[ASan] Make insertion of version mismatch guard configurable
[llvm-core.git] / lib / Transforms / InstCombine / InstCombineAndOrXor.cpp
blobb9e8df8c29acf31df2fd627a3781536ba66fe0aa
1 //===- InstCombineAndOrXor.cpp --------------------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the visitAnd, visitOr, and visitXor functions.
11 //===----------------------------------------------------------------------===//
13 #include "InstCombineInternal.h"
14 #include "llvm/Analysis/CmpInstAnalysis.h"
15 #include "llvm/Analysis/InstructionSimplify.h"
16 #include "llvm/Transforms/Utils/Local.h"
17 #include "llvm/IR/ConstantRange.h"
18 #include "llvm/IR/Intrinsics.h"
19 #include "llvm/IR/PatternMatch.h"
20 using namespace llvm;
21 using namespace PatternMatch;
23 #define DEBUG_TYPE "instcombine"
25 /// Similar to getICmpCode but for FCmpInst. This encodes a fcmp predicate into
26 /// a four bit mask.
27 static unsigned getFCmpCode(FCmpInst::Predicate CC) {
28 assert(FCmpInst::FCMP_FALSE <= CC && CC <= FCmpInst::FCMP_TRUE &&
29 "Unexpected FCmp predicate!");
30 // Take advantage of the bit pattern of FCmpInst::Predicate here.
31 // U L G E
32 static_assert(FCmpInst::FCMP_FALSE == 0, ""); // 0 0 0 0
33 static_assert(FCmpInst::FCMP_OEQ == 1, ""); // 0 0 0 1
34 static_assert(FCmpInst::FCMP_OGT == 2, ""); // 0 0 1 0
35 static_assert(FCmpInst::FCMP_OGE == 3, ""); // 0 0 1 1
36 static_assert(FCmpInst::FCMP_OLT == 4, ""); // 0 1 0 0
37 static_assert(FCmpInst::FCMP_OLE == 5, ""); // 0 1 0 1
38 static_assert(FCmpInst::FCMP_ONE == 6, ""); // 0 1 1 0
39 static_assert(FCmpInst::FCMP_ORD == 7, ""); // 0 1 1 1
40 static_assert(FCmpInst::FCMP_UNO == 8, ""); // 1 0 0 0
41 static_assert(FCmpInst::FCMP_UEQ == 9, ""); // 1 0 0 1
42 static_assert(FCmpInst::FCMP_UGT == 10, ""); // 1 0 1 0
43 static_assert(FCmpInst::FCMP_UGE == 11, ""); // 1 0 1 1
44 static_assert(FCmpInst::FCMP_ULT == 12, ""); // 1 1 0 0
45 static_assert(FCmpInst::FCMP_ULE == 13, ""); // 1 1 0 1
46 static_assert(FCmpInst::FCMP_UNE == 14, ""); // 1 1 1 0
47 static_assert(FCmpInst::FCMP_TRUE == 15, ""); // 1 1 1 1
48 return CC;
51 /// This is the complement of getICmpCode, which turns an opcode and two
52 /// operands into either a constant true or false, or a brand new ICmp
53 /// instruction. The sign is passed in to determine which kind of predicate to
54 /// use in the new icmp instruction.
55 static Value *getNewICmpValue(unsigned Code, bool Sign, Value *LHS, Value *RHS,
56 InstCombiner::BuilderTy &Builder) {
57 ICmpInst::Predicate NewPred;
58 if (Constant *TorF = getPredForICmpCode(Code, Sign, LHS->getType(), NewPred))
59 return TorF;
60 return Builder.CreateICmp(NewPred, LHS, RHS);
63 /// This is the complement of getFCmpCode, which turns an opcode and two
64 /// operands into either a FCmp instruction, or a true/false constant.
65 static Value *getFCmpValue(unsigned Code, Value *LHS, Value *RHS,
66 InstCombiner::BuilderTy &Builder) {
67 const auto Pred = static_cast<FCmpInst::Predicate>(Code);
68 assert(FCmpInst::FCMP_FALSE <= Pred && Pred <= FCmpInst::FCMP_TRUE &&
69 "Unexpected FCmp predicate!");
70 if (Pred == FCmpInst::FCMP_FALSE)
71 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
72 if (Pred == FCmpInst::FCMP_TRUE)
73 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
74 return Builder.CreateFCmp(Pred, LHS, RHS);
77 /// Transform BITWISE_OP(BSWAP(A),BSWAP(B)) or
78 /// BITWISE_OP(BSWAP(A), Constant) to BSWAP(BITWISE_OP(A, B))
79 /// \param I Binary operator to transform.
80 /// \return Pointer to node that must replace the original binary operator, or
81 /// null pointer if no transformation was made.
82 static Value *SimplifyBSwap(BinaryOperator &I,
83 InstCombiner::BuilderTy &Builder) {
84 assert(I.isBitwiseLogicOp() && "Unexpected opcode for bswap simplifying");
86 Value *OldLHS = I.getOperand(0);
87 Value *OldRHS = I.getOperand(1);
89 Value *NewLHS;
90 if (!match(OldLHS, m_BSwap(m_Value(NewLHS))))
91 return nullptr;
93 Value *NewRHS;
94 const APInt *C;
96 if (match(OldRHS, m_BSwap(m_Value(NewRHS)))) {
97 // OP( BSWAP(x), BSWAP(y) ) -> BSWAP( OP(x, y) )
98 if (!OldLHS->hasOneUse() && !OldRHS->hasOneUse())
99 return nullptr;
100 // NewRHS initialized by the matcher.
101 } else if (match(OldRHS, m_APInt(C))) {
102 // OP( BSWAP(x), CONSTANT ) -> BSWAP( OP(x, BSWAP(CONSTANT) ) )
103 if (!OldLHS->hasOneUse())
104 return nullptr;
105 NewRHS = ConstantInt::get(I.getType(), C->byteSwap());
106 } else
107 return nullptr;
109 Value *BinOp = Builder.CreateBinOp(I.getOpcode(), NewLHS, NewRHS);
110 Function *F = Intrinsic::getDeclaration(I.getModule(), Intrinsic::bswap,
111 I.getType());
112 return Builder.CreateCall(F, BinOp);
115 /// This handles expressions of the form ((val OP C1) & C2). Where
116 /// the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'.
117 Instruction *InstCombiner::OptAndOp(BinaryOperator *Op,
118 ConstantInt *OpRHS,
119 ConstantInt *AndRHS,
120 BinaryOperator &TheAnd) {
121 Value *X = Op->getOperand(0);
123 switch (Op->getOpcode()) {
124 default: break;
125 case Instruction::Add:
126 if (Op->hasOneUse()) {
127 // Adding a one to a single bit bit-field should be turned into an XOR
128 // of the bit. First thing to check is to see if this AND is with a
129 // single bit constant.
130 const APInt &AndRHSV = AndRHS->getValue();
132 // If there is only one bit set.
133 if (AndRHSV.isPowerOf2()) {
134 // Ok, at this point, we know that we are masking the result of the
135 // ADD down to exactly one bit. If the constant we are adding has
136 // no bits set below this bit, then we can eliminate the ADD.
137 const APInt& AddRHS = OpRHS->getValue();
139 // Check to see if any bits below the one bit set in AndRHSV are set.
140 if ((AddRHS & (AndRHSV - 1)).isNullValue()) {
141 // If not, the only thing that can effect the output of the AND is
142 // the bit specified by AndRHSV. If that bit is set, the effect of
143 // the XOR is to toggle the bit. If it is clear, then the ADD has
144 // no effect.
145 if ((AddRHS & AndRHSV).isNullValue()) { // Bit is not set, noop
146 TheAnd.setOperand(0, X);
147 return &TheAnd;
148 } else {
149 // Pull the XOR out of the AND.
150 Value *NewAnd = Builder.CreateAnd(X, AndRHS);
151 NewAnd->takeName(Op);
152 return BinaryOperator::CreateXor(NewAnd, AndRHS);
157 break;
159 return nullptr;
162 /// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise
163 /// (V < Lo || V >= Hi). This method expects that Lo < Hi. IsSigned indicates
164 /// whether to treat V, Lo, and Hi as signed or not.
165 Value *InstCombiner::insertRangeTest(Value *V, const APInt &Lo, const APInt &Hi,
166 bool isSigned, bool Inside) {
167 assert((isSigned ? Lo.slt(Hi) : Lo.ult(Hi)) &&
168 "Lo is not < Hi in range emission code!");
170 Type *Ty = V->getType();
172 // V >= Min && V < Hi --> V < Hi
173 // V < Min || V >= Hi --> V >= Hi
174 ICmpInst::Predicate Pred = Inside ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE;
175 if (isSigned ? Lo.isMinSignedValue() : Lo.isMinValue()) {
176 Pred = isSigned ? ICmpInst::getSignedPredicate(Pred) : Pred;
177 return Builder.CreateICmp(Pred, V, ConstantInt::get(Ty, Hi));
180 // V >= Lo && V < Hi --> V - Lo u< Hi - Lo
181 // V < Lo || V >= Hi --> V - Lo u>= Hi - Lo
182 Value *VMinusLo =
183 Builder.CreateSub(V, ConstantInt::get(Ty, Lo), V->getName() + ".off");
184 Constant *HiMinusLo = ConstantInt::get(Ty, Hi - Lo);
185 return Builder.CreateICmp(Pred, VMinusLo, HiMinusLo);
188 /// Classify (icmp eq (A & B), C) and (icmp ne (A & B), C) as matching patterns
189 /// that can be simplified.
190 /// One of A and B is considered the mask. The other is the value. This is
191 /// described as the "AMask" or "BMask" part of the enum. If the enum contains
192 /// only "Mask", then both A and B can be considered masks. If A is the mask,
193 /// then it was proven that (A & C) == C. This is trivial if C == A or C == 0.
194 /// If both A and C are constants, this proof is also easy.
195 /// For the following explanations, we assume that A is the mask.
197 /// "AllOnes" declares that the comparison is true only if (A & B) == A or all
198 /// bits of A are set in B.
199 /// Example: (icmp eq (A & 3), 3) -> AMask_AllOnes
201 /// "AllZeros" declares that the comparison is true only if (A & B) == 0 or all
202 /// bits of A are cleared in B.
203 /// Example: (icmp eq (A & 3), 0) -> Mask_AllZeroes
205 /// "Mixed" declares that (A & B) == C and C might or might not contain any
206 /// number of one bits and zero bits.
207 /// Example: (icmp eq (A & 3), 1) -> AMask_Mixed
209 /// "Not" means that in above descriptions "==" should be replaced by "!=".
210 /// Example: (icmp ne (A & 3), 3) -> AMask_NotAllOnes
212 /// If the mask A contains a single bit, then the following is equivalent:
213 /// (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
214 /// (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
215 enum MaskedICmpType {
216 AMask_AllOnes = 1,
217 AMask_NotAllOnes = 2,
218 BMask_AllOnes = 4,
219 BMask_NotAllOnes = 8,
220 Mask_AllZeros = 16,
221 Mask_NotAllZeros = 32,
222 AMask_Mixed = 64,
223 AMask_NotMixed = 128,
224 BMask_Mixed = 256,
225 BMask_NotMixed = 512
228 /// Return the set of patterns (from MaskedICmpType) that (icmp SCC (A & B), C)
229 /// satisfies.
230 static unsigned getMaskedICmpType(Value *A, Value *B, Value *C,
231 ICmpInst::Predicate Pred) {
232 ConstantInt *ACst = dyn_cast<ConstantInt>(A);
233 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
234 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
235 bool IsEq = (Pred == ICmpInst::ICMP_EQ);
236 bool IsAPow2 = (ACst && !ACst->isZero() && ACst->getValue().isPowerOf2());
237 bool IsBPow2 = (BCst && !BCst->isZero() && BCst->getValue().isPowerOf2());
238 unsigned MaskVal = 0;
239 if (CCst && CCst->isZero()) {
240 // if C is zero, then both A and B qualify as mask
241 MaskVal |= (IsEq ? (Mask_AllZeros | AMask_Mixed | BMask_Mixed)
242 : (Mask_NotAllZeros | AMask_NotMixed | BMask_NotMixed));
243 if (IsAPow2)
244 MaskVal |= (IsEq ? (AMask_NotAllOnes | AMask_NotMixed)
245 : (AMask_AllOnes | AMask_Mixed));
246 if (IsBPow2)
247 MaskVal |= (IsEq ? (BMask_NotAllOnes | BMask_NotMixed)
248 : (BMask_AllOnes | BMask_Mixed));
249 return MaskVal;
252 if (A == C) {
253 MaskVal |= (IsEq ? (AMask_AllOnes | AMask_Mixed)
254 : (AMask_NotAllOnes | AMask_NotMixed));
255 if (IsAPow2)
256 MaskVal |= (IsEq ? (Mask_NotAllZeros | AMask_NotMixed)
257 : (Mask_AllZeros | AMask_Mixed));
258 } else if (ACst && CCst && ConstantExpr::getAnd(ACst, CCst) == CCst) {
259 MaskVal |= (IsEq ? AMask_Mixed : AMask_NotMixed);
262 if (B == C) {
263 MaskVal |= (IsEq ? (BMask_AllOnes | BMask_Mixed)
264 : (BMask_NotAllOnes | BMask_NotMixed));
265 if (IsBPow2)
266 MaskVal |= (IsEq ? (Mask_NotAllZeros | BMask_NotMixed)
267 : (Mask_AllZeros | BMask_Mixed));
268 } else if (BCst && CCst && ConstantExpr::getAnd(BCst, CCst) == CCst) {
269 MaskVal |= (IsEq ? BMask_Mixed : BMask_NotMixed);
272 return MaskVal;
275 /// Convert an analysis of a masked ICmp into its equivalent if all boolean
276 /// operations had the opposite sense. Since each "NotXXX" flag (recording !=)
277 /// is adjacent to the corresponding normal flag (recording ==), this just
278 /// involves swapping those bits over.
279 static unsigned conjugateICmpMask(unsigned Mask) {
280 unsigned NewMask;
281 NewMask = (Mask & (AMask_AllOnes | BMask_AllOnes | Mask_AllZeros |
282 AMask_Mixed | BMask_Mixed))
283 << 1;
285 NewMask |= (Mask & (AMask_NotAllOnes | BMask_NotAllOnes | Mask_NotAllZeros |
286 AMask_NotMixed | BMask_NotMixed))
287 >> 1;
289 return NewMask;
292 // Adapts the external decomposeBitTestICmp for local use.
293 static bool decomposeBitTestICmp(Value *LHS, Value *RHS, CmpInst::Predicate &Pred,
294 Value *&X, Value *&Y, Value *&Z) {
295 APInt Mask;
296 if (!llvm::decomposeBitTestICmp(LHS, RHS, Pred, X, Mask))
297 return false;
299 Y = ConstantInt::get(X->getType(), Mask);
300 Z = ConstantInt::get(X->getType(), 0);
301 return true;
304 /// Handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E).
305 /// Return the pattern classes (from MaskedICmpType) for the left hand side and
306 /// the right hand side as a pair.
307 /// LHS and RHS are the left hand side and the right hand side ICmps and PredL
308 /// and PredR are their predicates, respectively.
309 static
310 Optional<std::pair<unsigned, unsigned>>
311 getMaskedTypeForICmpPair(Value *&A, Value *&B, Value *&C,
312 Value *&D, Value *&E, ICmpInst *LHS,
313 ICmpInst *RHS,
314 ICmpInst::Predicate &PredL,
315 ICmpInst::Predicate &PredR) {
316 // vectors are not (yet?) supported. Don't support pointers either.
317 if (!LHS->getOperand(0)->getType()->isIntegerTy() ||
318 !RHS->getOperand(0)->getType()->isIntegerTy())
319 return None;
321 // Here comes the tricky part:
322 // LHS might be of the form L11 & L12 == X, X == L21 & L22,
323 // and L11 & L12 == L21 & L22. The same goes for RHS.
324 // Now we must find those components L** and R**, that are equal, so
325 // that we can extract the parameters A, B, C, D, and E for the canonical
326 // above.
327 Value *L1 = LHS->getOperand(0);
328 Value *L2 = LHS->getOperand(1);
329 Value *L11, *L12, *L21, *L22;
330 // Check whether the icmp can be decomposed into a bit test.
331 if (decomposeBitTestICmp(L1, L2, PredL, L11, L12, L2)) {
332 L21 = L22 = L1 = nullptr;
333 } else {
334 // Look for ANDs in the LHS icmp.
335 if (!match(L1, m_And(m_Value(L11), m_Value(L12)))) {
336 // Any icmp can be viewed as being trivially masked; if it allows us to
337 // remove one, it's worth it.
338 L11 = L1;
339 L12 = Constant::getAllOnesValue(L1->getType());
342 if (!match(L2, m_And(m_Value(L21), m_Value(L22)))) {
343 L21 = L2;
344 L22 = Constant::getAllOnesValue(L2->getType());
348 // Bail if LHS was a icmp that can't be decomposed into an equality.
349 if (!ICmpInst::isEquality(PredL))
350 return None;
352 Value *R1 = RHS->getOperand(0);
353 Value *R2 = RHS->getOperand(1);
354 Value *R11, *R12;
355 bool Ok = false;
356 if (decomposeBitTestICmp(R1, R2, PredR, R11, R12, R2)) {
357 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
358 A = R11;
359 D = R12;
360 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
361 A = R12;
362 D = R11;
363 } else {
364 return None;
366 E = R2;
367 R1 = nullptr;
368 Ok = true;
369 } else {
370 if (!match(R1, m_And(m_Value(R11), m_Value(R12)))) {
371 // As before, model no mask as a trivial mask if it'll let us do an
372 // optimization.
373 R11 = R1;
374 R12 = Constant::getAllOnesValue(R1->getType());
377 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
378 A = R11;
379 D = R12;
380 E = R2;
381 Ok = true;
382 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
383 A = R12;
384 D = R11;
385 E = R2;
386 Ok = true;
390 // Bail if RHS was a icmp that can't be decomposed into an equality.
391 if (!ICmpInst::isEquality(PredR))
392 return None;
394 // Look for ANDs on the right side of the RHS icmp.
