1 //===- InstCombineAndOrXor.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 visitAnd, visitOr, and visitXor functions. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "InstCombineInternal.h" 14 #include "llvm/Analysis/CmpInstAnalysis.h" 15 #include "llvm/Analysis/InstructionSimplify.h" 16 #include "llvm/IR/ConstantRange.h" 17 #include "llvm/IR/Intrinsics.h" 18 #include "llvm/IR/PatternMatch.h" 19 #include "llvm/Transforms/InstCombine/InstCombiner.h" 20 #include "llvm/Transforms/Utils/Local.h" 21 22 using namespace llvm; 23 using namespace PatternMatch; 24 25 #define DEBUG_TYPE "instcombine" 26 27 /// This is the complement of getICmpCode, which turns an opcode and two 28 /// operands into either a constant true or false, or a brand new ICmp 29 /// instruction. The sign is passed in to determine which kind of predicate to 30 /// use in the new icmp instruction. 31 static Value *getNewICmpValue(unsigned Code, bool Sign, Value *LHS, Value *RHS, 32 InstCombiner::BuilderTy &Builder) { 33 ICmpInst::Predicate NewPred; 34 if (Constant *TorF = getPredForICmpCode(Code, Sign, LHS->getType(), NewPred)) 35 return TorF; 36 return Builder.CreateICmp(NewPred, LHS, RHS); 37 } 38 39 /// This is the complement of getFCmpCode, which turns an opcode and two 40 /// operands into either a FCmp instruction, or a true/false constant. 41 static Value *getFCmpValue(unsigned Code, Value *LHS, Value *RHS, 42 InstCombiner::BuilderTy &Builder) { 43 FCmpInst::Predicate NewPred; 44 if (Constant *TorF = getPredForFCmpCode(Code, LHS->getType(), NewPred)) 45 return TorF; 46 return Builder.CreateFCmp(NewPred, LHS, RHS); 47 } 48 49 /// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise 50 /// (V < Lo || V >= Hi). This method expects that Lo < Hi. IsSigned indicates 51 /// whether to treat V, Lo, and Hi as signed or not. 52 Value *InstCombinerImpl::insertRangeTest(Value *V, const APInt &Lo, 53 const APInt &Hi, bool isSigned, 54 bool Inside) { 55 assert((isSigned ? Lo.slt(Hi) : Lo.ult(Hi)) && 56 "Lo is not < Hi in range emission code!"); 57 58 Type *Ty = V->getType(); 59 60 // V >= Min && V < Hi --> V < Hi 61 // V < Min || V >= Hi --> V >= Hi 62 ICmpInst::Predicate Pred = Inside ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE; 63 if (isSigned ? Lo.isMinSignedValue() : Lo.isMinValue()) { 64 Pred = isSigned ? ICmpInst::getSignedPredicate(Pred) : Pred; 65 return Builder.CreateICmp(Pred, V, ConstantInt::get(Ty, Hi)); 66 } 67 68 // V >= Lo && V < Hi --> V - Lo u< Hi - Lo 69 // V < Lo || V >= Hi --> V - Lo u>= Hi - Lo 70 Value *VMinusLo = 71 Builder.CreateSub(V, ConstantInt::get(Ty, Lo), V->getName() + ".off"); 72 Constant *HiMinusLo = ConstantInt::get(Ty, Hi - Lo); 73 return Builder.CreateICmp(Pred, VMinusLo, HiMinusLo); 74 } 75 76 /// Classify (icmp eq (A & B), C) and (icmp ne (A & B), C) as matching patterns 77 /// that can be simplified. 78 /// One of A and B is considered the mask. The other is the value. This is 79 /// described as the "AMask" or "BMask" part of the enum. If the enum contains 80 /// only "Mask", then both A and B can be considered masks. If A is the mask, 81 /// then it was proven that (A & C) == C. This is trivial if C == A or C == 0. 82 /// If both A and C are constants, this proof is also easy. 83 /// For the following explanations, we assume that A is the mask. 84 /// 85 /// "AllOnes" declares that the comparison is true only if (A & B) == A or all 86 /// bits of A are set in B. 87 /// Example: (icmp eq (A & 3), 3) -> AMask_AllOnes 88 /// 89 /// "AllZeros" declares that the comparison is true only if (A & B) == 0 or all 90 /// bits of A are cleared in B. 91 /// Example: (icmp eq (A & 3), 0) -> Mask_AllZeroes 92 /// 93 /// "Mixed" declares that (A & B) == C and C might or might not contain any 94 /// number of one bits and zero bits. 95 /// Example: (icmp eq (A & 3), 1) -> AMask_Mixed 96 /// 97 /// "Not" means that in above descriptions "==" should be replaced by "!=". 98 /// Example: (icmp ne (A & 3), 3) -> AMask_NotAllOnes 99 /// 100 /// If the mask A contains a single bit, then the following is equivalent: 101 /// (icmp eq (A & B), A) equals (icmp ne (A & B), 0) 102 /// (icmp ne (A & B), A) equals (icmp eq (A & B), 0) 103 enum MaskedICmpType { 104 AMask_AllOnes = 1, 105 AMask_NotAllOnes = 2, 106 BMask_AllOnes = 4, 107 BMask_NotAllOnes = 8, 108 Mask_AllZeros = 16, 109 Mask_NotAllZeros = 32, 110 AMask_Mixed = 64, 111 AMask_NotMixed = 128, 112 BMask_Mixed = 256, 113 BMask_NotMixed = 512 114 }; 115 116 /// Return the set of patterns (from MaskedICmpType) that (icmp SCC (A & B), C) 117 /// satisfies. 118 static unsigned getMaskedICmpType(Value *A, Value *B, Value *C, 119 ICmpInst::Predicate Pred) { 120 const APInt *ConstA = nullptr, *ConstB = nullptr, *ConstC = nullptr; 121 match(A, m_APInt(ConstA)); 122 match(B, m_APInt(ConstB)); 123 match(C, m_APInt(ConstC)); 124 bool IsEq = (Pred == ICmpInst::ICMP_EQ); 125 bool IsAPow2 = ConstA && ConstA->isPowerOf2(); 126 bool IsBPow2 = ConstB && ConstB->isPowerOf2(); 127 unsigned MaskVal = 0; 128 if (ConstC && ConstC->isZero()) { 129 // if C is zero, then both A and B qualify as mask 130 MaskVal |= (IsEq ? (Mask_AllZeros | AMask_Mixed | BMask_Mixed) 131 : (Mask_NotAllZeros | AMask_NotMixed | BMask_NotMixed)); 132 if (IsAPow2) 133 MaskVal |= (IsEq ? (AMask_NotAllOnes | AMask_NotMixed) 134 : (AMask_AllOnes | AMask_Mixed)); 135 if (IsBPow2) 136 MaskVal |= (IsEq ? (BMask_NotAllOnes | BMask_NotMixed) 137 : (BMask_AllOnes | BMask_Mixed)); 138 return MaskVal; 139 } 140 141 if (A == C) { 142 MaskVal |= (IsEq ? (AMask_AllOnes | AMask_Mixed) 143 : (AMask_NotAllOnes | AMask_NotMixed)); 144 if (IsAPow2) 145 MaskVal |= (IsEq ? (Mask_NotAllZeros | AMask_NotMixed) 146 : (Mask_AllZeros | AMask_Mixed)); 147 } else if (ConstA && ConstC && ConstC->isSubsetOf(*ConstA)) { 148 MaskVal |= (IsEq ? AMask_Mixed : AMask_NotMixed); 149 } 150 151 if (B == C) { 152 MaskVal |= (IsEq ? (BMask_AllOnes | BMask_Mixed) 153 : (BMask_NotAllOnes | BMask_NotMixed)); 154 if (IsBPow2) 155 MaskVal |= (IsEq ? (Mask_NotAllZeros | BMask_NotMixed) 156 : (Mask_AllZeros | BMask_Mixed)); 157 } else if (ConstB && ConstC && ConstC->isSubsetOf(*ConstB)) { 158 MaskVal |= (IsEq ? BMask_Mixed : BMask_NotMixed); 159 } 160 161 return MaskVal; 162 } 163 164 /// Convert an analysis of a masked ICmp into its equivalent if all boolean 165 /// operations had the opposite sense. Since each "NotXXX" flag (recording !=) 166 /// is adjacent to the corresponding normal flag (recording ==), this just 167 /// involves swapping those bits over. 168 static unsigned conjugateICmpMask(unsigned Mask) { 169 unsigned NewMask; 170 NewMask = (Mask & (AMask_AllOnes | BMask_AllOnes | Mask_AllZeros | 171 AMask_Mixed | BMask_Mixed)) 172 << 1; 173 174 NewMask |= (Mask & (AMask_NotAllOnes | BMask_NotAllOnes | Mask_NotAllZeros | 175 AMask_NotMixed | BMask_NotMixed)) 176 >> 1; 177 178 return NewMask; 179 } 180 181 // Adapts the external decomposeBitTestICmp for local use. 182 static bool decomposeBitTestICmp(Value *LHS, Value *RHS, CmpInst::Predicate &Pred, 183 Value *&X, Value *&Y, Value *&Z) { 184 APInt Mask; 185 if (!llvm::decomposeBitTestICmp(LHS, RHS, Pred, X, Mask)) 186 return false; 187 188 Y = ConstantInt::get(X->getType(), Mask); 189 Z = ConstantInt::get(X->getType(), 0); 190 return true; 191 } 192 193 /// Handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E). 194 /// Return the pattern classes (from MaskedICmpType) for the left hand side and 195 /// the right hand side as a pair. 196 /// LHS and RHS are the left hand side and the right hand side ICmps and PredL 197 /// and PredR are their predicates, respectively. 198 static std::optional<std::pair<unsigned, unsigned>> getMaskedTypeForICmpPair( 199 Value *&A, Value *&B, Value *&C, Value *&D, Value *&E, ICmpInst *LHS, 200 ICmpInst *RHS, ICmpInst::Predicate &PredL, ICmpInst::Predicate &PredR) { 201 // Don't allow pointers. Splat vectors are fine. 202 if (!LHS->getOperand(0)->getType()->isIntOrIntVectorTy() || 203 !RHS->getOperand(0)->getType()->isIntOrIntVectorTy()) 204 return std::nullopt; 205 206 // Here comes the tricky part: 207 // LHS might be of the form L11 & L12 == X, X == L21 & L22, 208 // and L11 & L12 == L21 & L22. The same goes for RHS. 209 // Now we must find those components L** and R**, that are equal, so 210 // that we can extract the parameters A, B, C, D, and E for the canonical 211 // above. 212 Value *L1 = LHS->getOperand(0); 213 Value *L2 = LHS->getOperand(1); 214 Value *L11, *L12, *L21, *L22; 215 // Check whether the icmp can be decomposed into a bit test. 216 if (decomposeBitTestICmp(L1, L2, PredL, L11, L12, L2)) { 217 L21 = L22 = L1 = nullptr; 218 } else { 219 // Look for ANDs in the LHS icmp. 220 if (!match(L1, m_And(m_Value(L11), m_Value(L12)))) { 221 // Any icmp can be viewed as being trivially masked; if it allows us to 222 // remove one, it's worth it. 223 L11 = L1; 224 L12 = Constant::getAllOnesValue(L1->getType()); 225 } 226 227 if (!match(L2, m_And(m_Value(L21), m_Value(L22)))) { 228 L21 = L2; 229 L22 = Constant::getAllOnesValue(L2->getType()); 230 } 231 } 232 233 // Bail if LHS was a icmp that can't be decomposed into an equality. 234 if (!ICmpInst::isEquality(PredL)) 235 return std::nullopt; 236 237 Value *R1 = RHS->getOperand(0); 238 Value *R2 = RHS->getOperand(1); 239 Value *R11, *R12; 240 bool Ok = false; 241 if (decomposeBitTestICmp(R1, R2, PredR, R11, R12, R2)) { 242 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) { 243 A = R11; 244 D = R12; 245 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) { 246 A = R12; 247 D = R11; 248 } else { 249 return std::nullopt; 250 } 251 E = R2; 252 R1 = nullptr; 253 Ok = true; 254 } else { 255 if (!match(R1, m_And(m_Value(R11), m_Value(R12)))) { 256 // As before, model no mask as a trivial mask if it'll let us do an 257 // optimization. 258 R11 = R1; 259 R12 = Constant::getAllOnesValue(R1->getType()); 260 } 261 262 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) { 263 A = R11; 264 D = R12; 265 E = R2; 266 Ok = true; 267 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) { 268 A = R12; 269 D = R11; 270 E = R2; 271 Ok = true; 272 } 273 } 274 275 // Bail if RHS was a icmp that can't be decomposed into an equality. 276 if (!ICmpInst::isEquality(PredR)) 277 return std::nullopt; 278 279 // Look for ANDs on the right side of the RHS icmp. 280 if (!Ok) { 281 if (!match(R2, m_And(m_Value(R11), m_Value(R12)))) { 282 R11 = R2; 283 R12 = Constant::getAllOnesValue(R2->getType()); 284 } 285 286 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) { 287 A = R11; 288 D = R12; 289 E = R1; 290 Ok = true; 291 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) { 292 A = R12; 293 D = R11; 294 E = R1; 295 Ok = true; 296 } else { 297 return std::nullopt; 298 } 299 300 assert(Ok && "Failed to find AND on the right side of the RHS icmp."); 301 } 302 303 if (L11 == A) { 304 B = L12; 305 C = L2; 306 } else if (L12 == A) { 307 B = L11; 308 C = L2; 309 } else if (L21 == A) { 310 B = L22; 311 C = L1; 312 } else if (L22 == A) { 313 B = L21; 314 C = L1; 315 } 316 317 unsigned LeftType = getMaskedICmpType(A, B, C, PredL); 318 unsigned RightType = getMaskedICmpType(A, D, E, PredR); 319 return std::optional<std::pair<unsigned, unsigned>>( 320 std::make_pair(LeftType, RightType)); 321 } 322 323 /// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E) into a single 324 /// (icmp(A & X) ==/!= Y), where the left-hand side is of type Mask_NotAllZeros 325 /// and the right hand side is of type BMask_Mixed. For example, 326 /// (icmp (A & 12) != 0) & (icmp (A & 15) == 8) -> (icmp (A & 15) == 8). 327 /// Also used for logical and/or, must be poison safe. 328 static Value *foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed( 329 ICmpInst *LHS, ICmpInst *RHS, bool IsAnd, Value *A, Value *B, Value *C, 330 Value *D, Value *E, ICmpInst::Predicate PredL, ICmpInst::Predicate PredR, 331 InstCombiner::BuilderTy &Builder) { 332 // We are given the canonical form: 333 // (icmp ne (A & B), 0) & (icmp eq (A & D), E). 334 // where D & E == E. 335 // 336 // If IsAnd is false, we get it in negated form: 337 // (icmp eq (A & B), 0) | (icmp ne (A & D), E) -> 338 // !((icmp ne (A & B), 0) & (icmp eq (A & D), E)). 339 // 340 // We currently handle the case of B, C, D, E are constant. 341 // 342 const APInt *BCst, *CCst, *DCst, *OrigECst; 343 if (!match(B, m_APInt(BCst)) || !match(C, m_APInt(CCst)) || 344 !match(D, m_APInt(DCst)) || !match(E, m_APInt(OrigECst))) 345 return nullptr; 346 347 ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE; 348 349 // Update E to the canonical form when D is a power of two and RHS is 350 // canonicalized as, 351 // (icmp ne (A & D), 0) -> (icmp eq (A & D), D) or 352 // (icmp ne (A & D), D) -> (icmp eq (A & D), 0). 353 APInt ECst = *OrigECst; 354 if (PredR != NewCC) 355 ECst ^= *DCst; 356 357 // If B or D is zero, skip because if LHS or RHS can be trivially folded by 358 // other folding rules and this pattern won't apply any more. 359 if (*BCst == 0 || *DCst == 0) 360 return nullptr; 361 362 // If B and D don't intersect, ie. (B & D) == 0, no folding because we can't 363 // deduce anything from it. 364 // For example, 365 // (icmp ne (A & 12), 0) & (icmp eq (A & 3), 1) -> no folding. 366 if ((*BCst & *DCst) == 0) 367 return nullptr; 368 369 // If the following two conditions are met: 370 // 371 // 1. mask B covers only a single bit that's not covered by mask D, that is, 372 // (B & (B ^ D)) is a power of 2 (in other words, B minus the intersection of 373 // B and D has only one bit set) and, 374 // 375 // 2. RHS (and E) indicates that the rest of B's bits are zero (in other 376 // words, the intersection of B and D is zero), that is, ((B & D) & E) == 0 377 // 378 // then that single bit in B must be one and thus the whole expression can be 379 // folded to 380 // (A & (B | D)) == (B & (B ^ D)) | E. 381 // 382 // For example, 383 // (icmp ne (A & 12), 0) & (icmp eq (A & 7), 1) -> (icmp eq (A & 15), 9) 384 // (icmp ne (A & 15), 0) & (icmp eq (A & 7), 0) -> (icmp eq (A & 15), 8) 385 if ((((*BCst & *DCst) & ECst) == 0) && 386 (*BCst & (*BCst ^ *DCst)).isPowerOf2()) { 387 APInt BorD = *BCst | *DCst; 388 APInt BandBxorDorE = (*BCst & (*BCst ^ *DCst)) | ECst; 389 Value *NewMask = ConstantInt::get(A->getType(), BorD); 390 Value *NewMaskedValue = ConstantInt::get(A->getType(), BandBxorDorE); 391 Value *NewAnd = Builder.CreateAnd(A, NewMask); 392 return Builder.CreateICmp(NewCC, NewAnd, NewMaskedValue); 393 } 394 395 auto IsSubSetOrEqual = [](const APInt *C1, const APInt *C2) { 396 return (*C1 & *C2) == *C1; 397 }; 398 auto IsSuperSetOrEqual = [](const APInt *C1, const APInt *C2) { 399 return (*C1 & *C2) == *C2; 400 }; 401 402 // In the following, we consider only the cases where B is a superset of D, B 403 // is a subset of D, or B == D because otherwise there's at least one bit 404 // covered by B but not D, in which case we can't deduce much from it, so 405 // no folding (aside from the single must-be-one bit case right above.) 406 // For example, 407 // (icmp ne (A & 14), 0) & (icmp eq (A & 3), 1) -> no folding. 408 if (!IsSubSetOrEqual(BCst, DCst) && !IsSuperSetOrEqual(BCst, DCst)) 409 return nullptr; 410 411 // At this point, either B is a superset of D, B is a subset of D or B == D. 412 413 // If E is zero, if B is a subset of (or equal to) D, LHS and RHS contradict 414 // and the whole expression becomes false (or true if negated), otherwise, no 415 // folding. 416 // For example, 417 // (icmp ne (A & 3), 0) & (icmp eq (A & 7), 0) -> false. 418 // (icmp ne (A & 15), 0) & (icmp eq (A & 3), 0) -> no folding. 419 if (ECst.isZero()) { 420 if (IsSubSetOrEqual(BCst, DCst)) 421 return ConstantInt::get(LHS->getType(), !IsAnd); 422 return nullptr; 423 } 424 425 // At this point, B, D, E aren't zero and (B & D) == B, (B & D) == D or B == 426 // D. If B is a superset of (or equal to) D, since E is not zero, LHS is 427 // subsumed by RHS (RHS implies LHS.) So the whole expression becomes 428 // RHS. For example, 429 // (icmp ne (A & 255), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8). 430 // (icmp ne (A & 15), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8). 431 if (IsSuperSetOrEqual(BCst, DCst)) 432 return RHS; 433 // Otherwise, B is a subset of D. If B and E have a common bit set, 434 // ie. (B & E) != 0, then LHS is subsumed by RHS. For example. 435 // (icmp ne (A & 12), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8). 436 assert(IsSubSetOrEqual(BCst, DCst) && "Precondition due to above code"); 437 if ((*BCst & ECst) != 0) 438 return RHS; 439 // Otherwise, LHS and RHS contradict and the whole expression becomes false 440 // (or true if negated.) For example, 441 // (icmp ne (A & 7), 0) & (icmp eq (A & 15), 8) -> false. 442 // (icmp ne (A & 6), 0) & (icmp eq (A & 15), 8) -> false. 443 return ConstantInt::get(LHS->getType(), !IsAnd); 444 } 445 446 /// Try to fold (icmp(A & B) ==/!= 0) &/| (icmp(A & D) ==/!= E) into a single 447 /// (icmp(A & X) ==/!= Y), where the left-hand side and the right hand side 448 /// aren't of the common mask pattern type. 449 /// Also used for logical and/or, must be poison safe. 450 static Value *foldLogOpOfMaskedICmpsAsymmetric( 451 ICmpInst *LHS, ICmpInst *RHS, bool IsAnd, Value *A, Value *B, Value *C, 452 Value *D, Value *E, ICmpInst::Predicate PredL, ICmpInst::Predicate PredR, 453 unsigned LHSMask, unsigned RHSMask, InstCombiner::BuilderTy &Builder) { 454 assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) && 455 "Expected equality predicates for masked type of icmps."); 456 // Handle Mask_NotAllZeros-BMask_Mixed cases. 457 // (icmp ne/eq (A & B), C) &/| (icmp eq/ne (A & D), E), or 458 // (icmp eq/ne (A & B), C) &/| (icmp ne/eq (A & D), E) 459 // which gets swapped to 460 // (icmp ne/eq (A & D), E) &/| (icmp eq/ne (A & B), C). 461 if (!IsAnd) { 462 LHSMask = conjugateICmpMask(LHSMask); 463 RHSMask = conjugateICmpMask(RHSMask); 464 } 465 if ((LHSMask & Mask_NotAllZeros) && (RHSMask & BMask_Mixed)) { 466 if (Value *V = foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed( 467 LHS, RHS, IsAnd, A, B, C, D, E, 468 PredL, PredR, Builder)) { 469 return V; 470 } 471 } else if ((LHSMask & BMask_Mixed) && (RHSMask & Mask_NotAllZeros)) { 472 if (Value *V = foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed( 473 RHS, LHS, IsAnd, A, D, E, B, C, 474 PredR, PredL, Builder)) { 475 return V; 476 } 477 } 478 return nullptr; 479 } 480 481 /// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E) 482 /// into a single (icmp(A & X) ==/!= Y). 483 static Value *foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS, bool IsAnd, 484 bool IsLogical, 485 InstCombiner::BuilderTy &Builder) { 486 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr; 487 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate(); 488 std::optional<std::pair<unsigned, unsigned>> MaskPair = 489 getMaskedTypeForICmpPair(A, B, C, D, E, LHS, RHS, PredL, PredR); 490 if (!MaskPair) 491 return nullptr; 492 assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) && 493 "Expected equality predicates for masked type of icmps."); 494 unsigned LHSMask = MaskPair->first; 495 unsigned RHSMask = MaskPair->second; 496 unsigned Mask = LHSMask & RHSMask; 497 if (Mask == 0) { 498 // Even if the two sides don't share a common pattern, check if folding can 499 // still happen. 500 if (Value *V = foldLogOpOfMaskedICmpsAsymmetric( 501 LHS, RHS, IsAnd, A, B, C, D, E, PredL, PredR, LHSMask, RHSMask, 502 Builder)) 503 return V; 504 return nullptr; 505 } 506 507 // In full generality: 508 // (icmp (A & B) Op C) | (icmp (A & D) Op E) 509 // == ![ (icmp (A & B) !Op C) & (icmp (A & D) !Op E) ] 510 // 511 // If the latter can be converted into (icmp (A & X) Op Y) then the former is 512 // equivalent to (icmp (A & X) !Op Y). 513 // 514 // Therefore, we can pretend for the rest of this function that we're dealing 515 // with the conjunction, provided we flip the sense of any comparisons (both 516 // input and output). 517 518 // In most cases we're going to produce an EQ for the "&&" case. 519 ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE; 520 if (!IsAnd) { 521 // Convert the masking analysis into its equivalent with negated 522 // comparisons. 523 Mask = conjugateICmpMask(Mask); 524 } 525 526 if (Mask & Mask_AllZeros) { 527 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0) 528 // -> (icmp eq (A & (B|D)), 0) 529 if (IsLogical && !isGuaranteedNotToBeUndefOrPoison(D)) 530 return nullptr; // TODO: Use freeze? 531 Value *NewOr = Builder.CreateOr(B, D); 532 Value *NewAnd = Builder.CreateAnd(A, NewOr); 533 // We can't use C as zero because we might actually handle 534 // (icmp ne (A & B), B) & (icmp ne (A & D), D) 535 // with B and D, having a single bit set. 536 Value *Zero = Constant::getNullValue(A->getType()); 537 return Builder.CreateICmp(NewCC, NewAnd, Zero); 538 } 539 if (Mask & BMask_AllOnes) { 540 // (icmp eq (A & B), B) & (icmp eq (A & D), D) 541 // -> (icmp eq (A & (B|D)), (B|D)) 542 if (IsLogical && !isGuaranteedNotToBeUndefOrPoison(D)) 543 return nullptr; // TODO: Use freeze? 544 Value *NewOr = Builder.CreateOr(B, D); 545 Value *NewAnd = Builder.CreateAnd(A, NewOr); 546 return Builder.CreateICmp(NewCC, NewAnd, NewOr); 547 } 548 if (Mask & AMask_AllOnes) { 549 // (icmp eq (A & B), A) & (icmp eq (A & D), A) 550 // -> (icmp eq (A & (B&D)), A) 551 if (IsLogical && !isGuaranteedNotToBeUndefOrPoison(D)) 552 return nullptr; // TODO: Use freeze? 553 Value *NewAnd1 = Builder.CreateAnd(B, D); 554 Value *NewAnd2 = Builder.CreateAnd(A, NewAnd1); 555 return Builder.CreateICmp(NewCC, NewAnd2, A); 556 } 557 558 // Remaining cases assume at least that B and D are constant, and depend on 559 // their actual values. This isn't strictly necessary, just a "handle the 560 // easy cases for now" decision. 561 const APInt *ConstB, *ConstD; 562 if (!match(B, m_APInt(ConstB)) || !match(D, m_APInt(ConstD))) 563 return nullptr; 564 565 if (Mask & (Mask_NotAllZeros | BMask_NotAllOnes)) { 566 // (icmp ne (A & B), 0) & (icmp ne (A & D), 0) and 567 // (icmp ne (A & B), B) & (icmp ne (A & D), D) 568 // -> (icmp ne (A & B), 0) or (icmp ne (A & D), 0) 569 // Only valid if one of the masks is a superset of the other (check "B&D" is 570 // the same as either B or D). 571 APInt NewMask = *ConstB & *ConstD; 572 if (NewMask == *ConstB) 573 return LHS; 574 else if (NewMask == *ConstD) 575 return RHS; 576 } 577 578 if (Mask & AMask_NotAllOnes) { 579 // (icmp ne (A & B), B) & (icmp ne (A & D), D) 580 // -> (icmp ne (A & B), A) or (icmp ne (A & D), A) 581 // Only valid if one of the masks is a superset of the other (check "B|D" is 582 // the same as either B or D). 583 APInt NewMask = *ConstB | *ConstD; 584 if (NewMask == *ConstB) 585 return LHS; 586 else if (NewMask == *ConstD) 587 return RHS; 588 } 589 590 if (Mask & (BMask_Mixed | BMask_NotMixed)) { 591 // Mixed: 592 // (icmp eq (A & B), C) & (icmp eq (A & D), E) 593 // We already know that B & C == C && D & E == E. 594 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of 595 // C and E, which are shared by both the mask B and the mask D, don't 596 // contradict, then we can transform to 597 // -> (icmp eq (A & (B|D)), (C|E)) 598 // Currently, we only handle the case of B, C, D, and E being constant. 599 // We can't simply use C and E because we might actually handle 600 // (icmp ne (A & B), B) & (icmp eq (A & D), D) 601 // with B and D, having a single bit set. 602 603 // NotMixed: 604 // (icmp ne (A & B), C) & (icmp ne (A & D), E) 605 // -> (icmp ne (A & (B & D)), (C & E)) 606 // Check the intersection (B & D) for inequality. 607 // Assume that (B & D) == B || (B & D) == D, i.e B/D is a subset of D/B 608 // and (B & D) & (C ^ E) == 0, bits of C and E, which are shared by both the 609 // B and the D, don't contradict. 