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/OverflowInstAnalysis.h" 22 #include "llvm/Analysis/ValueTracking.h" 23 #include "llvm/IR/BasicBlock.h" 24 #include "llvm/IR/Constant.h" 25 #include "llvm/IR/Constants.h" 26 #include "llvm/IR/DerivedTypes.h" 27 #include "llvm/IR/IRBuilder.h" 28 #include "llvm/IR/InstrTypes.h" 29 #include "llvm/IR/Instruction.h" 30 #include "llvm/IR/Instructions.h" 31 #include "llvm/IR/IntrinsicInst.h" 32 #include "llvm/IR/Intrinsics.h" 33 #include "llvm/IR/Operator.h" 34 #include "llvm/IR/PatternMatch.h" 35 #include "llvm/IR/Type.h" 36 #include "llvm/IR/User.h" 37 #include "llvm/IR/Value.h" 38 #include "llvm/Support/Casting.h" 39 #include "llvm/Support/ErrorHandling.h" 40 #include "llvm/Support/KnownBits.h" 41 #include "llvm/Transforms/InstCombine/InstCombineWorklist.h" 42 #include "llvm/Transforms/InstCombine/InstCombiner.h" 43 #include <cassert> 44 #include <utility> 45 46 using namespace llvm; 47 using namespace PatternMatch; 48 49 #define DEBUG_TYPE "instcombine" 50 51 static Value *createMinMax(InstCombiner::BuilderTy &Builder, 52 SelectPatternFlavor SPF, Value *A, Value *B) { 53 CmpInst::Predicate Pred = getMinMaxPred(SPF); 54 assert(CmpInst::isIntPredicate(Pred) && "Expected integer predicate"); 55 return Builder.CreateSelect(Builder.CreateICmp(Pred, A, B), A, B); 56 } 57 58 /// Replace a select operand based on an equality comparison with the identity 59 /// constant of a binop. 60 static Instruction *foldSelectBinOpIdentity(SelectInst &Sel, 61 const TargetLibraryInfo &TLI, 62 InstCombinerImpl &IC) { 63 // The select condition must be an equality compare with a constant operand. 64 Value *X; 65 Constant *C; 66 CmpInst::Predicate Pred; 67 if (!match(Sel.getCondition(), m_Cmp(Pred, m_Value(X), m_Constant(C)))) 68 return nullptr; 69 70 bool IsEq; 71 if (ICmpInst::isEquality(Pred)) 72 IsEq = Pred == ICmpInst::ICMP_EQ; 73 else if (Pred == FCmpInst::FCMP_OEQ) 74 IsEq = true; 75 else if (Pred == FCmpInst::FCMP_UNE) 76 IsEq = false; 77 else 78 return nullptr; 79 80 // A select operand must be a binop. 81 BinaryOperator *BO; 82 if (!match(Sel.getOperand(IsEq ? 1 : 2), m_BinOp(BO))) 83 return nullptr; 84 85 // The compare constant must be the identity constant for that binop. 86 // If this a floating-point compare with 0.0, any zero constant will do. 87 Type *Ty = BO->getType(); 88 Constant *IdC = ConstantExpr::getBinOpIdentity(BO->getOpcode(), Ty, true); 89 if (IdC != C) { 90 if (!IdC || !CmpInst::isFPPredicate(Pred)) 91 return nullptr; 92 if (!match(IdC, m_AnyZeroFP()) || !match(C, m_AnyZeroFP())) 93 return nullptr; 94 } 95 96 // Last, match the compare variable operand with a binop operand. 97 Value *Y; 98 if (!BO->isCommutative() && !match(BO, m_BinOp(m_Value(Y), m_Specific(X)))) 99 return nullptr; 100 if (!match(BO, m_c_BinOp(m_Value(Y), m_Specific(X)))) 101 return nullptr; 102 103 // +0.0 compares equal to -0.0, and so it does not behave as required for this 104 // transform. Bail out if we can not exclude that possibility. 105 if (isa<FPMathOperator>(BO)) 106 if (!BO->hasNoSignedZeros() && !CannotBeNegativeZero(Y, &TLI)) 107 return nullptr; 108 109 // BO = binop Y, X 110 // S = { select (cmp eq X, C), BO, ? } or { select (cmp ne X, C), ?, BO } 111 // => 112 // S = { select (cmp eq X, C), Y, ? } or { select (cmp ne X, C), ?, Y } 113 return IC.replaceOperand(Sel, IsEq ? 1 : 2, Y); 114 } 115 116 /// This folds: 117 /// select (icmp eq (and X, C1)), TC, FC 118 /// iff C1 is a power 2 and the difference between TC and FC is a power-of-2. 119 /// To something like: 120 /// (shr (and (X, C1)), (log2(C1) - log2(TC-FC))) + FC 121 /// Or: 122 /// (shl (and (X, C1)), (log2(TC-FC) - log2(C1))) + FC 123 /// With some variations depending if FC is larger than TC, or the shift 124 /// isn't needed, or the bit widths don't match. 125 static Value *foldSelectICmpAnd(SelectInst &Sel, ICmpInst *Cmp, 126 InstCombiner::BuilderTy &Builder) { 127 const APInt *SelTC, *SelFC; 128 if (!match(Sel.getTrueValue(), m_APInt(SelTC)) || 129 !match(Sel.getFalseValue(), m_APInt(SelFC))) 130 return nullptr; 131 132 // If this is a vector select, we need a vector compare. 133 Type *SelType = Sel.getType(); 134 if (SelType->isVectorTy() != Cmp->getType()->isVectorTy()) 135 return nullptr; 136 137 Value *V; 138 APInt AndMask; 139 bool CreateAnd = false; 140 ICmpInst::Predicate Pred = Cmp->getPredicate(); 141 if (ICmpInst::isEquality(Pred)) { 142 if (!match(Cmp->getOperand(1), m_Zero())) 143 return nullptr; 144 145 V = Cmp->getOperand(0); 146 const APInt *AndRHS; 147 if (!match(V, m_And(m_Value(), m_Power2(AndRHS)))) 148 return nullptr; 149 150 AndMask = *AndRHS; 151 } else if (decomposeBitTestICmp(Cmp->getOperand(0), Cmp->getOperand(1), 152 Pred, V, AndMask)) { 153 assert(ICmpInst::isEquality(Pred) && "Not equality test?"); 154 if (!AndMask.isPowerOf2()) 155 return nullptr; 156 157 CreateAnd = true; 158 } else { 159 return nullptr; 160 } 161 162 // In general, when both constants are non-zero, we would need an offset to 163 // replace the select. This would require more instructions than we started 164 // with. But there's one special-case that we handle here because it can 165 // simplify/reduce the instructions. 166 APInt TC = *SelTC; 167 APInt FC = *SelFC; 168 if (!TC.isNullValue() && !FC.isNullValue()) { 169 // If the select constants differ by exactly one bit and that's the same 170 // bit that is masked and checked by the select condition, the select can 171 // be replaced by bitwise logic to set/clear one bit of the constant result. 172 if (TC.getBitWidth() != AndMask.getBitWidth() || (TC ^ FC) != AndMask) 173 return nullptr; 174 if (CreateAnd) { 175 // If we have to create an 'and', then we must kill the cmp to not 176 // increase the instruction count. 177 if (!Cmp->hasOneUse()) 178 return nullptr; 179 V = Builder.CreateAnd(V, ConstantInt::get(SelType, AndMask)); 180 } 181 bool ExtraBitInTC = TC.ugt(FC); 182 if (Pred == ICmpInst::ICMP_EQ) { 183 // If the masked bit in V is clear, clear or set the bit in the result: 184 // (V & AndMaskC) == 0 ? TC : FC --> (V & AndMaskC) ^ TC 185 // (V & AndMaskC) == 0 ? TC : FC --> (V & AndMaskC) | TC 186 Constant *C = ConstantInt::get(SelType, TC); 187 return ExtraBitInTC ? Builder.CreateXor(V, C) : Builder.CreateOr(V, C); 188 } 189 if (Pred == ICmpInst::ICMP_NE) { 190 // If the masked bit in V is set, set or clear the bit in the result: 191 // (V & AndMaskC) != 0 ? TC : FC --> (V & AndMaskC) | FC 192 // (V & AndMaskC) != 0 ? TC : FC --> (V & AndMaskC) ^ FC 193 Constant *C = ConstantInt::get(SelType, FC); 194 return ExtraBitInTC ? Builder.CreateOr(V, C) : Builder.CreateXor(V, C); 195 } 196 llvm_unreachable("Only expecting equality predicates"); 197 } 198 199 // Make sure one of the select arms is a power-of-2. 200 if (!TC.isPowerOf2() && !FC.isPowerOf2()) 201 return nullptr; 202 203 // Determine which shift is needed to transform result of the 'and' into the 204 // desired result. 205 const APInt &ValC = !TC.isNullValue() ? TC : FC; 206 unsigned ValZeros = ValC.logBase2(); 207 unsigned AndZeros = AndMask.logBase2(); 208 209 // Insert the 'and' instruction on the input to the truncate. 210 if (CreateAnd) 211 V = Builder.CreateAnd(V, ConstantInt::get(V->getType(), AndMask)); 212 213 // If types don't match, we can still convert the select by introducing a zext 214 // or a trunc of the 'and'. 215 if (ValZeros > AndZeros) { 216 V = Builder.CreateZExtOrTrunc(V, SelType); 217 V = Builder.CreateShl(V, ValZeros - AndZeros); 218 } else if (ValZeros < AndZeros) { 219 V = Builder.CreateLShr(V, AndZeros - ValZeros); 220 V = Builder.CreateZExtOrTrunc(V, SelType); 221 } else { 222 V = Builder.CreateZExtOrTrunc(V, SelType); 223 } 224 225 // Okay, now we know that everything is set up, we just don't know whether we 226 // have a icmp_ne or icmp_eq and whether the true or false val is the zero. 227 bool ShouldNotVal = !TC.isNullValue(); 228 ShouldNotVal ^= Pred == ICmpInst::ICMP_NE; 229 if (ShouldNotVal) 230 V = Builder.CreateXor(V, ValC); 231 232 return V; 233 } 234 235 /// We want to turn code that looks like this: 236 /// %C = or %A, %B 237 /// %D = select %cond, %C, %A 238 /// into: 239 /// %C = select %cond, %B, 0 240 /// %D = or %A, %C 241 /// 242 /// Assuming that the specified instruction is an operand to the select, return 243 /// a bitmask indicating which operands of this instruction are foldable if they 244 /// equal the other incoming value of the select. 245 static unsigned getSelectFoldableOperands(BinaryOperator *I) { 246 switch (I->getOpcode()) { 247 case Instruction::Add: 248 case Instruction::Mul: 249 case Instruction::And: 250 case Instruction::Or: 251 case Instruction::Xor: 252 return 3; // Can fold through either operand. 253 case Instruction::Sub: // Can only fold on the amount subtracted. 254 case Instruction::Shl: // Can only fold on the shift amount. 255 case Instruction::LShr: 256 case Instruction::AShr: 257 return 1; 258 default: 259 return 0; // Cannot fold 260 } 261 } 262 263 /// We have (select c, TI, FI), and we know that TI and FI have the same opcode. 264 Instruction *InstCombinerImpl::foldSelectOpOp(SelectInst &SI, Instruction *TI, 265 Instruction *FI) { 266 // Don't break up min/max patterns. The hasOneUse checks below prevent that 267 // for most cases, but vector min/max with bitcasts can be transformed. If the 268 // one-use restrictions are eased for other patterns, we still don't want to 269 // obfuscate min/max. 270 if ((match(&SI, m_SMin(m_Value(), m_Value())) || 271 match(&SI, m_SMax(m_Value(), m_Value())) || 272 match(&SI, m_UMin(m_Value(), m_Value())) || 273 match(&SI, m_UMax(m_Value(), m_Value())))) 274 return nullptr; 275 276 // If this is a cast from the same type, merge. 277 Value *Cond = SI.getCondition(); 278 Type *CondTy = Cond->getType(); 279 if (TI->getNumOperands() == 1 && TI->isCast()) { 280 Type *FIOpndTy = FI->getOperand(0)->getType(); 281 if (TI->getOperand(0)->getType() != FIOpndTy) 282 return nullptr; 283 284 // The select condition may be a vector. We may only change the operand 285 // type if the vector width remains the same (and matches the condition). 286 if (auto *CondVTy = dyn_cast<VectorType>(CondTy)) { 287 if (!FIOpndTy->isVectorTy() || 288 CondVTy->getElementCount() != 289 cast<VectorType>(FIOpndTy)->getElementCount()) 290 return nullptr; 291 292 // TODO: If the backend knew how to deal with casts better, we could 293 // remove this limitation. For now, there's too much potential to create 294 // worse codegen by promoting the select ahead of size-altering casts 295 // (PR28160). 296 // 297 // Note that ValueTracking's matchSelectPattern() looks through casts 298 // without checking 'hasOneUse' when it matches min/max patterns, so this 299 // transform may end up happening anyway. 300 if (TI->getOpcode() != Instruction::BitCast && 301 (!TI->hasOneUse() || !FI->hasOneUse())) 302 return nullptr; 303 } else if (!TI->hasOneUse() || !FI->hasOneUse()) { 304 // TODO: The one-use restrictions for a scalar select could be eased if 305 // the fold of a select in visitLoadInst() was enhanced to match a pattern 306 // that includes a cast. 307 return nullptr; 308 } 309 310 // Fold this by inserting a select from the input values. 311 Value *NewSI = 312 Builder.CreateSelect(Cond, TI->getOperand(0), FI->getOperand(0), 313 SI.getName() + ".v", &SI); 314 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI, 315 TI->getType()); 316 } 317 318 // Cond ? -X : -Y --> -(Cond ? X : Y) 319 Value *X, *Y; 320 if (match(TI, m_FNeg(m_Value(X))) && match(FI, m_FNeg(m_Value(Y))) && 321 (TI->hasOneUse() || FI->hasOneUse())) { 322 Value *NewSel = Builder.CreateSelect(Cond, X, Y, SI.getName() + ".v", &SI); 323 return UnaryOperator::CreateFNegFMF(NewSel, TI); 324 } 325 326 // Min/max intrinsic with a common operand can have the common operand pulled 327 // after the select. This is the same transform as below for binops, but 328 // specialized for intrinsic matching and without the restrictive uses clause. 329 auto *TII = dyn_cast<IntrinsicInst>(TI); 330 auto *FII = dyn_cast<IntrinsicInst>(FI); 331 if (TII && FII && TII->getIntrinsicID() == FII->getIntrinsicID() && 332 (TII->hasOneUse() || FII->hasOneUse())) { 333 Value *T0, *T1, *F0, *F1; 334 if (match(TII, m_MaxOrMin(m_Value(T0), m_Value(T1))) && 335 match(FII, m_MaxOrMin(m_Value(F0), m_Value(F1)))) { 336 if (T0 == F0) { 337 Value *NewSel = Builder.CreateSelect(Cond, T1, F1, "minmaxop", &SI); 338 return CallInst::Create(TII->getCalledFunction(), {NewSel, T0}); 339 } 340 if (T0 == F1) { 341 Value *NewSel = Builder.CreateSelect(Cond, T1, F0, "minmaxop", &SI); 342 return CallInst::Create(TII->getCalledFunction(), {NewSel, T0}); 343 } 344 if (T1 == F0) { 345 Value *NewSel = Builder.CreateSelect(Cond, T0, F1, "minmaxop", &SI); 346 return CallInst::Create(TII->getCalledFunction(), {NewSel, T1}); 347 } 348 if (T1 == F1) { 349 Value *NewSel = Builder.CreateSelect(Cond, T0, F0, "minmaxop", &SI); 350 return CallInst::Create(TII->getCalledFunction(), {NewSel, T1}); 351 } 352 } 353 } 354 355 // Only handle binary operators (including two-operand getelementptr) with 356 // one-use here. As with the cast case above, it may be possible to relax the 357 // one-use constraint, but that needs be examined carefully since it may not 358 // reduce the total number of instructions. 359 if (TI->getNumOperands() != 2 || FI->getNumOperands() != 2 || 360 (!isa<BinaryOperator>(TI) && !isa<GetElementPtrInst>(TI)) || 361 !TI->hasOneUse() || !FI->hasOneUse()) 362 return nullptr; 363 364 // Figure out if the operations have any operands in common. 365 Value *MatchOp, *OtherOpT, *OtherOpF; 366 bool MatchIsOpZero; 367 if (TI->getOperand(0) == FI->getOperand(0)) { 368 MatchOp = TI->getOperand(0); 369 OtherOpT = TI->getOperand(1); 370 OtherOpF = FI->getOperand(1); 371 MatchIsOpZero = true; 372 } else if (TI->getOperand(1) == FI->getOperand(1)) { 373 MatchOp = TI->getOperand(1); 374 OtherOpT = TI->getOperand(0); 375 OtherOpF = FI->getOperand(0); 376 MatchIsOpZero = false; 377 } else if (!TI->isCommutative()) { 378 return nullptr; 379 } else if (TI->getOperand(0) == FI->getOperand(1)) { 380 MatchOp = TI->getOperand(0); 381 OtherOpT = TI->getOperand(1); 382 OtherOpF = FI->getOperand(0); 383 MatchIsOpZero = true; 384 } else if (TI->getOperand(1) == FI->getOperand(0)) { 385 MatchOp = TI->getOperand(1); 386 OtherOpT = TI->getOperand(0); 387 OtherOpF = FI->getOperand(1); 388 MatchIsOpZero = true; 389 } else { 390 return nullptr; 391 } 392 393 // If the select condition is a vector, the operands of the original select's 394 // operands also must be vectors. This may not be the case for getelementptr 395 // for example. 396 if (CondTy->isVectorTy() && (!OtherOpT->getType()->isVectorTy() || 397 !OtherOpF->getType()->isVectorTy())) 398 return nullptr; 399 400 // If we reach here, they do have operations in common. 401 Value *NewSI = Builder.CreateSelect(Cond, OtherOpT, OtherOpF, 402 SI.getName() + ".v", &SI); 403 Value *Op0 = MatchIsOpZero ? MatchOp : NewSI; 404 Value *Op1 = MatchIsOpZero ? NewSI : MatchOp; 405 if (auto *BO = dyn_cast<BinaryOperator>(TI)) { 406 BinaryOperator *NewBO = BinaryOperator::Create(BO->getOpcode(), Op0, Op1); 407 NewBO->copyIRFlags(TI); 408 NewBO->andIRFlags(FI); 409 return NewBO; 410 } 411 if (auto *TGEP = dyn_cast<GetElementPtrInst>(TI)) { 412 auto *FGEP = cast<GetElementPtrInst>(FI); 413 Type *ElementType = TGEP->getResultElementType(); 414 return TGEP->isInBounds() && FGEP->isInBounds() 415 ? GetElementPtrInst::CreateInBounds(ElementType, Op0, {Op1}) 416 : GetElementPtrInst::Create(ElementType, Op0, {Op1}); 417 } 418 llvm_unreachable("Expected BinaryOperator or GEP"); 419 return nullptr; 420 } 421 422 static bool isSelect01(const APInt &C1I, const APInt &C2I) { 423 if (!C1I.isNullValue() && !C2I.isNullValue()) // One side must be zero. 424 return false; 425 return C1I.isOneValue() || C1I.isAllOnesValue() || 426 C2I.isOneValue() || C2I.isAllOnesValue(); 427 } 428 429 /// Try to fold the select into one of the operands to allow further 430 /// optimization. 