1 //===- InstCombineAndOrXor.cpp --------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visitAnd, visitOr, and visitXor functions.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombine.h"
15 #include "llvm/Analysis/InstructionSimplify.h"
16 #include "llvm/IR/Intrinsics.h"
17 #include "llvm/Support/ConstantRange.h"
18 #include "llvm/Support/PatternMatch.h"
19 #include "llvm/Transforms/Utils/CmpInstAnalysis.h"
21 using namespace PatternMatch;
24 /// AddOne - Add one to a ConstantInt.
25 static Constant *AddOne(ConstantInt *C) {
26 return ConstantInt::get(C->getContext(), C->getValue() + 1);
28 /// SubOne - Subtract one from a ConstantInt.
29 static Constant *SubOne(ConstantInt *C) {
30 return ConstantInt::get(C->getContext(), C->getValue()-1);
33 /// isFreeToInvert - Return true if the specified value is free to invert (apply
34 /// ~ to). This happens in cases where the ~ can be eliminated.
35 static inline bool isFreeToInvert(Value *V) {
37 if (BinaryOperator::isNot(V))
40 // Constants can be considered to be not'ed values.
41 if (isa<ConstantInt>(V))
44 // Compares can be inverted if they have a single use.
45 if (CmpInst *CI = dyn_cast<CmpInst>(V))
46 return CI->hasOneUse();
51 static inline Value *dyn_castNotVal(Value *V) {
52 // If this is not(not(x)) don't return that this is a not: we want the two
53 // not's to be folded first.
54 if (BinaryOperator::isNot(V)) {
55 Value *Operand = BinaryOperator::getNotArgument(V);
56 if (!isFreeToInvert(Operand))
60 // Constants can be considered to be not'ed values...
61 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
62 return ConstantInt::get(C->getType(), ~C->getValue());
66 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
67 /// predicate into a three bit mask. It also returns whether it is an ordered
68 /// predicate by reference.
69 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
72 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
73 case FCmpInst::FCMP_UNO: return 0; // 000
74 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
75 case FCmpInst::FCMP_UGT: return 1; // 001
76 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
77 case FCmpInst::FCMP_UEQ: return 2; // 010
78 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
79 case FCmpInst::FCMP_UGE: return 3; // 011
80 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
81 case FCmpInst::FCMP_ULT: return 4; // 100
82 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
83 case FCmpInst::FCMP_UNE: return 5; // 101
84 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
85 case FCmpInst::FCMP_ULE: return 6; // 110
88 // Not expecting FCMP_FALSE and FCMP_TRUE;
89 llvm_unreachable("Unexpected FCmp predicate!");
93 /// getNewICmpValue - This is the complement of getICmpCode, which turns an
94 /// opcode and two operands into either a constant true or false, or a brand
95 /// new ICmp instruction. The sign is passed in to determine which kind
96 /// of predicate to use in the new icmp instruction.
97 static Value *getNewICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS,
98 InstCombiner::BuilderTy *Builder) {
99 ICmpInst::Predicate NewPred;
100 if (Value *NewConstant = getICmpValue(Sign, Code, LHS, RHS, NewPred))
102 return Builder->CreateICmp(NewPred, LHS, RHS);
105 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
106 /// opcode and two operands into either a FCmp instruction. isordered is passed
107 /// in to determine which kind of predicate to use in the new fcmp instruction.
108 static Value *getFCmpValue(bool isordered, unsigned code,
109 Value *LHS, Value *RHS,
110 InstCombiner::BuilderTy *Builder) {
111 CmpInst::Predicate Pred;
113 default: llvm_unreachable("Illegal FCmp code!");
114 case 0: Pred = isordered ? FCmpInst::FCMP_ORD : FCmpInst::FCMP_UNO; break;
115 case 1: Pred = isordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; break;
116 case 2: Pred = isordered ? FCmpInst::FCMP_OEQ : FCmpInst::FCMP_UEQ; break;
117 case 3: Pred = isordered ? FCmpInst::FCMP_OGE : FCmpInst::FCMP_UGE; break;
118 case 4: Pred = isordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; break;
119 case 5: Pred = isordered ? FCmpInst::FCMP_ONE : FCmpInst::FCMP_UNE; break;
120 case 6: Pred = isordered ? FCmpInst::FCMP_OLE : FCmpInst::FCMP_ULE; break;
122 if (!isordered) return ConstantInt::getTrue(LHS->getContext());
123 Pred = FCmpInst::FCMP_ORD; break;
125 return Builder->CreateFCmp(Pred, LHS, RHS);
128 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
129 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
130 // guaranteed to be a binary operator.
131 Instruction *InstCombiner::OptAndOp(Instruction *Op,
134 BinaryOperator &TheAnd) {
135 Value *X = Op->getOperand(0);
136 Constant *Together = 0;
138 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
140 switch (Op->getOpcode()) {
141 case Instruction::Xor:
142 if (Op->hasOneUse()) {
143 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
144 Value *And = Builder->CreateAnd(X, AndRHS);
146 return BinaryOperator::CreateXor(And, Together);
149 case Instruction::Or:
150 if (Op->hasOneUse()){
151 if (Together != OpRHS) {
152 // (X | C1) & C2 --> (X | (C1&C2)) & C2
153 Value *Or = Builder->CreateOr(X, Together);
155 return BinaryOperator::CreateAnd(Or, AndRHS);
158 ConstantInt *TogetherCI = dyn_cast<ConstantInt>(Together);
159 if (TogetherCI && !TogetherCI->isZero()){
160 // (X | C1) & C2 --> (X & (C2^(C1&C2))) | C1
161 // NOTE: This reduces the number of bits set in the & mask, which
162 // can expose opportunities for store narrowing.
163 Together = ConstantExpr::getXor(AndRHS, Together);
164 Value *And = Builder->CreateAnd(X, Together);
166 return BinaryOperator::CreateOr(And, OpRHS);
171 case Instruction::Add:
172 if (Op->hasOneUse()) {
173 // Adding a one to a single bit bit-field should be turned into an XOR
174 // of the bit. First thing to check is to see if this AND is with a
175 // single bit constant.
176 const APInt &AndRHSV = AndRHS->getValue();
178 // If there is only one bit set.
179 if (AndRHSV.isPowerOf2()) {
180 // Ok, at this point, we know that we are masking the result of the
181 // ADD down to exactly one bit. If the constant we are adding has
182 // no bits set below this bit, then we can eliminate the ADD.
183 const APInt& AddRHS = OpRHS->getValue();
185 // Check to see if any bits below the one bit set in AndRHSV are set.
186 if ((AddRHS & (AndRHSV-1)) == 0) {
187 // If not, the only thing that can effect the output of the AND is
188 // the bit specified by AndRHSV. If that bit is set, the effect of
189 // the XOR is to toggle the bit. If it is clear, then the ADD has
191 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
192 TheAnd.setOperand(0, X);
195 // Pull the XOR out of the AND.
196 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
197 NewAnd->takeName(Op);
198 return BinaryOperator::CreateXor(NewAnd, AndRHS);
205 case Instruction::Shl: {
206 // We know that the AND will not produce any of the bits shifted in, so if
207 // the anded constant includes them, clear them now!
209 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
210 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
211 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
212 ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShlMask);
214 if (CI->getValue() == ShlMask)
215 // Masking out bits that the shift already masks.
216 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
218 if (CI != AndRHS) { // Reducing bits set in and.
219 TheAnd.setOperand(1, CI);
224 case Instruction::LShr: {
225 // We know that the AND will not produce any of the bits shifted in, so if
226 // the anded constant includes them, clear them now! This only applies to
227 // unsigned shifts, because a signed shr may bring in set bits!
229 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
230 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
231 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
232 ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShrMask);
234 if (CI->getValue() == ShrMask)
235 // Masking out bits that the shift already masks.
236 return ReplaceInstUsesWith(TheAnd, Op);
239 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
244 case Instruction::AShr:
246 // See if this is shifting in some sign extension, then masking it out
248 if (Op->hasOneUse()) {
249 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
250 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
251 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
252 Constant *C = Builder->getInt(AndRHS->getValue() & ShrMask);
253 if (C == AndRHS) { // Masking out bits shifted in.
254 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
255 // Make the argument unsigned.
256 Value *ShVal = Op->getOperand(0);
257 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
258 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
266 /// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise
267 /// (V < Lo || V >= Hi). In practice, we emit the more efficient
268 /// (V-Lo) \<u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
269 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
270 /// insert new instructions.
271 Value *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
272 bool isSigned, bool Inside) {
273 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
274 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
275 "Lo is not <= Hi in range emission code!");
278 if (Lo == Hi) // Trivially false.
279 return Builder->getFalse();
281 // V >= Min && V < Hi --> V < Hi
282 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
283 ICmpInst::Predicate pred = (isSigned ?
284 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
285 return Builder->CreateICmp(pred, V, Hi);
288 // Emit V-Lo <u Hi-Lo
289 Constant *NegLo = ConstantExpr::getNeg(Lo);
290 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
291 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
292 return Builder->CreateICmpULT(Add, UpperBound);
295 if (Lo == Hi) // Trivially true.
