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.
10 //
11 //===----------------------------------------------------------------------===//
12
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;
22
23 #define DEBUG_TYPE "instcombine"
24
25 /// Similar to getICmpCode but for FCmpInst. This encodes a fcmp predicate into
26 /// a four bit mask.
getFCmpCode(FCmpInst::Predicate CC)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;
49 }
50
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.
getNewICmpValue(unsigned Code,bool Sign,Value * LHS,Value * RHS,InstCombiner::BuilderTy & Builder)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);
61 }
62
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.
getFCmpValue(unsigned Code,Value * LHS,Value * RHS,InstCombiner::BuilderTy & Builder)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);
75 }
76
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.
SimplifyBSwap(BinaryOperator & I,InstCombiner::BuilderTy & Builder)82 static Value *SimplifyBSwap(BinaryOperator &I,
83 InstCombiner::BuilderTy &Builder) {
84 assert(I.isBitwiseLogicOp() && "Unexpected opcode for bswap simplifying");
85
86 Value *OldLHS = I.getOperand(0);
87 Value *OldRHS = I.getOperand(1);
88
89 Value *NewLHS;
90 if (!match(OldLHS, m_BSwap(m_Value(NewLHS))))
91 return nullptr;
92
93 Value *NewRHS;
94 const APInt *C;
95
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;
108
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);
113 }
114
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'.
OptAndOp(BinaryOperator * Op,ConstantInt * OpRHS,ConstantInt * AndRHS,BinaryOperator & TheAnd)117 Instruction *InstCombiner::OptAndOp(BinaryOperator *Op,
118 ConstantInt *OpRHS,
119 ConstantInt *AndRHS,
120 BinaryOperator &TheAnd) {
121 Value *X = Op->getOperand(0);
122
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();
131
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();
138
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 return replaceOperand(TheAnd, 0, X);
147 } else {
148 // Pull the XOR out of the AND.
149 Value *NewAnd = Builder.CreateAnd(X, AndRHS);
150 NewAnd->takeName(Op);
151 return BinaryOperator::CreateXor(NewAnd, AndRHS);
152 }
153 }
154 }
155 }
156 break;
157 }
158 return nullptr;
159 }
160
161 /// Emit a computation of: (V >= Lo && V < Hi) if Inside is true, otherwise
162 /// (V < Lo || V >= Hi). This method expects that Lo < Hi. IsSigned indicates
163 /// whether to treat V, Lo, and Hi as signed or not.
insertRangeTest(Value * V,const APInt & Lo,const APInt & Hi,bool isSigned,bool Inside)164 Value *InstCombiner::insertRangeTest(Value *V, const APInt &Lo, const APInt &Hi,
165 bool isSigned, bool Inside) {
166 assert((isSigned ? Lo.slt(Hi) : Lo.ult(Hi)) &&
167 "Lo is not < Hi in range emission code!");
168
169 Type *Ty = V->getType();
170
171 // V >= Min && V < Hi --> V < Hi
172 // V < Min || V >= Hi --> V >= Hi
173 ICmpInst::Predicate Pred = Inside ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_UGE;
174 if (isSigned ? Lo.isMinSignedValue() : Lo.isMinValue()) {
175 Pred = isSigned ? ICmpInst::getSignedPredicate(Pred) : Pred;
176 return Builder.CreateICmp(Pred, V, ConstantInt::get(Ty, Hi));
177 }
178
179 // V >= Lo && V < Hi --> V - Lo u< Hi - Lo
180 // V < Lo || V >= Hi --> V - Lo u>= Hi - Lo
181 Value *VMinusLo =
182 Builder.CreateSub(V, ConstantInt::get(Ty, Lo), V->getName() + ".off");
183 Constant *HiMinusLo = ConstantInt::get(Ty, Hi - Lo);
184 return Builder.CreateICmp(Pred, VMinusLo, HiMinusLo);
185 }
186
187 /// Classify (icmp eq (A & B), C) and (icmp ne (A & B), C) as matching patterns
188 /// that can be simplified.
189 /// One of A and B is considered the mask. The other is the value. This is
190 /// described as the "AMask" or "BMask" part of the enum. If the enum contains
191 /// only "Mask", then both A and B can be considered masks. If A is the mask,
192 /// then it was proven that (A & C) == C. This is trivial if C == A or C == 0.
193 /// If both A and C are constants, this proof is also easy.
194 /// For the following explanations, we assume that A is the mask.
195 ///
196 /// "AllOnes" declares that the comparison is true only if (A & B) == A or all
197 /// bits of A are set in B.
198 /// Example: (icmp eq (A & 3), 3) -> AMask_AllOnes
199 ///
200 /// "AllZeros" declares that the comparison is true only if (A & B) == 0 or all
201 /// bits of A are cleared in B.
202 /// Example: (icmp eq (A & 3), 0) -> Mask_AllZeroes
203 ///
204 /// "Mixed" declares that (A & B) == C and C might or might not contain any
205 /// number of one bits and zero bits.
206 /// Example: (icmp eq (A & 3), 1) -> AMask_Mixed
207 ///
208 /// "Not" means that in above descriptions "==" should be replaced by "!=".
209 /// Example: (icmp ne (A & 3), 3) -> AMask_NotAllOnes
210 ///
211 /// If the mask A contains a single bit, then the following is equivalent:
212 /// (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
213 /// (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
214 enum MaskedICmpType {
215 AMask_AllOnes = 1,
216 AMask_NotAllOnes = 2,
217 BMask_AllOnes = 4,
218 BMask_NotAllOnes = 8,
219 Mask_AllZeros = 16,
220 Mask_NotAllZeros = 32,
221 AMask_Mixed = 64,
222 AMask_NotMixed = 128,
223 BMask_Mixed = 256,
224 BMask_NotMixed = 512
225 };
226
227 /// Return the set of patterns (from MaskedICmpType) that (icmp SCC (A & B), C)
228 /// satisfies.
getMaskedICmpType(Value * A,Value * B,Value * C,ICmpInst::Predicate Pred)229 static unsigned getMaskedICmpType(Value *A, Value *B, Value *C,
230 ICmpInst::Predicate Pred) {
231 ConstantInt *ACst = dyn_cast<ConstantInt>(A);
232 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
233 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
234 bool IsEq = (Pred == ICmpInst::ICMP_EQ);
235 bool IsAPow2 = (ACst && !ACst->isZero() && ACst->getValue().isPowerOf2());
236 bool IsBPow2 = (BCst && !BCst->isZero() && BCst->getValue().isPowerOf2());
237 unsigned MaskVal = 0;
238 if (CCst && CCst->isZero()) {
239 // if C is zero, then both A and B qualify as mask
240 MaskVal |= (IsEq ? (Mask_AllZeros | AMask_Mixed | BMask_Mixed)
241 : (Mask_NotAllZeros | AMask_NotMixed | BMask_NotMixed));
242 if (IsAPow2)
243 MaskVal |= (IsEq ? (AMask_NotAllOnes | AMask_NotMixed)
244 : (AMask_AllOnes | AMask_Mixed));
245 if (IsBPow2)
246 MaskVal |= (IsEq ? (BMask_NotAllOnes | BMask_NotMixed)
247 : (BMask_AllOnes | BMask_Mixed));
248 return MaskVal;
249 }
250
251 if (A == C) {
252 MaskVal |= (IsEq ? (AMask_AllOnes | AMask_Mixed)
253 : (AMask_NotAllOnes | AMask_NotMixed));
254 if (IsAPow2)
255 MaskVal |= (IsEq ? (Mask_NotAllZeros | AMask_NotMixed)
256 : (Mask_AllZeros | AMask_Mixed));
257 } else if (ACst && CCst && ConstantExpr::getAnd(ACst, CCst) == CCst) {
258 MaskVal |= (IsEq ? AMask_Mixed : AMask_NotMixed);
259 }
260
261 if (B == C) {
262 MaskVal |= (IsEq ? (BMask_AllOnes | BMask_Mixed)
263 : (BMask_NotAllOnes | BMask_NotMixed));
264 if (IsBPow2)
265 MaskVal |= (IsEq ? (Mask_NotAllZeros | BMask_NotMixed)
266 : (Mask_AllZeros | BMask_Mixed));
267 } else if (BCst && CCst && ConstantExpr::getAnd(BCst, CCst) == CCst) {
268 MaskVal |= (IsEq ? BMask_Mixed : BMask_NotMixed);
269 }
270
271 return MaskVal;
272 }
273
274 /// Convert an analysis of a masked ICmp into its equivalent if all boolean
275 /// operations had the opposite sense. Since each "NotXXX" flag (recording !=)
276 /// is adjacent to the corresponding normal flag (recording ==), this just
277 /// involves swapping those bits over.
conjugateICmpMask(unsigned Mask)278 static unsigned conjugateICmpMask(unsigned Mask) {
279 unsigned NewMask;
280 NewMask = (Mask & (AMask_AllOnes | BMask_AllOnes | Mask_AllZeros |
281 AMask_Mixed | BMask_Mixed))
282 << 1;
283
284 NewMask |= (Mask & (AMask_NotAllOnes | BMask_NotAllOnes | Mask_NotAllZeros |
285 AMask_NotMixed | BMask_NotMixed))
286 >> 1;
287
288 return NewMask;
289 }
290
291 // Adapts the external decomposeBitTestICmp for local use.
decomposeBitTestICmp(Value * LHS,Value * RHS,CmpInst::Predicate & Pred,Value * & X,Value * & Y,Value * & Z)292 static bool decomposeBitTestICmp(Value *LHS, Value *RHS, CmpInst::Predicate &Pred,
293 Value *&X, Value *&Y, Value *&Z) {
294 APInt Mask;
295 if (!llvm::decomposeBitTestICmp(LHS, RHS, Pred, X, Mask))
296 return false;
297
298 Y = ConstantInt::get(X->getType(), Mask);
299 Z = ConstantInt::get(X->getType(), 0);
300 return true;
301 }
302
303 /// Handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E).
304 /// Return the pattern classes (from MaskedICmpType) for the left hand side and
305 /// the right hand side as a pair.
306 /// LHS and RHS are the left hand side and the right hand side ICmps and PredL
307 /// and PredR are their predicates, respectively.
308 static
309 Optional<std::pair<unsigned, unsigned>>
getMaskedTypeForICmpPair(Value * & A,Value * & B,Value * & C,Value * & D,Value * & E,ICmpInst * LHS,ICmpInst * RHS,ICmpInst::Predicate & PredL,ICmpInst::Predicate & PredR)310 getMaskedTypeForICmpPair(Value *&A, Value *&B, Value *&C,
311 Value *&D, Value *&E, ICmpInst *LHS,
312 ICmpInst *RHS,
313 ICmpInst::Predicate &PredL,
314 ICmpInst::Predicate &PredR) {
315 // vectors are not (yet?) supported. Don't support pointers either.
316 if (!LHS->getOperand(0)->getType()->isIntegerTy() ||
317 !RHS->getOperand(0)->getType()->isIntegerTy())
318 return None;
319
320 // Here comes the tricky part:
321 // LHS might be of the form L11 & L12 == X, X == L21 & L22,
322 // and L11 & L12 == L21 & L22. The same goes for RHS.
323 // Now we must find those components L** and R**, that are equal, so
324 // that we can extract the parameters A, B, C, D, and E for the canonical
325 // above.
326 Value *L1 = LHS->getOperand(0);
327 Value *L2 = LHS->getOperand(1);
328 Value *L11, *L12, *L21, *L22;
329 // Check whether the icmp can be decomposed into a bit test.
330 if (decomposeBitTestICmp(L1, L2, PredL, L11, L12, L2)) {
331 L21 = L22 = L1 = nullptr;
332 } else {
333 // Look for ANDs in the LHS icmp.
334 if (!match(L1, m_And(m_Value(L11), m_Value(L12)))) {
335 // Any icmp can be viewed as being trivially masked; if it allows us to
336 // remove one, it's worth it.
337 L11 = L1;
338 L12 = Constant::getAllOnesValue(L1->getType());
339 }
340
341 if (!match(L2, m_And(m_Value(L21), m_Value(L22)))) {
342 L21 = L2;
343 L22 = Constant::getAllOnesValue(L2->getType());
344 }
345 }
346
347 // Bail if LHS was a icmp that can't be decomposed into an equality.
348 if (!ICmpInst::isEquality(PredL))
349 return None;
350
351 Value *R1 = RHS->getOperand(0);
352 Value *R2 = RHS->getOperand(1);
353 Value *R11, *R12;
354 bool Ok = false;
355 if (decomposeBitTestICmp(R1, R2, PredR, R11, R12, R2)) {
356 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
357 A = R11;
358 D = R12;
359 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
360 A = R12;
361 D = R11;
362 } else {
363 return None;
364 }
365 E = R2;
366 R1 = nullptr;
367 Ok = true;
368 } else {
369 if (!match(R1, m_And(m_Value(R11), m_Value(R12)))) {
370 // As before, model no mask as a trivial mask if it'll let us do an
371 // optimization.
372 R11 = R1;
373 R12 = Constant::getAllOnesValue(R1->getType());
374 }
375
376 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
377 A = R11;
378 D = R12;
379 E = R2;
380 Ok = true;
381 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
382 A = R12;
383 D = R11;
384 E = R2;
385 Ok = true;
386 }
387 }
388
389 // Bail if RHS was a icmp that can't be decomposed into an equality.
390 if (!ICmpInst::isEquality(PredR))
391 return None;
392
393 // Look for ANDs on the right side of the RHS icmp.
394 if (!Ok) {
395 if (!match(R2, m_And(m_Value(R11), m_Value(R12)))) {
396 R11 = R2;
397 R12 = Constant::getAllOnesValue(R2->getType());
398 }
399
400 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
401 A = R11;
402 D = R12;
403 E = R1;
404 Ok = true;
405 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
406 A = R12;
407 D = R11;
408 E = R1;
409 Ok = true;
410 } else {
411 return None;
412 }
413 }
414 if (!Ok)
415 return None;
416
417 if (L11 == A) {
418 B = L12;
419 C = L2;
420 } else if (L12 == A) {
421 B = L11;
422 C = L2;
423 } else if (L21 == A) {
424 B = L22;
425 C = L1;
426 } else if (L22 == A) {
427 B = L21;
428 C = L1;
429 }
430
431 unsigned LeftType = getMaskedICmpType(A, B, C, PredL);
432 unsigned RightType = getMaskedICmpType(A, D, E, PredR);
433 return Optional<std::pair<unsigned, unsigned>>(std::make_pair(LeftType, RightType));
434 }
435
436 /// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E) into a single
437 /// (icmp(A & X) ==/!= Y), where the left-hand side is of type Mask_NotAllZeros
438 /// and the right hand side is of type BMask_Mixed. For example,
439 /// (icmp (A & 12) != 0) & (icmp (A & 15) == 8) -> (icmp (A & 15) == 8).
foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(ICmpInst * LHS,ICmpInst * RHS,bool IsAnd,Value * A,Value * B,Value * C,Value * D,Value * E,ICmpInst::Predicate PredL,ICmpInst::Predicate PredR,llvm::InstCombiner::BuilderTy & Builder)440 static Value * foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(
441 ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
442 Value *A, Value *B, Value *C, Value *D, Value *E,
443 ICmpInst::Predicate PredL, ICmpInst::Predicate PredR,
444 llvm::InstCombiner::BuilderTy &Builder) {
445 // We are given the canonical form:
446 // (icmp ne (A & B), 0) & (icmp eq (A & D), E).
447 // where D & E == E.
448 //
449 // If IsAnd is false, we get it in negated form:
450 // (icmp eq (A & B), 0) | (icmp ne (A & D), E) ->
451 // !((icmp ne (A & B), 0) & (icmp eq (A & D), E)).
452 //
453 // We currently handle the case of B, C, D, E are constant.
454 //
455 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
456 if (!BCst)
457 return nullptr;
458 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
459 if (!CCst)
460 return nullptr;
461 ConstantInt *DCst = dyn_cast<ConstantInt>(D);
462 if (!DCst)
463 return nullptr;
464 ConstantInt *ECst = dyn_cast<ConstantInt>(E);
465 if (!ECst)
466 return nullptr;
467
468 ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
469
470 // Update E to the canonical form when D is a power of two and RHS is
471 // canonicalized as,
472 // (icmp ne (A & D), 0) -> (icmp eq (A & D), D) or
473 // (icmp ne (A & D), D) -> (icmp eq (A & D), 0).
474 if (PredR != NewCC)
475 ECst = cast<ConstantInt>(ConstantExpr::getXor(DCst, ECst));
476
477 // If B or D is zero, skip because if LHS or RHS can be trivially folded by
478 // other folding rules and this pattern won't apply any more.
479 if (BCst->getValue() == 0 || DCst->getValue() == 0)
480 return nullptr;
481
482 // If B and D don't intersect, ie. (B & D) == 0, no folding because we can't
483 // deduce anything from it.
484 // For example,
485 // (icmp ne (A & 12), 0) & (icmp eq (A & 3), 1) -> no folding.
486 if ((BCst->getValue() & DCst->getValue()) == 0)
487 return nullptr;
488
489 // If the following two conditions are met:
490 //
491 // 1. mask B covers only a single bit that's not covered by mask D, that is,
492 // (B & (B ^ D)) is a power of 2 (in other words, B minus the intersection of
493 // B and D has only one bit set) and,
494 //
495 // 2. RHS (and E) indicates that the rest of B's bits are zero (in other
496 // words, the intersection of B and D is zero), that is, ((B & D) & E) == 0
497 //
498 // then that single bit in B must be one and thus the whole expression can be
499 // folded to
500 // (A & (B | D)) == (B & (B ^ D)) | E.
501 //
502 // For example,
503 // (icmp ne (A & 12), 0) & (icmp eq (A & 7), 1) -> (icmp eq (A & 15), 9)
504 // (icmp ne (A & 15), 0) & (icmp eq (A & 7), 0) -> (icmp eq (A & 15), 8)
505 if ((((BCst->getValue() & DCst->getValue()) & ECst->getValue()) == 0) &&
506 (BCst->getValue() & (BCst->getValue() ^ DCst->getValue())).isPowerOf2()) {
507 APInt BorD = BCst->getValue() | DCst->getValue();
508 APInt BandBxorDorE = (BCst->getValue() & (BCst->getValue() ^ DCst->getValue())) |
509 ECst->getValue();
510 Value *NewMask = ConstantInt::get(BCst->getType(), BorD);
511 Value *NewMaskedValue = ConstantInt::get(BCst->getType(), BandBxorDorE);
512 Value *NewAnd = Builder.CreateAnd(A, NewMask);
513 return Builder.CreateICmp(NewCC, NewAnd, NewMaskedValue);
514 }
515
516 auto IsSubSetOrEqual = [](ConstantInt *C1, ConstantInt *C2) {
517 return (C1->getValue() & C2->getValue()) == C1->getValue();
518 };
519 auto IsSuperSetOrEqual = [](ConstantInt *C1, ConstantInt *C2) {
520 return (C1->getValue() & C2->getValue()) == C2->getValue();
521 };
522
523 // In the following, we consider only the cases where B is a superset of D, B
524 // is a subset of D, or B == D because otherwise there's at least one bit
525 // covered by B but not D, in which case we can't deduce much from it, so
526 // no folding (aside from the single must-be-one bit case right above.)
527 // For example,
528 // (icmp ne (A & 14), 0) & (icmp eq (A & 3), 1) -> no folding.
529 if (!IsSubSetOrEqual(BCst, DCst) && !IsSuperSetOrEqual(BCst, DCst))
530 return nullptr;
531
532 // At this point, either B is a superset of D, B is a subset of D or B == D.
533
534 // If E is zero, if B is a subset of (or equal to) D, LHS and RHS contradict
535 // and the whole expression becomes false (or true if negated), otherwise, no
536 // folding.
537 // For example,
538 // (icmp ne (A & 3), 0) & (icmp eq (A & 7), 0) -> false.
539 // (icmp ne (A & 15), 0) & (icmp eq (A & 3), 0) -> no folding.
540 if (ECst->isZero()) {
541 if (IsSubSetOrEqual(BCst, DCst))
542 return ConstantInt::get(LHS->getType(), !IsAnd);
543 return nullptr;
544 }
545
546 // At this point, B, D, E aren't zero and (B & D) == B, (B & D) == D or B ==
547 // D. If B is a superset of (or equal to) D, since E is not zero, LHS is
548 // subsumed by RHS (RHS implies LHS.) So the whole expression becomes
549 // RHS. For example,
550 // (icmp ne (A & 255), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).
551 // (icmp ne (A & 15), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).
552 if (IsSuperSetOrEqual(BCst, DCst))
553 return RHS;
554 // Otherwise, B is a subset of D. If B and E have a common bit set,
555 // ie. (B & E) != 0, then LHS is subsumed by RHS. For example.
556 // (icmp ne (A & 12), 0) & (icmp eq (A & 15), 8) -> (icmp eq (A & 15), 8).
557 assert(IsSubSetOrEqual(BCst, DCst) && "Precondition due to above code");
558 if ((BCst->getValue() & ECst->getValue()) != 0)
559 return RHS;
560 // Otherwise, LHS and RHS contradict and the whole expression becomes false
561 // (or true if negated.) For example,
562 // (icmp ne (A & 7), 0) & (icmp eq (A & 15), 8) -> false.
563 // (icmp ne (A & 6), 0) & (icmp eq (A & 15), 8) -> false.