610 // Note that we can assume (~B & C) == 0 && (~D & E) == 0, previous 611 // operation should delete these icmps if it hadn't been met. 612 613 const APInt *OldConstC, *OldConstE; 614 if (!match(C, m_APInt(OldConstC)) || !match(E, m_APInt(OldConstE))) 615 return nullptr; 616 617 auto FoldBMixed = [&](ICmpInst::Predicate CC, bool IsNot) -> Value * { 618 CC = IsNot ? CmpInst::getInversePredicate(CC) : CC; 619 const APInt ConstC = PredL != CC ? *ConstB ^ *OldConstC : *OldConstC; 620 const APInt ConstE = PredR != CC ? *ConstD ^ *OldConstE : *OldConstE; 621 622 if (((*ConstB & *ConstD) & (ConstC ^ ConstE)).getBoolValue()) 623 return IsNot ? nullptr : ConstantInt::get(LHS->getType(), !IsAnd); 624 625 if (IsNot && !ConstB->isSubsetOf(*ConstD) && !ConstD->isSubsetOf(*ConstB)) 626 return nullptr; 627 628 APInt BD, CE; 629 if (IsNot) { 630 BD = *ConstB & *ConstD; 631 CE = ConstC & ConstE; 632 } else { 633 BD = *ConstB | *ConstD; 634 CE = ConstC | ConstE; 635 } 636 Value *NewAnd = Builder.CreateAnd(A, BD); 637 Value *CEVal = ConstantInt::get(A->getType(), CE); 638 return Builder.CreateICmp(CC, CEVal, NewAnd); 639 }; 640 641 if (Mask & BMask_Mixed) 642 return FoldBMixed(NewCC, false); 643 if (Mask & BMask_NotMixed) // can be else also 644 return FoldBMixed(NewCC, true); 645 } 646 return nullptr; 647 } 648 649 /// Try to fold a signed range checked with lower bound 0 to an unsigned icmp. 650 /// Example: (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n 651 /// If \p Inverted is true then the check is for the inverted range, e.g. 652 /// (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n 653 Value *InstCombinerImpl::simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1, 654 bool Inverted) { 655 // Check the lower range comparison, e.g. x >= 0 656 // InstCombine already ensured that if there is a constant it's on the RHS. 657 ConstantInt *RangeStart = dyn_cast<ConstantInt>(Cmp0->getOperand(1)); 658 if (!RangeStart) 659 return nullptr; 660 661 ICmpInst::Predicate Pred0 = (Inverted ? Cmp0->getInversePredicate() : 662 Cmp0->getPredicate()); 663 664 // Accept x > -1 or x >= 0 (after potentially inverting the predicate). 665 if (!((Pred0 == ICmpInst::ICMP_SGT && RangeStart->isMinusOne()) || 666 (Pred0 == ICmpInst::ICMP_SGE && RangeStart->isZero()))) 667 return nullptr; 668 669 ICmpInst::Predicate Pred1 = (Inverted ? Cmp1->getInversePredicate() : 670 Cmp1->getPredicate()); 671 672 Value *Input = Cmp0->getOperand(0); 673 Value *RangeEnd; 674 if (Cmp1->getOperand(0) == Input) { 675 // For the upper range compare we have: icmp x, n 676 RangeEnd = Cmp1->getOperand(1); 677 } else if (Cmp1->getOperand(1) == Input) { 678 // For the upper range compare we have: icmp n, x 679 RangeEnd = Cmp1->getOperand(0); 680 Pred1 = ICmpInst::getSwappedPredicate(Pred1); 681 } else { 682 return nullptr; 683 } 684 685 // Check the upper range comparison, e.g. x < n 686 ICmpInst::Predicate NewPred; 687 switch (Pred1) { 688 case ICmpInst::ICMP_SLT: NewPred = ICmpInst::ICMP_ULT; break; 689 case ICmpInst::ICMP_SLE: NewPred = ICmpInst::ICMP_ULE; break; 690 default: return nullptr; 691 } 692 693 // This simplification is only valid if the upper range is not negative. 694 KnownBits Known = computeKnownBits(RangeEnd, /*Depth=*/0, Cmp1); 695 if (!Known.isNonNegative()) 696 return nullptr; 697 698 if (Inverted) 699 NewPred = ICmpInst::getInversePredicate(NewPred); 700 701 return Builder.CreateICmp(NewPred, Input, RangeEnd); 702 } 703 704 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2) 705 // Fold (!iszero(A & K1) & !iszero(A & K2)) -> (A & (K1 | K2)) == (K1 | K2) 706 Value *InstCombinerImpl::foldAndOrOfICmpsOfAndWithPow2(ICmpInst *LHS, 707 ICmpInst *RHS, 708 Instruction *CxtI, 709 bool IsAnd, 710 bool IsLogical) { 711 CmpInst::Predicate Pred = IsAnd ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ; 712 if (LHS->getPredicate() != Pred || RHS->getPredicate() != Pred) 713 return nullptr; 714 715 if (!match(LHS->getOperand(1), m_Zero()) || 716 !match(RHS->getOperand(1), m_Zero())) 717 return nullptr; 718 719 Value *L1, *L2, *R1, *R2; 720 if (match(LHS->getOperand(0), m_And(m_Value(L1), m_Value(L2))) && 721 match(RHS->getOperand(0), m_And(m_Value(R1), m_Value(R2)))) { 722 if (L1 == R2 || L2 == R2) 723 std::swap(R1, R2); 724 if (L2 == R1) 725 std::swap(L1, L2); 726 727 if (L1 == R1 && 728 isKnownToBeAPowerOfTwo(L2, false, 0, CxtI) && 729 isKnownToBeAPowerOfTwo(R2, false, 0, CxtI)) { 730 // If this is a logical and/or, then we must prevent propagation of a 731 // poison value from the RHS by inserting freeze. 732 if (IsLogical) 733 R2 = Builder.CreateFreeze(R2); 734 Value *Mask = Builder.CreateOr(L2, R2); 735 Value *Masked = Builder.CreateAnd(L1, Mask); 736 auto NewPred = IsAnd ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE; 737 return Builder.CreateICmp(NewPred, Masked, Mask); 738 } 739 } 740 741 return nullptr; 742 } 743 744 /// General pattern: 745 /// X & Y 746 /// 747 /// Where Y is checking that all the high bits (covered by a mask 4294967168) 748 /// are uniform, i.e. %arg & 4294967168 can be either 4294967168 or 0 749 /// Pattern can be one of: 750 /// %t = add i32 %arg, 128 751 /// %r = icmp ult i32 %t, 256 752 /// Or 753 /// %t0 = shl i32 %arg, 24 754 /// %t1 = ashr i32 %t0, 24 755 /// %r = icmp eq i32 %t1, %arg 756 /// Or 757 /// %t0 = trunc i32 %arg to i8 758 /// %t1 = sext i8 %t0 to i32 759 /// %r = icmp eq i32 %t1, %arg 760 /// This pattern is a signed truncation check. 761 /// 762 /// And X is checking that some bit in that same mask is zero. 763 /// I.e. can be one of: 764 /// %r = icmp sgt i32 %arg, -1 765 /// Or 766 /// %t = and i32 %arg, 2147483648 767 /// %r = icmp eq i32 %t, 0 768 /// 769 /// Since we are checking that all the bits in that mask are the same, 770 /// and a particular bit is zero, what we are really checking is that all the 771 /// masked bits are zero. 772 /// So this should be transformed to: 773 /// %r = icmp ult i32 %arg, 128 774 static Value *foldSignedTruncationCheck(ICmpInst *ICmp0, ICmpInst *ICmp1, 775 Instruction &CxtI, 776 InstCombiner::BuilderTy &Builder) { 777 assert(CxtI.getOpcode() == Instruction::And); 778 779 // Match icmp ult (add %arg, C01), C1 (C1 == C01 << 1; powers of two) 780 auto tryToMatchSignedTruncationCheck = [](ICmpInst *ICmp, Value *&X, 781 APInt &SignBitMask) -> bool { 782 CmpInst::Predicate Pred; 783 const APInt *I01, *I1; // powers of two; I1 == I01 << 1 784 if (!(match(ICmp, 785 m_ICmp(Pred, m_Add(m_Value(X), m_Power2(I01)), m_Power2(I1))) && 786 Pred == ICmpInst::ICMP_ULT && I1->ugt(*I01) && I01->shl(1) == *I1)) 787 return false; 788 // Which bit is the new sign bit as per the 'signed truncation' pattern? 789 SignBitMask = *I01; 790 return true; 791 }; 792 793 // One icmp needs to be 'signed truncation check'. 794 // We need to match this first, else we will mismatch commutative cases. 795 Value *X1; 796 APInt HighestBit; 797 ICmpInst *OtherICmp; 798 if (tryToMatchSignedTruncationCheck(ICmp1, X1, HighestBit)) 799 OtherICmp = ICmp0; 800 else if (tryToMatchSignedTruncationCheck(ICmp0, X1, HighestBit)) 801 OtherICmp = ICmp1; 802 else 803 return nullptr; 804 805 assert(HighestBit.isPowerOf2() && "expected to be power of two (non-zero)"); 806 807 // Try to match/decompose into: icmp eq (X & Mask), 0 808 auto tryToDecompose = [](ICmpInst *ICmp, Value *&X, 809 APInt &UnsetBitsMask) -> bool { 810 CmpInst::Predicate Pred = ICmp->getPredicate(); 811 // Can it be decomposed into icmp eq (X & Mask), 0 ? 812 if (llvm::decomposeBitTestICmp(ICmp->getOperand(0), ICmp->getOperand(1), 813 Pred, X, UnsetBitsMask, 814 /*LookThroughTrunc=*/false) && 815 Pred == ICmpInst::ICMP_EQ) 816 return true; 817 // Is it icmp eq (X & Mask), 0 already? 818 const APInt *Mask; 819 if (match(ICmp, m_ICmp(Pred, m_And(m_Value(X), m_APInt(Mask)), m_Zero())) && 820 Pred == ICmpInst::ICMP_EQ) { 821 UnsetBitsMask = *Mask; 822 return true; 823 } 824 return false; 825 }; 826 827 // And the other icmp needs to be decomposable into a bit test. 828 Value *X0; 829 APInt UnsetBitsMask; 830 if (!tryToDecompose(OtherICmp, X0, UnsetBitsMask)) 831 return nullptr; 832 833 assert(!UnsetBitsMask.isZero() && "empty mask makes no sense."); 834 835 // Are they working on the same value? 836 Value *X; 837 if (X1 == X0) { 838 // Ok as is. 839 X = X1; 840 } else if (match(X0, m_Trunc(m_Specific(X1)))) { 841 UnsetBitsMask = UnsetBitsMask.zext(X1->getType()->getScalarSizeInBits()); 842 X = X1; 843 } else 844 return nullptr; 845 846 // So which bits should be uniform as per the 'signed truncation check'? 847 // (all the bits starting with (i.e. including) HighestBit) 848 APInt SignBitsMask = ~(HighestBit - 1U); 849 850 // UnsetBitsMask must have some common bits with SignBitsMask, 851 if (!UnsetBitsMask.intersects(SignBitsMask)) 852 return nullptr; 853 854 // Does UnsetBitsMask contain any bits outside of SignBitsMask? 855 if (!UnsetBitsMask.isSubsetOf(SignBitsMask)) { 856 APInt OtherHighestBit = (~UnsetBitsMask) + 1U; 857 if (!OtherHighestBit.isPowerOf2()) 858 return nullptr; 859 HighestBit = APIntOps::umin(HighestBit, OtherHighestBit); 860 } 861 // Else, if it does not, then all is ok as-is. 862 863 // %r = icmp ult %X, SignBit 864 return Builder.CreateICmpULT(X, ConstantInt::get(X->getType(), HighestBit), 865 CxtI.getName() + ".simplified"); 866 } 867 868 /// Fold (icmp eq ctpop(X) 1) | (icmp eq X 0) into (icmp ult ctpop(X) 2) and 869 /// fold (icmp ne ctpop(X) 1) & (icmp ne X 0) into (icmp ugt ctpop(X) 1). 870 /// Also used for logical and/or, must be poison safe. 871 static Value *foldIsPowerOf2OrZero(ICmpInst *Cmp0, ICmpInst *Cmp1, bool IsAnd, 872 InstCombiner::BuilderTy &Builder) { 873 CmpInst::Predicate Pred0, Pred1; 874 Value *X; 875 if (!match(Cmp0, m_ICmp(Pred0, m_Intrinsic<Intrinsic::ctpop>(m_Value(X)), 876 m_SpecificInt(1))) || 877 !match(Cmp1, m_ICmp(Pred1, m_Specific(X), m_ZeroInt()))) 878 return nullptr; 879 880 Value *CtPop = Cmp0->getOperand(0); 881 if (IsAnd && Pred0 == ICmpInst::ICMP_NE && Pred1 == ICmpInst::ICMP_NE) 882 return Builder.CreateICmpUGT(CtPop, ConstantInt::get(CtPop->getType(), 1)); 883 if (!IsAnd && Pred0 == ICmpInst::ICMP_EQ && Pred1 == ICmpInst::ICMP_EQ) 884 return Builder.CreateICmpULT(CtPop, ConstantInt::get(CtPop->getType(), 2)); 885 886 return nullptr; 887 } 888 889 /// Reduce a pair of compares that check if a value has exactly 1 bit set. 890 /// Also used for logical and/or, must be poison safe. 891 static Value *foldIsPowerOf2(ICmpInst *Cmp0, ICmpInst *Cmp1, bool JoinedByAnd, 892 InstCombiner::BuilderTy &Builder) { 893 // Handle 'and' / 'or' commutation: make the equality check the first operand. 894 if (JoinedByAnd && Cmp1->getPredicate() == ICmpInst::ICMP_NE) 895 std::swap(Cmp0, Cmp1); 896 else if (!JoinedByAnd && Cmp1->getPredicate() == ICmpInst::ICMP_EQ) 897 std::swap(Cmp0, Cmp1); 898 899 // (X != 0) && (ctpop(X) u< 2) --> ctpop(X) == 1 900 CmpInst::Predicate Pred0, Pred1; 901 Value *X; 902 if (JoinedByAnd && match(Cmp0, m_ICmp(Pred0, m_Value(X), m_ZeroInt())) && 903 match(Cmp1, m_ICmp(Pred1, m_Intrinsic<Intrinsic::ctpop>(m_Specific(X)), 904 m_SpecificInt(2))) && 905 Pred0 == ICmpInst::ICMP_NE && Pred1 == ICmpInst::ICMP_ULT) { 906 Value *CtPop = Cmp1->getOperand(0); 907 return Builder.CreateICmpEQ(CtPop, ConstantInt::get(CtPop->getType(), 1)); 908 } 909 // (X == 0) || (ctpop(X) u> 1) --> ctpop(X) != 1 910 if (!JoinedByAnd && match(Cmp0, m_ICmp(Pred0, m_Value(X), m_ZeroInt())) && 911 match(Cmp1, m_ICmp(Pred1, m_Intrinsic<Intrinsic::ctpop>(m_Specific(X)), 912 m_SpecificInt(1))) && 913 Pred0 == ICmpInst::ICMP_EQ && Pred1 == ICmpInst::ICMP_UGT) { 914 Value *CtPop = Cmp1->getOperand(0); 915 return Builder.CreateICmpNE(CtPop, ConstantInt::get(CtPop->getType(), 1)); 916 } 917 return nullptr; 918 } 919 920 /// Try to fold (icmp(A & B) == 0) & (icmp(A & D) != E) into (icmp A u< D) iff 921 /// B is a contiguous set of ones starting from the most significant bit 922 /// (negative power of 2), D and E are equal, and D is a contiguous set of ones 923 /// starting at the most significant zero bit in B. Parameter B supports masking 924 /// using undef/poison in either scalar or vector values. 925 static Value *foldNegativePower2AndShiftedMask( 926 Value *A, Value *B, Value *D, Value *E, ICmpInst::Predicate PredL, 927 ICmpInst::Predicate PredR, InstCombiner::BuilderTy &Builder) { 928 assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) && 929 "Expected equality predicates for masked type of icmps."); 930 if (PredL != ICmpInst::ICMP_EQ || PredR != ICmpInst::ICMP_NE) 931 return nullptr; 932 933 if (!match(B, m_NegatedPower2()) || !match(D, m_ShiftedMask()) || 934 !match(E, m_ShiftedMask())) 935 return nullptr; 936 937 // Test scalar arguments for conversion. B has been validated earlier to be a 938 // negative power of two and thus is guaranteed to have one or more contiguous 939 // ones starting from the MSB followed by zero or more contiguous zeros. D has 940 // been validated earlier to be a shifted set of one or more contiguous ones. 941 // In order to match, B leading ones and D leading zeros should be equal. The 942 // predicate that B be a negative power of 2 prevents the condition of there 943 // ever being zero leading ones. Thus 0 == 0 cannot occur. The predicate that 944 // D always be a shifted mask prevents the condition of D equaling 0. This 945 // prevents matching the condition where B contains the maximum number of 946 // leading one bits (-1) and D contains the maximum number of leading zero 947 // bits (0). 948 auto isReducible = [](const Value *B, const Value *D, const Value *E) { 949 const APInt *BCst, *DCst, *ECst; 950 return match(B, m_APIntAllowUndef(BCst)) && match(D, m_APInt(DCst)) && 951 match(E, m_APInt(ECst)) && *DCst == *ECst && 952 (isa<UndefValue>(B) || 953 (BCst->countLeadingOnes() == DCst->countLeadingZeros())); 954 }; 955 956 // Test vector type arguments for conversion. 957 if (const auto *BVTy = dyn_cast<VectorType>(B->getType())) { 958 const auto *BFVTy = dyn_cast<FixedVectorType>(BVTy); 959 const auto *BConst = dyn_cast<Constant>(B); 960 const auto *DConst = dyn_cast<Constant>(D); 961 const auto *EConst = dyn_cast<Constant>(E); 962 963 if (!BFVTy || !BConst || !DConst || !EConst) 964 return nullptr; 965 966 for (unsigned I = 0; I != BFVTy->getNumElements(); ++I) { 967 const auto *BElt = BConst->getAggregateElement(I); 968 const auto *DElt = DConst->getAggregateElement(I); 969 const auto *EElt = EConst->getAggregateElement(I); 970 971 if (!BElt || !DElt || !EElt) 972 return nullptr; 973 if (!isReducible(BElt, DElt, EElt)) 974 return nullptr; 975 } 976 } else { 977 // Test scalar type arguments for conversion. 978 if (!isReducible(B, D, E)) 979 return nullptr; 980 } 981 return Builder.CreateICmp(ICmpInst::ICMP_ULT, A, D); 982 } 983 984 /// Try to fold ((icmp X u< P) & (icmp(X & M) != M)) or ((icmp X s> -1) & 985 /// (icmp(X & M) != M)) into (icmp X u< M). Where P is a power of 2, M < P, and 986 /// M is a contiguous shifted mask starting at the right most significant zero 987 /// bit in P. SGT is supported as when P is the largest representable power of 988 /// 2, an earlier optimization converts the expression into (icmp X s> -1). 989 /// Parameter P supports masking using undef/poison in either scalar or vector 990 /// values. 991 static Value *foldPowerOf2AndShiftedMask(ICmpInst *Cmp0, ICmpInst *Cmp1, 992 bool JoinedByAnd, 993 InstCombiner::BuilderTy &Builder) { 994 if (!JoinedByAnd) 995 return nullptr; 996 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr; 997 ICmpInst::Predicate CmpPred0 = Cmp0->getPredicate(), 998 CmpPred1 = Cmp1->getPredicate(); 999 // Assuming P is a 2^n, getMaskedTypeForICmpPair will normalize (icmp X u< 1000 // 2^n) into (icmp (X & ~(2^n-1)) == 0) and (icmp X s> -1) into (icmp (X & 1001 // SignMask) == 0). 1002 std::optional<std::pair<unsigned, unsigned>> MaskPair = 1003 getMaskedTypeForICmpPair(A, B, C, D, E, Cmp0, Cmp1, CmpPred0, CmpPred1); 1004 if (!MaskPair) 1005 return nullptr; 1006 1007 const auto compareBMask = BMask_NotMixed | BMask_NotAllOnes; 1008 unsigned CmpMask0 = MaskPair->first; 1009 unsigned CmpMask1 = MaskPair->second; 1010 if ((CmpMask0 & Mask_AllZeros) && (CmpMask1 == compareBMask)) { 1011 if (Value *V = foldNegativePower2AndShiftedMask(A, B, D, E, CmpPred0, 1012 CmpPred1, Builder)) 1013 return V; 1014 } else if ((CmpMask0 == compareBMask) && (CmpMask1 & Mask_AllZeros)) { 1015 if (Value *V = foldNegativePower2AndShiftedMask(A, D, B, C, CmpPred1, 1016 CmpPred0, Builder)) 1017 return V; 1018 } 1019 return nullptr; 1020 } 1021 1022 /// Commuted variants are assumed to be handled by calling this function again 1023 /// with the parameters swapped. 1024 static Value *foldUnsignedUnderflowCheck(ICmpInst *ZeroICmp, 1025 ICmpInst *UnsignedICmp, bool IsAnd, 1026 const SimplifyQuery &Q, 1027 InstCombiner::BuilderTy &Builder) { 1028 Value *ZeroCmpOp; 1029 ICmpInst::Predicate EqPred; 1030 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(ZeroCmpOp), m_Zero())) || 1031 !ICmpInst::isEquality(EqPred)) 1032 return nullptr; 1033 1034 auto IsKnownNonZero = [&](Value *V) { 1035 return isKnownNonZero(V, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT); 1036 }; 1037 1038 ICmpInst::Predicate UnsignedPred; 1039 1040 Value *A, *B; 1041 if (match(UnsignedICmp, 1042 m_c_ICmp(UnsignedPred, m_Specific(ZeroCmpOp), m_Value(A))) && 1043 match(ZeroCmpOp, m_c_Add(m_Specific(A), m_Value(B))) && 1044 (ZeroICmp->hasOneUse() || UnsignedICmp->hasOneUse())) { 1045 auto GetKnownNonZeroAndOther = [&](Value *&NonZero, Value *&Other) { 1046 if (!IsKnownNonZero(NonZero)) 1047 std::swap(NonZero, Other); 1048 return IsKnownNonZero(NonZero); 1049 }; 1050 1051 // Given ZeroCmpOp = (A + B) 1052 // ZeroCmpOp < A && ZeroCmpOp != 0 --> (0-X) < Y iff 1053 // ZeroCmpOp >= A || ZeroCmpOp == 0 --> (0-X) >= Y iff 1054 // with X being the value (A/B) that is known to be non-zero, 1055 // and Y being remaining value. 1056 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE && 1057 IsAnd && GetKnownNonZeroAndOther(B, A)) 1058 return Builder.CreateICmpULT(Builder.CreateNeg(B), A); 1059 if (UnsignedPred == ICmpInst::ICMP_UGE && EqPred == ICmpInst::ICMP_EQ && 1060 !IsAnd && GetKnownNonZeroAndOther(B, A)) 1061 return Builder.CreateICmpUGE(Builder.CreateNeg(B), A); 1062 } 1063 1064 return nullptr; 1065 } 1066 1067 struct IntPart { 1068 Value *From; 1069 unsigned StartBit; 1070 unsigned NumBits; 1071 }; 1072 1073 /// Match an extraction of bits from an integer. 1074 static std::optional<IntPart> matchIntPart(Value *V) { 1075 Value *X; 1076 if (!match(V, m_OneUse(m_Trunc(m_Value(X))))) 1077 return std::nullopt; 1078 1079 unsigned NumOriginalBits = X->getType()->getScalarSizeInBits(); 1080 unsigned NumExtractedBits = V->getType()->getScalarSizeInBits(); 1081 Value *Y; 1082 const APInt *Shift; 1083 // For a trunc(lshr Y, Shift) pattern, make sure we're only extracting bits 1084 // from Y, not any shifted-in zeroes. 1085 if (match(X, m_OneUse(m_LShr(m_Value(Y), m_APInt(Shift)))) && 1086 Shift->ule(NumOriginalBits - NumExtractedBits)) 1087 return {{Y, (unsigned)Shift->getZExtValue(), NumExtractedBits}}; 1088 return {{X, 0, NumExtractedBits}}; 1089 } 1090 1091 /// Materialize an extraction of bits from an integer in IR. 1092 static Value *extractIntPart(const IntPart &P, IRBuilderBase &Builder) { 1093 Value *V = P.From; 1094 if (P.StartBit) 1095 V = Builder.CreateLShr(V, P.StartBit); 1096 Type *TruncTy = V->getType()->getWithNewBitWidth(P.NumBits); 1097 if (TruncTy != V->getType()) 1098 V = Builder.CreateTrunc(V, TruncTy); 1099 return V; 1100 } 1101 1102 /// (icmp eq X0, Y0) & (icmp eq X1, Y1) -> icmp eq X01, Y01 1103 /// (icmp ne X0, Y0) | (icmp ne X1, Y1) -> icmp ne X01, Y01 1104 /// where X0, X1 and Y0, Y1 are adjacent parts extracted from an integer. 1105 Value *InstCombinerImpl::foldEqOfParts(ICmpInst *Cmp0, ICmpInst *Cmp1, 1106 bool IsAnd) { 1107 if (!Cmp0->hasOneUse() || !Cmp1->hasOneUse()) 1108 return nullptr; 1109 1110 CmpInst::Predicate Pred = IsAnd ? CmpInst::ICMP_EQ : CmpInst::ICMP_NE; 1111 auto GetMatchPart = [&](ICmpInst *Cmp, 1112 unsigned OpNo) -> std::optional<IntPart> { 1113 if (Pred == Cmp->getPredicate()) 1114 return matchIntPart(Cmp->getOperand(OpNo)); 1115 1116 const APInt *C; 1117 // (icmp eq (lshr x, C), (lshr y, C)) gets optimized to: 1118 // (icmp ult (xor x, y), 1 << C) so also look for that. 1119 if (Pred == CmpInst::ICMP_EQ && Cmp->getPredicate() == CmpInst::ICMP_ULT) { 1120 if (!match(Cmp->getOperand(1), m_Power2(C)) || 1121 !match(Cmp->getOperand(0), m_Xor(m_Value(), m_Value()))) 1122 return std::nullopt; 1123 } 1124 1125 // (icmp ne (lshr x, C), (lshr y, C)) gets optimized to: 1126 // (icmp ugt (xor x, y), (1 << C) - 1) so also look for that. 1127 else if (Pred == CmpInst::ICMP_NE && 1128 Cmp->getPredicate() == CmpInst::ICMP_UGT) { 1129 if (!match(Cmp->getOperand(1), m_LowBitMask(C)) || 1130 !match(Cmp->getOperand(0), m_Xor(m_Value(), m_Value()))) 1131 return std::nullopt; 1132 } else { 1133 return std::nullopt; 1134 } 1135 1136 unsigned From = Pred == CmpInst::ICMP_NE ? C->popcount() : C->countr_zero(); 1137 Instruction *I = cast<Instruction>(Cmp->getOperand(0)); 1138 return {{I->getOperand(OpNo), From, C->getBitWidth() - From}}; 1139 }; 1140 1141 std::optional<IntPart> L0 = GetMatchPart(Cmp0, 0); 1142 std::optional<IntPart> R0 = GetMatchPart(Cmp0, 1); 1143 std::optional<IntPart> L1 = GetMatchPart(Cmp1, 0); 1144 std::optional<IntPart> R1 = GetMatchPart(Cmp1, 1); 1145 if (!L0 || !R0 || !L1 || !R1) 1146 return nullptr; 1147 1148 // Make sure the LHS/RHS compare a part of the same value, possibly after 1149 // an operand swap. 1150 if (L0->From != L1->From || R0->From != R1->From) { 1151 if (L0->From != R1->From || R0->From != L1->From) 1152 return nullptr; 1153 std::swap(L1, R1); 1154 } 1155 1156 // Make sure the extracted parts are adjacent, canonicalizing to L0/R0 being 1157 // the low part and L1/R1 being the high part. 1158 if (L0->StartBit + L0->NumBits != L1->StartBit || 1159 R0->StartBit + R0->NumBits != R1->StartBit) { 1160 if (L1->StartBit + L1->NumBits != L0->StartBit || 1161 R1->StartBit + R1->NumBits != R0->StartBit) 1162 return nullptr; 1163 std::swap(L0, L1); 1164 std::swap(R0, R1); 1165 } 1166 1167 // We can simplify to a comparison of these larger parts of the integers. 1168 IntPart L = {L0->From, L0->StartBit, L0->NumBits + L1->NumBits}; 1169 IntPart R = {R0->From, R0->StartBit, R0->NumBits + R1->NumBits}; 1170 Value *LValue = extractIntPart(L, Builder); 1171 Value *RValue = extractIntPart(R, Builder); 1172 return Builder.CreateICmp(Pred, LValue, RValue); 1173 } 1174 1175 /// Reduce logic-of-compares with equality to a constant by substituting a 1176 /// common operand with the constant. Callers are expected to call this with 1177 /// Cmp0/Cmp1 switched to handle logic op commutativity. 1178 static Value *foldAndOrOfICmpsWithConstEq(ICmpInst *Cmp0, ICmpInst *Cmp1, 1179 bool IsAnd, bool IsLogical, 1180 InstCombiner::BuilderTy &Builder, 1181 const SimplifyQuery &Q) { 1182 // Match an equality compare with a non-poison constant as Cmp0. 1183 // Also, give up if the compare can be constant-folded to avoid looping. 1184 ICmpInst::Predicate Pred0; 1185 Value *X; 1186 Constant *C; 1187 if (!match(Cmp0, m_ICmp(Pred0, m_Value(X), m_Constant(C))) || 1188 !isGuaranteedNotToBeUndefOrPoison(C) || isa<Constant>(X)) 1189 return nullptr; 1190 if ((IsAnd && Pred0 != ICmpInst::ICMP_EQ) || 1191 (!IsAnd && Pred0 != ICmpInst::ICMP_NE)) 1192 return nullptr; 1193 1194 // The other compare must include a common operand (X). Canonicalize the 1195 // common operand as operand 1 (Pred1 is swapped if the common operand was 1196 // operand 0). 1197 Value *Y; 1198 ICmpInst::Predicate Pred1; 1199 if (!match(Cmp1, m_c_ICmp(Pred1, m_Value(Y), m_Deferred(X)))) 1200 return nullptr; 1201 1202 // Replace variable with constant value equivalence to remove a variable use: 1203 // (X == C) && (Y Pred1 X) --> (X == C) && (Y Pred1 C) 1204 // (X != C) || (Y Pred1 X) --> (X != C) || (Y Pred1 C) 1205 // Can think of the 'or' substitution with the 'and' bool equivalent: 1206 // A || B --> A || (!A && B) 1207 Value *SubstituteCmp = simplifyICmpInst(Pred1, Y, C, Q); 1208 if (!SubstituteCmp) { 1209 // If we need to create a new instruction, require that the old compare can 1210 // be removed. 1211 if (!Cmp1->hasOneUse()) 1212 return nullptr; 1213 SubstituteCmp = Builder.CreateICmp(Pred1, Y, C); 1214 } 1215 if (IsLogical) 1216 return IsAnd ? Builder.CreateLogicalAnd(Cmp0, SubstituteCmp) 1217 : Builder.CreateLogicalOr(Cmp0, SubstituteCmp); 1218 return Builder.CreateBinOp(IsAnd ? Instruction::And : Instruction::Or, Cmp0, 1219 SubstituteCmp); 1220 } 1221 1222 /// Fold (icmp Pred1 V1, C1) & (icmp Pred2 V2, C2) 1223 /// or (icmp Pred1 V1, C1) | (icmp Pred2 V2, C2) 1224 /// into a single comparison using range-based reasoning. 