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 "InstCombine.h"
15 #include "llvm/Analysis/InstructionSimplify.h"
16 #include "llvm/IR/ConstantRange.h"
17 #include "llvm/IR/Intrinsics.h"
18 #include "llvm/IR/PatternMatch.h"
19 #include "llvm/Transforms/Utils/CmpInstAnalysis.h"
21 using namespace PatternMatch;
23 #define DEBUG_TYPE "instcombine"
25 /// isFreeToInvert - Return true if the specified value is free to invert (apply
26 /// ~ to). This happens in cases where the ~ can be eliminated.
27 static inline bool isFreeToInvert(Value *V) {
29 if (BinaryOperator::isNot(V))
32 // Constants can be considered to be not'ed values.
33 if (isa<ConstantInt>(V))
36 // Compares can be inverted if they have a single use.
37 if (CmpInst *CI = dyn_cast<CmpInst>(V))
38 return CI->hasOneUse();
43 static inline Value *dyn_castNotVal(Value *V) {
44 // If this is not(not(x)) don't return that this is a not: we want the two
45 // not's to be folded first.
46 if (BinaryOperator::isNot(V)) {
47 Value *Operand = BinaryOperator::getNotArgument(V);
48 if (!isFreeToInvert(Operand))
52 // Constants can be considered to be not'ed values...
53 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
54 return ConstantInt::get(C->getType(), ~C->getValue());
58 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
59 /// predicate into a three bit mask. It also returns whether it is an ordered
60 /// predicate by reference.
61 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
64 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
65 case FCmpInst::FCMP_UNO: return 0; // 000
66 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
67 case FCmpInst::FCMP_UGT: return 1; // 001
68 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
69 case FCmpInst::FCMP_UEQ: return 2; // 010
70 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
71 case FCmpInst::FCMP_UGE: return 3; // 011
72 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
73 case FCmpInst::FCMP_ULT: return 4; // 100
74 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
75 case FCmpInst::FCMP_UNE: return 5; // 101
76 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
77 case FCmpInst::FCMP_ULE: return 6; // 110
80 // Not expecting FCMP_FALSE and FCMP_TRUE;
81 llvm_unreachable("Unexpected FCmp predicate!");
85 /// getNewICmpValue - This is the complement of getICmpCode, which turns an
86 /// opcode and two operands into either a constant true or false, or a brand
87 /// new ICmp instruction. The sign is passed in to determine which kind
88 /// of predicate to use in the new icmp instruction.
89 static Value *getNewICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS,
90 InstCombiner::BuilderTy *Builder) {
91 ICmpInst::Predicate NewPred;
92 if (Value *NewConstant = getICmpValue(Sign, Code, LHS, RHS, NewPred))
94 return Builder->CreateICmp(NewPred, LHS, RHS);
97 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
98 /// opcode and two operands into either a FCmp instruction. isordered is passed
99 /// in to determine which kind of predicate to use in the new fcmp instruction.
100 static Value *getFCmpValue(bool isordered, unsigned code,
101 Value *LHS, Value *RHS,
102 InstCombiner::BuilderTy *Builder) {
103 CmpInst::Predicate Pred;
105 default: llvm_unreachable("Illegal FCmp code!");
106 case 0: Pred = isordered ? FCmpInst::FCMP_ORD : FCmpInst::FCMP_UNO; break;
107 case 1: Pred = isordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; break;
108 case 2: Pred = isordered ? FCmpInst::FCMP_OEQ : FCmpInst::FCMP_UEQ; break;
109 case 3: Pred = isordered ? FCmpInst::FCMP_OGE : FCmpInst::FCMP_UGE; break;
110 case 4: Pred = isordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; break;
111 case 5: Pred = isordered ? FCmpInst::FCMP_ONE : FCmpInst::FCMP_UNE; break;
112 case 6: Pred = isordered ? FCmpInst::FCMP_OLE : FCmpInst::FCMP_ULE; break;
114 if (!isordered) return ConstantInt::getTrue(LHS->getContext());
115 Pred = FCmpInst::FCMP_ORD; break;
117 return Builder->CreateFCmp(Pred, LHS, RHS);
120 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
121 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
122 // guaranteed to be a binary operator.
123 Instruction *InstCombiner::OptAndOp(Instruction *Op,
126 BinaryOperator &TheAnd) {
127 Value *X = Op->getOperand(0);
128 Constant *Together = nullptr;
130 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
132 switch (Op->getOpcode()) {
133 case Instruction::Xor:
134 if (Op->hasOneUse()) {
135 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
136 Value *And = Builder->CreateAnd(X, AndRHS);
138 return BinaryOperator::CreateXor(And, Together);
141 case Instruction::Or:
142 if (Op->hasOneUse()){
143 if (Together != OpRHS) {
144 // (X | C1) & C2 --> (X | (C1&C2)) & C2
145 Value *Or = Builder->CreateOr(X, Together);
147 return BinaryOperator::CreateAnd(Or, AndRHS);
150 ConstantInt *TogetherCI = dyn_cast<ConstantInt>(Together);
151 if (TogetherCI && !TogetherCI->isZero()){
152 // (X | C1) & C2 --> (X & (C2^(C1&C2))) | C1
153 // NOTE: This reduces the number of bits set in the & mask, which
154 // can expose opportunities for store narrowing.
155 Together = ConstantExpr::getXor(AndRHS, Together);
156 Value *And = Builder->CreateAnd(X, Together);
158 return BinaryOperator::CreateOr(And, OpRHS);
163 case Instruction::Add:
164 if (Op->hasOneUse()) {
165 // Adding a one to a single bit bit-field should be turned into an XOR
166 // of the bit. First thing to check is to see if this AND is with a
167 // single bit constant.
168 const APInt &AndRHSV = AndRHS->getValue();
170 // If there is only one bit set.
171 if (AndRHSV.isPowerOf2()) {
172 // Ok, at this point, we know that we are masking the result of the
173 // ADD down to exactly one bit. If the constant we are adding has
174 // no bits set below this bit, then we can eliminate the ADD.
175 const APInt& AddRHS = OpRHS->getValue();
177 // Check to see if any bits below the one bit set in AndRHSV are set.
178 if ((AddRHS & (AndRHSV-1)) == 0) {
179 // If not, the only thing that can effect the output of the AND is
180 // the bit specified by AndRHSV. If that bit is set, the effect of
181 // the XOR is to toggle the bit. If it is clear, then the ADD has
183 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
184 TheAnd.setOperand(0, X);
187 // Pull the XOR out of the AND.
188 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
189 NewAnd->takeName(Op);
190 return BinaryOperator::CreateXor(NewAnd, AndRHS);
197 case Instruction::Shl: {
198 // We know that the AND will not produce any of the bits shifted in, so if
199 // the anded constant includes them, clear them now!
201 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
202 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
203 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
204 ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShlMask);
206 if (CI->getValue() == ShlMask)
207 // Masking out bits that the shift already masks.
208 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
210 if (CI != AndRHS) { // Reducing bits set in and.
211 TheAnd.setOperand(1, CI);
216 case Instruction::LShr: {
217 // We know that the AND will not produce any of the bits shifted in, so if
218 // the anded constant includes them, clear them now! This only applies to
219 // unsigned shifts, because a signed shr may bring in set bits!
221 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
222 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
223 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
224 ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShrMask);
226 if (CI->getValue() == ShrMask)
227 // Masking out bits that the shift already masks.
228 return ReplaceInstUsesWith(TheAnd, Op);
231 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
236 case Instruction::AShr:
238 // See if this is shifting in some sign extension, then masking it out
240 if (Op->hasOneUse()) {
241 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
242 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
243 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
244 Constant *C = Builder->getInt(AndRHS->getValue() & ShrMask);
245 if (C == AndRHS) { // Masking out bits shifted in.
246 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
247 // Make the argument unsigned.
248 Value *ShVal = Op->getOperand(0);
249 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
250 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
258 /// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise
259 /// (V < Lo || V >= Hi). In practice, we emit the more efficient
260 /// (V-Lo) \<u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
261 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
262 /// insert new instructions.
263 Value *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
264 bool isSigned, bool Inside) {
265 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
266 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
267 "Lo is not <= Hi in range emission code!");
270 if (Lo == Hi) // Trivially false.
271 return Builder->getFalse();
273 // V >= Min && V < Hi --> V < Hi
274 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
275 ICmpInst::Predicate pred = (isSigned ?
276 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
277 return Builder->CreateICmp(pred, V, Hi);
280 // Emit V-Lo <u Hi-Lo
281 Constant *NegLo = ConstantExpr::getNeg(Lo);
282 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
283 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
284 return Builder->CreateICmpULT(Add, UpperBound);
287 if (Lo == Hi) // Trivially true.
288 return Builder->getTrue();
290 // V < Min || V >= Hi -> V > Hi-1
291 Hi = SubOne(cast<ConstantInt>(Hi));
292 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
293 ICmpInst::Predicate pred = (isSigned ?
294 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
295 return Builder->CreateICmp(pred, V, Hi);
298 // Emit V-Lo >u Hi-1-Lo
299 // Note that Hi has already had one subtracted from it, above.
300 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
301 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
302 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
303 return Builder->CreateICmpUGT(Add, LowerBound);
306 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
307 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
308 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
309 // not, since all 1s are not contiguous.
310 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
311 const APInt& V = Val->getValue();
312 uint32_t BitWidth = Val->getType()->getBitWidth();
313 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
315 // look for the first zero bit after the run of ones
316 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
317 // look for the first non-zero bit
318 ME = V.getActiveBits();
322 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
323 /// where isSub determines whether the operator is a sub. If we can fold one of
324 /// the following xforms:
326 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
327 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
328 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
330 /// return (A +/- B).
332 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
333 ConstantInt *Mask, bool isSub,
335 Instruction *LHSI = dyn_cast<Instruction>(LHS);
336 if (!LHSI || LHSI->getNumOperands() != 2 ||
337 !isa<ConstantInt>(LHSI->getOperand(1))) return nullptr;
339 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
341 switch (LHSI->getOpcode()) {
342 default: return nullptr;
343 case Instruction::And:
344 if (ConstantExpr::getAnd(N, Mask) == Mask) {
345 // If the AndRHS is a power of two minus one (0+1+), this is simple.
346 if ((Mask->getValue().countLeadingZeros() +
347 Mask->getValue().countPopulation()) ==
348 Mask->getValue().getBitWidth())
351 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
352 // part, we don't need any explicit masks to take them out of A. If that
353 // is all N is, ignore it.
354 uint32_t MB = 0, ME = 0;
355 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
356 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
357 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
358 if (MaskedValueIsZero(RHS, Mask, 0, &I))
363 case Instruction::Or:
364 case Instruction::Xor:
365 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
366 if ((Mask->getValue().countLeadingZeros() +
367 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
368 && ConstantExpr::getAnd(N, Mask)->isNullValue())
374 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
375 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
378 /// enum for classifying (icmp eq (A & B), C) and (icmp ne (A & B), C)
379 /// One of A and B is considered the mask, the other the value. This is
380 /// described as the "AMask" or "BMask" part of the enum. If the enum
381 /// contains only "Mask", then both A and B can be considered masks.
382 /// If A is the mask, then it was proven, that (A & C) == C. This
383 /// is trivial if C == A, or C == 0. If both A and C are constants, this
384 /// proof is also easy.
385 /// For the following explanations we assume that A is the mask.
386 /// The part "AllOnes" declares, that the comparison is true only
387 /// if (A & B) == A, or all bits of A are set in B.
388 /// Example: (icmp eq (A & 3), 3) -> FoldMskICmp_AMask_AllOnes
389 /// The part "AllZeroes" declares, that the comparison is true only
390 /// if (A & B) == 0, or all bits of A are cleared in B.