296 return Builder->getTrue();
298 // V < Min || V >= Hi -> V > Hi-1
299 Hi = SubOne(cast<ConstantInt>(Hi));
300 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
301 ICmpInst::Predicate pred = (isSigned ?
302 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
303 return Builder->CreateICmp(pred, V, Hi);
306 // Emit V-Lo >u Hi-1-Lo
307 // Note that Hi has already had one subtracted from it, above.
308 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
309 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
310 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
311 return Builder->CreateICmpUGT(Add, LowerBound);
314 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
315 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
316 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
317 // not, since all 1s are not contiguous.
318 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
319 const APInt& V = Val->getValue();
320 uint32_t BitWidth = Val->getType()->getBitWidth();
321 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
323 // look for the first zero bit after the run of ones
324 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
325 // look for the first non-zero bit
326 ME = V.getActiveBits();
330 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
331 /// where isSub determines whether the operator is a sub. If we can fold one of
332 /// the following xforms:
334 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
335 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
336 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
338 /// return (A +/- B).
340 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
341 ConstantInt *Mask, bool isSub,
343 Instruction *LHSI = dyn_cast<Instruction>(LHS);
344 if (!LHSI || LHSI->getNumOperands() != 2 ||
345 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
347 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
349 switch (LHSI->getOpcode()) {
351 case Instruction::And:
352 if (ConstantExpr::getAnd(N, Mask) == Mask) {
353 // If the AndRHS is a power of two minus one (0+1+), this is simple.
354 if ((Mask->getValue().countLeadingZeros() +
355 Mask->getValue().countPopulation()) ==
356 Mask->getValue().getBitWidth())
359 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
360 // part, we don't need any explicit masks to take them out of A. If that
361 // is all N is, ignore it.
362 uint32_t MB = 0, ME = 0;
363 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
364 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
365 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
366 if (MaskedValueIsZero(RHS, Mask))
371 case Instruction::Or:
372 case Instruction::Xor:
373 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
374 if ((Mask->getValue().countLeadingZeros() +
375 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
376 && ConstantExpr::getAnd(N, Mask)->isNullValue())
382 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
383 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
386 /// enum for classifying (icmp eq (A & B), C) and (icmp ne (A & B), C)
387 /// One of A and B is considered the mask, the other the value. This is
388 /// described as the "AMask" or "BMask" part of the enum. If the enum
389 /// contains only "Mask", then both A and B can be considered masks.
390 /// If A is the mask, then it was proven, that (A & C) == C. This
391 /// is trivial if C == A, or C == 0. If both A and C are constants, this
392 /// proof is also easy.
393 /// For the following explanations we assume that A is the mask.
394 /// The part "AllOnes" declares, that the comparison is true only
395 /// if (A & B) == A, or all bits of A are set in B.
396 /// Example: (icmp eq (A & 3), 3) -> FoldMskICmp_AMask_AllOnes
397 /// The part "AllZeroes" declares, that the comparison is true only
398 /// if (A & B) == 0, or all bits of A are cleared in B.
399 /// Example: (icmp eq (A & 3), 0) -> FoldMskICmp_Mask_AllZeroes
400 /// The part "Mixed" declares, that (A & B) == C and C might or might not
401 /// contain any number of one bits and zero bits.
402 /// Example: (icmp eq (A & 3), 1) -> FoldMskICmp_AMask_Mixed
403 /// The Part "Not" means, that in above descriptions "==" should be replaced
405 /// Example: (icmp ne (A & 3), 3) -> FoldMskICmp_AMask_NotAllOnes
406 /// If the mask A contains a single bit, then the following is equivalent:
407 /// (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
408 /// (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
409 enum MaskedICmpType {
410 FoldMskICmp_AMask_AllOnes = 1,
411 FoldMskICmp_AMask_NotAllOnes = 2,
412 FoldMskICmp_BMask_AllOnes = 4,
413 FoldMskICmp_BMask_NotAllOnes = 8,
414 FoldMskICmp_Mask_AllZeroes = 16,
415 FoldMskICmp_Mask_NotAllZeroes = 32,
416 FoldMskICmp_AMask_Mixed = 64,
417 FoldMskICmp_AMask_NotMixed = 128,
418 FoldMskICmp_BMask_Mixed = 256,
419 FoldMskICmp_BMask_NotMixed = 512
422 /// return the set of pattern classes (from MaskedICmpType)
423 /// that (icmp SCC (A & B), C) satisfies
424 static unsigned getTypeOfMaskedICmp(Value* A, Value* B, Value* C,
425 ICmpInst::Predicate SCC)
427 ConstantInt *ACst = dyn_cast<ConstantInt>(A);
428 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
429 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
430 bool icmp_eq = (SCC == ICmpInst::ICMP_EQ);
431 bool icmp_abit = (ACst != 0 && !ACst->isZero() &&
432 ACst->getValue().isPowerOf2());
433 bool icmp_bbit = (BCst != 0 && !BCst->isZero() &&
434 BCst->getValue().isPowerOf2());
436 if (CCst != 0 && CCst->isZero()) {
437 // if C is zero, then both A and B qualify as mask
438 result |= (icmp_eq ? (FoldMskICmp_Mask_AllZeroes |
439 FoldMskICmp_Mask_AllZeroes |
440 FoldMskICmp_AMask_Mixed |
441 FoldMskICmp_BMask_Mixed)
442 : (FoldMskICmp_Mask_NotAllZeroes |
443 FoldMskICmp_Mask_NotAllZeroes |
444 FoldMskICmp_AMask_NotMixed |
445 FoldMskICmp_BMask_NotMixed));
447 result |= (icmp_eq ? (FoldMskICmp_AMask_NotAllOnes |
448 FoldMskICmp_AMask_NotMixed)
449 : (FoldMskICmp_AMask_AllOnes |
450 FoldMskICmp_AMask_Mixed));
452 result |= (icmp_eq ? (FoldMskICmp_BMask_NotAllOnes |
453 FoldMskICmp_BMask_NotMixed)
454 : (FoldMskICmp_BMask_AllOnes |
455 FoldMskICmp_BMask_Mixed));
459 result |= (icmp_eq ? (FoldMskICmp_AMask_AllOnes |
460 FoldMskICmp_AMask_Mixed)
461 : (FoldMskICmp_AMask_NotAllOnes |
462 FoldMskICmp_AMask_NotMixed));
464 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
465 FoldMskICmp_AMask_NotMixed)
466 : (FoldMskICmp_Mask_AllZeroes |
467 FoldMskICmp_AMask_Mixed));
468 } else if (ACst != 0 && CCst != 0 &&
469 ConstantExpr::getAnd(ACst, CCst) == CCst) {
470 result |= (icmp_eq ? FoldMskICmp_AMask_Mixed
471 : FoldMskICmp_AMask_NotMixed);
474 result |= (icmp_eq ? (FoldMskICmp_BMask_AllOnes |
475 FoldMskICmp_BMask_Mixed)
476 : (FoldMskICmp_BMask_NotAllOnes |
477 FoldMskICmp_BMask_NotMixed));
479 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
480 FoldMskICmp_BMask_NotMixed)
481 : (FoldMskICmp_Mask_AllZeroes |
482 FoldMskICmp_BMask_Mixed));
483 } else if (BCst != 0 && CCst != 0 &&
484 ConstantExpr::getAnd(BCst, CCst) == CCst) {
485 result |= (icmp_eq ? FoldMskICmp_BMask_Mixed
486 : FoldMskICmp_BMask_NotMixed);
491 /// decomposeBitTestICmp - Decompose an icmp into the form ((X & Y) pred Z)
492 /// if possible. The returned predicate is either == or !=. Returns false if
493 /// decomposition fails.
494 static bool decomposeBitTestICmp(const ICmpInst *I, ICmpInst::Predicate &Pred,
495 Value *&X, Value *&Y, Value *&Z) {
496 // X < 0 is equivalent to (X & SignBit) != 0.
497 if (I->getPredicate() == ICmpInst::ICMP_SLT)
498 if (ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1)))
500 X = I->getOperand(0);
501 Y = ConstantInt::get(I->getContext(),
502 APInt::getSignBit(C->getBitWidth()));
503 Pred = ICmpInst::ICMP_NE;
508 // X > -1 is equivalent to (X & SignBit) == 0.
509 if (I->getPredicate() == ICmpInst::ICMP_SGT)
510 if (ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1)))
511 if (C->isAllOnesValue()) {
512 X = I->getOperand(0);
513 Y = ConstantInt::get(I->getContext(),
514 APInt::getSignBit(C->getBitWidth()));
515 Pred = ICmpInst::ICMP_EQ;
516 Z = ConstantInt::getNullValue(C->getType());
523 /// foldLogOpOfMaskedICmpsHelper:
524 /// handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
525 /// return the set of pattern classes (from MaskedICmpType)
526 /// that both LHS and RHS satisfy
527 static unsigned foldLogOpOfMaskedICmpsHelper(Value*& A,
528 Value*& B, Value*& C,
529 Value*& D, Value*& E,
530 ICmpInst *LHS, ICmpInst *RHS,
531 ICmpInst::Predicate &LHSCC,
532 ICmpInst::Predicate &RHSCC) {
533 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) return 0;
534 // vectors are not (yet?) supported
535 if (LHS->getOperand(0)->getType()->isVectorTy()) return 0;
537 // Here comes the tricky part:
538 // LHS might be of the form L11 & L12 == X, X == L21 & L22,
539 // and L11 & L12 == L21 & L22. The same goes for RHS.
