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 pratice, 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);
761 // From here on, we only handle:
762 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
763 if (Val != Val2) return 0;
765 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
766 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
767 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
768 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
769 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
772 // Make a constant range that's the intersection of the two icmp ranges.
773 // If the intersection is empty, we know that the result is false.
774 ConstantRange LHSRange =
775 ConstantRange::makeICmpRegion(LHSCC, LHSCst->getValue());
776 ConstantRange RHSRange =
777 ConstantRange::makeICmpRegion(RHSCC, RHSCst->getValue());
779 if (LHSRange.intersectWith(RHSRange).isEmptySet())
780 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
782 // We can't fold (ugt x, C) & (sgt x, C2).
783 if (!PredicatesFoldable(LHSCC, RHSCC))
786 // Ensure that the larger constant is on the RHS.
788 if (CmpInst::isSigned(LHSCC) ||
789 (ICmpInst::isEquality(LHSCC) &&
790 CmpInst::isSigned(RHSCC)))
791 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
793 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
797 std::swap(LHSCst, RHSCst);
798 std::swap(LHSCC, RHSCC);
801 // At this point, we know we have two icmp instructions
802 // comparing a value against two constants and and'ing the result
803 // together. Because of the above check, we know that we only have
804 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
805 // (from the icmp folding check above), that the two constants
806 // are not equal and that the larger constant is on the RHS
807 assert(LHSCst != RHSCst && "Compares not folded above?");
810 default: llvm_unreachable("Unknown integer condition code!");
811 case ICmpInst::ICMP_EQ:
813 default: llvm_unreachable("Unknown integer condition code!");
814 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
815 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
816 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
819 case ICmpInst::ICMP_NE:
821 default: llvm_unreachable("Unknown integer condition code!");
822 case ICmpInst::ICMP_ULT:
823 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
824 return Builder->CreateICmpULT(Val, LHSCst);
825 break; // (X != 13 & X u< 15) -> no change
826 case ICmpInst::ICMP_SLT:
827 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
828 return Builder->CreateICmpSLT(Val, LHSCst);
829 break; // (X != 13 & X s< 15) -> no change
830 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
831 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
832 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
834 case ICmpInst::ICMP_NE:
835 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
836 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
837 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
838 return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1));
840 break; // (X != 13 & X != 15) -> no change
843 case ICmpInst::ICMP_ULT:
845 default: llvm_unreachable("Unknown integer condition code!");
846 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
847 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
848 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
849 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
851 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
852 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
854 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
858 case ICmpInst::ICMP_SLT:
860 default: llvm_unreachable("Unknown integer condition code!");
861 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
863 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
864 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
866 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
870 case ICmpInst::ICMP_UGT:
872 default: llvm_unreachable("Unknown integer condition code!");
873 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
874 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
876 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
878 case ICmpInst::ICMP_NE:
879 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
880 return Builder->CreateICmp(LHSCC, Val, RHSCst);
881 break; // (X u> 13 & X != 15) -> no change
882 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
883 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
884 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
888 case ICmpInst::ICMP_SGT:
890 default: llvm_unreachable("Unknown integer condition code!");
891 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
892 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
894 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
896 case ICmpInst::ICMP_NE:
897 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
898 return Builder->CreateICmp(LHSCC, Val, RHSCst);
899 break; // (X s> 13 & X != 15) -> no change
900 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
901 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true);
902 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
911 /// FoldAndOfFCmps - Optimize (fcmp)&(fcmp). NOTE: Unlike the rest of
912 /// instcombine, this returns a Value which should already be inserted into the
914 Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
915 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
916 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
917 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
918 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
919 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
920 // If either of the constants are nans, then the whole thing returns
922 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
923 return ConstantInt::getFalse(LHS->getContext());
924 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
927 // Handle vector zeros. This occurs because the canonical form of
928 // "fcmp ord x,x" is "fcmp ord x, 0".
929 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
930 isa<ConstantAggregateZero>(RHS->getOperand(1)))
931 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
935 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
936 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
937 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
940 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
941 // Swap RHS operands to match LHS.
942 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
943 std::swap(Op1LHS, Op1RHS);
946 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
947 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
949 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
950 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
951 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
952 if (Op0CC == FCmpInst::FCMP_TRUE)
954 if (Op1CC == FCmpInst::FCMP_TRUE)
959 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
960 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
963 std::swap(Op0Pred, Op1Pred);
964 std::swap(Op0Ordered, Op1Ordered);
967 // uno && ueq -> uno && (uno || eq) -> ueq
968 // ord && olt -> ord && (ord && lt) -> olt
969 if (Op0Ordered == Op1Ordered)
972 // uno && oeq -> uno && (ord && eq) -> false
973 // uno && ord -> false
975 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
976 // ord && ueq -> ord && (uno || eq) -> oeq
977 return getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS, Builder);
985 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
986 bool Changed = SimplifyAssociativeOrCommutative(I);
987 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
989 if (Value *V = SimplifyAndInst(Op0, Op1, TD))
990 return ReplaceInstUsesWith(I, V);
992 // (A|B)&(A|C) -> A|(B&C) etc
993 if (Value *V = SimplifyUsingDistributiveLaws(I))
994 return ReplaceInstUsesWith(I, V);
996 // See if we can simplify any instructions used by the instruction whose sole
997 // purpose is to compute bits we don't care about.