564 return ConstantInt::get(LHS->getType(), !IsAnd);
565 }
566
567 /// Try to fold (icmp(A & B) ==/!= 0) &/| (icmp(A & D) ==/!= E) into a single
568 /// (icmp(A & X) ==/!= Y), where the left-hand side and the right hand side
569 /// aren't of the common mask pattern type.
foldLogOpOfMaskedICmpsAsymmetric(ICmpInst * LHS,ICmpInst * RHS,bool IsAnd,Value * A,Value * B,Value * C,Value * D,Value * E,ICmpInst::Predicate PredL,ICmpInst::Predicate PredR,unsigned LHSMask,unsigned RHSMask,llvm::InstCombiner::BuilderTy & Builder)570 static Value *foldLogOpOfMaskedICmpsAsymmetric(
571 ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
572 Value *A, Value *B, Value *C, Value *D, Value *E,
573 ICmpInst::Predicate PredL, ICmpInst::Predicate PredR,
574 unsigned LHSMask, unsigned RHSMask,
575 llvm::InstCombiner::BuilderTy &Builder) {
576 assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&
577 "Expected equality predicates for masked type of icmps.");
578 // Handle Mask_NotAllZeros-BMask_Mixed cases.
579 // (icmp ne/eq (A & B), C) &/| (icmp eq/ne (A & D), E), or
580 // (icmp eq/ne (A & B), C) &/| (icmp ne/eq (A & D), E)
581 // which gets swapped to
582 // (icmp ne/eq (A & D), E) &/| (icmp eq/ne (A & B), C).
583 if (!IsAnd) {
584 LHSMask = conjugateICmpMask(LHSMask);
585 RHSMask = conjugateICmpMask(RHSMask);
586 }
587 if ((LHSMask & Mask_NotAllZeros) && (RHSMask & BMask_Mixed)) {
588 if (Value *V = foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(
589 LHS, RHS, IsAnd, A, B, C, D, E,
590 PredL, PredR, Builder)) {
591 return V;
592 }
593 } else if ((LHSMask & BMask_Mixed) && (RHSMask & Mask_NotAllZeros)) {
594 if (Value *V = foldLogOpOfMaskedICmps_NotAllZeros_BMask_Mixed(
595 RHS, LHS, IsAnd, A, D, E, B, C,
596 PredR, PredL, Builder)) {
597 return V;
598 }
599 }
600 return nullptr;
601 }
602
603 /// Try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
604 /// into a single (icmp(A & X) ==/!= Y).
foldLogOpOfMaskedICmps(ICmpInst * LHS,ICmpInst * RHS,bool IsAnd,llvm::InstCombiner::BuilderTy & Builder)605 static Value *foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS, bool IsAnd,
606 llvm::InstCombiner::BuilderTy &Builder) {
607 Value *A = nullptr, *B = nullptr, *C = nullptr, *D = nullptr, *E = nullptr;
608 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
609 Optional<std::pair<unsigned, unsigned>> MaskPair =
610 getMaskedTypeForICmpPair(A, B, C, D, E, LHS, RHS, PredL, PredR);
611 if (!MaskPair)
612 return nullptr;
613 assert(ICmpInst::isEquality(PredL) && ICmpInst::isEquality(PredR) &&
614 "Expected equality predicates for masked type of icmps.");
615 unsigned LHSMask = MaskPair->first;
616 unsigned RHSMask = MaskPair->second;
617 unsigned Mask = LHSMask & RHSMask;
618 if (Mask == 0) {
619 // Even if the two sides don't share a common pattern, check if folding can
620 // still happen.
621 if (Value *V = foldLogOpOfMaskedICmpsAsymmetric(
622 LHS, RHS, IsAnd, A, B, C, D, E, PredL, PredR, LHSMask, RHSMask,
623 Builder))
624 return V;
625 return nullptr;
626 }
627
628 // In full generality:
629 // (icmp (A & B) Op C) | (icmp (A & D) Op E)
630 // == ![ (icmp (A & B) !Op C) & (icmp (A & D) !Op E) ]
631 //
632 // If the latter can be converted into (icmp (A & X) Op Y) then the former is
633 // equivalent to (icmp (A & X) !Op Y).
634 //
635 // Therefore, we can pretend for the rest of this function that we're dealing
636 // with the conjunction, provided we flip the sense of any comparisons (both
637 // input and output).
638
639 // In most cases we're going to produce an EQ for the "&&" case.
640 ICmpInst::Predicate NewCC = IsAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
641 if (!IsAnd) {
642 // Convert the masking analysis into its equivalent with negated
643 // comparisons.
644 Mask = conjugateICmpMask(Mask);
645 }
646
647 if (Mask & Mask_AllZeros) {
648 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
649 // -> (icmp eq (A & (B|D)), 0)
650 Value *NewOr = Builder.CreateOr(B, D);
651 Value *NewAnd = Builder.CreateAnd(A, NewOr);
652 // We can't use C as zero because we might actually handle
653 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
654 // with B and D, having a single bit set.
655 Value *Zero = Constant::getNullValue(A->getType());
656 return Builder.CreateICmp(NewCC, NewAnd, Zero);
657 }
658 if (Mask & BMask_AllOnes) {
659 // (icmp eq (A & B), B) & (icmp eq (A & D), D)
660 // -> (icmp eq (A & (B|D)), (B|D))
661 Value *NewOr = Builder.CreateOr(B, D);
662 Value *NewAnd = Builder.CreateAnd(A, NewOr);
663 return Builder.CreateICmp(NewCC, NewAnd, NewOr);
664 }
665 if (Mask & AMask_AllOnes) {
666 // (icmp eq (A & B), A) & (icmp eq (A & D), A)
667 // -> (icmp eq (A & (B&D)), A)
668 Value *NewAnd1 = Builder.CreateAnd(B, D);
669 Value *NewAnd2 = Builder.CreateAnd(A, NewAnd1);
670 return Builder.CreateICmp(NewCC, NewAnd2, A);
671 }
672
673 // Remaining cases assume at least that B and D are constant, and depend on
674 // their actual values. This isn't strictly necessary, just a "handle the
675 // easy cases for now" decision.
676 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
677 if (!BCst)
678 return nullptr;
679 ConstantInt *DCst = dyn_cast<ConstantInt>(D);
680 if (!DCst)
681 return nullptr;
682
683 if (Mask & (Mask_NotAllZeros | BMask_NotAllOnes)) {
684 // (icmp ne (A & B), 0) & (icmp ne (A & D), 0) and
685 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
686 // -> (icmp ne (A & B), 0) or (icmp ne (A & D), 0)
687 // Only valid if one of the masks is a superset of the other (check "B&D" is
688 // the same as either B or D).
689 APInt NewMask = BCst->getValue() & DCst->getValue();
690
691 if (NewMask == BCst->getValue())
692 return LHS;
693 else if (NewMask == DCst->getValue())
694 return RHS;
695 }
696
697 if (Mask & AMask_NotAllOnes) {
698 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
699 // -> (icmp ne (A & B), A) or (icmp ne (A & D), A)
700 // Only valid if one of the masks is a superset of the other (check "B|D" is
701 // the same as either B or D).
702 APInt NewMask = BCst->getValue() | DCst->getValue();
703
704 if (NewMask == BCst->getValue())
705 return LHS;
706 else if (NewMask == DCst->getValue())
707 return RHS;
708 }
709
710 if (Mask & BMask_Mixed) {
711 // (icmp eq (A & B), C) & (icmp eq (A & D), E)
712 // We already know that B & C == C && D & E == E.
713 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
714 // C and E, which are shared by both the mask B and the mask D, don't
715 // contradict, then we can transform to
716 // -> (icmp eq (A & (B|D)), (C|E))
717 // Currently, we only handle the case of B, C, D, and E being constant.
718 // We can't simply use C and E because we might actually handle
719 // (icmp ne (A & B), B) & (icmp eq (A & D), D)
720 // with B and D, having a single bit set.
721 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
722 if (!CCst)
723 return nullptr;
724 ConstantInt *ECst = dyn_cast<ConstantInt>(E);
725 if (!ECst)
726 return nullptr;
727 if (PredL != NewCC)
728 CCst = cast<ConstantInt>(ConstantExpr::getXor(BCst, CCst));
729 if (PredR != NewCC)
730 ECst = cast<ConstantInt>(ConstantExpr::getXor(DCst, ECst));
731
732 // If there is a conflict, we should actually return a false for the
733 // whole construct.
734 if (((BCst->getValue() & DCst->getValue()) &
735 (CCst->getValue() ^ ECst->getValue())).getBoolValue())
736 return ConstantInt::get(LHS->getType(), !IsAnd);
737
738 Value *NewOr1 = Builder.CreateOr(B, D);
739 Value *NewOr2 = ConstantExpr::getOr(CCst, ECst);
740 Value *NewAnd = Builder.CreateAnd(A, NewOr1);
741 return Builder.CreateICmp(NewCC, NewAnd, NewOr2);
742 }
743
744 return nullptr;
745 }
746
747 /// Try to fold a signed range checked with lower bound 0 to an unsigned icmp.
748 /// Example: (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
749 /// If \p Inverted is true then the check is for the inverted range, e.g.
750 /// (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
simplifyRangeCheck(ICmpInst * Cmp0,ICmpInst * Cmp1,bool Inverted)751 Value *InstCombiner::simplifyRangeCheck(ICmpInst *Cmp0, ICmpInst *Cmp1,
752 bool Inverted) {
753 // Check the lower range comparison, e.g. x >= 0
754 // InstCombine already ensured that if there is a constant it's on the RHS.
755 ConstantInt *RangeStart = dyn_cast<ConstantInt>(Cmp0->getOperand(1));
756 if (!RangeStart)
757 return nullptr;
758
759 ICmpInst::Predicate Pred0 = (Inverted ? Cmp0->getInversePredicate() :
760 Cmp0->getPredicate());
761
762 // Accept x > -1 or x >= 0 (after potentially inverting the predicate).
763 if (!((Pred0 == ICmpInst::ICMP_SGT && RangeStart->isMinusOne()) ||
764 (Pred0 == ICmpInst::ICMP_SGE && RangeStart->isZero())))
765 return nullptr;
766
767 ICmpInst::Predicate Pred1 = (Inverted ? Cmp1->getInversePredicate() :
768 Cmp1->getPredicate());
769
770 Value *Input = Cmp0->getOperand(0);
771 Value *RangeEnd;
772 if (Cmp1->getOperand(0) == Input) {
773 // For the upper range compare we have: icmp x, n
774 RangeEnd = Cmp1->getOperand(1);
775 } else if (Cmp1->getOperand(1) == Input) {
776 // For the upper range compare we have: icmp n, x
777 RangeEnd = Cmp1->getOperand(0);
778 Pred1 = ICmpInst::getSwappedPredicate(Pred1);
779 } else {
780 return nullptr;
781 }
782
783 // Check the upper range comparison, e.g. x < n
784 ICmpInst::Predicate NewPred;
785 switch (Pred1) {
786 case ICmpInst::ICMP_SLT: NewPred = ICmpInst::ICMP_ULT; break;
787 case ICmpInst::ICMP_SLE: NewPred = ICmpInst::ICMP_ULE; break;
788 default: return nullptr;
789 }
790
791 // This simplification is only valid if the upper range is not negative.
792 KnownBits Known = computeKnownBits(RangeEnd, /*Depth=*/0, Cmp1);
793 if (!Known.isNonNegative())
794 return nullptr;
795
796 if (Inverted)
797 NewPred = ICmpInst::getInversePredicate(NewPred);
798
799 return Builder.CreateICmp(NewPred, Input, RangeEnd);
800 }
801
802 static Value *
foldAndOrOfEqualityCmpsWithConstants(ICmpInst * LHS,ICmpInst * RHS,bool JoinedByAnd,InstCombiner::BuilderTy & Builder)803 foldAndOrOfEqualityCmpsWithConstants(ICmpInst *LHS, ICmpInst *RHS,
804 bool JoinedByAnd,
805 InstCombiner::BuilderTy &Builder) {
806 Value *X = LHS->getOperand(0);
807 if (X != RHS->getOperand(0))
808 return nullptr;
809
810 const APInt *C1, *C2;
811 if (!match(LHS->getOperand(1), m_APInt(C1)) ||
812 !match(RHS->getOperand(1), m_APInt(C2)))
813 return nullptr;
814
815 // We only handle (X != C1 && X != C2) and (X == C1 || X == C2).
816 ICmpInst::Predicate Pred = LHS->getPredicate();
817 if (Pred != RHS->getPredicate())
818 return nullptr;
819 if (JoinedByAnd && Pred != ICmpInst::ICMP_NE)
820 return nullptr;
821 if (!JoinedByAnd && Pred != ICmpInst::ICMP_EQ)
822 return nullptr;
823
824 // The larger unsigned constant goes on the right.
825 if (C1->ugt(*C2))
826 std::swap(C1, C2);
827
828 APInt Xor = *C1 ^ *C2;
829 if (Xor.isPowerOf2()) {
830 // If LHSC and RHSC differ by only one bit, then set that bit in X and
831 // compare against the larger constant:
832 // (X == C1 || X == C2) --> (X | (C1 ^ C2)) == C2
833 // (X != C1 && X != C2) --> (X | (C1 ^ C2)) != C2
834 // We choose an 'or' with a Pow2 constant rather than the inverse mask with
835 // 'and' because that may lead to smaller codegen from a smaller constant.
836 Value *Or = Builder.CreateOr(X, ConstantInt::get(X->getType(), Xor));
837 return Builder.CreateICmp(Pred, Or, ConstantInt::get(X->getType(), *C2));
838 }
839
840 // Special case: get the ordering right when the values wrap around zero.
841 // Ie, we assumed the constants were unsigned when swapping earlier.
842 if (C1->isNullValue() && C2->isAllOnesValue())
843 std::swap(C1, C2);
844
845 if (*C1 == *C2 - 1) {
846 // (X == 13 || X == 14) --> X - 13 <=u 1
847 // (X != 13 && X != 14) --> X - 13 >u 1
848 // An 'add' is the canonical IR form, so favor that over a 'sub'.
849 Value *Add = Builder.CreateAdd(X, ConstantInt::get(X->getType(), -(*C1)));
850 auto NewPred = JoinedByAnd ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_ULE;
851 return Builder.CreateICmp(NewPred, Add, ConstantInt::get(X->getType(), 1));
852 }
853
854 return nullptr;
855 }
856
857 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2)
858 // Fold (!iszero(A & K1) & !iszero(A & K2)) -> (A & (K1 | K2)) == (K1 | K2)
foldAndOrOfICmpsOfAndWithPow2(ICmpInst * LHS,ICmpInst * RHS,BinaryOperator & Logic)859 Value *InstCombiner::foldAndOrOfICmpsOfAndWithPow2(ICmpInst *LHS, ICmpInst *RHS,
860 BinaryOperator &Logic) {
861 bool JoinedByAnd = Logic.getOpcode() == Instruction::And;
862 assert((JoinedByAnd || Logic.getOpcode() == Instruction::Or) &&
863 "Wrong opcode");
864 ICmpInst::Predicate Pred = LHS->getPredicate();
865 if (Pred != RHS->getPredicate())
866 return nullptr;
867 if (JoinedByAnd && Pred != ICmpInst::ICMP_NE)
868 return nullptr;
869 if (!JoinedByAnd && Pred != ICmpInst::ICMP_EQ)
870 return nullptr;
871
872 // TODO support vector splats
873 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHS->getOperand(1));
874 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS->getOperand(1));
875 if (!LHSC || !RHSC || !LHSC->isZero() || !RHSC->isZero())
876 return nullptr;
877
878 Value *A, *B, *C, *D;
879 if (match(LHS->getOperand(0), m_And(m_Value(A), m_Value(B))) &&
880 match(RHS->getOperand(0), m_And(m_Value(C), m_Value(D)))) {
881 if (A == D || B == D)
882 std::swap(C, D);
883 if (B == C)
884 std::swap(A, B);
885
886 if (A == C &&
887 isKnownToBeAPowerOfTwo(B, false, 0, &Logic) &&
888 isKnownToBeAPowerOfTwo(D, false, 0, &Logic)) {
889 Value *Mask = Builder.CreateOr(B, D);
890 Value *Masked = Builder.CreateAnd(A, Mask);
891 auto NewPred = JoinedByAnd ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE;
892 return Builder.CreateICmp(NewPred, Masked, Mask);
893 }
894 }
895
896 return nullptr;
897 }
898
899 /// General pattern:
900 /// X & Y
901 ///
902 /// Where Y is checking that all the high bits (covered by a mask 4294967168)
903 /// are uniform, i.e. %arg & 4294967168 can be either 4294967168 or 0
904 /// Pattern can be one of:
905 /// %t = add i32 %arg, 128
906 /// %r = icmp ult i32 %t, 256
907 /// Or
908 /// %t0 = shl i32 %arg, 24
909 /// %t1 = ashr i32 %t0, 24
910 /// %r = icmp eq i32 %t1, %arg
911 /// Or
912 /// %t0 = trunc i32 %arg to i8
913 /// %t1 = sext i8 %t0 to i32
914 /// %r = icmp eq i32 %t1, %arg
915 /// This pattern is a signed truncation check.
916 ///
917 /// And X is checking that some bit in that same mask is zero.
918 /// I.e. can be one of:
919 /// %r = icmp sgt i32 %arg, -1
920 /// Or
921 /// %t = and i32 %arg, 2147483648
922 /// %r = icmp eq i32 %t, 0
923 ///
924 /// Since we are checking that all the bits in that mask are the same,
925 /// and a particular bit is zero, what we are really checking is that all the
926 /// masked bits are zero.
927 /// So this should be transformed to:
928 /// %r = icmp ult i32 %arg, 128
foldSignedTruncationCheck(ICmpInst * ICmp0,ICmpInst * ICmp1,Instruction & CxtI,InstCombiner::BuilderTy & Builder)929 static Value *foldSignedTruncationCheck(ICmpInst *ICmp0, ICmpInst *ICmp1,
930 Instruction &CxtI,
931 InstCombiner::BuilderTy &Builder) {
932 assert(CxtI.getOpcode() == Instruction::And);
933
934 // Match icmp ult (add %arg, C01), C1 (C1 == C01 << 1; powers of two)
935 auto tryToMatchSignedTruncationCheck = [](ICmpInst *ICmp, Value *&X,
936 APInt &SignBitMask) -> bool {
937 CmpInst::Predicate Pred;
938 const APInt *I01, *I1; // powers of two; I1 == I01 << 1
939 if (!(match(ICmp,
940 m_ICmp(Pred, m_Add(m_Value(X), m_Power2(I01)), m_Power2(I1))) &&
941 Pred == ICmpInst::ICMP_ULT && I1->ugt(*I01) && I01->shl(1) == *I1))
942 return false;
943 // Which bit is the new sign bit as per the 'signed truncation' pattern?
944 SignBitMask = *I01;
945 return true;
946 };
947
948 // One icmp needs to be 'signed truncation check'.
949 // We need to match this first, else we will mismatch commutative cases.
950 Value *X1;
951 APInt HighestBit;
952 ICmpInst *OtherICmp;
953 if (tryToMatchSignedTruncationCheck(ICmp1, X1, HighestBit))
954 OtherICmp = ICmp0;
955 else if (tryToMatchSignedTruncationCheck(ICmp0, X1, HighestBit))
956 OtherICmp = ICmp1;
957 else
958 return nullptr;
959
960 assert(HighestBit.isPowerOf2() && "expected to be power of two (non-zero)");
961
962 // Try to match/decompose into: icmp eq (X & Mask), 0
963 auto tryToDecompose = [](ICmpInst *ICmp, Value *&X,
964 APInt &UnsetBitsMask) -> bool {
965 CmpInst::Predicate Pred = ICmp->getPredicate();
966 // Can it be decomposed into icmp eq (X & Mask), 0 ?
967 if (llvm::decomposeBitTestICmp(ICmp->getOperand(0), ICmp->getOperand(1),
968 Pred, X, UnsetBitsMask,
969 /*LookThroughTrunc=*/false) &&
970 Pred == ICmpInst::ICMP_EQ)
971 return true;
972 // Is it icmp eq (X & Mask), 0 already?
973 const APInt *Mask;
974 if (match(ICmp, m_ICmp(Pred, m_And(m_Value(X), m_APInt(Mask)), m_Zero())) &&
975 Pred == ICmpInst::ICMP_EQ) {
976 UnsetBitsMask = *Mask;
977 return true;
978 }
979 return false;
980 };
981
982 // And the other icmp needs to be decomposable into a bit test.
983 Value *X0;
984 APInt UnsetBitsMask;
985 if (!tryToDecompose(OtherICmp, X0, UnsetBitsMask))
986 return nullptr;
987
988 assert(!UnsetBitsMask.isNullValue() && "empty mask makes no sense.");
989
990 // Are they working on the same value?
991 Value *X;
992 if (X1 == X0) {
993 // Ok as is.
994 X = X1;
995 } else if (match(X0, m_Trunc(m_Specific(X1)))) {
996 UnsetBitsMask = UnsetBitsMask.zext(X1->getType()->getScalarSizeInBits());
997 X = X1;
998 } else
999 return nullptr;
1000
1001 // So which bits should be uniform as per the 'signed truncation check'?
1002 // (all the bits starting with (i.e. including) HighestBit)
1003 APInt SignBitsMask = ~(HighestBit - 1U);
1004
1005 // UnsetBitsMask must have some common bits with SignBitsMask,
1006 if (!UnsetBitsMask.intersects(SignBitsMask))
1007 return nullptr;
1008
1009 // Does UnsetBitsMask contain any bits outside of SignBitsMask?