431 Instruction *InstCombinerImpl::foldSelectIntoOp(SelectInst &SI, Value *TrueVal, 432 Value *FalseVal) { 433 // See the comment above GetSelectFoldableOperands for a description of the 434 // transformation we are doing here. 435 if (auto *TVI = dyn_cast<BinaryOperator>(TrueVal)) { 436 if (TVI->hasOneUse() && !isa<Constant>(FalseVal)) { 437 if (unsigned SFO = getSelectFoldableOperands(TVI)) { 438 unsigned OpToFold = 0; 439 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) { 440 OpToFold = 1; 441 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) { 442 OpToFold = 2; 443 } 444 445 if (OpToFold) { 446 Constant *C = ConstantExpr::getBinOpIdentity(TVI->getOpcode(), 447 TVI->getType(), true); 448 Value *OOp = TVI->getOperand(2-OpToFold); 449 // Avoid creating select between 2 constants unless it's selecting 450 // between 0, 1 and -1. 451 const APInt *OOpC; 452 bool OOpIsAPInt = match(OOp, m_APInt(OOpC)); 453 if (!isa<Constant>(OOp) || 454 (OOpIsAPInt && isSelect01(C->getUniqueInteger(), *OOpC))) { 455 Value *NewSel = Builder.CreateSelect(SI.getCondition(), OOp, C); 456 NewSel->takeName(TVI); 457 BinaryOperator *BO = BinaryOperator::Create(TVI->getOpcode(), 458 FalseVal, NewSel); 459 BO->copyIRFlags(TVI); 460 return BO; 461 } 462 } 463 } 464 } 465 } 466 467 if (auto *FVI = dyn_cast<BinaryOperator>(FalseVal)) { 468 if (FVI->hasOneUse() && !isa<Constant>(TrueVal)) { 469 if (unsigned SFO = getSelectFoldableOperands(FVI)) { 470 unsigned OpToFold = 0; 471 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) { 472 OpToFold = 1; 473 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) { 474 OpToFold = 2; 475 } 476 477 if (OpToFold) { 478 Constant *C = ConstantExpr::getBinOpIdentity(FVI->getOpcode(), 479 FVI->getType(), true); 480 Value *OOp = FVI->getOperand(2-OpToFold); 481 // Avoid creating select between 2 constants unless it's selecting 482 // between 0, 1 and -1. 483 const APInt *OOpC; 484 bool OOpIsAPInt = match(OOp, m_APInt(OOpC)); 485 if (!isa<Constant>(OOp) || 486 (OOpIsAPInt && isSelect01(C->getUniqueInteger(), *OOpC))) { 487 Value *NewSel = Builder.CreateSelect(SI.getCondition(), C, OOp); 488 NewSel->takeName(FVI); 489 BinaryOperator *BO = BinaryOperator::Create(FVI->getOpcode(), 490 TrueVal, NewSel); 491 BO->copyIRFlags(FVI); 492 return BO; 493 } 494 } 495 } 496 } 497 } 498 499 return nullptr; 500 } 501 502 /// We want to turn: 503 /// (select (icmp eq (and X, Y), 0), (and (lshr X, Z), 1), 1) 504 /// into: 505 /// zext (icmp ne i32 (and X, (or Y, (shl 1, Z))), 0) 506 /// Note: 507 /// Z may be 0 if lshr is missing. 508 /// Worst-case scenario is that we will replace 5 instructions with 5 different 509 /// instructions, but we got rid of select. 510 static Instruction *foldSelectICmpAndAnd(Type *SelType, const ICmpInst *Cmp, 511 Value *TVal, Value *FVal, 512 InstCombiner::BuilderTy &Builder) { 513 if (!(Cmp->hasOneUse() && Cmp->getOperand(0)->hasOneUse() && 514 Cmp->getPredicate() == ICmpInst::ICMP_EQ && 515 match(Cmp->getOperand(1), m_Zero()) && match(FVal, m_One()))) 516 return nullptr; 517 518 // The TrueVal has general form of: and %B, 1 519 Value *B; 520 if (!match(TVal, m_OneUse(m_And(m_Value(B), m_One())))) 521 return nullptr; 522 523 // Where %B may be optionally shifted: lshr %X, %Z. 524 Value *X, *Z; 525 const bool HasShift = match(B, m_OneUse(m_LShr(m_Value(X), m_Value(Z)))); 526 if (!HasShift) 527 X = B; 528 529 Value *Y; 530 if (!match(Cmp->getOperand(0), m_c_And(m_Specific(X), m_Value(Y)))) 531 return nullptr; 532 533 // ((X & Y) == 0) ? ((X >> Z) & 1) : 1 --> (X & (Y | (1 << Z))) != 0 534 // ((X & Y) == 0) ? (X & 1) : 1 --> (X & (Y | 1)) != 0 535 Constant *One = ConstantInt::get(SelType, 1); 536 Value *MaskB = HasShift ? Builder.CreateShl(One, Z) : One; 537 Value *FullMask = Builder.CreateOr(Y, MaskB); 538 Value *MaskedX = Builder.CreateAnd(X, FullMask); 539 Value *ICmpNeZero = Builder.CreateIsNotNull(MaskedX); 540 return new ZExtInst(ICmpNeZero, SelType); 541 } 542 543 /// We want to turn: 544 /// (select (icmp sgt x, C), lshr (X, Y), ashr (X, Y)); iff C s>= -1 545 /// (select (icmp slt x, C), ashr (X, Y), lshr (X, Y)); iff C s>= 0 546 /// into: 547 /// ashr (X, Y) 548 static Value *foldSelectICmpLshrAshr(const ICmpInst *IC, Value *TrueVal, 549 Value *FalseVal, 550 InstCombiner::BuilderTy &Builder) { 551 ICmpInst::Predicate Pred = IC->getPredicate(); 552 Value *CmpLHS = IC->getOperand(0); 553 Value *CmpRHS = IC->getOperand(1); 554 if (!CmpRHS->getType()->isIntOrIntVectorTy()) 555 return nullptr; 556 557 Value *X, *Y; 558 unsigned Bitwidth = CmpRHS->getType()->getScalarSizeInBits(); 559 if ((Pred != ICmpInst::ICMP_SGT || 560 !match(CmpRHS, 561 m_SpecificInt_ICMP(ICmpInst::ICMP_SGE, APInt(Bitwidth, -1)))) && 562 (Pred != ICmpInst::ICMP_SLT || 563 !match(CmpRHS, 564 m_SpecificInt_ICMP(ICmpInst::ICMP_SGE, APInt(Bitwidth, 0))))) 565 return nullptr; 566 567 // Canonicalize so that ashr is in FalseVal. 568 if (Pred == ICmpInst::ICMP_SLT) 569 std::swap(TrueVal, FalseVal); 570 571 if (match(TrueVal, m_LShr(m_Value(X), m_Value(Y))) && 572 match(FalseVal, m_AShr(m_Specific(X), m_Specific(Y))) && 573 match(CmpLHS, m_Specific(X))) { 574 const auto *Ashr = cast<Instruction>(FalseVal); 575 // if lshr is not exact and ashr is, this new ashr must not be exact. 576 bool IsExact = Ashr->isExact() && cast<Instruction>(TrueVal)->isExact(); 577 return Builder.CreateAShr(X, Y, IC->getName(), IsExact); 578 } 579 580 return nullptr; 581 } 582 583 /// We want to turn: 584 /// (select (icmp eq (and X, C1), 0), Y, (or Y, C2)) 585 /// into: 586 /// (or (shl (and X, C1), C3), Y) 587 /// iff: 588 /// C1 and C2 are both powers of 2 589 /// where: 590 /// C3 = Log(C2) - Log(C1) 591 /// 592 /// This transform handles cases where: 593 /// 1. The icmp predicate is inverted 594 /// 2. The select operands are reversed 595 /// 3. The magnitude of C2 and C1 are flipped 596 static Value *foldSelectICmpAndOr(const ICmpInst *IC, Value *TrueVal, 597 Value *FalseVal, 598 InstCombiner::BuilderTy &Builder) { 599 // Only handle integer compares. Also, if this is a vector select, we need a 600 // vector compare. 601 if (!TrueVal->getType()->isIntOrIntVectorTy() || 602 TrueVal->getType()->isVectorTy() != IC->getType()->isVectorTy()) 603 return nullptr; 604 605 Value *CmpLHS = IC->getOperand(0); 606 Value *CmpRHS = IC->getOperand(1); 607 608 Value *V; 609 unsigned C1Log; 610 bool IsEqualZero; 611 bool NeedAnd = false; 612 if (IC->isEquality()) { 613 if (!match(CmpRHS, m_Zero())) 614 return nullptr; 615 616 const APInt *C1; 617 if (!match(CmpLHS, m_And(m_Value(), m_Power2(C1)))) 618 return nullptr; 619 620 V = CmpLHS; 621 C1Log = C1->logBase2(); 622 IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_EQ; 623 } else if (IC->getPredicate() == ICmpInst::ICMP_SLT || 624 IC->getPredicate() == ICmpInst::ICMP_SGT) { 625 // We also need to recognize (icmp slt (trunc (X)), 0) and 626 // (icmp sgt (trunc (X)), -1). 627 IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_SGT; 628 if ((IsEqualZero && !match(CmpRHS, m_AllOnes())) || 629 (!IsEqualZero && !match(CmpRHS, m_Zero()))) 630 return nullptr; 631 632 if (!match(CmpLHS, m_OneUse(m_Trunc(m_Value(V))))) 633 return nullptr; 634 635 C1Log = CmpLHS->getType()->getScalarSizeInBits() - 1; 636 NeedAnd = true; 637 } else { 638 return nullptr; 639 } 640 641 const APInt *C2; 642 bool OrOnTrueVal = false; 643 bool OrOnFalseVal = match(FalseVal, m_Or(m_Specific(TrueVal), m_Power2(C2))); 644 if (!OrOnFalseVal) 645 OrOnTrueVal = match(TrueVal, m_Or(m_Specific(FalseVal), m_Power2(C2))); 646 647 if (!OrOnFalseVal && !OrOnTrueVal) 648 return nullptr; 649 650 Value *Y = OrOnFalseVal ? TrueVal : FalseVal; 651 652 unsigned C2Log = C2->logBase2(); 653 654 bool NeedXor = (!IsEqualZero && OrOnFalseVal) || (IsEqualZero && OrOnTrueVal); 655 bool NeedShift = C1Log != C2Log; 656 bool NeedZExtTrunc = Y->getType()->getScalarSizeInBits() != 657 V->getType()->getScalarSizeInBits(); 658 659 // Make sure we don't create more instructions than we save. 660 Value *Or = OrOnFalseVal ? FalseVal : TrueVal; 661 if ((NeedShift + NeedXor + NeedZExtTrunc) > 662 (IC->hasOneUse() + Or->hasOneUse())) 663 return nullptr; 664 665 if (NeedAnd) { 666 // Insert the AND instruction on the input to the truncate. 667 APInt C1 = APInt::getOneBitSet(V->getType()->getScalarSizeInBits(), C1Log); 668 V = Builder.CreateAnd(V, ConstantInt::get(V->getType(), C1)); 669 } 670 671 if (C2Log > C1Log) { 672 V = Builder.CreateZExtOrTrunc(V, Y->getType()); 673 V = Builder.CreateShl(V, C2Log - C1Log); 674 } else if (C1Log > C2Log) { 675 V = Builder.CreateLShr(V, C1Log - C2Log); 676 V = Builder.CreateZExtOrTrunc(V, Y->getType()); 677 } else 678 V = Builder.CreateZExtOrTrunc(V, Y->getType()); 679 680 if (NeedXor) 681 V = Builder.CreateXor(V, *C2); 682 683 return Builder.CreateOr(V, Y); 684 } 685 686 /// Canonicalize a set or clear of a masked set of constant bits to 687 /// select-of-constants form. 688 static Instruction *foldSetClearBits(SelectInst &Sel, 689 InstCombiner::BuilderTy &Builder) { 690 Value *Cond = Sel.getCondition(); 691 Value *T = Sel.getTrueValue(); 692 Value *F = Sel.getFalseValue(); 693 Type *Ty = Sel.getType(); 694 Value *X; 695 const APInt *NotC, *C; 696 697 // Cond ? (X & ~C) : (X | C) --> (X & ~C) | (Cond ? 0 : C) 698 if (match(T, m_And(m_Value(X), m_APInt(NotC))) && 699 match(F, m_OneUse(m_Or(m_Specific(X), m_APInt(C)))) && *NotC == ~(*C)) { 700 Constant *Zero = ConstantInt::getNullValue(Ty); 701 Constant *OrC = ConstantInt::get(Ty, *C); 702 Value *NewSel = Builder.CreateSelect(Cond, Zero, OrC, "masksel", &Sel); 703 return BinaryOperator::CreateOr(T, NewSel); 704 } 705 706 // Cond ? (X | C) : (X & ~C) --> (X & ~C) | (Cond ? C : 0) 707 if (match(F, m_And(m_Value(X), m_APInt(NotC))) && 708 match(T, m_OneUse(m_Or(m_Specific(X), m_APInt(C)))) && *NotC == ~(*C)) { 709 Constant *Zero = ConstantInt::getNullValue(Ty); 710 Constant *OrC = ConstantInt::get(Ty, *C); 711 Value *NewSel = Builder.CreateSelect(Cond, OrC, Zero, "masksel", &Sel); 712 return BinaryOperator::CreateOr(F, NewSel); 713 } 714 715 return nullptr; 716 } 717 718 /// Transform patterns such as (a > b) ? a - b : 0 into usub.sat(a, b). 719 /// There are 8 commuted/swapped variants of this pattern. 720 /// TODO: Also support a - UMIN(a,b) patterns. 721 static Value *canonicalizeSaturatedSubtract(const ICmpInst *ICI, 722 const Value *TrueVal, 723 const Value *FalseVal, 724 InstCombiner::BuilderTy &Builder) { 725 ICmpInst::Predicate Pred = ICI->getPredicate(); 726 if (!ICmpInst::isUnsigned(Pred)) 727 return nullptr; 728 729 // (b > a) ? 0 : a - b -> (b <= a) ? a - b : 0 730 if (match(TrueVal, m_Zero())) { 731 Pred = ICmpInst::getInversePredicate(Pred); 732 std::swap(TrueVal, FalseVal); 733 } 734 if (!match(FalseVal, m_Zero())) 735 return nullptr; 736 737 Value *A = ICI->getOperand(0); 738 Value *B = ICI->getOperand(1); 739 if (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_ULT) { 740 // (b < a) ? a - b : 0 -> (a > b) ? a - b : 0 741 std::swap(A, B); 742 Pred = ICmpInst::getSwappedPredicate(Pred); 743 } 744 745 assert((Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_UGT) && 746 "Unexpected isUnsigned predicate!"); 747 748 // Ensure the sub is of the form: 749 // (a > b) ? a - b : 0 -> usub.sat(a, b) 750 // (a > b) ? b - a : 0 -> -usub.sat(a, b) 751 // Checking for both a-b and a+(-b) as a constant. 752 bool IsNegative = false; 753 const APInt *C; 754 if (match(TrueVal, m_Sub(m_Specific(B), m_Specific(A))) || 755 (match(A, m_APInt(C)) && 756 match(TrueVal, m_Add(m_Specific(B), m_SpecificInt(-*C))))) 757 IsNegative = true; 758 else if (!match(TrueVal, m_Sub(m_Specific(A), m_Specific(B))) && 759 !(match(B, m_APInt(C)) && 760 match(TrueVal, m_Add(m_Specific(A), m_SpecificInt(-*C))))) 761 return nullptr; 762 763 // If we are adding a negate and the sub and icmp are used anywhere else, we 764 // would end up with more instructions. 765 if (IsNegative && !TrueVal->hasOneUse() && !ICI->hasOneUse()) 766 return nullptr; 767 768 // (a > b) ? a - b : 0 -> usub.sat(a, b) 769 // (a > b) ? b - a : 0 -> -usub.sat(a, b) 770 Value *Result = Builder.CreateBinaryIntrinsic(Intrinsic::usub_sat, A, B); 771 if (IsNegative) 772 Result = Builder.CreateNeg(Result); 773 return Result; 774 } 775 776 static Value *canonicalizeSaturatedAdd(ICmpInst *Cmp, Value *TVal, Value *FVal, 777 InstCombiner::BuilderTy &Builder) { 778 if (!Cmp->hasOneUse()) 779 return nullptr; 780 781 // Match unsigned saturated add with constant. 782 Value *Cmp0 = Cmp->getOperand(0); 783 Value *Cmp1 = Cmp->getOperand(1); 784 ICmpInst::Predicate Pred = Cmp->getPredicate(); 785 Value *X; 786 const APInt *C, *CmpC; 787 if (Pred == ICmpInst::ICMP_ULT && 788 match(TVal, m_Add(m_Value(X), m_APInt(C))) && X == Cmp0 && 789 match(FVal, m_AllOnes()) && match(Cmp1, m_APInt(CmpC)) && *CmpC == ~*C) { 790 // (X u< ~C) ? (X + C) : -1 --> uadd.sat(X, C) 791 return Builder.CreateBinaryIntrinsic( 792 Intrinsic::uadd_sat, X, ConstantInt::get(X->getType(), *C)); 793 } 794 795 // Match unsigned saturated add of 2 variables with an unnecessary 'not'. 796 // There are 8 commuted variants. 797 // Canonicalize -1 (saturated result) to true value of the select. 798 if (match(FVal, m_AllOnes())) { 799 std::swap(TVal, FVal); 800 Pred = CmpInst::getInversePredicate(Pred); 801 } 802 if (!match(TVal, m_AllOnes())) 803 return nullptr; 804 805 // Canonicalize predicate to less-than or less-or-equal-than. 806 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) { 807 std::swap(Cmp0, Cmp1); 808 Pred = CmpInst::getSwappedPredicate(Pred); 809 } 810 if (Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_ULE) 811 return nullptr; 812 813 // Match unsigned saturated add of 2 variables with an unnecessary 'not'. 814 // Strictness of the comparison is irrelevant. 815 Value *Y; 816 if (match(Cmp0, m_Not(m_Value(X))) && 817 match(FVal, m_c_Add(m_Specific(X), m_Value(Y))) && Y == Cmp1) { 818 // (~X u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y) 819 // (~X u< Y) ? -1 : (Y + X) --> uadd.sat(X, Y) 820 return Builder.CreateBinaryIntrinsic(Intrinsic::uadd_sat, X, Y); 821 } 822 // The 'not' op may be included in the sum but not the compare. 823 // Strictness of the comparison is irrelevant. 824 X = Cmp0; 825 Y = Cmp1; 826 if (match(FVal, m_c_Add(m_Not(m_Specific(X)), m_Specific(Y)))) { 827 // (X u< Y) ? -1 : (~X + Y) --> uadd.sat(~X, Y) 828 // (X u< Y) ? -1 : (Y + ~X) --> uadd.sat(Y, ~X) 829 BinaryOperator *BO = cast<BinaryOperator>(FVal); 830 return Builder.CreateBinaryIntrinsic( 831 Intrinsic::uadd_sat, BO->getOperand(0), BO->getOperand(1)); 832 } 833 // The overflow may be detected via the add wrapping round. 834 // This is only valid for strict comparison! 835 if (Pred == ICmpInst::ICMP_ULT && 836 match(Cmp0, m_c_Add(m_Specific(Cmp1), m_Value(Y))) && 837 match(FVal, m_c_Add(m_Specific(Cmp1), m_Specific(Y)))) { 838 // ((X + Y) u< X) ? -1 : (X + Y) --> uadd.sat(X, Y) 839 // ((X + Y) u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y) 840 return Builder.CreateBinaryIntrinsic(Intrinsic::uadd_sat, Cmp1, Y); 841 } 842 843 return nullptr; 844 } 845 846 /// Fold the following code sequence: 847 /// \code 848 /// int a = ctlz(x & -x); 849 // x ? 31 - a : a; 850 /// \code 851 /// 852 /// into: 853 /// cttz(x) 854 static Instruction *foldSelectCtlzToCttz(ICmpInst *ICI, Value *TrueVal, 855 Value *FalseVal, 856 InstCombiner::BuilderTy &Builder) { 857 unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits(); 858 if (!ICI->isEquality() || !match(ICI->getOperand(1), m_Zero())) 859 return nullptr; 860 861 if (ICI->getPredicate() == ICmpInst::ICMP_NE) 862 std::swap(TrueVal, FalseVal); 863 864 if (!match(FalseVal, 865 m_Xor(m_Deferred(TrueVal), m_SpecificInt(BitWidth - 1)))) 866 return nullptr; 867 868 if (!match(TrueVal, m_Intrinsic<Intrinsic::ctlz>())) 869 return nullptr; 870 871 Value *X = ICI->getOperand(0); 872 auto *II = cast<IntrinsicInst>(TrueVal); 873 if (!match(II->getOperand(0), m_c_And(m_Specific(X), m_Neg(m_Specific(X))))) 874 return nullptr; 875 876 Function *F = Intrinsic::getDeclaration(II->getModule(), Intrinsic::cttz, 877 II->getType()); 878 return CallInst::Create(F, {X, II->getArgOperand(1)}); 879 } 880 881 /// Attempt to fold a cttz/ctlz followed by a icmp plus select into a single 882 /// call to cttz/ctlz with flag 'is_zero_undef' cleared. 