391 /// Example: (icmp eq (A & 3), 0) -> FoldMskICmp_Mask_AllZeroes
392 /// The part "Mixed" declares, that (A & B) == C and C might or might not
393 /// contain any number of one bits and zero bits.
394 /// Example: (icmp eq (A & 3), 1) -> FoldMskICmp_AMask_Mixed
395 /// The Part "Not" means, that in above descriptions "==" should be replaced
397 /// Example: (icmp ne (A & 3), 3) -> FoldMskICmp_AMask_NotAllOnes
398 /// If the mask A contains a single bit, then the following is equivalent:
399 /// (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
400 /// (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
401 enum MaskedICmpType {
402 FoldMskICmp_AMask_AllOnes = 1,
403 FoldMskICmp_AMask_NotAllOnes = 2,
404 FoldMskICmp_BMask_AllOnes = 4,
405 FoldMskICmp_BMask_NotAllOnes = 8,
406 FoldMskICmp_Mask_AllZeroes = 16,
407 FoldMskICmp_Mask_NotAllZeroes = 32,
408 FoldMskICmp_AMask_Mixed = 64,
409 FoldMskICmp_AMask_NotMixed = 128,
410 FoldMskICmp_BMask_Mixed = 256,
411 FoldMskICmp_BMask_NotMixed = 512
414 /// return the set of pattern classes (from MaskedICmpType)
415 /// that (icmp SCC (A & B), C) satisfies
416 static unsigned getTypeOfMaskedICmp(Value* A, Value* B, Value* C,
417 ICmpInst::Predicate SCC)
419 ConstantInt *ACst = dyn_cast<ConstantInt>(A);
420 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
421 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
422 bool icmp_eq = (SCC == ICmpInst::ICMP_EQ);
423 bool icmp_abit = (ACst && !ACst->isZero() &&
424 ACst->getValue().isPowerOf2());
425 bool icmp_bbit = (BCst && !BCst->isZero() &&
426 BCst->getValue().isPowerOf2());
428 if (CCst && CCst->isZero()) {
429 // if C is zero, then both A and B qualify as mask
430 result |= (icmp_eq ? (FoldMskICmp_Mask_AllZeroes |
431 FoldMskICmp_Mask_AllZeroes |
432 FoldMskICmp_AMask_Mixed |
433 FoldMskICmp_BMask_Mixed)
434 : (FoldMskICmp_Mask_NotAllZeroes |
435 FoldMskICmp_Mask_NotAllZeroes |
436 FoldMskICmp_AMask_NotMixed |
437 FoldMskICmp_BMask_NotMixed));
439 result |= (icmp_eq ? (FoldMskICmp_AMask_NotAllOnes |
440 FoldMskICmp_AMask_NotMixed)
441 : (FoldMskICmp_AMask_AllOnes |
442 FoldMskICmp_AMask_Mixed));
444 result |= (icmp_eq ? (FoldMskICmp_BMask_NotAllOnes |
445 FoldMskICmp_BMask_NotMixed)
446 : (FoldMskICmp_BMask_AllOnes |
447 FoldMskICmp_BMask_Mixed));
451 result |= (icmp_eq ? (FoldMskICmp_AMask_AllOnes |
452 FoldMskICmp_AMask_Mixed)
453 : (FoldMskICmp_AMask_NotAllOnes |
454 FoldMskICmp_AMask_NotMixed));
456 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
457 FoldMskICmp_AMask_NotMixed)
458 : (FoldMskICmp_Mask_AllZeroes |
459 FoldMskICmp_AMask_Mixed));
460 } else if (ACst && CCst &&
461 ConstantExpr::getAnd(ACst, CCst) == CCst) {
462 result |= (icmp_eq ? FoldMskICmp_AMask_Mixed
463 : FoldMskICmp_AMask_NotMixed);
466 result |= (icmp_eq ? (FoldMskICmp_BMask_AllOnes |
467 FoldMskICmp_BMask_Mixed)
468 : (FoldMskICmp_BMask_NotAllOnes |
469 FoldMskICmp_BMask_NotMixed));
471 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
472 FoldMskICmp_BMask_NotMixed)
473 : (FoldMskICmp_Mask_AllZeroes |
474 FoldMskICmp_BMask_Mixed));
475 } else if (BCst && CCst &&
476 ConstantExpr::getAnd(BCst, CCst) == CCst) {
477 result |= (icmp_eq ? FoldMskICmp_BMask_Mixed
478 : FoldMskICmp_BMask_NotMixed);
483 /// Convert an analysis of a masked ICmp into its equivalent if all boolean
484 /// operations had the opposite sense. Since each "NotXXX" flag (recording !=)
485 /// is adjacent to the corresponding normal flag (recording ==), this just
486 /// involves swapping those bits over.
487 static unsigned conjugateICmpMask(unsigned Mask) {
489 NewMask = (Mask & (FoldMskICmp_AMask_AllOnes | FoldMskICmp_BMask_AllOnes |
490 FoldMskICmp_Mask_AllZeroes | FoldMskICmp_AMask_Mixed |
491 FoldMskICmp_BMask_Mixed))
495 (Mask & (FoldMskICmp_AMask_NotAllOnes | FoldMskICmp_BMask_NotAllOnes |
496 FoldMskICmp_Mask_NotAllZeroes | FoldMskICmp_AMask_NotMixed |
497 FoldMskICmp_BMask_NotMixed))
503 /// decomposeBitTestICmp - Decompose an icmp into the form ((X & Y) pred Z)
504 /// if possible. The returned predicate is either == or !=. Returns false if
505 /// decomposition fails.
506 static bool decomposeBitTestICmp(const ICmpInst *I, ICmpInst::Predicate &Pred,
507 Value *&X, Value *&Y, Value *&Z) {
508 ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1));
512 switch (I->getPredicate()) {
515 case ICmpInst::ICMP_SLT:
516 // X < 0 is equivalent to (X & SignBit) != 0.
519 Y = ConstantInt::get(I->getContext(), APInt::getSignBit(C->getBitWidth()));
520 Pred = ICmpInst::ICMP_NE;
522 case ICmpInst::ICMP_SGT:
523 // X > -1 is equivalent to (X & SignBit) == 0.
524 if (!C->isAllOnesValue())
526 Y = ConstantInt::get(I->getContext(), APInt::getSignBit(C->getBitWidth()));
527 Pred = ICmpInst::ICMP_EQ;
529 case ICmpInst::ICMP_ULT:
530 // X <u 2^n is equivalent to (X & ~(2^n-1)) == 0.
531 if (!C->getValue().isPowerOf2())
533 Y = ConstantInt::get(I->getContext(), -C->getValue());
534 Pred = ICmpInst::ICMP_EQ;
536 case ICmpInst::ICMP_UGT:
537 // X >u 2^n-1 is equivalent to (X & ~(2^n-1)) != 0.
538 if (!(C->getValue() + 1).isPowerOf2())
540 Y = ConstantInt::get(I->getContext(), ~C->getValue());
541 Pred = ICmpInst::ICMP_NE;
545 X = I->getOperand(0);
546 Z = ConstantInt::getNullValue(C->getType());
550 /// foldLogOpOfMaskedICmpsHelper:
551 /// handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
552 /// return the set of pattern classes (from MaskedICmpType)
553 /// that both LHS and RHS satisfy
554 static unsigned foldLogOpOfMaskedICmpsHelper(Value*& A,
555 Value*& B, Value*& C,
556 Value*& D, Value*& E,
557 ICmpInst *LHS, ICmpInst *RHS,
558 ICmpInst::Predicate &LHSCC,
559 ICmpInst::Predicate &RHSCC) {
560 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) return 0;
561 // vectors are not (yet?) supported
562 if (LHS->getOperand(0)->getType()->isVectorTy()) return 0;
564 // Here comes the tricky part:
565 // LHS might be of the form L11 & L12 == X, X == L21 & L22,
566 // and L11 & L12 == L21 & L22. The same goes for RHS.
567 // Now we must find those components L** and R**, that are equal, so
568 // that we can extract the parameters A, B, C, D, and E for the canonical
570 Value *L1 = LHS->getOperand(0);
571 Value *L2 = LHS->getOperand(1);
572 Value *L11,*L12,*L21,*L22;
573 // Check whether the icmp can be decomposed into a bit test.
574 if (decomposeBitTestICmp(LHS, LHSCC, L11, L12, L2)) {
575 L21 = L22 = L1 = nullptr;
577 // Look for ANDs in the LHS icmp.
578 if (!L1->getType()->isIntegerTy()) {
579 // You can icmp pointers, for example. They really aren't masks.
581 } else if (!match(L1, m_And(m_Value(L11), m_Value(L12)))) {
582 // Any icmp can be viewed as being trivially masked; if it allows us to
583 // remove one, it's worth it.
585 L12 = Constant::getAllOnesValue(L1->getType());
588 if (!L2->getType()->isIntegerTy()) {
589 // You can icmp pointers, for example. They really aren't masks.
591 } else if (!match(L2, m_And(m_Value(L21), m_Value(L22)))) {
593 L22 = Constant::getAllOnesValue(L2->getType());
597 // Bail if LHS was a icmp that can't be decomposed into an equality.
598 if (!ICmpInst::isEquality(LHSCC))
601 Value *R1 = RHS->getOperand(0);
602 Value *R2 = RHS->getOperand(1);
605 if (decomposeBitTestICmp(RHS, RHSCC, R11, R12, R2)) {
606 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
608 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
613 E = R2; R1 = nullptr; ok = true;
614 } else if (R1->getType()->isIntegerTy()) {
615 if (!match(R1, m_And(m_Value(R11), m_Value(R12)))) {
616 // As before, model no mask as a trivial mask if it'll let us do an
619 R12 = Constant::getAllOnesValue(R1->getType());
622 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
623 A = R11; D = R12; E = R2; ok = true;
624 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
625 A = R12; D = R11; E = R2; ok = true;
629 // Bail if RHS was a icmp that can't be decomposed into an equality.
630 if (!ICmpInst::isEquality(RHSCC))
633 // Look for ANDs in on the right side of the RHS icmp.
634 if (!ok && R2->getType()->isIntegerTy()) {
635 if (!match(R2, m_And(m_Value(R11), m_Value(R12)))) {
637 R12 = Constant::getAllOnesValue(R2->getType());
640 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
641 A = R11; D = R12; E = R1; ok = true;
642 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
643 A = R12; D = R11; E = R1; ok = true;
653 } else if (L12 == A) {
655 } else if (L21 == A) {
657 } else if (L22 == A) {
661 unsigned left_type = getTypeOfMaskedICmp(A, B, C, LHSCC);
662 unsigned right_type = getTypeOfMaskedICmp(A, D, E, RHSCC);
663 return left_type & right_type;
665 /// foldLogOpOfMaskedICmps:
666 /// try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
667 /// into a single (icmp(A & X) ==/!= Y)
668 static Value *foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
669 llvm::InstCombiner::BuilderTy *Builder) {
670 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr;
671 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
672 unsigned mask = foldLogOpOfMaskedICmpsHelper(A, B, C, D, E, LHS, RHS,
674 if (mask == 0) return nullptr;
675 assert(ICmpInst::isEquality(LHSCC) && ICmpInst::isEquality(RHSCC) &&
676 "foldLogOpOfMaskedICmpsHelper must return an equality predicate.");
678 // In full generality:
679 // (icmp (A & B) Op C) | (icmp (A & D) Op E)
680 // == ![ (icmp (A & B) !Op C) & (icmp (A & D) !Op E) ]
682 // If the latter can be converted into (icmp (A & X) Op Y) then the former is
683 // equivalent to (icmp (A & X) !Op Y).
685 // Therefore, we can pretend for the rest of this function that we're dealing
686 // with the conjunction, provided we flip the sense of any comparisons (both
687 // input and output).
689 // In most cases we're going to produce an EQ for the "&&" case.
