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/ConstantRange.h"
18 #include "llvm/Support/PatternMatch.h"
20 using namespace PatternMatch;
23 /// AddOne - Add one to a ConstantInt.
24 static Constant *AddOne(Constant *C) {
25 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
27 /// SubOne - Subtract one from a ConstantInt.
28 static Constant *SubOne(ConstantInt *C) {
29 return ConstantInt::get(C->getContext(), C->getValue()-1);
32 /// isFreeToInvert - Return true if the specified value is free to invert (apply
33 /// ~ to). This happens in cases where the ~ can be eliminated.
34 static inline bool isFreeToInvert(Value *V) {
36 if (BinaryOperator::isNot(V))
39 // Constants can be considered to be not'ed values.
40 if (isa<ConstantInt>(V))
43 // Compares can be inverted if they have a single use.
44 if (CmpInst *CI = dyn_cast<CmpInst>(V))
45 return CI->hasOneUse();
50 static inline Value *dyn_castNotVal(Value *V) {
51 // If this is not(not(x)) don't return that this is a not: we want the two
52 // not's to be folded first.
53 if (BinaryOperator::isNot(V)) {
54 Value *Operand = BinaryOperator::getNotArgument(V);
55 if (!isFreeToInvert(Operand))
59 // Constants can be considered to be not'ed values...
60 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
61 return ConstantInt::get(C->getType(), ~C->getValue());
66 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
67 /// are carefully arranged to allow folding of expressions such as:
69 /// (A < B) | (A > B) --> (A != B)
71 /// Note that this is only valid if the first and second predicates have the
72 /// same sign. Is illegal to do: (A u< B) | (A s> B)
74 /// Three bits are used to represent the condition, as follows:
79 /// <=> Value Definition
80 /// 000 0 Always false
89 static unsigned getICmpCode(const ICmpInst *ICI) {
90 switch (ICI->getPredicate()) {
92 case ICmpInst::ICMP_UGT: return 1; // 001
93 case ICmpInst::ICMP_SGT: return 1; // 001
94 case ICmpInst::ICMP_EQ: return 2; // 010
95 case ICmpInst::ICMP_UGE: return 3; // 011
96 case ICmpInst::ICMP_SGE: return 3; // 011
97 case ICmpInst::ICMP_ULT: return 4; // 100
98 case ICmpInst::ICMP_SLT: return 4; // 100
99 case ICmpInst::ICMP_NE: return 5; // 101
100 case ICmpInst::ICMP_ULE: return 6; // 110
101 case ICmpInst::ICMP_SLE: return 6; // 110
104 llvm_unreachable("Invalid ICmp predicate!");
109 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
110 /// predicate into a three bit mask. It also returns whether it is an ordered
111 /// predicate by reference.
112 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
115 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
116 case FCmpInst::FCMP_UNO: return 0; // 000
117 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
118 case FCmpInst::FCMP_UGT: return 1; // 001
119 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
120 case FCmpInst::FCMP_UEQ: return 2; // 010
121 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
122 case FCmpInst::FCMP_UGE: return 3; // 011
123 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
124 case FCmpInst::FCMP_ULT: return 4; // 100
125 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
126 case FCmpInst::FCMP_UNE: return 5; // 101
127 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
128 case FCmpInst::FCMP_ULE: return 6; // 110
131 // Not expecting FCMP_FALSE and FCMP_TRUE;
132 llvm_unreachable("Unexpected FCmp predicate!");
137 /// getICmpValue - This is the complement of getICmpCode, which turns an
138 /// opcode and two operands into either a constant true or false, or a brand
139 /// new ICmp instruction. The sign is passed in to determine which kind
140 /// of predicate to use in the new icmp instruction.
141 static Value *getICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS,
142 InstCombiner::BuilderTy *Builder) {
143 CmpInst::Predicate Pred;
145 default: assert(0 && "Illegal ICmp code!");
147 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
148 case 1: Pred = Sign ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; break;
149 case 2: Pred = ICmpInst::ICMP_EQ; break;
150 case 3: Pred = Sign ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; break;
151 case 4: Pred = Sign ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; break;
152 case 5: Pred = ICmpInst::ICMP_NE; break;
153 case 6: Pred = Sign ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; break;
155 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
157 return Builder->CreateICmp(Pred, LHS, RHS);
160 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
161 /// opcode and two operands into either a FCmp instruction. isordered is passed
162 /// in to determine which kind of predicate to use in the new fcmp instruction.
163 static Value *getFCmpValue(bool isordered, unsigned code,
164 Value *LHS, Value *RHS,
165 InstCombiner::BuilderTy *Builder) {
166 CmpInst::Predicate Pred;
168 default: assert(0 && "Illegal FCmp code!");
169 case 0: Pred = isordered ? FCmpInst::FCMP_ORD : FCmpInst::FCMP_UNO; break;
170 case 1: Pred = isordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; break;
171 case 2: Pred = isordered ? FCmpInst::FCMP_OEQ : FCmpInst::FCMP_UEQ; break;
172 case 3: Pred = isordered ? FCmpInst::FCMP_OGE : FCmpInst::FCMP_UGE; break;
173 case 4: Pred = isordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; break;
174 case 5: Pred = isordered ? FCmpInst::FCMP_ONE : FCmpInst::FCMP_UNE; break;
175 case 6: Pred = isordered ? FCmpInst::FCMP_OLE : FCmpInst::FCMP_ULE; break;
177 if (!isordered) return ConstantInt::getTrue(LHS->getContext());
178 Pred = FCmpInst::FCMP_ORD; break;
180 return Builder->CreateFCmp(Pred, LHS, RHS);
183 /// PredicatesFoldable - Return true if both predicates match sign or if at
184 /// least one of them is an equality comparison (which is signless).
185 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
186 return (CmpInst::isSigned(p1) == CmpInst::isSigned(p2)) ||
187 (CmpInst::isSigned(p1) && ICmpInst::isEquality(p2)) ||
188 (CmpInst::isSigned(p2) && ICmpInst::isEquality(p1));
191 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
192 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
193 // guaranteed to be a binary operator.
194 Instruction *InstCombiner::OptAndOp(Instruction *Op,
197 BinaryOperator &TheAnd) {
198 Value *X = Op->getOperand(0);
199 Constant *Together = 0;
201 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
203 switch (Op->getOpcode()) {
204 case Instruction::Xor:
205 if (Op->hasOneUse()) {
206 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
207 Value *And = Builder->CreateAnd(X, AndRHS);
209 return BinaryOperator::CreateXor(And, Together);
212 case Instruction::Or:
213 if (Op->hasOneUse()){
214 if (Together != OpRHS) {
215 // (X | C1) & C2 --> (X | (C1&C2)) & C2
216 Value *Or = Builder->CreateOr(X, Together);
218 return BinaryOperator::CreateAnd(Or, AndRHS);
221 ConstantInt *TogetherCI = dyn_cast<ConstantInt>(Together);
222 if (TogetherCI && !TogetherCI->isZero()){
223 // (X | C1) & C2 --> (X & (C2^(C1&C2))) | C1
224 // NOTE: This reduces the number of bits set in the & mask, which
225 // can expose opportunities for store narrowing.
226 Together = ConstantExpr::getXor(AndRHS, Together);
227 Value *And = Builder->CreateAnd(X, Together);
229 return BinaryOperator::CreateOr(And, OpRHS);
234 case Instruction::Add:
235 if (Op->hasOneUse()) {
236 // Adding a one to a single bit bit-field should be turned into an XOR
237 // of the bit. First thing to check is to see if this AND is with a
238 // single bit constant.
239 const APInt &AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
241 // If there is only one bit set.
242 if (AndRHSV.isPowerOf2()) {
243 // Ok, at this point, we know that we are masking the result of the
244 // ADD down to exactly one bit. If the constant we are adding has
245 // no bits set below this bit, then we can eliminate the ADD.
246 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
248 // Check to see if any bits below the one bit set in AndRHSV are set.
249 if ((AddRHS & (AndRHSV-1)) == 0) {
250 // If not, the only thing that can effect the output of the AND is
251 // the bit specified by AndRHSV. If that bit is set, the effect of
252 // the XOR is to toggle the bit. If it is clear, then the ADD has
254 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
255 TheAnd.setOperand(0, X);
258 // Pull the XOR out of the AND.
259 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
260 NewAnd->takeName(Op);
261 return BinaryOperator::CreateXor(NewAnd, AndRHS);
268 case Instruction::Shl: {
269 // We know that the AND will not produce any of the bits shifted in, so if
270 // the anded constant includes them, clear them now!
