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/Intrinsics.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/Support/PatternMatch.h"
19 using namespace PatternMatch;
22 /// AddOne - Add one to a ConstantInt.
23 static Constant *AddOne(Constant *C) {
24 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
26 /// SubOne - Subtract one from a ConstantInt.
27 static Constant *SubOne(ConstantInt *C) {
28 return ConstantInt::get(C->getContext(), C->getValue()-1);
31 /// isFreeToInvert - Return true if the specified value is free to invert (apply
32 /// ~ to). This happens in cases where the ~ can be eliminated.
33 static inline bool isFreeToInvert(Value *V) {
35 if (BinaryOperator::isNot(V))
38 // Constants can be considered to be not'ed values.
39 if (isa<ConstantInt>(V))
42 // Compares can be inverted if they have a single use.
43 if (CmpInst *CI = dyn_cast<CmpInst>(V))
44 return CI->hasOneUse();
49 static inline Value *dyn_castNotVal(Value *V) {
50 // If this is not(not(x)) don't return that this is a not: we want the two
51 // not's to be folded first.
52 if (BinaryOperator::isNot(V)) {
53 Value *Operand = BinaryOperator::getNotArgument(V);
54 if (!isFreeToInvert(Operand))
58 // Constants can be considered to be not'ed values...
59 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
60 return ConstantInt::get(C->getType(), ~C->getValue());
65 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
66 /// are carefully arranged to allow folding of expressions such as:
68 /// (A < B) | (A > B) --> (A != B)
70 /// Note that this is only valid if the first and second predicates have the
71 /// same sign. Is illegal to do: (A u< B) | (A s> B)
73 /// Three bits are used to represent the condition, as follows:
78 /// <=> Value Definition
79 /// 000 0 Always false
88 static unsigned getICmpCode(const ICmpInst *ICI) {
89 switch (ICI->getPredicate()) {
91 case ICmpInst::ICMP_UGT: return 1; // 001
92 case ICmpInst::ICMP_SGT: return 1; // 001
93 case ICmpInst::ICMP_EQ: return 2; // 010
94 case ICmpInst::ICMP_UGE: return 3; // 011
95 case ICmpInst::ICMP_SGE: return 3; // 011
96 case ICmpInst::ICMP_ULT: return 4; // 100
97 case ICmpInst::ICMP_SLT: return 4; // 100
98 case ICmpInst::ICMP_NE: return 5; // 101
99 case ICmpInst::ICMP_ULE: return 6; // 110
100 case ICmpInst::ICMP_SLE: return 6; // 110
103 llvm_unreachable("Invalid ICmp predicate!");
108 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
109 /// predicate into a three bit mask. It also returns whether it is an ordered
110 /// predicate by reference.
111 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
114 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
115 case FCmpInst::FCMP_UNO: return 0; // 000
116 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
117 case FCmpInst::FCMP_UGT: return 1; // 001
118 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
119 case FCmpInst::FCMP_UEQ: return 2; // 010
120 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
121 case FCmpInst::FCMP_UGE: return 3; // 011
122 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
123 case FCmpInst::FCMP_ULT: return 4; // 100
124 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
125 case FCmpInst::FCMP_UNE: return 5; // 101
126 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
127 case FCmpInst::FCMP_ULE: return 6; // 110
130 // Not expecting FCMP_FALSE and FCMP_TRUE;
131 llvm_unreachable("Unexpected FCmp predicate!");
136 /// getICmpValue - This is the complement of getICmpCode, which turns an
137 /// opcode and two operands into either a constant true or false, or a brand
138 /// new ICmp instruction. The sign is passed in to determine which kind
139 /// of predicate to use in the new icmp instruction.
140 static Value *getICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS,
141 InstCombiner::BuilderTy *Builder) {
142 CmpInst::Predicate Pred;
144 default: assert(0 && "Illegal ICmp code!");
146 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
147 case 1: Pred = Sign ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; break;
148 case 2: Pred = ICmpInst::ICMP_EQ; break;
149 case 3: Pred = Sign ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; break;
150 case 4: Pred = Sign ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; break;
151 case 5: Pred = ICmpInst::ICMP_NE; break;
152 case 6: Pred = Sign ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; break;
154 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
156 return Builder->CreateICmp(Pred, LHS, RHS);
159 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
160 /// opcode and two operands into either a FCmp instruction. isordered is passed
161 /// in to determine which kind of predicate to use in the new fcmp instruction.
162 static Value *getFCmpValue(bool isordered, unsigned code,
163 Value *LHS, Value *RHS,
164 InstCombiner::BuilderTy *Builder) {
165 CmpInst::Predicate Pred;
167 default: assert(0 && "Illegal FCmp code!");
168 case 0: Pred = isordered ? FCmpInst::FCMP_ORD : FCmpInst::FCMP_UNO; break;
169 case 1: Pred = isordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; break;
170 case 2: Pred = isordered ? FCmpInst::FCMP_OEQ : FCmpInst::FCMP_UEQ; break;
171 case 3: Pred = isordered ? FCmpInst::FCMP_OGE : FCmpInst::FCMP_UGE; break;
172 case 4: Pred = isordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; break;
173 case 5: Pred = isordered ? FCmpInst::FCMP_ONE : FCmpInst::FCMP_UNE; break;
174 case 6: Pred = isordered ? FCmpInst::FCMP_OLE : FCmpInst::FCMP_ULE; break;
176 if (!isordered) return ConstantInt::getTrue(LHS->getContext());
177 Pred = FCmpInst::FCMP_ORD; break;
179 return Builder->CreateFCmp(Pred, LHS, RHS);
182 /// PredicatesFoldable - Return true if both predicates match sign or if at
183 /// least one of them is an equality comparison (which is signless).
184 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
185 return (CmpInst::isSigned(p1) == CmpInst::isSigned(p2)) ||
186 (CmpInst::isSigned(p1) && ICmpInst::isEquality(p2)) ||
187 (CmpInst::isSigned(p2) && ICmpInst::isEquality(p1));
190 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
191 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
192 // guaranteed to be a binary operator.
193 Instruction *InstCombiner::OptAndOp(Instruction *Op,
196 BinaryOperator &TheAnd) {
197 Value *X = Op->getOperand(0);
198 Constant *Together = 0;
200 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
202 switch (Op->getOpcode()) {
203 case Instruction::Xor:
204 if (Op->hasOneUse()) {
205 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
206 Value *And = Builder->CreateAnd(X, AndRHS);
208 return BinaryOperator::CreateXor(And, Together);
211 case Instruction::Or:
212 if (Op->hasOneUse()){
213 if (Together != OpRHS) {
214 // (X | C1) & C2 --> (X | (C1&C2)) & C2
215 Value *Or = Builder->CreateOr(X, Together);
217 return BinaryOperator::CreateAnd(Or, AndRHS);
220 ConstantInt *TogetherCI = dyn_cast<ConstantInt>(Together);
221 if (TogetherCI && !TogetherCI->isZero()){
222 // (X | C1) & C2 --> (X & (C2^(C1&C2))) | C1
223 // NOTE: This reduces the number of bits set in the & mask, which
224 // can expose opportunities for store narrowing.
225 Together = ConstantExpr::getXor(AndRHS, Together);
226 Value *And = Builder->CreateAnd(X, Together);
228 return BinaryOperator::CreateOr(And, OpRHS);
233 case Instruction::Add:
234 if (Op->hasOneUse()) {
235 // Adding a one to a single bit bit-field should be turned into an XOR
236 // of the bit. First thing to check is to see if this AND is with a
237 // single bit constant.
238 const APInt &AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
240 // If there is only one bit set.
241 if (AndRHSV.isPowerOf2()) {
242 // Ok, at this point, we know that we are masking the result of the
243 // ADD down to exactly one bit. If the constant we are adding has
244 // no bits set below this bit, then we can eliminate the ADD.
245 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
247 // Check to see if any bits below the one bit set in AndRHSV are set.
248 if ((AddRHS & (AndRHSV-1)) == 0) {
249 // If not, the only thing that can effect the output of the AND is
250 // the bit specified by AndRHSV. If that bit is set, the effect of
251 // the XOR is to toggle the bit. If it is clear, then the ADD has
253 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
254 TheAnd.setOperand(0, X);
257 // Pull the XOR out of the AND.