690 ICmpInst::Predicate NEWCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
692 // Convert the masking analysis into its equivalent with negated
694 mask = conjugateICmpMask(mask);
697 if (mask & FoldMskICmp_Mask_AllZeroes) {
698 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
699 // -> (icmp eq (A & (B|D)), 0)
700 Value *newOr = Builder->CreateOr(B, D);
701 Value *newAnd = Builder->CreateAnd(A, newOr);
702 // we can't use C as zero, because we might actually handle
703 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
704 // with B and D, having a single bit set
705 Value *zero = Constant::getNullValue(A->getType());
706 return Builder->CreateICmp(NEWCC, newAnd, zero);
708 if (mask & FoldMskICmp_BMask_AllOnes) {
709 // (icmp eq (A & B), B) & (icmp eq (A & D), D)
710 // -> (icmp eq (A & (B|D)), (B|D))
711 Value *newOr = Builder->CreateOr(B, D);
712 Value *newAnd = Builder->CreateAnd(A, newOr);
713 return Builder->CreateICmp(NEWCC, newAnd, newOr);
715 if (mask & FoldMskICmp_AMask_AllOnes) {
716 // (icmp eq (A & B), A) & (icmp eq (A & D), A)
717 // -> (icmp eq (A & (B&D)), A)
718 Value *newAnd1 = Builder->CreateAnd(B, D);
719 Value *newAnd = Builder->CreateAnd(A, newAnd1);
720 return Builder->CreateICmp(NEWCC, newAnd, A);
723 // Remaining cases assume at least that B and D are constant, and depend on
724 // their actual values. This isn't strictly, necessary, just a "handle the
725 // easy cases for now" decision.
726 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
727 if (!BCst) return nullptr;
728 ConstantInt *DCst = dyn_cast<ConstantInt>(D);
729 if (!DCst) return nullptr;
731 if (mask & (FoldMskICmp_Mask_NotAllZeroes | FoldMskICmp_BMask_NotAllOnes)) {
732 // (icmp ne (A & B), 0) & (icmp ne (A & D), 0) and
733 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
734 // -> (icmp ne (A & B), 0) or (icmp ne (A & D), 0)
735 // Only valid if one of the masks is a superset of the other (check "B&D" is
736 // the same as either B or D).
737 APInt NewMask = BCst->getValue() & DCst->getValue();
739 if (NewMask == BCst->getValue())
741 else if (NewMask == DCst->getValue())
744 if (mask & FoldMskICmp_AMask_NotAllOnes) {
745 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
746 // -> (icmp ne (A & B), A) or (icmp ne (A & D), A)
747 // Only valid if one of the masks is a superset of the other (check "B|D" is
748 // the same as either B or D).
749 APInt NewMask = BCst->getValue() | DCst->getValue();
751 if (NewMask == BCst->getValue())
753 else if (NewMask == DCst->getValue())
756 if (mask & FoldMskICmp_BMask_Mixed) {
757 // (icmp eq (A & B), C) & (icmp eq (A & D), E)
758 // We already know that B & C == C && D & E == E.
759 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
760 // C and E, which are shared by both the mask B and the mask D, don't
761 // contradict, then we can transform to
762 // -> (icmp eq (A & (B|D)), (C|E))
763 // Currently, we only handle the case of B, C, D, and E being constant.
764 // we can't simply use C and E, because we might actually handle
765 // (icmp ne (A & B), B) & (icmp eq (A & D), D)
766 // with B and D, having a single bit set
767 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
768 if (!CCst) return nullptr;
769 ConstantInt *ECst = dyn_cast<ConstantInt>(E);
770 if (!ECst) return nullptr;
772 CCst = cast<ConstantInt>(ConstantExpr::getXor(BCst, CCst));
774 ECst = cast<ConstantInt>(ConstantExpr::getXor(DCst, ECst));
775 // if there is a conflict we should actually return a false for the
777 if (((BCst->getValue() & DCst->getValue()) &
778 (CCst->getValue() ^ ECst->getValue())) != 0)
779 return ConstantInt::get(LHS->getType(), !IsAnd);
780 Value *newOr1 = Builder->CreateOr(B, D);
781 Value *newOr2 = ConstantExpr::getOr(CCst, ECst);
782 Value *newAnd = Builder->CreateAnd(A, newOr1);
783 return Builder->CreateICmp(NEWCC, newAnd, newOr2);
788 /// Try to fold a signed range checked with lower bound 0 to an unsigned icmp.
789 /// Example: (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
790 /// If \p Inverted is true then the check is for the inverted range, e.g.
791 /// (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
792 Value *InstCombiner::simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1,
794 // Check the lower range comparison, e.g. x >= 0
795 // InstCombine already ensured that if there is a constant it's on the RHS.
796 ConstantInt *RangeStart = dyn_cast<ConstantInt>(Cmp0->getOperand(1));
800 ICmpInst::Predicate Pred0 = (Inverted ? Cmp0->getInversePredicate() :
801 Cmp0->getPredicate());
803 // Accept x > -1 or x >= 0 (after potentially inverting the predicate).
804 if (!((Pred0 == ICmpInst::ICMP_SGT && RangeStart->isMinusOne()) ||
805 (Pred0 == ICmpInst::ICMP_SGE && RangeStart->isZero())))
808 ICmpInst::Predicate Pred1 = (Inverted ? Cmp1->getInversePredicate() :
809 Cmp1->getPredicate());
811 Value *Input = Cmp0->getOperand(0);
813 if (Cmp1->getOperand(0) == Input) {
814 // For the upper range compare we have: icmp x, n
815 RangeEnd = Cmp1->getOperand(1);
816 } else if (Cmp1->getOperand(1) == Input) {
817 // For the upper range compare we have: icmp n, x
818 RangeEnd = Cmp1->getOperand(0);
819 Pred1 = ICmpInst::getSwappedPredicate(Pred1);
824 // Check the upper range comparison, e.g. x < n
825 ICmpInst::Predicate NewPred;
827 case ICmpInst::ICMP_SLT: NewPred = ICmpInst::ICMP_ULT; break;
828 case ICmpInst::ICMP_SLE: NewPred = ICmpInst::ICMP_ULE; break;
829 default: return nullptr;
832 // This simplification is only valid if the upper range is not negative.
833 bool IsNegative, IsNotNegative;
834 ComputeSignBit(RangeEnd, IsNotNegative, IsNegative, DL, 0, AT,
840 NewPred = ICmpInst::getInversePredicate(NewPred);
842 return Builder->CreateICmp(NewPred, Input, RangeEnd);
845 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
846 Value *InstCombiner::FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
847 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
849 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
850 if (PredicatesFoldable(LHSCC, RHSCC)) {
851 if (LHS->getOperand(0) == RHS->getOperand(1) &&
852 LHS->getOperand(1) == RHS->getOperand(0))
854 if (LHS->getOperand(0) == RHS->getOperand(0) &&
855 LHS->getOperand(1) == RHS->getOperand(1)) {
856 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
857 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
858 bool isSigned = LHS->isSigned() || RHS->isSigned();
859 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
863 // handle (roughly): (icmp eq (A & B), C) & (icmp eq (A & D), E)
864 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, true, Builder))
867 // E.g. (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
868 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/false))
871 // E.g. (icmp slt x, n) & (icmp sge x, 0) --> icmp ult x, n
872 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/false))
875 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
876 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
877 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
878 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
879 if (!LHSCst || !RHSCst) return nullptr;
881 if (LHSCst == RHSCst && LHSCC == RHSCC) {
882 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
883 // where C is a power of 2
884 if (LHSCC == ICmpInst::ICMP_ULT &&
885 LHSCst->getValue().isPowerOf2()) {
886 Value *NewOr = Builder->CreateOr(Val, Val2);
887 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
890 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
891 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
892 Value *NewOr = Builder->CreateOr(Val, Val2);
893 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
897 // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2
898 // where CMAX is the all ones value for the truncated type,
899 // iff the lower bits of C2 and CA are zero.
900 if (LHSCC == ICmpInst::ICMP_EQ && LHSCC == RHSCC &&
901 LHS->hasOneUse() && RHS->hasOneUse()) {
903 ConstantInt *AndCst, *SmallCst = nullptr, *BigCst = nullptr;
905 // (trunc x) == C1 & (and x, CA) == C2
906 // (and x, CA) == C2 & (trunc x) == C1
907 if (match(Val2, m_Trunc(m_Value(V))) &&
908 match(Val, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
911 } else if (match(Val, m_Trunc(m_Value(V))) &&
912 match(Val2, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
917 if (SmallCst && BigCst) {
918 unsigned BigBitSize = BigCst->getType()->getBitWidth();
919 unsigned SmallBitSize = SmallCst->getType()->getBitWidth();
921 // Check that the low bits are zero.
922 APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize);
923 if ((Low & AndCst->getValue()) == 0 && (Low & BigCst->getValue()) == 0) {
924 Value *NewAnd = Builder->CreateAnd(V, Low | AndCst->getValue());
925 APInt N = SmallCst->getValue().zext(BigBitSize) | BigCst->getValue();
926 Value *NewVal = ConstantInt::get(AndCst->getType()->getContext(), N);
927 return Builder->CreateICmp(LHSCC, NewAnd, NewVal);
932 // From here on, we only handle:
933 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
934 if (Val != Val2) return nullptr;
936 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
937 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
938 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
939 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
940 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
943 // Make a constant range that's the intersection of the two icmp ranges.
944 // If the intersection is empty, we know that the result is false.
945 ConstantRange LHSRange =
946 ConstantRange::makeICmpRegion(LHSCC, LHSCst->getValue());
947 ConstantRange RHSRange =
948 ConstantRange::makeICmpRegion(RHSCC, RHSCst->getValue());
950 if (LHSRange.intersectWith(RHSRange).isEmptySet())
951 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
953 // We can't fold (ugt x, C) & (sgt x, C2).
954 if (!PredicatesFoldable(LHSCC, RHSCC))
957 // Ensure that the larger constant is on the RHS.
959 if (CmpInst::isSigned(LHSCC) ||
960 (ICmpInst::isEquality(LHSCC) &&
961 CmpInst::isSigned(RHSCC)))
962 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
964 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
968 std::swap(LHSCst, RHSCst);
969 std::swap(LHSCC, RHSCC);
972 // At this point, we know we have two icmp instructions
973 // comparing a value against two constants and and'ing the result
974 // together. Because of the above check, we know that we only have
975 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
976 // (from the icmp folding check above), that the two constants
977 // are not equal and that the larger constant is on the RHS
978 assert(LHSCst != RHSCst && "Compares not folded above?");
981 default: llvm_unreachable("Unknown integer condition code!");
982 case ICmpInst::ICMP_EQ:
984 default: llvm_unreachable("Unknown integer condition code!");
985 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
986 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
987 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
990 case ICmpInst::ICMP_NE:
992 default: llvm_unreachable("Unknown integer condition code!");
993 case ICmpInst::ICMP_ULT:
994 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
995 return Builder->CreateICmpULT(Val, LHSCst);
996 if (LHSCst->isNullValue()) // (X != 0 & X u< 14) -> X-1 u< 13
997 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
998 break; // (X != 13 & X u< 15) -> no change
999 case ICmpInst::ICMP_SLT:
1000 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
1001 return Builder->CreateICmpSLT(Val, LHSCst);
1002 break; // (X != 13 & X s< 15) -> no change
1003 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
1004 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
1005 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
1007 case ICmpInst::ICMP_NE:
1008 // Special case to get the ordering right when the values wrap around
1010 if (LHSCst->getValue() == 0 && RHSCst->getValue().isAllOnesValue())
1011 std::swap(LHSCst, RHSCst);
1012 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
1013 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1014 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1015 return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1),
1016 Val->getName()+".cmp");
1018 break; // (X != 13 & X != 15) -> no change
1021 case ICmpInst::ICMP_ULT:
1023 default: llvm_unreachable("Unknown integer condition code!");
1024 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
1025 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
1026 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
1027 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
1029 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
1030 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
1032 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
1036 case ICmpInst::ICMP_SLT:
1038 default: llvm_unreachable("Unknown integer condition code!");
1039 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
1041 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
1042 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
1044 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
1048 case ICmpInst::ICMP_UGT:
1050 default: llvm_unreachable("Unknown integer condition code!");
1051 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
1052 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
1054 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
1056 case ICmpInst::ICMP_NE:
1057 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
1058 return Builder->CreateICmp(LHSCC, Val, RHSCst);
1059 break; // (X u> 13 & X != 15) -> no change
1060 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
1061 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
1062 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
1066 case ICmpInst::ICMP_SGT:
1068 default: llvm_unreachable("Unknown integer condition code!");
1069 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
1070 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
1072 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
1074 case ICmpInst::ICMP_NE:
1075 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
1076 return Builder->CreateICmp(LHSCC, Val, RHSCst);
1077 break; // (X s> 13 & X != 15) -> no change
1078 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
1079 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true);
1080 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
1089 /// FoldAndOfFCmps - Optimize (fcmp)&(fcmp). NOTE: Unlike the rest of
1090 /// instcombine, this returns a Value which should already be inserted into the
1092 Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
1093 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
1094 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
1095 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType())
1098 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
1099 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
1100 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
1101 // If either of the constants are nans, then the whole thing returns
1103 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
1104 return Builder->getFalse();
1105 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
1108 // Handle vector zeros. This occurs because the canonical form of
1109 // "fcmp ord x,x" is "fcmp ord x, 0".