998 if (SimplifyDemandedInstructionBits(I))
1001 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
1002 const APInt &AndRHSMask = AndRHS->getValue();
1004 // Optimize a variety of ((val OP C1) & C2) combinations...
1005 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1006 Value *Op0LHS = Op0I->getOperand(0);
1007 Value *Op0RHS = Op0I->getOperand(1);
1008 switch (Op0I->getOpcode()) {
1010 case Instruction::Xor:
1011 case Instruction::Or: {
1012 // If the mask is only needed on one incoming arm, push it up.
1013 if (!Op0I->hasOneUse()) break;
1015 APInt NotAndRHS(~AndRHSMask);
1016 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
1017 // Not masking anything out for the LHS, move to RHS.
1018 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
1019 Op0RHS->getName()+".masked");
1020 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
1022 if (!isa<Constant>(Op0RHS) &&
1023 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
1024 // Not masking anything out for the RHS, move to LHS.
1025 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
1026 Op0LHS->getName()+".masked");
1027 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
1032 case Instruction::Add:
1033 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
1034 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1035 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1036 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
1037 return BinaryOperator::CreateAnd(V, AndRHS);
1038 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
1039 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
1042 case Instruction::Sub:
1043 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
1044 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1045 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1046 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
1047 return BinaryOperator::CreateAnd(V, AndRHS);
1049 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
1050 // has 1's for all bits that the subtraction with A might affect.
1051 if (Op0I->hasOneUse() && !match(Op0LHS, m_Zero())) {
1052 uint32_t BitWidth = AndRHSMask.getBitWidth();
1053 uint32_t Zeros = AndRHSMask.countLeadingZeros();
1054 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
1056 if (MaskedValueIsZero(Op0LHS, Mask)) {
1057 Value *NewNeg = Builder->CreateNeg(Op0RHS);
1058 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
1063 case Instruction::Shl:
1064 case Instruction::LShr:
1065 // (1 << x) & 1 --> zext(x == 0)
1066 // (1 >> x) & 1 --> zext(x == 0)
1067 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
1069 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
1070 return new ZExtInst(NewICmp, I.getType());
1075 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
1076 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
1080 // If this is an integer truncation, and if the source is an 'and' with
1081 // immediate, transform it. This frequently occurs for bitfield accesses.
1083 Value *X = 0; ConstantInt *YC = 0;
1084 if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) {
1085 // Change: and (trunc (and X, YC) to T), C2
1086 // into : and (trunc X to T), trunc(YC) & C2
1087 // This will fold the two constants together, which may allow
1088 // other simplifications.
1089 Value *NewCast = Builder->CreateTrunc(X, I.getType(), "and.shrunk");
1090 Constant *C3 = ConstantExpr::getTrunc(YC, I.getType());
1091 C3 = ConstantExpr::getAnd(C3, AndRHS);
1092 return BinaryOperator::CreateAnd(NewCast, C3);
1096 // Try to fold constant and into select arguments.
1097 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1098 if (Instruction *R = FoldOpIntoSelect(I, SI))
1100 if (isa<PHINode>(Op0))
1101 if (Instruction *NV = FoldOpIntoPhi(I))
1106 // (~A & ~B) == (~(A | B)) - De Morgan's Law
1107 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1108 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1109 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1110 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
1111 I.getName()+".demorgan");
1112 return BinaryOperator::CreateNot(Or);
1116 Value *A = 0, *B = 0, *C = 0, *D = 0;
1117 // (A|B) & ~(A&B) -> A^B
1118 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1119 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1120 ((A == C && B == D) || (A == D && B == C)))
1121 return BinaryOperator::CreateXor(A, B);
1123 // ~(A&B) & (A|B) -> A^B
1124 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1125 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1126 ((A == C && B == D) || (A == D && B == C)))
1127 return BinaryOperator::CreateXor(A, B);
1129 if (Op0->hasOneUse() &&
1130 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
1131 if (A == Op1) { // (A^B)&A -> A&(A^B)
1132 I.swapOperands(); // Simplify below
1133 std::swap(Op0, Op1);
1134 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
1135 cast<BinaryOperator>(Op0)->swapOperands();
1136 I.swapOperands(); // Simplify below
1137 std::swap(Op0, Op1);
1141 if (Op1->hasOneUse() &&
1142 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
1143 if (B == Op0) { // B&(A^B) -> B&(B^A)
1144 cast<BinaryOperator>(Op1)->swapOperands();
1147 // Notice that the patten (A&(~B)) is actually (A&(-1^B)), so if
1148 // A is originally -1 (or a vector of -1 and undefs), then we enter
1149 // an endless loop. By checking that A is non-constant we ensure that
1150 // we will never get to the loop.
1151 if (A == Op0 && !isa<Constant>(A)) // A&(A^B) -> A & ~B
1152 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B, "tmp"));
1155 // (A&((~A)|B)) -> A&B
1156 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
1157 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
1158 return BinaryOperator::CreateAnd(A, Op1);
1159 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
1160 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
1161 return BinaryOperator::CreateAnd(A, Op0);
1164 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1))
1165 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
1166 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1167 return ReplaceInstUsesWith(I, Res);
1169 // If and'ing two fcmp, try combine them into one.