1010 if (!UnsetBitsMask.isSubsetOf(SignBitsMask)) {
1011 APInt OtherHighestBit = (~UnsetBitsMask) + 1U;
1012 if (!OtherHighestBit.isPowerOf2())
1013 return nullptr;
1014 HighestBit = APIntOps::umin(HighestBit, OtherHighestBit);
1015 }
1016 // Else, if it does not, then all is ok as-is.
1017
1018 // %r = icmp ult %X, SignBit
1019 return Builder.CreateICmpULT(X, ConstantInt::get(X->getType(), HighestBit),
1020 CxtI.getName() + ".simplified");
1021 }
1022
1023 /// Reduce a pair of compares that check if a value has exactly 1 bit set.
foldIsPowerOf2(ICmpInst * Cmp0,ICmpInst * Cmp1,bool JoinedByAnd,InstCombiner::BuilderTy & Builder)1024 static Value *foldIsPowerOf2(ICmpInst *Cmp0, ICmpInst *Cmp1, bool JoinedByAnd,
1025 InstCombiner::BuilderTy &Builder) {
1026 // Handle 'and' / 'or' commutation: make the equality check the first operand.
1027 if (JoinedByAnd && Cmp1->getPredicate() == ICmpInst::ICMP_NE)
1028 std::swap(Cmp0, Cmp1);
1029 else if (!JoinedByAnd && Cmp1->getPredicate() == ICmpInst::ICMP_EQ)
1030 std::swap(Cmp0, Cmp1);
1031
1032 // (X != 0) && (ctpop(X) u< 2) --> ctpop(X) == 1
1033 CmpInst::Predicate Pred0, Pred1;
1034 Value *X;
1035 if (JoinedByAnd && match(Cmp0, m_ICmp(Pred0, m_Value(X), m_ZeroInt())) &&
1036 match(Cmp1, m_ICmp(Pred1, m_Intrinsic<Intrinsic::ctpop>(m_Specific(X)),
1037 m_SpecificInt(2))) &&
1038 Pred0 == ICmpInst::ICMP_NE && Pred1 == ICmpInst::ICMP_ULT) {
1039 Value *CtPop = Cmp1->getOperand(0);
1040 return Builder.CreateICmpEQ(CtPop, ConstantInt::get(CtPop->getType(), 1));
1041 }
1042 // (X == 0) || (ctpop(X) u> 1) --> ctpop(X) != 1
1043 if (!JoinedByAnd && match(Cmp0, m_ICmp(Pred0, m_Value(X), m_ZeroInt())) &&
1044 match(Cmp1, m_ICmp(Pred1, m_Intrinsic<Intrinsic::ctpop>(m_Specific(X)),
1045 m_SpecificInt(1))) &&
1046 Pred0 == ICmpInst::ICMP_EQ && Pred1 == ICmpInst::ICMP_UGT) {
1047 Value *CtPop = Cmp1->getOperand(0);
1048 return Builder.CreateICmpNE(CtPop, ConstantInt::get(CtPop->getType(), 1));
1049 }
1050 return nullptr;
1051 }
1052
1053 /// Commuted variants are assumed to be handled by calling this function again
1054 /// with the parameters swapped.
foldUnsignedUnderflowCheck(ICmpInst * ZeroICmp,ICmpInst * UnsignedICmp,bool IsAnd,const SimplifyQuery & Q,InstCombiner::BuilderTy & Builder)1055 static Value *foldUnsignedUnderflowCheck(ICmpInst *ZeroICmp,
1056 ICmpInst *UnsignedICmp, bool IsAnd,
1057 const SimplifyQuery &Q,
1058 InstCombiner::BuilderTy &Builder) {
1059 Value *ZeroCmpOp;
1060 ICmpInst::Predicate EqPred;
1061 if (!match(ZeroICmp, m_ICmp(EqPred, m_Value(ZeroCmpOp), m_Zero())) ||
1062 !ICmpInst::isEquality(EqPred))
1063 return nullptr;
1064
1065 auto IsKnownNonZero = [&](Value *V) {
1066 return isKnownNonZero(V, Q.DL, /*Depth=*/0, Q.AC, Q.CxtI, Q.DT);
1067 };
1068
1069 ICmpInst::Predicate UnsignedPred;
1070
1071 Value *A, *B;
1072 if (match(UnsignedICmp,
1073 m_c_ICmp(UnsignedPred, m_Specific(ZeroCmpOp), m_Value(A))) &&
1074 match(ZeroCmpOp, m_c_Add(m_Specific(A), m_Value(B))) &&
1075 (ZeroICmp->hasOneUse() || UnsignedICmp->hasOneUse())) {
1076 auto GetKnownNonZeroAndOther = [&](Value *&NonZero, Value *&Other) {
1077 if (!IsKnownNonZero(NonZero))
1078 std::swap(NonZero, Other);
1079 return IsKnownNonZero(NonZero);
1080 };
1081
1082 // Given ZeroCmpOp = (A + B)
1083 // ZeroCmpOp <= A && ZeroCmpOp != 0 --> (0-B) < A
1084 // ZeroCmpOp > A || ZeroCmpOp == 0 --> (0-B) >= A
1085 //
1086 // ZeroCmpOp < A && ZeroCmpOp != 0 --> (0-X) < Y iff
1087 // ZeroCmpOp >= A || ZeroCmpOp == 0 --> (0-X) >= Y iff
1088 // with X being the value (A/B) that is known to be non-zero,
1089 // and Y being remaining value.
1090 if (UnsignedPred == ICmpInst::ICMP_ULE && EqPred == ICmpInst::ICMP_NE &&
1091 IsAnd)
1092 return Builder.CreateICmpULT(Builder.CreateNeg(B), A);
1093 if (UnsignedPred == ICmpInst::ICMP_ULT && EqPred == ICmpInst::ICMP_NE &&
1094 IsAnd && GetKnownNonZeroAndOther(B, A))
1095 return Builder.CreateICmpULT(Builder.CreateNeg(B), A);
1096 if (UnsignedPred == ICmpInst::ICMP_UGT && EqPred == ICmpInst::ICMP_EQ &&
1097 !IsAnd)
1098 return Builder.CreateICmpUGE(Builder.CreateNeg(B), A);
1099 if (UnsignedPred == ICmpInst::ICMP_UGE && EqPred == ICmpInst::ICMP_EQ &&
1100 !IsAnd && GetKnownNonZeroAndOther(B, A))
1101 return Builder.CreateICmpUGE(Builder.CreateNeg(B), A);
1102 }
1103
1104 Value *Base, *Offset;
1105 if (!match(ZeroCmpOp, m_Sub(m_Value(Base), m_Value(Offset))))
1106 return nullptr;
1107
1108 if (!match(UnsignedICmp,
1109 m_c_ICmp(UnsignedPred, m_Specific(Base), m_Specific(Offset))) ||
1110 !ICmpInst::isUnsigned(UnsignedPred))
1111 return nullptr;
1112
1113 // Base >=/> Offset && (Base - Offset) != 0 <--> Base > Offset
1114 // (no overflow and not null)
1115 if ((UnsignedPred == ICmpInst::ICMP_UGE ||
1116 UnsignedPred == ICmpInst::ICMP_UGT) &&
1117 EqPred == ICmpInst::ICMP_NE && IsAnd)
1118 return Builder.CreateICmpUGT(Base, Offset);
1119
1120 // Base <=/< Offset || (Base - Offset) == 0 <--> Base <= Offset
1121 // (overflow or null)
1122 if ((UnsignedPred == ICmpInst::ICMP_ULE ||
1123 UnsignedPred == ICmpInst::ICMP_ULT) &&
1124 EqPred == ICmpInst::ICMP_EQ && !IsAnd)
1125 return Builder.CreateICmpULE(Base, Offset);
1126
1127 // Base <= Offset && (Base - Offset) != 0 --> Base < Offset
1128 if (UnsignedPred == ICmpInst::ICMP_ULE && EqPred == ICmpInst::ICMP_NE &&
1129 IsAnd)
1130 return Builder.CreateICmpULT(Base, Offset);
1131
1132 // Base > Offset || (Base - Offset) == 0 --> Base >= Offset
1133 if (UnsignedPred == ICmpInst::ICMP_UGT && EqPred == ICmpInst::ICMP_EQ &&
1134 !IsAnd)
1135 return Builder.CreateICmpUGE(Base, Offset);
1136
1137 return nullptr;
1138 }
1139
1140 /// Reduce logic-of-compares with equality to a constant by substituting a
1141 /// common operand with the constant. Callers are expected to call this with
1142 /// Cmp0/Cmp1 switched to handle logic op commutativity.
foldAndOrOfICmpsWithConstEq(ICmpInst * Cmp0,ICmpInst * Cmp1,BinaryOperator & Logic,InstCombiner::BuilderTy & Builder,const SimplifyQuery & Q)1143 static Value *foldAndOrOfICmpsWithConstEq(ICmpInst *Cmp0, ICmpInst *Cmp1,
1144 BinaryOperator &Logic,
1145 InstCombiner::BuilderTy &Builder,
1146 const SimplifyQuery &Q) {
1147 bool IsAnd = Logic.getOpcode() == Instruction::And;
1148 assert((IsAnd || Logic.getOpcode() == Instruction::Or) && "Wrong logic op");
1149
1150 // Match an equality compare with a non-poison constant as Cmp0.
1151 ICmpInst::Predicate Pred0;
1152 Value *X;
1153 Constant *C;
1154 if (!match(Cmp0, m_ICmp(Pred0, m_Value(X), m_Constant(C))) ||
1155 !isGuaranteedNotToBeUndefOrPoison(C))
1156 return nullptr;
1157 if ((IsAnd && Pred0 != ICmpInst::ICMP_EQ) ||
1158 (!IsAnd && Pred0 != ICmpInst::ICMP_NE))
1159 return nullptr;
1160
1161 // The other compare must include a common operand (X). Canonicalize the
1162 // common operand as operand 1 (Pred1 is swapped if the common operand was
1163 // operand 0).
1164 Value *Y;
1165 ICmpInst::Predicate Pred1;
1166 if (!match(Cmp1, m_c_ICmp(Pred1, m_Value(Y), m_Deferred(X))))
1167 return nullptr;
1168
1169 // Replace variable with constant value equivalence to remove a variable use:
1170 // (X == C) && (Y Pred1 X) --> (X == C) && (Y Pred1 C)
1171 // (X != C) || (Y Pred1 X) --> (X != C) || (Y Pred1 C)
1172 // Can think of the 'or' substitution with the 'and' bool equivalent:
1173 // A || B --> A || (!A && B)
1174 Value *SubstituteCmp = SimplifyICmpInst(Pred1, Y, C, Q);
1175 if (!SubstituteCmp) {
1176 // If we need to create a new instruction, require that the old compare can
1177 // be removed.
1178 if (!Cmp1->hasOneUse())
1179 return nullptr;
1180 SubstituteCmp = Builder.CreateICmp(Pred1, Y, C);
1181 }
1182 return Builder.CreateBinOp(Logic.getOpcode(), Cmp0, SubstituteCmp);
1183 }
1184
1185 /// Fold (icmp)&(icmp) if possible.
foldAndOfICmps(ICmpInst * LHS,ICmpInst * RHS,BinaryOperator & And)1186 Value *InstCombiner::foldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS,
1187 BinaryOperator &And) {
1188 const SimplifyQuery Q = SQ.getWithInstruction(&And);
1189
1190 // Fold (!iszero(A & K1) & !iszero(A & K2)) -> (A & (K1 | K2)) == (K1 | K2)
1191 // if K1 and K2 are a one-bit mask.
1192 if (Value *V = foldAndOrOfICmpsOfAndWithPow2(LHS, RHS, And))
1193 return V;
1194
1195 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1196
1197 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
1198 if (predicatesFoldable(PredL, PredR)) {
1199 if (LHS->getOperand(0) == RHS->getOperand(1) &&
1200 LHS->getOperand(1) == RHS->getOperand(0))
1201 LHS->swapOperands();
1202 if (LHS->getOperand(0) == RHS->getOperand(0) &&
1203 LHS->getOperand(1) == RHS->getOperand(1)) {
1204 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1205 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
1206 bool IsSigned = LHS->isSigned() || RHS->isSigned();
1207 return getNewICmpValue(Code, IsSigned, Op0, Op1, Builder);
1208 }
1209 }
1210
1211 // handle (roughly): (icmp eq (A & B), C) & (icmp eq (A & D), E)
1212 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, true, Builder))
1213 return V;
1214
1215 if (Value *V = foldAndOrOfICmpsWithConstEq(LHS, RHS, And, Builder, Q))
1216 return V;
1217 if (Value *V = foldAndOrOfICmpsWithConstEq(RHS, LHS, And, Builder, Q))
1218 return V;
1219
1220 // E.g. (icmp sge x, 0) & (icmp slt x, n) --> icmp ult x, n
1221 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/false))
1222 return V;
1223
1224 // E.g. (icmp slt x, n) & (icmp sge x, 0) --> icmp ult x, n
1225 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/false))
1226 return V;
1227
1228 if (Value *V = foldAndOrOfEqualityCmpsWithConstants(LHS, RHS, true, Builder))
1229 return V;
1230
1231 if (Value *V = foldSignedTruncationCheck(LHS, RHS, And, Builder))
1232 return V;
1233
1234 if (Value *V = foldIsPowerOf2(LHS, RHS, true /* JoinedByAnd */, Builder))
1235 return V;
1236
1237 if (Value *X =
1238 foldUnsignedUnderflowCheck(LHS, RHS, /*IsAnd=*/true, Q, Builder))
1239 return X;
1240 if (Value *X =
1241 foldUnsignedUnderflowCheck(RHS, LHS, /*IsAnd=*/true, Q, Builder))
1242 return X;
1243
1244 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
1245 Value *LHS0 = LHS->getOperand(0), *RHS0 = RHS->getOperand(0);
1246 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHS->getOperand(1));
1247 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS->getOperand(1));
1248 if (!LHSC || !RHSC)
1249 return nullptr;
1250
1251 if (LHSC == RHSC && PredL == PredR) {
1252 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
1253 // where C is a power of 2 or
1254 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
1255 if ((PredL == ICmpInst::ICMP_ULT && LHSC->getValue().isPowerOf2()) ||
1256 (PredL == ICmpInst::ICMP_EQ && LHSC->isZero())) {
1257 Value *NewOr = Builder.CreateOr(LHS0, RHS0);
1258 return Builder.CreateICmp(PredL, NewOr, LHSC);
1259 }
1260 }
1261
1262 // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2
1263 // where CMAX is the all ones value for the truncated type,
1264 // iff the lower bits of C2 and CA are zero.
1265 if (PredL == ICmpInst::ICMP_EQ && PredL == PredR && LHS->hasOneUse() &&
1266 RHS->hasOneUse()) {
1267 Value *V;
1268 ConstantInt *AndC, *SmallC = nullptr, *BigC = nullptr;
1269
1270 // (trunc x) == C1 & (and x, CA) == C2
1271 // (and x, CA) == C2 & (trunc x) == C1
1272 if (match(RHS0, m_Trunc(m_Value(V))) &&
1273 match(LHS0, m_And(m_Specific(V), m_ConstantInt(AndC)))) {
1274 SmallC = RHSC;
1275 BigC = LHSC;
1276 } else if (match(LHS0, m_Trunc(m_Value(V))) &&
1277 match(RHS0, m_And(m_Specific(V), m_ConstantInt(AndC)))) {
1278 SmallC = LHSC;
1279 BigC = RHSC;
1280 }
1281
1282 if (SmallC && BigC) {
1283 unsigned BigBitSize = BigC->getType()->getBitWidth();
1284 unsigned SmallBitSize = SmallC->getType()->getBitWidth();
1285
1286 // Check that the low bits are zero.
1287 APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize);
1288 if ((Low & AndC->getValue()).isNullValue() &&
1289 (Low & BigC->getValue()).isNullValue()) {
1290 Value *NewAnd = Builder.CreateAnd(V, Low | AndC->getValue());
1291 APInt N = SmallC->getValue().zext(BigBitSize) | BigC->getValue();
1292 Value *NewVal = ConstantInt::get(AndC->getType()->getContext(), N);
1293 return Builder.CreateICmp(PredL, NewAnd, NewVal);
1294 }
1295 }
1296 }
1297
1298 // From here on, we only handle:
1299 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
1300 if (LHS0 != RHS0)
1301 return nullptr;
1302
1303 // ICMP_[US][GL]E X, C is folded to ICMP_[US][GL]T elsewhere.
1304 if (PredL == ICmpInst::ICMP_UGE || PredL == ICmpInst::ICMP_ULE ||
1305 PredR == ICmpInst::ICMP_UGE || PredR == ICmpInst::ICMP_ULE ||
1306 PredL == ICmpInst::ICMP_SGE || PredL == ICmpInst::ICMP_SLE ||
1307 PredR == ICmpInst::ICMP_SGE || PredR == ICmpInst::ICMP_SLE)
1308 return nullptr;
1309
1310 // We can't fold (ugt x, C) & (sgt x, C2).
1311 if (!predicatesFoldable(PredL, PredR))
1312 return nullptr;
1313
1314 // Ensure that the larger constant is on the RHS.
1315 bool ShouldSwap;
1316 if (CmpInst::isSigned(PredL) ||
1317 (ICmpInst::isEquality(PredL) && CmpInst::isSigned(PredR)))
1318 ShouldSwap = LHSC->getValue().sgt(RHSC->getValue());
1319 else
1320 ShouldSwap = LHSC->getValue().ugt(RHSC->getValue());
1321
1322 if (ShouldSwap) {
1323 std::swap(LHS, RHS);
1324 std::swap(LHSC, RHSC);
1325 std::swap(PredL, PredR);
1326 }
1327
1328 // At this point, we know we have two icmp instructions
1329 // comparing a value against two constants and and'ing the result
1330 // together. Because of the above check, we know that we only have
1331 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
1332 // (from the icmp folding check above), that the two constants
1333 // are not equal and that the larger constant is on the RHS
1334 assert(LHSC != RHSC && "Compares not folded above?");
1335
1336 switch (PredL) {
1337 default:
1338 llvm_unreachable("Unknown integer condition code!");
1339 case ICmpInst::ICMP_NE:
1340 switch (PredR) {
1341 default:
1342 llvm_unreachable("Unknown integer condition code!");
1343 case ICmpInst::ICMP_ULT:
1344 // (X != 13 & X u< 14) -> X < 13
1345 if (LHSC->getValue() == (RHSC->getValue() - 1))
1346 return Builder.CreateICmpULT(LHS0, LHSC);
1347 if (LHSC->isZero()) // (X != 0 & X u< C) -> X-1 u< C-1
1348 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(),
1349 false, true);
1350 break; // (X != 13 & X u< 15) -> no change
1351 case ICmpInst::ICMP_SLT:
1352 // (X != 13 & X s< 14) -> X < 13
1353 if (LHSC->getValue() == (RHSC->getValue() - 1))
1354 return Builder.CreateICmpSLT(LHS0, LHSC);
1355 // (X != INT_MIN & X s< C) -> X-(INT_MIN+1) u< (C-(INT_MIN+1))
1356 if (LHSC->isMinValue(true))
1357 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(),
1358 true, true);
1359 break; // (X != 13 & X s< 15) -> no change
1360 case ICmpInst::ICMP_NE:
1361 // Potential folds for this case should already be handled.
1362 break;
1363 }
1364 break;
1365 case ICmpInst::ICMP_UGT:
1366 switch (PredR) {
1367 default:
1368 llvm_unreachable("Unknown integer condition code!");
1369 case ICmpInst::ICMP_NE:
1370 // (X u> 13 & X != 14) -> X u> 14
1371 if (RHSC->getValue() == (LHSC->getValue() + 1))
1372 return Builder.CreateICmp(PredL, LHS0, RHSC);
1373 // X u> C & X != UINT_MAX -> (X-(C+1)) u< UINT_MAX-(C+1)
1374 if (RHSC->isMaxValue(false))
1375 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(),
1376 false, true);
1377 break; // (X u> 13 & X != 15) -> no change
1378 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) u< 1
1379 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(),
1380 false, true);
1381 }
1382 break;
1383 case ICmpInst::ICMP_SGT:
1384 switch (PredR) {
1385 default:
1386 llvm_unreachable("Unknown integer condition code!");
1387 case ICmpInst::ICMP_NE:
1388 // (X s> 13 & X != 14) -> X s> 14
1389 if (RHSC->getValue() == (LHSC->getValue() + 1))
1390 return Builder.CreateICmp(PredL, LHS0, RHSC);
1391 // X s> C & X != INT_MAX -> (X-(C+1)) u< INT_MAX-(C+1)
1392 if (RHSC->isMaxValue(true))
1393 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(),
1394 true, true);
1395 break; // (X s> 13 & X != 15) -> no change
1396 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) u< 1
1397 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue(), true,
1398 true);
1399 }
1400 break;
1401 }
1402
1403 return nullptr;
1404 }
1405
foldLogicOfFCmps(FCmpInst * LHS,FCmpInst * RHS,bool IsAnd)1406 Value *InstCombiner::foldLogicOfFCmps(FCmpInst *LHS, FCmpInst *RHS, bool IsAnd) {
1407 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
1408 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
1409 FCmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
1410
1411 if (LHS0 == RHS1 && RHS0 == LHS1) {
1412 // Swap RHS operands to match LHS.
