1 //===- InstCombineAndOrXor.cpp --------------------------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements the visitAnd, visitOr, and visitXor functions. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "InstCombine.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/Utils/CmpInstAnalysis.h" 20 using namespace llvm; 21 using namespace PatternMatch; 22 23 #define DEBUG_TYPE "instcombine" 24 25 /// isFreeToInvert - Return true if the specified value is free to invert (apply 26 /// ~ to). This happens in cases where the ~ can be eliminated. 27 static inline bool isFreeToInvert(Value *V) { 28 // ~(~(X)) -> X. 29 if (BinaryOperator::isNot(V)) 30 return true; 31 32 // Constants can be considered to be not'ed values. 33 if (isa<ConstantInt>(V)) 34 return true; 35 36 // Compares can be inverted if they have a single use. 37 if (CmpInst *CI = dyn_cast<CmpInst>(V)) 38 return CI->hasOneUse(); 39 40 return false; 41 } 42 43 static inline Value *dyn_castNotVal(Value *V) { 44 // If this is not(not(x)) don't return that this is a not: we want the two 45 // not's to be folded first. 46 if (BinaryOperator::isNot(V)) { 47 Value *Operand = BinaryOperator::getNotArgument(V); 48 if (!isFreeToInvert(Operand)) 49 return Operand; 50 } 51 52 // Constants can be considered to be not'ed values... 53 if (ConstantInt *C = dyn_cast<ConstantInt>(V)) 54 return ConstantInt::get(C->getType(), ~C->getValue()); 55 return nullptr; 56 } 57 58 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp 59 /// predicate into a three bit mask. It also returns whether it is an ordered 60 /// predicate by reference. 61 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) { 62 isOrdered = false; 63 switch (CC) { 64 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000 65 case FCmpInst::FCMP_UNO: return 0; // 000 66 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001 67 case FCmpInst::FCMP_UGT: return 1; // 001 68 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010 69 case FCmpInst::FCMP_UEQ: return 2; // 010 70 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011 71 case FCmpInst::FCMP_UGE: return 3; // 011 72 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100 73 case FCmpInst::FCMP_ULT: return 4; // 100 74 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101 75 case FCmpInst::FCMP_UNE: return 5; // 101 76 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110 77 case FCmpInst::FCMP_ULE: return 6; // 110 78 // True -> 7 79 default: 80 // Not expecting FCMP_FALSE and FCMP_TRUE; 81 llvm_unreachable("Unexpected FCmp predicate!"); 82 } 83 } 84 85 /// getNewICmpValue - This is the complement of getICmpCode, which turns an 86 /// opcode and two operands into either a constant true or false, or a brand 87 /// new ICmp instruction. The sign is passed in to determine which kind 88 /// of predicate to use in the new icmp instruction. 89 static Value *getNewICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS, 90 InstCombiner::BuilderTy *Builder) { 91 ICmpInst::Predicate NewPred; 92 if (Value *NewConstant = getICmpValue(Sign, Code, LHS, RHS, NewPred)) 93 return NewConstant; 94 return Builder->CreateICmp(NewPred, LHS, RHS); 95 } 96 97 /// getFCmpValue - This is the complement of getFCmpCode, which turns an 98 /// opcode and two operands into either a FCmp instruction. isordered is passed 99 /// in to determine which kind of predicate to use in the new fcmp instruction. 100 static Value *getFCmpValue(bool isordered, unsigned code, 101 Value *LHS, Value *RHS, 102 InstCombiner::BuilderTy *Builder) { 103 CmpInst::Predicate Pred; 104 switch (code) { 105 default: llvm_unreachable("Illegal FCmp code!"); 106 case 0: Pred = isordered ? FCmpInst::FCMP_ORD : FCmpInst::FCMP_UNO; break; 107 case 1: Pred = isordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; break; 108 case 2: Pred = isordered ? FCmpInst::FCMP_OEQ : FCmpInst::FCMP_UEQ; break; 109 case 3: Pred = isordered ? FCmpInst::FCMP_OGE : FCmpInst::FCMP_UGE; break; 110 case 4: Pred = isordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; break; 111 case 5: Pred = isordered ? FCmpInst::FCMP_ONE : FCmpInst::FCMP_UNE; break; 112 case 6: Pred = isordered ? FCmpInst::FCMP_OLE : FCmpInst::FCMP_ULE; break; 113 case 7: 114 if (!isordered) return ConstantInt::getTrue(LHS->getContext()); 115 Pred = FCmpInst::FCMP_ORD; break; 116 } 117 return Builder->CreateFCmp(Pred, LHS, RHS); 118 } 119 120 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where 121 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is 122 // guaranteed to be a binary operator. 123 Instruction *InstCombiner::OptAndOp(Instruction *Op, 124 ConstantInt *OpRHS, 125 ConstantInt *AndRHS, 126 BinaryOperator &TheAnd) { 127 Value *X = Op->getOperand(0); 128 Constant *Together = nullptr; 129 if (!Op->isShift()) 130 Together = ConstantExpr::getAnd(AndRHS, OpRHS); 131 132 switch (Op->getOpcode()) { 133 case Instruction::Xor: 134 if (Op->hasOneUse()) { 135 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2) 136 Value *And = Builder->CreateAnd(X, AndRHS); 137 And->takeName(Op); 138 return BinaryOperator::CreateXor(And, Together); 139 } 140 break; 141 case Instruction::Or: 142 if (Op->hasOneUse()){ 143 if (Together != OpRHS) { 144 // (X | C1) & C2 --> (X | (C1&C2)) & C2 145 Value *Or = Builder->CreateOr(X, Together); 146 Or->takeName(Op); 147 return BinaryOperator::CreateAnd(Or, AndRHS); 148 } 149 150 ConstantInt *TogetherCI = dyn_cast<ConstantInt>(Together); 151 if (TogetherCI && !TogetherCI->isZero()){ 152 // (X | C1) & C2 --> (X & (C2^(C1&C2))) | C1 153 // NOTE: This reduces the number of bits set in the & mask, which 154 // can expose opportunities for store narrowing. 155 Together = ConstantExpr::getXor(AndRHS, Together); 156 Value *And = Builder->CreateAnd(X, Together); 157 And->takeName(Op); 158 return BinaryOperator::CreateOr(And, OpRHS); 159 } 160 } 161 162 break; 163 case Instruction::Add: 164 if (Op->hasOneUse()) { 165 // Adding a one to a single bit bit-field should be turned into an XOR 166 // of the bit. First thing to check is to see if this AND is with a 167 // single bit constant. 168 const APInt &AndRHSV = AndRHS->getValue(); 169 170 // If there is only one bit set. 171 if (AndRHSV.isPowerOf2()) { 172 // Ok, at this point, we know that we are masking the result of the 173 // ADD down to exactly one bit. If the constant we are adding has 174 // no bits set below this bit, then we can eliminate the ADD. 175 const APInt& AddRHS = OpRHS->getValue(); 176 177 // Check to see if any bits below the one bit set in AndRHSV are set. 178 if ((AddRHS & (AndRHSV-1)) == 0) { 179 // If not, the only thing that can effect the output of the AND is 180 // the bit specified by AndRHSV. If that bit is set, the effect of 181 // the XOR is to toggle the bit. If it is clear, then the ADD has 182 // no effect. 183 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop 184 TheAnd.setOperand(0, X); 185 return &TheAnd; 186 } else { 187 // Pull the XOR out of the AND. 188 Value *NewAnd = Builder->CreateAnd(X, AndRHS); 189 NewAnd->takeName(Op); 190 return BinaryOperator::CreateXor(NewAnd, AndRHS); 191 } 192 } 193 } 194 } 195 break; 196 197 case Instruction::Shl: { 198 // We know that the AND will not produce any of the bits shifted in, so if 199 // the anded constant includes them, clear them now! 200 // 201 uint32_t BitWidth = AndRHS->getType()->getBitWidth(); 202 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth); 203 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal)); 204 ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShlMask); 205 206 if (CI->getValue() == ShlMask) 207 // Masking out bits that the shift already masks. 208 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and. 209 210 if (CI != AndRHS) { // Reducing bits set in and. 211 TheAnd.setOperand(1, CI); 212 return &TheAnd; 213 } 214 break; 215 } 216 case Instruction::LShr: { 217 // We know that the AND will not produce any of the bits shifted in, so if 218 // the anded constant includes them, clear them now! This only applies to 219 // unsigned shifts, because a signed shr may bring in set bits! 220 // 221 uint32_t BitWidth = AndRHS->getType()->getBitWidth(); 222 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth); 223 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal)); 224 ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShrMask); 225 226 if (CI->getValue() == ShrMask) 227 // Masking out bits that the shift already masks. 228 return ReplaceInstUsesWith(TheAnd, Op); 229 230 if (CI != AndRHS) { 231 TheAnd.setOperand(1, CI); // Reduce bits set in and cst. 232 return &TheAnd; 233 } 234 break; 235 } 236 case Instruction::AShr: 237 // Signed shr. 238 // See if this is shifting in some sign extension, then masking it out 239 // with an and. 240 if (Op->hasOneUse()) { 241 uint32_t BitWidth = AndRHS->getType()->getBitWidth(); 242 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth); 243 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal)); 244 Constant *C = Builder->getInt(AndRHS->getValue() & ShrMask); 245 if (C == AndRHS) { // Masking out bits shifted in. 246 // (Val ashr C1) & C2 -> (Val lshr C1) & C2 247 // Make the argument unsigned. 248 Value *ShVal = Op->getOperand(0); 249 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName()); 250 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName()); 251 } 252 } 253 break; 254 } 255 return nullptr; 256 } 257 258 /// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise 259 /// (V < Lo || V >= Hi). In practice, we emit the more efficient 260 /// (V-Lo) \<u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates 261 /// whether to treat the V, Lo and HI as signed or not. IB is the location to 262 /// insert new instructions. 263 Value *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi, 264 bool isSigned, bool Inside) { 265 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ? 266 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() && 267 "Lo is not <= Hi in range emission code!"); 268 269 if (Inside) { 270 if (Lo == Hi) // Trivially false. 271 return Builder->getFalse(); 272 273 // V >= Min && V < Hi --> V < Hi 274 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) { 275 ICmpInst::Predicate pred = (isSigned ? 276 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT); 277 return Builder->CreateICmp(pred, V, Hi); 278 } 279 280 // Emit V-Lo <u Hi-Lo 281 Constant *NegLo = ConstantExpr::getNeg(Lo); 282 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off"); 283 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi); 284 return Builder->CreateICmpULT(Add, UpperBound); 285 } 286 287 if (Lo == Hi) // Trivially true. 288 return Builder->getTrue(); 289 290 // V < Min || V >= Hi -> V > Hi-1 291 Hi = SubOne(cast<ConstantInt>(Hi)); 292 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) { 293 ICmpInst::Predicate pred = (isSigned ? 294 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT); 295 return Builder->CreateICmp(pred, V, Hi); 296 } 297 298 // Emit V-Lo >u Hi-1-Lo 299 // Note that Hi has already had one subtracted from it, above. 300 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo)); 301 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off"); 302 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi); 303 return Builder->CreateICmpUGT(Add, LowerBound); 304 } 305 306 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with 307 // any number of 0s on either side. The 1s are allowed to wrap from LSB to 308 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is 309 // not, since all 1s are not contiguous. 310 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) { 311 const APInt& V = Val->getValue(); 312 uint32_t BitWidth = Val->getType()->getBitWidth(); 313 if (!APIntOps::isShiftedMask(BitWidth, V)) return false; 314 315 // look for the first zero bit after the run of ones 316 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros(); 317 // look for the first non-zero bit 318 ME = V.getActiveBits(); 319 return true; 320 } 321 322 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask, 323 /// where isSub determines whether the operator is a sub. If we can fold one of 324 /// the following xforms: 325 /// 326 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask 327 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0 328 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0 329 /// 330 /// return (A +/- B). 