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 "InstCombineInternal.h"
15 #include "llvm/Analysis/InstructionSimplify.h"
16 #include "llvm/IR/ConstantRange.h"
17 #include "llvm/IR/Intrinsics.h"
18 #include "llvm/IR/PatternMatch.h"
19 #include "llvm/Transforms/Utils/CmpInstAnalysis.h"
20 #include "llvm/Transforms/Utils/Local.h"
22 using namespace PatternMatch;
24 #define DEBUG_TYPE "instcombine"
26 static inline Value *dyn_castNotVal(Value *V) {
27 // If this is not(not(x)) don't return that this is a not: we want the two
28 // not's to be folded first.
29 if (BinaryOperator::isNot(V)) {
30 Value *Operand = BinaryOperator::getNotArgument(V);
31 if (!IsFreeToInvert(Operand, Operand->hasOneUse()))
35 // Constants can be considered to be not'ed values...
36 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
37 return ConstantInt::get(C->getType(), ~C->getValue());
41 /// Similar to getICmpCode but for FCmpInst. This encodes a fcmp predicate into
43 static unsigned getFCmpCode(FCmpInst::Predicate CC) {
44 assert(FCmpInst::FCMP_FALSE <= CC && CC <= FCmpInst::FCMP_TRUE &&
45 "Unexpected FCmp predicate!");
46 // Take advantage of the bit pattern of FCmpInst::Predicate here.
48 static_assert(FCmpInst::FCMP_FALSE == 0, ""); // 0 0 0 0
49 static_assert(FCmpInst::FCMP_OEQ == 1, ""); // 0 0 0 1
50 static_assert(FCmpInst::FCMP_OGT == 2, ""); // 0 0 1 0
51 static_assert(FCmpInst::FCMP_OGE == 3, ""); // 0 0 1 1
52 static_assert(FCmpInst::FCMP_OLT == 4, ""); // 0 1 0 0
53 static_assert(FCmpInst::FCMP_OLE == 5, ""); // 0 1 0 1
54 static_assert(FCmpInst::FCMP_ONE == 6, ""); // 0 1 1 0
55 static_assert(FCmpInst::FCMP_ORD == 7, ""); // 0 1 1 1
56 static_assert(FCmpInst::FCMP_UNO == 8, ""); // 1 0 0 0
57 static_assert(FCmpInst::FCMP_UEQ == 9, ""); // 1 0 0 1
58 static_assert(FCmpInst::FCMP_UGT == 10, ""); // 1 0 1 0
59 static_assert(FCmpInst::FCMP_UGE == 11, ""); // 1 0 1 1
60 static_assert(FCmpInst::FCMP_ULT == 12, ""); // 1 1 0 0
61 static_assert(FCmpInst::FCMP_ULE == 13, ""); // 1 1 0 1
62 static_assert(FCmpInst::FCMP_UNE == 14, ""); // 1 1 1 0
63 static_assert(FCmpInst::FCMP_TRUE == 15, ""); // 1 1 1 1
67 /// This is the complement of getICmpCode, which turns an opcode and two
68 /// operands into either a constant true or false, or a brand new ICmp
69 /// instruction. The sign is passed in to determine which kind of predicate to
70 /// use in the new icmp instruction.
71 static Value *getNewICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS,
72 InstCombiner::BuilderTy *Builder) {
73 ICmpInst::Predicate NewPred;
74 if (Value *NewConstant = getICmpValue(Sign, Code, LHS, RHS, NewPred))
76 return Builder->CreateICmp(NewPred, LHS, RHS);
79 /// This is the complement of getFCmpCode, which turns an opcode and two
80 /// operands into either a FCmp instruction, or a true/false constant.
81 static Value *getFCmpValue(unsigned Code, Value *LHS, Value *RHS,
82 InstCombiner::BuilderTy *Builder) {
83 const auto Pred = static_cast<FCmpInst::Predicate>(Code);
84 assert(FCmpInst::FCMP_FALSE <= Pred && Pred <= FCmpInst::FCMP_TRUE &&
85 "Unexpected FCmp predicate!");
86 if (Pred == FCmpInst::FCMP_FALSE)
87 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
88 if (Pred == FCmpInst::FCMP_TRUE)
89 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
90 return Builder->CreateFCmp(Pred, LHS, RHS);
93 /// \brief Transform BITWISE_OP(BSWAP(A),BSWAP(B)) to BSWAP(BITWISE_OP(A, B))
94 /// \param I Binary operator to transform.
95 /// \return Pointer to node that must replace the original binary operator, or
96 /// null pointer if no transformation was made.
97 Value *InstCombiner::SimplifyBSwap(BinaryOperator &I) {
98 IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
101 if (I.getType()->isVectorTy())
104 // Can only do bitwise ops.
105 if (!I.isBitwiseLogicOp())
108 Value *OldLHS = I.getOperand(0);
109 Value *OldRHS = I.getOperand(1);
110 ConstantInt *ConstLHS = dyn_cast<ConstantInt>(OldLHS);
111 ConstantInt *ConstRHS = dyn_cast<ConstantInt>(OldRHS);
112 IntrinsicInst *IntrLHS = dyn_cast<IntrinsicInst>(OldLHS);
113 IntrinsicInst *IntrRHS = dyn_cast<IntrinsicInst>(OldRHS);
114 bool IsBswapLHS = (IntrLHS && IntrLHS->getIntrinsicID() == Intrinsic::bswap);
115 bool IsBswapRHS = (IntrRHS && IntrRHS->getIntrinsicID() == Intrinsic::bswap);
117 if (!IsBswapLHS && !IsBswapRHS)
120 if (!IsBswapLHS && !ConstLHS)
123 if (!IsBswapRHS && !ConstRHS)
126 /// OP( BSWAP(x), BSWAP(y) ) -> BSWAP( OP(x, y) )
127 /// OP( BSWAP(x), CONSTANT ) -> BSWAP( OP(x, BSWAP(CONSTANT) ) )
128 Value *NewLHS = IsBswapLHS ? IntrLHS->getOperand(0) :
129 Builder->getInt(ConstLHS->getValue().byteSwap());
131 Value *NewRHS = IsBswapRHS ? IntrRHS->getOperand(0) :
132 Builder->getInt(ConstRHS->getValue().byteSwap());
134 Value *BinOp = Builder->CreateBinOp(I.getOpcode(), NewLHS, NewRHS);
135 Function *F = Intrinsic::getDeclaration(I.getModule(), Intrinsic::bswap, ITy);
136 return Builder->CreateCall(F, BinOp);
139 /// This handles expressions of the form ((val OP C1) & C2). Where
140 /// the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
141 /// guaranteed to be a binary operator.
142 Instruction *InstCombiner::OptAndOp(Instruction *Op,
145 BinaryOperator &TheAnd) {
146 Value *X = Op->getOperand(0);
147 Constant *Together = nullptr;
149 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
151 switch (Op->getOpcode()) {
152 case Instruction::Xor:
153 if (Op->hasOneUse()) {
154 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
155 Value *And = Builder->CreateAnd(X, AndRHS);
157 return BinaryOperator::CreateXor(And, Together);
160 case Instruction::Or:
161 if (Op->hasOneUse()){
162 if (Together != OpRHS) {
163 // (X | C1) & C2 --> (X | (C1&C2)) & C2
164 Value *Or = Builder->CreateOr(X, Together);
166 return BinaryOperator::CreateAnd(Or, AndRHS);
169 ConstantInt *TogetherCI = dyn_cast<ConstantInt>(Together);
170 if (TogetherCI && !TogetherCI->isZero()){
171 // (X | C1) & C2 --> (X & (C2^(C1&C2))) | C1
172 // NOTE: This reduces the number of bits set in the & mask, which
173 // can expose opportunities for store narrowing.
174 Together = ConstantExpr::getXor(AndRHS, Together);
175 Value *And = Builder->CreateAnd(X, Together);
177 return BinaryOperator::CreateOr(And, OpRHS);
182 case Instruction::Add:
183 if (Op->hasOneUse()) {
184 // Adding a one to a single bit bit-field should be turned into an XOR
185 // of the bit. First thing to check is to see if this AND is with a
186 // single bit constant.
187 const APInt &AndRHSV = AndRHS->getValue();
189 // If there is only one bit set.
190 if (AndRHSV.isPowerOf2()) {
191 // Ok, at this point, we know that we are masking the result of the
192 // ADD down to exactly one bit. If the constant we are adding has
193 // no bits set below this bit, then we can eliminate the ADD.
194 const APInt& AddRHS = OpRHS->getValue();
196 // Check to see if any bits below the one bit set in AndRHSV are set.
197 if ((AddRHS & (AndRHSV-1)) == 0) {
198 // If not, the only thing that can effect the output of the AND is
199 // the bit specified by AndRHSV. If that bit is set, the effect of
200 // the XOR is to toggle the bit. If it is clear, then the ADD has
202 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
203 TheAnd.setOperand(0, X);
206 // Pull the XOR out of the AND.
207 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
208 NewAnd->takeName(Op);
209 return BinaryOperator::CreateXor(NewAnd, AndRHS);
216 case Instruction::Shl: {
217 // We know that the AND will not produce any of the bits shifted in, so if
218 // the anded constant includes them, clear them now!
220 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
221 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
222 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
223 ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShlMask);
225 if (CI->getValue() == ShlMask)
226 // Masking out bits that the shift already masks.
227 return replaceInstUsesWith(TheAnd, Op); // No need for the and.
229 if (CI != AndRHS) { // Reducing bits set in and.
230 TheAnd.setOperand(1, CI);
235 case Instruction::LShr: {
236 // We know that the AND will not produce any of the bits shifted in, so if
237 // the anded constant includes them, clear them now! This only applies to
238 // unsigned shifts, because a signed shr may bring in set bits!
240 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
241 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
242 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
243 ConstantInt *CI = Builder->getInt(AndRHS->getValue() & ShrMask);
245 if (CI->getValue() == ShrMask)
246 // Masking out bits that the shift already masks.
247 return replaceInstUsesWith(TheAnd, Op);
250 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
255 case Instruction::AShr:
257 // See if this is shifting in some sign extension, then masking it out
259 if (Op->hasOneUse()) {
260 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
261 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
262 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
263 Constant *C = Builder->getInt(AndRHS->getValue() & ShrMask);
264 if (C == AndRHS) { // Masking out bits shifted in.
265 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
266 // Make the argument unsigned.
267 Value *ShVal = Op->getOperand(0);
268 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
269 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
277 /// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise
278 /// (V < Lo || V >= Hi). This method expects that Lo <= Hi. IsSigned indicates
279 /// whether to treat V, Lo, and Hi as signed or not.
280 Value *InstCombiner::insertRangeTest(Value *V, const APInt &Lo, const APInt &Hi,
281 bool isSigned, bool Inside) {
282 assert((isSigned ? Lo.sle(Hi) : Lo.ule(Hi)) &&
283 "Lo is not <= Hi in range emission code!");
285 Type *Ty = V->getType();
287 return Inside ? ConstantInt::getFalse(Ty) : ConstantInt::getTrue(Ty);
289 // V >= Min && V < Hi --> V < Hi
290 // V < Min || V >= Hi --> V >= Hi
291 ICmpInst::Predicate Pred = Inside ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE;
292 if (isSigned ? Lo.isMinSignedValue() : Lo.isMinValue()) {
293 Pred = isSigned ? ICmpInst::getSignedPredicate(Pred) : Pred;
294 return Builder->CreateICmp(Pred, V, ConstantInt::get(Ty, Hi));
297 // V >= Lo && V < Hi --> V - Lo u< Hi - Lo
298 // V < Lo || V >= Hi --> V - Lo u>= Hi - Lo
300 Builder->CreateSub(V, ConstantInt::get(Ty, Lo), V->getName() + ".off");
301 Constant *HiMinusLo = ConstantInt::get(Ty, Hi - Lo);
302 return Builder->CreateICmp(Pred, VMinusLo, HiMinusLo);
305 /// Returns true iff Val consists of one contiguous run of 1s with any number
306 /// of 0s on either side. The 1s are allowed to wrap from LSB to MSB,
307 /// so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
308 /// not, since all 1s are not contiguous.
309 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
310 const APInt& V = Val->getValue();
311 uint32_t BitWidth = Val->getType()->getBitWidth();
312 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
314 // look for the first zero bit after the run of ones
315 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
316 // look for the first non-zero bit
317 ME = V.getActiveBits();
321 /// This is part of an expression (LHS +/- RHS) & Mask, where isSub determines
322 /// whether the operator is a sub. If we can fold one of the following xforms:
324 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
325 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
326 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
328 /// return (A +/- B).
