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/CmpInstAnalysis.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/Transforms/Utils/Local.h"
18 #include "llvm/IR/ConstantRange.h"
19 #include "llvm/IR/Intrinsics.h"
20 #include "llvm/IR/PatternMatch.h"
22 using namespace PatternMatch;
24 #define DEBUG_TYPE "instcombine"
26 /// Similar to getICmpCode but for FCmpInst. This encodes a fcmp predicate into
28 static unsigned getFCmpCode(FCmpInst::Predicate CC) {
29 assert(FCmpInst::FCMP_FALSE <= CC && CC <= FCmpInst::FCMP_TRUE &&
30 "Unexpected FCmp predicate!");
31 // Take advantage of the bit pattern of FCmpInst::Predicate here.
33 static_assert(FCmpInst::FCMP_FALSE == 0, ""); // 0 0 0 0
34 static_assert(FCmpInst::FCMP_OEQ == 1, ""); // 0 0 0 1
35 static_assert(FCmpInst::FCMP_OGT == 2, ""); // 0 0 1 0
36 static_assert(FCmpInst::FCMP_OGE == 3, ""); // 0 0 1 1
37 static_assert(FCmpInst::FCMP_OLT == 4, ""); // 0 1 0 0
38 static_assert(FCmpInst::FCMP_OLE == 5, ""); // 0 1 0 1
39 static_assert(FCmpInst::FCMP_ONE == 6, ""); // 0 1 1 0
40 static_assert(FCmpInst::FCMP_ORD == 7, ""); // 0 1 1 1
41 static_assert(FCmpInst::FCMP_UNO == 8, ""); // 1 0 0 0
42 static_assert(FCmpInst::FCMP_UEQ == 9, ""); // 1 0 0 1
43 static_assert(FCmpInst::FCMP_UGT == 10, ""); // 1 0 1 0
44 static_assert(FCmpInst::FCMP_UGE == 11, ""); // 1 0 1 1
45 static_assert(FCmpInst::FCMP_ULT == 12, ""); // 1 1 0 0
46 static_assert(FCmpInst::FCMP_ULE == 13, ""); // 1 1 0 1
47 static_assert(FCmpInst::FCMP_UNE == 14, ""); // 1 1 1 0
48 static_assert(FCmpInst::FCMP_TRUE == 15, ""); // 1 1 1 1
52 /// This is the complement of getICmpCode, which turns an opcode and two
53 /// operands into either a constant true or false, or a brand new ICmp
54 /// instruction. The sign is passed in to determine which kind of predicate to
55 /// use in the new icmp instruction.
56 static Value *getNewICmpValue(unsigned Code, bool Sign, Value *LHS, Value *RHS,
57 InstCombiner::BuilderTy &Builder) {
58 ICmpInst::Predicate NewPred;
59 if (Constant *TorF = getPredForICmpCode(Code, Sign, LHS->getType(), NewPred))
61 return Builder.CreateICmp(NewPred, LHS, RHS);
64 /// This is the complement of getFCmpCode, which turns an opcode and two
65 /// operands into either a FCmp instruction, or a true/false constant.
66 static Value *getFCmpValue(unsigned Code, Value *LHS, Value *RHS,
67 InstCombiner::BuilderTy &Builder) {
68 const auto Pred = static_cast<FCmpInst::Predicate>(Code);
69 assert(FCmpInst::FCMP_FALSE <= Pred && Pred <= FCmpInst::FCMP_TRUE &&
70 "Unexpected FCmp predicate!");
71 if (Pred == FCmpInst::FCMP_FALSE)
72 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
73 if (Pred == FCmpInst::FCMP_TRUE)
74 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
75 return Builder.CreateFCmp(Pred, LHS, RHS);
78 /// Transform BITWISE_OP(BSWAP(A),BSWAP(B)) or
79 /// BITWISE_OP(BSWAP(A), Constant) to BSWAP(BITWISE_OP(A, B))
80 /// \param I Binary operator to transform.
81 /// \return Pointer to node that must replace the original binary operator, or
82 /// null pointer if no transformation was made.
83 static Value *SimplifyBSwap(BinaryOperator &I,
84 InstCombiner::BuilderTy &Builder) {
85 assert(I.isBitwiseLogicOp() && "Unexpected opcode for bswap simplifying");
87 Value *OldLHS = I.getOperand(0);
88 Value *OldRHS = I.getOperand(1);
91 if (!match(OldLHS, m_BSwap(m_Value(NewLHS))))
97 if (match(OldRHS, m_BSwap(m_Value(NewRHS)))) {
98 // OP( BSWAP(x), BSWAP(y) ) -> BSWAP( OP(x, y) )
99 if (!OldLHS->hasOneUse() && !OldRHS->hasOneUse())
101 // NewRHS initialized by the matcher.
102 } else if (match(OldRHS, m_APInt(C))) {
103 // OP( BSWAP(x), CONSTANT ) -> BSWAP( OP(x, BSWAP(CONSTANT) ) )
104 if (!OldLHS->hasOneUse())
106 NewRHS = ConstantInt::get(I.getType(), C->byteSwap());
110 Value *BinOp = Builder.CreateBinOp(I.getOpcode(), NewLHS, NewRHS);
111 Function *F = Intrinsic::getDeclaration(I.getModule(), Intrinsic::bswap,
113 return Builder.CreateCall(F, BinOp);
116 /// This handles expressions of the form ((val OP C1) & C2). Where
117 /// the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'.
118 Instruction *InstCombiner::OptAndOp(BinaryOperator *Op,
121 BinaryOperator &TheAnd) {
122 Value *X = Op->getOperand(0);
124 switch (Op->getOpcode()) {
126 case Instruction::Add:
127 if (Op->hasOneUse()) {
128 // Adding a one to a single bit bit-field should be turned into an XOR
129 // of the bit. First thing to check is to see if this AND is with a
130 // single bit constant.
131 const APInt &AndRHSV = AndRHS->getValue();
133 // If there is only one bit set.
134 if (AndRHSV.isPowerOf2()) {
135 // Ok, at this point, we know that we are masking the result of the
136 // ADD down to exactly one bit. If the constant we are adding has
137 // no bits set below this bit, then we can eliminate the ADD.
138 const APInt& AddRHS = OpRHS->getValue();
140 // Check to see if any bits below the one bit set in AndRHSV are set.
141 if ((AddRHS & (AndRHSV - 1)).isNullValue()) {
142 // If not, the only thing that can effect the output of the AND is
143 // the bit specified by AndRHSV. If that bit is set, the effect of
144 // the XOR is to toggle the bit. If it is clear, then the ADD has
146 if ((AddRHS & AndRHSV).isNullValue()) { // Bit is not set, noop
147 TheAnd.setOperand(0, X);
150 // Pull the XOR out of the AND.
151 Value *NewAnd = Builder.CreateAnd(X, AndRHS);
152 NewAnd->takeName(Op);
153 return BinaryOperator::CreateXor(NewAnd, AndRHS);
163 /// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise
164 /// (V < Lo || V >= Hi). This method expects that Lo <= Hi. IsSigned indicates
165 /// whether to treat V, Lo, and Hi as signed or not.
166 Value *InstCombiner::insertRangeTest(Value *V, const APInt &Lo, const APInt &Hi,
167 bool isSigned, bool Inside) {
168 assert((isSigned ? Lo.sle(Hi) : Lo.ule(Hi)) &&
169 "Lo is not <= Hi in range emission code!");
171 Type *Ty = V->getType();
173 return Inside ? ConstantInt::getFalse(Ty) : ConstantInt::getTrue(Ty);
175 // V >= Min && V < Hi --> V < Hi
176 // V < Min || V >= Hi --> V >= Hi
177 ICmpInst::Predicate Pred = Inside ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE;
178 if (isSigned ? Lo.isMinSignedValue() : Lo.isMinValue()) {
179 Pred = isSigned ? ICmpInst::getSignedPredicate(Pred) : Pred;
180 return Builder.CreateICmp(Pred, V, ConstantInt::get(Ty, Hi));
183 // V >= Lo && V < Hi --> V - Lo u< Hi - Lo
184 // V < Lo || V >= Hi --> V - Lo u>= Hi - Lo
186 Builder.CreateSub(V, ConstantInt::get(Ty, Lo), V->getName() + ".off");
187 Constant *HiMinusLo = ConstantInt::get(Ty, Hi - Lo);
188 return Builder.CreateICmp(Pred, VMinusLo, HiMinusLo);
191 /// Classify (icmp eq (A & B), C) and (icmp ne (A & B), C) as matching patterns
192 /// that can be simplified.
193 /// One of A and B is considered the mask. The other is the value. This is
194 /// described as the "AMask" or "BMask" part of the enum. If the enum contains
195 /// only "Mask", then both A and B can be considered masks. If A is the mask,
196 /// then it was proven that (A & C) == C. This is trivial if C == A or C == 0.
197 /// If both A and C are constants, this proof is also easy.
198 /// For the following explanations, we assume that A is the mask.
200 /// "AllOnes" declares that the comparison is true only if (A & B) == A or all
201 /// bits of A are set in B.
202 /// Example: (icmp eq (A & 3), 3) -> AMask_AllOnes
204 /// "AllZeros" declares that the comparison is true only if (A & B) == 0 or all
205 /// bits of A are cleared in B.
206 /// Example: (icmp eq (A & 3), 0) -> Mask_AllZeroes
208 /// "Mixed" declares that (A & B) == C and C might or might not contain any
209 /// number of one bits and zero bits.
210 /// Example: (icmp eq (A & 3), 1) -> AMask_Mixed
212 /// "Not" means that in above descriptions "==" should be replaced by "!=".
213 /// Example: (icmp ne (A & 3), 3) -> AMask_NotAllOnes
215 /// If the mask A contains a single bit, then the following is equivalent:
216 /// (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
217 /// (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
218 enum MaskedICmpType {
220 AMask_NotAllOnes = 2,
222 BMask_NotAllOnes = 8,
224 Mask_NotAllZeros = 32,
226 AMask_NotMixed = 128,
231 /// Return the set of patterns (from MaskedICmpType) that (icmp SCC (A & B), C)
233 static unsigned getMaskedICmpType(Value *A, Value *B, Value *C,
234 ICmpInst::Predicate Pred) {
235 ConstantInt *ACst = dyn_cast<ConstantInt>(A);
236 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
237 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
238 bool IsEq = (Pred == ICmpInst::ICMP_EQ);
239 bool IsAPow2 = (ACst && !ACst->isZero() && ACst->getValue().isPowerOf2());
240 bool IsBPow2 = (BCst && !BCst->isZero() && BCst->getValue().isPowerOf2());
241 unsigned MaskVal = 0;
242 if (CCst && CCst->isZero()) {
243 // if C is zero, then both A and B qualify as mask
244 MaskVal |= (IsEq ? (Mask_AllZeros | AMask_Mixed | BMask_Mixed)
245 : (Mask_NotAllZeros | AMask_NotMixed | BMask_NotMixed));
247 MaskVal |= (IsEq ? (AMask_NotAllOnes | AMask_NotMixed)
248 : (AMask_AllOnes | AMask_Mixed));
250 MaskVal |= (IsEq ? (BMask_NotAllOnes | BMask_NotMixed)
251 : (BMask_AllOnes | BMask_Mixed));
256 MaskVal |= (IsEq ? (AMask_AllOnes | AMask_Mixed)
257 : (AMask_NotAllOnes | AMask_NotMixed));
259 MaskVal |= (IsEq ? (Mask_NotAllZeros | AMask_NotMixed)
260 : (Mask_AllZeros | AMask_Mixed));
261 } else if (ACst && CCst && ConstantExpr::getAnd(ACst, CCst) == CCst) {
262 MaskVal |= (IsEq ? AMask_Mixed : AMask_NotMixed);
266 MaskVal |= (IsEq ? (BMask_AllOnes | BMask_Mixed)
267 : (BMask_NotAllOnes | BMask_NotMixed));
269 MaskVal |= (IsEq ? (Mask_NotAllZeros | BMask_NotMixed)
270 : (Mask_AllZeros | BMask_Mixed));
271 } else if (BCst && CCst && ConstantExpr::getAnd(BCst, CCst) == CCst) {
272 MaskVal |= (IsEq ? BMask_Mixed : BMask_NotMixed);
278 /// Convert an analysis of a masked ICmp into its equivalent if all boolean
279 /// operations had the opposite sense. Since each "NotXXX" flag (recording !=)
280 /// is adjacent to the corresponding normal flag (recording ==), this just
281 /// involves swapping those bits over.
282 static unsigned conjugateICmpMask(unsigned Mask) {
284 NewMask = (Mask & (AMask_AllOnes | BMask_AllOnes | Mask_AllZeros |
285 AMask_Mixed | BMask_Mixed))
288 NewMask |= (Mask & (AMask_NotAllOnes | BMask_NotAllOnes | Mask_NotAllZeros |
289 AMask_NotMixed | BMask_NotMixed))
295 // Adapts the external decomposeBitTestICmp for local use.
296 static bool decomposeBitTestICmp(Value *LHS, Value *RHS, CmpInst::Predicate &Pred,
297 Value *&X, Value *&Y, Value *&Z) {
299 if (!llvm::decomposeBitTestICmp(LHS, RHS, Pred, X, Mask))
302 Y = ConstantInt::get(X->getType(), Mask);
303 Z = ConstantInt::get(X->getType(), 0);
307 /// Handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E).
308 /// Return the pattern classes (from MaskedICmpType) for the left hand side and
309 /// the right hand side as a pair.
310 /// LHS and RHS are the left hand side and the right hand side ICmps and PredL
311 /// and PredR are their predicates, respectively.
313 Optional<std::pair<unsigned, unsigned>>
314 getMaskedTypeForICmpPair(Value *&A, Value *&B, Value *&C,
315 Value *&D, Value *&E, ICmpInst *LHS,
317 ICmpInst::Predicate &PredL,
318 ICmpInst::Predicate &PredR) {
319 // vectors are not (yet?) supported. Don't support pointers either.
320 if (!LHS->getOperand(0)->getType()->isIntegerTy() ||
321 !RHS->getOperand(0)->getType()->isIntegerTy())
324 // Here comes the tricky part:
325 // LHS might be of the form L11 & L12 == X, X == L21 & L22,
326 // and L11 & L12 == L21 & L22. The same goes for RHS.
327 // Now we must find those components L** and R**, that are equal, so
328 // that we can extract the parameters A, B, C, D, and E for the canonical
330 Value *L1 = LHS->getOperand(0);
331 Value *L2 = LHS->getOperand(1);
332 Value *L11, *L12, *L21, *L22;
333 // Check whether the icmp can be decomposed into a bit test.
334 if (decomposeBitTestICmp(L1, L2, PredL, L11, L12, L2)) {
335 L21 = L22 = L1 = nullptr;
337 // Look for ANDs in the LHS icmp.
338 if (!match(L1, m_And(m_Value(L11), m_Value(L12)))) {
339 // Any icmp can be viewed as being trivially masked; if it allows us to
340 // remove one, it's worth it.
342 L12 = Constant::getAllOnesValue(L1->getType());
345 if (!match(L2, m_And(m_Value(L21), m_Value(L22)))) {
347 L22 = Constant::getAllOnesValue(L2->getType());
351 // Bail if LHS was a icmp that can't be decomposed into an equality.
352 if (!ICmpInst::isEquality(PredL))
355 Value *R1 = RHS->getOperand(0);
356 Value *R2 = RHS->getOperand(1);
359 if (decomposeBitTestICmp(R1, R2, PredR, R11, R12, R2)) {
360 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
363 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
373 if (!match(R1, m_And(m_Value(R11), m_Value(R12)))) {
374 // As before, model no mask as a trivial mask if it'll let us do an
377 R12 = Constant::getAllOnesValue(R1->getType());
380 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
385 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
393 // Bail if RHS was a icmp that can't be decomposed into an equality.
