1 //===- InstCombineAndOrXor.cpp --------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visitAnd, visitOr, and visitXor functions.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombineInternal.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"
21 using namespace PatternMatch;
23 #define DEBUG_TYPE "instcombine"
25 static inline Value *dyn_castNotVal(Value *V) {
26 // If this is not(not(x)) don't return that this is a not: we want the two
27 // not's to be folded first.
28 if (BinaryOperator::isNot(V)) {
29 Value *Operand = BinaryOperator::getNotArgument(V);
30 if (!IsFreeToInvert(Operand, Operand->hasOneUse()))
34 // Constants can be considered to be not'ed values...
35 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
36 return ConstantInt::get(C->getType(), ~C->getValue());
40 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
41 /// predicate into a three bit mask. It also returns whether it is an ordered
42 /// predicate by reference.
43 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
46 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
47 case FCmpInst::FCMP_UNO: return 0; // 000
48 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
49 case FCmpInst::FCMP_UGT: return 1; // 001
50 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
51 case FCmpInst::FCMP_UEQ: return 2; // 010
52 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
53 case FCmpInst::FCMP_UGE: return 3; // 011
54 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
55 case FCmpInst::FCMP_ULT: return 4; // 100
56 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
57 case FCmpInst::FCMP_UNE: return 5; // 101
58 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
59 case FCmpInst::FCMP_ULE: return 6; // 110
62 // Not expecting FCMP_FALSE and FCMP_TRUE;
63 llvm_unreachable("Unexpected FCmp predicate!");
67 /// getNewICmpValue - This is the complement of getICmpCode, which turns an
68 /// opcode and two operands into either a constant true or false, or a brand
69 /// new ICmp instruction. The sign is passed in to determine which kind
70 /// of predicate to use in the new icmp instruction.
71 static Value *getNewICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS,
72 InstCombiner::BuilderTy *Builder) {
73 ICmpInst::Predicate NewPred;
74 if (Value *NewConstant = getICmpValue(Sign, Code, LHS, RHS, NewPred))
76 return Builder->CreateICmp(NewPred, LHS, RHS);
79 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
80 /// opcode and two operands into either a FCmp instruction. isordered is passed
81 /// in to determine which kind of predicate to use in the new fcmp instruction.
82 static Value *getFCmpValue(bool isordered, unsigned code,
83 Value *LHS, Value *RHS,
84 InstCombiner::BuilderTy *Builder) {
85 CmpInst::Predicate Pred;
87 default: llvm_unreachable("Illegal FCmp code!");
88 case 0: Pred = isordered ? FCmpInst::FCMP_ORD : FCmpInst::FCMP_UNO; break;
89 case 1: Pred = isordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; break;
90 case 2: Pred = isordered ? FCmpInst::FCMP_OEQ : FCmpInst::FCMP_UEQ; break;
91 case 3: Pred = isordered ? FCmpInst::FCMP_OGE : FCmpInst::FCMP_UGE; break;
92 case 4: Pred = isordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; break;
93 case 5: Pred = isordered ? FCmpInst::FCMP_ONE : FCmpInst::FCMP_UNE; break;
94 case 6: Pred = isordered ? FCmpInst::FCMP_OLE : FCmpInst::FCMP_ULE; break;
97 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
98 Pred = FCmpInst::FCMP_ORD; break;
100 return Builder->CreateFCmp(Pred, LHS, RHS);
103 /// \brief Transform BITWISE_OP(BSWAP(A),BSWAP(B)) to BSWAP(BITWISE_OP(A, B))
104 /// \param I Binary operator to transform.
105 /// \return Pointer to node that must replace the original binary operator, or
106 /// null pointer if no transformation was made.
107 Value *InstCombiner::SimplifyBSwap(BinaryOperator &I) {
108 IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
111 if (I.getType()->isVectorTy()) return nullptr;
113 // Can only do bitwise ops.
114 unsigned Op = I.getOpcode();
115 if (Op != Instruction::And && Op != Instruction::Or &&
116 Op != Instruction::Xor)
119 Value *OldLHS = I.getOperand(0);
120 Value *OldRHS = I.getOperand(1);
121 ConstantInt *ConstLHS = dyn_cast<ConstantInt>(OldLHS);
122 ConstantInt *ConstRHS = dyn_cast<ConstantInt>(OldRHS);
123 IntrinsicInst *IntrLHS = dyn_cast<IntrinsicInst>(OldLHS);
124 IntrinsicInst *IntrRHS = dyn_cast<IntrinsicInst>(OldRHS);
125 bool IsBswapLHS = (IntrLHS && IntrLHS->getIntrinsicID() == Intrinsic::bswap);
126 bool IsBswapRHS = (IntrRHS && IntrRHS->getIntrinsicID() == Intrinsic::bswap);
128 if (!IsBswapLHS && !IsBswapRHS)
131 if (!IsBswapLHS && !ConstLHS)
134 if (!IsBswapRHS && !ConstRHS)
137 /// OP( BSWAP(x), BSWAP(y) ) -> BSWAP( OP(x, y) )
138 /// OP( BSWAP(x), CONSTANT ) -> BSWAP( OP(x, BSWAP(CONSTANT) ) )
139 Value *NewLHS = IsBswapLHS ? IntrLHS->getOperand(0) :
140 Builder->getInt(ConstLHS->getValue().byteSwap());
142 Value *NewRHS = IsBswapRHS ? IntrRHS->getOperand(0) :
143 Builder->getInt(ConstRHS->getValue().byteSwap());
145 Value *BinOp = nullptr;
146 if (Op == Instruction::And)
147 BinOp = Builder->CreateAnd(NewLHS, NewRHS);
148 else if (Op == Instruction::Or)
149 BinOp = Builder->CreateOr(NewLHS, NewRHS);
150 else //if (Op == Instruction::Xor)
151 BinOp = Builder->CreateXor(NewLHS, NewRHS);
153 Module *M = I.getParent()->getParent()->getParent();
154 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, ITy);
155 return Builder->CreateCall(F, BinOp);
158 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
159 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
160 // guaranteed to be a binary operator.
161 Instruction *InstCombiner::OptAndOp(Instruction *Op,
164 BinaryOperator &TheAnd) {
165 Value *X = Op->getOperand(0);
166 Constant *Together = nullptr;
168 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
170 switch (Op->getOpcode()) {
171 case Instruction::Xor:
172 if (Op->hasOneUse()) {
173 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
174 Value *And = Builder->CreateAnd(X, AndRHS);
176 return BinaryOperator::CreateXor(And, Together);
179 case Instruction::Or:
180 if (Op->hasOneUse()){
181 if (Together != OpRHS) {
182 // (X | C1) & C2 --> (X | (C1&C2)) & C2
183 Value *Or = Builder->CreateOr(X, Together);
185 return BinaryOperator::CreateAnd(Or, AndRHS);
188 ConstantInt *TogetherCI = dyn_cast<ConstantInt>(Together);
189 if (TogetherCI && !TogetherCI->isZero()){
190 // (X | C1) & C2 --> (X & (C2^(C1&C2))) | C1
191 // NOTE: This reduces the number of bits set in the & mask, which
192 // can expose opportunities for store narrowing.
193 Together = ConstantExpr::getXor(AndRHS, Together);
194 Value *And = Builder->CreateAnd(X, Together);
196 return BinaryOperator::CreateOr(And, OpRHS);
201 case Instruction::Add:
202 if (Op->hasOneUse()) {
203 // Adding a one to a single bit bit-field should be turned into an XOR
204 // of the bit. First thing to check is to see if this AND is with a
205 // single bit constant.
206 const APInt &AndRHSV = AndRHS->getValue();
208 // If there is only one bit set.
209 if (AndRHSV.isPowerOf2()) {
210 // Ok, at this point, we know that we are masking the result of the
211 // ADD down to exactly one bit. If the constant we are adding has
212 // no bits set below this bit, then we can eliminate the ADD.
213 const APInt& AddRHS = OpRHS->getValue();
215 // Check to see if any bits below the one bit set in AndRHSV are set.
216 if ((AddRHS & (AndRHSV-1)) == 0) {
217 // If not, the only thing that can effect the output of the AND is
218 // the bit specified by AndRHSV. If that bit is set, the effect of
219 // the XOR is to toggle the bit. If it is clear, then the ADD has
221 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
222 TheAnd.setOperand(0, X);
225 // Pull the XOR out of the AND.
226 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
227 NewAnd->takeName(Op);
228 return BinaryOperator::CreateXor(NewAnd, AndRHS);
235 case Instruction::Shl: {
236 // We know that the AND will not produce any of the bits shifted in, so if
237 // the anded constant includes them, clear them now!
239 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
240 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
241 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
242 ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShlMask);
244 if (CI->getValue() == ShlMask)
245 // Masking out bits that the shift already masks.
246 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
248 if (CI != AndRHS) { // Reducing bits set in and.
249 TheAnd.setOperand(1, CI);
254 case Instruction::LShr: {
255 // We know that the AND will not produce any of the bits shifted in, so if
256 // the anded constant includes them, clear them now! This only applies to
257 // unsigned shifts, because a signed shr may bring in set bits!
259 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
260 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
261 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
262 ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShrMask);
264 if (CI->getValue() == ShrMask)
265 // Masking out bits that the shift already masks.
266 return ReplaceInstUsesWith(TheAnd, Op);
269 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
274 case Instruction::AShr:
276 // See if this is shifting in some sign extension, then masking it out
278 if (Op->hasOneUse()) {
279 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
280 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
281 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
282 Constant *C = Builder->getInt(AndRHS->getValue() & ShrMask);
283 if (C == AndRHS) { // Masking out bits shifted in.
284 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
285 // Make the argument unsigned.
286 Value *ShVal = Op->getOperand(0);
287 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
288 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
296 /// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise
297 /// (V < Lo || V >= Hi). In practice, we emit the more efficient
298 /// (V-Lo) \<u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
299 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
300 /// insert new instructions.
301 Value *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
302 bool isSigned, bool Inside) {
303 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
304 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
305 "Lo is not <= Hi in range emission code!");
308 if (Lo == Hi) // Trivially false.
309 return Builder->getFalse();
311 // V >= Min && V < Hi --> V < Hi
312 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
313 ICmpInst::Predicate pred = (isSigned ?
314 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
315 return Builder->CreateICmp(pred, V, Hi);
318 // Emit V-Lo <u Hi-Lo
319 Constant *NegLo = ConstantExpr::getNeg(Lo);
320 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
321 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
322 return Builder->CreateICmpULT(Add, UpperBound);
325 if (Lo == Hi) // Trivially true.
326 return Builder->getTrue();
328 // V < Min || V >= Hi -> V > Hi-1
329 Hi = SubOne(cast<ConstantInt>(Hi));
330 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
331 ICmpInst::Predicate pred = (isSigned ?
332 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
333 return Builder->CreateICmp(pred, V, Hi);
336 // Emit V-Lo >u Hi-1-Lo
337 // Note that Hi has already had one subtracted from it, above.
338 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
339 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
340 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
341 return Builder->CreateICmpUGT(Add, LowerBound);
344 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
345 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
346 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
347 // not, since all 1s are not contiguous.
348 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
349 const APInt& V = Val->getValue();
350 uint32_t BitWidth = Val->getType()->getBitWidth();
351 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
353 // look for the first zero bit after the run of ones
354 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
355 // look for the first non-zero bit
356 ME = V.getActiveBits();
360 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
361 /// where isSub determines whether the operator is a sub. If we can fold one of
362 /// the following xforms:
364 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
365 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
366 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
368 /// return (A +/- B).
370 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
371 ConstantInt *Mask, bool isSub,
373 Instruction *LHSI = dyn_cast<Instruction>(LHS);
374 if (!LHSI || LHSI->getNumOperands() != 2 ||
375 !isa<ConstantInt>(LHSI->getOperand(1))) return nullptr;
377 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
379 switch (LHSI->getOpcode()) {
380 default: return nullptr;
381 case Instruction::And:
382 if (ConstantExpr::getAnd(N, Mask) == Mask) {
383 // If the AndRHS is a power of two minus one (0+1+), this is simple.