1413 PredR = FCmpInst::getSwappedPredicate(PredR);
1414 std::swap(RHS0, RHS1);
1415 }
1416
1417 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
1418 // Suppose the relation between x and y is R, where R is one of
1419 // U(1000), L(0100), G(0010) or E(0001), and CC0 and CC1 are the bitmasks for
1420 // testing the desired relations.
1421 //
1422 // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this:
1423 // bool(R & CC0) && bool(R & CC1)
1424 // = bool((R & CC0) & (R & CC1))
1425 // = bool(R & (CC0 & CC1)) <= by re-association, commutation, and idempotency
1426 //
1427 // Since (R & CC0) and (R & CC1) are either R or 0, we actually have this:
1428 // bool(R & CC0) || bool(R & CC1)
1429 // = bool((R & CC0) | (R & CC1))
1430 // = bool(R & (CC0 | CC1)) <= by reversed distribution (contribution? ;)
1431 if (LHS0 == RHS0 && LHS1 == RHS1) {
1432 unsigned FCmpCodeL = getFCmpCode(PredL);
1433 unsigned FCmpCodeR = getFCmpCode(PredR);
1434 unsigned NewPred = IsAnd ? FCmpCodeL & FCmpCodeR : FCmpCodeL | FCmpCodeR;
1435 return getFCmpValue(NewPred, LHS0, LHS1, Builder);
1436 }
1437
1438 if ((PredL == FCmpInst::FCMP_ORD && PredR == FCmpInst::FCMP_ORD && IsAnd) ||
1439 (PredL == FCmpInst::FCMP_UNO && PredR == FCmpInst::FCMP_UNO && !IsAnd)) {
1440 if (LHS0->getType() != RHS0->getType())
1441 return nullptr;
1442
1443 // FCmp canonicalization ensures that (fcmp ord/uno X, X) and
1444 // (fcmp ord/uno X, C) will be transformed to (fcmp X, +0.0).
1445 if (match(LHS1, m_PosZeroFP()) && match(RHS1, m_PosZeroFP()))
1446 // Ignore the constants because they are obviously not NANs:
1447 // (fcmp ord x, 0.0) & (fcmp ord y, 0.0) -> (fcmp ord x, y)
1448 // (fcmp uno x, 0.0) | (fcmp uno y, 0.0) -> (fcmp uno x, y)
1449 return Builder.CreateFCmp(PredL, LHS0, RHS0);
1450 }
1451
1452 return nullptr;
1453 }
1454
1455 /// This a limited reassociation for a special case (see above) where we are
1456 /// checking if two values are either both NAN (unordered) or not-NAN (ordered).
1457 /// This could be handled more generally in '-reassociation', but it seems like
1458 /// an unlikely pattern for a large number of logic ops and fcmps.
reassociateFCmps(BinaryOperator & BO,InstCombiner::BuilderTy & Builder)1459 static Instruction *reassociateFCmps(BinaryOperator &BO,
1460 InstCombiner::BuilderTy &Builder) {
1461 Instruction::BinaryOps Opcode = BO.getOpcode();
1462 assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&
1463 "Expecting and/or op for fcmp transform");
1464
1465 // There are 4 commuted variants of the pattern. Canonicalize operands of this
1466 // logic op so an fcmp is operand 0 and a matching logic op is operand 1.
1467 Value *Op0 = BO.getOperand(0), *Op1 = BO.getOperand(1), *X;
1468 FCmpInst::Predicate Pred;
1469 if (match(Op1, m_FCmp(Pred, m_Value(), m_AnyZeroFP())))
1470 std::swap(Op0, Op1);
1471
1472 // Match inner binop and the predicate for combining 2 NAN checks into 1.
1473 BinaryOperator *BO1;
1474 FCmpInst::Predicate NanPred = Opcode == Instruction::And ? FCmpInst::FCMP_ORD
1475 : FCmpInst::FCMP_UNO;
1476 if (!match(Op0, m_FCmp(Pred, m_Value(X), m_AnyZeroFP())) || Pred != NanPred ||
1477 !match(Op1, m_BinOp(BO1)) || BO1->getOpcode() != Opcode)
1478 return nullptr;
1479
1480 // The inner logic op must have a matching fcmp operand.
1481 Value *BO10 = BO1->getOperand(0), *BO11 = BO1->getOperand(1), *Y;
1482 if (!match(BO10, m_FCmp(Pred, m_Value(Y), m_AnyZeroFP())) ||
1483 Pred != NanPred || X->getType() != Y->getType())
1484 std::swap(BO10, BO11);
1485
1486 if (!match(BO10, m_FCmp(Pred, m_Value(Y), m_AnyZeroFP())) ||
1487 Pred != NanPred || X->getType() != Y->getType())
1488 return nullptr;
1489
1490 // and (fcmp ord X, 0), (and (fcmp ord Y, 0), Z) --> and (fcmp ord X, Y), Z
1491 // or (fcmp uno X, 0), (or (fcmp uno Y, 0), Z) --> or (fcmp uno X, Y), Z
1492 Value *NewFCmp = Builder.CreateFCmp(Pred, X, Y);
1493 if (auto *NewFCmpInst = dyn_cast<FCmpInst>(NewFCmp)) {
1494 // Intersect FMF from the 2 source fcmps.
1495 NewFCmpInst->copyIRFlags(Op0);
1496 NewFCmpInst->andIRFlags(BO10);
1497 }
1498 return BinaryOperator::Create(Opcode, NewFCmp, BO11);
1499 }
1500
1501 /// Match De Morgan's Laws:
1502 /// (~A & ~B) == (~(A | B))
1503 /// (~A | ~B) == (~(A & B))
matchDeMorgansLaws(BinaryOperator & I,InstCombiner::BuilderTy & Builder)1504 static Instruction *matchDeMorgansLaws(BinaryOperator &I,
1505 InstCombiner::BuilderTy &Builder) {
1506 auto Opcode = I.getOpcode();
1507 assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&
1508 "Trying to match De Morgan's Laws with something other than and/or");
1509
1510 // Flip the logic operation.
1511 Opcode = (Opcode == Instruction::And) ? Instruction::Or : Instruction::And;
1512
1513 Value *A, *B;
1514 if (match(I.getOperand(0), m_OneUse(m_Not(m_Value(A)))) &&
1515 match(I.getOperand(1), m_OneUse(m_Not(m_Value(B)))) &&
1516 !isFreeToInvert(A, A->hasOneUse()) &&
1517 !isFreeToInvert(B, B->hasOneUse())) {
1518 Value *AndOr = Builder.CreateBinOp(Opcode, A, B, I.getName() + ".demorgan");
1519 return BinaryOperator::CreateNot(AndOr);
1520 }
1521
1522 return nullptr;
1523 }
1524
shouldOptimizeCast(CastInst * CI)1525 bool InstCombiner::shouldOptimizeCast(CastInst *CI) {
1526 Value *CastSrc = CI->getOperand(0);
1527
1528 // Noop casts and casts of constants should be eliminated trivially.
1529 if (CI->getSrcTy() == CI->getDestTy() || isa<Constant>(CastSrc))
1530 return false;
1531
1532 // If this cast is paired with another cast that can be eliminated, we prefer
1533 // to have it eliminated.
1534 if (const auto *PrecedingCI = dyn_cast<CastInst>(CastSrc))
1535 if (isEliminableCastPair(PrecedingCI, CI))
1536 return false;
1537
1538 return true;
1539 }
1540
1541 /// Fold {and,or,xor} (cast X), C.
foldLogicCastConstant(BinaryOperator & Logic,CastInst * Cast,InstCombiner::BuilderTy & Builder)1542 static Instruction *foldLogicCastConstant(BinaryOperator &Logic, CastInst *Cast,
1543 InstCombiner::BuilderTy &Builder) {
1544 Constant *C = dyn_cast<Constant>(Logic.getOperand(1));
1545 if (!C)
1546 return nullptr;
1547
1548 auto LogicOpc = Logic.getOpcode();
1549 Type *DestTy = Logic.getType();
1550 Type *SrcTy = Cast->getSrcTy();
1551
1552 // Move the logic operation ahead of a zext or sext if the constant is
1553 // unchanged in the smaller source type. Performing the logic in a smaller
1554 // type may provide more information to later folds, and the smaller logic
1555 // instruction may be cheaper (particularly in the case of vectors).
1556 Value *X;
1557 if (match(Cast, m_OneUse(m_ZExt(m_Value(X))))) {
1558 Constant *TruncC = ConstantExpr::getTrunc(C, SrcTy);
1559 Constant *ZextTruncC = ConstantExpr::getZExt(TruncC, DestTy);
1560 if (ZextTruncC == C) {
1561 // LogicOpc (zext X), C --> zext (LogicOpc X, C)
1562 Value *NewOp = Builder.CreateBinOp(LogicOpc, X, TruncC);
1563 return new ZExtInst(NewOp, DestTy);
1564 }
1565 }
1566
1567 if (match(Cast, m_OneUse(m_SExt(m_Value(X))))) {
1568 Constant *TruncC = ConstantExpr::getTrunc(C, SrcTy);
1569 Constant *SextTruncC = ConstantExpr::getSExt(TruncC, DestTy);
1570 if (SextTruncC == C) {
1571 // LogicOpc (sext X), C --> sext (LogicOpc X, C)
1572 Value *NewOp = Builder.CreateBinOp(LogicOpc, X, TruncC);
1573 return new SExtInst(NewOp, DestTy);
1574 }
1575 }
1576
1577 return nullptr;
1578 }
1579
1580 /// Fold {and,or,xor} (cast X), Y.
foldCastedBitwiseLogic(BinaryOperator & I)1581 Instruction *InstCombiner::foldCastedBitwiseLogic(BinaryOperator &I) {
1582 auto LogicOpc = I.getOpcode();
1583 assert(I.isBitwiseLogicOp() && "Unexpected opcode for bitwise logic folding");
1584
1585 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1586 CastInst *Cast0 = dyn_cast<CastInst>(Op0);
1587 if (!Cast0)
1588 return nullptr;
1589
1590 // This must be a cast from an integer or integer vector source type to allow
1591 // transformation of the logic operation to the source type.
1592 Type *DestTy = I.getType();
1593 Type *SrcTy = Cast0->getSrcTy();
1594 if (!SrcTy->isIntOrIntVectorTy())
1595 return nullptr;
1596
1597 if (Instruction *Ret = foldLogicCastConstant(I, Cast0, Builder))
1598 return Ret;
1599
1600 CastInst *Cast1 = dyn_cast<CastInst>(Op1);
1601 if (!Cast1)
1602 return nullptr;
1603
1604 // Both operands of the logic operation are casts. The casts must be of the
1605 // same type for reduction.
1606 auto CastOpcode = Cast0->getOpcode();
1607 if (CastOpcode != Cast1->getOpcode() || SrcTy != Cast1->getSrcTy())
1608 return nullptr;
1609
1610 Value *Cast0Src = Cast0->getOperand(0);
1611 Value *Cast1Src = Cast1->getOperand(0);
1612
1613 // fold logic(cast(A), cast(B)) -> cast(logic(A, B))
1614 if (shouldOptimizeCast(Cast0) && shouldOptimizeCast(Cast1)) {
1615 Value *NewOp = Builder.CreateBinOp(LogicOpc, Cast0Src, Cast1Src,
1616 I.getName());
1617 return CastInst::Create(CastOpcode, NewOp, DestTy);
1618 }
1619
1620 // For now, only 'and'/'or' have optimizations after this.
1621 if (LogicOpc == Instruction::Xor)
1622 return nullptr;
1623
1624 // If this is logic(cast(icmp), cast(icmp)), try to fold this even if the
1625 // cast is otherwise not optimizable. This happens for vector sexts.
1626 ICmpInst *ICmp0 = dyn_cast<ICmpInst>(Cast0Src);
1627 ICmpInst *ICmp1 = dyn_cast<ICmpInst>(Cast1Src);
1628 if (ICmp0 && ICmp1) {
1629 Value *Res = LogicOpc == Instruction::And ? foldAndOfICmps(ICmp0, ICmp1, I)
1630 : foldOrOfICmps(ICmp0, ICmp1, I);
1631 if (Res)
1632 return CastInst::Create(CastOpcode, Res, DestTy);
1633 return nullptr;
1634 }
1635
1636 // If this is logic(cast(fcmp), cast(fcmp)), try to fold this even if the
1637 // cast is otherwise not optimizable. This happens for vector sexts.
1638 FCmpInst *FCmp0 = dyn_cast<FCmpInst>(Cast0Src);
1639 FCmpInst *FCmp1 = dyn_cast<FCmpInst>(Cast1Src);
1640 if (FCmp0 && FCmp1)
1641 if (Value *R = foldLogicOfFCmps(FCmp0, FCmp1, LogicOpc == Instruction::And))
1642 return CastInst::Create(CastOpcode, R, DestTy);
1643
1644 return nullptr;
1645 }
1646
foldAndToXor(BinaryOperator & I,InstCombiner::BuilderTy & Builder)1647 static Instruction *foldAndToXor(BinaryOperator &I,
1648 InstCombiner::BuilderTy &Builder) {
1649 assert(I.getOpcode() == Instruction::And);
1650 Value *Op0 = I.getOperand(0);
1651 Value *Op1 = I.getOperand(1);
1652 Value *A, *B;
1653
1654 // Operand complexity canonicalization guarantees that the 'or' is Op0.
1655 // (A | B) & ~(A & B) --> A ^ B
1656 // (A | B) & ~(B & A) --> A ^ B
1657 if (match(&I, m_BinOp(m_Or(m_Value(A), m_Value(B)),
1658 m_Not(m_c_And(m_Deferred(A), m_Deferred(B))))))
1659 return BinaryOperator::CreateXor(A, B);
1660
1661 // (A | ~B) & (~A | B) --> ~(A ^ B)
1662 // (A | ~B) & (B | ~A) --> ~(A ^ B)
1663 // (~B | A) & (~A | B) --> ~(A ^ B)
1664 // (~B | A) & (B | ~A) --> ~(A ^ B)
1665 if (Op0->hasOneUse() || Op1->hasOneUse())
1666 if (match(&I, m_BinOp(m_c_Or(m_Value(A), m_Not(m_Value(B))),
1667 m_c_Or(m_Not(m_Deferred(A)), m_Deferred(B)))))
1668 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
1669
1670 return nullptr;
1671 }
1672
foldOrToXor(BinaryOperator & I,InstCombiner::BuilderTy & Builder)1673 static Instruction *foldOrToXor(BinaryOperator &I,
1674 InstCombiner::BuilderTy &Builder) {
1675 assert(I.getOpcode() == Instruction::Or);
1676 Value *Op0 = I.getOperand(0);
1677 Value *Op1 = I.getOperand(1);
1678 Value *A, *B;
1679
1680 // Operand complexity canonicalization guarantees that the 'and' is Op0.
1681 // (A & B) | ~(A | B) --> ~(A ^ B)
1682 // (A & B) | ~(B | A) --> ~(A ^ B)
1683 if (Op0->hasOneUse() || Op1->hasOneUse())
1684 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
1685 match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B)))))
1686 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
1687
1688 // (A & ~B) | (~A & B) --> A ^ B
1689 // (A & ~B) | (B & ~A) --> A ^ B
1690 // (~B & A) | (~A & B) --> A ^ B
1691 // (~B & A) | (B & ~A) --> A ^ B
1692 if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) &&
1693 match(Op1, m_c_And(m_Not(m_Specific(A)), m_Specific(B))))
1694 return BinaryOperator::CreateXor(A, B);
1695
1696 return nullptr;
1697 }
1698
1699 /// Return true if a constant shift amount is always less than the specified
1700 /// bit-width. If not, the shift could create poison in the narrower type.
canNarrowShiftAmt(Constant * C,unsigned BitWidth)1701 static bool canNarrowShiftAmt(Constant *C, unsigned BitWidth) {
1702 if (auto *ScalarC = dyn_cast<ConstantInt>(C))
1703 return ScalarC->getZExtValue() < BitWidth;
1704
1705 if (C->getType()->isVectorTy()) {
1706 // Check each element of a constant vector.
1707 unsigned NumElts = cast<VectorType>(C->getType())->getNumElements();
1708 for (unsigned i = 0; i != NumElts; ++i) {
1709 Constant *Elt = C->getAggregateElement(i);
1710 if (!Elt)
1711 return false;
1712 if (isa<UndefValue>(Elt))
1713 continue;
1714 auto *CI = dyn_cast<ConstantInt>(Elt);
1715 if (!CI || CI->getZExtValue() >= BitWidth)
1716 return false;
1717 }
1718 return true;
1719 }
1720
1721 // The constant is a constant expression or unknown.
1722 return false;
1723 }
1724
1725 /// Try to use narrower ops (sink zext ops) for an 'and' with binop operand and
1726 /// a common zext operand: and (binop (zext X), C), (zext X).
narrowMaskedBinOp(BinaryOperator & And)1727 Instruction *InstCombiner::narrowMaskedBinOp(BinaryOperator &And) {
1728 // This transform could also apply to {or, and, xor}, but there are better
1729 // folds for those cases, so we don't expect those patterns here. AShr is not
1730 // handled because it should always be transformed to LShr in this sequence.
1731 // The subtract transform is different because it has a constant on the left.
1732 // Add/mul commute the constant to RHS; sub with constant RHS becomes add.
1733 Value *Op0 = And.getOperand(0), *Op1 = And.getOperand(1);
1734 Constant *C;
1735 if (!match(Op0, m_OneUse(m_Add(m_Specific(Op1), m_Constant(C)))) &&
1736 !match(Op0, m_OneUse(m_Mul(m_Specific(Op1), m_Constant(C)))) &&
1737 !match(Op0, m_OneUse(m_LShr(m_Specific(Op1), m_Constant(C)))) &&
1738 !match(Op0, m_OneUse(m_Shl(m_Specific(Op1), m_Constant(C)))) &&
1739 !match(Op0, m_OneUse(m_Sub(m_Constant(C), m_Specific(Op1)))))
1740 return nullptr;
1741
1742 Value *X;
1743 if (!match(Op1, m_ZExt(m_Value(X))) || Op1->hasNUsesOrMore(3))
1744 return nullptr;
1745
1746 Type *Ty = And.getType();
1747 if (!isa<VectorType>(Ty) && !shouldChangeType(Ty, X->getType()))
1748 return nullptr;
1749
1750 // If we're narrowing a shift, the shift amount must be safe (less than the
1751 // width) in the narrower type. If the shift amount is greater, instsimplify
1752 // usually handles that case, but we can't guarantee/assert it.
1753 Instruction::BinaryOps Opc = cast<BinaryOperator>(Op0)->getOpcode();
1754 if (Opc == Instruction::LShr || Opc == Instruction::Shl)
1755 if (!canNarrowShiftAmt(C, X->getType()->getScalarSizeInBits()))
1756 return nullptr;
1757
1758 // and (sub C, (zext X)), (zext X) --> zext (and (sub C', X), X)
1759 // and (binop (zext X), C), (zext X) --> zext (and (binop X, C'), X)
1760 Value *NewC = ConstantExpr::getTrunc(C, X->getType());
1761 Value *NewBO = Opc == Instruction::Sub ? Builder.CreateBinOp(Opc, NewC, X)
1762 : Builder.CreateBinOp(Opc, X, NewC);
1763 return new ZExtInst(Builder.CreateAnd(NewBO, X), Ty);
1764 }
1765
1766 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
1767 // here. We should standardize that construct where it is needed or choose some
1768 // other way to ensure that commutated variants of patterns are not missed.
visitAnd(BinaryOperator & I)1769 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
1770 if (Value *V = SimplifyAndInst(I.getOperand(0), I.getOperand(1),
1771 SQ.getWithInstruction(&I)))
1772 return replaceInstUsesWith(I, V);
1773
1774 if (SimplifyAssociativeOrCommutative(I))
1775 return &I;
1776
1777 if (Instruction *X = foldVectorBinop(I))
1778 return X;
1779
1780 // See if we can simplify any instructions used by the instruction whose sole
1781 // purpose is to compute bits we don't care about.
1782 if (SimplifyDemandedInstructionBits(I))
1783 return &I;
1784
1785 // Do this before using distributive laws to catch simple and/or/not patterns.
1786 if (Instruction *Xor = foldAndToXor(I, Builder))
1787 return Xor;
1788
1789 // (A|B)&(A|C) -> A|(B&C) etc
1790 if (Value *V = SimplifyUsingDistributiveLaws(I))
1791 return replaceInstUsesWith(I, V);
1792
1793 if (Value *V = SimplifyBSwap(I, Builder))
1794 return replaceInstUsesWith(I, V);
1795
1796 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1797 const APInt *C;
1798 if (match(Op1, m_APInt(C))) {
1799 Value *X, *Y;
1800 if (match(Op0, m_OneUse(m_LogicalShift(m_One(), m_Value(X)))) &&
1801 C->isOneValue()) {
1802 // (1 << X) & 1 --> zext(X == 0)
1803 // (1 >> X) & 1 --> zext(X == 0)
1804 Value *IsZero = Builder.CreateICmpEQ(X, ConstantInt::get(I.getType(), 0));
1805 return new ZExtInst(IsZero, I.getType());
1806 }
1807
1808 const APInt *XorC;
1809 if (match(Op0, m_OneUse(m_Xor(m_Value(X), m_APInt(XorC))))) {
1810 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
1811 Constant *NewC = ConstantInt::get(I.getType(), *C & *XorC);
1812 Value *And = Builder.CreateAnd(X, Op1);
1813 And->takeName(Op0);
1814 return BinaryOperator::CreateXor(And, NewC);
1815 }
1816
1817 const APInt *OrC;
1818 if (match(Op0, m_OneUse(m_Or(m_Value(X), m_APInt(OrC))))) {
1819 // (X | C1) & C2 --> (X & C2^(C1&C2)) | (C1&C2)
1820 // NOTE: This reduces the number of bits set in the & mask, which
1821 // can expose opportunities for store narrowing for scalars.