258 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
259 NewAnd->takeName(Op);
260 return BinaryOperator::CreateXor(NewAnd, AndRHS);
267 case Instruction::Shl: {
268 // We know that the AND will not produce any of the bits shifted in, so if
269 // the anded constant includes them, clear them now!
271 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
272 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
273 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
274 ConstantInt *CI = ConstantInt::get(AndRHS->getContext(),
275 AndRHS->getValue() & ShlMask);
277 if (CI->getValue() == ShlMask)
278 // Masking out bits that the shift already masks.
279 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
281 if (CI != AndRHS) { // Reducing bits set in and.
282 TheAnd.setOperand(1, CI);
287 case Instruction::LShr: {
288 // We know that the AND will not produce any of the bits shifted in, so if
289 // the anded constant includes them, clear them now! This only applies to
290 // unsigned shifts, because a signed shr may bring in set bits!
292 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
293 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
294 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
295 ConstantInt *CI = ConstantInt::get(Op->getContext(),
296 AndRHS->getValue() & ShrMask);
298 if (CI->getValue() == ShrMask)
299 // Masking out bits that the shift already masks.
300 return ReplaceInstUsesWith(TheAnd, Op);
303 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
308 case Instruction::AShr:
310 // See if this is shifting in some sign extension, then masking it out
312 if (Op->hasOneUse()) {
313 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
314 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
315 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
316 Constant *C = ConstantInt::get(Op->getContext(),
317 AndRHS->getValue() & ShrMask);
318 if (C == AndRHS) { // Masking out bits shifted in.
319 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
320 // Make the argument unsigned.
321 Value *ShVal = Op->getOperand(0);
322 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
323 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
332 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
333 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
334 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
335 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
336 /// insert new instructions.
337 Value *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
338 bool isSigned, bool Inside) {
339 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
340 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
341 "Lo is not <= Hi in range emission code!");
344 if (Lo == Hi) // Trivially false.
345 return ConstantInt::getFalse(V->getContext());
347 // V >= Min && V < Hi --> V < Hi
348 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
349 ICmpInst::Predicate pred = (isSigned ?
350 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
351 return Builder->CreateICmp(pred, V, Hi);
354 // Emit V-Lo <u Hi-Lo
355 Constant *NegLo = ConstantExpr::getNeg(Lo);
356 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
357 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
358 return Builder->CreateICmpULT(Add, UpperBound);
361 if (Lo == Hi) // Trivially true.
362 return ConstantInt::getTrue(V->getContext());
364 // V < Min || V >= Hi -> V > Hi-1
365 Hi = SubOne(cast<ConstantInt>(Hi));
366 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
367 ICmpInst::Predicate pred = (isSigned ?
368 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
369 return Builder->CreateICmp(pred, V, Hi);
372 // Emit V-Lo >u Hi-1-Lo
373 // Note that Hi has already had one subtracted from it, above.
374 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
375 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
376 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
377 return Builder->CreateICmpUGT(Add, LowerBound);
380 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
381 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
382 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
383 // not, since all 1s are not contiguous.
384 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
385 const APInt& V = Val->getValue();
386 uint32_t BitWidth = Val->getType()->getBitWidth();
387 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
389 // look for the first zero bit after the run of ones
390 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
391 // look for the first non-zero bit
392 ME = V.getActiveBits();
396 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
397 /// where isSub determines whether the operator is a sub. If we can fold one of
398 /// the following xforms:
400 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
401 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
402 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
404 /// return (A +/- B).
406 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
407 ConstantInt *Mask, bool isSub,
409 Instruction *LHSI = dyn_cast<Instruction>(LHS);
410 if (!LHSI || LHSI->getNumOperands() != 2 ||
411 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
413 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
415 switch (LHSI->getOpcode()) {
417 case Instruction::And:
418 if (ConstantExpr::getAnd(N, Mask) == Mask) {
419 // If the AndRHS is a power of two minus one (0+1+), this is simple.
420 if ((Mask->getValue().countLeadingZeros() +
421 Mask->getValue().countPopulation()) ==
422 Mask->getValue().getBitWidth())
425 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
426 // part, we don't need any explicit masks to take them out of A. If that
427 // is all N is, ignore it.
428 uint32_t MB = 0, ME = 0;
429 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
430 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
431 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
432 if (MaskedValueIsZero(RHS, Mask))
437 case Instruction::Or:
438 case Instruction::Xor:
439 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
440 if ((Mask->getValue().countLeadingZeros() +
441 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
442 && ConstantExpr::getAnd(N, Mask)->isNullValue())
448 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
449 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
452 /// enum for classifying (icmp eq (A & B), C) and (icmp ne (A & B), C)
453 /// One of A and B is considered the mask, the other the value. This is
454 /// described as the "AMask" or "BMask" part of the enum. If the enum
455 /// contains only "Mask", then both A and B can be considered masks.
456 /// If A is the mask, then it was proven, that (A & C) == C. This
457 /// is trivial if C == A, or C == 0. If both A and C are constants, this
458 /// proof is also easy.
459 /// For the following explanations we assume that A is the mask.
460 /// The part "AllOnes" declares, that the comparison is true only
461 /// if (A & B) == A, or all bits of A are set in B.
462 /// Example: (icmp eq (A & 3), 3) -> FoldMskICmp_AMask_AllOnes
463 /// The part "AllZeroes" declares, that the comparison is true only
464 /// if (A & B) == 0, or all bits of A are cleared in B.
465 /// Example: (icmp eq (A & 3), 0) -> FoldMskICmp_Mask_AllZeroes
466 /// The part "Mixed" declares, that (A & B) == C and C might or might not
467 /// contain any number of one bits and zero bits.