1110 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
1111 isa<ConstantAggregateZero>(RHS->getOperand(1)))
1112 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
1116 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
1117 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
1118 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
1121 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
1122 // Swap RHS operands to match LHS.
1123 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
1124 std::swap(Op1LHS, Op1RHS);
1127 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
1128 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
1130 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
1131 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
1132 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
1133 if (Op0CC == FCmpInst::FCMP_TRUE)
1135 if (Op1CC == FCmpInst::FCMP_TRUE)
1140 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
1141 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
1142 // uno && ord -> false
1143 if (Op0Pred == 0 && Op1Pred == 0 && Op0Ordered != Op1Ordered)
1144 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
1146 std::swap(LHS, RHS);
1147 std::swap(Op0Pred, Op1Pred);
1148 std::swap(Op0Ordered, Op1Ordered);
1151 // uno && ueq -> uno && (uno || eq) -> uno
1152 // ord && olt -> ord && (ord && lt) -> olt
1153 if (!Op0Ordered && (Op0Ordered == Op1Ordered))
1155 if (Op0Ordered && (Op0Ordered == Op1Ordered))
1158 // uno && oeq -> uno && (ord && eq) -> false
1160 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
1161 // ord && ueq -> ord && (uno || eq) -> oeq
1162 return getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS, Builder);
1169 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
1170 bool Changed = SimplifyAssociativeOrCommutative(I);
1171 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1173 if (Value *V = SimplifyVectorOp(I))
1174 return ReplaceInstUsesWith(I, V);
1176 if (Value *V = SimplifyAndInst(Op0, Op1, DL, TLI, DT, AT))
1177 return ReplaceInstUsesWith(I, V);
1179 // (A|B)&(A|C) -> A|(B&C) etc
1180 if (Value *V = SimplifyUsingDistributiveLaws(I))
1181 return ReplaceInstUsesWith(I, V);
1183 // See if we can simplify any instructions used by the instruction whose sole
1184 // purpose is to compute bits we don't care about.
1185 if (SimplifyDemandedInstructionBits(I))
1188 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
1189 const APInt &AndRHSMask = AndRHS->getValue();
1191 // Optimize a variety of ((val OP C1) & C2) combinations...
1192 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1193 Value *Op0LHS = Op0I->getOperand(0);
1194 Value *Op0RHS = Op0I->getOperand(1);
1195 switch (Op0I->getOpcode()) {
1197 case Instruction::Xor:
1198 case Instruction::Or: {
1199 // If the mask is only needed on one incoming arm, push it up.
1200 if (!Op0I->hasOneUse()) break;
1202 APInt NotAndRHS(~AndRHSMask);
1203 if (MaskedValueIsZero(Op0LHS, NotAndRHS, 0, &I)) {
1204 // Not masking anything out for the LHS, move to RHS.
1205 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
1206 Op0RHS->getName()+".masked");
1207 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
1209 if (!isa<Constant>(Op0RHS) &&
1210 MaskedValueIsZero(Op0RHS, NotAndRHS, 0, &I)) {
1211 // Not masking anything out for the RHS, move to LHS.
1212 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
1213 Op0LHS->getName()+".masked");
1214 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
1219 case Instruction::Add:
1220 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
1221 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1222 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1223 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
1224 return BinaryOperator::CreateAnd(V, AndRHS);
1225 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
1226 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
1229 case Instruction::Sub:
1230 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
1231 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1232 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1233 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
1234 return BinaryOperator::CreateAnd(V, AndRHS);
1236 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
1237 // has 1's for all bits that the subtraction with A might affect.
1238 if (Op0I->hasOneUse() && !match(Op0LHS, m_Zero())) {
1239 uint32_t BitWidth = AndRHSMask.getBitWidth();
1240 uint32_t Zeros = AndRHSMask.countLeadingZeros();
1241 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
1243 if (MaskedValueIsZero(Op0LHS, Mask, 0, &I)) {
1244 Value *NewNeg = Builder->CreateNeg(Op0RHS);
1245 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
1250 case Instruction::Shl:
1251 case Instruction::LShr:
1252 // (1 << x) & 1 --> zext(x == 0)
1253 // (1 >> x) & 1 --> zext(x == 0)
1254 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
1256 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
1257 return new ZExtInst(NewICmp, I.getType());
1262 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
1263 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
1267 // If this is an integer truncation, and if the source is an 'and' with
1268 // immediate, transform it. This frequently occurs for bitfield accesses.
1270 Value *X = nullptr; ConstantInt *YC = nullptr;
1271 if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) {
1272 // Change: and (trunc (and X, YC) to T), C2
1273 // into : and (trunc X to T), trunc(YC) & C2
1274 // This will fold the two constants together, which may allow
1275 // other simplifications.
1276 Value *NewCast = Builder->CreateTrunc(X, I.getType(), "and.shrunk");
1277 Constant *C3 = ConstantExpr::getTrunc(YC, I.getType());
1278 C3 = ConstantExpr::getAnd(C3, AndRHS);
1279 return BinaryOperator::CreateAnd(NewCast, C3);
1283 // Try to fold constant and into select arguments.
1284 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1285 if (Instruction *R = FoldOpIntoSelect(I, SI))
1287 if (isa<PHINode>(Op0))
1288 if (Instruction *NV = FoldOpIntoPhi(I))
1293 // (~A & ~B) == (~(A | B)) - De Morgan's Law
1294 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1295 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1296 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1297 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
1298 I.getName()+".demorgan");
1299 return BinaryOperator::CreateNot(Or);
1303 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
1304 // (A|B) & ~(A&B) -> A^B
1305 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1306 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1307 ((A == C && B == D) || (A == D && B == C)))
1308 return BinaryOperator::CreateXor(A, B);
1310 // ~(A&B) & (A|B) -> A^B
1311 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1312 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1313 ((A == C && B == D) || (A == D && B == C)))
1314 return BinaryOperator::CreateXor(A, B);
1316 // A&(A^B) => A & ~B
1318 Value *tmpOp0 = Op0;
1319 Value *tmpOp1 = Op1;
1320 if (Op0->hasOneUse() &&
1321 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
1322 if (A == Op1 || B == Op1 ) {
1329 if (tmpOp1->hasOneUse() &&
1330 match(tmpOp1, m_Xor(m_Value(A), m_Value(B)))) {
1334 // Notice that the patten (A&(~B)) is actually (A&(-1^B)), so if
1335 // A is originally -1 (or a vector of -1 and undefs), then we enter
1336 // an endless loop. By checking that A is non-constant we ensure that
1337 // we will never get to the loop.
1338 if (A == tmpOp0 && !isa<Constant>(A)) // A&(A^B) -> A & ~B
1339 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B));
1343 // (A&((~A)|B)) -> A&B
1344 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
1345 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
1346 return BinaryOperator::CreateAnd(A, Op1);
1347 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
1348 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
1349 return BinaryOperator::CreateAnd(A, Op0);
1351 // (A ^ B) & ((B ^ C) ^ A) -> (A ^ B) & ~C
1352 if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
1353 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
1354 if (Op1->hasOneUse() || cast<BinaryOperator>(Op1)->hasOneUse())
1355 return BinaryOperator::CreateAnd(Op0, Builder->CreateNot(C));
1357 // ((A ^ C) ^ B) & (B ^ A) -> (B ^ A) & ~C
1358 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
1359 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
1360 if (Op0->hasOneUse() || cast<BinaryOperator>(Op0)->hasOneUse())
1361 return BinaryOperator::CreateAnd(Op1, Builder->CreateNot(C));
1363 // (A | B) & ((~A) ^ B) -> (A & B)
1364 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1365 match(Op1, m_Xor(m_Not(m_Specific(A)), m_Specific(B))))
1366 return BinaryOperator::CreateAnd(A, B);
1368 // ((~A) ^ B) & (A | B) -> (A & B)
1369 if (match(Op0, m_Xor(m_Not(m_Value(A)), m_Value(B))) &&
1370 match(Op1, m_Or(m_Specific(A), m_Specific(B))))
1371 return BinaryOperator::CreateAnd(A, B);
1375 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
1376 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
1378 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1379 return ReplaceInstUsesWith(I, Res);
1381 // TODO: Make this recursive; it's a little tricky because an arbitrary
1382 // number of 'and' instructions might have to be created.
1384 if (LHS && match(Op1, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
1385 if (auto *Cmp = dyn_cast<ICmpInst>(X))
1386 if (Value *Res = FoldAndOfICmps(LHS, Cmp))
1387 return ReplaceInstUsesWith(I, Builder->CreateAnd(Res, Y));
1388 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
1389 if (Value *Res = FoldAndOfICmps(LHS, Cmp))
1390 return ReplaceInstUsesWith(I, Builder->CreateAnd(Res, X));
1392 if (RHS && match(Op0, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
1393 if (auto *Cmp = dyn_cast<ICmpInst>(X))
1394 if (Value *Res = FoldAndOfICmps(Cmp, RHS))
1395 return ReplaceInstUsesWith(I, Builder->CreateAnd(Res, Y));
1396 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
1397 if (Value *Res = FoldAndOfICmps(Cmp, RHS))
1398 return ReplaceInstUsesWith(I, Builder->CreateAnd(Res, X));
1402 // If and'ing two fcmp, try combine them into one.
1403 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1404 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1405 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1406 return ReplaceInstUsesWith(I, Res);
1409 // fold (and (cast A), (cast B)) -> (cast (and A, B))
1410 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
1411 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) {
1412 Type *SrcTy = Op0C->getOperand(0)->getType();
1413 if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ?
1414 SrcTy == Op1C->getOperand(0)->getType() &&
1415 SrcTy->isIntOrIntVectorTy()) {
1416 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
1418 // Only do this if the casts both really cause code to be generated.
1419 if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
1420 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
1421 Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName());
1422 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
1425 // If this is and(cast(icmp), cast(icmp)), try to fold this even if the
1426 // cast is otherwise not optimizable. This happens for vector sexts.
1427 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
1428 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
1429 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1430 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1432 // If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the
1433 // cast is otherwise not optimizable. This happens for vector sexts.