331 /// 332 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS, 333 ConstantInt *Mask, bool isSub, 334 Instruction &I) { 335 Instruction *LHSI = dyn_cast<Instruction>(LHS); 336 if (!LHSI || LHSI->getNumOperands() != 2 || 337 !isa<ConstantInt>(LHSI->getOperand(1))) return nullptr; 338 339 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1)); 340 341 switch (LHSI->getOpcode()) { 342 default: return nullptr; 343 case Instruction::And: 344 if (ConstantExpr::getAnd(N, Mask) == Mask) { 345 // If the AndRHS is a power of two minus one (0+1+), this is simple. 346 if ((Mask->getValue().countLeadingZeros() + 347 Mask->getValue().countPopulation()) == 348 Mask->getValue().getBitWidth()) 349 break; 350 351 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+ 352 // part, we don't need any explicit masks to take them out of A. If that 353 // is all N is, ignore it. 354 uint32_t MB = 0, ME = 0; 355 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive 356 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth(); 357 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1)); 358 if (MaskedValueIsZero(RHS, Mask)) 359 break; 360 } 361 } 362 return nullptr; 363 case Instruction::Or: 364 case Instruction::Xor: 365 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0 366 if ((Mask->getValue().countLeadingZeros() + 367 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth() 368 && ConstantExpr::getAnd(N, Mask)->isNullValue()) 369 break; 370 return nullptr; 371 } 372 373 if (isSub) 374 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold"); 375 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold"); 376 } 377 378 /// enum for classifying (icmp eq (A & B), C) and (icmp ne (A & B), C) 379 /// One of A and B is considered the mask, the other the value. This is 380 /// described as the "AMask" or "BMask" part of the enum. If the enum 381 /// contains only "Mask", then both A and B can be considered masks. 382 /// If A is the mask, then it was proven, that (A & C) == C. This 383 /// is trivial if C == A, or C == 0. If both A and C are constants, this 384 /// proof is also easy. 385 /// For the following explanations we assume that A is the mask. 386 /// The part "AllOnes" declares, that the comparison is true only 387 /// if (A & B) == A, or all bits of A are set in B. 388 /// Example: (icmp eq (A & 3), 3) -> FoldMskICmp_AMask_AllOnes 389 /// The part "AllZeroes" declares, that the comparison is true only 390 /// if (A & B) == 0, or all bits of A are cleared in B. 391 /// Example: (icmp eq (A & 3), 0) -> FoldMskICmp_Mask_AllZeroes 392 /// The part "Mixed" declares, that (A & B) == C and C might or might not 393 /// contain any number of one bits and zero bits. 394 /// Example: (icmp eq (A & 3), 1) -> FoldMskICmp_AMask_Mixed 395 /// The Part "Not" means, that in above descriptions "==" should be replaced 396 /// by "!=". 397 /// Example: (icmp ne (A & 3), 3) -> FoldMskICmp_AMask_NotAllOnes 398 /// If the mask A contains a single bit, then the following is equivalent: 399 /// (icmp eq (A & B), A) equals (icmp ne (A & B), 0) 400 /// (icmp ne (A & B), A) equals (icmp eq (A & B), 0) 401 enum MaskedICmpType { 402 FoldMskICmp_AMask_AllOnes = 1, 403 FoldMskICmp_AMask_NotAllOnes = 2, 404 FoldMskICmp_BMask_AllOnes = 4, 405 FoldMskICmp_BMask_NotAllOnes = 8, 406 FoldMskICmp_Mask_AllZeroes = 16, 407 FoldMskICmp_Mask_NotAllZeroes = 32, 408 FoldMskICmp_AMask_Mixed = 64, 409 FoldMskICmp_AMask_NotMixed = 128, 410 FoldMskICmp_BMask_Mixed = 256, 411 FoldMskICmp_BMask_NotMixed = 512 412 }; 413 414 /// return the set of pattern classes (from MaskedICmpType) 415 /// that (icmp SCC (A & B), C) satisfies 416 static unsigned getTypeOfMaskedICmp(Value* A, Value* B, Value* C, 417 ICmpInst::Predicate SCC) 418 { 419 ConstantInt *ACst = dyn_cast<ConstantInt>(A); 420 ConstantInt *BCst = dyn_cast<ConstantInt>(B); 421 ConstantInt *CCst = dyn_cast<ConstantInt>(C); 422 bool icmp_eq = (SCC == ICmpInst::ICMP_EQ); 423 bool icmp_abit = (ACst && !ACst->isZero() && 424 ACst->getValue().isPowerOf2()); 425 bool icmp_bbit = (BCst && !BCst->isZero() && 426 BCst->getValue().isPowerOf2()); 427 unsigned result = 0; 428 if (CCst && CCst->isZero()) { 429 // if C is zero, then both A and B qualify as mask 430 result |= (icmp_eq ? (FoldMskICmp_Mask_AllZeroes | 431 FoldMskICmp_Mask_AllZeroes | 432 FoldMskICmp_AMask_Mixed | 433 FoldMskICmp_BMask_Mixed) 434 : (FoldMskICmp_Mask_NotAllZeroes | 435 FoldMskICmp_Mask_NotAllZeroes | 436 FoldMskICmp_AMask_NotMixed | 437 FoldMskICmp_BMask_NotMixed)); 438 if (icmp_abit) 439 result |= (icmp_eq ? (FoldMskICmp_AMask_NotAllOnes | 440 FoldMskICmp_AMask_NotMixed) 441 : (FoldMskICmp_AMask_AllOnes | 442 FoldMskICmp_AMask_Mixed)); 443 if (icmp_bbit) 444 result |= (icmp_eq ? (FoldMskICmp_BMask_NotAllOnes | 445 FoldMskICmp_BMask_NotMixed) 446 : (FoldMskICmp_BMask_AllOnes | 447 FoldMskICmp_BMask_Mixed)); 448 return result; 449 } 450 if (A == C) { 451 result |= (icmp_eq ? (FoldMskICmp_AMask_AllOnes | 452 FoldMskICmp_AMask_Mixed) 453 : (FoldMskICmp_AMask_NotAllOnes | 454 FoldMskICmp_AMask_NotMixed)); 455 if (icmp_abit) 456 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes | 457 FoldMskICmp_AMask_NotMixed) 458 : (FoldMskICmp_Mask_AllZeroes | 459 FoldMskICmp_AMask_Mixed)); 460 } else if (ACst && CCst && 461 ConstantExpr::getAnd(ACst, CCst) == CCst) { 462 result |= (icmp_eq ? FoldMskICmp_AMask_Mixed 463 : FoldMskICmp_AMask_NotMixed); 464 } 465 if (B == C) { 466 result |= (icmp_eq ? (FoldMskICmp_BMask_AllOnes | 467 FoldMskICmp_BMask_Mixed) 468 : (FoldMskICmp_BMask_NotAllOnes | 469 FoldMskICmp_BMask_NotMixed)); 470 if (icmp_bbit) 471 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes | 472 FoldMskICmp_BMask_NotMixed) 473 : (FoldMskICmp_Mask_AllZeroes | 474 FoldMskICmp_BMask_Mixed)); 475 } else if (BCst && CCst && 476 ConstantExpr::getAnd(BCst, CCst) == CCst) { 477 result |= (icmp_eq ? FoldMskICmp_BMask_Mixed 478 : FoldMskICmp_BMask_NotMixed); 479 } 480 return result; 481 } 482 483 /// Convert an analysis of a masked ICmp into its equivalent if all boolean 484 /// operations had the opposite sense. Since each "NotXXX" flag (recording !=) 485 /// is adjacent to the corresponding normal flag (recording ==), this just 486 /// involves swapping those bits over. 487 static unsigned conjugateICmpMask(unsigned Mask) { 488 unsigned NewMask; 489 NewMask = (Mask & (FoldMskICmp_AMask_AllOnes | FoldMskICmp_BMask_AllOnes | 490 FoldMskICmp_Mask_AllZeroes | FoldMskICmp_AMask_Mixed | 491 FoldMskICmp_BMask_Mixed)) 492 << 1; 493 494 NewMask |= 495 (Mask & (FoldMskICmp_AMask_NotAllOnes | FoldMskICmp_BMask_NotAllOnes | 496 FoldMskICmp_Mask_NotAllZeroes | FoldMskICmp_AMask_NotMixed | 497 FoldMskICmp_BMask_NotMixed)) 498 >> 1; 499 500 return NewMask; 501 } 502 503 /// decomposeBitTestICmp - Decompose an icmp into the form ((X & Y) pred Z) 504 /// if possible. The returned predicate is either == or !=. Returns false if 505 /// decomposition fails. 506 static bool decomposeBitTestICmp(const ICmpInst *I, ICmpInst::Predicate &Pred, 507 Value *&X, Value *&Y, Value *&Z) { 508 ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1)); 509 if (!C) 510 return false; 511 512 switch (I->getPredicate()) { 513 default: 514 return false; 515 case ICmpInst::ICMP_SLT: 516 // X < 0 is equivalent to (X & SignBit) != 0. 517 if (!C->isZero()) 518 return false; 519 Y = ConstantInt::get(I->getContext(), APInt::getSignBit(C->getBitWidth())); 520 Pred = ICmpInst::ICMP_NE; 521 break; 522 case ICmpInst::ICMP_SGT: 523 // X > -1 is equivalent to (X & SignBit) == 0. 524 if (!C->isAllOnesValue()) 525 return false; 526 Y = ConstantInt::get(I->getContext(), APInt::getSignBit(C->getBitWidth())); 527 Pred = ICmpInst::ICMP_EQ; 528 break; 529 case ICmpInst::ICMP_ULT: 530 // X <u 2^n is equivalent to (X & ~(2^n-1)) == 0. 531 if (!C->getValue().isPowerOf2()) 532 return false; 533 Y = ConstantInt::get(I->getContext(), -C->getValue()); 534 Pred = ICmpInst::ICMP_EQ; 535 break; 536 case ICmpInst::ICMP_UGT: 537 // X >u 2^n-1 is equivalent to (X & ~(2^n-1)) != 0. 538 if (!(C->getValue() + 1).isPowerOf2()) 539 return false; 540 Y = ConstantInt::get(I->getContext(), ~C->getValue()); 541 Pred = ICmpInst::ICMP_NE; 542 break; 543 } 544 545 X = I->getOperand(0); 546 Z = ConstantInt::getNullValue(C->getType()); 547 return true; 548 } 549 550 /// foldLogOpOfMaskedICmpsHelper: 551 /// handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E) 552 /// return the set of pattern classes (from MaskedICmpType) 553 /// that both LHS and RHS satisfy 554 static unsigned foldLogOpOfMaskedICmpsHelper(Value*& A, 555 Value*& B, Value*& C, 556 Value*& D, Value*& E, 557 ICmpInst *LHS, ICmpInst *RHS, 558 ICmpInst::Predicate &LHSCC, 559 ICmpInst::Predicate &RHSCC) { 560 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) return 0; 561 // vectors are not (yet?) supported 562 if (LHS->getOperand(0)->getType()->isVectorTy()) return 0; 563 564 // Here comes the tricky part: 565 // LHS might be of the form L11 & L12 == X, X == L21 & L22, 566 // and L11 & L12 == L21 & L22. The same goes for RHS. 567 // Now we must find those components L** and R**, that are equal, so 568 // that we can extract the parameters A, B, C, D, and E for the canonical 569 // above. 570 Value *L1 = LHS->getOperand(0); 571 Value *L2 = LHS->getOperand(1); 572 Value *L11,*L12,*L21,*L22; 573 // Check whether the icmp can be decomposed into a bit test. 574 if (decomposeBitTestICmp(LHS, LHSCC, L11, L12, L2)) { 575 L21 = L22 = L1 = nullptr; 576 } else { 577 // Look for ANDs in the LHS icmp. 578 if (!L1->getType()->isIntegerTy()) { 579 // You can icmp pointers, for example. They really aren't masks. 580 L11 = L12 = nullptr; 581 } else if (!match(L1, m_And(m_Value(L11), m_Value(L12)))) { 582 // Any icmp can be viewed as being trivially masked; if it allows us to 583 // remove one, it's worth it. 584 L11 = L1; 585 L12 = Constant::getAllOnesValue(L1->getType()); 586 } 587 588 if (!L2->getType()->isIntegerTy()) { 589 // You can icmp pointers, for example. They really aren't masks. 590 L21 = L22 = nullptr; 591 } else if (!match(L2, m_And(m_Value(L21), m_Value(L22)))) { 592 L21 = L2; 593 L22 = Constant::getAllOnesValue(L2->getType()); 594 } 595 } 596 597 // Bail if LHS was a icmp that can't be decomposed into an equality. 598 if (!ICmpInst::isEquality(LHSCC)) 599 return 0; 600 601 Value *R1 = RHS->getOperand(0); 602 Value *R2 = RHS->getOperand(1); 603 Value *R11,*R12; 604 bool ok = false; 605 if (decomposeBitTestICmp(RHS, RHSCC, R11, R12, R2)) { 606 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) { 607 A = R11; D = R12; 608 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) { 609 A = R12; D = R11; 610 } else { 611 return 0; 612 } 613 E = R2; R1 = nullptr; ok = true; 614 } else if (R1->getType()->isIntegerTy()) { 615 if (!match(R1, m_And(m_Value(R11), m_Value(R12)))) { 616 // As before, model no mask as a trivial mask if it'll let us do an 617 // optimisation. 618 R11 = R1; 619 R12 = Constant::getAllOnesValue(R1->getType()); 620 } 621 622 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) { 623 A = R11; D = R12; E = R2; ok = true; 624 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) { 625 A = R12; D = R11; E = R2; ok = true; 626 } 627 } 628 629 // Bail if RHS was a icmp that can't be decomposed into an equality. 630 if (!ICmpInst::isEquality(RHSCC)) 631 return 0; 632 633 // Look for ANDs in on the right side of the RHS icmp. 634 if (!ok && R2->getType()->isIntegerTy()) { 635 if (!match(R2, m_And(m_Value(R11), m_Value(R12)))) { 636 R11 = R2; 637 R12 = Constant::getAllOnesValue(R2->getType()); 638 } 639 640 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) { 641 A = R11; D = R12; E = R1; ok = true; 642 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) { 643 A = R12; D = R11; E = R1; ok = true; 644 } else { 645 return 0; 646 } 647 } 648 if (!ok) 649 return 0; 650 651 if (L11 == A) { 652 B = L12; C = L2; 653 } else if (L12 == A) { 654 B = L11; C = L2; 655 } else if (L21 == A) { 656 B = L22; C = L1; 657 } else if (L22 == A) { 658 B = L21; C = L1; 659 } 660 661 unsigned left_type = getTypeOfMaskedICmp(A, B, C, LHSCC); 662 unsigned right_type = getTypeOfMaskedICmp(A, D, E, RHSCC); 663 return left_type & right_type; 664 } 665 /// foldLogOpOfMaskedICmps: 666 /// try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E) 667 /// into a single (icmp(A & X) ==/!