && 305 (!TI->hasOneUse() || !FI->hasOneUse())) 306 return nullptr; 307 } else if (!TI->hasOneUse() || !FI->hasOneUse()) { 308 // TODO: The one-use restrictions for a scalar select could be eased if 309 // the fold of a select in visitLoadInst() was enhanced to match a pattern 310 // that includes a cast. 311 return nullptr; 312 } 313 314 // Fold this by inserting a select from the input values. 315 Value *NewSI = 316 Builder.CreateSelect(Cond, TI->getOperand(0), FI->getOperand(0), 317 SI.getName() + ".v", &SI); 318 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI, 319 TI->getType()); 320 } 321 322 // Cond ? -X : -Y --> -(Cond ? X : Y) 323 Value *X, *Y; 324 if (match(TI, m_FNeg(m_Value(X))) && match(FI, m_FNeg(m_Value(Y))) && 325 (TI->hasOneUse() || FI->hasOneUse())) { 326 Value *NewSel = Builder.CreateSelect(Cond, X, Y, SI.getName() + ".v", &SI); 327 return UnaryOperator::CreateFNegFMF(NewSel, TI); 328 } 329 330 // Only handle binary operators (including two-operand getelementptr) with 331 // one-use here. As with the cast case above, it may be possible to relax the 332 // one-use constraint, but that needs be examined carefully since it may not 333 // reduce the total number of instructions. 334 if (TI->getNumOperands() != 2 || FI->getNumOperands() != 2 || 335 (!isa<BinaryOperator>(TI) && !isa<GetElementPtrInst>(TI)) || 336 !TI->hasOneUse() || !FI->hasOneUse()) 337 return nullptr; 338 339 // Figure out if the operations have any operands in common. 340 Value *MatchOp, *OtherOpT, *OtherOpF; 341 bool MatchIsOpZero; 342 if (TI->getOperand(0) == FI->getOperand(0)) { 343 MatchOp = TI->getOperand(0); 344 OtherOpT = TI->getOperand(1); 345 OtherOpF = FI->getOperand(1); 346 MatchIsOpZero = true; 347 } else if (TI->getOperand(1) == FI->getOperand(1)) { 348 MatchOp = TI->getOperand(1); 349 OtherOpT = TI->getOperand(0); 350 OtherOpF = FI->getOperand(0); 351 MatchIsOpZero = false; 352 } else if (!TI->isCommutative()) { 353 return nullptr; 354 } else if (TI->getOperand(0) == FI->getOperand(1)) { 355 MatchOp = TI->getOperand(0); 356 OtherOpT = TI->getOperand(1); 357 OtherOpF = FI->getOperand(0); 358 MatchIsOpZero = true; 359 } else if (TI->getOperand(1) == FI->getOperand(0)) { 360 MatchOp = TI->getOperand(1); 361 OtherOpT = TI->getOperand(0); 362 OtherOpF = FI->getOperand(1); 363 MatchIsOpZero = true; 364 } else { 365 return nullptr; 366 } 367 368 // If the select condition is a vector, the operands of the original select's 369 // operands also must be vectors. This may not be the case for getelementptr 370 // for example. 371 if (CondTy->isVectorTy() && (!OtherOpT->getType()->isVectorTy() || 372 !OtherOpF->getType()->isVectorTy())) 373 return nullptr; 374 375 // If we reach here, they do have operations in common. 376 Value *NewSI = Builder.CreateSelect(Cond, OtherOpT, OtherOpF, 377 SI.getName() + ".v", &SI); 378 Value *Op0 = MatchIsOpZero ? MatchOp : NewSI; 379 Value *Op1 = MatchIsOpZero ? NewSI : MatchOp; 380 if (auto *BO = dyn_cast<BinaryOperator>(TI)) { 381 BinaryOperator *NewBO = BinaryOperator::Create(BO->getOpcode(), Op0, Op1); 382 NewBO->copyIRFlags(TI); 383 NewBO->andIRFlags(FI); 384 return NewBO; 385 } 386 if (auto *TGEP = dyn_cast<GetElementPtrInst>(TI)) { 387 auto *FGEP = cast<GetElementPtrInst>(FI); 388 Type *ElementType = TGEP->getResultElementType(); 389 return TGEP->isInBounds() && FGEP->isInBounds() 390 ? GetElementPtrInst::CreateInBounds(ElementType, Op0, {Op1}) 391 : GetElementPtrInst::Create(ElementType, Op0, {Op1}); 392 } 393 llvm_unreachable("Expected BinaryOperator or GEP"); 394 return nullptr; 395 } 396 397 static bool isSelect01(const APInt &C1I, const APInt &C2I) { 398 if (!C1I.isNullValue() && !C2I.isNullValue()) // One side must be zero. 399 return false; 400 return C1I.isOneValue() || C1I.isAllOnesValue() || 401 C2I.isOneValue() || C2I.isAllOnesValue(); 402 } 403 404 /// Try to fold the select into one of the operands to allow further 405 /// optimization. 406 Instruction *InstCombinerImpl::foldSelectIntoOp(SelectInst &SI, Value *TrueVal, 407 Value *FalseVal) { 408 // See the comment above GetSelectFoldableOperands for a description of the 409 // transformation we are doing here. 410 if (auto *TVI = dyn_cast<BinaryOperator>(TrueVal)) { 411 if (TVI->hasOneUse() && !isa<Constant>(FalseVal)) { 412 if (unsigned SFO = getSelectFoldableOperands(TVI)) { 413 unsigned OpToFold = 0; 414 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) { 415 OpToFold = 1; 416 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) { 417 OpToFold = 2; 418 } 419 420 if (OpToFold) { 421 Constant *C = ConstantExpr::getBinOpIdentity(TVI->getOpcode(), 422 TVI->getType(), true); 423 Value *OOp = TVI->getOperand(2-OpToFold); 424 // Avoid creating select between 2 constants unless it's selecting 425 // between 0, 1 and -1. 426 const APInt *OOpC; 427 bool OOpIsAPInt = match(OOp, m_APInt(OOpC)); 428 if (!isa<Constant>(OOp) || 429 (OOpIsAPInt && isSelect01(C->getUniqueInteger(), *OOpC))) { 430 Value *NewSel = Builder.CreateSelect(SI.getCondition(), OOp, C); 431 NewSel->takeName(TVI); 432 BinaryOperator *BO = BinaryOperator::Create(TVI->getOpcode(), 433 FalseVal, NewSel); 434 BO->copyIRFlags(TVI); 435 return BO; 436 } 437 } 438 } 439 } 440 } 441 442 if (auto *FVI = dyn_cast<BinaryOperator>(FalseVal)) { 443 if (FVI->hasOneUse() && !isa<Constant>(TrueVal)) { 444 if (unsigned SFO = getSelectFoldableOperands(FVI)) { 445 unsigned OpToFold = 0; 446 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) { 447 OpToFold = 1; 448 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) { 449 OpToFold = 2; 450 } 451 452 if (OpToFold) { 453 Constant *C = ConstantExpr::getBinOpIdentity(FVI->getOpcode(), 454 FVI->getType(), true); 455 Value *OOp = FVI->getOperand(2-OpToFold); 456 // Avoid creating select between 2 constants unless it's selecting 457 // between 0, 1 and -1. 458 const APInt *OOpC; 459 bool OOpIsAPInt = match(OOp, m_APInt(OOpC)); 460 if (!isa<Constant>(OOp) || 461 (OOpIsAPInt && isSelect01(C->getUniqueInteger(), *OOpC))) { 462 Value *NewSel = Builder.CreateSelect(SI.getCondition(), C, OOp); 463 NewSel->takeName(FVI); 464 BinaryOperator *BO = BinaryOperator::Create(FVI->getOpcode(), 465 TrueVal, NewSel); 466 BO->copyIRFlags(FVI); 467 return BO; 468 } 469 } 470 } 471 } 472 } 473 474 return nullptr; 475 } 476 477 /// We want to turn: 478 /// (select (icmp eq (and X, Y), 0), (and (lshr X, Z), 1), 1) 479 /// into: 480 /// zext (icmp ne i32 (and X, (or Y, (shl 1, Z))), 0) 481 /// Note: 482 /// Z may be 0 if lshr is missing. 483 /// Worst-case scenario is that we will replace 5 instructions with 5 different 484 /// instructions, but we got rid of select. 485 static Instruction *foldSelectICmpAndAnd(Type *SelType, const ICmpInst *Cmp, 486 Value *TVal, Value *FVal, 487 InstCombiner::BuilderTy &Builder) { 488 if (!(Cmp->hasOneUse() && Cmp->getOperand(0)->hasOneUse() && 489 Cmp->getPredicate() == ICmpInst::ICMP_EQ && 490 match(Cmp->getOperand(1), m_Zero()) && match(FVal, m_One()))) 491 return nullptr; 492 493 // The TrueVal has general form of: and %B, 1 494 Value *B; 495 if (!match(TVal, m_OneUse(m_And(m_Value(B), m_One())))) 496 return nullptr; 497 498 // Where %B may be optionally shifted: lshr %X, %Z. 499 Value *X, *Z; 500 const bool HasShift = match(B, m_OneUse(m_LShr(m_Value(X), m_Value(Z)))); 501 if (!HasShift) 502 X = B; 503 504 Value *Y; 505 if (!match(Cmp->getOperand(0), m_c_And(m_Specific(X), m_Value(Y)))) 506 return nullptr; 507 508 // ((X & Y) == 0) ? ((X >> Z) & 1) : 1 --> (X & (Y | (1 << Z))) != 0 509 // ((X & Y) == 0) ? (X & 1) : 1 --> (X & (Y | 1)) != 0 510 Constant *One = ConstantInt::get(SelType, 1); 511 Value *MaskB = HasShift ? Builder.CreateShl(One, Z) : One; 512 Value *FullMask = Builder.CreateOr(Y, MaskB); 513 Value *MaskedX = Builder.CreateAnd(X, FullMask); 514 Value *ICmpNeZero = Builder.CreateIsNotNull(MaskedX); 515 return new ZExtInst(ICmpNeZero, SelType); 516 } 517 518 /// We want to turn: 519 /// (select (icmp sgt x, C), lshr (X, Y), ashr (X, Y)); iff C s>= -1 520 /// (select (icmp slt x, C), ashr (X, Y), lshr (X, Y)); iff C s>= 0 521 /// into: 522 /// ashr (X, Y) 523 static Value *foldSelectICmpLshrAshr(const ICmpInst *IC, Value *TrueVal, 524 Value *FalseVal, 525 InstCombiner::BuilderTy &Builder) { 526 ICmpInst::Predicate Pred = IC->getPredicate(); 527 Value *CmpLHS = IC->getOperand(0); 528 Value *CmpRHS = IC->getOperand(1); 529 if (!CmpRHS->getType()->isIntOrIntVectorTy()) 530 return nullptr; 531 532 Value *X, *Y; 533 unsigned Bitwidth = CmpRHS->getType()->getScalarSizeInBits(); 534 if ((Pred != ICmpInst::ICMP_SGT || 535 !match(CmpRHS, 536 m_SpecificInt_ICMP(ICmpInst::ICMP_SGE, APInt(Bitwidth, -1)))) && 537 (Pred != ICmpInst::ICMP_SLT || 538 !match(CmpRHS, 539 m_SpecificInt_ICMP(ICmpInst::ICMP_SGE, APInt(Bitwidth, 0))))) 540 return nullptr; 541 542 // Canonicalize so that ashr is in FalseVal. 543 if (Pred == ICmpInst::ICMP_SLT) 544 std::swap(TrueVal, FalseVal); 545 546 if (match(TrueVal, m_LShr(m_Value(X), m_Value(Y))) && 547 match(FalseVal, m_AShr(m_Specific(X), m_Specific(Y))) && 548 match(CmpLHS, m_Specific(X))) { 549 const auto *Ashr = cast<Instruction>(FalseVal); 550 // if lshr is not exact and ashr is, this new ashr must not be exact. 551 bool IsExact = Ashr->isExact() && cast<Instruction>(TrueVal)->isExact(); 552 return Builder.CreateAShr(X, Y, IC->getName(), IsExact); 553 } 554 555 return nullptr; 556 } 557 558 /// We want to turn: 559 /// (select (icmp eq (and X, C1), 0), Y, (or Y, C2)) 560 /// into: 561 /// (or (shl (and X, C1), C3), Y) 562 /// iff: 563 /// C1 and C2 are both powers of 2 564 /// where: 565 /// C3 = Log(C2) - Log(C1) 566 /// 567 /// This transform handles cases where: 568 /// 1. The icmp predicate is inverted 569 /// 2. The select operands are reversed 570 /// 3. The magnitude of C2 and C1 are flipped 571 static Value *foldSelectICmpAndOr(const ICmpInst *IC, Value *TrueVal, 572 Value *FalseVal, 573 InstCombiner::BuilderTy &Builder) { 574 // Only handle integer compares. Also, if this is a vector select, we need a 575 // vector compare. 576 if (!TrueVal->getType()->isIntOrIntVectorTy() || 577 TrueVal->getType()->isVectorTy() != IC->getType()->isVectorTy()) 578 return nullptr; 579 580 Value *CmpLHS = IC->getOperand(0); 581 Value *CmpRHS = IC->getOperand(1); 582 583 Value *V; 584 unsigned C1Log; 585 bool IsEqualZero; 586 bool NeedAnd = false; 587 if (IC->isEquality()) { 588 if (!match(CmpRHS, m_Zero())) 589 return nullptr; 590 591 const APInt *C1; 592 if (!match(CmpLHS, m_And(m_Value(), m_Power2(C1)))) 593 return nullptr; 594 595 V = CmpLHS; 596 C1Log = C1->logBase2(); 597 IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_EQ; 598 } else if (IC->getPredicate() == ICmpInst::ICMP_SLT || 599 IC->getPredicate() == ICmpInst::ICMP_SGT) { 600 // We also need to recognize (icmp slt (trunc (X)), 0) and 601 // (icmp sgt (trunc (X)), -1). 602 IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_SGT; 603 if ((IsEqualZero && !match(CmpRHS, m_AllOnes())) || 604 (!IsEqualZero && !match(CmpRHS, m_Zero()))) 605 return nullptr; 606 607 if (!match(CmpLHS, m_OneUse(m_Trunc(m_Value(V))))) 608 return nullptr; 609 610 C1Log = CmpLHS->getType()->getScalarSizeInBits() - 1; 611 NeedAnd = true; 612 } else { 613 return nullptr; 614 } 615 616 const APInt *C2; 617 bool OrOnTrueVal = false; 618 bool OrOnFalseVal = match(FalseVal, m_Or(m_Specific(TrueVal), m_Power2(C2))); 619 if (!OrOnFalseVal) 620 OrOnTrueVal = match(TrueVal, m_Or(m_Specific(FalseVal), m_Power2(C2))); 621 622 if (!OrOnFalseVal && !OrOnTrueVal) 623 return nullptr; 624 625 Value *Y = OrOnFalseVal ? TrueVal : FalseVal; 626 627 unsigned C2Log = C2->logBase2(); 628 629 bool NeedXor = (!IsEqualZero && OrOnFalseVal) || (IsEqualZero && OrOnTrueVal); 630 bool NeedShift = C1Log != C2Log; 631 bool NeedZExtTrunc = Y->getType()->getScalarSizeInBits() != 632 V->getType()->getScalarSizeInBits(); 633 634 // Make sure we don't create more instructions than we save. 635 Value *Or = OrOnFalseVal ? FalseVal : TrueVal; 636 if ((NeedShift + NeedXor + NeedZExtTrunc) > 637 (IC->hasOneUse() + Or->hasOneUse())) 638 return nullptr; 639 640 if (NeedAnd) { 641 // Insert the AND instruction on the input to the truncate. 642 APInt C1 = APInt::getOneBitSet(V->getType()->getScalarSizeInBits(), C1Log); 643 V = Builder.CreateAnd(V, ConstantInt::get(V->getType(), C1)); 644 } 645 646 if (C2Log > C1Log) { 647 V = Builder.CreateZExtOrTrunc(V, Y->getType()); 648 V = Builder.CreateShl(V, C2Log - C1Log); 649 } else if (C1Log > C2Log) { 650 V = Builder.CreateLShr(V, C1Log - C2Log); 651 V = Builder.CreateZExtOrTrunc(V, Y->getType()); 652 } else 653 V = Builder.CreateZExtOrTrunc(V, Y->getType()); 654 655 if (NeedXor) 656 V = Builder.CreateXor(V, *C2); 657 658 return Builder.CreateOr(V, Y); 659 } 660 661 /// Canonicalize a set or clear of a masked set of constant bits to 662 /// select-of-constants form. 663 static Instruction *foldSetClearBits(SelectInst &Sel, 664 InstCombiner::BuilderTy &Builder) { 665 Value *Cond = Sel.getCondition(); 666 Value *T = Sel.getTrueValue(); 667 Value *F = Sel.getFalseValue(); 668 Type *Ty = Sel.getType(); 669 Value *X; 670 const APInt *NotC, *C; 671 672 // Cond ? (X & ~C) : (X | C) --> (X & ~C) | (Cond ? 