394 if (!ICmpInst::isEquality(PredR))
397 // Look for ANDs on the right side of the RHS icmp.
399 if (!match(R2, m_And(m_Value(R11), m_Value(R12)))) {
401 R12 = Constant::getAllOnesValue(R2->getType());
404 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
409 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
424 } else if (L12 == A) {
427 } else if (L21 == A) {
430 } else if (L22 == A) {
435 unsigned LeftType = getMaskedICmpType(A, B, C, PredL);
436 unsigned RightType = getMaskedICmpType(A, D, E, PredR);
437 return Optional<std::pair<unsigned, unsigned>>(std::make_pair(LeftType, RightType));
440 /// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E) into a single
441 /// (icmp(A & X) ==/!= Y), where the left-hand side is of type Mask_NotAllZeros
442 /// and the right hand side is of type BMask_Mixed. For example,
443 /// (icmp (A & 12) != 0) & (icmp (A & 15) == 8) -> (icmp (A & 15) == 8).
444 static Value * foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(
445 ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
446 Value *A, Value *B, Value *C, Value *D, Value *E,
447 ICmpInst::Predicate PredL, ICmpInst::Predicate PredR,
448 llvm::InstCombiner::BuilderTy &Builder) {
449 // We are given the canonical form:
450 // (icmp ne (A & B), 0) & (icmp eq (A & D), E).
453 // If IsAnd is false, we get it in negated form:
454 // (icmp eq (A & B), 0) | (icmp ne (A & D), E) ->
455 // !((icmp ne (A & B), 0) & (icmp eq (A & D), E)).
457 // We currently handle the case of B, C, D, E are constant.
459 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
462 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
465 ConstantInt *DCst = dyn_cast<ConstantInt>(D);
468 ConstantInt *ECst = dyn_cast<ConstantInt>(E);
472 ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
474 // Update E to the canonical form when D is a power of two and RHS is
476 // (icmp ne (A & D), 0) -> (icmp eq (A & D), D) or
477 // (icmp ne (A & D), D) -> (icmp eq (A & D), 0).
479 ECst = cast<ConstantInt>(ConstantExpr::getXor(DCst, ECst));
481 // If B or D is zero, skip because if LHS or RHS can be trivially folded by
482 // other folding rules and this pattern won't apply any more.
483 if (BCst->getValue() == 0 || DCst->getValue() == 0)
486 // If B and D don't intersect, ie. (B & D) == 0, no folding because we can't
487 // deduce anything from it.
489 // (icmp ne (A & 12), 0) & (icmp eq (A & 3), 1) -> no folding.
490 if ((BCst->getValue() & DCst->getValue()) == 0)
493 // If the following two conditions are met:
495 // 1. mask B covers only a single bit that's not covered by mask D, that is,
496 // (B & (B ^ D)) is a power of 2 (in other words, B minus the intersection of
497 // B and D has only one bit set) and,
499 // 2. RHS (and E) indicates that the rest of B's bits are zero (in other
500 // words, the intersection of B and D is zero), that is, ((B & D) & E) == 0
502 // then that single bit in B must be one and thus the whole expression can be
504 // (A & (B | D)) == (B & (B ^ D)) | E.
507 // (icmp ne (A & 12), 0) & (icmp eq (A & 7), 1) -> (icmp eq (A & 15), 9)
508 // (icmp ne (A & 15), 0) & (icmp eq (A & 7), 0) -> (icmp eq (A & 15), 8)
509 if ((((BCst->getValue() & DCst->getValue()) & ECst->getValue()) == 0) &&
510 (BCst->getValue() & (BCst->getValue() ^ DCst->getValue())).isPowerOf2()) {
511 APInt BorD = BCst->getValue() | DCst->getValue();
512 APInt BandBxorDorE = (BCst->getValue() & (BCst->getValue() ^ DCst->getValue())) |
514 Value *NewMask = ConstantInt::get(BCst->getType(), BorD);
515 Value *NewMaskedValue = ConstantInt::get(BCst->getType(), BandBxorDorE);
516 Value *NewAnd = Builder.CreateAnd(A, NewMask);
517 return Builder.CreateICmp(NewCC, NewAnd, NewMaskedValue);
520 auto IsSubSetOrEqual = [](ConstantInt *C1, ConstantInt *C2) {
521 return (C1->getValue() & C2->getValue()) == C1->getValue();
523 auto IsSuperSetOrEqual = [](ConstantInt *C1, ConstantInt *C2) {
524 return (C1->getValue() & C2->getValue()) == C2->getValue();
527 // In the following, we consider only the cases where B is a superset of D, B
528 // is a subset of D, or B == D because otherwise there's at least one bit
529 // covered by B but not D, in which case we can't deduce much from it, so
530 // no folding (aside from the single must-be-one bit case right above.)
532 // (icmp ne (A & 14), 0) & (icmp eq (A & 3), 1) -> no folding.
533 if (!IsSubSetOrEqual(BCst, DCst) && !IsSuperSetOrEqual(BCst, DCst))
536 // At this point, either B is a superset of D, B is a subset of D or B == D.
538 // If E is zero, if B is a subset of (or equal to) D, LHS and RHS contradict
539 // and the whole expression becomes false (or true if negated), otherwise, no
542 // (icmp ne (A & 3), 0) & (icmp eq (A & 7), 0) -> false.
543 // (icmp ne (A & 15), 0) & (icmp eq (A & 3), 0) -> no folding.
544 if (ECst->isZero()) {
545 if (IsSubSetOrEqual(BCst, DCst))
546 return ConstantInt::get(LHS->getType(), !IsAnd);
550 // At this point, B, D, E aren't zero and (B & D) == B, (B & D) == D or B ==
551 // D. If B is a superset of (or equal to) D, since E is not zero, LHS is
552 // subsumed by RHS (RHS implies LHS.) So the whole expression becomes
554 // (icmp ne (A & 255), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).
555 // (icmp ne (A & 15), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).
556 if (IsSuperSetOrEqual(BCst, DCst))
558 // Otherwise, B is a subset of D. If B and E have a common bit set,
559 // ie. (B & E) != 0, then LHS is subsumed by RHS. For example.
560 // (icmp ne (A & 12), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).
561 assert(IsSubSetOrEqual(BCst, DCst) && "Precondition due to above code");
562 if ((BCst->getValue() & ECst->getValue()) != 0)
564 // Otherwise, LHS and RHS contradict and the whole expression becomes false
565 // (or true if negated.) For example,
566 // (icmp ne (A & 7), 0) & (icmp eq (A & 15), 8) -> false.
567 // (icmp ne (A & 6), 0) & (icmp eq (A & 15), 8) -> false.
568 return ConstantInt::get(LHS->getType(), !IsAnd);
571 /// Try to fold (icmp(A & B) ==/!= 0) &/| (icmp(A & D) ==/!= E) into a single
572 /// (icmp(A & X) ==/!= Y), where the left-hand side and the right hand side
573 /// aren't of the common mask pattern type.
574 static Value *foldLogOpOfMaskedICmpsAsymmetric(
575 ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
576 Value *A, Value *B, Value *C, Value *D, Value *E,
577 ICmpInst::Predicate PredL, ICmpInst::Predicate PredR,
578 unsigned LHSMask, unsigned RHSMask,
579 llvm::InstCombiner::BuilderTy &Builder) {
580 assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&
581 "Expected equality predicates for masked type of icmps.");
582 // Handle Mask_NotAllZeros-BMask_Mixed cases.
583 // (icmp ne/eq (A & B), C) &/| (icmp eq/ne (A & D), E), or
584 // (icmp eq/ne (A & B), C) &/| (icmp ne/eq (A & D), E)
585 // which gets swapped to
586 // (icmp ne/eq (A & D), E) &/| (icmp eq/ne (A & B), C).
588 LHSMask = conjugateICmpMask(LHSMask);
589 RHSMask = conjugateICmpMask(RHSMask);
591 if ((LHSMask & Mask_NotAllZeros) && (RHSMask & BMask_Mixed)) {
592 if (Value *V = foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(
593 LHS, RHS, IsAnd, A, B, C, D, E,
594 PredL, PredR, Builder)) {
597 } else if ((LHSMask & BMask_Mixed) && (RHSMask & Mask_NotAllZeros)) {
598 if (Value *V = foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(
599 RHS, LHS, IsAnd, A, D, E, B, C,
600 PredR, PredL, Builder)) {
607 /// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
608 /// into a single (icmp(A & X) ==/!= Y).
609 static Value *foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
610 llvm::InstCombiner::BuilderTy &Builder) {
611 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr;
612 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
613 Optional<std::pair<unsigned, unsigned>> MaskPair =
614 getMaskedTypeForICmpPair(A, B, C, D, E, LHS, RHS, PredL, PredR);
617 assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&
618 "Expected equality predicates for masked type of icmps.");
619 unsigned LHSMask = MaskPair->first;
620 unsigned RHSMask = MaskPair->second;
621 unsigned Mask = LHSMask & RHSMask;
623 // Even if the two sides don't share a common pattern, check if folding can
625 if (Value *V = foldLogOpOfMaskedICmpsAsymmetric(
626 LHS, RHS, IsAnd, A, B, C, D, E, PredL, PredR, LHSMask, RHSMask,
632 // In full generality:
633 // (icmp (A & B) Op C) | (icmp (A & D) Op E)
634 // == ![ (icmp (A & B) !Op C) & (icmp (A & D) !Op E) ]
636 // If the latter can be converted into (icmp (A & X) Op Y) then the former is
637 // equivalent to (icmp (A & X) !Op Y).
639 // Therefore, we can pretend for the rest of this function that we're dealing
640 // with the conjunction, provided we flip the sense of any comparisons (both
641 // input and output).
643 // In most cases we're going to produce an EQ for the "&&" case.
644 ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
646 // Convert the masking analysis into its equivalent with negated
648 Mask = conjugateICmpMask(Mask);
651 if (Mask & Mask_AllZeros) {
652 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
653 // -> (icmp eq (A & (B|D)), 0)
654 Value *NewOr = Builder.CreateOr(B, D);
655 Value *NewAnd = Builder.CreateAnd(A, NewOr);
656 // We can't use C as zero because we might actually handle
657 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
658 // with B and D, having a single bit set.
659 Value *Zero = Constant::getNullValue(A->getType());
660 return Builder.CreateICmp(NewCC, NewAnd, Zero);
662 if (Mask & BMask_AllOnes) {
663 // (icmp eq (A & B), B) & (icmp eq (A & D), D)
664 // -> (icmp eq (A & (B|D)), (B|D))
665 Value *NewOr = Builder.CreateOr(B, D);
666 Value *NewAnd = Builder.CreateAnd(A, NewOr);
667 return Builder.CreateICmp(NewCC, NewAnd, NewOr);
669 if (Mask & AMask_AllOnes) {
670 // (icmp eq (A & B), A) & (icmp eq (A & D), A)
671 // -> (icmp eq (A & (B&D)), A)
672 Value *NewAnd1 = Builder.CreateAnd(B, D);
673 Value *NewAnd2 = Builder.CreateAnd(A, NewAnd1);
674 return Builder.CreateICmp(NewCC, NewAnd2, A);
677 // Remaining cases assume at least that B and D are constant, and depend on
678 // their actual values. This isn't strictly necessary, just a "handle the
679 // easy cases for now" decision.
680 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
683 ConstantInt *DCst = dyn_cast<ConstantInt>(D);
687 if (Mask & (Mask_NotAllZeros | BMask_NotAllOnes)) {
688 // (icmp ne (A & B), 0) & (icmp ne (A & D), 0) and
689 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
690 // -> (icmp ne (A & B), 0) or (icmp ne (A & D), 0)
691 // Only valid if one of the masks is a superset of the other (check "B&D" is
692 // the same as either B or D).
693 APInt NewMask = BCst->getValue() & DCst->getValue();
695 if (NewMask == BCst->getValue())
697 else if (NewMask == DCst->getValue())
701 if (Mask & AMask_NotAllOnes) {
702 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
703 // -> (icmp ne (A & B), A) or (icmp ne (A & D), A)
704 // Only valid if one of the masks is a superset of the other (check "B|D" is
705 // the same as either B or D).
706 APInt NewMask = BCst->getValue() | DCst->getValue();
708 if (NewMask == BCst->getValue())
710 else if (NewMask == DCst->getValue())
714 if (Mask & BMask_Mixed) {
715 // (icmp eq (A & B), C) & (icmp eq (A & D), E)
716 // We already know that B & C == C && D & E == E.
717 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
718 // C and E, which are shared by both the mask B and the mask D, don't
719 // contradict, then we can transform to
720 // -> (icmp eq (A & (B|D)), (C|E))
721 // Currently, we only handle the case of B, C, D, and E being constant.
722 // We can't simply use C and E because we might actually handle
723 // (icmp ne (A & B), B) & (icmp eq (A & D), D)
724 // with B and D, having a single bit set.
725 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
728 ConstantInt *ECst = dyn_cast<ConstantInt>(E);
732 CCst = cast<ConstantInt>(ConstantExpr::getXor(BCst, CCst));
734 ECst = cast<ConstantInt>(ConstantExpr::getXor(DCst, ECst));
736 // If there is a conflict, we should actually return a false for the
738 if (((BCst->getValue() & DCst->getValue()) &
739 (CCst->getValue() ^ ECst->getValue())).getBoolValue())
740 return ConstantInt::get(LHS->getType(), !IsAnd);
742 Value *NewOr1 = Builder.CreateOr(B, D);
743 Value *NewOr2 = ConstantExpr::getOr(CCst, ECst);
744 Value *NewAnd = Builder.CreateAnd(A, NewOr1);
745 return Builder.CreateICmp(NewCC, NewAnd, NewOr2);
751 /// Try to fold a signed range checked with lower bound 0 to an unsigned icmp.
752 /// Example: (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
753 /// If \p Inverted is true then the check is for the inverted range, e.g.
754 /// (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
755 Value *InstCombiner::simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1,
757 // Check the lower range comparison, e.g. x >= 0
758 // InstCombine already ensured that if there is a constant it's on the RHS.
759 ConstantInt *RangeStart = dyn_cast<ConstantInt>(Cmp0->getOperand(1));
763 ICmpInst::Predicate Pred0 = (Inverted ? Cmp0->getInversePredicate() :
764 Cmp0->getPredicate());
766 // Accept x > -1 or x >= 0 (after potentially inverting the predicate).
767 if (!((Pred0 == ICmpInst::ICMP_SGT && RangeStart->isMinusOne()) ||
768 (Pred0 == ICmpInst::ICMP_SGE && RangeStart->isZero())))
771 ICmpInst::Predicate Pred1 = (Inverted ? Cmp1->getInversePredicate() :
772 Cmp1->getPredicate());
774 Value *Input = Cmp0->getOperand(0);
776 if (Cmp1->getOperand(0) == Input) {
777 // For the upper range compare we have: icmp x, n
778 RangeEnd = Cmp1->getOperand(1);
779 } else if (Cmp1->getOperand(1) == Input) {
780 // For the upper range compare we have: icmp n, x
781 RangeEnd = Cmp1->getOperand(0);
782 Pred1 = ICmpInst::getSwappedPredicate(Pred1);
787 // Check the upper range comparison, e.g. x < n
788 ICmpInst::Predicate NewPred;
790 case ICmpInst::ICMP_SLT: NewPred = ICmpInst::ICMP_ULT; break;
791 case ICmpInst::ICMP_SLE: NewPred = ICmpInst::ICMP_ULE; break;
792 default: return nullptr;
795 // This simplification is only valid if the upper range is not negative.