395 if (!Ok) {
396 if (!match(R2, m_And(m_Value(R11), m_Value(R12)))) {
397 R11 = R2;
398 R12 = Constant::getAllOnesValue(R2->getType());
401 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
402 A = R11;
403 D = R12;
404 E = R1;
405 Ok = true;
406 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
407 A = R12;
408 D = R11;
409 E = R1;
410 Ok = true;
411 } else {
412 return None;
415 if (!Ok)
416 return None;
418 if (L11 == A) {
419 B = L12;
420 C = L2;
421 } else if (L12 == A) {
422 B = L11;
423 C = L2;
424 } else if (L21 == A) {
425 B = L22;
426 C = L1;
427 } else if (L22 == A) {
428 B = L21;
429 C = L1;
432 unsigned LeftType = getMaskedICmpType(A, B, C, PredL);
433 unsigned RightType = getMaskedICmpType(A, D, E, PredR);
434 return Optional<std::pair<unsigned, unsigned>>(std::make_pair(LeftType, RightType));
437 /// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E) into a single
438 /// (icmp(A & X) ==/!= Y), where the left-hand side is of type Mask_NotAllZeros
439 /// and the right hand side is of type BMask_Mixed. For example,
440 /// (icmp (A & 12) != 0) & (icmp (A & 15) == 8) -> (icmp (A & 15) == 8).
441 static Value * foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(
442 ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
443 Value *A, Value *B, Value *C, Value *D, Value *E,
444 ICmpInst::Predicate PredL, ICmpInst::Predicate PredR,
445 llvm::InstCombiner::BuilderTy &Builder) {
446 // We are given the canonical form:
447 // (icmp ne (A & B), 0) & (icmp eq (A & D), E).
448 // where D & E == E.
450 // If IsAnd is false, we get it in negated form:
451 // (icmp eq (A & B), 0) | (icmp ne (A & D), E) ->
452 // !((icmp ne (A & B), 0) & (icmp eq (A & D), E)).
454 // We currently handle the case of B, C, D, E are constant.
456 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
457 if (!BCst)
458 return nullptr;
459 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
460 if (!CCst)
461 return nullptr;
462 ConstantInt *DCst = dyn_cast<ConstantInt>(D);
463 if (!DCst)
464 return nullptr;
465 ConstantInt *ECst = dyn_cast<ConstantInt>(E);
466 if (!ECst)
467 return nullptr;
469 ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
471 // Update E to the canonical form when D is a power of two and RHS is
472 // canonicalized as,
473 // (icmp ne (A & D), 0) -> (icmp eq (A & D), D) or
474 // (icmp ne (A & D), D) -> (icmp eq (A & D), 0).
475 if (PredR != NewCC)
476 ECst = cast<ConstantInt>(ConstantExpr::getXor(DCst, ECst));
478 // If B or D is zero, skip because if LHS or RHS can be trivially folded by
479 // other folding rules and this pattern won't apply any more.
480 if (BCst->getValue() == 0 || DCst->getValue() == 0)
481 return nullptr;
483 // If B and D don't intersect, ie. (B & D) == 0, no folding because we can't
484 // deduce anything from it.
485 // For example,
486 // (icmp ne (A & 12), 0) & (icmp eq (A & 3), 1) -> no folding.
487 if ((BCst->getValue() & DCst->getValue()) == 0)
488 return nullptr;
490 // If the following two conditions are met:
492 // 1. mask B covers only a single bit that's not covered by mask D, that is,
493 // (B & (B ^ D)) is a power of 2 (in other words, B minus the intersection of
494 // B and D has only one bit set) and,
496 // 2. RHS (and E) indicates that the rest of B's bits are zero (in other
497 // words, the intersection of B and D is zero), that is, ((B & D) & E) == 0
499 // then that single bit in B must be one and thus the whole expression can be
500 // folded to
501 // (A & (B | D)) == (B & (B ^ D)) | E.
503 // For example,
504 // (icmp ne (A & 12), 0) & (icmp eq (A & 7), 1) -> (icmp eq (A & 15), 9)
505 // (icmp ne (A & 15), 0) & (icmp eq (A & 7), 0) -> (icmp eq (A & 15), 8)
506 if ((((BCst->getValue() & DCst->getValue()) & ECst->getValue()) == 0) &&
507 (BCst->getValue() & (BCst->getValue() ^ DCst->getValue())).isPowerOf2()) {
508 APInt BorD = BCst->getValue() | DCst->getValue();
509 APInt BandBxorDorE = (BCst->getValue() & (BCst->getValue() ^ DCst->getValue())) |
510 ECst->getValue();
511 Value *NewMask = ConstantInt::get(BCst->getType(), BorD);
512 Value *NewMaskedValue = ConstantInt::get(BCst->getType(), BandBxorDorE);
513 Value *NewAnd = Builder.CreateAnd(A, NewMask);
514 return Builder.CreateICmp(NewCC, NewAnd, NewMaskedValue);
517 auto IsSubSetOrEqual = [](ConstantInt *C1, ConstantInt *C2) {
518 return (C1->getValue() & C2->getValue()) == C1->getValue();
520 auto IsSuperSetOrEqual = [](ConstantInt *C1, ConstantInt *C2) {
521 return (C1->getValue() & C2->getValue()) == C2->getValue();
524 // In the following, we consider only the cases where B is a superset of D, B
525 // is a subset of D, or B == D because otherwise there's at least one bit
526 // covered by B but not D, in which case we can't deduce much from it, so
527 // no folding (aside from the single must-be-one bit case right above.)
528 // For example,
529 // (icmp ne (A & 14), 0) & (icmp eq (A & 3), 1) -> no folding.
530 if (!IsSubSetOrEqual(BCst, DCst) && !IsSuperSetOrEqual(BCst, DCst))
531 return nullptr;
533 // At this point, either B is a superset of D, B is a subset of D or B == D.
535 // If E is zero, if B is a subset of (or equal to) D, LHS and RHS contradict
536 // and the whole expression becomes false (or true if negated), otherwise, no
537 // folding.
538 // For example,
539 // (icmp ne (A & 3), 0) & (icmp eq (A & 7), 0) -> false.
540 // (icmp ne (A & 15), 0) & (icmp eq (A & 3), 0) -> no folding.
541 if (ECst->isZero()) {
542 if (IsSubSetOrEqual(BCst, DCst))
543 return ConstantInt::get(LHS->getType(), !IsAnd);
544 return nullptr;
547 // At this point, B, D, E aren't zero and (B & D) == B, (B & D) == D or B ==
548 // D. If B is a superset of (or equal to) D, since E is not zero, LHS is
549 // subsumed by RHS (RHS implies LHS.) So the whole expression becomes
550 // RHS. For example,
551 // (icmp ne (A & 255), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).
552 // (icmp ne (A & 15), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).
553 if (IsSuperSetOrEqual(BCst, DCst))
554 return RHS;
555 // Otherwise, B is a subset of D. If B and E have a common bit set,
556 // ie. (B & E) != 0, then LHS is subsumed by RHS. For example.
557 // (icmp ne (A & 12), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).
558 assert(IsSubSetOrEqual(BCst, DCst) && "Precondition due to above code");
559 if ((BCst->getValue() & ECst->getValue()) != 0)
560 return RHS;
561 // Otherwise, LHS and RHS contradict and the whole expression becomes false
562 // (or true if negated.) For example,
563 // (icmp ne (A & 7), 0) & (icmp eq (A & 15), 8) -> false.
564 // (icmp ne (A & 6), 0) & (icmp eq (A & 15), 8) -> false.
565 return ConstantInt::get(LHS->getType(), !IsAnd);
568 /// Try to fold (icmp(A & B) ==/!= 0) &/| (icmp(A & D) ==/!= E) into a single
569 /// (icmp(A & X) ==/!= Y), where the left-hand side and the right hand side
570 /// aren't of the common mask pattern type.
571 static Value *foldLogOpOfMaskedICmpsAsymmetric(
572 ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
573 Value *A, Value *B, Value *C, Value *D, Value *E,
574 ICmpInst::Predicate PredL, ICmpInst::Predicate PredR,
575 unsigned LHSMask, unsigned RHSMask,
576 llvm::InstCombiner::BuilderTy &Builder) {
577 assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&
578 "Expected equality predicates for masked type of icmps.");
579 // Handle Mask_NotAllZeros-BMask_Mixed cases.
580 // (icmp ne/eq (A & B), C) &/| (icmp eq/ne (A & D), E), or
581 // (icmp eq/ne (A & B), C) &/| (icmp ne/eq (A & D), E)
582 // which gets swapped to
583 // (icmp ne/eq (A & D), E) &/| (icmp eq/ne (A & B), C).
584 if (!IsAnd) {
585 LHSMask = conjugateICmpMask(LHSMask);
586 RHSMask = conjugateICmpMask(RHSMask);
588 if ((LHSMask & Mask_NotAllZeros) && (RHSMask & BMask_Mixed)) {
589 if (Value *V = foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(
590 LHS, RHS, IsAnd, A, B, C, D, E,
591 PredL, PredR, Builder)) {
592 return V;
594 } else if ((LHSMask & BMask_Mixed) && (RHSMask & Mask_NotAllZeros)) {
595 if (Value *V = foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(
596 RHS, LHS, IsAnd, A, D, E, B, C,
597 PredR, PredL, Builder)) {
598 return V;
601 return nullptr;
604 /// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
605 /// into a single (icmp(A & X) ==/!= Y).
606 static Value *foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
607 llvm::InstCombiner::BuilderTy &Builder) {
608 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr;
609 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
610 Optional<std::pair<unsigned, unsigned>> MaskPair =
611 getMaskedTypeForICmpPair(A, B, C, D, E, LHS, RHS, PredL, PredR);
612 if (!MaskPair)
613 return nullptr;
614 assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&
615 "Expected equality predicates for masked type of icmps.");
616 unsigned LHSMask = MaskPair->first;
617 unsigned RHSMask = MaskPair->second;
618 unsigned Mask = LHSMask & RHSMask;
619 if (Mask == 0) {
620 // Even if the two sides don't share a common pattern, check if folding can
621 // still happen.
622 if (Value *V = foldLogOpOfMaskedICmpsAsymmetric(
623 LHS, RHS, IsAnd, A, B, C, D, E, PredL, PredR, LHSMask, RHSMask,
624 Builder))
625 return V;
626 return nullptr;
629 // In full generality:
630 // (icmp (A & B) Op C) | (icmp (A & D) Op E)
631 // == ![ (icmp (A & B) !Op C) & (icmp (A & D) !Op E) ]
633 // If the latter can be converted into (icmp (A & X) Op Y) then the former is
634 // equivalent to (icmp (A & X) !Op Y).
636 // Therefore, we can pretend for the rest of this function that we're dealing
637 // with the conjunction, provided we flip the sense of any comparisons (both
638 // input and output).
640 // In most cases we're going to produce an EQ for the "&&" case.
641 ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
642 if (!IsAnd) {
643 // Convert the masking analysis into its equivalent with negated
644 // comparisons.
645 Mask = conjugateICmpMask(Mask);
648 if (Mask & Mask_AllZeros) {
649 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
650 // -> (icmp eq (A & (B|D)), 0)
651 Value *NewOr = Builder.CreateOr(B, D);
652 Value *NewAnd = Builder.CreateAnd(A, NewOr);
653 // We can't use C as zero because we might actually handle
654 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
655 // with B and D, having a single bit set.
656 Value *Zero = Constant::getNullValue(A->getType());
657 return Builder.CreateICmp(NewCC, NewAnd, Zero);
659 if (Mask & BMask_AllOnes) {
660 // (icmp eq (A & B), B) & (icmp eq (A & D), D)
661 // -> (icmp eq (A & (B|D)), (B|D))
662 Value *NewOr = Builder.CreateOr(B, D);
663 Value *NewAnd = Builder.CreateAnd(A, NewOr);
664 return Builder.CreateICmp(NewCC, NewAnd, NewOr);
666 if (Mask & AMask_AllOnes) {
667 // (icmp eq (A & B), A) & (icmp eq (A & D), A)
668 // -> (icmp eq (A & (B&D)), A)
669 Value *NewAnd1 = Builder.CreateAnd(B, D);
670 Value *NewAnd2 = Builder.CreateAnd(A, NewAnd1);
671 return Builder.CreateICmp(NewCC, NewAnd2, A);
674 // Remaining cases assume at least that B and D are constant, and depend on
675 // their actual values. This isn't strictly necessary, just a "handle the
676 // easy cases for now" decision.
677 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
678 if (!BCst)
679 return nullptr;
680 ConstantInt *DCst = dyn_cast<ConstantInt>(D);
681 if (!DCst)
682 return nullptr;
684 if (Mask & (Mask_NotAllZeros | BMask_NotAllOnes)) {
685 // (icmp ne (A & B), 0) & (icmp ne (A & D), 0) and
686 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
687 // -> (icmp ne (A & B), 0) or (icmp ne (A & D), 0)
688 // Only valid if one of the masks is a superset of the other (check "B&D" is
689 // the same as either B or D).
690 APInt NewMask = BCst->getValue() & DCst->getValue();
692 if (NewMask == BCst->getValue())
693 return LHS;
694 else if (NewMask == DCst->getValue())
695 return RHS;
698 if (Mask & AMask_NotAllOnes) {
699 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
700 // -> (icmp ne (A & B), A) or (icmp ne (A & D), A)
701 // Only valid if one of the masks is a superset of the other (check "B|D" is
702 // the same as either B or D).
703 APInt NewMask = BCst->getValue() | DCst->getValue();
705 if (NewMask == BCst->getValue())
706 return LHS;
707 else if (NewMask == DCst->getValue())
708 return RHS;
711 if (Mask & BMask_Mixed) {
712 // (icmp eq (A & B), C) & (icmp eq (A & D), E)
713 // We already know that B & C == C && D & E == E.
714 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
715 // C and E, which are shared by both the mask B and the mask D, don't
716 // contradict, then we can transform to
717 // -> (icmp eq (A & (B|D)), (C|E))
718 // Currently, we only handle the case of B, C, D, and E being constant.
719 // We can't simply use C and E because we might actually handle
720 // (icmp ne (A & B), B) & (icmp eq (A & D), D)
721 // with B and D, having a single bit set.
722 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
723 if (!CCst)
724 return nullptr;
725 ConstantInt *ECst = dyn_cast<ConstantInt>(E);
726 if (!ECst)
727 return nullptr;
728 if (PredL != NewCC)
729 CCst = cast<ConstantInt>(ConstantExpr::getXor(BCst, CCst));
730 if (PredR != NewCC)
731 ECst = cast<ConstantInt>(ConstantExpr::getXor(DCst, ECst));
733 // If there is a conflict, we should actually return a false for the
734 // whole construct.
735 if (((BCst->getValue() & DCst->getValue()) &
736 (CCst->getValue() ^ ECst->getValue())).getBoolValue())
737 return ConstantInt::get(LHS->getType(), !IsAnd);
739 Value *NewOr1 = Builder.CreateOr(B, D);
740 Value *NewOr2 = ConstantExpr::getOr(CCst, ECst);
741 Value *NewAnd = Builder.CreateAnd(A, NewOr1);
742 return Builder.CreateICmp(NewCC, NewAnd, NewOr2);
745 return nullptr;
748 /// Try to fold a signed range checked with lower bound 0 to an unsigned icmp.
749 /// Example: (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
750 /// If \p Inverted is true then the check is for the inverted range, e.g.
751 /// (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
752 Value *InstCombiner::simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1,
753 bool Inverted) {
754 // Check the lower range comparison, e.g. x >= 0
755 // InstCombine already ensured that if there is a constant it's on the RHS.
756 ConstantInt *RangeStart = dyn_cast<ConstantInt>(Cmp0->getOperand(1));
757 if (!RangeStart)
758 return nullptr;
760 ICmpInst::Predicate Pred0 = (Inverted ? Cmp0->getInversePredicate() :
761 Cmp0->getPredicate());
763 // Accept x > -1 or x >= 0 (after potentially inverting the predicate).
764 if (!((Pred0 == ICmpInst::ICMP_SGT && RangeStart->isMinusOne()) ||
765 (Pred0 == ICmpInst::ICMP_SGE && RangeStart->isZero())))
766 return nullptr;
768 ICmpInst::Predicate Pred1 = (Inverted ? Cmp1->getInversePredicate() :
769 Cmp1->getPredicate());
771 Value *Input = Cmp0->getOperand(0);
772 Value *RangeEnd;
773 if (Cmp1->getOperand(0) == Input) {
774 // For the upper range compare we have: icmp x, n
775 RangeEnd = Cmp1->getOperand(1);
776 } else if (Cmp1->getOperand(1) == Input) {
777 // For the upper range compare we have: icmp n, x
778 RangeEnd = Cmp1->getOperand(0);
779 Pred1 = ICmpInst::getSwappedPredicate(Pred1);
780 } else {
781 return nullptr;
784 // Check the upper range comparison, e.g. x < n
785 ICmpInst::Predicate NewPred;
786 switch (Pred1) {
787 case ICmpInst::ICMP_SLT: NewPred = ICmpInst::ICMP_ULT; break;
788 case ICmpInst::ICMP_SLE: NewPred = ICmpInst::ICMP_ULE; break;
789 default: return nullptr;
792 // This simplification is only valid if the upper range is not negative.
793 KnownBits Known = computeKnownBits(RangeEnd, /*Depth=*/0, Cmp1);
794 if (!Known.isNonNegative())
795 return nullptr;
797 if (Inverted)
798 NewPred = ICmpInst::getInversePredicate(NewPred);
800 return Builder.CreateICmp(NewPred, Input, RangeEnd);
803 static Value *
804 foldAndOrOfEqualityCmpsWithConstants(ICmpInst *LHS, ICmpInst *RHS,
805 bool JoinedByAnd,
806 InstCombiner::BuilderTy &Builder) {
807 Value *X = LHS->getOperand(0);
808 if (X != RHS->getOperand(0))
809 return nullptr;
811 const APInt *C1, *C2;
812 if (!match(LHS->getOperand(1), m_APInt(C1)) ||
813 !match(RHS->getOperand(1), m_APInt(C2)))
814 return nullptr;
816 // We only handle (X != C1 && X != C2) and (X == C1 || X == C2).
817 ICmpInst::Predicate Pred = LHS->getPredicate();
818 if (Pred != RHS->getPredicate())
819 return nullptr;
820 if (JoinedByAnd && Pred != ICmpInst::ICMP_NE)
821 return nullptr;
822 if (!JoinedByAnd && Pred != ICmpInst::ICMP_EQ)
823 return nullptr;
825 // The larger unsigned constant goes on the right.