540 // Now we must find those components L** and R**, that are equal, so
541 // that we can extract the parameters A, B, C, D, and E for the canonical
543 Value *L1 = LHS->getOperand(0);
544 Value *L2 = LHS->getOperand(1);
545 Value *L11,*L12,*L21,*L22;
546 // Check whether the icmp can be decomposed into a bit test.
547 if (decomposeBitTestICmp(LHS, LHSCC, L11, L12, L2)) {
550 // Look for ANDs in the LHS icmp.
551 if (match(L1, m_And(m_Value(L11), m_Value(L12)))) {
552 if (!match(L2, m_And(m_Value(L21), m_Value(L22))))
555 if (!match(L2, m_And(m_Value(L11), m_Value(L12))))
562 // Bail if LHS was a icmp that can't be decomposed into an equality.
563 if (!ICmpInst::isEquality(LHSCC))
566 Value *R1 = RHS->getOperand(0);
567 Value *R2 = RHS->getOperand(1);
570 if (decomposeBitTestICmp(RHS, RHSCC, R11, R12, R2)) {
571 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
573 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
578 E = R2; R1 = 0; ok = true;
579 } else if (match(R1, m_And(m_Value(R11), m_Value(R12)))) {
580 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
581 A = R11; D = R12; E = R2; ok = true;
582 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
583 A = R12; D = R11; E = R2; ok = true;
587 // Bail if RHS was a icmp that can't be decomposed into an equality.
588 if (!ICmpInst::isEquality(RHSCC))
591 // Look for ANDs in on the right side of the RHS icmp.
592 if (!ok && match(R2, m_And(m_Value(R11), m_Value(R12)))) {
593 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
594 A = R11; D = R12; E = R1; ok = true;
595 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
596 A = R12; D = R11; E = R1; ok = true;
606 } else if (L12 == A) {
608 } else if (L21 == A) {
610 } else if (L22 == A) {
614 unsigned left_type = getTypeOfMaskedICmp(A, B, C, LHSCC);
615 unsigned right_type = getTypeOfMaskedICmp(A, D, E, RHSCC);
616 return left_type & right_type;
618 /// foldLogOpOfMaskedICmps:
619 /// try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
620 /// into a single (icmp(A & X) ==/!= Y)
621 static Value* foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS,
622 ICmpInst::Predicate NEWCC,
623 llvm::InstCombiner::BuilderTy* Builder) {
624 Value *A = 0, *B = 0, *C = 0, *D = 0, *E = 0;
625 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
626 unsigned mask = foldLogOpOfMaskedICmpsHelper(A, B, C, D, E, LHS, RHS,
628 if (mask == 0) return 0;
629 assert(ICmpInst::isEquality(LHSCC) && ICmpInst::isEquality(RHSCC) &&
630 "foldLogOpOfMaskedICmpsHelper must return an equality predicate.");
632 if (NEWCC == ICmpInst::ICMP_NE)
633 mask >>= 1; // treat "Not"-states as normal states
635 if (mask & FoldMskICmp_Mask_AllZeroes) {
636 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
637 // -> (icmp eq (A & (B|D)), 0)
638 Value* newOr = Builder->CreateOr(B, D);
639 Value* newAnd = Builder->CreateAnd(A, newOr);
640 // we can't use C as zero, because we might actually handle
641 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
642 // with B and D, having a single bit set
643 Value* zero = Constant::getNullValue(A->getType());
644 return Builder->CreateICmp(NEWCC, newAnd, zero);
646 if (mask & FoldMskICmp_BMask_AllOnes) {
647 // (icmp eq (A & B), B) & (icmp eq (A & D), D)
648 // -> (icmp eq (A & (B|D)), (B|D))
649 Value* newOr = Builder->CreateOr(B, D);
650 Value* newAnd = Builder->CreateAnd(A, newOr);
651 return Builder->CreateICmp(NEWCC, newAnd, newOr);
653 if (mask & FoldMskICmp_AMask_AllOnes) {
654 // (icmp eq (A & B), A) & (icmp eq (A & D), A)
655 // -> (icmp eq (A & (B&D)), A)
656 Value* newAnd1 = Builder->CreateAnd(B, D);
657 Value* newAnd = Builder->CreateAnd(A, newAnd1);
658 return Builder->CreateICmp(NEWCC, newAnd, A);
660 if (mask & FoldMskICmp_BMask_Mixed) {
661 // (icmp eq (A & B), C) & (icmp eq (A & D), E)
662 // We already know that B & C == C && D & E == E.
663 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
664 // C and E, which are shared by both the mask B and the mask D, don't
665 // contradict, then we can transform to
666 // -> (icmp eq (A & (B|D)), (C|E))
667 // Currently, we only handle the case of B, C, D, and E being constant.
668 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
669 if (BCst == 0) return 0;
670 ConstantInt *DCst = dyn_cast<ConstantInt>(D);
671 if (DCst == 0) return 0;
672 // we can't simply use C and E, because we might actually handle
673 // (icmp ne (A & B), B) & (icmp eq (A & D), D)
674 // with B and D, having a single bit set
676 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
677 if (CCst == 0) return 0;
679 CCst = dyn_cast<ConstantInt>( ConstantExpr::getXor(BCst, CCst) );
680 ConstantInt *ECst = dyn_cast<ConstantInt>(E);
681 if (ECst == 0) return 0;
683 ECst = dyn_cast<ConstantInt>( ConstantExpr::getXor(DCst, ECst) );
684 ConstantInt* MCst = dyn_cast<ConstantInt>(
685 ConstantExpr::getAnd(ConstantExpr::getAnd(BCst, DCst),
686 ConstantExpr::getXor(CCst, ECst)) );
687 // if there is a conflict we should actually return a false for the
691 Value *newOr1 = Builder->CreateOr(B, D);
692 Value *newOr2 = ConstantExpr::getOr(CCst, ECst);
693 Value *newAnd = Builder->CreateAnd(A, newOr1);
694 return Builder->CreateICmp(NEWCC, newAnd, newOr2);
699 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
700 Value *InstCombiner::FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
701 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
703 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
704 if (PredicatesFoldable(LHSCC, RHSCC)) {
705 if (LHS->getOperand(0) == RHS->getOperand(1) &&
706 LHS->getOperand(1) == RHS->getOperand(0))
708 if (LHS->getOperand(0) == RHS->getOperand(0) &&
709 LHS->getOperand(1) == RHS->getOperand(1)) {
710 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
711 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
712 bool isSigned = LHS->isSigned() || RHS->isSigned();
713 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
717 // handle (roughly): (icmp eq (A & B), C) & (icmp eq (A & D), E)
718 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, ICmpInst::ICMP_EQ, Builder))
721 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
722 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
723 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
724 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
725 if (LHSCst == 0 || RHSCst == 0) return 0;
727 if (LHSCst == RHSCst && LHSCC == RHSCC) {
728 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
729 // where C is a power of 2
730 if (LHSCC == ICmpInst::ICMP_ULT &&
731 LHSCst->getValue().isPowerOf2()) {
732 Value *NewOr = Builder->CreateOr(Val, Val2);
733 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
736 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
737 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
738 Value *NewOr = Builder->CreateOr(Val, Val2);
739 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
743 // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2
744 // where CMAX is the all ones value for the truncated type,
745 // iff the lower bits of C2 and CA are zero.
746 if (LHSCC == ICmpInst::ICMP_EQ && LHSCC == RHSCC &&
747 LHS->hasOneUse() && RHS->hasOneUse()) {
749 ConstantInt *AndCst, *SmallCst = 0, *BigCst = 0;
751 // (trunc x) == C1 & (and x, CA) == C2
752 // (and x, CA) == C2 & (trunc x) == C1
753 if (match(Val2, m_Trunc(m_Value(V))) &&
754 match(Val, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
757 } else if (match(Val, m_Trunc(m_Value(V))) &&
758 match(Val2, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
763 if (SmallCst && BigCst) {
764 unsigned BigBitSize = BigCst->getType()->getBitWidth();
765 unsigned SmallBitSize = SmallCst->getType()->getBitWidth();
767 // Check that the low bits are zero.
768 APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize);
769 if ((Low & AndCst->getValue()) == 0 && (Low & BigCst->getValue()) == 0) {
770 Value *NewAnd = Builder->CreateAnd(V, Low | AndCst->getValue());
771 APInt N = SmallCst->getValue().zext(BigBitSize) | BigCst->getValue();
772 Value *NewVal = ConstantInt::get(AndCst->getType()->getContext(), N);
773 return Builder->CreateICmp(LHSCC, NewAnd, NewVal);
778 // From here on, we only handle:
779 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
780 if (Val != Val2) return 0;
782 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
783 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
784 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
785 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
786 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
789 // Make a constant range that's the intersection of the two icmp ranges.
790 // If the intersection is empty, we know that the result is false.