0 : C) 673 if (match(T, m_And(m_Value(X), m_APInt(NotC))) && 674 match(F, m_OneUse(m_Or(m_Specific(X), m_APInt(C)))) && *NotC == ~(*C)) { 675 Constant *Zero = ConstantInt::getNullValue(Ty); 676 Constant *OrC = ConstantInt::get(Ty, *C); 677 Value *NewSel = Builder.CreateSelect(Cond, Zero, OrC, "masksel", &Sel); 678 return BinaryOperator::CreateOr(T, NewSel); 679 } 680 681 // Cond ? (X | C) : (X & ~C) --> (X & ~C) | (Cond ? C : 0) 682 if (match(F, m_And(m_Value(X), m_APInt(NotC))) && 683 match(T, m_OneUse(m_Or(m_Specific(X), m_APInt(C)))) && *NotC == ~(*C)) { 684 Constant *Zero = ConstantInt::getNullValue(Ty); 685 Constant *OrC = ConstantInt::get(Ty, *C); 686 Value *NewSel = Builder.CreateSelect(Cond, OrC, Zero, "masksel", &Sel); 687 return BinaryOperator::CreateOr(F, NewSel); 688 } 689 690 return nullptr; 691 } 692 693 /// Transform patterns such as (a > b) ? a - b : 0 into usub.sat(a, b). 694 /// There are 8 commuted/swapped variants of this pattern. 695 /// TODO: Also support a - UMIN(a,b) patterns. 696 static Value *canonicalizeSaturatedSubtract(const ICmpInst *ICI, 697 const Value *TrueVal, 698 const Value *FalseVal, 699 InstCombiner::BuilderTy &Builder) { 700 ICmpInst::Predicate Pred = ICI->getPredicate(); 701 if (!ICmpInst::isUnsigned(Pred)) 702 return nullptr; 703 704 // (b > a) ? 0 : a - b -> (b <= a) ? a - b : 0 705 if (match(TrueVal, m_Zero())) { 706 Pred = ICmpInst::getInversePredicate(Pred); 707 std::swap(TrueVal, FalseVal); 708 } 709 if (!match(FalseVal, m_Zero())) 710 return nullptr; 711 712 Value *A = ICI->getOperand(0); 713 Value *B = ICI->getOperand(1); 714 if (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_ULT) { 715 // (b < a) ? a - b : 0 -> (a > b) ? a - b : 0 716 std::swap(A, B); 717 Pred = ICmpInst::getSwappedPredicate(Pred); 718 } 719 720 assert((Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_UGT) && 721 "Unexpected isUnsigned predicate!"); 722 723 // Ensure the sub is of the form: 724 // (a > b) ? a - b : 0 -> usub.sat(a, b) 725 // (a > b) ? b - a : 0 -> -usub.sat(a, b) 726 // Checking for both a-b and a+(-b) as a constant. 727 bool IsNegative = false; 728 const APInt *C; 729 if (match(TrueVal, m_Sub(m_Specific(B), m_Specific(A))) || 730 (match(A, m_APInt(C)) && 731 match(TrueVal, m_Add(m_Specific(B), m_SpecificInt(-*C))))) 732 IsNegative = true; 733 else if (!match(TrueVal, m_Sub(m_Specific(A), m_Specific(B))) && 734 !(match(B, m_APInt(C)) && 735 match(TrueVal, m_Add(m_Specific(A), m_SpecificInt(-*C))))) 736 return nullptr; 737 738 // If we are adding a negate and the sub and icmp are used anywhere else, we 739 // would end up with more instructions. 740 if (IsNegative && !TrueVal->hasOneUse() && !ICI->hasOneUse()) 741 return nullptr; 742 743 // (a > b) ? a - b : 0 -> usub.sat(a, b) 744 // (a > b) ? b - a : 0 -> -usub.sat(a, b) 745 Value *Result = Builder.CreateBinaryIntrinsic(Intrinsic::usub_sat, A, B); 746 if (IsNegative) 747 Result = Builder.CreateNeg(Result); 748 return Result; 749 } 750 751 static Value *canonicalizeSaturatedAdd(ICmpInst *Cmp, Value *TVal, Value *FVal, 752 InstCombiner::BuilderTy &Builder) { 753 if (!Cmp->hasOneUse()) 754 return nullptr; 755 756 // Match unsigned saturated add with constant. 757 Value *Cmp0 = Cmp->getOperand(0); 758 Value *Cmp1 = Cmp->getOperand(1); 759 ICmpInst::Predicate Pred = Cmp->getPredicate(); 760 Value *X; 761 const APInt *C, *CmpC; 762 if (Pred == ICmpInst::ICMP_ULT && 763 match(TVal, m_Add(m_Value(X), m_APInt(C))) && X == Cmp0 && 764 match(FVal, m_AllOnes()) && match(Cmp1, m_APInt(CmpC)) && *CmpC == ~*C) { 765 // (X u< ~C) ? (X + C) : -1 --> uadd.sat(X, C) 766 return Builder.CreateBinaryIntrinsic( 767 Intrinsic::uadd_sat, X, ConstantInt::get(X->getType(), *C)); 768 } 769 770 // Match unsigned saturated add of 2 variables with an unnecessary 'not'. 771 // There are 8 commuted variants. 772 // Canonicalize -1 (saturated result) to true value of the select. 773 if (match(FVal, m_AllOnes())) { 774 std::swap(TVal, FVal); 775 Pred = CmpInst::getInversePredicate(Pred); 776 } 777 if (!match(TVal, m_AllOnes())) 778 return nullptr; 779 780 // Canonicalize predicate to less-than or less-or-equal-than. 781 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) { 782 std::swap(Cmp0, Cmp1); 783 Pred = CmpInst::getSwappedPredicate(Pred); 784 } 785 if (Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_ULE) 786 return nullptr; 787 788 // Match unsigned saturated add of 2 variables with an unnecessary 'not'. 789 // Strictness of the comparison is irrelevant. 790 Value *Y; 791 if (match(Cmp0, m_Not(m_Value(X))) && 792 match(FVal, m_c_Add(m_Specific(X), m_Value(Y))) && Y == Cmp1) { 793 // (~X u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y) 794 // (~X u< Y) ? -1 : (Y + X) --> uadd.sat(X, Y) 795 return Builder.CreateBinaryIntrinsic(Intrinsic::uadd_sat, X, Y); 796 } 797 // The 'not' op may be included in the sum but not the compare. 798 // Strictness of the comparison is irrelevant. 799 X = Cmp0; 800 Y = Cmp1; 801 if (match(FVal, m_c_Add(m_Not(m_Specific(X)), m_Specific(Y)))) { 802 // (X u< Y) ? -1 : (~X + Y) --> uadd.sat(~X, Y) 803 // (X u< Y) ? -1 : (Y + ~X) --> uadd.sat(Y, ~X) 804 BinaryOperator *BO = cast<BinaryOperator>(FVal); 805 return Builder.CreateBinaryIntrinsic( 806 Intrinsic::uadd_sat, BO->getOperand(0), BO->getOperand(1)); 807 } 808 // The overflow may be detected via the add wrapping round. 809 // This is only valid for strict comparison! 810 if (Pred == ICmpInst::ICMP_ULT && 811 match(Cmp0, m_c_Add(m_Specific(Cmp1), m_Value(Y))) && 812 match(FVal, m_c_Add(m_Specific(Cmp1), m_Specific(Y)))) { 813 // ((X + Y) u< X) ? -1 : (X + Y) --> uadd.sat(X, Y) 814 // ((X + Y) u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y) 815 return Builder.CreateBinaryIntrinsic(Intrinsic::uadd_sat, Cmp1, Y); 816 } 817 818 return nullptr; 819 } 820 821 /// Fold the following code sequence: 822 /// \code 823 /// int a = ctlz(x & -x); 824 // x ? 31 - a : a; 825 /// \code 826 /// 827 /// into: 828 /// cttz(x) 829 static Instruction *foldSelectCtlzToCttz(ICmpInst *ICI, Value *TrueVal, 830 Value *FalseVal, 831 InstCombiner::BuilderTy &Builder) { 832 unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits(); 833 if (!ICI->isEquality() || !match(ICI->getOperand(1), m_Zero())) 834 return nullptr; 835 836 if (ICI->getPredicate() == ICmpInst::ICMP_NE) 837 std::swap(TrueVal, FalseVal); 838 839 if (!match(FalseVal, 840 m_Xor(m_Deferred(TrueVal), m_SpecificInt(BitWidth - 1)))) 841 return nullptr; 842 843 if (!match(TrueVal, m_Intrinsic<Intrinsic::ctlz>())) 844 return nullptr; 845 846 Value *X = ICI->getOperand(0); 847 auto *II = cast<IntrinsicInst>(TrueVal); 848 if (!match(II->getOperand(0), m_c_And(m_Specific(X), m_Neg(m_Specific(X))))) 849 return nullptr; 850 851 Function *F = Intrinsic::getDeclaration(II->getModule(), Intrinsic::cttz, 852 II->getType()); 853 return CallInst::Create(F, {X, II->getArgOperand(1)}); 854 } 855 856 /// Attempt to fold a cttz/ctlz followed by a icmp plus select into a single 857 /// call to cttz/ctlz with flag 'is_zero_undef' cleared. 858 /// 859 /// For example, we can fold the following code sequence: 860 /// \code 861 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 true) 862 /// %1 = icmp ne i32 %x, 0 863 /// %2 = select i1 %1, i32 %0, i32 32 864 /// \code 865 /// 866 /// into: 867 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 false) 868 static Value *foldSelectCttzCtlz(ICmpInst *ICI, Value *TrueVal, Value *FalseVal, 869 InstCombiner::BuilderTy &Builder) { 870 ICmpInst::Predicate Pred = ICI->getPredicate(); 871 Value *CmpLHS = ICI->getOperand(0); 872 Value *CmpRHS = ICI->getOperand(1); 873 874 // Check if the condition value compares a value for equality against zero. 875 if (!ICI->isEquality() || !match(CmpRHS, m_Zero())) 876 return nullptr; 877 878 Value *SelectArg = FalseVal; 879 Value *ValueOnZero = TrueVal; 880 if (Pred == ICmpInst::ICMP_NE) 881 std::swap(SelectArg, ValueOnZero); 882 883 // Skip zero extend/truncate. 884 Value *Count = nullptr; 885 if (!match(SelectArg, m_ZExt(m_Value(Count))) && 886 !match(SelectArg, m_Trunc(m_Value(Count)))) 887 Count = SelectArg; 888 889 // Check that 'Count' is a call to intrinsic cttz/ctlz. Also check that the 890 // input to the cttz/ctlz is used as LHS for the compare instruction. 891 if (!match(Count, m_Intrinsic<Intrinsic::cttz>(m_Specific(CmpLHS))) && 892 !match(Count, m_Intrinsic<Intrinsic::ctlz>(m_Specific(CmpLHS)))) 893 return nullptr; 894 895 IntrinsicInst *II = cast<IntrinsicInst>(Count); 896 897 // Check if the value propagated on zero is a constant number equal to the 898 // sizeof in bits of 'Count'. 899 unsigned SizeOfInBits = Count->getType()->getScalarSizeInBits(); 900 if (match(ValueOnZero, m_SpecificInt(SizeOfInBits))) { 901 // Explicitly clear the 'undef_on_zero' flag. It's always valid to go from 902 // true to false on this flag, so we can replace it for all users. 903 II->setArgOperand(1, ConstantInt::getFalse(II->getContext())); 904 return SelectArg; 905 } 906 907 // The ValueOnZero is not the bitwidth. But if the cttz/ctlz (and optional 908 // zext/trunc) have one use (ending at the select), the cttz/ctlz result will 909 // not be used if the input is zero. Relax to 'undef_on_zero' for that case. 910 if (II->hasOneUse() && SelectArg->hasOneUse() && 911 !match(II->getArgOperand(1), m_One())) 912 II->setArgOperand(1, ConstantInt::getTrue(II->getContext())); 913 914 return nullptr; 915 } 916 917 /// Return true if we find and adjust an icmp+select pattern where the compare 918 /// is with a constant that can be incremented or decremented to match the 919 /// minimum or maximum idiom. 920 static bool adjustMinMax(SelectInst &Sel, ICmpInst &Cmp) { 921 ICmpInst::Predicate Pred = Cmp.getPredicate(); 922 Value *CmpLHS = Cmp.getOperand(0); 923 Value *CmpRHS = Cmp.getOperand(1); 924 Value *TrueVal = Sel.getTrueValue(); 925 Value *FalseVal = Sel.getFalseValue(); 926 927 // We may move or edit the compare, so make sure the select is the only user. 928 const APInt *CmpC; 929 if (!Cmp.hasOneUse() || !match(CmpRHS, m_APInt(CmpC))) 930 return false; 931 932 // These transforms only work for selects of integers or vector selects of 933 // integer vectors. 934 Type *SelTy = Sel.getType(); 935 auto *SelEltTy = dyn_cast<IntegerType>(SelTy->getScalarType()); 936 if (!SelEltTy || SelTy->isVectorTy() != Cmp.getType()->isVectorTy()) 937 return false; 938 939 Constant *AdjustedRHS; 940 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT) 941 AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC + 1); 942 else if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT) 943 AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC - 1); 944 else 945 return false; 946 947 // X > C ? X : C+1 --> X < C+1 ? C+1 : X 948 // X < C ? X : C-1 --> X > C-1 ? C-1 : X 949 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) || 950 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) { 951 ; // Nothing to do here. Values match without any sign/zero extension. 952 } 953 // Types do not match. Instead of calculating this with mixed types, promote 954 // all to the larger type. This enables scalar evolution to analyze this 955 // expression. 956 else if (CmpRHS->getType()->getScalarSizeInBits() < SelEltTy->getBitWidth()) { 957 Constant *SextRHS = ConstantExpr::getSExt(AdjustedRHS, SelTy); 958 959 // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X 960 // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X 961 // X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X 962 // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X 963 if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) && SextRHS == FalseVal) { 964 CmpLHS = TrueVal; 965 AdjustedRHS = SextRHS; 966 } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) && 967 SextRHS == TrueVal) { 968 CmpLHS = FalseVal; 969 AdjustedRHS = SextRHS; 970 } else if (Cmp.isUnsigned()) { 971 Constant *ZextRHS = ConstantExpr::getZExt(AdjustedRHS, SelTy); 972 // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X 973 // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X 974 // zext + signed compare cannot be changed: 975 // 0xff <s 0x00, but 0x00ff >s 0x0000 976 if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) && ZextRHS == FalseVal) { 977 CmpLHS = TrueVal; 978 AdjustedRHS = ZextRHS; 979 } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) && 980 ZextRHS == TrueVal) { 981 CmpLHS = FalseVal; 982 AdjustedRHS = ZextRHS; 983 } else { 984 return false; 985 } 986 } else { 987 return false; 988 } 989 } else { 990 return false; 991 } 992 993 Pred = ICmpInst::getSwappedPredicate(Pred); 994 CmpRHS = AdjustedRHS; 995 std::swap(FalseVal, TrueVal); 996 Cmp.setPredicate(Pred); 997 Cmp.setOperand(0, CmpLHS); 998 Cmp.setOperand(1, CmpRHS); 999 Sel.setOperand(1, TrueVal); 1000 Sel.setOperand(2, FalseVal); 1001 Sel.swapProfMetadata(); 1002 1003 // Move the compare instruction right before the select instruction. Otherwise 1004 // the sext/zext value may be defined after the compare instruction uses it. 1005 Cmp.moveBefore(&Sel); 1006 1007 return true; 1008 } 1009 1010 /// If this is an integer min/max (icmp + select) with a constant operand, 1011 /// create the canonical icmp for the min/max operation and canonicalize the 1012 /// constant to the 'false' operand of the select: 1013 /// select (icmp Pred X, C1), C2, X --> select (icmp Pred' X, C2), X, C2 1014 /// Note: if C1 != C2, this will change the icmp constant to the existing 1015 /// constant operand of the select. 1016 static Instruction *canonicalizeMinMaxWithConstant(SelectInst &Sel, 1017 ICmpInst &Cmp, 1018 InstCombinerImpl &IC) { 1019 if (!Cmp.hasOneUse() || !isa<Constant>(Cmp.getOperand(1))) 1020 return nullptr; 1021 1022 // Canonicalize the compare predicate based on whether we have min or max. 1023 Value *LHS, *RHS; 1024 SelectPatternResult SPR = matchSelectPattern(&Sel, LHS, RHS); 1025 if (!SelectPatternResult::isMinOrMax(SPR.Flavor)) 1026 return nullptr; 1027 1028 // Is this already canonical? 