1170 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1171 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1172 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1173 return ReplaceInstUsesWith(I, Res);
1176 // fold (and (cast A), (cast B)) -> (cast (and A, B))
1177 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
1178 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) {
1179 const Type *SrcTy = Op0C->getOperand(0)->getType();
1180 if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ?
1181 SrcTy == Op1C->getOperand(0)->getType() &&
1182 SrcTy->isIntOrIntVectorTy()) {
1183 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
1185 // Only do this if the casts both really cause code to be generated.
1186 if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
1187 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
1188 Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName());
1189 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
1192 // If this is and(cast(icmp), cast(icmp)), try to fold this even if the
1193 // cast is otherwise not optimizable. This happens for vector sexts.
1194 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
1195 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
1196 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1197 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1199 // If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the
1200 // cast is otherwise not optimizable. This happens for vector sexts.
1201 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
1202 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
1203 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1204 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1208 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
1209 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
1210 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
1211 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
1212 SI0->getOperand(1) == SI1->getOperand(1) &&
1213 (SI0->hasOneUse() || SI1->hasOneUse())) {
1215 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
1217 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
1218 SI1->getOperand(1));
1222 return Changed ? &I : 0;
1225 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
1226 /// capable of providing pieces of a bswap. The subexpression provides pieces
1227 /// of a bswap if it is proven that each of the non-zero bytes in the output of
1228 /// the expression came from the corresponding "byte swapped" byte in some other
1229 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
1230 /// we know that the expression deposits the low byte of %X into the high byte
1231 /// of the bswap result and that all other bytes are zero. This expression is
1232 /// accepted, the high byte of ByteValues is set to X to indicate a correct
1235 /// This function returns true if the match was unsuccessful and false if so.
1236 /// On entry to the function the "OverallLeftShift" is a signed integer value
1237 /// indicating the number of bytes that the subexpression is later shifted. For
1238 /// example, if the expression is later right shifted by 16 bits, the
1239 /// OverallLeftShift value would be -2 on entry. This is used to specify which
1240 /// byte of ByteValues is actually being set.
1242 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
1243 /// byte is masked to zero by a user. For example, in (X & 255), X will be
1244 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
1245 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
1246 /// always in the local (OverallLeftShift) coordinate space.
1248 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
1249 SmallVector<Value*, 8> &ByteValues) {
1250 if (Instruction *I = dyn_cast<Instruction>(V)) {
1251 // If this is an or instruction, it may be an inner node of the bswap.
1252 if (I->getOpcode() == Instruction::Or) {
1253 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1255 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
1259 // If this is a logical shift by a constant multiple of 8, recurse with
1260 // OverallLeftShift and ByteMask adjusted.
1261 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
1263 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
1264 // Ensure the shift amount is defined and of a byte value.
1265 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
1268 unsigned ByteShift = ShAmt >> 3;
1269 if (I->getOpcode() == Instruction::Shl) {
1270 // X << 2 -> collect(X, +2)
1271 OverallLeftShift += ByteShift;
1272 ByteMask >>= ByteShift;
1274 // X >>u 2 -> collect(X, -2)
1275 OverallLeftShift -= ByteShift;
1276 ByteMask <<= ByteShift;
1277 ByteMask &= (~0U >> (32-ByteValues.size()));
1280 if (OverallLeftShift >= (int)ByteValues.size()) return true;
1281 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
1283 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1287 // If this is a logical 'and' with a mask that clears bytes, clear the
1288 // corresponding bytes in ByteMask.
1289 if (I->getOpcode() == Instruction::And &&
1290 isa<ConstantInt>(I->getOperand(1))) {
1291 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
1292 unsigned NumBytes = ByteValues.size();
1293 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
1294 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
1296 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
1297 // If this byte is masked out by a later operation, we don't care what
1299 if ((ByteMask & (1 << i)) == 0)
1302 // If the AndMask is all zeros for this byte, clear the bit.
1303 APInt MaskB = AndMask & Byte;
1305 ByteMask &= ~(1U << i);
1309 // If the AndMask is not all ones for this byte, it's not a bytezap.
1313 // Otherwise, this byte is kept.
1316 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1321 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
1322 // the input value to the bswap. Some observations: 1) if more than one byte
1323 // is demanded from this input, then it could not be successfully assembled
1324 // into a byteswap. At least one of the two bytes would not be aligned with
1325 // their ultimate destination.
1326 if (!isPowerOf2_32(ByteMask)) return true;
1327 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
1329 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
1330 // is demanded, it needs to go into byte 0 of the result. This means that the
1331 // byte needs to be shifted until it lands in the right byte bucket. The
1332 // shift amount depends on the position: if the byte is coming from the high
1333 // part of the value (e.g. byte 3) then it must be shifted right. If from the
1334 // low part, it must be shifted left.
1335 unsigned DestByteNo = InputByteNo + OverallLeftShift;
1336 if (InputByteNo < ByteValues.size()/2) {
1337 if (ByteValues.size()-1-DestByteNo != InputByteNo)
1340 if (ByteValues.size()-1-DestByteNo != InputByteNo)
1344 // If the destination byte value is already defined, the values are or'd
1345 // together, which isn't a bswap (unless it's an or of the same bits).