468 /// Example: (icmp eq (A & 3), 1) -> FoldMskICmp_AMask_Mixed
469 /// The Part "Not" means, that in above descriptions "==" should be replaced
471 /// Example: (icmp ne (A & 3), 3) -> FoldMskICmp_AMask_NotAllOnes
472 /// If the mask A contains a single bit, then the following is equivalent:
473 /// (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
474 /// (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
475 enum MaskedICmpType {
476 FoldMskICmp_AMask_AllOnes = 1,
477 FoldMskICmp_AMask_NotAllOnes = 2,
478 FoldMskICmp_BMask_AllOnes = 4,
479 FoldMskICmp_BMask_NotAllOnes = 8,
480 FoldMskICmp_Mask_AllZeroes = 16,
481 FoldMskICmp_Mask_NotAllZeroes = 32,
482 FoldMskICmp_AMask_Mixed = 64,
483 FoldMskICmp_AMask_NotMixed = 128,
484 FoldMskICmp_BMask_Mixed = 256,
485 FoldMskICmp_BMask_NotMixed = 512
488 /// return the set of pattern classes (from MaskedICmpType)
489 /// that (icmp SCC (A & B), C) satisfies
490 static unsigned getTypeOfMaskedICmp(Value* A, Value* B, Value* C,
491 ICmpInst::Predicate SCC)
493 ConstantInt *ACst = dyn_cast<ConstantInt>(A);
494 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
495 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
496 bool icmp_eq = (SCC == ICmpInst::ICMP_EQ);
497 bool icmp_abit = (ACst != 0 && !ACst->isZero() &&
498 ACst->getValue().isPowerOf2());
499 bool icmp_bbit = (BCst != 0 && !BCst->isZero() &&
500 BCst->getValue().isPowerOf2());
502 if (CCst != 0 && CCst->isZero()) {
503 // if C is zero, then both A and B qualify as mask
504 result |= (icmp_eq ? (FoldMskICmp_Mask_AllZeroes |
505 FoldMskICmp_Mask_AllZeroes |
506 FoldMskICmp_AMask_Mixed |
507 FoldMskICmp_BMask_Mixed)
508 : (FoldMskICmp_Mask_NotAllZeroes |
509 FoldMskICmp_Mask_NotAllZeroes |
510 FoldMskICmp_AMask_NotMixed |
511 FoldMskICmp_BMask_NotMixed));
513 result |= (icmp_eq ? (FoldMskICmp_AMask_NotAllOnes |
514 FoldMskICmp_AMask_NotMixed)
515 : (FoldMskICmp_AMask_AllOnes |
516 FoldMskICmp_AMask_Mixed));
518 result |= (icmp_eq ? (FoldMskICmp_BMask_NotAllOnes |
519 FoldMskICmp_BMask_NotMixed)
520 : (FoldMskICmp_BMask_AllOnes |
521 FoldMskICmp_BMask_Mixed));
525 result |= (icmp_eq ? (FoldMskICmp_AMask_AllOnes |
526 FoldMskICmp_AMask_Mixed)
527 : (FoldMskICmp_AMask_NotAllOnes |
528 FoldMskICmp_AMask_NotMixed));
530 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
531 FoldMskICmp_AMask_NotMixed)
532 : (FoldMskICmp_Mask_AllZeroes |
533 FoldMskICmp_AMask_Mixed));
535 else if (ACst != 0 && CCst != 0 &&
536 ConstantExpr::getAnd(ACst, CCst) == CCst) {
537 result |= (icmp_eq ? FoldMskICmp_AMask_Mixed
538 : FoldMskICmp_AMask_NotMixed);
542 result |= (icmp_eq ? (FoldMskICmp_BMask_AllOnes |
543 FoldMskICmp_BMask_Mixed)
544 : (FoldMskICmp_BMask_NotAllOnes |
545 FoldMskICmp_BMask_NotMixed));
547 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
548 FoldMskICmp_BMask_NotMixed)
549 : (FoldMskICmp_Mask_AllZeroes |
550 FoldMskICmp_BMask_Mixed));
552 else if (BCst != 0 && CCst != 0 &&
553 ConstantExpr::getAnd(BCst, CCst) == CCst) {
554 result |= (icmp_eq ? FoldMskICmp_BMask_Mixed
555 : FoldMskICmp_BMask_NotMixed);
560 /// foldLogOpOfMaskedICmpsHelper:
561 /// handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
562 /// return the set of pattern classes (from MaskedICmpType)
563 /// that both LHS and RHS satisfy
564 static unsigned foldLogOpOfMaskedICmpsHelper(Value*& A,
565 Value*& B, Value*& C,
566 Value*& D, Value*& E,
567 ICmpInst *LHS, ICmpInst *RHS) {
568 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
569 if (LHSCC != ICmpInst::ICMP_EQ && LHSCC != ICmpInst::ICMP_NE) return 0;
570 if (RHSCC != ICmpInst::ICMP_EQ && RHSCC != ICmpInst::ICMP_NE) return 0;
571 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) return 0;
572 // vectors are not (yet?) supported
573 if (LHS->getOperand(0)->getType()->isVectorTy()) return 0;
575 // Here comes the tricky part:
576 // LHS might be of the form L11 & L12 == X, X == L21 & L22,
577 // and L11 & L12 == L21 & L22. The same goes for RHS.
578 // Now we must find those components L** and R**, that are equal, so
579 // that we can extract the parameters A, B, C, D, and E for the canonical
581 Value *L1 = LHS->getOperand(0);
582 Value *L2 = LHS->getOperand(1);
583 Value *L11,*L12,*L21,*L22;
584 if (match(L1, m_And(m_Value(L11), m_Value(L12)))) {
585 if (!match(L2, m_And(m_Value(L21), m_Value(L22))))
589 if (!match(L2, m_And(m_Value(L11), m_Value(L12))))
595 Value *R1 = RHS->getOperand(0);
596 Value *R2 = RHS->getOperand(1);
599 if (match(R1, m_And(m_Value(R11), m_Value(R12)))) {
600 if (R11 != 0 && (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22)) {
601 A = R11; D = R12; E = R2; ok = true;
604 if (R12 != 0 && (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22)) {
605 A = R12; D = R11; E = R2; ok = true;
608 if (!ok && match(R2, m_And(m_Value(R11), m_Value(R12)))) {
609 if (R11 != 0 && (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22)) {
610 A = R11; D = R12; E = R1; ok = true;
613 if (R12 != 0 && (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22)) {
614 A = R12; D = R11; E = R1; ok = true;
635 unsigned left_type = getTypeOfMaskedICmp(A, B, C, LHSCC);
636 unsigned right_type = getTypeOfMaskedICmp(A, D, E, RHSCC);
637 return left_type & right_type;
639 /// foldLogOpOfMaskedICmps:
640 /// try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
641 /// into a single (icmp(A & X) ==/!= Y)
642 static Value* foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS,
643 ICmpInst::Predicate NEWCC,
644 llvm::InstCombiner::BuilderTy* Builder) {
645 Value *A = 0, *B = 0, *C = 0, *D = 0, *E = 0;
646 unsigned mask = foldLogOpOfMaskedICmpsHelper(A, B, C, D, E, LHS, RHS);
647 if (mask == 0) return 0;
649 if (NEWCC == ICmpInst::ICMP_NE)
650 mask >>= 1; // treat "Not"-states as normal states
652 if (mask & FoldMskICmp_Mask_AllZeroes) {
653 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
654 // -> (icmp eq (A & (B|D)), 0)
655 Value* newOr = Builder->CreateOr(B, D);
656 Value* newAnd = Builder->CreateAnd(A, newOr);
657 // we can't use C as zero, because we might actually handle
658 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
659 // with B and D, having a single bit set
660 Value* zero = Constant::getNullValue(A->getType());
661 return Builder->CreateICmp(NEWCC, newAnd, zero);
663 else if (mask & FoldMskICmp_BMask_AllOnes) {
664 // (icmp eq (A & B), B) & (icmp eq (A & D), D)
665 // -> (icmp eq (A & (B|D)), (B|D))
666 Value* newOr = Builder->CreateOr(B, D);
667 Value* newAnd = Builder->CreateAnd(A, newOr);
668 return Builder->CreateICmp(NEWCC, newAnd, newOr);
670 else if (mask & FoldMskICmp_AMask_AllOnes) {
671 // (icmp eq (A & B), A) & (icmp eq (A & D), A)
672 // -> (icmp eq (A & (B&D)), A)
673 Value* newAnd1 = Builder->CreateAnd(B, D);
674 Value* newAnd = Builder->CreateAnd(A, newAnd1);
675 return Builder->CreateICmp(NEWCC, newAnd, A);
677 else if (mask & FoldMskICmp_BMask_Mixed) {
678 // (icmp eq (A & B), C) & (icmp eq (A & D), E)
679 // We already know that B & C == C && D & E == E.
680 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
681 // C and E, which are shared by both the mask B and the mask D, don't
682 // contradict, then we can transform to
683 // -> (icmp eq (A & (B|D)), (C|E))
684 // Currently, we only handle the case of B, C, D, and E being constant.
685 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
686 if (BCst == 0) return 0;
687 ConstantInt *DCst = dyn_cast<ConstantInt>(D);
688 if (DCst == 0) return 0;
689 // we can't simply use C and E, because we might actually handle
690 // (icmp ne (A & B), B) & (icmp eq (A & D), D)
691 // with B and D, having a single bit set
693 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
694 if (CCst == 0) return 0;
695 if (LHS->getPredicate() != NEWCC)
696 CCst = dyn_cast<ConstantInt>( ConstantExpr::getXor(BCst, CCst) );
697 ConstantInt *ECst = dyn_cast<ConstantInt>(E);
698 if (ECst == 0) return 0;
699 if (RHS->getPredicate() != NEWCC)
700 ECst = dyn_cast<ConstantInt>( ConstantExpr::getXor(DCst, ECst) );
701 ConstantInt* MCst = dyn_cast<ConstantInt>(
702 ConstantExpr::getAnd(ConstantExpr::getAnd(BCst, DCst),
703 ConstantExpr::getXor(CCst, ECst)) );
704 // if there is a conflict we should actually return a false for the
708 Value *newOr1 = Builder->CreateOr(B, D);
709 Value *newOr2 = ConstantExpr::getOr(CCst, ECst);
710 Value *newAnd = Builder->CreateAnd(A, newOr1);
711 return Builder->CreateICmp(NEWCC, newAnd, newOr2);
716 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
717 Value *InstCombiner::FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
718 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
720 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
721 if (PredicatesFoldable(LHSCC, RHSCC)) {
722 if (LHS->getOperand(0) == RHS->getOperand(1) &&
723 LHS->getOperand(1) == RHS->getOperand(0))
725 if (LHS->getOperand(0) == RHS->getOperand(0) &&
726 LHS->getOperand(1) == RHS->getOperand(1)) {
727 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
728 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
729 bool isSigned = LHS->isSigned() || RHS->isSigned();
730 return getICmpValue(isSigned, Code, Op0, Op1, Builder);
734 // handle (roughly): (icmp eq (A & B), C) & (icmp eq (A & D), E)
735 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, ICmpInst::ICMP_EQ, Builder))
738 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
739 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
740 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
741 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
742 if (LHSCst == 0 || RHSCst == 0) return 0;
744 if (LHSCst == RHSCst && LHSCC == RHSCC) {
745 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
746 // where C is a power of 2
747 if (LHSCC == ICmpInst::ICMP_ULT &&
748 LHSCst->getValue().isPowerOf2()) {
749 Value *NewOr = Builder->CreateOr(Val, Val2);
750 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
753 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
754 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
755 Value *NewOr = Builder->CreateOr(Val, Val2);
756 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
760 // From here on, we only handle:
761 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
762 if (Val != Val2) return 0;
764 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
765 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
766 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
767 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
768 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
771 // We can't fold (ugt x, C) & (sgt x, C2).
