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 = cast<ConstantInt>(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 = cast<ConstantInt>(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 = ConstantInt::get(AndRHS->getContext(),
213 AndRHS->getValue() & ShlMask);
215 if (CI->getValue() == ShlMask)
216 // Masking out bits that the shift already masks.
217 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
219 if (CI != AndRHS) { // Reducing bits set in and.
220 TheAnd.setOperand(1, CI);
225 case Instruction::LShr: {
226 // We know that the AND will not produce any of the bits shifted in, so if
227 // the anded constant includes them, clear them now! This only applies to
228 // unsigned shifts, because a signed shr may bring in set bits!
230 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
231 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
232 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
233 ConstantInt *CI = ConstantInt::get(Op->getContext(),
234 AndRHS->getValue() & ShrMask);
236 if (CI->getValue() == ShrMask)
237 // Masking out bits that the shift already masks.
238 return ReplaceInstUsesWith(TheAnd, Op);
241 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
246 case Instruction::AShr:
248 // See if this is shifting in some sign extension, then masking it out
250 if (Op->hasOneUse()) {
251 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
252 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
253 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
254 Constant *C = ConstantInt::get(Op->getContext(),
255 AndRHS->getValue() & ShrMask);
256 if (C == AndRHS) { // Masking out bits shifted in.
257 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
258 // Make the argument unsigned.
259 Value *ShVal = Op->getOperand(0);
260 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
261 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
269 /// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise
270 /// (V < Lo || V >= Hi). In practice, we emit the more efficient
271 /// (V-Lo) \<u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
272 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
273 /// insert new instructions.
274 Value *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
275 bool isSigned, bool Inside) {
276 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
277 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
278 "Lo is not <= Hi in range emission code!");
281 if (Lo == Hi) // Trivially false.
282 return ConstantInt::getFalse(V->getContext());
284 // V >= Min && V < Hi --> V < Hi
285 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
286 ICmpInst::Predicate pred = (isSigned ?
287 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
288 return Builder->CreateICmp(pred, V, Hi);
291 // Emit V-Lo <u Hi-Lo
292 Constant *NegLo = ConstantExpr::getNeg(Lo);
293 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
294 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
295 return Builder->CreateICmpULT(Add, UpperBound);
298 if (Lo == Hi) // Trivially true.
299 return ConstantInt::getTrue(V->getContext());
301 // V < Min || V >= Hi -> V > Hi-1
302 Hi = SubOne(cast<ConstantInt>(Hi));
303 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
304 ICmpInst::Predicate pred = (isSigned ?
305 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
306 return Builder->CreateICmp(pred, V, Hi);
309 // Emit V-Lo >u Hi-1-Lo
310 // Note that Hi has already had one subtracted from it, above.
311 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
312 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
313 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
314 return Builder->CreateICmpUGT(Add, LowerBound);
317 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
318 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
319 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
320 // not, since all 1s are not contiguous.
321 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
322 const APInt& V = Val->getValue();
323 uint32_t BitWidth = Val->getType()->getBitWidth();
324 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
326 // look for the first zero bit after the run of ones
327 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
328 // look for the first non-zero bit
329 ME = V.getActiveBits();
333 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
334 /// where isSub determines whether the operator is a sub. If we can fold one of
335 /// the following xforms:
337 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
338 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
339 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
341 /// return (A +/- B).
343 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
344 ConstantInt *Mask, bool isSub,
346 Instruction *LHSI = dyn_cast<Instruction>(LHS);
347 if (!LHSI || LHSI->getNumOperands() != 2 ||
348 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
350 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
352 switch (LHSI->getOpcode()) {
354 case Instruction::And:
355 if (ConstantExpr::getAnd(N, Mask) == Mask) {
356 // If the AndRHS is a power of two minus one (0+1+), this is simple.
357 if ((Mask->getValue().countLeadingZeros() +
358 Mask->getValue().countPopulation()) ==
359 Mask->getValue().getBitWidth())
362 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
363 // part, we don't need any explicit masks to take them out of A. If that
364 // is all N is, ignore it.
365 uint32_t MB = 0, ME = 0;
366 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
367 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
368 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
369 if (MaskedValueIsZero(RHS, Mask))
374 case Instruction::Or:
375 case Instruction::Xor:
376 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
377 if ((Mask->getValue().countLeadingZeros() +
378 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
379 && ConstantExpr::getAnd(N, Mask)->isNullValue())
385 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
386 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
389 /// enum for classifying (icmp eq (A & B), C) and (icmp ne (A & B), C)
390 /// One of A and B is considered the mask, the other the value. This is
391 /// described as the "AMask" or "BMask" part of the enum. If the enum
392 /// contains only "Mask", then both A and B can be considered masks.
393 /// If A is the mask, then it was proven, that (A & C) == C. This
394 /// is trivial if C == A, or C == 0. If both A and C are constants, this
395 /// proof is also easy.
396 /// For the following explanations we assume that A is the mask.
397 /// The part "AllOnes" declares, that the comparison is true only
398 /// if (A & B) == A, or all bits of A are set in B.
399 /// Example: (icmp eq (A & 3), 3) -> FoldMskICmp_AMask_AllOnes
400 /// The part "AllZeroes" declares, that the comparison is true only
401 /// if (A & B) == 0, or all bits of A are cleared in B.
402 /// Example: (icmp eq (A & 3), 0) -> FoldMskICmp_Mask_AllZeroes
403 /// The part "Mixed" declares, that (A & B) == C and C might or might not
404 /// contain any number of one bits and zero bits.
405 /// Example: (icmp eq (A & 3), 1) -> FoldMskICmp_AMask_Mixed
406 /// The Part "Not" means, that in above descriptions "==" should be replaced
408 /// Example: (icmp ne (A & 3), 3) -> FoldMskICmp_AMask_NotAllOnes
409 /// If the mask A contains a single bit, then the following is equivalent:
410 /// (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
411 /// (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
412 enum MaskedICmpType {
413 FoldMskICmp_AMask_AllOnes = 1,
414 FoldMskICmp_AMask_NotAllOnes = 2,
415 FoldMskICmp_BMask_AllOnes = 4,
416 FoldMskICmp_BMask_NotAllOnes = 8,
417 FoldMskICmp_Mask_AllZeroes = 16,
418 FoldMskICmp_Mask_NotAllZeroes = 32,
419 FoldMskICmp_AMask_Mixed = 64,
420 FoldMskICmp_AMask_NotMixed = 128,
421 FoldMskICmp_BMask_Mixed = 256,
422 FoldMskICmp_BMask_NotMixed = 512
425 /// return the set of pattern classes (from MaskedICmpType)
426 /// that (icmp SCC (A & B), C) satisfies
427 static unsigned getTypeOfMaskedICmp(Value* A, Value* B, Value* C,
428 ICmpInst::Predicate SCC)
430 ConstantInt *ACst = dyn_cast<ConstantInt>(A);
431 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
432 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
433 bool icmp_eq = (SCC == ICmpInst::ICMP_EQ);
434 bool icmp_abit = (ACst != 0 && !ACst->isZero() &&
435 ACst->getValue().isPowerOf2());
436 bool icmp_bbit = (BCst != 0 && !BCst->isZero() &&
437 BCst->getValue().isPowerOf2());
439 if (CCst != 0 && CCst->isZero()) {
440 // if C is zero, then both A and B qualify as mask
441 result |= (icmp_eq ? (FoldMskICmp_Mask_AllZeroes |
442 FoldMskICmp_Mask_AllZeroes |
443 FoldMskICmp_AMask_Mixed |
444 FoldMskICmp_BMask_Mixed)
445 : (FoldMskICmp_Mask_NotAllZeroes |
446 FoldMskICmp_Mask_NotAllZeroes |
447 FoldMskICmp_AMask_NotMixed |
448 FoldMskICmp_BMask_NotMixed));
450 result |= (icmp_eq ? (FoldMskICmp_AMask_NotAllOnes |
451 FoldMskICmp_AMask_NotMixed)
452 : (FoldMskICmp_AMask_AllOnes |
453 FoldMskICmp_AMask_Mixed));
455 result |= (icmp_eq ? (FoldMskICmp_BMask_NotAllOnes |
456 FoldMskICmp_BMask_NotMixed)
457 : (FoldMskICmp_BMask_AllOnes |
458 FoldMskICmp_BMask_Mixed));
462 result |= (icmp_eq ? (FoldMskICmp_AMask_AllOnes |
463 FoldMskICmp_AMask_Mixed)
464 : (FoldMskICmp_AMask_NotAllOnes |
465 FoldMskICmp_AMask_NotMixed));
467 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
468 FoldMskICmp_AMask_NotMixed)
469 : (FoldMskICmp_Mask_AllZeroes |
470 FoldMskICmp_AMask_Mixed));
471 } else if (ACst != 0 && CCst != 0 &&
472 ConstantExpr::getAnd(ACst, CCst) == CCst) {
473 result |= (icmp_eq ? FoldMskICmp_AMask_Mixed
474 : FoldMskICmp_AMask_NotMixed);
477 result |= (icmp_eq ? (FoldMskICmp_BMask_AllOnes |
478 FoldMskICmp_BMask_Mixed)
479 : (FoldMskICmp_BMask_NotAllOnes |
480 FoldMskICmp_BMask_NotMixed));
482 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
483 FoldMskICmp_BMask_NotMixed)
484 : (FoldMskICmp_Mask_AllZeroes |
485 FoldMskICmp_BMask_Mixed));
486 } else if (BCst != 0 && CCst != 0 &&
487 ConstantExpr::getAnd(BCst, CCst) == CCst) {
488 result |= (icmp_eq ? FoldMskICmp_BMask_Mixed
489 : FoldMskICmp_BMask_NotMixed);
494 /// decomposeBitTestICmp - Decompose an icmp into the form ((X & Y) pred Z)
495 /// if possible. The returned predicate is either == or !=. Returns false if
496 /// decomposition fails.