384 if ((Mask->getValue().countLeadingZeros() +
385 Mask->getValue().countPopulation()) ==
386 Mask->getValue().getBitWidth())
389 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
390 // part, we don't need any explicit masks to take them out of A. If that
391 // is all N is, ignore it.
392 uint32_t MB = 0, ME = 0;
393 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
394 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
395 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
396 if (MaskedValueIsZero(RHS, Mask, 0, &I))
401 case Instruction::Or:
402 case Instruction::Xor:
403 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
404 if ((Mask->getValue().countLeadingZeros() +
405 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
406 && ConstantExpr::getAnd(N, Mask)->isNullValue())
412 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
413 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
416 /// enum for classifying (icmp eq (A & B), C) and (icmp ne (A & B), C)
417 /// One of A and B is considered the mask, the other the value. This is
418 /// described as the "AMask" or "BMask" part of the enum. If the enum
419 /// contains only "Mask", then both A and B can be considered masks.
420 /// If A is the mask, then it was proven, that (A & C) == C. This
421 /// is trivial if C == A, or C == 0. If both A and C are constants, this
422 /// proof is also easy.
423 /// For the following explanations we assume that A is the mask.
424 /// The part "AllOnes" declares, that the comparison is true only
425 /// if (A & B) == A, or all bits of A are set in B.
426 /// Example: (icmp eq (A & 3), 3) -> FoldMskICmp_AMask_AllOnes
427 /// The part "AllZeroes" declares, that the comparison is true only
428 /// if (A & B) == 0, or all bits of A are cleared in B.
429 /// Example: (icmp eq (A & 3), 0) -> FoldMskICmp_Mask_AllZeroes
430 /// The part "Mixed" declares, that (A & B) == C and C might or might not
431 /// contain any number of one bits and zero bits.
432 /// Example: (icmp eq (A & 3), 1) -> FoldMskICmp_AMask_Mixed
433 /// The Part "Not" means, that in above descriptions "==" should be replaced
435 /// Example: (icmp ne (A & 3), 3) -> FoldMskICmp_AMask_NotAllOnes
436 /// If the mask A contains a single bit, then the following is equivalent:
437 /// (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
438 /// (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
439 enum MaskedICmpType {
440 FoldMskICmp_AMask_AllOnes = 1,
441 FoldMskICmp_AMask_NotAllOnes = 2,
442 FoldMskICmp_BMask_AllOnes = 4,
443 FoldMskICmp_BMask_NotAllOnes = 8,
444 FoldMskICmp_Mask_AllZeroes = 16,
445 FoldMskICmp_Mask_NotAllZeroes = 32,
446 FoldMskICmp_AMask_Mixed = 64,
447 FoldMskICmp_AMask_NotMixed = 128,
448 FoldMskICmp_BMask_Mixed = 256,
449 FoldMskICmp_BMask_NotMixed = 512
452 /// return the set of pattern classes (from MaskedICmpType)
453 /// that (icmp SCC (A & B), C) satisfies
454 static unsigned getTypeOfMaskedICmp(Value* A, Value* B, Value* C,
455 ICmpInst::Predicate SCC)
457 ConstantInt *ACst = dyn_cast<ConstantInt>(A);
458 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
459 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
460 bool icmp_eq = (SCC == ICmpInst::ICMP_EQ);
461 bool icmp_abit = (ACst && !ACst->isZero() &&
462 ACst->getValue().isPowerOf2());
463 bool icmp_bbit = (BCst && !BCst->isZero() &&
464 BCst->getValue().isPowerOf2());
466 if (CCst && CCst->isZero()) {
467 // if C is zero, then both A and B qualify as mask
468 result |= (icmp_eq ? (FoldMskICmp_Mask_AllZeroes |
469 FoldMskICmp_Mask_AllZeroes |
470 FoldMskICmp_AMask_Mixed |
471 FoldMskICmp_BMask_Mixed)
472 : (FoldMskICmp_Mask_NotAllZeroes |
473 FoldMskICmp_Mask_NotAllZeroes |
474 FoldMskICmp_AMask_NotMixed |
475 FoldMskICmp_BMask_NotMixed));
477 result |= (icmp_eq ? (FoldMskICmp_AMask_NotAllOnes |
478 FoldMskICmp_AMask_NotMixed)
479 : (FoldMskICmp_AMask_AllOnes |
480 FoldMskICmp_AMask_Mixed));
482 result |= (icmp_eq ? (FoldMskICmp_BMask_NotAllOnes |
483 FoldMskICmp_BMask_NotMixed)
484 : (FoldMskICmp_BMask_AllOnes |
485 FoldMskICmp_BMask_Mixed));
489 result |= (icmp_eq ? (FoldMskICmp_AMask_AllOnes |
490 FoldMskICmp_AMask_Mixed)
491 : (FoldMskICmp_AMask_NotAllOnes |
492 FoldMskICmp_AMask_NotMixed));
494 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
495 FoldMskICmp_AMask_NotMixed)
496 : (FoldMskICmp_Mask_AllZeroes |
497 FoldMskICmp_AMask_Mixed));
498 } else if (ACst && CCst &&
499 ConstantExpr::getAnd(ACst, CCst) == CCst) {
500 result |= (icmp_eq ? FoldMskICmp_AMask_Mixed
501 : FoldMskICmp_AMask_NotMixed);
504 result |= (icmp_eq ? (FoldMskICmp_BMask_AllOnes |
505 FoldMskICmp_BMask_Mixed)
506 : (FoldMskICmp_BMask_NotAllOnes |
507 FoldMskICmp_BMask_NotMixed));
509 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
510 FoldMskICmp_BMask_NotMixed)
511 : (FoldMskICmp_Mask_AllZeroes |
512 FoldMskICmp_BMask_Mixed));
513 } else if (BCst && CCst &&
514 ConstantExpr::getAnd(BCst, CCst) == CCst) {
515 result |= (icmp_eq ? FoldMskICmp_BMask_Mixed
516 : FoldMskICmp_BMask_NotMixed);
521 /// Convert an analysis of a masked ICmp into its equivalent if all boolean
522 /// operations had the opposite sense. Since each "NotXXX" flag (recording !=)
523 /// is adjacent to the corresponding normal flag (recording ==), this just
524 /// involves swapping those bits over.
525 static unsigned conjugateICmpMask(unsigned Mask) {
527 NewMask = (Mask & (FoldMskICmp_AMask_AllOnes | FoldMskICmp_BMask_AllOnes |
528 FoldMskICmp_Mask_AllZeroes | FoldMskICmp_AMask_Mixed |
529 FoldMskICmp_BMask_Mixed))
533 (Mask & (FoldMskICmp_AMask_NotAllOnes | FoldMskICmp_BMask_NotAllOnes |
534 FoldMskICmp_Mask_NotAllZeroes | FoldMskICmp_AMask_NotMixed |
535 FoldMskICmp_BMask_NotMixed))
541 /// decomposeBitTestICmp - Decompose an icmp into the form ((X & Y) pred Z)
542 /// if possible. The returned predicate is either == or !=. Returns false if
543 /// decomposition fails.
544 static bool decomposeBitTestICmp(const ICmpInst *I, ICmpInst::Predicate &Pred,
545 Value *&X, Value *&Y, Value *&Z) {
546 ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1));
550 switch (I->getPredicate()) {
553 case ICmpInst::ICMP_SLT:
554 // X < 0 is equivalent to (X & SignBit) != 0.
557 Y = ConstantInt::get(I->getContext(), APInt::getSignBit(C->getBitWidth()));
558 Pred = ICmpInst::ICMP_NE;
560 case ICmpInst::ICMP_SGT:
561 // X > -1 is equivalent to (X & SignBit) == 0.
562 if (!C->isAllOnesValue())
564 Y = ConstantInt::get(I->getContext(), APInt::getSignBit(C->getBitWidth()));
565 Pred = ICmpInst::ICMP_EQ;
567 case ICmpInst::ICMP_ULT:
568 // X <u 2^n is equivalent to (X & ~(2^n-1)) == 0.
569 if (!C->getValue().isPowerOf2())
571 Y = ConstantInt::get(I->getContext(), -C->getValue());
572 Pred = ICmpInst::ICMP_EQ;
574 case ICmpInst::ICMP_UGT:
575 // X >u 2^n-1 is equivalent to (X & ~(2^n-1)) != 0.
576 if (!(C->getValue() + 1).isPowerOf2())
578 Y = ConstantInt::get(I->getContext(), ~C->getValue());
579 Pred = ICmpInst::ICMP_NE;
583 X = I->getOperand(0);
584 Z = ConstantInt::getNullValue(C->getType());
588 /// foldLogOpOfMaskedICmpsHelper:
589 /// handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
590 /// return the set of pattern classes (from MaskedICmpType)
591 /// that both LHS and RHS satisfy
592 static unsigned foldLogOpOfMaskedICmpsHelper(Value*& A,
593 Value*& B, Value*& C,
594 Value*& D, Value*& E,
595 ICmpInst *LHS, ICmpInst *RHS,
596 ICmpInst::Predicate &LHSCC,
597 ICmpInst::Predicate &RHSCC) {
598 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) return 0;
599 // vectors are not (yet?) supported
600 if (LHS->getOperand(0)->getType()->isVectorTy()) return 0;
602 // Here comes the tricky part:
603 // LHS might be of the form L11 & L12 == X, X == L21 & L22,
604 // and L11 & L12 == L21 & L22. The same goes for RHS.
605 // Now we must find those components L** and R**, that are equal, so
606 // that we can extract the parameters A, B, C, D, and E for the canonical
608 Value *L1 = LHS->getOperand(0);
609 Value *L2 = LHS->getOperand(1);
610 Value *L11,*L12,*L21,*L22;
611 // Check whether the icmp can be decomposed into a bit test.
612 if (decomposeBitTestICmp(LHS, LHSCC, L11, L12, L2)) {
613 L21 = L22 = L1 = nullptr;
615 // Look for ANDs in the LHS icmp.
616 if (!L1->getType()->isIntegerTy()) {
617 // You can icmp pointers, for example. They really aren't masks.
619 } else if (!match(L1, m_And(m_Value(L11), m_Value(L12)))) {
620 // Any icmp can be viewed as being trivially masked; if it allows us to
621 // remove one, it's worth it.
623 L12 = Constant::getAllOnesValue(L1->getType());
626 if (!L2->getType()->isIntegerTy()) {
627 // You can icmp pointers, for example. They really aren't masks.
629 } else if (!match(L2, m_And(m_Value(L21), m_Value(L22)))) {
631 L22 = Constant::getAllOnesValue(L2->getType());
635 // Bail if LHS was a icmp that can't be decomposed into an equality.
636 if (!ICmpInst::isEquality(LHSCC))
639 Value *R1 = RHS->getOperand(0);
640 Value *R2 = RHS->getOperand(1);
643 if (decomposeBitTestICmp(RHS, RHSCC, R11, R12, R2)) {
644 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
646 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
651 E = R2; R1 = nullptr; ok = true;
652 } else if (R1->getType()->isIntegerTy()) {
653 if (!match(R1, m_And(m_Value(R11), m_Value(R12)))) {
654 // As before, model no mask as a trivial mask if it'll let us do an
657 R12 = Constant::getAllOnesValue(R1->getType());
660 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
661 A = R11; D = R12; E = R2; ok = true;
662 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
663 A = R12; D = R11; E = R2; ok = true;
667 // Bail if RHS was a icmp that can't be decomposed into an equality.
668 if (!ICmpInst::isEquality(RHSCC))
671 // Look for ANDs in on the right side of the RHS icmp.
