1 //===- InstCombineAndOrXor.cpp --------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visitAnd, visitOr, and visitXor functions.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombine.h"
15 #include "llvm/Intrinsics.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/Support/PatternMatch.h"
19 using namespace PatternMatch;
22 /// AddOne - Add one to a ConstantInt.
23 static Constant *AddOne(Constant *C) {
24 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
26 /// SubOne - Subtract one from a ConstantInt.
27 static Constant *SubOne(ConstantInt *C) {
28 return ConstantInt::get(C->getContext(), C->getValue()-1);
31 /// isFreeToInvert - Return true if the specified value is free to invert (apply
32 /// ~ to). This happens in cases where the ~ can be eliminated.
33 static inline bool isFreeToInvert(Value *V) {
35 if (BinaryOperator::isNot(V))
38 // Constants can be considered to be not'ed values.
39 if (isa<ConstantInt>(V))
42 // Compares can be inverted if they have a single use.
43 if (CmpInst *CI = dyn_cast<CmpInst>(V))
44 return CI->hasOneUse();
49 static inline Value *dyn_castNotVal(Value *V) {
50 // If this is not(not(x)) don't return that this is a not: we want the two
51 // not's to be folded first.
52 if (BinaryOperator::isNot(V)) {
53 Value *Operand = BinaryOperator::getNotArgument(V);
54 if (!isFreeToInvert(Operand))
58 // Constants can be considered to be not'ed values...
59 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
60 return ConstantInt::get(C->getType(), ~C->getValue());
65 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
66 /// are carefully arranged to allow folding of expressions such as:
68 /// (A < B) | (A > B) --> (A != B)
70 /// Note that this is only valid if the first and second predicates have the
71 /// same sign. Is illegal to do: (A u< B) | (A s> B)
73 /// Three bits are used to represent the condition, as follows:
78 /// <=> Value Definition
79 /// 000 0 Always false
88 static unsigned getICmpCode(const ICmpInst *ICI) {
89 switch (ICI->getPredicate()) {
91 case ICmpInst::ICMP_UGT: return 1; // 001
92 case ICmpInst::ICMP_SGT: return 1; // 001
93 case ICmpInst::ICMP_EQ: return 2; // 010
94 case ICmpInst::ICMP_UGE: return 3; // 011
95 case ICmpInst::ICMP_SGE: return 3; // 011
96 case ICmpInst::ICMP_ULT: return 4; // 100
97 case ICmpInst::ICMP_SLT: return 4; // 100
98 case ICmpInst::ICMP_NE: return 5; // 101
99 case ICmpInst::ICMP_ULE: return 6; // 110
100 case ICmpInst::ICMP_SLE: return 6; // 110
103 llvm_unreachable("Invalid ICmp predicate!");
108 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
109 /// predicate into a three bit mask. It also returns whether it is an ordered
110 /// predicate by reference.
111 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
114 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
115 case FCmpInst::FCMP_UNO: return 0; // 000
116 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
117 case FCmpInst::FCMP_UGT: return 1; // 001
118 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
119 case FCmpInst::FCMP_UEQ: return 2; // 010
120 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
121 case FCmpInst::FCMP_UGE: return 3; // 011
122 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
123 case FCmpInst::FCMP_ULT: return 4; // 100
124 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
125 case FCmpInst::FCMP_UNE: return 5; // 101
126 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
127 case FCmpInst::FCMP_ULE: return 6; // 110
130 // Not expecting FCMP_FALSE and FCMP_TRUE;
131 llvm_unreachable("Unexpected FCmp predicate!");
136 /// getICmpValue - This is the complement of getICmpCode, which turns an
137 /// opcode and two operands into either a constant true or false, or a brand
138 /// new ICmp instruction. The sign is passed in to determine which kind
139 /// of predicate to use in the new icmp instruction.
140 static Value *getICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS,
141 InstCombiner::BuilderTy *Builder) {
142 CmpInst::Predicate Pred;
144 default: assert(0 && "Illegal ICmp code!");
146 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
147 case 1: Pred = Sign ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; break;
148 case 2: Pred = ICmpInst::ICMP_EQ; break;
149 case 3: Pred = Sign ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; break;
150 case 4: Pred = Sign ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; break;
151 case 5: Pred = ICmpInst::ICMP_NE; break;
152 case 6: Pred = Sign ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; break;
154 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
156 return Builder->CreateICmp(Pred, LHS, RHS);
159 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
160 /// opcode and two operands into either a FCmp instruction. isordered is passed
161 /// in to determine which kind of predicate to use in the new fcmp instruction.
162 static Value *getFCmpValue(bool isordered, unsigned code,
163 Value *LHS, Value *RHS,
164 InstCombiner::BuilderTy *Builder) {
165 CmpInst::Predicate Pred;
167 default: assert(0 && "Illegal FCmp code!");
168 case 0: Pred = isordered ? FCmpInst::FCMP_ORD : FCmpInst::FCMP_UNO; break;
169 case 1: Pred = isordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; break;
170 case 2: Pred = isordered ? FCmpInst::FCMP_OEQ : FCmpInst::FCMP_UEQ; break;
171 case 3: Pred = isordered ? FCmpInst::FCMP_OGE : FCmpInst::FCMP_UGE; break;
172 case 4: Pred = isordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; break;
173 case 5: Pred = isordered ? FCmpInst::FCMP_ONE : FCmpInst::FCMP_UNE; break;
174 case 6: Pred = isordered ? FCmpInst::FCMP_OLE : FCmpInst::FCMP_ULE; break;
175 case 7: return ConstantInt::getTrue(LHS->getContext());
177 return Builder->CreateFCmp(Pred, LHS, RHS);
180 /// PredicatesFoldable - Return true if both predicates match sign or if at
181 /// least one of them is an equality comparison (which is signless).
182 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
183 return (CmpInst::isSigned(p1) == CmpInst::isSigned(p2)) ||
184 (CmpInst::isSigned(p1) && ICmpInst::isEquality(p2)) ||
185 (CmpInst::isSigned(p2) && ICmpInst::isEquality(p1));
188 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
189 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
190 // guaranteed to be a binary operator.
191 Instruction *InstCombiner::OptAndOp(Instruction *Op,
194 BinaryOperator &TheAnd) {
195 Value *X = Op->getOperand(0);
196 Constant *Together = 0;
198 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
200 switch (Op->getOpcode()) {
201 case Instruction::Xor:
202 if (Op->hasOneUse()) {
203 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
204 Value *And = Builder->CreateAnd(X, AndRHS);
206 return BinaryOperator::CreateXor(And, Together);
209 case Instruction::Or:
210 if (Together == AndRHS) // (X | C) & C --> C
211 return ReplaceInstUsesWith(TheAnd, AndRHS);
213 if (Op->hasOneUse() && Together != OpRHS) {
214 // (X | C1) & C2 --> (X | (C1&C2)) & C2
215 Value *Or = Builder->CreateOr(X, Together);
217 return BinaryOperator::CreateAnd(Or, AndRHS);
220 case Instruction::Add:
221 if (Op->hasOneUse()) {
222 // Adding a one to a single bit bit-field should be turned into an XOR
223 // of the bit. First thing to check is to see if this AND is with a
224 // single bit constant.