772 if (!PredicatesFoldable(LHSCC, RHSCC))
775 // Ensure that the larger constant is on the RHS.
777 if (CmpInst::isSigned(LHSCC) ||
778 (ICmpInst::isEquality(LHSCC) &&
779 CmpInst::isSigned(RHSCC)))
780 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
782 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
786 std::swap(LHSCst, RHSCst);
787 std::swap(LHSCC, RHSCC);
790 // At this point, we know we have two icmp instructions
791 // comparing a value against two constants and and'ing the result
792 // together. Because of the above check, we know that we only have
793 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
794 // (from the icmp folding check above), that the two constants
795 // are not equal and that the larger constant is on the RHS
796 assert(LHSCst != RHSCst && "Compares not folded above?");
799 default: llvm_unreachable("Unknown integer condition code!");
800 case ICmpInst::ICMP_EQ:
802 default: llvm_unreachable("Unknown integer condition code!");
803 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
804 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
805 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
806 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
807 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
808 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
809 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
812 case ICmpInst::ICMP_NE:
814 default: llvm_unreachable("Unknown integer condition code!");
815 case ICmpInst::ICMP_ULT:
816 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
817 return Builder->CreateICmpULT(Val, LHSCst);
818 break; // (X != 13 & X u< 15) -> no change
819 case ICmpInst::ICMP_SLT:
820 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
821 return Builder->CreateICmpSLT(Val, LHSCst);
822 break; // (X != 13 & X s< 15) -> no change
823 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
824 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
825 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
827 case ICmpInst::ICMP_NE:
828 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
829 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
830 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
831 return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1));
833 break; // (X != 13 & X != 15) -> no change
836 case ICmpInst::ICMP_ULT:
838 default: llvm_unreachable("Unknown integer condition code!");
839 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
840 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
841 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
842 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
844 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
845 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
847 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
851 case ICmpInst::ICMP_SLT:
853 default: llvm_unreachable("Unknown integer condition code!");
854 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
855 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
856 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
857 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
859 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
860 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
862 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
866 case ICmpInst::ICMP_UGT:
868 default: llvm_unreachable("Unknown integer condition code!");
869 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
870 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
872 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
874 case ICmpInst::ICMP_NE:
875 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
876 return Builder->CreateICmp(LHSCC, Val, RHSCst);
877 break; // (X u> 13 & X != 15) -> no change
878 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
879 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
880 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
884 case ICmpInst::ICMP_SGT:
886 default: llvm_unreachable("Unknown integer condition code!");
887 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
888 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
890 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
892 case ICmpInst::ICMP_NE:
893 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
894 return Builder->CreateICmp(LHSCC, Val, RHSCst);
895 break; // (X s> 13 & X != 15) -> no change
896 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
897 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true);
898 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
907 /// FoldAndOfFCmps - Optimize (fcmp)&(fcmp). NOTE: Unlike the rest of
908 /// instcombine, this returns a Value which should already be inserted into the
910 Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
911 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
912 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
913 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
914 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
915 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
916 // If either of the constants are nans, then the whole thing returns
918 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
919 return ConstantInt::getFalse(LHS->getContext());
920 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
923 // Handle vector zeros. This occurs because the canonical form of
924 // "fcmp ord x,x" is "fcmp ord x, 0".
925 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
926 isa<ConstantAggregateZero>(RHS->getOperand(1)))
927 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
931 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
932 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
933 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
936 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
937 // Swap RHS operands to match LHS.
938 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
939 std::swap(Op1LHS, Op1RHS);
942 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
943 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
945 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
946 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
947 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
948 if (Op0CC == FCmpInst::FCMP_TRUE)
950 if (Op1CC == FCmpInst::FCMP_TRUE)
955 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
956 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
959 std::swap(Op0Pred, Op1Pred);
960 std::swap(Op0Ordered, Op1Ordered);
963 // uno && ueq -> uno && (uno || eq) -> ueq
964 // ord && olt -> ord && (ord && lt) -> olt
965 if (Op0Ordered == Op1Ordered)
968 // uno && oeq -> uno && (ord && eq) -> false
969 // uno && ord -> false
971 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
972 // ord && ueq -> ord && (uno || eq) -> oeq
973 return getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS, Builder);
981 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
982 bool Changed = SimplifyAssociativeOrCommutative(I);
983 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
985 if (Value *V = SimplifyAndInst(Op0, Op1, TD))
986 return ReplaceInstUsesWith(I, V);
988 // (A|B)&(A|C) -> A|(B&C) etc
989 if (Value *V = SimplifyUsingDistributiveLaws(I))
990 return ReplaceInstUsesWith(I, V);
992 // See if we can simplify any instructions used by the instruction whose sole
993 // purpose is to compute bits we don't care about.
994 if (SimplifyDemandedInstructionBits(I))
997 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
998 const APInt &AndRHSMask = AndRHS->getValue();
1000 // Optimize a variety of ((val OP C1) & C2) combinations...
1001 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1002 Value *Op0LHS = Op0I->getOperand(0);
1003 Value *Op0RHS = Op0I->getOperand(1);
1004 switch (Op0I->getOpcode()) {
1006 case Instruction::Xor:
1007 case Instruction::Or: {
1008 // If the mask is only needed on one incoming arm, push it up.
1009 if (!Op0I->hasOneUse()) break;
1011 APInt NotAndRHS(~AndRHSMask);
1012 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
1013 // Not masking anything out for the LHS, move to RHS.
1014 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
1015 Op0RHS->getName()+".masked");
1016 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
1018 if (!isa<Constant>(Op0RHS) &&
1019 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
1020 // Not masking anything out for the RHS, move to LHS.
1021 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
1022 Op0LHS->getName()+".masked");
1023 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
1028 case Instruction::Add:
1029 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
1030 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1031 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1032 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
1033 return BinaryOperator::CreateAnd(V, AndRHS);
1034 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
1035 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
1038 case Instruction::Sub:
1039 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
1040 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1041 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1042 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
1043 return BinaryOperator::CreateAnd(V, AndRHS);
1045 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
1046 // has 1's for all bits that the subtraction with A might affect.
1047 if (Op0I->hasOneUse() && !match(Op0LHS, m_Zero())) {
1048 uint32_t BitWidth = AndRHSMask.getBitWidth();
1049 uint32_t Zeros = AndRHSMask.countLeadingZeros();
1050 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
1052 if (MaskedValueIsZero(Op0LHS, Mask)) {
1053 Value *NewNeg = Builder->CreateNeg(Op0RHS);
1054 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
1059 case Instruction::Shl:
1060 case Instruction::LShr:
1061 // (1 << x) & 1 --> zext(x == 0)
1062 // (1 >> x) & 1 --> zext(x == 0)
1063 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
1065 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
1066 return new ZExtInst(NewICmp, I.getType());
1071 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
1072 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
1076 // If this is an integer truncation, and if the source is an 'and' with
1077 // immediate, transform it. This frequently occurs for bitfield accesses.