272 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
273 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
274 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
275 ConstantInt *CI = ConstantInt::get(AndRHS->getContext(),
276 AndRHS->getValue() & ShlMask);
278 if (CI->getValue() == ShlMask)
279 // Masking out bits that the shift already masks.
280 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
282 if (CI != AndRHS) { // Reducing bits set in and.
283 TheAnd.setOperand(1, CI);
288 case Instruction::LShr: {
289 // We know that the AND will not produce any of the bits shifted in, so if
290 // the anded constant includes them, clear them now! This only applies to
291 // unsigned shifts, because a signed shr may bring in set bits!
293 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
294 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
295 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
296 ConstantInt *CI = ConstantInt::get(Op->getContext(),
297 AndRHS->getValue() & ShrMask);
299 if (CI->getValue() == ShrMask)
300 // Masking out bits that the shift already masks.
301 return ReplaceInstUsesWith(TheAnd, Op);
304 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
309 case Instruction::AShr:
311 // See if this is shifting in some sign extension, then masking it out
313 if (Op->hasOneUse()) {
314 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
315 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
316 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
317 Constant *C = ConstantInt::get(Op->getContext(),
318 AndRHS->getValue() & ShrMask);
319 if (C == AndRHS) { // Masking out bits shifted in.
320 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
321 // Make the argument unsigned.
322 Value *ShVal = Op->getOperand(0);
323 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
324 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
333 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
334 /// true, otherwise (V < Lo || V >= Hi). In practice, we emit the more efficient
335 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
336 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
337 /// insert new instructions.
338 Value *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
339 bool isSigned, bool Inside) {
340 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
341 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
342 "Lo is not <= Hi in range emission code!");
345 if (Lo == Hi) // Trivially false.
346 return ConstantInt::getFalse(V->getContext());
348 // V >= Min && V < Hi --> V < Hi
349 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
350 ICmpInst::Predicate pred = (isSigned ?
351 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
352 return Builder->CreateICmp(pred, V, Hi);
355 // Emit V-Lo <u Hi-Lo
356 Constant *NegLo = ConstantExpr::getNeg(Lo);
357 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
358 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
359 return Builder->CreateICmpULT(Add, UpperBound);
362 if (Lo == Hi) // Trivially true.
363 return ConstantInt::getTrue(V->getContext());
365 // V < Min || V >= Hi -> V > Hi-1
366 Hi = SubOne(cast<ConstantInt>(Hi));
367 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
368 ICmpInst::Predicate pred = (isSigned ?
369 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
370 return Builder->CreateICmp(pred, V, Hi);
373 // Emit V-Lo >u Hi-1-Lo
374 // Note that Hi has already had one subtracted from it, above.
375 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
376 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
377 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
378 return Builder->CreateICmpUGT(Add, LowerBound);
381 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
382 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
383 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
384 // not, since all 1s are not contiguous.
385 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
386 const APInt& V = Val->getValue();
387 uint32_t BitWidth = Val->getType()->getBitWidth();
388 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
390 // look for the first zero bit after the run of ones
391 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
392 // look for the first non-zero bit
393 ME = V.getActiveBits();
397 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
398 /// where isSub determines whether the operator is a sub. If we can fold one of
399 /// the following xforms:
401 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
402 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
403 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
405 /// return (A +/- B).
407 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
408 ConstantInt *Mask, bool isSub,
410 Instruction *LHSI = dyn_cast<Instruction>(LHS);
411 if (!LHSI || LHSI->getNumOperands() != 2 ||
412 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
414 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
416 switch (LHSI->getOpcode()) {
418 case Instruction::And:
419 if (ConstantExpr::getAnd(N, Mask) == Mask) {
420 // If the AndRHS is a power of two minus one (0+1+), this is simple.
421 if ((Mask->getValue().countLeadingZeros() +
422 Mask->getValue().countPopulation()) ==
423 Mask->getValue().getBitWidth())
426 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
427 // part, we don't need any explicit masks to take them out of A. If that
428 // is all N is, ignore it.
429 uint32_t MB = 0, ME = 0;
430 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
431 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
432 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
433 if (MaskedValueIsZero(RHS, Mask))
438 case Instruction::Or:
439 case Instruction::Xor:
440 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
441 if ((Mask->getValue().countLeadingZeros() +
442 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
443 && ConstantExpr::getAnd(N, Mask)->isNullValue())
449 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
450 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
453 /// enum for classifying (icmp eq (A & B), C) and (icmp ne (A & B), C)
454 /// One of A and B is considered the mask, the other the value. This is
455 /// described as the "AMask" or "BMask" part of the enum. If the enum
456 /// contains only "Mask", then both A and B can be considered masks.
457 /// If A is the mask, then it was proven, that (A & C) == C. This
458 /// is trivial if C == A, or C == 0. If both A and C are constants, this
459 /// proof is also easy.
460 /// For the following explanations we assume that A is the mask.
461 /// The part "AllOnes" declares, that the comparison is true only
462 /// if (A & B) == A, or all bits of A are set in B.
463 /// Example: (icmp eq (A & 3), 3) -> FoldMskICmp_AMask_AllOnes
464 /// The part "AllZeroes" declares, that the comparison is true only
465 /// if (A & B) == 0, or all bits of A are cleared in B.
466 /// Example: (icmp eq (A & 3), 0) -> FoldMskICmp_Mask_AllZeroes
467 /// The part "Mixed" declares, that (A & B) == C and C might or might not
468 /// contain any number of one bits and zero bits.
469 /// Example: (icmp eq (A & 3), 1) -> FoldMskICmp_AMask_Mixed
470 /// The Part "Not" means, that in above descriptions "==" should be replaced
472 /// Example: (icmp ne (A & 3), 3) -> FoldMskICmp_AMask_NotAllOnes
473 /// If the mask A contains a single bit, then the following is equivalent:
474 /// (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
475 /// (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
476 enum MaskedICmpType {
477 FoldMskICmp_AMask_AllOnes = 1,
478 FoldMskICmp_AMask_NotAllOnes = 2,
479 FoldMskICmp_BMask_AllOnes = 4,
480 FoldMskICmp_BMask_NotAllOnes = 8,
481 FoldMskICmp_Mask_AllZeroes = 16,
482 FoldMskICmp_Mask_NotAllZeroes = 32,
483 FoldMskICmp_AMask_Mixed = 64,
484 FoldMskICmp_AMask_NotMixed = 128,
485 FoldMskICmp_BMask_Mixed = 256,
486 FoldMskICmp_BMask_NotMixed = 512
489 /// return the set of pattern classes (from MaskedICmpType)
490 /// that (icmp SCC (A & B), C) satisfies
491 static unsigned getTypeOfMaskedICmp(Value* A, Value* B, Value* C,
492 ICmpInst::Predicate SCC)
494 ConstantInt *ACst = dyn_cast<ConstantInt>(A);
495 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
496 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
497 bool icmp_eq = (SCC == ICmpInst::ICMP_EQ);
498 bool icmp_abit = (ACst != 0 && !ACst->isZero() &&
499 ACst->getValue().isPowerOf2());
500 bool icmp_bbit = (BCst != 0 && !BCst->isZero() &&
501 BCst->getValue().isPowerOf2());
503 if (CCst != 0 && CCst->isZero()) {
504 // if C is zero, then both A and B qualify as mask
505 result |= (icmp_eq ? (FoldMskICmp_Mask_AllZeroes |
506 FoldMskICmp_Mask_AllZeroes |
507 FoldMskICmp_AMask_Mixed |
508 FoldMskICmp_BMask_Mixed)
509 : (FoldMskICmp_Mask_NotAllZeroes |
510 FoldMskICmp_Mask_NotAllZeroes |
511 FoldMskICmp_AMask_NotMixed |
512 FoldMskICmp_BMask_NotMixed));
514 result |= (icmp_eq ? (FoldMskICmp_AMask_NotAllOnes |
515 FoldMskICmp_AMask_NotMixed)
516 : (FoldMskICmp_AMask_AllOnes |
517 FoldMskICmp_AMask_Mixed));
519 result |= (icmp_eq ? (FoldMskICmp_BMask_NotAllOnes |
520 FoldMskICmp_BMask_NotMixed)
521 : (FoldMskICmp_BMask_AllOnes |
522 FoldMskICmp_BMask_Mixed));
526 result |= (icmp_eq ? (FoldMskICmp_AMask_AllOnes |
527 FoldMskICmp_AMask_Mixed)
528 : (FoldMskICmp_AMask_NotAllOnes |
529 FoldMskICmp_AMask_NotMixed));
531 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
532 FoldMskICmp_AMask_NotMixed)
533 : (FoldMskICmp_Mask_AllZeroes |
534 FoldMskICmp_AMask_Mixed));
536 else if (ACst != 0 && CCst != 0 &&
537 ConstantExpr::getAnd(ACst, CCst) == CCst) {
538 result |= (icmp_eq ? FoldMskICmp_AMask_Mixed
539 : FoldMskICmp_AMask_NotMixed);
543 result |= (icmp_eq ? (FoldMskICmp_BMask_AllOnes |
544 FoldMskICmp_BMask_Mixed)
545 : (FoldMskICmp_BMask_NotAllOnes |
546 FoldMskICmp_BMask_NotMixed));
548 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
549 FoldMskICmp_BMask_NotMixed)
550 : (FoldMskICmp_Mask_AllZeroes |
551 FoldMskICmp_BMask_Mixed));
553 else if (BCst != 0 && CCst != 0 &&
554 ConstantExpr::getAnd(BCst, CCst) == CCst) {
555 result |= (icmp_eq ? FoldMskICmp_BMask_Mixed
556 : FoldMskICmp_BMask_NotMixed);
561 /// foldLogOpOfMaskedICmpsHelper:
562 /// handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
563 /// return the set of pattern classes (from MaskedICmpType)
564 /// that both LHS and RHS satisfy
565 static unsigned foldLogOpOfMaskedICmpsHelper(Value*& A,
566 Value*& B, Value*& C,
567 Value*& D, Value*& E,
568 ICmpInst *LHS, ICmpInst *RHS) {
569 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
570 if (LHSCC != ICmpInst::ICMP_EQ && LHSCC != ICmpInst::ICMP_NE) return 0;
571 if (RHSCC != ICmpInst::ICMP_EQ && RHSCC != ICmpInst::ICMP_NE) return 0;
572 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) return 0;
573 // vectors are not (yet?) supported
574 if (LHS->getOperand(0)->getType()->isVectorTy()) return 0;
576 // Here comes the tricky part:
577 // LHS might be of the form L11 & L12 == X, X == L21 & L22,
578 // and L11 & L12 == L21 & L22. The same goes for RHS.