672 if (!ok && R2->getType()->isIntegerTy()) {
673 if (!match(R2, m_And(m_Value(R11), m_Value(R12)))) {
675 R12 = Constant::getAllOnesValue(R2->getType());
678 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
679 A = R11; D = R12; E = R1; ok = true;
680 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
681 A = R12; D = R11; E = R1; ok = true;
691 } else if (L12 == A) {
693 } else if (L21 == A) {
695 } else if (L22 == A) {
699 unsigned left_type = getTypeOfMaskedICmp(A, B, C, LHSCC);
700 unsigned right_type = getTypeOfMaskedICmp(A, D, E, RHSCC);
701 return left_type & right_type;
703 /// foldLogOpOfMaskedICmps:
704 /// try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
705 /// into a single (icmp(A & X) ==/!= Y)
706 static Value *foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
707 llvm::InstCombiner::BuilderTy *Builder) {
708 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr;
709 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
710 unsigned mask = foldLogOpOfMaskedICmpsHelper(A, B, C, D, E, LHS, RHS,
712 if (mask == 0) return nullptr;
713 assert(ICmpInst::isEquality(LHSCC) && ICmpInst::isEquality(RHSCC) &&
714 "foldLogOpOfMaskedICmpsHelper must return an equality predicate.");
716 // In full generality:
717 // (icmp (A & B) Op C) | (icmp (A & D) Op E)
718 // == ![ (icmp (A & B) !Op C) & (icmp (A & D) !Op E) ]
720 // If the latter can be converted into (icmp (A & X) Op Y) then the former is
721 // equivalent to (icmp (A & X) !Op Y).
723 // Therefore, we can pretend for the rest of this function that we're dealing
724 // with the conjunction, provided we flip the sense of any comparisons (both
725 // input and output).
727 // In most cases we're going to produce an EQ for the "&&" case.
728 ICmpInst::Predicate NEWCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
730 // Convert the masking analysis into its equivalent with negated
732 mask = conjugateICmpMask(mask);
735 if (mask & FoldMskICmp_Mask_AllZeroes) {
736 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
737 // -> (icmp eq (A & (B|D)), 0)
738 Value *newOr = Builder->CreateOr(B, D);
739 Value *newAnd = Builder->CreateAnd(A, newOr);
740 // we can't use C as zero, because we might actually handle
741 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
742 // with B and D, having a single bit set
743 Value *zero = Constant::getNullValue(A->getType());
744 return Builder->CreateICmp(NEWCC, newAnd, zero);
746 if (mask & FoldMskICmp_BMask_AllOnes) {
747 // (icmp eq (A & B), B) & (icmp eq (A & D), D)
748 // -> (icmp eq (A & (B|D)), (B|D))
749 Value *newOr = Builder->CreateOr(B, D);
750 Value *newAnd = Builder->CreateAnd(A, newOr);
751 return Builder->CreateICmp(NEWCC, newAnd, newOr);
753 if (mask & FoldMskICmp_AMask_AllOnes) {
754 // (icmp eq (A & B), A) & (icmp eq (A & D), A)
755 // -> (icmp eq (A & (B&D)), A)
756 Value *newAnd1 = Builder->CreateAnd(B, D);
757 Value *newAnd = Builder->CreateAnd(A, newAnd1);
758 return Builder->CreateICmp(NEWCC, newAnd, A);
761 // Remaining cases assume at least that B and D are constant, and depend on
762 // their actual values. This isn't strictly, necessary, just a "handle the
763 // easy cases for now" decision.
764 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
765 if (!BCst) return nullptr;
766 ConstantInt *DCst = dyn_cast<ConstantInt>(D);
767 if (!DCst) return nullptr;
769 if (mask & (FoldMskICmp_Mask_NotAllZeroes | FoldMskICmp_BMask_NotAllOnes)) {
770 // (icmp ne (A & B), 0) & (icmp ne (A & D), 0) and
771 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
772 // -> (icmp ne (A & B), 0) or (icmp ne (A & D), 0)
773 // Only valid if one of the masks is a superset of the other (check "B&D" is
774 // the same as either B or D).
775 APInt NewMask = BCst->getValue() & DCst->getValue();
777 if (NewMask == BCst->getValue())
779 else if (NewMask == DCst->getValue())
782 if (mask & FoldMskICmp_AMask_NotAllOnes) {
783 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
784 // -> (icmp ne (A & B), A) or (icmp ne (A & D), A)
785 // Only valid if one of the masks is a superset of the other (check "B|D" is
786 // the same as either B or D).
787 APInt NewMask = BCst->getValue() | DCst->getValue();
789 if (NewMask == BCst->getValue())
791 else if (NewMask == DCst->getValue())
794 if (mask & FoldMskICmp_BMask_Mixed) {
795 // (icmp eq (A & B), C) & (icmp eq (A & D), E)
796 // We already know that B & C == C && D & E == E.
797 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
798 // C and E, which are shared by both the mask B and the mask D, don't
799 // contradict, then we can transform to
800 // -> (icmp eq (A & (B|D)), (C|E))
801 // Currently, we only handle the case of B, C, D, and E being constant.
802 // we can't simply use C and E, because we might actually handle
803 // (icmp ne (A & B), B) & (icmp eq (A & D), D)
804 // with B and D, having a single bit set
805 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
806 if (!CCst) return nullptr;
807 ConstantInt *ECst = dyn_cast<ConstantInt>(E);
808 if (!ECst) return nullptr;
810 CCst = cast<ConstantInt>(ConstantExpr::getXor(BCst, CCst));
812 ECst = cast<ConstantInt>(ConstantExpr::getXor(DCst, ECst));
813 // if there is a conflict we should actually return a false for the
815 if (((BCst->getValue() & DCst->getValue()) &
816 (CCst->getValue() ^ ECst->getValue())) != 0)
817 return ConstantInt::get(LHS->getType(), !IsAnd);
818 Value *newOr1 = Builder->CreateOr(B, D);
819 Value *newOr2 = ConstantExpr::getOr(CCst, ECst);
820 Value *newAnd = Builder->CreateAnd(A, newOr1);
821 return Builder->CreateICmp(NEWCC, newAnd, newOr2);
826 /// Try to fold a signed range checked with lower bound 0 to an unsigned icmp.
827 /// Example: (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
828 /// If \p Inverted is true then the check is for the inverted range, e.g.
829 /// (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
830 Value *InstCombiner::simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1,
832 // Check the lower range comparison, e.g. x >= 0
833 // InstCombine already ensured that if there is a constant it's on the RHS.
834 ConstantInt *RangeStart = dyn_cast<ConstantInt>(Cmp0->getOperand(1));
838 ICmpInst::Predicate Pred0 = (Inverted ? Cmp0->getInversePredicate() :
839 Cmp0->getPredicate());
841 // Accept x > -1 or x >= 0 (after potentially inverting the predicate).
842 if (!((Pred0 == ICmpInst::ICMP_SGT && RangeStart->isMinusOne()) ||
843 (Pred0 == ICmpInst::ICMP_SGE && RangeStart->isZero())))
846 ICmpInst::Predicate Pred1 = (Inverted ? Cmp1->getInversePredicate() :
847 Cmp1->getPredicate());
849 Value *Input = Cmp0->getOperand(0);
851 if (Cmp1->getOperand(0) == Input) {
852 // For the upper range compare we have: icmp x, n
853 RangeEnd = Cmp1->getOperand(1);
854 } else if (Cmp1->getOperand(1) == Input) {
855 // For the upper range compare we have: icmp n, x
856 RangeEnd = Cmp1->getOperand(0);
857 Pred1 = ICmpInst::getSwappedPredicate(Pred1);
862 // Check the upper range comparison, e.g. x < n
863 ICmpInst::Predicate NewPred;
865 case ICmpInst::ICMP_SLT: NewPred = ICmpInst::ICMP_ULT; break;
866 case ICmpInst::ICMP_SLE: NewPred = ICmpInst::ICMP_ULE; break;
867 default: return nullptr;
870 // This simplification is only valid if the upper range is not negative.
871 bool IsNegative, IsNotNegative;
872 ComputeSignBit(RangeEnd, IsNotNegative, IsNegative, /*Depth=*/0, Cmp1);
877 NewPred = ICmpInst::getInversePredicate(NewPred);
879 return Builder->CreateICmp(NewPred, Input, RangeEnd);
882 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
883 Value *InstCombiner::FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
884 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
886 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
887 if (PredicatesFoldable(LHSCC, RHSCC)) {
888 if (LHS->getOperand(0) == RHS->getOperand(1) &&
889 LHS->getOperand(1) == RHS->getOperand(0))
891 if (LHS->getOperand(0) == RHS->getOperand(0) &&
892 LHS->getOperand(1) == RHS->getOperand(1)) {
893 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
894 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
895 bool isSigned = LHS->isSigned() || RHS->isSigned();
896 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
900 // handle (roughly): (icmp eq (A & B), C) & (icmp eq (A & D), E)
901 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, true, Builder))
904 // E.g. (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
905 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/false))
908 // E.g. (icmp slt x, n) & (icmp sge x, 0) --> icmp ult x, n
909 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/false))
912 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
913 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
914 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
915 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
916 if (!LHSCst || !RHSCst) return nullptr;
918 if (LHSCst == RHSCst && LHSCC == RHSCC) {
919 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
920 // where C is a power of 2
921 if (LHSCC == ICmpInst::ICMP_ULT &&
922 LHSCst->getValue().isPowerOf2()) {
923 Value *NewOr = Builder->CreateOr(Val, Val2);
924 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
927 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
928 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
929 Value *NewOr = Builder->CreateOr(Val, Val2);
930 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
934 // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2
935 // where CMAX is the all ones value for the truncated type,
936 // iff the lower bits of C2 and CA are zero.
937 if (LHSCC == ICmpInst::ICMP_EQ && LHSCC == RHSCC &&
938 LHS->hasOneUse() && RHS->hasOneUse()) {
940 ConstantInt *AndCst, *SmallCst = nullptr, *BigCst = nullptr;
942 // (trunc x) == C1 & (and x, CA) == C2
943 // (and x, CA) == C2 & (trunc x) == C1
944 if (match(Val2, m_Trunc(m_Value(V))) &&
945 match(Val, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
948 } else if (match(Val, m_Trunc(m_Value(V))) &&
949 match(Val2, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
954 if (SmallCst && BigCst) {
955 unsigned BigBitSize = BigCst->getType()->getBitWidth();
956 unsigned SmallBitSize = SmallCst->getType()->getBitWidth();
958 // Check that the low bits are zero.
959 APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize);
960 if ((Low & AndCst->getValue()) == 0 && (Low & BigCst->getValue()) == 0) {
961 Value *NewAnd = Builder->CreateAnd(V, Low | AndCst->getValue());
962 APInt N = SmallCst->getValue().zext(BigBitSize) | BigCst->getValue();
963 Value *NewVal = ConstantInt::get(AndCst->getType()->getContext(), N);
964 return Builder->CreateICmp(LHSCC, NewAnd, NewVal);
969 // From here on, we only handle:
970 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
971 if (Val != Val2) return nullptr;
973 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
974 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
975 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
976 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
977 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
980 // Make a constant range that's the intersection of the two icmp ranges.
981 // If the intersection is empty, we know that the result is false.
982 ConstantRange LHSRange =
983 ConstantRange::makeAllowedICmpRegion(LHSCC, LHSCst->getValue());
984 ConstantRange RHSRange =
985 ConstantRange::makeAllowedICmpRegion(RHSCC, RHSCst->getValue());
987 if (LHSRange.intersectWith(RHSRange).isEmptySet())
988 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
990 // We can't fold (ugt x, C) & (sgt x, C2).
991 if (!PredicatesFoldable(LHSCC, RHSCC))
994 // Ensure that the larger constant is on the RHS.