791 ConstantRange LHSRange =
792 ConstantRange::makeICmpRegion(LHSCC, LHSCst->getValue());
793 ConstantRange RHSRange =
794 ConstantRange::makeICmpRegion(RHSCC, RHSCst->getValue());
796 if (LHSRange.intersectWith(RHSRange).isEmptySet())
797 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
799 // We can't fold (ugt x, C) & (sgt x, C2).
800 if (!PredicatesFoldable(LHSCC, RHSCC))
803 // Ensure that the larger constant is on the RHS.
805 if (CmpInst::isSigned(LHSCC) ||
806 (ICmpInst::isEquality(LHSCC) &&
807 CmpInst::isSigned(RHSCC)))
808 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
810 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
814 std::swap(LHSCst, RHSCst);
815 std::swap(LHSCC, RHSCC);
818 // At this point, we know we have two icmp instructions
819 // comparing a value against two constants and and'ing the result
820 // together. Because of the above check, we know that we only have
821 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
822 // (from the icmp folding check above), that the two constants
823 // are not equal and that the larger constant is on the RHS
824 assert(LHSCst != RHSCst && "Compares not folded above?");
827 default: llvm_unreachable("Unknown integer condition code!");
828 case ICmpInst::ICMP_EQ:
830 default: llvm_unreachable("Unknown integer condition code!");
831 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
832 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
833 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
836 case ICmpInst::ICMP_NE:
838 default: llvm_unreachable("Unknown integer condition code!");
839 case ICmpInst::ICMP_ULT:
840 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
841 return Builder->CreateICmpULT(Val, LHSCst);
842 break; // (X != 13 & X u< 15) -> no change
843 case ICmpInst::ICMP_SLT:
844 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
845 return Builder->CreateICmpSLT(Val, LHSCst);
846 break; // (X != 13 & X s< 15) -> no change
847 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
848 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
849 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
851 case ICmpInst::ICMP_NE:
852 // Special case to get the ordering right when the values wrap around
854 if (LHSCst->getValue() == 0 && RHSCst->getValue().isAllOnesValue())
855 std::swap(LHSCst, RHSCst);
856 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
857 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
858 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
859 return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1),
860 Val->getName()+".cmp");
862 break; // (X != 13 & X != 15) -> no change
865 case ICmpInst::ICMP_ULT:
867 default: llvm_unreachable("Unknown integer condition code!");
868 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
869 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
870 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
871 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
873 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
874 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
876 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
880 case ICmpInst::ICMP_SLT:
882 default: llvm_unreachable("Unknown integer condition code!");
883 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
885 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
886 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
888 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
892 case ICmpInst::ICMP_UGT:
894 default: llvm_unreachable("Unknown integer condition code!");
895 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
896 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
898 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
900 case ICmpInst::ICMP_NE:
901 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
902 return Builder->CreateICmp(LHSCC, Val, RHSCst);
903 break; // (X u> 13 & X != 15) -> no change
904 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
905 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
906 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
910 case ICmpInst::ICMP_SGT:
912 default: llvm_unreachable("Unknown integer condition code!");
913 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
914 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
916 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
918 case ICmpInst::ICMP_NE:
919 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
920 return Builder->CreateICmp(LHSCC, Val, RHSCst);
921 break; // (X s> 13 & X != 15) -> no change
922 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
923 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true);
924 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
933 /// FoldAndOfFCmps - Optimize (fcmp)&(fcmp). NOTE: Unlike the rest of
934 /// instcombine, this returns a Value which should already be inserted into the
936 Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
937 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
938 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
939 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType())
942 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
943 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
944 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
945 // If either of the constants are nans, then the whole thing returns
947 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
948 return Builder->getFalse();
949 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
952 // Handle vector zeros. This occurs because the canonical form of
953 // "fcmp ord x,x" is "fcmp ord x, 0".
954 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
955 isa<ConstantAggregateZero>(RHS->getOperand(1)))
956 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
960 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
961 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
962 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
965 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
966 // Swap RHS operands to match LHS.
967 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
968 std::swap(Op1LHS, Op1RHS);
971 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
972 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
974 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
975 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
976 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
977 if (Op0CC == FCmpInst::FCMP_TRUE)
979 if (Op1CC == FCmpInst::FCMP_TRUE)
984 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
985 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
986 // uno && ord -> false
987 if (Op0Pred == 0 && Op1Pred == 0 && Op0Ordered != Op1Ordered)
988 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
991 std::swap(Op0Pred, Op1Pred);
992 std::swap(Op0Ordered, Op1Ordered);
995 // uno && ueq -> uno && (uno || eq) -> uno
996 // ord && olt -> ord && (ord && lt) -> olt
997 if (!Op0Ordered && (Op0Ordered == Op1Ordered))
999 if (Op0Ordered && (Op0Ordered == Op1Ordered))
1002 // uno && oeq -> uno && (ord && eq) -> false
1004 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
1005 // ord && ueq -> ord && (uno || eq) -> oeq
1006 return getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS, Builder);
1014 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
1015 bool Changed = SimplifyAssociativeOrCommutative(I);
1016 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1018 if (Value *V = SimplifyAndInst(Op0, Op1, TD))
1019 return ReplaceInstUsesWith(I, V);
1021 // (A|B)&(A|C) -> A|(B&C) etc
1022 if (Value *V = SimplifyUsingDistributiveLaws(I))
1023 return ReplaceInstUsesWith(I, V);
1025 // See if we can simplify any instructions used by the instruction whose sole
1026 // purpose is to compute bits we don't care about.
1027 if (SimplifyDemandedInstructionBits(I))
1030 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
1031 const APInt &AndRHSMask = AndRHS->getValue();
1033 // Optimize a variety of ((val OP C1) & C2) combinations...
1034 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1035 Value *Op0LHS = Op0I->getOperand(0);
1036 Value *Op0RHS = Op0I->getOperand(1);
1037 switch (Op0I->getOpcode()) {
1039 case Instruction::Xor:
1040 case Instruction::Or: {
1041 // If the mask is only needed on one incoming arm, push it up.
1042 if (!Op0I->hasOneUse()) break;
1044 APInt NotAndRHS(~AndRHSMask);
1045 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
1046 // Not masking anything out for the LHS, move to RHS.
1047 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
1048 Op0RHS->getName()+".masked");
1049 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
1051 if (!isa<Constant>(Op0RHS) &&
1052 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
1053 // Not masking anything out for the RHS, move to LHS.
1054 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
1055 Op0LHS->getName()+".masked");
1056 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
1061 case Instruction::Add:
1062 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
1063 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1064 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1065 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
1066 return BinaryOperator::CreateAnd(V, AndRHS);
1067 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
1068 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
1071 case Instruction::Sub:
1072 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
1073 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1074 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1075 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
1076 return BinaryOperator::CreateAnd(V, AndRHS);
1078 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
1079 // has 1's for all bits that the subtraction with A might affect.
1080 if (Op0I->hasOneUse() && !match(Op0LHS, m_Zero())) {
1081 uint32_t BitWidth = AndRHSMask.getBitWidth();
1082 uint32_t Zeros = AndRHSMask.countLeadingZeros();
1083 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
1085 if (MaskedValueIsZero(Op0LHS, Mask)) {
1086 Value *NewNeg = Builder->CreateNeg(Op0RHS);
1087 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
1092 case Instruction::Shl:
1093 case Instruction::LShr:
1094 // (1 << x) & 1 --> zext(x == 0)
1095 // (1 >> x) & 1 --> zext(x == 0)
1096 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
1098 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
1099 return new ZExtInst(NewICmp, I.getType());
1104 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
1105 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
1109 // If this is an integer truncation, and if the source is an 'and' with
1110 // immediate, transform it. This frequently occurs for bitfield accesses.
1112 Value *X = 0; ConstantInt *YC = 0;
1113 if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) {
1114 // Change: and (trunc (and X, YC) to T), C2
1115 // into : and (trunc X to T), trunc(YC) & C2
1116 // This will fold the two constants together, which may allow
1117 // other simplifications.
1118 Value *NewCast = Builder->CreateTrunc(X, I.getType(), "and.shrunk");
1119 Constant *C3 = ConstantExpr::getTrunc(YC, I.getType());
1120 C3 = ConstantExpr::getAnd(C3, AndRHS);
1121 return BinaryOperator::CreateAnd(NewCast, C3);
1125 // Try to fold constant and into select arguments.
1126 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1127 if (Instruction *R = FoldOpIntoSelect(I, SI))
1129 if (isa<PHINode>(Op0))
1130 if (Instruction *NV = FoldOpIntoPhi(I))
1135 // (~A & ~B) == (~(A | B)) - De Morgan's Law
1136 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1137 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1138 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1139 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
1140 I.getName()+".demorgan");
1141 return BinaryOperator::CreateNot(Or);
1145 Value *A = 0, *B = 0, *C = 0, *D = 0;
1146 // (A|B) & ~(A&B) -> A^B
1147 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1148 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1149 ((A == C && B == D) || (A == D && B == C)))
1150 return BinaryOperator::CreateXor(A, B);
1152 // ~(A&B) & (A|B) -> A^B
1153 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1154 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1155 ((A == C && B == D) || (A == D && B == C)))
1156 return BinaryOperator::CreateXor(A, B);
1158 // A&(A^B) => A & ~B
1160 Value *tmpOp0 = Op0;
1161 Value *tmpOp1 = Op1;
1162 if (Op0->hasOneUse() &&
1163 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
1164 if (A == Op1 || B == Op1 ) {
1171 if (tmpOp1->hasOneUse() &&
1172 match(tmpOp1, m_Xor(m_Value(A), m_Value(B)))) {
1176 // Notice that the patten (A&(~B)) is actually (A&(-1^B)), so if
1177 // A is originally -1 (or a vector of -1 and undefs), then we enter
1178 // an endless loop. By checking that A is non-constant we ensure that
1179 // we will never get to the loop.