330 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
331 ConstantInt *Mask, bool isSub,
333 Instruction *LHSI = dyn_cast<Instruction>(LHS);
334 if (!LHSI || LHSI->getNumOperands() != 2 ||
335 !isa<ConstantInt>(LHSI->getOperand(1))) return nullptr;
337 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
339 switch (LHSI->getOpcode()) {
340 default: return nullptr;
341 case Instruction::And:
342 if (ConstantExpr::getAnd(N, Mask) == Mask) {
343 // If the AndRHS is a power of two minus one (0+1+), this is simple.
344 if ((Mask->getValue().countLeadingZeros() +
345 Mask->getValue().countPopulation()) ==
346 Mask->getValue().getBitWidth())
349 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
350 // part, we don't need any explicit masks to take them out of A. If that
351 // is all N is, ignore it.
352 uint32_t MB = 0, ME = 0;
353 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
354 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
355 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
356 if (MaskedValueIsZero(RHS, Mask, 0, &I))
361 case Instruction::Or:
362 case Instruction::Xor:
363 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
364 if ((Mask->getValue().countLeadingZeros() +
365 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
366 && ConstantExpr::getAnd(N, Mask)->isNullValue())
372 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
373 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
376 /// enum for classifying (icmp eq (A & B), C) and (icmp ne (A & B), C)
377 /// One of A and B is considered the mask, the other the value. This is
378 /// described as the "AMask" or "BMask" part of the enum. If the enum
379 /// contains only "Mask", then both A and B can be considered masks.
380 /// If A is the mask, then it was proven, that (A & C) == C. This
381 /// is trivial if C == A, or C == 0. If both A and C are constants, this
382 /// proof is also easy.
383 /// For the following explanations we assume that A is the mask.
384 /// The part "AllOnes" declares, that the comparison is true only
385 /// if (A & B) == A, or all bits of A are set in B.
386 /// Example: (icmp eq (A & 3), 3) -> FoldMskICmp_AMask_AllOnes
387 /// The part "AllZeroes" declares, that the comparison is true only
388 /// if (A & B) == 0, or all bits of A are cleared in B.
389 /// Example: (icmp eq (A & 3), 0) -> FoldMskICmp_Mask_AllZeroes
390 /// The part "Mixed" declares, that (A & B) == C and C might or might not
391 /// contain any number of one bits and zero bits.
392 /// Example: (icmp eq (A & 3), 1) -> FoldMskICmp_AMask_Mixed
393 /// The Part "Not" means, that in above descriptions "==" should be replaced
395 /// Example: (icmp ne (A & 3), 3) -> FoldMskICmp_AMask_NotAllOnes
396 /// If the mask A contains a single bit, then the following is equivalent:
397 /// (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
398 /// (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
399 enum MaskedICmpType {
400 FoldMskICmp_AMask_AllOnes = 1,
401 FoldMskICmp_AMask_NotAllOnes = 2,
402 FoldMskICmp_BMask_AllOnes = 4,
403 FoldMskICmp_BMask_NotAllOnes = 8,
404 FoldMskICmp_Mask_AllZeroes = 16,
405 FoldMskICmp_Mask_NotAllZeroes = 32,
406 FoldMskICmp_AMask_Mixed = 64,
407 FoldMskICmp_AMask_NotMixed = 128,
408 FoldMskICmp_BMask_Mixed = 256,
409 FoldMskICmp_BMask_NotMixed = 512
412 /// Return the set of pattern classes (from MaskedICmpType)
413 /// that (icmp SCC (A & B), C) satisfies.
414 static unsigned getTypeOfMaskedICmp(Value* A, Value* B, Value* C,
415 ICmpInst::Predicate SCC)
417 ConstantInt *ACst = dyn_cast<ConstantInt>(A);
418 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
419 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
420 bool icmp_eq = (SCC == ICmpInst::ICMP_EQ);
421 bool icmp_abit = (ACst && !ACst->isZero() &&
422 ACst->getValue().isPowerOf2());
423 bool icmp_bbit = (BCst && !BCst->isZero() &&
424 BCst->getValue().isPowerOf2());
426 if (CCst && CCst->isZero()) {
427 // if C is zero, then both A and B qualify as mask
428 result |= (icmp_eq ? (FoldMskICmp_Mask_AllZeroes |
429 FoldMskICmp_AMask_Mixed |
430 FoldMskICmp_BMask_Mixed)
431 : (FoldMskICmp_Mask_NotAllZeroes |
432 FoldMskICmp_AMask_NotMixed |
433 FoldMskICmp_BMask_NotMixed));
435 result |= (icmp_eq ? (FoldMskICmp_AMask_NotAllOnes |
436 FoldMskICmp_AMask_NotMixed)
437 : (FoldMskICmp_AMask_AllOnes |
438 FoldMskICmp_AMask_Mixed));
440 result |= (icmp_eq ? (FoldMskICmp_BMask_NotAllOnes |
441 FoldMskICmp_BMask_NotMixed)
442 : (FoldMskICmp_BMask_AllOnes |
443 FoldMskICmp_BMask_Mixed));
447 result |= (icmp_eq ? (FoldMskICmp_AMask_AllOnes |
448 FoldMskICmp_AMask_Mixed)
449 : (FoldMskICmp_AMask_NotAllOnes |
450 FoldMskICmp_AMask_NotMixed));
452 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
453 FoldMskICmp_AMask_NotMixed)
454 : (FoldMskICmp_Mask_AllZeroes |
455 FoldMskICmp_AMask_Mixed));
456 } else if (ACst && CCst &&
457 ConstantExpr::getAnd(ACst, CCst) == CCst) {
458 result |= (icmp_eq ? FoldMskICmp_AMask_Mixed
459 : FoldMskICmp_AMask_NotMixed);
462 result |= (icmp_eq ? (FoldMskICmp_BMask_AllOnes |
463 FoldMskICmp_BMask_Mixed)
464 : (FoldMskICmp_BMask_NotAllOnes |
465 FoldMskICmp_BMask_NotMixed));
467 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
468 FoldMskICmp_BMask_NotMixed)
469 : (FoldMskICmp_Mask_AllZeroes |
470 FoldMskICmp_BMask_Mixed));
471 } else if (BCst && CCst &&
472 ConstantExpr::getAnd(BCst, CCst) == CCst) {
473 result |= (icmp_eq ? FoldMskICmp_BMask_Mixed
474 : FoldMskICmp_BMask_NotMixed);
479 /// Convert an analysis of a masked ICmp into its equivalent if all boolean
480 /// operations had the opposite sense. Since each "NotXXX" flag (recording !=)
481 /// is adjacent to the corresponding normal flag (recording ==), this just
482 /// involves swapping those bits over.
483 static unsigned conjugateICmpMask(unsigned Mask) {
485 NewMask = (Mask & (FoldMskICmp_AMask_AllOnes | FoldMskICmp_BMask_AllOnes |
486 FoldMskICmp_Mask_AllZeroes | FoldMskICmp_AMask_Mixed |
487 FoldMskICmp_BMask_Mixed))
491 (Mask & (FoldMskICmp_AMask_NotAllOnes | FoldMskICmp_BMask_NotAllOnes |
492 FoldMskICmp_Mask_NotAllZeroes | FoldMskICmp_AMask_NotMixed |
493 FoldMskICmp_BMask_NotMixed))
499 /// Handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
500 /// Return the set of pattern classes (from MaskedICmpType)
501 /// that both LHS and RHS satisfy.
502 static unsigned foldLogOpOfMaskedICmpsHelper(Value*& A,
503 Value*& B, Value*& C,
504 Value*& D, Value*& E,
505 ICmpInst *LHS, ICmpInst *RHS,
506 ICmpInst::Predicate &LHSCC,
507 ICmpInst::Predicate &RHSCC) {
508 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) return 0;
509 // vectors are not (yet?) supported
510 if (LHS->getOperand(0)->getType()->isVectorTy()) return 0;
512 // Here comes the tricky part:
513 // LHS might be of the form L11 & L12 == X, X == L21 & L22,
514 // and L11 & L12 == L21 & L22. The same goes for RHS.
515 // Now we must find those components L** and R**, that are equal, so
516 // that we can extract the parameters A, B, C, D, and E for the canonical
518 Value *L1 = LHS->getOperand(0);
519 Value *L2 = LHS->getOperand(1);
520 Value *L11,*L12,*L21,*L22;
521 // Check whether the icmp can be decomposed into a bit test.
522 if (decomposeBitTestICmp(LHS, LHSCC, L11, L12, L2)) {
523 L21 = L22 = L1 = nullptr;
525 // Look for ANDs in the LHS icmp.
526 if (!L1->getType()->isIntegerTy()) {
527 // You can icmp pointers, for example. They really aren't masks.
529 } else if (!match(L1, m_And(m_Value(L11), m_Value(L12)))) {
530 // Any icmp can be viewed as being trivially masked; if it allows us to
531 // remove one, it's worth it.
533 L12 = Constant::getAllOnesValue(L1->getType());
536 if (!L2->getType()->isIntegerTy()) {
537 // You can icmp pointers, for example. They really aren't masks.
539 } else if (!match(L2, m_And(m_Value(L21), m_Value(L22)))) {
541 L22 = Constant::getAllOnesValue(L2->getType());
545 // Bail if LHS was a icmp that can't be decomposed into an equality.
546 if (!ICmpInst::isEquality(LHSCC))
549 Value *R1 = RHS->getOperand(0);
550 Value *R2 = RHS->getOperand(1);
553 if (decomposeBitTestICmp(RHS, RHSCC, R11, R12, R2)) {
554 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
556 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
561 E = R2; R1 = nullptr; ok = true;
562 } else if (R1->getType()->isIntegerTy()) {
563 if (!match(R1, m_And(m_Value(R11), m_Value(R12)))) {
564 // As before, model no mask as a trivial mask if it'll let us do an
567 R12 = Constant::getAllOnesValue(R1->getType());
570 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
571 A = R11; D = R12; E = R2; ok = true;
572 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
573 A = R12; D = R11; E = R2; ok = true;
577 // Bail if RHS was a icmp that can't be decomposed into an equality.
578 if (!ICmpInst::isEquality(RHSCC))
581 // Look for ANDs on the right side of the RHS icmp.
582 if (!ok && R2->getType()->isIntegerTy()) {
583 if (!match(R2, m_And(m_Value(R11), m_Value(R12)))) {
585 R12 = Constant::getAllOnesValue(R2->getType());
588 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
589 A = R11; D = R12; E = R1; ok = true;
590 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
591 A = R12; D = R11; E = R1; ok = true;
601 } else if (L12 == A) {
603 } else if (L21 == A) {
605 } else if (L22 == A) {
609 unsigned LeftType = getTypeOfMaskedICmp(A, B, C, LHSCC);
610 unsigned RightType = getTypeOfMaskedICmp(A, D, E, RHSCC);
611 return LeftType & RightType;
614 /// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
615 /// into a single (icmp(A & X) ==/!= Y).
616 static Value *foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
617 llvm::InstCombiner::BuilderTy *Builder) {
618 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr;
619 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
620 unsigned Mask = foldLogOpOfMaskedICmpsHelper(A, B, C, D, E, LHS, RHS,
622 if (Mask == 0) return nullptr;
623 assert(ICmpInst::isEquality(LHSCC) && ICmpInst::isEquality(RHSCC) &&
624 "foldLogOpOfMaskedICmpsHelper must return an equality predicate.");
626 // In full generality:
627 // (icmp (A & B) Op C) | (icmp (A & D) Op E)
628 // == ![ (icmp (A & B) !Op C) & (icmp (A & D) !Op E) ]
630 // If the latter can be converted into (icmp (A & X) Op Y) then the former is
631 // equivalent to (icmp (A & X) !Op Y).
633 // Therefore, we can pretend for the rest of this function that we're dealing
634 // with the conjunction, provided we flip the sense of any comparisons (both
635 // input and output).
637 // In most cases we're going to produce an EQ for the "&&" case.
638 ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
640 // Convert the masking analysis into its equivalent with negated
642 Mask = conjugateICmpMask(Mask);
645 if (Mask & FoldMskICmp_Mask_AllZeroes) {
646 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
647 // -> (icmp eq (A & (B|D)), 0)
648 Value *NewOr = Builder->CreateOr(B, D);
649 Value *NewAnd = Builder->CreateAnd(A, NewOr);
650 // We can't use C as zero because we might actually handle
651 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
652 // with B and D, having a single bit set.
