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 /// Similar to getICmpCode but for FCmpInst. This encodes a fcmp predicate into 28 /// a four bit mask. 29 static unsigned getFCmpCode(FCmpInst::Predicate CC) { 30 assert(FCmpInst::FCMP_FALSE <= CC && CC <= FCmpInst::FCMP_TRUE && 31 "Unexpected FCmp predicate!"); 32 // Take advantage of the bit pattern of FCmpInst::Predicate here. 33 // U L G E 34 static_assert(FCmpInst::FCMP_FALSE == 0, ""); // 0 0 0 0 35 static_assert(FCmpInst::FCMP_OEQ == 1, ""); // 0 0 0 1 36 static_assert(FCmpInst::FCMP_OGT == 2, ""); // 0 0 1 0 37 static_assert(FCmpInst::FCMP_OGE == 3, ""); // 0 0 1 1 38 static_assert(FCmpInst::FCMP_OLT == 4, ""); // 0 1 0 0 39 static_assert(FCmpInst::FCMP_OLE == 5, ""); // 0 1 0 1 40 static_assert(FCmpInst::FCMP_ONE == 6, ""); // 0 1 1 0 41 static_assert(FCmpInst::FCMP_ORD == 7, ""); // 0 1 1 1 42 static_assert(FCmpInst::FCMP_UNO == 8, ""); // 1 0 0 0 43 static_assert(FCmpInst::FCMP_UEQ == 9, ""); // 1 0 0 1 44 static_assert(FCmpInst::FCMP_UGT == 10, ""); // 1 0 1 0 45 static_assert(FCmpInst::FCMP_UGE == 11, ""); // 1 0 1 1 46 static_assert(FCmpInst::FCMP_ULT == 12, ""); // 1 1 0 0 47 static_assert(FCmpInst::FCMP_ULE == 13, ""); // 1 1 0 1 48 static_assert(FCmpInst::FCMP_UNE == 14, ""); // 1 1 1 0 49 static_assert(FCmpInst::FCMP_TRUE == 15, ""); // 1 1 1 1 50 return CC; 51 } 52 53 /// This is the complement of getICmpCode, which turns an opcode and two 54 /// operands into either a constant true or false, or a brand new ICmp 55 /// instruction. The sign is passed in to determine which kind of predicate to 56 /// use in the new icmp instruction. 57 static Value *getNewICmpValue(unsigned Code, bool Sign, Value *LHS, Value *RHS, 58 InstCombiner::BuilderTy &Builder) { 59 ICmpInst::Predicate NewPred; 60 if (Constant *TorF = getPredForICmpCode(Code, Sign, LHS->getType(), NewPred)) 61 return TorF; 62 return Builder.CreateICmp(NewPred, LHS, RHS); 63 } 64 65 /// This is the complement of getFCmpCode, which turns an opcode and two 66 /// operands into either a FCmp instruction, or a true/false constant. 67 static Value *getFCmpValue(unsigned Code, Value *LHS, Value *RHS, 68 InstCombiner::BuilderTy &Builder) { 69 const auto Pred = static_cast<FCmpInst::Predicate>(Code); 70 assert(FCmpInst::FCMP_FALSE <= Pred && Pred <= FCmpInst::FCMP_TRUE && 71 "Unexpected FCmp predicate!"); 72 if (Pred == FCmpInst::FCMP_FALSE) 73 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 74 if (Pred == FCmpInst::FCMP_TRUE) 75 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1); 76 return Builder.CreateFCmp(Pred, LHS, RHS); 77 } 78 79 /// Transform BITWISE_OP(BSWAP(A),BSWAP(B)) or 80 /// BITWISE_OP(BSWAP(A), Constant) to BSWAP(BITWISE_OP(A, B)) 81 /// \param I Binary operator to transform. 82 /// \return Pointer to node that must replace the original binary operator, or 83 /// null pointer if no transformation was made. 84 static Value *SimplifyBSwap(BinaryOperator &I, 85 InstCombiner::BuilderTy &Builder) { 86 assert(I.isBitwiseLogicOp() && "Unexpected opcode for bswap simplifying"); 87 88 Value *OldLHS = I.getOperand(0); 89 Value *OldRHS = I.getOperand(1); 90 91 Value *NewLHS; 92 if (!match(OldLHS, m_BSwap(m_Value(NewLHS)))) 93 return nullptr; 94 95 Value *NewRHS; 96 const APInt *C; 97 98 if (match(OldRHS, m_BSwap(m_Value(NewRHS)))) { 99 // OP( BSWAP(x), BSWAP(y) ) -> BSWAP( OP(x, y) ) 100 if (!OldLHS->hasOneUse() && !OldRHS->hasOneUse()) 101 return nullptr; 102 // NewRHS initialized by the matcher. 103 } else if (match(OldRHS, m_APInt(C))) { 104 // OP( BSWAP(x), CONSTANT ) -> BSWAP( OP(x, BSWAP(CONSTANT) ) ) 105 if (!OldLHS->hasOneUse()) 106 return nullptr; 107 NewRHS = ConstantInt::get(I.getType(), C->byteSwap()); 108 } else 109 return nullptr; 110 111 Value *BinOp = Builder.CreateBinOp(I.getOpcode(), NewLHS, NewRHS); 112 Function *F = Intrinsic::getDeclaration(I.getModule(), Intrinsic::bswap, 113 I.getType()); 114 return Builder.CreateCall(F, BinOp); 115 } 116 117 /// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise 118 /// (V < Lo || V >= Hi). This method expects that Lo < Hi. IsSigned indicates 119 /// whether to treat V, Lo, and Hi as signed or not. 120 Value *InstCombinerImpl::insertRangeTest(Value *V, const APInt &Lo, 121 const APInt &Hi, bool isSigned, 122 bool Inside) { 123 assert((isSigned ? Lo.slt(Hi) : Lo.ult(Hi)) && 124 "Lo is not < Hi in range emission code!"); 125 126 Type *Ty = V->getType(); 127 128 // V >= Min && V < Hi --> V < Hi 129 // V < Min || V >= Hi --> V >= Hi 130 ICmpInst::Predicate Pred = Inside ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE; 131 if (isSigned ? Lo.isMinSignedValue() : Lo.isMinValue()) { 132 Pred = isSigned ? ICmpInst::getSignedPredicate(Pred) : Pred; 133 return Builder.CreateICmp(Pred, V, ConstantInt::get(Ty, Hi)); 134 } 135 136 // V >= Lo && V < Hi --> V - Lo u< Hi - Lo 137 // V < Lo || V >= Hi --> V - Lo u>= Hi - Lo 138 Value *VMinusLo = 139 Builder.CreateSub(V, ConstantInt::get(Ty, Lo), V->getName() + ".off"); 140 Constant *HiMinusLo = ConstantInt::get(Ty, Hi - Lo); 141 return Builder.CreateICmp(Pred, VMinusLo, HiMinusLo); 142 } 143 144 /// Classify (icmp eq (A & B), C) and (icmp ne (A & B), C) as matching patterns 145 /// that can be simplified. 146 /// One of A and B is considered the mask. The other is the value. This is 147 /// described as the "AMask" or "BMask" part of the enum. If the enum contains 148 /// only "Mask", then both A and B can be considered masks. If A is the mask, 149 /// then it was proven that (A & C) == C. This is trivial if C == A or C == 0. 150 /// If both A and C are constants, this proof is also easy. 151 /// For the following explanations, we assume that A is the mask. 152 /// 153 /// "AllOnes" declares that the comparison is true only if (A & B) == A or all 154 /// bits of A are set in B. 155 /// Example: (icmp eq (A & 3), 3) -> AMask_AllOnes 156 /// 157 /// "AllZeros" declares that the comparison is true only if (A & B) == 0 or all 158 /// bits of A are cleared in B. 159 /// Example: (icmp eq (A & 3), 0) -> Mask_AllZeroes 160 /// 161 /// "Mixed" declares that (A & B) == C and C might or might not contain any 162 /// number of one bits and zero bits. 163 /// Example: (icmp eq (A & 3), 1) -> AMask_Mixed 164 /// 165 /// "Not" means that in above descriptions "==" should be replaced by "!=". 166 /// Example: (icmp ne (A & 3), 3) -> AMask_NotAllOnes 167 /// 168 /// If the mask A contains a single bit, then the following is equivalent: 169 /// (icmp eq (A & B), A) equals (icmp ne (A & B), 0) 170 /// (icmp ne (A & B), A) equals (icmp eq (A & B), 0) 171 enum MaskedICmpType { 172 AMask_AllOnes = 1, 173 AMask_NotAllOnes = 2, 174 BMask_AllOnes = 4, 175 BMask_NotAllOnes = 8, 176 Mask_AllZeros = 16, 177 Mask_NotAllZeros = 32, 178 AMask_Mixed = 64, 179 AMask_NotMixed = 128, 180 BMask_Mixed = 256, 181 BMask_NotMixed = 512 182 }; 183 184 /// Return the set of patterns (from MaskedICmpType) that (icmp SCC (A & B), C) 185 /// satisfies. 186 static unsigned getMaskedICmpType(Value *A, Value *B, Value *C, 187 ICmpInst::Predicate Pred) { 188 ConstantInt *ACst = dyn_cast<ConstantInt>(A); 189 ConstantInt *BCst = dyn_cast<ConstantInt>(B); 190 ConstantInt *CCst = dyn_cast<ConstantInt>(C); 191 bool IsEq = (Pred == ICmpInst::ICMP_EQ); 192 bool IsAPow2 = (ACst && !ACst->isZero() && ACst->getValue().isPowerOf2()); 193 bool IsBPow2 = (BCst && !BCst->isZero() && BCst->getValue().isPowerOf2()); 194 unsigned MaskVal = 0; 195 if (CCst && CCst->isZero()) { 196 // if C is zero, then both A and B qualify as mask 197 MaskVal |= (IsEq ? (Mask_AllZeros | AMask_Mixed | BMask_Mixed) 198 : (Mask_NotAllZeros | AMask_NotMixed | BMask_NotMixed)); 199 if (IsAPow2) 200 MaskVal |= (IsEq ? (AMask_NotAllOnes | AMask_NotMixed) 201 : (AMask_AllOnes | AMask_Mixed)); 202 if (IsBPow2) 203 MaskVal |= (IsEq ? (BMask_NotAllOnes | BMask_NotMixed) 204 : (BMask_AllOnes | BMask_Mixed)); 205 return MaskVal; 206 } 207 208 if (A == C) { 209 MaskVal |= (IsEq ? (AMask_AllOnes | AMask_Mixed) 210 : (AMask_NotAllOnes | AMask_NotMixed)); 211 if (IsAPow2) 212 MaskVal |= (IsEq ? (Mask_NotAllZeros | AMask_NotMixed) 213 : (Mask_AllZeros | AMask_Mixed)); 214 } else if (ACst && CCst && ConstantExpr::getAnd(ACst, CCst) == CCst) { 215 MaskVal |= (IsEq ? AMask_Mixed : AMask_NotMixed); 216 } 217 218 if (B == C) { 219 MaskVal |= (IsEq ? (BMask_AllOnes | BMask_Mixed) 220 : (BMask_NotAllOnes | BMask_NotMixed)); 221 if (IsBPow2) 222 MaskVal |= (IsEq ? (Mask_NotAllZeros | BMask_NotMixed) 223 : (Mask_AllZeros | BMask_Mixed)); 224 } else if (BCst && CCst && ConstantExpr::getAnd(BCst, CCst) == CCst) { 225 MaskVal |= (IsEq ? BMask_Mixed : BMask_NotMixed); 226 } 227 228 return MaskVal; 229 } 230 231 /// Convert an analysis of a masked ICmp into its equivalent if all boolean 232 /// operations had the opposite sense. Since each "NotXXX" flag (recording !=) 233 /// is adjacent to the corresponding normal flag (recording ==), this just 234 /// involves swapping those bits over. 235 static unsigned conjugateICmpMask(unsigned Mask) { 236 unsigned NewMask; 237 NewMask = (Mask & (AMask_AllOnes | BMask_AllOnes | Mask_AllZeros | 238 AMask_Mixed | BMask_Mixed)) 239 << 1; 240 241 NewMask |= (Mask & (AMask_NotAllOnes | BMask_NotAllOnes | Mask_NotAllZeros | 242 AMask_NotMixed | BMask_NotMixed)) 243 >> 1; 244 245 return NewMask; 246 } 247 248 // Adapts the external decomposeBitTestICmp for local use. 249 static bool decomposeBitTestICmp(Value *LHS, Value *RHS, CmpInst::Predicate &Pred, 250 Value *&X, Value *&Y, Value *&Z) { 251 APInt Mask; 252 if (!llvm::decomposeBitTestICmp(LHS, RHS, Pred, X, Mask)) 253 return false; 254 255 Y = ConstantInt::get(X->getType(), Mask); 256 Z = ConstantInt::get(X->getType(), 0); 257 return true; 258 } 259 260 /// Handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E). 261 /// Return the pattern classes (from MaskedICmpType) for the left hand side and 262 /// the right hand side as a pair. 263 /// LHS and RHS are the left hand side and the right hand side ICmps and PredL 264 /// and PredR are their predicates, respectively. 265 static 266 Optional<std::pair<unsigned, unsigned>> 267 getMaskedTypeForICmpPair(Value *&A, Value *&B, Value *&C, 268 Value *&D, Value *&E, ICmpInst *LHS, 269 ICmpInst *RHS, 270 ICmpInst::Predicate &PredL, 271 ICmpInst::Predicate &PredR) { 272 // vectors are not (yet?) supported. Don't support pointers either. 273 if (!LHS->getOperand(0)->getType()->isIntegerTy() || 274 !RHS->getOperand(0)->getType()->isIntegerTy()) 275 return None; 276 277 // Here comes the tricky part: 278 // LHS might be of the form L11 & L12 == X, X == L21 & L22, 279 // and L11 & L12 == L21 & L22. The same goes for RHS. 280 // Now we must find those components L** and R**, that are equal, so 281 // that we can extract the parameters A, B, C, D, and E for the canonical 282 // above. 283 Value *L1 = LHS->getOperand(0); 284 Value *L2 = LHS->getOperand(1); 285 Value *L11, *L12, *L21, *L22; 286 // Check whether the icmp can be decomposed into a bit test. 287 if (decomposeBitTestICmp(L1, L2, PredL, L11, L12, L2)) { 288 L21 = L22 = L1 = nullptr; 289 } else { 290 // Look for ANDs in the LHS icmp. 291 if (!match(L1, m_And(m_Value(L11), m_Value(L12)))) { 292 // Any icmp can be viewed as being trivially masked; if it allows us to 293 // remove one, it's worth it. 294 L11 = L1; 295 L12 = Constant::getAllOnesValue(L1->getType()); 296 } 297 298 if (!match(L2, m_And(m_Value(L21), m_Value(L22)))) { 299 L21 = L2; 300 L22 = Constant::getAllOnesValue(L2->getType()); 301 } 302 } 303 304 // Bail if LHS was a icmp that can't be decomposed into an equality. 305 if (!ICmpInst::isEquality(PredL)) 306 return None; 307 308 Value *R1 = RHS->getOperand(0); 309 Value *R2 = RHS->getOperand(1); 310 Value *R11, *R12; 311 bool Ok = false; 312 if (decomposeBitTestICmp(R1, R2, PredR, R11, R12, R2)) { 313 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) { 314 A = R11; 315 D = R12; 316 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) { 317 A = R12; 318 D = R11; 319 } else { 320 return None; 321 } 322 E = R2; 323 R1 = nullptr; 324 Ok = true; 325 } else { 326 if (!match(R1, m_And(m_Value(R11), m_Value(R12)))) { 327 // As before, model no mask as a trivial mask if it'll let us do an 328 // optimization. 329 R11 = R1; 330 R12 = Constant::getAllOnesValue(R1->getType()); 331 } 332 333 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) { 334 A = R11; 335 D = R12; 336 E = R2; 337 Ok = true; 338 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) { 339 A = R12; 340 D = R11; 341 E = R2; 342 Ok = true; 343 } 344 } 345 346 // Bail if RHS was a icmp that can't be decomposed into an equality. 347 if (!ICmpInst::isEquality(PredR)) 348 return None; 349 350 // Look for ANDs on the right side of the RHS icmp. 351 if (!Ok) { 352 if (!match(R2, m_And(m_Value(R11), m_Value(R12)))) { 353 R11 = R2; 354 R12 = Constant::getAllOnesValue(R2->getType()); 355 } 356 357 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) { 358 A = R11; 359 D = R12; 360 E = R1; 361 Ok = true; 362 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) { 363 A = R12; 364 D = R11; 365 E = R1; 366 Ok = true; 367 } else { 368 return None; 369 } 370 } 371 if (!Ok) 372 return None; 373 374 if (L11 == A) { 375 B = L12; 376 C = L2; 377 } else if (L12 == A) { 378 B = L11; 379 C = L2; 380 } else if (L21 == A) { 381 B = L22; 382 C = L1; 383 } else if (L22 == A) { 384 B = L21; 385 C = L1; 386 } 387 388 unsigned LeftType = getMaskedICmpType(A, B, C, PredL); 389 unsigned RightType = getMaskedICmpType(A, D, E, PredR); 390 return Optional<std::pair<unsigned, unsigned>>(std::make_pair(LeftType, RightType)); 391 } 392 393 /// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E) into a single 394 /// (icmp(A & X) ==/!= Y), where the left-hand side is of type Mask_NotAllZeros 395 /// and the right hand side is of type BMask_Mixed. For example, 396 /// (icmp (A & 12) != 0) & (icmp (A & 15) == 8) -> (icmp (A & 15) == 8). 397 static Value *foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed( 398 ICmpInst *LHS, ICmpInst *RHS, bool IsAnd, Value *A, Value *B, Value *C, 399 Value *D, Value *E, ICmpInst::Predicate PredL, ICmpInst::Predicate PredR, 400 InstCombiner::BuilderTy &Builder) { 401 // We are given the canonical form: 402 // (icmp ne (A & B), 0) & (icmp eq (A & D), E). 403 // where D & E == E. 404 // 405 // If IsAnd is false, we get it in negated form: 406 // (icmp eq (A & B), 0) | (icmp ne (A & D), E) -> 407 // !((icmp ne (A & B), 0) & (icmp eq (A & D), E)). 408 // 409 // We currently handle the case of B, C, D, E are constant. 410 // 411 ConstantInt *BCst, *CCst, *DCst, *ECst; 412 if (!match(B, m_ConstantInt(BCst)) || !match(C, m_ConstantInt(CCst)) || 413 !match(D, m_ConstantInt(DCst)) || !match(E, m_ConstantInt(ECst))) 414 return nullptr; 415 416 ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE; 417 418 // Update E to the canonical form when D is a power of two and RHS is 419 // canonicalized as, 420 // (icmp ne (A & D), 0) -> (icmp eq (A & D), D) or 421 // (icmp ne (A & D), D) -> (icmp eq (A & D), 0). 422 if (PredR != NewCC) 423 ECst = cast<ConstantInt>(ConstantExpr::getXor(DCst, ECst)); 424 425 // If B or D is zero, skip because if LHS or RHS can be trivially folded by 426 // other folding rules and this pattern won't apply any more. 427 if (BCst->getValue() == 0 || DCst->getValue() == 0) 428 return nullptr; 429 430 // If B and D don't intersect, ie. (B & D) == 0, no folding because we can't 431 // deduce anything from it. 432 // For example, 433 // (icmp ne (A & 12), 0) & (icmp eq (A & 3), 1) -> no folding. 434 if ((BCst->getValue() & DCst->getValue()) == 0) 435 return nullptr; 436 437 // If the following two conditions are met: 438 // 439 // 1. mask B covers only a single bit that's not covered by mask D, that is, 440 // (B & (B ^ D)) is a power of 2 (in other words, B minus the intersection of 441 // B and D has only one bit set) and, 442 // 443 // 2. RHS (and E) indicates that the rest of B's bits are zero (in other 444 // words, the intersection of B and D is zero), that is, ((B & D) & E) == 0 445 // 446 // then that single bit in B must be one and thus the whole expression can be 447 // folded to 448 // (A & (B | D)) == (B & (B ^ D)) | E. 449 // 450 // For example, 451 // (icmp ne (A & 12), 0) & (icmp eq (A & 7), 1) -> (icmp eq (A & 15), 9) 452 // (icmp ne (A & 15), 0) & (icmp eq (A & 7), 0) -> (icmp eq (A & 15), 8) 453 if ((((BCst->getValue() & DCst->getValue()) & ECst->getValue()) == 0) && 454 (BCst->getValue() & (BCst->getValue() ^ DCst->getValue())).isPowerOf2()) { 455 APInt BorD = BCst->getValue() | DCst->getValue(); 456 APInt BandBxorDorE = (BCst->getValue() & (BCst->getValue() ^ DCst->getValue())) | 457 ECst->getValue(); 458 Value *NewMask = ConstantInt::get(BCst->getType(), BorD); 459 Value *NewMaskedValue = ConstantInt::get(BCst->getType(), BandBxorDorE); 460 Value *NewAnd = Builder.CreateAnd(A, NewMask); 461 return Builder.CreateICmp(NewCC, NewAnd, NewMaskedValue); 462 } 463 464 auto IsSubSetOrEqual = [](ConstantInt *C1, ConstantInt *C2) { 465 return (C1->getValue() & C2->getValue()) == C1->getValue(); 466 }; 467 auto IsSuperSetOrEqual = [](ConstantInt *C1, ConstantInt *C2) { 468 return (C1->getValue() & C2->getValue()) == C2->getValue(); 469 }; 470 471 // In the following, we consider only the cases where B is a superset of D, B 472 // is a subset of D, or B == D because otherwise there's at least one bit 473 // covered by B but not D, in which case we can't deduce much from it, so 474 // no folding (aside from the single must-be-one bit case right above.) 475 // For example, 476 // (icmp ne (A & 14), 0) & (icmp eq (A & 3), 1) -> no folding. 477 if (!IsSubSetOrEqual(BCst, DCst) && !IsSuperSetOrEqual(BCst, DCst)) 478 return nullptr; 479 480 // At this point, either B is a superset of D, B is a subset of D or B == D. 481 482 // If E is zero, if B is a subset of (or equal to) D, LHS and RHS contradict 483 // and the whole expression becomes false (or true if negated), otherwise, no 484 // folding. 