883 /// 884 /// For example, we can fold the following code sequence: 885 /// \code 886 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 true) 887 /// %1 = icmp ne i32 %x, 0 888 /// %2 = select i1 %1, i32 %0, i32 32 889 /// \code 890 /// 891 /// into: 892 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 false) 893 static Value *foldSelectCttzCtlz(ICmpInst *ICI, Value *TrueVal, Value *FalseVal, 894 InstCombiner::BuilderTy &Builder) { 895 ICmpInst::Predicate Pred = ICI->getPredicate(); 896 Value *CmpLHS = ICI->getOperand(0); 897 Value *CmpRHS = ICI->getOperand(1); 898 899 // Check if the condition value compares a value for equality against zero. 900 if (!ICI->isEquality() || !match(CmpRHS, m_Zero())) 901 return nullptr; 902 903 Value *SelectArg = FalseVal; 904 Value *ValueOnZero = TrueVal; 905 if (Pred == ICmpInst::ICMP_NE) 906 std::swap(SelectArg, ValueOnZero); 907 908 // Skip zero extend/truncate. 909 Value *Count = nullptr; 910 if (!match(SelectArg, m_ZExt(m_Value(Count))) && 911 !match(SelectArg, m_Trunc(m_Value(Count)))) 912 Count = SelectArg; 913 914 // Check that 'Count' is a call to intrinsic cttz/ctlz. Also check that the 915 // input to the cttz/ctlz is used as LHS for the compare instruction. 916 if (!match(Count, m_Intrinsic<Intrinsic::cttz>(m_Specific(CmpLHS))) && 917 !match(Count, m_Intrinsic<Intrinsic::ctlz>(m_Specific(CmpLHS)))) 918 return nullptr; 919 920 IntrinsicInst *II = cast<IntrinsicInst>(Count); 921 922 // Check if the value propagated on zero is a constant number equal to the 923 // sizeof in bits of 'Count'. 924 unsigned SizeOfInBits = Count->getType()->getScalarSizeInBits(); 925 if (match(ValueOnZero, m_SpecificInt(SizeOfInBits))) { 926 // Explicitly clear the 'undef_on_zero' flag. It's always valid to go from 927 // true to false on this flag, so we can replace it for all users. 928 II->setArgOperand(1, ConstantInt::getFalse(II->getContext())); 929 return SelectArg; 930 } 931 932 // The ValueOnZero is not the bitwidth. But if the cttz/ctlz (and optional 933 // zext/trunc) have one use (ending at the select), the cttz/ctlz result will 934 // not be used if the input is zero. Relax to 'undef_on_zero' for that case. 935 if (II->hasOneUse() && SelectArg->hasOneUse() && 936 !match(II->getArgOperand(1), m_One())) 937 II->setArgOperand(1, ConstantInt::getTrue(II->getContext())); 938 939 return nullptr; 940 } 941 942 /// Return true if we find and adjust an icmp+select pattern where the compare 943 /// is with a constant that can be incremented or decremented to match the 944 /// minimum or maximum idiom. 945 static bool adjustMinMax(SelectInst &Sel, ICmpInst &Cmp) { 946 ICmpInst::Predicate Pred = Cmp.getPredicate(); 947 Value *CmpLHS = Cmp.getOperand(0); 948 Value *CmpRHS = Cmp.getOperand(1); 949 Value *TrueVal = Sel.getTrueValue(); 950 Value *FalseVal = Sel.getFalseValue(); 951 952 // We may move or edit the compare, so make sure the select is the only user. 953 const APInt *CmpC; 954 if (!Cmp.hasOneUse() || !match(CmpRHS, m_APInt(CmpC))) 955 return false; 956 957 // These transforms only work for selects of integers or vector selects of 958 // integer vectors. 959 Type *SelTy = Sel.getType(); 960 auto *SelEltTy = dyn_cast<IntegerType>(SelTy->getScalarType()); 961 if (!SelEltTy || SelTy->isVectorTy() != Cmp.getType()->isVectorTy()) 962 return false; 963 964 Constant *AdjustedRHS; 965 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT) 966 AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC + 1); 967 else if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT) 968 AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC - 1); 969 else 970 return false; 971 972 // X > C ? X : C+1 --> X < C+1 ? C+1 : X 973 // X < C ? X : C-1 --> X > C-1 ? C-1 : X 974 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) || 975 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) { 976 ; // Nothing to do here. Values match without any sign/zero extension. 977 } 978 // Types do not match. Instead of calculating this with mixed types, promote 979 // all to the larger type. This enables scalar evolution to analyze this 980 // expression. 981 else if (CmpRHS->getType()->getScalarSizeInBits() < SelEltTy->getBitWidth()) { 982 Constant *SextRHS = ConstantExpr::getSExt(AdjustedRHS, SelTy); 983 984 // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X 985 // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X 986 // X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X 987 // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X 988 if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) && SextRHS == FalseVal) { 989 CmpLHS = TrueVal; 990 AdjustedRHS = SextRHS; 991 } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) && 992 SextRHS == TrueVal) { 993 CmpLHS = FalseVal; 994 AdjustedRHS = SextRHS; 995 } else if (Cmp.isUnsigned()) { 996 Constant *ZextRHS = ConstantExpr::getZExt(AdjustedRHS, SelTy); 997 // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X 998 // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X 999 // zext + signed compare cannot be changed: 1000 // 0xff <s 0x00, but 0x00ff >s 0x0000 1001 if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) && ZextRHS == FalseVal) { 1002 CmpLHS = TrueVal; 1003 AdjustedRHS = ZextRHS; 1004 } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) && 1005 ZextRHS == TrueVal) { 1006 CmpLHS = FalseVal; 1007 AdjustedRHS = ZextRHS; 1008 } else { 1009 return false; 1010 } 1011 } else { 1012 return false; 1013 } 1014 } else { 1015 return false; 1016 } 1017 1018 Pred = ICmpInst::getSwappedPredicate(Pred); 1019 CmpRHS = AdjustedRHS; 1020 std::swap(FalseVal, TrueVal); 1021 Cmp.setPredicate(Pred); 1022 Cmp.setOperand(0, CmpLHS); 1023 Cmp.setOperand(1, CmpRHS); 1024 Sel.setOperand(1, TrueVal); 1025 Sel.setOperand(2, FalseVal); 1026 Sel.swapProfMetadata(); 1027 1028 // Move the compare instruction right before the select instruction. Otherwise 1029 // the sext/zext value may be defined after the compare instruction uses it. 1030 Cmp.moveBefore(&Sel); 1031 1032 return true; 1033 } 1034 1035 /// If this is an integer min/max (icmp + select) with a constant operand, 1036 /// create the canonical icmp for the min/max operation and canonicalize the 1037 /// constant to the 'false' operand of the select: 1038 /// select (icmp Pred X, C1), C2, X --> select (icmp Pred' X, C2), X, C2 1039 /// Note: if C1 != C2, this will change the icmp constant to the existing 1040 /// constant operand of the select. 1041 static Instruction *canonicalizeMinMaxWithConstant(SelectInst &Sel, 1042 ICmpInst &Cmp, 1043 InstCombinerImpl &IC) { 1044 if (!Cmp.hasOneUse() || !isa<Constant>(Cmp.getOperand(1))) 1045 return nullptr; 1046 1047 // Canonicalize the compare predicate based on whether we have min or max. 1048 Value *LHS, *RHS; 1049 SelectPatternResult SPR = matchSelectPattern(&Sel, LHS, RHS); 1050 if (!SelectPatternResult::isMinOrMax(SPR.Flavor)) 1051 return nullptr; 1052 1053 // Is this already canonical? 1054 ICmpInst::Predicate CanonicalPred = getMinMaxPred(SPR.Flavor); 1055 if (Cmp.getOperand(0) == LHS && Cmp.getOperand(1) == RHS && 1056 Cmp.getPredicate() == CanonicalPred) 1057 return nullptr; 1058 1059 // Bail out on unsimplified X-0 operand (due to some worklist management bug), 1060 // as this may cause an infinite combine loop. Let the sub be folded first. 1061 if (match(LHS, m_Sub(m_Value(), m_Zero())) || 1062 match(RHS, m_Sub(m_Value(), m_Zero()))) 1063 return nullptr; 1064 1065 // Create the canonical compare and plug it into the select. 1066 IC.replaceOperand(Sel, 0, IC.Builder.CreateICmp(CanonicalPred, LHS, RHS)); 1067 1068 // If the select operands did not change, we're done. 1069 if (Sel.getTrueValue() == LHS && Sel.getFalseValue() == RHS) 1070 return &Sel; 1071 1072 // If we are swapping the select operands, swap the metadata too. 1073 assert(Sel.getTrueValue() == RHS && Sel.getFalseValue() == LHS && 1074 "Unexpected results from matchSelectPattern"); 1075 Sel.swapValues(); 1076 Sel.swapProfMetadata(); 1077 return &Sel; 1078 } 1079 1080 static Instruction *canonicalizeAbsNabs(SelectInst &Sel, ICmpInst &Cmp, 1081 InstCombinerImpl &IC) { 1082 if (!Cmp.hasOneUse() || !isa<Constant>(Cmp.getOperand(1))) 1083 return nullptr; 1084 1085 Value *LHS, *RHS; 1086 SelectPatternFlavor SPF = matchSelectPattern(&Sel, LHS, RHS).Flavor; 1087 if (SPF != SelectPatternFlavor::SPF_ABS && 1088 SPF != SelectPatternFlavor::SPF_NABS) 1089 return nullptr; 1090 1091 // Note that NSW flag can only be propagated for normal, non-negated abs! 1092 bool IntMinIsPoison = SPF == SelectPatternFlavor::SPF_ABS && 1093 match(RHS, m_NSWNeg(m_Specific(LHS))); 1094 Constant *IntMinIsPoisonC = 1095 ConstantInt::get(Type::getInt1Ty(Sel.getContext()), IntMinIsPoison); 1096 Instruction *Abs = 1097 IC.Builder.CreateBinaryIntrinsic(Intrinsic::abs, LHS, IntMinIsPoisonC); 1098 1099 if (SPF == SelectPatternFlavor::SPF_NABS) 1100 return BinaryOperator::CreateNeg(Abs); // Always without NSW flag! 1101 1102 return IC.replaceInstUsesWith(Sel, Abs); 1103 } 1104 1105 /// If we have a select with an equality comparison, then we know the value in 1106 /// one of the arms of the select. See if substituting this value into an arm 1107 /// and simplifying the result yields the same value as the other arm. 1108 /// 1109 /// To make this transform safe, we must drop poison-generating flags 1110 /// (nsw, etc) if we simplified to a binop because the select may be guarding 1111 /// that poison from propagating. If the existing binop already had no 1112 /// poison-generating flags, then this transform can be done by instsimplify. 1113 /// 1114 /// Consider: 1115 /// %cmp = icmp eq i32 %x, 2147483647 1116 /// %add = add nsw i32 %x, 1 1117 /// %sel = select i1 %cmp, i32 -2147483648, i32 %add 1118 /// 1119 /// We can't replace %sel with %add unless we strip away the flags. 1120 /// TODO: Wrapping flags could be preserved in some cases with better analysis. 1121 Instruction *InstCombinerImpl::foldSelectValueEquivalence(SelectInst &Sel, 1122 ICmpInst &Cmp) { 1123 // Value equivalence substitution requires an all-or-nothing replacement. 1124 // It does not make sense for a vector compare where each lane is chosen 1125 // independently. 1126 if (!Cmp.isEquality() || Cmp.getType()->isVectorTy()) 1127 return nullptr; 1128 1129 // Canonicalize the pattern to ICMP_EQ by swapping the select operands. 1130 Value *TrueVal = Sel.getTrueValue(), *FalseVal = Sel.getFalseValue(); 1131 bool Swapped = false; 1132 if (Cmp.getPredicate() == ICmpInst::ICMP_NE) { 1133 std::swap(TrueVal, FalseVal); 1134 Swapped = true; 1135 } 1136 1137 // In X == Y ? f(X) : Z, try to evaluate f(Y) and replace the operand. 1138 // Make sure Y cannot be undef though, as we might pick different values for 1139 // undef in the icmp and in f(Y). Additionally, take care to avoid replacing 1140 // X == Y ? X : Z with X == Y ? Y : Z, as that would lead to an infinite 1141 // replacement cycle. 1142 Value *CmpLHS = Cmp.getOperand(0), *CmpRHS = Cmp.getOperand(1); 1143 if (TrueVal != CmpLHS && 1144 isGuaranteedNotToBeUndefOrPoison(CmpRHS, SQ.AC, &Sel, &DT)) { 1145 if (Value *V = simplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, SQ, 1146 /* AllowRefinement */ true)) 1147 return replaceOperand(Sel, Swapped ? 2 : 1, V); 1148 1149 // Even if TrueVal does not simplify, we can directly replace a use of 1150 // CmpLHS with CmpRHS, as long as the instruction is not used anywhere 1151 // else and is safe to speculatively execute (we may end up executing it 1152 // with different operands, which should not cause side-effects or trigger 1153 // undefined behavior). Only do this if CmpRHS is a constant, as 1154 // profitability is not clear for other cases. 1155 // FIXME: The replacement could be performed recursively. 1156 if (match(CmpRHS, m_ImmConstant()) && !match(CmpLHS, m_ImmConstant())) 1157 if (auto *I = dyn_cast<Instruction>(TrueVal)) 1158 if (I->hasOneUse() && isSafeToSpeculativelyExecute(I)) 1159 for (Use &U : I->operands()) 1160 if (U == CmpLHS) { 1161 replaceUse(U, CmpRHS); 1162 return &Sel; 1163 } 1164 } 1165 if (TrueVal != CmpRHS && 1166 isGuaranteedNotToBeUndefOrPoison(CmpLHS, SQ.AC, &Sel, &DT)) 1167 if (Value *V = simplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, SQ, 1168 /* AllowRefinement */ true)) 1169 return replaceOperand(Sel, Swapped ? 2 : 1, V); 1170 1171 auto *FalseInst = dyn_cast<Instruction>(FalseVal); 1172 if (!FalseInst) 1173 return nullptr; 1174 1175 // InstSimplify already performed this fold if it was possible subject to 1176 // current poison-generating flags. Try the transform again with 1177 // poison-generating flags temporarily dropped. 1178 bool WasNUW = false, WasNSW = false, WasExact = false, WasInBounds = false; 1179 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(FalseVal)) { 1180 WasNUW = OBO->hasNoUnsignedWrap(); 1181 WasNSW = OBO->hasNoSignedWrap(); 1182 FalseInst->setHasNoUnsignedWrap(false); 1183 FalseInst->setHasNoSignedWrap(false); 1184 } 1185 if (auto *PEO = dyn_cast<PossiblyExactOperator>(FalseVal)) { 1186 WasExact = PEO->isExact(); 1187 FalseInst->setIsExact(false); 1188 } 1189 if (auto *GEP = dyn_cast<GetElementPtrInst>(FalseVal)) { 1190 WasInBounds = GEP->isInBounds(); 1191 GEP->setIsInBounds(false); 1192 } 1193 1194 // Try each equivalence substitution possibility. 1195 // We have an 'EQ' comparison, so the select's false value will propagate. 1196 // Example: 1197 // (X == 42) ? 43 : (X + 1) --> (X == 42) ? (X + 1) : (X + 1) --> X + 1 1198 if (simplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, SQ, 1199 /* AllowRefinement */ false) == TrueVal || 1200 simplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, SQ, 1201 /* AllowRefinement */ false) == TrueVal) { 1202 return replaceInstUsesWith(Sel, FalseVal); 1203 } 1204 1205 // Restore poison-generating flags if the transform did not apply. 1206 if (WasNUW) 1207 FalseInst->setHasNoUnsignedWrap(); 1208 if (WasNSW) 1209 FalseInst->setHasNoSignedWrap(); 1210 if (WasExact) 1211 FalseInst->setIsExact(); 1212 if (WasInBounds) 1213 cast<GetElementPtrInst>(FalseInst)->setIsInBounds(); 1214 1215 return nullptr; 1216 } 1217 1218 // See if this is a pattern like: 1219 // %old_cmp1 = icmp slt i32 %x, C2 1220 // %old_replacement = select i1 %old_cmp1, i32 %target_low, i32 %target_high 1221 // %old_x_offseted = add i32 %x, C1 1222 // %old_cmp0 = icmp ult i32 %old_x_offseted, C0 1223 // %r = select i1 %old_cmp0, i32 %x, i32 %old_replacement 1224 // This can be rewritten as more canonical pattern: 1225 // %new_cmp1 = icmp slt i32 %x, -C1 1226 // %new_cmp2 = icmp sge i32 %x, C0-C1 1227 // %new_clamped_low = select i1 %new_cmp1, i32 %target_low, i32 %x 1228 // %r = select i1 %new_cmp2, i32 %target_high, i32 %new_clamped_low 1229 // Iff -C1 s<= C2 s<= C0-C1 1230 // Also ULT predicate can also be UGT iff C0 != -1 (+invert result) 1231 // SLT predicate can also be SGT iff C2 != INT_MAX (+invert res.) 1232 static Instruction *canonicalizeClampLike(SelectInst &Sel0, ICmpInst &Cmp0, 1233 InstCombiner::BuilderTy &Builder) { 1234 Value *X = Sel0.getTrueValue(); 1235 Value *Sel1 = Sel0.getFalseValue(); 1236 1237 // First match the condition of the outermost select. 1238 // Said condition must be one-use. 1239 if (!Cmp0.hasOneUse()) 1240 return nullptr; 1241 Value *Cmp00 = Cmp0.getOperand(0); 1242 Constant *C0; 1243 if (!match(Cmp0.getOperand(1), 1244 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C0)))) 1245 return nullptr; 1246 // Canonicalize Cmp0 into the form we expect. 1247 // FIXME: we shouldn't care about lanes that are 'undef' in the end? 1248 switch (Cmp0.getPredicate()) { 1249 case ICmpInst::Predicate::ICMP_ULT: 1250 break; // Great! 1251 case ICmpInst::Predicate::ICMP_ULE: 1252 // We'd have to increment C0 by one, and for that it must not have all-ones 1253 // element, but then it would have been canonicalized to 'ult' before 1254 // we get here. So we can't do anything useful with 'ule'. 1255 return nullptr; 1256 case ICmpInst::Predicate::ICMP_UGT: 1257 // We want to canonicalize it to 'ult', so we'll need to increment C0, 1258 // which again means it must not have any all-ones elements. 1259 if (!match(C0, 1260 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_NE, 1261 APInt::getAllOnesValue( 1262 C0->getType()->getScalarSizeInBits())))) 1263 return nullptr; // Can't do, have all-ones element[s]. 