1029 ICmpInst::Predicate CanonicalPred = getMinMaxPred(SPR.Flavor); 1030 if (Cmp.getOperand(0) == LHS && Cmp.getOperand(1) == RHS && 1031 Cmp.getPredicate() == CanonicalPred) 1032 return nullptr; 1033 1034 // Bail out on unsimplified X-0 operand (due to some worklist management bug), 1035 // as this may cause an infinite combine loop. Let the sub be folded first. 1036 if (match(LHS, m_Sub(m_Value(), m_Zero())) || 1037 match(RHS, m_Sub(m_Value(), m_Zero()))) 1038 return nullptr; 1039 1040 // Create the canonical compare and plug it into the select. 1041 IC.replaceOperand(Sel, 0, IC.Builder.CreateICmp(CanonicalPred, LHS, RHS)); 1042 1043 // If the select operands did not change, we're done. 1044 if (Sel.getTrueValue() == LHS && Sel.getFalseValue() == RHS) 1045 return &Sel; 1046 1047 // If we are swapping the select operands, swap the metadata too. 1048 assert(Sel.getTrueValue() == RHS && Sel.getFalseValue() == LHS && 1049 "Unexpected results from matchSelectPattern"); 1050 Sel.swapValues(); 1051 Sel.swapProfMetadata(); 1052 return &Sel; 1053 } 1054 1055 static Instruction *canonicalizeAbsNabs(SelectInst &Sel, ICmpInst &Cmp, 1056 InstCombinerImpl &IC) { 1057 if (!Cmp.hasOneUse() || !isa<Constant>(Cmp.getOperand(1))) 1058 return nullptr; 1059 1060 Value *LHS, *RHS; 1061 SelectPatternFlavor SPF = matchSelectPattern(&Sel, LHS, RHS).Flavor; 1062 if (SPF != SelectPatternFlavor::SPF_ABS && 1063 SPF != SelectPatternFlavor::SPF_NABS) 1064 return nullptr; 1065 1066 // Note that NSW flag can only be propagated for normal, non-negated abs! 1067 bool IntMinIsPoison = SPF == SelectPatternFlavor::SPF_ABS && 1068 match(RHS, m_NSWNeg(m_Specific(LHS))); 1069 Constant *IntMinIsPoisonC = 1070 ConstantInt::get(Type::getInt1Ty(Sel.getContext()), IntMinIsPoison); 1071 Instruction *Abs = 1072 IC.Builder.CreateBinaryIntrinsic(Intrinsic::abs, LHS, IntMinIsPoisonC); 1073 1074 if (SPF == SelectPatternFlavor::SPF_NABS) 1075 return BinaryOperator::CreateNeg(Abs); // Always without NSW flag! 1076 1077 return IC.replaceInstUsesWith(Sel, Abs); 1078 } 1079 1080 /// If we have a select with an equality comparison, then we know the value in 1081 /// one of the arms of the select. See if substituting this value into an arm 1082 /// and simplifying the result yields the same value as the other arm. 1083 /// 1084 /// To make this transform safe, we must drop poison-generating flags 1085 /// (nsw, etc) if we simplified to a binop because the select may be guarding 1086 /// that poison from propagating. If the existing binop already had no 1087 /// poison-generating flags, then this transform can be done by instsimplify. 1088 /// 1089 /// Consider: 1090 /// %cmp = icmp eq i32 %x, 2147483647 1091 /// %add = add nsw i32 %x, 1 1092 /// %sel = select i1 %cmp, i32 -2147483648, i32 %add 1093 /// 1094 /// We can't replace %sel with %add unless we strip away the flags. 1095 /// TODO: Wrapping flags could be preserved in some cases with better analysis. 1096 Instruction *InstCombinerImpl::foldSelectValueEquivalence(SelectInst &Sel, 1097 ICmpInst &Cmp) { 1098 // Value equivalence substitution requires an all-or-nothing replacement. 1099 // It does not make sense for a vector compare where each lane is chosen 1100 // independently. 1101 if (!Cmp.isEquality() || Cmp.getType()->isVectorTy()) 1102 return nullptr; 1103 1104 // Canonicalize the pattern to ICMP_EQ by swapping the select operands. 1105 Value *TrueVal = Sel.getTrueValue(), *FalseVal = Sel.getFalseValue(); 1106 bool Swapped = false; 1107 if (Cmp.getPredicate() == ICmpInst::ICMP_NE) { 1108 std::swap(TrueVal, FalseVal); 1109 Swapped = true; 1110 } 1111 1112 // In X == Y ? f(X) : Z, try to evaluate f(Y) and replace the operand. 1113 // Make sure Y cannot be undef though, as we might pick different values for 1114 // undef in the icmp and in f(Y). Additionally, take care to avoid replacing 1115 // X == Y ? X : Z with X == Y ? Y : Z, as that would lead to an infinite 1116 // replacement cycle. 1117 Value *CmpLHS = Cmp.getOperand(0), *CmpRHS = Cmp.getOperand(1); 1118 if (TrueVal != CmpLHS && 1119 isGuaranteedNotToBeUndefOrPoison(CmpRHS, SQ.AC, &Sel, &DT)) { 1120 if (Value *V = SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, SQ, 1121 /* AllowRefinement */ true)) 1122 return replaceOperand(Sel, Swapped ? 2 : 1, V); 1123 1124 // Even if TrueVal does not simplify, we can directly replace a use of 1125 // CmpLHS with CmpRHS, as long as the instruction is not used anywhere 1126 // else and is safe to speculatively execute (we may end up executing it 1127 // with different operands, which should not cause side-effects or trigger 1128 // undefined behavior). Only do this if CmpRHS is a constant, as 1129 // profitability is not clear for other cases. 1130 // FIXME: The replacement could be performed recursively. 1131 if (match(CmpRHS, m_ImmConstant()) && !match(CmpLHS, m_ImmConstant())) 1132 if (auto *I = dyn_cast<Instruction>(TrueVal)) 1133 if (I->hasOneUse() && isSafeToSpeculativelyExecute(I)) 1134 for (Use &U : I->operands()) 1135 if (U == CmpLHS) { 1136 replaceUse(U, CmpRHS); 1137 return &Sel; 1138 } 1139 } 1140 if (TrueVal != CmpRHS && 1141 isGuaranteedNotToBeUndefOrPoison(CmpLHS, SQ.AC, &Sel, &DT)) 1142 if (Value *V = SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, SQ, 1143 /* AllowRefinement */ true)) 1144 return replaceOperand(Sel, Swapped ? 2 : 1, V); 1145 1146 auto *FalseInst = dyn_cast<Instruction>(FalseVal); 1147 if (!FalseInst) 1148 return nullptr; 1149 1150 // InstSimplify already performed this fold if it was possible subject to 1151 // current poison-generating flags. Try the transform again with 1152 // poison-generating flags temporarily dropped. 1153 bool WasNUW = false, WasNSW = false, WasExact = false, WasInBounds = false; 1154 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(FalseVal)) { 1155 WasNUW = OBO->hasNoUnsignedWrap(); 1156 WasNSW = OBO->hasNoSignedWrap(); 1157 FalseInst->setHasNoUnsignedWrap(false); 1158 FalseInst->setHasNoSignedWrap(false); 1159 } 1160 if (auto *PEO = dyn_cast<PossiblyExactOperator>(FalseVal)) { 1161 WasExact = PEO->isExact(); 1162 FalseInst->setIsExact(false); 1163 } 1164 if (auto *GEP = dyn_cast<GetElementPtrInst>(FalseVal)) { 1165 WasInBounds = GEP->isInBounds(); 1166 GEP->setIsInBounds(false); 1167 } 1168 1169 // Try each equivalence substitution possibility. 1170 // We have an 'EQ' comparison, so the select's false value will propagate. 1171 // Example: 1172 // (X == 42) ? 43 : (X + 1) --> (X == 42) ? (X + 1) : (X + 1) --> X + 1 1173 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, SQ, 1174 /* AllowRefinement */ false) == TrueVal || 1175 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, SQ, 1176 /* AllowRefinement */ false) == TrueVal) { 1177 return replaceInstUsesWith(Sel, FalseVal); 1178 } 1179 1180 // Restore poison-generating flags if the transform did not apply. 1181 if (WasNUW) 1182 FalseInst->setHasNoUnsignedWrap(); 1183 if (WasNSW) 1184 FalseInst->setHasNoSignedWrap(); 1185 if (WasExact) 1186 FalseInst->setIsExact(); 1187 if (WasInBounds) 1188 cast<GetElementPtrInst>(FalseInst)->setIsInBounds(); 1189 1190 return nullptr; 1191 } 1192 1193 // See if this is a pattern like: 1194 // %old_cmp1 = icmp slt i32 %x, C2 1195 // %old_replacement = select i1 %old_cmp1, i32 %target_low, i32 %target_high 1196 // %old_x_offseted = add i32 %x, C1 1197 // %old_cmp0 = icmp ult i32 %old_x_offseted, C0 1198 // %r = select i1 %old_cmp0, i32 %x, i32 %old_replacement 1199 // This can be rewritten as more canonical pattern: 1200 // %new_cmp1 = icmp slt i32 %x, -C1 1201 // %new_cmp2 = icmp sge i32 %x, C0-C1 1202 // %new_clamped_low = select i1 %new_cmp1, i32 %target_low, i32 %x 1203 // %r = select i1 %new_cmp2, i32 %target_high, i32 %new_clamped_low 1204 // Iff -C1 s<= C2 s<= C0-C1 1205 // Also ULT predicate can also be UGT iff C0 != -1 (+invert result) 1206 // SLT predicate can also be SGT iff C2 != INT_MAX (+invert res.) 1207 static Instruction *canonicalizeClampLike(SelectInst &Sel0, ICmpInst &Cmp0, 1208 InstCombiner::BuilderTy &Builder) { 1209 Value *X = Sel0.getTrueValue(); 1210 Value *Sel1 = Sel0.getFalseValue(); 1211 1212 // First match the condition of the outermost select. 1213 // Said condition must be one-use. 1214 if (!Cmp0.hasOneUse()) 1215 return nullptr; 1216 Value *Cmp00 = Cmp0.getOperand(0); 1217 Constant *C0; 1218 if (!match(Cmp0.getOperand(1), 1219 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C0)))) 1220 return nullptr; 1221 // Canonicalize Cmp0 into the form we expect. 1222 // FIXME: we shouldn't care about lanes that are 'undef' in the end? 1223 switch (Cmp0.getPredicate()) { 1224 case ICmpInst::Predicate::ICMP_ULT: 1225 break; // Great! 1226 case ICmpInst::Predicate::ICMP_ULE: 1227 // We'd have to increment C0 by one, and for that it must not have all-ones 1228 // element, but then it would have been canonicalized to 'ult' before 1229 // we get here. So we can't do anything useful with 'ule'. 1230 return nullptr; 1231 case ICmpInst::Predicate::ICMP_UGT: 1232 // We want to canonicalize it to 'ult', so we'll need to increment C0, 1233 // which again means it must not have any all-ones elements. 1234 if (!match(C0, 1235 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_NE, 1236 APInt::getAllOnesValue( 1237 C0->getType()->getScalarSizeInBits())))) 1238 return nullptr; // Can't do, have all-ones element[s]. 1239 C0 = InstCombiner::AddOne(C0); 1240 std::swap(X, Sel1); 1241 break; 1242 case ICmpInst::Predicate::ICMP_UGE: 1243 // The only way we'd get this predicate if this `icmp` has extra uses, 1244 // but then we won't be able to do this fold. 1245 return nullptr; 1246 default: 1247 return nullptr; // Unknown predicate. 1248 } 1249 1250 // Now that we've canonicalized the ICmp, we know the X we expect; 1251 // the select in other hand should be one-use. 1252 if (!Sel1->hasOneUse()) 1253 return nullptr; 1254 1255 // We now can finish matching the condition of the outermost select: 1256 // it should either be the X itself, or an addition of some constant to X. 1257 Constant *C1; 1258 if (Cmp00 == X) 1259 C1 = ConstantInt::getNullValue(Sel0.getType()); 1260 else if (!match(Cmp00, 1261 m_Add(m_Specific(X), 1262 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C1))))) 1263 return nullptr; 1264 1265 Value *Cmp1; 1266 ICmpInst::Predicate Pred1; 1267 Constant *C2; 1268 Value *ReplacementLow, *ReplacementHigh; 1269 if (!match(Sel1, m_Select(m_Value(Cmp1), m_Value(ReplacementLow), 1270 m_Value(ReplacementHigh))) || 1271 !match(Cmp1, 1272 m_ICmp(Pred1, m_Specific(X), 1273 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C2))))) 1274 return nullptr; 1275 1276 if (!Cmp1->hasOneUse() && (Cmp00 == X || !Cmp00->hasOneUse())) 1277 return nullptr; // Not enough one-use instructions for the fold. 1278 // FIXME: this restriction could be relaxed if Cmp1 can be reused as one of 1279 // two comparisons we'll need to build. 1280 1281 // Canonicalize Cmp1 into the form we expect. 1282 // FIXME: we shouldn't care about lanes that are 'undef' in the end? 1283 switch (Pred1) { 1284 case ICmpInst::Predicate::ICMP_SLT: 1285 break; 1286 case ICmpInst::Predicate::ICMP_SLE: 1287 // We'd have to increment C2 by one, and for that it must not have signed 1288 // max element, but then it would have been canonicalized to 'slt' before 1289 // we get here. So we can't do anything useful with 'sle'. 1290 return nullptr; 1291 case ICmpInst::Predicate::ICMP_SGT: 1292 // We want to canonicalize it to 'slt', so we'll need to increment C2, 1293 // which again means it must not have any signed max elements. 1294 if (!match(C2, 1295 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_NE, 1296 APInt::getSignedMaxValue( 1297 C2->getType()->getScalarSizeInBits())))) 1298 return nullptr; // Can't do, have signed max element[s]. 1299 C2 = InstCombiner::AddOne(C2); 1300 LLVM_FALLTHROUGH; 1301 case ICmpInst::Predicate::ICMP_SGE: 1302 // Also non-canonical, but here we don't need to change C2, 1303 // so we don't have any restrictions on C2, so we can just handle it. 1304 std::swap(ReplacementLow, ReplacementHigh); 1305 break; 1306 default: 1307 return nullptr; // Unknown predicate. 1308 } 1309 1310 // The thresholds of this clamp-like pattern. 1311 auto *ThresholdLowIncl = ConstantExpr::getNeg(C1); 1312 auto *ThresholdHighExcl = ConstantExpr::getSub(C0, C1); 1313 1314 // The fold has a precondition 1: C2 s>= ThresholdLow 1315 auto *Precond1 = ConstantExpr::getICmp(ICmpInst::Predicate::ICMP_SGE, C2, 1316 ThresholdLowIncl); 1317 if (!match(Precond1, m_One())) 1318 return nullptr; 1319 // The fold has a precondition 2: C2 s<= ThresholdHigh 1320 auto *Precond2 = ConstantExpr::getICmp(ICmpInst::Predicate::ICMP_SLE, C2, 1321 ThresholdHighExcl); 1322 if (!match(Precond2, m_One())) 1323 return nullptr; 1324 1325 // All good, finally emit the new pattern. 1326 Value *ShouldReplaceLow = Builder.CreateICmpSLT(X, ThresholdLowIncl); 1327 Value *ShouldReplaceHigh = Builder.CreateICmpSGE(X, ThresholdHighExcl); 1328 Value *MaybeReplacedLow = 1329 Builder.CreateSelect(ShouldReplaceLow, ReplacementLow, X); 1330 Instruction *MaybeReplacedHigh = 1331 SelectInst::Create(ShouldReplaceHigh, ReplacementHigh, MaybeReplacedLow); 1332 1333 return MaybeReplacedHigh; 1334 } 1335 1336 // If we have 1337 // %cmp = icmp [canonical predicate] i32 %x, C0 1338 // %r = select i1 %cmp, i32 %y, i32 C1 1339 // Where C0 != C1 and %x may be different from %y, see if the constant that we 1340 // will have if we flip the strictness of the predicate (i.e. without changing 1341 // the result) is identical to the C1 in select. If it matches we can change 1342 // original comparison to one with swapped predicate, reuse the constant, 1343 // and swap the hands of select. 