579 // Now we must find those components L** and R**, that are equal, so
580 // that we can extract the parameters A, B, C, D, and E for the canonical
582 Value *L1 = LHS->getOperand(0);
583 Value *L2 = LHS->getOperand(1);
584 Value *L11,*L12,*L21,*L22;
585 if (match(L1, m_And(m_Value(L11), m_Value(L12)))) {
586 if (!match(L2, m_And(m_Value(L21), m_Value(L22))))
590 if (!match(L2, m_And(m_Value(L11), m_Value(L12))))
596 Value *R1 = RHS->getOperand(0);
597 Value *R2 = RHS->getOperand(1);
600 if (match(R1, m_And(m_Value(R11), m_Value(R12)))) {
601 if (R11 != 0 && (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22)) {
602 A = R11; D = R12; E = R2; ok = true;
605 if (R12 != 0 && (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22)) {
606 A = R12; D = R11; E = R2; ok = true;
609 if (!ok && match(R2, m_And(m_Value(R11), m_Value(R12)))) {
610 if (R11 != 0 && (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22)) {
611 A = R11; D = R12; E = R1; ok = true;
614 if (R12 != 0 && (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22)) {
615 A = R12; D = R11; E = R1; ok = true;
636 unsigned left_type = getTypeOfMaskedICmp(A, B, C, LHSCC);
637 unsigned right_type = getTypeOfMaskedICmp(A, D, E, RHSCC);
638 return left_type & right_type;
640 /// foldLogOpOfMaskedICmps:
641 /// try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
642 /// into a single (icmp(A & X) ==/!= Y)
643 static Value* foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS,
644 ICmpInst::Predicate NEWCC,
645 llvm::InstCombiner::BuilderTy* Builder) {
646 Value *A = 0, *B = 0, *C = 0, *D = 0, *E = 0;
647 unsigned mask = foldLogOpOfMaskedICmpsHelper(A, B, C, D, E, LHS, RHS);
648 if (mask == 0) return 0;
650 if (NEWCC == ICmpInst::ICMP_NE)
651 mask >>= 1; // treat "Not"-states as normal states
653 if (mask & FoldMskICmp_Mask_AllZeroes) {
654 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
655 // -> (icmp eq (A & (B|D)), 0)
656 Value* newOr = Builder->CreateOr(B, D);
657 Value* newAnd = Builder->CreateAnd(A, newOr);
658 // we can't use C as zero, because we might actually handle
659 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
660 // with B and D, having a single bit set
661 Value* zero = Constant::getNullValue(A->getType());
662 return Builder->CreateICmp(NEWCC, newAnd, zero);
664 else if (mask & FoldMskICmp_BMask_AllOnes) {
665 // (icmp eq (A & B), B) & (icmp eq (A & D), D)
666 // -> (icmp eq (A & (B|D)), (B|D))
667 Value* newOr = Builder->CreateOr(B, D);
668 Value* newAnd = Builder->CreateAnd(A, newOr);
669 return Builder->CreateICmp(NEWCC, newAnd, newOr);
671 else if (mask & FoldMskICmp_AMask_AllOnes) {
672 // (icmp eq (A & B), A) & (icmp eq (A & D), A)
673 // -> (icmp eq (A & (B&D)), A)
674 Value* newAnd1 = Builder->CreateAnd(B, D);
675 Value* newAnd = Builder->CreateAnd(A, newAnd1);
676 return Builder->CreateICmp(NEWCC, newAnd, A);
678 else if (mask & FoldMskICmp_BMask_Mixed) {
679 // (icmp eq (A & B), C) & (icmp eq (A & D), E)
680 // We already know that B & C == C && D & E == E.
681 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
682 // C and E, which are shared by both the mask B and the mask D, don't
683 // contradict, then we can transform to
684 // -> (icmp eq (A & (B|D)), (C|E))
685 // Currently, we only handle the case of B, C, D, and E being constant.
686 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
687 if (BCst == 0) return 0;
688 ConstantInt *DCst = dyn_cast<ConstantInt>(D);
689 if (DCst == 0) return 0;
690 // we can't simply use C and E, because we might actually handle
691 // (icmp ne (A & B), B) & (icmp eq (A & D), D)
692 // with B and D, having a single bit set
694 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
695 if (CCst == 0) return 0;
696 if (LHS->getPredicate() != NEWCC)
697 CCst = dyn_cast<ConstantInt>( ConstantExpr::getXor(BCst, CCst) );
698 ConstantInt *ECst = dyn_cast<ConstantInt>(E);
699 if (ECst == 0) return 0;
700 if (RHS->getPredicate() != NEWCC)
701 ECst = dyn_cast<ConstantInt>( ConstantExpr::getXor(DCst, ECst) );
702 ConstantInt* MCst = dyn_cast<ConstantInt>(
703 ConstantExpr::getAnd(ConstantExpr::getAnd(BCst, DCst),
704 ConstantExpr::getXor(CCst, ECst)) );
705 // if there is a conflict we should actually return a false for the
709 Value *newOr1 = Builder->CreateOr(B, D);
710 Value *newOr2 = ConstantExpr::getOr(CCst, ECst);
711 Value *newAnd = Builder->CreateAnd(A, newOr1);
712 return Builder->CreateICmp(NEWCC, newAnd, newOr2);
717 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
718 Value *InstCombiner::FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
719 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
721 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
722 if (PredicatesFoldable(LHSCC, RHSCC)) {
723 if (LHS->getOperand(0) == RHS->getOperand(1) &&
724 LHS->getOperand(1) == RHS->getOperand(0))
726 if (LHS->getOperand(0) == RHS->getOperand(0) &&
727 LHS->getOperand(1) == RHS->getOperand(1)) {
728 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
729 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
730 bool isSigned = LHS->isSigned() || RHS->isSigned();
731 return getICmpValue(isSigned, Code, Op0, Op1, Builder);
735 // handle (roughly): (icmp eq (A & B), C) & (icmp eq (A & D), E)
736 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, ICmpInst::ICMP_EQ, Builder))
739 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
740 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
741 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
742 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
743 if (LHSCst == 0 || RHSCst == 0) return 0;
745 if (LHSCst == RHSCst && LHSCC == RHSCC) {
746 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
747 // where C is a power of 2
748 if (LHSCC == ICmpInst::ICMP_ULT &&
749 LHSCst->getValue().isPowerOf2()) {
750 Value *NewOr = Builder->CreateOr(Val, Val2);
751 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
754 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
755 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
756 Value *NewOr = Builder->CreateOr(Val, Val2);
757 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
760 // (icmp slt A, 0) & (icmp slt B, 0) --> (icmp slt (A&B), 0)
761 if (LHSCC == ICmpInst::ICMP_SLT && LHSCst->isZero()) {
762 Value *NewAnd = Builder->CreateAnd(Val, Val2);
763 return Builder->CreateICmp(LHSCC, NewAnd, LHSCst);
766 // (icmp sgt A, -1) & (icmp sgt B, -1) --> (icmp sgt (A|B), -1)
767 if (LHSCC == ICmpInst::ICMP_SGT && LHSCst->isAllOnesValue()) {
768 Value *NewOr = Builder->CreateOr(Val, Val2);
769 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
773 // From here on, we only handle:
774 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
775 if (Val != Val2) return 0;
777 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
778 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
779 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
780 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
781 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
784 // Make a constant range that's the intersection of the two icmp ranges.