1264 C0 = InstCombiner::AddOne(C0); 1265 std::swap(X, Sel1); 1266 break; 1267 case ICmpInst::Predicate::ICMP_UGE: 1268 // The only way we'd get this predicate if this `icmp` has extra uses, 1269 // but then we won't be able to do this fold. 1270 return nullptr; 1271 default: 1272 return nullptr; // Unknown predicate. 1273 } 1274 1275 // Now that we've canonicalized the ICmp, we know the X we expect; 1276 // the select in other hand should be one-use. 1277 if (!Sel1->hasOneUse()) 1278 return nullptr; 1279 1280 // We now can finish matching the condition of the outermost select: 1281 // it should either be the X itself, or an addition of some constant to X. 1282 Constant *C1; 1283 if (Cmp00 == X) 1284 C1 = ConstantInt::getNullValue(Sel0.getType()); 1285 else if (!match(Cmp00, 1286 m_Add(m_Specific(X), 1287 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C1))))) 1288 return nullptr; 1289 1290 Value *Cmp1; 1291 ICmpInst::Predicate Pred1; 1292 Constant *C2; 1293 Value *ReplacementLow, *ReplacementHigh; 1294 if (!match(Sel1, m_Select(m_Value(Cmp1), m_Value(ReplacementLow), 1295 m_Value(ReplacementHigh))) || 1296 !match(Cmp1, 1297 m_ICmp(Pred1, m_Specific(X), 1298 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C2))))) 1299 return nullptr; 1300 1301 if (!Cmp1->hasOneUse() && (Cmp00 == X || !Cmp00->hasOneUse())) 1302 return nullptr; // Not enough one-use instructions for the fold. 1303 // FIXME: this restriction could be relaxed if Cmp1 can be reused as one of 1304 // two comparisons we'll need to build. 1305 1306 // Canonicalize Cmp1 into the form we expect. 1307 // FIXME: we shouldn't care about lanes that are 'undef' in the end? 1308 switch (Pred1) { 1309 case ICmpInst::Predicate::ICMP_SLT: 1310 break; 1311 case ICmpInst::Predicate::ICMP_SLE: 1312 // We'd have to increment C2 by one, and for that it must not have signed 1313 // max element, but then it would have been canonicalized to 'slt' before 1314 // we get here. So we can't do anything useful with 'sle'. 1315 return nullptr; 1316 case ICmpInst::Predicate::ICMP_SGT: 1317 // We want to canonicalize it to 'slt', so we'll need to increment C2, 1318 // which again means it must not have any signed max elements. 1319 if (!match(C2, 1320 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_NE, 1321 APInt::getSignedMaxValue( 1322 C2->getType()->getScalarSizeInBits())))) 1323 return nullptr; // Can't do, have signed max element[s]. 1324 C2 = InstCombiner::AddOne(C2); 1325 LLVM_FALLTHROUGH; 1326 case ICmpInst::Predicate::ICMP_SGE: 1327 // Also non-canonical, but here we don't need to change C2, 1328 // so we don't have any restrictions on C2, so we can just handle it. 1329 std::swap(ReplacementLow, ReplacementHigh); 1330 break; 1331 default: 1332 return nullptr; // Unknown predicate. 1333 } 1334 1335 // The thresholds of this clamp-like pattern. 1336 auto *ThresholdLowIncl = ConstantExpr::getNeg(C1); 1337 auto *ThresholdHighExcl = ConstantExpr::getSub(C0, C1); 1338 1339 // The fold has a precondition 1: C2 s>= ThresholdLow 1340 auto *Precond1 = ConstantExpr::getICmp(ICmpInst::Predicate::ICMP_SGE, C2, 1341 ThresholdLowIncl); 1342 if (!match(Precond1, m_One())) 1343 return nullptr; 1344 // The fold has a precondition 2: C2 s<= ThresholdHigh 1345 auto *Precond2 = ConstantExpr::getICmp(ICmpInst::Predicate::ICMP_SLE, C2, 1346 ThresholdHighExcl); 1347 if (!match(Precond2, m_One())) 1348 return nullptr; 1349 1350 // All good, finally emit the new pattern. 1351 Value *ShouldReplaceLow = Builder.CreateICmpSLT(X, ThresholdLowIncl); 1352 Value *ShouldReplaceHigh = Builder.CreateICmpSGE(X, ThresholdHighExcl); 1353 Value *MaybeReplacedLow = 1354 Builder.CreateSelect(ShouldReplaceLow, ReplacementLow, X); 1355 Instruction *MaybeReplacedHigh = 1356 SelectInst::Create(ShouldReplaceHigh, ReplacementHigh, MaybeReplacedLow); 1357 1358 return MaybeReplacedHigh; 1359 } 1360 1361 // If we have 1362 // %cmp = icmp [canonical predicate] i32 %x, C0 1363 // %r = select i1 %cmp, i32 %y, i32 C1 1364 // Where C0 != C1 and %x may be different from %y, see if the constant that we 1365 // will have if we flip the strictness of the predicate (i.e. without changing 1366 // the result) is identical to the C1 in select. If it matches we can change 1367 // original comparison to one with swapped predicate, reuse the constant, 1368 // and swap the hands of select. 1369 static Instruction * 1370 tryToReuseConstantFromSelectInComparison(SelectInst &Sel, ICmpInst &Cmp, 1371 InstCombinerImpl &IC) { 1372 ICmpInst::Predicate Pred; 1373 Value *X; 1374 Constant *C0; 1375 if (!match(&Cmp, m_OneUse(m_ICmp( 1376 Pred, m_Value(X), 1377 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C0)))))) 1378 return nullptr; 1379 1380 // If comparison predicate is non-relational, we won't be able to do anything. 1381 if (ICmpInst::isEquality(Pred)) 1382 return nullptr; 1383 1384 // If comparison predicate is non-canonical, then we certainly won't be able 1385 // to make it canonical; canonicalizeCmpWithConstant() already tried. 1386 if (!InstCombiner::isCanonicalPredicate(Pred)) 1387 return nullptr; 1388 1389 // If the [input] type of comparison and select type are different, lets abort 1390 // for now. We could try to compare constants with trunc/[zs]ext though. 1391 if (C0->getType() != Sel.getType()) 1392 return nullptr; 1393 1394 // FIXME: are there any magic icmp predicate+constant pairs we must not touch? 1395 1396 Value *SelVal0, *SelVal1; // We do not care which one is from where. 1397 match(&Sel, m_Select(m_Value(), m_Value(SelVal0), m_Value(SelVal1))); 1398 // At least one of these values we are selecting between must be a constant 1399 // else we'll never succeed. 1400 if (!match(SelVal0, m_AnyIntegralConstant()) && 1401 !match(SelVal1, m_AnyIntegralConstant())) 1402 return nullptr; 1403 1404 // Does this constant C match any of the `select` values? 1405 auto MatchesSelectValue = [SelVal0, SelVal1](Constant *C) { 1406 return C->isElementWiseEqual(SelVal0) || C->isElementWiseEqual(SelVal1); 1407 }; 1408 1409 // If C0 *already* matches true/false value of select, we are done. 1410 if (MatchesSelectValue(C0)) 1411 return nullptr; 1412 1413 // Check the constant we'd have with flipped-strictness predicate. 1414 auto FlippedStrictness = 1415 InstCombiner::getFlippedStrictnessPredicateAndConstant(Pred, C0); 1416 if (!FlippedStrictness) 1417 return nullptr; 1418 1419 // If said constant doesn't match either, then there is no hope, 1420 if (!MatchesSelectValue(FlippedStrictness->second)) 1421 return nullptr; 1422 1423 // It matched! Lets insert the new comparison just before select. 1424 InstCombiner::BuilderTy::InsertPointGuard Guard(IC.Builder); 1425 IC.Builder.SetInsertPoint(&Sel); 1426 1427 Pred = ICmpInst::getSwappedPredicate(Pred); // Yes, swapped. 1428 Value *NewCmp = IC.Builder.CreateICmp(Pred, X, FlippedStrictness->second, 1429 Cmp.getName() + ".inv"); 1430 IC.replaceOperand(Sel, 0, NewCmp); 1431 Sel.swapValues(); 1432 Sel.swapProfMetadata(); 1433 1434 return &Sel; 1435 } 1436 1437 /// Visit a SelectInst that has an ICmpInst as its first operand. 1438 Instruction *InstCombinerImpl::foldSelectInstWithICmp(SelectInst &SI, 1439 ICmpInst *ICI) { 1440 if (Instruction *NewSel = foldSelectValueEquivalence(SI, *ICI)) 1441 return NewSel; 1442 1443 if (Instruction *NewSel = canonicalizeMinMaxWithConstant(SI, *ICI, *this)) 1444 return NewSel; 1445 1446 if (Instruction *NewAbs = canonicalizeAbsNabs(SI, *ICI, *this)) 1447 return NewAbs; 1448 1449 if (Instruction *NewAbs = canonicalizeClampLike(SI, *ICI, Builder)) 1450 return NewAbs; 1451 1452 if (Instruction *NewSel = 1453 tryToReuseConstantFromSelectInComparison(SI, *ICI, *this)) 1454 return NewSel; 1455 1456 bool Changed = adjustMinMax(SI, *ICI); 1457 1458 if (Value *V = foldSelectICmpAnd(SI, ICI, Builder)) 1459 return replaceInstUsesWith(SI, V); 1460 1461 // NOTE: if we wanted to, this is where to detect integer MIN/MAX 1462 Value *TrueVal = SI.getTrueValue(); 1463 Value *FalseVal = SI.getFalseValue(); 1464 ICmpInst::Predicate Pred = ICI->getPredicate(); 1465 Value *CmpLHS = ICI->getOperand(0); 1466 Value *CmpRHS = ICI->getOperand(1); 1467 if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) { 1468 if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) { 1469 // Transform (X == C) ? X : Y -> (X == C) ? C : Y 1470 SI.setOperand(1, CmpRHS); 1471 Changed = true; 1472 } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) { 1473 // Transform (X != C) ? Y : X -> (X != C) ? Y : C 1474 SI.setOperand(2, CmpRHS); 1475 Changed = true; 1476 } 1477 } 1478 1479 // FIXME: This code is nearly duplicated in InstSimplify. Using/refactoring 1480 // decomposeBitTestICmp() might help. 1481 { 1482 unsigned BitWidth = 1483 DL.getTypeSizeInBits(TrueVal->getType()->getScalarType()); 1484 APInt MinSignedValue = APInt::getSignedMinValue(BitWidth); 1485 Value *X; 1486 const APInt *Y, *C; 1487 bool TrueWhenUnset; 1488 bool IsBitTest = false; 1489 if (ICmpInst::isEquality(Pred) && 1490 match(CmpLHS, m_And(m_Value(X), m_Power2(Y))) && 1491 match(CmpRHS, m_Zero())) { 1492 IsBitTest = true; 1493 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ; 1494 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) { 1495 X = CmpLHS; 1496 Y = &MinSignedValue; 1497 IsBitTest = true; 1498 TrueWhenUnset = false; 1499 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) { 1500 X = CmpLHS; 1501 Y = &MinSignedValue; 1502 IsBitTest = true; 1503 TrueWhenUnset = true; 1504 } 1505 if (IsBitTest) { 1506 Value *V = nullptr; 1507 // (X & Y) == 0 ? X : X ^ Y --> X & ~Y 1508 if (TrueWhenUnset && TrueVal == X && 1509 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 1510 V = Builder.CreateAnd(X, ~(*Y)); 1511 // (X & Y) != 0 ? X ^ Y : X --> X & ~Y 1512 else if (!TrueWhenUnset && FalseVal == X && 1513 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 1514 V = Builder.CreateAnd(X, ~(*Y)); 1515 // (X & Y) == 0 ? X ^ Y : X --> X | Y 1516 else if (TrueWhenUnset && FalseVal == X && 1517 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 1518 V = Builder.CreateOr(X, *Y); 1519 // (X & Y) != 0 ? X : X ^ Y --> X | Y 1520 else if (!TrueWhenUnset && TrueVal == X && 1521 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 1522 V = Builder.CreateOr(X, *Y); 1523 1524 if (V) 1525 return replaceInstUsesWith(SI, V); 1526 } 1527 } 1528 1529 if (Instruction *V = 1530 foldSelectICmpAndAnd(SI.getType(), ICI, TrueVal, FalseVal, Builder)) 1531 return V; 1532 1533 if (Instruction *V = foldSelectCtlzToCttz(ICI, TrueVal, FalseVal, Builder)) 1534 return V; 1535 1536 if (Value *V = foldSelectICmpAndOr(ICI, TrueVal, FalseVal, Builder)) 1537 return replaceInstUsesWith(SI, V); 1538 1539 if (Value *V = foldSelectICmpLshrAshr(ICI, TrueVal, FalseVal, Builder)) 1540 return replaceInstUsesWith(SI, V); 1541 1542 if (Value *V = foldSelectCttzCtlz(ICI, TrueVal, FalseVal, Builder)) 1543 return replaceInstUsesWith(SI, V); 1544 1545 if (Value *V = canonicalizeSaturatedSubtract(ICI, TrueVal, FalseVal, Builder)) 1546 return replaceInstUsesWith(SI, V); 1547 1548 if (Value *V = canonicalizeSaturatedAdd(ICI, TrueVal, FalseVal, Builder)) 1549 return replaceInstUsesWith(SI, V); 1550 1551 return Changed ? &SI : nullptr; 1552 } 1553 1554 /// SI is a select whose condition is a PHI node (but the two may be in 1555 /// different blocks). See if the true/false values (V) are live in all of the 1556 /// predecessor blocks of the PHI. For example, cases like this can't be mapped: 1557 /// 1558 /// X = phi [ C1, BB1], [C2, BB2] 1559 /// Y = add 1560 /// Z = select X, Y, 0 1561 /// 1562 /// because Y is not live in BB1/BB2. 1563 static bool canSelectOperandBeMappingIntoPredBlock(const Value *V, 1564 const SelectInst &SI) { 1565 // If the value is a non-instruction value like a constant or argument, it 1566 // can always be mapped. 1567 const Instruction *I = dyn_cast<Instruction>(V); 1568 if (!I) return true; 1569 1570 // If V is a PHI node defined in the same block as the condition PHI, we can 1571 // map the arguments. 1572 const PHINode *CondPHI = cast<PHINode>(SI.getCondition()); 1573 1574 if (const PHINode *VP = dyn_cast<PHINode>(I)) 1575 if (VP->getParent() == CondPHI->getParent()) 1576 return true; 1577 1578 // Otherwise, if the PHI and select are defined in the same block and if V is 1579 // defined in a different block, then we can transform it. 1580 if (SI.getParent() == CondPHI->getParent() && 1581 I->getParent() != CondPHI->getParent()) 1582 return true; 1583 1584 // Otherwise we have a 'hard' case and we can't tell without doing more 1585 // detailed dominator based analysis, punt. 1586 return false; 1587 } 1588 1589 /// We have an SPF (e.g. a min or max) of an SPF of the form: 1590 /// SPF2(SPF1(A, B), C) 1591 Instruction *InstCombinerImpl::foldSPFofSPF(Instruction *Inner, 1592 SelectPatternFlavor SPF1, Value *A, 1593 Value *B, Instruction &Outer, 1594 SelectPatternFlavor SPF2, 1595 Value *C) { 1596 if (Outer.getType() != Inner->getType()) 1597 return nullptr; 1598 1599 if (C == A || C == B) { 1600 // MAX(MAX(A, B), B) -> MAX(A, B) 1601 // MIN(MIN(a, b), a) -> MIN(a, b) 1602 // TODO: This could be done in instsimplify. 1603 if (SPF1 == SPF2 && SelectPatternResult::isMinOrMax(SPF1)) 1604 return replaceInstUsesWith(Outer, Inner); 1605 1606 // MAX(MIN(a, b), a) -> a 1607 // MIN(MAX(a, b), a) -> a 1608 // TODO: This could be done in instsimplify. 1609 if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) || 1610 (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) || 1611 (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) || 1612 (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN)) 1613 return replaceInstUsesWith(Outer, C); 1614 } 1615 1616 if (SPF1 == SPF2) { 1617 const APInt *CB, *CC; 1618 if (match(B, m_APInt(CB)) && match(C, m_APInt(CC))) { 1619 // MIN(MIN(A, 23), 97) -> MIN(A, 23) 1620 // MAX(MAX(A, 97), 23) -> MAX(A, 97) 1621 // TODO: This could be done in instsimplify. 1622 if ((SPF1 == SPF_UMIN && CB->ule(*CC)) || 1623 (SPF1 == SPF_SMIN && CB->sle(*CC)) || 1624 (SPF1 == SPF_UMAX && CB->uge(*CC)) || 1625 (SPF1 == SPF_SMAX && CB->sge(*CC))) 1626 return replaceInstUsesWith(Outer, Inner); 1627 1628 // MIN(MIN(A, 97), 23) -> MIN(A, 23) 1629 // MAX(MAX(A, 23), 97) -> MAX(A, 97) 1630 if ((SPF1 == SPF_UMIN && CB->ugt(*CC)) || 1631 (SPF1 == SPF_SMIN && CB->sgt(*CC)) || 1632 (SPF1 == SPF_UMAX && CB->ult(*CC)) || 1633 (SPF1 == SPF_SMAX && CB->slt(*CC))) { 1634 Outer.replaceUsesOfWith(Inner, A); 1635 return &Outer; 1636 } 1637 } 1638 } 1639 1640 // max(max(A, B), min(A, B)) --> max(A, B) 1641 // min(min(A, B), max(A, B)) --> min(A, B) 1642 // TODO: This could be done in instsimplify. 1643 if (SPF1 == SPF2 && 1644 ((SPF1 == SPF_UMIN && match(C, m_c_UMax(m_Specific(A), m_Specific(B)))) || 1645 (SPF1 == SPF_SMIN && match(C, m_c_SMax(m_Specific(A), m_Specific(B)))) || 1646 (SPF1 == SPF_UMAX && match(C, m_c_UMin(m_Specific(A), m_Specific(B)))) || 1647 (SPF1 == SPF_SMAX && match(C, m_c_SMin(m_Specific(A), m_Specific(B)))))) 1648 return replaceInstUsesWith(Outer, Inner); 1649 1650 // ABS(ABS(X)) -> ABS(X) 1651 // NABS(NABS(X)) -> NABS(X) 1652 // TODO: This could be done in instsimplify. 1653 if (SPF1 == SPF2 && (SPF1 == SPF_ABS || SPF1 == SPF_NABS)) { 1654 return replaceInstUsesWith(Outer, Inner); 1655 } 1656 1657 // ABS(NABS(X)) -> ABS(X) 1658 // NABS(ABS(X)) -> NABS(X) 1659 if ((SPF1 == SPF_ABS && SPF2 == SPF_NABS) || 1660 (SPF1 == SPF_NABS && SPF2 == SPF_ABS)) { 1661 SelectInst *SI = cast<SelectInst>(Inner); 1662 Value *NewSI = 1663 Builder.CreateSelect(SI->getCondition(), SI->getFalseValue(), 1664 SI->getTrueValue(), SI->getName(), SI); 1665 return replaceInstUsesWith(Outer, NewSI); 1666 } 1667 1668 auto IsFreeOrProfitableToInvert = 1669 [&](Value *V, Value *&NotV, bool &ElidesXor) { 1670 if (match(V, m_Not(m_Value(NotV)))) { 1671 // If V has at most 2 uses then we can get rid of the xor operation 1672 // entirely. 