1344 static Instruction * 1345 tryToReuseConstantFromSelectInComparison(SelectInst &Sel, ICmpInst &Cmp, 1346 InstCombinerImpl &IC) { 1347 ICmpInst::Predicate Pred; 1348 Value *X; 1349 Constant *C0; 1350 if (!match(&Cmp, m_OneUse(m_ICmp( 1351 Pred, m_Value(X), 1352 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C0)))))) 1353 return nullptr; 1354 1355 // If comparison predicate is non-relational, we won't be able to do anything. 1356 if (ICmpInst::isEquality(Pred)) 1357 return nullptr; 1358 1359 // If comparison predicate is non-canonical, then we certainly won't be able 1360 // to make it canonical; canonicalizeCmpWithConstant() already tried. 1361 if (!InstCombiner::isCanonicalPredicate(Pred)) 1362 return nullptr; 1363 1364 // If the [input] type of comparison and select type are different, lets abort 1365 // for now. We could try to compare constants with trunc/[zs]ext though. 1366 if (C0->getType() != Sel.getType()) 1367 return nullptr; 1368 1369 // FIXME: are there any magic icmp predicate+constant pairs we must not touch? 1370 1371 Value *SelVal0, *SelVal1; // We do not care which one is from where. 1372 match(&Sel, m_Select(m_Value(), m_Value(SelVal0), m_Value(SelVal1))); 1373 // At least one of these values we are selecting between must be a constant 1374 // else we'll never succeed. 1375 if (!match(SelVal0, m_AnyIntegralConstant()) && 1376 !match(SelVal1, m_AnyIntegralConstant())) 1377 return nullptr; 1378 1379 // Does this constant C match any of the `select` values? 1380 auto MatchesSelectValue = [SelVal0, SelVal1](Constant *C) { 1381 return C->isElementWiseEqual(SelVal0) || C->isElementWiseEqual(SelVal1); 1382 }; 1383 1384 // If C0 *already* matches true/false value of select, we are done. 1385 if (MatchesSelectValue(C0)) 1386 return nullptr; 1387 1388 // Check the constant we'd have with flipped-strictness predicate. 1389 auto FlippedStrictness = 1390 InstCombiner::getFlippedStrictnessPredicateAndConstant(Pred, C0); 1391 if (!FlippedStrictness) 1392 return nullptr; 1393 1394 // If said constant doesn't match either, then there is no hope, 1395 if (!MatchesSelectValue(FlippedStrictness->second)) 1396 return nullptr; 1397 1398 // It matched! Lets insert the new comparison just before select. 1399 InstCombiner::BuilderTy::InsertPointGuard Guard(IC.Builder); 1400 IC.Builder.SetInsertPoint(&Sel); 1401 1402 Pred = ICmpInst::getSwappedPredicate(Pred); // Yes, swapped. 1403 Value *NewCmp = IC.Builder.CreateICmp(Pred, X, FlippedStrictness->second, 1404 Cmp.getName() + ".inv"); 1405 IC.replaceOperand(Sel, 0, NewCmp); 1406 Sel.swapValues(); 1407 Sel.swapProfMetadata(); 1408 1409 return &Sel; 1410 } 1411 1412 /// Visit a SelectInst that has an ICmpInst as its first operand. 1413 Instruction *InstCombinerImpl::foldSelectInstWithICmp(SelectInst &SI, 1414 ICmpInst *ICI) { 1415 if (Instruction *NewSel = foldSelectValueEquivalence(SI, *ICI)) 1416 return NewSel; 1417 1418 if (Instruction *NewSel = canonicalizeMinMaxWithConstant(SI, *ICI, *this)) 1419 return NewSel; 1420 1421 if (Instruction *NewAbs = canonicalizeAbsNabs(SI, *ICI, *this)) 1422 return NewAbs; 1423 1424 if (Instruction *NewAbs = canonicalizeClampLike(SI, *ICI, Builder)) 1425 return NewAbs; 1426 1427 if (Instruction *NewSel = 1428 tryToReuseConstantFromSelectInComparison(SI, *ICI, *this)) 1429 return NewSel; 1430 1431 bool Changed = adjustMinMax(SI, *ICI); 1432 1433 if (Value *V = foldSelectICmpAnd(SI, ICI, Builder)) 1434 return replaceInstUsesWith(SI, V); 1435 1436 // NOTE: if we wanted to, this is where to detect integer MIN/MAX 1437 Value *TrueVal = SI.getTrueValue(); 1438 Value *FalseVal = SI.getFalseValue(); 1439 ICmpInst::Predicate Pred = ICI->getPredicate(); 1440 Value *CmpLHS = ICI->getOperand(0); 1441 Value *CmpRHS = ICI->getOperand(1); 1442 if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) { 1443 if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) { 1444 // Transform (X == C) ? X : Y -> (X == C) ? C : Y 1445 SI.setOperand(1, CmpRHS); 1446 Changed = true; 1447 } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) { 1448 // Transform (X != C) ? Y : X -> (X != C) ? Y : C 1449 SI.setOperand(2, CmpRHS); 1450 Changed = true; 1451 } 1452 } 1453 1454 // FIXME: This code is nearly duplicated in InstSimplify. Using/refactoring 1455 // decomposeBitTestICmp() might help. 1456 { 1457 unsigned BitWidth = 1458 DL.getTypeSizeInBits(TrueVal->getType()->getScalarType()); 1459 APInt MinSignedValue = APInt::getSignedMinValue(BitWidth); 1460 Value *X; 1461 const APInt *Y, *C; 1462 bool TrueWhenUnset; 1463 bool IsBitTest = false; 1464 if (ICmpInst::isEquality(Pred) && 1465 match(CmpLHS, m_And(m_Value(X), m_Power2(Y))) && 1466 match(CmpRHS, m_Zero())) { 1467 IsBitTest = true; 1468 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ; 1469 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) { 1470 X = CmpLHS; 1471 Y = &MinSignedValue; 1472 IsBitTest = true; 1473 TrueWhenUnset = false; 1474 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) { 1475 X = CmpLHS; 1476 Y = &MinSignedValue; 1477 IsBitTest = true; 1478 TrueWhenUnset = true; 1479 } 1480 if (IsBitTest) { 1481 Value *V = nullptr; 1482 // (X & Y) == 0 ? X : X ^ Y --> X & ~Y 1483 if (TrueWhenUnset && TrueVal == X && 1484 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 1485 V = Builder.CreateAnd(X, ~(*Y)); 1486 // (X & Y) != 0 ? X ^ Y : X --> X & ~Y 1487 else if (!TrueWhenUnset && FalseVal == X && 1488 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 1489 V = Builder.CreateAnd(X, ~(*Y)); 1490 // (X & Y) == 0 ? X ^ Y : X --> X | Y 1491 else if (TrueWhenUnset && FalseVal == X && 1492 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 1493 V = Builder.CreateOr(X, *Y); 1494 // (X & Y) != 0 ? X : X ^ Y --> X | Y 1495 else if (!TrueWhenUnset && TrueVal == X && 1496 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 1497 V = Builder.CreateOr(X, *Y); 1498 1499 if (V) 1500 return replaceInstUsesWith(SI, V); 1501 } 1502 } 1503 1504 if (Instruction *V = 1505 foldSelectICmpAndAnd(SI.getType(), ICI, TrueVal, FalseVal, Builder)) 1506 return V; 1507 1508 if (Instruction *V = foldSelectCtlzToCttz(ICI, TrueVal, FalseVal, Builder)) 1509 return V; 1510 1511 if (Value *V = foldSelectICmpAndOr(ICI, TrueVal, FalseVal, Builder)) 1512 return replaceInstUsesWith(SI, V); 1513 1514 if (Value *V = foldSelectICmpLshrAshr(ICI, TrueVal, FalseVal, Builder)) 1515 return replaceInstUsesWith(SI, V); 1516 1517 if (Value *V = foldSelectCttzCtlz(ICI, TrueVal, FalseVal, Builder)) 1518 return replaceInstUsesWith(SI, V); 1519 1520 if (Value *V = canonicalizeSaturatedSubtract(ICI, TrueVal, FalseVal, Builder)) 1521 return replaceInstUsesWith(SI, V); 1522 1523 if (Value *V = canonicalizeSaturatedAdd(ICI, TrueVal, FalseVal, Builder)) 1524 return replaceInstUsesWith(SI, V); 1525 1526 return Changed ? &SI : nullptr; 1527 } 1528 1529 /// SI is a select whose condition is a PHI node (but the two may be in 1530 /// different blocks). See if the true/false values (V) are live in all of the 1531 /// predecessor blocks of the PHI. For example, cases like this can't be mapped: 1532 /// 1533 /// X = phi [ C1, BB1], [C2, BB2] 1534 /// Y = add 1535 /// Z = select X, Y, 0 1536 /// 1537 /// because Y is not live in BB1/BB2. 1538 static bool canSelectOperandBeMappingIntoPredBlock(const Value *V, 1539 const SelectInst &SI) { 1540 // If the value is a non-instruction value like a constant or argument, it 1541 // can always be mapped. 1542 const Instruction *I = dyn_cast<Instruction>(V); 1543 if (!I) return true; 1544 1545 // If V is a PHI node defined in the same block as the condition PHI, we can 1546 // map the arguments. 1547 const PHINode *CondPHI = cast<PHINode>(SI.getCondition()); 1548 1549 if (const PHINode *VP = dyn_cast<PHINode>(I)) 1550 if (VP->getParent() == CondPHI->getParent()) 1551 return true; 1552 1553 // Otherwise, if the PHI and select are defined in the same block and if V is 1554 // defined in a different block, then we can transform it. 1555 if (SI.getParent() == CondPHI->getParent() && 1556 I->getParent() != CondPHI->getParent()) 1557 return true; 1558 1559 // Otherwise we have a 'hard' case and we can't tell without doing more 1560 // detailed dominator based analysis, punt. 1561 return false; 1562 } 1563 1564 /// We have an SPF (e.g. a min or max) of an SPF of the form: 1565 /// SPF2(SPF1(A, B), C) 1566 Instruction *InstCombinerImpl::foldSPFofSPF(Instruction *Inner, 1567 SelectPatternFlavor SPF1, Value *A, 1568 Value *B, Instruction &Outer, 1569 SelectPatternFlavor SPF2, 1570 Value *C) { 1571 if (Outer.getType() != Inner->getType()) 1572 return nullptr; 1573 1574 if (C == A || C == B) { 1575 // MAX(MAX(A, B), B) -> MAX(A, B) 1576 // MIN(MIN(a, b), a) -> MIN(a, b) 1577 // TODO: This could be done in instsimplify. 1578 if (SPF1 == SPF2 && SelectPatternResult::isMinOrMax(SPF1)) 1579 return replaceInstUsesWith(Outer, Inner); 1580 1581 // MAX(MIN(a, b), a) -> a 1582 // MIN(MAX(a, b), a) -> a 1583 // TODO: This could be done in instsimplify. 1584 if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) || 1585 (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) || 1586 (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) || 1587 (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN)) 1588 return replaceInstUsesWith(Outer, C); 1589 } 1590 1591 if (SPF1 == SPF2) { 1592 const APInt *CB, *CC; 1593 if (match(B, m_APInt(CB)) && match(C, m_APInt(CC))) { 1594 // MIN(MIN(A, 23), 97) -> MIN(A, 23) 1595 // MAX(MAX(A, 97), 23) -> MAX(A, 97) 1596 // TODO: This could be done in instsimplify. 1597 if ((SPF1 == SPF_UMIN && CB->ule(*CC)) || 1598 (SPF1 == SPF_SMIN && CB->sle(*CC)) || 1599 (SPF1 == SPF_UMAX && CB->uge(*CC)) || 1600 (SPF1 == SPF_SMAX && CB->sge(*CC))) 1601 return replaceInstUsesWith(Outer, Inner); 1602 1603 // MIN(MIN(A, 97), 23) -> MIN(A, 23) 1604 // MAX(MAX(A, 23), 97) -> MAX(A, 97) 1605 if ((SPF1 == SPF_UMIN && CB->ugt(*CC)) || 1606 (SPF1 == SPF_SMIN && CB->sgt(*CC)) || 1607 (SPF1 == SPF_UMAX && CB->ult(*CC)) || 1608 (SPF1 == SPF_SMAX && CB->slt(*CC))) { 1609 Outer.replaceUsesOfWith(Inner, A); 1610 return &Outer; 1611 } 1612 } 1613 } 1614 1615 // max(max(A, B), min(A, B)) --> max(A, B) 1616 // min(min(A, B), max(A, B)) --> min(A, B) 1617 // TODO: This could be done in instsimplify. 1618 if (SPF1 == SPF2 && 1619 ((SPF1 == SPF_UMIN && match(C, m_c_UMax(m_Specific(A), m_Specific(B)))) || 1620 (SPF1 == SPF_SMIN && match(C, m_c_SMax(m_Specific(A), m_Specific(B)))) || 1621 (SPF1 == SPF_UMAX && match(C, m_c_UMin(m_Specific(A), m_Specific(B)))) || 1622 (SPF1 == SPF_SMAX && match(C, m_c_SMin(m_Specific(A), m_Specific(B)))))) 1623 return replaceInstUsesWith(Outer, Inner); 1624 1625 // ABS(ABS(X)) -> ABS(X) 1626 // NABS(NABS(X)) -> NABS(X) 1627 // TODO: This could be done in instsimplify. 1628 if (SPF1 == SPF2 && (SPF1 == SPF_ABS || SPF1 == SPF_NABS)) { 1629 return replaceInstUsesWith(Outer, Inner); 1630 } 1631 1632 // ABS(NABS(X)) -> ABS(X) 1633 // NABS(ABS(X)) -> NABS(X) 1634 if ((SPF1 == SPF_ABS && SPF2 == SPF_NABS) || 1635 (SPF1 == SPF_NABS && SPF2 == SPF_ABS)) { 1636 SelectInst *SI = cast<SelectInst>(Inner); 1637 Value *NewSI = 1638 Builder.CreateSelect(SI->getCondition(), SI->getFalseValue(), 1639 SI->getTrueValue(), SI->getName(), SI); 1640 return replaceInstUsesWith(Outer, NewSI); 1641 } 1642 1643 auto IsFreeOrProfitableToInvert = 1644 [&](Value *V, Value *&NotV, bool &ElidesXor) { 1645 if (match(V, m_Not(m_Value(NotV)))) { 1646 // If V has at most 2 uses then we can get rid of the xor operation 1647 // entirely. 1648 ElidesXor |= !V->hasNUsesOrMore(3); 1649 return true; 1650 } 1651 1652 if (isFreeToInvert(V, !V->hasNUsesOrMore(3))) { 1653 NotV = nullptr; 1654 return true; 1655 } 1656 1657 return false; 1658 }; 1659 1660 Value *NotA, *NotB, *NotC; 1661 bool ElidesXor = false; 1662 1663 // MIN(MIN(~A, ~B), ~C) == ~MAX(MAX(A, B), C) 1664 // MIN(MAX(~A, ~B), ~C) == ~MAX(MIN(A, B), C) 1665 // MAX(MIN(~A, ~B), ~C) == ~MIN(MAX(A, B), C) 1666 // MAX(MAX(~A, ~B), ~C) == ~MIN(MIN(A, B), C) 1667 // 1668 // This transform is performance neutral if we can elide at least one xor from 1669 // the set of three operands, since we'll be tacking on an xor at the very 1670 // end. 1671 if (SelectPatternResult::isMinOrMax(SPF1) && 1672 SelectPatternResult::isMinOrMax(SPF2) && 1673 IsFreeOrProfitableToInvert(A, NotA, ElidesXor) && 1674 IsFreeOrProfitableToInvert(B, NotB, ElidesXor) && 1675 IsFreeOrProfitableToInvert(C, NotC, ElidesXor) && ElidesXor) { 1676 if (!NotA) 1677 NotA = Builder.CreateNot(A); 1678 if (!NotB) 1679 NotB = Builder.CreateNot(B); 1680 if (!NotC) 1681 NotC = Builder.CreateNot(C); 1682 1683 Value *NewInner = createMinMax(Builder, getInverseMinMaxFlavor(SPF1), NotA, 1684 NotB); 1685 Value *NewOuter = Builder.CreateNot( 1686 createMinMax(Builder, getInverseMinMaxFlavor(SPF2), NewInner, NotC)); 1687 return replaceInstUsesWith(Outer, NewOuter); 1688 } 1689 1690 return nullptr; 1691 } 1692 1693 /// Turn select C, (X + Y), (X - Y) --> (X + (select C, Y, (-Y))). 