485 // For example, 486 // (icmp ne (A & 3), 0) & (icmp eq (A & 7), 0) -> false. 487 // (icmp ne (A & 15), 0) & (icmp eq (A & 3), 0) -> no folding. 488 if (ECst->isZero()) { 489 if (IsSubSetOrEqual(BCst, DCst)) 490 return ConstantInt::get(LHS->getType(), !IsAnd); 491 return nullptr; 492 } 493 494 // At this point, B, D, E aren't zero and (B & D) == B, (B & D) == D or B == 495 // D. If B is a superset of (or equal to) D, since E is not zero, LHS is 496 // subsumed by RHS (RHS implies LHS.) So the whole expression becomes 497 // RHS. For example, 498 // (icmp ne (A & 255), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8). 499 // (icmp ne (A & 15), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8). 500 if (IsSuperSetOrEqual(BCst, DCst)) 501 return RHS; 502 // Otherwise, B is a subset of D. If B and E have a common bit set, 503 // ie. (B & E) != 0, then LHS is subsumed by RHS. For example. 504 // (icmp ne (A & 12), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8). 505 assert(IsSubSetOrEqual(BCst, DCst) && "Precondition due to above code"); 506 if ((BCst->getValue() & ECst->getValue()) != 0) 507 return RHS; 508 // Otherwise, LHS and RHS contradict and the whole expression becomes false 509 // (or true if negated.) For example, 510 // (icmp ne (A & 7), 0) & (icmp eq (A & 15), 8) -> false. 511 // (icmp ne (A & 6), 0) & (icmp eq (A & 15), 8) -> false. 512 return ConstantInt::get(LHS->getType(), !IsAnd); 513 } 514 515 /// Try to fold (icmp(A & B) ==/!= 0) &/| (icmp(A & D) ==/!= E) into a single 516 /// (icmp(A & X) ==/!= Y), where the left-hand side and the right hand side 517 /// aren't of the common mask pattern type. 518 static Value *foldLogOpOfMaskedICmpsAsymmetric( 519 ICmpInst *LHS, ICmpInst *RHS, bool IsAnd, Value *A, Value *B, Value *C, 520 Value *D, Value *E, ICmpInst::Predicate PredL, ICmpInst::Predicate PredR, 521 unsigned LHSMask, unsigned RHSMask, InstCombiner::BuilderTy &Builder) { 522 assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) && 523 "Expected equality predicates for masked type of icmps."); 524 // Handle Mask_NotAllZeros-BMask_Mixed cases. 525 // (icmp ne/eq (A & B), C) &/| (icmp eq/ne (A & D), E), or 526 // (icmp eq/ne (A & B), C) &/| (icmp ne/eq (A & D), E) 527 // which gets swapped to 528 // (icmp ne/eq (A & D), E) &/| (icmp eq/ne (A & B), C). 529 if (!IsAnd) { 530 LHSMask = conjugateICmpMask(LHSMask); 531 RHSMask = conjugateICmpMask(RHSMask); 532 } 533 if ((LHSMask & Mask_NotAllZeros) && (RHSMask & BMask_Mixed)) { 534 if (Value *V = foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed( 535 LHS, RHS, IsAnd, A, B, C, D, E, 536 PredL, PredR, Builder)) { 537 return V; 538 } 539 } else if ((LHSMask & BMask_Mixed) && (RHSMask & Mask_NotAllZeros)) { 540 if (Value *V = foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed( 541 RHS, LHS, IsAnd, A, D, E, B, C, 542 PredR, PredL, Builder)) { 543 return V; 544 } 545 } 546 return nullptr; 547 } 548 549 /// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E) 550 /// into a single (icmp(A & X) ==/!= Y). 551 static Value *foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS, bool IsAnd, 552 InstCombiner::BuilderTy &Builder) { 553 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr; 554 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate(); 555 Optional<std::pair<unsigned, unsigned>> MaskPair = 556 getMaskedTypeForICmpPair(A, B, C, D, E, LHS, RHS, PredL, PredR); 557 if (!MaskPair) 558 return nullptr; 559 assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) && 560 "Expected equality predicates for masked type of icmps."); 561 unsigned LHSMask = MaskPair->first; 562 unsigned RHSMask = MaskPair->second; 563 unsigned Mask = LHSMask & RHSMask; 564 if (Mask == 0) { 565 // Even if the two sides don't share a common pattern, check if folding can 566 // still happen. 567 if (Value *V = foldLogOpOfMaskedICmpsAsymmetric( 568 LHS, RHS, IsAnd, A, B, C, D, E, PredL, PredR, LHSMask, RHSMask, 569 Builder)) 570 return V; 571 return nullptr; 572 } 573 574 // In full generality: 575 // (icmp (A & B) Op C) | (icmp (A & D) Op E) 576 // == ![ (icmp (A & B) !Op C) & (icmp (A & D) !Op E) ] 577 // 578 // If the latter can be converted into (icmp (A & X) Op Y) then the former is 579 // equivalent to (icmp (A & X) !Op Y). 580 // 581 // Therefore, we can pretend for the rest of this function that we're dealing 582 // with the conjunction, provided we flip the sense of any comparisons (both 583 // input and output). 584 585 // In most cases we're going to produce an EQ for the "&&" case. 586 ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE; 587 if (!IsAnd) { 588 // Convert the masking analysis into its equivalent with negated 589 // comparisons. 590 Mask = conjugateICmpMask(Mask); 591 } 592 593 if (Mask & Mask_AllZeros) { 594 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0) 595 // -> (icmp eq (A & (B|D)), 0) 596 Value *NewOr = Builder.CreateOr(B, D); 597 Value *NewAnd = Builder.CreateAnd(A, NewOr); 598 // We can't use C as zero because we might actually handle 599 // (icmp ne (A & B), B) & (icmp ne (A & D), D) 600 // with B and D, having a single bit set. 601 Value *Zero = Constant::getNullValue(A->getType()); 602 return Builder.CreateICmp(NewCC, NewAnd, Zero); 603 } 604 if (Mask & BMask_AllOnes) { 605 // (icmp eq (A & B), B) & (icmp eq (A & D), D) 606 // -> (icmp eq (A & (B|D)), (B|D)) 607 Value *NewOr = Builder.CreateOr(B, D); 608 Value *NewAnd = Builder.CreateAnd(A, NewOr); 609 return Builder.CreateICmp(NewCC, NewAnd, NewOr); 610 } 611 if (Mask & AMask_AllOnes) { 612 // (icmp eq (A & B), A) & (icmp eq (A & D), A) 613 // -> (icmp eq (A & (B&D)), A) 614 Value *NewAnd1 = Builder.CreateAnd(B, D); 615 Value *NewAnd2 = Builder.CreateAnd(A, NewAnd1); 616 return Builder.CreateICmp(NewCC, NewAnd2, A); 617 } 618 619 // Remaining cases assume at least that B and D are constant, and depend on 620 // their actual values. This isn't strictly necessary, just a "handle the 621 // easy cases for now" decision. 622 ConstantInt *BCst, *DCst; 623 if (!match(B, m_ConstantInt(BCst)) || !match(D, m_ConstantInt(DCst))) 624 return nullptr; 625 626 if (Mask & (Mask_NotAllZeros | BMask_NotAllOnes)) { 627 // (icmp ne (A & B), 0) & (icmp ne (A & D), 0) and 628 // (icmp ne (A & B), B) & (icmp ne (A & D), D) 629 // -> (icmp ne (A & B), 0) or (icmp ne (A & D), 0) 630 // Only valid if one of the masks is a superset of the other (check "B&D" is 631 // the same as either B or D). 632 APInt NewMask = BCst->getValue() & DCst->getValue(); 633 634 if (NewMask == BCst->getValue()) 635 return LHS; 636 else if (NewMask == DCst->getValue()) 637 return RHS; 638 } 639 640 if (Mask & AMask_NotAllOnes) { 641 // (icmp ne (A & B), B) & (icmp ne (A & D), D) 642 // -> (icmp ne (A & B), A) or (icmp ne (A & D), A) 643 // Only valid if one of the masks is a superset of the other (check "B|D" is 644 // the same as either B or D). 645 APInt NewMask = BCst->getValue() | DCst->getValue(); 646 647 if (NewMask == BCst->getValue()) 648 return LHS; 649 else if (NewMask == DCst->getValue()) 650 return RHS; 651 } 652 653 if (Mask & BMask_Mixed) { 654 // (icmp eq (A & B), C) & (icmp eq (A & D), E) 655 // We already know that B & C == C && D & E == E. 656 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of 657 // C and E, which are shared by both the mask B and the mask D, don't 658 // contradict, then we can transform to 659 // -> (icmp eq (A & (B|D)), (C|E)) 660 // Currently, we only handle the case of B, C, D, and E being constant. 661 // We can't simply use C and E because we might actually handle 662 // (icmp ne (A & B), B) & (icmp eq (A & D), D) 663 // with B and D, having a single bit set. 664 ConstantInt *CCst, *ECst; 665 if (!match(C, m_ConstantInt(CCst)) || !match(E, m_ConstantInt(ECst))) 666 return nullptr; 667 if (PredL != NewCC) 668 CCst = cast<ConstantInt>(ConstantExpr::getXor(BCst, CCst)); 669 if (PredR != NewCC) 670 ECst = cast<ConstantInt>(ConstantExpr::getXor(DCst, ECst)); 671 672 // If there is a conflict, we should actually return a false for the 673 // whole construct. 674 if (((BCst->getValue() & DCst->getValue()) & 675 (CCst->getValue() ^ ECst->getValue())).getBoolValue()) 676 return ConstantInt::get(LHS->getType(), !IsAnd); 677 678 Value *NewOr1 = Builder.CreateOr(B, D); 679 Value *NewOr2 = ConstantExpr::getOr(CCst, ECst); 680 Value *NewAnd = Builder.CreateAnd(A, NewOr1); 681 return Builder.CreateICmp(NewCC, NewAnd, NewOr2); 682 } 683 684 return nullptr; 685 } 686 687 /// Try to fold a signed range checked with lower bound 0 to an unsigned icmp. 688 /// Example: (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n 689 /// If \p Inverted is true then the check is for the inverted range, e.g. 690 /// (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n 691 Value *InstCombinerImpl::simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1, 692 bool Inverted) { 693 // Check the lower range comparison, e.g. x >= 0 694 // InstCombine already ensured that if there is a constant it's on the RHS. 695 ConstantInt *RangeStart = dyn_cast<ConstantInt>(Cmp0->getOperand(1)); 696 if (!RangeStart) 697 return nullptr; 698 699 ICmpInst::Predicate Pred0 = (Inverted ? Cmp0->getInversePredicate() : 700 Cmp0->getPredicate()); 701 702 // Accept x > -1 or x >= 0 (after potentially inverting the predicate). 703 if (!((Pred0 == ICmpInst::ICMP_SGT && RangeStart->isMinusOne()) || 704 (Pred0 == ICmpInst::ICMP_SGE && RangeStart->isZero()))) 705 return nullptr; 706 707 ICmpInst::Predicate Pred1 = (Inverted ? Cmp1->getInversePredicate() : 708 Cmp1->getPredicate()); 709 710 Value *Input = Cmp0->getOperand(0); 711 Value *RangeEnd; 712 if (Cmp1->getOperand(0) == Input) { 713 // For the upper range compare we have: icmp x, n 714 RangeEnd = Cmp1->getOperand(1); 715 } else if (Cmp1->getOperand(1) == Input) { 716 // For the upper range compare we have: icmp n, x 717 RangeEnd = Cmp1->getOperand(0); 718 Pred1 = ICmpInst::getSwappedPredicate(Pred1); 719 } else { 720 return nullptr; 721 } 722 723 // Check the upper range comparison, e.g. x < n 724 ICmpInst::Predicate NewPred; 725 switch (Pred1) { 726 case ICmpInst::ICMP_SLT: NewPred = ICmpInst::ICMP_ULT; break; 727 case ICmpInst::ICMP_SLE: NewPred = ICmpInst::ICMP_ULE; break; 728 default: return nullptr; 729 } 730 731 // This simplification is only valid if the upper range is not negative. 732 KnownBits Known = computeKnownBits(RangeEnd, /*Depth=*/0, Cmp1); 733 if (!Known.isNonNegative()) 734 return nullptr; 735 736 if (Inverted) 737 NewPred = ICmpInst::getInversePredicate(NewPred); 738 739 return Builder.CreateICmp(NewPred, Input, RangeEnd); 740 } 741 742 static Value * 743 foldAndOrOfEqualityCmpsWithConstants(ICmpInst *LHS, ICmpInst *RHS, 744 bool JoinedByAnd, 745 InstCombiner::BuilderTy &Builder) { 746 Value *X = LHS->getOperand(0); 747 if (X != RHS->getOperand(0)) 748 return nullptr; 749 750 const APInt *C1, *C2; 751 if (!match(LHS->getOperand(1), m_APInt(C1)) || 752 !match(RHS->getOperand(1), m_APInt(C2))) 753 return nullptr; 754 755 // We only handle (X != C1 && X != C2) and (X == C1 || X == C2). 756 ICmpInst::Predicate Pred = LHS->getPredicate(); 757 if (Pred != RHS->getPredicate()) 758 return nullptr; 759 if (JoinedByAnd && Pred != ICmpInst::ICMP_NE) 760 return nullptr; 761 if (!JoinedByAnd && Pred != ICmpInst::ICMP_EQ) 762 return nullptr; 763 764 // The larger unsigned constant goes on the right. 765 if (C1->ugt(*C2)) 766 std::swap(C1, C2); 767 768 APInt Xor = *C1 ^ *C2; 769 if (Xor.isPowerOf2()) { 770 // If LHSC and RHSC differ by only one bit, then set that bit in X and 771 // compare against the larger constant: 772 // (X == C1 || X == C2) --> (X | (C1 ^ C2)) == C2 773 // (X != C1 && X != C2) --> (X | (C1 ^ C2)) != C2 774 // We choose an 'or' with a Pow2 constant rather than the inverse mask with 775 // 'and' because that may lead to smaller codegen from a smaller constant. 776 Value *Or = Builder.CreateOr(X, ConstantInt::get(X->getType(), Xor)); 777 return Builder.CreateICmp(Pred, Or, ConstantInt::get(X->getType(), *C2)); 778 } 779 780 // Special case: get the ordering right when the values wrap around zero. 781 // Ie, we assumed the constants were unsigned when swapping earlier. 782 if (C1->isNullValue() && C2->isAllOnesValue()) 783 std::swap(C1, C2); 784 785 if (*C1 == *C2 - 1) { 786 // (X == 13 || X == 14) --> X - 13 <=u 1 787 // (X != 13 && X != 14) --> X - 13 >u 1 788 // An 'add' is the canonical IR form, so favor that over a 'sub'. 789 Value *Add = Builder.CreateAdd(X, ConstantInt::get(X->getType(), -(*C1))); 790 auto NewPred = JoinedByAnd ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_ULE; 791 return Builder.CreateICmp(NewPred, Add, ConstantInt::get(X->getType(), 1)); 792 } 793 794 return nullptr; 795 } 796 797 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2) 798 // Fold (!iszero(A & K1) & !iszero(A & K2)) -> (A & (K1 | K2)) == (K1 | K2) 799 Value *InstCombinerImpl::foldAndOrOfICmpsOfAndWithPow2(ICmpInst *LHS, 800 ICmpInst *RHS, 801 BinaryOperator &Logic) { 802 bool JoinedByAnd = Logic.getOpcode() == Instruction::And; 803 assert((JoinedByAnd || Logic.getOpcode() == Instruction::Or) && 804 "Wrong opcode"); 805 ICmpInst::Predicate Pred = LHS->getPredicate(); 806 if (Pred != RHS->getPredicate()) 807 return nullptr; 808 if (JoinedByAnd && Pred != ICmpInst::ICMP_NE) 809 return nullptr; 810 if (!JoinedByAnd && Pred != ICmpInst::ICMP_EQ) 811 return nullptr; 812 813 if (!match(LHS->getOperand(1), m_Zero()) || 814 !match(RHS->getOperand(1), m_Zero())) 815 return nullptr; 816 817 Value *A, *B, *C, *D; 818 if (match(LHS->getOperand(0), m_And(m_Value(A), m_Value(B))) && 819 match(RHS->getOperand(0), m_And(m_Value(C), m_Value(D)))) { 820 if (A == D || B == D) 821 std::swap(C, D); 822 if (B == C) 823 std::swap(A, B); 824 825 if (A == C && 826 isKnownToBeAPowerOfTwo(B, false, 0, &Logic) && 827 isKnownToBeAPowerOfTwo(D, false, 0, &Logic)) { 828 Value *Mask = Builder.CreateOr(B, D); 829 Value *Masked = Builder.CreateAnd(A, Mask); 830 auto NewPred = JoinedByAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE; 831 return Builder.CreateICmp(NewPred, Masked, Mask); 832 } 833 } 834 835 return nullptr; 836 } 837 838 /// General pattern: 839 /// X & Y 840 /// 841 /// Where Y is checking that all the high bits (covered by a mask 4294967168) 842 /// are uniform, i.e. %arg & 4294967168 can be either 4294967168 or 0 843 /// Pattern can be one of: 844 /// %t = add i32 %arg, 128 845 /// %r = icmp ult i32 %t, 256 846 /// Or 847 /// %t0 = shl i32 %arg, 24 848 /// %t1 = ashr i32 %t0, 24 849 /// %r = icmp eq i32 %t1, %arg 850 /// Or 851 /// %t0 = trunc i32 %arg to i8 852 /// %t1 = sext i8 %t0 to i32 853 /// %r = icmp eq i32 %t1, %arg 854 /// This pattern is a signed truncation check. 855 /// 856 /// And X is checking that some bit in that same mask is zero. 857 /// I.e. can be one of: 858 /// %r = icmp sgt i32 %arg, -1 859 /// Or 860 /// %t = and i32 %arg, 2147483648 861 /// %r = icmp eq i32 %t, 0 862 /// 863 /// Since we are checking that all the bits in that mask are the same, 864 /// and a particular bit is zero, what we are really checking is that all the 865 /// masked bits are zero. 866 /// So this should be transformed to: 867 /// %r = icmp ult i32 %arg, 128 868 static Value *foldSignedTruncationCheck(ICmpInst *ICmp0, ICmpInst *ICmp1, 869 Instruction &CxtI, 870 InstCombiner::BuilderTy &Builder) { 871 assert(CxtI.getOpcode() == Instruction::And); 872 873 // Match icmp ult (add %arg, C01), C1 (C1 == C01 << 1; powers of two) 874 auto tryToMatchSignedTruncationCheck = [](ICmpInst *ICmp, Value *&X, 875 APInt &SignBitMask) -> bool { 876 CmpInst::Predicate Pred; 877 const APInt *I01, *I1; // powers of two; I1 == I01 << 1 878 if (!(match(ICmp, 879 m_ICmp(Pred, m_Add(m_Value(X), m_Power2(I01)), m_Power2(I1))) && 880 Pred == ICmpInst::ICMP_ULT && I1->ugt(*I01) && I01->shl(1) == *I1)) 881 return false; 882 // Which bit is the new sign bit as per the 'signed truncation' pattern? 883 SignBitMask = *I01; 884 return true; 885 }; 886 887 // One icmp needs to be 'signed truncation check'. 888 // We need to match this first, else we will mismatch commutative cases. 889 Value *X1; 890 APInt HighestBit; 891 ICmpInst *OtherICmp; 892 if (tryToMatchSignedTruncationCheck(ICmp1, X1, HighestBit)) 893 OtherICmp = ICmp0; 894 else if (tryToMatchSignedTruncationCheck(ICmp0, X1, HighestBit)) 895 OtherICmp = ICmp1; 896 else 897 return nullptr; 898 899 assert(HighestBit.isPowerOf2() && "expected to be power of two (non-zero)"); 900 901 // Try to match/decompose into: icmp eq (X & Mask), 0 902 auto tryToDecompose = [](ICmpInst *ICmp, Value *&X, 903 APInt &UnsetBitsMask) -> bool { 904 CmpInst::Predicate Pred = ICmp->getPredicate(); 905 // Can it be decomposed into icmp eq (X & Mask), 0 ? 906 if (llvm::decomposeBitTestICmp(ICmp->getOperand(0), ICmp->getOperand(1), 907 Pred, X, UnsetBitsMask, 908 /*LookThroughTrunc=*/false) && 909 Pred == ICmpInst::ICMP_EQ) 910 return true; 911 // Is it icmp eq (X & Mask), 0 already? 912 const APInt *Mask; 913 if (match(ICmp, m_ICmp(Pred, m_And(m_Value(X), m_APInt(Mask)), m_Zero())) && 914 Pred == ICmpInst::ICMP_EQ) { 915 UnsetBitsMask = *Mask; 916 return true; 917 } 918 return false; 919 }; 920 921 // And the other icmp needs to be decomposable into a bit test. 922 Value *X0; 923 APInt UnsetBitsMask; 924 if (!tryToDecompose(OtherICmp, X0, UnsetBitsMask)) 925 return nullptr; 926 927 assert(!UnsetBitsMask.isNullValue() && "empty mask makes no sense."); 928 929 // Are they working on the same value? 930 Value *X; 931 if (X1 == X0) { 932 // Ok as is. 933 X = X1; 934 } else if (match(X0, m_Trunc(m_Specific(X1)))) { 935 UnsetBitsMask = UnsetBitsMask.zext(X1->getType()->getScalarSizeInBits()); 936 X = X1; 937 } else 938 return nullptr; 939 940 // So which bits should be uniform as per the 'signed truncation check'? 