1225 /// NOTE: This is also used for logical and/or, must be poison-safe! 1226 Value *InstCombinerImpl::foldAndOrOfICmpsUsingRanges(ICmpInst *ICmp1, 1227 ICmpInst *ICmp2, 1228 bool IsAnd) { 1229 ICmpInst::Predicate Pred1, Pred2; 1230 Value *V1, *V2; 1231 const APInt *C1, *C2; 1232 if (!match(ICmp1, m_ICmp(Pred1, m_Value(V1), m_APInt(C1))) || 1233 !match(ICmp2, m_ICmp(Pred2, m_Value(V2), m_APInt(C2)))) 1234 return nullptr; 1235 1236 // Look through add of a constant offset on V1, V2, or both operands. This 1237 // allows us to interpret the V + C' < C'' range idiom into a proper range. 1238 const APInt *Offset1 = nullptr, *Offset2 = nullptr; 1239 if (V1 != V2) { 1240 Value *X; 1241 if (match(V1, m_Add(m_Value(X), m_APInt(Offset1)))) 1242 V1 = X; 1243 if (match(V2, m_Add(m_Value(X), m_APInt(Offset2)))) 1244 V2 = X; 1245 } 1246 1247 if (V1 != V2) 1248 return nullptr; 1249 1250 ConstantRange CR1 = ConstantRange::makeExactICmpRegion( 1251 IsAnd ? ICmpInst::getInversePredicate(Pred1) : Pred1, *C1); 1252 if (Offset1) 1253 CR1 = CR1.subtract(*Offset1); 1254 1255 ConstantRange CR2 = ConstantRange::makeExactICmpRegion( 1256 IsAnd ? ICmpInst::getInversePredicate(Pred2) : Pred2, *C2); 1257 if (Offset2) 1258 CR2 = CR2.subtract(*Offset2); 1259 1260 Type *Ty = V1->getType(); 1261 Value *NewV = V1; 1262 std::optional<ConstantRange> CR = CR1.exactUnionWith(CR2); 1263 if (!CR) { 1264 if (!(ICmp1->hasOneUse() && ICmp2->hasOneUse()) || CR1.isWrappedSet() || 1265 CR2.isWrappedSet()) 1266 return nullptr; 1267 1268 // Check whether we have equal-size ranges that only differ by one bit. 1269 // In that case we can apply a mask to map one range onto the other. 1270 APInt LowerDiff = CR1.getLower() ^ CR2.getLower(); 1271 APInt UpperDiff = (CR1.getUpper() - 1) ^ (CR2.getUpper() - 1); 1272 APInt CR1Size = CR1.getUpper() - CR1.getLower(); 1273 if (!LowerDiff.isPowerOf2() || LowerDiff != UpperDiff || 1274 CR1Size != CR2.getUpper() - CR2.getLower()) 1275 return nullptr; 1276 1277 CR = CR1.getLower().ult(CR2.getLower()) ? CR1 : CR2; 1278 NewV = Builder.CreateAnd(NewV, ConstantInt::get(Ty, ~LowerDiff)); 1279 } 1280 1281 if (IsAnd) 1282 CR = CR->inverse(); 1283 1284 CmpInst::Predicate NewPred; 1285 APInt NewC, Offset; 1286 CR->getEquivalentICmp(NewPred, NewC, Offset); 1287 1288 if (Offset != 0) 1289 NewV = Builder.CreateAdd(NewV, ConstantInt::get(Ty, Offset)); 1290 return Builder.CreateICmp(NewPred, NewV, ConstantInt::get(Ty, NewC)); 1291 } 1292 1293 /// Ignore all operations which only change the sign of a value, returning the 1294 /// underlying magnitude value. 1295 static Value *stripSignOnlyFPOps(Value *Val) { 1296 match(Val, m_FNeg(m_Value(Val))); 1297 match(Val, m_FAbs(m_Value(Val))); 1298 match(Val, m_CopySign(m_Value(Val), m_Value())); 1299 return Val; 1300 } 1301 1302 /// Matches canonical form of isnan, fcmp ord x, 0 1303 static bool matchIsNotNaN(FCmpInst::Predicate P, Value *LHS, Value *RHS) { 1304 return P == FCmpInst::FCMP_ORD && match(RHS, m_AnyZeroFP()); 1305 } 1306 1307 /// Matches fcmp u__ x, +/-inf 1308 static bool matchUnorderedInfCompare(FCmpInst::Predicate P, Value *LHS, 1309 Value *RHS) { 1310 return FCmpInst::isUnordered(P) && match(RHS, m_Inf()); 1311 } 1312 1313 /// and (fcmp ord x, 0), (fcmp u* x, inf) -> fcmp o* x, inf 1314 /// 1315 /// Clang emits this pattern for doing an isfinite check in __builtin_isnormal. 1316 static Value *matchIsFiniteTest(InstCombiner::BuilderTy &Builder, FCmpInst *LHS, 1317 FCmpInst *RHS) { 1318 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1); 1319 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1); 1320 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate(); 1321 1322 if (!matchIsNotNaN(PredL, LHS0, LHS1) || 1323 !matchUnorderedInfCompare(PredR, RHS0, RHS1)) 1324 return nullptr; 1325 1326 IRBuilder<>::FastMathFlagGuard FMFG(Builder); 1327 FastMathFlags FMF = LHS->getFastMathFlags(); 1328 FMF &= RHS->getFastMathFlags(); 1329 Builder.setFastMathFlags(FMF); 1330 1331 return Builder.CreateFCmp(FCmpInst::getOrderedPredicate(PredR), RHS0, RHS1); 1332 } 1333 1334 Value *InstCombinerImpl::foldLogicOfFCmps(FCmpInst *LHS, FCmpInst *RHS, 1335 bool IsAnd, bool IsLogicalSelect) { 1336 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1); 1337 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1); 1338 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate(); 1339 1340 if (LHS0 == RHS1 && RHS0 == LHS1) { 1341 // Swap RHS operands to match LHS. 1342 PredR = FCmpInst::getSwappedPredicate(PredR); 1343 std::swap(RHS0, RHS1); 1344 } 1345 1346 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y). 1347 // Suppose the relation between x and y is R, where R is one of 1348 // U(1000), L(0100), G(0010) or E(0001), and CC0 and CC1 are the bitmasks for 1349 // testing the desired relations. 1350 // 1351 // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this: 1352 // bool(R & CC0) && bool(R & CC1) 1353 // = bool((R & CC0) & (R & CC1)) 1354 // = bool(R & (CC0 & CC1)) <= by re-association, commutation, and idempotency 1355 // 1356 // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this: 1357 // bool(R & CC0) || bool(R & CC1) 1358 // = bool((R & CC0) | (R & CC1)) 1359 // = bool(R & (CC0 | CC1)) <= by reversed distribution (contribution? ;) 1360 if (LHS0 == RHS0 && LHS1 == RHS1) { 1361 unsigned FCmpCodeL = getFCmpCode(PredL); 1362 unsigned FCmpCodeR = getFCmpCode(PredR); 1363 unsigned NewPred = IsAnd ? FCmpCodeL & FCmpCodeR : FCmpCodeL | FCmpCodeR; 1364 1365 // Intersect the fast math flags. 1366 // TODO: We can union the fast math flags unless this is a logical select. 1367 IRBuilder<>::FastMathFlagGuard FMFG(Builder); 1368 FastMathFlags FMF = LHS->getFastMathFlags(); 1369 FMF &= RHS->getFastMathFlags(); 1370 Builder.setFastMathFlags(FMF); 1371 1372 return getFCmpValue(NewPred, LHS0, LHS1, Builder); 1373 } 1374 1375 // This transform is not valid for a logical select. 1376 if (!IsLogicalSelect && 1377 ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) || 1378 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && 1379 !IsAnd))) { 1380 if (LHS0->getType() != RHS0->getType()) 1381 return nullptr; 1382 1383 // FCmp canonicalization ensures that (fcmp ord/uno X, X) and 1384 // (fcmp ord/uno X, C) will be transformed to (fcmp X, +0.0). 1385 if (match(LHS1, m_PosZeroFP()) && match(RHS1, m_PosZeroFP())) 1386 // Ignore the constants because they are obviously not NANs: 1387 // (fcmp ord x, 0.0) & (fcmp ord y, 0.0) -> (fcmp ord x, y) 1388 // (fcmp uno x, 0.0) | (fcmp uno y, 0.0) -> (fcmp uno x, y) 1389 return Builder.CreateFCmp(PredL, LHS0, RHS0); 1390 } 1391 1392 if (IsAnd && stripSignOnlyFPOps(LHS0) == stripSignOnlyFPOps(RHS0)) { 1393 // and (fcmp ord x, 0), (fcmp u* x, inf) -> fcmp o* x, inf 1394 // and (fcmp ord x, 0), (fcmp u* fabs(x), inf) -> fcmp o* x, inf 1395 if (Value *Left = matchIsFiniteTest(Builder, LHS, RHS)) 1396 return Left; 1397 if (Value *Right = matchIsFiniteTest(Builder, RHS, LHS)) 1398 return Right; 1399 } 1400 1401 // Turn at least two fcmps with constants into llvm.is.fpclass. 1402 // 1403 // If we can represent a combined value test with one class call, we can 1404 // potentially eliminate 4-6 instructions. If we can represent a test with a 1405 // single fcmp with fneg and fabs, that's likely a better canonical form. 1406 if (LHS->hasOneUse() && RHS->hasOneUse()) { 1407 auto [ClassValRHS, ClassMaskRHS] = 1408 fcmpToClassTest(PredR, *RHS->getFunction(), RHS0, RHS1); 1409 if (ClassValRHS) { 1410 auto [ClassValLHS, ClassMaskLHS] = 1411 fcmpToClassTest(PredL, *LHS->getFunction(), LHS0, LHS1); 1412 if (ClassValLHS == ClassValRHS) { 1413 unsigned CombinedMask = IsAnd ? (ClassMaskLHS & ClassMaskRHS) 1414 : (ClassMaskLHS | ClassMaskRHS); 1415 return Builder.CreateIntrinsic( 1416 Intrinsic::is_fpclass, {ClassValLHS->getType()}, 1417 {ClassValLHS, Builder.getInt32(CombinedMask)}); 1418 } 1419 } 1420 } 1421 1422 return nullptr; 1423 } 1424 1425 /// Match an fcmp against a special value that performs a test possible by 1426 /// llvm.is.fpclass. 1427 static bool matchIsFPClassLikeFCmp(Value *Op, Value *&ClassVal, 1428 uint64_t &ClassMask) { 1429 auto *FCmp = dyn_cast<FCmpInst>(Op); 1430 if (!FCmp || !FCmp->hasOneUse()) 1431 return false; 1432 1433 std::tie(ClassVal, ClassMask) = 1434 fcmpToClassTest(FCmp->getPredicate(), *FCmp->getParent()->getParent(), 1435 FCmp->getOperand(0), FCmp->getOperand(1)); 1436 return ClassVal != nullptr; 1437 } 1438 1439 /// or (is_fpclass x, mask0), (is_fpclass x, mask1) 1440 /// -> is_fpclass x, (mask0 | mask1) 1441 /// and (is_fpclass x, mask0), (is_fpclass x, mask1) 1442 /// -> is_fpclass x, (mask0 & mask1) 1443 /// xor (is_fpclass x, mask0), (is_fpclass x, mask1) 1444 /// -> is_fpclass x, (mask0 ^ mask1) 1445 Instruction *InstCombinerImpl::foldLogicOfIsFPClass(BinaryOperator &BO, 1446 Value *Op0, Value *Op1) { 1447 Value *ClassVal0 = nullptr; 1448 Value *ClassVal1 = nullptr; 1449 uint64_t ClassMask0, ClassMask1; 1450 1451 // Restrict to folding one fcmp into one is.fpclass for now, don't introduce a 1452 // new class. 1453 // 1454 // TODO: Support forming is.fpclass out of 2 separate fcmps when codegen is 1455 // better. 1456 1457 bool IsLHSClass = 1458 match(Op0, m_OneUse(m_Intrinsic<Intrinsic::is_fpclass>( 1459 m_Value(ClassVal0), m_ConstantInt(ClassMask0)))); 1460 bool IsRHSClass = 1461 match(Op1, m_OneUse(m_Intrinsic<Intrinsic::is_fpclass>( 1462 m_Value(ClassVal1), m_ConstantInt(ClassMask1)))); 1463 if ((((IsLHSClass || matchIsFPClassLikeFCmp(Op0, ClassVal0, ClassMask0)) && 1464 (IsRHSClass || matchIsFPClassLikeFCmp(Op1, ClassVal1, ClassMask1)))) && 1465 ClassVal0 == ClassVal1) { 1466 unsigned NewClassMask; 1467 switch (BO.getOpcode()) { 1468 case Instruction::And: 1469 NewClassMask = ClassMask0 & ClassMask1; 1470 break; 1471 case Instruction::Or: 1472 NewClassMask = ClassMask0 | ClassMask1; 1473 break; 1474 case Instruction::Xor: 1475 NewClassMask = ClassMask0 ^ ClassMask1; 1476 break; 1477 default: 1478 llvm_unreachable("not a binary logic operator"); 1479 } 1480 1481 if (IsLHSClass) { 1482 auto *II = cast<IntrinsicInst>(Op0); 1483 II->setArgOperand( 1484 1, ConstantInt::get(II->getArgOperand(1)->getType(), NewClassMask)); 1485 return replaceInstUsesWith(BO, II); 1486 } 1487 1488 if (IsRHSClass) { 1489 auto *II = cast<IntrinsicInst>(Op1); 1490 II->setArgOperand( 1491 1, ConstantInt::get(II->getArgOperand(1)->getType(), NewClassMask)); 1492 return replaceInstUsesWith(BO, II); 1493 } 1494 1495 CallInst *NewClass = 1496 Builder.CreateIntrinsic(Intrinsic::is_fpclass, {ClassVal0->getType()}, 1497 {ClassVal0, Builder.getInt32(NewClassMask)}); 1498 return replaceInstUsesWith(BO, NewClass); 1499 } 1500 1501 return nullptr; 1502 } 1503 1504 /// Look for the pattern that conditionally negates a value via math operations: 1505 /// cond.splat = sext i1 cond 1506 /// sub = add cond.splat, x 1507 /// xor = xor sub, cond.splat 1508 /// and rewrite it to do the same, but via logical operations: 1509 /// value.neg = sub 0, value 1510 /// cond = select i1 neg, value.neg, value 1511 Instruction *InstCombinerImpl::canonicalizeConditionalNegationViaMathToSelect( 1512 BinaryOperator &I) { 1513 assert(I.getOpcode() == BinaryOperator::Xor && "Only for xor!"); 1514 Value *Cond, *X; 1515 // As per complexity ordering, `xor` is not commutative here. 1516 if (!match(&I, m_c_BinOp(m_OneUse(m_Value()), m_Value())) || 1517 !match(I.getOperand(1), m_SExt(m_Value(Cond))) || 1518 !Cond->getType()->isIntOrIntVectorTy(1) || 1519 !match(I.getOperand(0), m_c_Add(m_SExt(m_Deferred(Cond)), m_Value(X)))) 1520 return nullptr; 1521 return SelectInst::Create(Cond, Builder.CreateNeg(X, X->getName() + ".neg"), 1522 X); 1523 } 1524 1525 /// This a limited reassociation for a special case (see above) where we are 1526 /// checking if two values are either both NAN (unordered) or not-NAN (ordered). 1527 /// This could be handled more generally in '-reassociation', but it seems like 1528 /// an unlikely pattern for a large number of logic ops and fcmps. 1529 static Instruction *reassociateFCmps(BinaryOperator &BO, 1530 InstCombiner::BuilderTy &Builder) { 1531 Instruction::BinaryOps Opcode = BO.getOpcode(); 1532 assert((Opcode == Instruction::And || Opcode == Instruction::Or) && 1533 "Expecting and/or op for fcmp transform"); 1534 1535 // There are 4 commuted variants of the pattern. Canonicalize operands of this 1536 // logic op so an fcmp is operand 0 and a matching logic op is operand 1. 1537 Value *Op0 = BO.getOperand(0), *Op1 = BO.getOperand(1), *X; 1538 FCmpInst::Predicate Pred; 1539 if (match(Op1, m_FCmp(Pred, m_Value(), m_AnyZeroFP()))) 1540 std::swap(Op0, Op1); 1541 1542 // Match inner binop and the predicate for combining 2 NAN checks into 1. 1543 Value *BO10, *BO11; 1544 FCmpInst::Predicate NanPred = Opcode == Instruction::And ? FCmpInst::FCMP_ORD 1545 : FCmpInst::FCMP_UNO; 1546 if (!match(Op0, m_FCmp(Pred, m_Value(X), m_AnyZeroFP())) || Pred != NanPred || 1547 !match(Op1, m_BinOp(Opcode, m_Value(BO10), m_Value(BO11)))) 1548 return nullptr; 1549 1550 // The inner logic op must have a matching fcmp operand. 1551 Value *Y; 1552 if (!match(BO10, m_FCmp(Pred, m_Value(Y), m_AnyZeroFP())) || 1553 Pred != NanPred || X->getType() != Y->getType()) 1554 std::swap(BO10, BO11); 1555 1556 if (!match(BO10, m_FCmp(Pred, m_Value(Y), m_AnyZeroFP())) || 1557 Pred != NanPred || X->getType() != Y->getType()) 1558 return nullptr; 1559 1560 // and (fcmp ord X, 0), (and (fcmp ord Y, 0), Z) --> and (fcmp ord X, Y), Z 1561 // or (fcmp uno X, 0), (or (fcmp uno Y, 0), Z) --> or (fcmp uno X, Y), Z 1562 Value *NewFCmp = Builder.CreateFCmp(Pred, X, Y); 1563 if (auto *NewFCmpInst = dyn_cast<FCmpInst>(NewFCmp)) { 1564 // Intersect FMF from the 2 source fcmps. 1565 NewFCmpInst->copyIRFlags(Op0); 1566 NewFCmpInst->andIRFlags(BO10); 1567 } 1568 return BinaryOperator::Create(Opcode, NewFCmp, BO11); 1569 } 1570 1571 /// Match variations of De Morgan's Laws: 1572 /// (~A & ~B) == (~(A | B)) 1573 /// (~A | ~B) == (~(A & B)) 1574 static Instruction *matchDeMorgansLaws(BinaryOperator &I, 1575 InstCombiner &IC) { 1576 const Instruction::BinaryOps Opcode = I.getOpcode(); 1577 assert((Opcode == Instruction::And || Opcode == Instruction::Or) && 1578 "Trying to match De Morgan's Laws with something other than and/or"); 1579 1580 // Flip the logic operation. 1581 const Instruction::BinaryOps FlippedOpcode = 1582 (Opcode == Instruction::And) ? Instruction::Or : Instruction::And; 1583 1584 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1585 Value *A, *B; 1586 if (match(Op0, m_OneUse(m_Not(m_Value(A)))) && 1587 match(Op1, m_OneUse(m_Not(m_Value(B)))) && 1588 !IC.isFreeToInvert(A, A->hasOneUse()) && 1589 !IC.isFreeToInvert(B, B->hasOneUse())) { 1590 Value *AndOr = 1591 IC.Builder.CreateBinOp(FlippedOpcode, A, B, I.getName() + ".demorgan"); 1592 return BinaryOperator::CreateNot(AndOr); 1593 } 1594 1595 // The 'not' ops may require reassociation. 1596 // (A & ~B) & ~C --> A & ~(B | C) 1597 // (~B & A) & ~C --> A & ~(B | C) 1598 // (A | ~B) | ~C --> A | ~(B & C) 1599 // (~B | A) | ~C --> A | ~(B & C) 1600 Value *C; 1601 if (match(Op0, m_OneUse(m_c_BinOp(Opcode, m_Value(A), m_Not(m_Value(B))))) && 1602 match(Op1, m_Not(m_Value(C)))) { 1603 Value *FlippedBO = IC.Builder.CreateBinOp(FlippedOpcode, B, C); 1604 return BinaryOperator::Create(Opcode, A, IC.Builder.CreateNot(FlippedBO)); 1605 } 1606 1607 return nullptr; 1608 } 1609 1610 bool InstCombinerImpl::shouldOptimizeCast(CastInst *CI) { 1611 Value *CastSrc = CI->getOperand(0); 1612 1613 // Noop casts and casts of constants should be eliminated trivially. 1614 if (CI->getSrcTy() == CI->getDestTy() || isa<Constant>(CastSrc)) 1615 return false; 1616 1617 // If this cast is paired with another cast that can be eliminated, we prefer 1618 // to have it eliminated. 1619 if (const auto *PrecedingCI = dyn_cast<CastInst>(CastSrc)) 1620 if (isEliminableCastPair(PrecedingCI, CI)) 1621 return false; 1622 1623 return true; 1624 } 1625 1626 /// Fold {and,or,xor} (cast X), C. 1627 static Instruction *foldLogicCastConstant(BinaryOperator &Logic, CastInst *Cast, 1628 InstCombinerImpl &IC) { 1629 Constant *C = dyn_cast<Constant>(Logic.getOperand(1)); 1630 if (!C) 1631 return nullptr; 1632 1633 auto LogicOpc = Logic.getOpcode(); 1634 Type *DestTy = Logic.getType(); 1635 Type *SrcTy = Cast->getSrcTy(); 1636 1637 // Move the logic operation ahead of a zext or sext if the constant is 1638 // unchanged in the smaller source type. Performing the logic in a smaller 1639 // type may provide more information to later folds, and the smaller logic 1640 // instruction may be cheaper (particularly in the case of vectors). 1641 Value *X; 1642 if (match(Cast, m_OneUse(m_ZExt(m_Value(X))))) { 1643 if (Constant *TruncC = IC.getLosslessUnsignedTrunc(C, SrcTy)) { 1644 // LogicOpc (zext X), C --> zext (LogicOpc X, C) 1645 Value *NewOp = IC.Builder.CreateBinOp(LogicOpc, X, TruncC); 1646 return new ZExtInst(NewOp, DestTy); 1647 } 1648 } 1649 1650 if (match(Cast, m_OneUse(m_SExt(m_Value(X))))) { 1651 if (Constant *TruncC = IC.getLosslessSignedTrunc(C, SrcTy)) { 1652 // LogicOpc (sext X), C --> sext (LogicOpc X, C) 1653 Value *NewOp = IC.Builder.CreateBinOp(LogicOpc, X, TruncC); 1654 return new SExtInst(NewOp, DestTy); 1655 } 1656 } 1657 1658 return nullptr; 1659 } 1660 1661 /// Fold {and,or,xor} (cast X), Y. 1662 Instruction *InstCombinerImpl::foldCastedBitwiseLogic(BinaryOperator &I) { 1663 auto LogicOpc = I.getOpcode(); 1664 assert(I.isBitwiseLogicOp() && "Unexpected opcode for bitwise logic folding"); 1665 1666 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1667 1668 // fold bitwise(A >> BW - 1, zext(icmp)) (BW is the scalar bits of the 1669 // type of A) 1670 // -> bitwise(zext(A < 0), zext(icmp)) 1671 // -> zext(bitwise(A < 0, icmp)) 1672 auto FoldBitwiseICmpZeroWithICmp = [&](Value *Op0, 1673 Value *Op1) -> Instruction * { 1674 ICmpInst::Predicate Pred; 1675 Value *A; 1676 bool IsMatched = 1677 match(Op0, 1678 m_OneUse(m_LShr( 1679 m_Value(A), 1680 m_SpecificInt(Op0->getType()->getScalarSizeInBits() - 1)))) && 1681 match(Op1, m_OneUse(m_ZExt(m_ICmp(Pred, m_Value(), m_Value())))); 1682 1683 if (!IsMatched) 1684 return nullptr; 1685 1686 auto *ICmpL = 1687 Builder.CreateICmpSLT(A, Constant::getNullValue(A->getType())); 1688 auto *ICmpR = cast<ZExtInst>(Op1)->getOperand(0); 1689 auto *BitwiseOp = Builder.CreateBinOp(LogicOpc, ICmpL, ICmpR); 1690 1691 return new ZExtInst(BitwiseOp, Op0->getType()); 1692 }; 1693 1694 if (auto *Ret = FoldBitwiseICmpZeroWithICmp(Op0, Op1)) 1695 return Ret; 1696 1697 if (auto *Ret = FoldBitwiseICmpZeroWithICmp(Op1, Op0)) 1698 return Ret; 1699 1700 CastInst *Cast0 = dyn_cast<CastInst>(Op0); 1701 if (!Cast0) 1702 return nullptr; 1703 1704 // This must be a cast from an integer or integer vector source type to allow 1705 // transformation of the logic operation to the source type. 1706 Type *DestTy = I.getType(); 1707 Type *SrcTy = Cast0->getSrcTy(); 1708 if (!SrcTy->isIntOrIntVectorTy()) 1709 return nullptr; 1710 1711 if (Instruction *Ret = foldLogicCastConstant(I, Cast0, *this)) 1712 return Ret; 1713 1714 CastInst *Cast1 = dyn_cast<CastInst>(Op1); 1715 if (!Cast1) 1716 return nullptr; 1717 1718 // Both operands of the logic operation are casts. The casts must be the 1719 // same kind for reduction. 1720 Instruction::CastOps CastOpcode = Cast0->getOpcode(); 1721 if (CastOpcode != Cast1->getOpcode()) 1722 return nullptr; 1723 1724 // If the source types do not match, but the casts are matching extends, we 1725 // can still narrow the logic op. 1726 if (SrcTy != Cast1->getSrcTy()) { 1727 Value *X, *Y; 1728 if (match(Cast0, m_OneUse(m_ZExtOrSExt(m_Value(X)))) && 1729 match(Cast1, m_OneUse(m_ZExtOrSExt(m_Value(Y))))) { 1730 // Cast the narrower source to the wider source type. 1731 unsigned XNumBits = X->getType()->getScalarSizeInBits(); 1732 unsigned YNumBits = Y->getType()->getScalarSizeInBits(); 1733 if (XNumBits < YNumBits) 1734 X = Builder.CreateCast(CastOpcode, X, Y->getType()); 1735 else 1736 Y = Builder.CreateCast(CastOpcode, Y, X->getType()); 1737 // Do the logic op in the intermediate width, then widen more. 1738 Value *NarrowLogic = Builder.CreateBinOp(LogicOpc, X, Y); 1739 return CastInst::Create(CastOpcode, NarrowLogic, DestTy); 1740 } 1741 1742 // Give up for other cast opcodes. 1743 return nullptr; 1744 } 1745 1746 Value *Cast0Src = Cast0->getOperand(0); 1747 Value *Cast1Src = Cast1->getOperand(0); 1748 1749 // fold logic(cast(A), cast(B)) -> cast(logic(A, B)) 1750 if ((Cast0->hasOneUse() || Cast1->hasOneUse()) && 1751 shouldOptimizeCast(Cast0) && shouldOptimizeCast(Cast1)) { 1752 Value *NewOp = Builder.CreateBinOp(LogicOpc, Cast0Src, Cast1Src, 1753 I.getName()); 1754 return CastInst::Create(CastOpcode, NewOp, DestTy); 1755 } 1756 1757 return nullptr; 1758 } 1759 1760 static Instruction *foldAndToXor(BinaryOperator &I, 1761 InstCombiner::BuilderTy &Builder) { 1762 assert(I.getOpcode() == Instruction::And); 1763 Value *Op0 = I.getOperand(0); 1764 Value *Op1 = I.getOperand(1); 1765 Value *A, *B; 1766 1767 // Operand complexity canonicalization guarantees that the 'or' is Op0. 1768 // (A | B) & ~(A & B) --> A ^ B 1769 // (A | B) & ~(B & A) --> A ^ B 1770 if (match(&I, m_BinOp(m_Or(m_Value(A), m_Value(B)), 1771 m_Not(m_c_And(m_Deferred(A), m_Deferred(B)))))) 1772 return BinaryOperator::CreateXor(A, B); 1773 1774 // (A | ~B) & (~A | B) --> ~(A ^ B) 1775 // (A | ~B) & (B | ~A) --> ~(A ^ B) 1776 // (~B | A) & (~A | B) --> ~(A ^ B) 1777 // (~B | A) & (B | ~A) --> ~(A ^ B) 1778 if (Op0->hasOneUse() || Op1->hasOneUse()) 1779 if (match(&I, m_BinOp(m_c_Or(m_Value(A), m_Not(m_Value(B))), 1780 m_c_Or(m_Not(m_Deferred(A)), m_Deferred(B))))) 1781 return BinaryOperator::CreateNot(Builder.CreateXor(A, B)); 1782 1783 return nullptr; 1784 } 1785 1786 static Instruction *foldOrToXor(BinaryOperator &I, 1787 InstCombiner::BuilderTy &Builder) { 1788 assert(I.getOpcode() == Instruction::Or); 1789 Value *Op0 = I.getOperand(0); 1790 Value *Op1 = I.getOperand(1); 1791 Value *A, *B; 1792 1793 // Operand complexity canonicalization guarantees that the 'and' is Op0. 1794 // (A & B) | ~(A | B) --> ~(A ^ B) 1795 // (A & B) | ~(B | A) --> ~(A ^ B) 1796 if (Op0->hasOneUse() || Op1->hasOneUse()) 1797 if (match(Op0, m_And(m_Value(A), m_Value(B))) && 1798 match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B))))) 1799 return BinaryOperator::CreateNot(Builder.CreateXor(A, B)); 1800 1801 // Operand complexity canonicalization guarantees that the 'xor' is Op0. 1802 // (A ^ B) | ~(A | B) --> ~(A & B) 1803 // (A ^ B) | ~(B | A) --> ~(A & B) 1804 if (Op0->hasOneUse() || Op1->hasOneUse()) 1805 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) && 1806 match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B))))) 1807 return BinaryOperator::CreateNot(Builder.CreateAnd(A, B)); 1808 1809 // (A & ~B) | (~A & B) --> A ^ B 1810 // (A & ~B) | (B & ~A) --> A ^ B 1811 // (~B & A) | (~A & B) --> A ^ B 1812 // (~B & A) | (B & ~A) --> A ^ B 1813 if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) && 1814 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))) 1815 return BinaryOperator::CreateXor(A, B); 1816 1817 return nullptr; 1818 } 1819 1820 /// Return true if a constant shift amount is always less than the specified 1821 /// bit-width. If not, the shift could create poison in the narrower type. 1822 static bool canNarrowShiftAmt(Constant *C, unsigned BitWidth) { 1823 APInt Threshold(C->getType()->getScalarSizeInBits(), BitWidth); 1824 return match(C, m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, Threshold)); 1825 } 1826 1827 /// Try to use narrower ops (sink zext ops) for an 'and' with binop operand and 1828 /// a common zext operand: and (binop (zext X), C), (zext X). 1829 Instruction *InstCombinerImpl::narrowMaskedBinOp(BinaryOperator &And) { 1830 // This transform could also apply to {or, and, xor}, but there are better 1831 // folds for those cases, so we don't expect those patterns here. AShr is not 1832 // handled because it should always be transformed to LShr in this sequence. 