796 KnownBits Known = computeKnownBits(RangeEnd, /*Depth=*/0, Cmp1);
797 if (!Known.isNonNegative())
801 NewPred = ICmpInst::getInversePredicate(NewPred);
803 return Builder.CreateICmp(NewPred, Input, RangeEnd);
807 foldAndOrOfEqualityCmpsWithConstants(ICmpInst *LHS, ICmpInst *RHS,
809 InstCombiner::BuilderTy &Builder) {
810 Value *X = LHS->getOperand(0);
811 if (X != RHS->getOperand(0))
814 const APInt *C1, *C2;
815 if (!match(LHS->getOperand(1), m_APInt(C1)) ||
816 !match(RHS->getOperand(1), m_APInt(C2)))
819 // We only handle (X != C1 && X != C2) and (X == C1 || X == C2).
820 ICmpInst::Predicate Pred = LHS->getPredicate();
821 if (Pred != RHS->getPredicate())
823 if (JoinedByAnd && Pred != ICmpInst::ICMP_NE)
825 if (!JoinedByAnd && Pred != ICmpInst::ICMP_EQ)
828 // The larger unsigned constant goes on the right.
832 APInt Xor = *C1 ^ *C2;
833 if (Xor.isPowerOf2()) {
834 // If LHSC and RHSC differ by only one bit, then set that bit in X and
835 // compare against the larger constant:
836 // (X == C1 || X == C2) --> (X | (C1 ^ C2)) == C2
837 // (X != C1 && X != C2) --> (X | (C1 ^ C2)) != C2
838 // We choose an 'or' with a Pow2 constant rather than the inverse mask with
839 // 'and' because that may lead to smaller codegen from a smaller constant.
840 Value *Or = Builder.CreateOr(X, ConstantInt::get(X->getType(), Xor));
841 return Builder.CreateICmp(Pred, Or, ConstantInt::get(X->getType(), *C2));
844 // Special case: get the ordering right when the values wrap around zero.
845 // Ie, we assumed the constants were unsigned when swapping earlier.
846 if (C1->isNullValue() && C2->isAllOnesValue())
849 if (*C1 == *C2 - 1) {
850 // (X == 13 || X == 14) --> X - 13 <=u 1
851 // (X != 13 && X != 14) --> X - 13 >u 1
852 // An 'add' is the canonical IR form, so favor that over a 'sub'.
853 Value *Add = Builder.CreateAdd(X, ConstantInt::get(X->getType(), -(*C1)));
854 auto NewPred = JoinedByAnd ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_ULE;
855 return Builder.CreateICmp(NewPred, Add, ConstantInt::get(X->getType(), 1));
861 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2)
862 // Fold (!iszero(A & K1) & !iszero(A & K2)) -> (A & (K1 | K2)) == (K1 | K2)
863 Value *InstCombiner::foldAndOrOfICmpsOfAndWithPow2(ICmpInst *LHS, ICmpInst *RHS,
866 ICmpInst::Predicate Pred = LHS->getPredicate();
867 if (Pred != RHS->getPredicate())
869 if (JoinedByAnd && Pred != ICmpInst::ICMP_NE)
871 if (!JoinedByAnd && Pred != ICmpInst::ICMP_EQ)
874 // TODO support vector splats
875 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHS->getOperand(1));
876 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS->getOperand(1));
877 if (!LHSC || !RHSC || !LHSC->isZero() || !RHSC->isZero())
880 Value *A, *B, *C, *D;
881 if (match(LHS->getOperand(0), m_And(m_Value(A), m_Value(B))) &&
882 match(RHS->getOperand(0), m_And(m_Value(C), m_Value(D)))) {
883 if (A == D || B == D)
889 isKnownToBeAPowerOfTwo(B, false, 0, &CxtI) &&
890 isKnownToBeAPowerOfTwo(D, false, 0, &CxtI)) {
891 Value *Mask = Builder.CreateOr(B, D);
892 Value *Masked = Builder.CreateAnd(A, Mask);
893 auto NewPred = JoinedByAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
894 return Builder.CreateICmp(NewPred, Masked, Mask);
904 /// Where Y is checking that all the high bits (covered by a mask 4294967168)
905 /// are uniform, i.e. %arg & 4294967168 can be either 4294967168 or 0
906 /// Pattern can be one of:
907 /// %t = add i32 %arg, 128
908 /// %r = icmp ult i32 %t, 256
910 /// %t0 = shl i32 %arg, 24
911 /// %t1 = ashr i32 %t0, 24
912 /// %r = icmp eq i32 %t1, %arg
914 /// %t0 = trunc i32 %arg to i8
915 /// %t1 = sext i8 %t0 to i32
916 /// %r = icmp eq i32 %t1, %arg
917 /// This pattern is a signed truncation check.
919 /// And X is checking that some bit in that same mask is zero.
920 /// I.e. can be one of:
921 /// %r = icmp sgt i32 %arg, -1
923 /// %t = and i32 %arg, 2147483648
924 /// %r = icmp eq i32 %t, 0
926 /// Since we are checking that all the bits in that mask are the same,
927 /// and a particular bit is zero, what we are really checking is that all the
928 /// masked bits are zero.
929 /// So this should be transformed to:
930 /// %r = icmp ult i32 %arg, 128
931 static Value *foldSignedTruncationCheck(ICmpInst *ICmp0, ICmpInst *ICmp1,
933 InstCombiner::BuilderTy &Builder) {
934 assert(CxtI.getOpcode() == Instruction::And);
936 // Match icmp ult (add %arg, C01), C1 (C1 == C01 << 1; powers of two)
937 auto tryToMatchSignedTruncationCheck = [](ICmpInst *ICmp, Value *&X,
938 APInt &SignBitMask) -> bool {
939 CmpInst::Predicate Pred;
940 const APInt *I01, *I1; // powers of two; I1 == I01 << 1
942 m_ICmp(Pred, m_Add(m_Value(X), m_Power2(I01)), m_Power2(I1))) &&
943 Pred == ICmpInst::ICMP_ULT && I1->ugt(*I01) && I01->shl(1) == *I1))
945 // Which bit is the new sign bit as per the 'signed truncation' pattern?
950 // One icmp needs to be 'signed truncation check'.
951 // We need to match this first, else we will mismatch commutative cases.
955 if (tryToMatchSignedTruncationCheck(ICmp1, X1, HighestBit))
957 else if (tryToMatchSignedTruncationCheck(ICmp0, X1, HighestBit))
962 assert(HighestBit.isPowerOf2() && "expected to be power of two (non-zero)");
964 // Try to match/decompose into: icmp eq (X & Mask), 0
965 auto tryToDecompose = [](ICmpInst *ICmp, Value *&X,
966 APInt &UnsetBitsMask) -> bool {
967 CmpInst::Predicate Pred = ICmp->getPredicate();
968 // Can it be decomposed into icmp eq (X & Mask), 0 ?
969 if (llvm::decomposeBitTestICmp(ICmp->getOperand(0), ICmp->getOperand(1),
970 Pred, X, UnsetBitsMask,
971 /*LookThruTrunc=*/false) &&
972 Pred == ICmpInst::ICMP_EQ)
974 // Is it icmp eq (X & Mask), 0 already?
976 if (match(ICmp, m_ICmp(Pred, m_And(m_Value(X), m_APInt(Mask)), m_Zero())) &&
977 Pred == ICmpInst::ICMP_EQ) {
978 UnsetBitsMask = *Mask;
984 // And the other icmp needs to be decomposable into a bit test.
987 if (!tryToDecompose(OtherICmp, X0, UnsetBitsMask))
990 assert(!UnsetBitsMask.isNullValue() && "empty mask makes no sense.");
992 // Are they working on the same value?
997 } else if (match(X0, m_Trunc(m_Specific(X1)))) {
998 UnsetBitsMask = UnsetBitsMask.zext(X1->getType()->getScalarSizeInBits());
1003 // So which bits should be uniform as per the 'signed truncation check'?
1004 // (all the bits starting with (i.e. including) HighestBit)
1005 APInt SignBitsMask = ~(HighestBit - 1U);
1007 // UnsetBitsMask must have some common bits with SignBitsMask,
1008 if (!UnsetBitsMask.intersects(SignBitsMask))
1011 // Does UnsetBitsMask contain any bits outside of SignBitsMask?
1012 if (!UnsetBitsMask.isSubsetOf(SignBitsMask)) {
1013 APInt OtherHighestBit = (~UnsetBitsMask) + 1U;
1014 if (!OtherHighestBit.isPowerOf2())
1016 HighestBit = APIntOps::umin(HighestBit, OtherHighestBit);
1018 // Else, if it does not, then all is ok as-is.
1020 // %r = icmp ult %X, SignBit
1021 return Builder.CreateICmpULT(X, ConstantInt::get(X->getType(), HighestBit),
1022 CxtI.getName() + ".simplified");
1025 /// Fold (icmp)&(icmp) if possible.
1026 Value *InstCombiner::foldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS,
1027 Instruction &CxtI) {
1028 // Fold (!iszero(A & K1) & !iszero(A & K2)) -> (A & (K1 | K2)) == (K1 | K2)
1029 // if K1 and K2 are a one-bit mask.
1030 if (Value *V = foldAndOrOfICmpsOfAndWithPow2(LHS, RHS, true, CxtI))
1033 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1035 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
1036 if (predicatesFoldable(PredL, PredR)) {
1037 if (LHS->getOperand(0) == RHS->getOperand(1) &&
1038 LHS->getOperand(1) == RHS->getOperand(0))
1039 LHS->swapOperands();
1040 if (LHS->getOperand(0) == RHS->getOperand(0) &&
1041 LHS->getOperand(1) == RHS->getOperand(1)) {
1042 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1043 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
1044 bool IsSigned = LHS->isSigned() || RHS->isSigned();
1045 return getNewICmpValue(Code, IsSigned, Op0, Op1, Builder);
1049 // handle (roughly): (icmp eq (A & B), C) & (icmp eq (A & D), E)
1050 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, true, Builder))
1053 // E.g. (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
1054 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/false))
1057 // E.g. (icmp slt x, n) & (icmp sge x, 0) --> icmp ult x, n
1058 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/false))
1061 if (Value *V = foldAndOrOfEqualityCmpsWithConstants(LHS, RHS, true, Builder))
1064 if (Value *V = foldSignedTruncationCheck(LHS, RHS, CxtI, Builder))
1067 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
1068 Value *LHS0 = LHS->getOperand(0), *RHS0 = RHS->getOperand(0);
1069 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHS->getOperand(1));
1070 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS->getOperand(1));
1074 if (LHSC == RHSC && PredL == PredR) {
1075 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
1076 // where C is a power of 2 or
1077 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
1078 if ((PredL == ICmpInst::ICMP_ULT && LHSC->getValue().isPowerOf2()) ||
1079 (PredL == ICmpInst::ICMP_EQ && LHSC->isZero())) {
1080 Value *NewOr = Builder.CreateOr(LHS0, RHS0);
1081 return Builder.CreateICmp(PredL, NewOr, LHSC);
1085 // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2
1086 // where CMAX is the all ones value for the truncated type,
1087 // iff the lower bits of C2 and CA are zero.
1088 if (PredL == ICmpInst::ICMP_EQ && PredL == PredR && LHS->hasOneUse() &&
1091 ConstantInt *AndC, *SmallC = nullptr, *BigC = nullptr;
1093 // (trunc x) == C1 & (and x, CA) == C2
1094 // (and x, CA) == C2 & (trunc x) == C1
1095 if (match(RHS0, m_Trunc(m_Value(V))) &&
1096 match(LHS0, m_And(m_Specific(V), m_ConstantInt(AndC)))) {
1099 } else if (match(LHS0, m_Trunc(m_Value(V))) &&
1100 match(RHS0, m_And(m_Specific(V), m_ConstantInt(AndC)))) {
1105 if (SmallC && BigC) {
1106 unsigned BigBitSize = BigC->getType()->getBitWidth();
1107 unsigned SmallBitSize = SmallC->getType()->getBitWidth();
1109 // Check that the low bits are zero.
1110 APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize);
1111 if ((Low & AndC->getValue()).isNullValue() &&
1112 (Low & BigC->getValue()).isNullValue()) {
1113 Value *NewAnd = Builder.CreateAnd(V, Low | AndC->getValue());
1114 APInt N = SmallC->getValue().zext(BigBitSize) | BigC->getValue();
1115 Value *NewVal = ConstantInt::get(AndC->getType()->getContext(), N);
1116 return Builder.CreateICmp(PredL, NewAnd, NewVal);
1121 // From here on, we only handle:
1122 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
1126 // ICMP_[US][GL]E X, C is folded to ICMP_[US][GL]T elsewhere.
1127 if (PredL == ICmpInst::ICMP_UGE || PredL == ICmpInst::ICMP_ULE ||
1128 PredR == ICmpInst::ICMP_UGE || PredR == ICmpInst::ICMP_ULE ||
1129 PredL == ICmpInst::ICMP_SGE || PredL == ICmpInst::ICMP_SLE ||
1130 PredR == ICmpInst::ICMP_SGE || PredR == ICmpInst::ICMP_SLE)
1133 // We can't fold (ugt x, C) & (sgt x, C2).
1134 if (!predicatesFoldable(PredL, PredR))
1137 // Ensure that the larger constant is on the RHS.
1139 if (CmpInst::isSigned(PredL) ||
1140 (ICmpInst::isEquality(PredL) && CmpInst::isSigned(PredR)))
1141 ShouldSwap = LHSC->getValue().sgt(RHSC->getValue());
1143 ShouldSwap = LHSC->getValue().ugt(RHSC->getValue());
1146 std::swap(LHS, RHS);
1147 std::swap(LHSC, RHSC);
1148 std::swap(PredL, PredR);
1151 // At this point, we know we have two icmp instructions
1152 // comparing a value against two constants and and'ing the result
1153 // together. Because of the above check, we know that we only have
1154 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
1155 // (from the icmp folding check above), that the two constants
1156 // are not equal and that the larger constant is on the RHS
1157 assert(LHSC != RHSC && "Compares not folded above?");
1161 llvm_unreachable("Unknown integer condition code!");
1162 case ICmpInst::ICMP_NE:
1165 llvm_unreachable("Unknown integer condition code!");
1166 case ICmpInst::ICMP_ULT:
1167 if (LHSC == SubOne(RHSC)) // (X != 13 & X u< 14) -> X < 13
1168 return Builder.CreateICmpULT(LHS0, LHSC);
1169 if (LHSC->isZero()) // (X != 0 & X u< 14) -> X-1 u< 13
1170 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(),
1172 break; // (X != 13 & X u< 15) -> no change
1173 case ICmpInst::ICMP_SLT:
1174 if (LHSC == SubOne(RHSC)) // (X != 13 & X s< 14) -> X < 13
1175 return Builder.CreateICmpSLT(LHS0, LHSC);
1176 break; // (X != 13 & X s< 15) -> no change
1177 case ICmpInst::ICMP_NE:
1178 // Potential folds for this case should already be handled.