1694 /// This is even legal for FP. 1695 static Instruction *foldAddSubSelect(SelectInst &SI, 1696 InstCombiner::BuilderTy &Builder) { 1697 Value *CondVal = SI.getCondition(); 1698 Value *TrueVal = SI.getTrueValue(); 1699 Value *FalseVal = SI.getFalseValue(); 1700 auto *TI = dyn_cast<Instruction>(TrueVal); 1701 auto *FI = dyn_cast<Instruction>(FalseVal); 1702 if (!TI || !FI || !TI->hasOneUse() || !FI->hasOneUse()) 1703 return nullptr; 1704 1705 Instruction *AddOp = nullptr, *SubOp = nullptr; 1706 if ((TI->getOpcode() == Instruction::Sub && 1707 FI->getOpcode() == Instruction::Add) || 1708 (TI->getOpcode() == Instruction::FSub && 1709 FI->getOpcode() == Instruction::FAdd)) { 1710 AddOp = FI; 1711 SubOp = TI; 1712 } else if ((FI->getOpcode() == Instruction::Sub && 1713 TI->getOpcode() == Instruction::Add) || 1714 (FI->getOpcode() == Instruction::FSub && 1715 TI->getOpcode() == Instruction::FAdd)) { 1716 AddOp = TI; 1717 SubOp = FI; 1718 } 1719 1720 if (AddOp) { 1721 Value *OtherAddOp = nullptr; 1722 if (SubOp->getOperand(0) == AddOp->getOperand(0)) { 1723 OtherAddOp = AddOp->getOperand(1); 1724 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) { 1725 OtherAddOp = AddOp->getOperand(0); 1726 } 1727 1728 if (OtherAddOp) { 1729 // So at this point we know we have (Y -> OtherAddOp): 1730 // select C, (add X, Y), (sub X, Z) 1731 Value *NegVal; // Compute -Z 1732 if (SI.getType()->isFPOrFPVectorTy()) { 1733 NegVal = Builder.CreateFNeg(SubOp->getOperand(1)); 1734 if (Instruction *NegInst = dyn_cast<Instruction>(NegVal)) { 1735 FastMathFlags Flags = AddOp->getFastMathFlags(); 1736 Flags &= SubOp->getFastMathFlags(); 1737 NegInst->setFastMathFlags(Flags); 1738 } 1739 } else { 1740 NegVal = Builder.CreateNeg(SubOp->getOperand(1)); 1741 } 1742 1743 Value *NewTrueOp = OtherAddOp; 1744 Value *NewFalseOp = NegVal; 1745 if (AddOp != TI) 1746 std::swap(NewTrueOp, NewFalseOp); 1747 Value *NewSel = Builder.CreateSelect(CondVal, NewTrueOp, NewFalseOp, 1748 SI.getName() + ".p", &SI); 1749 1750 if (SI.getType()->isFPOrFPVectorTy()) { 1751 Instruction *RI = 1752 BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel); 1753 1754 FastMathFlags Flags = AddOp->getFastMathFlags(); 1755 Flags &= SubOp->getFastMathFlags(); 1756 RI->setFastMathFlags(Flags); 1757 return RI; 1758 } else 1759 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel); 1760 } 1761 } 1762 return nullptr; 1763 } 1764 1765 /// Turn X + Y overflows ? -1 : X + Y -> uadd_sat X, Y 1766 /// And X - Y overflows ? 0 : X - Y -> usub_sat X, Y 1767 /// Along with a number of patterns similar to: 1768 /// X + Y overflows ? (X < 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1769 /// X - Y overflows ? (X > 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1770 static Instruction * 1771 foldOverflowingAddSubSelect(SelectInst &SI, InstCombiner::BuilderTy &Builder) { 1772 Value *CondVal = SI.getCondition(); 1773 Value *TrueVal = SI.getTrueValue(); 1774 Value *FalseVal = SI.getFalseValue(); 1775 1776 WithOverflowInst *II; 1777 if (!match(CondVal, m_ExtractValue<1>(m_WithOverflowInst(II))) || 1778 !match(FalseVal, m_ExtractValue<0>(m_Specific(II)))) 1779 return nullptr; 1780 1781 Value *X = II->getLHS(); 1782 Value *Y = II->getRHS(); 1783 1784 auto IsSignedSaturateLimit = [&](Value *Limit, bool IsAdd) { 1785 Type *Ty = Limit->getType(); 1786 1787 ICmpInst::Predicate Pred; 1788 Value *TrueVal, *FalseVal, *Op; 1789 const APInt *C; 1790 if (!match(Limit, m_Select(m_ICmp(Pred, m_Value(Op), m_APInt(C)), 1791 m_Value(TrueVal), m_Value(FalseVal)))) 1792 return false; 1793 1794 auto IsZeroOrOne = [](const APInt &C) { 1795 return C.isNullValue() || C.isOneValue(); 1796 }; 1797 auto IsMinMax = [&](Value *Min, Value *Max) { 1798 APInt MinVal = APInt::getSignedMinValue(Ty->getScalarSizeInBits()); 1799 APInt MaxVal = APInt::getSignedMaxValue(Ty->getScalarSizeInBits()); 1800 return match(Min, m_SpecificInt(MinVal)) && 1801 match(Max, m_SpecificInt(MaxVal)); 1802 }; 1803 1804 if (Op != X && Op != Y) 1805 return false; 1806 1807 if (IsAdd) { 1808 // X + Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1809 // X + Y overflows ? (X <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1810 // X + Y overflows ? (Y <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1811 // X + Y overflows ? (Y <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1812 if (Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C) && 1813 IsMinMax(TrueVal, FalseVal)) 1814 return true; 1815 // X + Y overflows ? (X >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1816 // X + Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1817 // X + Y overflows ? (Y >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1818 // X + Y overflows ? (Y >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1819 if (Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 1) && 1820 IsMinMax(FalseVal, TrueVal)) 1821 return true; 1822 } else { 1823 // X - Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1824 // X - Y overflows ? (X <s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1825 if (Op == X && Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C + 1) && 1826 IsMinMax(TrueVal, FalseVal)) 1827 return true; 1828 // X - Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1829 // X - Y overflows ? (X >s -2 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1830 if (Op == X && Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 2) && 1831 IsMinMax(FalseVal, TrueVal)) 1832 return true; 1833 // X - Y overflows ? (Y <s 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1834 // X - Y overflows ? (Y <s 1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1835 if (Op == Y && Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C) && 1836 IsMinMax(FalseVal, TrueVal)) 1837 return true; 1838 // X - Y overflows ? (Y >s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1839 // X - Y overflows ? (Y >s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1840 if (Op == Y && Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 1) && 1841 IsMinMax(TrueVal, FalseVal)) 1842 return true; 1843 } 1844 1845 return false; 1846 }; 1847 1848 Intrinsic::ID NewIntrinsicID; 1849 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow && 1850 match(TrueVal, m_AllOnes())) 1851 // X + Y overflows ? -1 : X + Y -> uadd_sat X, Y 1852 NewIntrinsicID = Intrinsic::uadd_sat; 1853 else if (II->getIntrinsicID() == Intrinsic::usub_with_overflow && 1854 match(TrueVal, m_Zero())) 1855 // X - Y overflows ? 0 : X - Y -> usub_sat X, Y 1856 NewIntrinsicID = Intrinsic::usub_sat; 1857 else if (II->getIntrinsicID() == Intrinsic::sadd_with_overflow && 1858 IsSignedSaturateLimit(TrueVal, /*IsAdd=*/true)) 1859 // X + Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1860 // X + Y overflows ? (X <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1861 // X + Y overflows ? (X >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1862 // X + Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1863 // X + Y overflows ? (Y <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1864 // X + Y overflows ? (Y <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1865 // X + Y overflows ? (Y >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1866 // X + Y overflows ? (Y >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1867 NewIntrinsicID = Intrinsic::sadd_sat; 1868 else if (II->getIntrinsicID() == Intrinsic::ssub_with_overflow && 1869 IsSignedSaturateLimit(TrueVal, /*IsAdd=*/false)) 1870 // X - Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1871 // X - Y overflows ? (X <s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1872 // X - Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1873 // X - Y overflows ? (X >s -2 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1874 // X - Y overflows ? (Y <s 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1875 // X - Y overflows ? (Y <s 1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1876 // X - Y overflows ? (Y >s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1877 // X - Y overflows ? (Y >s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1878 NewIntrinsicID = Intrinsic::ssub_sat; 1879 else 1880 return nullptr; 1881 1882 Function *F = 1883 Intrinsic::getDeclaration(SI.getModule(), NewIntrinsicID, SI.getType()); 1884 return CallInst::Create(F, {X, Y}); 1885 } 1886 1887 Instruction *InstCombinerImpl::foldSelectExtConst(SelectInst &Sel) { 1888 Constant *C; 1889 if (!match(Sel.getTrueValue(), m_Constant(C)) && 1890 !match(Sel.getFalseValue(), m_Constant(C))) 1891 return nullptr; 1892 1893 Instruction *ExtInst; 1894 if (!match(Sel.getTrueValue(), m_Instruction(ExtInst)) && 1895 !match(Sel.getFalseValue(), m_Instruction(ExtInst))) 1896 return nullptr; 1897 1898 auto ExtOpcode = ExtInst->getOpcode(); 1899 if (ExtOpcode != Instruction::ZExt && ExtOpcode != Instruction::SExt) 1900 return nullptr; 1901 1902 // If we are extending from a boolean type or if we can create a select that 1903 // has the same size operands as its condition, try to narrow the select. 1904 Value *X = ExtInst->getOperand(0); 1905 Type *SmallType = X->getType(); 1906 Value *Cond = Sel.getCondition(); 1907 auto *Cmp = dyn_cast<CmpInst>(Cond); 1908 if (!SmallType->isIntOrIntVectorTy(1) && 1909 (!Cmp || Cmp->getOperand(0)->getType() != SmallType)) 1910 return nullptr; 1911 1912 // If the constant is the same after truncation to the smaller type and 1913 // extension to the original type, we can narrow the select. 1914 Type *SelType = Sel.getType(); 1915 Constant *TruncC = ConstantExpr::getTrunc(C, SmallType); 1916 Constant *ExtC = ConstantExpr::getCast(ExtOpcode, TruncC, SelType); 1917 if (ExtC == C && ExtInst->hasOneUse()) { 1918 Value *TruncCVal = cast<Value>(TruncC); 1919 if (ExtInst == Sel.getFalseValue()) 1920 std::swap(X, TruncCVal); 1921 1922 // select Cond, (ext X), C --> ext(select Cond, X, C') 1923 // select Cond, C, (ext X) --> ext(select Cond, C', X) 1924 Value *NewSel = Builder.CreateSelect(Cond, X, TruncCVal, "narrow", &Sel); 1925 return CastInst::Create(Instruction::CastOps(ExtOpcode), NewSel, SelType); 1926 } 1927 1928 // If one arm of the select is the extend of the condition, replace that arm 1929 // with the extension of the appropriate known bool value. 1930 if (Cond == X) { 1931 if (ExtInst == Sel.getTrueValue()) { 1932 // select X, (sext X), C --> select X, -1, C 1933 // select X, (zext X), C --> select X, 1, C 1934 Constant *One = ConstantInt::getTrue(SmallType); 1935 Constant *AllOnesOrOne = ConstantExpr::getCast(ExtOpcode, One, SelType); 1936 return SelectInst::Create(Cond, AllOnesOrOne, C, "", nullptr, &Sel); 1937 } else { 1938 // select X, C, (sext X) --> select X, C, 0 1939 // select X, C, (zext X) --> select X, C, 0 1940 Constant *Zero = ConstantInt::getNullValue(SelType); 1941 return SelectInst::Create(Cond, C, Zero, "", nullptr, &Sel); 1942 } 1943 } 1944 1945 return nullptr; 1946 } 1947 1948 /// Try to transform a vector select with a constant condition vector into a 1949 /// shuffle for easier combining with other shuffles and insert/extract. 1950 static Instruction *canonicalizeSelectToShuffle(SelectInst &SI) { 1951 Value *CondVal = SI.getCondition(); 1952 Constant *CondC; 1953 auto *CondValTy = dyn_cast<FixedVectorType>(CondVal->getType()); 1954 if (!CondValTy || !match(CondVal, m_Constant(CondC))) 1955 return nullptr; 1956 1957 unsigned NumElts = CondValTy->getNumElements(); 1958 SmallVector<int, 16> Mask; 1959 Mask.reserve(NumElts); 1960 for (unsigned i = 0; i != NumElts; ++i) { 1961 Constant *Elt = CondC->getAggregateElement(i); 1962 if (!Elt) 1963 return nullptr; 1964 1965 if (Elt->isOneValue()) { 1966 // If the select condition element is true, choose from the 1st vector. 1967 Mask.push_back(i); 1968 } else if (Elt->isNullValue()) { 1969 // If the select condition element is false, choose from the 2nd vector. 1970 Mask.push_back(i + NumElts); 1971 } else if (isa<UndefValue>(Elt)) { 1972 // Undef in a select condition (choose one of the operands) does not mean 1973 // the same thing as undef in a shuffle mask (any value is acceptable), so 1974 // give up. 1975 return nullptr; 1976 } else { 1977 // Bail out on a constant expression. 1978 return nullptr; 1979 } 1980 } 1981 1982 return new ShuffleVectorInst(SI.getTrueValue(), SI.getFalseValue(), Mask); 1983 } 1984 1985 /// If we have a select of vectors with a scalar condition, try to convert that 1986 /// to a vector select by splatting the condition. A splat may get folded with 1987 /// other operations in IR and having all operands of a select be vector types 1988 /// is likely better for vector codegen. 1989 static Instruction *canonicalizeScalarSelectOfVecs(SelectInst &Sel, 1990 InstCombinerImpl &IC) { 1991 auto *Ty = dyn_cast<VectorType>(Sel.getType()); 1992 if (!Ty) 1993 return nullptr; 1994 1995 // We can replace a single-use extract with constant index. 1996 Value *Cond = Sel.getCondition(); 1997 if (!match(Cond, m_OneUse(m_ExtractElt(m_Value(), m_ConstantInt())))) 1998 return nullptr; 1999 2000 // select (extelt V, Index), T, F --> select (splat V, Index), T, F 2001 // Splatting the extracted condition reduces code (we could directly create a 2002 // splat shuffle of the source vector to eliminate the intermediate step). 