1673 ElidesXor |= !V->hasNUsesOrMore(3); 1674 return true; 1675 } 1676 1677 if (isFreeToInvert(V, !V->hasNUsesOrMore(3))) { 1678 NotV = nullptr; 1679 return true; 1680 } 1681 1682 return false; 1683 }; 1684 1685 Value *NotA, *NotB, *NotC; 1686 bool ElidesXor = false; 1687 1688 // MIN(MIN(~A, ~B), ~C) == ~MAX(MAX(A, B), C) 1689 // MIN(MAX(~A, ~B), ~C) == ~MAX(MIN(A, B), C) 1690 // MAX(MIN(~A, ~B), ~C) == ~MIN(MAX(A, B), C) 1691 // MAX(MAX(~A, ~B), ~C) == ~MIN(MIN(A, B), C) 1692 // 1693 // This transform is performance neutral if we can elide at least one xor from 1694 // the set of three operands, since we'll be tacking on an xor at the very 1695 // end. 1696 if (SelectPatternResult::isMinOrMax(SPF1) && 1697 SelectPatternResult::isMinOrMax(SPF2) && 1698 IsFreeOrProfitableToInvert(A, NotA, ElidesXor) && 1699 IsFreeOrProfitableToInvert(B, NotB, ElidesXor) && 1700 IsFreeOrProfitableToInvert(C, NotC, ElidesXor) && ElidesXor) { 1701 if (!NotA) 1702 NotA = Builder.CreateNot(A); 1703 if (!NotB) 1704 NotB = Builder.CreateNot(B); 1705 if (!NotC) 1706 NotC = Builder.CreateNot(C); 1707 1708 Value *NewInner = createMinMax(Builder, getInverseMinMaxFlavor(SPF1), NotA, 1709 NotB); 1710 Value *NewOuter = Builder.CreateNot( 1711 createMinMax(Builder, getInverseMinMaxFlavor(SPF2), NewInner, NotC)); 1712 return replaceInstUsesWith(Outer, NewOuter); 1713 } 1714 1715 return nullptr; 1716 } 1717 1718 /// Turn select C, (X + Y), (X - Y) --> (X + (select C, Y, (-Y))). 1719 /// This is even legal for FP. 1720 static Instruction *foldAddSubSelect(SelectInst &SI, 1721 InstCombiner::BuilderTy &Builder) { 1722 Value *CondVal = SI.getCondition(); 1723 Value *TrueVal = SI.getTrueValue(); 1724 Value *FalseVal = SI.getFalseValue(); 1725 auto *TI = dyn_cast<Instruction>(TrueVal); 1726 auto *FI = dyn_cast<Instruction>(FalseVal); 1727 if (!TI || !FI || !TI->hasOneUse() || !FI->hasOneUse()) 1728 return nullptr; 1729 1730 Instruction *AddOp = nullptr, *SubOp = nullptr; 1731 if ((TI->getOpcode() == Instruction::Sub && 1732 FI->getOpcode() == Instruction::Add) || 1733 (TI->getOpcode() == Instruction::FSub && 1734 FI->getOpcode() == Instruction::FAdd)) { 1735 AddOp = FI; 1736 SubOp = TI; 1737 } else if ((FI->getOpcode() == Instruction::Sub && 1738 TI->getOpcode() == Instruction::Add) || 1739 (FI->getOpcode() == Instruction::FSub && 1740 TI->getOpcode() == Instruction::FAdd)) { 1741 AddOp = TI; 1742 SubOp = FI; 1743 } 1744 1745 if (AddOp) { 1746 Value *OtherAddOp = nullptr; 1747 if (SubOp->getOperand(0) == AddOp->getOperand(0)) { 1748 OtherAddOp = AddOp->getOperand(1); 1749 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) { 1750 OtherAddOp = AddOp->getOperand(0); 1751 } 1752 1753 if (OtherAddOp) { 1754 // So at this point we know we have (Y -> OtherAddOp): 1755 // select C, (add X, Y), (sub X, Z) 1756 Value *NegVal; // Compute -Z 1757 if (SI.getType()->isFPOrFPVectorTy()) { 1758 NegVal = Builder.CreateFNeg(SubOp->getOperand(1)); 1759 if (Instruction *NegInst = dyn_cast<Instruction>(NegVal)) { 1760 FastMathFlags Flags = AddOp->getFastMathFlags(); 1761 Flags &= SubOp->getFastMathFlags(); 1762 NegInst->setFastMathFlags(Flags); 1763 } 1764 } else { 1765 NegVal = Builder.CreateNeg(SubOp->getOperand(1)); 1766 } 1767 1768 Value *NewTrueOp = OtherAddOp; 1769 Value *NewFalseOp = NegVal; 1770 if (AddOp != TI) 1771 std::swap(NewTrueOp, NewFalseOp); 1772 Value *NewSel = Builder.CreateSelect(CondVal, NewTrueOp, NewFalseOp, 1773 SI.getName() + ".p", &SI); 1774 1775 if (SI.getType()->isFPOrFPVectorTy()) { 1776 Instruction *RI = 1777 BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel); 1778 1779 FastMathFlags Flags = AddOp->getFastMathFlags(); 1780 Flags &= SubOp->getFastMathFlags(); 1781 RI->setFastMathFlags(Flags); 1782 return RI; 1783 } else 1784 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel); 1785 } 1786 } 1787 return nullptr; 1788 } 1789 1790 /// Turn X + Y overflows ? -1 : X + Y -> uadd_sat X, Y 1791 /// And X - Y overflows ? 0 : X - Y -> usub_sat X, Y 1792 /// Along with a number of patterns similar to: 1793 /// X + Y overflows ? (X < 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1794 /// X - Y overflows ? (X > 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1795 static Instruction * 1796 foldOverflowingAddSubSelect(SelectInst &SI, InstCombiner::BuilderTy &Builder) { 1797 Value *CondVal = SI.getCondition(); 1798 Value *TrueVal = SI.getTrueValue(); 1799 Value *FalseVal = SI.getFalseValue(); 1800 1801 WithOverflowInst *II; 1802 if (!match(CondVal, m_ExtractValue<1>(m_WithOverflowInst(II))) || 1803 !match(FalseVal, m_ExtractValue<0>(m_Specific(II)))) 1804 return nullptr; 1805 1806 Value *X = II->getLHS(); 1807 Value *Y = II->getRHS(); 1808 1809 auto IsSignedSaturateLimit = [&](Value *Limit, bool IsAdd) { 1810 Type *Ty = Limit->getType(); 1811 1812 ICmpInst::Predicate Pred; 1813 Value *TrueVal, *FalseVal, *Op; 1814 const APInt *C; 1815 if (!match(Limit, m_Select(m_ICmp(Pred, m_Value(Op), m_APInt(C)), 1816 m_Value(TrueVal), m_Value(FalseVal)))) 1817 return false; 1818 1819 auto IsZeroOrOne = [](const APInt &C) { 1820 return C.isNullValue() || C.isOneValue(); 1821 }; 1822 auto IsMinMax = [&](Value *Min, Value *Max) { 1823 APInt MinVal = APInt::getSignedMinValue(Ty->getScalarSizeInBits()); 1824 APInt MaxVal = APInt::getSignedMaxValue(Ty->getScalarSizeInBits()); 1825 return match(Min, m_SpecificInt(MinVal)) && 1826 match(Max, m_SpecificInt(MaxVal)); 1827 }; 1828 1829 if (Op != X && Op != Y) 1830 return false; 1831 1832 if (IsAdd) { 1833 // X + Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1834 // X + Y overflows ? (X <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1835 // X + Y overflows ? (Y <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1836 // X + Y overflows ? (Y <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1837 if (Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C) && 1838 IsMinMax(TrueVal, FalseVal)) 1839 return true; 1840 // X + Y overflows ? (X >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1841 // X + Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1842 // X + Y overflows ? (Y >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1843 // X + Y overflows ? (Y >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1844 if (Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 1) && 1845 IsMinMax(FalseVal, TrueVal)) 1846 return true; 1847 } else { 1848 // X - Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1849 // X - Y overflows ? (X <s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1850 if (Op == X && Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C + 1) && 1851 IsMinMax(TrueVal, FalseVal)) 1852 return true; 1853 // X - Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1854 // X - Y overflows ? (X >s -2 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1855 if (Op == X && Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 2) && 1856 IsMinMax(FalseVal, TrueVal)) 1857 return true; 1858 // X - Y overflows ? (Y <s 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1859 // X - Y overflows ? (Y <s 1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1860 if (Op == Y && Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C) && 1861 IsMinMax(FalseVal, TrueVal)) 1862 return true; 1863 // X - Y overflows ? (Y >s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1864 // X - Y overflows ? (Y >s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1865 if (Op == Y && Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 1) && 1866 IsMinMax(TrueVal, FalseVal)) 1867 return true; 1868 } 1869 1870 return false; 1871 }; 1872 1873 Intrinsic::ID NewIntrinsicID; 1874 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow && 1875 match(TrueVal, m_AllOnes())) 1876 // X + Y overflows ? -1 : X + Y -> uadd_sat X, Y 1877 NewIntrinsicID = Intrinsic::uadd_sat; 1878 else if (II->getIntrinsicID() == Intrinsic::usub_with_overflow && 1879 match(TrueVal, m_Zero())) 1880 // X - Y overflows ? 0 : X - Y -> usub_sat X, Y 1881 NewIntrinsicID = Intrinsic::usub_sat; 1882 else if (II->getIntrinsicID() == Intrinsic::sadd_with_overflow && 1883 IsSignedSaturateLimit(TrueVal, /*IsAdd=*/true)) 1884 // X + Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1885 // X + Y overflows ? (X <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1886 // X + Y overflows ? (X >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1887 // X + Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1888 // X + Y overflows ? (Y <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1889 // X + Y overflows ? (Y <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1890 // X + Y overflows ? (Y >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1891 // X + Y overflows ? (Y >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1892 NewIntrinsicID = Intrinsic::sadd_sat; 1893 else if (II->getIntrinsicID() == Intrinsic::ssub_with_overflow && 1894 IsSignedSaturateLimit(TrueVal, /*IsAdd=*/false)) 1895 // X - Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1896 // X - Y overflows ? (X <s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1897 // X - Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1898 // X - Y overflows ? (X >s -2 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1899 // X - Y overflows ? (Y <s 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1900 // X - Y overflows ? (Y <s 1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1901 // X - Y overflows ? (Y >s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1902 // X - Y overflows ? (Y >s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1903 NewIntrinsicID = Intrinsic::ssub_sat; 1904 else 1905 return nullptr; 1906 1907 Function *F = 1908 Intrinsic::getDeclaration(SI.getModule(), NewIntrinsicID, SI.getType()); 1909 return CallInst::Create(F, {X, Y}); 1910 } 1911 1912 Instruction *InstCombinerImpl::foldSelectExtConst(SelectInst &Sel) { 1913 Constant *C; 1914 if (!match(Sel.getTrueValue(), m_Constant(C)) && 1915 !match(Sel.getFalseValue(), m_Constant(C))) 1916 return nullptr; 1917 1918 Instruction *ExtInst; 1919 if (!match(Sel.getTrueValue(), m_Instruction(ExtInst)) && 1920 !match(Sel.getFalseValue(), m_Instruction(ExtInst))) 1921 return nullptr; 1922 1923 auto ExtOpcode = ExtInst->getOpcode(); 1924 if (ExtOpcode != Instruction::ZExt && ExtOpcode != Instruction::SExt) 1925 return nullptr; 1926 1927 // If we are extending from a boolean type or if we can create a select that 1928 // has the same size operands as its condition, try to narrow the select. 1929 Value *X = ExtInst->getOperand(0); 1930 Type *SmallType = X->getType(); 1931 Value *Cond = Sel.getCondition(); 1932 auto *Cmp = dyn_cast<CmpInst>(Cond); 1933 if (!SmallType->isIntOrIntVectorTy(1) && 1934 (!Cmp || Cmp->getOperand(0)->getType() != SmallType)) 1935 return nullptr; 1936 1937 // If the constant is the same after truncation to the smaller type and 1938 // extension to the original type, we can narrow the select. 1939 Type *SelType = Sel.getType(); 1940 Constant *TruncC = ConstantExpr::getTrunc(C, SmallType); 1941 Constant *ExtC = ConstantExpr::getCast(ExtOpcode, TruncC, SelType); 1942 if (ExtC == C && ExtInst->hasOneUse()) { 1943 Value *TruncCVal = cast<Value>(TruncC); 1944 if (ExtInst == Sel.getFalseValue()) 1945 std::swap(X, TruncCVal); 1946 1947 // select Cond, (ext X), C --> ext(select Cond, X, C') 1948 // select Cond, C, (ext X) --> ext(select Cond, C', X) 1949 Value *NewSel = Builder.CreateSelect(Cond, X, TruncCVal, "narrow", &Sel); 1950 return CastInst::Create(Instruction::CastOps(ExtOpcode), NewSel, SelType); 1951 } 1952 1953 // If one arm of the select is the extend of the condition, replace that arm 1954 // with the extension of the appropriate known bool value. 1955 if (Cond == X) { 1956 if (ExtInst == Sel.getTrueValue()) { 1957 // select X, (sext X), C --> select X, -1, C 1958 // select X, (zext X), C --> select X, 1, C 1959 Constant *One = ConstantInt::getTrue(SmallType); 1960 Constant *AllOnesOrOne = ConstantExpr::getCast(ExtOpcode, One, SelType); 1961 return SelectInst::Create(Cond, AllOnesOrOne, C, "", nullptr, &Sel); 1962 } else { 1963 // select X, C, (sext X) --> select X, C, 0 1964 // select X, C, (zext X) --> select X, C, 0 1965 Constant *Zero = ConstantInt::getNullValue(SelType); 1966 return SelectInst::Create(Cond, C, Zero, "", nullptr, &Sel); 1967 } 1968 } 1969 1970 return nullptr; 1971 } 1972 1973 /// Try to transform a vector select with a constant condition vector into a 1974 /// shuffle for easier combining with other shuffles and insert/extract. 1975 static Instruction *canonicalizeSelectToShuffle(SelectInst &SI) { 1976 Value *CondVal = SI.getCondition(); 1977 Constant *CondC; 1978 auto *CondValTy = dyn_cast<FixedVectorType>(CondVal->getType()); 1979 if (!CondValTy || !match(CondVal, m_Constant(CondC))) 1980 return nullptr; 1981 1982 unsigned NumElts = CondValTy->getNumElements(); 1983 SmallVector<int, 16> Mask; 1984 Mask.reserve(NumElts); 1985 for (unsigned i = 0; i != NumElts; ++i) { 1986 Constant *Elt = CondC->getAggregateElement(i); 1987 if (!Elt) 1988 return nullptr; 1989 1990 if (Elt->isOneValue()) { 1991 // If the select condition element is true, choose from the 1st vector. 1992 Mask.push_back(i); 1993 } else if (Elt->isNullValue()) { 1994 // If the select condition element is false, choose from the 2nd vector. 1995 Mask.push_back(i + NumElts); 1996 } else if (isa<UndefValue>(Elt)) { 1997 // Undef in a select condition (choose one of the operands) does not mean 1998 // the same thing as undef in a shuffle mask (any value is acceptable), so 1999 // give up. 2000 return nullptr; 2001 } else { 2002 // Bail out on a constant expression. 2003 return nullptr; 2004 } 2005 } 2006 2007 return new ShuffleVectorInst(SI.getTrueValue(), SI.getFalseValue(), Mask); 2008 } 2009 2010 /// If we have a select of vectors with a scalar condition, try to convert that 2011 /// to a vector select by splatting the condition. A splat may get folded with 2012 /// other operations in IR and having all operands of a select be vector types 2013 /// is likely better for vector codegen. 2014 static Instruction *canonicalizeScalarSelectOfVecs(SelectInst &Sel, 2015 InstCombinerImpl &IC) { 2016 auto *Ty = dyn_cast<VectorType>(Sel.getType()); 2017 if (!Ty) 2018 return nullptr; 2019 2020 // We can replace a single-use extract with constant index. 2021 Value *Cond = Sel.getCondition(); 2022 if (!match(Cond, m_OneUse(m_ExtractElt(m_Value(), m_ConstantInt())))) 2023 return nullptr; 2024 2025 // select (extelt V, Index), T, F --> select (splat V, Index), T, F 2026 // Splatting the extracted condition reduces code (we could directly create a 2027 // splat shuffle of the source vector to eliminate the intermediate step). 2028 return IC.replaceOperand( 2029 Sel, 0, IC.Builder.CreateVectorSplat(Ty->getElementCount(), Cond)); 2030 } 2031 2032 /// Reuse bitcasted operands between a compare and select: 2033 /// select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) --> 2034 /// bitcast (select (cmp (bitcast C), (bitcast D)), (bitcast C), (bitcast D)) 2035 static Instruction *foldSelectCmpBitcasts(SelectInst &Sel, 2036 InstCombiner::BuilderTy &Builder) { 2037 Value *Cond = Sel.getCondition(); 2038 Value *TVal = Sel.getTrueValue(); 2039 Value *FVal = Sel.getFalseValue(); 2040 2041 CmpInst::Predicate Pred; 2042 Value *A, *B; 2043 if (!match(Cond, m_Cmp(Pred, m_Value(A), m_Value(B)))) 2044 return nullptr; 2045 2046 // The select condition is a compare instruction. If the select's true/false 2047 // values are already the same as the compare operands, there's nothing to do. 2048 if (TVal == A || TVal == B || FVal == A || FVal == B) 2049 return nullptr; 2050 2051 Value *C, *D; 2052 if (!match(A, m_BitCast(m_Value(C))) || !match(B, m_BitCast(m_Value(D)))) 2053 return nullptr; 2054 2055 // select (cmp (bitcast C), (bitcast D)), (bitcast TSrc), (bitcast FSrc) 2056 Value *TSrc, *FSrc; 2057 if (!match(TVal, m_BitCast(m_Value(TSrc))) || 2058 !match(FVal, m_BitCast(m_Value(FSrc)))) 2059 return nullptr; 2060 2061 // If the select true/false values are *different bitcasts* of the same source 2062 // operands, make the select operands the same as the compare operands and 2063 // cast the result. This is the canonical select form for min/max. 2064 Value *NewSel; 2065 if (TSrc == C && FSrc == D) { 2066 // select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) --> 2067 // bitcast (select (cmp A, B), A, B) 2068 NewSel = Builder.CreateSelect(Cond, A, B, "", &Sel); 2069 } else if (TSrc == D && FSrc == C) { 2070 // select (cmp (bitcast C), (bitcast D)), (bitcast' D), (bitcast' C) --> 2071 // bitcast (select (cmp A, B), B, A) 2072 NewSel = Builder.CreateSelect(Cond, B, A, "", &Sel); 2073 } else { 2074 return nullptr; 2075 } 2076 return CastInst::CreateBitOrPointerCast(NewSel, Sel.getType()); 2077 } 2078 2079 /// Try to eliminate select instructions that test the returned flag of cmpxchg 2080 /// instructions. 