1182 case ICmpInst::ICMP_UGT:
1185 llvm_unreachable("Unknown integer condition code!");
1186 case ICmpInst::ICMP_NE:
1187 if (RHSC == AddOne(LHSC)) // (X u> 13 & X != 14) -> X u> 14
1188 return Builder.CreateICmp(PredL, LHS0, RHSC);
1189 break; // (X u> 13 & X != 15) -> no change
1190 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
1191 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(),
1195 case ICmpInst::ICMP_SGT:
1198 llvm_unreachable("Unknown integer condition code!");
1199 case ICmpInst::ICMP_NE:
1200 if (RHSC == AddOne(LHSC)) // (X s> 13 & X != 14) -> X s> 14
1201 return Builder.CreateICmp(PredL, LHS0, RHSC);
1202 break; // (X s> 13 & X != 15) -> no change
1203 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
1204 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(), true,
1213 Value *InstCombiner::foldLogicOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1214 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1215 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1216 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1218 if (LHS0 == RHS1 && RHS0 == LHS1) {
1219 // Swap RHS operands to match LHS.
1220 PredR = FCmpInst::getSwappedPredicate(PredR);
1221 std::swap(RHS0, RHS1);
1224 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
1225 // Suppose the relation between x and y is R, where R is one of
1226 // U(1000), L(0100), G(0010) or E(0001), and CC0 and CC1 are the bitmasks for
1227 // testing the desired relations.
1229 // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this:
1230 // bool(R & CC0) && bool(R & CC1)
1231 // = bool((R & CC0) & (R & CC1))
1232 // = bool(R & (CC0 & CC1)) <= by re-association, commutation, and idempotency
1234 // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this:
1235 // bool(R & CC0) || bool(R & CC1)
1236 // = bool((R & CC0) | (R & CC1))
1237 // = bool(R & (CC0 | CC1)) <= by reversed distribution (contribution? ;)
1238 if (LHS0 == RHS0 && LHS1 == RHS1) {
1239 unsigned FCmpCodeL = getFCmpCode(PredL);
1240 unsigned FCmpCodeR = getFCmpCode(PredR);
1241 unsigned NewPred = IsAnd ? FCmpCodeL & FCmpCodeR : FCmpCodeL | FCmpCodeR;
1242 return getFCmpValue(NewPred, LHS0, LHS1, Builder);
1245 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1246 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1247 if (LHS0->getType() != RHS0->getType())
1250 // FCmp canonicalization ensures that (fcmp ord/uno X, X) and
1251 // (fcmp ord/uno X, C) will be transformed to (fcmp X, +0.0).
1252 if (match(LHS1, m_PosZeroFP()) && match(RHS1, m_PosZeroFP()))
1253 // Ignore the constants because they are obviously not NANs:
1254 // (fcmp ord x, 0.0) & (fcmp ord y, 0.0) -> (fcmp ord x, y)
1255 // (fcmp uno x, 0.0) | (fcmp uno y, 0.0) -> (fcmp uno x, y)
1256 return Builder.CreateFCmp(PredL, LHS0, RHS0);
1262 /// Match De Morgan's Laws:
1263 /// (~A & ~B) == (~(A | B))
1264 /// (~A | ~B) == (~(A & B))
1265 static Instruction *matchDeMorgansLaws(BinaryOperator &I,
1266 InstCombiner::BuilderTy &Builder) {
1267 auto Opcode = I.getOpcode();
1268 assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&
1269 "Trying to match De Morgan's Laws with something other than and/or");
1271 // Flip the logic operation.
1272 Opcode = (Opcode == Instruction::And) ? Instruction::Or : Instruction::And;
1275 if (match(I.getOperand(0), m_OneUse(m_Not(m_Value(A)))) &&
1276 match(I.getOperand(1), m_OneUse(m_Not(m_Value(B)))) &&
1277 !IsFreeToInvert(A, A->hasOneUse()) &&
1278 !IsFreeToInvert(B, B->hasOneUse())) {
1279 Value *AndOr = Builder.CreateBinOp(Opcode, A, B, I.getName() + ".demorgan");
1280 return BinaryOperator::CreateNot(AndOr);
1286 bool InstCombiner::shouldOptimizeCast(CastInst *CI) {
1287 Value *CastSrc = CI->getOperand(0);
1289 // Noop casts and casts of constants should be eliminated trivially.
1290 if (CI->getSrcTy() == CI->getDestTy() || isa<Constant>(CastSrc))
1293 // If this cast is paired with another cast that can be eliminated, we prefer
1294 // to have it eliminated.
1295 if (const auto *PrecedingCI = dyn_cast<CastInst>(CastSrc))
1296 if (isEliminableCastPair(PrecedingCI, CI))
1302 /// Fold {and,or,xor} (cast X), C.
1303 static Instruction *foldLogicCastConstant(BinaryOperator &Logic, CastInst *Cast,
1304 InstCombiner::BuilderTy &Builder) {
1305 Constant *C = dyn_cast<Constant>(Logic.getOperand(1));
1309 auto LogicOpc = Logic.getOpcode();
1310 Type *DestTy = Logic.getType();
1311 Type *SrcTy = Cast->getSrcTy();
1313 // Move the logic operation ahead of a zext or sext if the constant is
1314 // unchanged in the smaller source type. Performing the logic in a smaller
1315 // type may provide more information to later folds, and the smaller logic
1316 // instruction may be cheaper (particularly in the case of vectors).
1318 if (match(Cast, m_OneUse(m_ZExt(m_Value(X))))) {
1319 Constant *TruncC = ConstantExpr::getTrunc(C, SrcTy);
1320 Constant *ZextTruncC = ConstantExpr::getZExt(TruncC, DestTy);
1321 if (ZextTruncC == C) {
1322 // LogicOpc (zext X), C --> zext (LogicOpc X, C)
1323 Value *NewOp = Builder.CreateBinOp(LogicOpc, X, TruncC);
1324 return new ZExtInst(NewOp, DestTy);
1328 if (match(Cast, m_OneUse(m_SExt(m_Value(X))))) {
1329 Constant *TruncC = ConstantExpr::getTrunc(C, SrcTy);
1330 Constant *SextTruncC = ConstantExpr::getSExt(TruncC, DestTy);
1331 if (SextTruncC == C) {
1332 // LogicOpc (sext X), C --> sext (LogicOpc X, C)
1333 Value *NewOp = Builder.CreateBinOp(LogicOpc, X, TruncC);
1334 return new SExtInst(NewOp, DestTy);
1341 /// Fold {and,or,xor} (cast X), Y.
1342 Instruction *InstCombiner::foldCastedBitwiseLogic(BinaryOperator &I) {
1343 auto LogicOpc = I.getOpcode();
1344 assert(I.isBitwiseLogicOp() && "Unexpected opcode for bitwise logic folding");
1346 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1347 CastInst *Cast0 = dyn_cast<CastInst>(Op0);
1351 // This must be a cast from an integer or integer vector source type to allow
1352 // transformation of the logic operation to the source type.
1353 Type *DestTy = I.getType();
1354 Type *SrcTy = Cast0->getSrcTy();
1355 if (!SrcTy->isIntOrIntVectorTy())
1358 if (Instruction *Ret = foldLogicCastConstant(I, Cast0, Builder))
1361 CastInst *Cast1 = dyn_cast<CastInst>(Op1);
1365 // Both operands of the logic operation are casts. The casts must be of the
1366 // same type for reduction.
1367 auto CastOpcode = Cast0->getOpcode();
1368 if (CastOpcode != Cast1->getOpcode() || SrcTy != Cast1->getSrcTy())
1371 Value *Cast0Src = Cast0->getOperand(0);
1372 Value *Cast1Src = Cast1->getOperand(0);
1374 // fold logic(cast(A), cast(B)) -> cast(logic(A, B))
1375 if (shouldOptimizeCast(Cast0) && shouldOptimizeCast(Cast1)) {
1376 Value *NewOp = Builder.CreateBinOp(LogicOpc, Cast0Src, Cast1Src,
1378 return CastInst::Create(CastOpcode, NewOp, DestTy);
1381 // For now, only 'and'/'or' have optimizations after this.
1382 if (LogicOpc == Instruction::Xor)
1385 // If this is logic(cast(icmp), cast(icmp)), try to fold this even if the
1386 // cast is otherwise not optimizable. This happens for vector sexts.
1387 ICmpInst *ICmp0 = dyn_cast<ICmpInst>(Cast0Src);
1388 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(Cast1Src);
1389 if (ICmp0 && ICmp1) {
1390 Value *Res = LogicOpc == Instruction::And ? foldAndOfICmps(ICmp0, ICmp1, I)
1391 : foldOrOfICmps(ICmp0, ICmp1, I);
1393 return CastInst::Create(CastOpcode, Res, DestTy);
1397 // If this is logic(cast(fcmp), cast(fcmp)), try to fold this even if the
1398 // cast is otherwise not optimizable. This happens for vector sexts.
1399 FCmpInst *FCmp0 = dyn_cast<FCmpInst>(Cast0Src);
1400 FCmpInst *FCmp1 = dyn_cast<FCmpInst>(Cast1Src);
1402 if (Value *R = foldLogicOfFCmps(FCmp0, FCmp1, LogicOpc == Instruction::And))
1403 return CastInst::Create(CastOpcode, R, DestTy);
1408 static Instruction *foldAndToXor(BinaryOperator &I,
1409 InstCombiner::BuilderTy &Builder) {
1410 assert(I.getOpcode() == Instruction::And);
1411 Value *Op0 = I.getOperand(0);
1412 Value *Op1 = I.getOperand(1);
1415 // Operand complexity canonicalization guarantees that the 'or' is Op0.
1416 // (A | B) & ~(A & B) --> A ^ B
1417 // (A | B) & ~(B & A) --> A ^ B
1418 if (match(&I, m_BinOp(m_Or(m_Value(A), m_Value(B)),
1419 m_Not(m_c_And(m_Deferred(A), m_Deferred(B))))))
1420 return BinaryOperator::CreateXor(A, B);
1422 // (A | ~B) & (~A | B) --> ~(A ^ B)
1423 // (A | ~B) & (B | ~A) --> ~(A ^ B)
1424 // (~B | A) & (~A | B) --> ~(A ^ B)
1425 // (~B | A) & (B | ~A) --> ~(A ^ B)
1426 if (Op0->hasOneUse() || Op1->hasOneUse())
1427 if (match(&I, m_BinOp(m_c_Or(m_Value(A), m_Not(m_Value(B))),
1428 m_c_Or(m_Not(m_Deferred(A)), m_Deferred(B)))))
1429 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
1434 static Instruction *foldOrToXor(BinaryOperator &I,
1435 InstCombiner::BuilderTy &Builder) {
1436 assert(I.getOpcode() == Instruction::Or);
1437 Value *Op0 = I.getOperand(0);
1438 Value *Op1 = I.getOperand(1);
1441 // Operand complexity canonicalization guarantees that the 'and' is Op0.
1442 // (A & B) | ~(A | B) --> ~(A ^ B)
1443 // (A & B) | ~(B | A) --> ~(A ^ B)
1444 if (Op0->hasOneUse() || Op1->hasOneUse())
1445 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1446 match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B)))))
1447 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
1449 // (A & ~B) | (~A & B) --> A ^ B
1450 // (A & ~B) | (B & ~A) --> A ^ B
1451 // (~B & A) | (~A & B) --> A ^ B
1452 // (~B & A) | (B & ~A) --> A ^ B
1453 if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) &&
1454 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B))))
1455 return BinaryOperator::CreateXor(A, B);
1460 /// Return true if a constant shift amount is always less than the specified
1461 /// bit-width. If not, the shift could create poison in the narrower type.
1462 static bool canNarrowShiftAmt(Constant *C, unsigned BitWidth) {
1463 if (auto *ScalarC = dyn_cast<ConstantInt>(C))
1464 return ScalarC->getZExtValue() < BitWidth;
1466 if (C->getType()->isVectorTy()) {
1467 // Check each element of a constant vector.
1468 unsigned NumElts = C->getType()->getVectorNumElements();
1469 for (unsigned i = 0; i != NumElts; ++i) {
1470 Constant *Elt = C->getAggregateElement(i);
1473 if (isa<UndefValue>(Elt))
1475 auto *CI = dyn_cast<ConstantInt>(Elt);
1476 if (!CI || CI->getZExtValue() >= BitWidth)
1482 // The constant is a constant expression or unknown.
1486 /// Try to use narrower ops (sink zext ops) for an 'and' with binop operand and
1487 /// a common zext operand: and (binop (zext X), C), (zext X).
1488 Instruction *InstCombiner::narrowMaskedBinOp(BinaryOperator &And) {
1489 // This transform could also apply to {or, and, xor}, but there are better
1490 // folds for those cases, so we don't expect those patterns here. AShr is not
1491 // handled because it should always be transformed to LShr in this sequence.
1492 // The subtract transform is different because it has a constant on the left.
1493 // Add/mul commute the constant to RHS; sub with constant RHS becomes add.
1494 Value *Op0 = And.getOperand(0), *Op1 = And.getOperand(1);
1496 if (!match(Op0, m_OneUse(m_Add(m_Specific(Op1), m_Constant(C)))) &&
1497 !match(Op0, m_OneUse(m_Mul(m_Specific(Op1), m_Constant(C)))) &&
1498 !match(Op0, m_OneUse(m_LShr(m_Specific(Op1), m_Constant(C)))) &&
1499 !match(Op0, m_OneUse(m_Shl(m_Specific(Op1), m_Constant(C)))) &&
1500 !match(Op0, m_OneUse(m_Sub(m_Constant(C), m_Specific(Op1)))))
1504 if (!match(Op1, m_ZExt(m_Value(X))) || Op1->hasNUsesOrMore(3))
1507 Type *Ty = And.getType();
1508 if (!isa<VectorType>(Ty) && !shouldChangeType(Ty, X->getType()))
1511 // If we're narrowing a shift, the shift amount must be safe (less than the
1512 // width) in the narrower type. If the shift amount is greater, instsimplify
1513 // usually handles that case, but we can't guarantee/assert it.
1514 Instruction::BinaryOps Opc = cast<BinaryOperator>(Op0)->getOpcode();
1515 if (Opc == Instruction::LShr || Opc == Instruction::Shl)
1516 if (!canNarrowShiftAmt(C, X->getType()->getScalarSizeInBits()))
1519 // and (sub C, (zext X)), (zext X) --> zext (and (sub C', X), X)
1520 // and (binop (zext X), C), (zext X) --> zext (and (binop X, C'), X)
1521 Value *NewC = ConstantExpr::getTrunc(C, X->getType());
1522 Value *NewBO = Opc == Instruction::Sub ? Builder.CreateBinOp(Opc, NewC, X)
1523 : Builder.CreateBinOp(Opc, X, NewC);
1524 return new ZExtInst(Builder.CreateAnd(NewBO, X), Ty);
1527 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
1528 // here. We should standardize that construct where it is needed or choose some
1529 // other way to ensure that commutated variants of patterns are not missed.
1530 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
1531 if (Value *V = SimplifyAndInst(I.getOperand(0), I.getOperand(1),
1532 SQ.getWithInstruction(&I)))
1533 return replaceInstUsesWith(I, V);
1535 if (SimplifyAssociativeOrCommutative(I))
1538 if (Instruction *X = foldVectorBinop(I))
1541 // See if we can simplify any instructions used by the instruction whose sole
1542 // purpose is to compute bits we don't care about.
1543 if (SimplifyDemandedInstructionBits(I))
1546 // Do this before using distributive laws to catch simple and/or/not patterns.