1180 if (A == tmpOp0 && !isa<Constant>(A)) // A&(A^B) -> A & ~B
1181 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B));
1185 // (A&((~A)|B)) -> A&B
1186 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
1187 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
1188 return BinaryOperator::CreateAnd(A, Op1);
1189 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
1190 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
1191 return BinaryOperator::CreateAnd(A, Op0);
1194 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1))
1195 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
1196 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1197 return ReplaceInstUsesWith(I, Res);
1199 // If and'ing two fcmp, try combine them into one.
1200 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1201 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1202 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1203 return ReplaceInstUsesWith(I, Res);
1206 // fold (and (cast A), (cast B)) -> (cast (and A, B))
1207 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
1208 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) {
1209 Type *SrcTy = Op0C->getOperand(0)->getType();
1210 if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ?
1211 SrcTy == Op1C->getOperand(0)->getType() &&
1212 SrcTy->isIntOrIntVectorTy()) {
1213 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
1215 // Only do this if the casts both really cause code to be generated.
1216 if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
1217 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
1218 Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName());
1219 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
1222 // If this is and(cast(icmp), cast(icmp)), try to fold this even if the
1223 // cast is otherwise not optimizable. This happens for vector sexts.
1224 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
1225 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
1226 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1227 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1229 // If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the
1230 // cast is otherwise not optimizable. This happens for vector sexts.
1231 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
1232 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
1233 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1234 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1238 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
1239 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
1240 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
1241 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
1242 SI0->getOperand(1) == SI1->getOperand(1) &&
1243 (SI0->hasOneUse() || SI1->hasOneUse())) {
1245 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
1247 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
1248 SI1->getOperand(1));
1254 bool OpsSwapped = false;
1255 // Canonicalize SExt or Not to the LHS
1256 if (match(Op1, m_SExt(m_Value())) ||
1257 match(Op1, m_Not(m_Value()))) {
1258 std::swap(Op0, Op1);
1262 // Fold (and (sext bool to A), B) --> (select bool, B, 0)
1263 if (match(Op0, m_SExt(m_Value(X))) &&
1264 X->getType()->getScalarType()->isIntegerTy(1)) {
1265 Value *Zero = Constant::getNullValue(Op1->getType());
1266 return SelectInst::Create(X, Op1, Zero);
1269 // Fold (and ~(sext bool to A), B) --> (select bool, 0, B)
1270 if (match(Op0, m_Not(m_SExt(m_Value(X)))) &&
1271 X->getType()->getScalarType()->isIntegerTy(1)) {
1272 Value *Zero = Constant::getNullValue(Op0->getType());
1273 return SelectInst::Create(X, Zero, Op1);
1277 std::swap(Op0, Op1);
1280 return Changed ? &I : 0;
1283 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
1284 /// capable of providing pieces of a bswap. The subexpression provides pieces
1285 /// of a bswap if it is proven that each of the non-zero bytes in the output of
1286 /// the expression came from the corresponding "byte swapped" byte in some other
1287 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
1288 /// we know that the expression deposits the low byte of %X into the high byte
1289 /// of the bswap result and that all other bytes are zero. This expression is
1290 /// accepted, the high byte of ByteValues is set to X to indicate a correct
1293 /// This function returns true if the match was unsuccessful and false if so.
1294 /// On entry to the function the "OverallLeftShift" is a signed integer value
1295 /// indicating the number of bytes that the subexpression is later shifted. For
1296 /// example, if the expression is later right shifted by 16 bits, the
1297 /// OverallLeftShift value would be -2 on entry. This is used to specify which
1298 /// byte of ByteValues is actually being set.
1300 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
1301 /// byte is masked to zero by a user. For example, in (X & 255), X will be
1302 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
1303 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
1304 /// always in the local (OverallLeftShift) coordinate space.
1306 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
1307 SmallVectorImpl<Value *> &ByteValues) {
1308 if (Instruction *I = dyn_cast<Instruction>(V)) {
1309 // If this is an or instruction, it may be an inner node of the bswap.
1310 if (I->getOpcode() == Instruction::Or) {
1311 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1313 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
1317 // If this is a logical shift by a constant multiple of 8, recurse with
1318 // OverallLeftShift and ByteMask adjusted.
1319 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
1321 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
1322 // Ensure the shift amount is defined and of a byte value.
1323 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
1326 unsigned ByteShift = ShAmt >> 3;
1327 if (I->getOpcode() == Instruction::Shl) {
1328 // X << 2 -> collect(X, +2)
1329 OverallLeftShift += ByteShift;
1330 ByteMask >>= ByteShift;
1332 // X >>u 2 -> collect(X, -2)
1333 OverallLeftShift -= ByteShift;
1334 ByteMask <<= ByteShift;
1335 ByteMask &= (~0U >> (32-ByteValues.size()));
1338 if (OverallLeftShift >= (int)ByteValues.size()) return true;
1339 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
1341 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1345 // If this is a logical 'and' with a mask that clears bytes, clear the
1346 // corresponding bytes in ByteMask.
1347 if (I->getOpcode() == Instruction::And &&
1348 isa<ConstantInt>(I->getOperand(1))) {
1349 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
1350 unsigned NumBytes = ByteValues.size();
1351 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
1352 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
1354 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
1355 // If this byte is masked out by a later operation, we don't care what
1357 if ((ByteMask & (1 << i)) == 0)
1360 // If the AndMask is all zeros for this byte, clear the bit.
1361 APInt MaskB = AndMask & Byte;
1363 ByteMask &= ~(1U << i);
1367 // If the AndMask is not all ones for this byte, it's not a bytezap.
1371 // Otherwise, this byte is kept.
1374 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1379 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
1380 // the input value to the bswap. Some observations: 1) if more than one byte
1381 // is demanded from this input, then it could not be successfully assembled
1382 // into a byteswap. At least one of the two bytes would not be aligned with
1383 // their ultimate destination.
1384 if (!isPowerOf2_32(ByteMask)) return true;
1385 unsigned InputByteNo = countTrailingZeros(ByteMask);
1387 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
1388 // is demanded, it needs to go into byte 0 of the result. This means that the
1389 // byte needs to be shifted until it lands in the right byte bucket. The
1390 // shift amount depends on the position: if the byte is coming from the high
1391 // part of the value (e.g. byte 3) then it must be shifted right. If from the
1392 // low part, it must be shifted left.
1393 unsigned DestByteNo = InputByteNo + OverallLeftShift;
1394 if (ByteValues.size()-1-DestByteNo != InputByteNo)
1397 // If the destination byte value is already defined, the values are or'd
1398 // together, which isn't a bswap (unless it's an or of the same bits).
1399 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
1401 ByteValues[DestByteNo] = V;
1405 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
1406 /// If so, insert the new bswap intrinsic and return it.
1407 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
1408 IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
1409 if (!ITy || ITy->getBitWidth() % 16 ||
1410 // ByteMask only allows up to 32-byte values.
1411 ITy->getBitWidth() > 32*8)
1412 return 0; // Can only bswap pairs of bytes. Can't do vectors.
1414 /// ByteValues - For each byte of the result, we keep track of which value
1415 /// defines each byte.
1416 SmallVector<Value*, 8> ByteValues;
1417 ByteValues.resize(ITy->getBitWidth()/8);
1419 // Try to find all the pieces corresponding to the bswap.
1420 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
1421 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
1424 // Check to see if all of the bytes come from the same value.
1425 Value *V = ByteValues[0];
1426 if (V == 0) return 0; // Didn't find a byte? Must be zero.
1428 // Check to make sure that all of the bytes come from the same value.
1429 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
1430 if (ByteValues[i] != V)
1432 Module *M = I.getParent()->getParent()->getParent();
1433 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, ITy);
1434 return CallInst::Create(F, V);
1437 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
1438 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
1439 /// we can simplify this expression to "cond ? C : D or B".
1440 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
1441 Value *C, Value *D) {
1442 // If A is not a select of -1/0, this cannot match.