1079 Value *X = 0; ConstantInt *YC = 0;
1080 if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) {
1081 // Change: and (trunc (and X, YC) to T), C2
1082 // into : and (trunc X to T), trunc(YC) & C2
1083 // This will fold the two constants together, which may allow
1084 // other simplifications.
1085 Value *NewCast = Builder->CreateTrunc(X, I.getType(), "and.shrunk");
1086 Constant *C3 = ConstantExpr::getTrunc(YC, I.getType());
1087 C3 = ConstantExpr::getAnd(C3, AndRHS);
1088 return BinaryOperator::CreateAnd(NewCast, C3);
1092 // Try to fold constant and into select arguments.
1093 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1094 if (Instruction *R = FoldOpIntoSelect(I, SI))
1096 if (isa<PHINode>(Op0))
1097 if (Instruction *NV = FoldOpIntoPhi(I))
1102 // (~A & ~B) == (~(A | B)) - De Morgan's Law
1103 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1104 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1105 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1106 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
1107 I.getName()+".demorgan");
1108 return BinaryOperator::CreateNot(Or);
1112 Value *A = 0, *B = 0, *C = 0, *D = 0;
1113 // (A|B) & ~(A&B) -> A^B
1114 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1115 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1116 ((A == C && B == D) || (A == D && B == C)))
1117 return BinaryOperator::CreateXor(A, B);
1119 // ~(A&B) & (A|B) -> A^B
1120 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1121 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1122 ((A == C && B == D) || (A == D && B == C)))
1123 return BinaryOperator::CreateXor(A, B);
1125 if (Op0->hasOneUse() &&
1126 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
1127 if (A == Op1) { // (A^B)&A -> A&(A^B)
1128 I.swapOperands(); // Simplify below
1129 std::swap(Op0, Op1);
1130 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
1131 cast<BinaryOperator>(Op0)->swapOperands();
1132 I.swapOperands(); // Simplify below
1133 std::swap(Op0, Op1);
1137 if (Op1->hasOneUse() &&
1138 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
1139 if (B == Op0) { // B&(A^B) -> B&(B^A)
1140 cast<BinaryOperator>(Op1)->swapOperands();
1143 if (A == Op0) // A&(A^B) -> A & ~B
1144 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B, "tmp"));
1147 // (A&((~A)|B)) -> A&B
1148 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
1149 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
1150 return BinaryOperator::CreateAnd(A, Op1);
1151 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
1152 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
1153 return BinaryOperator::CreateAnd(A, Op0);
1156 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1))
1157 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
1158 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1159 return ReplaceInstUsesWith(I, Res);
1161 // If and'ing two fcmp, try combine them into one.
1162 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1163 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1164 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1165 return ReplaceInstUsesWith(I, Res);
1168 // fold (and (cast A), (cast B)) -> (cast (and A, B))
1169 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
1170 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) {
1171 const Type *SrcTy = Op0C->getOperand(0)->getType();
1172 if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ?
1173 SrcTy == Op1C->getOperand(0)->getType() &&
1174 SrcTy->isIntOrIntVectorTy()) {
1175 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
1177 // Only do this if the casts both really cause code to be generated.
1178 if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
1179 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
1180 Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName());
1181 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
1184 // If this is and(cast(icmp), cast(icmp)), try to fold this even if the
1185 // cast is otherwise not optimizable. This happens for vector sexts.
1186 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
1187 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
1188 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1189 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1191 // If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the
1192 // cast is otherwise not optimizable. This happens for vector sexts.
1193 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
1194 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
1195 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1196 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1200 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
1201 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
1202 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
1203 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
1204 SI0->getOperand(1) == SI1->getOperand(1) &&
1205 (SI0->hasOneUse() || SI1->hasOneUse())) {
1207 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
1209 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
1210 SI1->getOperand(1));
1214 return Changed ? &I : 0;
1217 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
1218 /// capable of providing pieces of a bswap. The subexpression provides pieces
1219 /// of a bswap if it is proven that each of the non-zero bytes in the output of
1220 /// the expression came from the corresponding "byte swapped" byte in some other
1221 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
1222 /// we know that the expression deposits the low byte of %X into the high byte
1223 /// of the bswap result and that all other bytes are zero. This expression is
1224 /// accepted, the high byte of ByteValues is set to X to indicate a correct
1227 /// This function returns true if the match was unsuccessful and false if so.
1228 /// On entry to the function the "OverallLeftShift" is a signed integer value
1229 /// indicating the number of bytes that the subexpression is later shifted. For
1230 /// example, if the expression is later right shifted by 16 bits, the
1231 /// OverallLeftShift value would be -2 on entry. This is used to specify which
1232 /// byte of ByteValues is actually being set.
1234 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
1235 /// byte is masked to zero by a user. For example, in (X & 255), X will be
1236 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
1237 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
1238 /// always in the local (OverallLeftShift) coordinate space.
1240 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
1241 SmallVector<Value*, 8> &ByteValues) {
1242 if (Instruction *I = dyn_cast<Instruction>(V)) {
1243 // If this is an or instruction, it may be an inner node of the bswap.
1244 if (I->getOpcode() == Instruction::Or) {
1245 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1247 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
1251 // If this is a logical shift by a constant multiple of 8, recurse with
1252 // OverallLeftShift and ByteMask adjusted.
1253 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
1255 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
1256 // Ensure the shift amount is defined and of a byte value.
1257 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
1260 unsigned ByteShift = ShAmt >> 3;
1261 if (I->getOpcode() == Instruction::Shl) {
1262 // X << 2 -> collect(X, +2)
1263 OverallLeftShift += ByteShift;
1264 ByteMask >>= ByteShift;
1266 // X >>u 2 -> collect(X, -2)
1267 OverallLeftShift -= ByteShift;
1268 ByteMask <<= ByteShift;
1269 ByteMask &= (~0U >> (32-ByteValues.size()));
1272 if (OverallLeftShift >= (int)ByteValues.size()) return true;
1273 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
1275 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1279 // If this is a logical 'and' with a mask that clears bytes, clear the
1280 // corresponding bytes in ByteMask.
1281 if (I->getOpcode() == Instruction::And &&
1282 isa<ConstantInt>(I->getOperand(1))) {
1283 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
1284 unsigned NumBytes = ByteValues.size();
1285 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
1286 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
1288 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
1289 // If this byte is masked out by a later operation, we don't care what
1291 if ((ByteMask & (1 << i)) == 0)
1294 // If the AndMask is all zeros for this byte, clear the bit.
1295 APInt MaskB = AndMask & Byte;
1297 ByteMask &= ~(1U << i);
1301 // If the AndMask is not all ones for this byte, it's not a bytezap.
1305 // Otherwise, this byte is kept.
1308 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1313 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
1314 // the input value to the bswap. Some observations: 1) if more than one byte
1315 // is demanded from this input, then it could not be successfully assembled
1316 // into a byteswap. At least one of the two bytes would not be aligned with
1317 // their ultimate destination.
1318 if (!isPowerOf2_32(ByteMask)) return true;
1319 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
1321 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
1322 // is demanded, it needs to go into byte 0 of the result. This means that the
1323 // byte needs to be shifted until it lands in the right byte bucket. The
1324 // shift amount depends on the position: if the byte is coming from the high
1325 // part of the value (e.g. byte 3) then it must be shifted right. If from the
1326 // low part, it must be shifted left.
1327 unsigned DestByteNo = InputByteNo + OverallLeftShift;
1328 if (InputByteNo < ByteValues.size()/2) {
1329 if (ByteValues.size()-1-DestByteNo != InputByteNo)
1332 if (ByteValues.size()-1-DestByteNo != InputByteNo)
1336 // If the destination byte value is already defined, the values are or'd
1337 // together, which isn't a bswap (unless it's an or of the same bits).
1338 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
1340 ByteValues[DestByteNo] = V;
1344 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
1345 /// If so, insert the new bswap intrinsic and return it.