1547 if (Instruction *Xor = foldAndToXor(I, Builder))
1550 // (A|B)&(A|C) -> A|(B&C) etc
1551 if (Value *V = SimplifyUsingDistributiveLaws(I))
1552 return replaceInstUsesWith(I, V);
1554 if (Value *V = SimplifyBSwap(I, Builder))
1555 return replaceInstUsesWith(I, V);
1557 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1559 if (match(Op1, m_APInt(C))) {
1561 if (match(Op0, m_OneUse(m_LogicalShift(m_One(), m_Value(X)))) &&
1563 // (1 << X) & 1 --> zext(X == 0)
1564 // (1 >> X) & 1 --> zext(X == 0)
1565 Value *IsZero = Builder.CreateICmpEQ(X, ConstantInt::get(I.getType(), 0));
1566 return new ZExtInst(IsZero, I.getType());
1570 if (match(Op0, m_OneUse(m_Xor(m_Value(X), m_APInt(XorC))))) {
1571 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
1572 Constant *NewC = ConstantInt::get(I.getType(), *C & *XorC);
1573 Value *And = Builder.CreateAnd(X, Op1);
1575 return BinaryOperator::CreateXor(And, NewC);
1579 if (match(Op0, m_OneUse(m_Or(m_Value(X), m_APInt(OrC))))) {
1580 // (X | C1) & C2 --> (X & C2^(C1&C2)) | (C1&C2)
1581 // NOTE: This reduces the number of bits set in the & mask, which
1582 // can expose opportunities for store narrowing for scalars.
1583 // NOTE: SimplifyDemandedBits should have already removed bits from C1
1584 // that aren't set in C2. Meaning we can replace (C1&C2) with C1 in
1585 // above, but this feels safer.
1586 APInt Together = *C & *OrC;
1587 Value *And = Builder.CreateAnd(X, ConstantInt::get(I.getType(),
1590 return BinaryOperator::CreateOr(And, ConstantInt::get(I.getType(),
1594 // If the mask is only needed on one incoming arm, push the 'and' op up.
1595 if (match(Op0, m_OneUse(m_Xor(m_Value(X), m_Value(Y)))) ||
1596 match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
1597 APInt NotAndMask(~(*C));
1598 BinaryOperator::BinaryOps BinOp = cast<BinaryOperator>(Op0)->getOpcode();
1599 if (MaskedValueIsZero(X, NotAndMask, 0, &I)) {
1600 // Not masking anything out for the LHS, move mask to RHS.
1601 // and ({x}or X, Y), C --> {x}or X, (and Y, C)
1602 Value *NewRHS = Builder.CreateAnd(Y, Op1, Y->getName() + ".masked");
1603 return BinaryOperator::Create(BinOp, X, NewRHS);
1605 if (!isa<Constant>(Y) && MaskedValueIsZero(Y, NotAndMask, 0, &I)) {
1606 // Not masking anything out for the RHS, move mask to LHS.
1607 // and ({x}or X, Y), C --> {x}or (and X, C), Y
1608 Value *NewLHS = Builder.CreateAnd(X, Op1, X->getName() + ".masked");
1609 return BinaryOperator::Create(BinOp, NewLHS, Y);
1615 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
1616 const APInt &AndRHSMask = AndRHS->getValue();
1618 // Optimize a variety of ((val OP C1) & C2) combinations...
1619 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1620 // ((C1 OP zext(X)) & C2) -> zext((C1-X) & C2) if C2 fits in the bitwidth
1621 // of X and OP behaves well when given trunc(C1) and X.
1622 switch (Op0I->getOpcode()) {
1625 case Instruction::Xor:
1626 case Instruction::Or:
1627 case Instruction::Mul:
1628 case Instruction::Add:
1629 case Instruction::Sub:
1632 if (match(Op0I, m_c_BinOp(m_ZExt(m_Value(X)), m_ConstantInt(C1)))) {
1633 if (AndRHSMask.isIntN(X->getType()->getScalarSizeInBits())) {
1634 auto *TruncC1 = ConstantExpr::getTrunc(C1, X->getType());
1636 Value *Op0LHS = Op0I->getOperand(0);
1637 if (isa<ZExtInst>(Op0LHS))
1638 BinOp = Builder.CreateBinOp(Op0I->getOpcode(), X, TruncC1);
1640 BinOp = Builder.CreateBinOp(Op0I->getOpcode(), TruncC1, X);
1641 auto *TruncC2 = ConstantExpr::getTrunc(AndRHS, X->getType());
1642 auto *And = Builder.CreateAnd(BinOp, TruncC2);
1643 return new ZExtInst(And, I.getType());
1648 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
1649 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
1653 // If this is an integer truncation, and if the source is an 'and' with
1654 // immediate, transform it. This frequently occurs for bitfield accesses.
1656 Value *X = nullptr; ConstantInt *YC = nullptr;
1657 if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) {
1658 // Change: and (trunc (and X, YC) to T), C2
1659 // into : and (trunc X to T), trunc(YC) & C2
1660 // This will fold the two constants together, which may allow
1661 // other simplifications.
1662 Value *NewCast = Builder.CreateTrunc(X, I.getType(), "and.shrunk");
1663 Constant *C3 = ConstantExpr::getTrunc(YC, I.getType());
1664 C3 = ConstantExpr::getAnd(C3, AndRHS);
1665 return BinaryOperator::CreateAnd(NewCast, C3);
1670 if (Instruction *Z = narrowMaskedBinOp(I))
1673 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))
1676 if (Instruction *DeMorgan = matchDeMorgansLaws(I, Builder))
1681 // A & (A ^ B) --> A & ~B
1682 if (match(Op1, m_OneUse(m_c_Xor(m_Specific(Op0), m_Value(B)))))
1683 return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(B));
1684 // (A ^ B) & A --> A & ~B
1685 if (match(Op0, m_OneUse(m_c_Xor(m_Specific(Op1), m_Value(B)))))
1686 return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(B));
1688 // (A ^ B) & ((B ^ C) ^ A) -> (A ^ B) & ~C
1689 if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
1690 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
1691 if (Op1->hasOneUse() || IsFreeToInvert(C, C->hasOneUse()))
1692 return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(C));
1694 // ((A ^ C) ^ B) & (B ^ A) -> (B ^ A) & ~C
1695 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
1696 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
1697 if (Op0->hasOneUse() || IsFreeToInvert(C, C->hasOneUse()))
1698 return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(C));
1700 // (A | B) & ((~A) ^ B) -> (A & B)
1701 // (A | B) & (B ^ (~A)) -> (A & B)
1702 // (B | A) & ((~A) ^ B) -> (A & B)
1703 // (B | A) & (B ^ (~A)) -> (A & B)
1704 if (match(Op1, m_c_Xor(m_Not(m_Value(A)), m_Value(B))) &&
1705 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1706 return BinaryOperator::CreateAnd(A, B);
1708 // ((~A) ^ B) & (A | B) -> (A & B)
1709 // ((~A) ^ B) & (B | A) -> (A & B)
1710 // (B ^ (~A)) & (A | B) -> (A & B)
1711 // (B ^ (~A)) & (B | A) -> (A & B)
1712 if (match(Op0, m_c_Xor(m_Not(m_Value(A)), m_Value(B))) &&
1713 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
1714 return BinaryOperator::CreateAnd(A, B);
1718 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
1719 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
1721 if (Value *Res = foldAndOfICmps(LHS, RHS, I))
1722 return replaceInstUsesWith(I, Res);
1724 // TODO: Make this recursive; it's a little tricky because an arbitrary
1725 // number of 'and' instructions might have to be created.
1727 if (LHS && match(Op1, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
1728 if (auto *Cmp = dyn_cast<ICmpInst>(X))
1729 if (Value *Res = foldAndOfICmps(LHS, Cmp, I))
1730 return replaceInstUsesWith(I, Builder.CreateAnd(Res, Y));
1731 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
1732 if (Value *Res = foldAndOfICmps(LHS, Cmp, I))
1733 return replaceInstUsesWith(I, Builder.CreateAnd(Res, X));
1735 if (RHS && match(Op0, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
1736 if (auto *Cmp = dyn_cast<ICmpInst>(X))
1737 if (Value *Res = foldAndOfICmps(Cmp, RHS, I))
1738 return replaceInstUsesWith(I, Builder.CreateAnd(Res, Y));
1739 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
1740 if (Value *Res = foldAndOfICmps(Cmp, RHS, I))
1741 return replaceInstUsesWith(I, Builder.CreateAnd(Res, X));
1745 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1746 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1747 if (Value *Res = foldLogicOfFCmps(LHS, RHS, true))
1748 return replaceInstUsesWith(I, Res);
1750 if (Instruction *CastedAnd = foldCastedBitwiseLogic(I))
1753 // and(sext(A), B) / and(B, sext(A)) --> A ? B : 0, where A is i1 or <N x i1>.
1755 if (match(Op0, m_OneUse(m_SExt(m_Value(A)))) &&
1756 A->getType()->isIntOrIntVectorTy(1))
1757 return SelectInst::Create(A, Op1, Constant::getNullValue(I.getType()));
1758 if (match(Op1, m_OneUse(m_SExt(m_Value(A)))) &&
1759 A->getType()->isIntOrIntVectorTy(1))
1760 return SelectInst::Create(A, Op0, Constant::getNullValue(I.getType()));
1765 Instruction *InstCombiner::matchBSwap(BinaryOperator &Or) {
1766 assert(Or.getOpcode() == Instruction::Or && "bswap requires an 'or'");
1767 Value *Op0 = Or.getOperand(0), *Op1 = Or.getOperand(1);
1769 // Look through zero extends.
1770 if (Instruction *Ext = dyn_cast<ZExtInst>(Op0))
1771 Op0 = Ext->getOperand(0);
1773 if (Instruction *Ext = dyn_cast<ZExtInst>(Op1))
1774 Op1 = Ext->getOperand(0);
1776 // (A | B) | C and A | (B | C) -> bswap if possible.
1777 bool OrOfOrs = match(Op0, m_Or(m_Value(), m_Value())) ||
1778 match(Op1, m_Or(m_Value(), m_Value()));
1780 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
1781 bool OrOfShifts = match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
1782 match(Op1, m_LogicalShift(m_Value(), m_Value()));
1784 // (A & B) | (C & D) -> bswap if possible.
1785 bool OrOfAnds = match(Op0, m_And(m_Value(), m_Value())) &&
1786 match(Op1, m_And(m_Value(), m_Value()));
1788 // (A << B) | (C & D) -> bswap if possible.
1789 // The bigger pattern here is ((A & C1) << C2) | ((B >> C2) & C1), which is a
1790 // part of the bswap idiom for specific values of C1, C2 (e.g. C1 = 16711935,
1792 // This pattern can occur when the operands of the 'or' are not canonicalized
1793 // for some reason (not having only one use, for example).
1794 bool OrOfAndAndSh = (match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
1795 match(Op1, m_And(m_Value(), m_Value()))) ||
1796 (match(Op0, m_And(m_Value(), m_Value())) &&
1797 match(Op1, m_LogicalShift(m_Value(), m_Value())));
1799 if (!OrOfOrs && !OrOfShifts && !OrOfAnds && !OrOfAndAndSh)
1802 SmallVector<Instruction*, 4> Insts;
1803 if (!recognizeBSwapOrBitReverseIdiom(&Or, true, false, Insts))
1805 Instruction *LastInst = Insts.pop_back_val();
1806 LastInst->removeFromParent();
1808 for (auto *Inst : Insts)
1813 /// Transform UB-safe variants of bitwise rotate to the funnel shift intrinsic.
1814 static Instruction *matchRotate(Instruction &Or) {
1815 // TODO: Can we reduce the code duplication between this and the related
1816 // rotate matching code under visitSelect and visitTrunc?
1817 unsigned Width = Or.getType()->getScalarSizeInBits();
1818 if (!isPowerOf2_32(Width))
1821 // First, find an or'd pair of opposite shifts with the same shifted operand:
1822 // or (lshr ShVal, ShAmt0), (shl ShVal, ShAmt1)
1823 Value *Or0 = Or.getOperand(0), *Or1 = Or.getOperand(1);
1824 Value *ShVal, *ShAmt0, *ShAmt1;
1825 if (!match(Or0, m_OneUse(m_LogicalShift(m_Value(ShVal), m_Value(ShAmt0)))) ||
1826 !match(Or1, m_OneUse(m_LogicalShift(m_Specific(ShVal), m_Value(ShAmt1)))))
1829 auto ShiftOpcode0 = cast<BinaryOperator>(Or0)->getOpcode();
1830 auto ShiftOpcode1 = cast<BinaryOperator>(Or1)->getOpcode();
1831 if (ShiftOpcode0 == ShiftOpcode1)
1834 // Match the shift amount operands for a rotate pattern. This always matches
1835 // a subtraction on the R operand.
1836 auto matchShiftAmount = [](Value *L, Value *R, unsigned Width) -> Value * {
1837 // The shift amount may be masked with negation:
1838 // (shl ShVal, (X & (Width - 1))) | (lshr ShVal, ((-X) & (Width - 1)))
1840 unsigned Mask = Width - 1;
1841 if (match(L, m_And(m_Value(X), m_SpecificInt(Mask))) &&
1842 match(R, m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask))))
1848 Value *ShAmt = matchShiftAmount(ShAmt0, ShAmt1, Width);
1849 bool SubIsOnLHS = false;
1851 ShAmt = matchShiftAmount(ShAmt1, ShAmt0, Width);
1857 bool IsFshl = (!SubIsOnLHS && ShiftOpcode0 == BinaryOperator::Shl) ||
1858 (SubIsOnLHS && ShiftOpcode1 == BinaryOperator::Shl);
1859 Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr;
1860 Function *F = Intrinsic::getDeclaration(Or.getModule(), IID, Or.getType());
1861 return IntrinsicInst::Create(F, { ShVal, ShVal, ShAmt });
1864 /// If all elements of two constant vectors are 0/-1 and inverses, return true.
1865 static bool areInverseVectorBitmasks(Constant *C1, Constant *C2) {
1866 unsigned NumElts = C1->getType()->getVectorNumElements();
1867 for (unsigned i = 0; i != NumElts; ++i) {
1868 Constant *EltC1 = C1->getAggregateElement(i);
1869 Constant *EltC2 = C2->getAggregateElement(i);
1870 if (!EltC1 || !EltC2)
1873 // One element must be all ones, and the other must be all zeros.
1874 if (!((match(EltC1, m_Zero()) && match(EltC2, m_AllOnes())) ||
1875 (match(EltC2, m_Zero()) && match(EltC1, m_AllOnes()))))
1881 /// We have an expression of the form (A & C) | (B & D). If A is a scalar or
1882 /// vector composed of all-zeros or all-ones values and is the bitwise 'not' of
1883 /// B, it can be used as the condition operand of a select instruction.
1884 Value *InstCombiner::getSelectCondition(Value *A, Value *B) {
1885 // Step 1: We may have peeked through bitcasts in the caller.
1886 // Exit immediately if we don't have (vector) integer types.
1887 Type *Ty = A->getType();
1888 if (!Ty->isIntOrIntVectorTy() || !B->getType()->isIntOrIntVectorTy())
1891 // Step 2: We need 0 or all-1's bitmasks.
1892 if (ComputeNumSignBits(A) != Ty->getScalarSizeInBits())
1895 // Step 3: If B is the 'not' value of A, we have our answer.
1896 if (match(A, m_Not(m_Specific(B)))) {
1897 // If these are scalars or vectors of i1, A can be used directly.
1898 if (Ty->isIntOrIntVectorTy(1))
1900 return Builder.CreateTrunc(A, CmpInst::makeCmpResultType(Ty));
1903 // If both operands are constants, see if the constants are inverse bitmasks.
1904 Constant *AConst, *BConst;
1905 if (match(A, m_Constant(AConst)) && match(B, m_Constant(BConst)))
1906 if (AConst == ConstantExpr::getNot(BConst))
1907 return Builder.CreateZExtOrTrunc(A, CmpInst::makeCmpResultType(Ty));
1909 // Look for more complex patterns. The 'not' op may be hidden behind various
1910 // casts. Look through sexts and bitcasts to find the booleans.