2003 return IC.replaceOperand( 2004 Sel, 0, IC.Builder.CreateVectorSplat(Ty->getElementCount(), Cond)); 2005 } 2006 2007 /// Reuse bitcasted operands between a compare and select: 2008 /// select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) --> 2009 /// bitcast (select (cmp (bitcast C), (bitcast D)), (bitcast C), (bitcast D)) 2010 static Instruction *foldSelectCmpBitcasts(SelectInst &Sel, 2011 InstCombiner::BuilderTy &Builder) { 2012 Value *Cond = Sel.getCondition(); 2013 Value *TVal = Sel.getTrueValue(); 2014 Value *FVal = Sel.getFalseValue(); 2015 2016 CmpInst::Predicate Pred; 2017 Value *A, *B; 2018 if (!match(Cond, m_Cmp(Pred, m_Value(A), m_Value(B)))) 2019 return nullptr; 2020 2021 // The select condition is a compare instruction. If the select's true/false 2022 // values are already the same as the compare operands, there's nothing to do. 2023 if (TVal == A || TVal == B || FVal == A || FVal == B) 2024 return nullptr; 2025 2026 Value *C, *D; 2027 if (!match(A, m_BitCast(m_Value(C))) || !match(B, m_BitCast(m_Value(D)))) 2028 return nullptr; 2029 2030 // select (cmp (bitcast C), (bitcast D)), (bitcast TSrc), (bitcast FSrc) 2031 Value *TSrc, *FSrc; 2032 if (!match(TVal, m_BitCast(m_Value(TSrc))) || 2033 !match(FVal, m_BitCast(m_Value(FSrc)))) 2034 return nullptr; 2035 2036 // If the select true/false values are *different bitcasts* of the same source 2037 // operands, make the select operands the same as the compare operands and 2038 // cast the result. This is the canonical select form for min/max. 2039 Value *NewSel; 2040 if (TSrc == C && FSrc == D) { 2041 // select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) --> 2042 // bitcast (select (cmp A, B), A, B) 2043 NewSel = Builder.CreateSelect(Cond, A, B, "", &Sel); 2044 } else if (TSrc == D && FSrc == C) { 2045 // select (cmp (bitcast C), (bitcast D)), (bitcast' D), (bitcast' C) --> 2046 // bitcast (select (cmp A, B), B, A) 2047 NewSel = Builder.CreateSelect(Cond, B, A, "", &Sel); 2048 } else { 2049 return nullptr; 2050 } 2051 return CastInst::CreateBitOrPointerCast(NewSel, Sel.getType()); 2052 } 2053 2054 /// Try to eliminate select instructions that test the returned flag of cmpxchg 2055 /// instructions. 2056 /// 2057 /// If a select instruction tests the returned flag of a cmpxchg instruction and 2058 /// selects between the returned value of the cmpxchg instruction its compare 2059 /// operand, the result of the select will always be equal to its false value. 2060 /// For example: 2061 /// 2062 /// %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst 2063 /// %1 = extractvalue { i64, i1 } %0, 1 2064 /// %2 = extractvalue { i64, i1 } %0, 0 2065 /// %3 = select i1 %1, i64 %compare, i64 %2 2066 /// ret i64 %3 2067 /// 2068 /// The returned value of the cmpxchg instruction (%2) is the original value 2069 /// located at %ptr prior to any update. If the cmpxchg operation succeeds, %2 2070 /// must have been equal to %compare. Thus, the result of the select is always 2071 /// equal to %2, and the code can be simplified to: 2072 /// 2073 /// %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst 2074 /// %1 = extractvalue { i64, i1 } %0, 0 2075 /// ret i64 %1 2076 /// 2077 static Value *foldSelectCmpXchg(SelectInst &SI) { 2078 // A helper that determines if V is an extractvalue instruction whose 2079 // aggregate operand is a cmpxchg instruction and whose single index is equal 2080 // to I. If such conditions are true, the helper returns the cmpxchg 2081 // instruction; otherwise, a nullptr is returned. 2082 auto isExtractFromCmpXchg = [](Value *V, unsigned I) -> AtomicCmpXchgInst * { 2083 auto *Extract = dyn_cast<ExtractValueInst>(V); 2084 if (!Extract) 2085 return nullptr; 2086 if (Extract->getIndices()[0] != I) 2087 return nullptr; 2088 return dyn_cast<AtomicCmpXchgInst>(Extract->getAggregateOperand()); 2089 }; 2090 2091 // If the select has a single user, and this user is a select instruction that 2092 // we can simplify, skip the cmpxchg simplification for now. 2093 if (SI.hasOneUse()) 2094 if (auto *Select = dyn_cast<SelectInst>(SI.user_back())) 2095 if (Select->getCondition() == SI.getCondition()) 2096 if (Select->getFalseValue() == SI.getTrueValue() || 2097 Select->getTrueValue() == SI.getFalseValue()) 2098 return nullptr; 2099 2100 // Ensure the select condition is the returned flag of a cmpxchg instruction. 2101 auto *CmpXchg = isExtractFromCmpXchg(SI.getCondition(), 1); 2102 if (!CmpXchg) 2103 return nullptr; 2104 2105 // Check the true value case: The true value of the select is the returned 2106 // value of the same cmpxchg used by the condition, and the false value is the 2107 // cmpxchg instruction's compare operand. 2108 if (auto *X = isExtractFromCmpXchg(SI.getTrueValue(), 0)) 2109 if (X == CmpXchg && X->getCompareOperand() == SI.getFalseValue()) 2110 return SI.getFalseValue(); 2111 2112 // Check the false value case: The false value of the select is the returned 2113 // value of the same cmpxchg used by the condition, and the true value is the 2114 // cmpxchg instruction's compare operand. 2115 if (auto *X = isExtractFromCmpXchg(SI.getFalseValue(), 0)) 2116 if (X == CmpXchg && X->getCompareOperand() == SI.getTrueValue()) 2117 return SI.getFalseValue(); 2118 2119 return nullptr; 2120 } 2121 2122 static Instruction *moveAddAfterMinMax(SelectPatternFlavor SPF, Value *X, 2123 Value *Y, 2124 InstCombiner::BuilderTy &Builder) { 2125 assert(SelectPatternResult::isMinOrMax(SPF) && "Expected min/max pattern"); 2126 bool IsUnsigned = SPF == SelectPatternFlavor::SPF_UMIN || 2127 SPF == SelectPatternFlavor::SPF_UMAX; 2128 // TODO: If InstSimplify could fold all cases where C2 <= C1, we could change 2129 // the constant value check to an assert. 2130 Value *A; 2131 const APInt *C1, *C2; 2132 if (IsUnsigned && match(X, m_NUWAdd(m_Value(A), m_APInt(C1))) && 2133 match(Y, m_APInt(C2)) && C2->uge(*C1) && X->hasNUses(2)) { 2134 // umin (add nuw A, C1), C2 --> add nuw (umin A, C2 - C1), C1 2135 // umax (add nuw A, C1), C2 --> add nuw (umax A, C2 - C1), C1 2136 Value *NewMinMax = createMinMax(Builder, SPF, A, 2137 ConstantInt::get(X->getType(), *C2 - *C1)); 2138 return BinaryOperator::CreateNUW(BinaryOperator::Add, NewMinMax, 2139 ConstantInt::get(X->getType(), *C1)); 2140 } 2141 2142 if (!IsUnsigned && match(X, m_NSWAdd(m_Value(A), m_APInt(C1))) && 2143 match(Y, m_APInt(C2)) && X->hasNUses(2)) { 2144 bool Overflow; 2145 APInt Diff = C2->ssub_ov(*C1, Overflow); 2146 if (!Overflow) { 2147 // smin (add nsw A, C1), C2 --> add nsw (smin A, C2 - C1), C1 2148 // smax (add nsw A, C1), C2 --> add nsw (smax A, C2 - C1), C1 2149 Value *NewMinMax = createMinMax(Builder, SPF, A, 2150 ConstantInt::get(X->getType(), Diff)); 2151 return BinaryOperator::CreateNSW(BinaryOperator::Add, NewMinMax, 2152 ConstantInt::get(X->getType(), *C1)); 2153 } 2154 } 2155 2156 return nullptr; 2157 } 2158 2159 /// Match a sadd_sat or ssub_sat which is using min/max to clamp the value. 2160 Instruction *InstCombinerImpl::matchSAddSubSat(SelectInst &MinMax1) { 2161 Type *Ty = MinMax1.getType(); 2162 2163 // We are looking for a tree of: 2164 // max(INT_MIN, min(INT_MAX, add(sext(A), sext(B)))) 2165 // Where the min and max could be reversed 2166 Instruction *MinMax2; 2167 BinaryOperator *AddSub; 2168 const APInt *MinValue, *MaxValue; 2169 if (match(&MinMax1, m_SMin(m_Instruction(MinMax2), m_APInt(MaxValue)))) { 2170 if (!match(MinMax2, m_SMax(m_BinOp(AddSub), m_APInt(MinValue)))) 2171 return nullptr; 2172 } else if (match(&MinMax1, 2173 m_SMax(m_Instruction(MinMax2), m_APInt(MinValue)))) { 2174 if (!match(MinMax2, m_SMin(m_BinOp(AddSub), m_APInt(MaxValue)))) 2175 return nullptr; 2176 } else 2177 return nullptr; 2178 2179 // Check that the constants clamp a saturate, and that the new type would be 2180 // sensible to convert to. 2181 if (!(*MaxValue + 1).isPowerOf2() || -*MinValue != *MaxValue + 1) 2182 return nullptr; 2183 // In what bitwidth can this be treated as saturating arithmetics? 2184 unsigned NewBitWidth = (*MaxValue + 1).logBase2() + 1; 2185 // FIXME: This isn't quite right for vectors, but using the scalar type is a 2186 // good first approximation for what should be done there. 2187 if (!shouldChangeType(Ty->getScalarType()->getIntegerBitWidth(), NewBitWidth)) 2188 return nullptr; 2189 2190 // Also make sure that the number of uses is as expected. The "3"s are for the 2191 // the two items of min/max (the compare and the select). 2192 if (MinMax2->hasNUsesOrMore(3) || AddSub->hasNUsesOrMore(3)) 2193 return nullptr; 2194 2195 // Create the new type (which can be a vector type) 2196 Type *NewTy = Ty->getWithNewBitWidth(NewBitWidth); 2197 // Match the two extends from the add/sub 2198 Value *A, *B; 2199 if(!match(AddSub, m_BinOp(m_SExt(m_Value(A)), m_SExt(m_Value(B))))) 2200 return nullptr; 2201 // And check the incoming values are of a type smaller than or equal to the 2202 // size of the saturation. Otherwise the higher bits can cause different 2203 // results. 2204 if (A->getType()->getScalarSizeInBits() > NewBitWidth || 2205 B->getType()->getScalarSizeInBits() > NewBitWidth) 2206 return nullptr; 2207 2208 Intrinsic::ID IntrinsicID; 2209 if (AddSub->getOpcode() == Instruction::Add) 2210 IntrinsicID = Intrinsic::sadd_sat; 2211 else if (AddSub->getOpcode() == Instruction::Sub) 2212 IntrinsicID = Intrinsic::ssub_sat; 2213 else 2214 return nullptr; 2215 2216 // Finally create and return the sat intrinsic, truncated to the new type 2217 Function *F = Intrinsic::getDeclaration(MinMax1.getModule(), IntrinsicID, NewTy); 2218 Value *AT = Builder.CreateSExt(A, NewTy); 2219 Value *BT = Builder.CreateSExt(B, NewTy); 2220 Value *Sat = Builder.CreateCall(F, {AT, BT}); 2221 return CastInst::Create(Instruction::SExt, Sat, Ty); 2222 } 2223 2224 /// Reduce a sequence of min/max with a common operand. 2225 static Instruction *factorizeMinMaxTree(SelectPatternFlavor SPF, Value *LHS, 2226 Value *RHS, 2227 InstCombiner::BuilderTy &Builder) { 2228 assert(SelectPatternResult::isMinOrMax(SPF) && "Expected a min/max"); 2229 // TODO: Allow FP min/max with nnan/nsz. 2230 if (!LHS->getType()->isIntOrIntVectorTy()) 2231 return nullptr; 2232 2233 // Match 3 of the same min/max ops. Example: umin(umin(), umin()). 2234 Value *A, *B, *C, *D; 2235 SelectPatternResult L = matchSelectPattern(LHS, A, B); 2236 SelectPatternResult R = matchSelectPattern(RHS, C, D); 2237 if (SPF != L.Flavor || L.Flavor != R.Flavor) 2238 return nullptr; 2239 2240 // Look for a common operand. The use checks are different than usual because 2241 // a min/max pattern typically has 2 uses of each op: 1 by the cmp and 1 by 2242 // the select. 2243 Value *MinMaxOp = nullptr; 2244 Value *ThirdOp = nullptr; 2245 if (!LHS->hasNUsesOrMore(3) && RHS->hasNUsesOrMore(3)) { 2246 // If the LHS is only used in this chain and the RHS is used outside of it, 2247 // reuse the RHS min/max because that will eliminate the LHS. 2248 if (D == A || C == A) { 2249 // min(min(a, b), min(c, a)) --> min(min(c, a), b) 2250 // min(min(a, b), min(a, d)) --> min(min(a, d), b) 2251 MinMaxOp = RHS; 2252 ThirdOp = B; 2253 } else if (D == B || C == B) { 2254 // min(min(a, b), min(c, b)) --> min(min(c, b), a) 2255 // min(min(a, b), min(b, d)) --> min(min(b, d), a) 2256 MinMaxOp = RHS; 2257 ThirdOp = A; 2258 } 2259 } else if (!RHS->hasNUsesOrMore(3)) { 2260 // Reuse the LHS. This will eliminate the RHS. 2261 if (D == A || D == B) { 2262 // min(min(a, b), min(c, a)) --> min(min(a, b), c) 2263 // min(min(a, b), min(c, b)) --> min(min(a, b), c) 2264 MinMaxOp = LHS; 2265 ThirdOp = C; 2266 } else if (C == A || C == B) { 2267 // min(min(a, b), min(b, d)) --> min(min(a, b), d) 2268 // min(min(a, b), min(c, b)) --> min(min(a, b), d) 2269 MinMaxOp = LHS; 2270 ThirdOp = D; 2271 } 2272 } 2273 if (!MinMaxOp || !ThirdOp) 2274 return nullptr; 2275 2276 CmpInst::Predicate P = getMinMaxPred(SPF); 2277 Value *CmpABC = Builder.CreateICmp(P, MinMaxOp, ThirdOp); 2278 return SelectInst::Create(CmpABC, MinMaxOp, ThirdOp); 2279 } 2280 2281 /// Try to reduce a funnel/rotate pattern that includes a compare and select 2282 /// into a funnel shift intrinsic. Example: 2283 /// rotl32(a, b) --> (b == 0 ? a : ((a >> (32 - b)) | (a << b))) 2284 /// --> call llvm.fshl.i32(a, a, b) 2285 /// fshl32(a, b, c) --> (c == 0 ? a : ((b >> (32 - c)) | (a << c))) 2286 /// --> call llvm.fshl.i32(a, b, c) 2287 /// fshr32(a, b, c) --> (c == 0 ? b : ((a >> (32 - c)) | (b << c))) 2288 /// --> call llvm.fshr.i32(a, b, c) 2289 static Instruction *foldSelectFunnelShift(SelectInst &Sel, 2290 InstCombiner::BuilderTy &Builder) { 2291 // This must be a power-of-2 type for a bitmasking transform to be valid. 2292 unsigned Width = Sel.getType()->getScalarSizeInBits(); 2293 if (!isPowerOf2_32(Width)) 2294 return nullptr; 2295 2296 BinaryOperator *Or0, *Or1; 2297 if (!match(Sel.getFalseValue(), m_OneUse(m_Or(m_BinOp(Or0), m_BinOp(Or1))))) 2298 return nullptr; 2299 2300 Value *SV0, *SV1, *SA0, *SA1; 2301 if (!match(Or0, m_OneUse(m_LogicalShift(m_Value(SV0), 2302 m_ZExtOrSelf(m_Value(SA0))))) || 2303 !match(Or1, m_OneUse(m_LogicalShift(m_Value(SV1), 2304 m_ZExtOrSelf(m_Value(SA1))))) || 2305 Or0->getOpcode() == Or1->getOpcode()) 2306 return nullptr; 2307 2308 // Canonicalize to or(shl(SV0, SA0), lshr(SV1, SA1)). 2309 if (Or0->getOpcode() == BinaryOperator::LShr) { 2310 std::swap(Or0, Or1); 2311 std::swap(SV0, SV1); 2312 std::swap(SA0, SA1); 2313 } 2314 assert(Or0->getOpcode() == BinaryOperator::Shl && 2315 Or1->getOpcode() == BinaryOperator::LShr && 2316 "Illegal or(shift,shift) pair"); 2317 2318 // Check the shift amounts to see if they are an opposite pair. 2319 Value *ShAmt; 2320 if (match(SA1, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(SA0))))) 2321 ShAmt = SA0; 2322 else if (match(SA0, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(SA1))))) 2323 ShAmt = SA1; 2324 else 2325 return nullptr; 2326 2327 // We should now have this pattern: 2328 // select ?, TVal, (or (shl SV0, SA0), (lshr SV1, SA1)) 2329 // The false value of the select must be a funnel-shift of the true value: 2330 // IsFShl -> TVal must be SV0 else TVal must be SV1. 