826 if (C1->ugt(*C2))
827 std::swap(C1, C2);
829 APInt Xor = *C1 ^ *C2;
830 if (Xor.isPowerOf2()) {
831 // If LHSC and RHSC differ by only one bit, then set that bit in X and
832 // compare against the larger constant:
833 // (X == C1 || X == C2) --> (X | (C1 ^ C2)) == C2
834 // (X != C1 && X != C2) --> (X | (C1 ^ C2)) != C2
835 // We choose an 'or' with a Pow2 constant rather than the inverse mask with
836 // 'and' because that may lead to smaller codegen from a smaller constant.
837 Value *Or = Builder.CreateOr(X, ConstantInt::get(X->getType(), Xor));
838 return Builder.CreateICmp(Pred, Or, ConstantInt::get(X->getType(), *C2));
841 // Special case: get the ordering right when the values wrap around zero.
842 // Ie, we assumed the constants were unsigned when swapping earlier.
843 if (C1->isNullValue() && C2->isAllOnesValue())
844 std::swap(C1, C2);
846 if (*C1 == *C2 - 1) {
847 // (X == 13 || X == 14) --> X - 13 <=u 1
848 // (X != 13 && X != 14) --> X - 13 >u 1
849 // An 'add' is the canonical IR form, so favor that over a 'sub'.
850 Value *Add = Builder.CreateAdd(X, ConstantInt::get(X->getType(), -(*C1)));
851 auto NewPred = JoinedByAnd ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_ULE;
852 return Builder.CreateICmp(NewPred, Add, ConstantInt::get(X->getType(), 1));
855 return nullptr;
858 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2)
859 // Fold (!iszero(A & K1) & !iszero(A & K2)) -> (A & (K1 | K2)) == (K1 | K2)
860 Value *InstCombiner::foldAndOrOfICmpsOfAndWithPow2(ICmpInst *LHS, ICmpInst *RHS,
861 bool JoinedByAnd,
862 Instruction &CxtI) {
863 ICmpInst::Predicate Pred = LHS->getPredicate();
864 if (Pred != RHS->getPredicate())
865 return nullptr;
866 if (JoinedByAnd && Pred != ICmpInst::ICMP_NE)
867 return nullptr;
868 if (!JoinedByAnd && Pred != ICmpInst::ICMP_EQ)
869 return nullptr;
871 // TODO support vector splats
872 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHS->getOperand(1));
873 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS->getOperand(1));
874 if (!LHSC || !RHSC || !LHSC->isZero() || !RHSC->isZero())
875 return nullptr;
877 Value *A, *B, *C, *D;
878 if (match(LHS->getOperand(0), m_And(m_Value(A), m_Value(B))) &&
879 match(RHS->getOperand(0), m_And(m_Value(C), m_Value(D)))) {
880 if (A == D || B == D)
881 std::swap(C, D);
882 if (B == C)
883 std::swap(A, B);
885 if (A == C &&
886 isKnownToBeAPowerOfTwo(B, false, 0, &CxtI) &&
887 isKnownToBeAPowerOfTwo(D, false, 0, &CxtI)) {
888 Value *Mask = Builder.CreateOr(B, D);
889 Value *Masked = Builder.CreateAnd(A, Mask);
890 auto NewPred = JoinedByAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
891 return Builder.CreateICmp(NewPred, Masked, Mask);
895 return nullptr;
898 /// General pattern:
899 /// X & Y
901 /// Where Y is checking that all the high bits (covered by a mask 4294967168)
902 /// are uniform, i.e. %arg & 4294967168 can be either 4294967168 or 0
903 /// Pattern can be one of:
904 /// %t = add i32 %arg, 128
905 /// %r = icmp ult i32 %t, 256
906 /// Or
907 /// %t0 = shl i32 %arg, 24
908 /// %t1 = ashr i32 %t0, 24
909 /// %r = icmp eq i32 %t1, %arg
910 /// Or
911 /// %t0 = trunc i32 %arg to i8
912 /// %t1 = sext i8 %t0 to i32
913 /// %r = icmp eq i32 %t1, %arg
914 /// This pattern is a signed truncation check.
916 /// And X is checking that some bit in that same mask is zero.
917 /// I.e. can be one of:
918 /// %r = icmp sgt i32 %arg, -1
919 /// Or
920 /// %t = and i32 %arg, 2147483648
921 /// %r = icmp eq i32 %t, 0
923 /// Since we are checking that all the bits in that mask are the same,
924 /// and a particular bit is zero, what we are really checking is that all the
925 /// masked bits are zero.
926 /// So this should be transformed to:
927 /// %r = icmp ult i32 %arg, 128
928 static Value *foldSignedTruncationCheck(ICmpInst *ICmp0, ICmpInst *ICmp1,
929 Instruction &CxtI,
930 InstCombiner::BuilderTy &Builder) {
931 assert(CxtI.getOpcode() == Instruction::And);
933 // Match icmp ult (add %arg, C01), C1 (C1 == C01 << 1; powers of two)
934 auto tryToMatchSignedTruncationCheck = [](ICmpInst *ICmp, Value *&X,
935 APInt &SignBitMask) -> bool {
936 CmpInst::Predicate Pred;
937 const APInt *I01, *I1; // powers of two; I1 == I01 << 1
938 if (!(match(ICmp,
939 m_ICmp(Pred, m_Add(m_Value(X), m_Power2(I01)), m_Power2(I1))) &&
940 Pred == ICmpInst::ICMP_ULT && I1->ugt(*I01) && I01->shl(1) == *I1))
941 return false;
942 // Which bit is the new sign bit as per the 'signed truncation' pattern?
943 SignBitMask = *I01;
944 return true;
947 // One icmp needs to be 'signed truncation check'.
948 // We need to match this first, else we will mismatch commutative cases.
949 Value *X1;
950 APInt HighestBit;
951 ICmpInst *OtherICmp;
952 if (tryToMatchSignedTruncationCheck(ICmp1, X1, HighestBit))
953 OtherICmp = ICmp0;
954 else if (tryToMatchSignedTruncationCheck(ICmp0, X1, HighestBit))
955 OtherICmp = ICmp1;
956 else
957 return nullptr;
959 assert(HighestBit.isPowerOf2() && "expected to be power of two (non-zero)");
961 // Try to match/decompose into: icmp eq (X & Mask), 0
962 auto tryToDecompose = [](ICmpInst *ICmp, Value *&X,
963 APInt &UnsetBitsMask) -> bool {
964 CmpInst::Predicate Pred = ICmp->getPredicate();
965 // Can it be decomposed into icmp eq (X & Mask), 0 ?
966 if (llvm::decomposeBitTestICmp(ICmp->getOperand(0), ICmp->getOperand(1),
967 Pred, X, UnsetBitsMask,
968 /*LookThroughTrunc=*/false) &&
969 Pred == ICmpInst::ICMP_EQ)
970 return true;
971 // Is it icmp eq (X & Mask), 0 already?
972 const APInt *Mask;
973 if (match(ICmp, m_ICmp(Pred, m_And(m_Value(X), m_APInt(Mask)), m_Zero())) &&
974 Pred == ICmpInst::ICMP_EQ) {
975 UnsetBitsMask = *Mask;
976 return true;
978 return false;
981 // And the other icmp needs to be decomposable into a bit test.
982 Value *X0;
983 APInt UnsetBitsMask;
984 if (!tryToDecompose(OtherICmp, X0, UnsetBitsMask))
985 return nullptr;
987 assert(!UnsetBitsMask.isNullValue() && "empty mask makes no sense.");
989 // Are they working on the same value?
990 Value *X;
991 if (X1 == X0) {
992 // Ok as is.
993 X = X1;
994 } else if (match(X0, m_Trunc(m_Specific(X1)))) {
995 UnsetBitsMask = UnsetBitsMask.zext(X1->getType()->getScalarSizeInBits());
996 X = X1;
997 } else
998 return nullptr;
1000 // So which bits should be uniform as per the 'signed truncation check'?
1001 // (all the bits starting with (i.e. including) HighestBit)
1002 APInt SignBitsMask = ~(HighestBit - 1U);
1004 // UnsetBitsMask must have some common bits with SignBitsMask,
1005 if (!UnsetBitsMask.intersects(SignBitsMask))
1006 return nullptr;
1008 // Does UnsetBitsMask contain any bits outside of SignBitsMask?
1009 if (!UnsetBitsMask.isSubsetOf(SignBitsMask)) {
1010 APInt OtherHighestBit = (~UnsetBitsMask) + 1U;
1011 if (!OtherHighestBit.isPowerOf2())
1012 return nullptr;
1013 HighestBit = APIntOps::umin(HighestBit, OtherHighestBit);
1015 // Else, if it does not, then all is ok as-is.
1017 // %r = icmp ult %X, SignBit
1018 return Builder.CreateICmpULT(X, ConstantInt::get(X->getType(), HighestBit),
1019 CxtI.getName() + ".simplified");
1022 /// Reduce a pair of compares that check if a value has exactly 1 bit set.
1023 static Value *foldIsPowerOf2(ICmpInst *Cmp0, ICmpInst *Cmp1, bool JoinedByAnd,
1024 InstCombiner::BuilderTy &Builder) {
1025 // Handle 'and' / 'or' commutation: make the equality check the first operand.
1026 if (JoinedByAnd && Cmp1->getPredicate() == ICmpInst::ICMP_NE)
1027 std::swap(Cmp0, Cmp1);
1028 else if (!JoinedByAnd && Cmp1->getPredicate() == ICmpInst::ICMP_EQ)
1029 std::swap(Cmp0, Cmp1);
1031 // (X != 0) && (ctpop(X) u< 2) --> ctpop(X) == 1
1032 CmpInst::Predicate Pred0, Pred1;
1033 Value *X;
1034 if (JoinedByAnd && match(Cmp0, m_ICmp(Pred0, m_Value(X), m_ZeroInt())) &&
1035 match(Cmp1, m_ICmp(Pred1, m_Intrinsic<Intrinsic::ctpop>(m_Specific(X)),
1036 m_SpecificInt(2))) &&
1037 Pred0 == ICmpInst::ICMP_NE && Pred1 == ICmpInst::ICMP_ULT) {
1038 Value *CtPop = Cmp1->getOperand(0);
1039 return Builder.CreateICmpEQ(CtPop, ConstantInt::get(CtPop->getType(), 1));
1041 // (X == 0) || (ctpop(X) u> 1) --> ctpop(X) != 1
1042 if (!JoinedByAnd && match(Cmp0, m_ICmp(Pred0, m_Value(X), m_ZeroInt())) &&
1043 match(Cmp1, m_ICmp(Pred1, m_Intrinsic<Intrinsic::ctpop>(m_Specific(X)),
1044 m_SpecificInt(1))) &&
1045 Pred0 == ICmpInst::ICMP_EQ && Pred1 == ICmpInst::ICMP_UGT) {
1046 Value *CtPop = Cmp1->getOperand(0);
1047 return Builder.CreateICmpNE(CtPop, ConstantInt::get(CtPop->getType(), 1));
1049 return nullptr;
1052 /// Fold (icmp)&(icmp) if possible.
1053 Value *InstCombiner::foldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS,
1054 Instruction &CxtI) {
1055 // Fold (!iszero(A & K1) & !iszero(A & K2)) -> (A & (K1 | K2)) == (K1 | K2)
1056 // if K1 and K2 are a one-bit mask.
1057 if (Value *V = foldAndOrOfICmpsOfAndWithPow2(LHS, RHS, true, CxtI))
1058 return V;
1060 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1062 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
1063 if (predicatesFoldable(PredL, PredR)) {
1064 if (LHS->getOperand(0) == RHS->getOperand(1) &&
1065 LHS->getOperand(1) == RHS->getOperand(0))
1066 LHS->swapOperands();
1067 if (LHS->getOperand(0) == RHS->getOperand(0) &&
1068 LHS->getOperand(1) == RHS->getOperand(1)) {
1069 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1070 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
1071 bool IsSigned = LHS->isSigned() || RHS->isSigned();
1072 return getNewICmpValue(Code, IsSigned, Op0, Op1, Builder);
1076 // handle (roughly): (icmp eq (A & B), C) & (icmp eq (A & D), E)
1077 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, true, Builder))
1078 return V;
1080 // E.g. (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
1081 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/false))
1082 return V;
1084 // E.g. (icmp slt x, n) & (icmp sge x, 0) --> icmp ult x, n
1085 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/false))
1086 return V;
1088 if (Value *V = foldAndOrOfEqualityCmpsWithConstants(LHS, RHS, true, Builder))
1089 return V;
1091 if (Value *V = foldSignedTruncationCheck(LHS, RHS, CxtI, Builder))
1092 return V;
1094 if (Value *V = foldIsPowerOf2(LHS, RHS, true /* JoinedByAnd */, Builder))
1095 return V;
1097 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
1098 Value *LHS0 = LHS->getOperand(0), *RHS0 = RHS->getOperand(0);
1099 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHS->getOperand(1));
1100 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS->getOperand(1));
1101 if (!LHSC || !RHSC)
1102 return nullptr;
1104 if (LHSC == RHSC && PredL == PredR) {
1105 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
1106 // where C is a power of 2 or
1107 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
1108 if ((PredL == ICmpInst::ICMP_ULT && LHSC->getValue().isPowerOf2()) ||
1109 (PredL == ICmpInst::ICMP_EQ && LHSC->isZero())) {
1110 Value *NewOr = Builder.CreateOr(LHS0, RHS0);
1111 return Builder.CreateICmp(PredL, NewOr, LHSC);
1115 // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2
1116 // where CMAX is the all ones value for the truncated type,
1117 // iff the lower bits of C2 and CA are zero.
1118 if (PredL == ICmpInst::ICMP_EQ && PredL == PredR && LHS->hasOneUse() &&
1119 RHS->hasOneUse()) {
1120 Value *V;
1121 ConstantInt *AndC, *SmallC = nullptr, *BigC = nullptr;
1123 // (trunc x) == C1 & (and x, CA) == C2
1124 // (and x, CA) == C2 & (trunc x) == C1
1125 if (match(RHS0, m_Trunc(m_Value(V))) &&
1126 match(LHS0, m_And(m_Specific(V), m_ConstantInt(AndC)))) {
1127 SmallC = RHSC;
1128 BigC = LHSC;
1129 } else if (match(LHS0, m_Trunc(m_Value(V))) &&
1130 match(RHS0, m_And(m_Specific(V), m_ConstantInt(AndC)))) {
1131 SmallC = LHSC;
1132 BigC = RHSC;
1135 if (SmallC && BigC) {
1136 unsigned BigBitSize = BigC->getType()->getBitWidth();
1137 unsigned SmallBitSize = SmallC->getType()->getBitWidth();
1139 // Check that the low bits are zero.
1140 APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize);
1141 if ((Low & AndC->getValue()).isNullValue() &&
1142 (Low & BigC->getValue()).isNullValue()) {
1143 Value *NewAnd = Builder.CreateAnd(V, Low | AndC->getValue());
1144 APInt N = SmallC->getValue().zext(BigBitSize) | BigC->getValue();
1145 Value *NewVal = ConstantInt::get(AndC->getType()->getContext(), N);
1146 return Builder.CreateICmp(PredL, NewAnd, NewVal);
1151 // From here on, we only handle:
1152 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
1153 if (LHS0 != RHS0)
1154 return nullptr;
1156 // ICMP_[US][GL]E X, C is folded to ICMP_[US][GL]T elsewhere.
1157 if (PredL == ICmpInst::ICMP_UGE || PredL == ICmpInst::ICMP_ULE ||
1158 PredR == ICmpInst::ICMP_UGE || PredR == ICmpInst::ICMP_ULE ||
1159 PredL == ICmpInst::ICMP_SGE || PredL == ICmpInst::ICMP_SLE ||
1160 PredR == ICmpInst::ICMP_SGE || PredR == ICmpInst::ICMP_SLE)
1161 return nullptr;
1163 // We can't fold (ugt x, C) & (sgt x, C2).
1164 if (!predicatesFoldable(PredL, PredR))
1165 return nullptr;
1167 // Ensure that the larger constant is on the RHS.
1168 bool ShouldSwap;
1169 if (CmpInst::isSigned(PredL) ||
1170 (ICmpInst::isEquality(PredL) && CmpInst::isSigned(PredR)))
1171 ShouldSwap = LHSC->getValue().sgt(RHSC->getValue());
1172 else
1173 ShouldSwap = LHSC->getValue().ugt(RHSC->getValue());
1175 if (ShouldSwap) {
1176 std::swap(LHS, RHS);
1177 std::swap(LHSC, RHSC);
1178 std::swap(PredL, PredR);
1181 // At this point, we know we have two icmp instructions
1182 // comparing a value against two constants and and'ing the result
1183 // together. Because of the above check, we know that we only have
1184 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
1185 // (from the icmp folding check above), that the two constants
1186 // are not equal and that the larger constant is on the RHS
1187 assert(LHSC != RHSC && "Compares not folded above?");
1189 switch (PredL) {
1190 default:
1191 llvm_unreachable("Unknown integer condition code!");
1192 case ICmpInst::ICMP_NE:
1193 switch (PredR) {
1194 default:
1195 llvm_unreachable("Unknown integer condition code!");
1196 case ICmpInst::ICMP_ULT:
1197 // (X != 13 & X u< 14) -> X < 13
1198 if (LHSC->getValue() == (RHSC->getValue() - 1))
1199 return Builder.CreateICmpULT(LHS0, LHSC);
1200 if (LHSC->isZero()) // (X != 0 & X u< C) -> X-1 u< C-1
1201 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(),
1202 false, true);
1203 break; // (X != 13 & X u< 15) -> no change
1204 case ICmpInst::ICMP_SLT:
1205 // (X != 13 & X s< 14) -> X < 13
1206 if (LHSC->getValue() == (RHSC->getValue() - 1))
1207 return Builder.CreateICmpSLT(LHS0, LHSC);
1208 // (X != INT_MIN & X s< C) -> X-(INT_MIN+1) u< (C-(INT_MIN+1))
1209 if (LHSC->isMinValue(true))
1210 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(),
1211 true, true);
1212 break; // (X != 13 & X s< 15) -> no change
1213 case ICmpInst::ICMP_NE:
1214 // Potential folds for this case should already be handled.