1434 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
1435 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
1436 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1437 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1443 bool OpsSwapped = false;
1444 // Canonicalize SExt or Not to the LHS
1445 if (match(Op1, m_SExt(m_Value())) ||
1446 match(Op1, m_Not(m_Value()))) {
1447 std::swap(Op0, Op1);
1451 // Fold (and (sext bool to A), B) --> (select bool, B, 0)
1452 if (match(Op0, m_SExt(m_Value(X))) &&
1453 X->getType()->getScalarType()->isIntegerTy(1)) {
1454 Value *Zero = Constant::getNullValue(Op1->getType());
1455 return SelectInst::Create(X, Op1, Zero);
1458 // Fold (and ~(sext bool to A), B) --> (select bool, 0, B)
1459 if (match(Op0, m_Not(m_SExt(m_Value(X)))) &&
1460 X->getType()->getScalarType()->isIntegerTy(1)) {
1461 Value *Zero = Constant::getNullValue(Op0->getType());
1462 return SelectInst::Create(X, Zero, Op1);
1466 std::swap(Op0, Op1);
1469 return Changed ? &I : nullptr;
1472 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
1473 /// capable of providing pieces of a bswap. The subexpression provides pieces
1474 /// of a bswap if it is proven that each of the non-zero bytes in the output of
1475 /// the expression came from the corresponding "byte swapped" byte in some other
1476 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
1477 /// we know that the expression deposits the low byte of %X into the high byte
1478 /// of the bswap result and that all other bytes are zero. This expression is
1479 /// accepted, the high byte of ByteValues is set to X to indicate a correct
1482 /// This function returns true if the match was unsuccessful and false if so.
1483 /// On entry to the function the "OverallLeftShift" is a signed integer value
1484 /// indicating the number of bytes that the subexpression is later shifted. For
1485 /// example, if the expression is later right shifted by 16 bits, the
1486 /// OverallLeftShift value would be -2 on entry. This is used to specify which
1487 /// byte of ByteValues is actually being set.
1489 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
1490 /// byte is masked to zero by a user. For example, in (X & 255), X will be
1491 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
1492 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
1493 /// always in the local (OverallLeftShift) coordinate space.
1495 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
1496 SmallVectorImpl<Value *> &ByteValues) {
1497 if (Instruction *I = dyn_cast<Instruction>(V)) {
1498 // If this is an or instruction, it may be an inner node of the bswap.
1499 if (I->getOpcode() == Instruction::Or) {
1500 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1502 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
1506 // If this is a logical shift by a constant multiple of 8, recurse with
1507 // OverallLeftShift and ByteMask adjusted.
1508 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
1510 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
1511 // Ensure the shift amount is defined and of a byte value.
1512 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
1515 unsigned ByteShift = ShAmt >> 3;
1516 if (I->getOpcode() == Instruction::Shl) {
1517 // X << 2 -> collect(X, +2)
1518 OverallLeftShift += ByteShift;
1519 ByteMask >>= ByteShift;
1521 // X >>u 2 -> collect(X, -2)
1522 OverallLeftShift -= ByteShift;
1523 ByteMask <<= ByteShift;
1524 ByteMask &= (~0U >> (32-ByteValues.size()));
1527 if (OverallLeftShift >= (int)ByteValues.size()) return true;
1528 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
1530 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1534 // If this is a logical 'and' with a mask that clears bytes, clear the
1535 // corresponding bytes in ByteMask.
1536 if (I->getOpcode() == Instruction::And &&
1537 isa<ConstantInt>(I->getOperand(1))) {
1538 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
1539 unsigned NumBytes = ByteValues.size();
1540 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
1541 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
1543 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
1544 // If this byte is masked out by a later operation, we don't care what
1546 if ((ByteMask & (1 << i)) == 0)
1549 // If the AndMask is all zeros for this byte, clear the bit.
1550 APInt MaskB = AndMask & Byte;
1552 ByteMask &= ~(1U << i);
1556 // If the AndMask is not all ones for this byte, it's not a bytezap.
1560 // Otherwise, this byte is kept.
1563 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1568 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
1569 // the input value to the bswap. Some observations: 1) if more than one byte
1570 // is demanded from this input, then it could not be successfully assembled
1571 // into a byteswap. At least one of the two bytes would not be aligned with
1572 // their ultimate destination.
1573 if (!isPowerOf2_32(ByteMask)) return true;
1574 unsigned InputByteNo = countTrailingZeros(ByteMask);
1576 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
1577 // is demanded, it needs to go into byte 0 of the result. This means that the
1578 // byte needs to be shifted until it lands in the right byte bucket. The
1579 // shift amount depends on the position: if the byte is coming from the high
1580 // part of the value (e.g. byte 3) then it must be shifted right. If from the
1581 // low part, it must be shifted left.
1582 unsigned DestByteNo = InputByteNo + OverallLeftShift;
1583 if (ByteValues.size()-1-DestByteNo != InputByteNo)
1586 // If the destination byte value is already defined, the values are or'd
1587 // together, which isn't a bswap (unless it's an or of the same bits).
1588 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
1590 ByteValues[DestByteNo] = V;
1594 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
1595 /// If so, insert the new bswap intrinsic and return it.
1596 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
1597 IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
1598 if (!ITy || ITy->getBitWidth() % 16 ||
1599 // ByteMask only allows up to 32-byte values.
1600 ITy->getBitWidth() > 32*8)
1601 return nullptr; // Can only bswap pairs of bytes. Can't do vectors.
1603 /// ByteValues - For each byte of the result, we keep track of which value
1604 /// defines each byte.
1605 SmallVector<Value*, 8> ByteValues;
1606 ByteValues.resize(ITy->getBitWidth()/8);
1608 // Try to find all the pieces corresponding to the bswap.
1609 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
1610 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
1613 // Check to see if all of the bytes come from the same value.
1614 Value *V = ByteValues[0];
1615 if (!V) return nullptr; // Didn't find a byte? Must be zero.
1617 // Check to make sure that all of the bytes come from the same value.
1618 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
1619 if (ByteValues[i] != V)
1621 Module *M = I.getParent()->getParent()->getParent();
1622 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, ITy);
1623 return CallInst::Create(F, V);
1626 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
1627 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
1628 /// we can simplify this expression to "cond ? C : D or B".
1629 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
1630 Value *C, Value *D) {
1631 // If A is not a select of -1/0, this cannot match.
1632 Value *Cond = nullptr;
1633 if (!match(A, m_SExt(m_Value(Cond))) ||
1634 !Cond->getType()->isIntegerTy(1))
1637 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
1638 if (match(D, m_Not(m_SExt(m_Specific(Cond)))))
1639 return SelectInst::Create(Cond, C, B);
1640 if (match(D, m_SExt(m_Not(m_Specific(Cond)))))
1641 return SelectInst::Create(Cond, C, B);
1643 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
1644 if (match(B, m_Not(m_SExt(m_Specific(Cond)))))
1645 return SelectInst::Create(Cond, C, D);
1646 if (match(B, m_SExt(m_Not(m_Specific(Cond)))))
1647 return SelectInst::Create(Cond, C, D);
1651 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
1652 Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS,
1653 Instruction *CxtI) {
1654 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
1656 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2)
1657 // if K1 and K2 are a one-bit mask.
1658 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
1659 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
1661 if (LHS->getPredicate() == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero() &&
1662 RHS->getPredicate() == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) {
1664 BinaryOperator *LAnd = dyn_cast<BinaryOperator>(LHS->getOperand(0));
1665 BinaryOperator *RAnd = dyn_cast<BinaryOperator>(RHS->getOperand(0));
1666 if (LAnd && RAnd && LAnd->hasOneUse() && RHS->hasOneUse() &&
1667 LAnd->getOpcode() == Instruction::And &&
1668 RAnd->getOpcode() == Instruction::And) {
1670 Value *Mask = nullptr;
1671 Value *Masked = nullptr;
1672 if (LAnd->getOperand(0) == RAnd->getOperand(0) &&
1673 isKnownToBeAPowerOfTwo(LAnd->getOperand(1), false, 0, AT, CxtI, DT) &&
1674 isKnownToBeAPowerOfTwo(RAnd->getOperand(1), false, 0, AT, CxtI, DT)) {
1675 Mask = Builder->CreateOr(LAnd->getOperand(1), RAnd->getOperand(1));
1676 Masked = Builder->CreateAnd(LAnd->getOperand(0), Mask);
1677 } else if (LAnd->getOperand(1) == RAnd->getOperand(1) &&
1678 isKnownToBeAPowerOfTwo(LAnd->getOperand(0),
1679 false, 0, AT, CxtI, DT) &&
1680 isKnownToBeAPowerOfTwo(RAnd->getOperand(0),
1681 false, 0, AT, CxtI, DT)) {
1682 Mask = Builder->CreateOr(LAnd->getOperand(0), RAnd->getOperand(0));
1683 Masked = Builder->CreateAnd(LAnd->getOperand(1), Mask);
1687 return Builder->CreateICmp(ICmpInst::ICMP_NE, Masked, Mask);
1691 // Fold (icmp ult/ule (A + C1), C3) | (icmp ult/ule (A + C2), C3)
1692 // --> (icmp ult/ule ((A & ~(C1 ^ C2)) + max(C1, C2)), C3)
1693 // The original condition actually refers to the following two ranges:
1694 // [MAX_UINT-C1+1, MAX_UINT-C1+1+C3] and [MAX_UINT-C2+1, MAX_UINT-C2+1+C3]
1695 // We can fold these two ranges if:
1696 // 1) C1 and C2 is unsigned greater than C3.
1697 // 2) The two ranges are separated.
1698 // 3) C1 ^ C2 is one-bit mask.
1699 // 4) LowRange1 ^ LowRange2 and HighRange1 ^ HighRange2 are one-bit mask.
1700 // This implies all values in the two ranges differ by exactly one bit.