225 const APInt &AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
227 // If there is only one bit set.
228 if (AndRHSV.isPowerOf2()) {
229 // Ok, at this point, we know that we are masking the result of the
230 // ADD down to exactly one bit. If the constant we are adding has
231 // no bits set below this bit, then we can eliminate the ADD.
232 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
234 // Check to see if any bits below the one bit set in AndRHSV are set.
235 if ((AddRHS & (AndRHSV-1)) == 0) {
236 // If not, the only thing that can effect the output of the AND is
237 // the bit specified by AndRHSV. If that bit is set, the effect of
238 // the XOR is to toggle the bit. If it is clear, then the ADD has
240 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
241 TheAnd.setOperand(0, X);
244 // Pull the XOR out of the AND.
245 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
246 NewAnd->takeName(Op);
247 return BinaryOperator::CreateXor(NewAnd, AndRHS);
254 case Instruction::Shl: {
255 // We know that the AND will not produce any of the bits shifted in, so if
256 // the anded constant includes them, clear them now!
258 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
259 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
260 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
261 ConstantInt *CI = ConstantInt::get(AndRHS->getContext(),
262 AndRHS->getValue() & ShlMask);
264 if (CI->getValue() == ShlMask) {
265 // Masking out bits that the shift already masks
266 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
267 } else if (CI != AndRHS) { // Reducing bits set in and.
268 TheAnd.setOperand(1, CI);
273 case Instruction::LShr: {
274 // We know that the AND will not produce any of the bits shifted in, so if
275 // the anded constant includes them, clear them now! This only applies to
276 // unsigned shifts, because a signed shr may bring in set bits!
278 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
279 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
280 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
281 ConstantInt *CI = ConstantInt::get(Op->getContext(),
282 AndRHS->getValue() & ShrMask);
284 if (CI->getValue() == ShrMask) {
285 // Masking out bits that the shift already masks.
286 return ReplaceInstUsesWith(TheAnd, Op);
287 } else if (CI != AndRHS) {
288 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
293 case Instruction::AShr:
295 // See if this is shifting in some sign extension, then masking it out
297 if (Op->hasOneUse()) {
298 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
299 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
300 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
301 Constant *C = ConstantInt::get(Op->getContext(),
302 AndRHS->getValue() & ShrMask);
303 if (C == AndRHS) { // Masking out bits shifted in.
304 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
305 // Make the argument unsigned.
306 Value *ShVal = Op->getOperand(0);
307 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
308 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
317 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
318 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
319 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
320 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
321 /// insert new instructions.
322 Value *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
323 bool isSigned, bool Inside) {
324 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
325 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
326 "Lo is not <= Hi in range emission code!");
329 if (Lo == Hi) // Trivially false.
330 return ConstantInt::getFalse(V->getContext());
332 // V >= Min && V < Hi --> V < Hi
333 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
334 ICmpInst::Predicate pred = (isSigned ?
335 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
336 return Builder->CreateICmp(pred, V, Hi);
339 // Emit V-Lo <u Hi-Lo
340 Constant *NegLo = ConstantExpr::getNeg(Lo);
341 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
342 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
343 return Builder->CreateICmpULT(Add, UpperBound);
346 if (Lo == Hi) // Trivially true.
347 return ConstantInt::getTrue(V->getContext());
349 // V < Min || V >= Hi -> V > Hi-1
350 Hi = SubOne(cast<ConstantInt>(Hi));
351 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
352 ICmpInst::Predicate pred = (isSigned ?
353 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
354 return Builder->CreateICmp(pred, V, Hi);
357 // Emit V-Lo >u Hi-1-Lo
358 // Note that Hi has already had one subtracted from it, above.
359 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
360 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
361 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
362 return Builder->CreateICmpUGT(Add, LowerBound);
365 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
366 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
367 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
368 // not, since all 1s are not contiguous.
369 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
370 const APInt& V = Val->getValue();
371 uint32_t BitWidth = Val->getType()->getBitWidth();
372 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
374 // look for the first zero bit after the run of ones
375 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
376 // look for the first non-zero bit
377 ME = V.getActiveBits();
381 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
382 /// where isSub determines whether the operator is a sub. If we can fold one of
383 /// the following xforms:
385 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
386 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
387 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
389 /// return (A +/- B).
391 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
392 ConstantInt *Mask, bool isSub,
394 Instruction *LHSI = dyn_cast<Instruction>(LHS);
395 if (!LHSI || LHSI->getNumOperands() != 2 ||
396 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
398 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
400 switch (LHSI->getOpcode()) {
402 case Instruction::And:
403 if (ConstantExpr::getAnd(N, Mask) == Mask) {
404 // If the AndRHS is a power of two minus one (0+1+), this is simple.
405 if ((Mask->getValue().countLeadingZeros() +
406 Mask->getValue().countPopulation()) ==
407 Mask->getValue().getBitWidth())
410 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
411 // part, we don't need any explicit masks to take them out of A. If that
412 // is all N is, ignore it.
413 uint32_t MB = 0, ME = 0;
414 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
415 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
416 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
417 if (MaskedValueIsZero(RHS, Mask))
422 case Instruction::Or:
423 case Instruction::Xor:
424 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
425 if ((Mask->getValue().countLeadingZeros() +
426 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
427 && ConstantExpr::getAnd(N, Mask)->isNullValue())
433 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
434 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
437 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
438 Value *InstCombiner::FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
439 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
441 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
442 if (PredicatesFoldable(LHSCC, RHSCC)) {
443 if (LHS->getOperand(0) == RHS->getOperand(1) &&
444 LHS->getOperand(1) == RHS->getOperand(0))
446 if (LHS->getOperand(0) == RHS->getOperand(0) &&
447 LHS->getOperand(1) == RHS->getOperand(1)) {
448 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
449 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
450 bool isSigned = LHS->isSigned() || RHS->isSigned();
451 return getICmpValue(isSigned, Code, Op0, Op1, Builder);
455 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
456 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
457 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
458 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
459 if (LHSCst == 0 || RHSCst == 0) return 0;
461 if (LHSCst == RHSCst && LHSCC == RHSCC) {
462 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
463 // where C is a power of 2
464 if (LHSCC == ICmpInst::ICMP_ULT &&
465 LHSCst->getValue().isPowerOf2()) {
466 Value *NewOr = Builder->CreateOr(Val, Val2);
467 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
470 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
471 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
472 Value *NewOr = Builder->CreateOr(Val, Val2);
473 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
476 // (icmp ne (A & C1), 0) & (icmp ne (A & C2), 0) -->
477 // (icmp eq (A & (C1|C2)), (C1|C2)) where C1 and C2 are non-zero POT
478 if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
479 Value *Op1 = 0, *Op2 = 0;
480 ConstantInt *CI1 = 0, *CI2 = 0;
481 if (match(LHS->getOperand(0), m_And(m_Value(Op1), m_ConstantInt(CI1))) &&
482 match(RHS->getOperand(0), m_And(m_Value(Op2), m_ConstantInt(CI2)))) {
483 if (Op1 == Op2 && !CI1->isZero() && !CI2->isZero() &&
484 CI1->getValue().isPowerOf2() && CI2->getValue().isPowerOf2()) {
485 Constant *ConstOr = ConstantExpr::getOr(CI1, CI2);
486 Value *NewAnd = Builder->CreateAnd(Op1, ConstOr);
487 return Builder->CreateICmp(ICmpInst::ICMP_EQ, NewAnd, ConstOr);
493 // From here on, we only handle:
494 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
495 if (Val != Val2) return 0;
497 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
498 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
499 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
500 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
501 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
504 // We can't fold (ugt x, C) & (sgt x, C2).