497 static bool decomposeBitTestICmp(const ICmpInst *I, ICmpInst::Predicate &Pred,
498 Value *&X, Value *&Y, Value *&Z) {
499 // X < 0 is equivalent to (X & SignBit) != 0.
500 if (I->getPredicate() == ICmpInst::ICMP_SLT)
501 if (ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1)))
503 X = I->getOperand(0);
504 Y = ConstantInt::get(I->getContext(),
505 APInt::getSignBit(C->getBitWidth()));
506 Pred = ICmpInst::ICMP_NE;
511 // X > -1 is equivalent to (X & SignBit) == 0.
512 if (I->getPredicate() == ICmpInst::ICMP_SGT)
513 if (ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1)))
514 if (C->isAllOnesValue()) {
515 X = I->getOperand(0);
516 Y = ConstantInt::get(I->getContext(),
517 APInt::getSignBit(C->getBitWidth()));
518 Pred = ICmpInst::ICMP_EQ;
519 Z = ConstantInt::getNullValue(C->getType());
526 /// foldLogOpOfMaskedICmpsHelper:
527 /// handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
528 /// return the set of pattern classes (from MaskedICmpType)
529 /// that both LHS and RHS satisfy
530 static unsigned foldLogOpOfMaskedICmpsHelper(Value*& A,
531 Value*& B, Value*& C,
532 Value*& D, Value*& E,
533 ICmpInst *LHS, ICmpInst *RHS,
534 ICmpInst::Predicate &LHSCC,
535 ICmpInst::Predicate &RHSCC) {
536 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) return 0;
537 // vectors are not (yet?) supported
538 if (LHS->getOperand(0)->getType()->isVectorTy()) return 0;
540 // Here comes the tricky part:
541 // LHS might be of the form L11 & L12 == X, X == L21 & L22,
542 // and L11 & L12 == L21 & L22. The same goes for RHS.
543 // Now we must find those components L** and R**, that are equal, so
544 // that we can extract the parameters A, B, C, D, and E for the canonical
546 Value *L1 = LHS->getOperand(0);
547 Value *L2 = LHS->getOperand(1);
548 Value *L11,*L12,*L21,*L22;
549 // Check whether the icmp can be decomposed into a bit test.
550 if (decomposeBitTestICmp(LHS, LHSCC, L11, L12, L2)) {
553 // Look for ANDs in the LHS icmp.
554 if (match(L1, m_And(m_Value(L11), m_Value(L12)))) {
555 if (!match(L2, m_And(m_Value(L21), m_Value(L22))))
558 if (!match(L2, m_And(m_Value(L11), m_Value(L12))))
565 // Bail if LHS was a icmp that can't be decomposed into an equality.
566 if (!ICmpInst::isEquality(LHSCC))
569 Value *R1 = RHS->getOperand(0);
570 Value *R2 = RHS->getOperand(1);
573 if (decomposeBitTestICmp(RHS, RHSCC, R11, R12, R2)) {
574 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
576 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
581 E = R2; R1 = 0; ok = true;
582 } else if (match(R1, m_And(m_Value(R11), m_Value(R12)))) {
583 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
584 A = R11; D = R12; E = R2; ok = true;
585 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
586 A = R12; D = R11; E = R2; ok = true;
590 // Bail if RHS was a icmp that can't be decomposed into an equality.
591 if (!ICmpInst::isEquality(RHSCC))
594 // Look for ANDs in on the right side of the RHS icmp.
595 if (!ok && match(R2, m_And(m_Value(R11), m_Value(R12)))) {
596 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
597 A = R11; D = R12; E = R1; ok = true;
598 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
599 A = R12; D = R11; E = R1; ok = true;
609 } else if (L12 == A) {
611 } else if (L21 == A) {
613 } else if (L22 == A) {
617 unsigned left_type = getTypeOfMaskedICmp(A, B, C, LHSCC);
618 unsigned right_type = getTypeOfMaskedICmp(A, D, E, RHSCC);
619 return left_type & right_type;
621 /// foldLogOpOfMaskedICmps:
622 /// try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
623 /// into a single (icmp(A & X) ==/!= Y)
624 static Value* foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS,
625 ICmpInst::Predicate NEWCC,
626 llvm::InstCombiner::BuilderTy* Builder) {
627 Value *A = 0, *B = 0, *C = 0, *D = 0, *E = 0;
628 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
629 unsigned mask = foldLogOpOfMaskedICmpsHelper(A, B, C, D, E, LHS, RHS,
631 if (mask == 0) return 0;
632 assert(ICmpInst::isEquality(LHSCC) && ICmpInst::isEquality(RHSCC) &&
633 "foldLogOpOfMaskedICmpsHelper must return an equality predicate.");
635 if (NEWCC == ICmpInst::ICMP_NE)
636 mask >>= 1; // treat "Not"-states as normal states
638 if (mask & FoldMskICmp_Mask_AllZeroes) {
639 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
640 // -> (icmp eq (A & (B|D)), 0)
641 Value* newOr = Builder->CreateOr(B, D);
642 Value* newAnd = Builder->CreateAnd(A, newOr);
643 // we can't use C as zero, because we might actually handle
644 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
645 // with B and D, having a single bit set
646 Value* zero = Constant::getNullValue(A->getType());
647 return Builder->CreateICmp(NEWCC, newAnd, zero);
649 if (mask & FoldMskICmp_BMask_AllOnes) {
650 // (icmp eq (A & B), B) & (icmp eq (A & D), D)
651 // -> (icmp eq (A & (B|D)), (B|D))
652 Value* newOr = Builder->CreateOr(B, D);
653 Value* newAnd = Builder->CreateAnd(A, newOr);
654 return Builder->CreateICmp(NEWCC, newAnd, newOr);
656 if (mask & FoldMskICmp_AMask_AllOnes) {
657 // (icmp eq (A & B), A) & (icmp eq (A & D), A)
658 // -> (icmp eq (A & (B&D)), A)
659 Value* newAnd1 = Builder->CreateAnd(B, D);
660 Value* newAnd = Builder->CreateAnd(A, newAnd1);
661 return Builder->CreateICmp(NEWCC, newAnd, A);
663 if (mask & FoldMskICmp_BMask_Mixed) {
664 // (icmp eq (A & B), C) & (icmp eq (A & D), E)
665 // We already know that B & C == C && D & E == E.
666 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
667 // C and E, which are shared by both the mask B and the mask D, don't
668 // contradict, then we can transform to
669 // -> (icmp eq (A & (B|D)), (C|E))
670 // Currently, we only handle the case of B, C, D, and E being constant.