1346 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
1347 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
1348 if (!ITy || ITy->getBitWidth() % 16 ||
1349 // ByteMask only allows up to 32-byte values.
1350 ITy->getBitWidth() > 32*8)
1351 return 0; // Can only bswap pairs of bytes. Can't do vectors.
1353 /// ByteValues - For each byte of the result, we keep track of which value
1354 /// defines each byte.
1355 SmallVector<Value*, 8> ByteValues;
1356 ByteValues.resize(ITy->getBitWidth()/8);
1358 // Try to find all the pieces corresponding to the bswap.
1359 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
1360 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
1363 // Check to see if all of the bytes come from the same value.
1364 Value *V = ByteValues[0];
1365 if (V == 0) return 0; // Didn't find a byte? Must be zero.
1367 // Check to make sure that all of the bytes come from the same value.
1368 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
1369 if (ByteValues[i] != V)
1371 const Type *Tys[] = { ITy };
1372 Module *M = I.getParent()->getParent()->getParent();
1373 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
1374 return CallInst::Create(F, V);
1377 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
1378 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
1379 /// we can simplify this expression to "cond ? C : D or B".
1380 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
1381 Value *C, Value *D) {
1382 // If A is not a select of -1/0, this cannot match.
1384 if (!match(A, m_SExt(m_Value(Cond))) ||
1385 !Cond->getType()->isIntegerTy(1))
1388 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
1389 if (match(D, m_Not(m_SExt(m_Specific(Cond)))))
1390 return SelectInst::Create(Cond, C, B);
1391 if (match(D, m_SExt(m_Not(m_Specific(Cond)))))
1392 return SelectInst::Create(Cond, C, B);
1394 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
1395 if (match(B, m_Not(m_SExt(m_Specific(Cond)))))
1396 return SelectInst::Create(Cond, C, D);
1397 if (match(B, m_SExt(m_Not(m_Specific(Cond)))))
1398 return SelectInst::Create(Cond, C, D);
1402 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
1403 Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
1404 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
1406 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
1407 if (PredicatesFoldable(LHSCC, RHSCC)) {
1408 if (LHS->getOperand(0) == RHS->getOperand(1) &&
1409 LHS->getOperand(1) == RHS->getOperand(0))
1410 LHS->swapOperands();
1411 if (LHS->getOperand(0) == RHS->getOperand(0) &&
1412 LHS->getOperand(1) == RHS->getOperand(1)) {
1413 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1414 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
1415 bool isSigned = LHS->isSigned() || RHS->isSigned();
1416 return getICmpValue(isSigned, Code, Op0, Op1, Builder);
1420 // handle (roughly):
1421 // (icmp ne (A & B), C) | (icmp ne (A & D), E)
1422 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, ICmpInst::ICMP_NE, Builder))
1425 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
1426 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
1427 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
1428 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
1429 if (LHSCst == 0 || RHSCst == 0) return 0;
1431 if (LHSCst == RHSCst && LHSCC == RHSCC) {
1432 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
1433 if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
1434 Value *NewOr = Builder->CreateOr(Val, Val2);
1435 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
1439 // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
1440 // iff C2 + CA == C1.
1441 if (LHSCC == ICmpInst::ICMP_ULT && RHSCC == ICmpInst::ICMP_EQ) {
1442 ConstantInt *AddCst;
1443 if (match(Val, m_Add(m_Specific(Val2), m_ConstantInt(AddCst))))
1444 if (RHSCst->getValue() + AddCst->getValue() == LHSCst->getValue())
1445 return Builder->CreateICmpULE(Val, LHSCst);
1448 // From here on, we only handle:
1449 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
1450 if (Val != Val2) return 0;
1452 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
1453 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
1454 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
1455 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
1456 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
1459 // We can't fold (ugt x, C) | (sgt x, C2).
1460 if (!PredicatesFoldable(LHSCC, RHSCC))
1463 // Ensure that the larger constant is on the RHS.
1465 if (CmpInst::isSigned(LHSCC) ||
1466 (ICmpInst::isEquality(LHSCC) &&
1467 CmpInst::isSigned(RHSCC)))
1468 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
1470 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
1473 std::swap(LHS, RHS);
1474 std::swap(LHSCst, RHSCst);
1475 std::swap(LHSCC, RHSCC);
1478 // At this point, we know we have two icmp instructions
1479 // comparing a value against two constants and or'ing the result
1480 // together. Because of the above check, we know that we only have
1481 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
1482 // icmp folding check above), that the two constants are not
1484 assert(LHSCst != RHSCst && "Compares not folded above?");
1487 default: llvm_unreachable("Unknown integer condition code!");
1488 case ICmpInst::ICMP_EQ:
1490 default: llvm_unreachable("Unknown integer condition code!");
1491 case ICmpInst::ICMP_EQ:
1492 if (LHSCst == SubOne(RHSCst)) {
1493 // (X == 13 | X == 14) -> X-13 <u 2
1494 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1495 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1496 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
1497 return Builder->CreateICmpULT(Add, AddCST);
1499 break; // (X == 13 | X == 15) -> no change
1500 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
1501 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
1503 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
1504 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
1505 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
1509 case ICmpInst::ICMP_NE:
1511 default: llvm_unreachable("Unknown integer condition code!");
1512 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
1513 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
1514 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
1516 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
1517 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
1518 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
1519 return ConstantInt::getTrue(LHS->getContext());
1522 case ICmpInst::ICMP_ULT:
1524 default: llvm_unreachable("Unknown integer condition code!");
1525 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
1527 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
1528 // If RHSCst is [us]MAXINT, it is always false. Not handling
1529 // this can cause overflow.
1530 if (RHSCst->isMaxValue(false))
1532 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false);
1533 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
1535 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
1536 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
1538 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
1542 case ICmpInst::ICMP_SLT:
1544 default: llvm_unreachable("Unknown integer condition code!");
1545 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
1547 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
1548 // If RHSCst is [us]MAXINT, it is always false. Not handling
1549 // this can cause overflow.
1550 if (RHSCst->isMaxValue(true))
1552 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false);
1553 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
1555 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
1556 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
1558 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
1562 case ICmpInst::ICMP_UGT:
1564 default: llvm_unreachable("Unknown integer condition code!");
1565 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
1566 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
1568 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
1570 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
1571 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
1572 return ConstantInt::getTrue(LHS->getContext());
1573 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
1577 case ICmpInst::ICMP_SGT:
1579 default: llvm_unreachable("Unknown integer condition code!");
1580 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
1581 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
1583 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
1585 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
1586 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
1587 return ConstantInt::getTrue(LHS->getContext());
1588 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
1596 /// FoldOrOfFCmps - Optimize (fcmp)|(fcmp). NOTE: Unlike the rest of
1597 /// instcombine, this returns a Value which should already be inserted into the
1599 Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
1600 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
1601 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
1602 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
1603 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
1604 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
1605 // If either of the constants are nans, then the whole thing returns
1607 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
1608 return ConstantInt::getTrue(LHS->getContext());
1610 // Otherwise, no need to compare the two constants, compare the
1612 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
1615 // Handle vector zeros. This occurs because the canonical form of
1616 // "fcmp uno x,x" is "fcmp uno x, 0".
1617 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
1618 isa<ConstantAggregateZero>(RHS->getOperand(1)))
1619 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
1624 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
1625 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
1626 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
1628 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
1629 // Swap RHS operands to match LHS.
1630 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
1631 std::swap(Op1LHS, Op1RHS);
1633 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
1634 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
1636 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
1637 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
1638 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
1639 if (Op0CC == FCmpInst::FCMP_FALSE)
1641 if (Op1CC == FCmpInst::FCMP_FALSE)
1645 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
1646 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
1647 if (Op0Ordered == Op1Ordered) {
1648 // If both are ordered or unordered, return a new fcmp with
1649 // or'ed predicates.
1650 return getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS, Builder);
1656 /// FoldOrWithConstants - This helper function folds:
1658 /// ((A | B) & C1) | (B & C2)
1664 /// when the XOR of the two constants is "all ones" (-1).