505 if (!PredicatesFoldable(LHSCC, RHSCC))
508 // Ensure that the larger constant is on the RHS.
510 if (CmpInst::isSigned(LHSCC) ||
511 (ICmpInst::isEquality(LHSCC) &&
512 CmpInst::isSigned(RHSCC)))
513 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
515 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
519 std::swap(LHSCst, RHSCst);
520 std::swap(LHSCC, RHSCC);
523 // At this point, we know we have two icmp instructions
524 // comparing a value against two constants and and'ing the result
525 // together. Because of the above check, we know that we only have
526 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
527 // (from the icmp folding check above), that the two constants
528 // are not equal and that the larger constant is on the RHS
529 assert(LHSCst != RHSCst && "Compares not folded above?");
532 default: llvm_unreachable("Unknown integer condition code!");
533 case ICmpInst::ICMP_EQ:
535 default: llvm_unreachable("Unknown integer condition code!");
536 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
537 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
538 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
539 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
540 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
541 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
542 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
545 case ICmpInst::ICMP_NE:
547 default: llvm_unreachable("Unknown integer condition code!");
548 case ICmpInst::ICMP_ULT:
549 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
550 return Builder->CreateICmpULT(Val, LHSCst);
551 break; // (X != 13 & X u< 15) -> no change
552 case ICmpInst::ICMP_SLT:
553 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
554 return Builder->CreateICmpSLT(Val, LHSCst);
555 break; // (X != 13 & X s< 15) -> no change
556 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
557 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
558 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
560 case ICmpInst::ICMP_NE:
561 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
562 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
563 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
564 return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1));
566 break; // (X != 13 & X != 15) -> no change
569 case ICmpInst::ICMP_ULT:
571 default: llvm_unreachable("Unknown integer condition code!");
572 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
573 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
574 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
575 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
577 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
578 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
580 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
584 case ICmpInst::ICMP_SLT:
586 default: llvm_unreachable("Unknown integer condition code!");
587 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
588 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
589 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
590 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
592 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
593 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
595 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
599 case ICmpInst::ICMP_UGT:
601 default: llvm_unreachable("Unknown integer condition code!");
602 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
603 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
605 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
607 case ICmpInst::ICMP_NE:
608 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
609 return Builder->CreateICmp(LHSCC, Val, RHSCst);
610 break; // (X u> 13 & X != 15) -> no change
611 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
612 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
613 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
617 case ICmpInst::ICMP_SGT:
619 default: llvm_unreachable("Unknown integer condition code!");
620 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
621 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
623 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
625 case ICmpInst::ICMP_NE:
626 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
627 return Builder->CreateICmp(LHSCC, Val, RHSCst);
628 break; // (X s> 13 & X != 15) -> no change
629 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
630 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true);
631 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
640 /// FoldAndOfFCmps - Optimize (fcmp)&(fcmp). NOTE: Unlike the rest of
641 /// instcombine, this returns a Value which should already be inserted into the
643 Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
644 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
645 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
646 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
647 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
648 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
649 // If either of the constants are nans, then the whole thing returns
651 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
652 return ConstantInt::getFalse(LHS->getContext());
653 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
656 // Handle vector zeros. This occurs because the canonical form of
657 // "fcmp ord x,x" is "fcmp ord x, 0".
658 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
659 isa<ConstantAggregateZero>(RHS->getOperand(1)))
660 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
664 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
665 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
666 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
669 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
670 // Swap RHS operands to match LHS.
671 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
672 std::swap(Op1LHS, Op1RHS);
675 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
676 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
678 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
679 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
680 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
681 if (Op0CC == FCmpInst::FCMP_TRUE)
683 if (Op1CC == FCmpInst::FCMP_TRUE)
688 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
689 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
692 std::swap(Op0Pred, Op1Pred);
693 std::swap(Op0Ordered, Op1Ordered);
696 // uno && ueq -> uno && (uno || eq) -> ueq
697 // ord && olt -> ord && (ord && lt) -> olt
698 if (Op0Ordered == Op1Ordered)
701 // uno && oeq -> uno && (ord && eq) -> false
702 // uno && ord -> false
704 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
705 // ord && ueq -> ord && (uno || eq) -> oeq
706 return getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS, Builder);
714 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
715 bool Changed = SimplifyCommutative(I);
716 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
718 if (Value *V = SimplifyAndInst(Op0, Op1, TD))
719 return ReplaceInstUsesWith(I, V);
721 // See if we can simplify any instructions used by the instruction whose sole
722 // purpose is to compute bits we don't care about.
723 if (SimplifyDemandedInstructionBits(I))
726 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
727 const APInt &AndRHSMask = AndRHS->getValue();
728 APInt NotAndRHS(~AndRHSMask);
730 // Optimize a variety of ((val OP C1) & C2) combinations...
731 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
732 Value *Op0LHS = Op0I->getOperand(0);
733 Value *Op0RHS = Op0I->getOperand(1);
734 switch (Op0I->getOpcode()) {
736 case Instruction::Xor:
737 case Instruction::Or:
738 // If the mask is only needed on one incoming arm, push it up.
739 if (!Op0I->hasOneUse()) break;
741 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
742 // Not masking anything out for the LHS, move to RHS.
743 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
744 Op0RHS->getName()+".masked");
745 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
747 if (!isa<Constant>(Op0RHS) &&
748 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
749 // Not masking anything out for the RHS, move to LHS.
750 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
751 Op0LHS->getName()+".masked");
752 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
756 case Instruction::Add:
757 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
758 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
759 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
760 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
761 return BinaryOperator::CreateAnd(V, AndRHS);
762 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
763 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
766 case Instruction::Sub:
767 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
768 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
769 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
770 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
771 return BinaryOperator::CreateAnd(V, AndRHS);
773 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
774 // has 1's for all bits that the subtraction with A might affect.
775 if (Op0I->hasOneUse()) {
776 uint32_t BitWidth = AndRHSMask.getBitWidth();
777 uint32_t Zeros = AndRHSMask.countLeadingZeros();
778 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
780 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
781 if (!(A && A->isZero()) && // avoid infinite recursion.