2331 bool IsFshl = (ShAmt == SA0); 2332 Value *TVal = Sel.getTrueValue(); 2333 if ((IsFshl && TVal != SV0) || (!IsFshl && TVal != SV1)) 2334 return nullptr; 2335 2336 // Finally, see if the select is filtering out a shift-by-zero. 2337 Value *Cond = Sel.getCondition(); 2338 ICmpInst::Predicate Pred; 2339 if (!match(Cond, m_OneUse(m_ICmp(Pred, m_Specific(ShAmt), m_ZeroInt()))) || 2340 Pred != ICmpInst::ICMP_EQ) 2341 return nullptr; 2342 2343 // If this is not a rotate then the select was blocking poison from the 2344 // 'shift-by-zero' non-TVal, but a funnel shift won't - so freeze it. 2345 if (SV0 != SV1) { 2346 if (IsFshl && !llvm::isGuaranteedNotToBePoison(SV1)) 2347 SV1 = Builder.CreateFreeze(SV1); 2348 else if (!IsFshl && !llvm::isGuaranteedNotToBePoison(SV0)) 2349 SV0 = Builder.CreateFreeze(SV0); 2350 } 2351 2352 // This is a funnel/rotate that avoids shift-by-bitwidth UB in a suboptimal way. 2353 // Convert to funnel shift intrinsic. 2354 Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr; 2355 Function *F = Intrinsic::getDeclaration(Sel.getModule(), IID, Sel.getType()); 2356 ShAmt = Builder.CreateZExt(ShAmt, Sel.getType()); 2357 return IntrinsicInst::Create(F, { SV0, SV1, ShAmt }); 2358 } 2359 2360 static Instruction *foldSelectToCopysign(SelectInst &Sel, 2361 InstCombiner::BuilderTy &Builder) { 2362 Value *Cond = Sel.getCondition(); 2363 Value *TVal = Sel.getTrueValue(); 2364 Value *FVal = Sel.getFalseValue(); 2365 Type *SelType = Sel.getType(); 2366 2367 // Match select ?, TC, FC where the constants are equal but negated. 2368 // TODO: Generalize to handle a negated variable operand? 2369 const APFloat *TC, *FC; 2370 if (!match(TVal, m_APFloat(TC)) || !match(FVal, m_APFloat(FC)) || 2371 !abs(*TC).bitwiseIsEqual(abs(*FC))) 2372 return nullptr; 2373 2374 assert(TC != FC && "Expected equal select arms to simplify"); 2375 2376 Value *X; 2377 const APInt *C; 2378 bool IsTrueIfSignSet; 2379 ICmpInst::Predicate Pred; 2380 if (!match(Cond, m_OneUse(m_ICmp(Pred, m_BitCast(m_Value(X)), m_APInt(C)))) || 2381 !InstCombiner::isSignBitCheck(Pred, *C, IsTrueIfSignSet) || 2382 X->getType() != SelType) 2383 return nullptr; 2384 2385 // If needed, negate the value that will be the sign argument of the copysign: 2386 // (bitcast X) < 0 ? -TC : TC --> copysign(TC, X) 2387 // (bitcast X) < 0 ? TC : -TC --> copysign(TC, -X) 2388 // (bitcast X) >= 0 ? -TC : TC --> copysign(TC, -X) 2389 // (bitcast X) >= 0 ? TC : -TC --> copysign(TC, X) 2390 if (IsTrueIfSignSet ^ TC->isNegative()) 2391 X = Builder.CreateFNegFMF(X, &Sel); 2392 2393 // Canonicalize the magnitude argument as the positive constant since we do 2394 // not care about its sign. 2395 Value *MagArg = TC->isNegative() ? FVal : TVal; 2396 Function *F = Intrinsic::getDeclaration(Sel.getModule(), Intrinsic::copysign, 2397 Sel.getType()); 2398 Instruction *CopySign = IntrinsicInst::Create(F, { MagArg, X }); 2399 CopySign->setFastMathFlags(Sel.getFastMathFlags()); 2400 return CopySign; 2401 } 2402 2403 Instruction *InstCombinerImpl::foldVectorSelect(SelectInst &Sel) { 2404 auto *VecTy = dyn_cast<FixedVectorType>(Sel.getType()); 2405 if (!VecTy) 2406 return nullptr; 2407 2408 unsigned NumElts = VecTy->getNumElements(); 2409 APInt UndefElts(NumElts, 0); 2410 APInt AllOnesEltMask(APInt::getAllOnesValue(NumElts)); 2411 if (Value *V = SimplifyDemandedVectorElts(&Sel, AllOnesEltMask, UndefElts)) { 2412 if (V != &Sel) 2413 return replaceInstUsesWith(Sel, V); 2414 return &Sel; 2415 } 2416 2417 // A select of a "select shuffle" with a common operand can be rearranged 2418 // to select followed by "select shuffle". Because of poison, this only works 2419 // in the case of a shuffle with no undefined mask elements. 2420 Value *Cond = Sel.getCondition(); 2421 Value *TVal = Sel.getTrueValue(); 2422 Value *FVal = Sel.getFalseValue(); 2423 Value *X, *Y; 2424 ArrayRef<int> Mask; 2425 if (match(TVal, m_OneUse(m_Shuffle(m_Value(X), m_Value(Y), m_Mask(Mask)))) && 2426 !is_contained(Mask, UndefMaskElem) && 2427 cast<ShuffleVectorInst>(TVal)->isSelect()) { 2428 if (X == FVal) { 2429 // select Cond, (shuf_sel X, Y), X --> shuf_sel X, (select Cond, Y, X) 2430 Value *NewSel = Builder.CreateSelect(Cond, Y, X, "sel", &Sel); 2431 return new ShuffleVectorInst(X, NewSel, Mask); 2432 } 2433 if (Y == FVal) { 2434 // select Cond, (shuf_sel X, Y), Y --> shuf_sel (select Cond, X, Y), Y 2435 Value *NewSel = Builder.CreateSelect(Cond, X, Y, "sel", &Sel); 2436 return new ShuffleVectorInst(NewSel, Y, Mask); 2437 } 2438 } 2439 if (match(FVal, m_OneUse(m_Shuffle(m_Value(X), m_Value(Y), m_Mask(Mask)))) && 2440 !is_contained(Mask, UndefMaskElem) && 2441 cast<ShuffleVectorInst>(FVal)->isSelect()) { 2442 if (X == TVal) { 2443 // select Cond, X, (shuf_sel X, Y) --> shuf_sel X, (select Cond, X, Y) 2444 Value *NewSel = Builder.CreateSelect(Cond, X, Y, "sel", &Sel); 2445 return new ShuffleVectorInst(X, NewSel, Mask); 2446 } 2447 if (Y == TVal) { 2448 // select Cond, Y, (shuf_sel X, Y) --> shuf_sel (select Cond, Y, X), Y 2449 Value *NewSel = Builder.CreateSelect(Cond, Y, X, "sel", &Sel); 2450 return new ShuffleVectorInst(NewSel, Y, Mask); 2451 } 2452 } 2453 2454 return nullptr; 2455 } 2456 2457 static Instruction *foldSelectToPhiImpl(SelectInst &Sel, BasicBlock *BB, 2458 const DominatorTree &DT, 2459 InstCombiner::BuilderTy &Builder) { 2460 // Find the block's immediate dominator that ends with a conditional branch 2461 // that matches select's condition (maybe inverted). 2462 auto *IDomNode = DT[BB]->getIDom(); 2463 if (!IDomNode) 2464 return nullptr; 2465 BasicBlock *IDom = IDomNode->getBlock(); 2466 2467 Value *Cond = Sel.getCondition(); 2468 Value *IfTrue, *IfFalse; 2469 BasicBlock *TrueSucc, *FalseSucc; 2470 if (match(IDom->getTerminator(), 2471 m_Br(m_Specific(Cond), m_BasicBlock(TrueSucc), 2472 m_BasicBlock(FalseSucc)))) { 2473 IfTrue = Sel.getTrueValue(); 2474 IfFalse = Sel.getFalseValue(); 2475 } else if (match(IDom->getTerminator(), 2476 m_Br(m_Not(m_Specific(Cond)), m_BasicBlock(TrueSucc), 2477 m_BasicBlock(FalseSucc)))) { 2478 IfTrue = Sel.getFalseValue(); 2479 IfFalse = Sel.getTrueValue(); 2480 } else 2481 return nullptr; 2482 2483 // Make sure the branches are actually different. 2484 if (TrueSucc == FalseSucc) 2485 return nullptr; 2486 2487 // We want to replace select %cond, %a, %b with a phi that takes value %a 2488 // for all incoming edges that are dominated by condition `%cond == true`, 2489 // and value %b for edges dominated by condition `%cond == false`. If %a 2490 // or %b are also phis from the same basic block, we can go further and take 2491 // their incoming values from the corresponding blocks. 2492 BasicBlockEdge TrueEdge(IDom, TrueSucc); 2493 BasicBlockEdge FalseEdge(IDom, FalseSucc); 2494 DenseMap<BasicBlock *, Value *> Inputs; 2495 for (auto *Pred : predecessors(BB)) { 2496 // Check implication. 2497 BasicBlockEdge Incoming(Pred, BB); 2498 if (DT.dominates(TrueEdge, Incoming)) 2499 Inputs[Pred] = IfTrue->DoPHITranslation(BB, Pred); 2500 else if (DT.dominates(FalseEdge, Incoming)) 2501 Inputs[Pred] = IfFalse->DoPHITranslation(BB, Pred); 2502 else 2503 return nullptr; 2504 // Check availability. 2505 if (auto *Insn = dyn_cast<Instruction>(Inputs[Pred])) 2506 if (!DT.dominates(Insn, Pred->getTerminator())) 2507 return nullptr; 2508 } 2509 2510 Builder.SetInsertPoint(&*BB->begin()); 2511 auto *PN = Builder.CreatePHI(Sel.getType(), Inputs.size()); 2512 for (auto *Pred : predecessors(BB)) 2513 PN->addIncoming(Inputs[Pred], Pred); 2514 PN->takeName(&Sel); 2515 return PN; 2516 } 2517 2518 static Instruction *foldSelectToPhi(SelectInst &Sel, const DominatorTree &DT, 2519 InstCombiner::BuilderTy &Builder) { 2520 // Try to replace this select with Phi in one of these blocks. 2521 SmallSetVector<BasicBlock *, 4> CandidateBlocks; 2522 CandidateBlocks.insert(Sel.getParent()); 2523 for (Value *V : Sel.operands()) 2524 if (auto *I = dyn_cast<Instruction>(V)) 2525 CandidateBlocks.insert(I->getParent()); 2526 2527 for (BasicBlock *BB : CandidateBlocks) 2528 if (auto *PN = foldSelectToPhiImpl(Sel, BB, DT, Builder)) 2529 return PN; 2530 return nullptr; 2531 } 2532 2533 static Value *foldSelectWithFrozenICmp(SelectInst &Sel, InstCombiner::BuilderTy &Builder) { 2534 FreezeInst *FI = dyn_cast<FreezeInst>(Sel.getCondition()); 2535 if (!FI) 2536 return nullptr; 2537 2538 Value *Cond = FI->getOperand(0); 2539 Value *TrueVal = Sel.getTrueValue(), *FalseVal = Sel.getFalseValue(); 2540 2541 // select (freeze(x == y)), x, y --> y 2542 // select (freeze(x != y)), x, y --> x 2543 // The freeze should be only used by this select. Otherwise, remaining uses of 2544 // the freeze can observe a contradictory value. 2545 // c = freeze(x == y) ; Let's assume that y = poison & x = 42; c is 0 or 1 2546 // a = select c, x, y ; 2547 // f(a, c) ; f(poison, 1) cannot happen, but if a is folded 2548 // ; to y, this can happen. 2549 CmpInst::Predicate Pred; 2550 if (FI->hasOneUse() && 2551 match(Cond, m_c_ICmp(Pred, m_Specific(TrueVal), m_Specific(FalseVal))) && 2552 (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE)) { 2553 return Pred == ICmpInst::ICMP_EQ ? FalseVal : TrueVal; 2554 } 2555 2556 return nullptr; 2557 } 2558 2559 Instruction *InstCombinerImpl::visitSelectInst(SelectInst &SI) { 2560 Value *CondVal = SI.getCondition(); 2561 Value *TrueVal = SI.getTrueValue(); 2562 Value *FalseVal = SI.getFalseValue(); 2563 Type *SelType = SI.getType(); 2564 2565 // FIXME: Remove this workaround when freeze related patches are done. 2566 // For select with undef operand which feeds into an equality comparison, 2567 // don't simplify it so loop unswitch can know the equality comparison 2568 // may have an undef operand. This is a workaround for PR31652 caused by 2569 // descrepancy about branch on undef between LoopUnswitch and GVN. 2570 if (isa<UndefValue>(TrueVal) || isa<UndefValue>(FalseVal)) { 2571 if (llvm::any_of(SI.users(), [&](User *U) { 2572 ICmpInst *CI = dyn_cast<ICmpInst>(U); 2573 if (CI && CI->isEquality()) 2574 return true; 2575 return false; 2576 })) { 2577 return nullptr; 2578 } 2579 } 2580 2581 if (Value *V = SimplifySelectInst(CondVal, TrueVal, FalseVal, 2582 SQ.getWithInstruction(&SI))) 2583 return replaceInstUsesWith(SI, V); 2584 2585 if (Instruction *I = canonicalizeSelectToShuffle(SI)) 2586 return I; 2587 2588 if (Instruction *I = canonicalizeScalarSelectOfVecs(SI, *this)) 2589 return I; 2590 2591 CmpInst::Predicate Pred; 2592 2593 if (SelType->isIntOrIntVectorTy(1) && 2594 TrueVal->getType() == CondVal->getType()) { 2595 if (match(TrueVal, m_One()) && 2596 (EnableUnsafeSelectTransform || impliesPoison(FalseVal, CondVal))) { 2597 // Change: A = select B, true, C --> A = or B, C 2598 return BinaryOperator::CreateOr(CondVal, FalseVal); 2599 } 2600 if (match(FalseVal, m_Zero()) && 2601 (EnableUnsafeSelectTransform || impliesPoison(TrueVal, CondVal))) { 2602 // Change: A = select B, C, false --> A = and B, C 2603 return BinaryOperator::CreateAnd(CondVal, TrueVal); 2604 } 2605 2606 // select a, false, b -> select !a, b, false 2607 if (match(TrueVal, m_Zero())) { 2608 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName()); 2609 return SelectInst::Create(NotCond, FalseVal, 2610 ConstantInt::getFalse(SelType)); 2611 } 2612 // select a, b, true -> select !a, true, b 2613 if (match(FalseVal, m_One())) { 2614 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName()); 2615 return SelectInst::Create(NotCond, ConstantInt::getTrue(SelType), 2616 TrueVal); 2617 } 2618 2619 // select a, a, b -> select a, true, b 2620 if (CondVal == TrueVal) 2621 return replaceOperand(SI, 1, ConstantInt::getTrue(SelType)); 2622 // select a, b, a -> select a, b, false 2623 if (CondVal == FalseVal) 2624 return replaceOperand(SI, 2, ConstantInt::getFalse(SelType)); 2625 2626 // select a, !a, b -> select !a, b, false 2627 if (match(TrueVal, m_Not(m_Specific(CondVal)))) 2628 return SelectInst::Create(TrueVal, FalseVal, 2629 ConstantInt::getFalse(SelType)); 2630 // select a, b, !a -> select !a, true, b 2631 if (match(FalseVal, m_Not(m_Specific(CondVal)))) 2632 return SelectInst::Create(FalseVal, ConstantInt::getTrue(SelType), 2633 TrueVal); 2634 } 2635 2636 // Selecting between two integer or vector splat integer constants? 2637 // 2638 // Note that we don't handle a scalar select of vectors: 2639 // select i1 %c, <2 x i8> <1, 1>, <2 x i8> <0, 0> 2640 // because that may need 3 instructions to splat the condition value: 2641 // extend, insertelement, shufflevector. 2642 // 2643 // Do not handle i1 TrueVal and FalseVal otherwise would result in 2644 // zext/sext i1 to i1. 2645 if (SelType->isIntOrIntVectorTy() && !SelType->isIntOrIntVectorTy(1) && 2646 CondVal->getType()->isVectorTy() == SelType->isVectorTy()) { 2647 // select C, 1, 0 -> zext C to int 2648 if (match(TrueVal, m_One()) && match(FalseVal, m_Zero())) 2649 return new ZExtInst(CondVal, SelType); 2650 2651 // select C, -1, 0 -> sext C to int 2652 if (match(TrueVal, m_AllOnes()) && match(FalseVal, m_Zero())) 2653 return new SExtInst(CondVal, SelType); 2654 2655 // select C, 0, 1 -> zext !C to int 2656 if (match(TrueVal, m_Zero()) && match(FalseVal, m_One())) { 2657 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName()); 2658 return new ZExtInst(NotCond, SelType); 2659 } 2660 2661 // select C, 0, -1 -> sext !C to int 2662 if (match(TrueVal, m_Zero()) && match(FalseVal, m_AllOnes())) { 2663 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName()); 2664 return new SExtInst(NotCond, SelType); 2665 } 2666 } 2667 2668 // See if we are selecting two values based on a comparison of the two values. 2669 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) { 2670 Value *Cmp0 = FCI->getOperand(0), *Cmp1 = FCI->getOperand(1); 2671 if ((Cmp0 == TrueVal && Cmp1 == FalseVal) || 2672 (Cmp0 == FalseVal && Cmp1 == TrueVal)) { 2673 // Canonicalize to use ordered comparisons by swapping the select 2674 // operands. 