785 // If the intersection is empty, we know that the result is false.
786 ConstantRange LHSRange =
787 ConstantRange::makeICmpRegion(LHSCC, LHSCst->getValue());
788 ConstantRange RHSRange =
789 ConstantRange::makeICmpRegion(RHSCC, RHSCst->getValue());
791 if (LHSRange.intersectWith(RHSRange).isEmptySet())
792 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
794 // We can't fold (ugt x, C) & (sgt x, C2).
795 if (!PredicatesFoldable(LHSCC, RHSCC))
798 // Ensure that the larger constant is on the RHS.
800 if (CmpInst::isSigned(LHSCC) ||
801 (ICmpInst::isEquality(LHSCC) &&
802 CmpInst::isSigned(RHSCC)))
803 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
805 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
809 std::swap(LHSCst, RHSCst);
810 std::swap(LHSCC, RHSCC);
813 // At this point, we know we have two icmp instructions
814 // comparing a value against two constants and and'ing the result
815 // together. Because of the above check, we know that we only have
816 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
817 // (from the icmp folding check above), that the two constants
818 // are not equal and that the larger constant is on the RHS
819 assert(LHSCst != RHSCst && "Compares not folded above?");
822 default: llvm_unreachable("Unknown integer condition code!");
823 case ICmpInst::ICMP_EQ:
825 default: llvm_unreachable("Unknown integer condition code!");
826 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
827 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
828 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
831 case ICmpInst::ICMP_NE:
833 default: llvm_unreachable("Unknown integer condition code!");
834 case ICmpInst::ICMP_ULT:
835 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
836 return Builder->CreateICmpULT(Val, LHSCst);
837 break; // (X != 13 & X u< 15) -> no change
838 case ICmpInst::ICMP_SLT:
839 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
840 return Builder->CreateICmpSLT(Val, LHSCst);
841 break; // (X != 13 & X s< 15) -> no change
842 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
843 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
844 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
846 case ICmpInst::ICMP_NE:
847 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
848 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
849 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
850 return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1));
852 break; // (X != 13 & X != 15) -> no change
855 case ICmpInst::ICMP_ULT:
857 default: llvm_unreachable("Unknown integer condition code!");
858 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
859 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
860 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
861 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
863 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
864 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
866 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
870 case ICmpInst::ICMP_SLT:
872 default: llvm_unreachable("Unknown integer condition code!");
873 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
875 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
876 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
878 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
882 case ICmpInst::ICMP_UGT:
884 default: llvm_unreachable("Unknown integer condition code!");
885 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
886 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
888 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
890 case ICmpInst::ICMP_NE:
891 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
892 return Builder->CreateICmp(LHSCC, Val, RHSCst);
893 break; // (X u> 13 & X != 15) -> no change
894 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
895 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
896 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
900 case ICmpInst::ICMP_SGT:
902 default: llvm_unreachable("Unknown integer condition code!");
903 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
904 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
906 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
908 case ICmpInst::ICMP_NE:
909 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
910 return Builder->CreateICmp(LHSCC, Val, RHSCst);
911 break; // (X s> 13 & X != 15) -> no change
912 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
913 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true);
914 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
923 /// FoldAndOfFCmps - Optimize (fcmp)&(fcmp). NOTE: Unlike the rest of
924 /// instcombine, this returns a Value which should already be inserted into the
926 Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
927 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
928 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
929 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
930 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
931 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
932 // If either of the constants are nans, then the whole thing returns
934 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
935 return ConstantInt::getFalse(LHS->getContext());
936 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
939 // Handle vector zeros. This occurs because the canonical form of
940 // "fcmp ord x,x" is "fcmp ord x, 0".
941 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
942 isa<ConstantAggregateZero>(RHS->getOperand(1)))
943 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
947 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
948 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
949 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
952 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
953 // Swap RHS operands to match LHS.
954 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
955 std::swap(Op1LHS, Op1RHS);
958 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
959 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
961 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
962 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
963 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
964 if (Op0CC == FCmpInst::FCMP_TRUE)
966 if (Op1CC == FCmpInst::FCMP_TRUE)
971 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
972 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
975 std::swap(Op0Pred, Op1Pred);
976 std::swap(Op0Ordered, Op1Ordered);
979 // uno && ueq -> uno && (uno || eq) -> ueq
980 // ord && olt -> ord && (ord && lt) -> olt
981 if (Op0Ordered == Op1Ordered)
984 // uno && oeq -> uno && (ord && eq) -> false
985 // uno && ord -> false
987 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
988 // ord && ueq -> ord && (uno || eq) -> oeq
989 return getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS, Builder);
997 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
998 bool Changed = SimplifyAssociativeOrCommutative(I);
999 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1001 if (Value *V = SimplifyAndInst(Op0, Op1, TD))
1002 return ReplaceInstUsesWith(I, V);
1004 // (A|B)&(A|C) -> A|(B&C) etc
1005 if (Value *V = SimplifyUsingDistributiveLaws(I))
1006 return ReplaceInstUsesWith(I, V);
1008 // See if we can simplify any instructions used by the instruction whose sole
1009 // purpose is to compute bits we don't care about.
1010 if (SimplifyDemandedInstructionBits(I))
1013 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
1014 const APInt &AndRHSMask = AndRHS->getValue();
1016 // Optimize a variety of ((val OP C1) & C2) combinations...
1017 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1018 Value *Op0LHS = Op0I->getOperand(0);
1019 Value *Op0RHS = Op0I->getOperand(1);
1020 switch (Op0I->getOpcode()) {
1022 case Instruction::Xor:
1023 case Instruction::Or: {
1024 // If the mask is only needed on one incoming arm, push it up.
1025 if (!Op0I->hasOneUse()) break;
1027 APInt NotAndRHS(~AndRHSMask);
1028 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
1029 // Not masking anything out for the LHS, move to RHS.
1030 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
1031 Op0RHS->getName()+".masked");
1032 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
1034 if (!isa<Constant>(Op0RHS) &&
1035 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
1036 // Not masking anything out for the RHS, move to LHS.
1037 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
1038 Op0LHS->getName()+".masked");
1039 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
1044 case Instruction::Add:
1045 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
1046 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1047 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1048 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
1049 return BinaryOperator::CreateAnd(V, AndRHS);
1050 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
1051 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
1054 case Instruction::Sub:
1055 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
1056 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1057 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1058 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
1059 return BinaryOperator::CreateAnd(V, AndRHS);
1061 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
1062 // has 1's for all bits that the subtraction with A might affect.
1063 if (Op0I->hasOneUse() && !match(Op0LHS, m_Zero())) {
1064 uint32_t BitWidth = AndRHSMask.getBitWidth();
1065 uint32_t Zeros = AndRHSMask.countLeadingZeros();
1066 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
1068 if (MaskedValueIsZero(Op0LHS, Mask)) {
1069 Value *NewNeg = Builder->CreateNeg(Op0RHS);
1070 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
1075 case Instruction::Shl:
1076 case Instruction::LShr:
1077 // (1 << x) & 1 --> zext(x == 0)
1078 // (1 >> x) & 1 --> zext(x == 0)
1079 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
1081 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
1082 return new ZExtInst(NewICmp, I.getType());
1087 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
1088 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
1092 // If this is an integer truncation, and if the source is an 'and' with
1093 // immediate, transform it. This frequently occurs for bitfield accesses.
1095 Value *X = 0; ConstantInt *YC = 0;
1096 if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) {
1097 // Change: and (trunc (and X, YC) to T), C2
1098 // into : and (trunc X to T), trunc(YC) & C2
1099 // This will fold the two constants together, which may allow
1100 // other simplifications.