1822 // NOTE: SimplifyDemandedBits should have already removed bits from C1
1823 // that aren't set in C2. Meaning we can replace (C1&C2) with C1 in
1824 // above, but this feels safer.
1825 APInt Together = *C & *OrC;
1826 Value *And = Builder.CreateAnd(X, ConstantInt::get(I.getType(),
1827 Together ^ *C));
1828 And->takeName(Op0);
1829 return BinaryOperator::CreateOr(And, ConstantInt::get(I.getType(),
1830 Together));
1831 }
1832
1833 // If the mask is only needed on one incoming arm, push the 'and' op up.
1834 if (match(Op0, m_OneUse(m_Xor(m_Value(X), m_Value(Y)))) ||
1835 match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
1836 APInt NotAndMask(~(*C));
1837 BinaryOperator::BinaryOps BinOp = cast<BinaryOperator>(Op0)->getOpcode();
1838 if (MaskedValueIsZero(X, NotAndMask, 0, &I)) {
1839 // Not masking anything out for the LHS, move mask to RHS.
1840 // and ({x}or X, Y), C --> {x}or X, (and Y, C)
1841 Value *NewRHS = Builder.CreateAnd(Y, Op1, Y->getName() + ".masked");
1842 return BinaryOperator::Create(BinOp, X, NewRHS);
1843 }
1844 if (!isa<Constant>(Y) && MaskedValueIsZero(Y, NotAndMask, 0, &I)) {
1845 // Not masking anything out for the RHS, move mask to LHS.
1846 // and ({x}or X, Y), C --> {x}or (and X, C), Y
1847 Value *NewLHS = Builder.CreateAnd(X, Op1, X->getName() + ".masked");
1848 return BinaryOperator::Create(BinOp, NewLHS, Y);
1849 }
1850 }
1851 const APInt *ShiftC;
1852 if (match(Op0, m_OneUse(m_SExt(m_AShr(m_Value(X), m_APInt(ShiftC)))))) {
1853 unsigned Width = I.getType()->getScalarSizeInBits();
1854 if (*C == APInt::getLowBitsSet(Width, Width - ShiftC->getZExtValue())) {
1855 // We are clearing high bits that were potentially set by sext+ashr:
1856 // and (sext (ashr X, ShiftC)), C --> lshr (sext X), ShiftC
1857 Value *Sext = Builder.CreateSExt(X, I.getType());
1858 Constant *ShAmtC = ConstantInt::get(I.getType(), ShiftC->zext(Width));
1859 return BinaryOperator::CreateLShr(Sext, ShAmtC);
1860 }
1861 }
1862 }
1863
1864 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
1865 const APInt &AndRHSMask = AndRHS->getValue();
1866
1867 // Optimize a variety of ((val OP C1) & C2) combinations...
1868 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1869 // ((C1 OP zext(X)) & C2) -> zext((C1-X) & C2) if C2 fits in the bitwidth
1870 // of X and OP behaves well when given trunc(C1) and X.
1871 // TODO: Do this for vectors by using m_APInt isntead of m_ConstantInt.
1872 switch (Op0I->getOpcode()) {
1873 default:
1874 break;
1875 case Instruction::Xor:
1876 case Instruction::Or:
1877 case Instruction::Mul:
1878 case Instruction::Add:
1879 case Instruction::Sub:
1880 Value *X;
1881 ConstantInt *C1;
1882 // TODO: The one use restrictions could be relaxed a little if the AND
1883 // is going to be removed.
1884 if (match(Op0I, m_OneUse(m_c_BinOp(m_OneUse(m_ZExt(m_Value(X))),
1885 m_ConstantInt(C1))))) {
1886 if (AndRHSMask.isIntN(X->getType()->getScalarSizeInBits())) {
1887 auto *TruncC1 = ConstantExpr::getTrunc(C1, X->getType());
1888 Value *BinOp;
1889 Value *Op0LHS = Op0I->getOperand(0);
1890 if (isa<ZExtInst>(Op0LHS))
1891 BinOp = Builder.CreateBinOp(Op0I->getOpcode(), X, TruncC1);
1892 else
1893 BinOp = Builder.CreateBinOp(Op0I->getOpcode(), TruncC1, X);
1894 auto *TruncC2 = ConstantExpr::getTrunc(AndRHS, X->getType());
1895 auto *And = Builder.CreateAnd(BinOp, TruncC2);
1896 return new ZExtInst(And, I.getType());
1897 }
1898 }
1899 }
1900
1901 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
1902 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
1903 return Res;
1904 }
1905
1906 // If this is an integer truncation, and if the source is an 'and' with
1907 // immediate, transform it. This frequently occurs for bitfield accesses.
1908 {
1909 Value *X = nullptr; ConstantInt *YC = nullptr;
1910 if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) {
1911 // Change: and (trunc (and X, YC) to T), C2
1912 // into : and (trunc X to T), trunc(YC) & C2
1913 // This will fold the two constants together, which may allow
1914 // other simplifications.
1915 Value *NewCast = Builder.CreateTrunc(X, I.getType(), "and.shrunk");
1916 Constant *C3 = ConstantExpr::getTrunc(YC, I.getType());
1917 C3 = ConstantExpr::getAnd(C3, AndRHS);
1918 return BinaryOperator::CreateAnd(NewCast, C3);
1919 }
1920 }
1921 }
1922
1923 if (Instruction *Z = narrowMaskedBinOp(I))
1924 return Z;
1925
1926 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))
1927 return FoldedLogic;
1928
1929 if (Instruction *DeMorgan = matchDeMorgansLaws(I, Builder))
1930 return DeMorgan;
1931
1932 {
1933 Value *A, *B, *C;
1934 // A & (A ^ B) --> A & ~B
1935 if (match(Op1, m_OneUse(m_c_Xor(m_Specific(Op0), m_Value(B)))))
1936 return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(B));
1937 // (A ^ B) & A --> A & ~B
1938 if (match(Op0, m_OneUse(m_c_Xor(m_Specific(Op1), m_Value(B)))))
1939 return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(B));
1940
1941 // (A ^ B) & ((B ^ C) ^ A) -> (A ^ B) & ~C
1942 if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
1943 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
1944 if (Op1->hasOneUse() || isFreeToInvert(C, C->hasOneUse()))
1945 return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(C));
1946
1947 // ((A ^ C) ^ B) & (B ^ A) -> (B ^ A) & ~C
1948 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
1949 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
1950 if (Op0->hasOneUse() || isFreeToInvert(C, C->hasOneUse()))
1951 return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(C));
1952
1953 // (A | B) & ((~A) ^ B) -> (A & B)
1954 // (A | B) & (B ^ (~A)) -> (A & B)
1955 // (B | A) & ((~A) ^ B) -> (A & B)
1956 // (B | A) & (B ^ (~A)) -> (A & B)
1957 if (match(Op1, m_c_Xor(m_Not(m_Value(A)), m_Value(B))) &&
1958 match(Op0, m_c_Or(m_Specific(A), m_Specific(B))))
1959 return BinaryOperator::CreateAnd(A, B);
1960
1961 // ((~A) ^ B) & (A | B) -> (A & B)
1962 // ((~A) ^ B) & (B | A) -> (A & B)
1963 // (B ^ (~A)) & (A | B) -> (A & B)
1964 // (B ^ (~A)) & (B | A) -> (A & B)
1965 if (match(Op0, m_c_Xor(m_Not(m_Value(A)), m_Value(B))) &&
1966 match(Op1, m_c_Or(m_Specific(A), m_Specific(B))))
1967 return BinaryOperator::CreateAnd(A, B);
1968 }
1969
1970 {
1971 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
1972 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
1973 if (LHS && RHS)
1974 if (Value *Res = foldAndOfICmps(LHS, RHS, I))
1975 return replaceInstUsesWith(I, Res);
1976
1977 // TODO: Make this recursive; it's a little tricky because an arbitrary
1978 // number of 'and' instructions might have to be created.
1979 Value *X, *Y;
1980 if (LHS && match(Op1, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
1981 if (auto *Cmp = dyn_cast<ICmpInst>(X))
1982 if (Value *Res = foldAndOfICmps(LHS, Cmp, I))
1983 return replaceInstUsesWith(I, Builder.CreateAnd(Res, Y));
1984 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
1985 if (Value *Res = foldAndOfICmps(LHS, Cmp, I))
1986 return replaceInstUsesWith(I, Builder.CreateAnd(Res, X));
1987 }
1988 if (RHS && match(Op0, m_OneUse(m_And(m_Value(X), m_Value(Y))))) {
1989 if (auto *Cmp = dyn_cast<ICmpInst>(X))
1990 if (Value *Res = foldAndOfICmps(Cmp, RHS, I))
1991 return replaceInstUsesWith(I, Builder.CreateAnd(Res, Y));
1992 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
1993 if (Value *Res = foldAndOfICmps(Cmp, RHS, I))
1994 return replaceInstUsesWith(I, Builder.CreateAnd(Res, X));
1995 }
1996 }
1997
1998 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1999 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
2000 if (Value *Res = foldLogicOfFCmps(LHS, RHS, true))
2001 return replaceInstUsesWith(I, Res);
2002
2003 if (Instruction *FoldedFCmps = reassociateFCmps(I, Builder))
2004 return FoldedFCmps;
2005
2006 if (Instruction *CastedAnd = foldCastedBitwiseLogic(I))
2007 return CastedAnd;
2008
2009 // and(sext(A), B) / and(B, sext(A)) --> A ? B : 0, where A is i1 or <N x i1>.
2010 Value *A;
2011 if (match(Op0, m_OneUse(m_SExt(m_Value(A)))) &&
2012 A->getType()->isIntOrIntVectorTy(1))
2013 return SelectInst::Create(A, Op1, Constant::getNullValue(I.getType()));
2014 if (match(Op1, m_OneUse(m_SExt(m_Value(A)))) &&
2015 A->getType()->isIntOrIntVectorTy(1))
2016 return SelectInst::Create(A, Op0, Constant::getNullValue(I.getType()));
2017
2018 // and(ashr(subNSW(Y, X), ScalarSizeInBits(Y)-1), X) --> X s> Y ? X : 0.
2019 {
2020 Value *X, *Y;
2021 const APInt *ShAmt;
2022 Type *Ty = I.getType();
2023 if (match(&I, m_c_And(m_OneUse(m_AShr(m_NSWSub(m_Value(Y), m_Value(X)),
2024 m_APInt(ShAmt))),
2025 m_Deferred(X))) &&
2026 *ShAmt == Ty->getScalarSizeInBits() - 1) {
2027 Value *NewICmpInst = Builder.CreateICmpSGT(X, Y);
2028 return SelectInst::Create(NewICmpInst, X, ConstantInt::getNullValue(Ty));
2029 }
2030 }
2031
2032 return nullptr;
2033 }
2034
matchBSwap(BinaryOperator & Or)2035 Instruction *InstCombiner::matchBSwap(BinaryOperator &Or) {
2036 assert(Or.getOpcode() == Instruction::Or && "bswap requires an 'or'");
2037 Value *Op0 = Or.getOperand(0), *Op1 = Or.getOperand(1);
2038
2039 // Look through zero extends.
2040 if (Instruction *Ext = dyn_cast<ZExtInst>(Op0))
2041 Op0 = Ext->getOperand(0);
2042
2043 if (Instruction *Ext = dyn_cast<ZExtInst>(Op1))
2044 Op1 = Ext->getOperand(0);
2045
2046 // (A | B) | C and A | (B | C) -> bswap if possible.
2047 bool OrOfOrs = match(Op0, m_Or(m_Value(), m_Value())) ||
2048 match(Op1, m_Or(m_Value(), m_Value()));
2049
2050 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
2051 bool OrOfShifts = match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
2052 match(Op1, m_LogicalShift(m_Value(), m_Value()));
2053
2054 // (A & B) | (C & D) -> bswap if possible.
2055 bool OrOfAnds = match(Op0, m_And(m_Value(), m_Value())) &&
2056 match(Op1, m_And(m_Value(), m_Value()));
2057
2058 // (A << B) | (C & D) -> bswap if possible.
2059 // The bigger pattern here is ((A & C1) << C2) | ((B >> C2) & C1), which is a
2060 // part of the bswap idiom for specific values of C1, C2 (e.g. C1 = 16711935,
2061 // C2 = 8 for i32).
2062 // This pattern can occur when the operands of the 'or' are not canonicalized
2063 // for some reason (not having only one use, for example).
2064 bool OrOfAndAndSh = (match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
2065 match(Op1, m_And(m_Value(), m_Value()))) ||
2066 (match(Op0, m_And(m_Value(), m_Value())) &&
2067 match(Op1, m_LogicalShift(m_Value(), m_Value())));
2068
2069 if (!OrOfOrs && !OrOfShifts && !OrOfAnds && !OrOfAndAndSh)
2070 return nullptr;
2071
2072 SmallVector<Instruction*, 4> Insts;
2073 if (!recognizeBSwapOrBitReverseIdiom(&Or, true, false, Insts))
2074 return nullptr;
2075 Instruction *LastInst = Insts.pop_back_val();
2076 LastInst->removeFromParent();
2077
2078 for (auto *Inst : Insts)
2079 Worklist.push(Inst);
2080 return LastInst;
2081 }
2082
2083 /// Transform UB-safe variants of bitwise rotate to the funnel shift intrinsic.
matchRotate(Instruction & Or)2084 static Instruction *matchRotate(Instruction &Or) {
2085 // TODO: Can we reduce the code duplication between this and the related
2086 // rotate matching code under visitSelect and visitTrunc?
2087 unsigned Width = Or.getType()->getScalarSizeInBits();
2088 if (!isPowerOf2_32(Width))
2089 return nullptr;
2090
2091 // First, find an or'd pair of opposite shifts with the same shifted operand:
2092 // or (lshr ShVal, ShAmt0), (shl ShVal, ShAmt1)
2093 BinaryOperator *Or0, *Or1;
2094 if (!match(Or.getOperand(0), m_BinOp(Or0)) ||
2095 !match(Or.getOperand(1), m_BinOp(Or1)))
2096 return nullptr;
2097
2098 Value *ShVal, *ShAmt0, *ShAmt1;
2099 if (!match(Or0, m_OneUse(m_LogicalShift(m_Value(ShVal), m_Value(ShAmt0)))) ||
2100 !match(Or1, m_OneUse(m_LogicalShift(m_Specific(ShVal), m_Value(ShAmt1)))))
2101 return nullptr;
2102
2103 BinaryOperator::BinaryOps ShiftOpcode0 = Or0->getOpcode();
2104 BinaryOperator::BinaryOps ShiftOpcode1 = Or1->getOpcode();
2105 if (ShiftOpcode0 == ShiftOpcode1)
2106 return nullptr;
2107
2108 // Match the shift amount operands for a rotate pattern. This always matches
2109 // a subtraction on the R operand.
2110 auto matchShiftAmount = [](Value *L, Value *R, unsigned Width) -> Value * {
2111 // The shift amount may be masked with negation:
2112 // (shl ShVal, (X & (Width - 1))) | (lshr ShVal, ((-X) & (Width - 1)))
2113 Value *X;
2114 unsigned Mask = Width - 1;
2115 if (match(L, m_And(m_Value(X), m_SpecificInt(Mask))) &&
2116 match(R, m_And(m_Neg(m_Specific(X)), m_SpecificInt(Mask))))
2117 return X;
2118
2119 // Similar to above, but the shift amount may be extended after masking,
2120 // so return the extended value as the parameter for the intrinsic.
2121 if (match(L, m_ZExt(m_And(m_Value(X), m_SpecificInt(Mask)))) &&
2122 match(R, m_And(m_Neg(m_ZExt(m_And(m_Specific(X), m_SpecificInt(Mask)))),
2123 m_SpecificInt(Mask))))
2124 return L;
2125
2126 return nullptr;
2127 };
2128
2129 Value *ShAmt = matchShiftAmount(ShAmt0, ShAmt1, Width);
2130 bool SubIsOnLHS = false;
2131 if (!ShAmt) {
2132 ShAmt = matchShiftAmount(ShAmt1, ShAmt0, Width);
2133 SubIsOnLHS = true;
2134 }
2135 if (!ShAmt)
2136 return nullptr;
2137
2138 bool IsFshl = (!SubIsOnLHS && ShiftOpcode0 == BinaryOperator::Shl) ||
2139 (SubIsOnLHS && ShiftOpcode1 == BinaryOperator::Shl);
2140 Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr;
2141 Function *F = Intrinsic::getDeclaration(Or.getModule(), IID, Or.getType());
2142 return IntrinsicInst::Create(F, { ShVal, ShVal, ShAmt });
2143 }
2144
2145 /// Attempt to combine or(zext(x),shl(zext(y),bw/2) concat packing patterns.
matchOrConcat(Instruction & Or,InstCombiner::BuilderTy & Builder)2146 static Instruction *matchOrConcat(Instruction &Or,
2147 InstCombiner::BuilderTy &Builder) {
2148 assert(Or.getOpcode() == Instruction::Or && "bswap requires an 'or'");
2149 Value *Op0 = Or.getOperand(0), *Op1 = Or.getOperand(1);
2150 Type *Ty = Or.getType();
2151
2152 unsigned Width = Ty->getScalarSizeInBits();
2153 if ((Width & 1) != 0)
2154 return nullptr;
2155 unsigned HalfWidth = Width / 2;
2156
2157 // Canonicalize zext (lower half) to LHS.
2158 if (!isa<ZExtInst>(Op0))
2159 std::swap(Op0, Op1);
2160
2161 // Find lower/upper half.
2162 Value *LowerSrc, *ShlVal, *UpperSrc;
2163 const APInt *C;
2164 if (!match(Op0, m_OneUse(m_ZExt(m_Value(LowerSrc)))) ||
2165 !match(Op1, m_OneUse(m_Shl(m_Value(ShlVal), m_APInt(C)))) ||
2166 !match(ShlVal, m_OneUse(m_ZExt(m_Value(UpperSrc)))))
2167 return nullptr;
2168 if (*C != HalfWidth || LowerSrc->getType() != UpperSrc->getType() ||
2169 LowerSrc->getType()->getScalarSizeInBits() != HalfWidth)
2170 return nullptr;
2171
2172 auto ConcatIntrinsicCalls = [&](Intrinsic::ID id, Value *Lo, Value *Hi) {
2173 Value *NewLower = Builder.CreateZExt(Lo, Ty);
2174 Value *NewUpper = Builder.CreateZExt(Hi, Ty);
2175 NewUpper = Builder.CreateShl(NewUpper, HalfWidth);
2176 Value *BinOp = Builder.CreateOr(NewLower, NewUpper);
2177 Function *F = Intrinsic::getDeclaration(Or.getModule(), id, Ty);
2178 return Builder.CreateCall(F, BinOp);
2179 };
2180
2181 // BSWAP: Push the concat down, swapping the lower/upper sources.
2182 // concat(bswap(x),bswap(y)) -> bswap(concat(x,y))
2183 Value *LowerBSwap, *UpperBSwap;
2184 if (match(LowerSrc, m_BSwap(m_Value(LowerBSwap))) &&
2185 match(UpperSrc, m_BSwap(m_Value(UpperBSwap))))
2186 return ConcatIntrinsicCalls(Intrinsic::bswap, UpperBSwap, LowerBSwap);
2187
2188 // BITREVERSE: Push the concat down, swapping the lower/upper sources.
2189 // concat(bitreverse(x),bitreverse(y)) -> bitreverse(concat(x,y))
2190 Value *LowerBRev, *UpperBRev;
2191 if (match(LowerSrc, m_BitReverse(m_Value(LowerBRev))) &&
2192 match(UpperSrc, m_BitReverse(m_Value(UpperBRev))))
2193 return ConcatIntrinsicCalls(Intrinsic::bitreverse, UpperBRev, LowerBRev);
2194
2195 return nullptr;
2196 }
2197
2198 /// If all elements of two constant vectors are 0/-1 and inverses, return true.
areInverseVectorBitmasks(Constant * C1,Constant * C2)2199 static bool areInverseVectorBitmasks(Constant *C1, Constant *C2) {
2200 unsigned NumElts = cast<VectorType>(C1->getType())->getNumElements();
2201 for (unsigned i = 0; i != NumElts; ++i) {
2202 Constant *EltC1 = C1->getAggregateElement(i);
2203 Constant *EltC2 = C2->getAggregateElement(i);
2204 if (!EltC1 || !EltC2)
2205 return false;
2206
2207 // One element must be all ones, and the other must be all zeros.
2208 if (!((match(EltC1, m_Zero()) && match(EltC2, m_AllOnes())) ||
2209 (match(EltC2, m_Zero()) && match(EltC1, m_AllOnes()))))
2210 return false;
2211 }
2212 return true;
2213 }
2214
2215 /// We have an expression of the form (A & C) | (B & D). If A is a scalar or
2216 /// vector composed of all-zeros or all-ones values and is the bitwise 'not' of
2217 /// B, it can be used as the condition operand of a select instruction.
getSelectCondition(Value * A,Value * B)2218 Value *InstCombiner::getSelectCondition(Value *A, Value *B) {
2219 // Step 1: We may have peeked through bitcasts in the caller.