1833 // The subtract transform is different because it has a constant on the left. 1834 // Add/mul commute the constant to RHS; sub with constant RHS becomes add. 1835 Value *Op0 = And.getOperand(0), *Op1 = And.getOperand(1); 1836 Constant *C; 1837 if (!match(Op0, m_OneUse(m_Add(m_Specific(Op1), m_Constant(C)))) && 1838 !match(Op0, m_OneUse(m_Mul(m_Specific(Op1), m_Constant(C)))) && 1839 !match(Op0, m_OneUse(m_LShr(m_Specific(Op1), m_Constant(C)))) && 1840 !match(Op0, m_OneUse(m_Shl(m_Specific(Op1), m_Constant(C)))) && 1841 !match(Op0, m_OneUse(m_Sub(m_Constant(C), m_Specific(Op1))))) 1842 return nullptr; 1843 1844 Value *X; 1845 if (!match(Op1, m_ZExt(m_Value(X))) || Op1->hasNUsesOrMore(3)) 1846 return nullptr; 1847 1848 Type *Ty = And.getType(); 1849 if (!isa<VectorType>(Ty) && !shouldChangeType(Ty, X->getType())) 1850 return nullptr; 1851 1852 // If we're narrowing a shift, the shift amount must be safe (less than the 1853 // width) in the narrower type. If the shift amount is greater, instsimplify 1854 // usually handles that case, but we can't guarantee/assert it. 1855 Instruction::BinaryOps Opc = cast<BinaryOperator>(Op0)->getOpcode(); 1856 if (Opc == Instruction::LShr || Opc == Instruction::Shl) 1857 if (!canNarrowShiftAmt(C, X->getType()->getScalarSizeInBits())) 1858 return nullptr; 1859 1860 // and (sub C, (zext X)), (zext X) --> zext (and (sub C', X), X) 1861 // and (binop (zext X), C), (zext X) --> zext (and (binop X, C'), X) 1862 Value *NewC = ConstantExpr::getTrunc(C, X->getType()); 1863 Value *NewBO = Opc == Instruction::Sub ? Builder.CreateBinOp(Opc, NewC, X) 1864 : Builder.CreateBinOp(Opc, X, NewC); 1865 return new ZExtInst(Builder.CreateAnd(NewBO, X), Ty); 1866 } 1867 1868 /// Try folding relatively complex patterns for both And and Or operations 1869 /// with all And and Or swapped. 1870 static Instruction *foldComplexAndOrPatterns(BinaryOperator &I, 1871 InstCombiner::BuilderTy &Builder) { 1872 const Instruction::BinaryOps Opcode = I.getOpcode(); 1873 assert(Opcode == Instruction::And || Opcode == Instruction::Or); 1874 1875 // Flip the logic operation. 1876 const Instruction::BinaryOps FlippedOpcode = 1877 (Opcode == Instruction::And) ? Instruction::Or : Instruction::And; 1878 1879 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1880 Value *A, *B, *C, *X, *Y, *Dummy; 1881 1882 // Match following expressions: 1883 // (~(A | B) & C) 1884 // (~(A & B) | C) 1885 // Captures X = ~(A | B) or ~(A & B) 1886 const auto matchNotOrAnd = 1887 [Opcode, FlippedOpcode](Value *Op, auto m_A, auto m_B, auto m_C, 1888 Value *&X, bool CountUses = false) -> bool { 1889 if (CountUses && !Op->hasOneUse()) 1890 return false; 1891 1892 if (match(Op, m_c_BinOp(FlippedOpcode, 1893 m_CombineAnd(m_Value(X), 1894 m_Not(m_c_BinOp(Opcode, m_A, m_B))), 1895 m_C))) 1896 return !CountUses || X->hasOneUse(); 1897 1898 return false; 1899 }; 1900 1901 // (~(A | B) & C) | ... --> ... 1902 // (~(A & B) | C) & ... --> ... 1903 // TODO: One use checks are conservative. We just need to check that a total 1904 // number of multiple used values does not exceed reduction 1905 // in operations. 1906 if (matchNotOrAnd(Op0, m_Value(A), m_Value(B), m_Value(C), X)) { 1907 // (~(A | B) & C) | (~(A | C) & B) --> (B ^ C) & ~A 1908 // (~(A & B) | C) & (~(A & C) | B) --> ~((B ^ C) & A) 1909 if (matchNotOrAnd(Op1, m_Specific(A), m_Specific(C), m_Specific(B), Dummy, 1910 true)) { 1911 Value *Xor = Builder.CreateXor(B, C); 1912 return (Opcode == Instruction::Or) 1913 ? BinaryOperator::CreateAnd(Xor, Builder.CreateNot(A)) 1914 : BinaryOperator::CreateNot(Builder.CreateAnd(Xor, A)); 1915 } 1916 1917 // (~(A | B) & C) | (~(B | C) & A) --> (A ^ C) & ~B 1918 // (~(A & B) | C) & (~(B & C) | A) --> ~((A ^ C) & B) 1919 if (matchNotOrAnd(Op1, m_Specific(B), m_Specific(C), m_Specific(A), Dummy, 1920 true)) { 1921 Value *Xor = Builder.CreateXor(A, C); 1922 return (Opcode == Instruction::Or) 1923 ? BinaryOperator::CreateAnd(Xor, Builder.CreateNot(B)) 1924 : BinaryOperator::CreateNot(Builder.CreateAnd(Xor, B)); 1925 } 1926 1927 // (~(A | B) & C) | ~(A | C) --> ~((B & C) | A) 1928 // (~(A & B) | C) & ~(A & C) --> ~((B | C) & A) 1929 if (match(Op1, m_OneUse(m_Not(m_OneUse( 1930 m_c_BinOp(Opcode, m_Specific(A), m_Specific(C))))))) 1931 return BinaryOperator::CreateNot(Builder.CreateBinOp( 1932 Opcode, Builder.CreateBinOp(FlippedOpcode, B, C), A)); 1933 1934 // (~(A | B) & C) | ~(B | C) --> ~((A & C) | B) 1935 // (~(A & B) | C) & ~(B & C) --> ~((A | C) & B) 1936 if (match(Op1, m_OneUse(m_Not(m_OneUse( 1937 m_c_BinOp(Opcode, m_Specific(B), m_Specific(C))))))) 1938 return BinaryOperator::CreateNot(Builder.CreateBinOp( 1939 Opcode, Builder.CreateBinOp(FlippedOpcode, A, C), B)); 1940 1941 // (~(A | B) & C) | ~(C | (A ^ B)) --> ~((A | B) & (C | (A ^ B))) 1942 // Note, the pattern with swapped and/or is not handled because the 1943 // result is more undefined than a source: 1944 // (~(A & B) | C) & ~(C & (A ^ B)) --> (A ^ B ^ C) | ~(A | C) is invalid. 1945 if (Opcode == Instruction::Or && Op0->hasOneUse() && 1946 match(Op1, m_OneUse(m_Not(m_CombineAnd( 1947 m_Value(Y), 1948 m_c_BinOp(Opcode, m_Specific(C), 1949 m_c_Xor(m_Specific(A), m_Specific(B)))))))) { 1950 // X = ~(A | B) 1951 // Y = (C | (A ^ B) 1952 Value *Or = cast<BinaryOperator>(X)->getOperand(0); 1953 return BinaryOperator::CreateNot(Builder.CreateAnd(Or, Y)); 1954 } 1955 } 1956 1957 // (~A & B & C) | ... --> ... 1958 // (~A | B | C) | ... --> ... 1959 // TODO: One use checks are conservative. We just need to check that a total 1960 // number of multiple used values does not exceed reduction 1961 // in operations. 1962 if (match(Op0, 1963 m_OneUse(m_c_BinOp(FlippedOpcode, 1964 m_BinOp(FlippedOpcode, m_Value(B), m_Value(C)), 1965 m_CombineAnd(m_Value(X), m_Not(m_Value(A)))))) || 1966 match(Op0, m_OneUse(m_c_BinOp( 1967 FlippedOpcode, 1968 m_c_BinOp(FlippedOpcode, m_Value(C), 1969 m_CombineAnd(m_Value(X), m_Not(m_Value(A)))), 1970 m_Value(B))))) { 1971 // X = ~A 1972 // (~A & B & C) | ~(A | B | C) --> ~(A | (B ^ C)) 1973 // (~A | B | C) & ~(A & B & C) --> (~A | (B ^ C)) 1974 if (match(Op1, m_OneUse(m_Not(m_c_BinOp( 1975 Opcode, m_c_BinOp(Opcode, m_Specific(A), m_Specific(B)), 1976 m_Specific(C))))) || 1977 match(Op1, m_OneUse(m_Not(m_c_BinOp( 1978 Opcode, m_c_BinOp(Opcode, m_Specific(B), m_Specific(C)), 1979 m_Specific(A))))) || 1980 match(Op1, m_OneUse(m_Not(m_c_BinOp( 1981 Opcode, m_c_BinOp(Opcode, m_Specific(A), m_Specific(C)), 1982 m_Specific(B)))))) { 1983 Value *Xor = Builder.CreateXor(B, C); 1984 return (Opcode == Instruction::Or) 1985 ? BinaryOperator::CreateNot(Builder.CreateOr(Xor, A)) 1986 : BinaryOperator::CreateOr(Xor, X); 1987 } 1988 1989 // (~A & B & C) | ~(A | B) --> (C | ~B) & ~A 1990 // (~A | B | C) & ~(A & B) --> (C & ~B) | ~A 1991 if (match(Op1, m_OneUse(m_Not(m_OneUse( 1992 m_c_BinOp(Opcode, m_Specific(A), m_Specific(B))))))) 1993 return BinaryOperator::Create( 1994 FlippedOpcode, Builder.CreateBinOp(Opcode, C, Builder.CreateNot(B)), 1995 X); 1996 1997 // (~A & B & C) | ~(A | C) --> (B | ~C) & ~A 1998 // (~A | B | C) & ~(A & C) --> (B & ~C) | ~A 1999 if (match(Op1, m_OneUse(m_Not(m_OneUse( 2000 m_c_BinOp(Opcode, m_Specific(A), m_Specific(C))))))) 2001 return BinaryOperator::Create( 2002 FlippedOpcode, Builder.CreateBinOp(Opcode, B, Builder.CreateNot(C)), 2003 X); 2004 } 2005 2006 return nullptr; 2007 } 2008 2009 /// Try to reassociate a pair of binops so that values with one use only are 2010 /// part of the same instruction. This may enable folds that are limited with 2011 /// multi-use restrictions and makes it more likely to match other patterns that 2012 /// are looking for a common operand. 2013 static Instruction *reassociateForUses(BinaryOperator &BO, 2014 InstCombinerImpl::BuilderTy &Builder) { 2015 Instruction::BinaryOps Opcode = BO.getOpcode(); 2016 Value *X, *Y, *Z; 2017 if (match(&BO, 2018 m_c_BinOp(Opcode, m_OneUse(m_BinOp(Opcode, m_Value(X), m_Value(Y))), 2019 m_OneUse(m_Value(Z))))) { 2020 if (!isa<Constant>(X) && !isa<Constant>(Y) && !isa<Constant>(Z)) { 2021 // (X op Y) op Z --> (Y op Z) op X 2022 if (!X->hasOneUse()) { 2023 Value *YZ = Builder.CreateBinOp(Opcode, Y, Z); 2024 return BinaryOperator::Create(Opcode, YZ, X); 2025 } 2026 // (X op Y) op Z --> (X op Z) op Y 2027 if (!Y->hasOneUse()) { 2028 Value *XZ = Builder.CreateBinOp(Opcode, X, Z); 2029 return BinaryOperator::Create(Opcode, XZ, Y); 2030 } 2031 } 2032 } 2033 2034 return nullptr; 2035 } 2036 2037 // Match 2038 // (X + C2) | C 2039 // (X + C2) ^ C 2040 // (X + C2) & C 2041 // and convert to do the bitwise logic first: 2042 // (X | C) + C2 2043 // (X ^ C) + C2 2044 // (X & C) + C2 2045 // iff bits affected by logic op are lower than last bit affected by math op 2046 static Instruction *canonicalizeLogicFirst(BinaryOperator &I, 2047 InstCombiner::BuilderTy &Builder) { 2048 Type *Ty = I.getType(); 2049 Instruction::BinaryOps OpC = I.getOpcode(); 2050 Value *Op0 = I.getOperand(0); 2051 Value *Op1 = I.getOperand(1); 2052 Value *X; 2053 const APInt *C, *C2; 2054 2055 if (!(match(Op0, m_OneUse(m_Add(m_Value(X), m_APInt(C2)))) && 2056 match(Op1, m_APInt(C)))) 2057 return nullptr; 2058 2059 unsigned Width = Ty->getScalarSizeInBits(); 2060 unsigned LastOneMath = Width - C2->countr_zero(); 2061 2062 switch (OpC) { 2063 case Instruction::And: 2064 if (C->countl_one() < LastOneMath) 2065 return nullptr; 2066 break; 2067 case Instruction::Xor: 2068 case Instruction::Or: 2069 if (C->countl_zero() < LastOneMath) 2070 return nullptr; 2071 break; 2072 default: 2073 llvm_unreachable("Unexpected BinaryOp!"); 2074 } 2075 2076 Value *NewBinOp = Builder.CreateBinOp(OpC, X, ConstantInt::get(Ty, *C)); 2077 return BinaryOperator::CreateWithCopiedFlags(Instruction::Add, NewBinOp, 2078 ConstantInt::get(Ty, *C2), Op0); 2079 } 2080 2081 // binop(shift(ShiftedC1, ShAmt), shift(ShiftedC2, add(ShAmt, AddC))) -> 2082 // shift(binop(ShiftedC1, shift(ShiftedC2, AddC)), ShAmt) 2083 // where both shifts are the same and AddC is a valid shift amount. 2084 Instruction *InstCombinerImpl::foldBinOpOfDisplacedShifts(BinaryOperator &I) { 2085 assert((I.isBitwiseLogicOp() || I.getOpcode() == Instruction::Add) && 2086 "Unexpected opcode"); 2087 2088 Value *ShAmt; 2089 Constant *ShiftedC1, *ShiftedC2, *AddC; 2090 Type *Ty = I.getType(); 2091 unsigned BitWidth = Ty->getScalarSizeInBits(); 2092 if (!match(&I, m_c_BinOp(m_Shift(m_ImmConstant(ShiftedC1), m_Value(ShAmt)), 2093 m_Shift(m_ImmConstant(ShiftedC2), 2094 m_AddLike(m_Deferred(ShAmt), 2095 m_ImmConstant(AddC)))))) 2096 return nullptr; 2097 2098 // Make sure the add constant is a valid shift amount. 2099 if (!match(AddC, 2100 m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, APInt(BitWidth, BitWidth)))) 2101 return nullptr; 2102 2103 // Avoid constant expressions. 2104 auto *Op0Inst = dyn_cast<Instruction>(I.getOperand(0)); 2105 auto *Op1Inst = dyn_cast<Instruction>(I.getOperand(1)); 2106 if (!Op0Inst || !Op1Inst) 2107 return nullptr; 2108 2109 // Both shifts must be the same. 2110 Instruction::BinaryOps ShiftOp = 2111 static_cast<Instruction::BinaryOps>(Op0Inst->getOpcode()); 2112 if (ShiftOp != Op1Inst->getOpcode()) 2113 return nullptr; 2114 2115 // For adds, only left shifts are supported. 2116 if (I.getOpcode() == Instruction::Add && ShiftOp != Instruction::Shl) 2117 return nullptr; 2118 2119 Value *NewC = Builder.CreateBinOp( 2120 I.getOpcode(), ShiftedC1, Builder.CreateBinOp(ShiftOp, ShiftedC2, AddC)); 2121 return BinaryOperator::Create(ShiftOp, NewC, ShAmt); 2122 } 2123 2124 // Fold and/or/xor with two equal intrinsic IDs: 2125 // bitwise(fshl (A, B, ShAmt), fshl(C, D, ShAmt)) 2126 // -> fshl(bitwise(A, C), bitwise(B, D), ShAmt) 2127 // bitwise(fshr (A, B, ShAmt), fshr(C, D, ShAmt)) 2128 // -> fshr(bitwise(A, C), bitwise(B, D), ShAmt) 2129 // bitwise(bswap(A), bswap(B)) -> bswap(bitwise(A, B)) 2130 // bitwise(bswap(A), C) -> bswap(bitwise(A, bswap(C))) 2131 // bitwise(bitreverse(A), bitreverse(B)) -> bitreverse(bitwise(A, B)) 2132 // bitwise(bitreverse(A), C) -> bitreverse(bitwise(A, bitreverse(C))) 2133 static Instruction * 2134 foldBitwiseLogicWithIntrinsics(BinaryOperator &I, 2135 InstCombiner::BuilderTy &Builder) { 2136 assert(I.isBitwiseLogicOp() && "Should and/or/xor"); 2137 if (!I.getOperand(0)->hasOneUse()) 2138 return nullptr; 2139 IntrinsicInst *X = dyn_cast<IntrinsicInst>(I.getOperand(0)); 2140 if (!X) 2141 return nullptr; 2142 2143 IntrinsicInst *Y = dyn_cast<IntrinsicInst>(I.getOperand(1)); 2144 if (Y && (!Y->hasOneUse() || X->getIntrinsicID() != Y->getIntrinsicID())) 2145 return nullptr; 2146 2147 Intrinsic::ID IID = X->getIntrinsicID(); 2148 const APInt *RHSC; 2149 // Try to match constant RHS. 2150 if (!Y && (!(IID == Intrinsic::bswap || IID == Intrinsic::bitreverse) || 2151 !match(I.getOperand(1), m_APInt(RHSC)))) 2152 return nullptr; 2153 2154 switch (IID) { 2155 case Intrinsic::fshl: 2156 case Intrinsic::fshr: { 2157 if (X->getOperand(2) != Y->getOperand(2)) 2158 return nullptr; 2159 Value *NewOp0 = 2160 Builder.CreateBinOp(I.getOpcode(), X->getOperand(0), Y->getOperand(0)); 2161 Value *NewOp1 = 2162 Builder.CreateBinOp(I.getOpcode(), X->getOperand(1), Y->getOperand(1)); 2163 Function *F = Intrinsic::getDeclaration(I.getModule(), IID, I.getType()); 2164 return CallInst::Create(F, {NewOp0, NewOp1, X->getOperand(2)}); 2165 } 2166 case Intrinsic::bswap: 2167 case Intrinsic::bitreverse: { 2168 Value *NewOp0 = Builder.CreateBinOp( 2169 I.getOpcode(), X->getOperand(0), 2170 Y ? Y->getOperand(0) 2171 : ConstantInt::get(I.getType(), IID == Intrinsic::bswap 2172 ? RHSC->byteSwap() 2173 : RHSC->reverseBits())); 2174 Function *F = Intrinsic::getDeclaration(I.getModule(), IID, I.getType()); 2175 return CallInst::Create(F, {NewOp0}); 2176 } 2177 default: 2178 return nullptr; 2179 } 2180 } 2181 2182 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches 2183 // here. We should standardize that construct where it is needed or choose some 2184 // other way to ensure that commutated variants of patterns are not missed. 2185 Instruction *InstCombinerImpl::visitAnd(BinaryOperator &I) { 2186 Type *Ty = I.getType(); 2187 2188 if (Value *V = simplifyAndInst(I.getOperand(0), I.getOperand(1), 2189 SQ.getWithInstruction(&I))) 2190 return replaceInstUsesWith(I, V); 2191 2192 if (SimplifyAssociativeOrCommutative(I)) 2193 return &I; 2194 2195 if (Instruction *X = foldVectorBinop(I)) 2196 return X; 2197 2198 if (Instruction *Phi = foldBinopWithPhiOperands(I)) 2199 return Phi; 2200 2201 // See if we can simplify any instructions used by the instruction whose sole 2202 // purpose is to compute bits we don't care about. 2203 if (SimplifyDemandedInstructionBits(I)) 2204 return &I; 2205 2206 // Do this before using distributive laws to catch simple and/or/not patterns. 2207 if (Instruction *Xor = foldAndToXor(I, Builder)) 2208 return Xor; 2209 2210 if (Instruction *X = foldComplexAndOrPatterns(I, Builder)) 2211 return X; 2212 2213 // (A|B)&(A|C) -> A|(B&C) etc 2214 if (Value *V = foldUsingDistributiveLaws(I)) 2215 return replaceInstUsesWith(I, V); 2216 2217 if (Instruction *R = foldBinOpShiftWithShift(I)) 2218 return R; 2219 2220 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 2221 2222 Value *X, *Y; 2223 if (match(Op0, m_OneUse(m_LogicalShift(m_One(), m_Value(X)))) && 2224 match(Op1, m_One())) { 2225 // (1 << X) & 1 --> zext(X == 0) 2226 // (1 >> X) & 1 --> zext(X == 0) 2227 Value *IsZero = Builder.CreateICmpEQ(X, ConstantInt::get(Ty, 0)); 2228 return new ZExtInst(IsZero, Ty); 2229 } 2230 2231 // (-(X & 1)) & Y --> (X & 1) == 0 ? 0 : Y 2232 Value *Neg; 2233 if (match(&I, 2234 m_c_And(m_CombineAnd(m_Value(Neg), 2235 m_OneUse(m_Neg(m_And(m_Value(), m_One())))), 2236 m_Value(Y)))) { 2237 Value *Cmp = Builder.CreateIsNull(Neg); 2238 return SelectInst::Create(Cmp, ConstantInt::getNullValue(Ty), Y); 2239 } 2240 2241 // Canonicalize: 2242 // (X +/- Y) & Y --> ~X & Y when Y is a power of 2. 2243 if (match(&I, m_c_And(m_Value(Y), m_OneUse(m_CombineOr( 2244 m_c_Add(m_Value(X), m_Deferred(Y)), 2245 m_Sub(m_Value(X), m_Deferred(Y)))))) && 2246 isKnownToBeAPowerOfTwo(Y, /*OrZero*/ true, /*Depth*/ 0, &I)) 2247 return BinaryOperator::CreateAnd(Builder.CreateNot(X), Y); 2248 2249 const APInt *C; 2250 if (match(Op1, m_APInt(C))) { 2251 const APInt *XorC; 2252 if (match(Op0, m_OneUse(m_Xor(m_Value(X), m_APInt(XorC))))) { 2253 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2) 2254 Constant *NewC = ConstantInt::get(Ty, *C & *XorC); 2255 Value *And = Builder.CreateAnd(X, Op1); 2256 And->takeName(Op0); 2257 return BinaryOperator::CreateXor(And, NewC); 2258 } 2259 2260 const APInt *OrC; 2261 if (match(Op0, m_OneUse(m_Or(m_Value(X), m_APInt(OrC))))) { 2262 // (X | C1) & C2 --> (X & C2^(C1&C2)) | (C1&C2) 2263 // NOTE: This reduces the number of bits set in the & mask, which 2264 // can expose opportunities for store narrowing for scalars. 2265 // NOTE: SimplifyDemandedBits should have already removed bits from C1 2266 // that aren't set in C2. Meaning we can replace (C1&C2) with C1 in 2267 // above, but this feels safer. 2268 APInt Together = *C & *OrC; 2269 Value *And = Builder.CreateAnd(X, ConstantInt::get(Ty, Together ^ *C)); 2270 And->takeName(Op0); 2271 return BinaryOperator::CreateOr(And, ConstantInt::get(Ty, Together)); 2272 } 2273 2274 unsigned Width = Ty->getScalarSizeInBits(); 2275 const APInt *ShiftC; 2276 if (match(Op0, m_OneUse(m_SExt(m_AShr(m_Value(X), m_APInt(ShiftC))))) && 2277 ShiftC->ult(Width)) { 2278 if (*C == APInt::getLowBitsSet(Width, Width - ShiftC->getZExtValue())) { 2279 // We are clearing high bits that were potentially set by sext+ashr: 2280 // and (sext (ashr X, ShiftC)), C --> lshr (sext X), ShiftC 2281 Value *Sext = Builder.CreateSExt(X, Ty); 2282 Constant *ShAmtC = ConstantInt::get(Ty, ShiftC->zext(Width)); 2283 return BinaryOperator::CreateLShr(Sext, ShAmtC); 2284 } 2285 } 2286 2287 // If this 'and' clears the sign-bits added by ashr, replace with lshr: 2288 // and (ashr X, ShiftC), C --> lshr X, ShiftC 2289 if (match(Op0, m_AShr(m_Value(X), m_APInt(ShiftC))) && ShiftC->ult(Width) && 2290 C->isMask(Width - ShiftC->getZExtValue())) 2291 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, *ShiftC)); 2292 2293 const APInt *AddC; 2294 if (match(Op0, m_Add(m_Value(X), m_APInt(AddC)))) { 2295 // If we are masking the result of the add down to exactly one bit and 2296 // the constant we are adding has no bits set below that bit, then the 2297 // add is flipping a single bit. Example: 2298 // (X + 4) & 4 --> (X & 4) ^ 4 2299 if (Op0->hasOneUse() && C->isPowerOf2() && (*AddC & (*C - 1)) == 0) { 2300 assert((*C & *AddC) != 0 && "Expected common bit"); 2301 Value *NewAnd = Builder.CreateAnd(X, Op1); 2302 return BinaryOperator::CreateXor(NewAnd, Op1); 2303 } 2304 } 2305 2306 // ((C1 OP zext(X)) & C2) -> zext((C1 OP X) & C2) if C2 fits in the 2307 // bitwidth of X and OP behaves well when given trunc(C1) and X. 2308 auto isNarrowableBinOpcode = [](BinaryOperator *B) { 2309 switch (B->getOpcode()) { 2310 case Instruction::Xor: 2311 case Instruction::Or: 2312 case Instruction::Mul: 2313 case Instruction::Add: 2314 case Instruction::Sub: 2315 return true; 2316 default: 2317 return false; 2318 } 2319 }; 2320 BinaryOperator *BO; 2321 if (match(Op0, m_OneUse(m_BinOp(BO))) && isNarrowableBinOpcode(BO)) { 2322 Instruction::BinaryOps BOpcode = BO->getOpcode(); 2323 Value *X; 2324 const APInt *C1; 2325 // TODO: The one-use restrictions could be relaxed a little if the AND 2326 // is going to be removed. 2327 // Try to narrow the 'and' and a binop with constant operand: 2328 // and (bo (zext X), C1), C --> zext (and (bo X, TruncC1), TruncC) 2329 if (match(BO, m_c_BinOp(m_OneUse(m_ZExt(m_Value(X))), m_APInt(C1))) && 2330 C->isIntN(X->getType()->getScalarSizeInBits())) { 2331 unsigned XWidth = X->getType()->getScalarSizeInBits(); 2332 Constant *TruncC1 = ConstantInt::get(X->getType(), C1->trunc(XWidth)); 2333 Value *BinOp = isa<ZExtInst>(BO->getOperand(0)) 2334 ? Builder.CreateBinOp(BOpcode, X, TruncC1) 2335 : Builder.CreateBinOp(BOpcode, TruncC1, X); 2336 Constant *TruncC = ConstantInt::get(X->getType(), C->trunc(XWidth)); 2337 Value *And = Builder.CreateAnd(BinOp, TruncC); 2338 return new ZExtInst(And, Ty); 2339 } 2340 2341 // Similar to above: if the mask matches the zext input width, then the 2342 // 'and' can be eliminated, so we can truncate the other variable op: 2343 // and (bo (zext X), Y), C --> zext (bo X, (trunc Y)) 2344 if (isa<Instruction>(BO->getOperand(0)) && 2345 match(BO->getOperand(0), m_OneUse(m_ZExt(m_Value(X)))) && 2346 C->isMask(X->getType()->getScalarSizeInBits())) { 2347 Y = BO->getOperand(1); 2348 Value *TrY = Builder.CreateTrunc(Y, X->getType(), Y->getName() + ".tr"); 2349 Value *NewBO = 2350 Builder.CreateBinOp(BOpcode, X, TrY, BO->getName() + ".narrow"); 2351 return new ZExtInst(NewBO, Ty); 2352 } 2353 // and (bo Y, (zext X)), C --> zext (bo (trunc Y), X) 2354 if (isa<Instruction>(BO->getOperand(1)) && 2355 match(BO->getOperand(1), m_OneUse(m_ZExt(m_Value(X)))) && 2356 C->isMask(X->getType()->getScalarSizeInBits())) { 2357 Y = BO->getOperand(0); 2358 Value *TrY = Builder.CreateTrunc(Y, X->getType(), Y->getName() + ".tr"); 2359 Value *NewBO = 2360 Builder.CreateBinOp(BOpcode, TrY, X, BO->getName() + ".narrow"); 2361 return new ZExtInst(NewBO, Ty); 2362 } 2363 } 2364 2365 // This is intentionally placed after the narrowing transforms for 2366 // efficiency (transform directly to the narrow logic op if possible). 2367 // If the mask is only needed on one incoming arm, push the 'and' op up. 2368 if (match(Op0, m_OneUse(m_Xor(m_Value(X), m_Value(Y)))) || 2369 match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) { 2370 APInt NotAndMask(~(*C)); 2371 BinaryOperator::BinaryOps BinOp = cast<BinaryOperator>(Op0)->getOpcode(); 2372 if (MaskedValueIsZero(X, NotAndMask, 0, &I)) { 2373 // Not masking anything out for the LHS, move mask to RHS. 2374 // and ({x}or X, Y), C --> {x}or X, (and Y, C) 2375 Value *NewRHS = Builder.CreateAnd(Y, Op1, Y->getName() + ".masked"); 2376 return BinaryOperator::Create(BinOp, X, NewRHS); 2377 } 2378 if (!isa<Constant>(Y) && MaskedValueIsZero(Y, NotAndMask, 0, &I)) { 2379 // Not masking anything out for the RHS, move mask to LHS. 2380 // and ({x}or X, Y), C --> {x}or (and X, C), Y 2381 Value *NewLHS = Builder.CreateAnd(X, Op1, X->getName() + ".masked"); 2382 return BinaryOperator::Create(BinOp, NewLHS, Y); 2383 } 2384 } 2385 2386 // When the mask is a power-of-2 constant and op0 is a shifted-power-of-2 2387 // constant, test if the shift amount equals the offset bit index: 2388 // (ShiftC << X) & C --> X == (log2(C) - log2(ShiftC)) ? C : 0 2389 // (ShiftC >> X) & C --> X == (log2(ShiftC) - log2(C)) ? C : 0 2390 if (C->isPowerOf2() && 2391 match(Op0, m_OneUse(m_LogicalShift(m_Power2(ShiftC), m_Value(X))))) { 2392 int Log2ShiftC = ShiftC->exactLogBase2(); 2393 int Log2C = C->exactLogBase2(); 2394 bool IsShiftLeft = 2395 cast<BinaryOperator>(Op0)->getOpcode() == Instruction::Shl; 2396 int BitNum = IsShiftLeft ? Log2C - Log2ShiftC : Log2ShiftC - Log2C; 2397 assert(BitNum >= 0 && "Expected demanded bits to handle impossible mask"); 2398 Value *Cmp = Builder.CreateICmpEQ(X, ConstantInt::get(Ty, BitNum)); 2399 return SelectInst::Create(Cmp, ConstantInt::get(Ty, *C), 2400 ConstantInt::getNullValue(Ty)); 2401 } 2402 2403 Constant *C1, *C2; 2404 const APInt *C3 = C; 2405 Value *X; 2406 if (C3->isPowerOf2()) { 2407 Constant *Log2C3 = ConstantInt::get(Ty, C3->countr_zero()); 2408 if (match(Op0, m_OneUse(m_LShr(m_Shl(m_ImmConstant(C1), m_Value(X)), 2409 m_ImmConstant(C2)))) && 2410 match(C1, m_Power2())) { 2411 Constant *Log2C1 = ConstantExpr::getExactLogBase2(C1); 2412 Constant *LshrC = ConstantExpr::getAdd(C2, Log2C3); 2413 KnownBits KnownLShrc = computeKnownBits(LshrC, 0, nullptr); 2414 if (KnownLShrc.getMaxValue().ult(Width)) { 2415 // iff C1,C3 is pow2 and C2 + cttz(C3) < BitWidth: 2416 // ((C1 << X) >> C2) & C3 -> X == (cttz(C3)+C2-cttz(C1)) ? C3 : 0 2417 Constant *CmpC = ConstantExpr::getSub(LshrC, Log2C1); 2418 Value *Cmp = Builder.CreateICmpEQ(X, CmpC); 2419 return SelectInst::Create(Cmp, ConstantInt::get(Ty, *C3), 2420 ConstantInt::getNullValue(Ty)); 2421 } 2422 } 2423 2424 if (match(Op0, m_OneUse(m_Shl(m_LShr(m_ImmConstant(C1), m_Value(X)), 2425 m_ImmConstant(C2)))) && 2426 match(C1, m_Power2())) { 2427 Constant *Log2C1 = ConstantExpr::getExactLogBase2(C1); 2428 Constant *Cmp = 2429 ConstantExpr::getCompare(ICmpInst::ICMP_ULT, Log2C3, C2); 2430 if (Cmp->isZeroValue()) { 2431 // iff C1,C3 is pow2 and Log2(C3) >= C2: 2432 // ((C1 >> X) << C2) & C3 -> X == (cttz(C1)+C2-cttz(C3)) ? C3 : 0 2433 Constant *ShlC = ConstantExpr::getAdd(C2, Log2C1); 2434 Constant *CmpC = ConstantExpr::getSub(ShlC, Log2C3); 2435 Value *Cmp = Builder.CreateICmpEQ(X, CmpC); 2436 return SelectInst::Create(Cmp, ConstantInt::get(Ty, *C3), 2437 ConstantInt::getNullValue(Ty)); 2438 } 2439 } 2440 } 2441 } 2442 2443 // If we are clearing the sign bit of a floating-point value, convert this to 2444 // fabs, then cast back to integer. 