1346 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
1348 ByteValues[DestByteNo] = V;
1352 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
1353 /// If so, insert the new bswap intrinsic and return it.
1354 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
1355 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
1356 if (!ITy || ITy->getBitWidth() % 16 ||
1357 // ByteMask only allows up to 32-byte values.
1358 ITy->getBitWidth() > 32*8)
1359 return 0; // Can only bswap pairs of bytes. Can't do vectors.
1361 /// ByteValues - For each byte of the result, we keep track of which value
1362 /// defines each byte.
1363 SmallVector<Value*, 8> ByteValues;
1364 ByteValues.resize(ITy->getBitWidth()/8);
1366 // Try to find all the pieces corresponding to the bswap.
1367 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
1368 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
1371 // Check to see if all of the bytes come from the same value.
1372 Value *V = ByteValues[0];
1373 if (V == 0) return 0; // Didn't find a byte? Must be zero.
1375 // Check to make sure that all of the bytes come from the same value.
1376 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
1377 if (ByteValues[i] != V)
1379 const Type *Tys[] = { ITy };
1380 Module *M = I.getParent()->getParent()->getParent();
1381 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
1382 return CallInst::Create(F, V);
1385 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
1386 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
1387 /// we can simplify this expression to "cond ? C : D or B".
1388 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
1389 Value *C, Value *D) {
1390 // If A is not a select of -1/0, this cannot match.
1392 if (!match(A, m_SExt(m_Value(Cond))) ||
1393 !Cond->getType()->isIntegerTy(1))
1396 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
1397 if (match(D, m_Not(m_SExt(m_Specific(Cond)))))
1398 return SelectInst::Create(Cond, C, B);
1399 if (match(D, m_SExt(m_Not(m_Specific(Cond)))))
1400 return SelectInst::Create(Cond, C, B);
1402 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
1403 if (match(B, m_Not(m_SExt(m_Specific(Cond)))))
1404 return SelectInst::Create(Cond, C, D);
1405 if (match(B, m_SExt(m_Not(m_Specific(Cond)))))
1406 return SelectInst::Create(Cond, C, D);
1410 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
1411 Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
1412 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
1414 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
1415 if (PredicatesFoldable(LHSCC, RHSCC)) {
1416 if (LHS->getOperand(0) == RHS->getOperand(1) &&
1417 LHS->getOperand(1) == RHS->getOperand(0))
1418 LHS->swapOperands();
1419 if (LHS->getOperand(0) == RHS->getOperand(0) &&
1420 LHS->getOperand(1) == RHS->getOperand(1)) {
1421 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1422 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
1423 bool isSigned = LHS->isSigned() || RHS->isSigned();
1424 return getICmpValue(isSigned, Code, Op0, Op1, Builder);
1428 // handle (roughly):
1429 // (icmp ne (A & B), C) | (icmp ne (A & D), E)
1430 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, ICmpInst::ICMP_NE, Builder))
1433 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
1434 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
1435 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
1436 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
1437 if (LHSCst == 0 || RHSCst == 0) return 0;
1439 if (LHSCst == RHSCst && LHSCC == RHSCC) {
1440 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
1441 if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
1442 Value *NewOr = Builder->CreateOr(Val, Val2);
1443 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
1447 // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
1448 // iff C2 + CA == C1.
1449 if (LHSCC == ICmpInst::ICMP_ULT && RHSCC == ICmpInst::ICMP_EQ) {
1450 ConstantInt *AddCst;
1451 if (match(Val, m_Add(m_Specific(Val2), m_ConstantInt(AddCst))))
1452 if (RHSCst->getValue() + AddCst->getValue() == LHSCst->getValue())
1453 return Builder->CreateICmpULE(Val, LHSCst);
1456 // From here on, we only handle:
1457 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
1458 if (Val != Val2) return 0;
1460 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
1461 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
1462 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
1463 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
1464 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
1467 // We can't fold (ugt x, C) | (sgt x, C2).
1468 if (!PredicatesFoldable(LHSCC, RHSCC))
1471 // Ensure that the larger constant is on the RHS.
1473 if (CmpInst::isSigned(LHSCC) ||
1474 (ICmpInst::isEquality(LHSCC) &&
1475 CmpInst::isSigned(RHSCC)))
1476 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
1478 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
1481 std::swap(LHS, RHS);
1482 std::swap(LHSCst, RHSCst);
1483 std::swap(LHSCC, RHSCC);
1486 // At this point, we know we have two icmp instructions
1487 // comparing a value against two constants and or'ing the result
1488 // together. Because of the above check, we know that we only have
1489 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
1490 // icmp folding check above), that the two constants are not
1492 assert(LHSCst != RHSCst && "Compares not folded above?");
1495 default: llvm_unreachable("Unknown integer condition code!");
1496 case ICmpInst::ICMP_EQ:
1498 default: llvm_unreachable("Unknown integer condition code!");
1499 case ICmpInst::ICMP_EQ:
1500 if (LHSCst == SubOne(RHSCst)) {
1501 // (X == 13 | X == 14) -> X-13 <u 2
1502 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1503 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1504 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
1505 return Builder->CreateICmpULT(Add, AddCST);
1507 break; // (X == 13 | X == 15) -> no change
1508 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
1509 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
1511 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
1512 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
1513 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
1517 case ICmpInst::ICMP_NE:
1519 default: llvm_unreachable("Unknown integer condition code!");
1520 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
1521 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
1522 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
1524 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
1525 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
1526 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
1527 return ConstantInt::getTrue(LHS->getContext());
1530 case ICmpInst::ICMP_ULT:
1532 default: llvm_unreachable("Unknown integer condition code!");
1533 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
1535 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
1536 // If RHSCst is [us]MAXINT, it is always false. Not handling
1537 // this can cause overflow.