1215 break;
1217 break;
1218 case ICmpInst::ICMP_UGT:
1219 switch (PredR) {
1220 default:
1221 llvm_unreachable("Unknown integer condition code!");
1222 case ICmpInst::ICMP_NE:
1223 // (X u> 13 & X != 14) -> X u> 14
1224 if (RHSC->getValue() == (LHSC->getValue() + 1))
1225 return Builder.CreateICmp(PredL, LHS0, RHSC);
1226 // X u> C & X != UINT_MAX -> (X-(C+1)) u< UINT_MAX-(C+1)
1227 if (RHSC->isMaxValue(false))
1228 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(),
1229 false, true);
1230 break; // (X u> 13 & X != 15) -> no change
1231 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) u< 1
1232 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(),
1233 false, true);
1235 break;
1236 case ICmpInst::ICMP_SGT:
1237 switch (PredR) {
1238 default:
1239 llvm_unreachable("Unknown integer condition code!");
1240 case ICmpInst::ICMP_NE:
1241 // (X s> 13 & X != 14) -> X s> 14
1242 if (RHSC->getValue() == (LHSC->getValue() + 1))
1243 return Builder.CreateICmp(PredL, LHS0, RHSC);
1244 // X s> C & X != INT_MAX -> (X-(C+1)) u< INT_MAX-(C+1)
1245 if (RHSC->isMaxValue(true))
1246 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(),
1247 true, true);
1248 break; // (X s> 13 & X != 15) -> no change
1249 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) u< 1
1250 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(), true,
1251 true);
1253 break;
1256 return nullptr;
1259 Value *InstCombiner::foldLogicOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1260 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1261 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1262 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1264 if (LHS0 == RHS1 && RHS0 == LHS1) {
1265 // Swap RHS operands to match LHS.
1266 PredR = FCmpInst::getSwappedPredicate(PredR);
1267 std::swap(RHS0, RHS1);
1270 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
1271 // Suppose the relation between x and y is R, where R is one of
1272 // U(1000), L(0100), G(0010) or E(0001), and CC0 and CC1 are the bitmasks for
1273 // testing the desired relations.
1275 // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this:
1276 // bool(R & CC0) && bool(R & CC1)
1277 // = bool((R & CC0) & (R & CC1))
1278 // = bool(R & (CC0 & CC1)) <= by re-association, commutation, and idempotency
1280 // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this:
1281 // bool(R & CC0) || bool(R & CC1)
1282 // = bool((R & CC0) | (R & CC1))
1283 // = bool(R & (CC0 | CC1)) <= by reversed distribution (contribution? ;)
1284 if (LHS0 == RHS0 && LHS1 == RHS1) {
1285 unsigned FCmpCodeL = getFCmpCode(PredL);
1286 unsigned FCmpCodeR = getFCmpCode(PredR);
1287 unsigned NewPred = IsAnd ? FCmpCodeL & FCmpCodeR : FCmpCodeL | FCmpCodeR;
1288 return getFCmpValue(NewPred, LHS0, LHS1, Builder);
1291 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1292 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1293 if (LHS0->getType() != RHS0->getType())
1294 return nullptr;
1296 // FCmp canonicalization ensures that (fcmp ord/uno X, X) and
1297 // (fcmp ord/uno X, C) will be transformed to (fcmp X, +0.0).
1298 if (match(LHS1, m_PosZeroFP()) && match(RHS1, m_PosZeroFP()))
1299 // Ignore the constants because they are obviously not NANs:
1300 // (fcmp ord x, 0.0) & (fcmp ord y, 0.0) -> (fcmp ord x, y)
1301 // (fcmp uno x, 0.0) | (fcmp uno y, 0.0) -> (fcmp uno x, y)
1302 return Builder.CreateFCmp(PredL, LHS0, RHS0);
1305 return nullptr;
1308 /// This a limited reassociation for a special case (see above) where we are
1309 /// checking if two values are either both NAN (unordered) or not-NAN (ordered).
1310 /// This could be handled more generally in '-reassociation', but it seems like
1311 /// an unlikely pattern for a large number of logic ops and fcmps.
1312 static Instruction *reassociateFCmps(BinaryOperator &BO,
1313 InstCombiner::BuilderTy &Builder) {
1314 Instruction::BinaryOps Opcode = BO.getOpcode();
1315 assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&
1316 "Expecting and/or op for fcmp transform");
1318 // There are 4 commuted variants of the pattern. Canonicalize operands of this
1319 // logic op so an fcmp is operand 0 and a matching logic op is operand 1.
1320 Value *Op0 = BO.getOperand(0), *Op1 = BO.getOperand(1), *X;
1321 FCmpInst::Predicate Pred;
1322 if (match(Op1, m_FCmp(Pred, m_Value(), m_AnyZeroFP())))
1323 std::swap(Op0, Op1);
1325 // Match inner binop and the predicate for combining 2 NAN checks into 1.
1326 BinaryOperator *BO1;
1327 FCmpInst::Predicate NanPred = Opcode == Instruction::And ? FCmpInst::FCMP_ORD
1328 : FCmpInst::FCMP_UNO;
1329 if (!match(Op0, m_FCmp(Pred, m_Value(X), m_AnyZeroFP())) || Pred != NanPred ||
1330 !match(Op1, m_BinOp(BO1)) || BO1->getOpcode() != Opcode)
1331 return nullptr;
1333 // The inner logic op must have a matching fcmp operand.
1334 Value *BO10 = BO1->getOperand(0), *BO11 = BO1->getOperand(1), *Y;
1335 if (!match(BO10, m_FCmp(Pred, m_Value(Y), m_AnyZeroFP())) ||
1336 Pred != NanPred || X->getType() != Y->getType())
1337 std::swap(BO10, BO11);
1339 if (!match(BO10, m_FCmp(Pred, m_Value(Y), m_AnyZeroFP())) ||
1340 Pred != NanPred || X->getType() != Y->getType())
1341 return nullptr;
1343 // and (fcmp ord X, 0), (and (fcmp ord Y, 0), Z) --> and (fcmp ord X, Y), Z
1344 // or (fcmp uno X, 0), (or (fcmp uno Y, 0), Z) --> or (fcmp uno X, Y), Z
1345 Value *NewFCmp = Builder.CreateFCmp(Pred, X, Y);
1346 if (auto *NewFCmpInst = dyn_cast<FCmpInst>(NewFCmp)) {
1347 // Intersect FMF from the 2 source fcmps.
1348 NewFCmpInst->copyIRFlags(Op0);
1349 NewFCmpInst->andIRFlags(BO10);
1351 return BinaryOperator::Create(Opcode, NewFCmp, BO11);
1354 /// Match De Morgan's Laws:
1355 /// (~A & ~B) == (~(A | B))
1356 /// (~A | ~B) == (~(A & B))
1357 static Instruction *matchDeMorgansLaws(BinaryOperator &I,
1358 InstCombiner::BuilderTy &Builder) {
1359 auto Opcode = I.getOpcode();
1360 assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&
1361 "Trying to match De Morgan's Laws with something other than and/or");
1363 // Flip the logic operation.
1364 Opcode = (Opcode == Instruction::And) ? Instruction::Or : Instruction::And;
1366 Value *A, *B;
1367 if (match(I.getOperand(0), m_OneUse(m_Not(m_Value(A)))) &&
1368 match(I.getOperand(1), m_OneUse(m_Not(m_Value(B)))) &&
1369 !isFreeToInvert(A, A->hasOneUse()) &&
1370 !isFreeToInvert(B, B->hasOneUse())) {
1371 Value *AndOr = Builder.CreateBinOp(Opcode, A, B, I.getName() + ".demorgan");
1372 return BinaryOperator::CreateNot(AndOr);
1375 return nullptr;
1378 bool InstCombiner::shouldOptimizeCast(CastInst *CI) {
1379 Value *CastSrc = CI->getOperand(0);
1381 // Noop casts and casts of constants should be eliminated trivially.
1382 if (CI->getSrcTy() == CI->getDestTy() || isa<Constant>(CastSrc))
1383 return false;
1385 // If this cast is paired with another cast that can be eliminated, we prefer
1386 // to have it eliminated.
1387 if (const auto *PrecedingCI = dyn_cast<CastInst>(CastSrc))
1388 if (isEliminableCastPair(PrecedingCI, CI))
1389 return false;
1391 return true;
1394 /// Fold {and,or,xor} (cast X), C.
1395 static Instruction *foldLogicCastConstant(BinaryOperator &Logic, CastInst *Cast,
1396 InstCombiner::BuilderTy &Builder) {
1397 Constant *C = dyn_cast<Constant>(Logic.getOperand(1));
1398 if (!C)
1399 return nullptr;
1401 auto LogicOpc = Logic.getOpcode();
1402 Type *DestTy = Logic.getType();
1403 Type *SrcTy = Cast->getSrcTy();
1405 // Move the logic operation ahead of a zext or sext if the constant is
1406 // unchanged in the smaller source type. Performing the logic in a smaller
1407 // type may provide more information to later folds, and the smaller logic
1408 // instruction may be cheaper (particularly in the case of vectors).
1409 Value *X;
1410 if (match(Cast, m_OneUse(m_ZExt(m_Value(X))))) {
1411 Constant *TruncC = ConstantExpr::getTrunc(C, SrcTy);
1412 Constant *ZextTruncC = ConstantExpr::getZExt(TruncC, DestTy);
1413 if (ZextTruncC == C) {
1414 // LogicOpc (zext X), C --> zext (LogicOpc X, C)
1415 Value *NewOp = Builder.CreateBinOp(LogicOpc, X, TruncC);
1416 return new ZExtInst(NewOp, DestTy);
1420 if (match(Cast, m_OneUse(m_SExt(m_Value(X))))) {
1421 Constant *TruncC = ConstantExpr::getTrunc(C, SrcTy);
1422 Constant *SextTruncC = ConstantExpr::getSExt(TruncC, DestTy);
1423 if (SextTruncC == C) {
1424 // LogicOpc (sext X), C --> sext (LogicOpc X, C)
1425 Value *NewOp = Builder.CreateBinOp(LogicOpc, X, TruncC);
1426 return new SExtInst(NewOp, DestTy);
1430 return nullptr;
1433 /// Fold {and,or,xor} (cast X), Y.
1434 Instruction *InstCombiner::foldCastedBitwiseLogic(BinaryOperator &I) {
1435 auto LogicOpc = I.getOpcode();
1436 assert(I.isBitwiseLogicOp() && "Unexpected opcode for bitwise logic folding");
1438 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1439 CastInst *Cast0 = dyn_cast<CastInst>(Op0);
1440 if (!Cast0)
1441 return nullptr;
1443 // This must be a cast from an integer or integer vector source type to allow
1444 // transformation of the logic operation to the source type.
1445 Type *DestTy = I.getType();
1446 Type *SrcTy = Cast0->getSrcTy();
1447 if (!SrcTy->isIntOrIntVectorTy())
1448 return nullptr;
1450 if (Instruction *Ret = foldLogicCastConstant(I, Cast0, Builder))
1451 return Ret;
1453 CastInst *Cast1 = dyn_cast<CastInst>(Op1);
1454 if (!Cast1)
1455 return nullptr;
1457 // Both operands of the logic operation are casts. The casts must be of the
1458 // same type for reduction.
1459 auto CastOpcode = Cast0->getOpcode();
1460 if (CastOpcode != Cast1->getOpcode() || SrcTy != Cast1->getSrcTy())
1461 return nullptr;
1463 Value *Cast0Src = Cast0->getOperand(0);
1464 Value *Cast1Src = Cast1->getOperand(0);
1466 // fold logic(cast(A), cast(B)) -> cast(logic(A, B))
1467 if (shouldOptimizeCast(Cast0) && shouldOptimizeCast(Cast1)) {
1468 Value *NewOp = Builder.CreateBinOp(LogicOpc, Cast0Src, Cast1Src,
1469 I.getName());
1470 return CastInst::Create(CastOpcode, NewOp, DestTy);
1473 // For now, only 'and'/'or' have optimizations after this.
1474 if (LogicOpc == Instruction::Xor)
1475 return nullptr;
1477 // If this is logic(cast(icmp), cast(icmp)), try to fold this even if the
1478 // cast is otherwise not optimizable. This happens for vector sexts.
1479 ICmpInst *ICmp0 = dyn_cast<ICmpInst>(Cast0Src);
1480 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(Cast1Src);
1481 if (ICmp0 && ICmp1) {
1482 Value *Res = LogicOpc == Instruction::And ? foldAndOfICmps(ICmp0, ICmp1, I)
1483 : foldOrOfICmps(ICmp0, ICmp1, I);
1484 if (Res)
1485 return CastInst::Create(CastOpcode, Res, DestTy);
1486 return nullptr;
1489 // If this is logic(cast(fcmp), cast(fcmp)), try to fold this even if the
1490 // cast is otherwise not optimizable. This happens for vector sexts.
1491 FCmpInst *FCmp0 = dyn_cast<FCmpInst>(Cast0Src);
1492 FCmpInst *FCmp1 = dyn_cast<FCmpInst>(Cast1Src);
1493 if (FCmp0 && FCmp1)
1494 if (Value *R = foldLogicOfFCmps(FCmp0, FCmp1, LogicOpc == Instruction::And))
1495 return CastInst::Create(CastOpcode, R, DestTy);
1497 return nullptr;
1500 static Instruction *foldAndToXor(BinaryOperator &I,
1501 InstCombiner::BuilderTy &Builder) {
1502 assert(I.getOpcode() == Instruction::And);
1503 Value *Op0 = I.getOperand(0);
1504 Value *Op1 = I.getOperand(1);
1505 Value *A, *B;
1507 // Operand complexity canonicalization guarantees that the 'or' is Op0.
1508 // (A | B) & ~(A & B) --> A ^ B
1509 // (A | B) & ~(B & A) --> A ^ B
1510 if (match(&I, m_BinOp(m_Or(m_Value(A), m_Value(B)),
1511 m_Not(m_c_And(m_Deferred(A), m_Deferred(B))))))
1512 return BinaryOperator::CreateXor(A, B);
1514 // (A | ~B) & (~A | B) --> ~(A ^ B)
1515 // (A | ~B) & (B | ~A) --> ~(A ^ B)
1516 // (~B | A) & (~A | B) --> ~(A ^ B)
1517 // (~B | A) & (B | ~A) --> ~(A ^ B)
1518 if (Op0->hasOneUse() || Op1->hasOneUse())
1519 if (match(&I, m_BinOp(m_c_Or(m_Value(A), m_Not(m_Value(B))),
1520 m_c_Or(m_Not(m_Deferred(A)), m_Deferred(B)))))
1521 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
1523 return nullptr;
1526 static Instruction *foldOrToXor(BinaryOperator &I,
1527 InstCombiner::BuilderTy &Builder) {
1528 assert(I.getOpcode() == Instruction::Or);
1529 Value *Op0 = I.getOperand(0);
1530 Value *Op1 = I.getOperand(1);
1531 Value *A, *B;
1533 // Operand complexity canonicalization guarantees that the 'and' is Op0.
1534 // (A & B) | ~(A | B) --> ~(A ^ B)
1535 // (A & B) | ~(B | A) --> ~(A ^ B)
1536 if (Op0->hasOneUse() || Op1->hasOneUse())
1537 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1538 match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B)))))
1539 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
1541 // (A & ~B) | (~A & B) --> A ^ B
1542 // (A & ~B) | (B & ~A) --> A ^ B
1543 // (~B & A) | (~A & B) --> A ^ B
1544 // (~B & A) | (B & ~A) --> A ^ B
1545 if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) &&
1546 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B))))
1547 return BinaryOperator::CreateXor(A, B);
1549 return nullptr;
1552 /// Return true if a constant shift amount is always less than the specified
1553 /// bit-width. If not, the shift could create poison in the narrower type.
1554 static bool canNarrowShiftAmt(Constant *C, unsigned BitWidth) {
1555 if (auto *ScalarC = dyn_cast<ConstantInt>(C))
1556 return ScalarC->getZExtValue() < BitWidth;
1558 if (C->getType()->isVectorTy()) {
1559 // Check each element of a constant vector.
1560 unsigned NumElts = C->getType()->getVectorNumElements();
1561 for (unsigned i = 0; i != NumElts; ++i) {
1562 Constant *Elt = C->getAggregateElement(i);
1563 if (!Elt)
1564 return false;
1565 if (isa<UndefValue>(Elt))
1566 continue;
1567 auto *CI = dyn_cast<ConstantInt>(Elt);
1568 if (!CI || CI->getZExtValue() >= BitWidth)
1569 return false;
1571 return true;
1574 // The constant is a constant expression or unknown.
1575 return false;
1578 /// Try to use narrower ops (sink zext ops) for an 'and' with binop operand and
1579 /// a common zext operand: and (binop (zext X), C), (zext X).
1580 Instruction *InstCombiner::narrowMaskedBinOp(BinaryOperator &And) {
1581 // This transform could also apply to {or, and, xor}, but there are better
1582 // folds for those cases, so we don't expect those patterns here. AShr is not
1583 // handled because it should always be transformed to LShr in this sequence.
1584 // The subtract transform is different because it has a constant on the left.
1585 // Add/mul commute the constant to RHS; sub with constant RHS becomes add.
1586 Value *Op0 = And.getOperand(0), *Op1 = And.getOperand(1);
1587 Constant *C;
1588 if (!match(Op0, m_OneUse(m_Add(m_Specific(Op1), m_Constant(C)))) &&
1589 !match(Op0, m_OneUse(m_Mul(m_Specific(Op1), m_Constant(C)))) &&
1590 !match(Op0, m_OneUse(m_LShr(m_Specific(Op1), m_Constant(C)))) &&
1591 !match(Op0, m_OneUse(m_Shl(m_Specific(Op1), m_Constant(C)))) &&
1592 !match(Op0, m_OneUse(m_Sub(m_Constant(C), m_Specific(Op1)))))
1593 return nullptr;
1595 Value *X;
1596 if (!match(Op1, m_ZExt(m_Value(X))) || Op1->hasNUsesOrMore(3))
1597 return nullptr;
1599 Type *Ty = And.getType();
1600 if (!isa<VectorType>(Ty) && !shouldChangeType(Ty, X->getType()))
1601 return nullptr;
1603 // If we're narrowing a shift, the shift amount must be safe (less than the
1604 // width) in the narrower type. If the shift amount is greater, instsimplify
1605 // usually handles that case, but we can't guarantee/assert it.