2445 // 2446 // This is a generous interpretation for noimplicitfloat, this is not a true 2447 // floating-point operation. 2448 // 2449 // Assumes any IEEE-represented type has the sign bit in the high bit. 2450 // TODO: Unify with APInt matcher. This version allows undef unlike m_APInt 2451 Value *CastOp; 2452 if (match(Op0, m_BitCast(m_Value(CastOp))) && 2453 match(Op1, m_MaxSignedValue()) && 2454 !Builder.GetInsertBlock()->getParent()->hasFnAttribute( 2455 Attribute::NoImplicitFloat)) { 2456 Type *EltTy = CastOp->getType()->getScalarType(); 2457 if (EltTy->isFloatingPointTy() && EltTy->isIEEE() && 2458 EltTy->getPrimitiveSizeInBits() == 2459 I.getType()->getScalarType()->getPrimitiveSizeInBits()) { 2460 Value *FAbs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, CastOp); 2461 return new BitCastInst(FAbs, I.getType()); 2462 } 2463 } 2464 2465 if (match(&I, m_And(m_OneUse(m_Shl(m_ZExt(m_Value(X)), m_Value(Y))), 2466 m_SignMask())) && 2467 match(Y, m_SpecificInt_ICMP( 2468 ICmpInst::Predicate::ICMP_EQ, 2469 APInt(Ty->getScalarSizeInBits(), 2470 Ty->getScalarSizeInBits() - 2471 X->getType()->getScalarSizeInBits())))) { 2472 auto *SExt = Builder.CreateSExt(X, Ty, X->getName() + ".signext"); 2473 auto *SanitizedSignMask = cast<Constant>(Op1); 2474 // We must be careful with the undef elements of the sign bit mask, however: 2475 // the mask elt can be undef iff the shift amount for that lane was undef, 2476 // otherwise we need to sanitize undef masks to zero. 2477 SanitizedSignMask = Constant::replaceUndefsWith( 2478 SanitizedSignMask, ConstantInt::getNullValue(Ty->getScalarType())); 2479 SanitizedSignMask = 2480 Constant::mergeUndefsWith(SanitizedSignMask, cast<Constant>(Y)); 2481 return BinaryOperator::CreateAnd(SExt, SanitizedSignMask); 2482 } 2483 2484 if (Instruction *Z = narrowMaskedBinOp(I)) 2485 return Z; 2486 2487 if (I.getType()->isIntOrIntVectorTy(1)) { 2488 if (auto *SI0 = dyn_cast<SelectInst>(Op0)) { 2489 if (auto *R = 2490 foldAndOrOfSelectUsingImpliedCond(Op1, *SI0, /* IsAnd */ true)) 2491 return R; 2492 } 2493 if (auto *SI1 = dyn_cast<SelectInst>(Op1)) { 2494 if (auto *R = 2495 foldAndOrOfSelectUsingImpliedCond(Op0, *SI1, /* IsAnd */ true)) 2496 return R; 2497 } 2498 } 2499 2500 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I)) 2501 return FoldedLogic; 2502 2503 if (Instruction *DeMorgan = matchDeMorgansLaws(I, *this)) 2504 return DeMorgan; 2505 2506 { 2507 Value *A, *B, *C; 2508 // A & (A ^ B) --> A & ~B 2509 if (match(Op1, m_OneUse(m_c_Xor(m_Specific(Op0), m_Value(B))))) 2510 return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(B)); 2511 // (A ^ B) & A --> A & ~B 2512 if (match(Op0, m_OneUse(m_c_Xor(m_Specific(Op1), m_Value(B))))) 2513 return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(B)); 2514 2515 // A & ~(A ^ B) --> A & B 2516 if (match(Op1, m_Not(m_c_Xor(m_Specific(Op0), m_Value(B))))) 2517 return BinaryOperator::CreateAnd(Op0, B); 2518 // ~(A ^ B) & A --> A & B 2519 if (match(Op0, m_Not(m_c_Xor(m_Specific(Op1), m_Value(B))))) 2520 return BinaryOperator::CreateAnd(Op1, B); 2521 2522 // (A ^ B) & ((B ^ C) ^ A) -> (A ^ B) & ~C 2523 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) && 2524 match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A)))) { 2525 Value *NotC = Op1->hasOneUse() 2526 ? Builder.CreateNot(C) 2527 : getFreelyInverted(C, C->hasOneUse(), &Builder); 2528 if (NotC != nullptr) 2529 return BinaryOperator::CreateAnd(Op0, NotC); 2530 } 2531 2532 // ((A ^ C) ^ B) & (B ^ A) -> (B ^ A) & ~C 2533 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))) && 2534 match(Op1, m_Xor(m_Specific(B), m_Specific(A)))) { 2535 Value *NotC = Op0->hasOneUse() 2536 ? Builder.CreateNot(C) 2537 : getFreelyInverted(C, C->hasOneUse(), &Builder); 2538 if (NotC != nullptr) 2539 return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(C)); 2540 } 2541 2542 // (A | B) & (~A ^ B) -> A & B 2543 // (A | B) & (B ^ ~A) -> A & B 2544 // (B | A) & (~A ^ B) -> A & B 2545 // (B | A) & (B ^ ~A) -> A & B 2546 if (match(Op1, m_c_Xor(m_Not(m_Value(A)), m_Value(B))) && 2547 match(Op0, m_c_Or(m_Specific(A), m_Specific(B)))) 2548 return BinaryOperator::CreateAnd(A, B); 2549 2550 // (~A ^ B) & (A | B) -> A & B 2551 // (~A ^ B) & (B | A) -> A & B 2552 // (B ^ ~A) & (A | B) -> A & B 2553 // (B ^ ~A) & (B | A) -> A & B 2554 if (match(Op0, m_c_Xor(m_Not(m_Value(A)), m_Value(B))) && 2555 match(Op1, m_c_Or(m_Specific(A), m_Specific(B)))) 2556 return BinaryOperator::CreateAnd(A, B); 2557 2558 // (~A | B) & (A ^ B) -> ~A & B 2559 // (~A | B) & (B ^ A) -> ~A & B 2560 // (B | ~A) & (A ^ B) -> ~A & B 2561 // (B | ~A) & (B ^ A) -> ~A & B 2562 if (match(Op0, m_c_Or(m_Not(m_Value(A)), m_Value(B))) && 2563 match(Op1, m_c_Xor(m_Specific(A), m_Specific(B)))) 2564 return BinaryOperator::CreateAnd(Builder.CreateNot(A), B); 2565 2566 // (A ^ B) & (~A | B) -> ~A & B 2567 // (B ^ A) & (~A | B) -> ~A & B 2568 // (A ^ B) & (B | ~A) -> ~A & B 2569 // (B ^ A) & (B | ~A) -> ~A & B 2570 if (match(Op1, m_c_Or(m_Not(m_Value(A)), m_Value(B))) && 2571 match(Op0, m_c_Xor(m_Specific(A), m_Specific(B)))) 2572 return BinaryOperator::CreateAnd(Builder.CreateNot(A), B); 2573 } 2574 2575 { 2576 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0); 2577 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1); 2578 if (LHS && RHS) 2579 if (Value *Res = foldAndOrOfICmps(LHS, RHS, I, /* IsAnd */ true)) 2580 return replaceInstUsesWith(I, Res); 2581 2582 // TODO: Make this recursive; it's a little tricky because an arbitrary 2583 // number of 'and' instructions might have to be created. 2584 if (LHS && match(Op1, m_OneUse(m_LogicalAnd(m_Value(X), m_Value(Y))))) { 2585 bool IsLogical = isa<SelectInst>(Op1); 2586 // LHS & (X && Y) --> (LHS && X) && Y 2587 if (auto *Cmp = dyn_cast<ICmpInst>(X)) 2588 if (Value *Res = 2589 foldAndOrOfICmps(LHS, Cmp, I, /* IsAnd */ true, IsLogical)) 2590 return replaceInstUsesWith(I, IsLogical 2591 ? Builder.CreateLogicalAnd(Res, Y) 2592 : Builder.CreateAnd(Res, Y)); 2593 // LHS & (X && Y) --> X && (LHS & Y) 2594 if (auto *Cmp = dyn_cast<ICmpInst>(Y)) 2595 if (Value *Res = foldAndOrOfICmps(LHS, Cmp, I, /* IsAnd */ true, 2596 /* IsLogical */ false)) 2597 return replaceInstUsesWith(I, IsLogical 2598 ? Builder.CreateLogicalAnd(X, Res) 2599 : Builder.CreateAnd(X, Res)); 2600 } 2601 if (RHS && match(Op0, m_OneUse(m_LogicalAnd(m_Value(X), m_Value(Y))))) { 2602 bool IsLogical = isa<SelectInst>(Op0); 2603 // (X && Y) & RHS --> (X && RHS) && Y 2604 if (auto *Cmp = dyn_cast<ICmpInst>(X)) 2605 if (Value *Res = 2606 foldAndOrOfICmps(Cmp, RHS, I, /* IsAnd */ true, IsLogical)) 2607 return replaceInstUsesWith(I, IsLogical 2608 ? Builder.CreateLogicalAnd(Res, Y) 2609 : Builder.CreateAnd(Res, Y)); 2610 // (X && Y) & RHS --> X && (Y & RHS) 2611 if (auto *Cmp = dyn_cast<ICmpInst>(Y)) 2612 if (Value *Res = foldAndOrOfICmps(Cmp, RHS, I, /* IsAnd */ true, 2613 /* IsLogical */ false)) 2614 return replaceInstUsesWith(I, IsLogical 2615 ? Builder.CreateLogicalAnd(X, Res) 2616 : Builder.CreateAnd(X, Res)); 2617 } 2618 } 2619 2620 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) 2621 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) 2622 if (Value *Res = foldLogicOfFCmps(LHS, RHS, /*IsAnd*/ true)) 2623 return replaceInstUsesWith(I, Res); 2624 2625 if (Instruction *FoldedFCmps = reassociateFCmps(I, Builder)) 2626 return FoldedFCmps; 2627 2628 if (Instruction *CastedAnd = foldCastedBitwiseLogic(I)) 2629 return CastedAnd; 2630 2631 if (Instruction *Sel = foldBinopOfSextBoolToSelect(I)) 2632 return Sel; 2633 2634 // and(sext(A), B) / and(B, sext(A)) --> A ? B : 0, where A is i1 or <N x i1>. 2635 // TODO: Move this into foldBinopOfSextBoolToSelect as a more generalized fold 2636 // with binop identity constant. But creating a select with non-constant 2637 // arm may not be reversible due to poison semantics. Is that a good 2638 // canonicalization? 2639 Value *A, *B; 2640 if (match(&I, m_c_And(m_OneUse(m_SExt(m_Value(A))), m_Value(B))) && 2641 A->getType()->isIntOrIntVectorTy(1)) 2642 return SelectInst::Create(A, B, Constant::getNullValue(Ty)); 2643 2644 // Similarly, a 'not' of the bool translates to a swap of the select arms: 2645 // ~sext(A) & B / B & ~sext(A) --> A ? 0 : B 2646 if (match(&I, m_c_And(m_Not(m_SExt(m_Value(A))), m_Value(B))) && 2647 A->getType()->isIntOrIntVectorTy(1)) 2648 return SelectInst::Create(A, Constant::getNullValue(Ty), B); 2649 2650 // and(zext(A), B) -> A ? (B & 1) : 0 2651 if (match(&I, m_c_And(m_OneUse(m_ZExt(m_Value(A))), m_Value(B))) && 2652 A->getType()->isIntOrIntVectorTy(1)) 2653 return SelectInst::Create(A, Builder.CreateAnd(B, ConstantInt::get(Ty, 1)), 2654 Constant::getNullValue(Ty)); 2655 2656 // (-1 + A) & B --> A ? 0 : B where A is 0/1. 2657 if (match(&I, m_c_And(m_OneUse(m_Add(m_ZExtOrSelf(m_Value(A)), m_AllOnes())), 2658 m_Value(B)))) { 2659 if (A->getType()->isIntOrIntVectorTy(1)) 2660 return SelectInst::Create(A, Constant::getNullValue(Ty), B); 2661 if (computeKnownBits(A, /* Depth */ 0, &I).countMaxActiveBits() <= 1) { 2662 return SelectInst::Create( 2663 Builder.CreateICmpEQ(A, Constant::getNullValue(A->getType())), B, 2664 Constant::getNullValue(Ty)); 2665 } 2666 } 2667 2668 // (iN X s>> (N-1)) & Y --> (X s< 0) ? Y : 0 -- with optional sext 2669 if (match(&I, m_c_And(m_OneUse(m_SExtOrSelf( 2670 m_AShr(m_Value(X), m_APIntAllowUndef(C)))), 2671 m_Value(Y))) && 2672 *C == X->getType()->getScalarSizeInBits() - 1) { 2673 Value *IsNeg = Builder.CreateIsNeg(X, "isneg"); 2674 return SelectInst::Create(IsNeg, Y, ConstantInt::getNullValue(Ty)); 2675 } 2676 // If there's a 'not' of the shifted value, swap the select operands: 2677 // ~(iN X s>> (N-1)) & Y --> (X s< 0) ? 0 : Y -- with optional sext 2678 if (match(&I, m_c_And(m_OneUse(m_SExtOrSelf( 2679 m_Not(m_AShr(m_Value(X), m_APIntAllowUndef(C))))), 2680 m_Value(Y))) && 2681 *C == X->getType()->getScalarSizeInBits() - 1) { 2682 Value *IsNeg = Builder.CreateIsNeg(X, "isneg"); 2683 return SelectInst::Create(IsNeg, ConstantInt::getNullValue(Ty), Y); 2684 } 2685 2686 // (~x) & y --> ~(x | (~y)) iff that gets rid of inversions 2687 if (sinkNotIntoOtherHandOfLogicalOp(I)) 2688 return &I; 2689 2690 // An and recurrence w/loop invariant step is equivelent to (and start, step) 2691 PHINode *PN = nullptr; 2692 Value *Start = nullptr, *Step = nullptr; 2693 if (matchSimpleRecurrence(&I, PN, Start, Step) && DT.dominates(Step, PN)) 2694 return replaceInstUsesWith(I, Builder.CreateAnd(Start, Step)); 2695 2696 if (Instruction *R = reassociateForUses(I, Builder)) 2697 return R; 2698 2699 if (Instruction *Canonicalized = canonicalizeLogicFirst(I, Builder)) 2700 return Canonicalized; 2701 2702 if (Instruction *Folded = foldLogicOfIsFPClass(I, Op0, Op1)) 2703 return Folded; 2704 2705 if (Instruction *Res = foldBinOpOfDisplacedShifts(I)) 2706 return Res; 2707 2708 if (Instruction *Res = foldBitwiseLogicWithIntrinsics(I, Builder)) 2709 return Res; 2710 2711 return nullptr; 2712 } 2713 2714 Instruction *InstCombinerImpl::matchBSwapOrBitReverse(Instruction &I, 2715 bool MatchBSwaps, 2716 bool MatchBitReversals) { 2717 SmallVector<Instruction *, 4> Insts; 2718 if (!recognizeBSwapOrBitReverseIdiom(&I, MatchBSwaps, MatchBitReversals, 2719 Insts)) 2720 return nullptr; 2721 Instruction *LastInst = Insts.pop_back_val(); 2722 LastInst->removeFromParent(); 2723 2724 for (auto *Inst : Insts) 2725 Worklist.push(Inst); 2726 return LastInst; 2727 } 2728 2729 /// Match UB-safe variants of the funnel shift intrinsic. 2730 static Instruction *matchFunnelShift(Instruction &Or, InstCombinerImpl &IC, 2731 const DominatorTree &DT) { 2732 // TODO: Can we reduce the code duplication between this and the related 2733 // rotate matching code under visitSelect and visitTrunc? 2734 unsigned Width = Or.getType()->getScalarSizeInBits(); 2735 2736 Instruction *Or0, *Or1; 2737 if (!match(Or.getOperand(0), m_Instruction(Or0)) || 2738 !match(Or.getOperand(1), m_Instruction(Or1))) 2739 return nullptr; 2740 2741 bool IsFshl = true; // Sub on LSHR. 2742 SmallVector<Value *, 3> FShiftArgs; 2743 2744 // First, find an or'd pair of opposite shifts: 2745 // or (lshr ShVal0, ShAmt0), (shl ShVal1, ShAmt1) 2746 if (isa<BinaryOperator>(Or0) && isa<BinaryOperator>(Or1)) { 2747 Value *ShVal0, *ShVal1, *ShAmt0, *ShAmt1; 2748 if (!match(Or0, 2749 m_OneUse(m_LogicalShift(m_Value(ShVal0), m_Value(ShAmt0)))) || 2750 !match(Or1, 2751 m_OneUse(m_LogicalShift(m_Value(ShVal1), m_Value(ShAmt1)))) || 2752 Or0->getOpcode() == Or1->getOpcode()) 2753 return nullptr; 2754 2755 // Canonicalize to or(shl(ShVal0, ShAmt0), lshr(ShVal1, ShAmt1)). 2756 if (Or0->getOpcode() == BinaryOperator::LShr) { 2757 std::swap(Or0, Or1); 2758 std::swap(ShVal0, ShVal1); 2759 std::swap(ShAmt0, ShAmt1); 2760 } 2761 assert(Or0->getOpcode() == BinaryOperator::Shl && 2762 Or1->getOpcode() == BinaryOperator::LShr && 2763 "Illegal or(shift,shift) pair"); 2764 2765 // Match the shift amount operands for a funnel shift pattern. This always 2766 // matches a subtraction on the R operand. 2767 auto matchShiftAmount = [&](Value *L, Value *R, unsigned Width) -> Value * { 2768 // Check for constant shift amounts that sum to the bitwidth. 2769 const APInt *LI, *RI; 2770 if (match(L, m_APIntAllowUndef(LI)) && match(R, m_APIntAllowUndef(RI))) 2771 if (LI->ult(Width) && RI->ult(Width) && (*LI + *RI) == Width) 2772 return ConstantInt::get(L->getType(), *LI); 2773 2774 Constant *LC, *RC; 2775 if (match(L, m_Constant(LC)) && match(R, m_Constant(RC)) && 2776 match(L, 2777 m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, APInt(Width, Width))) && 2778 match(R, 2779 m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, APInt(Width, Width))) && 2780 match(ConstantExpr::getAdd(LC, RC), m_SpecificIntAllowUndef(Width))) 2781 return ConstantExpr::mergeUndefsWith(LC, RC); 2782 2783 // (shl ShVal, X) | (lshr ShVal, (Width - x)) iff X < Width. 2784 // We limit this to X < Width in case the backend re-expands the 2785 // intrinsic, and has to reintroduce a shift modulo operation (InstCombine 2786 // might remove it after this fold). This still doesn't guarantee that the 2787 // final codegen will match this original pattern. 2788 if (match(R, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(L))))) { 2789 KnownBits KnownL = IC.computeKnownBits(L, /*Depth*/ 0, &Or); 2790 return KnownL.getMaxValue().ult(Width) ? L : nullptr; 2791 } 2792 2793 // For non-constant cases, the following patterns currently only work for 2794 // rotation patterns. 2795 // TODO: Add general funnel-shift compatible patterns. 2796 if (ShVal0 != ShVal1) 2797 return nullptr; 2798 2799 // For non-constant cases we don't support non-pow2 shift masks. 2800 // TODO: Is it worth matching urem as well? 2801 if (!isPowerOf2_32(Width)) 2802 return nullptr; 2803 2804 // The shift amount may be masked with negation: 2805 // (shl ShVal, (X & (Width - 1))) | (lshr ShVal, ((-X) & (Width - 1))) 2806 Value *X; 2807 unsigned Mask = Width - 1; 2808 if (match(L, m_And(m_Value(X), m_SpecificInt(Mask))) && 2809 match(R, m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask)))) 2810 return X; 2811 2812 // (shl ShVal, X) | (lshr ShVal, ((-X) & (Width - 1))) 2813 if (match(R, m_And(m_Neg(m_Specific(L)), m_SpecificInt(Mask)))) 2814 return L; 2815 2816 // Similar to above, but the shift amount may be extended after masking, 2817 // so return the extended value as the parameter for the intrinsic. 2818 if (match(L, m_ZExt(m_And(m_Value(X), m_SpecificInt(Mask)))) && 2819 match(R, 2820 m_And(m_Neg(m_ZExt(m_And(m_Specific(X), m_SpecificInt(Mask)))), 2821 m_SpecificInt(Mask)))) 2822 return L; 2823 2824 if (match(L, m_ZExt(m_And(m_Value(X), m_SpecificInt(Mask)))) && 2825 match(R, m_ZExt(m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask))))) 2826 return L; 2827 2828 return nullptr; 2829 }; 2830 2831 Value *ShAmt = matchShiftAmount(ShAmt0, ShAmt1, Width); 2832 if (!ShAmt) { 2833 ShAmt = matchShiftAmount(ShAmt1, ShAmt0, Width); 2834 IsFshl = false; // Sub on SHL. 2835 } 2836 if (!ShAmt) 2837 return nullptr; 2838 2839 FShiftArgs = {ShVal0, ShVal1, ShAmt}; 2840 } else if (isa<ZExtInst>(Or0) || isa<ZExtInst>(Or1)) { 2841 // If there are two 'or' instructions concat variables in opposite order: 2842 // 2843 // Slot1 and Slot2 are all zero bits. 2844 // | Slot1 | Low | Slot2 | High | 2845 // LowHigh = or (shl (zext Low), ZextLowShlAmt), (zext High) 2846 // | Slot2 | High | Slot1 | Low | 2847 // HighLow = or (shl (zext High), ZextHighShlAmt), (zext Low) 2848 // 2849 // the latter 'or' can be safely convert to 2850 // -> HighLow = fshl LowHigh, LowHigh, ZextHighShlAmt 2851 // if ZextLowShlAmt + ZextHighShlAmt == Width. 2852 if (!isa<ZExtInst>(Or1)) 2853 std::swap(Or0, Or1); 2854 2855 Value *High, *ZextHigh, *Low; 2856 const APInt *ZextHighShlAmt; 2857 if (!match(Or0, 2858 m_OneUse(m_Shl(m_Value(ZextHigh), m_APInt(ZextHighShlAmt))))) 2859 return nullptr; 2860 2861 if (!match(Or1, m_ZExt(m_Value(Low))) || 2862 !match(ZextHigh, m_ZExt(m_Value(High)))) 2863 return nullptr; 2864 2865 unsigned HighSize = High->getType()->getScalarSizeInBits(); 2866 unsigned LowSize = Low->getType()->getScalarSizeInBits(); 2867 // Make sure High does not overlap with Low and most significant bits of 2868 // High aren't shifted out. 2869 if (ZextHighShlAmt->ult(LowSize) || ZextHighShlAmt->ugt(Width - HighSize)) 2870 return nullptr; 2871 2872 for (User *U : ZextHigh->users()) { 2873 Value *X, *Y; 2874 if (!match(U, m_Or(m_Value(X), m_Value(Y)))) 2875 continue; 2876 2877 if (!isa<ZExtInst>(Y)) 2878 std::swap(X, Y); 2879 2880 const APInt *ZextLowShlAmt; 2881 if (!match(X, m_Shl(m_Specific(Or1), m_APInt(ZextLowShlAmt))) || 2882 !match(Y, m_Specific(ZextHigh)) || !DT.dominates(U, &Or)) 2883 continue; 2884 2885 // HighLow is good concat. If sum of two shifts amount equals to Width, 2886 // LowHigh must also be a good concat. 2887 if (*ZextLowShlAmt + *ZextHighShlAmt != Width) 2888 continue; 2889 2890 // Low must not overlap with High and most significant bits of Low must 2891 // not be shifted out. 2892 assert(ZextLowShlAmt->uge(HighSize) && 2893 ZextLowShlAmt->ule(Width - LowSize) && "Invalid concat"); 2894 2895 FShiftArgs = {U, U, ConstantInt::get(Or0->getType(), *ZextHighShlAmt)}; 2896 break; 2897 } 2898 } 2899 2900 if (FShiftArgs.empty()) 2901 return nullptr; 2902 2903 Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr; 2904 Function *F = Intrinsic::getDeclaration(Or.getModule(), IID, Or.getType()); 2905 return CallInst::Create(F, FShiftArgs); 2906 } 2907 2908 /// Attempt to combine or(zext(x),shl(zext(y),bw/2) concat packing patterns. 2909 static Instruction *matchOrConcat(Instruction &Or, 2910 InstCombiner::BuilderTy &Builder) { 2911 assert(Or.getOpcode() == Instruction::Or && "bswap requires an 'or'"); 2912 Value *Op0 = Or.getOperand(0), *Op1 = Or.getOperand(1); 2913 Type *Ty = Or.getType(); 2914 2915 unsigned Width = Ty->getScalarSizeInBits(); 2916 if ((Width & 1) != 0) 2917 return nullptr; 2918 unsigned HalfWidth = Width / 2; 2919 2920 // Canonicalize zext (lower half) to LHS. 2921 if (!isa<ZExtInst>(Op0)) 2922 std::swap(Op0, Op1); 2923 2924 // Find lower/upper half. 2925 Value *LowerSrc, *ShlVal, *UpperSrc; 2926 const APInt *C; 2927 if (!match(Op0, m_OneUse(m_ZExt(m_Value(LowerSrc)))) || 2928 !match(Op1, m_OneUse(m_Shl(m_Value(ShlVal), m_APInt(C)))) || 2929 !match(ShlVal, m_OneUse(m_ZExt(m_Value(UpperSrc))))) 2930 return nullptr; 2931 if (*C != HalfWidth || LowerSrc->getType() != UpperSrc->getType() || 2932 LowerSrc->getType()->getScalarSizeInBits() != HalfWidth) 2933 return nullptr; 2934 2935 auto ConcatIntrinsicCalls = [&](Intrinsic::ID id, Value *Lo, Value *Hi) { 2936 Value *NewLower = Builder.CreateZExt(Lo, Ty); 2937 Value *NewUpper = Builder.CreateZExt(Hi, Ty); 2938 NewUpper = Builder.CreateShl(NewUpper, HalfWidth); 2939 Value *BinOp = Builder.CreateOr(NewLower, NewUpper); 2940 Function *F = Intrinsic::getDeclaration(Or.getModule(), id, Ty); 2941 return Builder.CreateCall(F, BinOp); 2942 }; 2943 2944 // BSWAP: Push the concat down, swapping the lower/upper sources. 2945 // concat(bswap(x),bswap(y)) -> bswap(concat(x,y)) 2946 Value *LowerBSwap, *UpperBSwap; 2947 if (match(LowerSrc, m_BSwap(m_Value(LowerBSwap))) && 2948 match(UpperSrc, m_BSwap(m_Value(UpperBSwap)))) 2949 return ConcatIntrinsicCalls(Intrinsic::bswap, UpperBSwap, LowerBSwap); 2950 2951 // BITREVERSE: Push the concat down, swapping the lower/upper sources. 2952 // concat(bitreverse(x),bitreverse(y)) -> bitreverse(concat(x,y)) 2953 Value *LowerBRev, *UpperBRev; 2954 if (match(LowerSrc, m_BitReverse(m_Value(LowerBRev))) && 2955 match(UpperSrc, m_BitReverse(m_Value(UpperBRev)))) 2956 return ConcatIntrinsicCalls(Intrinsic::bitreverse, UpperBRev, LowerBRev); 2957 2958 return nullptr; 2959 } 2960 2961 /// If all elements of two constant vectors are 0/-1 and inverses, return true. 2962 static bool areInverseVectorBitmasks(Constant *C1, Constant *C2) { 2963 unsigned NumElts = cast<FixedVectorType>(C1->getType())->getNumElements(); 2964 for (unsigned i = 0; i != NumElts; ++i) { 2965 Constant *EltC1 = C1->getAggregateElement(i); 2966 Constant *EltC2 = C2->getAggregateElement(i); 2967 if (!EltC1 || !EltC2) 2968 return false; 2969 2970 // One element must be all ones, and the other must be all zeros. 2971 if (!((match(EltC1, m_Zero()) && match(EltC2, m_AllOnes())) || 2972 (match(EltC2, m_Zero()) && match(EltC1, m_AllOnes())))) 2973 return false; 2974 } 2975 return true; 2976 } 2977 2978 /// We have an expression of the form (A & C) | (B & D). If A is a scalar or 2979 /// vector composed of all-zeros or all-ones values and is the bitwise 'not' of 2980 /// B, it can be used as the condition operand of a select instruction. 2981 /// We will detect (A & C) | ~(B | D) when the flag ABIsTheSame enabled. 2982 Value *InstCombinerImpl::getSelectCondition(Value *A, Value *B, 2983 bool ABIsTheSame) { 2984 // We may have peeked through bitcasts in the caller. 2985 // Exit immediately if we don't have (vector) integer types. 2986 Type *Ty = A->getType(); 2987 if (!Ty->isIntOrIntVectorTy() || !B->getType()->isIntOrIntVectorTy()) 2988 return nullptr; 2989 2990 // If A is the 'not' operand of B and has enough signbits, we have our answer. 2991 if (ABIsTheSame ? (A == B) : match(B, m_Not(m_Specific(A)))) { 2992 // If these are scalars or vectors of i1, A can be used directly. 2993 if (Ty->isIntOrIntVectorTy(1)) 2994 return A; 2995 2996 // If we look through a vector bitcast, the caller will bitcast the operands 2997 // to match the condition's number of bits (N x i1). 2998 // To make this poison-safe, disallow bitcast from wide element to narrow 2999 // element. That could allow poison in lanes where it was not present in the 3000 // original code. 3001 A = peekThroughBitcast(A); 3002 if (A->getType()->isIntOrIntVectorTy()) { 3003 unsigned NumSignBits = ComputeNumSignBits(A); 3004 if (NumSignBits == A->getType()->getScalarSizeInBits() && 3005 NumSignBits <= Ty->getScalarSizeInBits()) 3006 return Builder.CreateTrunc(A, CmpInst::makeCmpResultType(A->getType())); 3007 } 3008 return nullptr; 3009 } 3010 3011 // TODO: add support for sext and constant case 3012 if (ABIsTheSame) 3013 return nullptr; 3014 3015 // If both operands are constants, see if the constants are inverse bitmasks. 3016 Constant *AConst, *BConst; 3017 if (match(A, m_Constant(AConst)) && match(B, m_Constant(BConst))) 3018 if (AConst == ConstantExpr::getNot(BConst) && 3019 ComputeNumSignBits(A) == Ty->getScalarSizeInBits()) 3020 return Builder.CreateZExtOrTrunc(A, CmpInst::makeCmpResultType(Ty)); 3021 3022 // Look for more complex patterns. The 'not' op may be hidden behind various 3023 // casts. Look through sexts and bitcasts to find the booleans. 3024 Value *Cond; 3025 Value *NotB; 3026 if (match(A, m_SExt(m_Value(Cond))) && 3027 Cond->getType()->isIntOrIntVectorTy(1)) { 3028 // A = sext i1 Cond; B = sext (not (i1 Cond)) 3029 if (match(B, m_SExt(m_Not(m_Specific(Cond))))) 3030 return Cond; 3031 3032 // A = sext i1 Cond; B = not ({bitcast} (sext (i1 Cond))) 3033 // TODO: The one-use checks are unnecessary or misplaced. If the caller 3034 // checked for uses on logic ops/casts, that should be enough to 3035 // make this transform worthwhile. 