2220 // Exit immediately if we don't have (vector) integer types.
2221 Type *Ty = A->getType();
2222 if (!Ty->isIntOrIntVectorTy() || !B->getType()->isIntOrIntVectorTy())
2223 return nullptr;
2224
2225 // Step 2: We need 0 or all-1's bitmasks.
2226 if (ComputeNumSignBits(A) != Ty->getScalarSizeInBits())
2227 return nullptr;
2228
2229 // Step 3: If B is the 'not' value of A, we have our answer.
2230 if (match(A, m_Not(m_Specific(B)))) {
2231 // If these are scalars or vectors of i1, A can be used directly.
2232 if (Ty->isIntOrIntVectorTy(1))
2233 return A;
2234 return Builder.CreateTrunc(A, CmpInst::makeCmpResultType(Ty));
2235 }
2236
2237 // If both operands are constants, see if the constants are inverse bitmasks.
2238 Constant *AConst, *BConst;
2239 if (match(A, m_Constant(AConst)) && match(B, m_Constant(BConst)))
2240 if (AConst == ConstantExpr::getNot(BConst))
2241 return Builder.CreateZExtOrTrunc(A, CmpInst::makeCmpResultType(Ty));
2242
2243 // Look for more complex patterns. The 'not' op may be hidden behind various
2244 // casts. Look through sexts and bitcasts to find the booleans.
2245 Value *Cond;
2246 Value *NotB;
2247 if (match(A, m_SExt(m_Value(Cond))) &&
2248 Cond->getType()->isIntOrIntVectorTy(1) &&
2249 match(B, m_OneUse(m_Not(m_Value(NotB))))) {
2250 NotB = peekThroughBitcast(NotB, true);
2251 if (match(NotB, m_SExt(m_Specific(Cond))))
2252 return Cond;
2253 }
2254
2255 // All scalar (and most vector) possibilities should be handled now.
2256 // Try more matches that only apply to non-splat constant vectors.
2257 if (!Ty->isVectorTy())
2258 return nullptr;
2259
2260 // If both operands are xor'd with constants using the same sexted boolean
2261 // operand, see if the constants are inverse bitmasks.
2262 // TODO: Use ConstantExpr::getNot()?
2263 if (match(A, (m_Xor(m_SExt(m_Value(Cond)), m_Constant(AConst)))) &&
2264 match(B, (m_Xor(m_SExt(m_Specific(Cond)), m_Constant(BConst)))) &&
2265 Cond->getType()->isIntOrIntVectorTy(1) &&
2266 areInverseVectorBitmasks(AConst, BConst)) {
2267 AConst = ConstantExpr::getTrunc(AConst, CmpInst::makeCmpResultType(Ty));
2268 return Builder.CreateXor(Cond, AConst);
2269 }
2270 return nullptr;
2271 }
2272
2273 /// We have an expression of the form (A & C) | (B & D). Try to simplify this
2274 /// to "A' ? C : D", where A' is a boolean or vector of booleans.
matchSelectFromAndOr(Value * A,Value * C,Value * B,Value * D)2275 Value *InstCombiner::matchSelectFromAndOr(Value *A, Value *C, Value *B,
2276 Value *D) {
2277 // The potential condition of the select may be bitcasted. In that case, look
2278 // through its bitcast and the corresponding bitcast of the 'not' condition.
2279 Type *OrigType = A->getType();
2280 A = peekThroughBitcast(A, true);
2281 B = peekThroughBitcast(B, true);
2282 if (Value *Cond = getSelectCondition(A, B)) {
2283 // ((bc Cond) & C) | ((bc ~Cond) & D) --> bc (select Cond, (bc C), (bc D))
2284 // The bitcasts will either all exist or all not exist. The builder will
2285 // not create unnecessary casts if the types already match.
2286 Value *BitcastC = Builder.CreateBitCast(C, A->getType());
2287 Value *BitcastD = Builder.CreateBitCast(D, A->getType());
2288 Value *Select = Builder.CreateSelect(Cond, BitcastC, BitcastD);
2289 return Builder.CreateBitCast(Select, OrigType);
2290 }
2291
2292 return nullptr;
2293 }
2294
2295 /// Fold (icmp)|(icmp) if possible.
foldOrOfICmps(ICmpInst * LHS,ICmpInst * RHS,BinaryOperator & Or)2296 Value *InstCombiner::foldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS,
2297 BinaryOperator &Or) {
2298 const SimplifyQuery Q = SQ.getWithInstruction(&Or);
2299
2300 // Fold (iszero(A & K1) | iszero(A & K2)) -> (A & (K1 | K2)) != (K1 | K2)
2301 // if K1 and K2 are a one-bit mask.
2302 if (Value *V = foldAndOrOfICmpsOfAndWithPow2(LHS, RHS, Or))
2303 return V;
2304
2305 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
2306
2307 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHS->getOperand(1));
2308 ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS->getOperand(1));
2309
2310 // Fold (icmp ult/ule (A + C1), C3) | (icmp ult/ule (A + C2), C3)
2311 // --> (icmp ult/ule ((A & ~(C1 ^ C2)) + max(C1, C2)), C3)
2312 // The original condition actually refers to the following two ranges:
2313 // [MAX_UINT-C1+1, MAX_UINT-C1+1+C3] and [MAX_UINT-C2+1, MAX_UINT-C2+1+C3]
2314 // We can fold these two ranges if:
2315 // 1) C1 and C2 is unsigned greater than C3.
2316 // 2) The two ranges are separated.
2317 // 3) C1 ^ C2 is one-bit mask.
2318 // 4) LowRange1 ^ LowRange2 and HighRange1 ^ HighRange2 are one-bit mask.
2319 // This implies all values in the two ranges differ by exactly one bit.
2320
2321 if ((PredL == ICmpInst::ICMP_ULT || PredL == ICmpInst::ICMP_ULE) &&
2322 PredL == PredR && LHSC && RHSC && LHS->hasOneUse() && RHS->hasOneUse() &&
2323 LHSC->getType() == RHSC->getType() &&
2324 LHSC->getValue() == (RHSC->getValue())) {
2325
2326 Value *LAdd = LHS->getOperand(0);
2327 Value *RAdd = RHS->getOperand(0);
2328
2329 Value *LAddOpnd, *RAddOpnd;
2330 ConstantInt *LAddC, *RAddC;
2331 if (match(LAdd, m_Add(m_Value(LAddOpnd), m_ConstantInt(LAddC))) &&
2332 match(RAdd, m_Add(m_Value(RAddOpnd), m_ConstantInt(RAddC))) &&
2333 LAddC->getValue().ugt(LHSC->getValue()) &&
2334 RAddC->getValue().ugt(LHSC->getValue())) {
2335
2336 APInt DiffC = LAddC->getValue() ^ RAddC->getValue();
2337 if (LAddOpnd == RAddOpnd && DiffC.isPowerOf2()) {
2338 ConstantInt *MaxAddC = nullptr;
2339 if (LAddC->getValue().ult(RAddC->getValue()))
2340 MaxAddC = RAddC;
2341 else
2342 MaxAddC = LAddC;
2343
2344 APInt RRangeLow = -RAddC->getValue();
2345 APInt RRangeHigh = RRangeLow + LHSC->getValue();
2346 APInt LRangeLow = -LAddC->getValue();
2347 APInt LRangeHigh = LRangeLow + LHSC->getValue();
2348 APInt LowRangeDiff = RRangeLow ^ LRangeLow;
2349 APInt HighRangeDiff = RRangeHigh ^ LRangeHigh;
2350 APInt RangeDiff = LRangeLow.sgt(RRangeLow) ? LRangeLow - RRangeLow
2351 : RRangeLow - LRangeLow;
2352
2353 if (LowRangeDiff.isPowerOf2() && LowRangeDiff == HighRangeDiff &&
2354 RangeDiff.ugt(LHSC->getValue())) {
2355 Value *MaskC = ConstantInt::get(LAddC->getType(), ~DiffC);
2356
2357 Value *NewAnd = Builder.CreateAnd(LAddOpnd, MaskC);
2358 Value *NewAdd = Builder.CreateAdd(NewAnd, MaxAddC);
2359 return Builder.CreateICmp(LHS->getPredicate(), NewAdd, LHSC);
2360 }
2361 }
2362 }
2363 }
2364
2365 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
2366 if (predicatesFoldable(PredL, PredR)) {
2367 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2368 LHS->getOperand(1) == RHS->getOperand(0))
2369 LHS->swapOperands();
2370 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2371 LHS->getOperand(1) == RHS->getOperand(1)) {
2372 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2373 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
2374 bool IsSigned = LHS->isSigned() || RHS->isSigned();
2375 return getNewICmpValue(Code, IsSigned, Op0, Op1, Builder);
2376 }
2377 }
2378
2379 // handle (roughly):
2380 // (icmp ne (A & B), C) | (icmp ne (A & D), E)
2381 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, false, Builder))
2382 return V;
2383
2384 Value *LHS0 = LHS->getOperand(0), *RHS0 = RHS->getOperand(0);
2385 if (LHS->hasOneUse() || RHS->hasOneUse()) {
2386 // (icmp eq B, 0) | (icmp ult A, B) -> (icmp ule A, B-1)
2387 // (icmp eq B, 0) | (icmp ugt B, A) -> (icmp ule A, B-1)
2388 Value *A = nullptr, *B = nullptr;
2389 if (PredL == ICmpInst::ICMP_EQ && LHSC && LHSC->isZero()) {
2390 B = LHS0;
2391 if (PredR == ICmpInst::ICMP_ULT && LHS0 == RHS->getOperand(1))
2392 A = RHS0;
2393 else if (PredR == ICmpInst::ICMP_UGT && LHS0 == RHS0)
2394 A = RHS->getOperand(1);
2395 }
2396 // (icmp ult A, B) | (icmp eq B, 0) -> (icmp ule A, B-1)
2397 // (icmp ugt B, A) | (icmp eq B, 0) -> (icmp ule A, B-1)
2398 else if (PredR == ICmpInst::ICMP_EQ && RHSC && RHSC->isZero()) {
2399 B = RHS0;
2400 if (PredL == ICmpInst::ICMP_ULT && RHS0 == LHS->getOperand(1))
2401 A = LHS0;
2402 else if (PredL == ICmpInst::ICMP_UGT && LHS0 == RHS0)
2403 A = LHS->getOperand(1);
2404 }
2405 if (A && B)
2406 return Builder.CreateICmp(
2407 ICmpInst::ICMP_UGE,
2408 Builder.CreateAdd(B, ConstantInt::getSigned(B->getType(), -1)), A);
2409 }
2410
2411 if (Value *V = foldAndOrOfICmpsWithConstEq(LHS, RHS, Or, Builder, Q))
2412 return V;
2413 if (Value *V = foldAndOrOfICmpsWithConstEq(RHS, LHS, Or, Builder, Q))
2414 return V;
2415
2416 // E.g. (icmp slt x, 0) | (icmp sgt x, n) --> icmp ugt x, n
2417 if (Value *V = simplifyRangeCheck(LHS, RHS, /*Inverted=*/true))
2418 return V;
2419
2420 // E.g. (icmp sgt x, n) | (icmp slt x, 0) --> icmp ugt x, n
2421 if (Value *V = simplifyRangeCheck(RHS, LHS, /*Inverted=*/true))
2422 return V;
2423
2424 if (Value *V = foldAndOrOfEqualityCmpsWithConstants(LHS, RHS, false, Builder))
2425 return V;
2426
2427 if (Value *V = foldIsPowerOf2(LHS, RHS, false /* JoinedByAnd */, Builder))
2428 return V;
2429
2430 if (Value *X =
2431 foldUnsignedUnderflowCheck(LHS, RHS, /*IsAnd=*/false, Q, Builder))
2432 return X;
2433 if (Value *X =
2434 foldUnsignedUnderflowCheck(RHS, LHS, /*IsAnd=*/false, Q, Builder))
2435 return X;
2436
2437 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
2438 if (!LHSC || !RHSC)
2439 return nullptr;
2440
2441 if (LHSC == RHSC && PredL == PredR) {
2442 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
2443 if (PredL == ICmpInst::ICMP_NE && LHSC->isZero()) {
2444 Value *NewOr = Builder.CreateOr(LHS0, RHS0);
2445 return Builder.CreateICmp(PredL, NewOr, LHSC);
2446 }
2447 }
2448
2449 // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
2450 // iff C2 + CA == C1.
2451 if (PredL == ICmpInst::ICMP_ULT && PredR == ICmpInst::ICMP_EQ) {
2452 ConstantInt *AddC;
2453 if (match(LHS0, m_Add(m_Specific(RHS0), m_ConstantInt(AddC))))
2454 if (RHSC->getValue() + AddC->getValue() == LHSC->getValue())
2455 return Builder.CreateICmpULE(LHS0, LHSC);
2456 }
2457
2458 // From here on, we only handle:
2459 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
2460 if (LHS0 != RHS0)
2461 return nullptr;
2462
2463 // ICMP_[US][GL]E X, C is folded to ICMP_[US][GL]T elsewhere.
2464 if (PredL == ICmpInst::ICMP_UGE || PredL == ICmpInst::ICMP_ULE ||
2465 PredR == ICmpInst::ICMP_UGE || PredR == ICmpInst::ICMP_ULE ||
2466 PredL == ICmpInst::ICMP_SGE || PredL == ICmpInst::ICMP_SLE ||
2467 PredR == ICmpInst::ICMP_SGE || PredR == ICmpInst::ICMP_SLE)
2468 return nullptr;
2469
2470 // We can't fold (ugt x, C) | (sgt x, C2).
2471 if (!predicatesFoldable(PredL, PredR))
2472 return nullptr;
2473
2474 // Ensure that the larger constant is on the RHS.
2475 bool ShouldSwap;
2476 if (CmpInst::isSigned(PredL) ||
2477 (ICmpInst::isEquality(PredL) && CmpInst::isSigned(PredR)))
2478 ShouldSwap = LHSC->getValue().sgt(RHSC->getValue());
2479 else
2480 ShouldSwap = LHSC->getValue().ugt(RHSC->getValue());
2481
2482 if (ShouldSwap) {
2483 std::swap(LHS, RHS);
2484 std::swap(LHSC, RHSC);
2485 std::swap(PredL, PredR);
2486 }
2487
2488 // At this point, we know we have two icmp instructions
2489 // comparing a value against two constants and or'ing the result
2490 // together. Because of the above check, we know that we only have
2491 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
2492 // icmp folding check above), that the two constants are not
2493 // equal.
2494 assert(LHSC != RHSC && "Compares not folded above?");
2495
2496 switch (PredL) {
2497 default:
2498 llvm_unreachable("Unknown integer condition code!");
2499 case ICmpInst::ICMP_EQ:
2500 switch (PredR) {
2501 default:
2502 llvm_unreachable("Unknown integer condition code!");
2503 case ICmpInst::ICMP_EQ:
2504 // Potential folds for this case should already be handled.
2505 break;
2506 case ICmpInst::ICMP_UGT:
2507 // (X == 0 || X u> C) -> (X-1) u>= C
2508 if (LHSC->isMinValue(false))
2509 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue() + 1,
2510 false, false);
2511 // (X == 13 | X u> 14) -> no change
2512 break;
2513 case ICmpInst::ICMP_SGT:
2514 // (X == INT_MIN || X s> C) -> (X-(INT_MIN+1)) u>= C-INT_MIN
2515 if (LHSC->isMinValue(true))
2516 return insertRangeTest(LHS0, LHSC->getValue() + 1, RHSC->getValue() + 1,
2517 true, false);
2518 // (X == 13 | X s> 14) -> no change
2519 break;
2520 }
2521 break;
2522 case ICmpInst::ICMP_ULT:
2523 switch (PredR) {
2524 default:
2525 llvm_unreachable("Unknown integer condition code!");
2526 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
2527 // (X u< C || X == UINT_MAX) => (X-C) u>= UINT_MAX-C
2528 if (RHSC->isMaxValue(false))
2529 return insertRangeTest(LHS0, LHSC->getValue(), RHSC->getValue(),
2530 false, false);
2531 break;
2532 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
2533 assert(!RHSC->isMaxValue(false) && "Missed icmp simplification");
2534 return insertRangeTest(LHS0, LHSC->getValue(), RHSC->getValue() + 1,
2535 false, false);
2536 }
2537 break;
2538 case ICmpInst::ICMP_SLT:
2539 switch (PredR) {
2540 default:
2541 llvm_unreachable("Unknown integer condition code!");
2542 case ICmpInst::ICMP_EQ:
2543 // (X s< C || X == INT_MAX) => (X-C) u>= INT_MAX-C
2544 if (RHSC->isMaxValue(true))
2545 return insertRangeTest(LHS0, LHSC->getValue(), RHSC->getValue(),
2546 true, false);
2547 // (X s< 13 | X == 14) -> no change
2548 break;
2549 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) u> 2
2550 assert(!RHSC->isMaxValue(true) && "Missed icmp simplification");
2551 return insertRangeTest(LHS0, LHSC->getValue(), RHSC->getValue() + 1, true,
2552 false);
2553 }
2554 break;
2555 }
2556 return nullptr;
2557 }
2558
2559 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
2560 // here. We should standardize that construct where it is needed or choose some
2561 // other way to ensure that commutated variants of patterns are not missed.
visitOr(BinaryOperator & I)2562 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
2563 if (Value *V = SimplifyOrInst(I.getOperand(0), I.getOperand(1),
2564 SQ.getWithInstruction(&I)))
2565 return replaceInstUsesWith(I, V);
2566
2567 if (SimplifyAssociativeOrCommutative(I))
2568 return &I;
2569
2570 if (Instruction *X = foldVectorBinop(I))
2571 return X;
2572
2573 // See if we can simplify any instructions used by the instruction whose sole
2574 // purpose is to compute bits we don't care about.
2575 if (SimplifyDemandedInstructionBits(I))
2576 return &I;
2577
2578 // Do this before using distributive laws to catch simple and/or/not patterns.
2579 if (Instruction *Xor = foldOrToXor(I, Builder))
2580 return Xor;
2581
2582 // (A&B)|(A&C) -> A&(B|C) etc
2583 if (Value *V = SimplifyUsingDistributiveLaws(I))
2584 return replaceInstUsesWith(I, V);
2585
2586 if (Value *V = SimplifyBSwap(I, Builder))
2587 return replaceInstUsesWith(I, V);
2588
2589 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))
2590 return FoldedLogic;
2591
2592 if (Instruction *BSwap = matchBSwap(I))
2593 return BSwap;
2594
2595 if (Instruction *Rotate = matchRotate(I))
2596 return Rotate;
2597
2598 if (Instruction *Concat = matchOrConcat(I, Builder))
2599 return replaceInstUsesWith(I, Concat);
2600
2601 Value *X, *Y;
2602 const APInt *CV;
2603 if (match(&I, m_c_Or(m_OneUse(m_Xor(m_Value(X), m_APInt(CV))), m_Value(Y))) &&
2604 !CV->isAllOnesValue() && MaskedValueIsZero(Y, *CV, 0, &I)) {
2605 // (X ^ C) | Y -> (X | Y) ^ C iff Y & C == 0
2606 // The check for a 'not' op is for efficiency (if Y is known zero --> ~X).
2607 Value *Or = Builder.CreateOr(X, Y);
2608 return BinaryOperator::CreateXor(Or, ConstantInt::get(I.getType(), *CV));
2609 }
2610
2611 // (A & C)|(B & D)
2612 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2613 Value *A, *B, *C, *D;
2614 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
2615 match(Op1, m_And(m_Value(B), m_Value(D)))) {
2616 ConstantInt *C1 = dyn_cast<ConstantInt>(C);
2617 ConstantInt *C2 = dyn_cast<ConstantInt>(D);
2618 if (C1 && C2) { // (A & C1)|(B & C2)
2619 Value *V1 = nullptr, *V2 = nullptr;
2620 if ((C1->getValue() & C2->getValue()).isNullValue()) {
2621 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
2622 // iff (C1&C2) == 0 and (N&~C1) == 0
2623 if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
2624 ((V1 == B &&
2625 MaskedValueIsZero(V2, ~C1->getValue(), 0, &I)) || // (V|N)
2626 (V2 == B &&
2627 MaskedValueIsZero(V1, ~C1->getValue(), 0, &I)))) // (N|V)
2628 return BinaryOperator::CreateAnd(A,
2629 Builder.getInt(C1->getValue()|C2->getValue()));
2630 // Or commutes, try both ways.