2081 /// 2082 /// If a select instruction tests the returned flag of a cmpxchg instruction and 2083 /// selects between the returned value of the cmpxchg instruction its compare 2084 /// operand, the result of the select will always be equal to its false value. 2085 /// For example: 2086 /// 2087 /// %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst 2088 /// %1 = extractvalue { i64, i1 } %0, 1 2089 /// %2 = extractvalue { i64, i1 } %0, 0 2090 /// %3 = select i1 %1, i64 %compare, i64 %2 2091 /// ret i64 %3 2092 /// 2093 /// The returned value of the cmpxchg instruction (%2) is the original value 2094 /// located at %ptr prior to any update. If the cmpxchg operation succeeds, %2 2095 /// must have been equal to %compare. Thus, the result of the select is always 2096 /// equal to %2, and the code can be simplified to: 2097 /// 2098 /// %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst 2099 /// %1 = extractvalue { i64, i1 } %0, 0 2100 /// ret i64 %1 2101 /// 2102 static Value *foldSelectCmpXchg(SelectInst &SI) { 2103 // A helper that determines if V is an extractvalue instruction whose 2104 // aggregate operand is a cmpxchg instruction and whose single index is equal 2105 // to I. If such conditions are true, the helper returns the cmpxchg 2106 // instruction; otherwise, a nullptr is returned. 2107 auto isExtractFromCmpXchg = [](Value *V, unsigned I) -> AtomicCmpXchgInst * { 2108 auto *Extract = dyn_cast<ExtractValueInst>(V); 2109 if (!Extract) 2110 return nullptr; 2111 if (Extract->getIndices()[0] != I) 2112 return nullptr; 2113 return dyn_cast<AtomicCmpXchgInst>(Extract->getAggregateOperand()); 2114 }; 2115 2116 // If the select has a single user, and this user is a select instruction that 2117 // we can simplify, skip the cmpxchg simplification for now. 2118 if (SI.hasOneUse()) 2119 if (auto *Select = dyn_cast<SelectInst>(SI.user_back())) 2120 if (Select->getCondition() == SI.getCondition()) 2121 if (Select->getFalseValue() == SI.getTrueValue() || 2122 Select->getTrueValue() == SI.getFalseValue()) 2123 return nullptr; 2124 2125 // Ensure the select condition is the returned flag of a cmpxchg instruction. 2126 auto *CmpXchg = isExtractFromCmpXchg(SI.getCondition(), 1); 2127 if (!CmpXchg) 2128 return nullptr; 2129 2130 // Check the true value case: The true value of the select is the returned 2131 // value of the same cmpxchg used by the condition, and the false value is the 2132 // cmpxchg instruction's compare operand. 2133 if (auto *X = isExtractFromCmpXchg(SI.getTrueValue(), 0)) 2134 if (X == CmpXchg && X->getCompareOperand() == SI.getFalseValue()) 2135 return SI.getFalseValue(); 2136 2137 // Check the false value case: The false value of the select is the returned 2138 // value of the same cmpxchg used by the condition, and the true value is the 2139 // cmpxchg instruction's compare operand. 2140 if (auto *X = isExtractFromCmpXchg(SI.getFalseValue(), 0)) 2141 if (X == CmpXchg && X->getCompareOperand() == SI.getTrueValue()) 2142 return SI.getFalseValue(); 2143 2144 return nullptr; 2145 } 2146 2147 static Instruction *moveAddAfterMinMax(SelectPatternFlavor SPF, Value *X, 2148 Value *Y, 2149 InstCombiner::BuilderTy &Builder) { 2150 assert(SelectPatternResult::isMinOrMax(SPF) && "Expected min/max pattern"); 2151 bool IsUnsigned = SPF == SelectPatternFlavor::SPF_UMIN || 2152 SPF == SelectPatternFlavor::SPF_UMAX; 2153 // TODO: If InstSimplify could fold all cases where C2 <= C1, we could change 2154 // the constant value check to an assert. 2155 Value *A; 2156 const APInt *C1, *C2; 2157 if (IsUnsigned && match(X, m_NUWAdd(m_Value(A), m_APInt(C1))) && 2158 match(Y, m_APInt(C2)) && C2->uge(*C1) && X->hasNUses(2)) { 2159 // umin (add nuw A, C1), C2 --> add nuw (umin A, C2 - C1), C1 2160 // umax (add nuw A, C1), C2 --> add nuw (umax A, C2 - C1), C1 2161 Value *NewMinMax = createMinMax(Builder, SPF, A, 2162 ConstantInt::get(X->getType(), *C2 - *C1)); 2163 return BinaryOperator::CreateNUW(BinaryOperator::Add, NewMinMax, 2164 ConstantInt::get(X->getType(), *C1)); 2165 } 2166 2167 if (!IsUnsigned && match(X, m_NSWAdd(m_Value(A), m_APInt(C1))) && 2168 match(Y, m_APInt(C2)) && X->hasNUses(2)) { 2169 bool Overflow; 2170 APInt Diff = C2->ssub_ov(*C1, Overflow); 2171 if (!Overflow) { 2172 // smin (add nsw A, C1), C2 --> add nsw (smin A, C2 - C1), C1 2173 // smax (add nsw A, C1), C2 --> add nsw (smax A, C2 - C1), C1 2174 Value *NewMinMax = createMinMax(Builder, SPF, A, 2175 ConstantInt::get(X->getType(), Diff)); 2176 return BinaryOperator::CreateNSW(BinaryOperator::Add, NewMinMax, 2177 ConstantInt::get(X->getType(), *C1)); 2178 } 2179 } 2180 2181 return nullptr; 2182 } 2183 2184 /// Match a sadd_sat or ssub_sat which is using min/max to clamp the value. 2185 Instruction *InstCombinerImpl::matchSAddSubSat(SelectInst &MinMax1) { 2186 Type *Ty = MinMax1.getType(); 2187 2188 // We are looking for a tree of: 2189 // max(INT_MIN, min(INT_MAX, add(sext(A), sext(B)))) 2190 // Where the min and max could be reversed 2191 Instruction *MinMax2; 2192 BinaryOperator *AddSub; 2193 const APInt *MinValue, *MaxValue; 2194 if (match(&MinMax1, m_SMin(m_Instruction(MinMax2), m_APInt(MaxValue)))) { 2195 if (!match(MinMax2, m_SMax(m_BinOp(AddSub), m_APInt(MinValue)))) 2196 return nullptr; 2197 } else if (match(&MinMax1, 2198 m_SMax(m_Instruction(MinMax2), m_APInt(MinValue)))) { 2199 if (!match(MinMax2, m_SMin(m_BinOp(AddSub), m_APInt(MaxValue)))) 2200 return nullptr; 2201 } else 2202 return nullptr; 2203 2204 // Check that the constants clamp a saturate, and that the new type would be 2205 // sensible to convert to. 2206 if (!(*MaxValue + 1).isPowerOf2() || -*MinValue != *MaxValue + 1) 2207 return nullptr; 2208 // In what bitwidth can this be treated as saturating arithmetics? 2209 unsigned NewBitWidth = (*MaxValue + 1).logBase2() + 1; 2210 // FIXME: This isn't quite right for vectors, but using the scalar type is a 2211 // good first approximation for what should be done there. 2212 if (!shouldChangeType(Ty->getScalarType()->getIntegerBitWidth(), NewBitWidth)) 2213 return nullptr; 2214 2215 // Also make sure that the number of uses is as expected. The "3"s are for the 2216 // the two items of min/max (the compare and the select). 2217 if (MinMax2->hasNUsesOrMore(3) || AddSub->hasNUsesOrMore(3)) 2218 return nullptr; 2219 2220 // Create the new type (which can be a vector type) 2221 Type *NewTy = Ty->getWithNewBitWidth(NewBitWidth); 2222 // Match the two extends from the add/sub 2223 Value *A, *B; 2224 if(!match(AddSub, m_BinOp(m_SExt(m_Value(A)), m_SExt(m_Value(B))))) 2225 return nullptr; 2226 // And check the incoming values are of a type smaller than or equal to the 2227 // size of the saturation. Otherwise the higher bits can cause different 2228 // results. 2229 if (A->getType()->getScalarSizeInBits() > NewBitWidth || 2230 B->getType()->getScalarSizeInBits() > NewBitWidth) 2231 return nullptr; 2232 2233 Intrinsic::ID IntrinsicID; 2234 if (AddSub->getOpcode() == Instruction::Add) 2235 IntrinsicID = Intrinsic::sadd_sat; 2236 else if (AddSub->getOpcode() == Instruction::Sub) 2237 IntrinsicID = Intrinsic::ssub_sat; 2238 else 2239 return nullptr; 2240 2241 // Finally create and return the sat intrinsic, truncated to the new type 2242 Function *F = Intrinsic::getDeclaration(MinMax1.getModule(), IntrinsicID, NewTy); 2243 Value *AT = Builder.CreateSExt(A, NewTy); 2244 Value *BT = Builder.CreateSExt(B, NewTy); 2245 Value *Sat = Builder.CreateCall(F, {AT, BT}); 2246 return CastInst::Create(Instruction::SExt, Sat, Ty); 2247 } 2248 2249 /// Reduce a sequence of min/max with a common operand. 2250 static Instruction *factorizeMinMaxTree(SelectPatternFlavor SPF, Value *LHS, 2251 Value *RHS, 2252 InstCombiner::BuilderTy &Builder) { 2253 assert(SelectPatternResult::isMinOrMax(SPF) && "Expected a min/max"); 2254 // TODO: Allow FP min/max with nnan/nsz. 2255 if (!LHS->getType()->isIntOrIntVectorTy()) 2256 return nullptr; 2257 2258 // Match 3 of the same min/max ops. Example: umin(umin(), umin()). 2259 Value *A, *B, *C, *D; 2260 SelectPatternResult L = matchSelectPattern(LHS, A, B); 2261 SelectPatternResult R = matchSelectPattern(RHS, C, D); 2262 if (SPF != L.Flavor || L.Flavor != R.Flavor) 2263 return nullptr; 2264 2265 // Look for a common operand. The use checks are different than usual because 2266 // a min/max pattern typically has 2 uses of each op: 1 by the cmp and 1 by 2267 // the select. 2268 Value *MinMaxOp = nullptr; 2269 Value *ThirdOp = nullptr; 2270 if (!LHS->hasNUsesOrMore(3) && RHS->hasNUsesOrMore(3)) { 2271 // If the LHS is only used in this chain and the RHS is used outside of it, 2272 // reuse the RHS min/max because that will eliminate the LHS. 2273 if (D == A || C == A) { 2274 // min(min(a, b), min(c, a)) --> min(min(c, a), b) 2275 // min(min(a, b), min(a, d)) --> min(min(a, d), b) 2276 MinMaxOp = RHS; 2277 ThirdOp = B; 2278 } else if (D == B || C == B) { 2279 // min(min(a, b), min(c, b)) --> min(min(c, b), a) 2280 // min(min(a, b), min(b, d)) --> min(min(b, d), a) 2281 MinMaxOp = RHS; 2282 ThirdOp = A; 2283 } 2284 } else if (!RHS->hasNUsesOrMore(3)) { 2285 // Reuse the LHS. This will eliminate the RHS. 2286 if (D == A || D == B) { 2287 // min(min(a, b), min(c, a)) --> min(min(a, b), c) 2288 // min(min(a, b), min(c, b)) --> min(min(a, b), c) 2289 MinMaxOp = LHS; 2290 ThirdOp = C; 2291 } else if (C == A || C == B) { 2292 // min(min(a, b), min(b, d)) --> min(min(a, b), d) 2293 // min(min(a, b), min(c, b)) --> min(min(a, b), d) 2294 MinMaxOp = LHS; 2295 ThirdOp = D; 2296 } 2297 } 2298 if (!MinMaxOp || !ThirdOp) 2299 return nullptr; 2300 2301 CmpInst::Predicate P = getMinMaxPred(SPF); 2302 Value *CmpABC = Builder.CreateICmp(P, MinMaxOp, ThirdOp); 2303 return SelectInst::Create(CmpABC, MinMaxOp, ThirdOp); 2304 } 2305 2306 /// Try to reduce a funnel/rotate pattern that includes a compare and select 2307 /// into a funnel shift intrinsic. Example: 2308 /// rotl32(a, b) --> (b == 0 ? a : ((a >> (32 - b)) | (a << b))) 2309 /// --> call llvm.fshl.i32(a, a, b) 2310 /// fshl32(a, b, c) --> (c == 0 ? a : ((b >> (32 - c)) | (a << c))) 2311 /// --> call llvm.fshl.i32(a, b, c) 2312 /// fshr32(a, b, c) --> (c == 0 ? b : ((a >> (32 - c)) | (b << c))) 2313 /// --> call llvm.fshr.i32(a, b, c) 2314 static Instruction *foldSelectFunnelShift(SelectInst &Sel, 2315 InstCombiner::BuilderTy &Builder) { 2316 // This must be a power-of-2 type for a bitmasking transform to be valid. 2317 unsigned Width = Sel.getType()->getScalarSizeInBits(); 2318 if (!isPowerOf2_32(Width)) 2319 return nullptr; 2320 2321 BinaryOperator *Or0, *Or1; 2322 if (!match(Sel.getFalseValue(), m_OneUse(m_Or(m_BinOp(Or0), m_BinOp(Or1))))) 2323 return nullptr; 2324 2325 Value *SV0, *SV1, *SA0, *SA1; 2326 if (!match(Or0, m_OneUse(m_LogicalShift(m_Value(SV0), 2327 m_ZExtOrSelf(m_Value(SA0))))) || 2328 !match(Or1, m_OneUse(m_LogicalShift(m_Value(SV1), 2329 m_ZExtOrSelf(m_Value(SA1))))) || 2330 Or0->getOpcode() == Or1->getOpcode()) 2331 return nullptr; 2332 2333 // Canonicalize to or(shl(SV0, SA0), lshr(SV1, SA1)). 2334 if (Or0->getOpcode() == BinaryOperator::LShr) { 2335 std::swap(Or0, Or1); 2336 std::swap(SV0, SV1); 2337 std::swap(SA0, SA1); 2338 } 2339 assert(Or0->getOpcode() == BinaryOperator::Shl && 2340 Or1->getOpcode() == BinaryOperator::LShr && 2341 "Illegal or(shift,shift) pair"); 2342 2343 // Check the shift amounts to see if they are an opposite pair. 2344 Value *ShAmt; 2345 if (match(SA1, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(SA0))))) 2346 ShAmt = SA0; 2347 else if (match(SA0, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(SA1))))) 2348 ShAmt = SA1; 2349 else 2350 return nullptr; 2351 2352 // We should now have this pattern: 2353 // select ?, TVal, (or (shl SV0, SA0), (lshr SV1, SA1)) 2354 // The false value of the select must be a funnel-shift of the true value: 2355 // IsFShl -> TVal must be SV0 else TVal must be SV1. 2356 bool IsFshl = (ShAmt == SA0); 2357 Value *TVal = Sel.getTrueValue(); 2358 if ((IsFshl && TVal != SV0) || (!IsFshl && TVal != SV1)) 2359 return nullptr; 2360 2361 // Finally, see if the select is filtering out a shift-by-zero. 2362 Value *Cond = Sel.getCondition(); 2363 ICmpInst::Predicate Pred; 2364 if (!match(Cond, m_OneUse(m_ICmp(Pred, m_Specific(ShAmt), m_ZeroInt()))) || 2365 Pred != ICmpInst::ICMP_EQ) 2366 return nullptr; 2367 2368 // If this is not a rotate then the select was blocking poison from the 2369 // 'shift-by-zero' non-TVal, but a funnel shift won't - so freeze it. 2370 if (SV0 != SV1) { 2371 if (IsFshl && !llvm::isGuaranteedNotToBePoison(SV1)) 2372 SV1 = Builder.CreateFreeze(SV1); 2373 else if (!IsFshl && !llvm::isGuaranteedNotToBePoison(SV0)) 2374 SV0 = Builder.CreateFreeze(SV0); 2375 } 2376 2377 // This is a funnel/rotate that avoids shift-by-bitwidth UB in a suboptimal way. 2378 // Convert to funnel shift intrinsic. 2379 Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr; 2380 Function *F = Intrinsic::getDeclaration(Sel.getModule(), IID, Sel.getType()); 2381 ShAmt = Builder.CreateZExt(ShAmt, Sel.getType()); 2382 return CallInst::Create(F, { SV0, SV1, ShAmt }); 2383 } 2384 2385 static Instruction *foldSelectToCopysign(SelectInst &Sel, 2386 InstCombiner::BuilderTy &Builder) { 2387 Value *Cond = Sel.getCondition(); 2388 Value *TVal = Sel.getTrueValue(); 2389 Value *FVal = Sel.getFalseValue(); 2390 Type *SelType = Sel.getType(); 2391 2392 // Match select ?, TC, FC where the constants are equal but negated. 2393 // TODO: Generalize to handle a negated variable operand? 2394 const APFloat *TC, *FC; 2395 if (!match(TVal, m_APFloat(TC)) || !match(FVal, m_APFloat(FC)) || 2396 !abs(*TC).bitwiseIsEqual(abs(*FC))) 2397 return nullptr; 2398 2399 assert(TC != FC && "Expected equal select arms to simplify"); 2400 2401 Value *X; 2402 const APInt *C; 2403 bool IsTrueIfSignSet; 2404 ICmpInst::Predicate Pred; 2405 if (!match(Cond, m_OneUse(m_ICmp(Pred, m_BitCast(m_Value(X)), m_APInt(C)))) || 2406 !InstCombiner::isSignBitCheck(Pred, *C, IsTrueIfSignSet) || 2407 X->getType() != SelType) 2408 return nullptr; 2409 2410 // If needed, negate the value that will be the sign argument of the copysign: 2411 // (bitcast X) < 0 ? -TC : TC --> copysign(TC, X) 2412 // (bitcast X) < 0 ? TC : -TC --> copysign(TC, -X) 2413 // (bitcast X) >= 0 ? -TC : TC --> copysign(TC, -X) 2414 // (bitcast X) >= 0 ? TC : -TC --> copysign(TC, X) 2415 if (IsTrueIfSignSet ^ TC->isNegative()) 2416 X = Builder.CreateFNegFMF(X, &Sel); 2417 2418 // Canonicalize the magnitude argument as the positive constant since we do 2419 // not care about its sign. 2420 Value *MagArg = TC->isNegative() ? FVal : TVal; 2421 Function *F = Intrinsic::getDeclaration(Sel.getModule(), Intrinsic::copysign, 2422 Sel.getType()); 2423 Instruction *CopySign = CallInst::Create(F, { MagArg, X }); 2424 CopySign->setFastMathFlags(Sel.getFastMathFlags()); 2425 return CopySign; 2426 } 2427 2428 Instruction *InstCombinerImpl::foldVectorSelect(SelectInst &Sel) { 2429 auto *VecTy = dyn_cast<FixedVectorType>(Sel.getType()); 2430 if (!VecTy) 2431 return nullptr; 2432 2433 unsigned NumElts = VecTy->getNumElements(); 2434 APInt UndefElts(NumElts, 0); 2435 APInt AllOnesEltMask(APInt::getAllOnesValue(NumElts)); 2436 if (Value *V = SimplifyDemandedVectorElts(&Sel, AllOnesEltMask, UndefElts)) { 2437 if (V != &Sel) 2438 return replaceInstUsesWith(Sel, V); 2439 return &Sel; 2440 } 2441 2442 // A select of a "select shuffle" with a common operand can be rearranged 2443 // to select followed by "select shuffle". Because of poison, this only works 2444 // in the case of a shuffle with no undefined mask elements. 