1702 if ((LHSCC == ICmpInst::ICMP_ULT || LHSCC == ICmpInst::ICMP_ULE) &&
1703 LHSCC == RHSCC && LHSCst && RHSCst && LHS->hasOneUse() &&
1704 RHS->hasOneUse() && LHSCst->getType() == RHSCst->getType() &&
1705 LHSCst->getValue() == (RHSCst->getValue())) {
1707 Value *LAdd = LHS->getOperand(0);
1708 Value *RAdd = RHS->getOperand(0);
1710 Value *LAddOpnd, *RAddOpnd;
1711 ConstantInt *LAddCst, *RAddCst;
1712 if (match(LAdd, m_Add(m_Value(LAddOpnd), m_ConstantInt(LAddCst))) &&
1713 match(RAdd, m_Add(m_Value(RAddOpnd), m_ConstantInt(RAddCst))) &&
1714 LAddCst->getValue().ugt(LHSCst->getValue()) &&
1715 RAddCst->getValue().ugt(LHSCst->getValue())) {
1717 APInt DiffCst = LAddCst->getValue() ^ RAddCst->getValue();
1718 if (LAddOpnd == RAddOpnd && DiffCst.isPowerOf2()) {
1719 ConstantInt *MaxAddCst = nullptr;
1720 if (LAddCst->getValue().ult(RAddCst->getValue()))
1721 MaxAddCst = RAddCst;
1723 MaxAddCst = LAddCst;
1725 APInt RRangeLow = -RAddCst->getValue();
1726 APInt RRangeHigh = RRangeLow + LHSCst->getValue();
1727 APInt LRangeLow = -LAddCst->getValue();
1728 APInt LRangeHigh = LRangeLow + LHSCst->getValue();
1729 APInt LowRangeDiff = RRangeLow ^ LRangeLow;
1730 APInt HighRangeDiff = RRangeHigh ^ LRangeHigh;
1731 APInt RangeDiff = LRangeLow.sgt(RRangeLow) ? LRangeLow - RRangeLow
1732 : RRangeLow - LRangeLow;
1734 if (LowRangeDiff.isPowerOf2() && LowRangeDiff == HighRangeDiff &&
1735 RangeDiff.ugt(LHSCst->getValue())) {
1736 Value *MaskCst = ConstantInt::get(LAddCst->getType(), ~DiffCst);
1738 Value *NewAnd = Builder->CreateAnd(LAddOpnd, MaskCst);
1739 Value *NewAdd = Builder->CreateAdd(NewAnd, MaxAddCst);
1740 return (Builder->CreateICmp(LHS->getPredicate(), NewAdd, LHSCst));
1746 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
1747 if (PredicatesFoldable(LHSCC, RHSCC)) {
1748 if (LHS->getOperand(0) == RHS->getOperand(1) &&
1749 LHS->getOperand(1) == RHS->getOperand(0))
1750 LHS->swapOperands();
1751 if (LHS->getOperand(0) == RHS->getOperand(0) &&
1752 LHS->getOperand(1) == RHS->getOperand(1)) {
1753 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1754 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
1755 bool isSigned = LHS->isSigned() || RHS->isSigned();
1756 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
1760 // handle (roughly):
1761 // (icmp ne (A & B), C) | (icmp ne (A & D), E)
1762 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, false, Builder))
1765 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
1766 if (LHS->hasOneUse() || RHS->hasOneUse()) {
1767 // (icmp eq B, 0) | (icmp ult A, B) -> (icmp ule A, B-1)
1768 // (icmp eq B, 0) | (icmp ugt B, A) -> (icmp ule A, B-1)
1769 Value *A = nullptr, *B = nullptr;
1770 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero()) {
1772 if (RHSCC == ICmpInst::ICMP_ULT && Val == RHS->getOperand(1))
1774 else if (RHSCC == ICmpInst::ICMP_UGT && Val == Val2)
1775 A = RHS->getOperand(1);
1777 // (icmp ult A, B) | (icmp eq B, 0) -> (icmp ule A, B-1)
1778 // (icmp ugt B, A) | (icmp eq B, 0) -> (icmp ule A, B-1)
1779 else if (RHSCC == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) {
1781 if (LHSCC == ICmpInst::ICMP_ULT && Val2 == LHS->getOperand(1))
1783 else if (LHSCC == ICmpInst::ICMP_UGT && Val2 == Val)
1784 A = LHS->getOperand(1);
1787 return Builder->CreateICmp(
1789 Builder->CreateAdd(B, ConstantInt::getSigned(B->getType(), -1)), A);
1792 // E.g. (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
1793 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/true))
1796 // E.g. (icmp sgt x, n) | (icmp slt x, 0) --> icmp ugt x, n
1797 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/true))
1800 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
1801 if (!LHSCst || !RHSCst) return nullptr;
1803 if (LHSCst == RHSCst && LHSCC == RHSCC) {
1804 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
1805 if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
1806 Value *NewOr = Builder->CreateOr(Val, Val2);
1807 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
1811 // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
1812 // iff C2 + CA == C1.
1813 if (LHSCC == ICmpInst::ICMP_ULT && RHSCC == ICmpInst::ICMP_EQ) {
1814 ConstantInt *AddCst;
1815 if (match(Val, m_Add(m_Specific(Val2), m_ConstantInt(AddCst))))
1816 if (RHSCst->getValue() + AddCst->getValue() == LHSCst->getValue())
1817 return Builder->CreateICmpULE(Val, LHSCst);
1820 // From here on, we only handle:
1821 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
1822 if (Val != Val2) return nullptr;
1824 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
1825 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
1826 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
1827 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
1828 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
1831 // We can't fold (ugt x, C) | (sgt x, C2).
1832 if (!PredicatesFoldable(LHSCC, RHSCC))
1835 // Ensure that the larger constant is on the RHS.
1837 if (CmpInst::isSigned(LHSCC) ||
1838 (ICmpInst::isEquality(LHSCC) &&
1839 CmpInst::isSigned(RHSCC)))
1840 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
1842 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
1845 std::swap(LHS, RHS);
1846 std::swap(LHSCst, RHSCst);
1847 std::swap(LHSCC, RHSCC);
1850 // At this point, we know we have two icmp instructions
1851 // comparing a value against two constants and or'ing the result
1852 // together. Because of the above check, we know that we only have
1853 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
1854 // icmp folding check above), that the two constants are not
1856 assert(LHSCst != RHSCst && "Compares not folded above?");
1859 default: llvm_unreachable("Unknown integer condition code!");
1860 case ICmpInst::ICMP_EQ:
1862 default: llvm_unreachable("Unknown integer condition code!");
1863 case ICmpInst::ICMP_EQ:
1864 if (LHS->getOperand(0) == RHS->getOperand(0)) {
1865 // if LHSCst and RHSCst differ only by one bit:
1866 // (A == C1 || A == C2) -> (A & ~(C1 ^ C2)) == C1
1867 assert(LHSCst->getValue().ule(LHSCst->getValue()));
1869 APInt Xor = LHSCst->getValue() ^ RHSCst->getValue();
1870 if (Xor.isPowerOf2()) {
1871 Value *NegCst = Builder->getInt(~Xor);
1872 Value *And = Builder->CreateAnd(LHS->getOperand(0), NegCst);
1873 return Builder->CreateICmp(ICmpInst::ICMP_EQ, And, LHSCst);
1877 if (LHSCst == SubOne(RHSCst)) {
1878 // (X == 13 | X == 14) -> X-13 <u 2
1879 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1880 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1881 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
1882 return Builder->CreateICmpULT(Add, AddCST);
1885 break; // (X == 13 | X == 15) -> no change
1886 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
1887 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
1889 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
1890 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
1891 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
1895 case ICmpInst::ICMP_NE:
1897 default: llvm_unreachable("Unknown integer condition code!");
1898 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
1899 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
1900 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
1902 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
1903 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
1904 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
1905 return Builder->getTrue();
1907 case ICmpInst::ICMP_ULT:
1909 default: llvm_unreachable("Unknown integer condition code!");
1910 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
1912 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
1913 // If RHSCst is [us]MAXINT, it is always false. Not handling
1914 // this can cause overflow.
1915 if (RHSCst->isMaxValue(false))
1917 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false);
1918 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
1920 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
1921 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
1923 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
1927 case ICmpInst::ICMP_SLT:
1929 default: llvm_unreachable("Unknown integer condition code!");
1930 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
1932 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
1933 // If RHSCst is [us]MAXINT, it is always false. Not handling
1934 // this can cause overflow.
1935 if (RHSCst->isMaxValue(true))
1937 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false);
1938 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
1940 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
1941 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
1943 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
1947 case ICmpInst::ICMP_UGT:
1949 default: llvm_unreachable("Unknown integer condition code!");
1950 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
1951 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
1953 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
1955 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
1956 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
1957 return Builder->getTrue();
1958 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
1962 case ICmpInst::ICMP_SGT:
1964 default: llvm_unreachable("Unknown integer condition code!");
1965 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
1966 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
1968 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
1970 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
1971 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
1972 return Builder->getTrue();
1973 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
1981 /// FoldOrOfFCmps - Optimize (fcmp)|(fcmp). NOTE: Unlike the rest of
1982 /// instcombine, this returns a Value which should already be inserted into the
1984 Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
1985 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
1986 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
1987 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
1988 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
1989 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
1990 // If either of the constants are nans, then the whole thing returns
1992 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
1993 return Builder->getTrue();
1995 // Otherwise, no need to compare the two constants, compare the
1997 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
2000 // Handle vector zeros. This occurs because the canonical form of
2001 // "fcmp uno x,x" is "fcmp uno x, 0".
2002 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
2003 isa<ConstantAggregateZero>(RHS->getOperand(1)))
2004 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
2009 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
2010 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
2011 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
2013 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
2014 // Swap RHS operands to match LHS.
2015 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
2016 std::swap(Op1LHS, Op1RHS);
2018 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
2019 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
2021 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
2022 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
2023 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
2024 if (Op0CC == FCmpInst::FCMP_FALSE)
2026 if (Op1CC == FCmpInst::FCMP_FALSE)
2030 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
2031 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
2032 if (Op0Ordered == Op1Ordered) {
2033 // If both are ordered or unordered, return a new fcmp with
2034 // or'ed predicates.
2035 return getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS, Builder);
2041 /// FoldOrWithConstants - This helper function folds:
2043 /// ((A | B) & C1) | (B & C2)
2049 /// when the XOR of the two constants is "all ones" (-1).
2050 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
2051 Value *A, Value *B, Value *C) {
2052 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
2053 if (!CI1) return nullptr;
2055 Value *V1 = nullptr;
2056 ConstantInt *CI2 = nullptr;
2057 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return nullptr;
2059 APInt Xor = CI1->getValue() ^ CI2->getValue();
2060 if (!Xor.isAllOnesValue()) return nullptr;
2062 if (V1 == A || V1 == B) {
2063 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
2064 return BinaryOperator::CreateOr(NewOp, V1);
2070 /// \brief This helper function folds:
2072 /// ((A | B) & C1) ^ (B & C2)
2078 /// when the XOR of the two constants is "all ones" (-1).
2079 Instruction *InstCombiner::FoldXorWithConstants(BinaryOperator &I, Value *Op,
2080 Value *A, Value *B, Value *C) {
2081 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
2085 Value *V1 = nullptr;
2086 ConstantInt *CI2 = nullptr;
2087 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2))))
2090 APInt Xor = CI1->getValue() ^ CI2->getValue();
2091 if (!Xor.isAllOnesValue())
2094 if (V1 == A || V1 == B) {
2095 Value *NewOp = Builder->CreateAnd(V1 == A ? B : A, CI1);
2096 return BinaryOperator::CreateXor(NewOp, V1);
2102 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
2103 bool Changed = SimplifyAssociativeOrCommutative(I);
2104 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2106 if (Value *V = SimplifyVectorOp(I))
2107 return ReplaceInstUsesWith(I, V);
2109 if (Value *V = SimplifyOrInst(Op0, Op1, DL, TLI, DT, AT))
2110 return ReplaceInstUsesWith(I, V);
2112 // (A&B)|(A&C) -> A&(B|C) etc
2113 if (Value *V = SimplifyUsingDistributiveLaws(I))
2114 return ReplaceInstUsesWith(I, V);
2116 // See if we can simplify any instructions used by the instruction whose sole
2117 // purpose is to compute bits we don't care about.
2118 if (SimplifyDemandedInstructionBits(I))
2121 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2122 ConstantInt *C1 = nullptr; Value *X = nullptr;
2123 // (X & C1) | C2 --> (X | C2) & (C1|C2)
2124 // iff (C1 & C2) == 0.
2125 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
2126 (RHS->getValue() & C1->getValue()) != 0 &&
2128 Value *Or = Builder->CreateOr(X, RHS);
2130 return BinaryOperator::CreateAnd(Or,
2131 Builder->getInt(RHS->getValue() | C1->getValue()));
2134 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
2135 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
2137 Value *Or = Builder->CreateOr(X, RHS);
2139 return BinaryOperator::CreateXor(Or,
2140 Builder->getInt(C1->getValue() & ~RHS->getValue()));
2143 // Try to fold constant and into select arguments.
2144 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2145 if (Instruction *R = FoldOpIntoSelect(I, SI))
2148 if (isa<PHINode>(Op0))
2149 if (Instruction *NV = FoldOpIntoPhi(I))
2153 Value *A = nullptr, *B = nullptr;
2154 ConstantInt *C1 = nullptr, *C2 = nullptr;
2156 // (A | B) | C and A | (B | C) -> bswap if possible.