1606 Instruction::BinaryOps Opc = cast<BinaryOperator>(Op0)->getOpcode();
1607 if (Opc == Instruction::LShr || Opc == Instruction::Shl)
1608 if (!canNarrowShiftAmt(C, X->getType()->getScalarSizeInBits()))
1609 return nullptr;
1611 // and (sub C, (zext X)), (zext X) --> zext (and (sub C', X), X)
1612 // and (binop (zext X), C), (zext X) --> zext (and (binop X, C'), X)
1613 Value *NewC = ConstantExpr::getTrunc(C, X->getType());
1614 Value *NewBO = Opc == Instruction::Sub ? Builder.CreateBinOp(Opc, NewC, X)
1615 : Builder.CreateBinOp(Opc, X, NewC);
1616 return new ZExtInst(Builder.CreateAnd(NewBO, X), Ty);
1619 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
1620 // here. We should standardize that construct where it is needed or choose some
1621 // other way to ensure that commutated variants of patterns are not missed.
1622 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
1623 if (Value *V = SimplifyAndInst(I.getOperand(0), I.getOperand(1),
1624 SQ.getWithInstruction(&I)))
1625 return replaceInstUsesWith(I, V);
1627 if (SimplifyAssociativeOrCommutative(I))
1628 return &I;
1630 if (Instruction *X = foldVectorBinop(I))
1631 return X;
1633 // See if we can simplify any instructions used by the instruction whose sole
1634 // purpose is to compute bits we don't care about.
1635 if (SimplifyDemandedInstructionBits(I))
1636 return &I;
1638 // Do this before using distributive laws to catch simple and/or/not patterns.
1639 if (Instruction *Xor = foldAndToXor(I, Builder))
1640 return Xor;
1642 // (A|B)&(A|C) -> A|(B&C) etc
1643 if (Value *V = SimplifyUsingDistributiveLaws(I))
1644 return replaceInstUsesWith(I, V);
1646 if (Value *V = SimplifyBSwap(I, Builder))
1647 return replaceInstUsesWith(I, V);
1649 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1650 const APInt *C;
1651 if (match(Op1, m_APInt(C))) {
1652 Value *X, *Y;
1653 if (match(Op0, m_OneUse(m_LogicalShift(m_One(), m_Value(X)))) &&
1654 C->isOneValue()) {
1655 // (1 << X) & 1 --> zext(X == 0)
1656 // (1 >> X) & 1 --> zext(X == 0)
1657 Value *IsZero = Builder.CreateICmpEQ(X, ConstantInt::get(I.getType(), 0));
1658 return new ZExtInst(IsZero, I.getType());
1661 const APInt *XorC;
1662 if (match(Op0, m_OneUse(m_Xor(m_Value(X), m_APInt(XorC))))) {
1663 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
1664 Constant *NewC = ConstantInt::get(I.getType(), *C & *XorC);
1665 Value *And = Builder.CreateAnd(X, Op1);
1666 And->takeName(Op0);
1667 return BinaryOperator::CreateXor(And, NewC);
1670 const APInt *OrC;
1671 if (match(Op0, m_OneUse(m_Or(m_Value(X), m_APInt(OrC))))) {
1672 // (X | C1) & C2 --> (X & C2^(C1&C2)) | (C1&C2)
1673 // NOTE: This reduces the number of bits set in the & mask, which
1674 // can expose opportunities for store narrowing for scalars.
1675 // NOTE: SimplifyDemandedBits should have already removed bits from C1
1676 // that aren't set in C2. Meaning we can replace (C1&C2) with C1 in
1677 // above, but this feels safer.
1678 APInt Together = *C & *OrC;
1679 Value *And = Builder.CreateAnd(X, ConstantInt::get(I.getType(),
1680 Together ^ *C));
1681 And->takeName(Op0);
1682 return BinaryOperator::CreateOr(And, ConstantInt::get(I.getType(),
1683 Together));
1686 // If the mask is only needed on one incoming arm, push the 'and' op up.
1687 if (match(Op0, m_OneUse(m_Xor(m_Value(X), m_Value(Y)))) ||
1688 match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
1689 APInt NotAndMask(~(*C));
1690 BinaryOperator::BinaryOps BinOp = cast<BinaryOperator>(Op0)->getOpcode();
1691 if (MaskedValueIsZero(X, NotAndMask, 0, &I)) {
1692 // Not masking anything out for the LHS, move mask to RHS.
1693 // and ({x}or X, Y), C --> {x}or X, (and Y, C)
1694 Value *NewRHS = Builder.CreateAnd(Y, Op1, Y->getName() + ".masked");
1695 return BinaryOperator::Create(BinOp, X, NewRHS);
1697 if (!isa<Constant>(Y) && MaskedValueIsZero(Y, NotAndMask, 0, &I)) {
1698 // Not masking anything out for the RHS, move mask to LHS.
1699 // and ({x}or X, Y), C --> {x}or (and X, C), Y
1700 Value *NewLHS = Builder.CreateAnd(X, Op1, X->getName() + ".masked");
1701 return BinaryOperator::Create(BinOp, NewLHS, Y);
1707 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
1708 const APInt &AndRHSMask = AndRHS->getValue();
1710 // Optimize a variety of ((val OP C1) & C2) combinations...
1711 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1712 // ((C1 OP zext(X)) & C2) -> zext((C1-X) & C2) if C2 fits in the bitwidth
1713 // of X and OP behaves well when given trunc(C1) and X.
1714 // TODO: Do this for vectors by using m_APInt isntead of m_ConstantInt.
1715 switch (Op0I->getOpcode()) {
1716 default:
1717 break;
1718 case Instruction::Xor:
1719 case Instruction::Or:
1720 case Instruction::Mul:
1721 case Instruction::Add:
1722 case Instruction::Sub:
1723 Value *X;
1724 ConstantInt *C1;
1725 // TODO: The one use restrictions could be relaxed a little if the AND
1726 // is going to be removed.
1727 if (match(Op0I, m_OneUse(m_c_BinOp(m_OneUse(m_ZExt(m_Value(X))),
1728 m_ConstantInt(C1))))) {
1729 if (AndRHSMask.isIntN(X->getType()->getScalarSizeInBits())) {
1730 auto *TruncC1 = ConstantExpr::getTrunc(C1, X->getType());
1731 Value *BinOp;
1732 Value *Op0LHS = Op0I->getOperand(0);
1733 if (isa<ZExtInst>(Op0LHS))
1734 BinOp = Builder.CreateBinOp(Op0I->getOpcode(), X, TruncC1);
1735 else
1736 BinOp = Builder.CreateBinOp(Op0I->getOpcode(), TruncC1, X);
1737 auto *TruncC2 = ConstantExpr::getTrunc(AndRHS, X->getType());
1738 auto *And = Builder.CreateAnd(BinOp, TruncC2);
1739 return new ZExtInst(And, I.getType());
1744 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
1745 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
1746 return Res;
1749 // If this is an integer truncation, and if the source is an 'and' with
1750 // immediate, transform it. This frequently occurs for bitfield accesses.
1752 Value *X = nullptr; ConstantInt *YC = nullptr;
1753 if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) {
1754 // Change: and (trunc (and X, YC) to T), C2
1755 // into : and (trunc X to T), trunc(YC) & C2
1756 // This will fold the two constants together, which may allow
1757 // other simplifications.
1758 Value *NewCast = Builder.CreateTrunc(X, I.getType(), "and.shrunk");
1759 Constant *C3 = ConstantExpr::getTrunc(YC, I.getType());
1760 C3 = ConstantExpr::getAnd(C3, AndRHS);
1761 return BinaryOperator::CreateAnd(NewCast, C3);
1766 if (Instruction *Z = narrowMaskedBinOp(I))
1767 return Z;
1769 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))
1770 return FoldedLogic;
1772 if (Instruction *DeMorgan = matchDeMorgansLaws(I, Builder))
1773 return DeMorgan;
1776 Value *A, *B, *C;
1777 // A & (A ^ B) --> A & ~B
1778 if (match(Op1, m_OneUse(m_c_Xor(m_Specific(Op0), m_Value(B)))))
1779 return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(B));
1780 // (A ^ B) & A --> A & ~B
1781 if (match(Op0, m_OneUse(m_c_Xor(m_Specific(Op1), m_Value(B)))))
1782 return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(B));
1784 // (A ^ B) & ((B ^ C) ^ A) -> (A ^ B) & ~C
1785 if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
1786 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
1787 if (Op1->hasOneUse() || isFreeToInvert(C, C->hasOneUse()))
1788 return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(C));
1790 // ((A ^ C) ^ B) & (B ^ A) -> (B ^ A) & ~C
1791 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
1792 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
1793 if (Op0->hasOneUse() || isFreeToInvert(C, C->hasOneUse()))
1794 return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(C));
1796 // (A | B) & ((~A) ^ B) -> (A & B)
1797 // (A | B) & (B ^ (~A)) -> (A & B)
1798 // (B | A) & ((~A) ^ B) -> (A & B)
1799 // (B | A) & (B ^ (~A)) -> (A & B)
1800 if (match(Op1, m_c_Xor(m_Not(m_Value(A)), m_Value(B))) &&
1801 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1802 return BinaryOperator::CreateAnd(A, B);
1804 // ((~A) ^ B) & (A | B) -> (A & B)
1805 // ((~A) ^ B) & (B | A) -> (A & B)
1806 // (B ^ (~A)) & (A | B) -> (A & B)
1807 // (B ^ (~A)) & (B | A) -> (A & B)
1808 if (match(Op0, m_c_Xor(m_Not(m_Value(A)), m_Value(B))) &&
1809 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
1810 return BinaryOperator::CreateAnd(A, B);
1814 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
1815 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
1816 if (LHS && RHS)
1817 if (Value *Res = foldAndOfICmps(LHS, RHS, I))
1818 return replaceInstUsesWith(I, Res);
1820 // TODO: Make this recursive; it's a little tricky because an arbitrary
1821 // number of 'and' instructions might have to be created.
1822 Value *X, *Y;
1823 if (LHS && match(Op1, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
1824 if (auto *Cmp = dyn_cast<ICmpInst>(X))
1825 if (Value *Res = foldAndOfICmps(LHS, Cmp, I))
1826 return replaceInstUsesWith(I, Builder.CreateAnd(Res, Y));
1827 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
1828 if (Value *Res = foldAndOfICmps(LHS, Cmp, I))
1829 return replaceInstUsesWith(I, Builder.CreateAnd(Res, X));
1831 if (RHS && match(Op0, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
1832 if (auto *Cmp = dyn_cast<ICmpInst>(X))
1833 if (Value *Res = foldAndOfICmps(Cmp, RHS, I))
1834 return replaceInstUsesWith(I, Builder.CreateAnd(Res, Y));
1835 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
1836 if (Value *Res = foldAndOfICmps(Cmp, RHS, I))
1837 return replaceInstUsesWith(I, Builder.CreateAnd(Res, X));
1841 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1842 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1843 if (Value *Res = foldLogicOfFCmps(LHS, RHS, true))
1844 return replaceInstUsesWith(I, Res);
1846 if (Instruction *FoldedFCmps = reassociateFCmps(I, Builder))
1847 return FoldedFCmps;
1849 if (Instruction *CastedAnd = foldCastedBitwiseLogic(I))
1850 return CastedAnd;
1852 // and(sext(A), B) / and(B, sext(A)) --> A ? B : 0, where A is i1 or <N x i1>.
1853 Value *A;
1854 if (match(Op0, m_OneUse(m_SExt(m_Value(A)))) &&
1855 A->getType()->isIntOrIntVectorTy(1))
1856 return SelectInst::Create(A, Op1, Constant::getNullValue(I.getType()));
1857 if (match(Op1, m_OneUse(m_SExt(m_Value(A)))) &&
1858 A->getType()->isIntOrIntVectorTy(1))
1859 return SelectInst::Create(A, Op0, Constant::getNullValue(I.getType()));
1861 return nullptr;
1864 Instruction *InstCombiner::matchBSwap(BinaryOperator &Or) {
1865 assert(Or.getOpcode() == Instruction::Or && "bswap requires an 'or'");
1866 Value *Op0 = Or.getOperand(0), *Op1 = Or.getOperand(1);
1868 // Look through zero extends.
1869 if (Instruction *Ext = dyn_cast<ZExtInst>(Op0))
1870 Op0 = Ext->getOperand(0);
1872 if (Instruction *Ext = dyn_cast<ZExtInst>(Op1))
1873 Op1 = Ext->getOperand(0);
1875 // (A | B) | C and A | (B | C) -> bswap if possible.
1876 bool OrOfOrs = match(Op0, m_Or(m_Value(), m_Value())) ||
1877 match(Op1, m_Or(m_Value(), m_Value()));
1879 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
1880 bool OrOfShifts = match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
1881 match(Op1, m_LogicalShift(m_Value(), m_Value()));
1883 // (A & B) | (C & D) -> bswap if possible.
1884 bool OrOfAnds = match(Op0, m_And(m_Value(), m_Value())) &&
1885 match(Op1, m_And(m_Value(), m_Value()));
1887 // (A << B) | (C & D) -> bswap if possible.
1888 // The bigger pattern here is ((A & C1) << C2) | ((B >> C2) & C1), which is a
1889 // part of the bswap idiom for specific values of C1, C2 (e.g. C1 = 16711935,
1890 // C2 = 8 for i32).
1891 // This pattern can occur when the operands of the 'or' are not canonicalized
1892 // for some reason (not having only one use, for example).
1893 bool OrOfAndAndSh = (match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
1894 match(Op1, m_And(m_Value(), m_Value()))) ||
1895 (match(Op0, m_And(m_Value(), m_Value())) &&
1896 match(Op1, m_LogicalShift(m_Value(), m_Value())));
1898 if (!OrOfOrs && !OrOfShifts && !OrOfAnds && !OrOfAndAndSh)
1899 return nullptr;
1901 SmallVector<Instruction*, 4> Insts;
1902 if (!recognizeBSwapOrBitReverseIdiom(&Or, true, false, Insts))
1903 return nullptr;
1904 Instruction *LastInst = Insts.pop_back_val();
1905 LastInst->removeFromParent();
1907 for (auto *Inst : Insts)
1908 Worklist.Add(Inst);
1909 return LastInst;
1912 /// Transform UB-safe variants of bitwise rotate to the funnel shift intrinsic.
1913 static Instruction *matchRotate(Instruction &Or) {
1914 // TODO: Can we reduce the code duplication between this and the related
1915 // rotate matching code under visitSelect and visitTrunc?
1916 unsigned Width = Or.getType()->getScalarSizeInBits();
1917 if (!isPowerOf2_32(Width))
1918 return nullptr;
1920 // First, find an or'd pair of opposite shifts with the same shifted operand:
1921 // or (lshr ShVal, ShAmt0), (shl ShVal, ShAmt1)
1922 BinaryOperator *Or0, *Or1;
1923 if (!match(Or.getOperand(0), m_BinOp(Or0)) ||
1924 !match(Or.getOperand(1), m_BinOp(Or1)))
1925 return nullptr;
1927 Value *ShVal, *ShAmt0, *ShAmt1;
1928 if (!match(Or0, m_OneUse(m_LogicalShift(m_Value(ShVal), m_Value(ShAmt0)))) ||
1929 !match(Or1, m_OneUse(m_LogicalShift(m_Specific(ShVal), m_Value(ShAmt1)))))
1930 return nullptr;
1932 BinaryOperator::BinaryOps ShiftOpcode0 = Or0->getOpcode();
1933 BinaryOperator::BinaryOps ShiftOpcode1 = Or1->getOpcode();
1934 if (ShiftOpcode0 == ShiftOpcode1)
1935 return nullptr;
1937 // Match the shift amount operands for a rotate pattern. This always matches
1938 // a subtraction on the R operand.
1939 auto matchShiftAmount = [](Value *L, Value *R, unsigned Width) -> Value * {
1940 // The shift amount may be masked with negation:
1941 // (shl ShVal, (X & (Width - 1))) | (lshr ShVal, ((-X) & (Width - 1)))
1942 Value *X;
1943 unsigned Mask = Width - 1;
1944 if (match(L, m_And(m_Value(X), m_SpecificInt(Mask))) &&
1945 match(R, m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask))))
1946 return X;
1948 // Similar to above, but the shift amount may be extended after masking,
1949 // so return the extended value as the parameter for the intrinsic.
1950 if (match(L, m_ZExt(m_And(m_Value(X), m_SpecificInt(Mask)))) &&
1951 match(R, m_And(m_Neg(m_ZExt(m_And(m_Specific(X), m_SpecificInt(Mask)))),
1952 m_SpecificInt(Mask))))
1953 return L;
1955 return nullptr;
1958 Value *ShAmt = matchShiftAmount(ShAmt0, ShAmt1, Width);
1959 bool SubIsOnLHS = false;
1960 if (!ShAmt) {
1961 ShAmt = matchShiftAmount(ShAmt1, ShAmt0, Width);
1962 SubIsOnLHS = true;
1964 if (!ShAmt)
1965 return nullptr;
1967 bool IsFshl = (!SubIsOnLHS && ShiftOpcode0 == BinaryOperator::Shl) ||
1968 (SubIsOnLHS && ShiftOpcode1 == BinaryOperator::Shl);
1969 Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr;
1970 Function *F = Intrinsic::getDeclaration(Or.getModule(), IID, Or.getType());
1971 return IntrinsicInst::Create(F, { ShVal, ShVal, ShAmt });
1974 /// If all elements of two constant vectors are 0/-1 and inverses, return true.
1975 static bool areInverseVectorBitmasks(Constant *C1, Constant *C2) {
1976 unsigned NumElts = C1->getType()->getVectorNumElements();
1977 for (unsigned i = 0; i != NumElts; ++i) {
1978 Constant *EltC1 = C1->getAggregateElement(i);
1979 Constant *EltC2 = C2->getAggregateElement(i);
1980 if (!EltC1 || !EltC2)
1981 return false;
1983 // One element must be all ones, and the other must be all zeros.