2445 Value *Cond = Sel.getCondition(); 2446 Value *TVal = Sel.getTrueValue(); 2447 Value *FVal = Sel.getFalseValue(); 2448 Value *X, *Y; 2449 ArrayRef<int> Mask; 2450 if (match(TVal, m_OneUse(m_Shuffle(m_Value(X), m_Value(Y), m_Mask(Mask)))) && 2451 !is_contained(Mask, UndefMaskElem) && 2452 cast<ShuffleVectorInst>(TVal)->isSelect()) { 2453 if (X == FVal) { 2454 // select Cond, (shuf_sel X, Y), X --> shuf_sel X, (select Cond, Y, X) 2455 Value *NewSel = Builder.CreateSelect(Cond, Y, X, "sel", &Sel); 2456 return new ShuffleVectorInst(X, NewSel, Mask); 2457 } 2458 if (Y == FVal) { 2459 // select Cond, (shuf_sel X, Y), Y --> shuf_sel (select Cond, X, Y), Y 2460 Value *NewSel = Builder.CreateSelect(Cond, X, Y, "sel", &Sel); 2461 return new ShuffleVectorInst(NewSel, Y, Mask); 2462 } 2463 } 2464 if (match(FVal, m_OneUse(m_Shuffle(m_Value(X), m_Value(Y), m_Mask(Mask)))) && 2465 !is_contained(Mask, UndefMaskElem) && 2466 cast<ShuffleVectorInst>(FVal)->isSelect()) { 2467 if (X == TVal) { 2468 // select Cond, X, (shuf_sel X, Y) --> shuf_sel X, (select Cond, X, Y) 2469 Value *NewSel = Builder.CreateSelect(Cond, X, Y, "sel", &Sel); 2470 return new ShuffleVectorInst(X, NewSel, Mask); 2471 } 2472 if (Y == TVal) { 2473 // select Cond, Y, (shuf_sel X, Y) --> shuf_sel (select Cond, Y, X), Y 2474 Value *NewSel = Builder.CreateSelect(Cond, Y, X, "sel", &Sel); 2475 return new ShuffleVectorInst(NewSel, Y, Mask); 2476 } 2477 } 2478 2479 return nullptr; 2480 } 2481 2482 static Instruction *foldSelectToPhiImpl(SelectInst &Sel, BasicBlock *BB, 2483 const DominatorTree &DT, 2484 InstCombiner::BuilderTy &Builder) { 2485 // Find the block's immediate dominator that ends with a conditional branch 2486 // that matches select's condition (maybe inverted). 2487 auto *IDomNode = DT[BB]->getIDom(); 2488 if (!IDomNode) 2489 return nullptr; 2490 BasicBlock *IDom = IDomNode->getBlock(); 2491 2492 Value *Cond = Sel.getCondition(); 2493 Value *IfTrue, *IfFalse; 2494 BasicBlock *TrueSucc, *FalseSucc; 2495 if (match(IDom->getTerminator(), 2496 m_Br(m_Specific(Cond), m_BasicBlock(TrueSucc), 2497 m_BasicBlock(FalseSucc)))) { 2498 IfTrue = Sel.getTrueValue(); 2499 IfFalse = Sel.getFalseValue(); 2500 } else if (match(IDom->getTerminator(), 2501 m_Br(m_Not(m_Specific(Cond)), m_BasicBlock(TrueSucc), 2502 m_BasicBlock(FalseSucc)))) { 2503 IfTrue = Sel.getFalseValue(); 2504 IfFalse = Sel.getTrueValue(); 2505 } else 2506 return nullptr; 2507 2508 // Make sure the branches are actually different. 2509 if (TrueSucc == FalseSucc) 2510 return nullptr; 2511 2512 // We want to replace select %cond, %a, %b with a phi that takes value %a 2513 // for all incoming edges that are dominated by condition `%cond == true`, 2514 // and value %b for edges dominated by condition `%cond == false`. If %a 2515 // or %b are also phis from the same basic block, we can go further and take 2516 // their incoming values from the corresponding blocks. 2517 BasicBlockEdge TrueEdge(IDom, TrueSucc); 2518 BasicBlockEdge FalseEdge(IDom, FalseSucc); 2519 DenseMap<BasicBlock *, Value *> Inputs; 2520 for (auto *Pred : predecessors(BB)) { 2521 // Check implication. 2522 BasicBlockEdge Incoming(Pred, BB); 2523 if (DT.dominates(TrueEdge, Incoming)) 2524 Inputs[Pred] = IfTrue->DoPHITranslation(BB, Pred); 2525 else if (DT.dominates(FalseEdge, Incoming)) 2526 Inputs[Pred] = IfFalse->DoPHITranslation(BB, Pred); 2527 else 2528 return nullptr; 2529 // Check availability. 2530 if (auto *Insn = dyn_cast<Instruction>(Inputs[Pred])) 2531 if (!DT.dominates(Insn, Pred->getTerminator())) 2532 return nullptr; 2533 } 2534 2535 Builder.SetInsertPoint(&*BB->begin()); 2536 auto *PN = Builder.CreatePHI(Sel.getType(), Inputs.size()); 2537 for (auto *Pred : predecessors(BB)) 2538 PN->addIncoming(Inputs[Pred], Pred); 2539 PN->takeName(&Sel); 2540 return PN; 2541 } 2542 2543 static Instruction *foldSelectToPhi(SelectInst &Sel, const DominatorTree &DT, 2544 InstCombiner::BuilderTy &Builder) { 2545 // Try to replace this select with Phi in one of these blocks. 2546 SmallSetVector<BasicBlock *, 4> CandidateBlocks; 2547 CandidateBlocks.insert(Sel.getParent()); 2548 for (Value *V : Sel.operands()) 2549 if (auto *I = dyn_cast<Instruction>(V)) 2550 CandidateBlocks.insert(I->getParent()); 2551 2552 for (BasicBlock *BB : CandidateBlocks) 2553 if (auto *PN = foldSelectToPhiImpl(Sel, BB, DT, Builder)) 2554 return PN; 2555 return nullptr; 2556 } 2557 2558 static Value *foldSelectWithFrozenICmp(SelectInst &Sel, InstCombiner::BuilderTy &Builder) { 2559 FreezeInst *FI = dyn_cast<FreezeInst>(Sel.getCondition()); 2560 if (!FI) 2561 return nullptr; 2562 2563 Value *Cond = FI->getOperand(0); 2564 Value *TrueVal = Sel.getTrueValue(), *FalseVal = Sel.getFalseValue(); 2565 2566 // select (freeze(x == y)), x, y --> y 2567 // select (freeze(x != y)), x, y --> x 2568 // The freeze should be only used by this select. Otherwise, remaining uses of 2569 // the freeze can observe a contradictory value. 2570 // c = freeze(x == y) ; Let's assume that y = poison & x = 42; c is 0 or 1 2571 // a = select c, x, y ; 2572 // f(a, c) ; f(poison, 1) cannot happen, but if a is folded 2573 // ; to y, this can happen. 2574 CmpInst::Predicate Pred; 2575 if (FI->hasOneUse() && 2576 match(Cond, m_c_ICmp(Pred, m_Specific(TrueVal), m_Specific(FalseVal))) && 2577 (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE)) { 2578 return Pred == ICmpInst::ICMP_EQ ? FalseVal : TrueVal; 2579 } 2580 2581 return nullptr; 2582 } 2583 2584 Instruction *InstCombinerImpl::foldAndOrOfSelectUsingImpliedCond(Value *Op, 2585 SelectInst &SI, 2586 bool IsAnd) { 2587 Value *CondVal = SI.getCondition(); 2588 Value *A = SI.getTrueValue(); 2589 Value *B = SI.getFalseValue(); 2590 2591 assert(Op->getType()->isIntOrIntVectorTy(1) && 2592 "Op must be either i1 or vector of i1."); 2593 2594 Optional<bool> Res = isImpliedCondition(Op, CondVal, DL, IsAnd); 2595 if (!Res) 2596 return nullptr; 2597 2598 Value *Zero = Constant::getNullValue(A->getType()); 2599 Value *One = Constant::getAllOnesValue(A->getType()); 2600 2601 if (*Res == true) { 2602 if (IsAnd) 2603 // select op, (select cond, A, B), false => select op, A, false 2604 // and op, (select cond, A, B) => select op, A, false 2605 // if op = true implies condval = true. 2606 return SelectInst::Create(Op, A, Zero); 2607 else 2608 // select op, true, (select cond, A, B) => select op, true, A 2609 // or op, (select cond, A, B) => select op, true, A 2610 // if op = false implies condval = true. 2611 return SelectInst::Create(Op, One, A); 2612 } else { 2613 if (IsAnd) 2614 // select op, (select cond, A, B), false => select op, B, false 2615 // and op, (select cond, A, B) => select op, B, false 2616 // if op = true implies condval = false. 2617 return SelectInst::Create(Op, B, Zero); 2618 else 2619 // select op, true, (select cond, A, B) => select op, true, B 2620 // or op, (select cond, A, B) => select op, true, B 2621 // if op = false implies condval = false. 2622 return SelectInst::Create(Op, One, B); 2623 } 2624 } 2625 2626 Instruction *InstCombinerImpl::visitSelectInst(SelectInst &SI) { 2627 Value *CondVal = SI.getCondition(); 2628 Value *TrueVal = SI.getTrueValue(); 2629 Value *FalseVal = SI.getFalseValue(); 2630 Type *SelType = SI.getType(); 2631 2632 // FIXME: Remove this workaround when freeze related patches are done. 2633 // For select with undef operand which feeds into an equality comparison, 2634 // don't simplify it so loop unswitch can know the equality comparison 2635 // may have an undef operand. This is a workaround for PR31652 caused by 2636 // descrepancy about branch on undef between LoopUnswitch and GVN. 2637 if (match(TrueVal, m_Undef()) || match(FalseVal, m_Undef())) { 2638 if (llvm::any_of(SI.users(), [&](User *U) { 2639 ICmpInst *CI = dyn_cast<ICmpInst>(U); 2640 if (CI && CI->isEquality()) 2641 return true; 2642 return false; 2643 })) { 2644 return nullptr; 2645 } 2646 } 2647 2648 if (Value *V = SimplifySelectInst(CondVal, TrueVal, FalseVal, 2649 SQ.getWithInstruction(&SI))) 2650 return replaceInstUsesWith(SI, V); 2651 2652 if (Instruction *I = canonicalizeSelectToShuffle(SI)) 2653 return I; 2654 2655 if (Instruction *I = canonicalizeScalarSelectOfVecs(SI, *this)) 2656 return I; 2657 2658 CmpInst::Predicate Pred; 2659 2660 // Avoid potential infinite loops by checking for non-constant condition. 2661 // TODO: Can we assert instead by improving canonicalizeSelectToShuffle()? 2662 // Scalar select must have simplified? 2663 if (SelType->isIntOrIntVectorTy(1) && !isa<Constant>(CondVal) && 2664 TrueVal->getType() == CondVal->getType()) { 2665 // Folding select to and/or i1 isn't poison safe in general. impliesPoison 2666 // checks whether folding it does not convert a well-defined value into 2667 // poison. 2668 if (match(TrueVal, m_One()) && impliesPoison(FalseVal, CondVal)) { 2669 // Change: A = select B, true, C --> A = or B, C 2670 return BinaryOperator::CreateOr(CondVal, FalseVal); 2671 } 2672 if (match(FalseVal, m_Zero()) && impliesPoison(TrueVal, CondVal)) { 2673 // Change: A = select B, C, false --> A = and B, C 2674 return BinaryOperator::CreateAnd(CondVal, TrueVal); 2675 } 2676 2677 auto *One = ConstantInt::getTrue(SelType); 2678 auto *Zero = ConstantInt::getFalse(SelType); 2679 2680 // We match the "full" 0 or 1 constant here to avoid a potential infinite 2681 // loop with vectors that may have undefined/poison elements. 2682 // select a, false, b -> select !a, b, false 2683 if (match(TrueVal, m_Specific(Zero))) { 2684 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName()); 2685 return SelectInst::Create(NotCond, FalseVal, Zero); 2686 } 2687 // select a, b, true -> select !a, true, b 2688 if (match(FalseVal, m_Specific(One))) { 2689 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName()); 2690 return SelectInst::Create(NotCond, One, TrueVal); 2691 } 2692 2693 // select a, a, b -> select a, true, b 2694 if (CondVal == TrueVal) 2695 return replaceOperand(SI, 1, One); 2696 // select a, b, a -> select a, b, false 2697 if (CondVal == FalseVal) 2698 return replaceOperand(SI, 2, Zero); 2699 2700 // select a, !a, b -> select !a, b, false 2701 if (match(TrueVal, m_Not(m_Specific(CondVal)))) 2702 return SelectInst::Create(TrueVal, FalseVal, Zero); 2703 // select a, b, !a -> select !a, true, b 2704 if (match(FalseVal, m_Not(m_Specific(CondVal)))) 2705 return SelectInst::Create(FalseVal, One, TrueVal); 2706 2707 Value *A, *B; 2708 2709 // DeMorgan in select form: !a && !b --> !(a || b) 2710 // select !a, !b, false --> not (select a, true, b) 2711 if (match(&SI, m_LogicalAnd(m_Not(m_Value(A)), m_Not(m_Value(B)))) && 2712 (CondVal->hasOneUse() || TrueVal->hasOneUse()) && 2713 !match(A, m_ConstantExpr()) && !match(B, m_ConstantExpr())) 2714 return BinaryOperator::CreateNot(Builder.CreateSelect(A, One, B)); 2715 2716 // DeMorgan in select form: !a || !b --> !(a && b) 2717 // select !a, true, !b --> not (select a, b, false) 2718 if (match(&SI, m_LogicalOr(m_Not(m_Value(A)), m_Not(m_Value(B)))) && 2719 (CondVal->hasOneUse() || FalseVal->hasOneUse()) && 2720 !match(A, m_ConstantExpr()) && !match(B, m_ConstantExpr())) 2721 return BinaryOperator::CreateNot(Builder.CreateSelect(A, B, Zero)); 2722 2723 // select (select a, true, b), true, b -> select a, true, b 2724 if (match(CondVal, m_Select(m_Value(A), m_One(), m_Value(B))) && 2725 match(TrueVal, m_One()) && match(FalseVal, m_Specific(B))) 2726 return replaceOperand(SI, 0, A); 2727 // select (select a, b, false), b, false -> select a, b, false 2728 if (match(CondVal, m_Select(m_Value(A), m_Value(B), m_Zero())) && 2729 match(TrueVal, m_Specific(B)) && match(FalseVal, m_Zero())) 2730 return replaceOperand(SI, 0, A); 2731 2732 if (!SelType->isVectorTy()) { 2733 if (Value *S = simplifyWithOpReplaced(TrueVal, CondVal, One, SQ, 2734 /* AllowRefinement */ true)) 2735 return replaceOperand(SI, 1, S); 2736 if (Value *S = simplifyWithOpReplaced(FalseVal, CondVal, Zero, SQ, 2737 /* AllowRefinement */ true)) 2738 return replaceOperand(SI, 2, S); 2739 } 2740 2741 if (match(FalseVal, m_Zero()) || match(TrueVal, m_One())) { 2742 Use *Y = nullptr; 2743 bool IsAnd = match(FalseVal, m_Zero()) ? true : false; 2744 Value *Op1 = IsAnd ? TrueVal : FalseVal; 2745 if (isCheckForZeroAndMulWithOverflow(CondVal, Op1, IsAnd, Y)) { 2746 auto *FI = new FreezeInst(*Y, (*Y)->getName() + ".fr"); 2747 InsertNewInstBefore(FI, *cast<Instruction>(Y->getUser())); 2748 replaceUse(*Y, FI); 2749 return replaceInstUsesWith(SI, Op1); 2750 } 2751 2752 if (auto *Op1SI = dyn_cast<SelectInst>(Op1)) 2753 if (auto *I = foldAndOrOfSelectUsingImpliedCond(CondVal, *Op1SI, 2754 /* IsAnd */ IsAnd)) 2755 return I; 2756 2757 if (auto *ICmp0 = dyn_cast<ICmpInst>(CondVal)) 2758 if (auto *ICmp1 = dyn_cast<ICmpInst>(Op1)) 2759 if (auto *V = foldAndOrOfICmpsOfAndWithPow2(ICmp0, ICmp1, &SI, IsAnd, 2760 /* IsLogical */ true)) 2761 return replaceInstUsesWith(SI, V); 2762 } 2763 2764 // select (select a, true, b), c, false -> select a, c, false 2765 // select c, (select a, true, b), false -> select c, a, false 2766 // if c implies that b is false. 2767 if (match(CondVal, m_Select(m_Value(A), m_One(), m_Value(B))) && 2768 match(FalseVal, m_Zero())) { 2769 Optional<bool> Res = isImpliedCondition(TrueVal, B, DL); 2770 if (Res && *Res == false) 2771 return replaceOperand(SI, 0, A); 2772 } 2773 if (match(TrueVal, m_Select(m_Value(A), m_One(), m_Value(B))) && 2774 match(FalseVal, m_Zero())) { 2775 Optional<bool> Res = isImpliedCondition(CondVal, B, DL); 2776 if (Res && *Res == false) 2777 return replaceOperand(SI, 1, A); 2778 } 2779 // select c, true, (select a, b, false) -> select c, true, a 2780 // select (select a, b, false), true, c -> select a, true, c 2781 // if c = false implies that b = true 2782 if (match(TrueVal, m_One()) && 2783 match(FalseVal, m_Select(m_Value(A), m_Value(B), m_Zero()))) { 2784 Optional<bool> Res = isImpliedCondition(CondVal, B, DL, false); 2785 if (Res && *Res == true) 2786 return replaceOperand(SI, 2, A); 2787 } 2788 if (match(CondVal, m_Select(m_Value(A), m_Value(B), m_Zero())) && 2789 match(TrueVal, m_One())) { 2790 Optional<bool> Res = isImpliedCondition(FalseVal, B, DL, false); 2791 if (Res && *Res == true) 2792 return replaceOperand(SI, 0, A); 2793 } 2794 2795 // sel (sel c, a, false), true, (sel !c, b, false) -> sel c, a, b 2796 // sel (sel !c, a, false), true, (sel c, b, false) -> sel c, b, a 2797 Value *C1, *C2; 2798 if (match(CondVal, m_Select(m_Value(C1), m_Value(A), m_Zero())) && 2799 match(TrueVal, m_One()) && 2800 match(FalseVal, m_Select(m_Value(C2), m_Value(B), m_Zero()))) { 2801 if (match(C2, m_Not(m_Specific(C1)))) // first case 2802 return SelectInst::Create(C1, A, B); 2803 else if (match(C1, m_Not(m_Specific(C2)))) // second case 2804 return SelectInst::Create(C2, B, A); 2805 } 2806 } 2807 2808 // Selecting between two integer or vector splat integer constants? 2809 // 2810 // Note that we don't handle a scalar select of vectors: 2811 // select i1 %c, <2 x i8> <1, 1>, <2 x i8> <0, 0> 2812 // because that may need 3 instructions to splat the condition value: 2813 // extend, insertelement, shufflevector. 2814 // 2815 // Do not handle i1 TrueVal and FalseVal otherwise would result in 2816 // zext/sext i1 to i1. 2817 if (SelType->isIntOrIntVectorTy() && !SelType->isIntOrIntVectorTy(1) && 2818 CondVal->getType()->isVectorTy() == SelType->isVectorTy()) { 2819 // select C, 1, 0 -> zext C to int 2820 if (match(TrueVal, m_One()) && match(FalseVal, m_Zero())) 2821 return new ZExtInst(CondVal, SelType); 2822 2823 // select C, -1, 0 -> sext C to int 2824 if (match(TrueVal, m_AllOnes()) && match(FalseVal, m_Zero())) 2825 return new SExtInst(CondVal, SelType); 2826 2827 // select C, 0, 1 -> zext !C to int 2828 if (match(TrueVal, m_Zero()) && match(FalseVal, m_One())) { 2829 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName()); 2830 return new ZExtInst(NotCond, SelType); 2831 } 2832 2833 // select C, 0, -1 -> sext !C to int 2834 if (match(TrueVal, m_Zero()) && match(FalseVal, m_AllOnes())) { 2835 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName()); 2836 return new SExtInst(NotCond, SelType); 2837 } 2838 } 2839 2840 if (auto *FCmp = dyn_cast<FCmpInst>(CondVal)) { 2841 Value *Cmp0 = FCmp->getOperand(0), *Cmp1 = FCmp->getOperand(1); 2842 // Are we selecting a value based on a comparison of the two values? 2843 if ((Cmp0 == TrueVal && Cmp1 == FalseVal) || 2844 (Cmp0 == FalseVal && Cmp1 == TrueVal)) { 2845 // Canonicalize to use ordered comparisons by swapping the select 2846 // operands. 