1538 if (RHSCst->isMaxValue(false))
1540 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false);
1541 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
1543 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
1544 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
1546 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
1550 case ICmpInst::ICMP_SLT:
1552 default: llvm_unreachable("Unknown integer condition code!");
1553 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
1555 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
1556 // If RHSCst is [us]MAXINT, it is always false. Not handling
1557 // this can cause overflow.
1558 if (RHSCst->isMaxValue(true))
1560 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false);
1561 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
1563 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
1564 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
1566 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
1570 case ICmpInst::ICMP_UGT:
1572 default: llvm_unreachable("Unknown integer condition code!");
1573 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
1574 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
1576 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
1578 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
1579 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
1580 return ConstantInt::getTrue(LHS->getContext());
1581 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
1585 case ICmpInst::ICMP_SGT:
1587 default: llvm_unreachable("Unknown integer condition code!");
1588 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
1589 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
1591 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
1593 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
1594 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
1595 return ConstantInt::getTrue(LHS->getContext());
1596 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
1604 /// FoldOrOfFCmps - Optimize (fcmp)|(fcmp). NOTE: Unlike the rest of
1605 /// instcombine, this returns a Value which should already be inserted into the
1607 Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
1608 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
1609 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
1610 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
1611 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
1612 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
1613 // If either of the constants are nans, then the whole thing returns
1615 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
1616 return ConstantInt::getTrue(LHS->getContext());
1618 // Otherwise, no need to compare the two constants, compare the
1620 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
1623 // Handle vector zeros. This occurs because the canonical form of
1624 // "fcmp uno x,x" is "fcmp uno x, 0".
1625 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
1626 isa<ConstantAggregateZero>(RHS->getOperand(1)))
1627 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
1632 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
1633 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
1634 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
1636 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
1637 // Swap RHS operands to match LHS.
1638 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
1639 std::swap(Op1LHS, Op1RHS);
1641 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
1642 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
1644 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
1645 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
1646 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
1647 if (Op0CC == FCmpInst::FCMP_FALSE)
1649 if (Op1CC == FCmpInst::FCMP_FALSE)
1653 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
1654 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
1655 if (Op0Ordered == Op1Ordered) {
1656 // If both are ordered or unordered, return a new fcmp with
1657 // or'ed predicates.
1658 return getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS, Builder);
1664 /// FoldOrWithConstants - This helper function folds:
1666 /// ((A | B) & C1) | (B & C2)
1672 /// when the XOR of the two constants is "all ones" (-1).
1673 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
1674 Value *A, Value *B, Value *C) {
1675 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
1679 ConstantInt *CI2 = 0;
1680 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
1682 APInt Xor = CI1->getValue() ^ CI2->getValue();
1683 if (!Xor.isAllOnesValue()) return 0;
1685 if (V1 == A || V1 == B) {
1686 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
1687 return BinaryOperator::CreateOr(NewOp, V1);
1693 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
1694 bool Changed = SimplifyAssociativeOrCommutative(I);
1695 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1697 if (Value *V = SimplifyOrInst(Op0, Op1, TD))
1698 return ReplaceInstUsesWith(I, V);
1700 // (A&B)|(A&C) -> A&(B|C) etc
1701 if (Value *V = SimplifyUsingDistributiveLaws(I))
1702 return ReplaceInstUsesWith(I, V);
1704 // See if we can simplify any instructions used by the instruction whose sole
1705 // purpose is to compute bits we don't care about.
1706 if (SimplifyDemandedInstructionBits(I))
1709 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1710 ConstantInt *C1 = 0; Value *X = 0;
1711 // (X & C1) | C2 --> (X | C2) & (C1|C2)
1712 // iff (C1 & C2) == 0.
1713 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
1714 (RHS->getValue() & C1->getValue()) != 0 &&
1716 Value *Or = Builder->CreateOr(X, RHS);
1718 return BinaryOperator::CreateAnd(Or,
1719 ConstantInt::get(I.getContext(),
1720 RHS->getValue() | C1->getValue()));
1723 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
1724 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
1726 Value *Or = Builder->CreateOr(X, RHS);
1728 return BinaryOperator::CreateXor(Or,
1729 ConstantInt::get(I.getContext(),
1730 C1->getValue() & ~RHS->getValue()));
1733 // Try to fold constant and into select arguments.
1734 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1735 if (Instruction *R = FoldOpIntoSelect(I, SI))
1738 if (isa<PHINode>(Op0))
1739 if (Instruction *NV = FoldOpIntoPhi(I))
1743 Value *A = 0, *B = 0;
1744 ConstantInt *C1 = 0, *C2 = 0;
1746 // (A | B) | C and A | (B | C) -> bswap if possible.