996 if (CmpInst::isSigned(LHSCC) ||
997 (ICmpInst::isEquality(LHSCC) &&
998 CmpInst::isSigned(RHSCC)))
999 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
1001 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
1004 std::swap(LHS, RHS);
1005 std::swap(LHSCst, RHSCst);
1006 std::swap(LHSCC, RHSCC);
1009 // At this point, we know we have two icmp instructions
1010 // comparing a value against two constants and and'ing the result
1011 // together. Because of the above check, we know that we only have
1012 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
1013 // (from the icmp folding check above), that the two constants
1014 // are not equal and that the larger constant is on the RHS
1015 assert(LHSCst != RHSCst && "Compares not folded above?");
1018 default: llvm_unreachable("Unknown integer condition code!");
1019 case ICmpInst::ICMP_EQ:
1021 default: llvm_unreachable("Unknown integer condition code!");
1022 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
1023 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
1024 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
1027 case ICmpInst::ICMP_NE:
1029 default: llvm_unreachable("Unknown integer condition code!");
1030 case ICmpInst::ICMP_ULT:
1031 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
1032 return Builder->CreateICmpULT(Val, LHSCst);
1033 if (LHSCst->isNullValue()) // (X != 0 & X u< 14) -> X-1 u< 13
1034 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
1035 break; // (X != 13 & X u< 15) -> no change
1036 case ICmpInst::ICMP_SLT:
1037 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
1038 return Builder->CreateICmpSLT(Val, LHSCst);
1039 break; // (X != 13 & X s< 15) -> no change
1040 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
1041 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
1042 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
1044 case ICmpInst::ICMP_NE:
1045 // Special case to get the ordering right when the values wrap around
1047 if (LHSCst->getValue() == 0 && RHSCst->getValue().isAllOnesValue())
1048 std::swap(LHSCst, RHSCst);
1049 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
1050 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1051 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1052 return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1),
1053 Val->getName()+".cmp");
1055 break; // (X != 13 & X != 15) -> no change
1058 case ICmpInst::ICMP_ULT:
1060 default: llvm_unreachable("Unknown integer condition code!");
1061 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
1062 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
1063 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
1064 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
1066 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
1067 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
1069 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
1073 case ICmpInst::ICMP_SLT:
1075 default: llvm_unreachable("Unknown integer condition code!");
1076 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
1078 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
1079 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
1081 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
1085 case ICmpInst::ICMP_UGT:
1087 default: llvm_unreachable("Unknown integer condition code!");
1088 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
1089 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
1091 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
1093 case ICmpInst::ICMP_NE:
1094 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
1095 return Builder->CreateICmp(LHSCC, Val, RHSCst);
1096 break; // (X u> 13 & X != 15) -> no change
1097 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
1098 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
1099 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
1103 case ICmpInst::ICMP_SGT:
1105 default: llvm_unreachable("Unknown integer condition code!");
1106 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
1107 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
1109 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
1111 case ICmpInst::ICMP_NE:
1112 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
1113 return Builder->CreateICmp(LHSCC, Val, RHSCst);
1114 break; // (X s> 13 & X != 15) -> no change
1115 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
1116 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true);
1117 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
1126 /// FoldAndOfFCmps - Optimize (fcmp)&(fcmp). NOTE: Unlike the rest of
1127 /// instcombine, this returns a Value which should already be inserted into the
1129 Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
1130 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
1131 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
1132 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType())
1135 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
1136 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
1137 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
1138 // If either of the constants are nans, then the whole thing returns
1140 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
1141 return Builder->getFalse();
1142 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
1145 // Handle vector zeros. This occurs because the canonical form of
1146 // "fcmp ord x,x" is "fcmp ord x, 0".
1147 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
1148 isa<ConstantAggregateZero>(RHS->getOperand(1)))
1149 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
1153 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
1154 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
1155 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
1158 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
1159 // Swap RHS operands to match LHS.
1160 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
1161 std::swap(Op1LHS, Op1RHS);
1164 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
1165 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
1167 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
1168 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
1169 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
1170 if (Op0CC == FCmpInst::FCMP_TRUE)
1172 if (Op1CC == FCmpInst::FCMP_TRUE)
1177 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
1178 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
1179 // uno && ord -> false
1180 if (Op0Pred == 0 && Op1Pred == 0 && Op0Ordered != Op1Ordered)
1181 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
1183 std::swap(LHS, RHS);
1184 std::swap(Op0Pred, Op1Pred);
1185 std::swap(Op0Ordered, Op1Ordered);
1188 // uno && ueq -> uno && (uno || eq) -> uno
1189 // ord && olt -> ord && (ord && lt) -> olt
1190 if (!Op0Ordered && (Op0Ordered == Op1Ordered))
1192 if (Op0Ordered && (Op0Ordered == Op1Ordered))
1195 // uno && oeq -> uno && (ord && eq) -> false
1197 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
1198 // ord && ueq -> ord && (uno || eq) -> oeq
1199 return getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS, Builder);
1206 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
1207 bool Changed = SimplifyAssociativeOrCommutative(I);
1208 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1210 if (Value *V = SimplifyVectorOp(I))
1211 return ReplaceInstUsesWith(I, V);
1213 if (Value *V = SimplifyAndInst(Op0, Op1, DL, TLI, DT, AC))
1214 return ReplaceInstUsesWith(I, V);
1216 // (A|B)&(A|C) -> A|(B&C) etc
1217 if (Value *V = SimplifyUsingDistributiveLaws(I))
1218 return ReplaceInstUsesWith(I, V);
1220 // See if we can simplify any instructions used by the instruction whose sole
1221 // purpose is to compute bits we don't care about.
1222 if (SimplifyDemandedInstructionBits(I))
1225 if (Value *V = SimplifyBSwap(I))
1226 return ReplaceInstUsesWith(I, V);
1228 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
1229 const APInt &AndRHSMask = AndRHS->getValue();
1231 // Optimize a variety of ((val OP C1) & C2) combinations...
1232 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1233 Value *Op0LHS = Op0I->getOperand(0);
1234 Value *Op0RHS = Op0I->getOperand(1);
1235 switch (Op0I->getOpcode()) {
1237 case Instruction::Xor:
1238 case Instruction::Or: {
1239 // If the mask is only needed on one incoming arm, push it up.
1240 if (!Op0I->hasOneUse()) break;
1242 APInt NotAndRHS(~AndRHSMask);
1243 if (MaskedValueIsZero(Op0LHS, NotAndRHS, 0, &I)) {
1244 // Not masking anything out for the LHS, move to RHS.
1245 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
1246 Op0RHS->getName()+".masked");
1247 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
1249 if (!isa<Constant>(Op0RHS) &&
1250 MaskedValueIsZero(Op0RHS, NotAndRHS, 0, &I)) {
1251 // Not masking anything out for the RHS, move to LHS.
1252 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
1253 Op0LHS->getName()+".masked");
1254 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
1259 case Instruction::Add:
1260 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
1261 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1262 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1263 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
1264 return BinaryOperator::CreateAnd(V, AndRHS);
1265 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
1266 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
1269 case Instruction::Sub:
1270 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
1271 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1272 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1273 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
1274 return BinaryOperator::CreateAnd(V, AndRHS);
1277 if (AndRHSMask == 1 && match(Op0LHS, m_Zero()))
1278 return BinaryOperator::CreateAnd(Op0RHS, AndRHS);
1280 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
1281 // has 1's for all bits that the subtraction with A might affect.
1282 if (Op0I->hasOneUse() && !match(Op0LHS, m_Zero())) {
1283 uint32_t BitWidth = AndRHSMask.getBitWidth();
1284 uint32_t Zeros = AndRHSMask.countLeadingZeros();
1285 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
1287 if (MaskedValueIsZero(Op0LHS, Mask, 0, &I)) {
1288 Value *NewNeg = Builder->CreateNeg(Op0RHS);
1289 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
1294 case Instruction::Shl:
1295 case Instruction::LShr:
1296 // (1 << x) & 1 --> zext(x == 0)
1297 // (1 >> x) & 1 --> zext(x == 0)
1298 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
1300 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
1301 return new ZExtInst(NewICmp, I.getType());
1306 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
1307 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
1311 // If this is an integer truncation, and if the source is an 'and' with
1312 // immediate, transform it. This frequently occurs for bitfield accesses.
1314 Value *X = nullptr; ConstantInt *YC = nullptr;
1315 if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) {
1316 // Change: and (trunc (and X, YC) to T), C2
1317 // into : and (trunc X to T), trunc(YC) & C2
1318 // This will fold the two constants together, which may allow
1319 // other simplifications.
1320 Value *NewCast = Builder->CreateTrunc(X, I.getType(), "and.shrunk");
1321 Constant *C3 = ConstantExpr::getTrunc(YC, I.getType());
1322 C3 = ConstantExpr::getAnd(C3, AndRHS);
1323 return BinaryOperator::CreateAnd(NewCast, C3);
1327 // Try to fold constant and into select arguments.
1328 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1329 if (Instruction *R = FoldOpIntoSelect(I, SI))
1331 if (isa<PHINode>(Op0))
1332 if (Instruction *NV = FoldOpIntoPhi(I))
1337 // (~A & ~B) == (~(A | B)) - De Morgan's Law
1338 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1339 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1340 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1341 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
1342 I.getName()+".demorgan");
1343 return BinaryOperator::CreateNot(Or);
1347 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
1348 // (A|B) & ~(A&B) -> A^B
1349 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1350 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1351 ((A == C && B == D) || (A == D && B == C)))
1352 return BinaryOperator::CreateXor(A, B);
1354 // ~(A&B) & (A|B) -> A^B
1355 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1356 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1357 ((A == C && B == D) || (A == D && B == C)))
1358 return BinaryOperator::CreateXor(A, B);
1360 // A&(A^B) => A & ~B
1362 Value *tmpOp0 = Op0;
1363 Value *tmpOp1 = Op1;
1364 if (Op0->hasOneUse() &&
1365 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
1366 if (A == Op1 || B == Op1 ) {
1373 if (tmpOp1->hasOneUse() &&
1374 match(tmpOp1, m_Xor(m_Value(A), m_Value(B)))) {
1378 // Notice that the patten (A&(~B)) is actually (A&(-1^B)), so if
1379 // A is originally -1 (or a vector of -1 and undefs), then we enter
1380 // an endless loop. By checking that A is non-constant we ensure that
1381 // we will never get to the loop.
1382 if (A == tmpOp0 && !isa<Constant>(A)) // A&(A^B) -> A & ~B
1383 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B));
1387 // (A&((~A)|B)) -> A&B
1388 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
1389 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
1390 return BinaryOperator::CreateAnd(A, Op1);
1391 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
1392 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
1393 return BinaryOperator::CreateAnd(A, Op0);
1395 // (A ^ B) & ((B ^ C) ^ A) -> (A ^ B) & ~C
1396 if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
1397 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
1398 if (Op1->hasOneUse() || cast<BinaryOperator>(Op1)->hasOneUse())
1399 return BinaryOperator::CreateAnd(Op0, Builder->CreateNot(C));
1401 // ((A ^ C) ^ B) & (B ^ A) -> (B ^ A) & ~C
1402 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
1403 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
1404 if (Op0->hasOneUse() || cast<BinaryOperator>(Op0)->hasOneUse())
1405 return BinaryOperator::CreateAnd(Op1, Builder->CreateNot(C));
1407 // (A | B) & ((~A) ^ B) -> (A & B)
1408 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1409 match(Op1, m_Xor(m_Not(m_Specific(A)), m_Specific(B))))
1410 return BinaryOperator::CreateAnd(A, B);
1412 // ((~A) ^ B) & (A | B) -> (A & B)
1413 if (match(Op0, m_Xor(m_Not(m_Value(A)), m_Value(B))) &&
1414 match(Op1, m_Or(m_Specific(A), m_Specific(B))))
1415 return BinaryOperator::CreateAnd(A, B);
1419 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
1420 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
1422 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1423 return ReplaceInstUsesWith(I, Res);
1425 // TODO: Make this recursive; it's a little tricky because an arbitrary
1426 // number of 'and' instructions might have to be created.
1428 if (LHS && match(Op1, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
1429 if (auto *Cmp = dyn_cast<ICmpInst>(X))
1430 if (Value *Res = FoldAndOfICmps(LHS, Cmp))
1431 return ReplaceInstUsesWith(I, Builder->CreateAnd(Res, Y));
1432 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
1433 if (Value *Res = FoldAndOfICmps(LHS, Cmp))
1434 return ReplaceInstUsesWith(I, Builder->CreateAnd(Res, X));
1436 if (RHS && match(Op0, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
1437 if (auto *Cmp = dyn_cast<ICmpInst>(X))
1438 if (Value *Res = FoldAndOfICmps(Cmp, RHS))
1439 return ReplaceInstUsesWith(I, Builder->CreateAnd(Res, Y));
1440 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
1441 if (Value *Res = FoldAndOfICmps(Cmp, RHS))
1442 return ReplaceInstUsesWith(I, Builder->CreateAnd(Res, X));
1446 // If and'ing two fcmp, try combine them into one.