782 MaskedValueIsZero(Op0LHS, Mask)) {
783 Value *NewNeg = Builder->CreateNeg(Op0RHS);
784 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
789 case Instruction::Shl:
790 case Instruction::LShr:
791 // (1 << x) & 1 --> zext(x == 0)
792 // (1 >> x) & 1 --> zext(x == 0)
793 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
795 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
796 return new ZExtInst(NewICmp, I.getType());
801 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
802 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
804 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
805 // If this is an integer truncation or change from signed-to-unsigned, and
806 // if the source is an and/or with immediate, transform it. This
807 // frequently occurs for bitfield accesses.
808 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
809 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
810 CastOp->getNumOperands() == 2)
811 if (ConstantInt *AndCI =dyn_cast<ConstantInt>(CastOp->getOperand(1))){
812 if (CastOp->getOpcode() == Instruction::And) {
813 // Change: and (cast (and X, C1) to T), C2
814 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
815 // This will fold the two constants together, which may allow
816 // other simplifications.
817 Value *NewCast = Builder->CreateTruncOrBitCast(
818 CastOp->getOperand(0), I.getType(),
819 CastOp->getName()+".shrunk");
820 // trunc_or_bitcast(C1)&C2
821 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
822 C3 = ConstantExpr::getAnd(C3, AndRHS);
823 return BinaryOperator::CreateAnd(NewCast, C3);
824 } else if (CastOp->getOpcode() == Instruction::Or) {
825 // Change: and (cast (or X, C1) to T), C2
826 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
827 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
828 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS)
830 return ReplaceInstUsesWith(I, AndRHS);
836 // Try to fold constant and into select arguments.
837 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
838 if (Instruction *R = FoldOpIntoSelect(I, SI))
840 if (isa<PHINode>(Op0))
841 if (Instruction *NV = FoldOpIntoPhi(I))
846 // (~A & ~B) == (~(A | B)) - De Morgan's Law
847 if (Value *Op0NotVal = dyn_castNotVal(Op0))
848 if (Value *Op1NotVal = dyn_castNotVal(Op1))
849 if (Op0->hasOneUse() && Op1->hasOneUse()) {
850 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
851 I.getName()+".demorgan");
852 return BinaryOperator::CreateNot(Or);
856 Value *A = 0, *B = 0, *C = 0, *D = 0;
857 // (A|B) & ~(A&B) -> A^B
858 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
859 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
860 ((A == C && B == D) || (A == D && B == C)))
861 return BinaryOperator::CreateXor(A, B);
863 // ~(A&B) & (A|B) -> A^B
864 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
865 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
866 ((A == C && B == D) || (A == D && B == C)))
867 return BinaryOperator::CreateXor(A, B);
869 if (Op0->hasOneUse() &&
870 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
871 if (A == Op1) { // (A^B)&A -> A&(A^B)
872 I.swapOperands(); // Simplify below
874 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
875 cast<BinaryOperator>(Op0)->swapOperands();
876 I.swapOperands(); // Simplify below
881 if (Op1->hasOneUse() &&
882 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
883 if (B == Op0) { // B&(A^B) -> B&(B^A)
884 cast<BinaryOperator>(Op1)->swapOperands();
887 if (A == Op0) // A&(A^B) -> A & ~B
888 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B, "tmp"));
891 // (A&((~A)|B)) -> A&B
892 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
893 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
894 return BinaryOperator::CreateAnd(A, Op1);
895 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
896 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
897 return BinaryOperator::CreateAnd(A, Op0);
900 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1))
901 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
902 if (Value *Res = FoldAndOfICmps(LHS, RHS))
903 return ReplaceInstUsesWith(I, Res);
905 // If and'ing two fcmp, try combine them into one.
906 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
907 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
908 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
909 return ReplaceInstUsesWith(I, Res);
912 // fold (and (cast A), (cast B)) -> (cast (and A, B))
913 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
914 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) {
915 const Type *SrcTy = Op0C->getOperand(0)->getType();
916 if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ?
917 SrcTy == Op1C->getOperand(0)->getType() &&
918 SrcTy->isIntOrIntVectorTy()) {
919 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
921 // Only do this if the casts both really cause code to be generated.
922 if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
923 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
924 Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName());
925 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
928 // If this is and(cast(icmp), cast(icmp)), try to fold this even if the
929 // cast is otherwise not optimizable. This happens for vector sexts.
930 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
931 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
932 if (Value *Res = FoldAndOfICmps(LHS, RHS))
933 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
935 // If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the
936 // cast is otherwise not optimizable. This happens for vector sexts.
937 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
938 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
939 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
940 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
944 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
945 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
946 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
947 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
948 SI0->getOperand(1) == SI1->getOperand(1) &&
949 (SI0->hasOneUse() || SI1->hasOneUse())) {
951 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
953 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
958 return Changed ? &I : 0;
961 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
962 /// capable of providing pieces of a bswap. The subexpression provides pieces
963 /// of a bswap if it is proven that each of the non-zero bytes in the output of
964 /// the expression came from the corresponding "byte swapped" byte in some other
965 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
966 /// we know that the expression deposits the low byte of %X into the high byte
967 /// of the bswap result and that all other bytes are zero. This expression is
968 /// accepted, the high byte of ByteValues is set to X to indicate a correct
971 /// This function returns true if the match was unsuccessful and false if so.
972 /// On entry to the function the "OverallLeftShift" is a signed integer value
973 /// indicating the number of bytes that the subexpression is later shifted. For
974 /// example, if the expression is later right shifted by 16 bits, the
975 /// OverallLeftShift value would be -2 on entry. This is used to specify which
976 /// byte of ByteValues is actually being set.
978 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
979 /// byte is masked to zero by a user. For example, in (X & 255), X will be
980 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
981 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
982 /// always in the local (OverallLeftShift) coordinate space.
984 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
985 SmallVector<Value*, 8> &ByteValues) {
986 if (Instruction *I = dyn_cast<Instruction>(V)) {
987 // If this is an or instruction, it may be an inner node of the bswap.
988 if (I->getOpcode() == Instruction::Or) {
989 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
991 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
995 // If this is a logical shift by a constant multiple of 8, recurse with
996 // OverallLeftShift and ByteMask adjusted.
997 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
999 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
1000 // Ensure the shift amount is defined and of a byte value.
1001 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
1004 unsigned ByteShift = ShAmt >> 3;
1005 if (I->getOpcode() == Instruction::Shl) {
1006 // X << 2 -> collect(X, +2)
1007 OverallLeftShift += ByteShift;
1008 ByteMask >>= ByteShift;
1010 // X >>u 2 -> collect(X, -2)
1011 OverallLeftShift -= ByteShift;
1012 ByteMask <<= ByteShift;
1013 ByteMask &= (~0U >> (32-ByteValues.size()));
1016 if (OverallLeftShift >= (int)ByteValues.size()) return true;
1017 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
1019 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1023 // If this is a logical 'and' with a mask that clears bytes, clear the
1024 // corresponding bytes in ByteMask.