1913 if (match(A, m_SExt(m_Value(Cond))) &&
1914 Cond->getType()->isIntOrIntVectorTy(1) &&
1915 match(B, m_OneUse(m_Not(m_Value(NotB))))) {
1916 NotB = peekThroughBitcast(NotB, true);
1917 if (match(NotB, m_SExt(m_Specific(Cond))))
1921 // All scalar (and most vector) possibilities should be handled now.
1922 // Try more matches that only apply to non-splat constant vectors.
1923 if (!Ty->isVectorTy())
1926 // If both operands are xor'd with constants using the same sexted boolean
1927 // operand, see if the constants are inverse bitmasks.
1928 // TODO: Use ConstantExpr::getNot()?
1929 if (match(A, (m_Xor(m_SExt(m_Value(Cond)), m_Constant(AConst)))) &&
1930 match(B, (m_Xor(m_SExt(m_Specific(Cond)), m_Constant(BConst)))) &&
1931 Cond->getType()->isIntOrIntVectorTy(1) &&
1932 areInverseVectorBitmasks(AConst, BConst)) {
1933 AConst = ConstantExpr::getTrunc(AConst, CmpInst::makeCmpResultType(Ty));
1934 return Builder.CreateXor(Cond, AConst);
1939 /// We have an expression of the form (A & C) | (B & D). Try to simplify this
1940 /// to "A' ? C : D", where A' is a boolean or vector of booleans.
1941 Value *InstCombiner::matchSelectFromAndOr(Value *A, Value *C, Value *B,
1943 // The potential condition of the select may be bitcasted. In that case, look
1944 // through its bitcast and the corresponding bitcast of the 'not' condition.
1945 Type *OrigType = A->getType();
1946 A = peekThroughBitcast(A, true);
1947 B = peekThroughBitcast(B, true);
1948 if (Value *Cond = getSelectCondition(A, B)) {
1949 // ((bc Cond) & C) | ((bc ~Cond) & D) --> bc (select Cond, (bc C), (bc D))
1950 // The bitcasts will either all exist or all not exist. The builder will
1951 // not create unnecessary casts if the types already match.
1952 Value *BitcastC = Builder.CreateBitCast(C, A->getType());
1953 Value *BitcastD = Builder.CreateBitCast(D, A->getType());
1954 Value *Select = Builder.CreateSelect(Cond, BitcastC, BitcastD);
1955 return Builder.CreateBitCast(Select, OrigType);
1961 /// Fold (icmp)|(icmp) if possible.
1962 Value *InstCombiner::foldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS,
1963 Instruction &CxtI) {
1964 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2)
1965 // if K1 and K2 are a one-bit mask.
1966 if (Value *V = foldAndOrOfICmpsOfAndWithPow2(LHS, RHS, false, CxtI))
1969 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1971 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHS->getOperand(1));
1972 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS->getOperand(1));
1974 // Fold (icmp ult/ule (A + C1), C3) | (icmp ult/ule (A + C2), C3)
1975 // --> (icmp ult/ule ((A & ~(C1 ^ C2)) + max(C1, C2)), C3)
1976 // The original condition actually refers to the following two ranges:
1977 // [MAX_UINT-C1+1, MAX_UINT-C1+1+C3] and [MAX_UINT-C2+1, MAX_UINT-C2+1+C3]
1978 // We can fold these two ranges if:
1979 // 1) C1 and C2 is unsigned greater than C3.
1980 // 2) The two ranges are separated.
1981 // 3) C1 ^ C2 is one-bit mask.
1982 // 4) LowRange1 ^ LowRange2 and HighRange1 ^ HighRange2 are one-bit mask.
1983 // This implies all values in the two ranges differ by exactly one bit.
1985 if ((PredL == ICmpInst::ICMP_ULT || PredL == ICmpInst::ICMP_ULE) &&
1986 PredL == PredR && LHSC && RHSC && LHS->hasOneUse() && RHS->hasOneUse() &&
1987 LHSC->getType() == RHSC->getType() &&
1988 LHSC->getValue() == (RHSC->getValue())) {
1990 Value *LAdd = LHS->getOperand(0);
1991 Value *RAdd = RHS->getOperand(0);
1993 Value *LAddOpnd, *RAddOpnd;
1994 ConstantInt *LAddC, *RAddC;
1995 if (match(LAdd, m_Add(m_Value(LAddOpnd), m_ConstantInt(LAddC))) &&
1996 match(RAdd, m_Add(m_Value(RAddOpnd), m_ConstantInt(RAddC))) &&
1997 LAddC->getValue().ugt(LHSC->getValue()) &&
1998 RAddC->getValue().ugt(LHSC->getValue())) {
2000 APInt DiffC = LAddC->getValue() ^ RAddC->getValue();
2001 if (LAddOpnd == RAddOpnd && DiffC.isPowerOf2()) {
2002 ConstantInt *MaxAddC = nullptr;
2003 if (LAddC->getValue().ult(RAddC->getValue()))
2008 APInt RRangeLow = -RAddC->getValue();
2009 APInt RRangeHigh = RRangeLow + LHSC->getValue();
2010 APInt LRangeLow = -LAddC->getValue();
2011 APInt LRangeHigh = LRangeLow + LHSC->getValue();
2012 APInt LowRangeDiff = RRangeLow ^ LRangeLow;
2013 APInt HighRangeDiff = RRangeHigh ^ LRangeHigh;
2014 APInt RangeDiff = LRangeLow.sgt(RRangeLow) ? LRangeLow - RRangeLow
2015 : RRangeLow - LRangeLow;
2017 if (LowRangeDiff.isPowerOf2() && LowRangeDiff == HighRangeDiff &&
2018 RangeDiff.ugt(LHSC->getValue())) {
2019 Value *MaskC = ConstantInt::get(LAddC->getType(), ~DiffC);
2021 Value *NewAnd = Builder.CreateAnd(LAddOpnd, MaskC);
2022 Value *NewAdd = Builder.CreateAdd(NewAnd, MaxAddC);
2023 return Builder.CreateICmp(LHS->getPredicate(), NewAdd, LHSC);
2029 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
2030 if (predicatesFoldable(PredL, PredR)) {
2031 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2032 LHS->getOperand(1) == RHS->getOperand(0))
2033 LHS->swapOperands();
2034 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2035 LHS->getOperand(1) == RHS->getOperand(1)) {
2036 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2037 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
2038 bool IsSigned = LHS->isSigned() || RHS->isSigned();
2039 return getNewICmpValue(Code, IsSigned, Op0, Op1, Builder);
2043 // handle (roughly):
2044 // (icmp ne (A & B), C) | (icmp ne (A & D), E)
2045 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, false, Builder))
2048 Value *LHS0 = LHS->getOperand(0), *RHS0 = RHS->getOperand(0);
2049 if (LHS->hasOneUse() || RHS->hasOneUse()) {
2050 // (icmp eq B, 0) | (icmp ult A, B) -> (icmp ule A, B-1)
2051 // (icmp eq B, 0) | (icmp ugt B, A) -> (icmp ule A, B-1)
2052 Value *A = nullptr, *B = nullptr;
2053 if (PredL == ICmpInst::ICMP_EQ && LHSC && LHSC->isZero()) {
2055 if (PredR == ICmpInst::ICMP_ULT && LHS0 == RHS->getOperand(1))
2057 else if (PredR == ICmpInst::ICMP_UGT && LHS0 == RHS0)
2058 A = RHS->getOperand(1);
2060 // (icmp ult A, B) | (icmp eq B, 0) -> (icmp ule A, B-1)
2061 // (icmp ugt B, A) | (icmp eq B, 0) -> (icmp ule A, B-1)
2062 else if (PredR == ICmpInst::ICMP_EQ && RHSC && RHSC->isZero()) {
2064 if (PredL == ICmpInst::ICMP_ULT && RHS0 == LHS->getOperand(1))
2066 else if (PredL == ICmpInst::ICMP_UGT && LHS0 == RHS0)
2067 A = LHS->getOperand(1);
2070 return Builder.CreateICmp(
2072 Builder.CreateAdd(B, ConstantInt::getSigned(B->getType(), -1)), A);
2075 // E.g. (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
2076 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/true))
2079 // E.g. (icmp sgt x, n) | (icmp slt x, 0) --> icmp ugt x, n
2080 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/true))
2083 if (Value *V = foldAndOrOfEqualityCmpsWithConstants(LHS, RHS, false, Builder))
2086 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
2090 if (LHSC == RHSC && PredL == PredR) {
2091 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
2092 if (PredL == ICmpInst::ICMP_NE && LHSC->isZero()) {
2093 Value *NewOr = Builder.CreateOr(LHS0, RHS0);
2094 return Builder.CreateICmp(PredL, NewOr, LHSC);
2098 // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
2099 // iff C2 + CA == C1.
2100 if (PredL == ICmpInst::ICMP_ULT && PredR == ICmpInst::ICMP_EQ) {
2102 if (match(LHS0, m_Add(m_Specific(RHS0), m_ConstantInt(AddC))))
2103 if (RHSC->getValue() + AddC->getValue() == LHSC->getValue())
2104 return Builder.CreateICmpULE(LHS0, LHSC);
2107 // From here on, we only handle:
2108 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
2112 // ICMP_[US][GL]E X, C is folded to ICMP_[US][GL]T elsewhere.
2113 if (PredL == ICmpInst::ICMP_UGE || PredL == ICmpInst::ICMP_ULE ||
2114 PredR == ICmpInst::ICMP_UGE || PredR == ICmpInst::ICMP_ULE ||
2115 PredL == ICmpInst::ICMP_SGE || PredL == ICmpInst::ICMP_SLE ||
2116 PredR == ICmpInst::ICMP_SGE || PredR == ICmpInst::ICMP_SLE)
2119 // We can't fold (ugt x, C) | (sgt x, C2).
2120 if (!predicatesFoldable(PredL, PredR))
2123 // Ensure that the larger constant is on the RHS.
2125 if (CmpInst::isSigned(PredL) ||
2126 (ICmpInst::isEquality(PredL) && CmpInst::isSigned(PredR)))
2127 ShouldSwap = LHSC->getValue().sgt(RHSC->getValue());
2129 ShouldSwap = LHSC->getValue().ugt(RHSC->getValue());
2132 std::swap(LHS, RHS);
2133 std::swap(LHSC, RHSC);
2134 std::swap(PredL, PredR);
2137 // At this point, we know we have two icmp instructions
2138 // comparing a value against two constants and or'ing the result
2139 // together. Because of the above check, we know that we only have
2140 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
2141 // icmp folding check above), that the two constants are not
2143 assert(LHSC != RHSC && "Compares not folded above?");
2147 llvm_unreachable("Unknown integer condition code!");
2148 case ICmpInst::ICMP_EQ:
2151 llvm_unreachable("Unknown integer condition code!");
2152 case ICmpInst::ICMP_EQ:
2153 // Potential folds for this case should already be handled.
2155 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
2156 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
2160 case ICmpInst::ICMP_ULT:
2163 llvm_unreachable("Unknown integer condition code!");
2164 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
2166 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
2167 assert(!RHSC->isMaxValue(false) && "Missed icmp simplification");
2168 return insertRangeTest(LHS0, LHSC->getValue(), RHSC->getValue() + 1,
2172 case ICmpInst::ICMP_SLT:
2175 llvm_unreachable("Unknown integer condition code!");
2176 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
2178 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
2179 assert(!RHSC->isMaxValue(true) && "Missed icmp simplification");
2180 return insertRangeTest(LHS0, LHSC->getValue(), RHSC->getValue() + 1, true,
2188 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
2189 // here. We should standardize that construct where it is needed or choose some
2190 // other way to ensure that commutated variants of patterns are not missed.
2191 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
2192 if (Value *V = SimplifyOrInst(I.getOperand(0), I.getOperand(1),
2193 SQ.getWithInstruction(&I)))
2194 return replaceInstUsesWith(I, V);
2196 if (SimplifyAssociativeOrCommutative(I))
2199 if (Instruction *X = foldVectorBinop(I))
2202 // See if we can simplify any instructions used by the instruction whose sole
2203 // purpose is to compute bits we don't care about.
2204 if (SimplifyDemandedInstructionBits(I))
2207 // Do this before using distributive laws to catch simple and/or/not patterns.
2208 if (Instruction *Xor = foldOrToXor(I, Builder))
2211 // (A&B)|(A&C) -> A&(B|C) etc
2212 if (Value *V = SimplifyUsingDistributiveLaws(I))
2213 return replaceInstUsesWith(I, V);
2215 if (Value *V = SimplifyBSwap(I, Builder))
2216 return replaceInstUsesWith(I, V);
2218 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))
2221 if (Instruction *BSwap = matchBSwap(I))
2224 if (Instruction *Rotate = matchRotate(I))
2229 if (match(&I, m_c_Or(m_OneUse(m_Xor(m_Value(X), m_APInt(CV))), m_Value(Y))) &&
2230 !CV->isAllOnesValue() && MaskedValueIsZero(Y, *CV, 0, &I)) {
2231 // (X ^ C) | Y -> (X | Y) ^ C iff Y & C == 0
2232 // The check for a 'not' op is for efficiency (if Y is known zero --> ~X).
2233 Value *Or = Builder.CreateOr(X, Y);
2234 return BinaryOperator::CreateXor(Or, ConstantInt::get(I.getType(), *CV));
2238 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2239 Value *A, *B, *C, *D;
2240 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
2241 match(Op1, m_And(m_Value(B), m_Value(D)))) {
2242 ConstantInt *C1 = dyn_cast<ConstantInt>(C);
2243 ConstantInt *C2 = dyn_cast<ConstantInt>(D);
2244 if (C1 && C2) { // (A & C1)|(B & C2)
2245 Value *V1 = nullptr, *V2 = nullptr;
2246 if ((C1->getValue() & C2->getValue()).isNullValue()) {
2247 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
2248 // iff (C1&C2) == 0 and (N&~C1) == 0
2249 if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
2251 MaskedValueIsZero(V2, ~C1->getValue(), 0, &I)) || // (V|N)
2253 MaskedValueIsZero(V1, ~C1->getValue(), 0, &I)))) // (N|V)
2254 return BinaryOperator::CreateAnd(A,
2255 Builder.getInt(C1->getValue()|C2->getValue()));
2256 // Or commutes, try both ways.
2257 if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
2259 MaskedValueIsZero(V2, ~C2->getValue(), 0, &I)) || // (V|N)
2261 MaskedValueIsZero(V1, ~C2->getValue(), 0, &I)))) // (N|V)
2262 return BinaryOperator::CreateAnd(B,
2263 Builder.getInt(C1->getValue()|C2->getValue()));
2265 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
2266 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
2267 ConstantInt *C3 = nullptr, *C4 = nullptr;
2268 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
2269 (C3->getValue() & ~C1->getValue()).isNullValue() &&
2270 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
2271 (C4->getValue() & ~C2->getValue()).isNullValue()) {
2272 V2 = Builder.CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
2273 return BinaryOperator::CreateAnd(V2,
2274 Builder.getInt(C1->getValue()|C2->getValue()));
2278 if (C1->getValue() == ~C2->getValue()) {
2281 // ((X|B)&C1)|(B&C2) -> (X&C1) | B iff C1 == ~C2
2282 if (match(A, m_c_Or(m_Value(X), m_Specific(B))))
2283 return BinaryOperator::CreateOr(Builder.CreateAnd(X, C1), B);
2284 // (A&C2)|((X|A)&C1) -> (X&C2) | A iff C1 == ~C2
2285 if (match(B, m_c_Or(m_Specific(A), m_Value(X))))
2286 return BinaryOperator::CreateOr(Builder.CreateAnd(X, C2), A);
2288 // ((X^B)&C1)|(B&C2) -> (X&C1) ^ B iff C1 == ~C2
2289 if (match(A, m_c_Xor(m_Value(X), m_Specific(B))))
2290 return BinaryOperator::CreateXor(Builder.CreateAnd(X, C1), B);
2291 // (A&C2)|((X^A)&C1) -> (X&C2) ^ A iff C1 == ~C2
2292 if (match(B, m_c_Xor(m_Specific(A), m_Value(X))))
2293 return BinaryOperator::CreateXor(Builder.CreateAnd(X, C2), A);
2297 // Don't try to form a select if it's unlikely that we'll get rid of at
2298 // least one of the operands. A select is generally more expensive than the
2299 // 'or' that it is replacing.