1101 Value *NewCast = Builder->CreateTrunc(X, I.getType(), "and.shrunk");
1102 Constant *C3 = ConstantExpr::getTrunc(YC, I.getType());
1103 C3 = ConstantExpr::getAnd(C3, AndRHS);
1104 return BinaryOperator::CreateAnd(NewCast, C3);
1108 // Try to fold constant and into select arguments.
1109 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1110 if (Instruction *R = FoldOpIntoSelect(I, SI))
1112 if (isa<PHINode>(Op0))
1113 if (Instruction *NV = FoldOpIntoPhi(I))
1118 // (~A & ~B) == (~(A | B)) - De Morgan's Law
1119 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1120 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1121 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1122 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
1123 I.getName()+".demorgan");
1124 return BinaryOperator::CreateNot(Or);
1128 Value *A = 0, *B = 0, *C = 0, *D = 0;
1129 // (A|B) & ~(A&B) -> A^B
1130 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1131 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1132 ((A == C && B == D) || (A == D && B == C)))
1133 return BinaryOperator::CreateXor(A, B);
1135 // ~(A&B) & (A|B) -> A^B
1136 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1137 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1138 ((A == C && B == D) || (A == D && B == C)))
1139 return BinaryOperator::CreateXor(A, B);
1141 if (Op0->hasOneUse() &&
1142 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
1143 if (A == Op1) { // (A^B)&A -> A&(A^B)
1144 I.swapOperands(); // Simplify below
1145 std::swap(Op0, Op1);
1146 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
1147 cast<BinaryOperator>(Op0)->swapOperands();
1148 I.swapOperands(); // Simplify below
1149 std::swap(Op0, Op1);
1153 if (Op1->hasOneUse() &&
1154 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
1155 if (B == Op0) { // B&(A^B) -> B&(B^A)
1156 cast<BinaryOperator>(Op1)->swapOperands();
1159 // Notice that the patten (A&(~B)) is actually (A&(-1^B)), so if
1160 // A is originally -1 (or a vector of -1 and undefs), then we enter
1161 // an endless loop. By checking that A is non-constant we ensure that
1162 // we will never get to the loop.
1163 if (A == Op0 && !isa<Constant>(A)) // A&(A^B) -> A & ~B
1164 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B, "tmp"));
1167 // (A&((~A)|B)) -> A&B
1168 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
1169 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
1170 return BinaryOperator::CreateAnd(A, Op1);
1171 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
1172 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
1173 return BinaryOperator::CreateAnd(A, Op0);
1176 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1))
1177 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
1178 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1179 return ReplaceInstUsesWith(I, Res);
1181 // If and'ing two fcmp, try combine them into one.
1182 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1183 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1184 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1185 return ReplaceInstUsesWith(I, Res);
1188 // fold (and (cast A), (cast B)) -> (cast (and A, B))
1189 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
1190 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) {
1191 const Type *SrcTy = Op0C->getOperand(0)->getType();
1192 if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ?
1193 SrcTy == Op1C->getOperand(0)->getType() &&
1194 SrcTy->isIntOrIntVectorTy()) {
1195 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
1197 // Only do this if the casts both really cause code to be generated.
1198 if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
1199 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
1200 Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName());
1201 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
1204 // If this is and(cast(icmp), cast(icmp)), try to fold this even if the
1205 // cast is otherwise not optimizable. This happens for vector sexts.
1206 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
1207 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
1208 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1209 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1211 // If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the
1212 // cast is otherwise not optimizable. This happens for vector sexts.
1213 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
1214 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
1215 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1216 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1220 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
1221 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
1222 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
1223 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
1224 SI0->getOperand(1) == SI1->getOperand(1) &&
1225 (SI0->hasOneUse() || SI1->hasOneUse())) {
1227 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
1229 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
1230 SI1->getOperand(1));
1234 return Changed ? &I : 0;
1237 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
1238 /// capable of providing pieces of a bswap. The subexpression provides pieces
1239 /// of a bswap if it is proven that each of the non-zero bytes in the output of
1240 /// the expression came from the corresponding "byte swapped" byte in some other
1241 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
1242 /// we know that the expression deposits the low byte of %X into the high byte
1243 /// of the bswap result and that all other bytes are zero. This expression is
1244 /// accepted, the high byte of ByteValues is set to X to indicate a correct
1247 /// This function returns true if the match was unsuccessful and false if so.
1248 /// On entry to the function the "OverallLeftShift" is a signed integer value
1249 /// indicating the number of bytes that the subexpression is later shifted. For
1250 /// example, if the expression is later right shifted by 16 bits, the
1251 /// OverallLeftShift value would be -2 on entry. This is used to specify which
1252 /// byte of ByteValues is actually being set.
1254 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
1255 /// byte is masked to zero by a user. For example, in (X & 255), X will be
1256 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
1257 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
1258 /// always in the local (OverallLeftShift) coordinate space.
1260 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
1261 SmallVector<Value*, 8> &ByteValues) {
1262 if (Instruction *I = dyn_cast<Instruction>(V)) {
1263 // If this is an or instruction, it may be an inner node of the bswap.
1264 if (I->getOpcode() == Instruction::Or) {
1265 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1267 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
1271 // If this is a logical shift by a constant multiple of 8, recurse with
1272 // OverallLeftShift and ByteMask adjusted.
1273 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
1275 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
1276 // Ensure the shift amount is defined and of a byte value.
1277 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
1280 unsigned ByteShift = ShAmt >> 3;
1281 if (I->getOpcode() == Instruction::Shl) {
1282 // X << 2 -> collect(X, +2)
1283 OverallLeftShift += ByteShift;
1284 ByteMask >>= ByteShift;
1286 // X >>u 2 -> collect(X, -2)
1287 OverallLeftShift -= ByteShift;
1288 ByteMask <<= ByteShift;
1289 ByteMask &= (~0U >> (32-ByteValues.size()));
1292 if (OverallLeftShift >= (int)ByteValues.size()) return true;
1293 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
1295 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1299 // If this is a logical 'and' with a mask that clears bytes, clear the
1300 // corresponding bytes in ByteMask.
1301 if (I->getOpcode() == Instruction::And &&
1302 isa<ConstantInt>(I->getOperand(1))) {
1303 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
1304 unsigned NumBytes = ByteValues.size();
1305 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
1306 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
1308 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
1309 // If this byte is masked out by a later operation, we don't care what
1311 if ((ByteMask & (1 << i)) == 0)
1314 // If the AndMask is all zeros for this byte, clear the bit.
1315 APInt MaskB = AndMask & Byte;
1317 ByteMask &= ~(1U << i);
1321 // If the AndMask is not all ones for this byte, it's not a bytezap.
1325 // Otherwise, this byte is kept.
1328 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1333 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
1334 // the input value to the bswap. Some observations: 1) if more than one byte
1335 // is demanded from this input, then it could not be successfully assembled
1336 // into a byteswap. At least one of the two bytes would not be aligned with
1337 // their ultimate destination.
1338 if (!isPowerOf2_32(ByteMask)) return true;
1339 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
1341 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
1342 // is demanded, it needs to go into byte 0 of the result. This means that the
1343 // byte needs to be shifted until it lands in the right byte bucket. The
1344 // shift amount depends on the position: if the byte is coming from the high
1345 // part of the value (e.g. byte 3) then it must be shifted right. If from the
1346 // low part, it must be shifted left.
1347 unsigned DestByteNo = InputByteNo + OverallLeftShift;
1348 if (InputByteNo < ByteValues.size()/2) {
1349 if (ByteValues.size()-1-DestByteNo != InputByteNo)
1352 if (ByteValues.size()-1-DestByteNo != InputByteNo)
1356 // If the destination byte value is already defined, the values are or'd
1357 // together, which isn't a bswap (unless it's an or of the same bits).
1358 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
1360 ByteValues[DestByteNo] = V;
1364 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
1365 /// If so, insert the new bswap intrinsic and return it.
1366 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
1367 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
1368 if (!ITy || ITy->getBitWidth() % 16 ||
1369 // ByteMask only allows up to 32-byte values.
1370 ITy->getBitWidth() > 32*8)
1371 return 0; // Can only bswap pairs of bytes. Can't do vectors.
1373 /// ByteValues - For each byte of the result, we keep track of which value
1374 /// defines each byte.
1375 SmallVector<Value*, 8> ByteValues;
1376 ByteValues.resize(ITy->getBitWidth()/8);
1378 // Try to find all the pieces corresponding to the bswap.
1379 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
1380 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
1383 // Check to see if all of the bytes come from the same value.
1384 Value *V = ByteValues[0];
1385 if (V == 0) return 0; // Didn't find a byte? Must be zero.
1387 // Check to make sure that all of the bytes come from the same value.