1444 if (!match(A, m_SExt(m_Value(Cond))) ||
1445 !Cond->getType()->isIntegerTy(1))
1448 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
1449 if (match(D, m_Not(m_SExt(m_Specific(Cond)))))
1450 return SelectInst::Create(Cond, C, B);
1451 if (match(D, m_SExt(m_Not(m_Specific(Cond)))))
1452 return SelectInst::Create(Cond, C, B);
1454 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
1455 if (match(B, m_Not(m_SExt(m_Specific(Cond)))))
1456 return SelectInst::Create(Cond, C, D);
1457 if (match(B, m_SExt(m_Not(m_Specific(Cond)))))
1458 return SelectInst::Create(Cond, C, D);
1462 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
1463 Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
1464 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
1466 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
1467 if (PredicatesFoldable(LHSCC, RHSCC)) {
1468 if (LHS->getOperand(0) == RHS->getOperand(1) &&
1469 LHS->getOperand(1) == RHS->getOperand(0))
1470 LHS->swapOperands();
1471 if (LHS->getOperand(0) == RHS->getOperand(0) &&
1472 LHS->getOperand(1) == RHS->getOperand(1)) {
1473 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1474 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
1475 bool isSigned = LHS->isSigned() || RHS->isSigned();
1476 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
1480 // handle (roughly):
1481 // (icmp ne (A & B), C) | (icmp ne (A & D), E)
1482 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, ICmpInst::ICMP_NE, Builder))
1485 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
1486 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
1487 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
1489 if (LHS->hasOneUse() || RHS->hasOneUse()) {
1490 // (icmp eq B, 0) | (icmp ult A, B) -> (icmp ule A, B-1)
1491 // (icmp eq B, 0) | (icmp ugt B, A) -> (icmp ule A, B-1)
1492 Value *A = 0, *B = 0;
1493 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero()) {
1495 if (RHSCC == ICmpInst::ICMP_ULT && Val == RHS->getOperand(1))
1497 else if (RHSCC == ICmpInst::ICMP_UGT && Val == Val2)
1498 A = RHS->getOperand(1);
1500 // (icmp ult A, B) | (icmp eq B, 0) -> (icmp ule A, B-1)
1501 // (icmp ugt B, A) | (icmp eq B, 0) -> (icmp ule A, B-1)
1502 else if (RHSCC == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) {
1504 if (LHSCC == ICmpInst::ICMP_ULT && Val2 == LHS->getOperand(1))
1506 else if (LHSCC == ICmpInst::ICMP_UGT && Val2 == Val)
1507 A = LHS->getOperand(1);
1510 return Builder->CreateICmp(
1512 Builder->CreateAdd(B, ConstantInt::getSigned(B->getType(), -1)), A);
1515 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
1516 if (LHSCst == 0 || RHSCst == 0) return 0;
1518 if (LHSCst == RHSCst && LHSCC == RHSCC) {
1519 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
1520 if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
1521 Value *NewOr = Builder->CreateOr(Val, Val2);
1522 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
1526 // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
1527 // iff C2 + CA == C1.
1528 if (LHSCC == ICmpInst::ICMP_ULT && RHSCC == ICmpInst::ICMP_EQ) {
1529 ConstantInt *AddCst;
1530 if (match(Val, m_Add(m_Specific(Val2), m_ConstantInt(AddCst))))
1531 if (RHSCst->getValue() + AddCst->getValue() == LHSCst->getValue())
1532 return Builder->CreateICmpULE(Val, LHSCst);
1535 // From here on, we only handle:
1536 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
1537 if (Val != Val2) return 0;
1539 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
1540 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
1541 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
1542 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
1543 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
1546 // We can't fold (ugt x, C) | (sgt x, C2).
1547 if (!PredicatesFoldable(LHSCC, RHSCC))
1550 // Ensure that the larger constant is on the RHS.
1552 if (CmpInst::isSigned(LHSCC) ||
1553 (ICmpInst::isEquality(LHSCC) &&
1554 CmpInst::isSigned(RHSCC)))
1555 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
1557 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
1560 std::swap(LHS, RHS);
1561 std::swap(LHSCst, RHSCst);
1562 std::swap(LHSCC, RHSCC);
1565 // At this point, we know we have two icmp instructions
1566 // comparing a value against two constants and or'ing the result
1567 // together. Because of the above check, we know that we only have
1568 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
1569 // icmp folding check above), that the two constants are not
1571 assert(LHSCst != RHSCst && "Compares not folded above?");
1574 default: llvm_unreachable("Unknown integer condition code!");
1575 case ICmpInst::ICMP_EQ:
1577 default: llvm_unreachable("Unknown integer condition code!");
1578 case ICmpInst::ICMP_EQ:
1579 if (LHS->getOperand(0) == RHS->getOperand(0)) {
1580 // if LHSCst and RHSCst differ only by one bit:
1581 // (A == C1 || A == C2) -> (A & ~(C1 ^ C2)) == C1
1582 assert(LHSCst->getValue().ule(LHSCst->getValue()));
1584 APInt Xor = LHSCst->getValue() ^ RHSCst->getValue();
1585 if (Xor.isPowerOf2()) {
1586 Value *NegCst = Builder->getInt(~Xor);
1587 Value *And = Builder->CreateAnd(LHS->getOperand(0), NegCst);
1588 return Builder->CreateICmp(ICmpInst::ICMP_EQ, And, LHSCst);
1592 if (LHSCst == SubOne(RHSCst)) {
1593 // (X == 13 | X == 14) -> X-13 <u 2
1594 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1595 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1596 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
1597 return Builder->CreateICmpULT(Add, AddCST);
1600 break; // (X == 13 | X == 15) -> no change
1601 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
1602 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
1604 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
1605 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
1606 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
1610 case ICmpInst::ICMP_NE:
1612 default: llvm_unreachable("Unknown integer condition code!");
1613 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
1614 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
1615 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
1617 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
1618 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
1619 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
1620 return Builder->getTrue();
1622 case ICmpInst::ICMP_ULT:
1624 default: llvm_unreachable("Unknown integer condition code!");
1625 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
1627 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
1628 // If RHSCst is [us]MAXINT, it is always false. Not handling
1629 // this can cause overflow.
1630 if (RHSCst->isMaxValue(false))
1632 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false);
1633 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
1635 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
1636 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
1638 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
1642 case ICmpInst::ICMP_SLT:
1644 default: llvm_unreachable("Unknown integer condition code!");
1645 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
1647 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
1648 // If RHSCst is [us]MAXINT, it is always false. Not handling
1649 // this can cause overflow.
1650 if (RHSCst->isMaxValue(true))
1652 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false);
1653 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
1655 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
1656 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
1658 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
1662 case ICmpInst::ICMP_UGT:
1664 default: llvm_unreachable("Unknown integer condition code!");
1665 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
1666 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
1668 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
1670 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
1671 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
1672 return Builder->getTrue();
1673 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
1677 case ICmpInst::ICMP_SGT:
1679 default: llvm_unreachable("Unknown integer condition code!");
1680 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
1681 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
1683 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
1685 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
1686 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
1687 return Builder->getTrue();
1688 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
1696 /// FoldOrOfFCmps - Optimize (fcmp)|(fcmp). NOTE: Unlike the rest of
1697 /// instcombine, this returns a Value which should already be inserted into the
1699 Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
1700 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
1701 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
1702 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
1703 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
1704 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
1705 // If either of the constants are nans, then the whole thing returns
1707 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
1708 return Builder->getTrue();
1710 // Otherwise, no need to compare the two constants, compare the
1712 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
1715 // Handle vector zeros. This occurs because the canonical form of
1716 // "fcmp uno x,x" is "fcmp uno x, 0".
1717 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
1718 isa<ConstantAggregateZero>(RHS->getOperand(1)))
1719 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
1724 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
1725 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
1726 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
1728 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
1729 // Swap RHS operands to match LHS.
1730 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
1731 std::swap(Op1LHS, Op1RHS);
1733 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
1734 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
1736 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
1737 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
1738 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
1739 if (Op0CC == FCmpInst::FCMP_FALSE)
1741 if (Op1CC == FCmpInst::FCMP_FALSE)
1745 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
1746 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
1747 if (Op0Ordered == Op1Ordered) {
1748 // If both are ordered or unordered, return a new fcmp with
1749 // or'ed predicates.
1750 return getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS, Builder);
1756 /// FoldOrWithConstants - This helper function folds:
1758 /// ((A | B) & C1) | (B & C2)
1764 /// when the XOR of the two constants is "all ones" (-1).
1765 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
1766 Value *A, Value *B, Value *C) {
1767 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
1771 ConstantInt *CI2 = 0;
1772 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
1774 APInt Xor = CI1->getValue() ^ CI2->getValue();
1775 if (!Xor.isAllOnesValue()) return 0;
1777 if (V1 == A || V1 == B) {
1778 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
1779 return BinaryOperator::CreateOr(NewOp, V1);
1785 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
1786 bool Changed = SimplifyAssociativeOrCommutative(I);
1787 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1789 if (Value *V = SimplifyOrInst(Op0, Op1, TD))
1790 return ReplaceInstUsesWith(I, V);
1792 // (A&B)|(A&C) -> A&(B|C) etc
1793 if (Value *V = SimplifyUsingDistributiveLaws(I))
1794 return ReplaceInstUsesWith(I, V);
1796 // See if we can simplify any instructions used by the instruction whose sole
1797 // purpose is to compute bits we don't care about.
1798 if (SimplifyDemandedInstructionBits(I))
1801 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1802 ConstantInt *C1 = 0; Value *X = 0;
1803 // (X & C1) | C2 --> (X | C2) & (C1|C2)
1804 // iff (C1 & C2) == 0.