1447 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1448 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1449 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1450 return ReplaceInstUsesWith(I, Res);
1453 // fold (and (cast A), (cast B)) -> (cast (and A, B))
1454 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
1455 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) {
1456 Type *SrcTy = Op0C->getOperand(0)->getType();
1457 if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ?
1458 SrcTy == Op1C->getOperand(0)->getType() &&
1459 SrcTy->isIntOrIntVectorTy()) {
1460 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
1462 // Only do this if the casts both really cause code to be generated.
1463 if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
1464 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
1465 Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName());
1466 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
1469 // If this is and(cast(icmp), cast(icmp)), try to fold this even if the
1470 // cast is otherwise not optimizable. This happens for vector sexts.
1471 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
1472 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
1473 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1474 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1476 // If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the
1477 // cast is otherwise not optimizable. This happens for vector sexts.
1478 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
1479 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
1480 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1481 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1487 bool OpsSwapped = false;
1488 // Canonicalize SExt or Not to the LHS
1489 if (match(Op1, m_SExt(m_Value())) ||
1490 match(Op1, m_Not(m_Value()))) {
1491 std::swap(Op0, Op1);
1495 // Fold (and (sext bool to A), B) --> (select bool, B, 0)
1496 if (match(Op0, m_SExt(m_Value(X))) &&
1497 X->getType()->getScalarType()->isIntegerTy(1)) {
1498 Value *Zero = Constant::getNullValue(Op1->getType());
1499 return SelectInst::Create(X, Op1, Zero);
1502 // Fold (and ~(sext bool to A), B) --> (select bool, 0, B)
1503 if (match(Op0, m_Not(m_SExt(m_Value(X)))) &&
1504 X->getType()->getScalarType()->isIntegerTy(1)) {
1505 Value *Zero = Constant::getNullValue(Op0->getType());
1506 return SelectInst::Create(X, Zero, Op1);
1510 std::swap(Op0, Op1);
1513 return Changed ? &I : nullptr;
1516 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
1517 /// capable of providing pieces of a bswap. The subexpression provides pieces
1518 /// of a bswap if it is proven that each of the non-zero bytes in the output of
1519 /// the expression came from the corresponding "byte swapped" byte in some other
1520 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
1521 /// we know that the expression deposits the low byte of %X into the high byte
1522 /// of the bswap result and that all other bytes are zero. This expression is
1523 /// accepted, the high byte of ByteValues is set to X to indicate a correct
1526 /// This function returns true if the match was unsuccessful and false if so.
1527 /// On entry to the function the "OverallLeftShift" is a signed integer value
1528 /// indicating the number of bytes that the subexpression is later shifted. For
1529 /// example, if the expression is later right shifted by 16 bits, the
1530 /// OverallLeftShift value would be -2 on entry. This is used to specify which
1531 /// byte of ByteValues is actually being set.
1533 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
1534 /// byte is masked to zero by a user. For example, in (X & 255), X will be
1535 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
1536 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
1537 /// always in the local (OverallLeftShift) coordinate space.
1539 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
1540 SmallVectorImpl<Value *> &ByteValues) {
1541 if (Instruction *I = dyn_cast<Instruction>(V)) {
1542 // If this is an or instruction, it may be an inner node of the bswap.
1543 if (I->getOpcode() == Instruction::Or) {
1544 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1546 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
1550 // If this is a logical shift by a constant multiple of 8, recurse with
1551 // OverallLeftShift and ByteMask adjusted.
1552 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
1554 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
1555 // Ensure the shift amount is defined and of a byte value.
1556 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
1559 unsigned ByteShift = ShAmt >> 3;
1560 if (I->getOpcode() == Instruction::Shl) {
1561 // X << 2 -> collect(X, +2)
1562 OverallLeftShift += ByteShift;
1563 ByteMask >>= ByteShift;
1565 // X >>u 2 -> collect(X, -2)
1566 OverallLeftShift -= ByteShift;
1567 ByteMask <<= ByteShift;
1568 ByteMask &= (~0U >> (32-ByteValues.size()));
1571 if (OverallLeftShift >= (int)ByteValues.size()) return true;
1572 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
1574 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1578 // If this is a logical 'and' with a mask that clears bytes, clear the
1579 // corresponding bytes in ByteMask.
1580 if (I->getOpcode() == Instruction::And &&
1581 isa<ConstantInt>(I->getOperand(1))) {
1582 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
1583 unsigned NumBytes = ByteValues.size();
1584 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
1585 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
1587 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
1588 // If this byte is masked out by a later operation, we don't care what
1590 if ((ByteMask & (1 << i)) == 0)
1593 // If the AndMask is all zeros for this byte, clear the bit.
1594 APInt MaskB = AndMask & Byte;
1596 ByteMask &= ~(1U << i);
1600 // If the AndMask is not all ones for this byte, it's not a bytezap.
1604 // Otherwise, this byte is kept.
1607 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1612 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
1613 // the input value to the bswap. Some observations: 1) if more than one byte
1614 // is demanded from this input, then it could not be successfully assembled
1615 // into a byteswap. At least one of the two bytes would not be aligned with
1616 // their ultimate destination.
1617 if (!isPowerOf2_32(ByteMask)) return true;
1618 unsigned InputByteNo = countTrailingZeros(ByteMask);
1620 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
1621 // is demanded, it needs to go into byte 0 of the result. This means that the
1622 // byte needs to be shifted until it lands in the right byte bucket. The
1623 // shift amount depends on the position: if the byte is coming from the high
1624 // part of the value (e.g. byte 3) then it must be shifted right. If from the
1625 // low part, it must be shifted left.
1626 unsigned DestByteNo = InputByteNo + OverallLeftShift;
1627 if (ByteValues.size()-1-DestByteNo != InputByteNo)
1630 // If the destination byte value is already defined, the values are or'd
1631 // together, which isn't a bswap (unless it's an or of the same bits).
1632 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
1634 ByteValues[DestByteNo] = V;
1638 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
1639 /// If so, insert the new bswap intrinsic and return it.
1640 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
1641 IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
1642 if (!ITy || ITy->getBitWidth() % 16 ||
1643 // ByteMask only allows up to 32-byte values.
1644 ITy->getBitWidth() > 32*8)
1645 return nullptr; // Can only bswap pairs of bytes. Can't do vectors.
1647 /// ByteValues - For each byte of the result, we keep track of which value
1648 /// defines each byte.
1649 SmallVector<Value*, 8> ByteValues;
1650 ByteValues.resize(ITy->getBitWidth()/8);
1652 // Try to find all the pieces corresponding to the bswap.
1653 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
1654 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
1657 // Check to see if all of the bytes come from the same value.
1658 Value *V = ByteValues[0];
1659 if (!V) return nullptr; // Didn't find a byte? Must be zero.
1661 // Check to make sure that all of the bytes come from the same value.
1662 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
1663 if (ByteValues[i] != V)
1665 Module *M = I.getParent()->getParent()->getParent();
1666 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, ITy);
1667 return CallInst::Create(F, V);
1670 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
1671 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
1672 /// we can simplify this expression to "cond ? C : D or B".
1673 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
1674 Value *C, Value *D) {
1675 // If A is not a select of -1/0, this cannot match.
1676 Value *Cond = nullptr;
1677 if (!match(A, m_SExt(m_Value(Cond))) ||
1678 !Cond->getType()->isIntegerTy(1))
1681 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
1682 if (match(D, m_Not(m_SExt(m_Specific(Cond)))))
1683 return SelectInst::Create(Cond, C, B);
1684 if (match(D, m_SExt(m_Not(m_Specific(Cond)))))
1685 return SelectInst::Create(Cond, C, B);
1687 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
1688 if (match(B, m_Not(m_SExt(m_Specific(Cond)))))
1689 return SelectInst::Create(Cond, C, D);
1690 if (match(B, m_SExt(m_Not(m_Specific(Cond)))))
1691 return SelectInst::Create(Cond, C, D);
1695 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
1696 Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS,
1697 Instruction *CxtI) {
1698 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
1700 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2)
1701 // if K1 and K2 are a one-bit mask.
1702 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
1703 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
1705 if (LHS->getPredicate() == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero() &&
1706 RHS->getPredicate() == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) {
1708 BinaryOperator *LAnd = dyn_cast<BinaryOperator>(LHS->getOperand(0));
1709 BinaryOperator *RAnd = dyn_cast<BinaryOperator>(RHS->getOperand(0));
1710 if (LAnd && RAnd && LAnd->hasOneUse() && RHS->hasOneUse() &&
1711 LAnd->getOpcode() == Instruction::And &&
1712 RAnd->getOpcode() == Instruction::And) {
1714 Value *Mask = nullptr;
1715 Value *Masked = nullptr;
1716 if (LAnd->getOperand(0) == RAnd->getOperand(0) &&
1717 isKnownToBeAPowerOfTwo(LAnd->getOperand(1), DL, false, 0, AC, CxtI,
1719 isKnownToBeAPowerOfTwo(RAnd->getOperand(1), DL, false, 0, AC, CxtI,
1721 Mask = Builder->CreateOr(LAnd->getOperand(1), RAnd->getOperand(1));
1722 Masked = Builder->CreateAnd(LAnd->getOperand(0), Mask);
1723 } else if (LAnd->getOperand(1) == RAnd->getOperand(1) &&
1724 isKnownToBeAPowerOfTwo(LAnd->getOperand(0), DL, false, 0, AC,
1726 isKnownToBeAPowerOfTwo(RAnd->getOperand(0), DL, false, 0, AC,
1728 Mask = Builder->CreateOr(LAnd->getOperand(0), RAnd->getOperand(0));
1729 Masked = Builder->CreateAnd(LAnd->getOperand(1), Mask);
1733 return Builder->CreateICmp(ICmpInst::ICMP_NE, Masked, Mask);
1737 // Fold (icmp ult/ule (A + C1), C3) | (icmp ult/ule (A + C2), C3)
1738 // --> (icmp ult/ule ((A & ~(C1 ^ C2)) + max(C1, C2)), C3)
1739 // The original condition actually refers to the following two ranges:
1740 // [MAX_UINT-C1+1, MAX_UINT-C1+1+C3] and [MAX_UINT-C2+1, MAX_UINT-C2+1+C3]
1741 // We can fold these two ranges if:
1742 // 1) C1 and C2 is unsigned greater than C3.
1743 // 2) The two ranges are separated.
1744 // 3) C1 ^ C2 is one-bit mask.
1745 // 4) LowRange1 ^ LowRange2 and HighRange1 ^ HighRange2 are one-bit mask.
1746 // This implies all values in the two ranges differ by exactly one bit.