671 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
672 if (BCst == 0) return 0;
673 ConstantInt *DCst = dyn_cast<ConstantInt>(D);
674 if (DCst == 0) return 0;
675 // we can't simply use C and E, because we might actually handle
676 // (icmp ne (A & B), B) & (icmp eq (A & D), D)
677 // with B and D, having a single bit set
679 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
680 if (CCst == 0) return 0;
682 CCst = dyn_cast<ConstantInt>( ConstantExpr::getXor(BCst, CCst) );
683 ConstantInt *ECst = dyn_cast<ConstantInt>(E);
684 if (ECst == 0) return 0;
686 ECst = dyn_cast<ConstantInt>( ConstantExpr::getXor(DCst, ECst) );
687 ConstantInt* MCst = dyn_cast<ConstantInt>(
688 ConstantExpr::getAnd(ConstantExpr::getAnd(BCst, DCst),
689 ConstantExpr::getXor(CCst, ECst)) );
690 // if there is a conflict we should actually return a false for the
694 Value *newOr1 = Builder->CreateOr(B, D);
695 Value *newOr2 = ConstantExpr::getOr(CCst, ECst);
696 Value *newAnd = Builder->CreateAnd(A, newOr1);
697 return Builder->CreateICmp(NEWCC, newAnd, newOr2);
702 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
703 Value *InstCombiner::FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
704 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
706 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
707 if (PredicatesFoldable(LHSCC, RHSCC)) {
708 if (LHS->getOperand(0) == RHS->getOperand(1) &&
709 LHS->getOperand(1) == RHS->getOperand(0))
711 if (LHS->getOperand(0) == RHS->getOperand(0) &&
712 LHS->getOperand(1) == RHS->getOperand(1)) {
713 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
714 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
715 bool isSigned = LHS->isSigned() || RHS->isSigned();
716 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
720 // handle (roughly): (icmp eq (A & B), C) & (icmp eq (A & D), E)
721 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, ICmpInst::ICMP_EQ, Builder))
724 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
725 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
726 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
727 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
728 if (LHSCst == 0 || RHSCst == 0) return 0;
730 if (LHSCst == RHSCst && LHSCC == RHSCC) {
731 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
732 // where C is a power of 2
733 if (LHSCC == ICmpInst::ICMP_ULT &&
734 LHSCst->getValue().isPowerOf2()) {
735 Value *NewOr = Builder->CreateOr(Val, Val2);
736 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
739 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
740 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
741 Value *NewOr = Builder->CreateOr(Val, Val2);
742 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
746 // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2
747 // where CMAX is the all ones value for the truncated type,
748 // iff the lower bits of C2 and CA are zero.
749 if (LHSCC == ICmpInst::ICMP_EQ && LHSCC == RHSCC &&
750 LHS->hasOneUse() && RHS->hasOneUse()) {
752 ConstantInt *AndCst, *SmallCst = 0, *BigCst = 0;
754 // (trunc x) == C1 & (and x, CA) == C2
755 // (and x, CA) == C2 & (trunc x) == C1
756 if (match(Val2, m_Trunc(m_Value(V))) &&
757 match(Val, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
760 } else if (match(Val, m_Trunc(m_Value(V))) &&
761 match(Val2, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
766 if (SmallCst && BigCst) {
767 unsigned BigBitSize = BigCst->getType()->getBitWidth();
768 unsigned SmallBitSize = SmallCst->getType()->getBitWidth();
770 // Check that the low bits are zero.
771 APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize);
772 if ((Low & AndCst->getValue()) == 0 && (Low & BigCst->getValue()) == 0) {
773 Value *NewAnd = Builder->CreateAnd(V, Low | AndCst->getValue());
774 APInt N = SmallCst->getValue().zext(BigBitSize) | BigCst->getValue();
775 Value *NewVal = ConstantInt::get(AndCst->getType()->getContext(), N);
776 return Builder->CreateICmp(LHSCC, NewAnd, NewVal);
781 // From here on, we only handle:
782 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
783 if (Val != Val2) return 0;
785 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
786 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
787 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
788 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
789 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
792 // Make a constant range that's the intersection of the two icmp ranges.
793 // If the intersection is empty, we know that the result is false.
794 ConstantRange LHSRange =
795 ConstantRange::makeICmpRegion(LHSCC, LHSCst->getValue());
796 ConstantRange RHSRange =
797 ConstantRange::makeICmpRegion(RHSCC, RHSCst->getValue());
799 if (LHSRange.intersectWith(RHSRange).isEmptySet())
800 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
802 // We can't fold (ugt x, C) & (sgt x, C2).
803 if (!PredicatesFoldable(LHSCC, RHSCC))
806 // Ensure that the larger constant is on the RHS.
808 if (CmpInst::isSigned(LHSCC) ||
809 (ICmpInst::isEquality(LHSCC) &&
810 CmpInst::isSigned(RHSCC)))
811 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
813 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
817 std::swap(LHSCst, RHSCst);
818 std::swap(LHSCC, RHSCC);
821 // At this point, we know we have two icmp instructions
822 // comparing a value against two constants and and'ing the result
823 // together. Because of the above check, we know that we only have
824 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
825 // (from the icmp folding check above), that the two constants
826 // are not equal and that the larger constant is on the RHS
827 assert(LHSCst != RHSCst && "Compares not folded above?");
830 default: llvm_unreachable("Unknown integer condition code!");
831 case ICmpInst::ICMP_EQ:
833 default: llvm_unreachable("Unknown integer condition code!");
834 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
835 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
836 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
839 case ICmpInst::ICMP_NE:
841 default: llvm_unreachable("Unknown integer condition code!");
842 case ICmpInst::ICMP_ULT:
843 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
844 return Builder->CreateICmpULT(Val, LHSCst);
845 break; // (X != 13 & X u< 15) -> no change
846 case ICmpInst::ICMP_SLT:
847 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
848 return Builder->CreateICmpSLT(Val, LHSCst);
849 break; // (X != 13 & X s< 15) -> no change
850 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
851 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
852 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
854 case ICmpInst::ICMP_NE:
855 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
856 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
857 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
858 return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1));
860 break; // (X != 13 & X != 15) -> no change
863 case ICmpInst::ICMP_ULT:
865 default: llvm_unreachable("Unknown integer condition code!");
866 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
867 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
868 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
869 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
871 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
872 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
874 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
878 case ICmpInst::ICMP_SLT:
880 default: llvm_unreachable("Unknown integer condition code!");
881 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
883 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
884 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
886 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
890 case ICmpInst::ICMP_UGT:
892 default: llvm_unreachable("Unknown integer condition code!");
893 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
894 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
896 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
898 case ICmpInst::ICMP_NE:
899 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
900 return Builder->CreateICmp(LHSCC, Val, RHSCst);
901 break; // (X u> 13 & X != 15) -> no change
902 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
903 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
904 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
908 case ICmpInst::ICMP_SGT:
910 default: llvm_unreachable("Unknown integer condition code!");
911 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
912 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
914 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
916 case ICmpInst::ICMP_NE:
917 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
918 return Builder->CreateICmp(LHSCC, Val, RHSCst);
919 break; // (X s> 13 & X != 15) -> no change
920 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
921 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true);
922 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
931 /// FoldAndOfFCmps - Optimize (fcmp)&(fcmp). NOTE: Unlike the rest of
932 /// instcombine, this returns a Value which should already be inserted into the
934 Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
935 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
936 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
937 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
938 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
939 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
940 // If either of the constants are nans, then the whole thing returns
942 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
943 return ConstantInt::getFalse(LHS->getContext());
944 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
947 // Handle vector zeros. This occurs because the canonical form of
948 // "fcmp ord x,x" is "fcmp ord x, 0".
949 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
950 isa<ConstantAggregateZero>(RHS->getOperand(1)))
951 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
955 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
956 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
957 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
960 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
961 // Swap RHS operands to match LHS.