3036 if (match(B, m_OneUse(m_Not(m_Value(NotB))))) { 3037 NotB = peekThroughBitcast(NotB, true); 3038 if (match(NotB, m_SExt(m_Specific(Cond)))) 3039 return Cond; 3040 } 3041 } 3042 3043 // All scalar (and most vector) possibilities should be handled now. 3044 // Try more matches that only apply to non-splat constant vectors. 3045 if (!Ty->isVectorTy()) 3046 return nullptr; 3047 3048 // If both operands are xor'd with constants using the same sexted boolean 3049 // operand, see if the constants are inverse bitmasks. 3050 // TODO: Use ConstantExpr::getNot()? 3051 if (match(A, (m_Xor(m_SExt(m_Value(Cond)), m_Constant(AConst)))) && 3052 match(B, (m_Xor(m_SExt(m_Specific(Cond)), m_Constant(BConst)))) && 3053 Cond->getType()->isIntOrIntVectorTy(1) && 3054 areInverseVectorBitmasks(AConst, BConst)) { 3055 AConst = ConstantExpr::getTrunc(AConst, CmpInst::makeCmpResultType(Ty)); 3056 return Builder.CreateXor(Cond, AConst); 3057 } 3058 return nullptr; 3059 } 3060 3061 /// We have an expression of the form (A & C) | (B & D). Try to simplify this 3062 /// to "A' ? C : D", where A' is a boolean or vector of booleans. 3063 /// When InvertFalseVal is set to true, we try to match the pattern 3064 /// where we have peeked through a 'not' op and A and B are the same: 3065 /// (A & C) | ~(A | D) --> (A & C) | (~A & ~D) --> A' ? C : ~D 3066 Value *InstCombinerImpl::matchSelectFromAndOr(Value *A, Value *C, Value *B, 3067 Value *D, bool InvertFalseVal) { 3068 // The potential condition of the select may be bitcasted. In that case, look 3069 // through its bitcast and the corresponding bitcast of the 'not' condition. 3070 Type *OrigType = A->getType(); 3071 A = peekThroughBitcast(A, true); 3072 B = peekThroughBitcast(B, true); 3073 if (Value *Cond = getSelectCondition(A, B, InvertFalseVal)) { 3074 // ((bc Cond) & C) | ((bc ~Cond) & D) --> bc (select Cond, (bc C), (bc D)) 3075 // If this is a vector, we may need to cast to match the condition's length. 3076 // The bitcasts will either all exist or all not exist. The builder will 3077 // not create unnecessary casts if the types already match. 3078 Type *SelTy = A->getType(); 3079 if (auto *VecTy = dyn_cast<VectorType>(Cond->getType())) { 3080 // For a fixed or scalable vector get N from <{vscale x} N x iM> 3081 unsigned Elts = VecTy->getElementCount().getKnownMinValue(); 3082 // For a fixed or scalable vector, get the size in bits of N x iM; for a 3083 // scalar this is just M. 3084 unsigned SelEltSize = SelTy->getPrimitiveSizeInBits().getKnownMinValue(); 3085 Type *EltTy = Builder.getIntNTy(SelEltSize / Elts); 3086 SelTy = VectorType::get(EltTy, VecTy->getElementCount()); 3087 } 3088 Value *BitcastC = Builder.CreateBitCast(C, SelTy); 3089 if (InvertFalseVal) 3090 D = Builder.CreateNot(D); 3091 Value *BitcastD = Builder.CreateBitCast(D, SelTy); 3092 Value *Select = Builder.CreateSelect(Cond, BitcastC, BitcastD); 3093 return Builder.CreateBitCast(Select, OrigType); 3094 } 3095 3096 return nullptr; 3097 } 3098 3099 // (icmp eq X, C) | (icmp ult Other, (X - C)) -> (icmp ule Other, (X - (C + 1))) 3100 // (icmp ne X, C) & (icmp uge Other, (X - C)) -> (icmp ugt Other, (X - (C + 1))) 3101 static Value *foldAndOrOfICmpEqConstantAndICmp(ICmpInst *LHS, ICmpInst *RHS, 3102 bool IsAnd, bool IsLogical, 3103 IRBuilderBase &Builder) { 3104 Value *LHS0 = LHS->getOperand(0); 3105 Value *RHS0 = RHS->getOperand(0); 3106 Value *RHS1 = RHS->getOperand(1); 3107 3108 ICmpInst::Predicate LPred = 3109 IsAnd ? LHS->getInversePredicate() : LHS->getPredicate(); 3110 ICmpInst::Predicate RPred = 3111 IsAnd ? RHS->getInversePredicate() : RHS->getPredicate(); 3112 3113 const APInt *CInt; 3114 if (LPred != ICmpInst::ICMP_EQ || 3115 !match(LHS->getOperand(1), m_APIntAllowUndef(CInt)) || 3116 !LHS0->getType()->isIntOrIntVectorTy() || 3117 !(LHS->hasOneUse() || RHS->hasOneUse())) 3118 return nullptr; 3119 3120 auto MatchRHSOp = [LHS0, CInt](const Value *RHSOp) { 3121 return match(RHSOp, 3122 m_Add(m_Specific(LHS0), m_SpecificIntAllowUndef(-*CInt))) || 3123 (CInt->isZero() && RHSOp == LHS0); 3124 }; 3125 3126 Value *Other; 3127 if (RPred == ICmpInst::ICMP_ULT && MatchRHSOp(RHS1)) 3128 Other = RHS0; 3129 else if (RPred == ICmpInst::ICMP_UGT && MatchRHSOp(RHS0)) 3130 Other = RHS1; 3131 else 3132 return nullptr; 3133 3134 if (IsLogical) 3135 Other = Builder.CreateFreeze(Other); 3136 3137 return Builder.CreateICmp( 3138 IsAnd ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE, 3139 Builder.CreateSub(LHS0, ConstantInt::get(LHS0->getType(), *CInt + 1)), 3140 Other); 3141 } 3142 3143 /// Fold (icmp)&(icmp) or (icmp)|(icmp) if possible. 3144 /// If IsLogical is true, then the and/or is in select form and the transform 3145 /// must be poison-safe. 3146 Value *InstCombinerImpl::foldAndOrOfICmps(ICmpInst *LHS, ICmpInst *RHS, 3147 Instruction &I, bool IsAnd, 3148 bool IsLogical) { 3149 const SimplifyQuery Q = SQ.getWithInstruction(&I); 3150 3151 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2) 3152 // Fold (!iszero(A & K1) & !iszero(A & K2)) -> (A & (K1 | K2)) == (K1 | K2) 3153 // if K1 and K2 are a one-bit mask. 3154 if (Value *V = foldAndOrOfICmpsOfAndWithPow2(LHS, RHS, &I, IsAnd, IsLogical)) 3155 return V; 3156 3157 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate(); 3158 Value *LHS0 = LHS->getOperand(0), *RHS0 = RHS->getOperand(0); 3159 Value *LHS1 = LHS->getOperand(1), *RHS1 = RHS->getOperand(1); 3160 const APInt *LHSC = nullptr, *RHSC = nullptr; 3161 match(LHS1, m_APInt(LHSC)); 3162 match(RHS1, m_APInt(RHSC)); 3163 3164 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B) 3165 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B) 3166 if (predicatesFoldable(PredL, PredR)) { 3167 if (LHS0 == RHS1 && LHS1 == RHS0) { 3168 PredL = ICmpInst::getSwappedPredicate(PredL); 3169 std::swap(LHS0, LHS1); 3170 } 3171 if (LHS0 == RHS0 && LHS1 == RHS1) { 3172 unsigned Code = IsAnd ? getICmpCode(PredL) & getICmpCode(PredR) 3173 : getICmpCode(PredL) | getICmpCode(PredR); 3174 bool IsSigned = LHS->isSigned() || RHS->isSigned(); 3175 return getNewICmpValue(Code, IsSigned, LHS0, LHS1, Builder); 3176 } 3177 } 3178 3179 // handle (roughly): 3180 // (icmp ne (A & B), C) | (icmp ne (A & D), E) 3181 // (icmp eq (A & B), C) & (icmp eq (A & D), E) 3182 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, IsAnd, IsLogical, Builder)) 3183 return V; 3184 3185 if (Value *V = 3186 foldAndOrOfICmpEqConstantAndICmp(LHS, RHS, IsAnd, IsLogical, Builder)) 3187 return V; 3188 // We can treat logical like bitwise here, because both operands are used on 3189 // the LHS, and as such poison from both will propagate. 3190 if (Value *V = foldAndOrOfICmpEqConstantAndICmp(RHS, LHS, IsAnd, 3191 /*IsLogical*/ false, Builder)) 3192 return V; 3193 3194 if (Value *V = 3195 foldAndOrOfICmpsWithConstEq(LHS, RHS, IsAnd, IsLogical, Builder, Q)) 3196 return V; 3197 // We can convert this case to bitwise and, because both operands are used 3198 // on the LHS, and as such poison from both will propagate. 3199 if (Value *V = foldAndOrOfICmpsWithConstEq(RHS, LHS, IsAnd, 3200 /*IsLogical*/ false, Builder, Q)) 3201 return V; 3202 3203 if (Value *V = foldIsPowerOf2OrZero(LHS, RHS, IsAnd, Builder)) 3204 return V; 3205 if (Value *V = foldIsPowerOf2OrZero(RHS, LHS, IsAnd, Builder)) 3206 return V; 3207 3208 // TODO: One of these directions is fine with logical and/or, the other could 3209 // be supported by inserting freeze. 3210 if (!IsLogical) { 3211 // E.g. (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n 3212 // E.g. (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n 3213 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/!IsAnd)) 3214 return V; 3215 3216 // E.g. (icmp sgt x, n) | (icmp slt x, 0) --> icmp ugt x, n 3217 // E.g. (icmp slt x, n) & (icmp sge x, 0) --> icmp ult x, n 3218 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/!IsAnd)) 3219 return V; 3220 } 3221 3222 // TODO: Add conjugated or fold, check whether it is safe for logical and/or. 3223 if (IsAnd && !IsLogical) 3224 if (Value *V = foldSignedTruncationCheck(LHS, RHS, I, Builder)) 3225 return V; 3226 3227 if (Value *V = foldIsPowerOf2(LHS, RHS, IsAnd, Builder)) 3228 return V; 3229 3230 if (Value *V = foldPowerOf2AndShiftedMask(LHS, RHS, IsAnd, Builder)) 3231 return V; 3232 3233 // TODO: Verify whether this is safe for logical and/or. 3234 if (!IsLogical) { 3235 if (Value *X = foldUnsignedUnderflowCheck(LHS, RHS, IsAnd, Q, Builder)) 3236 return X; 3237 if (Value *X = foldUnsignedUnderflowCheck(RHS, LHS, IsAnd, Q, Builder)) 3238 return X; 3239 } 3240 3241 if (Value *X = foldEqOfParts(LHS, RHS, IsAnd)) 3242 return X; 3243 3244 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0) 3245 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0) 3246 // TODO: Remove this and below when foldLogOpOfMaskedICmps can handle undefs. 3247 if (!IsLogical && PredL == (IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE) && 3248 PredL == PredR && match(LHS1, m_ZeroInt()) && match(RHS1, m_ZeroInt()) && 3249 LHS0->getType() == RHS0->getType()) { 3250 Value *NewOr = Builder.CreateOr(LHS0, RHS0); 3251 return Builder.CreateICmp(PredL, NewOr, 3252 Constant::getNullValue(NewOr->getType())); 3253 } 3254 3255 // (icmp ne A, -1) | (icmp ne B, -1) --> (icmp ne (A&B), -1) 3256 // (icmp eq A, -1) & (icmp eq B, -1) --> (icmp eq (A&B), -1) 3257 if (!IsLogical && PredL == (IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE) && 3258 PredL == PredR && match(LHS1, m_AllOnes()) && match(RHS1, m_AllOnes()) && 3259 LHS0->getType() == RHS0->getType()) { 3260 Value *NewAnd = Builder.CreateAnd(LHS0, RHS0); 3261 return Builder.CreateICmp(PredL, NewAnd, 3262 Constant::getAllOnesValue(LHS0->getType())); 3263 } 3264 3265 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2). 3266 if (!LHSC || !RHSC) 3267 return nullptr; 3268 3269 // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2 3270 // (trunc x) != C1 | (and x, CA) != C2 -> (and x, CA|CMAX) != C1|C2 3271 // where CMAX is the all ones value for the truncated type, 3272 // iff the lower bits of C2 and CA are zero. 3273 if (PredL == (IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE) && 3274 PredL == PredR && LHS->hasOneUse() && RHS->hasOneUse()) { 3275 Value *V; 3276 const APInt *AndC, *SmallC = nullptr, *BigC = nullptr; 3277 3278 // (trunc x) == C1 & (and x, CA) == C2 3279 // (and x, CA) == C2 & (trunc x) == C1 3280 if (match(RHS0, m_Trunc(m_Value(V))) && 3281 match(LHS0, m_And(m_Specific(V), m_APInt(AndC)))) { 3282 SmallC = RHSC; 3283 BigC = LHSC; 3284 } else if (match(LHS0, m_Trunc(m_Value(V))) && 3285 match(RHS0, m_And(m_Specific(V), m_APInt(AndC)))) { 3286 SmallC = LHSC; 3287 BigC = RHSC; 3288 } 3289 3290 if (SmallC && BigC) { 3291 unsigned BigBitSize = BigC->getBitWidth(); 3292 unsigned SmallBitSize = SmallC->getBitWidth(); 3293 3294 // Check that the low bits are zero. 3295 APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize); 3296 if ((Low & *AndC).isZero() && (Low & *BigC).isZero()) { 3297 Value *NewAnd = Builder.CreateAnd(V, Low | *AndC); 3298 APInt N = SmallC->zext(BigBitSize) | *BigC; 3299 Value *NewVal = ConstantInt::get(NewAnd->getType(), N); 3300 return Builder.CreateICmp(PredL, NewAnd, NewVal); 3301 } 3302 } 3303 } 3304 3305 // Match naive pattern (and its inverted form) for checking if two values 3306 // share same sign. An example of the pattern: 3307 // (icmp slt (X & Y), 0) | (icmp sgt (X | Y), -1) -> (icmp sgt (X ^ Y), -1) 3308 // Inverted form (example): 3309 // (icmp slt (X | Y), 0) & (icmp sgt (X & Y), -1) -> (icmp slt (X ^ Y), 0) 3310 bool TrueIfSignedL, TrueIfSignedR; 3311 if (isSignBitCheck(PredL, *LHSC, TrueIfSignedL) && 3312 isSignBitCheck(PredR, *RHSC, TrueIfSignedR) && 3313 (RHS->hasOneUse() || LHS->hasOneUse())) { 3314 Value *X, *Y; 3315 if (IsAnd) { 3316 if ((TrueIfSignedL && !TrueIfSignedR && 3317 match(LHS0, m_Or(m_Value(X), m_Value(Y))) && 3318 match(RHS0, m_c_And(m_Specific(X), m_Specific(Y)))) || 3319 (!TrueIfSignedL && TrueIfSignedR && 3320 match(LHS0, m_And(m_Value(X), m_Value(Y))) && 3321 match(RHS0, m_c_Or(m_Specific(X), m_Specific(Y))))) { 3322 Value *NewXor = Builder.CreateXor(X, Y); 3323 return Builder.CreateIsNeg(NewXor); 3324 } 3325 } else { 3326 if ((TrueIfSignedL && !TrueIfSignedR && 3327 match(LHS0, m_And(m_Value(X), m_Value(Y))) && 3328 match(RHS0, m_c_Or(m_Specific(X), m_Specific(Y)))) || 3329 (!TrueIfSignedL && TrueIfSignedR && 3330 match(LHS0, m_Or(m_Value(X), m_Value(Y))) && 3331 match(RHS0, m_c_And(m_Specific(X), m_Specific(Y))))) { 3332 Value *NewXor = Builder.CreateXor(X, Y); 3333 return Builder.CreateIsNotNeg(NewXor); 3334 } 3335 } 3336 } 3337 3338 return foldAndOrOfICmpsUsingRanges(LHS, RHS, IsAnd); 3339 } 3340 3341 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches 3342 // here. We should standardize that construct where it is needed or choose some 3343 // other way to ensure that commutated variants of patterns are not missed. 3344 Instruction *InstCombinerImpl::visitOr(BinaryOperator &I) { 3345 if (Value *V = simplifyOrInst(I.getOperand(0), I.getOperand(1), 3346 SQ.getWithInstruction(&I))) 3347 return replaceInstUsesWith(I, V); 3348 3349 if (SimplifyAssociativeOrCommutative(I)) 3350 return &I; 3351 3352 if (Instruction *X = foldVectorBinop(I)) 3353 return X; 3354 3355 if (Instruction *Phi = foldBinopWithPhiOperands(I)) 3356 return Phi; 3357 3358 // See if we can simplify any instructions used by the instruction whose sole 3359 // purpose is to compute bits we don't care about. 3360 if (SimplifyDemandedInstructionBits(I)) 3361 return &I; 3362 3363 // Do this before using distributive laws to catch simple and/or/not patterns. 3364 if (Instruction *Xor = foldOrToXor(I, Builder)) 3365 return Xor; 3366 3367 if (Instruction *X = foldComplexAndOrPatterns(I, Builder)) 3368 return X; 3369 3370 // (A&B)|(A&C) -> A&(B|C) etc 3371 if (Value *V = foldUsingDistributiveLaws(I)) 3372 return replaceInstUsesWith(I, V); 3373 3374 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 3375 Type *Ty = I.getType(); 3376 if (Ty->isIntOrIntVectorTy(1)) { 3377 if (auto *SI0 = dyn_cast<SelectInst>(Op0)) { 3378 if (auto *R = 3379 foldAndOrOfSelectUsingImpliedCond(Op1, *SI0, /* IsAnd */ false)) 3380 return R; 3381 } 3382 if (auto *SI1 = dyn_cast<SelectInst>(Op1)) { 3383 if (auto *R = 3384 foldAndOrOfSelectUsingImpliedCond(Op0, *SI1, /* IsAnd */ false)) 3385 return R; 3386 } 3387 } 3388 3389 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I)) 3390 return FoldedLogic; 3391 3392 if (Instruction *BitOp = matchBSwapOrBitReverse(I, /*MatchBSwaps*/ true, 3393 /*MatchBitReversals*/ true)) 3394 return BitOp; 3395 3396 if (Instruction *Funnel = matchFunnelShift(I, *this, DT)) 3397 return Funnel; 3398 3399 if (Instruction *Concat = matchOrConcat(I, Builder)) 3400 return replaceInstUsesWith(I, Concat); 3401 3402 if (Instruction *R = foldBinOpShiftWithShift(I)) 3403 return R; 3404 3405 if (Instruction *R = tryFoldInstWithCtpopWithNot(&I)) 3406 return R; 3407 3408 Value *X, *Y; 3409 const APInt *CV; 3410 if (match(&I, m_c_Or(m_OneUse(m_Xor(m_Value(X), m_APInt(CV))), m_Value(Y))) && 3411 !CV->isAllOnes() && MaskedValueIsZero(Y, *CV, 0, &I)) { 3412 // (X ^ C) | Y -> (X | Y) ^ C iff Y & C == 0 3413 // The check for a 'not' op is for efficiency (if Y is known zero --> ~X). 3414 Value *Or = Builder.CreateOr(X, Y); 3415 return BinaryOperator::CreateXor(Or, ConstantInt::get(Ty, *CV)); 3416 } 3417 3418 // If the operands have no common bits set: 3419 // or (mul X, Y), X --> add (mul X, Y), X --> mul X, (Y + 1) 3420 if (match(&I, m_c_DisjointOr(m_OneUse(m_Mul(m_Value(X), m_Value(Y))), 3421 m_Deferred(X)))) { 3422 Value *IncrementY = Builder.CreateAdd(Y, ConstantInt::get(Ty, 1)); 3423 return BinaryOperator::CreateMul(X, IncrementY); 3424 } 3425 3426 // X | (X ^ Y) --> X | Y (4 commuted patterns) 3427 if (match(&I, m_c_Or(m_Value(X), m_c_Xor(m_Deferred(X), m_Value(Y))))) 3428 return BinaryOperator::CreateOr(X, Y); 3429 3430 // (A & C) | (B & D) 3431 Value *A, *B, *C, *D; 3432 if (match(Op0, m_And(m_Value(A), m_Value(C))) && 3433 match(Op1, m_And(m_Value(B), m_Value(D)))) { 3434 3435 // (A & C0) | (B & C1) 3436 const APInt *C0, *C1; 3437 if (match(C, m_APInt(C0)) && match(D, m_APInt(C1))) { 3438 Value *X; 3439 if (*C0 == ~*C1) { 3440 // ((X | B) & MaskC) | (B & ~MaskC) -> (X & MaskC) | B 3441 if (match(A, m_c_Or(m_Value(X), m_Specific(B)))) 3442 return BinaryOperator::CreateOr(Builder.CreateAnd(X, *C0), B); 3443 // (A & MaskC) | ((X | A) & ~MaskC) -> (X & ~MaskC) | A 3444 if (match(B, m_c_Or(m_Specific(A), m_Value(X)))) 3445 return BinaryOperator::CreateOr(Builder.CreateAnd(X, *C1), A); 3446 3447 // ((X ^ B) & MaskC) | (B & ~MaskC) -> (X & MaskC) ^ B 3448 if (match(A, m_c_Xor(m_Value(X), m_Specific(B)))) 3449 return BinaryOperator::CreateXor(Builder.CreateAnd(X, *C0), B); 3450 // (A & MaskC) | ((X ^ A) & ~MaskC) -> (X & ~MaskC) ^ A 3451 if (match(B, m_c_Xor(m_Specific(A), m_Value(X)))) 3452 return BinaryOperator::CreateXor(Builder.CreateAnd(X, *C1), A); 3453 } 3454 3455 if ((*C0 & *C1).isZero()) { 3456 // ((X | B) & C0) | (B & C1) --> (X | B) & (C0 | C1) 3457 // iff (C0 & C1) == 0 and (X & ~C0) == 0 3458 if (match(A, m_c_Or(m_Value(X), m_Specific(B))) && 3459 MaskedValueIsZero(X, ~*C0, 0, &I)) { 3460 Constant *C01 = ConstantInt::get(Ty, *C0 | *C1); 3461 return BinaryOperator::CreateAnd(A, C01); 3462 } 3463 // (A & C0) | ((X | A) & C1) --> (X | A) & (C0 | C1) 3464 // iff (C0 & C1) == 0 and (X & ~C1) == 0 3465 if (match(B, m_c_Or(m_Value(X), m_Specific(A))) && 3466 MaskedValueIsZero(X, ~*C1, 0, &I)) { 3467 Constant *C01 = ConstantInt::get(Ty, *C0 | *C1); 3468 return BinaryOperator::CreateAnd(B, C01); 3469 } 3470 // ((X | C2) & C0) | ((X | C3) & C1) --> (X | C2 | C3) & (C0 | C1) 3471 // iff (C0 & C1) == 0 and (C2 & ~C0) == 0 and (C3 & ~C1) == 0. 3472 const APInt *C2, *C3; 3473 if (match(A, m_Or(m_Value(X), m_APInt(C2))) && 3474 match(B, m_Or(m_Specific(X), m_APInt(C3))) && 3475 (*C2 & ~*C0).isZero() && (*C3 & ~*C1).isZero()) { 3476 Value *Or = Builder.CreateOr(X, *C2 | *C3, "bitfield"); 3477 Constant *C01 = ConstantInt::get(Ty, *C0 | *C1); 3478 return BinaryOperator::CreateAnd(Or, C01); 3479 } 3480 } 3481 } 3482 3483 // Don't try to form a select if it's unlikely that we'll get rid of at 3484 // least one of the operands. A select is generally more expensive than the 3485 // 'or' that it is replacing. 3486 if (Op0->hasOneUse() || Op1->hasOneUse()) { 3487 // (Cond & C) | (~Cond & D) -> Cond ? C : D, and commuted variants. 3488 if (Value *V = matchSelectFromAndOr(A, C, B, D)) 3489 return replaceInstUsesWith(I, V); 3490 if (Value *V = matchSelectFromAndOr(A, C, D, B)) 3491 return replaceInstUsesWith(I, V); 3492 if (Value *V = matchSelectFromAndOr(C, A, B, D)) 3493 return replaceInstUsesWith(I, V); 3494 if (Value *V = matchSelectFromAndOr(C, A, D, B)) 3495 return replaceInstUsesWith(I, V); 3496 if (Value *V = matchSelectFromAndOr(B, D, A, C)) 3497 return replaceInstUsesWith(I, V); 3498 if (Value *V = matchSelectFromAndOr(B, D, C, A)) 3499 return replaceInstUsesWith(I, V); 3500 if (Value *V = matchSelectFromAndOr(D, B, A, C)) 3501 return replaceInstUsesWith(I, V); 3502 if (Value *V = matchSelectFromAndOr(D, B, C, A)) 3503 return replaceInstUsesWith(I, V); 3504 } 3505 } 3506 3507 if (match(Op0, m_And(m_Value(A), m_Value(C))) && 3508 match(Op1, m_Not(m_Or(m_Value(B), m_Value(D)))) && 3509 (Op0->hasOneUse() || Op1->hasOneUse())) { 3510 // (Cond & C) | ~(Cond | D) -> Cond ? C : ~D 3511 if (Value *V = matchSelectFromAndOr(A, C, B, D, true)) 3512 return replaceInstUsesWith(I, V); 3513 if (Value *V = matchSelectFromAndOr(A, C, D, B, true)) 3514 return replaceInstUsesWith(I, V); 3515 if (Value *V = matchSelectFromAndOr(C, A, B, D, true)) 3516 return replaceInstUsesWith(I, V); 3517 if (Value *V = matchSelectFromAndOr(C, A, D, B, true)) 3518 return replaceInstUsesWith(I, V); 3519 } 3520 3521 // (A ^ B) | ((B ^ C) ^ A) -> (A ^ B) | C 3522 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) 3523 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A)))) 3524 return BinaryOperator::CreateOr(Op0, C); 3525 3526 // ((A ^ C) ^ B) | (B ^ A) -> (B ^ A) | C 3527 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B)))) 3528 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A)))) 3529 return BinaryOperator::CreateOr(Op1, C); 3530 3531 // ((A & B) ^ C) | B -> C | B 3532 if (match(Op0, m_c_Xor(m_c_And(m_Value(A), m_Specific(Op1)), m_Value(C)))) 3533 return BinaryOperator::CreateOr(C, Op1); 3534 3535 // B | ((A & B) ^ C) -> B | C 3536 if (match(Op1, m_c_Xor(m_c_And(m_Value(A), m_Specific(Op0)), m_Value(C)))) 3537 return BinaryOperator::CreateOr(Op0, C); 3538 3539 // ((B | C) & A) | B -> B | (A & C) 3540 if (match(Op0, m_c_And(m_c_Or(m_Specific(Op1), m_Value(C)), m_Value(A)))) 3541 return BinaryOperator::CreateOr(Op1, Builder.CreateAnd(A, C)); 3542 3543 // B | ((B | C) & A) -> B | (A & C) 3544 if (match(Op1, m_c_And(m_c_Or(m_Specific(Op0), m_Value(C)), m_Value(A)))) 3545 return BinaryOperator::CreateOr(Op0, Builder.CreateAnd(A, C)); 3546 3547 if (Instruction *DeMorgan = matchDeMorgansLaws(I, *this)) 3548 return DeMorgan; 3549 3550 // Canonicalize xor to the RHS. 3551 bool SwappedForXor = false; 3552 if (match(Op0, m_Xor(m_Value(), m_Value()))) { 3553 std::swap(Op0, Op1); 3554 SwappedForXor = true; 3555 } 3556 3557 if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) { 3558 // (A | ?) | (A ^ B) --> (A | ?) | B 3559 // (B | ?) | (A ^ B) --> (B | ?) | A 3560 if (match(Op0, m_c_Or(m_Specific(A), m_Value()))) 3561 return BinaryOperator::CreateOr(Op0, B); 3562 if (match(Op0, m_c_Or(m_Specific(B), m_Value()))) 3563 return BinaryOperator::CreateOr(Op0, A); 3564 3565 // (A & B) | (A ^ B) --> A | B 3566 // (B & A) | (A ^ B) --> A | B 3567 if (match(Op0, m_And(m_Specific(A), m_Specific(B))) || 3568 match(Op0, m_And(m_Specific(B), m_Specific(A)))) 3569 return BinaryOperator::CreateOr(A, B); 3570 3571 // ~A | (A ^ B) --> ~(A & B) 3572 // ~B | (A ^ B) --> ~(A & B) 3573 // The swap above should always make Op0 the 'not'. 3574 if ((Op0->hasOneUse() || Op1->hasOneUse()) && 3575 (match(Op0, m_Not(m_Specific(A))) || match(Op0, m_Not(m_Specific(B))))) 3576 return BinaryOperator::CreateNot(Builder.CreateAnd(A, B)); 3577 3578 // Same as above, but peek through an 'and' to the common operand: 3579 // ~(A & ?) | (A ^ B) --> ~((A & ?) & B) 3580 // ~(B & ?) | (A ^ B) --> ~((B & ?) & A) 3581 Instruction *And; 3582 if ((Op0->hasOneUse() || Op1->hasOneUse()) && 3583 match(Op0, m_Not(m_CombineAnd(m_Instruction(And), 3584 m_c_And(m_Specific(A), m_Value()))))) 3585 return BinaryOperator::CreateNot(Builder.CreateAnd(And, B)); 3586 if ((Op0->hasOneUse() || Op1->hasOneUse()) && 3587 match(Op0, m_Not(m_CombineAnd(m_Instruction(And), 3588 m_c_And(m_Specific(B), m_Value()))))) 3589 return BinaryOperator::CreateNot(Builder.CreateAnd(And, A)); 3590 3591 // (~A | C) | (A ^ B) --> ~(A & B) | C 3592 // (~B | C) | (A ^ B) --> ~(A & B) | C 3593 if (Op0->hasOneUse() && Op1->hasOneUse() && 3594 (match(Op0, m_c_Or(m_Not(m_Specific(A)), m_Value(C))) || 3595 match(Op0, m_c_Or(m_Not(m_Specific(B)), m_Value(C))))) { 3596 Value *Nand = Builder.CreateNot(Builder.CreateAnd(A, B), "nand"); 3597 return BinaryOperator::CreateOr(Nand, C); 3598 } 3599 3600 // A | (~A ^ B) --> ~B | A 3601 // B | (A ^ ~B) --> ~A | B 3602 if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) { 3603 Value *NotB = Builder.CreateNot(B, B->getName() + ".not"); 3604 return BinaryOperator::CreateOr(NotB, Op0); 3605 } 3606 if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) { 3607 Value *NotA = Builder.CreateNot(A, A->getName() + ".not"); 3608 return BinaryOperator::CreateOr(NotA, Op0); 3609 } 3610 } 3611 3612 // A | ~(A | B) -> A | ~B 3613 // A | ~(A ^ B) -> A | ~B 3614 if (match(Op1, m_Not(m_Value(A)))) 3615 if (BinaryOperator *B = dyn_cast<BinaryOperator>(A)) 3616 if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) && 3617 Op1->hasOneUse() && (B->getOpcode() == Instruction::Or || 3618 B->getOpcode() == Instruction::Xor)) { 3619 Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) : 3620 B->getOperand(0); 3621 Value *Not = Builder.CreateNot(NotOp, NotOp->getName() + ".not"); 3622 return BinaryOperator::CreateOr(Not, Op0); 3623 } 3624 3625 if (SwappedForXor) 3626 std::swap(Op0, Op1); 3627 3628 { 3629 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0); 3630 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1); 3631 if (LHS && RHS) 3632 if (Value *Res = foldAndOrOfICmps(LHS, RHS, I, /* IsAnd */ false)) 3633 return replaceInstUsesWith(I, Res); 3634 3635 // TODO: Make this recursive; it's a little tricky because an arbitrary 3636 // number of 'or' instructions might have to be created. 