2631 if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
2632 ((V1 == A &&
2633 MaskedValueIsZero(V2, ~C2->getValue(), 0, &I)) || // (V|N)
2634 (V2 == A &&
2635 MaskedValueIsZero(V1, ~C2->getValue(), 0, &I)))) // (N|V)
2636 return BinaryOperator::CreateAnd(B,
2637 Builder.getInt(C1->getValue()|C2->getValue()));
2638
2639 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
2640 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
2641 ConstantInt *C3 = nullptr, *C4 = nullptr;
2642 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
2643 (C3->getValue() & ~C1->getValue()).isNullValue() &&
2644 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
2645 (C4->getValue() & ~C2->getValue()).isNullValue()) {
2646 V2 = Builder.CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
2647 return BinaryOperator::CreateAnd(V2,
2648 Builder.getInt(C1->getValue()|C2->getValue()));
2649 }
2650 }
2651
2652 if (C1->getValue() == ~C2->getValue()) {
2653 Value *X;
2654
2655 // ((X|B)&C1)|(B&C2) -> (X&C1) | B iff C1 == ~C2
2656 if (match(A, m_c_Or(m_Value(X), m_Specific(B))))
2657 return BinaryOperator::CreateOr(Builder.CreateAnd(X, C1), B);
2658 // (A&C2)|((X|A)&C1) -> (X&C2) | A iff C1 == ~C2
2659 if (match(B, m_c_Or(m_Specific(A), m_Value(X))))
2660 return BinaryOperator::CreateOr(Builder.CreateAnd(X, C2), A);
2661
2662 // ((X^B)&C1)|(B&C2) -> (X&C1) ^ B iff C1 == ~C2
2663 if (match(A, m_c_Xor(m_Value(X), m_Specific(B))))
2664 return BinaryOperator::CreateXor(Builder.CreateAnd(X, C1), B);
2665 // (A&C2)|((X^A)&C1) -> (X&C2) ^ A iff C1 == ~C2
2666 if (match(B, m_c_Xor(m_Specific(A), m_Value(X))))
2667 return BinaryOperator::CreateXor(Builder.CreateAnd(X, C2), A);
2668 }
2669 }
2670
2671 // Don't try to form a select if it's unlikely that we'll get rid of at
2672 // least one of the operands. A select is generally more expensive than the
2673 // 'or' that it is replacing.
2674 if (Op0->hasOneUse() || Op1->hasOneUse()) {
2675 // (Cond & C) | (~Cond & D) -> Cond ? C : D, and commuted variants.
2676 if (Value *V = matchSelectFromAndOr(A, C, B, D))
2677 return replaceInstUsesWith(I, V);
2678 if (Value *V = matchSelectFromAndOr(A, C, D, B))
2679 return replaceInstUsesWith(I, V);
2680 if (Value *V = matchSelectFromAndOr(C, A, B, D))
2681 return replaceInstUsesWith(I, V);
2682 if (Value *V = matchSelectFromAndOr(C, A, D, B))
2683 return replaceInstUsesWith(I, V);
2684 if (Value *V = matchSelectFromAndOr(B, D, A, C))
2685 return replaceInstUsesWith(I, V);
2686 if (Value *V = matchSelectFromAndOr(B, D, C, A))
2687 return replaceInstUsesWith(I, V);
2688 if (Value *V = matchSelectFromAndOr(D, B, A, C))
2689 return replaceInstUsesWith(I, V);
2690 if (Value *V = matchSelectFromAndOr(D, B, C, A))
2691 return replaceInstUsesWith(I, V);
2692 }
2693 }
2694
2695 // (A ^ B) | ((B ^ C) ^ A) -> (A ^ B) | C
2696 if (match(Op0, m_Xor(m_Value(A), m_Value(B))))
2697 if (match(Op1, m_Xor(m_Xor(m_Specific(B), m_Value(C)), m_Specific(A))))
2698 return BinaryOperator::CreateOr(Op0, C);
2699
2700 // ((A ^ C) ^ B) | (B ^ A) -> (B ^ A) | C
2701 if (match(Op0, m_Xor(m_Xor(m_Value(A), m_Value(C)), m_Value(B))))
2702 if (match(Op1, m_Xor(m_Specific(B), m_Specific(A))))
2703 return BinaryOperator::CreateOr(Op1, C);
2704
2705 // ((B | C) & A) | B -> B | (A & C)
2706 if (match(Op0, m_And(m_Or(m_Specific(Op1), m_Value(C)), m_Value(A))))
2707 return BinaryOperator::CreateOr(Op1, Builder.CreateAnd(A, C));
2708
2709 if (Instruction *DeMorgan = matchDeMorgansLaws(I, Builder))
2710 return DeMorgan;
2711
2712 // Canonicalize xor to the RHS.
2713 bool SwappedForXor = false;
2714 if (match(Op0, m_Xor(m_Value(), m_Value()))) {
2715 std::swap(Op0, Op1);
2716 SwappedForXor = true;
2717 }
2718
2719 // A | ( A ^ B) -> A | B
2720 // A | (~A ^ B) -> A | ~B
2721 // (A & B) | (A ^ B)
2722 if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
2723 if (Op0 == A || Op0 == B)
2724 return BinaryOperator::CreateOr(A, B);
2725
2726 if (match(Op0, m_And(m_Specific(A), m_Specific(B))) ||
2727 match(Op0, m_And(m_Specific(B), m_Specific(A))))
2728 return BinaryOperator::CreateOr(A, B);
2729
2730 if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
2731 Value *Not = Builder.CreateNot(B, B->getName() + ".not");
2732 return BinaryOperator::CreateOr(Not, Op0);
2733 }
2734 if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
2735 Value *Not = Builder.CreateNot(A, A->getName() + ".not");
2736 return BinaryOperator::CreateOr(Not, Op0);
2737 }
2738 }
2739
2740 // A | ~(A | B) -> A | ~B
2741 // A | ~(A ^ B) -> A | ~B
2742 if (match(Op1, m_Not(m_Value(A))))
2743 if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
2744 if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
2745 Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
2746 B->getOpcode() == Instruction::Xor)) {
2747 Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
2748 B->getOperand(0);
2749 Value *Not = Builder.CreateNot(NotOp, NotOp->getName() + ".not");
2750 return BinaryOperator::CreateOr(Not, Op0);
2751 }
2752
2753 if (SwappedForXor)
2754 std::swap(Op0, Op1);
2755
2756 {
2757 ICmpInst *LHS = dyn_cast<ICmpInst>(Op0);
2758 ICmpInst *RHS = dyn_cast<ICmpInst>(Op1);
2759 if (LHS && RHS)
2760 if (Value *Res = foldOrOfICmps(LHS, RHS, I))
2761 return replaceInstUsesWith(I, Res);
2762
2763 // TODO: Make this recursive; it's a little tricky because an arbitrary
2764 // number of 'or' instructions might have to be created.
2765 Value *X, *Y;
2766 if (LHS && match(Op1, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2767 if (auto *Cmp = dyn_cast<ICmpInst>(X))
2768 if (Value *Res = foldOrOfICmps(LHS, Cmp, I))
2769 return replaceInstUsesWith(I, Builder.CreateOr(Res, Y));
2770 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2771 if (Value *Res = foldOrOfICmps(LHS, Cmp, I))
2772 return replaceInstUsesWith(I, Builder.CreateOr(Res, X));
2773 }
2774 if (RHS && match(Op0, m_OneUse(m_Or(m_Value(X), m_Value(Y))))) {
2775 if (auto *Cmp = dyn_cast<ICmpInst>(X))
2776 if (Value *Res = foldOrOfICmps(Cmp, RHS, I))
2777 return replaceInstUsesWith(I, Builder.CreateOr(Res, Y));
2778 if (auto *Cmp = dyn_cast<ICmpInst>(Y))
2779 if (Value *Res = foldOrOfICmps(Cmp, RHS, I))
2780 return replaceInstUsesWith(I, Builder.CreateOr(Res, X));
2781 }
2782 }
2783
2784 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
2785 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
2786 if (Value *Res = foldLogicOfFCmps(LHS, RHS, false))
2787 return replaceInstUsesWith(I, Res);
2788
2789 if (Instruction *FoldedFCmps = reassociateFCmps(I, Builder))
2790 return FoldedFCmps;
2791
2792 if (Instruction *CastedOr = foldCastedBitwiseLogic(I))
2793 return CastedOr;
2794
2795 // or(sext(A), B) / or(B, sext(A)) --> A ? -1 : B, where A is i1 or <N x i1>.
2796 if (match(Op0, m_OneUse(m_SExt(m_Value(A)))) &&
2797 A->getType()->isIntOrIntVectorTy(1))
2798 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1);
2799 if (match(Op1, m_OneUse(m_SExt(m_Value(A)))) &&
2800 A->getType()->isIntOrIntVectorTy(1))
2801 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0);
2802
2803 // Note: If we've gotten to the point of visiting the outer OR, then the
2804 // inner one couldn't be simplified. If it was a constant, then it won't
2805 // be simplified by a later pass either, so we try swapping the inner/outer
2806 // ORs in the hopes that we'll be able to simplify it this way.
2807 // (X|C) | V --> (X|V) | C
2808 ConstantInt *CI;
2809 if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
2810 match(Op0, m_Or(m_Value(A), m_ConstantInt(CI)))) {
2811 Value *Inner = Builder.CreateOr(A, Op1);
2812 Inner->takeName(Op0);
2813 return BinaryOperator::CreateOr(Inner, CI);
2814 }
2815
2816 // Change (or (bool?A:B),(bool?C:D)) --> (bool?(or A,C):(or B,D))
2817 // Since this OR statement hasn't been optimized further yet, we hope
2818 // that this transformation will allow the new ORs to be optimized.
2819 {
2820 Value *X = nullptr, *Y = nullptr;
2821 if (Op0->hasOneUse() && Op1->hasOneUse() &&
2822 match(Op0, m_Select(m_Value(X), m_Value(A), m_Value(B))) &&
2823 match(Op1, m_Select(m_Value(Y), m_Value(C), m_Value(D))) && X == Y) {
2824 Value *orTrue = Builder.CreateOr(A, C);
2825 Value *orFalse = Builder.CreateOr(B, D);
2826 return SelectInst::Create(X, orTrue, orFalse);
2827 }
2828 }
2829
2830 // or(ashr(subNSW(Y, X), ScalarSizeInBits(Y)-1), X) --> X s> Y ? -1 : X.
2831 {
2832 Value *X, *Y;
2833 const APInt *ShAmt;
2834 Type *Ty = I.getType();
2835 if (match(&I, m_c_Or(m_OneUse(m_AShr(m_NSWSub(m_Value(Y), m_Value(X)),
2836 m_APInt(ShAmt))),
2837 m_Deferred(X))) &&
2838 *ShAmt == Ty->getScalarSizeInBits() - 1) {
2839 Value *NewICmpInst = Builder.CreateICmpSGT(X, Y);
2840 return SelectInst::Create(NewICmpInst, ConstantInt::getAllOnesValue(Ty),
2841 X);
2842 }
2843 }
2844
2845 if (Instruction *V =
2846 canonicalizeCondSignextOfHighBitExtractToSignextHighBitExtract(I))
2847 return V;
2848
2849 CmpInst::Predicate Pred;
2850 Value *Mul, *Ov, *MulIsNotZero, *UMulWithOv;
2851 // Check if the OR weakens the overflow condition for umul.with.overflow by
2852 // treating any non-zero result as overflow. In that case, we overflow if both
2853 // umul.with.overflow operands are != 0, as in that case the result can only
2854 // be 0, iff the multiplication overflows.
2855 if (match(&I,
2856 m_c_Or(m_CombineAnd(m_ExtractValue<1>(m_Value(UMulWithOv)),
2857 m_Value(Ov)),
2858 m_CombineAnd(m_ICmp(Pred,
2859 m_CombineAnd(m_ExtractValue<0>(
2860 m_Deferred(UMulWithOv)),
2861 m_Value(Mul)),
2862 m_ZeroInt()),
2863 m_Value(MulIsNotZero)))) &&
2864 (Ov->hasOneUse() || (MulIsNotZero->hasOneUse() && Mul->hasOneUse())) &&
2865 Pred == CmpInst::ICMP_NE) {
2866 Value *A, *B;
2867 if (match(UMulWithOv, m_Intrinsic<Intrinsic::umul_with_overflow>(
2868 m_Value(A), m_Value(B)))) {
2869 Value *NotNullA = Builder.CreateIsNotNull(A);
2870 Value *NotNullB = Builder.CreateIsNotNull(B);
2871 return BinaryOperator::CreateAnd(NotNullA, NotNullB);
2872 }
2873 }
2874
2875 return nullptr;
2876 }
2877
2878 /// A ^ B can be specified using other logic ops in a variety of patterns. We
2879 /// can fold these early and efficiently by morphing an existing instruction.
foldXorToXor(BinaryOperator & I,InstCombiner::BuilderTy & Builder)2880 static Instruction *foldXorToXor(BinaryOperator &I,
2881 InstCombiner::BuilderTy &Builder) {
2882 assert(I.getOpcode() == Instruction::Xor);
2883 Value *Op0 = I.getOperand(0);
2884 Value *Op1 = I.getOperand(1);
2885 Value *A, *B;
2886
2887 // There are 4 commuted variants for each of the basic patterns.
2888
2889 // (A & B) ^ (A | B) -> A ^ B
2890 // (A & B) ^ (B | A) -> A ^ B
2891 // (A | B) ^ (A & B) -> A ^ B
2892 // (A | B) ^ (B & A) -> A ^ B
2893 if (match(&I, m_c_Xor(m_And(m_Value(A), m_Value(B)),
2894 m_c_Or(m_Deferred(A), m_Deferred(B)))))
2895 return BinaryOperator::CreateXor(A, B);
2896
2897 // (A | ~B) ^ (~A | B) -> A ^ B
2898 // (~B | A) ^ (~A | B) -> A ^ B
2899 // (~A | B) ^ (A | ~B) -> A ^ B
2900 // (B | ~A) ^ (A | ~B) -> A ^ B
2901 if (match(&I, m_Xor(m_c_Or(m_Value(A), m_Not(m_Value(B))),
2902 m_c_Or(m_Not(m_Deferred(A)), m_Deferred(B)))))
2903 return BinaryOperator::CreateXor(A, B);
2904
2905 // (A & ~B) ^ (~A & B) -> A ^ B
2906 // (~B & A) ^ (~A & B) -> A ^ B
2907 // (~A & B) ^ (A & ~B) -> A ^ B
2908 // (B & ~A) ^ (A & ~B) -> A ^ B
2909 if (match(&I, m_Xor(m_c_And(m_Value(A), m_Not(m_Value(B))),
2910 m_c_And(m_Not(m_Deferred(A)), m_Deferred(B)))))
2911 return BinaryOperator::CreateXor(A, B);
2912
2913 // For the remaining cases we need to get rid of one of the operands.
2914 if (!Op0->hasOneUse() && !Op1->hasOneUse())
2915 return nullptr;
2916
2917 // (A | B) ^ ~(A & B) -> ~(A ^ B)
2918 // (A | B) ^ ~(B & A) -> ~(A ^ B)
2919 // (A & B) ^ ~(A | B) -> ~(A ^ B)
2920 // (A & B) ^ ~(B | A) -> ~(A ^ B)
2921 // Complexity sorting ensures the not will be on the right side.
2922 if ((match(Op0, m_Or(m_Value(A), m_Value(B))) &&
2923 match(Op1, m_Not(m_c_And(m_Specific(A), m_Specific(B))))) ||
2924 (match(Op0, m_And(m_Value(A), m_Value(B))) &&
2925 match(Op1, m_Not(m_c_Or(m_Specific(A), m_Specific(B))))))
2926 return BinaryOperator::CreateNot(Builder.CreateXor(A, B));
2927
2928 return nullptr;
2929 }
2930
foldXorOfICmps(ICmpInst * LHS,ICmpInst * RHS,BinaryOperator & I)2931 Value *InstCombiner::foldXorOfICmps(ICmpInst *LHS, ICmpInst *RHS,
2932 BinaryOperator &I) {
2933 assert(I.getOpcode() == Instruction::Xor && I.getOperand(0) == LHS &&
2934 I.getOperand(1) == RHS && "Should be 'xor' with these operands");
2935
2936 if (predicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
2937 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2938 LHS->getOperand(1) == RHS->getOperand(0))
2939 LHS->swapOperands();
2940 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2941 LHS->getOperand(1) == RHS->getOperand(1)) {
2942 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
2943 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2944 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
2945 bool IsSigned = LHS->isSigned() || RHS->isSigned();
2946 return getNewICmpValue(Code, IsSigned, Op0, Op1, Builder);
2947 }
2948 }
2949
2950 // TODO: This can be generalized to compares of non-signbits using
2951 // decomposeBitTestICmp(). It could be enhanced more by using (something like)
2952 // foldLogOpOfMaskedICmps().
2953 ICmpInst::Predicate PredL = LHS->getPredicate(), PredR = RHS->getPredicate();
2954 Value *LHS0 = LHS->getOperand(0), *LHS1 = LHS->getOperand(1);
2955 Value *RHS0 = RHS->getOperand(0), *RHS1 = RHS->getOperand(1);
2956 if ((LHS->hasOneUse() || RHS->hasOneUse()) &&
2957 LHS0->getType() == RHS0->getType() &&
2958 LHS0->getType()->isIntOrIntVectorTy()) {
2959 // (X > -1) ^ (Y > -1) --> (X ^ Y) < 0
2960 // (X < 0) ^ (Y < 0) --> (X ^ Y) < 0
2961 if ((PredL == CmpInst::ICMP_SGT && match(LHS1, m_AllOnes()) &&
2962 PredR == CmpInst::ICMP_SGT && match(RHS1, m_AllOnes())) ||
2963 (PredL == CmpInst::ICMP_SLT && match(LHS1, m_Zero()) &&
2964 PredR == CmpInst::ICMP_SLT && match(RHS1, m_Zero()))) {
2965 Value *Zero = ConstantInt::getNullValue(LHS0->getType());
2966 return Builder.CreateICmpSLT(Builder.CreateXor(LHS0, RHS0), Zero);
2967 }
2968 // (X > -1) ^ (Y < 0) --> (X ^ Y) > -1
2969 // (X < 0) ^ (Y > -1) --> (X ^ Y) > -1
2970 if ((PredL == CmpInst::ICMP_SGT && match(LHS1, m_AllOnes()) &&
2971 PredR == CmpInst::ICMP_SLT && match(RHS1, m_Zero())) ||
2972 (PredL == CmpInst::ICMP_SLT && match(LHS1, m_Zero()) &&
2973 PredR == CmpInst::ICMP_SGT && match(RHS1, m_AllOnes()))) {
2974 Value *MinusOne = ConstantInt::getAllOnesValue(LHS0->getType());
2975 return Builder.CreateICmpSGT(Builder.CreateXor(LHS0, RHS0), MinusOne);
2976 }
2977 }
2978
2979 // Instead of trying to imitate the folds for and/or, decompose this 'xor'
2980 // into those logic ops. That is, try to turn this into an and-of-icmps
2981 // because we have many folds for that pattern.
2982 //
2983 // This is based on a truth table definition of xor:
2984 // X ^ Y --> (X | Y) & !(X & Y)
2985 if (Value *OrICmp = SimplifyBinOp(Instruction::Or, LHS, RHS, SQ)) {
2986 // TODO: If OrICmp is true, then the definition of xor simplifies to !(X&Y).
2987 // TODO: If OrICmp is false, the whole thing is false (InstSimplify?).
2988 if (Value *AndICmp = SimplifyBinOp(Instruction::And, LHS, RHS, SQ)) {
2989 // TODO: Independently handle cases where the 'and' side is a constant.
2990 ICmpInst *X = nullptr, *Y = nullptr;
2991 if (OrICmp == LHS && AndICmp == RHS) {
2992 // (LHS | RHS) & !(LHS & RHS) --> LHS & !RHS --> X & !Y
2993 X = LHS;
2994 Y = RHS;
2995 }
2996 if (OrICmp == RHS && AndICmp == LHS) {
2997 // !(LHS & RHS) & (LHS | RHS) --> !LHS & RHS --> !Y & X
2998 X = RHS;
2999 Y = LHS;
3000 }
3001 if (X && Y && (Y->hasOneUse() || canFreelyInvertAllUsersOf(Y, &I))) {
3002 // Invert the predicate of 'Y', thus inverting its output.
3003 Y->setPredicate(Y->getInversePredicate());
3004 // So, are there other uses of Y?
3005 if (!Y->hasOneUse()) {
3006 // We need to adapt other uses of Y though. Get a value that matches
3007 // the original value of Y before inversion. While this increases
3008 // immediate instruction count, we have just ensured that all the
3009 // users are freely-invertible, so that 'not' *will* get folded away.
3010 BuilderTy::InsertPointGuard Guard(Builder);
3011 // Set insertion point to right after the Y.
3012 Builder.SetInsertPoint(Y->getParent(), ++(Y->getIterator()));
3013 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");
3014 // Replace all uses of Y (excluding the one in NotY!) with NotY.
3015 Worklist.pushUsersToWorkList(*Y);
3016 Y->replaceUsesWithIf(NotY,
3017 [NotY](Use &U) { return U.getUser() != NotY; });
3018 }
3019 // All done.
3020 return Builder.CreateAnd(LHS, RHS);
3021 }
3022 }
3023 }
3024
3025 return nullptr;
3026 }
3027
3028 /// If we have a masked merge, in the canonical form of:
3029 /// (assuming that A only has one use.)
3030 /// | A | |B|
3031 /// ((x ^ y) & M) ^ y
3032 /// | D |
3033 /// * If M is inverted:
3034 /// | D |
3035 /// ((x ^ y) & ~M) ^ y
3036 /// We can canonicalize by swapping the final xor operand
3037 /// to eliminate the 'not' of the mask.
3038 /// ((x ^ y) & M) ^ x
3039 /// * If M is a constant, and D has one use, we transform to 'and' / 'or' ops
3040 /// because that shortens the dependency chain and improves analysis:
3041 /// (x & M) | (y & ~M)
visitMaskedMerge(BinaryOperator & I,InstCombiner::BuilderTy & Builder)3042 static Instruction *visitMaskedMerge(BinaryOperator &I,
3043 InstCombiner::BuilderTy &Builder) {
3044 Value *B, *X, *D;
3045 Value *M;
3046 if (!match(&I, m_c_Xor(m_Value(B),
3047 m_OneUse(m_c_And(
3048 m_CombineAnd(m_c_Xor(m_Deferred(B), m_Value(X)),
3049 m_Value(D)),
3050 m_Value(M))))))
3051 return nullptr;
3052
3053 Value *NotM;
3054 if (match(M, m_Not(m_Value(NotM)))) {
3055 // De-invert the mask and swap the value in B part.