2675 // 2676 // e.g. 2677 // (X ugt Y) ? X : Y -> (X ole Y) ? Y : X 2678 if (FCI->hasOneUse() && FCmpInst::isUnordered(FCI->getPredicate())) { 2679 FCmpInst::Predicate InvPred = FCI->getInversePredicate(); 2680 IRBuilder<>::FastMathFlagGuard FMFG(Builder); 2681 // FIXME: The FMF should propagate from the select, not the fcmp. 2682 Builder.setFastMathFlags(FCI->getFastMathFlags()); 2683 Value *NewCond = Builder.CreateFCmp(InvPred, Cmp0, Cmp1, 2684 FCI->getName() + ".inv"); 2685 Value *NewSel = Builder.CreateSelect(NewCond, FalseVal, TrueVal); 2686 return replaceInstUsesWith(SI, NewSel); 2687 } 2688 2689 // NOTE: if we wanted to, this is where to detect MIN/MAX 2690 } 2691 } 2692 2693 // Canonicalize select with fcmp to fabs(). -0.0 makes this tricky. We need 2694 // fast-math-flags (nsz) or fsub with +0.0 (not fneg) for this to work. We 2695 // also require nnan because we do not want to unintentionally change the 2696 // sign of a NaN value. 2697 // FIXME: These folds should test/propagate FMF from the select, not the 2698 // fsub or fneg. 2699 // (X <= +/-0.0) ? (0.0 - X) : X --> fabs(X) 2700 Instruction *FSub; 2701 if (match(CondVal, m_FCmp(Pred, m_Specific(FalseVal), m_AnyZeroFP())) && 2702 match(TrueVal, m_FSub(m_PosZeroFP(), m_Specific(FalseVal))) && 2703 match(TrueVal, m_Instruction(FSub)) && FSub->hasNoNaNs() && 2704 (Pred == FCmpInst::FCMP_OLE || Pred == FCmpInst::FCMP_ULE)) { 2705 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, FalseVal, FSub); 2706 return replaceInstUsesWith(SI, Fabs); 2707 } 2708 // (X > +/-0.0) ? X : (0.0 - X) --> fabs(X) 2709 if (match(CondVal, m_FCmp(Pred, m_Specific(TrueVal), m_AnyZeroFP())) && 2710 match(FalseVal, m_FSub(m_PosZeroFP(), m_Specific(TrueVal))) && 2711 match(FalseVal, m_Instruction(FSub)) && FSub->hasNoNaNs() && 2712 (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_UGT)) { 2713 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, TrueVal, FSub); 2714 return replaceInstUsesWith(SI, Fabs); 2715 } 2716 // With nnan and nsz: 2717 // (X < +/-0.0) ? -X : X --> fabs(X) 2718 // (X <= +/-0.0) ? -X : X --> fabs(X) 2719 Instruction *FNeg; 2720 if (match(CondVal, m_FCmp(Pred, m_Specific(FalseVal), m_AnyZeroFP())) && 2721 match(TrueVal, m_FNeg(m_Specific(FalseVal))) && 2722 match(TrueVal, m_Instruction(FNeg)) && 2723 FNeg->hasNoNaNs() && FNeg->hasNoSignedZeros() && 2724 (Pred == FCmpInst::FCMP_OLT || Pred == FCmpInst::FCMP_OLE || 2725 Pred == FCmpInst::FCMP_ULT || Pred == FCmpInst::FCMP_ULE)) { 2726 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, FalseVal, FNeg); 2727 return replaceInstUsesWith(SI, Fabs); 2728 } 2729 // With nnan and nsz: 2730 // (X > +/-0.0) ? X : -X --> fabs(X) 2731 // (X >= +/-0.0) ? X : -X --> fabs(X) 2732 if (match(CondVal, m_FCmp(Pred, m_Specific(TrueVal), m_AnyZeroFP())) && 2733 match(FalseVal, m_FNeg(m_Specific(TrueVal))) && 2734 match(FalseVal, m_Instruction(FNeg)) && 2735 FNeg->hasNoNaNs() && FNeg->hasNoSignedZeros() && 2736 (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_OGE || 2737 Pred == FCmpInst::FCMP_UGT || Pred == FCmpInst::FCMP_UGE)) { 2738 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, TrueVal, FNeg); 2739 return replaceInstUsesWith(SI, Fabs); 2740 } 2741 2742 // See if we are selecting two values based on a comparison of the two values. 2743 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) 2744 if (Instruction *Result = foldSelectInstWithICmp(SI, ICI)) 2745 return Result; 2746 2747 if (Instruction *Add = foldAddSubSelect(SI, Builder)) 2748 return Add; 2749 if (Instruction *Add = foldOverflowingAddSubSelect(SI, Builder)) 2750 return Add; 2751 if (Instruction *Or = foldSetClearBits(SI, Builder)) 2752 return Or; 2753 2754 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z)) 2755 auto *TI = dyn_cast<Instruction>(TrueVal); 2756 auto *FI = dyn_cast<Instruction>(FalseVal); 2757 if (TI && FI && TI->getOpcode() == FI->getOpcode()) 2758 if (Instruction *IV = foldSelectOpOp(SI, TI, FI)) 2759 return IV; 2760 2761 if (Instruction *I = foldSelectExtConst(SI)) 2762 return I; 2763 2764 // See if we can fold the select into one of our operands. 2765 if (SelType->isIntOrIntVectorTy() || SelType->isFPOrFPVectorTy()) { 2766 if (Instruction *FoldI = foldSelectIntoOp(SI, TrueVal, FalseVal)) 2767 return FoldI; 2768 2769 Value *LHS, *RHS; 2770 Instruction::CastOps CastOp; 2771 SelectPatternResult SPR = matchSelectPattern(&SI, LHS, RHS, &CastOp); 2772 auto SPF = SPR.Flavor; 2773 if (SPF) { 2774 Value *LHS2, *RHS2; 2775 if (SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor) 2776 if (Instruction *R = foldSPFofSPF(cast<Instruction>(LHS), SPF2, LHS2, 2777 RHS2, SI, SPF, RHS)) 2778 return R; 2779 if (SelectPatternFlavor SPF2 = matchSelectPattern(RHS, LHS2, RHS2).Flavor) 2780 if (Instruction *R = foldSPFofSPF(cast<Instruction>(RHS), SPF2, LHS2, 2781 RHS2, SI, SPF, LHS)) 2782 return R; 2783 // TODO. 2784 // ABS(-X) -> ABS(X) 2785 } 2786 2787 if (SelectPatternResult::isMinOrMax(SPF)) { 2788 // Canonicalize so that 2789 // - type casts are outside select patterns. 2790 // - float clamp is transformed to min/max pattern 2791 2792 bool IsCastNeeded = LHS->getType() != SelType; 2793 Value *CmpLHS = cast<CmpInst>(CondVal)->getOperand(0); 2794 Value *CmpRHS = cast<CmpInst>(CondVal)->getOperand(1); 2795 if (IsCastNeeded || 2796 (LHS->getType()->isFPOrFPVectorTy() && 2797 ((CmpLHS != LHS && CmpLHS != RHS) || 2798 (CmpRHS != LHS && CmpRHS != RHS)))) { 2799 CmpInst::Predicate MinMaxPred = getMinMaxPred(SPF, SPR.Ordered); 2800 2801 Value *Cmp; 2802 if (CmpInst::isIntPredicate(MinMaxPred)) { 2803 Cmp = Builder.CreateICmp(MinMaxPred, LHS, RHS); 2804 } else { 2805 IRBuilder<>::FastMathFlagGuard FMFG(Builder); 2806 auto FMF = 2807 cast<FPMathOperator>(SI.getCondition())->getFastMathFlags(); 2808 Builder.setFastMathFlags(FMF); 2809 Cmp = Builder.CreateFCmp(MinMaxPred, LHS, RHS); 2810 } 2811 2812 Value *NewSI = Builder.CreateSelect(Cmp, LHS, RHS, SI.getName(), &SI); 2813 if (!IsCastNeeded) 2814 return replaceInstUsesWith(SI, NewSI); 2815 2816 Value *NewCast = Builder.CreateCast(CastOp, NewSI, SelType); 2817 return replaceInstUsesWith(SI, NewCast); 2818 } 2819 2820 // MAX(~a, ~b) -> ~MIN(a, b) 2821 // MAX(~a, C) -> ~MIN(a, ~C) 2822 // MIN(~a, ~b) -> ~MAX(a, b) 2823 // MIN(~a, C) -> ~MAX(a, ~C) 2824 auto moveNotAfterMinMax = [&](Value *X, Value *Y) -> Instruction * { 2825 Value *A; 2826 if (match(X, m_Not(m_Value(A))) && !X->hasNUsesOrMore(3) && 2827 !isFreeToInvert(A, A->hasOneUse()) && 2828 // Passing false to only consider m_Not and constants. 2829 isFreeToInvert(Y, false)) { 2830 Value *B = Builder.CreateNot(Y); 2831 Value *NewMinMax = createMinMax(Builder, getInverseMinMaxFlavor(SPF), 2832 A, B); 2833 // Copy the profile metadata. 2834 if (MDNode *MD = SI.getMetadata(LLVMContext::MD_prof)) { 2835 cast<SelectInst>(NewMinMax)->setMetadata(LLVMContext::MD_prof, MD); 2836 // Swap the metadata if the operands are swapped. 2837 if (X == SI.getFalseValue() && Y == SI.getTrueValue()) 2838 cast<SelectInst>(NewMinMax)->swapProfMetadata(); 2839 } 2840 2841 return BinaryOperator::CreateNot(NewMinMax); 2842 } 2843 2844 return nullptr; 2845 }; 2846 2847 if (Instruction *I = moveNotAfterMinMax(LHS, RHS)) 2848 return I; 2849 if (Instruction *I = moveNotAfterMinMax(RHS, LHS)) 2850 return I; 2851 2852 if (Instruction *I = moveAddAfterMinMax(SPF, LHS, RHS, Builder)) 2853 return I; 2854 2855 if (Instruction *I = factorizeMinMaxTree(SPF, LHS, RHS, Builder)) 2856 return I; 2857 if (Instruction *I = matchSAddSubSat(SI)) 2858 return I; 2859 } 2860 } 2861 2862 // Canonicalize select of FP values where NaN and -0.0 are not valid as 2863 // minnum/maxnum intrinsics. 2864 if (isa<FPMathOperator>(SI) && SI.hasNoNaNs() && SI.hasNoSignedZeros()) { 2865 Value *X, *Y; 2866 if (match(&SI, m_OrdFMax(m_Value(X), m_Value(Y)))) 2867 return replaceInstUsesWith( 2868 SI, Builder.CreateBinaryIntrinsic(Intrinsic::maxnum, X, Y, &SI)); 2869 2870 if (match(&SI, m_OrdFMin(m_Value(X), m_Value(Y)))) 2871 return replaceInstUsesWith( 2872 SI, Builder.CreateBinaryIntrinsic(Intrinsic::minnum, X, Y, &SI)); 2873 } 2874 2875 // See if we can fold the select into a phi node if the condition is a select. 2876 if (auto *PN = dyn_cast<PHINode>(SI.getCondition())) 2877 // The true/false values have to be live in the PHI predecessor's blocks. 2878 if (canSelectOperandBeMappingIntoPredBlock(TrueVal, SI) && 2879 canSelectOperandBeMappingIntoPredBlock(FalseVal, SI)) 2880 if (Instruction *NV = foldOpIntoPhi(SI, PN)) 2881 return NV; 2882 2883 if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) { 2884 if (TrueSI->getCondition()->getType() == CondVal->getType()) { 2885 // select(C, select(C, a, b), c) -> select(C, a, c) 2886 if (TrueSI->getCondition() == CondVal) { 2887 if (SI.getTrueValue() == TrueSI->getTrueValue()) 2888 return nullptr; 2889 return replaceOperand(SI, 1, TrueSI->getTrueValue()); 2890 } 2891 // select(C0, select(C1, a, b), b) -> select(C0&C1, a, b) 2892 // We choose this as normal form to enable folding on the And and 2893 // shortening paths for the values (this helps getUnderlyingObjects() for 2894 // example). 2895 if (TrueSI->getFalseValue() == FalseVal && TrueSI->hasOneUse()) { 2896 Value *And = Builder.CreateAnd(CondVal, TrueSI->getCondition()); 2897 replaceOperand(SI, 0, And); 2898 replaceOperand(SI, 1, TrueSI->getTrueValue()); 2899 return &SI; 2900 } 2901 } 2902 } 2903 if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) { 2904 if (FalseSI->getCondition()->getType() == CondVal->getType()) { 2905 // select(C, a, select(C, b, c)) -> select(C, a, c) 2906 if (FalseSI->getCondition() == CondVal) { 2907 if (SI.getFalseValue() == FalseSI->getFalseValue()) 2908 return nullptr; 2909 return replaceOperand(SI, 2, FalseSI->getFalseValue()); 2910 } 2911 // select(C0, a, select(C1, a, b)) -> select(C0|C1, a, b) 2912 if (FalseSI->getTrueValue() == TrueVal && FalseSI->hasOneUse()) { 2913 Value *Or = Builder.CreateOr(CondVal, FalseSI->getCondition()); 2914 replaceOperand(SI, 0, Or); 2915 replaceOperand(SI, 2, FalseSI->getFalseValue()); 2916 return &SI; 2917 } 2918 } 2919 } 2920 2921 auto canMergeSelectThroughBinop = [](BinaryOperator *BO) { 2922 // The select might be preventing a division by 0. 2923 switch (BO->getOpcode()) { 2924 default: 2925 return true; 2926 case Instruction::SRem: 2927 case Instruction::URem: 2928 case Instruction::SDiv: 2929 case Instruction::UDiv: 2930 return false; 2931 } 2932 }; 2933 2934 // Try to simplify a binop sandwiched between 2 selects with the same 2935 // condition. 2936 // select(C, binop(select(C, X, Y), W), Z) -> select(C, binop(X, W), Z) 2937 BinaryOperator *TrueBO; 2938 if (match(TrueVal, m_OneUse(m_BinOp(TrueBO))) && 2939 canMergeSelectThroughBinop(TrueBO)) { 2940 if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(0))) { 2941 if (TrueBOSI->getCondition() == CondVal) { 2942 replaceOperand(*TrueBO, 0, TrueBOSI->getTrueValue()); 2943 Worklist.push(TrueBO); 2944 return &SI; 2945 } 2946 } 2947 if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(1))) { 2948 if (TrueBOSI->getCondition() == CondVal) { 2949 replaceOperand(*TrueBO, 1, TrueBOSI->getTrueValue()); 2950 Worklist.push(TrueBO); 2951 return &SI; 2952 } 2953 } 2954 } 2955 2956 // select(C, Z, binop(select(C, X, Y), W)) -> select(C, Z, binop(Y, W)) 2957 BinaryOperator *FalseBO; 2958 if (match(FalseVal, m_OneUse(m_BinOp(FalseBO))) && 2959 canMergeSelectThroughBinop(FalseBO)) { 2960 if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(0))) { 2961 if (FalseBOSI->getCondition() == CondVal) { 2962 replaceOperand(*FalseBO, 0, FalseBOSI->getFalseValue()); 2963 Worklist.push(FalseBO); 2964 return &SI; 2965 } 2966 } 2967 if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(1))) { 2968 if (FalseBOSI->getCondition() == CondVal) { 2969 replaceOperand(*FalseBO, 1, FalseBOSI->getFalseValue()); 2970 Worklist.push(FalseBO); 2971 return &SI; 2972 } 2973 } 2974 } 2975 2976 Value *NotCond; 2977 if (match(CondVal, m_Not(m_Value(NotCond))) && 2978 !InstCombiner::shouldAvoidAbsorbingNotIntoSelect(SI)) { 2979 replaceOperand(SI, 0, NotCond); 2980 SI.swapValues(); 2981 SI.swapProfMetadata(); 2982 return &SI; 2983 } 2984 2985 if (Instruction *I = foldVectorSelect(SI)) 2986 return I; 2987 2988 // If we can compute the condition, there's no need for a select. 2989 // Like the above fold, we are attempting to reduce compile-time cost by 2990 // putting this fold here with limitations rather than in InstSimplify. 2991 // The motivation for this call into value tracking is to take advantage of 2992 // the assumption cache, so make sure that is populated. 2993 if (!CondVal->getType()->isVectorTy() && !AC.assumptions().empty()) { 2994 KnownBits Known(1); 2995 computeKnownBits(CondVal, Known, 0, &SI); 2996 if (Known.One.isOneValue()) 2997 return replaceInstUsesWith(SI, TrueVal); 2998 if (Known.Zero.isOneValue()) 2999 return replaceInstUsesWith(SI, FalseVal); 3000 } 3001 3002 if (Instruction *BitCastSel = foldSelectCmpBitcasts(SI, Builder)) 3003 return BitCastSel; 3004 3005 // Simplify selects that test the returned flag of cmpxchg instructions. 3006 if (Value *V = foldSelectCmpXchg(SI)) 3007 return replaceInstUsesWith(SI, V); 3008 3009 if (Instruction *Select = foldSelectBinOpIdentity(SI, TLI, *this)) 3010 return Select; 3011 3012 if (Instruction *Funnel = foldSelectFunnelShift(SI, Builder)) 3013 return Funnel; 3014 3015 if (Instruction *Copysign = foldSelectToCopysign(SI, Builder)) 3016 return Copysign; 3017 3018 if (Instruction *PN = foldSelectToPhi(SI, DT, Builder)) 3019 return replaceInstUsesWith(SI, PN); 3020 3021 if (Value *Fr = foldSelectWithFrozenICmp(SI, Builder)) 3022 return replaceInstUsesWith(SI, Fr); 3023 3024 return nullptr; 3025 } 3026