1805 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
1806 (RHS->getValue() & C1->getValue()) != 0 &&
1808 Value *Or = Builder->CreateOr(X, RHS);
1810 return BinaryOperator::CreateAnd(Or,
1811 Builder->getInt(RHS->getValue() | C1->getValue()));
1814 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
1815 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
1817 Value *Or = Builder->CreateOr(X, RHS);
1819 return BinaryOperator::CreateXor(Or,
1820 Builder->getInt(C1->getValue() & ~RHS->getValue()));
1823 // Try to fold constant and into select arguments.
1824 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1825 if (Instruction *R = FoldOpIntoSelect(I, SI))
1828 if (isa<PHINode>(Op0))
1829 if (Instruction *NV = FoldOpIntoPhi(I))
1833 Value *A = 0, *B = 0;
1834 ConstantInt *C1 = 0, *C2 = 0;
1836 // (A | B) | C and A | (B | C) -> bswap if possible.
1837 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
1838 if (match(Op0, m_Or(m_Value(), m_Value())) ||
1839 match(Op1, m_Or(m_Value(), m_Value())) ||
1840 (match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
1841 match(Op1, m_LogicalShift(m_Value(), m_Value())))) {
1842 if (Instruction *BSwap = MatchBSwap(I))
1846 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
1847 if (Op0->hasOneUse() &&
1848 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
1849 MaskedValueIsZero(Op1, C1->getValue())) {
1850 Value *NOr = Builder->CreateOr(A, Op1);
1852 return BinaryOperator::CreateXor(NOr, C1);
1855 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
1856 if (Op1->hasOneUse() &&
1857 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
1858 MaskedValueIsZero(Op0, C1->getValue())) {
1859 Value *NOr = Builder->CreateOr(A, Op0);
1861 return BinaryOperator::CreateXor(NOr, C1);
1865 Value *C = 0, *D = 0;
1866 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1867 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1868 Value *V1 = 0, *V2 = 0;
1869 C1 = dyn_cast<ConstantInt>(C);
1870 C2 = dyn_cast<ConstantInt>(D);
1871 if (C1 && C2) { // (A & C1)|(B & C2)
1872 // If we have: ((V + N) & C1) | (V & C2)
1873 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1874 // replace with V+N.
1875 if (C1->getValue() == ~C2->getValue()) {
1876 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
1877 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1878 // Add commutes, try both ways.
1879 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
1880 return ReplaceInstUsesWith(I, A);
1881 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
1882 return ReplaceInstUsesWith(I, A);
1884 // Or commutes, try both ways.
1885 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
1886 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1887 // Add commutes, try both ways.
1888 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
1889 return ReplaceInstUsesWith(I, B);
1890 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
1891 return ReplaceInstUsesWith(I, B);
1895 if ((C1->getValue() & C2->getValue()) == 0) {
1896 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
1897 // iff (C1&C2) == 0 and (N&~C1) == 0
1898 if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
1899 ((V1 == B && MaskedValueIsZero(V2, ~C1->getValue())) || // (V|N)
1900 (V2 == B && MaskedValueIsZero(V1, ~C1->getValue())))) // (N|V)
1901 return BinaryOperator::CreateAnd(A,
1902 Builder->getInt(C1->getValue()|C2->getValue()));
1903 // Or commutes, try both ways.
1904 if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
1905 ((V1 == A && MaskedValueIsZero(V2, ~C2->getValue())) || // (V|N)
1906 (V2 == A && MaskedValueIsZero(V1, ~C2->getValue())))) // (N|V)
1907 return BinaryOperator::CreateAnd(B,
1908 Builder->getInt(C1->getValue()|C2->getValue()));
1910 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
1911 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
1912 ConstantInt *C3 = 0, *C4 = 0;
1913 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
1914 (C3->getValue() & ~C1->getValue()) == 0 &&
1915 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
1916 (C4->getValue() & ~C2->getValue()) == 0) {
1917 V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
1918 return BinaryOperator::CreateAnd(V2,
1919 Builder->getInt(C1->getValue()|C2->getValue()));
1924 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants.
1925 // Don't do this for vector select idioms, the code generator doesn't handle
1927 if (!I.getType()->isVectorTy()) {
1928 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
1930 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
1932 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
1934 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
1938 // ((A&~B)|(~A&B)) -> A^B
1939 if ((match(C, m_Not(m_Specific(D))) &&
1940 match(B, m_Not(m_Specific(A)))))
1941 return BinaryOperator::CreateXor(A, D);
1942 // ((~B&A)|(~A&B)) -> A^B
1943 if ((match(A, m_Not(m_Specific(D))) &&
1944 match(B, m_Not(m_Specific(C)))))
1945 return BinaryOperator::CreateXor(C, D);
1946 // ((A&~B)|(B&~A)) -> A^B
1947 if ((match(C, m_Not(m_Specific(B))) &&
1948 match(D, m_Not(m_Specific(A)))))
1949 return BinaryOperator::CreateXor(A, B);
1950 // ((~B&A)|(B&~A)) -> A^B
1951 if ((match(A, m_Not(m_Specific(B))) &&
1952 match(D, m_Not(m_Specific(C)))))
1953 return BinaryOperator::CreateXor(C, B);
1955 // ((A|B)&1)|(B&-2) -> (A&1) | B
1956 if (match(A, m_Or(m_Value(V1), m_Specific(B))) ||
1957 match(A, m_Or(m_Specific(B), m_Value(V1)))) {
1958 Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C);
1959 if (Ret) return Ret;
1961 // (B&-2)|((A|B)&1) -> (A&1) | B
1962 if (match(B, m_Or(m_Specific(A), m_Value(V1))) ||
1963 match(B, m_Or(m_Value(V1), m_Specific(A)))) {
1964 Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D);
1965 if (Ret) return Ret;
1969 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
1970 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
1971 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
1972 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
1973 SI0->getOperand(1) == SI1->getOperand(1) &&
1974 (SI0->hasOneUse() || SI1->hasOneUse())) {
1975 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
1977 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
1978 SI1->getOperand(1));
1982 // (~A | ~B) == (~(A & B)) - De Morgan's Law
1983 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1984 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1985 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1986 Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
1987 I.getName()+".demorgan");
1988 return BinaryOperator::CreateNot(And);
1991 // Canonicalize xor to the RHS.
1992 bool SwappedForXor = false;
1993 if (match(Op0, m_Xor(m_Value(), m_Value()))) {
1994 std::swap(Op0, Op1);
1995 SwappedForXor = true;
1998 // A | ( A ^ B) -> A | B
1999 // A | (~A ^ B) -> A | ~B
2000 // (A & B) | (A ^ B)
2001 if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
2002 if (Op0 == A || Op0 == B)
2003 return BinaryOperator::CreateOr(A, B);
2005 if (match(Op0, m_And(m_Specific(A), m_Specific(B))) ||
2006 match(Op0, m_And(m_Specific(B), m_Specific(A))))
2007 return BinaryOperator::CreateOr(A, B);
2009 if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
2010 Value *Not = Builder->CreateNot(B, B->getName()+".not");
2011 return BinaryOperator::CreateOr(Not, Op0);
2013 if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
2014 Value *Not = Builder->CreateNot(A, A->getName()+".not");
2015 return BinaryOperator::CreateOr(Not, Op0);
2019 // A | ~(A | B) -> A | ~B
2020 // A | ~(A ^ B) -> A | ~B
2021 if (match(Op1, m_Not(m_Value(A))))
2022 if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
2023 if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
2024 Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
2025 B->getOpcode() == Instruction::Xor)) {
2026 Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
2028 Value *Not = Builder->CreateNot(NotOp, NotOp->getName()+".not");
2029 return BinaryOperator::CreateOr(Not, Op0);
2033 std::swap(Op0, Op1);
2035 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
2036 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
2037 if (Value *Res = FoldOrOfICmps(LHS, RHS))
2038 return ReplaceInstUsesWith(I, Res);
2040 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
2041 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
2042 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
2043 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
2044 return ReplaceInstUsesWith(I, Res);
2046 // fold (or (cast A), (cast B)) -> (cast (or A, B))
2047 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2048 CastInst *Op1C = dyn_cast<CastInst>(Op1);
2049 if (Op1C && Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
2050 Type *SrcTy = Op0C->getOperand(0)->getType();
2051 if (SrcTy == Op1C->getOperand(0)->getType() &&
2052 SrcTy->isIntOrIntVectorTy()) {
2053 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
2055 if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) &&
2056 // Only do this if the casts both really cause code to be
2058 ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
2059 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
2060 Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName());
2061 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2064 // If this is or(cast(icmp), cast(icmp)), try to fold this even if the
2065 // cast is otherwise not optimizable. This happens for vector sexts.
2066 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
2067 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
2068 if (Value *Res = FoldOrOfICmps(LHS, RHS))
2069 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
2071 // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the
2072 // cast is otherwise not optimizable. This happens for vector sexts.
2073 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
2074 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
2075 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
2076 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
2081 // or(sext(A), B) -> A ? -1 : B where A is an i1
2082 // or(A, sext(B)) -> B ? -1 : A where B is an i1
2083 if (match(Op0, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
2084 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1);
2085 if (match(Op1, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
2086 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0);
2088 // Note: If we've gotten to the point of visiting the outer OR, then the
2089 // inner one couldn't be simplified. If it was a constant, then it won't
2090 // be simplified by a later pass either, so we try swapping the inner/outer
2091 // ORs in the hopes that we'll be able to simplify it this way.