1984 if (!((match(EltC1, m_Zero()) && match(EltC2, m_AllOnes())) ||
1985 (match(EltC2, m_Zero()) && match(EltC1, m_AllOnes()))))
1986 return false;
1988 return true;
1991 /// We have an expression of the form (A & C) | (B & D). If A is a scalar or
1992 /// vector composed of all-zeros or all-ones values and is the bitwise 'not' of
1993 /// B, it can be used as the condition operand of a select instruction.
1994 Value *InstCombiner::getSelectCondition(Value *A, Value *B) {
1995 // Step 1: We may have peeked through bitcasts in the caller.
1996 // Exit immediately if we don't have (vector) integer types.
1997 Type *Ty = A->getType();
1998 if (!Ty->isIntOrIntVectorTy() || !B->getType()->isIntOrIntVectorTy())
1999 return nullptr;
2001 // Step 2: We need 0 or all-1's bitmasks.
2002 if (ComputeNumSignBits(A) != Ty->getScalarSizeInBits())
2003 return nullptr;
2005 // Step 3: If B is the 'not' value of A, we have our answer.
2006 if (match(A, m_Not(m_Specific(B)))) {
2007 // If these are scalars or vectors of i1, A can be used directly.
2008 if (Ty->isIntOrIntVectorTy(1))
2009 return A;
2010 return Builder.CreateTrunc(A, CmpInst::makeCmpResultType(Ty));
2013 // If both operands are constants, see if the constants are inverse bitmasks.
2014 Constant *AConst, *BConst;
2015 if (match(A, m_Constant(AConst)) && match(B, m_Constant(BConst)))
2016 if (AConst == ConstantExpr::getNot(BConst))
2017 return Builder.CreateZExtOrTrunc(A, CmpInst::makeCmpResultType(Ty));
2019 // Look for more complex patterns. The 'not' op may be hidden behind various
2020 // casts. Look through sexts and bitcasts to find the booleans.
2021 Value *Cond;
2022 Value *NotB;
2023 if (match(A, m_SExt(m_Value(Cond))) &&
2024 Cond->getType()->isIntOrIntVectorTy(1) &&
2025 match(B, m_OneUse(m_Not(m_Value(NotB))))) {
2026 NotB = peekThroughBitcast(NotB, true);
2027 if (match(NotB, m_SExt(m_Specific(Cond))))
2028 return Cond;
2031 // All scalar (and most vector) possibilities should be handled now.
2032 // Try more matches that only apply to non-splat constant vectors.
2033 if (!Ty->isVectorTy())
2034 return nullptr;
2036 // If both operands are xor'd with constants using the same sexted boolean
2037 // operand, see if the constants are inverse bitmasks.
2038 // TODO: Use ConstantExpr::getNot()?
2039 if (match(A, (m_Xor(m_SExt(m_Value(Cond)), m_Constant(AConst)))) &&
2040 match(B, (m_Xor(m_SExt(m_Specific(Cond)), m_Constant(BConst)))) &&
2041 Cond->getType()->isIntOrIntVectorTy(1) &&
2042 areInverseVectorBitmasks(AConst, BConst)) {
2043 AConst = ConstantExpr::getTrunc(AConst, CmpInst::makeCmpResultType(Ty));
2044 return Builder.CreateXor(Cond, AConst);
2046 return nullptr;
2049 /// We have an expression of the form (A & C) | (B & D). Try to simplify this
2050 /// to "A' ? C : D", where A' is a boolean or vector of booleans.
2051 Value *InstCombiner::matchSelectFromAndOr(Value *A, Value *C, Value *B,
2052 Value *D) {
2053 // The potential condition of the select may be bitcasted. In that case, look
2054 // through its bitcast and the corresponding bitcast of the 'not' condition.
2055 Type *OrigType = A->getType();
2056 A = peekThroughBitcast(A, true);
2057 B = peekThroughBitcast(B, true);
2058 if (Value *Cond = getSelectCondition(A, B)) {
2059 // ((bc Cond) & C) | ((bc ~Cond) & D) --> bc (select Cond, (bc C), (bc D))
2060 // The bitcasts will either all exist or all not exist. The builder will
2061 // not create unnecessary casts if the types already match.
2062 Value *BitcastC = Builder.CreateBitCast(C, A->getType());
2063 Value *BitcastD = Builder.CreateBitCast(D, A->getType());
2064 Value *Select = Builder.CreateSelect(Cond, BitcastC, BitcastD);
2065 return Builder.CreateBitCast(Select, OrigType);
2068 return nullptr;
2071 /// Fold (icmp)|(icmp) if possible.
2072 Value *InstCombiner::foldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS,
2073 Instruction &CxtI) {
2074 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2)
2075 // if K1 and K2 are a one-bit mask.
2076 if (Value *V = foldAndOrOfICmpsOfAndWithPow2(LHS, RHS, false, CxtI))
2077 return V;
2079 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
2081 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHS->getOperand(1));
2082 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS->getOperand(1));
2084 // Fold (icmp ult/ule (A + C1), C3) | (icmp ult/ule (A + C2), C3)
2085 // --> (icmp ult/ule ((A & ~(C1 ^ C2)) + max(C1, C2)), C3)
2086 // The original condition actually refers to the following two ranges:
2087 // [MAX_UINT-C1+1, MAX_UINT-C1+1+C3] and [MAX_UINT-C2+1, MAX_UINT-C2+1+C3]
2088 // We can fold these two ranges if:
2089 // 1) C1 and C2 is unsigned greater than C3.
2090 // 2) The two ranges are separated.
2091 // 3) C1 ^ C2 is one-bit mask.
2092 // 4) LowRange1 ^ LowRange2 and HighRange1 ^ HighRange2 are one-bit mask.
2093 // This implies all values in the two ranges differ by exactly one bit.
2095 if ((PredL == ICmpInst::ICMP_ULT || PredL == ICmpInst::ICMP_ULE) &&
2096 PredL == PredR && LHSC && RHSC && LHS->hasOneUse() && RHS->hasOneUse() &&
2097 LHSC->getType() == RHSC->getType() &&
2098 LHSC->getValue() == (RHSC->getValue())) {
2100 Value *LAdd = LHS->getOperand(0);
2101 Value *RAdd = RHS->getOperand(0);
2103 Value *LAddOpnd, *RAddOpnd;
2104 ConstantInt *LAddC, *RAddC;
2105 if (match(LAdd, m_Add(m_Value(LAddOpnd), m_ConstantInt(LAddC))) &&
2106 match(RAdd, m_Add(m_Value(RAddOpnd), m_ConstantInt(RAddC))) &&
2107 LAddC->getValue().ugt(LHSC->getValue()) &&
2108 RAddC->getValue().ugt(LHSC->getValue())) {
2110 APInt DiffC = LAddC->getValue() ^ RAddC->getValue();
2111 if (LAddOpnd == RAddOpnd && DiffC.isPowerOf2()) {
2112 ConstantInt *MaxAddC = nullptr;
2113 if (LAddC->getValue().ult(RAddC->getValue()))
2114 MaxAddC = RAddC;
2115 else
2116 MaxAddC = LAddC;
2118 APInt RRangeLow = -RAddC->getValue();
2119 APInt RRangeHigh = RRangeLow + LHSC->getValue();
2120 APInt LRangeLow = -LAddC->getValue();
2121 APInt LRangeHigh = LRangeLow + LHSC->getValue();
2122 APInt LowRangeDiff = RRangeLow ^ LRangeLow;
2123 APInt HighRangeDiff = RRangeHigh ^ LRangeHigh;
2124 APInt RangeDiff = LRangeLow.sgt(RRangeLow) ? LRangeLow - RRangeLow
2125 : RRangeLow - LRangeLow;
2127 if (LowRangeDiff.isPowerOf2() && LowRangeDiff == HighRangeDiff &&
2128 RangeDiff.ugt(LHSC->getValue())) {
2129 Value *MaskC = ConstantInt::get(LAddC->getType(), ~DiffC);
2131 Value *NewAnd = Builder.CreateAnd(LAddOpnd, MaskC);
2132 Value *NewAdd = Builder.CreateAdd(NewAnd, MaxAddC);
2133 return Builder.CreateICmp(LHS->getPredicate(), NewAdd, LHSC);
2139 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
2140 if (predicatesFoldable(PredL, PredR)) {
2141 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2142 LHS->getOperand(1) == RHS->getOperand(0))
2143 LHS->swapOperands();
2144 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2145 LHS->getOperand(1) == RHS->getOperand(1)) {
2146 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2147 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
2148 bool IsSigned = LHS->isSigned() || RHS->isSigned();
2149 return getNewICmpValue(Code, IsSigned, Op0, Op1, Builder);
2153 // handle (roughly):
2154 // (icmp ne (A & B), C) | (icmp ne (A & D), E)
2155 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, false, Builder))
2156 return V;
2158 Value *LHS0 = LHS->getOperand(0), *RHS0 = RHS->getOperand(0);
2159 if (LHS->hasOneUse() || RHS->hasOneUse()) {
2160 // (icmp eq B, 0) | (icmp ult A, B) -> (icmp ule A, B-1)
2161 // (icmp eq B, 0) | (icmp ugt B, A) -> (icmp ule A, B-1)
2162 Value *A = nullptr, *B = nullptr;
2163 if (PredL == ICmpInst::ICMP_EQ && LHSC && LHSC->isZero()) {
2164 B = LHS0;
2165 if (PredR == ICmpInst::ICMP_ULT && LHS0 == RHS->getOperand(1))
2166 A = RHS0;
2167 else if (PredR == ICmpInst::ICMP_UGT && LHS0 == RHS0)
2168 A = RHS->getOperand(1);
2170 // (icmp ult A, B) | (icmp eq B, 0) -> (icmp ule A, B-1)
2171 // (icmp ugt B, A) | (icmp eq B, 0) -> (icmp ule A, B-1)
2172 else if (PredR == ICmpInst::ICMP_EQ && RHSC && RHSC->isZero()) {
2173 B = RHS0;
2174 if (PredL == ICmpInst::ICMP_ULT && RHS0 == LHS->getOperand(1))
2175 A = LHS0;
2176 else if (PredL == ICmpInst::ICMP_UGT && LHS0 == RHS0)
2177 A = LHS->getOperand(1);
2179 if (A && B)
2180 return Builder.CreateICmp(
2181 ICmpInst::ICMP_UGE,
2182 Builder.CreateAdd(B, ConstantInt::getSigned(B->getType(), -1)), A);
2185 // E.g. (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
2186 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/true))
2187 return V;
2189 // E.g. (icmp sgt x, n) | (icmp slt x, 0) --> icmp ugt x, n
2190 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/true))
2191 return V;
2193 if (Value *V = foldAndOrOfEqualityCmpsWithConstants(LHS, RHS, false, Builder))
2194 return V;
2196 if (Value *V = foldIsPowerOf2(LHS, RHS, false /* JoinedByAnd */, Builder))
2197 return V;
2199 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
2200 if (!LHSC || !RHSC)
2201 return nullptr;
2203 if (LHSC == RHSC && PredL == PredR) {
2204 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
2205 if (PredL == ICmpInst::ICMP_NE && LHSC->isZero()) {
2206 Value *NewOr = Builder.CreateOr(LHS0, RHS0);
2207 return Builder.CreateICmp(PredL, NewOr, LHSC);
2211 // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
2212 // iff C2 + CA == C1.
2213 if (PredL == ICmpInst::ICMP_ULT && PredR == ICmpInst::ICMP_EQ) {
2214 ConstantInt *AddC;
2215 if (match(LHS0, m_Add(m_Specific(RHS0), m_ConstantInt(AddC))))
2216 if (RHSC->getValue() + AddC->getValue() == LHSC->getValue())
2217 return Builder.CreateICmpULE(LHS0, LHSC);
2220 // From here on, we only handle:
2221 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
2222 if (LHS0 != RHS0)
2223 return nullptr;
2225 // ICMP_[US][GL]E X, C is folded to ICMP_[US][GL]T elsewhere.
2226 if (PredL == ICmpInst::ICMP_UGE || PredL == ICmpInst::ICMP_ULE ||
2227 PredR == ICmpInst::ICMP_UGE || PredR == ICmpInst::ICMP_ULE ||
2228 PredL == ICmpInst::ICMP_SGE || PredL == ICmpInst::ICMP_SLE ||
2229 PredR == ICmpInst::ICMP_SGE || PredR == ICmpInst::ICMP_SLE)
2230 return nullptr;
2232 // We can't fold (ugt x, C) | (sgt x, C2).
2233 if (!predicatesFoldable(PredL, PredR))
2234 return nullptr;
2236 // Ensure that the larger constant is on the RHS.
2237 bool ShouldSwap;
2238 if (CmpInst::isSigned(PredL) ||
2239 (ICmpInst::isEquality(PredL) && CmpInst::isSigned(PredR)))
2240 ShouldSwap = LHSC->getValue().sgt(RHSC->getValue());
2241 else
2242 ShouldSwap = LHSC->getValue().ugt(RHSC->getValue());
2244 if (ShouldSwap) {
2245 std::swap(LHS, RHS);
2246 std::swap(LHSC, RHSC);
2247 std::swap(PredL, PredR);
2250 // At this point, we know we have two icmp instructions
2251 // comparing a value against two constants and or'ing the result
2252 // together. Because of the above check, we know that we only have
2253 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
2254 // icmp folding check above), that the two constants are not
2255 // equal.
2256 assert(LHSC != RHSC && "Compares not folded above?");
2258 switch (PredL) {
2259 default:
2260 llvm_unreachable("Unknown integer condition code!");
2261 case ICmpInst::ICMP_EQ:
2262 switch (PredR) {
2263 default:
2264 llvm_unreachable("Unknown integer condition code!");
2265 case ICmpInst::ICMP_EQ:
2266 // Potential folds for this case should already be handled.
2267 break;
2268 case ICmpInst::ICMP_UGT:
2269 // (X == 0 || X u> C) -> (X-1) u>= C
2270 if (LHSC->isMinValue(false))
2271 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue() + 1,
2272 false, false);
2273 // (X == 13 | X u> 14) -> no change
2274 break;
2275 case ICmpInst::ICMP_SGT:
2276 // (X == INT_MIN || X s> C) -> (X-(INT_MIN+1)) u>= C-INT_MIN
2277 if (LHSC->isMinValue(true))
2278 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue() + 1,
2279 true, false);
2280 // (X == 13 | X s> 14) -> no change
2281 break;
2283 break;
2284 case ICmpInst::ICMP_ULT:
2285 switch (PredR) {
2286 default:
2287 llvm_unreachable("Unknown integer condition code!");
2288 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
2289 // (X u< C || X == UINT_MAX) => (X-C) u>= UINT_MAX-C
2290 if (RHSC->isMaxValue(false))
2291 return insertRangeTest(LHS0, LHSC->getValue(), RHSC->getValue(),
2292 false, false);
2293 break;
2294 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
2295 assert(!RHSC->isMaxValue(false) && "Missed icmp simplification");
2296 return insertRangeTest(LHS0, LHSC->getValue(), RHSC->getValue() + 1,
2297 false, false);
2299 break;
2300 case ICmpInst::ICMP_SLT:
2301 switch (PredR) {
2302 default:
2303 llvm_unreachable("Unknown integer condition code!");
2304 case ICmpInst::ICMP_EQ:
2305 // (X s< C || X == INT_MAX) => (X-C) u>= INT_MAX-C
2306 if (RHSC->isMaxValue(true))
2307 return insertRangeTest(LHS0, LHSC->getValue(), RHSC->getValue(),
2308 true, false);
2309 // (X s< 13 | X == 14) -> no change
2310 break;
2311 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) u> 2
2312 assert(!RHSC->isMaxValue(true) && "Missed icmp simplification");
2313 return insertRangeTest(LHS0, LHSC->getValue(), RHSC->getValue() + 1, true,
2314 false);
2316 break;
2318 return nullptr;
2321 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
2322 // here. We should standardize that construct where it is needed or choose some
2323 // other way to ensure that commutated variants of patterns are not missed.
2324 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
2325 if (Value *V = SimplifyOrInst(I.getOperand(0), I.getOperand(1),
2326 SQ.getWithInstruction(&I)))
2327 return replaceInstUsesWith(I, V);
2329 if (SimplifyAssociativeOrCommutative(I))
2330 return &I;
2332 if (Instruction *X = foldVectorBinop(I))
2333 return X;
2335 // See if we can simplify any instructions used by the instruction whose sole
2336 // purpose is to compute bits we don't care about.
2337 if (SimplifyDemandedInstructionBits(I))
2338 return &I;
2340 // Do this before using distributive laws to catch simple and/or/not patterns.
2341 if (Instruction *Xor = foldOrToXor(I, Builder))
2342 return Xor;
2344 // (A&B)|(A&C) -> A&(B|C) etc
2345 if (Value *V = SimplifyUsingDistributiveLaws(I))
2346 return replaceInstUsesWith(I, V);
2348 if (Value *V = SimplifyBSwap(I, Builder))
2349 return replaceInstUsesWith(I, V);
2351 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))
2352 return FoldedLogic;
2354 if (Instruction *BSwap = matchBSwap(I))
2355 return BSwap;
2357 if (Instruction *Rotate = matchRotate(I))
2358 return Rotate;
2360 Value *X, *Y;
2361 const APInt *CV;
2362 if (match(&I, m_c_Or(m_OneUse(m_Xor(m_Value(X), m_APInt(CV))), m_Value(Y))) &&
2363 !CV->isAllOnesValue() && MaskedValueIsZero(Y, *CV, 0, &I)) {
2364 // (X ^ C) | Y -> (X | Y) ^ C iff Y & C == 0
2365 // The check for a 'not' op is for efficiency (if Y is known zero --> ~X).
2366 Value *Or = Builder.CreateOr(X, Y);
2367 return BinaryOperator::CreateXor(Or, ConstantInt::get(I.getType(), *CV));
2370 // (A & C)|(B & D)
2371 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2372 Value *A, *B, *C, *D;
2373 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
2374 match(Op1, m_And(m_Value(B), m_Value(D)))) {
2375 ConstantInt *C1 = dyn_cast<ConstantInt>(C);
2376 ConstantInt *C2 = dyn_cast<ConstantInt>(D);
2377 if (C1 && C2) { // (A & C1)|(B & C2)
2378 Value *V1 = nullptr, *V2 = nullptr;
2379 if ((C1->getValue() & C2->getValue()).isNullValue()) {
2380 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
2381 // iff (C1&C2) == 0 and (N&~C1) == 0
2382 if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
2383 ((V1 == B &&
2384 MaskedValueIsZero(V2, ~C1->getValue(), 0, &I)) || // (V|N)
2385 (V2 == B &&
2386 MaskedValueIsZero(V1, ~C1->getValue(), 0, &I)))) // (N|V)
2387 return BinaryOperator::CreateAnd(A,
2388 Builder.getInt(C1->getValue()|C2->getValue()));
2389 // Or commutes, try both ways.