941 // (all the bits starting with (i.e. including) HighestBit) 942 APInt SignBitsMask = ~(HighestBit - 1U); 943 944 // UnsetBitsMask must have some common bits with SignBitsMask, 945 if (!UnsetBitsMask.intersects(SignBitsMask)) 946 return nullptr; 947 948 // Does UnsetBitsMask contain any bits outside of SignBitsMask? 949 if (!UnsetBitsMask.isSubsetOf(SignBitsMask)) { 950 APInt OtherHighestBit = (~UnsetBitsMask) + 1U; 951 if (!OtherHighestBit.isPowerOf2()) 952 return nullptr; 953 HighestBit = APIntOps::umin(HighestBit, OtherHighestBit); 954 } 955 // Else, if it does not, then all is ok as-is. 956 957 // %r = icmp ult %X, SignBit 958 return Builder.CreateICmpULT(X, ConstantInt::get(X->getType(), HighestBit), 959 CxtI.getName() + ".simplified"); 960 } 961 962 /// Reduce a pair of compares that check if a value has exactly 1 bit set. 963 static Value *foldIsPowerOf2(ICmpInst *Cmp0, ICmpInst *Cmp1, bool JoinedByAnd, 964 InstCombiner::BuilderTy &Builder) { 965 // Handle 'and' / 'or' commutation: make the equality check the first operand. 966 if (JoinedByAnd && Cmp1->getPredicate() == ICmpInst::ICMP_NE) 967 std::swap(Cmp0, Cmp1); 968 else if (!JoinedByAnd && Cmp1->getPredicate() == ICmpInst::ICMP_EQ) 969 std::swap(Cmp0, Cmp1); 970 971 // (X != 0) && (ctpop(X) u< 2) --> ctpop(X) == 1 972 CmpInst::Predicate Pred0, Pred1; 973 Value *X; 974 if (JoinedByAnd && match(Cmp0, m_ICmp(Pred0, m_Value(X), m_ZeroInt())) && 975 match(Cmp1, m_ICmp(Pred1, m_Intrinsic<Intrinsic::ctpop>(m_Specific(X)), 976 m_SpecificInt(2))) && 977 Pred0 == ICmpInst::ICMP_NE && Pred1 == ICmpInst::ICMP_ULT) { 978 Value *CtPop = Cmp1->getOperand(0); 979 return Builder.CreateICmpEQ(CtPop, ConstantInt::get(CtPop->getType(), 1)); 980 } 981 // (X == 0) || (ctpop(X) u> 1) --> ctpop(X) != 1 982 if (!JoinedByAnd && match(Cmp0, m_ICmp(Pred0, m_Value(X), m_ZeroInt())) && 983 match(Cmp1, m_ICmp(Pred1, m_Intrinsic<Intrinsic::ctpop>(m_Specific(X)), 984 m_SpecificInt(1))) && 985 Pred0 == ICmpInst::ICMP_EQ && Pred1 == ICmpInst::ICMP_UGT) { 986 Value *CtPop = Cmp1->getOperand(0); 987 return Builder.CreateICmpNE(CtPop, ConstantInt::get(CtPop->getType(), 1)); 988 } 989 return nullptr; 990 } 991 992 /// Commuted variants are assumed to be handled by calling this function again 993 /// with the parameters swapped. 994 static Value *foldUnsignedUnderflowCheck(ICmpInst *ZeroICmp, 995 ICmpInst *UnsignedICmp, bool IsAnd, 996 const SimplifyQuery &Q, 997 InstCombiner::BuilderTy &Builder) { 998 Value *ZeroCmpOp; 999 ICmpInst::Predicate EqPred; 1000 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(ZeroCmpOp), m_Zero())) || 1001 !ICmpInst::isEquality(EqPred)) 1002 return nullptr; 1003 1004 auto IsKnownNonZero = [&](Value *V) { 1005 return isKnownNonZero(V, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT); 1006 }; 1007 1008 ICmpInst::Predicate UnsignedPred; 1009 1010 Value *A, *B; 1011 if (match(UnsignedICmp, 1012 m_c_ICmp(UnsignedPred, m_Specific(ZeroCmpOp), m_Value(A))) && 1013 match(ZeroCmpOp, m_c_Add(m_Specific(A), m_Value(B))) && 1014 (ZeroICmp->hasOneUse() || UnsignedICmp->hasOneUse())) { 1015 auto GetKnownNonZeroAndOther = [&](Value *&NonZero, Value *&Other) { 1016 if (!IsKnownNonZero(NonZero)) 1017 std::swap(NonZero, Other); 1018 return IsKnownNonZero(NonZero); 1019 }; 1020 1021 // Given ZeroCmpOp = (A + B) 1022 // ZeroCmpOp <= A && ZeroCmpOp != 0 --> (0-B) < A 1023 // ZeroCmpOp > A || ZeroCmpOp == 0 --> (0-B) >= A 1024 // 1025 // ZeroCmpOp < A && ZeroCmpOp != 0 --> (0-X) < Y iff 1026 // ZeroCmpOp >= A || ZeroCmpOp == 0 --> (0-X) >= Y iff 1027 // with X being the value (A/B) that is known to be non-zero, 1028 // and Y being remaining value. 1029 if (UnsignedPred == ICmpInst::ICMP_ULE && EqPred == ICmpInst::ICMP_NE && 1030 IsAnd) 1031 return Builder.CreateICmpULT(Builder.CreateNeg(B), A); 1032 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE && 1033 IsAnd && GetKnownNonZeroAndOther(B, A)) 1034 return Builder.CreateICmpULT(Builder.CreateNeg(B), A); 1035 if (UnsignedPred == ICmpInst::ICMP_UGT && EqPred == ICmpInst::ICMP_EQ && 1036 !IsAnd) 1037 return Builder.CreateICmpUGE(Builder.CreateNeg(B), A); 1038 if (UnsignedPred == ICmpInst::ICMP_UGE && EqPred == ICmpInst::ICMP_EQ && 1039 !IsAnd && GetKnownNonZeroAndOther(B, A)) 1040 return Builder.CreateICmpUGE(Builder.CreateNeg(B), A); 1041 } 1042 1043 Value *Base, *Offset; 1044 if (!match(ZeroCmpOp, m_Sub(m_Value(Base), m_Value(Offset)))) 1045 return nullptr; 1046 1047 if (!match(UnsignedICmp, 1048 m_c_ICmp(UnsignedPred, m_Specific(Base), m_Specific(Offset))) || 1049 !ICmpInst::isUnsigned(UnsignedPred)) 1050 return nullptr; 1051 1052 // Base >=/> Offset && (Base - Offset) != 0 <--> Base > Offset 1053 // (no overflow and not null) 1054 if ((UnsignedPred == ICmpInst::ICMP_UGE || 1055 UnsignedPred == ICmpInst::ICMP_UGT) && 1056 EqPred == ICmpInst::ICMP_NE && IsAnd) 1057 return Builder.CreateICmpUGT(Base, Offset); 1058 1059 // Base <=/< Offset || (Base - Offset) == 0 <--> Base <= Offset 1060 // (overflow or null) 1061 if ((UnsignedPred == ICmpInst::ICMP_ULE || 1062 UnsignedPred == ICmpInst::ICMP_ULT) && 1063 EqPred == ICmpInst::ICMP_EQ && !IsAnd) 1064 return Builder.CreateICmpULE(Base, Offset); 1065 1066 // Base <= Offset && (Base - Offset) != 0 --> Base < Offset 1067 if (UnsignedPred == ICmpInst::ICMP_ULE && EqPred == ICmpInst::ICMP_NE && 1068 IsAnd) 1069 return Builder.CreateICmpULT(Base, Offset); 1070 1071 // Base > Offset || (Base - Offset) == 0 --> Base >= Offset 1072 if (UnsignedPred == ICmpInst::ICMP_UGT && EqPred == ICmpInst::ICMP_EQ && 1073 !IsAnd) 1074 return Builder.CreateICmpUGE(Base, Offset); 1075 1076 return nullptr; 1077 } 1078 1079 /// Reduce logic-of-compares with equality to a constant by substituting a 1080 /// common operand with the constant. Callers are expected to call this with 1081 /// Cmp0/Cmp1 switched to handle logic op commutativity. 1082 static Value *foldAndOrOfICmpsWithConstEq(ICmpInst *Cmp0, ICmpInst *Cmp1, 1083 BinaryOperator &Logic, 1084 InstCombiner::BuilderTy &Builder, 1085 const SimplifyQuery &Q) { 1086 bool IsAnd = Logic.getOpcode() == Instruction::And; 1087 assert((IsAnd || Logic.getOpcode() == Instruction::Or) && "Wrong logic op"); 1088 1089 // Match an equality compare with a non-poison constant as Cmp0. 1090 // Also, give up if the compare can be constant-folded to avoid looping. 1091 ICmpInst::Predicate Pred0; 1092 Value *X; 1093 Constant *C; 1094 if (!match(Cmp0, m_ICmp(Pred0, m_Value(X), m_Constant(C))) || 1095 !isGuaranteedNotToBeUndefOrPoison(C) || isa<Constant>(X)) 1096 return nullptr; 1097 if ((IsAnd && Pred0 != ICmpInst::ICMP_EQ) || 1098 (!IsAnd && Pred0 != ICmpInst::ICMP_NE)) 1099 return nullptr; 1100 1101 // The other compare must include a common operand (X). Canonicalize the 1102 // common operand as operand 1 (Pred1 is swapped if the common operand was 1103 // operand 0). 1104 Value *Y; 1105 ICmpInst::Predicate Pred1; 1106 if (!match(Cmp1, m_c_ICmp(Pred1, m_Value(Y), m_Deferred(X)))) 1107 return nullptr; 1108 1109 // Replace variable with constant value equivalence to remove a variable use: 1110 // (X == C) && (Y Pred1 X) --> (X == C) && (Y Pred1 C) 1111 // (X != C) || (Y Pred1 X) --> (X != C) || (Y Pred1 C) 1112 // Can think of the 'or' substitution with the 'and' bool equivalent: 1113 // A || B --> A || (!A && B) 1114 Value *SubstituteCmp = SimplifyICmpInst(Pred1, Y, C, Q); 1115 if (!SubstituteCmp) { 1116 // If we need to create a new instruction, require that the old compare can 1117 // be removed. 1118 if (!Cmp1->hasOneUse()) 1119 return nullptr; 1120 SubstituteCmp = Builder.CreateICmp(Pred1, Y, C); 1121 } 1122 return Builder.CreateBinOp(Logic.getOpcode(), Cmp0, SubstituteCmp); 1123 } 1124 1125 /// Fold (icmp)&(icmp) if possible. 1126 Value *InstCombinerImpl::foldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS, 1127 BinaryOperator &And) { 1128 const SimplifyQuery Q = SQ.getWithInstruction(&And); 1129 1130 // Fold (!iszero(A & K1) & !iszero(A & K2)) -> (A & (K1 | K2)) == (K1 | K2) 1131 // if K1 and K2 are a one-bit mask. 1132 if (Value *V = foldAndOrOfICmpsOfAndWithPow2(LHS, RHS, And)) 1133 return V; 1134 1135 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate(); 1136 1137 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B) 1138 if (predicatesFoldable(PredL, PredR)) { 1139 if (LHS->getOperand(0) == RHS->getOperand(1) && 1140 LHS->getOperand(1) == RHS->getOperand(0)) 1141 LHS->swapOperands(); 1142 if (LHS->getOperand(0) == RHS->getOperand(0) && 1143 LHS->getOperand(1) == RHS->getOperand(1)) { 1144 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1); 1145 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS); 1146 bool IsSigned = LHS->isSigned() || RHS->isSigned(); 1147 return getNewICmpValue(Code, IsSigned, Op0, Op1, Builder); 1148 } 1149 } 1150 1151 // handle (roughly): (icmp eq (A & B), C) & (icmp eq (A & D), E) 1152 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, true, Builder)) 1153 return V; 1154 1155 if (Value *V = foldAndOrOfICmpsWithConstEq(LHS, RHS, And, Builder, Q)) 1156 return V; 1157 if (Value *V = foldAndOrOfICmpsWithConstEq(RHS, LHS, And, Builder, Q)) 1158 return V; 1159 1160 // E.g. (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n 1161 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/false)) 1162 return V; 1163 1164 // E.g. (icmp slt x, n) & (icmp sge x, 0) --> icmp ult x, n 1165 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/false)) 1166 return V; 1167 1168 if (Value *V = foldAndOrOfEqualityCmpsWithConstants(LHS, RHS, true, Builder)) 1169 return V; 1170 1171 if (Value *V = foldSignedTruncationCheck(LHS, RHS, And, Builder)) 1172 return V; 1173 1174 if (Value *V = foldIsPowerOf2(LHS, RHS, true /* JoinedByAnd */, Builder)) 1175 return V; 1176 1177 if (Value *X = 1178 foldUnsignedUnderflowCheck(LHS, RHS, /*IsAnd=*/true, Q, Builder)) 1179 return X; 1180 if (Value *X = 1181 foldUnsignedUnderflowCheck(RHS, LHS, /*IsAnd=*/true, Q, Builder)) 1182 return X; 1183 1184 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2). 1185 Value *LHS0 = LHS->getOperand(0), *RHS0 = RHS->getOperand(0); 1186 1187 ConstantInt *LHSC, *RHSC; 1188 if (!match(LHS->getOperand(1), m_ConstantInt(LHSC)) || 1189 !match(RHS->getOperand(1), m_ConstantInt(RHSC))) 1190 return nullptr; 1191 1192 if (LHSC == RHSC && PredL == PredR) { 1193 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C) 1194 // where C is a power of 2 or 1195 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0) 1196 if ((PredL == ICmpInst::ICMP_ULT && LHSC->getValue().isPowerOf2()) || 1197 (PredL == ICmpInst::ICMP_EQ && LHSC->isZero())) { 1198 Value *NewOr = Builder.CreateOr(LHS0, RHS0); 1199 return Builder.CreateICmp(PredL, NewOr, LHSC); 1200 } 1201 } 1202 1203 // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2 1204 // where CMAX is the all ones value for the truncated type, 1205 // iff the lower bits of C2 and CA are zero. 1206 if (PredL == ICmpInst::ICMP_EQ && PredL == PredR && LHS->hasOneUse() && 1207 RHS->hasOneUse()) { 1208 Value *V; 1209 ConstantInt *AndC, *SmallC = nullptr, *BigC = nullptr; 1210 1211 // (trunc x) == C1 & (and x, CA) == C2 1212 // (and x, CA) == C2 & (trunc x) == C1 1213 if (match(RHS0, m_Trunc(m_Value(V))) && 1214 match(LHS0, m_And(m_Specific(V), m_ConstantInt(AndC)))) { 1215 SmallC = RHSC; 1216 BigC = LHSC; 1217 } else if (match(LHS0, m_Trunc(m_Value(V))) && 1218 match(RHS0, m_And(m_Specific(V), m_ConstantInt(AndC)))) { 1219 SmallC = LHSC; 1220 BigC = RHSC; 1221 } 1222 1223 if (SmallC && BigC) { 1224 unsigned BigBitSize = BigC->getType()->getBitWidth(); 1225 unsigned SmallBitSize = SmallC->getType()->getBitWidth(); 1226 1227 // Check that the low bits are zero. 1228 APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize); 1229 if ((Low & AndC->getValue()).isNullValue() && 1230 (Low & BigC->getValue()).isNullValue()) { 1231 Value *NewAnd = Builder.CreateAnd(V, Low | AndC->getValue()); 1232 APInt N = SmallC->getValue().zext(BigBitSize) | BigC->getValue(); 1233 Value *NewVal = ConstantInt::get(AndC->getType()->getContext(), N); 1234 return Builder.CreateICmp(PredL, NewAnd, NewVal); 1235 } 1236 } 1237 } 1238 1239 // From here on, we only handle: 1240 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler. 1241 if (LHS0 != RHS0) 1242 return nullptr; 1243 1244 // ICMP_[US][GL]E X, C is folded to ICMP_[US][GL]T elsewhere. 1245 if (PredL == ICmpInst::ICMP_UGE || PredL == ICmpInst::ICMP_ULE || 1246 PredR == ICmpInst::ICMP_UGE || PredR == ICmpInst::ICMP_ULE || 1247 PredL == ICmpInst::ICMP_SGE || PredL == ICmpInst::ICMP_SLE || 1248 PredR == ICmpInst::ICMP_SGE || PredR == ICmpInst::ICMP_SLE) 1249 return nullptr; 1250 1251 // We can't fold (ugt x, C) & (sgt x, C2). 1252 if (!predicatesFoldable(PredL, PredR)) 1253 return nullptr; 1254 1255 // Ensure that the larger constant is on the RHS. 1256 bool ShouldSwap; 1257 if (CmpInst::isSigned(PredL) || 1258 (ICmpInst::isEquality(PredL) && CmpInst::isSigned(PredR))) 1259 ShouldSwap = LHSC->getValue().sgt(RHSC->getValue()); 1260 else 1261 ShouldSwap = LHSC->getValue().ugt(RHSC->getValue()); 1262 1263 if (ShouldSwap) { 1264 std::swap(LHS, RHS); 1265 std::swap(LHSC, RHSC); 1266 std::swap(PredL, PredR); 1267 } 1268 1269 // At this point, we know we have two icmp instructions 1270 // comparing a value against two constants and and'ing the result 1271 // together. Because of the above check, we know that we only have 1272 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know 1273 // (from the icmp folding check above), that the two constants 1274 // are not equal and that the larger constant is on the RHS 1275 assert(LHSC != RHSC && "Compares not folded above?"); 1276 1277 switch (PredL) { 1278 default: 1279 llvm_unreachable("Unknown integer condition code!"); 1280 case ICmpInst::ICMP_NE: 1281 switch (PredR) { 1282 default: 1283 llvm_unreachable("Unknown integer condition code!"); 1284 case ICmpInst::ICMP_ULT: 1285 // (X != 13 & X u< 14) -> X < 13 1286 if (LHSC->getValue() == (RHSC->getValue() - 1)) 1287 return Builder.CreateICmpULT(LHS0, LHSC); 1288 if (LHSC->isZero()) // (X != 0 & X u< C) -> X-1 u< C-1 1289 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(), 1290 false, true); 1291 break; // (X != 13 & X u< 15) -> no change 1292 case ICmpInst::ICMP_SLT: 1293 // (X != 13 & X s< 14) -> X < 13 1294 if (LHSC->getValue() == (RHSC->getValue() - 1)) 1295 return Builder.CreateICmpSLT(LHS0, LHSC); 1296 // (X != INT_MIN & X s< C) -> X-(INT_MIN+1) u< (C-(INT_MIN+1)) 1297 if (LHSC->isMinValue(true)) 1298 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(), 1299 true, true); 1300 break; // (X != 13 & X s< 15) -> no change 1301 case ICmpInst::ICMP_NE: 1302 // Potential folds for this case should already be handled. 1303 break; 1304 } 1305 break; 1306 case ICmpInst::ICMP_UGT: 1307 switch (PredR) { 1308 default: 1309 llvm_unreachable("Unknown integer condition code!"); 1310 case ICmpInst::ICMP_NE: 1311 // (X u> 13 & X != 14) -> X u> 14 1312 if (RHSC->getValue() == (LHSC->getValue() + 1)) 1313 return Builder.CreateICmp(PredL, LHS0, RHSC); 1314 // X u> C & X != UINT_MAX -> (X-(C+1)) u< UINT_MAX-(C+1) 1315 if (RHSC->isMaxValue(false)) 1316 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(), 1317 false, true); 1318 break; // (X u> 13 & X != 15) -> no change 1319 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) u< 1 1320 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(), 1321 false, true); 1322 } 1323 break; 1324 case ICmpInst::ICMP_SGT: 1325 switch (PredR) { 1326 default: 1327 llvm_unreachable("Unknown integer condition code!"); 1328 case ICmpInst::ICMP_NE: 1329 // (X s> 13 & X != 14) -> X s> 14 1330 if (RHSC->getValue() == (LHSC->getValue() + 1)) 1331 return Builder.CreateICmp(PredL, LHS0, RHSC); 1332 // X s> C & X != INT_MAX -> (X-(C+1)) u< INT_MAX-(C+1) 1333 if (RHSC->isMaxValue(true)) 1334 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(), 1335 true, true); 1336 break; // (X s> 13 & X != 15) -> no change 1337 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) u< 1 1338 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(), true, 1339 true); 1340 } 1341 break; 1342 } 1343 1344 return nullptr; 1345 } 1346 1347 Value *InstCombinerImpl::foldLogicOfFCmps(FCmpInst *LHS, FCmpInst *RHS, 1348 bool IsAnd) { 1349 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1); 1350 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1); 1351 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate(); 1352 1353 if (LHS0 == RHS1 && RHS0 == LHS1) { 1354 // Swap RHS operands to match LHS. 1355 PredR = FCmpInst::getSwappedPredicate(PredR); 1356 std::swap(RHS0, RHS1); 1357 } 1358 1359 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y). 1360 // Suppose the relation between x and y is R, where R is one of 1361 // U(1000), L(0100), G(0010) or E(0001), and CC0 and CC1 are the bitmasks for 1362 // testing the desired relations. 1363 // 1364 // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this: 1365 // bool(R & CC0) && bool(R & CC1) 1366 // = bool((R & CC0) & (R & CC1)) 1367 // = bool(R & (CC0 & CC1)) <= by re-association, commutation, and idempotency 1368 // 1369 // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this: 1370 // bool(R & CC0) || bool(R & CC1) 1371 // = bool((R & CC0) | (R & CC1)) 1372 // = bool(R & (CC0 | CC1)) <= by reversed distribution (contribution? ;) 1373 if (LHS0 == RHS0 && LHS1 == RHS1) { 1374 unsigned FCmpCodeL = getFCmpCode(PredL); 1375 unsigned FCmpCodeR = getFCmpCode(PredR); 1376 unsigned NewPred = IsAnd ? FCmpCodeL & FCmpCodeR : FCmpCodeL | FCmpCodeR; 1377 return getFCmpValue(NewPred, LHS0, LHS1, Builder); 1378 } 1379 1380 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) || 1381 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) { 1382 if (LHS0->getType() != RHS0->getType()) 1383 return nullptr; 1384 1385 // FCmp canonicalization ensures that (fcmp ord/uno X, X) and 1386 // (fcmp ord/uno X, C) will be transformed to (fcmp X, +0.0). 