2157 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
2158 if (match(Op0, m_Or(m_Value(), m_Value())) ||
2159 match(Op1, m_Or(m_Value(), m_Value())) ||
2160 (match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
2161 match(Op1, m_LogicalShift(m_Value(), m_Value())))) {
2162 if (Instruction *BSwap = MatchBSwap(I))
2166 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
2167 if (Op0->hasOneUse() &&
2168 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2169 MaskedValueIsZero(Op1, C1->getValue(), 0, &I)) {
2170 Value *NOr = Builder->CreateOr(A, Op1);
2172 return BinaryOperator::CreateXor(NOr, C1);
2175 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
2176 if (Op1->hasOneUse() &&
2177 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2178 MaskedValueIsZero(Op0, C1->getValue(), 0, &I)) {
2179 Value *NOr = Builder->CreateOr(A, Op0);
2181 return BinaryOperator::CreateXor(NOr, C1);
2184 // ((~A & B) | A) -> (A | B)
2185 if (match(Op0, m_And(m_Not(m_Value(A)), m_Value(B))) &&
2186 match(Op1, m_Specific(A)))
2187 return BinaryOperator::CreateOr(A, B);
2189 // ((A & B) | ~A) -> (~A | B)
2190 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2191 match(Op1, m_Not(m_Specific(A))))
2192 return BinaryOperator::CreateOr(Builder->CreateNot(A), B);
2194 // (A & (~B)) | (A ^ B) -> (A ^ B)
2195 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2196 match(Op1, m_Xor(m_Specific(A), m_Specific(B))))
2197 return BinaryOperator::CreateXor(A, B);
2199 // (A ^ B) | ( A & (~B)) -> (A ^ B)
2200 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
2201 match(Op1, m_And(m_Specific(A), m_Not(m_Specific(B)))))
2202 return BinaryOperator::CreateXor(A, B);
2205 Value *C = nullptr, *D = nullptr;
2206 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
2207 match(Op1, m_And(m_Value(B), m_Value(D)))) {
2208 Value *V1 = nullptr, *V2 = nullptr;
2209 C1 = dyn_cast<ConstantInt>(C);
2210 C2 = dyn_cast<ConstantInt>(D);
2211 if (C1 && C2) { // (A & C1)|(B & C2)
2212 if ((C1->getValue() & C2->getValue()) == 0) {
2213 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
2214 // iff (C1&C2) == 0 and (N&~C1) == 0
2215 if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
2217 MaskedValueIsZero(V2, ~C1->getValue(), 0, &I)) || // (V|N)
2219 MaskedValueIsZero(V1, ~C1->getValue(), 0, &I)))) // (N|V)
2220 return BinaryOperator::CreateAnd(A,
2221 Builder->getInt(C1->getValue()|C2->getValue()));
2222 // Or commutes, try both ways.
2223 if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
2225 MaskedValueIsZero(V2, ~C2->getValue(), 0, &I)) || // (V|N)
2227 MaskedValueIsZero(V1, ~C2->getValue(), 0, &I)))) // (N|V)
2228 return BinaryOperator::CreateAnd(B,
2229 Builder->getInt(C1->getValue()|C2->getValue()));
2231 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
2232 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
2233 ConstantInt *C3 = nullptr, *C4 = nullptr;
2234 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
2235 (C3->getValue() & ~C1->getValue()) == 0 &&
2236 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
2237 (C4->getValue() & ~C2->getValue()) == 0) {
2238 V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
2239 return BinaryOperator::CreateAnd(V2,
2240 Builder->getInt(C1->getValue()|C2->getValue()));
2245 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants.
2246 // Don't do this for vector select idioms, the code generator doesn't handle
2248 if (!I.getType()->isVectorTy()) {
2249 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
2251 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
2253 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
2255 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
2259 // ((A&~B)|(~A&B)) -> A^B
2260 if ((match(C, m_Not(m_Specific(D))) &&
2261 match(B, m_Not(m_Specific(A)))))
2262 return BinaryOperator::CreateXor(A, D);
2263 // ((~B&A)|(~A&B)) -> A^B
2264 if ((match(A, m_Not(m_Specific(D))) &&
2265 match(B, m_Not(m_Specific(C)))))
2266 return BinaryOperator::CreateXor(C, D);
2267 // ((A&~B)|(B&~A)) -> A^B
2268 if ((match(C, m_Not(m_Specific(B))) &&
2269 match(D, m_Not(m_Specific(A)))))
2270 return BinaryOperator::CreateXor(A, B);
2271 // ((~B&A)|(B&~A)) -> A^B
2272 if ((match(A, m_Not(m_Specific(B))) &&
2273 match(D, m_Not(m_Specific(C)))))
2274 return BinaryOperator::CreateXor(C, B);
2276 // ((A|B)&1)|(B&-2) -> (A&1) | B
2277 if (match(A, m_Or(m_Value(V1), m_Specific(B))) ||
2278 match(A, m_Or(m_Specific(B), m_Value(V1)))) {
2279 Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C);
2280 if (Ret) return Ret;
2282 // (B&-2)|((A|B)&1) -> (A&1) | B
2283 if (match(B, m_Or(m_Specific(A), m_Value(V1))) ||
2284 match(B, m_Or(m_Value(V1), m_Specific(A)))) {
2285 Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D);
2286 if (Ret) return Ret;
2288 // ((A^B)&1)|(B&-2) -> (A&1) ^ B
2289 if (match(A, m_Xor(m_Value(V1), m_Specific(B))) ||
2290 match(A, m_Xor(m_Specific(B), m_Value(V1)))) {
2291 Instruction *Ret = FoldXorWithConstants(I, Op1, V1, B, C);
2292 if (Ret) return Ret;
2294 // (B&-2)|((A^B)&1) -> (A&1) ^ B
2295 if (match(B, m_Xor(m_Specific(A), m_Value(V1))) ||
2296 match(B, m_Xor(m_Value(V1), m_Specific(A)))) {
2297 Instruction *Ret = FoldXorWithConstants(I, Op0, A, V1, D);
2298 if (Ret) return Ret;
2302 // (A ^ B) | ((B ^ C) ^ A) -> (A ^ B) | C
2303 if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
2304 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
2305 if (Op1->hasOneUse() || cast<BinaryOperator>(Op1)->hasOneUse())
2306 return BinaryOperator::CreateOr(Op0, C);
2308 // ((A ^ C) ^ B) | (B ^ A) -> (B ^ A) | C
2309 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
2310 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
2311 if (Op0->hasOneUse() || cast<BinaryOperator>(Op0)->hasOneUse())
2312 return BinaryOperator::CreateOr(Op1, C);
2314 // ((B | C) & A) | B -> B | (A & C)
2315 if (match(Op0, m_And(m_Or(m_Specific(Op1), m_Value(C)), m_Value(A))))
2316 return BinaryOperator::CreateOr(Op1, Builder->CreateAnd(A, C));
2318 // (~A | ~B) == (~(A & B)) - De Morgan's Law
2319 if (Value *Op0NotVal = dyn_castNotVal(Op0))
2320 if (Value *Op1NotVal = dyn_castNotVal(Op1))
2321 if (Op0->hasOneUse() && Op1->hasOneUse()) {
2322 Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
2323 I.getName()+".demorgan");
2324 return BinaryOperator::CreateNot(And);
2327 // Canonicalize xor to the RHS.
2328 bool SwappedForXor = false;
2329 if (match(Op0, m_Xor(m_Value(), m_Value()))) {
2330 std::swap(Op0, Op1);
2331 SwappedForXor = true;
2334 // A | ( A ^ B) -> A | B
2335 // A | (~A ^ B) -> A | ~B
2336 // (A & B) | (A ^ B)
2337 if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
2338 if (Op0 == A || Op0 == B)
2339 return BinaryOperator::CreateOr(A, B);
2341 if (match(Op0, m_And(m_Specific(A), m_Specific(B))) ||
2342 match(Op0, m_And(m_Specific(B), m_Specific(A))))
2343 return BinaryOperator::CreateOr(A, B);
2345 if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
2346 Value *Not = Builder->CreateNot(B, B->getName()+".not");
2347 return BinaryOperator::CreateOr(Not, Op0);
2349 if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
2350 Value *Not = Builder->CreateNot(A, A->getName()+".not");
2351 return BinaryOperator::CreateOr(Not, Op0);
2355 // A | ~(A | B) -> A | ~B
2356 // A | ~(A ^ B) -> A | ~B
2357 if (match(Op1, m_Not(m_Value(A))))
2358 if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
2359 if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
2360 Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
2361 B->getOpcode() == Instruction::Xor)) {
2362 Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
2364 Value *Not = Builder->CreateNot(NotOp, NotOp->getName()+".not");
2365 return BinaryOperator::CreateOr(Not, Op0);
2368 // (A & B) | ((~A) ^ B) -> (~A ^ B)
2369 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2370 match(Op1, m_Xor(m_Not(m_Specific(A)), m_Specific(B))))
2371 return BinaryOperator::CreateXor(Builder->CreateNot(A), B);
2373 // ((~A) ^ B) | (A & B) -> (~A ^ B)
2374 if (match(Op0, m_Xor(m_Not(m_Value(A)), m_Value(B))) &&
2375 match(Op1, m_And(m_Specific(A), m_Specific(B))))
2376 return BinaryOperator::CreateXor(Builder->CreateNot(A), B);
2379 std::swap(Op0, Op1);
2382 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
2383 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
2385 if (Value *Res = FoldOrOfICmps(LHS, RHS, &I))
2386 return ReplaceInstUsesWith(I, Res);
2388 // TODO: Make this recursive; it's a little tricky because an arbitrary
2389 // number of 'or' instructions might have to be created.
2391 if (LHS && match(Op1, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2392 if (auto *Cmp = dyn_cast<ICmpInst>(X))
2393 if (Value *Res = FoldOrOfICmps(LHS, Cmp, &I))
2394 return ReplaceInstUsesWith(I, Builder->CreateOr(Res, Y));
2395 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2396 if (Value *Res = FoldOrOfICmps(LHS, Cmp, &I))
2397 return ReplaceInstUsesWith(I, Builder->CreateOr(Res, X));
2399 if (RHS && match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2400 if (auto *Cmp = dyn_cast<ICmpInst>(X))
2401 if (Value *Res = FoldOrOfICmps(Cmp, RHS, &I))
2402 return ReplaceInstUsesWith(I, Builder->CreateOr(Res, Y));
2403 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2404 if (Value *Res = FoldOrOfICmps(Cmp, RHS, &I))
2405 return ReplaceInstUsesWith(I, Builder->CreateOr(Res, X));
2409 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
2410 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
2411 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
2412 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
2413 return ReplaceInstUsesWith(I, Res);
2415 // fold (or (cast A), (cast B)) -> (cast (or A, B))
2416 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2417 CastInst *Op1C = dyn_cast<CastInst>(Op1);
2418 if (Op1C && Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
2419 Type *SrcTy = Op0C->getOperand(0)->getType();
2420 if (SrcTy == Op1C->getOperand(0)->getType() &&
2421 SrcTy->isIntOrIntVectorTy()) {
2422 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
2424 if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) &&
2425 // Only do this if the casts both really cause code to be
2427 ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
2428 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
2429 Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName());
2430 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2433 // If this is or(cast(icmp), cast(icmp)), try to fold this even if the
2434 // cast is otherwise not optimizable. This happens for vector sexts.
2435 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
2436 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
2437 if (Value *Res = FoldOrOfICmps(LHS, RHS, &I))
2438 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
2440 // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the
2441 // cast is otherwise not optimizable. This happens for vector sexts.
2442 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
2443 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
2444 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
2445 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
2450 // or(sext(A), B) -> A ? -1 : B where A is an i1
2451 // or(A, sext(B)) -> B ? -1 : A where B is an i1
2452 if (match(Op0, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
2453 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1);
2454 if (match(Op1, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
2455 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0);
2457 // Note: If we've gotten to the point of visiting the outer OR, then the
2458 // inner one couldn't be simplified. If it was a constant, then it won't
2459 // be simplified by a later pass either, so we try swapping the inner/outer
2460 // ORs in the hopes that we'll be able to simplify it this way.