653 Value *Zero = Constant::getNullValue(A->getType());
654 return Builder->CreateICmp(NewCC, NewAnd, Zero);
656 if (Mask & FoldMskICmp_BMask_AllOnes) {
657 // (icmp eq (A & B), B) & (icmp eq (A & D), D)
658 // -> (icmp eq (A & (B|D)), (B|D))
659 Value *NewOr = Builder->CreateOr(B, D);
660 Value *NewAnd = Builder->CreateAnd(A, NewOr);
661 return Builder->CreateICmp(NewCC, NewAnd, NewOr);
663 if (Mask & FoldMskICmp_AMask_AllOnes) {
664 // (icmp eq (A & B), A) & (icmp eq (A & D), A)
665 // -> (icmp eq (A & (B&D)), A)
666 Value *NewAnd1 = Builder->CreateAnd(B, D);
667 Value *NewAnd2 = Builder->CreateAnd(A, NewAnd1);
668 return Builder->CreateICmp(NewCC, NewAnd2, A);
671 // Remaining cases assume at least that B and D are constant, and depend on
672 // their actual values. This isn't strictly necessary, just a "handle the
673 // easy cases for now" decision.
674 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
675 if (!BCst) return nullptr;
676 ConstantInt *DCst = dyn_cast<ConstantInt>(D);
677 if (!DCst) return nullptr;
679 if (Mask & (FoldMskICmp_Mask_NotAllZeroes | FoldMskICmp_BMask_NotAllOnes)) {
680 // (icmp ne (A & B), 0) & (icmp ne (A & D), 0) and
681 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
682 // -> (icmp ne (A & B), 0) or (icmp ne (A & D), 0)
683 // Only valid if one of the masks is a superset of the other (check "B&D" is
684 // the same as either B or D).
685 APInt NewMask = BCst->getValue() & DCst->getValue();
687 if (NewMask == BCst->getValue())
689 else if (NewMask == DCst->getValue())
692 if (Mask & FoldMskICmp_AMask_NotAllOnes) {
693 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
694 // -> (icmp ne (A & B), A) or (icmp ne (A & D), A)
695 // Only valid if one of the masks is a superset of the other (check "B|D" is
696 // the same as either B or D).
697 APInt NewMask = BCst->getValue() | DCst->getValue();
699 if (NewMask == BCst->getValue())
701 else if (NewMask == DCst->getValue())
704 if (Mask & FoldMskICmp_BMask_Mixed) {
705 // (icmp eq (A & B), C) & (icmp eq (A & D), E)
706 // We already know that B & C == C && D & E == E.
707 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
708 // C and E, which are shared by both the mask B and the mask D, don't
709 // contradict, then we can transform to
710 // -> (icmp eq (A & (B|D)), (C|E))
711 // Currently, we only handle the case of B, C, D, and E being constant.
712 // We can't simply use C and E because we might actually handle
713 // (icmp ne (A & B), B) & (icmp eq (A & D), D)
714 // with B and D, having a single bit set.
715 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
716 if (!CCst) return nullptr;
717 ConstantInt *ECst = dyn_cast<ConstantInt>(E);
718 if (!ECst) return nullptr;
720 CCst = cast<ConstantInt>(ConstantExpr::getXor(BCst, CCst));
722 ECst = cast<ConstantInt>(ConstantExpr::getXor(DCst, ECst));
723 // If there is a conflict, we should actually return a false for the
725 if (((BCst->getValue() & DCst->getValue()) &
726 (CCst->getValue() ^ ECst->getValue())) != 0)
727 return ConstantInt::get(LHS->getType(), !IsAnd);
728 Value *NewOr1 = Builder->CreateOr(B, D);
729 Value *NewOr2 = ConstantExpr::getOr(CCst, ECst);
730 Value *NewAnd = Builder->CreateAnd(A, NewOr1);
731 return Builder->CreateICmp(NewCC, NewAnd, NewOr2);
736 /// Try to fold a signed range checked with lower bound 0 to an unsigned icmp.
737 /// Example: (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
738 /// If \p Inverted is true then the check is for the inverted range, e.g.
739 /// (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
740 Value *InstCombiner::simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1,
742 // Check the lower range comparison, e.g. x >= 0
743 // InstCombine already ensured that if there is a constant it's on the RHS.
744 ConstantInt *RangeStart = dyn_cast<ConstantInt>(Cmp0->getOperand(1));
748 ICmpInst::Predicate Pred0 = (Inverted ? Cmp0->getInversePredicate() :
749 Cmp0->getPredicate());
751 // Accept x > -1 or x >= 0 (after potentially inverting the predicate).
752 if (!((Pred0 == ICmpInst::ICMP_SGT && RangeStart->isMinusOne()) ||
753 (Pred0 == ICmpInst::ICMP_SGE && RangeStart->isZero())))
756 ICmpInst::Predicate Pred1 = (Inverted ? Cmp1->getInversePredicate() :
757 Cmp1->getPredicate());
759 Value *Input = Cmp0->getOperand(0);
761 if (Cmp1->getOperand(0) == Input) {
762 // For the upper range compare we have: icmp x, n
763 RangeEnd = Cmp1->getOperand(1);
764 } else if (Cmp1->getOperand(1) == Input) {
765 // For the upper range compare we have: icmp n, x
766 RangeEnd = Cmp1->getOperand(0);
767 Pred1 = ICmpInst::getSwappedPredicate(Pred1);
772 // Check the upper range comparison, e.g. x < n
773 ICmpInst::Predicate NewPred;
775 case ICmpInst::ICMP_SLT: NewPred = ICmpInst::ICMP_ULT; break;
776 case ICmpInst::ICMP_SLE: NewPred = ICmpInst::ICMP_ULE; break;
777 default: return nullptr;
780 // This simplification is only valid if the upper range is not negative.
781 bool IsNegative, IsNotNegative;
782 ComputeSignBit(RangeEnd, IsNotNegative, IsNegative, /*Depth=*/0, Cmp1);
787 NewPred = ICmpInst::getInversePredicate(NewPred);
789 return Builder->CreateICmp(NewPred, Input, RangeEnd);
792 /// Fold (icmp)&(icmp) if possible.
793 Value *InstCombiner::FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
794 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
796 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
797 if (PredicatesFoldable(LHSCC, RHSCC)) {
798 if (LHS->getOperand(0) == RHS->getOperand(1) &&
799 LHS->getOperand(1) == RHS->getOperand(0))
801 if (LHS->getOperand(0) == RHS->getOperand(0) &&
802 LHS->getOperand(1) == RHS->getOperand(1)) {
803 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
804 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
805 bool isSigned = LHS->isSigned() || RHS->isSigned();
806 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
810 // handle (roughly): (icmp eq (A & B), C) & (icmp eq (A & D), E)
811 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, true, Builder))
814 // E.g. (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
815 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/false))
818 // E.g. (icmp slt x, n) & (icmp sge x, 0) --> icmp ult x, n
819 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/false))
822 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
823 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
824 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
825 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
826 if (!LHSCst || !RHSCst) return nullptr;
828 if (LHSCst == RHSCst && LHSCC == RHSCC) {
829 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
830 // where C is a power of 2 or
831 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
832 if ((LHSCC == ICmpInst::ICMP_ULT && LHSCst->getValue().isPowerOf2()) ||
833 (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero())) {
834 Value *NewOr = Builder->CreateOr(Val, Val2);
835 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
839 // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2
840 // where CMAX is the all ones value for the truncated type,
841 // iff the lower bits of C2 and CA are zero.
842 if (LHSCC == ICmpInst::ICMP_EQ && LHSCC == RHSCC &&
843 LHS->hasOneUse() && RHS->hasOneUse()) {
845 ConstantInt *AndCst, *SmallCst = nullptr, *BigCst = nullptr;
847 // (trunc x) == C1 & (and x, CA) == C2
848 // (and x, CA) == C2 & (trunc x) == C1
849 if (match(Val2, m_Trunc(m_Value(V))) &&
850 match(Val, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
853 } else if (match(Val, m_Trunc(m_Value(V))) &&
854 match(Val2, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
859 if (SmallCst && BigCst) {
860 unsigned BigBitSize = BigCst->getType()->getBitWidth();
861 unsigned SmallBitSize = SmallCst->getType()->getBitWidth();
863 // Check that the low bits are zero.
864 APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize);
865 if ((Low & AndCst->getValue()) == 0 && (Low & BigCst->getValue()) == 0) {
866 Value *NewAnd = Builder->CreateAnd(V, Low | AndCst->getValue());
867 APInt N = SmallCst->getValue().zext(BigBitSize) | BigCst->getValue();
868 Value *NewVal = ConstantInt::get(AndCst->getType()->getContext(), N);
869 return Builder->CreateICmp(LHSCC, NewAnd, NewVal);
874 // From here on, we only handle:
875 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
876 if (Val != Val2) return nullptr;
878 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
879 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
880 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
881 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
882 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
885 // We can't fold (ugt x, C) & (sgt x, C2).
886 if (!PredicatesFoldable(LHSCC, RHSCC))
889 // Ensure that the larger constant is on the RHS.
891 if (CmpInst::isSigned(LHSCC) ||
892 (ICmpInst::isEquality(LHSCC) &&
893 CmpInst::isSigned(RHSCC)))
894 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
896 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
900 std::swap(LHSCst, RHSCst);
901 std::swap(LHSCC, RHSCC);
904 // At this point, we know we have two icmp instructions
905 // comparing a value against two constants and and'ing the result
906 // together. Because of the above check, we know that we only have
907 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
908 // (from the icmp folding check above), that the two constants
909 // are not equal and that the larger constant is on the RHS
910 assert(LHSCst != RHSCst && "Compares not folded above?");
913 default: llvm_unreachable("Unknown integer condition code!");
914 case ICmpInst::ICMP_EQ:
916 default: llvm_unreachable("Unknown integer condition code!");
917 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
918 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
919 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
922 case ICmpInst::ICMP_NE:
924 default: llvm_unreachable("Unknown integer condition code!");
925 case ICmpInst::ICMP_ULT:
926 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
927 return Builder->CreateICmpULT(Val, LHSCst);
928 if (LHSCst->isNullValue()) // (X != 0 & X u< 14) -> X-1 u< 13
929 return insertRangeTest(Val, LHSCst->getValue() + 1, RHSCst->getValue(),
931 break; // (X != 13 & X u< 15) -> no change
932 case ICmpInst::ICMP_SLT:
933 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
934 return Builder->CreateICmpSLT(Val, LHSCst);
935 break; // (X != 13 & X s< 15) -> no change
936 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
937 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
938 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
940 case ICmpInst::ICMP_NE:
941 // Special case to get the ordering right when the values wrap around
943 if (LHSCst->getValue() == 0 && RHSCst->getValue().isAllOnesValue())
944 std::swap(LHSCst, RHSCst);
945 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
946 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
947 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
948 return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1),
949 Val->getName()+".cmp");
951 break; // (X != 13 & X != 15) -> no change
954 case ICmpInst::ICMP_ULT:
956 default: llvm_unreachable("Unknown integer condition code!");
957 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
958 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
959 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
960 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
962 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
963 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
965 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
969 case ICmpInst::ICMP_SLT:
971 default: llvm_unreachable("Unknown integer condition code!");
972 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
974 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
975 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
977 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
981 case ICmpInst::ICMP_UGT:
983 default: llvm_unreachable("Unknown integer condition code!");
984 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
985 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
987 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
989 case ICmpInst::ICMP_NE:
990 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
991 return Builder->CreateICmp(LHSCC, Val, RHSCst);
992 break; // (X u> 13 & X != 15) -> no change
993 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
994 return insertRangeTest(Val, LHSCst->getValue() + 1, RHSCst->getValue(),
996 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
1000 case ICmpInst::ICMP_SGT:
1002 default: llvm_unreachable("Unknown integer condition code!");
1003 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
1004 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
1006 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
1008 case ICmpInst::ICMP_NE:
1009 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
1010 return Builder->CreateICmp(LHSCC, Val, RHSCst);
1011 break; // (X s> 13 & X != 15) -> no change
1012 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
1013 return insertRangeTest(Val, LHSCst->getValue() + 1, RHSCst->getValue(),
1015 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
1024 /// Optimize (fcmp)&(fcmp). NOTE: Unlike the rest of instcombine, this returns
1025 /// a Value which should already be inserted into the function.
1026 Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
1027 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
1028 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
1029 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
1031 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
1032 // Swap RHS operands to match LHS.
1033 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
1034 std::swap(Op1LHS, Op1RHS);
1037 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
1038 // Suppose the relation between x and y is R, where R is one of
1039 // U(1000), L(0100), G(0010) or E(0001), and CC0 and CC1 are the bitmasks for
1040 // testing the desired relations.