3637 Value *X, *Y; 3638 if (LHS && match(Op1, m_OneUse(m_LogicalOr(m_Value(X), m_Value(Y))))) { 3639 bool IsLogical = isa<SelectInst>(Op1); 3640 // LHS | (X || Y) --> (LHS || X) || Y 3641 if (auto *Cmp = dyn_cast<ICmpInst>(X)) 3642 if (Value *Res = 3643 foldAndOrOfICmps(LHS, Cmp, I, /* IsAnd */ false, IsLogical)) 3644 return replaceInstUsesWith(I, IsLogical 3645 ? Builder.CreateLogicalOr(Res, Y) 3646 : Builder.CreateOr(Res, Y)); 3647 // LHS | (X || Y) --> X || (LHS | Y) 3648 if (auto *Cmp = dyn_cast<ICmpInst>(Y)) 3649 if (Value *Res = foldAndOrOfICmps(LHS, Cmp, I, /* IsAnd */ false, 3650 /* IsLogical */ false)) 3651 return replaceInstUsesWith(I, IsLogical 3652 ? Builder.CreateLogicalOr(X, Res) 3653 : Builder.CreateOr(X, Res)); 3654 } 3655 if (RHS && match(Op0, m_OneUse(m_LogicalOr(m_Value(X), m_Value(Y))))) { 3656 bool IsLogical = isa<SelectInst>(Op0); 3657 // (X || Y) | RHS --> (X || RHS) || Y 3658 if (auto *Cmp = dyn_cast<ICmpInst>(X)) 3659 if (Value *Res = 3660 foldAndOrOfICmps(Cmp, RHS, I, /* IsAnd */ false, IsLogical)) 3661 return replaceInstUsesWith(I, IsLogical 3662 ? Builder.CreateLogicalOr(Res, Y) 3663 : Builder.CreateOr(Res, Y)); 3664 // (X || Y) | RHS --> X || (Y | RHS) 3665 if (auto *Cmp = dyn_cast<ICmpInst>(Y)) 3666 if (Value *Res = foldAndOrOfICmps(Cmp, RHS, I, /* IsAnd */ false, 3667 /* IsLogical */ false)) 3668 return replaceInstUsesWith(I, IsLogical 3669 ? Builder.CreateLogicalOr(X, Res) 3670 : Builder.CreateOr(X, Res)); 3671 } 3672 } 3673 3674 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) 3675 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) 3676 if (Value *Res = foldLogicOfFCmps(LHS, RHS, /*IsAnd*/ false)) 3677 return replaceInstUsesWith(I, Res); 3678 3679 if (Instruction *FoldedFCmps = reassociateFCmps(I, Builder)) 3680 return FoldedFCmps; 3681 3682 if (Instruction *CastedOr = foldCastedBitwiseLogic(I)) 3683 return CastedOr; 3684 3685 if (Instruction *Sel = foldBinopOfSextBoolToSelect(I)) 3686 return Sel; 3687 3688 // or(sext(A), B) / or(B, sext(A)) --> A ? -1 : B, where A is i1 or <N x i1>. 3689 // TODO: Move this into foldBinopOfSextBoolToSelect as a more generalized fold 3690 // with binop identity constant. But creating a select with non-constant 3691 // arm may not be reversible due to poison semantics. Is that a good 3692 // canonicalization? 3693 if (match(&I, m_c_Or(m_OneUse(m_SExt(m_Value(A))), m_Value(B))) && 3694 A->getType()->isIntOrIntVectorTy(1)) 3695 return SelectInst::Create(A, ConstantInt::getAllOnesValue(Ty), B); 3696 3697 // Note: If we've gotten to the point of visiting the outer OR, then the 3698 // inner one couldn't be simplified. If it was a constant, then it won't 3699 // be simplified by a later pass either, so we try swapping the inner/outer 3700 // ORs in the hopes that we'll be able to simplify it this way. 3701 // (X|C) | V --> (X|V) | C 3702 ConstantInt *CI; 3703 if (Op0->hasOneUse() && !match(Op1, m_ConstantInt()) && 3704 match(Op0, m_Or(m_Value(A), m_ConstantInt(CI)))) { 3705 Value *Inner = Builder.CreateOr(A, Op1); 3706 Inner->takeName(Op0); 3707 return BinaryOperator::CreateOr(Inner, CI); 3708 } 3709 3710 // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D)) 3711 // Since this OR statement hasn't been optimized further yet, we hope 3712 // that this transformation will allow the new ORs to be optimized. 3713 { 3714 Value *X = nullptr, *Y = nullptr; 3715 if (Op0->hasOneUse() && Op1->hasOneUse() && 3716 match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) && 3717 match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) { 3718 Value *orTrue = Builder.CreateOr(A, C); 3719 Value *orFalse = Builder.CreateOr(B, D); 3720 return SelectInst::Create(X, orTrue, orFalse); 3721 } 3722 } 3723 3724 // or(ashr(subNSW(Y, X), ScalarSizeInBits(Y) - 1), X) --> X s> Y ? -1 : X. 3725 { 3726 Value *X, *Y; 3727 if (match(&I, m_c_Or(m_OneUse(m_AShr( 3728 m_NSWSub(m_Value(Y), m_Value(X)), 3729 m_SpecificInt(Ty->getScalarSizeInBits() - 1))), 3730 m_Deferred(X)))) { 3731 Value *NewICmpInst = Builder.CreateICmpSGT(X, Y); 3732 Value *AllOnes = ConstantInt::getAllOnesValue(Ty); 3733 return SelectInst::Create(NewICmpInst, AllOnes, X); 3734 } 3735 } 3736 3737 { 3738 // ((A & B) ^ A) | ((A & B) ^ B) -> A ^ B 3739 // (A ^ (A & B)) | (B ^ (A & B)) -> A ^ B 3740 // ((A & B) ^ B) | ((A & B) ^ A) -> A ^ B 3741 // (B ^ (A & B)) | (A ^ (A & B)) -> A ^ B 3742 const auto TryXorOpt = [&](Value *Lhs, Value *Rhs) -> Instruction * { 3743 if (match(Lhs, m_c_Xor(m_And(m_Value(A), m_Value(B)), m_Deferred(A))) && 3744 match(Rhs, 3745 m_c_Xor(m_And(m_Specific(A), m_Specific(B)), m_Deferred(B)))) { 3746 return BinaryOperator::CreateXor(A, B); 3747 } 3748 return nullptr; 3749 }; 3750 3751 if (Instruction *Result = TryXorOpt(Op0, Op1)) 3752 return Result; 3753 if (Instruction *Result = TryXorOpt(Op1, Op0)) 3754 return Result; 3755 } 3756 3757 if (Instruction *V = 3758 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I)) 3759 return V; 3760 3761 CmpInst::Predicate Pred; 3762 Value *Mul, *Ov, *MulIsNotZero, *UMulWithOv; 3763 // Check if the OR weakens the overflow condition for umul.with.overflow by 3764 // treating any non-zero result as overflow. In that case, we overflow if both 3765 // umul.with.overflow operands are != 0, as in that case the result can only 3766 // be 0, iff the multiplication overflows. 3767 if (match(&I, 3768 m_c_Or(m_CombineAnd(m_ExtractValue<1>(m_Value(UMulWithOv)), 3769 m_Value(Ov)), 3770 m_CombineAnd(m_ICmp(Pred, 3771 m_CombineAnd(m_ExtractValue<0>( 3772 m_Deferred(UMulWithOv)), 3773 m_Value(Mul)), 3774 m_ZeroInt()), 3775 m_Value(MulIsNotZero)))) && 3776 (Ov->hasOneUse() || (MulIsNotZero->hasOneUse() && Mul->hasOneUse())) && 3777 Pred == CmpInst::ICMP_NE) { 3778 Value *A, *B; 3779 if (match(UMulWithOv, m_Intrinsic<Intrinsic::umul_with_overflow>( 3780 m_Value(A), m_Value(B)))) { 3781 Value *NotNullA = Builder.CreateIsNotNull(A); 3782 Value *NotNullB = Builder.CreateIsNotNull(B); 3783 return BinaryOperator::CreateAnd(NotNullA, NotNullB); 3784 } 3785 } 3786 3787 /// Res, Overflow = xxx_with_overflow X, C1 3788 /// Try to canonicalize the pattern "Overflow | icmp pred Res, C2" into 3789 /// "Overflow | icmp pred X, C2 +/- C1". 3790 const WithOverflowInst *WO; 3791 const Value *WOV; 3792 const APInt *C1, *C2; 3793 if (match(&I, m_c_Or(m_CombineAnd(m_ExtractValue<1>(m_CombineAnd( 3794 m_WithOverflowInst(WO), m_Value(WOV))), 3795 m_Value(Ov)), 3796 m_OneUse(m_ICmp(Pred, m_ExtractValue<0>(m_Deferred(WOV)), 3797 m_APInt(C2))))) && 3798 (WO->getBinaryOp() == Instruction::Add || 3799 WO->getBinaryOp() == Instruction::Sub) && 3800 (ICmpInst::isEquality(Pred) || 3801 WO->isSigned() == ICmpInst::isSigned(Pred)) && 3802 match(WO->getRHS(), m_APInt(C1))) { 3803 bool Overflow; 3804 APInt NewC = WO->getBinaryOp() == Instruction::Add 3805 ? (ICmpInst::isSigned(Pred) ? C2->ssub_ov(*C1, Overflow) 3806 : C2->usub_ov(*C1, Overflow)) 3807 : (ICmpInst::isSigned(Pred) ? C2->sadd_ov(*C1, Overflow) 3808 : C2->uadd_ov(*C1, Overflow)); 3809 if (!Overflow || ICmpInst::isEquality(Pred)) { 3810 Value *NewCmp = Builder.CreateICmp( 3811 Pred, WO->getLHS(), ConstantInt::get(WO->getLHS()->getType(), NewC)); 3812 return BinaryOperator::CreateOr(Ov, NewCmp); 3813 } 3814 } 3815 3816 // (~x) | y --> ~(x & (~y)) iff that gets rid of inversions 3817 if (sinkNotIntoOtherHandOfLogicalOp(I)) 3818 return &I; 3819 3820 // Improve "get low bit mask up to and including bit X" pattern: 3821 // (1 << X) | ((1 << X) + -1) --> -1 l>> (bitwidth(x) - 1 - X) 3822 if (match(&I, m_c_Or(m_Add(m_Shl(m_One(), m_Value(X)), m_AllOnes()), 3823 m_Shl(m_One(), m_Deferred(X)))) && 3824 match(&I, m_c_Or(m_OneUse(m_Value()), m_Value()))) { 3825 Value *Sub = Builder.CreateSub( 3826 ConstantInt::get(Ty, Ty->getScalarSizeInBits() - 1), X); 3827 return BinaryOperator::CreateLShr(Constant::getAllOnesValue(Ty), Sub); 3828 } 3829 3830 // An or recurrence w/loop invariant step is equivelent to (or start, step) 3831 PHINode *PN = nullptr; 3832 Value *Start = nullptr, *Step = nullptr; 3833 if (matchSimpleRecurrence(&I, PN, Start, Step) && DT.dominates(Step, PN)) 3834 return replaceInstUsesWith(I, Builder.CreateOr(Start, Step)); 3835 3836 // (A & B) | (C | D) or (C | D) | (A & B) 3837 // Can be combined if C or D is of type (A/B & X) 3838 if (match(&I, m_c_Or(m_OneUse(m_And(m_Value(A), m_Value(B))), 3839 m_OneUse(m_Or(m_Value(C), m_Value(D)))))) { 3840 // (A & B) | (C | ?) -> C | (? | (A & B)) 3841 // (A & B) | (C | ?) -> C | (? | (A & B)) 3842 // (A & B) | (C | ?) -> C | (? | (A & B)) 3843 // (A & B) | (C | ?) -> C | (? | (A & B)) 3844 // (C | ?) | (A & B) -> C | (? | (A & B)) 3845 // (C | ?) | (A & B) -> C | (? | (A & B)) 3846 // (C | ?) | (A & B) -> C | (? | (A & B)) 3847 // (C | ?) | (A & B) -> C | (? | (A & B)) 3848 if (match(D, m_OneUse(m_c_And(m_Specific(A), m_Value()))) || 3849 match(D, m_OneUse(m_c_And(m_Specific(B), m_Value())))) 3850 return BinaryOperator::CreateOr( 3851 C, Builder.CreateOr(D, Builder.CreateAnd(A, B))); 3852 // (A & B) | (? | D) -> (? | (A & B)) | D 3853 // (A & B) | (? | D) -> (? | (A & B)) | D 3854 // (A & B) | (? | D) -> (? | (A & B)) | D 3855 // (A & B) | (? | D) -> (? | (A & B)) | D 3856 // (? | D) | (A & B) -> (? | (A & B)) | D 3857 // (? | D) | (A & B) -> (? | (A & B)) | D 3858 // (? | D) | (A & B) -> (? | (A & B)) | D 3859 // (? | D) | (A & B) -> (? | (A & B)) | D 3860 if (match(C, m_OneUse(m_c_And(m_Specific(A), m_Value()))) || 3861 match(C, m_OneUse(m_c_And(m_Specific(B), m_Value())))) 3862 return BinaryOperator::CreateOr( 3863 Builder.CreateOr(C, Builder.CreateAnd(A, B)), D); 3864 } 3865 3866 if (Instruction *R = reassociateForUses(I, Builder)) 3867 return R; 3868 3869 if (Instruction *Canonicalized = canonicalizeLogicFirst(I, Builder)) 3870 return Canonicalized; 3871 3872 if (Instruction *Folded = foldLogicOfIsFPClass(I, Op0, Op1)) 3873 return Folded; 3874 3875 if (Instruction *Res = foldBinOpOfDisplacedShifts(I)) 3876 return Res; 3877 3878 // If we are setting the sign bit of a floating-point value, convert 3879 // this to fneg(fabs), then cast back to integer. 3880 // 3881 // If the result isn't immediately cast back to a float, this will increase 3882 // the number of instructions. This is still probably a better canonical form 3883 // as it enables FP value tracking. 3884 // 3885 // Assumes any IEEE-represented type has the sign bit in the high bit. 3886 // 3887 // This is generous interpretation of noimplicitfloat, this is not a true 3888 // floating-point operation. 3889 Value *CastOp; 3890 if (match(Op0, m_BitCast(m_Value(CastOp))) && match(Op1, m_SignMask()) && 3891 !Builder.GetInsertBlock()->getParent()->hasFnAttribute( 3892 Attribute::NoImplicitFloat)) { 3893 Type *EltTy = CastOp->getType()->getScalarType(); 3894 if (EltTy->isFloatingPointTy() && EltTy->isIEEE() && 3895 EltTy->getPrimitiveSizeInBits() == 3896 I.getType()->getScalarType()->getPrimitiveSizeInBits()) { 3897 Value *FAbs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, CastOp); 3898 Value *FNegFAbs = Builder.CreateFNeg(FAbs); 3899 return new BitCastInst(FNegFAbs, I.getType()); 3900 } 3901 } 3902 3903 // (X & C1) | C2 -> X & (C1 | C2) iff (X & C2) == C2 3904 if (match(Op0, m_OneUse(m_And(m_Value(X), m_APInt(C1)))) && 3905 match(Op1, m_APInt(C2))) { 3906 KnownBits KnownX = computeKnownBits(X, /*Depth*/ 0, &I); 3907 if ((KnownX.One & *C2) == *C2) 3908 return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, *C1 | *C2)); 3909 } 3910 3911 if (Instruction *Res = foldBitwiseLogicWithIntrinsics(I, Builder)) 3912 return Res; 3913 3914 return nullptr; 3915 } 3916 3917 /// A ^ B can be specified using other logic ops in a variety of patterns. We 3918 /// can fold these early and efficiently by morphing an existing instruction. 3919 static Instruction *foldXorToXor(BinaryOperator &I, 3920 InstCombiner::BuilderTy &Builder) { 3921 assert(I.getOpcode() == Instruction::Xor); 3922 Value *Op0 = I.getOperand(0); 3923 Value *Op1 = I.getOperand(1); 3924 Value *A, *B; 3925 3926 // There are 4 commuted variants for each of the basic patterns. 3927 3928 // (A & B) ^ (A | B) -> A ^ B 3929 // (A & B) ^ (B | A) -> A ^ B 3930 // (A | B) ^ (A & B) -> A ^ B 3931 // (A | B) ^ (B & A) -> A ^ B 3932 if (match(&I, m_c_Xor(m_And(m_Value(A), m_Value(B)), 3933 m_c_Or(m_Deferred(A), m_Deferred(B))))) 3934 return BinaryOperator::CreateXor(A, B); 3935 3936 // (A | ~B) ^ (~A | B) -> A ^ B 3937 // (~B | A) ^ (~A | B) -> A ^ B 3938 // (~A | B) ^ (A | ~B) -> A ^ B 3939 // (B | ~A) ^ (A | ~B) -> A ^ B 3940 if (match(&I, m_Xor(m_c_Or(m_Value(A), m_Not(m_Value(B))), 3941 m_c_Or(m_Not(m_Deferred(A)), m_Deferred(B))))) 3942 return BinaryOperator::CreateXor(A, B); 3943 3944 // (A & ~B) ^ (~A & B) -> A ^ B 3945 // (~B & A) ^ (~A & B) -> A ^ B 3946 // (~A & B) ^ (A & ~B) -> A ^ B 3947 // (B & ~A) ^ (A & ~B) -> A ^ B 3948 if (match(&I, m_Xor(m_c_And(m_Value(A), m_Not(m_Value(B))), 3949 m_c_And(m_Not(m_Deferred(A)), m_Deferred(B))))) 3950 return BinaryOperator::CreateXor(A, B); 3951 3952 // For the remaining cases we need to get rid of one of the operands. 3953 if (!Op0->hasOneUse() && !Op1->hasOneUse()) 3954 return nullptr; 3955 3956 // (A | B) ^ ~(A & B) -> ~(A ^ B) 3957 // (A | B) ^ ~(B & A) -> ~(A ^ B) 3958 // (A & B) ^ ~(A | B) -> ~(A ^ B) 3959 // (A & B) ^ ~(B | A) -> ~(A ^ B) 3960 // Complexity sorting ensures the not will be on the right side. 3961 if ((match(Op0, m_Or(m_Value(A), m_Value(B))) && 3962 match(Op1, m_Not(m_c_And(m_Specific(A), m_Specific(B))))) || 3963 (match(Op0, m_And(m_Value(A), m_Value(B))) && 3964 match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B)))))) 3965 return BinaryOperator::CreateNot(Builder.CreateXor(A, B)); 3966 3967 return nullptr; 3968 } 3969 3970 Value *InstCombinerImpl::foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS, 3971 BinaryOperator &I) { 3972 assert(I.getOpcode() == Instruction::Xor && I.getOperand(0) == LHS && 3973 I.getOperand(1) == RHS && "Should be 'xor' with these operands"); 3974 3975 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate(); 3976 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1); 3977 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1); 3978 3979 if (predicatesFoldable(PredL, PredR)) { 3980 if (LHS0 == RHS1 && LHS1 == RHS0) { 3981 std::swap(LHS0, LHS1); 3982 PredL = ICmpInst::getSwappedPredicate(PredL); 3983 } 3984 if (LHS0 == RHS0 && LHS1 == RHS1) { 3985 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B) 3986 unsigned Code = getICmpCode(PredL) ^ getICmpCode(PredR); 3987 bool IsSigned = LHS->isSigned() || RHS->isSigned(); 3988 return getNewICmpValue(Code, IsSigned, LHS0, LHS1, Builder); 3989 } 3990 } 3991 3992 // TODO: This can be generalized to compares of non-signbits using 3993 // decomposeBitTestICmp(). It could be enhanced more by using (something like) 3994 // foldLogOpOfMaskedICmps(). 3995 const APInt *LC, *RC; 3996 if (match(LHS1, m_APInt(LC)) && match(RHS1, m_APInt(RC)) && 3997 LHS0->getType() == RHS0->getType() && 3998 LHS0->getType()->isIntOrIntVectorTy()) { 3999 // Convert xor of signbit tests to signbit test of xor'd values: 4000 // (X > -1) ^ (Y > -1) --> (X ^ Y) < 0 4001 // (X < 0) ^ (Y < 0) --> (X ^ Y) < 0 4002 // (X > -1) ^ (Y < 0) --> (X ^ Y) > -1 4003 // (X < 0) ^ (Y > -1) --> (X ^ Y) > -1 4004 bool TrueIfSignedL, TrueIfSignedR; 4005 if ((LHS->hasOneUse() || RHS->hasOneUse()) && 4006 isSignBitCheck(PredL, *LC, TrueIfSignedL) && 4007 isSignBitCheck(PredR, *RC, TrueIfSignedR)) { 4008 Value *XorLR = Builder.CreateXor(LHS0, RHS0); 4009 return TrueIfSignedL == TrueIfSignedR ? Builder.CreateIsNeg(XorLR) : 4010 Builder.CreateIsNotNeg(XorLR); 4011 } 4012 4013 // Fold (icmp pred1 X, C1) ^ (icmp pred2 X, C2) 4014 // into a single comparison using range-based reasoning. 4015 if (LHS0 == RHS0) { 4016 ConstantRange CR1 = ConstantRange::makeExactICmpRegion(PredL, *LC); 4017 ConstantRange CR2 = ConstantRange::makeExactICmpRegion(PredR, *RC); 4018 auto CRUnion = CR1.exactUnionWith(CR2); 4019 auto CRIntersect = CR1.exactIntersectWith(CR2); 4020 if (CRUnion && CRIntersect) 4021 if (auto CR = CRUnion->exactIntersectWith(CRIntersect->inverse())) { 4022 if (CR->isFullSet()) 4023 return ConstantInt::getTrue(I.getType()); 4024 if (CR->isEmptySet()) 4025 return ConstantInt::getFalse(I.getType()); 4026 4027 CmpInst::Predicate NewPred; 4028 APInt NewC, Offset; 4029 CR->getEquivalentICmp(NewPred, NewC, Offset); 4030 4031 if ((Offset.isZero() && (LHS->hasOneUse() || RHS->hasOneUse())) || 4032 (LHS->hasOneUse() && RHS->hasOneUse())) { 4033 Value *NewV = LHS0; 4034 Type *Ty = LHS0->getType(); 4035 if (!Offset.isZero()) 4036 NewV = Builder.CreateAdd(NewV, ConstantInt::get(Ty, Offset)); 4037 return Builder.CreateICmp(NewPred, NewV, 4038 ConstantInt::get(Ty, NewC)); 4039 } 4040 } 4041 } 4042 } 4043 4044 // Instead of trying to imitate the folds for and/or, decompose this 'xor' 4045 // into those logic ops. That is, try to turn this into an and-of-icmps 4046 // because we have many folds for that pattern. 4047 // 4048 // This is based on a truth table definition of xor: 4049 // X ^ Y --> (X | Y) & !(X & Y) 4050 if (Value *OrICmp = simplifyBinOp(Instruction::Or, LHS, RHS, SQ)) { 4051 // TODO: If OrICmp is true, then the definition of xor simplifies to !(X&Y). 4052 // TODO: If OrICmp is false, the whole thing is false (InstSimplify?). 4053 if (Value *AndICmp = simplifyBinOp(Instruction::And, LHS, RHS, SQ)) { 4054 // TODO: Independently handle cases where the 'and' side is a constant. 4055 ICmpInst *X = nullptr, *Y = nullptr; 4056 if (OrICmp == LHS && AndICmp == RHS) { 4057 // (LHS | RHS) & !(LHS & RHS) --> LHS & !RHS --> X & !Y 4058 X = LHS; 4059 Y = RHS; 4060 } 4061 if (OrICmp == RHS && AndICmp == LHS) { 4062 // !(LHS & RHS) & (LHS | RHS) --> !LHS & RHS --> !Y & X 4063 X = RHS; 4064 Y = LHS; 4065 } 4066 if (X && Y && (Y->hasOneUse() || canFreelyInvertAllUsersOf(Y, &I))) { 4067 // Invert the predicate of 'Y', thus inverting its output. 4068 Y->setPredicate(Y->getInversePredicate()); 4069 // So, are there other uses of Y? 4070 if (!Y->hasOneUse()) { 4071 // We need to adapt other uses of Y though. Get a value that matches 4072 // the original value of Y before inversion. While this increases 4073 // immediate instruction count, we have just ensured that all the 4074 // users are freely-invertible, so that 'not' *will* get folded away. 4075 BuilderTy::InsertPointGuard Guard(Builder); 4076 // Set insertion point to right after the Y. 4077 Builder.SetInsertPoint(Y->getParent(), ++(Y->getIterator())); 4078 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not"); 4079 // Replace all uses of Y (excluding the one in NotY!) with NotY. 4080 Worklist.pushUsersToWorkList(*Y); 4081 Y->replaceUsesWithIf(NotY, 4082 [NotY](Use &U) { return U.getUser() != NotY; }); 4083 } 4084 // All done. 4085 return Builder.CreateAnd(LHS, RHS); 4086 } 4087 } 4088 } 4089 4090 return nullptr; 4091 } 4092 4093 /// If we have a masked merge, in the canonical form of: 4094 /// (assuming that A only has one use.) 4095 /// | A | |B| 4096 /// ((x ^ y) & M) ^ y 4097 /// | D | 4098 /// * If M is inverted: 4099 /// | D | 4100 /// ((x ^ y) & ~M) ^ y 4101 /// We can canonicalize by swapping the final xor operand 4102 /// to eliminate the 'not' of the mask. 4103 /// ((x ^ y) & M) ^ x 4104 /// * If M is a constant, and D has one use, we transform to 'and' / 'or' ops 4105 /// because that shortens the dependency chain and improves analysis: 4106 /// (x & M) | (y & ~M) 4107 static Instruction *visitMaskedMerge(BinaryOperator &I, 4108 InstCombiner::BuilderTy &Builder) { 4109 Value *B, *X, *D; 4110 Value *M; 4111 if (!match(&I, m_c_Xor(m_Value(B), 4112 m_OneUse(m_c_And( 4113 m_CombineAnd(m_c_Xor(m_Deferred(B), m_Value(X)), 4114 m_Value(D)), 4115 m_Value(M)))))) 4116 return nullptr; 4117 4118 Value *NotM; 4119 if (match(M, m_Not(m_Value(NotM)))) { 4120 // De-invert the mask and swap the value in B part. 4121 Value *NewA = Builder.CreateAnd(D, NotM); 4122 return BinaryOperator::CreateXor(NewA, X); 4123 } 4124 4125 Constant *C; 4126 if (D->hasOneUse() && match(M, m_Constant(C))) { 4127 // Propagating undef is unsafe. Clamp undef elements to -1. 4128 Type *EltTy = C->getType()->getScalarType(); 4129 C = Constant::replaceUndefsWith(C, ConstantInt::getAllOnesValue(EltTy)); 4130 // Unfold. 4131 Value *LHS = Builder.CreateAnd(X, C); 4132 Value *NotC = Builder.CreateNot(C); 4133 Value *RHS = Builder.CreateAnd(B, NotC); 4134 return BinaryOperator::CreateOr(LHS, RHS); 4135 } 4136 4137 return nullptr; 4138 } 4139 4140 static Instruction *foldNotXor(BinaryOperator &I, 4141 InstCombiner::BuilderTy &Builder) { 4142 Value *X, *Y; 4143 // FIXME: one-use check is not needed in general, but currently we are unable 4144 // to fold 'not' into 'icmp', if that 'icmp' has multiple uses. (D35182) 4145 if (!match(&I, m_Not(m_OneUse(m_Xor(m_Value(X), m_Value(Y)))))) 4146 return nullptr; 4147 4148 auto hasCommonOperand = [](Value *A, Value *B, Value *C, Value *D) { 4149 return A == C || A == D || B == C || B == D; 4150 }; 4151 4152 Value *A, *B, *C, *D; 4153 // Canonicalize ~((A & B) ^ (A | ?)) -> (A & B) | ~(A | ?) 4154 // 4 commuted variants 4155 if (match(X, m_And(m_Value(A), m_Value(B))) && 4156 match(Y, m_Or(m_Value(C), m_Value(D))) && hasCommonOperand(A, B, C, D)) { 4157 Value *NotY = Builder.CreateNot(Y); 4158 return BinaryOperator::CreateOr(X, NotY); 4159 }; 4160 4161 // Canonicalize ~((A | ?) ^ (A & B)) -> (A & B) | ~(A | ?) 4162 // 4 commuted variants 4163 if (match(Y, m_And(m_Value(A), m_Value(B))) && 4164 match(X, m_Or(m_Value(C), m_Value(D))) && hasCommonOperand(A, B, C, D)) { 4165 Value *NotX = Builder.CreateNot(X); 4166 return BinaryOperator::CreateOr(Y, NotX); 4167 }; 4168 4169 return nullptr; 4170 } 4171 4172 /// Canonicalize a shifty way to code absolute value to the more common pattern 4173 /// that uses negation and select. 4174 static Instruction *canonicalizeAbs(BinaryOperator &Xor, 4175 InstCombiner::BuilderTy &Builder) { 4176 assert(Xor.getOpcode() == Instruction::Xor && "Expected an xor instruction."); 4177 4178 // There are 4 potential commuted variants. Move the 'ashr' candidate to Op1. 4179 // We're relying on the fact that we only do this transform when the shift has 4180 // exactly 2 uses and the add has exactly 1 use (otherwise, we might increase 4181 // instructions). 4182 Value *Op0 = Xor.getOperand(0), *Op1 = Xor.getOperand(1); 4183 if (Op0->hasNUses(2)) 4184 std::swap(Op0, Op1); 4185 4186 Type *Ty = Xor.getType(); 4187 Value *A; 4188 const APInt *ShAmt; 4189 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) && 4190 Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 && 4191 match(Op0, m_OneUse(m_c_Add(m_Specific(A), m_Specific(Op1))))) { 4192 // Op1 = ashr i32 A, 31 ; smear the sign bit 4193 // xor (add A, Op1), Op1 ; add -1 and flip bits if negative 4194 // --> (A < 0) ? -A : A 4195 Value *IsNeg = Builder.CreateIsNeg(A); 4196 // Copy the nuw/nsw flags from the add to the negate. 4197 auto *Add = cast<BinaryOperator>(Op0); 4198 Value *NegA = Builder.CreateNeg(A, "", Add->hasNoUnsignedWrap(), 4199 Add->hasNoSignedWrap()); 4200 return SelectInst::Create(IsNeg, NegA, A); 4201 } 4202 return nullptr; 4203 } 4204 4205 static bool canFreelyInvert(InstCombiner &IC, Value *Op, 4206 Instruction *IgnoredUser) { 4207 auto *I = dyn_cast<Instruction>(Op); 4208 return I && IC.isFreeToInvert(I, /*WillInvertAllUses=*/true) && 4209 IC.canFreelyInvertAllUsersOf(I, IgnoredUser); 4210 } 4211 4212 static Value *freelyInvert(InstCombinerImpl &IC, Value *Op, 4213 Instruction *IgnoredUser) { 4214 auto *I = cast<Instruction>(Op); 4215 IC.Builder.SetInsertPoint(*I->getInsertionPointAfterDef()); 4216 Value *NotOp = IC.Builder.CreateNot(Op, Op->getName() + ".not"); 4217 Op->replaceUsesWithIf(NotOp, 4218 [NotOp](Use &U) { return U.getUser() != NotOp; }); 4219 IC.freelyInvertAllUsersOf(NotOp, IgnoredUser); 4220 return NotOp; 4221 } 4222 4223 // Transform 4224 // z = ~(x &/| y) 4225 // into: 4226 // z = ((~x) |/& (~y)) 4227 // iff both x and y are free to invert and all uses of z can be freely updated. 4228 bool InstCombinerImpl::sinkNotIntoLogicalOp(Instruction &I) { 4229 Value *Op0, *Op1; 4230 if (!match(&I, m_LogicalOp(m_Value(Op0), m_Value(Op1)))) 4231 return false; 4232 4233 // If this logic op has not been simplified yet, just bail out and let that 4234 // happen first. Otherwise, the code below may wrongly invert. 