962 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
963 std::swap(Op1LHS, Op1RHS);
966 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
967 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
969 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
970 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
971 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
972 if (Op0CC == FCmpInst::FCMP_TRUE)
974 if (Op1CC == FCmpInst::FCMP_TRUE)
979 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
980 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
981 // uno && ord -> false
982 if (Op0Pred == 0 && Op1Pred == 0 && Op0Ordered != Op1Ordered)
983 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
986 std::swap(Op0Pred, Op1Pred);
987 std::swap(Op0Ordered, Op1Ordered);
990 // uno && ueq -> uno && (uno || eq) -> uno
991 // ord && olt -> ord && (ord && lt) -> olt
992 if (!Op0Ordered && (Op0Ordered == Op1Ordered))
994 if (Op0Ordered && (Op0Ordered == Op1Ordered))
997 // uno && oeq -> uno && (ord && eq) -> false
999 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
1000 // ord && ueq -> ord && (uno || eq) -> oeq
1001 return getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS, Builder);
1009 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
1010 bool Changed = SimplifyAssociativeOrCommutative(I);
1011 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1013 if (Value *V = SimplifyAndInst(Op0, Op1, TD))
1014 return ReplaceInstUsesWith(I, V);
1016 // (A|B)&(A|C) -> A|(B&C) etc
1017 if (Value *V = SimplifyUsingDistributiveLaws(I))
1018 return ReplaceInstUsesWith(I, V);
1020 // See if we can simplify any instructions used by the instruction whose sole
1021 // purpose is to compute bits we don't care about.
1022 if (SimplifyDemandedInstructionBits(I))
1025 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
1026 const APInt &AndRHSMask = AndRHS->getValue();
1028 // Optimize a variety of ((val OP C1) & C2) combinations...
1029 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1030 Value *Op0LHS = Op0I->getOperand(0);
1031 Value *Op0RHS = Op0I->getOperand(1);
1032 switch (Op0I->getOpcode()) {
1034 case Instruction::Xor:
1035 case Instruction::Or: {
1036 // If the mask is only needed on one incoming arm, push it up.
1037 if (!Op0I->hasOneUse()) break;
1039 APInt NotAndRHS(~AndRHSMask);
1040 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
1041 // Not masking anything out for the LHS, move to RHS.
1042 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
1043 Op0RHS->getName()+".masked");
1044 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
1046 if (!isa<Constant>(Op0RHS) &&
1047 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
1048 // Not masking anything out for the RHS, move to LHS.
1049 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
1050 Op0LHS->getName()+".masked");
1051 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
1056 case Instruction::Add:
1057 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
1058 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1059 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1060 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
1061 return BinaryOperator::CreateAnd(V, AndRHS);
1062 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
1063 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
1066 case Instruction::Sub:
1067 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
1068 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1069 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1070 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
1071 return BinaryOperator::CreateAnd(V, AndRHS);
1073 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
1074 // has 1's for all bits that the subtraction with A might affect.
1075 if (Op0I->hasOneUse() && !match(Op0LHS, m_Zero())) {
1076 uint32_t BitWidth = AndRHSMask.getBitWidth();
1077 uint32_t Zeros = AndRHSMask.countLeadingZeros();
1078 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
1080 if (MaskedValueIsZero(Op0LHS, Mask)) {
1081 Value *NewNeg = Builder->CreateNeg(Op0RHS);
1082 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
1087 case Instruction::Shl:
1088 case Instruction::LShr:
1089 // (1 << x) & 1 --> zext(x == 0)
1090 // (1 >> x) & 1 --> zext(x == 0)
1091 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
1093 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
1094 return new ZExtInst(NewICmp, I.getType());
1099 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
1100 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
1104 // If this is an integer truncation, and if the source is an 'and' with
1105 // immediate, transform it. This frequently occurs for bitfield accesses.
1107 Value *X = 0; ConstantInt *YC = 0;
1108 if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) {
1109 // Change: and (trunc (and X, YC) to T), C2
1110 // into : and (trunc X to T), trunc(YC) & C2
1111 // This will fold the two constants together, which may allow
1112 // other simplifications.
1113 Value *NewCast = Builder->CreateTrunc(X, I.getType(), "and.shrunk");
1114 Constant *C3 = ConstantExpr::getTrunc(YC, I.getType());
1115 C3 = ConstantExpr::getAnd(C3, AndRHS);
1116 return BinaryOperator::CreateAnd(NewCast, C3);
1120 // Try to fold constant and into select arguments.
1121 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1122 if (Instruction *R = FoldOpIntoSelect(I, SI))
1124 if (isa<PHINode>(Op0))
1125 if (Instruction *NV = FoldOpIntoPhi(I))
1130 // (~A & ~B) == (~(A | B)) - De Morgan's Law
1131 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1132 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1133 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1134 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
1135 I.getName()+".demorgan");
1136 return BinaryOperator::CreateNot(Or);
1140 Value *A = 0, *B = 0, *C = 0, *D = 0;
1141 // (A|B) & ~(A&B) -> A^B
1142 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1143 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1144 ((A == C && B == D) || (A == D && B == C)))
1145 return BinaryOperator::CreateXor(A, B);
1147 // ~(A&B) & (A|B) -> A^B
1148 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1149 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1150 ((A == C && B == D) || (A == D && B == C)))
1151 return BinaryOperator::CreateXor(A, B);
1153 // A&(A^B) => A & ~B
1155 Value *tmpOp0 = Op0;
1156 Value *tmpOp1 = Op1;
1157 if (Op0->hasOneUse() &&
1158 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
1159 if (A == Op1 || B == Op1 ) {
1166 if (tmpOp1->hasOneUse() &&
1167 match(tmpOp1, m_Xor(m_Value(A), m_Value(B)))) {
1171 // Notice that the patten (A&(~B)) is actually (A&(-1^B)), so if
1172 // A is originally -1 (or a vector of -1 and undefs), then we enter
1173 // an endless loop. By checking that A is non-constant we ensure that
1174 // we will never get to the loop.
1175 if (A == tmpOp0 && !isa<Constant>(A)) // A&(A^B) -> A & ~B
1176 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B));
1180 // (A&((~A)|B)) -> A&B
1181 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
1182 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
1183 return BinaryOperator::CreateAnd(A, Op1);
1184 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
1185 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
1186 return BinaryOperator::CreateAnd(A, Op0);
1189 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1))
1190 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
1191 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1192 return ReplaceInstUsesWith(I, Res);
1194 // If and'ing two fcmp, try combine them into one.
1195 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1196 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1197 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1198 return ReplaceInstUsesWith(I, Res);
1201 // fold (and (cast A), (cast B)) -> (cast (and A, B))
1202 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
1203 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) {
1204 Type *SrcTy = Op0C->getOperand(0)->getType();
1205 if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ?
1206 SrcTy == Op1C->getOperand(0)->getType() &&
1207 SrcTy->isIntOrIntVectorTy()) {
1208 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
1210 // Only do this if the casts both really cause code to be generated.
1211 if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
1212 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
1213 Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName());
1214 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
1217 // If this is and(cast(icmp), cast(icmp)), try to fold this even if the
1218 // cast is otherwise not optimizable. This happens for vector sexts.
1219 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
1220 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
1221 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1222 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1224 // If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the
1225 // cast is otherwise not optimizable. This happens for vector sexts.
1226 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
1227 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
1228 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1229 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1233 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
1234 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
1235 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
1236 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
1237 SI0->getOperand(1) == SI1->getOperand(1) &&
1238 (SI0->hasOneUse() || SI1->hasOneUse())) {
1240 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
1242 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
1243 SI1->getOperand(1));
1249 bool OpsSwapped = false;
1250 // Canonicalize SExt or Not to the LHS
1251 if (match(Op1, m_SExt(m_Value())) ||
1252 match(Op1, m_Not(m_Value()))) {
1253 std::swap(Op0, Op1);
1257 // Fold (and (sext bool to A), B) --> (select bool, B, 0)
1258 if (match(Op0, m_SExt(m_Value(X))) &&
1259 X->getType()->getScalarType()->isIntegerTy(1)) {
1260 Value *Zero = Constant::getNullValue(Op1->getType());
1261 return SelectInst::Create(X, Op1, Zero);
1264 // Fold (and ~(sext bool to A), B) --> (select bool, 0, B)
1265 if (match(Op0, m_Not(m_SExt(m_Value(X)))) &&
1266 X->getType()->getScalarType()->isIntegerTy(1)) {
1267 Value *Zero = Constant::getNullValue(Op0->getType());
1268 return SelectInst::Create(X, Zero, Op1);
1272 std::swap(Op0, Op1);
1275 return Changed ? &I : 0;
1278 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
1279 /// capable of providing pieces of a bswap. The subexpression provides pieces
1280 /// of a bswap if it is proven that each of the non-zero bytes in the output of
1281 /// the expression came from the corresponding "byte swapped" byte in some other
1282 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
1283 /// we know that the expression deposits the low byte of %X into the high byte
1284 /// of the bswap result and that all other bytes are zero. This expression is
1285 /// accepted, the high byte of ByteValues is set to X to indicate a correct
1288 /// This function returns true if the match was unsuccessful and false if so.