1042 // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this:
1043 // bool(R & CC0) && bool(R & CC1)
1044 // = bool((R & CC0) & (R & CC1))
1045 // = bool(R & (CC0 & CC1)) <= by re-association, commutation, and idempotency
1046 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS)
1047 return getFCmpValue(getFCmpCode(Op0CC) & getFCmpCode(Op1CC), Op0LHS, Op0RHS,
1050 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
1051 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
1052 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType())
1055 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
1056 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
1057 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
1058 // If either of the constants are nans, then the whole thing returns
1060 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
1061 return Builder->getFalse();
1062 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
1065 // Handle vector zeros. This occurs because the canonical form of
1066 // "fcmp ord x,x" is "fcmp ord x, 0".
1067 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
1068 isa<ConstantAggregateZero>(RHS->getOperand(1)))
1069 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
1076 /// Match De Morgan's Laws:
1077 /// (~A & ~B) == (~(A | B))
1078 /// (~A | ~B) == (~(A & B))
1079 static Instruction *matchDeMorgansLaws(BinaryOperator &I,
1080 InstCombiner::BuilderTy *Builder) {
1081 auto Opcode = I.getOpcode();
1082 assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&
1083 "Trying to match De Morgan's Laws with something other than and/or");
1084 // Flip the logic operation.
1085 if (Opcode == Instruction::And)
1086 Opcode = Instruction::Or;
1088 Opcode = Instruction::And;
1090 Value *Op0 = I.getOperand(0);
1091 Value *Op1 = I.getOperand(1);
1092 // TODO: Use pattern matchers instead of dyn_cast.
1093 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1094 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1095 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1096 Value *LogicOp = Builder->CreateBinOp(Opcode, Op0NotVal, Op1NotVal,
1097 I.getName() + ".demorgan");
1098 return BinaryOperator::CreateNot(LogicOp);
1104 bool InstCombiner::shouldOptimizeCast(CastInst *CI) {
1105 Value *CastSrc = CI->getOperand(0);
1107 // Noop casts and casts of constants should be eliminated trivially.
1108 if (CI->getSrcTy() == CI->getDestTy() || isa<Constant>(CastSrc))
1111 // If this cast is paired with another cast that can be eliminated, we prefer
1112 // to have it eliminated.
1113 if (const auto *PrecedingCI = dyn_cast<CastInst>(CastSrc))
1114 if (isEliminableCastPair(PrecedingCI, CI))
1117 // If this is a vector sext from a compare, then we don't want to break the
1118 // idiom where each element of the extended vector is either zero or all ones.
1119 if (CI->getOpcode() == Instruction::SExt &&
1120 isa<CmpInst>(CastSrc) && CI->getDestTy()->isVectorTy())
1126 /// Fold {and,or,xor} (cast X), C.
1127 static Instruction *foldLogicCastConstant(BinaryOperator &Logic, CastInst *Cast,
1128 InstCombiner::BuilderTy *Builder) {
1130 if (!match(Logic.getOperand(1), m_Constant(C)))
1133 auto LogicOpc = Logic.getOpcode();
1134 Type *DestTy = Logic.getType();
1135 Type *SrcTy = Cast->getSrcTy();
1137 // If the first operand is bitcast, move the logic operation ahead of the
1138 // bitcast (do the logic operation in the original type). This can eliminate
1139 // bitcasts and allow combines that would otherwise be impeded by the bitcast.
1141 if (match(Cast, m_BitCast(m_Value(X)))) {
1142 Value *NewConstant = ConstantExpr::getBitCast(C, SrcTy);
1143 Value *NewOp = Builder->CreateBinOp(LogicOpc, X, NewConstant);
1144 return CastInst::CreateBitOrPointerCast(NewOp, DestTy);
1147 // Similarly, move the logic operation ahead of a zext if the constant is
1148 // unchanged in the smaller source type. Performing the logic in a smaller
1149 // type may provide more information to later folds, and the smaller logic
1150 // instruction may be cheaper (particularly in the case of vectors).
1151 if (match(Cast, m_OneUse(m_ZExt(m_Value(X))))) {
1152 Constant *TruncC = ConstantExpr::getTrunc(C, SrcTy);
1153 Constant *ZextTruncC = ConstantExpr::getZExt(TruncC, DestTy);
1154 if (ZextTruncC == C) {
1155 // LogicOpc (zext X), C --> zext (LogicOpc X, C)
1156 Value *NewOp = Builder->CreateBinOp(LogicOpc, X, TruncC);
1157 return new ZExtInst(NewOp, DestTy);
1164 /// Fold {and,or,xor} (cast X), Y.
1165 Instruction *InstCombiner::foldCastedBitwiseLogic(BinaryOperator &I) {
1166 auto LogicOpc = I.getOpcode();
1167 assert(I.isBitwiseLogicOp() && "Unexpected opcode for bitwise logic folding");
1169 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1170 CastInst *Cast0 = dyn_cast<CastInst>(Op0);
1174 // This must be a cast from an integer or integer vector source type to allow
1175 // transformation of the logic operation to the source type.
1176 Type *DestTy = I.getType();
1177 Type *SrcTy = Cast0->getSrcTy();
1178 if (!SrcTy->isIntOrIntVectorTy())
1181 if (Instruction *Ret = foldLogicCastConstant(I, Cast0, Builder))
1184 CastInst *Cast1 = dyn_cast<CastInst>(Op1);
1188 // Both operands of the logic operation are casts. The casts must be of the
1189 // same type for reduction.
1190 auto CastOpcode = Cast0->getOpcode();
1191 if (CastOpcode != Cast1->getOpcode() || SrcTy != Cast1->getSrcTy())
1194 Value *Cast0Src = Cast0->getOperand(0);
1195 Value *Cast1Src = Cast1->getOperand(0);
1197 // fold logic(cast(A), cast(B)) -> cast(logic(A, B))
1198 if (shouldOptimizeCast(Cast0) && shouldOptimizeCast(Cast1)) {
1199 Value *NewOp = Builder->CreateBinOp(LogicOpc, Cast0Src, Cast1Src,
1201 return CastInst::Create(CastOpcode, NewOp, DestTy);
1204 // For now, only 'and'/'or' have optimizations after this.
1205 if (LogicOpc == Instruction::Xor)
1208 // If this is logic(cast(icmp), cast(icmp)), try to fold this even if the
1209 // cast is otherwise not optimizable. This happens for vector sexts.
1210 ICmpInst *ICmp0 = dyn_cast<ICmpInst>(Cast0Src);
1211 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(Cast1Src);
1212 if (ICmp0 && ICmp1) {
1213 Value *Res = LogicOpc == Instruction::And ? FoldAndOfICmps(ICmp0, ICmp1)
1214 : FoldOrOfICmps(ICmp0, ICmp1, &I);
1216 return CastInst::Create(CastOpcode, Res, DestTy);
1220 // If this is logic(cast(fcmp), cast(fcmp)), try to fold this even if the
1221 // cast is otherwise not optimizable. This happens for vector sexts.
1222 FCmpInst *FCmp0 = dyn_cast<FCmpInst>(Cast0Src);
1223 FCmpInst *FCmp1 = dyn_cast<FCmpInst>(Cast1Src);
1224 if (FCmp0 && FCmp1) {
1225 Value *Res = LogicOpc == Instruction::And ? FoldAndOfFCmps(FCmp0, FCmp1)
1226 : FoldOrOfFCmps(FCmp0, FCmp1);
1228 return CastInst::Create(CastOpcode, Res, DestTy);
1235 static Instruction *foldBoolSextMaskToSelect(BinaryOperator &I) {
1236 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1238 // Canonicalize SExt or Not to the LHS
1239 if (match(Op1, m_SExt(m_Value())) || match(Op1, m_Not(m_Value()))) {
1240 std::swap(Op0, Op1);
1243 // Fold (and (sext bool to A), B) --> (select bool, B, 0)
1245 if (match(Op0, m_SExt(m_Value(X))) &&
1246 X->getType()->getScalarType()->isIntegerTy(1)) {
1247 Value *Zero = Constant::getNullValue(Op1->getType());
1248 return SelectInst::Create(X, Op1, Zero);
1251 // Fold (and ~(sext bool to A), B) --> (select bool, 0, B)
1252 if (match(Op0, m_Not(m_SExt(m_Value(X)))) &&
1253 X->getType()->getScalarType()->isIntegerTy(1)) {
1254 Value *Zero = Constant::getNullValue(Op0->getType());
1255 return SelectInst::Create(X, Zero, Op1);
1261 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
1262 // here. We should standardize that construct where it is needed or choose some
1263 // other way to ensure that commutated variants of patterns are not missed.
1264 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
1265 bool Changed = SimplifyAssociativeOrCommutative(I);
1266 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1268 if (Value *V = SimplifyVectorOp(I))
1269 return replaceInstUsesWith(I, V);
1271 if (Value *V = SimplifyAndInst(Op0, Op1, DL, &TLI, &DT, &AC))
1272 return replaceInstUsesWith(I, V);
1274 // (A|B)&(A|C) -> A|(B&C) etc
1275 if (Value *V = SimplifyUsingDistributiveLaws(I))
1276 return replaceInstUsesWith(I, V);
1278 // See if we can simplify any instructions used by the instruction whose sole
1279 // purpose is to compute bits we don't care about.
1280 if (SimplifyDemandedInstructionBits(I))
1283 if (Value *V = SimplifyBSwap(I))
1284 return replaceInstUsesWith(I, V);
1286 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
1287 const APInt &AndRHSMask = AndRHS->getValue();
1289 // Optimize a variety of ((val OP C1) & C2) combinations...
1290 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1291 Value *Op0LHS = Op0I->getOperand(0);
1292 Value *Op0RHS = Op0I->getOperand(1);
1293 switch (Op0I->getOpcode()) {
1295 case Instruction::Xor:
1296 case Instruction::Or: {
1297 // If the mask is only needed on one incoming arm, push it up.
1298 if (!Op0I->hasOneUse()) break;
1300 APInt NotAndRHS(~AndRHSMask);
1301 if (MaskedValueIsZero(Op0LHS, NotAndRHS, 0, &I)) {
1302 // Not masking anything out for the LHS, move to RHS.
1303 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
1304 Op0RHS->getName()+".masked");
1305 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
1307 if (!isa<Constant>(Op0RHS) &&
1308 MaskedValueIsZero(Op0RHS, NotAndRHS, 0, &I)) {
1309 // Not masking anything out for the RHS, move to LHS.
1310 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
1311 Op0LHS->getName()+".masked");
1312 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
1317 case Instruction::Add:
1318 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
1319 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1320 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1321 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
1322 return BinaryOperator::CreateAnd(V, AndRHS);
1323 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
1324 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
1327 case Instruction::Sub:
1328 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
1329 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1330 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1331 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
1332 return BinaryOperator::CreateAnd(V, AndRHS);
1335 if (AndRHSMask == 1 && match(Op0LHS, m_Zero()))
1336 return BinaryOperator::CreateAnd(Op0RHS, AndRHS);
1338 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
1339 // has 1's for all bits that the subtraction with A might affect.
1340 if (Op0I->hasOneUse() && !match(Op0LHS, m_Zero())) {
1341 uint32_t BitWidth = AndRHSMask.getBitWidth();
1342 uint32_t Zeros = AndRHSMask.countLeadingZeros();
1343 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
1345 if (MaskedValueIsZero(Op0LHS, Mask, 0, &I)) {
1346 Value *NewNeg = Builder->CreateNeg(Op0RHS);
1347 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
1352 case Instruction::Shl:
1353 case Instruction::LShr:
1354 // (1 << x) & 1 --> zext(x == 0)
1355 // (1 >> x) & 1 --> zext(x == 0)
1356 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
1358 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
1359 return new ZExtInst(NewICmp, I.getType());
1364 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
1365 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
1369 // If this is an integer truncation, and if the source is an 'and' with
1370 // immediate, transform it. This frequently occurs for bitfield accesses.
1372 Value *X = nullptr; ConstantInt *YC = nullptr;
1373 if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) {
1374 // Change: and (trunc (and X, YC) to T), C2
1375 // into : and (trunc X to T), trunc(YC) & C2
1376 // This will fold the two constants together, which may allow
1377 // other simplifications.
1378 Value *NewCast = Builder->CreateTrunc(X, I.getType(), "and.shrunk");
1379 Constant *C3 = ConstantExpr::getTrunc(YC, I.getType());
1380 C3 = ConstantExpr::getAnd(C3, AndRHS);
1381 return BinaryOperator::CreateAnd(NewCast, C3);
1385 // Try to fold constant and into select arguments.