1025 if (I->getOpcode() == Instruction::And &&
1026 isa<ConstantInt>(I->getOperand(1))) {
1027 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
1028 unsigned NumBytes = ByteValues.size();
1029 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
1030 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
1032 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
1033 // If this byte is masked out by a later operation, we don't care what
1035 if ((ByteMask & (1 << i)) == 0)
1038 // If the AndMask is all zeros for this byte, clear the bit.
1039 APInt MaskB = AndMask & Byte;
1041 ByteMask &= ~(1U << i);
1045 // If the AndMask is not all ones for this byte, it's not a bytezap.
1049 // Otherwise, this byte is kept.
1052 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1057 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
1058 // the input value to the bswap. Some observations: 1) if more than one byte
1059 // is demanded from this input, then it could not be successfully assembled
1060 // into a byteswap. At least one of the two bytes would not be aligned with
1061 // their ultimate destination.
1062 if (!isPowerOf2_32(ByteMask)) return true;
1063 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
1065 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
1066 // is demanded, it needs to go into byte 0 of the result. This means that the
1067 // byte needs to be shifted until it lands in the right byte bucket. The
1068 // shift amount depends on the position: if the byte is coming from the high
1069 // part of the value (e.g. byte 3) then it must be shifted right. If from the
1070 // low part, it must be shifted left.
1071 unsigned DestByteNo = InputByteNo + OverallLeftShift;
1072 if (InputByteNo < ByteValues.size()/2) {
1073 if (ByteValues.size()-1-DestByteNo != InputByteNo)
1076 if (ByteValues.size()-1-DestByteNo != InputByteNo)
1080 // If the destination byte value is already defined, the values are or'd
1081 // together, which isn't a bswap (unless it's an or of the same bits).
1082 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
1084 ByteValues[DestByteNo] = V;
1088 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
1089 /// If so, insert the new bswap intrinsic and return it.
1090 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
1091 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
1092 if (!ITy || ITy->getBitWidth() % 16 ||
1093 // ByteMask only allows up to 32-byte values.
1094 ITy->getBitWidth() > 32*8)
1095 return 0; // Can only bswap pairs of bytes. Can't do vectors.
1097 /// ByteValues - For each byte of the result, we keep track of which value
1098 /// defines each byte.
1099 SmallVector<Value*, 8> ByteValues;
1100 ByteValues.resize(ITy->getBitWidth()/8);
1102 // Try to find all the pieces corresponding to the bswap.
1103 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
1104 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
1107 // Check to see if all of the bytes come from the same value.
1108 Value *V = ByteValues[0];
1109 if (V == 0) return 0; // Didn't find a byte? Must be zero.
1111 // Check to make sure that all of the bytes come from the same value.
1112 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
1113 if (ByteValues[i] != V)
1115 const Type *Tys[] = { ITy };
1116 Module *M = I.getParent()->getParent()->getParent();
1117 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
1118 return CallInst::Create(F, V);
1121 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
1122 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
1123 /// we can simplify this expression to "cond ? C : D or B".
1124 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
1125 Value *C, Value *D) {
1126 // If A is not a select of -1/0, this cannot match.
1128 if (!match(A, m_SExt(m_Value(Cond))) ||
1129 !Cond->getType()->isIntegerTy(1))
1132 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
1133 if (match(D, m_Not(m_SExt(m_Specific(Cond)))))
1134 return SelectInst::Create(Cond, C, B);
1135 if (match(D, m_SExt(m_Not(m_Specific(Cond)))))
1136 return SelectInst::Create(Cond, C, B);
1138 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
1139 if (match(B, m_Not(m_SExt(m_Specific(Cond)))))
1140 return SelectInst::Create(Cond, C, D);
1141 if (match(B, m_SExt(m_Not(m_Specific(Cond)))))
1142 return SelectInst::Create(Cond, C, D);
1146 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
1147 Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
1148 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
1150 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
1151 if (PredicatesFoldable(LHSCC, RHSCC)) {
1152 if (LHS->getOperand(0) == RHS->getOperand(1) &&
1153 LHS->getOperand(1) == RHS->getOperand(0))
1154 LHS->swapOperands();
1155 if (LHS->getOperand(0) == RHS->getOperand(0) &&
1156 LHS->getOperand(1) == RHS->getOperand(1)) {
1157 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1158 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
1159 bool isSigned = LHS->isSigned() || RHS->isSigned();
1160 return getICmpValue(isSigned, Code, Op0, Op1, Builder);
1164 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
1165 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
1166 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
1167 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
1168 if (LHSCst == 0 || RHSCst == 0) return 0;
1170 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
1171 if (LHSCst == RHSCst && LHSCC == RHSCC &&
1172 LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
1173 Value *NewOr = Builder->CreateOr(Val, Val2);
1174 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
1177 // (icmp eq (A & C1), 0) | (icmp eq (A & C2), 0) -->
1178 // (icmp ne (A & (C1|C2)), (C1|C2)) where C1 and C2 are non-zero POT
1179 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
1180 Value *Op1 = 0, *Op2 = 0;
1181 ConstantInt *CI1 = 0, *CI2 = 0;
1182 if (match(LHS->getOperand(0), m_And(m_Value(Op1), m_ConstantInt(CI1))) &&
1183 match(RHS->getOperand(0), m_And(m_Value(Op2), m_ConstantInt(CI2)))) {
1184 if (Op1 == Op2 && !CI1->isZero() && !CI2->isZero() &&
1185 CI1->getValue().isPowerOf2() && CI2->getValue().isPowerOf2()) {
1186 Constant *ConstOr = ConstantExpr::getOr(CI1, CI2);
1187 Value *NewAnd = Builder->CreateAnd(Op1, ConstOr);
1188 return Builder->CreateICmp(ICmpInst::ICMP_NE, NewAnd, ConstOr);
1193 // From here on, we only handle:
1194 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
1195 if (Val != Val2) return 0;
1197 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
1198 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
1199 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
1200 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
1201 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
1204 // We can't fold (ugt x, C) | (sgt x, C2).
1205 if (!PredicatesFoldable(LHSCC, RHSCC))
1208 // Ensure that the larger constant is on the RHS.