1665 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
1666 Value *A, Value *B, Value *C) {
1667 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
1671 ConstantInt *CI2 = 0;
1672 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
1674 APInt Xor = CI1->getValue() ^ CI2->getValue();
1675 if (!Xor.isAllOnesValue()) return 0;
1677 if (V1 == A || V1 == B) {
1678 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
1679 return BinaryOperator::CreateOr(NewOp, V1);
1685 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
1686 bool Changed = SimplifyAssociativeOrCommutative(I);
1687 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1689 if (Value *V = SimplifyOrInst(Op0, Op1, TD))
1690 return ReplaceInstUsesWith(I, V);
1692 // (A&B)|(A&C) -> A&(B|C) etc
1693 if (Value *V = SimplifyUsingDistributiveLaws(I))
1694 return ReplaceInstUsesWith(I, V);
1696 // See if we can simplify any instructions used by the instruction whose sole
1697 // purpose is to compute bits we don't care about.
1698 if (SimplifyDemandedInstructionBits(I))
1701 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1702 ConstantInt *C1 = 0; Value *X = 0;
1703 // (X & C1) | C2 --> (X | C2) & (C1|C2)
1704 // iff (C1 & C2) == 0.
1705 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
1706 (RHS->getValue() & C1->getValue()) != 0 &&
1708 Value *Or = Builder->CreateOr(X, RHS);
1710 return BinaryOperator::CreateAnd(Or,
1711 ConstantInt::get(I.getContext(),
1712 RHS->getValue() | C1->getValue()));
1715 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
1716 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
1718 Value *Or = Builder->CreateOr(X, RHS);
1720 return BinaryOperator::CreateXor(Or,
1721 ConstantInt::get(I.getContext(),
1722 C1->getValue() & ~RHS->getValue()));
1725 // Try to fold constant and into select arguments.
1726 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1727 if (Instruction *R = FoldOpIntoSelect(I, SI))
1730 if (isa<PHINode>(Op0))
1731 if (Instruction *NV = FoldOpIntoPhi(I))
1735 Value *A = 0, *B = 0;
1736 ConstantInt *C1 = 0, *C2 = 0;
1738 // (A | B) | C and A | (B | C) -> bswap if possible.
1739 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
1740 if (match(Op0, m_Or(m_Value(), m_Value())) ||
1741 match(Op1, m_Or(m_Value(), m_Value())) ||
1742 (match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
1743 match(Op1, m_LogicalShift(m_Value(), m_Value())))) {
1744 if (Instruction *BSwap = MatchBSwap(I))
1748 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
1749 if (Op0->hasOneUse() &&
1750 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
1751 MaskedValueIsZero(Op1, C1->getValue())) {
1752 Value *NOr = Builder->CreateOr(A, Op1);
1754 return BinaryOperator::CreateXor(NOr, C1);
1757 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
1758 if (Op1->hasOneUse() &&
1759 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
1760 MaskedValueIsZero(Op0, C1->getValue())) {
1761 Value *NOr = Builder->CreateOr(A, Op0);
1763 return BinaryOperator::CreateXor(NOr, C1);
1767 Value *C = 0, *D = 0;
1768 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1769 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1770 Value *V1 = 0, *V2 = 0;
1771 C1 = dyn_cast<ConstantInt>(C);
1772 C2 = dyn_cast<ConstantInt>(D);
1773 if (C1 && C2) { // (A & C1)|(B & C2)
1774 // If we have: ((V + N) & C1) | (V & C2)
1775 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1776 // replace with V+N.
1777 if (C1->getValue() == ~C2->getValue()) {
1778 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
1779 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1780 // Add commutes, try both ways.
1781 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
1782 return ReplaceInstUsesWith(I, A);
1783 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
1784 return ReplaceInstUsesWith(I, A);
1786 // Or commutes, try both ways.
1787 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
1788 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1789 // Add commutes, try both ways.
1790 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
1791 return ReplaceInstUsesWith(I, B);
1792 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
1793 return ReplaceInstUsesWith(I, B);
1797 if ((C1->getValue() & C2->getValue()) == 0) {
1798 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
1799 // iff (C1&C2) == 0 and (N&~C1) == 0
1800 if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
1801 ((V1 == B && MaskedValueIsZero(V2, ~C1->getValue())) || // (V|N)
1802 (V2 == B && MaskedValueIsZero(V1, ~C1->getValue())))) // (N|V)
1803 return BinaryOperator::CreateAnd(A,
1804 ConstantInt::get(A->getContext(),
1805 C1->getValue()|C2->getValue()));
1806 // Or commutes, try both ways.
1807 if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
1808 ((V1 == A && MaskedValueIsZero(V2, ~C2->getValue())) || // (V|N)
1809 (V2 == A && MaskedValueIsZero(V1, ~C2->getValue())))) // (N|V)
1810 return BinaryOperator::CreateAnd(B,
1811 ConstantInt::get(B->getContext(),
1812 C1->getValue()|C2->getValue()));
1814 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
1815 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
1816 ConstantInt *C3 = 0, *C4 = 0;
1817 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
1818 (C3->getValue() & ~C1->getValue()) == 0 &&
1819 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
1820 (C4->getValue() & ~C2->getValue()) == 0) {
1821 V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
1822 return BinaryOperator::CreateAnd(V2,
1823 ConstantInt::get(B->getContext(),
1824 C1->getValue()|C2->getValue()));
1829 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants.
1830 // Don't do this for vector select idioms, the code generator doesn't handle
1832 if (!I.getType()->isVectorTy()) {
1833 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
1835 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
1837 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
1839 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
1843 // ((A&~B)|(~A&B)) -> A^B
1844 if ((match(C, m_Not(m_Specific(D))) &&
1845 match(B, m_Not(m_Specific(A)))))
1846 return BinaryOperator::CreateXor(A, D);
1847 // ((~B&A)|(~A&B)) -> A^B
1848 if ((match(A, m_Not(m_Specific(D))) &&
1849 match(B, m_Not(m_Specific(C)))))
1850 return BinaryOperator::CreateXor(C, D);
1851 // ((A&~B)|(B&~A)) -> A^B
1852 if ((match(C, m_Not(m_Specific(B))) &&
1853 match(D, m_Not(m_Specific(A)))))
1854 return BinaryOperator::CreateXor(A, B);
1855 // ((~B&A)|(B&~A)) -> A^B
1856 if ((match(A, m_Not(m_Specific(B))) &&
1857 match(D, m_Not(m_Specific(C)))))
1858 return BinaryOperator::CreateXor(C, B);
1860 // ((A|B)&1)|(B&-2) -> (A&1) | B
1861 if (match(A, m_Or(m_Value(V1), m_Specific(B))) ||
1862 match(A, m_Or(m_Specific(B), m_Value(V1)))) {
1863 Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C);
1864 if (Ret) return Ret;
1866 // (B&-2)|((A|B)&1) -> (A&1) | B
1867 if (match(B, m_Or(m_Specific(A), m_Value(V1))) ||
1868 match(B, m_Or(m_Value(V1), m_Specific(A)))) {
1869 Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D);
1870 if (Ret) return Ret;
1874 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
1875 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
1876 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
1877 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
1878 SI0->getOperand(1) == SI1->getOperand(1) &&
1879 (SI0->hasOneUse() || SI1->hasOneUse())) {
1880 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
1882 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
1883 SI1->getOperand(1));
1887 // (~A | ~B) == (~(A & B)) - De Morgan's Law
1888 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1889 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1890 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1891 Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
1892 I.getName()+".demorgan");
1893 return BinaryOperator::CreateNot(And);
1896 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
1897 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
1898 if (Value *Res = FoldOrOfICmps(LHS, RHS))
1899 return ReplaceInstUsesWith(I, Res);
1901 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
1902 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1903 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1904 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
1905 return ReplaceInstUsesWith(I, Res);
1907 // fold (or (cast A), (cast B)) -> (cast (or A, B))
1908 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
1909 CastInst *Op1C = dyn_cast<CastInst>(Op1);
1910 if (Op1C && Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
1911 const Type *SrcTy = Op0C->getOperand(0)->getType();
1912 if (SrcTy == Op1C->getOperand(0)->getType() &&
1913 SrcTy->isIntOrIntVectorTy()) {
1914 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
1916 if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) &&
1917 // Only do this if the casts both really cause code to be
1919 ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
1920 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
1921 Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName());
1922 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
1925 // If this is or(cast(icmp), cast(icmp)), try to fold this even if the
1926 // cast is otherwise not optimizable. This happens for vector sexts.