2300 if (Op0->hasOneUse() || Op1->hasOneUse()) {
2301 // (Cond & C) | (~Cond & D) -> Cond ? C : D, and commuted variants.
2302 if (Value *V = matchSelectFromAndOr(A, C, B, D))
2303 return replaceInstUsesWith(I, V);
2304 if (Value *V = matchSelectFromAndOr(A, C, D, B))
2305 return replaceInstUsesWith(I, V);
2306 if (Value *V = matchSelectFromAndOr(C, A, B, D))
2307 return replaceInstUsesWith(I, V);
2308 if (Value *V = matchSelectFromAndOr(C, A, D, B))
2309 return replaceInstUsesWith(I, V);
2310 if (Value *V = matchSelectFromAndOr(B, D, A, C))
2311 return replaceInstUsesWith(I, V);
2312 if (Value *V = matchSelectFromAndOr(B, D, C, A))
2313 return replaceInstUsesWith(I, V);
2314 if (Value *V = matchSelectFromAndOr(D, B, A, C))
2315 return replaceInstUsesWith(I, V);
2316 if (Value *V = matchSelectFromAndOr(D, B, C, A))
2317 return replaceInstUsesWith(I, V);
2321 // (A ^ B) | ((B ^ C) ^ A) -> (A ^ B) | C
2322 if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
2323 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
2324 return BinaryOperator::CreateOr(Op0, C);
2326 // ((A ^ C) ^ B) | (B ^ A) -> (B ^ A) | C
2327 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
2328 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
2329 return BinaryOperator::CreateOr(Op1, C);
2331 // ((B | C) & A) | B -> B | (A & C)
2332 if (match(Op0, m_And(m_Or(m_Specific(Op1), m_Value(C)), m_Value(A))))
2333 return BinaryOperator::CreateOr(Op1, Builder.CreateAnd(A, C));
2335 if (Instruction *DeMorgan = matchDeMorgansLaws(I, Builder))
2338 // Canonicalize xor to the RHS.
2339 bool SwappedForXor = false;
2340 if (match(Op0, m_Xor(m_Value(), m_Value()))) {
2341 std::swap(Op0, Op1);
2342 SwappedForXor = true;
2345 // A | ( A ^ B) -> A | B
2346 // A | (~A ^ B) -> A | ~B
2347 // (A & B) | (A ^ B)
2348 if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
2349 if (Op0 == A || Op0 == B)
2350 return BinaryOperator::CreateOr(A, B);
2352 if (match(Op0, m_And(m_Specific(A), m_Specific(B))) ||
2353 match(Op0, m_And(m_Specific(B), m_Specific(A))))
2354 return BinaryOperator::CreateOr(A, B);
2356 if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
2357 Value *Not = Builder.CreateNot(B, B->getName() + ".not");
2358 return BinaryOperator::CreateOr(Not, Op0);
2360 if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
2361 Value *Not = Builder.CreateNot(A, A->getName() + ".not");
2362 return BinaryOperator::CreateOr(Not, Op0);
2366 // A | ~(A | B) -> A | ~B
2367 // A | ~(A ^ B) -> A | ~B
2368 if (match(Op1, m_Not(m_Value(A))))
2369 if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
2370 if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
2371 Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
2372 B->getOpcode() == Instruction::Xor)) {
2373 Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
2375 Value *Not = Builder.CreateNot(NotOp, NotOp->getName() + ".not");
2376 return BinaryOperator::CreateOr(Not, Op0);
2380 std::swap(Op0, Op1);
2383 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
2384 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
2386 if (Value *Res = foldOrOfICmps(LHS, RHS, I))
2387 return replaceInstUsesWith(I, Res);
2389 // TODO: Make this recursive; it's a little tricky because an arbitrary
2390 // number of 'or' instructions might have to be created.
2392 if (LHS && match(Op1, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2393 if (auto *Cmp = dyn_cast<ICmpInst>(X))
2394 if (Value *Res = foldOrOfICmps(LHS, Cmp, I))
2395 return replaceInstUsesWith(I, Builder.CreateOr(Res, Y));
2396 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2397 if (Value *Res = foldOrOfICmps(LHS, Cmp, I))
2398 return replaceInstUsesWith(I, Builder.CreateOr(Res, X));
2400 if (RHS && match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2401 if (auto *Cmp = dyn_cast<ICmpInst>(X))
2402 if (Value *Res = foldOrOfICmps(Cmp, RHS, I))
2403 return replaceInstUsesWith(I, Builder.CreateOr(Res, Y));
2404 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2405 if (Value *Res = foldOrOfICmps(Cmp, RHS, I))
2406 return replaceInstUsesWith(I, Builder.CreateOr(Res, X));
2410 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
2411 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
2412 if (Value *Res = foldLogicOfFCmps(LHS, RHS, false))
2413 return replaceInstUsesWith(I, Res);
2415 if (Instruction *CastedOr = foldCastedBitwiseLogic(I))
2418 // or(sext(A), B) / or(B, sext(A)) --> A ? -1 : B, where A is i1 or <N x i1>.
2419 if (match(Op0, m_OneUse(m_SExt(m_Value(A)))) &&
2420 A->getType()->isIntOrIntVectorTy(1))
2421 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1);
2422 if (match(Op1, m_OneUse(m_SExt(m_Value(A)))) &&
2423 A->getType()->isIntOrIntVectorTy(1))
2424 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0);
2426 // Note: If we've gotten to the point of visiting the outer OR, then the
2427 // inner one couldn't be simplified. If it was a constant, then it won't
2428 // be simplified by a later pass either, so we try swapping the inner/outer
2429 // ORs in the hopes that we'll be able to simplify it this way.
2430 // (X|C) | V --> (X|V) | C
2432 if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
2433 match(Op0, m_Or(m_Value(A), m_ConstantInt(CI)))) {
2434 Value *Inner = Builder.CreateOr(A, Op1);
2435 Inner->takeName(Op0);
2436 return BinaryOperator::CreateOr(Inner, CI);
2439 // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D))
2440 // Since this OR statement hasn't been optimized further yet, we hope
2441 // that this transformation will allow the new ORs to be optimized.
2443 Value *X = nullptr, *Y = nullptr;
2444 if (Op0->hasOneUse() && Op1->hasOneUse() &&
2445 match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) &&
2446 match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) {
2447 Value *orTrue = Builder.CreateOr(A, C);
2448 Value *orFalse = Builder.CreateOr(B, D);
2449 return SelectInst::Create(X, orTrue, orFalse);
2456 /// A ^ B can be specified using other logic ops in a variety of patterns. We
2457 /// can fold these early and efficiently by morphing an existing instruction.
2458 static Instruction *foldXorToXor(BinaryOperator &I,
2459 InstCombiner::BuilderTy &Builder) {
2460 assert(I.getOpcode() == Instruction::Xor);
2461 Value *Op0 = I.getOperand(0);
2462 Value *Op1 = I.getOperand(1);
2465 // There are 4 commuted variants for each of the basic patterns.
2467 // (A & B) ^ (A | B) -> A ^ B
2468 // (A & B) ^ (B | A) -> A ^ B
2469 // (A | B) ^ (A & B) -> A ^ B
2470 // (A | B) ^ (B & A) -> A ^ B
2471 if (match(&I, m_c_Xor(m_And(m_Value(A), m_Value(B)),
2472 m_c_Or(m_Deferred(A), m_Deferred(B))))) {
2478 // (A | ~B) ^ (~A | B) -> A ^ B
2479 // (~B | A) ^ (~A | B) -> A ^ B
2480 // (~A | B) ^ (A | ~B) -> A ^ B
2481 // (B | ~A) ^ (A | ~B) -> A ^ B
2482 if (match(&I, m_Xor(m_c_Or(m_Value(A), m_Not(m_Value(B))),
2483 m_c_Or(m_Not(m_Deferred(A)), m_Deferred(B))))) {
2489 // (A & ~B) ^ (~A & B) -> A ^ B
2490 // (~B & A) ^ (~A & B) -> A ^ B
2491 // (~A & B) ^ (A & ~B) -> A ^ B
2492 // (B & ~A) ^ (A & ~B) -> A ^ B
2493 if (match(&I, m_Xor(m_c_And(m_Value(A), m_Not(m_Value(B))),
2494 m_c_And(m_Not(m_Deferred(A)), m_Deferred(B))))) {
2500 // For the remaining cases we need to get rid of one of the operands.
2501 if (!Op0->hasOneUse() && !Op1->hasOneUse())
2504 // (A | B) ^ ~(A & B) -> ~(A ^ B)
2505 // (A | B) ^ ~(B & A) -> ~(A ^ B)
2506 // (A & B) ^ ~(A | B) -> ~(A ^ B)
2507 // (A & B) ^ ~(B | A) -> ~(A ^ B)
2508 // Complexity sorting ensures the not will be on the right side.
2509 if ((match(Op0, m_Or(m_Value(A), m_Value(B))) &&
2510 match(Op1, m_Not(m_c_And(m_Specific(A), m_Specific(B))))) ||
2511 (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2512 match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B))))))
2513 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
2518 Value *InstCombiner::foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
2519 if (predicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
2520 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2521 LHS->getOperand(1) == RHS->getOperand(0))
2522 LHS->swapOperands();
2523 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2524 LHS->getOperand(1) == RHS->getOperand(1)) {
2525 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
2526 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2527 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
2528 bool IsSigned = LHS->isSigned() || RHS->isSigned();
2529 return getNewICmpValue(Code, IsSigned, Op0, Op1, Builder);
2533 // TODO: This can be generalized to compares of non-signbits using
2534 // decomposeBitTestICmp(). It could be enhanced more by using (something like)
2535 // foldLogOpOfMaskedICmps().
2536 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
2537 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
2538 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
2539 if ((LHS->hasOneUse() || RHS->hasOneUse()) &&
2540 LHS0->getType() == RHS0->getType() &&
2541 LHS0->getType()->isIntOrIntVectorTy()) {
2542 // (X > -1) ^ (Y > -1) --> (X ^ Y) < 0
2543 // (X < 0) ^ (Y < 0) --> (X ^ Y) < 0
2544 if ((PredL == CmpInst::ICMP_SGT && match(LHS1, m_AllOnes()) &&
2545 PredR == CmpInst::ICMP_SGT && match(RHS1, m_AllOnes())) ||
2546 (PredL == CmpInst::ICMP_SLT && match(LHS1, m_Zero()) &&
2547 PredR == CmpInst::ICMP_SLT && match(RHS1, m_Zero()))) {
2548 Value *Zero = ConstantInt::getNullValue(LHS0->getType());
2549 return Builder.CreateICmpSLT(Builder.CreateXor(LHS0, RHS0), Zero);
2551 // (X > -1) ^ (Y < 0) --> (X ^ Y) > -1
2552 // (X < 0) ^ (Y > -1) --> (X ^ Y) > -1
2553 if ((PredL == CmpInst::ICMP_SGT && match(LHS1, m_AllOnes()) &&
2554 PredR == CmpInst::ICMP_SLT && match(RHS1, m_Zero())) ||
2555 (PredL == CmpInst::ICMP_SLT && match(LHS1, m_Zero()) &&
2556 PredR == CmpInst::ICMP_SGT && match(RHS1, m_AllOnes()))) {
2557 Value *MinusOne = ConstantInt::getAllOnesValue(LHS0->getType());
2558 return Builder.CreateICmpSGT(Builder.CreateXor(LHS0, RHS0), MinusOne);
2562 // Instead of trying to imitate the folds for and/or, decompose this 'xor'
2563 // into those logic ops. That is, try to turn this into an and-of-icmps
2564 // because we have many folds for that pattern.
2566 // This is based on a truth table definition of xor:
2567 // X ^ Y --> (X | Y) & !(X & Y)
2568 if (Value *OrICmp = SimplifyBinOp(Instruction::Or, LHS, RHS, SQ)) {
2569 // TODO: If OrICmp is true, then the definition of xor simplifies to !(X&Y).
2570 // TODO: If OrICmp is false, the whole thing is false (InstSimplify?).
2571 if (Value *AndICmp = SimplifyBinOp(Instruction::And, LHS, RHS, SQ)) {
2572 // TODO: Independently handle cases where the 'and' side is a constant.
2573 if (OrICmp == LHS && AndICmp == RHS && RHS->hasOneUse()) {
2574 // (LHS | RHS) & !(LHS & RHS) --> LHS & !RHS
2575 RHS->setPredicate(RHS->getInversePredicate());
2576 return Builder.CreateAnd(LHS, RHS);
2578 if (OrICmp == RHS && AndICmp == LHS && LHS->hasOneUse()) {
2579 // !(LHS & RHS) & (LHS | RHS) --> !LHS & RHS
2580 LHS->setPredicate(LHS->getInversePredicate());
2581 return Builder.CreateAnd(LHS, RHS);
2589 /// If we have a masked merge, in the canonical form of:
2590 /// (assuming that A only has one use.)
2592 /// ((x ^ y) & M) ^ y
2594 /// * If M is inverted:
2596 /// ((x ^ y) & ~M) ^ y
2597 /// We can canonicalize by swapping the final xor operand
2598 /// to eliminate the 'not' of the mask.
2599 /// ((x ^ y) & M) ^ x
2600 /// * If M is a constant, and D has one use, we transform to 'and' / 'or' ops
2601 /// because that shortens the dependency chain and improves analysis:
2602 /// (x & M) | (y & ~M)
2603 static Instruction *visitMaskedMerge(BinaryOperator &I,
2604 InstCombiner::BuilderTy &Builder) {
2607 if (!match(&I, m_c_Xor(m_Value(B),
2609 m_CombineAnd(m_c_Xor(m_Deferred(B), m_Value(X)),
2615 if (match(M, m_Not(m_Value(NotM)))) {
2616 // De-invert the mask and swap the value in B part.
2617 Value *NewA = Builder.CreateAnd(D, NotM);
2618 return BinaryOperator::CreateXor(NewA, X);
2622 if (D->hasOneUse() && match(M, m_Constant(C))) {
2624 Value *LHS = Builder.CreateAnd(X, C);
2625 Value *NotC = Builder.CreateNot(C);
2626 Value *RHS = Builder.CreateAnd(B, NotC);
2627 return BinaryOperator::CreateOr(LHS, RHS);
2639 static Instruction *sinkNotIntoXor(BinaryOperator &I,
2640 InstCombiner::BuilderTy &Builder) {
2642 // FIXME: one-use check is not needed in general, but currently we are unable
2643 // to fold 'not' into 'icmp', if that 'icmp' has multiple uses. (D35182)
2644 if (!match(&I, m_Not(m_OneUse(m_Xor(m_Value(X), m_Value(Y))))))
2647 // We only want to do the transform if it is free to do.
2648 if (IsFreeToInvert(X, X->hasOneUse())) {
2650 } else if (IsFreeToInvert(Y, Y->hasOneUse())) {
2655 Value *NotX = Builder.CreateNot(X, X->getName() + ".not");
2656 return BinaryOperator::CreateXor(NotX, Y, I.getName() + ".demorgan");
2659 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
2660 // here. We should standardize that construct where it is needed or choose some
2661 // other way to ensure that commutated variants of patterns are not missed.