= Y) 668 static Value* foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS, bool IsAnd, 669 llvm::InstCombiner::BuilderTy* Builder) { 670 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr; 671 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate(); 672 unsigned mask = foldLogOpOfMaskedICmpsHelper(A, B, C, D, E, LHS, RHS, 673 LHSCC, RHSCC); 674 if (mask == 0) return nullptr; 675 assert(ICmpInst::isEquality(LHSCC) && ICmpInst::isEquality(RHSCC) && 676 "foldLogOpOfMaskedICmpsHelper must return an equality predicate."); 677 678 // In full generality: 679 // (icmp (A & B) Op C) | (icmp (A & D) Op E) 680 // == ![ (icmp (A & B) !Op C) & (icmp (A & D) !Op E) ] 681 // 682 // If the latter can be converted into (icmp (A & X) Op Y) then the former is 683 // equivalent to (icmp (A & X) !Op Y). 684 // 685 // Therefore, we can pretend for the rest of this function that we're dealing 686 // with the conjunction, provided we flip the sense of any comparisons (both 687 // input and output). 688 689 // In most cases we're going to produce an EQ for the "&&" case. 690 ICmpInst::Predicate NEWCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE; 691 if (!IsAnd) { 692 // Convert the masking analysis into its equivalent with negated 693 // comparisons. 694 mask = conjugateICmpMask(mask); 695 } 696 697 if (mask & FoldMskICmp_Mask_AllZeroes) { 698 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0) 699 // -> (icmp eq (A & (B|D)), 0) 700 Value* newOr = Builder->CreateOr(B, D); 701 Value* newAnd = Builder->CreateAnd(A, newOr); 702 // we can't use C as zero, because we might actually handle 703 // (icmp ne (A & B), B) & (icmp ne (A & D), D) 704 // with B and D, having a single bit set 705 Value* zero = Constant::getNullValue(A->getType()); 706 return Builder->CreateICmp(NEWCC, newAnd, zero); 707 } 708 if (mask & FoldMskICmp_BMask_AllOnes) { 709 // (icmp eq (A & B), B) & (icmp eq (A & D), D) 710 // -> (icmp eq (A & (B|D)), (B|D)) 711 Value* newOr = Builder->CreateOr(B, D); 712 Value* newAnd = Builder->CreateAnd(A, newOr); 713 return Builder->CreateICmp(NEWCC, newAnd, newOr); 714 } 715 if (mask & FoldMskICmp_AMask_AllOnes) { 716 // (icmp eq (A & B), A) & (icmp eq (A & D), A) 717 // -> (icmp eq (A & (B&D)), A) 718 Value* newAnd1 = Builder->CreateAnd(B, D); 719 Value* newAnd = Builder->CreateAnd(A, newAnd1); 720 return Builder->CreateICmp(NEWCC, newAnd, A); 721 } 722 723 // Remaining cases assume at least that B and D are constant, and depend on 724 // their actual values. This isn't strictly, necessary, just a "handle the 725 // easy cases for now" decision. 726 ConstantInt *BCst = dyn_cast<ConstantInt>(B); 727 if (!BCst) return nullptr; 728 ConstantInt *DCst = dyn_cast<ConstantInt>(D); 729 if (!DCst) return nullptr; 730 731 if (mask & (FoldMskICmp_Mask_NotAllZeroes | FoldMskICmp_BMask_NotAllOnes)) { 732 // (icmp ne (A & B), 0) & (icmp ne (A & D), 0) and 733 // (icmp ne (A & B), B) & (icmp ne (A & D), D) 734 // -> (icmp ne (A & B), 0) or (icmp ne (A & D), 0) 735 // Only valid if one of the masks is a superset of the other (check "B&D" is 736 // the same as either B or D). 737 APInt NewMask = BCst->getValue() & DCst->getValue(); 738 739 if (NewMask == BCst->getValue()) 740 return LHS; 741 else if (NewMask == DCst->getValue()) 742 return RHS; 743 } 744 if (mask & FoldMskICmp_AMask_NotAllOnes) { 745 // (icmp ne (A & B), B) & (icmp ne (A & D), D) 746 // -> (icmp ne (A & B), A) or (icmp ne (A & D), A) 747 // Only valid if one of the masks is a superset of the other (check "B|D" is 748 // the same as either B or D). 749 APInt NewMask = BCst->getValue() | DCst->getValue(); 750 751 if (NewMask == BCst->getValue()) 752 return LHS; 753 else if (NewMask == DCst->getValue()) 754 return RHS; 755 } 756 if (mask & FoldMskICmp_BMask_Mixed) { 757 // (icmp eq (A & B), C) & (icmp eq (A & D), E) 758 // We already know that B & C == C && D & E == E. 759 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of 760 // C and E, which are shared by both the mask B and the mask D, don't 761 // contradict, then we can transform to 762 // -> (icmp eq (A & (B|D)), (C|E)) 763 // Currently, we only handle the case of B, C, D, and E being constant. 764 // we can't simply use C and E, because we might actually handle 765 // (icmp ne (A & B), B) & (icmp eq (A & D), D) 766 // with B and D, having a single bit set 767 ConstantInt *CCst = dyn_cast<ConstantInt>(C); 768 if (!CCst) return nullptr; 769 if (LHSCC != NEWCC) 770 CCst = dyn_cast<ConstantInt>( ConstantExpr::getXor(BCst, CCst) ); 771 ConstantInt *ECst = dyn_cast<ConstantInt>(E); 772 if (!ECst) return nullptr; 773 if (RHSCC != NEWCC) 774 ECst = dyn_cast<ConstantInt>( ConstantExpr::getXor(DCst, ECst) ); 775 ConstantInt* MCst = dyn_cast<ConstantInt>( 776 ConstantExpr::getAnd(ConstantExpr::getAnd(BCst, DCst), 777 ConstantExpr::getXor(CCst, ECst)) ); 778 // if there is a conflict we should actually return a false for the 779 // whole construct 780 if (!MCst->isZero()) 781 return nullptr; 782 Value *newOr1 = Builder->CreateOr(B, D); 783 Value *newOr2 = ConstantExpr::getOr(CCst, ECst); 784 Value *newAnd = Builder->CreateAnd(A, newOr1); 785 return Builder->CreateICmp(NEWCC, newAnd, newOr2); 786 } 787 return nullptr; 788 } 789 790 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible. 791 Value *InstCombiner::FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS) { 792 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate(); 793 794 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B) 795 if (PredicatesFoldable(LHSCC, RHSCC)) { 796 if (LHS->getOperand(0) == RHS->getOperand(1) && 797 LHS->getOperand(1) == RHS->getOperand(0)) 798 LHS->swapOperands(); 799 if (LHS->getOperand(0) == RHS->getOperand(0) && 800 LHS->getOperand(1) == RHS->getOperand(1)) { 801 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1); 802 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS); 803 bool isSigned = LHS->isSigned() || RHS->isSigned(); 804 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder); 805 } 806 } 807 808 // handle (roughly): (icmp eq (A & B), C) & (icmp eq (A & D), E) 809 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, true, Builder)) 810 return V; 811 812 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2). 813 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0); 814 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1)); 815 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1)); 816 if (!LHSCst || !RHSCst) return nullptr; 817 818 if (LHSCst == RHSCst && LHSCC == RHSCC) { 819 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C) 820 // where C is a power of 2 821 if (LHSCC == ICmpInst::ICMP_ULT && 822 LHSCst->getValue().isPowerOf2()) { 823 Value *NewOr = Builder->CreateOr(Val, Val2); 824 return Builder->CreateICmp(LHSCC, NewOr, LHSCst); 825 } 826 827 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0) 828 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) { 829 Value *NewOr = Builder->CreateOr(Val, Val2); 830 return Builder->CreateICmp(LHSCC, NewOr, LHSCst); 831 } 832 } 833 834 // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2 835 // where CMAX is the all ones value for the truncated type, 836 // iff the lower bits of C2 and CA are zero. 837 if (LHSCC == ICmpInst::ICMP_EQ && LHSCC == RHSCC && 838 LHS->hasOneUse() && RHS->hasOneUse()) { 839 Value *V; 840 ConstantInt *AndCst, *SmallCst = nullptr, *BigCst = nullptr; 841 842 // (trunc x) == C1 & (and x, CA) == C2 843 // (and x, CA) == C2 & (trunc x) == C1 844 if (match(Val2, m_Trunc(m_Value(V))) && 845 match(Val, m_And(m_Specific(V), m_ConstantInt(AndCst)))) { 846 SmallCst = RHSCst; 847 BigCst = LHSCst; 848 } else if (match(Val, m_Trunc(m_Value(V))) && 849 match(Val2, m_And(m_Specific(V), m_ConstantInt(AndCst)))) { 850 SmallCst = LHSCst; 851 BigCst = RHSCst; 852 } 853 854 if (SmallCst && BigCst) { 855 unsigned BigBitSize = BigCst->getType()->getBitWidth(); 856 unsigned SmallBitSize = SmallCst->getType()->getBitWidth(); 857 858 // Check that the low bits are zero. 859 APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize); 860 if ((Low & AndCst->getValue()) == 0 && (Low & BigCst->getValue()) == 0) { 861 Value *NewAnd = Builder->CreateAnd(V, Low | AndCst->getValue()); 862 APInt N = SmallCst->getValue().zext(BigBitSize) | BigCst->getValue(); 863 Value *NewVal = ConstantInt::get(AndCst->getType()->getContext(), N); 864 return Builder->CreateICmp(LHSCC, NewAnd, NewVal); 865 } 866 } 867 } 868 869 // From here on, we only handle: 870 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler. 871 if (Val != Val2) return nullptr; 872 873 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere. 874 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE || 875 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE || 876 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE || 877 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE) 878 return nullptr; 879 880 // Make a constant range that's the intersection of the two icmp ranges. 881 // If the intersection is empty, we know that the result is false. 882 ConstantRange LHSRange = 883 ConstantRange::makeICmpRegion(LHSCC, LHSCst->getValue()); 884 ConstantRange RHSRange = 885 ConstantRange::makeICmpRegion(RHSCC, RHSCst->getValue()); 886 887 if (LHSRange.intersectWith(RHSRange).isEmptySet()) 888 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 889 890 // We can't fold (ugt x, C) & (sgt x, C2). 891 if (!PredicatesFoldable(LHSCC, RHSCC)) 892 return nullptr; 893 894 // Ensure that the larger constant is on the RHS. 895 bool ShouldSwap; 896 if (CmpInst::isSigned(LHSCC) || 897 (ICmpInst::isEquality(LHSCC) && 898 CmpInst::isSigned(RHSCC))) 899 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue()); 900 else 901 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue()); 902 903 if (ShouldSwap) { 904 std::swap(LHS, RHS); 905 std::swap(LHSCst, RHSCst); 906 std::swap(LHSCC, RHSCC); 907 } 908 909 // At this point, we know we have two icmp instructions 910 // comparing a value against two constants and and'ing the result 911 // together. Because of the above check, we know that we only have 912 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know 913 // (from the icmp folding check above), that the two constants 914 // are not equal and that the larger constant is on the RHS 915 assert(LHSCst != RHSCst && "Compares not folded above?"); 916 917 switch (LHSCC) { 918 default: llvm_unreachable("Unknown integer condition code!"); 919 case ICmpInst::ICMP_EQ: 920 switch (RHSCC) { 921 default: llvm_unreachable("Unknown integer condition code!"); 922 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13 923 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13 924 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13 925 return LHS; 926 } 927 case ICmpInst::ICMP_NE: 928 switch (RHSCC) { 929 default: llvm_unreachable("Unknown integer condition code!"); 930 case ICmpInst::ICMP_ULT: 931 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13 932 return Builder->CreateICmpULT(Val, LHSCst); 933 break; // (X != 13 & X u< 15) -> no change 934 case ICmpInst::ICMP_SLT: 935 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13 936 return Builder->CreateICmpSLT(Val, LHSCst); 937 break; // (X != 13 & X s< 15) -> no change 938 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15 939 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15 940 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15 941 return RHS; 942 case ICmpInst::ICMP_NE: 943 // Special case to get the ordering right when the values wrap around 944 // zero. 945 if (LHSCst->getValue() == 0 && RHSCst->getValue().isAllOnesValue()) 