2390 if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
2391 ((V1 == A &&
2392 MaskedValueIsZero(V2, ~C2->getValue(), 0, &I)) || // (V|N)
2393 (V2 == A &&
2394 MaskedValueIsZero(V1, ~C2->getValue(), 0, &I)))) // (N|V)
2395 return BinaryOperator::CreateAnd(B,
2396 Builder.getInt(C1->getValue()|C2->getValue()));
2398 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
2399 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
2400 ConstantInt *C3 = nullptr, *C4 = nullptr;
2401 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
2402 (C3->getValue() & ~C1->getValue()).isNullValue() &&
2403 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
2404 (C4->getValue() & ~C2->getValue()).isNullValue()) {
2405 V2 = Builder.CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
2406 return BinaryOperator::CreateAnd(V2,
2407 Builder.getInt(C1->getValue()|C2->getValue()));
2411 if (C1->getValue() == ~C2->getValue()) {
2412 Value *X;
2414 // ((X|B)&C1)|(B&C2) -> (X&C1) | B iff C1 == ~C2
2415 if (match(A, m_c_Or(m_Value(X), m_Specific(B))))
2416 return BinaryOperator::CreateOr(Builder.CreateAnd(X, C1), B);
2417 // (A&C2)|((X|A)&C1) -> (X&C2) | A iff C1 == ~C2
2418 if (match(B, m_c_Or(m_Specific(A), m_Value(X))))
2419 return BinaryOperator::CreateOr(Builder.CreateAnd(X, C2), A);
2421 // ((X^B)&C1)|(B&C2) -> (X&C1) ^ B iff C1 == ~C2
2422 if (match(A, m_c_Xor(m_Value(X), m_Specific(B))))
2423 return BinaryOperator::CreateXor(Builder.CreateAnd(X, C1), B);
2424 // (A&C2)|((X^A)&C1) -> (X&C2) ^ A iff C1 == ~C2
2425 if (match(B, m_c_Xor(m_Specific(A), m_Value(X))))
2426 return BinaryOperator::CreateXor(Builder.CreateAnd(X, C2), A);
2430 // Don't try to form a select if it's unlikely that we'll get rid of at
2431 // least one of the operands. A select is generally more expensive than the
2432 // 'or' that it is replacing.
2433 if (Op0->hasOneUse() || Op1->hasOneUse()) {
2434 // (Cond & C) | (~Cond & D) -> Cond ? C : D, and commuted variants.
2435 if (Value *V = matchSelectFromAndOr(A, C, B, D))
2436 return replaceInstUsesWith(I, V);
2437 if (Value *V = matchSelectFromAndOr(A, C, D, B))
2438 return replaceInstUsesWith(I, V);
2439 if (Value *V = matchSelectFromAndOr(C, A, B, D))
2440 return replaceInstUsesWith(I, V);
2441 if (Value *V = matchSelectFromAndOr(C, A, D, B))
2442 return replaceInstUsesWith(I, V);
2443 if (Value *V = matchSelectFromAndOr(B, D, A, C))
2444 return replaceInstUsesWith(I, V);
2445 if (Value *V = matchSelectFromAndOr(B, D, C, A))
2446 return replaceInstUsesWith(I, V);
2447 if (Value *V = matchSelectFromAndOr(D, B, A, C))
2448 return replaceInstUsesWith(I, V);
2449 if (Value *V = matchSelectFromAndOr(D, B, C, A))
2450 return replaceInstUsesWith(I, V);
2454 // (A ^ B) | ((B ^ C) ^ A) -> (A ^ B) | C
2455 if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
2456 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
2457 return BinaryOperator::CreateOr(Op0, C);
2459 // ((A ^ C) ^ B) | (B ^ A) -> (B ^ A) | C
2460 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
2461 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
2462 return BinaryOperator::CreateOr(Op1, C);
2464 // ((B | C) & A) | B -> B | (A & C)
2465 if (match(Op0, m_And(m_Or(m_Specific(Op1), m_Value(C)), m_Value(A))))
2466 return BinaryOperator::CreateOr(Op1, Builder.CreateAnd(A, C));
2468 if (Instruction *DeMorgan = matchDeMorgansLaws(I, Builder))
2469 return DeMorgan;
2471 // Canonicalize xor to the RHS.
2472 bool SwappedForXor = false;
2473 if (match(Op0, m_Xor(m_Value(), m_Value()))) {
2474 std::swap(Op0, Op1);
2475 SwappedForXor = true;
2478 // A | ( A ^ B) -> A | B
2479 // A | (~A ^ B) -> A | ~B
2480 // (A & B) | (A ^ B)
2481 if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
2482 if (Op0 == A || Op0 == B)
2483 return BinaryOperator::CreateOr(A, B);
2485 if (match(Op0, m_And(m_Specific(A), m_Specific(B))) ||
2486 match(Op0, m_And(m_Specific(B), m_Specific(A))))
2487 return BinaryOperator::CreateOr(A, B);
2489 if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
2490 Value *Not = Builder.CreateNot(B, B->getName() + ".not");
2491 return BinaryOperator::CreateOr(Not, Op0);
2493 if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
2494 Value *Not = Builder.CreateNot(A, A->getName() + ".not");
2495 return BinaryOperator::CreateOr(Not, Op0);
2499 // A | ~(A | B) -> A | ~B
2500 // A | ~(A ^ B) -> A | ~B
2501 if (match(Op1, m_Not(m_Value(A))))
2502 if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
2503 if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
2504 Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
2505 B->getOpcode() == Instruction::Xor)) {
2506 Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
2507 B->getOperand(0);
2508 Value *Not = Builder.CreateNot(NotOp, NotOp->getName() + ".not");
2509 return BinaryOperator::CreateOr(Not, Op0);
2512 if (SwappedForXor)
2513 std::swap(Op0, Op1);
2516 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
2517 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
2518 if (LHS && RHS)
2519 if (Value *Res = foldOrOfICmps(LHS, RHS, I))
2520 return replaceInstUsesWith(I, Res);
2522 // TODO: Make this recursive; it's a little tricky because an arbitrary
2523 // number of 'or' instructions might have to be created.
2524 Value *X, *Y;
2525 if (LHS && match(Op1, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2526 if (auto *Cmp = dyn_cast<ICmpInst>(X))
2527 if (Value *Res = foldOrOfICmps(LHS, Cmp, I))
2528 return replaceInstUsesWith(I, Builder.CreateOr(Res, Y));
2529 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2530 if (Value *Res = foldOrOfICmps(LHS, Cmp, I))
2531 return replaceInstUsesWith(I, Builder.CreateOr(Res, X));
2533 if (RHS && match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2534 if (auto *Cmp = dyn_cast<ICmpInst>(X))
2535 if (Value *Res = foldOrOfICmps(Cmp, RHS, I))
2536 return replaceInstUsesWith(I, Builder.CreateOr(Res, Y));
2537 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2538 if (Value *Res = foldOrOfICmps(Cmp, RHS, I))
2539 return replaceInstUsesWith(I, Builder.CreateOr(Res, X));
2543 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
2544 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
2545 if (Value *Res = foldLogicOfFCmps(LHS, RHS, false))
2546 return replaceInstUsesWith(I, Res);
2548 if (Instruction *FoldedFCmps = reassociateFCmps(I, Builder))
2549 return FoldedFCmps;
2551 if (Instruction *CastedOr = foldCastedBitwiseLogic(I))
2552 return CastedOr;
2554 // or(sext(A), B) / or(B, sext(A)) --> A ? -1 : B, where A is i1 or <N x i1>.
2555 if (match(Op0, m_OneUse(m_SExt(m_Value(A)))) &&
2556 A->getType()->isIntOrIntVectorTy(1))
2557 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1);
2558 if (match(Op1, m_OneUse(m_SExt(m_Value(A)))) &&
2559 A->getType()->isIntOrIntVectorTy(1))
2560 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0);
2562 // Note: If we've gotten to the point of visiting the outer OR, then the
2563 // inner one couldn't be simplified. If it was a constant, then it won't
2564 // be simplified by a later pass either, so we try swapping the inner/outer
2565 // ORs in the hopes that we'll be able to simplify it this way.
2566 // (X|C) | V --> (X|V) | C
2567 ConstantInt *CI;
2568 if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
2569 match(Op0, m_Or(m_Value(A), m_ConstantInt(CI)))) {
2570 Value *Inner = Builder.CreateOr(A, Op1);
2571 Inner->takeName(Op0);
2572 return BinaryOperator::CreateOr(Inner, CI);
2575 // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D))
2576 // Since this OR statement hasn't been optimized further yet, we hope
2577 // that this transformation will allow the new ORs to be optimized.
2579 Value *X = nullptr, *Y = nullptr;
2580 if (Op0->hasOneUse() && Op1->hasOneUse() &&
2581 match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) &&
2582 match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) {
2583 Value *orTrue = Builder.CreateOr(A, C);
2584 Value *orFalse = Builder.CreateOr(B, D);
2585 return SelectInst::Create(X, orTrue, orFalse);
2589 return nullptr;
2592 /// A ^ B can be specified using other logic ops in a variety of patterns. We
2593 /// can fold these early and efficiently by morphing an existing instruction.
2594 static Instruction *foldXorToXor(BinaryOperator &I,
2595 InstCombiner::BuilderTy &Builder) {
2596 assert(I.getOpcode() == Instruction::Xor);
2597 Value *Op0 = I.getOperand(0);
2598 Value *Op1 = I.getOperand(1);
2599 Value *A, *B;
2601 // There are 4 commuted variants for each of the basic patterns.
2603 // (A & B) ^ (A | B) -> A ^ B
2604 // (A & B) ^ (B | A) -> A ^ B
2605 // (A | B) ^ (A & B) -> A ^ B
2606 // (A | B) ^ (B & A) -> A ^ B
2607 if (match(&I, m_c_Xor(m_And(m_Value(A), m_Value(B)),
2608 m_c_Or(m_Deferred(A), m_Deferred(B))))) {
2609 I.setOperand(0, A);
2610 I.setOperand(1, B);
2611 return &I;
2614 // (A | ~B) ^ (~A | B) -> A ^ B
2615 // (~B | A) ^ (~A | B) -> A ^ B
2616 // (~A | B) ^ (A | ~B) -> A ^ B
2617 // (B | ~A) ^ (A | ~B) -> A ^ B
2618 if (match(&I, m_Xor(m_c_Or(m_Value(A), m_Not(m_Value(B))),
2619 m_c_Or(m_Not(m_Deferred(A)), m_Deferred(B))))) {
2620 I.setOperand(0, A);
2621 I.setOperand(1, B);
2622 return &I;
2625 // (A & ~B) ^ (~A & B) -> A ^ B
2626 // (~B & A) ^ (~A & B) -> A ^ B
2627 // (~A & B) ^ (A & ~B) -> A ^ B
2628 // (B & ~A) ^ (A & ~B) -> A ^ B
2629 if (match(&I, m_Xor(m_c_And(m_Value(A), m_Not(m_Value(B))),
2630 m_c_And(m_Not(m_Deferred(A)), m_Deferred(B))))) {
2631 I.setOperand(0, A);
2632 I.setOperand(1, B);
2633 return &I;
2636 // For the remaining cases we need to get rid of one of the operands.
2637 if (!Op0->hasOneUse() && !Op1->hasOneUse())
2638 return nullptr;
2640 // (A | B) ^ ~(A & B) -> ~(A ^ B)
2641 // (A | B) ^ ~(B & A) -> ~(A ^ B)
2642 // (A & B) ^ ~(A | B) -> ~(A ^ B)
2643 // (A & B) ^ ~(B | A) -> ~(A ^ B)
2644 // Complexity sorting ensures the not will be on the right side.
2645 if ((match(Op0, m_Or(m_Value(A), m_Value(B))) &&
2646 match(Op1, m_Not(m_c_And(m_Specific(A), m_Specific(B))))) ||
2647 (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2648 match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B))))))
2649 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
2651 return nullptr;
2654 Value *InstCombiner::foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS,
2655 BinaryOperator &I) {
2656 assert(I.getOpcode() == Instruction::Xor && I.getOperand(0) == LHS &&
2657 I.getOperand(1) == RHS && "Should be 'xor' with these operands");
2659 if (predicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
2660 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2661 LHS->getOperand(1) == RHS->getOperand(0))
2662 LHS->swapOperands();
2663 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2664 LHS->getOperand(1) == RHS->getOperand(1)) {
2665 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
2666 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2667 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
2668 bool IsSigned = LHS->isSigned() || RHS->isSigned();
2669 return getNewICmpValue(Code, IsSigned, Op0, Op1, Builder);
2673 // TODO: This can be generalized to compares of non-signbits using
2674 // decomposeBitTestICmp(). It could be enhanced more by using (something like)
2675 // foldLogOpOfMaskedICmps().
2676 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
2677 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
2678 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
2679 if ((LHS->hasOneUse() || RHS->hasOneUse()) &&
2680 LHS0->getType() == RHS0->getType() &&
2681 LHS0->getType()->isIntOrIntVectorTy()) {
2682 // (X > -1) ^ (Y > -1) --> (X ^ Y) < 0
2683 // (X < 0) ^ (Y < 0) --> (X ^ Y) < 0
2684 if ((PredL == CmpInst::ICMP_SGT && match(LHS1, m_AllOnes()) &&
2685 PredR == CmpInst::ICMP_SGT && match(RHS1, m_AllOnes())) ||
2686 (PredL == CmpInst::ICMP_SLT && match(LHS1, m_Zero()) &&
2687 PredR == CmpInst::ICMP_SLT && match(RHS1, m_Zero()))) {
2688 Value *Zero = ConstantInt::getNullValue(LHS0->getType());
2689 return Builder.CreateICmpSLT(Builder.CreateXor(LHS0, RHS0), Zero);
2691 // (X > -1) ^ (Y < 0) --> (X ^ Y) > -1
2692 // (X < 0) ^ (Y > -1) --> (X ^ Y) > -1
2693 if ((PredL == CmpInst::ICMP_SGT && match(LHS1, m_AllOnes()) &&
2694 PredR == CmpInst::ICMP_SLT && match(RHS1, m_Zero())) ||
2695 (PredL == CmpInst::ICMP_SLT && match(LHS1, m_Zero()) &&
2696 PredR == CmpInst::ICMP_SGT && match(RHS1, m_AllOnes()))) {
2697 Value *MinusOne = ConstantInt::getAllOnesValue(LHS0->getType());
2698 return Builder.CreateICmpSGT(Builder.CreateXor(LHS0, RHS0), MinusOne);
2702 // Instead of trying to imitate the folds for and/or, decompose this 'xor'
2703 // into those logic ops. That is, try to turn this into an and-of-icmps
2704 // because we have many folds for that pattern.
2706 // This is based on a truth table definition of xor:
2707 // X ^ Y --> (X | Y) & !(X & Y)
2708 if (Value *OrICmp = SimplifyBinOp(Instruction::Or, LHS, RHS, SQ)) {
2709 // TODO: If OrICmp is true, then the definition of xor simplifies to !(X&Y).
2710 // TODO: If OrICmp is false, the whole thing is false (InstSimplify?).
2711 if (Value *AndICmp = SimplifyBinOp(Instruction::And, LHS, RHS, SQ)) {
2712 // TODO: Independently handle cases where the 'and' side is a constant.
2713 ICmpInst *X = nullptr, *Y = nullptr;
2714 if (OrICmp == LHS && AndICmp == RHS) {
2715 // (LHS | RHS) & !(LHS & RHS) --> LHS & !RHS --> X & !Y
2716 X = LHS;
2717 Y = RHS;
2719 if (OrICmp == RHS && AndICmp == LHS) {
2720 // !(LHS & RHS) & (LHS | RHS) --> !LHS & RHS --> !Y & X
2721 X = RHS;
2722 Y = LHS;
2724 if (X && Y && (Y->hasOneUse() || canFreelyInvertAllUsersOf(Y, &I))) {
2725 // Invert the predicate of 'Y', thus inverting its output.
2726 Y->setPredicate(Y->getInversePredicate());
2727 // So, are there other uses of Y?
2728 if (!Y->hasOneUse()) {
2729 // We need to adapt other uses of Y though. Get a value that matches
2730 // the original value of Y before inversion. While this increases
2731 // immediate instruction count, we have just ensured that all the
2732 // users are freely-invertible, so that 'not' *will* get folded away.
2733 BuilderTy::InsertPointGuard Guard(Builder);
2734 // Set insertion point to right after the Y.
2735 Builder.SetInsertPoint(Y->getParent(), ++(Y->getIterator()));
2736 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");
2737 // Replace all uses of Y (excluding the one in NotY!) with NotY.
2738 Y->replaceUsesWithIf(NotY,
2739 [NotY](Use &U) { return U.getUser() != NotY; });
2741 // All done.
2742 return Builder.CreateAnd(LHS, RHS);
2747 return nullptr;
2750 /// If we have a masked merge, in the canonical form of:
2751 /// (assuming that A only has one use.)
2752 /// | A | |B|
2753 /// ((x ^ y) & M) ^ y
2754 /// | D |
2755 /// * If M is inverted:
2756 /// | D |
2757 /// ((x ^ y) & ~M) ^ y
2758 /// We can canonicalize by swapping the final xor operand
2759 /// to eliminate the 'not' of the mask.
2760 /// ((x ^ y) & M) ^ x
2761 /// * If M is a constant, and D has one use, we transform to 'and' / 'or' ops
2762 /// because that shortens the dependency chain and improves analysis:
2763 /// (x & M) | (y & ~M)
2764 static Instruction *visitMaskedMerge(BinaryOperator &I,
2765 InstCombiner::BuilderTy &Builder) {
2766 Value *B, *X, *D;
2767 Value *M;
2768 if (!match(&I, m_c_Xor(m_Value(B),
2769 m_OneUse(m_c_And(
2770 m_CombineAnd(m_c_Xor(m_Deferred(B), m_Value(X)),
2771 m_Value(D)),
2772 m_Value(M))))))
2773 return nullptr;
2775 Value *NotM;
2776 if (match(M, m_Not(m_Value(NotM)))) {
2777 // De-invert the mask and swap the value in B part.