1387 if (match(LHS1, m_PosZeroFP()) && match(RHS1, m_PosZeroFP())) 1388 // Ignore the constants because they are obviously not NANs: 1389 // (fcmp ord x, 0.0) & (fcmp ord y, 0.0) -> (fcmp ord x, y) 1390 // (fcmp uno x, 0.0) | (fcmp uno y, 0.0) -> (fcmp uno x, y) 1391 return Builder.CreateFCmp(PredL, LHS0, RHS0); 1392 } 1393 1394 return nullptr; 1395 } 1396 1397 /// This a limited reassociation for a special case (see above) where we are 1398 /// checking if two values are either both NAN (unordered) or not-NAN (ordered). 1399 /// This could be handled more generally in '-reassociation', but it seems like 1400 /// an unlikely pattern for a large number of logic ops and fcmps. 1401 static Instruction *reassociateFCmps(BinaryOperator &BO, 1402 InstCombiner::BuilderTy &Builder) { 1403 Instruction::BinaryOps Opcode = BO.getOpcode(); 1404 assert((Opcode == Instruction::And || Opcode == Instruction::Or) && 1405 "Expecting and/or op for fcmp transform"); 1406 1407 // There are 4 commuted variants of the pattern. Canonicalize operands of this 1408 // logic op so an fcmp is operand 0 and a matching logic op is operand 1. 1409 Value *Op0 = BO.getOperand(0), *Op1 = BO.getOperand(1), *X; 1410 FCmpInst::Predicate Pred; 1411 if (match(Op1, m_FCmp(Pred, m_Value(), m_AnyZeroFP()))) 1412 std::swap(Op0, Op1); 1413 1414 // Match inner binop and the predicate for combining 2 NAN checks into 1. 1415 BinaryOperator *BO1; 1416 FCmpInst::Predicate NanPred = Opcode == Instruction::And ? FCmpInst::FCMP_ORD 1417 : FCmpInst::FCMP_UNO; 1418 if (!match(Op0, m_FCmp(Pred, m_Value(X), m_AnyZeroFP())) || Pred != NanPred || 1419 !match(Op1, m_BinOp(BO1)) || BO1->getOpcode() != Opcode) 1420 return nullptr; 1421 1422 // The inner logic op must have a matching fcmp operand. 1423 Value *BO10 = BO1->getOperand(0), *BO11 = BO1->getOperand(1), *Y; 1424 if (!match(BO10, m_FCmp(Pred, m_Value(Y), m_AnyZeroFP())) || 1425 Pred != NanPred || X->getType() != Y->getType()) 1426 std::swap(BO10, BO11); 1427 1428 if (!match(BO10, m_FCmp(Pred, m_Value(Y), m_AnyZeroFP())) || 1429 Pred != NanPred || X->getType() != Y->getType()) 1430 return nullptr; 1431 1432 // and (fcmp ord X, 0), (and (fcmp ord Y, 0), Z) --> and (fcmp ord X, Y), Z 1433 // or (fcmp uno X, 0), (or (fcmp uno Y, 0), Z) --> or (fcmp uno X, Y), Z 1434 Value *NewFCmp = Builder.CreateFCmp(Pred, X, Y); 1435 if (auto *NewFCmpInst = dyn_cast<FCmpInst>(NewFCmp)) { 1436 // Intersect FMF from the 2 source fcmps. 1437 NewFCmpInst->copyIRFlags(Op0); 1438 NewFCmpInst->andIRFlags(BO10); 1439 } 1440 return BinaryOperator::Create(Opcode, NewFCmp, BO11); 1441 } 1442 1443 /// Match De Morgan's Laws: 1444 /// (~A & ~B) == (~(A | B)) 1445 /// (~A | ~B) == (~(A & B)) 1446 static Instruction *matchDeMorgansLaws(BinaryOperator &I, 1447 InstCombiner::BuilderTy &Builder) { 1448 auto Opcode = I.getOpcode(); 1449 assert((Opcode == Instruction::And || Opcode == Instruction::Or) && 1450 "Trying to match De Morgan's Laws with something other than and/or"); 1451 1452 // Flip the logic operation. 1453 Opcode = (Opcode == Instruction::And) ? Instruction::Or : Instruction::And; 1454 1455 Value *A, *B; 1456 if (match(I.getOperand(0), m_OneUse(m_Not(m_Value(A)))) && 1457 match(I.getOperand(1), m_OneUse(m_Not(m_Value(B)))) && 1458 !InstCombiner::isFreeToInvert(A, A->hasOneUse()) && 1459 !InstCombiner::isFreeToInvert(B, B->hasOneUse())) { 1460 Value *AndOr = Builder.CreateBinOp(Opcode, A, B, I.getName() + ".demorgan"); 1461 return BinaryOperator::CreateNot(AndOr); 1462 } 1463 1464 return nullptr; 1465 } 1466 1467 bool InstCombinerImpl::shouldOptimizeCast(CastInst *CI) { 1468 Value *CastSrc = CI->getOperand(0); 1469 1470 // Noop casts and casts of constants should be eliminated trivially. 1471 if (CI->getSrcTy() == CI->getDestTy() || isa<Constant>(CastSrc)) 1472 return false; 1473 1474 // If this cast is paired with another cast that can be eliminated, we prefer 1475 // to have it eliminated. 1476 if (const auto *PrecedingCI = dyn_cast<CastInst>(CastSrc)) 1477 if (isEliminableCastPair(PrecedingCI, CI)) 1478 return false; 1479 1480 return true; 1481 } 1482 1483 /// Fold {and,or,xor} (cast X), C. 1484 static Instruction *foldLogicCastConstant(BinaryOperator &Logic, CastInst *Cast, 1485 InstCombiner::BuilderTy &Builder) { 1486 Constant *C = dyn_cast<Constant>(Logic.getOperand(1)); 1487 if (!C) 1488 return nullptr; 1489 1490 auto LogicOpc = Logic.getOpcode(); 1491 Type *DestTy = Logic.getType(); 1492 Type *SrcTy = Cast->getSrcTy(); 1493 1494 // Move the logic operation ahead of a zext or sext if the constant is 1495 // unchanged in the smaller source type. Performing the logic in a smaller 1496 // type may provide more information to later folds, and the smaller logic 1497 // instruction may be cheaper (particularly in the case of vectors). 1498 Value *X; 1499 if (match(Cast, m_OneUse(m_ZExt(m_Value(X))))) { 1500 Constant *TruncC = ConstantExpr::getTrunc(C, SrcTy); 1501 Constant *ZextTruncC = ConstantExpr::getZExt(TruncC, DestTy); 1502 if (ZextTruncC == C) { 1503 // LogicOpc (zext X), C --> zext (LogicOpc X, C) 1504 Value *NewOp = Builder.CreateBinOp(LogicOpc, X, TruncC); 1505 return new ZExtInst(NewOp, DestTy); 1506 } 1507 } 1508 1509 if (match(Cast, m_OneUse(m_SExt(m_Value(X))))) { 1510 Constant *TruncC = ConstantExpr::getTrunc(C, SrcTy); 1511 Constant *SextTruncC = ConstantExpr::getSExt(TruncC, DestTy); 1512 if (SextTruncC == C) { 1513 // LogicOpc (sext X), C --> sext (LogicOpc X, C) 1514 Value *NewOp = Builder.CreateBinOp(LogicOpc, X, TruncC); 1515 return new SExtInst(NewOp, DestTy); 1516 } 1517 } 1518 1519 return nullptr; 1520 } 1521 1522 /// Fold {and,or,xor} (cast X), Y. 1523 Instruction *InstCombinerImpl::foldCastedBitwiseLogic(BinaryOperator &I) { 1524 auto LogicOpc = I.getOpcode(); 1525 assert(I.isBitwiseLogicOp() && "Unexpected opcode for bitwise logic folding"); 1526 1527 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1528 CastInst *Cast0 = dyn_cast<CastInst>(Op0); 1529 if (!Cast0) 1530 return nullptr; 1531 1532 // This must be a cast from an integer or integer vector source type to allow 1533 // transformation of the logic operation to the source type. 1534 Type *DestTy = I.getType(); 1535 Type *SrcTy = Cast0->getSrcTy(); 1536 if (!SrcTy->isIntOrIntVectorTy()) 1537 return nullptr; 1538 1539 if (Instruction *Ret = foldLogicCastConstant(I, Cast0, Builder)) 1540 return Ret; 1541 1542 CastInst *Cast1 = dyn_cast<CastInst>(Op1); 1543 if (!Cast1) 1544 return nullptr; 1545 1546 // Both operands of the logic operation are casts. The casts must be of the 1547 // same type for reduction. 1548 auto CastOpcode = Cast0->getOpcode(); 1549 if (CastOpcode != Cast1->getOpcode() || SrcTy != Cast1->getSrcTy()) 1550 return nullptr; 1551 1552 Value *Cast0Src = Cast0->getOperand(0); 1553 Value *Cast1Src = Cast1->getOperand(0); 1554 1555 // fold logic(cast(A), cast(B)) -> cast(logic(A, B)) 1556 if (shouldOptimizeCast(Cast0) && shouldOptimizeCast(Cast1)) { 1557 Value *NewOp = Builder.CreateBinOp(LogicOpc, Cast0Src, Cast1Src, 1558 I.getName()); 1559 return CastInst::Create(CastOpcode, NewOp, DestTy); 1560 } 1561 1562 // For now, only 'and'/'or' have optimizations after this. 1563 if (LogicOpc == Instruction::Xor) 1564 return nullptr; 1565 1566 // If this is logic(cast(icmp), cast(icmp)), try to fold this even if the 1567 // cast is otherwise not optimizable. This happens for vector sexts. 1568 ICmpInst *ICmp0 = dyn_cast<ICmpInst>(Cast0Src); 1569 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(Cast1Src); 1570 if (ICmp0 && ICmp1) { 1571 Value *Res = LogicOpc == Instruction::And ? foldAndOfICmps(ICmp0, ICmp1, I) 1572 : foldOrOfICmps(ICmp0, ICmp1, I); 1573 if (Res) 1574 return CastInst::Create(CastOpcode, Res, DestTy); 1575 return nullptr; 1576 } 1577 1578 // If this is logic(cast(fcmp), cast(fcmp)), try to fold this even if the 1579 // cast is otherwise not optimizable. This happens for vector sexts. 1580 FCmpInst *FCmp0 = dyn_cast<FCmpInst>(Cast0Src); 1581 FCmpInst *FCmp1 = dyn_cast<FCmpInst>(Cast1Src); 1582 if (FCmp0 && FCmp1) 1583 if (Value *R = foldLogicOfFCmps(FCmp0, FCmp1, LogicOpc == Instruction::And)) 1584 return CastInst::Create(CastOpcode, R, DestTy); 1585 1586 return nullptr; 1587 } 1588 1589 static Instruction *foldAndToXor(BinaryOperator &I, 1590 InstCombiner::BuilderTy &Builder) { 1591 assert(I.getOpcode() == Instruction::And); 1592 Value *Op0 = I.getOperand(0); 1593 Value *Op1 = I.getOperand(1); 1594 Value *A, *B; 1595 1596 // Operand complexity canonicalization guarantees that the 'or' is Op0. 1597 // (A | B) & ~(A & B) --> A ^ B 1598 // (A | B) & ~(B & A) --> A ^ B 1599 if (match(&I, m_BinOp(m_Or(m_Value(A), m_Value(B)), 1600 m_Not(m_c_And(m_Deferred(A), m_Deferred(B)))))) 1601 return BinaryOperator::CreateXor(A, B); 1602 1603 // (A | ~B) & (~A | B) --> ~(A ^ B) 1604 // (A | ~B) & (B | ~A) --> ~(A ^ B) 1605 // (~B | A) & (~A | B) --> ~(A ^ B) 1606 // (~B | A) & (B | ~A) --> ~(A ^ B) 1607 if (Op0->hasOneUse() || Op1->hasOneUse()) 1608 if (match(&I, m_BinOp(m_c_Or(m_Value(A), m_Not(m_Value(B))), 1609 m_c_Or(m_Not(m_Deferred(A)), m_Deferred(B))))) 1610 return BinaryOperator::CreateNot(Builder.CreateXor(A, B)); 1611 1612 return nullptr; 1613 } 1614 1615 static Instruction *foldOrToXor(BinaryOperator &I, 1616 InstCombiner::BuilderTy &Builder) { 1617 assert(I.getOpcode() == Instruction::Or); 1618 Value *Op0 = I.getOperand(0); 1619 Value *Op1 = I.getOperand(1); 1620 Value *A, *B; 1621 1622 // Operand complexity canonicalization guarantees that the 'and' is Op0. 1623 // (A & B) | ~(A | B) --> ~(A ^ B) 1624 // (A & B) | ~(B | A) --> ~(A ^ B) 1625 if (Op0->hasOneUse() || Op1->hasOneUse()) 1626 if (match(Op0, m_And(m_Value(A), m_Value(B))) && 1627 match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B))))) 1628 return BinaryOperator::CreateNot(Builder.CreateXor(A, B)); 1629 1630 // Operand complexity canonicalization guarantees that the 'xor' is Op0. 1631 // (A ^ B) | ~(A | B) --> ~(A & B) 1632 // (A ^ B) | ~(B | A) --> ~(A & B) 1633 if (Op0->hasOneUse() || Op1->hasOneUse()) 1634 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) && 1635 match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B))))) 1636 return BinaryOperator::CreateNot(Builder.CreateAnd(A, B)); 1637 1638 // (A & ~B) | (~A & B) --> A ^ B 1639 // (A & ~B) | (B & ~A) --> A ^ B 1640 // (~B & A) | (~A & B) --> A ^ B 1641 // (~B & A) | (B & ~A) --> A ^ B 1642 if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) && 1643 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B)))) 1644 return BinaryOperator::CreateXor(A, B); 1645 1646 return nullptr; 1647 } 1648 1649 /// Return true if a constant shift amount is always less than the specified 1650 /// bit-width. If not, the shift could create poison in the narrower type. 1651 static bool canNarrowShiftAmt(Constant *C, unsigned BitWidth) { 1652 APInt Threshold(C->getType()->getScalarSizeInBits(), BitWidth); 1653 return match(C, m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, Threshold)); 1654 } 1655 1656 /// Try to use narrower ops (sink zext ops) for an 'and' with binop operand and 1657 /// a common zext operand: and (binop (zext X), C), (zext X). 1658 Instruction *InstCombinerImpl::narrowMaskedBinOp(BinaryOperator &And) { 1659 // This transform could also apply to {or, and, xor}, but there are better 1660 // folds for those cases, so we don't expect those patterns here. AShr is not 1661 // handled because it should always be transformed to LShr in this sequence. 1662 // The subtract transform is different because it has a constant on the left. 1663 // Add/mul commute the constant to RHS; sub with constant RHS becomes add. 1664 Value *Op0 = And.getOperand(0), *Op1 = And.getOperand(1); 1665 Constant *C; 1666 if (!match(Op0, m_OneUse(m_Add(m_Specific(Op1), m_Constant(C)))) && 1667 !match(Op0, m_OneUse(m_Mul(m_Specific(Op1), m_Constant(C)))) && 1668 !match(Op0, m_OneUse(m_LShr(m_Specific(Op1), m_Constant(C)))) && 1669 !match(Op0, m_OneUse(m_Shl(m_Specific(Op1), m_Constant(C)))) && 1670 !match(Op0, m_OneUse(m_Sub(m_Constant(C), m_Specific(Op1))))) 1671 return nullptr; 1672 1673 Value *X; 1674 if (!match(Op1, m_ZExt(m_Value(X))) || Op1->hasNUsesOrMore(3)) 1675 return nullptr; 1676 1677 Type *Ty = And.getType(); 1678 if (!isa<VectorType>(Ty) && !shouldChangeType(Ty, X->getType())) 1679 return nullptr; 1680 1681 // If we're narrowing a shift, the shift amount must be safe (less than the 1682 // width) in the narrower type. If the shift amount is greater, instsimplify 1683 // usually handles that case, but we can't guarantee/assert it. 1684 Instruction::BinaryOps Opc = cast<BinaryOperator>(Op0)->getOpcode(); 1685 if (Opc == Instruction::LShr || Opc == Instruction::Shl) 1686 if (!canNarrowShiftAmt(C, X->getType()->getScalarSizeInBits())) 1687 return nullptr; 1688 1689 // and (sub C, (zext X)), (zext X) --> zext (and (sub C', X), X) 1690 // and (binop (zext X), C), (zext X) --> zext (and (binop X, C'), X) 1691 Value *NewC = ConstantExpr::getTrunc(C, X->getType()); 1692 Value *NewBO = Opc == Instruction::Sub ? Builder.CreateBinOp(Opc, NewC, X) 1693 : Builder.CreateBinOp(Opc, X, NewC); 1694 return new ZExtInst(Builder.CreateAnd(NewBO, X), Ty); 1695 } 1696 1697 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches 1698 // here. We should standardize that construct where it is needed or choose some 1699 // other way to ensure that commutated variants of patterns are not missed. 1700 Instruction *InstCombinerImpl::visitAnd(BinaryOperator &I) { 1701 Type *Ty = I.getType(); 1702 1703 if (Value *V = SimplifyAndInst(I.getOperand(0), I.getOperand(1), 1704 SQ.getWithInstruction(&I))) 1705 return replaceInstUsesWith(I, V); 1706 1707 if (SimplifyAssociativeOrCommutative(I)) 1708 return &I; 1709 1710 if (Instruction *X = foldVectorBinop(I)) 1711 return X; 1712 1713 // See if we can simplify any instructions used by the instruction whose sole 1714 // purpose is to compute bits we don't care about. 1715 if (SimplifyDemandedInstructionBits(I)) 1716 return &I; 1717 1718 // Do this before using distributive laws to catch simple and/or/not patterns. 1719 if (Instruction *Xor = foldAndToXor(I, Builder)) 1720 return Xor; 1721 1722 // (A|B)&(A|C) -> A|(B&C) etc 1723 if (Value *V = SimplifyUsingDistributiveLaws(I)) 1724 return replaceInstUsesWith(I, V); 1725 1726 if (Value *V = SimplifyBSwap(I, Builder)) 1727 return replaceInstUsesWith(I, V); 1728 1729 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1730 1731 Value *X, *Y; 1732 if (match(Op0, m_OneUse(m_LogicalShift(m_One(), m_Value(X)))) && 1733 match(Op1, m_One())) { 1734 // (1 << X) & 1 --> zext(X == 0) 1735 // (1 >> X) & 1 --> zext(X == 0) 1736 Value *IsZero = Builder.CreateICmpEQ(X, ConstantInt::get(Ty, 0)); 1737 return new ZExtInst(IsZero, Ty); 1738 } 1739 1740 const APInt *C; 1741 if (match(Op1, m_APInt(C))) { 1742 const APInt *XorC; 1743 if (match(Op0, m_OneUse(m_Xor(m_Value(X), m_APInt(XorC))))) { 1744 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2) 1745 Constant *NewC = ConstantInt::get(Ty, *C & *XorC); 1746 Value *And = Builder.CreateAnd(X, Op1); 1747 And->takeName(Op0); 1748 return BinaryOperator::CreateXor(And, NewC); 1749 } 1750 1751 const APInt *OrC; 1752 if (match(Op0, m_OneUse(m_Or(m_Value(X), m_APInt(OrC))))) { 1753 // (X | C1) & C2 --> (X & C2^(C1&C2)) | (C1&C2) 1754 // NOTE: This reduces the number of bits set in the & mask, which 1755 // can expose opportunities for store narrowing for scalars. 1756 // NOTE: SimplifyDemandedBits should have already removed bits from C1 1757 // that aren't set in C2. Meaning we can replace (C1&C2) with C1 in 1758 // above, but this feels safer. 1759 APInt Together = *C & *OrC; 1760 Value *And = Builder.CreateAnd(X, ConstantInt::get(Ty, Together ^ *C)); 1761 And->takeName(Op0); 1762 return BinaryOperator::CreateOr(And, ConstantInt::get(Ty, Together)); 1763 } 1764 1765 // If the mask is only needed on one incoming arm, push the 'and' op up. 1766 if (match(Op0, m_OneUse(m_Xor(m_Value(X), m_Value(Y)))) || 1767 match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) { 1768 APInt NotAndMask(~(*C)); 1769 BinaryOperator::BinaryOps BinOp = cast<BinaryOperator>(Op0)->getOpcode(); 1770 if (MaskedValueIsZero(X, NotAndMask, 0, &I)) { 1771 // Not masking anything out for the LHS, move mask to RHS. 1772 // and ({x}or X, Y), C --> {x}or X, (and Y, C) 1773 Value *NewRHS = Builder.CreateAnd(Y, Op1, Y->getName() + ".masked"); 1774 return BinaryOperator::Create(BinOp, X, NewRHS); 1775 } 1776 if (!isa<Constant>(Y) && MaskedValueIsZero(Y, NotAndMask, 0, &I)) { 1777 // Not masking anything out for the RHS, move mask to LHS. 1778 // and ({x}or X, Y), C --> {x}or (and X, C), Y 1779 Value *NewLHS = Builder.CreateAnd(X, Op1, X->getName() + ".masked"); 1780 return BinaryOperator::Create(BinOp, NewLHS, Y); 1781 } 1782 } 1783 1784 unsigned Width = Ty->getScalarSizeInBits(); 1785 const APInt *ShiftC; 1786 if (match(Op0, m_OneUse(m_SExt(m_AShr(m_Value(X), m_APInt(ShiftC)))))) { 1787 if (*C == APInt::getLowBitsSet(Width, Width - ShiftC->getZExtValue())) { 1788 // We are clearing high bits that were potentially set by sext+ashr: 1789 // and (sext (ashr X, ShiftC)), C --> lshr (sext X), ShiftC 1790 Value *Sext = Builder.CreateSExt(X, Ty); 1791 Constant *ShAmtC = ConstantInt::get(Ty, ShiftC->zext(Width)); 1792 return BinaryOperator::CreateLShr(Sext, ShAmtC); 1793 } 1794 } 1795 1796 const APInt *AddC; 1797 if (match(Op0, m_Add(m_Value(X), m_APInt(AddC)))) { 1798 // If we add zeros to every bit below a mask, the add has no effect: 1799 // (X + AddC) & LowMaskC --> X & LowMaskC 1800 unsigned Ctlz = C->countLeadingZeros(); 1801 APInt LowMask(APInt::getLowBitsSet(Width, Width - Ctlz)); 1802 if ((*AddC & LowMask).isNullValue()) 1803 return BinaryOperator::CreateAnd(X, Op1); 1804 1805 // If we are masking the result of the add down to exactly one bit and 1806 // the constant we are adding has no bits set below that bit, then the 1807 // add is flipping a single bit. Example: 1808 // (X + 4) & 4 --> (X & 4) ^ 4 1809 if (Op0->hasOneUse() && C->isPowerOf2() && (*AddC & (*C - 1)) == 0) { 1810 assert((*C & *AddC) != 0 && "Expected common bit"); 1811 Value *NewAnd = Builder.CreateAnd(X, Op1); 1812 return BinaryOperator::CreateXor(NewAnd, Op1); 1813 } 1814 } 1815 } 1816 1817 ConstantInt *AndRHS; 1818 if (match(Op1, m_ConstantInt(AndRHS))) { 1819 const APInt &AndRHSMask = AndRHS->getValue(); 1820 1821 // Optimize a variety of ((val OP C1) & C2) combinations... 