1289 /// On entry to the function the "OverallLeftShift" is a signed integer value
1290 /// indicating the number of bytes that the subexpression is later shifted. For
1291 /// example, if the expression is later right shifted by 16 bits, the
1292 /// OverallLeftShift value would be -2 on entry. This is used to specify which
1293 /// byte of ByteValues is actually being set.
1295 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
1296 /// byte is masked to zero by a user. For example, in (X & 255), X will be
1297 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
1298 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
1299 /// always in the local (OverallLeftShift) coordinate space.
1301 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
1302 SmallVector<Value*, 8> &ByteValues) {
1303 if (Instruction *I = dyn_cast<Instruction>(V)) {
1304 // If this is an or instruction, it may be an inner node of the bswap.
1305 if (I->getOpcode() == Instruction::Or) {
1306 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1308 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
1312 // If this is a logical shift by a constant multiple of 8, recurse with
1313 // OverallLeftShift and ByteMask adjusted.
1314 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
1316 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
1317 // Ensure the shift amount is defined and of a byte value.
1318 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
1321 unsigned ByteShift = ShAmt >> 3;
1322 if (I->getOpcode() == Instruction::Shl) {
1323 // X << 2 -> collect(X, +2)
1324 OverallLeftShift += ByteShift;
1325 ByteMask >>= ByteShift;
1327 // X >>u 2 -> collect(X, -2)
1328 OverallLeftShift -= ByteShift;
1329 ByteMask <<= ByteShift;
1330 ByteMask &= (~0U >> (32-ByteValues.size()));
1333 if (OverallLeftShift >= (int)ByteValues.size()) return true;
1334 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
1336 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1340 // If this is a logical 'and' with a mask that clears bytes, clear the
1341 // corresponding bytes in ByteMask.
1342 if (I->getOpcode() == Instruction::And &&
1343 isa<ConstantInt>(I->getOperand(1))) {
1344 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
1345 unsigned NumBytes = ByteValues.size();
1346 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
1347 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
1349 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
1350 // If this byte is masked out by a later operation, we don't care what
1352 if ((ByteMask & (1 << i)) == 0)
1355 // If the AndMask is all zeros for this byte, clear the bit.
1356 APInt MaskB = AndMask & Byte;
1358 ByteMask &= ~(1U << i);
1362 // If the AndMask is not all ones for this byte, it's not a bytezap.
1366 // Otherwise, this byte is kept.
1369 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1374 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
1375 // the input value to the bswap. Some observations: 1) if more than one byte
1376 // is demanded from this input, then it could not be successfully assembled
1377 // into a byteswap. At least one of the two bytes would not be aligned with
1378 // their ultimate destination.
1379 if (!isPowerOf2_32(ByteMask)) return true;
1380 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
1382 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
1383 // is demanded, it needs to go into byte 0 of the result. This means that the
1384 // byte needs to be shifted until it lands in the right byte bucket. The
1385 // shift amount depends on the position: if the byte is coming from the high
1386 // part of the value (e.g. byte 3) then it must be shifted right. If from the
1387 // low part, it must be shifted left.
1388 unsigned DestByteNo = InputByteNo + OverallLeftShift;
1389 if (ByteValues.size()-1-DestByteNo != InputByteNo)
1392 // If the destination byte value is already defined, the values are or'd
1393 // together, which isn't a bswap (unless it's an or of the same bits).
1394 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
1396 ByteValues[DestByteNo] = V;
1400 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
1401 /// If so, insert the new bswap intrinsic and return it.
1402 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
1403 IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
1404 if (!ITy || ITy->getBitWidth() % 16 ||
1405 // ByteMask only allows up to 32-byte values.
1406 ITy->getBitWidth() > 32*8)
1407 return 0; // Can only bswap pairs of bytes. Can't do vectors.
1409 /// ByteValues - For each byte of the result, we keep track of which value
1410 /// defines each byte.
1411 SmallVector<Value*, 8> ByteValues;
1412 ByteValues.resize(ITy->getBitWidth()/8);
1414 // Try to find all the pieces corresponding to the bswap.
1415 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
1416 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
1419 // Check to see if all of the bytes come from the same value.
1420 Value *V = ByteValues[0];
1421 if (V == 0) return 0; // Didn't find a byte? Must be zero.
1423 // Check to make sure that all of the bytes come from the same value.
1424 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
1425 if (ByteValues[i] != V)
1427 Module *M = I.getParent()->getParent()->getParent();
1428 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, ITy);
1429 return CallInst::Create(F, V);
1432 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
1433 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
1434 /// we can simplify this expression to "cond ? C : D or B".
1435 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
1436 Value *C, Value *D) {
1437 // If A is not a select of -1/0, this cannot match.
1439 if (!match(A, m_SExt(m_Value(Cond))) ||
1440 !Cond->getType()->isIntegerTy(1))
1443 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
1444 if (match(D, m_Not(m_SExt(m_Specific(Cond)))))
1445 return SelectInst::Create(Cond, C, B);
1446 if (match(D, m_SExt(m_Not(m_Specific(Cond)))))
1447 return SelectInst::Create(Cond, C, B);
1449 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
1450 if (match(B, m_Not(m_SExt(m_Specific(Cond)))))
1451 return SelectInst::Create(Cond, C, D);
1452 if (match(B, m_SExt(m_Not(m_Specific(Cond)))))
1453 return SelectInst::Create(Cond, C, D);
1457 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
1458 Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
1459 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
1461 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
1462 if (PredicatesFoldable(LHSCC, RHSCC)) {
1463 if (LHS->getOperand(0) == RHS->getOperand(1) &&
1464 LHS->getOperand(1) == RHS->getOperand(0))
1465 LHS->swapOperands();
1466 if (LHS->getOperand(0) == RHS->getOperand(0) &&
1467 LHS->getOperand(1) == RHS->getOperand(1)) {
1468 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1469 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
1470 bool isSigned = LHS->isSigned() || RHS->isSigned();
1471 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
1475 // handle (roughly):
1476 // (icmp ne (A & B), C) | (icmp ne (A & D), E)
1477 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, ICmpInst::ICMP_NE, Builder))
1480 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
1481 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
1482 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
1483 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
1484 if (LHSCst == 0 || RHSCst == 0) return 0;
1486 if (LHSCst == RHSCst && LHSCC == RHSCC) {
1487 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
1488 if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
1489 Value *NewOr = Builder->CreateOr(Val, Val2);
1490 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
1494 // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
1495 // iff C2 + CA == C1.
1496 if (LHSCC == ICmpInst::ICMP_ULT && RHSCC == ICmpInst::ICMP_EQ) {
1497 ConstantInt *AddCst;
1498 if (match(Val, m_Add(m_Specific(Val2), m_ConstantInt(AddCst))))
1499 if (RHSCst->getValue() + AddCst->getValue() == LHSCst->getValue())
1500 return Builder->CreateICmpULE(Val, LHSCst);
1503 // From here on, we only handle:
1504 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
1505 if (Val != Val2) return 0;
1507 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
1508 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
1509 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
1510 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
1511 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
1514 // We can't fold (ugt x, C) | (sgt x, C2).
1515 if (!PredicatesFoldable(LHSCC, RHSCC))
1518 // Ensure that the larger constant is on the RHS.