1747 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
1748 if (match(Op0, m_Or(m_Value(), m_Value())) ||
1749 match(Op1, m_Or(m_Value(), m_Value())) ||
1750 (match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
1751 match(Op1, m_LogicalShift(m_Value(), m_Value())))) {
1752 if (Instruction *BSwap = MatchBSwap(I))
1756 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
1757 if (Op0->hasOneUse() &&
1758 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
1759 MaskedValueIsZero(Op1, C1->getValue())) {
1760 Value *NOr = Builder->CreateOr(A, Op1);
1762 return BinaryOperator::CreateXor(NOr, C1);
1765 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
1766 if (Op1->hasOneUse() &&
1767 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
1768 MaskedValueIsZero(Op0, C1->getValue())) {
1769 Value *NOr = Builder->CreateOr(A, Op0);
1771 return BinaryOperator::CreateXor(NOr, C1);
1775 Value *C = 0, *D = 0;
1776 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1777 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1778 Value *V1 = 0, *V2 = 0;
1779 C1 = dyn_cast<ConstantInt>(C);
1780 C2 = dyn_cast<ConstantInt>(D);
1781 if (C1 && C2) { // (A & C1)|(B & C2)
1782 // If we have: ((V + N) & C1) | (V & C2)
1783 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1784 // replace with V+N.
1785 if (C1->getValue() == ~C2->getValue()) {
1786 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
1787 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1788 // Add commutes, try both ways.
1789 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
1790 return ReplaceInstUsesWith(I, A);
1791 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
1792 return ReplaceInstUsesWith(I, A);
1794 // Or commutes, try both ways.
1795 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
1796 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1797 // Add commutes, try both ways.
1798 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
1799 return ReplaceInstUsesWith(I, B);
1800 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
1801 return ReplaceInstUsesWith(I, B);
1805 if ((C1->getValue() & C2->getValue()) == 0) {
1806 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
1807 // iff (C1&C2) == 0 and (N&~C1) == 0
1808 if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
1809 ((V1 == B && MaskedValueIsZero(V2, ~C1->getValue())) || // (V|N)
1810 (V2 == B && MaskedValueIsZero(V1, ~C1->getValue())))) // (N|V)
1811 return BinaryOperator::CreateAnd(A,
1812 ConstantInt::get(A->getContext(),
1813 C1->getValue()|C2->getValue()));
1814 // Or commutes, try both ways.
1815 if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
1816 ((V1 == A && MaskedValueIsZero(V2, ~C2->getValue())) || // (V|N)
1817 (V2 == A && MaskedValueIsZero(V1, ~C2->getValue())))) // (N|V)
1818 return BinaryOperator::CreateAnd(B,
1819 ConstantInt::get(B->getContext(),
1820 C1->getValue()|C2->getValue()));
1822 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
1823 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
1824 ConstantInt *C3 = 0, *C4 = 0;
1825 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
1826 (C3->getValue() & ~C1->getValue()) == 0 &&
1827 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
1828 (C4->getValue() & ~C2->getValue()) == 0) {
1829 V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
1830 return BinaryOperator::CreateAnd(V2,
1831 ConstantInt::get(B->getContext(),
1832 C1->getValue()|C2->getValue()));
1837 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants.
1838 // Don't do this for vector select idioms, the code generator doesn't handle
1840 if (!I.getType()->isVectorTy()) {
1841 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
1843 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
1845 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
1847 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
1851 // ((A&~B)|(~A&B)) -> A^B
1852 if ((match(C, m_Not(m_Specific(D))) &&
1853 match(B, m_Not(m_Specific(A)))))
1854 return BinaryOperator::CreateXor(A, D);
1855 // ((~B&A)|(~A&B)) -> A^B
1856 if ((match(A, m_Not(m_Specific(D))) &&
1857 match(B, m_Not(m_Specific(C)))))
1858 return BinaryOperator::CreateXor(C, D);
1859 // ((A&~B)|(B&~A)) -> A^B
1860 if ((match(C, m_Not(m_Specific(B))) &&
1861 match(D, m_Not(m_Specific(A)))))
1862 return BinaryOperator::CreateXor(A, B);
1863 // ((~B&A)|(B&~A)) -> A^B
1864 if ((match(A, m_Not(m_Specific(B))) &&
1865 match(D, m_Not(m_Specific(C)))))
1866 return BinaryOperator::CreateXor(C, B);
1868 // ((A|B)&1)|(B&-2) -> (A&1) | B
1869 if (match(A, m_Or(m_Value(V1), m_Specific(B))) ||
1870 match(A, m_Or(m_Specific(B), m_Value(V1)))) {
1871 Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C);
1872 if (Ret) return Ret;
1874 // (B&-2)|((A|B)&1) -> (A&1) | B
1875 if (match(B, m_Or(m_Specific(A), m_Value(V1))) ||
1876 match(B, m_Or(m_Value(V1), m_Specific(A)))) {
1877 Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D);
1878 if (Ret) return Ret;
1882 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
1883 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
1884 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
1885 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
1886 SI0->getOperand(1) == SI1->getOperand(1) &&
1887 (SI0->hasOneUse() || SI1->hasOneUse())) {
1888 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
1890 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
1891 SI1->getOperand(1));
1895 // (~A | ~B) == (~(A & B)) - De Morgan's Law
1896 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1897 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1898 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1899 Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
1900 I.getName()+".demorgan");
1901 return BinaryOperator::CreateNot(And);
1904 // Canonicalize xor to the RHS.