1388 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
1389 if (ByteValues[i] != V)
1391 const Type *Tys[] = { ITy };
1392 Module *M = I.getParent()->getParent()->getParent();
1393 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
1394 return CallInst::Create(F, V);
1397 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
1398 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
1399 /// we can simplify this expression to "cond ? C : D or B".
1400 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
1401 Value *C, Value *D) {
1402 // If A is not a select of -1/0, this cannot match.
1404 if (!match(A, m_SExt(m_Value(Cond))) ||
1405 !Cond->getType()->isIntegerTy(1))
1408 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
1409 if (match(D, m_Not(m_SExt(m_Specific(Cond)))))
1410 return SelectInst::Create(Cond, C, B);
1411 if (match(D, m_SExt(m_Not(m_Specific(Cond)))))
1412 return SelectInst::Create(Cond, C, B);
1414 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
1415 if (match(B, m_Not(m_SExt(m_Specific(Cond)))))
1416 return SelectInst::Create(Cond, C, D);
1417 if (match(B, m_SExt(m_Not(m_Specific(Cond)))))
1418 return SelectInst::Create(Cond, C, D);
1422 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
1423 Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
1424 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
1426 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
1427 if (PredicatesFoldable(LHSCC, RHSCC)) {
1428 if (LHS->getOperand(0) == RHS->getOperand(1) &&
1429 LHS->getOperand(1) == RHS->getOperand(0))
1430 LHS->swapOperands();
1431 if (LHS->getOperand(0) == RHS->getOperand(0) &&
1432 LHS->getOperand(1) == RHS->getOperand(1)) {
1433 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1434 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
1435 bool isSigned = LHS->isSigned() || RHS->isSigned();
1436 return getICmpValue(isSigned, Code, Op0, Op1, Builder);
1440 // handle (roughly):
1441 // (icmp ne (A & B), C) | (icmp ne (A & D), E)
1442 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, ICmpInst::ICMP_NE, Builder))
1445 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
1446 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
1447 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
1448 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
1449 if (LHSCst == 0 || RHSCst == 0) return 0;
1451 if (LHSCst == RHSCst && LHSCC == RHSCC) {
1452 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
1453 if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
1454 Value *NewOr = Builder->CreateOr(Val, Val2);
1455 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
1458 // (icmp slt A, 0) | (icmp slt B, 0) --> (icmp slt (A|B), 0)
1459 if (LHSCC == ICmpInst::ICMP_SLT && LHSCst->isZero()) {
1460 Value *NewOr = Builder->CreateOr(Val, Val2);
1461 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
1464 // (icmp sgt A, -1) | (icmp sgt B, -1) --> (icmp sgt (A&B), -1)
1465 if (LHSCC == ICmpInst::ICMP_SGT && LHSCst->isAllOnesValue()) {
1466 Value *NewAnd = Builder->CreateAnd(Val, Val2);
1467 return Builder->CreateICmp(LHSCC, NewAnd, LHSCst);
1471 // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
1472 // iff C2 + CA == C1.
1473 if (LHSCC == ICmpInst::ICMP_ULT && RHSCC == ICmpInst::ICMP_EQ) {
1474 ConstantInt *AddCst;
1475 if (match(Val, m_Add(m_Specific(Val2), m_ConstantInt(AddCst))))
1476 if (RHSCst->getValue() + AddCst->getValue() == LHSCst->getValue())
1477 return Builder->CreateICmpULE(Val, LHSCst);
1480 // From here on, we only handle:
1481 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
1482 if (Val != Val2) return 0;
1484 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
1485 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
1486 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
1487 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
1488 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
1491 // We can't fold (ugt x, C) | (sgt x, C2).
1492 if (!PredicatesFoldable(LHSCC, RHSCC))
1495 // Ensure that the larger constant is on the RHS.
1497 if (CmpInst::isSigned(LHSCC) ||
1498 (ICmpInst::isEquality(LHSCC) &&
1499 CmpInst::isSigned(RHSCC)))
1500 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
1502 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
1505 std::swap(LHS, RHS);
1506 std::swap(LHSCst, RHSCst);
1507 std::swap(LHSCC, RHSCC);
1510 // At this point, we know we have two icmp instructions
1511 // comparing a value against two constants and or'ing the result
1512 // together. Because of the above check, we know that we only have
1513 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
1514 // icmp folding check above), that the two constants are not
1516 assert(LHSCst != RHSCst && "Compares not folded above?");
1519 default: llvm_unreachable("Unknown integer condition code!");
1520 case ICmpInst::ICMP_EQ:
1522 default: llvm_unreachable("Unknown integer condition code!");
1523 case ICmpInst::ICMP_EQ:
1524 if (LHSCst == SubOne(RHSCst)) {
1525 // (X == 13 | X == 14) -> X-13 <u 2
1526 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1527 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1528 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
1529 return Builder->CreateICmpULT(Add, AddCST);
1531 break; // (X == 13 | X == 15) -> no change
1532 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
1533 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
1535 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
1536 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
1537 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
1541 case ICmpInst::ICMP_NE:
1543 default: llvm_unreachable("Unknown integer condition code!");
1544 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
1545 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
1546 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
1548 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
1549 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
1550 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
1551 return ConstantInt::getTrue(LHS->getContext());
1554 case ICmpInst::ICMP_ULT:
1556 default: llvm_unreachable("Unknown integer condition code!");
1557 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
1559 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
1560 // If RHSCst is [us]MAXINT, it is always false. Not handling
1561 // this can cause overflow.
1562 if (RHSCst->isMaxValue(false))
1564 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false);
1565 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
1567 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
1568 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
1570 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
1574 case ICmpInst::ICMP_SLT:
1576 default: llvm_unreachable("Unknown integer condition code!");
1577 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
1579 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
1580 // If RHSCst is [us]MAXINT, it is always false. Not handling
1581 // this can cause overflow.
1582 if (RHSCst->isMaxValue(true))
1584 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false);
1585 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
1587 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
1588 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
1590 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
1594 case ICmpInst::ICMP_UGT:
1596 default: llvm_unreachable("Unknown integer condition code!");
1597 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
1598 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
1600 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
1602 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
1603 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
1604 return ConstantInt::getTrue(LHS->getContext());
1605 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
1609 case ICmpInst::ICMP_SGT:
1611 default: llvm_unreachable("Unknown integer condition code!");
1612 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
1613 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
1615 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
1617 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
1618 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
1619 return ConstantInt::getTrue(LHS->getContext());
1620 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
1628 /// FoldOrOfFCmps - Optimize (fcmp)|(fcmp). NOTE: Unlike the rest of
1629 /// instcombine, this returns a Value which should already be inserted into the
1631 Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
1632 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
1633 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
1634 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
1635 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
1636 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
1637 // If either of the constants are nans, then the whole thing returns
1639 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
1640 return ConstantInt::getTrue(LHS->getContext());
1642 // Otherwise, no need to compare the two constants, compare the
1644 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
1647 // Handle vector zeros. This occurs because the canonical form of
1648 // "fcmp uno x,x" is "fcmp uno x, 0".
1649 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
1650 isa<ConstantAggregateZero>(RHS->getOperand(1)))
1651 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
1656 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
1657 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
1658 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
1660 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
1661 // Swap RHS operands to match LHS.
1662 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
1663 std::swap(Op1LHS, Op1RHS);
1665 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
1666 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
1668 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
1669 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
1670 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
1671 if (Op0CC == FCmpInst::FCMP_FALSE)
1673 if (Op1CC == FCmpInst::FCMP_FALSE)
1677 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
1678 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
1679 if (Op0Ordered == Op1Ordered) {
1680 // If both are ordered or unordered, return a new fcmp with
1681 // or'ed predicates.
1682 return getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS, Builder);
1688 /// FoldOrWithConstants - This helper function folds:
1690 /// ((A | B) & C1) | (B & C2)
1696 /// when the XOR of the two constants is "all ones" (-1).
1697 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
1698 Value *A, Value *B, Value *C) {
1699 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
1703 ConstantInt *CI2 = 0;
1704 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
1706 APInt Xor = CI1->getValue() ^ CI2->getValue();
1707 if (!Xor.isAllOnesValue()) return 0;
1709 if (V1 == A || V1 == B) {
1710 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
1711 return BinaryOperator::CreateOr(NewOp, V1);
1717 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
1718 bool Changed = SimplifyAssociativeOrCommutative(I);
1719 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1721 if (Value *V = SimplifyOrInst(Op0, Op1, TD))
1722 return ReplaceInstUsesWith(I, V);
1724 // (A&B)|(A&C) -> A&(B|C) etc
1725 if (Value *V = SimplifyUsingDistributiveLaws(I))
1726 return ReplaceInstUsesWith(I, V);
1728 // See if we can simplify any instructions used by the instruction whose sole
1729 // purpose is to compute bits we don't care about.