1386 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1387 if (Instruction *R = FoldOpIntoSelect(I, SI))
1389 if (isa<PHINode>(Op0))
1390 if (Instruction *NV = FoldOpIntoPhi(I))
1394 if (Instruction *DeMorgan = matchDeMorgansLaws(I, Builder))
1398 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr;
1399 // (A|B) & ~(A&B) -> A^B
1400 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1401 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1402 ((A == C && B == D) || (A == D && B == C)))
1403 return BinaryOperator::CreateXor(A, B);
1405 // ~(A&B) & (A|B) -> A^B
1406 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1407 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1408 ((A == C && B == D) || (A == D && B == C)))
1409 return BinaryOperator::CreateXor(A, B);
1411 // A&(A^B) => A & ~B
1413 Value *tmpOp0 = Op0;
1414 Value *tmpOp1 = Op1;
1415 if (match(Op0, m_OneUse(m_Xor(m_Value(A), m_Value(B))))) {
1416 if (A == Op1 || B == Op1 ) {
1423 if (match(tmpOp1, m_OneUse(m_Xor(m_Value(A), m_Value(B))))) {
1427 // Notice that the pattern (A&(~B)) is actually (A&(-1^B)), so if
1428 // A is originally -1 (or a vector of -1 and undefs), then we enter
1429 // an endless loop. By checking that A is non-constant we ensure that
1430 // we will never get to the loop.
1431 if (A == tmpOp0 && !isa<Constant>(A)) // A&(A^B) -> A & ~B
1432 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B));
1436 // (A&((~A)|B)) -> A&B
1437 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
1438 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
1439 return BinaryOperator::CreateAnd(A, Op1);
1440 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
1441 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
1442 return BinaryOperator::CreateAnd(A, Op0);
1444 // (A ^ B) & ((B ^ C) ^ A) -> (A ^ B) & ~C
1445 if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
1446 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
1447 if (Op1->hasOneUse() || cast<BinaryOperator>(Op1)->hasOneUse())
1448 return BinaryOperator::CreateAnd(Op0, Builder->CreateNot(C));
1450 // ((A ^ C) ^ B) & (B ^ A) -> (B ^ A) & ~C
1451 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
1452 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
1453 if (Op0->hasOneUse() || cast<BinaryOperator>(Op0)->hasOneUse())
1454 return BinaryOperator::CreateAnd(Op1, Builder->CreateNot(C));
1456 // (A | B) & ((~A) ^ B) -> (A & B)
1457 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1458 match(Op1, m_Xor(m_Not(m_Specific(A)), m_Specific(B))))
1459 return BinaryOperator::CreateAnd(A, B);
1461 // ((~A) ^ B) & (A | B) -> (A & B)
1462 // ((~A) ^ B) & (B | A) -> (A & B)
1463 if (match(Op0, m_Xor(m_Not(m_Value(A)), m_Value(B))) &&
1464 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
1465 return BinaryOperator::CreateAnd(A, B);
1469 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
1470 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
1472 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1473 return replaceInstUsesWith(I, Res);
1475 // TODO: Make this recursive; it's a little tricky because an arbitrary
1476 // number of 'and' instructions might have to be created.
1478 if (LHS && match(Op1, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
1479 if (auto *Cmp = dyn_cast<ICmpInst>(X))
1480 if (Value *Res = FoldAndOfICmps(LHS, Cmp))
1481 return replaceInstUsesWith(I, Builder->CreateAnd(Res, Y));
1482 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
1483 if (Value *Res = FoldAndOfICmps(LHS, Cmp))
1484 return replaceInstUsesWith(I, Builder->CreateAnd(Res, X));
1486 if (RHS && match(Op0, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
1487 if (auto *Cmp = dyn_cast<ICmpInst>(X))
1488 if (Value *Res = FoldAndOfICmps(Cmp, RHS))
1489 return replaceInstUsesWith(I, Builder->CreateAnd(Res, Y));
1490 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
1491 if (Value *Res = FoldAndOfICmps(Cmp, RHS))
1492 return replaceInstUsesWith(I, Builder->CreateAnd(Res, X));
1496 // If and'ing two fcmp, try combine them into one.
1497 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1498 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1499 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1500 return replaceInstUsesWith(I, Res);
1502 if (Instruction *CastedAnd = foldCastedBitwiseLogic(I))
1505 if (Instruction *Select = foldBoolSextMaskToSelect(I))
1508 return Changed ? &I : nullptr;
1511 /// Given an OR instruction, check to see if this is a bswap idiom. If so,
1512 /// insert the new intrinsic and return it.
1513 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
1514 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1516 // Look through zero extends.
1517 if (Instruction *Ext = dyn_cast<ZExtInst>(Op0))
1518 Op0 = Ext->getOperand(0);
1520 if (Instruction *Ext = dyn_cast<ZExtInst>(Op1))
1521 Op1 = Ext->getOperand(0);
1523 // (A | B) | C and A | (B | C) -> bswap if possible.
1524 bool OrOfOrs = match(Op0, m_Or(m_Value(), m_Value())) ||
1525 match(Op1, m_Or(m_Value(), m_Value()));
1527 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
1528 bool OrOfShifts = match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
1529 match(Op1, m_LogicalShift(m_Value(), m_Value()));
1531 // (A & B) | (C & D) -> bswap if possible.
1532 bool OrOfAnds = match(Op0, m_And(m_Value(), m_Value())) &&
1533 match(Op1, m_And(m_Value(), m_Value()));
1535 if (!OrOfOrs && !OrOfShifts && !OrOfAnds)
1538 SmallVector<Instruction*, 4> Insts;
1539 if (!recognizeBSwapOrBitReverseIdiom(&I, true, false, Insts))
1541 Instruction *LastInst = Insts.pop_back_val();
1542 LastInst->removeFromParent();
1544 for (auto *Inst : Insts)
1549 /// If all elements of two constant vectors are 0/-1 and inverses, return true.
1550 static bool areInverseVectorBitmasks(Constant *C1, Constant *C2) {
1551 unsigned NumElts = C1->getType()->getVectorNumElements();
1552 for (unsigned i = 0; i != NumElts; ++i) {
1553 Constant *EltC1 = C1->getAggregateElement(i);
1554 Constant *EltC2 = C2->getAggregateElement(i);
1555 if (!EltC1 || !EltC2)
1558 // One element must be all ones, and the other must be all zeros.
1559 // FIXME: Allow undef elements.
1560 if (!((match(EltC1, m_Zero()) && match(EltC2, m_AllOnes())) ||
1561 (match(EltC2, m_Zero()) && match(EltC1, m_AllOnes()))))
1567 /// We have an expression of the form (A & C) | (B & D). If A is a scalar or
1568 /// vector composed of all-zeros or all-ones values and is the bitwise 'not' of
1569 /// B, it can be used as the condition operand of a select instruction.
1570 static Value *getSelectCondition(Value *A, Value *B,
1571 InstCombiner::BuilderTy &Builder) {
1572 // If these are scalars or vectors of i1, A can be used directly.
1573 Type *Ty = A->getType();
1574 if (match(A, m_Not(m_Specific(B))) && Ty->getScalarType()->isIntegerTy(1))
1577 // If A and B are sign-extended, look through the sexts to find the booleans.
1579 if (match(A, m_SExt(m_Value(Cond))) &&
1580 Cond->getType()->getScalarType()->isIntegerTy(1) &&
1581 match(B, m_CombineOr(m_Not(m_SExt(m_Specific(Cond))),
1582 m_SExt(m_Not(m_Specific(Cond))))))
1585 // All scalar (and most vector) possibilities should be handled now.
1586 // Try more matches that only apply to non-splat constant vectors.
1587 if (!Ty->isVectorTy())
1590 // If both operands are constants, see if the constants are inverse bitmasks.
1592 if (match(A, m_Constant(AC)) && match(B, m_Constant(BC)) &&
1593 areInverseVectorBitmasks(AC, BC))
1594 return ConstantExpr::getTrunc(AC, CmpInst::makeCmpResultType(Ty));
1596 // If both operands are xor'd with constants using the same sexted boolean
1597 // operand, see if the constants are inverse bitmasks.
1598 if (match(A, (m_Xor(m_SExt(m_Value(Cond)), m_Constant(AC)))) &&
1599 match(B, (m_Xor(m_SExt(m_Specific(Cond)), m_Constant(BC)))) &&
1600 Cond->getType()->getScalarType()->isIntegerTy(1) &&
1601 areInverseVectorBitmasks(AC, BC)) {
1602 AC = ConstantExpr::getTrunc(AC, CmpInst::makeCmpResultType(Ty));
1603 return Builder.CreateXor(Cond, AC);
1608 /// We have an expression of the form (A & C) | (B & D). Try to simplify this
1609 /// to "A' ? C : D", where A' is a boolean or vector of booleans.
1610 static Value *matchSelectFromAndOr(Value *A, Value *C, Value *B, Value *D,
1611 InstCombiner::BuilderTy &Builder) {
1612 // The potential condition of the select may be bitcasted. In that case, look
1613 // through its bitcast and the corresponding bitcast of the 'not' condition.
1614 Type *OrigType = A->getType();
1616 if (match(A, m_OneUse(m_BitCast(m_Value(SrcA)))) &&
1617 match(B, m_OneUse(m_BitCast(m_Value(SrcB))))) {
1622 if (Value *Cond = getSelectCondition(A, B, Builder)) {
1623 // ((bc Cond) & C) | ((bc ~Cond) & D) --> bc (select Cond, (bc C), (bc D))
1624 // The bitcasts will either all exist or all not exist. The builder will
1625 // not create unnecessary casts if the types already match.
1626 Value *BitcastC = Builder.CreateBitCast(C, A->getType());
1627 Value *BitcastD = Builder.CreateBitCast(D, A->getType());
1628 Value *Select = Builder.CreateSelect(Cond, BitcastC, BitcastD);
1629 return Builder.CreateBitCast(Select, OrigType);
1635 /// Fold (icmp)|(icmp) if possible.
1636 Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS,
1637 Instruction *CxtI) {
1638 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
1640 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2)
1641 // if K1 and K2 are a one-bit mask.
1642 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
1643 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
1645 if (LHS->getPredicate() == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero() &&
1646 RHS->getPredicate() == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) {
1648 BinaryOperator *LAnd = dyn_cast<BinaryOperator>(LHS->getOperand(0));
1649 BinaryOperator *RAnd = dyn_cast<BinaryOperator>(RHS->getOperand(0));
1650 if (LAnd && RAnd && LAnd->hasOneUse() && RHS->hasOneUse() &&
1651 LAnd->getOpcode() == Instruction::And &&
1652 RAnd->getOpcode() == Instruction::And) {
1654 Value *Mask = nullptr;
1655 Value *Masked = nullptr;
1656 if (LAnd->getOperand(0) == RAnd->getOperand(0) &&
1657 isKnownToBeAPowerOfTwo(LAnd->getOperand(1), DL, false, 0, &AC, CxtI,
1659 isKnownToBeAPowerOfTwo(RAnd->getOperand(1), DL, false, 0, &AC, CxtI,
1661 Mask = Builder->CreateOr(LAnd->getOperand(1), RAnd->getOperand(1));
1662 Masked = Builder->CreateAnd(LAnd->getOperand(0), Mask);
1663 } else if (LAnd->getOperand(1) == RAnd->getOperand(1) &&
1664 isKnownToBeAPowerOfTwo(LAnd->getOperand(0), DL, false, 0, &AC,
1666 isKnownToBeAPowerOfTwo(RAnd->getOperand(0), DL, false, 0, &AC,
1668 Mask = Builder->CreateOr(LAnd->getOperand(0), RAnd->getOperand(0));
1669 Masked = Builder->CreateAnd(LAnd->getOperand(1), Mask);
1673 return Builder->CreateICmp(ICmpInst::ICMP_NE, Masked, Mask);
1677 // Fold (icmp ult/ule (A + C1), C3) | (icmp ult/ule (A + C2), C3)
1678 // --> (icmp ult/ule ((A & ~(C1 ^ C2)) + max(C1, C2)), C3)
1679 // The original condition actually refers to the following two ranges:
1680 // [MAX_UINT-C1+1, MAX_UINT-C1+1+C3] and [MAX_UINT-C2+1, MAX_UINT-C2+1+C3]
1681 // We can fold these two ranges if:
1682 // 1) C1 and C2 is unsigned greater than C3.
1683 // 2) The two ranges are separated.
1684 // 3) C1 ^ C2 is one-bit mask.
1685 // 4) LowRange1 ^ LowRange2 and HighRange1 ^ HighRange2 are one-bit mask.
1686 // This implies all values in the two ranges differ by exactly one bit.