2847 // 2848 // e.g. 2849 // (X ugt Y) ? X : Y -> (X ole Y) ? Y : X 2850 if (FCmp->hasOneUse() && FCmpInst::isUnordered(FCmp->getPredicate())) { 2851 FCmpInst::Predicate InvPred = FCmp->getInversePredicate(); 2852 IRBuilder<>::FastMathFlagGuard FMFG(Builder); 2853 // FIXME: The FMF should propagate from the select, not the fcmp. 2854 Builder.setFastMathFlags(FCmp->getFastMathFlags()); 2855 Value *NewCond = Builder.CreateFCmp(InvPred, Cmp0, Cmp1, 2856 FCmp->getName() + ".inv"); 2857 Value *NewSel = Builder.CreateSelect(NewCond, FalseVal, TrueVal); 2858 return replaceInstUsesWith(SI, NewSel); 2859 } 2860 2861 // NOTE: if we wanted to, this is where to detect MIN/MAX 2862 } 2863 } 2864 2865 // Canonicalize select with fcmp to fabs(). -0.0 makes this tricky. We need 2866 // fast-math-flags (nsz) or fsub with +0.0 (not fneg) for this to work. We 2867 // also require nnan because we do not want to unintentionally change the 2868 // sign of a NaN value. 2869 // (X <= +/-0.0) ? (0.0 - X) : X --> fabs(X) 2870 Instruction *FSub; 2871 if (match(CondVal, m_FCmp(Pred, m_Specific(FalseVal), m_AnyZeroFP())) && 2872 match(TrueVal, m_FSub(m_PosZeroFP(), m_Specific(FalseVal))) && 2873 match(TrueVal, m_Instruction(FSub)) && FSub->hasNoNaNs() && 2874 (Pred == FCmpInst::FCMP_OLE || Pred == FCmpInst::FCMP_ULE)) { 2875 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, FalseVal, &SI); 2876 return replaceInstUsesWith(SI, Fabs); 2877 } 2878 // (X > +/-0.0) ? X : (0.0 - X) --> fabs(X) 2879 if (match(CondVal, m_FCmp(Pred, m_Specific(TrueVal), m_AnyZeroFP())) && 2880 match(FalseVal, m_FSub(m_PosZeroFP(), m_Specific(TrueVal))) && 2881 match(FalseVal, m_Instruction(FSub)) && FSub->hasNoNaNs() && 2882 (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_UGT)) { 2883 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, TrueVal, &SI); 2884 return replaceInstUsesWith(SI, Fabs); 2885 } 2886 // With nnan and nsz: 2887 // (X < +/-0.0) ? -X : X --> fabs(X) 2888 // (X <= +/-0.0) ? -X : X --> fabs(X) 2889 Instruction *FNeg; 2890 if (match(CondVal, m_FCmp(Pred, m_Specific(FalseVal), m_AnyZeroFP())) && 2891 match(TrueVal, m_FNeg(m_Specific(FalseVal))) && 2892 match(TrueVal, m_Instruction(FNeg)) && FNeg->hasNoNaNs() && 2893 FNeg->hasNoSignedZeros() && SI.hasNoSignedZeros() && 2894 (Pred == FCmpInst::FCMP_OLT || Pred == FCmpInst::FCMP_OLE || 2895 Pred == FCmpInst::FCMP_ULT || Pred == FCmpInst::FCMP_ULE)) { 2896 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, FalseVal, &SI); 2897 return replaceInstUsesWith(SI, Fabs); 2898 } 2899 // With nnan and nsz: 2900 // (X > +/-0.0) ? X : -X --> fabs(X) 2901 // (X >= +/-0.0) ? X : -X --> fabs(X) 2902 if (match(CondVal, m_FCmp(Pred, m_Specific(TrueVal), m_AnyZeroFP())) && 2903 match(FalseVal, m_FNeg(m_Specific(TrueVal))) && 2904 match(FalseVal, m_Instruction(FNeg)) && FNeg->hasNoNaNs() && 2905 FNeg->hasNoSignedZeros() && SI.hasNoSignedZeros() && 2906 (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_OGE || 2907 Pred == FCmpInst::FCMP_UGT || Pred == FCmpInst::FCMP_UGE)) { 2908 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, TrueVal, &SI); 2909 return replaceInstUsesWith(SI, Fabs); 2910 } 2911 2912 // See if we are selecting two values based on a comparison of the two values. 2913 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) 2914 if (Instruction *Result = foldSelectInstWithICmp(SI, ICI)) 2915 return Result; 2916 2917 if (Instruction *Add = foldAddSubSelect(SI, Builder)) 2918 return Add; 2919 if (Instruction *Add = foldOverflowingAddSubSelect(SI, Builder)) 2920 return Add; 2921 if (Instruction *Or = foldSetClearBits(SI, Builder)) 2922 return Or; 2923 2924 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z)) 2925 auto *TI = dyn_cast<Instruction>(TrueVal); 2926 auto *FI = dyn_cast<Instruction>(FalseVal); 2927 if (TI && FI && TI->getOpcode() == FI->getOpcode()) 2928 if (Instruction *IV = foldSelectOpOp(SI, TI, FI)) 2929 return IV; 2930 2931 if (Instruction *I = foldSelectExtConst(SI)) 2932 return I; 2933 2934 // Fold (select C, (gep Ptr, Idx), Ptr) -> (gep Ptr, (select C, Idx, 0)) 2935 // Fold (select C, Ptr, (gep Ptr, Idx)) -> (gep Ptr, (select C, 0, Idx)) 2936 auto SelectGepWithBase = [&](GetElementPtrInst *Gep, Value *Base, 2937 bool Swap) -> GetElementPtrInst * { 2938 Value *Ptr = Gep->getPointerOperand(); 2939 if (Gep->getNumOperands() != 2 || Gep->getPointerOperand() != Base || 2940 !Gep->hasOneUse()) 2941 return nullptr; 2942 Type *ElementType = Gep->getResultElementType(); 2943 Value *Idx = Gep->getOperand(1); 2944 Value *NewT = Idx; 2945 Value *NewF = Constant::getNullValue(Idx->getType()); 2946 if (Swap) 2947 std::swap(NewT, NewF); 2948 Value *NewSI = 2949 Builder.CreateSelect(CondVal, NewT, NewF, SI.getName() + ".idx", &SI); 2950 return GetElementPtrInst::Create(ElementType, Ptr, {NewSI}); 2951 }; 2952 if (auto *TrueGep = dyn_cast<GetElementPtrInst>(TrueVal)) 2953 if (auto *NewGep = SelectGepWithBase(TrueGep, FalseVal, false)) 2954 return NewGep; 2955 if (auto *FalseGep = dyn_cast<GetElementPtrInst>(FalseVal)) 2956 if (auto *NewGep = SelectGepWithBase(FalseGep, TrueVal, true)) 2957 return NewGep; 2958 2959 // See if we can fold the select into one of our operands. 2960 if (SelType->isIntOrIntVectorTy() || SelType->isFPOrFPVectorTy()) { 2961 if (Instruction *FoldI = foldSelectIntoOp(SI, TrueVal, FalseVal)) 2962 return FoldI; 2963 2964 Value *LHS, *RHS; 2965 Instruction::CastOps CastOp; 2966 SelectPatternResult SPR = matchSelectPattern(&SI, LHS, RHS, &CastOp); 2967 auto SPF = SPR.Flavor; 2968 if (SPF) { 2969 Value *LHS2, *RHS2; 2970 if (SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor) 2971 if (Instruction *R = foldSPFofSPF(cast<Instruction>(LHS), SPF2, LHS2, 2972 RHS2, SI, SPF, RHS)) 2973 return R; 2974 if (SelectPatternFlavor SPF2 = matchSelectPattern(RHS, LHS2, RHS2).Flavor) 2975 if (Instruction *R = foldSPFofSPF(cast<Instruction>(RHS), SPF2, LHS2, 2976 RHS2, SI, SPF, LHS)) 2977 return R; 2978 // TODO. 2979 // ABS(-X) -> ABS(X) 2980 } 2981 2982 if (SelectPatternResult::isMinOrMax(SPF)) { 2983 // Canonicalize so that 2984 // - type casts are outside select patterns. 2985 // - float clamp is transformed to min/max pattern 2986 2987 bool IsCastNeeded = LHS->getType() != SelType; 2988 Value *CmpLHS = cast<CmpInst>(CondVal)->getOperand(0); 2989 Value *CmpRHS = cast<CmpInst>(CondVal)->getOperand(1); 2990 if (IsCastNeeded || 2991 (LHS->getType()->isFPOrFPVectorTy() && 2992 ((CmpLHS != LHS && CmpLHS != RHS) || 2993 (CmpRHS != LHS && CmpRHS != RHS)))) { 2994 CmpInst::Predicate MinMaxPred = getMinMaxPred(SPF, SPR.Ordered); 2995 2996 Value *Cmp; 2997 if (CmpInst::isIntPredicate(MinMaxPred)) { 2998 Cmp = Builder.CreateICmp(MinMaxPred, LHS, RHS); 2999 } else { 3000 IRBuilder<>::FastMathFlagGuard FMFG(Builder); 3001 auto FMF = 3002 cast<FPMathOperator>(SI.getCondition())->getFastMathFlags(); 3003 Builder.setFastMathFlags(FMF); 3004 Cmp = Builder.CreateFCmp(MinMaxPred, LHS, RHS); 3005 } 3006 3007 Value *NewSI = Builder.CreateSelect(Cmp, LHS, RHS, SI.getName(), &SI); 3008 if (!IsCastNeeded) 3009 return replaceInstUsesWith(SI, NewSI); 3010 3011 Value *NewCast = Builder.CreateCast(CastOp, NewSI, SelType); 3012 return replaceInstUsesWith(SI, NewCast); 3013 } 3014 3015 // MAX(~a, ~b) -> ~MIN(a, b) 3016 // MAX(~a, C) -> ~MIN(a, ~C) 3017 // MIN(~a, ~b) -> ~MAX(a, b) 3018 // MIN(~a, C) -> ~MAX(a, ~C) 3019 auto moveNotAfterMinMax = [&](Value *X, Value *Y) -> Instruction * { 3020 Value *A; 3021 if (match(X, m_Not(m_Value(A))) && !X->hasNUsesOrMore(3) && 3022 !isFreeToInvert(A, A->hasOneUse()) && 3023 // Passing false to only consider m_Not and constants. 3024 isFreeToInvert(Y, false)) { 3025 Value *B = Builder.CreateNot(Y); 3026 Value *NewMinMax = createMinMax(Builder, getInverseMinMaxFlavor(SPF), 3027 A, B); 3028 // Copy the profile metadata. 3029 if (MDNode *MD = SI.getMetadata(LLVMContext::MD_prof)) { 3030 cast<SelectInst>(NewMinMax)->setMetadata(LLVMContext::MD_prof, MD); 3031 // Swap the metadata if the operands are swapped. 3032 if (X == SI.getFalseValue() && Y == SI.getTrueValue()) 3033 cast<SelectInst>(NewMinMax)->swapProfMetadata(); 3034 } 3035 3036 return BinaryOperator::CreateNot(NewMinMax); 3037 } 3038 3039 return nullptr; 3040 }; 3041 3042 if (Instruction *I = moveNotAfterMinMax(LHS, RHS)) 3043 return I; 3044 if (Instruction *I = moveNotAfterMinMax(RHS, LHS)) 3045 return I; 3046 3047 if (Instruction *I = moveAddAfterMinMax(SPF, LHS, RHS, Builder)) 3048 return I; 3049 3050 if (Instruction *I = factorizeMinMaxTree(SPF, LHS, RHS, Builder)) 3051 return I; 3052 if (Instruction *I = matchSAddSubSat(SI)) 3053 return I; 3054 } 3055 } 3056 3057 // Canonicalize select of FP values where NaN and -0.0 are not valid as 3058 // minnum/maxnum intrinsics. 3059 if (isa<FPMathOperator>(SI) && SI.hasNoNaNs() && SI.hasNoSignedZeros()) { 3060 Value *X, *Y; 3061 if (match(&SI, m_OrdFMax(m_Value(X), m_Value(Y)))) 3062 return replaceInstUsesWith( 3063 SI, Builder.CreateBinaryIntrinsic(Intrinsic::maxnum, X, Y, &SI)); 3064 3065 if (match(&SI, m_OrdFMin(m_Value(X), m_Value(Y)))) 3066 return replaceInstUsesWith( 3067 SI, Builder.CreateBinaryIntrinsic(Intrinsic::minnum, X, Y, &SI)); 3068 } 3069 3070 // See if we can fold the select into a phi node if the condition is a select. 3071 if (auto *PN = dyn_cast<PHINode>(SI.getCondition())) 3072 // The true/false values have to be live in the PHI predecessor's blocks. 3073 if (canSelectOperandBeMappingIntoPredBlock(TrueVal, SI) && 3074 canSelectOperandBeMappingIntoPredBlock(FalseVal, SI)) 3075 if (Instruction *NV = foldOpIntoPhi(SI, PN)) 3076 return NV; 3077 3078 if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) { 3079 if (TrueSI->getCondition()->getType() == CondVal->getType()) { 3080 // select(C, select(C, a, b), c) -> select(C, a, c) 3081 if (TrueSI->getCondition() == CondVal) { 3082 if (SI.getTrueValue() == TrueSI->getTrueValue()) 3083 return nullptr; 3084 return replaceOperand(SI, 1, TrueSI->getTrueValue()); 3085 } 3086 // select(C0, select(C1, a, b), b) -> select(C0&C1, a, b) 3087 // We choose this as normal form to enable folding on the And and 3088 // shortening paths for the values (this helps getUnderlyingObjects() for 3089 // example). 3090 if (TrueSI->getFalseValue() == FalseVal && TrueSI->hasOneUse()) { 3091 Value *And = Builder.CreateLogicalAnd(CondVal, TrueSI->getCondition()); 3092 replaceOperand(SI, 0, And); 3093 replaceOperand(SI, 1, TrueSI->getTrueValue()); 3094 return &SI; 3095 } 3096 } 3097 } 3098 if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) { 3099 if (FalseSI->getCondition()->getType() == CondVal->getType()) { 3100 // select(C, a, select(C, b, c)) -> select(C, a, c) 3101 if (FalseSI->getCondition() == CondVal) { 3102 if (SI.getFalseValue() == FalseSI->getFalseValue()) 3103 return nullptr; 3104 return replaceOperand(SI, 2, FalseSI->getFalseValue()); 3105 } 3106 // select(C0, a, select(C1, a, b)) -> select(C0|C1, a, b) 3107 if (FalseSI->getTrueValue() == TrueVal && FalseSI->hasOneUse()) { 3108 Value *Or = Builder.CreateLogicalOr(CondVal, FalseSI->getCondition()); 3109 replaceOperand(SI, 0, Or); 3110 replaceOperand(SI, 2, FalseSI->getFalseValue()); 3111 return &SI; 3112 } 3113 } 3114 } 3115 3116 auto canMergeSelectThroughBinop = [](BinaryOperator *BO) { 3117 // The select might be preventing a division by 0. 3118 switch (BO->getOpcode()) { 3119 default: 3120 return true; 3121 case Instruction::SRem: 3122 case Instruction::URem: 3123 case Instruction::SDiv: 3124 case Instruction::UDiv: 3125 return false; 3126 } 3127 }; 3128 3129 // Try to simplify a binop sandwiched between 2 selects with the same 3130 // condition. 3131 // select(C, binop(select(C, X, Y), W), Z) -> select(C, binop(X, W), Z) 3132 BinaryOperator *TrueBO; 3133 if (match(TrueVal, m_OneUse(m_BinOp(TrueBO))) && 3134 canMergeSelectThroughBinop(TrueBO)) { 3135 if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(0))) { 3136 if (TrueBOSI->getCondition() == CondVal) { 3137 replaceOperand(*TrueBO, 0, TrueBOSI->getTrueValue()); 3138 Worklist.push(TrueBO); 3139 return &SI; 3140 } 3141 } 3142 if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(1))) { 3143 if (TrueBOSI->getCondition() == CondVal) { 3144 replaceOperand(*TrueBO, 1, TrueBOSI->getTrueValue()); 3145 Worklist.push(TrueBO); 3146 return &SI; 3147 } 3148 } 3149 } 3150 3151 // select(C, Z, binop(select(C, X, Y), W)) -> select(C, Z, binop(Y, W)) 3152 BinaryOperator *FalseBO; 3153 if (match(FalseVal, m_OneUse(m_BinOp(FalseBO))) && 3154 canMergeSelectThroughBinop(FalseBO)) { 3155 if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(0))) { 3156 if (FalseBOSI->getCondition() == CondVal) { 3157 replaceOperand(*FalseBO, 0, FalseBOSI->getFalseValue()); 3158 Worklist.push(FalseBO); 3159 return &SI; 3160 } 3161 } 3162 if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(1))) { 3163 if (FalseBOSI->getCondition() == CondVal) { 3164 replaceOperand(*FalseBO, 1, FalseBOSI->getFalseValue()); 3165 Worklist.push(FalseBO); 3166 return &SI; 3167 } 3168 } 3169 } 3170 3171 Value *NotCond; 3172 if (match(CondVal, m_Not(m_Value(NotCond))) && 3173 !InstCombiner::shouldAvoidAbsorbingNotIntoSelect(SI)) { 3174 replaceOperand(SI, 0, NotCond); 3175 SI.swapValues(); 3176 SI.swapProfMetadata(); 3177 return &SI; 3178 } 3179 3180 if (Instruction *I = foldVectorSelect(SI)) 3181 return I; 3182 3183 // If we can compute the condition, there's no need for a select. 3184 // Like the above fold, we are attempting to reduce compile-time cost by 3185 // putting this fold here with limitations rather than in InstSimplify. 3186 // The motivation for this call into value tracking is to take advantage of 3187 // the assumption cache, so make sure that is populated. 3188 if (!CondVal->getType()->isVectorTy() && !AC.assumptions().empty()) { 3189 KnownBits Known(1); 3190 computeKnownBits(CondVal, Known, 0, &SI); 3191 if (Known.One.isOneValue()) 3192 return replaceInstUsesWith(SI, TrueVal); 3193 if (Known.Zero.isOneValue()) 3194 return replaceInstUsesWith(SI, FalseVal); 3195 } 3196 3197 if (Instruction *BitCastSel = foldSelectCmpBitcasts(SI, Builder)) 3198 return BitCastSel; 3199 3200 // Simplify selects that test the returned flag of cmpxchg instructions. 3201 if (Value *V = foldSelectCmpXchg(SI)) 3202 return replaceInstUsesWith(SI, V); 3203 3204 if (Instruction *Select = foldSelectBinOpIdentity(SI, TLI, *this)) 3205 return Select; 3206 3207 if (Instruction *Funnel = foldSelectFunnelShift(SI, Builder)) 3208 return Funnel; 3209 3210 if (Instruction *Copysign = foldSelectToCopysign(SI, Builder)) 3211 return Copysign; 3212 3213 if (Instruction *PN = foldSelectToPhi(SI, DT, Builder)) 3214 return replaceInstUsesWith(SI, PN); 3215 3216 if (Value *Fr = foldSelectWithFrozenICmp(SI, Builder)) 3217 return replaceInstUsesWith(SI, Fr); 3218 3219 // select(mask, mload(,,mask,0), 0) -> mload(,,mask,0) 3220 // Load inst is intentionally not checked for hasOneUse() 3221 if (match(FalseVal, m_Zero()) && 3222 match(TrueVal, m_MaskedLoad(m_Value(), m_Value(), m_Specific(CondVal), 3223 m_CombineOr(m_Undef(), m_Zero())))) { 3224 auto *MaskedLoad = cast<IntrinsicInst>(TrueVal); 3225 if (isa<UndefValue>(MaskedLoad->getArgOperand(3))) 3226 MaskedLoad->setArgOperand(3, FalseVal /* Zero */); 3227 return replaceInstUsesWith(SI, MaskedLoad); 3228 } 3229 3230 Value *Mask; 3231 if (match(TrueVal, m_Zero()) && 3232 match(FalseVal, m_MaskedLoad(m_Value(), m_Value(), m_Value(Mask), 3233 m_CombineOr(m_Undef(), m_Zero()))) && 3234 (CondVal->getType() == Mask->getType())) { 3235 // We can remove the select by ensuring the load zeros all lanes the 3236 // select would have. We determine this by proving there is no overlap 3237 // between the load and select masks. 3238 // (i.e (load_mask & select_mask) == 0 == no overlap) 3239 bool CanMergeSelectIntoLoad = false; 3240 if (Value *V = SimplifyAndInst(CondVal, Mask, SQ.getWithInstruction(&SI))) 3241 CanMergeSelectIntoLoad = match(V, m_Zero()); 3242 3243 if (CanMergeSelectIntoLoad) { 3244 auto *MaskedLoad = cast<IntrinsicInst>(FalseVal); 3245 if (isa<UndefValue>(MaskedLoad->getArgOperand(3))) 3246 MaskedLoad->setArgOperand(3, TrueVal /* Zero */); 3247 return replaceInstUsesWith(SI, MaskedLoad); 3248 } 3249 } 3250 3251 return nullptr; 3252 } 3253