4235 if (Op0 == Op1) 4236 return false; 4237 4238 Instruction::BinaryOps NewOpc = 4239 match(&I, m_LogicalAnd()) ? Instruction::Or : Instruction::And; 4240 bool IsBinaryOp = isa<BinaryOperator>(I); 4241 4242 // Can our users be adapted? 4243 if (!InstCombiner::canFreelyInvertAllUsersOf(&I, /*IgnoredUser=*/nullptr)) 4244 return false; 4245 4246 // And can the operands be adapted? 4247 if (!canFreelyInvert(*this, Op0, &I) || !canFreelyInvert(*this, Op1, &I)) 4248 return false; 4249 4250 Op0 = freelyInvert(*this, Op0, &I); 4251 Op1 = freelyInvert(*this, Op1, &I); 4252 4253 Builder.SetInsertPoint(*I.getInsertionPointAfterDef()); 4254 Value *NewLogicOp; 4255 if (IsBinaryOp) 4256 NewLogicOp = Builder.CreateBinOp(NewOpc, Op0, Op1, I.getName() + ".not"); 4257 else 4258 NewLogicOp = 4259 Builder.CreateLogicalOp(NewOpc, Op0, Op1, I.getName() + ".not"); 4260 4261 replaceInstUsesWith(I, NewLogicOp); 4262 // We can not just create an outer `not`, it will most likely be immediately 4263 // folded back, reconstructing our initial pattern, and causing an 4264 // infinite combine loop, so immediately manually fold it away. 4265 freelyInvertAllUsersOf(NewLogicOp); 4266 return true; 4267 } 4268 4269 // Transform 4270 // z = (~x) &/| y 4271 // into: 4272 // z = ~(x |/& (~y)) 4273 // iff y is free to invert and all uses of z can be freely updated. 4274 bool InstCombinerImpl::sinkNotIntoOtherHandOfLogicalOp(Instruction &I) { 4275 Value *Op0, *Op1; 4276 if (!match(&I, m_LogicalOp(m_Value(Op0), m_Value(Op1)))) 4277 return false; 4278 Instruction::BinaryOps NewOpc = 4279 match(&I, m_LogicalAnd()) ? Instruction::Or : Instruction::And; 4280 bool IsBinaryOp = isa<BinaryOperator>(I); 4281 4282 Value *NotOp0 = nullptr; 4283 Value *NotOp1 = nullptr; 4284 Value **OpToInvert = nullptr; 4285 if (match(Op0, m_Not(m_Value(NotOp0))) && canFreelyInvert(*this, Op1, &I)) { 4286 Op0 = NotOp0; 4287 OpToInvert = &Op1; 4288 } else if (match(Op1, m_Not(m_Value(NotOp1))) && 4289 canFreelyInvert(*this, Op0, &I)) { 4290 Op1 = NotOp1; 4291 OpToInvert = &Op0; 4292 } else 4293 return false; 4294 4295 // And can our users be adapted? 4296 if (!InstCombiner::canFreelyInvertAllUsersOf(&I, /*IgnoredUser=*/nullptr)) 4297 return false; 4298 4299 *OpToInvert = freelyInvert(*this, *OpToInvert, &I); 4300 4301 Builder.SetInsertPoint(*I.getInsertionPointAfterDef()); 4302 Value *NewBinOp; 4303 if (IsBinaryOp) 4304 NewBinOp = Builder.CreateBinOp(NewOpc, Op0, Op1, I.getName() + ".not"); 4305 else 4306 NewBinOp = Builder.CreateLogicalOp(NewOpc, Op0, Op1, I.getName() + ".not"); 4307 replaceInstUsesWith(I, NewBinOp); 4308 // We can not just create an outer `not`, it will most likely be immediately 4309 // folded back, reconstructing our initial pattern, and causing an 4310 // infinite combine loop, so immediately manually fold it away. 4311 freelyInvertAllUsersOf(NewBinOp); 4312 return true; 4313 } 4314 4315 Instruction *InstCombinerImpl::foldNot(BinaryOperator &I) { 4316 Value *NotOp; 4317 if (!match(&I, m_Not(m_Value(NotOp)))) 4318 return nullptr; 4319 4320 // Apply DeMorgan's Law for 'nand' / 'nor' logic with an inverted operand. 4321 // We must eliminate the and/or (one-use) for these transforms to not increase 4322 // the instruction count. 4323 // 4324 // ~(~X & Y) --> (X | ~Y) 4325 // ~(Y & ~X) --> (X | ~Y) 4326 // 4327 // Note: The logical matches do not check for the commuted patterns because 4328 // those are handled via SimplifySelectsFeedingBinaryOp(). 4329 Type *Ty = I.getType(); 4330 Value *X, *Y; 4331 if (match(NotOp, m_OneUse(m_c_And(m_Not(m_Value(X)), m_Value(Y))))) { 4332 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not"); 4333 return BinaryOperator::CreateOr(X, NotY); 4334 } 4335 if (match(NotOp, m_OneUse(m_LogicalAnd(m_Not(m_Value(X)), m_Value(Y))))) { 4336 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not"); 4337 return SelectInst::Create(X, ConstantInt::getTrue(Ty), NotY); 4338 } 4339 4340 // ~(~X | Y) --> (X & ~Y) 4341 // ~(Y | ~X) --> (X & ~Y) 4342 if (match(NotOp, m_OneUse(m_c_Or(m_Not(m_Value(X)), m_Value(Y))))) { 4343 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not"); 4344 return BinaryOperator::CreateAnd(X, NotY); 4345 } 4346 if (match(NotOp, m_OneUse(m_LogicalOr(m_Not(m_Value(X)), m_Value(Y))))) { 4347 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not"); 4348 return SelectInst::Create(X, NotY, ConstantInt::getFalse(Ty)); 4349 } 4350 4351 // Is this a 'not' (~) fed by a binary operator? 4352 BinaryOperator *NotVal; 4353 if (match(NotOp, m_BinOp(NotVal))) { 4354 // ~((-X) | Y) --> (X - 1) & (~Y) 4355 if (match(NotVal, 4356 m_OneUse(m_c_Or(m_OneUse(m_Neg(m_Value(X))), m_Value(Y))))) { 4357 Value *DecX = Builder.CreateAdd(X, ConstantInt::getAllOnesValue(Ty)); 4358 Value *NotY = Builder.CreateNot(Y); 4359 return BinaryOperator::CreateAnd(DecX, NotY); 4360 } 4361 4362 // ~(~X >>s Y) --> (X >>s Y) 4363 if (match(NotVal, m_AShr(m_Not(m_Value(X)), m_Value(Y)))) 4364 return BinaryOperator::CreateAShr(X, Y); 4365 4366 // Treat lshr with non-negative operand as ashr. 4367 // ~(~X >>u Y) --> (X >>s Y) iff X is known negative 4368 if (match(NotVal, m_LShr(m_Not(m_Value(X)), m_Value(Y))) && 4369 isKnownNegative(X, SQ.getWithInstruction(NotVal))) 4370 return BinaryOperator::CreateAShr(X, Y); 4371 4372 // Bit-hack form of a signbit test for iN type: 4373 // ~(X >>s (N - 1)) --> sext i1 (X > -1) to iN 4374 unsigned FullShift = Ty->getScalarSizeInBits() - 1; 4375 if (match(NotVal, m_OneUse(m_AShr(m_Value(X), m_SpecificInt(FullShift))))) { 4376 Value *IsNotNeg = Builder.CreateIsNotNeg(X, "isnotneg"); 4377 return new SExtInst(IsNotNeg, Ty); 4378 } 4379 4380 // If we are inverting a right-shifted constant, we may be able to eliminate 4381 // the 'not' by inverting the constant and using the opposite shift type. 4382 // Canonicalization rules ensure that only a negative constant uses 'ashr', 4383 // but we must check that in case that transform has not fired yet. 4384 4385 // ~(C >>s Y) --> ~C >>u Y (when inverting the replicated sign bits) 4386 Constant *C; 4387 if (match(NotVal, m_AShr(m_Constant(C), m_Value(Y))) && 4388 match(C, m_Negative())) { 4389 // We matched a negative constant, so propagating undef is unsafe. 4390 // Clamp undef elements to -1. 4391 Type *EltTy = Ty->getScalarType(); 4392 C = Constant::replaceUndefsWith(C, ConstantInt::getAllOnesValue(EltTy)); 4393 return BinaryOperator::CreateLShr(ConstantExpr::getNot(C), Y); 4394 } 4395 4396 // ~(C >>u Y) --> ~C >>s Y (when inverting the replicated sign bits) 4397 if (match(NotVal, m_LShr(m_Constant(C), m_Value(Y))) && 4398 match(C, m_NonNegative())) { 4399 // We matched a non-negative constant, so propagating undef is unsafe. 4400 // Clamp undef elements to 0. 4401 Type *EltTy = Ty->getScalarType(); 4402 C = Constant::replaceUndefsWith(C, ConstantInt::getNullValue(EltTy)); 4403 return BinaryOperator::CreateAShr(ConstantExpr::getNot(C), Y); 4404 } 4405 4406 // ~(X + C) --> ~C - X 4407 if (match(NotVal, m_c_Add(m_Value(X), m_ImmConstant(C)))) 4408 return BinaryOperator::CreateSub(ConstantExpr::getNot(C), X); 4409 4410 // ~(X - Y) --> ~X + Y 4411 // FIXME: is it really beneficial to sink the `not` here? 4412 if (match(NotVal, m_Sub(m_Value(X), m_Value(Y)))) 4413 if (isa<Constant>(X) || NotVal->hasOneUse()) 4414 return BinaryOperator::CreateAdd(Builder.CreateNot(X), Y); 4415 4416 // ~(~X + Y) --> X - Y 4417 if (match(NotVal, m_c_Add(m_Not(m_Value(X)), m_Value(Y)))) 4418 return BinaryOperator::CreateWithCopiedFlags(Instruction::Sub, X, Y, 4419 NotVal); 4420 } 4421 4422 // not (cmp A, B) = !cmp A, B 4423 CmpInst::Predicate Pred; 4424 if (match(NotOp, m_Cmp(Pred, m_Value(), m_Value())) && 4425 (NotOp->hasOneUse() || 4426 InstCombiner::canFreelyInvertAllUsersOf(cast<Instruction>(NotOp), 4427 /*IgnoredUser=*/nullptr))) { 4428 cast<CmpInst>(NotOp)->setPredicate(CmpInst::getInversePredicate(Pred)); 4429 freelyInvertAllUsersOf(NotOp); 4430 return &I; 4431 } 4432 4433 // Move a 'not' ahead of casts of a bool to enable logic reduction: 4434 // not (bitcast (sext i1 X)) --> bitcast (sext (not i1 X)) 4435 if (match(NotOp, m_OneUse(m_BitCast(m_OneUse(m_SExt(m_Value(X)))))) && X->getType()->isIntOrIntVectorTy(1)) { 4436 Type *SextTy = cast<BitCastOperator>(NotOp)->getSrcTy(); 4437 Value *NotX = Builder.CreateNot(X); 4438 Value *Sext = Builder.CreateSExt(NotX, SextTy); 4439 return CastInst::CreateBitOrPointerCast(Sext, Ty); 4440 } 4441 4442 if (auto *NotOpI = dyn_cast<Instruction>(NotOp)) 4443 if (sinkNotIntoLogicalOp(*NotOpI)) 4444 return &I; 4445 4446 // Eliminate a bitwise 'not' op of 'not' min/max by inverting the min/max: 4447 // ~min(~X, ~Y) --> max(X, Y) 4448 // ~max(~X, Y) --> min(X, ~Y) 4449 auto *II = dyn_cast<IntrinsicInst>(NotOp); 4450 if (II && II->hasOneUse()) { 4451 if (match(NotOp, m_c_MaxOrMin(m_Not(m_Value(X)), m_Value(Y)))) { 4452 Intrinsic::ID InvID = getInverseMinMaxIntrinsic(II->getIntrinsicID()); 4453 Value *NotY = Builder.CreateNot(Y); 4454 Value *InvMaxMin = Builder.CreateBinaryIntrinsic(InvID, X, NotY); 4455 return replaceInstUsesWith(I, InvMaxMin); 4456 } 4457 4458 if (II->getIntrinsicID() == Intrinsic::is_fpclass) { 4459 ConstantInt *ClassMask = cast<ConstantInt>(II->getArgOperand(1)); 4460 II->setArgOperand( 4461 1, ConstantInt::get(ClassMask->getType(), 4462 ~ClassMask->getZExtValue() & fcAllFlags)); 4463 return replaceInstUsesWith(I, II); 4464 } 4465 } 4466 4467 if (NotOp->hasOneUse()) { 4468 // Pull 'not' into operands of select if both operands are one-use compares 4469 // or one is one-use compare and the other one is a constant. 4470 // Inverting the predicates eliminates the 'not' operation. 4471 // Example: 4472 // not (select ?, (cmp TPred, ?, ?), (cmp FPred, ?, ?) --> 4473 // select ?, (cmp InvTPred, ?, ?), (cmp InvFPred, ?, ?) 4474 // not (select ?, (cmp TPred, ?, ?), true --> 4475 // select ?, (cmp InvTPred, ?, ?), false 4476 if (auto *Sel = dyn_cast<SelectInst>(NotOp)) { 4477 Value *TV = Sel->getTrueValue(); 4478 Value *FV = Sel->getFalseValue(); 4479 auto *CmpT = dyn_cast<CmpInst>(TV); 4480 auto *CmpF = dyn_cast<CmpInst>(FV); 4481 bool InvertibleT = (CmpT && CmpT->hasOneUse()) || isa<Constant>(TV); 4482 bool InvertibleF = (CmpF && CmpF->hasOneUse()) || isa<Constant>(FV); 4483 if (InvertibleT && InvertibleF) { 4484 if (CmpT) 4485 CmpT->setPredicate(CmpT->getInversePredicate()); 4486 else 4487 Sel->setTrueValue(ConstantExpr::getNot(cast<Constant>(TV))); 4488 if (CmpF) 4489 CmpF->setPredicate(CmpF->getInversePredicate()); 4490 else 4491 Sel->setFalseValue(ConstantExpr::getNot(cast<Constant>(FV))); 4492 return replaceInstUsesWith(I, Sel); 4493 } 4494 } 4495 } 4496 4497 if (Instruction *NewXor = foldNotXor(I, Builder)) 4498 return NewXor; 4499 4500 // TODO: Could handle multi-use better by checking if all uses of NotOp (other 4501 // than I) can be inverted. 4502 if (Value *R = getFreelyInverted(NotOp, NotOp->hasOneUse(), &Builder)) 4503 return replaceInstUsesWith(I, R); 4504 4505 return nullptr; 4506 } 4507 4508 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches 4509 // here. We should standardize that construct where it is needed or choose some 4510 // other way to ensure that commutated variants of patterns are not missed. 4511 Instruction *InstCombinerImpl::visitXor(BinaryOperator &I) { 4512 if (Value *V = simplifyXorInst(I.getOperand(0), I.getOperand(1), 4513 SQ.getWithInstruction(&I))) 4514 return replaceInstUsesWith(I, V); 4515 4516 if (SimplifyAssociativeOrCommutative(I)) 4517 return &I; 4518 4519 if (Instruction *X = foldVectorBinop(I)) 4520 return X; 4521 4522 if (Instruction *Phi = foldBinopWithPhiOperands(I)) 4523 return Phi; 4524 4525 if (Instruction *NewXor = foldXorToXor(I, Builder)) 4526 return NewXor; 4527 4528 // (A&B)^(A&C) -> A&(B^C) etc 4529 if (Value *V = foldUsingDistributiveLaws(I)) 4530 return replaceInstUsesWith(I, V); 4531 4532 // See if we can simplify any instructions used by the instruction whose sole 4533 // purpose is to compute bits we don't care about. 4534 if (SimplifyDemandedInstructionBits(I)) 4535 return &I; 4536 4537 if (Instruction *R = foldNot(I)) 4538 return R; 4539 4540 if (Instruction *R = foldBinOpShiftWithShift(I)) 4541 return R; 4542 4543 // Fold (X & M) ^ (Y & ~M) -> (X & M) | (Y & ~M) 4544 // This it a special case in haveNoCommonBitsSet, but the computeKnownBits 4545 // calls in there are unnecessary as SimplifyDemandedInstructionBits should 4546 // have already taken care of those cases. 4547 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 4548 Value *M; 4549 if (match(&I, m_c_Xor(m_c_And(m_Not(m_Value(M)), m_Value()), 4550 m_c_And(m_Deferred(M), m_Value())))) 4551 return BinaryOperator::CreateDisjointOr(Op0, Op1); 4552 4553 if (Instruction *Xor = visitMaskedMerge(I, Builder)) 4554 return Xor; 4555 4556 Value *X, *Y; 4557 Constant *C1; 4558 if (match(Op1, m_Constant(C1))) { 4559 Constant *C2; 4560 4561 if (match(Op0, m_OneUse(m_Or(m_Value(X), m_ImmConstant(C2)))) && 4562 match(C1, m_ImmConstant())) { 4563 // (X | C2) ^ C1 --> (X & ~C2) ^ (C1^C2) 4564 C2 = Constant::replaceUndefsWith( 4565 C2, Constant::getAllOnesValue(C2->getType()->getScalarType())); 4566 Value *And = Builder.CreateAnd( 4567 X, Constant::mergeUndefsWith(ConstantExpr::getNot(C2), C1)); 4568 return BinaryOperator::CreateXor( 4569 And, Constant::mergeUndefsWith(ConstantExpr::getXor(C1, C2), C1)); 4570 } 4571 4572 // Use DeMorgan and reassociation to eliminate a 'not' op. 4573 if (match(Op0, m_OneUse(m_Or(m_Not(m_Value(X)), m_Constant(C2))))) { 4574 // (~X | C2) ^ C1 --> ((X & ~C2) ^ -1) ^ C1 --> (X & ~C2) ^ ~C1 4575 Value *And = Builder.CreateAnd(X, ConstantExpr::getNot(C2)); 4576 return BinaryOperator::CreateXor(And, ConstantExpr::getNot(C1)); 4577 } 4578 if (match(Op0, m_OneUse(m_And(m_Not(m_Value(X)), m_Constant(C2))))) { 4579 // (~X & C2) ^ C1 --> ((X | ~C2) ^ -1) ^ C1 --> (X | ~C2) ^ ~C1 4580 Value *Or = Builder.CreateOr(X, ConstantExpr::getNot(C2)); 4581 return BinaryOperator::CreateXor(Or, ConstantExpr::getNot(C1)); 4582 } 4583 4584 // Convert xor ([trunc] (ashr X, BW-1)), C => 4585 // select(X >s -1, C, ~C) 4586 // The ashr creates "AllZeroOrAllOne's", which then optionally inverses the 4587 // constant depending on whether this input is less than 0. 4588 const APInt *CA; 4589 if (match(Op0, m_OneUse(m_TruncOrSelf( 4590 m_AShr(m_Value(X), m_APIntAllowUndef(CA))))) && 4591 *CA == X->getType()->getScalarSizeInBits() - 1 && 4592 !match(C1, m_AllOnes())) { 4593 assert(!C1->isZeroValue() && "Unexpected xor with 0"); 4594 Value *IsNotNeg = Builder.CreateIsNotNeg(X); 4595 return SelectInst::Create(IsNotNeg, Op1, Builder.CreateNot(Op1)); 4596 } 4597 } 4598 4599 Type *Ty = I.getType(); 4600 { 4601 const APInt *RHSC; 4602 if (match(Op1, m_APInt(RHSC))) { 4603 Value *X; 4604 const APInt *C; 4605 // (C - X) ^ signmaskC --> (C + signmaskC) - X 4606 if (RHSC->isSignMask() && match(Op0, m_Sub(m_APInt(C), m_Value(X)))) 4607 return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C + *RHSC), X); 4608 4609 // (X + C) ^ signmaskC --> X + (C + signmaskC) 4610 if (RHSC->isSignMask() && match(Op0, m_Add(m_Value(X), m_APInt(C)))) 4611 return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C + *RHSC)); 4612 4613 // (X | C) ^ RHSC --> X ^ (C ^ RHSC) iff X & C == 0 4614 if (match(Op0, m_Or(m_Value(X), m_APInt(C))) && 4615 MaskedValueIsZero(X, *C, 0, &I)) 4616 return BinaryOperator::CreateXor(X, ConstantInt::get(Ty, *C ^ *RHSC)); 4617 4618 // When X is a power-of-two or zero and zero input is poison: 4619 // ctlz(i32 X) ^ 31 --> cttz(X) 4620 // cttz(i32 X) ^ 31 --> ctlz(X) 4621 auto *II = dyn_cast<IntrinsicInst>(Op0); 4622 if (II && II->hasOneUse() && *RHSC == Ty->getScalarSizeInBits() - 1) { 4623 Intrinsic::ID IID = II->getIntrinsicID(); 4624 if ((IID == Intrinsic::ctlz || IID == Intrinsic::cttz) && 4625 match(II->getArgOperand(1), m_One()) && 4626 isKnownToBeAPowerOfTwo(II->getArgOperand(0), /*OrZero */ true)) { 4627 IID = (IID == Intrinsic::ctlz) ? Intrinsic::cttz : Intrinsic::ctlz; 4628 Function *F = Intrinsic::getDeclaration(II->getModule(), IID, Ty); 4629 return CallInst::Create(F, {II->getArgOperand(0), Builder.getTrue()}); 4630 } 4631 } 4632 4633 // If RHSC is inverting the remaining bits of shifted X, 4634 // canonicalize to a 'not' before the shift to help SCEV and codegen: 4635 // (X << C) ^ RHSC --> ~X << C 4636 if (match(Op0, m_OneUse(m_Shl(m_Value(X), m_APInt(C)))) && 4637 *RHSC == APInt::getAllOnes(Ty->getScalarSizeInBits()).shl(*C)) { 4638 Value *NotX = Builder.CreateNot(X); 4639 return BinaryOperator::CreateShl(NotX, ConstantInt::get(Ty, *C)); 4640 } 4641 // (X >>u C) ^ RHSC --> ~X >>u C 4642 if (match(Op0, m_OneUse(m_LShr(m_Value(X), m_APInt(C)))) && 4643 *RHSC == APInt::getAllOnes(Ty->getScalarSizeInBits()).lshr(*C)) { 4644 Value *NotX = Builder.CreateNot(X); 4645 return BinaryOperator::CreateLShr(NotX, ConstantInt::get(Ty, *C)); 4646 } 4647 // TODO: We could handle 'ashr' here as well. That would be matching 4648 // a 'not' op and moving it before the shift. Doing that requires 4649 // preventing the inverse fold in canShiftBinOpWithConstantRHS(). 4650 } 4651 4652 // If we are XORing the sign bit of a floating-point value, convert 4653 // this to fneg, then cast back to integer. 4654 // 4655 // This is generous interpretation of noimplicitfloat, this is not a true 4656 // floating-point operation. 4657 // 4658 // Assumes any IEEE-represented type has the sign bit in the high bit. 4659 // TODO: Unify with APInt matcher. This version allows undef unlike m_APInt 4660 Value *CastOp; 4661 if (match(Op0, m_BitCast(m_Value(CastOp))) && match(Op1, m_SignMask()) && 4662 !Builder.GetInsertBlock()->getParent()->hasFnAttribute( 4663 Attribute::NoImplicitFloat)) { 4664 Type *EltTy = CastOp->getType()->getScalarType(); 4665 if (EltTy->isFloatingPointTy() && EltTy->isIEEE() && 4666 EltTy->getPrimitiveSizeInBits() == 4667 I.getType()->getScalarType()->getPrimitiveSizeInBits()) { 4668 Value *FNeg = Builder.CreateFNeg(CastOp); 4669 return new BitCastInst(FNeg, I.getType()); 4670 } 4671 } 4672 } 4673 4674 // FIXME: This should not be limited to scalar (pull into APInt match above). 4675 { 4676 Value *X; 4677 ConstantInt *C1, *C2, *C3; 4678 // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3) 4679 if (match(Op1, m_ConstantInt(C3)) && 4680 match(Op0, m_LShr(m_Xor(m_Value(X), m_ConstantInt(C1)), 4681 m_ConstantInt(C2))) && 4682 Op0->hasOneUse()) { 4683 // fold (C1 >> C2) ^ C3 4684 APInt FoldConst = C1->getValue().lshr(C2->getValue()); 4685 FoldConst ^= C3->getValue(); 4686 // Prepare the two operands. 4687 auto *Opnd0 = Builder.CreateLShr(X, C2); 4688 Opnd0->takeName(Op0); 4689 return BinaryOperator::CreateXor(Opnd0, ConstantInt::get(Ty, FoldConst)); 4690 } 4691 } 4692 4693 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I)) 4694 return FoldedLogic; 4695 4696 // Y ^ (X | Y) --> X & ~Y 4697 // Y ^ (Y | X) --> X & ~Y 4698 if (match(Op1, m_OneUse(m_c_Or(m_Value(X), m_Specific(Op0))))) 4699 return BinaryOperator::CreateAnd(X, Builder.CreateNot(Op0)); 4700 // (X | Y) ^ Y --> X & ~Y 4701 // (Y | X) ^ Y --> X & ~Y 4702 if (match(Op0, m_OneUse(m_c_Or(m_Value(X), m_Specific(Op1))))) 4703 return BinaryOperator::CreateAnd(X, Builder.CreateNot(Op1)); 4704 4705 // Y ^ (X & Y) --> ~X & Y 4706 // Y ^ (Y & X) --> ~X & Y 4707 if (match(Op1, m_OneUse(m_c_And(m_Value(X), m_Specific(Op0))))) 4708 return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(X)); 4709 // (X & Y) ^ Y --> ~X & Y 4710 // (Y & X) ^ Y --> ~X & Y 4711 // Canonical form is (X & C) ^ C; don't touch that. 4712 // TODO: A 'not' op is better for analysis and codegen, but demanded bits must 4713 // be fixed to prefer that (otherwise we get infinite looping). 4714 if (!match(Op1, m_Constant()) && 4715 match(Op0, m_OneUse(m_c_And(m_Value(X), m_Specific(Op1))))) 4716 return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(X)); 4717 4718 Value *A, *B, *C; 4719 // (A ^ B) ^ (A | C) --> (~A & C) ^ B -- There are 4 commuted variants. 4720 if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_Value(A), m_Value(B))), 4721 m_OneUse(m_c_Or(m_Deferred(A), m_Value(C)))))) 4722 return BinaryOperator::CreateXor( 4723 Builder.CreateAnd(Builder.CreateNot(A), C), B); 4724 4725 // (A ^ B) ^ (B | C) --> (~B & C) ^ A -- There are 4 commuted variants. 4726 if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_Value(A), m_Value(B))), 4727 m_OneUse(m_c_Or(m_Deferred(B), m_Value(C)))))) 4728 return BinaryOperator::CreateXor( 4729 Builder.CreateAnd(Builder.CreateNot(B), C), A); 4730 4731 // (A & B) ^ (A ^ B) -> (A | B) 4732 if (match(Op0, m_And(m_Value(A), m_Value(B))) && 4733 match(Op1, m_c_Xor(m_Specific(A), m_Specific(B)))) 4734 return BinaryOperator::CreateOr(A, B); 4735 // (A ^ B) ^ (A & B) -> (A | B) 4736 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) && 4737 match(Op1, m_c_And(m_Specific(A), m_Specific(B)))) 4738 return BinaryOperator::CreateOr(A, B); 4739 4740 // (A & ~B) ^ ~A -> ~(A & B) 4741 // (~B & A) ^ ~A -> ~(A & B) 4742 if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) && 4743 match(Op1, m_Not(m_Specific(A)))) 4744 return BinaryOperator::CreateNot(Builder.CreateAnd(A, B)); 4745 4746 // (~A & B) ^ A --> A | B -- There are 4 commuted variants. 4747 if (match(&I, m_c_Xor(m_c_And(m_Not(m_Value(A)), m_Value(B)), m_Deferred(A)))) 4748 return BinaryOperator::CreateOr(A, B); 4749 4750 // (~A | B) ^ A --> ~(A & B) 4751 if (match(Op0, m_OneUse(m_c_Or(m_Not(m_Specific(Op1)), m_Value(B))))) 4752 return BinaryOperator::CreateNot(Builder.CreateAnd(Op1, B)); 4753 4754 // A ^ (~A | B) --> ~(A & B) 4755 if (match(Op1, m_OneUse(m_c_Or(m_Not(m_Specific(Op0)), m_Value(B))))) 4756 return BinaryOperator::CreateNot(Builder.CreateAnd(Op0, B)); 4757 4758 // (A | B) ^ (A | C) --> (B ^ C) & ~A -- There are 4 commuted variants. 4759 // TODO: Loosen one-use restriction if common operand is a constant. 4760 Value *D; 4761 if (match(Op0, m_OneUse(m_Or(m_Value(A), m_Value(B)))) && 4762 match(Op1, m_OneUse(m_Or(m_Value(C), m_Value(D))))) { 4763 if (B == C || B == D) 4764 std::swap(A, B); 4765 if (A == C) 4766 std::swap(C, D); 4767 if (A == D) { 4768 Value *NotA = Builder.CreateNot(A); 4769 return BinaryOperator::CreateAnd(Builder.CreateXor(B, C), NotA); 4770 } 4771 } 4772 4773 // (A & B) ^ (A | C) --> A ? ~B : C -- There are 4 commuted variants. 4774 if (I.getType()->isIntOrIntVectorTy(1) && 4775 match(Op0, m_OneUse(m_LogicalAnd(m_Value(A), m_Value(B)))) && 4776 match(Op1, m_OneUse(m_LogicalOr(m_Value(C), m_Value(D))))) { 4777 bool NeedFreeze = isa<SelectInst>(Op0) && isa<SelectInst>(Op1) && B == D; 4778 if (B == C || B == D) 4779 std::swap(A, B); 4780 if (A == C) 4781 std::swap(C, D); 4782 if (A == D) { 4783 if (NeedFreeze) 4784 A = Builder.CreateFreeze(A); 4785 Value *NotB = Builder.CreateNot(B); 4786 return SelectInst::Create(A, NotB, C); 4787 } 4788 } 4789 4790 if (auto *LHS = dyn_cast<ICmpInst>(I.getOperand(0))) 4791 if (auto *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) 4792 if (Value *V = foldXorOfICmps(LHS, RHS, I)) 4793 return replaceInstUsesWith(I, V); 4794 4795 if (Instruction *CastedXor = foldCastedBitwiseLogic(I)) 4796 return CastedXor; 4797 4798 if (Instruction *Abs = canonicalizeAbs(I, Builder)) 4799 return Abs; 4800 4801 // Otherwise, if all else failed, try to hoist the xor-by-constant: 4802 // (X ^ C) ^ Y --> (X ^ Y) ^ C 4803 // Just like we do in other places, we completely avoid the fold 4804 // for constantexprs, at least to avoid endless combine loop. 4805 if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_CombineAnd(m_Value(X), 4806 m_Unless(m_ConstantExpr())), 4807 m_ImmConstant(C1))), 4808 m_Value(Y)))) 4809 return BinaryOperator::CreateXor(Builder.CreateXor(X, Y), C1); 4810 4811 if (Instruction *R = reassociateForUses(I, Builder)) 4812 return R; 4813 4814 if (Instruction *Canonicalized = canonicalizeLogicFirst(I, Builder)) 4815 return Canonicalized; 4816 4817 if (Instruction *Folded = foldLogicOfIsFPClass(I, Op0, Op1)) 4818 return Folded; 4819 4820 if (Instruction *Folded = canonicalizeConditionalNegationViaMathToSelect(I)) 4821 return Folded; 4822 4823 if (Instruction *Res = foldBinOpOfDisplacedShifts(I)) 4824 return Res; 4825 4826 if (Instruction *Res = foldBitwiseLogicWithIntrinsics(I, Builder)) 4827 return Res; 4828 4829 return nullptr; 4830 } 4831