1905 if (match(Op0, m_Xor(m_Value(), m_Value())))
1906 std::swap(Op0, Op1);
1908 // A | ( A ^ B) -> A | B
1909 // A | (~A ^ B) -> A | ~B
1910 if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
1911 if (Op0 == A || Op0 == B)
1912 return BinaryOperator::CreateOr(A, B);
1914 if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
1915 Value *Not = Builder->CreateNot(B, B->getName()+".not");
1916 return BinaryOperator::CreateOr(Not, Op0);
1918 if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
1919 Value *Not = Builder->CreateNot(A, A->getName()+".not");
1920 return BinaryOperator::CreateOr(Not, Op0);
1924 // A | ~(A | B) -> A | ~B
1925 // A | ~(A ^ B) -> A | ~B
1926 if (match(Op1, m_Not(m_Value(A))))
1927 if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
1928 if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
1929 Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
1930 B->getOpcode() == Instruction::Xor)) {
1931 Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
1933 Value *Not = Builder->CreateNot(NotOp, NotOp->getName()+".not");
1934 return BinaryOperator::CreateOr(Not, Op0);
1937 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
1938 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
1939 if (Value *Res = FoldOrOfICmps(LHS, RHS))
1940 return ReplaceInstUsesWith(I, Res);
1942 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
1943 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1944 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1945 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
1946 return ReplaceInstUsesWith(I, Res);
1948 // fold (or (cast A), (cast B)) -> (cast (or A, B))
1949 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
1950 CastInst *Op1C = dyn_cast<CastInst>(Op1);
1951 if (Op1C && Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
1952 const Type *SrcTy = Op0C->getOperand(0)->getType();
1953 if (SrcTy == Op1C->getOperand(0)->getType() &&
1954 SrcTy->isIntOrIntVectorTy()) {
1955 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
1957 if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) &&
1958 // Only do this if the casts both really cause code to be
1960 ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
1961 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
1962 Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName());
1963 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
1966 // If this is or(cast(icmp), cast(icmp)), try to fold this even if the
1967 // cast is otherwise not optimizable. This happens for vector sexts.
1968 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
1969 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
1970 if (Value *Res = FoldOrOfICmps(LHS, RHS))
1971 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1973 // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the
1974 // cast is otherwise not optimizable. This happens for vector sexts.
1975 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
1976 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
1977 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
1978 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1983 // Note: If we've gotten to the point of visiting the outer OR, then the
1984 // inner one couldn't be simplified. If it was a constant, then it won't
1985 // be simplified by a later pass either, so we try swapping the inner/outer
1986 // ORs in the hopes that we'll be able to simplify it this way.
1987 // (X|C) | V --> (X|V) | C
1988 if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
1989 match(Op0, m_Or(m_Value(A), m_ConstantInt(C1)))) {
1990 Value *Inner = Builder->CreateOr(A, Op1);
1991 Inner->takeName(Op0);
1992 return BinaryOperator::CreateOr(Inner, C1);
1995 return Changed ? &I : 0;
1998 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
1999 bool Changed = SimplifyAssociativeOrCommutative(I);
2000 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2002 if (Value *V = SimplifyXorInst(Op0, Op1, TD))
2003 return ReplaceInstUsesWith(I, V);
2005 // (A&B)^(A&C) -> A&(B^C) etc
2006 if (Value *V = SimplifyUsingDistributiveLaws(I))
2007 return ReplaceInstUsesWith(I, V);
2009 // See if we can simplify any instructions used by the instruction whose sole
2010 // purpose is to compute bits we don't care about.
2011 if (SimplifyDemandedInstructionBits(I))
2014 // Is this a ~ operation?