1688 if ((LHSCC == ICmpInst::ICMP_ULT || LHSCC == ICmpInst::ICMP_ULE) &&
1689 LHSCC == RHSCC && LHSCst && RHSCst && LHS->hasOneUse() &&
1690 RHS->hasOneUse() && LHSCst->getType() == RHSCst->getType() &&
1691 LHSCst->getValue() == (RHSCst->getValue())) {
1693 Value *LAdd = LHS->getOperand(0);
1694 Value *RAdd = RHS->getOperand(0);
1696 Value *LAddOpnd, *RAddOpnd;
1697 ConstantInt *LAddCst, *RAddCst;
1698 if (match(LAdd, m_Add(m_Value(LAddOpnd), m_ConstantInt(LAddCst))) &&
1699 match(RAdd, m_Add(m_Value(RAddOpnd), m_ConstantInt(RAddCst))) &&
1700 LAddCst->getValue().ugt(LHSCst->getValue()) &&
1701 RAddCst->getValue().ugt(LHSCst->getValue())) {
1703 APInt DiffCst = LAddCst->getValue() ^ RAddCst->getValue();
1704 if (LAddOpnd == RAddOpnd && DiffCst.isPowerOf2()) {
1705 ConstantInt *MaxAddCst = nullptr;
1706 if (LAddCst->getValue().ult(RAddCst->getValue()))
1707 MaxAddCst = RAddCst;
1709 MaxAddCst = LAddCst;
1711 APInt RRangeLow = -RAddCst->getValue();
1712 APInt RRangeHigh = RRangeLow + LHSCst->getValue();
1713 APInt LRangeLow = -LAddCst->getValue();
1714 APInt LRangeHigh = LRangeLow + LHSCst->getValue();
1715 APInt LowRangeDiff = RRangeLow ^ LRangeLow;
1716 APInt HighRangeDiff = RRangeHigh ^ LRangeHigh;
1717 APInt RangeDiff = LRangeLow.sgt(RRangeLow) ? LRangeLow - RRangeLow
1718 : RRangeLow - LRangeLow;
1720 if (LowRangeDiff.isPowerOf2() && LowRangeDiff == HighRangeDiff &&
1721 RangeDiff.ugt(LHSCst->getValue())) {
1722 Value *MaskCst = ConstantInt::get(LAddCst->getType(), ~DiffCst);
1724 Value *NewAnd = Builder->CreateAnd(LAddOpnd, MaskCst);
1725 Value *NewAdd = Builder->CreateAdd(NewAnd, MaxAddCst);
1726 return (Builder->CreateICmp(LHS->getPredicate(), NewAdd, LHSCst));
1732 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
1733 if (PredicatesFoldable(LHSCC, RHSCC)) {
1734 if (LHS->getOperand(0) == RHS->getOperand(1) &&
1735 LHS->getOperand(1) == RHS->getOperand(0))
1736 LHS->swapOperands();
1737 if (LHS->getOperand(0) == RHS->getOperand(0) &&
1738 LHS->getOperand(1) == RHS->getOperand(1)) {
1739 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1740 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
1741 bool isSigned = LHS->isSigned() || RHS->isSigned();
1742 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
1746 // handle (roughly):
1747 // (icmp ne (A & B), C) | (icmp ne (A & D), E)
1748 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, false, Builder))
1751 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
1752 if (LHS->hasOneUse() || RHS->hasOneUse()) {
1753 // (icmp eq B, 0) | (icmp ult A, B) -> (icmp ule A, B-1)
1754 // (icmp eq B, 0) | (icmp ugt B, A) -> (icmp ule A, B-1)
1755 Value *A = nullptr, *B = nullptr;
1756 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst && LHSCst->isZero()) {
1758 if (RHSCC == ICmpInst::ICMP_ULT && Val == RHS->getOperand(1))
1760 else if (RHSCC == ICmpInst::ICMP_UGT && Val == Val2)
1761 A = RHS->getOperand(1);
1763 // (icmp ult A, B) | (icmp eq B, 0) -> (icmp ule A, B-1)
1764 // (icmp ugt B, A) | (icmp eq B, 0) -> (icmp ule A, B-1)
1765 else if (RHSCC == ICmpInst::ICMP_EQ && RHSCst && RHSCst->isZero()) {
1767 if (LHSCC == ICmpInst::ICMP_ULT && Val2 == LHS->getOperand(1))
1769 else if (LHSCC == ICmpInst::ICMP_UGT && Val2 == Val)
1770 A = LHS->getOperand(1);
1773 return Builder->CreateICmp(
1775 Builder->CreateAdd(B, ConstantInt::getSigned(B->getType(), -1)), A);
1778 // E.g. (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
1779 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/true))
1782 // E.g. (icmp sgt x, n) | (icmp slt x, 0) --> icmp ugt x, n
1783 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/true))
1786 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
1787 if (!LHSCst || !RHSCst) return nullptr;
1789 if (LHSCst == RHSCst && LHSCC == RHSCC) {
1790 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
1791 if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
1792 Value *NewOr = Builder->CreateOr(Val, Val2);
1793 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
1797 // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
1798 // iff C2 + CA == C1.
1799 if (LHSCC == ICmpInst::ICMP_ULT && RHSCC == ICmpInst::ICMP_EQ) {
1800 ConstantInt *AddCst;
1801 if (match(Val, m_Add(m_Specific(Val2), m_ConstantInt(AddCst))))
1802 if (RHSCst->getValue() + AddCst->getValue() == LHSCst->getValue())
1803 return Builder->CreateICmpULE(Val, LHSCst);
1806 // From here on, we only handle:
1807 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
1808 if (Val != Val2) return nullptr;
1810 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
1811 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
1812 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
1813 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
1814 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
1817 // We can't fold (ugt x, C) | (sgt x, C2).
1818 if (!PredicatesFoldable(LHSCC, RHSCC))
1821 // Ensure that the larger constant is on the RHS.
1823 if (CmpInst::isSigned(LHSCC) ||
1824 (ICmpInst::isEquality(LHSCC) &&
1825 CmpInst::isSigned(RHSCC)))
1826 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
1828 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
1831 std::swap(LHS, RHS);
1832 std::swap(LHSCst, RHSCst);
1833 std::swap(LHSCC, RHSCC);
1836 // At this point, we know we have two icmp instructions
1837 // comparing a value against two constants and or'ing the result
1838 // together. Because of the above check, we know that we only have
1839 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
1840 // icmp folding check above), that the two constants are not
1842 assert(LHSCst != RHSCst && "Compares not folded above?");
1845 default: llvm_unreachable("Unknown integer condition code!");
1846 case ICmpInst::ICMP_EQ:
1848 default: llvm_unreachable("Unknown integer condition code!");
1849 case ICmpInst::ICMP_EQ:
1850 if (LHS->getOperand(0) == RHS->getOperand(0)) {
1851 // if LHSCst and RHSCst differ only by one bit:
1852 // (A == C1 || A == C2) -> (A | (C1 ^ C2)) == C2
1853 assert(LHSCst->getValue().ule(LHSCst->getValue()));
1855 APInt Xor = LHSCst->getValue() ^ RHSCst->getValue();
1856 if (Xor.isPowerOf2()) {
1857 Value *Cst = Builder->getInt(Xor);
1858 Value *Or = Builder->CreateOr(LHS->getOperand(0), Cst);
1859 return Builder->CreateICmp(ICmpInst::ICMP_EQ, Or, RHSCst);
1863 if (LHSCst == SubOne(RHSCst)) {
1864 // (X == 13 | X == 14) -> X-13 <u 2
1865 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1866 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1867 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
1868 return Builder->CreateICmpULT(Add, AddCST);
1871 break; // (X == 13 | X == 15) -> no change
1872 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
1873 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
1875 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
1876 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
1877 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
1881 case ICmpInst::ICMP_NE:
1883 default: llvm_unreachable("Unknown integer condition code!");
1884 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
1885 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
1886 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
1888 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
1889 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
1890 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
1891 return Builder->getTrue();
1893 case ICmpInst::ICMP_ULT:
1895 default: llvm_unreachable("Unknown integer condition code!");
1896 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
1898 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
1899 // If RHSCst is [us]MAXINT, it is always false. Not handling
1900 // this can cause overflow.
1901 if (RHSCst->isMaxValue(false))
1903 return insertRangeTest(Val, LHSCst->getValue(), RHSCst->getValue() + 1,
1905 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
1907 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
1908 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
1910 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
1914 case ICmpInst::ICMP_SLT:
1916 default: llvm_unreachable("Unknown integer condition code!");
1917 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
1919 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
1920 // If RHSCst is [us]MAXINT, it is always false. Not handling
1921 // this can cause overflow.
1922 if (RHSCst->isMaxValue(true))
1924 return insertRangeTest(Val, LHSCst->getValue(), RHSCst->getValue() + 1,
1926 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
1928 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
1929 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
1931 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
1935 case ICmpInst::ICMP_UGT:
1937 default: llvm_unreachable("Unknown integer condition code!");
1938 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
1939 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
1941 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
1943 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
1944 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
1945 return Builder->getTrue();
1946 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
1950 case ICmpInst::ICMP_SGT:
1952 default: llvm_unreachable("Unknown integer condition code!");
1953 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
1954 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
1956 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
1958 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
1959 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
1960 return Builder->getTrue();
1961 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
1969 /// Optimize (fcmp)|(fcmp). NOTE: Unlike the rest of instcombine, this returns
1970 /// a Value which should already be inserted into the function.
1971 Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
1972 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
1973 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
1974 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
1976 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
1977 // Swap RHS operands to match LHS.
1978 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
1979 std::swap(Op1LHS, Op1RHS);
1982 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
1983 // This is a similar transformation to the one in FoldAndOfFCmps.
1985 // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this:
1986 // bool(R & CC0) || bool(R & CC1)
1987 // = bool((R & CC0) | (R & CC1))
1988 // = bool(R & (CC0 | CC1)) <= by reversed distribution (contribution? ;)
1989 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS)
1990 return getFCmpValue(getFCmpCode(Op0CC) | getFCmpCode(Op1CC), Op0LHS, Op0RHS,
1993 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
1994 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
1995 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
1996 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
1997 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
1998 // If either of the constants are nans, then the whole thing returns
2000 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
2001 return Builder->getTrue();
2003 // Otherwise, no need to compare the two constants, compare the
2005 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
2008 // Handle vector zeros. This occurs because the canonical form of
2009 // "fcmp uno x,x" is "fcmp uno x, 0".
2010 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
2011 isa<ConstantAggregateZero>(RHS->getOperand(1)))
2012 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
2020 /// This helper function folds:
2022 /// ((A | B) & C1) | (B & C2)
2028 /// when the XOR of the two constants is "all ones" (-1).
2029 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
2030 Value *A, Value *B, Value *C) {
2031 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
2032 if (!CI1) return nullptr;
2034 Value *V1 = nullptr;
2035 ConstantInt *CI2 = nullptr;
2036 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return nullptr;
2038 APInt Xor = CI1->getValue() ^ CI2->getValue();
2039 if (!Xor.isAllOnesValue()) return nullptr;
2041 if (V1 == A || V1 == B) {
2042 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
2043 return BinaryOperator::CreateOr(NewOp, V1);
2049 /// \brief This helper function folds:
2051 /// ((A | B) & C1) ^ (B & C2)
2057 /// when the XOR of the two constants is "all ones" (-1).
2058 Instruction *InstCombiner::FoldXorWithConstants(BinaryOperator &I, Value *Op,
2059 Value *A, Value *B, Value *C) {
2060 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
2064 Value *V1 = nullptr;
2065 ConstantInt *CI2 = nullptr;
2066 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2))))
2069 APInt Xor = CI1->getValue() ^ CI2->getValue();
2070 if (!Xor.isAllOnesValue())
2073 if (V1 == A || V1 == B) {
2074 Value *NewOp = Builder->CreateAnd(V1 == A ? B : A, CI1);
2075 return BinaryOperator::CreateXor(NewOp, V1);
2081 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
2082 // here. We should standardize that construct where it is needed or choose some
2083 // other way to ensure that commutated variants of patterns are not missed.
2084 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
2085 bool Changed = SimplifyAssociativeOrCommutative(I);
2086 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2088 if (Value *V = SimplifyVectorOp(I))
2089 return replaceInstUsesWith(I, V);
2091 if (Value *V = SimplifyOrInst(Op0, Op1, DL, &TLI, &DT, &AC))
2092 return replaceInstUsesWith(I, V);
2094 // (A&B)|(A&C) -> A&(B|C) etc
2095 if (Value *V = SimplifyUsingDistributiveLaws(I))
2096 return replaceInstUsesWith(I, V);
2098 // See if we can simplify any instructions used by the instruction whose sole
2099 // purpose is to compute bits we don't care about.
2100 if (SimplifyDemandedInstructionBits(I))
2103 if (Value *V = SimplifyBSwap(I))
2104 return replaceInstUsesWith(I, V);
2106 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2107 ConstantInt *C1 = nullptr; Value *X = nullptr;
2108 // (X & C1) | C2 --> (X | C2) & (C1|C2)
2109 // iff (C1 & C2) == 0.