1748 if ((LHSCC == ICmpInst::ICMP_ULT || LHSCC == ICmpInst::ICMP_ULE) &&
1749 LHSCC == RHSCC && LHSCst && RHSCst && LHS->hasOneUse() &&
1750 RHS->hasOneUse() && LHSCst->getType() == RHSCst->getType() &&
1751 LHSCst->getValue() == (RHSCst->getValue())) {
1753 Value *LAdd = LHS->getOperand(0);
1754 Value *RAdd = RHS->getOperand(0);
1756 Value *LAddOpnd, *RAddOpnd;
1757 ConstantInt *LAddCst, *RAddCst;
1758 if (match(LAdd, m_Add(m_Value(LAddOpnd), m_ConstantInt(LAddCst))) &&
1759 match(RAdd, m_Add(m_Value(RAddOpnd), m_ConstantInt(RAddCst))) &&
1760 LAddCst->getValue().ugt(LHSCst->getValue()) &&
1761 RAddCst->getValue().ugt(LHSCst->getValue())) {
1763 APInt DiffCst = LAddCst->getValue() ^ RAddCst->getValue();
1764 if (LAddOpnd == RAddOpnd && DiffCst.isPowerOf2()) {
1765 ConstantInt *MaxAddCst = nullptr;
1766 if (LAddCst->getValue().ult(RAddCst->getValue()))
1767 MaxAddCst = RAddCst;
1769 MaxAddCst = LAddCst;
1771 APInt RRangeLow = -RAddCst->getValue();
1772 APInt RRangeHigh = RRangeLow + LHSCst->getValue();
1773 APInt LRangeLow = -LAddCst->getValue();
1774 APInt LRangeHigh = LRangeLow + LHSCst->getValue();
1775 APInt LowRangeDiff = RRangeLow ^ LRangeLow;
1776 APInt HighRangeDiff = RRangeHigh ^ LRangeHigh;
1777 APInt RangeDiff = LRangeLow.sgt(RRangeLow) ? LRangeLow - RRangeLow
1778 : RRangeLow - LRangeLow;
1780 if (LowRangeDiff.isPowerOf2() && LowRangeDiff == HighRangeDiff &&
1781 RangeDiff.ugt(LHSCst->getValue())) {
1782 Value *MaskCst = ConstantInt::get(LAddCst->getType(), ~DiffCst);
1784 Value *NewAnd = Builder->CreateAnd(LAddOpnd, MaskCst);
1785 Value *NewAdd = Builder->CreateAdd(NewAnd, MaxAddCst);
1786 return (Builder->CreateICmp(LHS->getPredicate(), NewAdd, LHSCst));
1792 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
1793 if (PredicatesFoldable(LHSCC, RHSCC)) {
1794 if (LHS->getOperand(0) == RHS->getOperand(1) &&
1795 LHS->getOperand(1) == RHS->getOperand(0))
1796 LHS->swapOperands();
1797 if (LHS->getOperand(0) == RHS->getOperand(0) &&
1798 LHS->getOperand(1) == RHS->getOperand(1)) {
1799 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1800 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
1801 bool isSigned = LHS->isSigned() || RHS->isSigned();
1802 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
1806 // handle (roughly):
1807 // (icmp ne (A & B), C) | (icmp ne (A & D), E)
1808 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, false, Builder))
1811 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
1812 if (LHS->hasOneUse() || RHS->hasOneUse()) {
1813 // (icmp eq B, 0) | (icmp ult A, B) -> (icmp ule A, B-1)
1814 // (icmp eq B, 0) | (icmp ugt B, A) -> (icmp ule A, B-1)
1815 Value *A = nullptr, *B = nullptr;
1816 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero()) {
1818 if (RHSCC == ICmpInst::ICMP_ULT && Val == RHS->getOperand(1))
1820 else if (RHSCC == ICmpInst::ICMP_UGT && Val == Val2)
1821 A = RHS->getOperand(1);
1823 // (icmp ult A, B) | (icmp eq B, 0) -> (icmp ule A, B-1)
1824 // (icmp ugt B, A) | (icmp eq B, 0) -> (icmp ule A, B-1)
1825 else if (RHSCC == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) {
1827 if (LHSCC == ICmpInst::ICMP_ULT && Val2 == LHS->getOperand(1))
1829 else if (LHSCC == ICmpInst::ICMP_UGT && Val2 == Val)
1830 A = LHS->getOperand(1);
1833 return Builder->CreateICmp(
1835 Builder->CreateAdd(B, ConstantInt::getSigned(B->getType(), -1)), A);
1838 // E.g. (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
1839 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/true))
1842 // E.g. (icmp sgt x, n) | (icmp slt x, 0) --> icmp ugt x, n
1843 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/true))
1846 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
1847 if (!LHSCst || !RHSCst) return nullptr;
1849 if (LHSCst == RHSCst && LHSCC == RHSCC) {
1850 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
1851 if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
1852 Value *NewOr = Builder->CreateOr(Val, Val2);
1853 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
1857 // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
1858 // iff C2 + CA == C1.
1859 if (LHSCC == ICmpInst::ICMP_ULT && RHSCC == ICmpInst::ICMP_EQ) {
1860 ConstantInt *AddCst;
1861 if (match(Val, m_Add(m_Specific(Val2), m_ConstantInt(AddCst))))
1862 if (RHSCst->getValue() + AddCst->getValue() == LHSCst->getValue())
1863 return Builder->CreateICmpULE(Val, LHSCst);
1866 // From here on, we only handle:
1867 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
1868 if (Val != Val2) return nullptr;
1870 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
1871 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
1872 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
1873 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
1874 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
1877 // We can't fold (ugt x, C) | (sgt x, C2).
1878 if (!PredicatesFoldable(LHSCC, RHSCC))
1881 // Ensure that the larger constant is on the RHS.
1883 if (CmpInst::isSigned(LHSCC) ||
1884 (ICmpInst::isEquality(LHSCC) &&
1885 CmpInst::isSigned(RHSCC)))
1886 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
1888 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
1891 std::swap(LHS, RHS);
1892 std::swap(LHSCst, RHSCst);
1893 std::swap(LHSCC, RHSCC);
1896 // At this point, we know we have two icmp instructions
1897 // comparing a value against two constants and or'ing the result
1898 // together. Because of the above check, we know that we only have
1899 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
1900 // icmp folding check above), that the two constants are not
1902 assert(LHSCst != RHSCst && "Compares not folded above?");
1905 default: llvm_unreachable("Unknown integer condition code!");
1906 case ICmpInst::ICMP_EQ:
1908 default: llvm_unreachable("Unknown integer condition code!");
1909 case ICmpInst::ICMP_EQ:
1910 if (LHS->getOperand(0) == RHS->getOperand(0)) {
1911 // if LHSCst and RHSCst differ only by one bit:
1912 // (A == C1 || A == C2) -> (A & ~(C1 ^ C2)) == C1
1913 assert(LHSCst->getValue().ule(LHSCst->getValue()));
1915 APInt Xor = LHSCst->getValue() ^ RHSCst->getValue();
1916 if (Xor.isPowerOf2()) {
1917 Value *NegCst = Builder->getInt(~Xor);
1918 Value *And = Builder->CreateAnd(LHS->getOperand(0), NegCst);
1919 return Builder->CreateICmp(ICmpInst::ICMP_EQ, And, LHSCst);
1923 if (LHSCst == SubOne(RHSCst)) {
1924 // (X == 13 | X == 14) -> X-13 <u 2
1925 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1926 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1927 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
1928 return Builder->CreateICmpULT(Add, AddCST);
1931 break; // (X == 13 | X == 15) -> no change
1932 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
1933 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
1935 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
1936 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
1937 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
1941 case ICmpInst::ICMP_NE:
1943 default: llvm_unreachable("Unknown integer condition code!");
1944 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
1945 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
1946 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
1948 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
1949 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
1950 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
1951 return Builder->getTrue();
1953 case ICmpInst::ICMP_ULT:
1955 default: llvm_unreachable("Unknown integer condition code!");
1956 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
1958 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
1959 // If RHSCst is [us]MAXINT, it is always false. Not handling
1960 // this can cause overflow.
1961 if (RHSCst->isMaxValue(false))
1963 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false);
1964 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
1966 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
1967 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
1969 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
1973 case ICmpInst::ICMP_SLT:
1975 default: llvm_unreachable("Unknown integer condition code!");
1976 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
1978 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
1979 // If RHSCst is [us]MAXINT, it is always false. Not handling
1980 // this can cause overflow.
1981 if (RHSCst->isMaxValue(true))
1983 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false);
1984 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
1986 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
1987 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
1989 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
1993 case ICmpInst::ICMP_UGT:
1995 default: llvm_unreachable("Unknown integer condition code!");
1996 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
1997 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
1999 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
2001 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
2002 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
2003 return Builder->getTrue();
2004 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
2008 case ICmpInst::ICMP_SGT:
2010 default: llvm_unreachable("Unknown integer condition code!");
2011 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
2012 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
2014 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
2016 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
2017 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
2018 return Builder->getTrue();
2019 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
2027 /// FoldOrOfFCmps - Optimize (fcmp)|(fcmp). NOTE: Unlike the rest of
2028 /// instcombine, this returns a Value which should already be inserted into the
2030 Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
2031 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
2032 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
2033 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
2034 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
2035 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
2036 // If either of the constants are nans, then the whole thing returns
2038 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
2039 return Builder->getTrue();
2041 // Otherwise, no need to compare the two constants, compare the
2043 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
2046 // Handle vector zeros. This occurs because the canonical form of
2047 // "fcmp uno x,x" is "fcmp uno x, 0".
2048 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
2049 isa<ConstantAggregateZero>(RHS->getOperand(1)))
2050 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
2055 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
2056 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
2057 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
2059 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
2060 // Swap RHS operands to match LHS.
2061 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
2062 std::swap(Op1LHS, Op1RHS);
2064 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
2065 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
2067 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
2068 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
2069 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
2070 if (Op0CC == FCmpInst::FCMP_FALSE)
2072 if (Op1CC == FCmpInst::FCMP_FALSE)
2076 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
2077 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
2078 if (Op0Ordered == Op1Ordered) {
2079 // If both are ordered or unordered, return a new fcmp with
2080 // or'ed predicates.
2081 return getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS, Builder);
2087 /// FoldOrWithConstants - This helper function folds:
2089 /// ((A | B) & C1) | (B & C2)
2095 /// when the XOR of the two constants is "all ones" (-1).
2096 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
2097 Value *A, Value *B, Value *C) {
2098 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
2099 if (!CI1) return nullptr;
2101 Value *V1 = nullptr;
2102 ConstantInt *CI2 = nullptr;
2103 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return nullptr;
2105 APInt Xor = CI1->getValue() ^ CI2->getValue();
2106 if (!Xor.isAllOnesValue()) return nullptr;
2108 if (V1 == A || V1 == B) {
2109 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
2110 return BinaryOperator::CreateOr(NewOp, V1);
2116 /// \brief This helper function folds:
2118 /// ((A | B) & C1) ^ (B & C2)
2124 /// when the XOR of the two constants is "all ones" (-1).
2125 Instruction *InstCombiner::FoldXorWithConstants(BinaryOperator &I, Value *Op,
2126 Value *A, Value *B, Value *C) {
2127 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
2131 Value *V1 = nullptr;
2132 ConstantInt *CI2 = nullptr;
2133 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2))))
2136 APInt Xor = CI1->getValue() ^ CI2->getValue();
2137 if (!Xor.isAllOnesValue())
2140 if (V1 == A || V1 == B) {
2141 Value *NewOp = Builder->CreateAnd(V1 == A ? B : A, CI1);
2142 return BinaryOperator::CreateXor(NewOp, V1);
2148 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
2149 bool Changed = SimplifyAssociativeOrCommutative(I);
2150 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2152 if (Value *V = SimplifyVectorOp(I))
2153 return ReplaceInstUsesWith(I, V);
2155 if (Value *V = SimplifyOrInst(Op0, Op1, DL, TLI, DT, AC))
2156 return ReplaceInstUsesWith(I, V);
2158 // (A&B)|(A&C) -> A&(B|C) etc
2159 if (Value *V = SimplifyUsingDistributiveLaws(I))
2160 return ReplaceInstUsesWith(I, V);
2162 // See if we can simplify any instructions used by the instruction whose sole
2163 // purpose is to compute bits we don't care about.
2164 if (SimplifyDemandedInstructionBits(I))
2167 if (Value *V = SimplifyBSwap(I))
2168 return ReplaceInstUsesWith(I, V);
2170 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2171 ConstantInt *C1 = nullptr; Value *X = nullptr;
2172 // (X & C1) | C2 --> (X | C2) & (C1|C2)
2173 // iff (C1 & C2) == 0.
2174 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
2175 (RHS->getValue() & C1->getValue()) != 0 &&
2177 Value *Or = Builder->CreateOr(X, RHS);
2179 return BinaryOperator::CreateAnd(Or,
2180 Builder->getInt(RHS->getValue() | C1->getValue()));
2183 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
2184 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
2186 Value *Or = Builder->CreateOr(X, RHS);
2188 return BinaryOperator::CreateXor(Or,
2189 Builder->getInt(C1->getValue() & ~RHS->getValue()));
2192 // Try to fold constant and into select arguments.
2193 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2194 if (Instruction *R = FoldOpIntoSelect(I, SI))
2197 if (isa<PHINode>(Op0))
2198 if (Instruction *NV = FoldOpIntoPhi(I))
2202 Value *A = nullptr, *B = nullptr;
2203 ConstantInt *C1 = nullptr, *C2 = nullptr;
2205 // (A | B) | C and A | (B | C) -> bswap if possible.