3056 Value *NewA = Builder.CreateAnd(D, NotM);
3057 return BinaryOperator::CreateXor(NewA, X);
3058 }
3059
3060 Constant *C;
3061 if (D->hasOneUse() && match(M, m_Constant(C))) {
3062 // Propagating undef is unsafe. Clamp undef elements to -1.
3063 Type *EltTy = C->getType()->getScalarType();
3064 C = Constant::replaceUndefsWith(C, ConstantInt::getAllOnesValue(EltTy));
3065 // Unfold.
3066 Value *LHS = Builder.CreateAnd(X, C);
3067 Value *NotC = Builder.CreateNot(C);
3068 Value *RHS = Builder.CreateAnd(B, NotC);
3069 return BinaryOperator::CreateOr(LHS, RHS);
3070 }
3071
3072 return nullptr;
3073 }
3074
3075 // Transform
3076 // ~(x ^ y)
3077 // into:
3078 // (~x) ^ y
3079 // or into
3080 // x ^ (~y)
sinkNotIntoXor(BinaryOperator & I,InstCombiner::BuilderTy & Builder)3081 static Instruction *sinkNotIntoXor(BinaryOperator &I,
3082 InstCombiner::BuilderTy &Builder) {
3083 Value *X, *Y;
3084 // FIXME: one-use check is not needed in general, but currently we are unable
3085 // to fold 'not' into 'icmp', if that 'icmp' has multiple uses. (D35182)
3086 if (!match(&I, m_Not(m_OneUse(m_Xor(m_Value(X), m_Value(Y))))))
3087 return nullptr;
3088
3089 // We only want to do the transform if it is free to do.
3090 if (isFreeToInvert(X, X->hasOneUse())) {
3091 // Ok, good.
3092 } else if (isFreeToInvert(Y, Y->hasOneUse())) {
3093 std::swap(X, Y);
3094 } else
3095 return nullptr;
3096
3097 Value *NotX = Builder.CreateNot(X, X->getName() + ".not");
3098 return BinaryOperator::CreateXor(NotX, Y, I.getName() + ".demorgan");
3099 }
3100
3101 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
3102 // here. We should standardize that construct where it is needed or choose some
3103 // other way to ensure that commutated variants of patterns are not missed.
visitXor(BinaryOperator & I)3104 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3105 if (Value *V = SimplifyXorInst(I.getOperand(0), I.getOperand(1),
3106 SQ.getWithInstruction(&I)))
3107 return replaceInstUsesWith(I, V);
3108
3109 if (SimplifyAssociativeOrCommutative(I))
3110 return &I;
3111
3112 if (Instruction *X = foldVectorBinop(I))
3113 return X;
3114
3115 if (Instruction *NewXor = foldXorToXor(I, Builder))
3116 return NewXor;
3117
3118 // (A&B)^(A&C) -> A&(B^C) etc
3119 if (Value *V = SimplifyUsingDistributiveLaws(I))
3120 return replaceInstUsesWith(I, V);
3121
3122 // See if we can simplify any instructions used by the instruction whose sole
3123 // purpose is to compute bits we don't care about.
3124 if (SimplifyDemandedInstructionBits(I))
3125 return &I;
3126
3127 if (Value *V = SimplifyBSwap(I, Builder))
3128 return replaceInstUsesWith(I, V);
3129
3130 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3131
3132 // Fold (X & M) ^ (Y & ~M) -> (X & M) | (Y & ~M)
3133 // This it a special case in haveNoCommonBitsSet, but the computeKnownBits
3134 // calls in there are unnecessary as SimplifyDemandedInstructionBits should
3135 // have already taken care of those cases.
3136 Value *M;
3137 if (match(&I, m_c_Xor(m_c_And(m_Not(m_Value(M)), m_Value()),
3138 m_c_And(m_Deferred(M), m_Value()))))
3139 return BinaryOperator::CreateOr(Op0, Op1);
3140
3141 // Apply DeMorgan's Law for 'nand' / 'nor' logic with an inverted operand.
3142 Value *X, *Y;
3143
3144 // We must eliminate the and/or (one-use) for these transforms to not increase
3145 // the instruction count.
3146 // ~(~X & Y) --> (X | ~Y)
3147 // ~(Y & ~X) --> (X | ~Y)
3148 if (match(&I, m_Not(m_OneUse(m_c_And(m_Not(m_Value(X)), m_Value(Y)))))) {
3149 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");
3150 return BinaryOperator::CreateOr(X, NotY);
3151 }
3152 // ~(~X | Y) --> (X & ~Y)
3153 // ~(Y | ~X) --> (X & ~Y)
3154 if (match(&I, m_Not(m_OneUse(m_c_Or(m_Not(m_Value(X)), m_Value(Y)))))) {
3155 Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");
3156 return BinaryOperator::CreateAnd(X, NotY);
3157 }
3158
3159 if (Instruction *Xor = visitMaskedMerge(I, Builder))
3160 return Xor;
3161
3162 // Is this a 'not' (~) fed by a binary operator?
3163 BinaryOperator *NotVal;
3164 if (match(&I, m_Not(m_BinOp(NotVal)))) {
3165 if (NotVal->getOpcode() == Instruction::And ||
3166 NotVal->getOpcode() == Instruction::Or) {
3167 // Apply DeMorgan's Law when inverts are free:
3168 // ~(X & Y) --> (~X | ~Y)
3169 // ~(X | Y) --> (~X & ~Y)
3170 if (isFreeToInvert(NotVal->getOperand(0),
3171 NotVal->getOperand(0)->hasOneUse()) &&
3172 isFreeToInvert(NotVal->getOperand(1),
3173 NotVal->getOperand(1)->hasOneUse())) {
3174 Value *NotX = Builder.CreateNot(NotVal->getOperand(0), "notlhs");
3175 Value *NotY = Builder.CreateNot(NotVal->getOperand(1), "notrhs");
3176 if (NotVal->getOpcode() == Instruction::And)
3177 return BinaryOperator::CreateOr(NotX, NotY);
3178 return BinaryOperator::CreateAnd(NotX, NotY);
3179 }
3180 }
3181
3182 // ~(X - Y) --> ~X + Y
3183 if (match(NotVal, m_Sub(m_Value(X), m_Value(Y))))
3184 if (isa<Constant>(X) || NotVal->hasOneUse())
3185 return BinaryOperator::CreateAdd(Builder.CreateNot(X), Y);
3186
3187 // ~(~X >>s Y) --> (X >>s Y)
3188 if (match(NotVal, m_AShr(m_Not(m_Value(X)), m_Value(Y))))
3189 return BinaryOperator::CreateAShr(X, Y);
3190
3191 // If we are inverting a right-shifted constant, we may be able to eliminate
3192 // the 'not' by inverting the constant and using the opposite shift type.
3193 // Canonicalization rules ensure that only a negative constant uses 'ashr',
3194 // but we must check that in case that transform has not fired yet.
3195
3196 // ~(C >>s Y) --> ~C >>u Y (when inverting the replicated sign bits)
3197 Constant *C;
3198 if (match(NotVal, m_AShr(m_Constant(C), m_Value(Y))) &&
3199 match(C, m_Negative())) {
3200 // We matched a negative constant, so propagating undef is unsafe.
3201 // Clamp undef elements to -1.
3202 Type *EltTy = C->getType()->getScalarType();
3203 C = Constant::replaceUndefsWith(C, ConstantInt::getAllOnesValue(EltTy));
3204 return BinaryOperator::CreateLShr(ConstantExpr::getNot(C), Y);
3205 }
3206
3207 // ~(C >>u Y) --> ~C >>s Y (when inverting the replicated sign bits)
3208 if (match(NotVal, m_LShr(m_Constant(C), m_Value(Y))) &&
3209 match(C, m_NonNegative())) {
3210 // We matched a non-negative constant, so propagating undef is unsafe.
3211 // Clamp undef elements to 0.
3212 Type *EltTy = C->getType()->getScalarType();
3213 C = Constant::replaceUndefsWith(C, ConstantInt::getNullValue(EltTy));
3214 return BinaryOperator::CreateAShr(ConstantExpr::getNot(C), Y);
3215 }
3216
3217 // ~(X + C) --> -(C + 1) - X
3218 if (match(Op0, m_Add(m_Value(X), m_Constant(C))))
3219 return BinaryOperator::CreateSub(ConstantExpr::getNeg(AddOne(C)), X);
3220 }
3221
3222 // Use DeMorgan and reassociation to eliminate a 'not' op.
3223 Constant *C1;
3224 if (match(Op1, m_Constant(C1))) {
3225 Constant *C2;
3226 if (match(Op0, m_OneUse(m_Or(m_Not(m_Value(X)), m_Constant(C2))))) {
3227 // (~X | C2) ^ C1 --> ((X & ~C2) ^ -1) ^ C1 --> (X & ~C2) ^ ~C1
3228 Value *And = Builder.CreateAnd(X, ConstantExpr::getNot(C2));
3229 return BinaryOperator::CreateXor(And, ConstantExpr::getNot(C1));
3230 }
3231 if (match(Op0, m_OneUse(m_And(m_Not(m_Value(X)), m_Constant(C2))))) {
3232 // (~X & C2) ^ C1 --> ((X | ~C2) ^ -1) ^ C1 --> (X | ~C2) ^ ~C1
3233 Value *Or = Builder.CreateOr(X, ConstantExpr::getNot(C2));
3234 return BinaryOperator::CreateXor(Or, ConstantExpr::getNot(C1));
3235 }
3236 }
3237
3238 // not (cmp A, B) = !cmp A, B
3239 CmpInst::Predicate Pred;
3240 if (match(&I, m_Not(m_OneUse(m_Cmp(Pred, m_Value(), m_Value()))))) {
3241 cast<CmpInst>(Op0)->setPredicate(CmpInst::getInversePredicate(Pred));
3242 return replaceInstUsesWith(I, Op0);
3243 }
3244
3245 {
3246 const APInt *RHSC;
3247 if (match(Op1, m_APInt(RHSC))) {
3248 Value *X;
3249 const APInt *C;
3250 if (RHSC->isSignMask() && match(Op0, m_Sub(m_APInt(C), m_Value(X)))) {
3251 // (C - X) ^ signmask -> (C + signmask - X)
3252 Constant *NewC = ConstantInt::get(I.getType(), *C + *RHSC);
3253 return BinaryOperator::CreateSub(NewC, X);
3254 }
3255 if (RHSC->isSignMask() && match(Op0, m_Add(m_Value(X), m_APInt(C)))) {
3256 // (X + C) ^ signmask -> (X + C + signmask)
3257 Constant *NewC = ConstantInt::get(I.getType(), *C + *RHSC);
3258 return BinaryOperator::CreateAdd(X, NewC);
3259 }
3260
3261 // (X|C1)^C2 -> X^(C1^C2) iff X&~C1 == 0
3262 if (match(Op0, m_Or(m_Value(X), m_APInt(C))) &&
3263 MaskedValueIsZero(X, *C, 0, &I)) {
3264 Constant *NewC = ConstantInt::get(I.getType(), *C ^ *RHSC);
3265 return BinaryOperator::CreateXor(X, NewC);
3266 }
3267 }
3268 }
3269
3270 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(Op1)) {
3271 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3272 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
3273 if (Op0I->getOpcode() == Instruction::LShr) {
3274 // ((X^C1) >> C2) ^ C3 -> (X>>C2) ^ ((C1>>C2)^C3)
3275 // E1 = "X ^ C1"
3276 BinaryOperator *E1;
3277 ConstantInt *C1;
3278 if (Op0I->hasOneUse() &&
3279 (E1 = dyn_cast<BinaryOperator>(Op0I->getOperand(0))) &&
3280 E1->getOpcode() == Instruction::Xor &&
3281 (C1 = dyn_cast<ConstantInt>(E1->getOperand(1)))) {
3282 // fold (C1 >> C2) ^ C3
3283 ConstantInt *C2 = Op0CI, *C3 = RHSC;
3284 APInt FoldConst = C1->getValue().lshr(C2->getValue());
3285 FoldConst ^= C3->getValue();
3286 // Prepare the two operands.
3287 Value *Opnd0 = Builder.CreateLShr(E1->getOperand(0), C2);
3288 Opnd0->takeName(Op0I);
3289 cast<Instruction>(Opnd0)->setDebugLoc(I.getDebugLoc());
3290 Value *FoldVal = ConstantInt::get(Opnd0->getType(), FoldConst);
3291
3292 return BinaryOperator::CreateXor(Opnd0, FoldVal);
3293 }
3294 }
3295 }
3296 }
3297 }
3298
3299 if (Instruction *FoldedLogic = foldBinOpIntoSelectOrPhi(I))
3300 return FoldedLogic;
3301
3302 // Y ^ (X | Y) --> X & ~Y
3303 // Y ^ (Y | X) --> X & ~Y
3304 if (match(Op1, m_OneUse(m_c_Or(m_Value(X), m_Specific(Op0)))))
3305 return BinaryOperator::CreateAnd(X, Builder.CreateNot(Op0));
3306 // (X | Y) ^ Y --> X & ~Y
3307 // (Y | X) ^ Y --> X & ~Y
3308 if (match(Op0, m_OneUse(m_c_Or(m_Value(X), m_Specific(Op1)))))
3309 return BinaryOperator::CreateAnd(X, Builder.CreateNot(Op1));
3310
3311 // Y ^ (X & Y) --> ~X & Y
3312 // Y ^ (Y & X) --> ~X & Y
3313 if (match(Op1, m_OneUse(m_c_And(m_Value(X), m_Specific(Op0)))))
3314 return BinaryOperator::CreateAnd(Op0, Builder.CreateNot(X));
3315 // (X & Y) ^ Y --> ~X & Y
3316 // (Y & X) ^ Y --> ~X & Y
3317 // Canonical form is (X & C) ^ C; don't touch that.
3318 // TODO: A 'not' op is better for analysis and codegen, but demanded bits must
3319 // be fixed to prefer that (otherwise we get infinite looping).
3320 if (!match(Op1, m_Constant()) &&
3321 match(Op0, m_OneUse(m_c_And(m_Value(X), m_Specific(Op1)))))
3322 return BinaryOperator::CreateAnd(Op1, Builder.CreateNot(X));
3323
3324 Value *A, *B, *C;
3325 // (A ^ B) ^ (A | C) --> (~A & C) ^ B -- There are 4 commuted variants.
3326 if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_Value(A), m_Value(B))),
3327 m_OneUse(m_c_Or(m_Deferred(A), m_Value(C))))))
3328 return BinaryOperator::CreateXor(
3329 Builder.CreateAnd(Builder.CreateNot(A), C), B);
3330
3331 // (A ^ B) ^ (B | C) --> (~B & C) ^ A -- There are 4 commuted variants.
3332 if (match(&I, m_c_Xor(m_OneUse(m_Xor(m_Value(A), m_Value(B))),
3333 m_OneUse(m_c_Or(m_Deferred(B), m_Value(C))))))
3334 return BinaryOperator::CreateXor(
3335 Builder.CreateAnd(Builder.CreateNot(B), C), A);
3336
3337 // (A & B) ^ (A ^ B) -> (A | B)
3338 if (match(Op0, m_And(m_Value(A), m_Value(B))) &&
3339 match(Op1, m_c_Xor(m_Specific(A), m_Specific(B))))
3340 return BinaryOperator::CreateOr(A, B);
3341 // (A ^ B) ^ (A & B) -> (A | B)
3342 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
3343 match(Op1, m_c_And(m_Specific(A), m_Specific(B))))
3344 return BinaryOperator::CreateOr(A, B);
3345
3346 // (A & ~B) ^ ~A -> ~(A & B)
3347 // (~B & A) ^ ~A -> ~(A & B)
3348 if (match(Op0, m_c_And(m_Value(A), m_Not(m_Value(B)))) &&
3349 match(Op1, m_Not(m_Specific(A))))
3350 return BinaryOperator::CreateNot(Builder.CreateAnd(A, B));
3351
3352 if (auto *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
3353 if (auto *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
3354 if (Value *V = foldXorOfICmps(LHS, RHS, I))
3355 return replaceInstUsesWith(I, V);
3356
3357 if (Instruction *CastedXor = foldCastedBitwiseLogic(I))
3358 return CastedXor;
3359
3360 // Canonicalize a shifty way to code absolute value to the common pattern.
3361 // There are 4 potential commuted variants. Move the 'ashr' candidate to Op1.
3362 // We're relying on the fact that we only do this transform when the shift has
3363 // exactly 2 uses and the add has exactly 1 use (otherwise, we might increase
3364 // instructions).
3365 if (Op0->hasNUses(2))
3366 std::swap(Op0, Op1);
3367
3368 const APInt *ShAmt;
3369 Type *Ty = I.getType();
3370 if (match(Op1, m_AShr(m_Value(A), m_APInt(ShAmt))) &&
3371 Op1->hasNUses(2) && *ShAmt == Ty->getScalarSizeInBits() - 1 &&
3372 match(Op0, m_OneUse(m_c_Add(m_Specific(A), m_Specific(Op1))))) {
3373 // B = ashr i32 A, 31 ; smear the sign bit
3374 // xor (add A, B), B ; add -1 and flip bits if negative
3375 // --> (A < 0) ? -A : A
3376 Value *Cmp = Builder.CreateICmpSLT(A, ConstantInt::getNullValue(Ty));
3377 // Copy the nuw/nsw flags from the add to the negate.
3378 auto *Add = cast<BinaryOperator>(Op0);
3379 Value *Neg = Builder.CreateNeg(A, "", Add->hasNoUnsignedWrap(),
3380 Add->hasNoSignedWrap());
3381 return SelectInst::Create(Cmp, Neg, A);
3382 }
3383
3384 // Eliminate a bitwise 'not' op of 'not' min/max by inverting the min/max:
3385 //
3386 // %notx = xor i32 %x, -1
3387 // %cmp1 = icmp sgt i32 %notx, %y
3388 // %smax = select i1 %cmp1, i32 %notx, i32 %y
3389 // %res = xor i32 %smax, -1
3390 // =>
3391 // %noty = xor i32 %y, -1
3392 // %cmp2 = icmp slt %x, %noty
3393 // %res = select i1 %cmp2, i32 %x, i32 %noty
3394 //
3395 // Same is applicable for smin/umax/umin.
3396 if (match(Op1, m_AllOnes()) && Op0->hasOneUse()) {
3397 Value *LHS, *RHS;
3398 SelectPatternFlavor SPF = matchSelectPattern(Op0, LHS, RHS).Flavor;
3399 if (SelectPatternResult::isMinOrMax(SPF)) {
3400 // It's possible we get here before the not has been simplified, so make
3401 // sure the input to the not isn't freely invertible.
3402 if (match(LHS, m_Not(m_Value(X))) && !isFreeToInvert(X, X->hasOneUse())) {
3403 Value *NotY = Builder.CreateNot(RHS);
3404 return SelectInst::Create(
3405 Builder.CreateICmp(getInverseMinMaxPred(SPF), X, NotY), X, NotY);
3406 }
3407
3408 // It's possible we get here before the not has been simplified, so make
3409 // sure the input to the not isn't freely invertible.
3410 if (match(RHS, m_Not(m_Value(Y))) && !isFreeToInvert(Y, Y->hasOneUse())) {
3411 Value *NotX = Builder.CreateNot(LHS);
3412 return SelectInst::Create(
3413 Builder.CreateICmp(getInverseMinMaxPred(SPF), NotX, Y), NotX, Y);
3414 }
3415
3416 // If both sides are freely invertible, then we can get rid of the xor
3417 // completely.
3418 if (isFreeToInvert(LHS, !LHS->hasNUsesOrMore(3)) &&
3419 isFreeToInvert(RHS, !RHS->hasNUsesOrMore(3))) {
3420 Value *NotLHS = Builder.CreateNot(LHS);
3421 Value *NotRHS = Builder.CreateNot(RHS);
3422 return SelectInst::Create(
3423 Builder.CreateICmp(getInverseMinMaxPred(SPF), NotLHS, NotRHS),
3424 NotLHS, NotRHS);
3425 }
3426 }
3427
3428 // Pull 'not' into operands of select if both operands are one-use compares.
3429 // Inverting the predicates eliminates the 'not' operation.
3430 // Example:
3431 // not (select ?, (cmp TPred, ?, ?), (cmp FPred, ?, ?) -->
3432 // select ?, (cmp InvTPred, ?, ?), (cmp InvFPred, ?, ?)
3433 // TODO: Canonicalize by hoisting 'not' into an arm of the select if only
3434 // 1 select operand is a cmp?
3435 if (auto *Sel = dyn_cast<SelectInst>(Op0)) {
3436 auto *CmpT = dyn_cast<CmpInst>(Sel->getTrueValue());
3437 auto *CmpF = dyn_cast<CmpInst>(Sel->getFalseValue());
3438 if (CmpT && CmpF && CmpT->hasOneUse() && CmpF->hasOneUse()) {
3439 CmpT->setPredicate(CmpT->getInversePredicate());
3440 CmpF->setPredicate(CmpF->getInversePredicate());
3441 return replaceInstUsesWith(I, Sel);
3442 }
3443 }
3444 }
3445
3446 if (Instruction *NewXor = sinkNotIntoXor(I, Builder))
3447 return NewXor;
3448
3449 return nullptr;
3450 }
3451