946 std::swap(LHSCst, RHSCst); 947 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1 948 Constant *AddCST = ConstantExpr::getNeg(LHSCst); 949 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off"); 950 return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1), 951 Val->getName()+".cmp"); 952 } 953 break; // (X != 13 & X != 15) -> no change 954 } 955 break; 956 case ICmpInst::ICMP_ULT: 957 switch (RHSCC) { 958 default: llvm_unreachable("Unknown integer condition code!"); 959 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false 960 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false 961 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 962 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change 963 break; 964 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13 965 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13 966 return LHS; 967 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change 968 break; 969 } 970 break; 971 case ICmpInst::ICMP_SLT: 972 switch (RHSCC) { 973 default: llvm_unreachable("Unknown integer condition code!"); 974 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change 975 break; 976 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13 977 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13 978 return LHS; 979 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change 980 break; 981 } 982 break; 983 case ICmpInst::ICMP_UGT: 984 switch (RHSCC) { 985 default: llvm_unreachable("Unknown integer condition code!"); 986 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15 987 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15 988 return RHS; 989 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change 990 break; 991 case ICmpInst::ICMP_NE: 992 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14 993 return Builder->CreateICmp(LHSCC, Val, RHSCst); 994 break; // (X u> 13 & X != 15) -> no change 995 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1 996 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true); 997 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change 998 break; 999 } 1000 break; 1001 case ICmpInst::ICMP_SGT: 1002 switch (RHSCC) { 1003 default: llvm_unreachable("Unknown integer condition code!"); 1004 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15 1005 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15 1006 return RHS; 1007 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change 1008 break; 1009 case ICmpInst::ICMP_NE: 1010 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14 1011 return Builder->CreateICmp(LHSCC, Val, RHSCst); 1012 break; // (X s> 13 & X != 15) -> no change 1013 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1 1014 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true); 1015 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change 1016 break; 1017 } 1018 break; 1019 } 1020 1021 return nullptr; 1022 } 1023 1024 /// FoldAndOfFCmps - Optimize (fcmp)&(fcmp). NOTE: Unlike the rest of 1025 /// instcombine, this returns a Value which should already be inserted into the 1026 /// function. 1027 Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) { 1028 if (LHS->getPredicate() == FCmpInst::FCMP_ORD && 1029 RHS->getPredicate() == FCmpInst::FCMP_ORD) { 1030 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) 1031 return nullptr; 1032 1033 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y) 1034 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1))) 1035 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) { 1036 // If either of the constants are nans, then the whole thing returns 1037 // false. 1038 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN()) 1039 return Builder->getFalse(); 1040 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0)); 1041 } 1042 1043 // Handle vector zeros. This occurs because the canonical form of 1044 // "fcmp ord x,x" is "fcmp ord x, 0". 1045 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) && 1046 isa<ConstantAggregateZero>(RHS->getOperand(1))) 1047 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0)); 1048 return nullptr; 1049 } 1050 1051 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1); 1052 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1); 1053 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate(); 1054 1055 1056 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) { 1057 // Swap RHS operands to match LHS. 1058 Op1CC = FCmpInst::getSwappedPredicate(Op1CC); 1059 std::swap(Op1LHS, Op1RHS); 1060 } 1061 1062 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) { 1063 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y). 1064 if (Op0CC == Op1CC) 1065 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS); 1066 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE) 1067 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 1068 if (Op0CC == FCmpInst::FCMP_TRUE) 1069 return RHS; 1070 if (Op1CC == FCmpInst::FCMP_TRUE) 1071 return LHS; 1072 1073 bool Op0Ordered; 1074 bool Op1Ordered; 1075 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered); 1076 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered); 1077 // uno && ord -> false 1078 if (Op0Pred == 0 && Op1Pred == 0 && Op0Ordered != Op1Ordered) 1079 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 1080 if (Op1Pred == 0) { 1081 std::swap(LHS, RHS); 1082 std::swap(Op0Pred, Op1Pred); 1083 std::swap(Op0Ordered, Op1Ordered); 1084 } 1085 if (Op0Pred == 0) { 1086 // uno && ueq -> uno && (uno || eq) -> uno 1087 // ord && olt -> ord && (ord && lt) -> olt 1088 if (!Op0Ordered && (Op0Ordered == Op1Ordered)) 1089 return LHS; 1090 if (Op0Ordered && (Op0Ordered == Op1Ordered)) 1091 return RHS; 1092 1093 // uno && oeq -> uno && (ord && eq) -> false 1094 if (!Op0Ordered) 1095 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0); 1096 // ord && ueq -> ord && (uno || eq) -> oeq 1097 return getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS, Builder); 1098 } 1099 } 1100 1101 return nullptr; 1102 } 1103 1104 1105 Instruction *InstCombiner::visitAnd(BinaryOperator &I) { 1106 bool Changed = SimplifyAssociativeOrCommutative(I); 1107 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1108 1109 if (Value *V = SimplifyVectorOp(I)) 1110 return ReplaceInstUsesWith(I, V); 1111 1112 if (Value *V = SimplifyAndInst(Op0, Op1, DL)) 1113 return ReplaceInstUsesWith(I, V); 1114 1115 // (A|B)&(A|C) -> A|(B&C) etc 1116 if (Value *V = SimplifyUsingDistributiveLaws(I)) 1117 return ReplaceInstUsesWith(I, V); 1118 1119 // See if we can simplify any instructions used by the instruction whose sole 1120 // purpose is to compute bits we don't care about. 1121 if (SimplifyDemandedInstructionBits(I)) 1122 return &I; 1123 1124 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) { 1125 const APInt &AndRHSMask = AndRHS->getValue(); 1126 1127 // Optimize a variety of ((val OP C1) & C2) combinations... 1128 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) { 1129 Value *Op0LHS = Op0I->getOperand(0); 1130 Value *Op0RHS = Op0I->getOperand(1); 1131 switch (Op0I->getOpcode()) { 1132 default: break; 1133 case Instruction::Xor: 1134 case Instruction::Or: { 1135 // If the mask is only needed on one incoming arm, push it up. 1136 if (!Op0I->hasOneUse()) break; 1137 1138 APInt NotAndRHS(~AndRHSMask); 1139 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) { 1140 // Not masking anything out for the LHS, move to RHS. 1141 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS, 1142 Op0RHS->getName()+".masked"); 1143 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS); 1144 } 1145 if (!isa<Constant>(Op0RHS) && 1146 MaskedValueIsZero(Op0RHS, NotAndRHS)) { 1147 // Not masking anything out for the RHS, move to LHS. 1148 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS, 1149 Op0LHS->getName()+".masked"); 1150 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS); 1151 } 1152 1153 break; 1154 } 1155 case Instruction::Add: 1156 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS. 1157 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0 1158 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0 1159 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I)) 1160 return BinaryOperator::CreateAnd(V, AndRHS); 1161 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I)) 1162 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes 1163 break; 1164 1165 case Instruction::Sub: 1166 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS. 1167 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0 1168 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0 1169 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I)) 1170 return BinaryOperator::CreateAnd(V, AndRHS); 1171 1172 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS 1173 // has 1's for all bits that the subtraction with A might affect. 1174 if (Op0I->hasOneUse() && !match(Op0LHS, m_Zero())) { 1175 uint32_t BitWidth = AndRHSMask.getBitWidth(); 1176 uint32_t Zeros = AndRHSMask.countLeadingZeros(); 1177 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros); 1178 1179 if (MaskedValueIsZero(Op0LHS, Mask)) { 1180 Value *NewNeg = Builder->CreateNeg(Op0RHS); 1181 return BinaryOperator::CreateAnd(NewNeg, AndRHS); 1182 } 1183 } 1184 break; 1185 1186 case Instruction::Shl: 1187 case Instruction::LShr: 1188 // (1 << x) & 1 --> zext(x == 0) 1189 // (1 >> x) & 1 --> zext(x == 0) 1190 if (AndRHSMask == 1 && Op0LHS == AndRHS) { 1191 Value *NewICmp = 1192 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType())); 1193 return new ZExtInst(NewICmp, I.getType()); 1194 } 1195 break; 1196 } 1197 1198 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) 1199 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I)) 1200 return Res; 1201 } 1202 1203 // If this is an integer truncation, and if the source is an 'and' with 1204 // immediate, transform it. This frequently occurs for bitfield accesses. 1205 { 1206 Value *X = nullptr; ConstantInt *YC = nullptr; 1207 if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) { 1208 // Change: and (trunc (and X, YC) to T), C2 1209 // into : and (trunc X to T), trunc(YC) & C2 1210 // This will fold the two constants together, which may allow 1211 // other simplifications. 1212 Value *NewCast = Builder->CreateTrunc(X, I.getType(), "and.shrunk"); 1213 Constant *C3 = ConstantExpr::getTrunc(YC, I.getType()); 1214 C3 = ConstantExpr::getAnd(C3, AndRHS); 1215 return BinaryOperator::CreateAnd(NewCast, C3); 1216 } 1217 } 1218 1219 // Try to fold constant and into select arguments. 1220 if (SelectInst *SI = dyn_cast<SelectInst>(Op0)) 1221 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1222 return R; 1223 if (isa<PHINode>(Op0)) 1224 if (Instruction *NV = FoldOpIntoPhi(I)) 1225 return NV; 1226 } 1227 1228 1229 // (~A & ~B) == (~(A | B)) - De Morgan's Law 1230 if (Value *Op0NotVal = dyn_castNotVal(Op0)) 1231 if (Value *Op1NotVal = dyn_castNotVal(Op1)) 1232 if (Op0->hasOneUse() && Op1->hasOneUse()) { 1233 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal, 1234 I.getName()+".demorgan"); 1235 return BinaryOperator::CreateNot(Or); 1236 } 1237 1238 { 1239 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr; 1240 // (A|B) & ~(A&B) -> A^B 1241 if (match(Op0, m_Or(m_Value(A), m_Value(B))) && 1242 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) && 1243 ((A == C && B == D) || (A == D && B == C))) 1244 return BinaryOperator::CreateXor(A, B); 1245 1246 // ~(A&B) & (A|B) -> A^B 1247 if (match(Op1, m_Or(m_Value(A), m_Value(B))) && 1248 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) && 1249 ((A == C && B == D) || (A == D && B == C))) 1250 return BinaryOperator::CreateXor(A, B); 1251 1252 // A&(A^B) => A & ~B 1253 { 1254 Value *tmpOp0 = Op0; 1255 Value *tmpOp1 = Op1; 1256 if (Op0->hasOneUse() && 1257 match(Op0, m_Xor(m_Value(A), m_Value(B)))) { 1258 if (A == Op1 || B == Op1 ) { 1259 tmpOp1 = Op0; 1260 tmpOp0 = Op1; 1261 // Simplify below 1262 } 1263 } 1264 1265 if (tmpOp1->hasOneUse() && 1266 match(tmpOp1, m_Xor(m_Value(A), m_Value(B)))) { 1267 if (B == tmpOp0) { 1268 std::swap(A, B); 1269 } 1270 // Notice that the patten (A&(~B)) is actually (A&(-1^B)), so if 1271 // A is originally -1 (or a vector of -1 and undefs), then we enter 1272 // an endless loop. By checking that A is non-constant we ensure that 1273 // we will never get to the loop. 1274 if (A == tmpOp0 && !isa<Constant>(A)) // A&(A^B) -> A & ~B 1275 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B)); 1276 } 1277 } 1278 1279 // (A&((~A)|B)) -> A&B 1280 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) || 1281 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1))))) 1282 return BinaryOperator::CreateAnd(A, Op1); 1283 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) || 1284 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0))))) 1285 return BinaryOperator::CreateAnd(A, Op0); 1286 } 1287 1288 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) 1289 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0)) 1290 if (Value *Res = FoldAndOfICmps(LHS, RHS)) 1291 return ReplaceInstUsesWith(I, Res); 1292 1293 // If and'ing two fcmp, try combine them into one. 1294 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) 1295 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) 1296 if (Value *Res = FoldAndOfFCmps(LHS, RHS)) 1297 return ReplaceInstUsesWith(I, Res); 1298 1299 1300 // fold (and (cast A), (cast B)) -> (cast (and A, B)) 1301 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) 1302 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) { 1303 Type *SrcTy = Op0C->getOperand(0)->getType(); 1304 if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ? 