2206 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
2207 if (match(Op0, m_Or(m_Value(), m_Value())) ||
2208 match(Op1, m_Or(m_Value(), m_Value())) ||
2209 (match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
2210 match(Op1, m_LogicalShift(m_Value(), m_Value())))) {
2211 if (Instruction *BSwap = MatchBSwap(I))
2215 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
2216 if (Op0->hasOneUse() &&
2217 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2218 MaskedValueIsZero(Op1, C1->getValue(), 0, &I)) {
2219 Value *NOr = Builder->CreateOr(A, Op1);
2221 return BinaryOperator::CreateXor(NOr, C1);
2224 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
2225 if (Op1->hasOneUse() &&
2226 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2227 MaskedValueIsZero(Op0, C1->getValue(), 0, &I)) {
2228 Value *NOr = Builder->CreateOr(A, Op0);
2230 return BinaryOperator::CreateXor(NOr, C1);
2233 // ((~A & B) | A) -> (A | B)
2234 if (match(Op0, m_And(m_Not(m_Value(A)), m_Value(B))) &&
2235 match(Op1, m_Specific(A)))
2236 return BinaryOperator::CreateOr(A, B);
2238 // ((A & B) | ~A) -> (~A | B)
2239 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2240 match(Op1, m_Not(m_Specific(A))))
2241 return BinaryOperator::CreateOr(Builder->CreateNot(A), B);
2243 // (A & (~B)) | (A ^ B) -> (A ^ B)
2244 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2245 match(Op1, m_Xor(m_Specific(A), m_Specific(B))))
2246 return BinaryOperator::CreateXor(A, B);
2248 // (A ^ B) | ( A & (~B)) -> (A ^ B)
2249 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
2250 match(Op1, m_And(m_Specific(A), m_Not(m_Specific(B)))))
2251 return BinaryOperator::CreateXor(A, B);
2254 Value *C = nullptr, *D = nullptr;
2255 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
2256 match(Op1, m_And(m_Value(B), m_Value(D)))) {
2257 Value *V1 = nullptr, *V2 = nullptr;
2258 C1 = dyn_cast<ConstantInt>(C);
2259 C2 = dyn_cast<ConstantInt>(D);
2260 if (C1 && C2) { // (A & C1)|(B & C2)
2261 if ((C1->getValue() & C2->getValue()) == 0) {
2262 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
2263 // iff (C1&C2) == 0 and (N&~C1) == 0
2264 if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
2266 MaskedValueIsZero(V2, ~C1->getValue(), 0, &I)) || // (V|N)
2268 MaskedValueIsZero(V1, ~C1->getValue(), 0, &I)))) // (N|V)
2269 return BinaryOperator::CreateAnd(A,
2270 Builder->getInt(C1->getValue()|C2->getValue()));
2271 // Or commutes, try both ways.
2272 if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
2274 MaskedValueIsZero(V2, ~C2->getValue(), 0, &I)) || // (V|N)
2276 MaskedValueIsZero(V1, ~C2->getValue(), 0, &I)))) // (N|V)
2277 return BinaryOperator::CreateAnd(B,
2278 Builder->getInt(C1->getValue()|C2->getValue()));
2280 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
2281 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
2282 ConstantInt *C3 = nullptr, *C4 = nullptr;
2283 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
2284 (C3->getValue() & ~C1->getValue()) == 0 &&
2285 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
2286 (C4->getValue() & ~C2->getValue()) == 0) {
2287 V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
2288 return BinaryOperator::CreateAnd(V2,
2289 Builder->getInt(C1->getValue()|C2->getValue()));
2294 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants.
2295 // Don't do this for vector select idioms, the code generator doesn't handle
2297 if (!I.getType()->isVectorTy()) {
2298 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
2300 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
2302 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
2304 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
2308 // ((A&~B)|(~A&B)) -> A^B
2309 if ((match(C, m_Not(m_Specific(D))) &&
2310 match(B, m_Not(m_Specific(A)))))
2311 return BinaryOperator::CreateXor(A, D);
2312 // ((~B&A)|(~A&B)) -> A^B
2313 if ((match(A, m_Not(m_Specific(D))) &&
2314 match(B, m_Not(m_Specific(C)))))
2315 return BinaryOperator::CreateXor(C, D);
2316 // ((A&~B)|(B&~A)) -> A^B
2317 if ((match(C, m_Not(m_Specific(B))) &&
2318 match(D, m_Not(m_Specific(A)))))
2319 return BinaryOperator::CreateXor(A, B);
2320 // ((~B&A)|(B&~A)) -> A^B
2321 if ((match(A, m_Not(m_Specific(B))) &&
2322 match(D, m_Not(m_Specific(C)))))
2323 return BinaryOperator::CreateXor(C, B);
2325 // ((A|B)&1)|(B&-2) -> (A&1) | B
2326 if (match(A, m_Or(m_Value(V1), m_Specific(B))) ||
2327 match(A, m_Or(m_Specific(B), m_Value(V1)))) {
2328 Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C);
2329 if (Ret) return Ret;
2331 // (B&-2)|((A|B)&1) -> (A&1) | B
2332 if (match(B, m_Or(m_Specific(A), m_Value(V1))) ||
2333 match(B, m_Or(m_Value(V1), m_Specific(A)))) {
2334 Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D);
2335 if (Ret) return Ret;
2337 // ((A^B)&1)|(B&-2) -> (A&1) ^ B
2338 if (match(A, m_Xor(m_Value(V1), m_Specific(B))) ||
2339 match(A, m_Xor(m_Specific(B), m_Value(V1)))) {
2340 Instruction *Ret = FoldXorWithConstants(I, Op1, V1, B, C);
2341 if (Ret) return Ret;
2343 // (B&-2)|((A^B)&1) -> (A&1) ^ B
2344 if (match(B, m_Xor(m_Specific(A), m_Value(V1))) ||
2345 match(B, m_Xor(m_Value(V1), m_Specific(A)))) {
2346 Instruction *Ret = FoldXorWithConstants(I, Op0, A, V1, D);
2347 if (Ret) return Ret;
2351 // (A ^ B) | ((B ^ C) ^ A) -> (A ^ B) | C
2352 if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
2353 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
2354 if (Op1->hasOneUse() || cast<BinaryOperator>(Op1)->hasOneUse())
2355 return BinaryOperator::CreateOr(Op0, C);
2357 // ((A ^ C) ^ B) | (B ^ A) -> (B ^ A) | C
2358 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
2359 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
2360 if (Op0->hasOneUse() || cast<BinaryOperator>(Op0)->hasOneUse())
2361 return BinaryOperator::CreateOr(Op1, C);
2363 // ((B | C) & A) | B -> B | (A & C)
2364 if (match(Op0, m_And(m_Or(m_Specific(Op1), m_Value(C)), m_Value(A))))
2365 return BinaryOperator::CreateOr(Op1, Builder->CreateAnd(A, C));
2367 // (~A | ~B) == (~(A & B)) - De Morgan's Law
2368 if (Value *Op0NotVal = dyn_castNotVal(Op0))
2369 if (Value *Op1NotVal = dyn_castNotVal(Op1))
2370 if (Op0->hasOneUse() && Op1->hasOneUse()) {
2371 Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
2372 I.getName()+".demorgan");
2373 return BinaryOperator::CreateNot(And);
2376 // Canonicalize xor to the RHS.
2377 bool SwappedForXor = false;
2378 if (match(Op0, m_Xor(m_Value(), m_Value()))) {
2379 std::swap(Op0, Op1);
2380 SwappedForXor = true;
2383 // A | ( A ^ B) -> A | B
2384 // A | (~A ^ B) -> A | ~B
2385 // (A & B) | (A ^ B)
2386 if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
2387 if (Op0 == A || Op0 == B)
2388 return BinaryOperator::CreateOr(A, B);
2390 if (match(Op0, m_And(m_Specific(A), m_Specific(B))) ||
2391 match(Op0, m_And(m_Specific(B), m_Specific(A))))
2392 return BinaryOperator::CreateOr(A, B);
2394 if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
2395 Value *Not = Builder->CreateNot(B, B->getName()+".not");
2396 return BinaryOperator::CreateOr(Not, Op0);
2398 if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
2399 Value *Not = Builder->CreateNot(A, A->getName()+".not");
2400 return BinaryOperator::CreateOr(Not, Op0);
2404 // A | ~(A | B) -> A | ~B
2405 // A | ~(A ^ B) -> A | ~B
2406 if (match(Op1, m_Not(m_Value(A))))
2407 if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
2408 if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
2409 Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
2410 B->getOpcode() == Instruction::Xor)) {
2411 Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
2413 Value *Not = Builder->CreateNot(NotOp, NotOp->getName()+".not");
2414 return BinaryOperator::CreateOr(Not, Op0);
2417 // (A & B) | ((~A) ^ B) -> (~A ^ B)
2418 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2419 match(Op1, m_Xor(m_Not(m_Specific(A)), m_Specific(B))))
2420 return BinaryOperator::CreateXor(Builder->CreateNot(A), B);
2422 // ((~A) ^ B) | (A & B) -> (~A ^ B)
2423 if (match(Op0, m_Xor(m_Not(m_Value(A)), m_Value(B))) &&
2424 match(Op1, m_And(m_Specific(A), m_Specific(B))))
2425 return BinaryOperator::CreateXor(Builder->CreateNot(A), B);
2428 std::swap(Op0, Op1);
2431 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
2432 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
2434 if (Value *Res = FoldOrOfICmps(LHS, RHS, &I))
2435 return ReplaceInstUsesWith(I, Res);
2437 // TODO: Make this recursive; it's a little tricky because an arbitrary
2438 // number of 'or' instructions might have to be created.
2440 if (LHS && match(Op1, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2441 if (auto *Cmp = dyn_cast<ICmpInst>(X))
2442 if (Value *Res = FoldOrOfICmps(LHS, Cmp, &I))
2443 return ReplaceInstUsesWith(I, Builder->CreateOr(Res, Y));
2444 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2445 if (Value *Res = FoldOrOfICmps(LHS, Cmp, &I))
2446 return ReplaceInstUsesWith(I, Builder->CreateOr(Res, X));
2448 if (RHS && match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2449 if (auto *Cmp = dyn_cast<ICmpInst>(X))
2450 if (Value *Res = FoldOrOfICmps(Cmp, RHS, &I))
2451 return ReplaceInstUsesWith(I, Builder->CreateOr(Res, Y));
2452 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2453 if (Value *Res = FoldOrOfICmps(Cmp, RHS, &I))
2454 return ReplaceInstUsesWith(I, Builder->CreateOr(Res, X));
2458 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
2459 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
2460 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
2461 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
2462 return ReplaceInstUsesWith(I, Res);
2464 // fold (or (cast A), (cast B)) -> (cast (or A, B))
2465 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2466 CastInst *Op1C = dyn_cast<CastInst>(Op1);
2467 if (Op1C && Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
2468 Type *SrcTy = Op0C->getOperand(0)->getType();
2469 if (SrcTy == Op1C->getOperand(0)->getType() &&
2470 SrcTy->isIntOrIntVectorTy()) {
2471 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
2473 if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) &&
2474 // Only do this if the casts both really cause code to be
2476 ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
2477 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
2478 Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName());
2479 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2482 // If this is or(cast(icmp), cast(icmp)), try to fold this even if the
2483 // cast is otherwise not optimizable. This happens for vector sexts.
2484 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
2485 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
2486 if (Value *Res = FoldOrOfICmps(LHS, RHS, &I))
2487 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
2489 // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the
2490 // cast is otherwise not optimizable. This happens for vector sexts.
2491 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
2492 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
2493 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
2494 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
2499 // or(sext(A), B) -> A ? -1 : B where A is an i1
2500 // or(A, sext(B)) -> B ? -1 : A where B is an i1
2501 if (match(Op0, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
2502 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1);
2503 if (match(Op1, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
2504 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0);
2506 // Note: If we've gotten to the point of visiting the outer OR, then the
2507 // inner one couldn't be simplified. If it was a constant, then it won't
2508 // be simplified by a later pass either, so we try swapping the inner/outer
2509 // ORs in the hopes that we'll be able to simplify it this way.