1730 if (SimplifyDemandedInstructionBits(I))
1733 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1734 ConstantInt *C1 = 0; Value *X = 0;
1735 // (X & C1) | C2 --> (X | C2) & (C1|C2)
1736 // iff (C1 & C2) == 0.
1737 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
1738 (RHS->getValue() & C1->getValue()) != 0 &&
1740 Value *Or = Builder->CreateOr(X, RHS);
1742 return BinaryOperator::CreateAnd(Or,
1743 ConstantInt::get(I.getContext(),
1744 RHS->getValue() | C1->getValue()));
1747 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
1748 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
1750 Value *Or = Builder->CreateOr(X, RHS);
1752 return BinaryOperator::CreateXor(Or,
1753 ConstantInt::get(I.getContext(),
1754 C1->getValue() & ~RHS->getValue()));
1757 // Try to fold constant and into select arguments.
1758 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1759 if (Instruction *R = FoldOpIntoSelect(I, SI))
1762 if (isa<PHINode>(Op0))
1763 if (Instruction *NV = FoldOpIntoPhi(I))
1767 Value *A = 0, *B = 0;
1768 ConstantInt *C1 = 0, *C2 = 0;
1770 // (A | B) | C and A | (B | C) -> bswap if possible.
1771 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
1772 if (match(Op0, m_Or(m_Value(), m_Value())) ||
1773 match(Op1, m_Or(m_Value(), m_Value())) ||
1774 (match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
1775 match(Op1, m_LogicalShift(m_Value(), m_Value())))) {
1776 if (Instruction *BSwap = MatchBSwap(I))
1780 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
1781 if (Op0->hasOneUse() &&
1782 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
1783 MaskedValueIsZero(Op1, C1->getValue())) {
1784 Value *NOr = Builder->CreateOr(A, Op1);
1786 return BinaryOperator::CreateXor(NOr, C1);
1789 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
1790 if (Op1->hasOneUse() &&
1791 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
1792 MaskedValueIsZero(Op0, C1->getValue())) {
1793 Value *NOr = Builder->CreateOr(A, Op0);
1795 return BinaryOperator::CreateXor(NOr, C1);
1799 Value *C = 0, *D = 0;
1800 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1801 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1802 Value *V1 = 0, *V2 = 0;
1803 C1 = dyn_cast<ConstantInt>(C);
1804 C2 = dyn_cast<ConstantInt>(D);
1805 if (C1 && C2) { // (A & C1)|(B & C2)
1806 // If we have: ((V + N) & C1) | (V & C2)
1807 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1808 // replace with V+N.
1809 if (C1->getValue() == ~C2->getValue()) {
1810 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
1811 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1812 // Add commutes, try both ways.
1813 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
1814 return ReplaceInstUsesWith(I, A);
1815 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
1816 return ReplaceInstUsesWith(I, A);
1818 // Or commutes, try both ways.
1819 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
1820 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1821 // Add commutes, try both ways.
1822 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
1823 return ReplaceInstUsesWith(I, B);
1824 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
1825 return ReplaceInstUsesWith(I, B);
1829 if ((C1->getValue() & C2->getValue()) == 0) {
1830 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
1831 // iff (C1&C2) == 0 and (N&~C1) == 0
1832 if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
1833 ((V1 == B && MaskedValueIsZero(V2, ~C1->getValue())) || // (V|N)
1834 (V2 == B && MaskedValueIsZero(V1, ~C1->getValue())))) // (N|V)
1835 return BinaryOperator::CreateAnd(A,
1836 ConstantInt::get(A->getContext(),
1837 C1->getValue()|C2->getValue()));
1838 // Or commutes, try both ways.
1839 if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
1840 ((V1 == A && MaskedValueIsZero(V2, ~C2->getValue())) || // (V|N)
1841 (V2 == A && MaskedValueIsZero(V1, ~C2->getValue())))) // (N|V)
1842 return BinaryOperator::CreateAnd(B,
1843 ConstantInt::get(B->getContext(),
1844 C1->getValue()|C2->getValue()));
1846 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
1847 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
1848 ConstantInt *C3 = 0, *C4 = 0;
1849 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
1850 (C3->getValue() & ~C1->getValue()) == 0 &&
1851 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
1852 (C4->getValue() & ~C2->getValue()) == 0) {
1853 V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
1854 return BinaryOperator::CreateAnd(V2,
1855 ConstantInt::get(B->getContext(),
1856 C1->getValue()|C2->getValue()));
1861 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants.
1862 // Don't do this for vector select idioms, the code generator doesn't handle
1864 if (!I.getType()->isVectorTy()) {
1865 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
1867 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
1869 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
1871 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
1875 // ((A&~B)|(~A&B)) -> A^B
1876 if ((match(C, m_Not(m_Specific(D))) &&
1877 match(B, m_Not(m_Specific(A)))))
1878 return BinaryOperator::CreateXor(A, D);
1879 // ((~B&A)|(~A&B)) -> A^B
1880 if ((match(A, m_Not(m_Specific(D))) &&
1881 match(B, m_Not(m_Specific(C)))))
1882 return BinaryOperator::CreateXor(C, D);
1883 // ((A&~B)|(B&~A)) -> A^B
1884 if ((match(C, m_Not(m_Specific(B))) &&
1885 match(D, m_Not(m_Specific(A)))))
1886 return BinaryOperator::CreateXor(A, B);
1887 // ((~B&A)|(B&~A)) -> A^B
1888 if ((match(A, m_Not(m_Specific(B))) &&
1889 match(D, m_Not(m_Specific(C)))))
1890 return BinaryOperator::CreateXor(C, B);
1892 // ((A|B)&1)|(B&-2) -> (A&1) | B
1893 if (match(A, m_Or(m_Value(V1), m_Specific(B))) ||
1894 match(A, m_Or(m_Specific(B), m_Value(V1)))) {
1895 Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C);
1896 if (Ret) return Ret;
1898 // (B&-2)|((A|B)&1) -> (A&1) | B
1899 if (match(B, m_Or(m_Specific(A), m_Value(V1))) ||
1900 match(B, m_Or(m_Value(V1), m_Specific(A)))) {
1901 Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D);
1902 if (Ret) return Ret;
1906 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
1907 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
1908 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
1909 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
1910 SI0->getOperand(1) == SI1->getOperand(1) &&
1911 (SI0->hasOneUse() || SI1->hasOneUse())) {
1912 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
1914 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
1915 SI1->getOperand(1));
1919 // (~A | ~B) == (~(A & B)) - De Morgan's Law
1920 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1921 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1922 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1923 Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
1924 I.getName()+".demorgan");
1925 return BinaryOperator::CreateNot(And);
1928 // Canonicalize xor to the RHS.
1929 if (match(Op0, m_Xor(m_Value(), m_Value())))
1930 std::swap(Op0, Op1);
1932 // A | ( A ^ B) -> A | B
1933 // A | (~A ^ B) -> A | ~B
1934 if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
1935 if (Op0 == A || Op0 == B)
1936 return BinaryOperator::CreateOr(A, B);
1938 if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
1939 Value *Not = Builder->CreateNot(B, B->getName()+".not");
1940 return BinaryOperator::CreateOr(Not, Op0);
1942 if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
1943 Value *Not = Builder->CreateNot(A, A->getName()+".not");
1944 return BinaryOperator::CreateOr(Not, Op0);
1948 // A | ~(A | B) -> A | ~B
1949 // A | ~(A ^ B) -> A | ~B
1950 if (match(Op1, m_Not(m_Value(A))))
1951 if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
1952 if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
1953 Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
1954 B->getOpcode() == Instruction::Xor)) {
1955 Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
1957 Value *Not = Builder->CreateNot(NotOp, NotOp->getName()+".not");
1958 return BinaryOperator::CreateOr(Not, Op0);
1961 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
1962 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
1963 if (Value *Res = FoldOrOfICmps(LHS, RHS))
1964 return ReplaceInstUsesWith(I, Res);
1966 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
1967 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1968 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1969 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
1970 return ReplaceInstUsesWith(I, Res);
1972 // fold (or (cast A), (cast B)) -> (cast (or A, B))
1973 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
1974 CastInst *Op1C = dyn_cast<CastInst>(Op1);
1975 if (Op1C && Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
1976 const Type *SrcTy = Op0C->getOperand(0)->getType();
1977 if (SrcTy == Op1C->getOperand(0)->getType() &&
1978 SrcTy->isIntOrIntVectorTy()) {
1979 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
1981 if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) &&
1982 // Only do this if the casts both really cause code to be
1984 ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
1985 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
1986 Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName());
1987 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
1990 // If this is or(cast(icmp), cast(icmp)), try to fold this even if the
1991 // cast is otherwise not optimizable. This happens for vector sexts.