1927 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
1928 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
1929 if (Value *Res = FoldOrOfICmps(LHS, RHS))
1930 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1932 // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the
1933 // cast is otherwise not optimizable. This happens for vector sexts.
1934 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
1935 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
1936 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
1937 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1942 // Note: If we've gotten to the point of visiting the outer OR, then the
1943 // inner one couldn't be simplified. If it was a constant, then it won't
1944 // be simplified by a later pass either, so we try swapping the inner/outer
1945 // ORs in the hopes that we'll be able to simplify it this way.
1946 // (X|C) | V --> (X|V) | C
1947 if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
1948 match(Op0, m_Or(m_Value(A), m_ConstantInt(C1)))) {
1949 Value *Inner = Builder->CreateOr(A, Op1);
1950 Inner->takeName(Op0);
1951 return BinaryOperator::CreateOr(Inner, C1);
1954 return Changed ? &I : 0;
1957 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
1958 bool Changed = SimplifyAssociativeOrCommutative(I);
1959 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1961 if (Value *V = SimplifyXorInst(Op0, Op1, TD))
1962 return ReplaceInstUsesWith(I, V);
1964 // (A&B)^(A&C) -> A&(B^C) etc
1965 if (Value *V = SimplifyUsingDistributiveLaws(I))
1966 return ReplaceInstUsesWith(I, V);
1968 // See if we can simplify any instructions used by the instruction whose sole
1969 // purpose is to compute bits we don't care about.
1970 if (SimplifyDemandedInstructionBits(I))
1973 // Is this a ~ operation?
1974 if (Value *NotOp = dyn_castNotVal(&I)) {
1975 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
1976 if (Op0I->getOpcode() == Instruction::And ||
1977 Op0I->getOpcode() == Instruction::Or) {
1978 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
1979 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
1980 if (dyn_castNotVal(Op0I->getOperand(1)))
1981 Op0I->swapOperands();
1982 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
1984 Builder->CreateNot(Op0I->getOperand(1),
1985 Op0I->getOperand(1)->getName()+".not");
1986 if (Op0I->getOpcode() == Instruction::And)
1987 return BinaryOperator::CreateOr(Op0NotVal, NotY);
1988 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
1991 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
1992 // ~(X | Y) === (~X & ~Y) - De Morgan's Law
1993 if (isFreeToInvert(Op0I->getOperand(0)) &&
1994 isFreeToInvert(Op0I->getOperand(1))) {
1996 Builder->CreateNot(Op0I->getOperand(0), "notlhs");
1998 Builder->CreateNot(Op0I->getOperand(1), "notrhs");
1999 if (Op0I->getOpcode() == Instruction::And)
2000 return BinaryOperator::CreateOr(NotX, NotY);
2001 return BinaryOperator::CreateAnd(NotX, NotY);
2004 } else if (Op0I->getOpcode() == Instruction::AShr) {
2005 // ~(~X >>s Y) --> (X >>s Y)
2006 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0)))
2007 return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1));
2013 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2014 if (RHS->isOne() && Op0->hasOneUse())
2015 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
2016 if (CmpInst *CI = dyn_cast<CmpInst>(Op0))
2017 return CmpInst::Create(CI->getOpcode(),
2018 CI->getInversePredicate(),
2019 CI->getOperand(0), CI->getOperand(1));
2021 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
2022 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2023 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
2024 if (CI->hasOneUse() && Op0C->hasOneUse()) {
2025 Instruction::CastOps Opcode = Op0C->getOpcode();
2026 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
2027 (RHS == ConstantExpr::getCast(Opcode,
2028 ConstantInt::getTrue(I.getContext()),
2029 Op0C->getDestTy()))) {
2030 CI->setPredicate(CI->getInversePredicate());
2031 return CastInst::Create(Opcode, CI, Op0C->getType());
2037 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2038 // ~(c-X) == X-c-1 == X+(-c-1)
2039 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
2040 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
2041 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
2042 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
2043 ConstantInt::get(I.getType(), 1));
2044 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
2047 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2048 if (Op0I->getOpcode() == Instruction::Add) {
2049 // ~(X-c) --> (-c-1)-X
2050 if (RHS->isAllOnesValue()) {
2051 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
2052 return BinaryOperator::CreateSub(
2053 ConstantExpr::getSub(NegOp0CI,
2054 ConstantInt::get(I.getType(), 1)),
2055 Op0I->getOperand(0));
2056 } else if (RHS->getValue().isSignBit()) {
2057 // (X + C) ^ signbit -> (X + C + signbit)
2058 Constant *C = ConstantInt::get(I.getContext(),
2059 RHS->getValue() + Op0CI->getValue());
2060 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
2063 } else if (Op0I->getOpcode() == Instruction::Or) {
2064 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
2065 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
2066 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
2067 // Anything in both C1 and C2 is known to be zero, remove it from
2069 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
2070 NewRHS = ConstantExpr::getAnd(NewRHS,
2071 ConstantExpr::getNot(CommonBits));
2073 I.setOperand(0, Op0I->getOperand(0));
2074 I.setOperand(1, NewRHS);
2081 // Try to fold constant and into select arguments.
2082 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2083 if (Instruction *R = FoldOpIntoSelect(I, SI))
2085 if (isa<PHINode>(Op0))
2086 if (Instruction *NV = FoldOpIntoPhi(I))
2090 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
2093 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2094 if (A == Op0) { // B^(B|A) == (A|B)^B
2095 Op1I->swapOperands();
2097 std::swap(Op0, Op1);
2098 } else if (B == Op0) { // B^(A|B) == (A|B)^B
2099 I.swapOperands(); // Simplified below.
2100 std::swap(Op0, Op1);
2102 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
2104 if (A == Op0) { // A^(A&B) -> A^(B&A)
2105 Op1I->swapOperands();
2108 if (B == Op0) { // A^(B&A) -> (B&A)^A
2109 I.swapOperands(); // Simplified below.
2110 std::swap(Op0, Op1);
2115 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
2118 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2119 Op0I->hasOneUse()) {
2120 if (A == Op1) // (B|A)^B == (A|B)^B
2122 if (B == Op1) // (A|B)^B == A & ~B
2123 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1, "tmp"));
2124 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2126 if (A == Op1) // (A&B)^A -> (B&A)^A
2128 if (B == Op1 && // (B&A)^A == ~B & A
2129 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
2130 return BinaryOperator::CreateAnd(Builder->CreateNot(A, "tmp"), Op1);
2135 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
2136 if (Op0I && Op1I && Op0I->isShift() &&
2137 Op0I->getOpcode() == Op1I->getOpcode() &&
2138 Op0I->getOperand(1) == Op1I->getOperand(1) &&
2139 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
2141 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
2143 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
2144 Op1I->getOperand(1));
2148 Value *A, *B, *C, *D;
2149 // (A & B)^(A | B) -> A ^ B
2150 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2151 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
2152 if ((A == C && B == D) || (A == D && B == C))
2153 return BinaryOperator::CreateXor(A, B);
2155 // (A | B)^(A & B) -> A ^ B
2156 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2157 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
2158 if ((A == C && B == D) || (A == D && B == C))
2159 return BinaryOperator::CreateXor(A, B);
2163 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
2164 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
2165 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
2166 if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
2167 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2168 LHS->getOperand(1) == RHS->getOperand(0))
2169 LHS->swapOperands();
2170 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2171 LHS->getOperand(1) == RHS->getOperand(1)) {
2172 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2173 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
2174 bool isSigned = LHS->isSigned() || RHS->isSigned();
2175 return ReplaceInstUsesWith(I,
2176 getICmpValue(isSigned, Code, Op0, Op1, Builder));
2180 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
2181 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2182 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
2183 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
2184 const Type *SrcTy = Op0C->getOperand(0)->getType();
2185 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() &&
2186 // Only do this if the casts both really cause code to be generated.
2187 ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0),
2189 ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0),
2191 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
2192 Op1C->getOperand(0), I.getName());
2193 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2198 return Changed ? &I : 0;