1305 SrcTy == Op1C->getOperand(0)->getType() && 1306 SrcTy->isIntOrIntVectorTy()) { 1307 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0); 1308 1309 // Only do this if the casts both really cause code to be generated. 1310 if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) && 1311 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) { 1312 Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName()); 1313 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType()); 1314 } 1315 1316 // If this is and(cast(icmp), cast(icmp)), try to fold this even if the 1317 // cast is otherwise not optimizable. This happens for vector sexts. 1318 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp)) 1319 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp)) 1320 if (Value *Res = FoldAndOfICmps(LHS, RHS)) 1321 return CastInst::Create(Op0C->getOpcode(), Res, I.getType()); 1322 1323 // If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the 1324 // cast is otherwise not optimizable. This happens for vector sexts. 1325 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp)) 1326 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp)) 1327 if (Value *Res = FoldAndOfFCmps(LHS, RHS)) 1328 return CastInst::Create(Op0C->getOpcode(), Res, I.getType()); 1329 } 1330 } 1331 1332 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts. 1333 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) { 1334 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0)) 1335 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() && 1336 SI0->getOperand(1) == SI1->getOperand(1) && 1337 (SI0->hasOneUse() || SI1->hasOneUse())) { 1338 Value *NewOp = 1339 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0), 1340 SI0->getName()); 1341 return BinaryOperator::Create(SI1->getOpcode(), NewOp, 1342 SI1->getOperand(1)); 1343 } 1344 } 1345 1346 { 1347 Value *X = nullptr; 1348 bool OpsSwapped = false; 1349 // Canonicalize SExt or Not to the LHS 1350 if (match(Op1, m_SExt(m_Value())) || 1351 match(Op1, m_Not(m_Value()))) { 1352 std::swap(Op0, Op1); 1353 OpsSwapped = true; 1354 } 1355 1356 // Fold (and (sext bool to A), B) --> (select bool, B, 0) 1357 if (match(Op0, m_SExt(m_Value(X))) && 1358 X->getType()->getScalarType()->isIntegerTy(1)) { 1359 Value *Zero = Constant::getNullValue(Op1->getType()); 1360 return SelectInst::Create(X, Op1, Zero); 1361 } 1362 1363 // Fold (and ~(sext bool to A), B) --> (select bool, 0, B) 1364 if (match(Op0, m_Not(m_SExt(m_Value(X)))) && 1365 X->getType()->getScalarType()->isIntegerTy(1)) { 1366 Value *Zero = Constant::getNullValue(Op0->getType()); 1367 return SelectInst::Create(X, Zero, Op1); 1368 } 1369 1370 if (OpsSwapped) 1371 std::swap(Op0, Op1); 1372 } 1373 1374 return Changed ? &I : nullptr; 1375 } 1376 1377 /// CollectBSwapParts - Analyze the specified subexpression and see if it is 1378 /// capable of providing pieces of a bswap. The subexpression provides pieces 1379 /// of a bswap if it is proven that each of the non-zero bytes in the output of 1380 /// the expression came from the corresponding "byte swapped" byte in some other 1381 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then 1382 /// we know that the expression deposits the low byte of %X into the high byte 1383 /// of the bswap result and that all other bytes are zero. This expression is 1384 /// accepted, the high byte of ByteValues is set to X to indicate a correct 1385 /// match. 1386 /// 1387 /// This function returns true if the match was unsuccessful and false if so. 1388 /// On entry to the function the "OverallLeftShift" is a signed integer value 1389 /// indicating the number of bytes that the subexpression is later shifted. For 1390 /// example, if the expression is later right shifted by 16 bits, the 1391 /// OverallLeftShift value would be -2 on entry. This is used to specify which 1392 /// byte of ByteValues is actually being set. 1393 /// 1394 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding 1395 /// byte is masked to zero by a user. For example, in (X & 255), X will be 1396 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits 1397 /// this function to working on up to 32-byte (256 bit) values. ByteMask is 1398 /// always in the local (OverallLeftShift) coordinate space. 1399 /// 1400 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask, 1401 SmallVectorImpl<Value *> &ByteValues) { 1402 if (Instruction *I = dyn_cast<Instruction>(V)) { 1403 // If this is an or instruction, it may be an inner node of the bswap. 1404 if (I->getOpcode() == Instruction::Or) { 1405 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask, 1406 ByteValues) || 1407 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask, 1408 ByteValues); 1409 } 1410 1411 // If this is a logical shift by a constant multiple of 8, recurse with 1412 // OverallLeftShift and ByteMask adjusted. 1413 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) { 1414 unsigned ShAmt = 1415 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U); 1416 // Ensure the shift amount is defined and of a byte value. 1417 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size())) 1418 return true; 1419 1420 unsigned ByteShift = ShAmt >> 3; 1421 if (I->getOpcode() == Instruction::Shl) { 1422 // X << 2 -> collect(X, +2) 1423 OverallLeftShift += ByteShift; 1424 ByteMask >>= ByteShift; 1425 } else { 1426 // X >>u 2 -> collect(X, -2) 1427 OverallLeftShift -= ByteShift; 1428 ByteMask <<= ByteShift; 1429 ByteMask &= (~0U >> (32-ByteValues.size())); 1430 } 1431 1432 if (OverallLeftShift >= (int)ByteValues.size()) return true; 1433 if (OverallLeftShift <= -(int)ByteValues.size()) return true; 1434 1435 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask, 1436 ByteValues); 1437 } 1438 1439 // If this is a logical 'and' with a mask that clears bytes, clear the 1440 // corresponding bytes in ByteMask. 1441 if (I->getOpcode() == Instruction::And && 1442 isa<ConstantInt>(I->getOperand(1))) { 1443 // Scan every byte of the and mask, seeing if the byte is either 0 or 255. 1444 unsigned NumBytes = ByteValues.size(); 1445 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255); 1446 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue(); 1447 1448 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) { 1449 // If this byte is masked out by a later operation, we don't care what 1450 // the and mask is. 1451 if ((ByteMask & (1 << i)) == 0) 1452 continue; 1453 1454 // If the AndMask is all zeros for this byte, clear the bit. 1455 APInt MaskB = AndMask & Byte; 1456 if (MaskB == 0) { 1457 ByteMask &= ~(1U << i); 1458 continue; 1459 } 1460 1461 // If the AndMask is not all ones for this byte, it's not a bytezap. 1462 if (MaskB != Byte) 1463 return true; 1464 1465 // Otherwise, this byte is kept. 1466 } 1467 1468 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask, 1469 ByteValues); 1470 } 1471 } 1472 1473 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be 1474 // the input value to the bswap. Some observations: 1) if more than one byte 1475 // is demanded from this input, then it could not be successfully assembled 1476 // into a byteswap. At least one of the two bytes would not be aligned with 1477 // their ultimate destination. 1478 if (!isPowerOf2_32(ByteMask)) return true; 1479 unsigned InputByteNo = countTrailingZeros(ByteMask); 1480 1481 // 2) The input and ultimate destinations must line up: if byte 3 of an i32 1482 // is demanded, it needs to go into byte 0 of the result. This means that the 1483 // byte needs to be shifted until it lands in the right byte bucket. The 1484 // shift amount depends on the position: if the byte is coming from the high 1485 // part of the value (e.g. byte 3) then it must be shifted right. If from the 1486 // low part, it must be shifted left. 1487 unsigned DestByteNo = InputByteNo + OverallLeftShift; 1488 if (ByteValues.size()-1-DestByteNo != InputByteNo) 1489 return true; 1490 1491 // If the destination byte value is already defined, the values are or'd 1492 // together, which isn't a bswap (unless it's an or of the same bits). 1493 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V) 1494 return true; 1495 ByteValues[DestByteNo] = V; 1496 return false; 1497 } 1498 1499 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom. 1500 /// If so, insert the new bswap intrinsic and return it. 1501 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) { 1502 IntegerType *ITy = dyn_cast<IntegerType>(I.getType()); 1503 if (!ITy || ITy->getBitWidth() % 16 || 1504 // ByteMask only allows up to 32-byte values. 1505 ITy->getBitWidth() > 32*8) 1506 return nullptr; // Can only bswap pairs of bytes. Can't do vectors. 1507 1508 /// ByteValues - For each byte of the result, we keep track of which value 1509 /// defines each byte. 1510 SmallVector<Value*, 8> ByteValues; 1511 ByteValues.resize(ITy->getBitWidth()/8); 1512 1513 // Try to find all the pieces corresponding to the bswap. 1514 uint32_t ByteMask = ~0U >> (32-ByteValues.size()); 1515 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues)) 1516 return nullptr; 1517 1518 // Check to see if all of the bytes come from the same value. 1519 Value *V = ByteValues[0]; 1520 if (!V) return nullptr; // Didn't find a byte? Must be zero. 1521 1522 // Check to make sure that all of the bytes come from the same value. 1523 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i) 1524 if (ByteValues[i] != V) 1525 return nullptr; 1526 Module *M = I.getParent()->getParent()->getParent(); 1527 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, ITy); 1528 return CallInst::Create(F, V); 1529 } 1530 1531 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check 1532 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then 1533 /// we can simplify this expression to "cond ? C : D or B". 1534 static Instruction *MatchSelectFromAndOr(Value *A, Value *B, 1535 Value *C, Value *D) { 1536 // If A is not a select of -1/0, this cannot match. 