1992 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
1993 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
1994 if (Value *Res = FoldOrOfICmps(LHS, RHS))
1995 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1997 // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the
1998 // cast is otherwise not optimizable. This happens for vector sexts.
1999 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
2000 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
2001 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
2002 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
2007 // or(sext(A), B) -> A ? -1 : B where A is an i1
2008 // or(A, sext(B)) -> B ? -1 : A where B is an i1
2009 if (match(Op0, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
2010 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1);
2011 if (match(Op1, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
2012 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0);
2014 // Note: If we've gotten to the point of visiting the outer OR, then the
2015 // inner one couldn't be simplified. If it was a constant, then it won't
2016 // be simplified by a later pass either, so we try swapping the inner/outer
2017 // ORs in the hopes that we'll be able to simplify it this way.
2018 // (X|C) | V --> (X|V) | C
2019 if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
2020 match(Op0, m_Or(m_Value(A), m_ConstantInt(C1)))) {
2021 Value *Inner = Builder->CreateOr(A, Op1);
2022 Inner->takeName(Op0);
2023 return BinaryOperator::CreateOr(Inner, C1);
2026 return Changed ? &I : 0;
2029 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
2030 bool Changed = SimplifyAssociativeOrCommutative(I);
2031 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2033 if (Value *V = SimplifyXorInst(Op0, Op1, TD))
2034 return ReplaceInstUsesWith(I, V);
2036 // (A&B)^(A&C) -> A&(B^C) etc
2037 if (Value *V = SimplifyUsingDistributiveLaws(I))
2038 return ReplaceInstUsesWith(I, V);
2040 // See if we can simplify any instructions used by the instruction whose sole
2041 // purpose is to compute bits we don't care about.
2042 if (SimplifyDemandedInstructionBits(I))
2045 // Is this a ~ operation?
2046 if (Value *NotOp = dyn_castNotVal(&I)) {
2047 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
2048 if (Op0I->getOpcode() == Instruction::And ||
2049 Op0I->getOpcode() == Instruction::Or) {
2050 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
2051 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
2052 if (dyn_castNotVal(Op0I->getOperand(1)))
2053 Op0I->swapOperands();
2054 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
2056 Builder->CreateNot(Op0I->getOperand(1),
2057 Op0I->getOperand(1)->getName()+".not");
2058 if (Op0I->getOpcode() == Instruction::And)
2059 return BinaryOperator::CreateOr(Op0NotVal, NotY);
2060 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
2063 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
2064 // ~(X | Y) === (~X & ~Y) - De Morgan's Law
2065 if (isFreeToInvert(Op0I->getOperand(0)) &&
2066 isFreeToInvert(Op0I->getOperand(1))) {
2068 Builder->CreateNot(Op0I->getOperand(0), "notlhs");
2070 Builder->CreateNot(Op0I->getOperand(1), "notrhs");
2071 if (Op0I->getOpcode() == Instruction::And)
2072 return BinaryOperator::CreateOr(NotX, NotY);
2073 return BinaryOperator::CreateAnd(NotX, NotY);
2076 } else if (Op0I->getOpcode() == Instruction::AShr) {
2077 // ~(~X >>s Y) --> (X >>s Y)
2078 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0)))
2079 return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1));
2085 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2086 if (RHS->isOne() && Op0->hasOneUse())
2087 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
2088 if (CmpInst *CI = dyn_cast<CmpInst>(Op0))
2089 return CmpInst::Create(CI->getOpcode(),
2090 CI->getInversePredicate(),
2091 CI->getOperand(0), CI->getOperand(1));
2093 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
2094 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2095 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
2096 if (CI->hasOneUse() && Op0C->hasOneUse()) {
2097 Instruction::CastOps Opcode = Op0C->getOpcode();
2098 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
2099 (RHS == ConstantExpr::getCast(Opcode,
2100 ConstantInt::getTrue(I.getContext()),
2101 Op0C->getDestTy()))) {
2102 CI->setPredicate(CI->getInversePredicate());
2103 return CastInst::Create(Opcode, CI, Op0C->getType());
2109 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2110 // ~(c-X) == X-c-1 == X+(-c-1)
2111 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
2112 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
2113 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
2114 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
2115 ConstantInt::get(I.getType(), 1));
2116 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
2119 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2120 if (Op0I->getOpcode() == Instruction::Add) {
2121 // ~(X-c) --> (-c-1)-X
2122 if (RHS->isAllOnesValue()) {
2123 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
2124 return BinaryOperator::CreateSub(
2125 ConstantExpr::getSub(NegOp0CI,
2126 ConstantInt::get(I.getType(), 1)),
2127 Op0I->getOperand(0));
2128 } else if (RHS->getValue().isSignBit()) {
2129 // (X + C) ^ signbit -> (X + C + signbit)
2130 Constant *C = ConstantInt::get(I.getContext(),
2131 RHS->getValue() + Op0CI->getValue());
2132 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
2135 } else if (Op0I->getOpcode() == Instruction::Or) {
2136 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
2137 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
2138 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
2139 // Anything in both C1 and C2 is known to be zero, remove it from
2141 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
2142 NewRHS = ConstantExpr::getAnd(NewRHS,
2143 ConstantExpr::getNot(CommonBits));
2145 I.setOperand(0, Op0I->getOperand(0));
2146 I.setOperand(1, NewRHS);
2153 // Try to fold constant and into select arguments.
2154 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2155 if (Instruction *R = FoldOpIntoSelect(I, SI))
2157 if (isa<PHINode>(Op0))
2158 if (Instruction *NV = FoldOpIntoPhi(I))
2162 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
2165 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2166 if (A == Op0) { // B^(B|A) == (A|B)^B
2167 Op1I->swapOperands();
2169 std::swap(Op0, Op1);
2170 } else if (B == Op0) { // B^(A|B) == (A|B)^B
2171 I.swapOperands(); // Simplified below.
2172 std::swap(Op0, Op1);
2174 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
2176 if (A == Op0) { // A^(A&B) -> A^(B&A)
2177 Op1I->swapOperands();
2180 if (B == Op0) { // A^(B&A) -> (B&A)^A
2181 I.swapOperands(); // Simplified below.
2182 std::swap(Op0, Op1);
2187 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
2190 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2191 Op0I->hasOneUse()) {
2192 if (A == Op1) // (B|A)^B == (A|B)^B
2194 if (B == Op1) // (A|B)^B == A & ~B
2195 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1, "tmp"));
2196 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2198 if (A == Op1) // (A&B)^A -> (B&A)^A
2200 if (B == Op1 && // (B&A)^A == ~B & A
2201 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
2202 return BinaryOperator::CreateAnd(Builder->CreateNot(A, "tmp"), Op1);
2207 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
2208 if (Op0I && Op1I && Op0I->isShift() &&
2209 Op0I->getOpcode() == Op1I->getOpcode() &&
2210 Op0I->getOperand(1) == Op1I->getOperand(1) &&
2211 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
2213 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
2215 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
2216 Op1I->getOperand(1));
2220 Value *A, *B, *C, *D;
2221 // (A & B)^(A | B) -> A ^ B
2222 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2223 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
2224 if ((A == C && B == D) || (A == D && B == C))
2225 return BinaryOperator::CreateXor(A, B);
2227 // (A | B)^(A & B) -> A ^ B
2228 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2229 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
2230 if ((A == C && B == D) || (A == D && B == C))
2231 return BinaryOperator::CreateXor(A, B);
2235 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
2236 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
2237 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
2238 if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
2239 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2240 LHS->getOperand(1) == RHS->getOperand(0))
2241 LHS->swapOperands();
2242 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2243 LHS->getOperand(1) == RHS->getOperand(1)) {
2244 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2245 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
2246 bool isSigned = LHS->isSigned() || RHS->isSigned();
2247 return ReplaceInstUsesWith(I,
2248 getICmpValue(isSigned, Code, Op0, Op1, Builder));
2252 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
2253 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2254 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
2255 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
2256 const Type *SrcTy = Op0C->getOperand(0)->getType();
2257 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() &&
2258 // Only do this if the casts both really cause code to be generated.
2259 ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0),
2261 ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0),
2263 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
2264 Op1C->getOperand(0), I.getName());
2265 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2270 return Changed ? &I : 0;