2110 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
2111 (RHS->getValue() & C1->getValue()) != 0 &&
2113 Value *Or = Builder->CreateOr(X, RHS);
2115 return BinaryOperator::CreateAnd(Or,
2116 Builder->getInt(RHS->getValue() | C1->getValue()));
2119 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
2120 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
2122 Value *Or = Builder->CreateOr(X, RHS);
2124 return BinaryOperator::CreateXor(Or,
2125 Builder->getInt(C1->getValue() & ~RHS->getValue()));
2128 // Try to fold constant and into select arguments.
2129 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2130 if (Instruction *R = FoldOpIntoSelect(I, SI))
2133 if (isa<PHINode>(Op0))
2134 if (Instruction *NV = FoldOpIntoPhi(I))
2138 // Given an OR instruction, check to see if this is a bswap.
2139 if (Instruction *BSwap = MatchBSwap(I))
2142 Value *A = nullptr, *B = nullptr;
2143 ConstantInt *C1 = nullptr, *C2 = nullptr;
2145 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
2146 if (Op0->hasOneUse() &&
2147 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2148 MaskedValueIsZero(Op1, C1->getValue(), 0, &I)) {
2149 Value *NOr = Builder->CreateOr(A, Op1);
2151 return BinaryOperator::CreateXor(NOr, C1);
2154 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
2155 if (Op1->hasOneUse() &&
2156 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
2157 MaskedValueIsZero(Op0, C1->getValue(), 0, &I)) {
2158 Value *NOr = Builder->CreateOr(A, Op0);
2160 return BinaryOperator::CreateXor(NOr, C1);
2163 // ((~A & B) | A) -> (A | B)
2164 if (match(Op0, m_And(m_Not(m_Value(A)), m_Value(B))) &&
2165 match(Op1, m_Specific(A)))
2166 return BinaryOperator::CreateOr(A, B);
2168 // ((A & B) | ~A) -> (~A | B)
2169 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2170 match(Op1, m_Not(m_Specific(A))))
2171 return BinaryOperator::CreateOr(Builder->CreateNot(A), B);
2173 // (A & ~B) | (A ^ B) -> (A ^ B)
2174 // (~B & A) | (A ^ B) -> (A ^ B)
2175 if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) &&
2176 match(Op1, m_Xor(m_Specific(A), m_Specific(B))))
2177 return BinaryOperator::CreateXor(A, B);
2179 // Commute the 'or' operands.
2180 // (A ^ B) | (A & ~B) -> (A ^ B)
2181 // (A ^ B) | (~B & A) -> (A ^ B)
2182 if (match(Op1, m_c_And(m_Value(A), m_Not(m_Value(B)))) &&
2183 match(Op0, m_Xor(m_Specific(A), m_Specific(B))))
2184 return BinaryOperator::CreateXor(A, B);
2187 Value *C = nullptr, *D = nullptr;
2188 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
2189 match(Op1, m_And(m_Value(B), m_Value(D)))) {
2190 Value *V1 = nullptr, *V2 = nullptr;
2191 C1 = dyn_cast<ConstantInt>(C);
2192 C2 = dyn_cast<ConstantInt>(D);
2193 if (C1 && C2) { // (A & C1)|(B & C2)
2194 if ((C1->getValue() & C2->getValue()) == 0) {
2195 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
2196 // iff (C1&C2) == 0 and (N&~C1) == 0
2197 if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
2199 MaskedValueIsZero(V2, ~C1->getValue(), 0, &I)) || // (V|N)
2201 MaskedValueIsZero(V1, ~C1->getValue(), 0, &I)))) // (N|V)
2202 return BinaryOperator::CreateAnd(A,
2203 Builder->getInt(C1->getValue()|C2->getValue()));
2204 // Or commutes, try both ways.
2205 if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
2207 MaskedValueIsZero(V2, ~C2->getValue(), 0, &I)) || // (V|N)
2209 MaskedValueIsZero(V1, ~C2->getValue(), 0, &I)))) // (N|V)
2210 return BinaryOperator::CreateAnd(B,
2211 Builder->getInt(C1->getValue()|C2->getValue()));
2213 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
2214 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
2215 ConstantInt *C3 = nullptr, *C4 = nullptr;
2216 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
2217 (C3->getValue() & ~C1->getValue()) == 0 &&
2218 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
2219 (C4->getValue() & ~C2->getValue()) == 0) {
2220 V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
2221 return BinaryOperator::CreateAnd(V2,
2222 Builder->getInt(C1->getValue()|C2->getValue()));
2227 // Don't try to form a select if it's unlikely that we'll get rid of at
2228 // least one of the operands. A select is generally more expensive than the
2229 // 'or' that it is replacing.
2230 if (Op0->hasOneUse() || Op1->hasOneUse()) {
2231 // (Cond & C) | (~Cond & D) -> Cond ? C : D, and commuted variants.
2232 if (Value *V = matchSelectFromAndOr(A, C, B, D, *Builder))
2233 return replaceInstUsesWith(I, V);
2234 if (Value *V = matchSelectFromAndOr(A, C, D, B, *Builder))
2235 return replaceInstUsesWith(I, V);
2236 if (Value *V = matchSelectFromAndOr(C, A, B, D, *Builder))
2237 return replaceInstUsesWith(I, V);
2238 if (Value *V = matchSelectFromAndOr(C, A, D, B, *Builder))
2239 return replaceInstUsesWith(I, V);
2240 if (Value *V = matchSelectFromAndOr(B, D, A, C, *Builder))
2241 return replaceInstUsesWith(I, V);
2242 if (Value *V = matchSelectFromAndOr(B, D, C, A, *Builder))
2243 return replaceInstUsesWith(I, V);
2244 if (Value *V = matchSelectFromAndOr(D, B, A, C, *Builder))
2245 return replaceInstUsesWith(I, V);
2246 if (Value *V = matchSelectFromAndOr(D, B, C, A, *Builder))
2247 return replaceInstUsesWith(I, V);
2250 // ((A&~B)|(~A&B)) -> A^B
2251 if ((match(C, m_Not(m_Specific(D))) &&
2252 match(B, m_Not(m_Specific(A)))))
2253 return BinaryOperator::CreateXor(A, D);
2254 // ((~B&A)|(~A&B)) -> A^B
2255 if ((match(A, m_Not(m_Specific(D))) &&
2256 match(B, m_Not(m_Specific(C)))))
2257 return BinaryOperator::CreateXor(C, D);
2258 // ((A&~B)|(B&~A)) -> A^B
2259 if ((match(C, m_Not(m_Specific(B))) &&
2260 match(D, m_Not(m_Specific(A)))))
2261 return BinaryOperator::CreateXor(A, B);
2262 // ((~B&A)|(B&~A)) -> A^B
2263 if ((match(A, m_Not(m_Specific(B))) &&
2264 match(D, m_Not(m_Specific(C)))))
2265 return BinaryOperator::CreateXor(C, B);
2267 // ((A|B)&1)|(B&-2) -> (A&1) | B
2268 if (match(A, m_Or(m_Value(V1), m_Specific(B))) ||
2269 match(A, m_Or(m_Specific(B), m_Value(V1)))) {
2270 Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C);
2271 if (Ret) return Ret;
2273 // (B&-2)|((A|B)&1) -> (A&1) | B
2274 if (match(B, m_Or(m_Specific(A), m_Value(V1))) ||
2275 match(B, m_Or(m_Value(V1), m_Specific(A)))) {
2276 Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D);
2277 if (Ret) return Ret;
2279 // ((A^B)&1)|(B&-2) -> (A&1) ^ B
2280 if (match(A, m_Xor(m_Value(V1), m_Specific(B))) ||
2281 match(A, m_Xor(m_Specific(B), m_Value(V1)))) {
2282 Instruction *Ret = FoldXorWithConstants(I, Op1, V1, B, C);
2283 if (Ret) return Ret;
2285 // (B&-2)|((A^B)&1) -> (A&1) ^ B
2286 if (match(B, m_Xor(m_Specific(A), m_Value(V1))) ||
2287 match(B, m_Xor(m_Value(V1), m_Specific(A)))) {
2288 Instruction *Ret = FoldXorWithConstants(I, Op0, A, V1, D);
2289 if (Ret) return Ret;
2293 // (A ^ B) | ((B ^ C) ^ A) -> (A ^ B) | C
2294 if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
2295 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
2296 if (Op1->hasOneUse() || cast<BinaryOperator>(Op1)->hasOneUse())
2297 return BinaryOperator::CreateOr(Op0, C);
2299 // ((A ^ C) ^ B) | (B ^ A) -> (B ^ A) | C
2300 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
2301 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
2302 if (Op0->hasOneUse() || cast<BinaryOperator>(Op0)->hasOneUse())
2303 return BinaryOperator::CreateOr(Op1, C);
2305 // ((B | C) & A) | B -> B | (A & C)
2306 if (match(Op0, m_And(m_Or(m_Specific(Op1), m_Value(C)), m_Value(A))))
2307 return BinaryOperator::CreateOr(Op1, Builder->CreateAnd(A, C));
2309 if (Instruction *DeMorgan = matchDeMorgansLaws(I, Builder))
2312 // Canonicalize xor to the RHS.
2313 bool SwappedForXor = false;
2314 if (match(Op0, m_Xor(m_Value(), m_Value()))) {
2315 std::swap(Op0, Op1);
2316 SwappedForXor = true;
2319 // A | ( A ^ B) -> A | B
2320 // A | (~A ^ B) -> A | ~B
2321 // (A & B) | (A ^ B)
2322 if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
2323 if (Op0 == A || Op0 == B)
2324 return BinaryOperator::CreateOr(A, B);
2326 if (match(Op0, m_And(m_Specific(A), m_Specific(B))) ||
2327 match(Op0, m_And(m_Specific(B), m_Specific(A))))
2328 return BinaryOperator::CreateOr(A, B);
2330 if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
2331 Value *Not = Builder->CreateNot(B, B->getName()+".not");
2332 return BinaryOperator::CreateOr(Not, Op0);
2334 if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
2335 Value *Not = Builder->CreateNot(A, A->getName()+".not");
2336 return BinaryOperator::CreateOr(Not, Op0);
2340 // A | ~(A | B) -> A | ~B
2341 // A | ~(A ^ B) -> A | ~B
2342 if (match(Op1, m_Not(m_Value(A))))
2343 if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
2344 if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
2345 Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
2346 B->getOpcode() == Instruction::Xor)) {
2347 Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
2349 Value *Not = Builder->CreateNot(NotOp, NotOp->getName()+".not");
2350 return BinaryOperator::CreateOr(Not, Op0);
2353 // (A & B) | (~A ^ B) -> (~A ^ B)
2354 // (A & B) | (B ^ ~A) -> (~A ^ B)
2355 // (B & A) | (~A ^ B) -> (~A ^ B)
2356 // (B & A) | (B ^ ~A) -> (~A ^ B)
2357 // The match order is important: match the xor first because the 'not'
2358 // operation defines 'A'. We do not need to match the xor as Op0 because the
2359 // xor was canonicalized to Op1 above.
2360 if (match(Op1, m_c_Xor(m_Not(m_Value(A)), m_Value(B))) &&
2361 match(Op0, m_c_And(m_Specific(A), m_Specific(B))))
2362 return BinaryOperator::CreateXor(Builder->CreateNot(A), B);
2365 std::swap(Op0, Op1);
2368 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
2369 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
2371 if (Value *Res = FoldOrOfICmps(LHS, RHS, &I))
2372 return replaceInstUsesWith(I, Res);
2374 // TODO: Make this recursive; it's a little tricky because an arbitrary
2375 // number of 'or' instructions might have to be created.
2377 if (LHS && match(Op1, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2378 if (auto *Cmp = dyn_cast<ICmpInst>(X))
2379 if (Value *Res = FoldOrOfICmps(LHS, Cmp, &I))
2380 return replaceInstUsesWith(I, Builder->CreateOr(Res, Y));
2381 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2382 if (Value *Res = FoldOrOfICmps(LHS, Cmp, &I))
2383 return replaceInstUsesWith(I, Builder->CreateOr(Res, X));
2385 if (RHS && match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2386 if (auto *Cmp = dyn_cast<ICmpInst>(X))
2387 if (Value *Res = FoldOrOfICmps(Cmp, RHS, &I))
2388 return replaceInstUsesWith(I, Builder->CreateOr(Res, Y));
2389 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2390 if (Value *Res = FoldOrOfICmps(Cmp, RHS, &I))
2391 return replaceInstUsesWith(I, Builder->CreateOr(Res, X));
2395 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
2396 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
2397 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
2398 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
2399 return replaceInstUsesWith(I, Res);
2401 if (Instruction *CastedOr = foldCastedBitwiseLogic(I))
2404 // or(sext(A), B) / or(B, sext(A)) --> A ? -1 : B, where A is i1 or <N x i1>.