1537 Value *Cond = nullptr; 1538 if (!match(A, m_SExt(m_Value(Cond))) || 1539 !Cond->getType()->isIntegerTy(1)) 1540 return nullptr; 1541 1542 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B. 1543 if (match(D, m_Not(m_SExt(m_Specific(Cond))))) 1544 return SelectInst::Create(Cond, C, B); 1545 if (match(D, m_SExt(m_Not(m_Specific(Cond))))) 1546 return SelectInst::Create(Cond, C, B); 1547 1548 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D. 1549 if (match(B, m_Not(m_SExt(m_Specific(Cond))))) 1550 return SelectInst::Create(Cond, C, D); 1551 if (match(B, m_SExt(m_Not(m_Specific(Cond))))) 1552 return SelectInst::Create(Cond, C, D); 1553 return nullptr; 1554 } 1555 1556 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible. 1557 Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS) { 1558 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate(); 1559 1560 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2) 1561 // if K1 and K2 are a one-bit mask. 1562 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1)); 1563 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1)); 1564 1565 if (LHS->getPredicate() == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero() && 1566 RHS->getPredicate() == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) { 1567 1568 BinaryOperator *LAnd = dyn_cast<BinaryOperator>(LHS->getOperand(0)); 1569 BinaryOperator *RAnd = dyn_cast<BinaryOperator>(RHS->getOperand(0)); 1570 if (LAnd && RAnd && LAnd->hasOneUse() && RHS->hasOneUse() && 1571 LAnd->getOpcode() == Instruction::And && 1572 RAnd->getOpcode() == Instruction::And) { 1573 1574 Value *Mask = nullptr; 1575 Value *Masked = nullptr; 1576 if (LAnd->getOperand(0) == RAnd->getOperand(0) && 1577 isKnownToBeAPowerOfTwo(LAnd->getOperand(1)) && 1578 isKnownToBeAPowerOfTwo(RAnd->getOperand(1))) { 1579 Mask = Builder->CreateOr(LAnd->getOperand(1), RAnd->getOperand(1)); 1580 Masked = Builder->CreateAnd(LAnd->getOperand(0), Mask); 1581 } else if (LAnd->getOperand(1) == RAnd->getOperand(1) && 1582 isKnownToBeAPowerOfTwo(LAnd->getOperand(0)) && 1583 isKnownToBeAPowerOfTwo(RAnd->getOperand(0))) { 1584 Mask = Builder->CreateOr(LAnd->getOperand(0), RAnd->getOperand(0)); 1585 Masked = Builder->CreateAnd(LAnd->getOperand(1), Mask); 1586 } 1587 1588 if (Masked) 1589 return Builder->CreateICmp(ICmpInst::ICMP_NE, Masked, Mask); 1590 } 1591 } 1592 1593 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B) 1594 if (PredicatesFoldable(LHSCC, RHSCC)) { 1595 if (LHS->getOperand(0) == RHS->getOperand(1) && 1596 LHS->getOperand(1) == RHS->getOperand(0)) 1597 LHS->swapOperands(); 1598 if (LHS->getOperand(0) == RHS->getOperand(0) && 1599 LHS->getOperand(1) == RHS->getOperand(1)) { 1600 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1); 1601 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS); 1602 bool isSigned = LHS->isSigned() || RHS->isSigned(); 1603 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder); 1604 } 1605 } 1606 1607 // handle (roughly): 1608 // (icmp ne (A & B), C) | (icmp ne (A & D), E) 1609 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, false, Builder)) 1610 return V; 1611 1612 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0); 1613 if (LHS->hasOneUse() || RHS->hasOneUse()) { 1614 // (icmp eq B, 0) | (icmp ult A, B) -> (icmp ule A, B-1) 1615 // (icmp eq B, 0) | (icmp ugt B, A) -> (icmp ule A, B-1) 1616 Value *A = nullptr, *B = nullptr; 1617 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero()) { 1618 B = Val; 1619 if (RHSCC == ICmpInst::ICMP_ULT && Val == RHS->getOperand(1)) 1620 A = Val2; 1621 else if (RHSCC == ICmpInst::ICMP_UGT && Val == Val2) 1622 A = RHS->getOperand(1); 1623 } 1624 // (icmp ult A, B) | (icmp eq B, 0) -> (icmp ule A, B-1) 1625 // (icmp ugt B, A) | (icmp eq B, 0) -> (icmp ule A, B-1) 1626 else if (RHSCC == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) { 1627 B = Val2; 1628 if (LHSCC == ICmpInst::ICMP_ULT && Val2 == LHS->getOperand(1)) 1629 A = Val; 1630 else if (LHSCC == ICmpInst::ICMP_UGT && Val2 == Val) 1631 A = LHS->getOperand(1); 1632 } 1633 if (A && B) 1634 return Builder->CreateICmp( 1635 ICmpInst::ICMP_UGE, 1636 Builder->CreateAdd(B, ConstantInt::getSigned(B->getType(), -1)), A); 1637 } 1638 1639 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2). 1640 if (!LHSCst || !RHSCst) return nullptr; 1641 1642 if (LHSCst == RHSCst && LHSCC == RHSCC) { 1643 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0) 1644 if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) { 1645 Value *NewOr = Builder->CreateOr(Val, Val2); 1646 return Builder->CreateICmp(LHSCC, NewOr, LHSCst); 1647 } 1648 } 1649 1650 // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1) 1651 // iff C2 + CA == C1. 1652 if (LHSCC == ICmpInst::ICMP_ULT && RHSCC == ICmpInst::ICMP_EQ) { 1653 ConstantInt *AddCst; 1654 if (match(Val, m_Add(m_Specific(Val2), m_ConstantInt(AddCst)))) 1655 if (RHSCst->getValue() + AddCst->getValue() == LHSCst->getValue()) 1656 return Builder->CreateICmpULE(Val, LHSCst); 1657 } 1658 1659 // From here on, we only handle: 1660 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler. 1661 if (Val != Val2) return nullptr; 1662 1663 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere. 1664 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE || 1665 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE || 1666 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE || 1667 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE) 1668 return nullptr; 1669 1670 // We can't fold (ugt x, C) | (sgt x, C2). 1671 if (!PredicatesFoldable(LHSCC, RHSCC)) 1672 return nullptr; 1673 1674 // Ensure that the larger constant is on the RHS. 1675 bool ShouldSwap; 1676 if (CmpInst::isSigned(LHSCC) || 1677 (ICmpInst::isEquality(LHSCC) && 1678 CmpInst::isSigned(RHSCC))) 1679 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue()); 1680 else 1681 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue()); 1682 1683 if (ShouldSwap) { 1684 std::swap(LHS, RHS); 1685 std::swap(LHSCst, RHSCst); 1686 std::swap(LHSCC, RHSCC); 1687 } 1688 1689 // At this point, we know we have two icmp instructions 1690 // comparing a value against two constants and or'ing the result 1691 // together. Because of the above check, we know that we only have 1692 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the 1693 // icmp folding check above), that the two constants are not 1694 // equal. 1695 assert(LHSCst != RHSCst && "Compares not folded above?"); 1696 1697 switch (LHSCC) { 1698 default: llvm_unreachable("Unknown integer condition code!"); 1699 case ICmpInst::ICMP_EQ: 1700 switch (RHSCC) { 1701 default: llvm_unreachable("Unknown integer condition code!"); 1702 case ICmpInst::ICMP_EQ: 1703 if (LHS->getOperand(0) == RHS->getOperand(0)) { 1704 // if LHSCst and RHSCst differ only by one bit: 1705 // (A == C1 || A == C2) -> (A & ~(C1 ^ C2)) == C1 1706 assert(LHSCst->getValue().ule(LHSCst->getValue())); 1707 1708 APInt Xor = LHSCst->getValue() ^ RHSCst->getValue(); 1709 if (Xor.isPowerOf2()) { 1710 Value *NegCst = Builder->getInt(~Xor); 1711 Value *And = Builder->CreateAnd(LHS->getOperand(0), NegCst); 1712 return Builder->CreateICmp(ICmpInst::ICMP_EQ, And, LHSCst); 1713 } 1714 } 1715 1716 if (LHSCst == SubOne(RHSCst)) { 1717 // (X == 13 | X == 14) -> X-13 <u 2 1718 Constant *AddCST = ConstantExpr::getNeg(LHSCst); 1719 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off"); 1720 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst); 1721 return Builder->CreateICmpULT(Add, AddCST); 1722 } 1723 1724 break; // (X == 13 | X == 15) -> no change 1725 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change 1726 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change 1727 break; 1728 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15 1729 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15 1730 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15 1731 return RHS; 1732 } 1733 break; 1734 case ICmpInst::ICMP_NE: 1735 switch (RHSCC) { 1736 default: llvm_unreachable("Unknown integer condition code!"); 1737 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13 1738 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13 1739 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13 1740 return LHS; 1741 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true 1742 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true 1743 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true 1744 return Builder->getTrue(); 1745 } 1746 case ICmpInst::ICMP_ULT: 1747 switch (RHSCC) { 1748 default: llvm_unreachable("Unknown integer condition code!"); 1749 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change 1750 break; 1751 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2 1752 // If RHSCst is [us]MAXINT, it is always false. Not handling 1753 // this can cause overflow. 1754 if (RHSCst->isMaxValue(false)) 1755 return LHS; 1756 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false); 1757 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change 1758 break; 1759 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15 1760 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15 1761 return RHS; 1762 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change 1763 break; 1764 } 1765 break; 1766 case ICmpInst::ICMP_SLT: 1767 switch (RHSCC) { 1768 default: llvm_unreachable("Unknown integer condition code!"); 1769 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change 1770 break; 1771 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2 1772 // If RHSCst is [us]MAXINT, it is always false. Not handling 1773 // this can cause overflow. 1774 if (RHSCst->isMaxValue(true)) 1775 return LHS; 1776 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false); 1777 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change 1778 break; 1779 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15 1780 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15 1781 return RHS; 1782 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change 1783 break; 1784 } 1785 break; 1786 case ICmpInst::ICMP_UGT: 1787 switch (RHSCC) { 1788 default: llvm_unreachable("Unknown integer condition code!"); 1789 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13 1790 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13 1791 return LHS; 1792 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change 1793 break; 1794 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true 1795 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true 1796 return Builder->getTrue(); 1797 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change 1798 break; 1799 } 1800 break; 1801 case ICmpInst::ICMP_SGT: 1802 switch (RHSCC) { 1803 default: llvm_unreachable("Unknown integer condition code!"); 1804 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13 1805 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13 1806 return LHS; 1807 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change 1808 break; 1809 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true 1810 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true 1811 return Builder->getTrue(); 1812 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change 1813 break; 1814 } 1815 break; 1816 } 1817 return nullptr; 1818 } 1819 1820 /// FoldOrOfFCmps - Optimize (fcmp)|(fcmp). NOTE: Unlike the rest of 1821 /// instcombine, this returns a Value which should already be inserted into the 1822 /// function. 1823 Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) { 1824 if (LHS->getPredicate() == FCmpInst::FCMP_UNO && 1825 RHS->getPredicate() == FCmpInst::FCMP_UNO && 1826 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) { 1827 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1))) 1828 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) { 1829 // If either of the constants are nans, then the whole thing returns 1830 // true. 1831 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN()) 1832 return Builder->getTrue(); 1833 1834 // Otherwise, no need to compare the two constants, compare the 1835 // rest. 1836 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0)); 1837 } 1838 1839 // Handle vector zeros. This occurs because the canonical form of 1840 // "fcmp uno x,x" is "fcmp uno x, 0". 1841 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) && 1842 isa<ConstantAggregateZero>(RHS->getOperand(1))) 1843 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0)); 1844 1845 return nullptr; 1846 } 1847 1848 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1); 1849 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1); 1850 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate(); 1851 1852 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) { 1853 // Swap RHS operands to match LHS. 1854 Op1CC = FCmpInst::getSwappedPredicate(Op1CC); 1855 std::swap(Op1LHS, Op1RHS); 1856 } 1857 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) { 1858 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y). 