2092 // (X|C) | V --> (X|V) | C
2093 if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
2094 match(Op0, m_Or(m_Value(A), m_ConstantInt(C1)))) {
2095 Value *Inner = Builder->CreateOr(A, Op1);
2096 Inner->takeName(Op0);
2097 return BinaryOperator::CreateOr(Inner, C1);
2100 // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D))
2101 // Since this OR statement hasn't been optimized further yet, we hope
2102 // that this transformation will allow the new ORs to be optimized.
2104 Value *X = 0, *Y = 0;
2105 if (Op0->hasOneUse() && Op1->hasOneUse() &&
2106 match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) &&
2107 match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) {
2108 Value *orTrue = Builder->CreateOr(A, C);
2109 Value *orFalse = Builder->CreateOr(B, D);
2110 return SelectInst::Create(X, orTrue, orFalse);
2114 return Changed ? &I : 0;
2117 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
2118 bool Changed = SimplifyAssociativeOrCommutative(I);
2119 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2121 if (Value *V = SimplifyXorInst(Op0, Op1, TD))
2122 return ReplaceInstUsesWith(I, V);
2124 // (A&B)^(A&C) -> A&(B^C) etc
2125 if (Value *V = SimplifyUsingDistributiveLaws(I))
2126 return ReplaceInstUsesWith(I, V);
2128 // See if we can simplify any instructions used by the instruction whose sole
2129 // purpose is to compute bits we don't care about.
2130 if (SimplifyDemandedInstructionBits(I))
2133 // Is this a ~ operation?
2134 if (Value *NotOp = dyn_castNotVal(&I)) {
2135 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
2136 if (Op0I->getOpcode() == Instruction::And ||
2137 Op0I->getOpcode() == Instruction::Or) {
2138 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
2139 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
2140 if (dyn_castNotVal(Op0I->getOperand(1)))
2141 Op0I->swapOperands();
2142 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
2144 Builder->CreateNot(Op0I->getOperand(1),
2145 Op0I->getOperand(1)->getName()+".not");
2146 if (Op0I->getOpcode() == Instruction::And)
2147 return BinaryOperator::CreateOr(Op0NotVal, NotY);
2148 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
2151 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
2152 // ~(X | Y) === (~X & ~Y) - De Morgan's Law
2153 if (isFreeToInvert(Op0I->getOperand(0)) &&
2154 isFreeToInvert(Op0I->getOperand(1))) {
2156 Builder->CreateNot(Op0I->getOperand(0), "notlhs");
2158 Builder->CreateNot(Op0I->getOperand(1), "notrhs");
2159 if (Op0I->getOpcode() == Instruction::And)
2160 return BinaryOperator::CreateOr(NotX, NotY);
2161 return BinaryOperator::CreateAnd(NotX, NotY);
2164 } else if (Op0I->getOpcode() == Instruction::AShr) {
2165 // ~(~X >>s Y) --> (X >>s Y)
2166 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0)))
2167 return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1));
2173 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2174 if (RHS->isOne() && Op0->hasOneUse())
2175 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
2176 if (CmpInst *CI = dyn_cast<CmpInst>(Op0))
2177 return CmpInst::Create(CI->getOpcode(),
2178 CI->getInversePredicate(),
2179 CI->getOperand(0), CI->getOperand(1));
2181 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
2182 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2183 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
2184 if (CI->hasOneUse() && Op0C->hasOneUse()) {
2185 Instruction::CastOps Opcode = Op0C->getOpcode();
2186 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
2187 (RHS == ConstantExpr::getCast(Opcode, Builder->getTrue(),
2188 Op0C->getDestTy()))) {
2189 CI->setPredicate(CI->getInversePredicate());
2190 return CastInst::Create(Opcode, CI, Op0C->getType());
2196 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2197 // ~(c-X) == X-c-1 == X+(-c-1)
2198 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
2199 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
2200 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
2201 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
2202 ConstantInt::get(I.getType(), 1));
2203 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
2206 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2207 if (Op0I->getOpcode() == Instruction::Add) {
2208 // ~(X-c) --> (-c-1)-X
2209 if (RHS->isAllOnesValue()) {
2210 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
2211 return BinaryOperator::CreateSub(
2212 ConstantExpr::getSub(NegOp0CI,
2213 ConstantInt::get(I.getType(), 1)),
2214 Op0I->getOperand(0));
2215 } else if (RHS->getValue().isSignBit()) {
2216 // (X + C) ^ signbit -> (X + C + signbit)
2217 Constant *C = Builder->getInt(RHS->getValue() + Op0CI->getValue());
2218 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
2221 } else if (Op0I->getOpcode() == Instruction::Or) {
2222 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
2223 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
2224 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
2225 // Anything in both C1 and C2 is known to be zero, remove it from
2227 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
2228 NewRHS = ConstantExpr::getAnd(NewRHS,
2229 ConstantExpr::getNot(CommonBits));
2231 I.setOperand(0, Op0I->getOperand(0));
2232 I.setOperand(1, NewRHS);
2235 } else if (Op0I->getOpcode() == Instruction::LShr) {
2236 // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3)
2240 if (Op0I->hasOneUse() &&
2241 (E1 = dyn_cast<BinaryOperator>(Op0I->getOperand(0))) &&
2242 E1->getOpcode() == Instruction::Xor &&
2243 (C1 = dyn_cast<ConstantInt>(E1->getOperand(1)))) {
2244 // fold (C1 >> C2) ^ C3
2245 ConstantInt *C2 = Op0CI, *C3 = RHS;
2246 APInt FoldConst = C1->getValue().lshr(C2->getValue());
2247 FoldConst ^= C3->getValue();
2248 // Prepare the two operands.
2249 Value *Opnd0 = Builder->CreateLShr(E1->getOperand(0), C2);
2250 Opnd0->takeName(Op0I);
2251 cast<Instruction>(Opnd0)->setDebugLoc(I.getDebugLoc());
2252 Value *FoldVal = ConstantInt::get(Opnd0->getType(), FoldConst);
2254 return BinaryOperator::CreateXor(Opnd0, FoldVal);
2260 // Try to fold constant and into select arguments.
2261 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2262 if (Instruction *R = FoldOpIntoSelect(I, SI))
2264 if (isa<PHINode>(Op0))
2265 if (Instruction *NV = FoldOpIntoPhi(I))
2269 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
2272 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2273 if (A == Op0) { // B^(B|A) == (A|B)^B
2274 Op1I->swapOperands();
2276 std::swap(Op0, Op1);
2277 } else if (B == Op0) { // B^(A|B) == (A|B)^B
2278 I.swapOperands(); // Simplified below.
2279 std::swap(Op0, Op1);
2281 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
2283 if (A == Op0) { // A^(A&B) -> A^(B&A)
2284 Op1I->swapOperands();
2287 if (B == Op0) { // A^(B&A) -> (B&A)^A
2288 I.swapOperands(); // Simplified below.
2289 std::swap(Op0, Op1);
2294 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
2297 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2298 Op0I->hasOneUse()) {
2299 if (A == Op1) // (B|A)^B == (A|B)^B
2301 if (B == Op1) // (A|B)^B == A & ~B
2302 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1));
2303 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2305 if (A == Op1) // (A&B)^A -> (B&A)^A
2307 if (B == Op1 && // (B&A)^A == ~B & A
2308 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
2309 return BinaryOperator::CreateAnd(Builder->CreateNot(A), Op1);
2314 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
2315 if (Op0I && Op1I && Op0I->isShift() &&
2316 Op0I->getOpcode() == Op1I->getOpcode() &&
2317 Op0I->getOperand(1) == Op1I->getOperand(1) &&
2318 (Op0I->hasOneUse() || Op1I->hasOneUse())) {
2320 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
2322 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
2323 Op1I->getOperand(1));
2327 Value *A, *B, *C, *D;
2328 // (A & B)^(A | B) -> A ^ B
2329 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2330 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
2331 if ((A == C && B == D) || (A == D && B == C))
2332 return BinaryOperator::CreateXor(A, B);
2334 // (A | B)^(A & B) -> A ^ B
2335 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2336 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
2337 if ((A == C && B == D) || (A == D && B == C))
2338 return BinaryOperator::CreateXor(A, B);
2342 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
2343 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
2344 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
2345 if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
2346 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2347 LHS->getOperand(1) == RHS->getOperand(0))
2348 LHS->swapOperands();
2349 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2350 LHS->getOperand(1) == RHS->getOperand(1)) {
2351 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2352 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
2353 bool isSigned = LHS->isSigned() || RHS->isSigned();
2354 return ReplaceInstUsesWith(I,
2355 getNewICmpValue(isSigned, Code, Op0, Op1,
2360 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
2361 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2362 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
2363 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
2364 Type *SrcTy = Op0C->getOperand(0)->getType();
2365 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() &&
2366 // Only do this if the casts both really cause code to be generated.
2367 ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0),
2369 ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0),
2371 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
2372 Op1C->getOperand(0), I.getName());
2373 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2378 return Changed ? &I : 0;