2015 if (Value *NotOp = dyn_castNotVal(&I)) {
2016 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
2017 if (Op0I->getOpcode() == Instruction::And ||
2018 Op0I->getOpcode() == Instruction::Or) {
2019 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
2020 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
2021 if (dyn_castNotVal(Op0I->getOperand(1)))
2022 Op0I->swapOperands();
2023 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
2025 Builder->CreateNot(Op0I->getOperand(1),
2026 Op0I->getOperand(1)->getName()+".not");
2027 if (Op0I->getOpcode() == Instruction::And)
2028 return BinaryOperator::CreateOr(Op0NotVal, NotY);
2029 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
2032 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
2033 // ~(X | Y) === (~X & ~Y) - De Morgan's Law
2034 if (isFreeToInvert(Op0I->getOperand(0)) &&
2035 isFreeToInvert(Op0I->getOperand(1))) {
2037 Builder->CreateNot(Op0I->getOperand(0), "notlhs");
2039 Builder->CreateNot(Op0I->getOperand(1), "notrhs");
2040 if (Op0I->getOpcode() == Instruction::And)
2041 return BinaryOperator::CreateOr(NotX, NotY);
2042 return BinaryOperator::CreateAnd(NotX, NotY);
2045 } else if (Op0I->getOpcode() == Instruction::AShr) {
2046 // ~(~X >>s Y) --> (X >>s Y)
2047 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0)))
2048 return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1));
2054 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2055 if (RHS->isOne() && Op0->hasOneUse())
2056 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
2057 if (CmpInst *CI = dyn_cast<CmpInst>(Op0))
2058 return CmpInst::Create(CI->getOpcode(),
2059 CI->getInversePredicate(),
2060 CI->getOperand(0), CI->getOperand(1));
2062 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
2063 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2064 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
2065 if (CI->hasOneUse() && Op0C->hasOneUse()) {
2066 Instruction::CastOps Opcode = Op0C->getOpcode();
2067 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
2068 (RHS == ConstantExpr::getCast(Opcode,
2069 ConstantInt::getTrue(I.getContext()),
2070 Op0C->getDestTy()))) {
2071 CI->setPredicate(CI->getInversePredicate());
2072 return CastInst::Create(Opcode, CI, Op0C->getType());
2078 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2079 // ~(c-X) == X-c-1 == X+(-c-1)
2080 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
2081 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
2082 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
2083 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
2084 ConstantInt::get(I.getType(), 1));
2085 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
2088 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2089 if (Op0I->getOpcode() == Instruction::Add) {
2090 // ~(X-c) --> (-c-1)-X
2091 if (RHS->isAllOnesValue()) {
2092 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
2093 return BinaryOperator::CreateSub(
2094 ConstantExpr::getSub(NegOp0CI,
2095 ConstantInt::get(I.getType(), 1)),
2096 Op0I->getOperand(0));
2097 } else if (RHS->getValue().isSignBit()) {
2098 // (X + C) ^ signbit -> (X + C + signbit)
2099 Constant *C = ConstantInt::get(I.getContext(),
2100 RHS->getValue() + Op0CI->getValue());
2101 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
2104 } else if (Op0I->getOpcode() == Instruction::Or) {
2105 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
2106 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
2107 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
2108 // Anything in both C1 and C2 is known to be zero, remove it from
2110 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
2111 NewRHS = ConstantExpr::getAnd(NewRHS,
2112 ConstantExpr::getNot(CommonBits));
2114 I.setOperand(0, Op0I->getOperand(0));
2115 I.setOperand(1, NewRHS);
2122 // Try to fold constant and into select arguments.
2123 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2124 if (Instruction *R = FoldOpIntoSelect(I, SI))
2126 if (isa<PHINode>(Op0))
2127 if (Instruction *NV = FoldOpIntoPhi(I))
2131 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
2134 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2135 if (A == Op0) { // B^(B|A) == (A|B)^B
2136 Op1I->swapOperands();
2138 std::swap(Op0, Op1);
2139 } else if (B == Op0) { // B^(A|B) == (A|B)^B
2140 I.swapOperands(); // Simplified below.
2141 std::swap(Op0, Op1);
2143 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
2145 if (A == Op0) { // A^(A&B) -> A^(B&A)
2146 Op1I->swapOperands();
2149 if (B == Op0) { // A^(B&A) -> (B&A)^A
2150 I.swapOperands(); // Simplified below.
2151 std::swap(Op0, Op1);
2156 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
2159 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2160 Op0I->hasOneUse()) {
2161 if (A == Op1) // (B|A)^B == (A|B)^B
2163 if (B == Op1) // (A|B)^B == A & ~B
2164 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1, "tmp"));
2165 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2167 if (A == Op1) // (A&B)^A -> (B&A)^A
2169 if (B == Op1 && // (B&A)^A == ~B & A
2170 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
2171 return BinaryOperator::CreateAnd(Builder->CreateNot(A, "tmp"), Op1);
2176 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
2177 if (Op0I && Op1I && Op0I->isShift() &&
2178 Op0I->getOpcode() == Op1I->getOpcode() &&
2179 Op0I->getOperand(1) == Op1I->getOperand(1) &&
2180 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
2182 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
2184 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
2185 Op1I->getOperand(1));
2189 Value *A, *B, *C, *D;
2190 // (A & B)^(A | B) -> A ^ B
2191 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2192 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
2193 if ((A == C && B == D) || (A == D && B == C))
2194 return BinaryOperator::CreateXor(A, B);
2196 // (A | B)^(A & B) -> A ^ B
2197 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2198 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
2199 if ((A == C && B == D) || (A == D && B == C))
2200 return BinaryOperator::CreateXor(A, B);
2204 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
2205 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
2206 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
2207 if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
2208 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2209 LHS->getOperand(1) == RHS->getOperand(0))
2210 LHS->swapOperands();
2211 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2212 LHS->getOperand(1) == RHS->getOperand(1)) {
2213 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2214 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
2215 bool isSigned = LHS->isSigned() || RHS->isSigned();
2216 return ReplaceInstUsesWith(I,
2217 getICmpValue(isSigned, Code, Op0, Op1, Builder));
2221 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
2222 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2223 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
2224 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
2225 const Type *SrcTy = Op0C->getOperand(0)->getType();
2226 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() &&
2227 // Only do this if the casts both really cause code to be generated.
2228 ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0),
2230 ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0),
2232 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
2233 Op1C->getOperand(0), I.getName());
2234 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2239 return Changed ? &I : 0;