1859 if (Op0CC == Op1CC) 1860 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS); 1861 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE) 1862 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1); 1863 if (Op0CC == FCmpInst::FCMP_FALSE) 1864 return RHS; 1865 if (Op1CC == FCmpInst::FCMP_FALSE) 1866 return LHS; 1867 bool Op0Ordered; 1868 bool Op1Ordered; 1869 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered); 1870 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered); 1871 if (Op0Ordered == Op1Ordered) { 1872 // If both are ordered or unordered, return a new fcmp with 1873 // or'ed predicates. 1874 return getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS, Builder); 1875 } 1876 } 1877 return nullptr; 1878 } 1879 1880 /// FoldOrWithConstants - This helper function folds: 1881 /// 1882 /// ((A | B) & C1) | (B & C2) 1883 /// 1884 /// into: 1885 /// 1886 /// (A & C1) | B 1887 /// 1888 /// when the XOR of the two constants is "all ones" (-1). 1889 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op, 1890 Value *A, Value *B, Value *C) { 1891 ConstantInt *CI1 = dyn_cast<ConstantInt>(C); 1892 if (!CI1) return nullptr; 1893 1894 Value *V1 = nullptr; 1895 ConstantInt *CI2 = nullptr; 1896 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return nullptr; 1897 1898 APInt Xor = CI1->getValue() ^ CI2->getValue(); 1899 if (!Xor.isAllOnesValue()) return nullptr; 1900 1901 if (V1 == A || V1 == B) { 1902 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1); 1903 return BinaryOperator::CreateOr(NewOp, V1); 1904 } 1905 1906 return nullptr; 1907 } 1908 1909 Instruction *InstCombiner::visitOr(BinaryOperator &I) { 1910 bool Changed = SimplifyAssociativeOrCommutative(I); 1911 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 1912 1913 if (Value *V = SimplifyVectorOp(I)) 1914 return ReplaceInstUsesWith(I, V); 1915 1916 if (Value *V = SimplifyOrInst(Op0, Op1, DL)) 1917 return ReplaceInstUsesWith(I, V); 1918 1919 // (A&B)|(A&C) -> A&(B|C) etc 1920 if (Value *V = SimplifyUsingDistributiveLaws(I)) 1921 return ReplaceInstUsesWith(I, V); 1922 1923 // See if we can simplify any instructions used by the instruction whose sole 1924 // purpose is to compute bits we don't care about. 1925 if (SimplifyDemandedInstructionBits(I)) 1926 return &I; 1927 1928 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) { 1929 ConstantInt *C1 = nullptr; Value *X = nullptr; 1930 // (X & C1) | C2 --> (X | C2) & (C1|C2) 1931 // iff (C1 & C2) == 0. 1932 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && 1933 (RHS->getValue() & C1->getValue()) != 0 && 1934 Op0->hasOneUse()) { 1935 Value *Or = Builder->CreateOr(X, RHS); 1936 Or->takeName(Op0); 1937 return BinaryOperator::CreateAnd(Or, 1938 Builder->getInt(RHS->getValue() | C1->getValue())); 1939 } 1940 1941 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2) 1942 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && 1943 Op0->hasOneUse()) { 1944 Value *Or = Builder->CreateOr(X, RHS); 1945 Or->takeName(Op0); 1946 return BinaryOperator::CreateXor(Or, 1947 Builder->getInt(C1->getValue() & ~RHS->getValue())); 1948 } 1949 1950 // Try to fold constant and into select arguments. 1951 if (SelectInst *SI = dyn_cast<SelectInst>(Op0)) 1952 if (Instruction *R = FoldOpIntoSelect(I, SI)) 1953 return R; 1954 1955 if (isa<PHINode>(Op0)) 1956 if (Instruction *NV = FoldOpIntoPhi(I)) 1957 return NV; 1958 } 1959 1960 Value *A = nullptr, *B = nullptr; 1961 ConstantInt *C1 = nullptr, *C2 = nullptr; 1962 1963 // (A | B) | C and A | (B | C) -> bswap if possible. 1964 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible. 1965 if (match(Op0, m_Or(m_Value(), m_Value())) || 1966 match(Op1, m_Or(m_Value(), m_Value())) || 1967 (match(Op0, m_LogicalShift(m_Value(), m_Value())) && 1968 match(Op1, m_LogicalShift(m_Value(), m_Value())))) { 1969 if (Instruction *BSwap = MatchBSwap(I)) 1970 return BSwap; 1971 } 1972 1973 // (X^C)|Y -> (X|Y)^C iff Y&C == 0 1974 if (Op0->hasOneUse() && 1975 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) && 1976 MaskedValueIsZero(Op1, C1->getValue())) { 1977 Value *NOr = Builder->CreateOr(A, Op1); 1978 NOr->takeName(Op0); 1979 return BinaryOperator::CreateXor(NOr, C1); 1980 } 1981 1982 // Y|(X^C) -> (X|Y)^C iff Y&C == 0 1983 if (Op1->hasOneUse() && 1984 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) && 1985 MaskedValueIsZero(Op0, C1->getValue())) { 1986 Value *NOr = Builder->CreateOr(A, Op0); 1987 NOr->takeName(Op0); 1988 return BinaryOperator::CreateXor(NOr, C1); 1989 } 1990 1991 // (A & C)|(B & D) 1992 Value *C = nullptr, *D = nullptr; 1993 if (match(Op0, m_And(m_Value(A), m_Value(C))) && 1994 match(Op1, m_And(m_Value(B), m_Value(D)))) { 1995 Value *V1 = nullptr, *V2 = nullptr; 1996 C1 = dyn_cast<ConstantInt>(C); 1997 C2 = dyn_cast<ConstantInt>(D); 1998 if (C1 && C2) { // (A & C1)|(B & C2) 1999 if ((C1->getValue() & C2->getValue()) == 0) { 2000 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2) 2001 // iff (C1&C2) == 0 and (N&~C1) == 0 2002 if (match(A, m_Or(m_Value(V1), m_Value(V2))) && 2003 ((V1 == B && MaskedValueIsZero(V2, ~C1->getValue())) || // (V|N) 2004 (V2 == B && MaskedValueIsZero(V1, ~C1->getValue())))) // (N|V) 2005 return BinaryOperator::CreateAnd(A, 2006 Builder->getInt(C1->getValue()|C2->getValue())); 2007 // Or commutes, try both ways. 2008 if (match(B, m_Or(m_Value(V1), m_Value(V2))) && 2009 ((V1 == A && MaskedValueIsZero(V2, ~C2->getValue())) || // (V|N) 2010 (V2 == A && MaskedValueIsZero(V1, ~C2->getValue())))) // (N|V) 2011 return BinaryOperator::CreateAnd(B, 2012 Builder->getInt(C1->getValue()|C2->getValue())); 2013 2014 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2) 2015 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0. 2016 ConstantInt *C3 = nullptr, *C4 = nullptr; 2017 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) && 2018 (C3->getValue() & ~C1->getValue()) == 0 && 2019 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) && 2020 (C4->getValue() & ~C2->getValue()) == 0) { 2021 V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield"); 2022 return BinaryOperator::CreateAnd(V2, 2023 Builder->getInt(C1->getValue()|C2->getValue())); 2024 } 2025 } 2026 } 2027 2028 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants. 2029 // Don't do this for vector select idioms, the code generator doesn't handle 2030 // them well yet. 2031 if (!I.getType()->isVectorTy()) { 2032 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D)) 2033 return Match; 2034 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C)) 2035 return Match; 2036 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D)) 2037 return Match; 2038 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C)) 2039 return Match; 2040 } 2041 2042 // ((A&~B)|(~A&B)) -> A^B 2043 if ((match(C, m_Not(m_Specific(D))) && 2044 match(B, m_Not(m_Specific(A))))) 2045 return BinaryOperator::CreateXor(A, D); 2046 // ((~B&A)|(~A&B)) -> A^B 2047 if ((match(A, m_Not(m_Specific(D))) && 2048 match(B, m_Not(m_Specific(C))))) 2049 return BinaryOperator::CreateXor(C, D); 2050 // ((A&~B)|(B&~A)) -> A^B 2051 if ((match(C, m_Not(m_Specific(B))) && 2052 match(D, m_Not(m_Specific(A))))) 2053 return BinaryOperator::CreateXor(A, B); 2054 // ((~B&A)|(B&~A)) -> A^B 2055 if ((match(A, m_Not(m_Specific(B))) && 2056 match(D, m_Not(m_Specific(C))))) 2057 return BinaryOperator::CreateXor(C, B); 2058 2059 // ((A|B)&1)|(B&-2) -> (A&1) | B 2060 if (match(A, m_Or(m_Value(V1), m_Specific(B))) || 2061 match(A, m_Or(m_Specific(B), m_Value(V1)))) { 2062 Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C); 2063 if (Ret) return Ret; 2064 } 2065 // (B&-2)|((A|B)&1) -> (A&1) | B 2066 if (match(B, m_Or(m_Specific(A), m_Value(V1))) || 2067 match(B, m_Or(m_Value(V1), m_Specific(A)))) { 2068 Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D); 2069 if (Ret) return Ret; 2070 } 2071 } 2072 2073 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts. 2074 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) { 2075 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0)) 2076 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() && 2077 SI0->getOperand(1) == SI1->getOperand(1) && 2078 (SI0->hasOneUse() || SI1->hasOneUse())) { 2079 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0), 2080 SI0->getName()); 2081 return BinaryOperator::Create(SI1->getOpcode(), NewOp, 2082 SI1->getOperand(1)); 2083 } 2084 } 2085 2086 // (~A | ~B) == (~(A & B)) - De Morgan's Law 2087 if (Value *Op0NotVal = dyn_castNotVal(Op0)) 2088 if (Value *Op1NotVal = dyn_castNotVal(Op1)) 2089 if (Op0->hasOneUse() && Op1->hasOneUse()) { 2090 Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal, 2091 I.getName()+".demorgan"); 2092 return BinaryOperator::CreateNot(And); 2093 } 2094 2095 // Canonicalize xor to the RHS. 2096 bool SwappedForXor = false; 2097 if (match(Op0, m_Xor(m_Value(), m_Value()))) { 2098 std::swap(Op0, Op1); 2099 SwappedForXor = true; 2100 } 2101 2102 // A | ( A ^ B) -> A | B 2103 // A | (~A ^ B) -> A | ~B 2104 // (A & B) | (A ^ B) 2105 if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) { 2106 if (Op0 == A || Op0 == B) 2107 return BinaryOperator::CreateOr(A, B); 2108 2109 if (match(Op0, m_And(m_Specific(A), m_Specific(B))) || 2110 match(Op0, m_And(m_Specific(B), m_Specific(A)))) 2111 return BinaryOperator::CreateOr(A, B); 2112 2113 if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) { 2114 Value *Not = Builder->CreateNot(B, B->getName()+".not"); 2115 return BinaryOperator::CreateOr(Not, Op0); 2116 } 2117 if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) { 2118 Value *Not = Builder->CreateNot(A, A->getName()+".not"); 2119 return BinaryOperator::CreateOr(Not, Op0); 2120 } 2121 } 2122 2123 // A | ~(A | B) -> A | ~B 2124 // A | ~(A ^ B) -> A | ~B 2125 if (match(Op1, m_Not(m_Value(A)))) 2126 if (BinaryOperator *B = dyn_cast<BinaryOperator>(A)) 2127 if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) && 2128 Op1->hasOneUse() && (B->getOpcode() == Instruction::Or || 2129 B->getOpcode() == Instruction::Xor)) { 2130 Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) : 2131 B->getOperand(0); 2132 Value *Not = Builder->CreateNot(NotOp, NotOp->getName()+".not"); 2133 return BinaryOperator::CreateOr(Not, Op0); 2134 } 2135 2136 if (SwappedForXor) 2137 std::swap(Op0, Op1); 2138 2139 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) 2140 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0))) 2141 if (Value *Res = FoldOrOfICmps(LHS, RHS)) 2142 return ReplaceInstUsesWith(I, Res); 2143 2144 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y) 2145 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) 2146 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) 2147 if (Value *Res = FoldOrOfFCmps(LHS, RHS)) 2148 return ReplaceInstUsesWith(I, Res); 2149 2150 // fold (or (cast A), (cast B)) -> (cast (or A, B)) 2151 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) { 2152 CastInst *Op1C = dyn_cast<CastInst>(Op1); 2153 if (Op1C && Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ? 2154 Type *SrcTy = Op0C->getOperand(0)->getType(); 2155 if (SrcTy == Op1C->getOperand(0)->getType() && 2156 SrcTy->isIntOrIntVectorTy()) { 2157 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0); 2158 2159 if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) && 2160 // Only do this if the casts both really cause code to be 2161 // generated. 2162 ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) && 2163 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) { 2164 Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName()); 2165 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType()); 2166 } 2167 2168 // If this is or(cast(icmp), cast(icmp)), try to fold this even if the 2169 // cast is otherwise not optimizable. This happens for vector sexts. 