1822 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) { 1823 // ((C1 OP zext(X)) & C2) -> zext((C1-X) & C2) if C2 fits in the bitwidth 1824 // of X and OP behaves well when given trunc(C1) and X. 1825 // TODO: Do this for vectors by using m_APInt instead of m_ConstantInt. 1826 switch (Op0I->getOpcode()) { 1827 default: 1828 break; 1829 case Instruction::Xor: 1830 case Instruction::Or: 1831 case Instruction::Mul: 1832 case Instruction::Add: 1833 case Instruction::Sub: 1834 Value *X; 1835 ConstantInt *C1; 1836 // TODO: The one use restrictions could be relaxed a little if the AND 1837 // is going to be removed. 1838 if (match(Op0I, m_OneUse(m_c_BinOp(m_OneUse(m_ZExt(m_Value(X))), 1839 m_ConstantInt(C1))))) { 1840 if (AndRHSMask.isIntN(X->getType()->getScalarSizeInBits())) { 1841 auto *TruncC1 = ConstantExpr::getTrunc(C1, X->getType()); 1842 Value *BinOp; 1843 Value *Op0LHS = Op0I->getOperand(0); 1844 if (isa<ZExtInst>(Op0LHS)) 1845 BinOp = Builder.CreateBinOp(Op0I->getOpcode(), X, TruncC1); 1846 else 1847 BinOp = Builder.CreateBinOp(Op0I->getOpcode(), TruncC1, X); 1848 auto *TruncC2 = ConstantExpr::getTrunc(AndRHS, X->getType()); 1849 auto *And = Builder.CreateAnd(BinOp, TruncC2); 1850 return new ZExtInst(And, Ty); 1851 } 1852 } 1853 } 1854 } 1855 } 1856 1857 if (match(&I, m_And(m_OneUse(m_Shl(m_ZExt(m_Value(X)), m_Value(Y))), 1858 m_SignMask())) && 1859 match(Y, m_SpecificInt_ICMP( 1860 ICmpInst::Predicate::ICMP_EQ, 1861 APInt(Ty->getScalarSizeInBits(), 1862 Ty->getScalarSizeInBits() - 1863 X->getType()->getScalarSizeInBits())))) { 1864 auto *SExt = Builder.CreateSExt(X, Ty, X->getName() + ".signext"); 1865 auto *SanitizedSignMask = cast<Constant>(Op1); 1866 // We must be careful with the undef elements of the sign bit mask, however: 1867 // the mask elt can be undef iff the shift amount for that lane was undef, 1868 // otherwise we need to sanitize undef masks to zero. 1869 SanitizedSignMask = Constant::replaceUndefsWith( 1870 SanitizedSignMask, ConstantInt::getNullValue(Ty->getScalarType())); 1871 SanitizedSignMask = 1872 Constant::mergeUndefsWith(SanitizedSignMask, cast<Constant>(Y)); 1873 return BinaryOperator::CreateAnd(SExt, SanitizedSignMask); 1874 } 1875 1876 if (Instruction *Z = narrowMaskedBinOp(I)) 1877 return Z; 1878 1879 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I)) 1880 return FoldedLogic; 1881 1882 if (Instruction *DeMorgan = matchDeMorgansLaws(I, Builder)) 1883 return DeMorgan; 1884 1885 { 1886 Value *A, *B, *C; 1887 // A & (A ^ B) --> A & ~B 1888 if (match(Op1, m_OneUse(m_c_Xor(m_Specific(Op0), m_Value(B))))) 1889 return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(B)); 1890 // (A ^ B) & A --> A & ~B 1891 if (match(Op0, m_OneUse(m_c_Xor(m_Specific(Op1), m_Value(B))))) 1892 return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(B)); 1893 1894 // A & ~(A ^ B) --> A & B 1895 if (match(Op1, m_Not(m_c_Xor(m_Specific(Op0), m_Value(B))))) 1896 return BinaryOperator::CreateAnd(Op0, B); 1897 // ~(A ^ B) & A --> A & B 1898 if (match(Op0, m_Not(m_c_Xor(m_Specific(Op1), m_Value(B))))) 1899 return BinaryOperator::CreateAnd(Op1, B); 1900 1901 // (A ^ B) & ((B ^ C) ^ A) -> (A ^ B) & ~C 1902 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) 1903 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A)))) 1904 if (Op1->hasOneUse() || isFreeToInvert(C, C->hasOneUse())) 1905 return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(C)); 1906 1907 // ((A ^ C) ^ B) & (B ^ A) -> (B ^ A) & ~C 1908 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B)))) 1909 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A)))) 1910 if (Op0->hasOneUse() || isFreeToInvert(C, C->hasOneUse())) 1911 return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(C)); 1912 1913 // (A | B) & ((~A) ^ B) -> (A & B) 1914 // (A | B) & (B ^ (~A)) -> (A & B) 1915 // (B | A) & ((~A) ^ B) -> (A & B) 1916 // (B | A) & (B ^ (~A)) -> (A & B) 1917 if (match(Op1, m_c_Xor(m_Not(m_Value(A)), m_Value(B))) && 1918 match(Op0, m_c_Or(m_Specific(A), m_Specific(B)))) 1919 return BinaryOperator::CreateAnd(A, B); 1920 1921 // ((~A) ^ B) & (A | B) -> (A & B) 1922 // ((~A) ^ B) & (B | A) -> (A & B) 1923 // (B ^ (~A)) & (A | B) -> (A & B) 1924 // (B ^ (~A)) & (B | A) -> (A & B) 1925 if (match(Op0, m_c_Xor(m_Not(m_Value(A)), m_Value(B))) && 1926 match(Op1, m_c_Or(m_Specific(A), m_Specific(B)))) 1927 return BinaryOperator::CreateAnd(A, B); 1928 } 1929 1930 { 1931 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0); 1932 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1); 1933 if (LHS && RHS) 1934 if (Value *Res = foldAndOfICmps(LHS, RHS, I)) 1935 return replaceInstUsesWith(I, Res); 1936 1937 // TODO: Make this recursive; it's a little tricky because an arbitrary 1938 // number of 'and' instructions might have to be created. 1939 if (LHS && match(Op1, m_OneUse(m_And(m_Value(X), m_Value(Y))))) { 1940 if (auto *Cmp = dyn_cast<ICmpInst>(X)) 1941 if (Value *Res = foldAndOfICmps(LHS, Cmp, I)) 1942 return replaceInstUsesWith(I, Builder.CreateAnd(Res, Y)); 1943 if (auto *Cmp = dyn_cast<ICmpInst>(Y)) 1944 if (Value *Res = foldAndOfICmps(LHS, Cmp, I)) 1945 return replaceInstUsesWith(I, Builder.CreateAnd(Res, X)); 1946 } 1947 if (RHS && match(Op0, m_OneUse(m_And(m_Value(X), m_Value(Y))))) { 1948 if (auto *Cmp = dyn_cast<ICmpInst>(X)) 1949 if (Value *Res = foldAndOfICmps(Cmp, RHS, I)) 1950 return replaceInstUsesWith(I, Builder.CreateAnd(Res, Y)); 1951 if (auto *Cmp = dyn_cast<ICmpInst>(Y)) 1952 if (Value *Res = foldAndOfICmps(Cmp, RHS, I)) 1953 return replaceInstUsesWith(I, Builder.CreateAnd(Res, X)); 1954 } 1955 } 1956 1957 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) 1958 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) 1959 if (Value *Res = foldLogicOfFCmps(LHS, RHS, true)) 1960 return replaceInstUsesWith(I, Res); 1961 1962 if (Instruction *FoldedFCmps = reassociateFCmps(I, Builder)) 1963 return FoldedFCmps; 1964 1965 if (Instruction *CastedAnd = foldCastedBitwiseLogic(I)) 1966 return CastedAnd; 1967 1968 // and(sext(A), B) / and(B, sext(A)) --> A ? B : 0, where A is i1 or <N x i1>. 1969 Value *A; 1970 if (match(Op0, m_OneUse(m_SExt(m_Value(A)))) && 1971 A->getType()->isIntOrIntVectorTy(1)) 1972 return SelectInst::Create(A, Op1, Constant::getNullValue(Ty)); 1973 if (match(Op1, m_OneUse(m_SExt(m_Value(A)))) && 1974 A->getType()->isIntOrIntVectorTy(1)) 1975 return SelectInst::Create(A, Op0, Constant::getNullValue(Ty)); 1976 1977 // and(ashr(subNSW(Y, X), ScalarSizeInBits(Y)-1), X) --> X s> Y ? X : 0. 1978 if (match(&I, m_c_And(m_OneUse(m_AShr( 1979 m_NSWSub(m_Value(Y), m_Value(X)), 1980 m_SpecificInt(Ty->getScalarSizeInBits() - 1))), 1981 m_Deferred(X)))) { 1982 Value *NewICmpInst = Builder.CreateICmpSGT(X, Y); 1983 return SelectInst::Create(NewICmpInst, X, ConstantInt::getNullValue(Ty)); 1984 } 1985 1986 // (~x) & y --> ~(x | (~y)) iff that gets rid of inversions 1987 if (sinkNotIntoOtherHandOfAndOrOr(I)) 1988 return &I; 1989 1990 return nullptr; 1991 } 1992 1993 Instruction *InstCombinerImpl::matchBSwapOrBitReverse(BinaryOperator &Or, 1994 bool MatchBSwaps, 1995 bool MatchBitReversals) { 1996 assert(Or.getOpcode() == Instruction::Or && "bswap requires an 'or'"); 1997 Value *Op0 = Or.getOperand(0), *Op1 = Or.getOperand(1); 1998 1999 // Look through zero extends. 2000 if (Instruction *Ext = dyn_cast<ZExtInst>(Op0)) 2001 Op0 = Ext->getOperand(0); 2002 2003 if (Instruction *Ext = dyn_cast<ZExtInst>(Op1)) 2004 Op1 = Ext->getOperand(0); 2005 2006 // (A | B) | C and A | (B | C) -> bswap if possible. 2007 bool OrWithOrs = match(Op0, m_Or(m_Value(), m_Value())) || 2008 match(Op1, m_Or(m_Value(), m_Value())); 2009 2010 // (A >> B) | C and (A << B) | C -> bswap if possible. 2011 bool OrWithShifts = match(Op0, m_LogicalShift(m_Value(), m_Value())) || 2012 match(Op1, m_LogicalShift(m_Value(), m_Value())); 2013 2014 // (A & B) | C and A | (B & C) -> bswap if possible. 2015 bool OrWithAnds = match(Op0, m_And(m_Value(), m_Value())) || 2016 match(Op1, m_And(m_Value(), m_Value())); 2017 2018 // fshl(A,B,C) | D and A | fshl(B,C,D) -> bswap if possible. 2019 // fshr(A,B,C) | D and A | fshr(B,C,D) -> bswap if possible. 2020 bool OrWithFunnels = match(Op0, m_FShl(m_Value(), m_Value(), m_Value())) || 2021 match(Op0, m_FShr(m_Value(), m_Value(), m_Value())) || 2022 match(Op0, m_FShl(m_Value(), m_Value(), m_Value())) || 2023 match(Op0, m_FShr(m_Value(), m_Value(), m_Value())); 2024 2025 // TODO: Do we need all these filtering checks or should we just rely on 2026 // recognizeBSwapOrBitReverseIdiom + collectBitParts to reject them quickly? 2027 if (!OrWithOrs && !OrWithShifts && !OrWithAnds && !OrWithFunnels) 2028 return nullptr; 2029 2030 SmallVector<Instruction *, 4> Insts; 2031 if (!recognizeBSwapOrBitReverseIdiom(&Or, MatchBSwaps, MatchBitReversals, 2032 Insts)) 2033 return nullptr; 2034 Instruction *LastInst = Insts.pop_back_val(); 2035 LastInst->removeFromParent(); 2036 2037 for (auto *Inst : Insts) 2038 Worklist.push(Inst); 2039 return LastInst; 2040 } 2041 2042 /// Match UB-safe variants of the funnel shift intrinsic. 2043 static Instruction *matchFunnelShift(Instruction &Or, InstCombinerImpl &IC) { 2044 // TODO: Can we reduce the code duplication between this and the related 2045 // rotate matching code under visitSelect and visitTrunc? 2046 unsigned Width = Or.getType()->getScalarSizeInBits(); 2047 2048 // First, find an or'd pair of opposite shifts: 2049 // or (lshr ShVal0, ShAmt0), (shl ShVal1, ShAmt1) 2050 BinaryOperator *Or0, *Or1; 2051 if (!match(Or.getOperand(0), m_BinOp(Or0)) || 2052 !match(Or.getOperand(1), m_BinOp(Or1))) 2053 return nullptr; 2054 2055 Value *ShVal0, *ShVal1, *ShAmt0, *ShAmt1; 2056 if (!match(Or0, m_OneUse(m_LogicalShift(m_Value(ShVal0), m_Value(ShAmt0)))) || 2057 !match(Or1, m_OneUse(m_LogicalShift(m_Value(ShVal1), m_Value(ShAmt1)))) || 2058 Or0->getOpcode() == Or1->getOpcode()) 2059 return nullptr; 2060 2061 // Canonicalize to or(shl(ShVal0, ShAmt0), lshr(ShVal1, ShAmt1)). 2062 if (Or0->getOpcode() == BinaryOperator::LShr) { 2063 std::swap(Or0, Or1); 2064 std::swap(ShVal0, ShVal1); 2065 std::swap(ShAmt0, ShAmt1); 2066 } 2067 assert(Or0->getOpcode() == BinaryOperator::Shl && 2068 Or1->getOpcode() == BinaryOperator::LShr && 2069 "Illegal or(shift,shift) pair"); 2070 2071 // Match the shift amount operands for a funnel shift pattern. This always 2072 // matches a subtraction on the R operand. 2073 auto matchShiftAmount = [&](Value *L, Value *R, unsigned Width) -> Value * { 2074 // Check for constant shift amounts that sum to the bitwidth. 2075 const APInt *LI, *RI; 2076 if (match(L, m_APIntAllowUndef(LI)) && match(R, m_APIntAllowUndef(RI))) 2077 if (LI->ult(Width) && RI->ult(Width) && (*LI + *RI) == Width) 2078 return ConstantInt::get(L->getType(), *LI); 2079 2080 Constant *LC, *RC; 2081 if (match(L, m_Constant(LC)) && match(R, m_Constant(RC)) && 2082 match(L, m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, APInt(Width, Width))) && 2083 match(R, m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, APInt(Width, Width))) && 2084 match(ConstantExpr::getAdd(LC, RC), m_SpecificIntAllowUndef(Width))) 2085 return ConstantExpr::mergeUndefsWith(LC, RC); 2086 2087 // (shl ShVal, X) | (lshr ShVal, (Width - x)) iff X < Width. 2088 // We limit this to X < Width in case the backend re-expands the intrinsic, 2089 // and has to reintroduce a shift modulo operation (InstCombine might remove 2090 // it after this fold). This still doesn't guarantee that the final codegen 2091 // will match this original pattern. 2092 if (match(R, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(L))))) { 2093 KnownBits KnownL = IC.computeKnownBits(L, /*Depth*/ 0, &Or); 2094 return KnownL.getMaxValue().ult(Width) ? L : nullptr; 2095 } 2096 2097 // For non-constant cases, the following patterns currently only work for 2098 // rotation patterns. 2099 // TODO: Add general funnel-shift compatible patterns. 2100 if (ShVal0 != ShVal1) 2101 return nullptr; 2102 2103 // For non-constant cases we don't support non-pow2 shift masks. 2104 // TODO: Is it worth matching urem as well? 2105 if (!isPowerOf2_32(Width)) 2106 return nullptr; 2107 2108 // The shift amount may be masked with negation: 2109 // (shl ShVal, (X & (Width - 1))) | (lshr ShVal, ((-X) & (Width - 1))) 2110 Value *X; 2111 unsigned Mask = Width - 1; 2112 if (match(L, m_And(m_Value(X), m_SpecificInt(Mask))) && 2113 match(R, m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask)))) 2114 return X; 2115 2116 // Similar to above, but the shift amount may be extended after masking, 2117 // so return the extended value as the parameter for the intrinsic. 2118 if (match(L, m_ZExt(m_And(m_Value(X), m_SpecificInt(Mask)))) && 2119 match(R, m_And(m_Neg(m_ZExt(m_And(m_Specific(X), m_SpecificInt(Mask)))), 2120 m_SpecificInt(Mask)))) 2121 return L; 2122 2123 if (match(L, m_ZExt(m_And(m_Value(X), m_SpecificInt(Mask)))) && 2124 match(R, m_ZExt(m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask))))) 2125 return L; 2126 2127 return nullptr; 2128 }; 2129 2130 Value *ShAmt = matchShiftAmount(ShAmt0, ShAmt1, Width); 2131 bool IsFshl = true; // Sub on LSHR. 2132 if (!ShAmt) { 2133 ShAmt = matchShiftAmount(ShAmt1, ShAmt0, Width); 2134 IsFshl = false; // Sub on SHL. 2135 } 2136 if (!ShAmt) 2137 return nullptr; 2138 2139 Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr; 2140 Function *F = Intrinsic::getDeclaration(Or.getModule(), IID, Or.getType()); 2141 return IntrinsicInst::Create(F, {ShVal0, ShVal1, ShAmt}); 2142 } 2143 2144 /// Attempt to combine or(zext(x),shl(zext(y),bw/2) concat packing patterns. 2145 static Instruction *matchOrConcat(Instruction &Or, 2146 InstCombiner::BuilderTy &Builder) { 2147 assert(Or.getOpcode() == Instruction::Or && "bswap requires an 'or'"); 2148 Value *Op0 = Or.getOperand(0), *Op1 = Or.getOperand(1); 2149 Type *Ty = Or.getType(); 2150 2151 unsigned Width = Ty->getScalarSizeInBits(); 2152 if ((Width & 1) != 0) 2153 return nullptr; 2154 unsigned HalfWidth = Width / 2; 2155 2156 // Canonicalize zext (lower half) to LHS. 2157 if (!isa<ZExtInst>(Op0)) 2158 std::swap(Op0, Op1); 2159 2160 // Find lower/upper half. 2161 Value *LowerSrc, *ShlVal, *UpperSrc; 2162 const APInt *C; 2163 if (!match(Op0, m_OneUse(m_ZExt(m_Value(LowerSrc)))) || 2164 !match(Op1, m_OneUse(m_Shl(m_Value(ShlVal), m_APInt(C)))) || 2165 !match(ShlVal, m_OneUse(m_ZExt(m_Value(UpperSrc))))) 2166 return nullptr; 2167 if (*C != HalfWidth || LowerSrc->getType() != UpperSrc->getType() || 2168 LowerSrc->getType()->getScalarSizeInBits() != HalfWidth) 2169 return nullptr; 2170 2171 auto ConcatIntrinsicCalls = [&](Intrinsic::ID id, Value *Lo, Value *Hi) { 2172 Value *NewLower = Builder.CreateZExt(Lo, Ty); 2173 Value *NewUpper = Builder.CreateZExt(Hi, Ty); 2174 NewUpper = Builder.CreateShl(NewUpper, HalfWidth); 2175 Value *BinOp = Builder.CreateOr(NewLower, NewUpper); 2176 Function *F = Intrinsic::getDeclaration(Or.getModule(), id, Ty); 2177 return Builder.CreateCall(F, BinOp); 2178 }; 2179 2180 // BSWAP: Push the concat down, swapping the lower/upper sources. 2181 // concat(bswap(x),bswap(y)) -> bswap(concat(x,y)) 2182 Value *LowerBSwap, *UpperBSwap; 2183 if (match(LowerSrc, m_BSwap(m_Value(LowerBSwap))) && 2184 match(UpperSrc, m_BSwap(m_Value(UpperBSwap)))) 2185 return ConcatIntrinsicCalls(Intrinsic::bswap, UpperBSwap, LowerBSwap); 2186 2187 // BITREVERSE: Push the concat down, swapping the lower/upper sources. 2188 // concat(bitreverse(x),bitreverse(y)) -> bitreverse(concat(x,y)) 2189 Value *LowerBRev, *UpperBRev; 2190 if (match(LowerSrc, m_BitReverse(m_Value(LowerBRev))) && 2191 match(UpperSrc, m_BitReverse(m_Value(UpperBRev)))) 2192 return ConcatIntrinsicCalls(Intrinsic::bitreverse, UpperBRev, LowerBRev); 2193 2194 return nullptr; 2195 } 2196 2197 /// If all elements of two constant vectors are 0/-1 and inverses, return true. 2198 static bool areInverseVectorBitmasks(Constant *C1, Constant *C2) { 2199 unsigned NumElts = cast<FixedVectorType>(C1->getType())->getNumElements(); 2200 for (unsigned i = 0; i != NumElts; ++i) { 2201 Constant *EltC1 = C1->getAggregateElement(i); 2202 Constant *EltC2 = C2->getAggregateElement(i); 2203 if (!EltC1 || !EltC2) 2204 return false; 2205 2206 // One element must be all ones, and the other must be all zeros. 2207 if (!((match(EltC1, m_Zero()) && match(EltC2, m_AllOnes())) || 2208 (match(EltC2, m_Zero()) && match(EltC1, m_AllOnes())))) 2209 return false; 2210 } 2211 return true; 2212 } 2213 2214 /// We have an expression of the form (A & C) | (B & D). If A is a scalar or 2215 /// vector composed of all-zeros or all-ones values and is the bitwise 'not' of 2216 /// B, it can be used as the condition operand of a select instruction. 2217 Value *InstCombinerImpl::getSelectCondition(Value *A, Value *B) { 2218 // Step 1: We may have peeked through bitcasts in the caller. 2219 // Exit immediately if we don't have (vector) integer types. 2220 Type *Ty = A->getType(); 2221 if (!Ty->isIntOrIntVectorTy() || !B->getType()->isIntOrIntVectorTy()) 2222 return nullptr; 2223 2224 // Step 2: We need 0 or all-1's bitmasks. 