2405 if (match(Op0, m_OneUse(m_SExt(m_Value(A)))) &&
2406 A->getType()->getScalarType()->isIntegerTy(1))
2407 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1);
2408 if (match(Op1, m_OneUse(m_SExt(m_Value(A)))) &&
2409 A->getType()->getScalarType()->isIntegerTy(1))
2410 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0);
2412 // Note: If we've gotten to the point of visiting the outer OR, then the
2413 // inner one couldn't be simplified. If it was a constant, then it won't
2414 // be simplified by a later pass either, so we try swapping the inner/outer
2415 // ORs in the hopes that we'll be able to simplify it this way.
2416 // (X|C) | V --> (X|V) | C
2417 if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
2418 match(Op0, m_Or(m_Value(A), m_ConstantInt(C1)))) {
2419 Value *Inner = Builder->CreateOr(A, Op1);
2420 Inner->takeName(Op0);
2421 return BinaryOperator::CreateOr(Inner, C1);
2424 // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D))
2425 // Since this OR statement hasn't been optimized further yet, we hope
2426 // that this transformation will allow the new ORs to be optimized.
2428 Value *X = nullptr, *Y = nullptr;
2429 if (Op0->hasOneUse() && Op1->hasOneUse() &&
2430 match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) &&
2431 match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) {
2432 Value *orTrue = Builder->CreateOr(A, C);
2433 Value *orFalse = Builder->CreateOr(B, D);
2434 return SelectInst::Create(X, orTrue, orFalse);
2438 return Changed ? &I : nullptr;
2441 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
2442 // here. We should standardize that construct where it is needed or choose some
2443 // other way to ensure that commutated variants of patterns are not missed.
2444 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
2445 bool Changed = SimplifyAssociativeOrCommutative(I);
2446 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2448 if (Value *V = SimplifyVectorOp(I))
2449 return replaceInstUsesWith(I, V);
2451 if (Value *V = SimplifyXorInst(Op0, Op1, DL, &TLI, &DT, &AC))
2452 return replaceInstUsesWith(I, V);
2454 // (A&B)^(A&C) -> A&(B^C) etc
2455 if (Value *V = SimplifyUsingDistributiveLaws(I))
2456 return replaceInstUsesWith(I, V);
2458 // See if we can simplify any instructions used by the instruction whose sole
2459 // purpose is to compute bits we don't care about.
2460 if (SimplifyDemandedInstructionBits(I))
2463 if (Value *V = SimplifyBSwap(I))
2464 return replaceInstUsesWith(I, V);
2466 // Is this a ~ operation?
2467 if (Value *NotOp = dyn_castNotVal(&I)) {
2468 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
2469 if (Op0I->getOpcode() == Instruction::And ||
2470 Op0I->getOpcode() == Instruction::Or) {
2471 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
2472 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
2473 if (dyn_castNotVal(Op0I->getOperand(1)))
2474 Op0I->swapOperands();
2475 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
2477 Builder->CreateNot(Op0I->getOperand(1),
2478 Op0I->getOperand(1)->getName()+".not");
2479 if (Op0I->getOpcode() == Instruction::And)
2480 return BinaryOperator::CreateOr(Op0NotVal, NotY);
2481 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
2484 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
2485 // ~(X | Y) === (~X & ~Y) - De Morgan's Law
2486 if (IsFreeToInvert(Op0I->getOperand(0),
2487 Op0I->getOperand(0)->hasOneUse()) &&
2488 IsFreeToInvert(Op0I->getOperand(1),
2489 Op0I->getOperand(1)->hasOneUse())) {
2491 Builder->CreateNot(Op0I->getOperand(0), "notlhs");
2493 Builder->CreateNot(Op0I->getOperand(1), "notrhs");
2494 if (Op0I->getOpcode() == Instruction::And)
2495 return BinaryOperator::CreateOr(NotX, NotY);
2496 return BinaryOperator::CreateAnd(NotX, NotY);
2499 } else if (Op0I->getOpcode() == Instruction::AShr) {
2500 // ~(~X >>s Y) --> (X >>s Y)
2501 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0)))
2502 return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1));
2507 if (Constant *RHS = dyn_cast<Constant>(Op1)) {
2508 if (RHS->isAllOnesValue() && Op0->hasOneUse())
2509 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
2510 if (CmpInst *CI = dyn_cast<CmpInst>(Op0))
2511 return CmpInst::Create(CI->getOpcode(),
2512 CI->getInversePredicate(),
2513 CI->getOperand(0), CI->getOperand(1));
2516 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2517 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
2518 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2519 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
2520 if (CI->hasOneUse() && Op0C->hasOneUse()) {
2521 Instruction::CastOps Opcode = Op0C->getOpcode();
2522 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
2523 (RHS == ConstantExpr::getCast(Opcode, Builder->getTrue(),
2524 Op0C->getDestTy()))) {
2525 CI->setPredicate(CI->getInversePredicate());
2526 return CastInst::Create(Opcode, CI, Op0C->getType());
2532 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2533 // ~(c-X) == X-c-1 == X+(-c-1)
2534 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
2535 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
2536 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
2537 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
2538 ConstantInt::get(I.getType(), 1));
2539 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
2542 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2543 if (Op0I->getOpcode() == Instruction::Add) {
2544 // ~(X-c) --> (-c-1)-X
2545 if (RHS->isAllOnesValue()) {
2546 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
2547 return BinaryOperator::CreateSub(
2548 ConstantExpr::getSub(NegOp0CI,
2549 ConstantInt::get(I.getType(), 1)),
2550 Op0I->getOperand(0));
2551 } else if (RHS->getValue().isSignBit()) {
2552 // (X + C) ^ signbit -> (X + C + signbit)
2553 Constant *C = Builder->getInt(RHS->getValue() + Op0CI->getValue());
2554 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
2557 } else if (Op0I->getOpcode() == Instruction::Or) {
2558 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
2559 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue(),
2561 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
2562 // Anything in both C1 and C2 is known to be zero, remove it from
2564 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
2565 NewRHS = ConstantExpr::getAnd(NewRHS,
2566 ConstantExpr::getNot(CommonBits));
2568 I.setOperand(0, Op0I->getOperand(0));
2569 I.setOperand(1, NewRHS);
2572 } else if (Op0I->getOpcode() == Instruction::LShr) {
2573 // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3)
2577 if (Op0I->hasOneUse() &&
2578 (E1 = dyn_cast<BinaryOperator>(Op0I->getOperand(0))) &&
2579 E1->getOpcode() == Instruction::Xor &&
2580 (C1 = dyn_cast<ConstantInt>(E1->getOperand(1)))) {
2581 // fold (C1 >> C2) ^ C3
2582 ConstantInt *C2 = Op0CI, *C3 = RHS;
2583 APInt FoldConst = C1->getValue().lshr(C2->getValue());
2584 FoldConst ^= C3->getValue();
2585 // Prepare the two operands.
2586 Value *Opnd0 = Builder->CreateLShr(E1->getOperand(0), C2);
2587 Opnd0->takeName(Op0I);
2588 cast<Instruction>(Opnd0)->setDebugLoc(I.getDebugLoc());
2589 Value *FoldVal = ConstantInt::get(Opnd0->getType(), FoldConst);
2591 return BinaryOperator::CreateXor(Opnd0, FoldVal);
2597 // Try to fold constant and into select arguments.
2598 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2599 if (Instruction *R = FoldOpIntoSelect(I, SI))
2601 if (isa<PHINode>(Op0))
2602 if (Instruction *NV = FoldOpIntoPhi(I))
2606 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
2609 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2610 if (A == Op0) { // B^(B|A) == (A|B)^B
2611 Op1I->swapOperands();
2613 std::swap(Op0, Op1);
2614 } else if (B == Op0) { // B^(A|B) == (A|B)^B
2615 I.swapOperands(); // Simplified below.
2616 std::swap(Op0, Op1);
2618 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
2620 if (A == Op0) { // A^(A&B) -> A^(B&A)
2621 Op1I->swapOperands();
2624 if (B == Op0) { // A^(B&A) -> (B&A)^A
2625 I.swapOperands(); // Simplified below.
2626 std::swap(Op0, Op1);
2631 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
2634 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2635 Op0I->hasOneUse()) {
2636 if (A == Op1) // (B|A)^B == (A|B)^B
2638 if (B == Op1) // (A|B)^B == A & ~B
2639 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1));
2640 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2642 if (A == Op1) // (A&B)^A -> (B&A)^A
2644 if (B == Op1 && // (B&A)^A == ~B & A
2645 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
2646 return BinaryOperator::CreateAnd(Builder->CreateNot(A), Op1);
2652 Value *A, *B, *C, *D;
2653 // (A & B)^(A | B) -> A ^ B
2654 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2655 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
2656 if ((A == C && B == D) || (A == D && B == C))
2657 return BinaryOperator::CreateXor(A, B);
2659 // (A | B)^(A & B) -> A ^ B
2660 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2661 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
2662 if ((A == C && B == D) || (A == D && B == C))
2663 return BinaryOperator::CreateXor(A, B);
2665 // (A | ~B) ^ (~A | B) -> A ^ B
2666 // (~B | A) ^ (~A | B) -> A ^ B
2667 if (match(Op0I, m_c_Or(m_Value(A), m_Not(m_Value(B)))) &&
2668 match(Op1I, m_Or(m_Not(m_Specific(A)), m_Specific(B))))
2669 return BinaryOperator::CreateXor(A, B);
2671 // (~A | B) ^ (A | ~B) -> A ^ B
2672 if (match(Op0I, m_Or(m_Not(m_Value(A)), m_Value(B))) &&
2673 match(Op1I, m_Or(m_Specific(A), m_Not(m_Specific(B))))) {
2674 return BinaryOperator::CreateXor(A, B);
2676 // (A & ~B) ^ (~A & B) -> A ^ B
2677 // (~B & A) ^ (~A & B) -> A ^ B
2678 if (match(Op0I, m_c_And(m_Value(A), m_Not(m_Value(B)))) &&
2679 match(Op1I, m_And(m_Not(m_Specific(A)), m_Specific(B))))
2680 return BinaryOperator::CreateXor(A, B);
2682 // (~A & B) ^ (A & ~B) -> A ^ B
2683 if (match(Op0I, m_And(m_Not(m_Value(A)), m_Value(B))) &&
2684 match(Op1I, m_And(m_Specific(A), m_Not(m_Specific(B))))) {
2685 return BinaryOperator::CreateXor(A, B);
2687 // (A ^ C)^(A | B) -> ((~A) & B) ^ C
2688 if (match(Op0I, m_Xor(m_Value(D), m_Value(C))) &&
2689 match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2691 return BinaryOperator::CreateXor(
2692 Builder->CreateAnd(Builder->CreateNot(A), B), C);
2694 return BinaryOperator::CreateXor(
2695 Builder->CreateAnd(Builder->CreateNot(B), A), C);
2697 // (A | B)^(A ^ C) -> ((~A) & B) ^ C
2698 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2699 match(Op1I, m_Xor(m_Value(D), m_Value(C)))) {
2701 return BinaryOperator::CreateXor(
2702 Builder->CreateAnd(Builder->CreateNot(A), B), C);
2704 return BinaryOperator::CreateXor(
2705 Builder->CreateAnd(Builder->CreateNot(B), A), C);
2707 // (A & B) ^ (A ^ B) -> (A | B)
2708 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2709 match(Op1I, m_Xor(m_Specific(A), m_Specific(B))))
2710 return BinaryOperator::CreateOr(A, B);
2711 // (A ^ B) ^ (A & B) -> (A | B)
2712 if (match(Op0I, m_Xor(m_Value(A), m_Value(B))) &&
2713 match(Op1I, m_And(m_Specific(A), m_Specific(B))))
2714 return BinaryOperator::CreateOr(A, B);
2717 // (A & ~B) ^ ~A -> ~(A & B)
2718 // (~B & A) ^ ~A -> ~(A & B)
2720 if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) &&
2721 match(Op1, m_Not(m_Specific(A))))
2722 return BinaryOperator::CreateNot(Builder->CreateAnd(A, B));
2724 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
2725 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
2726 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
2727 if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
2728 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2729 LHS->getOperand(1) == RHS->getOperand(0))
2730 LHS->swapOperands();
2731 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2732 LHS->getOperand(1) == RHS->getOperand(1)) {
2733 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2734 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
2735 bool isSigned = LHS->isSigned() || RHS->isSigned();
2736 return replaceInstUsesWith(I,
2737 getNewICmpValue(isSigned, Code, Op0, Op1,
2742 if (Instruction *CastedXor = foldCastedBitwiseLogic(I))
2745 return Changed ? &I : nullptr;