1210 if (CmpInst::isSigned(LHSCC) ||
1211 (ICmpInst::isEquality(LHSCC) &&
1212 CmpInst::isSigned(RHSCC)))
1213 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
1215 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
1218 std::swap(LHS, RHS);
1219 std::swap(LHSCst, RHSCst);
1220 std::swap(LHSCC, RHSCC);
1223 // At this point, we know we have two icmp instructions
1224 // comparing a value against two constants and or'ing the result
1225 // together. Because of the above check, we know that we only have
1226 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
1227 // icmp folding check above), that the two constants are not
1229 assert(LHSCst != RHSCst && "Compares not folded above?");
1232 default: llvm_unreachable("Unknown integer condition code!");
1233 case ICmpInst::ICMP_EQ:
1235 default: llvm_unreachable("Unknown integer condition code!");
1236 case ICmpInst::ICMP_EQ:
1237 if (LHSCst == SubOne(RHSCst)) {
1238 // (X == 13 | X == 14) -> X-13 <u 2
1239 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1240 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1241 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
1242 return Builder->CreateICmpULT(Add, AddCST);
1244 break; // (X == 13 | X == 15) -> no change
1245 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
1246 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
1248 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
1249 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
1250 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
1254 case ICmpInst::ICMP_NE:
1256 default: llvm_unreachable("Unknown integer condition code!");
1257 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
1258 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
1259 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
1261 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
1262 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
1263 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
1264 return ConstantInt::getTrue(LHS->getContext());
1267 case ICmpInst::ICMP_ULT:
1269 default: llvm_unreachable("Unknown integer condition code!");
1270 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
1272 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
1273 // If RHSCst is [us]MAXINT, it is always false. Not handling
1274 // this can cause overflow.
1275 if (RHSCst->isMaxValue(false))
1277 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false);
1278 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
1280 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
1281 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
1283 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
1287 case ICmpInst::ICMP_SLT:
1289 default: llvm_unreachable("Unknown integer condition code!");
1290 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
1292 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
1293 // If RHSCst is [us]MAXINT, it is always false. Not handling
1294 // this can cause overflow.
1295 if (RHSCst->isMaxValue(true))
1297 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false);
1298 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
1300 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
1301 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
1303 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
1307 case ICmpInst::ICMP_UGT:
1309 default: llvm_unreachable("Unknown integer condition code!");
1310 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
1311 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
1313 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
1315 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
1316 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
1317 return ConstantInt::getTrue(LHS->getContext());
1318 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
1322 case ICmpInst::ICMP_SGT:
1324 default: llvm_unreachable("Unknown integer condition code!");
1325 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
1326 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
1328 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
1330 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
1331 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
1332 return ConstantInt::getTrue(LHS->getContext());
1333 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
1341 /// FoldOrOfFCmps - Optimize (fcmp)|(fcmp). NOTE: Unlike the rest of
1342 /// instcombine, this returns a Value which should already be inserted into the
1344 Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
1345 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
1346 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
1347 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
1348 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
1349 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
1350 // If either of the constants are nans, then the whole thing returns
1352 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
1353 return ConstantInt::getTrue(LHS->getContext());
1355 // Otherwise, no need to compare the two constants, compare the
1357 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
1360 // Handle vector zeros. This occurs because the canonical form of
1361 // "fcmp uno x,x" is "fcmp uno x, 0".
1362 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
1363 isa<ConstantAggregateZero>(RHS->getOperand(1)))
1364 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
1369 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
1370 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
1371 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
1373 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
1374 // Swap RHS operands to match LHS.
1375 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
1376 std::swap(Op1LHS, Op1RHS);
1378 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
1379 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
1381 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
1382 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
1383 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
1384 if (Op0CC == FCmpInst::FCMP_FALSE)
1386 if (Op1CC == FCmpInst::FCMP_FALSE)
1390 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
1391 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
1392 if (Op0Ordered == Op1Ordered) {
1393 // If both are ordered or unordered, return a new fcmp with
1394 // or'ed predicates.
1395 return getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS, Builder);
1401 /// FoldOrWithConstants - This helper function folds:
1403 /// ((A | B) & C1) | (B & C2)
1409 /// when the XOR of the two constants is "all ones" (-1).
1410 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
1411 Value *A, Value *B, Value *C) {
1412 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
1416 ConstantInt *CI2 = 0;
1417 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
1419 APInt Xor = CI1->getValue() ^ CI2->getValue();
1420 if (!Xor.isAllOnesValue()) return 0;
1422 if (V1 == A || V1 == B) {
1423 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
1424 return BinaryOperator::CreateOr(NewOp, V1);
1430 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
1431 bool Changed = SimplifyCommutative(I);
1432 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1434 if (Value *V = SimplifyOrInst(Op0, Op1, TD))
1435 return ReplaceInstUsesWith(I, V);
1437 // See if we can simplify any instructions used by the instruction whose sole
1438 // purpose is to compute bits we don't care about.
1439 if (SimplifyDemandedInstructionBits(I))
1442 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1443 ConstantInt *C1 = 0; Value *X = 0;
1444 // (X & C1) | C2 --> (X | C2) & (C1|C2)
1445 // iff (C1 & C2) == 0.
1446 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
1447 (RHS->getValue() & C1->getValue()) != 0 &&
1449 Value *Or = Builder->CreateOr(X, RHS);
1451 return BinaryOperator::CreateAnd(Or,
1452 ConstantInt::get(I.getContext(),
1453 RHS->getValue() | C1->getValue()));
1456 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
1457 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
1459 Value *Or = Builder->CreateOr(X, RHS);
1461 return BinaryOperator::CreateXor(Or,
1462 ConstantInt::get(I.getContext(),
1463 C1->getValue() & ~RHS->getValue()));
1466 // Try to fold constant and into select arguments.
1467 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1468 if (Instruction *R = FoldOpIntoSelect(I, SI))
1471 if (isa<PHINode>(Op0))
1472 if (Instruction *NV = FoldOpIntoPhi(I))
1476 Value *A = 0, *B = 0;
1477 ConstantInt *C1 = 0, *C2 = 0;
1479 // (A | B) | C and A | (B | C) -> bswap if possible.
1480 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
1481 if (match(Op0, m_Or(m_Value(), m_Value())) ||
1482 match(Op1, m_Or(m_Value(), m_Value())) ||
1483 (match(Op0, m_Shift(m_Value(), m_Value())) &&
1484 match(Op1, m_Shift(m_Value(), m_Value())))) {
1485 if (Instruction *BSwap = MatchBSwap(I))
1489 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
1490 if (Op0->hasOneUse() &&
1491 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
1492 MaskedValueIsZero(Op1, C1->getValue())) {
1493 Value *NOr = Builder->CreateOr(A, Op1);
1495 return BinaryOperator::CreateXor(NOr, C1);
1498 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
1499 if (Op1->hasOneUse() &&
1500 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
1501 MaskedValueIsZero(Op0, C1->getValue())) {
1502 Value *NOr = Builder->CreateOr(A, Op0);
1504 return BinaryOperator::CreateXor(NOr, C1);
1508 Value *C = 0, *D = 0;
1509 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1510 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1511 Value *V1 = 0, *V2 = 0, *V3 = 0;
1512 C1 = dyn_cast<ConstantInt>(C);
1513 C2 = dyn_cast<ConstantInt>(D);
1514 if (C1 && C2) { // (A & C1)|(B & C2)
1515 // If we have: ((V + N) & C1) | (V & C2)
1516 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1517 // replace with V+N.
1518 if (C1->getValue() == ~C2->getValue()) {
1519 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
1520 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1521 // Add commutes, try both ways.