2225 if (ComputeNumSignBits(A) != Ty->getScalarSizeInBits()) 2226 return nullptr; 2227 2228 // Step 3: If B is the 'not' value of A, we have our answer. 2229 if (match(A, m_Not(m_Specific(B)))) { 2230 // If these are scalars or vectors of i1, A can be used directly. 2231 if (Ty->isIntOrIntVectorTy(1)) 2232 return A; 2233 return Builder.CreateTrunc(A, CmpInst::makeCmpResultType(Ty)); 2234 } 2235 2236 // If both operands are constants, see if the constants are inverse bitmasks. 2237 Constant *AConst, *BConst; 2238 if (match(A, m_Constant(AConst)) && match(B, m_Constant(BConst))) 2239 if (AConst == ConstantExpr::getNot(BConst)) 2240 return Builder.CreateZExtOrTrunc(A, CmpInst::makeCmpResultType(Ty)); 2241 2242 // Look for more complex patterns. The 'not' op may be hidden behind various 2243 // casts. Look through sexts and bitcasts to find the booleans. 2244 Value *Cond; 2245 Value *NotB; 2246 if (match(A, m_SExt(m_Value(Cond))) && 2247 Cond->getType()->isIntOrIntVectorTy(1) && 2248 match(B, m_OneUse(m_Not(m_Value(NotB))))) { 2249 NotB = peekThroughBitcast(NotB, true); 2250 if (match(NotB, m_SExt(m_Specific(Cond)))) 2251 return Cond; 2252 } 2253 2254 // All scalar (and most vector) possibilities should be handled now. 2255 // Try more matches that only apply to non-splat constant vectors. 2256 if (!Ty->isVectorTy()) 2257 return nullptr; 2258 2259 // If both operands are xor'd with constants using the same sexted boolean 2260 // operand, see if the constants are inverse bitmasks. 2261 // TODO: Use ConstantExpr::getNot()? 2262 if (match(A, (m_Xor(m_SExt(m_Value(Cond)), m_Constant(AConst)))) && 2263 match(B, (m_Xor(m_SExt(m_Specific(Cond)), m_Constant(BConst)))) && 2264 Cond->getType()->isIntOrIntVectorTy(1) && 2265 areInverseVectorBitmasks(AConst, BConst)) { 2266 AConst = ConstantExpr::getTrunc(AConst, CmpInst::makeCmpResultType(Ty)); 2267 return Builder.CreateXor(Cond, AConst); 2268 } 2269 return nullptr; 2270 } 2271 2272 /// We have an expression of the form (A & C) | (B & D). Try to simplify this 2273 /// to "A' ? C : D", where A' is a boolean or vector of booleans. 2274 Value *InstCombinerImpl::matchSelectFromAndOr(Value *A, Value *C, Value *B, 2275 Value *D) { 2276 // The potential condition of the select may be bitcasted. In that case, look 2277 // through its bitcast and the corresponding bitcast of the 'not' condition. 2278 Type *OrigType = A->getType(); 2279 A = peekThroughBitcast(A, true); 2280 B = peekThroughBitcast(B, true); 2281 if (Value *Cond = getSelectCondition(A, B)) { 2282 // ((bc Cond) & C) | ((bc ~Cond) & D) --> bc (select Cond, (bc C), (bc D)) 2283 // The bitcasts will either all exist or all not exist. The builder will 2284 // not create unnecessary casts if the types already match. 2285 Value *BitcastC = Builder.CreateBitCast(C, A->getType()); 2286 Value *BitcastD = Builder.CreateBitCast(D, A->getType()); 2287 Value *Select = Builder.CreateSelect(Cond, BitcastC, BitcastD); 2288 return Builder.CreateBitCast(Select, OrigType); 2289 } 2290 2291 return nullptr; 2292 } 2293 2294 /// Fold (icmp)|(icmp) if possible. 2295 Value *InstCombinerImpl::foldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS, 2296 BinaryOperator &Or) { 2297 const SimplifyQuery Q = SQ.getWithInstruction(&Or); 2298 2299 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2) 2300 // if K1 and K2 are a one-bit mask. 2301 if (Value *V = foldAndOrOfICmpsOfAndWithPow2(LHS, RHS, Or)) 2302 return V; 2303 2304 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate(); 2305 Value *LHS0 = LHS->getOperand(0), *RHS0 = RHS->getOperand(0); 2306 Value *LHS1 = LHS->getOperand(1), *RHS1 = RHS->getOperand(1); 2307 auto *LHSC = dyn_cast<ConstantInt>(LHS1); 2308 auto *RHSC = dyn_cast<ConstantInt>(RHS1); 2309 2310 // Fold (icmp ult/ule (A + C1), C3) | (icmp ult/ule (A + C2), C3) 2311 // --> (icmp ult/ule ((A & ~(C1 ^ C2)) + max(C1, C2)), C3) 2312 // The original condition actually refers to the following two ranges: 2313 // [MAX_UINT-C1+1, MAX_UINT-C1+1+C3] and [MAX_UINT-C2+1, MAX_UINT-C2+1+C3] 2314 // We can fold these two ranges if: 2315 // 1) C1 and C2 is unsigned greater than C3. 2316 // 2) The two ranges are separated. 2317 // 3) C1 ^ C2 is one-bit mask. 2318 // 4) LowRange1 ^ LowRange2 and HighRange1 ^ HighRange2 are one-bit mask. 2319 // This implies all values in the two ranges differ by exactly one bit. 2320 if ((PredL == ICmpInst::ICMP_ULT || PredL == ICmpInst::ICMP_ULE) && 2321 PredL == PredR && LHSC && RHSC && LHS->hasOneUse() && RHS->hasOneUse() && 2322 LHSC->getType() == RHSC->getType() && 2323 LHSC->getValue() == (RHSC->getValue())) { 2324 2325 Value *AddOpnd; 2326 ConstantInt *LAddC, *RAddC; 2327 if (match(LHS0, m_Add(m_Value(AddOpnd), m_ConstantInt(LAddC))) && 2328 match(RHS0, m_Add(m_Specific(AddOpnd), m_ConstantInt(RAddC))) && 2329 LAddC->getValue().ugt(LHSC->getValue()) && 2330 RAddC->getValue().ugt(LHSC->getValue())) { 2331 2332 APInt DiffC = LAddC->getValue() ^ RAddC->getValue(); 2333 if (DiffC.isPowerOf2()) { 2334 ConstantInt *MaxAddC = nullptr; 2335 if (LAddC->getValue().ult(RAddC->getValue())) 2336 MaxAddC = RAddC; 2337 else 2338 MaxAddC = LAddC; 2339 2340 APInt RRangeLow = -RAddC->getValue(); 2341 APInt RRangeHigh = RRangeLow + LHSC->getValue(); 2342 APInt LRangeLow = -LAddC->getValue(); 2343 APInt LRangeHigh = LRangeLow + LHSC->getValue(); 2344 APInt LowRangeDiff = RRangeLow ^ LRangeLow; 2345 APInt HighRangeDiff = RRangeHigh ^ LRangeHigh; 2346 APInt RangeDiff = LRangeLow.sgt(RRangeLow) ? LRangeLow - RRangeLow 2347 : RRangeLow - LRangeLow; 2348 2349 if (LowRangeDiff.isPowerOf2() && LowRangeDiff == HighRangeDiff && 2350 RangeDiff.ugt(LHSC->getValue())) { 2351 Value *MaskC = ConstantInt::get(LAddC->getType(), ~DiffC); 2352 2353 Value *NewAnd = Builder.CreateAnd(AddOpnd, MaskC); 2354 Value *NewAdd = Builder.CreateAdd(NewAnd, MaxAddC); 2355 return Builder.CreateICmp(LHS->getPredicate(), NewAdd, LHSC); 2356 } 2357 } 2358 } 2359 } 2360 2361 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B) 2362 if (predicatesFoldable(PredL, PredR)) { 2363 if (LHS0 == RHS1 && LHS1 == RHS0) 2364 LHS->swapOperands(); 2365 if (LHS0 == RHS0 && LHS1 == RHS1) { 2366 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS); 2367 bool IsSigned = LHS->isSigned() || RHS->isSigned(); 2368 return getNewICmpValue(Code, IsSigned, LHS0, LHS1, Builder); 2369 } 2370 } 2371 2372 // handle (roughly): 2373 // (icmp ne (A & B), C) | (icmp ne (A & D), E) 2374 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, false, Builder)) 2375 return V; 2376 2377 if (LHS->hasOneUse() || RHS->hasOneUse()) { 2378 // (icmp eq B, 0) | (icmp ult A, B) -> (icmp ule A, B-1) 2379 // (icmp eq B, 0) | (icmp ugt B, A) -> (icmp ule A, B-1) 2380 Value *A = nullptr, *B = nullptr; 2381 if (PredL == ICmpInst::ICMP_EQ && match(LHS1, m_Zero())) { 2382 B = LHS0; 2383 if (PredR == ICmpInst::ICMP_ULT && LHS0 == RHS1) 2384 A = RHS0; 2385 else if (PredR == ICmpInst::ICMP_UGT && LHS0 == RHS0) 2386 A = RHS1; 2387 } 2388 // (icmp ult A, B) | (icmp eq B, 0) -> (icmp ule A, B-1) 2389 // (icmp ugt B, A) | (icmp eq B, 0) -> (icmp ule A, B-1) 2390 else if (PredR == ICmpInst::ICMP_EQ && match(RHS1, m_Zero())) { 2391 B = RHS0; 2392 if (PredL == ICmpInst::ICMP_ULT && RHS0 == LHS1) 2393 A = LHS0; 2394 else if (PredL == ICmpInst::ICMP_UGT && RHS0 == LHS0) 2395 A = LHS1; 2396 } 2397 if (A && B && B->getType()->isIntOrIntVectorTy()) 2398 return Builder.CreateICmp( 2399 ICmpInst::ICMP_UGE, 2400 Builder.CreateAdd(B, Constant::getAllOnesValue(B->getType())), A); 2401 } 2402 2403 if (Value *V = foldAndOrOfICmpsWithConstEq(LHS, RHS, Or, Builder, Q)) 2404 return V; 2405 if (Value *V = foldAndOrOfICmpsWithConstEq(RHS, LHS, Or, Builder, Q)) 2406 return V; 2407 2408 // E.g. (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n 2409 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/true)) 2410 return V; 2411 2412 // E.g. (icmp sgt x, n) | (icmp slt x, 0) --> icmp ugt x, n 2413 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/true)) 2414 return V; 2415 2416 if (Value *V = foldAndOrOfEqualityCmpsWithConstants(LHS, RHS, false, Builder)) 2417 return V; 2418 2419 if (Value *V = foldIsPowerOf2(LHS, RHS, false /* JoinedByAnd */, Builder)) 2420 return V; 2421 2422 if (Value *X = 2423 foldUnsignedUnderflowCheck(LHS, RHS, /*IsAnd=*/false, Q, Builder)) 2424 return X; 2425 if (Value *X = 2426 foldUnsignedUnderflowCheck(RHS, LHS, /*IsAnd=*/false, Q, Builder)) 2427 return X; 2428 2429 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0) 2430 // TODO: Remove this when foldLogOpOfMaskedICmps can handle vectors. 2431 if (PredL == ICmpInst::ICMP_NE && match(LHS1, m_Zero()) && 2432 PredR == ICmpInst::ICMP_NE && match(RHS1, m_Zero()) && 2433 LHS0->getType()->isIntOrIntVectorTy() && 2434 LHS0->getType() == RHS0->getType()) { 2435 Value *NewOr = Builder.CreateOr(LHS0, RHS0); 2436 return Builder.CreateICmp(PredL, NewOr, 2437 Constant::getNullValue(NewOr->getType())); 2438 } 2439 2440 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2). 2441 if (!LHSC || !RHSC) 2442 return nullptr; 2443 2444 // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1) 2445 // iff C2 + CA == C1. 2446 if (PredL == ICmpInst::ICMP_ULT && PredR == ICmpInst::ICMP_EQ) { 2447 ConstantInt *AddC; 2448 if (match(LHS0, m_Add(m_Specific(RHS0), m_ConstantInt(AddC)))) 2449 if (RHSC->getValue() + AddC->getValue() == LHSC->getValue()) 2450 return Builder.CreateICmpULE(LHS0, LHSC); 2451 } 2452 2453 // From here on, we only handle: 2454 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler. 2455 if (LHS0 != RHS0) 2456 return nullptr; 2457 2458 // ICMP_[US][GL]E X, C is folded to ICMP_[US][GL]T elsewhere. 2459 if (PredL == ICmpInst::ICMP_UGE || PredL == ICmpInst::ICMP_ULE || 2460 PredR == ICmpInst::ICMP_UGE || PredR == ICmpInst::ICMP_ULE || 2461 PredL == ICmpInst::ICMP_SGE || PredL == ICmpInst::ICMP_SLE || 2462 PredR == ICmpInst::ICMP_SGE || PredR == ICmpInst::ICMP_SLE) 2463 return nullptr; 2464 2465 // We can't fold (ugt x, C) | (sgt x, C2). 2466 if (!predicatesFoldable(PredL, PredR)) 2467 return nullptr; 2468 2469 // Ensure that the larger constant is on the RHS. 2470 bool ShouldSwap; 2471 if (CmpInst::isSigned(PredL) || 2472 (ICmpInst::isEquality(PredL) && CmpInst::isSigned(PredR))) 2473 ShouldSwap = LHSC->getValue().sgt(RHSC->getValue()); 2474 else 2475 ShouldSwap = LHSC->getValue().ugt(RHSC->getValue()); 2476 2477 if (ShouldSwap) { 2478 std::swap(LHS, RHS); 2479 std::swap(LHSC, RHSC); 2480 std::swap(PredL, PredR); 2481 } 2482 2483 // At this point, we know we have two icmp instructions 2484 // comparing a value against two constants and or'ing the result 2485 // together. Because of the above check, we know that we only have 2486 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the 2487 // icmp folding check above), that the two constants are not 2488 // equal. 2489 assert(LHSC != RHSC && "Compares not folded above?"); 2490 2491 switch (PredL) { 2492 default: 2493 llvm_unreachable("Unknown integer condition code!"); 2494 case ICmpInst::ICMP_EQ: 2495 switch (PredR) { 2496 default: 2497 llvm_unreachable("Unknown integer condition code!"); 2498 case ICmpInst::ICMP_EQ: 2499 // Potential folds for this case should already be handled. 2500 break; 2501 case ICmpInst::ICMP_UGT: 2502 // (X == 0 || X u> C) -> (X-1) u>= C 2503 if (LHSC->isMinValue(false)) 2504 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue() + 1, 2505 false, false); 2506 // (X == 13 | X u> 14) -> no change 2507 break; 2508 case ICmpInst::ICMP_SGT: 2509 // (X == INT_MIN || X s> C) -> (X-(INT_MIN+1)) u>= C-INT_MIN 2510 if (LHSC->isMinValue(true)) 2511 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue() + 1, 2512 true, false); 2513 // (X == 13 | X s> 14) -> no change 2514 break; 2515 } 2516 break; 2517 case ICmpInst::ICMP_ULT: 2518 switch (PredR) { 2519 default: 2520 llvm_unreachable("Unknown integer condition code!"); 2521 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change 2522 // (X u< C || X == UINT_MAX) => (X-C) u>= UINT_MAX-C 2523 if (RHSC->isMaxValue(false)) 2524 return insertRangeTest(LHS0, LHSC->getValue(), RHSC->getValue(), 2525 false, false); 2526 break; 2527 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2 2528 assert(!RHSC->isMaxValue(false) && "Missed icmp simplification"); 2529 return insertRangeTest(LHS0, LHSC->getValue(), RHSC->getValue() + 1, 2530 false, false); 2531 } 2532 break; 2533 case ICmpInst::ICMP_SLT: 2534 switch (PredR) { 2535 default: 2536 llvm_unreachable("Unknown integer condition code!"); 2537 case ICmpInst::ICMP_EQ: 2538 // (X s< C || X == INT_MAX) => (X-C) u>= INT_MAX-C 2539 if (RHSC->isMaxValue(true)) 2540 return insertRangeTest(LHS0, LHSC->getValue(), RHSC->getValue(), 2541 true, false); 2542 // (X s< 13 | X == 14) -> no change 2543 break; 2544 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) u> 2 2545 assert(!RHSC->isMaxValue(true) && "Missed icmp simplification"); 2546 return insertRangeTest(LHS0, LHSC->getValue(), RHSC->getValue() + 1, true, 2547 false); 2548 } 2549 break; 2550 } 2551 return nullptr; 2552 } 2553 2554 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches 2555 // here. We should standardize that construct where it is needed or choose some 2556 // other way to ensure that commutated variants of patterns are not missed. 2557 Instruction *InstCombinerImpl::visitOr(BinaryOperator &I) { 2558 if (Value *V = SimplifyOrInst(I.getOperand(0), I.getOperand(1), 2559 SQ.getWithInstruction(&I))) 2560 return replaceInstUsesWith(I, V); 2561 2562 if (SimplifyAssociativeOrCommutative(I)) 2563 return &I; 2564 2565 if (Instruction *X = foldVectorBinop(I)) 2566 return X; 2567 2568 // See if we can simplify any instructions used by the instruction whose sole 2569 // purpose is to compute bits we don't care about. 2570 if (SimplifyDemandedInstructionBits(I)) 2571 return &I; 2572 2573 // Do this before using distributive laws to catch simple and/or/not patterns. 2574 if (Instruction *Xor = foldOrToXor(I, Builder)) 2575 return Xor; 2576 2577 // (A&B)|(A&C) -> A&(B|C) etc 2578 if (Value *V = SimplifyUsingDistributiveLaws(I)) 2579 return replaceInstUsesWith(I, V); 2580 2581 if (Value *V = SimplifyBSwap(I, Builder)) 2582 return replaceInstUsesWith(I, V); 2583 2584 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I)) 2585 return FoldedLogic; 2586 2587 if (Instruction *BSwap = matchBSwapOrBitReverse(I, /*MatchBSwaps*/ true, 2588 /*MatchBitReversals*/ false)) 2589 return BSwap; 2590 2591 if (Instruction *Funnel = matchFunnelShift(I, *this)) 2592 return Funnel; 2593 2594 if (Instruction *Concat = matchOrConcat(I, Builder)) 2595 return replaceInstUsesWith(I, Concat); 2596 2597 Value *X, *Y; 2598 const APInt *CV; 2599 if (match(&I, m_c_Or(m_OneUse(m_Xor(m_Value(X), m_APInt(CV))), m_Value(Y))) && 2600 !CV->isAllOnesValue() && MaskedValueIsZero(Y, *CV, 0, &I)) { 2601 // (X ^ C) | Y -> (X | Y) ^ C iff Y & C == 0 2602 // The check for a 'not' op is for efficiency (if Y is known zero --> ~X). 2603 Value *Or = Builder.CreateOr(X, Y); 2604 return BinaryOperator::CreateXor(Or, ConstantInt::get(I.getType(), *CV)); 2605 } 2606 2607 // (A & C)|(B & D) 2608 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 2609 Value *A, *B, *C, *D; 2610 if (match(Op0, m_And(m_Value(A), m_Value(C))) && 2611 match(Op1, m_And(m_Value(B), m_Value(D)))) { 2612 // (A & C1)|(B & C2) 2613 ConstantInt *C1, *C2; 2614 if (match(C, m_ConstantInt(C1)) && match(D, m_ConstantInt(C2))) { 2615 Value *V1 = nullptr, *V2 = nullptr; 2616 if ((C1->getValue() & C2->getValue()).isNullValue()) { 2617 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2) 2618 // iff (C1&C2) == 0 and (N&~C1) == 0 2619 if (match(A, m_Or(m_Value(V1), m_Value(V2))) && 2620 ((V1 == B && 2621 MaskedValueIsZero(V2, ~C1->getValue(), 0, &I)) || // (V|N) 2622 (V2 == B && 2623 MaskedValueIsZero(V1, ~C1->getValue(), 0, &I)))) // (N|V) 2624 return BinaryOperator::CreateAnd(A, 2625 Builder.getInt(C1->getValue()|C2->getValue())); 2626 // Or commutes, try both ways. 2627 if (match(B, m_Or(m_Value(V1), m_Value(V2))) && 2628 ((V1 == A && 2629 MaskedValueIsZero(V2, ~C2->getValue(), 0, &I)) || // (V|N) 2630 (V2 == A && 2631 MaskedValueIsZero(V1, ~C2->getValue(), 0, &I)))) // (N|V) 2632 return BinaryOperator::CreateAnd(B, 2633 Builder.getInt(C1->getValue()|C2->getValue())); 2634 2635 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2) 2636 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0. 2637 ConstantInt *C3 = nullptr, *C4 = nullptr; 2638 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) && 2639 (C3->getValue() & ~C1->getValue()).isNullValue() && 2640 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) && 2641 (C4->getValue() & ~C2->getValue()).isNullValue()) { 2642 V2 = Builder.CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield"); 2643 return BinaryOperator::CreateAnd(V2, 2644 Builder.getInt(C1->getValue()|C2->getValue())); 2645 } 2646 } 2647 2648 if (C1->getValue() == ~C2->getValue()) { 2649 Value *X; 2650 2651 // ((X|B)&C1)|(B&C2) -> (X&C1) | B iff C1 == ~C2 2652 if (match(A, m_c_Or(m_Value(X), m_Specific(B)))) 2653 return BinaryOperator::CreateOr(Builder.CreateAnd(X, C1), B); 2654 // (A&C2)|((X|A)&C1) -> (X&C2) | A iff C1 == ~C2 2655 if (match(B, m_c_Or(m_Specific(A), m_Value(X)))) 2656 return BinaryOperator::CreateOr(Builder.CreateAnd(X, C2), A); 2657 2658 // ((X^B)&C1)|(B&C2) -> (X&C1) ^ B iff C1 == ~C2 2659 if (match(A, m_c_Xor(m_Value(X), m_Specific(B)))) 2660 return BinaryOperator::CreateXor(Builder.CreateAnd(X, C1), B); 2661 // (A&C2)|((X^A)&C1) -> (X&C2) ^ A iff C1 == ~C2 2662 if (match(B, m_c_Xor(m_Specific(A), m_Value(X)))) 2663 return BinaryOperator::CreateXor(Builder.CreateAnd(X, C2), A); 2664 } 2665 } 2666 2667 // Don't try to form a select if it's unlikely that we'll get rid of at 2668 // least one of the operands. A select is generally more expensive than the 2669 // 'or' that it is replacing. 