2778 Value *NewA = Builder.CreateAnd(D, NotM);
2779 return BinaryOperator::CreateXor(NewA, X);
2782 Constant *C;
2783 if (D->hasOneUse() && match(M, m_Constant(C))) {
2784 // Unfold.
2785 Value *LHS = Builder.CreateAnd(X, C);
2786 Value *NotC = Builder.CreateNot(C);
2787 Value *RHS = Builder.CreateAnd(B, NotC);
2788 return BinaryOperator::CreateOr(LHS, RHS);
2791 return nullptr;
2794 // Transform
2795 // ~(x ^ y)
2796 // into:
2797 // (~x) ^ y
2798 // or into
2799 // x ^ (~y)
2800 static Instruction *sinkNotIntoXor(BinaryOperator &I,
2801 InstCombiner::BuilderTy &Builder) {
2802 Value *X, *Y;
2803 // FIXME: one-use check is not needed in general, but currently we are unable
2804 // to fold 'not' into 'icmp', if that 'icmp' has multiple uses. (D35182)
2805 if (!match(&I, m_Not(m_OneUse(m_Xor(m_Value(X), m_Value(Y))))))
2806 return nullptr;
2808 // We only want to do the transform if it is free to do.
2809 if (isFreeToInvert(X, X->hasOneUse())) {
2810 // Ok, good.
2811 } else if (isFreeToInvert(Y, Y->hasOneUse())) {
2812 std::swap(X, Y);
2813 } else
2814 return nullptr;
2816 Value *NotX = Builder.CreateNot(X, X->getName() + ".not");
2817 return BinaryOperator::CreateXor(NotX, Y, I.getName() + ".demorgan");
2820 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
2821 // here. We should standardize that construct where it is needed or choose some
2822 // other way to ensure that commutated variants of patterns are not missed.
2823 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
2824 if (Value *V = SimplifyXorInst(I.getOperand(0), I.getOperand(1),
2825 SQ.getWithInstruction(&I)))
2826 return replaceInstUsesWith(I, V);
2828 if (SimplifyAssociativeOrCommutative(I))
2829 return &I;
2831 if (Instruction *X = foldVectorBinop(I))
2832 return X;
2834 if (Instruction *NewXor = foldXorToXor(I, Builder))
2835 return NewXor;
2837 // (A&B)^(A&C) -> A&(B^C) etc
2838 if (Value *V = SimplifyUsingDistributiveLaws(I))
2839 return replaceInstUsesWith(I, V);
2841 // See if we can simplify any instructions used by the instruction whose sole
2842 // purpose is to compute bits we don't care about.
2843 if (SimplifyDemandedInstructionBits(I))
2844 return &I;
2846 if (Value *V = SimplifyBSwap(I, Builder))
2847 return replaceInstUsesWith(I, V);
2849 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2851 // Fold (X & M) ^ (Y & ~M) -> (X & M) | (Y & ~M)
2852 // This it a special case in haveNoCommonBitsSet, but the computeKnownBits
2853 // calls in there are unnecessary as SimplifyDemandedInstructionBits should
2854 // have already taken care of those cases.
2855 Value *M;
2856 if (match(&I, m_c_Xor(m_c_And(m_Not(m_Value(M)), m_Value()),
2857 m_c_And(m_Deferred(M), m_Value()))))
2858 return BinaryOperator::CreateOr(Op0, Op1);
2860 // Apply DeMorgan's Law for 'nand' / 'nor' logic with an inverted operand.
2861 Value *X, *Y;
2863 // We must eliminate the and/or (one-use) for these transforms to not increase
2864 // the instruction count.
2865 // ~(~X & Y) --> (X | ~Y)
2866 // ~(Y & ~X) --> (X | ~Y)
2867 if (match(&I, m_Not(m_OneUse(m_c_And(m_Not(m_Value(X)), m_Value(Y)))))) {
2868 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");
2869 return BinaryOperator::CreateOr(X, NotY);
2871 // ~(~X | Y) --> (X & ~Y)
2872 // ~(Y | ~X) --> (X & ~Y)
2873 if (match(&I, m_Not(m_OneUse(m_c_Or(m_Not(m_Value(X)), m_Value(Y)))))) {
2874 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");
2875 return BinaryOperator::CreateAnd(X, NotY);
2878 if (Instruction *Xor = visitMaskedMerge(I, Builder))
2879 return Xor;
2881 // Is this a 'not' (~) fed by a binary operator?
2882 BinaryOperator *NotVal;
2883 if (match(&I, m_Not(m_BinOp(NotVal)))) {
2884 if (NotVal->getOpcode() == Instruction::And ||
2885 NotVal->getOpcode() == Instruction::Or) {
2886 // Apply DeMorgan's Law when inverts are free:
2887 // ~(X & Y) --> (~X | ~Y)
2888 // ~(X | Y) --> (~X & ~Y)
2889 if (isFreeToInvert(NotVal->getOperand(0),
2890 NotVal->getOperand(0)->hasOneUse()) &&
2891 isFreeToInvert(NotVal->getOperand(1),
2892 NotVal->getOperand(1)->hasOneUse())) {
2893 Value *NotX = Builder.CreateNot(NotVal->getOperand(0), "notlhs");
2894 Value *NotY = Builder.CreateNot(NotVal->getOperand(1), "notrhs");
2895 if (NotVal->getOpcode() == Instruction::And)
2896 return BinaryOperator::CreateOr(NotX, NotY);
2897 return BinaryOperator::CreateAnd(NotX, NotY);
2901 // ~(X - Y) --> ~X + Y
2902 if (match(NotVal, m_Sub(m_Value(X), m_Value(Y))))
2903 if (isa<Constant>(X) || NotVal->hasOneUse())
2904 return BinaryOperator::CreateAdd(Builder.CreateNot(X), Y);
2906 // ~(~X >>s Y) --> (X >>s Y)
2907 if (match(NotVal, m_AShr(m_Not(m_Value(X)), m_Value(Y))))
2908 return BinaryOperator::CreateAShr(X, Y);
2910 // If we are inverting a right-shifted constant, we may be able to eliminate
2911 // the 'not' by inverting the constant and using the opposite shift type.
2912 // Canonicalization rules ensure that only a negative constant uses 'ashr',
2913 // but we must check that in case that transform has not fired yet.
2915 // ~(C >>s Y) --> ~C >>u Y (when inverting the replicated sign bits)
2916 Constant *C;
2917 if (match(NotVal, m_AShr(m_Constant(C), m_Value(Y))) &&
2918 match(C, m_Negative()))
2919 return BinaryOperator::CreateLShr(ConstantExpr::getNot(C), Y);
2921 // ~(C >>u Y) --> ~C >>s Y (when inverting the replicated sign bits)
2922 if (match(NotVal, m_LShr(m_Constant(C), m_Value(Y))) &&
2923 match(C, m_NonNegative()))
2924 return BinaryOperator::CreateAShr(ConstantExpr::getNot(C), Y);
2926 // ~(X + C) --> -(C + 1) - X
2927 if (match(Op0, m_Add(m_Value(X), m_Constant(C))))
2928 return BinaryOperator::CreateSub(ConstantExpr::getNeg(AddOne(C)), X);
2931 // Use DeMorgan and reassociation to eliminate a 'not' op.
2932 Constant *C1;
2933 if (match(Op1, m_Constant(C1))) {
2934 Constant *C2;
2935 if (match(Op0, m_OneUse(m_Or(m_Not(m_Value(X)), m_Constant(C2))))) {
2936 // (~X | C2) ^ C1 --> ((X & ~C2) ^ -1) ^ C1 --> (X & ~C2) ^ ~C1
2937 Value *And = Builder.CreateAnd(X, ConstantExpr::getNot(C2));
2938 return BinaryOperator::CreateXor(And, ConstantExpr::getNot(C1));
2940 if (match(Op0, m_OneUse(m_And(m_Not(m_Value(X)), m_Constant(C2))))) {
2941 // (~X & C2) ^ C1 --> ((X | ~C2) ^ -1) ^ C1 --> (X | ~C2) ^ ~C1
2942 Value *Or = Builder.CreateOr(X, ConstantExpr::getNot(C2));
2943 return BinaryOperator::CreateXor(Or, ConstantExpr::getNot(C1));
2947 // not (cmp A, B) = !cmp A, B
2948 CmpInst::Predicate Pred;
2949 if (match(&I, m_Not(m_OneUse(m_Cmp(Pred, m_Value(), m_Value()))))) {
2950 cast<CmpInst>(Op0)->setPredicate(CmpInst::getInversePredicate(Pred));
2951 return replaceInstUsesWith(I, Op0);
2955 const APInt *RHSC;
2956 if (match(Op1, m_APInt(RHSC))) {
2957 Value *X;
2958 const APInt *C;
2959 if (RHSC->isSignMask() && match(Op0, m_Sub(m_APInt(C), m_Value(X)))) {
2960 // (C - X) ^ signmask -> (C + signmask - X)
2961 Constant *NewC = ConstantInt::get(I.getType(), *C + *RHSC);
2962 return BinaryOperator::CreateSub(NewC, X);
2964 if (RHSC->isSignMask() && match(Op0, m_Add(m_Value(X), m_APInt(C)))) {
2965 // (X + C) ^ signmask -> (X + C + signmask)
2966 Constant *NewC = ConstantInt::get(I.getType(), *C + *RHSC);
2967 return BinaryOperator::CreateAdd(X, NewC);
2970 // (X|C1)^C2 -> X^(C1^C2) iff X&~C1 == 0
2971 if (match(Op0, m_Or(m_Value(X), m_APInt(C))) &&
2972 MaskedValueIsZero(X, *C, 0, &I)) {
2973 Constant *NewC = ConstantInt::get(I.getType(), *C ^ *RHSC);
2974 Worklist.Add(cast<Instruction>(Op0));
2975 I.setOperand(0, X);
2976 I.setOperand(1, NewC);
2977 return &I;
2982 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1)) {
2983 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2984 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2985 if (Op0I->getOpcode() == Instruction::LShr) {
2986 // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3)
2987 // E1 = "X ^ C1"
2988 BinaryOperator *E1;
2989 ConstantInt *C1;
2990 if (Op0I->hasOneUse() &&
2991 (E1 = dyn_cast<BinaryOperator>(Op0I->getOperand(0))) &&
2992 E1->getOpcode() == Instruction::Xor &&
2993 (C1 = dyn_cast<ConstantInt>(E1->getOperand(1)))) {
2994 // fold (C1 >> C2) ^ C3
2995 ConstantInt *C2 = Op0CI, *C3 = RHSC;
2996 APInt FoldConst = C1->getValue().lshr(C2->getValue());
2997 FoldConst ^= C3->getValue();
2998 // Prepare the two operands.
2999 Value *Opnd0 = Builder.CreateLShr(E1->getOperand(0), C2);
3000 Opnd0->takeName(Op0I);
3001 cast<Instruction>(Opnd0)->setDebugLoc(I.getDebugLoc());
3002 Value *FoldVal = ConstantInt::get(Opnd0->getType(), FoldConst);
3004 return BinaryOperator::CreateXor(Opnd0, FoldVal);
3011 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))
3012 return FoldedLogic;
3014 // Y ^ (X | Y) --> X & ~Y
3015 // Y ^ (Y | X) --> X & ~Y
3016 if (match(Op1, m_OneUse(m_c_Or(m_Value(X), m_Specific(Op0)))))
3017 return BinaryOperator::CreateAnd(X, Builder.CreateNot(Op0));
3018 // (X | Y) ^ Y --> X & ~Y
3019 // (Y | X) ^ Y --> X & ~Y
3020 if (match(Op0, m_OneUse(m_c_Or(m_Value(X), m_Specific(Op1)))))
3021 return BinaryOperator::CreateAnd(X, Builder.CreateNot(Op1));
3023 // Y ^ (X & Y) --> ~X & Y
3024 // Y ^ (Y & X) --> ~X & Y
3025 if (match(Op1, m_OneUse(m_c_And(m_Value(X), m_Specific(Op0)))))
3026 return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(X));
3027 // (X & Y) ^ Y --> ~X & Y
3028 // (Y & X) ^ Y --> ~X & Y
3029 // Canonical form is (X & C) ^ C; don't touch that.
3030 // TODO: A 'not' op is better for analysis and codegen, but demanded bits must
3031 // be fixed to prefer that (otherwise we get infinite looping).
3032 if (!match(Op1, m_Constant()) &&
3033 match(Op0, m_OneUse(m_c_And(m_Value(X), m_Specific(Op1)))))
3034 return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(X));
3036 Value *A, *B, *C;
3037 // (A ^ B) ^ (A | C) --> (~A & C) ^ B -- There are 4 commuted variants.
3038 if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_Value(A), m_Value(B))),
3039 m_OneUse(m_c_Or(m_Deferred(A), m_Value(C))))))
3040 return BinaryOperator::CreateXor(
3041 Builder.CreateAnd(Builder.CreateNot(A), C), B);
3043 // (A ^ B) ^ (B | C) --> (~B & C) ^ A -- There are 4 commuted variants.
3044 if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_Value(A), m_Value(B))),
3045 m_OneUse(m_c_Or(m_Deferred(B), m_Value(C))))))
3046 return BinaryOperator::CreateXor(
3047 Builder.CreateAnd(Builder.CreateNot(B), C), A);
3049 // (A & B) ^ (A ^ B) -> (A | B)
3050 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
3051 match(Op1, m_c_Xor(m_Specific(A), m_Specific(B))))
3052 return BinaryOperator::CreateOr(A, B);
3053 // (A ^ B) ^ (A & B) -> (A | B)
3054 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
3055 match(Op1, m_c_And(m_Specific(A), m_Specific(B))))
3056 return BinaryOperator::CreateOr(A, B);
3058 // (A & ~B) ^ ~A -> ~(A & B)
3059 // (~B & A) ^ ~A -> ~(A & B)
3060 if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) &&
3061 match(Op1, m_Not(m_Specific(A))))
3062 return BinaryOperator::CreateNot(Builder.CreateAnd(A, B));
3064 if (auto *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
3065 if (auto *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
3066 if (Value *V = foldXorOfICmps(LHS, RHS, I))
3067 return replaceInstUsesWith(I, V);
3069 if (Instruction *CastedXor = foldCastedBitwiseLogic(I))
3070 return CastedXor;
3072 // Canonicalize a shifty way to code absolute value to the common pattern.
3073 // There are 4 potential commuted variants. Move the 'ashr' candidate to Op1.
3074 // We're relying on the fact that we only do this transform when the shift has
3075 // exactly 2 uses and the add has exactly 1 use (otherwise, we might increase
3076 // instructions).
3077 if (Op0->hasNUses(2))
3078 std::swap(Op0, Op1);
3080 const APInt *ShAmt;
3081 Type *Ty = I.getType();
3082 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
3083 Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 &&
3084 match(Op0, m_OneUse(m_c_Add(m_Specific(A), m_Specific(Op1))))) {
3085 // B = ashr i32 A, 31 ; smear the sign bit
3086 // xor (add A, B), B ; add -1 and flip bits if negative
3087 // --> (A < 0) ? -A : A
3088 Value *Cmp = Builder.CreateICmpSLT(A, ConstantInt::getNullValue(Ty));
3089 // Copy the nuw/nsw flags from the add to the negate.
3090 auto *Add = cast<BinaryOperator>(Op0);
3091 Value *Neg = Builder.CreateNeg(A, "", Add->hasNoUnsignedWrap(),
3092 Add->hasNoSignedWrap());
3093 return SelectInst::Create(Cmp, Neg, A);
3096 // Eliminate a bitwise 'not' op of 'not' min/max by inverting the min/max:
3098 // %notx = xor i32 %x, -1
3099 // %cmp1 = icmp sgt i32 %notx, %y
3100 // %smax = select i1 %cmp1, i32 %notx, i32 %y
3101 // %res = xor i32 %smax, -1
3102 // =>
3103 // %noty = xor i32 %y, -1
3104 // %cmp2 = icmp slt %x, %noty
3105 // %res = select i1 %cmp2, i32 %x, i32 %noty
3107 // Same is applicable for smin/umax/umin.
3108 if (match(Op1, m_AllOnes()) && Op0->hasOneUse()) {
3109 Value *LHS, *RHS;
3110 SelectPatternFlavor SPF = matchSelectPattern(Op0, LHS, RHS).Flavor;
3111 if (SelectPatternResult::isMinOrMax(SPF)) {
3112 // It's possible we get here before the not has been simplified, so make
3113 // sure the input to the not isn't freely invertible.
3114 if (match(LHS, m_Not(m_Value(X))) && !isFreeToInvert(X, X->hasOneUse())) {
3115 Value *NotY = Builder.CreateNot(RHS);
3116 return SelectInst::Create(
3117 Builder.CreateICmp(getInverseMinMaxPred(SPF), X, NotY), X, NotY);
3120 // It's possible we get here before the not has been simplified, so make
3121 // sure the input to the not isn't freely invertible.
3122 if (match(RHS, m_Not(m_Value(Y))) && !isFreeToInvert(Y, Y->hasOneUse())) {
3123 Value *NotX = Builder.CreateNot(LHS);
3124 return SelectInst::Create(
3125 Builder.CreateICmp(getInverseMinMaxPred(SPF), NotX, Y), NotX, Y);
3128 // If both sides are freely invertible, then we can get rid of the xor
3129 // completely.
3130 if (isFreeToInvert(LHS, !LHS->hasNUsesOrMore(3)) &&
3131 isFreeToInvert(RHS, !RHS->hasNUsesOrMore(3))) {
3132 Value *NotLHS = Builder.CreateNot(LHS);
3133 Value *NotRHS = Builder.CreateNot(RHS);
3134 return SelectInst::Create(
3135 Builder.CreateICmp(getInverseMinMaxPred(SPF), NotLHS, NotRHS),
3136 NotLHS, NotRHS);
3141 if (Instruction *NewXor = sinkNotIntoXor(I, Builder))
3142 return NewXor;
3144 return nullptr;