2510 // (X|C) | V --> (X|V) | C
2511 if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
2512 match(Op0, m_Or(m_Value(A), m_ConstantInt(C1)))) {
2513 Value *Inner = Builder->CreateOr(A, Op1);
2514 Inner->takeName(Op0);
2515 return BinaryOperator::CreateOr(Inner, C1);
2518 // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D))
2519 // Since this OR statement hasn't been optimized further yet, we hope
2520 // that this transformation will allow the new ORs to be optimized.
2522 Value *X = nullptr, *Y = nullptr;
2523 if (Op0->hasOneUse() && Op1->hasOneUse() &&
2524 match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) &&
2525 match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) {
2526 Value *orTrue = Builder->CreateOr(A, C);
2527 Value *orFalse = Builder->CreateOr(B, D);
2528 return SelectInst::Create(X, orTrue, orFalse);
2532 return Changed ? &I : nullptr;
2535 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
2536 bool Changed = SimplifyAssociativeOrCommutative(I);
2537 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2539 if (Value *V = SimplifyVectorOp(I))
2540 return ReplaceInstUsesWith(I, V);
2542 if (Value *V = SimplifyXorInst(Op0, Op1, DL, TLI, DT, AC))
2543 return ReplaceInstUsesWith(I, V);
2545 // (A&B)^(A&C) -> A&(B^C) etc
2546 if (Value *V = SimplifyUsingDistributiveLaws(I))
2547 return ReplaceInstUsesWith(I, V);
2549 // See if we can simplify any instructions used by the instruction whose sole
2550 // purpose is to compute bits we don't care about.
2551 if (SimplifyDemandedInstructionBits(I))
2554 if (Value *V = SimplifyBSwap(I))
2555 return ReplaceInstUsesWith(I, V);
2557 // Is this a ~ operation?
2558 if (Value *NotOp = dyn_castNotVal(&I)) {
2559 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
2560 if (Op0I->getOpcode() == Instruction::And ||
2561 Op0I->getOpcode() == Instruction::Or) {
2562 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
2563 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
2564 if (dyn_castNotVal(Op0I->getOperand(1)))
2565 Op0I->swapOperands();
2566 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
2568 Builder->CreateNot(Op0I->getOperand(1),
2569 Op0I->getOperand(1)->getName()+".not");
2570 if (Op0I->getOpcode() == Instruction::And)
2571 return BinaryOperator::CreateOr(Op0NotVal, NotY);
2572 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
2575 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
2576 // ~(X | Y) === (~X & ~Y) - De Morgan's Law
2577 if (IsFreeToInvert(Op0I->getOperand(0),
2578 Op0I->getOperand(0)->hasOneUse()) &&
2579 IsFreeToInvert(Op0I->getOperand(1),
2580 Op0I->getOperand(1)->hasOneUse())) {
2582 Builder->CreateNot(Op0I->getOperand(0), "notlhs");
2584 Builder->CreateNot(Op0I->getOperand(1), "notrhs");
2585 if (Op0I->getOpcode() == Instruction::And)
2586 return BinaryOperator::CreateOr(NotX, NotY);
2587 return BinaryOperator::CreateAnd(NotX, NotY);
2590 } else if (Op0I->getOpcode() == Instruction::AShr) {
2591 // ~(~X >>s Y) --> (X >>s Y)
2592 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0)))
2593 return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1));
2598 if (Constant *RHS = dyn_cast<Constant>(Op1)) {
2599 if (RHS->isAllOnesValue() && Op0->hasOneUse())
2600 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
2601 if (CmpInst *CI = dyn_cast<CmpInst>(Op0))
2602 return CmpInst::Create(CI->getOpcode(),
2603 CI->getInversePredicate(),
2604 CI->getOperand(0), CI->getOperand(1));
2607 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2608 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
2609 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2610 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
2611 if (CI->hasOneUse() && Op0C->hasOneUse()) {
2612 Instruction::CastOps Opcode = Op0C->getOpcode();
2613 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
2614 (RHS == ConstantExpr::getCast(Opcode, Builder->getTrue(),
2615 Op0C->getDestTy()))) {
2616 CI->setPredicate(CI->getInversePredicate());
2617 return CastInst::Create(Opcode, CI, Op0C->getType());
2623 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2624 // ~(c-X) == X-c-1 == X+(-c-1)
2625 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
2626 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
2627 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
2628 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
2629 ConstantInt::get(I.getType(), 1));
2630 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
2633 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2634 if (Op0I->getOpcode() == Instruction::Add) {
2635 // ~(X-c) --> (-c-1)-X
2636 if (RHS->isAllOnesValue()) {
2637 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
2638 return BinaryOperator::CreateSub(
2639 ConstantExpr::getSub(NegOp0CI,
2640 ConstantInt::get(I.getType(), 1)),
2641 Op0I->getOperand(0));
2642 } else if (RHS->getValue().isSignBit()) {
2643 // (X + C) ^ signbit -> (X + C + signbit)
2644 Constant *C = Builder->getInt(RHS->getValue() + Op0CI->getValue());
2645 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
2648 } else if (Op0I->getOpcode() == Instruction::Or) {
2649 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
2650 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue(),
2652 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
2653 // Anything in both C1 and C2 is known to be zero, remove it from
2655 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
2656 NewRHS = ConstantExpr::getAnd(NewRHS,
2657 ConstantExpr::getNot(CommonBits));
2659 I.setOperand(0, Op0I->getOperand(0));
2660 I.setOperand(1, NewRHS);
2663 } else if (Op0I->getOpcode() == Instruction::LShr) {
2664 // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3)
2668 if (Op0I->hasOneUse() &&
2669 (E1 = dyn_cast<BinaryOperator>(Op0I->getOperand(0))) &&
2670 E1->getOpcode() == Instruction::Xor &&
2671 (C1 = dyn_cast<ConstantInt>(E1->getOperand(1)))) {
2672 // fold (C1 >> C2) ^ C3
2673 ConstantInt *C2 = Op0CI, *C3 = RHS;
2674 APInt FoldConst = C1->getValue().lshr(C2->getValue());
2675 FoldConst ^= C3->getValue();
2676 // Prepare the two operands.
2677 Value *Opnd0 = Builder->CreateLShr(E1->getOperand(0), C2);
2678 Opnd0->takeName(Op0I);
2679 cast<Instruction>(Opnd0)->setDebugLoc(I.getDebugLoc());
2680 Value *FoldVal = ConstantInt::get(Opnd0->getType(), FoldConst);
2682 return BinaryOperator::CreateXor(Opnd0, FoldVal);
2688 // Try to fold constant and into select arguments.
2689 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2690 if (Instruction *R = FoldOpIntoSelect(I, SI))
2692 if (isa<PHINode>(Op0))
2693 if (Instruction *NV = FoldOpIntoPhi(I))
2697 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
2700 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2701 if (A == Op0) { // B^(B|A) == (A|B)^B
2702 Op1I->swapOperands();
2704 std::swap(Op0, Op1);
2705 } else if (B == Op0) { // B^(A|B) == (A|B)^B
2706 I.swapOperands(); // Simplified below.
2707 std::swap(Op0, Op1);
2709 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
2711 if (A == Op0) { // A^(A&B) -> A^(B&A)
2712 Op1I->swapOperands();
2715 if (B == Op0) { // A^(B&A) -> (B&A)^A
2716 I.swapOperands(); // Simplified below.
2717 std::swap(Op0, Op1);
2722 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
2725 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2726 Op0I->hasOneUse()) {
2727 if (A == Op1) // (B|A)^B == (A|B)^B
2729 if (B == Op1) // (A|B)^B == A & ~B
2730 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1));
2731 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2733 if (A == Op1) // (A&B)^A -> (B&A)^A
2735 if (B == Op1 && // (B&A)^A == ~B & A
2736 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
2737 return BinaryOperator::CreateAnd(Builder->CreateNot(A), Op1);
2743 Value *A, *B, *C, *D;
2744 // (A & B)^(A | B) -> A ^ B
2745 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2746 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
2747 if ((A == C && B == D) || (A == D && B == C))
2748 return BinaryOperator::CreateXor(A, B);
2750 // (A | B)^(A & B) -> A ^ B
2751 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2752 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
2753 if ((A == C && B == D) || (A == D && B == C))
2754 return BinaryOperator::CreateXor(A, B);
2756 // (A | ~B) ^ (~A | B) -> A ^ B
2757 if (match(Op0I, m_Or(m_Value(A), m_Not(m_Value(B)))) &&
2758 match(Op1I, m_Or(m_Not(m_Specific(A)), m_Specific(B)))) {
2759 return BinaryOperator::CreateXor(A, B);
2761 // (~A | B) ^ (A | ~B) -> A ^ B
2762 if (match(Op0I, m_Or(m_Not(m_Value(A)), m_Value(B))) &&
2763 match(Op1I, m_Or(m_Specific(A), m_Not(m_Specific(B))))) {
2764 return BinaryOperator::CreateXor(A, B);
2766 // (A & ~B) ^ (~A & B) -> A ^ B
2767 if (match(Op0I, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2768 match(Op1I, m_And(m_Not(m_Specific(A)), m_Specific(B)))) {
2769 return BinaryOperator::CreateXor(A, B);
2771 // (~A & B) ^ (A & ~B) -> A ^ B
2772 if (match(Op0I, m_And(m_Not(m_Value(A)), m_Value(B))) &&
2773 match(Op1I, m_And(m_Specific(A), m_Not(m_Specific(B))))) {
2774 return BinaryOperator::CreateXor(A, B);
2776 // (A ^ C)^(A | B) -> ((~A) & B) ^ C
2777 if (match(Op0I, m_Xor(m_Value(D), m_Value(C))) &&
2778 match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2780 return BinaryOperator::CreateXor(
2781 Builder->CreateAnd(Builder->CreateNot(A), B), C);
2783 return BinaryOperator::CreateXor(
2784 Builder->CreateAnd(Builder->CreateNot(B), A), C);
2786 // (A | B)^(A ^ C) -> ((~A) & B) ^ C
2787 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2788 match(Op1I, m_Xor(m_Value(D), m_Value(C)))) {
2790 return BinaryOperator::CreateXor(
2791 Builder->CreateAnd(Builder->CreateNot(A), B), C);
2793 return BinaryOperator::CreateXor(
2794 Builder->CreateAnd(Builder->CreateNot(B), A), C);
2796 // (A & B) ^ (A ^ B) -> (A | B)
2797 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2798 match(Op1I, m_Xor(m_Specific(A), m_Specific(B))))
2799 return BinaryOperator::CreateOr(A, B);
2800 // (A ^ B) ^ (A & B) -> (A | B)
2801 if (match(Op0I, m_Xor(m_Value(A), m_Value(B))) &&
2802 match(Op1I, m_And(m_Specific(A), m_Specific(B))))
2803 return BinaryOperator::CreateOr(A, B);
2806 Value *A = nullptr, *B = nullptr;
2807 // (A & ~B) ^ (~A) -> ~(A & B)
2808 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2809 match(Op1, m_Not(m_Specific(A))))
2810 return BinaryOperator::CreateNot(Builder->CreateAnd(A, B));
2812 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
2813 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
2814 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
2815 if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
2816 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2817 LHS->getOperand(1) == RHS->getOperand(0))
2818 LHS->swapOperands();
2819 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2820 LHS->getOperand(1) == RHS->getOperand(1)) {
2821 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2822 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
2823 bool isSigned = LHS->isSigned() || RHS->isSigned();
2824 return ReplaceInstUsesWith(I,
2825 getNewICmpValue(isSigned, Code, Op0, Op1,
2830 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
2831 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2832 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
2833 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
2834 Type *SrcTy = Op0C->getOperand(0)->getType();
2835 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() &&
2836 // Only do this if the casts both really cause code to be generated.
2837 ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0),
2839 ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0),
2841 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
2842 Op1C->getOperand(0), I.getName());
2843 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2848 return Changed ? &I : nullptr;