1522 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
1523 return ReplaceInstUsesWith(I, A);
1524 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
1525 return ReplaceInstUsesWith(I, A);
1527 // Or commutes, try both ways.
1528 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
1529 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1530 // Add commutes, try both ways.
1531 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
1532 return ReplaceInstUsesWith(I, B);
1533 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
1534 return ReplaceInstUsesWith(I, B);
1538 if ((C1->getValue() & C2->getValue()) == 0) {
1539 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
1540 // iff (C1&C2) == 0 and (N&~C1) == 0
1541 if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
1542 ((V1 == B && MaskedValueIsZero(V2, ~C1->getValue())) || // (V|N)
1543 (V2 == B && MaskedValueIsZero(V1, ~C1->getValue())))) // (N|V)
1544 return BinaryOperator::CreateAnd(A,
1545 ConstantInt::get(A->getContext(),
1546 C1->getValue()|C2->getValue()));
1547 // Or commutes, try both ways.
1548 if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
1549 ((V1 == A && MaskedValueIsZero(V2, ~C2->getValue())) || // (V|N)
1550 (V2 == A && MaskedValueIsZero(V1, ~C2->getValue())))) // (N|V)
1551 return BinaryOperator::CreateAnd(B,
1552 ConstantInt::get(B->getContext(),
1553 C1->getValue()|C2->getValue()));
1555 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
1556 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
1557 ConstantInt *C3 = 0, *C4 = 0;
1558 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
1559 (C3->getValue() & ~C1->getValue()) == 0 &&
1560 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
1561 (C4->getValue() & ~C2->getValue()) == 0) {
1562 V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
1563 return BinaryOperator::CreateAnd(V2,
1564 ConstantInt::get(B->getContext(),
1565 C1->getValue()|C2->getValue()));
1570 // Check to see if we have any common things being and'ed. If so, find the
1571 // terms for V1 & (V2|V3).
1572 if (Op0->hasOneUse() || Op1->hasOneUse()) {
1574 if (A == B) // (A & C)|(A & D) == A & (C|D)
1575 V1 = A, V2 = C, V3 = D;
1576 else if (A == D) // (A & C)|(B & A) == A & (B|C)
1577 V1 = A, V2 = B, V3 = C;
1578 else if (C == B) // (A & C)|(C & D) == C & (A|D)
1579 V1 = C, V2 = A, V3 = D;
1580 else if (C == D) // (A & C)|(B & C) == C & (A|B)
1581 V1 = C, V2 = A, V3 = B;
1584 Value *Or = Builder->CreateOr(V2, V3, "tmp");
1585 return BinaryOperator::CreateAnd(V1, Or);
1589 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants.
1590 // Don't do this for vector select idioms, the code generator doesn't handle
1592 if (!I.getType()->isVectorTy()) {
1593 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
1595 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
1597 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
1599 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
1603 // ((A&~B)|(~A&B)) -> A^B
1604 if ((match(C, m_Not(m_Specific(D))) &&
1605 match(B, m_Not(m_Specific(A)))))
1606 return BinaryOperator::CreateXor(A, D);
1607 // ((~B&A)|(~A&B)) -> A^B
1608 if ((match(A, m_Not(m_Specific(D))) &&
1609 match(B, m_Not(m_Specific(C)))))
1610 return BinaryOperator::CreateXor(C, D);
1611 // ((A&~B)|(B&~A)) -> A^B
1612 if ((match(C, m_Not(m_Specific(B))) &&
1613 match(D, m_Not(m_Specific(A)))))
1614 return BinaryOperator::CreateXor(A, B);
1615 // ((~B&A)|(B&~A)) -> A^B
1616 if ((match(A, m_Not(m_Specific(B))) &&
1617 match(D, m_Not(m_Specific(C)))))
1618 return BinaryOperator::CreateXor(C, B);
1620 // ((A|B)&1)|(B&-2) -> (A&1) | B
1621 if (match(A, m_Or(m_Value(V1), m_Specific(B))) ||
1622 match(A, m_Or(m_Specific(B), m_Value(V1)))) {
1623 Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C);
1624 if (Ret) return Ret;
1626 // (B&-2)|((A|B)&1) -> (A&1) | B
1627 if (match(B, m_Or(m_Specific(A), m_Value(V1))) ||
1628 match(B, m_Or(m_Value(V1), m_Specific(A)))) {
1629 Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D);
1630 if (Ret) return Ret;
1634 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
1635 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
1636 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
1637 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
1638 SI0->getOperand(1) == SI1->getOperand(1) &&
1639 (SI0->hasOneUse() || SI1->hasOneUse())) {
1640 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
1642 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
1643 SI1->getOperand(1));
1647 // (~A | ~B) == (~(A & B)) - De Morgan's Law
1648 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1649 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1650 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1651 Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
1652 I.getName()+".demorgan");
1653 return BinaryOperator::CreateNot(And);
1656 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
1657 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
1658 if (Value *Res = FoldOrOfICmps(LHS, RHS))
1659 return ReplaceInstUsesWith(I, Res);
1661 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
1662 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1663 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1664 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
1665 return ReplaceInstUsesWith(I, Res);
1667 // fold (or (cast A), (cast B)) -> (cast (or A, B))
1668 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
1669 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
1670 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
1671 const Type *SrcTy = Op0C->getOperand(0)->getType();
1672 if (SrcTy == Op1C->getOperand(0)->getType() &&
1673 SrcTy->isIntOrIntVectorTy()) {
1674 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
1676 if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) &&
1677 // Only do this if the casts both really cause code to be
1679 ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
1680 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
1681 Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName());
1682 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
1685 // If this is or(cast(icmp), cast(icmp)), try to fold this even if the
1686 // cast is otherwise not optimizable. This happens for vector sexts.
1687 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
1688 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
1689 if (Value *Res = FoldOrOfICmps(LHS, RHS))
1690 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1692 // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the
1693 // cast is otherwise not optimizable. This happens for vector sexts.
1694 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
1695 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
1696 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
1697 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1702 return Changed ? &I : 0;
1705 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
1706 bool Changed = SimplifyCommutative(I);
1707 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1709 if (isa<UndefValue>(Op1)) {
1710 if (isa<UndefValue>(Op0))
1711 // Handle undef ^ undef -> 0 special case. This is a common
1713 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1714 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
1719 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1721 // See if we can simplify any instructions used by the instruction whose sole
1722 // purpose is to compute bits we don't care about.
1723 if (SimplifyDemandedInstructionBits(I))
1725 if (I.getType()->isVectorTy())
1726 if (isa<ConstantAggregateZero>(Op1))
1727 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
1729 // Is this a ~ operation?