2662 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
2663 if (Value *V = SimplifyXorInst(I.getOperand(0), I.getOperand(1),
2664 SQ.getWithInstruction(&I)))
2665 return replaceInstUsesWith(I, V);
2667 if (SimplifyAssociativeOrCommutative(I))
2670 if (Instruction *X = foldVectorBinop(I))
2673 if (Instruction *NewXor = foldXorToXor(I, Builder))
2676 // (A&B)^(A&C) -> A&(B^C) etc
2677 if (Value *V = SimplifyUsingDistributiveLaws(I))
2678 return replaceInstUsesWith(I, V);
2680 // See if we can simplify any instructions used by the instruction whose sole
2681 // purpose is to compute bits we don't care about.
2682 if (SimplifyDemandedInstructionBits(I))
2685 if (Value *V = SimplifyBSwap(I, Builder))
2686 return replaceInstUsesWith(I, V);
2688 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2690 // Fold (X & M) ^ (Y & ~M) -> (X & M) | (Y & ~M)
2691 // This it a special case in haveNoCommonBitsSet, but the computeKnownBits
2692 // calls in there are unnecessary as SimplifyDemandedInstructionBits should
2693 // have already taken care of those cases.
2695 if (match(&I, m_c_Xor(m_c_And(m_Not(m_Value(M)), m_Value()),
2696 m_c_And(m_Deferred(M), m_Value()))))
2697 return BinaryOperator::CreateOr(Op0, Op1);
2699 // Apply DeMorgan's Law for 'nand' / 'nor' logic with an inverted operand.
2702 // We must eliminate the and/or (one-use) for these transforms to not increase
2703 // the instruction count.
2704 // ~(~X & Y) --> (X | ~Y)
2705 // ~(Y & ~X) --> (X | ~Y)
2706 if (match(&I, m_Not(m_OneUse(m_c_And(m_Not(m_Value(X)), m_Value(Y)))))) {
2707 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");
2708 return BinaryOperator::CreateOr(X, NotY);
2710 // ~(~X | Y) --> (X & ~Y)
2711 // ~(Y | ~X) --> (X & ~Y)
2712 if (match(&I, m_Not(m_OneUse(m_c_Or(m_Not(m_Value(X)), m_Value(Y)))))) {
2713 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");
2714 return BinaryOperator::CreateAnd(X, NotY);
2717 if (Instruction *Xor = visitMaskedMerge(I, Builder))
2720 // Is this a 'not' (~) fed by a binary operator?
2721 BinaryOperator *NotVal;
2722 if (match(&I, m_Not(m_BinOp(NotVal)))) {
2723 if (NotVal->getOpcode() == Instruction::And ||
2724 NotVal->getOpcode() == Instruction::Or) {
2725 // Apply DeMorgan's Law when inverts are free:
2726 // ~(X & Y) --> (~X | ~Y)
2727 // ~(X | Y) --> (~X & ~Y)
2728 if (IsFreeToInvert(NotVal->getOperand(0),
2729 NotVal->getOperand(0)->hasOneUse()) &&
2730 IsFreeToInvert(NotVal->getOperand(1),
2731 NotVal->getOperand(1)->hasOneUse())) {
2732 Value *NotX = Builder.CreateNot(NotVal->getOperand(0), "notlhs");
2733 Value *NotY = Builder.CreateNot(NotVal->getOperand(1), "notrhs");
2734 if (NotVal->getOpcode() == Instruction::And)
2735 return BinaryOperator::CreateOr(NotX, NotY);
2736 return BinaryOperator::CreateAnd(NotX, NotY);
2740 // ~(X - Y) --> ~X + Y
2741 if (match(NotVal, m_Sub(m_Value(X), m_Value(Y))))
2742 if (isa<Constant>(X) || NotVal->hasOneUse())
2743 return BinaryOperator::CreateAdd(Builder.CreateNot(X), Y);
2745 // ~(~X >>s Y) --> (X >>s Y)
2746 if (match(NotVal, m_AShr(m_Not(m_Value(X)), m_Value(Y))))
2747 return BinaryOperator::CreateAShr(X, Y);
2749 // If we are inverting a right-shifted constant, we may be able to eliminate
2750 // the 'not' by inverting the constant and using the opposite shift type.
2751 // Canonicalization rules ensure that only a negative constant uses 'ashr',
2752 // but we must check that in case that transform has not fired yet.
2754 // ~(C >>s Y) --> ~C >>u Y (when inverting the replicated sign bits)
2756 if (match(NotVal, m_AShr(m_Constant(C), m_Value(Y))) &&
2757 match(C, m_Negative()))
2758 return BinaryOperator::CreateLShr(ConstantExpr::getNot(C), Y);
2760 // ~(C >>u Y) --> ~C >>s Y (when inverting the replicated sign bits)
2761 if (match(NotVal, m_LShr(m_Constant(C), m_Value(Y))) &&
2762 match(C, m_NonNegative()))
2763 return BinaryOperator::CreateAShr(ConstantExpr::getNot(C), Y);
2765 // ~(X + C) --> -(C + 1) - X
2766 if (match(Op0, m_Add(m_Value(X), m_Constant(C))))
2767 return BinaryOperator::CreateSub(ConstantExpr::getNeg(AddOne(C)), X);
2770 // Use DeMorgan and reassociation to eliminate a 'not' op.
2772 if (match(Op1, m_Constant(C1))) {
2774 if (match(Op0, m_OneUse(m_Or(m_Not(m_Value(X)), m_Constant(C2))))) {
2775 // (~X | C2) ^ C1 --> ((X & ~C2) ^ -1) ^ C1 --> (X & ~C2) ^ ~C1
2776 Value *And = Builder.CreateAnd(X, ConstantExpr::getNot(C2));
2777 return BinaryOperator::CreateXor(And, ConstantExpr::getNot(C1));
2779 if (match(Op0, m_OneUse(m_And(m_Not(m_Value(X)), m_Constant(C2))))) {
2780 // (~X & C2) ^ C1 --> ((X | ~C2) ^ -1) ^ C1 --> (X | ~C2) ^ ~C1
2781 Value *Or = Builder.CreateOr(X, ConstantExpr::getNot(C2));
2782 return BinaryOperator::CreateXor(Or, ConstantExpr::getNot(C1));
2786 // not (cmp A, B) = !cmp A, B
2787 CmpInst::Predicate Pred;
2788 if (match(&I, m_Not(m_OneUse(m_Cmp(Pred, m_Value(), m_Value()))))) {
2789 cast<CmpInst>(Op0)->setPredicate(CmpInst::getInversePredicate(Pred));
2790 return replaceInstUsesWith(I, Op0);
2795 if (match(Op1, m_APInt(RHSC))) {
2798 if (RHSC->isSignMask() && match(Op0, m_Sub(m_APInt(C), m_Value(X)))) {
2799 // (C - X) ^ signmask -> (C + signmask - X)
2800 Constant *NewC = ConstantInt::get(I.getType(), *C + *RHSC);
2801 return BinaryOperator::CreateSub(NewC, X);
2803 if (RHSC->isSignMask() && match(Op0, m_Add(m_Value(X), m_APInt(C)))) {
2804 // (X + C) ^ signmask -> (X + C + signmask)
2805 Constant *NewC = ConstantInt::get(I.getType(), *C + *RHSC);
2806 return BinaryOperator::CreateAdd(X, NewC);
2809 // (X|C1)^C2 -> X^(C1^C2) iff X&~C1 == 0
2810 if (match(Op0, m_Or(m_Value(X), m_APInt(C))) &&
2811 MaskedValueIsZero(X, *C, 0, &I)) {
2812 Constant *NewC = ConstantInt::get(I.getType(), *C ^ *RHSC);
2813 Worklist.Add(cast<Instruction>(Op0));
2815 I.setOperand(1, NewC);
2821 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1)) {
2822 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2823 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2824 if (Op0I->getOpcode() == Instruction::LShr) {
2825 // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3)
2829 if (Op0I->hasOneUse() &&
2830 (E1 = dyn_cast<BinaryOperator>(Op0I->getOperand(0))) &&
2831 E1->getOpcode() == Instruction::Xor &&
2832 (C1 = dyn_cast<ConstantInt>(E1->getOperand(1)))) {
2833 // fold (C1 >> C2) ^ C3
2834 ConstantInt *C2 = Op0CI, *C3 = RHSC;
2835 APInt FoldConst = C1->getValue().lshr(C2->getValue());
2836 FoldConst ^= C3->getValue();
2837 // Prepare the two operands.
2838 Value *Opnd0 = Builder.CreateLShr(E1->getOperand(0), C2);
2839 Opnd0->takeName(Op0I);
2840 cast<Instruction>(Opnd0)->setDebugLoc(I.getDebugLoc());
2841 Value *FoldVal = ConstantInt::get(Opnd0->getType(), FoldConst);
2843 return BinaryOperator::CreateXor(Opnd0, FoldVal);
2850 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))
2853 // Y ^ (X | Y) --> X & ~Y
2854 // Y ^ (Y | X) --> X & ~Y
2855 if (match(Op1, m_OneUse(m_c_Or(m_Value(X), m_Specific(Op0)))))
2856 return BinaryOperator::CreateAnd(X, Builder.CreateNot(Op0));
2857 // (X | Y) ^ Y --> X & ~Y
2858 // (Y | X) ^ Y --> X & ~Y
2859 if (match(Op0, m_OneUse(m_c_Or(m_Value(X), m_Specific(Op1)))))
2860 return BinaryOperator::CreateAnd(X, Builder.CreateNot(Op1));
2862 // Y ^ (X & Y) --> ~X & Y
2863 // Y ^ (Y & X) --> ~X & Y
2864 if (match(Op1, m_OneUse(m_c_And(m_Value(X), m_Specific(Op0)))))
2865 return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(X));
2866 // (X & Y) ^ Y --> ~X & Y
2867 // (Y & X) ^ Y --> ~X & Y
2868 // Canonical form is (X & C) ^ C; don't touch that.
2869 // TODO: A 'not' op is better for analysis and codegen, but demanded bits must
2870 // be fixed to prefer that (otherwise we get infinite looping).
2871 if (!match(Op1, m_Constant()) &&
2872 match(Op0, m_OneUse(m_c_And(m_Value(X), m_Specific(Op1)))))
2873 return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(X));
2876 // (A ^ B) ^ (A | C) --> (~A & C) ^ B -- There are 4 commuted variants.
2877 if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_Value(A), m_Value(B))),
2878 m_OneUse(m_c_Or(m_Deferred(A), m_Value(C))))))
2879 return BinaryOperator::CreateXor(
2880 Builder.CreateAnd(Builder.CreateNot(A), C), B);
2882 // (A ^ B) ^ (B | C) --> (~B & C) ^ A -- There are 4 commuted variants.
2883 if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_Value(A), m_Value(B))),
2884 m_OneUse(m_c_Or(m_Deferred(B), m_Value(C))))))
2885 return BinaryOperator::CreateXor(
2886 Builder.CreateAnd(Builder.CreateNot(B), C), A);
2888 // (A & B) ^ (A ^ B) -> (A | B)
2889 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2890 match(Op1, m_c_Xor(m_Specific(A), m_Specific(B))))
2891 return BinaryOperator::CreateOr(A, B);
2892 // (A ^ B) ^ (A & B) -> (A | B)
2893 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
2894 match(Op1, m_c_And(m_Specific(A), m_Specific(B))))
2895 return BinaryOperator::CreateOr(A, B);
2897 // (A & ~B) ^ ~A -> ~(A & B)
2898 // (~B & A) ^ ~A -> ~(A & B)
2899 if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) &&
2900 match(Op1, m_Not(m_Specific(A))))
2901 return BinaryOperator::CreateNot(Builder.CreateAnd(A, B));
2903 if (auto *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
2904 if (auto *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
2905 if (Value *V = foldXorOfICmps(LHS, RHS))
2906 return replaceInstUsesWith(I, V);
2908 if (Instruction *CastedXor = foldCastedBitwiseLogic(I))
2911 // Canonicalize a shifty way to code absolute value to the common pattern.
2912 // There are 4 potential commuted variants. Move the 'ashr' candidate to Op1.
2913 // We're relying on the fact that we only do this transform when the shift has
2914 // exactly 2 uses and the add has exactly 1 use (otherwise, we might increase
2916 if (Op0->hasNUses(2))
2917 std::swap(Op0, Op1);
2920 Type *Ty = I.getType();
2921 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
2922 Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 &&
2923 match(Op0, m_OneUse(m_c_Add(m_Specific(A), m_Specific(Op1))))) {
2924 // B = ashr i32 A, 31 ; smear the sign bit
2925 // xor (add A, B), B ; add -1 and flip bits if negative
2926 // --> (A < 0) ? -A : A
2927 Value *Cmp = Builder.CreateICmpSLT(A, ConstantInt::getNullValue(Ty));
2928 // Copy the nuw/nsw flags from the add to the negate.
2929 auto *Add = cast<BinaryOperator>(Op0);
2930 Value *Neg = Builder.CreateNeg(A, "", Add->hasNoUnsignedWrap(),
2931 Add->hasNoSignedWrap());
2932 return SelectInst::Create(Cmp, Neg, A);
2935 // Eliminate a bitwise 'not' op of 'not' min/max by inverting the min/max:
2937 // %notx = xor i32 %x, -1
2938 // %cmp1 = icmp sgt i32 %notx, %y
2939 // %smax = select i1 %cmp1, i32 %notx, i32 %y
2940 // %res = xor i32 %smax, -1
2942 // %noty = xor i32 %y, -1
2943 // %cmp2 = icmp slt %x, %noty
2944 // %res = select i1 %cmp2, i32 %x, i32 %noty
2946 // Same is applicable for smin/umax/umin.
2947 if (match(Op1, m_AllOnes()) && Op0->hasOneUse()) {
2949 SelectPatternFlavor SPF = matchSelectPattern(Op0, LHS, RHS).Flavor;
2950 if (SelectPatternResult::isMinOrMax(SPF)) {
2951 // It's possible we get here before the not has been simplified, so make
2952 // sure the input to the not isn't freely invertible.
2953 if (match(LHS, m_Not(m_Value(X))) && !IsFreeToInvert(X, X->hasOneUse())) {
2954 Value *NotY = Builder.CreateNot(RHS);
2955 return SelectInst::Create(
2956 Builder.CreateICmp(getInverseMinMaxPred(SPF), X, NotY), X, NotY);
2959 // It's possible we get here before the not has been simplified, so make
2960 // sure the input to the not isn't freely invertible.
2961 if (match(RHS, m_Not(m_Value(Y))) && !IsFreeToInvert(Y, Y->hasOneUse())) {
2962 Value *NotX = Builder.CreateNot(LHS);
2963 return SelectInst::Create(
2964 Builder.CreateICmp(getInverseMinMaxPred(SPF), NotX, Y), NotX, Y);
2967 // If both sides are freely invertible, then we can get rid of the xor
2969 if (IsFreeToInvert(LHS, !LHS->hasNUsesOrMore(3)) &&
2970 IsFreeToInvert(RHS, !RHS->hasNUsesOrMore(3))) {
2971 Value *NotLHS = Builder.CreateNot(LHS);
2972 Value *NotRHS = Builder.CreateNot(RHS);
2973 return SelectInst::Create(
2974 Builder.CreateICmp(getInverseMinMaxPred(SPF), NotLHS, NotRHS),
2980 if (Instruction *NewXor = sinkNotIntoXor(I, Builder))