2670 if (Op0->hasOneUse() || Op1->hasOneUse()) { 2671 // (Cond & C) | (~Cond & D) -> Cond ? C : D, and commuted variants. 2672 if (Value *V = matchSelectFromAndOr(A, C, B, D)) 2673 return replaceInstUsesWith(I, V); 2674 if (Value *V = matchSelectFromAndOr(A, C, D, B)) 2675 return replaceInstUsesWith(I, V); 2676 if (Value *V = matchSelectFromAndOr(C, A, B, D)) 2677 return replaceInstUsesWith(I, V); 2678 if (Value *V = matchSelectFromAndOr(C, A, D, B)) 2679 return replaceInstUsesWith(I, V); 2680 if (Value *V = matchSelectFromAndOr(B, D, A, C)) 2681 return replaceInstUsesWith(I, V); 2682 if (Value *V = matchSelectFromAndOr(B, D, C, A)) 2683 return replaceInstUsesWith(I, V); 2684 if (Value *V = matchSelectFromAndOr(D, B, A, C)) 2685 return replaceInstUsesWith(I, V); 2686 if (Value *V = matchSelectFromAndOr(D, B, C, A)) 2687 return replaceInstUsesWith(I, V); 2688 } 2689 } 2690 2691 // (A ^ B) | ((B ^ C) ^ A) -> (A ^ B) | C 2692 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) 2693 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A)))) 2694 return BinaryOperator::CreateOr(Op0, C); 2695 2696 // ((A ^ C) ^ B) | (B ^ A) -> (B ^ A) | C 2697 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B)))) 2698 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A)))) 2699 return BinaryOperator::CreateOr(Op1, C); 2700 2701 // ((B | C) & A) | B -> B | (A & C) 2702 if (match(Op0, m_And(m_Or(m_Specific(Op1), m_Value(C)), m_Value(A)))) 2703 return BinaryOperator::CreateOr(Op1, Builder.CreateAnd(A, C)); 2704 2705 if (Instruction *DeMorgan = matchDeMorgansLaws(I, Builder)) 2706 return DeMorgan; 2707 2708 // Canonicalize xor to the RHS. 2709 bool SwappedForXor = false; 2710 if (match(Op0, m_Xor(m_Value(), m_Value()))) { 2711 std::swap(Op0, Op1); 2712 SwappedForXor = true; 2713 } 2714 2715 // A | ( A ^ B) -> A | B 2716 // A | (~A ^ B) -> A | ~B 2717 // (A & B) | (A ^ B) 2718 if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) { 2719 if (Op0 == A || Op0 == B) 2720 return BinaryOperator::CreateOr(A, B); 2721 2722 if (match(Op0, m_And(m_Specific(A), m_Specific(B))) || 2723 match(Op0, m_And(m_Specific(B), m_Specific(A)))) 2724 return BinaryOperator::CreateOr(A, B); 2725 2726 if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) { 2727 Value *Not = Builder.CreateNot(B, B->getName() + ".not"); 2728 return BinaryOperator::CreateOr(Not, Op0); 2729 } 2730 if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) { 2731 Value *Not = Builder.CreateNot(A, A->getName() + ".not"); 2732 return BinaryOperator::CreateOr(Not, Op0); 2733 } 2734 } 2735 2736 // A | ~(A | B) -> A | ~B 2737 // A | ~(A ^ B) -> A | ~B 2738 if (match(Op1, m_Not(m_Value(A)))) 2739 if (BinaryOperator *B = dyn_cast<BinaryOperator>(A)) 2740 if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) && 2741 Op1->hasOneUse() && (B->getOpcode() == Instruction::Or || 2742 B->getOpcode() == Instruction::Xor)) { 2743 Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) : 2744 B->getOperand(0); 2745 Value *Not = Builder.CreateNot(NotOp, NotOp->getName() + ".not"); 2746 return BinaryOperator::CreateOr(Not, Op0); 2747 } 2748 2749 if (SwappedForXor) 2750 std::swap(Op0, Op1); 2751 2752 { 2753 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0); 2754 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1); 2755 if (LHS && RHS) 2756 if (Value *Res = foldOrOfICmps(LHS, RHS, I)) 2757 return replaceInstUsesWith(I, Res); 2758 2759 // TODO: Make this recursive; it's a little tricky because an arbitrary 2760 // number of 'or' instructions might have to be created. 2761 Value *X, *Y; 2762 if (LHS && match(Op1, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) { 2763 if (auto *Cmp = dyn_cast<ICmpInst>(X)) 2764 if (Value *Res = foldOrOfICmps(LHS, Cmp, I)) 2765 return replaceInstUsesWith(I, Builder.CreateOr(Res, Y)); 2766 if (auto *Cmp = dyn_cast<ICmpInst>(Y)) 2767 if (Value *Res = foldOrOfICmps(LHS, Cmp, I)) 2768 return replaceInstUsesWith(I, Builder.CreateOr(Res, X)); 2769 } 2770 if (RHS && match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) { 2771 if (auto *Cmp = dyn_cast<ICmpInst>(X)) 2772 if (Value *Res = foldOrOfICmps(Cmp, RHS, I)) 2773 return replaceInstUsesWith(I, Builder.CreateOr(Res, Y)); 2774 if (auto *Cmp = dyn_cast<ICmpInst>(Y)) 2775 if (Value *Res = foldOrOfICmps(Cmp, RHS, I)) 2776 return replaceInstUsesWith(I, Builder.CreateOr(Res, X)); 2777 } 2778 } 2779 2780 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) 2781 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) 2782 if (Value *Res = foldLogicOfFCmps(LHS, RHS, false)) 2783 return replaceInstUsesWith(I, Res); 2784 2785 if (Instruction *FoldedFCmps = reassociateFCmps(I, Builder)) 2786 return FoldedFCmps; 2787 2788 if (Instruction *CastedOr = foldCastedBitwiseLogic(I)) 2789 return CastedOr; 2790 2791 // or(sext(A), B) / or(B, sext(A)) --> A ? -1 : B, where A is i1 or <N x i1>. 2792 if (match(Op0, m_OneUse(m_SExt(m_Value(A)))) && 2793 A->getType()->isIntOrIntVectorTy(1)) 2794 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1); 2795 if (match(Op1, m_OneUse(m_SExt(m_Value(A)))) && 2796 A->getType()->isIntOrIntVectorTy(1)) 2797 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0); 2798 2799 // Note: If we've gotten to the point of visiting the outer OR, then the 2800 // inner one couldn't be simplified. If it was a constant, then it won't 2801 // be simplified by a later pass either, so we try swapping the inner/outer 2802 // ORs in the hopes that we'll be able to simplify it this way. 2803 // (X|C) | V --> (X|V) | C 2804 ConstantInt *CI; 2805 if (Op0->hasOneUse() && !match(Op1, m_ConstantInt()) && 2806 match(Op0, m_Or(m_Value(A), m_ConstantInt(CI)))) { 2807 Value *Inner = Builder.CreateOr(A, Op1); 2808 Inner->takeName(Op0); 2809 return BinaryOperator::CreateOr(Inner, CI); 2810 } 2811 2812 // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D)) 2813 // Since this OR statement hasn't been optimized further yet, we hope 2814 // that this transformation will allow the new ORs to be optimized. 2815 { 2816 Value *X = nullptr, *Y = nullptr; 2817 if (Op0->hasOneUse() && Op1->hasOneUse() && 2818 match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) && 2819 match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) { 2820 Value *orTrue = Builder.CreateOr(A, C); 2821 Value *orFalse = Builder.CreateOr(B, D); 2822 return SelectInst::Create(X, orTrue, orFalse); 2823 } 2824 } 2825 2826 // or(ashr(subNSW(Y, X), ScalarSizeInBits(Y) - 1), X) --> X s> Y ? -1 : X. 2827 { 2828 Value *X, *Y; 2829 Type *Ty = I.getType(); 2830 if (match(&I, m_c_Or(m_OneUse(m_AShr( 2831 m_NSWSub(m_Value(Y), m_Value(X)), 2832 m_SpecificInt(Ty->getScalarSizeInBits() - 1))), 2833 m_Deferred(X)))) { 2834 Value *NewICmpInst = Builder.CreateICmpSGT(X, Y); 2835 Value *AllOnes = ConstantInt::getAllOnesValue(Ty); 2836 return SelectInst::Create(NewICmpInst, AllOnes, X); 2837 } 2838 } 2839 2840 if (Instruction *V = 2841 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I)) 2842 return V; 2843 2844 CmpInst::Predicate Pred; 2845 Value *Mul, *Ov, *MulIsNotZero, *UMulWithOv; 2846 // Check if the OR weakens the overflow condition for umul.with.overflow by 2847 // treating any non-zero result as overflow. In that case, we overflow if both 2848 // umul.with.overflow operands are != 0, as in that case the result can only 2849 // be 0, iff the multiplication overflows. 2850 if (match(&I, 2851 m_c_Or(m_CombineAnd(m_ExtractValue<1>(m_Value(UMulWithOv)), 2852 m_Value(Ov)), 2853 m_CombineAnd(m_ICmp(Pred, 2854 m_CombineAnd(m_ExtractValue<0>( 2855 m_Deferred(UMulWithOv)), 2856 m_Value(Mul)), 2857 m_ZeroInt()), 2858 m_Value(MulIsNotZero)))) && 2859 (Ov->hasOneUse() || (MulIsNotZero->hasOneUse() && Mul->hasOneUse())) && 2860 Pred == CmpInst::ICMP_NE) { 2861 Value *A, *B; 2862 if (match(UMulWithOv, m_Intrinsic<Intrinsic::umul_with_overflow>( 2863 m_Value(A), m_Value(B)))) { 2864 Value *NotNullA = Builder.CreateIsNotNull(A); 2865 Value *NotNullB = Builder.CreateIsNotNull(B); 2866 return BinaryOperator::CreateAnd(NotNullA, NotNullB); 2867 } 2868 } 2869 2870 // (~x) | y --> ~(x & (~y)) iff that gets rid of inversions 2871 if (sinkNotIntoOtherHandOfAndOrOr(I)) 2872 return &I; 2873 2874 return nullptr; 2875 } 2876 2877 /// A ^ B can be specified using other logic ops in a variety of patterns. We 2878 /// can fold these early and efficiently by morphing an existing instruction. 2879 static Instruction *foldXorToXor(BinaryOperator &I, 2880 InstCombiner::BuilderTy &Builder) { 2881 assert(I.getOpcode() == Instruction::Xor); 2882 Value *Op0 = I.getOperand(0); 2883 Value *Op1 = I.getOperand(1); 2884 Value *A, *B; 2885 2886 // There are 4 commuted variants for each of the basic patterns. 2887 2888 // (A & B) ^ (A | B) -> A ^ B 2889 // (A & B) ^ (B | A) -> A ^ B 2890 // (A | B) ^ (A & B) -> A ^ B 2891 // (A | B) ^ (B & A) -> A ^ B 2892 if (match(&I, m_c_Xor(m_And(m_Value(A), m_Value(B)), 2893 m_c_Or(m_Deferred(A), m_Deferred(B))))) 2894 return BinaryOperator::CreateXor(A, B); 2895 2896 // (A | ~B) ^ (~A | B) -> A ^ B 2897 // (~B | A) ^ (~A | B) -> A ^ B 2898 // (~A | B) ^ (A | ~B) -> A ^ B 2899 // (B | ~A) ^ (A | ~B) -> A ^ B 2900 if (match(&I, m_Xor(m_c_Or(m_Value(A), m_Not(m_Value(B))), 2901 m_c_Or(m_Not(m_Deferred(A)), m_Deferred(B))))) 2902 return BinaryOperator::CreateXor(A, B); 2903 2904 // (A & ~B) ^ (~A & B) -> A ^ B 2905 // (~B & A) ^ (~A & B) -> A ^ B 2906 // (~A & B) ^ (A & ~B) -> A ^ B 2907 // (B & ~A) ^ (A & ~B) -> A ^ B 2908 if (match(&I, m_Xor(m_c_And(m_Value(A), m_Not(m_Value(B))), 2909 m_c_And(m_Not(m_Deferred(A)), m_Deferred(B))))) 2910 return BinaryOperator::CreateXor(A, B); 2911 2912 // For the remaining cases we need to get rid of one of the operands. 2913 if (!Op0->hasOneUse() && !Op1->hasOneUse()) 2914 return nullptr; 2915 2916 // (A | B) ^ ~(A & B) -> ~(A ^ B) 2917 // (A | B) ^ ~(B & A) -> ~(A ^ B) 2918 // (A & B) ^ ~(A | B) -> ~(A ^ B) 2919 // (A & B) ^ ~(B | A) -> ~(A ^ B) 2920 // Complexity sorting ensures the not will be on the right side. 2921 if ((match(Op0, m_Or(m_Value(A), m_Value(B))) && 2922 match(Op1, m_Not(m_c_And(m_Specific(A), m_Specific(B))))) || 2923 (match(Op0, m_And(m_Value(A), m_Value(B))) && 2924 match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B)))))) 2925 return BinaryOperator::CreateNot(Builder.CreateXor(A, B)); 2926 2927 return nullptr; 2928 } 2929 2930 Value *InstCombinerImpl::foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS, 2931 BinaryOperator &I) { 2932 assert(I.getOpcode() == Instruction::Xor && I.getOperand(0) == LHS && 2933 I.getOperand(1) == RHS && "Should be 'xor' with these operands"); 2934 2935 if (predicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) { 2936 if (LHS->getOperand(0) == RHS->getOperand(1) && 2937 LHS->getOperand(1) == RHS->getOperand(0)) 2938 LHS->swapOperands(); 2939 if (LHS->getOperand(0) == RHS->getOperand(0) && 2940 LHS->getOperand(1) == RHS->getOperand(1)) { 2941 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B) 2942 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1); 2943 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS); 2944 bool IsSigned = LHS->isSigned() || RHS->isSigned(); 2945 return getNewICmpValue(Code, IsSigned, Op0, Op1, Builder); 2946 } 2947 } 2948 2949 // TODO: This can be generalized to compares of non-signbits using 2950 // decomposeBitTestICmp(). It could be enhanced more by using (something like) 2951 // foldLogOpOfMaskedICmps(). 2952 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate(); 2953 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1); 2954 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1); 2955 if ((LHS->hasOneUse() || RHS->hasOneUse()) && 2956 LHS0->getType() == RHS0->getType() && 2957 LHS0->getType()->isIntOrIntVectorTy()) { 2958 // (X > -1) ^ (Y > -1) --> (X ^ Y) < 0 2959 // (X < 0) ^ (Y < 0) --> (X ^ Y) < 0 2960 if ((PredL == CmpInst::ICMP_SGT && match(LHS1, m_AllOnes()) && 2961 PredR == CmpInst::ICMP_SGT && match(RHS1, m_AllOnes())) || 2962 (PredL == CmpInst::ICMP_SLT && match(LHS1, m_Zero()) && 2963 PredR == CmpInst::ICMP_SLT && match(RHS1, m_Zero()))) { 2964 Value *Zero = ConstantInt::getNullValue(LHS0->getType()); 2965 return Builder.CreateICmpSLT(Builder.CreateXor(LHS0, RHS0), Zero); 2966 } 2967 // (X > -1) ^ (Y < 0) --> (X ^ Y) > -1 2968 // (X < 0) ^ (Y > -1) --> (X ^ Y) > -1 2969 if ((PredL == CmpInst::ICMP_SGT && match(LHS1, m_AllOnes()) && 2970 PredR == CmpInst::ICMP_SLT && match(RHS1, m_Zero())) || 2971 (PredL == CmpInst::ICMP_SLT && match(LHS1, m_Zero()) && 2972 PredR == CmpInst::ICMP_SGT && match(RHS1, m_AllOnes()))) { 2973 Value *MinusOne = ConstantInt::getAllOnesValue(LHS0->getType()); 2974 return Builder.CreateICmpSGT(Builder.CreateXor(LHS0, RHS0), MinusOne); 2975 } 2976 } 2977 2978 // Instead of trying to imitate the folds for and/or, decompose this 'xor' 2979 // into those logic ops. That is, try to turn this into an and-of-icmps 2980 // because we have many folds for that pattern. 2981 // 2982 // This is based on a truth table definition of xor: 2983 // X ^ Y --> (X | Y) & !(X & Y) 2984 if (Value *OrICmp = SimplifyBinOp(Instruction::Or, LHS, RHS, SQ)) { 2985 // TODO: If OrICmp is true, then the definition of xor simplifies to !(X&Y). 2986 // TODO: If OrICmp is false, the whole thing is false (InstSimplify?). 2987 if (Value *AndICmp = SimplifyBinOp(Instruction::And, LHS, RHS, SQ)) { 2988 // TODO: Independently handle cases where the 'and' side is a constant. 2989 ICmpInst *X = nullptr, *Y = nullptr; 2990 if (OrICmp == LHS && AndICmp == RHS) { 2991 // (LHS | RHS) & !(LHS & RHS) --> LHS & !RHS --> X & !Y 2992 X = LHS; 2993 Y = RHS; 2994 } 2995 if (OrICmp == RHS && AndICmp == LHS) { 2996 // !(LHS & RHS) & (LHS | RHS) --> !LHS & RHS --> !Y & X 2997 X = RHS; 2998 Y = LHS; 2999 } 3000 if (X && Y && (Y->hasOneUse() || canFreelyInvertAllUsersOf(Y, &I))) { 3001 // Invert the predicate of 'Y', thus inverting its output. 3002 Y->setPredicate(Y->getInversePredicate()); 3003 // So, are there other uses of Y? 3004 if (!Y->hasOneUse()) { 3005 // We need to adapt other uses of Y though. Get a value that matches 3006 // the original value of Y before inversion. While this increases 3007 // immediate instruction count, we have just ensured that all the 3008 // users are freely-invertible, so that 'not' *will* get folded away. 3009 BuilderTy::InsertPointGuard Guard(Builder); 3010 // Set insertion point to right after the Y. 3011 Builder.SetInsertPoint(Y->getParent(), ++(Y->getIterator())); 3012 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not"); 3013 // Replace all uses of Y (excluding the one in NotY!) with NotY. 3014 Worklist.pushUsersToWorkList(*Y); 3015 Y->replaceUsesWithIf(NotY, 3016 [NotY](Use &U) { return U.getUser() != NotY; }); 3017 } 3018 // All done. 3019 return Builder.CreateAnd(LHS, RHS); 3020 } 3021 } 3022 } 3023 3024 return nullptr; 3025 } 3026 3027 /// If we have a masked merge, in the canonical form of: 3028 /// (assuming that A only has one use.) 3029 /// | A | |B| 3030 /// ((x ^ y) & M) ^ y 3031 /// | D | 3032 /// * If M is inverted: 3033 /// | D | 3034 /// ((x ^ y) & ~M) ^ y 3035 /// We can canonicalize by swapping the final xor operand 3036 /// to eliminate the 'not' of the mask. 3037 /// ((x ^ y) & M) ^ x 3038 /// * If M is a constant, and D has one use, we transform to 'and' / 'or' ops 3039 /// because that shortens the dependency chain and improves analysis: 3040 /// (x & M) | (y & ~M) 3041 static Instruction *visitMaskedMerge(BinaryOperator &I, 3042 InstCombiner::BuilderTy &Builder) { 3043 Value *B, *X, *D; 3044 Value *M; 3045 if (!match(&I, m_c_Xor(m_Value(B), 3046 m_OneUse(m_c_And( 3047 m_CombineAnd(m_c_Xor(m_Deferred(B), m_Value(X)), 3048 m_Value(D)), 3049 m_Value(M)))))) 3050 return nullptr; 3051 3052 Value *NotM; 3053 if (match(M, m_Not(m_Value(NotM)))) { 3054 // De-invert the mask and swap the value in B part. 3055 Value *NewA = Builder.CreateAnd(D, NotM); 3056 return BinaryOperator::CreateXor(NewA, X); 3057 } 3058 3059 Constant *C; 3060 if (D->hasOneUse() && match(M, m_Constant(C))) { 3061 // Propagating undef is unsafe. Clamp undef elements to -1. 3062 Type *EltTy = C->getType()->getScalarType(); 3063 C = Constant::replaceUndefsWith(C, ConstantInt::getAllOnesValue(EltTy)); 3064 // Unfold. 3065 Value *LHS = Builder.CreateAnd(X, C); 3066 Value *NotC = Builder.CreateNot(C); 3067 Value *RHS = Builder.CreateAnd(B, NotC); 3068 return BinaryOperator::CreateOr(LHS, RHS); 3069 } 3070 3071 return nullptr; 3072 } 3073 3074 // Transform 3075 // ~(x ^ y) 3076 // into: 3077 // (~x) ^ y 3078 // or into 3079 // x ^ (~y) 3080 static Instruction *sinkNotIntoXor(BinaryOperator &I, 3081 InstCombiner::BuilderTy &Builder) { 3082 Value *X, *Y; 3083 // FIXME: one-use check is not needed in general, but currently we are unable 3084 // to fold 'not' into 'icmp', if that 'icmp' has multiple uses. (D35182) 3085 if (!match(&I, m_Not(m_OneUse(m_Xor(m_Value(X), m_Value(Y)))))) 3086 return nullptr; 3087 3088 // We only want to do the transform if it is free to do. 3089 if (InstCombiner::isFreeToInvert(X, X->hasOneUse())) { 3090 // Ok, good. 3091 } else if (InstCombiner::isFreeToInvert(Y, Y->hasOneUse())) { 3092 std::swap(X, Y); 3093 } else 3094 return nullptr; 3095 3096 Value *NotX = Builder.CreateNot(X, X->getName() + ".not"); 3097 return BinaryOperator::CreateXor(NotX, Y, I.getName() + ".demorgan"); 3098 } 3099 3100 // Transform 3101 // z = (~x) &/| y 3102 // into: 3103 // z = ~(x |/& (~y)) 3104 // iff y is free to invert and all uses of z can be freely updated. 3105 bool InstCombinerImpl::sinkNotIntoOtherHandOfAndOrOr(BinaryOperator &I) { 3106 Instruction::BinaryOps NewOpc; 3107 switch (I.getOpcode()) { 3108 case Instruction::And: 3109 NewOpc = Instruction::Or; 3110 break; 3111 case Instruction::Or: 3112 NewOpc = Instruction::And; 3113 break; 3114 default: 3115 return false; 3116 }; 3117 3118 Value *X, *Y; 3119 if (!match(&I, m_c_BinOp(m_Not(m_Value(X)), m_Value(Y)))) 3120 return false; 3121 3122 // Will we be able to fold the `not` into Y eventually? 