1520 if (CmpInst::isSigned(LHSCC) ||
1521 (ICmpInst::isEquality(LHSCC) &&
1522 CmpInst::isSigned(RHSCC)))
1523 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
1525 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
1528 std::swap(LHS, RHS);
1529 std::swap(LHSCst, RHSCst);
1530 std::swap(LHSCC, RHSCC);
1533 // At this point, we know we have two icmp instructions
1534 // comparing a value against two constants and or'ing the result
1535 // together. Because of the above check, we know that we only have
1536 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
1537 // icmp folding check above), that the two constants are not
1539 assert(LHSCst != RHSCst && "Compares not folded above?");
1542 default: llvm_unreachable("Unknown integer condition code!");
1543 case ICmpInst::ICMP_EQ:
1545 default: llvm_unreachable("Unknown integer condition code!");
1546 case ICmpInst::ICMP_EQ:
1547 if (LHSCst == SubOne(RHSCst)) {
1548 // (X == 13 | X == 14) -> X-13 <u 2
1549 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1550 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1551 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
1552 return Builder->CreateICmpULT(Add, AddCST);
1555 if (LHS->getOperand(0) == RHS->getOperand(0)) {
1556 // if LHSCst and RHSCst differ only by one bit:
1557 // (A == C1 || A == C2) -> (A & ~(C1 ^ C2)) == C1
1558 assert(LHSCst->getValue().ule(LHSCst->getValue()));
1560 APInt Xor = LHSCst->getValue() ^ RHSCst->getValue();
1561 if (Xor.isPowerOf2()) {
1562 Value *NegCst = Builder->getInt(~Xor);
1563 Value *And = Builder->CreateAnd(LHS->getOperand(0), NegCst);
1564 return Builder->CreateICmp(ICmpInst::ICMP_EQ, And, LHSCst);
1568 break; // (X == 13 | X == 15) -> no change
1569 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
1570 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
1572 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
1573 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
1574 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
1578 case ICmpInst::ICMP_NE:
1580 default: llvm_unreachable("Unknown integer condition code!");
1581 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
1582 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
1583 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
1585 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
1586 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
1587 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
1588 return ConstantInt::getTrue(LHS->getContext());
1590 case ICmpInst::ICMP_ULT:
1592 default: llvm_unreachable("Unknown integer condition code!");
1593 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
1595 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
1596 // If RHSCst is [us]MAXINT, it is always false. Not handling
1597 // this can cause overflow.
1598 if (RHSCst->isMaxValue(false))
1600 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false);
1601 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
1603 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
1604 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
1606 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
1610 case ICmpInst::ICMP_SLT:
1612 default: llvm_unreachable("Unknown integer condition code!");
1613 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
1615 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
1616 // If RHSCst is [us]MAXINT, it is always false. Not handling
1617 // this can cause overflow.
1618 if (RHSCst->isMaxValue(true))
1620 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false);
1621 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
1623 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
1624 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
1626 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
1630 case ICmpInst::ICMP_UGT:
1632 default: llvm_unreachable("Unknown integer condition code!");
1633 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
1634 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
1636 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
1638 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
1639 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
1640 return ConstantInt::getTrue(LHS->getContext());
1641 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
1645 case ICmpInst::ICMP_SGT:
1647 default: llvm_unreachable("Unknown integer condition code!");
1648 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
1649 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
1651 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
1653 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
1654 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
1655 return ConstantInt::getTrue(LHS->getContext());
1656 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
1664 /// FoldOrOfFCmps - Optimize (fcmp)|(fcmp). NOTE: Unlike the rest of
1665 /// instcombine, this returns a Value which should already be inserted into the
1667 Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
1668 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
1669 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
1670 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
1671 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
1672 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
1673 // If either of the constants are nans, then the whole thing returns
1675 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
1676 return ConstantInt::getTrue(LHS->getContext());
1678 // Otherwise, no need to compare the two constants, compare the
1680 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
1683 // Handle vector zeros. This occurs because the canonical form of
1684 // "fcmp uno x,x" is "fcmp uno x, 0".
1685 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
1686 isa<ConstantAggregateZero>(RHS->getOperand(1)))
1687 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
1692 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
1693 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
1694 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
1696 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
1697 // Swap RHS operands to match LHS.
1698 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
1699 std::swap(Op1LHS, Op1RHS);
1701 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
1702 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
1704 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
1705 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
1706 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
1707 if (Op0CC == FCmpInst::FCMP_FALSE)
1709 if (Op1CC == FCmpInst::FCMP_FALSE)
1713 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
1714 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
1715 if (Op0Ordered == Op1Ordered) {
1716 // If both are ordered or unordered, return a new fcmp with
1717 // or'ed predicates.
1718 return getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS, Builder);
1724 /// FoldOrWithConstants - This helper function folds:
1726 /// ((A | B) & C1) | (B & C2)
1732 /// when the XOR of the two constants is "all ones" (-1).
1733 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
1734 Value *A, Value *B, Value *C) {
1735 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
1739 ConstantInt *CI2 = 0;
1740 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
1742 APInt Xor = CI1->getValue() ^ CI2->getValue();
1743 if (!Xor.isAllOnesValue()) return 0;
1745 if (V1 == A || V1 == B) {
1746 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
1747 return BinaryOperator::CreateOr(NewOp, V1);
1753 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
1754 bool Changed = SimplifyAssociativeOrCommutative(I);
1755 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1757 if (Value *V = SimplifyOrInst(Op0, Op1, TD))
1758 return ReplaceInstUsesWith(I, V);
1760 // (A&B)|(A&C) -> A&(B|C) etc
1761 if (Value *V = SimplifyUsingDistributiveLaws(I))
1762 return ReplaceInstUsesWith(I, V);
1764 // See if we can simplify any instructions used by the instruction whose sole
1765 // purpose is to compute bits we don't care about.
1766 if (SimplifyDemandedInstructionBits(I))
1769 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1770 ConstantInt *C1 = 0; Value *X = 0;
1771 // (X & C1) | C2 --> (X | C2) & (C1|C2)
1772 // iff (C1 & C2) == 0.
1773 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
1774 (RHS->getValue() & C1->getValue()) != 0 &&
1776 Value *Or = Builder->CreateOr(X, RHS);
1778 return BinaryOperator::CreateAnd(Or,
1779 ConstantInt::get(I.getContext(),
1780 RHS->getValue() | C1->getValue()));
1783 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
1784 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
1786 Value *Or = Builder->CreateOr(X, RHS);
1788 return BinaryOperator::CreateXor(Or,
1789 ConstantInt::get(I.getContext(),
1790 C1->getValue() & ~RHS->getValue()));
1793 // Try to fold constant and into select arguments.
1794 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1795 if (Instruction *R = FoldOpIntoSelect(I, SI))
1798 if (isa<PHINode>(Op0))
1799 if (Instruction *NV = FoldOpIntoPhi(I))
1803 Value *A = 0, *B = 0;
1804 ConstantInt *C1 = 0, *C2 = 0;
1806 // (A | B) | C and A | (B | C) -> bswap if possible.
1807 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
1808 if (match(Op0, m_Or(m_Value(), m_Value())) ||
1809 match(Op1, m_Or(m_Value(), m_Value())) ||
1810 (match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
1811 match(Op1, m_LogicalShift(m_Value(), m_Value())))) {
1812 if (Instruction *BSwap = MatchBSwap(I))
1816 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
1817 if (Op0->hasOneUse() &&
1818 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
1819 MaskedValueIsZero(Op1, C1->getValue())) {
1820 Value *NOr = Builder->CreateOr(A, Op1);
1822 return BinaryOperator::CreateXor(NOr, C1);
1825 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
1826 if (Op1->hasOneUse() &&
1827 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
1828 MaskedValueIsZero(Op0, C1->getValue())) {
1829 Value *NOr = Builder->CreateOr(A, Op0);
1831 return BinaryOperator::CreateXor(NOr, C1);
1835 Value *C = 0, *D = 0;
1836 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1837 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1838 Value *V1 = 0, *V2 = 0;
1839 C1 = dyn_cast<ConstantInt>(C);
1840 C2 = dyn_cast<ConstantInt>(D);
1841 if (C1 && C2) { // (A & C1)|(B & C2)
1842 // If we have: ((V + N) & C1) | (V & C2)
1843 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1844 // replace with V+N.