1730 if (Value *NotOp = dyn_castNotVal(&I)) {
1731 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
1732 if (Op0I->getOpcode() == Instruction::And ||
1733 Op0I->getOpcode() == Instruction::Or) {
1734 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
1735 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
1736 if (dyn_castNotVal(Op0I->getOperand(1)))
1737 Op0I->swapOperands();
1738 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
1740 Builder->CreateNot(Op0I->getOperand(1),
1741 Op0I->getOperand(1)->getName()+".not");
1742 if (Op0I->getOpcode() == Instruction::And)
1743 return BinaryOperator::CreateOr(Op0NotVal, NotY);
1744 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
1747 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
1748 // ~(X | Y) === (~X & ~Y) - De Morgan's Law
1749 if (isFreeToInvert(Op0I->getOperand(0)) &&
1750 isFreeToInvert(Op0I->getOperand(1))) {
1752 Builder->CreateNot(Op0I->getOperand(0), "notlhs");
1754 Builder->CreateNot(Op0I->getOperand(1), "notrhs");
1755 if (Op0I->getOpcode() == Instruction::And)
1756 return BinaryOperator::CreateOr(NotX, NotY);
1757 return BinaryOperator::CreateAnd(NotX, NotY);
1760 } else if (Op0I->getOpcode() == Instruction::AShr) {
1761 // ~(~X >>s Y) --> (X >>s Y)
1762 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0)))
1763 return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1));
1769 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1770 if (RHS->isOne() && Op0->hasOneUse())
1771 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
1772 if (CmpInst *CI = dyn_cast<CmpInst>(Op0))
1773 return CmpInst::Create(CI->getOpcode(),
1774 CI->getInversePredicate(),
1775 CI->getOperand(0), CI->getOperand(1));
1777 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
1778 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
1779 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
1780 if (CI->hasOneUse() && Op0C->hasOneUse()) {
1781 Instruction::CastOps Opcode = Op0C->getOpcode();
1782 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
1783 (RHS == ConstantExpr::getCast(Opcode,
1784 ConstantInt::getTrue(I.getContext()),
1785 Op0C->getDestTy()))) {
1786 CI->setPredicate(CI->getInversePredicate());
1787 return CastInst::Create(Opcode, CI, Op0C->getType());
1793 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1794 // ~(c-X) == X-c-1 == X+(-c-1)
1795 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
1796 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
1797 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
1798 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
1799 ConstantInt::get(I.getType(), 1));
1800 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
1803 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
1804 if (Op0I->getOpcode() == Instruction::Add) {
1805 // ~(X-c) --> (-c-1)-X
1806 if (RHS->isAllOnesValue()) {
1807 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
1808 return BinaryOperator::CreateSub(
1809 ConstantExpr::getSub(NegOp0CI,
1810 ConstantInt::get(I.getType(), 1)),
1811 Op0I->getOperand(0));
1812 } else if (RHS->getValue().isSignBit()) {
1813 // (X + C) ^ signbit -> (X + C + signbit)
1814 Constant *C = ConstantInt::get(I.getContext(),
1815 RHS->getValue() + Op0CI->getValue());
1816 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
1819 } else if (Op0I->getOpcode() == Instruction::Or) {
1820 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
1821 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
1822 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
1823 // Anything in both C1 and C2 is known to be zero, remove it from
1825 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
1826 NewRHS = ConstantExpr::getAnd(NewRHS,
1827 ConstantExpr::getNot(CommonBits));
1829 I.setOperand(0, Op0I->getOperand(0));
1830 I.setOperand(1, NewRHS);
1837 // Try to fold constant and into select arguments.
1838 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1839 if (Instruction *R = FoldOpIntoSelect(I, SI))
1841 if (isa<PHINode>(Op0))
1842 if (Instruction *NV = FoldOpIntoPhi(I))
1846 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
1848 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
1850 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
1852 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
1855 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
1858 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
1859 if (A == Op0) { // B^(B|A) == (A|B)^B
1860 Op1I->swapOperands();
1862 std::swap(Op0, Op1);
1863 } else if (B == Op0) { // B^(A|B) == (A|B)^B
1864 I.swapOperands(); // Simplified below.
1865 std::swap(Op0, Op1);
1867 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
1868 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
1869 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
1870 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
1871 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
1873 if (A == Op0) { // A^(A&B) -> A^(B&A)
1874 Op1I->swapOperands();
1877 if (B == Op0) { // A^(B&A) -> (B&A)^A
1878 I.swapOperands(); // Simplified below.
1879 std::swap(Op0, Op1);
1884 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
1887 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
1888 Op0I->hasOneUse()) {
1889 if (A == Op1) // (B|A)^B == (A|B)^B
1891 if (B == Op1) // (A|B)^B == A & ~B
1892 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1, "tmp"));
1893 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
1894 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
1895 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
1896 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
1897 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
1899 if (A == Op1) // (A&B)^A -> (B&A)^A
1901 if (B == Op1 && // (B&A)^A == ~B & A
1902 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
1903 return BinaryOperator::CreateAnd(Builder->CreateNot(A, "tmp"), Op1);
1908 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
1909 if (Op0I && Op1I && Op0I->isShift() &&
1910 Op0I->getOpcode() == Op1I->getOpcode() &&
1911 Op0I->getOperand(1) == Op1I->getOperand(1) &&
1912 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
1914 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
1916 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
1917 Op1I->getOperand(1));
1921 Value *A, *B, *C, *D;
1922 // (A & B)^(A | B) -> A ^ B
1923 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
1924 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
1925 if ((A == C && B == D) || (A == D && B == C))
1926 return BinaryOperator::CreateXor(A, B);
1928 // (A | B)^(A & B) -> A ^ B
1929 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
1930 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
1931 if ((A == C && B == D) || (A == D && B == C))
1932 return BinaryOperator::CreateXor(A, B);
1936 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
1937 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
1938 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
1939 // (X & Y)^(X & Y) -> (Y^Z) & X
1940 Value *X = 0, *Y = 0, *Z = 0;
1942 X = A, Y = B, Z = D;
1944 X = A, Y = B, Z = C;
1946 X = B, Y = A, Z = D;
1948 X = B, Y = A, Z = C;
1951 Value *NewOp = Builder->CreateXor(Y, Z, Op0->getName());
1952 return BinaryOperator::CreateAnd(NewOp, X);
1957 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
1958 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
1959 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
1960 if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
1961 if (LHS->getOperand(0) == RHS->getOperand(1) &&
1962 LHS->getOperand(1) == RHS->getOperand(0))
1963 LHS->swapOperands();
1964 if (LHS->getOperand(0) == RHS->getOperand(0) &&
1965 LHS->getOperand(1) == RHS->getOperand(1)) {
1966 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1967 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
1968 bool isSigned = LHS->isSigned() || RHS->isSigned();
1969 return ReplaceInstUsesWith(I,
1970 getICmpValue(isSigned, Code, Op0, Op1, Builder));
1974 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
1975 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
1976 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
1977 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
1978 const Type *SrcTy = Op0C->getOperand(0)->getType();
1979 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() &&
1980 // Only do this if the casts both really cause code to be generated.
1981 ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0),
1983 ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0),
1985 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
1986 Op1C->getOperand(0), I.getName());
1987 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
1992 return Changed ? &I : 0;