2461 // (X|C) | V --> (X|V) | C
2462 if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
2463 match(Op0, m_Or(m_Value(A), m_ConstantInt(C1)))) {
2464 Value *Inner = Builder->CreateOr(A, Op1);
2465 Inner->takeName(Op0);
2466 return BinaryOperator::CreateOr(Inner, C1);
2469 // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D))
2470 // Since this OR statement hasn't been optimized further yet, we hope
2471 // that this transformation will allow the new ORs to be optimized.
2473 Value *X = nullptr, *Y = nullptr;
2474 if (Op0->hasOneUse() && Op1->hasOneUse() &&
2475 match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) &&
2476 match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) {
2477 Value *orTrue = Builder->CreateOr(A, C);
2478 Value *orFalse = Builder->CreateOr(B, D);
2479 return SelectInst::Create(X, orTrue, orFalse);
2483 return Changed ? &I : nullptr;
2486 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
2487 bool Changed = SimplifyAssociativeOrCommutative(I);
2488 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2490 if (Value *V = SimplifyVectorOp(I))
2491 return ReplaceInstUsesWith(I, V);
2493 if (Value *V = SimplifyXorInst(Op0, Op1, DL, TLI, DT, AT))
2494 return ReplaceInstUsesWith(I, V);
2496 // (A&B)^(A&C) -> A&(B^C) etc
2497 if (Value *V = SimplifyUsingDistributiveLaws(I))
2498 return ReplaceInstUsesWith(I, V);
2500 // See if we can simplify any instructions used by the instruction whose sole
2501 // purpose is to compute bits we don't care about.
2502 if (SimplifyDemandedInstructionBits(I))
2505 // Is this a ~ operation?
2506 if (Value *NotOp = dyn_castNotVal(&I)) {
2507 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
2508 if (Op0I->getOpcode() == Instruction::And ||
2509 Op0I->getOpcode() == Instruction::Or) {
2510 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
2511 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
2512 if (dyn_castNotVal(Op0I->getOperand(1)))
2513 Op0I->swapOperands();
2514 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
2516 Builder->CreateNot(Op0I->getOperand(1),
2517 Op0I->getOperand(1)->getName()+".not");
2518 if (Op0I->getOpcode() == Instruction::And)
2519 return BinaryOperator::CreateOr(Op0NotVal, NotY);
2520 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
2523 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
2524 // ~(X | Y) === (~X & ~Y) - De Morgan's Law
2525 if (isFreeToInvert(Op0I->getOperand(0)) &&
2526 isFreeToInvert(Op0I->getOperand(1))) {
2528 Builder->CreateNot(Op0I->getOperand(0), "notlhs");
2530 Builder->CreateNot(Op0I->getOperand(1), "notrhs");
2531 if (Op0I->getOpcode() == Instruction::And)
2532 return BinaryOperator::CreateOr(NotX, NotY);
2533 return BinaryOperator::CreateAnd(NotX, NotY);
2536 } else if (Op0I->getOpcode() == Instruction::AShr) {
2537 // ~(~X >>s Y) --> (X >>s Y)
2538 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0)))
2539 return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1));
2545 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2546 if (RHS->isOne() && Op0->hasOneUse())
2547 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
2548 if (CmpInst *CI = dyn_cast<CmpInst>(Op0))
2549 return CmpInst::Create(CI->getOpcode(),
2550 CI->getInversePredicate(),
2551 CI->getOperand(0), CI->getOperand(1));
2553 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
2554 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2555 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
2556 if (CI->hasOneUse() && Op0C->hasOneUse()) {
2557 Instruction::CastOps Opcode = Op0C->getOpcode();
2558 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
2559 (RHS == ConstantExpr::getCast(Opcode, Builder->getTrue(),
2560 Op0C->getDestTy()))) {
2561 CI->setPredicate(CI->getInversePredicate());
2562 return CastInst::Create(Opcode, CI, Op0C->getType());
2568 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2569 // ~(c-X) == X-c-1 == X+(-c-1)
2570 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
2571 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
2572 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
2573 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
2574 ConstantInt::get(I.getType(), 1));
2575 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
2578 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2579 if (Op0I->getOpcode() == Instruction::Add) {
2580 // ~(X-c) --> (-c-1)-X
2581 if (RHS->isAllOnesValue()) {
2582 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
2583 return BinaryOperator::CreateSub(
2584 ConstantExpr::getSub(NegOp0CI,
2585 ConstantInt::get(I.getType(), 1)),
2586 Op0I->getOperand(0));
2587 } else if (RHS->getValue().isSignBit()) {
2588 // (X + C) ^ signbit -> (X + C + signbit)
2589 Constant *C = Builder->getInt(RHS->getValue() + Op0CI->getValue());
2590 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
2593 } else if (Op0I->getOpcode() == Instruction::Or) {
2594 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
2595 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue(),
2597 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
2598 // Anything in both C1 and C2 is known to be zero, remove it from
2600 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
2601 NewRHS = ConstantExpr::getAnd(NewRHS,
2602 ConstantExpr::getNot(CommonBits));
2604 I.setOperand(0, Op0I->getOperand(0));
2605 I.setOperand(1, NewRHS);
2608 } else if (Op0I->getOpcode() == Instruction::LShr) {
2609 // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3)
2613 if (Op0I->hasOneUse() &&
2614 (E1 = dyn_cast<BinaryOperator>(Op0I->getOperand(0))) &&
2615 E1->getOpcode() == Instruction::Xor &&
2616 (C1 = dyn_cast<ConstantInt>(E1->getOperand(1)))) {
2617 // fold (C1 >> C2) ^ C3
2618 ConstantInt *C2 = Op0CI, *C3 = RHS;
2619 APInt FoldConst = C1->getValue().lshr(C2->getValue());
2620 FoldConst ^= C3->getValue();
2621 // Prepare the two operands.
2622 Value *Opnd0 = Builder->CreateLShr(E1->getOperand(0), C2);
2623 Opnd0->takeName(Op0I);
2624 cast<Instruction>(Opnd0)->setDebugLoc(I.getDebugLoc());
2625 Value *FoldVal = ConstantInt::get(Opnd0->getType(), FoldConst);
2627 return BinaryOperator::CreateXor(Opnd0, FoldVal);
2633 // Try to fold constant and into select arguments.
2634 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2635 if (Instruction *R = FoldOpIntoSelect(I, SI))
2637 if (isa<PHINode>(Op0))
2638 if (Instruction *NV = FoldOpIntoPhi(I))
2642 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
2645 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2646 if (A == Op0) { // B^(B|A) == (A|B)^B
2647 Op1I->swapOperands();
2649 std::swap(Op0, Op1);
2650 } else if (B == Op0) { // B^(A|B) == (A|B)^B
2651 I.swapOperands(); // Simplified below.
2652 std::swap(Op0, Op1);
2654 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
2656 if (A == Op0) { // A^(A&B) -> A^(B&A)
2657 Op1I->swapOperands();
2660 if (B == Op0) { // A^(B&A) -> (B&A)^A
2661 I.swapOperands(); // Simplified below.
2662 std::swap(Op0, Op1);
2667 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
2670 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2671 Op0I->hasOneUse()) {
2672 if (A == Op1) // (B|A)^B == (A|B)^B
2674 if (B == Op1) // (A|B)^B == A & ~B
2675 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1));
2676 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2678 if (A == Op1) // (A&B)^A -> (B&A)^A
2680 if (B == Op1 && // (B&A)^A == ~B & A
2681 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
2682 return BinaryOperator::CreateAnd(Builder->CreateNot(A), Op1);
2688 Value *A, *B, *C, *D;
2689 // (A & B)^(A | B) -> A ^ B
2690 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2691 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
2692 if ((A == C && B == D) || (A == D && B == C))
2693 return BinaryOperator::CreateXor(A, B);
2695 // (A | B)^(A & B) -> A ^ B
2696 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2697 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
2698 if ((A == C && B == D) || (A == D && B == C))
2699 return BinaryOperator::CreateXor(A, B);
2701 // (A | ~B) ^ (~A | B) -> A ^ B
2702 if (match(Op0I, m_Or(m_Value(A), m_Not(m_Value(B)))) &&
2703 match(Op1I, m_Or(m_Not(m_Specific(A)), m_Specific(B)))) {
2704 return BinaryOperator::CreateXor(A, B);
2706 // (~A | B) ^ (A | ~B) -> A ^ B
2707 if (match(Op0I, m_Or(m_Not(m_Value(A)), m_Value(B))) &&
2708 match(Op1I, m_Or(m_Specific(A), m_Not(m_Specific(B))))) {
2709 return BinaryOperator::CreateXor(A, B);
2711 // (A & ~B) ^ (~A & B) -> A ^ B
2712 if (match(Op0I, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2713 match(Op1I, m_And(m_Not(m_Specific(A)), m_Specific(B)))) {
2714 return BinaryOperator::CreateXor(A, B);
2716 // (~A & B) ^ (A & ~B) -> A ^ B
2717 if (match(Op0I, m_And(m_Not(m_Value(A)), m_Value(B))) &&
2718 match(Op1I, m_And(m_Specific(A), m_Not(m_Specific(B))))) {
2719 return BinaryOperator::CreateXor(A, B);
2721 // (A ^ C)^(A | B) -> ((~A) & B) ^ C
2722 if (match(Op0I, m_Xor(m_Value(D), m_Value(C))) &&
2723 match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2725 return BinaryOperator::CreateXor(
2726 Builder->CreateAnd(Builder->CreateNot(A), B), C);
2728 return BinaryOperator::CreateXor(
2729 Builder->CreateAnd(Builder->CreateNot(B), A), C);
2731 // (A | B)^(A ^ C) -> ((~A) & B) ^ C
2732 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2733 match(Op1I, m_Xor(m_Value(D), m_Value(C)))) {
2735 return BinaryOperator::CreateXor(
2736 Builder->CreateAnd(Builder->CreateNot(A), B), C);
2738 return BinaryOperator::CreateXor(
2739 Builder->CreateAnd(Builder->CreateNot(B), A), C);
2741 // (A & B) ^ (A ^ B) -> (A | B)
2742 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2743 match(Op1I, m_Xor(m_Specific(A), m_Specific(B))))
2744 return BinaryOperator::CreateOr(A, B);
2745 // (A ^ B) ^ (A & B) -> (A | B)
2746 if (match(Op0I, m_Xor(m_Value(A), m_Value(B))) &&
2747 match(Op1I, m_And(m_Specific(A), m_Specific(B))))
2748 return BinaryOperator::CreateOr(A, B);
2751 Value *A = nullptr, *B = nullptr;
2752 // (A & ~B) ^ (~A) -> ~(A & B)
2753 if (match(Op0, m_And(m_Value(A), m_Not(m_Value(B)))) &&
2754 match(Op1, m_Not(m_Specific(A))))
2755 return BinaryOperator::CreateNot(Builder->CreateAnd(A, B));
2757 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
2758 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
2759 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
2760 if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
2761 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2762 LHS->getOperand(1) == RHS->getOperand(0))
2763 LHS->swapOperands();
2764 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2765 LHS->getOperand(1) == RHS->getOperand(1)) {
2766 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2767 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
2768 bool isSigned = LHS->isSigned() || RHS->isSigned();
2769 return ReplaceInstUsesWith(I,
2770 getNewICmpValue(isSigned, Code, Op0, Op1,
2775 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
2776 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2777 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
2778 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
2779 Type *SrcTy = Op0C->getOperand(0)->getType();
2780 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() &&
2781 // Only do this if the casts both really cause code to be generated.
2782 ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0),
2784 ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0),
2786 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
2787 Op1C->getOperand(0), I.getName());
2788 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2793 return Changed ? &I : nullptr;