3123 if (!InstCombiner::isFreeToInvert(Y, Y->hasOneUse())) 3124 return false; 3125 3126 // And can our users be adapted? 3127 if (!InstCombiner::canFreelyInvertAllUsersOf(&I, /*IgnoredUser=*/nullptr)) 3128 return false; 3129 3130 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not"); 3131 Value *NewBinOp = 3132 BinaryOperator::Create(NewOpc, X, NotY, I.getName() + ".not"); 3133 Builder.Insert(NewBinOp); 3134 replaceInstUsesWith(I, NewBinOp); 3135 // We can not just create an outer `not`, it will most likely be immediately 3136 // folded back, reconstructing our initial pattern, and causing an 3137 // infinite combine loop, so immediately manually fold it away. 3138 freelyInvertAllUsersOf(NewBinOp); 3139 return true; 3140 } 3141 3142 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches 3143 // here. We should standardize that construct where it is needed or choose some 3144 // other way to ensure that commutated variants of patterns are not missed. 3145 Instruction *InstCombinerImpl::visitXor(BinaryOperator &I) { 3146 if (Value *V = SimplifyXorInst(I.getOperand(0), I.getOperand(1), 3147 SQ.getWithInstruction(&I))) 3148 return replaceInstUsesWith(I, V); 3149 3150 if (SimplifyAssociativeOrCommutative(I)) 3151 return &I; 3152 3153 if (Instruction *X = foldVectorBinop(I)) 3154 return X; 3155 3156 if (Instruction *NewXor = foldXorToXor(I, Builder)) 3157 return NewXor; 3158 3159 // (A&B)^(A&C) -> A&(B^C) etc 3160 if (Value *V = SimplifyUsingDistributiveLaws(I)) 3161 return replaceInstUsesWith(I, V); 3162 3163 // See if we can simplify any instructions used by the instruction whose sole 3164 // purpose is to compute bits we don't care about. 3165 if (SimplifyDemandedInstructionBits(I)) 3166 return &I; 3167 3168 if (Value *V = SimplifyBSwap(I, Builder)) 3169 return replaceInstUsesWith(I, V); 3170 3171 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 3172 Type *Ty = I.getType(); 3173 3174 // Fold (X & M) ^ (Y & ~M) -> (X & M) | (Y & ~M) 3175 // This it a special case in haveNoCommonBitsSet, but the computeKnownBits 3176 // calls in there are unnecessary as SimplifyDemandedInstructionBits should 3177 // have already taken care of those cases. 3178 Value *M; 3179 if (match(&I, m_c_Xor(m_c_And(m_Not(m_Value(M)), m_Value()), 3180 m_c_And(m_Deferred(M), m_Value())))) 3181 return BinaryOperator::CreateOr(Op0, Op1); 3182 3183 // Apply DeMorgan's Law for 'nand' / 'nor' logic with an inverted operand. 3184 Value *X, *Y; 3185 3186 // We must eliminate the and/or (one-use) for these transforms to not increase 3187 // the instruction count. 3188 // ~(~X & Y) --> (X | ~Y) 3189 // ~(Y & ~X) --> (X | ~Y) 3190 if (match(&I, m_Not(m_OneUse(m_c_And(m_Not(m_Value(X)), m_Value(Y)))))) { 3191 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not"); 3192 return BinaryOperator::CreateOr(X, NotY); 3193 } 3194 // ~(~X | Y) --> (X & ~Y) 3195 // ~(Y | ~X) --> (X & ~Y) 3196 if (match(&I, m_Not(m_OneUse(m_c_Or(m_Not(m_Value(X)), m_Value(Y)))))) { 3197 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not"); 3198 return BinaryOperator::CreateAnd(X, NotY); 3199 } 3200 3201 if (Instruction *Xor = visitMaskedMerge(I, Builder)) 3202 return Xor; 3203 3204 // Is this a 'not' (~) fed by a binary operator? 3205 BinaryOperator *NotVal; 3206 if (match(&I, m_Not(m_BinOp(NotVal)))) { 3207 if (NotVal->getOpcode() == Instruction::And || 3208 NotVal->getOpcode() == Instruction::Or) { 3209 // Apply DeMorgan's Law when inverts are free: 3210 // ~(X & Y) --> (~X | ~Y) 3211 // ~(X | Y) --> (~X & ~Y) 3212 if (isFreeToInvert(NotVal->getOperand(0), 3213 NotVal->getOperand(0)->hasOneUse()) && 3214 isFreeToInvert(NotVal->getOperand(1), 3215 NotVal->getOperand(1)->hasOneUse())) { 3216 Value *NotX = Builder.CreateNot(NotVal->getOperand(0), "notlhs"); 3217 Value *NotY = Builder.CreateNot(NotVal->getOperand(1), "notrhs"); 3218 if (NotVal->getOpcode() == Instruction::And) 3219 return BinaryOperator::CreateOr(NotX, NotY); 3220 return BinaryOperator::CreateAnd(NotX, NotY); 3221 } 3222 } 3223 3224 // ~(~X >>s Y) --> (X >>s Y) 3225 if (match(NotVal, m_AShr(m_Not(m_Value(X)), m_Value(Y)))) 3226 return BinaryOperator::CreateAShr(X, Y); 3227 3228 // If we are inverting a right-shifted constant, we may be able to eliminate 3229 // the 'not' by inverting the constant and using the opposite shift type. 3230 // Canonicalization rules ensure that only a negative constant uses 'ashr', 3231 // but we must check that in case that transform has not fired yet. 3232 3233 // ~(C >>s Y) --> ~C >>u Y (when inverting the replicated sign bits) 3234 Constant *C; 3235 if (match(NotVal, m_AShr(m_Constant(C), m_Value(Y))) && 3236 match(C, m_Negative())) { 3237 // We matched a negative constant, so propagating undef is unsafe. 3238 // Clamp undef elements to -1. 3239 Type *EltTy = Ty->getScalarType(); 3240 C = Constant::replaceUndefsWith(C, ConstantInt::getAllOnesValue(EltTy)); 3241 return BinaryOperator::CreateLShr(ConstantExpr::getNot(C), Y); 3242 } 3243 3244 // ~(C >>u Y) --> ~C >>s Y (when inverting the replicated sign bits) 3245 if (match(NotVal, m_LShr(m_Constant(C), m_Value(Y))) && 3246 match(C, m_NonNegative())) { 3247 // We matched a non-negative constant, so propagating undef is unsafe. 3248 // Clamp undef elements to 0. 3249 Type *EltTy = Ty->getScalarType(); 3250 C = Constant::replaceUndefsWith(C, ConstantInt::getNullValue(EltTy)); 3251 return BinaryOperator::CreateAShr(ConstantExpr::getNot(C), Y); 3252 } 3253 3254 // ~(X + C) --> ~C - X 3255 if (match(NotVal, m_c_Add(m_Value(X), m_ImmConstant(C)))) 3256 return BinaryOperator::CreateSub(ConstantExpr::getNot(C), X); 3257 3258 // ~(X - Y) --> ~X + Y 3259 // FIXME: is it really beneficial to sink the `not` here? 3260 if (match(NotVal, m_Sub(m_Value(X), m_Value(Y)))) 3261 if (isa<Constant>(X) || NotVal->hasOneUse()) 3262 return BinaryOperator::CreateAdd(Builder.CreateNot(X), Y); 3263 3264 // ~(~X + Y) --> X - Y 3265 if (match(NotVal, m_c_Add(m_Not(m_Value(X)), m_Value(Y)))) 3266 return BinaryOperator::CreateWithCopiedFlags(Instruction::Sub, X, Y, 3267 NotVal); 3268 } 3269 3270 // Use DeMorgan and reassociation to eliminate a 'not' op. 3271 Constant *C1; 3272 if (match(Op1, m_Constant(C1))) { 3273 Constant *C2; 3274 if (match(Op0, m_OneUse(m_Or(m_Not(m_Value(X)), m_Constant(C2))))) { 3275 // (~X | C2) ^ C1 --> ((X & ~C2) ^ -1) ^ C1 --> (X & ~C2) ^ ~C1 3276 Value *And = Builder.CreateAnd(X, ConstantExpr::getNot(C2)); 3277 return BinaryOperator::CreateXor(And, ConstantExpr::getNot(C1)); 3278 } 3279 if (match(Op0, m_OneUse(m_And(m_Not(m_Value(X)), m_Constant(C2))))) { 3280 // (~X & C2) ^ C1 --> ((X | ~C2) ^ -1) ^ C1 --> (X | ~C2) ^ ~C1 3281 Value *Or = Builder.CreateOr(X, ConstantExpr::getNot(C2)); 3282 return BinaryOperator::CreateXor(Or, ConstantExpr::getNot(C1)); 3283 } 3284 } 3285 3286 // not (cmp A, B) = !cmp A, B 3287 CmpInst::Predicate Pred; 3288 if (match(&I, m_Not(m_OneUse(m_Cmp(Pred, m_Value(), m_Value()))))) { 3289 cast<CmpInst>(Op0)->setPredicate(CmpInst::getInversePredicate(Pred)); 3290 return replaceInstUsesWith(I, Op0); 3291 } 3292 3293 { 3294 const APInt *RHSC; 3295 if (match(Op1, m_APInt(RHSC))) { 3296 Value *X; 3297 const APInt *C; 3298 // (C - X) ^ signmaskC --> (C + signmaskC) - X 3299 if (RHSC->isSignMask() && match(Op0, m_Sub(m_APInt(C), m_Value(X)))) 3300 return BinaryOperator::CreateSub(ConstantInt::get(Ty, *C + *RHSC), X); 3301 3302 // (X + C) ^ signmaskC --> X + (C + signmaskC) 3303 if (RHSC->isSignMask() && match(Op0, m_Add(m_Value(X), m_APInt(C)))) 3304 return BinaryOperator::CreateAdd(X, ConstantInt::get(Ty, *C + *RHSC)); 3305 3306 // (X | C) ^ RHSC --> X ^ (C ^ RHSC) iff X & C == 0 3307 if (match(Op0, m_Or(m_Value(X), m_APInt(C))) && 3308 MaskedValueIsZero(X, *C, 0, &I)) 3309 return BinaryOperator::CreateXor(X, ConstantInt::get(Ty, *C ^ *RHSC)); 3310 3311 // If RHSC is inverting the remaining bits of shifted X, 3312 // canonicalize to a 'not' before the shift to help SCEV and codegen: 3313 // (X << C) ^ RHSC --> ~X << C 3314 if (match(Op0, m_OneUse(m_Shl(m_Value(X), m_APInt(C)))) && 3315 *RHSC == APInt::getAllOnesValue(Ty->getScalarSizeInBits()).shl(*C)) { 3316 Value *NotX = Builder.CreateNot(X); 3317 return BinaryOperator::CreateShl(NotX, ConstantInt::get(Ty, *C)); 3318 } 3319 // (X >>u C) ^ RHSC --> ~X >>u C 3320 if (match(Op0, m_OneUse(m_LShr(m_Value(X), m_APInt(C)))) && 3321 *RHSC == APInt::getAllOnesValue(Ty->getScalarSizeInBits()).lshr(*C)) { 3322 Value *NotX = Builder.CreateNot(X); 3323 return BinaryOperator::CreateLShr(NotX, ConstantInt::get(Ty, *C)); 3324 } 3325 // TODO: We could handle 'ashr' here as well. That would be matching 3326 // a 'not' op and moving it before the shift. Doing that requires 3327 // preventing the inverse fold in canShiftBinOpWithConstantRHS(). 3328 } 3329 } 3330 3331 // FIXME: This should not be limited to scalar (pull into APInt match above). 3332 { 3333 Value *X; 3334 ConstantInt *C1, *C2, *C3; 3335 // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3) 3336 if (match(Op1, m_ConstantInt(C3)) && 3337 match(Op0, m_LShr(m_Xor(m_Value(X), m_ConstantInt(C1)), 3338 m_ConstantInt(C2))) && 3339 Op0->hasOneUse()) { 3340 // fold (C1 >> C2) ^ C3 3341 APInt FoldConst = C1->getValue().lshr(C2->getValue()); 3342 FoldConst ^= C3->getValue(); 3343 // Prepare the two operands. 3344 auto *Opnd0 = cast<Instruction>(Builder.CreateLShr(X, C2)); 3345 Opnd0->takeName(cast<Instruction>(Op0)); 3346 Opnd0->setDebugLoc(I.getDebugLoc()); 3347 return BinaryOperator::CreateXor(Opnd0, ConstantInt::get(Ty, FoldConst)); 3348 } 3349 } 3350 3351 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I)) 3352 return FoldedLogic; 3353 3354 // Y ^ (X | Y) --> X & ~Y 3355 // Y ^ (Y | X) --> X & ~Y 3356 if (match(Op1, m_OneUse(m_c_Or(m_Value(X), m_Specific(Op0))))) 3357 return BinaryOperator::CreateAnd(X, Builder.CreateNot(Op0)); 3358 // (X | Y) ^ Y --> X & ~Y 3359 // (Y | X) ^ Y --> X & ~Y 3360 if (match(Op0, m_OneUse(m_c_Or(m_Value(X), m_Specific(Op1))))) 3361 return BinaryOperator::CreateAnd(X, Builder.CreateNot(Op1)); 3362 3363 // Y ^ (X & Y) --> ~X & Y 3364 // Y ^ (Y & X) --> ~X & Y 3365 if (match(Op1, m_OneUse(m_c_And(m_Value(X), m_Specific(Op0))))) 3366 return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(X)); 3367 // (X & Y) ^ Y --> ~X & Y 3368 // (Y & X) ^ Y --> ~X & Y 3369 // Canonical form is (X & C) ^ C; don't touch that. 3370 // TODO: A 'not' op is better for analysis and codegen, but demanded bits must 3371 // be fixed to prefer that (otherwise we get infinite looping). 3372 if (!match(Op1, m_Constant()) && 3373 match(Op0, m_OneUse(m_c_And(m_Value(X), m_Specific(Op1))))) 3374 return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(X)); 3375 3376 Value *A, *B, *C; 3377 // (A ^ B) ^ (A | C) --> (~A & C) ^ B -- There are 4 commuted variants. 3378 if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_Value(A), m_Value(B))), 3379 m_OneUse(m_c_Or(m_Deferred(A), m_Value(C)))))) 3380 return BinaryOperator::CreateXor( 3381 Builder.CreateAnd(Builder.CreateNot(A), C), B); 3382 3383 // (A ^ B) ^ (B | C) --> (~B & C) ^ A -- There are 4 commuted variants. 3384 if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_Value(A), m_Value(B))), 3385 m_OneUse(m_c_Or(m_Deferred(B), m_Value(C)))))) 3386 return BinaryOperator::CreateXor( 3387 Builder.CreateAnd(Builder.CreateNot(B), C), A); 3388 3389 // (A & B) ^ (A ^ B) -> (A | B) 3390 if (match(Op0, m_And(m_Value(A), m_Value(B))) && 3391 match(Op1, m_c_Xor(m_Specific(A), m_Specific(B)))) 3392 return BinaryOperator::CreateOr(A, B); 3393 // (A ^ B) ^ (A & B) -> (A | B) 3394 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) && 3395 match(Op1, m_c_And(m_Specific(A), m_Specific(B)))) 3396 return BinaryOperator::CreateOr(A, B); 3397 3398 // (A & ~B) ^ ~A -> ~(A & B) 3399 // (~B & A) ^ ~A -> ~(A & B) 3400 if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) && 3401 match(Op1, m_Not(m_Specific(A)))) 3402 return BinaryOperator::CreateNot(Builder.CreateAnd(A, B)); 3403 3404 // (~A & B) ^ A --> A | B -- There are 4 commuted variants. 3405 if (match(&I, m_c_Xor(m_c_And(m_Not(m_Value(A)), m_Value(B)), m_Deferred(A)))) 3406 return BinaryOperator::CreateOr(A, B); 3407 3408 // (A | B) ^ (A | C) --> (B ^ C) & ~A -- There are 4 commuted variants. 3409 // TODO: Loosen one-use restriction if common operand is a constant. 3410 Value *D; 3411 if (match(Op0, m_OneUse(m_Or(m_Value(A), m_Value(B)))) && 3412 match(Op1, m_OneUse(m_Or(m_Value(C), m_Value(D))))) { 3413 if (B == C || B == D) 3414 std::swap(A, B); 3415 if (A == C) 3416 std::swap(C, D); 3417 if (A == D) { 3418 Value *NotA = Builder.CreateNot(A); 3419 return BinaryOperator::CreateAnd(Builder.CreateXor(B, C), NotA); 3420 } 3421 } 3422 3423 if (auto *LHS = dyn_cast<ICmpInst>(I.getOperand(0))) 3424 if (auto *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) 3425 if (Value *V = foldXorOfICmps(LHS, RHS, I)) 3426 return replaceInstUsesWith(I, V); 3427 3428 if (Instruction *CastedXor = foldCastedBitwiseLogic(I)) 3429 return CastedXor; 3430 3431 // Canonicalize a shifty way to code absolute value to the common pattern. 3432 // There are 4 potential commuted variants. Move the 'ashr' candidate to Op1. 3433 // We're relying on the fact that we only do this transform when the shift has 3434 // exactly 2 uses and the add has exactly 1 use (otherwise, we might increase 3435 // instructions). 3436 if (Op0->hasNUses(2)) 3437 std::swap(Op0, Op1); 3438 3439 const APInt *ShAmt; 3440 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) && 3441 Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 && 3442 match(Op0, m_OneUse(m_c_Add(m_Specific(A), m_Specific(Op1))))) { 3443 // B = ashr i32 A, 31 ; smear the sign bit 3444 // xor (add A, B), B ; add -1 and flip bits if negative 3445 // --> (A < 0) ? -A : A 3446 Value *Cmp = Builder.CreateICmpSLT(A, ConstantInt::getNullValue(Ty)); 3447 // Copy the nuw/nsw flags from the add to the negate. 3448 auto *Add = cast<BinaryOperator>(Op0); 3449 Value *Neg = Builder.CreateNeg(A, "", Add->hasNoUnsignedWrap(), 3450 Add->hasNoSignedWrap()); 3451 return SelectInst::Create(Cmp, Neg, A); 3452 } 3453 3454 // Eliminate a bitwise 'not' op of 'not' min/max by inverting the min/max: 3455 // 3456 // %notx = xor i32 %x, -1 3457 // %cmp1 = icmp sgt i32 %notx, %y 3458 // %smax = select i1 %cmp1, i32 %notx, i32 %y 3459 // %res = xor i32 %smax, -1 3460 // => 3461 // %noty = xor i32 %y, -1 3462 // %cmp2 = icmp slt %x, %noty 3463 // %res = select i1 %cmp2, i32 %x, i32 %noty 3464 // 3465 // Same is applicable for smin/umax/umin. 3466 if (match(Op1, m_AllOnes()) && Op0->hasOneUse()) { 3467 Value *LHS, *RHS; 3468 SelectPatternFlavor SPF = matchSelectPattern(Op0, LHS, RHS).Flavor; 3469 if (SelectPatternResult::isMinOrMax(SPF)) { 3470 // It's possible we get here before the not has been simplified, so make 3471 // sure the input to the not isn't freely invertible. 3472 if (match(LHS, m_Not(m_Value(X))) && !isFreeToInvert(X, X->hasOneUse())) { 3473 Value *NotY = Builder.CreateNot(RHS); 3474 return SelectInst::Create( 3475 Builder.CreateICmp(getInverseMinMaxPred(SPF), X, NotY), X, NotY); 3476 } 3477 3478 // It's possible we get here before the not has been simplified, so make 3479 // sure the input to the not isn't freely invertible. 3480 if (match(RHS, m_Not(m_Value(Y))) && !isFreeToInvert(Y, Y->hasOneUse())) { 3481 Value *NotX = Builder.CreateNot(LHS); 3482 return SelectInst::Create( 3483 Builder.CreateICmp(getInverseMinMaxPred(SPF), NotX, Y), NotX, Y); 3484 } 3485 3486 // If both sides are freely invertible, then we can get rid of the xor 3487 // completely. 3488 if (isFreeToInvert(LHS, !LHS->hasNUsesOrMore(3)) && 3489 isFreeToInvert(RHS, !RHS->hasNUsesOrMore(3))) { 3490 Value *NotLHS = Builder.CreateNot(LHS); 3491 Value *NotRHS = Builder.CreateNot(RHS); 3492 return SelectInst::Create( 3493 Builder.CreateICmp(getInverseMinMaxPred(SPF), NotLHS, NotRHS), 3494 NotLHS, NotRHS); 3495 } 3496 } 3497 3498 // Pull 'not' into operands of select if both operands are one-use compares 3499 // or one is one-use compare and the other one is a constant. 3500 // Inverting the predicates eliminates the 'not' operation. 3501 // Example: 3502 // not (select ?, (cmp TPred, ?, ?), (cmp FPred, ?, ?) --> 3503 // select ?, (cmp InvTPred, ?, ?), (cmp InvFPred, ?, ?) 3504 // not (select ?, (cmp TPred, ?, ?), true --> 3505 // select ?, (cmp InvTPred, ?, ?), false 3506 if (auto *Sel = dyn_cast<SelectInst>(Op0)) { 3507 Value *TV = Sel->getTrueValue(); 3508 Value *FV = Sel->getFalseValue(); 3509 auto *CmpT = dyn_cast<CmpInst>(TV); 3510 auto *CmpF = dyn_cast<CmpInst>(FV); 3511 bool InvertibleT = (CmpT && CmpT->hasOneUse()) || isa<Constant>(TV); 3512 bool InvertibleF = (CmpF && CmpF->hasOneUse()) || isa<Constant>(FV); 3513 if (InvertibleT && InvertibleF) { 3514 if (CmpT) 3515 CmpT->setPredicate(CmpT->getInversePredicate()); 3516 else 3517 Sel->setTrueValue(ConstantExpr::getNot(cast<Constant>(TV))); 3518 if (CmpF) 3519 CmpF->setPredicate(CmpF->getInversePredicate()); 3520 else 3521 Sel->setFalseValue(ConstantExpr::getNot(cast<Constant>(FV))); 3522 return replaceInstUsesWith(I, Sel); 3523 } 3524 } 3525 } 3526 3527 if (Instruction *NewXor = sinkNotIntoXor(I, Builder)) 3528 return NewXor; 3529 3530 // Otherwise, if all else failed, try to hoist the xor-by-constant: 3531 // (X ^ C) ^ Y --> (X ^ Y) ^ C 3532 // Just like we do in other places, we completely avoid the fold 3533 // for constantexprs, at least to avoid endless combine loop. 3534 if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_CombineAnd(m_Value(X), 3535 m_Unless(m_ConstantExpr())), 3536 m_ImmConstant(C1))), 3537 m_Value(Y)))) 3538 return BinaryOperator::CreateXor(Builder.CreateXor(X, Y), C1); 3539 3540 return nullptr; 3541 } 3542