1845 if (C1->getValue() == ~C2->getValue()) {
1846 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
1847 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1848 // Add commutes, try both ways.
1849 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
1850 return ReplaceInstUsesWith(I, A);
1851 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
1852 return ReplaceInstUsesWith(I, A);
1854 // Or commutes, try both ways.
1855 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
1856 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1857 // Add commutes, try both ways.
1858 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
1859 return ReplaceInstUsesWith(I, B);
1860 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
1861 return ReplaceInstUsesWith(I, B);
1865 if ((C1->getValue() & C2->getValue()) == 0) {
1866 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
1867 // iff (C1&C2) == 0 and (N&~C1) == 0
1868 if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
1869 ((V1 == B && MaskedValueIsZero(V2, ~C1->getValue())) || // (V|N)
1870 (V2 == B && MaskedValueIsZero(V1, ~C1->getValue())))) // (N|V)
1871 return BinaryOperator::CreateAnd(A,
1872 ConstantInt::get(A->getContext(),
1873 C1->getValue()|C2->getValue()));
1874 // Or commutes, try both ways.
1875 if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
1876 ((V1 == A && MaskedValueIsZero(V2, ~C2->getValue())) || // (V|N)
1877 (V2 == A && MaskedValueIsZero(V1, ~C2->getValue())))) // (N|V)
1878 return BinaryOperator::CreateAnd(B,
1879 ConstantInt::get(B->getContext(),
1880 C1->getValue()|C2->getValue()));
1882 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
1883 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
1884 ConstantInt *C3 = 0, *C4 = 0;
1885 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
1886 (C3->getValue() & ~C1->getValue()) == 0 &&
1887 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
1888 (C4->getValue() & ~C2->getValue()) == 0) {
1889 V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
1890 return BinaryOperator::CreateAnd(V2,
1891 ConstantInt::get(B->getContext(),
1892 C1->getValue()|C2->getValue()));
1897 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants.
1898 // Don't do this for vector select idioms, the code generator doesn't handle
1900 if (!I.getType()->isVectorTy()) {
1901 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
1903 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
1905 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
1907 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
1911 // ((A&~B)|(~A&B)) -> A^B
1912 if ((match(C, m_Not(m_Specific(D))) &&
1913 match(B, m_Not(m_Specific(A)))))
1914 return BinaryOperator::CreateXor(A, D);
1915 // ((~B&A)|(~A&B)) -> A^B
1916 if ((match(A, m_Not(m_Specific(D))) &&
1917 match(B, m_Not(m_Specific(C)))))
1918 return BinaryOperator::CreateXor(C, D);
1919 // ((A&~B)|(B&~A)) -> A^B
1920 if ((match(C, m_Not(m_Specific(B))) &&
1921 match(D, m_Not(m_Specific(A)))))
1922 return BinaryOperator::CreateXor(A, B);
1923 // ((~B&A)|(B&~A)) -> A^B
1924 if ((match(A, m_Not(m_Specific(B))) &&
1925 match(D, m_Not(m_Specific(C)))))
1926 return BinaryOperator::CreateXor(C, B);
1928 // ((A|B)&1)|(B&-2) -> (A&1) | B
1929 if (match(A, m_Or(m_Value(V1), m_Specific(B))) ||
1930 match(A, m_Or(m_Specific(B), m_Value(V1)))) {
1931 Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C);
1932 if (Ret) return Ret;
1934 // (B&-2)|((A|B)&1) -> (A&1) | B
1935 if (match(B, m_Or(m_Specific(A), m_Value(V1))) ||
1936 match(B, m_Or(m_Value(V1), m_Specific(A)))) {
1937 Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D);
1938 if (Ret) return Ret;
1942 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
1943 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
1944 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
1945 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
1946 SI0->getOperand(1) == SI1->getOperand(1) &&
1947 (SI0->hasOneUse() || SI1->hasOneUse())) {
1948 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
1950 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
1951 SI1->getOperand(1));
1955 // (~A | ~B) == (~(A & B)) - De Morgan's Law
1956 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1957 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1958 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1959 Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
1960 I.getName()+".demorgan");
1961 return BinaryOperator::CreateNot(And);
1964 // Canonicalize xor to the RHS.
1965 bool SwappedForXor = false;
1966 if (match(Op0, m_Xor(m_Value(), m_Value()))) {
1967 std::swap(Op0, Op1);
1968 SwappedForXor = true;
1971 // A | ( A ^ B) -> A | B
1972 // A | (~A ^ B) -> A | ~B
1973 // (A & B) | (A ^ B)
1974 if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
1975 if (Op0 == A || Op0 == B)
1976 return BinaryOperator::CreateOr(A, B);
1978 if (match(Op0, m_And(m_Specific(A), m_Specific(B))) ||
1979 match(Op0, m_And(m_Specific(B), m_Specific(A))))
1980 return BinaryOperator::CreateOr(A, B);
1982 if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
1983 Value *Not = Builder->CreateNot(B, B->getName()+".not");
1984 return BinaryOperator::CreateOr(Not, Op0);
1986 if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
1987 Value *Not = Builder->CreateNot(A, A->getName()+".not");
1988 return BinaryOperator::CreateOr(Not, Op0);
1992 // A | ~(A | B) -> A | ~B
1993 // A | ~(A ^ B) -> A | ~B
1994 if (match(Op1, m_Not(m_Value(A))))
1995 if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
1996 if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
1997 Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
1998 B->getOpcode() == Instruction::Xor)) {
1999 Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
2001 Value *Not = Builder->CreateNot(NotOp, NotOp->getName()+".not");
2002 return BinaryOperator::CreateOr(Not, Op0);
2006 std::swap(Op0, Op1);
2008 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
2009 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
2010 if (Value *Res = FoldOrOfICmps(LHS, RHS))
2011 return ReplaceInstUsesWith(I, Res);
2013 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
2014 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
2015 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
2016 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
2017 return ReplaceInstUsesWith(I, Res);
2019 // fold (or (cast A), (cast B)) -> (cast (or A, B))
2020 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2021 CastInst *Op1C = dyn_cast<CastInst>(Op1);
2022 if (Op1C && Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
2023 Type *SrcTy = Op0C->getOperand(0)->getType();
2024 if (SrcTy == Op1C->getOperand(0)->getType() &&
2025 SrcTy->isIntOrIntVectorTy()) {
2026 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
2028 if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) &&
2029 // Only do this if the casts both really cause code to be
2031 ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
2032 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
2033 Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName());
2034 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2037 // If this is or(cast(icmp), cast(icmp)), try to fold this even if the
2038 // cast is otherwise not optimizable. This happens for vector sexts.
2039 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
2040 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
2041 if (Value *Res = FoldOrOfICmps(LHS, RHS))
2042 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
2044 // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the
2045 // cast is otherwise not optimizable. This happens for vector sexts.
2046 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
2047 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
2048 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
2049 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
2054 // or(sext(A), B) -> A ? -1 : B where A is an i1
2055 // or(A, sext(B)) -> B ? -1 : A where B is an i1
2056 if (match(Op0, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
2057 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1);
2058 if (match(Op1, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
2059 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0);
2061 // Note: If we've gotten to the point of visiting the outer OR, then the
2062 // inner one couldn't be simplified. If it was a constant, then it won't
2063 // be simplified by a later pass either, so we try swapping the inner/outer
2064 // ORs in the hopes that we'll be able to simplify it this way.
2065 // (X|C) | V --> (X|V) | C
2066 if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
2067 match(Op0, m_Or(m_Value(A), m_ConstantInt(C1)))) {
2068 Value *Inner = Builder->CreateOr(A, Op1);
2069 Inner->takeName(Op0);
2070 return BinaryOperator::CreateOr(Inner, C1);
2073 // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D))
2074 // Since this OR statement hasn't been optimized further yet, we hope
2075 // that this transformation will allow the new ORs to be optimized.