2170 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp)) 2171 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp)) 2172 if (Value *Res = FoldOrOfICmps(LHS, RHS)) 2173 return CastInst::Create(Op0C->getOpcode(), Res, I.getType()); 2174 2175 // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the 2176 // cast is otherwise not optimizable. This happens for vector sexts. 2177 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp)) 2178 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp)) 2179 if (Value *Res = FoldOrOfFCmps(LHS, RHS)) 2180 return CastInst::Create(Op0C->getOpcode(), Res, I.getType()); 2181 } 2182 } 2183 } 2184 2185 // or(sext(A), B) -> A ? -1 : B where A is an i1 2186 // or(A, sext(B)) -> B ? -1 : A where B is an i1 2187 if (match(Op0, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1)) 2188 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1); 2189 if (match(Op1, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1)) 2190 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0); 2191 2192 // Note: If we've gotten to the point of visiting the outer OR, then the 2193 // inner one couldn't be simplified. If it was a constant, then it won't 2194 // be simplified by a later pass either, so we try swapping the inner/outer 2195 // ORs in the hopes that we'll be able to simplify it this way. 2196 // (X|C) | V --> (X|V) | C 2197 if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) && 2198 match(Op0, m_Or(m_Value(A), m_ConstantInt(C1)))) { 2199 Value *Inner = Builder->CreateOr(A, Op1); 2200 Inner->takeName(Op0); 2201 return BinaryOperator::CreateOr(Inner, C1); 2202 } 2203 2204 // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D)) 2205 // Since this OR statement hasn't been optimized further yet, we hope 2206 // that this transformation will allow the new ORs to be optimized. 2207 { 2208 Value *X = nullptr, *Y = nullptr; 2209 if (Op0->hasOneUse() && Op1->hasOneUse() && 2210 match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) && 2211 match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) { 2212 Value *orTrue = Builder->CreateOr(A, C); 2213 Value *orFalse = Builder->CreateOr(B, D); 2214 return SelectInst::Create(X, orTrue, orFalse); 2215 } 2216 } 2217 2218 return Changed ? &I : nullptr; 2219 } 2220 2221 Instruction *InstCombiner::visitXor(BinaryOperator &I) { 2222 bool Changed = SimplifyAssociativeOrCommutative(I); 2223 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1); 2224 2225 if (Value *V = SimplifyVectorOp(I)) 2226 return ReplaceInstUsesWith(I, V); 2227 2228 if (Value *V = SimplifyXorInst(Op0, Op1, DL)) 2229 return ReplaceInstUsesWith(I, V); 2230 2231 // (A&B)^(A&C) -> A&(B^C) etc 2232 if (Value *V = SimplifyUsingDistributiveLaws(I)) 2233 return ReplaceInstUsesWith(I, V); 2234 2235 // See if we can simplify any instructions used by the instruction whose sole 2236 // purpose is to compute bits we don't care about. 2237 if (SimplifyDemandedInstructionBits(I)) 2238 return &I; 2239 2240 // Is this a ~ operation? 2241 if (Value *NotOp = dyn_castNotVal(&I)) { 2242 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) { 2243 if (Op0I->getOpcode() == Instruction::And || 2244 Op0I->getOpcode() == Instruction::Or) { 2245 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law 2246 // ~(~X | Y) === (X & ~Y) - De Morgan's Law 2247 if (dyn_castNotVal(Op0I->getOperand(1))) 2248 Op0I->swapOperands(); 2249 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) { 2250 Value *NotY = 2251 Builder->CreateNot(Op0I->getOperand(1), 2252 Op0I->getOperand(1)->getName()+".not"); 2253 if (Op0I->getOpcode() == Instruction::And) 2254 return BinaryOperator::CreateOr(Op0NotVal, NotY); 2255 return BinaryOperator::CreateAnd(Op0NotVal, NotY); 2256 } 2257 2258 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law 2259 // ~(X | Y) === (~X & ~Y) - De Morgan's Law 2260 if (isFreeToInvert(Op0I->getOperand(0)) && 2261 isFreeToInvert(Op0I->getOperand(1))) { 2262 Value *NotX = 2263 Builder->CreateNot(Op0I->getOperand(0), "notlhs"); 2264 Value *NotY = 2265 Builder->CreateNot(Op0I->getOperand(1), "notrhs"); 2266 if (Op0I->getOpcode() == Instruction::And) 2267 return BinaryOperator::CreateOr(NotX, NotY); 2268 return BinaryOperator::CreateAnd(NotX, NotY); 2269 } 2270 2271 } else if (Op0I->getOpcode() == Instruction::AShr) { 2272 // ~(~X >>s Y) --> (X >>s Y) 2273 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) 2274 return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1)); 2275 } 2276 } 2277 } 2278 2279 2280 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) { 2281 if (RHS->isOne() && Op0->hasOneUse()) 2282 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B 2283 if (CmpInst *CI = dyn_cast<CmpInst>(Op0)) 2284 return CmpInst::Create(CI->getOpcode(), 2285 CI->getInversePredicate(), 2286 CI->getOperand(0), CI->getOperand(1)); 2287 2288 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp). 2289 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) { 2290 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) { 2291 if (CI->hasOneUse() && Op0C->hasOneUse()) { 2292 Instruction::CastOps Opcode = Op0C->getOpcode(); 2293 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) && 2294 (RHS == ConstantExpr::getCast(Opcode, Builder->getTrue(), 2295 Op0C->getDestTy()))) { 2296 CI->setPredicate(CI->getInversePredicate()); 2297 return CastInst::Create(Opcode, CI, Op0C->getType()); 2298 } 2299 } 2300 } 2301 } 2302 2303 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) { 2304 // ~(c-X) == X-c-1 == X+(-c-1) 2305 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue()) 2306 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) { 2307 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C); 2308 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C, 2309 ConstantInt::get(I.getType(), 1)); 2310 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS); 2311 } 2312 2313 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) { 2314 if (Op0I->getOpcode() == Instruction::Add) { 2315 // ~(X-c) --> (-c-1)-X 2316 if (RHS->isAllOnesValue()) { 2317 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI); 2318 return BinaryOperator::CreateSub( 2319 ConstantExpr::getSub(NegOp0CI, 2320 ConstantInt::get(I.getType(), 1)), 2321 Op0I->getOperand(0)); 2322 } else if (RHS->getValue().isSignBit()) { 2323 // (X + C) ^ signbit -> (X + C + signbit) 2324 Constant *C = Builder->getInt(RHS->getValue() + Op0CI->getValue()); 2325 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C); 2326 2327 } 2328 } else if (Op0I->getOpcode() == Instruction::Or) { 2329 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0 2330 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) { 2331 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS); 2332 // Anything in both C1 and C2 is known to be zero, remove it from 2333 // NewRHS. 2334 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS); 2335 NewRHS = ConstantExpr::getAnd(NewRHS, 2336 ConstantExpr::getNot(CommonBits)); 2337 Worklist.Add(Op0I); 2338 I.setOperand(0, Op0I->getOperand(0)); 2339 I.setOperand(1, NewRHS); 2340 return &I; 2341 } 2342 } else if (Op0I->getOpcode() == Instruction::LShr) { 2343 // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3) 2344 // E1 = "X ^ C1" 2345 BinaryOperator *E1; 2346 ConstantInt *C1; 2347 if (Op0I->hasOneUse() && 2348 (E1 = dyn_cast<BinaryOperator>(Op0I->getOperand(0))) && 2349 E1->getOpcode() == Instruction::Xor && 2350 (C1 = dyn_cast<ConstantInt>(E1->getOperand(1)))) { 2351 // fold (C1 >> C2) ^ C3 2352 ConstantInt *C2 = Op0CI, *C3 = RHS; 2353 APInt FoldConst = C1->getValue().lshr(C2->getValue()); 2354 FoldConst ^= C3->getValue(); 2355 // Prepare the two operands. 2356 Value *Opnd0 = Builder->CreateLShr(E1->getOperand(0), C2); 2357 Opnd0->takeName(Op0I); 2358 cast<Instruction>(Opnd0)->setDebugLoc(I.getDebugLoc()); 2359 Value *FoldVal = ConstantInt::get(Opnd0->getType(), FoldConst); 2360 2361 return BinaryOperator::CreateXor(Opnd0, FoldVal); 2362 } 2363 } 2364 } 2365 } 2366 2367 // Try to fold constant and into select arguments. 2368 if (SelectInst *SI = dyn_cast<SelectInst>(Op0)) 2369 if (Instruction *R = FoldOpIntoSelect(I, SI)) 2370 return R; 2371 if (isa<PHINode>(Op0)) 2372 if (Instruction *NV = FoldOpIntoPhi(I)) 2373 return NV; 2374 } 2375 2376 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1); 2377 if (Op1I) { 2378 Value *A, *B; 2379 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) { 2380 if (A == Op0) { // B^(B|A) == (A|B)^B 2381 Op1I->swapOperands(); 2382 I.swapOperands(); 2383 std::swap(Op0, Op1); 2384 } else if (B == Op0) { // B^(A|B) == (A|B)^B 2385 I.swapOperands(); // Simplified below. 2386 std::swap(Op0, Op1); 2387 } 2388 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && 2389 Op1I->hasOneUse()){ 2390 if (A == Op0) { // A^(A&B) -> A^(B&A) 2391 Op1I->swapOperands(); 2392 std::swap(A, B); 2393 } 2394 if (B == Op0) { // A^(B&A) -> (B&A)^A 2395 I.swapOperands(); // Simplified below. 2396 std::swap(Op0, Op1); 2397 } 2398 } 2399 } 2400 2401 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0); 2402 if (Op0I) { 2403 Value *A, *B; 2404 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && 2405 Op0I->hasOneUse()) { 2406 if (A == Op1) // (B|A)^B == (A|B)^B 2407 std::swap(A, B); 2408 if (B == Op1) // (A|B)^B == A & ~B 2409 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1)); 2410 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && 2411 Op0I->hasOneUse()){ 2412 if (A == Op1) // (A&B)^A -> (B&A)^A 2413 std::swap(A, B); 2414 if (B == Op1 && // (B&A)^A == ~B & A 2415 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C 2416 return BinaryOperator::CreateAnd(Builder->CreateNot(A), Op1); 2417 } 2418 } 2419 } 2420 2421 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts. 2422 if (Op0I && Op1I && Op0I->isShift() && 2423 Op0I->getOpcode() == Op1I->getOpcode() && 2424 Op0I->getOperand(1) == Op1I->getOperand(1) && 2425 (Op0I->hasOneUse() || Op1I->hasOneUse())) { 2426 Value *NewOp = 2427 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0), 2428 Op0I->getName()); 2429 return BinaryOperator::Create(Op1I->getOpcode(), NewOp, 2430 Op1I->getOperand(1)); 2431 } 2432 2433 if (Op0I && Op1I) { 2434 Value *A, *B, *C, *D; 2435 // (A & B)^(A | B) -> A ^ B 2436 if (match(Op0I, m_And(m_Value(A), m_Value(B))) && 2437 match(Op1I, m_Or(m_Value(C), m_Value(D)))) { 2438 if ((A == C && B == D) || (A == D && B == C)) 2439 return BinaryOperator::CreateXor(A, B); 2440 } 2441 // (A | B)^(A & B) -> A ^ B 2442 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && 2443 match(Op1I, m_And(m_Value(C), m_Value(D)))) { 2444 if ((A == C && B == D) || (A == D && B == C)) 2445 return BinaryOperator::CreateXor(A, B); 2446 } 2447 } 2448 2449 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B) 2450 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) 2451 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0))) 2452 if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) { 2453 if (LHS->getOperand(0) == RHS->getOperand(1) && 2454 LHS->getOperand(1) == RHS->getOperand(0)) 2455 LHS->swapOperands(); 2456 if (LHS->getOperand(0) == RHS->getOperand(0) && 2457 LHS->getOperand(1) == RHS->getOperand(1)) { 2458 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1); 2459 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS); 2460 bool isSigned = LHS->isSigned() || RHS->isSigned(); 2461 return ReplaceInstUsesWith(I, 2462 getNewICmpValue(isSigned, Code, Op0, Op1, 2463 Builder)); 2464 } 2465 } 2466 2467 // fold (xor (cast A), (cast B)) -> (cast (xor A, B)) 2468 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) { 2469 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) 2470 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind? 2471 Type *SrcTy = Op0C->getOperand(0)->getType(); 2472 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() && 2473 // Only do this if the casts both really cause code to be generated. 2474 ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0), 2475 I.getType()) && 2476 ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0), 2477 I.getType())) { 2478 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0), 2479 Op1C->getOperand(0), I.getName()); 2480 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType()); 2481 } 2482 } 2483 } 2484 2485 return Changed ? &I : nullptr; 2486 } 2487