2077 Value *X = 0, *Y = 0;
2078 if (Op0->hasOneUse() && Op1->hasOneUse() &&
2079 match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) &&
2080 match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) {
2081 Value *orTrue = Builder->CreateOr(A, C);
2082 Value *orFalse = Builder->CreateOr(B, D);
2083 return SelectInst::Create(X, orTrue, orFalse);
2087 return Changed ? &I : 0;
2090 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
2091 bool Changed = SimplifyAssociativeOrCommutative(I);
2092 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2094 if (Value *V = SimplifyXorInst(Op0, Op1, TD))
2095 return ReplaceInstUsesWith(I, V);
2097 // (A&B)^(A&C) -> A&(B^C) etc
2098 if (Value *V = SimplifyUsingDistributiveLaws(I))
2099 return ReplaceInstUsesWith(I, V);
2101 // See if we can simplify any instructions used by the instruction whose sole
2102 // purpose is to compute bits we don't care about.
2103 if (SimplifyDemandedInstructionBits(I))
2106 // Is this a ~ operation?
2107 if (Value *NotOp = dyn_castNotVal(&I)) {
2108 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
2109 if (Op0I->getOpcode() == Instruction::And ||
2110 Op0I->getOpcode() == Instruction::Or) {
2111 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
2112 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
2113 if (dyn_castNotVal(Op0I->getOperand(1)))
2114 Op0I->swapOperands();
2115 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
2117 Builder->CreateNot(Op0I->getOperand(1),
2118 Op0I->getOperand(1)->getName()+".not");
2119 if (Op0I->getOpcode() == Instruction::And)
2120 return BinaryOperator::CreateOr(Op0NotVal, NotY);
2121 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
2124 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
2125 // ~(X | Y) === (~X & ~Y) - De Morgan's Law
2126 if (isFreeToInvert(Op0I->getOperand(0)) &&
2127 isFreeToInvert(Op0I->getOperand(1))) {
2129 Builder->CreateNot(Op0I->getOperand(0), "notlhs");
2131 Builder->CreateNot(Op0I->getOperand(1), "notrhs");
2132 if (Op0I->getOpcode() == Instruction::And)
2133 return BinaryOperator::CreateOr(NotX, NotY);
2134 return BinaryOperator::CreateAnd(NotX, NotY);
2137 } else if (Op0I->getOpcode() == Instruction::AShr) {
2138 // ~(~X >>s Y) --> (X >>s Y)
2139 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0)))
2140 return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1));
2146 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2147 if (RHS->isOne() && Op0->hasOneUse())
2148 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
2149 if (CmpInst *CI = dyn_cast<CmpInst>(Op0))
2150 return CmpInst::Create(CI->getOpcode(),
2151 CI->getInversePredicate(),
2152 CI->getOperand(0), CI->getOperand(1));
2154 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
2155 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2156 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
2157 if (CI->hasOneUse() && Op0C->hasOneUse()) {
2158 Instruction::CastOps Opcode = Op0C->getOpcode();
2159 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
2160 (RHS == ConstantExpr::getCast(Opcode,
2161 ConstantInt::getTrue(I.getContext()),
2162 Op0C->getDestTy()))) {
2163 CI->setPredicate(CI->getInversePredicate());
2164 return CastInst::Create(Opcode, CI, Op0C->getType());
2170 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2171 // ~(c-X) == X-c-1 == X+(-c-1)
2172 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
2173 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
2174 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
2175 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
2176 ConstantInt::get(I.getType(), 1));
2177 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
2180 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2181 if (Op0I->getOpcode() == Instruction::Add) {
2182 // ~(X-c) --> (-c-1)-X
2183 if (RHS->isAllOnesValue()) {
2184 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
2185 return BinaryOperator::CreateSub(
2186 ConstantExpr::getSub(NegOp0CI,
2187 ConstantInt::get(I.getType(), 1)),
2188 Op0I->getOperand(0));
2189 } else if (RHS->getValue().isSignBit()) {
2190 // (X + C) ^ signbit -> (X + C + signbit)
2191 Constant *C = ConstantInt::get(I.getContext(),
2192 RHS->getValue() + Op0CI->getValue());
2193 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
2196 } else if (Op0I->getOpcode() == Instruction::Or) {
2197 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
2198 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
2199 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
2200 // Anything in both C1 and C2 is known to be zero, remove it from
2202 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
2203 NewRHS = ConstantExpr::getAnd(NewRHS,
2204 ConstantExpr::getNot(CommonBits));
2206 I.setOperand(0, Op0I->getOperand(0));
2207 I.setOperand(1, NewRHS);
2210 } else if (Op0I->getOpcode() == Instruction::LShr) {
2211 // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3)
2215 if (Op0I->hasOneUse() &&
2216 (E1 = dyn_cast<BinaryOperator>(Op0I->getOperand(0))) &&
2217 E1->getOpcode() == Instruction::Xor &&
2218 (C1 = dyn_cast<ConstantInt>(E1->getOperand(1)))) {
2219 // fold (C1 >> C2) ^ C3
2220 ConstantInt *C2 = Op0CI, *C3 = RHS;
2221 APInt FoldConst = C1->getValue().lshr(C2->getValue());
2222 FoldConst ^= C3->getValue();
2223 // Prepare the two operands.
2224 Value *Opnd0 = Builder->CreateLShr(E1->getOperand(0), C2);
2225 Opnd0->takeName(Op0I);
2226 cast<Instruction>(Opnd0)->setDebugLoc(I.getDebugLoc());
2227 Value *FoldVal = ConstantInt::get(Opnd0->getType(), FoldConst);
2229 return BinaryOperator::CreateXor(Opnd0, FoldVal);
2235 // Try to fold constant and into select arguments.
2236 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2237 if (Instruction *R = FoldOpIntoSelect(I, SI))
2239 if (isa<PHINode>(Op0))
2240 if (Instruction *NV = FoldOpIntoPhi(I))
2244 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
2247 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2248 if (A == Op0) { // B^(B|A) == (A|B)^B
2249 Op1I->swapOperands();
2251 std::swap(Op0, Op1);
2252 } else if (B == Op0) { // B^(A|B) == (A|B)^B
2253 I.swapOperands(); // Simplified below.
2254 std::swap(Op0, Op1);
2256 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
2258 if (A == Op0) { // A^(A&B) -> A^(B&A)
2259 Op1I->swapOperands();
2262 if (B == Op0) { // A^(B&A) -> (B&A)^A
2263 I.swapOperands(); // Simplified below.
2264 std::swap(Op0, Op1);
2269 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
2272 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2273 Op0I->hasOneUse()) {
2274 if (A == Op1) // (B|A)^B == (A|B)^B
2276 if (B == Op1) // (A|B)^B == A & ~B
2277 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1));
2278 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2280 if (A == Op1) // (A&B)^A -> (B&A)^A
2282 if (B == Op1 && // (B&A)^A == ~B & A
2283 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
2284 return BinaryOperator::CreateAnd(Builder->CreateNot(A), Op1);
2289 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
2290 if (Op0I && Op1I && Op0I->isShift() &&
2291 Op0I->getOpcode() == Op1I->getOpcode() &&
2292 Op0I->getOperand(1) == Op1I->getOperand(1) &&
2293 (Op0I->hasOneUse() || Op1I->hasOneUse())) {
2295 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
2297 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
2298 Op1I->getOperand(1));
2302 Value *A, *B, *C, *D;
2303 // (A & B)^(A | B) -> A ^ B
2304 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2305 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
2306 if ((A == C && B == D) || (A == D && B == C))
2307 return BinaryOperator::CreateXor(A, B);
2309 // (A | B)^(A & B) -> A ^ B
2310 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2311 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
2312 if ((A == C && B == D) || (A == D && B == C))
2313 return BinaryOperator::CreateXor(A, B);
2317 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
2318 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
2319 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
2320 if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
2321 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2322 LHS->getOperand(1) == RHS->getOperand(0))
2323 LHS->swapOperands();
2324 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2325 LHS->getOperand(1) == RHS->getOperand(1)) {
2326 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2327 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
2328 bool isSigned = LHS->isSigned() || RHS->isSigned();
2329 return ReplaceInstUsesWith(I,
2330 getNewICmpValue(isSigned, Code, Op0, Op1,
2335 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
2336 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2337 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
2338 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
2339 Type *SrcTy = Op0C->getOperand(0)->getType();
2340 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() &&
2341 // Only do this if the casts both really cause code to be generated.
2342 ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0),
2344 ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0),
2346 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
2347 Op1C->getOperand(0), I.getName());
2348 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2353 return Changed ? &I : 0;